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 2023/01/20 03:59:40 UTC
[tvm-site] branch asf-site updated: deploying docs (apache/tvm@6c2d485a011fbfbd426353c6fc1254f3385d826e)
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 a4b4639943 deploying docs (apache/tvm@6c2d485a011fbfbd426353c6fc1254f3385d826e)
a4b4639943 is described below
commit a4b463994377f7dc06a24a57957baba139bbf2ff
Author: tvm-bot <95...@users.noreply.github.com>
AuthorDate: Fri Jan 20 03:59:34 2023 +0000
deploying docs (apache/tvm@6c2d485a011fbfbd426353c6fc1254f3385d826e)
---
docs/_images/sphx_glr_micro_train_001.png | Bin 329141 -> 282079 bytes
docs/_images/sphx_glr_micro_train_thumb.png | Bin 22792 -> 22031 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_adreno.rst.txt | 2 +-
.../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 | 20 +-
.../extend_tvm/bring_your_own_datatypes.rst.txt | 2 +-
.../extend_tvm/low_level_custom_pass.rst.txt | 26 +-
.../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 | 34 +-
.../how_to/optimize_operators/opt_gemm.rst.txt | 139 +--
.../optimize_operators/sg_execution_times.rst.txt | 8 +-
.../sg_execution_times.rst.txt | 14 +-
.../tune_conv2d_layer_cuda.rst.txt | 1175 +++++++++++++++-----
.../tune_network_cuda.rst.txt | 4 +-
.../tune_network_x86.rst.txt | 4 +-
.../tune_sparse_x86.rst.txt | 473 ++++++--
.../tune_with_autotvm/sg_execution_times.rst.txt | 6 +-
.../tune_with_autotvm/tune_conv2d_cuda.rst.txt | 335 +-----
.../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 | 16 +-
.../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 | 7 +-
.../work_with_schedules/sg_execution_times.rst.txt | 14 +-
.../how_to/work_with_schedules/tensorize.rst.txt | 46 +-
.../tutorials/autotvm/sg_execution_times.rst.txt | 6 +-
.../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 | 45 +-
.../vta/tutorials/optimize/convolution_opt.rst.txt | 51 +-
.../tutorials/optimize/matrix_multiply_opt.rst.txt | 45 +-
.../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 | 28 +-
.../tutorial/auto_scheduler_matmul_x86.rst.txt | 37 +-
docs/_sources/tutorial/autotvm_matmul_x86.rst.txt | 20 +-
docs/_sources/tutorial/autotvm_relay_x86.rst.txt | 54 +-
.../tutorial/cross_compilation_and_rpc.rst.txt | 2 +-
docs/_sources/tutorial/intro_topi.rst.txt | 109 +-
docs/_sources/tutorial/sg_execution_times.rst.txt | 22 +-
.../tutorial/tensor_expr_get_started.rst.txt | 168 ++-
.../tutorial/tensor_ir_blitz_course.rst.txt | 20 +-
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 | 14 +-
docs/how_to/compile_models/from_pytorch.html | 8 +-
docs/how_to/compile_models/from_tensorflow.html | 2 +-
docs/how_to/compile_models/sg_execution_times.html | 22 +-
.../deploy_models/deploy_model_on_adreno.html | 2 +-
.../deploy_models/deploy_model_on_android.html | 2 +-
.../deploy_object_detection_pytorch.html | 38 +-
docs/how_to/deploy_models/deploy_prequantized.html | 9 +-
.../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 | 20 +-
.../extend_tvm/bring_your_own_datatypes.html | 2 +-
docs/how_to/extend_tvm/low_level_custom_pass.html | 26 +-
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 | 34 +-
docs/how_to/optimize_operators/opt_gemm.html | 139 +--
.../optimize_operators/sg_execution_times.html | 8 +-
.../sg_execution_times.html | 14 +-
.../tune_conv2d_layer_cuda.html | 1175 +++++++++++++++-----
.../tune_with_autoscheduler/tune_network_cuda.html | 4 +-
.../tune_with_autoscheduler/tune_network_x86.html | 4 +-
.../tune_with_autoscheduler/tune_sparse_x86.html | 473 ++++++--
.../tune_with_autotvm/sg_execution_times.html | 6 +-
.../how_to/tune_with_autotvm/tune_conv2d_cuda.html | 335 +-----
docs/how_to/work_with_microtvm/micro_autotune.html | 16 +-
docs/how_to/work_with_microtvm/micro_pytorch.html | 5 +-
docs/how_to/work_with_microtvm/micro_train.html | 16 +-
.../work_with_microtvm/sg_execution_times.html | 16 +-
.../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 | 7 +-
.../work_with_schedules/sg_execution_times.html | 14 +-
docs/how_to/work_with_schedules/tensorize.html | 46 +-
docs/install/nnpack.html | 12 +-
docs/reference/api/python/auto_scheduler.html | 4 +-
.../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 | 6 +-
.../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 | 45 +-
.../vta/tutorials/optimize/convolution_opt.html | 51 +-
.../tutorials/optimize/matrix_multiply_opt.html | 45 +-
.../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 | 28 +-
docs/tutorial/auto_scheduler_matmul_x86.html | 33 +-
docs/tutorial/autotvm_matmul_x86.html | 20 +-
docs/tutorial/autotvm_relay_x86.html | 264 ++---
docs/tutorial/cross_compilation_and_rpc.html | 2 +-
docs/tutorial/intro_topi.html | 109 +-
docs/tutorial/sg_execution_times.html | 22 +-
docs/tutorial/tensor_expr_get_started.html | 164 ++-
docs/tutorial/tensor_ir_blitz_course.html | 20 +-
144 files changed, 4033 insertions(+), 2899 deletions(-)
diff --git a/docs/_images/sphx_glr_micro_train_001.png b/docs/_images/sphx_glr_micro_train_001.png
index 9c86215278..d501408c74 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 839a4f5975..12aef69638 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 520dc78408..dbddc1d263 100644
--- a/docs/_sources/how_to/compile_models/from_darknet.rst.txt
+++ b/docs/_sources/how_to/compile_models/from_darknet.rst.txt
@@ -318,7 +318,7 @@ The process is no different from other examples.
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 1 minutes 16.257 seconds)
+ **Total running time of the script:** ( 1 minutes 13.727 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 afdcda9a9c..bd9874d4d7 100644
--- a/docs/_sources/how_to/compile_models/from_keras.rst.txt
+++ b/docs/_sources/how_to/compile_models/from_keras.rst.txt
@@ -232,7 +232,7 @@ 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 958ms/step
+
1/1 [==============================] - ETA: 0s
1/1 [==============================] - 1s 938ms/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 6a0580eee4..16b3bea35b 100644
--- a/docs/_sources/how_to/compile_models/from_mxnet.rst.txt
+++ b/docs/_sources/how_to/compile_models/from_mxnet.rst.txt
@@ -116,7 +116,7 @@ In this section, we download a pretrained imagenet model and classify an image.
.. code-block:: none
- Downloading /workspace/.mxnet/models/resnet18_v1-a0666292.zip74b370c8-6b03-402b-8791-7ba7f0d5f123 from https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/models/resnet18_v1-a0666292.zip...
+ Downloading /workspace/.mxnet/models/resnet18_v1-a0666292.zip281edc03-6311-457c-b054-e2f5809f278a 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 96c7e6a257..0e48ba6be1 100644
--- a/docs/_sources/how_to/compile_models/from_oneflow.rst.txt
+++ b/docs/_sources/how_to/compile_models/from_oneflow.rst.txt
@@ -121,7 +121,7 @@ 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, 51.7MB/s]
39%|###8 | 16.0M/41.5M [00:00<00:00, 48.4MB/s]
58%|#####7 | 24.0M/41.5M [00:00<00:00, 45.0MB/s]
77%|#######7 | 32.0M/41.5M [00:00<00:00, 47.3MB/s]
90%|########9 | 37.2M/41.5M [00:00<00:00, 44.8MB/s]
100%|#########9| 41.4M/41.5M [00:00<00:00, 43.5MB/s]
100%|##########| 41.5M/41.5M [00:00<00:00, 45.3MB/s]
+
0%| | 0.00/41.5M [00:00<?, ?B/s]
15%|#5 | 6.33M/41.5M [00:00<00:00, 51.0MB/s]
33%|###3 | 13.9M/41.5M [00:00<00:00, 65.7MB/s]
49%|####8 | 20.3M/41.5M [00:00<00:00, 51.4MB/s]
62%|######1 | 25.6M/41.5M [00:00<00:00, 50.1MB/s]
77%|#######7 | 32.0M/41.5M [00:00<00:00, 52.8MB/s]
92%|#########2| 38.3M/41.5M [00:00<00:00, 45.2MB/s]
100%|##########| 41.5M/41.5M [00:00<00:00, 49.6MB/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 686e7a5c10..bb6b0ff25e 100644
--- a/docs/_sources/how_to/compile_models/from_pytorch.rst.txt
+++ b/docs/_sources/how_to/compile_models/from_pytorch.rst.txt
@@ -101,7 +101,7 @@ 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]
24%|##3 | 10.6M/44.7M [00:00<00:00, 112MB/s]
51%|#####1 | 22.9M/44.7M [00:00<00:00, 122MB/s]
77%|#######7 | 34.5M/44.7M [00:00<00:00, 108MB/s]
100%|##########| 44.7M/44.7M [00:00<00:00, 108MB/s]
+
0%| | 0.00/44.7M [00:00<?, ?B/s]
23%|##2 | 10.1M/44.7M [00:00<00:00, 74.9MB/s]
50%|##### | 22.4M/44.7M [00:00<00:00, 102MB/s]
73%|#######3 | 32.6M/44.7M [00:00<00:00, 87.4MB/s]
100%|##########| 44.7M/44.7M [00:00<00:00, 98.9MB/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 8769ede750..1f6ecbe67a 100644
--- a/docs/_sources/how_to/compile_models/from_tensorflow.rst.txt
+++ b/docs/_sources/how_to/compile_models/from_tensorflow.rst.txt
@@ -424,7 +424,7 @@ Run the corresponding model on tensorflow
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 1 minutes 20.531 seconds)
+ **Total running time of the script:** ( 1 minutes 18.160 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 e16f971ac7..2bcd8f169b 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
=================
-**06:19.340** total execution time for **how_to_compile_models** files:
+**06:07.945** total execution time for **how_to_compile_models** files:
+-----------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_compile_models_from_tensorflow.py` (``from_tensorflow.py``) | 01:20.531 | 0.0 MB |
+| :ref:`sphx_glr_how_to_compile_models_from_tensorflow.py` (``from_tensorflow.py``) | 01:18.160 | 0.0 MB |
+-----------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_compile_models_from_darknet.py` (``from_darknet.py``) | 01:16.257 | 0.0 MB |
+| :ref:`sphx_glr_how_to_compile_models_from_darknet.py` (``from_darknet.py``) | 01:13.727 | 0.0 MB |
+-----------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_compile_models_from_paddle.py` (``from_paddle.py``) | 00:52.501 | 0.0 MB |
+| :ref:`sphx_glr_how_to_compile_models_from_paddle.py` (``from_paddle.py``) | 00:50.305 | 0.0 MB |
+-----------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_compile_models_from_oneflow.py` (``from_oneflow.py``) | 00:35.465 | 0.0 MB |
+| :ref:`sphx_glr_how_to_compile_models_from_oneflow.py` (``from_oneflow.py``) | 00:34.517 | 0.0 MB |
+-----------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_compile_models_from_mxnet.py` (``from_mxnet.py``) | 00:30.357 | 0.0 MB |
+| :ref:`sphx_glr_how_to_compile_models_from_mxnet.py` (``from_mxnet.py``) | 00:29.818 | 0.0 MB |
+-----------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_compile_models_from_coreml.py` (``from_coreml.py``) | 00:30.149 | 0.0 MB |
+| :ref:`sphx_glr_how_to_compile_models_from_coreml.py` (``from_coreml.py``) | 00:29.515 | 0.0 MB |
+-----------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_compile_models_from_tflite.py` (``from_tflite.py``) | 00:26.165 | 0.0 MB |
+| :ref:`sphx_glr_how_to_compile_models_from_tflite.py` (``from_tflite.py``) | 00:26.227 | 0.0 MB |
+-----------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_compile_models_from_pytorch.py` (``from_pytorch.py``) | 00:24.732 | 0.0 MB |
+| :ref:`sphx_glr_how_to_compile_models_from_pytorch.py` (``from_pytorch.py``) | 00:23.431 | 0.0 MB |
+-----------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_compile_models_from_keras.py` (``from_keras.py``) | 00:20.531 | 0.0 MB |
+| :ref:`sphx_glr_how_to_compile_models_from_keras.py` (``from_keras.py``) | 00:19.695 | 0.0 MB |
+-----------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_compile_models_from_onnx.py` (``from_onnx.py``) | 00:02.651 | 0.0 MB |
+| :ref:`sphx_glr_how_to_compile_models_from_onnx.py` (``from_onnx.py``) | 00:02.550 | 0.0 MB |
+-----------------------------------------------------------------------------------+-----------+--------+
diff --git a/docs/_sources/how_to/deploy_models/deploy_model_on_adreno.rst.txt b/docs/_sources/how_to/deploy_models/deploy_model_on_adreno.rst.txt
index 32d6a1720c..caa6e47528 100644
--- a/docs/_sources/how_to/deploy_models/deploy_model_on_adreno.rst.txt
+++ b/docs/_sources/how_to/deploy_models/deploy_model_on_adreno.rst.txt
@@ -727,7 +727,7 @@ well as provides information about the model's performance
Evaluate inference time cost...
Execution time summary:
mean (ms) median (ms) max (ms) min (ms) std (ms)
- 2550.8269 2549.1817 2566.5008 2545.9899 5.4542
+ 2758.3525 2758.4177 2760.3811 2755.1281 1.6989
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 8aae35985b..6d921eceb2 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
@@ -437,7 +437,7 @@ Execute on TVM
Evaluate inference time cost...
Execution time summary:
mean (ms) median (ms) max (ms) min (ms) std (ms)
- 16.0102 15.8558 16.7513 15.6779 0.3372
+ 15.6184 15.5163 16.1542 15.4111 0.2631
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 0d9e58e5d8..29723b06b7 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
@@ -130,7 +130,7 @@ 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]
5%|4 | 7.99M/170M [00:00<00:02, 74.9MB/s]
9%|8 | 15.1M/170M [00:00<00:02, 59.8MB/s]
13%|#3 | 22.9M/170M [00:00<00:02, 68.1MB/s]
18%|#7 | 30.3M/170M [00:00<00:02, 53.7MB/s]
21%|##1 | 35.9M/170M [00:00<00:03, 46.6MB/s]
24%|##3 | 40.7M/170M [00:00<00:02, 47.0MB/s]
33%|###2 | 56.0M/170M [00:00<00:01, 74.1MB/s]
38%|###7 | 64.0M/170M [00:01<00:01, 74.4MB/s]
43%|####3 | 73.4M/170M [00:01<00:01, 80.8MB/s]
52%|#####1 | 88.0M/170M [00:01<00:00, 96.1MB/s]
61%|######1 | 104M/170M [00:01<00:00, 99.5MB/s]
67%|######6 | 114M/170M [00:01<00:00, 88.9MB/s]
72%|#######1 | 122M/170M [00:01<00:00, 87.8MB/s]
77%|#######6 | 131M/170M [00:01<00:00, 70.9MB/s]
81%|########1 | 138M/170M [00:01<00:00, 72.0MB/s]
86%|########5 | 145M/170M [00:02<00:00, 70.0MB/s]
94%|#########4| 160M/170M [00:02<00:00, 84.1MB/s]
100%|##########| 170M/170M [00:02<00:00, 78.1MB/s]
+
0%| | 0.00/170M [00:00<?, ?B/s]
6%|5 | 10.1M/170M [00:00<00:01, 86.0MB/s]
12%|#2 | 21.2M/170M [00:00<00:01, 97.2MB/s]
18%|#7 | 30.5M/170M [00:00<00:01, 97.6MB/s]
24%|##3 | 40.0M/170M [00:00<00:01, 80.4MB/s]
32%|###1 | 53.9M/170M [00:00<00:01, 101MB/s]
38%|###7 | 64.1M/170M [00:00<00:01, 83.8MB/s]
43%|####2 | 72.8M/170M [00:00<00:01, 81.9MB/s]
48%|####7 | 81.0M/170M [00:01<00:01, 75.3MB/s]
53%|#####3 | 90.6M/170M [00:01<00:01, 81.0MB/s]
59%|#####8 | 99.7M/170M [00:01<00:00, 84.7MB/s]
64%|######3 | 108M/170M [00:01<00:00, 76.1MB/s]
71%|####### | 120M/170M [00:01<00:00, 82.0MB/s]
75%|#######5 | 128M/170M [00:01<00:00, 74.7MB/s]
80%|######## | 136M/170M [00:01<00:00, 67.7MB/s]
86%|########5 | 146M/170M [00:01<00:00, 75.6MB/s]
90%|######### | 153M/170M [00:02<00:00, 64.2MB/s]
95%|#########5| 162M/170M [00:02<00:00, 68.0MB/s]
100%|##########| 170M/170M [00:02<00:00, 78.6MB/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').
@@ -299,7 +299,7 @@ Get boxes with score larger than 0.9
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 3 minutes 28.344 seconds)
+ **Total running time of the script:** ( 3 minutes 21.632 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 448c51b8f9..8ae9f80c01 100644
--- a/docs/_sources/how_to/deploy_models/deploy_prequantized.rst.txt
+++ b/docs/_sources/how_to/deploy_models/deploy_prequantized.rst.txt
@@ -227,7 +227,7 @@ 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]
47%|####6 | 6.30M/13.6M [00:00<00:00, 53.4MB/s]
84%|########4 | 11.4M/13.6M [00:00<00:00, 50.5MB/s]
100%|##########| 13.6M/13.6M [00:00<00:00, 39.7MB/s]
+
0%| | 0.00/13.6M [00:00<?, ?B/s]
59%|#####8 | 7.99M/13.6M [00:00<00:00, 76.8MB/s]
100%|##########| 13.6M/13.6M [00:00<00:00, 103MB/s]
@@ -409,7 +409,7 @@ 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)
- 90.2068 90.1677 90.8147 90.0469 0.1458
+ 90.1720 90.1005 91.6585 89.8012 0.3037
@@ -458,7 +458,7 @@ TODO
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 1 minutes 12.919 seconds)
+ **Total running time of the script:** ( 1 minutes 10.916 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 29b3b11e05..02451a16d4 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
@@ -423,7 +423,7 @@ 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)
- 119.6592 119.5643 126.8230 119.0097 0.8249
+ 118.3536 118.3627 119.9356 117.1745 0.4004
@@ -460,7 +460,7 @@ 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 32.610 seconds)
+ **Total running time of the script:** ( 2 minutes 25.221 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 405d588589..2338c7b8df 100644
--- a/docs/_sources/how_to/deploy_models/deploy_quantized.rst.txt
+++ b/docs/_sources/how_to/deploy_models/deploy_quantized.rst.txt
@@ -257,7 +257,7 @@ We create a Relay VM to build and execute the model.
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 1 minutes 40.483 seconds)
+ **Total running time of the script:** ( 1 minutes 32.344 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 96a3e40b6b..f9d56cb7a3 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
@@ -170,7 +170,7 @@ 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]
5%|5 | 6645/132723 [00:00<00:01, 66444.68KB/s]
12%|#1 | 15500/132723 [00:00<00:01, 79443.08KB/s]
18%|#8 | 24276/132723 [00:00<00:01, 83238.71KB/s]
25%|##4 | 33050/132723 [00:00<00:01, 85011.97KB/s]
31%|###1 | 41552/132723 [00:00<00:01, 84301.38KB/s]
38%|###7 | 49984/132723 [00:00<00:01, 81095.71KB/s]
44%|####4 | 58717/132723 [00:00<00:00, 83069.29KB/s]
51%|##### | 67526/132723 [00:00<00:00, 84632.76KB/s]
57%|#####7 | 76271/132723 [00:00<00:00, 85499.51KB/s]
64%|######4 | 85023/132723 [00:01<00:00, 86114.52KB/s]
71%|####### | 93817/132723 [00:01<00:00, 86665.19KB/s]
77%|#######7 | 102637/132723 [00:01<00:00, 87119.68KB/s]
84%|########3 | 111464/132723 [00:01<00:00, 87464.64KB/s]
91%|######### | 120251/132723 [00:01<00:00, 87584.54KB/s]
97%|#########7| 129109/132723 [00:01<00:00, 87880.14KB/s]
100%|#######
###| 132723/132723 [00:01<00:00, 85230.00KB/s]
+
0%| | 0/132723 [00:00<?, ?KB/s]
3%|3 | 4409/132723 [00:00<00:02, 43293.70KB/s]
8%|8 | 10936/132723 [00:00<00:02, 56120.51KB/s]
14%|#3 | 18550/132723 [00:00<00:01, 65219.72KB/s]
20%|#9 | 26096/132723 [00:00<00:01, 69251.04KB/s]
25%|##5 | 33696/132723 [00:00<00:01, 71679.02KB/s]
31%|###1 | 41335/132723 [00:00<00:01, 73276.82KB/s]
37%|###6 | 48867/132723 [00:00<00:01, 73941.19KB/s]
43%|####2 | 56468/132723 [00:00<00:01, 74596.86KB/s]
48%|####8 | 64162/132723 [00:00<00:00, 75327.52KB/s]
54%|#####4 | 71814/132723 [00:01<00:00, 75692.40KB/s]
60%|#####9 | 79443/132723 [00:01<00:00, 75873.67KB/s]
66%|######5 | 87088/132723 [00:01<00:00, 76047.17KB/s]
71%|#######1 | 94694/132723 [00:01<00:00, 76034.39KB/s]
77%|#######7 | 102397/132723 [00:01<00:00, 76333.30KB/s]
83%|########2 | 110031/132723 [00:01<00:00, 76333.69KB/s]
89%|########8
| 117703/132723 [00:01<00:00, 76447.67KB/s]
94%|#########4| 125369/132723 [00:01<00:00, 76510.11KB/s]
100%|##########| 132723/132723 [00:01<00:00, 73840.87KB/s]
@@ -246,7 +246,7 @@ Display result
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 3 minutes 29.026 seconds)
+ **Total running time of the script:** ( 3 minutes 26.377 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 0e3033bcb8..469332e837 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
@@ -5,26 +5,26 @@
Computation times
=================
-**14:51.021** total execution time for **how_to_deploy_models** files:
+**14:22.337** total execution time for **how_to_deploy_models** files:
+------------------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_deploy_models_deploy_ssd_gluoncv.py` (``deploy_ssd_gluoncv.py``) | 03:29.026 | 0.0 MB |
+| :ref:`sphx_glr_how_to_deploy_models_deploy_ssd_gluoncv.py` (``deploy_ssd_gluoncv.py``) | 03:26.377 | 0.0 MB |
+------------------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_deploy_models_deploy_object_detection_pytorch.py` (``deploy_object_detection_pytorch.py``) | 03:28.344 | 0.0 MB |
+| :ref:`sphx_glr_how_to_deploy_models_deploy_object_detection_pytorch.py` (``deploy_object_detection_pytorch.py``) | 03:21.632 | 0.0 MB |
+------------------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_deploy_models_deploy_prequantized_tflite.py` (``deploy_prequantized_tflite.py``) | 02:32.610 | 0.0 MB |
+| :ref:`sphx_glr_how_to_deploy_models_deploy_prequantized_tflite.py` (``deploy_prequantized_tflite.py``) | 02:25.221 | 0.0 MB |
+------------------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_deploy_models_deploy_quantized.py` (``deploy_quantized.py``) | 01:40.483 | 0.0 MB |
+| :ref:`sphx_glr_how_to_deploy_models_deploy_quantized.py` (``deploy_quantized.py``) | 01:32.344 | 0.0 MB |
+------------------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_deploy_models_deploy_prequantized.py` (``deploy_prequantized.py``) | 01:12.919 | 0.0 MB |
+| :ref:`sphx_glr_how_to_deploy_models_deploy_prequantized.py` (``deploy_prequantized.py``) | 01:10.916 | 0.0 MB |
+------------------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_deploy_models_deploy_model_on_adreno.py` (``deploy_model_on_adreno.py``) | 00:53.319 | 0.0 MB |
+| :ref:`sphx_glr_how_to_deploy_models_deploy_model_on_adreno.py` (``deploy_model_on_adreno.py``) | 00:55.272 | 0.0 MB |
+------------------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_deploy_models_deploy_model_on_android.py` (``deploy_model_on_android.py``) | 00:39.824 | 0.0 MB |
+| :ref:`sphx_glr_how_to_deploy_models_deploy_model_on_android.py` (``deploy_model_on_android.py``) | 00:38.670 | 0.0 MB |
+------------------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_deploy_models_deploy_model_on_nano.py` (``deploy_model_on_nano.py``) | 00:27.427 | 0.0 MB |
+| :ref:`sphx_glr_how_to_deploy_models_deploy_model_on_nano.py` (``deploy_model_on_nano.py``) | 00:26.085 | 0.0 MB |
+------------------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_deploy_models_deploy_model_on_rasp.py` (``deploy_model_on_rasp.py``) | 00:27.063 | 0.0 MB |
+| :ref:`sphx_glr_how_to_deploy_models_deploy_model_on_rasp.py` (``deploy_model_on_rasp.py``) | 00:25.814 | 0.0 MB |
+------------------------------------------------------------------------------------------------------------------+-----------+--------+
| :ref:`sphx_glr_how_to_deploy_models_deploy_sparse.py` (``deploy_sparse.py``) | 00:00.006 | 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 5b1d109e8d..3ead13366d 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
@@ -463,7 +463,7 @@ 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.zipcb3154e2-9390-47b4-8140-e0c63d32f07a from https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/models/mobilenet0.25-9f83e440.zip...
+ Downloading /workspace/.mxnet/models/mobilenet0.25-9f83e440.zip07fb9aa0-cc7e-474e-ab93-9e3af8a9a86f 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 35427c1612..56fb26ff7c 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
@@ -93,16 +93,13 @@ our customized lowering pass to manipulate the IR directly instead of using sche
@I.ir_module
class Module:
@T.prim_func
- def main(a: T.handle, b: T.handle, c: T.handle):
+ def main(a: T.Buffer((128,), "float32"), b: T.Buffer((128,), "float32"), c: T.Buffer((128,), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- a_1 = T.match_buffer(a, (128,))
- b_1 = T.match_buffer(b, (128,))
- c_1 = T.match_buffer(c, (128,))
for i in range(128):
- c_2 = T.buffer_decl((128,), data=c_1.data)
- a_2 = T.buffer_decl((128,), data=a_1.data)
- b_2 = T.buffer_decl((128,), data=b_1.data)
- c_2[i] = a_2[i] + b_2[i]
+ c_1 = T.buffer_decl((128,), data=c.data)
+ a_1 = T.buffer_decl((128,), data=a.data)
+ b_1 = T.buffer_decl((128,), data=b.data)
+ c_1[i] = a_1[i] + b_1[i]
@@ -253,17 +250,14 @@ Thus, a good place to put this transformation pass is just after Phase 1.
@I.ir_module
class Module:
@T.prim_func
- def main(a: T.handle, b: T.handle, c: T.handle):
+ def main(a: T.Buffer((128,), "float32"), b: T.Buffer((128,), "float32"), c: T.Buffer((128,), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- a_1 = T.match_buffer(a, (128,))
- b_1 = T.match_buffer(b, (128,))
- c_1 = T.match_buffer(c, (128,))
for i_outer in range(16):
cse_var_1: T.int32 = i_outer * 8
- c_2 = T.buffer_decl((128,), data=c_1.data)
- a_2 = T.buffer_decl((128,), data=a_1.data)
- b_2 = T.buffer_decl((128,), data=b_1.data)
- c_2[cse_var_1:cse_var_1 + 8] = a_2[cse_var_1:cse_var_1 + 8] + b_2[cse_var_1:cse_var_1 + 8]
+ c_1 = T.buffer_decl((128,), data=c.data)
+ a_1 = T.buffer_decl((128,), data=a.data)
+ b_1 = T.buffer_decl((128,), data=b.data)
+ c_1[cse_var_1:cse_var_1 + 8] = a_1[cse_var_1:cse_var_1 + 8] + b_1[cse_var_1:cse_var_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 760ff0207a..a5decbd58b 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
=================
-**00:52.015** total execution time for **how_to_extend_tvm** files:
+**00:50.546** 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``) | 00:48.296 | 0.0 MB |
+| :ref:`sphx_glr_how_to_extend_tvm_bring_your_own_datatypes.py` (``bring_your_own_datatypes.py``) | 00:46.963 | 0.0 MB |
+-------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_extend_tvm_use_pass_instrument.py` (``use_pass_instrument.py``) | 00:02.653 | 0.0 MB |
+| :ref:`sphx_glr_how_to_extend_tvm_use_pass_instrument.py` (``use_pass_instrument.py``) | 00:02.544 | 0.0 MB |
+-------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_extend_tvm_use_pass_infra.py` (``use_pass_infra.py``) | 00:01.060 | 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_low_level_custom_pass.py` (``low_level_custom_pass.py``) | 00:00.007 | 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 a67020cf5a..edaff2d669 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
@@ -220,10 +220,10 @@ profile the execution time of each passes.
.. code-block:: none
Printing results of timing profile...
- InferType: 18158us [18158us] (48.55%; 48.55%)
- FoldScaleAxis: 19244us [7us] (51.45%; 51.45%)
- FoldConstant: 19237us [1733us] (51.43%; 99.96%)
- InferType: 17505us [17505us] (46.80%; 90.99%)
+ InferType: 17614us [17614us] (48.44%; 48.44%)
+ FoldScaleAxis: 18746us [6us] (51.56%; 51.56%)
+ FoldConstant: 18740us [1648us] (51.54%; 99.97%)
+ InferType: 17091us [17091us] (47.00%; 91.20%)
@@ -262,10 +262,10 @@ Refer to following sections and :py:func:`tvm.instrument.pass_instrument` for th
.. code-block:: none
Printing results of timing profile...
- InferType: 17363us [17363us] (47.85%; 47.85%)
- FoldScaleAxis: 18921us [5us] (52.15%; 52.15%)
- FoldConstant: 18916us [1709us] (52.13%; 99.97%)
- InferType: 17208us [17208us] (47.42%; 90.97%)
+ InferType: 17137us [17137us] (47.94%; 47.94%)
+ FoldScaleAxis: 18611us [5us] (52.06%; 52.06%)
+ FoldConstant: 18606us [1663us] (52.05%; 99.97%)
+ InferType: 16943us [16943us] (47.40%; 91.06%)
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 027d52359e..afada960c6 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
@@ -331,7 +331,7 @@ latency of convolution.
.. code-block:: none
- Convolution: 44.062721 ms
+ Convolution: 54.217983 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 9701ef674d..e93ff5c59c 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
@@ -435,11 +435,8 @@ one time.
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, W: T.handle, Conv: T.handle):
+ def main(A: T.Buffer((16, 14, 14, 16, 16, 16), "float16"), W: T.Buffer((3, 3, 16, 32, 16, 16), "float16"), Conv: T.Buffer((16, 14, 14, 32, 16, 16), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (16, 14, 14, 16, 16, 16), "float16")
- W_1 = T.match_buffer(W, (3, 3, 16, 32, 16, 16), "float16")
- Conv_1 = T.match_buffer(Conv, (16, 14, 14, 32, 16, 16))
blockIdx_z = T.env_thread("blockIdx.z")
T.launch_thread(blockIdx_z, 196)
Conv_wmma_accumulator = T.allocate([2048], "float32", "wmma.accumulator")
@@ -465,13 +462,13 @@ one time.
cse_var_2: T.int32 = ax3 * 256
cse_var_1: T.int32 = ax4_ax5_fused_outer * 32
T.launch_thread(threadIdx_x, 32)
- A_2 = T.buffer_decl((12845056,), "float16", data=A_1.data)
- Apad_shared_1[threadIdx_y * 3072 + threadIdx_z * 1536 + ax2 * 512 + cse_var_2 + cse_var_1 + threadIdx_x] = T.if_then_else(1 <= blockIdx_z // 14 + kh and blockIdx_z // 14 + kh < 15 and 1 <= ax2 + blockIdx_z % 14 and ax2 + blockIdx_z % 14 < 15, A_2[blockIdx_x * 6422528 + threadIdx_y * 1605632 + threadIdx_z * 802816 + kh * 57344 + blockIdx_z * 4096 + ax2 * 4096 + ic_outer * 512 + cse_var_2 + cse_var_1 + threadIdx_x - 61440], T.float16(0))
+ A_1 = T.buffer_decl((12845056,), "float16", data=A.data)
+ Apad_shared_1[threadIdx_y * 3072 + threadIdx_z * 1536 + ax2 * 512 + cse_var_2 + cse_var_1 + threadIdx_x] = T.if_then_else(1 <= blockIdx_z // 14 + kh and blockIdx_z // 14 + kh < 15 and 1 <= ax2 + blockIdx_z % 14 and ax2 + blockIdx_z % 14 < 15, A_1[blockIdx_x * 6422528 + threadIdx_y * 1605632 + threadIdx_z * 802816 + kh * 57344 + blockIdx_z * 4096 + ax2 * 4096 + ic_outer * 512 + cse_var_2 + cse_var_1 + threadIdx_x - 61440], T.float16(0))
W_shared_1 = T.buffer_decl((12288,), "float16", data=W_shared, scope="shared")
for ax1, ax2 in T.grid(3, 2):
T.launch_thread(threadIdx_x, 32)
- W_2 = T.buffer_decl((1179648,), "float16", data=W_1.data)
- W_shared_1[ax1 * 4096 + ax2 * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8:ax1 * 4096 + ax2 * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8 + 8] = W_2[kh * 393216 + ax1 * 131072 + ic_outer * 16384 + ax2 * 8192 + blockIdx_y * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8:kh * 393216 + ax1 * 131072 + ic_outer * 16384 + ax2 * 8192 + blockIdx_y * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8 + 8]
+ W_1 = T.buffer_decl((1179648,), "float16", data=W.data)
+ W_shared_1[ax1 * 4096 + ax2 * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8:ax1 * 4096 + ax2 * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8 + 8] = W_1[kh * 393216 + ax1 * 131072 + ic_outer * 16384 + ax2 * 8192 + blockIdx_y * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8:kh * 393216 + ax1 * 131072 + ic_outer * 16384 + ax2 * 8192 + blockIdx_y * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8 + 8]
for ic_inner, kw in T.grid(2, 3):
Apad_shared_wmma_matrix_a_1 = T.buffer_decl((512,), "float16", data=Apad_shared_wmma_matrix_a, scope="wmma.matrix_a")
for ax0, ax4, ax5 in T.grid(2, 16, 16):
@@ -490,8 +487,8 @@ one time.
for n_inner, o_inner, nn, oo in T.grid(2, 4, 16, 16):
cse_var_10: T.int32 = o_inner * 256
cse_var_9: T.int32 = nn * 16
- Conv_2 = T.buffer_decl((25690112,), data=Conv_1.data)
- Conv_2[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_1 = T.buffer_decl((25690112,), data=Conv.data)
+ Conv_1[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]
@@ -525,11 +522,8 @@ by mapping the 2D convolution to tensor intrinsics
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, W: T.handle, Conv: T.handle):
+ def main(A: T.Buffer((16, 14, 14, 16, 16, 16), "float16"), W: T.Buffer((3, 3, 16, 32, 16, 16), "float16"), Conv: T.Buffer((16, 14, 14, 32, 16, 16), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (16, 14, 14, 16, 16, 16), "float16")
- W_1 = T.match_buffer(W, (3, 3, 16, 32, 16, 16), "float16")
- Conv_1 = T.match_buffer(Conv, (16, 14, 14, 32, 16, 16))
blockIdx_z = T.env_thread("blockIdx.z")
T.launch_thread(blockIdx_z, 196)
Conv_wmma_accumulator = T.allocate([2048], "float32", "wmma.accumulator")
@@ -554,13 +548,13 @@ by mapping the 2D convolution to tensor intrinsics
cse_var_1: T.int32 = ax4_ax5_fused_outer * 32
T.launch_thread(threadIdx_x, 32)
Apad_shared_1 = T.buffer_decl((12288,), "float16", data=Apad_shared, scope="shared")
- A_2 = T.buffer_decl((12845056,), "float16", data=A_1.data)
- Apad_shared_1[threadIdx_y * 3072 + threadIdx_z * 1536 + ax2 * 512 + cse_var_2 + cse_var_1 + threadIdx_x] = T.if_then_else(1 <= blockIdx_z // 14 + kh and blockIdx_z // 14 + kh < 15 and 1 <= ax2 + blockIdx_z % 14 and ax2 + blockIdx_z % 14 < 15, A_2[blockIdx_x * 6422528 + threadIdx_y * 1605632 + threadIdx_z * 802816 + kh * 57344 + blockIdx_z * 4096 + ax2 * 4096 + ic_outer * 512 + cse_var_2 + cse_var_1 + threadIdx_x - 61440], T.float16(0))
+ A_1 = T.buffer_decl((12845056,), "float16", data=A.data)
+ Apad_shared_1[threadIdx_y * 3072 + threadIdx_z * 1536 + ax2 * 512 + cse_var_2 + cse_var_1 + threadIdx_x] = T.if_then_else(1 <= blockIdx_z // 14 + kh and blockIdx_z // 14 + kh < 15 and 1 <= ax2 + blockIdx_z % 14 and ax2 + blockIdx_z % 14 < 15, A_1[blockIdx_x * 6422528 + threadIdx_y * 1605632 + threadIdx_z * 802816 + kh * 57344 + blockIdx_z * 4096 + ax2 * 4096 + ic_outer * 512 + cse_var_2 + cse_var_1 + threadIdx_x - 61440], T.float16(0))
for ax1, ax2 in T.grid(3, 2):
T.launch_thread(threadIdx_x, 32)
W_shared_1 = T.buffer_decl((12288,), "float16", data=W_shared, scope="shared")
- W_2 = T.buffer_decl((1179648,), "float16", data=W_1.data)
- W_shared_1[ax1 * 4096 + ax2 * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8:ax1 * 4096 + ax2 * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8 + 8] = W_2[kh * 393216 + ax1 * 131072 + ic_outer * 16384 + ax2 * 8192 + blockIdx_y * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8:kh * 393216 + ax1 * 131072 + ic_outer * 16384 + ax2 * 8192 + blockIdx_y * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8 + 8]
+ W_1 = T.buffer_decl((1179648,), "float16", data=W.data)
+ W_shared_1[ax1 * 4096 + ax2 * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8:ax1 * 4096 + ax2 * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8 + 8] = W_1[kh * 393216 + ax1 * 131072 + ic_outer * 16384 + ax2 * 8192 + blockIdx_y * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8:kh * 393216 + ax1 * 131072 + ic_outer * 16384 + ax2 * 8192 + blockIdx_y * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8 + 8]
for ic_inner, kw in T.grid(2, 3):
for ax0 in range(2):
T.tvm_load_matrix_sync(Apad_shared_wmma_matrix_a, 16, 16, 16, ax0, T.tvm_access_ptr(T.type_annotation("float16"), Apad_shared, threadIdx_y * 3072 + ax0 * 1536 + kw * 512 + ic_inner * 256, 256, 1), 16, "row_major")
@@ -570,7 +564,7 @@ by mapping the 2D convolution to tensor intrinsics
cse_var_3: T.int32 = n_c * 4 + o_c
T.tvm_mma_sync(Conv_wmma_accumulator, cse_var_3, Apad_shared_wmma_matrix_a, n_c, W_shared_wmma_matrix_b, o_c, Conv_wmma_accumulator, cse_var_3)
for n_inner, o_inner in T.grid(2, 4):
- T.tvm_store_matrix_sync(Conv_wmma_accumulator, 16, 16, 16, n_inner * 4 + o_inner, T.tvm_access_ptr(T.type_annotation("float32"), Conv_1.data, blockIdx_x * 12845056 + threadIdx_y * 3211264 + n_inner * 1605632 + blockIdx_z * 8192 + blockIdx_y * 2048 + threadIdx_z * 1024 + o_inner * 256, 256, 2), 16, "row_major")
+ T.tvm_store_matrix_sync(Conv_wmma_accumulator, 16, 16, 16, n_inner * 4 + o_inner, T.tvm_access_ptr(T.type_annotation("float32"), Conv.data, blockIdx_x * 12845056 + threadIdx_y * 3211264 + n_inner * 1605632 + blockIdx_z * 8192 + blockIdx_y * 2048 + threadIdx_z * 1024 + o_inner * 256, 256, 2), 16, "row_major")
@@ -608,7 +602,7 @@ be able to run on our build server
.. code-block:: none
- conv2d with tensor core: 10.772263 ms
+ conv2d with tensor core: 6.685200 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 df0ed7c798..02f2bbe7f7 100644
--- a/docs/_sources/how_to/optimize_operators/opt_gemm.rst.txt
+++ b/docs/_sources/how_to/optimize_operators/opt_gemm.rst.txt
@@ -134,8 +134,8 @@ Then we write a baseline implementation, the simplest way to write a matrix mult
.. code-block:: none
- Numpy running time: 0.018628
- Baseline: 3.397229
+ Numpy running time: 0.018060
+ Baseline: 3.197752
@@ -163,20 +163,17 @@ Here is the generated IR using our baseline schedule.
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 1024), "float32"), B: T.Buffer((1024, 1024), "float32"), C: T.Buffer((1024, 1024), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 1024))
- B_1 = T.match_buffer(B, (1024, 1024))
- C_1 = T.match_buffer(C, (1024, 1024))
for m, n in T.grid(1024, 1024):
- C_2 = T.buffer_decl((1048576,), data=C_1.data)
- C_2[m * 1024 + n] = T.float32(0)
+ C_1 = T.buffer_decl((1048576,), data=C.data)
+ C_1[m * 1024 + n] = T.float32(0)
for k in range(1024):
cse_var_2: T.int32 = m * 1024
cse_var_1: T.int32 = cse_var_2 + n
- A_2 = T.buffer_decl((1048576,), data=A_1.data)
- B_2 = T.buffer_decl((1048576,), data=B_1.data)
- C_2[cse_var_1] = C_2[cse_var_1] + A_2[cse_var_2 + k] * B_2[k * 1024 + n]
+ A_1 = T.buffer_decl((1048576,), data=A.data)
+ B_1 = T.buffer_decl((1048576,), data=B.data)
+ C_1[cse_var_1] = C_1[cse_var_1] + A_1[cse_var_2 + k] * B_1[k * 1024 + n]
@@ -227,7 +224,7 @@ fill 32 * 32 * sizeof(float) which is 4KB in the cache whose total size is 32KB
.. code-block:: none
- Opt1: 0.295933
+ Opt1: 0.296106
@@ -254,22 +251,19 @@ Here is the generated IR after blocking.
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 1024), "float32"), B: T.Buffer((1024, 1024), "float32"), C: T.Buffer((1024, 1024), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 1024))
- B_1 = T.match_buffer(B, (1024, 1024))
- C_1 = T.match_buffer(C, (1024, 1024))
for m_outer, n_outer in T.grid(32, 32):
- C_2 = T.buffer_decl((1048576,), data=C_1.data)
+ C_1 = T.buffer_decl((1048576,), data=C.data)
for m_inner_init, n_inner_init in T.grid(32, 32):
- C_2[m_outer * 32768 + m_inner_init * 1024 + n_outer * 32 + n_inner_init] = T.float32(0)
+ C_1[m_outer * 32768 + m_inner_init * 1024 + n_outer * 32 + n_inner_init] = T.float32(0)
for k_outer, k_inner, m_inner, n_inner in T.grid(256, 4, 32, 32):
cse_var_3: T.int32 = n_outer * 32
cse_var_2: T.int32 = m_outer * 32768 + m_inner * 1024
cse_var_1: T.int32 = cse_var_2 + cse_var_3 + n_inner
- A_2 = T.buffer_decl((1048576,), data=A_1.data)
- B_2 = T.buffer_decl((1048576,), data=B_1.data)
- C_2[cse_var_1] = C_2[cse_var_1] + A_2[cse_var_2 + k_outer * 4 + k_inner] * B_2[k_outer * 4096 + k_inner * 1024 + cse_var_3 + n_inner]
+ A_1 = T.buffer_decl((1048576,), data=A.data)
+ B_1 = T.buffer_decl((1048576,), data=B.data)
+ C_1[cse_var_1] = C_1[cse_var_1] + A_1[cse_var_2 + k_outer * 4 + k_inner] * B_1[k_outer * 4096 + k_inner * 1024 + cse_var_3 + n_inner]
@@ -318,7 +312,7 @@ In this tutorial, we chose to vectorize the inner loop row data since it is cach
.. code-block:: none
- Opt2: 0.332504
+ Opt2: 0.333387
@@ -345,22 +339,19 @@ Here is the generated IR after vectorization.
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 1024), "float32"), B: T.Buffer((1024, 1024), "float32"), C: T.Buffer((1024, 1024), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 1024))
- B_1 = T.match_buffer(B, (1024, 1024))
- C_1 = T.match_buffer(C, (1024, 1024))
for m_outer, n_outer in T.grid(32, 32):
- C_2 = T.buffer_decl((1048576,), data=C_1.data)
+ C_1 = T.buffer_decl((1048576,), data=C.data)
for m_inner_init in range(32):
- C_2[m_outer * 32768 + m_inner_init * 1024 + n_outer * 32:m_outer * 32768 + m_inner_init * 1024 + n_outer * 32 + 32] = T.Broadcast(T.float32(0), 32)
+ C_1[m_outer * 32768 + m_inner_init * 1024 + n_outer * 32:m_outer * 32768 + m_inner_init * 1024 + n_outer * 32 + 32] = T.Broadcast(T.float32(0), 32)
for k_outer, k_inner, m_inner in T.grid(256, 4, 32):
cse_var_3: T.int32 = n_outer * 32
cse_var_2: T.int32 = m_outer * 32768 + m_inner * 1024
cse_var_1: T.int32 = cse_var_2 + cse_var_3
- A_2 = T.buffer_decl((1048576,), data=A_1.data)
- B_2 = T.buffer_decl((1048576,), data=B_1.data)
- C_2[cse_var_1:cse_var_1 + 32] = C_2[cse_var_1:cse_var_1 + 32] + T.Broadcast(A_2[cse_var_2 + k_outer * 4 + k_inner], 32) * B_2[k_outer * 4096 + k_inner * 1024 + cse_var_3:k_outer * 4096 + k_inner * 1024 + cse_var_3 + 32]
+ A_1 = T.buffer_decl((1048576,), data=A.data)
+ B_1 = T.buffer_decl((1048576,), data=B.data)
+ C_1[cse_var_1:cse_var_1 + 32] = C_1[cse_var_1:cse_var_1 + 32] + T.Broadcast(A_1[cse_var_2 + k_outer * 4 + k_inner], 32) * B_1[k_outer * 4096 + k_inner * 1024 + cse_var_3:k_outer * 4096 + k_inner * 1024 + cse_var_3 + 32]
@@ -406,7 +397,7 @@ the access pattern for A matrix is more cache friendly.
.. code-block:: none
- Opt3: 0.116739
+ Opt3: 0.114968
@@ -433,22 +424,19 @@ Here is the generated IR after loop permutation.
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 1024), "float32"), B: T.Buffer((1024, 1024), "float32"), C: T.Buffer((1024, 1024), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 1024))
- B_1 = T.match_buffer(B, (1024, 1024))
- C_1 = T.match_buffer(C, (1024, 1024))
for m_outer, n_outer in T.grid(32, 32):
- C_2 = T.buffer_decl((1048576,), data=C_1.data)
+ C_1 = T.buffer_decl((1048576,), data=C.data)
for m_inner_init in range(32):
- C_2[m_outer * 32768 + m_inner_init * 1024 + n_outer * 32:m_outer * 32768 + m_inner_init * 1024 + n_outer * 32 + 32] = T.Broadcast(T.float32(0), 32)
+ C_1[m_outer * 32768 + m_inner_init * 1024 + n_outer * 32:m_outer * 32768 + m_inner_init * 1024 + n_outer * 32 + 32] = T.Broadcast(T.float32(0), 32)
for k_outer, m_inner, k_inner in T.grid(256, 32, 4):
cse_var_3: T.int32 = n_outer * 32
cse_var_2: T.int32 = m_outer * 32768 + m_inner * 1024
cse_var_1: T.int32 = cse_var_2 + cse_var_3
- A_2 = T.buffer_decl((1048576,), data=A_1.data)
- B_2 = T.buffer_decl((1048576,), data=B_1.data)
- C_2[cse_var_1:cse_var_1 + 32] = C_2[cse_var_1:cse_var_1 + 32] + T.Broadcast(A_2[cse_var_2 + k_outer * 4 + k_inner], 32) * B_2[k_outer * 4096 + k_inner * 1024 + cse_var_3:k_outer * 4096 + k_inner * 1024 + cse_var_3 + 32]
+ A_1 = T.buffer_decl((1048576,), data=A.data)
+ B_1 = T.buffer_decl((1048576,), data=B.data)
+ C_1[cse_var_1:cse_var_1 + 32] = C_1[cse_var_1:cse_var_1 + 32] + T.Broadcast(A_1[cse_var_2 + k_outer * 4 + k_inner], 32) * B_1[k_outer * 4096 + k_inner * 1024 + cse_var_3:k_outer * 4096 + k_inner * 1024 + cse_var_3 + 32]
@@ -523,7 +511,7 @@ flattening.
.. code-block:: none
- Opt4: 0.109848
+ Opt4: 0.108361
@@ -550,27 +538,24 @@ Here is the generated IR after array packing.
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 1024), "float32"), B: T.Buffer((1024, 1024), "float32"), C: T.Buffer((1024, 1024), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 1024))
- B_1 = T.match_buffer(B, (1024, 1024))
- C_1 = T.match_buffer(C, (1024, 1024))
packedB = T.allocate([32768], "float32x32", "global")
packedB_1 = T.buffer_decl((32768,), "float32x32", data=packedB)
for bigN in T.parallel(32):
for k in range(1024):
- B_2 = T.buffer_decl((1048576,), data=B_1.data)
- packedB_1[bigN * 1024 + k] = B_2[k * 1024 + bigN * 32:k * 1024 + bigN * 32 + 32]
+ B_1 = T.buffer_decl((1048576,), data=B.data)
+ packedB_1[bigN * 1024 + k] = B_1[k * 1024 + bigN * 32:k * 1024 + bigN * 32 + 32]
for m_outer, n_outer in T.grid(32, 32):
- C_2 = T.buffer_decl((1048576,), data=C_1.data)
+ C_1 = T.buffer_decl((1048576,), data=C.data)
for m_inner_init in range(32):
- C_2[m_outer * 32768 + m_inner_init * 1024 + n_outer * 32:m_outer * 32768 + m_inner_init * 1024 + n_outer * 32 + 32] = T.Broadcast(T.float32(0), 32)
+ C_1[m_outer * 32768 + m_inner_init * 1024 + n_outer * 32:m_outer * 32768 + m_inner_init * 1024 + n_outer * 32 + 32] = T.Broadcast(T.float32(0), 32)
for k_outer, m_inner, k_inner in T.grid(256, 32, 4):
cse_var_3: T.int32 = m_outer * 32768 + m_inner * 1024
cse_var_2: T.int32 = k_outer * 4
cse_var_1: T.int32 = cse_var_3 + n_outer * 32
- A_2 = T.buffer_decl((1048576,), data=A_1.data)
- C_2[cse_var_1:cse_var_1 + 32] = C_2[cse_var_1:cse_var_1 + 32] + T.Broadcast(A_2[cse_var_3 + cse_var_2 + k_inner], 32) * packedB_1[n_outer * 1024 + cse_var_2 + k_inner]
+ A_1 = T.buffer_decl((1048576,), data=A.data)
+ C_1[cse_var_1:cse_var_1 + 32] = C_1[cse_var_1:cse_var_1 + 32] + T.Broadcast(A_1[cse_var_3 + cse_var_2 + k_inner], 32) * packedB_1[n_outer * 1024 + cse_var_2 + k_inner]
@@ -635,7 +620,7 @@ write to C when all the block results are ready.
.. code-block:: none
- Opt5: 0.112016
+ Opt5: 0.111469
@@ -662,18 +647,15 @@ Here is the generated IR after blocking.
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 1024), "float32"), B: T.Buffer((1024, 1024), "float32"), C: T.Buffer((1024, 1024), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 1024))
- B_1 = T.match_buffer(B, (1024, 1024))
- C_1 = T.match_buffer(C, (1024, 1024))
packedB = T.allocate([32768], "float32x32", "global")
C_global = T.allocate([1024], "float32", "global")
packedB_1 = T.buffer_decl((32768,), "float32x32", data=packedB)
for bigN in T.parallel(32):
for k in range(1024):
- B_2 = T.buffer_decl((1048576,), data=B_1.data)
- packedB_1[bigN * 1024 + k] = B_2[k * 1024 + bigN * 32:k * 1024 + bigN * 32 + 32]
+ B_1 = T.buffer_decl((1048576,), data=B.data)
+ packedB_1[bigN * 1024 + k] = B_1[k * 1024 + bigN * 32:k * 1024 + bigN * 32 + 32]
for m_outer, n_outer in T.grid(32, 32):
C_global_1 = T.buffer_decl((1024,), data=C_global)
for m_c_init in range(32):
@@ -683,14 +665,14 @@ Here is the generated IR after blocking.
cse_var_3: T.int32 = m_c * 32
cse_var_2: T.int32 = n_outer * 1024 + cse_var_4
cse_var_1: T.int32 = m_outer * 32768 + m_c * 1024 + cse_var_4
- A_2 = T.buffer_decl((1048576,), data=A_1.data)
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1], 32) * packedB_1[cse_var_2]
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1 + 1], 32) * packedB_1[cse_var_2 + 1]
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1 + 2], 32) * packedB_1[cse_var_2 + 2]
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1 + 3], 32) * packedB_1[cse_var_2 + 3]
+ A_1 = T.buffer_decl((1048576,), data=A.data)
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1], 32) * packedB_1[cse_var_2]
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1 + 1], 32) * packedB_1[cse_var_2 + 1]
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1 + 2], 32) * packedB_1[cse_var_2 + 2]
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1 + 3], 32) * packedB_1[cse_var_2 + 3]
for m_inner, n_inner in T.grid(32, 32):
- C_2 = T.buffer_decl((1048576,), data=C_1.data)
- C_2[m_outer * 32768 + m_inner * 1024 + n_outer * 32 + n_inner] = C_global_1[m_inner * 32 + n_inner]
+ C_1 = T.buffer_decl((1048576,), data=C.data)
+ C_1[m_outer * 32768 + m_inner * 1024 + n_outer * 32 + n_inner] = C_global_1[m_inner * 32 + n_inner]
@@ -748,7 +730,7 @@ Furthermore, we can also utilize multi-core processors to do the thread-level pa
.. code-block:: none
- Opt6: 0.147469
+ Opt6: 0.146623
@@ -775,17 +757,14 @@ Here is the generated IR after parallelization.
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 1024), "float32"), B: T.Buffer((1024, 1024), "float32"), C: T.Buffer((1024, 1024), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 1024))
- B_1 = T.match_buffer(B, (1024, 1024))
- C_1 = T.match_buffer(C, (1024, 1024))
packedB = T.allocate([32768], "float32x32", "global")
packedB_1 = T.buffer_decl((32768,), "float32x32", data=packedB)
for bigN in T.parallel(32):
for k in range(1024):
- B_2 = T.buffer_decl((1048576,), data=B_1.data)
- packedB_1[bigN * 1024 + k] = B_2[k * 1024 + bigN * 32:k * 1024 + bigN * 32 + 32]
+ B_1 = T.buffer_decl((1048576,), data=B.data)
+ packedB_1[bigN * 1024 + k] = B_1[k * 1024 + bigN * 32:k * 1024 + bigN * 32 + 32]
for m_outer in T.parallel(32):
C_global = T.allocate([1024], "float32", "global")
for n_outer in range(32):
@@ -797,14 +776,14 @@ Here is the generated IR after parallelization.
cse_var_3: T.int32 = m_c * 32
cse_var_2: T.int32 = n_outer * 1024 + cse_var_4
cse_var_1: T.int32 = m_outer * 32768 + m_c * 1024 + cse_var_4
- A_2 = T.buffer_decl((1048576,), data=A_1.data)
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1], 32) * packedB_1[cse_var_2]
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1 + 1], 32) * packedB_1[cse_var_2 + 1]
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1 + 2], 32) * packedB_1[cse_var_2 + 2]
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1 + 3], 32) * packedB_1[cse_var_2 + 3]
+ A_1 = T.buffer_decl((1048576,), data=A.data)
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1], 32) * packedB_1[cse_var_2]
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1 + 1], 32) * packedB_1[cse_var_2 + 1]
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1 + 2], 32) * packedB_1[cse_var_2 + 2]
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1 + 3], 32) * packedB_1[cse_var_2 + 3]
for m_inner, n_inner in T.grid(32, 32):
- C_2 = T.buffer_decl((1048576,), data=C_1.data)
- C_2[m_outer * 32768 + m_inner * 1024 + n_outer * 32 + n_inner] = C_global_1[m_inner * 32 + n_inner]
+ C_1 = T.buffer_decl((1048576,), data=C.data)
+ C_1[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 151dc4eced..9ea002bb02 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
=================
-**00:34.839** total execution time for **how_to_optimize_operators** files:
+**00:33.932** total execution time for **how_to_optimize_operators** files:
+-----------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_optimize_operators_opt_gemm.py` (``opt_gemm.py``) | 00:32.242 | 0.0 MB |
+| :ref:`sphx_glr_how_to_optimize_operators_opt_gemm.py` (``opt_gemm.py``) | 00:31.481 | 0.0 MB |
+-----------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_optimize_operators_opt_conv_tensorcore.py` (``opt_conv_tensorcore.py``) | 00:01.513 | 0.0 MB |
+| :ref:`sphx_glr_how_to_optimize_operators_opt_conv_tensorcore.py` (``opt_conv_tensorcore.py``) | 00:01.426 | 0.0 MB |
+-----------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_optimize_operators_opt_conv_cuda.py` (``opt_conv_cuda.py``) | 00:01.085 | 0.0 MB |
+| :ref:`sphx_glr_how_to_optimize_operators_opt_conv_cuda.py` (``opt_conv_cuda.py``) | 00:01.025 | 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 19f750c4c4..f2f08d6578 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
=================
-**09:28.448** total execution time for **how_to_tune_with_autoscheduler** files:
+**09:06.030** 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``) | 05:46.341 | 0.0 MB |
+| :ref:`sphx_glr_how_to_tune_with_autoscheduler_tune_conv2d_layer_cuda.py` (``tune_conv2d_layer_cuda.py``) | 05:27.031 | 0.0 MB |
+----------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_tune_with_autoscheduler_tune_network_x86.py` (``tune_network_x86.py``) | 01:38.610 | 0.0 MB |
+| :ref:`sphx_glr_how_to_tune_with_autoscheduler_tune_network_x86.py` (``tune_network_x86.py``) | 01:36.571 | 0.0 MB |
+----------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_tune_with_autoscheduler_tune_network_cuda.py` (``tune_network_cuda.py``) | 01:05.810 | 0.0 MB |
+| :ref:`sphx_glr_how_to_tune_with_autoscheduler_tune_network_cuda.py` (``tune_network_cuda.py``) | 01:04.732 | 0.0 MB |
+----------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_tune_with_autoscheduler_tune_sparse_x86.py` (``tune_sparse_x86.py``) | 00:31.182 | 0.0 MB |
+| :ref:`sphx_glr_how_to_tune_with_autoscheduler_tune_sparse_x86.py` (``tune_sparse_x86.py``) | 00:31.554 | 0.0 MB |
+----------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_tune_with_autoscheduler_tune_network_arm.py` (``tune_network_arm.py``) | 00:13.841 | 0.0 MB |
+| :ref:`sphx_glr_how_to_tune_with_autoscheduler_tune_network_arm.py` (``tune_network_arm.py``) | 00:13.565 | 0.0 MB |
+----------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_tune_with_autoscheduler_tune_network_mali.py` (``tune_network_mali.py``) | 00:12.664 | 0.0 MB |
+| :ref:`sphx_glr_how_to_tune_with_autoscheduler_tune_network_mali.py` (``tune_network_mali.py``) | 00:12.577 | 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 ee25ed6dc0..c1d24be76f 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
@@ -238,19 +238,15 @@ cooperative fetching, unrolling and operator fusion.
@I.ir_module
class Module:
@T.prim_func
- def main(data: T.handle, kernel: T.handle, bias: T.handle, compute: T.handle):
+ def main(data: T.Buffer((1, 512, 7, 7), "float32"), kernel: T.Buffer((512, 512, 3, 3), "float32"), bias: T.Buffer((1, 512, 1, 1), "float32"), compute: T.Buffer((1, 512, 7, 7), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- data_1 = T.match_buffer(data, (1, 512, 7, 7))
- kernel_1 = T.match_buffer(kernel, (512, 512, 3, 3))
- bias_1 = T.match_buffer(bias, (1, 512, 1, 1))
- compute_1 = T.match_buffer(compute, (1, 512, 7, 7))
blockIdx_x = T.env_thread("blockIdx.x")
- T.launch_thread(blockIdx_x, 16)
+ T.launch_thread(blockIdx_x, 28)
conv2d_nchw = T.allocate([14], "float32", "local")
- pad_temp_shared = T.allocate([2592], "float32", "shared")
- kernel_shared = T.allocate([9216], "float32", "shared")
+ pad_temp_shared = T.allocate([72], "float32", "shared")
+ kernel_shared = T.allocate([3072], "float32", "shared")
threadIdx_x = T.env_thread("threadIdx.x")
- T.launch_thread(threadIdx_x, 112)
+ T.launch_thread(threadIdx_x, 64)
conv2d_nchw_1 = T.buffer_decl((14,), data=conv2d_nchw, scope="local", align=32)
conv2d_nchw_1[0] = T.float32(0)
conv2d_nchw_1[1] = T.float32(0)
@@ -266,146 +262,460 @@ cooperative fetching, unrolling and operator fusion.
conv2d_nchw_1[11] = T.float32(0)
conv2d_nchw_1[12] = T.float32(0)
conv2d_nchw_1[13] = T.float32(0)
- for rc_outer_outer in range(16):
- cse_var_1: T.int32 = rc_outer_outer * 1568
+ for rc_outer_outer, ry_outer_outer in T.grid(64, 3):
+ cse_var_2: T.int32 = rc_outer_outer * 72
+ cse_var_1: T.int32 = ry_outer_outer * 3
threadIdx_x_1 = T.env_thread("threadIdx.x")
- pad_temp_shared_1 = T.buffer_decl((2592,), data=pad_temp_shared, scope="shared")
- data_2 = T.buffer_decl((25088,), data=data_1.data)
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1] = T.if_then_else(9 <= threadIdx_x_1 % 81 and threadIdx_x_1 % 81 < 72 and 1 <= threadIdx_x_1 % 9 and threadIdx_x_1 % 9 < 8, data_2[cse_var_1 + threadIdx_x_1 // 81 * 49 + threadIdx_x_1 % 81 // 9 * 7 + threadIdx_x_1 % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 112] = T.if_then_else(9 <= (threadIdx_x_1 + 31) % 81 and (threadIdx_x_1 + 31) % 81 < 72 and 1 <= (threadIdx_x_1 + 4) % 9 and (threadIdx_x_1 + 4) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 112) // 81 * 49 + (threadIdx_x_1 + 31) % 81 // 9 * 7 + (threadIdx_x_1 + 4) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 224] = T.if_then_else(9 <= (threadIdx_x_1 + 62) % 81 and (threadIdx_x_1 + 62) % 81 < 72 and 1 <= (threadIdx_x_1 + 8) % 9 and (threadIdx_x_1 + 8) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 224) // 81 * 49 + (threadIdx_x_1 + 62) % 81 // 9 * 7 + (threadIdx_x_1 + 8) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 336] = T.if_then_else(9 <= (threadIdx_x_1 + 12) % 81 and (threadIdx_x_1 + 12) % 81 < 72 and 1 <= (threadIdx_x_1 + 3) % 9 and (threadIdx_x_1 + 3) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 336) // 81 * 49 + (threadIdx_x_1 + 12) % 81 // 9 * 7 + (threadIdx_x_1 + 3) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 448] = T.if_then_else(9 <= (threadIdx_x_1 + 43) % 81 and (threadIdx_x_1 + 43) % 81 < 72 and 1 <= (threadIdx_x_1 + 7) % 9 and (threadIdx_x_1 + 7) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 448) // 81 * 49 + (threadIdx_x_1 + 43) % 81 // 9 * 7 + (threadIdx_x_1 + 7) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 560] = T.if_then_else(9 <= (threadIdx_x_1 + 74) % 81 and (threadIdx_x_1 + 74) % 81 < 72 and 1 <= (threadIdx_x_1 + 2) % 9 and (threadIdx_x_1 + 2) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 560) // 81 * 49 + (threadIdx_x_1 + 74) % 81 // 9 * 7 + (threadIdx_x_1 + 2) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 672] = T.if_then_else(9 <= (threadIdx_x_1 + 24) % 81 and (threadIdx_x_1 + 24) % 81 < 72 and 1 <= (threadIdx_x_1 + 6) % 9 and (threadIdx_x_1 + 6) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 672) // 81 * 49 + (threadIdx_x_1 + 24) % 81 // 9 * 7 + (threadIdx_x_1 + 6) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 784] = T.if_then_else(9 <= (threadIdx_x_1 + 55) % 81 and (threadIdx_x_1 + 55) % 81 < 72 and 1 <= (threadIdx_x_1 + 1) % 9 and (threadIdx_x_1 + 1) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 784) // 81 * 49 + (threadIdx_x_1 + 55) % 81 // 9 * 7 + (threadIdx_x_1 + 1) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 896] = T.if_then_else(9 <= (threadIdx_x_1 + 5) % 81 and (threadIdx_x_1 + 5) % 81 < 72 and 1 <= (threadIdx_x_1 + 5) % 9 and (threadIdx_x_1 + 5) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 896) // 81 * 49 + (threadIdx_x_1 + 5) % 81 // 9 * 7 + (threadIdx_x_1 + 5) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 1008] = T.if_then_else(1 <= (threadIdx_x_1 // 9 + 4) % 9 and (threadIdx_x_1 + 36) % 81 < 72 and 1 <= threadIdx_x_1 % 9 and threadIdx_x_1 % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 1008) // 81 * 49 + (threadIdx_x_1 // 9 + 4) % 9 * 7 + threadIdx_x_1 % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 1120] = T.if_then_else(9 <= (threadIdx_x_1 + 67) % 81 and (threadIdx_x_1 + 67) % 81 < 72 and 1 <= (threadIdx_x_1 + 4) % 9 and (threadIdx_x_1 + 4) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 1120) // 81 * 49 + (threadIdx_x_1 + 67) % 81 // 9 * 7 + (threadIdx_x_1 + 4) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 1232] = T.if_then_else(9 <= (threadIdx_x_1 + 17) % 81 and (threadIdx_x_1 + 17) % 81 < 72 and 1 <= (threadIdx_x_1 + 8) % 9 and (threadIdx_x_1 + 8) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 1232) // 81 * 49 + (threadIdx_x_1 + 17) % 81 // 9 * 7 + (threadIdx_x_1 + 8) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 1344] = T.if_then_else(9 <= (threadIdx_x_1 + 48) % 81 and (threadIdx_x_1 + 48) % 81 < 72 and 1 <= (threadIdx_x_1 + 3) % 9 and (threadIdx_x_1 + 3) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 1344) // 81 * 49 + (threadIdx_x_1 + 48) % 81 // 9 * 7 + (threadIdx_x_1 + 3) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 1456] = T.if_then_else(9 <= (threadIdx_x_1 + 79) % 81 and (threadIdx_x_1 + 79) % 81 < 72 and 1 <= (threadIdx_x_1 + 7) % 9 and (threadIdx_x_1 + 7) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 1456) // 81 * 49 + (threadIdx_x_1 + 79) % 81 // 9 * 7 + (threadIdx_x_1 + 7) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 1568] = T.if_then_else(9 <= (threadIdx_x_1 + 29) % 81 and (threadIdx_x_1 + 29) % 81 < 72 and 1 <= (threadIdx_x_1 + 2) % 9 and (threadIdx_x_1 + 2) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 1568) // 81 * 49 + (threadIdx_x_1 + 29) % 81 // 9 * 7 + (threadIdx_x_1 + 2) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 1680] = T.if_then_else(9 <= (threadIdx_x_1 + 60) % 81 and (threadIdx_x_1 + 60) % 81 < 72 and 1 <= (threadIdx_x_1 + 6) % 9 and (threadIdx_x_1 + 6) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 1680) // 81 * 49 + (threadIdx_x_1 + 60) % 81 // 9 * 7 + (threadIdx_x_1 + 6) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 1792] = T.if_then_else(9 <= (threadIdx_x_1 + 10) % 81 and (threadIdx_x_1 + 10) % 81 < 72 and 1 <= (threadIdx_x_1 + 1) % 9 and (threadIdx_x_1 + 1) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 1792) // 81 * 49 + (threadIdx_x_1 + 10) % 81 // 9 * 7 + (threadIdx_x_1 + 1) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 1904] = T.if_then_else(9 <= (threadIdx_x_1 + 41) % 81 and (threadIdx_x_1 + 41) % 81 < 72 and 1 <= (threadIdx_x_1 + 5) % 9 and (threadIdx_x_1 + 5) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 1904) // 81 * 49 + (threadIdx_x_1 + 41) % 81 // 9 * 7 + (threadIdx_x_1 + 5) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 2016] = T.if_then_else(1 <= (threadIdx_x_1 // 9 + 8) % 9 and (threadIdx_x_1 + 72) % 81 < 72 and 1 <= threadIdx_x_1 % 9 and threadIdx_x_1 % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 2016) // 81 * 49 + (threadIdx_x_1 // 9 + 8) % 9 * 7 + threadIdx_x_1 % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 2128] = T.if_then_else(9 <= (threadIdx_x_1 + 22) % 81 and (threadIdx_x_1 + 22) % 81 < 72 and 1 <= (threadIdx_x_1 + 4) % 9 and (threadIdx_x_1 + 4) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 2128) // 81 * 49 + (threadIdx_x_1 + 22) % 81 // 9 * 7 + (threadIdx_x_1 + 4) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 2240] = T.if_then_else(9 <= (threadIdx_x_1 + 53) % 81 and (threadIdx_x_1 + 53) % 81 < 72 and 1 <= (threadIdx_x_1 + 8) % 9 and (threadIdx_x_1 + 8) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 2240) // 81 * 49 + (threadIdx_x_1 + 53) % 81 // 9 * 7 + (threadIdx_x_1 + 8) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 2352] = T.if_then_else(9 <= (threadIdx_x_1 + 3) % 81 and (threadIdx_x_1 + 3) % 81 < 72 and 1 <= (threadIdx_x_1 + 3) % 9 and (threadIdx_x_1 + 3) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 2352) // 81 * 49 + (threadIdx_x_1 + 3) % 81 // 9 * 7 + (threadIdx_x_1 + 3) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 2464] = T.if_then_else(9 <= (threadIdx_x_1 + 34) % 81 and (threadIdx_x_1 + 34) % 81 < 72 and 1 <= (threadIdx_x_1 + 7) % 9 and (threadIdx_x_1 + 7) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 2464) // 81 * 49 + (threadIdx_x_1 + 34) % 81 // 9 * 7 + (threadIdx_x_1 + 7) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- if T.likely(threadIdx_x_1 < 16):
- pad_temp_shared_1[threadIdx_x_1 + 2576] = T.if_then_else(threadIdx_x_1 < 7 and 1 <= (threadIdx_x_1 + 2) % 9 and (threadIdx_x_1 + 2) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 2576) // 81 * 49 + (threadIdx_x_1 + 65) % 81 // 9 * 7 + (threadIdx_x_1 + 2) - 8], T.float32(0))
- kernel_shared_1 = T.buffer_decl((9216,), data=kernel_shared, scope="shared")
- for ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer in range(4):
- threadIdx_x_2 = T.env_thread("threadIdx.x")
- T.launch_thread(threadIdx_x_2, 112)
- kernel_2 = T.buffer_decl((2359296,), data=kernel_1.data)
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 32 + threadIdx_x_2 * 8) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2 + threadIdx_x_2 * 2) % 3 * 3]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 1] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 32 + threadIdx_x_2 * 8) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2 + threadIdx_x_2 * 2) % 3 * 3 + 1]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 2] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 32 + threadIdx_x_2 * 8) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2 + threadIdx_x_2 * 2) % 3 * 3 + 2]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 3] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 1) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 1) % 3 * 3]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 4] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 1) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 1) % 3 * 3 + 1]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 5] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 1) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 1) % 3 * 3 + 2]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 6] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 2) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 2) % 3 * 3]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 7] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 2) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 2) % 3 * 3 + 1]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 8] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 2) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 2) % 3 * 3 + 2]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 9] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + ((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8) // 3 + 1) % 32 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2 + threadIdx_x_2 * 2) % 3 * 3]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 10] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + ((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8) // 3 + 1) % 32 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2 + threadIdx_x_2 * 2) % 3 * 3 + 1]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 11] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + ((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8) // 3 + 1) % 32 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2 + threadIdx_x_2 * 2) % 3 * 3 + 2]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 12] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 4) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 1) % 3 * 3]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 13] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 4) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 1) % 3 * 3 + 1]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 14] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 4) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 1) % 3 * 3 + 2]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 15] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 5) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 2) % 3 * 3]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 16] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 5) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 2) % 3 * 3 + 1]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 17] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 5) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 2) % 3 * 3 + 2]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 18] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + ((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8) // 3 + 2) % 32 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2 + threadIdx_x_2 * 2) % 3 * 3]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 19] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + ((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8) // 3 + 2) % 32 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2 + threadIdx_x_2 * 2) % 3 * 3 + 1]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 20] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + ((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8) // 3 + 2) % 32 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2 + threadIdx_x_2 * 2) % 3 * 3 + 2]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 21] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 7) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 1) % 3 * 3]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 22] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 7) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 1) % 3 * 3 + 1]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 23] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 7) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 1) % 3 * 3 + 2]
- for rc_outer_inner, rx_outer_inner, ff_outer_inner, rc_inner in T.grid(2, 3, 2, 16):
- cse_var_8: T.int32 = ff_outer_inner * 7
- cse_var_7: T.int32 = cse_var_8 + 6
- cse_var_6: T.int32 = cse_var_8 + 5
- cse_var_5: T.int32 = cse_var_8 + 4
- cse_var_4: T.int32 = cse_var_8 + 3
- cse_var_3: T.int32 = cse_var_8 + 2
- cse_var_2: T.int32 = cse_var_8 + 1
- conv2d_nchw_1[cse_var_8] = conv2d_nchw_1[cse_var_8] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner]
- conv2d_nchw_1[cse_var_2] = conv2d_nchw_1[cse_var_2] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 9] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner]
- conv2d_nchw_1[cse_var_3] = conv2d_nchw_1[cse_var_3] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 18] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner]
- conv2d_nchw_1[cse_var_4] = conv2d_nchw_1[cse_var_4] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 27] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner]
- conv2d_nchw_1[cse_var_5] = conv2d_nchw_1[cse_var_5] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 36] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner]
- conv2d_nchw_1[cse_var_6] = conv2d_nchw_1[cse_var_6] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 45] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner]
- conv2d_nchw_1[cse_var_7] = conv2d_nchw_1[cse_var_7] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 54] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner]
- conv2d_nchw_1[cse_var_8] = conv2d_nchw_1[cse_var_8] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 9] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 3]
- conv2d_nchw_1[cse_var_2] = conv2d_nchw_1[cse_var_2] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 18] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 3]
- conv2d_nchw_1[cse_var_3] = conv2d_nchw_1[cse_var_3] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 27] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 3]
- conv2d_nchw_1[cse_var_4] = conv2d_nchw_1[cse_var_4] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 36] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 3]
- conv2d_nchw_1[cse_var_5] = conv2d_nchw_1[cse_var_5] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 45] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 3]
- conv2d_nchw_1[cse_var_6] = conv2d_nchw_1[cse_var_6] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 54] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 3]
- conv2d_nchw_1[cse_var_7] = conv2d_nchw_1[cse_var_7] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 63] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 3]
- conv2d_nchw_1[cse_var_8] = conv2d_nchw_1[cse_var_8] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 18] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 6]
- conv2d_nchw_1[cse_var_2] = conv2d_nchw_1[cse_var_2] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 27] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 6]
- conv2d_nchw_1[cse_var_3] = conv2d_nchw_1[cse_var_3] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 36] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 6]
- conv2d_nchw_1[cse_var_4] = conv2d_nchw_1[cse_var_4] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 45] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 6]
- conv2d_nchw_1[cse_var_5] = conv2d_nchw_1[cse_var_5] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 54] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 6]
- conv2d_nchw_1[cse_var_6] = conv2d_nchw_1[cse_var_6] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 63] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 6]
- conv2d_nchw_1[cse_var_7] = conv2d_nchw_1[cse_var_7] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 72] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 6]
- for i1_inner, i2_inner in T.grid(2, 7):
- compute_2 = T.buffer_decl((25088,), data=compute_1.data)
- bias_2 = T.buffer_decl((512,), data=bias_1.data)
- compute_2[blockIdx_x * 1568 + threadIdx_x // 7 * 98 + i1_inner * 49 + i2_inner * 7 + threadIdx_x % 7] = T.max(conv2d_nchw_1[i1_inner * 7 + i2_inner] + bias_2[blockIdx_x * 32 + threadIdx_x // 7 * 2 + i1_inner], T.float32(0))
+ pad_temp_shared_1 = T.buffer_decl((72,), data=pad_temp_shared, scope="shared")
+ with T.launch_thread(threadIdx_x_1, 64):
+ data_1 = T.buffer_decl((25088,), data=data.data)
+ if T.likely(threadIdx_x_1 < 18):
+ pad_temp_shared_1[threadIdx_x_1 * 4] = T.if_then_else(1 <= ry_outer_outer + blockIdx_x % 7 and ry_outer_outer + blockIdx_x % 7 < 8 and 1 <= threadIdx_x_1 * 4 % 9 and threadIdx_x_1 * 4 % 9 < 8, data_1[rc_outer_outer * 392 + threadIdx_x_1 * 4 // 9 * 49 + ry_outer_outer * 7 + blockIdx_x % 7 * 7 + threadIdx_x_1 * 4 % 9 - 8], T.float32(0))
+ if T.likely(threadIdx_x_1 < 18):
+ pad_temp_shared_1[threadIdx_x_1 * 4 + 1] = T.if_then_else(1 <= ry_outer_outer + blockIdx_x % 7 and ry_outer_outer + blockIdx_x % 7 < 8 and 1 <= (threadIdx_x_1 * 4 + 1) % 9 and (threadIdx_x_1 * 4 + 1) % 9 < 8, data_1[rc_outer_outer * 392 + (threadIdx_x_1 * 4 + 1) // 9 * 49 + ry_outer_outer * 7 + blockIdx_x % 7 * 7 + (threadIdx_x_1 * 4 + 1) % 9 - 8], T.float32(0))
+ if T.likely(threadIdx_x_1 < 18):
+ pad_temp_shared_1[threadIdx_x_1 * 4 + 2] = T.if_then_else(1 <= ry_outer_outer + blockIdx_x % 7 and ry_outer_outer + blockIdx_x % 7 < 8 and 1 <= (threadIdx_x_1 * 4 + 2) % 9 and (threadIdx_x_1 * 4 + 2) % 9 < 8, data_1[rc_outer_outer * 392 + (threadIdx_x_1 * 4 + 2) // 9 * 49 + ry_outer_outer * 7 + blockIdx_x % 7 * 7 + (threadIdx_x_1 * 4 + 2) % 9 - 8], T.float32(0))
+ if T.likely(threadIdx_x_1 < 18):
+ pad_temp_shared_1[threadIdx_x_1 * 4 + 3] = T.if_then_else(1 <= ry_outer_outer + blockIdx_x % 7 and ry_outer_outer + blockIdx_x % 7 < 8 and 1 <= (threadIdx_x_1 * 4 + 3) % 9 and (threadIdx_x_1 * 4 + 3) % 9 < 8, data_1[rc_outer_outer * 392 + (threadIdx_x_1 * 4 + 3) // 9 * 49 + ry_outer_outer * 7 + blockIdx_x % 7 * 7 + (threadIdx_x_1 * 4 + 3) % 9 - 8], T.float32(0))
+ threadIdx_x_2 = T.env_thread("threadIdx.x")
+ kernel_shared_1 = T.buffer_decl((3072,), data=kernel_shared, scope="shared")
+ kernel_1 = T.buffer_decl((2359296,), data=kernel.data)
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 64] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 64) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 128] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 128) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 192] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 36864]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 256] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 256) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 320] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 320) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 384] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 73728]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 448] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 448) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 512] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 512) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 576] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 110592]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 640] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 640) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 704] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 704) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 768] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 147456]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 832] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 832) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 896] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 896) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 960] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 184320]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1024] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 1024) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1088] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 1088) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1152] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 221184]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1216] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 1216) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1280] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 1280) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1344] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 258048]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1408] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 1408) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1472] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 1472) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1536] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 294912]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1600] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 1600) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1664] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 1664) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1728] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 331776]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1792] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 1792) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1856] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 1856) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1920] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 368640]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1984] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 1984) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2048] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 2048) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2112] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 405504]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2176] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 2176) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2240] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 2240) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2304] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 442368]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2368] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 2368) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2432] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 2432) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2496] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 479232]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2560] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 2560) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2624] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 2624) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2688] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 516096]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2752] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 2752) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2816] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 2816) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2880] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 552960]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2944] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 2944) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 3008] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 3008) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (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]
+ for i1_inner, i3_inner in T.grid(2, 7):
+ compute_1 = T.buffer_decl((25088,), data=compute.data)
+ bias_1 = T.buffer_decl((512,), data=bias.data)
+ compute_1[blockIdx_x // 7 * 6272 + threadIdx_x * 98 + i1_inner * 49 + blockIdx_x % 7 * 7 + i3_inner] = T.max(conv2d_nchw_1[i1_inner * 7 + i3_inner] + bias_1[blockIdx_x // 7 * 128 + threadIdx_x * 2 + i1_inner], T.float32(0))
@@ -455,7 +765,7 @@ We build the binary and check its correctness and performance.
.. code-block:: none
- Execution time of this operator: 0.325 ms
+ Execution time of this operator: 0.355 ms
@@ -505,19 +815,19 @@ They can be used for debugging and learning the behavior of the auto-scheduler.
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=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=16)
+ 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=7)
+ 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=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=16)
- 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=1)
conv2d_nchw_rx_o_o, conv2d_nchw_rx_o_i = s[conv2d_nchw].split(conv2d_nchw_rx_o_i, factor=3)
@@ -526,13 +836,13 @@ They can be used for debugging and learning the behavior of the auto-scheduler.
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=2)
- compute_i1_o_o_i, compute_i1_o_i = s[compute].split(compute_i1_o_i, factor=16)
+ 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=7)
+ 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=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)
@@ -550,16 +860,16 @@ They can be used for debugging and learning the behavior of the auto-scheduler.
compute_i0_o_i_i1_o_i_fused_i2_o_i_fused_i3_o_i_fused = s[compute].fuse(compute_i0_o_i, compute_i1_o_i, compute_i2_o_i, compute_i3_o_i)
s[compute].bind(compute_i0_o_i_i1_o_i_fused_i2_o_i_fused_i3_o_i_fused, te.thread_axis("threadIdx.x"))
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=24)
+ 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=112)
+ 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=112)
+ 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", 64)
+ 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)
CUDA source code:
@@ -577,10 +887,10 @@ 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__(112) default_function_kernel0(float* __restrict__ data, float* __restrict__ kernel, float* __restrict__ compute, float* __restrict__ bias) {
+ 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[2592];
- __shared__ float kernel_shared[9216];
+ __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;
@@ -595,142 +905,411 @@ They can be used for debugging and learning the behavior of the auto-scheduler.
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 < 16; ++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 * 1568) + ((((int)threadIdx.x) / 81) * 49)) + (((((int)threadIdx.x) % 81) / 9) * 7)) + (((int)threadIdx.x) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 112)] = (((((9 <= ((((int)threadIdx.x) + 31) % 81)) && (((((int)threadIdx.x) + 31) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 4) % 9))) && (((((int)threadIdx.x) + 4) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 112) / 81) * 49)) + ((((((int)threadIdx.x) + 31) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 4) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 224)] = (((((9 <= ((((int)threadIdx.x) + 62) % 81)) && (((((int)threadIdx.x) + 62) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 8) % 9))) && (((((int)threadIdx.x) + 8) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 224) / 81) * 49)) + ((((((int)threadIdx.x) + 62) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 8) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 336)] = (((((9 <= ((((int)threadIdx.x) + 12) % 81)) && (((((int)threadIdx.x) + 12) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 3) % 9))) && (((((int)threadIdx.x) + 3) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 336) / 81) * 49)) + ((((((int)threadIdx.x) + 12) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 3) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 448)] = (((((9 <= ((((int)threadIdx.x) + 43) % 81)) && (((((int)threadIdx.x) + 43) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 7) % 9))) && (((((int)threadIdx.x) + 7) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 448) / 81) * 49)) + ((((((int)threadIdx.x) + 43) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 7) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 560)] = (((((9 <= ((((int)threadIdx.x) + 74) % 81)) && (((((int)threadIdx.x) + 74) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 2) % 9))) && (((((int)threadIdx.x) + 2) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 560) / 81) * 49)) + ((((((int)threadIdx.x) + 74) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 2) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 672)] = (((((9 <= ((((int)threadIdx.x) + 24) % 81)) && (((((int)threadIdx.x) + 24) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 6) % 9))) && (((((int)threadIdx.x) + 6) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 672) / 81) * 49)) + ((((((int)threadIdx.x) + 24) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 6) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 784)] = (((((9 <= ((((int)threadIdx.x) + 55) % 81)) && (((((int)threadIdx.x) + 55) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 1) % 9))) && (((((int)threadIdx.x) + 1) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 784) / 81) * 49)) + ((((((int)threadIdx.x) + 55) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 1) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 896)] = (((((9 <= ((((int)threadIdx.x) + 5) % 81)) && (((((int)threadIdx.x) + 5) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 5) % 9))) && (((((int)threadIdx.x) + 5) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 896) / 81) * 49)) + ((((((int)threadIdx.x) + 5) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 5) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 1008)] = (((((1 <= (((((int)threadIdx.x) / 9) + 4) % 9)) && (((((int)threadIdx.x) + 36) % 81) < 72)) && (1 <= (((int)threadIdx.x) % 9))) && ((((int)threadIdx.x) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 1008) / 81) * 49)) + ((((((int)threadIdx.x) / 9) + 4) % 9) * 7)) + (((int)threadIdx.x) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 1120)] = (((((9 <= ((((int)threadIdx.x) + 67) % 81)) && (((((int)threadIdx.x) + 67) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 4) % 9))) && (((((int)threadIdx.x) + 4) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 1120) / 81) * 49)) + ((((((int)threadIdx.x) + 67) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 4) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 1232)] = (((((9 <= ((((int)threadIdx.x) + 17) % 81)) && (((((int)threadIdx.x) + 17) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 8) % 9))) && (((((int)threadIdx.x) + 8) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 1232) / 81) * 49)) + ((((((int)threadIdx.x) + 17) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 8) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 1344)] = (((((9 <= ((((int)threadIdx.x) + 48) % 81)) && (((((int)threadIdx.x) + 48) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 3) % 9))) && (((((int)threadIdx.x) + 3) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 1344) / 81) * 49)) + ((((((int)threadIdx.x) + 48) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 3) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 1456)] = (((((9 <= ((((int)threadIdx.x) + 79) % 81)) && (((((int)threadIdx.x) + 79) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 7) % 9))) && (((((int)threadIdx.x) + 7) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 1456) / 81) * 49)) + ((((((int)threadIdx.x) + 79) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 7) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 1568)] = (((((9 <= ((((int)threadIdx.x) + 29) % 81)) && (((((int)threadIdx.x) + 29) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 2) % 9))) && (((((int)threadIdx.x) + 2) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 1568) / 81) * 49)) + ((((((int)threadIdx.x) + 29) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 2) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 1680)] = (((((9 <= ((((int)threadIdx.x) + 60) % 81)) && (((((int)threadIdx.x) + 60) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 6) % 9))) && (((((int)threadIdx.x) + 6) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 1680) / 81) * 49)) + ((((((int)threadIdx.x) + 60) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 6) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 1792)] = (((((9 <= ((((int)threadIdx.x) + 10) % 81)) && (((((int)threadIdx.x) + 10) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 1) % 9))) && (((((int)threadIdx.x) + 1) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 1792) / 81) * 49)) + ((((((int)threadIdx.x) + 10) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 1) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 1904)] = (((((9 <= ((((int)threadIdx.x) + 41) % 81)) && (((((int)threadIdx.x) + 41) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 5) % 9))) && (((((int)threadIdx.x) + 5) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 1904) / 81) * 49)) + ((((((int)threadIdx.x) + 41) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 5) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 2016)] = (((((1 <= (((((int)threadIdx.x) / 9) + 8) % 9)) && (((((int)threadIdx.x) + 72) % 81) < 72)) && (1 <= (((int)threadIdx.x) % 9))) && ((((int)threadIdx.x) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 2016) / 81) * 49)) + ((((((int)threadIdx.x) / 9) + 8) % 9) * 7)) + (((int)threadIdx.x) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 2128)] = (((((9 <= ((((int)threadIdx.x) + 22) % 81)) && (((((int)threadIdx.x) + 22) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 4) % 9))) && (((((int)threadIdx.x) + 4) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 2128) / 81) * 49)) + ((((((int)threadIdx.x) + 22) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 4) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 2240)] = (((((9 <= ((((int)threadIdx.x) + 53) % 81)) && (((((int)threadIdx.x) + 53) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 8) % 9))) && (((((int)threadIdx.x) + 8) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 2240) / 81) * 49)) + ((((((int)threadIdx.x) + 53) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 8) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 2352)] = (((((9 <= ((((int)threadIdx.x) + 3) % 81)) && (((((int)threadIdx.x) + 3) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 3) % 9))) && (((((int)threadIdx.x) + 3) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 2352) / 81) * 49)) + ((((((int)threadIdx.x) + 3) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 3) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 2464)] = (((((9 <= ((((int)threadIdx.x) + 34) % 81)) && (((((int)threadIdx.x) + 34) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 7) % 9))) && (((((int)threadIdx.x) + 7) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 2464) / 81) * 49)) + ((((((int)threadIdx.x) + 34) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 7) % 9)) - 8)] : 0.000000e+00f);
- if (((int)threadIdx.x) < 16) {
- pad_temp_shared[(((int)threadIdx.x) + 2576)] = ((((((int)threadIdx.x) < 7) && (1 <= ((((int)threadIdx.x) + 2) % 9))) && (((((int)threadIdx.x) + 2) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 2576) / 81) * 49)) + (((((int)threadIdx.x) + 65) / 9) * 7)) + ((int)threadIdx.x)) - 6)] : 0.000000e+00f);
- }
- for (int ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer = 0; ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer < 4; ++ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer) {
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24))] = kernel[(((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 32) + (((int)threadIdx.x) * 8)) % 96) / 3) * 9)) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2) + (((int)threadIdx.x) * 2)) % 3) * 3))];
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 1)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 32) + (((int)threadIdx.x) * 8)) % 96) / 3) * 9)) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2) + (((int)threadIdx.x) * 2)) % 3) * 3)) + 1)];
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 2)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 32) + (((int)threadIdx.x) * 8)) % 96) / 3) * 9)) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2) + (((int)threadIdx.x) * 2)) % 3) * 3)) + 2)];
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 3)] = kernel[(((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 1) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 1) % [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 4)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 1) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 1) [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 5)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 1) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 1) [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 6)] = kernel[(((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 2) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 2) % [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 7)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 2) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 2) [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 8)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 2) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 2) [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 9)] = kernel[(((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) / 3) + 1) & 31) * 9)) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2) + (((int)threadIdx.x) * 2)) % 3) * 3))];
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 10)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) / 3) + 1) & 31) * 9)) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2) + (((int)threadIdx.x) * 2)) % 3) * [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 11)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) / 3) + 1) & 31) * 9)) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2) + (((int)threadIdx.x) * 2)) % 3) * [...]
+ 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 (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 12)] = kernel[(((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 4) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 1) [...]
+ 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 (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 13)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 4) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 1) [...]
+ 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 (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 14)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 4) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 1) [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 15)] = kernel[(((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 5) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 2) [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 16)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 5) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 2) [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 17)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 5) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 2) [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 18)] = kernel[(((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) / 3) + 2) & 31) * 9)) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2) + (((int)threadIdx.x) * 2)) % 3) * 3))];
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 19)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) / 3) + 2) & 31) * 9)) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2) + (((int)threadIdx.x) * 2)) % 3) * [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 20)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) / 3) + 2) & 31) * 9)) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2) + (((int)threadIdx.x) * 2)) % 3) * [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 21)] = kernel[(((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 7) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 1) [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 22)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 7) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 1) [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 23)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 7) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 1) [...]
- }
- }
- __syncthreads();
- for (int rc_outer_inner = 0; rc_outer_inner < 2; ++rc_outer_inner) {
- for (int rx_outer_inner = 0; rx_outer_inner < 3; ++rx_outer_inner) {
- for (int ff_outer_inner = 0; ff_outer_inner < 2; ++ff_outer_inner) {
- for (int rc_inner = 0; rc_inner < 16; ++rc_inner) {
- conv2d_nchw[(ff_outer_inner * 7)] = (conv2d_nchw[(ff_outer_inner * 7)] + (pad_temp_shared[((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7))] * kernel_shared[((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner)]));
- conv2d_nchw[((ff_outer_inner * 7) + 1)] = (conv2d_nchw[((ff_outer_inner * 7) + 1)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 9)] * kernel_shared[((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner)]));
- conv2d_nchw[((ff_outer_inner * 7) + 2)] = (conv2d_nchw[((ff_outer_inner * 7) + 2)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 18)] * kernel_shared[((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner)]));
- conv2d_nchw[((ff_outer_inner * 7) + 3)] = (conv2d_nchw[((ff_outer_inner * 7) + 3)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 27)] * kernel_shared[((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner)]));
- conv2d_nchw[((ff_outer_inner * 7) + 4)] = (conv2d_nchw[((ff_outer_inner * 7) + 4)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 36)] * kernel_shared[((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner)]));
- conv2d_nchw[((ff_outer_inner * 7) + 5)] = (conv2d_nchw[((ff_outer_inner * 7) + 5)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 45)] * kernel_shared[((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner)]));
- conv2d_nchw[((ff_outer_inner * 7) + 6)] = (conv2d_nchw[((ff_outer_inner * 7) + 6)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 54)] * kernel_shared[((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner)]));
- conv2d_nchw[(ff_outer_inner * 7)] = (conv2d_nchw[(ff_outer_inner * 7)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 9)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 3)]));
- conv2d_nchw[((ff_outer_inner * 7) + 1)] = (conv2d_nchw[((ff_outer_inner * 7) + 1)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 18)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 3)]));
- conv2d_nchw[((ff_outer_inner * 7) + 2)] = (conv2d_nchw[((ff_outer_inner * 7) + 2)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 27)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 3)]));
- conv2d_nchw[((ff_outer_inner * 7) + 3)] = (conv2d_nchw[((ff_outer_inner * 7) + 3)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 36)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 3)]));
- conv2d_nchw[((ff_outer_inner * 7) + 4)] = (conv2d_nchw[((ff_outer_inner * 7) + 4)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 45)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 3)]));
- conv2d_nchw[((ff_outer_inner * 7) + 5)] = (conv2d_nchw[((ff_outer_inner * 7) + 5)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 54)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 3)]));
- conv2d_nchw[((ff_outer_inner * 7) + 6)] = (conv2d_nchw[((ff_outer_inner * 7) + 6)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 63)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 3)]));
- conv2d_nchw[(ff_outer_inner * 7)] = (conv2d_nchw[(ff_outer_inner * 7)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 18)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 6)]));
- conv2d_nchw[((ff_outer_inner * 7) + 1)] = (conv2d_nchw[((ff_outer_inner * 7) + 1)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 27)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 6)]));
- conv2d_nchw[((ff_outer_inner * 7) + 2)] = (conv2d_nchw[((ff_outer_inner * 7) + 2)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 36)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 6)]));
- conv2d_nchw[((ff_outer_inner * 7) + 3)] = (conv2d_nchw[((ff_outer_inner * 7) + 3)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 45)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 6)]));
- conv2d_nchw[((ff_outer_inner * 7) + 4)] = (conv2d_nchw[((ff_outer_inner * 7) + 4)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 54)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 6)]));
- conv2d_nchw[((ff_outer_inner * 7) + 5)] = (conv2d_nchw[((ff_outer_inner * 7) + 5)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 63)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 6)]));
- conv2d_nchw[((ff_outer_inner * 7) + 6)] = (conv2d_nchw[((ff_outer_inner * 7) + 6)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 72)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 6)]));
- }
- }
+ 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)]));
}
}
for (int i1_inner = 0; i1_inner < 2; ++i1_inner) {
- for (int i2_inner = 0; i2_inner < 7; ++i2_inner) {
- compute[(((((((int)blockIdx.x) * 1568) + ((((int)threadIdx.x) / 7) * 98)) + (i1_inner * 49)) + (i2_inner * 7)) + (((int)threadIdx.x) % 7))] = max((conv2d_nchw[((i1_inner * 7) + i2_inner)] + bias[(((((int)blockIdx.x) * 32) + ((((int)threadIdx.x) / 7) * 2)) + i1_inner)]), 0.000000e+00f);
+ 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);
}
}
}
@@ -793,7 +1372,7 @@ 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 46.341 seconds)
+ **Total running time of the script:** ( 5 minutes 27.031 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 ec56826e0b..1bab3f78c4 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
@@ -647,7 +647,7 @@ 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)
- 7.8969 7.8932 7.9050 7.8925 0.0057
+ 7.8663 7.8661 7.8708 7.8619 0.0036
@@ -675,7 +675,7 @@ Other Tips
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 1 minutes 5.810 seconds)
+ **Total running time of the script:** ( 1 minutes 4.732 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 3b43fde243..fa4ff17b5e 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
@@ -666,7 +666,7 @@ 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)
- 751.3712 751.1382 753.5222 749.4532 1.6693
+ 750.0263 750.0990 750.6909 749.2892 0.5746
@@ -694,7 +694,7 @@ Other Tips
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 1 minutes 38.610 seconds)
+ **Total running time of the script:** ( 1 minutes 36.571 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 6e2dc06608..f541801123 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
@@ -384,82 +384,409 @@ layout transformation, parallelization, vectorization, unrolling, and operator f
@I.ir_module
class Module:
@T.prim_func
- def main(placeholder: T.handle, placeholder_1: T.handle, placeholder_2: T.handle, placeholder_3: T.handle, placeholder_4: T.handle, compute: T.handle):
+ def main(placeholder: T.Buffer((128, 256), "float32"), placeholder_1: T.Buffer((4916, 16, 1), "float32"), placeholder_2: T.Buffer((4916,), "int32"), placeholder_3: T.Buffer((33,), "int32"), placeholder_4: T.Buffer((128, 512), "float32"), compute: T.Buffer((128, 512), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- placeholder_5 = T.match_buffer(placeholder, (128, 256))
- placeholder_6 = T.match_buffer(placeholder_1, (4916, 16, 1))
- placeholder_7 = T.match_buffer(placeholder_2, (4916,), "int32")
- placeholder_8 = T.match_buffer(placeholder_3, (33,), "int32")
- placeholder_9 = T.match_buffer(placeholder_4, (128, 512))
- compute_1 = T.match_buffer(compute, (128, 512))
for i0_outer_i1_outer_fused in T.parallel(128):
- compute_2 = T.allocate([512], "float32", "global")
- compute_3 = T.buffer_decl((512,), data=compute_2)
+ compute_1 = T.allocate([512], "float32", "global")
+ compute_2 = T.buffer_decl((512,), data=compute_1)
for i_outer_inner, nb_j_inner in T.grid(2, 2):
- for i_inner_init in range(8):
- cse_var_1: T.int32 = i_outer_inner * 256 + i_inner_init * 32 + nb_j_inner * 16
- compute_3[cse_var_1] = T.float32(0)
- compute_3[cse_var_1 + 1] = T.float32(0)
- compute_3[cse_var_1 + 2] = T.float32(0)
- compute_3[cse_var_1 + 3] = T.float32(0)
- compute_3[cse_var_1 + 4] = T.float32(0)
- compute_3[cse_var_1 + 5] = T.float32(0)
- compute_3[cse_var_1 + 6] = T.float32(0)
- compute_3[cse_var_1 + 7] = T.float32(0)
- compute_3[cse_var_1 + 8] = T.float32(0)
- compute_3[cse_var_1 + 9] = T.float32(0)
- compute_3[cse_var_1 + 10] = T.float32(0)
- compute_3[cse_var_1 + 11] = T.float32(0)
- compute_3[cse_var_1 + 12] = T.float32(0)
- compute_3[cse_var_1 + 13] = T.float32(0)
- compute_3[cse_var_1 + 14] = T.float32(0)
- compute_3[cse_var_1 + 15] = T.float32(0)
- for elem_idx, i_inner in T.grid(T.let(cse_var_2, i0_outer_i1_outer_fused % 16 * 2 + nb_j_inner, placeholder_10[cse_var_2 + 1] - placeholder_10[cse_var_2]), 8):
- cse_var_2 = T.var("int32")
- placeholder_10 = T.buffer_decl((33,), "int32", data=placeholder_8.data)
- cse_var_21: T.int32 = elem_idx * 16
- cse_var_20: T.int32 = i0_outer_i1_outer_fused % 16 * 2 + nb_j_inner
- cse_var_19: T.int32 = i_outer_inner * 256 + i_inner * 32 + nb_j_inner * 16
- cse_var_18: T.int32 = i0_outer_i1_outer_fused // 16 * 4096 + i_outer_inner * 2048 + i_inner * 256
- cse_var_17: T.int32 = cse_var_19 + 9
- cse_var_16: T.int32 = cse_var_19 + 8
- cse_var_15: T.int32 = cse_var_19 + 7
- cse_var_14: T.int32 = cse_var_19 + 6
- cse_var_13: T.int32 = cse_var_19 + 5
- cse_var_12: T.int32 = cse_var_19 + 4
- cse_var_11: T.int32 = cse_var_19 + 3
- cse_var_10: T.int32 = cse_var_19 + 2
- cse_var_9: T.int32 = cse_var_19 + 15
- cse_var_8: T.int32 = cse_var_19 + 14
- cse_var_7: T.int32 = cse_var_19 + 13
- cse_var_6: T.int32 = cse_var_19 + 12
- cse_var_5: T.int32 = cse_var_19 + 11
- cse_var_4: T.int32 = cse_var_19 + 10
- cse_var_3: T.int32 = cse_var_19 + 1
- placeholder_11 = T.buffer_decl((78656,), data=placeholder_6.data)
- placeholder_12 = T.buffer_decl((32768,), data=placeholder_5.data)
- placeholder_13 = T.buffer_decl((4916,), "int32", data=placeholder_7.data)
- compute_3[cse_var_19] = compute_3[cse_var_19] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_3] = compute_3[cse_var_3] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 1] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_10] = compute_3[cse_var_10] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 2] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_11] = compute_3[cse_var_11] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 3] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_12] = compute_3[cse_var_12] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 4] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_13] = compute_3[cse_var_13] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 5] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_14] = compute_3[cse_var_14] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 6] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_15] = compute_3[cse_var_15] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 7] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_16] = compute_3[cse_var_16] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 8] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_17] = compute_3[cse_var_17] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 9] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_4] = compute_3[cse_var_4] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 10] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_5] = compute_3[cse_var_5] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 11] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_6] = compute_3[cse_var_6] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 12] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_7] = compute_3[cse_var_7] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 13] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_8] = compute_3[cse_var_8] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 14] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_9] = compute_3[cse_var_9] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 15] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
+ cse_var_2: T.int32 = i_outer_inner * 256 + nb_j_inner * 16
+ cse_var_1: T.int32 = i0_outer_i1_outer_fused % 16 * 2 + nb_j_inner
+ compute_2[cse_var_2] = T.float32(0)
+ compute_2[cse_var_2 + 1] = T.float32(0)
+ compute_2[cse_var_2 + 2] = T.float32(0)
+ compute_2[cse_var_2 + 3] = T.float32(0)
+ compute_2[cse_var_2 + 4] = T.float32(0)
+ compute_2[cse_var_2 + 5] = T.float32(0)
+ compute_2[cse_var_2 + 6] = T.float32(0)
+ compute_2[cse_var_2 + 7] = T.float32(0)
+ compute_2[cse_var_2 + 8] = T.float32(0)
+ compute_2[cse_var_2 + 9] = T.float32(0)
+ compute_2[cse_var_2 + 10] = T.float32(0)
+ compute_2[cse_var_2 + 11] = T.float32(0)
+ compute_2[cse_var_2 + 12] = T.float32(0)
+ compute_2[cse_var_2 + 13] = T.float32(0)
+ compute_2[cse_var_2 + 14] = T.float32(0)
+ compute_2[cse_var_2 + 15] = T.float32(0)
+ compute_2[cse_var_2 + 32] = T.float32(0)
+ compute_2[cse_var_2 + 33] = T.float32(0)
+ compute_2[cse_var_2 + 34] = T.float32(0)
+ compute_2[cse_var_2 + 35] = T.float32(0)
+ compute_2[cse_var_2 + 36] = T.float32(0)
+ compute_2[cse_var_2 + 37] = T.float32(0)
+ compute_2[cse_var_2 + 38] = T.float32(0)
+ compute_2[cse_var_2 + 39] = T.float32(0)
+ compute_2[cse_var_2 + 40] = T.float32(0)
+ compute_2[cse_var_2 + 41] = T.float32(0)
+ compute_2[cse_var_2 + 42] = T.float32(0)
+ compute_2[cse_var_2 + 43] = T.float32(0)
+ compute_2[cse_var_2 + 44] = T.float32(0)
+ compute_2[cse_var_2 + 45] = T.float32(0)
+ compute_2[cse_var_2 + 46] = T.float32(0)
+ compute_2[cse_var_2 + 47] = T.float32(0)
+ compute_2[cse_var_2 + 64] = T.float32(0)
+ compute_2[cse_var_2 + 65] = T.float32(0)
+ compute_2[cse_var_2 + 66] = T.float32(0)
+ compute_2[cse_var_2 + 67] = T.float32(0)
+ compute_2[cse_var_2 + 68] = T.float32(0)
+ compute_2[cse_var_2 + 69] = T.float32(0)
+ compute_2[cse_var_2 + 70] = T.float32(0)
+ compute_2[cse_var_2 + 71] = T.float32(0)
+ compute_2[cse_var_2 + 72] = T.float32(0)
+ compute_2[cse_var_2 + 73] = T.float32(0)
+ compute_2[cse_var_2 + 74] = T.float32(0)
+ compute_2[cse_var_2 + 75] = T.float32(0)
+ compute_2[cse_var_2 + 76] = T.float32(0)
+ compute_2[cse_var_2 + 77] = T.float32(0)
+ compute_2[cse_var_2 + 78] = T.float32(0)
+ compute_2[cse_var_2 + 79] = T.float32(0)
+ compute_2[cse_var_2 + 96] = T.float32(0)
+ compute_2[cse_var_2 + 97] = T.float32(0)
+ compute_2[cse_var_2 + 98] = T.float32(0)
+ compute_2[cse_var_2 + 99] = T.float32(0)
+ compute_2[cse_var_2 + 100] = T.float32(0)
+ compute_2[cse_var_2 + 101] = T.float32(0)
+ compute_2[cse_var_2 + 102] = T.float32(0)
+ compute_2[cse_var_2 + 103] = T.float32(0)
+ compute_2[cse_var_2 + 104] = T.float32(0)
+ compute_2[cse_var_2 + 105] = T.float32(0)
+ compute_2[cse_var_2 + 106] = T.float32(0)
+ compute_2[cse_var_2 + 107] = T.float32(0)
+ compute_2[cse_var_2 + 108] = T.float32(0)
+ compute_2[cse_var_2 + 109] = T.float32(0)
+ compute_2[cse_var_2 + 110] = T.float32(0)
+ compute_2[cse_var_2 + 111] = T.float32(0)
+ compute_2[cse_var_2 + 128] = T.float32(0)
+ compute_2[cse_var_2 + 129] = T.float32(0)
+ compute_2[cse_var_2 + 130] = T.float32(0)
+ compute_2[cse_var_2 + 131] = T.float32(0)
+ compute_2[cse_var_2 + 132] = T.float32(0)
+ compute_2[cse_var_2 + 133] = T.float32(0)
+ compute_2[cse_var_2 + 134] = T.float32(0)
+ compute_2[cse_var_2 + 135] = T.float32(0)
+ compute_2[cse_var_2 + 136] = T.float32(0)
+ compute_2[cse_var_2 + 137] = T.float32(0)
+ compute_2[cse_var_2 + 138] = T.float32(0)
+ compute_2[cse_var_2 + 139] = T.float32(0)
+ compute_2[cse_var_2 + 140] = T.float32(0)
+ compute_2[cse_var_2 + 141] = T.float32(0)
+ compute_2[cse_var_2 + 142] = T.float32(0)
+ compute_2[cse_var_2 + 143] = T.float32(0)
+ compute_2[cse_var_2 + 160] = T.float32(0)
+ compute_2[cse_var_2 + 161] = T.float32(0)
+ compute_2[cse_var_2 + 162] = T.float32(0)
+ compute_2[cse_var_2 + 163] = T.float32(0)
+ compute_2[cse_var_2 + 164] = T.float32(0)
+ compute_2[cse_var_2 + 165] = T.float32(0)
+ compute_2[cse_var_2 + 166] = T.float32(0)
+ compute_2[cse_var_2 + 167] = T.float32(0)
+ compute_2[cse_var_2 + 168] = T.float32(0)
+ compute_2[cse_var_2 + 169] = T.float32(0)
+ compute_2[cse_var_2 + 170] = T.float32(0)
+ compute_2[cse_var_2 + 171] = T.float32(0)
+ compute_2[cse_var_2 + 172] = T.float32(0)
+ compute_2[cse_var_2 + 173] = T.float32(0)
+ compute_2[cse_var_2 + 174] = T.float32(0)
+ compute_2[cse_var_2 + 175] = T.float32(0)
+ compute_2[cse_var_2 + 192] = T.float32(0)
+ compute_2[cse_var_2 + 193] = T.float32(0)
+ compute_2[cse_var_2 + 194] = T.float32(0)
+ compute_2[cse_var_2 + 195] = T.float32(0)
+ compute_2[cse_var_2 + 196] = T.float32(0)
+ compute_2[cse_var_2 + 197] = T.float32(0)
+ compute_2[cse_var_2 + 198] = T.float32(0)
+ compute_2[cse_var_2 + 199] = T.float32(0)
+ compute_2[cse_var_2 + 200] = T.float32(0)
+ compute_2[cse_var_2 + 201] = T.float32(0)
+ compute_2[cse_var_2 + 202] = T.float32(0)
+ compute_2[cse_var_2 + 203] = T.float32(0)
+ compute_2[cse_var_2 + 204] = T.float32(0)
+ compute_2[cse_var_2 + 205] = T.float32(0)
+ compute_2[cse_var_2 + 206] = T.float32(0)
+ compute_2[cse_var_2 + 207] = T.float32(0)
+ compute_2[cse_var_2 + 224] = T.float32(0)
+ compute_2[cse_var_2 + 225] = T.float32(0)
+ compute_2[cse_var_2 + 226] = T.float32(0)
+ compute_2[cse_var_2 + 227] = T.float32(0)
+ compute_2[cse_var_2 + 228] = T.float32(0)
+ compute_2[cse_var_2 + 229] = T.float32(0)
+ compute_2[cse_var_2 + 230] = T.float32(0)
+ compute_2[cse_var_2 + 231] = T.float32(0)
+ compute_2[cse_var_2 + 232] = T.float32(0)
+ compute_2[cse_var_2 + 233] = T.float32(0)
+ compute_2[cse_var_2 + 234] = T.float32(0)
+ compute_2[cse_var_2 + 235] = T.float32(0)
+ compute_2[cse_var_2 + 236] = T.float32(0)
+ compute_2[cse_var_2 + 237] = T.float32(0)
+ compute_2[cse_var_2 + 238] = T.float32(0)
+ compute_2[cse_var_2 + 239] = T.float32(0)
+ for elem_idx in range(placeholder_5[cse_var_1 + 1] - placeholder_5[cse_var_1]):
+ placeholder_5 = T.buffer_decl((33,), "int32", data=placeholder_3.data)
+ cse_var_131: T.int32 = elem_idx * 16
+ cse_var_130: T.int32 = cse_var_2 + 99
+ cse_var_129: T.int32 = cse_var_2 + 98
+ cse_var_128: T.int32 = cse_var_2 + 97
+ cse_var_127: T.int32 = cse_var_2 + 96
+ cse_var_126: T.int32 = cse_var_2 + 9
+ cse_var_125: T.int32 = cse_var_2 + 8
+ cse_var_124: T.int32 = cse_var_2 + 79
+ cse_var_123: T.int32 = cse_var_2 + 78
+ cse_var_122: T.int32 = cse_var_2 + 77
+ cse_var_121: T.int32 = cse_var_2 + 76
+ cse_var_120: T.int32 = cse_var_2 + 75
+ cse_var_119: T.int32 = cse_var_2 + 74
+ cse_var_118: T.int32 = cse_var_2 + 73
+ cse_var_117: T.int32 = cse_var_2 + 72
+ cse_var_116: T.int32 = cse_var_2 + 71
+ cse_var_115: T.int32 = cse_var_2 + 70
+ cse_var_114: T.int32 = cse_var_2 + 7
+ cse_var_113: T.int32 = cse_var_2 + 69
+ cse_var_112: T.int32 = cse_var_2 + 68
+ cse_var_111: T.int32 = cse_var_2 + 67
+ cse_var_110: T.int32 = cse_var_2 + 66
+ cse_var_109: T.int32 = cse_var_2 + 65
+ cse_var_108: T.int32 = cse_var_2 + 64
+ cse_var_107: T.int32 = cse_var_2 + 6
+ cse_var_106: T.int32 = cse_var_2 + 5
+ cse_var_105: T.int32 = cse_var_2 + 47
+ cse_var_104: T.int32 = cse_var_2 + 46
+ cse_var_103: T.int32 = cse_var_2 + 45
+ cse_var_102: T.int32 = cse_var_2 + 44
+ cse_var_101: T.int32 = cse_var_2 + 43
+ cse_var_100: T.int32 = cse_var_2 + 42
+ cse_var_99: T.int32 = cse_var_2 + 41
+ cse_var_98: T.int32 = cse_var_2 + 40
+ cse_var_97: T.int32 = cse_var_2 + 4
+ cse_var_96: T.int32 = cse_var_2 + 39
+ cse_var_95: T.int32 = cse_var_2 + 38
+ cse_var_94: T.int32 = cse_var_2 + 37
+ cse_var_93: T.int32 = cse_var_2 + 36
+ cse_var_92: T.int32 = cse_var_2 + 35
+ cse_var_91: T.int32 = cse_var_2 + 34
+ cse_var_90: T.int32 = cse_var_2 + 33
+ cse_var_89: T.int32 = cse_var_2 + 32
+ cse_var_88: T.int32 = cse_var_2 + 3
+ cse_var_87: T.int32 = cse_var_2 + 239
+ cse_var_86: T.int32 = cse_var_2 + 238
+ cse_var_85: T.int32 = cse_var_2 + 237
+ cse_var_84: T.int32 = cse_var_2 + 236
+ cse_var_83: T.int32 = cse_var_2 + 235
+ cse_var_82: T.int32 = cse_var_2 + 234
+ cse_var_81: T.int32 = cse_var_2 + 233
+ cse_var_80: T.int32 = cse_var_2 + 232
+ cse_var_79: T.int32 = cse_var_2 + 231
+ cse_var_78: T.int32 = cse_var_2 + 230
+ cse_var_77: T.int32 = cse_var_2 + 229
+ cse_var_76: T.int32 = cse_var_2 + 228
+ cse_var_75: T.int32 = cse_var_2 + 227
+ cse_var_74: T.int32 = cse_var_2 + 226
+ cse_var_73: T.int32 = cse_var_2 + 225
+ cse_var_72: T.int32 = cse_var_2 + 224
+ cse_var_71: T.int32 = cse_var_2 + 207
+ cse_var_70: T.int32 = cse_var_2 + 206
+ cse_var_69: T.int32 = cse_var_2 + 205
+ cse_var_68: T.int32 = cse_var_2 + 204
+ cse_var_67: T.int32 = cse_var_2 + 203
+ cse_var_66: T.int32 = cse_var_2 + 202
+ cse_var_65: T.int32 = cse_var_2 + 201
+ cse_var_64: T.int32 = cse_var_2 + 200
+ cse_var_63: T.int32 = cse_var_2 + 2
+ cse_var_62: T.int32 = cse_var_2 + 199
+ cse_var_61: T.int32 = cse_var_2 + 198
+ cse_var_60: T.int32 = cse_var_2 + 197
+ cse_var_59: T.int32 = cse_var_2 + 196
+ cse_var_58: T.int32 = cse_var_2 + 195
+ cse_var_57: T.int32 = cse_var_2 + 194
+ cse_var_56: T.int32 = cse_var_2 + 193
+ cse_var_55: T.int32 = cse_var_2 + 192
+ cse_var_54: T.int32 = cse_var_2 + 175
+ cse_var_53: T.int32 = cse_var_2 + 174
+ cse_var_52: T.int32 = cse_var_2 + 173
+ cse_var_51: T.int32 = cse_var_2 + 172
+ cse_var_50: T.int32 = cse_var_2 + 171
+ cse_var_49: T.int32 = cse_var_2 + 170
+ cse_var_48: T.int32 = cse_var_2 + 169
+ cse_var_47: T.int32 = cse_var_2 + 168
+ cse_var_46: T.int32 = cse_var_2 + 167
+ cse_var_45: T.int32 = cse_var_2 + 166
+ cse_var_44: T.int32 = cse_var_2 + 165
+ cse_var_43: T.int32 = cse_var_2 + 164
+ cse_var_42: T.int32 = cse_var_2 + 163
+ cse_var_41: T.int32 = cse_var_2 + 162
+ cse_var_40: T.int32 = cse_var_2 + 161
+ cse_var_39: T.int32 = cse_var_2 + 160
+ cse_var_38: T.int32 = cse_var_2 + 15
+ cse_var_37: T.int32 = cse_var_2 + 143
+ cse_var_36: T.int32 = cse_var_2 + 142
+ cse_var_35: T.int32 = cse_var_2 + 141
+ cse_var_34: T.int32 = cse_var_2 + 140
+ cse_var_33: T.int32 = cse_var_2 + 14
+ cse_var_32: T.int32 = cse_var_2 + 139
+ cse_var_31: T.int32 = cse_var_2 + 138
+ cse_var_30: T.int32 = cse_var_2 + 137
+ cse_var_29: T.int32 = cse_var_2 + 136
+ cse_var_28: T.int32 = cse_var_2 + 135
+ cse_var_27: T.int32 = cse_var_2 + 134
+ cse_var_26: T.int32 = cse_var_2 + 133
+ cse_var_25: T.int32 = cse_var_2 + 132
+ cse_var_24: T.int32 = cse_var_2 + 131
+ cse_var_23: T.int32 = cse_var_2 + 130
+ cse_var_22: T.int32 = cse_var_2 + 13
+ cse_var_21: T.int32 = cse_var_2 + 129
+ cse_var_20: T.int32 = cse_var_2 + 128
+ cse_var_19: T.int32 = cse_var_2 + 12
+ cse_var_18: T.int32 = cse_var_2 + 111
+ cse_var_17: T.int32 = cse_var_2 + 110
+ cse_var_16: T.int32 = cse_var_2 + 11
+ cse_var_15: T.int32 = cse_var_2 + 109
+ cse_var_14: T.int32 = cse_var_2 + 108
+ cse_var_13: T.int32 = cse_var_2 + 107
+ cse_var_12: T.int32 = cse_var_2 + 106
+ cse_var_11: T.int32 = cse_var_2 + 105
+ cse_var_10: T.int32 = cse_var_2 + 104
+ cse_var_9: T.int32 = cse_var_2 + 103
+ cse_var_8: T.int32 = cse_var_2 + 102
+ cse_var_7: T.int32 = cse_var_2 + 101
+ cse_var_6: T.int32 = cse_var_2 + 100
+ cse_var_5: T.int32 = cse_var_2 + 10
+ cse_var_4: T.int32 = cse_var_2 + 1
+ cse_var_3: T.int32 = i0_outer_i1_outer_fused // 16 * 4096 + i_outer_inner * 2048
+ placeholder_6 = T.buffer_decl((78656,), data=placeholder_1.data)
+ placeholder_7 = T.buffer_decl((32768,), data=placeholder.data)
+ placeholder_8 = T.buffer_decl((4916,), "int32", data=placeholder_2.data)
+ compute_2[cse_var_2] = compute_2[cse_var_2] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_4] = compute_2[cse_var_4] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 1] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_63] = compute_2[cse_var_63] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 2] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_88] = compute_2[cse_var_88] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 3] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_97] = compute_2[cse_var_97] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 4] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_106] = compute_2[cse_var_106] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 5] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_107] = compute_2[cse_var_107] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 6] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_114] = compute_2[cse_var_114] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 7] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_125] = compute_2[cse_var_125] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 8] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_126] = compute_2[cse_var_126] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 9] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_5] = compute_2[cse_var_5] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 10] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_16] = compute_2[cse_var_16] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 11] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_19] = compute_2[cse_var_19] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 12] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_22] = compute_2[cse_var_22] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 13] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_33] = compute_2[cse_var_33] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 14] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_38] = compute_2[cse_var_38] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 15] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_89] = compute_2[cse_var_89] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_90] = compute_2[cse_var_90] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 1] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_91] = compute_2[cse_var_91] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 2] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_92] = compute_2[cse_var_92] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 3] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_93] = compute_2[cse_var_93] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 4] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_94] = compute_2[cse_var_94] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 5] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_95] = compute_2[cse_var_95] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 6] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_96] = compute_2[cse_var_96] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 7] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_98] = compute_2[cse_var_98] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 8] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_99] = compute_2[cse_var_99] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 9] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_100] = compute_2[cse_var_100] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 10] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_101] = compute_2[cse_var_101] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 11] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_102] = compute_2[cse_var_102] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 12] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_103] = compute_2[cse_var_103] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 13] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_104] = compute_2[cse_var_104] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 14] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_105] = compute_2[cse_var_105] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 15] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_108] = compute_2[cse_var_108] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_109] = compute_2[cse_var_109] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 1] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_110] = compute_2[cse_var_110] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 2] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_111] = compute_2[cse_var_111] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 3] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_112] = compute_2[cse_var_112] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 4] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_113] = compute_2[cse_var_113] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 5] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_115] = compute_2[cse_var_115] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 6] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_116] = compute_2[cse_var_116] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 7] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_117] = compute_2[cse_var_117] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 8] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_118] = compute_2[cse_var_118] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 9] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_119] = compute_2[cse_var_119] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 10] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_120] = compute_2[cse_var_120] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 11] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_121] = compute_2[cse_var_121] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 12] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_122] = compute_2[cse_var_122] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 13] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_123] = compute_2[cse_var_123] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 14] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_124] = compute_2[cse_var_124] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 15] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_127] = compute_2[cse_var_127] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_128] = compute_2[cse_var_128] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 1] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_129] = compute_2[cse_var_129] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 2] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_130] = compute_2[cse_var_130] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 3] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_6] = compute_2[cse_var_6] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 4] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_7] = compute_2[cse_var_7] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 5] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_8] = compute_2[cse_var_8] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 6] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_9] = compute_2[cse_var_9] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 7] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_10] = compute_2[cse_var_10] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 8] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_11] = compute_2[cse_var_11] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 9] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_12] = compute_2[cse_var_12] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 10] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_13] = compute_2[cse_var_13] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 11] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_14] = compute_2[cse_var_14] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 12] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_15] = compute_2[cse_var_15] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 13] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_17] = compute_2[cse_var_17] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 14] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_18] = compute_2[cse_var_18] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 15] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_20] = compute_2[cse_var_20] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_21] = compute_2[cse_var_21] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 1] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_23] = compute_2[cse_var_23] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 2] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_24] = compute_2[cse_var_24] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 3] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_25] = compute_2[cse_var_25] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 4] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_26] = compute_2[cse_var_26] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 5] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_27] = compute_2[cse_var_27] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 6] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_28] = compute_2[cse_var_28] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 7] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_29] = compute_2[cse_var_29] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 8] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_30] = compute_2[cse_var_30] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 9] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_31] = compute_2[cse_var_31] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 10] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_32] = compute_2[cse_var_32] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 11] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_34] = compute_2[cse_var_34] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 12] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_35] = compute_2[cse_var_35] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 13] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_36] = compute_2[cse_var_36] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 14] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_37] = compute_2[cse_var_37] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 15] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_39] = compute_2[cse_var_39] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_40] = compute_2[cse_var_40] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 1] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_41] = compute_2[cse_var_41] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 2] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_42] = compute_2[cse_var_42] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 3] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_43] = compute_2[cse_var_43] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 4] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_44] = compute_2[cse_var_44] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 5] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_45] = compute_2[cse_var_45] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 6] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_46] = compute_2[cse_var_46] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 7] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_47] = compute_2[cse_var_47] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 8] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_48] = compute_2[cse_var_48] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 9] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_49] = compute_2[cse_var_49] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 10] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_50] = compute_2[cse_var_50] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 11] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_51] = compute_2[cse_var_51] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 12] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_52] = compute_2[cse_var_52] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 13] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_53] = compute_2[cse_var_53] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 14] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_54] = compute_2[cse_var_54] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 15] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_55] = compute_2[cse_var_55] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_56] = compute_2[cse_var_56] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 1] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_57] = compute_2[cse_var_57] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 2] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_58] = compute_2[cse_var_58] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 3] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_59] = compute_2[cse_var_59] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 4] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_60] = compute_2[cse_var_60] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 5] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_61] = compute_2[cse_var_61] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 6] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_62] = compute_2[cse_var_62] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 7] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_64] = compute_2[cse_var_64] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 8] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_65] = compute_2[cse_var_65] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 9] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_66] = compute_2[cse_var_66] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 10] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_67] = compute_2[cse_var_67] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 11] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_68] = compute_2[cse_var_68] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 12] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_69] = compute_2[cse_var_69] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 13] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_70] = compute_2[cse_var_70] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 14] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_71] = compute_2[cse_var_71] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 15] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_72] = compute_2[cse_var_72] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_73] = compute_2[cse_var_73] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 1] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_74] = compute_2[cse_var_74] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 2] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_75] = compute_2[cse_var_75] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 3] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_76] = compute_2[cse_var_76] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 4] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_77] = compute_2[cse_var_77] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 5] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_78] = compute_2[cse_var_78] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 6] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_79] = compute_2[cse_var_79] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 7] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_80] = compute_2[cse_var_80] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 8] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_81] = compute_2[cse_var_81] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 9] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_82] = compute_2[cse_var_82] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 10] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_83] = compute_2[cse_var_83] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 11] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_84] = compute_2[cse_var_84] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 12] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_85] = compute_2[cse_var_85] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 13] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_86] = compute_2[cse_var_86] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 14] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_87] = compute_2[cse_var_87] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 15] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
for i0_inner in range(16):
- cse_var_22: T.int32 = i0_outer_i1_outer_fused // 16 * 8192 + i0_inner * 512 + i0_outer_i1_outer_fused % 16 * 32
- compute_4 = T.buffer_decl((65536,), data=compute_1.data)
- placeholder_10 = T.buffer_decl((65536,), data=placeholder_9.data)
- compute_4[cse_var_22:cse_var_22 + 32] = T.max(compute_3[i0_inner * 32:i0_inner * 32 + 32] + placeholder_10[cse_var_22:cse_var_22 + 32], T.Broadcast(T.float32(0), 32))
+ cse_var_132: T.int32 = i0_outer_i1_outer_fused // 16 * 8192 + i0_inner * 512 + i0_outer_i1_outer_fused % 16 * 32
+ compute_3 = T.buffer_decl((65536,), data=compute.data)
+ placeholder_5 = T.buffer_decl((65536,), data=placeholder_4.data)
+ compute_3[cse_var_132:cse_var_132 + 32] = T.max(compute_2[i0_inner * 32:i0_inner * 32 + 32] + placeholder_5[cse_var_132:cse_var_132 + 32], T.Broadcast(T.float32(0), 32))
@@ -509,7 +836,7 @@ We build the binary and check its correctness and performance.
.. code-block:: none
- Execution time of this operator: 1.894 ms
+ Execution time of this operator: 2.768 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 51a55380d3..9137835e06 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
@@ -5,12 +5,12 @@
Computation times
=================
-**00:24.192** total execution time for **how_to_tune_with_autotvm** files:
+**00:41.552** 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``) | 00:24.157 | 0.0 MB |
+| :ref:`sphx_glr_how_to_tune_with_autotvm_tune_conv2d_cuda.py` (``tune_conv2d_cuda.py``) | 00:41.519 | 0.0 MB |
+--------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_tune_with_autotvm_tune_relay_x86.py` (``tune_relay_x86.py``) | 00:00.021 | 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_cuda.py` (``tune_relay_cuda.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 60fd7976b5..194b082ce6 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
@@ -268,9 +268,8 @@ for this template
waiting for device...
device available
Get devices for measurement successfully!
- No: 1 GFLOPS: 50.48/50.48 result: MeasureResult(costs=(0.004586107448275862,), error_no=MeasureErrorNo.NO_ERROR, all_cost=3.907130241394043, timestamp=1674175038.7889755) [('tile_f', [-1, 2, 1, 8]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 8, 4]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9371037
- No: 2 GFLOPS: 93.65/93.65 result: MeasureResult(costs=(0.002471886048780488,), error_no=MeasureErrorNo.NO_ERROR, all_cost=2.077481985092163, timestamp=1674175039.5116773) [('tile_f', [-1, 1, 16, 2]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 64, 2]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 1)],None,5284705
- No: 3 GFLOPS: 0.00/93.65 result: Traceback (most recent call last):
+ No: 1 GFLOPS: 7.75/7.75 result: MeasureResult(costs=(0.0298710955,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.783884048461914, timestamp=1674185426.5654202) [('tile_f', [-1, 1, 2, 32]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 2, 2]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,41550
+ No: 2 GFLOPS: 0.00/7.75 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -392,9 +391,8 @@ 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 875, 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, 1]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 64, 8]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,1666969
- No: 4 GFLOPS: 213.99/213.99 result: MeasureResult(costs=(0.0010818329677419354,), error_no=MeasureErrorNo.NO_ERROR, all_cost=2.1306028366088867, timestamp=1674175042.3393323) [('tile_f', [-1, 4, 8, 1]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 2, 4]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2782589
- No: 5 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 1, 128, 2]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 128, 2]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,1224397
+ No: 3 GFLOPS: 0.00/7.75 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -516,8 +514,8 @@ 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 875, 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, 4]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 8, 64]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,7526534
- No: 6 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 1, 8, 4]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 16, 8]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,10372241
+ No: 4 GFLOPS: 0.00/7.75 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -639,8 +637,8 @@ 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 875, 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, 2, 32]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 16, 8]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9404753
- No: 7 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 2, 16, 4]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 16, 4]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,10342327
+ No: 5 GFLOPS: 0.00/7.75 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -762,8 +760,9 @@ 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 875, 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, 64]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 4, 2]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,1011105
- No: 8 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 2, 128, 2]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 8, 16]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,4392618
+ No: 6 GFLOPS: 127.76/127.76 result: MeasureResult(costs=(0.0018119425555555557,), error_no=MeasureErrorNo.NO_ERROR, all_cost=4.839144229888916, timestamp=1674185434.4324708) [('tile_f', [-1, 1, 2, 16]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 16, 1]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9502410
+ No: 7 GFLOPS: 0.00/127.76 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -885,8 +884,8 @@ 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 875, 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, 1, 2]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 256, 2]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 1)],None,5484881
- No: 9 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 4, 8, 4]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 32, 1]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 1)],None,6022183
+ No: 8 GFLOPS: 0.00/127.76 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -1008,8 +1007,8 @@ 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 875, 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, 1, 1, 7]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 16, 2]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,8375713
- No: 10 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 32, 2, 2]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 1, 64]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2289829
+ No: 9 GFLOPS: 0.00/127.76 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -1131,8 +1130,26 @@ 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 875, 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, 16, 2]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 64, 4]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9187286
- No: 11 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 2, 4, 4]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 1, 512]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,7162656
+ No: 10 GFLOPS: 0.00/127.76 result: Traceback (most recent call last):
+ File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 142, in build
+ res = future.result()
+ File "/usr/lib/python3.7/concurrent/futures/_base.py", line 435, in result
+ return self.__get_result()
+ File "/usr/lib/python3.7/concurrent/futures/_base.py", line 384, in __get_result
+ raise self._exception
+ File "/usr/lib/python3.7/concurrent/futures/thread.py", line 57, in run
+ result = self.fn(*self.args, **self.kwargs)
+ File "/workspace/python/tvm/contrib/popen_pool.py", line 432, in <lambda>
+ worker = lambda *args: self._worker_run(*args)
+ File "/workspace/python/tvm/contrib/popen_pool.py", line 401, in _worker_run
+ return proc.recv()
+ File "/workspace/python/tvm/contrib/popen_pool.py", line 309, in recv
+ raise TimeoutError()
+ TimeoutError
+
+ [('tile_f', [-1, 4, 1, 16]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 1, 8]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9195726
+ No: 11 GFLOPS: 0.00/127.76 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -1254,8 +1271,8 @@ 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 875, 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, 1]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 8, 16]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2262743
- No: 12 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 64, 2, 4]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 128, 1]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,5058574
+ No: 12 GFLOPS: 0.00/127.76 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -1377,8 +1394,10 @@ 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 875, 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, 16, 2]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 32, 16]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,1107565
- No: 13 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 1, 4, 16]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 128, 2]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2190495
+ No: 13 GFLOPS: 2.71/127.76 result: MeasureResult(costs=(0.08529100775000001,), error_no=MeasureErrorNo.NO_ERROR, all_cost=5.805093050003052, timestamp=1674185450.5576441) [('tile_f', [-1, 1, 2, 8]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 2, 64]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,7325263
+ No: 14 GFLOPS: 299.03/299.03 result: MeasureResult(costs=(0.000774184811594203,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.5905015468597412, timestamp=1674185451.5622723) [('tile_f', [-1, 2, 32, 1]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 8, 2]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,436081
+ No: 15 GFLOPS: 0.00/299.03 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -1500,8 +1519,8 @@ 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 875, 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, 8, 32]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 512, 1]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,420837
- No: 14 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 128, 2, 2]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 8, 32]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2864031
+ No: 16 GFLOPS: 0.00/299.03 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -1623,8 +1642,8 @@ 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 875, 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, 32, 8, 1]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 8, 32]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,5185872
- No: 15 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 8, 4, 16]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 32, 1]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 1)],None,5632838
+ No: 17 GFLOPS: 0.00/299.03 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -1746,8 +1765,8 @@ 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 875, 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, 4, 1]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 16, 16]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,3232479
- No: 16 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 2, 8, 16]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 1, 256]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2313260
+ No: 18 GFLOPS: 0.00/299.03 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -1869,8 +1888,8 @@ 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 875, 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, 32]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 64, 4]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2606310
- No: 17 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 256, 1, 1]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 16, 32]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9836868
+ No: 19 GFLOPS: 0.00/299.03 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -1992,8 +2011,8 @@ 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 875, 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, 16, 16]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 1, 8]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 0), ('unroll_explicit', 1)],None,6485563
- No: 18 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 16, 1, 4]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 512, 1]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,7970924
+ No: 20 GFLOPS: 0.00/299.03 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -2115,253 +2134,7 @@ 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 875, 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, 1, 7, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 4, 32]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 1)],None,5571827
- No: 19 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
- func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, 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:1730
- 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:1670
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1630
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1645
- 13: operator()
- at ../src/driver/driver_api.cc:395
- 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:381
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:276
- 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:451
- 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:1749
- 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:1693
- 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:1617
- 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 875, 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:1730
- 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:1670
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1630
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1645
- 13: operator()
- at ../src/driver/driver_api.cc:395
- 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:381
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:276
- 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:451
- 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:1749
- 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:1693
- 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:1617
- 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 875, 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, 7, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 4, 32]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9441890
- No: 20 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
- func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, 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:1730
- 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:1670
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1630
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1645
- 13: operator()
- at ../src/driver/driver_api.cc:395
- 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:381
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:276
- 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:451
- 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:1749
- 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:1693
- 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:1617
- 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 875, 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:1730
- 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:1670
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1630
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1645
- 13: operator()
- at ../src/driver/driver_api.cc:395
- 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:381
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:276
- 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:451
- 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:1749
- 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:1693
- 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:1617
- 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 875, 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, 1, 4]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 16, 4]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,4341582
+ tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 2, 1, 32]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 256, 2]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,3935986
@@ -2416,9 +2189,9 @@ and measure running time.
Finish loading 20 records
Best config:
- [('tile_f', [-1, 4, 8, 1]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 2, 4]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2782589
+ [('tile_f', [-1, 2, 32, 1]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 8, 2]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,436081
Finish loading 20 records
- Time cost of this operator: 0.001132
+ Time cost of this operator: 0.001133
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 bff02587b5..b6057f0766 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
@@ -363,10 +363,10 @@ 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 310.8 98.686 (1, 2, 10, 10, 3) 2 1 [310.8]
- tvmgen_default_fused_layout_transform_1 tvmgen_default_fused_layout_transform_1 3.186 1.012 (1, 6, 10, 10) 1 1 [3.186]
- tvmgen_default_fused_layout_transform tvmgen_default_fused_layout_transform 0.953 0.303 (1, 1, 10, 10, 3) 1 1 [0.953]
- Total_time - 314.939 - - - - -
+ tvmgen_default_fused_nn_contrib_conv2d_NCHWc tvmgen_default_fused_nn_contrib_conv2d_NCHWc 307.8 98.719 (1, 2, 10, 10, 3) 2 1 [307.8]
+ tvmgen_default_fused_layout_transform_1 tvmgen_default_fused_layout_transform_1 3.021 0.969 (1, 6, 10, 10) 1 1 [3.021]
+ tvmgen_default_fused_layout_transform tvmgen_default_fused_layout_transform 0.972 0.312 (1, 1, 10, 10, 3) 1 1 [0.972]
+ Total_time - 311.793 - - - - -
@@ -431,10 +431,10 @@ 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 102.8 97.511 (1, 6, 10, 10, 1) 2 1 [102.8]
- tvmgen_default_fused_layout_transform_1 tvmgen_default_fused_layout_transform_1 1.773 1.682 (1, 6, 10, 10) 1 1 [1.773]
- tvmgen_default_fused_layout_transform tvmgen_default_fused_layout_transform 0.851 0.808 (1, 3, 10, 10, 1) 1 1 [0.851]
- Total_time - 105.424 - - - - -
+ tvmgen_default_fused_nn_contrib_conv2d_NCHWc tvmgen_default_fused_nn_contrib_conv2d_NCHWc 104.8 97.555 (1, 6, 10, 10, 1) 2 1 [104.8]
+ tvmgen_default_fused_layout_transform_1 tvmgen_default_fused_layout_transform_1 1.776 1.653 (1, 6, 10, 10) 1 1 [1.776]
+ tvmgen_default_fused_layout_transform tvmgen_default_fused_layout_transform 0.851 0.792 (1, 3, 10, 10, 1) 1 1 [0.851]
+ Total_time - 107.426 - - - - -
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 10d497e3b1..2cd0784826 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
@@ -117,7 +117,7 @@ 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]
61%|###### | 2.09M/3.42M [00:00<00:00, 19.9MB/s]
100%|##########| 3.42M/3.42M [00:00<00:00, 30.4MB/s]
+
0%| | 0.00/3.42M [00:00<?, ?B/s]
100%|##########| 3.42M/3.42M [00:00<00:00, 137MB/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.
@@ -322,7 +322,7 @@ Look up prediction top 1 index in 1000 class synset.
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 1 minutes 9.130 seconds)
+ **Total running time of the script:** ( 1 minutes 7.086 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 53b217a1f2..12be13df9d 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
@@ -218,7 +218,7 @@ take about **2 minutes** to download the Stanford Cars, while COCO 2017 validati
.. code-block:: none
- '/tmp/tmp9frj0wj1/images/random'
+ '/tmp/tmpiql0qhf6/images/random'
@@ -309,7 +309,7 @@ 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: [0.0, 1.0], [1.0, 0.0], [1.0, 0.0], [0.0, 1.0], [1.0, 0.0], [0.0, 1.0], [1.0, 0.0], [1.0, 0.0], [0.0, 1.0], [1.0, 0.0]
+ :alt: [0.0, 1.0], [0.0, 1.0], [0.0, 1.0], [0.0, 1.0], [0.0, 1.0], [0.0, 1.0], [0.0, 1.0], [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
@@ -318,8 +318,8 @@ objects to other stuff? We can display some examples from our datasets using ``m
.. code-block:: none
- /tmp/tmp9frj0wj1/images/target contains 8144 images
- /tmp/tmp9frj0wj1/images/random contains 5000 images
+ /tmp/tmpiql0qhf6/images/target contains 8144 images
+ /tmp/tmpiql0qhf6/images/random contains 5000 images
@@ -494,13 +494,13 @@ the time on our validation set).
.. code-block:: none
Epoch 1/3
- 328/328 - 47s - loss: 0.2199 - accuracy: 0.9239 - val_loss: 0.1185 - val_accuracy: 0.9581 - 47s/epoch - 142ms/step
+ 328/328 - 47s - loss: 0.2138 - accuracy: 0.9251 - val_loss: 0.1147 - val_accuracy: 0.9562 - 47s/epoch - 143ms/step
Epoch 2/3
- 328/328 - 43s - loss: 0.1003 - accuracy: 0.9617 - val_loss: 0.1516 - val_accuracy: 0.9486 - 43s/epoch - 132ms/step
+ 328/328 - 43s - loss: 0.0960 - accuracy: 0.9631 - val_loss: 0.0898 - val_accuracy: 0.9671 - 43s/epoch - 131ms/step
Epoch 3/3
- 328/328 - 43s - loss: 0.0737 - accuracy: 0.9728 - val_loss: 0.1310 - val_accuracy: 0.9615 - 43s/epoch - 132ms/step
+ 328/328 - 43s - loss: 0.0588 - accuracy: 0.9798 - val_loss: 0.1360 - val_accuracy: 0.9528 - 43s/epoch - 131ms/step
- <keras.callbacks.History object at 0x7fe689e4a810>
+ <keras.callbacks.History object at 0x7f6b69639fd0>
@@ -857,7 +857,7 @@ 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 29.197 seconds)
+ **Total running time of the script:** ( 4 minutes 19.926 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 5bc9adbb3c..be1453c589 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
@@ -5,23 +5,23 @@
Computation times
=================
-**06:42.755** total execution time for **how_to_work_with_microtvm** files:
+**06:30.155** total execution time for **how_to_work_with_microtvm** files:
+---------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_microtvm_micro_train.py` (``micro_train.py``) | 04:29.197 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_microtvm_micro_train.py` (``micro_train.py``) | 04:19.926 | 0.0 MB |
+---------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_microtvm_micro_pytorch.py` (``micro_pytorch.py``) | 01:09.130 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_microtvm_micro_pytorch.py` (``micro_pytorch.py``) | 01:07.086 | 0.0 MB |
+---------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_microtvm_micro_autotune.py` (``micro_autotune.py``) | 00:51.744 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_microtvm_micro_autotune.py` (``micro_autotune.py``) | 00:50.864 | 0.0 MB |
+---------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_microtvm_micro_aot.py` (``micro_aot.py``) | 00:08.855 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_microtvm_micro_aot.py` (``micro_aot.py``) | 00:08.512 | 0.0 MB |
+---------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_microtvm_micro_tflite.py` (``micro_tflite.py``) | 00:03.827 | 0.0 MB |
-+---------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_microtvm_micro_tvmc.py` (``micro_tvmc.py``) | 00:00.000 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_microtvm_micro_tflite.py` (``micro_tflite.py``) | 00:03.768 | 0.0 MB |
+---------------------------------------------------------------------------------------------+-----------+--------+
| :ref:`sphx_glr_how_to_work_with_microtvm_micro_ethosu.py` (``micro_ethosu.py``) | 00:00.000 | 0.0 MB |
+---------------------------------------------------------------------------------------------+-----------+--------+
+| :ref:`sphx_glr_how_to_work_with_microtvm_micro_tvmc.py` (``micro_tvmc.py``) | 00:00.000 | 0.0 MB |
++---------------------------------------------------------------------------------------------+-----------+--------+
| :ref:`sphx_glr_how_to_work_with_microtvm_micro_reference_vm.py` (``micro_reference_vm.py``) | 00:00.000 | 0.0 MB |
+---------------------------------------------------------------------------------------------+-----------+--------+
| :ref:`sphx_glr_how_to_work_with_microtvm_micro_mlperftiny.py` (``micro_mlperftiny.py``) | 00:00.000 | 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 0a737689d9..1660fe8fc2 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
=================
-**00:45.175** total execution time for **how_to_work_with_relay** files:
+**00:43.680** 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``) | 00:32.926 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_relay_using_pipeline_executor.py` (``using_pipeline_executor.py``) | 00:31.973 | 0.0 MB |
+----------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_relay_using_external_lib.py` (``using_external_lib.py``) | 00:10.433 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_relay_using_external_lib.py` (``using_external_lib.py``) | 00:10.161 | 0.0 MB |
+----------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_relay_build_gcn.py` (``build_gcn.py``) | 00:01.810 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_relay_build_gcn.py` (``build_gcn.py``) | 00:01.541 | 0.0 MB |
+----------------------------------------------------------------------------------------------------+-----------+--------+
| :ref:`sphx_glr_how_to_work_with_relay_using_relay_viz.py` (``using_relay_viz.py``) | 00:00.006 | 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 71015ef04f..c192de79a8 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
@@ -264,7 +264,7 @@ The following example customizes CUDA lowering rule for :code:`exp`.
.. code-block:: none
- <function my_cuda_math_rule at 0x7fe68a33d290>
+ <function my_cuda_math_rule at 0x7f6bc0351440>
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 1f162cc8d6..42e8be7aaf 100644
--- a/docs/_sources/how_to/work_with_schedules/reduction.rst.txt
+++ b/docs/_sources/how_to/work_with_schedules/reduction.rst.txt
@@ -483,21 +483,20 @@ Here is an example for 2D convolution with filter size = [3, 3] and strides = [1
@I.ir_module
class Module:
@T.prim_func
- def main(Input: T.handle, Filter: T.handle, Output: T.handle):
+ def main(Input: T.handle, Filter: T.Buffer((3, 3), "float32"), Output: T.handle):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
n = T.var("int32")
stride = T.var("int32")
stride_1 = T.var("int32")
Input_1 = T.match_buffer(Input, (n, n), strides=(stride, stride_1), type="auto")
- Filter_1 = T.match_buffer(Filter, (3, 3))
Output_1 = T.match_buffer(Output, (n - 2, n - 2))
for i, j in T.grid(n - 2, n - 2):
Output_2 = T.buffer_decl(((n - 2) * (n - 2),), data=Output_1.data)
Output_2[i * (n - 2) + j] = T.float32(0)
for di, dj in T.grid(3, 3):
Input_2 = T.buffer_decl((stride * n,), data=Input_1.data, type="auto")
- Filter_2 = T.buffer_decl((9,), data=Filter_1.data)
- Output_2[i * (n - 2) + j] = Output_2[i * (n - 2) + j] + Input_2[(i + di) * stride + (j + dj) * stride_1] * Filter_2[di * 3 + dj]
+ Filter_1 = T.buffer_decl((9,), data=Filter.data)
+ Output_2[i * (n - 2) + j] = Output_2[i * (n - 2) + j] + Input_2[(i + di) * stride + (j + dj) * stride_1] * Filter_1[di * 3 + dj]
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 50bd7c02e8..44e3b6bd55 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
=================
-**00:07.592** total execution time for **how_to_work_with_schedules** files:
+**00:07.523** total execution time for **how_to_work_with_schedules** files:
+------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_schedules_intrin_math.py` (``intrin_math.py``) | 00:05.077 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_schedules_intrin_math.py` (``intrin_math.py``) | 00:05.064 | 0.0 MB |
+------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_schedules_tensorize.py` (``tensorize.py``) | 00:01.155 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_schedules_tensorize.py` (``tensorize.py``) | 00:01.119 | 0.0 MB |
+------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_schedules_reduction.py` (``reduction.py``) | 00:00.577 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_schedules_reduction.py` (``reduction.py``) | 00:00.570 | 0.0 MB |
+------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_schedules_scan.py` (``scan.py``) | 00:00.562 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_schedules_scan.py` (``scan.py``) | 00:00.549 | 0.0 MB |
+------------------------------------------------------------------------------------------------+-----------+--------+
| :ref:`sphx_glr_how_to_work_with_schedules_extern_op.py` (``extern_op.py``) | 00:00.116 | 0.0 MB |
+------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_how_to_work_with_schedules_schedule_primitives.py` (``schedule_primitives.py``) | 00:00.050 | 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_tedd.py` (``tedd.py``) | 00:00.032 | 0.0 MB |
+| :ref:`sphx_glr_how_to_work_with_schedules_tedd.py` (``tedd.py``) | 00:00.031 | 0.0 MB |
+------------------------------------------------------------------------------------------------+-----------+--------+
| :ref:`sphx_glr_how_to_work_with_schedules_tuple_inputs.py` (``tuple_inputs.py``) | 00:00.024 | 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 43a87f122c..3b448c4bb0 100644
--- a/docs/_sources/how_to/work_with_schedules/tensorize.rst.txt
+++ b/docs/_sources/how_to/work_with_schedules/tensorize.rst.txt
@@ -90,19 +90,16 @@ The following lines describe the computation :code:`A * B^T` in TVM.
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 64), "float32"), B: T.Buffer((512, 64), "float32"), C: T.Buffer((1024, 512), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 64))
- B_1 = T.match_buffer(B, (512, 64))
- C_1 = T.match_buffer(C, (1024, 512))
for i, j in T.grid(1024, 512):
- C_2 = T.buffer_decl((524288,), data=C_1.data)
- C_2[i * 512 + j] = T.float32(0)
+ C_1 = T.buffer_decl((524288,), data=C.data)
+ C_1[i * 512 + j] = T.float32(0)
for k in range(64):
cse_var_1: T.int32 = i * 512 + j
- A_2 = T.buffer_decl((65536,), data=A_1.data)
- B_2 = T.buffer_decl((32768,), data=B_1.data)
- C_2[cse_var_1] = C_2[cse_var_1] + A_2[i * 64 + k] * B_2[j * 64 + k]
+ A_1 = T.buffer_decl((65536,), data=A.data)
+ B_1 = T.buffer_decl((32768,), data=B.data)
+ C_1[cse_var_1] = C_1[cse_var_1] + A_1[i * 64 + k] * B_1[j * 64 + k]
@@ -140,19 +137,16 @@ Thus we break down the matmul loops to make the innermost loops a (16x64) GEMV.
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 64), "float32"), B: T.Buffer((512, 64), "float32"), C: T.Buffer((1024, 512), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 64))
- B_1 = T.match_buffer(B, (512, 64))
- C_1 = T.match_buffer(C, (1024, 512))
for i, j_outer, j_inner in T.grid(1024, 32, 16):
- C_2 = T.buffer_decl((524288,), data=C_1.data)
- C_2[i * 512 + j_outer * 16 + j_inner] = T.float32(0)
+ C_1 = T.buffer_decl((524288,), data=C.data)
+ C_1[i * 512 + j_outer * 16 + j_inner] = T.float32(0)
for k in range(64):
cse_var_1: T.int32 = i * 512 + j_outer * 16 + j_inner
- A_2 = T.buffer_decl((65536,), data=A_1.data)
- B_2 = T.buffer_decl((32768,), data=B_1.data)
- C_2[cse_var_1] = C_2[cse_var_1] + A_2[i * 64 + k] * B_2[j_outer * 1024 + j_inner * 64 + k]
+ A_1 = T.buffer_decl((65536,), data=A.data)
+ B_1 = T.buffer_decl((32768,), data=B.data)
+ C_1[cse_var_1] = C_1[cse_var_1] + A_1[i * 64 + k] * B_1[j_outer * 1024 + j_inner * 64 + k]
@@ -260,13 +254,10 @@ such placeholder can be put to let TVM automatically bind the inferred value for
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 64), "float32"), B: T.Buffer((512, 64), "float32"), C: T.Buffer((1024, 512), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 64))
- B_1 = T.match_buffer(B, (512, 64))
- C_1 = T.match_buffer(C, (1024, 512))
for i, j_outer in T.grid(1024, 32):
- T.call_extern("int32", "gemv_update", T.tvm_access_ptr(T.type_annotation("float32"), C_1.data, i * 512 + j_outer * 16, 16, 2), T.tvm_access_ptr(T.type_annotation("float32"), A_1.data, i * 64, 64, 1), T.tvm_access_ptr(T.type_annotation("float32"), B_1.data, j_outer * 1024, 1024, 1), 16, 64, 64)
+ T.call_extern("int32", "gemv_update", T.tvm_access_ptr(T.type_annotation("float32"), C.data, i * 512 + j_outer * 16, 16, 2), T.tvm_access_ptr(T.type_annotation("float32"), A.data, i * 64, 64, 1), T.tvm_access_ptr(T.type_annotation("float32"), B.data, j_outer * 1024, 1024, 1), 16, 64, 64)
@@ -334,15 +325,12 @@ The importing needs to happen before the tensorized GEMV being executed.
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 64), "float32"), B: T.Buffer((512, 64), "float32"), C: T.Buffer((1024, 512), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 64))
- B_1 = T.match_buffer(B, (512, 64))
- C_1 = T.match_buffer(C, (1024, 512))
i = T.var("int32")
- T.attr(T.iter_var(i, None, "DataPar", ""), "pragma_import_llvm", "; ModuleID = '/tmp/tmpgb2rg8u1/input0.cc'\nsource_filename = \"/tmp/tmpgb2rg8u1/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 float*, [...]
+ T.attr(T.iter_var(i, None, "DataPar", ""), "pragma_import_llvm", "; ModuleID = '/tmp/tmpqmz888xx/input0.cc'\nsource_filename = \"/tmp/tmpqmz888xx/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 float*, [...]
for i, j_outer in T.grid(1024, 32):
- T.call_extern("int32", "gemv_update", T.tvm_access_ptr(T.type_annotation("float32"), C_1.data, i * 512 + j_outer * 16, 16, 2), T.tvm_access_ptr(T.type_annotation("float32"), A_1.data, i * 64, 64, 1), T.tvm_access_ptr(T.type_annotation("float32"), B_1.data, j_outer * 1024, 1024, 1), 16, 64, 64)
+ T.call_extern("int32", "gemv_update", T.tvm_access_ptr(T.type_annotation("float32"), C.data, i * 512 + j_outer * 16, 16, 2), T.tvm_access_ptr(T.type_annotation("float32"), A.data, i * 64, 64, 1), T.tvm_access_ptr(T.type_annotation("float32"), B.data, j_outer * 1024, 1024, 1), 16, 64, 64)
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 1641e21bee..25d16a489b 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
=================
-**00:29.785** total execution time for **topic_vta_tutorials_autotvm** files:
+**00:29.175** total execution time for **topic_vta_tutorials_autotvm** files:
+---------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_topic_vta_tutorials_autotvm_tune_relay_vta.py` (``tune_relay_vta.py``) | 00:29.778 | 0.0 MB |
+| :ref:`sphx_glr_topic_vta_tutorials_autotvm_tune_relay_vta.py` (``tune_relay_vta.py``) | 00:29.169 | 0.0 MB |
+---------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_topic_vta_tutorials_autotvm_tune_alu_vta.py` (``tune_alu_vta.py``) | 00:00.007 | 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 d7d798679d..2756a84548 100644
--- a/docs/_sources/topic/vta/tutorials/frontend/deploy_classification.rst.txt
+++ b/docs/_sources/topic/vta/tutorials/frontend/deploy_classification.rst.txt
@@ -293,7 +293,7 @@ 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 31.53s!
+ resnet18_v1 inference graph built in 30.69s!
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 4a3da4c7cc..c21f909da6 100644
--- a/docs/_sources/topic/vta/tutorials/frontend/deploy_detection.rst.txt
+++ b/docs/_sources/topic/vta/tutorials/frontend/deploy_detection.rst.txt
@@ -337,7 +337,7 @@ 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 21.77s!
+ yolov3-tiny inference graph built in 21.09s!
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 263c94c995..d9669d45eb 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
=================
-**01:37.002** total execution time for **topic_vta_tutorials_frontend** files:
+**01:35.112** total execution time for **topic_vta_tutorials_frontend** files:
+------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_topic_vta_tutorials_frontend_deploy_detection.py` (``deploy_detection.py``) | 00:48.682 | 0.0 MB |
+| :ref:`sphx_glr_topic_vta_tutorials_frontend_deploy_detection.py` (``deploy_detection.py``) | 00:47.673 | 0.0 MB |
+------------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_topic_vta_tutorials_frontend_deploy_classification.py` (``deploy_classification.py``) | 00:48.320 | 0.0 MB |
+| :ref:`sphx_glr_topic_vta_tutorials_frontend_deploy_classification.py` (``deploy_classification.py``) | 00:47.439 | 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 a21a495d46..5b33acf6a8 100644
--- a/docs/_sources/topic/vta/tutorials/matrix_multiply.rst.txt
+++ b/docs/_sources/topic/vta/tutorials/matrix_multiply.rst.txt
@@ -349,24 +349,21 @@ After we construct the schedule, by default the schedule computes
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1, 16, 1, 16), "int8"), B: T.Buffer((16, 16, 16, 16), "int8"), C: T.Buffer((1, 16, 1, 16), "int8")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1, 16, 1, 16), "int8")
- B_1 = T.match_buffer(B, (16, 16, 16, 16), "int8")
- C_1 = T.match_buffer(C, (1, 16, 1, 16), "int8")
A_buf = T.allocate([256], "int8", "global")
B_buf = T.allocate([65536], "int8", "global")
C_buf = T.allocate([256], "int32", "global")
A_buf_1 = T.buffer_decl((256,), "int8", data=A_buf)
for i1, i3 in T.grid(16, 16):
cse_var_1: T.int32 = i1 * 16 + i3
- A_2 = T.buffer_decl((256,), "int8", data=A_1.data)
- A_buf_1[cse_var_1] = A_2[cse_var_1]
+ A_1 = T.buffer_decl((256,), "int8", data=A.data)
+ A_buf_1[cse_var_1] = A_1[cse_var_1]
B_buf_1 = T.buffer_decl((65536,), "int8", data=B_buf)
for i0, i1, i2, i3 in T.grid(16, 16, 16, 16):
cse_var_2: T.int32 = i0 * 4096 + i1 * 256 + i2 * 16 + i3
- B_2 = T.buffer_decl((65536,), "int8", data=B_1.data)
- B_buf_1[cse_var_2] = B_2[cse_var_2]
+ B_1 = T.buffer_decl((65536,), "int8", data=B.data)
+ B_buf_1[cse_var_2] = B_1[cse_var_2]
C_buf_1 = T.buffer_decl((256,), "int32", data=C_buf)
for co, ci in T.grid(16, 16):
C_buf_1[co * 16 + ci] = 0
@@ -375,8 +372,8 @@ After we construct the schedule, by default the schedule computes
C_buf_1[cse_var_3] = C_buf_1[cse_var_3] + T.Cast("int32", A_buf_1[ko * 16 + ki]) * T.Cast("int32", B_buf_1[co * 4096 + ko * 256 + ci * 16 + ki])
for i1, i3 in T.grid(16, 16):
cse_var_4: T.int32 = i1 * 16 + i3
- C_2 = T.buffer_decl((256,), "int8", data=C_1.data)
- C_2[cse_var_4] = T.Cast("int8", C_buf_1[cse_var_4])
+ C_1 = T.buffer_decl((256,), "int8", data=C.data)
+ C_1[cse_var_4] = T.Cast("int8", C_buf_1[cse_var_4])
@@ -497,11 +494,8 @@ moving the copy operations into the matrix multiplication loop.
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1, 16, 1, 16), "int8"), B: T.Buffer((16, 16, 16, 16), "int8"), C: T.Buffer((1, 16, 1, 16), "int8")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1, 16, 1, 16), "int8")
- B_1 = T.match_buffer(B, (16, 16, 16, 16), "int8")
- C_1 = T.match_buffer(C, (1, 16, 1, 16), "int8")
C_buf = T.allocate([256], "int32", "local.acc_buffer")
A_buf = T.allocate([16], "int8", "local.inp_buffer")
B_buf = T.allocate([16], "int8", "local.wgt_buffer")
@@ -513,14 +507,14 @@ moving the copy operations into the matrix multiplication loop.
A_buf_1 = T.buffer_decl((16,), "int8", data=A_buf, scope="local.inp_buffer", align=16)
with T.attr(T.iter_var(i0, None, "DataPar", ""), "pragma_dma_copy", 1):
for i3 in range(16):
- A_2 = T.buffer_decl((256,), "int8", data=A_1.data)
- A_buf_1[i3] = A_2[ko * 16 + i3]
+ A_1 = T.buffer_decl((256,), "int8", data=A.data)
+ A_buf_1[i3] = A_1[ko * 16 + i3]
i0_1 = T.var("int32")
B_buf_1 = T.buffer_decl((16,), "int8", data=B_buf, scope="local.wgt_buffer", align=256)
with T.attr(T.iter_var(i0_1, None, "DataPar", ""), "pragma_dma_copy", 1):
for i3 in range(16):
- B_2 = T.buffer_decl((65536,), "int8", data=B_1.data)
- B_buf_1[i3] = B_2[co * 4096 + ko * 256 + ci * 16 + i3]
+ B_1 = T.buffer_decl((65536,), "int8", data=B.data)
+ B_buf_1[i3] = B_1[co * 4096 + ko * 256 + ci * 16 + i3]
for ki in range(16):
cse_var_1: T.int32 = co * 16 + ci
C_buf_1[cse_var_1] = C_buf_1[cse_var_1] + T.Cast("int32", A_buf_1[ki]) * T.Cast("int32", B_buf_1[ki])
@@ -528,8 +522,8 @@ moving the copy operations into the matrix multiplication loop.
T.attr(T.iter_var(i0, None, "DataPar", ""), "pragma_dma_copy", 1)
for i1, i3 in T.grid(16, 16):
cse_var_2: T.int32 = i1 * 16 + i3
- C_2 = T.buffer_decl((256,), "int8", data=C_1.data)
- C_2[cse_var_2] = T.Cast("int8", C_buf_1[cse_var_2])
+ C_1 = T.buffer_decl((256,), "int8", data=C.data)
+ C_1[cse_var_2] = T.Cast("int8", C_buf_1[cse_var_2])
@@ -580,11 +574,8 @@ by the VTA runtime JIT compiler.
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1, 16, 1, 16), "int8"), B: T.Buffer((16, 16, 16, 16), "int8"), C: T.Buffer((1, 16, 1, 16), "int8")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1, 16, 1, 16), "int8")
- B_1 = T.match_buffer(B, (16, 16, 16, 16), "int8")
- C_1 = T.match_buffer(C, (1, 16, 1, 16), "int8")
vta = T.var("int32")
with T.attr(T.iter_var(vta, None, "ThreadIndex", "vta"), "coproc_scope", 2):
with T.attr(T.iter_var(vta, None, "ThreadIndex", "vta"), "coproc_uop_scope", "VTAPushGEMMOp"):
@@ -595,8 +586,8 @@ by the VTA runtime JIT compiler.
for ko in range(16):
with T.attr(T.iter_var(vta, None, "ThreadIndex", "vta"), "coproc_scope", 1):
T.tir.vta.coproc_dep_pop(2, 1)
- T.call_extern("int32", "VTALoadBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), A_1.data, ko, 1, 1, 1, 0, 0, 0, 0, 0, 2)
- T.call_extern("int32", "VTALoadBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), B_1.data, ko, 1, 16, 16, 0, 0, 0, 0, 0, 1)
+ T.call_extern("int32", "VTALoadBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), A.data, ko, 1, 1, 1, 0, 0, 0, 0, 0, 2)
+ T.call_extern("int32", "VTALoadBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), B.data, ko, 1, 16, 16, 0, 0, 0, 0, 0, 1)
T.tir.vta.coproc_dep_push(1, 2)
T.attr(T.iter_var(vta, None, "ThreadIndex", "vta"), "coproc_scope", 2)
T.tir.vta.coproc_dep_pop(1, 2)
@@ -609,7 +600,7 @@ by the VTA runtime JIT compiler.
T.tir.vta.coproc_dep_pop(2, 1)
with T.attr(T.iter_var(vta, None, "ThreadIndex", "vta"), "coproc_scope", 3):
T.tir.vta.coproc_dep_pop(2, 3)
- T.call_extern("int32", "VTAStoreBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), 0, 4, C_1.data, 0, 16, 1, 16)
+ T.call_extern("int32", "VTAStoreBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), 0, 4, C.data, 0, 16, 1, 16)
T.tir.vta.coproc_sync()
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 f8ab9652a4..826fd8005d 100644
--- a/docs/_sources/topic/vta/tutorials/optimize/convolution_opt.rst.txt
+++ b/docs/_sources/topic/vta/tutorials/optimize/convolution_opt.rst.txt
@@ -277,24 +277,21 @@ Those include:
@I.ir_module
class Module:
@T.prim_func
- def main(data: T.handle, kernel: T.handle, res: T.handle):
+ def main(data: T.Buffer((1, 16, 14, 14, 1, 16), "int8"), kernel: T.Buffer((16, 16, 3, 3, 16, 16), "int8"), res: T.Buffer((1, 16, 14, 14, 1, 16), "int8")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- data_1 = T.match_buffer(data, (1, 16, 14, 14, 1, 16), "int8")
- kernel_1 = T.match_buffer(kernel, (16, 16, 3, 3, 16, 16), "int8")
- res_1 = T.match_buffer(res, (1, 16, 14, 14, 1, 16), "int8")
data_buf = T.allocate([65536], "int8", "global")
kernel_buf = T.allocate([589824], "int8", "global")
res_conv = T.allocate([50176], "int32", "global")
data_buf_1 = T.buffer_decl((65536,), "int8", data=data_buf)
for i1, i2, i3, i5 in T.grid(16, 16, 16, 16):
cse_var_1: T.int32 = i3 * 16
- data_2 = T.buffer_decl((50176,), "int8", data=data_1.data)
- data_buf_1[i1 * 4096 + i2 * 256 + cse_var_1 + i5] = T.if_then_else(1 <= i2 and i2 < 15 and 1 <= i3 and i3 < 15, data_2[i1 * 3136 + i2 * 224 + cse_var_1 + i5 - 240], T.int8(0))
+ data_1 = T.buffer_decl((50176,), "int8", data=data.data)
+ data_buf_1[i1 * 4096 + i2 * 256 + cse_var_1 + i5] = T.if_then_else(1 <= i2 and i2 < 15 and 1 <= i3 and i3 < 15, data_1[i1 * 3136 + i2 * 224 + cse_var_1 + i5 - 240], T.int8(0))
kernel_buf_1 = T.buffer_decl((589824,), "int8", data=kernel_buf)
for i0, i1, i2, i3, i4, i5 in T.grid(16, 16, 3, 3, 16, 16):
cse_var_2: T.int32 = i0 * 36864 + i1 * 2304 + i2 * 768 + i3 * 256 + i4 * 16 + i5
- kernel_2 = T.buffer_decl((589824,), "int8", data=kernel_1.data)
- kernel_buf_1[cse_var_2] = kernel_2[cse_var_2]
+ kernel_1 = T.buffer_decl((589824,), "int8", data=kernel.data)
+ kernel_buf_1[cse_var_2] = kernel_1[cse_var_2]
res_conv_1 = T.buffer_decl((50176,), "int32", data=res_conv)
for co, i, j, ci in T.grid(16, 14, 14, 16):
res_conv_1[co * 3136 + i * 224 + j * 16 + ci] = 0
@@ -316,8 +313,8 @@ Those include:
res_conv_4[cse_var_7] = T.min(res_conv_3[cse_var_7], 127)
for i1, i2, i3, i5 in T.grid(16, 14, 14, 16):
cse_var_8: T.int32 = i1 * 3136 + i2 * 224 + i3 * 16 + i5
- res_2 = T.buffer_decl((50176,), "int8", data=res_1.data)
- res_2[cse_var_8] = T.Cast("int8", res_conv_4[cse_var_8])
+ res_1 = T.buffer_decl((50176,), "int8", data=res.data)
+ res_1[cse_var_8] = T.Cast("int8", res_conv_4[cse_var_8])
@@ -431,24 +428,21 @@ below.
@I.ir_module
class Module:
@T.prim_func
- def main(data: T.handle, kernel: T.handle, res: T.handle):
+ def main(data: T.Buffer((1, 16, 14, 14, 1, 16), "int8"), kernel: T.Buffer((16, 16, 3, 3, 16, 16), "int8"), res: T.Buffer((1, 16, 14, 14, 1, 16), "int8")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- data_1 = T.match_buffer(data, (1, 16, 14, 14, 1, 16), "int8")
- kernel_1 = T.match_buffer(kernel, (16, 16, 3, 3, 16, 16), "int8")
- res_1 = T.match_buffer(res, (1, 16, 14, 14, 1, 16), "int8")
data_buf = T.allocate([65536], "int8", "global")
kernel_buf = T.allocate([589824], "int8", "global")
res_conv = T.allocate([25088], "int32", "global")
data_buf_1 = T.buffer_decl((65536,), "int8", data=data_buf)
for i1, i2, i3, i5 in T.grid(16, 16, 16, 16):
cse_var_1: T.int32 = i3 * 16
- data_2 = T.buffer_decl((50176,), "int8", data=data_1.data)
- data_buf_1[i1 * 4096 + i2 * 256 + cse_var_1 + i5] = T.if_then_else(1 <= i2 and i2 < 15 and 1 <= i3 and i3 < 15, data_2[i1 * 3136 + i2 * 224 + cse_var_1 + i5 - 240], T.int8(0))
+ data_1 = T.buffer_decl((50176,), "int8", data=data.data)
+ data_buf_1[i1 * 4096 + i2 * 256 + cse_var_1 + i5] = T.if_then_else(1 <= i2 and i2 < 15 and 1 <= i3 and i3 < 15, data_1[i1 * 3136 + i2 * 224 + cse_var_1 + i5 - 240], T.int8(0))
kernel_buf_1 = T.buffer_decl((589824,), "int8", data=kernel_buf)
for i0, i1, i2, i3, i4, i5 in T.grid(16, 16, 3, 3, 16, 16):
cse_var_2: T.int32 = i0 * 36864 + i1 * 2304 + i2 * 768 + i3 * 256 + i4 * 16 + i5
- kernel_2 = T.buffer_decl((589824,), "int8", data=kernel_1.data)
- kernel_buf_1[cse_var_2] = kernel_2[cse_var_2]
+ kernel_1 = T.buffer_decl((589824,), "int8", data=kernel.data)
+ kernel_buf_1[cse_var_2] = kernel_1[cse_var_2]
for i2_outer in range(2):
res_conv_1 = T.buffer_decl((157351936,), "int32", data=res_conv)
for co_init, i_init, j_init, ci_init in T.grid(8, 7, 14, 16):
@@ -486,9 +480,9 @@ below.
cse_var_17: T.int32 = i3_inner * 16
cse_var_16: T.int32 = i1_inner * 1568 + cse_var_18 + cse_var_17 + i5
cse_var_15: T.int32 = i1_inner * 3136 + i2_outer * 1568 + cse_var_18 + cse_var_17 + i5
- res_2 = T.buffer_decl((50176,), "int8", data=res_1.data)
- res_2[cse_var_15] = T.Cast("int8", res_conv_4[cse_var_16])
- res_2[cse_var_15 + 25088] = T.Cast("int8", res_conv_4[cse_var_16 + 12544])
+ res_1 = T.buffer_decl((50176,), "int8", data=res.data)
+ res_1[cse_var_15] = T.Cast("int8", res_conv_4[cse_var_16])
+ res_1[cse_var_15 + 25088] = T.Cast("int8", res_conv_4[cse_var_16 + 12544])
@@ -573,11 +567,8 @@ and mapping the shift, and clipping computation to the vector ALU.
@I.ir_module
class Module:
@T.prim_func
- def main(data: T.handle, kernel: T.handle, res: T.handle):
+ def main(data: T.Buffer((1, 16, 14, 14, 1, 16), "int8"), kernel: T.Buffer((16, 16, 3, 3, 16, 16), "int8"), res: T.Buffer((1, 16, 14, 14, 1, 16), "int8")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- data_1 = T.match_buffer(data, (1, 16, 14, 14, 1, 16), "int8")
- kernel_1 = T.match_buffer(kernel, (16, 16, 3, 3, 16, 16), "int8")
- res_1 = T.match_buffer(res, (1, 16, 14, 14, 1, 16), "int8")
T.tir.vta.coproc_dep_push(3, 2)
T.tir.vta.coproc_dep_push(3, 2)
for i2_outer in range(2):
@@ -602,13 +593,13 @@ and mapping the shift, and clipping computation to the vector ALU.
cse_var_1: T.int32 = ic_outer * 196 + i2_outer * 98 + cse_var_4 * 14 - 14
with T.attr(T.iter_var(vta, None, "ThreadIndex", "vta"), "coproc_scope", 1):
T.tir.vta.coproc_dep_pop(2, 1)
- T.call_extern("int32", "VTALoadBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), data_1.data, cse_var_1, 14, cse_var_2, 14, 1, cse_var_4, 1, cse_var_3, 0, 2)
- T.call_extern("int32", "VTALoadBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), kernel_1.data, cse_var_5, 9, 8, 144, 0, 0, 0, 0, 0, 1)
+ T.call_extern("int32", "VTALoadBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), data.data, cse_var_1, 14, cse_var_2, 14, 1, cse_var_4, 1, cse_var_3, 0, 2)
+ T.call_extern("int32", "VTALoadBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), kernel.data, cse_var_5, 9, 8, 144, 0, 0, 0, 0, 0, 1)
T.tir.vta.coproc_dep_push(1, 2)
with T.attr(T.iter_var(vta, None, "ThreadIndex", "vta"), "coproc_scope", 1):
T.tir.vta.coproc_dep_pop(2, 1)
- T.call_extern("int32", "VTALoadBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), data_1.data, cse_var_1, 14, cse_var_2, 14, 1, cse_var_4, 1, cse_var_3, 144, 2)
- T.call_extern("int32", "VTALoadBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), kernel_1.data, cse_var_5 + 1152, 9, 8, 144, 0, 0, 0, 0, 72, 1)
+ T.call_extern("int32", "VTALoadBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), data.data, cse_var_1, 14, cse_var_2, 14, 1, cse_var_4, 1, cse_var_3, 144, 2)
+ T.call_extern("int32", "VTALoadBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), kernel.data, cse_var_5 + 1152, 9, 8, 144, 0, 0, 0, 0, 72, 1)
T.tir.vta.coproc_dep_push(1, 2)
for cthread_s in range(2):
T.attr(T.iter_var(vta, None, "ThreadIndex", "vta"), "coproc_scope", 2)
@@ -644,7 +635,7 @@ and mapping the shift, and clipping computation to the vector ALU.
T.tir.vta.coproc_dep_pop(2, 3)
for i1_inner, i2_inner, i3_inner in T.grid(8, 7, 14):
cse_var_8: T.int32 = i2_inner * 14
- T.call_extern("int32", "VTAStoreBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), cthread_s * 784 + i1_inner * 98 + cse_var_8 + i3_inner, 4, res_1.data, cthread_s * 1568 + i1_inner * 196 + i2_outer * 98 + cse_var_8 + i3_inner, 1, 1, 1)
+ T.call_extern("int32", "VTAStoreBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), cthread_s * 784 + i1_inner * 98 + cse_var_8 + i3_inner, 4, res.data, cthread_s * 1568 + i1_inner * 196 + i2_outer * 98 + cse_var_8 + i3_inner, 1, 1, 1)
T.tir.vta.coproc_dep_push(3, 2)
T.tir.vta.coproc_dep_pop(3, 2)
T.tir.vta.coproc_dep_pop(3, 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 d7eec75c7e..712b7f7a10 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
@@ -214,24 +214,21 @@ Those include:
@I.ir_module
class Module:
@T.prim_func
- def main(data: T.handle, weight: T.handle, res: T.handle):
+ def main(data: T.Buffer((1, 64, 1, 16), "int8"), weight: T.Buffer((64, 64, 16, 16), "int8"), res: T.Buffer((1, 64, 1, 16), "int8")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- data_1 = T.match_buffer(data, (1, 64, 1, 16), "int8")
- weight_1 = T.match_buffer(weight, (64, 64, 16, 16), "int8")
- res_1 = T.match_buffer(res, (1, 64, 1, 16), "int8")
data_buf = T.allocate([1024], "int8", "global")
weight_buf = T.allocate([1048576], "int8", "global")
res_gem = T.allocate([1024], "int32", "global")
data_buf_1 = T.buffer_decl((1024,), "int8", data=data_buf)
for i1, i3 in T.grid(64, 16):
cse_var_1: T.int32 = i1 * 16 + i3
- data_2 = T.buffer_decl((1024,), "int8", data=data_1.data)
- data_buf_1[cse_var_1] = data_2[cse_var_1]
+ data_1 = T.buffer_decl((1024,), "int8", data=data.data)
+ data_buf_1[cse_var_1] = data_1[cse_var_1]
weight_buf_1 = T.buffer_decl((1048576,), "int8", data=weight_buf)
for i0, i1, i2, i3 in T.grid(64, 64, 16, 16):
cse_var_2: T.int32 = i0 * 16384 + i1 * 256 + i2 * 16 + i3
- weight_2 = T.buffer_decl((1048576,), "int8", data=weight_1.data)
- weight_buf_1[cse_var_2] = weight_2[cse_var_2]
+ weight_1 = T.buffer_decl((1048576,), "int8", data=weight.data)
+ weight_buf_1[cse_var_2] = weight_1[cse_var_2]
res_gem_1 = T.buffer_decl((1024,), "int32", data=res_gem)
for co, ci in T.grid(64, 16):
res_gem_1[co * 16 + ci] = 0
@@ -252,8 +249,8 @@ Those include:
res_gem_4[cse_var_6] = T.min(res_gem_3[cse_var_6], 127)
for i1, i3 in T.grid(64, 16):
cse_var_7: T.int32 = i1 * 16 + i3
- res_2 = T.buffer_decl((1024,), "int8", data=res_1.data)
- res_2[cse_var_7] = T.Cast("int8", res_gem_4[cse_var_7])
+ res_1 = T.buffer_decl((1024,), "int8", data=res.data)
+ res_1[cse_var_7] = T.Cast("int8", res_gem_4[cse_var_7])
@@ -361,24 +358,21 @@ below:
@I.ir_module
class Module:
@T.prim_func
- def main(data: T.handle, weight: T.handle, res: T.handle):
+ def main(data: T.Buffer((1, 64, 1, 16), "int8"), weight: T.Buffer((64, 64, 16, 16), "int8"), res: T.Buffer((1, 64, 1, 16), "int8")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- data_1 = T.match_buffer(data, (1, 64, 1, 16), "int8")
- weight_1 = T.match_buffer(weight, (64, 64, 16, 16), "int8")
- res_1 = T.match_buffer(res, (1, 64, 1, 16), "int8")
data_buf = T.allocate([1024], "int8", "global")
weight_buf = T.allocate([1048576], "int8", "global")
res_gem = T.allocate([256], "int32", "global")
data_buf_1 = T.buffer_decl((1024,), "int8", data=data_buf)
for i1, i3 in T.grid(64, 16):
cse_var_1: T.int32 = i1 * 16 + i3
- data_2 = T.buffer_decl((1024,), "int8", data=data_1.data)
- data_buf_1[cse_var_1] = data_2[cse_var_1]
+ data_1 = T.buffer_decl((1024,), "int8", data=data.data)
+ data_buf_1[cse_var_1] = data_1[cse_var_1]
weight_buf_1 = T.buffer_decl((1048576,), "int8", data=weight_buf)
for i0, i1, i2, i3 in T.grid(64, 64, 16, 16):
cse_var_2: T.int32 = i0 * 16384 + i1 * 256 + i2 * 16 + i3
- weight_2 = T.buffer_decl((1048576,), "int8", data=weight_1.data)
- weight_buf_1[cse_var_2] = weight_2[cse_var_2]
+ weight_1 = T.buffer_decl((1048576,), "int8", data=weight.data)
+ weight_buf_1[cse_var_2] = weight_1[cse_var_2]
for i1_outer in range(4):
res_gem_1 = T.buffer_decl((256,), "int32", data=res_gem)
for co_init, ci_init in T.grid(16, 16):
@@ -400,8 +394,8 @@ below:
res_gem_4[cse_var_6] = T.min(res_gem_3[cse_var_6], 127)
for i1_inner, i3 in T.grid(16, 16):
cse_var_7: T.int32 = i1_inner * 16
- res_2 = T.buffer_decl((1024,), "int8", data=res_1.data)
- res_2[i1_outer * 256 + cse_var_7 + i3] = T.Cast("int8", res_gem_4[cse_var_7 + i3])
+ res_1 = T.buffer_decl((1024,), "int8", data=res.data)
+ res_1[i1_outer * 256 + cse_var_7 + i3] = T.Cast("int8", res_gem_4[cse_var_7 + i3])
@@ -486,11 +480,8 @@ and mapping the shift, and clipping computation to the vector ALU.
@I.ir_module
class Module:
@T.prim_func
- def main(data: T.handle, weight: T.handle, res: T.handle):
+ def main(data: T.Buffer((1, 64, 1, 16), "int8"), weight: T.Buffer((64, 64, 16, 16), "int8"), res: T.Buffer((1, 64, 1, 16), "int8")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- data_1 = T.match_buffer(data, (1, 64, 1, 16), "int8")
- weight_1 = T.match_buffer(weight, (64, 64, 16, 16), "int8")
- res_1 = T.match_buffer(res, (1, 64, 1, 16), "int8")
T.tir.vta.coproc_dep_push(3, 2)
for i1_outer in range(4):
vta = T.var("int32")
@@ -505,8 +496,8 @@ and mapping the shift, and clipping computation to the vector ALU.
cse_var_1: T.int32 = ic_outer * 16
with T.attr(T.iter_var(vta, None, "ThreadIndex", "vta"), "coproc_scope", 1):
T.tir.vta.coproc_dep_pop(2, 1)
- T.call_extern("int32", "VTALoadBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), data_1.data, cse_var_1, 16, 1, 16, 0, 0, 0, 0, 0, 2)
- T.call_extern("int32", "VTALoadBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), weight_1.data, i1_outer * 1024 + cse_var_1, 16, 16, 64, 0, 0, 0, 0, 0, 1)
+ T.call_extern("int32", "VTALoadBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), data.data, cse_var_1, 16, 1, 16, 0, 0, 0, 0, 0, 2)
+ T.call_extern("int32", "VTALoadBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), weight.data, i1_outer * 1024 + cse_var_1, 16, 16, 64, 0, 0, 0, 0, 0, 1)
T.tir.vta.coproc_dep_push(1, 2)
T.attr(T.iter_var(vta, None, "ThreadIndex", "vta"), "coproc_scope", 2)
T.tir.vta.coproc_dep_pop(1, 2)
@@ -535,7 +526,7 @@ and mapping the shift, and clipping computation to the vector ALU.
T.attr(T.iter_var(vta, None, "ThreadIndex", "vta"), "coproc_scope", 3)
T.tir.vta.coproc_dep_pop(2, 3)
for i1_inner in range(16):
- T.call_extern("int32", "VTAStoreBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), i1_inner, 4, res_1.data, i1_outer * 16 + i1_inner, 1, 1, 1)
+ T.call_extern("int32", "VTAStoreBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), i1_inner, 4, res.data, i1_outer * 16 + i1_inner, 1, 1, 1)
T.tir.vta.coproc_dep_push(3, 2)
T.tir.vta.coproc_sync()
T.tir.vta.coproc_dep_pop(3, 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 f2497c6afb..b53df5cf7c 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
=================
-**00:03.152** total execution time for **topic_vta_tutorials_optimize** files:
+**00:03.115** total execution time for **topic_vta_tutorials_optimize** files:
+--------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_topic_vta_tutorials_optimize_convolution_opt.py` (``convolution_opt.py``) | 00:02.693 | 0.0 MB |
+| :ref:`sphx_glr_topic_vta_tutorials_optimize_convolution_opt.py` (``convolution_opt.py``) | 00:02.666 | 0.0 MB |
+--------------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_topic_vta_tutorials_optimize_matrix_multiply_opt.py` (``matrix_multiply_opt.py``) | 00:00.459 | 0.0 MB |
+| :ref:`sphx_glr_topic_vta_tutorials_optimize_matrix_multiply_opt.py` (``matrix_multiply_opt.py``) | 00:00.448 | 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 6f24491772..fdcd43cb0d 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
=================
-**00:00.817** total execution time for **topic_vta_tutorials** files:
+**00:00.786** total execution time for **topic_vta_tutorials** files:
+---------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_topic_vta_tutorials_matrix_multiply.py` (``matrix_multiply.py``) | 00:00.434 | 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``) | 00:00.383 | 0.0 MB |
+| :ref:`sphx_glr_topic_vta_tutorials_vta_get_started.py` (``vta_get_started.py``) | 00:00.373 | 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 ae51f76072..afad956840 100644
--- a/docs/_sources/topic/vta/tutorials/vta_get_started.rst.txt
+++ b/docs/_sources/topic/vta/tutorials/vta_get_started.rst.txt
@@ -355,31 +355,28 @@ After we construct the schedule, by default the schedule computes
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1, 64, 1, 16), "int32"), B: T.Buffer((1, 64, 1, 16), "int32"), C: T.Buffer((1, 64, 1, 16), "int8")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1, 64, 1, 16), "int32")
- B_1 = T.match_buffer(B, (1, 64, 1, 16), "int32")
- C_1 = T.match_buffer(C, (1, 64, 1, 16), "int8")
A_buf = T.allocate([1024], "int32", "global")
B_buf = T.allocate([1024], "int32", "global")
A_buf_1 = T.buffer_decl((1024,), "int32", data=A_buf)
for i1, i3 in T.grid(64, 16):
cse_var_1: T.int32 = i1 * 16 + i3
- A_2 = T.buffer_decl((1024,), "int32", data=A_1.data)
- A_buf_1[cse_var_1] = A_2[cse_var_1]
+ A_1 = T.buffer_decl((1024,), "int32", data=A.data)
+ A_buf_1[cse_var_1] = A_1[cse_var_1]
B_buf_1 = T.buffer_decl((1024,), "int32", data=B_buf)
for i1, i3 in T.grid(64, 16):
cse_var_2: T.int32 = i1 * 16 + i3
- B_2 = T.buffer_decl((1024,), "int32", data=B_1.data)
- B_buf_1[cse_var_2] = B_2[cse_var_2]
+ B_1 = T.buffer_decl((1024,), "int32", data=B.data)
+ B_buf_1[cse_var_2] = B_1[cse_var_2]
A_buf_2 = T.buffer_decl((1024,), "int32", data=A_buf)
for i1, i3 in T.grid(64, 16):
cse_var_3: T.int32 = i1 * 16 + i3
A_buf_2[cse_var_3] = A_buf_1[cse_var_3] + B_buf_1[cse_var_3]
for i1, i3 in T.grid(64, 16):
cse_var_4: T.int32 = i1 * 16 + i3
- C_2 = T.buffer_decl((1024,), "int8", data=C_1.data)
- C_2[cse_var_4] = T.Cast("int8", A_buf_2[cse_var_4])
+ C_1 = T.buffer_decl((1024,), "int8", data=C.data)
+ C_1[cse_var_4] = T.Cast("int8", A_buf_2[cse_var_4])
@@ -490,15 +487,12 @@ with an :code:`env.alu` pragma.
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1, 64, 1, 16), "int32"), B: T.Buffer((1, 64, 1, 16), "int32"), C: T.Buffer((1, 64, 1, 16), "int8")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1, 64, 1, 16), "int32")
- B_1 = T.match_buffer(B, (1, 64, 1, 16), "int32")
- C_1 = T.match_buffer(C, (1, 64, 1, 16), "int8")
vta = T.var("int32")
with T.attr(T.iter_var(vta, None, "ThreadIndex", "vta"), "coproc_scope", 2):
- T.call_extern("int32", "VTALoadBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), A_1.data, 0, 64, 1, 64, 0, 0, 0, 0, 0, 3)
- T.call_extern("int32", "VTALoadBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), B_1.data, 0, 64, 1, 64, 0, 0, 0, 0, 64, 3)
+ T.call_extern("int32", "VTALoadBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), A.data, 0, 64, 1, 64, 0, 0, 0, 0, 0, 3)
+ T.call_extern("int32", "VTALoadBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), B.data, 0, 64, 1, 64, 0, 0, 0, 0, 64, 3)
with T.attr(T.iter_var(vta, None, "ThreadIndex", "vta"), "coproc_uop_scope", "VTAPushALUOp"):
T.call_extern("int32", "VTAUopLoopBegin", 64, 1, 1, 0)
T.tir.vta.uop_push(1, 0, 0, 64, 0, 2, 0, 0)
@@ -506,7 +500,7 @@ with an :code:`env.alu` pragma.
T.tir.vta.coproc_dep_push(2, 3)
with T.attr(T.iter_var(vta, None, "ThreadIndex", "vta"), "coproc_scope", 3):
T.tir.vta.coproc_dep_pop(2, 3)
- T.call_extern("int32", "VTAStoreBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), 0, 4, C_1.data, 0, 64, 1, 64)
+ T.call_extern("int32", "VTAStoreBuffer2D", T.tvm_thread_context(T.tir.vta.command_handle()), 0, 4, C.data, 0, 64, 1, 64)
T.tir.vta.coproc_sync()
diff --git a/docs/_sources/tutorial/auto_scheduler_matmul_x86.rst.txt b/docs/_sources/tutorial/auto_scheduler_matmul_x86.rst.txt
index 743c33ed7a..cd82899457 100644
--- a/docs/_sources/tutorial/auto_scheduler_matmul_x86.rst.txt
+++ b/docs/_sources/tutorial/auto_scheduler_matmul_x86.rst.txt
@@ -207,6 +207,13 @@ trials, we can load the best schedule from the log file and apply it.
+.. rst-class:: sphx-glr-script-out
+
+ .. code-block:: none
+
+ *E
+
+
@@ -239,18 +246,14 @@ operator fusion.
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle, out: T.handle):
+ def main(A: T.Buffer((1024, 1024), "float32"), B: T.Buffer((1024, 1024), "float32"), C: T.Buffer((1024, 1024), "float32"), out: T.Buffer((1024, 1024), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 1024))
- B_1 = T.match_buffer(B, (1024, 1024))
- C_1 = T.match_buffer(C, (1024, 1024))
- out_1 = T.match_buffer(out, (1024, 1024))
auto_scheduler_layout_transform = T.allocate([1048576], "float32", "global")
auto_scheduler_layout_transform_1 = T.buffer_decl((1048576,), data=auto_scheduler_layout_transform)
for ax0_ax1_fused_ax2_fused in T.parallel(128):
for ax4, ax6, ax7 in T.grid(256, 4, 8):
- B_2 = T.buffer_decl((1048576,), data=B_1.data)
- auto_scheduler_layout_transform_1[ax0_ax1_fused_ax2_fused * 8192 + ax4 * 32 + ax6 * 8 + ax7] = B_2[ax4 * 4096 + ax6 * 1024 + ax0_ax1_fused_ax2_fused * 8 + ax7]
+ B_1 = T.buffer_decl((1048576,), data=B.data)
+ auto_scheduler_layout_transform_1[ax0_ax1_fused_ax2_fused * 8192 + ax4 * 32 + ax6 * 8 + ax7] = B_1[ax4 * 4096 + ax6 * 1024 + ax0_ax1_fused_ax2_fused * 8 + ax7]
for i_outer_outer_j_outer_outer_fused in T.parallel(16384):
matmul = T.allocate([4], "float32x8", "global")
for i_outer_inner in range(2):
@@ -262,16 +265,16 @@ operator fusion.
for k_outer, k_inner in T.grid(256, 4):
cse_var_2: T.int32 = i_outer_outer_j_outer_outer_fused % 128 * 8192 + k_outer * 32 + k_inner * 8
cse_var_1: T.int32 = i_outer_outer_j_outer_outer_fused // 128 * 8192 + i_outer_inner * 4096 + k_outer * 4 + k_inner
- A_2 = T.buffer_decl((1048576,), data=A_1.data)
- matmul_1[0] = matmul_1[0] + T.Broadcast(A_2[cse_var_1], 8) * auto_scheduler_layout_transform_1[cse_var_2:cse_var_2 + 8]
- matmul_1[1] = matmul_1[1] + T.Broadcast(A_2[cse_var_1 + 1024], 8) * auto_scheduler_layout_transform_1[cse_var_2:cse_var_2 + 8]
- matmul_1[2] = matmul_1[2] + T.Broadcast(A_2[cse_var_1 + 2048], 8) * auto_scheduler_layout_transform_1[cse_var_2:cse_var_2 + 8]
- matmul_1[3] = matmul_1[3] + T.Broadcast(A_2[cse_var_1 + 3072], 8) * auto_scheduler_layout_transform_1[cse_var_2:cse_var_2 + 8]
+ A_1 = T.buffer_decl((1048576,), data=A.data)
+ matmul_1[0] = matmul_1[0] + T.Broadcast(A_1[cse_var_1], 8) * auto_scheduler_layout_transform_1[cse_var_2:cse_var_2 + 8]
+ matmul_1[1] = matmul_1[1] + T.Broadcast(A_1[cse_var_1 + 1024], 8) * auto_scheduler_layout_transform_1[cse_var_2:cse_var_2 + 8]
+ matmul_1[2] = matmul_1[2] + T.Broadcast(A_1[cse_var_1 + 2048], 8) * auto_scheduler_layout_transform_1[cse_var_2:cse_var_2 + 8]
+ matmul_1[3] = matmul_1[3] + T.Broadcast(A_1[cse_var_1 + 3072], 8) * auto_scheduler_layout_transform_1[cse_var_2:cse_var_2 + 8]
for i_inner in range(4):
cse_var_3: T.int32 = i_outer_outer_j_outer_outer_fused // 128 * 8192 + i_outer_inner * 4096 + i_inner * 1024 + i_outer_outer_j_outer_outer_fused % 128 * 8
- out_2 = T.buffer_decl((1048576,), data=out_1.data)
- C_2 = T.buffer_decl((1048576,), data=C_1.data)
- out_2[cse_var_3:cse_var_3 + 8] = matmul_1[i_inner] + C_2[cse_var_3:cse_var_3 + 8]
+ out_1 = T.buffer_decl((1048576,), data=out.data)
+ C_1 = T.buffer_decl((1048576,), data=C.data)
+ out_1[cse_var_3:cse_var_3 + 8] = matmul_1[i_inner] + C_1[cse_var_3:cse_var_3 + 8]
@@ -319,7 +322,7 @@ We build the binary and check its correctness and performance.
.. code-block:: none
- Execution time of this operator: 97.015 ms
+ Execution time of this operator: 96.538 ms
@@ -437,7 +440,7 @@ operations.
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 1 minutes 20.459 seconds)
+ **Total running time of the script:** ( 1 minutes 30.420 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 e9d292ee1d..3a04adbaf0 100644
--- a/docs/_sources/tutorial/autotvm_matmul_x86.rst.txt
+++ b/docs/_sources/tutorial/autotvm_matmul_x86.rst.txt
@@ -454,16 +454,16 @@ reduce variance, we take 5 measurements and average them.
waiting for device...
device available
Get devices for measurement successfully!
- No: 1 GFLOPS: 11.82/11.82 result: MeasureResult(costs=(0.0227024582,), error_no=MeasureErrorNo.NO_ERROR, all_cost=0.6222848892211914, timestamp=1674173522.5541093) [('tile_y', [-1, 256]), ('tile_x', [-1, 256])],None,88
- No: 2 GFLOPS: 0.50/11.82 result: MeasureResult(costs=(0.5333215642,), error_no=MeasureErrorNo.NO_ERROR, all_cost=8.763290643692017, timestamp=1674173531.341208) [('tile_y', [-1, 256]), ('tile_x', [-1, 1])],None,8
- No: 3 GFLOPS: 1.98/11.82 result: MeasureResult(costs=(0.13528974740000002,), error_no=MeasureErrorNo.NO_ERROR, all_cost=2.397447109222412, timestamp=1674173534.5228543) [('tile_y', [-1, 2]), ('tile_x', [-1, 4])],None,21
- No: 4 GFLOPS: 13.79/13.79 result: MeasureResult(costs=(0.0194591336,), error_no=MeasureErrorNo.NO_ERROR, all_cost=0.5597658157348633, timestamp=1674173535.0910823) [('tile_y', [-1, 128]), ('tile_x', [-1, 64])],None,67
- No: 5 GFLOPS: 7.38/13.79 result: MeasureResult(costs=(0.0363800002,), error_no=MeasureErrorNo.NO_ERROR, all_cost=0.9181301593780518, timestamp=1674173536.297884) [('tile_y', [-1, 1]), ('tile_x', [-1, 128])],None,70
- No: 6 GFLOPS: 9.81/13.79 result: MeasureResult(costs=(0.0273711362,), error_no=MeasureErrorNo.NO_ERROR, all_cost=0.9397749900817871, timestamp=1674173537.7699847) [('tile_y', [-1, 2]), ('tile_x', [-1, 128])],None,71
- No: 7 GFLOPS: 11.62/13.79 result: MeasureResult(costs=(0.0231036638,), error_no=MeasureErrorNo.NO_ERROR, all_cost=0.6368618011474609, timestamp=1674173539.1731124) [('tile_y', [-1, 32]), ('tile_x', [-1, 256])],None,85
- No: 8 GFLOPS: 9.00/13.79 result: MeasureResult(costs=(0.0298130234,), error_no=MeasureErrorNo.NO_ERROR, all_cost=0.786644697189331, timestamp=1674173539.9062722) [('tile_y', [-1, 512]), ('tile_x', [-1, 64])],None,69
- No: 9 GFLOPS: 11.59/13.79 result: MeasureResult(costs=(0.0231601226,), error_no=MeasureErrorNo.NO_ERROR, all_cost=0.6171057224273682, timestamp=1674173540.6359928) [('tile_y', [-1, 16]), ('tile_x', [-1, 256])],None,84
- No: 10 GFLOPS: 3.58/13.79 result: MeasureResult(costs=(0.074947897,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.402517318725586, timestamp=1674173542.083198) [('tile_y', [-1, 16]), ('tile_x', [-1, 8])],None,34
+ No: 1 GFLOPS: 4.18/4.18 result: MeasureResult(costs=(0.06415541500000001,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.287381649017334, timestamp=1674183947.8544009) [('tile_y', [-1, 16]), ('tile_x', [-1, 16])],None,44
+ No: 2 GFLOPS: 2.70/4.18 result: MeasureResult(costs=(0.099422066,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.8374905586242676, timestamp=1674183949.698981) [('tile_y', [-1, 512]), ('tile_x', [-1, 16])],None,49
+ No: 3 GFLOPS: 8.37/8.37 result: MeasureResult(costs=(0.032066305,), error_no=MeasureErrorNo.NO_ERROR, all_cost=0.7405209541320801, timestamp=1674183951.2165217) [('tile_y', [-1, 1]), ('tile_x', [-1, 32])],None,50
+ No: 4 GFLOPS: 3.71/8.37 result: MeasureResult(costs=(0.07238791139999999,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.3979723453521729, timestamp=1674183953.3796506) [('tile_y', [-1, 128]), ('tile_x', [-1, 16])],None,47
+ No: 5 GFLOPS: 10.89/10.89 result: MeasureResult(costs=(0.0246521962,), error_no=MeasureErrorNo.NO_ERROR, all_cost=0.6746640205383301, timestamp=1674183954.9166691) [('tile_y', [-1, 4]), ('tile_x', [-1, 128])],None,72
+ No: 6 GFLOPS: 1.77/10.89 result: MeasureResult(costs=(0.1514064862,), error_no=MeasureErrorNo.NO_ERROR, all_cost=2.660912036895752, timestamp=1674183957.5901525) [('tile_y', [-1, 16]), ('tile_x', [-1, 2])],None,14
+ No: 7 GFLOPS: 1.33/10.89 result: MeasureResult(costs=(0.20251314800000003,), error_no=MeasureErrorNo.NO_ERROR, all_cost=3.4556992053985596, timestamp=1674183961.0785851) [('tile_y', [-1, 2]), ('tile_x', [-1, 1])],None,1
+ No: 8 GFLOPS: 9.18/10.89 result: MeasureResult(costs=(0.029232778999999997,), error_no=MeasureErrorNo.NO_ERROR, all_cost=0.6821742057800293, timestamp=1674183961.7943165) [('tile_y', [-1, 512]), ('tile_x', [-1, 32])],None,59
+ No: 9 GFLOPS: 13.50/13.50 result: MeasureResult(costs=(0.019883790800000002,), error_no=MeasureErrorNo.NO_ERROR, all_cost=0.5237207412719727, timestamp=1674183962.4325068) [('tile_y', [-1, 256]), ('tile_x', [-1, 64])],None,68
+ No: 10 GFLOPS: 1.78/13.50 result: MeasureResult(costs=(0.1507603704,), error_no=MeasureErrorNo.NO_ERROR, all_cost=2.611154317855835, timestamp=1674183965.0960453) [('tile_y', [-1, 2]), ('tile_x', [-1, 2])],None,11
diff --git a/docs/_sources/tutorial/autotvm_relay_x86.rst.txt b/docs/_sources/tutorial/autotvm_relay_x86.rst.txt
index c46b3813de..3d52bf14a5 100644
--- a/docs/_sources/tutorial/autotvm_relay_x86.rst.txt
+++ b/docs/_sources/tutorial/autotvm_relay_x86.rst.txt
@@ -311,7 +311,7 @@ standard deviation.
.. code-block:: none
- {'mean': 512.2079894000011, 'median': 512.4371018000033, 'std': 2.4379111328350107}
+ {'mean': 477.48544495999687, 'median': 477.1313110499932, 'std': 1.163654270518335}
@@ -545,29 +545,29 @@ 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: 12.89/ 17.10 GFLOPS | Progress: (4/20) | 7.57 s
[Task 1/25] Current/Best: 9.09/ 17.97 GFLOPS | Progress: (8/20) | 12.27 s
[Task 1/25] Current/Best: 11.21/ 22.27 GFLOPS | Progress: (12/20) | 14.39 s
[Task 1/25] Current/Best: 5.63/ 22.27 GFLOPS | Progress: (16/20) | 18.13 s
[Task 1/25] Current/Best: 12.87/ 22.27 GFLOPS | Progress: (20/20) | 21.43 s Done.
-
[Task 2/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 2/25] Current/Best: 8.96/ 15.28 GFLOPS | Progress: (4/20) | 4.03 s
[Task 2/25] Current/Best: 3.30/ 19.94 GFLOPS | Progress: (8/20) | 5.70 s
[Task 2/25] Current/Best: 11.99/ 23.19 GFLOPS | Progress: (12/20) | 7.81 s
[Task 2/25] Current/Best: 10.68/ 23.19 GFLOPS | Progress: (16/20) | 9.45 s
[Task 2/25] Current/Best: 12.32/ 23.19 GFLOPS | Progress: (20/20) | 10.83 s Done.
-
[Task 3/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 3/25] Current/Best: 17.33/ 17.33 GFLOPS | Progress: (4/20) | 4.08 s
[Task 3/25] Current/Best: 3.10/ 22.78 GFLOPS | Progress: (8/20) | 7.84 s
[Task 3/25] Current/Best: 18.53/ 22.78 GFLOPS | Progress: (12/20) | 10.33 s
[Task 3/25] Current/Best: 12.38/ 22.78 GFLOPS | Progress: (16/20) | 13.01 s
[Task 3/25] Current/Best: 8.25/ 22.78 GFLOPS | Progress: (20/20) | 16.16 s Done.
-
[Task 4/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 4/25] Current/Best: 6.52/ 22.10 GFLOPS | Progress: (4/20) | 4.04 s
[Task 4/25] Current/Best: 14.61/ 22.10 GFLOPS | Progress: (8/20) | 6.40 s
[Task 4/25] Current/Best: 11.67/ 22.10 GFLOPS | Progress: (12/20) | 11.79 s
[Task 4/25] Current/Best: 17.22/ 22.10 GFLOPS | Progress: (16/20) | 13.58 s
[Task 4/25] Current/Best: 11.33/ 22.10 GFLOPS | Progress: (20/20) | 21.09 s Done.
-
[Task 5/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 5/25] Current/Best: 10.96/ 18.91 GFLOPS | Progress: (4/20) | 3.87 s
[Task 5/25] Current/Best: 16.94/ 18.91 GFLOPS | Progress: (8/20) | 6.47 s
[Task 5/25] Current/Best: 4.43/ 18.91 GFLOPS | Progress: (12/20) | 8.71 s
[Task 5/25] Current/Best: 4.38/ 18.91 GFLOPS | Progress: (16/20) | 11.28 s
[Task 5/25] Current/Best: 10.97/ 18.91 GFLOPS | Progress: (20/20) | 13.18 s Done.
-
[Task 6/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 6/25] Current/Best: 2.39/ 15.94 GFLOPS | Progress: (4/20) | 4.98 s
[Task 6/25] Current/Best: 5.92/ 15.94 GFLOPS | Progress: (8/20) | 9.66 s
[Task 6/25] Current/Best: 4.78/ 15.94 GFLOPS | Progress: (12/20) | 12.57 s
[Task 6/25] Current/Best: 17.79/ 17.79 GFLOPS | Progress: (16/20) | 19.57 s
[Task 6/25] Current/Best: 20.14/ 20.14 GFLOPS | Progress: (20/20) | 22.19 s Done.
-
[Task 7/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 7/25] Current/Best: 18.59/ 18.59 GFLOPS | Progress: (4/20) | 4.29 s
[Task 7/25] Current/Best: 11.17/ 18.59 GFLOPS | Progress: (8/20) | 6.72 s
[Task 7/25] Current/Best: 11.52/ 18.59 GFLOPS | Progress: (12/20) | 9.78 s
[Task 7/25] Current/Best: 14.23/ 22.64 GFLOPS | Progress: (16/20) | 12.62 s
[Task 7/25] Current/Best: 18.86/ 22.64 GFLOPS | Progress: (20/20) | 14.63 s Done.
-
[Task 8/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 8/25] Current/Best: 6.17/ 14.25 GFLOPS | Progress: (4/20) | 8.28 s
[Task 8/25] Current/Best: 9.73/ 15.95 GFLOPS | Progress: (8/20) | 12.96 s
[Task 8/25] Current/Best: 2.65/ 20.35 GFLOPS | Progress: (12/20) | 23.01 s
[Task 8/25] Current/Best: 18.64/ 20.35 GFLOPS | Progress: (16/20) | 25.38 s
[Task 8/25] Current/Best: 9.79/ 20.35 GFLOPS | Progress: (20/20) | 32.74 s Done.
-
[Task 9/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 9/25] Current/Best: 10.35/ 21.47 GFLOPS | Progress: (4/20) | 3.49 s
[Task 9/25] Current/Best: 7.04/ 21.47 GFLOPS | Progress: (8/20) | 9.50 s
[Task 9/25] Current/Best: 7.56/ 21.47 GFLOPS | Progress: (12/20) | 13.80 s
[Task 9/25] Current/Best: 18.41/ 22.63 GFLOPS | Progress: (16/20) | 16.64 s
[Task 9/25] Current/Best: 13.80/ 22.63 GFLOPS | Progress: (20/20) | 18.92 s Done.
-
[Task 10/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 10/25] Current/Best: 5.15/ 15.01 GFLOPS | Progress: (4/20) | 4.22 s
[Task 10/25] Current/Best: 10.88/ 16.14 GFLOPS | Progress: (8/20) | 6.48 s
[Task 10/25] Current/Best: 5.22/ 16.50 GFLOPS | Progress: (12/20) | 9.20 s
[Task 10/25] Current/Best: 13.77/ 18.22 GFLOPS | Progress: (16/20) | 11.45 s
[Task 10/25] Current/Best: 9.70/ 19.70 GFLOPS | Progress: (20/20) | 13.85 s Done.
-
[Task 11/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 11/25] Current/Best: 7.01/ 23.16 GFLOPS | Progress: (4/20) | 4.25 s
[Task 11/25] Current/Best: 6.17/ 23.95 GFLOPS | Progress: (8/20) | 6.87 s
[Task 11/25] Current/Best: 3.11/ 23.95 GFLOPS | Progress: (12/20) | 10.89 s
[Task 11/25] Current/Best: 6.18/ 23.95 GFLOPS | Progress: (16/20) | 14.62 s
[Task 11/25] Current/Best: 21.76/ 23.95 GFLOPS | Progress: (20/20) | 17.73 s Done.
-
[Task 12/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 12/25] Current/Best: 1.59/ 15.12 GFLOPS | Progress: (4/20) | 8.62 s
[Task 12/25] Current/Best: 14.48/ 18.00 GFLOPS | Progress: (8/20) | 15.03 s
[Task 12/25] Current/Best: 19.05/ 19.44 GFLOPS | Progress: (12/20) | 16.75 s
[Task 12/25] Current/Best: 6.10/ 19.44 GFLOPS | Progress: (16/20) | 20.20 s
[Task 12/25] Current/Best: 10.71/ 19.44 GFLOPS | Progress: (20/20) | 29.91 s Done.
-
[Task 13/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 13/25] Current/Best: 16.37/ 16.37 GFLOPS | Progress: (4/20) | 5.34 s
[Task 13/25] Current/Best: 10.87/ 16.37 GFLOPS | Progress: (8/20) | 8.61 s
[Task 13/25] Current/Best: 6.20/ 18.29 GFLOPS | Progress: (12/20) | 12.23 s
[Task 13/25] Current/Best: 14.75/ 18.29 GFLOPS | Progress: (16/20) | 16.23 s
[Task 13/25] Current/Best: 15.54/ 18.29 GFLOPS | Progress: (20/20) | 19.29 s Done.
-
[Task 14/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 14/25] Current/Best: 12.54/ 12.54 GFLOPS | Progress: (4/20) | 8.71 s
[Task 14/25] Current/Best: 8.35/ 20.71 GFLOPS | Progress: (8/20) | 15.62 s
[Task 14/25] Current/Best: 14.70/ 20.71 GFLOPS | Progress: (12/20) | 18.77 s
[Task 14/25] Current/Best: 4.66/ 20.71 GFLOPS | Progress: (16/20) | 21.04 s
[Task 14/25] Current/Best: 5.19/ 20.77 GFLOPS | Progress: (20/20) | 23.43 s
[Task 15/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 15/25] Current/Best: 16.31/ 19.92 GFLOPS | Progress: (4/20) | 4.57 s
[Task 15/25] Current/Best: 16.42/ 19.92 GFLOPS | Progress: (8/20) | 6.26 s Done.
-
[Task 15/25] Current/Best: 19.08/ 19.92 GFLOPS | Progress: (12/20) | 8.86 s
[Task 15/25] Current/Best: 10.70/ 23.81 GFLOPS | Progress: (16/20) | 11.51 s
[Task 15/25] Current/Best: 10.69/ 23.81 GFLOPS | Progress: (20/20) | 13.06 s Done.
-
[Task 16/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 16/25] Current/Best: 14.15/ 14.15 GFLOPS | Progress: (4/20) | 4.21 s
[Task 16/25] Current/Best: 1.58/ 14.15 GFLOPS | Progress: (8/20) | 7.42 s
[Task 16/25] Current/Best: 1.57/ 17.66 GFLOPS | Progress: (12/20) | 11.38 s
[Task 16/25] Current/Best: 4.73/ 20.82 GFLOPS | Progress: (16/20) | 13.25 s
[Task 16/25] Current/Best: 14.30/ 20.82 GFLOPS | Progress: (20/20) | 15.65 s Done.
-
[Task 17/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 17/25] Current/Best: 9.74/ 16.64 GFLOPS | Progress: (4/20) | 4.59 s
[Task 17/25] Current/Best: 10.10/ 19.31 GFLOPS | Progress: (8/20) | 7.31 s
[Task 17/25] Current/Best: 8.23/ 19.31 GFLOPS | Progress: (12/20) | 9.87 s
[Task 17/25] Current/Best: 16.50/ 19.31 GFLOPS | Progress: (16/20) | 12.68 s
[Task 17/25] Current/Best: 7.13/ 19.31 GFLOPS | Progress: (20/20) | 15.71 s Done.
-
[Task 18/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 18/25] Current/Best: 6.32/ 20.33 GFLOPS | Progress: (4/20) | 3.86 s
[Task 18/25] Current/Best: 17.07/ 20.33 GFLOPS | Progress: (8/20) | 6.32 s
[Task 18/25] Current/Best: 13.56/ 20.33 GFLOPS | Progress: (12/20) | 8.59 s
[Task 18/25] Current/Best: 9.44/ 21.40 GFLOPS | Progress: (16/20) | 10.83 s
[Task 18/25] Current/Best: 16.07/ 21.40 GFLOPS | Progress: (20/20) | 12.73 s Done.
-
[Task 19/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 19/25] Current/Best: 20.34/ 20.34 GFLOPS | Progress: (4/20) | 7.51 s
[Task 19/25] Current/Best: 13.27/ 20.34 GFLOPS | Progress: (8/20) | 10.21 s
[Task 19/25] Current/Best: 1.55/ 20.34 GFLOPS | Progress: (12/20) | 14.43 s
[Task 19/25] Current/Best: 11.04/ 20.34 GFLOPS | Progress: (16/20) | 20.60 s
[Task 19/25] Current/Best: 21.28/ 22.64 GFLOPS | Progress: (20/20) | 25.25 s Done.
-
[Task 20/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 20/25] Current/Best: 14.63/ 14.63 GFLOPS | Progress: (4/20) | 4.19 s
[Task 20/25] Current/Best: 19.49/ 19.49 GFLOPS | Progress: (8/20) | 6.23 s
[Task 20/25] Current/Best: 10.06/ 19.49 GFLOPS | Progress: (12/20) | 9.00 s
[Task 20/25] Current/Best: 18.32/ 21.26 GFLOPS | Progress: (16/20) | 12.32 s
[Task 20/25] Current/Best: 6.80/ 21.26 GFLOPS | Progress: (20/20) | 15.74 s
[Task 21/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 21/25] Current/Best: 7.59/ 15.88 GFLOPS | Progress: (4/20) | 5.65 s Done.
-
[Task 21/25] Current/Best: 8.97/ 17.91 GFLOPS | Progress: (8/20) | 7.97 s
[Task 21/25] Current/Best: 15.47/ 17.91 GFLOPS | Progress: (12/20) | 9.95 s
[Task 21/25] Current/Best: 13.77/ 18.76 GFLOPS | Progress: (16/20) | 12.39 s
[Task 21/25] Current/Best: 8.49/ 18.76 GFLOPS | Progress: (20/20) | 14.63 s
[Task 22/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 22/25] Current/Best: 10.94/ 16.93 GFLOPS | Progress: (4/20) | 4.31 s
[Task 22/25] Current/Best: 7.28/ 16.93 GFLOPS | Progress: (8/20) | 6.08 s
[Task 22/25] Current/Best: 4.73/ 21.34 GFLOPS | Progress: (12/20) | 8.25 s
[Task 22/25] Current/Best: 9.71/ 21.34 GFLOPS | Progress: (16/20) | 11.20 s
[Task 22/25] Current/Best: 15.43/ 21.34 GFLOPS | Progress: (20/20) | 13.21 s Done.
-
[Task 23/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 23/25] Current/Best: 5.57/ 18.21 GFLOPS | Progress: (4/20) | 5.41 s
[Task 23/25] Current/Best: 6.15/ 21.77 GFLOPS | Progress: (8/20) | 8.35 s
[Task 23/25] Current/Best: 5.02/ 21.77 GFLOPS | Progress: (12/20) | 11.21 s
[Task 23/25] Current/Best: 18.52/ 21.93 GFLOPS | Progress: (16/20) | 14.24 s
[Task 23/25] Current/Best: 6.10/ 21.93 GFLOPS | Progress: (20/20) | 17.55 s Done.
-
[Task 24/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 24/25] Current/Best: 6.31/ 6.31 GFLOPS | Progress: (4/20) | 12.75 s
[Task 24/25] Current/Best: 2.65/ 6.70 GFLOPS | Progress: (8/20) | 24.89 s
[Task 24/25] Current/Best: 7.64/ 7.64 GFLOPS | Progress: (12/20) | 35.86 s
[Task 24/25] Current/Best: 5.65/ 7.64 GFLOPS | Progress: (16/20) | 47.57 s
[Task 24/25] Current/Best: 8.22/ 9.97 GFLOPS | Progress: (20/20) | 59.43 s
[Task 25/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 25/25] Current/Best: 5.48/ 5.48 GFLOPS | Progress: (4/20) | 6.88 s
[Task 25/25] Current/Best: 1.54/ 8.82 GFLOPS | Progress: (8/20) | 8.49 s
[Task 25/25] Current/Best: 3.59/ 8.82 GFLOPS | Progress: (12/20) | 19.44 s
[Task 25/25] Current/Best: 7.74/ 8.82 GFLOPS | Progress: (16/20) | 30.38 s
[Task 25/25] Current/Best: 3.03/ 8.82 GFLOPS | Progress: (20/
20) | 32.49 s
+
[Task 1/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 1/25] Current/Best: 15.30/ 18.71 GFLOPS | Progress: (4/20) | 7.85 s
[Task 1/25] Current/Best: 12.14/ 18.71 GFLOPS | Progress: (8/20) | 11.20 s
[Task 1/25] Current/Best: 6.83/ 23.82 GFLOPS | Progress: (12/20) | 14.40 s
[Task 1/25] Current/Best: 16.58/ 23.82 GFLOPS | Progress: (16/20) | 17.92 s
[Task 1/25] Current/Best: 9.74/ 23.82 GFLOPS | Progress: (20/20) | 20.11 s Done.
+
[Task 2/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 2/25] Current/Best: 12.60/ 14.27 GFLOPS | Progress: (4/20) | 3.90 s
[Task 2/25] Current/Best: 6.61/ 19.78 GFLOPS | Progress: (8/20) | 5.90 s
[Task 2/25] Current/Best: 11.01/ 19.78 GFLOPS | Progress: (12/20) | 7.84 s
[Task 2/25] Current/Best: 6.19/ 19.78 GFLOPS | Progress: (16/20) | 10.69 s
[Task 2/25] Current/Best: 14.99/ 19.92 GFLOPS | Progress: (20/20) | 11.99 s Done.
+
[Task 3/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 3/25] Current/Best: 15.85/ 18.87 GFLOPS | Progress: (4/20) | 4.02 s
[Task 3/25] Current/Best: 18.06/ 18.87 GFLOPS | Progress: (8/20) | 6.62 s
[Task 3/25] Current/Best: 11.32/ 18.87 GFLOPS | Progress: (12/20) | 8.83 s
[Task 3/25] Current/Best: 7.72/ 21.24 GFLOPS | Progress: (16/20) | 11.17 s
[Task 3/25] Current/Best: 9.67/ 21.24 GFLOPS | Progress: (20/20) | 14.17 s Done.
+
[Task 4/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 4/25] Current/Best: 13.41/ 16.65 GFLOPS | Progress: (4/20) | 6.52 s
[Task 4/25] Current/Best: 14.24/ 16.65 GFLOPS | Progress: (8/20) | 8.66 s
[Task 4/25] Current/Best: 18.32/ 18.32 GFLOPS | Progress: (12/20) | 10.77 s
[Task 4/25] Current/Best: 11.23/ 18.32 GFLOPS | Progress: (16/20) | 13.68 s
[Task 4/25] Current/Best: 14.60/ 18.47 GFLOPS | Progress: (20/20) | 15.37 s Done.
+
[Task 5/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 5/25] Current/Best: 6.07/ 16.00 GFLOPS | Progress: (4/20) | 3.69 s
[Task 5/25] Current/Best: 16.29/ 23.13 GFLOPS | Progress: (8/20) | 5.62 s
[Task 5/25] Current/Best: 14.06/ 23.13 GFLOPS | Progress: (12/20) | 7.58 s
[Task 5/25] Current/Best: 12.15/ 23.13 GFLOPS | Progress: (16/20) | 9.75 s
[Task 5/25] Current/Best: 5.66/ 23.13 GFLOPS | Progress: (20/20) | 11.70 s Done.
+
[Task 6/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 6/25] Current/Best: 12.17/ 23.11 GFLOPS | Progress: (4/20) | 4.11 s
[Task 6/25] Current/Best: 13.31/ 23.11 GFLOPS | Progress: (8/20) | 6.51 s
[Task 6/25] Current/Best: 15.13/ 23.11 GFLOPS | Progress: (12/20) | 9.36 s
[Task 6/25] Current/Best: 4.13/ 23.11 GFLOPS | Progress: (16/20) | 13.20 s
[Task 6/25] Current/Best: 2.81/ 23.11 GFLOPS | Progress: (20/20) | 17.23 s Done.
+
[Task 7/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 7/25] Current/Best: 15.43/ 18.22 GFLOPS | Progress: (4/20) | 4.48 s
[Task 7/25] Current/Best: 13.84/ 18.22 GFLOPS | Progress: (8/20) | 6.79 s
[Task 7/25] Current/Best: 12.23/ 20.78 GFLOPS | Progress: (12/20) | 8.84 s
[Task 7/25] Current/Best: 23.29/ 23.29 GFLOPS | Progress: (16/20) | 10.72 s
[Task 7/25] Current/Best: 9.72/ 23.29 GFLOPS | Progress: (20/20) | 12.77 s Done.
+
[Task 8/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 8/25] Current/Best: 3.62/ 11.80 GFLOPS | Progress: (4/20) | 11.38 s
[Task 8/25] Current/Best: 7.33/ 12.44 GFLOPS | Progress: (8/20) | 16.47 s
[Task 8/25] Current/Best: 5.14/ 13.67 GFLOPS | Progress: (12/20) | 21.18 s
[Task 8/25] Current/Best: 10.57/ 13.67 GFLOPS | Progress: (16/20) | 26.95 s
[Task 8/25] Current/Best: 2.81/ 18.58 GFLOPS | Progress: (20/20) | 34.54 s Done.
+
[Task 9/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 9/25] Current/Best: 11.01/ 13.89 GFLOPS | Progress: (4/20) | 5.91 s
[Task 9/25] Current/Best: 12.57/ 22.75 GFLOPS | Progress: (8/20) | 16.29 s
[Task 9/25] Current/Best: 16.27/ 22.75 GFLOPS | Progress: (12/20) | 27.37 s
[Task 9/25] Current/Best: 16.64/ 22.75 GFLOPS | Progress: (16/20) | 30.03 s
[Task 9/25] Current/Best: 13.22/ 22.75 GFLOPS | Progress: (20/20) | 32.51 s
[Task 10/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 10/25] Current/Best: 7.16/ 12.91 GFLOPS | Progress: (4/20) | 3.86 s
[Task 10/25] Current/Best: 9.49/ 13.08 GFLOPS | Progress: (8/20) | 6.51 s
[Task 10/25] Current/Best: 9.40/ 22.06 GFLOPS | Progress: (12/20) | 8.95 s
[Task 10/25] Current/Best: 14.26/ 22.06 GFLOPS | Progress: (16/20) | 11.30 s
[Task 10/25] Current/Best: 11.07/ 22.57 GFLOPS | Progress: (20/20
) | 13.49 s Done.
+
[Task 11/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 11/25] Current/Best: 17.02/ 17.02 GFLOPS | Progress: (4/20) | 4.13 s
[Task 11/25] Current/Best: 10.07/ 22.61 GFLOPS | Progress: (8/20) | 6.58 s
[Task 11/25] Current/Best: 16.71/ 22.61 GFLOPS | Progress: (12/20) | 10.58 s
[Task 11/25] Current/Best: 19.73/ 22.61 GFLOPS | Progress: (16/20) | 12.70 s
[Task 11/25] Current/Best: 19.42/ 22.61 GFLOPS | Progress: (20/20) | 15.44 s Done.
+
[Task 12/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 12/25] Current/Best: 16.42/ 17.67 GFLOPS | Progress: (4/20) | 3.98 s
[Task 12/25] Current/Best: 13.24/ 18.53 GFLOPS | Progress: (8/20) | 7.47 s
[Task 12/25] Current/Best: 11.77/ 18.53 GFLOPS | Progress: (12/20) | 10.42 s
[Task 12/25] Current/Best: 4.57/ 18.53 GFLOPS | Progress: (16/20) | 12.81 s
[Task 12/25] Current/Best: 5.08/ 20.63 GFLOPS | Progress: (20/20) | 15.29 s Done.
+
[Task 13/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 13/25] Current/Best: 9.86/ 20.33 GFLOPS | Progress: (4/20) | 4.75 s
[Task 13/25] Current/Best: 21.82/ 21.82 GFLOPS | Progress: (8/20) | 7.72 s
[Task 13/25] Current/Best: 16.97/ 21.82 GFLOPS | Progress: (12/20) | 11.23 s
[Task 13/25] Current/Best: 12.25/ 23.21 GFLOPS | Progress: (16/20) | 13.92 s
[Task 13/25] Current/Best: 12.19/ 23.21 GFLOPS | Progress: (20/20) | 16.11 s Done.
+
[Task 14/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 14/25] Current/Best: 13.29/ 17.68 GFLOPS | Progress: (4/20) | 3.95 s
[Task 14/25] Current/Best: 8.59/ 17.68 GFLOPS | Progress: (8/20) | 6.34 s
[Task 14/25] Current/Best: 18.86/ 18.86 GFLOPS | Progress: (12/20) | 9.78 s Done.
+
[Task 14/25] Current/Best: 10.30/ 22.75 GFLOPS | Progress: (16/20) | 12.37 s
[Task 14/25] Current/Best: 4.55/ 22.75 GFLOPS | Progress: (20/20) | 15.91 s Done.
+
[Task 15/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 15/25] Current/Best: 8.87/ 12.81 GFLOPS | Progress: (4/20) | 4.85 s
[Task 15/25] Current/Best: 15.48/ 18.88 GFLOPS | Progress: (8/20) | 6.43 s
[Task 15/25] Current/Best: 13.28/ 18.88 GFLOPS | Progress: (12/20) | 9.06 s
[Task 15/25] Current/Best: 15.18/ 20.23 GFLOPS | Progress: (16/20) | 10.78 s
[Task 15/25] Current/Best: 10.95/ 20.23 GFLOPS | Progress: (20/20) | 13.09 s
[Task 16/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 16/25] Current/Best: 7.67/ 19.12 GFLOPS | Progress: (4/20) | 4.20 s
[Task 16/25] Current/Best: 17.02/ 19.12 GFLOPS | Progress: (8/20) | 6.60 s
[Task 16/25] Current/Best: 14.04/ 19.12 GFLOPS | Progress: (12/20) | 10.29 s
[Task 16/25] Current/Best: 18.09/ 19.12 GFLOPS | Progress: (16/20) | 11.79 s
[Task 16/25] Current/Best: 16.36/ 19.12 GFLOPS | Progress: (20/20)
| 15.41 s Done.
+
[Task 17/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 17/25] Current/Best: 6.22/ 20.51 GFLOPS | Progress: (4/20) | 4.28 s
[Task 17/25] Current/Best: 12.27/ 20.51 GFLOPS | Progress: (8/20) | 7.48 s
[Task 17/25] Current/Best: 15.02/ 20.51 GFLOPS | Progress: (12/20) | 9.84 s
[Task 17/25] Current/Best: 16.16/ 20.51 GFLOPS | Progress: (16/20) | 12.51 s
[Task 17/25] Current/Best: 16.17/ 20.51 GFLOPS | Progress: (20/20) | 15.24 s Done.
+
[Task 18/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 18/25] Current/Best: 17.44/ 17.44 GFLOPS | Progress: (4/20) | 6.42 s
[Task 18/25] Current/Best: 12.29/ 19.96 GFLOPS | Progress: (8/20) | 9.49 s
[Task 18/25] Current/Best: 20.27/ 20.27 GFLOPS | Progress: (12/20) | 11.30 s
[Task 18/25] Current/Best: 11.19/ 21.02 GFLOPS | Progress: (16/20) | 17.84 s
[Task 18/25] Current/Best: 7.88/ 21.02 GFLOPS | Progress: (20/20) | 21.90 s Done.
+
[Task 19/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 19/25] Current/Best: 21.09/ 21.09 GFLOPS | Progress: (4/20) | 5.47 s
[Task 19/25] Current/Best: 21.10/ 21.10 GFLOPS | Progress: (8/20) | 9.26 s
[Task 19/25] Current/Best: 18.07/ 21.10 GFLOPS | Progress: (12/20) | 12.50 s
[Task 19/25] Current/Best: 6.11/ 21.10 GFLOPS | Progress: (16/20) | 16.33 s
[Task 19/25] Current/Best: 20.04/ 21.10 GFLOPS | Progress: (20/20) | 19.24 s Done.
+
[Task 20/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 20/25] Current/Best: 18.49/ 20.04 GFLOPS | Progress: (4/20) | 4.58 s
[Task 20/25] Current/Best: 16.62/ 20.04 GFLOPS | Progress: (8/20) | 7.52 s
[Task 20/25] Current/Best: 6.20/ 20.04 GFLOPS | Progress: (12/20) | 9.30 s
[Task 20/25] Current/Best: 8.57/ 20.04 GFLOPS | Progress: (16/20) | 11.33 s Done.
+
[Task 20/25] Current/Best: 11.29/ 20.06 GFLOPS | Progress: (20/20) | 14.35 s
[Task 21/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 21/25] Current/Best: 9.45/ 18.85 GFLOPS | Progress: (4/20) | 3.79 s
[Task 21/25] Current/Best: 22.47/ 22.47 GFLOPS | Progress: (8/20) | 5.72 s
[Task 21/25] Current/Best: 10.95/ 22.47 GFLOPS | Progress: (12/20) | 8.91 s
[Task 21/25] Current/Best: 6.93/ 22.47 GFLOPS | Progress: (16/20) | 11.23 s
[Task 21/25] Current/Best: 17.10/ 22.47 GFLOPS | Progress: (20/20) | 12.82 s
[Task 22/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 22/25] Current/Best: 2.70/ 15.99 GFLOPS | Progress: (4/20) | 5.03 s
[Task 22/25] Current/Best: 12.40/ 15.99 GFLOPS | Progress: (8/20) | 7.28 s
[Task 22/25] Current/Best: 16.05/ 16.05 GFLOPS | Progress: (12/20) | 9.60 s
[Task 22/25] Current/Best: 16.57/ 18.99 GFLOPS | Progress: (16/20)
| 11.17 s
[Task 22/25] Current/Best: 11.23/ 18.99 GFLOPS | Progress: (20/20) | 13.02 s Done.
+
[Task 23/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 23/25] Current/Best: 8.38/ 18.13 GFLOPS | Progress: (4/20) | 5.09 s
[Task 23/25] Current/Best: 12.32/ 18.98 GFLOPS | Progress: (8/20) | 8.03 s
[Task 23/25] Current/Best: 11.99/ 18.98 GFLOPS | Progress: (12/20) | 10.97 s
[Task 23/25] Current/Best: 22.52/ 22.52 GFLOPS | Progress: (16/20) | 13.97 s
[Task 23/25] Current/Best: 14.05/ 22.52 GFLOPS | Progress: (20/20) | 16.69 s Done.
+
[Task 24/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 24/25] Current/Best: 3.61/ 9.73 GFLOPS | Progress: (4/20) | 12.76 s
[Task 24/25] Current/Best: 0.54/ 9.73 GFLOPS | Progress: (8/20) | 20.64 s
[Task 24/25] Current/Best: 3.01/ 10.69 GFLOPS | Progress: (12/20) | 22.00 s
[Task 24/25] Current/Best: 6.27/ 10.69 GFLOPS | Progress: (16/20) | 32.92 s Done.
+
[Task 24/25] Current/Best: 3.29/ 10.69 GFLOPS | Progress: (20/20) | 44.62 s
[Task 25/25] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/20) | 0.00 s
[Task 25/25] Current/Best: 5.42/ 6.08 GFLOPS | Progress: (4/20) | 13.52 s
[Task 25/25] Current/Best: 6.15/ 9.80 GFLOPS | Progress: (8/20) | 24.48 s
[Task 25/25] Current/Best: 7.84/ 9.80 GFLOPS | Progress: (12/20) | 26.91 s
[Task 25/25] Current/Best: 9.10/ 9.80 GFLOPS | Progress: (16/20) | 37.87 s
[Task 25/25] Current/Best: 7.96/ 9.80 GFLOPS | Progress: (20/20) | 40.04 s
@@ -721,8 +721,8 @@ improvement in comparing the optimized model to the unoptimized model.
.. code-block:: none
- optimized: {'mean': 404.5399479699995, 'median': 404.1449735499896, 'std': 3.42485888438619}
- unoptimized: {'mean': 512.2079894000011, 'median': 512.4371018000033, 'std': 2.4379111328350107}
+ optimized: {'mean': 413.88311280999915, 'median': 413.60596074999876, 'std': 1.6160087440403037}
+ unoptimized: {'mean': 477.48544495999687, 'median': 477.1313110499932, 'std': 1.163654270518335}
@@ -745,7 +745,7 @@ profiling/benchmarking.
.. rst-class:: sphx-glr-timing
- **Total running time of the script:** ( 12 minutes 11.809 seconds)
+ **Total running time of the script:** ( 11 minutes 26.752 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 83138ebd3e..535d194b18 100644
--- a/docs/_sources/tutorial/cross_compilation_and_rpc.rst.txt
+++ b/docs/_sources/tutorial/cross_compilation_and_rpc.rst.txt
@@ -274,7 +274,7 @@ device and returns the measured cost. Network overhead is excluded.
.. code-block:: none
- 1.28e-07 secs/op
+ 1.308e-07 secs/op
diff --git a/docs/_sources/tutorial/intro_topi.rst.txt b/docs/_sources/tutorial/intro_topi.rst.txt
index 0209fec57a..e9d91e8c1d 100644
--- a/docs/_sources/tutorial/intro_topi.rst.txt
+++ b/docs/_sources/tutorial/intro_topi.rst.txt
@@ -225,10 +225,8 @@ we can schedule the following series of operations ending with :code:`topi.sum`
@I.ir_module
class Module:
@T.prim_func
- def main(a: T.handle, b: T.handle):
+ def main(a: T.Buffer((100, 10, 10), "float32"), b: T.Buffer((10, 10), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- a_1 = T.match_buffer(a, (100, 10, 10))
- b_1 = T.match_buffer(b, (10, 10))
T_divide_red = T.allocate([1], "float32", "global")
threadIdx_x = T.env_thread("threadIdx.x")
T.launch_thread(threadIdx_x, 1024)
@@ -238,9 +236,9 @@ we can schedule the following series of operations ending with :code:`topi.sum`
T_divide_red_rf_1[0] = T.float32(0)
for k0_k1_fused_k2_fused_outer in range(10):
if T.likely(k0_k1_fused_k2_fused_outer * 64 + threadIdx_x // 16 < 625 and k0_k1_fused_k2_fused_outer * 64 + threadIdx_x // 16 < 625 and k0_k1_fused_k2_fused_outer * 64 + threadIdx_x // 16 < 625):
- a_2 = T.buffer_decl((10000,), data=a_1.data)
- b_2 = T.buffer_decl((100,), data=b_1.data)
- T_divide_red_rf_1[0] = T_divide_red_rf_1[0] + (a_2[k0_k1_fused_k2_fused_outer * 1024 + threadIdx_x] + b_2[(k0_k1_fused_k2_fused_outer * 12 + threadIdx_x // 2) % 50 // 5 * 10 + (k0_k1_fused_k2_fused_outer * 4 + threadIdx_x) % 10] + a_2[k0_k1_fused_k2_fused_outer * 1024 + threadIdx_x] * b_2[(k0_k1_fused_k2_fused_outer * 12 + threadIdx_x // 2) % 50 // 5 * 10 + (k0_k1_fused_k2_fused_outer * 4 + threadIdx_x) % 10]) * T.float32(0.5)
+ a_1 = T.buffer_decl((10000,), data=a.data)
+ b_1 = T.buffer_decl((100,), data=b.data)
+ T_divide_red_rf_1[0] = T_divide_red_rf_1[0] + (a_1[k0_k1_fused_k2_fused_outer * 1024 + threadIdx_x] + b_1[(k0_k1_fused_k2_fused_outer * 12 + threadIdx_x // 2) % 50 // 5 * 10 + (k0_k1_fused_k2_fused_outer * 4 + threadIdx_x) % 10] + a_1[k0_k1_fused_k2_fused_outer * 1024 + threadIdx_x] * b_1[(k0_k1_fused_k2_fused_outer * 12 + threadIdx_x // 2) % 50 // 5 * 10 + (k0_k1_fused_k2_fused_outer * 4 + threadIdx_x) % 10]) * T.float32(0.5)
reduce_temp0_1 = T.buffer_decl((1,), data=reduce_temp0, scope="local")
with T.attr(T.comm_reducer(lambda x, y: x + y, [T.float32(0)]), "reduce_scope", T.reinterpret("handle", T.uint64(0))):
T.tvm_thread_allreduce(T.uint32(1), T_divide_red_rf_1[0], True, reduce_temp0_1[0], threadIdx_x)
@@ -270,7 +268,7 @@ As you can see, scheduled stages of computation have been accumulated and we can
.. code-block:: none
- [stage(a, placeholder(a, 0x271001e0)), stage(b, placeholder(b, 0x270c0f10)), stage(T_add, compute(T_add, body=[a[ax0, ax1, ax2] + b[ax1, ax2]], axis=[T.iter_var(ax0, T.Range(0, 100), "DataPar", ""), T.iter_var(ax1, T.Range(0, 10), "DataPar", ""), T.iter_var(ax2, T.Range(0, 10), "DataPar", "")], reduce_axis=[], tag=broadcast, attrs={})), stage(T_multiply, compute(T_multiply, body=[a[ax0, ax1, ax2] * b[ax1, ax2]], axis=[T.iter_var(ax0, T.Range(0, 100), "DataPar", ""), T.iter_var(ax1, T [...]
+ [stage(a, placeholder(a, 0x22bcd1f0)), stage(b, placeholder(b, 0x21d8d400)), stage(T_add, compute(T_add, body=[a[ax0, ax1, ax2] + b[ax1, ax2]], axis=[T.iter_var(ax0, T.Range(0, 100), "DataPar", ""), T.iter_var(ax1, T.Range(0, 10), "DataPar", ""), T.iter_var(ax2, T.Range(0, 10), "DataPar", "")], reduce_axis=[], tag=broadcast, attrs={})), stage(T_multiply, compute(T_multiply, body=[a[ax0, ax1, ax2] * b[ax1, ax2]], axis=[T.iter_var(ax0, T.Range(0, 100), "DataPar", ""), T.iter_var(ax1, T [...]
@@ -328,9 +326,8 @@ TOPI also provides common neural nets operations such as _softmax_ with optimize
@I.ir_module
class Module:
@T.prim_func
- def main(tarray: T.handle):
+ def main(tarray: T.Buffer((512, 512), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- tarray_1 = T.match_buffer(tarray, (512, 512))
T_softmax_norm = T.allocate([65536], "float32x4", "global")
blockIdx_x = T.env_thread("blockIdx.x")
T.launch_thread(blockIdx_x, 512)
@@ -344,16 +341,16 @@ TOPI also provides common neural nets operations such as _softmax_ with optimize
with T.launch_thread(threadIdx_x, 32):
normal_reduce_temp0_2 = T.buffer_decl((1,), data=normal_reduce_temp0, scope="local")
normal_reduce_temp0_2[0] = T.float32(-3.4028234663852886e+38)
- tarray_2 = T.buffer_decl((262144,), data=tarray_1.data)
+ tarray_1 = T.buffer_decl((262144,), data=tarray.data)
for k_inner in range(16):
- normal_reduce_temp0_2[0] = T.max(normal_reduce_temp0_2[0], tarray_2[blockIdx_x * 512 + threadIdx_x * 16 + k_inner])
+ normal_reduce_temp0_2[0] = T.max(normal_reduce_temp0_2[0], tarray_1[blockIdx_x * 512 + threadIdx_x * 16 + k_inner])
with T.attr(T.comm_reducer(lambda x, y: T.max(x, y), [T.float32(-3.4028234663852886e+38)]), "reduce_scope", T.reinterpret("handle", T.uint64(0))):
reduce_temp0_2 = T.buffer_decl((1,), data=reduce_temp0, scope="local")
T.tvm_thread_allreduce(T.uint32(1), normal_reduce_temp0_2[0], True, reduce_temp0_2[0], threadIdx_x)
for i1_inner_outer in range(4):
cse_var_1: T.int32 = i1_inner_outer * 4
reduce_temp0_2 = T.buffer_decl((1,), data=reduce_temp0, scope="local", align=4)
- T_softmax_exp_1[threadIdx_x * 16 + cse_var_1:threadIdx_x * 16 + cse_var_1 + 4] = T.exp(tarray_2[blockIdx_x * 512 + threadIdx_x * 16 + cse_var_1:blockIdx_x * 512 + threadIdx_x * 16 + cse_var_1 + 4] - T.Broadcast(reduce_temp0_2[0], 4))
+ T_softmax_exp_1[threadIdx_x * 16 + cse_var_1:threadIdx_x * 16 + cse_var_1 + 4] = T.exp(tarray_1[blockIdx_x * 512 + threadIdx_x * 16 + cse_var_1:blockIdx_x * 512 + threadIdx_x * 16 + cse_var_1 + 4] - T.Broadcast(reduce_temp0_2[0], 4))
T.launch_thread(threadIdx_x, 32)
normal_reduce_temp0_2 = T.buffer_decl((1,), data=normal_reduce_temp0_1, scope="local")
normal_reduce_temp0_2[0] = T.float32(0)
@@ -409,10 +406,8 @@ We can fuse :code:`topi.nn.conv2d` and :code:`topi.nn.relu` together.
@I.ir_module
class Module:
@T.prim_func
- def main(placeholder: T.handle, placeholder_1: T.handle):
+ def main(placeholder: T.Buffer((1, 3, 224, 224), "float32"), placeholder_1: T.Buffer((10, 3, 5, 5), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- placeholder_2 = T.match_buffer(placeholder, (1, 3, 224, 224))
- placeholder_3 = T.match_buffer(placeholder_1, (10, 3, 5, 5))
compute = T.allocate([501760], "float32", "global")
blockIdx_z = T.env_thread("blockIdx.z")
T.launch_thread(blockIdx_z, 5)
@@ -449,27 +444,27 @@ We can fuse :code:`topi.nn.conv2d` and :code:`topi.nn.relu` together.
threadIdx_y_1 = T.env_thread("threadIdx.y")
threadIdx_x_1 = T.env_thread("threadIdx.x")
pad_temp_shared_1 = T.buffer_decl((112,), data=pad_temp_shared, scope="shared")
- placeholder_4 = T.buffer_decl((150528,), data=placeholder_2.data)
+ placeholder_2 = T.buffer_decl((150528,), data=placeholder.data)
with T.launch_thread(threadIdx_z_1, 1):
T.launch_thread(threadIdx_y_1, 1)
T.launch_thread(threadIdx_x_1, 16)
- pad_temp_shared_1[threadIdx_x_1 * 7] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226 and 1 <= blockIdx_x * 56 + threadIdx_x_1 * 7 // 2, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 450], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 1] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226 and 1 <= blockIdx_x * 56 + (threadIdx_x_1 * 7 + 1) // 2, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 449], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 2] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 448], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 3] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 447], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 4] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 446], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 5] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 445], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 6] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 444], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226 and 1 <= blockIdx_x * 56 + threadIdx_x_1 * 7 // 2, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 450], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 1] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226 and 1 <= blockIdx_x * 56 + (threadIdx_x_1 * 7 + 1) // 2, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 449], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 2] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 448], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 3] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 447], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 4] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 446], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 5] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 445], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 6] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 444], T.float32(0))
threadIdx_z_2 = T.env_thread("threadIdx.z")
threadIdx_y_2 = T.env_thread("threadIdx.y")
threadIdx_x_2 = T.env_thread("threadIdx.x")
placeholder_shared_1 = T.buffer_decl((2,), data=placeholder_shared, scope="shared", align=8)
- placeholder_5 = T.buffer_decl((750,), data=placeholder_3.data)
+ placeholder_3 = T.buffer_decl((750,), data=placeholder_1.data)
with T.launch_thread(threadIdx_z_2, 1):
T.launch_thread(threadIdx_y_2, 1)
T.launch_thread(threadIdx_x_2, 16)
if T.likely(threadIdx_x_2 < 2):
- placeholder_shared_1[threadIdx_x_2] = placeholder_5[blockIdx_z * 150 + threadIdx_x_2 * 75 + rc_outer * 25 + ry_outer * 5]
+ placeholder_shared_1[threadIdx_x_2] = placeholder_3[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]
conv2d_nchw_1[4] = conv2d_nchw_1[4] + pad_temp_shared_1[threadIdx_x + 32] * placeholder_shared_1[0]
@@ -487,18 +482,18 @@ We can fuse :code:`topi.nn.conv2d` and :code:`topi.nn.relu` together.
with T.launch_thread(threadIdx_z_1, 1):
T.launch_thread(threadIdx_y_1, 1)
T.launch_thread(threadIdx_x_1, 16)
- pad_temp_shared_1[threadIdx_x_1 * 7] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226 and 1 <= blockIdx_x * 56 + (threadIdx_x_1 * 7 + 1) // 2, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 449], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 1] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 448], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 2] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 447], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 3] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 446], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 4] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 445], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 5] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 444], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 6] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 443], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226 and 1 <= blockIdx_x * 56 + (threadIdx_x_1 * 7 + 1) // 2, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 449], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 1] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 448], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 2] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 447], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 3] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 446], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 4] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 445], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 5] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 444], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 6] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 443], T.float32(0))
with T.launch_thread(threadIdx_z_2, 1):
T.launch_thread(threadIdx_y_2, 1)
T.launch_thread(threadIdx_x_2, 16)
if T.likely(threadIdx_x_2 < 2):
- placeholder_shared_1[threadIdx_x_2] = placeholder_5[blockIdx_z * 150 + threadIdx_x_2 * 75 + rc_outer * 25 + ry_outer * 5 + 1]
+ placeholder_shared_1[threadIdx_x_2] = placeholder_3[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]
conv2d_nchw_1[4] = conv2d_nchw_1[4] + pad_temp_shared_1[threadIdx_x + 32] * placeholder_shared_1[0]
@@ -516,18 +511,18 @@ We can fuse :code:`topi.nn.conv2d` and :code:`topi.nn.relu` together.
with T.launch_thread(threadIdx_z_1, 1):
T.launch_thread(threadIdx_y_1, 1)
T.launch_thread(threadIdx_x_1, 16)
- pad_temp_shared_1[threadIdx_x_1 * 7] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 448], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 1] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 447], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 2] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 446], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 3] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 445], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 4] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 444], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 5] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 443], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 6] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 442], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 448], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 1] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 447], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 2] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 446], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 3] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 445], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 4] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 444], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 5] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 443], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 6] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 442], T.float32(0))
with T.launch_thread(threadIdx_z_2, 1):
T.launch_thread(threadIdx_y_2, 1)
T.launch_thread(threadIdx_x_2, 16)
if T.likely(threadIdx_x_2 < 2):
- placeholder_shared_1[threadIdx_x_2] = placeholder_5[blockIdx_z * 150 + threadIdx_x_2 * 75 + rc_outer * 25 + ry_outer * 5 + 2]
+ placeholder_shared_1[threadIdx_x_2] = placeholder_3[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]
conv2d_nchw_1[4] = conv2d_nchw_1[4] + pad_temp_shared_1[threadIdx_x + 32] * placeholder_shared_1[0]
@@ -545,18 +540,18 @@ We can fuse :code:`topi.nn.conv2d` and :code:`topi.nn.relu` together.
with T.launch_thread(threadIdx_z_1, 1):
T.launch_thread(threadIdx_y_1, 1)
T.launch_thread(threadIdx_x_1, 16)
- pad_temp_shared_1[threadIdx_x_1 * 7] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 447], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 1] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 446], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 2] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 445], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 3] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 444], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 4] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 443], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 5] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 442], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 6] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226 and blockIdx_x * 56 + (threadIdx_x_1 * 7 + 9) // 2 < 113, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 441], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 447], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 1] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 446], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 2] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 445], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 3] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 444], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 4] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 443], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 5] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 442], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 6] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226 and blockIdx_x * 56 + (threadIdx_x_1 * 7 + 9) // 2 < 113, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 441], T.float32(0))
with T.launch_thread(threadIdx_z_2, 1):
T.launch_thread(threadIdx_y_2, 1)
T.launch_thread(threadIdx_x_2, 16)
if T.likely(threadIdx_x_2 < 2):
- placeholder_shared_1[threadIdx_x_2] = placeholder_5[blockIdx_z * 150 + threadIdx_x_2 * 75 + rc_outer * 25 + ry_outer * 5 + 3]
+ placeholder_shared_1[threadIdx_x_2] = placeholder_3[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]
conv2d_nchw_1[4] = conv2d_nchw_1[4] + pad_temp_shared_1[threadIdx_x + 32] * placeholder_shared_1[0]
@@ -574,18 +569,18 @@ We can fuse :code:`topi.nn.conv2d` and :code:`topi.nn.relu` together.
with T.launch_thread(threadIdx_z_1, 1):
T.launch_thread(threadIdx_y_1, 1)
T.launch_thread(threadIdx_x_1, 16)
- pad_temp_shared_1[threadIdx_x_1 * 7] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 446], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 1] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 445], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 2] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 444], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 3] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 443], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 4] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 442], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 5] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226 and blockIdx_x * 56 + (threadIdx_x_1 * 7 + 9) // 2 < 113, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 441], T.float32(0))
- pad_temp_shared_1[threadIdx_x_1 * 7 + 6] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226 and blockIdx_x * 56 + threadIdx_x_1 * 7 // 2 < 108, placeholder_4[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 440], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 446], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 1] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 445], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 2] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 444], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 3] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 443], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 4] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 442], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 5] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226 and blockIdx_x * 56 + (threadIdx_x_1 * 7 + 9) // 2 < 113, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 441], T.float32(0))
+ pad_temp_shared_1[threadIdx_x_1 * 7 + 6] = T.if_then_else(2 <= blockIdx_y + ry_outer and blockIdx_y + ry_outer < 226 and blockIdx_x * 56 + threadIdx_x_1 * 7 // 2 < 108, placeholder_2[rc_outer * 50176 + blockIdx_y * 224 + ry_outer * 224 + blockIdx_x * 112 + threadIdx_x_1 * 7 - 440], T.float32(0))
with T.launch_thread(threadIdx_z_2, 1):
T.launch_thread(threadIdx_y_2, 1)
T.launch_thread(threadIdx_x_2, 16)
if T.likely(threadIdx_x_2 < 2):
- placeholder_shared_1[threadIdx_x_2] = placeholder_5[blockIdx_z * 150 + threadIdx_x_2 * 75 + rc_outer * 25 + ry_outer * 5 + 4]
+ placeholder_shared_1[threadIdx_x_2] = placeholder_3[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]
conv2d_nchw_1[4] = conv2d_nchw_1[4] + pad_temp_shared_1[threadIdx_x + 32] * 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 1f75f9fae2..dbb9de6f38 100644
--- a/docs/_sources/tutorial/sg_execution_times.rst.txt
+++ b/docs/_sources/tutorial/sg_execution_times.rst.txt
@@ -5,31 +5,31 @@
Computation times
=================
-**15:35.251** total execution time for **tutorial** files:
+**14:55.935** total execution time for **tutorial** files:
+------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_tutorial_autotvm_relay_x86.py` (``autotvm_relay_x86.py``) | 12:11.809 | 0.0 MB |
+| :ref:`sphx_glr_tutorial_autotvm_relay_x86.py` (``autotvm_relay_x86.py``) | 11:26.752 | 0.0 MB |
+------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_tutorial_auto_scheduler_matmul_x86.py` (``auto_scheduler_matmul_x86.py``) | 01:20.459 | 0.0 MB |
+| :ref:`sphx_glr_tutorial_auto_scheduler_matmul_x86.py` (``auto_scheduler_matmul_x86.py``) | 01:30.420 | 0.0 MB |
+------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_tutorial_tensor_expr_get_started.py` (``tensor_expr_get_started.py``) | 01:00.967 | 0.0 MB |
+| :ref:`sphx_glr_tutorial_tensor_expr_get_started.py` (``tensor_expr_get_started.py``) | 00:58.900 | 0.0 MB |
+------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_tutorial_relay_quick_start.py` (``relay_quick_start.py``) | 00:35.443 | 0.0 MB |
+| :ref:`sphx_glr_tutorial_relay_quick_start.py` (``relay_quick_start.py``) | 00:34.825 | 0.0 MB |
+------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_tutorial_autotvm_matmul_x86.py` (``autotvm_matmul_x86.py``) | 00:24.450 | 0.0 MB |
+| :ref:`sphx_glr_tutorial_autotvm_matmul_x86.py` (``autotvm_matmul_x86.py``) | 00:22.680 | 0.0 MB |
+------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_tutorial_tensor_ir_blitz_course.py` (``tensor_ir_blitz_course.py``) | 00:01.126 | 0.0 MB |
+| :ref:`sphx_glr_tutorial_tensor_ir_blitz_course.py` (``tensor_ir_blitz_course.py``) | 00:01.379 | 0.0 MB |
+------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_tutorial_intro_topi.py` (``intro_topi.py``) | 00:00.828 | 0.0 MB |
+| :ref:`sphx_glr_tutorial_intro_topi.py` (``intro_topi.py``) | 00:00.819 | 0.0 MB |
+------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_tutorial_cross_compilation_and_rpc.py` (``cross_compilation_and_rpc.py``) | 00:00.169 | 0.0 MB |
+| :ref:`sphx_glr_tutorial_cross_compilation_and_rpc.py` (``cross_compilation_and_rpc.py``) | 00:00.159 | 0.0 MB |
+------------------------------------------------------------------------------------------+-----------+--------+
| :ref:`sphx_glr_tutorial_uma.py` (``uma.py``) | 00:00.000 | 0.0 MB |
+------------------------------------------------------------------------------------------+-----------+--------+
-| :ref:`sphx_glr_tutorial_tvmc_python.py` (``tvmc_python.py``) | 00:00.000 | 0.0 MB |
-+------------------------------------------------------------------------------------------+-----------+--------+
| :ref:`sphx_glr_tutorial_introduction.py` (``introduction.py``) | 00:00.000 | 0.0 MB |
+------------------------------------------------------------------------------------------+-----------+--------+
+| :ref:`sphx_glr_tutorial_tvmc_python.py` (``tvmc_python.py``) | 00:00.000 | 0.0 MB |
++------------------------------------------------------------------------------------------+-----------+--------+
| :ref:`sphx_glr_tutorial_tvmc_command_line_driver.py` (``tvmc_command_line_driver.py``) | 00:00.000 | 0.0 MB |
+------------------------------------------------------------------------------------------+-----------+--------+
| :ref:`sphx_glr_tutorial_install.py` (``install.py``) | 00:00.000 | 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 1bd1f80608..7235d27187 100644
--- a/docs/_sources/tutorial/tensor_expr_get_started.rst.txt
+++ b/docs/_sources/tutorial/tensor_expr_get_started.rst.txt
@@ -389,7 +389,7 @@ compile and run this new schedule with the parallel operation applied:
.. code-block:: none
- parallel: 0.000007
+ parallel: 0.000011
@@ -498,10 +498,10 @@ We can now compare the different schedules
.. code-block:: none
Operator Timing Performance
- numpy 6.711370001539763e-06 1.0
- naive 6.6762e-06 0.9947596390108585
- parallel 6.9535000000000004e-06 1.036077581537702
- vector 2.4626500000000002e-05 3.6693700383602814
+ numpy 7.256909998432093e-06 1.0
+ naive 6.6563999999999995e-06 0.9172499040828896
+ parallel 1.126e-05 1.551624589864392
+ vector 2.45527e-05 3.3833546241175356
@@ -922,7 +922,7 @@ matrix multiplication.
.. code-block:: none
- Numpy running time: 0.018786
+ Numpy running time: 0.017994
@@ -980,7 +980,7 @@ optimizations.
.. code-block:: none
- none: 3.399821
+ none: 3.258496
@@ -1010,20 +1010,17 @@ nested loops over the indices of the A and B matrices.
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 1024), "float32"), B: T.Buffer((1024, 1024), "float32"), C: T.Buffer((1024, 1024), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 1024))
- B_1 = T.match_buffer(B, (1024, 1024))
- C_1 = T.match_buffer(C, (1024, 1024))
for x, y in T.grid(1024, 1024):
- C_2 = T.buffer_decl((1048576,), data=C_1.data)
- C_2[x * 1024 + y] = T.float32(0)
+ C_1 = T.buffer_decl((1048576,), data=C.data)
+ C_1[x * 1024 + y] = T.float32(0)
for k in range(1024):
cse_var_2: T.int32 = x * 1024
cse_var_1: T.int32 = cse_var_2 + y
- A_2 = T.buffer_decl((1048576,), data=A_1.data)
- B_2 = T.buffer_decl((1048576,), data=B_1.data)
- C_2[cse_var_1] = C_2[cse_var_1] + A_2[cse_var_2 + k] * B_2[k * 1024 + y]
+ A_1 = T.buffer_decl((1048576,), data=A.data)
+ B_1 = T.buffer_decl((1048576,), data=B.data)
+ C_1[cse_var_1] = C_1[cse_var_1] + A_1[cse_var_2 + k] * B_1[k * 1024 + y]
@@ -1080,7 +1077,7 @@ schedule.
.. code-block:: none
- blocking: 0.300759
+ blocking: 0.297530
@@ -1109,22 +1106,19 @@ internal representation and compare it to the original:
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 1024), "float32"), B: T.Buffer((1024, 1024), "float32"), C: T.Buffer((1024, 1024), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 1024))
- B_1 = T.match_buffer(B, (1024, 1024))
- C_1 = T.match_buffer(C, (1024, 1024))
for x_outer, y_outer in T.grid(32, 32):
- C_2 = T.buffer_decl((1048576,), data=C_1.data)
+ C_1 = T.buffer_decl((1048576,), data=C.data)
for x_inner_init, y_inner_init in T.grid(32, 32):
- C_2[x_outer * 32768 + x_inner_init * 1024 + y_outer * 32 + y_inner_init] = T.float32(0)
+ C_1[x_outer * 32768 + x_inner_init * 1024 + y_outer * 32 + y_inner_init] = T.float32(0)
for k_outer, k_inner, x_inner, y_inner in T.grid(256, 4, 32, 32):
cse_var_3: T.int32 = y_outer * 32
cse_var_2: T.int32 = x_outer * 32768 + x_inner * 1024
cse_var_1: T.int32 = cse_var_2 + cse_var_3 + y_inner
- A_2 = T.buffer_decl((1048576,), data=A_1.data)
- B_2 = T.buffer_decl((1048576,), data=B_1.data)
- C_2[cse_var_1] = C_2[cse_var_1] + A_2[cse_var_2 + k_outer * 4 + k_inner] * B_2[k_outer * 4096 + k_inner * 1024 + cse_var_3 + y_inner]
+ A_1 = T.buffer_decl((1048576,), data=A.data)
+ B_1 = T.buffer_decl((1048576,), data=B.data)
+ C_1[cse_var_1] = C_1[cse_var_1] + A_1[cse_var_2 + k_outer * 4 + k_inner] * B_1[k_outer * 4096 + k_inner * 1024 + cse_var_3 + y_inner]
@@ -1164,26 +1158,23 @@ already cache friendly from our previous optimizations.
.. code-block:: none
- vectorization: 0.344421
+ vectorization: 0.330118
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 1024), "float32"), B: T.Buffer((1024, 1024), "float32"), C: T.Buffer((1024, 1024), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 1024))
- B_1 = T.match_buffer(B, (1024, 1024))
- C_1 = T.match_buffer(C, (1024, 1024))
for x_outer, y_outer in T.grid(32, 32):
- C_2 = T.buffer_decl((1048576,), data=C_1.data)
+ C_1 = T.buffer_decl((1048576,), data=C.data)
for x_inner_init in range(32):
- C_2[x_outer * 32768 + x_inner_init * 1024 + y_outer * 32:x_outer * 32768 + x_inner_init * 1024 + y_outer * 32 + 32] = T.Broadcast(T.float32(0), 32)
+ C_1[x_outer * 32768 + x_inner_init * 1024 + y_outer * 32:x_outer * 32768 + x_inner_init * 1024 + y_outer * 32 + 32] = T.Broadcast(T.float32(0), 32)
for k_outer, k_inner, x_inner in T.grid(256, 4, 32):
cse_var_3: T.int32 = y_outer * 32
cse_var_2: T.int32 = x_outer * 32768 + x_inner * 1024
cse_var_1: T.int32 = cse_var_2 + cse_var_3
- A_2 = T.buffer_decl((1048576,), data=A_1.data)
- B_2 = T.buffer_decl((1048576,), data=B_1.data)
- C_2[cse_var_1:cse_var_1 + 32] = C_2[cse_var_1:cse_var_1 + 32] + T.Broadcast(A_2[cse_var_2 + k_outer * 4 + k_inner], 32) * B_2[k_outer * 4096 + k_inner * 1024 + cse_var_3:k_outer * 4096 + k_inner * 1024 + cse_var_3 + 32]
+ A_1 = T.buffer_decl((1048576,), data=A.data)
+ B_1 = T.buffer_decl((1048576,), data=B.data)
+ C_1[cse_var_1:cse_var_1 + 32] = C_1[cse_var_1:cse_var_1 + 32] + T.Broadcast(A_1[cse_var_2 + k_outer * 4 + k_inner], 32) * B_1[k_outer * 4096 + k_inner * 1024 + cse_var_3:k_outer * 4096 + k_inner * 1024 + cse_var_3 + 32]
@@ -1230,26 +1221,23 @@ more cache friendly.
.. code-block:: none
- loop permutation: 0.115957
+ loop permutation: 0.119767
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 1024), "float32"), B: T.Buffer((1024, 1024), "float32"), C: T.Buffer((1024, 1024), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 1024))
- B_1 = T.match_buffer(B, (1024, 1024))
- C_1 = T.match_buffer(C, (1024, 1024))
for x_outer, y_outer in T.grid(32, 32):
- C_2 = T.buffer_decl((1048576,), data=C_1.data)
+ C_1 = T.buffer_decl((1048576,), data=C.data)
for x_inner_init in range(32):
- C_2[x_outer * 32768 + x_inner_init * 1024 + y_outer * 32:x_outer * 32768 + x_inner_init * 1024 + y_outer * 32 + 32] = T.Broadcast(T.float32(0), 32)
+ C_1[x_outer * 32768 + x_inner_init * 1024 + y_outer * 32:x_outer * 32768 + x_inner_init * 1024 + y_outer * 32 + 32] = T.Broadcast(T.float32(0), 32)
for k_outer, x_inner, k_inner in T.grid(256, 32, 4):
cse_var_3: T.int32 = y_outer * 32
cse_var_2: T.int32 = x_outer * 32768 + x_inner * 1024
cse_var_1: T.int32 = cse_var_2 + cse_var_3
- A_2 = T.buffer_decl((1048576,), data=A_1.data)
- B_2 = T.buffer_decl((1048576,), data=B_1.data)
- C_2[cse_var_1:cse_var_1 + 32] = C_2[cse_var_1:cse_var_1 + 32] + T.Broadcast(A_2[cse_var_2 + k_outer * 4 + k_inner], 32) * B_2[k_outer * 4096 + k_inner * 1024 + cse_var_3:k_outer * 4096 + k_inner * 1024 + cse_var_3 + 32]
+ A_1 = T.buffer_decl((1048576,), data=A.data)
+ B_1 = T.buffer_decl((1048576,), data=B.data)
+ C_1[cse_var_1:cse_var_1 + 32] = C_1[cse_var_1:cse_var_1 + 32] + T.Broadcast(A_1[cse_var_2 + k_outer * 4 + k_inner], 32) * B_1[k_outer * 4096 + k_inner * 1024 + cse_var_3:k_outer * 4096 + k_inner * 1024 + cse_var_3 + 32]
@@ -1321,31 +1309,28 @@ optimized schedule.
.. code-block:: none
- array packing: 0.108478
+ array packing: 0.109910
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 1024), "float32"), B: T.Buffer((1024, 1024), "float32"), C: T.Buffer((1024, 1024), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 1024))
- B_1 = T.match_buffer(B, (1024, 1024))
- C_1 = T.match_buffer(C, (1024, 1024))
packedB = T.allocate([32768], "float32x32", "global")
packedB_1 = T.buffer_decl((32768,), "float32x32", data=packedB)
for x in T.parallel(32):
for y in range(1024):
- B_2 = T.buffer_decl((1048576,), data=B_1.data)
- packedB_1[x * 1024 + y] = B_2[y * 1024 + x * 32:y * 1024 + x * 32 + 32]
+ B_1 = T.buffer_decl((1048576,), data=B.data)
+ packedB_1[x * 1024 + y] = B_1[y * 1024 + x * 32:y * 1024 + x * 32 + 32]
for x_outer, y_outer in T.grid(32, 32):
- C_2 = T.buffer_decl((1048576,), data=C_1.data)
+ C_1 = T.buffer_decl((1048576,), data=C.data)
for x_inner_init in range(32):
- C_2[x_outer * 32768 + x_inner_init * 1024 + y_outer * 32:x_outer * 32768 + x_inner_init * 1024 + y_outer * 32 + 32] = T.Broadcast(T.float32(0), 32)
+ C_1[x_outer * 32768 + x_inner_init * 1024 + y_outer * 32:x_outer * 32768 + x_inner_init * 1024 + y_outer * 32 + 32] = T.Broadcast(T.float32(0), 32)
for k_outer, x_inner, k_inner in T.grid(256, 32, 4):
cse_var_3: T.int32 = x_outer * 32768 + x_inner * 1024
cse_var_2: T.int32 = k_outer * 4
cse_var_1: T.int32 = cse_var_3 + y_outer * 32
- A_2 = T.buffer_decl((1048576,), data=A_1.data)
- C_2[cse_var_1:cse_var_1 + 32] = C_2[cse_var_1:cse_var_1 + 32] + T.Broadcast(A_2[cse_var_3 + cse_var_2 + k_inner], 32) * packedB_1[y_outer * 1024 + cse_var_2 + k_inner]
+ A_1 = T.buffer_decl((1048576,), data=A.data)
+ C_1[cse_var_1:cse_var_1 + 32] = C_1[cse_var_1:cse_var_1 + 32] + T.Broadcast(A_1[cse_var_3 + cse_var_2 + k_inner], 32) * packedB_1[y_outer * 1024 + cse_var_2 + k_inner]
@@ -1404,22 +1389,19 @@ to `C` when all the block results are ready.
.. code-block:: none
- block caching: 0.110815
+ block caching: 0.110722
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 1024), "float32"), B: T.Buffer((1024, 1024), "float32"), C: T.Buffer((1024, 1024), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 1024))
- B_1 = T.match_buffer(B, (1024, 1024))
- C_1 = T.match_buffer(C, (1024, 1024))
packedB = T.allocate([32768], "float32x32", "global")
C_global = T.allocate([1024], "float32", "global")
packedB_1 = T.buffer_decl((32768,), "float32x32", data=packedB)
for x in T.parallel(32):
for y in range(1024):
- B_2 = T.buffer_decl((1048576,), data=B_1.data)
- packedB_1[x * 1024 + y] = B_2[y * 1024 + x * 32:y * 1024 + x * 32 + 32]
+ B_1 = T.buffer_decl((1048576,), data=B.data)
+ packedB_1[x * 1024 + y] = B_1[y * 1024 + x * 32:y * 1024 + x * 32 + 32]
for x_outer, y_outer in T.grid(32, 32):
C_global_1 = T.buffer_decl((1024,), data=C_global)
for x_c_init in range(32):
@@ -1429,14 +1411,14 @@ to `C` when all the block results are ready.
cse_var_3: T.int32 = x_c * 32
cse_var_2: T.int32 = y_outer * 1024 + cse_var_4
cse_var_1: T.int32 = x_outer * 32768 + x_c * 1024 + cse_var_4
- A_2 = T.buffer_decl((1048576,), data=A_1.data)
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1], 32) * packedB_1[cse_var_2]
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1 + 1], 32) * packedB_1[cse_var_2 + 1]
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1 + 2], 32) * packedB_1[cse_var_2 + 2]
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1 + 3], 32) * packedB_1[cse_var_2 + 3]
+ A_1 = T.buffer_decl((1048576,), data=A.data)
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1], 32) * packedB_1[cse_var_2]
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1 + 1], 32) * packedB_1[cse_var_2 + 1]
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1 + 2], 32) * packedB_1[cse_var_2 + 2]
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1 + 3], 32) * packedB_1[cse_var_2 + 3]
for x_inner, y_inner in T.grid(32, 32):
- C_2 = T.buffer_decl((1048576,), data=C_1.data)
- C_2[x_outer * 32768 + x_inner * 1024 + y_outer * 32 + y_inner] = C_global_1[x_inner * 32 + y_inner]
+ C_1 = T.buffer_decl((1048576,), data=C.data)
+ C_1[x_outer * 32768 + x_inner * 1024 + y_outer * 32 + y_inner] = C_global_1[x_inner * 32 + y_inner]
@@ -1478,21 +1460,18 @@ of thread-level parallelization.
.. code-block:: none
- parallelization: 0.146343
+ parallelization: 0.146423
@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 1024), "float32"), B: T.Buffer((1024, 1024), "float32"), C: T.Buffer((1024, 1024), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 1024))
- B_1 = T.match_buffer(B, (1024, 1024))
- C_1 = T.match_buffer(C, (1024, 1024))
packedB = T.allocate([32768], "float32x32", "global")
packedB_1 = T.buffer_decl((32768,), "float32x32", data=packedB)
for x in T.parallel(32):
for y in range(1024):
- B_2 = T.buffer_decl((1048576,), data=B_1.data)
- packedB_1[x * 1024 + y] = B_2[y * 1024 + x * 32:y * 1024 + x * 32 + 32]
+ B_1 = T.buffer_decl((1048576,), data=B.data)
+ packedB_1[x * 1024 + y] = B_1[y * 1024 + x * 32:y * 1024 + x * 32 + 32]
for x_outer in T.parallel(32):
C_global = T.allocate([1024], "float32", "global")
for y_outer in range(32):
@@ -1504,14 +1483,14 @@ of thread-level parallelization.
cse_var_3: T.int32 = x_c * 32
cse_var_2: T.int32 = y_outer * 1024 + cse_var_4
cse_var_1: T.int32 = x_outer * 32768 + x_c * 1024 + cse_var_4
- A_2 = T.buffer_decl((1048576,), data=A_1.data)
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1], 32) * packedB_1[cse_var_2]
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1 + 1], 32) * packedB_1[cse_var_2 + 1]
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1 + 2], 32) * packedB_1[cse_var_2 + 2]
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1 + 3], 32) * packedB_1[cse_var_2 + 3]
+ A_1 = T.buffer_decl((1048576,), data=A.data)
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1], 32) * packedB_1[cse_var_2]
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1 + 1], 32) * packedB_1[cse_var_2 + 1]
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1 + 2], 32) * packedB_1[cse_var_2 + 2]
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1 + 3], 32) * packedB_1[cse_var_2 + 3]
for x_inner, y_inner in T.grid(32, 32):
- C_2 = T.buffer_decl((1048576,), data=C_1.data)
- C_2[x_outer * 32768 + x_inner * 1024 + y_outer * 32 + y_inner] = C_global_1[x_inner * 32 + y_inner]
+ C_1 = T.buffer_decl((1048576,), data=C.data)
+ C_1[x_outer * 32768 + x_inner * 1024 + y_outer * 32 + y_inner] = C_global_1[x_inner * 32 + y_inner]
@@ -1548,13 +1527,13 @@ working, we can compare the results.
.. code-block:: none
Operator Timing Performance
- none 3.3998205908 1.0
- blocking 0.3007591159 0.08846323147576131
- vectorization 0.3444209805 0.10130563401845727
- loop permutation 0.1159568229 0.0341067476365612
- array packing 0.1084781797 0.03190703062201125
- block caching 0.11081503379999999 0.0325943769209082
- parallelization 0.1463433178 0.04304442363694387
+ none 3.2584962005 1.0
+ blocking 0.297529924 0.09130896760117305
+ vectorization 0.3301182801 0.10131000921509282
+ loop permutation 0.11976749480000001 0.03675545019252202
+ array packing 0.1099097601 0.03373020968480334
+ block caching 0.11072204810000001 0.033979492774308055
+ parallelization 0.1464230078 0.04493576140353704
@@ -1594,11 +1573,6 @@ operations with tunable parameters that allows you to automatically optimize
the computation for specific platforms.
-.. rst-class:: sphx-glr-timing
-
- **Total running time of the script:** ( 1 minutes 0.967 seconds)
-
-
.. _sphx_glr_download_tutorial_tensor_expr_get_started.py:
.. only:: html
diff --git a/docs/_sources/tutorial/tensor_ir_blitz_course.rst.txt b/docs/_sources/tutorial/tensor_ir_blitz_course.rst.txt
index 3fb6484e44..531d04128d 100644
--- a/docs/_sources/tutorial/tensor_ir_blitz_course.rst.txt
+++ b/docs/_sources/tutorial/tensor_ir_blitz_course.rst.txt
@@ -129,10 +129,8 @@ Following is a simple example for vector addition.
@I.ir_module
class Module:
@T.prim_func
- def main(a: T.handle, b: T.handle):
+ def main(A: T.Buffer((8,), "float32"), B: T.Buffer((8,), "float32")):
T.func_attr({"global_symbol": "main", "tir.noalias": True})
- A = T.match_buffer(a, (8,))
- B = T.match_buffer(b, (8,))
with T.block("root"):
T.reads()
T.writes()
@@ -177,10 +175,8 @@ to an IRModule.
@I.ir_module
class Module:
@T.prim_func
- def main(var_A: T.handle, var_B: T.handle):
+ def main(A: T.Buffer((8,), "float32"), B: T.Buffer((8,), "float32")):
T.func_attr({"global_symbol": "main", "tir.noalias": True})
- A = T.match_buffer(var_A, (8,))
- B = T.match_buffer(var_B, (8,))
with T.block("root"):
T.reads()
T.writes()
@@ -323,10 +319,8 @@ Tile the loop into 3 loops and print the result.
@I.ir_module
class Module:
@T.prim_func
- def main(a: T.handle, b: T.handle):
+ def main(A: T.Buffer((8,), "float32"), B: T.Buffer((8,), "float32")):
T.func_attr({"global_symbol": "main", "tir.noalias": True})
- A = T.match_buffer(a, (8,))
- B = T.match_buffer(b, (8,))
with T.block("root"):
T.reads()
T.writes()
@@ -363,10 +357,8 @@ We can also reorder the loops. Now we move loop `i_2` to outside of `i_1`.
@I.ir_module
class Module:
@T.prim_func
- def main(a: T.handle, b: T.handle):
+ def main(A: T.Buffer((8,), "float32"), B: T.Buffer((8,), "float32")):
T.func_attr({"global_symbol": "main", "tir.noalias": True})
- A = T.match_buffer(a, (8,))
- B = T.match_buffer(b, (8,))
with T.block("root"):
T.reads()
T.writes()
@@ -409,10 +401,8 @@ also use primitives and do incrementally transformation.
@I.ir_module
class Module:
@T.prim_func
- def main(a: T.handle, b: T.handle):
+ def main(A: T.Buffer((8,), "float32"), B: T.Buffer((8,), "float32")):
T.func_attr({"global_symbol": "main", "tir.noalias": True})
- A = T.match_buffer(a, (8,))
- B = T.match_buffer(b, (8,))
with T.block("root"):
T.reads()
T.writes()
diff --git a/docs/commit_hash b/docs/commit_hash
index 7bdb802766..9260decc08 100644
--- a/docs/commit_hash
+++ b/docs/commit_hash
@@ -1 +1 @@
-cfa65b26c1bd975daaef78c60b16989be0d23970
+6c2d485a011fbfbd426353c6fc1254f3385d826e
diff --git a/docs/how_to/compile_models/from_darknet.html b/docs/how_to/compile_models/from_darknet.html
index 4a1fa62995..bf01b558c0 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 16.257 seconds)</p>
+<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 1 minutes 13.727 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 20f7f0cc9d..bb9ebcf6a8 100644
--- a/docs/how_to/compile_models/from_keras.html
+++ b/docs/how_to/compile_models/from_keras.html
@@ -506,7 +506,7 @@ Tensorflow is also required since it’s used as the default backend of keras.</
<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 958ms/step
+1/1 [==============================] - 1s 938ms/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 6804b8360c..b6b7aac5e7 100644
--- a/docs/how_to/compile_models/from_mxnet.html
+++ b/docs/how_to/compile_models/from_mxnet.html
@@ -439,7 +439,7 @@
<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.zip74b370c8-6b03-402b-8791-7ba7f0d5f123 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.zip281edc03-6311-457c-b054-e2f5809f278a 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 876d3eb660..4b89f675d4 100644
--- a/docs/how_to/compile_models/from_oneflow.html
+++ b/docs/how_to/compile_models/from_oneflow.html
@@ -449,13 +449,13 @@ 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, 51.7MB/s]
- 39%|###8 | 16.0M/41.5M [00:00<00:00, 48.4MB/s]
- 58%|#####7 | 24.0M/41.5M [00:00<00:00, 45.0MB/s]
- 77%|#######7 | 32.0M/41.5M [00:00<00:00, 47.3MB/s]
- 90%|########9 | 37.2M/41.5M [00:00<00:00, 44.8MB/s]
-100%|#########9| 41.4M/41.5M [00:00<00:00, 43.5MB/s]
-100%|##########| 41.5M/41.5M [00:00<00:00, 45.3MB/s]
+ 15%|#5 | 6.33M/41.5M [00:00<00:00, 51.0MB/s]
+ 33%|###3 | 13.9M/41.5M [00:00<00:00, 65.7MB/s]
+ 49%|####8 | 20.3M/41.5M [00:00<00:00, 51.4MB/s]
+ 62%|######1 | 25.6M/41.5M [00:00<00:00, 50.1MB/s]
+ 77%|#######7 | 32.0M/41.5M [00:00<00:00, 52.8MB/s]
+ 92%|#########2| 38.3M/41.5M [00:00<00:00, 45.2MB/s]
+100%|##########| 41.5M/41.5M [00:00<00:00, 49.6MB/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 a919ad73f4..fd3b06245a 100644
--- a/docs/how_to/compile_models/from_pytorch.html
+++ b/docs/how_to/compile_models/from_pytorch.html
@@ -432,10 +432,10 @@ 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]
- 24%|##3 | 10.6M/44.7M [00:00<00:00, 112MB/s]
- 51%|#####1 | 22.9M/44.7M [00:00<00:00, 122MB/s]
- 77%|#######7 | 34.5M/44.7M [00:00<00:00, 108MB/s]
-100%|##########| 44.7M/44.7M [00:00<00:00, 108MB/s]
+ 23%|##2 | 10.1M/44.7M [00:00<00:00, 74.9MB/s]
+ 50%|##### | 22.4M/44.7M [00:00<00:00, 102MB/s]
+ 73%|#######3 | 32.6M/44.7M [00:00<00:00, 87.4MB/s]
+100%|##########| 44.7M/44.7M [00:00<00:00, 98.9MB/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 06b96f6c76..b2a35a7cc1 100644
--- a/docs/how_to/compile_models/from_tensorflow.html
+++ b/docs/how_to/compile_models/from_tensorflow.html
@@ -649,7 +649,7 @@ 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 20.531 seconds)</p>
+<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 1 minutes 18.160 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 b367a271a0..4f30f972c5 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>06:19.340</strong> total execution time for <strong>how_to_compile_models</strong> files:</p>
+<p><strong>06:07.945</strong> total execution time for <strong>how_to_compile_models</strong> files:</p>
<table class="docutils align-default">
<colgroup>
<col style="width: 81%" />
@@ -349,43 +349,43 @@
</colgroup>
<tbody>
<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>01:20.531</p></td>
+<td><p>01:18.160</p></td>
<td><p>0.0 MB</p></td>
</tr>
<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>01:16.257</p></td>
+<td><p>01:13.727</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>00:52.501</p></td>
+<td><p>00:50.305</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>00:35.465</p></td>
+<td><p>00:34.517</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>00:30.357</p></td>
+<td><p>00:29.818</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>00:30.149</p></td>
+<td><p>00:29.515</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>00:26.165</p></td>
+<td><p>00:26.227</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>00:24.732</p></td>
+<td><p>00:23.431</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>00:20.531</p></td>
+<td><p>00:19.695</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>00:02.651</p></td>
+<td><p>00:02.550</p></td>
<td><p>0.0 MB</p></td>
</tr>
</tbody>
diff --git a/docs/how_to/deploy_models/deploy_model_on_adreno.html b/docs/how_to/deploy_models/deploy_model_on_adreno.html
index 6db8005f14..cef60e39a3 100644
--- a/docs/how_to/deploy_models/deploy_model_on_adreno.html
+++ b/docs/how_to/deploy_models/deploy_model_on_adreno.html
@@ -920,7 +920,7 @@ Top5 predictions:
Evaluate inference time cost...
Execution time summary:
mean (ms) median (ms) max (ms) min (ms) std (ms)
- 2550.8269 2549.1817 2566.5008 2545.9899 5.4542
+ 2758.3525 2758.4177 2760.3811 2755.1281 1.6989
</pre></div>
</div>
<div class="sphx-glr-footer sphx-glr-footer-example docutils container" id="sphx-glr-download-how-to-deploy-models-deploy-model-on-adreno-py">
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 c996eb1624..01992373b2 100644
--- a/docs/how_to/deploy_models/deploy_model_on_android.html
+++ b/docs/how_to/deploy_models/deploy_model_on_android.html
@@ -662,7 +662,7 @@ to the remote android device.</p>
Evaluate inference time cost...
Execution time summary:
mean (ms) median (ms) max (ms) min (ms) std (ms)
- 16.0102 15.8558 16.7513 15.6779 0.3372
+ 15.6184 15.5163 16.1542 15.4111 0.2631
</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 c99332ed72..41d1422989 100644
--- a/docs/how_to/deploy_models/deploy_object_detection_pytorch.html
+++ b/docs/how_to/deploy_models/deploy_object_detection_pytorch.html
@@ -454,24 +454,24 @@ 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]
- 5%|4 | 7.99M/170M [00:00<00:02, 74.9MB/s]
- 9%|8 | 15.1M/170M [00:00<00:02, 59.8MB/s]
- 13%|#3 | 22.9M/170M [00:00<00:02, 68.1MB/s]
- 18%|#7 | 30.3M/170M [00:00<00:02, 53.7MB/s]
- 21%|##1 | 35.9M/170M [00:00<00:03, 46.6MB/s]
- 24%|##3 | 40.7M/170M [00:00<00:02, 47.0MB/s]
- 33%|###2 | 56.0M/170M [00:00<00:01, 74.1MB/s]
- 38%|###7 | 64.0M/170M [00:01<00:01, 74.4MB/s]
- 43%|####3 | 73.4M/170M [00:01<00:01, 80.8MB/s]
- 52%|#####1 | 88.0M/170M [00:01<00:00, 96.1MB/s]
- 61%|######1 | 104M/170M [00:01<00:00, 99.5MB/s]
- 67%|######6 | 114M/170M [00:01<00:00, 88.9MB/s]
- 72%|#######1 | 122M/170M [00:01<00:00, 87.8MB/s]
- 77%|#######6 | 131M/170M [00:01<00:00, 70.9MB/s]
- 81%|########1 | 138M/170M [00:01<00:00, 72.0MB/s]
- 86%|########5 | 145M/170M [00:02<00:00, 70.0MB/s]
- 94%|#########4| 160M/170M [00:02<00:00, 84.1MB/s]
-100%|##########| 170M/170M [00:02<00:00, 78.1MB/s]
+ 6%|5 | 10.1M/170M [00:00<00:01, 86.0MB/s]
+ 12%|#2 | 21.2M/170M [00:00<00:01, 97.2MB/s]
+ 18%|#7 | 30.5M/170M [00:00<00:01, 97.6MB/s]
+ 24%|##3 | 40.0M/170M [00:00<00:01, 80.4MB/s]
+ 32%|###1 | 53.9M/170M [00:00<00:01, 101MB/s]
+ 38%|###7 | 64.1M/170M [00:00<00:01, 83.8MB/s]
+ 43%|####2 | 72.8M/170M [00:00<00:01, 81.9MB/s]
+ 48%|####7 | 81.0M/170M [00:01<00:01, 75.3MB/s]
+ 53%|#####3 | 90.6M/170M [00:01<00:01, 81.0MB/s]
+ 59%|#####8 | 99.7M/170M [00:01<00:00, 84.7MB/s]
+ 64%|######3 | 108M/170M [00:01<00:00, 76.1MB/s]
+ 71%|####### | 120M/170M [00:01<00:00, 82.0MB/s]
+ 75%|#######5 | 128M/170M [00:01<00:00, 74.7MB/s]
+ 80%|######## | 136M/170M [00:01<00:00, 67.7MB/s]
+ 86%|########5 | 146M/170M [00:01<00:00, 75.6MB/s]
+ 90%|######### | 153M/170M [00:02<00:00, 64.2MB/s]
+ 95%|#########5| 162M/170M [00:02<00:00, 68.0MB/s]
+100%|##########| 170M/170M [00:02<00:00, 78.6MB/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=& [...]
@@ -569,7 +569,7 @@ 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 28.344 seconds)</p>
+<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 3 minutes 21.632 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 1450cffda8..45578ed3cd 100644
--- a/docs/how_to/deploy_models/deploy_prequantized.html
+++ b/docs/how_to/deploy_models/deploy_prequantized.html
@@ -495,9 +495,8 @@ 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]
- 47%|####6 | 6.30M/13.6M [00:00<00:00, 53.4MB/s]
- 84%|########4 | 11.4M/13.6M [00:00<00:00, 50.5MB/s]
-100%|##########| 13.6M/13.6M [00:00<00:00, 39.7MB/s]
+ 59%|#####8 | 7.99M/13.6M [00:00<00:00, 76.8MB/s]
+100%|##########| 13.6M/13.6M [00:00<00:00, 103MB/s]
</pre></div>
</div>
</div>
@@ -588,7 +587,7 @@ 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)
- 90.2068 90.1677 90.8147 90.0469 0.1458
+ 90.1720 90.1005 91.6585 89.8012 0.3037
</pre></div>
</div>
<div class="admonition note">
@@ -627,7 +626,7 @@ 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 12.919 seconds)</p>
+<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 1 minutes 10.916 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 bb2b0351de..6e678ab148 100644
--- a/docs/how_to/deploy_models/deploy_prequantized_tflite.html
+++ b/docs/how_to/deploy_models/deploy_prequantized_tflite.html
@@ -580,7 +580,7 @@ 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)
- 119.6592 119.5643 126.8230 119.0097 0.8249
+ 118.3536 118.3627 119.9356 117.1745 0.4004
</pre></div>
</div>
<div class="admonition note">
@@ -608,7 +608,7 @@ 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 32.610 seconds)</p>
+<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 2 minutes 25.221 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 9e0912f43a..9a88f9abac 100644
--- a/docs/how_to/deploy_models/deploy_quantized.html
+++ b/docs/how_to/deploy_models/deploy_quantized.html
@@ -521,7 +521,7 @@ 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 40.483 seconds)</p>
+<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 1 minutes 32.344 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 9c67559d98..47f5d0bb06 100644
--- a/docs/how_to/deploy_models/deploy_ssd_gluoncv.html
+++ b/docs/how_to/deploy_models/deploy_ssd_gluoncv.html
@@ -463,22 +463,24 @@ 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]
- 5%|5 | 6645/132723 [00:00<00:01, 66444.68KB/s]
- 12%|#1 | 15500/132723 [00:00<00:01, 79443.08KB/s]
- 18%|#8 | 24276/132723 [00:00<00:01, 83238.71KB/s]
- 25%|##4 | 33050/132723 [00:00<00:01, 85011.97KB/s]
- 31%|###1 | 41552/132723 [00:00<00:01, 84301.38KB/s]
- 38%|###7 | 49984/132723 [00:00<00:01, 81095.71KB/s]
- 44%|####4 | 58717/132723 [00:00<00:00, 83069.29KB/s]
- 51%|##### | 67526/132723 [00:00<00:00, 84632.76KB/s]
- 57%|#####7 | 76271/132723 [00:00<00:00, 85499.51KB/s]
- 64%|######4 | 85023/132723 [00:01<00:00, 86114.52KB/s]
- 71%|####### | 93817/132723 [00:01<00:00, 86665.19KB/s]
- 77%|#######7 | 102637/132723 [00:01<00:00, 87119.68KB/s]
- 84%|########3 | 111464/132723 [00:01<00:00, 87464.64KB/s]
- 91%|######### | 120251/132723 [00:01<00:00, 87584.54KB/s]
- 97%|#########7| 129109/132723 [00:01<00:00, 87880.14KB/s]
-100%|##########| 132723/132723 [00:01<00:00, 85230.00KB/s]
+ 3%|3 | 4409/132723 [00:00<00:02, 43293.70KB/s]
+ 8%|8 | 10936/132723 [00:00<00:02, 56120.51KB/s]
+ 14%|#3 | 18550/132723 [00:00<00:01, 65219.72KB/s]
+ 20%|#9 | 26096/132723 [00:00<00:01, 69251.04KB/s]
+ 25%|##5 | 33696/132723 [00:00<00:01, 71679.02KB/s]
+ 31%|###1 | 41335/132723 [00:00<00:01, 73276.82KB/s]
+ 37%|###6 | 48867/132723 [00:00<00:01, 73941.19KB/s]
+ 43%|####2 | 56468/132723 [00:00<00:01, 74596.86KB/s]
+ 48%|####8 | 64162/132723 [00:00<00:00, 75327.52KB/s]
+ 54%|#####4 | 71814/132723 [00:01<00:00, 75692.40KB/s]
+ 60%|#####9 | 79443/132723 [00:01<00:00, 75873.67KB/s]
+ 66%|######5 | 87088/132723 [00:01<00:00, 76047.17KB/s]
+ 71%|#######1 | 94694/132723 [00:01<00:00, 76034.39KB/s]
+ 77%|#######7 | 102397/132723 [00:01<00:00, 76333.30KB/s]
+ 83%|########2 | 110031/132723 [00:01<00:00, 76333.69KB/s]
+ 89%|########8 | 117703/132723 [00:01<00:00, 76447.67KB/s]
+ 94%|#########4| 125369/132723 [00:01<00:00, 76510.11KB/s]
+100%|##########| 132723/132723 [00:01<00:00, 73840.87KB/s]
</pre></div>
</div>
<p>Create TVM runtime and do inference
@@ -517,7 +519,7 @@ 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 29.026 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 26.377 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 2afdb80e1a..ba211fb55f 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>14:51.021</strong> total execution time for <strong>how_to_deploy_models</strong> files:</p>
+<p><strong>14:22.337</strong> total execution time for <strong>how_to_deploy_models</strong> files:</p>
<table class="docutils align-default">
<colgroup>
<col style="width: 86%" />
@@ -349,39 +349,39 @@
</colgroup>
<tbody>
<tr class="row-odd"><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>03:29.026</p></td>
+<td><p>03:26.377</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><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>03:28.344</p></td>
+<td><p>03:21.632</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>02:32.610</p></td>
+<td><p>02:25.221</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>01:40.483</p></td>
+<td><p>01:32.344</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>01:12.919</p></td>
+<td><p>01:10.916</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="deploy_model_on_adreno.html#sphx-glr-how-to-deploy-models-deploy-model-on-adreno-py"><span class="std std-ref">Deploy the Pretrained Model on Adreno</span></a> (<code class="docutils literal notranslate"><span class="pre">deploy_model_on_adreno.py</span></code>)</p></td>
-<td><p>00:53.319</p></td>
+<td><p>00:55.272</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-odd"><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>00:39.824</p></td>
+<td><p>00:38.670</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><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>00:27.427</p></td>
+<td><p>00:26.085</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-odd"><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>00:27.063</p></td>
+<td><p>00:25.814</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><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 ab570a2268..5835b09540 100644
--- a/docs/how_to/extend_tvm/bring_your_own_datatypes.html
+++ b/docs/how_to/extend_tvm/bring_your_own_datatypes.html
@@ -619,7 +619,7 @@ 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.zipcb3154e2-9390-47b4-8140-e0c63d32f07a 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.zip07fb9aa0-cc7e-474e-ab93-9e3af8a9a86f 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 29fd10cbb5..a6865ebf92 100644
--- a/docs/how_to/extend_tvm/low_level_custom_pass.html
+++ b/docs/how_to/extend_tvm/low_level_custom_pass.html
@@ -417,16 +417,13 @@ our customized lowering pass to manipulate the IR directly instead of using sche
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@I.ir_module
class Module:
@T.prim_func
- def main(a: T.handle, b: T.handle, c: T.handle):
+ def main(a: T.Buffer((128,), "float32"), b: T.Buffer((128,), "float32"), c: T.Buffer((128,), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- a_1 = T.match_buffer(a, (128,))
- b_1 = T.match_buffer(b, (128,))
- c_1 = T.match_buffer(c, (128,))
for i in range(128):
- c_2 = T.buffer_decl((128,), data=c_1.data)
- a_2 = T.buffer_decl((128,), data=a_1.data)
- b_2 = T.buffer_decl((128,), data=b_1.data)
- c_2[i] = a_2[i] + b_2[i]
+ c_1 = T.buffer_decl((128,), data=c.data)
+ a_1 = T.buffer_decl((128,), data=a.data)
+ b_1 = T.buffer_decl((128,), data=b.data)
+ c_1[i] = a_1[i] + b_1[i]
</pre></div>
</div>
</div>
@@ -528,17 +525,14 @@ called after each phase is done.</p>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@I.ir_module
class Module:
@T.prim_func
- def main(a: T.handle, b: T.handle, c: T.handle):
+ def main(a: T.Buffer((128,), "float32"), b: T.Buffer((128,), "float32"), c: T.Buffer((128,), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- a_1 = T.match_buffer(a, (128,))
- b_1 = T.match_buffer(b, (128,))
- c_1 = T.match_buffer(c, (128,))
for i_outer in range(16):
cse_var_1: T.int32 = i_outer * 8
- c_2 = T.buffer_decl((128,), data=c_1.data)
- a_2 = T.buffer_decl((128,), data=a_1.data)
- b_2 = T.buffer_decl((128,), data=b_1.data)
- c_2[cse_var_1:cse_var_1 + 8] = a_2[cse_var_1:cse_var_1 + 8] + b_2[cse_var_1:cse_var_1 + 8]
+ c_1 = T.buffer_decl((128,), data=c.data)
+ a_1 = T.buffer_decl((128,), data=a.data)
+ b_1 = T.buffer_decl((128,), data=b.data)
+ c_1[cse_var_1:cse_var_1 + 8] = a_1[cse_var_1:cse_var_1 + 8] + b_1[cse_var_1:cse_var_1 + 8]
</pre></div>
</div>
</div>
diff --git a/docs/how_to/extend_tvm/sg_execution_times.html b/docs/how_to/extend_tvm/sg_execution_times.html
index f7d6cab4ff..a8372c5eb6 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>00:52.015</strong> total execution time for <strong>how_to_extend_tvm</strong> files:</p>
+<p><strong>00:50.546</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>00:48.296</p></td>
+<td><p>00:46.963</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>00:02.653</p></td>
+<td><p>00:02.544</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>00:01.060</p></td>
+<td><p>00:01.033</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 3195fc97d4..81870de64a 100644
--- a/docs/how_to/extend_tvm/use_pass_instrument.html
+++ b/docs/how_to/extend_tvm/use_pass_instrument.html
@@ -526,10 +526,10 @@ 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: 18158us [18158us] (48.55%; 48.55%)
-FoldScaleAxis: 19244us [7us] (51.45%; 51.45%)
- FoldConstant: 19237us [1733us] (51.43%; 99.96%)
- InferType: 17505us [17505us] (46.80%; 90.99%)
+InferType: 17614us [17614us] (48.44%; 48.44%)
+FoldScaleAxis: 18746us [6us] (51.56%; 51.56%)
+ FoldConstant: 18740us [1648us] (51.54%; 99.97%)
+ InferType: 17091us [17091us] (47.00%; 91.20%)
</pre></div>
</div>
</div>
@@ -551,10 +551,10 @@ 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: 17363us [17363us] (47.85%; 47.85%)
-FoldScaleAxis: 18921us [5us] (52.15%; 52.15%)
- FoldConstant: 18916us [1709us] (52.13%; 99.97%)
- InferType: 17208us [17208us] (47.42%; 90.97%)
+InferType: 17137us [17137us] (47.94%; 47.94%)
+FoldScaleAxis: 18611us [5us] (52.06%; 52.06%)
+ FoldConstant: 18606us [1663us] (52.05%; 99.97%)
+ InferType: 16943us [16943us] (47.40%; 91.06%)
</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 7bab61c423..f8ac9dede1 100644
--- a/docs/how_to/optimize_operators/opt_conv_cuda.html
+++ b/docs/how_to/optimize_operators/opt_conv_cuda.html
@@ -575,7 +575,7 @@ 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: 44.062721 ms
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Convolution: 54.217983 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 179e3441eb..28f20bd46a 100644
--- a/docs/how_to/optimize_operators/opt_conv_tensorcore.html
+++ b/docs/how_to/optimize_operators/opt_conv_tensorcore.html
@@ -726,11 +726,8 @@ The only thing we should do is to make sure all threads in a warp can call Tenso
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, W: T.handle, Conv: T.handle):
+ def main(A: T.Buffer((16, 14, 14, 16, 16, 16), "float16"), W: T.Buffer((3, 3, 16, 32, 16, 16), "float16"), Conv: T.Buffer((16, 14, 14, 32, 16, 16), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (16, 14, 14, 16, 16, 16), "float16")
- W_1 = T.match_buffer(W, (3, 3, 16, 32, 16, 16), "float16")
- Conv_1 = T.match_buffer(Conv, (16, 14, 14, 32, 16, 16))
blockIdx_z = T.env_thread("blockIdx.z")
T.launch_thread(blockIdx_z, 196)
Conv_wmma_accumulator = T.allocate([2048], "float32", "wmma.accumulator")
@@ -756,13 +753,13 @@ class Module:
cse_var_2: T.int32 = ax3 * 256
cse_var_1: T.int32 = ax4_ax5_fused_outer * 32
T.launch_thread(threadIdx_x, 32)
- A_2 = T.buffer_decl((12845056,), "float16", data=A_1.data)
- Apad_shared_1[threadIdx_y * 3072 + threadIdx_z * 1536 + ax2 * 512 + cse_var_2 + cse_var_1 + threadIdx_x] = T.if_then_else(1 <= blockIdx_z // 14 + kh and blockIdx_z // 14 + kh < 15 and 1 <= ax2 + blockIdx_z % 14 and ax2 + blockIdx_z % 14 < 15, A_2[blockIdx_x * 6422528 + threadIdx_y * 1605632 + threadIdx_z * 802816 + kh * 57344 + blockIdx_z * 4096 + ax2 * 4096 + ic_outer * 512 + cse_var_2 + cse_var_1 + threadIdx_x - 61440], T.float16(0))
+ A_1 = T.buffer_decl((12845056,), "float16", data=A.data)
+ Apad_shared_1[threadIdx_y * 3072 + threadIdx_z * 1536 + ax2 * 512 + cse_var_2 + cse_var_1 + threadIdx_x] = T.if_then_else(1 <= blockIdx_z // 14 + kh and blockIdx_z // 14 + kh < 15 and 1 <= ax2 + blockIdx_z % 14 and ax2 + blockIdx_z % 14 < 15, A_1[blockIdx_x * 6422528 + threadIdx_y * 1605632 + threadIdx_z * 802816 + kh * 57344 + blockIdx_z * 4096 + ax2 * 4096 + ic_outer * 512 + cse_var_2 + cse_var_1 + threadIdx_x - 61440], T.float16(0))
W_shared_1 = T.buffer_decl((12288,), "float16", data=W_shared, scope="shared")
for ax1, ax2 in T.grid(3, 2):
T.launch_thread(threadIdx_x, 32)
- W_2 = T.buffer_decl((1179648,), "float16", data=W_1.data)
- W_shared_1[ax1 * 4096 + ax2 * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8:ax1 * 4096 + ax2 * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8 + 8] = W_2[kh * 393216 + ax1 * 131072 + ic_outer * 16384 + ax2 * 8192 + blockIdx_y * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8:kh * 393216 + ax1 * 131072 + ic_outer * 16384 + ax2 * 8192 + blockIdx_y * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8 + 8]
+ W_1 = T.buffer_decl((1179648,), "float16", data=W.data)
+ W_shared_1[ax1 * 4096 + ax2 * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8:ax1 * 4096 + ax2 * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8 + 8] = W_1[kh * 393216 + ax1 * 131072 + ic_outer * 16384 + ax2 * 8192 + blockIdx_y * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8:kh * 393216 + ax1 * 131072 + ic_outer * 16384 + ax2 * 8192 + blockIdx_y * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8 + 8]
for ic_inner, kw in T.grid(2, 3):
Apad_shared_wmma_matrix_a_1 = T.buffer_decl((512,), "float16", data=Apad_shared_wmma_matrix_a, scope="wmma.matrix_a")
for ax0, ax4, ax5 in T.grid(2, 16, 16):
@@ -781,8 +778,8 @@ class Module:
for n_inner, o_inner, nn, oo in T.grid(2, 4, 16, 16):
cse_var_10: T.int32 = o_inner * 256
cse_var_9: T.int32 = nn * 16
- Conv_2 = T.buffer_decl((25690112,), data=Conv_1.data)
- Conv_2[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_1 = T.buffer_decl((25690112,), data=Conv.data)
+ Conv_1[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]
</pre></div>
</div>
</div>
@@ -800,11 +797,8 @@ by mapping the 2D convolution to tensor intrinsics</p>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, W: T.handle, Conv: T.handle):
+ def main(A: T.Buffer((16, 14, 14, 16, 16, 16), "float16"), W: T.Buffer((3, 3, 16, 32, 16, 16), "float16"), Conv: T.Buffer((16, 14, 14, 32, 16, 16), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (16, 14, 14, 16, 16, 16), "float16")
- W_1 = T.match_buffer(W, (3, 3, 16, 32, 16, 16), "float16")
- Conv_1 = T.match_buffer(Conv, (16, 14, 14, 32, 16, 16))
blockIdx_z = T.env_thread("blockIdx.z")
T.launch_thread(blockIdx_z, 196)
Conv_wmma_accumulator = T.allocate([2048], "float32", "wmma.accumulator")
@@ -829,13 +823,13 @@ class Module:
cse_var_1: T.int32 = ax4_ax5_fused_outer * 32
T.launch_thread(threadIdx_x, 32)
Apad_shared_1 = T.buffer_decl((12288,), "float16", data=Apad_shared, scope="shared")
- A_2 = T.buffer_decl((12845056,), "float16", data=A_1.data)
- Apad_shared_1[threadIdx_y * 3072 + threadIdx_z * 1536 + ax2 * 512 + cse_var_2 + cse_var_1 + threadIdx_x] = T.if_then_else(1 <= blockIdx_z // 14 + kh and blockIdx_z // 14 + kh < 15 and 1 <= ax2 + blockIdx_z % 14 and ax2 + blockIdx_z % 14 < 15, A_2[blockIdx_x * 6422528 + threadIdx_y * 1605632 + threadIdx_z * 802816 + kh * 57344 + blockIdx_z * 4096 + ax2 * 4096 + ic_outer * 512 + cse_var_2 + cse_var_1 + threadIdx_x - 61440], T.float16(0))
+ A_1 = T.buffer_decl((12845056,), "float16", data=A.data)
+ Apad_shared_1[threadIdx_y * 3072 + threadIdx_z * 1536 + ax2 * 512 + cse_var_2 + cse_var_1 + threadIdx_x] = T.if_then_else(1 <= blockIdx_z // 14 + kh and blockIdx_z // 14 + kh < 15 and 1 <= ax2 + blockIdx_z % 14 and ax2 + blockIdx_z % 14 < 15, A_1[blockIdx_x * 6422528 + threadIdx_y * 1605632 + threadIdx_z * 802816 + kh * 57344 + blockIdx_z * 4096 + ax2 * 4096 + ic_outer * 512 + cse_var_2 + cse_var_1 + threadIdx_x - 61440], T.float16(0))
for ax1, ax2 in T.grid(3, 2):
T.launch_thread(threadIdx_x, 32)
W_shared_1 = T.buffer_decl((12288,), "float16", data=W_shared, scope="shared")
- W_2 = T.buffer_decl((1179648,), "float16", data=W_1.data)
- W_shared_1[ax1 * 4096 + ax2 * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8:ax1 * 4096 + ax2 * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8 + 8] = W_2[kh * 393216 + ax1 * 131072 + ic_outer * 16384 + ax2 * 8192 + blockIdx_y * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8:kh * 393216 + ax1 * 131072 + ic_outer * 16384 + ax2 * 8192 + blockIdx_y * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8 + 8]
+ W_1 = T.buffer_decl((1179648,), "float16", data=W.data)
+ W_shared_1[ax1 * 4096 + ax2 * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8:ax1 * 4096 + ax2 * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8 + 8] = W_1[kh * 393216 + ax1 * 131072 + ic_outer * 16384 + ax2 * 8192 + blockIdx_y * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8:kh * 393216 + ax1 * 131072 + ic_outer * 16384 + ax2 * 8192 + blockIdx_y * 2048 + threadIdx_y * 512 + threadIdx_z * 256 + threadIdx_x * 8 + 8]
for ic_inner, kw in T.grid(2, 3):
for ax0 in range(2):
T.tvm_load_matrix_sync(Apad_shared_wmma_matrix_a, 16, 16, 16, ax0, T.tvm_access_ptr(T.type_annotation("float16"), Apad_shared, threadIdx_y * 3072 + ax0 * 1536 + kw * 512 + ic_inner * 256, 256, 1), 16, "row_major")
@@ -845,7 +839,7 @@ class Module:
cse_var_3: T.int32 = n_c * 4 + o_c
T.tvm_mma_sync(Conv_wmma_accumulator, cse_var_3, Apad_shared_wmma_matrix_a, n_c, W_shared_wmma_matrix_b, o_c, Conv_wmma_accumulator, cse_var_3)
for n_inner, o_inner in T.grid(2, 4):
- T.tvm_store_matrix_sync(Conv_wmma_accumulator, 16, 16, 16, n_inner * 4 + o_inner, T.tvm_access_ptr(T.type_annotation("float32"), Conv_1.data, blockIdx_x * 12845056 + threadIdx_y * 3211264 + n_inner * 1605632 + blockIdx_z * 8192 + blockIdx_y * 2048 + threadIdx_z * 1024 + o_inner * 256, 256, 2), 16, "row_major")
+ T.tvm_store_matrix_sync(Conv_wmma_accumulator, 16, 16, 16, n_inner * 4 + o_inner, T.tvm_access_ptr(T.type_annotation("float32"), Conv.data, blockIdx_x * 12845056 + threadIdx_y * 3211264 + n_inner * 1605632 + blockIdx_z * 8192 + blockIdx_y * 2048 + threadIdx_z * 1024 + o_inner * 256, 256, 2), 16, "row_major")
</pre></div>
</div>
</div>
@@ -867,7 +861,7 @@ 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: 10.772263 ms
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>conv2d with tensor core: 6.685200 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 aa51fd460d..03c4eb6c55 100644
--- a/docs/how_to/optimize_operators/opt_gemm.html
+++ b/docs/how_to/optimize_operators/opt_gemm.html
@@ -472,8 +472,8 @@ 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: 0.018628
-Baseline: 3.397229
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Numpy running time: 0.018060
+Baseline: 3.197752
</pre></div>
</div>
<p>In TVM, we can always inspect lower level IR to debug or optimize our schedule.
@@ -484,20 +484,17 @@ Here is the generated IR using our baseline schedule.</p>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 1024), "float32"), B: T.Buffer((1024, 1024), "float32"), C: T.Buffer((1024, 1024), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 1024))
- B_1 = T.match_buffer(B, (1024, 1024))
- C_1 = T.match_buffer(C, (1024, 1024))
for m, n in T.grid(1024, 1024):
- C_2 = T.buffer_decl((1048576,), data=C_1.data)
- C_2[m * 1024 + n] = T.float32(0)
+ C_1 = T.buffer_decl((1048576,), data=C.data)
+ C_1[m * 1024 + n] = T.float32(0)
for k in range(1024):
cse_var_2: T.int32 = m * 1024
cse_var_1: T.int32 = cse_var_2 + n
- A_2 = T.buffer_decl((1048576,), data=A_1.data)
- B_2 = T.buffer_decl((1048576,), data=B_1.data)
- C_2[cse_var_1] = C_2[cse_var_1] + A_2[cse_var_2 + k] * B_2[k * 1024 + n]
+ A_1 = T.buffer_decl((1048576,), data=A.data)
+ B_1 = T.buffer_decl((1048576,), data=B.data)
+ C_1[cse_var_1] = C_1[cse_var_1] + A_1[cse_var_2 + k] * B_1[k * 1024 + n]
</pre></div>
</div>
</div>
@@ -532,7 +529,7 @@ 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: 0.295933
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Opt1: 0.296106
</pre></div>
</div>
<p>Here is the generated IR after blocking.</p>
@@ -542,22 +539,19 @@ fill 32 * 32 * sizeof(float) which is 4KB in the cache whose total size is 32KB
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 1024), "float32"), B: T.Buffer((1024, 1024), "float32"), C: T.Buffer((1024, 1024), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 1024))
- B_1 = T.match_buffer(B, (1024, 1024))
- C_1 = T.match_buffer(C, (1024, 1024))
for m_outer, n_outer in T.grid(32, 32):
- C_2 = T.buffer_decl((1048576,), data=C_1.data)
+ C_1 = T.buffer_decl((1048576,), data=C.data)
for m_inner_init, n_inner_init in T.grid(32, 32):
- C_2[m_outer * 32768 + m_inner_init * 1024 + n_outer * 32 + n_inner_init] = T.float32(0)
+ C_1[m_outer * 32768 + m_inner_init * 1024 + n_outer * 32 + n_inner_init] = T.float32(0)
for k_outer, k_inner, m_inner, n_inner in T.grid(256, 4, 32, 32):
cse_var_3: T.int32 = n_outer * 32
cse_var_2: T.int32 = m_outer * 32768 + m_inner * 1024
cse_var_1: T.int32 = cse_var_2 + cse_var_3 + n_inner
- A_2 = T.buffer_decl((1048576,), data=A_1.data)
- B_2 = T.buffer_decl((1048576,), data=B_1.data)
- C_2[cse_var_1] = C_2[cse_var_1] + A_2[cse_var_2 + k_outer * 4 + k_inner] * B_2[k_outer * 4096 + k_inner * 1024 + cse_var_3 + n_inner]
+ A_1 = T.buffer_decl((1048576,), data=A.data)
+ B_1 = T.buffer_decl((1048576,), data=B.data)
+ C_1[cse_var_1] = C_1[cse_var_1] + A_1[cse_var_2 + k_outer * 4 + k_inner] * B_1[k_outer * 4096 + k_inner * 1024 + cse_var_3 + n_inner]
</pre></div>
</div>
</div>
@@ -589,7 +583,7 @@ 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: 0.332504
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Opt2: 0.333387
</pre></div>
</div>
<p>Here is the generated IR after vectorization.</p>
@@ -599,22 +593,19 @@ vastly.</p>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 1024), "float32"), B: T.Buffer((1024, 1024), "float32"), C: T.Buffer((1024, 1024), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 1024))
- B_1 = T.match_buffer(B, (1024, 1024))
- C_1 = T.match_buffer(C, (1024, 1024))
for m_outer, n_outer in T.grid(32, 32):
- C_2 = T.buffer_decl((1048576,), data=C_1.data)
+ C_1 = T.buffer_decl((1048576,), data=C.data)
for m_inner_init in range(32):
- C_2[m_outer * 32768 + m_inner_init * 1024 + n_outer * 32:m_outer * 32768 + m_inner_init * 1024 + n_outer * 32 + 32] = T.Broadcast(T.float32(0), 32)
+ C_1[m_outer * 32768 + m_inner_init * 1024 + n_outer * 32:m_outer * 32768 + m_inner_init * 1024 + n_outer * 32 + 32] = T.Broadcast(T.float32(0), 32)
for k_outer, k_inner, m_inner in T.grid(256, 4, 32):
cse_var_3: T.int32 = n_outer * 32
cse_var_2: T.int32 = m_outer * 32768 + m_inner * 1024
cse_var_1: T.int32 = cse_var_2 + cse_var_3
- A_2 = T.buffer_decl((1048576,), data=A_1.data)
- B_2 = T.buffer_decl((1048576,), data=B_1.data)
- C_2[cse_var_1:cse_var_1 + 32] = C_2[cse_var_1:cse_var_1 + 32] + T.Broadcast(A_2[cse_var_2 + k_outer * 4 + k_inner], 32) * B_2[k_outer * 4096 + k_inner * 1024 + cse_var_3:k_outer * 4096 + k_inner * 1024 + cse_var_3 + 32]
+ A_1 = T.buffer_decl((1048576,), data=A.data)
+ B_1 = T.buffer_decl((1048576,), data=B.data)
+ C_1[cse_var_1:cse_var_1 + 32] = C_1[cse_var_1:cse_var_1 + 32] + T.Broadcast(A_1[cse_var_2 + k_outer * 4 + k_inner], 32) * B_1[k_outer * 4096 + k_inner * 1024 + cse_var_3:k_outer * 4096 + k_inner * 1024 + cse_var_3 + 32]
</pre></div>
</div>
</div>
@@ -644,7 +635,7 @@ 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: 0.116739
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Opt3: 0.114968
</pre></div>
</div>
<p>Here is the generated IR after loop permutation.</p>
@@ -654,22 +645,19 @@ the access pattern for A matrix is more cache friendly.</p>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 1024), "float32"), B: T.Buffer((1024, 1024), "float32"), C: T.Buffer((1024, 1024), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 1024))
- B_1 = T.match_buffer(B, (1024, 1024))
- C_1 = T.match_buffer(C, (1024, 1024))
for m_outer, n_outer in T.grid(32, 32):
- C_2 = T.buffer_decl((1048576,), data=C_1.data)
+ C_1 = T.buffer_decl((1048576,), data=C.data)
for m_inner_init in range(32):
- C_2[m_outer * 32768 + m_inner_init * 1024 + n_outer * 32:m_outer * 32768 + m_inner_init * 1024 + n_outer * 32 + 32] = T.Broadcast(T.float32(0), 32)
+ C_1[m_outer * 32768 + m_inner_init * 1024 + n_outer * 32:m_outer * 32768 + m_inner_init * 1024 + n_outer * 32 + 32] = T.Broadcast(T.float32(0), 32)
for k_outer, m_inner, k_inner in T.grid(256, 32, 4):
cse_var_3: T.int32 = n_outer * 32
cse_var_2: T.int32 = m_outer * 32768 + m_inner * 1024
cse_var_1: T.int32 = cse_var_2 + cse_var_3
- A_2 = T.buffer_decl((1048576,), data=A_1.data)
- B_2 = T.buffer_decl((1048576,), data=B_1.data)
- C_2[cse_var_1:cse_var_1 + 32] = C_2[cse_var_1:cse_var_1 + 32] + T.Broadcast(A_2[cse_var_2 + k_outer * 4 + k_inner], 32) * B_2[k_outer * 4096 + k_inner * 1024 + cse_var_3:k_outer * 4096 + k_inner * 1024 + cse_var_3 + 32]
+ A_1 = T.buffer_decl((1048576,), data=A.data)
+ B_1 = T.buffer_decl((1048576,), data=B.data)
+ C_1[cse_var_1:cse_var_1 + 32] = C_1[cse_var_1:cse_var_1 + 32] + T.Broadcast(A_1[cse_var_2 + k_outer * 4 + k_inner], 32) * B_1[k_outer * 4096 + k_inner * 1024 + cse_var_3:k_outer * 4096 + k_inner * 1024 + cse_var_3 + 32]
</pre></div>
</div>
</div>
@@ -721,7 +709,7 @@ 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: 0.109848
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Opt4: 0.108361
</pre></div>
</div>
<p>Here is the generated IR after array packing.</p>
@@ -731,27 +719,24 @@ flattening.</p>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 1024), "float32"), B: T.Buffer((1024, 1024), "float32"), C: T.Buffer((1024, 1024), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 1024))
- B_1 = T.match_buffer(B, (1024, 1024))
- C_1 = T.match_buffer(C, (1024, 1024))
packedB = T.allocate([32768], "float32x32", "global")
packedB_1 = T.buffer_decl((32768,), "float32x32", data=packedB)
for bigN in T.parallel(32):
for k in range(1024):
- B_2 = T.buffer_decl((1048576,), data=B_1.data)
- packedB_1[bigN * 1024 + k] = B_2[k * 1024 + bigN * 32:k * 1024 + bigN * 32 + 32]
+ B_1 = T.buffer_decl((1048576,), data=B.data)
+ packedB_1[bigN * 1024 + k] = B_1[k * 1024 + bigN * 32:k * 1024 + bigN * 32 + 32]
for m_outer, n_outer in T.grid(32, 32):
- C_2 = T.buffer_decl((1048576,), data=C_1.data)
+ C_1 = T.buffer_decl((1048576,), data=C.data)
for m_inner_init in range(32):
- C_2[m_outer * 32768 + m_inner_init * 1024 + n_outer * 32:m_outer * 32768 + m_inner_init * 1024 + n_outer * 32 + 32] = T.Broadcast(T.float32(0), 32)
+ C_1[m_outer * 32768 + m_inner_init * 1024 + n_outer * 32:m_outer * 32768 + m_inner_init * 1024 + n_outer * 32 + 32] = T.Broadcast(T.float32(0), 32)
for k_outer, m_inner, k_inner in T.grid(256, 32, 4):
cse_var_3: T.int32 = m_outer * 32768 + m_inner * 1024
cse_var_2: T.int32 = k_outer * 4
cse_var_1: T.int32 = cse_var_3 + n_outer * 32
- A_2 = T.buffer_decl((1048576,), data=A_1.data)
- C_2[cse_var_1:cse_var_1 + 32] = C_2[cse_var_1:cse_var_1 + 32] + T.Broadcast(A_2[cse_var_3 + cse_var_2 + k_inner], 32) * packedB_1[n_outer * 1024 + cse_var_2 + k_inner]
+ A_1 = T.buffer_decl((1048576,), data=A.data)
+ C_1[cse_var_1:cse_var_1 + 32] = C_1[cse_var_1:cse_var_1 + 32] + T.Broadcast(A_1[cse_var_3 + cse_var_2 + k_inner], 32) * packedB_1[n_outer * 1024 + cse_var_2 + k_inner]
</pre></div>
</div>
</div>
@@ -799,7 +784,7 @@ 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: 0.112016
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Opt5: 0.111469
</pre></div>
</div>
<p>Here is the generated IR after blocking.</p>
@@ -809,18 +794,15 @@ write to C when all the block results are ready.</p>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 1024), "float32"), B: T.Buffer((1024, 1024), "float32"), C: T.Buffer((1024, 1024), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 1024))
- B_1 = T.match_buffer(B, (1024, 1024))
- C_1 = T.match_buffer(C, (1024, 1024))
packedB = T.allocate([32768], "float32x32", "global")
C_global = T.allocate([1024], "float32", "global")
packedB_1 = T.buffer_decl((32768,), "float32x32", data=packedB)
for bigN in T.parallel(32):
for k in range(1024):
- B_2 = T.buffer_decl((1048576,), data=B_1.data)
- packedB_1[bigN * 1024 + k] = B_2[k * 1024 + bigN * 32:k * 1024 + bigN * 32 + 32]
+ B_1 = T.buffer_decl((1048576,), data=B.data)
+ packedB_1[bigN * 1024 + k] = B_1[k * 1024 + bigN * 32:k * 1024 + bigN * 32 + 32]
for m_outer, n_outer in T.grid(32, 32):
C_global_1 = T.buffer_decl((1024,), data=C_global)
for m_c_init in range(32):
@@ -830,14 +812,14 @@ class Module:
cse_var_3: T.int32 = m_c * 32
cse_var_2: T.int32 = n_outer * 1024 + cse_var_4
cse_var_1: T.int32 = m_outer * 32768 + m_c * 1024 + cse_var_4
- A_2 = T.buffer_decl((1048576,), data=A_1.data)
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1], 32) * packedB_1[cse_var_2]
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1 + 1], 32) * packedB_1[cse_var_2 + 1]
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1 + 2], 32) * packedB_1[cse_var_2 + 2]
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1 + 3], 32) * packedB_1[cse_var_2 + 3]
+ A_1 = T.buffer_decl((1048576,), data=A.data)
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1], 32) * packedB_1[cse_var_2]
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1 + 1], 32) * packedB_1[cse_var_2 + 1]
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1 + 2], 32) * packedB_1[cse_var_2 + 2]
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1 + 3], 32) * packedB_1[cse_var_2 + 3]
for m_inner, n_inner in T.grid(32, 32):
- C_2 = T.buffer_decl((1048576,), data=C_1.data)
- C_2[m_outer * 32768 + m_inner * 1024 + n_outer * 32 + n_inner] = C_global_1[m_inner * 32 + n_inner]
+ C_1 = T.buffer_decl((1048576,), data=C.data)
+ C_1[m_outer * 32768 + m_inner * 1024 + n_outer * 32 + n_inner] = C_global_1[m_inner * 32 + n_inner]
</pre></div>
</div>
</div>
@@ -879,7 +861,7 @@ class Module:
<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: 0.147469
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Opt6: 0.146623
</pre></div>
</div>
<p>Here is the generated IR after parallelization.</p>
@@ -889,17 +871,14 @@ class Module:
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 1024), "float32"), B: T.Buffer((1024, 1024), "float32"), C: T.Buffer((1024, 1024), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 1024))
- B_1 = T.match_buffer(B, (1024, 1024))
- C_1 = T.match_buffer(C, (1024, 1024))
packedB = T.allocate([32768], "float32x32", "global")
packedB_1 = T.buffer_decl((32768,), "float32x32", data=packedB)
for bigN in T.parallel(32):
for k in range(1024):
- B_2 = T.buffer_decl((1048576,), data=B_1.data)
- packedB_1[bigN * 1024 + k] = B_2[k * 1024 + bigN * 32:k * 1024 + bigN * 32 + 32]
+ B_1 = T.buffer_decl((1048576,), data=B.data)
+ packedB_1[bigN * 1024 + k] = B_1[k * 1024 + bigN * 32:k * 1024 + bigN * 32 + 32]
for m_outer in T.parallel(32):
C_global = T.allocate([1024], "float32", "global")
for n_outer in range(32):
@@ -911,14 +890,14 @@ class Module:
cse_var_3: T.int32 = m_c * 32
cse_var_2: T.int32 = n_outer * 1024 + cse_var_4
cse_var_1: T.int32 = m_outer * 32768 + m_c * 1024 + cse_var_4
- A_2 = T.buffer_decl((1048576,), data=A_1.data)
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1], 32) * packedB_1[cse_var_2]
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1 + 1], 32) * packedB_1[cse_var_2 + 1]
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1 + 2], 32) * packedB_1[cse_var_2 + 2]
- C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_2[cse_var_1 + 3], 32) * packedB_1[cse_var_2 + 3]
+ A_1 = T.buffer_decl((1048576,), data=A.data)
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1], 32) * packedB_1[cse_var_2]
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1 + 1], 32) * packedB_1[cse_var_2 + 1]
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1 + 2], 32) * packedB_1[cse_var_2 + 2]
+ C_global_1[cse_var_3:cse_var_3 + 32] = C_global_1[cse_var_3:cse_var_3 + 32] + T.Broadcast(A_1[cse_var_1 + 3], 32) * packedB_1[cse_var_2 + 3]
for m_inner, n_inner in T.grid(32, 32):
- C_2 = T.buffer_decl((1048576,), data=C_1.data)
- C_2[m_outer * 32768 + m_inner * 1024 + n_outer * 32 + n_inner] = C_global_1[m_inner * 32 + n_inner]
+ C_1 = T.buffer_decl((1048576,), data=C.data)
+ C_1[m_outer * 32768 + m_inner * 1024 + n_outer * 32 + n_inner] = C_global_1[m_inner * 32 + n_inner]
</pre></div>
</div>
</div>
diff --git a/docs/how_to/optimize_operators/sg_execution_times.html b/docs/how_to/optimize_operators/sg_execution_times.html
index 2fa03d94fa..cdf500a5be 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>00:34.839</strong> total execution time for <strong>how_to_optimize_operators</strong> files:</p>
+<p><strong>00:33.932</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>00:32.242</p></td>
+<td><p>00:31.481</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>00:01.513</p></td>
+<td><p>00:01.426</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>00:01.085</p></td>
+<td><p>00:01.025</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 b50fe393b9..e364ff44c8 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>09:28.448</strong> total execution time for <strong>how_to_tune_with_autoscheduler</strong> files:</p>
+<p><strong>09:06.030</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>05:46.341</p></td>
+<td><p>05:27.031</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>01:38.610</p></td>
+<td><p>01:36.571</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>01:05.810</p></td>
+<td><p>01:04.732</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>00:31.182</p></td>
+<td><p>00:31.554</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>00:13.841</p></td>
+<td><p>00:13.565</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>00:12.664</p></td>
+<td><p>00:12.577</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 e6844cd787..a4f618dbfb 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
@@ -500,19 +500,15 @@ cooperative fetching, unrolling and operator fusion.</p>
@I.ir_module
class Module:
@T.prim_func
- def main(data: T.handle, kernel: T.handle, bias: T.handle, compute: T.handle):
+ def main(data: T.Buffer((1, 512, 7, 7), "float32"), kernel: T.Buffer((512, 512, 3, 3), "float32"), bias: T.Buffer((1, 512, 1, 1), "float32"), compute: T.Buffer((1, 512, 7, 7), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- data_1 = T.match_buffer(data, (1, 512, 7, 7))
- kernel_1 = T.match_buffer(kernel, (512, 512, 3, 3))
- bias_1 = T.match_buffer(bias, (1, 512, 1, 1))
- compute_1 = T.match_buffer(compute, (1, 512, 7, 7))
blockIdx_x = T.env_thread("blockIdx.x")
- T.launch_thread(blockIdx_x, 16)
+ T.launch_thread(blockIdx_x, 28)
conv2d_nchw = T.allocate([14], "float32", "local")
- pad_temp_shared = T.allocate([2592], "float32", "shared")
- kernel_shared = T.allocate([9216], "float32", "shared")
+ pad_temp_shared = T.allocate([72], "float32", "shared")
+ kernel_shared = T.allocate([3072], "float32", "shared")
threadIdx_x = T.env_thread("threadIdx.x")
- T.launch_thread(threadIdx_x, 112)
+ T.launch_thread(threadIdx_x, 64)
conv2d_nchw_1 = T.buffer_decl((14,), data=conv2d_nchw, scope="local", align=32)
conv2d_nchw_1[0] = T.float32(0)
conv2d_nchw_1[1] = T.float32(0)
@@ -528,146 +524,460 @@ class Module:
conv2d_nchw_1[11] = T.float32(0)
conv2d_nchw_1[12] = T.float32(0)
conv2d_nchw_1[13] = T.float32(0)
- for rc_outer_outer in range(16):
- cse_var_1: T.int32 = rc_outer_outer * 1568
+ for rc_outer_outer, ry_outer_outer in T.grid(64, 3):
+ cse_var_2: T.int32 = rc_outer_outer * 72
+ cse_var_1: T.int32 = ry_outer_outer * 3
threadIdx_x_1 = T.env_thread("threadIdx.x")
- pad_temp_shared_1 = T.buffer_decl((2592,), data=pad_temp_shared, scope="shared")
- data_2 = T.buffer_decl((25088,), data=data_1.data)
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1] = T.if_then_else(9 <= threadIdx_x_1 % 81 and threadIdx_x_1 % 81 < 72 and 1 <= threadIdx_x_1 % 9 and threadIdx_x_1 % 9 < 8, data_2[cse_var_1 + threadIdx_x_1 // 81 * 49 + threadIdx_x_1 % 81 // 9 * 7 + threadIdx_x_1 % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 112] = T.if_then_else(9 <= (threadIdx_x_1 + 31) % 81 and (threadIdx_x_1 + 31) % 81 < 72 and 1 <= (threadIdx_x_1 + 4) % 9 and (threadIdx_x_1 + 4) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 112) // 81 * 49 + (threadIdx_x_1 + 31) % 81 // 9 * 7 + (threadIdx_x_1 + 4) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 224] = T.if_then_else(9 <= (threadIdx_x_1 + 62) % 81 and (threadIdx_x_1 + 62) % 81 < 72 and 1 <= (threadIdx_x_1 + 8) % 9 and (threadIdx_x_1 + 8) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 224) // 81 * 49 + (threadIdx_x_1 + 62) % 81 // 9 * 7 + (threadIdx_x_1 + 8) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 336] = T.if_then_else(9 <= (threadIdx_x_1 + 12) % 81 and (threadIdx_x_1 + 12) % 81 < 72 and 1 <= (threadIdx_x_1 + 3) % 9 and (threadIdx_x_1 + 3) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 336) // 81 * 49 + (threadIdx_x_1 + 12) % 81 // 9 * 7 + (threadIdx_x_1 + 3) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 448] = T.if_then_else(9 <= (threadIdx_x_1 + 43) % 81 and (threadIdx_x_1 + 43) % 81 < 72 and 1 <= (threadIdx_x_1 + 7) % 9 and (threadIdx_x_1 + 7) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 448) // 81 * 49 + (threadIdx_x_1 + 43) % 81 // 9 * 7 + (threadIdx_x_1 + 7) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 560] = T.if_then_else(9 <= (threadIdx_x_1 + 74) % 81 and (threadIdx_x_1 + 74) % 81 < 72 and 1 <= (threadIdx_x_1 + 2) % 9 and (threadIdx_x_1 + 2) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 560) // 81 * 49 + (threadIdx_x_1 + 74) % 81 // 9 * 7 + (threadIdx_x_1 + 2) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 672] = T.if_then_else(9 <= (threadIdx_x_1 + 24) % 81 and (threadIdx_x_1 + 24) % 81 < 72 and 1 <= (threadIdx_x_1 + 6) % 9 and (threadIdx_x_1 + 6) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 672) // 81 * 49 + (threadIdx_x_1 + 24) % 81 // 9 * 7 + (threadIdx_x_1 + 6) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 784] = T.if_then_else(9 <= (threadIdx_x_1 + 55) % 81 and (threadIdx_x_1 + 55) % 81 < 72 and 1 <= (threadIdx_x_1 + 1) % 9 and (threadIdx_x_1 + 1) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 784) // 81 * 49 + (threadIdx_x_1 + 55) % 81 // 9 * 7 + (threadIdx_x_1 + 1) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 896] = T.if_then_else(9 <= (threadIdx_x_1 + 5) % 81 and (threadIdx_x_1 + 5) % 81 < 72 and 1 <= (threadIdx_x_1 + 5) % 9 and (threadIdx_x_1 + 5) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 896) // 81 * 49 + (threadIdx_x_1 + 5) % 81 // 9 * 7 + (threadIdx_x_1 + 5) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 1008] = T.if_then_else(1 <= (threadIdx_x_1 // 9 + 4) % 9 and (threadIdx_x_1 + 36) % 81 < 72 and 1 <= threadIdx_x_1 % 9 and threadIdx_x_1 % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 1008) // 81 * 49 + (threadIdx_x_1 // 9 + 4) % 9 * 7 + threadIdx_x_1 % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 1120] = T.if_then_else(9 <= (threadIdx_x_1 + 67) % 81 and (threadIdx_x_1 + 67) % 81 < 72 and 1 <= (threadIdx_x_1 + 4) % 9 and (threadIdx_x_1 + 4) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 1120) // 81 * 49 + (threadIdx_x_1 + 67) % 81 // 9 * 7 + (threadIdx_x_1 + 4) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 1232] = T.if_then_else(9 <= (threadIdx_x_1 + 17) % 81 and (threadIdx_x_1 + 17) % 81 < 72 and 1 <= (threadIdx_x_1 + 8) % 9 and (threadIdx_x_1 + 8) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 1232) // 81 * 49 + (threadIdx_x_1 + 17) % 81 // 9 * 7 + (threadIdx_x_1 + 8) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 1344] = T.if_then_else(9 <= (threadIdx_x_1 + 48) % 81 and (threadIdx_x_1 + 48) % 81 < 72 and 1 <= (threadIdx_x_1 + 3) % 9 and (threadIdx_x_1 + 3) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 1344) // 81 * 49 + (threadIdx_x_1 + 48) % 81 // 9 * 7 + (threadIdx_x_1 + 3) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 1456] = T.if_then_else(9 <= (threadIdx_x_1 + 79) % 81 and (threadIdx_x_1 + 79) % 81 < 72 and 1 <= (threadIdx_x_1 + 7) % 9 and (threadIdx_x_1 + 7) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 1456) // 81 * 49 + (threadIdx_x_1 + 79) % 81 // 9 * 7 + (threadIdx_x_1 + 7) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 1568] = T.if_then_else(9 <= (threadIdx_x_1 + 29) % 81 and (threadIdx_x_1 + 29) % 81 < 72 and 1 <= (threadIdx_x_1 + 2) % 9 and (threadIdx_x_1 + 2) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 1568) // 81 * 49 + (threadIdx_x_1 + 29) % 81 // 9 * 7 + (threadIdx_x_1 + 2) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 1680] = T.if_then_else(9 <= (threadIdx_x_1 + 60) % 81 and (threadIdx_x_1 + 60) % 81 < 72 and 1 <= (threadIdx_x_1 + 6) % 9 and (threadIdx_x_1 + 6) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 1680) // 81 * 49 + (threadIdx_x_1 + 60) % 81 // 9 * 7 + (threadIdx_x_1 + 6) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 1792] = T.if_then_else(9 <= (threadIdx_x_1 + 10) % 81 and (threadIdx_x_1 + 10) % 81 < 72 and 1 <= (threadIdx_x_1 + 1) % 9 and (threadIdx_x_1 + 1) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 1792) // 81 * 49 + (threadIdx_x_1 + 10) % 81 // 9 * 7 + (threadIdx_x_1 + 1) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 1904] = T.if_then_else(9 <= (threadIdx_x_1 + 41) % 81 and (threadIdx_x_1 + 41) % 81 < 72 and 1 <= (threadIdx_x_1 + 5) % 9 and (threadIdx_x_1 + 5) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 1904) // 81 * 49 + (threadIdx_x_1 + 41) % 81 // 9 * 7 + (threadIdx_x_1 + 5) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 2016] = T.if_then_else(1 <= (threadIdx_x_1 // 9 + 8) % 9 and (threadIdx_x_1 + 72) % 81 < 72 and 1 <= threadIdx_x_1 % 9 and threadIdx_x_1 % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 2016) // 81 * 49 + (threadIdx_x_1 // 9 + 8) % 9 * 7 + threadIdx_x_1 % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 2128] = T.if_then_else(9 <= (threadIdx_x_1 + 22) % 81 and (threadIdx_x_1 + 22) % 81 < 72 and 1 <= (threadIdx_x_1 + 4) % 9 and (threadIdx_x_1 + 4) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 2128) // 81 * 49 + (threadIdx_x_1 + 22) % 81 // 9 * 7 + (threadIdx_x_1 + 4) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 2240] = T.if_then_else(9 <= (threadIdx_x_1 + 53) % 81 and (threadIdx_x_1 + 53) % 81 < 72 and 1 <= (threadIdx_x_1 + 8) % 9 and (threadIdx_x_1 + 8) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 2240) // 81 * 49 + (threadIdx_x_1 + 53) % 81 // 9 * 7 + (threadIdx_x_1 + 8) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 2352] = T.if_then_else(9 <= (threadIdx_x_1 + 3) % 81 and (threadIdx_x_1 + 3) % 81 < 72 and 1 <= (threadIdx_x_1 + 3) % 9 and (threadIdx_x_1 + 3) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 2352) // 81 * 49 + (threadIdx_x_1 + 3) % 81 // 9 * 7 + (threadIdx_x_1 + 3) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- pad_temp_shared_1[threadIdx_x_1 + 2464] = T.if_then_else(9 <= (threadIdx_x_1 + 34) % 81 and (threadIdx_x_1 + 34) % 81 < 72 and 1 <= (threadIdx_x_1 + 7) % 9 and (threadIdx_x_1 + 7) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 2464) // 81 * 49 + (threadIdx_x_1 + 34) % 81 // 9 * 7 + (threadIdx_x_1 + 7) % 9 - 8], T.float32(0))
- with T.launch_thread(threadIdx_x_1, 112):
- if T.likely(threadIdx_x_1 < 16):
- pad_temp_shared_1[threadIdx_x_1 + 2576] = T.if_then_else(threadIdx_x_1 < 7 and 1 <= (threadIdx_x_1 + 2) % 9 and (threadIdx_x_1 + 2) % 9 < 8, data_2[cse_var_1 + (threadIdx_x_1 + 2576) // 81 * 49 + (threadIdx_x_1 + 65) % 81 // 9 * 7 + (threadIdx_x_1 + 2) - 8], T.float32(0))
- kernel_shared_1 = T.buffer_decl((9216,), data=kernel_shared, scope="shared")
- for ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer in range(4):
- threadIdx_x_2 = T.env_thread("threadIdx.x")
- T.launch_thread(threadIdx_x_2, 112)
- kernel_2 = T.buffer_decl((2359296,), data=kernel_1.data)
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 32 + threadIdx_x_2 * 8) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2 + threadIdx_x_2 * 2) % 3 * 3]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 1] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 32 + threadIdx_x_2 * 8) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2 + threadIdx_x_2 * 2) % 3 * 3 + 1]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 2] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 32 + threadIdx_x_2 * 8) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2 + threadIdx_x_2 * 2) % 3 * 3 + 2]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 3] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 1) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 1) % 3 * 3]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 4] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 1) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 1) % 3 * 3 + 1]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 5] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 1) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 1) % 3 * 3 + 2]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 6] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 2) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 2) % 3 * 3]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 7] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 2) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 2) % 3 * 3 + 1]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 8] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 2) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 2) % 3 * 3 + 2]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 9] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + ((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8) // 3 + 1) % 32 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2 + threadIdx_x_2 * 2) % 3 * 3]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 10] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + ((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8) // 3 + 1) % 32 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2 + threadIdx_x_2 * 2) % 3 * 3 + 1]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 11] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + ((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8) // 3 + 1) % 32 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2 + threadIdx_x_2 * 2) % 3 * 3 + 2]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 12] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 4) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 1) % 3 * 3]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 13] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 4) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 1) % 3 * 3 + 1]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 14] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 4) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 1) % 3 * 3 + 2]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 15] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 5) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 2) % 3 * 3]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 16] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 5) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 2) % 3 * 3 + 1]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 17] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 5) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 2) % 3 * 3 + 2]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 18] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + ((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8) // 3 + 2) % 32 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2 + threadIdx_x_2 * 2) % 3 * 3]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 19] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + ((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8) // 3 + 2) % 32 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2 + threadIdx_x_2 * 2) % 3 * 3 + 1]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 20] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + ((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8) // 3 + 2) % 32 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2 + threadIdx_x_2 * 2) % 3 * 3 + 2]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 21] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 7) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 1) % 3 * 3]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 22] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 7) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 1) % 3 * 3 + 1]
- if T.likely(ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7 + threadIdx_x_2 // 16 < 24):
- kernel_shared_1[ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688 + threadIdx_x_2 * 24 + 23] = kernel_2[blockIdx_x * 147456 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28 + threadIdx_x_2 // 4) // 3 * 4608 + rc_outer_outer * 288 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896 + threadIdx_x_2 * 8 + 7) % 96 // 3 * 9 + (ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224 + threadIdx_x_2 * 2 + 1) % 3 * 3 + 2]
- for rc_outer_inner, rx_outer_inner, ff_outer_inner, rc_inner in T.grid(2, 3, 2, 16):
- cse_var_8: T.int32 = ff_outer_inner * 7
- cse_var_7: T.int32 = cse_var_8 + 6
- cse_var_6: T.int32 = cse_var_8 + 5
- cse_var_5: T.int32 = cse_var_8 + 4
- cse_var_4: T.int32 = cse_var_8 + 3
- cse_var_3: T.int32 = cse_var_8 + 2
- cse_var_2: T.int32 = cse_var_8 + 1
- conv2d_nchw_1[cse_var_8] = conv2d_nchw_1[cse_var_8] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner]
- conv2d_nchw_1[cse_var_2] = conv2d_nchw_1[cse_var_2] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 9] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner]
- conv2d_nchw_1[cse_var_3] = conv2d_nchw_1[cse_var_3] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 18] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner]
- conv2d_nchw_1[cse_var_4] = conv2d_nchw_1[cse_var_4] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 27] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner]
- conv2d_nchw_1[cse_var_5] = conv2d_nchw_1[cse_var_5] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 36] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner]
- conv2d_nchw_1[cse_var_6] = conv2d_nchw_1[cse_var_6] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 45] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner]
- conv2d_nchw_1[cse_var_7] = conv2d_nchw_1[cse_var_7] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 54] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner]
- conv2d_nchw_1[cse_var_8] = conv2d_nchw_1[cse_var_8] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 9] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 3]
- conv2d_nchw_1[cse_var_2] = conv2d_nchw_1[cse_var_2] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 18] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 3]
- conv2d_nchw_1[cse_var_3] = conv2d_nchw_1[cse_var_3] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 27] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 3]
- conv2d_nchw_1[cse_var_4] = conv2d_nchw_1[cse_var_4] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 36] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 3]
- conv2d_nchw_1[cse_var_5] = conv2d_nchw_1[cse_var_5] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 45] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 3]
- conv2d_nchw_1[cse_var_6] = conv2d_nchw_1[cse_var_6] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 54] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 3]
- conv2d_nchw_1[cse_var_7] = conv2d_nchw_1[cse_var_7] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 63] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 3]
- conv2d_nchw_1[cse_var_8] = conv2d_nchw_1[cse_var_8] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 18] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 6]
- conv2d_nchw_1[cse_var_2] = conv2d_nchw_1[cse_var_2] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 27] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 6]
- conv2d_nchw_1[cse_var_3] = conv2d_nchw_1[cse_var_3] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 36] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 6]
- conv2d_nchw_1[cse_var_4] = conv2d_nchw_1[cse_var_4] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 45] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 6]
- conv2d_nchw_1[cse_var_5] = conv2d_nchw_1[cse_var_5] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 54] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 6]
- conv2d_nchw_1[cse_var_6] = conv2d_nchw_1[cse_var_6] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 63] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 6]
- conv2d_nchw_1[cse_var_7] = conv2d_nchw_1[cse_var_7] + pad_temp_shared_1[rc_outer_inner * 1296 + rc_inner * 81 + rx_outer_inner + threadIdx_x % 7 + 72] * kernel_shared_1[threadIdx_x // 7 * 576 + ff_outer_inner * 288 + rc_outer_inner * 144 + rc_inner * 9 + rx_outer_inner + 6]
- for i1_inner, i2_inner in T.grid(2, 7):
- compute_2 = T.buffer_decl((25088,), data=compute_1.data)
- bias_2 = T.buffer_decl((512,), data=bias_1.data)
- compute_2[blockIdx_x * 1568 + threadIdx_x // 7 * 98 + i1_inner * 49 + i2_inner * 7 + threadIdx_x % 7] = T.max(conv2d_nchw_1[i1_inner * 7 + i2_inner] + bias_2[blockIdx_x * 32 + threadIdx_x // 7 * 2 + i1_inner], T.float32(0))
+ pad_temp_shared_1 = T.buffer_decl((72,), data=pad_temp_shared, scope="shared")
+ with T.launch_thread(threadIdx_x_1, 64):
+ data_1 = T.buffer_decl((25088,), data=data.data)
+ if T.likely(threadIdx_x_1 < 18):
+ pad_temp_shared_1[threadIdx_x_1 * 4] = T.if_then_else(1 <= ry_outer_outer + blockIdx_x % 7 and ry_outer_outer + blockIdx_x % 7 < 8 and 1 <= threadIdx_x_1 * 4 % 9 and threadIdx_x_1 * 4 % 9 < 8, data_1[rc_outer_outer * 392 + threadIdx_x_1 * 4 // 9 * 49 + ry_outer_outer * 7 + blockIdx_x % 7 * 7 + threadIdx_x_1 * 4 % 9 - 8], T.float32(0))
+ if T.likely(threadIdx_x_1 < 18):
+ pad_temp_shared_1[threadIdx_x_1 * 4 + 1] = T.if_then_else(1 <= ry_outer_outer + blockIdx_x % 7 and ry_outer_outer + blockIdx_x % 7 < 8 and 1 <= (threadIdx_x_1 * 4 + 1) % 9 and (threadIdx_x_1 * 4 + 1) % 9 < 8, data_1[rc_outer_outer * 392 + (threadIdx_x_1 * 4 + 1) // 9 * 49 + ry_outer_outer * 7 + blockIdx_x % 7 * 7 + (threadIdx_x_1 * 4 + 1) % 9 - 8], T.float32(0))
+ if T.likely(threadIdx_x_1 < 18):
+ pad_temp_shared_1[threadIdx_x_1 * 4 + 2] = T.if_then_else(1 <= ry_outer_outer + blockIdx_x % 7 and ry_outer_outer + blockIdx_x % 7 < 8 and 1 <= (threadIdx_x_1 * 4 + 2) % 9 and (threadIdx_x_1 * 4 + 2) % 9 < 8, data_1[rc_outer_outer * 392 + (threadIdx_x_1 * 4 + 2) // 9 * 49 + ry_outer_outer * 7 + blockIdx_x % 7 * 7 + (threadIdx_x_1 * 4 + 2) % 9 - 8], T.float32(0))
+ if T.likely(threadIdx_x_1 < 18):
+ pad_temp_shared_1[threadIdx_x_1 * 4 + 3] = T.if_then_else(1 <= ry_outer_outer + blockIdx_x % 7 and ry_outer_outer + blockIdx_x % 7 < 8 and 1 <= (threadIdx_x_1 * 4 + 3) % 9 and (threadIdx_x_1 * 4 + 3) % 9 < 8, data_1[rc_outer_outer * 392 + (threadIdx_x_1 * 4 + 3) // 9 * 49 + ry_outer_outer * 7 + blockIdx_x % 7 * 7 + (threadIdx_x_1 * 4 + 3) % 9 - 8], T.float32(0))
+ threadIdx_x_2 = T.env_thread("threadIdx.x")
+ kernel_shared_1 = T.buffer_decl((3072,), data=kernel_shared, scope="shared")
+ kernel_1 = T.buffer_decl((2359296,), data=kernel.data)
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 64] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 64) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 128] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 128) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 192] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 36864]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 256] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 256) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 320] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 320) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 384] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 73728]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 448] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 448) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 512] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 512) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 576] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 110592]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 640] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 640) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 704] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 704) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 768] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 147456]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 832] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 832) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 896] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 896) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 960] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 184320]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1024] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 1024) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1088] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 1088) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1152] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 221184]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1216] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 1216) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1280] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 1280) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1344] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 258048]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1408] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 1408) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1472] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 1472) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1536] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 294912]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1600] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 1600) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1664] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 1664) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1728] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 331776]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1792] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 1792) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1856] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 1856) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1920] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 368640]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 1984] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 1984) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2048] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 2048) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2112] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 405504]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2176] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 2176) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2240] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 2240) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2304] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 442368]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2368] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 2368) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2432] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 2432) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2496] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 479232]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2560] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 2560) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2624] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 2624) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2688] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 516096]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2752] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 2752) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2816] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 2816) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 2) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2880] = kernel_1[blockIdx_x // 7 * 589824 + threadIdx_x_2 // 24 * 4608 + cse_var_2 + threadIdx_x_2 % 24 // 3 * 9 + cse_var_1 + threadIdx_x_2 % 3 + 552960]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 2944] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 2944) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 16) % 24 // 3 * 9 + cse_var_1 + (threadIdx_x_2 + 1) % 3]
+ with T.launch_thread(threadIdx_x_2, 64):
+ kernel_shared_1[threadIdx_x_2 + 3008] = kernel_1[blockIdx_x // 7 * 589824 + (threadIdx_x_2 + 3008) // 24 * 4608 + cse_var_2 + (threadIdx_x_2 + 8) % 24 // 3 * 9 + cse_var_1 + (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]
+ for i1_inner, i3_inner in T.grid(2, 7):
+ compute_1 = T.buffer_decl((25088,), data=compute.data)
+ bias_1 = T.buffer_decl((512,), data=bias.data)
+ compute_1[blockIdx_x // 7 * 6272 + threadIdx_x * 98 + i1_inner * 49 + blockIdx_x % 7 * 7 + i3_inner] = T.max(conv2d_nchw_1[i1_inner * 7 + i3_inner] + bias_1[blockIdx_x // 7 * 128 + threadIdx_x * 2 + i1_inner], T.float32(0))
</pre></div>
</div>
</div>
@@ -701,7 +1011,7 @@ class Module:
<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: 0.325 ms
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Execution time of this operator: 0.355 ms
</pre></div>
</div>
</div>
@@ -732,19 +1042,19 @@ conv2d_nchw_nn_o_o_o_i, conv2d_nchw_nn_o_o_i = s[conv2d_nchw].split(conv2d_nchw_
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=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=16)
+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=7)
+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=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=16)
-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=1)
conv2d_nchw_rx_o_o, conv2d_nchw_rx_o_i = s[conv2d_nchw].split(conv2d_nchw_rx_o_i, factor=3)
@@ -753,13 +1063,13 @@ 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=2)
-compute_i1_o_o_i, compute_i1_o_i = s[compute].split(compute_i1_o_i, factor=16)
+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=7)
+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=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)
@@ -777,16 +1087,16 @@ s[compute].bind(compute_i0_o_o_i_i1_o_o_i_fused_i2_o_o_i_fused_i3_o_o_i_fused, t
compute_i0_o_i_i1_o_i_fused_i2_o_i_fused_i3_o_i_fused = s[compute].fuse(compute_i0_o_i, compute_i1_o_i, compute_i2_o_i, compute_i3_o_i)
s[compute].bind(compute_i0_o_i_i1_o_i_fused_i2_o_i_fused_i3_o_i_fused, te.thread_axis("threadIdx.x"))
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=24)
+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=112)
+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=112)
+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", 64)
+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)
CUDA source code:
@@ -804,10 +1114,10 @@ CUDA source code:
#define int64_t long long
#define uint64_t unsigned long long
#endif
-extern "C" __global__ void __launch_bounds__(112) default_function_kernel0(float* __restrict__ data, float* __restrict__ kernel, float* __restrict__ compute, float* __restrict__ bias) {
+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[2592];
- __shared__ float kernel_shared[9216];
+ __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;
@@ -822,142 +1132,411 @@ extern "C" __global__ void __launch_bounds__(112) default_function_ker
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 < 16; ++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 * 1568) + ((((int)threadIdx.x) / 81) * 49)) + (((((int)threadIdx.x) % 81) / 9) * 7)) + (((int)threadIdx.x) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 112)] = (((((9 <= ((((int)threadIdx.x) + 31) % 81)) && (((((int)threadIdx.x) + 31) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 4) % 9))) && (((((int)threadIdx.x) + 4) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 112) / 81) * 49)) + ((((((int)threadIdx.x) + 31) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 4) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 224)] = (((((9 <= ((((int)threadIdx.x) + 62) % 81)) && (((((int)threadIdx.x) + 62) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 8) % 9))) && (((((int)threadIdx.x) + 8) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 224) / 81) * 49)) + ((((((int)threadIdx.x) + 62) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 8) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 336)] = (((((9 <= ((((int)threadIdx.x) + 12) % 81)) && (((((int)threadIdx.x) + 12) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 3) % 9))) && (((((int)threadIdx.x) + 3) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 336) / 81) * 49)) + ((((((int)threadIdx.x) + 12) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 3) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 448)] = (((((9 <= ((((int)threadIdx.x) + 43) % 81)) && (((((int)threadIdx.x) + 43) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 7) % 9))) && (((((int)threadIdx.x) + 7) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 448) / 81) * 49)) + ((((((int)threadIdx.x) + 43) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 7) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 560)] = (((((9 <= ((((int)threadIdx.x) + 74) % 81)) && (((((int)threadIdx.x) + 74) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 2) % 9))) && (((((int)threadIdx.x) + 2) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 560) / 81) * 49)) + ((((((int)threadIdx.x) + 74) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 2) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 672)] = (((((9 <= ((((int)threadIdx.x) + 24) % 81)) && (((((int)threadIdx.x) + 24) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 6) % 9))) && (((((int)threadIdx.x) + 6) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 672) / 81) * 49)) + ((((((int)threadIdx.x) + 24) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 6) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 784)] = (((((9 <= ((((int)threadIdx.x) + 55) % 81)) && (((((int)threadIdx.x) + 55) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 1) % 9))) && (((((int)threadIdx.x) + 1) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 784) / 81) * 49)) + ((((((int)threadIdx.x) + 55) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 1) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 896)] = (((((9 <= ((((int)threadIdx.x) + 5) % 81)) && (((((int)threadIdx.x) + 5) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 5) % 9))) && (((((int)threadIdx.x) + 5) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 896) / 81) * 49)) + ((((((int)threadIdx.x) + 5) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 5) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 1008)] = (((((1 <= (((((int)threadIdx.x) / 9) + 4) % 9)) && (((((int)threadIdx.x) + 36) % 81) < 72)) && (1 <= (((int)threadIdx.x) % 9))) && ((((int)threadIdx.x) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 1008) / 81) * 49)) + ((((((int)threadIdx.x) / 9) + 4) % 9) * 7)) + (((int)threadIdx.x) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 1120)] = (((((9 <= ((((int)threadIdx.x) + 67) % 81)) && (((((int)threadIdx.x) + 67) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 4) % 9))) && (((((int)threadIdx.x) + 4) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 1120) / 81) * 49)) + ((((((int)threadIdx.x) + 67) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 4) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 1232)] = (((((9 <= ((((int)threadIdx.x) + 17) % 81)) && (((((int)threadIdx.x) + 17) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 8) % 9))) && (((((int)threadIdx.x) + 8) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 1232) / 81) * 49)) + ((((((int)threadIdx.x) + 17) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 8) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 1344)] = (((((9 <= ((((int)threadIdx.x) + 48) % 81)) && (((((int)threadIdx.x) + 48) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 3) % 9))) && (((((int)threadIdx.x) + 3) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 1344) / 81) * 49)) + ((((((int)threadIdx.x) + 48) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 3) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 1456)] = (((((9 <= ((((int)threadIdx.x) + 79) % 81)) && (((((int)threadIdx.x) + 79) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 7) % 9))) && (((((int)threadIdx.x) + 7) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 1456) / 81) * 49)) + ((((((int)threadIdx.x) + 79) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 7) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 1568)] = (((((9 <= ((((int)threadIdx.x) + 29) % 81)) && (((((int)threadIdx.x) + 29) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 2) % 9))) && (((((int)threadIdx.x) + 2) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 1568) / 81) * 49)) + ((((((int)threadIdx.x) + 29) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 2) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 1680)] = (((((9 <= ((((int)threadIdx.x) + 60) % 81)) && (((((int)threadIdx.x) + 60) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 6) % 9))) && (((((int)threadIdx.x) + 6) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 1680) / 81) * 49)) + ((((((int)threadIdx.x) + 60) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 6) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 1792)] = (((((9 <= ((((int)threadIdx.x) + 10) % 81)) && (((((int)threadIdx.x) + 10) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 1) % 9))) && (((((int)threadIdx.x) + 1) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 1792) / 81) * 49)) + ((((((int)threadIdx.x) + 10) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 1) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 1904)] = (((((9 <= ((((int)threadIdx.x) + 41) % 81)) && (((((int)threadIdx.x) + 41) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 5) % 9))) && (((((int)threadIdx.x) + 5) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 1904) / 81) * 49)) + ((((((int)threadIdx.x) + 41) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 5) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 2016)] = (((((1 <= (((((int)threadIdx.x) / 9) + 8) % 9)) && (((((int)threadIdx.x) + 72) % 81) < 72)) && (1 <= (((int)threadIdx.x) % 9))) && ((((int)threadIdx.x) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 2016) / 81) * 49)) + ((((((int)threadIdx.x) / 9) + 8) % 9) * 7)) + (((int)threadIdx.x) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 2128)] = (((((9 <= ((((int)threadIdx.x) + 22) % 81)) && (((((int)threadIdx.x) + 22) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 4) % 9))) && (((((int)threadIdx.x) + 4) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 2128) / 81) * 49)) + ((((((int)threadIdx.x) + 22) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 4) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 2240)] = (((((9 <= ((((int)threadIdx.x) + 53) % 81)) && (((((int)threadIdx.x) + 53) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 8) % 9))) && (((((int)threadIdx.x) + 8) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 2240) / 81) * 49)) + ((((((int)threadIdx.x) + 53) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 8) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 2352)] = (((((9 <= ((((int)threadIdx.x) + 3) % 81)) && (((((int)threadIdx.x) + 3) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 3) % 9))) && (((((int)threadIdx.x) + 3) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 2352) / 81) * 49)) + ((((((int)threadIdx.x) + 3) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 3) % 9)) - 8)] : 0.000000e+00f);
- pad_temp_shared[(((int)threadIdx.x) + 2464)] = (((((9 <= ((((int)threadIdx.x) + 34) % 81)) && (((((int)threadIdx.x) + 34) % 81) < 72)) && (1 <= ((((int)threadIdx.x) + 7) % 9))) && (((((int)threadIdx.x) + 7) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 2464) / 81) * 49)) + ((((((int)threadIdx.x) + 34) % 81) / 9) * 7)) + ((((int)threadIdx.x) + 7) % 9)) - 8)] : 0.000000e+00f);
- if (((int)threadIdx.x) < 16) {
- pad_temp_shared[(((int)threadIdx.x) + 2576)] = ((((((int)threadIdx.x) < 7) && (1 <= ((((int)threadIdx.x) + 2) % 9))) && (((((int)threadIdx.x) + 2) % 9) < 8)) ? data[(((((rc_outer_outer * 1568) + (((((int)threadIdx.x) + 2576) / 81) * 49)) + (((((int)threadIdx.x) + 65) / 9) * 7)) + ((int)threadIdx.x)) - 6)] : 0.000000e+00f);
- }
- for (int ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer = 0; ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer < 4; ++ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer) {
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24))] = kernel[(((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 32) + (((int)threadIdx.x) * 8)) % 96) / 3) * 9)) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2) + (((int)threadIdx.x) * 2)) % 3) * 3))];
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 1)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 32) + (((int)threadIdx.x) * 8)) % 96) / 3) * 9)) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2) + (((int)threadIdx.x) * 2)) % 3) * 3)) + 1)];
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 2)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 32) + (((int)threadIdx.x) * 8)) % 96) / 3) * 9)) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2) + (((int)threadIdx.x) * 2)) % 3) * 3)) + 2)];
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 3)] = kernel[(((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 1) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 1) [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 4)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 1) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 1 [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 5)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 1) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 1 [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 6)] = kernel[(((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 2) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 2) [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 7)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 2) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 2 [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 8)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 2) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 2 [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 9)] = kernel[(((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) / 3) + 1) & 31) * 9)) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2) + (((int)threadIdx.x) * 2)) % 3 [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 10)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) / 3) + 1) & 31) * 9)) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2) + (((int)threadIdx.x) * 2)) % [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 11)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) / 3) + 1) & 31) * 9)) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2) + (((int)threadIdx.x) * 2)) % [...]
+ 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 (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 12)] = kernel[(((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 4) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 1 [...]
+ 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 (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 13)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 4) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + [...]
+ 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 (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 14)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 4) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 15)] = kernel[(((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 5) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 2 [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 16)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 5) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 17)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 5) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 18)] = kernel[(((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) / 3) + 2) & 31) * 9)) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2) + (((int)threadIdx.x) * 2)) % [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 19)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) / 3) + 2) & 31) * 9)) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2) + (((int)threadIdx.x) * 2)) % [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 20)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) / 3) + 2) & 31) * 9)) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2) + (((int)threadIdx.x) * 2)) % [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 21)] = kernel[(((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 7) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + 1 [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 22)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 7) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + [...]
- }
- if (((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 7) + (((int)threadIdx.x) >> 4)) < 24) {
- kernel_shared[(((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 2688) + (((int)threadIdx.x) * 24)) + 23)] = kernel[((((((((int)blockIdx.x) * 147456) + ((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 28) + (((int)threadIdx.x) >> 2)) / 3) * 4608)) + (rc_outer_outer * 288)) + ((((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 896) + (((int)threadIdx.x) * 8)) + 7) % 96) / 3) * 9)) + (((((ax0_ax1_fused_ax2_fused_ax3_fused_outer_outer * 224) + (((int)threadIdx.x) * 2)) + [...]
- }
- }
- __syncthreads();
- for (int rc_outer_inner = 0; rc_outer_inner < 2; ++rc_outer_inner) {
- for (int rx_outer_inner = 0; rx_outer_inner < 3; ++rx_outer_inner) {
- for (int ff_outer_inner = 0; ff_outer_inner < 2; ++ff_outer_inner) {
- for (int rc_inner = 0; rc_inner < 16; ++rc_inner) {
- conv2d_nchw[(ff_outer_inner * 7)] = (conv2d_nchw[(ff_outer_inner * 7)] + (pad_temp_shared[((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7))] * kernel_shared[((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner)]));
- conv2d_nchw[((ff_outer_inner * 7) + 1)] = (conv2d_nchw[((ff_outer_inner * 7) + 1)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 9)] * kernel_shared[((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner)]));
- conv2d_nchw[((ff_outer_inner * 7) + 2)] = (conv2d_nchw[((ff_outer_inner * 7) + 2)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 18)] * kernel_shared[((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner)]));
- conv2d_nchw[((ff_outer_inner * 7) + 3)] = (conv2d_nchw[((ff_outer_inner * 7) + 3)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 27)] * kernel_shared[((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner)]));
- conv2d_nchw[((ff_outer_inner * 7) + 4)] = (conv2d_nchw[((ff_outer_inner * 7) + 4)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 36)] * kernel_shared[((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner)]));
- conv2d_nchw[((ff_outer_inner * 7) + 5)] = (conv2d_nchw[((ff_outer_inner * 7) + 5)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 45)] * kernel_shared[((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner)]));
- conv2d_nchw[((ff_outer_inner * 7) + 6)] = (conv2d_nchw[((ff_outer_inner * 7) + 6)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 54)] * kernel_shared[((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner)]));
- conv2d_nchw[(ff_outer_inner * 7)] = (conv2d_nchw[(ff_outer_inner * 7)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 9)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 3)]));
- conv2d_nchw[((ff_outer_inner * 7) + 1)] = (conv2d_nchw[((ff_outer_inner * 7) + 1)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 18)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 3)]));
- conv2d_nchw[((ff_outer_inner * 7) + 2)] = (conv2d_nchw[((ff_outer_inner * 7) + 2)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 27)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 3)]));
- conv2d_nchw[((ff_outer_inner * 7) + 3)] = (conv2d_nchw[((ff_outer_inner * 7) + 3)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 36)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 3)]));
- conv2d_nchw[((ff_outer_inner * 7) + 4)] = (conv2d_nchw[((ff_outer_inner * 7) + 4)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 45)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 3)]));
- conv2d_nchw[((ff_outer_inner * 7) + 5)] = (conv2d_nchw[((ff_outer_inner * 7) + 5)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 54)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 3)]));
- conv2d_nchw[((ff_outer_inner * 7) + 6)] = (conv2d_nchw[((ff_outer_inner * 7) + 6)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 63)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 3)]));
- conv2d_nchw[(ff_outer_inner * 7)] = (conv2d_nchw[(ff_outer_inner * 7)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 18)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 6)]));
- conv2d_nchw[((ff_outer_inner * 7) + 1)] = (conv2d_nchw[((ff_outer_inner * 7) + 1)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 27)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 6)]));
- conv2d_nchw[((ff_outer_inner * 7) + 2)] = (conv2d_nchw[((ff_outer_inner * 7) + 2)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 36)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 6)]));
- conv2d_nchw[((ff_outer_inner * 7) + 3)] = (conv2d_nchw[((ff_outer_inner * 7) + 3)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 45)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 6)]));
- conv2d_nchw[((ff_outer_inner * 7) + 4)] = (conv2d_nchw[((ff_outer_inner * 7) + 4)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 54)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 6)]));
- conv2d_nchw[((ff_outer_inner * 7) + 5)] = (conv2d_nchw[((ff_outer_inner * 7) + 5)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 63)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 6)]));
- conv2d_nchw[((ff_outer_inner * 7) + 6)] = (conv2d_nchw[((ff_outer_inner * 7) + 6)] + (pad_temp_shared[(((((rc_outer_inner * 1296) + (rc_inner * 81)) + rx_outer_inner) + (((int)threadIdx.x) % 7)) + 72)] * kernel_shared[(((((((((int)threadIdx.x) / 7) * 576) + (ff_outer_inner * 288)) + (rc_outer_inner * 144)) + (rc_inner * 9)) + rx_outer_inner) + 6)]));
- }
- }
+ 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)]));
}
}
for (int i1_inner = 0; i1_inner < 2; ++i1_inner) {
- for (int i2_inner = 0; i2_inner < 7; ++i2_inner) {
- compute[(((((((int)blockIdx.x) * 1568) + ((((int)threadIdx.x) / 7) * 98)) + (i1_inner * 49)) + (i2_inner * 7)) + (((int)threadIdx.x) % 7))] = max((conv2d_nchw[((i1_inner * 7) + i2_inner)] + bias[(((((int)blockIdx.x) * 32) + ((((int)threadIdx.x) / 7) * 2)) + i1_inner)]), 0.000000e+00f);
+ 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);
}
}
}
@@ -995,7 +1574,7 @@ 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 46.341 seconds)</p>
+<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 5 minutes 27.031 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 0493b7ff77..70b66e25b1 100644
--- a/docs/how_to/tune_with_autoscheduler/tune_network_cuda.html
+++ b/docs/how_to/tune_with_autoscheduler/tune_network_cuda.html
@@ -916,7 +916,7 @@ 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)
- 7.8969 7.8932 7.9050 7.8925 0.0057
+ 7.8663 7.8661 7.8708 7.8619 0.0036
</pre></div>
</div>
</div>
@@ -938,7 +938,7 @@ 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 5.810 seconds)</p>
+<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 1 minutes 4.732 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 de993870ea..da6fc105a6 100644
--- a/docs/how_to/tune_with_autoscheduler/tune_network_x86.html
+++ b/docs/how_to/tune_with_autoscheduler/tune_network_x86.html
@@ -935,7 +935,7 @@ 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)
- 751.3712 751.1382 753.5222 749.4532 1.6693
+ 750.0263 750.0990 750.6909 749.2892 0.5746
</pre></div>
</div>
</div>
@@ -957,7 +957,7 @@ 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 38.610 seconds)</p>
+<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 1 minutes 36.571 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 752980959e..1fe48bf1a3 100644
--- a/docs/how_to/tune_with_autoscheduler/tune_sparse_x86.html
+++ b/docs/how_to/tune_with_autoscheduler/tune_sparse_x86.html
@@ -627,82 +627,409 @@ layout transformation, parallelization, vectorization, unrolling, and operator f
@I.ir_module
class Module:
@T.prim_func
- def main(placeholder: T.handle, placeholder_1: T.handle, placeholder_2: T.handle, placeholder_3: T.handle, placeholder_4: T.handle, compute: T.handle):
+ def main(placeholder: T.Buffer((128, 256), "float32"), placeholder_1: T.Buffer((4916, 16, 1), "float32"), placeholder_2: T.Buffer((4916,), "int32"), placeholder_3: T.Buffer((33,), "int32"), placeholder_4: T.Buffer((128, 512), "float32"), compute: T.Buffer((128, 512), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- placeholder_5 = T.match_buffer(placeholder, (128, 256))
- placeholder_6 = T.match_buffer(placeholder_1, (4916, 16, 1))
- placeholder_7 = T.match_buffer(placeholder_2, (4916,), "int32")
- placeholder_8 = T.match_buffer(placeholder_3, (33,), "int32")
- placeholder_9 = T.match_buffer(placeholder_4, (128, 512))
- compute_1 = T.match_buffer(compute, (128, 512))
for i0_outer_i1_outer_fused in T.parallel(128):
- compute_2 = T.allocate([512], "float32", "global")
- compute_3 = T.buffer_decl((512,), data=compute_2)
+ compute_1 = T.allocate([512], "float32", "global")
+ compute_2 = T.buffer_decl((512,), data=compute_1)
for i_outer_inner, nb_j_inner in T.grid(2, 2):
- for i_inner_init in range(8):
- cse_var_1: T.int32 = i_outer_inner * 256 + i_inner_init * 32 + nb_j_inner * 16
- compute_3[cse_var_1] = T.float32(0)
- compute_3[cse_var_1 + 1] = T.float32(0)
- compute_3[cse_var_1 + 2] = T.float32(0)
- compute_3[cse_var_1 + 3] = T.float32(0)
- compute_3[cse_var_1 + 4] = T.float32(0)
- compute_3[cse_var_1 + 5] = T.float32(0)
- compute_3[cse_var_1 + 6] = T.float32(0)
- compute_3[cse_var_1 + 7] = T.float32(0)
- compute_3[cse_var_1 + 8] = T.float32(0)
- compute_3[cse_var_1 + 9] = T.float32(0)
- compute_3[cse_var_1 + 10] = T.float32(0)
- compute_3[cse_var_1 + 11] = T.float32(0)
- compute_3[cse_var_1 + 12] = T.float32(0)
- compute_3[cse_var_1 + 13] = T.float32(0)
- compute_3[cse_var_1 + 14] = T.float32(0)
- compute_3[cse_var_1 + 15] = T.float32(0)
- for elem_idx, i_inner in T.grid(T.let(cse_var_2, i0_outer_i1_outer_fused % 16 * 2 + nb_j_inner, placeholder_10[cse_var_2 + 1] - placeholder_10[cse_var_2]), 8):
- cse_var_2 = T.var("int32")
- placeholder_10 = T.buffer_decl((33,), "int32", data=placeholder_8.data)
- cse_var_21: T.int32 = elem_idx * 16
- cse_var_20: T.int32 = i0_outer_i1_outer_fused % 16 * 2 + nb_j_inner
- cse_var_19: T.int32 = i_outer_inner * 256 + i_inner * 32 + nb_j_inner * 16
- cse_var_18: T.int32 = i0_outer_i1_outer_fused // 16 * 4096 + i_outer_inner * 2048 + i_inner * 256
- cse_var_17: T.int32 = cse_var_19 + 9
- cse_var_16: T.int32 = cse_var_19 + 8
- cse_var_15: T.int32 = cse_var_19 + 7
- cse_var_14: T.int32 = cse_var_19 + 6
- cse_var_13: T.int32 = cse_var_19 + 5
- cse_var_12: T.int32 = cse_var_19 + 4
- cse_var_11: T.int32 = cse_var_19 + 3
- cse_var_10: T.int32 = cse_var_19 + 2
- cse_var_9: T.int32 = cse_var_19 + 15
- cse_var_8: T.int32 = cse_var_19 + 14
- cse_var_7: T.int32 = cse_var_19 + 13
- cse_var_6: T.int32 = cse_var_19 + 12
- cse_var_5: T.int32 = cse_var_19 + 11
- cse_var_4: T.int32 = cse_var_19 + 10
- cse_var_3: T.int32 = cse_var_19 + 1
- placeholder_11 = T.buffer_decl((78656,), data=placeholder_6.data)
- placeholder_12 = T.buffer_decl((32768,), data=placeholder_5.data)
- placeholder_13 = T.buffer_decl((4916,), "int32", data=placeholder_7.data)
- compute_3[cse_var_19] = compute_3[cse_var_19] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_3] = compute_3[cse_var_3] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 1] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_10] = compute_3[cse_var_10] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 2] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_11] = compute_3[cse_var_11] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 3] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_12] = compute_3[cse_var_12] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 4] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_13] = compute_3[cse_var_13] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 5] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_14] = compute_3[cse_var_14] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 6] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_15] = compute_3[cse_var_15] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 7] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_16] = compute_3[cse_var_16] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 8] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_17] = compute_3[cse_var_17] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 9] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_4] = compute_3[cse_var_4] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 10] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_5] = compute_3[cse_var_5] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 11] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_6] = compute_3[cse_var_6] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 12] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_7] = compute_3[cse_var_7] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 13] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_8] = compute_3[cse_var_8] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 14] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
- compute_3[cse_var_9] = compute_3[cse_var_9] + placeholder_11[placeholder_10[cse_var_20] * 16 + cse_var_21 + 15] * T.max(placeholder_12[cse_var_18 + placeholder_13[placeholder_10[cse_var_20] + elem_idx]], T.float32(0))
+ cse_var_2: T.int32 = i_outer_inner * 256 + nb_j_inner * 16
+ cse_var_1: T.int32 = i0_outer_i1_outer_fused % 16 * 2 + nb_j_inner
+ compute_2[cse_var_2] = T.float32(0)
+ compute_2[cse_var_2 + 1] = T.float32(0)
+ compute_2[cse_var_2 + 2] = T.float32(0)
+ compute_2[cse_var_2 + 3] = T.float32(0)
+ compute_2[cse_var_2 + 4] = T.float32(0)
+ compute_2[cse_var_2 + 5] = T.float32(0)
+ compute_2[cse_var_2 + 6] = T.float32(0)
+ compute_2[cse_var_2 + 7] = T.float32(0)
+ compute_2[cse_var_2 + 8] = T.float32(0)
+ compute_2[cse_var_2 + 9] = T.float32(0)
+ compute_2[cse_var_2 + 10] = T.float32(0)
+ compute_2[cse_var_2 + 11] = T.float32(0)
+ compute_2[cse_var_2 + 12] = T.float32(0)
+ compute_2[cse_var_2 + 13] = T.float32(0)
+ compute_2[cse_var_2 + 14] = T.float32(0)
+ compute_2[cse_var_2 + 15] = T.float32(0)
+ compute_2[cse_var_2 + 32] = T.float32(0)
+ compute_2[cse_var_2 + 33] = T.float32(0)
+ compute_2[cse_var_2 + 34] = T.float32(0)
+ compute_2[cse_var_2 + 35] = T.float32(0)
+ compute_2[cse_var_2 + 36] = T.float32(0)
+ compute_2[cse_var_2 + 37] = T.float32(0)
+ compute_2[cse_var_2 + 38] = T.float32(0)
+ compute_2[cse_var_2 + 39] = T.float32(0)
+ compute_2[cse_var_2 + 40] = T.float32(0)
+ compute_2[cse_var_2 + 41] = T.float32(0)
+ compute_2[cse_var_2 + 42] = T.float32(0)
+ compute_2[cse_var_2 + 43] = T.float32(0)
+ compute_2[cse_var_2 + 44] = T.float32(0)
+ compute_2[cse_var_2 + 45] = T.float32(0)
+ compute_2[cse_var_2 + 46] = T.float32(0)
+ compute_2[cse_var_2 + 47] = T.float32(0)
+ compute_2[cse_var_2 + 64] = T.float32(0)
+ compute_2[cse_var_2 + 65] = T.float32(0)
+ compute_2[cse_var_2 + 66] = T.float32(0)
+ compute_2[cse_var_2 + 67] = T.float32(0)
+ compute_2[cse_var_2 + 68] = T.float32(0)
+ compute_2[cse_var_2 + 69] = T.float32(0)
+ compute_2[cse_var_2 + 70] = T.float32(0)
+ compute_2[cse_var_2 + 71] = T.float32(0)
+ compute_2[cse_var_2 + 72] = T.float32(0)
+ compute_2[cse_var_2 + 73] = T.float32(0)
+ compute_2[cse_var_2 + 74] = T.float32(0)
+ compute_2[cse_var_2 + 75] = T.float32(0)
+ compute_2[cse_var_2 + 76] = T.float32(0)
+ compute_2[cse_var_2 + 77] = T.float32(0)
+ compute_2[cse_var_2 + 78] = T.float32(0)
+ compute_2[cse_var_2 + 79] = T.float32(0)
+ compute_2[cse_var_2 + 96] = T.float32(0)
+ compute_2[cse_var_2 + 97] = T.float32(0)
+ compute_2[cse_var_2 + 98] = T.float32(0)
+ compute_2[cse_var_2 + 99] = T.float32(0)
+ compute_2[cse_var_2 + 100] = T.float32(0)
+ compute_2[cse_var_2 + 101] = T.float32(0)
+ compute_2[cse_var_2 + 102] = T.float32(0)
+ compute_2[cse_var_2 + 103] = T.float32(0)
+ compute_2[cse_var_2 + 104] = T.float32(0)
+ compute_2[cse_var_2 + 105] = T.float32(0)
+ compute_2[cse_var_2 + 106] = T.float32(0)
+ compute_2[cse_var_2 + 107] = T.float32(0)
+ compute_2[cse_var_2 + 108] = T.float32(0)
+ compute_2[cse_var_2 + 109] = T.float32(0)
+ compute_2[cse_var_2 + 110] = T.float32(0)
+ compute_2[cse_var_2 + 111] = T.float32(0)
+ compute_2[cse_var_2 + 128] = T.float32(0)
+ compute_2[cse_var_2 + 129] = T.float32(0)
+ compute_2[cse_var_2 + 130] = T.float32(0)
+ compute_2[cse_var_2 + 131] = T.float32(0)
+ compute_2[cse_var_2 + 132] = T.float32(0)
+ compute_2[cse_var_2 + 133] = T.float32(0)
+ compute_2[cse_var_2 + 134] = T.float32(0)
+ compute_2[cse_var_2 + 135] = T.float32(0)
+ compute_2[cse_var_2 + 136] = T.float32(0)
+ compute_2[cse_var_2 + 137] = T.float32(0)
+ compute_2[cse_var_2 + 138] = T.float32(0)
+ compute_2[cse_var_2 + 139] = T.float32(0)
+ compute_2[cse_var_2 + 140] = T.float32(0)
+ compute_2[cse_var_2 + 141] = T.float32(0)
+ compute_2[cse_var_2 + 142] = T.float32(0)
+ compute_2[cse_var_2 + 143] = T.float32(0)
+ compute_2[cse_var_2 + 160] = T.float32(0)
+ compute_2[cse_var_2 + 161] = T.float32(0)
+ compute_2[cse_var_2 + 162] = T.float32(0)
+ compute_2[cse_var_2 + 163] = T.float32(0)
+ compute_2[cse_var_2 + 164] = T.float32(0)
+ compute_2[cse_var_2 + 165] = T.float32(0)
+ compute_2[cse_var_2 + 166] = T.float32(0)
+ compute_2[cse_var_2 + 167] = T.float32(0)
+ compute_2[cse_var_2 + 168] = T.float32(0)
+ compute_2[cse_var_2 + 169] = T.float32(0)
+ compute_2[cse_var_2 + 170] = T.float32(0)
+ compute_2[cse_var_2 + 171] = T.float32(0)
+ compute_2[cse_var_2 + 172] = T.float32(0)
+ compute_2[cse_var_2 + 173] = T.float32(0)
+ compute_2[cse_var_2 + 174] = T.float32(0)
+ compute_2[cse_var_2 + 175] = T.float32(0)
+ compute_2[cse_var_2 + 192] = T.float32(0)
+ compute_2[cse_var_2 + 193] = T.float32(0)
+ compute_2[cse_var_2 + 194] = T.float32(0)
+ compute_2[cse_var_2 + 195] = T.float32(0)
+ compute_2[cse_var_2 + 196] = T.float32(0)
+ compute_2[cse_var_2 + 197] = T.float32(0)
+ compute_2[cse_var_2 + 198] = T.float32(0)
+ compute_2[cse_var_2 + 199] = T.float32(0)
+ compute_2[cse_var_2 + 200] = T.float32(0)
+ compute_2[cse_var_2 + 201] = T.float32(0)
+ compute_2[cse_var_2 + 202] = T.float32(0)
+ compute_2[cse_var_2 + 203] = T.float32(0)
+ compute_2[cse_var_2 + 204] = T.float32(0)
+ compute_2[cse_var_2 + 205] = T.float32(0)
+ compute_2[cse_var_2 + 206] = T.float32(0)
+ compute_2[cse_var_2 + 207] = T.float32(0)
+ compute_2[cse_var_2 + 224] = T.float32(0)
+ compute_2[cse_var_2 + 225] = T.float32(0)
+ compute_2[cse_var_2 + 226] = T.float32(0)
+ compute_2[cse_var_2 + 227] = T.float32(0)
+ compute_2[cse_var_2 + 228] = T.float32(0)
+ compute_2[cse_var_2 + 229] = T.float32(0)
+ compute_2[cse_var_2 + 230] = T.float32(0)
+ compute_2[cse_var_2 + 231] = T.float32(0)
+ compute_2[cse_var_2 + 232] = T.float32(0)
+ compute_2[cse_var_2 + 233] = T.float32(0)
+ compute_2[cse_var_2 + 234] = T.float32(0)
+ compute_2[cse_var_2 + 235] = T.float32(0)
+ compute_2[cse_var_2 + 236] = T.float32(0)
+ compute_2[cse_var_2 + 237] = T.float32(0)
+ compute_2[cse_var_2 + 238] = T.float32(0)
+ compute_2[cse_var_2 + 239] = T.float32(0)
+ for elem_idx in range(placeholder_5[cse_var_1 + 1] - placeholder_5[cse_var_1]):
+ placeholder_5 = T.buffer_decl((33,), "int32", data=placeholder_3.data)
+ cse_var_131: T.int32 = elem_idx * 16
+ cse_var_130: T.int32 = cse_var_2 + 99
+ cse_var_129: T.int32 = cse_var_2 + 98
+ cse_var_128: T.int32 = cse_var_2 + 97
+ cse_var_127: T.int32 = cse_var_2 + 96
+ cse_var_126: T.int32 = cse_var_2 + 9
+ cse_var_125: T.int32 = cse_var_2 + 8
+ cse_var_124: T.int32 = cse_var_2 + 79
+ cse_var_123: T.int32 = cse_var_2 + 78
+ cse_var_122: T.int32 = cse_var_2 + 77
+ cse_var_121: T.int32 = cse_var_2 + 76
+ cse_var_120: T.int32 = cse_var_2 + 75
+ cse_var_119: T.int32 = cse_var_2 + 74
+ cse_var_118: T.int32 = cse_var_2 + 73
+ cse_var_117: T.int32 = cse_var_2 + 72
+ cse_var_116: T.int32 = cse_var_2 + 71
+ cse_var_115: T.int32 = cse_var_2 + 70
+ cse_var_114: T.int32 = cse_var_2 + 7
+ cse_var_113: T.int32 = cse_var_2 + 69
+ cse_var_112: T.int32 = cse_var_2 + 68
+ cse_var_111: T.int32 = cse_var_2 + 67
+ cse_var_110: T.int32 = cse_var_2 + 66
+ cse_var_109: T.int32 = cse_var_2 + 65
+ cse_var_108: T.int32 = cse_var_2 + 64
+ cse_var_107: T.int32 = cse_var_2 + 6
+ cse_var_106: T.int32 = cse_var_2 + 5
+ cse_var_105: T.int32 = cse_var_2 + 47
+ cse_var_104: T.int32 = cse_var_2 + 46
+ cse_var_103: T.int32 = cse_var_2 + 45
+ cse_var_102: T.int32 = cse_var_2 + 44
+ cse_var_101: T.int32 = cse_var_2 + 43
+ cse_var_100: T.int32 = cse_var_2 + 42
+ cse_var_99: T.int32 = cse_var_2 + 41
+ cse_var_98: T.int32 = cse_var_2 + 40
+ cse_var_97: T.int32 = cse_var_2 + 4
+ cse_var_96: T.int32 = cse_var_2 + 39
+ cse_var_95: T.int32 = cse_var_2 + 38
+ cse_var_94: T.int32 = cse_var_2 + 37
+ cse_var_93: T.int32 = cse_var_2 + 36
+ cse_var_92: T.int32 = cse_var_2 + 35
+ cse_var_91: T.int32 = cse_var_2 + 34
+ cse_var_90: T.int32 = cse_var_2 + 33
+ cse_var_89: T.int32 = cse_var_2 + 32
+ cse_var_88: T.int32 = cse_var_2 + 3
+ cse_var_87: T.int32 = cse_var_2 + 239
+ cse_var_86: T.int32 = cse_var_2 + 238
+ cse_var_85: T.int32 = cse_var_2 + 237
+ cse_var_84: T.int32 = cse_var_2 + 236
+ cse_var_83: T.int32 = cse_var_2 + 235
+ cse_var_82: T.int32 = cse_var_2 + 234
+ cse_var_81: T.int32 = cse_var_2 + 233
+ cse_var_80: T.int32 = cse_var_2 + 232
+ cse_var_79: T.int32 = cse_var_2 + 231
+ cse_var_78: T.int32 = cse_var_2 + 230
+ cse_var_77: T.int32 = cse_var_2 + 229
+ cse_var_76: T.int32 = cse_var_2 + 228
+ cse_var_75: T.int32 = cse_var_2 + 227
+ cse_var_74: T.int32 = cse_var_2 + 226
+ cse_var_73: T.int32 = cse_var_2 + 225
+ cse_var_72: T.int32 = cse_var_2 + 224
+ cse_var_71: T.int32 = cse_var_2 + 207
+ cse_var_70: T.int32 = cse_var_2 + 206
+ cse_var_69: T.int32 = cse_var_2 + 205
+ cse_var_68: T.int32 = cse_var_2 + 204
+ cse_var_67: T.int32 = cse_var_2 + 203
+ cse_var_66: T.int32 = cse_var_2 + 202
+ cse_var_65: T.int32 = cse_var_2 + 201
+ cse_var_64: T.int32 = cse_var_2 + 200
+ cse_var_63: T.int32 = cse_var_2 + 2
+ cse_var_62: T.int32 = cse_var_2 + 199
+ cse_var_61: T.int32 = cse_var_2 + 198
+ cse_var_60: T.int32 = cse_var_2 + 197
+ cse_var_59: T.int32 = cse_var_2 + 196
+ cse_var_58: T.int32 = cse_var_2 + 195
+ cse_var_57: T.int32 = cse_var_2 + 194
+ cse_var_56: T.int32 = cse_var_2 + 193
+ cse_var_55: T.int32 = cse_var_2 + 192
+ cse_var_54: T.int32 = cse_var_2 + 175
+ cse_var_53: T.int32 = cse_var_2 + 174
+ cse_var_52: T.int32 = cse_var_2 + 173
+ cse_var_51: T.int32 = cse_var_2 + 172
+ cse_var_50: T.int32 = cse_var_2 + 171
+ cse_var_49: T.int32 = cse_var_2 + 170
+ cse_var_48: T.int32 = cse_var_2 + 169
+ cse_var_47: T.int32 = cse_var_2 + 168
+ cse_var_46: T.int32 = cse_var_2 + 167
+ cse_var_45: T.int32 = cse_var_2 + 166
+ cse_var_44: T.int32 = cse_var_2 + 165
+ cse_var_43: T.int32 = cse_var_2 + 164
+ cse_var_42: T.int32 = cse_var_2 + 163
+ cse_var_41: T.int32 = cse_var_2 + 162
+ cse_var_40: T.int32 = cse_var_2 + 161
+ cse_var_39: T.int32 = cse_var_2 + 160
+ cse_var_38: T.int32 = cse_var_2 + 15
+ cse_var_37: T.int32 = cse_var_2 + 143
+ cse_var_36: T.int32 = cse_var_2 + 142
+ cse_var_35: T.int32 = cse_var_2 + 141
+ cse_var_34: T.int32 = cse_var_2 + 140
+ cse_var_33: T.int32 = cse_var_2 + 14
+ cse_var_32: T.int32 = cse_var_2 + 139
+ cse_var_31: T.int32 = cse_var_2 + 138
+ cse_var_30: T.int32 = cse_var_2 + 137
+ cse_var_29: T.int32 = cse_var_2 + 136
+ cse_var_28: T.int32 = cse_var_2 + 135
+ cse_var_27: T.int32 = cse_var_2 + 134
+ cse_var_26: T.int32 = cse_var_2 + 133
+ cse_var_25: T.int32 = cse_var_2 + 132
+ cse_var_24: T.int32 = cse_var_2 + 131
+ cse_var_23: T.int32 = cse_var_2 + 130
+ cse_var_22: T.int32 = cse_var_2 + 13
+ cse_var_21: T.int32 = cse_var_2 + 129
+ cse_var_20: T.int32 = cse_var_2 + 128
+ cse_var_19: T.int32 = cse_var_2 + 12
+ cse_var_18: T.int32 = cse_var_2 + 111
+ cse_var_17: T.int32 = cse_var_2 + 110
+ cse_var_16: T.int32 = cse_var_2 + 11
+ cse_var_15: T.int32 = cse_var_2 + 109
+ cse_var_14: T.int32 = cse_var_2 + 108
+ cse_var_13: T.int32 = cse_var_2 + 107
+ cse_var_12: T.int32 = cse_var_2 + 106
+ cse_var_11: T.int32 = cse_var_2 + 105
+ cse_var_10: T.int32 = cse_var_2 + 104
+ cse_var_9: T.int32 = cse_var_2 + 103
+ cse_var_8: T.int32 = cse_var_2 + 102
+ cse_var_7: T.int32 = cse_var_2 + 101
+ cse_var_6: T.int32 = cse_var_2 + 100
+ cse_var_5: T.int32 = cse_var_2 + 10
+ cse_var_4: T.int32 = cse_var_2 + 1
+ cse_var_3: T.int32 = i0_outer_i1_outer_fused // 16 * 4096 + i_outer_inner * 2048
+ placeholder_6 = T.buffer_decl((78656,), data=placeholder_1.data)
+ placeholder_7 = T.buffer_decl((32768,), data=placeholder.data)
+ placeholder_8 = T.buffer_decl((4916,), "int32", data=placeholder_2.data)
+ compute_2[cse_var_2] = compute_2[cse_var_2] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_4] = compute_2[cse_var_4] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 1] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_63] = compute_2[cse_var_63] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 2] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_88] = compute_2[cse_var_88] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 3] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_97] = compute_2[cse_var_97] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 4] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_106] = compute_2[cse_var_106] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 5] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_107] = compute_2[cse_var_107] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 6] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_114] = compute_2[cse_var_114] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 7] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_125] = compute_2[cse_var_125] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 8] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_126] = compute_2[cse_var_126] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 9] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_5] = compute_2[cse_var_5] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 10] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_16] = compute_2[cse_var_16] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 11] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_19] = compute_2[cse_var_19] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 12] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_22] = compute_2[cse_var_22] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 13] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_33] = compute_2[cse_var_33] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 14] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_38] = compute_2[cse_var_38] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 15] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx]], T.float32(0))
+ compute_2[cse_var_89] = compute_2[cse_var_89] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_90] = compute_2[cse_var_90] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 1] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_91] = compute_2[cse_var_91] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 2] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_92] = compute_2[cse_var_92] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 3] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_93] = compute_2[cse_var_93] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 4] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_94] = compute_2[cse_var_94] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 5] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_95] = compute_2[cse_var_95] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 6] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_96] = compute_2[cse_var_96] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 7] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_98] = compute_2[cse_var_98] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 8] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_99] = compute_2[cse_var_99] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 9] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_100] = compute_2[cse_var_100] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 10] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_101] = compute_2[cse_var_101] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 11] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_102] = compute_2[cse_var_102] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 12] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_103] = compute_2[cse_var_103] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 13] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_104] = compute_2[cse_var_104] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 14] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_105] = compute_2[cse_var_105] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 15] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 256], T.float32(0))
+ compute_2[cse_var_108] = compute_2[cse_var_108] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_109] = compute_2[cse_var_109] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 1] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_110] = compute_2[cse_var_110] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 2] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_111] = compute_2[cse_var_111] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 3] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_112] = compute_2[cse_var_112] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 4] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_113] = compute_2[cse_var_113] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 5] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_115] = compute_2[cse_var_115] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 6] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_116] = compute_2[cse_var_116] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 7] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_117] = compute_2[cse_var_117] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 8] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_118] = compute_2[cse_var_118] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 9] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_119] = compute_2[cse_var_119] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 10] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_120] = compute_2[cse_var_120] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 11] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_121] = compute_2[cse_var_121] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 12] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_122] = compute_2[cse_var_122] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 13] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_123] = compute_2[cse_var_123] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 14] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_124] = compute_2[cse_var_124] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 15] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 512], T.float32(0))
+ compute_2[cse_var_127] = compute_2[cse_var_127] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_128] = compute_2[cse_var_128] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 1] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_129] = compute_2[cse_var_129] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 2] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_130] = compute_2[cse_var_130] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 3] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_6] = compute_2[cse_var_6] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 4] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_7] = compute_2[cse_var_7] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 5] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_8] = compute_2[cse_var_8] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 6] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_9] = compute_2[cse_var_9] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 7] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_10] = compute_2[cse_var_10] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 8] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_11] = compute_2[cse_var_11] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 9] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_12] = compute_2[cse_var_12] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 10] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_13] = compute_2[cse_var_13] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 11] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_14] = compute_2[cse_var_14] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 12] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_15] = compute_2[cse_var_15] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 13] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_17] = compute_2[cse_var_17] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 14] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_18] = compute_2[cse_var_18] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 15] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 768], T.float32(0))
+ compute_2[cse_var_20] = compute_2[cse_var_20] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_21] = compute_2[cse_var_21] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 1] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_23] = compute_2[cse_var_23] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 2] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_24] = compute_2[cse_var_24] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 3] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_25] = compute_2[cse_var_25] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 4] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_26] = compute_2[cse_var_26] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 5] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_27] = compute_2[cse_var_27] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 6] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_28] = compute_2[cse_var_28] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 7] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_29] = compute_2[cse_var_29] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 8] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_30] = compute_2[cse_var_30] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 9] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_31] = compute_2[cse_var_31] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 10] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_32] = compute_2[cse_var_32] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 11] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_34] = compute_2[cse_var_34] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 12] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_35] = compute_2[cse_var_35] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 13] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_36] = compute_2[cse_var_36] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 14] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_37] = compute_2[cse_var_37] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 15] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1024], T.float32(0))
+ compute_2[cse_var_39] = compute_2[cse_var_39] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_40] = compute_2[cse_var_40] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 1] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_41] = compute_2[cse_var_41] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 2] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_42] = compute_2[cse_var_42] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 3] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_43] = compute_2[cse_var_43] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 4] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_44] = compute_2[cse_var_44] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 5] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_45] = compute_2[cse_var_45] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 6] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_46] = compute_2[cse_var_46] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 7] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_47] = compute_2[cse_var_47] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 8] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_48] = compute_2[cse_var_48] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 9] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_49] = compute_2[cse_var_49] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 10] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_50] = compute_2[cse_var_50] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 11] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_51] = compute_2[cse_var_51] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 12] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_52] = compute_2[cse_var_52] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 13] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_53] = compute_2[cse_var_53] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 14] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_54] = compute_2[cse_var_54] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 15] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1280], T.float32(0))
+ compute_2[cse_var_55] = compute_2[cse_var_55] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_56] = compute_2[cse_var_56] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 1] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_57] = compute_2[cse_var_57] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 2] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_58] = compute_2[cse_var_58] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 3] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_59] = compute_2[cse_var_59] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 4] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_60] = compute_2[cse_var_60] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 5] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_61] = compute_2[cse_var_61] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 6] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_62] = compute_2[cse_var_62] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 7] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_64] = compute_2[cse_var_64] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 8] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_65] = compute_2[cse_var_65] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 9] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_66] = compute_2[cse_var_66] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 10] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_67] = compute_2[cse_var_67] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 11] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_68] = compute_2[cse_var_68] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 12] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_69] = compute_2[cse_var_69] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 13] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_70] = compute_2[cse_var_70] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 14] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_71] = compute_2[cse_var_71] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 15] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1536], T.float32(0))
+ compute_2[cse_var_72] = compute_2[cse_var_72] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_73] = compute_2[cse_var_73] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 1] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_74] = compute_2[cse_var_74] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 2] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_75] = compute_2[cse_var_75] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 3] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_76] = compute_2[cse_var_76] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 4] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_77] = compute_2[cse_var_77] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 5] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_78] = compute_2[cse_var_78] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 6] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_79] = compute_2[cse_var_79] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 7] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_80] = compute_2[cse_var_80] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 8] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_81] = compute_2[cse_var_81] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 9] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_82] = compute_2[cse_var_82] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 10] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_83] = compute_2[cse_var_83] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 11] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_84] = compute_2[cse_var_84] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 12] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_85] = compute_2[cse_var_85] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 13] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_86] = compute_2[cse_var_86] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 14] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
+ compute_2[cse_var_87] = compute_2[cse_var_87] + placeholder_6[placeholder_5[cse_var_1] * 16 + cse_var_131 + 15] * T.max(placeholder_7[cse_var_3 + placeholder_8[placeholder_5[cse_var_1] + elem_idx] + 1792], T.float32(0))
for i0_inner in range(16):
- cse_var_22: T.int32 = i0_outer_i1_outer_fused // 16 * 8192 + i0_inner * 512 + i0_outer_i1_outer_fused % 16 * 32
- compute_4 = T.buffer_decl((65536,), data=compute_1.data)
- placeholder_10 = T.buffer_decl((65536,), data=placeholder_9.data)
- compute_4[cse_var_22:cse_var_22 + 32] = T.max(compute_3[i0_inner * 32:i0_inner * 32 + 32] + placeholder_10[cse_var_22:cse_var_22 + 32], T.Broadcast(T.float32(0), 32))
+ cse_var_132: T.int32 = i0_outer_i1_outer_fused // 16 * 8192 + i0_inner * 512 + i0_outer_i1_outer_fused % 16 * 32
+ compute_3 = T.buffer_decl((65536,), data=compute.data)
+ placeholder_5 = T.buffer_decl((65536,), data=placeholder_4.data)
+ compute_3[cse_var_132:cse_var_132 + 32] = T.max(compute_2[i0_inner * 32:i0_inner * 32 + 32] + placeholder_5[cse_var_132:cse_var_132 + 32], T.Broadcast(T.float32(0), 32))
</pre></div>
</div>
</div>
@@ -736,7 +1063,7 @@ class Module:
<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: 1.894 ms
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>Execution time of this operator: 2.768 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 f6625da8e5..acca44ffab 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>00:24.192</strong> total execution time for <strong>how_to_tune_with_autotvm</strong> files:</p>
+<p><strong>00:41.552</strong> total execution time for <strong>how_to_tune_with_autotvm</strong> files:</p>
<table class="docutils align-default">
<colgroup>
<col style="width: 84%" />
@@ -349,11 +349,11 @@
</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>00:24.157</p></td>
+<td><p>00:41.519</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>00:00.021</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_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>
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 cd0b1ed6eb..7373724608 100644
--- a/docs/how_to/tune_with_autotvm/tune_conv2d_cuda.html
+++ b/docs/how_to/tune_with_autotvm/tune_conv2d_cuda.html
@@ -568,9 +568,8 @@ for this template</p>
waiting for device...
device available
Get devices for measurement successfully!
-No: 1 GFLOPS: 50.48/50.48 result: MeasureResult(costs=(0.004586107448275862,), error_no=MeasureErrorNo.NO_ERROR, all_cost=3.907130241394043, timestamp=1674175038.7889755) [('tile_f', [-1, 2, 1, 8]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 8, 4]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9371037
-No: 2 GFLOPS: 93.65/93.65 result: MeasureResult(costs=(0.002471886048780488,), error_no=MeasureErrorNo.NO_ERROR, all_cost=2.077481985092163, timestamp=1674175039.5116773) [('tile_f', [-1, 1, 16, 2]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 64, 2]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 1)],None,5284705
-No: 3 GFLOPS: 0.00/93.65 result: Traceback (most recent call last):
+No: 1 GFLOPS: 7.75/7.75 result: MeasureResult(costs=(0.0298710955,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.783884048461914, timestamp=1674185426.5654202) [('tile_f', [-1, 1, 2, 32]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 2, 2]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,41550
+No: 2 GFLOPS: 0.00/7.75 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -692,9 +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 875, 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, 1]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 64, 8]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,1666969
-No: 4 GFLOPS: 213.99/213.99 result: MeasureResult(costs=(0.0010818329677419354,), error_no=MeasureErrorNo.NO_ERROR, all_cost=2.1306028366088867, timestamp=1674175042.3393323) [('tile_f', [-1, 4, 8, 1]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 2, 4]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2782589
-No: 5 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 1, 128, 2]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 128, 2]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,1224397
+No: 3 GFLOPS: 0.00/7.75 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -816,8 +814,8 @@ 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 875, 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, 4]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 8, 64]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,7526534
-No: 6 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 1, 8, 4]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 16, 8]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,10372241
+No: 4 GFLOPS: 0.00/7.75 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -939,8 +937,8 @@ 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 875, 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, 2, 32]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 16, 8]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9404753
-No: 7 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 2, 16, 4]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 16, 4]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,10342327
+No: 5 GFLOPS: 0.00/7.75 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -1062,8 +1060,9 @@ 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 875, 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, 64]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 4, 2]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,1011105
-No: 8 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 2, 128, 2]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 8, 16]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,4392618
+No: 6 GFLOPS: 127.76/127.76 result: MeasureResult(costs=(0.0018119425555555557,), error_no=MeasureErrorNo.NO_ERROR, all_cost=4.839144229888916, timestamp=1674185434.4324708) [('tile_f', [-1, 1, 2, 16]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 16, 1]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9502410
+No: 7 GFLOPS: 0.00/127.76 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -1185,8 +1184,8 @@ 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 875, 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, 1, 2]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 256, 2]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 1)],None,5484881
-No: 9 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 4, 8, 4]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 32, 1]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 1)],None,6022183
+No: 8 GFLOPS: 0.00/127.76 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -1308,8 +1307,8 @@ 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 875, 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, 1, 1, 7]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 16, 2]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,8375713
-No: 10 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 32, 2, 2]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 1, 64]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2289829
+No: 9 GFLOPS: 0.00/127.76 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -1431,8 +1430,26 @@ 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 875, 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, 16, 2]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 64, 4]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9187286
-No: 11 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 2, 4, 4]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 1, 512]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,7162656
+No: 10 GFLOPS: 0.00/127.76 result: Traceback (most recent call last):
+ File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 142, in build
+ res = future.result()
+ File "/usr/lib/python3.7/concurrent/futures/_base.py", line 435, in result
+ return self.__get_result()
+ File "/usr/lib/python3.7/concurrent/futures/_base.py", line 384, in __get_result
+ raise self._exception
+ File "/usr/lib/python3.7/concurrent/futures/thread.py", line 57, in run
+ result = self.fn(*self.args, **self.kwargs)
+ File "/workspace/python/tvm/contrib/popen_pool.py", line 432, in <lambda>
+ worker = lambda *args: self._worker_run(*args)
+ File "/workspace/python/tvm/contrib/popen_pool.py", line 401, in _worker_run
+ return proc.recv()
+ File "/workspace/python/tvm/contrib/popen_pool.py", line 309, in recv
+ raise TimeoutError()
+TimeoutError
+
+ [('tile_f', [-1, 4, 1, 16]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 1, 8]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9195726
+No: 11 GFLOPS: 0.00/127.76 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -1554,8 +1571,8 @@ 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 875, 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, 1]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 8, 16]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2262743
-No: 12 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 64, 2, 4]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 128, 1]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,5058574
+No: 12 GFLOPS: 0.00/127.76 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -1677,8 +1694,10 @@ 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 875, 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, 16, 2]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 32, 16]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,1107565
-No: 13 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 1, 4, 16]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 128, 2]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2190495
+No: 13 GFLOPS: 2.71/127.76 result: MeasureResult(costs=(0.08529100775000001,), error_no=MeasureErrorNo.NO_ERROR, all_cost=5.805093050003052, timestamp=1674185450.5576441) [('tile_f', [-1, 1, 2, 8]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 2, 64]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,7325263
+No: 14 GFLOPS: 299.03/299.03 result: MeasureResult(costs=(0.000774184811594203,), error_no=MeasureErrorNo.NO_ERROR, all_cost=1.5905015468597412, timestamp=1674185451.5622723) [('tile_f', [-1, 2, 32, 1]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 8, 2]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,436081
+No: 15 GFLOPS: 0.00/299.03 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -1800,8 +1819,8 @@ 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 875, 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, 8, 32]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 512, 1]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,420837
-No: 14 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 128, 2, 2]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 8, 32]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2864031
+No: 16 GFLOPS: 0.00/299.03 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -1923,8 +1942,8 @@ 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 875, 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, 32, 8, 1]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 8, 32]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,5185872
-No: 15 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 8, 4, 16]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 32, 1]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 1)],None,5632838
+No: 17 GFLOPS: 0.00/299.03 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -2046,8 +2065,8 @@ 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 875, 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, 4, 1]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 16, 16]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,3232479
-No: 16 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 2, 8, 16]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 1, 256]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2313260
+No: 18 GFLOPS: 0.00/299.03 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -2169,8 +2188,8 @@ 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 875, 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, 32]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 64, 4]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2606310
-No: 17 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 256, 1, 1]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 16, 32]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9836868
+No: 19 GFLOPS: 0.00/299.03 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -2292,8 +2311,8 @@ 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 875, 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, 16, 16]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 1, 8]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 0), ('unroll_explicit', 1)],None,6485563
-No: 18 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 16, 1, 4]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 512, 1]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,7970924
+No: 20 GFLOPS: 0.00/299.03 result: Traceback (most recent call last):
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, in _build_func_common
@@ -2415,253 +2434,7 @@ 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 875, 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, 1, 7, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 4, 32]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 1)],None,5571827
-No: 19 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
- func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, 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:1730
- 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:1670
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1630
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1645
- 13: operator()
- at ../src/driver/driver_api.cc:395
- 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:381
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:276
- 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:451
- 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:1749
- 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:1693
- 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:1617
- 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 875, 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:1730
- 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:1670
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1630
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1645
- 13: operator()
- at ../src/driver/driver_api.cc:395
- 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:381
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:276
- 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:451
- 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:1749
- 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:1693
- 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:1617
- 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 875, 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, 7, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 4, 32]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9441890
-No: 20 GFLOPS: 0.00/213.99 result: Traceback (most recent call last):
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 592, in __call__
- func, arg_info = _build_func_common(measure_input, self.runtime, **kwargs)
- File "/workspace/python/tvm/autotvm/measure/measure_methods.py", line 544, 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:1730
- 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:1670
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1630
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1645
- 13: operator()
- at ../src/driver/driver_api.cc:395
- 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:381
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:276
- 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:451
- 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:1749
- 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:1693
- 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:1617
- 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 875, 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:1730
- 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:1670
- 19: run<>
- at ../include/tvm/runtime/packed_func.h:1630
- 18: run<tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 17: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 16: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 15: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1630
- 14: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
- at ../include/tvm/runtime/packed_func.h:1645
- 13: operator()
- at ../src/driver/driver_api.cc:395
- 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:381
- 11: tvm::LowerWithPassList(tvm::IRModule, tvm::runtime::Array<tvm::transform::Pass, void>)
- at ../src/driver/driver_api.cc:276
- 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:451
- 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:1749
- 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:1693
- 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:1617
- 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 875, 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, 1, 4]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 16, 4]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,4341582
+tvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel [('tile_f', [-1, 2, 1, 32]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 256, 2]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,3935986
</pre></div>
</div>
<p>Finally we can inspect the best config from log file, check correctness,
@@ -2700,9 +2473,9 @@ 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, 4, 8, 1]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 2, 4]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2782589
+[('tile_f', [-1, 2, 32, 1]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 8, 2]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,436081
Finish loading 20 records
-Time cost of this operator: 0.001132
+Time cost of this operator: 0.001133
</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 c5bd15a4e9..ad7b4284a0 100644
--- a/docs/how_to/work_with_microtvm/micro_autotune.html
+++ b/docs/how_to/work_with_microtvm/micro_autotune.html
@@ -647,10 +647,10 @@ 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 310.8 98.686 (1, 2, 10, 10, 3) 2 1 [310.8]
-tvmgen_default_fused_layout_transform_1 tvmgen_default_fused_layout_transform_1 3.186 1.012 (1, 6, 10, 10) 1 1 [3.186]
-tvmgen_default_fused_layout_transform tvmgen_default_fused_layout_transform 0.953 0.303 (1, 1, 10, 10, 3) 1 1 [0.953]
-Total_time - 314.939 - - - - -
+tvmgen_default_fused_nn_contrib_conv2d_NCHWc tvmgen_default_fused_nn_contrib_conv2d_NCHWc 307.8 98.719 (1, 2, 10, 10, 3) 2 1 [307.8]
+tvmgen_default_fused_layout_transform_1 tvmgen_default_fused_layout_transform_1 3.021 0.969 (1, 6, 10, 10) 1 1 [3.021]
+tvmgen_default_fused_layout_transform tvmgen_default_fused_layout_transform 0.972 0.312 (1, 1, 10, 10, 3) 1 1 [0.972]
+Total_time - 311.793 - - - - -
</pre></div>
</div>
</div>
@@ -702,10 +702,10 @@ 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 102.8 97.511 (1, 6, 10, 10, 1) 2 1 [102.8]
-tvmgen_default_fused_layout_transform_1 tvmgen_default_fused_layout_transform_1 1.773 1.682 (1, 6, 10, 10) 1 1 [1.773]
-tvmgen_default_fused_layout_transform tvmgen_default_fused_layout_transform 0.851 0.808 (1, 3, 10, 10, 1) 1 1 [0.851]
-Total_time - 105.424 - - - - -
+tvmgen_default_fused_nn_contrib_conv2d_NCHWc tvmgen_default_fused_nn_contrib_conv2d_NCHWc 104.8 97.555 (1, 6, 10, 10, 1) 2 1 [104.8]
+tvmgen_default_fused_layout_transform_1 tvmgen_default_fused_layout_transform_1 1.776 1.653 (1, 6, 10, 10) 1 1 [1.776]
+tvmgen_default_fused_layout_transform tvmgen_default_fused_layout_transform 0.851 0.792 (1, 3, 10, 10, 1) 1 1 [0.851]
+Total_time - 107.426 - - - - -
</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 216499fa2f..8702f07960 100644
--- a/docs/how_to/work_with_microtvm/micro_pytorch.html
+++ b/docs/how_to/work_with_microtvm/micro_pytorch.html
@@ -454,8 +454,7 @@ 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]
- 61%|###### | 2.09M/3.42M [00:00<00:00, 19.9MB/s]
-100%|##########| 3.42M/3.42M [00:00<00:00, 30.4MB/s]
+100%|##########| 3.42M/3.42M [00:00<00:00, 137MB/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.
@@ -579,7 +578,7 @@ 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 9.130 seconds)</p>
+<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 1 minutes 7.086 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 1f9459f4d6..9d30173986 100644
--- a/docs/how_to/work_with_microtvm/micro_train.html
+++ b/docs/how_to/work_with_microtvm/micro_train.html
@@ -524,7 +524,7 @@ 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/tmp9frj0wj1/images/random'
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>'/tmp/tmpiql0qhf6/images/random'
</pre></div>
</div>
</div>
@@ -584,8 +584,8 @@ 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="[0.0, 1.0], [1.0, 0.0], [1.0, 0.0], [0.0, 1.0], [1.0, 0.0], [0.0, 1.0], [1.0, 0.0], [1.0, 0.0], [0.0, 1.0], [1.0, 0.0]" class = "sphx-glr-single-img"/><div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>/tmp/tmp9frj0wj1/images/target contains 8144 images
-/tmp/tmp9frj0wj1/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], [0.0, 1.0], [0.0, 1.0], [0.0, 1.0], [0.0, 1.0], [0.0, 1.0], [0.0, 1.0], [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/tmpiql0qhf6/images/target contains 8144 images
+/tmp/tmpiql0qhf6/images/random contains 5000 images
</pre></div>
</div>
</div>
@@ -697,13 +697,13 @@ 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: 0.2199 - accuracy: 0.9239 - val_loss: 0.1185 - val_accuracy: 0.9581 - 47s/epoch - 142ms/step
+328/328 - 47s - loss: 0.2138 - accuracy: 0.9251 - val_loss: 0.1147 - val_accuracy: 0.9562 - 47s/epoch - 143ms/step
Epoch 2/3
-328/328 - 43s - loss: 0.1003 - accuracy: 0.9617 - val_loss: 0.1516 - val_accuracy: 0.9486 - 43s/epoch - 132ms/step
+328/328 - 43s - loss: 0.0960 - accuracy: 0.9631 - val_loss: 0.0898 - val_accuracy: 0.9671 - 43s/epoch - 131ms/step
Epoch 3/3
-328/328 - 43s - loss: 0.0737 - accuracy: 0.9728 - val_loss: 0.1310 - val_accuracy: 0.9615 - 43s/epoch - 132ms/step
+328/328 - 43s - loss: 0.0588 - accuracy: 0.9798 - val_loss: 0.1360 - val_accuracy: 0.9528 - 43s/epoch - 131ms/step
-<keras.callbacks.History object at 0x7fe689e4a810>
+<keras.callbacks.History object at 0x7f6b69639fd0>
</pre></div>
</div>
</div>
@@ -963,7 +963,7 @@ 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 29.197 seconds)</p>
+<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 4 minutes 19.926 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 fad165773d..6b95e2c6b5 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>06:42.755</strong> total execution time for <strong>how_to_work_with_microtvm</strong> files:</p>
+<p><strong>06:30.155</strong> total execution time for <strong>how_to_work_with_microtvm</strong> files:</p>
<table class="docutils align-default">
<colgroup>
<col style="width: 83%" />
@@ -349,30 +349,30 @@
</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>04:29.197</p></td>
+<td><p>04:19.926</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>01:09.130</p></td>
+<td><p>01:07.086</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>00:51.744</p></td>
+<td><p>00:50.864</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>00:08.855</p></td>
+<td><p>00:08.512</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>00:03.827</p></td>
+<td><p>00:03.768</p></td>
<td><p>0.0 MB</p></td>
</tr>
-<tr class="row-even"><td><p><a class="reference internal" href="micro_tvmc.html#sphx-glr-how-to-work-with-microtvm-micro-tvmc-py"><span class="std std-ref">Executing a Tiny Model with TVMC Micro</span></a> (<code class="docutils literal notranslate"><span class="pre">micro_tvmc.py</span></code>)</p></td>
+<tr class="row-even"><td><p><a class="reference internal" href="micro_ethosu.html#sphx-glr-how-to-work-with-microtvm-micro-ethosu-py"><span class="std std-ref">Running TVM on bare metal Arm(R) Cortex(R)-M55 CPU and Ethos(TM)-U55 NPU with CMSIS-NN</span></a> (<code class="docutils literal notranslate"><span class="pre">micro_ethosu.py</span></code>)</p></td>
<td><p>00:00.000</p></td>
<td><p>0.0 MB</p></td>
</tr>
-<tr class="row-odd"><td><p><a class="reference internal" href="micro_ethosu.html#sphx-glr-how-to-work-with-microtvm-micro-ethosu-py"><span class="std std-ref">Running TVM on bare metal Arm(R) Cortex(R)-M55 CPU and Ethos(TM)-U55 NPU with CMSIS-NN</span></a> (<code class="docutils literal notranslate"><span class="pre">micro_ethosu.py</span></code>)</p></td>
+<tr class="row-odd"><td><p><a class="reference internal" href="micro_tvmc.html#sphx-glr-how-to-work-with-microtvm-micro-tvmc-py"><span class="std std-ref">Executing a Tiny Model with TVMC Micro</span></a> (<code class="docutils literal notranslate"><span class="pre">micro_tvmc.py</span></code>)</p></td>
<td><p>00:00.000</p></td>
<td><p>0.0 MB</p></td>
</tr>
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 c0c1660efb..1f2bdd12a7 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>00:45.175</strong> total execution time for <strong>how_to_work_with_relay</strong> files:</p>
+<p><strong>00:43.680</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>00:32.926</p></td>
+<td><p>00:31.973</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>00:10.433</p></td>
+<td><p>00:10.161</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>00:01.810</p></td>
+<td><p>00:01.541</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 a6be60d150..04d480eec7 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 0x7fe68a33d290>
+<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span><function my_cuda_math_rule at 0x7f6bc0351440>
</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 6a4f8ab5ff..4889c00975 100644
--- a/docs/how_to/work_with_schedules/reduction.html
+++ b/docs/how_to/work_with_schedules/reduction.html
@@ -682,21 +682,20 @@ Here is an example for 2D convolution with filter size = [3, 3] and strides = [1
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@I.ir_module
class Module:
@T.prim_func
- def main(Input: T.handle, Filter: T.handle, Output: T.handle):
+ def main(Input: T.handle, Filter: T.Buffer((3, 3), "float32"), Output: T.handle):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
n = T.var("int32")
stride = T.var("int32")
stride_1 = T.var("int32")
Input_1 = T.match_buffer(Input, (n, n), strides=(stride, stride_1), type="auto")
- Filter_1 = T.match_buffer(Filter, (3, 3))
Output_1 = T.match_buffer(Output, (n - 2, n - 2))
for i, j in T.grid(n - 2, n - 2):
Output_2 = T.buffer_decl(((n - 2) * (n - 2),), data=Output_1.data)
Output_2[i * (n - 2) + j] = T.float32(0)
for di, dj in T.grid(3, 3):
Input_2 = T.buffer_decl((stride * n,), data=Input_1.data, type="auto")
- Filter_2 = T.buffer_decl((9,), data=Filter_1.data)
- Output_2[i * (n - 2) + j] = Output_2[i * (n - 2) + j] + Input_2[(i + di) * stride + (j + dj) * stride_1] * Filter_2[di * 3 + dj]
+ Filter_1 = T.buffer_decl((9,), data=Filter.data)
+ Output_2[i * (n - 2) + j] = Output_2[i * (n - 2) + j] + Input_2[(i + di) * stride + (j + dj) * stride_1] * Filter_1[di * 3 + dj]
</pre></div>
</div>
</div>
diff --git a/docs/how_to/work_with_schedules/sg_execution_times.html b/docs/how_to/work_with_schedules/sg_execution_times.html
index 4828b5a431..80deb40a55 100644
--- a/docs/how_to/work_with_schedules/sg_execution_times.html
+++ b/docs/how_to/work_with_schedules/sg_execution_times.html
@@ -340,7 +340,7 @@
<div class="section" id="computation-times">
<span id="sphx-glr-how-to-work-with-schedules-sg-execution-times"></span><h1>Computation times<a class="headerlink" href="#computation-times" title="Permalink to this headline">¶</a></h1>
-<p><strong>00:07.592</strong> total execution time for <strong>how_to_work_with_schedules</strong> files:</p>
+<p><strong>00:07.523</strong> total execution time for <strong>how_to_work_with_schedules</strong> files:</p>
<table class="docutils align-default">
<colgroup>
<col style="width: 83%" />
@@ -349,19 +349,19 @@
</colgroup>
<tbody>
<tr class="row-odd"><td><p><a class="reference internal" href="intrin_math.html#sphx-glr-how-to-work-with-schedules-intrin-math-py"><span class="std std-ref">Intrinsics and Math Functions</span></a> (<code class="docutils literal notranslate"><span class="pre">intrin_math.py</span></code>)</p></td>
-<td><p>00:05.077</p></td>
+<td><p>00:05.064</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="tensorize.html#sphx-glr-how-to-work-with-schedules-tensorize-py"><span class="std std-ref">Use Tensorize to Leverage Hardware Intrinsics</span></a> (<code class="docutils literal notranslate"><span class="pre">tensorize.py</span></code>)</p></td>
-<td><p>00:01.155</p></td>
+<td><p>00:01.119</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="reduction.html#sphx-glr-how-to-work-with-schedules-reduction-py"><span class="std std-ref">Reduction</span></a> (<code class="docutils literal notranslate"><span class="pre">reduction.py</span></code>)</p></td>
-<td><p>00:00.577</p></td>
+<td><p>00:00.570</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="scan.html#sphx-glr-how-to-work-with-schedules-scan-py"><span class="std std-ref">Scan and Recurrent Kernel</span></a> (<code class="docutils literal notranslate"><span class="pre">scan.py</span></code>)</p></td>
-<td><p>00:00.562</p></td>
+<td><p>00:00.549</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="extern_op.html#sphx-glr-how-to-work-with-schedules-extern-op-py"><span class="std std-ref">External Tensor Functions</span></a> (<code class="docutils literal notranslate"><span class="pre">extern_op.py</span></code>)</p></td>
@@ -369,11 +369,11 @@
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="schedule_primitives.html#sphx-glr-how-to-work-with-schedules-schedule-primitives-py"><span class="std std-ref">Schedule Primitives in TVM</span></a> (<code class="docutils literal notranslate"><span class="pre">schedule_primitives.py</span></code>)</p></td>
-<td><p>00:00.050</p></td>
+<td><p>00:00.049</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="tedd.html#sphx-glr-how-to-work-with-schedules-tedd-py"><span class="std std-ref">Use Tensor Expression Debug Display (TEDD) for Visualization</span></a> (<code class="docutils literal notranslate"><span class="pre">tedd.py</span></code>)</p></td>
-<td><p>00:00.032</p></td>
+<td><p>00:00.031</p></td>
<td><p>0.0 MB</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="tuple_inputs.html#sphx-glr-how-to-work-with-schedules-tuple-inputs-py"><span class="std std-ref">Compute and Reduce with Tuple Inputs</span></a> (<code class="docutils literal notranslate"><span class="pre">tuple_inputs.py</span></code>)</p></td>
diff --git a/docs/how_to/work_with_schedules/tensorize.html b/docs/how_to/work_with_schedules/tensorize.html
index cc92110685..13a699c5cc 100644
--- a/docs/how_to/work_with_schedules/tensorize.html
+++ b/docs/how_to/work_with_schedules/tensorize.html
@@ -418,19 +418,16 @@ The following lines describe the computation <code class="code docutils literal
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 64), "float32"), B: T.Buffer((512, 64), "float32"), C: T.Buffer((1024, 512), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 64))
- B_1 = T.match_buffer(B, (512, 64))
- C_1 = T.match_buffer(C, (1024, 512))
for i, j in T.grid(1024, 512):
- C_2 = T.buffer_decl((524288,), data=C_1.data)
- C_2[i * 512 + j] = T.float32(0)
+ C_1 = T.buffer_decl((524288,), data=C.data)
+ C_1[i * 512 + j] = T.float32(0)
for k in range(64):
cse_var_1: T.int32 = i * 512 + j
- A_2 = T.buffer_decl((65536,), data=A_1.data)
- B_2 = T.buffer_decl((32768,), data=B_1.data)
- C_2[cse_var_1] = C_2[cse_var_1] + A_2[i * 64 + k] * B_2[j * 64 + k]
+ A_1 = T.buffer_decl((65536,), data=A.data)
+ B_1 = T.buffer_decl((32768,), data=B.data)
+ C_1[cse_var_1] = C_1[cse_var_1] + A_1[i * 64 + k] * B_1[j * 64 + k]
</pre></div>
</div>
</div>
@@ -452,19 +449,16 @@ Thus we break down the matmul loops to make the innermost loops a (16x64) GEMV.<
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 64), "float32"), B: T.Buffer((512, 64), "float32"), C: T.Buffer((1024, 512), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 64))
- B_1 = T.match_buffer(B, (512, 64))
- C_1 = T.match_buffer(C, (1024, 512))
for i, j_outer, j_inner in T.grid(1024, 32, 16):
- C_2 = T.buffer_decl((524288,), data=C_1.data)
- C_2[i * 512 + j_outer * 16 + j_inner] = T.float32(0)
+ C_1 = T.buffer_decl((524288,), data=C.data)
+ C_1[i * 512 + j_outer * 16 + j_inner] = T.float32(0)
for k in range(64):
cse_var_1: T.int32 = i * 512 + j_outer * 16 + j_inner
- A_2 = T.buffer_decl((65536,), data=A_1.data)
- B_2 = T.buffer_decl((32768,), data=B_1.data)
- C_2[cse_var_1] = C_2[cse_var_1] + A_2[i * 64 + k] * B_2[j_outer * 1024 + j_inner * 64 + k]
+ A_1 = T.buffer_decl((65536,), data=A.data)
+ B_1 = T.buffer_decl((32768,), data=B.data)
+ C_1[cse_var_1] = C_1[cse_var_1] + A_1[i * 64 + k] * B_1[j_outer * 1024 + j_inner * 64 + k]
</pre></div>
</div>
<p>As showed in the IR printed above,
@@ -538,13 +532,10 @@ such placeholder can be put to let TVM automatically bind the inferred value for
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 64), "float32"), B: T.Buffer((512, 64), "float32"), C: T.Buffer((1024, 512), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 64))
- B_1 = T.match_buffer(B, (512, 64))
- C_1 = T.match_buffer(C, (1024, 512))
for i, j_outer in T.grid(1024, 32):
- T.call_extern("int32", "gemv_update", T.tvm_access_ptr(T.type_annotation("float32"), C_1.data, i * 512 + j_outer * 16, 16, 2), T.tvm_access_ptr(T.type_annotation("float32"), A_1.data, i * 64, 64, 1), T.tvm_access_ptr(T.type_annotation("float32"), B_1.data, j_outer * 1024, 1024, 1), 16, 64, 64)
+ T.call_extern("int32", "gemv_update", T.tvm_access_ptr(T.type_annotation("float32"), C.data, i * 512 + j_outer * 16, 16, 2), T.tvm_access_ptr(T.type_annotation("float32"), A.data, i * 64, 64, 1), T.tvm_access_ptr(T.type_annotation("float32"), B.data, j_outer * 1024, 1024, 1), 16, 64, 64)
</pre></div>
</div>
<p>By tensorizing over <code class="code docutils literal notranslate"><span class="pre">yi</span></code>, the inner most two loops are
@@ -580,15 +571,12 @@ The importing needs to happen before the tensorized GEMV being executed.</p>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>@I.ir_module
class Module:
@T.prim_func
- def main(A: T.handle, B: T.handle, C: T.handle):
+ def main(A: T.Buffer((1024, 64), "float32"), B: T.Buffer((512, 64), "float32"), C: T.Buffer((1024, 512), "float32")):
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
- A_1 = T.match_buffer(A, (1024, 64))
- B_1 = T.match_buffer(B, (512, 64))
- C_1 = T.match_buffer(C, (1024, 512))
i = T.var("int32")
- T.attr(T.iter_var(i, None, "DataPar", ""), "pragma_import_llvm", "; ModuleID = '/tmp/tmpgb2rg8u1/input0.cc'\nsource_filename = \"/tmp/tmpgb2rg8u1/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 [...]
+ T.attr(T.iter_var(i, None, "DataPar", ""), "pragma_import_llvm", "; ModuleID = '/tmp/tmpqmz888xx/input0.cc'\nsource_filename = \"/tmp/tmpqmz888xx/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 [...]
for i, j_outer in T.grid(1024, 32):
- T.call_extern("int32", "gemv_update", T.tvm_access_ptr(T.type_annotation("float32"), C_1.data, i * 512 + j_outer * 16, 16, 2), T.tvm_access_ptr(T.type_annotation("float32"), A_1.data, i * 64, 64, 1), T.tvm_access_ptr(T.type_annotation("float32"), B_1.data, j_outer * 1024, 1024, 1), 16, 64, 64)
+ T.call_extern("int32", "gemv_update", T.tvm_access_ptr(T.type_annotation("float32"), C.data, i * 512 + j_outer * 16, 16, 2), T.tvm_access_ptr(T.type_annotation("float32"), A.data, i * 64, 64, 1), T.tvm_access_ptr(T.type_annotation("float32"), B.data, j_outer * 1024, 1024, 1), 16, 64, 64)
</pre></div>
</div>
<p>Finally we compare the tensorize version with that <code class="code docutils literal notranslate"><span class="pre">numpy.dot</span></code> produces,
diff --git a/docs/install/nnpack.html b/docs/install/nnpack.html
index 1ef28de467..23d2181e9d 100644
--- a/docs/install/nnpack.html
+++ b/docs/install/nnpack.html
@@ -229,7 +229,17 @@
<p class="caption" role="heading"><span class="caption-text">Getting Started</span></p>
<ul class="current">
<li class="toctree-l1 current"><a class="reference internal" href="index.html">Installing TVM</a><ul class="current">
-<li class="toctree-l2"><a class="reference internal" href="from_source.html">Install from Source</a></li>
+<li class="toctree-l2 current"><a class="reference internal" href="from_source.html">Install from Source</a><ul class="current">
+<li class="toctree-l3"><a class="reference internal" href="from_source.html#developers-get-source-from-github">Developers: Get Source from Github</a></li>
+<li class="toctree-l3"><a class="reference internal" href="from_source.html#build-the-shared-library">Build the Shared Library</a></li>
+<li class="toctree-l3"><a class="reference internal" href="from_source.html#python-package-installation">Python Package Installation</a></li>
+<li class="toctree-l3 current"><a class="reference internal" href="from_source.html#install-contrib-libraries">Install Contrib Libraries</a><ul class="current">
+<li class="toctree-l4 current"><a class="current reference internal" href="#">NNPACK Contrib Installation</a></li>
+</ul>
+</li>
+<li class="toctree-l3"><a class="reference internal" href="from_source.html#enable-c-tests">Enable C++ Tests</a></li>
+</ul>
+</li>
<li class="toctree-l2"><a class="reference internal" href="docker.html">Docker Images</a></li>
<li class="toctree-l2 current"><a class="current reference internal" href="#">NNPACK Contrib Installation</a><ul>
<li class="toctree-l3"><a class="reference internal" href="#conditions">Conditions</a></li>
diff --git a/docs/reference/api/python/auto_scheduler.html b/docs/reference/api/python/auto_scheduler.html
index ea853c881c..6cd526ae9b 100644
--- a/docs/reference/api/python/auto_scheduler.html
+++ b/docs/reference/api/python/auto_scheduler.html
@@ -1615,7 +1615,7 @@ history states as starting point to perform Evolutionary Search).</p></li>
<dl class="py class">
<dt class="sig sig-object py" id="tvm.auto_scheduler.SketchPolicy">
-<em class="property"><span class="pre">class</span> </em><span class="sig-prename descclassname"><span class="pre">tvm.auto_scheduler.</span></span><span class="sig-name descname"><span class="pre">SketchPolicy</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">task</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">program_cost_model</span></span><span class="o"><span class="pre">=</span></span><span class="defau [...]
+<em class="property"><span class="pre">class</span> </em><span class="sig-prename descclassname"><span class="pre">tvm.auto_scheduler.</span></span><span class="sig-name descname"><span class="pre">SketchPolicy</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">task</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">program_cost_model</span></span><span class="o"><span class="pre">=</span></span><span class="defau [...]
<dd><p>The search policy that searches in a hierarchical search space defined by sketches.
The policy randomly samples programs from the space defined by sketches and use evolutionary
search to fine-tune them.</p>
@@ -1899,7 +1899,7 @@ Candidates:
<dl class="py function">
<dt class="sig sig-object py" id="tvm.auto_scheduler.auto_schedule">
-<span class="sig-prename descclassname"><span class="pre">tvm.auto_scheduler.</span></span><span class="sig-name descname"><span class="pre">auto_schedule</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">task</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">search_policy</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em clas [...]
+<span class="sig-prename descclassname"><span class="pre">tvm.auto_scheduler.</span></span><span class="sig-name descname"><span class="pre">auto_schedule</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">task</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">search_policy</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em clas [...]
<dd><p>THIS API IS DEPRECATED.</p>
<p>Run auto scheduling search for a task.</p>
<dl class="field-list simple">
diff --git a/docs/reference/api/typedoc/classes/bytestreamreader.html b/docs/reference/api/typedoc/classes/bytestreamreader.html
index 280a098e18..1115707070 100644
--- a/docs/reference/api/typedoc/classes/bytestreamreader.html
+++ b/docs/reference/api/typedoc/classes/bytestreamreader.html
@@ -119,7 +119,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/rpc_server.ts#L43">rpc_server.ts:43</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/rpc_server.ts#L43">rpc_server.ts:43</a></li>
</ul>
</aside>
<h4 class="tsd-parameters-title">Parameters</h4>
@@ -141,7 +141,7 @@
<div class="tsd-signature tsd-kind-icon">bytes<span class="tsd-signature-symbol">:</span> <span class="tsd-signature-type">Uint8Array</span></div>
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/rpc_server.ts#L43">rpc_server.ts:43</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/rpc_server.ts#L43">rpc_server.ts:43</a></li>
</ul>
</aside>
</section>
@@ -151,7 +151,7 @@
<div class="tsd-signature tsd-kind-icon">offset<span class="tsd-signature-symbol">:</span> <span class="tsd-signature-type">number</span><span class="tsd-signature-symbol"> = 0</span></div>
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/rpc_server.ts#L42">rpc_server.ts:42</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/rpc_server.ts#L42">rpc_server.ts:42</a></li>
</ul>
</aside>
</section>
@@ -168,7 +168,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/rpc_server.ts#L63">rpc_server.ts:63</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/rpc_server.ts#L63">rpc_server.ts:63</a></li>
</ul>
</aside>
<h4 class="tsd-returns-title">Returns <span class="tsd-signature-type">Uint8Array</span></h4>
@@ -185,7 +185,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/rpc_server.ts#L49">rpc_server.ts:49</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/rpc_server.ts#L49">rpc_server.ts:49</a></li>
</ul>
</aside>
<h4 class="tsd-returns-title">Returns <span class="tsd-signature-type">number</span></h4>
@@ -202,7 +202,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/rpc_server.ts#L57">rpc_server.ts:57</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/rpc_server.ts#L57">rpc_server.ts:57</a></li>
</ul>
</aside>
<h4 class="tsd-returns-title">Returns <span class="tsd-signature-type">number</span></h4>
diff --git a/docs/reference/api/typedoc/classes/cachedcallstack.html b/docs/reference/api/typedoc/classes/cachedcallstack.html
index ca7755865b..efc5a8a758 100644
--- a/docs/reference/api/typedoc/classes/cachedcallstack.html
+++ b/docs/reference/api/typedoc/classes/cachedcallstack.html
@@ -144,7 +144,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/memory.ts#L223">memory.ts:223</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/memory.ts#L223">memory.ts:223</a></li>
</ul>
</aside>
<h4 class="tsd-parameters-title">Parameters</h4>
@@ -172,7 +172,7 @@
<div class="tsd-signature tsd-kind-icon">temp<wbr>Args<span class="tsd-signature-symbol">:</span> <span class="tsd-signature-type">Array</span><span class="tsd-signature-symbol"><</span><a href="../interfaces/disposable.html" class="tsd-signature-type">Disposable</a><span class="tsd-signature-symbol">></span><span class="tsd-signature-symbol"> = []</span></div>
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/memory.ts#L208">memory.ts:208</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/memory.ts#L208">memory.ts:208</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -194,7 +194,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/memory.ts#L312">memory.ts:312</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/memory.ts#L312">memory.ts:312</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -226,7 +226,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/memory.ts#L284">memory.ts:284</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/memory.ts#L284">memory.ts:284</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -262,7 +262,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/memory.ts#L388">memory.ts:388</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/memory.ts#L388">memory.ts:388</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -300,7 +300,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/memory.ts#L376">memory.ts:376</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/memory.ts#L376">memory.ts:376</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -340,7 +340,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/memory.ts#L267">memory.ts:267</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/memory.ts#L267">memory.ts:267</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -373,7 +373,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/memory.ts#L243">memory.ts:243</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/memory.ts#L243">memory.ts:243</a></li>
</ul>
</aside>
<h4 class="tsd-returns-title">Returns <span class="tsd-signature-type">void</span></h4>
@@ -390,7 +390,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/memory.ts#L321">memory.ts:321</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/memory.ts#L321">memory.ts:321</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -422,7 +422,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/memory.ts#L252">memory.ts:252</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/memory.ts#L252">memory.ts:252</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -444,7 +444,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/memory.ts#L359">memory.ts:359</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/memory.ts#L359">memory.ts:359</a></li>
</ul>
</aside>
<h4 class="tsd-parameters-title">Parameters</h4>
@@ -470,7 +470,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/memory.ts#L342">memory.ts:342</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/memory.ts#L342">memory.ts:342</a></li>
</ul>
</aside>
<h4 class="tsd-parameters-title">Parameters</h4>
@@ -496,7 +496,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/memory.ts#L350">memory.ts:350</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/memory.ts#L350">memory.ts:350</a></li>
</ul>
</aside>
<h4 class="tsd-parameters-title">Parameters</h4>
@@ -522,7 +522,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/memory.ts#L326">memory.ts:326</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/memory.ts#L326">memory.ts:326</a></li>
</ul>
</aside>
<h4 class="tsd-parameters-title">Parameters</h4>
@@ -548,7 +548,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/memory.ts#L363">memory.ts:363</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/memory.ts#L363">memory.ts:363</a></li>
</ul>
</aside>
<h4 class="tsd-parameters-title">Parameters</h4>
@@ -574,7 +574,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/memory.ts#L346">memory.ts:346</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/memory.ts#L346">memory.ts:346</a></li>
</ul>
</aside>
<h4 class="tsd-parameters-title">Parameters</h4>
@@ -600,7 +600,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/memory.ts#L334">memory.ts:334</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/memory.ts#L334">memory.ts:334</a></li>
</ul>
</aside>
<h4 class="tsd-parameters-title">Parameters</h4>
diff --git a/docs/reference/api/typedoc/classes/dldatatype.html b/docs/reference/api/typedoc/classes/dldatatype.html
index 5834071eb2..a373e77471 100644
--- a/docs/reference/api/typedoc/classes/dldatatype.html
+++ b/docs/reference/api/typedoc/classes/dldatatype.html
@@ -119,7 +119,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L262">runtime.ts:262</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L262">runtime.ts:262</a></li>
</ul>
</aside>
<h4 class="tsd-parameters-title">Parameters</h4>
@@ -147,7 +147,7 @@
<div class="tsd-signature tsd-kind-icon">bits<span class="tsd-signature-symbol">:</span> <span class="tsd-signature-type">number</span></div>
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L260">runtime.ts:260</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L260">runtime.ts:260</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -162,7 +162,7 @@
<div class="tsd-signature tsd-kind-icon">code<span class="tsd-signature-symbol">:</span> <span class="tsd-signature-type">number</span></div>
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L258">runtime.ts:258</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L258">runtime.ts:258</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -177,7 +177,7 @@
<div class="tsd-signature tsd-kind-icon">lanes<span class="tsd-signature-symbol">:</span> <span class="tsd-signature-type">number</span></div>
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L262">runtime.ts:262</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L262">runtime.ts:262</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -199,7 +199,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L279">runtime.ts:279</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L279">runtime.ts:279</a></li>
</ul>
</aside>
<h4 class="tsd-returns-title">Returns <span class="tsd-signature-type">number</span></h4>
@@ -216,7 +216,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L270">runtime.ts:270</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L270">runtime.ts:270</a></li>
</ul>
</aside>
<h4 class="tsd-returns-title">Returns <span class="tsd-signature-type">string</span></h4>
diff --git a/docs/reference/api/typedoc/classes/dldevice.html b/docs/reference/api/typedoc/classes/dldevice.html
index c9e08f7ad0..79ed400577 100644
--- a/docs/reference/api/typedoc/classes/dldevice.html
+++ b/docs/reference/api/typedoc/classes/dldevice.html
@@ -118,7 +118,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L202">runtime.ts:202</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L202">runtime.ts:202</a></li>
</ul>
</aside>
<h4 class="tsd-parameters-title">Parameters</h4>
@@ -146,7 +146,7 @@
<div class="tsd-signature tsd-kind-icon">device<wbr>Id<span class="tsd-signature-symbol">:</span> <span class="tsd-signature-type">number</span></div>
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L200">runtime.ts:200</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L200">runtime.ts:200</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -161,7 +161,7 @@
<div class="tsd-signature tsd-kind-icon">device<wbr>Type<span class="tsd-signature-symbol">:</span> <span class="tsd-signature-type">number</span></div>
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L198">runtime.ts:198</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L198">runtime.ts:198</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -183,7 +183,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L223">runtime.ts:223</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L223">runtime.ts:223</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -205,7 +205,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L230">runtime.ts:230</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L230">runtime.ts:230</a></li>
</ul>
</aside>
<h4 class="tsd-returns-title">Returns <span class="tsd-signature-type">string</span></h4>
diff --git a/docs/reference/api/typedoc/classes/environment.html b/docs/reference/api/typedoc/classes/environment.html
index 62b18ccb78..07941173b7 100644
--- a/docs/reference/api/typedoc/classes/environment.html
+++ b/docs/reference/api/typedoc/classes/environment.html
@@ -125,7 +125,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/environment.ts#L86">environment.ts:86</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/environment.ts#L86">environment.ts:86</a></li>
</ul>
</aside>
<h4 class="tsd-parameters-title">Parameters</h4>
@@ -169,7 +169,7 @@
<aside class="tsd-sources">
<p>Implementation of <a href="../interfaces/libraryprovider.html">LibraryProvider</a>.<a href="../interfaces/libraryprovider.html#imports">imports</a></p>
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/environment.ts#L70">environment.ts:70</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/environment.ts#L70">environment.ts:70</a></li>
</ul>
</aside>
</section>
@@ -179,7 +179,7 @@
<div class="tsd-signature tsd-kind-icon">logger<span class="tsd-signature-symbol">:</span> <span class="tsd-signature-symbol">(</span>msg<span class="tsd-signature-symbol">: </span><span class="tsd-signature-type">string</span><span class="tsd-signature-symbol">)</span><span class="tsd-signature-symbol"> => </span><span class="tsd-signature-type">void</span></div>
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/environment.ts#L69">environment.ts:69</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/environment.ts#L69">environment.ts:69</a></li>
</ul>
</aside>
<div class="tsd-type-declaration">
@@ -210,7 +210,7 @@
<div class="tsd-signature tsd-kind-icon">packedCFunc<wbr>Table<span class="tsd-signature-symbol">:</span> <span class="tsd-signature-type">Array</span><span class="tsd-signature-symbol"><</span><span class="tsd-signature-type">ctypes.FTVMWasmPackedCFunc</span><span class="tsd-signature-symbol"> | </span><span class="tsd-signature-type">undefined</span><span class="tsd-signature-symbol">></span><span class="tsd-signature-symbol"> = [undefined,]</span></div>
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/environment.ts#L78">environment.ts:78</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/environment.ts#L78">environment.ts:78</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -228,7 +228,7 @@
<div class="tsd-signature tsd-kind-icon">packedCFunc<wbr>Table<wbr>Free<wbr>Id<span class="tsd-signature-symbol">:</span> <span class="tsd-signature-type">Array</span><span class="tsd-signature-symbol"><</span><span class="tsd-signature-type">number</span><span class="tsd-signature-symbol">></span><span class="tsd-signature-symbol"> = []</span></div>
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/environment.ts#L84">environment.ts:84</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/environment.ts#L84">environment.ts:84</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -250,7 +250,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/environment.ts#L105">environment.ts:105</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/environment.ts#L105">environment.ts:105</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
diff --git a/docs/reference/api/typedoc/classes/ffilibrary.html b/docs/reference/api/typedoc/classes/ffilibrary.html
index 65ef328348..489ac4c338 100644
--- a/docs/reference/api/typedoc/classes/ffilibrary.html
+++ b/docs/reference/api/typedoc/classes/ffilibrary.html
@@ -131,7 +131,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L49">runtime.ts:49</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L49">runtime.ts:49</a></li>
</ul>
</aside>
<h4 class="tsd-parameters-title">Parameters</h4>
@@ -156,7 +156,7 @@
<div class="tsd-signature tsd-kind-icon">exports<span class="tsd-signature-symbol">:</span> <span class="tsd-signature-type">Record</span><span class="tsd-signature-symbol"><</span><span class="tsd-signature-type">string</span><span class="tsd-signature-symbol">, </span><span class="tsd-signature-type">Function</span><span class="tsd-signature-symbol">></span></div>
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L46">runtime.ts:46</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L46">runtime.ts:46</a></li>
</ul>
</aside>
</section>
@@ -166,7 +166,7 @@
<div class="tsd-signature tsd-kind-icon">memory<span class="tsd-signature-symbol">:</span> <a href="memory.html" class="tsd-signature-type">Memory</a></div>
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L45">runtime.ts:45</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L45">runtime.ts:45</a></li>
</ul>
</aside>
</section>
@@ -176,7 +176,7 @@
<div class="tsd-signature tsd-kind-icon">wasm32<span class="tsd-signature-symbol">:</span> <span class="tsd-signature-type">boolean</span></div>
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L44">runtime.ts:44</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L44">runtime.ts:44</a></li>
</ul>
</aside>
</section>
@@ -186,7 +186,7 @@
<div class="tsd-signature tsd-kind-icon">webGPUContext<span class="tsd-signature-symbol">:</span> <a href="webgpucontext.html" class="tsd-signature-type">WebGPUContext</a></div>
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L47">runtime.ts:47</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L47">runtime.ts:47</a></li>
</ul>
</aside>
</section>
@@ -203,7 +203,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L76">runtime.ts:76</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L76">runtime.ts:76</a></li>
</ul>
</aside>
<h4 class="tsd-parameters-title">Parameters</h4>
@@ -226,7 +226,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L66">runtime.ts:66</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L66">runtime.ts:66</a></li>
</ul>
</aside>
<h4 class="tsd-returns-title">Returns <span class="tsd-signature-type">void</span></h4>
@@ -243,7 +243,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L84">runtime.ts:84</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L84">runtime.ts:84</a></li>
</ul>
</aside>
<h4 class="tsd-returns-title">Returns <a href="cachedcallstack.html" class="tsd-signature-type">CachedCallStack</a></h4>
@@ -260,7 +260,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L95">runtime.ts:95</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L95">runtime.ts:95</a></li>
</ul>
</aside>
<h4 class="tsd-parameters-title">Parameters</h4>
@@ -283,7 +283,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L72">runtime.ts:72</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L72">runtime.ts:72</a></li>
</ul>
</aside>
<h4 class="tsd-returns-title">Returns <span class="tsd-signature-type">number</span></h4>
diff --git a/docs/reference/api/typedoc/classes/graphexecutor.html b/docs/reference/api/typedoc/classes/graphexecutor.html
index 32703d3f5d..3b5fa3c8cf 100644
--- a/docs/reference/api/typedoc/classes/graphexecutor.html
+++ b/docs/reference/api/typedoc/classes/graphexecutor.html
@@ -130,7 +130,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L583">runtime.ts:583</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L583">runtime.ts:583</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -162,7 +162,7 @@
<div class="tsd-signature tsd-kind-icon">module<span class="tsd-signature-symbol">:</span> <a href="module.html" class="tsd-signature-type">Module</a></div>
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L579">runtime.ts:579</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L579">runtime.ts:579</a></li>
</ul>
</aside>
</section>
@@ -179,7 +179,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L654">runtime.ts:654</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L654">runtime.ts:654</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -224,7 +224,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L597">runtime.ts:597</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L597">runtime.ts:597</a></li>
</ul>
</aside>
<h4 class="tsd-returns-title">Returns <span class="tsd-signature-type">void</span></h4>
@@ -241,7 +241,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L631">runtime.ts:631</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L631">runtime.ts:631</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -279,7 +279,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L644">runtime.ts:644</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L644">runtime.ts:644</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -310,7 +310,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L621">runtime.ts:621</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L621">runtime.ts:621</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -332,7 +332,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L609">runtime.ts:609</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L609">runtime.ts:609</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
diff --git a/docs/reference/api/typedoc/classes/instance.html b/docs/reference/api/typedoc/classes/instance.html
index 034ace3eb8..7c70323483 100644
--- a/docs/reference/api/typedoc/classes/instance.html
+++ b/docs/reference/api/typedoc/classes/instance.html
@@ -139,7 +139,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L692">runtime.ts:692</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L692">runtime.ts:692</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -202,7 +202,7 @@
<div class="tsd-signature tsd-kind-icon">exports<span class="tsd-signature-symbol">:</span> <span class="tsd-signature-type">Record</span><span class="tsd-signature-symbol"><</span><span class="tsd-signature-type">string</span><span class="tsd-signature-symbol">, </span><span class="tsd-signature-type">Function</span><span class="tsd-signature-symbol">></span></div>
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L684">runtime.ts:684</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L684">runtime.ts:684</a></li>
</ul>
</aside>
</section>
@@ -212,7 +212,7 @@
<div class="tsd-signature tsd-kind-icon">memory<span class="tsd-signature-symbol">:</span> <a href="memory.html" class="tsd-signature-type">Memory</a></div>
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L683">runtime.ts:683</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L683">runtime.ts:683</a></li>
</ul>
</aside>
</section>
@@ -229,7 +229,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L932">runtime.ts:932</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L932">runtime.ts:932</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -260,7 +260,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L994">runtime.ts:994</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L994">runtime.ts:994</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -303,7 +303,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L924">runtime.ts:924</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L924">runtime.ts:924</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -341,7 +341,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L732">runtime.ts:732</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L732">runtime.ts:732</a></li>
</ul>
</aside>
<h4 class="tsd-returns-title">Returns <span class="tsd-signature-type">void</span></h4>
@@ -358,7 +358,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L952">runtime.ts:952</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L952">runtime.ts:952</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -402,7 +402,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L816">runtime.ts:816</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L816">runtime.ts:816</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -434,7 +434,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L1033">runtime.ts:1033</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L1033">runtime.ts:1033</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -465,7 +465,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L846">runtime.ts:846</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L846">runtime.ts:846</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -497,7 +497,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L750">runtime.ts:750</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L750">runtime.ts:750</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -520,7 +520,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L1013">runtime.ts:1013</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L1013">runtime.ts:1013</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -568,7 +568,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L789">runtime.ts:789</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L789">runtime.ts:789</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -608,7 +608,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L914">runtime.ts:914</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L914">runtime.ts:914</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -646,7 +646,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L1145">runtime.ts:1145</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L1145">runtime.ts:1145</a></li>
</ul>
</aside>
<div class="tsd-comment tsd-typography">
@@ -698,7 +698,7 @@
<li class="tsd-description">
<aside class="tsd-sources">
<ul>
- <li>Defined in <a href="https://github.com/apache/tvm/blob/cfa65b26c/web/src/runtime.ts#L740">runtime.ts:740</a></li>
+ <li>Defined in <a href="https://github.com/apache/tvm/blob/6c2d485a0/web/src/runtime.ts#L740">runtime.ts:740</a></li>
</ul>
... 3193 lines suppressed ...