You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@tvm.apache.org by GitBox <gi...@apache.org> on 2021/03/06 03:41:17 UTC
[GitHub] [tvm] FrozenGene commented on a change in pull request #7299: Introduce Apple BNNS backend
FrozenGene commented on a change in pull request #7299:
URL: https://github.com/apache/tvm/pull/7299#discussion_r588820469
##########
File path: docs/deploy/bnns.rst
##########
@@ -0,0 +1,183 @@
+.. Licensed to the Apache Software Foundation (ASF) under one
+ or more contributor license agreements. See the NOTICE file
+ distributed with this work for additional information
+ regarding copyright ownership. The ASF licenses this file
+ to you under the Apache License, Version 2.0 (the
+ "License"); you may not use this file except in compliance
+ with the License. You may obtain a copy of the License at
+
+.. http://www.apache.org/licenses/LICENSE-2.0
+
+.. Unless required by applicable law or agreed to in writing,
+ software distributed under the License is distributed on an
+ "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ KIND, either express or implied. See the License for the
+ specific language governing permissions and limitations
+ under the License.
+
+Relay BNNS Integration
+==========================
+**Author**: `Egor Churaev <https://github.com/echuraev>`_
+
+Introduction
+------------
+
+Apple BNNS library is a collection of functions that can be used to construct neural networks
+for inference (and train). It’s supported in macOS, iOS, tvOS, and watchOS. BNNS provides
+primitives executed on all CPU supported on those platforms and optimized for high performance
+and low-energy consumption. This integration will offload as many operators as possible from Relay to BNNS.
+
+BNNS runtime is a part of platform API and available on all modern Apple operating systems.
+Application using BNNS will not depends on any additional external dependencies.
+
+BNNS functions uses Apple private hardware capabilities which are not exposed yet by Apple. Example
+of such capabilities can be AMX Apple cpu extension.
+
+This guide will demonstrate how to build TVM with BNNS codegen and runtime enabled. It will also provide example
+code to compile and run models using BNNS runtime. Finally, we document the supported operators.
+
+Building TVM with BNNS support
+----------------------------------
+
+To turn on TVM BNNS codegen and TVM BNNS runtime you need to turn on the only USE_BNNS flag
+
+* USE_BNNS=ON/OFF - This flag will enable compiling a network with offloading subgraphs to BNNS primitives
+ and will link tvm library to the BNNS runtime module.
+
+Enabling of this flag will cause to search the default Accelerate Frameworks on current target SDK.
+The minimal versions of required SDK is macOS 11.0, iOS 14.0, tvOS 14.0 and watchOS 7.0.
+
+Example setting in config.cmake file:
+
+.. code:: cmake
+
+ set(USE_BNNS ON)
+
+BNNS partitioning of Relay graph
+----------------------------------------
+
+Operations to be offloaded on BNNS execution must be annotated before passing of module for compilation.
+All opps annotated by `partition_for_bnns` will be offloaded for BNNS execution. The rest of the ops
+will go through the LLVM compilation and code generation.
+
+Important note: BNNS support primitives only with constant weights. To satisfy this requirements we have
+to map constants to related tensor abstraction in relay representation. To freeze tensors and operate
+with them as with constants you may need to call ONNX importer with special flag "freeze_params=True"
Review comment:
as with -> as
##########
File path: python/tvm/relay/op/contrib/bnns.py
##########
@@ -0,0 +1,327 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+# pylint: disable=invalid-name, unused-argument
+"""BNNS library supported operators.
+Is a part of Accelerate framework on macOS/iOS platforms. Apple provide several APIs
+to handle tensor processing. Particularly:
+ * BNNS (basic neural )
+ * vDSP (1D and 2D tensor processing)
+"""
+import math
+import tvm.ir
+
+from tvm.relay import transform
+from tvm.relay.expr import const
+from tvm.relay.build_module import bind_params_by_name
+
+from .register import register_pattern_table, get_pattern_table
+from ...dataflow_pattern import wildcard, is_op, is_expr
+
+
+def partition_for_bnns(mod, params=None):
+ """Partition the graph greedily offloading supported
+ operators to BNNS.
+
+ Parameters
+ ----------
+ mod : Module
+ The module to run passes on.
+ params : Optional[Dict[str, NDArray]]
+ Constant input parameters.
+
+ Returns
+ -------
+ ret : annotated and partitioned module.
+ """
+ if params:
+ mod["main"] = bind_params_by_name(mod["main"], params)
+
+ seq = tvm.transform.Sequential(
+ [
+ transform.InferType(),
+ transform.FoldConstant(),
+ transform.FoldScaleAxis(),
+ transform.DynamicToStatic(),
+ transform.AlterOpLayout(),
+ # TODO(apeskov): WA. AlterOpLayout call lead to constants shape transformation
+ # Some expand_dims op may appears after constants. It breaks BNNS fusing.
+ # So we have to call FoldConstant right before bnns composite passes.
+ transform.FoldConstant(),
+ transform.MergeComposite(get_pattern_table("bnns")),
+ transform.AnnotateTarget("bnns"),
+ # If you no need in per layer performance statistic you can
+ # uncomment next line
+ # transform.MergeCompilerRegions(),
Review comment:
Does it have effect on end to end network performance?
##########
File path: docs/deploy/bnns.rst
##########
@@ -0,0 +1,183 @@
+.. Licensed to the Apache Software Foundation (ASF) under one
+ or more contributor license agreements. See the NOTICE file
+ distributed with this work for additional information
+ regarding copyright ownership. The ASF licenses this file
+ to you under the Apache License, Version 2.0 (the
+ "License"); you may not use this file except in compliance
+ with the License. You may obtain a copy of the License at
+
+.. http://www.apache.org/licenses/LICENSE-2.0
+
+.. Unless required by applicable law or agreed to in writing,
+ software distributed under the License is distributed on an
+ "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ KIND, either express or implied. See the License for the
+ specific language governing permissions and limitations
+ under the License.
+
+Relay BNNS Integration
+==========================
+**Author**: `Egor Churaev <https://github.com/echuraev>`_
+
+Introduction
+------------
+
+Apple BNNS library is a collection of functions that can be used to construct neural networks
+for inference (and train). It’s supported in macOS, iOS, tvOS, and watchOS. BNNS provides
+primitives executed on all CPU supported on those platforms and optimized for high performance
+and low-energy consumption. This integration will offload as many operators as possible from Relay to BNNS.
+
+BNNS runtime is a part of platform API and available on all modern Apple operating systems.
+Application using BNNS will not depends on any additional external dependencies.
+
+BNNS functions uses Apple private hardware capabilities which are not exposed yet by Apple. Example
+of such capabilities can be AMX Apple cpu extension.
+
+This guide will demonstrate how to build TVM with BNNS codegen and runtime enabled. It will also provide example
+code to compile and run models using BNNS runtime. Finally, we document the supported operators.
+
+Building TVM with BNNS support
+----------------------------------
+
+To turn on TVM BNNS codegen and TVM BNNS runtime you need to turn on the only USE_BNNS flag
+
+* USE_BNNS=ON/OFF - This flag will enable compiling a network with offloading subgraphs to BNNS primitives
+ and will link tvm library to the BNNS runtime module.
+
+Enabling of this flag will cause to search the default Accelerate Frameworks on current target SDK.
+The minimal versions of required SDK is macOS 11.0, iOS 14.0, tvOS 14.0 and watchOS 7.0.
+
+Example setting in config.cmake file:
+
+.. code:: cmake
+
+ set(USE_BNNS ON)
+
+BNNS partitioning of Relay graph
+----------------------------------------
+
+Operations to be offloaded on BNNS execution must be annotated before passing of module for compilation.
+All opps annotated by `partition_for_bnns` will be offloaded for BNNS execution. The rest of the ops
+will go through the LLVM compilation and code generation.
+
+Important note: BNNS support primitives only with constant weights. To satisfy this requirements we have
+to map constants to related tensor abstraction in relay representation. To freeze tensors and operate
+with them as with constants you may need to call ONNX importer with special flag "freeze_params=True"
+or performer binding manually. In general cases all relay importers don't do that by default.
+For your convenience "partition_for_bnns" can do this for you if params dictionary is passed as the argument.
+
+.. code:: python
+
+ from tvm.relay.op.contrib.bnns import partition_for_bnns
+ model = partition_for_bnns(model, params=params)
+
+
+Input data layout for operations to be offloaded to BNNS execution
+----------------------------------------
Review comment:
I wonder whether we have problem if we don't align `--------` with words, other places have same problem
##########
File path: docs/deploy/bnns.rst
##########
@@ -0,0 +1,183 @@
+.. Licensed to the Apache Software Foundation (ASF) under one
+ or more contributor license agreements. See the NOTICE file
+ distributed with this work for additional information
+ regarding copyright ownership. The ASF licenses this file
+ to you under the Apache License, Version 2.0 (the
+ "License"); you may not use this file except in compliance
+ with the License. You may obtain a copy of the License at
+
+.. http://www.apache.org/licenses/LICENSE-2.0
+
+.. Unless required by applicable law or agreed to in writing,
+ software distributed under the License is distributed on an
+ "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ KIND, either express or implied. See the License for the
+ specific language governing permissions and limitations
+ under the License.
+
+Relay BNNS Integration
+==========================
+**Author**: `Egor Churaev <https://github.com/echuraev>`_
+
+Introduction
+------------
+
+Apple BNNS library is a collection of functions that can be used to construct neural networks
+for inference (and train). It’s supported in macOS, iOS, tvOS, and watchOS. BNNS provides
+primitives executed on all CPU supported on those platforms and optimized for high performance
+and low-energy consumption. This integration will offload as many operators as possible from Relay to BNNS.
+
+BNNS runtime is a part of platform API and available on all modern Apple operating systems.
+Application using BNNS will not depends on any additional external dependencies.
+
+BNNS functions uses Apple private hardware capabilities which are not exposed yet by Apple. Example
+of such capabilities can be AMX Apple cpu extension.
+
+This guide will demonstrate how to build TVM with BNNS codegen and runtime enabled. It will also provide example
+code to compile and run models using BNNS runtime. Finally, we document the supported operators.
+
+Building TVM with BNNS support
+----------------------------------
+
+To turn on TVM BNNS codegen and TVM BNNS runtime you need to turn on the only USE_BNNS flag
+
+* USE_BNNS=ON/OFF - This flag will enable compiling a network with offloading subgraphs to BNNS primitives
+ and will link tvm library to the BNNS runtime module.
+
+Enabling of this flag will cause to search the default Accelerate Frameworks on current target SDK.
+The minimal versions of required SDK is macOS 11.0, iOS 14.0, tvOS 14.0 and watchOS 7.0.
+
+Example setting in config.cmake file:
+
+.. code:: cmake
+
+ set(USE_BNNS ON)
+
+BNNS partitioning of Relay graph
+----------------------------------------
+
+Operations to be offloaded on BNNS execution must be annotated before passing of module for compilation.
+All opps annotated by `partition_for_bnns` will be offloaded for BNNS execution. The rest of the ops
Review comment:
opps -> ops
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org