You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@tvm.apache.org by tq...@apache.org on 2022/11/16 17:01:48 UTC
[tvm-site] branch asf-site updated: deploying docs (apache/tvm@78b53221f8dd8c1d2bbeff9d34803db33ca254dd)
This is an automated email from the ASF dual-hosted git repository.
tqchen pushed a commit to branch asf-site
in repository https://gitbox.apache.org/repos/asf/tvm-site.git
The following commit(s) were added to refs/heads/asf-site by this push:
new 3256cd004b deploying docs (apache/tvm@78b53221f8dd8c1d2bbeff9d34803db33ca254dd)
3256cd004b is described below
commit 3256cd004b4005697988e27198343e46cb08db68
Author: tvm-bot <95...@users.noreply.github.com>
AuthorDate: Wed Nov 16 17:01:41 2022 +0000
deploying docs (apache/tvm@78b53221f8dd8c1d2bbeff9d34803db33ca254dd)
---
docs/_images/sphx_glr_micro_train_001.png | Bin 344037 -> 327199 bytes
docs/_images/sphx_glr_micro_train_thumb.png | Bin 24159 -> 22934 bytes
.../how_to/compile_models/from_darknet.rst.txt | 2 +-
.../how_to/compile_models/from_keras.rst.txt | 2 +-
.../how_to/compile_models/from_mxnet.rst.txt | 2 +-
.../how_to/compile_models/from_oneflow.rst.txt | 2 +-
.../how_to/compile_models/from_pytorch.rst.txt | 2 +-
.../how_to/compile_models/from_tensorflow.rst.txt | 2 +-
.../compile_models/sg_execution_times.rst.txt | 22 +-
.../deploy_models/deploy_model_on_android.rst.txt | 2 +-
.../deploy_object_detection_pytorch.rst.txt | 4 +-
.../deploy_models/deploy_prequantized.rst.txt | 6 +-
.../deploy_prequantized_tflite.rst.txt | 4 +-
.../how_to/deploy_models/deploy_quantized.rst.txt | 2 +-
.../deploy_models/deploy_ssd_gluoncv.rst.txt | 4 +-
.../deploy_models/sg_execution_times.rst.txt | 18 +-
.../extend_tvm/bring_your_own_datatypes.rst.txt | 2 +-
.../extend_tvm/low_level_custom_pass.rst.txt | 10 +-
.../how_to/extend_tvm/sg_execution_times.rst.txt | 8 +-
.../how_to/extend_tvm/use_pass_instrument.rst.txt | 16 +-
.../optimize_operators/opt_conv_cuda.rst.txt | 2 +-
.../optimize_operators/opt_conv_tensorcore.rst.txt | 30 +-
.../how_to/optimize_operators/opt_gemm.rst.txt | 125 ++-
.../optimize_operators/sg_execution_times.rst.txt | 8 +-
.../sg_execution_times.rst.txt | 14 +-
.../tune_conv2d_layer_cuda.rst.txt | 1121 ++++++++++++++++----
.../tune_network_cuda.rst.txt | 4 +-
.../tune_network_x86.rst.txt | 4 +-
.../tune_sparse_x86.rst.txt | 427 +-------
.../tune_with_autotvm/sg_execution_times.rst.txt | 8 +-
.../tune_with_autotvm/tune_conv2d_cuda.rst.txt | 678 +-----------
.../work_with_microtvm/micro_autotune.rst.txt | 16 +-
.../work_with_microtvm/micro_pytorch.rst.txt | 4 +-
.../how_to/work_with_microtvm/micro_train.rst.txt | 18 +-
.../work_with_microtvm/sg_execution_times.rst.txt | 12 +-
.../work_with_relay/sg_execution_times.rst.txt | 8 +-
.../how_to/work_with_schedules/intrin_math.rst.txt | 2 +-
.../how_to/work_with_schedules/reduction.rst.txt | 59 +-
.../how_to/work_with_schedules/scan.rst.txt | 41 +-
.../schedule_primitives.rst.txt | 115 +-
.../work_with_schedules/sg_execution_times.rst.txt | 18 +-
.../how_to/work_with_schedules/tensorize.rst.txt | 46 +-
.../work_with_schedules/tuple_inputs.rst.txt | 49 +-
.../tutorials/autotvm/sg_execution_times.rst.txt | 4 +-
.../frontend/deploy_classification.rst.txt | 2 +-
.../tutorials/frontend/deploy_detection.rst.txt | 2 +-
.../tutorials/frontend/sg_execution_times.rst.txt | 6 +-
.../topic/vta/tutorials/matrix_multiply.rst.txt | 39 +-
.../vta/tutorials/optimize/convolution_opt.rst.txt | 41 +-
.../tutorials/optimize/matrix_multiply_opt.rst.txt | 39 +-
.../tutorials/optimize/sg_execution_times.rst.txt | 6 +-
.../topic/vta/tutorials/sg_execution_times.rst.txt | 6 +-
.../topic/vta/tutorials/vta_get_started.rst.txt | 24 +-
.../tutorial/auto_scheduler_matmul_x86.rst.txt | 34 +-
docs/_sources/tutorial/autotvm_matmul_x86.rst.txt | 20 +-
docs/_sources/tutorial/autotvm_relay_x86.rst.txt | 59 +-
.../tutorial/cross_compilation_and_rpc.rst.txt | 2 +-
docs/_sources/tutorial/intro_topi.rst.txt | 121 +--
docs/_sources/tutorial/sg_execution_times.rst.txt | 22 +-
.../tutorial/tensor_expr_get_started.rst.txt | 175 ++-
docs/commit_hash | 2 +-
docs/how_to/compile_models/from_darknet.html | 2 +-
docs/how_to/compile_models/from_keras.html | 2 +-
docs/how_to/compile_models/from_mxnet.html | 2 +-
docs/how_to/compile_models/from_oneflow.html | 13 +-
docs/how_to/compile_models/from_pytorch.html | 11 +-
docs/how_to/compile_models/from_tensorflow.html | 2 +-
docs/how_to/compile_models/sg_execution_times.html | 26 +-
.../deploy_models/deploy_model_on_android.html | 2 +-
.../deploy_object_detection_pytorch.html | 49 +-
docs/how_to/deploy_models/deploy_prequantized.html | 8 +-
.../deploy_models/deploy_prequantized_tflite.html | 4 +-
docs/how_to/deploy_models/deploy_quantized.html | 2 +-
docs/how_to/deploy_models/deploy_ssd_gluoncv.html | 36 +-
docs/how_to/deploy_models/sg_execution_times.html | 18 +-
.../extend_tvm/bring_your_own_datatypes.html | 2 +-
docs/how_to/extend_tvm/low_level_custom_pass.html | 10 +-
docs/how_to/extend_tvm/sg_execution_times.html | 8 +-
docs/how_to/extend_tvm/use_pass_instrument.html | 16 +-
docs/how_to/optimize_operators/opt_conv_cuda.html | 2 +-
.../optimize_operators/opt_conv_tensorcore.html | 30 +-
docs/how_to/optimize_operators/opt_gemm.html | 125 ++-
.../optimize_operators/sg_execution_times.html | 8 +-
.../sg_execution_times.html | 14 +-
.../tune_conv2d_layer_cuda.html | 1121 ++++++++++++++++----
.../tune_with_autoscheduler/tune_network_cuda.html | 4 +-
.../tune_with_autoscheduler/tune_network_x86.html | 4 +-
.../tune_with_autoscheduler/tune_sparse_x86.html | 427 +-------
.../tune_with_autotvm/sg_execution_times.html | 10 +-
.../how_to/tune_with_autotvm/tune_conv2d_cuda.html | 678 +-----------
docs/how_to/work_with_microtvm/micro_autotune.html | 16 +-
docs/how_to/work_with_microtvm/micro_pytorch.html | 4 +-
docs/how_to/work_with_microtvm/micro_train.html | 16 +-
.../work_with_microtvm/sg_execution_times.html | 12 +-
.../how_to/work_with_relay/sg_execution_times.html | 8 +-
docs/how_to/work_with_schedules/intrin_math.html | 2 +-
docs/how_to/work_with_schedules/reduction.html | 59 +-
docs/how_to/work_with_schedules/scan.html | 41 +-
.../work_with_schedules/schedule_primitives.html | 115 +-
.../work_with_schedules/sg_execution_times.html | 18 +-
docs/how_to/work_with_schedules/tensorize.html | 46 +-
docs/how_to/work_with_schedules/tuple_inputs.html | 49 +-
docs/install/nnpack.html | 12 +-
.../reference/api/doxygen/arg__info_8h_source.html | 2 +-
.../api/doxygen/classtvm_1_1BaseFuncNode.html | 2 +-
.../classtvm_1_1BaseFuncNode__inherit__graph.svg | 207 ++--
.../api/doxygen/classtvm_1_1RelayExprNode.html | 2 +-
.../classtvm_1_1RelayExprNode__inherit__graph.svg | 635 ++++++-----
...ilder_1_1IRBuilderFrameNode__inherit__graph.svg | 35 +-
...uilder_1_1tir_1_1PrimFuncFrameNode-members.html | 35 +-
...1_1ir__builder_1_1tir_1_1PrimFuncFrameNode.html | 23 +-
...er_1_1tir_1_1PrimFuncFrameNode__coll__graph.svg | 54 +-
...1_1tir_1_1PrimFuncFrameNode__inherit__graph.svg | 149 ++-
...lder_1_1tir_1_1TIRFrameNode__inherit__graph.svg | 35 +-
.../classtvm_1_1tir_1_1PrimFunc-members.html | 2 +-
.../api/doxygen/classtvm_1_1tir_1_1PrimFunc.html | 17 +-
.../classtvm_1_1tir_1_1PrimFuncNode-members.html | 43 +-
.../doxygen/classtvm_1_1tir_1_1PrimFuncNode.html | 28 +-
...lasstvm_1_1tir_1_1PrimFuncNode__coll__graph.svg | 10 +-
...stvm_1_1tir_1_1PrimFuncNode__inherit__graph.svg | 175 ++-
.../api/doxygen/driver__api_8h_source.html | 2 +-
.../api/doxygen/extracted__task_8h_source.html | 2 +-
docs/reference/api/doxygen/functions_func_p.html | 2 +-
docs/reference/api/doxygen/functions_func_r.html | 2 +-
docs/reference/api/doxygen/functions_func_s.html | 2 +-
docs/reference/api/doxygen/functions_func_u.html | 2 +-
docs/reference/api/doxygen/functions_p.html | 8 +-
docs/reference/api/doxygen/functions_s.html | 4 +-
docs/reference/api/doxygen/functions_t.html | 12 +-
docs/reference/api/doxygen/functions_v.html | 6 +-
docs/reference/api/doxygen/functions_vars_p.html | 4 -
.../doxygen/ir__builder_2tir_2frame_8h_source.html | 185 ++--
.../api/doxygen/namespacemembers_func_p.html | 5 +-
docs/reference/api/doxygen/namespacemembers_p.html | 5 +-
...mespacetvm_1_1script_1_1ir__builder_1_1tir.html | 103 --
docs/reference/api/doxygen/search/all_11.js | 6 +-
docs/reference/api/doxygen/search/all_13.js | 10 +-
docs/reference/api/doxygen/search/all_14.js | 14 +-
docs/reference/api/doxygen/search/all_15.js | 8 +-
docs/reference/api/doxygen/search/all_16.js | 2 +-
docs/reference/api/doxygen/search/all_17.js | 2 +-
docs/reference/api/doxygen/search/all_18.js | 2 +-
docs/reference/api/doxygen/search/functions_10.js | 5 +-
docs/reference/api/doxygen/search/functions_12.js | 4 +-
docs/reference/api/doxygen/search/functions_13.js | 4 +-
docs/reference/api/doxygen/search/functions_15.js | 2 +-
docs/reference/api/doxygen/search/functions_16.js | 2 +-
docs/reference/api/doxygen/search/functions_17.js | 2 +-
docs/reference/api/doxygen/search/typedefs_f.js | 2 +-
docs/reference/api/doxygen/search/variables_f.js | 1 -
.../api/doxygen/stmt__functor_8h_source.html | 2 +-
.../api/doxygen/tir_2analysis_8h_source.html | 2 +-
.../api/doxygen/tir_2function_8h_source.html | 35 +-
docs/reference/api/doxygen/tir_2ir_8h.html | 9 +-
docs/reference/api/doxygen/tir_2ir_8h_source.html | 41 +-
docs/reference/api/python/auto_scheduler.html | 4 +-
docs/reference/api/python/tir.html | 6 +-
.../api/typedoc/classes/bytestreamreader.html | 12 +-
.../api/typedoc/classes/cachedcallstack.html | 34 +-
docs/reference/api/typedoc/classes/dldatatype.html | 12 +-
docs/reference/api/typedoc/classes/dldevice.html | 10 +-
.../reference/api/typedoc/classes/environment.html | 12 +-
docs/reference/api/typedoc/classes/ffilibrary.html | 20 +-
.../api/typedoc/classes/graphexecutor.html | 16 +-
docs/reference/api/typedoc/classes/instance.html | 40 +-
docs/reference/api/typedoc/classes/memory.html | 34 +-
docs/reference/api/typedoc/classes/module.html | 10 +-
docs/reference/api/typedoc/classes/ndarray.html | 22 +-
.../api/typedoc/classes/packedfunccell.html | 6 +-
docs/reference/api/typedoc/classes/rpcserver.html | 14 +-
docs/reference/api/typedoc/classes/scalar.html | 6 +-
.../api/typedoc/classes/webgpucontext.html | 12 +-
docs/reference/api/typedoc/enums/argtypecode.html | 30 +-
.../api/typedoc/enums/aynccallbackcode.html | 4 +-
.../api/typedoc/enums/dldatatypecode.html | 8 +-
.../api/typedoc/enums/rpcserverstate.html | 12 +-
docs/reference/api/typedoc/enums/sizeof.html | 18 +-
docs/reference/api/typedoc/index.html | 112 +-
.../api/typedoc/interfaces/disposable.html | 2 +-
.../api/typedoc/interfaces/functioninfo.html | 6 +-
.../api/typedoc/interfaces/libraryprovider.html | 4 +-
docs/searchindex.js | 2 +-
.../vta/tutorials/autotvm/sg_execution_times.html | 4 +-
.../tutorials/frontend/deploy_classification.html | 2 +-
.../vta/tutorials/frontend/deploy_detection.html | 2 +-
.../vta/tutorials/frontend/sg_execution_times.html | 6 +-
docs/topic/vta/tutorials/matrix_multiply.html | 39 +-
.../vta/tutorials/optimize/convolution_opt.html | 41 +-
.../tutorials/optimize/matrix_multiply_opt.html | 39 +-
.../vta/tutorials/optimize/sg_execution_times.html | 6 +-
docs/topic/vta/tutorials/sg_execution_times.html | 6 +-
docs/topic/vta/tutorials/vta_get_started.html | 24 +-
docs/tutorial/auto_scheduler_matmul_x86.html | 30 +-
docs/tutorial/autotvm_matmul_x86.html | 20 +-
docs/tutorial/autotvm_relay_x86.html | 272 +++--
docs/tutorial/cross_compilation_and_rpc.html | 2 +-
docs/tutorial/intro_topi.html | 121 +--
docs/tutorial/sg_execution_times.html | 22 +-
docs/tutorial/tensor_expr_get_started.html | 175 ++-
199 files changed, 4468 insertions(+), 5321 deletions(-)
diff --git a/docs/_images/sphx_glr_micro_train_001.png b/docs/_images/sphx_glr_micro_train_001.png
index 457bf652be..9acba7fd3b 100644
Binary files a/docs/_images/sphx_glr_micro_train_001.png and b/docs/_images/sphx_glr_micro_train_001.png differ
diff --git a/docs/_images/sphx_glr_micro_train_thumb.png b/docs/_images/sphx_glr_micro_train_thumb.png
index 13642b47f1..fb0f49ab60 100644
Binary files a/docs/_images/sphx_glr_micro_train_thumb.png and b/docs/_images/sphx_glr_micro_train_thumb.png differ
diff --git a/docs/_sources/how_to/compile_models/from_darknet.rst.txt b/docs/_sources/how_to/compile_models/from_darknet.rst.txt
index 9d952705a2..7a1b9de483 100644
--- a/docs/_sources/how_to/compile_models/from_darknet.rst.txt
+++ b/docs/_sources/how_to/compile_models/from_darknet.rst.txt
@@ -315,7 +315,7 @@ The process is no different from other examples.
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 1 minutes 13.973 seconds)
+ **Total running time of the script:** ( 1 minutes 13.629 seconds)
.. _sphx_glr_download_how_to_compile_models_from_darknet.py:
diff --git a/docs/_sources/how_to/compile_models/from_keras.rst.txt b/docs/_sources/how_to/compile_models/from_keras.rst.txt
index 03fd24086a..35ff6fe2ee 100644
--- a/docs/_sources/how_to/compile_models/from_keras.rst.txt
+++ b/docs/_sources/how_to/compile_models/from_keras.rst.txt
@@ -228,7 +228,7 @@ Look up prediction top 1 index in 1000 class synset.
.. code-block:: none
Relay top-1 id: 285, class name: Egyptian cat
-
1/1 [==============================] - ETA: 0s
1/1 [==============================] - 1s 974ms/step
+
1/1 [==============================] - ETA: 0s
1/1 [==============================] - 1s 958ms/step
Keras top-1 id: 285, class name: Egyptian cat
diff --git a/docs/_sources/how_to/compile_models/from_mxnet.rst.txt b/docs/_sources/how_to/compile_models/from_mxnet.rst.txt
index f122cb3349..cd89d79ae4 100644
--- a/docs/_sources/how_to/compile_models/from_mxnet.rst.txt
+++ b/docs/_sources/how_to/compile_models/from_mxnet.rst.txt
@@ -115,7 +115,7 @@ In this section, we download a pretrained imagenet model and classify an image.
.. code-block:: none
- Downloading /workspace/.mxnet/models/resnet18_v1-a0666292.zipf7e8c1d4-467d-4397-b979-f49ad05f57c3 from https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/models/resnet18_v1-a0666292.zip...
+ Downloading /workspace/.mxnet/models/resnet18_v1-a0666292.zip6172288d-0a35-409b-9577-c9b5f252dbdd from https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/models/resnet18_v1-a0666292.zip...
x (1, 3, 224, 224)
diff --git a/docs/_sources/how_to/compile_models/from_oneflow.rst.txt b/docs/_sources/how_to/compile_models/from_oneflow.rst.txt
index 242a91007a..617b4fbb2a 100644
--- a/docs/_sources/how_to/compile_models/from_oneflow.rst.txt
+++ b/docs/_sources/how_to/compile_models/from_oneflow.rst.txt
@@ -116,7 +116,7 @@ Load a pretrained OneFlow model and save model
.. code-block:: none
Downloading: "https://oneflow-public.oss-cn-beijing.aliyuncs.com/model_zoo/flowvision/classification/ResNet/resnet18.zip" to /workspace/.oneflow/flowvision_cache/resnet18.zip
-
0%| | 0.00/41.5M [00:00<?, ?B/s]
19%|#9 | 7.99M/41.5M [00:00<00:00, 37.1MB/s]
39%|###8 | 16.0M/41.5M [00:00<00:00, 46.9MB/s]
58%|#####7 | 24.0M/41.5M [00:00<00:00, 49.6MB/s]
77%|#######7 | 32.0M/41.5M [00:00<00:00, 55.8MB/s]
90%|######### | 37.5M/41.5M [00:00<00:00, 55.9MB/s]
100%|##########| 41.5M/41.5M [00:01<00:00, 43.1MB/s]
+
0%| | 0.00/41.5M [00:00<?, ?B/s]
15%|#5 | 6.33M/41.5M [00:00<00:00, 47.2MB/s]
26%|##6 | 10.8M/41.5M [00:00<00:00, 38.8MB/s]
39%|###8 | 16.0M/41.5M [00:00<00:00, 39.6MB/s]
58%|#####7 | 24.0M/41.5M [00:00<00:00, 38.3MB/s]
77%|#######7 | 32.0M/41.5M [00:00<00:00, 43.5MB/s]
87%|########7 | 36.2M/41.5M [00:00<00:00, 43.4MB/s]
100%|##########| 41.5M/41.5M [00:00<00:00, 44.2MB/s]
diff --git a/docs/_sources/how_to/compile_models/from_pytorch.rst.txt b/docs/_sources/how_to/compile_models/from_pytorch.rst.txt
index 12d23b3019..af9ad164a9 100644
--- a/docs/_sources/how_to/compile_models/from_pytorch.rst.txt
+++ b/docs/_sources/how_to/compile_models/from_pytorch.rst.txt
@@ -98,7 +98,7 @@ Load a pretrained PyTorch model
/venv/apache-tvm-py3.7/lib/python3.7/site-packages/torchvision/models/_utils.py:223: UserWarning: Arguments other than a weight enum or `None` for 'weights' are deprecated since 0.13 and will be removed in 0.15. The current behavior is equivalent to passing `weights=ResNet18_Weights.IMAGENET1K_V1`. You can also use `weights=ResNet18_Weights.DEFAULT` to get the most up-to-date weights.
warnings.warn(msg)
Downloading: "https://download.pytorch.org/models/resnet18-f37072fd.pth" to /workspace/.cache/torch/hub/checkpoints/resnet18-f37072fd.pth
-
0%| | 0.00/44.7M [00:00<?, ?B/s]
18%|#7 | 7.99M/44.7M [00:00<00:00, 63.7MB/s]
54%|#####3 | 24.0M/44.7M [00:00<00:00, 111MB/s]
78%|#######8 | 34.9M/44.7M [00:00<00:00, 73.9MB/s]
100%|##########| 44.7M/44.7M [00:00<00:00, 89.6MB/s]
+
0%| | 0.00/44.7M [00:00<?, ?B/s]
18%|#7 | 7.99M/44.7M [00:00<00:00, 39.7MB/s]
36%|###5 | 16.0M/44.7M [00:00<00:00, 46.9MB/s]
54%|#####3 | 24.1M/44.7M [00:00<00:00, 57.7MB/s]
68%|######7 | 30.3M/44.7M [00:00<00:00, 47.9MB/s]
79%|#######8 | 35.2M/44.7M [00:00<00:00, 43.7MB/s]
90%|########9 | 40.0M/44.7M [00:00<00:00, 39.0MB/s]
100%|##########| 44.7M/44.7M [00:01<00:00, 46.5MB/s]
diff --git a/docs/_sources/how_to/compile_models/from_tensorflow.rst.txt b/docs/_sources/how_to/compile_models/from_tensorflow.rst.txt
index 022f53d234..a939096eef 100644
--- a/docs/_sources/how_to/compile_models/from_tensorflow.rst.txt
+++ b/docs/_sources/how_to/compile_models/from_tensorflow.rst.txt
@@ -416,7 +416,7 @@ Run the corresponding model on tensorflow
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 1 minutes 10.762 seconds)
+ **Total running time of the script:** ( 1 minutes 14.751 seconds)
.. _sphx_glr_download_how_to_compile_models_from_tensorflow.py:
diff --git a/docs/_sources/how_to/compile_models/sg_execution_times.rst.txt b/docs/_sources/how_to/compile_models/sg_execution_times.rst.txt
index eb205aa3ab..6af56a208a 100644
--- a/docs/_sources/how_to/compile_models/sg_execution_times.rst.txt
+++ b/docs/_sources/how_to/compile_models/sg_execution_times.rst.txt
@@ -5,26 +5,26 @@
Computation times
=================
-**05:49.921** total execution time for **how_to_compile_models** files:
+**06:02.125** total execution time for **how_to_compile_models** files:
+-----------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_compile_models_from_darknet.py` (``from_darknet.py``) | 01:13.973 | 0.0 MB |
+| :ref:`sphx_glr_how_to_compile_models_from_tensorflow.py` (``from_tensorflow.py``) | 01:14.751 | 0.0 MB |
+-----------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_compile_models_from_tensorflow.py` (``from_tensorflow.py``) | 01:10.762 | 0.0 MB |
+| :ref:`sphx_glr_how_to_compile_models_from_darknet.py` (``from_darknet.py``) | 01:13.629 | 0.0 MB |
+-----------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_compile_models_from_paddle.py` (``from_paddle.py``) | 00:47.141 | 0.0 MB |
+| :ref:`sphx_glr_how_to_compile_models_from_paddle.py` (``from_paddle.py``) | 00:46.113 | 0.0 MB |
+-----------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_compile_models_from_oneflow.py` (``from_oneflow.py``) | 00:33.067 | 0.0 MB |
+| :ref:`sphx_glr_how_to_compile_models_from_oneflow.py` (``from_oneflow.py``) | 00:35.536 | 0.0 MB |
+-----------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_compile_models_from_mxnet.py` (``from_mxnet.py``) | 00:29.133 | 0.0 MB |
+| :ref:`sphx_glr_how_to_compile_models_from_mxnet.py` (``from_mxnet.py``) | 00:32.462 | 0.0 MB |
+-----------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_compile_models_from_coreml.py` (``from_coreml.py``) | 00:27.214 | 0.0 MB |
+| :ref:`sphx_glr_how_to_compile_models_from_coreml.py` (``from_coreml.py``) | 00:28.438 | 0.0 MB |
+-----------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_compile_models_from_tflite.py` (``from_tflite.py``) | 00:25.583 | 0.0 MB |
+| :ref:`sphx_glr_how_to_compile_models_from_tflite.py` (``from_tflite.py``) | 00:26.481 | 0.0 MB |
+-----------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_compile_models_from_pytorch.py` (``from_pytorch.py``) | 00:22.732 | 0.0 MB |
+| :ref:`sphx_glr_how_to_compile_models_from_pytorch.py` (``from_pytorch.py``) | 00:24.117 | 0.0 MB |
+-----------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_compile_models_from_keras.py` (``from_keras.py``) | 00:17.900 | 0.0 MB |
+| :ref:`sphx_glr_how_to_compile_models_from_keras.py` (``from_keras.py``) | 00:18.113 | 0.0 MB |
+-----------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_compile_models_from_onnx.py` (``from_onnx.py``) | 00:02.417 | 0.0 MB |
+| :ref:`sphx_glr_how_to_compile_models_from_onnx.py` (``from_onnx.py``) | 00:02.485 | 0.0 MB |
+-----------------------------------------------------------------------------------+-----------+--------+
diff --git a/docs/_sources/how_to/deploy_models/deploy_model_on_android.rst.txt b/docs/_sources/how_to/deploy_models/deploy_model_on_android.rst.txt
index b7aea20a5a..9b6c488e3e 100644
--- a/docs/_sources/how_to/deploy_models/deploy_model_on_android.rst.txt
+++ b/docs/_sources/how_to/deploy_models/deploy_model_on_android.rst.txt
@@ -433,7 +433,7 @@ Execute on TVM
Evaluate inference time cost...
Execution time summary:
mean (ms) median (ms) max (ms) min (ms) std (ms)
- 16.2353 16.2281 16.3931 16.1329 0.0897
+ 16.0401 16.0527 16.2899 15.8806 0.1304
diff --git a/docs/_sources/how_to/deploy_models/deploy_object_detection_pytorch.rst.txt b/docs/_sources/how_to/deploy_models/deploy_object_detection_pytorch.rst.txt
index c29999b1f5..501328ee2c 100644
--- a/docs/_sources/how_to/deploy_models/deploy_object_detection_pytorch.rst.txt
+++ b/docs/_sources/how_to/deploy_models/deploy_object_detection_pytorch.rst.txt
@@ -127,7 +127,7 @@ Load pre-trained maskrcnn from torchvision and do tracing
/venv/apache-tvm-py3.7/lib/python3.7/site-packages/torchvision/models/_utils.py:223: UserWarning: Arguments other than a weight enum or `None` for 'weights' are deprecated since 0.13 and will be removed in 0.15. The current behavior is equivalent to passing `weights=MaskRCNN_ResNet50_FPN_Weights.COCO_V1`. You can also use `weights=MaskRCNN_ResNet50_FPN_Weights.DEFAULT` to get the most up-to-date weights.
warnings.warn(msg)
Downloading: "https://download.pytorch.org/models/maskrcnn_resnet50_fpn_coco-bf2d0c1e.pth" to /workspace/.cache/torch/hub/checkpoints/maskrcnn_resnet50_fpn_coco-bf2d0c1e.pth
-
0%| | 0.00/170M [00:00<?, ?B/s]
1%|1 | 2.05M/170M [00:00<00:17, 9.96MB/s]
5%|4 | 7.99M/170M [00:00<00:05, 28.7MB/s]
9%|9 | 16.0M/170M [00:00<00:03, 41.3MB/s]
14%|#4 | 24.1M/170M [00:00<00:02, 51.9MB/s]
19%|#8 | 32.0M/170M [00:00<00:02, 53.5MB/s]
24%|##3 | 40.0M/170M [00:00<00:02, 54.7MB/s]
28%|##8 | 48.1M/170M [00:00<00:02, 62.0MB/s]
34%|###4 | 58.0M/170M [00:01<00:01, 61.2MB/s]
38%|###7 | 64.1M/170M [00:01<00:02, 51.7MB/s]
43%|####3 | 73.7M/170M [00:01<00:01, 62.6MB/s]
49%|####8 | 82.9M/170M [00:01<00:01, 70.9MB/s]
53%|#####3 | 90.3M/170M [00:01<00:01, 66.1MB/s]
59%|#####8 | 99.4M/170M [00:01<00:01, 73.4MB/s]
63%|######3 | 107M/170M [00:01<00:00, 76.2MB/s]
68%|######7 | 115M/170M [00:01<00:00, 76.7MB/s]
73%|#######2 | 123M/170M [00:02<00:00, 79.6MB/s]
77%|#######7 | 131M/170M [00:02<00:00, 63.7MB/s]
82%|########2 | 140M/170M [00:02<00:00, 70.4MB/s]
88%|########8 | 150M/170M [00:02<00:00, 78.7MB/s]
94%|#########4| 160M/170M [00:02<00:00, 81.3MB/s]
99%|#########8| 168M/170M [00:02<00:00, 78.0MB/s]
100%|##########| 170M/170M [00:02<00:00, 65.1MB/s]
+
0%| | 0.00/170M [00:00<?, ?B/s]
5%|4 | 7.99M/170M [00:00<00:02, 56.7MB/s]
9%|9 | 16.0M/170M [00:00<00:02, 63.9MB/s]
14%|#4 | 24.0M/170M [00:00<00:02, 53.8MB/s]
19%|#8 | 32.0M/170M [00:00<00:02, 50.6MB/s]
24%|##3 | 40.0M/170M [00:00<00:02, 55.8MB/s]
27%|##7 | 46.3M/170M [00:00<00:02, 53.3MB/s]
32%|###2 | 54.5M/170M [00:00<00:01, 61.5MB/s]
36%|###5 | 60.7M/170M [00:01<00:01, 62.3MB/s]
39%|###9 | 66.9M/170M [00:01<00:01, 62.1MB/s]
43%|####2 | 73.0M/170M [00:01<00:01, 59.0MB/s]
46%|####6 | 78.7M/170M [00:01<00:01, 59.2MB/s]
51%|##### | 86.3M/170M [00:01<00:01, 51.7MB/s]
54%|#####3 | 91.5M/170M [00:01<00:01, 45.6MB/s]
57%|#####6 | 96.1M/170M [00:01<00:01, 41.8MB/s]
60%|###### | 102M/170M [00:02<00:01, 46.2MB/s]
64%|######4 | 109M/170M [00:02<00:01, 52.0MB/s]
67%|######7 | 114M/170M [00:02<00:01, 48.1MB/s]
71%|####### | 120M/170M [00:02<00:01, 49.8MB/s]
75%|#######5 | 128M/170M [00:02<00:00, 56.6MB/s]
79%|#######9 | 134M/170M [00:02<00:00, 58.0MB/s]
82%|########2 | 140M/170M [00:02<00:00, 53.5MB/s]
86%|########5 | 146M/170M [00:02<00:00, 55.1MB/s]
89%|########9 | 152M/170M [00:03<00:00, 46.7MB/s]
94%|#########4| 160M/170M [00:03<00:00, 49.9MB/s]
100%|##########| 170M/170M [00:03<00:00, 54.5MB/s]
/venv/apache-tvm-py3.7/lib/python3.7/site-packages/torch/nn/functional.py:3897: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).
for i in range(dim)
/venv/apache-tvm-py3.7/lib/python3.7/site-packages/torchvision/models/detection/anchor_utils.py:124: UserWarning: __floordiv__ is deprecated, and its behavior will change in a future version of pytorch. It currently rounds toward 0 (like the 'trunc' function NOT 'floor'). This results in incorrect rounding for negative values. To keep the current behavior, use torch.div(a, b, rounding_mode='trunc'), or for actual floor division, use torch.div(a, b, rounding_mode='floor').
@@ -296,7 +296,7 @@ Get boxes with score larger than 0.9
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 3 minutes 16.638 seconds)
+ **Total running time of the script:** ( 3 minutes 21.658 seconds)
.. _sphx_glr_download_how_to_deploy_models_deploy_object_detection_pytorch.py:
diff --git a/docs/_sources/how_to/deploy_models/deploy_prequantized.rst.txt b/docs/_sources/how_to/deploy_models/deploy_prequantized.rst.txt
index 2b6c3f9aa5..7e72e1e3dd 100644
--- a/docs/_sources/how_to/deploy_models/deploy_prequantized.rst.txt
+++ b/docs/_sources/how_to/deploy_models/deploy_prequantized.rst.txt
@@ -236,7 +236,7 @@ training. Other models require a full post training calibration.
/venv/apache-tvm-py3.7/lib/python3.7/site-packages/torchvision/models/_utils.py:223: UserWarning: Arguments other than a weight enum or `None` for 'weights' are deprecated since 0.13 and will be removed in 0.15. The current behavior is equivalent to passing `weights=MobileNet_V2_Weights.IMAGENET1K_V1`. You can also use `weights=MobileNet_V2_Weights.DEFAULT` to get the most up-to-date weights.
warnings.warn(msg)
Downloading: "https://download.pytorch.org/models/mobilenet_v2-b0353104.pth" to /workspace/.cache/torch/hub/checkpoints/mobilenet_v2-b0353104.pth
-
0%| | 0.00/13.6M [00:00<?, ?B/s]
59%|#####8 | 7.99M/13.6M [00:00<00:00, 52.1MB/s]
100%|##########| 13.6M/13.6M [00:00<00:00, 76.1MB/s]
+
0%| | 0.00/13.6M [00:00<?, ?B/s]
59%|#####8 | 7.99M/13.6M [00:00<00:00, 59.0MB/s]
100%|##########| 13.6M/13.6M [00:00<00:00, 61.4MB/s]
@@ -418,7 +418,7 @@ Here we give an example of how to measure performance of TVM compiled models.
Execution time summary:
mean (ms) median (ms) max (ms) min (ms) std (ms)
- 90.3821 90.2990 92.0573 90.0415 0.2972
+ 90.5403 90.5017 94.8389 90.1184 0.4960
@@ -467,7 +467,7 @@ TODO
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 1 minutes 6.391 seconds)
+ **Total running time of the script:** ( 1 minutes 10.807 seconds)
.. _sphx_glr_download_how_to_deploy_models_deploy_prequantized.py:
diff --git a/docs/_sources/how_to/deploy_models/deploy_prequantized_tflite.rst.txt b/docs/_sources/how_to/deploy_models/deploy_prequantized_tflite.rst.txt
index edc0382322..a2c09ec2d4 100644
--- a/docs/_sources/how_to/deploy_models/deploy_prequantized_tflite.rst.txt
+++ b/docs/_sources/how_to/deploy_models/deploy_prequantized_tflite.rst.txt
@@ -432,7 +432,7 @@ Here we give an example of how to measure performance of TVM compiled models.
Execution time summary:
mean (ms) median (ms) max (ms) min (ms) std (ms)
- 120.0254 119.9875 121.7564 119.3263 0.3560
+ 119.4717 119.3346 124.6189 118.5417 0.7517
@@ -469,7 +469,7 @@ Here we give an example of how to measure performance of TVM compiled models.
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 2 minutes 29.810 seconds)
+ **Total running time of the script:** ( 2 minutes 30.840 seconds)
.. _sphx_glr_download_how_to_deploy_models_deploy_prequantized_tflite.py:
diff --git a/docs/_sources/how_to/deploy_models/deploy_quantized.rst.txt b/docs/_sources/how_to/deploy_models/deploy_quantized.rst.txt
index bf75bdb611..747dcfd459 100644
--- a/docs/_sources/how_to/deploy_models/deploy_quantized.rst.txt
+++ b/docs/_sources/how_to/deploy_models/deploy_quantized.rst.txt
@@ -253,7 +253,7 @@ We create a Relay VM to build and execute the model.
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 1 minutes 39.550 seconds)
+ **Total running time of the script:** ( 1 minutes 42.918 seconds)
.. _sphx_glr_download_how_to_deploy_models_deploy_quantized.py:
diff --git a/docs/_sources/how_to/deploy_models/deploy_ssd_gluoncv.rst.txt b/docs/_sources/how_to/deploy_models/deploy_ssd_gluoncv.rst.txt
index b817fab7a0..df63205b13 100644
--- a/docs/_sources/how_to/deploy_models/deploy_ssd_gluoncv.rst.txt
+++ b/docs/_sources/how_to/deploy_models/deploy_ssd_gluoncv.rst.txt
@@ -166,7 +166,7 @@ Convert and compile model for CPU.
data: None
input_sym_arg_type = in_param.infer_type()[0]
Downloading /workspace/.mxnet/models/ssd_512_resnet50_v1_voc-9c8b225a.zip from https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/models/ssd_512_resnet50_v1_voc-9c8b225a.zip...
-
0%| | 0/132723 [00:00<?, ?KB/s]
6%|5 | 7560/132723 [00:00<00:01, 75589.44KB/s]
11%|#1 | 15152/132723 [00:00<00:01, 66359.25KB/s]
16%|#6 | 21862/132723 [00:00<00:02, 54127.63KB/s]
23%|##3 | 30616/132723 [00:00<00:01, 65139.02KB/s]
28%|##8 | 37478/132723 [00:00<00:01, 51319.82KB/s]
35%|###4 | 46181/132723 [00:00<00:01, 60659.27KB/s]
41%|####1 | 54785/132723 [00:00<00:01, 67561.18KB/s]
48%|####7 | 63313/132723 [00:00<00:00, 72517.58KB/s]
54%|#####4 | 72176/132723 [00:01<00:00, 77117.75KB/s]
61%|######1 | 81030/132723 [00:01<00:00, 80424.93KB/s]
68%|######7 | 89889/132723 [00:01<00:00, 82814.05KB/s]
74%|#######4 | 98369/132723 [00:01<00:00, 69539.24KB/s]
81%|######## | 106916/132723 [00:01<00:00, 73678.02KB/s]
87%|########7 | 115725/132723 [00:01<00:00, 77577.02KB/s]
94%|#########3| 124494/132723 [00:01<00:00, 80396.11KB/s]
100%|########
##| 132723/132723 [00:01<00:00, 72410.91KB/s]
+
0%| | 0/132723 [00:00<?, ?KB/s]
5%|5 | 6912/132723 [00:00<00:01, 69106.06KB/s]
11%|#1 | 15152/132723 [00:00<00:01, 68122.83KB/s]
17%|#6 | 21972/132723 [00:00<00:02, 39543.87KB/s]
23%|##2 | 30354/132723 [00:00<00:02, 51109.35KB/s]
28%|##8 | 37497/132723 [00:00<00:01, 56165.70KB/s]
34%|###3 | 44486/132723 [00:00<00:01, 59980.29KB/s]
39%|###8 | 51150/132723 [00:00<00:01, 54000.24KB/s]
45%|####4 | 59588/132723 [00:01<00:01, 61946.16KB/s]
50%|####9 | 66313/132723 [00:01<00:01, 61297.42KB/s]
56%|#####6 | 74717/132723 [00:01<00:00, 67508.51KB/s]
62%|######1 | 81798/132723 [00:01<00:00, 62225.58KB/s]
67%|######6 | 88311/132723 [00:01<00:01, 43222.78KB/s]
73%|#######2 | 96688/132723 [00:01<00:00, 51636.36KB/s]
78%|#######7 | 102894/132723 [00:01<00:00, 53736.48KB/s]
84%|########3 | 111335/132723 [00:01<00:00, 61258.35KB/s]
89%|########9
| 118185/132723 [00:02<00:00, 50980.55KB/s]
95%|#########5| 126549/132723 [00:02<00:00, 58502.28KB/s]
100%|##########| 132723/132723 [00:02<00:00, 56851.09KB/s]
@@ -242,7 +242,7 @@ Display result
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 3 minutes 0.610 seconds)
+ **Total running time of the script:** ( 3 minutes 3.754 seconds)
.. _sphx_glr_download_how_to_deploy_models_deploy_ssd_gluoncv.py:
diff --git a/docs/_sources/how_to/deploy_models/sg_execution_times.rst.txt b/docs/_sources/how_to/deploy_models/sg_execution_times.rst.txt
index f141e43bcf..5817719fc2 100644
--- a/docs/_sources/how_to/deploy_models/sg_execution_times.rst.txt
+++ b/docs/_sources/how_to/deploy_models/sg_execution_times.rst.txt
@@ -5,24 +5,24 @@
Computation times
=================
-**13:00.700** total execution time for **how_to_deploy_models** files:
+**13:22.160** total execution time for **how_to_deploy_models** files:
+------------------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_deploy_models_deploy_object_detection_pytorch.py` (``deploy_object_detection_pytorch.py``) | 03:16.638 | 0.0 MB |
+| :ref:`sphx_glr_how_to_deploy_models_deploy_object_detection_pytorch.py` (``deploy_object_detection_pytorch.py``) | 03:21.658 | 0.0 MB |
+------------------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_deploy_models_deploy_ssd_gluoncv.py` (``deploy_ssd_gluoncv.py``) | 03:00.610 | 0.0 MB |
+| :ref:`sphx_glr_how_to_deploy_models_deploy_ssd_gluoncv.py` (``deploy_ssd_gluoncv.py``) | 03:03.754 | 0.0 MB |
+------------------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_deploy_models_deploy_prequantized_tflite.py` (``deploy_prequantized_tflite.py``) | 02:29.810 | 0.0 MB |
+| :ref:`sphx_glr_how_to_deploy_models_deploy_prequantized_tflite.py` (``deploy_prequantized_tflite.py``) | 02:30.840 | 0.0 MB |
+------------------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_deploy_models_deploy_quantized.py` (``deploy_quantized.py``) | 01:39.550 | 0.0 MB |
+| :ref:`sphx_glr_how_to_deploy_models_deploy_quantized.py` (``deploy_quantized.py``) | 01:42.918 | 0.0 MB |
+------------------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_deploy_models_deploy_prequantized.py` (``deploy_prequantized.py``) | 01:06.391 | 0.0 MB |
+| :ref:`sphx_glr_how_to_deploy_models_deploy_prequantized.py` (``deploy_prequantized.py``) | 01:10.807 | 0.0 MB |
+------------------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_deploy_models_deploy_model_on_android.py` (``deploy_model_on_android.py``) | 00:36.869 | 0.0 MB |
+| :ref:`sphx_glr_how_to_deploy_models_deploy_model_on_android.py` (``deploy_model_on_android.py``) | 00:38.804 | 0.0 MB |
+------------------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_deploy_models_deploy_model_on_nano.py` (``deploy_model_on_nano.py``) | 00:25.650 | 0.0 MB |
+| :ref:`sphx_glr_how_to_deploy_models_deploy_model_on_nano.py` (``deploy_model_on_nano.py``) | 00:26.989 | 0.0 MB |
+------------------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_deploy_models_deploy_model_on_rasp.py` (``deploy_model_on_rasp.py``) | 00:25.176 | 0.0 MB |
+| :ref:`sphx_glr_how_to_deploy_models_deploy_model_on_rasp.py` (``deploy_model_on_rasp.py``) | 00:26.383 | 0.0 MB |
+------------------------------------------------------------------------------------------------------------------+-----------+--------+
| :ref:`sphx_glr_how_to_deploy_models_deploy_sparse.py` (``deploy_sparse.py``) | 00:00.007 | 0.0 MB |
+------------------------------------------------------------------------------------------------------------------+-----------+--------+
diff --git a/docs/_sources/how_to/extend_tvm/bring_your_own_datatypes.rst.txt b/docs/_sources/how_to/extend_tvm/bring_your_own_datatypes.rst.txt
index 183149e0cf..16ba889de5 100644
--- a/docs/_sources/how_to/extend_tvm/bring_your_own_datatypes.rst.txt
+++ b/docs/_sources/how_to/extend_tvm/bring_your_own_datatypes.rst.txt
@@ -472,7 +472,7 @@ First let us define two helper functions to get the mobilenet model and a cat im
.. code-block:: none
- Downloading /workspace/.mxnet/models/mobilenet0.25-9f83e440.zip4f7f955f-332b-4872-85b7-b598746fbf97 from https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/models/mobilenet0.25-9f83e440.zip...
+ Downloading /workspace/.mxnet/models/mobilenet0.25-9f83e440.zipccbf2386-7e57-408a-8b03-16a9002a6478 from https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/models/mobilenet0.25-9f83e440.zip...
diff --git a/docs/_sources/how_to/extend_tvm/low_level_custom_pass.rst.txt b/docs/_sources/how_to/extend_tvm/low_level_custom_pass.rst.txt
index cb7ecc5c9b..28257ca020 100644
--- a/docs/_sources/how_to/extend_tvm/low_level_custom_pass.rst.txt
+++ b/docs/_sources/how_to/extend_tvm/low_level_custom_pass.rst.txt
@@ -91,10 +91,9 @@ our customized lowering pass to manipulate the IR directly instead of using sche
buffers = {a: Buffer(a_2: Pointer(float32), float32, [128], []),
b: Buffer(b_2: Pointer(float32), float32, [128], []),
c: Buffer(c_2: Pointer(float32), float32, [128], [])}
- buffer_map = {a_1: a, b_1: b, c_1: c}
- preflattened_buffer_map = {a_1: a_3: Buffer(a_2, float32, [128], []), b_1: b_3: Buffer(b_2, float32, [128], []), c_1: c_3: Buffer(c_2, float32, [128], [])} {
+ buffer_map = {a_1: a, b_1: b, c_1: c} {
for (i: int32, 0, 128) {
- c[i] = (a[i] + b[i])
+ c_3: Buffer(c_2, float32, [128], [])[i] = (a_3: Buffer(a_2, float32, [128], [])[i] + b_3: Buffer(b_2, float32, [128], [])[i])
}
}
@@ -251,11 +250,10 @@ Thus, a good place to put this transformation pass is just after Phase 1.
buffers = {a: Buffer(a_2: Pointer(float32), float32, [128], []),
b: Buffer(b_2: Pointer(float32), float32, [128], []),
c: Buffer(c_2: Pointer(float32), float32, [128], [])}
- buffer_map = {a_1: a, b_1: b, c_1: c}
- preflattened_buffer_map = {a_1: a_3: Buffer(a_2, float32, [128], []), b_1: b_3: Buffer(b_2, float32, [128], []), c_1: c_3: Buffer(c_2, float32, [128], [])} {
+ buffer_map = {a_1: a, b_1: b, c_1: c} {
for (i.outer: int32, 0, 16) {
let cse_var_1: int32 = (i.outer*8)
- c[ramp(cse_var_1, 1, 8)] = (a[ramp(cse_var_1, 1, 8)] + b[ramp(cse_var_1, 1, 8)])
+ c_3: Buffer(c_2, float32, [128], [])[ramp(cse_var_1, 1, 8)] = (a_3: Buffer(a_2, float32, [128], [])[ramp(cse_var_1, 1, 8)] + b_3: Buffer(b_2, float32, [128], [])[ramp(cse_var_1, 1, 8)])
}
}
diff --git a/docs/_sources/how_to/extend_tvm/sg_execution_times.rst.txt b/docs/_sources/how_to/extend_tvm/sg_execution_times.rst.txt
index 05296d43e6..defc7dc9e1 100644
--- a/docs/_sources/how_to/extend_tvm/sg_execution_times.rst.txt
+++ b/docs/_sources/how_to/extend_tvm/sg_execution_times.rst.txt
@@ -5,14 +5,14 @@
Computation times
=================
-**00:47.459** total execution time for **how_to_extend_tvm** files:
+**00:48.149** total execution time for **how_to_extend_tvm** files:
+-------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_extend_tvm_bring_your_own_datatypes.py` (``bring_your_own_datatypes.py``) | 00:44.024 | 0.0 MB |
+| :ref:`sphx_glr_how_to_extend_tvm_bring_your_own_datatypes.py` (``bring_your_own_datatypes.py``) | 00:44.665 | 0.0 MB |
+-------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_extend_tvm_use_pass_instrument.py` (``use_pass_instrument.py``) | 00:02.394 | 0.0 MB |
+| :ref:`sphx_glr_how_to_extend_tvm_use_pass_instrument.py` (``use_pass_instrument.py``) | 00:02.437 | 0.0 MB |
+-------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_extend_tvm_use_pass_infra.py` (``use_pass_infra.py``) | 00:01.033 | 0.0 MB |
+| :ref:`sphx_glr_how_to_extend_tvm_use_pass_infra.py` (``use_pass_infra.py``) | 00:01.039 | 0.0 MB |
+-------------------------------------------------------------------------------------------------+-----------+--------+
| :ref:`sphx_glr_how_to_extend_tvm_low_level_custom_pass.py` (``low_level_custom_pass.py``) | 00:00.008 | 0.0 MB |
+-------------------------------------------------------------------------------------------------+-----------+--------+
diff --git a/docs/_sources/how_to/extend_tvm/use_pass_instrument.rst.txt b/docs/_sources/how_to/extend_tvm/use_pass_instrument.rst.txt
index 20cf04786d..ee72902846 100644
--- a/docs/_sources/how_to/extend_tvm/use_pass_instrument.rst.txt
+++ b/docs/_sources/how_to/extend_tvm/use_pass_instrument.rst.txt
@@ -216,10 +216,10 @@ profile the execution time of each passes.
.. code-block:: none
Printing results of timing profile...
- InferType: 7142us [7142us] (44.30%; 44.30%)
- FoldScaleAxis: 8981us [7us] (55.70%; 55.70%)
- FoldConstant: 8975us [1662us] (55.66%; 99.93%)
- InferType: 7313us [7313us] (45.35%; 81.48%)
+ InferType: 7176us [7176us] (46.30%; 46.30%)
+ FoldScaleAxis: 8322us [6us] (53.70%; 53.70%)
+ FoldConstant: 8315us [1690us] (53.66%; 99.92%)
+ InferType: 6626us [6626us] (42.75%; 79.68%)
@@ -258,10 +258,10 @@ Refer to following sections and :py:func:`tvm.instrument.pass_instrument` for th
.. code-block:: none
Printing results of timing profile...
- InferType: 6707us [6707us] (45.33%; 45.33%)
- FoldScaleAxis: 8087us [5us] (54.67%; 54.67%)
- FoldConstant: 8083us [1680us] (54.63%; 99.94%)
- InferType: 6403us [6403us] (43.28%; 79.22%)
+ InferType: 6651us [6651us] (45.01%; 45.01%)
+ FoldScaleAxis: 8127us [5us] (54.99%; 54.99%)
+ FoldConstant: 8122us [1678us] (54.96%; 99.94%)
+ InferType: 6444us [6444us] (43.61%; 79.34%)
diff --git a/docs/_sources/how_to/optimize_operators/opt_conv_cuda.rst.txt b/docs/_sources/how_to/optimize_operators/opt_conv_cuda.rst.txt
index e4ddb3deb5..7a42515b07 100644
--- a/docs/_sources/how_to/optimize_operators/opt_conv_cuda.rst.txt
+++ b/docs/_sources/how_to/optimize_operators/opt_conv_cuda.rst.txt
@@ -340,7 +340,7 @@ latency of convolution.
.. code-block:: none
- Convolution: 47.685344 ms
+ Convolution: 41.190143 ms
diff --git a/docs/_sources/how_to/optimize_operators/opt_conv_tensorcore.rst.txt b/docs/_sources/how_to/optimize_operators/opt_conv_tensorcore.rst.txt
index aa50927ebd..fea31eb590 100644
--- a/docs/_sources/how_to/optimize_operators/opt_conv_tensorcore.rst.txt
+++ b/docs/_sources/how_to/optimize_operators/opt_conv_tensorcore.rst.txt
@@ -431,11 +431,10 @@ one time.
@main = primfn(A_1: handle, W_1: handle, Conv_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float16), float16, [12845056], []),
- W: Buffer(W_2: Pointer(float16), float16, [1179648], []),
- Conv: Buffer(Conv_2: Pointer(float32), float32, [25690112], [])}
- buffer_map = {A_1: A, W_1: W, Conv_1: Conv}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float16, [16, 14, 14, 16, 16, 16], []), W_1: W_3: Buffer(W_2, float16, [3, 3, 16, 32, 16, 16], []), Conv_1: Conv_3: Buffer(Conv_2, float32, [16, 14, 14, 32, 16, 16], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float16), float16, [16, 14, 14, 16, 16, 16], []),
+ W: Buffer(W_2: Pointer(float16), float16, [3, 3, 16, 32, 16, 16], []),
+ Conv: Buffer(Conv_2: Pointer(float32), float32, [16, 14, 14, 32, 16, 16], [])}
+ buffer_map = {A_1: A, W_1: W, Conv_1: Conv} {
attr [IterVar(blockIdx.z: int32, (nullptr), "ThreadIndex", "blockIdx.z")] "thread_extent" = 196;
allocate(Conv.wmma.accumulator: Pointer(wmma.accumulator float32), float32, [2048]), storage_scope = wmma.accumulator;
allocate(Apad.shared: Pointer(shared float16), float16, [12288]), storage_scope = shared;
@@ -463,14 +462,14 @@ one time.
let cse_var_2: int32 = (ax3*256)
let cse_var_1: int32 = (ax4.ax5.fused.outer*32)
attr [IterVar(threadIdx.x: int32, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 32;
- Apad.shared_1: Buffer(Apad.shared, float16, [12288], [], scope="shared")[((((((threadIdx.y*3072) + (threadIdx.z*1536)) + (ax2*512)) + cse_var_2) + cse_var_1) + threadIdx.x)] = @tir.if_then_else(((((1 <= (floordiv(blockIdx.z, 14) + kh)) && ((floordiv(blockIdx.z, 14) + kh) < 15)) && (1 <= (ax2 + floormod(blockIdx.z, 14)))) && ((ax2 + floormod(blockIdx.z, 14)) < 15)), A[(((((((((((blockIdx.x*6422528) + (threadIdx.y*1605632)) + (threadIdx.z*802816)) + (kh*57344)) + (blockId [...]
+ Apad.shared_1: Buffer(Apad.shared, float16, [12288], [], scope="shared")[((((((threadIdx.y*3072) + (threadIdx.z*1536)) + (ax2*512)) + cse_var_2) + cse_var_1) + threadIdx.x)] = @tir.if_then_else(((((1 <= (floordiv(blockIdx.z, 14) + kh)) && ((floordiv(blockIdx.z, 14) + kh) < 15)) && (1 <= (ax2 + floormod(blockIdx.z, 14)))) && ((ax2 + floormod(blockIdx.z, 14)) < 15)), A_3: Buffer(A_2, float16, [12845056], [])[(((((((((((blockIdx.x*6422528) + (threadIdx.y*1605632)) + (threa [...]
}
}
}
for (ax1: int32, 0, 3) {
for (ax2_1: int32, 0, 2) {
attr [IterVar(threadIdx.x, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 32;
- W.shared_1: Buffer(W.shared, float16, [12288], [], scope="shared")[ramp((((((ax1*4096) + (ax2_1*2048)) + (threadIdx.y*512)) + (threadIdx.z*256)) + (threadIdx.x*8)), 1, 8)] = W[ramp(((((((((kh*393216) + (ax1*131072)) + (ic.outer*16384)) + (ax2_1*8192)) + (blockIdx.y*2048)) + (threadIdx.y*512)) + (threadIdx.z*256)) + (threadIdx.x*8)), 1, 8)]
+ W.shared_1: Buffer(W.shared, float16, [12288], [], scope="shared")[ramp((((((ax1*4096) + (ax2_1*2048)) + (threadIdx.y*512)) + (threadIdx.z*256)) + (threadIdx.x*8)), 1, 8)] = W_3: Buffer(W_2, float16, [1179648], [])[ramp(((((((((kh*393216) + (ax1*131072)) + (ic.outer*16384)) + (ax2_1*8192)) + (blockIdx.y*2048)) + (threadIdx.y*512)) + (threadIdx.z*256)) + (threadIdx.x*8)), 1, 8)]
}
}
for (ic.inner: int32, 0, 2) {
@@ -516,7 +515,7 @@ one time.
for (oo: int32, 0, 16) {
let cse_var_10: int32 = (o.inner*256)
let cse_var_9: int32 = (nn*16)
- Conv[(((((((((blockIdx.x*12845056) + (threadIdx.y*3211264)) + (n.inner*1605632)) + (blockIdx.z*8192)) + (blockIdx.y*2048)) + (threadIdx.z*1024)) + cse_var_10) + cse_var_9) + oo)] = Conv.wmma.accumulator_1[((((n.inner*1024) + cse_var_10) + cse_var_9) + oo)]
+ Conv_3: Buffer(Conv_2, float32, [25690112], [])[(((((((((blockIdx.x*12845056) + (threadIdx.y*3211264)) + (n.inner*1605632)) + (blockIdx.z*8192)) + (blockIdx.y*2048)) + (threadIdx.z*1024)) + cse_var_10) + cse_var_9) + oo)] = Conv.wmma.accumulator_1[((((n.inner*1024) + cse_var_10) + cse_var_9) + oo)]
}
}
}
@@ -557,11 +556,10 @@ by mapping the 2D convolution to tensor intrinsics
@main = primfn(A_1: handle, W_1: handle, Conv_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float16), float16, [12845056], []),
- W: Buffer(W_2: Pointer(float16), float16, [1179648], []),
- Conv: Buffer(Conv_2: Pointer(float32), float32, [25690112], [])}
- buffer_map = {A_1: A, W_1: W, Conv_1: Conv}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float16, [16, 14, 14, 16, 16, 16], []), W_1: W_3: Buffer(W_2, float16, [3, 3, 16, 32, 16, 16], []), Conv_1: Conv_3: Buffer(Conv_2, float32, [16, 14, 14, 32, 16, 16], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float16), float16, [16, 14, 14, 16, 16, 16], []),
+ W: Buffer(W_2: Pointer(float16), float16, [3, 3, 16, 32, 16, 16], []),
+ Conv: Buffer(Conv_2: Pointer(float32), float32, [16, 14, 14, 32, 16, 16], [])}
+ buffer_map = {A_1: A, W_1: W, Conv_1: Conv} {
attr [IterVar(blockIdx.z: int32, (nullptr), "ThreadIndex", "blockIdx.z")] "thread_extent" = 196;
allocate(Conv.wmma.accumulator: Pointer(wmma.accumulator float32), float32, [2048]), storage_scope = wmma.accumulator;
allocate(Apad.shared: Pointer(shared float16), float16, [12288]), storage_scope = shared;
@@ -585,14 +583,14 @@ by mapping the 2D convolution to tensor intrinsics
let cse_var_2: int32 = (ax3*256)
let cse_var_1: int32 = (ax4.ax5.fused.outer*32)
attr [IterVar(threadIdx.x: int32, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 32;
- Apad.shared_1: Buffer(Apad.shared, float16, [12288], [], scope="shared")[((((((threadIdx.y*3072) + (threadIdx.z*1536)) + (ax2*512)) + cse_var_2) + cse_var_1) + threadIdx.x)] = @tir.if_then_else(((((1 <= (floordiv(blockIdx.z, 14) + kh)) && ((floordiv(blockIdx.z, 14) + kh) < 15)) && (1 <= (ax2 + floormod(blockIdx.z, 14)))) && ((ax2 + floormod(blockIdx.z, 14)) < 15)), A[(((((((((((blockIdx.x*6422528) + (threadIdx.y*1605632)) + (threadIdx.z*802816)) + (kh*57344)) + (blockId [...]
+ Apad.shared_1: Buffer(Apad.shared, float16, [12288], [], scope="shared")[((((((threadIdx.y*3072) + (threadIdx.z*1536)) + (ax2*512)) + cse_var_2) + cse_var_1) + threadIdx.x)] = @tir.if_then_else(((((1 <= (floordiv(blockIdx.z, 14) + kh)) && ((floordiv(blockIdx.z, 14) + kh) < 15)) && (1 <= (ax2 + floormod(blockIdx.z, 14)))) && ((ax2 + floormod(blockIdx.z, 14)) < 15)), A_3: Buffer(A_2, float16, [12845056], [])[(((((((((((blockIdx.x*6422528) + (threadIdx.y*1605632)) + (threa [...]
}
}
}
for (ax1: int32, 0, 3) {
for (ax2_1: int32, 0, 2) {
attr [IterVar(threadIdx.x, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 32;
- W.shared_1: Buffer(W.shared, float16, [12288], [], scope="shared")[ramp((((((ax1*4096) + (ax2_1*2048)) + (threadIdx.y*512)) + (threadIdx.z*256)) + (threadIdx.x*8)), 1, 8)] = W[ramp(((((((((kh*393216) + (ax1*131072)) + (ic.outer*16384)) + (ax2_1*8192)) + (blockIdx.y*2048)) + (threadIdx.y*512)) + (threadIdx.z*256)) + (threadIdx.x*8)), 1, 8)]
+ W.shared_1: Buffer(W.shared, float16, [12288], [], scope="shared")[ramp((((((ax1*4096) + (ax2_1*2048)) + (threadIdx.y*512)) + (threadIdx.z*256)) + (threadIdx.x*8)), 1, 8)] = W_3: Buffer(W_2, float16, [1179648], [])[ramp(((((((((kh*393216) + (ax1*131072)) + (ic.outer*16384)) + (ax2_1*8192)) + (blockIdx.y*2048)) + (threadIdx.y*512)) + (threadIdx.z*256)) + (threadIdx.x*8)), 1, 8)]
}
}
for (ic.inner: int32, 0, 2) {
@@ -659,7 +657,7 @@ be able to run on our build server
.. code-block:: none
- conv2d with tensor core: 12.128944 ms
+ conv2d with tensor core: 13.362189 ms
diff --git a/docs/_sources/how_to/optimize_operators/opt_gemm.rst.txt b/docs/_sources/how_to/optimize_operators/opt_gemm.rst.txt
index ba7f9580fa..906070cc83 100644
--- a/docs/_sources/how_to/optimize_operators/opt_gemm.rst.txt
+++ b/docs/_sources/how_to/optimize_operators/opt_gemm.rst.txt
@@ -143,8 +143,8 @@ Then we write a baseline implementation, the simplest way to write a matrix mult
.. code-block:: none
- Numpy running time: 0.019367
- Baseline: 3.423094
+ Numpy running time: 0.018437
+ Baseline: 3.542359
@@ -171,18 +171,17 @@ Here is the generated IR using our baseline schedule.
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [1048576], []),
- B: Buffer(B_2: Pointer(float32), float32, [1048576], []),
- C: Buffer(C_2: Pointer(float32), float32, [1048576], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 1024], []), B_1: B_3: Buffer(B_2, float32, [1024, 1024], []), C_1: C_3: Buffer(C_2, float32, [1024, 1024], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),
+ B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
for (m: int32, 0, 1024) {
for (n: int32, 0, 1024) {
- C[((m*1024) + n)] = 0f32
+ C_3: Buffer(C_2, float32, [1048576], [])[((m*1024) + n)] = 0f32
for (k: int32, 0, 1024) {
let cse_var_2: int32 = (m*1024)
let cse_var_1: int32 = (cse_var_2 + n)
- C[cse_var_1] = (C[cse_var_1] + (A[(cse_var_2 + k)]*B[((k*1024) + n)]))
+ C_3[cse_var_1] = (C_3[cse_var_1] + (A_3: Buffer(A_2, float32, [1048576], [])[(cse_var_2 + k)]*B_3: Buffer(B_2, float32, [1048576], [])[((k*1024) + n)]))
}
}
}
@@ -239,7 +238,7 @@ fill 32 * 32 * sizeof(float) which is 4KB in the cache whose total size is 32KB
.. code-block:: none
- Opt1: 0.302157
+ Opt1: 0.305990
@@ -265,16 +264,15 @@ Here is the generated IR after blocking.
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [1048576], []),
- B: Buffer(B_2: Pointer(float32), float32, [1048576], []),
- C: Buffer(C_2: Pointer(float32), float32, [1048576], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 1024], []), B_1: B_3: Buffer(B_2, float32, [1024, 1024], []), C_1: C_3: Buffer(C_2, float32, [1024, 1024], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),
+ B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
for (m.outer: int32, 0, 32) {
for (n.outer: int32, 0, 32) {
for (m.inner.init: int32, 0, 32) {
for (n.inner.init: int32, 0, 32) {
- C[((((m.outer*32768) + (m.inner.init*1024)) + (n.outer*32)) + n.inner.init)] = 0f32
+ C_3: Buffer(C_2, float32, [1048576], [])[((((m.outer*32768) + (m.inner.init*1024)) + (n.outer*32)) + n.inner.init)] = 0f32
}
}
for (k.outer: int32, 0, 256) {
@@ -284,7 +282,7 @@ Here is the generated IR after blocking.
let cse_var_3: int32 = (n.outer*32)
let cse_var_2: int32 = ((m.outer*32768) + (m.inner*1024))
let cse_var_1: int32 = ((cse_var_2 + cse_var_3) + n.inner)
- C[cse_var_1] = (C[cse_var_1] + (A[((cse_var_2 + (k.outer*4)) + k.inner)]*B[((((k.outer*4096) + (k.inner*1024)) + cse_var_3) + n.inner)]))
+ C_3[cse_var_1] = (C_3[cse_var_1] + (A_3: Buffer(A_2, float32, [1048576], [])[((cse_var_2 + (k.outer*4)) + k.inner)]*B_3: Buffer(B_2, float32, [1048576], [])[((((k.outer*4096) + (k.inner*1024)) + cse_var_3) + n.inner)]))
}
}
}
@@ -342,7 +340,7 @@ In this tutorial, we chose to vectorize the inner loop row data since it is cach
.. code-block:: none
- Opt2: 0.340464
+ Opt2: 0.335689
@@ -368,15 +366,14 @@ Here is the generated IR after vectorization.
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [1048576], []),
- B: Buffer(B_2: Pointer(float32), float32, [1048576], []),
- C: Buffer(C_2: Pointer(float32), float32, [1048576], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 1024], []), B_1: B_3: Buffer(B_2, float32, [1024, 1024], []), C_1: C_3: Buffer(C_2, float32, [1024, 1024], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),
+ B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
for (m.outer: int32, 0, 32) {
for (n.outer: int32, 0, 32) {
for (m.inner.init: int32, 0, 32) {
- C[ramp((((m.outer*32768) + (m.inner.init*1024)) + (n.outer*32)), 1, 32)] = broadcast(0f32, 32)
+ C_3: Buffer(C_2, float32, [1048576], [])[ramp((((m.outer*32768) + (m.inner.init*1024)) + (n.outer*32)), 1, 32)] = broadcast(0f32, 32)
}
for (k.outer: int32, 0, 256) {
for (k.inner: int32, 0, 4) {
@@ -384,7 +381,7 @@ Here is the generated IR after vectorization.
let cse_var_3: int32 = (n.outer*32)
let cse_var_2: int32 = ((m.outer*32768) + (m.inner*1024))
let cse_var_1: int32 = (cse_var_2 + cse_var_3)
- C[ramp(cse_var_1, 1, 32)] = (C[ramp(cse_var_1, 1, 32)] + (broadcast(A[((cse_var_2 + (k.outer*4)) + k.inner)], 32)*B[ramp((((k.outer*4096) + (k.inner*1024)) + cse_var_3), 1, 32)]))
+ C_3[ramp(cse_var_1, 1, 32)] = (C_3[ramp(cse_var_1, 1, 32)] + (broadcast(A_3: Buffer(A_2, float32, [1048576], [])[((cse_var_2 + (k.outer*4)) + k.inner)], 32)*B_3: Buffer(B_2, float32, [1048576], [])[ramp((((k.outer*4096) + (k.inner*1024)) + cse_var_3), 1, 32)]))
}
}
}
@@ -438,7 +435,7 @@ the access pattern for A matrix is more cache friendly.
.. code-block:: none
- Opt3: 0.116802
+ Opt3: 0.116123
@@ -464,15 +461,14 @@ Here is the generated IR after loop permutation.
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [1048576], []),
- B: Buffer(B_2: Pointer(float32), float32, [1048576], []),
- C: Buffer(C_2: Pointer(float32), float32, [1048576], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 1024], []), B_1: B_3: Buffer(B_2, float32, [1024, 1024], []), C_1: C_3: Buffer(C_2, float32, [1024, 1024], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),
+ B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
for (m.outer: int32, 0, 32) {
for (n.outer: int32, 0, 32) {
for (m.inner.init: int32, 0, 32) {
- C[ramp((((m.outer*32768) + (m.inner.init*1024)) + (n.outer*32)), 1, 32)] = broadcast(0f32, 32)
+ C_3: Buffer(C_2, float32, [1048576], [])[ramp((((m.outer*32768) + (m.inner.init*1024)) + (n.outer*32)), 1, 32)] = broadcast(0f32, 32)
}
for (k.outer: int32, 0, 256) {
for (m.inner: int32, 0, 32) {
@@ -480,7 +476,7 @@ Here is the generated IR after loop permutation.
let cse_var_3: int32 = (n.outer*32)
let cse_var_2: int32 = ((m.outer*32768) + (m.inner*1024))
let cse_var_1: int32 = (cse_var_2 + cse_var_3)
- C[ramp(cse_var_1, 1, 32)] = (C[ramp(cse_var_1, 1, 32)] + (broadcast(A[((cse_var_2 + (k.outer*4)) + k.inner)], 32)*B[ramp((((k.outer*4096) + (k.inner*1024)) + cse_var_3), 1, 32)]))
+ C_3[ramp(cse_var_1, 1, 32)] = (C_3[ramp(cse_var_1, 1, 32)] + (broadcast(A_3: Buffer(A_2, float32, [1048576], [])[((cse_var_2 + (k.outer*4)) + k.inner)], 32)*B_3: Buffer(B_2, float32, [1048576], [])[ramp((((k.outer*4096) + (k.inner*1024)) + cse_var_3), 1, 32)]))
}
}
}
@@ -563,7 +559,7 @@ flattening.
.. code-block:: none
- Opt4: 0.109638
+ Opt4: 0.110101
@@ -589,21 +585,20 @@ Here is the generated IR after array packing.
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [1048576], []),
- B: Buffer(B_2: Pointer(float32), float32, [1048576], []),
- C: Buffer(C_2: Pointer(float32), float32, [1048576], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 1024], []), B_1: B_3: Buffer(B_2, float32, [1024, 1024], []), C_1: C_3: Buffer(C_2, float32, [1024, 1024], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),
+ B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
allocate(packedB: Pointer(global float32x32), float32x32, [32768]), storage_scope = global {
for (bigN: int32, 0, 32) "parallel" {
for (k: int32, 0, 1024) {
- packedB_1: Buffer(packedB, float32x32, [32768], [])[((bigN*1024) + k)] = B[ramp(((k*1024) + (bigN*32)), 1, 32)]
+ packedB_1: Buffer(packedB, float32x32, [32768], [])[((bigN*1024) + k)] = B_3: Buffer(B_2, float32, [1048576], [])[ramp(((k*1024) + (bigN*32)), 1, 32)]
}
}
for (m.outer: int32, 0, 32) {
for (n.outer: int32, 0, 32) {
for (m.inner.init: int32, 0, 32) {
- C[ramp((((m.outer*32768) + (m.inner.init*1024)) + (n.outer*32)), 1, 32)] = broadcast(0f32, 32)
+ C_3: Buffer(C_2, float32, [1048576], [])[ramp((((m.outer*32768) + (m.inner.init*1024)) + (n.outer*32)), 1, 32)] = broadcast(0f32, 32)
}
for (k.outer: int32, 0, 256) {
for (m.inner: int32, 0, 32) {
@@ -611,7 +606,7 @@ Here is the generated IR after array packing.
let cse_var_3: int32 = ((m.outer*32768) + (m.inner*1024))
let cse_var_2: int32 = (k.outer*4)
let cse_var_1: int32 = (cse_var_3 + (n.outer*32))
- C[ramp(cse_var_1, 1, 32)] = (C[ramp(cse_var_1, 1, 32)] + (broadcast(A[((cse_var_3 + cse_var_2) + k.inner)], 32)*packedB_1[(((n.outer*1024) + cse_var_2) + k.inner)]))
+ C_3[ramp(cse_var_1, 1, 32)] = (C_3[ramp(cse_var_1, 1, 32)] + (broadcast(A_3: Buffer(A_2, float32, [1048576], [])[((cse_var_3 + cse_var_2) + k.inner)], 32)*packedB_1[(((n.outer*1024) + cse_var_2) + k.inner)]))
}
}
}
@@ -685,7 +680,7 @@ write to C when all the block results are ready.
.. code-block:: none
- Opt5: 0.115397
+ Opt5: 0.111848
@@ -711,16 +706,15 @@ Here is the generated IR after blocking.
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [1048576], []),
- B: Buffer(B_2: Pointer(float32), float32, [1048576], []),
- C: Buffer(C_2: Pointer(float32), float32, [1048576], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 1024], []), B_1: B_3: Buffer(B_2, float32, [1024, 1024], []), C_1: C_3: Buffer(C_2, float32, [1024, 1024], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),
+ B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
allocate(packedB: Pointer(global float32x32), float32x32, [32768]), storage_scope = global;
allocate(C.global: Pointer(global float32), float32, [1024]), storage_scope = global {
for (bigN: int32, 0, 32) "parallel" {
for (k: int32, 0, 1024) {
- packedB_1: Buffer(packedB, float32x32, [32768], [])[((bigN*1024) + k)] = B[ramp(((k*1024) + (bigN*32)), 1, 32)]
+ packedB_1: Buffer(packedB, float32x32, [32768], [])[((bigN*1024) + k)] = B_3: Buffer(B_2, float32, [1048576], [])[ramp(((k*1024) + (bigN*32)), 1, 32)]
}
}
for (m.outer: int32, 0, 32) {
@@ -735,16 +729,16 @@ Here is the generated IR after blocking.
let cse_var_2: int32 = ((n.outer*1024) + cse_var_4)
let cse_var_1: int32 = (((m.outer*32768) + (m.c*1024)) + cse_var_4)
{
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[cse_var_1], 32)*packedB_1[cse_var_2]))
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[(cse_var_1 + 1)], 32)*packedB_1[(cse_var_2 + 1)]))
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[(cse_var_1 + 2)], 32)*packedB_1[(cse_var_2 + 2)]))
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[(cse_var_1 + 3)], 32)*packedB_1[(cse_var_2 + 3)]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3: Buffer(A_2, float32, [1048576], [])[cse_var_1], 32)*packedB_1[cse_var_2]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3[(cse_var_1 + 1)], 32)*packedB_1[(cse_var_2 + 1)]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3[(cse_var_1 + 2)], 32)*packedB_1[(cse_var_2 + 2)]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3[(cse_var_1 + 3)], 32)*packedB_1[(cse_var_2 + 3)]))
}
}
}
for (m.inner: int32, 0, 32) {
for (n.inner: int32, 0, 32) {
- C[((((m.outer*32768) + (m.inner*1024)) + (n.outer*32)) + n.inner)] = C.global_1[((m.inner*32) + n.inner)]
+ C_3: Buffer(C_2, float32, [1048576], [])[((((m.outer*32768) + (m.inner*1024)) + (n.outer*32)) + n.inner)] = C.global_1[((m.inner*32) + n.inner)]
}
}
}
@@ -810,7 +804,7 @@ Furthermore, we can also utilize multi-core processors to do the thread-level pa
.. code-block:: none
- Opt6: 0.152455
+ Opt6: 0.147245
@@ -836,15 +830,14 @@ Here is the generated IR after parallelization.
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [1048576], []),
- B: Buffer(B_2: Pointer(float32), float32, [1048576], []),
- C: Buffer(C_2: Pointer(float32), float32, [1048576], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 1024], []), B_1: B_3: Buffer(B_2, float32, [1024, 1024], []), C_1: C_3: Buffer(C_2, float32, [1024, 1024], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),
+ B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
allocate(packedB: Pointer(global float32x32), float32x32, [32768]), storage_scope = global {
for (bigN: int32, 0, 32) "parallel" {
for (k: int32, 0, 1024) {
- packedB_1: Buffer(packedB, float32x32, [32768], [])[((bigN*1024) + k)] = B[ramp(((k*1024) + (bigN*32)), 1, 32)]
+ packedB_1: Buffer(packedB, float32x32, [32768], [])[((bigN*1024) + k)] = B_3: Buffer(B_2, float32, [1048576], [])[ramp(((k*1024) + (bigN*32)), 1, 32)]
}
}
for (m.outer: int32, 0, 32) "parallel" {
@@ -860,16 +853,16 @@ Here is the generated IR after parallelization.
let cse_var_2: int32 = ((n.outer*1024) + cse_var_4)
let cse_var_1: int32 = (((m.outer*32768) + (m.c*1024)) + cse_var_4)
{
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[cse_var_1], 32)*packedB_1[cse_var_2]))
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[(cse_var_1 + 1)], 32)*packedB_1[(cse_var_2 + 1)]))
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[(cse_var_1 + 2)], 32)*packedB_1[(cse_var_2 + 2)]))
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[(cse_var_1 + 3)], 32)*packedB_1[(cse_var_2 + 3)]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3: Buffer(A_2, float32, [1048576], [])[cse_var_1], 32)*packedB_1[cse_var_2]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3[(cse_var_1 + 1)], 32)*packedB_1[(cse_var_2 + 1)]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3[(cse_var_1 + 2)], 32)*packedB_1[(cse_var_2 + 2)]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3[(cse_var_1 + 3)], 32)*packedB_1[(cse_var_2 + 3)]))
}
}
}
for (m.inner: int32, 0, 32) {
for (n.inner: int32, 0, 32) {
- C[((((m.outer*32768) + (m.inner*1024)) + (n.outer*32)) + n.inner)] = C.global_1[((m.inner*32) + n.inner)]
+ C_3: Buffer(C_2, float32, [1048576], [])[((((m.outer*32768) + (m.inner*1024)) + (n.outer*32)) + n.inner)] = C.global_1[((m.inner*32) + n.inner)]
}
}
}
diff --git a/docs/_sources/how_to/optimize_operators/sg_execution_times.rst.txt b/docs/_sources/how_to/optimize_operators/sg_execution_times.rst.txt
index 05767f3aab..e86fca2eea 100644
--- a/docs/_sources/how_to/optimize_operators/sg_execution_times.rst.txt
+++ b/docs/_sources/how_to/optimize_operators/sg_execution_times.rst.txt
@@ -5,12 +5,12 @@
Computation times
=================
-**00:35.446** total execution time for **how_to_optimize_operators** files:
+**00:35.417** total execution time for **how_to_optimize_operators** files:
+-----------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_optimize_operators_opt_gemm.py` (``opt_gemm.py``) | 00:32.843 | 0.0 MB |
+| :ref:`sphx_glr_how_to_optimize_operators_opt_gemm.py` (``opt_gemm.py``) | 00:32.836 | 0.0 MB |
+-----------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_optimize_operators_opt_conv_tensorcore.py` (``opt_conv_tensorcore.py``) | 00:01.481 | 0.0 MB |
+| :ref:`sphx_glr_how_to_optimize_operators_opt_conv_tensorcore.py` (``opt_conv_tensorcore.py``) | 00:01.521 | 0.0 MB |
+-----------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_optimize_operators_opt_conv_cuda.py` (``opt_conv_cuda.py``) | 00:01.121 | 0.0 MB |
+| :ref:`sphx_glr_how_to_optimize_operators_opt_conv_cuda.py` (``opt_conv_cuda.py``) | 00:01.061 | 0.0 MB |
+-----------------------------------------------------------------------------------------------+-----------+--------+
diff --git a/docs/_sources/how_to/tune_with_autoscheduler/sg_execution_times.rst.txt b/docs/_sources/how_to/tune_with_autoscheduler/sg_execution_times.rst.txt
index e849ba3cbf..59f0199b5e 100644
--- a/docs/_sources/how_to/tune_with_autoscheduler/sg_execution_times.rst.txt
+++ b/docs/_sources/how_to/tune_with_autoscheduler/sg_execution_times.rst.txt
@@ -5,18 +5,18 @@
Computation times
=================
-**09:11.050** total execution time for **how_to_tune_with_autoscheduler** files:
+**09:43.171** total execution time for **how_to_tune_with_autoscheduler** files:
+----------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_tune_with_autoscheduler_tune_conv2d_layer_cuda.py` (``tune_conv2d_layer_cuda.py``) | 05:37.751 | 0.0 MB |
+| :ref:`sphx_glr_how_to_tune_with_autoscheduler_tune_conv2d_layer_cuda.py` (``tune_conv2d_layer_cuda.py``) | 06:07.710 | 0.0 MB |
+----------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_tune_with_autoscheduler_tune_network_x86.py` (``tune_network_x86.py``) | 01:34.377 | 0.0 MB |
+| :ref:`sphx_glr_how_to_tune_with_autoscheduler_tune_network_x86.py` (``tune_network_x86.py``) | 01:34.567 | 0.0 MB |
+----------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_tune_with_autoscheduler_tune_network_cuda.py` (``tune_network_cuda.py``) | 01:04.612 | 0.0 MB |
+| :ref:`sphx_glr_how_to_tune_with_autoscheduler_tune_network_cuda.py` (``tune_network_cuda.py``) | 01:04.692 | 0.0 MB |
+----------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_tune_with_autoscheduler_tune_sparse_x86.py` (``tune_sparse_x86.py``) | 00:30.535 | 0.0 MB |
+| :ref:`sphx_glr_how_to_tune_with_autoscheduler_tune_sparse_x86.py` (``tune_sparse_x86.py``) | 00:31.234 | 0.0 MB |
+----------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_tune_with_autoscheduler_tune_network_arm.py` (``tune_network_arm.py``) | 00:12.286 | 0.0 MB |
+| :ref:`sphx_glr_how_to_tune_with_autoscheduler_tune_network_arm.py` (``tune_network_arm.py``) | 00:12.963 | 0.0 MB |
+----------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_tune_with_autoscheduler_tune_network_mali.py` (``tune_network_mali.py``) | 00:11.490 | 0.0 MB |
+| :ref:`sphx_glr_how_to_tune_with_autoscheduler_tune_network_mali.py` (``tune_network_mali.py``) | 00:12.004 | 0.0 MB |
+----------------------------------------------------------------------------------------------------------+-----------+--------+
diff --git a/docs/_sources/how_to/tune_with_autoscheduler/tune_conv2d_layer_cuda.rst.txt b/docs/_sources/how_to/tune_with_autoscheduler/tune_conv2d_layer_cuda.rst.txt
index d30a903edd..a73cdc578d 100644
--- a/docs/_sources/how_to/tune_with_autoscheduler/tune_conv2d_layer_cuda.rst.txt
+++ b/docs/_sources/how_to/tune_with_autoscheduler/tune_conv2d_layer_cuda.rst.txt
@@ -234,109 +234,488 @@ cooperative fetching, unrolling and operator fusion.
Lowered TIR:
@main = primfn(data_1: handle, kernel_1: handle, bias_1: handle, compute_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {data: Buffer(data_2: Pointer(float32), float32, [25088], []),
- kernel: Buffer(kernel_2: Pointer(float32), float32, [2359296], []),
- bias: Buffer(bias_2: Pointer(float32), float32, [512], []),
- compute: Buffer(compute_2: Pointer(float32), float32, [25088], [])}
- buffer_map = {data_1: data, kernel_1: kernel, bias_1: bias, compute_1: compute}
- preflattened_buffer_map = {data_1: data_3: Buffer(data_2, float32, [1, 512, 7, 7], []), kernel_1: kernel_3: Buffer(kernel_2, float32, [512, 512, 3, 3], []), bias_1: bias_3: Buffer(bias_2, float32, [1, 512, 1, 1], []), compute_1: compute_3: Buffer(compute_2, float32, [1, 512, 7, 7], [])} {
- attr [IterVar(blockIdx.x: int32, (nullptr), "ThreadIndex", "blockIdx.x")] "thread_extent" = 64;
- allocate(conv2d_nchw: Pointer(local float32), float32, [4]), storage_scope = local;
- allocate(pad_temp.shared: Pointer(shared float32), float32, [162]), storage_scope = shared;
- allocate(kernel.shared: Pointer(shared float32), float32, [144]), storage_scope = shared;
- attr [IterVar(threadIdx.x: int32, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 98 {
- conv2d_nchw_1: Buffer(conv2d_nchw, float32, [4], [], scope="local", align=16)[0] = 0f32
+ buffers = {data: Buffer(data_2: Pointer(float32), float32, [1, 512, 7, 7], []),
+ kernel: Buffer(kernel_2: Pointer(float32), float32, [512, 512, 3, 3], []),
+ bias: Buffer(bias_2: Pointer(float32), float32, [1, 512, 1, 1], []),
+ compute: Buffer(compute_2: Pointer(float32), float32, [1, 512, 7, 7], [])}
+ buffer_map = {data_1: data, kernel_1: kernel, bias_1: bias, compute_1: compute} {
+ attr [IterVar(blockIdx.x: int32, (nullptr), "ThreadIndex", "blockIdx.x")] "thread_extent" = 28;
+ allocate(conv2d_nchw: Pointer(local float32), float32, [14]), storage_scope = local;
+ allocate(pad_temp.shared: Pointer(shared float32), float32, [72]), storage_scope = shared;
+ allocate(kernel.shared: Pointer(shared float32), float32, [3072]), storage_scope = shared;
+ attr [IterVar(threadIdx.x: int32, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64 {
+ conv2d_nchw_1: Buffer(conv2d_nchw, float32, [14], [], scope="local", align=32)[0] = 0f32
conv2d_nchw_1[1] = 0f32
conv2d_nchw_1[2] = 0f32
conv2d_nchw_1[3] = 0f32
- for (rc.outer.outer: int32, 0, 256) {
- attr [IterVar(threadIdx.x_1: int32, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 98;
- pad_temp.shared_1: Buffer(pad_temp.shared, float32, [162], [], scope="shared")[threadIdx.x_1] = @tir.if_then_else(((((9 <= floormod(threadIdx.x_1, 81)) && (floormod(threadIdx.x_1, 81) < 72)) && (1 <= floormod(threadIdx.x_1, 9))) && (floormod(threadIdx.x_1, 9) < 8)), data[(((((rc.outer.outer*98) + (floordiv(threadIdx.x_1, 81)*49)) + (floordiv(floormod(threadIdx.x_1, 81), 9)*7)) + floormod(threadIdx.x_1, 9)) - 8)], 0f32, dtype=float32)
- attr [IterVar(threadIdx.x_1, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 98;
- if @tir.likely((threadIdx.x_1 < 64), dtype=bool) {
- pad_temp.shared_1[(threadIdx.x_1 + 98)] = @tir.if_then_else((((threadIdx.x_1 < 55) && (1 <= floormod((threadIdx.x_1 + 8), 9))) && (floormod((threadIdx.x_1 + 8), 9) < 8)), data[(((((rc.outer.outer*98) + (floordiv((threadIdx.x_1 + 98), 81)*49)) + (floordiv((threadIdx.x_1 + 17), 9)*7)) + floormod((threadIdx.x_1 + 8), 9)) - 8)], 0f32, dtype=float32)
+ conv2d_nchw_1[4] = 0f32
+ conv2d_nchw_1[5] = 0f32
+ conv2d_nchw_1[6] = 0f32
+ conv2d_nchw_1[7] = 0f32
+ conv2d_nchw_1[8] = 0f32
+ conv2d_nchw_1[9] = 0f32
+ conv2d_nchw_1[10] = 0f32
+ conv2d_nchw_1[11] = 0f32
+ conv2d_nchw_1[12] = 0f32
+ conv2d_nchw_1[13] = 0f32
+ for (rc.outer.outer: int32, 0, 64) {
+ for (ry.outer.outer: int32, 0, 3) {
+ let cse_var_2: int32 = (rc.outer.outer*72)
+ let cse_var_1: int32 = (ry.outer.outer*3)
+ {
+ attr [IterVar(threadIdx.x_1: int32, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64 {
+ if @tir.likely((threadIdx.x_1 < 18), dtype=bool) {
+ pad_temp.shared_1: Buffer(pad_temp.shared, float32, [72], [], scope="shared")[(threadIdx.x_1*4)] = @tir.if_then_else(((((1 <= (ry.outer.outer + floormod(blockIdx.x, 7))) && ((ry.outer.outer + floormod(blockIdx.x, 7)) < 8)) && (1 <= floormod((threadIdx.x_1*4), 9))) && (floormod((threadIdx.x_1*4), 9) < 8)), data_3: Buffer(data_2, float32, [25088], [])[((((((rc.outer.outer*392) + (floordiv((threadIdx.x_1*4), 9)*49)) + (ry.outer.outer*7)) + (floormod(blockIdx.x, 7)*7)) + fl [...]
+ }
+ if @tir.likely((threadIdx.x_1 < 18), dtype=bool) {
+ pad_temp.shared_1[((threadIdx.x_1*4) + 1)] = @tir.if_then_else(((((1 <= (ry.outer.outer + floormod(blockIdx.x, 7))) && ((ry.outer.outer + floormod(blockIdx.x, 7)) < 8)) && (1 <= floormod(((threadIdx.x_1*4) + 1), 9))) && (floormod(((threadIdx.x_1*4) + 1), 9) < 8)), data_3[((((((rc.outer.outer*392) + (floordiv(((threadIdx.x_1*4) + 1), 9)*49)) + (ry.outer.outer*7)) + (floormod(blockIdx.x, 7)*7)) + floormod(((threadIdx.x_1*4) + 1), 9)) - 8)], 0f32, dtype=float32)
+ }
+ if @tir.likely((threadIdx.x_1 < 18), dtype=bool) {
+ pad_temp.shared_1[((threadIdx.x_1*4) + 2)] = @tir.if_then_else(((((1 <= (ry.outer.outer + floormod(blockIdx.x, 7))) && ((ry.outer.outer + floormod(blockIdx.x, 7)) < 8)) && (1 <= floormod(((threadIdx.x_1*4) + 2), 9))) && (floormod(((threadIdx.x_1*4) + 2), 9) < 8)), data_3[((((((rc.outer.outer*392) + (floordiv(((threadIdx.x_1*4) + 2), 9)*49)) + (ry.outer.outer*7)) + (floormod(blockIdx.x, 7)*7)) + floormod(((threadIdx.x_1*4) + 2), 9)) - 8)], 0f32, dtype=float32)
+ }
+ if @tir.likely((threadIdx.x_1 < 18), dtype=bool) {
+ pad_temp.shared_1[((threadIdx.x_1*4) + 3)] = @tir.if_then_else(((((1 <= (ry.outer.outer + floormod(blockIdx.x, 7))) && ((ry.outer.outer + floormod(blockIdx.x, 7)) < 8)) && (1 <= floormod(((threadIdx.x_1*4) + 3), 9))) && (floormod(((threadIdx.x_1*4) + 3), 9) < 8)), data_3[((((((rc.outer.outer*392) + (floordiv(((threadIdx.x_1*4) + 3), 9)*49)) + (ry.outer.outer*7)) + (floormod(blockIdx.x, 7)*7)) + floormod(((threadIdx.x_1*4) + 3), 9)) - 8)], 0f32, dtype=float32)
+ }
+ }
+ attr [IterVar(threadIdx.x_2: int32, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1: Buffer(kernel.shared, float32, [3072], [], scope="shared")[threadIdx.x_2] = kernel_3: Buffer(kernel_2, float32, [2359296], [])[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 64)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 64), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 128)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 128), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 192)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 36864)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 256)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 256), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 320)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 320), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 384)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 73728)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 448)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 448), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 512)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 512), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 576)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 110592)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 640)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 640), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 704)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 704), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 768)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 147456)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 832)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 832), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 896)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 896), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 960)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 184320)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1024)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 1024), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1088)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 1088), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1152)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 221184)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1216)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 1216), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1280)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 1280), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1344)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 258048)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1408)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 1408), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1472)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 1472), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1536)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 294912)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1600)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 1600), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1664)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 1664), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1728)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 331776)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1792)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 1792), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1856)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 1856), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1920)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 368640)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1984)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 1984), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2048)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 2048), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2112)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 405504)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2176)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 2176), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2240)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 2240), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2304)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 442368)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2368)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 2368), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2432)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 2432), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2496)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 479232)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2560)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 2560), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2624)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 2624), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2688)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 516096)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2752)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 2752), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2816)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 2816), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2880)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 552960)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2944)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 2944), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 3008)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 3008), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[0]*kernel.shared_1[(threadIdx.x*48)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[9]*kernel.shared_1[((threadIdx.x*48) + 3)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[1]*kernel.shared_1[(threadIdx.x*48)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[10]*kernel.shared_1[((threadIdx.x*48) + 3)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[2]*kernel.shared_1[(threadIdx.x*48)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[11]*kernel.shared_1[((threadIdx.x*48) + 3)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[3]*kernel.shared_1[(threadIdx.x*48)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[12]*kernel.shared_1[((threadIdx.x*48) + 3)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[4]*kernel.shared_1[(threadIdx.x*48)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[13]*kernel.shared_1[((threadIdx.x*48) + 3)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[5]*kernel.shared_1[(threadIdx.x*48)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[14]*kernel.shared_1[((threadIdx.x*48) + 3)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[6]*kernel.shared_1[(threadIdx.x*48)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[15]*kernel.shared_1[((threadIdx.x*48) + 3)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[0]*kernel.shared_1[((threadIdx.x*48) + 24)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[9]*kernel.shared_1[((threadIdx.x*48) + 27)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[1]*kernel.shared_1[((threadIdx.x*48) + 24)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[10]*kernel.shared_1[((threadIdx.x*48) + 27)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[2]*kernel.shared_1[((threadIdx.x*48) + 24)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[11]*kernel.shared_1[((threadIdx.x*48) + 27)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[3]*kernel.shared_1[((threadIdx.x*48) + 24)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[12]*kernel.shared_1[((threadIdx.x*48) + 27)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[4]*kernel.shared_1[((threadIdx.x*48) + 24)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[13]*kernel.shared_1[((threadIdx.x*48) + 27)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[5]*kernel.shared_1[((threadIdx.x*48) + 24)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[14]*kernel.shared_1[((threadIdx.x*48) + 27)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[6]*kernel.shared_1[((threadIdx.x*48) + 24)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[15]*kernel.shared_1[((threadIdx.x*48) + 27)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[1]*kernel.shared_1[((threadIdx.x*48) + 1)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[10]*kernel.shared_1[((threadIdx.x*48) + 4)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[2]*kernel.shared_1[((threadIdx.x*48) + 1)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[11]*kernel.shared_1[((threadIdx.x*48) + 4)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[3]*kernel.shared_1[((threadIdx.x*48) + 1)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[12]*kernel.shared_1[((threadIdx.x*48) + 4)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[4]*kernel.shared_1[((threadIdx.x*48) + 1)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[13]*kernel.shared_1[((threadIdx.x*48) + 4)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[5]*kernel.shared_1[((threadIdx.x*48) + 1)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[14]*kernel.shared_1[((threadIdx.x*48) + 4)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[6]*kernel.shared_1[((threadIdx.x*48) + 1)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[15]*kernel.shared_1[((threadIdx.x*48) + 4)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[7]*kernel.shared_1[((threadIdx.x*48) + 1)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[16]*kernel.shared_1[((threadIdx.x*48) + 4)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[1]*kernel.shared_1[((threadIdx.x*48) + 25)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[10]*kernel.shared_1[((threadIdx.x*48) + 28)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[2]*kernel.shared_1[((threadIdx.x*48) + 25)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[11]*kernel.shared_1[((threadIdx.x*48) + 28)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[3]*kernel.shared_1[((threadIdx.x*48) + 25)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[12]*kernel.shared_1[((threadIdx.x*48) + 28)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[4]*kernel.shared_1[((threadIdx.x*48) + 25)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[13]*kernel.shared_1[((threadIdx.x*48) + 28)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[5]*kernel.shared_1[((threadIdx.x*48) + 25)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[14]*kernel.shared_1[((threadIdx.x*48) + 28)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[6]*kernel.shared_1[((threadIdx.x*48) + 25)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[15]*kernel.shared_1[((threadIdx.x*48) + 28)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[7]*kernel.shared_1[((threadIdx.x*48) + 25)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[16]*kernel.shared_1[((threadIdx.x*48) + 28)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[2]*kernel.shared_1[((threadIdx.x*48) + 2)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[11]*kernel.shared_1[((threadIdx.x*48) + 5)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[3]*kernel.shared_1[((threadIdx.x*48) + 2)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[12]*kernel.shared_1[((threadIdx.x*48) + 5)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[4]*kernel.shared_1[((threadIdx.x*48) + 2)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[13]*kernel.shared_1[((threadIdx.x*48) + 5)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[5]*kernel.shared_1[((threadIdx.x*48) + 2)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[14]*kernel.shared_1[((threadIdx.x*48) + 5)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[6]*kernel.shared_1[((threadIdx.x*48) + 2)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[15]*kernel.shared_1[((threadIdx.x*48) + 5)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[7]*kernel.shared_1[((threadIdx.x*48) + 2)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[16]*kernel.shared_1[((threadIdx.x*48) + 5)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[8]*kernel.shared_1[((threadIdx.x*48) + 2)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[17]*kernel.shared_1[((threadIdx.x*48) + 5)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[2]*kernel.shared_1[((threadIdx.x*48) + 26)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[11]*kernel.shared_1[((threadIdx.x*48) + 29)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[3]*kernel.shared_1[((threadIdx.x*48) + 26)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[12]*kernel.shared_1[((threadIdx.x*48) + 29)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[4]*kernel.shared_1[((threadIdx.x*48) + 26)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[13]*kernel.shared_1[((threadIdx.x*48) + 29)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[5]*kernel.shared_1[((threadIdx.x*48) + 26)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[14]*kernel.shared_1[((threadIdx.x*48) + 29)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[6]*kernel.shared_1[((threadIdx.x*48) + 26)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[15]*kernel.shared_1[((threadIdx.x*48) + 29)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[7]*kernel.shared_1[((threadIdx.x*48) + 26)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[16]*kernel.shared_1[((threadIdx.x*48) + 29)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[8]*kernel.shared_1[((threadIdx.x*48) + 26)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[17]*kernel.shared_1[((threadIdx.x*48) + 29)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[18]*kernel.shared_1[((threadIdx.x*48) + 6)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[27]*kernel.shared_1[((threadIdx.x*48) + 9)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[19]*kernel.shared_1[((threadIdx.x*48) + 6)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[28]*kernel.shared_1[((threadIdx.x*48) + 9)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[20]*kernel.shared_1[((threadIdx.x*48) + 6)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[29]*kernel.shared_1[((threadIdx.x*48) + 9)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[21]*kernel.shared_1[((threadIdx.x*48) + 6)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[30]*kernel.shared_1[((threadIdx.x*48) + 9)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[22]*kernel.shared_1[((threadIdx.x*48) + 6)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[31]*kernel.shared_1[((threadIdx.x*48) + 9)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[23]*kernel.shared_1[((threadIdx.x*48) + 6)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[32]*kernel.shared_1[((threadIdx.x*48) + 9)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[24]*kernel.shared_1[((threadIdx.x*48) + 6)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[33]*kernel.shared_1[((threadIdx.x*48) + 9)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[18]*kernel.shared_1[((threadIdx.x*48) + 30)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[27]*kernel.shared_1[((threadIdx.x*48) + 33)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[19]*kernel.shared_1[((threadIdx.x*48) + 30)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[28]*kernel.shared_1[((threadIdx.x*48) + 33)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[20]*kernel.shared_1[((threadIdx.x*48) + 30)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[29]*kernel.shared_1[((threadIdx.x*48) + 33)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[21]*kernel.shared_1[((threadIdx.x*48) + 30)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[30]*kernel.shared_1[((threadIdx.x*48) + 33)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[22]*kernel.shared_1[((threadIdx.x*48) + 30)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[31]*kernel.shared_1[((threadIdx.x*48) + 33)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[23]*kernel.shared_1[((threadIdx.x*48) + 30)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[32]*kernel.shared_1[((threadIdx.x*48) + 33)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[24]*kernel.shared_1[((threadIdx.x*48) + 30)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[33]*kernel.shared_1[((threadIdx.x*48) + 33)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[19]*kernel.shared_1[((threadIdx.x*48) + 7)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[28]*kernel.shared_1[((threadIdx.x*48) + 10)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[20]*kernel.shared_1[((threadIdx.x*48) + 7)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[29]*kernel.shared_1[((threadIdx.x*48) + 10)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[21]*kernel.shared_1[((threadIdx.x*48) + 7)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[30]*kernel.shared_1[((threadIdx.x*48) + 10)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[22]*kernel.shared_1[((threadIdx.x*48) + 7)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[31]*kernel.shared_1[((threadIdx.x*48) + 10)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[23]*kernel.shared_1[((threadIdx.x*48) + 7)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[32]*kernel.shared_1[((threadIdx.x*48) + 10)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[24]*kernel.shared_1[((threadIdx.x*48) + 7)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[33]*kernel.shared_1[((threadIdx.x*48) + 10)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[25]*kernel.shared_1[((threadIdx.x*48) + 7)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[34]*kernel.shared_1[((threadIdx.x*48) + 10)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[19]*kernel.shared_1[((threadIdx.x*48) + 31)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[28]*kernel.shared_1[((threadIdx.x*48) + 34)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[20]*kernel.shared_1[((threadIdx.x*48) + 31)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[29]*kernel.shared_1[((threadIdx.x*48) + 34)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[21]*kernel.shared_1[((threadIdx.x*48) + 31)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[30]*kernel.shared_1[((threadIdx.x*48) + 34)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[22]*kernel.shared_1[((threadIdx.x*48) + 31)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[31]*kernel.shared_1[((threadIdx.x*48) + 34)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[23]*kernel.shared_1[((threadIdx.x*48) + 31)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[32]*kernel.shared_1[((threadIdx.x*48) + 34)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[24]*kernel.shared_1[((threadIdx.x*48) + 31)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[33]*kernel.shared_1[((threadIdx.x*48) + 34)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[25]*kernel.shared_1[((threadIdx.x*48) + 31)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[34]*kernel.shared_1[((threadIdx.x*48) + 34)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[20]*kernel.shared_1[((threadIdx.x*48) + 8)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[29]*kernel.shared_1[((threadIdx.x*48) + 11)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[21]*kernel.shared_1[((threadIdx.x*48) + 8)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[30]*kernel.shared_1[((threadIdx.x*48) + 11)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[22]*kernel.shared_1[((threadIdx.x*48) + 8)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[31]*kernel.shared_1[((threadIdx.x*48) + 11)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[23]*kernel.shared_1[((threadIdx.x*48) + 8)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[32]*kernel.shared_1[((threadIdx.x*48) + 11)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[24]*kernel.shared_1[((threadIdx.x*48) + 8)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[33]*kernel.shared_1[((threadIdx.x*48) + 11)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[25]*kernel.shared_1[((threadIdx.x*48) + 8)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[34]*kernel.shared_1[((threadIdx.x*48) + 11)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[26]*kernel.shared_1[((threadIdx.x*48) + 8)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[35]*kernel.shared_1[((threadIdx.x*48) + 11)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[20]*kernel.shared_1[((threadIdx.x*48) + 32)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[29]*kernel.shared_1[((threadIdx.x*48) + 35)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[21]*kernel.shared_1[((threadIdx.x*48) + 32)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[30]*kernel.shared_1[((threadIdx.x*48) + 35)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[22]*kernel.shared_1[((threadIdx.x*48) + 32)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[31]*kernel.shared_1[((threadIdx.x*48) + 35)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[23]*kernel.shared_1[((threadIdx.x*48) + 32)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[32]*kernel.shared_1[((threadIdx.x*48) + 35)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[24]*kernel.shared_1[((threadIdx.x*48) + 32)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[33]*kernel.shared_1[((threadIdx.x*48) + 35)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[25]*kernel.shared_1[((threadIdx.x*48) + 32)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[34]*kernel.shared_1[((threadIdx.x*48) + 35)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[26]*kernel.shared_1[((threadIdx.x*48) + 32)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[35]*kernel.shared_1[((threadIdx.x*48) + 35)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[36]*kernel.shared_1[((threadIdx.x*48) + 12)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[45]*kernel.shared_1[((threadIdx.x*48) + 15)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[37]*kernel.shared_1[((threadIdx.x*48) + 12)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[46]*kernel.shared_1[((threadIdx.x*48) + 15)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[38]*kernel.shared_1[((threadIdx.x*48) + 12)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[47]*kernel.shared_1[((threadIdx.x*48) + 15)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[39]*kernel.shared_1[((threadIdx.x*48) + 12)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[48]*kernel.shared_1[((threadIdx.x*48) + 15)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[40]*kernel.shared_1[((threadIdx.x*48) + 12)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[49]*kernel.shared_1[((threadIdx.x*48) + 15)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[41]*kernel.shared_1[((threadIdx.x*48) + 12)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[50]*kernel.shared_1[((threadIdx.x*48) + 15)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[42]*kernel.shared_1[((threadIdx.x*48) + 12)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[51]*kernel.shared_1[((threadIdx.x*48) + 15)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[36]*kernel.shared_1[((threadIdx.x*48) + 36)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[45]*kernel.shared_1[((threadIdx.x*48) + 39)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[37]*kernel.shared_1[((threadIdx.x*48) + 36)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[46]*kernel.shared_1[((threadIdx.x*48) + 39)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[38]*kernel.shared_1[((threadIdx.x*48) + 36)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[47]*kernel.shared_1[((threadIdx.x*48) + 39)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[39]*kernel.shared_1[((threadIdx.x*48) + 36)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[48]*kernel.shared_1[((threadIdx.x*48) + 39)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[40]*kernel.shared_1[((threadIdx.x*48) + 36)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[49]*kernel.shared_1[((threadIdx.x*48) + 39)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[41]*kernel.shared_1[((threadIdx.x*48) + 36)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[50]*kernel.shared_1[((threadIdx.x*48) + 39)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[42]*kernel.shared_1[((threadIdx.x*48) + 36)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[51]*kernel.shared_1[((threadIdx.x*48) + 39)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[37]*kernel.shared_1[((threadIdx.x*48) + 13)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[46]*kernel.shared_1[((threadIdx.x*48) + 16)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[38]*kernel.shared_1[((threadIdx.x*48) + 13)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[47]*kernel.shared_1[((threadIdx.x*48) + 16)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[39]*kernel.shared_1[((threadIdx.x*48) + 13)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[48]*kernel.shared_1[((threadIdx.x*48) + 16)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[40]*kernel.shared_1[((threadIdx.x*48) + 13)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[49]*kernel.shared_1[((threadIdx.x*48) + 16)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[41]*kernel.shared_1[((threadIdx.x*48) + 13)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[50]*kernel.shared_1[((threadIdx.x*48) + 16)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[42]*kernel.shared_1[((threadIdx.x*48) + 13)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[51]*kernel.shared_1[((threadIdx.x*48) + 16)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[43]*kernel.shared_1[((threadIdx.x*48) + 13)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[52]*kernel.shared_1[((threadIdx.x*48) + 16)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[37]*kernel.shared_1[((threadIdx.x*48) + 37)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[46]*kernel.shared_1[((threadIdx.x*48) + 40)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[38]*kernel.shared_1[((threadIdx.x*48) + 37)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[47]*kernel.shared_1[((threadIdx.x*48) + 40)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[39]*kernel.shared_1[((threadIdx.x*48) + 37)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[48]*kernel.shared_1[((threadIdx.x*48) + 40)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[40]*kernel.shared_1[((threadIdx.x*48) + 37)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[49]*kernel.shared_1[((threadIdx.x*48) + 40)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[41]*kernel.shared_1[((threadIdx.x*48) + 37)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[50]*kernel.shared_1[((threadIdx.x*48) + 40)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[42]*kernel.shared_1[((threadIdx.x*48) + 37)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[51]*kernel.shared_1[((threadIdx.x*48) + 40)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[43]*kernel.shared_1[((threadIdx.x*48) + 37)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[52]*kernel.shared_1[((threadIdx.x*48) + 40)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[38]*kernel.shared_1[((threadIdx.x*48) + 14)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[47]*kernel.shared_1[((threadIdx.x*48) + 17)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[39]*kernel.shared_1[((threadIdx.x*48) + 14)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[48]*kernel.shared_1[((threadIdx.x*48) + 17)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[40]*kernel.shared_1[((threadIdx.x*48) + 14)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[49]*kernel.shared_1[((threadIdx.x*48) + 17)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[41]*kernel.shared_1[((threadIdx.x*48) + 14)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[50]*kernel.shared_1[((threadIdx.x*48) + 17)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[42]*kernel.shared_1[((threadIdx.x*48) + 14)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[51]*kernel.shared_1[((threadIdx.x*48) + 17)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[43]*kernel.shared_1[((threadIdx.x*48) + 14)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[52]*kernel.shared_1[((threadIdx.x*48) + 17)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[44]*kernel.shared_1[((threadIdx.x*48) + 14)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[53]*kernel.shared_1[((threadIdx.x*48) + 17)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[38]*kernel.shared_1[((threadIdx.x*48) + 38)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[47]*kernel.shared_1[((threadIdx.x*48) + 41)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[39]*kernel.shared_1[((threadIdx.x*48) + 38)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[48]*kernel.shared_1[((threadIdx.x*48) + 41)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[40]*kernel.shared_1[((threadIdx.x*48) + 38)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[49]*kernel.shared_1[((threadIdx.x*48) + 41)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[41]*kernel.shared_1[((threadIdx.x*48) + 38)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[50]*kernel.shared_1[((threadIdx.x*48) + 41)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[42]*kernel.shared_1[((threadIdx.x*48) + 38)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[51]*kernel.shared_1[((threadIdx.x*48) + 41)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[43]*kernel.shared_1[((threadIdx.x*48) + 38)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[52]*kernel.shared_1[((threadIdx.x*48) + 41)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[44]*kernel.shared_1[((threadIdx.x*48) + 38)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[53]*kernel.shared_1[((threadIdx.x*48) + 41)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[54]*kernel.shared_1[((threadIdx.x*48) + 18)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[63]*kernel.shared_1[((threadIdx.x*48) + 21)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[55]*kernel.shared_1[((threadIdx.x*48) + 18)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[64]*kernel.shared_1[((threadIdx.x*48) + 21)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[56]*kernel.shared_1[((threadIdx.x*48) + 18)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[65]*kernel.shared_1[((threadIdx.x*48) + 21)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[57]*kernel.shared_1[((threadIdx.x*48) + 18)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[66]*kernel.shared_1[((threadIdx.x*48) + 21)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[58]*kernel.shared_1[((threadIdx.x*48) + 18)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[67]*kernel.shared_1[((threadIdx.x*48) + 21)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[59]*kernel.shared_1[((threadIdx.x*48) + 18)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[68]*kernel.shared_1[((threadIdx.x*48) + 21)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[60]*kernel.shared_1[((threadIdx.x*48) + 18)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[69]*kernel.shared_1[((threadIdx.x*48) + 21)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[54]*kernel.shared_1[((threadIdx.x*48) + 42)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[63]*kernel.shared_1[((threadIdx.x*48) + 45)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[55]*kernel.shared_1[((threadIdx.x*48) + 42)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[64]*kernel.shared_1[((threadIdx.x*48) + 45)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[56]*kernel.shared_1[((threadIdx.x*48) + 42)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[65]*kernel.shared_1[((threadIdx.x*48) + 45)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[57]*kernel.shared_1[((threadIdx.x*48) + 42)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[66]*kernel.shared_1[((threadIdx.x*48) + 45)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[58]*kernel.shared_1[((threadIdx.x*48) + 42)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[67]*kernel.shared_1[((threadIdx.x*48) + 45)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[59]*kernel.shared_1[((threadIdx.x*48) + 42)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[68]*kernel.shared_1[((threadIdx.x*48) + 45)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[60]*kernel.shared_1[((threadIdx.x*48) + 42)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[69]*kernel.shared_1[((threadIdx.x*48) + 45)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[55]*kernel.shared_1[((threadIdx.x*48) + 19)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[64]*kernel.shared_1[((threadIdx.x*48) + 22)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[56]*kernel.shared_1[((threadIdx.x*48) + 19)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[65]*kernel.shared_1[((threadIdx.x*48) + 22)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[57]*kernel.shared_1[((threadIdx.x*48) + 19)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[66]*kernel.shared_1[((threadIdx.x*48) + 22)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[58]*kernel.shared_1[((threadIdx.x*48) + 19)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[67]*kernel.shared_1[((threadIdx.x*48) + 22)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[59]*kernel.shared_1[((threadIdx.x*48) + 19)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[68]*kernel.shared_1[((threadIdx.x*48) + 22)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[60]*kernel.shared_1[((threadIdx.x*48) + 19)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[69]*kernel.shared_1[((threadIdx.x*48) + 22)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[61]*kernel.shared_1[((threadIdx.x*48) + 19)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[70]*kernel.shared_1[((threadIdx.x*48) + 22)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[55]*kernel.shared_1[((threadIdx.x*48) + 43)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[64]*kernel.shared_1[((threadIdx.x*48) + 46)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[56]*kernel.shared_1[((threadIdx.x*48) + 43)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[65]*kernel.shared_1[((threadIdx.x*48) + 46)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[57]*kernel.shared_1[((threadIdx.x*48) + 43)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[66]*kernel.shared_1[((threadIdx.x*48) + 46)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[58]*kernel.shared_1[((threadIdx.x*48) + 43)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[67]*kernel.shared_1[((threadIdx.x*48) + 46)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[59]*kernel.shared_1[((threadIdx.x*48) + 43)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[68]*kernel.shared_1[((threadIdx.x*48) + 46)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[60]*kernel.shared_1[((threadIdx.x*48) + 43)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[69]*kernel.shared_1[((threadIdx.x*48) + 46)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[61]*kernel.shared_1[((threadIdx.x*48) + 43)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[70]*kernel.shared_1[((threadIdx.x*48) + 46)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[56]*kernel.shared_1[((threadIdx.x*48) + 20)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[65]*kernel.shared_1[((threadIdx.x*48) + 23)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[57]*kernel.shared_1[((threadIdx.x*48) + 20)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[66]*kernel.shared_1[((threadIdx.x*48) + 23)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[58]*kernel.shared_1[((threadIdx.x*48) + 20)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[67]*kernel.shared_1[((threadIdx.x*48) + 23)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[59]*kernel.shared_1[((threadIdx.x*48) + 20)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[68]*kernel.shared_1[((threadIdx.x*48) + 23)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[60]*kernel.shared_1[((threadIdx.x*48) + 20)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[69]*kernel.shared_1[((threadIdx.x*48) + 23)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[61]*kernel.shared_1[((threadIdx.x*48) + 20)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[70]*kernel.shared_1[((threadIdx.x*48) + 23)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[62]*kernel.shared_1[((threadIdx.x*48) + 20)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[71]*kernel.shared_1[((threadIdx.x*48) + 23)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[56]*kernel.shared_1[((threadIdx.x*48) + 44)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[65]*kernel.shared_1[((threadIdx.x*48) + 47)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[57]*kernel.shared_1[((threadIdx.x*48) + 44)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[66]*kernel.shared_1[((threadIdx.x*48) + 47)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[58]*kernel.shared_1[((threadIdx.x*48) + 44)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[67]*kernel.shared_1[((threadIdx.x*48) + 47)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[59]*kernel.shared_1[((threadIdx.x*48) + 44)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[68]*kernel.shared_1[((threadIdx.x*48) + 47)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[60]*kernel.shared_1[((threadIdx.x*48) + 44)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[69]*kernel.shared_1[((threadIdx.x*48) + 47)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[61]*kernel.shared_1[((threadIdx.x*48) + 44)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[70]*kernel.shared_1[((threadIdx.x*48) + 47)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[62]*kernel.shared_1[((threadIdx.x*48) + 44)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[71]*kernel.shared_1[((threadIdx.x*48) + 47)]))
+ }
}
- attr [IterVar(threadIdx.x_2: int32, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 98;
- kernel.shared_1: Buffer(kernel.shared, float32, [144], [], scope="shared")[threadIdx.x_2] = kernel[((((blockIdx.x*36864) + (floordiv(threadIdx.x_2, 18)*4608)) + (rc.outer.outer*18)) + floormod(threadIdx.x_2, 18))]
- attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 98;
- if @tir.likely((threadIdx.x_2 < 46), dtype=bool) {
- kernel.shared_1[(threadIdx.x_2 + 98)] = kernel[(((((blockIdx.x*36864) + (floordiv((threadIdx.x_2 + 98), 18)*4608)) + (rc.outer.outer*18)) + (floordiv(floormod((threadIdx.x_2 + 8), 18), 3)*3)) + floormod((threadIdx.x_2 + 2), 3))]
- }
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7))]*kernel.shared_1[(floordiv(threadIdx.x, 49)*72)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7))]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 18)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 1)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 1)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 1)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 19)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 2)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 2)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 2)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 20)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 9)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 3)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 9)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 21)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 10)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 4)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 10)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 22)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 11)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 5)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 11)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 23)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 18)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 6)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 18)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 24)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 19)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 7)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 19)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 25)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 20)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 8)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 20)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 26)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7))]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 36)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7))]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 54)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 1)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 37)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 1)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 55)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 2)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 38)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 2)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 56)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 9)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 39)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 9)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 57)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 10)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 40)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 10)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 58)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 11)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 41)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 11)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 59)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 18)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 42)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 18)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 60)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 19)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 43)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 19)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 61)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 20)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 44)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 20)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 62)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 81)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 9)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 81)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 27)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 82)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 10)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 82)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 28)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 83)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 11)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 83)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 29)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 90)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 12)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 90)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 30)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 91)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 13)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 91)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 31)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 92)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 14)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 92)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 32)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 99)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 15)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 99)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 33)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 100)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 16)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 100)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 34)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 101)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 17)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 101)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 35)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 81)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 45)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 81)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 63)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 82)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 46)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 82)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 64)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 83)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 47)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 83)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 65)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 90)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 48)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 90)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 66)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 91)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 49)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 91)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 67)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 92)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 50)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 92)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 68)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 99)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 51)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 99)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 69)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 100)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 52)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 100)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 70)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 101)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 53)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 101)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 71)]))
}
- for (i1.inner: int32, 0, 4) {
- compute[((((blockIdx.x*392) + (floordiv(threadIdx.x, 49)*196)) + (i1.inner*49)) + floormod(threadIdx.x, 49))] = max((conv2d_nchw_1[i1.inner] + bias[(((blockIdx.x*8) + (floordiv(threadIdx.x, 49)*4)) + i1.inner)]), 0f32)
+ for (i1.inner: int32, 0, 2) {
+ for (i3.inner: int32, 0, 7) {
+ compute_3: Buffer(compute_2, float32, [25088], [])[(((((floordiv(blockIdx.x, 7)*6272) + (threadIdx.x*98)) + (i1.inner*49)) + (floormod(blockIdx.x, 7)*7)) + i3.inner)] = max((conv2d_nchw_1[((i1.inner*7) + i3.inner)] + bias_3: Buffer(bias_2, float32, [512], [])[(((floordiv(blockIdx.x, 7)*128) + (threadIdx.x*2)) + i1.inner)]), 0f32)
+ }
}
}
}
@@ -391,7 +770,7 @@ We build the binary and check its correctness and performance.
.. code-block:: none
- Execution time of this operator: 0.353 ms
+ Execution time of this operator: 0.349 ms
@@ -439,36 +818,36 @@ They can be used for debugging and learning the behavior of the auto-scheduler.
conv2d_nchw_nn_o_o_i, conv2d_nchw_nn_o_i = s[conv2d_nchw].split(conv2d_nchw_nn_o_i, factor=1)
conv2d_nchw_nn_o_o_o_i, conv2d_nchw_nn_o_o_i = s[conv2d_nchw].split(conv2d_nchw_nn_o_o_i, factor=1)
conv2d_nchw_nn_o_o_o_o, conv2d_nchw_nn_o_o_o_i = s[conv2d_nchw].split(conv2d_nchw_nn_o_o_o_i, factor=1)
- conv2d_nchw_ff_o_i, conv2d_nchw_ff_i = s[conv2d_nchw].split(conv2d_nchw_ff, factor=2)
+ conv2d_nchw_ff_o_i, conv2d_nchw_ff_i = s[conv2d_nchw].split(conv2d_nchw_ff, factor=1)
conv2d_nchw_ff_o_o_i, conv2d_nchw_ff_o_i = s[conv2d_nchw].split(conv2d_nchw_ff_o_i, factor=2)
- conv2d_nchw_ff_o_o_o_i, conv2d_nchw_ff_o_o_i = s[conv2d_nchw].split(conv2d_nchw_ff_o_o_i, factor=2)
+ conv2d_nchw_ff_o_o_o_i, conv2d_nchw_ff_o_o_i = s[conv2d_nchw].split(conv2d_nchw_ff_o_o_i, factor=64)
conv2d_nchw_ff_o_o_o_o, conv2d_nchw_ff_o_o_o_i = s[conv2d_nchw].split(conv2d_nchw_ff_o_o_o_i, factor=1)
conv2d_nchw_yy_o_i, conv2d_nchw_yy_i = s[conv2d_nchw].split(conv2d_nchw_yy, factor=1)
conv2d_nchw_yy_o_o_i, conv2d_nchw_yy_o_i = s[conv2d_nchw].split(conv2d_nchw_yy_o_i, factor=1)
- conv2d_nchw_yy_o_o_o_i, conv2d_nchw_yy_o_o_i = s[conv2d_nchw].split(conv2d_nchw_yy_o_o_i, factor=7)
+ conv2d_nchw_yy_o_o_o_i, conv2d_nchw_yy_o_o_i = s[conv2d_nchw].split(conv2d_nchw_yy_o_o_i, factor=1)
conv2d_nchw_yy_o_o_o_o, conv2d_nchw_yy_o_o_o_i = s[conv2d_nchw].split(conv2d_nchw_yy_o_o_o_i, factor=1)
conv2d_nchw_xx_o_i, conv2d_nchw_xx_i = s[conv2d_nchw].split(conv2d_nchw_xx, factor=1)
- conv2d_nchw_xx_o_o_i, conv2d_nchw_xx_o_i = s[conv2d_nchw].split(conv2d_nchw_xx_o_i, factor=1)
- conv2d_nchw_xx_o_o_o_i, conv2d_nchw_xx_o_o_i = s[conv2d_nchw].split(conv2d_nchw_xx_o_o_i, factor=7)
+ conv2d_nchw_xx_o_o_i, conv2d_nchw_xx_o_i = s[conv2d_nchw].split(conv2d_nchw_xx_o_i, factor=7)
+ conv2d_nchw_xx_o_o_o_i, conv2d_nchw_xx_o_o_i = s[conv2d_nchw].split(conv2d_nchw_xx_o_o_i, factor=1)
conv2d_nchw_xx_o_o_o_o, conv2d_nchw_xx_o_o_o_i = s[conv2d_nchw].split(conv2d_nchw_xx_o_o_o_i, factor=1)
- conv2d_nchw_rc_o_i, conv2d_nchw_rc_i = s[conv2d_nchw].split(conv2d_nchw_rc, factor=1)
- conv2d_nchw_rc_o_o, conv2d_nchw_rc_o_i = s[conv2d_nchw].split(conv2d_nchw_rc_o_i, factor=2)
- conv2d_nchw_ry_o_i, conv2d_nchw_ry_i = s[conv2d_nchw].split(conv2d_nchw_ry, factor=3)
+ conv2d_nchw_rc_o_i, conv2d_nchw_rc_i = s[conv2d_nchw].split(conv2d_nchw_rc, factor=2)
+ conv2d_nchw_rc_o_o, conv2d_nchw_rc_o_i = s[conv2d_nchw].split(conv2d_nchw_rc_o_i, factor=4)
+ conv2d_nchw_ry_o_i, conv2d_nchw_ry_i = s[conv2d_nchw].split(conv2d_nchw_ry, factor=1)
conv2d_nchw_ry_o_o, conv2d_nchw_ry_o_i = s[conv2d_nchw].split(conv2d_nchw_ry_o_i, factor=1)
- conv2d_nchw_rx_o_i, conv2d_nchw_rx_i = s[conv2d_nchw].split(conv2d_nchw_rx, factor=3)
- conv2d_nchw_rx_o_o, conv2d_nchw_rx_o_i = s[conv2d_nchw].split(conv2d_nchw_rx_o_i, factor=1)
+ conv2d_nchw_rx_o_i, conv2d_nchw_rx_i = s[conv2d_nchw].split(conv2d_nchw_rx, factor=1)
+ conv2d_nchw_rx_o_o, conv2d_nchw_rx_o_i = s[conv2d_nchw].split(conv2d_nchw_rx_o_i, factor=3)
s[conv2d_nchw].reorder(conv2d_nchw_nn_o_o_o_o, conv2d_nchw_ff_o_o_o_o, conv2d_nchw_yy_o_o_o_o, conv2d_nchw_xx_o_o_o_o, conv2d_nchw_nn_o_o_o_i, conv2d_nchw_ff_o_o_o_i, conv2d_nchw_yy_o_o_o_i, conv2d_nchw_xx_o_o_o_i, conv2d_nchw_nn_o_o_i, conv2d_nchw_ff_o_o_i, conv2d_nchw_yy_o_o_i, conv2d_nchw_xx_o_o_i, conv2d_nchw_rc_o_o, conv2d_nchw_ry_o_o, conv2d_nchw_rx_o_o, conv2d_nchw_rc_o_i, conv2d_nchw_ry_o_i, conv2d_nchw_rx_o_i, conv2d_nchw_nn_o_i, conv2d_nchw_ff_o_i, conv2d_nchw_yy_o_i, conv2 [...]
compute_i0_o_i, compute_i0_i = s[compute].split(compute_i0, factor=1)
compute_i0_o_o_i, compute_i0_o_i = s[compute].split(compute_i0_o_i, factor=1)
compute_i0_o_o_o, compute_i0_o_o_i = s[compute].split(compute_i0_o_o_i, factor=1)
- compute_i1_o_i, compute_i1_i = s[compute].split(compute_i1, factor=4)
- compute_i1_o_o_i, compute_i1_o_i = s[compute].split(compute_i1_o_i, factor=2)
+ compute_i1_o_i, compute_i1_i = s[compute].split(compute_i1, factor=2)
+ compute_i1_o_o_i, compute_i1_o_i = s[compute].split(compute_i1_o_i, factor=64)
compute_i1_o_o_o, compute_i1_o_o_i = s[compute].split(compute_i1_o_o_i, factor=1)
compute_i2_o_i, compute_i2_i = s[compute].split(compute_i2, factor=1)
- compute_i2_o_o_i, compute_i2_o_i = s[compute].split(compute_i2_o_i, factor=7)
+ compute_i2_o_o_i, compute_i2_o_i = s[compute].split(compute_i2_o_i, factor=1)
compute_i2_o_o_o, compute_i2_o_o_i = s[compute].split(compute_i2_o_o_i, factor=1)
- compute_i3_o_i, compute_i3_i = s[compute].split(compute_i3, factor=1)
- compute_i3_o_o_i, compute_i3_o_i = s[compute].split(compute_i3_o_i, factor=7)
+ compute_i3_o_i, compute_i3_i = s[compute].split(compute_i3, factor=7)
+ compute_i3_o_o_i, compute_i3_o_i = s[compute].split(compute_i3_o_i, factor=1)
compute_i3_o_o_o, compute_i3_o_o_i = s[compute].split(compute_i3_o_o_i, factor=1)
s[compute].reorder(compute_i0_o_o_o, compute_i1_o_o_o, compute_i2_o_o_o, compute_i3_o_o_o, compute_i0_o_o_i, compute_i1_o_o_i, compute_i2_o_o_i, compute_i3_o_o_i, compute_i0_o_i, compute_i1_o_i, compute_i2_o_i, compute_i3_o_i, compute_i0_i, compute_i1_i, compute_i2_i, compute_i3_i)
s[conv2d_nchw].compute_at(s[compute], compute_i3_o_i)
@@ -488,12 +867,12 @@ They can be used for debugging and learning the behavior of the auto-scheduler.
kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused = s[kernel_shared].fuse(kernel_shared_ax0, kernel_shared_ax1, kernel_shared_ax2, kernel_shared_ax3)
kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o, kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_i = s[kernel_shared].split(kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused, factor=1)
s[kernel_shared].vectorize(kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_i)
- kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_o, kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_i = s[kernel_shared].split(kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o, factor=98)
+ kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_o, kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_i = s[kernel_shared].split(kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o, factor=64)
s[kernel_shared].bind(kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_i, te.thread_axis("threadIdx.x"))
pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused = s[pad_temp_shared].fuse(pad_temp_shared_ax0, pad_temp_shared_ax1, pad_temp_shared_ax2, pad_temp_shared_ax3)
- pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o, pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_i = s[pad_temp_shared].split(pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused, factor=1)
+ pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o, pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_i = s[pad_temp_shared].split(pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused, factor=4)
s[pad_temp_shared].vectorize(pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_i)
- pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_o, pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_i = s[pad_temp_shared].split(pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o, factor=98)
+ pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_o, pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_i = s[pad_temp_shared].split(pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o, factor=64)
s[pad_temp_shared].bind(pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_i, te.thread_axis("threadIdx.x"))
s[conv2d_nchw].pragma(conv2d_nchw_nn_o_o_o_o, "auto_unroll_max_step", 512)
s[conv2d_nchw].pragma(conv2d_nchw_nn_o_o_o_o, "unroll_explicit", True)
@@ -513,100 +892,430 @@ They can be used for debugging and learning the behavior of the auto-scheduler.
#define int64_t long long
#define uint64_t unsigned long long
#endif
- extern "C" __global__ void __launch_bounds__(98) default_function_kernel0(float* __restrict__ data, float* __restrict__ kernel, float* __restrict__ compute, float* __restrict__ bias) {
- float conv2d_nchw[4];
- __shared__ float pad_temp_shared[162];
- __shared__ float kernel_shared[144];
+ extern "C" __global__ void __launch_bounds__(64) default_function_kernel0(float* __restrict__ data, float* __restrict__ kernel, float* __restrict__ compute, float* __restrict__ bias) {
+ float conv2d_nchw[14];
+ __shared__ float pad_temp_shared[72];
+ __shared__ float kernel_shared[3072];
conv2d_nchw[0] = 0.000000e+00f;
conv2d_nchw[1] = 0.000000e+00f;
conv2d_nchw[2] = 0.000000e+00f;
conv2d_nchw[3] = 0.000000e+00f;
- for (int rc_outer_outer = 0; rc_outer_outer < 256; ++rc_outer_outer) {
- __syncthreads();
- pad_temp_shared[((int)threadIdx.x)] = (((((9 <= (((int)threadIdx.x) % 81)) && ((((int)threadIdx.x) % 81) < 72)) && (1 <= (((int)threadIdx.x) % 9))) && ((((int)threadIdx.x) % 9) < 8)) ? data[(((((rc_outer_outer * 98) + ((((int)threadIdx.x) / 81) * 49)) + (((((int)threadIdx.x) % 81) / 9) * 7)) + (((int)threadIdx.x) % 9)) - 8)] : 0.000000e+00f);
- if (((int)threadIdx.x) < 64) {
- pad_temp_shared[(((int)threadIdx.x) + 98)] = ((((((int)threadIdx.x) < 55) && (1 <= ((((int)threadIdx.x) + 8) % 9))) && (((((int)threadIdx.x) + 8) % 9) < 8)) ? data[(((((rc_outer_outer * 98) + (((((int)threadIdx.x) + 98) / 81) * 49)) + (((((int)threadIdx.x) + 17) / 9) * 7)) + ((((int)threadIdx.x) + 8) % 9)) - 8)] : 0.000000e+00f);
- }
- kernel_shared[((int)threadIdx.x)] = kernel[((((((int)blockIdx.x) * 36864) + ((((int)threadIdx.x) / 18) * 4608)) + (rc_outer_outer * 18)) + (((int)threadIdx.x) % 18))];
- if (((int)threadIdx.x) < 46) {
- kernel_shared[(((int)threadIdx.x) + 98)] = kernel[(((((((int)blockIdx.x) * 36864) + (((((int)threadIdx.x) + 98) / 18) * 4608)) + (rc_outer_outer * 18)) + ((((((int)threadIdx.x) + 8) % 18) / 3) * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ conv2d_nchw[4] = 0.000000e+00f;
+ conv2d_nchw[5] = 0.000000e+00f;
+ conv2d_nchw[6] = 0.000000e+00f;
+ conv2d_nchw[7] = 0.000000e+00f;
+ conv2d_nchw[8] = 0.000000e+00f;
+ conv2d_nchw[9] = 0.000000e+00f;
+ conv2d_nchw[10] = 0.000000e+00f;
+ conv2d_nchw[11] = 0.000000e+00f;
+ conv2d_nchw[12] = 0.000000e+00f;
+ conv2d_nchw[13] = 0.000000e+00f;
+ for (int rc_outer_outer = 0; rc_outer_outer < 64; ++rc_outer_outer) {
+ for (int ry_outer_outer = 0; ry_outer_outer < 3; ++ry_outer_outer) {
+ __syncthreads();
+ if (((int)threadIdx.x) < 18) {
+ pad_temp_shared[(((int)threadIdx.x) * 4)] = (((((1 <= (ry_outer_outer + (((int)blockIdx.x) % 7))) && ((ry_outer_outer + (((int)blockIdx.x) % 7)) < 8)) && (1 <= ((((int)threadIdx.x) * 4) % 9))) && (((((int)threadIdx.x) * 4) % 9) < 8)) ? data[((((((rc_outer_outer * 392) + (((((int)threadIdx.x) * 4) / 9) * 49)) + (ry_outer_outer * 7)) + ((((int)blockIdx.x) % 7) * 7)) + ((((int)threadIdx.x) * 4) % 9)) - 8)] : 0.000000e+00f);
+ }
+ if (((int)threadIdx.x) < 18) {
+ pad_temp_shared[((((int)threadIdx.x) * 4) + 1)] = (((((1 <= (ry_outer_outer + (((int)blockIdx.x) % 7))) && ((ry_outer_outer + (((int)blockIdx.x) % 7)) < 8)) && (1 <= (((((int)threadIdx.x) * 4) + 1) % 9))) && ((((((int)threadIdx.x) * 4) + 1) % 9) < 8)) ? data[((((((rc_outer_outer * 392) + ((((((int)threadIdx.x) * 4) + 1) / 9) * 49)) + (ry_outer_outer * 7)) + ((((int)blockIdx.x) % 7) * 7)) + (((((int)threadIdx.x) * 4) + 1) % 9)) - 8)] : 0.000000e+00f);
+ }
+ if (((int)threadIdx.x) < 18) {
+ pad_temp_shared[((((int)threadIdx.x) * 4) + 2)] = (((((1 <= (ry_outer_outer + (((int)blockIdx.x) % 7))) && ((ry_outer_outer + (((int)blockIdx.x) % 7)) < 8)) && (1 <= (((((int)threadIdx.x) * 4) + 2) % 9))) && ((((((int)threadIdx.x) * 4) + 2) % 9) < 8)) ? data[((((((rc_outer_outer * 392) + ((((((int)threadIdx.x) * 4) + 2) / 9) * 49)) + (ry_outer_outer * 7)) + ((((int)blockIdx.x) % 7) * 7)) + (((((int)threadIdx.x) * 4) + 2) % 9)) - 8)] : 0.000000e+00f);
+ }
+ if (((int)threadIdx.x) < 18) {
+ pad_temp_shared[((((int)threadIdx.x) * 4) + 3)] = (((((1 <= (ry_outer_outer + (((int)blockIdx.x) % 7))) && ((ry_outer_outer + (((int)blockIdx.x) % 7)) < 8)) && (1 <= (((((int)threadIdx.x) * 4) + 3) % 9))) && ((((((int)threadIdx.x) * 4) + 3) % 9) < 8)) ? data[((((((rc_outer_outer * 392) + ((((((int)threadIdx.x) * 4) + 3) / 9) * 49)) + (ry_outer_outer * 7)) + ((((int)blockIdx.x) % 7) * 7)) + (((((int)threadIdx.x) * 4) + 3) % 9)) - 8)] : 0.000000e+00f);
+ }
+ kernel_shared[((int)threadIdx.x)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 64)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 64) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 128)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 128) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 192)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 36864)];
+ kernel_shared[(((int)threadIdx.x) + 256)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 256) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 320)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 320) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 384)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 73728)];
+ kernel_shared[(((int)threadIdx.x) + 448)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 448) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 512)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 512) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 576)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 110592)];
+ kernel_shared[(((int)threadIdx.x) + 640)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 640) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 704)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 704) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 768)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 147456)];
+ kernel_shared[(((int)threadIdx.x) + 832)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 832) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 896)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 896) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 960)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 184320)];
+ kernel_shared[(((int)threadIdx.x) + 1024)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1024) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 1088)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1088) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 1152)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 221184)];
+ kernel_shared[(((int)threadIdx.x) + 1216)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1216) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 1280)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1280) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 1344)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 258048)];
+ kernel_shared[(((int)threadIdx.x) + 1408)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1408) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 1472)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1472) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 1536)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 294912)];
+ kernel_shared[(((int)threadIdx.x) + 1600)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1600) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 1664)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1664) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 1728)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 331776)];
+ kernel_shared[(((int)threadIdx.x) + 1792)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1792) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 1856)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1856) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 1920)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 368640)];
+ kernel_shared[(((int)threadIdx.x) + 1984)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1984) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 2048)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2048) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 2112)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 405504)];
+ kernel_shared[(((int)threadIdx.x) + 2176)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2176) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 2240)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2240) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 2304)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 442368)];
+ kernel_shared[(((int)threadIdx.x) + 2368)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2368) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 2432)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2432) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 2496)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 479232)];
+ kernel_shared[(((int)threadIdx.x) + 2560)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2560) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 2624)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2624) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 2688)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 516096)];
+ kernel_shared[(((int)threadIdx.x) + 2752)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2752) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 2816)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2816) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 2880)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 552960)];
+ kernel_shared[(((int)threadIdx.x) + 2944)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2944) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 3008)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 3008) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ __syncthreads();
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[0] * kernel_shared[(((int)threadIdx.x) * 48)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[9] * kernel_shared[((((int)threadIdx.x) * 48) + 3)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[1] * kernel_shared[(((int)threadIdx.x) * 48)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[10] * kernel_shared[((((int)threadIdx.x) * 48) + 3)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[2] * kernel_shared[(((int)threadIdx.x) * 48)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[11] * kernel_shared[((((int)threadIdx.x) * 48) + 3)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[3] * kernel_shared[(((int)threadIdx.x) * 48)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[12] * kernel_shared[((((int)threadIdx.x) * 48) + 3)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[4] * kernel_shared[(((int)threadIdx.x) * 48)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[13] * kernel_shared[((((int)threadIdx.x) * 48) + 3)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[5] * kernel_shared[(((int)threadIdx.x) * 48)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[14] * kernel_shared[((((int)threadIdx.x) * 48) + 3)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[6] * kernel_shared[(((int)threadIdx.x) * 48)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[15] * kernel_shared[((((int)threadIdx.x) * 48) + 3)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[0] * kernel_shared[((((int)threadIdx.x) * 48) + 24)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[9] * kernel_shared[((((int)threadIdx.x) * 48) + 27)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[1] * kernel_shared[((((int)threadIdx.x) * 48) + 24)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[10] * kernel_shared[((((int)threadIdx.x) * 48) + 27)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[2] * kernel_shared[((((int)threadIdx.x) * 48) + 24)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[11] * kernel_shared[((((int)threadIdx.x) * 48) + 27)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[3] * kernel_shared[((((int)threadIdx.x) * 48) + 24)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[12] * kernel_shared[((((int)threadIdx.x) * 48) + 27)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[4] * kernel_shared[((((int)threadIdx.x) * 48) + 24)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[13] * kernel_shared[((((int)threadIdx.x) * 48) + 27)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[5] * kernel_shared[((((int)threadIdx.x) * 48) + 24)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[14] * kernel_shared[((((int)threadIdx.x) * 48) + 27)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[6] * kernel_shared[((((int)threadIdx.x) * 48) + 24)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[15] * kernel_shared[((((int)threadIdx.x) * 48) + 27)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[1] * kernel_shared[((((int)threadIdx.x) * 48) + 1)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[10] * kernel_shared[((((int)threadIdx.x) * 48) + 4)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[2] * kernel_shared[((((int)threadIdx.x) * 48) + 1)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[11] * kernel_shared[((((int)threadIdx.x) * 48) + 4)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[3] * kernel_shared[((((int)threadIdx.x) * 48) + 1)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[12] * kernel_shared[((((int)threadIdx.x) * 48) + 4)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[4] * kernel_shared[((((int)threadIdx.x) * 48) + 1)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[13] * kernel_shared[((((int)threadIdx.x) * 48) + 4)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[5] * kernel_shared[((((int)threadIdx.x) * 48) + 1)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[14] * kernel_shared[((((int)threadIdx.x) * 48) + 4)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[6] * kernel_shared[((((int)threadIdx.x) * 48) + 1)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[15] * kernel_shared[((((int)threadIdx.x) * 48) + 4)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[7] * kernel_shared[((((int)threadIdx.x) * 48) + 1)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[16] * kernel_shared[((((int)threadIdx.x) * 48) + 4)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[1] * kernel_shared[((((int)threadIdx.x) * 48) + 25)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[10] * kernel_shared[((((int)threadIdx.x) * 48) + 28)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[2] * kernel_shared[((((int)threadIdx.x) * 48) + 25)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[11] * kernel_shared[((((int)threadIdx.x) * 48) + 28)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[3] * kernel_shared[((((int)threadIdx.x) * 48) + 25)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[12] * kernel_shared[((((int)threadIdx.x) * 48) + 28)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[4] * kernel_shared[((((int)threadIdx.x) * 48) + 25)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[13] * kernel_shared[((((int)threadIdx.x) * 48) + 28)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[5] * kernel_shared[((((int)threadIdx.x) * 48) + 25)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[14] * kernel_shared[((((int)threadIdx.x) * 48) + 28)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[6] * kernel_shared[((((int)threadIdx.x) * 48) + 25)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[15] * kernel_shared[((((int)threadIdx.x) * 48) + 28)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[7] * kernel_shared[((((int)threadIdx.x) * 48) + 25)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[16] * kernel_shared[((((int)threadIdx.x) * 48) + 28)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[2] * kernel_shared[((((int)threadIdx.x) * 48) + 2)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[11] * kernel_shared[((((int)threadIdx.x) * 48) + 5)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[3] * kernel_shared[((((int)threadIdx.x) * 48) + 2)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[12] * kernel_shared[((((int)threadIdx.x) * 48) + 5)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[4] * kernel_shared[((((int)threadIdx.x) * 48) + 2)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[13] * kernel_shared[((((int)threadIdx.x) * 48) + 5)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[5] * kernel_shared[((((int)threadIdx.x) * 48) + 2)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[14] * kernel_shared[((((int)threadIdx.x) * 48) + 5)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[6] * kernel_shared[((((int)threadIdx.x) * 48) + 2)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[15] * kernel_shared[((((int)threadIdx.x) * 48) + 5)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[7] * kernel_shared[((((int)threadIdx.x) * 48) + 2)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[16] * kernel_shared[((((int)threadIdx.x) * 48) + 5)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[8] * kernel_shared[((((int)threadIdx.x) * 48) + 2)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[17] * kernel_shared[((((int)threadIdx.x) * 48) + 5)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[2] * kernel_shared[((((int)threadIdx.x) * 48) + 26)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[11] * kernel_shared[((((int)threadIdx.x) * 48) + 29)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[3] * kernel_shared[((((int)threadIdx.x) * 48) + 26)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[12] * kernel_shared[((((int)threadIdx.x) * 48) + 29)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[4] * kernel_shared[((((int)threadIdx.x) * 48) + 26)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[13] * kernel_shared[((((int)threadIdx.x) * 48) + 29)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[5] * kernel_shared[((((int)threadIdx.x) * 48) + 26)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[14] * kernel_shared[((((int)threadIdx.x) * 48) + 29)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[6] * kernel_shared[((((int)threadIdx.x) * 48) + 26)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[15] * kernel_shared[((((int)threadIdx.x) * 48) + 29)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[7] * kernel_shared[((((int)threadIdx.x) * 48) + 26)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[16] * kernel_shared[((((int)threadIdx.x) * 48) + 29)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[8] * kernel_shared[((((int)threadIdx.x) * 48) + 26)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[17] * kernel_shared[((((int)threadIdx.x) * 48) + 29)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[18] * kernel_shared[((((int)threadIdx.x) * 48) + 6)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[27] * kernel_shared[((((int)threadIdx.x) * 48) + 9)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[19] * kernel_shared[((((int)threadIdx.x) * 48) + 6)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[28] * kernel_shared[((((int)threadIdx.x) * 48) + 9)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[20] * kernel_shared[((((int)threadIdx.x) * 48) + 6)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[29] * kernel_shared[((((int)threadIdx.x) * 48) + 9)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[21] * kernel_shared[((((int)threadIdx.x) * 48) + 6)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[30] * kernel_shared[((((int)threadIdx.x) * 48) + 9)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[22] * kernel_shared[((((int)threadIdx.x) * 48) + 6)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[31] * kernel_shared[((((int)threadIdx.x) * 48) + 9)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[23] * kernel_shared[((((int)threadIdx.x) * 48) + 6)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[32] * kernel_shared[((((int)threadIdx.x) * 48) + 9)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[24] * kernel_shared[((((int)threadIdx.x) * 48) + 6)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[33] * kernel_shared[((((int)threadIdx.x) * 48) + 9)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[18] * kernel_shared[((((int)threadIdx.x) * 48) + 30)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[27] * kernel_shared[((((int)threadIdx.x) * 48) + 33)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[19] * kernel_shared[((((int)threadIdx.x) * 48) + 30)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[28] * kernel_shared[((((int)threadIdx.x) * 48) + 33)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[20] * kernel_shared[((((int)threadIdx.x) * 48) + 30)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[29] * kernel_shared[((((int)threadIdx.x) * 48) + 33)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[21] * kernel_shared[((((int)threadIdx.x) * 48) + 30)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[30] * kernel_shared[((((int)threadIdx.x) * 48) + 33)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[22] * kernel_shared[((((int)threadIdx.x) * 48) + 30)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[31] * kernel_shared[((((int)threadIdx.x) * 48) + 33)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[23] * kernel_shared[((((int)threadIdx.x) * 48) + 30)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[32] * kernel_shared[((((int)threadIdx.x) * 48) + 33)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[24] * kernel_shared[((((int)threadIdx.x) * 48) + 30)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[33] * kernel_shared[((((int)threadIdx.x) * 48) + 33)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[19] * kernel_shared[((((int)threadIdx.x) * 48) + 7)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[28] * kernel_shared[((((int)threadIdx.x) * 48) + 10)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[20] * kernel_shared[((((int)threadIdx.x) * 48) + 7)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[29] * kernel_shared[((((int)threadIdx.x) * 48) + 10)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[21] * kernel_shared[((((int)threadIdx.x) * 48) + 7)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[30] * kernel_shared[((((int)threadIdx.x) * 48) + 10)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[22] * kernel_shared[((((int)threadIdx.x) * 48) + 7)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[31] * kernel_shared[((((int)threadIdx.x) * 48) + 10)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[23] * kernel_shared[((((int)threadIdx.x) * 48) + 7)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[32] * kernel_shared[((((int)threadIdx.x) * 48) + 10)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[24] * kernel_shared[((((int)threadIdx.x) * 48) + 7)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[33] * kernel_shared[((((int)threadIdx.x) * 48) + 10)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[25] * kernel_shared[((((int)threadIdx.x) * 48) + 7)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[34] * kernel_shared[((((int)threadIdx.x) * 48) + 10)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[19] * kernel_shared[((((int)threadIdx.x) * 48) + 31)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[28] * kernel_shared[((((int)threadIdx.x) * 48) + 34)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[20] * kernel_shared[((((int)threadIdx.x) * 48) + 31)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[29] * kernel_shared[((((int)threadIdx.x) * 48) + 34)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[21] * kernel_shared[((((int)threadIdx.x) * 48) + 31)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[30] * kernel_shared[((((int)threadIdx.x) * 48) + 34)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[22] * kernel_shared[((((int)threadIdx.x) * 48) + 31)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[31] * kernel_shared[((((int)threadIdx.x) * 48) + 34)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[23] * kernel_shared[((((int)threadIdx.x) * 48) + 31)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[32] * kernel_shared[((((int)threadIdx.x) * 48) + 34)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[24] * kernel_shared[((((int)threadIdx.x) * 48) + 31)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[33] * kernel_shared[((((int)threadIdx.x) * 48) + 34)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[25] * kernel_shared[((((int)threadIdx.x) * 48) + 31)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[34] * kernel_shared[((((int)threadIdx.x) * 48) + 34)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[20] * kernel_shared[((((int)threadIdx.x) * 48) + 8)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[29] * kernel_shared[((((int)threadIdx.x) * 48) + 11)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[21] * kernel_shared[((((int)threadIdx.x) * 48) + 8)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[30] * kernel_shared[((((int)threadIdx.x) * 48) + 11)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[22] * kernel_shared[((((int)threadIdx.x) * 48) + 8)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[31] * kernel_shared[((((int)threadIdx.x) * 48) + 11)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[23] * kernel_shared[((((int)threadIdx.x) * 48) + 8)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[32] * kernel_shared[((((int)threadIdx.x) * 48) + 11)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[24] * kernel_shared[((((int)threadIdx.x) * 48) + 8)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[33] * kernel_shared[((((int)threadIdx.x) * 48) + 11)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[25] * kernel_shared[((((int)threadIdx.x) * 48) + 8)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[34] * kernel_shared[((((int)threadIdx.x) * 48) + 11)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[26] * kernel_shared[((((int)threadIdx.x) * 48) + 8)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[35] * kernel_shared[((((int)threadIdx.x) * 48) + 11)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[20] * kernel_shared[((((int)threadIdx.x) * 48) + 32)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[29] * kernel_shared[((((int)threadIdx.x) * 48) + 35)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[21] * kernel_shared[((((int)threadIdx.x) * 48) + 32)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[30] * kernel_shared[((((int)threadIdx.x) * 48) + 35)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[22] * kernel_shared[((((int)threadIdx.x) * 48) + 32)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[31] * kernel_shared[((((int)threadIdx.x) * 48) + 35)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[23] * kernel_shared[((((int)threadIdx.x) * 48) + 32)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[32] * kernel_shared[((((int)threadIdx.x) * 48) + 35)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[24] * kernel_shared[((((int)threadIdx.x) * 48) + 32)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[33] * kernel_shared[((((int)threadIdx.x) * 48) + 35)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[25] * kernel_shared[((((int)threadIdx.x) * 48) + 32)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[34] * kernel_shared[((((int)threadIdx.x) * 48) + 35)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[26] * kernel_shared[((((int)threadIdx.x) * 48) + 32)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[35] * kernel_shared[((((int)threadIdx.x) * 48) + 35)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[36] * kernel_shared[((((int)threadIdx.x) * 48) + 12)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[45] * kernel_shared[((((int)threadIdx.x) * 48) + 15)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[37] * kernel_shared[((((int)threadIdx.x) * 48) + 12)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[46] * kernel_shared[((((int)threadIdx.x) * 48) + 15)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[38] * kernel_shared[((((int)threadIdx.x) * 48) + 12)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[47] * kernel_shared[((((int)threadIdx.x) * 48) + 15)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[39] * kernel_shared[((((int)threadIdx.x) * 48) + 12)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[48] * kernel_shared[((((int)threadIdx.x) * 48) + 15)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[40] * kernel_shared[((((int)threadIdx.x) * 48) + 12)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[49] * kernel_shared[((((int)threadIdx.x) * 48) + 15)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[41] * kernel_shared[((((int)threadIdx.x) * 48) + 12)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[50] * kernel_shared[((((int)threadIdx.x) * 48) + 15)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[42] * kernel_shared[((((int)threadIdx.x) * 48) + 12)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[51] * kernel_shared[((((int)threadIdx.x) * 48) + 15)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[36] * kernel_shared[((((int)threadIdx.x) * 48) + 36)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[45] * kernel_shared[((((int)threadIdx.x) * 48) + 39)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[37] * kernel_shared[((((int)threadIdx.x) * 48) + 36)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[46] * kernel_shared[((((int)threadIdx.x) * 48) + 39)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[38] * kernel_shared[((((int)threadIdx.x) * 48) + 36)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[47] * kernel_shared[((((int)threadIdx.x) * 48) + 39)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[39] * kernel_shared[((((int)threadIdx.x) * 48) + 36)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[48] * kernel_shared[((((int)threadIdx.x) * 48) + 39)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[40] * kernel_shared[((((int)threadIdx.x) * 48) + 36)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[49] * kernel_shared[((((int)threadIdx.x) * 48) + 39)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[41] * kernel_shared[((((int)threadIdx.x) * 48) + 36)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[50] * kernel_shared[((((int)threadIdx.x) * 48) + 39)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[42] * kernel_shared[((((int)threadIdx.x) * 48) + 36)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[51] * kernel_shared[((((int)threadIdx.x) * 48) + 39)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[37] * kernel_shared[((((int)threadIdx.x) * 48) + 13)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[46] * kernel_shared[((((int)threadIdx.x) * 48) + 16)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[38] * kernel_shared[((((int)threadIdx.x) * 48) + 13)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[47] * kernel_shared[((((int)threadIdx.x) * 48) + 16)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[39] * kernel_shared[((((int)threadIdx.x) * 48) + 13)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[48] * kernel_shared[((((int)threadIdx.x) * 48) + 16)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[40] * kernel_shared[((((int)threadIdx.x) * 48) + 13)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[49] * kernel_shared[((((int)threadIdx.x) * 48) + 16)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[41] * kernel_shared[((((int)threadIdx.x) * 48) + 13)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[50] * kernel_shared[((((int)threadIdx.x) * 48) + 16)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[42] * kernel_shared[((((int)threadIdx.x) * 48) + 13)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[51] * kernel_shared[((((int)threadIdx.x) * 48) + 16)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[43] * kernel_shared[((((int)threadIdx.x) * 48) + 13)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[52] * kernel_shared[((((int)threadIdx.x) * 48) + 16)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[37] * kernel_shared[((((int)threadIdx.x) * 48) + 37)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[46] * kernel_shared[((((int)threadIdx.x) * 48) + 40)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[38] * kernel_shared[((((int)threadIdx.x) * 48) + 37)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[47] * kernel_shared[((((int)threadIdx.x) * 48) + 40)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[39] * kernel_shared[((((int)threadIdx.x) * 48) + 37)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[48] * kernel_shared[((((int)threadIdx.x) * 48) + 40)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[40] * kernel_shared[((((int)threadIdx.x) * 48) + 37)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[49] * kernel_shared[((((int)threadIdx.x) * 48) + 40)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[41] * kernel_shared[((((int)threadIdx.x) * 48) + 37)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[50] * kernel_shared[((((int)threadIdx.x) * 48) + 40)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[42] * kernel_shared[((((int)threadIdx.x) * 48) + 37)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[51] * kernel_shared[((((int)threadIdx.x) * 48) + 40)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[43] * kernel_shared[((((int)threadIdx.x) * 48) + 37)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[52] * kernel_shared[((((int)threadIdx.x) * 48) + 40)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[38] * kernel_shared[((((int)threadIdx.x) * 48) + 14)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[47] * kernel_shared[((((int)threadIdx.x) * 48) + 17)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[39] * kernel_shared[((((int)threadIdx.x) * 48) + 14)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[48] * kernel_shared[((((int)threadIdx.x) * 48) + 17)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[40] * kernel_shared[((((int)threadIdx.x) * 48) + 14)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[49] * kernel_shared[((((int)threadIdx.x) * 48) + 17)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[41] * kernel_shared[((((int)threadIdx.x) * 48) + 14)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[50] * kernel_shared[((((int)threadIdx.x) * 48) + 17)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[42] * kernel_shared[((((int)threadIdx.x) * 48) + 14)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[51] * kernel_shared[((((int)threadIdx.x) * 48) + 17)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[43] * kernel_shared[((((int)threadIdx.x) * 48) + 14)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[52] * kernel_shared[((((int)threadIdx.x) * 48) + 17)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[44] * kernel_shared[((((int)threadIdx.x) * 48) + 14)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[53] * kernel_shared[((((int)threadIdx.x) * 48) + 17)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[38] * kernel_shared[((((int)threadIdx.x) * 48) + 38)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[47] * kernel_shared[((((int)threadIdx.x) * 48) + 41)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[39] * kernel_shared[((((int)threadIdx.x) * 48) + 38)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[48] * kernel_shared[((((int)threadIdx.x) * 48) + 41)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[40] * kernel_shared[((((int)threadIdx.x) * 48) + 38)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[49] * kernel_shared[((((int)threadIdx.x) * 48) + 41)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[41] * kernel_shared[((((int)threadIdx.x) * 48) + 38)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[50] * kernel_shared[((((int)threadIdx.x) * 48) + 41)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[42] * kernel_shared[((((int)threadIdx.x) * 48) + 38)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[51] * kernel_shared[((((int)threadIdx.x) * 48) + 41)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[43] * kernel_shared[((((int)threadIdx.x) * 48) + 38)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[52] * kernel_shared[((((int)threadIdx.x) * 48) + 41)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[44] * kernel_shared[((((int)threadIdx.x) * 48) + 38)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[53] * kernel_shared[((((int)threadIdx.x) * 48) + 41)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[54] * kernel_shared[((((int)threadIdx.x) * 48) + 18)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[63] * kernel_shared[((((int)threadIdx.x) * 48) + 21)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[55] * kernel_shared[((((int)threadIdx.x) * 48) + 18)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[64] * kernel_shared[((((int)threadIdx.x) * 48) + 21)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[56] * kernel_shared[((((int)threadIdx.x) * 48) + 18)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[65] * kernel_shared[((((int)threadIdx.x) * 48) + 21)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[57] * kernel_shared[((((int)threadIdx.x) * 48) + 18)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[66] * kernel_shared[((((int)threadIdx.x) * 48) + 21)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[58] * kernel_shared[((((int)threadIdx.x) * 48) + 18)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[67] * kernel_shared[((((int)threadIdx.x) * 48) + 21)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[59] * kernel_shared[((((int)threadIdx.x) * 48) + 18)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[68] * kernel_shared[((((int)threadIdx.x) * 48) + 21)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[60] * kernel_shared[((((int)threadIdx.x) * 48) + 18)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[69] * kernel_shared[((((int)threadIdx.x) * 48) + 21)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[54] * kernel_shared[((((int)threadIdx.x) * 48) + 42)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[63] * kernel_shared[((((int)threadIdx.x) * 48) + 45)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[55] * kernel_shared[((((int)threadIdx.x) * 48) + 42)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[64] * kernel_shared[((((int)threadIdx.x) * 48) + 45)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[56] * kernel_shared[((((int)threadIdx.x) * 48) + 42)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[65] * kernel_shared[((((int)threadIdx.x) * 48) + 45)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[57] * kernel_shared[((((int)threadIdx.x) * 48) + 42)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[66] * kernel_shared[((((int)threadIdx.x) * 48) + 45)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[58] * kernel_shared[((((int)threadIdx.x) * 48) + 42)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[67] * kernel_shared[((((int)threadIdx.x) * 48) + 45)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[59] * kernel_shared[((((int)threadIdx.x) * 48) + 42)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[68] * kernel_shared[((((int)threadIdx.x) * 48) + 45)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[60] * kernel_shared[((((int)threadIdx.x) * 48) + 42)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[69] * kernel_shared[((((int)threadIdx.x) * 48) + 45)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[55] * kernel_shared[((((int)threadIdx.x) * 48) + 19)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[64] * kernel_shared[((((int)threadIdx.x) * 48) + 22)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[56] * kernel_shared[((((int)threadIdx.x) * 48) + 19)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[65] * kernel_shared[((((int)threadIdx.x) * 48) + 22)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[57] * kernel_shared[((((int)threadIdx.x) * 48) + 19)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[66] * kernel_shared[((((int)threadIdx.x) * 48) + 22)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[58] * kernel_shared[((((int)threadIdx.x) * 48) + 19)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[67] * kernel_shared[((((int)threadIdx.x) * 48) + 22)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[59] * kernel_shared[((((int)threadIdx.x) * 48) + 19)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[68] * kernel_shared[((((int)threadIdx.x) * 48) + 22)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[60] * kernel_shared[((((int)threadIdx.x) * 48) + 19)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[69] * kernel_shared[((((int)threadIdx.x) * 48) + 22)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[61] * kernel_shared[((((int)threadIdx.x) * 48) + 19)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[70] * kernel_shared[((((int)threadIdx.x) * 48) + 22)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[55] * kernel_shared[((((int)threadIdx.x) * 48) + 43)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[64] * kernel_shared[((((int)threadIdx.x) * 48) + 46)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[56] * kernel_shared[((((int)threadIdx.x) * 48) + 43)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[65] * kernel_shared[((((int)threadIdx.x) * 48) + 46)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[57] * kernel_shared[((((int)threadIdx.x) * 48) + 43)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[66] * kernel_shared[((((int)threadIdx.x) * 48) + 46)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[58] * kernel_shared[((((int)threadIdx.x) * 48) + 43)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[67] * kernel_shared[((((int)threadIdx.x) * 48) + 46)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[59] * kernel_shared[((((int)threadIdx.x) * 48) + 43)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[68] * kernel_shared[((((int)threadIdx.x) * 48) + 46)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[60] * kernel_shared[((((int)threadIdx.x) * 48) + 43)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[69] * kernel_shared[((((int)threadIdx.x) * 48) + 46)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[61] * kernel_shared[((((int)threadIdx.x) * 48) + 43)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[70] * kernel_shared[((((int)threadIdx.x) * 48) + 46)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[56] * kernel_shared[((((int)threadIdx.x) * 48) + 20)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[65] * kernel_shared[((((int)threadIdx.x) * 48) + 23)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[57] * kernel_shared[((((int)threadIdx.x) * 48) + 20)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[66] * kernel_shared[((((int)threadIdx.x) * 48) + 23)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[58] * kernel_shared[((((int)threadIdx.x) * 48) + 20)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[67] * kernel_shared[((((int)threadIdx.x) * 48) + 23)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[59] * kernel_shared[((((int)threadIdx.x) * 48) + 20)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[68] * kernel_shared[((((int)threadIdx.x) * 48) + 23)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[60] * kernel_shared[((((int)threadIdx.x) * 48) + 20)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[69] * kernel_shared[((((int)threadIdx.x) * 48) + 23)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[61] * kernel_shared[((((int)threadIdx.x) * 48) + 20)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[70] * kernel_shared[((((int)threadIdx.x) * 48) + 23)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[62] * kernel_shared[((((int)threadIdx.x) * 48) + 20)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[71] * kernel_shared[((((int)threadIdx.x) * 48) + 23)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[56] * kernel_shared[((((int)threadIdx.x) * 48) + 44)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[65] * kernel_shared[((((int)threadIdx.x) * 48) + 47)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[57] * kernel_shared[((((int)threadIdx.x) * 48) + 44)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[66] * kernel_shared[((((int)threadIdx.x) * 48) + 47)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[58] * kernel_shared[((((int)threadIdx.x) * 48) + 44)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[67] * kernel_shared[((((int)threadIdx.x) * 48) + 47)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[59] * kernel_shared[((((int)threadIdx.x) * 48) + 44)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[68] * kernel_shared[((((int)threadIdx.x) * 48) + 47)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[60] * kernel_shared[((((int)threadIdx.x) * 48) + 44)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[69] * kernel_shared[((((int)threadIdx.x) * 48) + 47)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[61] * kernel_shared[((((int)threadIdx.x) * 48) + 44)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[70] * kernel_shared[((((int)threadIdx.x) * 48) + 47)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[62] * kernel_shared[((((int)threadIdx.x) * 48) + 44)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[71] * kernel_shared[((((int)threadIdx.x) * 48) + 47)]));
}
- __syncthreads();
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7))] * kernel_shared[((((int)threadIdx.x) / 49) * 72)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7))] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 18)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 1)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 1)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 1)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 19)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 2)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 2)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 2)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 20)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 9)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 3)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 9)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 21)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 10)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 4)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 10)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 22)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 11)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 5)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 11)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 23)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 18)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 6)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 18)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 24)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 19)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 7)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 19)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 25)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 20)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 8)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 20)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 26)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7))] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 36)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7))] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 54)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 1)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 37)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 1)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 55)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 2)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 38)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 2)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 56)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 9)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 39)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 9)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 57)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 10)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 40)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 10)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 58)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 11)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 41)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 11)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 59)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 18)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 42)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 18)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 60)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 19)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 43)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 19)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 61)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 20)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 44)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 20)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 62)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 81)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 9)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 81)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 27)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 82)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 10)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 82)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 28)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 83)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 11)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 83)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 29)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 90)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 12)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 90)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 30)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 91)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 13)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 91)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 31)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 92)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 14)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 92)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 32)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 99)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 15)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 99)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 33)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 100)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 16)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 100)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 34)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 101)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 17)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 101)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 35)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 81)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 45)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 81)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 63)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 82)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 46)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 82)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 64)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 83)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 47)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 83)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 65)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 90)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 48)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 90)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 66)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 91)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 49)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 91)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 67)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 92)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 50)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 92)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 68)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 99)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 51)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 99)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 69)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 100)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 52)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 100)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 70)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 101)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 53)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 101)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 71)]));
}
- for (int i1_inner = 0; i1_inner < 4; ++i1_inner) {
- compute[((((((int)blockIdx.x) * 392) + ((((int)threadIdx.x) / 49) * 196)) + (i1_inner * 49)) + (((int)threadIdx.x) % 49))] = max((conv2d_nchw[i1_inner] + bias[(((((int)blockIdx.x) * 8) + ((((int)threadIdx.x) / 49) * 4)) + i1_inner)]), 0.000000e+00f);
+ for (int i1_inner = 0; i1_inner < 2; ++i1_inner) {
+ for (int i3_inner = 0; i3_inner < 7; ++i3_inner) {
+ compute[((((((((int)blockIdx.x) / 7) * 6272) + (((int)threadIdx.x) * 98)) + (i1_inner * 49)) + ((((int)blockIdx.x) % 7) * 7)) + i3_inner)] = max((conv2d_nchw[((i1_inner * 7) + i3_inner)] + bias[((((((int)blockIdx.x) / 7) * 128) + (((int)threadIdx.x) * 2)) + i1_inner)]), 0.000000e+00f);
+ }
}
}
@@ -668,7 +1377,7 @@ In the example below we resume the status and do more 5 trials.
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 5 minutes 37.751 seconds)
+ **Total running time of the script:** ( 6 minutes 7.710 seconds)
.. _sphx_glr_download_how_to_tune_with_autoscheduler_tune_conv2d_layer_cuda.py:
diff --git a/docs/_sources/how_to/tune_with_autoscheduler/tune_network_cuda.rst.txt b/docs/_sources/how_to/tune_with_autoscheduler/tune_network_cuda.rst.txt
index 5e02c95ab8..ef7db935ab 100644
--- a/docs/_sources/how_to/tune_with_autoscheduler/tune_network_cuda.rst.txt
+++ b/docs/_sources/how_to/tune_with_autoscheduler/tune_network_cuda.rst.txt
@@ -643,7 +643,7 @@ so we can read the log file and load the best schedules.
Evaluate inference time cost...
Execution time summary:
mean (ms) median (ms) max (ms) min (ms) std (ms)
- 8.1788 8.1757 8.1876 8.1730 0.0064
+ 8.1987 8.1995 8.2098 8.1868 0.0094
@@ -671,7 +671,7 @@ Other Tips
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 1 minutes 4.612 seconds)
+ **Total running time of the script:** ( 1 minutes 4.692 seconds)
.. _sphx_glr_download_how_to_tune_with_autoscheduler_tune_network_cuda.py:
diff --git a/docs/_sources/how_to/tune_with_autoscheduler/tune_network_x86.rst.txt b/docs/_sources/how_to/tune_with_autoscheduler/tune_network_x86.rst.txt
index defbaad2a6..638bd36263 100644
--- a/docs/_sources/how_to/tune_with_autoscheduler/tune_network_x86.rst.txt
+++ b/docs/_sources/how_to/tune_with_autoscheduler/tune_network_x86.rst.txt
@@ -662,7 +662,7 @@ so we can read the log file and load the best schedules.
Evaluate inference time cost...
Execution time summary:
mean (ms) median (ms) max (ms) min (ms) std (ms)
- 777.0653 776.8267 779.3890 774.9802 1.8078
+ 752.0948 752.5188 753.8207 749.9449 1.6104
@@ -690,7 +690,7 @@ Other Tips
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 1 minutes 34.377 seconds)
+ **Total running time of the script:** ( 1 minutes 34.567 seconds)
.. _sphx_glr_download_how_to_tune_with_autoscheduler_tune_network_x86.py:
diff --git a/docs/_sources/how_to/tune_with_autoscheduler/tune_sparse_x86.rst.txt b/docs/_sources/how_to/tune_with_autoscheduler/tune_sparse_x86.rst.txt
index 95a75fe5f6..3b4c7fa08c 100644
--- a/docs/_sources/how_to/tune_with_autoscheduler/tune_sparse_x86.rst.txt
+++ b/docs/_sources/how_to/tune_with_autoscheduler/tune_sparse_x86.rst.txt
@@ -379,416 +379,35 @@ layout transformation, parallelization, vectorization, unrolling, and operator f
Lowered TIR:
@main = primfn(placeholder_5: handle, placeholder_6: handle, placeholder_7: handle, placeholder_8: handle, placeholder_9: handle, compute_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {placeholder: Buffer(placeholder_10: Pointer(float32), float32, [32768], []),
- placeholder_1: Buffer(placeholder_11: Pointer(float32), float32, [78656], []),
+ buffers = {placeholder: Buffer(placeholder_10: Pointer(float32), float32, [128, 256], []),
+ placeholder_1: Buffer(placeholder_11: Pointer(float32), float32, [4916, 16, 1], []),
placeholder_2: Buffer(placeholder_12: Pointer(int32), int32, [4916], []),
placeholder_3: Buffer(placeholder_13: Pointer(int32), int32, [33], []),
- placeholder_4: Buffer(placeholder_14: Pointer(float32), float32, [65536], []),
- compute: Buffer(compute_2: Pointer(float32), float32, [65536], [])}
- buffer_map = {placeholder_5: placeholder, placeholder_6: placeholder_1, placeholder_7: placeholder_2, placeholder_8: placeholder_3, placeholder_9: placeholder_4, compute_1: compute}
- preflattened_buffer_map = {placeholder_6: placeholder_15: Buffer(placeholder_11, float32, [4916, 16, 1], []), placeholder_8: placeholder_16: Buffer(placeholder_13, int32, [33], []), compute_1: compute_3: Buffer(compute_2, float32, [128, 512], []), placeholder_7: placeholder_17: Buffer(placeholder_12, int32, [4916], []), placeholder_5: placeholder_18: Buffer(placeholder_10, float32, [128, 256], []), placeholder_9: placeholder_19: Buffer(placeholder_14, float32, [128, 512], [])} {
- for (i0.outer.i1.outer.fused: int32, 0, 64) "parallel" {
- allocate(compute_4: Pointer(global float32), float32, [1024]), storage_scope = global {
- for (i.outer.inner: int32, 0, 4) {
- for (nb_j.inner: int32, 0, 2) {
- let cse_var_2: int32 = ((i.outer.inner*256) + (nb_j.inner*16))
- let cse_var_1: int32 = ((floormod(i0.outer.i1.outer.fused, 16)*2) + nb_j.inner)
- {
- compute_5: Buffer(compute_4, float32, [1024], [])[cse_var_2] = 0f32
- compute_5[(cse_var_2 + 1)] = 0f32
- compute_5[(cse_var_2 + 2)] = 0f32
- compute_5[(cse_var_2 + 3)] = 0f32
- compute_5[(cse_var_2 + 4)] = 0f32
- compute_5[(cse_var_2 + 5)] = 0f32
- compute_5[(cse_var_2 + 6)] = 0f32
- compute_5[(cse_var_2 + 7)] = 0f32
- compute_5[(cse_var_2 + 8)] = 0f32
- compute_5[(cse_var_2 + 9)] = 0f32
- compute_5[(cse_var_2 + 10)] = 0f32
- compute_5[(cse_var_2 + 11)] = 0f32
- compute_5[(cse_var_2 + 12)] = 0f32
- compute_5[(cse_var_2 + 13)] = 0f32
- compute_5[(cse_var_2 + 14)] = 0f32
- compute_5[(cse_var_2 + 15)] = 0f32
- compute_5[(cse_var_2 + 32)] = 0f32
- compute_5[(cse_var_2 + 33)] = 0f32
- compute_5[(cse_var_2 + 34)] = 0f32
- compute_5[(cse_var_2 + 35)] = 0f32
- compute_5[(cse_var_2 + 36)] = 0f32
- compute_5[(cse_var_2 + 37)] = 0f32
- compute_5[(cse_var_2 + 38)] = 0f32
- compute_5[(cse_var_2 + 39)] = 0f32
- compute_5[(cse_var_2 + 40)] = 0f32
- compute_5[(cse_var_2 + 41)] = 0f32
- compute_5[(cse_var_2 + 42)] = 0f32
- compute_5[(cse_var_2 + 43)] = 0f32
- compute_5[(cse_var_2 + 44)] = 0f32
- compute_5[(cse_var_2 + 45)] = 0f32
- compute_5[(cse_var_2 + 46)] = 0f32
- compute_5[(cse_var_2 + 47)] = 0f32
- compute_5[(cse_var_2 + 64)] = 0f32
- compute_5[(cse_var_2 + 65)] = 0f32
- compute_5[(cse_var_2 + 66)] = 0f32
- compute_5[(cse_var_2 + 67)] = 0f32
- compute_5[(cse_var_2 + 68)] = 0f32
- compute_5[(cse_var_2 + 69)] = 0f32
- compute_5[(cse_var_2 + 70)] = 0f32
- compute_5[(cse_var_2 + 71)] = 0f32
- compute_5[(cse_var_2 + 72)] = 0f32
- compute_5[(cse_var_2 + 73)] = 0f32
- compute_5[(cse_var_2 + 74)] = 0f32
- compute_5[(cse_var_2 + 75)] = 0f32
- compute_5[(cse_var_2 + 76)] = 0f32
- compute_5[(cse_var_2 + 77)] = 0f32
- compute_5[(cse_var_2 + 78)] = 0f32
- compute_5[(cse_var_2 + 79)] = 0f32
- compute_5[(cse_var_2 + 96)] = 0f32
- compute_5[(cse_var_2 + 97)] = 0f32
- compute_5[(cse_var_2 + 98)] = 0f32
- compute_5[(cse_var_2 + 99)] = 0f32
- compute_5[(cse_var_2 + 100)] = 0f32
- compute_5[(cse_var_2 + 101)] = 0f32
- compute_5[(cse_var_2 + 102)] = 0f32
- compute_5[(cse_var_2 + 103)] = 0f32
- compute_5[(cse_var_2 + 104)] = 0f32
- compute_5[(cse_var_2 + 105)] = 0f32
- compute_5[(cse_var_2 + 106)] = 0f32
- compute_5[(cse_var_2 + 107)] = 0f32
- compute_5[(cse_var_2 + 108)] = 0f32
- compute_5[(cse_var_2 + 109)] = 0f32
- compute_5[(cse_var_2 + 110)] = 0f32
- compute_5[(cse_var_2 + 111)] = 0f32
- compute_5[(cse_var_2 + 128)] = 0f32
- compute_5[(cse_var_2 + 129)] = 0f32
- compute_5[(cse_var_2 + 130)] = 0f32
- compute_5[(cse_var_2 + 131)] = 0f32
- compute_5[(cse_var_2 + 132)] = 0f32
- compute_5[(cse_var_2 + 133)] = 0f32
- compute_5[(cse_var_2 + 134)] = 0f32
- compute_5[(cse_var_2 + 135)] = 0f32
- compute_5[(cse_var_2 + 136)] = 0f32
- compute_5[(cse_var_2 + 137)] = 0f32
- compute_5[(cse_var_2 + 138)] = 0f32
- compute_5[(cse_var_2 + 139)] = 0f32
- compute_5[(cse_var_2 + 140)] = 0f32
- compute_5[(cse_var_2 + 141)] = 0f32
- compute_5[(cse_var_2 + 142)] = 0f32
- compute_5[(cse_var_2 + 143)] = 0f32
- compute_5[(cse_var_2 + 160)] = 0f32
- compute_5[(cse_var_2 + 161)] = 0f32
- compute_5[(cse_var_2 + 162)] = 0f32
- compute_5[(cse_var_2 + 163)] = 0f32
- compute_5[(cse_var_2 + 164)] = 0f32
- compute_5[(cse_var_2 + 165)] = 0f32
- compute_5[(cse_var_2 + 166)] = 0f32
- compute_5[(cse_var_2 + 167)] = 0f32
- compute_5[(cse_var_2 + 168)] = 0f32
- compute_5[(cse_var_2 + 169)] = 0f32
- compute_5[(cse_var_2 + 170)] = 0f32
- compute_5[(cse_var_2 + 171)] = 0f32
- compute_5[(cse_var_2 + 172)] = 0f32
- compute_5[(cse_var_2 + 173)] = 0f32
- compute_5[(cse_var_2 + 174)] = 0f32
- compute_5[(cse_var_2 + 175)] = 0f32
- compute_5[(cse_var_2 + 192)] = 0f32
- compute_5[(cse_var_2 + 193)] = 0f32
- compute_5[(cse_var_2 + 194)] = 0f32
- compute_5[(cse_var_2 + 195)] = 0f32
- compute_5[(cse_var_2 + 196)] = 0f32
- compute_5[(cse_var_2 + 197)] = 0f32
- compute_5[(cse_var_2 + 198)] = 0f32
- compute_5[(cse_var_2 + 199)] = 0f32
- compute_5[(cse_var_2 + 200)] = 0f32
- compute_5[(cse_var_2 + 201)] = 0f32
- compute_5[(cse_var_2 + 202)] = 0f32
- compute_5[(cse_var_2 + 203)] = 0f32
- compute_5[(cse_var_2 + 204)] = 0f32
- compute_5[(cse_var_2 + 205)] = 0f32
- compute_5[(cse_var_2 + 206)] = 0f32
- compute_5[(cse_var_2 + 207)] = 0f32
- compute_5[(cse_var_2 + 224)] = 0f32
- compute_5[(cse_var_2 + 225)] = 0f32
- compute_5[(cse_var_2 + 226)] = 0f32
- compute_5[(cse_var_2 + 227)] = 0f32
- compute_5[(cse_var_2 + 228)] = 0f32
- compute_5[(cse_var_2 + 229)] = 0f32
- compute_5[(cse_var_2 + 230)] = 0f32
- compute_5[(cse_var_2 + 231)] = 0f32
- compute_5[(cse_var_2 + 232)] = 0f32
- compute_5[(cse_var_2 + 233)] = 0f32
- compute_5[(cse_var_2 + 234)] = 0f32
- compute_5[(cse_var_2 + 235)] = 0f32
- compute_5[(cse_var_2 + 236)] = 0f32
- compute_5[(cse_var_2 + 237)] = 0f32
- compute_5[(cse_var_2 + 238)] = 0f32
- compute_5[(cse_var_2 + 239)] = 0f32
- for (elem_idx: int32, 0, (placeholder_3[(cse_var_1 + 1)] - placeholder_3[cse_var_1])) {
- let cse_var_131: int32 = (elem_idx*16)
- let cse_var_130: int32 = (cse_var_2 + 99)
- let cse_var_129: int32 = (cse_var_2 + 98)
- let cse_var_128: int32 = (cse_var_2 + 97)
- let cse_var_127: int32 = (cse_var_2 + 96)
- let cse_var_126: int32 = (cse_var_2 + 9)
- let cse_var_125: int32 = (cse_var_2 + 8)
- let cse_var_124: int32 = (cse_var_2 + 79)
- let cse_var_123: int32 = (cse_var_2 + 78)
- let cse_var_122: int32 = (cse_var_2 + 77)
- let cse_var_121: int32 = (cse_var_2 + 76)
- let cse_var_120: int32 = (cse_var_2 + 75)
- let cse_var_119: int32 = (cse_var_2 + 74)
- let cse_var_118: int32 = (cse_var_2 + 73)
- let cse_var_117: int32 = (cse_var_2 + 72)
- let cse_var_116: int32 = (cse_var_2 + 71)
- let cse_var_115: int32 = (cse_var_2 + 70)
- let cse_var_114: int32 = (cse_var_2 + 7)
- let cse_var_113: int32 = (cse_var_2 + 69)
- let cse_var_112: int32 = (cse_var_2 + 68)
- let cse_var_111: int32 = (cse_var_2 + 67)
- let cse_var_110: int32 = (cse_var_2 + 66)
- let cse_var_109: int32 = (cse_var_2 + 65)
- let cse_var_108: int32 = (cse_var_2 + 64)
- let cse_var_107: int32 = (cse_var_2 + 6)
- let cse_var_106: int32 = (cse_var_2 + 5)
- let cse_var_105: int32 = (cse_var_2 + 47)
- let cse_var_104: int32 = (cse_var_2 + 46)
- let cse_var_103: int32 = (cse_var_2 + 45)
- let cse_var_102: int32 = (cse_var_2 + 44)
- let cse_var_101: int32 = (cse_var_2 + 43)
- let cse_var_100: int32 = (cse_var_2 + 42)
- let cse_var_99: int32 = (cse_var_2 + 41)
- let cse_var_98: int32 = (cse_var_2 + 40)
- let cse_var_97: int32 = (cse_var_2 + 4)
- let cse_var_96: int32 = (cse_var_2 + 39)
- let cse_var_95: int32 = (cse_var_2 + 38)
- let cse_var_94: int32 = (cse_var_2 + 37)
- let cse_var_93: int32 = (cse_var_2 + 36)
- let cse_var_92: int32 = (cse_var_2 + 35)
- let cse_var_91: int32 = (cse_var_2 + 34)
- let cse_var_90: int32 = (cse_var_2 + 33)
- let cse_var_89: int32 = (cse_var_2 + 32)
- let cse_var_88: int32 = (cse_var_2 + 3)
- let cse_var_87: int32 = (cse_var_2 + 239)
- let cse_var_86: int32 = (cse_var_2 + 238)
- let cse_var_85: int32 = (cse_var_2 + 237)
- let cse_var_84: int32 = (cse_var_2 + 236)
- let cse_var_83: int32 = (cse_var_2 + 235)
- let cse_var_82: int32 = (cse_var_2 + 234)
- let cse_var_81: int32 = (cse_var_2 + 233)
- let cse_var_80: int32 = (cse_var_2 + 232)
- let cse_var_79: int32 = (cse_var_2 + 231)
- let cse_var_78: int32 = (cse_var_2 + 230)
- let cse_var_77: int32 = (cse_var_2 + 229)
- let cse_var_76: int32 = (cse_var_2 + 228)
- let cse_var_75: int32 = (cse_var_2 + 227)
- let cse_var_74: int32 = (cse_var_2 + 226)
- let cse_var_73: int32 = (cse_var_2 + 225)
- let cse_var_72: int32 = (cse_var_2 + 224)
- let cse_var_71: int32 = (cse_var_2 + 207)
- let cse_var_70: int32 = (cse_var_2 + 206)
- let cse_var_69: int32 = (cse_var_2 + 205)
- let cse_var_68: int32 = (cse_var_2 + 204)
- let cse_var_67: int32 = (cse_var_2 + 203)
- let cse_var_66: int32 = (cse_var_2 + 202)
- let cse_var_65: int32 = (cse_var_2 + 201)
- let cse_var_64: int32 = (cse_var_2 + 200)
- let cse_var_63: int32 = (cse_var_2 + 2)
- let cse_var_62: int32 = (cse_var_2 + 199)
- let cse_var_61: int32 = (cse_var_2 + 198)
- let cse_var_60: int32 = (cse_var_2 + 197)
- let cse_var_59: int32 = (cse_var_2 + 196)
- let cse_var_58: int32 = (cse_var_2 + 195)
- let cse_var_57: int32 = (cse_var_2 + 194)
- let cse_var_56: int32 = (cse_var_2 + 193)
- let cse_var_55: int32 = (cse_var_2 + 192)
- let cse_var_54: int32 = (cse_var_2 + 175)
- let cse_var_53: int32 = (cse_var_2 + 174)
- let cse_var_52: int32 = (cse_var_2 + 173)
- let cse_var_51: int32 = (cse_var_2 + 172)
- let cse_var_50: int32 = (cse_var_2 + 171)
- let cse_var_49: int32 = (cse_var_2 + 170)
- let cse_var_48: int32 = (cse_var_2 + 169)
- let cse_var_47: int32 = (cse_var_2 + 168)
- let cse_var_46: int32 = (cse_var_2 + 167)
- let cse_var_45: int32 = (cse_var_2 + 166)
- let cse_var_44: int32 = (cse_var_2 + 165)
- let cse_var_43: int32 = (cse_var_2 + 164)
- let cse_var_42: int32 = (cse_var_2 + 163)
- let cse_var_41: int32 = (cse_var_2 + 162)
- let cse_var_40: int32 = (cse_var_2 + 161)
- let cse_var_39: int32 = (cse_var_2 + 160)
- let cse_var_38: int32 = (cse_var_2 + 15)
- let cse_var_37: int32 = (cse_var_2 + 143)
- let cse_var_36: int32 = (cse_var_2 + 142)
- let cse_var_35: int32 = (cse_var_2 + 141)
- let cse_var_34: int32 = (cse_var_2 + 140)
- let cse_var_33: int32 = (cse_var_2 + 14)
- let cse_var_32: int32 = (cse_var_2 + 139)
- let cse_var_31: int32 = (cse_var_2 + 138)
- let cse_var_30: int32 = (cse_var_2 + 137)
- let cse_var_29: int32 = (cse_var_2 + 136)
- let cse_var_28: int32 = (cse_var_2 + 135)
- let cse_var_27: int32 = (cse_var_2 + 134)
- let cse_var_26: int32 = (cse_var_2 + 133)
- let cse_var_25: int32 = (cse_var_2 + 132)
- let cse_var_24: int32 = (cse_var_2 + 131)
- let cse_var_23: int32 = (cse_var_2 + 130)
- let cse_var_22: int32 = (cse_var_2 + 13)
- let cse_var_21: int32 = (cse_var_2 + 129)
- let cse_var_20: int32 = (cse_var_2 + 128)
- let cse_var_19: int32 = (cse_var_2 + 12)
- let cse_var_18: int32 = (cse_var_2 + 111)
- let cse_var_17: int32 = (cse_var_2 + 110)
- let cse_var_16: int32 = (cse_var_2 + 11)
- let cse_var_15: int32 = (cse_var_2 + 109)
- let cse_var_14: int32 = (cse_var_2 + 108)
- let cse_var_13: int32 = (cse_var_2 + 107)
- let cse_var_12: int32 = (cse_var_2 + 106)
- let cse_var_11: int32 = (cse_var_2 + 105)
- let cse_var_10: int32 = (cse_var_2 + 104)
- let cse_var_9: int32 = (cse_var_2 + 103)
- let cse_var_8: int32 = (cse_var_2 + 102)
- let cse_var_7: int32 = (cse_var_2 + 101)
- let cse_var_6: int32 = (cse_var_2 + 100)
- let cse_var_5: int32 = (cse_var_2 + 10)
- let cse_var_4: int32 = (cse_var_2 + 1)
- let cse_var_3: int32 = ((floordiv(i0.outer.i1.outer.fused, 16)*8192) + (i.outer.inner*2048))
- {
- compute_5[cse_var_2] = (compute_5[cse_var_2] + (placeholder_1[((placeholder_3[cse_var_1]*16) + cse_var_131)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_4] = (compute_5[cse_var_4] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 1)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_63] = (compute_5[cse_var_63] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 2)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_88] = (compute_5[cse_var_88] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 3)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_97] = (compute_5[cse_var_97] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 4)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_106] = (compute_5[cse_var_106] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 5)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_107] = (compute_5[cse_var_107] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 6)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_114] = (compute_5[cse_var_114] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 7)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_125] = (compute_5[cse_var_125] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 8)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_126] = (compute_5[cse_var_126] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 9)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_5] = (compute_5[cse_var_5] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 10)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_16] = (compute_5[cse_var_16] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 11)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_19] = (compute_5[cse_var_19] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 12)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_22] = (compute_5[cse_var_22] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 13)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_33] = (compute_5[cse_var_33] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 14)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_38] = (compute_5[cse_var_38] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 15)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_89] = (compute_5[cse_var_89] + (placeholder_1[((placeholder_3[cse_var_1]*16) + cse_var_131)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_90] = (compute_5[cse_var_90] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 1)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_91] = (compute_5[cse_var_91] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 2)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_92] = (compute_5[cse_var_92] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 3)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_93] = (compute_5[cse_var_93] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 4)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_94] = (compute_5[cse_var_94] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 5)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_95] = (compute_5[cse_var_95] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 6)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_96] = (compute_5[cse_var_96] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 7)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_98] = (compute_5[cse_var_98] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 8)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_99] = (compute_5[cse_var_99] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 9)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_100] = (compute_5[cse_var_100] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 10)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_101] = (compute_5[cse_var_101] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 11)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_102] = (compute_5[cse_var_102] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 12)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_103] = (compute_5[cse_var_103] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 13)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_104] = (compute_5[cse_var_104] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 14)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_105] = (compute_5[cse_var_105] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 15)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_108] = (compute_5[cse_var_108] + (placeholder_1[((placeholder_3[cse_var_1]*16) + cse_var_131)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_109] = (compute_5[cse_var_109] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 1)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_110] = (compute_5[cse_var_110] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 2)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_111] = (compute_5[cse_var_111] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 3)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_112] = (compute_5[cse_var_112] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 4)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_113] = (compute_5[cse_var_113] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 5)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_115] = (compute_5[cse_var_115] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 6)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_116] = (compute_5[cse_var_116] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 7)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_117] = (compute_5[cse_var_117] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 8)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_118] = (compute_5[cse_var_118] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 9)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_119] = (compute_5[cse_var_119] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 10)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_120] = (compute_5[cse_var_120] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 11)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_121] = (compute_5[cse_var_121] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 12)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_122] = (compute_5[cse_var_122] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 13)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_123] = (compute_5[cse_var_123] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 14)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_124] = (compute_5[cse_var_124] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 15)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_127] = (compute_5[cse_var_127] + (placeholder_1[((placeholder_3[cse_var_1]*16) + cse_var_131)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_128] = (compute_5[cse_var_128] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 1)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_129] = (compute_5[cse_var_129] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 2)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_130] = (compute_5[cse_var_130] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 3)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_6] = (compute_5[cse_var_6] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 4)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_7] = (compute_5[cse_var_7] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 5)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_8] = (compute_5[cse_var_8] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 6)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_9] = (compute_5[cse_var_9] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 7)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_10] = (compute_5[cse_var_10] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 8)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_11] = (compute_5[cse_var_11] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 9)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_12] = (compute_5[cse_var_12] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 10)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_13] = (compute_5[cse_var_13] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 11)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_14] = (compute_5[cse_var_14] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 12)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_15] = (compute_5[cse_var_15] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 13)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_17] = (compute_5[cse_var_17] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 14)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_18] = (compute_5[cse_var_18] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 15)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_20] = (compute_5[cse_var_20] + (placeholder_1[((placeholder_3[cse_var_1]*16) + cse_var_131)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_21] = (compute_5[cse_var_21] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 1)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_23] = (compute_5[cse_var_23] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 2)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_24] = (compute_5[cse_var_24] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 3)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_25] = (compute_5[cse_var_25] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 4)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_26] = (compute_5[cse_var_26] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 5)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_27] = (compute_5[cse_var_27] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 6)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_28] = (compute_5[cse_var_28] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 7)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_29] = (compute_5[cse_var_29] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 8)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_30] = (compute_5[cse_var_30] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 9)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_31] = (compute_5[cse_var_31] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 10)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_32] = (compute_5[cse_var_32] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 11)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_34] = (compute_5[cse_var_34] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 12)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_35] = (compute_5[cse_var_35] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 13)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_36] = (compute_5[cse_var_36] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 14)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_37] = (compute_5[cse_var_37] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 15)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_39] = (compute_5[cse_var_39] + (placeholder_1[((placeholder_3[cse_var_1]*16) + cse_var_131)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_40] = (compute_5[cse_var_40] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 1)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_41] = (compute_5[cse_var_41] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 2)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_42] = (compute_5[cse_var_42] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 3)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_43] = (compute_5[cse_var_43] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 4)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_44] = (compute_5[cse_var_44] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 5)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_45] = (compute_5[cse_var_45] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 6)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_46] = (compute_5[cse_var_46] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 7)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_47] = (compute_5[cse_var_47] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 8)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_48] = (compute_5[cse_var_48] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 9)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_49] = (compute_5[cse_var_49] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 10)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_50] = (compute_5[cse_var_50] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 11)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_51] = (compute_5[cse_var_51] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 12)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_52] = (compute_5[cse_var_52] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 13)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_53] = (compute_5[cse_var_53] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 14)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_54] = (compute_5[cse_var_54] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 15)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_55] = (compute_5[cse_var_55] + (placeholder_1[((placeholder_3[cse_var_1]*16) + cse_var_131)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_56] = (compute_5[cse_var_56] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 1)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_57] = (compute_5[cse_var_57] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 2)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_58] = (compute_5[cse_var_58] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 3)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_59] = (compute_5[cse_var_59] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 4)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_60] = (compute_5[cse_var_60] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 5)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_61] = (compute_5[cse_var_61] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 6)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_62] = (compute_5[cse_var_62] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 7)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_64] = (compute_5[cse_var_64] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 8)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_65] = (compute_5[cse_var_65] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 9)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_66] = (compute_5[cse_var_66] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 10)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_67] = (compute_5[cse_var_67] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 11)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_68] = (compute_5[cse_var_68] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 12)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_69] = (compute_5[cse_var_69] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 13)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_70] = (compute_5[cse_var_70] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 14)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_71] = (compute_5[cse_var_71] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 15)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_72] = (compute_5[cse_var_72] + (placeholder_1[((placeholder_3[cse_var_1]*16) + cse_var_131)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_73] = (compute_5[cse_var_73] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 1)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_74] = (compute_5[cse_var_74] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 2)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_75] = (compute_5[cse_var_75] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 3)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_76] = (compute_5[cse_var_76] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 4)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_77] = (compute_5[cse_var_77] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 5)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_78] = (compute_5[cse_var_78] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 6)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_79] = (compute_5[cse_var_79] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 7)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_80] = (compute_5[cse_var_80] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 8)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_81] = (compute_5[cse_var_81] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 9)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_82] = (compute_5[cse_var_82] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 10)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_83] = (compute_5[cse_var_83] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 11)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_84] = (compute_5[cse_var_84] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 12)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_85] = (compute_5[cse_var_85] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 13)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_86] = (compute_5[cse_var_86] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 14)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_87] = (compute_5[cse_var_87] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 15)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
+ placeholder_4: Buffer(placeholder_14: Pointer(float32), float32, [128, 512], []),
+ compute: Buffer(compute_2: Pointer(float32), float32, [128, 512], [])}
+ buffer_map = {placeholder_5: placeholder, placeholder_6: placeholder_1, placeholder_7: placeholder_2, placeholder_8: placeholder_3, placeholder_9: placeholder_4, compute_1: compute} {
+ for (i0.outer.i1.outer.fused: int32, 0, 32) "parallel" {
+ allocate(compute_3: Pointer(global float32), float32, [2048]), storage_scope = global {
+ for (i.outer.inner: int32, 0, 2) {
+ for (i.inner.init: int32, 0, 64) {
+ for (j.init: int32, 0, 16) {
+ compute_4: Buffer(compute_3, float32, [2048], [])[(((i.outer.inner*1024) + (i.inner.init*16)) + j.init)] = 0f32
+ }
+ }
+ for (elem_idx: int32, 0, (placeholder_15: Buffer(placeholder_13, int32, [33], [])[(i0.outer.i1.outer.fused + 1)] - placeholder_15[i0.outer.i1.outer.fused])) {
+ for (i.inner: int32, 0, 64) {
+ for (j: int32, 0, 16) {
+ if @tir.likely((elem_idx < (placeholder_15[(i0.outer.i1.outer.fused + 1)] - placeholder_15[i0.outer.i1.outer.fused])), dtype=bool) {
+ let cse_var_1: int32 = (((i.outer.inner*1024) + (i.inner*16)) + j)
+ compute_4[cse_var_1] = (compute_4[cse_var_1] + (placeholder_16: Buffer(placeholder_11, float32, [78656], [])[(((placeholder_15[i0.outer.i1.outer.fused]*16) + (elem_idx*16)) + j)]*max(placeholder_17: Buffer(placeholder_10, float32, [32768], [])[(((i.outer.inner*16384) + (i.inner*256)) + placeholder_18: Buffer(placeholder_12, int32, [4916], [])[(placeholder_15[i0.outer.i1.outer.fused] + elem_idx)])], 0f32)))
}
}
}
}
}
- for (i0.inner: int32, 0, 32) {
- let cse_var_132: int32 = (((floordiv(i0.outer.i1.outer.fused, 16)*16384) + (i0.inner*512)) + (floormod(i0.outer.i1.outer.fused, 16)*32))
- compute[ramp(cse_var_132, 1, 32)] = max((compute_5[ramp((i0.inner*32), 1, 32)] + placeholder_4[ramp(cse_var_132, 1, 32)]), broadcast(0f32, 32))
+ for (i0.inner: int32, 0, 128) {
+ let cse_var_2: int32 = ((i0.inner*512) + (i0.outer.i1.outer.fused*16))
+ compute_5: Buffer(compute_2, float32, [65536], [])[ramp(cse_var_2, 1, 16)] = max((compute_4[ramp((i0.inner*16), 1, 16)] + placeholder_19: Buffer(placeholder_14, float32, [65536], [])[ramp(cse_var_2, 1, 16)]), broadcast(0f32, 16))
}
}
}
@@ -844,7 +463,7 @@ We build the binary and check its correctness and performance.
.. code-block:: none
- Execution time of this operator: 2.871 ms
+ Execution time of this operator: 1.807 ms
diff --git a/docs/_sources/how_to/tune_with_autotvm/sg_execution_times.rst.txt b/docs/_sources/how_to/tune_with_autotvm/sg_execution_times.rst.txt
index 416d62da8c..220922d074 100644
--- a/docs/_sources/how_to/tune_with_autotvm/sg_execution_times.rst.txt
+++ b/docs/_sources/how_to/tune_with_autotvm/sg_execution_times.rst.txt
@@ -5,16 +5,16 @@
Computation times
=================
-**00:23.851** total execution time for **how_to_tune_with_autotvm** files:
+**00:33.070** total execution time for **how_to_tune_with_autotvm** files:
+--------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_tune_with_autotvm_tune_conv2d_cuda.py` (``tune_conv2d_cuda.py``) | 00:23.816 | 0.0 MB |
+| :ref:`sphx_glr_how_to_tune_with_autotvm_tune_conv2d_cuda.py` (``tune_conv2d_cuda.py``) | 00:33.035 | 0.0 MB |
+--------------------------------------------------------------------------------------------------+-----------+--------+
| :ref:`sphx_glr_how_to_tune_with_autotvm_tune_relay_x86.py` (``tune_relay_x86.py``) | 00:00.020 | 0.0 MB |
+--------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_tune_with_autotvm_tune_relay_mobile_gpu.py` (``tune_relay_mobile_gpu.py``) | 00:00.005 | 0.0 MB |
-+--------------------------------------------------------------------------------------------------+-----------+--------+
| :ref:`sphx_glr_how_to_tune_with_autotvm_tune_relay_cuda.py` (``tune_relay_cuda.py``) | 00:00.005 | 0.0 MB |
+--------------------------------------------------------------------------------------------------+-----------+--------+
| :ref:`sphx_glr_how_to_tune_with_autotvm_tune_relay_arm.py` (``tune_relay_arm.py``) | 00:00.005 | 0.0 MB |
+--------------------------------------------------------------------------------------------------+-----------+--------+
+| :ref:`sphx_glr_how_to_tune_with_autotvm_tune_relay_mobile_gpu.py` (``tune_relay_mobile_gpu.py``) | 00:00.005 | 0.0 MB |
++--------------------------------------------------------------------------------------------------+-----------+--------+
diff --git a/docs/_sources/how_to/tune_with_autotvm/tune_conv2d_cuda.rst.txt b/docs/_sources/how_to/tune_with_autotvm/tune_conv2d_cuda.rst.txt
index ec33650992..71d5a1af55 100644
--- a/docs/_sources/how_to/tune_with_autotvm/tune_conv2d_cuda.rst.txt
+++ b/docs/_sources/how_to/tune_with_autotvm/tune_conv2d_cuda.rst.txt
@@ -265,9 +265,9 @@ for this template
waiting for device...
device available
Get devices for measurement successfully!
- No: 1 GFLOPS: 24.37/24.37 result: MeasureResult(costs=(0.0094983644375,), error_no=MeasureErrorNo.NO_ERROR, all_cost=3.947620153427124, timestamp=1668585629.6380484) [('tile_f', [-1, 16, 2, 1]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 64, 1]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,3699754
- No: 2 GFLOPS: 16.16/24.37 result: MeasureResult(costs=(0.014324914,), error_no=MeasureErrorNo.NO_ERROR, all_cost=3.4801909923553467, timestamp=1668585630.3499599) [('tile_f', [-1, 16, 8, 2]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 4, 2]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,3915643
- No: 3 GFLOPS: 0.00/24.37 result: Traceback (most recent call last):
+ No: 1 GFLOPS: 62.40/62.40 result: MeasureResult(costs=(0.003710196272727273,), error_no=MeasureErrorNo.NO_ERROR, all_cost=2.506129264831543, timestamp=1668612674.5134175) [('tile_f', [-1, 8, 4, 4]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 1, 4]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,4520238
+ No: 2 GFLOPS: 46.81/62.40 result: MeasureResult(costs=(0.004945171857142857,), error_no=MeasureErrorNo.NO_ERROR, all_cost=2.2848103046417236, timestamp=1668612676.0735888) [('tile_f', [-1, 4, 16, 2]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 1, 16]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,4185807
+ No: 3 GFLOPS: 0.00/62.40 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -389,8 +389,8 @@ for this template
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
- tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 16, 16, 2]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 16, 32]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,8482849
- No: 4 GFLOPS: 0.00/24.37 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 64, 4, 1]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 8, 8]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,4948265
+ No: 4 GFLOPS: 0.00/62.40 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -512,8 +512,9 @@ for this template
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
- tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 2, 1, 64]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 4, 4]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9755001
- No: 5 GFLOPS: 0.00/24.37 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 1, 32, 16]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 128, 4]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,3192384
+ No: 5 GFLOPS: 4.12/62.40 result: MeasureResult(costs=(0.05621964625,), error_no=MeasureErrorNo.NO_ERROR, all_cost=2.726212501525879, timestamp=1668612679.9325876) [('tile_f', [-1, 8, 4, 8]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 4, 4]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,74952
+ No: 6 GFLOPS: 0.00/62.40 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -635,9 +636,10 @@ for this template
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
- tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 1, 2, 16]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 4, 16]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,902170
- No: 6 GFLOPS: 74.52/74.52 result: MeasureResult(costs=(0.0031066027272727273,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.2222695350646973, timestamp=1668585633.5934608) [('tile_f', [-1, 8, 4, 4]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 4, 4]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,1818638
- No: 7 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 16, 4, 1]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 8, 8]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 1)],None,5528183
+ No: 7 GFLOPS: 200.02/200.02 result: MeasureResult(costs=(0.001157369597122302,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.7944960594177246, timestamp=1668612681.7724178) [('tile_f', [-1, 1, 16, 1]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 1, 32]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,8467394
+ No: 8 GFLOPS: 87.19/200.02 result: MeasureResult(costs=(0.0026551532195121953,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.1284987926483154, timestamp=1668612682.474605) [('tile_f', [-1, 1, 2, 2]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 4, 16]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,127224
+ No: 9 GFLOPS: 0.00/200.02 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -759,8 +761,10 @@ for this template
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
- tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 2, 1, 128]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 8, 32]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,7317191
- No: 8 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 1, 32, 1]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 2, 64]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,7907720
+ No: 10 GFLOPS: 46.96/200.02 result: MeasureResult(costs=(0.004929517181818182,), error_no=MeasureErrorNo.NO_ERROR, all_cost=3.243342876434326, timestamp=1668612685.9050357) [('tile_f', [-1, 1, 2, 1]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 16, 8]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,3789950
+ No: 11 GFLOPS: 413.73/413.73 result: MeasureResult(costs=(0.0005595513016759776,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.5997123718261719, timestamp=1668612686.585363) [('tile_f', [-1, 8, 4, 1]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 2, 2]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,7978322
+ No: 12 GFLOPS: 0.00/413.73 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -882,8 +886,8 @@ for this template
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
- tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 8, 8, 1]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 4, 64]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9071950
- No: 9 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 8, 16, 1]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 1, 16]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2832097
+ No: 13 GFLOPS: 0.00/413.73 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -1005,8 +1009,8 @@ for this template
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
- tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 1, 2, 8]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 64, 8]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,10184603
- No: 10 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 32, 4, 4]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 32, 4]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,7831240
+ No: 14 GFLOPS: 0.00/413.73 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -1128,8 +1132,8 @@ for this template
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
- tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 2, 4, 8]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 2, 32]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2276710
- No: 11 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 1, 2, 64]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 32, 1]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,4861104
+ No: 15 GFLOPS: 0.00/413.73 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -1251,8 +1255,8 @@ for this template
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
- tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 4, 8, 16]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 32, 4]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 0), ('unroll_explicit', 1)],None,6473901
- No: 12 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 16, 8, 1]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 128, 2]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,7030351
+ No: 16 GFLOPS: 0.00/413.73 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -1374,8 +1378,8 @@ for this template
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
- tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 8, 32, 2]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 2, 4]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9170793
- No: 13 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 64, 1, 4]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 512, 1]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2939086
+ No: 17 GFLOPS: 0.00/413.73 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -1497,8 +1501,8 @@ for this template
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
- tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 128, 1, 4]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 16, 4]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,8019767
- No: 14 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 1, 1, 256]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 4, 2]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,10112296
+ No: 18 GFLOPS: 0.00/413.73 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -1620,8 +1624,9 @@ for this template
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
- tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 8, 16, 2]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 1, 512]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,8708568
- No: 15 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 32, 8, 2]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 64, 4]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,4543744
+ No: 19 GFLOPS: 439.07/439.07 result: MeasureResult(costs=(0.000527258671875,), error_no=MeasureErrorNo.NO_ERROR, all_cost=2.0887837409973145, timestamp=1668612689.1139297) [('tile_f', [-1, 1, 2, 8]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 1, 4]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9168643
+ No: 20 GFLOPS: 0.00/439.07 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -1743,622 +1748,7 @@ for this template
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
- tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 64, 2, 2]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 8, 8]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2432170
- No: 16 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
- func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
- func = build(s, args, target_host=task.target_host, runtime=runtime)
- File "/workspace/python/tvm/driver/build_module.py", line 227, in build
- input_mod = lower(inputs, args, name=name, binds=binds)
- File "/workspace/python/tvm/driver/build_module.py", line 134, in lower
- return ffi.lower_schedule(inp, args, name, binds, simple_mode)
- File "tvm/_ffi/_cython/./packed_func.pxi", line 331, in tvm._ffi._cy3.core.PackedFuncBase.__call__
- File "tvm/_ffi/_cython/./packed_func.pxi", line 276, in tvm._ffi._cy3.core.FuncCall
- File "tvm/_ffi/_cython/./base.pxi", line 181, in tvm._ffi._cy3.core.CHECK_CALL
- tvm._ffi.base.TVMError: Traceback (most recent call last):
- 24: TVMFuncCall
- at ../src/runtime/c_runtime_api.cc:477
- 23: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 22: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 21: operator()
- at ../include/tvm/runtime/packed_func.h:1731
- 20: unpack_call<tvm::IRModule, 5, tvm::<lambda(tvm::te::Schedule, const tvm::runtime::Array<tvm::runtime::ObjectRef>&, const tvm::runtime::String&, const tvm::runtime::Map<tvm::te::Tensor, tvm::tir::Buffer>&, bool)> >
- at ../include/tvm/runtime/packed_func.h:1671
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1631
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1646
- 13: operator()
- at ../src/driver/driver_api.cc:388
- 12: tvm::LowerSchedule(tvm::te::Schedule, tvm::runtime::Array<tvm::runtime::ObjectRef, void> const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::unordered_map<tvm::te::Tensor, tvm::tir::Buffer, std::hash<tvm::te::Tensor>, std::equal_to<tvm::te::Tensor>, std::allocator<std::pair<tvm::te::Tensor const, tvm::tir::Buffer> > > const&, tvm::GlobalVarSupply, bool)
- at ../src/driver/driver_api.cc:374
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:269
- 10: tvm::transform::Pass::operator()(tvm::IRModule) const
- at ../src/ir/transform.cc:258
- 9: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 8: tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:453
- 7: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 6: tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/tir/ir/transform.cc:100
- 5: tvm::runtime::TypedPackedFunc<tvm::tir::PrimFunc (tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext)>::operator()(tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext) const
- at ../include/tvm/runtime/packed_func.h:1750
- 4: tvm::tir::PrimFunc tvm::runtime::detail::typed_packed_call_dispatcher<tvm::tir::PrimFunc>::run<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::runtime::PackedFunc const&, tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&)
- at ../include/tvm/runtime/packed_func.h:1694
- 3: tvm::runtime::TVMRetValue tvm::runtime::PackedFunc::operator()<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&) const
- at ../include/tvm/runtime/packed_func.h:1618
- 2: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 1: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 0: operator()
- at ../src/runtime/c_runtime_api.cc:534
- File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
- raise InstantiationError("Skipped because of invalid gpu kernel")
- tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel
-
- Traceback (most recent call last):
- 24: TVMFuncCall
- at ../src/runtime/c_runtime_api.cc:477
- 23: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 22: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 21: operator()
- at ../include/tvm/runtime/packed_func.h:1731
- 20: unpack_call<tvm::IRModule, 5, tvm::<lambda(tvm::te::Schedule, const tvm::runtime::Array<tvm::runtime::ObjectRef>&, const tvm::runtime::String&, const tvm::runtime::Map<tvm::te::Tensor, tvm::tir::Buffer>&, bool)> >
- at ../include/tvm/runtime/packed_func.h:1671
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1631
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1646
- 13: operator()
- at ../src/driver/driver_api.cc:388
- 12: tvm::LowerSchedule(tvm::te::Schedule, tvm::runtime::Array<tvm::runtime::ObjectRef, void> const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::unordered_map<tvm::te::Tensor, tvm::tir::Buffer, std::hash<tvm::te::Tensor>, std::equal_to<tvm::te::Tensor>, std::allocator<std::pair<tvm::te::Tensor const, tvm::tir::Buffer> > > const&, tvm::GlobalVarSupply, bool)
- at ../src/driver/driver_api.cc:374
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:269
- 10: tvm::transform::Pass::operator()(tvm::IRModule) const
- at ../src/ir/transform.cc:258
- 9: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 8: tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:453
- 7: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 6: tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/tir/ir/transform.cc:100
- 5: tvm::runtime::TypedPackedFunc<tvm::tir::PrimFunc (tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext)>::operator()(tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext) const
- at ../include/tvm/runtime/packed_func.h:1750
- 4: tvm::tir::PrimFunc tvm::runtime::detail::typed_packed_call_dispatcher<tvm::tir::PrimFunc>::run<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::runtime::PackedFunc const&, tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&)
- at ../include/tvm/runtime/packed_func.h:1694
- 3: tvm::runtime::TVMRetValue tvm::runtime::PackedFunc::operator()<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&) const
- at ../include/tvm/runtime/packed_func.h:1618
- 2: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 1: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 0: operator()
- at ../src/runtime/c_runtime_api.cc:534
- File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
- raise InstantiationError("Skipped because of invalid gpu kernel")
- tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 2, 2, 128]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 8, 8]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,4172514
- No: 17 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
- func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
- func = build(s, args, target_host=task.target_host, runtime=runtime)
- File "/workspace/python/tvm/driver/build_module.py", line 227, in build
- input_mod = lower(inputs, args, name=name, binds=binds)
- File "/workspace/python/tvm/driver/build_module.py", line 134, in lower
- return ffi.lower_schedule(inp, args, name, binds, simple_mode)
- File "tvm/_ffi/_cython/./packed_func.pxi", line 331, in tvm._ffi._cy3.core.PackedFuncBase.__call__
- File "tvm/_ffi/_cython/./packed_func.pxi", line 276, in tvm._ffi._cy3.core.FuncCall
- File "tvm/_ffi/_cython/./base.pxi", line 181, in tvm._ffi._cy3.core.CHECK_CALL
- tvm._ffi.base.TVMError: Traceback (most recent call last):
- 24: TVMFuncCall
- at ../src/runtime/c_runtime_api.cc:477
- 23: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 22: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 21: operator()
- at ../include/tvm/runtime/packed_func.h:1731
- 20: unpack_call<tvm::IRModule, 5, tvm::<lambda(tvm::te::Schedule, const tvm::runtime::Array<tvm::runtime::ObjectRef>&, const tvm::runtime::String&, const tvm::runtime::Map<tvm::te::Tensor, tvm::tir::Buffer>&, bool)> >
- at ../include/tvm/runtime/packed_func.h:1671
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1631
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1646
- 13: operator()
- at ../src/driver/driver_api.cc:388
- 12: tvm::LowerSchedule(tvm::te::Schedule, tvm::runtime::Array<tvm::runtime::ObjectRef, void> const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::unordered_map<tvm::te::Tensor, tvm::tir::Buffer, std::hash<tvm::te::Tensor>, std::equal_to<tvm::te::Tensor>, std::allocator<std::pair<tvm::te::Tensor const, tvm::tir::Buffer> > > const&, tvm::GlobalVarSupply, bool)
- at ../src/driver/driver_api.cc:374
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:269
- 10: tvm::transform::Pass::operator()(tvm::IRModule) const
- at ../src/ir/transform.cc:258
- 9: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 8: tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:453
- 7: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 6: tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/tir/ir/transform.cc:100
- 5: tvm::runtime::TypedPackedFunc<tvm::tir::PrimFunc (tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext)>::operator()(tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext) const
- at ../include/tvm/runtime/packed_func.h:1750
- 4: tvm::tir::PrimFunc tvm::runtime::detail::typed_packed_call_dispatcher<tvm::tir::PrimFunc>::run<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::runtime::PackedFunc const&, tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&)
- at ../include/tvm/runtime/packed_func.h:1694
- 3: tvm::runtime::TVMRetValue tvm::runtime::PackedFunc::operator()<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&) const
- at ../include/tvm/runtime/packed_func.h:1618
- 2: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 1: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 0: operator()
- at ../src/runtime/c_runtime_api.cc:534
- File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
- raise InstantiationError("Skipped because of invalid gpu kernel")
- tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel
-
- Traceback (most recent call last):
- 24: TVMFuncCall
- at ../src/runtime/c_runtime_api.cc:477
- 23: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 22: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 21: operator()
- at ../include/tvm/runtime/packed_func.h:1731
- 20: unpack_call<tvm::IRModule, 5, tvm::<lambda(tvm::te::Schedule, const tvm::runtime::Array<tvm::runtime::ObjectRef>&, const tvm::runtime::String&, const tvm::runtime::Map<tvm::te::Tensor, tvm::tir::Buffer>&, bool)> >
- at ../include/tvm/runtime/packed_func.h:1671
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1631
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1646
- 13: operator()
- at ../src/driver/driver_api.cc:388
- 12: tvm::LowerSchedule(tvm::te::Schedule, tvm::runtime::Array<tvm::runtime::ObjectRef, void> const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::unordered_map<tvm::te::Tensor, tvm::tir::Buffer, std::hash<tvm::te::Tensor>, std::equal_to<tvm::te::Tensor>, std::allocator<std::pair<tvm::te::Tensor const, tvm::tir::Buffer> > > const&, tvm::GlobalVarSupply, bool)
- at ../src/driver/driver_api.cc:374
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:269
- 10: tvm::transform::Pass::operator()(tvm::IRModule) const
- at ../src/ir/transform.cc:258
- 9: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 8: tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:453
- 7: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 6: tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/tir/ir/transform.cc:100
- 5: tvm::runtime::TypedPackedFunc<tvm::tir::PrimFunc (tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext)>::operator()(tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext) const
- at ../include/tvm/runtime/packed_func.h:1750
- 4: tvm::tir::PrimFunc tvm::runtime::detail::typed_packed_call_dispatcher<tvm::tir::PrimFunc>::run<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::runtime::PackedFunc const&, tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&)
- at ../include/tvm/runtime/packed_func.h:1694
- 3: tvm::runtime::TVMRetValue tvm::runtime::PackedFunc::operator()<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&) const
- at ../include/tvm/runtime/packed_func.h:1618
- 2: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 1: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 0: operator()
- at ../src/runtime/c_runtime_api.cc:534
- File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
- raise InstantiationError("Skipped because of invalid gpu kernel")
- tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 16, 32, 1]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 32, 8]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 1)],None,6116044
- No: 18 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
- func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
- func = build(s, args, target_host=task.target_host, runtime=runtime)
- File "/workspace/python/tvm/driver/build_module.py", line 227, in build
- input_mod = lower(inputs, args, name=name, binds=binds)
- File "/workspace/python/tvm/driver/build_module.py", line 134, in lower
- return ffi.lower_schedule(inp, args, name, binds, simple_mode)
- File "tvm/_ffi/_cython/./packed_func.pxi", line 331, in tvm._ffi._cy3.core.PackedFuncBase.__call__
- File "tvm/_ffi/_cython/./packed_func.pxi", line 276, in tvm._ffi._cy3.core.FuncCall
- File "tvm/_ffi/_cython/./base.pxi", line 181, in tvm._ffi._cy3.core.CHECK_CALL
- tvm._ffi.base.TVMError: Traceback (most recent call last):
- 24: TVMFuncCall
- at ../src/runtime/c_runtime_api.cc:477
- 23: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 22: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 21: operator()
- at ../include/tvm/runtime/packed_func.h:1731
- 20: unpack_call<tvm::IRModule, 5, tvm::<lambda(tvm::te::Schedule, const tvm::runtime::Array<tvm::runtime::ObjectRef>&, const tvm::runtime::String&, const tvm::runtime::Map<tvm::te::Tensor, tvm::tir::Buffer>&, bool)> >
- at ../include/tvm/runtime/packed_func.h:1671
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1631
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1646
- 13: operator()
- at ../src/driver/driver_api.cc:388
- 12: tvm::LowerSchedule(tvm::te::Schedule, tvm::runtime::Array<tvm::runtime::ObjectRef, void> const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::unordered_map<tvm::te::Tensor, tvm::tir::Buffer, std::hash<tvm::te::Tensor>, std::equal_to<tvm::te::Tensor>, std::allocator<std::pair<tvm::te::Tensor const, tvm::tir::Buffer> > > const&, tvm::GlobalVarSupply, bool)
- at ../src/driver/driver_api.cc:374
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:269
- 10: tvm::transform::Pass::operator()(tvm::IRModule) const
- at ../src/ir/transform.cc:258
- 9: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 8: tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:453
- 7: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 6: tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/tir/ir/transform.cc:100
- 5: tvm::runtime::TypedPackedFunc<tvm::tir::PrimFunc (tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext)>::operator()(tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext) const
- at ../include/tvm/runtime/packed_func.h:1750
- 4: tvm::tir::PrimFunc tvm::runtime::detail::typed_packed_call_dispatcher<tvm::tir::PrimFunc>::run<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::runtime::PackedFunc const&, tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&)
- at ../include/tvm/runtime/packed_func.h:1694
- 3: tvm::runtime::TVMRetValue tvm::runtime::PackedFunc::operator()<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&) const
- at ../include/tvm/runtime/packed_func.h:1618
- 2: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 1: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 0: operator()
- at ../src/runtime/c_runtime_api.cc:534
- File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
- raise InstantiationError("Skipped because of invalid gpu kernel")
- tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel
-
- Traceback (most recent call last):
- 24: TVMFuncCall
- at ../src/runtime/c_runtime_api.cc:477
- 23: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 22: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 21: operator()
- at ../include/tvm/runtime/packed_func.h:1731
- 20: unpack_call<tvm::IRModule, 5, tvm::<lambda(tvm::te::Schedule, const tvm::runtime::Array<tvm::runtime::ObjectRef>&, const tvm::runtime::String&, const tvm::runtime::Map<tvm::te::Tensor, tvm::tir::Buffer>&, bool)> >
- at ../include/tvm/runtime/packed_func.h:1671
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1631
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1646
- 13: operator()
- at ../src/driver/driver_api.cc:388
- 12: tvm::LowerSchedule(tvm::te::Schedule, tvm::runtime::Array<tvm::runtime::ObjectRef, void> const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::unordered_map<tvm::te::Tensor, tvm::tir::Buffer, std::hash<tvm::te::Tensor>, std::equal_to<tvm::te::Tensor>, std::allocator<std::pair<tvm::te::Tensor const, tvm::tir::Buffer> > > const&, tvm::GlobalVarSupply, bool)
- at ../src/driver/driver_api.cc:374
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:269
- 10: tvm::transform::Pass::operator()(tvm::IRModule) const
- at ../src/ir/transform.cc:258
- 9: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 8: tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:453
- 7: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 6: tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/tir/ir/transform.cc:100
- 5: tvm::runtime::TypedPackedFunc<tvm::tir::PrimFunc (tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext)>::operator()(tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext) const
- at ../include/tvm/runtime/packed_func.h:1750
- 4: tvm::tir::PrimFunc tvm::runtime::detail::typed_packed_call_dispatcher<tvm::tir::PrimFunc>::run<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::runtime::PackedFunc const&, tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&)
- at ../include/tvm/runtime/packed_func.h:1694
- 3: tvm::runtime::TVMRetValue tvm::runtime::PackedFunc::operator()<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&) const
- at ../include/tvm/runtime/packed_func.h:1618
- 2: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 1: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 0: operator()
- at ../src/runtime/c_runtime_api.cc:534
- File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
- raise InstantiationError("Skipped because of invalid gpu kernel")
- tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 16, 1, 4]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 256, 2]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,65444
- No: 19 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
- func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
- func = build(s, args, target_host=task.target_host, runtime=runtime)
- File "/workspace/python/tvm/driver/build_module.py", line 227, in build
- input_mod = lower(inputs, args, name=name, binds=binds)
- File "/workspace/python/tvm/driver/build_module.py", line 134, in lower
- return ffi.lower_schedule(inp, args, name, binds, simple_mode)
- File "tvm/_ffi/_cython/./packed_func.pxi", line 331, in tvm._ffi._cy3.core.PackedFuncBase.__call__
- File "tvm/_ffi/_cython/./packed_func.pxi", line 276, in tvm._ffi._cy3.core.FuncCall
- File "tvm/_ffi/_cython/./base.pxi", line 181, in tvm._ffi._cy3.core.CHECK_CALL
- tvm._ffi.base.TVMError: Traceback (most recent call last):
- 24: TVMFuncCall
- at ../src/runtime/c_runtime_api.cc:477
- 23: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 22: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 21: operator()
- at ../include/tvm/runtime/packed_func.h:1731
- 20: unpack_call<tvm::IRModule, 5, tvm::<lambda(tvm::te::Schedule, const tvm::runtime::Array<tvm::runtime::ObjectRef>&, const tvm::runtime::String&, const tvm::runtime::Map<tvm::te::Tensor, tvm::tir::Buffer>&, bool)> >
- at ../include/tvm/runtime/packed_func.h:1671
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1631
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1646
- 13: operator()
- at ../src/driver/driver_api.cc:388
- 12: tvm::LowerSchedule(tvm::te::Schedule, tvm::runtime::Array<tvm::runtime::ObjectRef, void> const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::unordered_map<tvm::te::Tensor, tvm::tir::Buffer, std::hash<tvm::te::Tensor>, std::equal_to<tvm::te::Tensor>, std::allocator<std::pair<tvm::te::Tensor const, tvm::tir::Buffer> > > const&, tvm::GlobalVarSupply, bool)
- at ../src/driver/driver_api.cc:374
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:269
- 10: tvm::transform::Pass::operator()(tvm::IRModule) const
- at ../src/ir/transform.cc:258
- 9: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 8: tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:453
- 7: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 6: tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/tir/ir/transform.cc:100
- 5: tvm::runtime::TypedPackedFunc<tvm::tir::PrimFunc (tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext)>::operator()(tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext) const
- at ../include/tvm/runtime/packed_func.h:1750
- 4: tvm::tir::PrimFunc tvm::runtime::detail::typed_packed_call_dispatcher<tvm::tir::PrimFunc>::run<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::runtime::PackedFunc const&, tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&)
- at ../include/tvm/runtime/packed_func.h:1694
- 3: tvm::runtime::TVMRetValue tvm::runtime::PackedFunc::operator()<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&) const
- at ../include/tvm/runtime/packed_func.h:1618
- 2: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 1: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 0: operator()
- at ../src/runtime/c_runtime_api.cc:534
- File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
- raise InstantiationError("Skipped because of invalid gpu kernel")
- tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel
-
- Traceback (most recent call last):
- 24: TVMFuncCall
- at ../src/runtime/c_runtime_api.cc:477
- 23: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 22: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 21: operator()
- at ../include/tvm/runtime/packed_func.h:1731
- 20: unpack_call<tvm::IRModule, 5, tvm::<lambda(tvm::te::Schedule, const tvm::runtime::Array<tvm::runtime::ObjectRef>&, const tvm::runtime::String&, const tvm::runtime::Map<tvm::te::Tensor, tvm::tir::Buffer>&, bool)> >
- at ../include/tvm/runtime/packed_func.h:1671
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1631
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1646
- 13: operator()
- at ../src/driver/driver_api.cc:388
- 12: tvm::LowerSchedule(tvm::te::Schedule, tvm::runtime::Array<tvm::runtime::ObjectRef, void> const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::unordered_map<tvm::te::Tensor, tvm::tir::Buffer, std::hash<tvm::te::Tensor>, std::equal_to<tvm::te::Tensor>, std::allocator<std::pair<tvm::te::Tensor const, tvm::tir::Buffer> > > const&, tvm::GlobalVarSupply, bool)
- at ../src/driver/driver_api.cc:374
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:269
- 10: tvm::transform::Pass::operator()(tvm::IRModule) const
- at ../src/ir/transform.cc:258
- 9: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 8: tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:453
- 7: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 6: tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/tir/ir/transform.cc:100
- 5: tvm::runtime::TypedPackedFunc<tvm::tir::PrimFunc (tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext)>::operator()(tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext) const
- at ../include/tvm/runtime/packed_func.h:1750
- 4: tvm::tir::PrimFunc tvm::runtime::detail::typed_packed_call_dispatcher<tvm::tir::PrimFunc>::run<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::runtime::PackedFunc const&, tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&)
- at ../include/tvm/runtime/packed_func.h:1694
- 3: tvm::runtime::TVMRetValue tvm::runtime::PackedFunc::operator()<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&) const
- at ../include/tvm/runtime/packed_func.h:1618
- 2: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 1: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 0: operator()
- at ../src/runtime/c_runtime_api.cc:534
- File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
- raise InstantiationError("Skipped because of invalid gpu kernel")
- tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 2, 4, 16]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 4, 32]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,537636
- No: 20 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
- func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
- func = build(s, args, target_host=task.target_host, runtime=runtime)
- File "/workspace/python/tvm/driver/build_module.py", line 227, in build
- input_mod = lower(inputs, args, name=name, binds=binds)
- File "/workspace/python/tvm/driver/build_module.py", line 134, in lower
- return ffi.lower_schedule(inp, args, name, binds, simple_mode)
- File "tvm/_ffi/_cython/./packed_func.pxi", line 331, in tvm._ffi._cy3.core.PackedFuncBase.__call__
- File "tvm/_ffi/_cython/./packed_func.pxi", line 276, in tvm._ffi._cy3.core.FuncCall
- File "tvm/_ffi/_cython/./base.pxi", line 181, in tvm._ffi._cy3.core.CHECK_CALL
- tvm._ffi.base.TVMError: Traceback (most recent call last):
- 24: TVMFuncCall
- at ../src/runtime/c_runtime_api.cc:477
- 23: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 22: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 21: operator()
- at ../include/tvm/runtime/packed_func.h:1731
- 20: unpack_call<tvm::IRModule, 5, tvm::<lambda(tvm::te::Schedule, const tvm::runtime::Array<tvm::runtime::ObjectRef>&, const tvm::runtime::String&, const tvm::runtime::Map<tvm::te::Tensor, tvm::tir::Buffer>&, bool)> >
- at ../include/tvm/runtime/packed_func.h:1671
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1631
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1646
- 13: operator()
- at ../src/driver/driver_api.cc:388
- 12: tvm::LowerSchedule(tvm::te::Schedule, tvm::runtime::Array<tvm::runtime::ObjectRef, void> const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::unordered_map<tvm::te::Tensor, tvm::tir::Buffer, std::hash<tvm::te::Tensor>, std::equal_to<tvm::te::Tensor>, std::allocator<std::pair<tvm::te::Tensor const, tvm::tir::Buffer> > > const&, tvm::GlobalVarSupply, bool)
- at ../src/driver/driver_api.cc:374
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:269
- 10: tvm::transform::Pass::operator()(tvm::IRModule) const
- at ../src/ir/transform.cc:258
- 9: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 8: tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:453
- 7: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 6: tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/tir/ir/transform.cc:100
- 5: tvm::runtime::TypedPackedFunc<tvm::tir::PrimFunc (tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext)>::operator()(tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext) const
- at ../include/tvm/runtime/packed_func.h:1750
- 4: tvm::tir::PrimFunc tvm::runtime::detail::typed_packed_call_dispatcher<tvm::tir::PrimFunc>::run<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::runtime::PackedFunc const&, tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&)
- at ../include/tvm/runtime/packed_func.h:1694
- 3: tvm::runtime::TVMRetValue tvm::runtime::PackedFunc::operator()<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&) const
- at ../include/tvm/runtime/packed_func.h:1618
- 2: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 1: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 0: operator()
- at ../src/runtime/c_runtime_api.cc:534
- File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
- raise InstantiationError("Skipped because of invalid gpu kernel")
- tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel
-
- Traceback (most recent call last):
- 24: TVMFuncCall
- at ../src/runtime/c_runtime_api.cc:477
- 23: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 22: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 21: operator()
- at ../include/tvm/runtime/packed_func.h:1731
- 20: unpack_call<tvm::IRModule, 5, tvm::<lambda(tvm::te::Schedule, const tvm::runtime::Array<tvm::runtime::ObjectRef>&, const tvm::runtime::String&, const tvm::runtime::Map<tvm::te::Tensor, tvm::tir::Buffer>&, bool)> >
- at ../include/tvm/runtime/packed_func.h:1671
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1631
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1646
- 13: operator()
- at ../src/driver/driver_api.cc:388
- 12: tvm::LowerSchedule(tvm::te::Schedule, tvm::runtime::Array<tvm::runtime::ObjectRef, void> const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::unordered_map<tvm::te::Tensor, tvm::tir::Buffer, std::hash<tvm::te::Tensor>, std::equal_to<tvm::te::Tensor>, std::allocator<std::pair<tvm::te::Tensor const, tvm::tir::Buffer> > > const&, tvm::GlobalVarSupply, bool)
- at ../src/driver/driver_api.cc:374
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:269
- 10: tvm::transform::Pass::operator()(tvm::IRModule) const
- at ../src/ir/transform.cc:258
- 9: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 8: tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:453
- 7: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 6: tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/tir/ir/transform.cc:100
- 5: tvm::runtime::TypedPackedFunc<tvm::tir::PrimFunc (tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext)>::operator()(tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext) const
- at ../include/tvm/runtime/packed_func.h:1750
- 4: tvm::tir::PrimFunc tvm::runtime::detail::typed_packed_call_dispatcher<tvm::tir::PrimFunc>::run<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::runtime::PackedFunc const&, tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&)
- at ../include/tvm/runtime/packed_func.h:1694
- 3: tvm::runtime::TVMRetValue tvm::runtime::PackedFunc::operator()<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&) const
- at ../include/tvm/runtime/packed_func.h:1618
- 2: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 1: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 0: operator()
- at ../src/runtime/c_runtime_api.cc:534
- File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
- raise InstantiationError("Skipped because of invalid gpu kernel")
- tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 4, 16, 4]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 1, 16]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,8446368
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 2, 128, 1]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 8, 32]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,10412870
@@ -2413,9 +1803,9 @@ and measure running time.
Finish loading 20 records
Best config:
- [('tile_f', [-1, 8, 4, 4]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 4, 4]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,1818638
+ [('tile_f', [-1, 1, 2, 8]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 1, 4]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9168643
Finish loading 20 records
- Time cost of this operator: 0.001609
+ Time cost of this operator: 0.000886
diff --git a/docs/_sources/how_to/work_with_microtvm/micro_autotune.rst.txt b/docs/_sources/how_to/work_with_microtvm/micro_autotune.rst.txt
index 1685d38891..70d4e25b46 100644
--- a/docs/_sources/how_to/work_with_microtvm/micro_autotune.rst.txt
+++ b/docs/_sources/how_to/work_with_microtvm/micro_autotune.rst.txt
@@ -327,10 +327,10 @@ Timing the untuned program
########## Build without Autotuning ##########
Node Name Ops Time(us) Time(%) Shape Inputs Outputs Measurements(us)
--------- --- -------- ------- ----- ------ ------- ----------------
- tvmgen_default_fused_nn_contrib_conv2d_NCHWc tvmgen_default_fused_nn_contrib_conv2d_NCHWc 311.4 98.717 (1, 2, 10, 10, 3) 2 1 [311.4]
- tvmgen_default_fused_layout_transform_1 tvmgen_default_fused_layout_transform_1 3.084 0.978 (1, 6, 10, 10) 1 1 [3.084]
- tvmgen_default_fused_layout_transform tvmgen_default_fused_layout_transform 0.963 0.305 (1, 1, 10, 10, 3) 1 1 [0.963]
- Total_time - 315.448 - - - - -
+ tvmgen_default_fused_nn_contrib_conv2d_NCHWc tvmgen_default_fused_nn_contrib_conv2d_NCHWc 311.8 98.714 (1, 2, 10, 10, 3) 2 1 [311.8]
+ tvmgen_default_fused_layout_transform_1 tvmgen_default_fused_layout_transform_1 3.102 0.982 (1, 6, 10, 10) 1 1 [3.102]
+ tvmgen_default_fused_layout_transform tvmgen_default_fused_layout_transform 0.962 0.305 (1, 1, 10, 10, 3) 1 1 [0.962]
+ Total_time - 315.863 - - - - -
@@ -394,10 +394,10 @@ Timing the tuned program
########## Build with Autotuning ##########
Node Name Ops Time(us) Time(%) Shape Inputs Outputs Measurements(us)
--------- --- -------- ------- ----- ------ ------- ----------------
- tvmgen_default_fused_nn_contrib_conv2d_NCHWc tvmgen_default_fused_nn_contrib_conv2d_NCHWc 100.2 97.298 (1, 6, 10, 10, 1) 2 1 [100.2]
- tvmgen_default_fused_layout_transform_1 tvmgen_default_fused_layout_transform_1 1.807 1.755 (1, 6, 10, 10) 1 1 [1.807]
- tvmgen_default_fused_layout_transform tvmgen_default_fused_layout_transform 0.976 0.947 (1, 1, 10, 10, 3) 1 1 [0.976]
- Total_time - 102.983 - - - - -
+ tvmgen_default_fused_nn_contrib_conv2d_NCHWc tvmgen_default_fused_nn_contrib_conv2d_NCHWc 106.6 97.612 (1, 6, 10, 10, 1) 2 1 [106.6]
+ tvmgen_default_fused_layout_transform_1 tvmgen_default_fused_layout_transform_1 1.77 1.621 (1, 6, 10, 10) 1 1 [1.77]
+ tvmgen_default_fused_layout_transform tvmgen_default_fused_layout_transform 0.838 0.767 (1, 3, 10, 10, 1) 1 1 [0.838]
+ Total_time - 109.208 - - - - -
diff --git a/docs/_sources/how_to/work_with_microtvm/micro_pytorch.rst.txt b/docs/_sources/how_to/work_with_microtvm/micro_pytorch.rst.txt
index bb97b9d919..8d471f7855 100644
--- a/docs/_sources/how_to/work_with_microtvm/micro_pytorch.rst.txt
+++ b/docs/_sources/how_to/work_with_microtvm/micro_pytorch.rst.txt
@@ -109,7 +109,7 @@ download a cat image and preprocess it to use as the model input.
/venv/apache-tvm-py3.7/lib/python3.7/site-packages/torch/ao/quantization/utils.py:281: UserWarning: must run observer before calling calculate_qparams. Returning default values.
"must run observer before calling calculate_qparams. " +
Downloading: "https://download.pytorch.org/models/quantized/mobilenet_v2_qnnpack_37f702c5.pth" to /workspace/.cache/torch/hub/checkpoints/mobilenet_v2_qnnpack_37f702c5.pth
-
0%| | 0.00/3.42M [00:00<?, ?B/s]
100%|##########| 3.42M/3.42M [00:00<00:00, 83.1MB/s]
+
0%| | 0.00/3.42M [00:00<?, ?B/s]
100%|##########| 3.42M/3.42M [00:00<00:00, 81.3MB/s]
/workspace/python/tvm/relay/frontend/pytorch_utils.py:47: DeprecationWarning: distutils Version classes are deprecated. Use packaging.version instead.
return LooseVersion(torch_ver) > ver
/venv/apache-tvm-py3.7/lib/python3.7/site-packages/setuptools/_distutils/version.py:346: DeprecationWarning: distutils Version classes are deprecated. Use packaging.version instead.
@@ -314,7 +314,7 @@ Look up prediction top 1 index in 1000 class synset.
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 1 minutes 4.172 seconds)
+ **Total running time of the script:** ( 1 minutes 6.472 seconds)
.. _sphx_glr_download_how_to_work_with_microtvm_micro_pytorch.py:
diff --git a/docs/_sources/how_to/work_with_microtvm/micro_train.rst.txt b/docs/_sources/how_to/work_with_microtvm/micro_train.rst.txt
index c26fb4f49b..230bcc304f 100644
--- a/docs/_sources/how_to/work_with_microtvm/micro_train.rst.txt
+++ b/docs/_sources/how_to/work_with_microtvm/micro_train.rst.txt
@@ -225,7 +225,7 @@ take about **2 minutes** to download the Stanford Cars, while COCO 2017 validati
.. code-block:: none
- '/tmp/tmp19qchhde/images/random'
+ '/tmp/tmpy5o6f33_/images/random'
@@ -316,7 +316,7 @@ objects to other stuff? We can display some examples from our datasets using ``m
.. image-sg:: /how_to/work_with_microtvm/images/sphx_glr_micro_train_001.png
- :alt: [1.0, 0.0], [0.0, 1.0], [1.0, 0.0], [0.0, 1.0], [1.0, 0.0], [0.0, 1.0], [0.0, 1.0], [1.0, 0.0], [0.0, 1.0], [0.0, 1.0]
+ :alt: [0.0, 1.0], [1.0, 0.0], [1.0, 0.0], [0.0, 1.0], [0.0, 1.0], [1.0, 0.0], [1.0, 0.0], [0.0, 1.0], [0.0, 1.0], [0.0, 1.0]
:srcset: /how_to/work_with_microtvm/images/sphx_glr_micro_train_001.png
:class: sphx-glr-single-img
@@ -325,8 +325,8 @@ objects to other stuff? We can display some examples from our datasets using ``m
.. code-block:: none
- /tmp/tmp19qchhde/images/target contains 8144 images
- /tmp/tmp19qchhde/images/random contains 5000 images
+ /tmp/tmpy5o6f33_/images/target contains 8144 images
+ /tmp/tmpy5o6f33_/images/random contains 5000 images
@@ -501,13 +501,13 @@ the time on our validation set).
.. code-block:: none
Epoch 1/3
- 328/328 - 47s - loss: 0.2178 - accuracy: 0.9269 - val_loss: 0.1431 - val_accuracy: 0.9528 - 47s/epoch - 144ms/step
+ 328/328 - 47s - loss: 0.2249 - accuracy: 0.9236 - val_loss: 0.1049 - val_accuracy: 0.9603 - 47s/epoch - 144ms/step
Epoch 2/3
- 328/328 - 43s - loss: 0.1014 - accuracy: 0.9611 - val_loss: 0.1025 - val_accuracy: 0.9656 - 43s/epoch - 132ms/step
+ 328/328 - 43s - loss: 0.1043 - accuracy: 0.9625 - val_loss: 0.1249 - val_accuracy: 0.9596 - 43s/epoch - 133ms/step
Epoch 3/3
- 328/328 - 43s - loss: 0.0720 - accuracy: 0.9741 - val_loss: 0.1185 - val_accuracy: 0.9619 - 43s/epoch - 132ms/step
+ 328/328 - 43s - loss: 0.0645 - accuracy: 0.9761 - val_loss: 0.1150 - val_accuracy: 0.9656 - 43s/epoch - 132ms/step
- <keras.callbacks.History object at 0x7f315a9d0890>
+ <keras.callbacks.History object at 0x7f09494f4550>
@@ -864,7 +864,7 @@ Arduino tutorial for how to do that `on GitHub <https://github.com/guberti/tvm-a
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 4 minutes 52.658 seconds)
+ **Total running time of the script:** ( 4 minutes 38.542 seconds)
.. _sphx_glr_download_how_to_work_with_microtvm_micro_train.py:
diff --git a/docs/_sources/how_to/work_with_microtvm/sg_execution_times.rst.txt b/docs/_sources/how_to/work_with_microtvm/sg_execution_times.rst.txt
index 1059173de3..588e43a29f 100644
--- a/docs/_sources/how_to/work_with_microtvm/sg_execution_times.rst.txt
+++ b/docs/_sources/how_to/work_with_microtvm/sg_execution_times.rst.txt
@@ -5,18 +5,18 @@
Computation times
=================
-**06:59.588** total execution time for **how_to_work_with_microtvm** files:
+**06:48.119** total execution time for **how_to_work_with_microtvm** files:
+---------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_microtvm_micro_train.py` (``micro_train.py``) | 04:52.658 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_microtvm_micro_train.py` (``micro_train.py``) | 04:38.542 | 0.0 MB |
+---------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_microtvm_micro_pytorch.py` (``micro_pytorch.py``) | 01:04.172 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_microtvm_micro_pytorch.py` (``micro_pytorch.py``) | 01:06.472 | 0.0 MB |
+---------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_microtvm_micro_autotune.py` (``micro_autotune.py``) | 00:50.525 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_microtvm_micro_autotune.py` (``micro_autotune.py``) | 00:50.409 | 0.0 MB |
+---------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_microtvm_micro_aot.py` (``micro_aot.py``) | 00:08.438 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_microtvm_micro_aot.py` (``micro_aot.py``) | 00:08.910 | 0.0 MB |
+---------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_microtvm_micro_tflite.py` (``micro_tflite.py``) | 00:03.794 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_microtvm_micro_tflite.py` (``micro_tflite.py``) | 00:03.784 | 0.0 MB |
+---------------------------------------------------------------------------------------------+-----------+--------+
| :ref:`sphx_glr_how_to_work_with_microtvm_micro_reference_vm.py` (``micro_reference_vm.py``) | 00:00.001 | 0.0 MB |
+---------------------------------------------------------------------------------------------+-----------+--------+
diff --git a/docs/_sources/how_to/work_with_relay/sg_execution_times.rst.txt b/docs/_sources/how_to/work_with_relay/sg_execution_times.rst.txt
index 562477be1f..2ed63129e6 100644
--- a/docs/_sources/how_to/work_with_relay/sg_execution_times.rst.txt
+++ b/docs/_sources/how_to/work_with_relay/sg_execution_times.rst.txt
@@ -5,14 +5,14 @@
Computation times
=================
-**00:43.405** total execution time for **how_to_work_with_relay** files:
+**00:44.620** total execution time for **how_to_work_with_relay** files:
+----------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_relay_using_pipeline_executor.py` (``using_pipeline_executor.py``) | 00:31.889 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_relay_using_pipeline_executor.py` (``using_pipeline_executor.py``) | 00:32.524 | 0.0 MB |
+----------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_relay_using_external_lib.py` (``using_external_lib.py``) | 00:09.793 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_relay_using_external_lib.py` (``using_external_lib.py``) | 00:10.560 | 0.0 MB |
+----------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_relay_build_gcn.py` (``build_gcn.py``) | 00:01.717 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_relay_build_gcn.py` (``build_gcn.py``) | 00:01.529 | 0.0 MB |
+----------------------------------------------------------------------------------------------------+-----------+--------+
| :ref:`sphx_glr_how_to_work_with_relay_using_relay_viz.py` (``using_relay_viz.py``) | 00:00.007 | 0.0 MB |
+----------------------------------------------------------------------------------------------------+-----------+--------+
diff --git a/docs/_sources/how_to/work_with_schedules/intrin_math.rst.txt b/docs/_sources/how_to/work_with_schedules/intrin_math.rst.txt
index 97ecd72d36..888f99bc9b 100644
--- a/docs/_sources/how_to/work_with_schedules/intrin_math.rst.txt
+++ b/docs/_sources/how_to/work_with_schedules/intrin_math.rst.txt
@@ -261,7 +261,7 @@ The following example customizes CUDA lowering rule for :code:`exp`.
.. code-block:: none
- <function my_cuda_math_rule at 0x7f315f6bb200>
+ <function my_cuda_math_rule at 0x7f0949fdbdd0>
diff --git a/docs/_sources/how_to/work_with_schedules/reduction.rst.txt b/docs/_sources/how_to/work_with_schedules/reduction.rst.txt
index 3086cec7be..65233eeaff 100644
--- a/docs/_sources/how_to/work_with_schedules/reduction.rst.txt
+++ b/docs/_sources/how_to/work_with_schedules/reduction.rst.txt
@@ -114,14 +114,13 @@ Before doing anything, let us print out the IR code of default schedule.
@main = primfn(A_1: handle, B_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [(stride: int32*n: int32)], [], type="auto"),
- B: Buffer(B_2: Pointer(float32), float32, [(stride_1: int32*n)], [], type="auto")}
- buffer_map = {A_1: A, B_1: B}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [n, m: int32], [stride, stride_2: int32], type="auto"), B_1: B_3: Buffer(B_2, float32, [n], [stride_1], type="auto")} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [n: int32, m: int32], [stride: int32, stride_1: int32], type="auto"),
+ B: Buffer(B_2: Pointer(float32), float32, [n], [stride_2: int32], type="auto")}
+ buffer_map = {A_1: A, B_1: B} {
for (i: int32, 0, n) {
- B[(i*stride_1)] = 0f32
+ B_3: Buffer(B_2, float32, [(stride_2*n)], [], type="auto")[(i*stride_2)] = 0f32
for (k: int32, 0, m) {
- B[(i*stride_1)] = (B[(i*stride_1)] + A[((i*stride) + (k*stride_2))])
+ B_3[(i*stride_2)] = (B_3[(i*stride_2)] + A_3: Buffer(A_2, float32, [(stride*n)], [], type="auto")[((i*stride) + (k*stride_1))])
}
}
}
@@ -158,21 +157,20 @@ axis by different factors. The result is a nested reduction.
@main = primfn(A_1: handle, B_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [(stride: int32*n: int32)], [], type="auto"),
- B: Buffer(B_2: Pointer(float32), float32, [(stride_1: int32*n)], [], type="auto")}
- buffer_map = {A_1: A, B_1: B}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [n, m: int32], [stride, stride_2: int32], type="auto"), B_1: B_3: Buffer(B_2, float32, [n], [stride_1], type="auto")} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [n: int32, m: int32], [stride: int32, stride_1: int32], type="auto"),
+ B: Buffer(B_2: Pointer(float32), float32, [n], [stride_2: int32], type="auto")}
+ buffer_map = {A_1: A, B_1: B} {
for (i.outer: int32, 0, floordiv((n + 31), 32)) {
for (i.inner: int32, 0, 32) {
if @tir.likely((((i.outer*32) + i.inner) < n), dtype=bool) {
- B[(((i.outer*32) + i.inner)*stride_1)] = 0f32
+ B_3: Buffer(B_2, float32, [(stride_2*n)], [], type="auto")[(((i.outer*32) + i.inner)*stride_2)] = 0f32
}
if @tir.likely((((i.outer*32) + i.inner) < n), dtype=bool) {
for (k.outer: int32, 0, floordiv((m + 15), 16)) {
for (k.inner: int32, 0, 16) {
if @tir.likely((((k.outer*16) + k.inner) < m), dtype=bool) {
let cse_var_1: int32 = ((i.outer*32) + i.inner)
- B[(cse_var_1*stride_1)] = (B[(cse_var_1*stride_1)] + A[((cse_var_1*stride) + (((k.outer*16) + k.inner)*stride_2))])
+ B_3[(cse_var_1*stride_2)] = (B_3[(cse_var_1*stride_2)] + A_3: Buffer(A_2, float32, [(stride*n)], [], type="auto")[((cse_var_1*stride) + (((k.outer*16) + k.inner)*stride_1))])
}
}
}
@@ -208,20 +206,19 @@ If we are building a GPU kernel, we can bind the rows of B to GPU threads.
@main = primfn(A_1: handle, B_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [(stride: int32*n: int32)], [], type="auto"),
- B: Buffer(B_2: Pointer(float32), float32, [(stride_1: int32*n)], [], type="auto")}
- buffer_map = {A_1: A, B_1: B}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [n, m: int32], [stride, stride_2: int32], type="auto"), B_1: B_3: Buffer(B_2, float32, [n], [stride_1], type="auto")} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [n: int32, m: int32], [stride: int32, stride_1: int32], type="auto"),
+ B: Buffer(B_2: Pointer(float32), float32, [n], [stride_2: int32], type="auto")}
+ buffer_map = {A_1: A, B_1: B} {
attr [IterVar(blockIdx.x: int32, (nullptr), "ThreadIndex", "blockIdx.x")] "thread_extent" = floordiv((n + 31), 32);
attr [IterVar(threadIdx.x: int32, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 32 {
if @tir.likely((((blockIdx.x*32) + threadIdx.x) < n), dtype=bool) {
- B[(((blockIdx.x*32) + threadIdx.x)*stride_1)] = 0f32
+ B_3: Buffer(B_2, float32, [(stride_2*n)], [], type="auto")[(((blockIdx.x*32) + threadIdx.x)*stride_2)] = 0f32
}
for (k.outer: int32, 0, floordiv((m + 15), 16)) {
for (k.inner: int32, 0, 16) {
if @tir.likely((((blockIdx.x*32) + threadIdx.x) < n), dtype=bool) {
if @tir.likely((((k.outer*16) + k.inner) < m), dtype=bool) {
- B[(((blockIdx.x*32) + threadIdx.x)*stride_1)] = (B[(((blockIdx.x*32) + threadIdx.x)*stride_1)] + A[((((blockIdx.x*32) + threadIdx.x)*stride) + (((k.outer*16) + k.inner)*stride_2))])
+ B_3[(((blockIdx.x*32) + threadIdx.x)*stride_2)] = (B_3[(((blockIdx.x*32) + threadIdx.x)*stride_2)] + A_3: Buffer(A_2, float32, [(stride*n)], [], type="auto")[((((blockIdx.x*32) + threadIdx.x)*stride) + (((k.outer*16) + k.inner)*stride_1))])
}
}
}
@@ -267,25 +264,24 @@ result B.rf. The factored dimension becomes the first dimension of B.rf.
@main = primfn(A_1: handle, B_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [(stride: int32*n: int32)], [], type="auto"),
- B: Buffer(B_2: Pointer(float32), float32, [(stride_1: int32*n)], [], type="auto")}
- buffer_map = {A_1: A, B_1: B}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [n, m: int32], [stride, stride_2: int32], type="auto"), B_1: B_3: Buffer(B_2, float32, [n], [stride_1], type="auto")} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [n: int32, m: int32], [stride: int32, stride_1: int32], type="auto"),
+ B: Buffer(B_2: Pointer(float32), float32, [n], [stride_2: int32], type="auto")}
+ buffer_map = {A_1: A, B_1: B} {
allocate(B.rf: Pointer(global float32), float32, [(n*16)]), storage_scope = global {
for (k.inner: int32, 0, 16) {
for (i: int32, 0, n) {
B.rf_1: Buffer(B.rf, float32, [(16*n)], [])[((k.inner*n) + i)] = 0f32
for (k.outer: int32, 0, floordiv((m + 15), 16)) {
if @tir.likely((((k.outer*16) + k.inner) < m), dtype=bool) {
- B.rf_1[((k.inner*n) + i)] = (B.rf_1[((k.inner*n) + i)] + A[((i*stride) + (((k.outer*16) + k.inner)*stride_2))])
+ B.rf_1[((k.inner*n) + i)] = (B.rf_1[((k.inner*n) + i)] + A_3: Buffer(A_2, float32, [(stride*n)], [], type="auto")[((i*stride) + (((k.outer*16) + k.inner)*stride_1))])
}
}
}
}
for (ax0: int32, 0, n) {
- B[(ax0*stride_1)] = 0f32
+ B_3: Buffer(B_2, float32, [(stride_2*n)], [], type="auto")[(ax0*stride_2)] = 0f32
for (k.inner.v: int32, 0, 16) {
- B[(ax0*stride_1)] = (B[(ax0*stride_1)] + B.rf_1[((k.inner.v*n) + ax0)])
+ B_3[(ax0*stride_2)] = (B_3[(ax0*stride_2)] + B.rf_1[((k.inner.v*n) + ax0)])
}
}
}
@@ -481,17 +477,16 @@ Here is an example for 2D convolution with filter size = [3, 3] and strides = [1
@main = primfn(Input_1: handle, Filter_1: handle, Output_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {Input: Buffer(Input_2: Pointer(float32), float32, [(stride: int32*n: int32)], [], type="auto"),
- Filter: Buffer(Filter_2: Pointer(float32), float32, [9], []),
- Output: Buffer(Output_2: Pointer(float32), float32, [((n - 2)*(n - 2))], [])}
- buffer_map = {Input_1: Input, Filter_1: Filter, Output_1: Output}
- preflattened_buffer_map = {Input_1: Input_3: Buffer(Input_2, float32, [n, n], [stride, stride_1: int32], type="auto"), Filter_1: Filter_3: Buffer(Filter_2, float32, [3, 3], []), Output_1: Output_3: Buffer(Output_2, float32, [(n - 2), (n - 2)], [])} {
+ buffers = {Input: Buffer(Input_2: Pointer(float32), float32, [n: int32, n], [stride: int32, stride_1: int32], type="auto"),
+ Filter: Buffer(Filter_2: Pointer(float32), float32, [3, 3], []),
+ Output: Buffer(Output_2: Pointer(float32), float32, [(n - 2), (n - 2)], [])}
+ buffer_map = {Input_1: Input, Filter_1: Filter, Output_1: Output} {
for (i: int32, 0, (n - 2)) {
for (j: int32, 0, (n - 2)) {
- Output[((i*(n - 2)) + j)] = 0f32
+ Output_3: Buffer(Output_2, float32, [((n - 2)*(n - 2))], [])[((i*(n - 2)) + j)] = 0f32
for (di: int32, 0, 3) {
for (dj: int32, 0, 3) {
- Output[((i*(n - 2)) + j)] = (Output[((i*(n - 2)) + j)] + (Input[(((i + di)*stride) + ((j + dj)*stride_1))]*Filter[((di*3) + dj)]))
+ Output_3[((i*(n - 2)) + j)] = (Output_3[((i*(n - 2)) + j)] + (Input_3: Buffer(Input_2, float32, [(stride*n)], [], type="auto")[(((i + di)*stride) + ((j + dj)*stride_1))]*Filter_3: Buffer(Filter_2, float32, [9], [])[((di*3) + dj)]))
}
}
}
diff --git a/docs/_sources/how_to/work_with_schedules/scan.rst.txt b/docs/_sources/how_to/work_with_schedules/scan.rst.txt
index 8451f4a55e..ce9cfa9eff 100644
--- a/docs/_sources/how_to/work_with_schedules/scan.rst.txt
+++ b/docs/_sources/how_to/work_with_schedules/scan.rst.txt
@@ -120,21 +120,20 @@ To split on the time iteration, user can schedule on scan_op.scan_axis instead.
@main = primfn(X_1: handle, scan_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {X: Buffer(X_2: Pointer(float32), float32, [(stride: int32*m: int32)], [], type="auto"),
- scan: Buffer(scan_2: Pointer(float32), float32, [(stride_1: int32*m)], [], type="auto")}
- buffer_map = {X_1: X, scan_1: scan}
- preflattened_buffer_map = {X_1: X_3: Buffer(X_2, float32, [m, n: int32], [stride, stride_2: int32], type="auto"), scan_1: scan_3: Buffer(scan_2, float32, [m, n], [stride_1, stride_3: int32], type="auto")} {
+ buffers = {X: Buffer(X_2: Pointer(float32), float32, [m: int32, n: int32], [stride: int32, stride_1: int32], type="auto"),
+ scan: Buffer(scan_2: Pointer(float32), float32, [m, n], [stride_2: int32, stride_3: int32], type="auto")}
+ buffer_map = {X_1: X, scan_1: scan} {
attr [IterVar(blockIdx.x: int32, (nullptr), "ThreadIndex", "blockIdx.x")] "thread_extent" = floordiv((n + 255), 256);
attr [IterVar(threadIdx.x: int32, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 256;
if @tir.likely((((blockIdx.x*256) + threadIdx.x) < n), dtype=bool) {
- scan[(((blockIdx.x*256) + threadIdx.x)*stride_3)] = X[(((blockIdx.x*256) + threadIdx.x)*stride_2)]
+ scan_3: Buffer(scan_2, float32, [(stride_2*m)], [], type="auto")[(((blockIdx.x*256) + threadIdx.x)*stride_3)] = X_3: Buffer(X_2, float32, [(stride*m)], [], type="auto")[(((blockIdx.x*256) + threadIdx.x)*stride_1)]
}
for (scan.idx: int32, 0, (m - 1)) {
attr [IterVar(blockIdx.x, (nullptr), "ThreadIndex", "blockIdx.x")] "thread_extent" = floordiv((n + 255), 256);
attr [IterVar(threadIdx.x, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 256;
if @tir.likely((((blockIdx.x*256) + threadIdx.x) < n), dtype=bool) {
let cse_var_1: int32 = (scan.idx + 1)
- scan[((cse_var_1*stride_1) + (((blockIdx.x*256) + threadIdx.x)*stride_3))] = (scan[((scan.idx*stride_1) + (((blockIdx.x*256) + threadIdx.x)*stride_3))] + X[((cse_var_1*stride) + (((blockIdx.x*256) + threadIdx.x)*stride_2))])
+ scan_3[((cse_var_1*stride_2) + (((blockIdx.x*256) + threadIdx.x)*stride_3))] = (scan_3[((scan.idx*stride_2) + (((blockIdx.x*256) + threadIdx.x)*stride_3))] + X_3[((cse_var_1*stride) + (((blockIdx.x*256) + threadIdx.x)*stride_1))])
}
}
}
@@ -231,26 +230,25 @@ the body of scan to be compute_at locations outside the scan loop.
@main = primfn(X_1: handle, scan_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {X: Buffer(X_2: Pointer(float32), float32, [(stride: int32*m: int32)], [], type="auto"),
- scan: Buffer(scan_2: Pointer(float32), float32, [(stride_1: int32*m)], [], type="auto")}
- buffer_map = {X_1: X, scan_1: scan}
- preflattened_buffer_map = {X_1: X_3: Buffer(X_2, float32, [m, n: int32], [stride, stride_2: int32], type="auto"), scan_1: scan_3: Buffer(scan_2, float32, [m, n], [stride_1, stride_3: int32], type="auto")} {
+ buffers = {X: Buffer(X_2: Pointer(float32), float32, [m: int32, n: int32], [stride: int32, stride_1: int32], type="auto"),
+ scan: Buffer(scan_2: Pointer(float32), float32, [m, n], [stride_2: int32, stride_3: int32], type="auto")}
+ buffer_map = {X_1: X, scan_1: scan} {
allocate(s1: Pointer(global float32), float32, [32]), storage_scope = global {
for (i: int32, 0, n) {
- scan[(i*stride_3)] = X[(i*stride_2)]
+ scan_3: Buffer(scan_2, float32, [(stride_2*m)], [], type="auto")[(i*stride_3)] = X_3: Buffer(X_2, float32, [(stride*m)], [], type="auto")[(i*stride_1)]
}
for (scan.idx: int32, 0, (m - 1)) {
for (i.outer: int32, 0, floordiv((n + 31), 32)) {
for (i_1: int32, 0, 32) {
if @tir.likely((((i.outer*32) + i_1) < n), dtype=bool) {
- s1_1: Buffer(s1, float32, [32], [])[i_1] = (scan[((scan.idx*stride_1) + (((i.outer*32) + i_1)*stride_3))]*2f32)
+ s1_1: Buffer(s1, float32, [32], [])[i_1] = (scan_3[((scan.idx*stride_2) + (((i.outer*32) + i_1)*stride_3))]*2f32)
}
}
for (i.inner: int32, 0, 32) {
if @tir.likely((((i.outer*32) + i.inner) < n), dtype=bool) {
let cse_var_2: int32 = (scan.idx + 1)
let cse_var_1: int32 = ((i.outer*32) + i.inner)
- scan[((cse_var_2*stride_1) + (cse_var_1*stride_3))] = (s1_1[i.inner] + X[((cse_var_2*stride) + (cse_var_1*stride_2))])
+ scan_3[((cse_var_2*stride_2) + (cse_var_1*stride_3))] = (s1_1[i.inner] + X_3[((cse_var_2*stride) + (cse_var_1*stride_1))])
}
}
}
@@ -302,24 +300,23 @@ The following example demonstrates how we can build recurrence with two states.
@main = primfn(X_1: handle, scan_2: handle, scan_3: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {X: Buffer(X_2: Pointer(float32), float32, [(stride: int32*m: int32)], [], type="auto"),
- scan: Buffer(scan_4: Pointer(float32), float32, [(stride_1: int32*m)], [], type="auto"),
- scan_1: Buffer(scan_5: Pointer(float32), float32, [(stride_2: int32*m)], [], type="auto")}
- buffer_map = {X_1: X, scan_2: scan, scan_3: scan_1}
- preflattened_buffer_map = {X_1: X_3: Buffer(X_2, float32, [m, n: int32], [stride, stride_3: int32], type="auto"), scan_2: scan_6: Buffer(scan_4, float32, [m, n], [stride_1, stride_4: int32], type="auto"), scan_3: scan_7: Buffer(scan_5, float32, [m, l: int32], [stride_2, stride_5: int32], type="auto")} {
+ buffers = {X: Buffer(X_2: Pointer(float32), float32, [m: int32, n: int32], [stride: int32, stride_1: int32], type="auto"),
+ scan: Buffer(scan_4: Pointer(float32), float32, [m, n], [stride_2: int32, stride_3: int32], type="auto"),
+ scan_1: Buffer(scan_5: Pointer(float32), float32, [m, l: int32], [stride_4: int32, stride_5: int32], type="auto")}
+ buffer_map = {X_1: X, scan_2: scan, scan_3: scan_1} {
for (i: int32, 0, n) {
- scan[(i*stride_4)] = X[(i*stride_3)]
+ scan_6: Buffer(scan_4, float32, [(stride_2*m)], [], type="auto")[(i*stride_3)] = X_3: Buffer(X_2, float32, [(stride*m)], [], type="auto")[(i*stride_1)]
}
for (i_1: int32, 0, l) {
- scan_1[(i_1*stride_5)] = 0f32
+ scan_7: Buffer(scan_5, float32, [(stride_4*m)], [], type="auto")[(i_1*stride_5)] = 0f32
}
for (scan.idx: int32, 0, (m - 1)) {
for (i_2: int32, 0, n) {
let cse_var_1: int32 = (scan.idx + 1)
- scan[((cse_var_1*stride_1) + (i_2*stride_4))] = (scan[((scan.idx*stride_1) + (i_2*stride_4))] + X[((cse_var_1*stride) + (i_2*stride_3))])
+ scan_6[((cse_var_1*stride_2) + (i_2*stride_3))] = (scan_6[((scan.idx*stride_2) + (i_2*stride_3))] + X_3[((cse_var_1*stride) + (i_2*stride_1))])
}
for (i_3: int32, 0, l) {
- scan_1[(((scan.idx + 1)*stride_2) + (i_3*stride_5))] = (scan_1[((scan.idx*stride_2) + (i_3*stride_5))] + scan[(scan.idx*stride_1)])
+ scan_7[(((scan.idx + 1)*stride_4) + (i_3*stride_5))] = (scan_7[((scan.idx*stride_4) + (i_3*stride_5))] + scan_6[(scan.idx*stride_2)])
}
}
}
diff --git a/docs/_sources/how_to/work_with_schedules/schedule_primitives.rst.txt b/docs/_sources/how_to/work_with_schedules/schedule_primitives.rst.txt
index ac9f23ab6a..ed0bd2eecc 100644
--- a/docs/_sources/how_to/work_with_schedules/schedule_primitives.rst.txt
+++ b/docs/_sources/how_to/work_with_schedules/schedule_primitives.rst.txt
@@ -106,14 +106,13 @@ schedule computes tensor in a serial manner in a row-major order.
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [(stride: int32*m: int32)], [], type="auto"),
- B: Buffer(B_2: Pointer(float32), float32, [(stride_1: int32*m)], [], type="auto"),
- C: Buffer(C_2: Pointer(float32), float32, [(stride_2: int32*m)], [], type="auto")}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [m, n: int32], [stride, stride_3: int32], type="auto"), B_1: B_3: Buffer(B_2, float32, [m, n], [stride_1, stride_4: int32], type="auto"), C_1: C_3: Buffer(C_2, float32, [m, n], [stride_2, stride_5: int32], type="auto")} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [m: int32, n: int32], [stride: int32, stride_1: int32], type="auto"),
+ B: Buffer(B_2: Pointer(float32), float32, [m, n], [stride_2: int32, stride_3: int32], type="auto"),
+ C: Buffer(C_2: Pointer(float32), float32, [m, n], [stride_4: int32, stride_5: int32], type="auto")}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
for (i: int32, 0, m) {
for (j: int32, 0, n) {
- C[((i*stride_2) + (j*stride_5))] = (A[((i*stride) + (j*stride_3))]*B[((i*stride_1) + (j*stride_4))])
+ C_3: Buffer(C_2, float32, [(stride_4*m)], [], type="auto")[((i*stride_4) + (j*stride_5))] = (A_3: Buffer(A_2, float32, [(stride*m)], [], type="auto")[((i*stride) + (j*stride_1))]*B_3: Buffer(B_2, float32, [(stride_2*m)], [], type="auto")[((i*stride_2) + (j*stride_3))])
}
}
}
@@ -157,15 +156,14 @@ split
@main = primfn(A_1: handle, B_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [(stride: int32*m: int32)], [], type="auto"),
- B: Buffer(B_2: Pointer(float32), float32, [(stride_1: int32*m)], [], type="auto")}
- buffer_map = {A_1: A, B_1: B}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [m], [stride], type="auto"), B_1: B_3: Buffer(B_2, float32, [m], [stride_1], type="auto")} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [m: int32], [stride: int32], type="auto"),
+ B: Buffer(B_2: Pointer(float32), float32, [m], [stride_1: int32], type="auto")}
+ buffer_map = {A_1: A, B_1: B} {
for (i.outer: int32, 0, floordiv((m + 31), 32)) {
for (i.inner: int32, 0, 32) {
if @tir.likely((((i.outer*32) + i.inner) < m), dtype=bool) {
let cse_var_1: int32 = ((i.outer*32) + i.inner)
- B[(cse_var_1*stride_1)] = (A[(cse_var_1*stride)]*2f32)
+ B_3: Buffer(B_2, float32, [(stride_1*m)], [], type="auto")[(cse_var_1*stride_1)] = (A_3: Buffer(A_2, float32, [(stride*m)], [], type="auto")[(cse_var_1*stride)]*2f32)
}
}
}
@@ -202,14 +200,13 @@ contrary with :code:`factor`.
@main = primfn(A_1: handle, B_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [(stride: int32*m: int32)], [], type="auto"),
- B: Buffer(B_2: Pointer(float32), float32, [(stride_1: int32*m)], [], type="auto")}
- buffer_map = {A_1: A, B_1: B}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [m], [stride], type="auto"), B_1: B_3: Buffer(B_2, float32, [m], [stride_1], type="auto")} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [m: int32], [stride: int32], type="auto"),
+ B: Buffer(B_2: Pointer(float32), float32, [m], [stride_1: int32], type="auto")}
+ buffer_map = {A_1: A, B_1: B} {
for (i.outer: int32, 0, 32) {
for (i.inner: int32, 0, floordiv((m + 31), 32)) {
if @tir.likely(((i.inner + (i.outer*floordiv((m + 31), 32))) < m), dtype=bool) {
- B[((i.inner + (i.outer*floordiv((m + 31), 32)))*stride_1)] = A[((i.inner + (i.outer*floordiv((m + 31), 32)))*stride)]
+ B_3: Buffer(B_2, float32, [(stride_1*m)], [], type="auto")[((i.inner + (i.outer*floordiv((m + 31), 32)))*stride_1)] = A_3: Buffer(A_2, float32, [(stride*m)], [], type="auto")[((i.inner + (i.outer*floordiv((m + 31), 32)))*stride)]
}
}
}
@@ -248,10 +245,9 @@ axes.
@main = primfn(A_1: handle, B_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [(stride: int32*m: int32)], [], type="auto"),
- B: Buffer(B_2: Pointer(float32), float32, [(stride_1: int32*m)], [], type="auto")}
- buffer_map = {A_1: A, B_1: B}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [m, n: int32], [stride, stride_2: int32], type="auto"), B_1: B_3: Buffer(B_2, float32, [m, n], [stride_1, stride_3: int32], type="auto")} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [m: int32, n: int32], [stride: int32, stride_1: int32], type="auto"),
+ B: Buffer(B_2: Pointer(float32), float32, [m, n], [stride_2: int32, stride_3: int32], type="auto")}
+ buffer_map = {A_1: A, B_1: B} {
for (i.outer: int32, 0, floordiv((m + 9), 10)) {
for (j.outer: int32, 0, floordiv((n + 4), 5)) {
for (i.inner: int32, 0, 10) {
@@ -260,7 +256,7 @@ axes.
if @tir.likely((((j.outer*5) + j.inner) < n), dtype=bool) {
let cse_var_2: int32 = ((j.outer*5) + j.inner)
let cse_var_1: int32 = ((i.outer*10) + i.inner)
- B[((cse_var_1*stride_1) + (cse_var_2*stride_3))] = A[((cse_var_1*stride) + (cse_var_2*stride_2))]
+ B_3: Buffer(B_2, float32, [(stride_2*m)], [], type="auto")[((cse_var_1*stride_2) + (cse_var_2*stride_3))] = A_3: Buffer(A_2, float32, [(stride*m)], [], type="auto")[((cse_var_1*stride) + (cse_var_2*stride_1))]
}
}
}
@@ -304,10 +300,9 @@ fuse
@main = primfn(A_1: handle, B_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [(stride: int32*m: int32)], [], type="auto"),
- B: Buffer(B_2: Pointer(float32), float32, [(stride_1: int32*m)], [], type="auto")}
- buffer_map = {A_1: A, B_1: B}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [m, n: int32], [stride, stride_2: int32], type="auto"), B_1: B_3: Buffer(B_2, float32, [m, n], [stride_1, stride_3: int32], type="auto")} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [m: int32, n: int32], [stride: int32, stride_1: int32], type="auto"),
+ B: Buffer(B_2: Pointer(float32), float32, [m, n], [stride_2: int32, stride_3: int32], type="auto")}
+ buffer_map = {A_1: A, B_1: B} {
for (i.outer: int32, 0, floordiv((m + 9), 10)) {
for (j.outer: int32, 0, floordiv((n + 4), 5)) {
for (i.inner.j.inner.fused: int32, 0, 50) {
@@ -315,7 +310,7 @@ fuse
if @tir.likely((((j.outer*5) + floormod(i.inner.j.inner.fused, 5)) < n), dtype=bool) {
let cse_var_2: int32 = ((j.outer*5) + floormod(i.inner.j.inner.fused, 5))
let cse_var_1: int32 = ((i.outer*10) + floordiv(i.inner.j.inner.fused, 5))
- B[((cse_var_1*stride_1) + (cse_var_2*stride_3))] = A[((cse_var_1*stride) + (cse_var_2*stride_2))]
+ B_3: Buffer(B_2, float32, [(stride_2*m)], [], type="auto")[((cse_var_1*stride_2) + (cse_var_2*stride_3))] = A_3: Buffer(A_2, float32, [(stride*m)], [], type="auto")[((cse_var_1*stride) + (cse_var_2*stride_1))]
}
}
}
@@ -358,10 +353,9 @@ reorder
@main = primfn(A_1: handle, B_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [(stride: int32*m: int32)], [], type="auto"),
- B: Buffer(B_2: Pointer(float32), float32, [(stride_1: int32*m)], [], type="auto")}
- buffer_map = {A_1: A, B_1: B}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [m, n: int32], [stride, stride_2: int32], type="auto"), B_1: B_3: Buffer(B_2, float32, [m, n], [stride_1, stride_3: int32], type="auto")} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [m: int32, n: int32], [stride: int32, stride_1: int32], type="auto"),
+ B: Buffer(B_2: Pointer(float32), float32, [m, n], [stride_2: int32, stride_3: int32], type="auto")}
+ buffer_map = {A_1: A, B_1: B} {
for (i.inner: int32, 0, 10) {
for (j.outer: int32, 0, floordiv((n + 4), 5)) {
for (i.outer: int32, 0, floordiv((m + 9), 10)) {
@@ -370,7 +364,7 @@ reorder
if @tir.likely((((j.outer*5) + j.inner) < n), dtype=bool) {
let cse_var_2: int32 = ((j.outer*5) + j.inner)
let cse_var_1: int32 = ((i.outer*10) + i.inner)
- B[((cse_var_1*stride_1) + (cse_var_2*stride_3))] = A[((cse_var_1*stride) + (cse_var_2*stride_2))]
+ B_3: Buffer(B_2, float32, [(stride_2*m)], [], type="auto")[((cse_var_1*stride_2) + (cse_var_2*stride_3))] = A_3: Buffer(A_2, float32, [(stride*m)], [], type="auto")[((cse_var_1*stride) + (cse_var_2*stride_1))]
}
}
}
@@ -414,14 +408,13 @@ in gpu programming.
@main = primfn(A_1: handle, B_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [(stride: int32*n: int32)], [], type="auto"),
- B: Buffer(B_2: Pointer(float32), float32, [(stride_1: int32*n)], [], type="auto")}
- buffer_map = {A_1: A, B_1: B}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [n], [stride], type="auto"), B_1: B_3: Buffer(B_2, float32, [n], [stride_1], type="auto")} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [n: int32], [stride: int32], type="auto"),
+ B: Buffer(B_2: Pointer(float32), float32, [n], [stride_1: int32], type="auto")}
+ buffer_map = {A_1: A, B_1: B} {
attr [IterVar(blockIdx.x: int32, (nullptr), "ThreadIndex", "blockIdx.x")] "thread_extent" = floordiv((n + 63), 64);
attr [IterVar(threadIdx.x: int32, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
if @tir.likely((((blockIdx.x*64) + threadIdx.x) < n), dtype=bool) {
- B[(((blockIdx.x*64) + threadIdx.x)*stride_1)] = (A[(((blockIdx.x*64) + threadIdx.x)*stride)]*2f32)
+ B_3: Buffer(B_2, float32, [(stride_1*n)], [], type="auto")[(((blockIdx.x*64) + threadIdx.x)*stride_1)] = (A_3: Buffer(A_2, float32, [(stride*n)], [], type="auto")[(((blockIdx.x*64) + threadIdx.x)*stride)]*2f32)
}
}
@@ -458,16 +451,15 @@ tensors at the root separately by default.
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [(stride: int32*m: int32)], [], type="auto"),
- B: Buffer(B_2: Pointer(float32), float32, [(stride_1: int32*m)], [], type="auto"),
- C: Buffer(C_2: Pointer(float32), float32, [(stride_2: int32*m)], [], type="auto")}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [m], [stride], type="auto"), B_1: B_3: Buffer(B_2, float32, [m], [stride_1], type="auto"), C_1: C_3: Buffer(C_2, float32, [m], [stride_2], type="auto")} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [m: int32], [stride: int32], type="auto"),
+ B: Buffer(B_2: Pointer(float32), float32, [m], [stride_1: int32], type="auto"),
+ C: Buffer(C_2: Pointer(float32), float32, [m], [stride_2: int32], type="auto")}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
for (i: int32, 0, m) {
- B[(i*stride_1)] = (A[(i*stride)] + 1f32)
+ B_3: Buffer(B_2, float32, [(stride_1*m)], [], type="auto")[(i*stride_1)] = (A_3: Buffer(A_2, float32, [(stride*m)], [], type="auto")[(i*stride)] + 1f32)
}
for (i_1: int32, 0, m) {
- C[(i_1*stride_2)] = (B[(i_1*stride_1)]*2f32)
+ C_3: Buffer(C_2, float32, [(stride_2*m)], [], type="auto")[(i_1*stride_2)] = (B_3[(i_1*stride_1)]*2f32)
}
}
@@ -503,14 +495,13 @@ of computation of `C`.
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [(stride: int32*m: int32)], [], type="auto"),
- B: Buffer(B_2: Pointer(float32), float32, [(stride_1: int32*m)], [], type="auto"),
- C: Buffer(C_2: Pointer(float32), float32, [(stride_2: int32*m)], [], type="auto")}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [m], [stride], type="auto"), B_1: B_3: Buffer(B_2, float32, [m], [stride_1], type="auto"), C_1: C_3: Buffer(C_2, float32, [m], [stride_2], type="auto")} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [m: int32], [stride: int32], type="auto"),
+ B: Buffer(B_2: Pointer(float32), float32, [m], [stride_1: int32], type="auto"),
+ C: Buffer(C_2: Pointer(float32), float32, [m], [stride_2: int32], type="auto")}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
for (i: int32, 0, m) {
- B[(i*stride_1)] = (A[(i*stride)] + 1f32)
- C[(i*stride_2)] = (B[(i*stride_1)]*2f32)
+ B_3: Buffer(B_2, float32, [(stride_1*m)], [], type="auto")[(i*stride_1)] = (A_3: Buffer(A_2, float32, [(stride*m)], [], type="auto")[(i*stride)] + 1f32)
+ C_3: Buffer(C_2, float32, [(stride_2*m)], [], type="auto")[(i*stride_2)] = (B_3[(i*stride_1)]*2f32)
}
}
@@ -549,13 +540,12 @@ tensor is required.
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [(stride: int32*m: int32)], [], type="auto"),
- B: Buffer(B_2: Pointer(float32), float32, [(stride_1: int32*m)], [], type="auto"),
- C: Buffer(C_2: Pointer(float32), float32, [(stride_2: int32*m)], [], type="auto")}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [m], [stride], type="auto"), B_1: B_3: Buffer(B_2, float32, [m], [stride_1], type="auto"), C_1: C_3: Buffer(C_2, float32, [m], [stride_2], type="auto")} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [m: int32], [stride: int32], type="auto"),
+ B: Buffer(B_2: Pointer(float32), float32, [m], [stride_1: int32], type="auto"),
+ C: Buffer(C_2: Pointer(float32), float32, [m], [stride_2: int32], type="auto")}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
for (i: int32, 0, m) {
- C[(i*stride_2)] = ((A[(i*stride)] + 1f32)*2f32)
+ C_3: Buffer(C_2, float32, [(stride_2*m)], [], type="auto")[(i*stride_2)] = ((A_3: Buffer(A_2, float32, [(stride*m)], [], type="auto")[(i*stride)] + 1f32)*2f32)
}
}
@@ -593,16 +583,15 @@ compute_root
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [(stride: int32*m: int32)], [], type="auto"),
- B: Buffer(B_2: Pointer(float32), float32, [(stride_1: int32*m)], [], type="auto"),
- C: Buffer(C_2: Pointer(float32), float32, [(stride_2: int32*m)], [], type="auto")}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [m], [stride], type="auto"), B_1: B_3: Buffer(B_2, float32, [m], [stride_1], type="auto"), C_1: C_3: Buffer(C_2, float32, [m], [stride_2], type="auto")} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [m: int32], [stride: int32], type="auto"),
+ B: Buffer(B_2: Pointer(float32), float32, [m], [stride_1: int32], type="auto"),
+ C: Buffer(C_2: Pointer(float32), float32, [m], [stride_2: int32], type="auto")}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
for (i: int32, 0, m) {
- B[(i*stride_1)] = (A[(i*stride)] + 1f32)
+ B_3: Buffer(B_2, float32, [(stride_1*m)], [], type="auto")[(i*stride_1)] = (A_3: Buffer(A_2, float32, [(stride*m)], [], type="auto")[(i*stride)] + 1f32)
}
for (i_1: int32, 0, m) {
- C[(i_1*stride_2)] = (B[(i_1*stride_1)]*2f32)
+ C_3: Buffer(C_2, float32, [(stride_2*m)], [], type="auto")[(i_1*stride_2)] = (B_3[(i_1*stride_1)]*2f32)
}
}
diff --git a/docs/_sources/how_to/work_with_schedules/sg_execution_times.rst.txt b/docs/_sources/how_to/work_with_schedules/sg_execution_times.rst.txt
index 3056f82a83..ff6659b4d2 100644
--- a/docs/_sources/how_to/work_with_schedules/sg_execution_times.rst.txt
+++ b/docs/_sources/how_to/work_with_schedules/sg_execution_times.rst.txt
@@ -5,22 +5,22 @@
Computation times
=================
-**00:04.605** total execution time for **how_to_work_with_schedules** files:
+**00:07.650** total execution time for **how_to_work_with_schedules** files:
+------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_schedules_intrin_math.py` (``intrin_math.py``) | 00:02.259 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_schedules_intrin_math.py` (``intrin_math.py``) | 00:05.341 | 0.0 MB |
+------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_schedules_tensorize.py` (``tensorize.py``) | 00:01.034 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_schedules_tensorize.py` (``tensorize.py``) | 00:00.986 | 0.0 MB |
+------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_schedules_reduction.py` (``reduction.py``) | 00:00.558 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_schedules_reduction.py` (``reduction.py``) | 00:00.555 | 0.0 MB |
+------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_schedules_scan.py` (``scan.py``) | 00:00.544 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_schedules_scan.py` (``scan.py``) | 00:00.542 | 0.0 MB |
+------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_schedules_extern_op.py` (``extern_op.py``) | 00:00.115 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_schedules_extern_op.py` (``extern_op.py``) | 00:00.117 | 0.0 MB |
+------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_schedules_schedule_primitives.py` (``schedule_primitives.py``) | 00:00.049 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_schedules_schedule_primitives.py` (``schedule_primitives.py``) | 00:00.055 | 0.0 MB |
+------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_schedules_tedd.py` (``tedd.py``) | 00:00.029 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_schedules_tedd.py` (``tedd.py``) | 00:00.033 | 0.0 MB |
+------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_schedules_tuple_inputs.py` (``tuple_inputs.py``) | 00:00.019 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_schedules_tuple_inputs.py` (``tuple_inputs.py``) | 00:00.021 | 0.0 MB |
+------------------------------------------------------------------------------------------------+-----------+--------+
diff --git a/docs/_sources/how_to/work_with_schedules/tensorize.rst.txt b/docs/_sources/how_to/work_with_schedules/tensorize.rst.txt
index 91484392d3..13908b1918 100644
--- a/docs/_sources/how_to/work_with_schedules/tensorize.rst.txt
+++ b/docs/_sources/how_to/work_with_schedules/tensorize.rst.txt
@@ -85,17 +85,16 @@ The following lines describe the computation :code:`A * B^T` in TVM.
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [65536], []),
- B: Buffer(B_2: Pointer(float32), float32, [32768], []),
- C: Buffer(C_2: Pointer(float32), float32, [524288], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 64], []), B_1: B_3: Buffer(B_2, float32, [512, 64], []), C_1: C_3: Buffer(C_2, float32, [1024, 512], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 64], []),
+ B: Buffer(B_2: Pointer(float32), float32, [512, 64], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 512], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
for (i: int32, 0, 1024) {
for (j: int32, 0, 512) {
- C[((i*512) + j)] = 0f32
+ C_3: Buffer(C_2, float32, [524288], [])[((i*512) + j)] = 0f32
for (k: int32, 0, 64) {
let cse_var_1: int32 = ((i*512) + j)
- C[cse_var_1] = (C[cse_var_1] + (A[((i*64) + k)]*B[((j*64) + k)]))
+ C_3[cse_var_1] = (C_3[cse_var_1] + (A_3: Buffer(A_2, float32, [65536], [])[((i*64) + k)]*B_3: Buffer(B_2, float32, [32768], [])[((j*64) + k)]))
}
}
}
@@ -138,18 +137,17 @@ Thus we break down the matmul loops to make the innermost loops a (16x64) GEMV.
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [65536], []),
- B: Buffer(B_2: Pointer(float32), float32, [32768], []),
- C: Buffer(C_2: Pointer(float32), float32, [524288], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 64], []), B_1: B_3: Buffer(B_2, float32, [512, 64], []), C_1: C_3: Buffer(C_2, float32, [1024, 512], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 64], []),
+ B: Buffer(B_2: Pointer(float32), float32, [512, 64], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 512], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
for (i: int32, 0, 1024) {
for (j.outer: int32, 0, 32) {
for (j.inner: int32, 0, 16) {
- C[(((i*512) + (j.outer*16)) + j.inner)] = 0f32
+ C_3: Buffer(C_2, float32, [524288], [])[(((i*512) + (j.outer*16)) + j.inner)] = 0f32
for (k: int32, 0, 64) {
let cse_var_1: int32 = (((i*512) + (j.outer*16)) + j.inner)
- C[cse_var_1] = (C[cse_var_1] + (A[((i*64) + k)]*B[(((j.outer*1024) + (j.inner*64)) + k)]))
+ C_3[cse_var_1] = (C_3[cse_var_1] + (A_3: Buffer(A_2, float32, [65536], [])[((i*64) + k)]*B_3: Buffer(B_2, float32, [32768], [])[(((j.outer*1024) + (j.inner*64)) + k)]))
}
}
}
@@ -263,11 +261,10 @@ such placeholder can be put to let TVM automatically bind the inferred value for
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [65536], []),
- B: Buffer(B_2: Pointer(float32), float32, [32768], []),
- C: Buffer(C_2: Pointer(float32), float32, [524288], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 64], []), B_1: B_3: Buffer(B_2, float32, [512, 64], []), C_1: C_3: Buffer(C_2, float32, [1024, 512], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 64], []),
+ B: Buffer(B_2: Pointer(float32), float32, [512, 64], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 512], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
for (i: int32, 0, 1024) {
for (j.outer: int32, 0, 32) {
@tir.call_extern("gemv_update", @tir.tvm_access_ptr(@tir.type_annotation(, dtype=float32), C_2, ((i*512) + (j.outer*16)), 16, 2, dtype=handle), @tir.tvm_access_ptr(@tir.type_annotation(, dtype=float32), A_2, (i*64), 64, 1, dtype=handle), @tir.tvm_access_ptr(@tir.type_annotation(, dtype=float32), B_2, (j.outer*1024), 1024, 1, dtype=handle), 16, 64, 64, dtype=int32)
@@ -342,12 +339,11 @@ The importing needs to happen before the tensorized GEMV being executed.
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [65536], []),
- B: Buffer(B_2: Pointer(float32), float32, [32768], []),
- C: Buffer(C_2: Pointer(float32), float32, [524288], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 64], []), B_1: B_3: Buffer(B_2, float32, [512, 64], []), C_1: C_3: Buffer(C_2, float32, [1024, 512], [])} {
- attr [IterVar(i: int32, (nullptr), "DataPar", "")] "pragma_import_llvm" = "; ModuleID = '/tmp/tmpg08za814/input0.cc'\nsource_filename = \"/tmp/tmpg08za814/input0.cc\"\ntarget datalayout = \"e-m:e-i64:64-f80:128-n8:16:32:64-S128\"\ntarget triple = \"x86_64-pc-linux-gnu\"\n\n; Function Attrs: noinline nounwind optnone uwtable\ndefine dso_local i32 @gemv_update(float*, float*, float*, i32, i32, i32) #0 {\n %7 = alloca float*, align 8\n %8 = alloca float*, align 8\n %9 = alloca floa [...]
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 64], []),
+ B: Buffer(B_2: Pointer(float32), float32, [512, 64], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 512], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
+ attr [IterVar(i: int32, (nullptr), "DataPar", "")] "pragma_import_llvm" = "; ModuleID = '/tmp/tmp9m2vzx82/input0.cc'\nsource_filename = \"/tmp/tmp9m2vzx82/input0.cc\"\ntarget datalayout = \"e-m:e-i64:64-f80:128-n8:16:32:64-S128\"\ntarget triple = \"x86_64-pc-linux-gnu\"\n\n; Function Attrs: noinline nounwind optnone uwtable\ndefine dso_local i32 @gemv_update(float*, float*, float*, i32, i32, i32) #0 {\n %7 = alloca float*, align 8\n %8 = alloca float*, align 8\n %9 = alloca floa [...]
for (i, 0, 1024) {
for (j.outer: int32, 0, 32) {
@tir.call_extern("gemv_update", @tir.tvm_access_ptr(@tir.type_annotation(, dtype=float32), C_2, ((i*512) + (j.outer*16)), 16, 2, dtype=handle), @tir.tvm_access_ptr(@tir.type_annotation(, dtype=float32), A_2, (i*64), 64, 1, dtype=handle), @tir.tvm_access_ptr(@tir.type_annotation(, dtype=float32), B_2, (j.outer*1024), 1024, 1, dtype=handle), 16, 64, 64, dtype=int32)
diff --git a/docs/_sources/how_to/work_with_schedules/tuple_inputs.rst.txt b/docs/_sources/how_to/work_with_schedules/tuple_inputs.rst.txt
index ed3334c46a..fd9d42507f 100644
--- a/docs/_sources/how_to/work_with_schedules/tuple_inputs.rst.txt
+++ b/docs/_sources/how_to/work_with_schedules/tuple_inputs.rst.txt
@@ -79,16 +79,15 @@ together in the next schedule procedure.
@main = primfn(A0_1: handle, A1_1: handle, B_2: handle, B_3: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A0: Buffer(A0_2: Pointer(float32), float32, [(stride: int32*m: int32)], [], type="auto"),
- A1: Buffer(A1_2: Pointer(float32), float32, [(stride_1: int32*m)], [], type="auto"),
- B: Buffer(B_4: Pointer(float32), float32, [(stride_2: int32*m)], [], type="auto"),
- B_1: Buffer(B_5: Pointer(float32), float32, [(stride_3: int32*m)], [], type="auto")}
- buffer_map = {A0_1: A0, A1_1: A1, B_2: B, B_3: B_1}
- preflattened_buffer_map = {A0_1: A0_3: Buffer(A0_2, float32, [m, n: int32], [stride, stride_4: int32], type="auto"), A1_1: A1_3: Buffer(A1_2, float32, [m, n], [stride_1, stride_5: int32], type="auto"), B_2: B_6: Buffer(B_4, float32, [m, n], [stride_2, stride_6: int32], type="auto"), B_3: B_7: Buffer(B_5, float32, [m, n], [stride_3, stride_7: int32], type="auto")} {
+ buffers = {A0: Buffer(A0_2: Pointer(float32), float32, [m: int32, n: int32], [stride: int32, stride_1: int32], type="auto"),
+ A1: Buffer(A1_2: Pointer(float32), float32, [m, n], [stride_2: int32, stride_3: int32], type="auto"),
+ B: Buffer(B_4: Pointer(float32), float32, [m, n], [stride_4: int32, stride_5: int32], type="auto"),
+ B_1: Buffer(B_5: Pointer(float32), float32, [m, n], [stride_6: int32, stride_7: int32], type="auto")}
+ buffer_map = {A0_1: A0, A1_1: A1, B_2: B, B_3: B_1} {
for (i: int32, 0, m) {
for (j: int32, 0, n) {
- B[((i*stride_2) + (j*stride_6))] = (A0[((i*stride) + (j*stride_4))] + 2f32)
- B_1[((i*stride_3) + (j*stride_7))] = (A1[((i*stride_1) + (j*stride_5))]*3f32)
+ B_6: Buffer(B_4, float32, [(stride_4*m)], [], type="auto")[((i*stride_4) + (j*stride_5))] = (A0_3: Buffer(A0_2, float32, [(stride*m)], [], type="auto")[((i*stride) + (j*stride_1))] + 2f32)
+ B_7: Buffer(B_5, float32, [(stride_6*m)], [], type="auto")[((i*stride_6) + (j*stride_7))] = (A1_3: Buffer(A1_2, float32, [(stride_2*m)], [], type="auto")[((i*stride_2) + (j*stride_3))]*3f32)
}
}
}
@@ -153,18 +152,17 @@ with :py:func:`te.comm_reducer` as below:
@main = primfn(idx_1: handle, val_1: handle, T_2: handle, T_3: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {idx: Buffer(idx_2: Pointer(int32), int32, [(stride: int32*m: int32)], [], type="auto"),
- val: Buffer(val_2: Pointer(int32), int32, [(stride_1: int32*m)], [], type="auto"),
- T: Buffer(T_4: Pointer(int32), int32, [(stride_2: int32*m)], [], type="auto"),
- T_1: Buffer(T_5: Pointer(int32), int32, [(stride_3: int32*m)], [], type="auto")}
- buffer_map = {idx_1: idx, val_1: val, T_2: T, T_3: T_1}
- preflattened_buffer_map = {idx_1: idx_3: Buffer(idx_2, int32, [m, n: int32], [stride, stride_4: int32], type="auto"), val_1: val_3: Buffer(val_2, int32, [m, n], [stride_1, stride_5: int32], type="auto"), T_2: T_6: Buffer(T_4, int32, [m], [stride_2], type="auto"), T_3: T_7: Buffer(T_5, int32, [m], [stride_3], type="auto")} {
+ buffers = {idx: Buffer(idx_2: Pointer(int32), int32, [m: int32, n: int32], [stride: int32, stride_1: int32], type="auto"),
+ val: Buffer(val_2: Pointer(int32), int32, [m, n], [stride_2: int32, stride_3: int32], type="auto"),
+ T: Buffer(T_4: Pointer(int32), int32, [m], [stride_4: int32], type="auto"),
+ T_1: Buffer(T_5: Pointer(int32), int32, [m], [stride_5: int32], type="auto")}
+ buffer_map = {idx_1: idx, val_1: val, T_2: T, T_3: T_1} {
for (i: int32, 0, m) {
- T[(i*stride_2)] = -1
- T_1[(i*stride_3)] = -2147483648
+ T_6: Buffer(T_4, int32, [(stride_4*m)], [], type="auto")[(i*stride_4)] = -1
+ T_7: Buffer(T_5, int32, [(stride_5*m)], [], type="auto")[(i*stride_5)] = -2147483648
for (k: int32, 0, n) {
- T[(i*stride_2)] = @tir.if_then_else((val[((i*stride_1) + (k*stride_5))] <= T_1[(i*stride_3)]), T[(i*stride_2)], idx[((i*stride) + (k*stride_4))], dtype=int32)
- T_1[(i*stride_3)] = @tir.if_then_else((val[((i*stride_1) + (k*stride_5))] <= T_1[(i*stride_3)]), T_1[(i*stride_3)], val[((i*stride_1) + (k*stride_5))], dtype=int32)
+ T_6[(i*stride_4)] = @tir.if_then_else((val_3: Buffer(val_2, int32, [(stride_2*m)], [], type="auto")[((i*stride_2) + (k*stride_3))] <= T_7[(i*stride_5)]), T_6[(i*stride_4)], idx_3: Buffer(idx_2, int32, [(stride*m)], [], type="auto")[((i*stride) + (k*stride_1))], dtype=int32)
+ T_7[(i*stride_5)] = @tir.if_then_else((val_3[((i*stride_2) + (k*stride_3))] <= T_7[(i*stride_5)]), T_7[(i*stride_5)], val_3[((i*stride_2) + (k*stride_3))], dtype=int32)
}
}
}
@@ -216,20 +214,19 @@ in terms of operation.
@main = primfn(A0_1: handle, A1_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A0: Buffer(A0_2: Pointer(float32), float32, [(stride: int32*m: int32)], [], type="auto"),
- A1: Buffer(A1_2: Pointer(float32), float32, [(stride_1: int32*m)], [], type="auto"),
- C: Buffer(C_2: Pointer(float32), float32, [(stride_2: int32*m)], [], type="auto")}
- buffer_map = {A0_1: A0, A1_1: A1, C_1: C}
- preflattened_buffer_map = {A0_1: A0_3: Buffer(A0_2, float32, [m, n: int32], [stride, stride_3: int32], type="auto"), A1_1: A1_3: Buffer(A1_2, float32, [m, n], [stride_1, stride_4: int32], type="auto"), C_1: C_3: Buffer(C_2, float32, [m, n], [stride_2, stride_5: int32], type="auto")} {
+ buffers = {A0: Buffer(A0_2: Pointer(float32), float32, [m: int32, n: int32], [stride: int32, stride_1: int32], type="auto"),
+ A1: Buffer(A1_2: Pointer(float32), float32, [m, n], [stride_2: int32, stride_3: int32], type="auto"),
+ C: Buffer(C_2: Pointer(float32), float32, [m, n], [stride_4: int32, stride_5: int32], type="auto")}
+ buffer_map = {A0_1: A0, A1_1: A1, C_1: C} {
allocate(B.v0: Pointer(global float32), float32, [n]), storage_scope = global;
allocate(B.v1: Pointer(global float32), float32, [n]), storage_scope = global;
for (i: int32, 0, m) {
for (j: int32, 0, n) {
- B.v0_1: Buffer(B.v0, float32, [n], [])[j] = (A0[((i*stride) + (j*stride_3))] + 2f32)
- B.v1_1: Buffer(B.v1, float32, [n], [])[j] = (A0[((i*stride) + (j*stride_3))]*3f32)
+ B.v0_1: Buffer(B.v0, float32, [n], [])[j] = (A0_3: Buffer(A0_2, float32, [(stride*m)], [], type="auto")[((i*stride) + (j*stride_1))] + 2f32)
+ B.v1_1: Buffer(B.v1, float32, [n], [])[j] = (A0_3[((i*stride) + (j*stride_1))]*3f32)
}
for (j_1: int32, 0, n) {
- C[((i*stride_2) + (j_1*stride_5))] = (A1[((i*stride_1) + (j_1*stride_4))] + B.v0_1[j_1])
+ C_3: Buffer(C_2, float32, [(stride_4*m)], [], type="auto")[((i*stride_4) + (j_1*stride_5))] = (A1_3: Buffer(A1_2, float32, [(stride_2*m)], [], type="auto")[((i*stride_2) + (j_1*stride_3))] + B.v0_1[j_1])
}
}
}
diff --git a/docs/_sources/topic/vta/tutorials/autotvm/sg_execution_times.rst.txt b/docs/_sources/topic/vta/tutorials/autotvm/sg_execution_times.rst.txt
index 118dc79e7a..e77273ec85 100644
--- a/docs/_sources/topic/vta/tutorials/autotvm/sg_execution_times.rst.txt
+++ b/docs/_sources/topic/vta/tutorials/autotvm/sg_execution_times.rst.txt
@@ -5,10 +5,10 @@
Computation times
=================
-**00:26.278** total execution time for **topic_vta_tutorials_autotvm** files:
+**00:27.455** total execution time for **topic_vta_tutorials_autotvm** files:
+---------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_topic_vta_tutorials_autotvm_tune_relay_vta.py` (``tune_relay_vta.py``) | 00:26.271 | 0.0 MB |
+| :ref:`sphx_glr_topic_vta_tutorials_autotvm_tune_relay_vta.py` (``tune_relay_vta.py``) | 00:27.449 | 0.0 MB |
+---------------------------------------------------------------------------------------+-----------+--------+
| :ref:`sphx_glr_topic_vta_tutorials_autotvm_tune_alu_vta.py` (``tune_alu_vta.py``) | 00:00.006 | 0.0 MB |
+---------------------------------------------------------------------------------------+-----------+--------+
diff --git a/docs/_sources/topic/vta/tutorials/frontend/deploy_classification.rst.txt b/docs/_sources/topic/vta/tutorials/frontend/deploy_classification.rst.txt
index 5f2d3da47c..3d98add499 100644
--- a/docs/_sources/topic/vta/tutorials/frontend/deploy_classification.rst.txt
+++ b/docs/_sources/topic/vta/tutorials/frontend/deploy_classification.rst.txt
@@ -289,7 +289,7 @@ The compilation steps are:
DeprecationWarning,
/workspace/vta/tutorials/frontend/deploy_classification.py:213: DeprecationWarning: legacy graph executor behavior of producing json / lib / params will be removed in the next release. Please see documents of tvm.contrib.graph_executor.GraphModule for the new recommended usage.
relay_prog, target=tvm.target.Target(target, host=env.target_host), params=params
- resnet18_v1 inference graph built in 29.07s!
+ resnet18_v1 inference graph built in 31.70s!
diff --git a/docs/_sources/topic/vta/tutorials/frontend/deploy_detection.rst.txt b/docs/_sources/topic/vta/tutorials/frontend/deploy_detection.rst.txt
index 2e978ce33d..7af79c1f81 100644
--- a/docs/_sources/topic/vta/tutorials/frontend/deploy_detection.rst.txt
+++ b/docs/_sources/topic/vta/tutorials/frontend/deploy_detection.rst.txt
@@ -333,7 +333,7 @@ The compilation steps are:
/workspace/python/tvm/relay/build_module.py:348: DeprecationWarning: Please use input parameter mod (tvm.IRModule) instead of deprecated parameter mod (tvm.relay.function.Function)
DeprecationWarning,
- yolov3-tiny inference graph built in 19.56s!
+ yolov3-tiny inference graph built in 20.16s!
diff --git a/docs/_sources/topic/vta/tutorials/frontend/sg_execution_times.rst.txt b/docs/_sources/topic/vta/tutorials/frontend/sg_execution_times.rst.txt
index 8033253c98..6476c30d66 100644
--- a/docs/_sources/topic/vta/tutorials/frontend/sg_execution_times.rst.txt
+++ b/docs/_sources/topic/vta/tutorials/frontend/sg_execution_times.rst.txt
@@ -5,10 +5,10 @@
Computation times
=================
-**01:41.002** total execution time for **topic_vta_tutorials_frontend** files:
+**01:44.215** total execution time for **topic_vta_tutorials_frontend** files:
+------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_topic_vta_tutorials_frontend_deploy_detection.py` (``deploy_detection.py``) | 00:51.877 | 0.0 MB |
+| :ref:`sphx_glr_topic_vta_tutorials_frontend_deploy_detection.py` (``deploy_detection.py``) | 00:52.496 | 0.0 MB |
+------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_topic_vta_tutorials_frontend_deploy_classification.py` (``deploy_classification.py``) | 00:49.125 | 0.0 MB |
+| :ref:`sphx_glr_topic_vta_tutorials_frontend_deploy_classification.py` (``deploy_classification.py``) | 00:51.719 | 0.0 MB |
+------------------------------------------------------------------------------------------------------+-----------+--------+
diff --git a/docs/_sources/topic/vta/tutorials/matrix_multiply.rst.txt b/docs/_sources/topic/vta/tutorials/matrix_multiply.rst.txt
index dceab9214e..d41ec2f1a9 100644
--- a/docs/_sources/topic/vta/tutorials/matrix_multiply.rst.txt
+++ b/docs/_sources/topic/vta/tutorials/matrix_multiply.rst.txt
@@ -344,18 +344,17 @@ After we construct the schedule, by default the schedule computes
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(int8), int8, [256], []),
- B: Buffer(B_2: Pointer(int8), int8, [65536], []),
- C: Buffer(C_2: Pointer(int8), int8, [256], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, int8, [1, 16, 1, 16], []), B_1: B_3: Buffer(B_2, int8, [16, 16, 16, 16], []), C_1: C_3: Buffer(C_2, int8, [1, 16, 1, 16], [])} {
+ buffers = {A: Buffer(A_2: Pointer(int8), int8, [1, 16, 1, 16], []),
+ B: Buffer(B_2: Pointer(int8), int8, [16, 16, 16, 16], []),
+ C: Buffer(C_2: Pointer(int8), int8, [1, 16, 1, 16], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
allocate(A_buf: Pointer(global int8), int8, [256]), storage_scope = global;
allocate(B_buf: Pointer(global int8), int8, [65536]), storage_scope = global;
allocate(C_buf: Pointer(global int32), int32, [256]), storage_scope = global {
for (i1: int32, 0, 16) {
for (i3: int32, 0, 16) {
let cse_var_1: int32 = ((i1*16) + i3)
- A_buf_1: Buffer(A_buf, int8, [256], [])[cse_var_1] = A[cse_var_1]
+ A_buf_1: Buffer(A_buf, int8, [256], [])[cse_var_1] = A_3: Buffer(A_2, int8, [256], [])[cse_var_1]
}
}
for (i0: int32, 0, 16) {
@@ -363,7 +362,7 @@ After we construct the schedule, by default the schedule computes
for (i2: int32, 0, 16) {
for (i3_1: int32, 0, 16) {
let cse_var_2: int32 = ((((i0*4096) + (i1_1*256)) + (i2*16)) + i3_1)
- B_buf_1: Buffer(B_buf, int8, [65536], [])[cse_var_2] = B[cse_var_2]
+ B_buf_1: Buffer(B_buf, int8, [65536], [])[cse_var_2] = B_3: Buffer(B_2, int8, [65536], [])[cse_var_2]
}
}
}
@@ -382,7 +381,7 @@ After we construct the schedule, by default the schedule computes
for (i1_2: int32, 0, 16) {
for (i3_2: int32, 0, 16) {
let cse_var_4: int32 = ((i1_2*16) + i3_2)
- C[cse_var_4] = cast(int8, C_buf_1[cse_var_4])
+ C_3: Buffer(C_2, int8, [256], [])[cse_var_4] = cast(int8, C_buf_1[cse_var_4])
}
}
}
@@ -508,11 +507,10 @@ moving the copy operations into the matrix multiplication loop.
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(int8), int8, [256], []),
- B: Buffer(B_2: Pointer(int8), int8, [65536], []),
- C: Buffer(C_2: Pointer(int8), int8, [256], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, int8, [1, 16, 1, 16], []), B_1: B_3: Buffer(B_2, int8, [16, 16, 16, 16], []), C_1: C_3: Buffer(C_2, int8, [1, 16, 1, 16], [])} {
+ buffers = {A: Buffer(A_2: Pointer(int8), int8, [1, 16, 1, 16], []),
+ B: Buffer(B_2: Pointer(int8), int8, [16, 16, 16, 16], []),
+ C: Buffer(C_2: Pointer(int8), int8, [1, 16, 1, 16], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
allocate(C_buf: Pointer(local.acc_buffer int32), int32, [256]), storage_scope = local.acc_buffer;
allocate(A_buf: Pointer(local.inp_buffer int8), int8, [16]), storage_scope = local.inp_buffer;
allocate(B_buf: Pointer(local.wgt_buffer int8), int8, [16]), storage_scope = local.wgt_buffer {
@@ -522,11 +520,11 @@ moving the copy operations into the matrix multiplication loop.
for (ko: int32, 0, 16) {
attr [IterVar(i0: int32, (nullptr), "DataPar", "")] "pragma_dma_copy" = 1;
for (i3: int32, 0, 16) {
- A_buf_1: Buffer(A_buf, int8, [16], [], scope="local.inp_buffer", align=16)[i3] = A[((ko*16) + i3)]
+ A_buf_1: Buffer(A_buf, int8, [16], [], scope="local.inp_buffer", align=16)[i3] = A_3: Buffer(A_2, int8, [256], [])[((ko*16) + i3)]
}
attr [IterVar(i0_1: int32, (nullptr), "DataPar", "")] "pragma_dma_copy" = 1;
for (i3_1: int32, 0, 16) {
- B_buf_1: Buffer(B_buf, int8, [16], [], scope="local.wgt_buffer", align=256)[i3_1] = B[((((co*4096) + (ko*256)) + (ci*16)) + i3_1)]
+ B_buf_1: Buffer(B_buf, int8, [16], [], scope="local.wgt_buffer", align=256)[i3_1] = B_3: Buffer(B_2, int8, [65536], [])[((((co*4096) + (ko*256)) + (ci*16)) + i3_1)]
}
for (ki: int32, 0, 16) {
let cse_var_1: int32 = ((co*16) + ci)
@@ -539,7 +537,7 @@ moving the copy operations into the matrix multiplication loop.
for (i1: int32, 0, 16) {
for (i3_2: int32, 0, 16) {
let cse_var_2: int32 = ((i1*16) + i3_2)
- C[cse_var_2] = cast(int8, C_buf_1[cse_var_2])
+ C_3: Buffer(C_2, int8, [256], [])[cse_var_2] = cast(int8, C_buf_1[cse_var_2])
}
}
}
@@ -595,11 +593,10 @@ by the VTA runtime JIT compiler.
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(int8), int8, [256], []),
- B: Buffer(B_2: Pointer(int8), int8, [65536], []),
- C: Buffer(C_2: Pointer(int8), int8, [256], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, int8, [1, 16, 1, 16], []), B_1: B_3: Buffer(B_2, int8, [16, 16, 16, 16], []), C_1: C_3: Buffer(C_2, int8, [1, 16, 1, 16], [])} {
+ buffers = {A: Buffer(A_2: Pointer(int8), int8, [1, 16, 1, 16], []),
+ B: Buffer(B_2: Pointer(int8), int8, [16, 16, 16, 16], []),
+ C: Buffer(C_2: Pointer(int8), int8, [1, 16, 1, 16], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
attr [IterVar(vta: int32, (nullptr), "ThreadIndex", "vta")] "coproc_scope" = 2 {
attr [IterVar(vta, (nullptr), "ThreadIndex", "vta")] "coproc_uop_scope" = "VTAPushGEMMOp" {
@tir.call_extern("VTAUopLoopBegin", 16, 1, 0, 0, dtype=int32)
diff --git a/docs/_sources/topic/vta/tutorials/optimize/convolution_opt.rst.txt b/docs/_sources/topic/vta/tutorials/optimize/convolution_opt.rst.txt
index 14ae316c35..bdb8246f06 100644
--- a/docs/_sources/topic/vta/tutorials/optimize/convolution_opt.rst.txt
+++ b/docs/_sources/topic/vta/tutorials/optimize/convolution_opt.rst.txt
@@ -272,11 +272,10 @@ Those include:
@main = primfn(data_1: handle, kernel_1: handle, res_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {data: Buffer(data_2: Pointer(int8), int8, [50176], []),
- kernel: Buffer(kernel_2: Pointer(int8), int8, [589824], []),
- res: Buffer(res_2: Pointer(int8), int8, [50176], [])}
- buffer_map = {data_1: data, kernel_1: kernel, res_1: res}
- preflattened_buffer_map = {data_1: data_3: Buffer(data_2, int8, [1, 16, 14, 14, 1, 16], []), kernel_1: kernel_3: Buffer(kernel_2, int8, [16, 16, 3, 3, 16, 16], []), res_1: res_3: Buffer(res_2, int8, [1, 16, 14, 14, 1, 16], [])} {
+ buffers = {data: Buffer(data_2: Pointer(int8), int8, [1, 16, 14, 14, 1, 16], []),
+ kernel: Buffer(kernel_2: Pointer(int8), int8, [16, 16, 3, 3, 16, 16], []),
+ res: Buffer(res_2: Pointer(int8), int8, [1, 16, 14, 14, 1, 16], [])}
+ buffer_map = {data_1: data, kernel_1: kernel, res_1: res} {
allocate(data_buf: Pointer(global int8), int8, [65536]), storage_scope = global;
allocate(kernel_buf: Pointer(global int8), int8, [589824]), storage_scope = global;
allocate(res_conv: Pointer(global int32), int32, [50176]), storage_scope = global {
@@ -285,7 +284,7 @@ Those include:
for (i3: int32, 0, 16) {
for (i5: int32, 0, 16) {
let cse_var_1: int32 = (i3*16)
- data_buf_1: Buffer(data_buf, int8, [65536], [])[((((i1*4096) + (i2*256)) + cse_var_1) + i5)] = @tir.if_then_else(((((1 <= i2) && (i2 < 15)) && (1 <= i3)) && (i3 < 15)), data[(((((i1*3136) + (i2*224)) + cse_var_1) + i5) - 240)], 0i8, dtype=int8)
+ data_buf_1: Buffer(data_buf, int8, [65536], [])[((((i1*4096) + (i2*256)) + cse_var_1) + i5)] = @tir.if_then_else(((((1 <= i2) && (i2 < 15)) && (1 <= i3)) && (i3 < 15)), data_3: Buffer(data_2, int8, [50176], [])[(((((i1*3136) + (i2*224)) + cse_var_1) + i5) - 240)], 0i8, dtype=int8)
}
}
}
@@ -297,7 +296,7 @@ Those include:
for (i4: int32, 0, 16) {
for (i5_1: int32, 0, 16) {
let cse_var_2: int32 = ((((((i0*36864) + (i1_1*2304)) + (i2_1*768)) + (i3_1*256)) + (i4*16)) + i5_1)
- kernel_buf_1: Buffer(kernel_buf, int8, [589824], [])[cse_var_2] = kernel[cse_var_2]
+ kernel_buf_1: Buffer(kernel_buf, int8, [589824], [])[cse_var_2] = kernel_3: Buffer(kernel_2, int8, [589824], [])[cse_var_2]
}
}
}
@@ -359,7 +358,7 @@ Those include:
for (i3_5: int32, 0, 14) {
for (i5_5: int32, 0, 16) {
let cse_var_8: int32 = ((((i1_5*3136) + (i2_5*224)) + (i3_5*16)) + i5_5)
- res[cse_var_8] = cast(int8, res_conv_4[cse_var_8])
+ res_3: Buffer(res_2, int8, [50176], [])[cse_var_8] = cast(int8, res_conv_4[cse_var_8])
}
}
}
@@ -480,11 +479,10 @@ below.
@main = primfn(data_1: handle, kernel_1: handle, res_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {data: Buffer(data_2: Pointer(int8), int8, [50176], []),
- kernel: Buffer(kernel_2: Pointer(int8), int8, [589824], []),
- res: Buffer(res_2: Pointer(int8), int8, [50176], [])}
- buffer_map = {data_1: data, kernel_1: kernel, res_1: res}
- preflattened_buffer_map = {data_1: data_3: Buffer(data_2, int8, [1, 16, 14, 14, 1, 16], []), kernel_1: kernel_3: Buffer(kernel_2, int8, [16, 16, 3, 3, 16, 16], []), res_1: res_3: Buffer(res_2, int8, [1, 16, 14, 14, 1, 16], [])} {
+ buffers = {data: Buffer(data_2: Pointer(int8), int8, [1, 16, 14, 14, 1, 16], []),
+ kernel: Buffer(kernel_2: Pointer(int8), int8, [16, 16, 3, 3, 16, 16], []),
+ res: Buffer(res_2: Pointer(int8), int8, [1, 16, 14, 14, 1, 16], [])}
+ buffer_map = {data_1: data, kernel_1: kernel, res_1: res} {
allocate(data_buf: Pointer(global int8), int8, [65536]), storage_scope = global;
allocate(kernel_buf: Pointer(global int8), int8, [589824]), storage_scope = global;
allocate(res_conv: Pointer(global int32), int32, [25088]), storage_scope = global {
@@ -493,7 +491,7 @@ below.
for (i3: int32, 0, 16) {
for (i5: int32, 0, 16) {
let cse_var_1: int32 = (i3*16)
- data_buf_1: Buffer(data_buf, int8, [65536], [])[((((i1*4096) + (i2*256)) + cse_var_1) + i5)] = @tir.if_then_else(((((1 <= i2) && (i2 < 15)) && (1 <= i3)) && (i3 < 15)), data[(((((i1*3136) + (i2*224)) + cse_var_1) + i5) - 240)], 0i8, dtype=int8)
+ data_buf_1: Buffer(data_buf, int8, [65536], [])[((((i1*4096) + (i2*256)) + cse_var_1) + i5)] = @tir.if_then_else(((((1 <= i2) && (i2 < 15)) && (1 <= i3)) && (i3 < 15)), data_3: Buffer(data_2, int8, [50176], [])[(((((i1*3136) + (i2*224)) + cse_var_1) + i5) - 240)], 0i8, dtype=int8)
}
}
}
@@ -505,7 +503,7 @@ below.
for (i4: int32, 0, 16) {
for (i5_1: int32, 0, 16) {
let cse_var_2: int32 = ((((((i0*36864) + (i1_1*2304)) + (i2_1*768)) + (i3_1*256)) + (i4*16)) + i5_1)
- kernel_buf_1: Buffer(kernel_buf, int8, [589824], [])[cse_var_2] = kernel[cse_var_2]
+ kernel_buf_1: Buffer(kernel_buf, int8, [589824], [])[cse_var_2] = kernel_3: Buffer(kernel_2, int8, [589824], [])[cse_var_2]
}
}
}
@@ -602,8 +600,8 @@ below.
let cse_var_16: int32 = ((((i1.inner*1568) + cse_var_18) + cse_var_17) + i5_5)
let cse_var_15: int32 = (((((i1.inner*3136) + (i2.outer*1568)) + cse_var_18) + cse_var_17) + i5_5)
{
- res[cse_var_15] = cast(int8, res_conv_4[cse_var_16])
- res[(cse_var_15 + 25088)] = cast(int8, res_conv_4[(cse_var_16 + 12544)])
+ res_3: Buffer(res_2, int8, [50176], [])[cse_var_15] = cast(int8, res_conv_4[cse_var_16])
+ res_3[(cse_var_15 + 25088)] = cast(int8, res_conv_4[(cse_var_16 + 12544)])
}
}
}
@@ -697,11 +695,10 @@ and mapping the shift, and clipping computation to the vector ALU.
@main = primfn(data_1: handle, kernel_1: handle, res_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {data: Buffer(data_2: Pointer(int8), int8, [50176], []),
- kernel: Buffer(kernel_2: Pointer(int8), int8, [589824], []),
- res: Buffer(res_2: Pointer(int8), int8, [50176], [])}
- buffer_map = {data_1: data, kernel_1: kernel, res_1: res}
- preflattened_buffer_map = {data_1: data_3: Buffer(data_2, int8, [1, 16, 14, 14, 1, 16], []), kernel_1: kernel_3: Buffer(kernel_2, int8, [16, 16, 3, 3, 16, 16], []), res_1: res_3: Buffer(res_2, int8, [1, 16, 14, 14, 1, 16], [])} {
+ buffers = {data: Buffer(data_2: Pointer(int8), int8, [1, 16, 14, 14, 1, 16], []),
+ kernel: Buffer(kernel_2: Pointer(int8), int8, [16, 16, 3, 3, 16, 16], []),
+ res: Buffer(res_2: Pointer(int8), int8, [1, 16, 14, 14, 1, 16], [])}
+ buffer_map = {data_1: data, kernel_1: kernel, res_1: res} {
@tir.vta.coproc_dep_push(3, 2, dtype=int32)
@tir.vta.coproc_dep_push(3, 2, dtype=int32)
for (i2.outer: int32, 0, 2) {
diff --git a/docs/_sources/topic/vta/tutorials/optimize/matrix_multiply_opt.rst.txt b/docs/_sources/topic/vta/tutorials/optimize/matrix_multiply_opt.rst.txt
index 5a7b34dba4..9ac9dc9638 100644
--- a/docs/_sources/topic/vta/tutorials/optimize/matrix_multiply_opt.rst.txt
+++ b/docs/_sources/topic/vta/tutorials/optimize/matrix_multiply_opt.rst.txt
@@ -209,18 +209,17 @@ Those include:
@main = primfn(data_1: handle, weight_1: handle, res_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {data: Buffer(data_2: Pointer(int8), int8, [1024], []),
- weight: Buffer(weight_2: Pointer(int8), int8, [1048576], []),
- res: Buffer(res_2: Pointer(int8), int8, [1024], [])}
- buffer_map = {data_1: data, weight_1: weight, res_1: res}
- preflattened_buffer_map = {data_1: data_3: Buffer(data_2, int8, [1, 64, 1, 16], []), weight_1: weight_3: Buffer(weight_2, int8, [64, 64, 16, 16], []), res_1: res_3: Buffer(res_2, int8, [1, 64, 1, 16], [])} {
+ buffers = {data: Buffer(data_2: Pointer(int8), int8, [1, 64, 1, 16], []),
+ weight: Buffer(weight_2: Pointer(int8), int8, [64, 64, 16, 16], []),
+ res: Buffer(res_2: Pointer(int8), int8, [1, 64, 1, 16], [])}
+ buffer_map = {data_1: data, weight_1: weight, res_1: res} {
allocate(data_buf: Pointer(global int8), int8, [1024]), storage_scope = global;
allocate(weight_buf: Pointer(global int8), int8, [1048576]), storage_scope = global;
allocate(res_gem: Pointer(global int32), int32, [1024]), storage_scope = global {
for (i1: int32, 0, 64) {
for (i3: int32, 0, 16) {
let cse_var_1: int32 = ((i1*16) + i3)
- data_buf_1: Buffer(data_buf, int8, [1024], [])[cse_var_1] = data[cse_var_1]
+ data_buf_1: Buffer(data_buf, int8, [1024], [])[cse_var_1] = data_3: Buffer(data_2, int8, [1024], [])[cse_var_1]
}
}
for (i0: int32, 0, 64) {
@@ -228,7 +227,7 @@ Those include:
for (i2: int32, 0, 16) {
for (i3_1: int32, 0, 16) {
let cse_var_2: int32 = ((((i0*16384) + (i1_1*256)) + (i2*16)) + i3_1)
- weight_buf_1: Buffer(weight_buf, int8, [1048576], [])[cse_var_2] = weight[cse_var_2]
+ weight_buf_1: Buffer(weight_buf, int8, [1048576], [])[cse_var_2] = weight_3: Buffer(weight_2, int8, [1048576], [])[cse_var_2]
}
}
}
@@ -265,7 +264,7 @@ Those include:
for (i1_5: int32, 0, 64) {
for (i3_5: int32, 0, 16) {
let cse_var_7: int32 = ((i1_5*16) + i3_5)
- res[cse_var_7] = cast(int8, res_gem_4[cse_var_7])
+ res_3: Buffer(res_2, int8, [1024], [])[cse_var_7] = cast(int8, res_gem_4[cse_var_7])
}
}
}
@@ -378,18 +377,17 @@ below:
@main = primfn(data_1: handle, weight_1: handle, res_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {data: Buffer(data_2: Pointer(int8), int8, [1024], []),
- weight: Buffer(weight_2: Pointer(int8), int8, [1048576], []),
- res: Buffer(res_2: Pointer(int8), int8, [1024], [])}
- buffer_map = {data_1: data, weight_1: weight, res_1: res}
- preflattened_buffer_map = {data_1: data_3: Buffer(data_2, int8, [1, 64, 1, 16], []), weight_1: weight_3: Buffer(weight_2, int8, [64, 64, 16, 16], []), res_1: res_3: Buffer(res_2, int8, [1, 64, 1, 16], [])} {
+ buffers = {data: Buffer(data_2: Pointer(int8), int8, [1, 64, 1, 16], []),
+ weight: Buffer(weight_2: Pointer(int8), int8, [64, 64, 16, 16], []),
+ res: Buffer(res_2: Pointer(int8), int8, [1, 64, 1, 16], [])}
+ buffer_map = {data_1: data, weight_1: weight, res_1: res} {
allocate(data_buf: Pointer(global int8), int8, [1024]), storage_scope = global;
allocate(weight_buf: Pointer(global int8), int8, [1048576]), storage_scope = global;
allocate(res_gem: Pointer(global int32), int32, [256]), storage_scope = global {
for (i1: int32, 0, 64) {
for (i3: int32, 0, 16) {
let cse_var_1: int32 = ((i1*16) + i3)
- data_buf_1: Buffer(data_buf, int8, [1024], [])[cse_var_1] = data[cse_var_1]
+ data_buf_1: Buffer(data_buf, int8, [1024], [])[cse_var_1] = data_3: Buffer(data_2, int8, [1024], [])[cse_var_1]
}
}
for (i0: int32, 0, 64) {
@@ -397,7 +395,7 @@ below:
for (i2: int32, 0, 16) {
for (i3_1: int32, 0, 16) {
let cse_var_2: int32 = ((((i0*16384) + (i1_1*256)) + (i2*16)) + i3_1)
- weight_buf_1: Buffer(weight_buf, int8, [1048576], [])[cse_var_2] = weight[cse_var_2]
+ weight_buf_1: Buffer(weight_buf, int8, [1048576], [])[cse_var_2] = weight_3: Buffer(weight_2, int8, [1048576], [])[cse_var_2]
}
}
}
@@ -441,7 +439,7 @@ below:
for (i1.inner: int32, 0, 16) {
for (i3_5: int32, 0, 16) {
let cse_var_7: int32 = (i1.inner*16)
- res[(((i1.outer*256) + cse_var_7) + i3_5)] = cast(int8, res_gem_4[(cse_var_7 + i3_5)])
+ res_3: Buffer(res_2, int8, [1024], [])[(((i1.outer*256) + cse_var_7) + i3_5)] = cast(int8, res_gem_4[(cse_var_7 + i3_5)])
}
}
}
@@ -532,11 +530,10 @@ and mapping the shift, and clipping computation to the vector ALU.
@main = primfn(data_1: handle, weight_1: handle, res_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {data: Buffer(data_2: Pointer(int8), int8, [1024], []),
- weight: Buffer(weight_2: Pointer(int8), int8, [1048576], []),
- res: Buffer(res_2: Pointer(int8), int8, [1024], [])}
- buffer_map = {data_1: data, weight_1: weight, res_1: res}
- preflattened_buffer_map = {data_1: data_3: Buffer(data_2, int8, [1, 64, 1, 16], []), weight_1: weight_3: Buffer(weight_2, int8, [64, 64, 16, 16], []), res_1: res_3: Buffer(res_2, int8, [1, 64, 1, 16], [])} {
+ buffers = {data: Buffer(data_2: Pointer(int8), int8, [1, 64, 1, 16], []),
+ weight: Buffer(weight_2: Pointer(int8), int8, [64, 64, 16, 16], []),
+ res: Buffer(res_2: Pointer(int8), int8, [1, 64, 1, 16], [])}
+ buffer_map = {data_1: data, weight_1: weight, res_1: res} {
@tir.vta.coproc_dep_push(3, 2, dtype=int32)
for (i1.outer: int32, 0, 4) {
attr [IterVar(vta: int32, (nullptr), "ThreadIndex", "vta")] "coproc_scope" = 2 {
diff --git a/docs/_sources/topic/vta/tutorials/optimize/sg_execution_times.rst.txt b/docs/_sources/topic/vta/tutorials/optimize/sg_execution_times.rst.txt
index cee8d525a8..8231ef25c4 100644
--- a/docs/_sources/topic/vta/tutorials/optimize/sg_execution_times.rst.txt
+++ b/docs/_sources/topic/vta/tutorials/optimize/sg_execution_times.rst.txt
@@ -5,10 +5,10 @@
Computation times
=================
-**00:03.085** total execution time for **topic_vta_tutorials_optimize** files:
+**00:03.108** total execution time for **topic_vta_tutorials_optimize** files:
+--------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_topic_vta_tutorials_optimize_convolution_opt.py` (``convolution_opt.py``) | 00:02.641 | 0.0 MB |
+| :ref:`sphx_glr_topic_vta_tutorials_optimize_convolution_opt.py` (``convolution_opt.py``) | 00:02.673 | 0.0 MB |
+--------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_topic_vta_tutorials_optimize_matrix_multiply_opt.py` (``matrix_multiply_opt.py``) | 00:00.444 | 0.0 MB |
+| :ref:`sphx_glr_topic_vta_tutorials_optimize_matrix_multiply_opt.py` (``matrix_multiply_opt.py``) | 00:00.434 | 0.0 MB |
+--------------------------------------------------------------------------------------------------+-----------+--------+
diff --git a/docs/_sources/topic/vta/tutorials/sg_execution_times.rst.txt b/docs/_sources/topic/vta/tutorials/sg_execution_times.rst.txt
index f11bb71e20..7e83f59ad3 100644
--- a/docs/_sources/topic/vta/tutorials/sg_execution_times.rst.txt
+++ b/docs/_sources/topic/vta/tutorials/sg_execution_times.rst.txt
@@ -5,10 +5,10 @@
Computation times
=================
-**00:00.795** total execution time for **topic_vta_tutorials** files:
+**00:00.767** total execution time for **topic_vta_tutorials** files:
+---------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_topic_vta_tutorials_matrix_multiply.py` (``matrix_multiply.py``) | 00:00.424 | 0.0 MB |
+| :ref:`sphx_glr_topic_vta_tutorials_matrix_multiply.py` (``matrix_multiply.py``) | 00:00.413 | 0.0 MB |
+---------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_topic_vta_tutorials_vta_get_started.py` (``vta_get_started.py``) | 00:00.371 | 0.0 MB |
+| :ref:`sphx_glr_topic_vta_tutorials_vta_get_started.py` (``vta_get_started.py``) | 00:00.354 | 0.0 MB |
+---------------------------------------------------------------------------------+-----------+--------+
diff --git a/docs/_sources/topic/vta/tutorials/vta_get_started.rst.txt b/docs/_sources/topic/vta/tutorials/vta_get_started.rst.txt
index ad56a95716..4d8cfda137 100644
--- a/docs/_sources/topic/vta/tutorials/vta_get_started.rst.txt
+++ b/docs/_sources/topic/vta/tutorials/vta_get_started.rst.txt
@@ -350,23 +350,22 @@ After we construct the schedule, by default the schedule computes
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(int32), int32, [1024], []),
- B: Buffer(B_2: Pointer(int32), int32, [1024], []),
- C: Buffer(C_2: Pointer(int8), int8, [1024], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, int32, [1, 64, 1, 16], []), B_1: B_3: Buffer(B_2, int32, [1, 64, 1, 16], []), C_1: C_3: Buffer(C_2, int8, [1, 64, 1, 16], [])} {
+ buffers = {A: Buffer(A_2: Pointer(int32), int32, [1, 64, 1, 16], []),
+ B: Buffer(B_2: Pointer(int32), int32, [1, 64, 1, 16], []),
+ C: Buffer(C_2: Pointer(int8), int8, [1, 64, 1, 16], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
allocate(A_buf: Pointer(global int32), int32, [1024]), storage_scope = global;
allocate(B_buf: Pointer(global int32), int32, [1024]), storage_scope = global {
for (i1: int32, 0, 64) {
for (i3: int32, 0, 16) {
let cse_var_1: int32 = ((i1*16) + i3)
- A_buf_1: Buffer(A_buf, int32, [1024], [])[cse_var_1] = A[cse_var_1]
+ A_buf_1: Buffer(A_buf, int32, [1024], [])[cse_var_1] = A_3: Buffer(A_2, int32, [1024], [])[cse_var_1]
}
}
for (i1_1: int32, 0, 64) {
for (i3_1: int32, 0, 16) {
let cse_var_2: int32 = ((i1_1*16) + i3_1)
- B_buf_1: Buffer(B_buf, int32, [1024], [])[cse_var_2] = B[cse_var_2]
+ B_buf_1: Buffer(B_buf, int32, [1024], [])[cse_var_2] = B_3: Buffer(B_2, int32, [1024], [])[cse_var_2]
}
}
for (i1_2: int32, 0, 64) {
@@ -378,7 +377,7 @@ After we construct the schedule, by default the schedule computes
for (i1_3: int32, 0, 64) {
for (i3_3: int32, 0, 16) {
let cse_var_4: int32 = ((i1_3*16) + i3_3)
- C[cse_var_4] = cast(int8, A_buf_2[cse_var_4])
+ C_3: Buffer(C_2, int8, [1024], [])[cse_var_4] = cast(int8, A_buf_2[cse_var_4])
}
}
}
@@ -494,11 +493,10 @@ with an :code:`env.alu` pragma.
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(int32), int32, [1024], []),
- B: Buffer(B_2: Pointer(int32), int32, [1024], []),
- C: Buffer(C_2: Pointer(int8), int8, [1024], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, int32, [1, 64, 1, 16], []), B_1: B_3: Buffer(B_2, int32, [1, 64, 1, 16], []), C_1: C_3: Buffer(C_2, int8, [1, 64, 1, 16], [])} {
+ buffers = {A: Buffer(A_2: Pointer(int32), int32, [1, 64, 1, 16], []),
+ B: Buffer(B_2: Pointer(int32), int32, [1, 64, 1, 16], []),
+ C: Buffer(C_2: Pointer(int8), int8, [1, 64, 1, 16], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
attr [IterVar(vta: int32, (nullptr), "ThreadIndex", "vta")] "coproc_scope" = 2 {
@tir.call_extern("VTALoadBuffer2D", @tir.tvm_thread_context(@tir.vta.command_handle(, dtype=handle), dtype=handle), A_2, 0, 64, 1, 64, 0, 0, 0, 0, 0, 3, dtype=int32)
@tir.call_extern("VTALoadBuffer2D", @tir.tvm_thread_context(@tir.vta.command_handle(, dtype=handle), dtype=handle), B_2, 0, 64, 1, 64, 0, 0, 0, 0, 64, 3, dtype=int32)
diff --git a/docs/_sources/tutorial/auto_scheduler_matmul_x86.rst.txt b/docs/_sources/tutorial/auto_scheduler_matmul_x86.rst.txt
index 5030fd8900..69aa0c496e 100644
--- a/docs/_sources/tutorial/auto_scheduler_matmul_x86.rst.txt
+++ b/docs/_sources/tutorial/auto_scheduler_matmul_x86.rst.txt
@@ -203,6 +203,13 @@ trials, we can load the best schedule from the log file and apply it.
+.. rst-class:: sphx-glr-script-out
+
+ .. code-block:: none
+
+
+ *E
+
@@ -234,18 +241,17 @@ operator fusion.
Lowered TIR:
@main = primfn(A_1: handle, B_1: handle, C_1: handle, out_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [1048576], []),
- B: Buffer(B_2: Pointer(float32), float32, [1048576], []),
- C: Buffer(C_2: Pointer(float32), float32, [1048576], []),
- out: Buffer(out_2: Pointer(float32), float32, [1048576], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C, out_1: out}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 1024], []), B_1: B_3: Buffer(B_2, float32, [1024, 1024], []), C_1: C_3: Buffer(C_2, float32, [1024, 1024], []), out_1: out_3: Buffer(out_2, float32, [1024, 1024], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),
+ B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], []),
+ out: Buffer(out_2: Pointer(float32), float32, [1024, 1024], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C, out_1: out} {
allocate(auto_scheduler_layout_transform: Pointer(global float32), float32, [1048576]), storage_scope = global {
for (ax0.ax1.fused.ax2.fused: int32, 0, 128) "parallel" {
for (ax4: int32, 0, 256) {
for (ax6: int32, 0, 4) {
for (ax7: int32, 0, 8) {
- auto_scheduler_layout_transform_1: Buffer(auto_scheduler_layout_transform, float32, [1048576], [])[((((ax0.ax1.fused.ax2.fused*8192) + (ax4*32)) + (ax6*8)) + ax7)] = B[((((ax4*4096) + (ax6*1024)) + (ax0.ax1.fused.ax2.fused*8)) + ax7)]
+ auto_scheduler_layout_transform_1: Buffer(auto_scheduler_layout_transform, float32, [1048576], [])[((((ax0.ax1.fused.ax2.fused*8192) + (ax4*32)) + (ax6*8)) + ax7)] = B_3: Buffer(B_2, float32, [1048576], [])[((((ax4*4096) + (ax6*1024)) + (ax0.ax1.fused.ax2.fused*8)) + ax7)]
}
}
}
@@ -262,16 +268,16 @@ operator fusion.
let cse_var_2: int32 = (((floormod(i.outer.outer.j.outer.outer.fused, 128)*8192) + (k.outer*32)) + (k.inner*8))
let cse_var_1: int32 = ((((floordiv(i.outer.outer.j.outer.outer.fused, 128)*8192) + (i.outer.inner*4096)) + (k.outer*4)) + k.inner)
{
- matmul_1[0] = (matmul_1[0] + (broadcast(A[cse_var_1], 8)*auto_scheduler_layout_transform_1[ramp(cse_var_2, 1, 8)]))
- matmul_1[1] = (matmul_1[1] + (broadcast(A[(cse_var_1 + 1024)], 8)*auto_scheduler_layout_transform_1[ramp(cse_var_2, 1, 8)]))
- matmul_1[2] = (matmul_1[2] + (broadcast(A[(cse_var_1 + 2048)], 8)*auto_scheduler_layout_transform_1[ramp(cse_var_2, 1, 8)]))
- matmul_1[3] = (matmul_1[3] + (broadcast(A[(cse_var_1 + 3072)], 8)*auto_scheduler_layout_transform_1[ramp(cse_var_2, 1, 8)]))
+ matmul_1[0] = (matmul_1[0] + (broadcast(A_3: Buffer(A_2, float32, [1048576], [])[cse_var_1], 8)*auto_scheduler_layout_transform_1[ramp(cse_var_2, 1, 8)]))
+ matmul_1[1] = (matmul_1[1] + (broadcast(A_3[(cse_var_1 + 1024)], 8)*auto_scheduler_layout_transform_1[ramp(cse_var_2, 1, 8)]))
+ matmul_1[2] = (matmul_1[2] + (broadcast(A_3[(cse_var_1 + 2048)], 8)*auto_scheduler_layout_transform_1[ramp(cse_var_2, 1, 8)]))
+ matmul_1[3] = (matmul_1[3] + (broadcast(A_3[(cse_var_1 + 3072)], 8)*auto_scheduler_layout_transform_1[ramp(cse_var_2, 1, 8)]))
}
}
}
for (i.inner: int32, 0, 4) {
let cse_var_3: int32 = ((((floordiv(i.outer.outer.j.outer.outer.fused, 128)*8192) + (i.outer.inner*4096)) + (i.inner*1024)) + (floormod(i.outer.outer.j.outer.outer.fused, 128)*8))
- out[ramp(cse_var_3, 1, 8)] = (matmul_1[i.inner] + C[ramp(cse_var_3, 1, 8)])
+ out_3: Buffer(out_2, float32, [1048576], [])[ramp(cse_var_3, 1, 8)] = (matmul_1[i.inner] + C_3: Buffer(C_2, float32, [1048576], [])[ramp(cse_var_3, 1, 8)])
}
}
}
@@ -326,7 +332,7 @@ We build the binary and check its correctness and performance.
.. code-block:: none
- Execution time of this operator: 95.039 ms
+ Execution time of this operator: 94.882 ms
@@ -444,7 +450,7 @@ operations.
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 1 minutes 12.406 seconds)
+ **Total running time of the script:** ( 1 minutes 33.236 seconds)
.. _sphx_glr_download_tutorial_auto_scheduler_matmul_x86.py:
diff --git a/docs/_sources/tutorial/autotvm_matmul_x86.rst.txt b/docs/_sources/tutorial/autotvm_matmul_x86.rst.txt
index 44862fe9d5..89e7902fd6 100644
--- a/docs/_sources/tutorial/autotvm_matmul_x86.rst.txt
+++ b/docs/_sources/tutorial/autotvm_matmul_x86.rst.txt
@@ -450,16 +450,16 @@ reduce variance, we take 5 measurements and average them.
waiting for device...
device available
Get devices for measurement successfully!
- No: 1 GFLOPS: 2.85/2.85 result: MeasureResult(costs=(0.0942681648,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.6793227195739746, timestamp=1668584257.23453) [('tile_y', [-1, 4]), ('tile_x', [-1, 8])],None,32
- No: 2 GFLOPS: 3.69/3.69 result: MeasureResult(costs=(0.072735489,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.3713977336883545, timestamp=1668584258.607186) [('tile_y', [-1, 256]), ('tile_x', [-1, 16])],None,48
- No: 3 GFLOPS: 4.39/4.39 result: MeasureResult(costs=(0.061116885200000005,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.1401047706604004, timestamp=1668584260.4989111) [('tile_y', [-1, 8]), ('tile_x', [-1, 16])],None,43
- No: 4 GFLOPS: 11.84/11.84 result: MeasureResult(costs=(0.0226792184,), error_no=MeasureErrorNo.NO_ERROR, all_cost=0.5028858184814453, timestamp=1668584261.0308566) [('tile_y', [-1, 2]), ('tile_x', [-1, 512])],None,91
- No: 5 GFLOPS: 2.09/11.84 result: MeasureResult(costs=(0.12822590279999999,), error_no=MeasureErrorNo.NO_ERROR, all_cost=2.198436975479126, timestamp=1668584263.3604555) [('tile_y', [-1, 8]), ('tile_x', [-1, 2])],None,13
- No: 6 GFLOPS: 10.67/11.84 result: MeasureResult(costs=(0.025148017,), error_no=MeasureErrorNo.NO_ERROR, all_cost=0.5394957065582275, timestamp=1668584264.6809936) [('tile_y', [-1, 512]), ('tile_x', [-1, 512])],None,99
- No: 7 GFLOPS: 2.79/11.84 result: MeasureResult(costs=(0.0960730784,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.6821808815002441, timestamp=1668584267.1267865) [('tile_y', [-1, 16]), ('tile_x', [-1, 4])],None,24
- No: 8 GFLOPS: 0.50/11.84 result: MeasureResult(costs=(0.5345131926,), error_no=MeasureErrorNo.NO_ERROR, all_cost=8.745623588562012, timestamp=1668584275.8864374) [('tile_y', [-1, 64]), ('tile_x', [-1, 1])],None,6
- No: 9 GFLOPS: 12.20/12.20 result: MeasureResult(costs=(0.022003847200000003,), error_no=MeasureErrorNo.NO_ERROR, all_cost=0.46866655349731445, timestamp=1668584276.494261) [('tile_y', [-1, 256]), ('tile_x', [-1, 512])],None,98
- No: 10 GFLOPS: 3.88/12.20 result: MeasureResult(costs=(0.0692155048,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.2420101165771484, timestamp=1668584277.7654698) [('tile_y', [-1, 32]), ('tile_x', [-1, 16])],None,45
+ No: 1 GFLOPS: 9.50/9.50 result: MeasureResult(costs=(0.028269254,), error_no=MeasureErrorNo.NO_ERROR, all_cost=0.6169953346252441, timestamp=1668611249.245705) [('tile_y', [-1, 4]), ('tile_x', [-1, 32])],None,52
+ No: 2 GFLOPS: 10.05/10.05 result: MeasureResult(costs=(0.026706745799999997,), error_no=MeasureErrorNo.NO_ERROR, all_cost=0.9295623302459717, timestamp=1668611250.5821307) [('tile_y', [-1, 16]), ('tile_x', [-1, 128])],None,74
+ No: 3 GFLOPS: 1.66/10.05 result: MeasureResult(costs=(0.161746875,), error_no=MeasureErrorNo.NO_ERROR, all_cost=2.7350432872772217, timestamp=1668611254.0759397) [('tile_y', [-1, 8]), ('tile_x', [-1, 1])],None,3
+ No: 4 GFLOPS: 3.68/10.05 result: MeasureResult(costs=(0.072891678,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.3188543319702148, timestamp=1668611256.1606019) [('tile_y', [-1, 128]), ('tile_x', [-1, 16])],None,47
+ No: 5 GFLOPS: 2.95/10.05 result: MeasureResult(costs=(0.09090782359999999,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.6086602210998535, timestamp=1668611257.922664) [('tile_y', [-1, 2]), ('tile_x', [-1, 8])],None,31
+ No: 6 GFLOPS: 2.86/10.05 result: MeasureResult(costs=(0.0937155482,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.6676971912384033, timestamp=1668611259.5862784) [('tile_y', [-1, 2]), ('tile_x', [-1, 16])],None,41
+ No: 7 GFLOPS: 3.61/10.05 result: MeasureResult(costs=(0.07428857520000001,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.3442738056182861, timestamp=1668611260.944852) [('tile_y', [-1, 8]), ('tile_x', [-1, 8])],None,33
+ No: 8 GFLOPS: 13.75/13.75 result: MeasureResult(costs=(0.0195251974,), error_no=MeasureErrorNo.NO_ERROR, all_cost=0.5363571643829346, timestamp=1668611261.464481) [('tile_y', [-1, 128]), ('tile_x', [-1, 64])],None,67
+ No: 9 GFLOPS: 9.56/13.75 result: MeasureResult(costs=(0.028064372200000003,), error_no=MeasureErrorNo.NO_ERROR, all_cost=0.6169466972351074, timestamp=1668611262.2148101) [('tile_y', [-1, 512]), ('tile_x', [-1, 128])],None,79
+ No: 10 GFLOPS: 13.15/13.75 result: MeasureResult(costs=(0.0204108224,), error_no=MeasureErrorNo.NO_ERROR, all_cost=0.5056636333465576, timestamp=1668611262.7110298) [('tile_y', [-1, 128]), ('tile_x', [-1, 128])],None,77
diff --git a/docs/_sources/tutorial/autotvm_relay_x86.rst.txt b/docs/_sources/tutorial/autotvm_relay_x86.rst.txt
index adb66bcbb9..4d21e88764 100644
--- a/docs/_sources/tutorial/autotvm_relay_x86.rst.txt
+++ b/docs/_sources/tutorial/autotvm_relay_x86.rst.txt
@@ -320,7 +320,7 @@ standard deviation.
.. code-block:: none
- {'mean': 515.8039834199826, 'median': 515.9055659999922, 'std': 1.7053726459433556}
+ {'mean': 515.9832460300095, 'median': 516.074852450015, 'std': 2.1019770271666354}
@@ -554,32 +554,29 @@ the tuning data to.
.. code-block:: none
-
[Task 1/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 1/25] Current/Best: 6.66/ 22.29 GFLOPS | Progress: (4/20) | 7.71 s
[Task 1/25] Current/Best: 21.20/ 22.29 GFLOPS | Progress: (8/20) | 11.23 s
[Task 1/25] Current/Best: 10.61/ 22.29 GFLOPS | Progress: (12/20) | 13.83 s
[Task 1/25] Current/Best: 18.10/ 22.29 GFLOPS | Progress: (16/20) | 15.81 s
[Task 1/25] Current/Best: 23.07/ 23.07 GFLOPS | Progress: (20/20) | 17.45 s Done.
-
[Task 2/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 2/25] Current/Best: 17.22/ 21.41 GFLOPS | Progress: (4/20) | 2.83 s
[Task 2/25] Current/Best: 11.24/ 21.41 GFLOPS | Progress: (8/20) | 4.48 s
[Task 2/25] Current/Best: 6.42/ 21.41 GFLOPS | Progress: (12/20) | 6.21 s
[Task 2/25] Current/Best: 14.07/ 21.41 GFLOPS | Progress: (16/20) | 7.74 s
[Task 2/25] Current/Best: 12.19/ 21.41 GFLOPS | Progress: (20/20) | 9.53 s Done.
-
[Task 3/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 3/25] Current/Best: 10.44/ 14.23 GFLOPS | Progress: (4/20) | 3.63 s
[Task 3/25] Current/Best: 12.70/ 17.81 GFLOPS | Progress: (8/20) | 5.70 s
[Task 3/25] Current/Best: 8.62/ 20.98 GFLOPS | Progress: (12/20) | 7.85 s
[Task 3/25] Current/Best: 13.95/ 20.98 GFLOPS | Progress: (16/20) | 9.86 s
[Task 3/25] Current/Best: 12.35/ 20.98 GFLOPS | Progress: (20/20) | 11.53 s Done.
-
[Task 4/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 4/25] Current/Best: 12.82/ 15.97 GFLOPS | Progress: (4/20) | 2.98 s
[Task 4/25] Current/Best: 17.11/ 18.98 GFLOPS | Progress: (8/20) | 4.39 s
[Task 4/25] Current/Best: 6.99/ 18.98 GFLOPS | Progress: (12/20) | 8.98 s
[Task 4/25] Current/Best: 7.47/ 21.33 GFLOPS | Progress: (16/20) | 11.63 s
[Task 4/25] Current/Best: 12.75/ 21.33 GFLOPS | Progress: (20/20) | 13.22 s Done.
-
[Task 5/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 5/25] Current/Best: 15.81/ 22.16 GFLOPS | Progress: (4/20) | 3.03 s
[Task 5/25] Current/Best: 13.04/ 22.16 GFLOPS | Progress: (8/20) | 4.98 s
[Task 5/25] Current/Best: 8.44/ 22.16 GFLOPS | Progress: (12/20) | 7.11 s
[Task 5/25] Current/Best: 11.35/ 22.16 GFLOPS | Progress: (16/20) | 9.24 s
[Task 5/25] Current/Best: 4.05/ 22.16 GFLOPS | Progress: (20/20) | 11.44 s Done.
-
[Task 6/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 6/25] Current/Best: 16.74/ 17.13 GFLOPS | Progress: (4/20) | 3.68 s
[Task 6/25] Current/Best: 9.41/ 18.66 GFLOPS | Progress: (8/20) | 6.65 s
[Task 6/25] Current/Best: 6.10/ 19.56 GFLOPS | Progress: (12/20) | 8.44 s
[Task 6/25] Current/Best: 13.57/ 19.56 GFLOPS | Progress: (16/20) | 11.20 s
[Task 6/25] Current/Best: 14.61/ 22.66 GFLOPS | Progress: (20/20) | 13.42 s Done.
-
[Task 7/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 7/25] Current/Best: 12.48/ 12.55 GFLOPS | Progress: (4/20) | 3.52 s
[Task 7/25] Current/Best: 11.60/ 12.55 GFLOPS | Progress: (8/20) | 6.54 s
[Task 7/25] Current/Best: 15.23/ 15.58 GFLOPS | Progress: (12/20) | 8.35 s
[Task 7/25] Current/Best: 15.06/ 15.58 GFLOPS | Progress: (16/20) | 10.64 s
[Task 7/25] Current/Best: 13.46/ 15.58 GFLOPS | Progress: (20/20) | 13.25 s Done.
-
[Task 8/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 8/25] Current/Best: 5.64/ 13.42 GFLOPS | Progress: (4/20) | 3.58 s
[Task 8/25] Current/Best: 14.96/ 14.96 GFLOPS | Progress: (8/20) | 6.14 s
[Task 8/25] Current/Best: 19.26/ 19.26 GFLOPS | Progress: (12/20) | 10.44 s
[Task 8/25] Current/Best: 2.88/ 19.26 GFLOPS | Progress: (16/20) | 19.47 s
[Task 8/25] Current/Best: 12.23/ 19.26 GFLOPS | Progress: (20/20) | 22.72 s Done.
-
[Task 9/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 9/25] Current/Best: 7.44/ 16.93 GFLOPS | Progress: (4/20) | 3.40 s
[Task 9/25] Current/Best: 15.76/ 16.93 GFLOPS | Progress: (8/20) | 7.35 s
[Task 9/25] Current/Best: 15.80/ 16.93 GFLOPS | Progress: (12/20) | 18.24 s
[Task 9/25] Current/Best: 12.41/ 16.93 GFLOPS | Progress: (16/20) | 24.81 s
[Task 9/25] Current/Best: 12.14/ 20.84 GFLOPS | Progress: (20/20) | 27.93 s
[Task 10/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 10/25] Current/Best: 16.30/ 20.25 GFLOPS | Progress: (4/20) | 2.97 s
[Task 10/25] Current/Best: 3.86/ 22.58 GFLOPS | Progress: (8/20) | 4.52 s
[Task 10/25] Current/Best: 4.28/ 22.58 GFLOPS | Progress: (12/20) | 6.26 s
[Task 10/25] Current/Best: 9.10/ 22.58 GFLOPS | Progress: (16/20) | 7.57 s
[Task 10/25] Current/Best: 10.21/ 22.58 GFLOPS | Progress: (20/20)
| 9.13 s Done.
-
[Task 11/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 11/25] Current/Best: 16.06/ 21.49 GFLOPS | Progress: (4/20) | 3.36 s
[Task 11/25] Current/Best: 17.35/ 21.49 GFLOPS | Progress: (8/20) | 5.84 s
[Task 11/25] Current/Best: 11.77/ 21.49 GFLOPS | Progress: (12/20) | 8.30 s
[Task 11/25] Current/Best: 8.97/ 21.49 GFLOPS | Progress: (16/20) | 10.98 s
[Task 11/25] Current/Best: 20.01/ 21.49 GFLOPS | Progress: (20/20) | 14.71 s Done.
-
[Task 12/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 12/25] Current/Best: 5.69/ 12.20 GFLOPS | Progress: (4/20) | 3.58 s
[Task 12/25] Current/Best: 14.16/ 17.96 GFLOPS | Progress: (8/20) | 6.49 s
[Task 12/25] Current/Best: 11.64/ 17.96 GFLOPS | Progress: (12/20) | 9.67 s
[Task 12/25] Current/Best: 7.77/ 21.31 GFLOPS | Progress: (16/20) | 13.11 s
[Task 12/25] Current/Best: 13.21/ 21.31 GFLOPS | Progress: (20/20) | 15.06 s Done.
-
[Task 13/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 13/25] Current/Best: 16.55/ 16.79 GFLOPS | Progress: (4/20) | 3.30 s
[Task 13/25] Current/Best: 14.48/ 16.79 GFLOPS | Progress: (8/20) | 7.87 s
[Task 13/25] Current/Best: 10.68/ 19.41 GFLOPS | Progress: (12/20) | 10.71 s
[Task 13/25] Current/Best: 6.19/ 19.41 GFLOPS | Progress: (16/20) | 13.42 s
[Task 13/25] Current/Best: 20.78/ 21.13 GFLOPS | Progress: (20/20) | 15.64 s Done.
-
[Task 14/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 14/25] Current/Best: 16.06/ 16.06 GFLOPS | Progress: (4/20) | 3.90 s
[Task 14/25] Current/Best: 14.41/ 21.37 GFLOPS | Progress: (8/20) | 6.21 s
[Task 14/25] Current/Best: 21.04/ 22.32 GFLOPS | Progress: (12/20) | 10.07 s
[Task 14/25] Current/Best: 10.75/ 22.32 GFLOPS | Progress: (16/20) | 13.75 s
[Task 14/25] Current/Best: 8.94/ 22.32 GFLOPS | Progress: (20/20) | 19.75 s
[Task 15/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 15/25] Current/Best: 16.22/ 17.19 GFLOPS | Progress: (4/20) | 2.86 s
[Task 15/25] Current/Best: 16.76/ 17.19 GFLOPS | Progress: (8/20) | 4.17 s Done.
+
[Task 1/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 1/25] Current/Best: 17.62/ 17.62 GFLOPS | Progress: (4/20) | 9.23 s
[Task 1/25] Current/Best: 22.20/ 22.93 GFLOPS | Progress: (8/20) | 11.31 s
[Task 1/25] Current/Best: 8.53/ 22.93 GFLOPS | Progress: (12/20) | 14.39 s
[Task 1/25] Current/Best: 13.52/ 22.93 GFLOPS | Progress: (16/20) | 16.87 s
[Task 1/25] Current/Best: 9.93/ 22.93 GFLOPS | Progress: (20/20) | 19.27 s Done.
+
[Task 2/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 2/25] Current/Best: 18.42/ 19.63 GFLOPS | Progress: (4/20) | 2.67 s
[Task 2/25] Current/Best: 12.58/ 19.63 GFLOPS | Progress: (8/20) | 3.74 s
[Task 2/25] Current/Best: 13.59/ 19.63 GFLOPS | Progress: (12/20) | 5.36 s
[Task 2/25] Current/Best: 6.25/ 19.63 GFLOPS | Progress: (16/20) | 7.11 s
[Task 2/25] Current/Best: 12.95/ 20.04 GFLOPS | Progress: (20/20) | 8.42 s Done.
+
[Task 3/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 3/25] Current/Best: 17.03/ 17.03 GFLOPS | Progress: (4/20) | 3.31 s
[Task 3/25] Current/Best: 5.74/ 20.14 GFLOPS | Progress: (8/20) | 5.95 s
[Task 3/25] Current/Best: 15.87/ 23.62 GFLOPS | Progress: (12/20) | 8.30 s
[Task 3/25] Current/Best: 17.11/ 23.62 GFLOPS | Progress: (16/20) | 10.95 s
[Task 3/25] Current/Best: 17.90/ 23.62 GFLOPS | Progress: (20/20) | 13.52 s Done.
+
[Task 4/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 4/25] Current/Best: 8.44/ 18.73 GFLOPS | Progress: (4/20) | 3.59 s
[Task 4/25] Current/Best: 5.10/ 18.73 GFLOPS | Progress: (8/20) | 11.01 s
[Task 4/25] Current/Best: 15.76/ 18.73 GFLOPS | Progress: (12/20) | 12.91 s
[Task 4/25] Current/Best: 11.95/ 18.73 GFLOPS | Progress: (16/20) | 14.61 s
[Task 4/25] Current/Best: 11.70/ 18.73 GFLOPS | Progress: (20/20) | 17.29 s Done.
+
[Task 5/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 5/25] Current/Best: 6.82/ 19.14 GFLOPS | Progress: (4/20) | 3.05 s
[Task 5/25] Current/Best: 4.12/ 19.14 GFLOPS | Progress: (8/20) | 4.96 s
[Task 5/25] Current/Best: 18.45/ 19.14 GFLOPS | Progress: (12/20) | 6.41 s
[Task 5/25] Current/Best: 6.80/ 19.14 GFLOPS | Progress: (16/20) | 8.18 s
[Task 5/25] Current/Best: 6.72/ 19.14 GFLOPS | Progress: (20/20) | 10.05 s Done.
+
[Task 6/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 6/25] Current/Best: 16.23/ 16.23 GFLOPS | Progress: (4/20) | 4.04 s
[Task 6/25] Current/Best: 20.57/ 20.57 GFLOPS | Progress: (8/20) | 5.80 s
[Task 6/25] Current/Best: 14.70/ 20.57 GFLOPS | Progress: (12/20) | 8.75 s
[Task 6/25] Current/Best: 12.13/ 20.57 GFLOPS | Progress: (16/20) | 11.07 s
[Task 6/25] Current/Best: 17.89/ 20.57 GFLOPS | Progress: (20/20) | 12.82 s Done.
+
[Task 7/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 7/25] Current/Best: 11.14/ 16.33 GFLOPS | Progress: (4/20) | 4.95 s
[Task 7/25] Current/Best: 15.37/ 16.66 GFLOPS | Progress: (8/20) | 6.95 s
[Task 7/25] Current/Best: 15.60/ 16.66 GFLOPS | Progress: (12/20) | 9.50 s
[Task 7/25] Current/Best: 17.45/ 17.45 GFLOPS | Progress: (16/20) | 12.01 s
[Task 7/25] Current/Best: 13.67/ 20.09 GFLOPS | Progress: (20/20) | 13.81 s Done.
+
[Task 8/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 8/25] Current/Best: 5.35/ 15.05 GFLOPS | Progress: (4/20) | 4.50 s
[Task 8/25] Current/Best: 16.11/ 16.11 GFLOPS | Progress: (8/20) | 15.62 s
[Task 8/25] Current/Best: 2.77/ 16.70 GFLOPS | Progress: (12/20) | 21.04 s
[Task 8/25] Current/Best: 15.39/ 16.70 GFLOPS | Progress: (16/20) | 28.78 s
[Task 8/25] Current/Best: 11.97/ 16.70 GFLOPS | Progress: (20/20) | 31.86 s
[Task 9/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 9/25] Current/Best: 13.79/ 13.79 GFLOPS | Progress: (4/20) | 5.40 s
[Task 9/25] Current/Best: 10.91/ 17.79 GFLOPS | Progress: (8/20) | 11.29 s
[Task 9/25] Current/Best: 12.27/ 22.39 GFLOPS | Progress: (12/20) | 15.97 s
[Task 9/25] Current/Best: 10.27/ 22.39 GFLOPS | Progress: (16/20) | 20.20 s
[Task 9/25] Current/Best: 17.87/ 22.39 GFLOPS | Progress: (20/
20) | 27.52 s Done.
+
[Task 10/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 10/25] Current/Best: 18.03/ 19.55 GFLOPS | Progress: (4/20) | 2.93 s
[Task 10/25] Current/Best: 8.00/ 19.55 GFLOPS | Progress: (8/20) | 4.37 s
[Task 10/25] Current/Best: 17.89/ 19.55 GFLOPS | Progress: (12/20) | 6.14 s
[Task 10/25] Current/Best: 1.61/ 19.55 GFLOPS | Progress: (16/20) | 8.15 s
[Task 10/25] Current/Best: 9.84/ 19.55 GFLOPS | Progress: (20/20) | 11.03 s Done.
+
[Task 11/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 11/25] Current/Best: 18.43/ 21.07 GFLOPS | Progress: (4/20) | 3.65 s
[Task 11/25] Current/Best: 17.97/ 21.07 GFLOPS | Progress: (8/20) | 6.36 s
[Task 11/25] Current/Best: 18.34/ 21.07 GFLOPS | Progress: (12/20) | 8.98 s
[Task 11/25] Current/Best: 6.28/ 22.11 GFLOPS | Progress: (16/20) | 10.99 s
[Task 11/25] Current/Best: 8.99/ 22.11 GFLOPS | Progress: (20/20) | 12.74 s Done.
+
[Task 12/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 12/25] Current/Best: 14.96/ 14.96 GFLOPS | Progress: (4/20) | 7.59 s
[Task 12/25] Current/Best: 13.11/ 14.96 GFLOPS | Progress: (8/20) | 10.25 s
[Task 12/25] Current/Best: 7.19/ 16.25 GFLOPS | Progress: (12/20) | 12.80 s
[Task 12/25] Current/Best: 13.40/ 16.83 GFLOPS | Progress: (16/20) | 15.64 s
[Task 12/25] Current/Best: 18.27/ 18.27 GFLOPS | Progress: (20/20) | 18.61 s Done.
+
[Task 13/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 13/25] Current/Best: 15.64/ 20.70 GFLOPS | Progress: (4/20) | 3.65 s
[Task 13/25] Current/Best: 6.99/ 20.70 GFLOPS | Progress: (8/20) | 7.70 s
[Task 13/25] Current/Best: 20.14/ 20.70 GFLOPS | Progress: (12/20) | 9.74 s
[Task 13/25] Current/Best: 18.66/ 20.70 GFLOPS | Progress: (16/20) | 12.46 s
[Task 13/25] Current/Best: 17.19/ 20.70 GFLOPS | Progress: (20/20) | 14.71 s Done.
+
[Task 14/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 14/25] Current/Best: 15.95/ 16.88 GFLOPS | Progress: (4/20) | 4.73 s
[Task 14/25] Current/Best: 10.84/ 16.88 GFLOPS | Progress: (8/20) | 6.69 s
[Task 14/25] Current/Best: 4.59/ 17.93 GFLOPS | Progress: (12/20) | 9.66 s
[Task 14/25] Current/Best: 16.22/ 17.93 GFLOPS | Progress: (16/20) | 13.34 s
[Task 14/25] Current/Best: 13.15/ 17.93 GFLOPS | Progress: (20/20) | 15.26 s Done.
+
[Task 15/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 15/25] Current/Best: 19.42/ 20.85 GFLOPS | Progress: (4/20) | 3.30 s
[Task 15/25] Current/Best: 12.02/ 20.85 GFLOPS | Progress: (8/20) | 8.36 s Done.
+
[Task 15/25] Current/Best: 16.04/ 20.85 GFLOPS | Progress: (12/20) | 10.34 s
[Task 15/25] Current/Best: 12.46/ 20.85 GFLOPS | Progress: (16/20) | 15.89 s
[Task 15/25] Current/Best: 13.47/ 21.22 GFLOPS | Progress: (20/20) | 17.92 s
[Task 16/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 16/25] Current/Best: 8.65/ 18.84 GFLOPS | Progress: (4/20) | 2.79 s
[Task 16/25] Current/Best: 15.93/ 19.96 GFLOPS | Progress: (8/20) | 4.03 s
[Task 16/25] Current/Best: 15.47/ 19.96 GFLOPS | Progress: (12/20) | 5.24 s
[Task 16/25] Current/Best: 18.49/ 19.96 GFLOPS | Progress: (16/20) | 6.97 s
[Task 16/25] Current/Best: 6.75/ 21.17 GFLOPS | Progress: (20/20) | 9.72 s Done.
+
[Task 17/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 17/25] Current/Best: 10.18/ 15.89 GFLOPS | Progress: (4/20) | 4.54 s
[Task 17/25] Current/Best: 20.77/ 23.74 GFLOPS | Progress: (8/20) | 6.15 s
[Task 17/25] Current/Best: 9.72/ 23.74 GFLOPS | Progress: (12/20) | 8.51 s
[Task 17/25] Current/Best: 17.01/ 23.74 GFLOPS | Progress: (16/20) | 10.72 s
[Task 17/25] Current/Best: 10.42/ 23.74 GFLOPS | Progress: (20/20) | 13.24 s Done.
+
[Task 18/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 18/25] Current/Best: 14.95/ 19.22 GFLOPS | Progress: (4/20) | 4.99 s
[Task 18/25] Current/Best: 17.63/ 19.22 GFLOPS | Progress: (8/20) | 6.65 s
[Task 18/25] Current/Best: 5.77/ 19.22 GFLOPS | Progress: (12/20) | 13.28 s
[Task 18/25] Current/Best: 9.45/ 19.22 GFLOPS | Progress: (16/20) | 17.04 s
[Task 18/25] Current/Best: 16.10/ 19.22 GFLOPS | Progress: (20/20) | 19.12 s Done.
+
[Task 19/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 19/25] Current/Best: 13.38/ 18.27 GFLOPS | Progress: (4/20) | 6.64 s
[Task 19/25] Current/Best: 19.57/ 19.57 GFLOPS | Progress: (8/20) | 9.94 s
[Task 19/25] Current/Best: 2.68/ 19.57 GFLOPS | Progress: (12/20) | 14.00 s
[Task 19/25] Current/Best: 8.18/ 20.74 GFLOPS | Progress: (16/20) | 17.23 s
[Task 19/25] Current/Best: 10.09/ 20.74 GFLOPS | Progress: (20/20) | 19.62 s Done.
+
[Task 20/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 20/25] Current/Best: 7.49/ 7.49 GFLOPS | Progress: (4/20) | 3.45 s
[Task 20/25] Current/Best: 16.03/ 16.76 GFLOPS | Progress: (8/20) | 9.41 s
[Task 20/25] Current/Best: 4.17/ 16.76 GFLOPS | Progress: (12/20) | 12.24 s
[Task 20/25] Current/Best: 10.20/ 16.76 GFLOPS | Progress: (16/20) | 14.23 s
[Task 20/25] Current/Best: 5.07/ 16.76 GFLOPS | Progress: (20/20) | 17.55 s
[Task 21/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 21/25] Current/Best: 10.31/ 13.11 GFLOPS | Progress: (4/20) | 4.96 s Done.
Done.
-
[Task 15/25] Current/Best: 16.99/ 21.35 GFLOPS | Progress: (12/20) | 6.01 s
[Task 15/25] Current/Best: 16.88/ 21.35 GFLOPS | Progress: (16/20) | 7.40 s
[Task 15/25] Current/Best: 18.87/ 22.28 GFLOPS | Progress: (20/20) | 9.21 s Done.
-
[Task 16/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 16/25] Current/Best: 11.81/ 20.64 GFLOPS | Progress: (4/20) | 4.14 s
[Task 16/25] Current/Best: 18.05/ 20.64 GFLOPS | Progress: (8/20) | 5.83 s
[Task 16/25] Current/Best: 9.31/ 20.64 GFLOPS | Progress: (12/20) | 7.77 s
[Task 16/25] Current/Best: 17.78/ 20.64 GFLOPS | Progress: (16/20) | 9.01 s
[Task 16/25] Current/Best: 9.29/ 20.64 GFLOPS | Progress: (20/20) | 10.86 s Done.
-
[Task 17/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 17/25] Current/Best: 17.91/ 19.69 GFLOPS | Progress: (4/20) | 3.52 s
[Task 17/25] Current/Best: 19.46/ 19.69 GFLOPS | Progress: (8/20) | 5.73 s
[Task 17/25] Current/Best: 19.56/ 19.69 GFLOPS | Progress: (12/20) | 8.71 s
[Task 17/25] Current/Best: 21.43/ 21.43 GFLOPS | Progress: (16/20) | 10.43 s
[Task 17/25] Current/Best: 3.09/ 21.43 GFLOPS | Progress: (20/20) | 13.00 s Done.
-
[Task 18/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 18/25] Current/Best: 17.42/ 17.42 GFLOPS | Progress: (4/20) | 3.97 s
[Task 18/25] Current/Best: 6.35/ 17.42 GFLOPS | Progress: (8/20) | 6.51 s
[Task 18/25] Current/Best: 5.08/ 17.42 GFLOPS | Progress: (12/20) | 12.52 s
[Task 18/25] Current/Best: 9.73/ 17.42 GFLOPS | Progress: (16/20) | 18.86 s
[Task 18/25] Current/Best: 12.38/ 17.42 GFLOPS | Progress: (20/20) | 23.71 s Done.
-
[Task 19/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 19/25] Current/Best: 5.23/ 9.61 GFLOPS | Progress: (4/20) | 7.69 s
[Task 19/25] Current/Best: 10.79/ 11.44 GFLOPS | Progress: (8/20) | 11.89 s
[Task 19/25] Current/Best: 17.61/ 17.61 GFLOPS | Progress: (12/20) | 15.23 s
[Task 19/25] Current/Best: 9.46/ 23.09 GFLOPS | Progress: (16/20) | 18.42 s
[Task 19/25] Current/Best: 18.65/ 23.09 GFLOPS | Progress: (20/20) | 21.68 s Done.
-
[Task 20/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 20/25] Current/Best: 11.13/ 13.99 GFLOPS | Progress: (4/20) | 3.62 s
[Task 20/25] Current/Best: 13.84/ 13.99 GFLOPS | Progress: (8/20) | 6.35 s
[Task 20/25] Current/Best: 4.86/ 13.99 GFLOPS | Progress: (12/20) | 9.87 s
[Task 20/25] Current/Best: 11.59/ 19.58 GFLOPS | Progress: (16/20) | 12.51 s
[Task 20/25] Current/Best: 10.50/ 21.04 GFLOPS | Progress: (20/20) | 15.72 s
[Task 21/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 21/25] Current/Best: 3.15/ 8.96 GFLOPS | Progress: (4/20) | 3.98 s
[Task 21/25] Current/Best: 10.02/ 10.02 GFLOPS | Progress: (8/20) | 6.99 s
[Task 21/25] Current/Best: 14.52/ 14.52 GFLOPS | Progress: (12/20) | 9.68 s Done.
-
[Task 21/25] Current/Best: 13.61/ 17.05 GFLOPS | Progress: (16/20) | 12.10 s
[Task 21/25] Current/Best: 16.30/ 22.08 GFLOPS | Progress: (20/20) | 13.63 s
[Task 22/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 22/25] Current/Best: 6.16/ 15.37 GFLOPS | Progress: (4/20) | 3.66 s
[Task 22/25] Current/Best: 17.71/ 17.71 GFLOPS | Progress: (8/20) | 4.99 s
[Task 22/25] Current/Best: 16.97/ 19.14 GFLOPS | Progress: (12/20) | 6.42 s
[Task 22/25] Current/Best: 10.10/ 19.14 GFLOPS | Progress: (16/20) | 8.11 s
[Task 22/25] Current/Best: 6.14/ 20.44 GFLOPS | Progress: (20/20) | 12.15 s Done.
-
[Task 23/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 23/25] Current/Best: 2.69/ 11.33 GFLOPS | Progress: (4/20) | 6.46 s
[Task 23/25] Current/Best: 19.17/ 19.17 GFLOPS | Progress: (8/20) | 9.51 s
[Task 23/25] Current/Best: 10.51/ 19.17 GFLOPS | Progress: (12/20) | 12.64 s
[Task 23/25] Current/Best: 17.50/ 23.10 GFLOPS | Progress: (16/20) | 14.41 s
[Task 23/25] Current/Best: 5.02/ 23.10 GFLOPS | Progress: (20/20) | 17.32 s Done.
-
[Task 24/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 24/25] Current/Best: 3.13/ 5.00 GFLOPS | Progress: (4/20) | 12.28 s
[Task 24/25] Current/Best: 3.29/ 5.00 GFLOPS | Progress: (8/20) | 23.54 s
[Task 24/25] Current/Best: 3.43/ 5.00 GFLOPS | Progress: (12/20) | 35.48 s
[Task 24/25] Current/Best: 8.21/ 8.21 GFLOPS | Progress: (16/20) | 45.94 s
[Task 24/25] Current/Best: 7.65/ 8.21 GFLOPS | Progress: (20/20) | 57.65 s
[Task 25/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s Done.
- Done.
-
[Task 25/25] Current/Best: 9.26/ 9.26 GFLOPS | Progress: (4/20) | 5.79 s
[Task 25/25] Current/Best: 1.52/ 9.26 GFLOPS | Progress: (8/20) | 15.75 s
[Task 25/25] Current/Best: 5.20/ 9.26 GFLOPS | Progress: (12/20) | 26.23 s
[Task 25/25] Current/Best: 1.55/ 9.26 GFLOPS | Progress: (16/20) | 36.97 s
[Task 25/25] Current/Best: 4.24/ 9.26 GFLOPS | Progress: (20/20) | 40.32 s Done.
-
+
[Task 21/25] Current/Best: 14.77/ 21.91 GFLOPS | Progress: (8/20) | 6.74 s
[Task 21/25] Current/Best: 12.92/ 21.91 GFLOPS | Progress: (12/20) | 9.14 s
[Task 21/25] Current/Best: 15.33/ 21.91 GFLOPS | Progress: (16/20) | 11.08 s
[Task 21/25] Current/Best: 8.90/ 21.91 GFLOPS | Progress: (20/20) | 13.09 s
[Task 22/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 22/25] Current/Best: 4.44/ 7.62 GFLOPS | Progress: (4/20) | 4.19 s
[Task 22/25] Current/Best: 14.70/ 17.83 GFLOPS | Progress: (8/20) | 6.71 s
[Task 22/25] Current/Best: 12.40/ 20.78 GFLOPS | Progress: (12/20) | 8.56 s
[Task 22/25] Current/Best: 20.28/ 20.78 GFLOPS | Progress: (16/20) | 11.73 s
[Task 22/25] Current/Best: 7.02/ 20.78 GFLOPS | Progress: (20/20) | 14.21 s Done.
+
[Task 23/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 23/25] Current/Best: 9.41/ 19.74 GFLOPS | Progress: (4/20) | 4.17 s
[Task 23/25] Current/Best: 9.27/ 19.74 GFLOPS | Progress: (8/20) | 6.75 s
[Task 23/25] Current/Best: 23.76/ 23.76 GFLOPS | Progress: (12/20) | 8.81 s
[Task 23/25] Current/Best: 3.08/ 23.76 GFLOPS | Progress: (16/20) | 11.64 s
[Task 23/25] Current/Best: 12.00/ 23.76 GFLOPS | Progress: (20/20) | 14.27 s Done.
+
[Task 24/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 24/25] Current/Best: 2.27/ 6.50 GFLOPS | Progress: (4/20) | 12.26 s
[Task 24/25] Current/Best: 7.90/ 7.90 GFLOPS | Progress: (8/20) | 23.59 s
[Task 24/25] Current/Best: 6.66/ 7.90 GFLOPS | Progress: (12/20) | 34.44 s
[Task 24/25] Current/Best: 8.96/ 8.96 GFLOPS | Progress: (16/20) | 45.11 s
[Task 24/25] Current/Best: 7.19/ 8.96 GFLOPS | Progress: (20/20) | 55.58 s
[Task 25/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 25/25] Current/Best: 9.06/ 9.06 GFLOPS | Progress: (4/20) | 11.59 s
[Task 25/25] Current/Best: 3.50/ 9.17 GFLOPS | Progress: (8/20) | 22.30 s
[Task 25/25] Current/Best: 7.17/ 9.17 GFLOPS | Progress: (12/20) | 24.08 s
[Task 25/25] Current/Best: 7.55/ 9.17 GFLOPS | Progress: (16/20) | 34.79 s
[Task 25/25] Current/Best: 4.37/ 9.17 GFLOPS | Progress: (2
0/20) | 36.67 s
@@ -675,8 +672,8 @@ Verify that the optimized model runs and produces the same results:
.. code-block:: none
- class='n02123045 tabby, tabby cat' with probability=0.621104
- class='n02123159 tiger cat' with probability=0.356378
+ class='n02123045 tabby, tabby cat' with probability=0.621103
+ class='n02123159 tiger cat' with probability=0.356379
class='n02124075 Egyptian cat' with probability=0.019712
class='n02129604 tiger, Panthera tigris' with probability=0.001215
class='n04040759 radiator' with probability=0.000262
@@ -733,8 +730,8 @@ improvement in comparing the optimized model to the unoptimized model.
.. code-block:: none
- optimized: {'mean': 410.8896473599907, 'median': 411.05055119996905, 'std': 4.679439781828764}
- unoptimized: {'mean': 515.8039834199826, 'median': 515.9055659999922, 'std': 1.7053726459433556}
+ optimized: {'mean': 417.24817383000754, 'median': 416.0994904000063, 'std': 3.345817977638771}
+ unoptimized: {'mean': 515.9832460300095, 'median': 516.074852450015, 'std': 2.1019770271666354}
@@ -757,7 +754,7 @@ profiling/benchmarking.
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 10 minutes 59.617 seconds)
+ **Total running time of the script:** ( 11 minutes 9.688 seconds)
.. _sphx_glr_download_tutorial_autotvm_relay_x86.py:
diff --git a/docs/_sources/tutorial/cross_compilation_and_rpc.rst.txt b/docs/_sources/tutorial/cross_compilation_and_rpc.rst.txt
index c1defdb20d..3734300c53 100644
--- a/docs/_sources/tutorial/cross_compilation_and_rpc.rst.txt
+++ b/docs/_sources/tutorial/cross_compilation_and_rpc.rst.txt
@@ -270,7 +270,7 @@ device and returns the measured cost. Network overhead is excluded.
.. code-block:: none
- 1.217e-07 secs/op
+ 1.137e-07 secs/op
diff --git a/docs/_sources/tutorial/intro_topi.rst.txt b/docs/_sources/tutorial/intro_topi.rst.txt
index 5eaea073ad..10a4589546 100644
--- a/docs/_sources/tutorial/intro_topi.rst.txt
+++ b/docs/_sources/tutorial/intro_topi.rst.txt
@@ -95,14 +95,13 @@ and to examine the IR code in human readable format, we can do
@main = primfn(A_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [(stride: int32*n: int32)], [], type="auto")}
- buffer_map = {A_1: A}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [n, m: int32], [stride, stride_1: int32], type="auto")} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [n: int32, m: int32], [stride: int32, stride_1: int32], type="auto")}
+ buffer_map = {A_1: A} {
allocate(B: Pointer(global float32), float32, [n]), storage_scope = global;
for (i: int32, 0, n) {
B_1: Buffer(B, float32, [n], [])[i] = 0f32
for (k: int32, 0, m) {
- B_1[i] = (B_1[i] + A[((i*stride) + (k*stride_1))])
+ B_1[i] = (B_1[i] + A_3: Buffer(A_2, float32, [(stride*n)], [], type="auto")[((i*stride) + (k*stride_1))])
}
}
}
@@ -137,14 +136,13 @@ Fortunately, we can replace those two lines with simple :code:`topi.sum` much li
@main = primfn(A_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [(stride: int32*n: int32)], [], type="auto")}
- buffer_map = {A_1: A}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [n, m: int32], [stride, stride_1: int32], type="auto")} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [n: int32, m: int32], [stride: int32, stride_1: int32], type="auto")}
+ buffer_map = {A_1: A} {
allocate(A_red: Pointer(global float32), float32, [n]), storage_scope = global;
for (ax0: int32, 0, n) {
A_red_1: Buffer(A_red, float32, [n], [])[ax0] = 0f32
for (k1: int32, 0, m) {
- A_red_1[ax0] = (A_red_1[ax0] + A[((ax0*stride) + (k1*stride_1))])
+ A_red_1[ax0] = (A_red_1[ax0] + A_3: Buffer(A_2, float32, [(stride*n)], [], type="auto")[((ax0*stride) + (k1*stride_1))])
}
}
}
@@ -217,10 +215,9 @@ we can schedule the following series of operations ending with :code:`topi.sum`
warnings.warn("Try specifying cuda arch by adding 'arch=sm_xx' to your target.")
@main = primfn(a_1: handle, b_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {a: Buffer(a_2: Pointer(float32), float32, [10000], []),
- b: Buffer(b_2: Pointer(float32), float32, [100], [])}
- buffer_map = {a_1: a, b_1: b}
- preflattened_buffer_map = {a_1: a_3: Buffer(a_2, float32, [100, 10, 10], []), b_1: b_3: Buffer(b_2, float32, [10, 10], [])} {
+ buffers = {a: Buffer(a_2: Pointer(float32), float32, [100, 10, 10], []),
+ b: Buffer(b_2: Pointer(float32), float32, [10, 10], [])}
+ buffer_map = {a_1: a, b_1: b} {
allocate(T_divide_red: Pointer(global float32), float32, [1]), storage_scope = global;
attr [IterVar(threadIdx.x: int32, [0:1024], "ThreadIndex", "threadIdx.x")] "thread_extent" = 1024;
allocate(T_divide_red.rf: Pointer(local float32), float32, [1]), storage_scope = local;
@@ -228,7 +225,7 @@ we can schedule the following series of operations ending with :code:`topi.sum`
T_divide_red.rf_1: Buffer(T_divide_red.rf, float32, [1], [], scope="local", align=4)[0] = 0f32
for (k0.k1.fused.k2.fused.outer: int32, 0, 10) {
if @tir.likely((((((k0.k1.fused.k2.fused.outer*64) + floordiv(threadIdx.x, 16)) < 625) && (((k0.k1.fused.k2.fused.outer*64) + floordiv(threadIdx.x, 16)) < 625)) && (((k0.k1.fused.k2.fused.outer*64) + floordiv(threadIdx.x, 16)) < 625)), dtype=bool) {
- T_divide_red.rf_1[0] = (T_divide_red.rf_1[0] + (((a[((k0.k1.fused.k2.fused.outer*1024) + threadIdx.x)] + b[((floordiv(floormod(((k0.k1.fused.k2.fused.outer*12) + floordiv(threadIdx.x, 2)), 50), 5)*10) + floormod(((k0.k1.fused.k2.fused.outer*4) + threadIdx.x), 10))]) + (a[((k0.k1.fused.k2.fused.outer*1024) + threadIdx.x)]*b[((floordiv(floormod(((k0.k1.fused.k2.fused.outer*12) + floordiv(threadIdx.x, 2)), 50), 5)*10) + floormod(((k0.k1.fused.k2.fused.outer*4) + threadIdx.x), 10 [...]
+ T_divide_red.rf_1[0] = (T_divide_red.rf_1[0] + (((a_3: Buffer(a_2, float32, [10000], [])[((k0.k1.fused.k2.fused.outer*1024) + threadIdx.x)] + b_3: Buffer(b_2, float32, [100], [])[((floordiv(floormod(((k0.k1.fused.k2.fused.outer*12) + floordiv(threadIdx.x, 2)), 50), 5)*10) + floormod(((k0.k1.fused.k2.fused.outer*4) + threadIdx.x), 10))]) + (a_3[((k0.k1.fused.k2.fused.outer*1024) + threadIdx.x)]*b_3[((floordiv(floormod(((k0.k1.fused.k2.fused.outer*12) + floordiv(threadIdx.x, 2) [...]
}
}
attr [meta[tir.CommReducer][0]] "reduce_scope" = @tir.reinterpret(0u64, dtype=handle);
@@ -263,7 +260,7 @@ As you can see, scheduled stages of computation have been accumulated and we can
.. code-block:: none
- [stage(a, placeholder(a, 0x216194d0)), stage(b, placeholder(b, 0xecb7f90)), stage(T_add, compute(T_add, body=[(a[ax0, ax1, ax2] + b[ax1, ax2])], axis=[iter_var(ax0, range(min=0, ext=100)), iter_var(ax1, range(min=0, ext=10)), iter_var(ax2, range(min=0, ext=10))], reduce_axis=[], tag=broadcast, attrs={})), stage(T_multiply, compute(T_multiply, body=[(a[ax0, ax1, ax2]*b[ax1, ax2])], axis=[iter_var(ax0, range(min=0, ext=100)), iter_var(ax1, range(min=0, ext=10)), iter_var(ax2, range(min [...]
+ [stage(a, placeholder(a, 0x2134c3f0)), stage(b, placeholder(b, 0x21ee7aa0)), stage(T_add, compute(T_add, body=[(a[ax0, ax1, ax2] + b[ax1, ax2])], axis=[iter_var(ax0, range(min=0, ext=100)), iter_var(ax1, range(min=0, ext=10)), iter_var(ax2, range(min=0, ext=10))], reduce_axis=[], tag=broadcast, attrs={})), stage(T_multiply, compute(T_multiply, body=[(a[ax0, ax1, ax2]*b[ax1, ax2])], axis=[iter_var(ax0, range(min=0, ext=100)), iter_var(ax1, range(min=0, ext=10)), iter_var(ax2, range(mi [...]
@@ -320,9 +317,8 @@ TOPI also provides common neural nets operations such as _softmax_ with optimize
@main = primfn(tarray_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {tarray: Buffer(tarray_2: Pointer(float32), float32, [262144], [])}
- buffer_map = {tarray_1: tarray}
- preflattened_buffer_map = {tarray_1: tarray_3: Buffer(tarray_2, float32, [512, 512], [])} {
+ buffers = {tarray: Buffer(tarray_2: Pointer(float32), float32, [512, 512], [])}
+ buffer_map = {tarray_1: tarray} {
allocate(T_softmax_norm: Pointer(global float32x4), float32x4, [65536]), storage_scope = global;
attr [IterVar(blockIdx.x: int32, (nullptr), "ThreadIndex", "blockIdx.x")] "thread_extent" = 512;
allocate(normal_reduce_temp0: Pointer(local float32), float32, [1]), storage_scope = local;
@@ -333,13 +329,13 @@ TOPI also provides common neural nets operations such as _softmax_ with optimize
attr [IterVar(threadIdx.x: int32, [0:32], "ThreadIndex", "threadIdx.x")] "thread_extent" = 32 {
normal_reduce_temp0_2: Buffer(normal_reduce_temp0, float32, [1], [], scope="local")[0] = -3.40282e+38f32
for (k.inner: int32, 0, 16) {
- normal_reduce_temp0_2[0] = max(normal_reduce_temp0_2[0], tarray[(((blockIdx.x*512) + (threadIdx.x*16)) + k.inner)])
+ normal_reduce_temp0_2[0] = max(normal_reduce_temp0_2[0], tarray_3: Buffer(tarray_2, float32, [262144], [])[(((blockIdx.x*512) + (threadIdx.x*16)) + k.inner)])
}
attr [meta[tir.CommReducer][0]] "reduce_scope" = @tir.reinterpret(0u64, dtype=handle);
@tir.tvm_thread_allreduce(1u32, normal_reduce_temp0_2[0], True, reduce_temp0_2: Buffer(reduce_temp0, float32, [1], [], scope="local")[0], threadIdx.x, dtype=handle)
for (i1.inner.outer: int32, 0, 4) {
let cse_var_1: int32 = (i1.inner.outer*4)
- T_softmax_exp_1: Buffer(T_softmax_exp, float32, [512], [], scope="warp")[ramp(((threadIdx.x*16) + cse_var_1), 1, 4)] = @tir.exp((tarray[ramp((((blockIdx.x*512) + (threadIdx.x*16)) + cse_var_1), 1, 4)] - broadcast(reduce_temp0_3: Buffer(reduce_temp0, float32, [1], [], scope="local", align=4)[0], 4)), dtype=float32x4)
+ T_softmax_exp_1: Buffer(T_softmax_exp, float32, [512], [], scope="warp")[ramp(((threadIdx.x*16) + cse_var_1), 1, 4)] = @tir.exp((tarray_3[ramp((((blockIdx.x*512) + (threadIdx.x*16)) + cse_var_1), 1, 4)] - broadcast(reduce_temp0_3: Buffer(reduce_temp0, float32, [1], [], scope="local", align=4)[0], 4)), dtype=float32x4)
}
}
attr [IterVar(threadIdx.x, [0:32], "ThreadIndex", "threadIdx.x")] "thread_extent" = 32 {
@@ -399,10 +395,9 @@ We can fuse :code:`topi.nn.conv2d` and :code:`topi.nn.relu` together.
@main = primfn(placeholder_2: handle, placeholder_3: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {placeholder: Buffer(placeholder_4: Pointer(float32), float32, [150528], []),
- placeholder_1: Buffer(placeholder_5: Pointer(float32), float32, [750], [])}
- buffer_map = {placeholder_2: placeholder, placeholder_3: placeholder_1}
- preflattened_buffer_map = {placeholder_2: placeholder_6: Buffer(placeholder_4, float32, [1, 3, 224, 224], []), placeholder_3: placeholder_7: Buffer(placeholder_5, float32, [10, 3, 5, 5], [])} {
+ buffers = {placeholder: Buffer(placeholder_4: Pointer(float32), float32, [1, 3, 224, 224], []),
+ placeholder_1: Buffer(placeholder_5: Pointer(float32), float32, [10, 3, 5, 5], [])}
+ buffer_map = {placeholder_2: placeholder, placeholder_3: placeholder_1} {
allocate(compute: Pointer(global float32), float32, [501760]), storage_scope = global;
attr [IterVar(blockIdx.z: int32, (nullptr), "ThreadIndex", "blockIdx.z")] "thread_extent" = 5;
allocate(conv2d_nchw: Pointer(local float32), float32, [14]), storage_scope = local;
@@ -432,19 +427,19 @@ We can fuse :code:`topi.nn.conv2d` and :code:`topi.nn.relu` together.
attr [IterVar(threadIdx.z_1: int32, (nullptr), "ThreadIndex", "threadIdx.z")] "thread_extent" = 1;
attr [IterVar(threadIdx.y_1: int32, (nullptr), "ThreadIndex", "threadIdx.y")] "thread_extent" = 1;
attr [IterVar(threadIdx.x_1: int32, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 16 {
- pad_temp.shared_1: Buffer(pad_temp.shared, float32, [112], [], scope="shared")[(threadIdx.x_1*7)] = @tir.if_then_else((((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)) && (1 <= ((blockIdx.x*56) + floordiv((threadIdx.x_1*7), 2)))), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 450)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 1)] = @tir.if_then_else((((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)) && (1 <= ((blockIdx.x*56) + floordiv(((threadIdx.x_1*7) + 1), 2)))), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 449)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 2)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 448)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 3)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 447)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 4)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 446)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 5)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 445)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 6)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 444)], 0f32, dtype=float32)
+ pad_temp.shared_1: Buffer(pad_temp.shared, float32, [112], [], scope="shared")[(threadIdx.x_1*7)] = @tir.if_then_else((((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)) && (1 <= ((blockIdx.x*56) + floordiv((threadIdx.x_1*7), 2)))), placeholder_6: Buffer(placeholder_4, float32, [150528], [])[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 450)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 1)] = @tir.if_then_else((((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)) && (1 <= ((blockIdx.x*56) + floordiv(((threadIdx.x_1*7) + 1), 2)))), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 449)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 2)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 448)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 3)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 447)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 4)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 446)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 5)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 445)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 6)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 444)], 0f32, dtype=float32)
}
attr [IterVar(threadIdx.z_2: int32, (nullptr), "ThreadIndex", "threadIdx.z")] "thread_extent" = 1;
attr [IterVar(threadIdx.y_2: int32, (nullptr), "ThreadIndex", "threadIdx.y")] "thread_extent" = 1;
attr [IterVar(threadIdx.x_2: int32, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 16;
if @tir.likely((threadIdx.x_2 < 2), dtype=bool) {
- placeholder.shared_1: Buffer(placeholder.shared, float32, [2], [], scope="shared", align=8)[threadIdx.x_2] = placeholder_1[((((blockIdx.z*150) + (threadIdx.x_2*75)) + (rc.outer*25)) + (ry.outer*5))]
+ placeholder.shared_1: Buffer(placeholder.shared, float32, [2], [], scope="shared", align=8)[threadIdx.x_2] = placeholder_7: Buffer(placeholder_5, float32, [750], [])[((((blockIdx.z*150) + (threadIdx.x_2*75)) + (rc.outer*25)) + (ry.outer*5))]
}
conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[threadIdx.x]*placeholder.shared_1[0]))
conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(threadIdx.x + 16)]*placeholder.shared_1[0]))
@@ -463,19 +458,19 @@ We can fuse :code:`topi.nn.conv2d` and :code:`topi.nn.relu` together.
attr [IterVar(threadIdx.z_1, (nullptr), "ThreadIndex", "threadIdx.z")] "thread_extent" = 1;
attr [IterVar(threadIdx.y_1, (nullptr), "ThreadIndex", "threadIdx.y")] "thread_extent" = 1;
attr [IterVar(threadIdx.x_1, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 16 {
- pad_temp.shared_1[(threadIdx.x_1*7)] = @tir.if_then_else((((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)) && (1 <= ((blockIdx.x*56) + floordiv(((threadIdx.x_1*7) + 1), 2)))), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 449)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 1)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 448)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 2)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 447)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 3)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 446)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 4)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 445)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 5)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 444)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 6)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 443)], 0f32, dtype=float32)
+ pad_temp.shared_1[(threadIdx.x_1*7)] = @tir.if_then_else((((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)) && (1 <= ((blockIdx.x*56) + floordiv(((threadIdx.x_1*7) + 1), 2)))), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 449)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 1)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 448)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 2)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 447)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 3)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 446)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 4)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 445)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 5)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 444)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 6)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 443)], 0f32, dtype=float32)
}
attr [IterVar(threadIdx.z_2, (nullptr), "ThreadIndex", "threadIdx.z")] "thread_extent" = 1;
attr [IterVar(threadIdx.y_2, (nullptr), "ThreadIndex", "threadIdx.y")] "thread_extent" = 1;
attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 16;
if @tir.likely((threadIdx.x_2 < 2), dtype=bool) {
- placeholder.shared_1[threadIdx.x_2] = placeholder_1[(((((blockIdx.z*150) + (threadIdx.x_2*75)) + (rc.outer*25)) + (ry.outer*5)) + 1)]
+ placeholder.shared_1[threadIdx.x_2] = placeholder_7[(((((blockIdx.z*150) + (threadIdx.x_2*75)) + (rc.outer*25)) + (ry.outer*5)) + 1)]
}
conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[threadIdx.x]*placeholder.shared_1[0]))
conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(threadIdx.x + 16)]*placeholder.shared_1[0]))
@@ -494,19 +489,19 @@ We can fuse :code:`topi.nn.conv2d` and :code:`topi.nn.relu` together.
attr [IterVar(threadIdx.z_1, (nullptr), "ThreadIndex", "threadIdx.z")] "thread_extent" = 1;
attr [IterVar(threadIdx.y_1, (nullptr), "ThreadIndex", "threadIdx.y")] "thread_extent" = 1;
attr [IterVar(threadIdx.x_1, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 16 {
- pad_temp.shared_1[(threadIdx.x_1*7)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 448)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 1)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 447)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 2)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 446)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 3)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 445)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 4)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 444)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 5)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 443)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 6)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 442)], 0f32, dtype=float32)
+ pad_temp.shared_1[(threadIdx.x_1*7)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 448)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 1)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 447)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 2)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 446)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 3)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 445)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 4)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 444)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 5)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 443)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 6)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 442)], 0f32, dtype=float32)
}
attr [IterVar(threadIdx.z_2, (nullptr), "ThreadIndex", "threadIdx.z")] "thread_extent" = 1;
attr [IterVar(threadIdx.y_2, (nullptr), "ThreadIndex", "threadIdx.y")] "thread_extent" = 1;
attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 16;
if @tir.likely((threadIdx.x_2 < 2), dtype=bool) {
- placeholder.shared_1[threadIdx.x_2] = placeholder_1[(((((blockIdx.z*150) + (threadIdx.x_2*75)) + (rc.outer*25)) + (ry.outer*5)) + 2)]
+ placeholder.shared_1[threadIdx.x_2] = placeholder_7[(((((blockIdx.z*150) + (threadIdx.x_2*75)) + (rc.outer*25)) + (ry.outer*5)) + 2)]
}
conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[threadIdx.x]*placeholder.shared_1[0]))
conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(threadIdx.x + 16)]*placeholder.shared_1[0]))
@@ -525,19 +520,19 @@ We can fuse :code:`topi.nn.conv2d` and :code:`topi.nn.relu` together.
attr [IterVar(threadIdx.z_1, (nullptr), "ThreadIndex", "threadIdx.z")] "thread_extent" = 1;
attr [IterVar(threadIdx.y_1, (nullptr), "ThreadIndex", "threadIdx.y")] "thread_extent" = 1;
attr [IterVar(threadIdx.x_1, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 16 {
- pad_temp.shared_1[(threadIdx.x_1*7)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 447)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 1)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 446)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 2)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 445)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 3)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 444)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 4)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 443)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 5)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 442)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 6)] = @tir.if_then_else((((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)) && (((blockIdx.x*56) + floordiv(((threadIdx.x_1*7) + 9), 2)) < 113)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 441)], 0f32, dtype=float32)
+ pad_temp.shared_1[(threadIdx.x_1*7)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 447)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 1)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 446)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 2)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 445)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 3)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 444)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 4)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 443)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 5)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 442)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 6)] = @tir.if_then_else((((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)) && (((blockIdx.x*56) + floordiv(((threadIdx.x_1*7) + 9), 2)) < 113)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 441)], 0f32, dtype=float32)
}
attr [IterVar(threadIdx.z_2, (nullptr), "ThreadIndex", "threadIdx.z")] "thread_extent" = 1;
attr [IterVar(threadIdx.y_2, (nullptr), "ThreadIndex", "threadIdx.y")] "thread_extent" = 1;
attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 16;
if @tir.likely((threadIdx.x_2 < 2), dtype=bool) {
- placeholder.shared_1[threadIdx.x_2] = placeholder_1[(((((blockIdx.z*150) + (threadIdx.x_2*75)) + (rc.outer*25)) + (ry.outer*5)) + 3)]
+ placeholder.shared_1[threadIdx.x_2] = placeholder_7[(((((blockIdx.z*150) + (threadIdx.x_2*75)) + (rc.outer*25)) + (ry.outer*5)) + 3)]
}
conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[threadIdx.x]*placeholder.shared_1[0]))
conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(threadIdx.x + 16)]*placeholder.shared_1[0]))
@@ -556,19 +551,19 @@ We can fuse :code:`topi.nn.conv2d` and :code:`topi.nn.relu` together.
attr [IterVar(threadIdx.z_1, (nullptr), "ThreadIndex", "threadIdx.z")] "thread_extent" = 1;
attr [IterVar(threadIdx.y_1, (nullptr), "ThreadIndex", "threadIdx.y")] "thread_extent" = 1;
attr [IterVar(threadIdx.x_1, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 16 {
- pad_temp.shared_1[(threadIdx.x_1*7)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 446)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 1)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 445)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 2)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 444)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 3)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 443)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 4)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 442)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 5)] = @tir.if_then_else((((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)) && (((blockIdx.x*56) + floordiv(((threadIdx.x_1*7) + 9), 2)) < 113)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 441)], 0f32, dtype=float32)
- pad_temp.shared_1[((threadIdx.x_1*7) + 6)] = @tir.if_then_else((((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)) && (((blockIdx.x*56) + floordiv((threadIdx.x_1*7), 2)) < 108)), placeholder[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 440)], 0f32, dtype=float32)
+ pad_temp.shared_1[(threadIdx.x_1*7)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 446)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 1)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 445)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 2)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 444)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 3)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 443)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 4)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 442)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 5)] = @tir.if_then_else((((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)) && (((blockIdx.x*56) + floordiv(((threadIdx.x_1*7) + 9), 2)) < 113)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 441)], 0f32, dtype=float32)
+ pad_temp.shared_1[((threadIdx.x_1*7) + 6)] = @tir.if_then_else((((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)) && (((blockIdx.x*56) + floordiv((threadIdx.x_1*7), 2)) < 108)), placeholder_6[((((((rc.outer*50176) + (blockIdx.y*224)) + (ry.outer*224)) + (blockIdx.x*112)) + (threadIdx.x_1*7)) - 440)], 0f32, dtype=float32)
}
attr [IterVar(threadIdx.z_2, (nullptr), "ThreadIndex", "threadIdx.z")] "thread_extent" = 1;
attr [IterVar(threadIdx.y_2, (nullptr), "ThreadIndex", "threadIdx.y")] "thread_extent" = 1;
attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 16;
if @tir.likely((threadIdx.x_2 < 2), dtype=bool) {
- placeholder.shared_1[threadIdx.x_2] = placeholder_1[(((((blockIdx.z*150) + (threadIdx.x_2*75)) + (rc.outer*25)) + (ry.outer*5)) + 4)]
+ placeholder.shared_1[threadIdx.x_2] = placeholder_7[(((((blockIdx.z*150) + (threadIdx.x_2*75)) + (rc.outer*25)) + (ry.outer*5)) + 4)]
}
conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[threadIdx.x]*placeholder.shared_1[0]))
conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(threadIdx.x + 16)]*placeholder.shared_1[0]))
diff --git a/docs/_sources/tutorial/sg_execution_times.rst.txt b/docs/_sources/tutorial/sg_execution_times.rst.txt
index aaae8baf85..51ea4ccf32 100644
--- a/docs/_sources/tutorial/sg_execution_times.rst.txt
+++ b/docs/_sources/tutorial/sg_execution_times.rst.txt
@@ -5,32 +5,32 @@
Computation times
=================
-**14:16.601** total execution time for **tutorial** files:
+**14:44.179** total execution time for **tutorial** files:
+------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_tutorial_autotvm_relay_x86.py` (``autotvm_relay_x86.py``) | 10:59.617 | 0.0 MB |
+| :ref:`sphx_glr_tutorial_autotvm_relay_x86.py` (``autotvm_relay_x86.py``) | 11:09.688 | 0.0 MB |
+------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_tutorial_auto_scheduler_matmul_x86.py` (``auto_scheduler_matmul_x86.py``) | 01:12.406 | 0.0 MB |
+| :ref:`sphx_glr_tutorial_auto_scheduler_matmul_x86.py` (``auto_scheduler_matmul_x86.py``) | 01:33.236 | 0.0 MB |
+------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_tutorial_tensor_expr_get_started.py` (``tensor_expr_get_started.py``) | 01:01.139 | 0.0 MB |
+| :ref:`sphx_glr_tutorial_tensor_expr_get_started.py` (``tensor_expr_get_started.py``) | 01:02.680 | 0.0 MB |
+------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_tutorial_relay_quick_start.py` (``relay_quick_start.py``) | 00:35.812 | 0.0 MB |
+| :ref:`sphx_glr_tutorial_relay_quick_start.py` (``relay_quick_start.py``) | 00:38.658 | 0.0 MB |
+------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_tutorial_autotvm_matmul_x86.py` (``autotvm_matmul_x86.py``) | 00:26.055 | 0.0 MB |
+| :ref:`sphx_glr_tutorial_autotvm_matmul_x86.py` (``autotvm_matmul_x86.py``) | 00:18.300 | 0.0 MB |
+------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_tutorial_intro_topi.py` (``intro_topi.py``) | 00:00.766 | 0.0 MB |
+| :ref:`sphx_glr_tutorial_intro_topi.py` (``intro_topi.py``) | 00:00.810 | 0.0 MB |
+------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_tutorial_tensor_ir_blitz_course.py` (``tensor_ir_blitz_course.py``) | 00:00.637 | 0.0 MB |
+| :ref:`sphx_glr_tutorial_tensor_ir_blitz_course.py` (``tensor_ir_blitz_course.py``) | 00:00.644 | 0.0 MB |
+------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_tutorial_cross_compilation_and_rpc.py` (``cross_compilation_and_rpc.py``) | 00:00.160 | 0.0 MB |
+| :ref:`sphx_glr_tutorial_cross_compilation_and_rpc.py` (``cross_compilation_and_rpc.py``) | 00:00.154 | 0.0 MB |
+------------------------------------------------------------------------------------------+-----------+--------+
| :ref:`sphx_glr_tutorial_introduction.py` (``introduction.py``) | 00:00.005 | 0.0 MB |
+------------------------------------------------------------------------------------------+-----------+--------+
| :ref:`sphx_glr_tutorial_uma.py` (``uma.py``) | 00:00.002 | 0.0 MB |
+------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_tutorial_tvmc_python.py` (``tvmc_python.py``) | 00:00.001 | 0.0 MB |
-+------------------------------------------------------------------------------------------+-----------+--------+
| :ref:`sphx_glr_tutorial_tvmc_command_line_driver.py` (``tvmc_command_line_driver.py``) | 00:00.001 | 0.0 MB |
+------------------------------------------------------------------------------------------+-----------+--------+
+| :ref:`sphx_glr_tutorial_tvmc_python.py` (``tvmc_python.py``) | 00:00.001 | 0.0 MB |
++------------------------------------------------------------------------------------------+-----------+--------+
| :ref:`sphx_glr_tutorial_install.py` (``install.py``) | 00:00.001 | 0.0 MB |
+------------------------------------------------------------------------------------------+-----------+--------+
diff --git a/docs/_sources/tutorial/tensor_expr_get_started.rst.txt b/docs/_sources/tutorial/tensor_expr_get_started.rst.txt
index 95c3c718c4..604eeab4e7 100644
--- a/docs/_sources/tutorial/tensor_expr_get_started.rst.txt
+++ b/docs/_sources/tutorial/tensor_expr_get_started.rst.txt
@@ -294,7 +294,7 @@ helper function to run a profile of the TVM generated code.
.. code-block:: none
- Numpy running time: 0.000007
+ Numpy running time: 0.000008
naive: 0.000007
@@ -354,13 +354,12 @@ return a readable C-style statement.
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [(stride: int32*n: int32)], [], type="auto"),
- B: Buffer(B_2: Pointer(float32), float32, [(stride_1: int32*n)], [], type="auto"),
- C: Buffer(C_2: Pointer(float32), float32, [(stride_2: int32*n)], [], type="auto")}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [n], [stride], type="auto"), B_1: B_3: Buffer(B_2, float32, [n], [stride_1], type="auto"), C_1: C_3: Buffer(C_2, float32, [n], [stride_2], type="auto")} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [n: int32], [stride: int32], type="auto"),
+ B: Buffer(B_2: Pointer(float32), float32, [n], [stride_1: int32], type="auto"),
+ C: Buffer(C_2: Pointer(float32), float32, [n], [stride_2: int32], type="auto")}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
for (i: int32, 0, n) "parallel" {
- C[(i*stride_2)] = (A[(i*stride)] + B[(i*stride_1)])
+ C_3: Buffer(C_2, float32, [(stride_2*n)], [], type="auto")[(i*stride_2)] = (A_3: Buffer(A_2, float32, [(stride*n)], [], type="auto")[(i*stride)] + B_3: Buffer(B_2, float32, [(stride_1*n)], [], type="auto")[(i*stride_1)])
}
}
@@ -394,7 +393,7 @@ compile and run this new schedule with the parallel operation applied:
.. code-block:: none
- parallel: 0.000007
+ parallel: 0.000008
@@ -452,16 +451,15 @@ factor to be the number of threads on your CPU.
vector: 0.000025
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [(stride: int32*n: int32)], [], type="auto"),
- B: Buffer(B_2: Pointer(float32), float32, [(stride_1: int32*n)], [], type="auto"),
- C: Buffer(C_2: Pointer(float32), float32, [(stride_2: int32*n)], [], type="auto")}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [n], [stride], type="auto"), B_1: B_3: Buffer(B_2, float32, [n], [stride_1], type="auto"), C_1: C_3: Buffer(C_2, float32, [n], [stride_2], type="auto")} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [n: int32], [stride: int32], type="auto"),
+ B: Buffer(B_2: Pointer(float32), float32, [n], [stride_1: int32], type="auto"),
+ C: Buffer(C_2: Pointer(float32), float32, [n], [stride_2: int32], type="auto")}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
for (i.outer: int32, 0, floordiv((n + 3), 4)) "parallel" {
for (i.inner.s: int32, 0, 4) {
if @tir.likely((((i.outer*4) + i.inner.s) < n), dtype=bool) {
let cse_var_1: int32 = ((i.outer*4) + i.inner.s)
- C[(cse_var_1*stride_2)] = (A[(cse_var_1*stride)] + B[(cse_var_1*stride_1)])
+ C_3: Buffer(C_2, float32, [(stride_2*n)], [], type="auto")[(cse_var_1*stride_2)] = (A_3: Buffer(A_2, float32, [(stride*n)], [], type="auto")[(cse_var_1*stride)] + B_3: Buffer(B_2, float32, [(stride_1*n)], [], type="auto")[(cse_var_1*stride_1)])
}
}
}
@@ -501,10 +499,10 @@ We can now compare the different schedules
.. code-block:: none
Operator Timing Performance
- numpy 7.098759997461457e-06 1.0
- naive 6.6883e-06 0.9421786343518813
- parallel 6.9717e-06 0.9821010996981312
- vector 2.4533900000000002e-05 3.456082471977277
+ numpy 7.648709997738479e-06 1.0
+ naive 6.6648e-06 0.8713626221899647
+ parallel 8.3591e-06 1.092877099860181
+ vector 2.4603500000000003e-05 3.216686213397371
@@ -925,7 +923,7 @@ matrix multiplication.
.. code-block:: none
- Numpy running time: 0.018547
+ Numpy running time: 0.018212
@@ -983,7 +981,7 @@ optimizations.
.. code-block:: none
- none: 3.401293
+ none: 3.540408
@@ -1012,18 +1010,17 @@ nested loops over the indices of the A and B matrices.
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [1048576], []),
- B: Buffer(B_2: Pointer(float32), float32, [1048576], []),
- C: Buffer(C_2: Pointer(float32), float32, [1048576], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 1024], []), B_1: B_3: Buffer(B_2, float32, [1024, 1024], []), C_1: C_3: Buffer(C_2, float32, [1024, 1024], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),
+ B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
for (x: int32, 0, 1024) {
for (y: int32, 0, 1024) {
- C[((x*1024) + y)] = 0f32
+ C_3: Buffer(C_2, float32, [1048576], [])[((x*1024) + y)] = 0f32
for (k: int32, 0, 1024) {
let cse_var_2: int32 = (x*1024)
let cse_var_1: int32 = (cse_var_2 + y)
- C[cse_var_1] = (C[cse_var_1] + (A[(cse_var_2 + k)]*B[((k*1024) + y)]))
+ C_3[cse_var_1] = (C_3[cse_var_1] + (A_3: Buffer(A_2, float32, [1048576], [])[(cse_var_2 + k)]*B_3: Buffer(B_2, float32, [1048576], [])[((k*1024) + y)]))
}
}
}
@@ -1086,7 +1083,7 @@ schedule.
.. code-block:: none
- blocking: 0.313142
+ blocking: 0.299061
@@ -1114,16 +1111,15 @@ internal representation and compare it to the original:
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [1048576], []),
- B: Buffer(B_2: Pointer(float32), float32, [1048576], []),
- C: Buffer(C_2: Pointer(float32), float32, [1048576], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 1024], []), B_1: B_3: Buffer(B_2, float32, [1024, 1024], []), C_1: C_3: Buffer(C_2, float32, [1024, 1024], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),
+ B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
for (x.outer: int32, 0, 32) {
for (y.outer: int32, 0, 32) {
for (x.inner.init: int32, 0, 32) {
for (y.inner.init: int32, 0, 32) {
- C[((((x.outer*32768) + (x.inner.init*1024)) + (y.outer*32)) + y.inner.init)] = 0f32
+ C_3: Buffer(C_2, float32, [1048576], [])[((((x.outer*32768) + (x.inner.init*1024)) + (y.outer*32)) + y.inner.init)] = 0f32
}
}
for (k.outer: int32, 0, 256) {
@@ -1133,7 +1129,7 @@ internal representation and compare it to the original:
let cse_var_3: int32 = (y.outer*32)
let cse_var_2: int32 = ((x.outer*32768) + (x.inner*1024))
let cse_var_1: int32 = ((cse_var_2 + cse_var_3) + y.inner)
- C[cse_var_1] = (C[cse_var_1] + (A[((cse_var_2 + (k.outer*4)) + k.inner)]*B[((((k.outer*4096) + (k.inner*1024)) + cse_var_3) + y.inner)]))
+ C_3[cse_var_1] = (C_3[cse_var_1] + (A_3: Buffer(A_2, float32, [1048576], [])[((cse_var_2 + (k.outer*4)) + k.inner)]*B_3: Buffer(B_2, float32, [1048576], [])[((((k.outer*4096) + (k.inner*1024)) + cse_var_3) + y.inner)]))
}
}
}
@@ -1182,18 +1178,17 @@ already cache friendly from our previous optimizations.
.. code-block:: none
- vectorization: 0.345544
+ vectorization: 0.337274
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [1048576], []),
- B: Buffer(B_2: Pointer(float32), float32, [1048576], []),
- C: Buffer(C_2: Pointer(float32), float32, [1048576], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 1024], []), B_1: B_3: Buffer(B_2, float32, [1024, 1024], []), C_1: C_3: Buffer(C_2, float32, [1024, 1024], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),
+ B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
for (x.outer: int32, 0, 32) {
for (y.outer: int32, 0, 32) {
for (x.inner.init: int32, 0, 32) {
- C[ramp((((x.outer*32768) + (x.inner.init*1024)) + (y.outer*32)), 1, 32)] = broadcast(0f32, 32)
+ C_3: Buffer(C_2, float32, [1048576], [])[ramp((((x.outer*32768) + (x.inner.init*1024)) + (y.outer*32)), 1, 32)] = broadcast(0f32, 32)
}
for (k.outer: int32, 0, 256) {
for (k.inner: int32, 0, 4) {
@@ -1201,7 +1196,7 @@ already cache friendly from our previous optimizations.
let cse_var_3: int32 = (y.outer*32)
let cse_var_2: int32 = ((x.outer*32768) + (x.inner*1024))
let cse_var_1: int32 = (cse_var_2 + cse_var_3)
- C[ramp(cse_var_1, 1, 32)] = (C[ramp(cse_var_1, 1, 32)] + (broadcast(A[((cse_var_2 + (k.outer*4)) + k.inner)], 32)*B[ramp((((k.outer*4096) + (k.inner*1024)) + cse_var_3), 1, 32)]))
+ C_3[ramp(cse_var_1, 1, 32)] = (C_3[ramp(cse_var_1, 1, 32)] + (broadcast(A_3: Buffer(A_2, float32, [1048576], [])[((cse_var_2 + (k.outer*4)) + k.inner)], 32)*B_3: Buffer(B_2, float32, [1048576], [])[ramp((((k.outer*4096) + (k.inner*1024)) + cse_var_3), 1, 32)]))
}
}
}
@@ -1256,18 +1251,17 @@ more cache friendly.
.. code-block:: none
- loop permutation: 0.115435
+ loop permutation: 0.118992
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [1048576], []),
- B: Buffer(B_2: Pointer(float32), float32, [1048576], []),
- C: Buffer(C_2: Pointer(float32), float32, [1048576], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 1024], []), B_1: B_3: Buffer(B_2, float32, [1024, 1024], []), C_1: C_3: Buffer(C_2, float32, [1024, 1024], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),
+ B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
for (x.outer: int32, 0, 32) {
for (y.outer: int32, 0, 32) {
for (x.inner.init: int32, 0, 32) {
- C[ramp((((x.outer*32768) + (x.inner.init*1024)) + (y.outer*32)), 1, 32)] = broadcast(0f32, 32)
+ C_3: Buffer(C_2, float32, [1048576], [])[ramp((((x.outer*32768) + (x.inner.init*1024)) + (y.outer*32)), 1, 32)] = broadcast(0f32, 32)
}
for (k.outer: int32, 0, 256) {
for (x.inner: int32, 0, 32) {
@@ -1275,7 +1269,7 @@ more cache friendly.
let cse_var_3: int32 = (y.outer*32)
let cse_var_2: int32 = ((x.outer*32768) + (x.inner*1024))
let cse_var_1: int32 = (cse_var_2 + cse_var_3)
- C[ramp(cse_var_1, 1, 32)] = (C[ramp(cse_var_1, 1, 32)] + (broadcast(A[((cse_var_2 + (k.outer*4)) + k.inner)], 32)*B[ramp((((k.outer*4096) + (k.inner*1024)) + cse_var_3), 1, 32)]))
+ C_3[ramp(cse_var_1, 1, 32)] = (C_3[ramp(cse_var_1, 1, 32)] + (broadcast(A_3: Buffer(A_2, float32, [1048576], [])[((cse_var_2 + (k.outer*4)) + k.inner)], 32)*B_3: Buffer(B_2, float32, [1048576], [])[ramp((((k.outer*4096) + (k.inner*1024)) + cse_var_3), 1, 32)]))
}
}
}
@@ -1355,24 +1349,23 @@ optimized schedule.
.. code-block:: none
- array packing: 0.109588
+ array packing: 0.109797
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [1048576], []),
- B: Buffer(B_2: Pointer(float32), float32, [1048576], []),
- C: Buffer(C_2: Pointer(float32), float32, [1048576], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 1024], []), B_1: B_3: Buffer(B_2, float32, [1024, 1024], []), C_1: C_3: Buffer(C_2, float32, [1024, 1024], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),
+ B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
allocate(packedB: Pointer(global float32x32), float32x32, [32768]), storage_scope = global {
for (x: int32, 0, 32) "parallel" {
for (y: int32, 0, 1024) {
- packedB_1: Buffer(packedB, float32x32, [32768], [])[((x*1024) + y)] = B[ramp(((y*1024) + (x*32)), 1, 32)]
+ packedB_1: Buffer(packedB, float32x32, [32768], [])[((x*1024) + y)] = B_3: Buffer(B_2, float32, [1048576], [])[ramp(((y*1024) + (x*32)), 1, 32)]
}
}
for (x.outer: int32, 0, 32) {
for (y.outer: int32, 0, 32) {
for (x.inner.init: int32, 0, 32) {
- C[ramp((((x.outer*32768) + (x.inner.init*1024)) + (y.outer*32)), 1, 32)] = broadcast(0f32, 32)
+ C_3: Buffer(C_2, float32, [1048576], [])[ramp((((x.outer*32768) + (x.inner.init*1024)) + (y.outer*32)), 1, 32)] = broadcast(0f32, 32)
}
for (k.outer: int32, 0, 256) {
for (x.inner: int32, 0, 32) {
@@ -1380,7 +1373,7 @@ optimized schedule.
let cse_var_3: int32 = ((x.outer*32768) + (x.inner*1024))
let cse_var_2: int32 = (k.outer*4)
let cse_var_1: int32 = (cse_var_3 + (y.outer*32))
- C[ramp(cse_var_1, 1, 32)] = (C[ramp(cse_var_1, 1, 32)] + (broadcast(A[((cse_var_3 + cse_var_2) + k.inner)], 32)*packedB_1[(((y.outer*1024) + cse_var_2) + k.inner)]))
+ C_3[ramp(cse_var_1, 1, 32)] = (C_3[ramp(cse_var_1, 1, 32)] + (broadcast(A_3: Buffer(A_2, float32, [1048576], [])[((cse_var_3 + cse_var_2) + k.inner)], 32)*packedB_1[(((y.outer*1024) + cse_var_2) + k.inner)]))
}
}
}
@@ -1448,19 +1441,18 @@ to `C` when all the block results are ready.
.. code-block:: none
- block caching: 0.110444
+ block caching: 0.110635
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [1048576], []),
- B: Buffer(B_2: Pointer(float32), float32, [1048576], []),
- C: Buffer(C_2: Pointer(float32), float32, [1048576], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 1024], []), B_1: B_3: Buffer(B_2, float32, [1024, 1024], []), C_1: C_3: Buffer(C_2, float32, [1024, 1024], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),
+ B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
allocate(packedB: Pointer(global float32x32), float32x32, [32768]), storage_scope = global;
allocate(C.global: Pointer(global float32), float32, [1024]), storage_scope = global {
for (x: int32, 0, 32) "parallel" {
for (y: int32, 0, 1024) {
- packedB_1: Buffer(packedB, float32x32, [32768], [])[((x*1024) + y)] = B[ramp(((y*1024) + (x*32)), 1, 32)]
+ packedB_1: Buffer(packedB, float32x32, [32768], [])[((x*1024) + y)] = B_3: Buffer(B_2, float32, [1048576], [])[ramp(((y*1024) + (x*32)), 1, 32)]
}
}
for (x.outer: int32, 0, 32) {
@@ -1475,16 +1467,16 @@ to `C` when all the block results are ready.
let cse_var_2: int32 = ((y.outer*1024) + cse_var_4)
let cse_var_1: int32 = (((x.outer*32768) + (x.c*1024)) + cse_var_4)
{
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[cse_var_1], 32)*packedB_1[cse_var_2]))
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[(cse_var_1 + 1)], 32)*packedB_1[(cse_var_2 + 1)]))
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[(cse_var_1 + 2)], 32)*packedB_1[(cse_var_2 + 2)]))
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[(cse_var_1 + 3)], 32)*packedB_1[(cse_var_2 + 3)]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3: Buffer(A_2, float32, [1048576], [])[cse_var_1], 32)*packedB_1[cse_var_2]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3[(cse_var_1 + 1)], 32)*packedB_1[(cse_var_2 + 1)]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3[(cse_var_1 + 2)], 32)*packedB_1[(cse_var_2 + 2)]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3[(cse_var_1 + 3)], 32)*packedB_1[(cse_var_2 + 3)]))
}
}
}
for (x.inner: int32, 0, 32) {
for (y.inner: int32, 0, 32) {
- C[((((x.outer*32768) + (x.inner*1024)) + (y.outer*32)) + y.inner)] = C.global_1[((x.inner*32) + y.inner)]
+ C_3: Buffer(C_2, float32, [1048576], [])[((((x.outer*32768) + (x.inner*1024)) + (y.outer*32)) + y.inner)] = C.global_1[((x.inner*32) + y.inner)]
}
}
}
@@ -1534,18 +1526,17 @@ of thread-level parallelization.
.. code-block:: none
- parallelization: 0.146092
+ parallelization: 0.144369
@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [1048576], []),
- B: Buffer(B_2: Pointer(float32), float32, [1048576], []),
- C: Buffer(C_2: Pointer(float32), float32, [1048576], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 1024], []), B_1: B_3: Buffer(B_2, float32, [1024, 1024], []), C_1: C_3: Buffer(C_2, float32, [1024, 1024], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),
+ B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
allocate(packedB: Pointer(global float32x32), float32x32, [32768]), storage_scope = global {
for (x: int32, 0, 32) "parallel" {
for (y: int32, 0, 1024) {
- packedB_1: Buffer(packedB, float32x32, [32768], [])[((x*1024) + y)] = B[ramp(((y*1024) + (x*32)), 1, 32)]
+ packedB_1: Buffer(packedB, float32x32, [32768], [])[((x*1024) + y)] = B_3: Buffer(B_2, float32, [1048576], [])[ramp(((y*1024) + (x*32)), 1, 32)]
}
}
for (x.outer: int32, 0, 32) "parallel" {
@@ -1561,16 +1552,16 @@ of thread-level parallelization.
let cse_var_2: int32 = ((y.outer*1024) + cse_var_4)
let cse_var_1: int32 = (((x.outer*32768) + (x.c*1024)) + cse_var_4)
{
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[cse_var_1], 32)*packedB_1[cse_var_2]))
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[(cse_var_1 + 1)], 32)*packedB_1[(cse_var_2 + 1)]))
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[(cse_var_1 + 2)], 32)*packedB_1[(cse_var_2 + 2)]))
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[(cse_var_1 + 3)], 32)*packedB_1[(cse_var_2 + 3)]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3: Buffer(A_2, float32, [1048576], [])[cse_var_1], 32)*packedB_1[cse_var_2]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3[(cse_var_1 + 1)], 32)*packedB_1[(cse_var_2 + 1)]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3[(cse_var_1 + 2)], 32)*packedB_1[(cse_var_2 + 2)]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3[(cse_var_1 + 3)], 32)*packedB_1[(cse_var_2 + 3)]))
}
}
}
for (x.inner: int32, 0, 32) {
for (y.inner: int32, 0, 32) {
- C[((((x.outer*32768) + (x.inner*1024)) + (y.outer*32)) + y.inner)] = C.global_1[((x.inner*32) + y.inner)]
+ C_3: Buffer(C_2, float32, [1048576], [])[((((x.outer*32768) + (x.inner*1024)) + (y.outer*32)) + y.inner)] = C.global_1[((x.inner*32) + y.inner)]
}
}
}
@@ -1615,13 +1606,13 @@ working, we can compare the results.
.. code-block:: none
Operator Timing Performance
- none 3.4012933316000002 1.0
- blocking 0.3131420486 0.09206558155120806
- vectorization 0.3455440729 0.1015919649415985
- loop permutation 0.11543491220000002 0.03393853482954331
- array packing 0.1095879703 0.03221949994193791
- block caching 0.1104435998 0.03247105998589257
- parallelization 0.14609199180000002 0.042951894340520456
+ none 3.5404082826999996 1.0
+ blocking 0.29906072260000005 0.08447068776257896
+ vectorization 0.3372738984 0.0952641253405912
+ loop permutation 0.11899243670000001 0.03360980632698485
+ array packing 0.10979653869999999 0.03101239459768367
+ block caching 0.110635423 0.031249340235874376
+ parallelization 0.14436925309999998 0.040777571842618265
@@ -1663,7 +1654,7 @@ the computation for specific platforms.
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 1 minutes 1.139 seconds)
+ **Total running time of the script:** ( 1 minutes 2.680 seconds)
.. _sphx_glr_download_tutorial_tensor_expr_get_started.py:
diff --git a/docs/commit_hash b/docs/commit_hash
index 3e8dfd8c54..944482e15b 100644
--- a/docs/commit_hash
+++ b/docs/commit_hash
@@ -1 +1 @@
-0d9b1850afa7553719789001aa69c2bd8a8f5651
+78b53221f8dd8c1d2bbeff9d34803db33ca254dd
diff --git a/docs/how_to/compile_models/from_darknet.html b/docs/how_to/compile_models/from_darknet.html
index 827ca9ecd3..aedfbbad50 100644
--- a/docs/how_to/compile_models/from_darknet.html
+++ b/docs/how_to/compile_models/from_darknet.html
@@ -585,7 +585,7 @@ class:['truck 0.9266'] left:471 top:83 right:689 bottom:169
class:['bicycle 0.9984'] left:111 top:113 right:577 bottom:447
</pre></div>
</div>
-<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 1 minutes 13.973 seconds)</p>
+<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 1 minutes 13.629 seconds)</p>
<div class="sphx-glr-footer sphx-glr-footer-example docutils container" id="sphx-glr-download-how-to-compile-models-from-darknet-py">
<div class="sphx-glr-download sphx-glr-download-python docutils container">
<p><a class="reference download internal" download="" href="../../_downloads/7716f96385bd5abb6e822041e285be54/from_darknet.py"><code class="xref download docutils literal notranslate"><span class="pre">Download</span> <span class="pre">Python</span> <span class="pre">source</span> <span class="pre">code:</span> <span class="pre">from_darknet.py</span></code></a></p>
diff --git a/docs/how_to/compile_models/from_keras.html b/docs/how_to/compile_models/from_keras.html
index cc206fb3e1..b1bb52252b 100644
--- a/docs/how_to/compile_models/from_keras.html
+++ b/docs/how_to/compile_models/from_keras.html
@@ -506,7 +506,7 @@ pip install -U tensorflow --user
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Relay top-1 id: 285, class name: Egyptian cat
1/1 [==============================] - ETA: 0s
-1/1 [==============================] - 1s 974ms/step
+1/1 [==============================] - 1s 958ms/step
Keras top-1 id: 285, class name: Egyptian cat
</pre></div>
</div>
diff --git a/docs/how_to/compile_models/from_mxnet.html b/docs/how_to/compile_models/from_mxnet.html
index 304054e467..001decc494 100644
--- a/docs/how_to/compile_models/from_mxnet.html
+++ b/docs/how_to/compile_models/from_mxnet.html
@@ -440,7 +440,7 @@ to download the full example code</p>
<span class="nb">print</span><span class="p">(</span><span class="s2">"x"</span><span class="p">,</span> <a href="https://docs.python.org/3/library/stdtypes.html#tuple" title="builtins.tuple" class="sphx-glr-backref-module-builtins sphx-glr-backref-type-py-class sphx-glr-backref-instance"><span class="n">x</span><span class="o">.</span><span class="n">shape</span></a><span class="p">)</span>
</pre></div>
</div>
-<img src="../../_images/sphx_glr_from_mxnet_001.png" srcset="../../_images/sphx_glr_from_mxnet_001.png" alt="from mxnet" class = "sphx-glr-single-img"/><div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Downloading /workspace/.mxnet/models/resnet18_v1-a0666292.zipf7e8c1d4-467d-4397-b979-f49ad05f57c3 from https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/models/resnet18_v1-a0666292.zip...
+<img src="../../_images/sphx_glr_from_mxnet_001.png" srcset="../../_images/sphx_glr_from_mxnet_001.png" alt="from mxnet" class = "sphx-glr-single-img"/><div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Downloading /workspace/.mxnet/models/resnet18_v1-a0666292.zip6172288d-0a35-409b-9577-c9b5f252dbdd from https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/models/resnet18_v1-a0666292.zip...
x (1, 3, 224, 224)
</pre></div>
</div>
diff --git a/docs/how_to/compile_models/from_oneflow.html b/docs/how_to/compile_models/from_oneflow.html
index c32dcb51a0..d2ace9f686 100644
--- a/docs/how_to/compile_models/from_oneflow.html
+++ b/docs/how_to/compile_models/from_oneflow.html
@@ -448,12 +448,13 @@ Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdo
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Downloading: "https://oneflow-public.oss-cn-beijing.aliyuncs.com/model_zoo/flowvision/classification/ResNet/resnet18.zip" to /workspace/.oneflow/flowvision_cache/resnet18.zip
0%| | 0.00/41.5M [00:00<?, ?B/s]
- 19%|#9 | 7.99M/41.5M [00:00<00:00, 37.1MB/s]
- 39%|###8 | 16.0M/41.5M [00:00<00:00, 46.9MB/s]
- 58%|#####7 | 24.0M/41.5M [00:00<00:00, 49.6MB/s]
- 77%|#######7 | 32.0M/41.5M [00:00<00:00, 55.8MB/s]
- 90%|######### | 37.5M/41.5M [00:00<00:00, 55.9MB/s]
-100%|##########| 41.5M/41.5M [00:01<00:00, 43.1MB/s]
+ 15%|#5 | 6.33M/41.5M [00:00<00:00, 47.2MB/s]
+ 26%|##6 | 10.8M/41.5M [00:00<00:00, 38.8MB/s]
+ 39%|###8 | 16.0M/41.5M [00:00<00:00, 39.6MB/s]
+ 58%|#####7 | 24.0M/41.5M [00:00<00:00, 38.3MB/s]
+ 77%|#######7 | 32.0M/41.5M [00:00<00:00, 43.5MB/s]
+ 87%|########7 | 36.2M/41.5M [00:00<00:00, 43.4MB/s]
+100%|##########| 41.5M/41.5M [00:00<00:00, 44.2MB/s]
</pre></div>
</div>
</div>
diff --git a/docs/how_to/compile_models/from_pytorch.html b/docs/how_to/compile_models/from_pytorch.html
index 021927d2cc..32a63ad589 100644
--- a/docs/how_to/compile_models/from_pytorch.html
+++ b/docs/how_to/compile_models/from_pytorch.html
@@ -431,10 +431,13 @@ be unstable.</p>
Downloading: "https://download.pytorch.org/models/resnet18-f37072fd.pth" to /workspace/.cache/torch/hub/checkpoints/resnet18-f37072fd.pth
0%| | 0.00/44.7M [00:00<?, ?B/s]
- 18%|#7 | 7.99M/44.7M [00:00<00:00, 63.7MB/s]
- 54%|#####3 | 24.0M/44.7M [00:00<00:00, 111MB/s]
- 78%|#######8 | 34.9M/44.7M [00:00<00:00, 73.9MB/s]
-100%|##########| 44.7M/44.7M [00:00<00:00, 89.6MB/s]
+ 18%|#7 | 7.99M/44.7M [00:00<00:00, 39.7MB/s]
+ 36%|###5 | 16.0M/44.7M [00:00<00:00, 46.9MB/s]
+ 54%|#####3 | 24.1M/44.7M [00:00<00:00, 57.7MB/s]
+ 68%|######7 | 30.3M/44.7M [00:00<00:00, 47.9MB/s]
+ 79%|#######8 | 35.2M/44.7M [00:00<00:00, 43.7MB/s]
+ 90%|########9 | 40.0M/44.7M [00:00<00:00, 39.0MB/s]
+100%|##########| 44.7M/44.7M [00:01<00:00, 46.5MB/s]
</pre></div>
</div>
</div>
diff --git a/docs/how_to/compile_models/from_tensorflow.html b/docs/how_to/compile_models/from_tensorflow.html
index b7be983556..a85fc16215 100644
--- a/docs/how_to/compile_models/from_tensorflow.html
+++ b/docs/how_to/compile_models/from_tensorflow.html
@@ -645,7 +645,7 @@ banana (score = 0.00022)
desk (score = 0.00019)
</pre></div>
</div>
-<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 1 minutes 10.762 seconds)</p>
+<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 1 minutes 14.751 seconds)</p>
<div class="sphx-glr-footer sphx-glr-footer-example docutils container" id="sphx-glr-download-how-to-compile-models-from-tensorflow-py">
<div class="sphx-glr-download sphx-glr-download-python docutils container">
<p><a class="reference download internal" download="" href="../../_downloads/7f1d3d1b878694c201c614c807cdebc8/from_tensorflow.py"><code class="xref download docutils literal notranslate"><span class="pre">Download</span> <span class="pre">Python</span> <span class="pre">source</span> <span class="pre">code:</span> <span class="pre">from_tensorflow.py</span></code></a></p>
diff --git a/docs/how_to/compile_models/sg_execution_times.html b/docs/how_to/compile_models/sg_execution_times.html
index 3df5299fd6..75f0c16580 100644
--- a/docs/how_to/compile_models/sg_execution_times.html
+++ b/docs/how_to/compile_models/sg_execution_times.html
@@ -340,7 +340,7 @@
<div class="section" id="computation-times">
<span id="sphx-glr-how-to-compile-models-sg-execution-times"></span><h1>Computation times<a class="headerlink" href="#computation-times" title="Permalink to this headline">¶</a></h1>
-<p><strong>05:49.921</strong> total execution time for <strong>how_to_compile_models</strong> files:</p>
+<p><strong>06:02.125</strong> total execution time for <strong>how_to_compile_models</strong> files:</p>
<table class="docutils align-default">
<colgroup>
<col style="width: 81%" />
@@ -348,44 +348,44 @@
<col style="width: 8%" />
</colgroup>
<tbody>
-<tr class="row-odd"><td><p><a class="reference internal" href="from_darknet.html#sphx-glr-how-to-compile-models-from-darknet-py"><span class="std std-ref">Compile YOLO-V2 and YOLO-V3 in DarkNet Models</span></a> (<code class="docutils literal notranslate"><span class="pre">from_darknet.py</span></code>)</p></td>
-<td><p>01:13.973</p></td>
+<tr class="row-odd"><td><p><a class="reference internal" href="from_tensorflow.html#sphx-glr-how-to-compile-models-from-tensorflow-py"><span class="std std-ref">Compile Tensorflow Models</span></a> (<code class="docutils literal notranslate"><span class="pre">from_tensorflow.py</span></code>)</p></td>
+<td><p>01:14.751</p></td>
<td><p>0.0 MB</p></td>
</tr>
-<tr class="row-even"><td><p><a class="reference internal" href="from_tensorflow.html#sphx-glr-how-to-compile-models-from-tensorflow-py"><span class="std std-ref">Compile Tensorflow Models</span></a> (<code class="docutils literal notranslate"><span class="pre">from_tensorflow.py</span></code>)</p></td>
-<td><p>01:10.762</p></td>
+<tr class="row-even"><td><p><a class="reference internal" href="from_darknet.html#sphx-glr-how-to-compile-models-from-darknet-py"><span class="std std-ref">Compile YOLO-V2 and YOLO-V3 in DarkNet Models</span></a> (<code class="docutils literal notranslate"><span class="pre">from_darknet.py</span></code>)</p></td>
+<td><p>01:13.629</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="from_paddle.html#sphx-glr-how-to-compile-models-from-paddle-py"><span class="std std-ref">Compile PaddlePaddle Models</span></a> (<code class="docutils literal notranslate"><span class="pre">from_paddle.py</span></code>)</p></td>
-<td><p>00:47.141</p></td>
+<td><p>00:46.113</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="from_oneflow.html#sphx-glr-how-to-compile-models-from-oneflow-py"><span class="std std-ref">Compile OneFlow Models</span></a> (<code class="docutils literal notranslate"><span class="pre">from_oneflow.py</span></code>)</p></td>
-<td><p>00:33.067</p></td>
+<td><p>00:35.536</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="from_mxnet.html#sphx-glr-how-to-compile-models-from-mxnet-py"><span class="std std-ref">Compile MXNet Models</span></a> (<code class="docutils literal notranslate"><span class="pre">from_mxnet.py</span></code>)</p></td>
-<td><p>00:29.133</p></td>
+<td><p>00:32.462</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="from_coreml.html#sphx-glr-how-to-compile-models-from-coreml-py"><span class="std std-ref">Compile CoreML Models</span></a> (<code class="docutils literal notranslate"><span class="pre">from_coreml.py</span></code>)</p></td>
-<td><p>00:27.214</p></td>
+<td><p>00:28.438</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="from_tflite.html#sphx-glr-how-to-compile-models-from-tflite-py"><span class="std std-ref">Compile TFLite Models</span></a> (<code class="docutils literal notranslate"><span class="pre">from_tflite.py</span></code>)</p></td>
-<td><p>00:25.583</p></td>
+<td><p>00:26.481</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="from_pytorch.html#sphx-glr-how-to-compile-models-from-pytorch-py"><span class="std std-ref">Compile PyTorch Models</span></a> (<code class="docutils literal notranslate"><span class="pre">from_pytorch.py</span></code>)</p></td>
-<td><p>00:22.732</p></td>
+<td><p>00:24.117</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="from_keras.html#sphx-glr-how-to-compile-models-from-keras-py"><span class="std std-ref">Compile Keras Models</span></a> (<code class="docutils literal notranslate"><span class="pre">from_keras.py</span></code>)</p></td>
-<td><p>00:17.900</p></td>
+<td><p>00:18.113</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="from_onnx.html#sphx-glr-how-to-compile-models-from-onnx-py"><span class="std std-ref">Compile ONNX Models</span></a> (<code class="docutils literal notranslate"><span class="pre">from_onnx.py</span></code>)</p></td>
-<td><p>00:02.417</p></td>
+<td><p>00:02.485</p></td>
<td><p>0.0 MB</p></td>
</tr>
</tbody>
diff --git a/docs/how_to/deploy_models/deploy_model_on_android.html b/docs/how_to/deploy_models/deploy_model_on_android.html
index cfdb9ef34c..423aa463e7 100644
--- a/docs/how_to/deploy_models/deploy_model_on_android.html
+++ b/docs/how_to/deploy_models/deploy_model_on_android.html
@@ -661,7 +661,7 @@ to the remote android device.</p>
Evaluate inference time cost...
Execution time summary:
mean (ms) median (ms) max (ms) min (ms) std (ms)
- 16.2353 16.2281 16.3931 16.1329 0.0897
+ 16.0401 16.0527 16.2899 15.8806 0.1304
</pre></div>
</div>
</div>
diff --git a/docs/how_to/deploy_models/deploy_object_detection_pytorch.html b/docs/how_to/deploy_models/deploy_object_detection_pytorch.html
index ace6490ebe..eff7109d40 100644
--- a/docs/how_to/deploy_models/deploy_object_detection_pytorch.html
+++ b/docs/how_to/deploy_models/deploy_object_detection_pytorch.html
@@ -453,28 +453,31 @@ be unstable.</p>
Downloading: "https://download.pytorch.org/models/maskrcnn_resnet50_fpn_coco-bf2d0c1e.pth" to /workspace/.cache/torch/hub/checkpoints/maskrcnn_resnet50_fpn_coco-bf2d0c1e.pth
0%| | 0.00/170M [00:00<?, ?B/s]
- 1%|1 | 2.05M/170M [00:00<00:17, 9.96MB/s]
- 5%|4 | 7.99M/170M [00:00<00:05, 28.7MB/s]
- 9%|9 | 16.0M/170M [00:00<00:03, 41.3MB/s]
- 14%|#4 | 24.1M/170M [00:00<00:02, 51.9MB/s]
- 19%|#8 | 32.0M/170M [00:00<00:02, 53.5MB/s]
- 24%|##3 | 40.0M/170M [00:00<00:02, 54.7MB/s]
- 28%|##8 | 48.1M/170M [00:00<00:02, 62.0MB/s]
- 34%|###4 | 58.0M/170M [00:01<00:01, 61.2MB/s]
- 38%|###7 | 64.1M/170M [00:01<00:02, 51.7MB/s]
- 43%|####3 | 73.7M/170M [00:01<00:01, 62.6MB/s]
- 49%|####8 | 82.9M/170M [00:01<00:01, 70.9MB/s]
- 53%|#####3 | 90.3M/170M [00:01<00:01, 66.1MB/s]
- 59%|#####8 | 99.4M/170M [00:01<00:01, 73.4MB/s]
- 63%|######3 | 107M/170M [00:01<00:00, 76.2MB/s]
- 68%|######7 | 115M/170M [00:01<00:00, 76.7MB/s]
- 73%|#######2 | 123M/170M [00:02<00:00, 79.6MB/s]
- 77%|#######7 | 131M/170M [00:02<00:00, 63.7MB/s]
- 82%|########2 | 140M/170M [00:02<00:00, 70.4MB/s]
- 88%|########8 | 150M/170M [00:02<00:00, 78.7MB/s]
- 94%|#########4| 160M/170M [00:02<00:00, 81.3MB/s]
- 99%|#########8| 168M/170M [00:02<00:00, 78.0MB/s]
-100%|##########| 170M/170M [00:02<00:00, 65.1MB/s]
+ 5%|4 | 7.99M/170M [00:00<00:02, 56.7MB/s]
+ 9%|9 | 16.0M/170M [00:00<00:02, 63.9MB/s]
+ 14%|#4 | 24.0M/170M [00:00<00:02, 53.8MB/s]
+ 19%|#8 | 32.0M/170M [00:00<00:02, 50.6MB/s]
+ 24%|##3 | 40.0M/170M [00:00<00:02, 55.8MB/s]
+ 27%|##7 | 46.3M/170M [00:00<00:02, 53.3MB/s]
+ 32%|###2 | 54.5M/170M [00:00<00:01, 61.5MB/s]
+ 36%|###5 | 60.7M/170M [00:01<00:01, 62.3MB/s]
+ 39%|###9 | 66.9M/170M [00:01<00:01, 62.1MB/s]
+ 43%|####2 | 73.0M/170M [00:01<00:01, 59.0MB/s]
+ 46%|####6 | 78.7M/170M [00:01<00:01, 59.2MB/s]
+ 51%|##### | 86.3M/170M [00:01<00:01, 51.7MB/s]
+ 54%|#####3 | 91.5M/170M [00:01<00:01, 45.6MB/s]
+ 57%|#####6 | 96.1M/170M [00:01<00:01, 41.8MB/s]
+ 60%|###### | 102M/170M [00:02<00:01, 46.2MB/s]
+ 64%|######4 | 109M/170M [00:02<00:01, 52.0MB/s]
+ 67%|######7 | 114M/170M [00:02<00:01, 48.1MB/s]
+ 71%|####### | 120M/170M [00:02<00:01, 49.8MB/s]
+ 75%|#######5 | 128M/170M [00:02<00:00, 56.6MB/s]
+ 79%|#######9 | 134M/170M [00:02<00:00, 58.0MB/s]
+ 82%|########2 | 140M/170M [00:02<00:00, 53.5MB/s]
+ 86%|########5 | 146M/170M [00:02<00:00, 55.1MB/s]
+ 89%|########9 | 152M/170M [00:03<00:00, 46.7MB/s]
+ 94%|#########4| 160M/170M [00:03<00:00, 49.9MB/s]
+100%|##########| 170M/170M [00:03<00:00, 54.5MB/s]
/venv/apache-tvm-py3.7/lib/python3.7/site-packages/torch/nn/functional.py:3897: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).
for i in range(dim)
/venv/apache-tvm-py3.7/lib/python3.7/site-packages/torchvision/models/detection/anchor_utils.py:124: UserWarning: __floordiv__ is deprecated, and its behavior will change in a future version of pytorch. It currently rounds toward 0 (like the 'trunc' function NOT 'floor'). This results in incorrect rounding for negative values. To keep the current behavior, use torch.div(a, b, rounding_mode='trunc'), or for actual floor division, use torch.div(a, b, rounding_mode=& [...]
@@ -572,7 +575,7 @@ torchvision rcnn models.</p>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Get 9 valid boxes
</pre></div>
</div>
-<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 3 minutes 16.638 seconds)</p>
+<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 3 minutes 21.658 seconds)</p>
<div class="sphx-glr-footer sphx-glr-footer-example docutils container" id="sphx-glr-download-how-to-deploy-models-deploy-object-detection-pytorch-py">
<div class="sphx-glr-download sphx-glr-download-python docutils container">
<p><a class="reference download internal" download="" href="../../_downloads/7795da4b258c8feff986668b95ef57ad/deploy_object_detection_pytorch.py"><code class="xref download docutils literal notranslate"><span class="pre">Download</span> <span class="pre">Python</span> <span class="pre">source</span> <span class="pre">code:</span> <span class="pre">deploy_object_detection_pytorch.py</span></code></a></p>
diff --git a/docs/how_to/deploy_models/deploy_prequantized.html b/docs/how_to/deploy_models/deploy_prequantized.html
index de66dc4e3c..d29b20b59f 100644
--- a/docs/how_to/deploy_models/deploy_prequantized.html
+++ b/docs/how_to/deploy_models/deploy_prequantized.html
@@ -497,8 +497,8 @@ training. Other models require a full post training calibration.</p>
Downloading: "https://download.pytorch.org/models/mobilenet_v2-b0353104.pth" to /workspace/.cache/torch/hub/checkpoints/mobilenet_v2-b0353104.pth
0%| | 0.00/13.6M [00:00<?, ?B/s]
- 59%|#####8 | 7.99M/13.6M [00:00<00:00, 52.1MB/s]
-100%|##########| 13.6M/13.6M [00:00<00:00, 76.1MB/s]
+ 59%|#####8 | 7.99M/13.6M [00:00<00:00, 59.0MB/s]
+100%|##########| 13.6M/13.6M [00:00<00:00, 61.4MB/s]
</pre></div>
</div>
</div>
@@ -589,7 +589,7 @@ output values are identical out of 1000 outputs from mobilenet v2.</p>
</div>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Execution time summary:
mean (ms) median (ms) max (ms) min (ms) std (ms)
- 90.3821 90.2990 92.0573 90.0415 0.2972
+ 90.5403 90.5017 94.8389 90.1184 0.4960
</pre></div>
</div>
<div class="admonition note">
@@ -628,7 +628,7 @@ This includes support for the VNNI 8 bit dot product instruction (CascadeLake or
<div class="section" id="deploy-a-quantized-tflite-model">
<h2>Deploy a quantized TFLite Model<a class="headerlink" href="#deploy-a-quantized-tflite-model" title="Permalink to this headline">¶</a></h2>
<p>TODO</p>
-<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 1 minutes 6.391 seconds)</p>
+<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 1 minutes 10.807 seconds)</p>
<div class="sphx-glr-footer sphx-glr-footer-example docutils container" id="sphx-glr-download-how-to-deploy-models-deploy-prequantized-py">
<div class="sphx-glr-download sphx-glr-download-python docutils container">
<p><a class="reference download internal" download="" href="../../_downloads/fb8217c13f4351224c6cf3aacf1a87fc/deploy_prequantized.py"><code class="xref download docutils literal notranslate"><span class="pre">Download</span> <span class="pre">Python</span> <span class="pre">source</span> <span class="pre">code:</span> <span class="pre">deploy_prequantized.py</span></code></a></p>
diff --git a/docs/how_to/deploy_models/deploy_prequantized_tflite.html b/docs/how_to/deploy_models/deploy_prequantized_tflite.html
index 2e4fd70cb6..e75177c4c0 100644
--- a/docs/how_to/deploy_models/deploy_prequantized_tflite.html
+++ b/docs/how_to/deploy_models/deploy_prequantized_tflite.html
@@ -582,7 +582,7 @@ TFLite Top-5 labels: [387 102 386 341 349]
</div>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Execution time summary:
mean (ms) median (ms) max (ms) min (ms) std (ms)
- 120.0254 119.9875 121.7564 119.3263 0.3560
+ 119.4717 119.3346 124.6189 118.5417 0.7517
</pre></div>
</div>
<div class="admonition note">
@@ -610,7 +610,7 @@ network for ARM CPU</span></a>.</p></li>
</ul>
</div></blockquote>
</div>
-<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 2 minutes 29.810 seconds)</p>
+<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 2 minutes 30.840 seconds)</p>
<div class="sphx-glr-footer sphx-glr-footer-example docutils container" id="sphx-glr-download-how-to-deploy-models-deploy-prequantized-tflite-py">
<div class="sphx-glr-download sphx-glr-download-python docutils container">
<p><a class="reference download internal" download="" href="../../_downloads/56691c7a27d45da61d112276334640d3/deploy_prequantized_tflite.py"><code class="xref download docutils literal notranslate"><span class="pre">Download</span> <span class="pre">Python</span> <span class="pre">source</span> <span class="pre">code:</span> <span class="pre">deploy_prequantized_tflite.py</span></code></a></p>
diff --git a/docs/how_to/deploy_models/deploy_quantized.html b/docs/how_to/deploy_models/deploy_quantized.html
index 65fe4714f3..719e8f7090 100644
--- a/docs/how_to/deploy_models/deploy_quantized.html
+++ b/docs/how_to/deploy_models/deploy_quantized.html
@@ -520,7 +520,7 @@ for calibration. But the accuracy might be impacted.</p>
DeprecationWarning,
</pre></div>
</div>
-<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 1 minutes 39.550 seconds)</p>
+<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 1 minutes 42.918 seconds)</p>
<div class="sphx-glr-footer sphx-glr-footer-example docutils container" id="sphx-glr-download-how-to-deploy-models-deploy-quantized-py">
<div class="sphx-glr-download sphx-glr-download-python docutils container">
<p><a class="reference download internal" download="" href="../../_downloads/7810ecf51bfc05f7d5e8a400ac3e815d/deploy_quantized.py"><code class="xref download docutils literal notranslate"><span class="pre">Download</span> <span class="pre">Python</span> <span class="pre">source</span> <span class="pre">code:</span> <span class="pre">deploy_quantized.py</span></code></a></p>
diff --git a/docs/how_to/deploy_models/deploy_ssd_gluoncv.html b/docs/how_to/deploy_models/deploy_ssd_gluoncv.html
index 87fe5f3b6b..6f44d235fe 100644
--- a/docs/how_to/deploy_models/deploy_ssd_gluoncv.html
+++ b/docs/how_to/deploy_models/deploy_ssd_gluoncv.html
@@ -462,22 +462,24 @@ to your device.</p>
Downloading /workspace/.mxnet/models/ssd_512_resnet50_v1_voc-9c8b225a.zip from https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/models/ssd_512_resnet50_v1_voc-9c8b225a.zip...
0%| | 0/132723 [00:00<?, ?KB/s]
- 6%|5 | 7560/132723 [00:00<00:01, 75589.44KB/s]
- 11%|#1 | 15152/132723 [00:00<00:01, 66359.25KB/s]
- 16%|#6 | 21862/132723 [00:00<00:02, 54127.63KB/s]
- 23%|##3 | 30616/132723 [00:00<00:01, 65139.02KB/s]
- 28%|##8 | 37478/132723 [00:00<00:01, 51319.82KB/s]
- 35%|###4 | 46181/132723 [00:00<00:01, 60659.27KB/s]
- 41%|####1 | 54785/132723 [00:00<00:01, 67561.18KB/s]
- 48%|####7 | 63313/132723 [00:00<00:00, 72517.58KB/s]
- 54%|#####4 | 72176/132723 [00:01<00:00, 77117.75KB/s]
- 61%|######1 | 81030/132723 [00:01<00:00, 80424.93KB/s]
- 68%|######7 | 89889/132723 [00:01<00:00, 82814.05KB/s]
- 74%|#######4 | 98369/132723 [00:01<00:00, 69539.24KB/s]
- 81%|######## | 106916/132723 [00:01<00:00, 73678.02KB/s]
- 87%|########7 | 115725/132723 [00:01<00:00, 77577.02KB/s]
- 94%|#########3| 124494/132723 [00:01<00:00, 80396.11KB/s]
-100%|##########| 132723/132723 [00:01<00:00, 72410.91KB/s]
+ 5%|5 | 6912/132723 [00:00<00:01, 69106.06KB/s]
+ 11%|#1 | 15152/132723 [00:00<00:01, 68122.83KB/s]
+ 17%|#6 | 21972/132723 [00:00<00:02, 39543.87KB/s]
+ 23%|##2 | 30354/132723 [00:00<00:02, 51109.35KB/s]
+ 28%|##8 | 37497/132723 [00:00<00:01, 56165.70KB/s]
+ 34%|###3 | 44486/132723 [00:00<00:01, 59980.29KB/s]
+ 39%|###8 | 51150/132723 [00:00<00:01, 54000.24KB/s]
+ 45%|####4 | 59588/132723 [00:01<00:01, 61946.16KB/s]
+ 50%|####9 | 66313/132723 [00:01<00:01, 61297.42KB/s]
+ 56%|#####6 | 74717/132723 [00:01<00:00, 67508.51KB/s]
+ 62%|######1 | 81798/132723 [00:01<00:00, 62225.58KB/s]
+ 67%|######6 | 88311/132723 [00:01<00:01, 43222.78KB/s]
+ 73%|#######2 | 96688/132723 [00:01<00:00, 51636.36KB/s]
+ 78%|#######7 | 102894/132723 [00:01<00:00, 53736.48KB/s]
+ 84%|########3 | 111335/132723 [00:01<00:00, 61258.35KB/s]
+ 89%|########9 | 118185/132723 [00:02<00:00, 50980.55KB/s]
+ 95%|#########5| 126549/132723 [00:02<00:00, 58502.28KB/s]
+100%|##########| 132723/132723 [00:02<00:00, 56851.09KB/s]
</pre></div>
</div>
<p>Create TVM runtime and do inference
@@ -516,7 +518,7 @@ Downloading /workspace/.mxnet/models/ssd_512_resnet50_v1_voc-9c8b225a.zip from h
<span class="n">plt</span><span class="o">.</span><span class="n">show</span><span class="p">()</span>
</pre></div>
</div>
-<img src="../../_images/sphx_glr_deploy_ssd_gluoncv_001.png" srcset="../../_images/sphx_glr_deploy_ssd_gluoncv_001.png" alt="deploy ssd gluoncv" class = "sphx-glr-single-img"/><p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 3 minutes 0.610 seconds)</p>
+<img src="../../_images/sphx_glr_deploy_ssd_gluoncv_001.png" srcset="../../_images/sphx_glr_deploy_ssd_gluoncv_001.png" alt="deploy ssd gluoncv" class = "sphx-glr-single-img"/><p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 3 minutes 3.754 seconds)</p>
<div class="sphx-glr-footer sphx-glr-footer-example docutils container" id="sphx-glr-download-how-to-deploy-models-deploy-ssd-gluoncv-py">
<div class="sphx-glr-download sphx-glr-download-python docutils container">
<p><a class="reference download internal" download="" href="../../_downloads/cccb17d28e5e8b2e94ea8cd5ec59f6ed/deploy_ssd_gluoncv.py"><code class="xref download docutils literal notranslate"><span class="pre">Download</span> <span class="pre">Python</span> <span class="pre">source</span> <span class="pre">code:</span> <span class="pre">deploy_ssd_gluoncv.py</span></code></a></p>
diff --git a/docs/how_to/deploy_models/sg_execution_times.html b/docs/how_to/deploy_models/sg_execution_times.html
index d60f484a28..d23124728d 100644
--- a/docs/how_to/deploy_models/sg_execution_times.html
+++ b/docs/how_to/deploy_models/sg_execution_times.html
@@ -340,7 +340,7 @@
<div class="section" id="computation-times">
<span id="sphx-glr-how-to-deploy-models-sg-execution-times"></span><h1>Computation times<a class="headerlink" href="#computation-times" title="Permalink to this headline">¶</a></h1>
-<p><strong>13:00.700</strong> total execution time for <strong>how_to_deploy_models</strong> files:</p>
+<p><strong>13:22.160</strong> total execution time for <strong>how_to_deploy_models</strong> files:</p>
<table class="docutils align-default">
<colgroup>
<col style="width: 86%" />
@@ -349,35 +349,35 @@
</colgroup>
<tbody>
<tr class="row-odd"><td><p><a class="reference internal" href="deploy_object_detection_pytorch.html#sphx-glr-how-to-deploy-models-deploy-object-detection-pytorch-py"><span class="std std-ref">Compile PyTorch Object Detection Models</span></a> (<code class="docutils literal notranslate"><span class="pre">deploy_object_detection_pytorch.py</span></code>)</p></td>
-<td><p>03:16.638</p></td>
+<td><p>03:21.658</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="deploy_ssd_gluoncv.html#sphx-glr-how-to-deploy-models-deploy-ssd-gluoncv-py"><span class="std std-ref">Deploy Single Shot Multibox Detector(SSD) model</span></a> (<code class="docutils literal notranslate"><span class="pre">deploy_ssd_gluoncv.py</span></code>)</p></td>
-<td><p>03:00.610</p></td>
+<td><p>03:03.754</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="deploy_prequantized_tflite.html#sphx-glr-how-to-deploy-models-deploy-prequantized-tflite-py"><span class="std std-ref">Deploy a Framework-prequantized Model with TVM - Part 3 (TFLite)</span></a> (<code class="docutils literal notranslate"><span class="pre">deploy_prequantized_tflite.py</span></code>)</p></td>
-<td><p>02:29.810</p></td>
+<td><p>02:30.840</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="deploy_quantized.html#sphx-glr-how-to-deploy-models-deploy-quantized-py"><span class="std std-ref">Deploy a Quantized Model on Cuda</span></a> (<code class="docutils literal notranslate"><span class="pre">deploy_quantized.py</span></code>)</p></td>
-<td><p>01:39.550</p></td>
+<td><p>01:42.918</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="deploy_prequantized.html#sphx-glr-how-to-deploy-models-deploy-prequantized-py"><span class="std std-ref">Deploy a Framework-prequantized Model with TVM</span></a> (<code class="docutils literal notranslate"><span class="pre">deploy_prequantized.py</span></code>)</p></td>
-<td><p>01:06.391</p></td>
+<td><p>01:10.807</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="deploy_model_on_android.html#sphx-glr-how-to-deploy-models-deploy-model-on-android-py"><span class="std std-ref">Deploy the Pretrained Model on Android</span></a> (<code class="docutils literal notranslate"><span class="pre">deploy_model_on_android.py</span></code>)</p></td>
-<td><p>00:36.869</p></td>
+<td><p>00:38.804</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="deploy_model_on_nano.html#sphx-glr-how-to-deploy-models-deploy-model-on-nano-py"><span class="std std-ref">Deploy the Pretrained Model on Jetson Nano</span></a> (<code class="docutils literal notranslate"><span class="pre">deploy_model_on_nano.py</span></code>)</p></td>
-<td><p>00:25.650</p></td>
+<td><p>00:26.989</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="deploy_model_on_rasp.html#sphx-glr-how-to-deploy-models-deploy-model-on-rasp-py"><span class="std std-ref">Deploy the Pretrained Model on Raspberry Pi</span></a> (<code class="docutils literal notranslate"><span class="pre">deploy_model_on_rasp.py</span></code>)</p></td>
-<td><p>00:25.176</p></td>
+<td><p>00:26.383</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="deploy_sparse.html#sphx-glr-how-to-deploy-models-deploy-sparse-py"><span class="std std-ref">Deploy a Hugging Face Pruned Model on CPU</span></a> (<code class="docutils literal notranslate"><span class="pre">deploy_sparse.py</span></code>)</p></td>
diff --git a/docs/how_to/extend_tvm/bring_your_own_datatypes.html b/docs/how_to/extend_tvm/bring_your_own_datatypes.html
index b8f7282cd3..a210593110 100644
--- a/docs/how_to/extend_tvm/bring_your_own_datatypes.html
+++ b/docs/how_to/extend_tvm/bring_your_own_datatypes.html
@@ -621,7 +621,7 @@ In this alpha state of the Bring Your Own Datatypes framework, we have not imple
<span class="n">module</span><span class="p">,</span> <a href="https://docs.python.org/3/library/stdtypes.html#dict" title="builtins.dict" class="sphx-glr-backref-module-builtins sphx-glr-backref-type-py-class sphx-glr-backref-instance"><span class="n">params</span></a> <span class="o">=</span> <span class="n">get_mobilenet</span><span class="p">()</span>
</pre></div>
</div>
-<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Downloading /workspace/.mxnet/models/mobilenet0.25-9f83e440.zip4f7f955f-332b-4872-85b7-b598746fbf97 from https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/models/mobilenet0.25-9f83e440.zip...
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Downloading /workspace/.mxnet/models/mobilenet0.25-9f83e440.zipccbf2386-7e57-408a-8b03-16a9002a6478 from https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/models/mobilenet0.25-9f83e440.zip...
</pre></div>
</div>
<p>It’s easy to execute MobileNet with native TVM:</p>
diff --git a/docs/how_to/extend_tvm/low_level_custom_pass.html b/docs/how_to/extend_tvm/low_level_custom_pass.html
index 459caf87a7..74a18730a7 100644
--- a/docs/how_to/extend_tvm/low_level_custom_pass.html
+++ b/docs/how_to/extend_tvm/low_level_custom_pass.html
@@ -418,10 +418,9 @@ our customized lowering pass to manipulate the IR directly instead of using sche
buffers = {a: Buffer(a_2: Pointer(float32), float32, [128], []),
b: Buffer(b_2: Pointer(float32), float32, [128], []),
c: Buffer(c_2: Pointer(float32), float32, [128], [])}
- buffer_map = {a_1: a, b_1: b, c_1: c}
- preflattened_buffer_map = {a_1: a_3: Buffer(a_2, float32, [128], []), b_1: b_3: Buffer(b_2, float32, [128], []), c_1: c_3: Buffer(c_2, float32, [128], [])} {
+ buffer_map = {a_1: a, b_1: b, c_1: c} {
for (i: int32, 0, 128) {
- c[i] = (a[i] + b[i])
+ c_3: Buffer(c_2, float32, [128], [])[i] = (a_3: Buffer(a_2, float32, [128], [])[i] + b_3: Buffer(b_2, float32, [128], [])[i])
}
}
</pre></div>
@@ -527,11 +526,10 @@ called after each phase is done.</p>
buffers = {a: Buffer(a_2: Pointer(float32), float32, [128], []),
b: Buffer(b_2: Pointer(float32), float32, [128], []),
c: Buffer(c_2: Pointer(float32), float32, [128], [])}
- buffer_map = {a_1: a, b_1: b, c_1: c}
- preflattened_buffer_map = {a_1: a_3: Buffer(a_2, float32, [128], []), b_1: b_3: Buffer(b_2, float32, [128], []), c_1: c_3: Buffer(c_2, float32, [128], [])} {
+ buffer_map = {a_1: a, b_1: b, c_1: c} {
for (i.outer: int32, 0, 16) {
let cse_var_1: int32 = (i.outer*8)
- c[ramp(cse_var_1, 1, 8)] = (a[ramp(cse_var_1, 1, 8)] + b[ramp(cse_var_1, 1, 8)])
+ c_3: Buffer(c_2, float32, [128], [])[ramp(cse_var_1, 1, 8)] = (a_3: Buffer(a_2, float32, [128], [])[ramp(cse_var_1, 1, 8)] + b_3: Buffer(b_2, float32, [128], [])[ramp(cse_var_1, 1, 8)])
}
}
</pre></div>
diff --git a/docs/how_to/extend_tvm/sg_execution_times.html b/docs/how_to/extend_tvm/sg_execution_times.html
index aad8ed6cce..f431be00d9 100644
--- a/docs/how_to/extend_tvm/sg_execution_times.html
+++ b/docs/how_to/extend_tvm/sg_execution_times.html
@@ -340,7 +340,7 @@
<div class="section" id="computation-times">
<span id="sphx-glr-how-to-extend-tvm-sg-execution-times"></span><h1>Computation times<a class="headerlink" href="#computation-times" title="Permalink to this headline">¶</a></h1>
-<p><strong>00:47.459</strong> total execution time for <strong>how_to_extend_tvm</strong> files:</p>
+<p><strong>00:48.149</strong> total execution time for <strong>how_to_extend_tvm</strong> files:</p>
<table class="docutils align-default">
<colgroup>
<col style="width: 84%" />
@@ -349,15 +349,15 @@
</colgroup>
<tbody>
<tr class="row-odd"><td><p><a class="reference internal" href="bring_your_own_datatypes.html#sphx-glr-how-to-extend-tvm-bring-your-own-datatypes-py"><span class="std std-ref">Bring Your Own Datatypes to TVM</span></a> (<code class="docutils literal notranslate"><span class="pre">bring_your_own_datatypes.py</span></code>)</p></td>
-<td><p>00:44.024</p></td>
+<td><p>00:44.665</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="use_pass_instrument.html#sphx-glr-how-to-extend-tvm-use-pass-instrument-py"><span class="std std-ref">How to Use TVM Pass Instrument</span></a> (<code class="docutils literal notranslate"><span class="pre">use_pass_instrument.py</span></code>)</p></td>
-<td><p>00:02.394</p></td>
+<td><p>00:02.437</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="use_pass_infra.html#sphx-glr-how-to-extend-tvm-use-pass-infra-py"><span class="std std-ref">How to Use TVM Pass Infra</span></a> (<code class="docutils literal notranslate"><span class="pre">use_pass_infra.py</span></code>)</p></td>
-<td><p>00:01.033</p></td>
+<td><p>00:01.039</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="low_level_custom_pass.html#sphx-glr-how-to-extend-tvm-low-level-custom-pass-py"><span class="std std-ref">Writing a Customized Pass</span></a> (<code class="docutils literal notranslate"><span class="pre">low_level_custom_pass.py</span></code>)</p></td>
diff --git a/docs/how_to/extend_tvm/use_pass_instrument.html b/docs/how_to/extend_tvm/use_pass_instrument.html
index 5df7ab7d07..f0042bca92 100644
--- a/docs/how_to/extend_tvm/use_pass_instrument.html
+++ b/docs/how_to/extend_tvm/use_pass_instrument.html
@@ -525,10 +525,10 @@ profile the execution time of each passes.</p>
</pre></div>
</div>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Printing results of timing profile...
-InferType: 7142us [7142us] (44.30%; 44.30%)
-FoldScaleAxis: 8981us [7us] (55.70%; 55.70%)
- FoldConstant: 8975us [1662us] (55.66%; 99.93%)
- InferType: 7313us [7313us] (45.35%; 81.48%)
+InferType: 7176us [7176us] (46.30%; 46.30%)
+FoldScaleAxis: 8322us [6us] (53.70%; 53.70%)
+ FoldConstant: 8315us [1690us] (53.66%; 99.92%)
+ InferType: 6626us [6626us] (42.75%; 79.68%)
</pre></div>
</div>
</div>
@@ -550,10 +550,10 @@ Refer to following sections and <a class="reference internal" href="../../refere
</pre></div>
</div>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Printing results of timing profile...
-InferType: 6707us [6707us] (45.33%; 45.33%)
-FoldScaleAxis: 8087us [5us] (54.67%; 54.67%)
- FoldConstant: 8083us [1680us] (54.63%; 99.94%)
- InferType: 6403us [6403us] (43.28%; 79.22%)
+InferType: 6651us [6651us] (45.01%; 45.01%)
+FoldScaleAxis: 8127us [5us] (54.99%; 54.99%)
+ FoldConstant: 8122us [1678us] (54.96%; 99.94%)
+ InferType: 6444us [6444us] (43.61%; 79.34%)
</pre></div>
</div>
<p>Register empty list to clear existing instruments.</p>
diff --git a/docs/how_to/optimize_operators/opt_conv_cuda.html b/docs/how_to/optimize_operators/opt_conv_cuda.html
index d73959941f..e8b02229ae 100644
--- a/docs/how_to/optimize_operators/opt_conv_cuda.html
+++ b/docs/how_to/optimize_operators/opt_conv_cuda.html
@@ -577,7 +577,7 @@ latency of convolution.</p>
<span class="nb">print</span><span class="p">(</span><span class="s2">"Convolution: </span><span class="si">%f</span><span class="s2"> ms"</span> <span class="o">%</span> <span class="p">(</span><span class="n">evaluator</span><span class="p">(</span><span class="n">a</span><span class="p">,</span> <span class="n">w</span><span class="p">,</span> <span class="n">b</span><span class="p">)</span><span class="o">.</span><span class="n">mean</span> <span class="o">*</span> <span cl [...]
</pre></div>
</div>
-<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Convolution: 47.685344 ms
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Convolution: 41.190143 ms
</pre></div>
</div>
<div class="sphx-glr-footer sphx-glr-footer-example docutils container" id="sphx-glr-download-how-to-optimize-operators-opt-conv-cuda-py">
diff --git a/docs/how_to/optimize_operators/opt_conv_tensorcore.html b/docs/how_to/optimize_operators/opt_conv_tensorcore.html
index 7a67713b5f..cac478e018 100644
--- a/docs/how_to/optimize_operators/opt_conv_tensorcore.html
+++ b/docs/how_to/optimize_operators/opt_conv_tensorcore.html
@@ -724,11 +724,10 @@ The only thing we should do is to make sure all threads in a warp can call Tenso
</div>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@main = primfn(A_1: handle, W_1: handle, Conv_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float16), float16, [12845056], []),
- W: Buffer(W_2: Pointer(float16), float16, [1179648], []),
- Conv: Buffer(Conv_2: Pointer(float32), float32, [25690112], [])}
- buffer_map = {A_1: A, W_1: W, Conv_1: Conv}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float16, [16, 14, 14, 16, 16, 16], []), W_1: W_3: Buffer(W_2, float16, [3, 3, 16, 32, 16, 16], []), Conv_1: Conv_3: Buffer(Conv_2, float32, [16, 14, 14, 32, 16, 16], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float16), float16, [16, 14, 14, 16, 16, 16], []),
+ W: Buffer(W_2: Pointer(float16), float16, [3, 3, 16, 32, 16, 16], []),
+ Conv: Buffer(Conv_2: Pointer(float32), float32, [16, 14, 14, 32, 16, 16], [])}
+ buffer_map = {A_1: A, W_1: W, Conv_1: Conv} {
attr [IterVar(blockIdx.z: int32, (nullptr), "ThreadIndex", "blockIdx.z")] "thread_extent" = 196;
allocate(Conv.wmma.accumulator: Pointer(wmma.accumulator float32), float32, [2048]), storage_scope = wmma.accumulator;
allocate(Apad.shared: Pointer(shared float16), float16, [12288]), storage_scope = shared;
@@ -756,14 +755,14 @@ The only thing we should do is to make sure all threads in a warp can call Tenso
let cse_var_2: int32 = (ax3*256)
let cse_var_1: int32 = (ax4.ax5.fused.outer*32)
attr [IterVar(threadIdx.x: int32, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 32;
- Apad.shared_1: Buffer(Apad.shared, float16, [12288], [], scope="shared")[((((((threadIdx.y*3072) + (threadIdx.z*1536)) + (ax2*512)) + cse_var_2) + cse_var_1) + threadIdx.x)] = @tir.if_then_else(((((1 <= (floordiv(blockIdx.z, 14) + kh)) && ((floordiv(blockIdx.z, 14) + kh) < 15)) && (1 <= (ax2 + floormod(blockIdx.z, 14)))) && ((ax2 + floormod(blockIdx.z, 14)) < 15)), A[(((((((((((blockIdx.x*6422528) + (threadIdx.y*1605632)) + (thr [...]
+ Apad.shared_1: Buffer(Apad.shared, float16, [12288], [], scope="shared")[((((((threadIdx.y*3072) + (threadIdx.z*1536)) + (ax2*512)) + cse_var_2) + cse_var_1) + threadIdx.x)] = @tir.if_then_else(((((1 <= (floordiv(blockIdx.z, 14) + kh)) && ((floordiv(blockIdx.z, 14) + kh) < 15)) && (1 <= (ax2 + floormod(blockIdx.z, 14)))) && ((ax2 + floormod(blockIdx.z, 14)) < 15)), A_3: Buffer(A_2, float16, [12845056], [])[(((((((((((blockIdx.x* [...]
}
}
}
for (ax1: int32, 0, 3) {
for (ax2_1: int32, 0, 2) {
attr [IterVar(threadIdx.x, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 32;
- W.shared_1: Buffer(W.shared, float16, [12288], [], scope="shared")[ramp((((((ax1*4096) + (ax2_1*2048)) + (threadIdx.y*512)) + (threadIdx.z*256)) + (threadIdx.x*8)), 1, 8)] = W[ramp(((((((((kh*393216) + (ax1*131072)) + (ic.outer*16384)) + (ax2_1*8192)) + (blockIdx.y*2048)) + (threadIdx.y*512)) + (threadIdx.z*256)) + (threadIdx.x*8)), 1, 8)]
+ W.shared_1: Buffer(W.shared, float16, [12288], [], scope="shared")[ramp((((((ax1*4096) + (ax2_1*2048)) + (threadIdx.y*512)) + (threadIdx.z*256)) + (threadIdx.x*8)), 1, 8)] = W_3: Buffer(W_2, float16, [1179648], [])[ramp(((((((((kh*393216) + (ax1*131072)) + (ic.outer*16384)) + (ax2_1*8192)) + (blockIdx.y*2048)) + (threadIdx.y*512)) + (threadIdx.z*256)) + (threadIdx.x*8)), 1, 8)]
}
}
for (ic.inner: int32, 0, 2) {
@@ -809,7 +808,7 @@ The only thing we should do is to make sure all threads in a warp can call Tenso
for (oo: int32, 0, 16) {
let cse_var_10: int32 = (o.inner*256)
let cse_var_9: int32 = (nn*16)
- Conv[(((((((((blockIdx.x*12845056) + (threadIdx.y*3211264)) + (n.inner*1605632)) + (blockIdx.z*8192)) + (blockIdx.y*2048)) + (threadIdx.z*1024)) + cse_var_10) + cse_var_9) + oo)] = Conv.wmma.accumulator_1[((((n.inner*1024) + cse_var_10) + cse_var_9) + oo)]
+ Conv_3: Buffer(Conv_2, float32, [25690112], [])[(((((((((blockIdx.x*12845056) + (threadIdx.y*3211264)) + (n.inner*1605632)) + (blockIdx.z*8192)) + (blockIdx.y*2048)) + (threadIdx.z*1024)) + cse_var_10) + cse_var_9) + oo)] = Conv.wmma.accumulator_1[((((n.inner*1024) + cse_var_10) + cse_var_9) + oo)]
}
}
}
@@ -832,11 +831,10 @@ by mapping the 2D convolution to tensor intrinsics</p>
</div>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@main = primfn(A_1: handle, W_1: handle, Conv_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float16), float16, [12845056], []),
- W: Buffer(W_2: Pointer(float16), float16, [1179648], []),
- Conv: Buffer(Conv_2: Pointer(float32), float32, [25690112], [])}
- buffer_map = {A_1: A, W_1: W, Conv_1: Conv}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float16, [16, 14, 14, 16, 16, 16], []), W_1: W_3: Buffer(W_2, float16, [3, 3, 16, 32, 16, 16], []), Conv_1: Conv_3: Buffer(Conv_2, float32, [16, 14, 14, 32, 16, 16], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float16), float16, [16, 14, 14, 16, 16, 16], []),
+ W: Buffer(W_2: Pointer(float16), float16, [3, 3, 16, 32, 16, 16], []),
+ Conv: Buffer(Conv_2: Pointer(float32), float32, [16, 14, 14, 32, 16, 16], [])}
+ buffer_map = {A_1: A, W_1: W, Conv_1: Conv} {
attr [IterVar(blockIdx.z: int32, (nullptr), "ThreadIndex", "blockIdx.z")] "thread_extent" = 196;
allocate(Conv.wmma.accumulator: Pointer(wmma.accumulator float32), float32, [2048]), storage_scope = wmma.accumulator;
allocate(Apad.shared: Pointer(shared float16), float16, [12288]), storage_scope = shared;
@@ -860,14 +858,14 @@ by mapping the 2D convolution to tensor intrinsics</p>
let cse_var_2: int32 = (ax3*256)
let cse_var_1: int32 = (ax4.ax5.fused.outer*32)
attr [IterVar(threadIdx.x: int32, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 32;
- Apad.shared_1: Buffer(Apad.shared, float16, [12288], [], scope="shared")[((((((threadIdx.y*3072) + (threadIdx.z*1536)) + (ax2*512)) + cse_var_2) + cse_var_1) + threadIdx.x)] = @tir.if_then_else(((((1 <= (floordiv(blockIdx.z, 14) + kh)) && ((floordiv(blockIdx.z, 14) + kh) < 15)) && (1 <= (ax2 + floormod(blockIdx.z, 14)))) && ((ax2 + floormod(blockIdx.z, 14)) < 15)), A[(((((((((((blockIdx.x*6422528) + (threadIdx.y*1605632)) + (thr [...]
+ Apad.shared_1: Buffer(Apad.shared, float16, [12288], [], scope="shared")[((((((threadIdx.y*3072) + (threadIdx.z*1536)) + (ax2*512)) + cse_var_2) + cse_var_1) + threadIdx.x)] = @tir.if_then_else(((((1 <= (floordiv(blockIdx.z, 14) + kh)) && ((floordiv(blockIdx.z, 14) + kh) < 15)) && (1 <= (ax2 + floormod(blockIdx.z, 14)))) && ((ax2 + floormod(blockIdx.z, 14)) < 15)), A_3: Buffer(A_2, float16, [12845056], [])[(((((((((((blockIdx.x* [...]
}
}
}
for (ax1: int32, 0, 3) {
for (ax2_1: int32, 0, 2) {
attr [IterVar(threadIdx.x, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 32;
- W.shared_1: Buffer(W.shared, float16, [12288], [], scope="shared")[ramp((((((ax1*4096) + (ax2_1*2048)) + (threadIdx.y*512)) + (threadIdx.z*256)) + (threadIdx.x*8)), 1, 8)] = W[ramp(((((((((kh*393216) + (ax1*131072)) + (ic.outer*16384)) + (ax2_1*8192)) + (blockIdx.y*2048)) + (threadIdx.y*512)) + (threadIdx.z*256)) + (threadIdx.x*8)), 1, 8)]
+ W.shared_1: Buffer(W.shared, float16, [12288], [], scope="shared")[ramp((((((ax1*4096) + (ax2_1*2048)) + (threadIdx.y*512)) + (threadIdx.z*256)) + (threadIdx.x*8)), 1, 8)] = W_3: Buffer(W_2, float16, [1179648], [])[ramp(((((((((kh*393216) + (ax1*131072)) + (ic.outer*16384)) + (ax2_1*8192)) + (blockIdx.y*2048)) + (threadIdx.y*512)) + (threadIdx.z*256)) + (threadIdx.x*8)), 1, 8)]
}
}
for (ic.inner: int32, 0, 2) {
@@ -916,7 +914,7 @@ be able to run on our build server</p>
<span class="nb">print</span><span class="p">(</span><span class="s2">"conv2d with tensor core: </span><span class="si">%f</span><span class="s2"> ms"</span> <span class="o">%</span> <span class="p">(</span><span class="n">evaluator</span><span class="p">(</span><span class="n">a</span><span class="p">,</span> <span class="n">w</span><span class="p">,</span> <span class="n">c</span><span class="p">)</span><span class="o">.</span><span class="n">mean</span> <span class="o">* [...]
</pre></div>
</div>
-<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>conv2d with tensor core: 12.128944 ms
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>conv2d with tensor core: 13.362189 ms
</pre></div>
</div>
</div>
diff --git a/docs/how_to/optimize_operators/opt_gemm.html b/docs/how_to/optimize_operators/opt_gemm.html
index 623edc2d75..a8d43ae6ad 100644
--- a/docs/how_to/optimize_operators/opt_gemm.html
+++ b/docs/how_to/optimize_operators/opt_gemm.html
@@ -474,8 +474,8 @@ Then we write a baseline implementation, the simplest way to write a matrix mult
<span class="nb">print</span><span class="p">(</span><span class="s2">"Baseline: </span><span class="si">%f</span><span class="s2">"</span> <span class="o">%</span> <span class="n">evaluator</span><span class="p">(</span><span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">,</span> <span class="n">c</span><span class="p">)</span><span class="o">.</span><span class="n">mean</span><span class="p">)</span>
</pre></div>
</div>
-<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Numpy running time: 0.019367
-Baseline: 3.423094
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Numpy running time: 0.018437
+Baseline: 3.542359
</pre></div>
</div>
<p>In TVM, we can always inspect lower level IR to debug or optimize our schedule.
@@ -485,18 +485,17 @@ Here is the generated IR using our baseline schedule.</p>
</div>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [1048576], []),
- B: Buffer(B_2: Pointer(float32), float32, [1048576], []),
- C: Buffer(C_2: Pointer(float32), float32, [1048576], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 1024], []), B_1: B_3: Buffer(B_2, float32, [1024, 1024], []), C_1: C_3: Buffer(C_2, float32, [1024, 1024], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),
+ B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
for (m: int32, 0, 1024) {
for (n: int32, 0, 1024) {
- C[((m*1024) + n)] = 0f32
+ C_3: Buffer(C_2, float32, [1048576], [])[((m*1024) + n)] = 0f32
for (k: int32, 0, 1024) {
let cse_var_2: int32 = (m*1024)
let cse_var_1: int32 = (cse_var_2 + n)
- C[cse_var_1] = (C[cse_var_1] + (A[(cse_var_2 + k)]*B[((k*1024) + n)]))
+ C_3[cse_var_1] = (C_3[cse_var_1] + (A_3: Buffer(A_2, float32, [1048576], [])[(cse_var_2 + k)]*B_3: Buffer(B_2, float32, [1048576], [])[((k*1024) + n)]))
}
}
}
@@ -535,7 +534,7 @@ fill 32 * 32 * sizeof(float) which is 4KB in the cache whose total size is 32KB
<span class="nb">print</span><span class="p">(</span><span class="s2">"Opt1: </span><span class="si">%f</span><span class="s2">"</span> <span class="o">%</span> <span class="n">evaluator</span><span class="p">(</span><span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">,</span> <span class="n">c</span><span class="p">)</span><span class="o">.</span><span class="n">mean</span><span class="p">)</span>
</pre></div>
</div>
-<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Opt1: 0.302157
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Opt1: 0.305990
</pre></div>
</div>
<p>Here is the generated IR after blocking.</p>
@@ -544,16 +543,15 @@ fill 32 * 32 * sizeof(float) which is 4KB in the cache whose total size is 32KB
</div>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [1048576], []),
- B: Buffer(B_2: Pointer(float32), float32, [1048576], []),
- C: Buffer(C_2: Pointer(float32), float32, [1048576], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 1024], []), B_1: B_3: Buffer(B_2, float32, [1024, 1024], []), C_1: C_3: Buffer(C_2, float32, [1024, 1024], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),
+ B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
for (m.outer: int32, 0, 32) {
for (n.outer: int32, 0, 32) {
for (m.inner.init: int32, 0, 32) {
for (n.inner.init: int32, 0, 32) {
- C[((((m.outer*32768) + (m.inner.init*1024)) + (n.outer*32)) + n.inner.init)] = 0f32
+ C_3: Buffer(C_2, float32, [1048576], [])[((((m.outer*32768) + (m.inner.init*1024)) + (n.outer*32)) + n.inner.init)] = 0f32
}
}
for (k.outer: int32, 0, 256) {
@@ -563,7 +561,7 @@ fill 32 * 32 * sizeof(float) which is 4KB in the cache whose total size is 32KB
let cse_var_3: int32 = (n.outer*32)
let cse_var_2: int32 = ((m.outer*32768) + (m.inner*1024))
let cse_var_1: int32 = ((cse_var_2 + cse_var_3) + n.inner)
- C[cse_var_1] = (C[cse_var_1] + (A[((cse_var_2 + (k.outer*4)) + k.inner)]*B[((((k.outer*4096) + (k.inner*1024)) + cse_var_3) + n.inner)]))
+ C_3[cse_var_1] = (C_3[cse_var_1] + (A_3: Buffer(A_2, float32, [1048576], [])[((cse_var_2 + (k.outer*4)) + k.inner)]*B_3: Buffer(B_2, float32, [1048576], [])[((((k.outer*4096) + (k.inner*1024)) + cse_var_3) + n.inner)]))
}
}
}
@@ -602,7 +600,7 @@ vastly.</p>
<span class="nb">print</span><span class="p">(</span><span class="s2">"Opt2: </span><span class="si">%f</span><span class="s2">"</span> <span class="o">%</span> <span class="n">evaluator</span><span class="p">(</span><span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">,</span> <span class="n">c</span><span class="p">)</span><span class="o">.</span><span class="n">mean</span><span class="p">)</span>
</pre></div>
</div>
-<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Opt2: 0.340464
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Opt2: 0.335689
</pre></div>
</div>
<p>Here is the generated IR after vectorization.</p>
@@ -611,15 +609,14 @@ vastly.</p>
</div>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [1048576], []),
- B: Buffer(B_2: Pointer(float32), float32, [1048576], []),
- C: Buffer(C_2: Pointer(float32), float32, [1048576], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 1024], []), B_1: B_3: Buffer(B_2, float32, [1024, 1024], []), C_1: C_3: Buffer(C_2, float32, [1024, 1024], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),
+ B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
for (m.outer: int32, 0, 32) {
for (n.outer: int32, 0, 32) {
for (m.inner.init: int32, 0, 32) {
- C[ramp((((m.outer*32768) + (m.inner.init*1024)) + (n.outer*32)), 1, 32)] = broadcast(0f32, 32)
+ C_3: Buffer(C_2, float32, [1048576], [])[ramp((((m.outer*32768) + (m.inner.init*1024)) + (n.outer*32)), 1, 32)] = broadcast(0f32, 32)
}
for (k.outer: int32, 0, 256) {
for (k.inner: int32, 0, 4) {
@@ -627,7 +624,7 @@ vastly.</p>
let cse_var_3: int32 = (n.outer*32)
let cse_var_2: int32 = ((m.outer*32768) + (m.inner*1024))
let cse_var_1: int32 = (cse_var_2 + cse_var_3)
- C[ramp(cse_var_1, 1, 32)] = (C[ramp(cse_var_1, 1, 32)] + (broadcast(A[((cse_var_2 + (k.outer*4)) + k.inner)], 32)*B[ramp((((k.outer*4096) + (k.inner*1024)) + cse_var_3), 1, 32)]))
+ C_3[ramp(cse_var_1, 1, 32)] = (C_3[ramp(cse_var_1, 1, 32)] + (broadcast(A_3: Buffer(A_2, float32, [1048576], [])[((cse_var_2 + (k.outer*4)) + k.inner)], 32)*B_3: Buffer(B_2, float32, [1048576], [])[ramp((((k.outer*4096) + (k.inner*1024)) + cse_var_3), 1, 32)]))
}
}
}
@@ -663,7 +660,7 @@ the access pattern for A matrix is more cache friendly.</p>
<span class="nb">print</span><span class="p">(</span><span class="s2">"Opt3: </span><span class="si">%f</span><span class="s2">"</span> <span class="o">%</span> <span class="n">evaluator</span><span class="p">(</span><span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">,</span> <span class="n">c</span><span class="p">)</span><span class="o">.</span><span class="n">mean</span><span class="p">)</span>
</pre></div>
</div>
-<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Opt3: 0.116802
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Opt3: 0.116123
</pre></div>
</div>
<p>Here is the generated IR after loop permutation.</p>
@@ -672,15 +669,14 @@ the access pattern for A matrix is more cache friendly.</p>
</div>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [1048576], []),
- B: Buffer(B_2: Pointer(float32), float32, [1048576], []),
- C: Buffer(C_2: Pointer(float32), float32, [1048576], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 1024], []), B_1: B_3: Buffer(B_2, float32, [1024, 1024], []), C_1: C_3: Buffer(C_2, float32, [1024, 1024], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),
+ B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
for (m.outer: int32, 0, 32) {
for (n.outer: int32, 0, 32) {
for (m.inner.init: int32, 0, 32) {
- C[ramp((((m.outer*32768) + (m.inner.init*1024)) + (n.outer*32)), 1, 32)] = broadcast(0f32, 32)
+ C_3: Buffer(C_2, float32, [1048576], [])[ramp((((m.outer*32768) + (m.inner.init*1024)) + (n.outer*32)), 1, 32)] = broadcast(0f32, 32)
}
for (k.outer: int32, 0, 256) {
for (m.inner: int32, 0, 32) {
@@ -688,7 +684,7 @@ the access pattern for A matrix is more cache friendly.</p>
let cse_var_3: int32 = (n.outer*32)
let cse_var_2: int32 = ((m.outer*32768) + (m.inner*1024))
let cse_var_1: int32 = (cse_var_2 + cse_var_3)
- C[ramp(cse_var_1, 1, 32)] = (C[ramp(cse_var_1, 1, 32)] + (broadcast(A[((cse_var_2 + (k.outer*4)) + k.inner)], 32)*B[ramp((((k.outer*4096) + (k.inner*1024)) + cse_var_3), 1, 32)]))
+ C_3[ramp(cse_var_1, 1, 32)] = (C_3[ramp(cse_var_1, 1, 32)] + (broadcast(A_3: Buffer(A_2, float32, [1048576], [])[((cse_var_2 + (k.outer*4)) + k.inner)], 32)*B_3: Buffer(B_2, float32, [1048576], [])[ramp((((k.outer*4096) + (k.inner*1024)) + cse_var_3), 1, 32)]))
}
}
}
@@ -746,7 +742,7 @@ flattening.</p>
<span class="nb">print</span><span class="p">(</span><span class="s2">"Opt4: </span><span class="si">%f</span><span class="s2">"</span> <span class="o">%</span> <span class="n">evaluator</span><span class="p">(</span><span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">,</span> <span class="n">c</span><span class="p">)</span><span class="o">.</span><span class="n">mean</span><span class="p">)</span>
</pre></div>
</div>
-<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Opt4: 0.109638
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Opt4: 0.110101
</pre></div>
</div>
<p>Here is the generated IR after array packing.</p>
@@ -755,21 +751,20 @@ flattening.</p>
</div>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [1048576], []),
- B: Buffer(B_2: Pointer(float32), float32, [1048576], []),
- C: Buffer(C_2: Pointer(float32), float32, [1048576], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 1024], []), B_1: B_3: Buffer(B_2, float32, [1024, 1024], []), C_1: C_3: Buffer(C_2, float32, [1024, 1024], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),
+ B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
allocate(packedB: Pointer(global float32x32), float32x32, [32768]), storage_scope = global {
for (bigN: int32, 0, 32) "parallel" {
for (k: int32, 0, 1024) {
- packedB_1: Buffer(packedB, float32x32, [32768], [])[((bigN*1024) + k)] = B[ramp(((k*1024) + (bigN*32)), 1, 32)]
+ packedB_1: Buffer(packedB, float32x32, [32768], [])[((bigN*1024) + k)] = B_3: Buffer(B_2, float32, [1048576], [])[ramp(((k*1024) + (bigN*32)), 1, 32)]
}
}
for (m.outer: int32, 0, 32) {
for (n.outer: int32, 0, 32) {
for (m.inner.init: int32, 0, 32) {
- C[ramp((((m.outer*32768) + (m.inner.init*1024)) + (n.outer*32)), 1, 32)] = broadcast(0f32, 32)
+ C_3: Buffer(C_2, float32, [1048576], [])[ramp((((m.outer*32768) + (m.inner.init*1024)) + (n.outer*32)), 1, 32)] = broadcast(0f32, 32)
}
for (k.outer: int32, 0, 256) {
for (m.inner: int32, 0, 32) {
@@ -777,7 +772,7 @@ flattening.</p>
let cse_var_3: int32 = ((m.outer*32768) + (m.inner*1024))
let cse_var_2: int32 = (k.outer*4)
let cse_var_1: int32 = (cse_var_3 + (n.outer*32))
- C[ramp(cse_var_1, 1, 32)] = (C[ramp(cse_var_1, 1, 32)] + (broadcast(A[((cse_var_3 + cse_var_2) + k.inner)], 32)*packedB_1[(((n.outer*1024) + cse_var_2) + k.inner)]))
+ C_3[ramp(cse_var_1, 1, 32)] = (C_3[ramp(cse_var_1, 1, 32)] + (broadcast(A_3: Buffer(A_2, float32, [1048576], [])[((cse_var_3 + cse_var_2) + k.inner)], 32)*packedB_1[(((n.outer*1024) + cse_var_2) + k.inner)]))
}
}
}
@@ -832,7 +827,7 @@ write to C when all the block results are ready.</p>
<span class="nb">print</span><span class="p">(</span><span class="s2">"Opt5: </span><span class="si">%f</span><span class="s2">"</span> <span class="o">%</span> <span class="n">evaluator</span><span class="p">(</span><span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">,</span> <span class="n">c</span><span class="p">)</span><span class="o">.</span><span class="n">mean</span><span class="p">)</span>
</pre></div>
</div>
-<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Opt5: 0.115397
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Opt5: 0.111848
</pre></div>
</div>
<p>Here is the generated IR after blocking.</p>
@@ -841,16 +836,15 @@ write to C when all the block results are ready.</p>
</div>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [1048576], []),
- B: Buffer(B_2: Pointer(float32), float32, [1048576], []),
- C: Buffer(C_2: Pointer(float32), float32, [1048576], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 1024], []), B_1: B_3: Buffer(B_2, float32, [1024, 1024], []), C_1: C_3: Buffer(C_2, float32, [1024, 1024], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),
+ B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
allocate(packedB: Pointer(global float32x32), float32x32, [32768]), storage_scope = global;
allocate(C.global: Pointer(global float32), float32, [1024]), storage_scope = global {
for (bigN: int32, 0, 32) "parallel" {
for (k: int32, 0, 1024) {
- packedB_1: Buffer(packedB, float32x32, [32768], [])[((bigN*1024) + k)] = B[ramp(((k*1024) + (bigN*32)), 1, 32)]
+ packedB_1: Buffer(packedB, float32x32, [32768], [])[((bigN*1024) + k)] = B_3: Buffer(B_2, float32, [1048576], [])[ramp(((k*1024) + (bigN*32)), 1, 32)]
}
}
for (m.outer: int32, 0, 32) {
@@ -865,16 +859,16 @@ write to C when all the block results are ready.</p>
let cse_var_2: int32 = ((n.outer*1024) + cse_var_4)
let cse_var_1: int32 = (((m.outer*32768) + (m.c*1024)) + cse_var_4)
{
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[cse_var_1], 32)*packedB_1[cse_var_2]))
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[(cse_var_1 + 1)], 32)*packedB_1[(cse_var_2 + 1)]))
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[(cse_var_1 + 2)], 32)*packedB_1[(cse_var_2 + 2)]))
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[(cse_var_1 + 3)], 32)*packedB_1[(cse_var_2 + 3)]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3: Buffer(A_2, float32, [1048576], [])[cse_var_1], 32)*packedB_1[cse_var_2]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3[(cse_var_1 + 1)], 32)*packedB_1[(cse_var_2 + 1)]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3[(cse_var_1 + 2)], 32)*packedB_1[(cse_var_2 + 2)]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3[(cse_var_1 + 3)], 32)*packedB_1[(cse_var_2 + 3)]))
}
}
}
for (m.inner: int32, 0, 32) {
for (n.inner: int32, 0, 32) {
- C[((((m.outer*32768) + (m.inner*1024)) + (n.outer*32)) + n.inner)] = C.global_1[((m.inner*32) + n.inner)]
+ C_3: Buffer(C_2, float32, [1048576], [])[((((m.outer*32768) + (m.inner*1024)) + (n.outer*32)) + n.inner)] = C.global_1[((m.inner*32) + n.inner)]
}
}
}
@@ -922,7 +916,7 @@ write to C when all the block results are ready.</p>
<span class="nb">print</span><span class="p">(</span><span class="s2">"Opt6: </span><span class="si">%f</span><span class="s2">"</span> <span class="o">%</span> <span class="n">opt6_time</span><span class="p">)</span>
</pre></div>
</div>
-<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Opt6: 0.152455
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Opt6: 0.147245
</pre></div>
</div>
<p>Here is the generated IR after parallelization.</p>
@@ -931,15 +925,14 @@ write to C when all the block results are ready.</p>
</div>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@main = primfn(A_1: handle, B_1: handle, C_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [1048576], []),
- B: Buffer(B_2: Pointer(float32), float32, [1048576], []),
- C: Buffer(C_2: Pointer(float32), float32, [1048576], [])}
- buffer_map = {A_1: A, B_1: B, C_1: C}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [1024, 1024], []), B_1: B_3: Buffer(B_2, float32, [1024, 1024], []), C_1: C_3: Buffer(C_2, float32, [1024, 1024], [])} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),
+ B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], []),
+ C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], [])}
+ buffer_map = {A_1: A, B_1: B, C_1: C} {
allocate(packedB: Pointer(global float32x32), float32x32, [32768]), storage_scope = global {
for (bigN: int32, 0, 32) "parallel" {
for (k: int32, 0, 1024) {
- packedB_1: Buffer(packedB, float32x32, [32768], [])[((bigN*1024) + k)] = B[ramp(((k*1024) + (bigN*32)), 1, 32)]
+ packedB_1: Buffer(packedB, float32x32, [32768], [])[((bigN*1024) + k)] = B_3: Buffer(B_2, float32, [1048576], [])[ramp(((k*1024) + (bigN*32)), 1, 32)]
}
}
for (m.outer: int32, 0, 32) "parallel" {
@@ -955,16 +948,16 @@ write to C when all the block results are ready.</p>
let cse_var_2: int32 = ((n.outer*1024) + cse_var_4)
let cse_var_1: int32 = (((m.outer*32768) + (m.c*1024)) + cse_var_4)
{
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[cse_var_1], 32)*packedB_1[cse_var_2]))
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[(cse_var_1 + 1)], 32)*packedB_1[(cse_var_2 + 1)]))
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[(cse_var_1 + 2)], 32)*packedB_1[(cse_var_2 + 2)]))
- C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A[(cse_var_1 + 3)], 32)*packedB_1[(cse_var_2 + 3)]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3: Buffer(A_2, float32, [1048576], [])[cse_var_1], 32)*packedB_1[cse_var_2]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3[(cse_var_1 + 1)], 32)*packedB_1[(cse_var_2 + 1)]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3[(cse_var_1 + 2)], 32)*packedB_1[(cse_var_2 + 2)]))
+ C.global_1[ramp(cse_var_3, 1, 32)] = (C.global_1[ramp(cse_var_3, 1, 32)] + (broadcast(A_3[(cse_var_1 + 3)], 32)*packedB_1[(cse_var_2 + 3)]))
}
}
}
for (m.inner: int32, 0, 32) {
for (n.inner: int32, 0, 32) {
- C[((((m.outer*32768) + (m.inner*1024)) + (n.outer*32)) + n.inner)] = C.global_1[((m.inner*32) + n.inner)]
+ C_3: Buffer(C_2, float32, [1048576], [])[((((m.outer*32768) + (m.inner*1024)) + (n.outer*32)) + n.inner)] = C.global_1[((m.inner*32) + n.inner)]
}
}
}
diff --git a/docs/how_to/optimize_operators/sg_execution_times.html b/docs/how_to/optimize_operators/sg_execution_times.html
index 727942bd80..a08aa0ef96 100644
--- a/docs/how_to/optimize_operators/sg_execution_times.html
+++ b/docs/how_to/optimize_operators/sg_execution_times.html
@@ -340,7 +340,7 @@
<div class="section" id="computation-times">
<span id="sphx-glr-how-to-optimize-operators-sg-execution-times"></span><h1>Computation times<a class="headerlink" href="#computation-times" title="Permalink to this headline">¶</a></h1>
-<p><strong>00:35.446</strong> total execution time for <strong>how_to_optimize_operators</strong> files:</p>
+<p><strong>00:35.417</strong> total execution time for <strong>how_to_optimize_operators</strong> files:</p>
<table class="docutils align-default">
<colgroup>
<col style="width: 83%" />
@@ -349,15 +349,15 @@
</colgroup>
<tbody>
<tr class="row-odd"><td><p><a class="reference internal" href="opt_gemm.html#sphx-glr-how-to-optimize-operators-opt-gemm-py"><span class="std std-ref">How to optimize GEMM on CPU</span></a> (<code class="docutils literal notranslate"><span class="pre">opt_gemm.py</span></code>)</p></td>
-<td><p>00:32.843</p></td>
+<td><p>00:32.836</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="opt_conv_tensorcore.html#sphx-glr-how-to-optimize-operators-opt-conv-tensorcore-py"><span class="std std-ref">How to optimize convolution using TensorCores</span></a> (<code class="docutils literal notranslate"><span class="pre">opt_conv_tensorcore.py</span></code>)</p></td>
-<td><p>00:01.481</p></td>
+<td><p>00:01.521</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="opt_conv_cuda.html#sphx-glr-how-to-optimize-operators-opt-conv-cuda-py"><span class="std std-ref">How to optimize convolution on GPU</span></a> (<code class="docutils literal notranslate"><span class="pre">opt_conv_cuda.py</span></code>)</p></td>
-<td><p>00:01.121</p></td>
+<td><p>00:01.061</p></td>
<td><p>0.0 MB</p></td>
</tr>
</tbody>
diff --git a/docs/how_to/tune_with_autoscheduler/sg_execution_times.html b/docs/how_to/tune_with_autoscheduler/sg_execution_times.html
index 0e561d0a31..d85ef7c18b 100644
--- a/docs/how_to/tune_with_autoscheduler/sg_execution_times.html
+++ b/docs/how_to/tune_with_autoscheduler/sg_execution_times.html
@@ -340,7 +340,7 @@
<div class="section" id="computation-times">
<span id="sphx-glr-how-to-tune-with-autoscheduler-sg-execution-times"></span><h1>Computation times<a class="headerlink" href="#computation-times" title="Permalink to this headline">¶</a></h1>
-<p><strong>09:11.050</strong> total execution time for <strong>how_to_tune_with_autoscheduler</strong> files:</p>
+<p><strong>09:43.171</strong> total execution time for <strong>how_to_tune_with_autoscheduler</strong> files:</p>
<table class="docutils align-default">
<colgroup>
<col style="width: 85%" />
@@ -349,27 +349,27 @@
</colgroup>
<tbody>
<tr class="row-odd"><td><p><a class="reference internal" href="tune_conv2d_layer_cuda.html#sphx-glr-how-to-tune-with-autoscheduler-tune-conv2d-layer-cuda-py"><span class="std std-ref">Auto-scheduling a Convolution Layer for GPU</span></a> (<code class="docutils literal notranslate"><span class="pre">tune_conv2d_layer_cuda.py</span></code>)</p></td>
-<td><p>05:37.751</p></td>
+<td><p>06:07.710</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="tune_network_x86.html#sphx-glr-how-to-tune-with-autoscheduler-tune-network-x86-py"><span class="std std-ref">Auto-scheduling a Neural Network for x86 CPU</span></a> (<code class="docutils literal notranslate"><span class="pre">tune_network_x86.py</span></code>)</p></td>
-<td><p>01:34.377</p></td>
+<td><p>01:34.567</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="tune_network_cuda.html#sphx-glr-how-to-tune-with-autoscheduler-tune-network-cuda-py"><span class="std std-ref">Auto-scheduling a Neural Network for NVIDIA GPU</span></a> (<code class="docutils literal notranslate"><span class="pre">tune_network_cuda.py</span></code>)</p></td>
-<td><p>01:04.612</p></td>
+<td><p>01:04.692</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="tune_sparse_x86.html#sphx-glr-how-to-tune-with-autoscheduler-tune-sparse-x86-py"><span class="std std-ref">Auto-scheduling Sparse Matrix Multiplication on CPU with Custom Sketch Rule</span></a> (<code class="docutils literal notranslate"><span class="pre">tune_sparse_x86.py</span></code>)</p></td>
-<td><p>00:30.535</p></td>
+<td><p>00:31.234</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="tune_network_arm.html#sphx-glr-how-to-tune-with-autoscheduler-tune-network-arm-py"><span class="std std-ref">Auto-scheduling a Neural Network for ARM CPU</span></a> (<code class="docutils literal notranslate"><span class="pre">tune_network_arm.py</span></code>)</p></td>
-<td><p>00:12.286</p></td>
+<td><p>00:12.963</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="tune_network_mali.html#sphx-glr-how-to-tune-with-autoscheduler-tune-network-mali-py"><span class="std std-ref">Auto-scheduling a Neural Network for mali GPU</span></a> (<code class="docutils literal notranslate"><span class="pre">tune_network_mali.py</span></code>)</p></td>
-<td><p>00:11.490</p></td>
+<td><p>00:12.004</p></td>
<td><p>0.0 MB</p></td>
</tr>
</tbody>
diff --git a/docs/how_to/tune_with_autoscheduler/tune_conv2d_layer_cuda.html b/docs/how_to/tune_with_autoscheduler/tune_conv2d_layer_cuda.html
index 74dbee8a8f..784debe5c7 100644
--- a/docs/how_to/tune_with_autoscheduler/tune_conv2d_layer_cuda.html
+++ b/docs/how_to/tune_with_autoscheduler/tune_conv2d_layer_cuda.html
@@ -498,109 +498,488 @@ cooperative fetching, unrolling and operator fusion.</p>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Lowered TIR:
@main = primfn(data_1: handle, kernel_1: handle, bias_1: handle, compute_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {data: Buffer(data_2: Pointer(float32), float32, [25088], []),
- kernel: Buffer(kernel_2: Pointer(float32), float32, [2359296], []),
- bias: Buffer(bias_2: Pointer(float32), float32, [512], []),
- compute: Buffer(compute_2: Pointer(float32), float32, [25088], [])}
- buffer_map = {data_1: data, kernel_1: kernel, bias_1: bias, compute_1: compute}
- preflattened_buffer_map = {data_1: data_3: Buffer(data_2, float32, [1, 512, 7, 7], []), kernel_1: kernel_3: Buffer(kernel_2, float32, [512, 512, 3, 3], []), bias_1: bias_3: Buffer(bias_2, float32, [1, 512, 1, 1], []), compute_1: compute_3: Buffer(compute_2, float32, [1, 512, 7, 7], [])} {
- attr [IterVar(blockIdx.x: int32, (nullptr), "ThreadIndex", "blockIdx.x")] "thread_extent" = 64;
- allocate(conv2d_nchw: Pointer(local float32), float32, [4]), storage_scope = local;
- allocate(pad_temp.shared: Pointer(shared float32), float32, [162]), storage_scope = shared;
- allocate(kernel.shared: Pointer(shared float32), float32, [144]), storage_scope = shared;
- attr [IterVar(threadIdx.x: int32, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 98 {
- conv2d_nchw_1: Buffer(conv2d_nchw, float32, [4], [], scope="local", align=16)[0] = 0f32
+ buffers = {data: Buffer(data_2: Pointer(float32), float32, [1, 512, 7, 7], []),
+ kernel: Buffer(kernel_2: Pointer(float32), float32, [512, 512, 3, 3], []),
+ bias: Buffer(bias_2: Pointer(float32), float32, [1, 512, 1, 1], []),
+ compute: Buffer(compute_2: Pointer(float32), float32, [1, 512, 7, 7], [])}
+ buffer_map = {data_1: data, kernel_1: kernel, bias_1: bias, compute_1: compute} {
+ attr [IterVar(blockIdx.x: int32, (nullptr), "ThreadIndex", "blockIdx.x")] "thread_extent" = 28;
+ allocate(conv2d_nchw: Pointer(local float32), float32, [14]), storage_scope = local;
+ allocate(pad_temp.shared: Pointer(shared float32), float32, [72]), storage_scope = shared;
+ allocate(kernel.shared: Pointer(shared float32), float32, [3072]), storage_scope = shared;
+ attr [IterVar(threadIdx.x: int32, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64 {
+ conv2d_nchw_1: Buffer(conv2d_nchw, float32, [14], [], scope="local", align=32)[0] = 0f32
conv2d_nchw_1[1] = 0f32
conv2d_nchw_1[2] = 0f32
conv2d_nchw_1[3] = 0f32
- for (rc.outer.outer: int32, 0, 256) {
- attr [IterVar(threadIdx.x_1: int32, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 98;
- pad_temp.shared_1: Buffer(pad_temp.shared, float32, [162], [], scope="shared")[threadIdx.x_1] = @tir.if_then_else(((((9 <= floormod(threadIdx.x_1, 81)) && (floormod(threadIdx.x_1, 81) < 72)) && (1 <= floormod(threadIdx.x_1, 9))) && (floormod(threadIdx.x_1, 9) < 8)), data[(((((rc.outer.outer*98) + (floordiv(threadIdx.x_1, 81)*49)) + (floordiv(floormod(threadIdx.x_1, 81), 9)*7)) + floormod(threadIdx.x_1, 9)) - 8)], 0f32, dtype=float32)
- attr [IterVar(threadIdx.x_1, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 98;
- if @tir.likely((threadIdx.x_1 < 64), dtype=bool) {
- pad_temp.shared_1[(threadIdx.x_1 + 98)] = @tir.if_then_else((((threadIdx.x_1 < 55) && (1 <= floormod((threadIdx.x_1 + 8), 9))) && (floormod((threadIdx.x_1 + 8), 9) < 8)), data[(((((rc.outer.outer*98) + (floordiv((threadIdx.x_1 + 98), 81)*49)) + (floordiv((threadIdx.x_1 + 17), 9)*7)) + floormod((threadIdx.x_1 + 8), 9)) - 8)], 0f32, dtype=float32)
+ conv2d_nchw_1[4] = 0f32
+ conv2d_nchw_1[5] = 0f32
+ conv2d_nchw_1[6] = 0f32
+ conv2d_nchw_1[7] = 0f32
+ conv2d_nchw_1[8] = 0f32
+ conv2d_nchw_1[9] = 0f32
+ conv2d_nchw_1[10] = 0f32
+ conv2d_nchw_1[11] = 0f32
+ conv2d_nchw_1[12] = 0f32
+ conv2d_nchw_1[13] = 0f32
+ for (rc.outer.outer: int32, 0, 64) {
+ for (ry.outer.outer: int32, 0, 3) {
+ let cse_var_2: int32 = (rc.outer.outer*72)
+ let cse_var_1: int32 = (ry.outer.outer*3)
+ {
+ attr [IterVar(threadIdx.x_1: int32, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64 {
+ if @tir.likely((threadIdx.x_1 < 18), dtype=bool) {
+ pad_temp.shared_1: Buffer(pad_temp.shared, float32, [72], [], scope="shared")[(threadIdx.x_1*4)] = @tir.if_then_else(((((1 <= (ry.outer.outer + floormod(blockIdx.x, 7))) && ((ry.outer.outer + floormod(blockIdx.x, 7)) < 8)) && (1 <= floormod((threadIdx.x_1*4), 9))) && (floormod((threadIdx.x_1*4), 9) < 8)), data_3: Buffer(data_2, float32, [25088], [])[((((((rc.outer.outer*392) + (floordiv((threadIdx.x_1*4), 9)*49)) + (ry.outer.out [...]
+ }
+ if @tir.likely((threadIdx.x_1 < 18), dtype=bool) {
+ pad_temp.shared_1[((threadIdx.x_1*4) + 1)] = @tir.if_then_else(((((1 <= (ry.outer.outer + floormod(blockIdx.x, 7))) && ((ry.outer.outer + floormod(blockIdx.x, 7)) < 8)) && (1 <= floormod(((threadIdx.x_1*4) + 1), 9))) && (floormod(((threadIdx.x_1*4) + 1), 9) < 8)), data_3[((((((rc.outer.outer*392) + (floordiv(((threadIdx.x_1*4) + 1), 9)*49)) + (ry.outer.outer*7)) + (floormod(blockIdx.x, 7)*7)) + floormod(((threadIdx.x_1*4) + 1), 9)) - 8)], [...]
+ }
+ if @tir.likely((threadIdx.x_1 < 18), dtype=bool) {
+ pad_temp.shared_1[((threadIdx.x_1*4) + 2)] = @tir.if_then_else(((((1 <= (ry.outer.outer + floormod(blockIdx.x, 7))) && ((ry.outer.outer + floormod(blockIdx.x, 7)) < 8)) && (1 <= floormod(((threadIdx.x_1*4) + 2), 9))) && (floormod(((threadIdx.x_1*4) + 2), 9) < 8)), data_3[((((((rc.outer.outer*392) + (floordiv(((threadIdx.x_1*4) + 2), 9)*49)) + (ry.outer.outer*7)) + (floormod(blockIdx.x, 7)*7)) + floormod(((threadIdx.x_1*4) + 2), 9)) - 8)], [...]
+ }
+ if @tir.likely((threadIdx.x_1 < 18), dtype=bool) {
+ pad_temp.shared_1[((threadIdx.x_1*4) + 3)] = @tir.if_then_else(((((1 <= (ry.outer.outer + floormod(blockIdx.x, 7))) && ((ry.outer.outer + floormod(blockIdx.x, 7)) < 8)) && (1 <= floormod(((threadIdx.x_1*4) + 3), 9))) && (floormod(((threadIdx.x_1*4) + 3), 9) < 8)), data_3[((((((rc.outer.outer*392) + (floordiv(((threadIdx.x_1*4) + 3), 9)*49)) + (ry.outer.outer*7)) + (floormod(blockIdx.x, 7)*7)) + floormod(((threadIdx.x_1*4) + 3), 9)) - 8)], [...]
+ }
+ }
+ attr [IterVar(threadIdx.x_2: int32, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1: Buffer(kernel.shared, float32, [3072], [], scope="shared")[threadIdx.x_2] = kernel_3: Buffer(kernel_2, float32, [2359296], [])[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 64)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 64), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 128)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 128), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 192)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 36864)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 256)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 256), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 320)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 320), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 384)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 73728)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 448)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 448), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 512)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 512), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 576)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 110592)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 640)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 640), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 704)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 704), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 768)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 147456)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 832)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 832), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 896)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 896), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 960)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 184320)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1024)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 1024), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1088)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 1088), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1152)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 221184)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1216)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 1216), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1280)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 1280), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1344)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 258048)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1408)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 1408), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1472)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 1472), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1536)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 294912)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1600)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 1600), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1664)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 1664), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1728)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 331776)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1792)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 1792), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1856)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 1856), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1920)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 368640)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 1984)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 1984), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2048)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 2048), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2112)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 405504)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2176)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 2176), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2240)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 2240), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2304)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 442368)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2368)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 2368), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2432)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 2432), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2496)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 479232)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2560)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 2560), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2624)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 2624), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2688)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 516096)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2752)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 2752), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2816)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 2816), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2880)] = kernel_3[(((((((floordiv(blockIdx.x, 7)*589824) + (floordiv(threadIdx.x_2, 24)*4608)) + cse_var_2) + (floordiv(floormod(threadIdx.x_2, 24), 3)*9)) + cse_var_1) + floormod(threadIdx.x_2, 3)) + 552960)]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 2944)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 2944), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 1), 3))]
+ attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64;
+ kernel.shared_1[(threadIdx.x_2 + 3008)] = kernel_3[((((((floordiv(blockIdx.x, 7)*589824) + (floordiv((threadIdx.x_2 + 3008), 24)*4608)) + cse_var_2) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)*9)) + cse_var_1) + floormod((threadIdx.x_2 + 2), 3))]
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[0]*kernel.shared_1[(threadIdx.x*48)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[9]*kernel.shared_1[((threadIdx.x*48) + 3)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[1]*kernel.shared_1[(threadIdx.x*48)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[10]*kernel.shared_1[((threadIdx.x*48) + 3)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[2]*kernel.shared_1[(threadIdx.x*48)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[11]*kernel.shared_1[((threadIdx.x*48) + 3)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[3]*kernel.shared_1[(threadIdx.x*48)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[12]*kernel.shared_1[((threadIdx.x*48) + 3)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[4]*kernel.shared_1[(threadIdx.x*48)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[13]*kernel.shared_1[((threadIdx.x*48) + 3)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[5]*kernel.shared_1[(threadIdx.x*48)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[14]*kernel.shared_1[((threadIdx.x*48) + 3)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[6]*kernel.shared_1[(threadIdx.x*48)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[15]*kernel.shared_1[((threadIdx.x*48) + 3)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[0]*kernel.shared_1[((threadIdx.x*48) + 24)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[9]*kernel.shared_1[((threadIdx.x*48) + 27)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[1]*kernel.shared_1[((threadIdx.x*48) + 24)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[10]*kernel.shared_1[((threadIdx.x*48) + 27)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[2]*kernel.shared_1[((threadIdx.x*48) + 24)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[11]*kernel.shared_1[((threadIdx.x*48) + 27)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[3]*kernel.shared_1[((threadIdx.x*48) + 24)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[12]*kernel.shared_1[((threadIdx.x*48) + 27)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[4]*kernel.shared_1[((threadIdx.x*48) + 24)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[13]*kernel.shared_1[((threadIdx.x*48) + 27)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[5]*kernel.shared_1[((threadIdx.x*48) + 24)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[14]*kernel.shared_1[((threadIdx.x*48) + 27)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[6]*kernel.shared_1[((threadIdx.x*48) + 24)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[15]*kernel.shared_1[((threadIdx.x*48) + 27)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[1]*kernel.shared_1[((threadIdx.x*48) + 1)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[10]*kernel.shared_1[((threadIdx.x*48) + 4)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[2]*kernel.shared_1[((threadIdx.x*48) + 1)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[11]*kernel.shared_1[((threadIdx.x*48) + 4)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[3]*kernel.shared_1[((threadIdx.x*48) + 1)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[12]*kernel.shared_1[((threadIdx.x*48) + 4)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[4]*kernel.shared_1[((threadIdx.x*48) + 1)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[13]*kernel.shared_1[((threadIdx.x*48) + 4)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[5]*kernel.shared_1[((threadIdx.x*48) + 1)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[14]*kernel.shared_1[((threadIdx.x*48) + 4)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[6]*kernel.shared_1[((threadIdx.x*48) + 1)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[15]*kernel.shared_1[((threadIdx.x*48) + 4)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[7]*kernel.shared_1[((threadIdx.x*48) + 1)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[16]*kernel.shared_1[((threadIdx.x*48) + 4)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[1]*kernel.shared_1[((threadIdx.x*48) + 25)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[10]*kernel.shared_1[((threadIdx.x*48) + 28)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[2]*kernel.shared_1[((threadIdx.x*48) + 25)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[11]*kernel.shared_1[((threadIdx.x*48) + 28)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[3]*kernel.shared_1[((threadIdx.x*48) + 25)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[12]*kernel.shared_1[((threadIdx.x*48) + 28)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[4]*kernel.shared_1[((threadIdx.x*48) + 25)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[13]*kernel.shared_1[((threadIdx.x*48) + 28)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[5]*kernel.shared_1[((threadIdx.x*48) + 25)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[14]*kernel.shared_1[((threadIdx.x*48) + 28)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[6]*kernel.shared_1[((threadIdx.x*48) + 25)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[15]*kernel.shared_1[((threadIdx.x*48) + 28)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[7]*kernel.shared_1[((threadIdx.x*48) + 25)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[16]*kernel.shared_1[((threadIdx.x*48) + 28)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[2]*kernel.shared_1[((threadIdx.x*48) + 2)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[11]*kernel.shared_1[((threadIdx.x*48) + 5)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[3]*kernel.shared_1[((threadIdx.x*48) + 2)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[12]*kernel.shared_1[((threadIdx.x*48) + 5)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[4]*kernel.shared_1[((threadIdx.x*48) + 2)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[13]*kernel.shared_1[((threadIdx.x*48) + 5)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[5]*kernel.shared_1[((threadIdx.x*48) + 2)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[14]*kernel.shared_1[((threadIdx.x*48) + 5)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[6]*kernel.shared_1[((threadIdx.x*48) + 2)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[15]*kernel.shared_1[((threadIdx.x*48) + 5)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[7]*kernel.shared_1[((threadIdx.x*48) + 2)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[16]*kernel.shared_1[((threadIdx.x*48) + 5)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[8]*kernel.shared_1[((threadIdx.x*48) + 2)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[17]*kernel.shared_1[((threadIdx.x*48) + 5)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[2]*kernel.shared_1[((threadIdx.x*48) + 26)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[11]*kernel.shared_1[((threadIdx.x*48) + 29)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[3]*kernel.shared_1[((threadIdx.x*48) + 26)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[12]*kernel.shared_1[((threadIdx.x*48) + 29)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[4]*kernel.shared_1[((threadIdx.x*48) + 26)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[13]*kernel.shared_1[((threadIdx.x*48) + 29)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[5]*kernel.shared_1[((threadIdx.x*48) + 26)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[14]*kernel.shared_1[((threadIdx.x*48) + 29)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[6]*kernel.shared_1[((threadIdx.x*48) + 26)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[15]*kernel.shared_1[((threadIdx.x*48) + 29)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[7]*kernel.shared_1[((threadIdx.x*48) + 26)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[16]*kernel.shared_1[((threadIdx.x*48) + 29)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[8]*kernel.shared_1[((threadIdx.x*48) + 26)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[17]*kernel.shared_1[((threadIdx.x*48) + 29)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[18]*kernel.shared_1[((threadIdx.x*48) + 6)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[27]*kernel.shared_1[((threadIdx.x*48) + 9)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[19]*kernel.shared_1[((threadIdx.x*48) + 6)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[28]*kernel.shared_1[((threadIdx.x*48) + 9)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[20]*kernel.shared_1[((threadIdx.x*48) + 6)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[29]*kernel.shared_1[((threadIdx.x*48) + 9)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[21]*kernel.shared_1[((threadIdx.x*48) + 6)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[30]*kernel.shared_1[((threadIdx.x*48) + 9)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[22]*kernel.shared_1[((threadIdx.x*48) + 6)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[31]*kernel.shared_1[((threadIdx.x*48) + 9)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[23]*kernel.shared_1[((threadIdx.x*48) + 6)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[32]*kernel.shared_1[((threadIdx.x*48) + 9)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[24]*kernel.shared_1[((threadIdx.x*48) + 6)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[33]*kernel.shared_1[((threadIdx.x*48) + 9)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[18]*kernel.shared_1[((threadIdx.x*48) + 30)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[27]*kernel.shared_1[((threadIdx.x*48) + 33)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[19]*kernel.shared_1[((threadIdx.x*48) + 30)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[28]*kernel.shared_1[((threadIdx.x*48) + 33)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[20]*kernel.shared_1[((threadIdx.x*48) + 30)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[29]*kernel.shared_1[((threadIdx.x*48) + 33)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[21]*kernel.shared_1[((threadIdx.x*48) + 30)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[30]*kernel.shared_1[((threadIdx.x*48) + 33)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[22]*kernel.shared_1[((threadIdx.x*48) + 30)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[31]*kernel.shared_1[((threadIdx.x*48) + 33)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[23]*kernel.shared_1[((threadIdx.x*48) + 30)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[32]*kernel.shared_1[((threadIdx.x*48) + 33)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[24]*kernel.shared_1[((threadIdx.x*48) + 30)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[33]*kernel.shared_1[((threadIdx.x*48) + 33)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[19]*kernel.shared_1[((threadIdx.x*48) + 7)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[28]*kernel.shared_1[((threadIdx.x*48) + 10)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[20]*kernel.shared_1[((threadIdx.x*48) + 7)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[29]*kernel.shared_1[((threadIdx.x*48) + 10)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[21]*kernel.shared_1[((threadIdx.x*48) + 7)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[30]*kernel.shared_1[((threadIdx.x*48) + 10)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[22]*kernel.shared_1[((threadIdx.x*48) + 7)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[31]*kernel.shared_1[((threadIdx.x*48) + 10)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[23]*kernel.shared_1[((threadIdx.x*48) + 7)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[32]*kernel.shared_1[((threadIdx.x*48) + 10)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[24]*kernel.shared_1[((threadIdx.x*48) + 7)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[33]*kernel.shared_1[((threadIdx.x*48) + 10)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[25]*kernel.shared_1[((threadIdx.x*48) + 7)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[34]*kernel.shared_1[((threadIdx.x*48) + 10)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[19]*kernel.shared_1[((threadIdx.x*48) + 31)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[28]*kernel.shared_1[((threadIdx.x*48) + 34)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[20]*kernel.shared_1[((threadIdx.x*48) + 31)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[29]*kernel.shared_1[((threadIdx.x*48) + 34)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[21]*kernel.shared_1[((threadIdx.x*48) + 31)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[30]*kernel.shared_1[((threadIdx.x*48) + 34)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[22]*kernel.shared_1[((threadIdx.x*48) + 31)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[31]*kernel.shared_1[((threadIdx.x*48) + 34)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[23]*kernel.shared_1[((threadIdx.x*48) + 31)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[32]*kernel.shared_1[((threadIdx.x*48) + 34)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[24]*kernel.shared_1[((threadIdx.x*48) + 31)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[33]*kernel.shared_1[((threadIdx.x*48) + 34)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[25]*kernel.shared_1[((threadIdx.x*48) + 31)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[34]*kernel.shared_1[((threadIdx.x*48) + 34)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[20]*kernel.shared_1[((threadIdx.x*48) + 8)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[29]*kernel.shared_1[((threadIdx.x*48) + 11)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[21]*kernel.shared_1[((threadIdx.x*48) + 8)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[30]*kernel.shared_1[((threadIdx.x*48) + 11)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[22]*kernel.shared_1[((threadIdx.x*48) + 8)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[31]*kernel.shared_1[((threadIdx.x*48) + 11)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[23]*kernel.shared_1[((threadIdx.x*48) + 8)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[32]*kernel.shared_1[((threadIdx.x*48) + 11)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[24]*kernel.shared_1[((threadIdx.x*48) + 8)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[33]*kernel.shared_1[((threadIdx.x*48) + 11)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[25]*kernel.shared_1[((threadIdx.x*48) + 8)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[34]*kernel.shared_1[((threadIdx.x*48) + 11)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[26]*kernel.shared_1[((threadIdx.x*48) + 8)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[35]*kernel.shared_1[((threadIdx.x*48) + 11)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[20]*kernel.shared_1[((threadIdx.x*48) + 32)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[29]*kernel.shared_1[((threadIdx.x*48) + 35)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[21]*kernel.shared_1[((threadIdx.x*48) + 32)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[30]*kernel.shared_1[((threadIdx.x*48) + 35)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[22]*kernel.shared_1[((threadIdx.x*48) + 32)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[31]*kernel.shared_1[((threadIdx.x*48) + 35)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[23]*kernel.shared_1[((threadIdx.x*48) + 32)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[32]*kernel.shared_1[((threadIdx.x*48) + 35)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[24]*kernel.shared_1[((threadIdx.x*48) + 32)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[33]*kernel.shared_1[((threadIdx.x*48) + 35)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[25]*kernel.shared_1[((threadIdx.x*48) + 32)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[34]*kernel.shared_1[((threadIdx.x*48) + 35)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[26]*kernel.shared_1[((threadIdx.x*48) + 32)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[35]*kernel.shared_1[((threadIdx.x*48) + 35)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[36]*kernel.shared_1[((threadIdx.x*48) + 12)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[45]*kernel.shared_1[((threadIdx.x*48) + 15)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[37]*kernel.shared_1[((threadIdx.x*48) + 12)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[46]*kernel.shared_1[((threadIdx.x*48) + 15)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[38]*kernel.shared_1[((threadIdx.x*48) + 12)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[47]*kernel.shared_1[((threadIdx.x*48) + 15)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[39]*kernel.shared_1[((threadIdx.x*48) + 12)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[48]*kernel.shared_1[((threadIdx.x*48) + 15)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[40]*kernel.shared_1[((threadIdx.x*48) + 12)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[49]*kernel.shared_1[((threadIdx.x*48) + 15)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[41]*kernel.shared_1[((threadIdx.x*48) + 12)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[50]*kernel.shared_1[((threadIdx.x*48) + 15)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[42]*kernel.shared_1[((threadIdx.x*48) + 12)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[51]*kernel.shared_1[((threadIdx.x*48) + 15)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[36]*kernel.shared_1[((threadIdx.x*48) + 36)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[45]*kernel.shared_1[((threadIdx.x*48) + 39)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[37]*kernel.shared_1[((threadIdx.x*48) + 36)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[46]*kernel.shared_1[((threadIdx.x*48) + 39)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[38]*kernel.shared_1[((threadIdx.x*48) + 36)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[47]*kernel.shared_1[((threadIdx.x*48) + 39)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[39]*kernel.shared_1[((threadIdx.x*48) + 36)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[48]*kernel.shared_1[((threadIdx.x*48) + 39)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[40]*kernel.shared_1[((threadIdx.x*48) + 36)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[49]*kernel.shared_1[((threadIdx.x*48) + 39)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[41]*kernel.shared_1[((threadIdx.x*48) + 36)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[50]*kernel.shared_1[((threadIdx.x*48) + 39)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[42]*kernel.shared_1[((threadIdx.x*48) + 36)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[51]*kernel.shared_1[((threadIdx.x*48) + 39)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[37]*kernel.shared_1[((threadIdx.x*48) + 13)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[46]*kernel.shared_1[((threadIdx.x*48) + 16)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[38]*kernel.shared_1[((threadIdx.x*48) + 13)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[47]*kernel.shared_1[((threadIdx.x*48) + 16)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[39]*kernel.shared_1[((threadIdx.x*48) + 13)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[48]*kernel.shared_1[((threadIdx.x*48) + 16)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[40]*kernel.shared_1[((threadIdx.x*48) + 13)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[49]*kernel.shared_1[((threadIdx.x*48) + 16)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[41]*kernel.shared_1[((threadIdx.x*48) + 13)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[50]*kernel.shared_1[((threadIdx.x*48) + 16)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[42]*kernel.shared_1[((threadIdx.x*48) + 13)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[51]*kernel.shared_1[((threadIdx.x*48) + 16)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[43]*kernel.shared_1[((threadIdx.x*48) + 13)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[52]*kernel.shared_1[((threadIdx.x*48) + 16)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[37]*kernel.shared_1[((threadIdx.x*48) + 37)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[46]*kernel.shared_1[((threadIdx.x*48) + 40)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[38]*kernel.shared_1[((threadIdx.x*48) + 37)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[47]*kernel.shared_1[((threadIdx.x*48) + 40)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[39]*kernel.shared_1[((threadIdx.x*48) + 37)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[48]*kernel.shared_1[((threadIdx.x*48) + 40)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[40]*kernel.shared_1[((threadIdx.x*48) + 37)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[49]*kernel.shared_1[((threadIdx.x*48) + 40)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[41]*kernel.shared_1[((threadIdx.x*48) + 37)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[50]*kernel.shared_1[((threadIdx.x*48) + 40)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[42]*kernel.shared_1[((threadIdx.x*48) + 37)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[51]*kernel.shared_1[((threadIdx.x*48) + 40)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[43]*kernel.shared_1[((threadIdx.x*48) + 37)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[52]*kernel.shared_1[((threadIdx.x*48) + 40)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[38]*kernel.shared_1[((threadIdx.x*48) + 14)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[47]*kernel.shared_1[((threadIdx.x*48) + 17)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[39]*kernel.shared_1[((threadIdx.x*48) + 14)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[48]*kernel.shared_1[((threadIdx.x*48) + 17)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[40]*kernel.shared_1[((threadIdx.x*48) + 14)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[49]*kernel.shared_1[((threadIdx.x*48) + 17)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[41]*kernel.shared_1[((threadIdx.x*48) + 14)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[50]*kernel.shared_1[((threadIdx.x*48) + 17)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[42]*kernel.shared_1[((threadIdx.x*48) + 14)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[51]*kernel.shared_1[((threadIdx.x*48) + 17)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[43]*kernel.shared_1[((threadIdx.x*48) + 14)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[52]*kernel.shared_1[((threadIdx.x*48) + 17)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[44]*kernel.shared_1[((threadIdx.x*48) + 14)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[53]*kernel.shared_1[((threadIdx.x*48) + 17)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[38]*kernel.shared_1[((threadIdx.x*48) + 38)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[47]*kernel.shared_1[((threadIdx.x*48) + 41)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[39]*kernel.shared_1[((threadIdx.x*48) + 38)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[48]*kernel.shared_1[((threadIdx.x*48) + 41)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[40]*kernel.shared_1[((threadIdx.x*48) + 38)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[49]*kernel.shared_1[((threadIdx.x*48) + 41)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[41]*kernel.shared_1[((threadIdx.x*48) + 38)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[50]*kernel.shared_1[((threadIdx.x*48) + 41)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[42]*kernel.shared_1[((threadIdx.x*48) + 38)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[51]*kernel.shared_1[((threadIdx.x*48) + 41)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[43]*kernel.shared_1[((threadIdx.x*48) + 38)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[52]*kernel.shared_1[((threadIdx.x*48) + 41)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[44]*kernel.shared_1[((threadIdx.x*48) + 38)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[53]*kernel.shared_1[((threadIdx.x*48) + 41)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[54]*kernel.shared_1[((threadIdx.x*48) + 18)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[63]*kernel.shared_1[((threadIdx.x*48) + 21)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[55]*kernel.shared_1[((threadIdx.x*48) + 18)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[64]*kernel.shared_1[((threadIdx.x*48) + 21)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[56]*kernel.shared_1[((threadIdx.x*48) + 18)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[65]*kernel.shared_1[((threadIdx.x*48) + 21)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[57]*kernel.shared_1[((threadIdx.x*48) + 18)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[66]*kernel.shared_1[((threadIdx.x*48) + 21)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[58]*kernel.shared_1[((threadIdx.x*48) + 18)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[67]*kernel.shared_1[((threadIdx.x*48) + 21)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[59]*kernel.shared_1[((threadIdx.x*48) + 18)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[68]*kernel.shared_1[((threadIdx.x*48) + 21)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[60]*kernel.shared_1[((threadIdx.x*48) + 18)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[69]*kernel.shared_1[((threadIdx.x*48) + 21)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[54]*kernel.shared_1[((threadIdx.x*48) + 42)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[63]*kernel.shared_1[((threadIdx.x*48) + 45)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[55]*kernel.shared_1[((threadIdx.x*48) + 42)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[64]*kernel.shared_1[((threadIdx.x*48) + 45)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[56]*kernel.shared_1[((threadIdx.x*48) + 42)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[65]*kernel.shared_1[((threadIdx.x*48) + 45)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[57]*kernel.shared_1[((threadIdx.x*48) + 42)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[66]*kernel.shared_1[((threadIdx.x*48) + 45)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[58]*kernel.shared_1[((threadIdx.x*48) + 42)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[67]*kernel.shared_1[((threadIdx.x*48) + 45)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[59]*kernel.shared_1[((threadIdx.x*48) + 42)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[68]*kernel.shared_1[((threadIdx.x*48) + 45)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[60]*kernel.shared_1[((threadIdx.x*48) + 42)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[69]*kernel.shared_1[((threadIdx.x*48) + 45)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[55]*kernel.shared_1[((threadIdx.x*48) + 19)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[64]*kernel.shared_1[((threadIdx.x*48) + 22)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[56]*kernel.shared_1[((threadIdx.x*48) + 19)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[65]*kernel.shared_1[((threadIdx.x*48) + 22)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[57]*kernel.shared_1[((threadIdx.x*48) + 19)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[66]*kernel.shared_1[((threadIdx.x*48) + 22)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[58]*kernel.shared_1[((threadIdx.x*48) + 19)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[67]*kernel.shared_1[((threadIdx.x*48) + 22)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[59]*kernel.shared_1[((threadIdx.x*48) + 19)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[68]*kernel.shared_1[((threadIdx.x*48) + 22)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[60]*kernel.shared_1[((threadIdx.x*48) + 19)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[69]*kernel.shared_1[((threadIdx.x*48) + 22)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[61]*kernel.shared_1[((threadIdx.x*48) + 19)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[70]*kernel.shared_1[((threadIdx.x*48) + 22)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[55]*kernel.shared_1[((threadIdx.x*48) + 43)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[64]*kernel.shared_1[((threadIdx.x*48) + 46)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[56]*kernel.shared_1[((threadIdx.x*48) + 43)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[65]*kernel.shared_1[((threadIdx.x*48) + 46)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[57]*kernel.shared_1[((threadIdx.x*48) + 43)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[66]*kernel.shared_1[((threadIdx.x*48) + 46)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[58]*kernel.shared_1[((threadIdx.x*48) + 43)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[67]*kernel.shared_1[((threadIdx.x*48) + 46)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[59]*kernel.shared_1[((threadIdx.x*48) + 43)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[68]*kernel.shared_1[((threadIdx.x*48) + 46)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[60]*kernel.shared_1[((threadIdx.x*48) + 43)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[69]*kernel.shared_1[((threadIdx.x*48) + 46)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[61]*kernel.shared_1[((threadIdx.x*48) + 43)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[70]*kernel.shared_1[((threadIdx.x*48) + 46)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[56]*kernel.shared_1[((threadIdx.x*48) + 20)]))
+ conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[65]*kernel.shared_1[((threadIdx.x*48) + 23)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[57]*kernel.shared_1[((threadIdx.x*48) + 20)]))
+ conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[66]*kernel.shared_1[((threadIdx.x*48) + 23)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[58]*kernel.shared_1[((threadIdx.x*48) + 20)]))
+ conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[67]*kernel.shared_1[((threadIdx.x*48) + 23)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[59]*kernel.shared_1[((threadIdx.x*48) + 20)]))
+ conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[68]*kernel.shared_1[((threadIdx.x*48) + 23)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[60]*kernel.shared_1[((threadIdx.x*48) + 20)]))
+ conv2d_nchw_1[4] = (conv2d_nchw_1[4] + (pad_temp.shared_1[69]*kernel.shared_1[((threadIdx.x*48) + 23)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[61]*kernel.shared_1[((threadIdx.x*48) + 20)]))
+ conv2d_nchw_1[5] = (conv2d_nchw_1[5] + (pad_temp.shared_1[70]*kernel.shared_1[((threadIdx.x*48) + 23)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[62]*kernel.shared_1[((threadIdx.x*48) + 20)]))
+ conv2d_nchw_1[6] = (conv2d_nchw_1[6] + (pad_temp.shared_1[71]*kernel.shared_1[((threadIdx.x*48) + 23)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[56]*kernel.shared_1[((threadIdx.x*48) + 44)]))
+ conv2d_nchw_1[7] = (conv2d_nchw_1[7] + (pad_temp.shared_1[65]*kernel.shared_1[((threadIdx.x*48) + 47)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[57]*kernel.shared_1[((threadIdx.x*48) + 44)]))
+ conv2d_nchw_1[8] = (conv2d_nchw_1[8] + (pad_temp.shared_1[66]*kernel.shared_1[((threadIdx.x*48) + 47)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[58]*kernel.shared_1[((threadIdx.x*48) + 44)]))
+ conv2d_nchw_1[9] = (conv2d_nchw_1[9] + (pad_temp.shared_1[67]*kernel.shared_1[((threadIdx.x*48) + 47)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[59]*kernel.shared_1[((threadIdx.x*48) + 44)]))
+ conv2d_nchw_1[10] = (conv2d_nchw_1[10] + (pad_temp.shared_1[68]*kernel.shared_1[((threadIdx.x*48) + 47)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[60]*kernel.shared_1[((threadIdx.x*48) + 44)]))
+ conv2d_nchw_1[11] = (conv2d_nchw_1[11] + (pad_temp.shared_1[69]*kernel.shared_1[((threadIdx.x*48) + 47)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[61]*kernel.shared_1[((threadIdx.x*48) + 44)]))
+ conv2d_nchw_1[12] = (conv2d_nchw_1[12] + (pad_temp.shared_1[70]*kernel.shared_1[((threadIdx.x*48) + 47)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[62]*kernel.shared_1[((threadIdx.x*48) + 44)]))
+ conv2d_nchw_1[13] = (conv2d_nchw_1[13] + (pad_temp.shared_1[71]*kernel.shared_1[((threadIdx.x*48) + 47)]))
+ }
}
- attr [IterVar(threadIdx.x_2: int32, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 98;
- kernel.shared_1: Buffer(kernel.shared, float32, [144], [], scope="shared")[threadIdx.x_2] = kernel[((((blockIdx.x*36864) + (floordiv(threadIdx.x_2, 18)*4608)) + (rc.outer.outer*18)) + floormod(threadIdx.x_2, 18))]
- attr [IterVar(threadIdx.x_2, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 98;
- if @tir.likely((threadIdx.x_2 < 46), dtype=bool) {
- kernel.shared_1[(threadIdx.x_2 + 98)] = kernel[(((((blockIdx.x*36864) + (floordiv((threadIdx.x_2 + 98), 18)*4608)) + (rc.outer.outer*18)) + (floordiv(floormod((threadIdx.x_2 + 8), 18), 3)*3)) + floormod((threadIdx.x_2 + 2), 3))]
- }
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7))]*kernel.shared_1[(floordiv(threadIdx.x, 49)*72)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7))]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 18)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 1)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 1)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 1)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 19)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 2)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 2)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 2)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 20)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 9)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 3)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 9)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 21)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 10)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 4)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 10)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 22)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 11)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 5)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 11)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 23)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 18)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 6)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 18)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 24)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 19)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 7)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 19)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 25)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 20)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 8)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 20)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 26)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7))]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 36)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7))]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 54)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 1)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 37)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 1)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 55)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 2)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 38)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 2)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 56)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 9)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 39)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 9)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 57)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 10)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 40)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 10)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 58)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 11)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 41)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 11)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 59)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 18)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 42)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 18)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 60)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 19)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 43)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 19)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 61)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 20)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 44)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 20)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 62)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 81)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 9)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 81)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 27)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 82)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 10)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 82)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 28)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 83)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 11)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 83)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 29)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 90)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 12)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 90)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 30)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 91)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 13)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 91)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 31)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 92)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 14)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 92)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 32)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 99)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 15)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 99)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 33)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 100)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 16)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 100)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 34)]))
- conv2d_nchw_1[0] = (conv2d_nchw_1[0] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 101)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 17)]))
- conv2d_nchw_1[1] = (conv2d_nchw_1[1] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 101)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 35)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 81)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 45)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 81)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 63)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 82)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 46)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 82)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 64)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 83)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 47)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 83)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 65)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 90)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 48)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 90)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 66)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 91)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 49)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 91)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 67)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 92)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 50)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 92)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 68)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 99)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 51)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 99)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 69)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 100)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 52)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 100)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 70)]))
- conv2d_nchw_1[2] = (conv2d_nchw_1[2] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 101)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 53)]))
- conv2d_nchw_1[3] = (conv2d_nchw_1[3] + (pad_temp.shared_1[(((floordiv(floormod(threadIdx.x, 49), 7)*9) + floormod(threadIdx.x, 7)) + 101)]*kernel.shared_1[((floordiv(threadIdx.x, 49)*72) + 71)]))
}
- for (i1.inner: int32, 0, 4) {
- compute[((((blockIdx.x*392) + (floordiv(threadIdx.x, 49)*196)) + (i1.inner*49)) + floormod(threadIdx.x, 49))] = max((conv2d_nchw_1[i1.inner] + bias[(((blockIdx.x*8) + (floordiv(threadIdx.x, 49)*4)) + i1.inner)]), 0f32)
+ for (i1.inner: int32, 0, 2) {
+ for (i3.inner: int32, 0, 7) {
+ compute_3: Buffer(compute_2, float32, [25088], [])[(((((floordiv(blockIdx.x, 7)*6272) + (threadIdx.x*98)) + (i1.inner*49)) + (floormod(blockIdx.x, 7)*7)) + i3.inner)] = max((conv2d_nchw_1[((i1.inner*7) + i3.inner)] + bias_3: Buffer(bias_2, float32, [512], [])[(((floordiv(blockIdx.x, 7)*128) + (threadIdx.x*2)) + i1.inner)]), 0f32)
+ }
}
}
}
@@ -637,7 +1016,7 @@ cooperative fetching, unrolling and operator fusion.</p>
<span class="p">)</span>
</pre></div>
</div>
-<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Execution time of this operator: 0.353 ms
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Execution time of this operator: 0.349 ms
</pre></div>
</div>
</div>
@@ -666,36 +1045,36 @@ conv2d_nchw_nn_o_i, conv2d_nchw_nn_i = s[conv2d_nchw].split(conv2d_nchw_nn, fact
conv2d_nchw_nn_o_o_i, conv2d_nchw_nn_o_i = s[conv2d_nchw].split(conv2d_nchw_nn_o_i, factor=1)
conv2d_nchw_nn_o_o_o_i, conv2d_nchw_nn_o_o_i = s[conv2d_nchw].split(conv2d_nchw_nn_o_o_i, factor=1)
conv2d_nchw_nn_o_o_o_o, conv2d_nchw_nn_o_o_o_i = s[conv2d_nchw].split(conv2d_nchw_nn_o_o_o_i, factor=1)
-conv2d_nchw_ff_o_i, conv2d_nchw_ff_i = s[conv2d_nchw].split(conv2d_nchw_ff, factor=2)
+conv2d_nchw_ff_o_i, conv2d_nchw_ff_i = s[conv2d_nchw].split(conv2d_nchw_ff, factor=1)
conv2d_nchw_ff_o_o_i, conv2d_nchw_ff_o_i = s[conv2d_nchw].split(conv2d_nchw_ff_o_i, factor=2)
-conv2d_nchw_ff_o_o_o_i, conv2d_nchw_ff_o_o_i = s[conv2d_nchw].split(conv2d_nchw_ff_o_o_i, factor=2)
+conv2d_nchw_ff_o_o_o_i, conv2d_nchw_ff_o_o_i = s[conv2d_nchw].split(conv2d_nchw_ff_o_o_i, factor=64)
conv2d_nchw_ff_o_o_o_o, conv2d_nchw_ff_o_o_o_i = s[conv2d_nchw].split(conv2d_nchw_ff_o_o_o_i, factor=1)
conv2d_nchw_yy_o_i, conv2d_nchw_yy_i = s[conv2d_nchw].split(conv2d_nchw_yy, factor=1)
conv2d_nchw_yy_o_o_i, conv2d_nchw_yy_o_i = s[conv2d_nchw].split(conv2d_nchw_yy_o_i, factor=1)
-conv2d_nchw_yy_o_o_o_i, conv2d_nchw_yy_o_o_i = s[conv2d_nchw].split(conv2d_nchw_yy_o_o_i, factor=7)
+conv2d_nchw_yy_o_o_o_i, conv2d_nchw_yy_o_o_i = s[conv2d_nchw].split(conv2d_nchw_yy_o_o_i, factor=1)
conv2d_nchw_yy_o_o_o_o, conv2d_nchw_yy_o_o_o_i = s[conv2d_nchw].split(conv2d_nchw_yy_o_o_o_i, factor=1)
conv2d_nchw_xx_o_i, conv2d_nchw_xx_i = s[conv2d_nchw].split(conv2d_nchw_xx, factor=1)
-conv2d_nchw_xx_o_o_i, conv2d_nchw_xx_o_i = s[conv2d_nchw].split(conv2d_nchw_xx_o_i, factor=1)
-conv2d_nchw_xx_o_o_o_i, conv2d_nchw_xx_o_o_i = s[conv2d_nchw].split(conv2d_nchw_xx_o_o_i, factor=7)
+conv2d_nchw_xx_o_o_i, conv2d_nchw_xx_o_i = s[conv2d_nchw].split(conv2d_nchw_xx_o_i, factor=7)
+conv2d_nchw_xx_o_o_o_i, conv2d_nchw_xx_o_o_i = s[conv2d_nchw].split(conv2d_nchw_xx_o_o_i, factor=1)
conv2d_nchw_xx_o_o_o_o, conv2d_nchw_xx_o_o_o_i = s[conv2d_nchw].split(conv2d_nchw_xx_o_o_o_i, factor=1)
-conv2d_nchw_rc_o_i, conv2d_nchw_rc_i = s[conv2d_nchw].split(conv2d_nchw_rc, factor=1)
-conv2d_nchw_rc_o_o, conv2d_nchw_rc_o_i = s[conv2d_nchw].split(conv2d_nchw_rc_o_i, factor=2)
-conv2d_nchw_ry_o_i, conv2d_nchw_ry_i = s[conv2d_nchw].split(conv2d_nchw_ry, factor=3)
+conv2d_nchw_rc_o_i, conv2d_nchw_rc_i = s[conv2d_nchw].split(conv2d_nchw_rc, factor=2)
+conv2d_nchw_rc_o_o, conv2d_nchw_rc_o_i = s[conv2d_nchw].split(conv2d_nchw_rc_o_i, factor=4)
+conv2d_nchw_ry_o_i, conv2d_nchw_ry_i = s[conv2d_nchw].split(conv2d_nchw_ry, factor=1)
conv2d_nchw_ry_o_o, conv2d_nchw_ry_o_i = s[conv2d_nchw].split(conv2d_nchw_ry_o_i, factor=1)
-conv2d_nchw_rx_o_i, conv2d_nchw_rx_i = s[conv2d_nchw].split(conv2d_nchw_rx, factor=3)
-conv2d_nchw_rx_o_o, conv2d_nchw_rx_o_i = s[conv2d_nchw].split(conv2d_nchw_rx_o_i, factor=1)
+conv2d_nchw_rx_o_i, conv2d_nchw_rx_i = s[conv2d_nchw].split(conv2d_nchw_rx, factor=1)
+conv2d_nchw_rx_o_o, conv2d_nchw_rx_o_i = s[conv2d_nchw].split(conv2d_nchw_rx_o_i, factor=3)
s[conv2d_nchw].reorder(conv2d_nchw_nn_o_o_o_o, conv2d_nchw_ff_o_o_o_o, conv2d_nchw_yy_o_o_o_o, conv2d_nchw_xx_o_o_o_o, conv2d_nchw_nn_o_o_o_i, conv2d_nchw_ff_o_o_o_i, conv2d_nchw_yy_o_o_o_i, conv2d_nchw_xx_o_o_o_i, conv2d_nchw_nn_o_o_i, conv2d_nchw_ff_o_o_i, conv2d_nchw_yy_o_o_i, conv2d_nchw_xx_o_o_i, conv2d_nchw_rc_o_o, conv2d_nchw_ry_o_o, conv2d_nchw_rx_o_o, conv2d_nchw_rc_o_i, conv2d_nchw_ry_o_i, conv2d_nchw_rx_o_i, conv2d_nchw_nn_o_i, conv2d_nchw_ff_o_i, conv2d_nchw_yy_o_i, conv2d_nc [...]
compute_i0_o_i, compute_i0_i = s[compute].split(compute_i0, factor=1)
compute_i0_o_o_i, compute_i0_o_i = s[compute].split(compute_i0_o_i, factor=1)
compute_i0_o_o_o, compute_i0_o_o_i = s[compute].split(compute_i0_o_o_i, factor=1)
-compute_i1_o_i, compute_i1_i = s[compute].split(compute_i1, factor=4)
-compute_i1_o_o_i, compute_i1_o_i = s[compute].split(compute_i1_o_i, factor=2)
+compute_i1_o_i, compute_i1_i = s[compute].split(compute_i1, factor=2)
+compute_i1_o_o_i, compute_i1_o_i = s[compute].split(compute_i1_o_i, factor=64)
compute_i1_o_o_o, compute_i1_o_o_i = s[compute].split(compute_i1_o_o_i, factor=1)
compute_i2_o_i, compute_i2_i = s[compute].split(compute_i2, factor=1)
-compute_i2_o_o_i, compute_i2_o_i = s[compute].split(compute_i2_o_i, factor=7)
+compute_i2_o_o_i, compute_i2_o_i = s[compute].split(compute_i2_o_i, factor=1)
compute_i2_o_o_o, compute_i2_o_o_i = s[compute].split(compute_i2_o_o_i, factor=1)
-compute_i3_o_i, compute_i3_i = s[compute].split(compute_i3, factor=1)
-compute_i3_o_o_i, compute_i3_o_i = s[compute].split(compute_i3_o_i, factor=7)
+compute_i3_o_i, compute_i3_i = s[compute].split(compute_i3, factor=7)
+compute_i3_o_o_i, compute_i3_o_i = s[compute].split(compute_i3_o_i, factor=1)
compute_i3_o_o_o, compute_i3_o_o_i = s[compute].split(compute_i3_o_o_i, factor=1)
s[compute].reorder(compute_i0_o_o_o, compute_i1_o_o_o, compute_i2_o_o_o, compute_i3_o_o_o, compute_i0_o_o_i, compute_i1_o_o_i, compute_i2_o_o_i, compute_i3_o_o_i, compute_i0_o_i, compute_i1_o_i, compute_i2_o_i, compute_i3_o_i, compute_i0_i, compute_i1_i, compute_i2_i, compute_i3_i)
s[conv2d_nchw].compute_at(s[compute], compute_i3_o_i)
@@ -715,12 +1094,12 @@ s[compute].bind(compute_i0_o_i_i1_o_i_fused_i2_o_i_fused_i3_o_i_fused, te.thread
kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused = s[kernel_shared].fuse(kernel_shared_ax0, kernel_shared_ax1, kernel_shared_ax2, kernel_shared_ax3)
kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o, kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_i = s[kernel_shared].split(kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused, factor=1)
s[kernel_shared].vectorize(kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_i)
-kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_o, kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_i = s[kernel_shared].split(kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o, factor=98)
+kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_o, kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_i = s[kernel_shared].split(kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o, factor=64)
s[kernel_shared].bind(kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_i, te.thread_axis("threadIdx.x"))
pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused = s[pad_temp_shared].fuse(pad_temp_shared_ax0, pad_temp_shared_ax1, pad_temp_shared_ax2, pad_temp_shared_ax3)
-pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o, pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_i = s[pad_temp_shared].split(pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused, factor=1)
+pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o, pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_i = s[pad_temp_shared].split(pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused, factor=4)
s[pad_temp_shared].vectorize(pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_i)
-pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_o, pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_i = s[pad_temp_shared].split(pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o, factor=98)
+pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_o, pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_i = s[pad_temp_shared].split(pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o, factor=64)
s[pad_temp_shared].bind(pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_i, te.thread_axis("threadIdx.x"))
s[conv2d_nchw].pragma(conv2d_nchw_nn_o_o_o_o, "auto_unroll_max_step", 512)
s[conv2d_nchw].pragma(conv2d_nchw_nn_o_o_o_o, "unroll_explicit", True)
@@ -740,100 +1119,430 @@ CUDA source code:
#define int64_t long long
#define uint64_t unsigned long long
#endif
-extern "C" __global__ void __launch_bounds__(98) default_function_kernel0(float* __restrict__ data, float* __restrict__ kernel, float* __restrict__ compute, float* __restrict__ bias) {
- float conv2d_nchw[4];
- __shared__ float pad_temp_shared[162];
- __shared__ float kernel_shared[144];
+extern "C" __global__ void __launch_bounds__(64) default_function_kernel0(float* __restrict__ data, float* __restrict__ kernel, float* __restrict__ compute, float* __restrict__ bias) {
+ float conv2d_nchw[14];
+ __shared__ float pad_temp_shared[72];
+ __shared__ float kernel_shared[3072];
conv2d_nchw[0] = 0.000000e+00f;
conv2d_nchw[1] = 0.000000e+00f;
conv2d_nchw[2] = 0.000000e+00f;
conv2d_nchw[3] = 0.000000e+00f;
- for (int rc_outer_outer = 0; rc_outer_outer < 256; ++rc_outer_outer) {
- __syncthreads();
- pad_temp_shared[((int)threadIdx.x)] = (((((9 <= (((int)threadIdx.x) % 81)) && ((((int)threadIdx.x) % 81) < 72)) && (1 <= (((int)threadIdx.x) % 9))) && ((((int)threadIdx.x) % 9) < 8)) ? data[(((((rc_outer_outer * 98) + ((((int)threadIdx.x) / 81) * 49)) + (((((int)threadIdx.x) % 81) / 9) * 7)) + (((int)threadIdx.x) % 9)) - 8)] : 0.000000e+00f);
- if (((int)threadIdx.x) < 64) {
- pad_temp_shared[(((int)threadIdx.x) + 98)] = ((((((int)threadIdx.x) < 55) && (1 <= ((((int)threadIdx.x) + 8) % 9))) && (((((int)threadIdx.x) + 8) % 9) < 8)) ? data[(((((rc_outer_outer * 98) + (((((int)threadIdx.x) + 98) / 81) * 49)) + (((((int)threadIdx.x) + 17) / 9) * 7)) + ((((int)threadIdx.x) + 8) % 9)) - 8)] : 0.000000e+00f);
- }
- kernel_shared[((int)threadIdx.x)] = kernel[((((((int)blockIdx.x) * 36864) + ((((int)threadIdx.x) / 18) * 4608)) + (rc_outer_outer * 18)) + (((int)threadIdx.x) % 18))];
- if (((int)threadIdx.x) < 46) {
- kernel_shared[(((int)threadIdx.x) + 98)] = kernel[(((((((int)blockIdx.x) * 36864) + (((((int)threadIdx.x) + 98) / 18) * 4608)) + (rc_outer_outer * 18)) + ((((((int)threadIdx.x) + 8) % 18) / 3) * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ conv2d_nchw[4] = 0.000000e+00f;
+ conv2d_nchw[5] = 0.000000e+00f;
+ conv2d_nchw[6] = 0.000000e+00f;
+ conv2d_nchw[7] = 0.000000e+00f;
+ conv2d_nchw[8] = 0.000000e+00f;
+ conv2d_nchw[9] = 0.000000e+00f;
+ conv2d_nchw[10] = 0.000000e+00f;
+ conv2d_nchw[11] = 0.000000e+00f;
+ conv2d_nchw[12] = 0.000000e+00f;
+ conv2d_nchw[13] = 0.000000e+00f;
+ for (int rc_outer_outer = 0; rc_outer_outer < 64; ++rc_outer_outer) {
+ for (int ry_outer_outer = 0; ry_outer_outer < 3; ++ry_outer_outer) {
+ __syncthreads();
+ if (((int)threadIdx.x) < 18) {
+ pad_temp_shared[(((int)threadIdx.x) * 4)] = (((((1 <= (ry_outer_outer + (((int)blockIdx.x) % 7))) && ((ry_outer_outer + (((int)blockIdx.x) % 7)) < 8)) && (1 <= ((((int)threadIdx.x) * 4) % 9))) && (((((int)threadIdx.x) * 4) % 9) < 8)) ? data[((((((rc_outer_outer * 392) + (((((int)threadIdx.x) * 4) / 9) * 49)) + (ry_outer_outer * 7)) + ((((int)blockIdx.x) % 7) * 7)) + ((((int)threadIdx.x) * 4) % 9)) - 8)] : 0.000000e+00f);
+ }
+ if (((int)threadIdx.x) < 18) {
+ pad_temp_shared[((((int)threadIdx.x) * 4) + 1)] = (((((1 <= (ry_outer_outer + (((int)blockIdx.x) % 7))) && ((ry_outer_outer + (((int)blockIdx.x) % 7)) < 8)) && (1 <= (((((int)threadIdx.x) * 4) + 1) % 9))) && ((((((int)threadIdx.x) * 4) + 1) % 9) < 8)) ? data[((((((rc_outer_outer * 392) + ((((((int)threadIdx.x) * 4) + 1) / 9) * 49)) + (ry_outer_outer * 7)) + ((((int)blockIdx.x) % 7) * 7)) + (((((int)threadIdx.x) * 4) + 1) % 9)) - 8)] : 0.000000e+00f);
+ }
+ if (((int)threadIdx.x) < 18) {
+ pad_temp_shared[((((int)threadIdx.x) * 4) + 2)] = (((((1 <= (ry_outer_outer + (((int)blockIdx.x) % 7))) && ((ry_outer_outer + (((int)blockIdx.x) % 7)) < 8)) && (1 <= (((((int)threadIdx.x) * 4) + 2) % 9))) && ((((((int)threadIdx.x) * 4) + 2) % 9) < 8)) ? data[((((((rc_outer_outer * 392) + ((((((int)threadIdx.x) * 4) + 2) / 9) * 49)) + (ry_outer_outer * 7)) + ((((int)blockIdx.x) % 7) * 7)) + (((((int)threadIdx.x) * 4) + 2) % 9)) - 8)] : 0.000000e+00f);
+ }
+ if (((int)threadIdx.x) < 18) {
+ pad_temp_shared[((((int)threadIdx.x) * 4) + 3)] = (((((1 <= (ry_outer_outer + (((int)blockIdx.x) % 7))) && ((ry_outer_outer + (((int)blockIdx.x) % 7)) < 8)) && (1 <= (((((int)threadIdx.x) * 4) + 3) % 9))) && ((((((int)threadIdx.x) * 4) + 3) % 9) < 8)) ? data[((((((rc_outer_outer * 392) + ((((((int)threadIdx.x) * 4) + 3) / 9) * 49)) + (ry_outer_outer * 7)) + ((((int)blockIdx.x) % 7) * 7)) + (((((int)threadIdx.x) * 4) + 3) % 9)) - 8)] : 0.000000e+00f);
+ }
+ kernel_shared[((int)threadIdx.x)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 64)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 64) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 128)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 128) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 192)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 36864)];
+ kernel_shared[(((int)threadIdx.x) + 256)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 256) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 320)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 320) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 384)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 73728)];
+ kernel_shared[(((int)threadIdx.x) + 448)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 448) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 512)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 512) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 576)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 110592)];
+ kernel_shared[(((int)threadIdx.x) + 640)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 640) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 704)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 704) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 768)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 147456)];
+ kernel_shared[(((int)threadIdx.x) + 832)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 832) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 896)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 896) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 960)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 184320)];
+ kernel_shared[(((int)threadIdx.x) + 1024)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1024) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 1088)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1088) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 1152)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 221184)];
+ kernel_shared[(((int)threadIdx.x) + 1216)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1216) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 1280)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1280) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 1344)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 258048)];
+ kernel_shared[(((int)threadIdx.x) + 1408)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1408) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 1472)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1472) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 1536)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 294912)];
+ kernel_shared[(((int)threadIdx.x) + 1600)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1600) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 1664)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1664) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 1728)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 331776)];
+ kernel_shared[(((int)threadIdx.x) + 1792)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1792) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 1856)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1856) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 1920)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 368640)];
+ kernel_shared[(((int)threadIdx.x) + 1984)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1984) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 2048)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2048) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 2112)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 405504)];
+ kernel_shared[(((int)threadIdx.x) + 2176)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2176) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 2240)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2240) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 2304)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 442368)];
+ kernel_shared[(((int)threadIdx.x) + 2368)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2368) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 2432)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2432) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 2496)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 479232)];
+ kernel_shared[(((int)threadIdx.x) + 2560)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2560) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 2624)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2624) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 2688)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 516096)];
+ kernel_shared[(((int)threadIdx.x) + 2752)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2752) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 2816)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2816) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 2880)] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 552960)];
+ kernel_shared[(((int)threadIdx.x) + 2944)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2944) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3))];
+ kernel_shared[(((int)threadIdx.x) + 3008)] = kernel[(((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 3008) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3))];
+ __syncthreads();
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[0] * kernel_shared[(((int)threadIdx.x) * 48)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[9] * kernel_shared[((((int)threadIdx.x) * 48) + 3)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[1] * kernel_shared[(((int)threadIdx.x) * 48)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[10] * kernel_shared[((((int)threadIdx.x) * 48) + 3)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[2] * kernel_shared[(((int)threadIdx.x) * 48)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[11] * kernel_shared[((((int)threadIdx.x) * 48) + 3)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[3] * kernel_shared[(((int)threadIdx.x) * 48)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[12] * kernel_shared[((((int)threadIdx.x) * 48) + 3)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[4] * kernel_shared[(((int)threadIdx.x) * 48)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[13] * kernel_shared[((((int)threadIdx.x) * 48) + 3)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[5] * kernel_shared[(((int)threadIdx.x) * 48)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[14] * kernel_shared[((((int)threadIdx.x) * 48) + 3)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[6] * kernel_shared[(((int)threadIdx.x) * 48)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[15] * kernel_shared[((((int)threadIdx.x) * 48) + 3)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[0] * kernel_shared[((((int)threadIdx.x) * 48) + 24)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[9] * kernel_shared[((((int)threadIdx.x) * 48) + 27)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[1] * kernel_shared[((((int)threadIdx.x) * 48) + 24)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[10] * kernel_shared[((((int)threadIdx.x) * 48) + 27)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[2] * kernel_shared[((((int)threadIdx.x) * 48) + 24)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[11] * kernel_shared[((((int)threadIdx.x) * 48) + 27)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[3] * kernel_shared[((((int)threadIdx.x) * 48) + 24)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[12] * kernel_shared[((((int)threadIdx.x) * 48) + 27)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[4] * kernel_shared[((((int)threadIdx.x) * 48) + 24)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[13] * kernel_shared[((((int)threadIdx.x) * 48) + 27)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[5] * kernel_shared[((((int)threadIdx.x) * 48) + 24)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[14] * kernel_shared[((((int)threadIdx.x) * 48) + 27)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[6] * kernel_shared[((((int)threadIdx.x) * 48) + 24)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[15] * kernel_shared[((((int)threadIdx.x) * 48) + 27)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[1] * kernel_shared[((((int)threadIdx.x) * 48) + 1)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[10] * kernel_shared[((((int)threadIdx.x) * 48) + 4)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[2] * kernel_shared[((((int)threadIdx.x) * 48) + 1)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[11] * kernel_shared[((((int)threadIdx.x) * 48) + 4)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[3] * kernel_shared[((((int)threadIdx.x) * 48) + 1)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[12] * kernel_shared[((((int)threadIdx.x) * 48) + 4)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[4] * kernel_shared[((((int)threadIdx.x) * 48) + 1)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[13] * kernel_shared[((((int)threadIdx.x) * 48) + 4)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[5] * kernel_shared[((((int)threadIdx.x) * 48) + 1)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[14] * kernel_shared[((((int)threadIdx.x) * 48) + 4)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[6] * kernel_shared[((((int)threadIdx.x) * 48) + 1)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[15] * kernel_shared[((((int)threadIdx.x) * 48) + 4)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[7] * kernel_shared[((((int)threadIdx.x) * 48) + 1)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[16] * kernel_shared[((((int)threadIdx.x) * 48) + 4)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[1] * kernel_shared[((((int)threadIdx.x) * 48) + 25)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[10] * kernel_shared[((((int)threadIdx.x) * 48) + 28)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[2] * kernel_shared[((((int)threadIdx.x) * 48) + 25)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[11] * kernel_shared[((((int)threadIdx.x) * 48) + 28)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[3] * kernel_shared[((((int)threadIdx.x) * 48) + 25)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[12] * kernel_shared[((((int)threadIdx.x) * 48) + 28)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[4] * kernel_shared[((((int)threadIdx.x) * 48) + 25)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[13] * kernel_shared[((((int)threadIdx.x) * 48) + 28)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[5] * kernel_shared[((((int)threadIdx.x) * 48) + 25)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[14] * kernel_shared[((((int)threadIdx.x) * 48) + 28)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[6] * kernel_shared[((((int)threadIdx.x) * 48) + 25)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[15] * kernel_shared[((((int)threadIdx.x) * 48) + 28)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[7] * kernel_shared[((((int)threadIdx.x) * 48) + 25)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[16] * kernel_shared[((((int)threadIdx.x) * 48) + 28)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[2] * kernel_shared[((((int)threadIdx.x) * 48) + 2)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[11] * kernel_shared[((((int)threadIdx.x) * 48) + 5)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[3] * kernel_shared[((((int)threadIdx.x) * 48) + 2)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[12] * kernel_shared[((((int)threadIdx.x) * 48) + 5)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[4] * kernel_shared[((((int)threadIdx.x) * 48) + 2)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[13] * kernel_shared[((((int)threadIdx.x) * 48) + 5)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[5] * kernel_shared[((((int)threadIdx.x) * 48) + 2)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[14] * kernel_shared[((((int)threadIdx.x) * 48) + 5)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[6] * kernel_shared[((((int)threadIdx.x) * 48) + 2)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[15] * kernel_shared[((((int)threadIdx.x) * 48) + 5)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[7] * kernel_shared[((((int)threadIdx.x) * 48) + 2)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[16] * kernel_shared[((((int)threadIdx.x) * 48) + 5)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[8] * kernel_shared[((((int)threadIdx.x) * 48) + 2)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[17] * kernel_shared[((((int)threadIdx.x) * 48) + 5)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[2] * kernel_shared[((((int)threadIdx.x) * 48) + 26)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[11] * kernel_shared[((((int)threadIdx.x) * 48) + 29)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[3] * kernel_shared[((((int)threadIdx.x) * 48) + 26)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[12] * kernel_shared[((((int)threadIdx.x) * 48) + 29)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[4] * kernel_shared[((((int)threadIdx.x) * 48) + 26)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[13] * kernel_shared[((((int)threadIdx.x) * 48) + 29)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[5] * kernel_shared[((((int)threadIdx.x) * 48) + 26)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[14] * kernel_shared[((((int)threadIdx.x) * 48) + 29)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[6] * kernel_shared[((((int)threadIdx.x) * 48) + 26)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[15] * kernel_shared[((((int)threadIdx.x) * 48) + 29)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[7] * kernel_shared[((((int)threadIdx.x) * 48) + 26)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[16] * kernel_shared[((((int)threadIdx.x) * 48) + 29)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[8] * kernel_shared[((((int)threadIdx.x) * 48) + 26)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[17] * kernel_shared[((((int)threadIdx.x) * 48) + 29)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[18] * kernel_shared[((((int)threadIdx.x) * 48) + 6)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[27] * kernel_shared[((((int)threadIdx.x) * 48) + 9)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[19] * kernel_shared[((((int)threadIdx.x) * 48) + 6)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[28] * kernel_shared[((((int)threadIdx.x) * 48) + 9)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[20] * kernel_shared[((((int)threadIdx.x) * 48) + 6)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[29] * kernel_shared[((((int)threadIdx.x) * 48) + 9)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[21] * kernel_shared[((((int)threadIdx.x) * 48) + 6)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[30] * kernel_shared[((((int)threadIdx.x) * 48) + 9)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[22] * kernel_shared[((((int)threadIdx.x) * 48) + 6)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[31] * kernel_shared[((((int)threadIdx.x) * 48) + 9)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[23] * kernel_shared[((((int)threadIdx.x) * 48) + 6)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[32] * kernel_shared[((((int)threadIdx.x) * 48) + 9)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[24] * kernel_shared[((((int)threadIdx.x) * 48) + 6)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[33] * kernel_shared[((((int)threadIdx.x) * 48) + 9)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[18] * kernel_shared[((((int)threadIdx.x) * 48) + 30)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[27] * kernel_shared[((((int)threadIdx.x) * 48) + 33)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[19] * kernel_shared[((((int)threadIdx.x) * 48) + 30)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[28] * kernel_shared[((((int)threadIdx.x) * 48) + 33)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[20] * kernel_shared[((((int)threadIdx.x) * 48) + 30)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[29] * kernel_shared[((((int)threadIdx.x) * 48) + 33)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[21] * kernel_shared[((((int)threadIdx.x) * 48) + 30)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[30] * kernel_shared[((((int)threadIdx.x) * 48) + 33)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[22] * kernel_shared[((((int)threadIdx.x) * 48) + 30)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[31] * kernel_shared[((((int)threadIdx.x) * 48) + 33)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[23] * kernel_shared[((((int)threadIdx.x) * 48) + 30)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[32] * kernel_shared[((((int)threadIdx.x) * 48) + 33)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[24] * kernel_shared[((((int)threadIdx.x) * 48) + 30)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[33] * kernel_shared[((((int)threadIdx.x) * 48) + 33)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[19] * kernel_shared[((((int)threadIdx.x) * 48) + 7)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[28] * kernel_shared[((((int)threadIdx.x) * 48) + 10)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[20] * kernel_shared[((((int)threadIdx.x) * 48) + 7)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[29] * kernel_shared[((((int)threadIdx.x) * 48) + 10)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[21] * kernel_shared[((((int)threadIdx.x) * 48) + 7)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[30] * kernel_shared[((((int)threadIdx.x) * 48) + 10)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[22] * kernel_shared[((((int)threadIdx.x) * 48) + 7)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[31] * kernel_shared[((((int)threadIdx.x) * 48) + 10)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[23] * kernel_shared[((((int)threadIdx.x) * 48) + 7)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[32] * kernel_shared[((((int)threadIdx.x) * 48) + 10)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[24] * kernel_shared[((((int)threadIdx.x) * 48) + 7)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[33] * kernel_shared[((((int)threadIdx.x) * 48) + 10)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[25] * kernel_shared[((((int)threadIdx.x) * 48) + 7)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[34] * kernel_shared[((((int)threadIdx.x) * 48) + 10)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[19] * kernel_shared[((((int)threadIdx.x) * 48) + 31)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[28] * kernel_shared[((((int)threadIdx.x) * 48) + 34)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[20] * kernel_shared[((((int)threadIdx.x) * 48) + 31)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[29] * kernel_shared[((((int)threadIdx.x) * 48) + 34)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[21] * kernel_shared[((((int)threadIdx.x) * 48) + 31)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[30] * kernel_shared[((((int)threadIdx.x) * 48) + 34)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[22] * kernel_shared[((((int)threadIdx.x) * 48) + 31)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[31] * kernel_shared[((((int)threadIdx.x) * 48) + 34)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[23] * kernel_shared[((((int)threadIdx.x) * 48) + 31)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[32] * kernel_shared[((((int)threadIdx.x) * 48) + 34)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[24] * kernel_shared[((((int)threadIdx.x) * 48) + 31)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[33] * kernel_shared[((((int)threadIdx.x) * 48) + 34)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[25] * kernel_shared[((((int)threadIdx.x) * 48) + 31)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[34] * kernel_shared[((((int)threadIdx.x) * 48) + 34)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[20] * kernel_shared[((((int)threadIdx.x) * 48) + 8)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[29] * kernel_shared[((((int)threadIdx.x) * 48) + 11)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[21] * kernel_shared[((((int)threadIdx.x) * 48) + 8)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[30] * kernel_shared[((((int)threadIdx.x) * 48) + 11)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[22] * kernel_shared[((((int)threadIdx.x) * 48) + 8)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[31] * kernel_shared[((((int)threadIdx.x) * 48) + 11)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[23] * kernel_shared[((((int)threadIdx.x) * 48) + 8)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[32] * kernel_shared[((((int)threadIdx.x) * 48) + 11)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[24] * kernel_shared[((((int)threadIdx.x) * 48) + 8)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[33] * kernel_shared[((((int)threadIdx.x) * 48) + 11)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[25] * kernel_shared[((((int)threadIdx.x) * 48) + 8)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[34] * kernel_shared[((((int)threadIdx.x) * 48) + 11)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[26] * kernel_shared[((((int)threadIdx.x) * 48) + 8)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[35] * kernel_shared[((((int)threadIdx.x) * 48) + 11)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[20] * kernel_shared[((((int)threadIdx.x) * 48) + 32)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[29] * kernel_shared[((((int)threadIdx.x) * 48) + 35)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[21] * kernel_shared[((((int)threadIdx.x) * 48) + 32)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[30] * kernel_shared[((((int)threadIdx.x) * 48) + 35)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[22] * kernel_shared[((((int)threadIdx.x) * 48) + 32)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[31] * kernel_shared[((((int)threadIdx.x) * 48) + 35)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[23] * kernel_shared[((((int)threadIdx.x) * 48) + 32)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[32] * kernel_shared[((((int)threadIdx.x) * 48) + 35)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[24] * kernel_shared[((((int)threadIdx.x) * 48) + 32)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[33] * kernel_shared[((((int)threadIdx.x) * 48) + 35)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[25] * kernel_shared[((((int)threadIdx.x) * 48) + 32)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[34] * kernel_shared[((((int)threadIdx.x) * 48) + 35)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[26] * kernel_shared[((((int)threadIdx.x) * 48) + 32)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[35] * kernel_shared[((((int)threadIdx.x) * 48) + 35)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[36] * kernel_shared[((((int)threadIdx.x) * 48) + 12)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[45] * kernel_shared[((((int)threadIdx.x) * 48) + 15)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[37] * kernel_shared[((((int)threadIdx.x) * 48) + 12)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[46] * kernel_shared[((((int)threadIdx.x) * 48) + 15)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[38] * kernel_shared[((((int)threadIdx.x) * 48) + 12)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[47] * kernel_shared[((((int)threadIdx.x) * 48) + 15)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[39] * kernel_shared[((((int)threadIdx.x) * 48) + 12)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[48] * kernel_shared[((((int)threadIdx.x) * 48) + 15)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[40] * kernel_shared[((((int)threadIdx.x) * 48) + 12)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[49] * kernel_shared[((((int)threadIdx.x) * 48) + 15)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[41] * kernel_shared[((((int)threadIdx.x) * 48) + 12)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[50] * kernel_shared[((((int)threadIdx.x) * 48) + 15)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[42] * kernel_shared[((((int)threadIdx.x) * 48) + 12)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[51] * kernel_shared[((((int)threadIdx.x) * 48) + 15)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[36] * kernel_shared[((((int)threadIdx.x) * 48) + 36)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[45] * kernel_shared[((((int)threadIdx.x) * 48) + 39)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[37] * kernel_shared[((((int)threadIdx.x) * 48) + 36)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[46] * kernel_shared[((((int)threadIdx.x) * 48) + 39)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[38] * kernel_shared[((((int)threadIdx.x) * 48) + 36)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[47] * kernel_shared[((((int)threadIdx.x) * 48) + 39)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[39] * kernel_shared[((((int)threadIdx.x) * 48) + 36)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[48] * kernel_shared[((((int)threadIdx.x) * 48) + 39)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[40] * kernel_shared[((((int)threadIdx.x) * 48) + 36)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[49] * kernel_shared[((((int)threadIdx.x) * 48) + 39)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[41] * kernel_shared[((((int)threadIdx.x) * 48) + 36)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[50] * kernel_shared[((((int)threadIdx.x) * 48) + 39)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[42] * kernel_shared[((((int)threadIdx.x) * 48) + 36)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[51] * kernel_shared[((((int)threadIdx.x) * 48) + 39)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[37] * kernel_shared[((((int)threadIdx.x) * 48) + 13)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[46] * kernel_shared[((((int)threadIdx.x) * 48) + 16)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[38] * kernel_shared[((((int)threadIdx.x) * 48) + 13)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[47] * kernel_shared[((((int)threadIdx.x) * 48) + 16)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[39] * kernel_shared[((((int)threadIdx.x) * 48) + 13)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[48] * kernel_shared[((((int)threadIdx.x) * 48) + 16)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[40] * kernel_shared[((((int)threadIdx.x) * 48) + 13)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[49] * kernel_shared[((((int)threadIdx.x) * 48) + 16)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[41] * kernel_shared[((((int)threadIdx.x) * 48) + 13)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[50] * kernel_shared[((((int)threadIdx.x) * 48) + 16)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[42] * kernel_shared[((((int)threadIdx.x) * 48) + 13)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[51] * kernel_shared[((((int)threadIdx.x) * 48) + 16)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[43] * kernel_shared[((((int)threadIdx.x) * 48) + 13)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[52] * kernel_shared[((((int)threadIdx.x) * 48) + 16)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[37] * kernel_shared[((((int)threadIdx.x) * 48) + 37)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[46] * kernel_shared[((((int)threadIdx.x) * 48) + 40)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[38] * kernel_shared[((((int)threadIdx.x) * 48) + 37)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[47] * kernel_shared[((((int)threadIdx.x) * 48) + 40)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[39] * kernel_shared[((((int)threadIdx.x) * 48) + 37)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[48] * kernel_shared[((((int)threadIdx.x) * 48) + 40)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[40] * kernel_shared[((((int)threadIdx.x) * 48) + 37)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[49] * kernel_shared[((((int)threadIdx.x) * 48) + 40)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[41] * kernel_shared[((((int)threadIdx.x) * 48) + 37)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[50] * kernel_shared[((((int)threadIdx.x) * 48) + 40)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[42] * kernel_shared[((((int)threadIdx.x) * 48) + 37)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[51] * kernel_shared[((((int)threadIdx.x) * 48) + 40)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[43] * kernel_shared[((((int)threadIdx.x) * 48) + 37)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[52] * kernel_shared[((((int)threadIdx.x) * 48) + 40)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[38] * kernel_shared[((((int)threadIdx.x) * 48) + 14)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[47] * kernel_shared[((((int)threadIdx.x) * 48) + 17)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[39] * kernel_shared[((((int)threadIdx.x) * 48) + 14)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[48] * kernel_shared[((((int)threadIdx.x) * 48) + 17)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[40] * kernel_shared[((((int)threadIdx.x) * 48) + 14)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[49] * kernel_shared[((((int)threadIdx.x) * 48) + 17)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[41] * kernel_shared[((((int)threadIdx.x) * 48) + 14)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[50] * kernel_shared[((((int)threadIdx.x) * 48) + 17)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[42] * kernel_shared[((((int)threadIdx.x) * 48) + 14)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[51] * kernel_shared[((((int)threadIdx.x) * 48) + 17)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[43] * kernel_shared[((((int)threadIdx.x) * 48) + 14)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[52] * kernel_shared[((((int)threadIdx.x) * 48) + 17)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[44] * kernel_shared[((((int)threadIdx.x) * 48) + 14)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[53] * kernel_shared[((((int)threadIdx.x) * 48) + 17)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[38] * kernel_shared[((((int)threadIdx.x) * 48) + 38)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[47] * kernel_shared[((((int)threadIdx.x) * 48) + 41)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[39] * kernel_shared[((((int)threadIdx.x) * 48) + 38)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[48] * kernel_shared[((((int)threadIdx.x) * 48) + 41)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[40] * kernel_shared[((((int)threadIdx.x) * 48) + 38)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[49] * kernel_shared[((((int)threadIdx.x) * 48) + 41)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[41] * kernel_shared[((((int)threadIdx.x) * 48) + 38)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[50] * kernel_shared[((((int)threadIdx.x) * 48) + 41)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[42] * kernel_shared[((((int)threadIdx.x) * 48) + 38)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[51] * kernel_shared[((((int)threadIdx.x) * 48) + 41)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[43] * kernel_shared[((((int)threadIdx.x) * 48) + 38)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[52] * kernel_shared[((((int)threadIdx.x) * 48) + 41)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[44] * kernel_shared[((((int)threadIdx.x) * 48) + 38)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[53] * kernel_shared[((((int)threadIdx.x) * 48) + 41)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[54] * kernel_shared[((((int)threadIdx.x) * 48) + 18)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[63] * kernel_shared[((((int)threadIdx.x) * 48) + 21)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[55] * kernel_shared[((((int)threadIdx.x) * 48) + 18)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[64] * kernel_shared[((((int)threadIdx.x) * 48) + 21)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[56] * kernel_shared[((((int)threadIdx.x) * 48) + 18)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[65] * kernel_shared[((((int)threadIdx.x) * 48) + 21)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[57] * kernel_shared[((((int)threadIdx.x) * 48) + 18)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[66] * kernel_shared[((((int)threadIdx.x) * 48) + 21)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[58] * kernel_shared[((((int)threadIdx.x) * 48) + 18)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[67] * kernel_shared[((((int)threadIdx.x) * 48) + 21)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[59] * kernel_shared[((((int)threadIdx.x) * 48) + 18)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[68] * kernel_shared[((((int)threadIdx.x) * 48) + 21)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[60] * kernel_shared[((((int)threadIdx.x) * 48) + 18)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[69] * kernel_shared[((((int)threadIdx.x) * 48) + 21)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[54] * kernel_shared[((((int)threadIdx.x) * 48) + 42)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[63] * kernel_shared[((((int)threadIdx.x) * 48) + 45)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[55] * kernel_shared[((((int)threadIdx.x) * 48) + 42)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[64] * kernel_shared[((((int)threadIdx.x) * 48) + 45)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[56] * kernel_shared[((((int)threadIdx.x) * 48) + 42)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[65] * kernel_shared[((((int)threadIdx.x) * 48) + 45)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[57] * kernel_shared[((((int)threadIdx.x) * 48) + 42)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[66] * kernel_shared[((((int)threadIdx.x) * 48) + 45)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[58] * kernel_shared[((((int)threadIdx.x) * 48) + 42)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[67] * kernel_shared[((((int)threadIdx.x) * 48) + 45)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[59] * kernel_shared[((((int)threadIdx.x) * 48) + 42)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[68] * kernel_shared[((((int)threadIdx.x) * 48) + 45)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[60] * kernel_shared[((((int)threadIdx.x) * 48) + 42)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[69] * kernel_shared[((((int)threadIdx.x) * 48) + 45)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[55] * kernel_shared[((((int)threadIdx.x) * 48) + 19)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[64] * kernel_shared[((((int)threadIdx.x) * 48) + 22)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[56] * kernel_shared[((((int)threadIdx.x) * 48) + 19)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[65] * kernel_shared[((((int)threadIdx.x) * 48) + 22)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[57] * kernel_shared[((((int)threadIdx.x) * 48) + 19)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[66] * kernel_shared[((((int)threadIdx.x) * 48) + 22)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[58] * kernel_shared[((((int)threadIdx.x) * 48) + 19)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[67] * kernel_shared[((((int)threadIdx.x) * 48) + 22)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[59] * kernel_shared[((((int)threadIdx.x) * 48) + 19)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[68] * kernel_shared[((((int)threadIdx.x) * 48) + 22)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[60] * kernel_shared[((((int)threadIdx.x) * 48) + 19)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[69] * kernel_shared[((((int)threadIdx.x) * 48) + 22)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[61] * kernel_shared[((((int)threadIdx.x) * 48) + 19)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[70] * kernel_shared[((((int)threadIdx.x) * 48) + 22)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[55] * kernel_shared[((((int)threadIdx.x) * 48) + 43)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[64] * kernel_shared[((((int)threadIdx.x) * 48) + 46)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[56] * kernel_shared[((((int)threadIdx.x) * 48) + 43)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[65] * kernel_shared[((((int)threadIdx.x) * 48) + 46)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[57] * kernel_shared[((((int)threadIdx.x) * 48) + 43)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[66] * kernel_shared[((((int)threadIdx.x) * 48) + 46)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[58] * kernel_shared[((((int)threadIdx.x) * 48) + 43)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[67] * kernel_shared[((((int)threadIdx.x) * 48) + 46)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[59] * kernel_shared[((((int)threadIdx.x) * 48) + 43)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[68] * kernel_shared[((((int)threadIdx.x) * 48) + 46)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[60] * kernel_shared[((((int)threadIdx.x) * 48) + 43)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[69] * kernel_shared[((((int)threadIdx.x) * 48) + 46)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[61] * kernel_shared[((((int)threadIdx.x) * 48) + 43)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[70] * kernel_shared[((((int)threadIdx.x) * 48) + 46)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[56] * kernel_shared[((((int)threadIdx.x) * 48) + 20)]));
+ conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[65] * kernel_shared[((((int)threadIdx.x) * 48) + 23)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[57] * kernel_shared[((((int)threadIdx.x) * 48) + 20)]));
+ conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[66] * kernel_shared[((((int)threadIdx.x) * 48) + 23)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[58] * kernel_shared[((((int)threadIdx.x) * 48) + 20)]));
+ conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[67] * kernel_shared[((((int)threadIdx.x) * 48) + 23)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[59] * kernel_shared[((((int)threadIdx.x) * 48) + 20)]));
+ conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[68] * kernel_shared[((((int)threadIdx.x) * 48) + 23)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[60] * kernel_shared[((((int)threadIdx.x) * 48) + 20)]));
+ conv2d_nchw[4] = (conv2d_nchw[4] + (pad_temp_shared[69] * kernel_shared[((((int)threadIdx.x) * 48) + 23)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[61] * kernel_shared[((((int)threadIdx.x) * 48) + 20)]));
+ conv2d_nchw[5] = (conv2d_nchw[5] + (pad_temp_shared[70] * kernel_shared[((((int)threadIdx.x) * 48) + 23)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[62] * kernel_shared[((((int)threadIdx.x) * 48) + 20)]));
+ conv2d_nchw[6] = (conv2d_nchw[6] + (pad_temp_shared[71] * kernel_shared[((((int)threadIdx.x) * 48) + 23)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[56] * kernel_shared[((((int)threadIdx.x) * 48) + 44)]));
+ conv2d_nchw[7] = (conv2d_nchw[7] + (pad_temp_shared[65] * kernel_shared[((((int)threadIdx.x) * 48) + 47)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[57] * kernel_shared[((((int)threadIdx.x) * 48) + 44)]));
+ conv2d_nchw[8] = (conv2d_nchw[8] + (pad_temp_shared[66] * kernel_shared[((((int)threadIdx.x) * 48) + 47)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[58] * kernel_shared[((((int)threadIdx.x) * 48) + 44)]));
+ conv2d_nchw[9] = (conv2d_nchw[9] + (pad_temp_shared[67] * kernel_shared[((((int)threadIdx.x) * 48) + 47)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[59] * kernel_shared[((((int)threadIdx.x) * 48) + 44)]));
+ conv2d_nchw[10] = (conv2d_nchw[10] + (pad_temp_shared[68] * kernel_shared[((((int)threadIdx.x) * 48) + 47)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[60] * kernel_shared[((((int)threadIdx.x) * 48) + 44)]));
+ conv2d_nchw[11] = (conv2d_nchw[11] + (pad_temp_shared[69] * kernel_shared[((((int)threadIdx.x) * 48) + 47)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[61] * kernel_shared[((((int)threadIdx.x) * 48) + 44)]));
+ conv2d_nchw[12] = (conv2d_nchw[12] + (pad_temp_shared[70] * kernel_shared[((((int)threadIdx.x) * 48) + 47)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[62] * kernel_shared[((((int)threadIdx.x) * 48) + 44)]));
+ conv2d_nchw[13] = (conv2d_nchw[13] + (pad_temp_shared[71] * kernel_shared[((((int)threadIdx.x) * 48) + 47)]));
}
- __syncthreads();
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7))] * kernel_shared[((((int)threadIdx.x) / 49) * 72)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7))] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 18)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 1)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 1)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 1)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 19)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 2)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 2)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 2)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 20)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 9)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 3)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 9)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 21)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 10)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 4)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 10)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 22)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 11)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 5)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 11)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 23)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 18)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 6)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 18)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 24)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 19)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 7)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 19)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 25)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 20)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 8)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 20)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 26)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7))] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 36)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7))] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 54)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 1)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 37)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 1)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 55)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 2)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 38)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 2)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 56)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 9)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 39)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 9)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 57)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 10)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 40)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 10)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 58)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 11)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 41)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 11)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 59)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 18)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 42)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 18)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 60)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 19)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 43)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 19)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 61)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 20)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 44)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 20)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 62)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 81)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 9)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 81)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 27)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 82)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 10)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 82)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 28)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 83)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 11)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 83)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 29)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 90)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 12)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 90)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 30)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 91)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 13)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 91)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 31)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 92)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 14)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 92)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 32)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 99)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 15)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 99)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 33)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 100)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 16)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 100)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 34)]));
- conv2d_nchw[0] = (conv2d_nchw[0] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 101)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 17)]));
- conv2d_nchw[1] = (conv2d_nchw[1] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 101)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 35)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 81)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 45)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 81)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 63)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 82)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 46)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 82)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 64)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 83)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 47)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 83)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 65)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 90)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 48)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 90)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 66)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 91)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 49)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 91)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 67)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 92)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 50)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 92)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 68)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 99)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 51)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 99)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 69)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 100)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 52)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 100)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 70)]));
- conv2d_nchw[2] = (conv2d_nchw[2] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 101)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 53)]));
- conv2d_nchw[3] = (conv2d_nchw[3] + (pad_temp_shared[(((((((int)threadIdx.x) % 49) / 7) * 9) + (((int)threadIdx.x) % 7)) + 101)] * kernel_shared[(((((int)threadIdx.x) / 49) * 72) + 71)]));
}
- for (int i1_inner = 0; i1_inner < 4; ++i1_inner) {
- compute[((((((int)blockIdx.x) * 392) + ((((int)threadIdx.x) / 49) * 196)) + (i1_inner * 49)) + (((int)threadIdx.x) % 49))] = max((conv2d_nchw[i1_inner] + bias[(((((int)blockIdx.x) * 8) + ((((int)threadIdx.x) / 49) * 4)) + i1_inner)]), 0.000000e+00f);
+ for (int i1_inner = 0; i1_inner < 2; ++i1_inner) {
+ for (int i3_inner = 0; i3_inner < 7; ++i3_inner) {
+ compute[((((((((int)blockIdx.x) / 7) * 6272) + (((int)threadIdx.x) * 98)) + (i1_inner * 49)) + ((((int)blockIdx.x) % 7) * 7)) + i3_inner)] = max((conv2d_nchw[((i1_inner * 7) + i3_inner)] + bias[((((((int)blockIdx.x) / 7) * 128) + (((int)threadIdx.x) * 2)) + i1_inner)]), 0.000000e+00f);
+ }
}
}
</pre></div>
@@ -870,7 +1579,7 @@ In the example below we resume the status and do more 5 trials.</p>
Get devices for measurement successfully!
</pre></div>
</div>
-<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 5 minutes 37.751 seconds)</p>
+<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 6 minutes 7.710 seconds)</p>
<div class="sphx-glr-footer sphx-glr-footer-example docutils container" id="sphx-glr-download-how-to-tune-with-autoscheduler-tune-conv2d-layer-cuda-py">
<div class="sphx-glr-download sphx-glr-download-python docutils container">
<p><a class="reference download internal" download="" href="../../_downloads/e3e540f3b477c0c52d8eb73e674e8ffd/tune_conv2d_layer_cuda.py"><code class="xref download docutils literal notranslate"><span class="pre">Download</span> <span class="pre">Python</span> <span class="pre">source</span> <span class="pre">code:</span> <span class="pre">tune_conv2d_layer_cuda.py</span></code></a></p>
diff --git a/docs/how_to/tune_with_autoscheduler/tune_network_cuda.html b/docs/how_to/tune_with_autoscheduler/tune_network_cuda.html
index b9702cf935..1e7eeb41cd 100644
--- a/docs/how_to/tune_with_autoscheduler/tune_network_cuda.html
+++ b/docs/how_to/tune_with_autoscheduler/tune_network_cuda.html
@@ -915,7 +915,7 @@ so we can read the log file and load the best schedules.</p>
Evaluate inference time cost...
Execution time summary:
mean (ms) median (ms) max (ms) min (ms) std (ms)
- 8.1788 8.1757 8.1876 8.1730 0.0064
+ 8.1987 8.1995 8.2098 8.1868 0.0094
</pre></div>
</div>
</div>
@@ -937,7 +937,7 @@ to learn how to use the RPC Tracker and RPC Server.
To use the RPC Tracker in auto-scheduler, replace the runner in <code class="code docutils literal notranslate"><span class="pre">TuningOptions</span></code>
with <a class="reference internal" href="../../reference/api/python/auto_scheduler.html#tvm.auto_scheduler.RPCRunner" title="tvm.auto_scheduler.RPCRunner"><code class="xref any py py-class docutils literal notranslate"><span class="pre">auto_scheduler.RPCRunner</span></code></a>.</p></li>
</ol>
-<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 1 minutes 4.612 seconds)</p>
+<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 1 minutes 4.692 seconds)</p>
<div class="sphx-glr-footer sphx-glr-footer-example docutils container" id="sphx-glr-download-how-to-tune-with-autoscheduler-tune-network-cuda-py">
<div class="sphx-glr-download sphx-glr-download-python docutils container">
<p><a class="reference download internal" download="" href="../../_downloads/eafe360d52540634c9eea0fa89e804bd/tune_network_cuda.py"><code class="xref download docutils literal notranslate"><span class="pre">Download</span> <span class="pre">Python</span> <span class="pre">source</span> <span class="pre">code:</span> <span class="pre">tune_network_cuda.py</span></code></a></p>
diff --git a/docs/how_to/tune_with_autoscheduler/tune_network_x86.html b/docs/how_to/tune_with_autoscheduler/tune_network_x86.html
index 5ec76dfcb2..b98474935d 100644
--- a/docs/how_to/tune_with_autoscheduler/tune_network_x86.html
+++ b/docs/how_to/tune_with_autoscheduler/tune_network_x86.html
@@ -934,7 +934,7 @@ so we can read the log file and load the best schedules.</p>
Evaluate inference time cost...
Execution time summary:
mean (ms) median (ms) max (ms) min (ms) std (ms)
- 777.0653 776.8267 779.3890 774.9802 1.8078
+ 752.0948 752.5188 753.8207 749.9449 1.6104
</pre></div>
</div>
</div>
@@ -956,7 +956,7 @@ to learn how to use the RPC Tracker and RPC Server.
To use the RPC Tracker in auto-scheduler, replace the runner in <code class="code docutils literal notranslate"><span class="pre">TuningOptions</span></code>
with <a class="reference internal" href="../../reference/api/python/auto_scheduler.html#tvm.auto_scheduler.RPCRunner" title="tvm.auto_scheduler.RPCRunner"><code class="xref any py py-class docutils literal notranslate"><span class="pre">auto_scheduler.RPCRunner</span></code></a>.</p></li>
</ol>
-<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 1 minutes 34.377 seconds)</p>
+<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 1 minutes 34.567 seconds)</p>
<div class="sphx-glr-footer sphx-glr-footer-example docutils container" id="sphx-glr-download-how-to-tune-with-autoscheduler-tune-network-x86-py">
<div class="sphx-glr-download sphx-glr-download-python docutils container">
<p><a class="reference download internal" download="" href="../../_downloads/e416b94ca1090b0897c0f6e0df95b911/tune_network_x86.py"><code class="xref download docutils literal notranslate"><span class="pre">Download</span> <span class="pre">Python</span> <span class="pre">source</span> <span class="pre">code:</span> <span class="pre">tune_network_x86.py</span></code></a></p>
diff --git a/docs/how_to/tune_with_autoscheduler/tune_sparse_x86.html b/docs/how_to/tune_with_autoscheduler/tune_sparse_x86.html
index 2411217ef0..f4c1cef722 100644
--- a/docs/how_to/tune_with_autoscheduler/tune_sparse_x86.html
+++ b/docs/how_to/tune_with_autoscheduler/tune_sparse_x86.html
@@ -625,416 +625,35 @@ layout transformation, parallelization, vectorization, unrolling, and operator f
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Lowered TIR:
@main = primfn(placeholder_5: handle, placeholder_6: handle, placeholder_7: handle, placeholder_8: handle, placeholder_9: handle, compute_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {placeholder: Buffer(placeholder_10: Pointer(float32), float32, [32768], []),
- placeholder_1: Buffer(placeholder_11: Pointer(float32), float32, [78656], []),
+ buffers = {placeholder: Buffer(placeholder_10: Pointer(float32), float32, [128, 256], []),
+ placeholder_1: Buffer(placeholder_11: Pointer(float32), float32, [4916, 16, 1], []),
placeholder_2: Buffer(placeholder_12: Pointer(int32), int32, [4916], []),
placeholder_3: Buffer(placeholder_13: Pointer(int32), int32, [33], []),
- placeholder_4: Buffer(placeholder_14: Pointer(float32), float32, [65536], []),
- compute: Buffer(compute_2: Pointer(float32), float32, [65536], [])}
- buffer_map = {placeholder_5: placeholder, placeholder_6: placeholder_1, placeholder_7: placeholder_2, placeholder_8: placeholder_3, placeholder_9: placeholder_4, compute_1: compute}
- preflattened_buffer_map = {placeholder_6: placeholder_15: Buffer(placeholder_11, float32, [4916, 16, 1], []), placeholder_8: placeholder_16: Buffer(placeholder_13, int32, [33], []), compute_1: compute_3: Buffer(compute_2, float32, [128, 512], []), placeholder_7: placeholder_17: Buffer(placeholder_12, int32, [4916], []), placeholder_5: placeholder_18: Buffer(placeholder_10, float32, [128, 256], []), placeholder_9: placeholder_19: Buffer(placeholder_14, float32, [128, 512], [])} {
- for (i0.outer.i1.outer.fused: int32, 0, 64) "parallel" {
- allocate(compute_4: Pointer(global float32), float32, [1024]), storage_scope = global {
- for (i.outer.inner: int32, 0, 4) {
- for (nb_j.inner: int32, 0, 2) {
- let cse_var_2: int32 = ((i.outer.inner*256) + (nb_j.inner*16))
- let cse_var_1: int32 = ((floormod(i0.outer.i1.outer.fused, 16)*2) + nb_j.inner)
- {
- compute_5: Buffer(compute_4, float32, [1024], [])[cse_var_2] = 0f32
- compute_5[(cse_var_2 + 1)] = 0f32
- compute_5[(cse_var_2 + 2)] = 0f32
- compute_5[(cse_var_2 + 3)] = 0f32
- compute_5[(cse_var_2 + 4)] = 0f32
- compute_5[(cse_var_2 + 5)] = 0f32
- compute_5[(cse_var_2 + 6)] = 0f32
- compute_5[(cse_var_2 + 7)] = 0f32
- compute_5[(cse_var_2 + 8)] = 0f32
- compute_5[(cse_var_2 + 9)] = 0f32
- compute_5[(cse_var_2 + 10)] = 0f32
- compute_5[(cse_var_2 + 11)] = 0f32
- compute_5[(cse_var_2 + 12)] = 0f32
- compute_5[(cse_var_2 + 13)] = 0f32
- compute_5[(cse_var_2 + 14)] = 0f32
- compute_5[(cse_var_2 + 15)] = 0f32
- compute_5[(cse_var_2 + 32)] = 0f32
- compute_5[(cse_var_2 + 33)] = 0f32
- compute_5[(cse_var_2 + 34)] = 0f32
- compute_5[(cse_var_2 + 35)] = 0f32
- compute_5[(cse_var_2 + 36)] = 0f32
- compute_5[(cse_var_2 + 37)] = 0f32
- compute_5[(cse_var_2 + 38)] = 0f32
- compute_5[(cse_var_2 + 39)] = 0f32
- compute_5[(cse_var_2 + 40)] = 0f32
- compute_5[(cse_var_2 + 41)] = 0f32
- compute_5[(cse_var_2 + 42)] = 0f32
- compute_5[(cse_var_2 + 43)] = 0f32
- compute_5[(cse_var_2 + 44)] = 0f32
- compute_5[(cse_var_2 + 45)] = 0f32
- compute_5[(cse_var_2 + 46)] = 0f32
- compute_5[(cse_var_2 + 47)] = 0f32
- compute_5[(cse_var_2 + 64)] = 0f32
- compute_5[(cse_var_2 + 65)] = 0f32
- compute_5[(cse_var_2 + 66)] = 0f32
- compute_5[(cse_var_2 + 67)] = 0f32
- compute_5[(cse_var_2 + 68)] = 0f32
- compute_5[(cse_var_2 + 69)] = 0f32
- compute_5[(cse_var_2 + 70)] = 0f32
- compute_5[(cse_var_2 + 71)] = 0f32
- compute_5[(cse_var_2 + 72)] = 0f32
- compute_5[(cse_var_2 + 73)] = 0f32
- compute_5[(cse_var_2 + 74)] = 0f32
- compute_5[(cse_var_2 + 75)] = 0f32
- compute_5[(cse_var_2 + 76)] = 0f32
- compute_5[(cse_var_2 + 77)] = 0f32
- compute_5[(cse_var_2 + 78)] = 0f32
- compute_5[(cse_var_2 + 79)] = 0f32
- compute_5[(cse_var_2 + 96)] = 0f32
- compute_5[(cse_var_2 + 97)] = 0f32
- compute_5[(cse_var_2 + 98)] = 0f32
- compute_5[(cse_var_2 + 99)] = 0f32
- compute_5[(cse_var_2 + 100)] = 0f32
- compute_5[(cse_var_2 + 101)] = 0f32
- compute_5[(cse_var_2 + 102)] = 0f32
- compute_5[(cse_var_2 + 103)] = 0f32
- compute_5[(cse_var_2 + 104)] = 0f32
- compute_5[(cse_var_2 + 105)] = 0f32
- compute_5[(cse_var_2 + 106)] = 0f32
- compute_5[(cse_var_2 + 107)] = 0f32
- compute_5[(cse_var_2 + 108)] = 0f32
- compute_5[(cse_var_2 + 109)] = 0f32
- compute_5[(cse_var_2 + 110)] = 0f32
- compute_5[(cse_var_2 + 111)] = 0f32
- compute_5[(cse_var_2 + 128)] = 0f32
- compute_5[(cse_var_2 + 129)] = 0f32
- compute_5[(cse_var_2 + 130)] = 0f32
- compute_5[(cse_var_2 + 131)] = 0f32
- compute_5[(cse_var_2 + 132)] = 0f32
- compute_5[(cse_var_2 + 133)] = 0f32
- compute_5[(cse_var_2 + 134)] = 0f32
- compute_5[(cse_var_2 + 135)] = 0f32
- compute_5[(cse_var_2 + 136)] = 0f32
- compute_5[(cse_var_2 + 137)] = 0f32
- compute_5[(cse_var_2 + 138)] = 0f32
- compute_5[(cse_var_2 + 139)] = 0f32
- compute_5[(cse_var_2 + 140)] = 0f32
- compute_5[(cse_var_2 + 141)] = 0f32
- compute_5[(cse_var_2 + 142)] = 0f32
- compute_5[(cse_var_2 + 143)] = 0f32
- compute_5[(cse_var_2 + 160)] = 0f32
- compute_5[(cse_var_2 + 161)] = 0f32
- compute_5[(cse_var_2 + 162)] = 0f32
- compute_5[(cse_var_2 + 163)] = 0f32
- compute_5[(cse_var_2 + 164)] = 0f32
- compute_5[(cse_var_2 + 165)] = 0f32
- compute_5[(cse_var_2 + 166)] = 0f32
- compute_5[(cse_var_2 + 167)] = 0f32
- compute_5[(cse_var_2 + 168)] = 0f32
- compute_5[(cse_var_2 + 169)] = 0f32
- compute_5[(cse_var_2 + 170)] = 0f32
- compute_5[(cse_var_2 + 171)] = 0f32
- compute_5[(cse_var_2 + 172)] = 0f32
- compute_5[(cse_var_2 + 173)] = 0f32
- compute_5[(cse_var_2 + 174)] = 0f32
- compute_5[(cse_var_2 + 175)] = 0f32
- compute_5[(cse_var_2 + 192)] = 0f32
- compute_5[(cse_var_2 + 193)] = 0f32
- compute_5[(cse_var_2 + 194)] = 0f32
- compute_5[(cse_var_2 + 195)] = 0f32
- compute_5[(cse_var_2 + 196)] = 0f32
- compute_5[(cse_var_2 + 197)] = 0f32
- compute_5[(cse_var_2 + 198)] = 0f32
- compute_5[(cse_var_2 + 199)] = 0f32
- compute_5[(cse_var_2 + 200)] = 0f32
- compute_5[(cse_var_2 + 201)] = 0f32
- compute_5[(cse_var_2 + 202)] = 0f32
- compute_5[(cse_var_2 + 203)] = 0f32
- compute_5[(cse_var_2 + 204)] = 0f32
- compute_5[(cse_var_2 + 205)] = 0f32
- compute_5[(cse_var_2 + 206)] = 0f32
- compute_5[(cse_var_2 + 207)] = 0f32
- compute_5[(cse_var_2 + 224)] = 0f32
- compute_5[(cse_var_2 + 225)] = 0f32
- compute_5[(cse_var_2 + 226)] = 0f32
- compute_5[(cse_var_2 + 227)] = 0f32
- compute_5[(cse_var_2 + 228)] = 0f32
- compute_5[(cse_var_2 + 229)] = 0f32
- compute_5[(cse_var_2 + 230)] = 0f32
- compute_5[(cse_var_2 + 231)] = 0f32
- compute_5[(cse_var_2 + 232)] = 0f32
- compute_5[(cse_var_2 + 233)] = 0f32
- compute_5[(cse_var_2 + 234)] = 0f32
- compute_5[(cse_var_2 + 235)] = 0f32
- compute_5[(cse_var_2 + 236)] = 0f32
- compute_5[(cse_var_2 + 237)] = 0f32
- compute_5[(cse_var_2 + 238)] = 0f32
- compute_5[(cse_var_2 + 239)] = 0f32
- for (elem_idx: int32, 0, (placeholder_3[(cse_var_1 + 1)] - placeholder_3[cse_var_1])) {
- let cse_var_131: int32 = (elem_idx*16)
- let cse_var_130: int32 = (cse_var_2 + 99)
- let cse_var_129: int32 = (cse_var_2 + 98)
- let cse_var_128: int32 = (cse_var_2 + 97)
- let cse_var_127: int32 = (cse_var_2 + 96)
- let cse_var_126: int32 = (cse_var_2 + 9)
- let cse_var_125: int32 = (cse_var_2 + 8)
- let cse_var_124: int32 = (cse_var_2 + 79)
- let cse_var_123: int32 = (cse_var_2 + 78)
- let cse_var_122: int32 = (cse_var_2 + 77)
- let cse_var_121: int32 = (cse_var_2 + 76)
- let cse_var_120: int32 = (cse_var_2 + 75)
- let cse_var_119: int32 = (cse_var_2 + 74)
- let cse_var_118: int32 = (cse_var_2 + 73)
- let cse_var_117: int32 = (cse_var_2 + 72)
- let cse_var_116: int32 = (cse_var_2 + 71)
- let cse_var_115: int32 = (cse_var_2 + 70)
- let cse_var_114: int32 = (cse_var_2 + 7)
- let cse_var_113: int32 = (cse_var_2 + 69)
- let cse_var_112: int32 = (cse_var_2 + 68)
- let cse_var_111: int32 = (cse_var_2 + 67)
- let cse_var_110: int32 = (cse_var_2 + 66)
- let cse_var_109: int32 = (cse_var_2 + 65)
- let cse_var_108: int32 = (cse_var_2 + 64)
- let cse_var_107: int32 = (cse_var_2 + 6)
- let cse_var_106: int32 = (cse_var_2 + 5)
- let cse_var_105: int32 = (cse_var_2 + 47)
- let cse_var_104: int32 = (cse_var_2 + 46)
- let cse_var_103: int32 = (cse_var_2 + 45)
- let cse_var_102: int32 = (cse_var_2 + 44)
- let cse_var_101: int32 = (cse_var_2 + 43)
- let cse_var_100: int32 = (cse_var_2 + 42)
- let cse_var_99: int32 = (cse_var_2 + 41)
- let cse_var_98: int32 = (cse_var_2 + 40)
- let cse_var_97: int32 = (cse_var_2 + 4)
- let cse_var_96: int32 = (cse_var_2 + 39)
- let cse_var_95: int32 = (cse_var_2 + 38)
- let cse_var_94: int32 = (cse_var_2 + 37)
- let cse_var_93: int32 = (cse_var_2 + 36)
- let cse_var_92: int32 = (cse_var_2 + 35)
- let cse_var_91: int32 = (cse_var_2 + 34)
- let cse_var_90: int32 = (cse_var_2 + 33)
- let cse_var_89: int32 = (cse_var_2 + 32)
- let cse_var_88: int32 = (cse_var_2 + 3)
- let cse_var_87: int32 = (cse_var_2 + 239)
- let cse_var_86: int32 = (cse_var_2 + 238)
- let cse_var_85: int32 = (cse_var_2 + 237)
- let cse_var_84: int32 = (cse_var_2 + 236)
- let cse_var_83: int32 = (cse_var_2 + 235)
- let cse_var_82: int32 = (cse_var_2 + 234)
- let cse_var_81: int32 = (cse_var_2 + 233)
- let cse_var_80: int32 = (cse_var_2 + 232)
- let cse_var_79: int32 = (cse_var_2 + 231)
- let cse_var_78: int32 = (cse_var_2 + 230)
- let cse_var_77: int32 = (cse_var_2 + 229)
- let cse_var_76: int32 = (cse_var_2 + 228)
- let cse_var_75: int32 = (cse_var_2 + 227)
- let cse_var_74: int32 = (cse_var_2 + 226)
- let cse_var_73: int32 = (cse_var_2 + 225)
- let cse_var_72: int32 = (cse_var_2 + 224)
- let cse_var_71: int32 = (cse_var_2 + 207)
- let cse_var_70: int32 = (cse_var_2 + 206)
- let cse_var_69: int32 = (cse_var_2 + 205)
- let cse_var_68: int32 = (cse_var_2 + 204)
- let cse_var_67: int32 = (cse_var_2 + 203)
- let cse_var_66: int32 = (cse_var_2 + 202)
- let cse_var_65: int32 = (cse_var_2 + 201)
- let cse_var_64: int32 = (cse_var_2 + 200)
- let cse_var_63: int32 = (cse_var_2 + 2)
- let cse_var_62: int32 = (cse_var_2 + 199)
- let cse_var_61: int32 = (cse_var_2 + 198)
- let cse_var_60: int32 = (cse_var_2 + 197)
- let cse_var_59: int32 = (cse_var_2 + 196)
- let cse_var_58: int32 = (cse_var_2 + 195)
- let cse_var_57: int32 = (cse_var_2 + 194)
- let cse_var_56: int32 = (cse_var_2 + 193)
- let cse_var_55: int32 = (cse_var_2 + 192)
- let cse_var_54: int32 = (cse_var_2 + 175)
- let cse_var_53: int32 = (cse_var_2 + 174)
- let cse_var_52: int32 = (cse_var_2 + 173)
- let cse_var_51: int32 = (cse_var_2 + 172)
- let cse_var_50: int32 = (cse_var_2 + 171)
- let cse_var_49: int32 = (cse_var_2 + 170)
- let cse_var_48: int32 = (cse_var_2 + 169)
- let cse_var_47: int32 = (cse_var_2 + 168)
- let cse_var_46: int32 = (cse_var_2 + 167)
- let cse_var_45: int32 = (cse_var_2 + 166)
- let cse_var_44: int32 = (cse_var_2 + 165)
- let cse_var_43: int32 = (cse_var_2 + 164)
- let cse_var_42: int32 = (cse_var_2 + 163)
- let cse_var_41: int32 = (cse_var_2 + 162)
- let cse_var_40: int32 = (cse_var_2 + 161)
- let cse_var_39: int32 = (cse_var_2 + 160)
- let cse_var_38: int32 = (cse_var_2 + 15)
- let cse_var_37: int32 = (cse_var_2 + 143)
- let cse_var_36: int32 = (cse_var_2 + 142)
- let cse_var_35: int32 = (cse_var_2 + 141)
- let cse_var_34: int32 = (cse_var_2 + 140)
- let cse_var_33: int32 = (cse_var_2 + 14)
- let cse_var_32: int32 = (cse_var_2 + 139)
- let cse_var_31: int32 = (cse_var_2 + 138)
- let cse_var_30: int32 = (cse_var_2 + 137)
- let cse_var_29: int32 = (cse_var_2 + 136)
- let cse_var_28: int32 = (cse_var_2 + 135)
- let cse_var_27: int32 = (cse_var_2 + 134)
- let cse_var_26: int32 = (cse_var_2 + 133)
- let cse_var_25: int32 = (cse_var_2 + 132)
- let cse_var_24: int32 = (cse_var_2 + 131)
- let cse_var_23: int32 = (cse_var_2 + 130)
- let cse_var_22: int32 = (cse_var_2 + 13)
- let cse_var_21: int32 = (cse_var_2 + 129)
- let cse_var_20: int32 = (cse_var_2 + 128)
- let cse_var_19: int32 = (cse_var_2 + 12)
- let cse_var_18: int32 = (cse_var_2 + 111)
- let cse_var_17: int32 = (cse_var_2 + 110)
- let cse_var_16: int32 = (cse_var_2 + 11)
- let cse_var_15: int32 = (cse_var_2 + 109)
- let cse_var_14: int32 = (cse_var_2 + 108)
- let cse_var_13: int32 = (cse_var_2 + 107)
- let cse_var_12: int32 = (cse_var_2 + 106)
- let cse_var_11: int32 = (cse_var_2 + 105)
- let cse_var_10: int32 = (cse_var_2 + 104)
- let cse_var_9: int32 = (cse_var_2 + 103)
- let cse_var_8: int32 = (cse_var_2 + 102)
- let cse_var_7: int32 = (cse_var_2 + 101)
- let cse_var_6: int32 = (cse_var_2 + 100)
- let cse_var_5: int32 = (cse_var_2 + 10)
- let cse_var_4: int32 = (cse_var_2 + 1)
- let cse_var_3: int32 = ((floordiv(i0.outer.i1.outer.fused, 16)*8192) + (i.outer.inner*2048))
- {
- compute_5[cse_var_2] = (compute_5[cse_var_2] + (placeholder_1[((placeholder_3[cse_var_1]*16) + cse_var_131)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_4] = (compute_5[cse_var_4] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 1)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_63] = (compute_5[cse_var_63] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 2)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_88] = (compute_5[cse_var_88] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 3)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_97] = (compute_5[cse_var_97] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 4)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_106] = (compute_5[cse_var_106] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 5)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_107] = (compute_5[cse_var_107] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 6)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_114] = (compute_5[cse_var_114] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 7)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_125] = (compute_5[cse_var_125] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 8)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_126] = (compute_5[cse_var_126] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 9)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_5] = (compute_5[cse_var_5] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 10)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_16] = (compute_5[cse_var_16] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 11)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_19] = (compute_5[cse_var_19] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 12)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_22] = (compute_5[cse_var_22] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 13)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_33] = (compute_5[cse_var_33] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 14)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_38] = (compute_5[cse_var_38] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 15)]*max(placeholder[(cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)])], 0f32)))
- compute_5[cse_var_89] = (compute_5[cse_var_89] + (placeholder_1[((placeholder_3[cse_var_1]*16) + cse_var_131)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_90] = (compute_5[cse_var_90] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 1)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_91] = (compute_5[cse_var_91] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 2)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_92] = (compute_5[cse_var_92] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 3)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_93] = (compute_5[cse_var_93] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 4)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_94] = (compute_5[cse_var_94] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 5)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_95] = (compute_5[cse_var_95] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 6)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_96] = (compute_5[cse_var_96] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 7)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_98] = (compute_5[cse_var_98] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 8)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_99] = (compute_5[cse_var_99] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 9)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_100] = (compute_5[cse_var_100] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 10)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_101] = (compute_5[cse_var_101] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 11)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_102] = (compute_5[cse_var_102] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 12)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_103] = (compute_5[cse_var_103] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 13)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_104] = (compute_5[cse_var_104] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 14)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_105] = (compute_5[cse_var_105] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 15)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 256)], 0f32)))
- compute_5[cse_var_108] = (compute_5[cse_var_108] + (placeholder_1[((placeholder_3[cse_var_1]*16) + cse_var_131)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_109] = (compute_5[cse_var_109] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 1)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_110] = (compute_5[cse_var_110] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 2)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_111] = (compute_5[cse_var_111] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 3)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_112] = (compute_5[cse_var_112] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 4)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_113] = (compute_5[cse_var_113] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 5)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_115] = (compute_5[cse_var_115] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 6)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_116] = (compute_5[cse_var_116] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 7)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_117] = (compute_5[cse_var_117] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 8)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_118] = (compute_5[cse_var_118] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 9)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_119] = (compute_5[cse_var_119] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 10)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_120] = (compute_5[cse_var_120] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 11)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_121] = (compute_5[cse_var_121] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 12)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_122] = (compute_5[cse_var_122] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 13)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_123] = (compute_5[cse_var_123] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 14)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_124] = (compute_5[cse_var_124] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 15)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 512)], 0f32)))
- compute_5[cse_var_127] = (compute_5[cse_var_127] + (placeholder_1[((placeholder_3[cse_var_1]*16) + cse_var_131)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_128] = (compute_5[cse_var_128] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 1)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_129] = (compute_5[cse_var_129] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 2)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_130] = (compute_5[cse_var_130] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 3)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_6] = (compute_5[cse_var_6] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 4)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_7] = (compute_5[cse_var_7] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 5)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_8] = (compute_5[cse_var_8] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 6)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_9] = (compute_5[cse_var_9] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 7)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_10] = (compute_5[cse_var_10] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 8)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_11] = (compute_5[cse_var_11] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 9)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_12] = (compute_5[cse_var_12] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 10)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_13] = (compute_5[cse_var_13] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 11)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_14] = (compute_5[cse_var_14] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 12)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_15] = (compute_5[cse_var_15] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 13)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_17] = (compute_5[cse_var_17] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 14)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_18] = (compute_5[cse_var_18] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 15)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 768)], 0f32)))
- compute_5[cse_var_20] = (compute_5[cse_var_20] + (placeholder_1[((placeholder_3[cse_var_1]*16) + cse_var_131)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_21] = (compute_5[cse_var_21] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 1)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_23] = (compute_5[cse_var_23] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 2)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_24] = (compute_5[cse_var_24] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 3)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_25] = (compute_5[cse_var_25] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 4)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_26] = (compute_5[cse_var_26] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 5)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_27] = (compute_5[cse_var_27] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 6)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_28] = (compute_5[cse_var_28] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 7)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_29] = (compute_5[cse_var_29] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 8)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_30] = (compute_5[cse_var_30] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 9)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_31] = (compute_5[cse_var_31] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 10)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_32] = (compute_5[cse_var_32] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 11)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_34] = (compute_5[cse_var_34] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 12)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_35] = (compute_5[cse_var_35] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 13)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_36] = (compute_5[cse_var_36] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 14)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_37] = (compute_5[cse_var_37] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 15)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1024)], 0f32)))
- compute_5[cse_var_39] = (compute_5[cse_var_39] + (placeholder_1[((placeholder_3[cse_var_1]*16) + cse_var_131)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_40] = (compute_5[cse_var_40] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 1)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_41] = (compute_5[cse_var_41] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 2)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_42] = (compute_5[cse_var_42] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 3)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_43] = (compute_5[cse_var_43] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 4)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_44] = (compute_5[cse_var_44] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 5)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_45] = (compute_5[cse_var_45] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 6)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_46] = (compute_5[cse_var_46] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 7)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_47] = (compute_5[cse_var_47] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 8)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_48] = (compute_5[cse_var_48] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 9)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_49] = (compute_5[cse_var_49] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 10)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_50] = (compute_5[cse_var_50] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 11)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_51] = (compute_5[cse_var_51] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 12)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_52] = (compute_5[cse_var_52] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 13)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_53] = (compute_5[cse_var_53] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 14)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_54] = (compute_5[cse_var_54] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 15)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1280)], 0f32)))
- compute_5[cse_var_55] = (compute_5[cse_var_55] + (placeholder_1[((placeholder_3[cse_var_1]*16) + cse_var_131)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_56] = (compute_5[cse_var_56] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 1)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_57] = (compute_5[cse_var_57] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 2)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_58] = (compute_5[cse_var_58] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 3)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_59] = (compute_5[cse_var_59] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 4)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_60] = (compute_5[cse_var_60] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 5)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_61] = (compute_5[cse_var_61] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 6)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_62] = (compute_5[cse_var_62] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 7)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_64] = (compute_5[cse_var_64] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 8)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_65] = (compute_5[cse_var_65] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 9)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_66] = (compute_5[cse_var_66] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 10)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_67] = (compute_5[cse_var_67] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 11)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_68] = (compute_5[cse_var_68] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 12)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_69] = (compute_5[cse_var_69] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 13)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_70] = (compute_5[cse_var_70] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 14)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_71] = (compute_5[cse_var_71] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 15)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1536)], 0f32)))
- compute_5[cse_var_72] = (compute_5[cse_var_72] + (placeholder_1[((placeholder_3[cse_var_1]*16) + cse_var_131)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_73] = (compute_5[cse_var_73] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 1)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_74] = (compute_5[cse_var_74] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 2)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_75] = (compute_5[cse_var_75] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 3)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_76] = (compute_5[cse_var_76] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 4)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_77] = (compute_5[cse_var_77] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 5)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_78] = (compute_5[cse_var_78] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 6)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_79] = (compute_5[cse_var_79] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 7)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_80] = (compute_5[cse_var_80] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 8)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_81] = (compute_5[cse_var_81] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 9)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_82] = (compute_5[cse_var_82] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 10)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_83] = (compute_5[cse_var_83] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 11)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_84] = (compute_5[cse_var_84] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 12)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_85] = (compute_5[cse_var_85] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 13)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_86] = (compute_5[cse_var_86] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 14)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
- compute_5[cse_var_87] = (compute_5[cse_var_87] + (placeholder_1[(((placeholder_3[cse_var_1]*16) + cse_var_131) + 15)]*max(placeholder[((cse_var_3 + placeholder_2[(placeholder_3[cse_var_1] + elem_idx)]) + 1792)], 0f32)))
+ placeholder_4: Buffer(placeholder_14: Pointer(float32), float32, [128, 512], []),
+ compute: Buffer(compute_2: Pointer(float32), float32, [128, 512], [])}
+ buffer_map = {placeholder_5: placeholder, placeholder_6: placeholder_1, placeholder_7: placeholder_2, placeholder_8: placeholder_3, placeholder_9: placeholder_4, compute_1: compute} {
+ for (i0.outer.i1.outer.fused: int32, 0, 32) "parallel" {
+ allocate(compute_3: Pointer(global float32), float32, [2048]), storage_scope = global {
+ for (i.outer.inner: int32, 0, 2) {
+ for (i.inner.init: int32, 0, 64) {
+ for (j.init: int32, 0, 16) {
+ compute_4: Buffer(compute_3, float32, [2048], [])[(((i.outer.inner*1024) + (i.inner.init*16)) + j.init)] = 0f32
+ }
+ }
+ for (elem_idx: int32, 0, (placeholder_15: Buffer(placeholder_13, int32, [33], [])[(i0.outer.i1.outer.fused + 1)] - placeholder_15[i0.outer.i1.outer.fused])) {
+ for (i.inner: int32, 0, 64) {
+ for (j: int32, 0, 16) {
+ if @tir.likely((elem_idx < (placeholder_15[(i0.outer.i1.outer.fused + 1)] - placeholder_15[i0.outer.i1.outer.fused])), dtype=bool) {
+ let cse_var_1: int32 = (((i.outer.inner*1024) + (i.inner*16)) + j)
+ compute_4[cse_var_1] = (compute_4[cse_var_1] + (placeholder_16: Buffer(placeholder_11, float32, [78656], [])[(((placeholder_15[i0.outer.i1.outer.fused]*16) + (elem_idx*16)) + j)]*max(placeholder_17: Buffer(placeholder_10, float32, [32768], [])[(((i.outer.inner*16384) + (i.inner*256)) + placeholder_18: Buffer(placeholder_12, int32, [4916], [])[(placeholder_15[i0.outer.i1.outer.fused] + elem_idx)])], 0f32)))
}
}
}
}
}
- for (i0.inner: int32, 0, 32) {
- let cse_var_132: int32 = (((floordiv(i0.outer.i1.outer.fused, 16)*16384) + (i0.inner*512)) + (floormod(i0.outer.i1.outer.fused, 16)*32))
- compute[ramp(cse_var_132, 1, 32)] = max((compute_5[ramp((i0.inner*32), 1, 32)] + placeholder_4[ramp(cse_var_132, 1, 32)]), broadcast(0f32, 32))
+ for (i0.inner: int32, 0, 128) {
+ let cse_var_2: int32 = ((i0.inner*512) + (i0.outer.i1.outer.fused*16))
+ compute_5: Buffer(compute_2, float32, [65536], [])[ramp(cse_var_2, 1, 16)] = max((compute_4[ramp((i0.inner*16), 1, 16)] + placeholder_19: Buffer(placeholder_14, float32, [65536], [])[ramp(cse_var_2, 1, 16)]), broadcast(0f32, 16))
}
}
}
@@ -1072,7 +691,7 @@ layout transformation, parallelization, vectorization, unrolling, and operator f
<span class="p">)</span>
</pre></div>
</div>
-<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Execution time of this operator: 2.871 ms
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Execution time of this operator: 1.807 ms
</pre></div>
</div>
<div class="admonition note">
diff --git a/docs/how_to/tune_with_autotvm/sg_execution_times.html b/docs/how_to/tune_with_autotvm/sg_execution_times.html
index 85b5c21b7f..4544a752eb 100644
--- a/docs/how_to/tune_with_autotvm/sg_execution_times.html
+++ b/docs/how_to/tune_with_autotvm/sg_execution_times.html
@@ -340,7 +340,7 @@
<div class="section" id="computation-times">
<span id="sphx-glr-how-to-tune-with-autotvm-sg-execution-times"></span><h1>Computation times<a class="headerlink" href="#computation-times" title="Permalink to this headline">¶</a></h1>
-<p><strong>00:23.851</strong> total execution time for <strong>how_to_tune_with_autotvm</strong> files:</p>
+<p><strong>00:33.070</strong> total execution time for <strong>how_to_tune_with_autotvm</strong> files:</p>
<table class="docutils align-default">
<colgroup>
<col style="width: 84%" />
@@ -349,22 +349,22 @@
</colgroup>
<tbody>
<tr class="row-odd"><td><p><a class="reference internal" href="tune_conv2d_cuda.html#sphx-glr-how-to-tune-with-autotvm-tune-conv2d-cuda-py"><span class="std std-ref">Tuning High Performance Convolution on NVIDIA GPUs</span></a> (<code class="docutils literal notranslate"><span class="pre">tune_conv2d_cuda.py</span></code>)</p></td>
-<td><p>00:23.816</p></td>
+<td><p>00:33.035</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="tune_relay_x86.html#sphx-glr-how-to-tune-with-autotvm-tune-relay-x86-py"><span class="std std-ref">Auto-tuning a Convolutional Network for x86 CPU</span></a> (<code class="docutils literal notranslate"><span class="pre">tune_relay_x86.py</span></code>)</p></td>
<td><p>00:00.020</p></td>
<td><p>0.0 MB</p></td>
</tr>
-<tr class="row-odd"><td><p><a class="reference internal" href="tune_relay_mobile_gpu.html#sphx-glr-how-to-tune-with-autotvm-tune-relay-mobile-gpu-py"><span class="std std-ref">Auto-tuning a Convolutional Network for Mobile GPU</span></a> (<code class="docutils literal notranslate"><span class="pre">tune_relay_mobile_gpu.py</span></code>)</p></td>
+<tr class="row-odd"><td><p><a class="reference internal" href="tune_relay_cuda.html#sphx-glr-how-to-tune-with-autotvm-tune-relay-cuda-py"><span class="std std-ref">Auto-tuning a Convolutional Network for NVIDIA GPU</span></a> (<code class="docutils literal notranslate"><span class="pre">tune_relay_cuda.py</span></code>)</p></td>
<td><p>00:00.005</p></td>
<td><p>0.0 MB</p></td>
</tr>
-<tr class="row-even"><td><p><a class="reference internal" href="tune_relay_cuda.html#sphx-glr-how-to-tune-with-autotvm-tune-relay-cuda-py"><span class="std std-ref">Auto-tuning a Convolutional Network for NVIDIA GPU</span></a> (<code class="docutils literal notranslate"><span class="pre">tune_relay_cuda.py</span></code>)</p></td>
+<tr class="row-even"><td><p><a class="reference internal" href="tune_relay_arm.html#sphx-glr-how-to-tune-with-autotvm-tune-relay-arm-py"><span class="std std-ref">Auto-tuning a Convolutional Network for ARM CPU</span></a> (<code class="docutils literal notranslate"><span class="pre">tune_relay_arm.py</span></code>)</p></td>
<td><p>00:00.005</p></td>
<td><p>0.0 MB</p></td>
</tr>
-<tr class="row-odd"><td><p><a class="reference internal" href="tune_relay_arm.html#sphx-glr-how-to-tune-with-autotvm-tune-relay-arm-py"><span class="std std-ref">Auto-tuning a Convolutional Network for ARM CPU</span></a> (<code class="docutils literal notranslate"><span class="pre">tune_relay_arm.py</span></code>)</p></td>
+<tr class="row-odd"><td><p><a class="reference internal" href="tune_relay_mobile_gpu.html#sphx-glr-how-to-tune-with-autotvm-tune-relay-mobile-gpu-py"><span class="std std-ref">Auto-tuning a Convolutional Network for Mobile GPU</span></a> (<code class="docutils literal notranslate"><span class="pre">tune_relay_mobile_gpu.py</span></code>)</p></td>
<td><p>00:00.005</p></td>
<td><p>0.0 MB</p></td>
</tr>
diff --git a/docs/how_to/tune_with_autotvm/tune_conv2d_cuda.html b/docs/how_to/tune_with_autotvm/tune_conv2d_cuda.html
index 45a77dd74e..89279b0b61 100644
--- a/docs/how_to/tune_with_autotvm/tune_conv2d_cuda.html
+++ b/docs/how_to/tune_with_autotvm/tune_conv2d_cuda.html
@@ -567,9 +567,9 @@ for this template</p>
waiting for device...
device available
Get devices for measurement successfully!
-No: 1 GFLOPS: 24.37/24.37 result: MeasureResult(costs=(0.0094983644375,), error_no=MeasureErrorNo.NO_ERROR, all_cost=3.947620153427124, timestamp=1668585629.6380484) [('tile_f', [-1, 16, 2, 1]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 64, 1]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,3699754
-No: 2 GFLOPS: 16.16/24.37 result: MeasureResult(costs=(0.014324914,), error_no=MeasureErrorNo.NO_ERROR, all_cost=3.4801909923553467, timestamp=1668585630.3499599) [('tile_f', [-1, 16, 8, 2]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 4, 2]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,3915643
-No: 3 GFLOPS: 0.00/24.37 result: Traceback (most recent call last):
+No: 1 GFLOPS: 62.40/62.40 result: MeasureResult(costs=(0.003710196272727273,), error_no=MeasureErrorNo.NO_ERROR, all_cost=2.506129264831543, timestamp=1668612674.5134175) [('tile_f', [-1, 8, 4, 4]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 1, 4]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,4520238
+No: 2 GFLOPS: 46.81/62.40 result: MeasureResult(costs=(0.004945171857142857,), error_no=MeasureErrorNo.NO_ERROR, all_cost=2.2848103046417236, timestamp=1668612676.0735888) [('tile_f', [-1, 4, 16, 2]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 1, 16]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,4185807
+No: 3 GFLOPS: 0.00/62.40 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -691,8 +691,8 @@ Traceback (most recent call last):
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
-tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 16, 16, 2]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 16, 32]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,8482849
-No: 4 GFLOPS: 0.00/24.37 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 64, 4, 1]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 8, 8]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,4948265
+No: 4 GFLOPS: 0.00/62.40 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -814,8 +814,9 @@ Traceback (most recent call last):
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
-tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 2, 1, 64]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 4, 4]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9755001
-No: 5 GFLOPS: 0.00/24.37 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 1, 32, 16]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 128, 4]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,3192384
+No: 5 GFLOPS: 4.12/62.40 result: MeasureResult(costs=(0.05621964625,), error_no=MeasureErrorNo.NO_ERROR, all_cost=2.726212501525879, timestamp=1668612679.9325876) [('tile_f', [-1, 8, 4, 8]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 4, 4]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,74952
+No: 6 GFLOPS: 0.00/62.40 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -937,9 +938,10 @@ Traceback (most recent call last):
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
-tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 1, 2, 16]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 4, 16]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,902170
-No: 6 GFLOPS: 74.52/74.52 result: MeasureResult(costs=(0.0031066027272727273,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.2222695350646973, timestamp=1668585633.5934608) [('tile_f', [-1, 8, 4, 4]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 4, 4]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,1818638
-No: 7 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 16, 4, 1]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 8, 8]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 1)],None,5528183
+No: 7 GFLOPS: 200.02/200.02 result: MeasureResult(costs=(0.001157369597122302,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.7944960594177246, timestamp=1668612681.7724178) [('tile_f', [-1, 1, 16, 1]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 1, 32]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,8467394
+No: 8 GFLOPS: 87.19/200.02 result: MeasureResult(costs=(0.0026551532195121953,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.1284987926483154, timestamp=1668612682.474605) [('tile_f', [-1, 1, 2, 2]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 4, 16]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,127224
+No: 9 GFLOPS: 0.00/200.02 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -1061,8 +1063,10 @@ Traceback (most recent call last):
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
-tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 2, 1, 128]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 8, 32]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,7317191
-No: 8 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 1, 32, 1]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 2, 64]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,7907720
+No: 10 GFLOPS: 46.96/200.02 result: MeasureResult(costs=(0.004929517181818182,), error_no=MeasureErrorNo.NO_ERROR, all_cost=3.243342876434326, timestamp=1668612685.9050357) [('tile_f', [-1, 1, 2, 1]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 16, 8]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,3789950
+No: 11 GFLOPS: 413.73/413.73 result: MeasureResult(costs=(0.0005595513016759776,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.5997123718261719, timestamp=1668612686.585363) [('tile_f', [-1, 8, 4, 1]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 2, 2]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,7978322
+No: 12 GFLOPS: 0.00/413.73 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -1184,8 +1188,8 @@ Traceback (most recent call last):
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
-tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 8, 8, 1]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 4, 64]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9071950
-No: 9 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 8, 16, 1]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 1, 16]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2832097
+No: 13 GFLOPS: 0.00/413.73 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -1307,8 +1311,8 @@ Traceback (most recent call last):
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
-tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 1, 2, 8]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 64, 8]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,10184603
-No: 10 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 32, 4, 4]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 32, 4]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,7831240
+No: 14 GFLOPS: 0.00/413.73 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -1430,8 +1434,8 @@ Traceback (most recent call last):
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
-tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 2, 4, 8]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 2, 32]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2276710
-No: 11 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 1, 2, 64]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 32, 1]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,4861104
+No: 15 GFLOPS: 0.00/413.73 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -1553,8 +1557,8 @@ Traceback (most recent call last):
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
-tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 4, 8, 16]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 32, 4]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 0), ('unroll_explicit', 1)],None,6473901
-No: 12 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 16, 8, 1]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 128, 2]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,7030351
+No: 16 GFLOPS: 0.00/413.73 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -1676,8 +1680,8 @@ Traceback (most recent call last):
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
-tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 8, 32, 2]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 2, 4]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9170793
-No: 13 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 64, 1, 4]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 512, 1]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2939086
+No: 17 GFLOPS: 0.00/413.73 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -1799,8 +1803,8 @@ Traceback (most recent call last):
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
-tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 128, 1, 4]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 16, 4]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,8019767
-No: 14 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 1, 1, 256]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 4, 2]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,10112296
+No: 18 GFLOPS: 0.00/413.73 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -1922,8 +1926,9 @@ Traceback (most recent call last):
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
-tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 8, 16, 2]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 1, 512]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,8708568
-No: 15 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 32, 8, 2]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 64, 4]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,4543744
+No: 19 GFLOPS: 439.07/439.07 result: MeasureResult(costs=(0.000527258671875,), error_no=MeasureErrorNo.NO_ERROR, all_cost=2.0887837409973145, timestamp=1668612689.1139297) [('tile_f', [-1, 1, 2, 8]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 1, 4]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9168643
+No: 20 GFLOPS: 0.00/439.07 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
@@ -2045,622 +2050,7 @@ Traceback (most recent call last):
File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
raise InstantiationError("Skipped because of invalid gpu kernel")
-tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 64, 2, 2]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 8, 8]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2432170
-No: 16 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
- func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
- func = build(s, args, target_host=task.target_host, runtime=runtime)
- File "/workspace/python/tvm/driver/build_module.py", line 227, in build
- input_mod = lower(inputs, args, name=name, binds=binds)
- File "/workspace/python/tvm/driver/build_module.py", line 134, in lower
- return ffi.lower_schedule(inp, args, name, binds, simple_mode)
- File "tvm/_ffi/_cython/./packed_func.pxi", line 331, in tvm._ffi._cy3.core.PackedFuncBase.__call__
- File "tvm/_ffi/_cython/./packed_func.pxi", line 276, in tvm._ffi._cy3.core.FuncCall
- File "tvm/_ffi/_cython/./base.pxi", line 181, in tvm._ffi._cy3.core.CHECK_CALL
-tvm._ffi.base.TVMError: Traceback (most recent call last):
- 24: TVMFuncCall
- at ../src/runtime/c_runtime_api.cc:477
- 23: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 22: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 21: operator()
- at ../include/tvm/runtime/packed_func.h:1731
- 20: unpack_call<tvm::IRModule, 5, tvm::<lambda(tvm::te::Schedule, const tvm::runtime::Array<tvm::runtime::ObjectRef>&, const tvm::runtime::String&, const tvm::runtime::Map<tvm::te::Tensor, tvm::tir::Buffer>&, bool)> >
- at ../include/tvm/runtime/packed_func.h:1671
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1631
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1646
- 13: operator()
- at ../src/driver/driver_api.cc:388
- 12: tvm::LowerSchedule(tvm::te::Schedule, tvm::runtime::Array<tvm::runtime::ObjectRef, void> const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::unordered_map<tvm::te::Tensor, tvm::tir::Buffer, std::hash<tvm::te::Tensor>, std::equal_to<tvm::te::Tensor>, std::allocator<std::pair<tvm::te::Tensor const, tvm::tir::Buffer> > > const&, tvm::GlobalVarSupply, bool)
- at ../src/driver/driver_api.cc:374
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:269
- 10: tvm::transform::Pass::operator()(tvm::IRModule) const
- at ../src/ir/transform.cc:258
- 9: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 8: tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:453
- 7: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 6: tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/tir/ir/transform.cc:100
- 5: tvm::runtime::TypedPackedFunc<tvm::tir::PrimFunc (tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext)>::operator()(tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext) const
- at ../include/tvm/runtime/packed_func.h:1750
- 4: tvm::tir::PrimFunc tvm::runtime::detail::typed_packed_call_dispatcher<tvm::tir::PrimFunc>::run<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::runtime::PackedFunc const&, tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&)
- at ../include/tvm/runtime/packed_func.h:1694
- 3: tvm::runtime::TVMRetValue tvm::runtime::PackedFunc::operator()<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&) const
- at ../include/tvm/runtime/packed_func.h:1618
- 2: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 1: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 0: operator()
- at ../src/runtime/c_runtime_api.cc:534
- File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
- raise InstantiationError("Skipped because of invalid gpu kernel")
-tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel
-
-Traceback (most recent call last):
- 24: TVMFuncCall
- at ../src/runtime/c_runtime_api.cc:477
- 23: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 22: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 21: operator()
- at ../include/tvm/runtime/packed_func.h:1731
- 20: unpack_call<tvm::IRModule, 5, tvm::<lambda(tvm::te::Schedule, const tvm::runtime::Array<tvm::runtime::ObjectRef>&, const tvm::runtime::String&, const tvm::runtime::Map<tvm::te::Tensor, tvm::tir::Buffer>&, bool)> >
- at ../include/tvm/runtime/packed_func.h:1671
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1631
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1646
- 13: operator()
- at ../src/driver/driver_api.cc:388
- 12: tvm::LowerSchedule(tvm::te::Schedule, tvm::runtime::Array<tvm::runtime::ObjectRef, void> const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::unordered_map<tvm::te::Tensor, tvm::tir::Buffer, std::hash<tvm::te::Tensor>, std::equal_to<tvm::te::Tensor>, std::allocator<std::pair<tvm::te::Tensor const, tvm::tir::Buffer> > > const&, tvm::GlobalVarSupply, bool)
- at ../src/driver/driver_api.cc:374
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:269
- 10: tvm::transform::Pass::operator()(tvm::IRModule) const
- at ../src/ir/transform.cc:258
- 9: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 8: tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:453
- 7: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 6: tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/tir/ir/transform.cc:100
- 5: tvm::runtime::TypedPackedFunc<tvm::tir::PrimFunc (tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext)>::operator()(tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext) const
- at ../include/tvm/runtime/packed_func.h:1750
- 4: tvm::tir::PrimFunc tvm::runtime::detail::typed_packed_call_dispatcher<tvm::tir::PrimFunc>::run<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::runtime::PackedFunc const&, tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&)
- at ../include/tvm/runtime/packed_func.h:1694
- 3: tvm::runtime::TVMRetValue tvm::runtime::PackedFunc::operator()<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&) const
- at ../include/tvm/runtime/packed_func.h:1618
- 2: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 1: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 0: operator()
- at ../src/runtime/c_runtime_api.cc:534
- File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
- raise InstantiationError("Skipped because of invalid gpu kernel")
-tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 2, 2, 128]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 8, 8]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,4172514
-No: 17 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
- func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
- func = build(s, args, target_host=task.target_host, runtime=runtime)
- File "/workspace/python/tvm/driver/build_module.py", line 227, in build
- input_mod = lower(inputs, args, name=name, binds=binds)
- File "/workspace/python/tvm/driver/build_module.py", line 134, in lower
- return ffi.lower_schedule(inp, args, name, binds, simple_mode)
- File "tvm/_ffi/_cython/./packed_func.pxi", line 331, in tvm._ffi._cy3.core.PackedFuncBase.__call__
- File "tvm/_ffi/_cython/./packed_func.pxi", line 276, in tvm._ffi._cy3.core.FuncCall
- File "tvm/_ffi/_cython/./base.pxi", line 181, in tvm._ffi._cy3.core.CHECK_CALL
-tvm._ffi.base.TVMError: Traceback (most recent call last):
- 24: TVMFuncCall
- at ../src/runtime/c_runtime_api.cc:477
- 23: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 22: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 21: operator()
- at ../include/tvm/runtime/packed_func.h:1731
- 20: unpack_call<tvm::IRModule, 5, tvm::<lambda(tvm::te::Schedule, const tvm::runtime::Array<tvm::runtime::ObjectRef>&, const tvm::runtime::String&, const tvm::runtime::Map<tvm::te::Tensor, tvm::tir::Buffer>&, bool)> >
- at ../include/tvm/runtime/packed_func.h:1671
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1631
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1646
- 13: operator()
- at ../src/driver/driver_api.cc:388
- 12: tvm::LowerSchedule(tvm::te::Schedule, tvm::runtime::Array<tvm::runtime::ObjectRef, void> const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::unordered_map<tvm::te::Tensor, tvm::tir::Buffer, std::hash<tvm::te::Tensor>, std::equal_to<tvm::te::Tensor>, std::allocator<std::pair<tvm::te::Tensor const, tvm::tir::Buffer> > > const&, tvm::GlobalVarSupply, bool)
- at ../src/driver/driver_api.cc:374
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:269
- 10: tvm::transform::Pass::operator()(tvm::IRModule) const
- at ../src/ir/transform.cc:258
- 9: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 8: tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:453
- 7: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 6: tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/tir/ir/transform.cc:100
- 5: tvm::runtime::TypedPackedFunc<tvm::tir::PrimFunc (tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext)>::operator()(tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext) const
- at ../include/tvm/runtime/packed_func.h:1750
- 4: tvm::tir::PrimFunc tvm::runtime::detail::typed_packed_call_dispatcher<tvm::tir::PrimFunc>::run<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::runtime::PackedFunc const&, tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&)
- at ../include/tvm/runtime/packed_func.h:1694
- 3: tvm::runtime::TVMRetValue tvm::runtime::PackedFunc::operator()<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&) const
- at ../include/tvm/runtime/packed_func.h:1618
- 2: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 1: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 0: operator()
- at ../src/runtime/c_runtime_api.cc:534
- File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
- raise InstantiationError("Skipped because of invalid gpu kernel")
-tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel
-
-Traceback (most recent call last):
- 24: TVMFuncCall
- at ../src/runtime/c_runtime_api.cc:477
- 23: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 22: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 21: operator()
- at ../include/tvm/runtime/packed_func.h:1731
- 20: unpack_call<tvm::IRModule, 5, tvm::<lambda(tvm::te::Schedule, const tvm::runtime::Array<tvm::runtime::ObjectRef>&, const tvm::runtime::String&, const tvm::runtime::Map<tvm::te::Tensor, tvm::tir::Buffer>&, bool)> >
- at ../include/tvm/runtime/packed_func.h:1671
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1631
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1646
- 13: operator()
- at ../src/driver/driver_api.cc:388
- 12: tvm::LowerSchedule(tvm::te::Schedule, tvm::runtime::Array<tvm::runtime::ObjectRef, void> const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::unordered_map<tvm::te::Tensor, tvm::tir::Buffer, std::hash<tvm::te::Tensor>, std::equal_to<tvm::te::Tensor>, std::allocator<std::pair<tvm::te::Tensor const, tvm::tir::Buffer> > > const&, tvm::GlobalVarSupply, bool)
- at ../src/driver/driver_api.cc:374
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:269
- 10: tvm::transform::Pass::operator()(tvm::IRModule) const
- at ../src/ir/transform.cc:258
- 9: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 8: tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:453
- 7: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 6: tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/tir/ir/transform.cc:100
- 5: tvm::runtime::TypedPackedFunc<tvm::tir::PrimFunc (tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext)>::operator()(tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext) const
- at ../include/tvm/runtime/packed_func.h:1750
- 4: tvm::tir::PrimFunc tvm::runtime::detail::typed_packed_call_dispatcher<tvm::tir::PrimFunc>::run<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::runtime::PackedFunc const&, tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&)
- at ../include/tvm/runtime/packed_func.h:1694
- 3: tvm::runtime::TVMRetValue tvm::runtime::PackedFunc::operator()<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&) const
- at ../include/tvm/runtime/packed_func.h:1618
- 2: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 1: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 0: operator()
- at ../src/runtime/c_runtime_api.cc:534
- File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
- raise InstantiationError("Skipped because of invalid gpu kernel")
-tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 16, 32, 1]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 32, 8]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 1)],None,6116044
-No: 18 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
- func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
- func = build(s, args, target_host=task.target_host, runtime=runtime)
- File "/workspace/python/tvm/driver/build_module.py", line 227, in build
- input_mod = lower(inputs, args, name=name, binds=binds)
- File "/workspace/python/tvm/driver/build_module.py", line 134, in lower
- return ffi.lower_schedule(inp, args, name, binds, simple_mode)
- File "tvm/_ffi/_cython/./packed_func.pxi", line 331, in tvm._ffi._cy3.core.PackedFuncBase.__call__
- File "tvm/_ffi/_cython/./packed_func.pxi", line 276, in tvm._ffi._cy3.core.FuncCall
- File "tvm/_ffi/_cython/./base.pxi", line 181, in tvm._ffi._cy3.core.CHECK_CALL
-tvm._ffi.base.TVMError: Traceback (most recent call last):
- 24: TVMFuncCall
- at ../src/runtime/c_runtime_api.cc:477
- 23: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 22: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 21: operator()
- at ../include/tvm/runtime/packed_func.h:1731
- 20: unpack_call<tvm::IRModule, 5, tvm::<lambda(tvm::te::Schedule, const tvm::runtime::Array<tvm::runtime::ObjectRef>&, const tvm::runtime::String&, const tvm::runtime::Map<tvm::te::Tensor, tvm::tir::Buffer>&, bool)> >
- at ../include/tvm/runtime/packed_func.h:1671
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1631
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1646
- 13: operator()
- at ../src/driver/driver_api.cc:388
- 12: tvm::LowerSchedule(tvm::te::Schedule, tvm::runtime::Array<tvm::runtime::ObjectRef, void> const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::unordered_map<tvm::te::Tensor, tvm::tir::Buffer, std::hash<tvm::te::Tensor>, std::equal_to<tvm::te::Tensor>, std::allocator<std::pair<tvm::te::Tensor const, tvm::tir::Buffer> > > const&, tvm::GlobalVarSupply, bool)
- at ../src/driver/driver_api.cc:374
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:269
- 10: tvm::transform::Pass::operator()(tvm::IRModule) const
- at ../src/ir/transform.cc:258
- 9: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 8: tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:453
- 7: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 6: tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/tir/ir/transform.cc:100
- 5: tvm::runtime::TypedPackedFunc<tvm::tir::PrimFunc (tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext)>::operator()(tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext) const
- at ../include/tvm/runtime/packed_func.h:1750
- 4: tvm::tir::PrimFunc tvm::runtime::detail::typed_packed_call_dispatcher<tvm::tir::PrimFunc>::run<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::runtime::PackedFunc const&, tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&)
- at ../include/tvm/runtime/packed_func.h:1694
- 3: tvm::runtime::TVMRetValue tvm::runtime::PackedFunc::operator()<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&) const
- at ../include/tvm/runtime/packed_func.h:1618
- 2: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 1: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 0: operator()
- at ../src/runtime/c_runtime_api.cc:534
- File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
- raise InstantiationError("Skipped because of invalid gpu kernel")
-tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel
-
-Traceback (most recent call last):
- 24: TVMFuncCall
- at ../src/runtime/c_runtime_api.cc:477
- 23: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 22: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 21: operator()
- at ../include/tvm/runtime/packed_func.h:1731
- 20: unpack_call<tvm::IRModule, 5, tvm::<lambda(tvm::te::Schedule, const tvm::runtime::Array<tvm::runtime::ObjectRef>&, const tvm::runtime::String&, const tvm::runtime::Map<tvm::te::Tensor, tvm::tir::Buffer>&, bool)> >
- at ../include/tvm/runtime/packed_func.h:1671
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1631
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1646
- 13: operator()
- at ../src/driver/driver_api.cc:388
- 12: tvm::LowerSchedule(tvm::te::Schedule, tvm::runtime::Array<tvm::runtime::ObjectRef, void> const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::unordered_map<tvm::te::Tensor, tvm::tir::Buffer, std::hash<tvm::te::Tensor>, std::equal_to<tvm::te::Tensor>, std::allocator<std::pair<tvm::te::Tensor const, tvm::tir::Buffer> > > const&, tvm::GlobalVarSupply, bool)
- at ../src/driver/driver_api.cc:374
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:269
- 10: tvm::transform::Pass::operator()(tvm::IRModule) const
- at ../src/ir/transform.cc:258
- 9: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 8: tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:453
- 7: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 6: tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/tir/ir/transform.cc:100
- 5: tvm::runtime::TypedPackedFunc<tvm::tir::PrimFunc (tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext)>::operator()(tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext) const
- at ../include/tvm/runtime/packed_func.h:1750
- 4: tvm::tir::PrimFunc tvm::runtime::detail::typed_packed_call_dispatcher<tvm::tir::PrimFunc>::run<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::runtime::PackedFunc const&, tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&)
- at ../include/tvm/runtime/packed_func.h:1694
- 3: tvm::runtime::TVMRetValue tvm::runtime::PackedFunc::operator()<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&) const
- at ../include/tvm/runtime/packed_func.h:1618
- 2: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 1: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 0: operator()
- at ../src/runtime/c_runtime_api.cc:534
- File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
- raise InstantiationError("Skipped because of invalid gpu kernel")
-tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 16, 1, 4]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 256, 2]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,65444
-No: 19 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
- func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
- func = build(s, args, target_host=task.target_host, runtime=runtime)
- File "/workspace/python/tvm/driver/build_module.py", line 227, in build
- input_mod = lower(inputs, args, name=name, binds=binds)
- File "/workspace/python/tvm/driver/build_module.py", line 134, in lower
- return ffi.lower_schedule(inp, args, name, binds, simple_mode)
- File "tvm/_ffi/_cython/./packed_func.pxi", line 331, in tvm._ffi._cy3.core.PackedFuncBase.__call__
- File "tvm/_ffi/_cython/./packed_func.pxi", line 276, in tvm._ffi._cy3.core.FuncCall
- File "tvm/_ffi/_cython/./base.pxi", line 181, in tvm._ffi._cy3.core.CHECK_CALL
-tvm._ffi.base.TVMError: Traceback (most recent call last):
- 24: TVMFuncCall
- at ../src/runtime/c_runtime_api.cc:477
- 23: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 22: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 21: operator()
- at ../include/tvm/runtime/packed_func.h:1731
- 20: unpack_call<tvm::IRModule, 5, tvm::<lambda(tvm::te::Schedule, const tvm::runtime::Array<tvm::runtime::ObjectRef>&, const tvm::runtime::String&, const tvm::runtime::Map<tvm::te::Tensor, tvm::tir::Buffer>&, bool)> >
- at ../include/tvm/runtime/packed_func.h:1671
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1631
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1646
- 13: operator()
- at ../src/driver/driver_api.cc:388
- 12: tvm::LowerSchedule(tvm::te::Schedule, tvm::runtime::Array<tvm::runtime::ObjectRef, void> const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::unordered_map<tvm::te::Tensor, tvm::tir::Buffer, std::hash<tvm::te::Tensor>, std::equal_to<tvm::te::Tensor>, std::allocator<std::pair<tvm::te::Tensor const, tvm::tir::Buffer> > > const&, tvm::GlobalVarSupply, bool)
- at ../src/driver/driver_api.cc:374
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:269
- 10: tvm::transform::Pass::operator()(tvm::IRModule) const
- at ../src/ir/transform.cc:258
- 9: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 8: tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:453
- 7: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 6: tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/tir/ir/transform.cc:100
- 5: tvm::runtime::TypedPackedFunc<tvm::tir::PrimFunc (tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext)>::operator()(tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext) const
- at ../include/tvm/runtime/packed_func.h:1750
- 4: tvm::tir::PrimFunc tvm::runtime::detail::typed_packed_call_dispatcher<tvm::tir::PrimFunc>::run<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::runtime::PackedFunc const&, tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&)
- at ../include/tvm/runtime/packed_func.h:1694
- 3: tvm::runtime::TVMRetValue tvm::runtime::PackedFunc::operator()<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&) const
- at ../include/tvm/runtime/packed_func.h:1618
- 2: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 1: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 0: operator()
- at ../src/runtime/c_runtime_api.cc:534
- File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
- raise InstantiationError("Skipped because of invalid gpu kernel")
-tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel
-
-Traceback (most recent call last):
- 24: TVMFuncCall
- at ../src/runtime/c_runtime_api.cc:477
- 23: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 22: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 21: operator()
- at ../include/tvm/runtime/packed_func.h:1731
- 20: unpack_call<tvm::IRModule, 5, tvm::<lambda(tvm::te::Schedule, const tvm::runtime::Array<tvm::runtime::ObjectRef>&, const tvm::runtime::String&, const tvm::runtime::Map<tvm::te::Tensor, tvm::tir::Buffer>&, bool)> >
- at ../include/tvm/runtime/packed_func.h:1671
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1631
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1646
- 13: operator()
- at ../src/driver/driver_api.cc:388
- 12: tvm::LowerSchedule(tvm::te::Schedule, tvm::runtime::Array<tvm::runtime::ObjectRef, void> const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::unordered_map<tvm::te::Tensor, tvm::tir::Buffer, std::hash<tvm::te::Tensor>, std::equal_to<tvm::te::Tensor>, std::allocator<std::pair<tvm::te::Tensor const, tvm::tir::Buffer> > > const&, tvm::GlobalVarSupply, bool)
- at ../src/driver/driver_api.cc:374
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:269
- 10: tvm::transform::Pass::operator()(tvm::IRModule) const
- at ../src/ir/transform.cc:258
- 9: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 8: tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:453
- 7: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 6: tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/tir/ir/transform.cc:100
- 5: tvm::runtime::TypedPackedFunc<tvm::tir::PrimFunc (tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext)>::operator()(tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext) const
- at ../include/tvm/runtime/packed_func.h:1750
- 4: tvm::tir::PrimFunc tvm::runtime::detail::typed_packed_call_dispatcher<tvm::tir::PrimFunc>::run<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::runtime::PackedFunc const&, tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&)
- at ../include/tvm/runtime/packed_func.h:1694
- 3: tvm::runtime::TVMRetValue tvm::runtime::PackedFunc::operator()<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&) const
- at ../include/tvm/runtime/packed_func.h:1618
- 2: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 1: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 0: operator()
- at ../src/runtime/c_runtime_api.cc:534
- File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
- raise InstantiationError("Skipped because of invalid gpu kernel")
-tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 2, 4, 16]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 4, 32]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,537636
-No: 20 GFLOPS: 0.00/74.52 result: Traceback (most recent call last):
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 588, in __call__
- func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 540, in _build_func_common
- func = build(s, args, target_host=task.target_host, runtime=runtime)
- File "/workspace/python/tvm/driver/build_module.py", line 227, in build
- input_mod = lower(inputs, args, name=name, binds=binds)
- File "/workspace/python/tvm/driver/build_module.py", line 134, in lower
- return ffi.lower_schedule(inp, args, name, binds, simple_mode)
- File "tvm/_ffi/_cython/./packed_func.pxi", line 331, in tvm._ffi._cy3.core.PackedFuncBase.__call__
- File "tvm/_ffi/_cython/./packed_func.pxi", line 276, in tvm._ffi._cy3.core.FuncCall
- File "tvm/_ffi/_cython/./base.pxi", line 181, in tvm._ffi._cy3.core.CHECK_CALL
-tvm._ffi.base.TVMError: Traceback (most recent call last):
- 24: TVMFuncCall
- at ../src/runtime/c_runtime_api.cc:477
- 23: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 22: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 21: operator()
- at ../include/tvm/runtime/packed_func.h:1731
- 20: unpack_call<tvm::IRModule, 5, tvm::<lambda(tvm::te::Schedule, const tvm::runtime::Array<tvm::runtime::ObjectRef>&, const tvm::runtime::String&, const tvm::runtime::Map<tvm::te::Tensor, tvm::tir::Buffer>&, bool)> >
- at ../include/tvm/runtime/packed_func.h:1671
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1631
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1646
- 13: operator()
- at ../src/driver/driver_api.cc:388
- 12: tvm::LowerSchedule(tvm::te::Schedule, tvm::runtime::Array<tvm::runtime::ObjectRef, void> const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::unordered_map<tvm::te::Tensor, tvm::tir::Buffer, std::hash<tvm::te::Tensor>, std::equal_to<tvm::te::Tensor>, std::allocator<std::pair<tvm::te::Tensor const, tvm::tir::Buffer> > > const&, tvm::GlobalVarSupply, bool)
- at ../src/driver/driver_api.cc:374
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:269
- 10: tvm::transform::Pass::operator()(tvm::IRModule) const
- at ../src/ir/transform.cc:258
- 9: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 8: tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:453
- 7: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 6: tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/tir/ir/transform.cc:100
- 5: tvm::runtime::TypedPackedFunc<tvm::tir::PrimFunc (tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext)>::operator()(tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext) const
- at ../include/tvm/runtime/packed_func.h:1750
- 4: tvm::tir::PrimFunc tvm::runtime::detail::typed_packed_call_dispatcher<tvm::tir::PrimFunc>::run<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::runtime::PackedFunc const&, tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&)
- at ../include/tvm/runtime/packed_func.h:1694
- 3: tvm::runtime::TVMRetValue tvm::runtime::PackedFunc::operator()<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&) const
- at ../include/tvm/runtime/packed_func.h:1618
- 2: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 1: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 0: operator()
- at ../src/runtime/c_runtime_api.cc:534
- File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
- raise InstantiationError("Skipped because of invalid gpu kernel")
-tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel
-
-Traceback (most recent call last):
- 24: TVMFuncCall
- at ../src/runtime/c_runtime_api.cc:477
- 23: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 22: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 21: operator()
- at ../include/tvm/runtime/packed_func.h:1731
- 20: unpack_call<tvm::IRModule, 5, tvm::<lambda(tvm::te::Schedule, const tvm::runtime::Array<tvm::runtime::ObjectRef>&, const tvm::runtime::String&, const tvm::runtime::Map<tvm::te::Tensor, tvm::tir::Buffer>&, bool)> >
- at ../include/tvm/runtime/packed_func.h:1671
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1631
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1631
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1646
- 13: operator()
- at ../src/driver/driver_api.cc:388
- 12: tvm::LowerSchedule(tvm::te::Schedule, tvm::runtime::Array<tvm::runtime::ObjectRef, void> const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::unordered_map<tvm::te::Tensor, tvm::tir::Buffer, std::hash<tvm::te::Tensor>, std::equal_to<tvm::te::Tensor>, std::allocator<std::pair<tvm::te::Tensor const, tvm::tir::Buffer> > > const&, tvm::GlobalVarSupply, bool)
- at ../src/driver/driver_api.cc:374
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:269
- 10: tvm::transform::Pass::operator()(tvm::IRModule) const
- at ../src/ir/transform.cc:258
- 9: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 8: tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:453
- 7: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/ir/transform.cc:274
- 6: tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
- at ../src/tir/ir/transform.cc:100
- 5: tvm::runtime::TypedPackedFunc<tvm::tir::PrimFunc (tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext)>::operator()(tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext) const
- at ../include/tvm/runtime/packed_func.h:1750
- 4: tvm::tir::PrimFunc tvm::runtime::detail::typed_packed_call_dispatcher<tvm::tir::PrimFunc>::run<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::runtime::PackedFunc const&, tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&)
- at ../include/tvm/runtime/packed_func.h:1694
- 3: tvm::runtime::TVMRetValue tvm::runtime::PackedFunc::operator()<tvm::tir::PrimFunc, tvm::IRModule, tvm::transform::PassContext>(tvm::tir::PrimFunc&&, tvm::IRModule&&, tvm::transform::PassContext&&) const
- at ../include/tvm/runtime/packed_func.h:1618
- 2: tvm::runtime::PackedFuncObj::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
- at ../include/tvm/runtime/packed_func.h:1217
- 1: Call
- at ../include/tvm/runtime/packed_func.h:1213
- 0: operator()
- at ../src/runtime/c_runtime_api.cc:534
- File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 871, in verify_pass
- raise InstantiationError("Skipped because of invalid gpu kernel")
-tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 4, 16, 4]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 1, 16]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,8446368
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 2, 128, 1]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 8, 32]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,10412870
</pre></div>
</div>
<p>Finally we can inspect the best config from log file, check correctness,
@@ -2699,9 +2089,9 @@ and measure running time.</p>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Finish loading 20 records
Best config:
-[('tile_f', [-1, 8, 4, 4]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 4, 4]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,1818638
+[('tile_f', [-1, 1, 2, 8]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 1, 4]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9168643
Finish loading 20 records
-Time cost of this operator: 0.001609
+Time cost of this operator: 0.000886
</pre></div>
</div>
<div class="sphx-glr-footer sphx-glr-footer-example docutils container" id="sphx-glr-download-how-to-tune-with-autotvm-tune-conv2d-cuda-py">
diff --git a/docs/how_to/work_with_microtvm/micro_autotune.html b/docs/how_to/work_with_microtvm/micro_autotune.html
index 7f0d8fd07a..b6a2141034 100644
--- a/docs/how_to/work_with_microtvm/micro_autotune.html
+++ b/docs/how_to/work_with_microtvm/micro_autotune.html
@@ -596,10 +596,10 @@ the tuned operator.</p>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>########## Build without Autotuning ##########
Node Name Ops Time(us) Time(%) Shape Inputs Outputs Measurements(us)
--------- --- -------- ------- ----- ------ ------- ----------------
-tvmgen_default_fused_nn_contrib_conv2d_NCHWc tvmgen_default_fused_nn_contrib_conv2d_NCHWc 311.4 98.717 (1, 2, 10, 10, 3) 2 1 [311.4]
-tvmgen_default_fused_layout_transform_1 tvmgen_default_fused_layout_transform_1 3.084 0.978 (1, 6, 10, 10) 1 1 [3.084]
-tvmgen_default_fused_layout_transform tvmgen_default_fused_layout_transform 0.963 0.305 (1, 1, 10, 10, 3) 1 1 [0.963]
-Total_time - 315.448 - - - - -
+tvmgen_default_fused_nn_contrib_conv2d_NCHWc tvmgen_default_fused_nn_contrib_conv2d_NCHWc 311.8 98.714 (1, 2, 10, 10, 3) 2 1 [311.8]
+tvmgen_default_fused_layout_transform_1 tvmgen_default_fused_layout_transform_1 3.102 0.982 (1, 6, 10, 10) 1 1 [3.102]
+tvmgen_default_fused_layout_transform tvmgen_default_fused_layout_transform 0.962 0.305 (1, 1, 10, 10, 3) 1 1 [0.962]
+Total_time - 315.863 - - - - -
</pre></div>
</div>
</div>
@@ -650,10 +650,10 @@ Total_time -
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>########## Build with Autotuning ##########
Node Name Ops Time(us) Time(%) Shape Inputs Outputs Measurements(us)
--------- --- -------- ------- ----- ------ ------- ----------------
-tvmgen_default_fused_nn_contrib_conv2d_NCHWc tvmgen_default_fused_nn_contrib_conv2d_NCHWc 100.2 97.298 (1, 6, 10, 10, 1) 2 1 [100.2]
-tvmgen_default_fused_layout_transform_1 tvmgen_default_fused_layout_transform_1 1.807 1.755 (1, 6, 10, 10) 1 1 [1.807]
-tvmgen_default_fused_layout_transform tvmgen_default_fused_layout_transform 0.976 0.947 (1, 1, 10, 10, 3) 1 1 [0.976]
-Total_time - 102.983 - - - - -
+tvmgen_default_fused_nn_contrib_conv2d_NCHWc tvmgen_default_fused_nn_contrib_conv2d_NCHWc 106.6 97.612 (1, 6, 10, 10, 1) 2 1 [106.6]
+tvmgen_default_fused_layout_transform_1 tvmgen_default_fused_layout_transform_1 1.77 1.621 (1, 6, 10, 10) 1 1 [1.77]
+tvmgen_default_fused_layout_transform tvmgen_default_fused_layout_transform 0.838 0.767 (1, 3, 10, 10, 1) 1 1 [0.838]
+Total_time - 109.208 - - - - -
</pre></div>
</div>
<div class="sphx-glr-footer sphx-glr-footer-example docutils container" id="sphx-glr-download-how-to-work-with-microtvm-micro-autotune-py">
diff --git a/docs/how_to/work_with_microtvm/micro_pytorch.html b/docs/how_to/work_with_microtvm/micro_pytorch.html
index b082e71b71..213b685950 100644
--- a/docs/how_to/work_with_microtvm/micro_pytorch.html
+++ b/docs/how_to/work_with_microtvm/micro_pytorch.html
@@ -440,7 +440,7 @@ download a cat image and preprocess it to use as the model input.</p>
Downloading: "https://download.pytorch.org/models/quantized/mobilenet_v2_qnnpack_37f702c5.pth" to /workspace/.cache/torch/hub/checkpoints/mobilenet_v2_qnnpack_37f702c5.pth
0%| | 0.00/3.42M [00:00<?, ?B/s]
-100%|##########| 3.42M/3.42M [00:00<00:00, 83.1MB/s]
+100%|##########| 3.42M/3.42M [00:00<00:00, 81.3MB/s]
/workspace/python/tvm/relay/frontend/pytorch_utils.py:47: DeprecationWarning: distutils Version classes are deprecated. Use packaging.version instead.
return LooseVersion(torch_ver) > ver
/venv/apache-tvm-py3.7/lib/python3.7/site-packages/setuptools/_distutils/version.py:346: DeprecationWarning: distutils Version classes are deprecated. Use packaging.version instead.
@@ -564,7 +564,7 @@ via the host <cite>main.cc`</cite> or if a Zephyr emulated board is selected as
Torch top-1 id: 282, class name: tiger cat
</pre></div>
</div>
-<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 1 minutes 4.172 seconds)</p>
+<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 1 minutes 6.472 seconds)</p>
<div class="sphx-glr-footer sphx-glr-footer-example docutils container" id="sphx-glr-download-how-to-work-with-microtvm-micro-pytorch-py">
<div class="sphx-glr-download sphx-glr-download-python docutils container">
<p><a class="reference download internal" download="" href="../../_downloads/12b9ecc04c41abaa12022061771821d1/micro_pytorch.py"><code class="xref download docutils literal notranslate"><span class="pre">Download</span> <span class="pre">Python</span> <span class="pre">source</span> <span class="pre">code:</span> <span class="pre">micro_pytorch.py</span></code></a></p>
diff --git a/docs/how_to/work_with_microtvm/micro_train.html b/docs/how_to/work_with_microtvm/micro_train.html
index a24596ff67..fe8e82864b 100644
--- a/docs/how_to/work_with_microtvm/micro_train.html
+++ b/docs/how_to/work_with_microtvm/micro_train.html
@@ -530,7 +530,7 @@ take about <strong>2 minutes</strong> to download the Stanford Cars, while COCO
<a href="https://docs.python.org/3/library/shutil.html#shutil.move" title="shutil.move" class="sphx-glr-backref-module-shutil sphx-glr-backref-type-py-function"><span class="n">shutil</span><span class="o">.</span><span class="n">move</span></a><span class="p">(</span><span class="sa">f</span><span class="s2">"</span><span class="si">{</span><a href="https://docs.python.org/3/library/stdtypes.html#str" title="builtins.str" class="sphx-glr-backref-module-builtins sphx-glr-backref-typ [...]
</pre></div>
</div>
-<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>'/tmp/tmp19qchhde/images/random'
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>'/tmp/tmpy5o6f33_/images/random'
</pre></div>
</div>
</div>
@@ -590,8 +590,8 @@ objects to other stuff? We can display some examples from our datasets using <co
<span class="n">plt</span><span class="o">.</span><span class="n">axis</span><span class="p">(</span><span class="s2">"off"</span><span class="p">)</span>
</pre></div>
</div>
-<img src="../../_images/sphx_glr_micro_train_001.png" srcset="../../_images/sphx_glr_micro_train_001.png" alt="[1.0, 0.0], [0.0, 1.0], [1.0, 0.0], [0.0, 1.0], [1.0, 0.0], [0.0, 1.0], [0.0, 1.0], [1.0, 0.0], [0.0, 1.0], [0.0, 1.0]" class = "sphx-glr-single-img"/><div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>/tmp/tmp19qchhde/images/target contains 8144 images
-/tmp/tmp19qchhde/images/random contains 5000 images
+<img src="../../_images/sphx_glr_micro_train_001.png" srcset="../../_images/sphx_glr_micro_train_001.png" alt="[0.0, 1.0], [1.0, 0.0], [1.0, 0.0], [0.0, 1.0], [0.0, 1.0], [1.0, 0.0], [1.0, 0.0], [0.0, 1.0], [0.0, 1.0], [0.0, 1.0]" class = "sphx-glr-single-img"/><div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>/tmp/tmpy5o6f33_/images/target contains 8144 images
+/tmp/tmpy5o6f33_/images/random contains 5000 images
</pre></div>
</div>
</div>
@@ -703,13 +703,13 @@ the time on our validation set).</p>
</pre></div>
</div>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Epoch 1/3
-328/328 - 47s - loss: 0.2178 - accuracy: 0.9269 - val_loss: 0.1431 - val_accuracy: 0.9528 - 47s/epoch - 144ms/step
+328/328 - 47s - loss: 0.2249 - accuracy: 0.9236 - val_loss: 0.1049 - val_accuracy: 0.9603 - 47s/epoch - 144ms/step
Epoch 2/3
-328/328 - 43s - loss: 0.1014 - accuracy: 0.9611 - val_loss: 0.1025 - val_accuracy: 0.9656 - 43s/epoch - 132ms/step
+328/328 - 43s - loss: 0.1043 - accuracy: 0.9625 - val_loss: 0.1249 - val_accuracy: 0.9596 - 43s/epoch - 133ms/step
Epoch 3/3
-328/328 - 43s - loss: 0.0720 - accuracy: 0.9741 - val_loss: 0.1185 - val_accuracy: 0.9619 - 43s/epoch - 132ms/step
+328/328 - 43s - loss: 0.0645 - accuracy: 0.9761 - val_loss: 0.1150 - val_accuracy: 0.9656 - 43s/epoch - 132ms/step
-<keras.callbacks.History object at 0x7f315a9d0890>
+<keras.callbacks.History object at 0x7f09494f4550>
</pre></div>
</div>
</div>
@@ -971,7 +971,7 @@ as intended.</p>
<p>From here, we could modify the model to read live images from the camera - we have another
Arduino tutorial for how to do that <a class="reference external" href="https://github.com/guberti/tvm-arduino-demos/tree/master/examples/person_detection">on GitHub</a>. Alternatively, we could also
<a class="reference external" href="https://tvm.apache.org/docs/how_to/work_with_microtvm/micro_autotune.html">use TVM’s autotuning capabilities</a> to dramatically improve the model’s performance.</p>
-<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 4 minutes 52.658 seconds)</p>
+<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 4 minutes 38.542 seconds)</p>
<div class="sphx-glr-footer sphx-glr-footer-example docutils container" id="sphx-glr-download-how-to-work-with-microtvm-micro-train-py">
<div class="sphx-glr-download sphx-glr-download-python docutils container">
<p><a class="reference download internal" download="" href="../../_downloads/b52cec46baf4f78d6bcd94cbe269c8a6/micro_train.py"><code class="xref download docutils literal notranslate"><span class="pre">Download</span> <span class="pre">Python</span> <span class="pre">source</span> <span class="pre">code:</span> <span class="pre">micro_train.py</span></code></a></p>
diff --git a/docs/how_to/work_with_microtvm/sg_execution_times.html b/docs/how_to/work_with_microtvm/sg_execution_times.html
index 4081d11e8e..1518cb0ff7 100644
--- a/docs/how_to/work_with_microtvm/sg_execution_times.html
+++ b/docs/how_to/work_with_microtvm/sg_execution_times.html
@@ -340,7 +340,7 @@
<div class="section" id="computation-times">
<span id="sphx-glr-how-to-work-with-microtvm-sg-execution-times"></span><h1>Computation times<a class="headerlink" href="#computation-times" title="Permalink to this headline">¶</a></h1>
-<p><strong>06:59.588</strong> total execution time for <strong>how_to_work_with_microtvm</strong> files:</p>
+<p><strong>06:48.119</strong> total execution time for <strong>how_to_work_with_microtvm</strong> files:</p>
<table class="docutils align-default">
<colgroup>
<col style="width: 83%" />
@@ -349,23 +349,23 @@
</colgroup>
<tbody>
<tr class="row-odd"><td><p><a class="reference internal" href="micro_train.html#sphx-glr-how-to-work-with-microtvm-micro-train-py"><span class="std std-ref">Training Vision Models for microTVM on Arduino</span></a> (<code class="docutils literal notranslate"><span class="pre">micro_train.py</span></code>)</p></td>
-<td><p>04:52.658</p></td>
+<td><p>04:38.542</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="micro_pytorch.html#sphx-glr-how-to-work-with-microtvm-micro-pytorch-py"><span class="std std-ref">microTVM PyTorch Tutorial</span></a> (<code class="docutils literal notranslate"><span class="pre">micro_pytorch.py</span></code>)</p></td>
-<td><p>01:04.172</p></td>
+<td><p>01:06.472</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="micro_autotune.html#sphx-glr-how-to-work-with-microtvm-micro-autotune-py"><span class="std std-ref">Autotuning with microTVM</span></a> (<code class="docutils literal notranslate"><span class="pre">micro_autotune.py</span></code>)</p></td>
-<td><p>00:50.525</p></td>
+<td><p>00:50.409</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="micro_aot.html#sphx-glr-how-to-work-with-microtvm-micro-aot-py"><span class="std std-ref">microTVM Host-Driven AoT</span></a> (<code class="docutils literal notranslate"><span class="pre">micro_aot.py</span></code>)</p></td>
-<td><p>00:08.438</p></td>
+<td><p>00:08.910</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="micro_tflite.html#sphx-glr-how-to-work-with-microtvm-micro-tflite-py"><span class="std std-ref">microTVM with TFLite Models</span></a> (<code class="docutils literal notranslate"><span class="pre">micro_tflite.py</span></code>)</p></td>
-<td><p>00:03.794</p></td>
+<td><p>00:03.784</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="micro_reference_vm.html#sphx-glr-how-to-work-with-microtvm-micro-reference-vm-py"><span class="std std-ref">microTVM Reference Virtual Machines</span></a> (<code class="docutils literal notranslate"><span class="pre">micro_reference_vm.py</span></code>)</p></td>
diff --git a/docs/how_to/work_with_relay/sg_execution_times.html b/docs/how_to/work_with_relay/sg_execution_times.html
index 34401aec3c..4df9fd21dd 100644
--- a/docs/how_to/work_with_relay/sg_execution_times.html
+++ b/docs/how_to/work_with_relay/sg_execution_times.html
@@ -340,7 +340,7 @@
<div class="section" id="computation-times">
<span id="sphx-glr-how-to-work-with-relay-sg-execution-times"></span><h1>Computation times<a class="headerlink" href="#computation-times" title="Permalink to this headline">¶</a></h1>
-<p><strong>00:43.405</strong> total execution time for <strong>how_to_work_with_relay</strong> files:</p>
+<p><strong>00:44.620</strong> total execution time for <strong>how_to_work_with_relay</strong> files:</p>
<table class="docutils align-default">
<colgroup>
<col style="width: 84%" />
@@ -349,15 +349,15 @@
</colgroup>
<tbody>
<tr class="row-odd"><td><p><a class="reference internal" href="using_pipeline_executor.html#sphx-glr-how-to-work-with-relay-using-pipeline-executor-py"><span class="std std-ref">Using Pipeline Executor in Relay</span></a> (<code class="docutils literal notranslate"><span class="pre">using_pipeline_executor.py</span></code>)</p></td>
-<td><p>00:31.889</p></td>
+<td><p>00:32.524</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="using_external_lib.html#sphx-glr-how-to-work-with-relay-using-external-lib-py"><span class="std std-ref">Using External Libraries in Relay</span></a> (<code class="docutils literal notranslate"><span class="pre">using_external_lib.py</span></code>)</p></td>
-<td><p>00:09.793</p></td>
+<td><p>00:10.560</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="build_gcn.html#sphx-glr-how-to-work-with-relay-build-gcn-py"><span class="std std-ref">Building a Graph Convolutional Network</span></a> (<code class="docutils literal notranslate"><span class="pre">build_gcn.py</span></code>)</p></td>
-<td><p>00:01.717</p></td>
+<td><p>00:01.529</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="using_relay_viz.html#sphx-glr-how-to-work-with-relay-using-relay-viz-py"><span class="std std-ref">Use Relay Visualizer to Visualize Relay</span></a> (<code class="docutils literal notranslate"><span class="pre">using_relay_viz.py</span></code>)</p></td>
diff --git a/docs/how_to/work_with_schedules/intrin_math.html b/docs/how_to/work_with_schedules/intrin_math.html
index b09af578ed..49e5bd0f55 100644
--- a/docs/how_to/work_with_schedules/intrin_math.html
+++ b/docs/how_to/work_with_schedules/intrin_math.html
@@ -535,7 +535,7 @@ The following example customizes CUDA lowering rule for <code class="code docuti
<a href="../../reference/api/python/ir.html#tvm.ir.register_intrin_lowering" title="tvm.ir.register_intrin_lowering" class="sphx-glr-backref-module-tvm-ir sphx-glr-backref-type-py-function"><span class="n">register_intrin_lowering</span></a><span class="p">(</span><span class="s2">"tir.exp"</span><span class="p">,</span> <span class="n">target</span><span class="o">=</span><span class="s2">"cuda"</span><span class="p">,</span> <span class="n">f</span><span class="o">= [...]
</pre></div>
</div>
-<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span><function my_cuda_math_rule at 0x7f315f6bb200>
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span><function my_cuda_math_rule at 0x7f0949fdbdd0>
</pre></div>
</div>
<p>Register the rule to TVM with override option to override existing rule.
diff --git a/docs/how_to/work_with_schedules/reduction.html b/docs/how_to/work_with_schedules/reduction.html
index e59c37ae01..c3478f0ec1 100644
--- a/docs/how_to/work_with_schedules/reduction.html
+++ b/docs/how_to/work_with_schedules/reduction.html
@@ -434,14 +434,13 @@ Before doing anything, let us print out the IR code of default schedule.</p>
</div>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@main = primfn(A_1: handle, B_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
- buffers = {A: Buffer(A_2: Pointer(float32), float32, [(stride: int32*n: int32)], [], type="auto"),
- B: Buffer(B_2: Pointer(float32), float32, [(stride_1: int32*n)], [], type="auto")}
- buffer_map = {A_1: A, B_1: B}
- preflattened_buffer_map = {A_1: A_3: Buffer(A_2, float32, [n, m: int32], [stride, stride_2: int32], type="auto"), B_1: B_3: Buffer(B_2, float32, [n], [stride_1], type="auto")} {
+ buffers = {A: Buffer(A_2: Pointer(float32), float32, [n: int32, m: int32], [stride: int32, stride_1: int32], type="auto"),
+ B: Buffer(B_2: Pointer(float32), float32, [n], [stride_2: int32], type="auto")}
+ buffer_map = {A_1: A, B_1: B} {
for (i: int32, 0, n) {
- B[(i*stride_1)] = 0f32
+ B_3: Buffer(B_2, float32, [(stride_2*n)], [], type="auto")[(i*stride_2)] = 0f32
for (k: int32, 0, m) {
- B[(i*stride_1)] = (B[(i*stride_1)] + A[((i*stride) + (k*stride_2))])
+ B_3[(i*stride_2)] = (B_3[(i*stride_2)] + A_3: Buffer(A_2, float32, [(stride*n)], [], type="auto")[((i*stride) + (k*stride_1))])
}
}
}
@@ -458,21 +457,20 @@ axis by different factors. The result is a nested reduction.</p>
</div>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@main = primfn(A_1: handle, B_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True}
... 7965 lines suppressed ...