saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-05-16 21:00:14 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions
onnxruntime/cmake/onnxruntime_training.cmake
Julius Tischbein 1391354265
Adding CUDNN Frontend and use for CUDA NN Convolution (#19470) 
### Description
Added CUDNN Frontend and used it for NHWC convolutions, and optionally
fuse activation.

#### Backward compatible 
- For model existed with FusedConv, model can still run. 
- If ORT is built with cuDNN 8, cuDNN frontend will not be built into
binary. Old kernels (using cudnn backend APIs) are used.

#### Major Changes
- For cuDNN 9, we will enable cudnn frontend to fuse convolution and
bias when a provider option `fuse_conv_bias=1`.
- Remove the fusion of FusedConv from graph transformer for CUDA
provider, so there will not be FusedConv be added to graph for CUDA EP
in the future.
- Update cmake files regarding to cudnn settings. The search order of
CUDNN installation in build are like the following:
  * environment variable `CUDNN_PATH`
* `onnxruntime_CUDNN_HOME` cmake extra defines. If a build starts from
build.py/build.sh, user can pass it through `--cudnn_home` parameter, or
by environment variable `CUDNN_HOME` if `--cudnn_home` not used.
* cudnn python package installation directory like
python3.xx/site-packages/nvidia/cudnn
  * CUDA installation path

#### Potential Issues

- If ORT is built with cuDNN 8, FusedConv fusion is no longer done
automatically, so some model might have performance regression. If user
still wants FusedConv operator for performance reason, they can still
have multiple ways to walkaround: like use older version of onnxruntime;
or use older version of ORT to save optimized onnx, then run with latest
version of ORT. We believe that majority users have moved to cudnn 9
when 1.20 release (since the default in ORT and PyTorch is cudnn 9 for 3
months when 1.20 release), so the impact is small.
- cuDNN graph uses TF32 by default, and user cannot disable TF32 through
the use_tf32 cuda provider option. If user encounters accuracy issue
(like in testing), user has to set environment variable
`NVIDIA_TF32_OVERRIDE=0` to disable TF32. Need update the document of
use_tf32 later.

#### Follow ups
This is one of PRs that target to enable NHWC convolution in CUDA EP by
default if device supports it. There are other changes will follow up to
make it possible.
(1) Enable `prefer_nhwc` by default for device with sm >= 70. 
(2) Change `fuse_conv_bias=1` by default after more testing.
(3) Add other NHWC operators (like Resize or UpSample).

### Motivation and Context

The new CUDNN Frontend library provides the functionality to fuse
operations and provides new heuristics for kernel selection. Here it
fuses the convolution with the pointwise bias operation. On the [NVIDIA
ResNet50](https://pytorch.org/hub/nvidia_deeplearningexamples_resnet50/)
we get a performance boost from 49.1144 ms to 42.4643 ms per inference
on a 2560x1440 input (`onnxruntime_perf_test -e cuda -I -q -r 100-d 1 -i
'prefer_nhwc|1' resnet50.onnx`).

---------

Co-authored-by: Tianlei Wu <tlwu@microsoft.com>
Co-authored-by: Maximilian Mueller <maximilianm@nvidia.com>
2024-08-02 15:16:42 -07:00

226 lines
11 KiB
CMake
Raw Blame History

# Copyright (c) Microsoft Corporation. All rights reserved.
# Licensed under the MIT License.
set (CXXOPTS ${cxxopts_SOURCE_DIR}/include)
# training lib
file(GLOB_RECURSE onnxruntime_training_srcs
"${ORTTRAINING_SOURCE_DIR}/core/framework/*.h"
"${ORTTRAINING_SOURCE_DIR}/core/framework/*.cc"
"${ORTTRAINING_SOURCE_DIR}/core/framework/tensorboard/*.h"
"${ORTTRAINING_SOURCE_DIR}/core/framework/tensorboard/*.cc"
"${ORTTRAINING_SOURCE_DIR}/core/framework/adasum/*"
"${ORTTRAINING_SOURCE_DIR}/core/framework/communication/*"
"${ORTTRAINING_SOURCE_DIR}/core/session/*.h"
"${ORTTRAINING_SOURCE_DIR}/core/session/*.cc"
"${ORTTRAINING_SOURCE_DIR}/core/agent/*.h"
"${ORTTRAINING_SOURCE_DIR}/core/agent/*.cc"
)
# This needs to be built in framework.cmake
file(GLOB_RECURSE onnxruntime_training_framework_excluded_srcs CONFIGURE_DEPENDS
"${ORTTRAINING_SOURCE_DIR}/core/framework/torch/*.h"
"${ORTTRAINING_SOURCE_DIR}/core/framework/torch/*.cc"
"${ORTTRAINING_SOURCE_DIR}/core/framework/triton/*.h"
"${ORTTRAINING_SOURCE_DIR}/core/framework/triton/*.cc"
)
list(REMOVE_ITEM onnxruntime_training_srcs ${onnxruntime_training_framework_excluded_srcs})
onnxruntime_add_static_library(onnxruntime_training ${onnxruntime_training_srcs})
add_dependencies(onnxruntime_training onnx tensorboard ${onnxruntime_EXTERNAL_DEPENDENCIES})
onnxruntime_add_include_to_target(onnxruntime_training onnxruntime_common onnx onnx_proto tensorboard ${PROTOBUF_LIB} flatbuffers::flatbuffers re2::re2 Boost::mp11 safeint_interface)
# fix event_writer.cc 4100 warning
if(WIN32)
target_compile_options(onnxruntime_training PRIVATE /wd4100)
endif()
target_include_directories(onnxruntime_training PRIVATE ${CMAKE_CURRENT_BINARY_DIR} ${ONNXRUNTIME_ROOT} ${ORTTRAINING_ROOT} ${eigen_INCLUDE_DIRS} PUBLIC ${onnxruntime_graph_header} ${MPI_CXX_INCLUDE_DIRS})
if (onnxruntime_USE_NCCL)
target_include_directories(onnxruntime_training PRIVATE ${NCCL_INCLUDE_DIRS})
endif()
if (onnxruntime_BUILD_UNIT_TESTS)
set_target_properties(onnxruntime_training PROPERTIES FOLDER "ONNXRuntime")
source_group(TREE ${ORTTRAINING_ROOT} FILES ${onnxruntime_training_srcs})
# training runner lib
file(GLOB_RECURSE onnxruntime_training_runner_srcs
"${ORTTRAINING_SOURCE_DIR}/models/runner/*.h"
"${ORTTRAINING_SOURCE_DIR}/models/runner/*.cc"
)
# perf test utils
set(onnxruntime_perf_test_src_dir ${TEST_SRC_DIR}/perftest)
set(onnxruntime_perf_test_src
"${onnxruntime_perf_test_src_dir}/utils.h")
if(WIN32)
list(APPEND onnxruntime_perf_test_src
"${onnxruntime_perf_test_src_dir}/windows/utils.cc")
else ()
list(APPEND onnxruntime_perf_test_src
"${onnxruntime_perf_test_src_dir}/posix/utils.cc")
endif()
onnxruntime_add_static_library(onnxruntime_training_runner ${onnxruntime_training_runner_srcs} ${onnxruntime_perf_test_src})
add_dependencies(onnxruntime_training_runner ${onnxruntime_EXTERNAL_DEPENDENCIES} onnx onnxruntime_providers)
if (onnxruntime_ENABLE_TRAINING_TORCH_INTEROP)
target_link_libraries(onnxruntime_training_runner PRIVATE Python::Python)
endif()
onnxruntime_add_include_to_target(onnxruntime_training_runner onnxruntime_training onnxruntime_framework onnxruntime_common onnx onnx_proto ${PROTOBUF_LIB} onnxruntime_training flatbuffers::flatbuffers Boost::mp11 safeint_interface)
target_include_directories(onnxruntime_training_runner PRIVATE ${CMAKE_CURRENT_BINARY_DIR} ${ONNXRUNTIME_ROOT} ${ORTTRAINING_ROOT} ${eigen_INCLUDE_DIRS} PUBLIC ${onnxruntime_graph_header})
target_link_libraries(onnxruntime_training_runner PRIVATE nlohmann_json::nlohmann_json)
if (onnxruntime_USE_NCCL)
target_include_directories(onnxruntime_training_runner PRIVATE ${NCCL_INCLUDE_DIRS})
endif()
if (onnxruntime_USE_ROCM)
add_definitions(-DUSE_ROCM=1)
target_include_directories(onnxruntime_training_runner PUBLIC ${onnxruntime_ROCM_HOME}/include)
endif()
check_cxx_compiler_flag(-Wno-maybe-uninitialized HAS_NO_MAYBE_UNINITIALIZED)
if(UNIX AND NOT APPLE)
if (HAS_NO_MAYBE_UNINITIALIZED)
target_compile_options(onnxruntime_training_runner PUBLIC "-Wno-maybe-uninitialized")
endif()
endif()
if (onnxruntime_USE_ROCM)
target_compile_options(onnxruntime_training_runner PUBLIC -D__HIP_PLATFORM_AMD__=1 -D__HIP_PLATFORM_HCC__=1)
endif()
set_target_properties(onnxruntime_training_runner PROPERTIES FOLDER "ONNXRuntimeTest")
source_group(TREE ${REPO_ROOT} FILES ${onnxruntime_training_runner_srcs} ${onnxruntime_perf_test_src})
# MNIST
file(GLOB_RECURSE training_mnist_src
"${ORTTRAINING_SOURCE_DIR}/models/mnist/*.h"
"${ORTTRAINING_SOURCE_DIR}/models/mnist/mnist_data_provider.cc"
"${ORTTRAINING_SOURCE_DIR}/models/mnist/main.cc"
)
onnxruntime_add_executable(onnxruntime_training_mnist ${training_mnist_src})
onnxruntime_add_include_to_target(onnxruntime_training_mnist onnxruntime_common onnx onnx_proto ${PROTOBUF_LIB} onnxruntime_training flatbuffers::flatbuffers Boost::mp11 safeint_interface)
target_include_directories(onnxruntime_training_mnist PUBLIC ${CMAKE_CURRENT_BINARY_DIR} ${ONNXRUNTIME_ROOT} ${ORTTRAINING_ROOT} ${eigen_INCLUDE_DIRS} ${CXXOPTS} ${extra_includes} ${onnxruntime_graph_header} ${onnxruntime_exec_src_dir} ${CMAKE_CURRENT_BINARY_DIR} ${CMAKE_CURRENT_BINARY_DIR}/onnx onnxruntime_training_runner)
set(ONNXRUNTIME_LIBS
onnxruntime_session
${onnxruntime_libs}
${PROVIDERS_MKLDNN}
${PROVIDERS_DML}
onnxruntime_optimizer
onnxruntime_providers
onnxruntime_util
onnxruntime_framework
)
if (onnxruntime_ENABLE_TRAINING_TORCH_INTEROP)
list(APPEND ONNXRUNTIME_LIBS Python::Python)
endif()
list(APPEND ONNXRUNTIME_LIBS
onnxruntime_graph
${ONNXRUNTIME_MLAS_LIBS}
onnxruntime_common
onnxruntime_flatbuffers
Boost::mp11 safeint_interface
)
if(UNIX AND NOT APPLE)
if (HAS_NO_MAYBE_UNINITIALIZED)
target_compile_options(onnxruntime_training_mnist PUBLIC "-Wno-maybe-uninitialized")
endif()
endif()
target_link_libraries(onnxruntime_training_mnist PRIVATE onnxruntime_training_runner onnxruntime_training ${ONNXRUNTIME_LIBS} ${onnxruntime_EXTERNAL_LIBRARIES})
set_target_properties(onnxruntime_training_mnist PROPERTIES FOLDER "ONNXRuntimeTest")
# squeezenet
# Disabling build for squeezenet, as no one is using this
#[[
file(GLOB_RECURSE training_squeezene_src
"${ORTTRAINING_SOURCE_DIR}/models/squeezenet/*.h"
"${ORTTRAINING_SOURCE_DIR}/models/squeezenet/*.cc"
)
onnxruntime_add_executable(onnxruntime_training_squeezenet ${training_squeezene_src})
onnxruntime_add_include_to_target(onnxruntime_training_squeezenet onnxruntime_common onnx onnx_proto ${PROTOBUF_LIB} onnxruntime_training flatbuffers::flatbuffers Boost::mp11 safeint_interface)
target_include_directories(onnxruntime_training_squeezenet PUBLIC ${ONNXRUNTIME_ROOT} ${ORTTRAINING_ROOT} ${eigen_INCLUDE_DIRS} ${extra_includes} ${onnxruntime_graph_header} ${onnxruntime_exec_src_dir} ${CMAKE_CURRENT_BINARY_DIR} ${CMAKE_CURRENT_BINARY_DIR}/onnx onnxruntime_training_runner)
if(UNIX AND NOT APPLE)
target_compile_options(onnxruntime_training_squeezenet PUBLIC "-Wno-maybe-uninitialized")
endif()
target_link_libraries(onnxruntime_training_squeezenet PRIVATE onnxruntime_training_runner onnxruntime_training ${ONNXRUNTIME_LIBS} ${onnxruntime_EXTERNAL_LIBRARIES})
set_target_properties(onnxruntime_training_squeezenet PROPERTIES FOLDER "ONNXRuntimeTest")
]]
# BERT
file(GLOB_RECURSE training_bert_src
"${ORTTRAINING_SOURCE_DIR}/models/bert/*.h"
"${ORTTRAINING_SOURCE_DIR}/models/bert/*.cc"
)
onnxruntime_add_executable(onnxruntime_training_bert ${training_bert_src})
if(UNIX AND NOT APPLE)
if (HAS_NO_MAYBE_UNINITIALIZED)
target_compile_options(onnxruntime_training_bert PUBLIC "-Wno-maybe-uninitialized")
endif()
endif()
onnxruntime_add_include_to_target(onnxruntime_training_bert onnxruntime_common onnx onnx_proto ${PROTOBUF_LIB} onnxruntime_training flatbuffers::flatbuffers Boost::mp11 safeint_interface)
target_include_directories(onnxruntime_training_bert PUBLIC ${CMAKE_CURRENT_BINARY_DIR} ${ONNXRUNTIME_ROOT} ${ORTTRAINING_ROOT} ${MPI_CXX_INCLUDE_DIRS} ${eigen_INCLUDE_DIRS} ${CXXOPTS} ${extra_includes} ${onnxruntime_graph_header} ${onnxruntime_exec_src_dir} ${CMAKE_CURRENT_BINARY_DIR} ${CMAKE_CURRENT_BINARY_DIR}/onnx onnxruntime_training_runner)
# ROCM provider sources are generated, need to add include directory for generated headers
if (onnxruntime_USE_ROCM)
target_include_directories(onnxruntime_training_bert PUBLIC ${CMAKE_CURRENT_BINARY_DIR}/amdgpu/onnxruntime)
endif()
target_link_libraries(onnxruntime_training_bert PRIVATE onnxruntime_training_runner onnxruntime_training ${ONNXRUNTIME_LIBS} ${onnxruntime_EXTERNAL_LIBRARIES})
set_target_properties(onnxruntime_training_bert PROPERTIES FOLDER "ONNXRuntimeTest")
# Pipeline
file(GLOB_RECURSE training_pipeline_poc_src
"${ORTTRAINING_SOURCE_DIR}/models/pipeline_poc/*.h"
"${ORTTRAINING_SOURCE_DIR}/models/pipeline_poc/*.cc"
)
onnxruntime_add_executable(onnxruntime_training_pipeline_poc ${training_pipeline_poc_src})
if(UNIX AND NOT APPLE)
if (HAS_NO_MAYBE_UNINITIALIZED)
target_compile_options(onnxruntime_training_pipeline_poc PUBLIC "-Wno-maybe-uninitialized")
endif()
endif()
onnxruntime_add_include_to_target(onnxruntime_training_pipeline_poc onnxruntime_common onnx onnx_proto ${PROTOBUF_LIB} onnxruntime_training flatbuffers::flatbuffers Boost::mp11 safeint_interface)
target_include_directories(onnxruntime_training_pipeline_poc PUBLIC ${CMAKE_CURRENT_BINARY_DIR} ${ONNXRUNTIME_ROOT} ${ORTTRAINING_ROOT} ${MPI_CXX_INCLUDE_DIRS} ${eigen_INCLUDE_DIRS} ${CXXOPTS} ${extra_includes} ${onnxruntime_graph_header} ${onnxruntime_exec_src_dir} ${CMAKE_CURRENT_BINARY_DIR} ${CMAKE_CURRENT_BINARY_DIR}/onnx onnxruntime_training_runner)
if (onnxruntime_USE_NCCL)
target_include_directories(onnxruntime_training_pipeline_poc PRIVATE ${NCCL_INCLUDE_DIRS})
endif()
target_link_libraries(onnxruntime_training_pipeline_poc PRIVATE onnxruntime_training_runner onnxruntime_training ${ONNXRUNTIME_LIBS} ${onnxruntime_EXTERNAL_LIBRARIES})
set_target_properties(onnxruntime_training_pipeline_poc PROPERTIES FOLDER "ONNXRuntimeTest")
# GPT-2
file(GLOB_RECURSE training_gpt2_src
"${ORTTRAINING_SOURCE_DIR}/models/gpt2/*.h"
"${ORTTRAINING_SOURCE_DIR}/models/gpt2/*.cc"
)
onnxruntime_add_executable(onnxruntime_training_gpt2 ${training_gpt2_src})
if(UNIX AND NOT APPLE)
if (HAS_NO_MAYBE_UNINITIALIZED)
target_compile_options(onnxruntime_training_gpt2 PUBLIC "-Wno-maybe-uninitialized")
endif()
endif()
target_include_directories(onnxruntime_training_gpt2 PUBLIC ${CMAKE_CURRENT_BINARY_DIR} ${ONNXRUNTIME_ROOT} ${ORTTRAINING_ROOT} ${MPI_CXX_INCLUDE_DIRS} ${eigen_INCLUDE_DIRS} ${CXXOPTS} ${extra_includes} ${onnxruntime_graph_header} ${onnxruntime_exec_src_dir} ${CMAKE_CURRENT_BINARY_DIR} ${CMAKE_CURRENT_BINARY_DIR}/onnx onnxruntime_training_runner)
target_link_libraries(onnxruntime_training_gpt2 PRIVATE onnxruntime_training_runner onnxruntime_training ${ONNXRUNTIME_LIBS} ${onnxruntime_EXTERNAL_LIBRARIES})
set_target_properties(onnxruntime_training_gpt2 PROPERTIES FOLDER "ONNXRuntimeTest")
endif()
Reference in a new issue View git blame Copy permalink