Create a separate component for graph optimization. (#421)

* Create a project for graph optimizer.

Move optimizer related code to the folder optimizer.

* Fix build failures.

* rebase and fix build failures.

* fix build failure.

* fix build failure with cuda path.

* fix python build failure.

* Move two transformers(memcpy and insert_cast) from framework to optimizer.

* rebase.

* SessionState should not depend on optimizer.

cmake/CMakeLists.txt

@@ -513,6 +513,7 @@ include(onnxruntime_framework.cmake)
 include(onnxruntime_util.cmake)
 #The next file will define 'onnxruntime_libs' cmake var
 include(onnxruntime_providers.cmake)
+include(onnxruntime_optimizer.cmake)
 include(onnxruntime_session.cmake)
 include(onnxruntime_mlas.cmake)
 

cmake/onnxruntime.cmake

@@ -49,6 +49,7 @@ target_link_libraries(onnxruntime PRIVATE
     ${onnxruntime_libs}
     ${PROVIDERS_CUDA}
     ${PROVIDERS_MKLDNN}
+    onnxruntime_optimizer
     onnxruntime_providers    
     onnxruntime_util
     ${onnxruntime_tvm_libs}

cmake/onnxruntime_optimizer.cmake

@@ -0,0 +1,16 @@
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License.
+
+file(GLOB onnxruntime_optimizer_srcs
+    "${ONNXRUNTIME_ROOT}/core/optimizer/*.h"
+    "${ONNXRUNTIME_ROOT}/core/optimizer/*.cc"
+    )
+
+source_group(TREE ${REPO_ROOT} FILES ${onnxruntime_optimizer_srcs})
+
+add_library(onnxruntime_optimizer ${onnxruntime_optimizer_srcs})
+install(DIRECTORY ${PROJECT_SOURCE_DIR}/../include/onnxruntime/core/optimizer  DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/onnxruntime/core)
+onnxruntime_add_include_to_target(onnxruntime_optimizer onnxruntime_common onnxruntime_framework gsl onnx onnx_proto protobuf::libprotobuf)
+target_include_directories(onnxruntime_optimizer PRIVATE ${ONNXRUNTIME_ROOT})
+add_dependencies(onnxruntime_optimizer ${onnxruntime_EXTERNAL_DEPENDENCIES})
+set_target_properties(onnxruntime_optimizer PROPERTIES FOLDER "ONNXRuntime")

cmake/onnxruntime_python.cmake

@@ -58,6 +58,7 @@ set(onnxruntime_pybind11_state_libs
     ${onnxruntime_libs}
     ${PROVIDERS_CUDA}
     ${PROVIDERS_MKLDNN}
+    onnxruntime_optimizer
     onnxruntime_providers
     onnxruntime_util
     ${onnxruntime_tvm_libs}

cmake/onnxruntime_unittests.cmake

@@ -135,6 +135,7 @@ set(onnxruntime_test_ir_libs
 
 set(onnxruntime_test_framework_libs
   onnxruntime_test_utils_for_framework
+  onnxruntime_optimizer
   onnxruntime_framework
   onnxruntime_util
   onnxruntime_graph
@@ -175,6 +176,7 @@ set(ONNXRUNTIME_TEST_LIBS
     ${onnxruntime_libs}
     ${PROVIDERS_CUDA}
     ${PROVIDERS_MKLDNN}
+    onnxruntime_optimizer
     onnxruntime_providers
     onnxruntime_util
     ${onnxruntime_tvm_libs}

include/onnxruntime/core/optimizer/graph_transformer.h

@@ -5,7 +5,7 @@
 
 #include "core/common/common.h"
 #include "core/graph/graph_viewer.h"
-#include "core/graph/rewrite_rule.h"
+#include "core/optimizer/rewrite_rule.h"
 
 namespace onnxruntime {
 

onnxruntime/core/framework/session_state_initializer.cc

@@ -10,11 +10,7 @@
 #include "core/common/logging/logging.h"
 
 #include "core/graph/graph_viewer.h"
-#include "core/graph/graph_transformer.h"
-#include "core/graph/graph_transformer_mgr.h"
-
 #include "core/framework/graph_partitioner.h"
-#include "core/framework/insert_cast_transformer.h"
 #include "core/framework/ml_value.h"
 #include "core/framework/ml_value_patterns_planner.h"
 #include "core/framework/mlvalue_name_idx_map.h"
@@ -22,7 +18,6 @@
 #include "core/framework/session_state.h"
 #include "core/framework/tensorutils.h"
 #include "core/framework/tensorprotoutils.h"
-#include "core/framework/transformer_memcpy.h"
 #include "core/framework/utils.h"
 
 namespace onnxruntime {

onnxruntime/core/optimizer/conv_activation_fusion.cc

@@ -1,10 +1,10 @@
 // Copyright (c) Microsoft Corporation. All rights reserved.
 // Licensed under the MIT License.
 
-#include "core/graph/initializer.h"
-#include "core/graph/conv_activation_fusion.h"
-#include "core/graph/graph_utils.h"
 #include <deque>
+#include "core/graph/graph_utils.h"
+#include "core/optimizer/initializer.h"
+#include "core/optimizer/conv_activation_fusion.h"
 
 using namespace onnx;
 using namespace ::onnxruntime::common;

onnxruntime/core/optimizer/conv_activation_fusion.h

@@ -3,7 +3,7 @@
 
 #pragma once
 
-#include "core/graph/graph_transformer.h"
+#include "core/optimizer/graph_transformer.h"
 
 namespace onnxruntime {
 

onnxruntime/core/optimizer/conv_add_fusion.cc

@@ -1,9 +1,9 @@
 // Copyright (c) Microsoft Corporation. All rights reserved.
 // Licensed under the MIT License.
 
-#include "core/graph/initializer.h"
 #include "core/graph/graph_utils.h"
-#include "core/graph/conv_add_fusion.h"
+#include "core/optimizer/initializer.h"
+#include "core/optimizer/conv_add_fusion.h"
 
 using namespace onnx;
 using namespace ::onnxruntime::common;

onnxruntime/core/optimizer/conv_add_fusion.h

@@ -3,7 +3,7 @@
 
 #pragma once
 
-#include "core/graph/graph_transformer.h"
+#include "core/optimizer/graph_transformer.h"
 
 namespace onnxruntime {
 

onnxruntime/core/optimizer/conv_bn_fusion.cc

@@ -1,9 +1,9 @@
 // Copyright (c) Microsoft Corporation. All rights reserved.
 // Licensed under the MIT License.
 
-#include "core/graph/initializer.h"
 #include "core/graph/graph_utils.h"
-#include "core/graph/conv_bn_fusion.h"
+#include "core/optimizer/initializer.h"
+#include "core/optimizer/conv_bn_fusion.h"
 
 using namespace onnx;
 using namespace ::onnxruntime::common;

onnxruntime/core/optimizer/conv_bn_fusion.h

@@ -3,7 +3,7 @@
 
 #pragma once
 
-#include "core/graph/graph_transformer.h"
+#include "core/optimizer/graph_transformer.h"
 
 namespace onnxruntime {
 

onnxruntime/core/optimizer/conv_mul_fusion.cc

@@ -1,9 +1,9 @@
 // Copyright (c) Microsoft Corporation. All rights reserved.
 // Licensed under the MIT License.
 
-#include "core/graph/initializer.h"
 #include "core/graph/graph_utils.h"
-#include "core/graph/conv_mul_fusion.h"
+#include "core/optimizer/initializer.h"
+#include "core/optimizer/conv_mul_fusion.h"
 
 using namespace onnx;
 using namespace ::onnxruntime::common;

onnxruntime/core/optimizer/conv_mul_fusion.h

@@ -2,7 +2,7 @@
 // Licensed under the MIT License.
 
 #pragma once
-#include "core/graph/graph_transformer.h"
+#include "core/optimizer/graph_transformer.h"
 
 namespace onnxruntime {
 

onnxruntime/core/optimizer/gemm_activation_fusion.cc

@@ -1,8 +1,8 @@
 // Copyright (c) Microsoft Corporation. All rights reserved.
 // Licensed under the MIT License.
 
-#include "core/graph/initializer.h"
-#include "core/graph/gemm_activation_fusion.h"
+#include "core/optimizer/initializer.h"
+#include "core/optimizer/gemm_activation_fusion.h"
 #include "core/graph/graph_utils.h"
 #include <deque>
 

onnxruntime/core/optimizer/gemm_activation_fusion.h

@@ -3,7 +3,7 @@
 
 #pragma once
 
-#include "core/graph/graph_transformer.h"
+#include "core/optimizer/graph_transformer.h"
 
 namespace onnxruntime {
 

onnxruntime/core/optimizer/graph_transformer.cc

@@ -1,7 +1,7 @@
 // Copyright (c) Microsoft Corporation. All rights reserved.
 // Licensed under the MIT License.
 
-#include "core/graph/graph_transformer.h"
+#include "core/optimizer/graph_transformer.h"
 
 using namespace ::onnxruntime::common;
 

onnxruntime/core/optimizer/graph_transformer_mgr.cc

@@ -1,7 +1,7 @@
 // Copyright (c) Microsoft Corporation. All rights reserved.
 // Licensed under the MIT License.
 
-#include "core/graph/graph_transformer_mgr.h"
+#include "core/optimizer/graph_transformer_mgr.h"
 using namespace onnxruntime;
 using namespace ::onnxruntime::common;
 

onnxruntime/core/optimizer/graph_transformer_mgr.h

@@ -3,7 +3,7 @@
 
 #pragma once
 
-#include "core/graph/graph_transformer.h"
+#include "core/optimizer/graph_transformer.h"
 
 namespace onnxruntime {
 // Manages a list of graph transformers. It is initialized with a list of graph

onnxruntime/core/optimizer/identity_elimination.cc

@@ -1,8 +1,12 @@
 // Copyright (c) Microsoft Corporation. All rights reserved.
 // Licensed under the MIT License.
 
-#include "core/graph/identity_elimination.h"
+#include "core/common/logging/logging.h"
+#include "core/graph/graph_viewer.h"
+#include "core/graph/op.h"
 #include "core/graph/graph_utils.h"
+#include "core/optimizer/rewrite_rule.h"
+#include "core/optimizer/identity_elimination.h"
 
 namespace onnxruntime {
 

onnxruntime/core/optimizer/identity_elimination.h

@@ -3,7 +3,7 @@
 
 #pragma once
 
-#include "core/graph/rewrite_rule.h"
+#include "core/optimizer/rewrite_rule.h"
 
 namespace onnxruntime {
 

onnxruntime/core/optimizer/insert_cast_transformer.cc

@@ -1,7 +1,7 @@
 // Copyright (c) Microsoft Corporation. All rights reserved.
 // Licensed under the MIT License.
 
-#include "core/framework/insert_cast_transformer.h"
+#include "core/optimizer/insert_cast_transformer.h"
 #include "core/framework/data_types.h"
 
 using namespace ONNX_NAMESPACE;

onnxruntime/core/optimizer/insert_cast_transformer.h

@@ -2,10 +2,10 @@
 // Licensed under the MIT License.
 
 #pragma once
-#include "core/graph/graph_viewer.h"
-#include "core/graph/graph_transformer.h"
 #include "core/common/common.h"
+#include "core/graph/graph_viewer.h"
 #include "core/framework/op_kernel.h"
+#include "core/optimizer/graph_transformer.h"
 
 namespace onnxruntime {
 class InsertCastTransformer : public onnxruntime::GraphTransformer {

onnxruntime/core/optimizer/matmul_add_fusion.cc

@@ -1,8 +1,8 @@
 // Copyright (c) Microsoft Corporation. All rights reserved.
 // Licensed under the MIT License.
 
-#include "core/graph/initializer.h"
-#include "core/graph/matmul_add_fusion.h"
+#include "core/optimizer/initializer.h"
+#include "core/optimizer/matmul_add_fusion.h"
 #include "core/graph/graph_utils.h"
 #include <deque>
 

onnxruntime/core/optimizer/matmul_add_fusion.h

@@ -3,7 +3,7 @@
 
 #pragma once
 
-#include "core/graph/graph_transformer.h"
+#include "core/optimizer/graph_transformer.h"
 
 namespace onnxruntime {
 

onnxruntime/core/optimizer/slice_elimination.cc

@@ -1,7 +1,7 @@
 // Copyright (c) Microsoft Corporation. All rights reserved.
 // Licensed under the MIT License.
 
-#include "core/graph/slice_elimination.h"
+#include "core/optimizer/slice_elimination.h"
 #include "core/graph/graph.h"
 #include "core/graph/graph_utils.h"
 #include "core/graph/op.h"

onnxruntime/core/optimizer/slice_elimination.h

@@ -3,7 +3,7 @@
 
 #pragma once
 
-#include "core/graph/rewrite_rule.h"
+#include "core/optimizer/rewrite_rule.h"
 
 namespace onnxruntime {
 

onnxruntime/core/optimizer/transformer_memcpy.h

@@ -6,7 +6,7 @@
 #include "core/common/common.h"
 #include "core/framework/op_kernel.h"
 #include "core/framework/kernel_registry_manager.h"
-#include "core/graph/graph_transformer.h"
+#include "core/optimizer/graph_transformer.h"
 
 namespace onnxruntime {
 

onnxruntime/core/optimizer/unsqueeze_elimination.cc

@@ -1,7 +1,7 @@
 // Copyright (c) Microsoft Corporation. All rights reserved.
 // Licensed under the MIT License.
 
-#include "core/graph/unsqueeze_elimination.h"
+#include "core/optimizer/unsqueeze_elimination.h"
 
 using namespace onnx;
 using namespace ::onnxruntime::common;

onnxruntime/core/optimizer/unsqueeze_elimination.h

@@ -3,7 +3,7 @@
 
 #pragma once
 
-#include "core/graph/graph_transformer.h"
+#include "core/optimizer/graph_transformer.h"
 
 namespace onnxruntime {
 

onnxruntime/core/providers/cpu/cpu_execution_provider.h

@@ -5,7 +5,6 @@
 
 #include "core/framework/allocatormgr.h"
 #include "core/framework/execution_provider.h"
-#include "core/graph/graph_transformer.h"
 #include "core/graph/constants.h"
 
 namespace onnxruntime {

onnxruntime/core/providers/cuda/cuda_execution_provider.cc

@@ -3,7 +3,6 @@
 
 #include "cuda_common.h"
 #include "cuda_execution_provider.h"
-#include "core/framework/transformer_memcpy.h"
 #include "core/framework/memcpy.h"
 #include "cuda_fence.h"
 #include "cuda_allocator.h"

onnxruntime/core/providers/cuda/cuda_execution_provider.h

@@ -4,7 +4,7 @@
 #pragma once
 #include "cuda_pch.h"
 #include "core/platform/ort_mutex.h"
-#include "core/graph/graph_transformer.h"
+#include "core/graph/constants.h"
 #include "core/framework/allocatormgr.h"
 #include "core/framework/execution_provider.h"
 #include "shared_inc/cuda_utils.h"

onnxruntime/core/providers/mkldnn/mkldnn_execution_provider.cc

@@ -4,7 +4,6 @@
 #include "mkldnn_allocator.h"
 #include "mkldnn_execution_provider.h"
 #include "core/framework/allocator.h"
-#include "core/framework/transformer_memcpy.h"
 #include "core/framework/memcpy.h"
 #include "core/framework/kernel_registry.h"
 #include "mkldnn_fwd.h"

onnxruntime/core/providers/mkldnn/mkldnn_execution_provider.h

@@ -9,9 +9,9 @@
 #include <memory.h>
 
 #include "core/platform/ort_mutex.h"
+#include "core/graph/constants.h"
 #include "core/framework/allocatormgr.h"
 #include "core/framework/execution_provider.h"
-#include "core/graph/graph_transformer.h"
 
 namespace mkldnn {
 struct memory;

onnxruntime/core/session/inference_session.cc

@@ -16,8 +16,6 @@
 #include "core/common/logging/logging.h"
 #include "core/common/task_thread_pool.h"
 #include "core/graph/graph_viewer.h"
-#include "core/graph/graph_transformer.h"
-#include "core/graph/graph_transformer_mgr.h"
 #include "core/graph/graph_utils.h"
 #include "core/graph/model.h"
 #include "core/framework/allocatormgr.h"
@@ -25,7 +23,6 @@
 #include "core/framework/environment.h"
 #include "core/framework/execution_frame.h"
 #include "core/framework/graph_partitioner.h"
-#include "core/framework/insert_cast_transformer.h"
 #include "core/framework/kernel_def_builder.h"
 #include "core/framework/kernel_registry.h"
 #include "core/framework/ml_value_patterns_planner.h"
@@ -37,8 +34,11 @@
 #include "core/framework/session_state_initializer.h"
 #include "core/framework/tensorprotoutils.h"
 #include "core/framework/tensorutils.h"
-#include "core/framework/transformer_memcpy.h"
 #include "core/framework/utils.h"
+#include "core/optimizer/graph_transformer.h"
+#include "core/optimizer/graph_transformer_mgr.h"
+#include "core/optimizer/insert_cast_transformer.h"
+#include "core/optimizer/transformer_memcpy.h"
 #include "core/platform/notification.h"
 #include "core/providers/cpu/cpu_execution_provider.h"
 #include "core/session/CustomOpsLoader.h"

onnxruntime/test/framework/insert_cast_transformer_test.cc

@@ -3,7 +3,7 @@
 
 #include "core/framework/allocatormgr.h"
 #include "core/framework/allocator.h"
-#include "core/framework/insert_cast_transformer.h"
+#include "core/optimizer/insert_cast_transformer.h"
 #include "core/graph/model.h"
 #include "gtest/gtest.h"
 #include "test_utils.h"

onnxruntime/test/framework/memcpy_transformer_test.cc

@@ -2,8 +2,9 @@
 // Licensed under the MIT License.
 
 #include <iterator>
-#include "core/framework/transformer_memcpy.h"
+
 #include "core/framework/execution_providers.h"
+#include "core/optimizer/transformer_memcpy.h"
 #include "core/graph/model.h"
 #include "gtest/gtest.h"
 #include "test_utils.h"

onnxruntime/test/ir/graph_transform_test.cc

@@ -4,19 +4,18 @@
 #include "core/session/inference_session.h"
 #include "core/graph/graph_viewer.h"
 #include "core/graph/model.h"
-#include "core/graph/graph_transformer.h"
-#include "core/graph/graph_transformer_mgr.h"
-#include "core/graph/identity_elimination.h"
-#include "core/graph/slice_elimination.h"
-#include "core/graph/unsqueeze_elimination.h"
-#include "core/graph/conv_bn_fusion.h"
-#include "core/graph/conv_mul_fusion.h"
-#include "core/graph/conv_add_fusion.h"
-#include "core/graph/conv_activation_fusion.h"
-#include "core/graph/matmul_add_fusion.h"
-#include "core/graph/gemm_activation_fusion.h"
+#include "core/optimizer/graph_transformer.h"
+#include "core/optimizer/graph_transformer_mgr.h"
+#include "core/optimizer/identity_elimination.h"
+#include "core/optimizer/slice_elimination.h"
+#include "core/optimizer/unsqueeze_elimination.h"
+#include "core/optimizer/conv_bn_fusion.h"
+#include "core/optimizer/conv_mul_fusion.h"
+#include "core/optimizer/conv_add_fusion.h"
+#include "core/optimizer/conv_activation_fusion.h"
+#include "core/optimizer/matmul_add_fusion.h"
+#include "core/optimizer/gemm_activation_fusion.h"
 #include "core/platform/env.h"
-
 #include "test/capturing_sink.h"
 #include "test/test_environment.h"
 #include "gtest/gtest.h"