saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-07-30 20:18:08 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Check kernel def hashes (#7120)

Browse source 
Add unit test for verifying kernel def hashes.
Add way to add new types to kernel definition without changing hash.
This commit is contained in:
Edward Chen 2021-04-01 17:42:58 -07:00 committed by GitHub
parent 15c67ddbf0
commit 0ebeaf529d
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
23 changed files with 3474 additions and 76 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

4
cmake/CMakeLists.txt
View file

@ -709,6 +709,10 @@ endif()
add_subdirectory(external/mp11 EXCLUDE_FROM_ALL)
set(JSON_BuildTests OFF CACHE INTERNAL "")
set(JSON_Install OFF CACHE INTERNAL "")
add_subdirectory(external/json EXCLUDE_FROM_ALL)
if(onnxruntime_PREFER_SYSTEM_LIB)
find_package(re2)
endif()

14
cmake/onnxruntime_providers.cmake
View file

@ -139,10 +139,20 @@ if (onnxruntime_ENABLE_TRAINING_OPS)
list(APPEND onnxruntime_providers_src ${onnxruntime_cpu_training_ops_srcs})
file(GLOB_RECURSE onnxruntime_cpu_full_training_only_srcs
"${ORTTRAINING_SOURCE_DIR}/training_ops/cpu/gist/*.h"
"${ORTTRAINING_SOURCE_DIR}/training_ops/cpu/collective/*.cc"
"${ORTTRAINING_SOURCE_DIR}/training_ops/cpu/collective/*.h"
"${ORTTRAINING_SOURCE_DIR}/training_ops/cpu/communication/*.cc"
"${ORTTRAINING_SOURCE_DIR}/training_ops/cpu/communication/*.h"
"${ORTTRAINING_SOURCE_DIR}/training_ops/cpu/controlflow/record.cc"
"${ORTTRAINING_SOURCE_DIR}/training_ops/cpu/controlflow/record.h"
"${ORTTRAINING_SOURCE_DIR}/training_ops/cpu/controlflow/wait.cc"
"${ORTTRAINING_SOURCE_DIR}/training_ops/cpu/controlflow/wait.h"
"${ORTTRAINING_SOURCE_DIR}/training_ops/cpu/controlflow/yield.cc"
"${ORTTRAINING_SOURCE_DIR}/training_ops/cpu/controlflow/yield.h"
"${ORTTRAINING_SOURCE_DIR}/training_ops/cpu/gist/*.cc"
"${ORTTRAINING_SOURCE_DIR}/training_ops/cpu/tensorboard/*.h"
"${ORTTRAINING_SOURCE_DIR}/training_ops/cpu/gist/*.h"
"${ORTTRAINING_SOURCE_DIR}/training_ops/cpu/tensorboard/*.cc"
"${ORTTRAINING_SOURCE_DIR}/training_ops/cpu/tensorboard/*.h"
)
list(REMOVE_ITEM onnxruntime_providers_src ${onnxruntime_cpu_full_training_only_srcs})

3
cmake/onnxruntime_session.cmake
View file

@ -15,7 +15,8 @@ onnxruntime_add_include_to_target(onnxruntime_session onnxruntime_common onnxrun
if(onnxruntime_ENABLE_INSTRUMENT)
target_compile_definitions(onnxruntime_session PUBLIC ONNXRUNTIME_ENABLE_INSTRUMENT)
endif()
target_include_directories(onnxruntime_session PRIVATE ${ONNXRUNTIME_ROOT} ${PROJECT_SOURCE_DIR}/external/json ${eigen_INCLUDE_DIRS})
target_include_directories(onnxruntime_session PRIVATE ${ONNXRUNTIME_ROOT} ${eigen_INCLUDE_DIRS})
target_link_libraries(onnxruntime_session PRIVATE nlohmann_json::nlohmann_json)
add_dependencies(onnxruntime_session ${onnxruntime_EXTERNAL_DEPENDENCIES})
set_target_properties(onnxruntime_session PROPERTIES FOLDER "ONNXRuntime")
if (onnxruntime_USE_CUDA)

3
cmake/onnxruntime_training.cmake
View file

@ -59,7 +59,8 @@ add_dependencies(onnxruntime_training_runner ${onnxruntime_EXTERNAL_DEPENDENCIES
onnxruntime_add_include_to_target(onnxruntime_training_runner onnxruntime_training onnxruntime_framework onnxruntime_common onnx onnx_proto protobuf::libprotobuf onnxruntime_training flatbuffers)
target_include_directories(onnxruntime_training_runner PRIVATE ${CMAKE_CURRENT_BINARY_DIR} ${ONNXRUNTIME_ROOT} ${ORTTRAINING_ROOT} ${eigen_INCLUDE_DIRS} ${PROJECT_SOURCE_DIR}/external/json PUBLIC ${onnxruntime_graph_header})
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_CUDA)
target_include_directories(onnxruntime_training_runner PUBLIC ${onnxruntime_CUDNN_HOME}/include ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
endif()

4
cmake/onnxruntime_unittests.cmake
View file

@ -607,7 +607,9 @@ endif()
AddTest(
TARGET onnxruntime_test_all
SOURCES ${all_tests} ${onnxruntime_unittest_main_src}
LIBS onnx_test_runner_common ${onnxruntime_test_providers_libs} ${onnxruntime_test_common_libs} re2::re2 onnx_test_data_proto
LIBS
onnx_test_runner_common ${onnxruntime_test_providers_libs} ${onnxruntime_test_common_libs} re2::re2
onnx_test_data_proto nlohmann_json::nlohmann_json
DEPENDS ${all_dependencies}
)

52
include/onnxruntime/core/framework/kernel_def_builder.h
View file

@ -53,7 +53,13 @@ class KernelDef {
return provider_type_;
}
// type constraints with types supported by default
const std::map<std::string, std::vector<MLDataType>>& TypeConstraints() const {
return default_type_constraints_;
}
// type constraints with types supported in this build
const std::map<std::string, std::vector<MLDataType>>& EnabledTypeConstraints() const {
return enabled_type_constraints_;
}
@ -127,16 +133,20 @@ class KernelDef {
// The type of the execution provider.
std::string provider_type_;
// The supported data types for inputs/outputs.
// The data types that are supported by default for inputs/outputs.
// Key is input/output name defined in op schema, Value are supported types.
// note: std::map as we need the order to be deterministic for the hash
// Note: supported_type_constraints_ are used to calculate the kernel hash so that the hash is
// Note: default_type_constraints_ are used to calculate the kernel hash so that the hash is
// stable across builds with and without kernel type reduction enabled.
std::map<std::string, std::vector<MLDataType>> supported_type_constraints_;
std::map<std::string, std::vector<MLDataType>> default_type_constraints_;
// the type constraints that are enabled in this build for the kernel
// the type constraints that are supported in this build (enabled) for the kernel
std::map<std::string, std::vector<MLDataType>> enabled_type_constraints_;
// optional alternate type constraints to use to calculate the hash instead of default_type_constraints_
// note: this provides a way to update the default type constraints while preserving the hash value
optional<std::map<std::string, std::vector<MLDataType>>> hash_type_constraints_;
// An element <i, j> means that output j reuses the memory of input i.
std::vector<std::pair<int, int>> inplace_map_;
@ -209,27 +219,41 @@ class KernelDefBuilder {
/**
Specify the set of types that this kernel supports. A further restriction
of the set of types specified in the op schema.
The arg name could be either op formal parameter name, say "X", or type
argument name specified in op schema, say "T".
If this build uses type reduction the enabled types can optionally be provided.
@param arg_name The arg name can be either op formal parameter name, say "X", or type
argument name specified in op schema, say "T".
@param default_types The types that are supported by default.
@param enabled_types The types that are supported in this build.
Possibly different from default_types when type reduction is enabled.
*/
KernelDefBuilder& TypeConstraint(const std::string& arg_name,
const std::vector<MLDataType>& supported_types);
const std::vector<MLDataType>& default_types);
KernelDefBuilder& TypeConstraint(const char* arg_name,
const std::vector<MLDataType>& supported_types);
const std::vector<MLDataType>& default_types);
KernelDefBuilder& TypeConstraint(const std::string& arg_name,
const std::vector<MLDataType>& supported_types,
const std::vector<MLDataType>& default_types,
const std::vector<MLDataType>& enabled_types);
KernelDefBuilder& TypeConstraint(const char* arg_name,
const std::vector<MLDataType>& supported_types,
const std::vector<MLDataType>& default_types,
const std::vector<MLDataType>& enabled_types);
/**
Like TypeConstraint but supports just a single type.
*/
KernelDefBuilder& TypeConstraint(const std::string& arg_name, MLDataType supported_type);
KernelDefBuilder& TypeConstraint(const char* arg_name, MLDataType supported_type);
KernelDefBuilder& TypeConstraint(const std::string& arg_name, MLDataType default_type);
KernelDefBuilder& TypeConstraint(const char* arg_name, MLDataType default_type);
/**
Specify the original set of types that this kernel supports by default to use when computing the kernel def hash.
The set of types supported by default may change over time, but the hash should stay the same.
*/
KernelDefBuilder& FixedTypeConstraintForHash(
const std::string& arg_name,
const std::vector<MLDataType>& default_types_for_hash);
KernelDefBuilder& FixedTypeConstraintForHash(
const char* arg_name,
const std::vector<MLDataType>& default_types_for_hash);
/**
Inplace mapping from inputs to outputs allowed.
@ -360,7 +384,7 @@ class KernelDefBuilder {
private:
KernelDefBuilder& TypeConstraintImpl(const std::string& arg_name,
const std::vector<MLDataType>& supported_types,
const std::vector<MLDataType>& default_types,
const std::vector<MLDataType>* enabled_types = nullptr);
// we own the KernelDef until Build() is called.

4
include/onnxruntime/core/framework/kernel_registry.h
View file

@ -8,6 +8,7 @@
namespace onnxruntime {
using KernelCreateMap = std::multimap<std::string, KernelCreateInfo>;
using KernelDefHashes = std::vector<std::pair<std::string, uint64_t>>;
/**
* Each provider has a KernelRegistry. Often, the KernelRegistry only belongs to that specific provider.
@ -61,6 +62,9 @@ class KernelRegistry {
}
#endif
// Get sorted kernel def key and hash pairs.
KernelDefHashes ExportKernelDefHashes() const;
private:
#if !defined(ORT_MINIMAL_BUILD)
// Check whether the types of inputs/outputs of the given node match the extra

60
onnxruntime/core/framework/kernel_def_builder.cc
View file

@ -56,8 +56,11 @@ void KernelDef::CalculateHash() {
hash_str(op_domain_);
hash_str(provider_type_);
// use the supported_type_constraints_ list for the hash so the value in an ORT format model is stable.
for (const auto& key_value : supported_type_constraints_) {
// use the hash_type_constraints_ or default_type_constraints_ list for the hash so the value in an ORT format model
// is stable.
const auto& hash_type_constraints =
hash_type_constraints_.has_value() ? *hash_type_constraints_ : default_type_constraints_;
for (const auto& key_value : hash_type_constraints) {
hash_str(key_value.first);
auto data_type_strings = DataTypeImpl::ToString(key_value.second);
// sort type constraint data type strings so that order does not matter
@ -94,9 +97,9 @@ bool KernelDef::IsConflict(const KernelDef& other) const {
//only one case they don't conflict:
//There is a type_constraint, it exists in both hands, but they don't overlap
//check types
const auto& other_types = other.TypeConstraints();
const auto& other_types = other.default_type_constraints_;
bool type_has_conflict = true;
for (const auto& it : supported_type_constraints_) {
for (const auto& it : default_type_constraints_) {
auto iter = other_types.find(it.first);
if (iter != other_types.end()) {
if (!AreVectorsOverlap(it.second, iter->second)) {
@ -173,46 +176,63 @@ KernelDefBuilder& KernelDefBuilder::Provider(const char* provider_type) {
}
KernelDefBuilder& KernelDefBuilder::TypeConstraintImpl(const std::string& arg_name,
const std::vector<MLDataType>& supported_types,
const std::vector<MLDataType>& default_types,
const std::vector<MLDataType>* enabled_types) {
// use the enabled types list if provided
kernel_def_->enabled_type_constraints_[arg_name] = enabled_types ? *enabled_types : supported_types;
kernel_def_->supported_type_constraints_[arg_name] = supported_types;
kernel_def_->enabled_type_constraints_[arg_name] = enabled_types ? *enabled_types : default_types;
kernel_def_->default_type_constraints_[arg_name] = default_types;
return *this;
}
KernelDefBuilder& KernelDefBuilder::TypeConstraint(const std::string& arg_name,
const std::vector<MLDataType>& supported_types) {
return TypeConstraintImpl(arg_name, supported_types, nullptr);
const std::vector<MLDataType>& default_types) {
return TypeConstraintImpl(arg_name, default_types, nullptr);
}
KernelDefBuilder& KernelDefBuilder::TypeConstraint(const char* arg_name,
const std::vector<MLDataType>& supported_types) {
return TypeConstraintImpl(arg_name, supported_types, nullptr);
const std::vector<MLDataType>& default_types) {
return TypeConstraintImpl(arg_name, default_types, nullptr);
}
KernelDefBuilder& KernelDefBuilder::TypeConstraint(const std::string& arg_name,
const std::vector<MLDataType>& supported_types,
const std::vector<MLDataType>& default_types,
const std::vector<MLDataType>& enabled_types) {
return TypeConstraintImpl(arg_name, supported_types, &enabled_types);
return TypeConstraintImpl(arg_name, default_types, &enabled_types);
}
KernelDefBuilder& KernelDefBuilder::TypeConstraint(const char* arg_name,
const std::vector<MLDataType>& supported_types,
const std::vector<MLDataType>& default_types,
const std::vector<MLDataType>& enabled_types) {
return TypeConstraintImpl(arg_name, supported_types, &enabled_types);
return TypeConstraintImpl(arg_name, default_types, &enabled_types);
}
KernelDefBuilder& KernelDefBuilder::TypeConstraint(const std::string& arg_name,
MLDataType supported_type) {
kernel_def_->enabled_type_constraints_[arg_name] = std::vector<MLDataType>{supported_type};
kernel_def_->supported_type_constraints_[arg_name] = std::vector<MLDataType>{supported_type};
MLDataType default_type) {
kernel_def_->enabled_type_constraints_[arg_name] = std::vector<MLDataType>{default_type};
kernel_def_->default_type_constraints_[arg_name] = std::vector<MLDataType>{default_type};
return *this;
}
KernelDefBuilder& KernelDefBuilder::TypeConstraint(const char* arg_name,
MLDataType supported_type) {
return TypeConstraint(std::string(arg_name), supported_type);
MLDataType default_type) {
return TypeConstraint(std::string(arg_name), default_type);
}
KernelDefBuilder& KernelDefBuilder::FixedTypeConstraintForHash(
const std::string& arg_name,
const std::vector<MLDataType>& default_types_for_hash) {
auto& hash_type_constraints = kernel_def_->hash_type_constraints_;
if (!hash_type_constraints.has_value()) {
hash_type_constraints.emplace();
}
(*hash_type_constraints)[arg_name] = default_types_for_hash;
return *this;
}
KernelDefBuilder& KernelDefBuilder::FixedTypeConstraintForHash(
const char* arg_name,
const std::vector<MLDataType>& default_types_for_hash) {
return FixedTypeConstraintForHash(std::string{arg_name}, default_types_for_hash);
}
KernelDefBuilder& KernelDefBuilder::MayInplace(const std::vector<std::pair<int, int>>& inplaces) {

20
onnxruntime/core/framework/kernel_registry.cc
View file

@ -1,9 +1,12 @@
// Copyright (c) Microsoft Corporation. All rights reserved.
// Licensed under the MIT License.
#include "core/framework/kernel_registry.h"
#include <algorithm>
#include <memory>
#include <unordered_map>
#include "core/framework/kernel_registry.h"
#include "core/framework/session_state.h"
using namespace ::onnxruntime::common;
@ -149,7 +152,7 @@ bool KernelRegistry::VerifyKernelDef(const onnxruntime::Node& node,
}
// check if type matches
auto& kernel_type_constraints = kernel_def.TypeConstraints();
auto& kernel_type_constraints = kernel_def.EnabledTypeConstraints();
// Note: The number of formal input/output parameters is N and the number of
// type constraints is M. We select between an O(N*M) and an O(N+M) approach.
@ -323,4 +326,17 @@ Status KernelRegistry::Register(KernelCreateInfo&& create_info) {
return Status::OK();
}
KernelDefHashes KernelRegistry::ExportKernelDefHashes() const {
KernelDefHashes result{};
result.reserve(kernel_creator_fn_map_.size());
std::transform(
kernel_creator_fn_map_.begin(), kernel_creator_fn_map_.end(),
std::back_inserter(result),
[](const KernelCreateMap::value_type& kvp) {
return std::make_pair(kvp.first, kvp.second.kernel_def->GetHash());
});
std::sort(result.begin(), result.end());
return result;
}
} // namespace onnxruntime

2
onnxruntime/core/providers/cpu/cpu_execution_provider.cc
View file

@ -1882,7 +1882,7 @@ Status RegisterOnnxMLOperatorKernels(KernelRegistry& kernel_registry) {
} // namespace ml
#endif
static Status RegisterCPUKernels(KernelRegistry& kernel_registry) {
Status RegisterCPUKernels(KernelRegistry& kernel_registry) {
ORT_RETURN_IF_ERROR(RegisterOnnxOperatorKernels(kernel_registry));
#ifndef DISABLE_ML_OPS
ORT_RETURN_IF_ERROR(::onnxruntime::ml::RegisterOnnxMLOperatorKernels(kernel_registry));

4
onnxruntime/core/providers/cpu/cpu_execution_provider.h
View file

@ -52,4 +52,8 @@ class CPUExecutionProvider : public IExecutionProvider {
private:
std::vector<FuseRuleFn> fuse_rules_;
};
// Registers all available CPU kernels
Status RegisterCPUKernels(KernelRegistry& kernel_registry);
} // namespace onnxruntime

2
onnxruntime/core/session/inference_session_utils.h
View file

@ -18,7 +18,7 @@
#pragma warning(push)
#pragma warning(disable : 28020)
#endif
#include "single_include/nlohmann/json.hpp"
#include "nlohmann/json.hpp"
#ifdef _WIN32
#pragma warning(pop)
#endif

18
onnxruntime/test/framework/kernel_def_test.cc
View file

@ -14,8 +14,6 @@ TEST(KernelDefTest, HashIgnoresTypeConstraintTypeOrdering) {
auto build_kernel_def = [](std::vector<MLDataType> type_constraint_types) {
return KernelDefBuilder{}
.SetName("MyOp")
.SetDomain("MyDomain")
.Provider("MyProvider")
.TypeConstraint("T", type_constraint_types)
.Build();
};
@ -26,5 +24,21 @@ TEST(KernelDefTest, HashIgnoresTypeConstraintTypeOrdering) {
ASSERT_EQ(a->GetHash(), b->GetHash());
}
TEST(KernelDefTest, HashUsesFixedTypeConstraint) {
const auto a =
KernelDefBuilder{}
.SetName("MyOp")
.TypeConstraint("T", BuildKernelDefConstraints<int, double>())
.Build();
const auto b =
KernelDefBuilder{}
.SetName("MyOp")
.TypeConstraint("T", BuildKernelDefConstraints<int, double, float>())
.FixedTypeConstraintForHash("T", BuildKernelDefConstraints<int, double>())
.Build();
ASSERT_EQ(a->GetHash(), b->GetHash());
}
} // namespace test
} // namespace onnxruntime

184
onnxruntime/test/providers/kernel_def_hash_test.cc Normal file
View file

@ -0,0 +1,184 @@
// Copyright (c) Microsoft Corporation. All rights reserved.
// Licensed under the MIT License.
/**
* IMPORTANT NOTE AT THE TOP OF THE FILE
*
* This file contains tests which verify expected kernel def hashes.
* It is important for these to remain stable so that ORT format models are
* backward compatible.
*
* If you are seeing a test failure from one of these tests, it is likely that
* some kernel definition changed in a way that updated its hash value.
* This is what we want to catch! Please update the kernel definition.
* If adding more supported types to an existing kernel definition, consider
* using KernelDefBuilder::FixedTypeConstraintForHash().
*
* For example:
* Say we have a kernel definition like this, which supports types int and
* double:
* KernelDefBuilder{}
* .TypeConstraint(
* "T", BuildKernelDefConstraints<int, double>())
* If we want to update the kernel definition to add support for float, we can
* change it to something like this:
* KernelDefBuilder{}
* .TypeConstraint(
* "T", BuildKernelDefConstraints<int, double, float>())
* .FixedTypeConstraintForHash(
* "T", BuildKernelDefConstraints<int, double>())
* In the updated kernel definition, the original types are specified with
* FixedTypeConstraintForHash().
*
* New kernel definitions should not use FixedTypeConstraintForHash().
* It is a way to keep the hash stable as kernel definitions change.
*
* It is also possible that you have added a new kernel definition and are
* seeing a message from one of these tests about updating the expected data.
* Please do that if appropriate.
*
* The expected value files are in this directory:
* onnxruntime/test/testdata/kernel_def_hashes
* The data is specified in JSON as an array of key-value arrays.
* Example data can be written to stdout with this test:
* KernelDefHashTest.DISABLED_PrintCpuKernelDefHashes
* Use the option --gtest_also_run_disabled_tests to enable it.
* Be careful about updating the expected values - as mentioned before, the
* values should be stable. Typically, we should only add new entries.
*
* In the unlikely event that we need to make a change to the kernel def
* hashing that breaks backward compatibility, the expected values may need to
* be updated.
*/
#include <algorithm>
#include <cinttypes>
#include <fstream>
#include <iostream>
#include "gtest/gtest.h"
#ifdef _WIN32
#pragma warning(push)
#pragma warning(disable : 28020)
#endif
#include "nlohmann/json.hpp"
#ifdef _WIN32
#pragma warning(pop)
#endif
#include "asserts.h"
#include "core/common/common.h"
#include "core/common/path_string.h"
#include "core/framework/kernel_registry.h"
#include "core/mlas/inc/mlas.h"
#include "core/platform/env_var_utils.h"
#include "core/providers/cpu/cpu_execution_provider.h"
using json = nlohmann::json;
namespace onnxruntime {
namespace test {
namespace {
// If set to 1, do strict checking of the kernel def hash values.
// With strict checking, the expected and actual values must match exactly.
// Otherwise, the expected values must be present in the actual values.
static constexpr const char* kStrictKernelDefHashCheckEnvVar =
"ORT_TEST_STRICT_KERNEL_DEF_HASH_CHECK";
std::string DumpKernelDefHashes(const onnxruntime::KernelDefHashes& kernel_def_hashes) {
const json j(kernel_def_hashes);
return j.dump(/* indent */ 4);
}
KernelDefHashes ParseKernelDefHashes(std::istream& in) {
KernelDefHashes kernel_def_hashes{};
const json j = json::parse(in);
j.get_to<onnxruntime::KernelDefHashes>(kernel_def_hashes);
return kernel_def_hashes;
}
void AppendKernelDefHashesFromFile(const PathString& path, KernelDefHashes& kernel_def_hashes) {
std::ifstream in{path};
ORT_ENFORCE(in, "Failed to open file: ", ToMBString(path));
const auto file_kernel_def_hashes = ParseKernelDefHashes(in);
kernel_def_hashes.insert(
kernel_def_hashes.end(), file_kernel_def_hashes.begin(), file_kernel_def_hashes.end());
}
void CheckKernelDefHashes(const KernelDefHashes& actual, const KernelDefHashes& expected, bool is_strict) {
ASSERT_TRUE(std::is_sorted(actual.begin(), actual.end()));
ASSERT_TRUE(std::is_sorted(expected.begin(), expected.end()));
constexpr const char* kNoteReference = "Note: Please read the note at the top of this file: " __FILE__;
KernelDefHashes expected_minus_actual{};
std::set_difference(expected.begin(), expected.end(), actual.begin(), actual.end(),
std::back_inserter(expected_minus_actual));
if (!expected_minus_actual.empty()) {
const auto message = MakeString(
"Some expected kernel def hashes were not found.\n",
kNoteReference, "\n",
DumpKernelDefHashes(expected_minus_actual));
ADD_FAILURE() << message;
}
KernelDefHashes actual_minus_expected{};
std::set_difference(actual.begin(), actual.end(), expected.begin(), expected.end(),
std::back_inserter(actual_minus_expected));
if (!actual_minus_expected.empty()) {
const auto message = MakeString(
"Unexpected kernel def hashes were found, please update the expected values as needed "
"(see the output below).\n",
kNoteReference, "\n",
DumpKernelDefHashes(actual_minus_expected));
if (is_strict) {
ADD_FAILURE() << message;
} else {
std::cerr << message << "\n";
}
}
}
} // namespace
TEST(KernelDefHashTest, DISABLED_PrintCpuKernelDefHashes) {
KernelRegistry kernel_registry{};
ASSERT_STATUS_OK(RegisterCPUKernels(kernel_registry));
const auto cpu_kernel_def_hashes = kernel_registry.ExportKernelDefHashes();
std::cout << DumpKernelDefHashes(cpu_kernel_def_hashes) << "\n";
}
TEST(KernelDefHashTest, ExpectedCpuKernelDefHashes) {
const bool is_strict = ParseEnvironmentVariableWithDefault<bool>(kStrictKernelDefHashCheckEnvVar, false);
const auto expected_cpu_kernel_def_hashes = []() {
KernelDefHashes result{};
AppendKernelDefHashesFromFile(ORT_TSTR("testdata/kernel_def_hashes/onnx.cpu.json"), result);
#if !defined(DISABLE_ML_OPS)
AppendKernelDefHashesFromFile(ORT_TSTR("testdata/kernel_def_hashes/onnx.ml.cpu.json"), result);
#endif // !DISABLE_ML_OPS
#if !defined(DISABLE_CONTRIB_OPS)
AppendKernelDefHashesFromFile(ORT_TSTR("testdata/kernel_def_hashes/contrib.cpu.json"), result);
// NCHWc kernels are enabled if MlasNchwcGetBlockSize() > 1
if (MlasNchwcGetBlockSize() > 1) {
AppendKernelDefHashesFromFile(ORT_TSTR("testdata/kernel_def_hashes/contrib.nchwc.cpu.json"), result);
}
#endif // !DISABLE_CONTRIB_OPS
#if defined(ENABLE_TRAINING_OPS)
AppendKernelDefHashesFromFile(ORT_TSTR("testdata/kernel_def_hashes/training_ops.cpu.json"), result);
#endif // ENABLE_TRAINING_OPS
// TODO also handle kernels enabled by these symbols: ML_FEATURIZERS, BUILD_MS_EXPERIMENTAL_OPS
std::sort(result.begin(), result.end());
return result;
}();
KernelRegistry kernel_registry{};
ASSERT_STATUS_OK(RegisterCPUKernels(kernel_registry));
auto cpu_kernel_def_hashes = kernel_registry.ExportKernelDefHashes();
CheckKernelDefHashes(cpu_kernel_def_hashes, expected_cpu_kernel_def_hashes, is_strict);
}
} // namespace test
} // namespace onnxruntime

250
onnxruntime/test/testdata/kernel_def_hashes/contrib.cpu.json vendored Normal file
View file

@ -0,0 +1,250 @@
[
[
"Affine ai.onnx CPUExecutionProvider",
7811918192248490408
],
[
"Crop ai.onnx CPUExecutionProvider",
6914973556202621376
],
[
"DynamicSlice ai.onnx CPUExecutionProvider",
5387668728763060584
],
[
"ImageScaler ai.onnx CPUExecutionProvider",
2013093418027264536
],
[
"LayerNormalization ai.onnx CPUExecutionProvider",
4058615579523172864
],
[
"LayerNormalization ai.onnx CPUExecutionProvider",
8466416990072218056
],
[
"MeanVarianceNormalization ai.onnx CPUExecutionProvider",
13114085849278607104
],
[
"ParametricSoftplus ai.onnx CPUExecutionProvider",
17971715260566574960
],
[
"Scale ai.onnx CPUExecutionProvider",
12599351089228483328
],
[
"ScaledTanh ai.onnx CPUExecutionProvider",
15584477984618710520
],
[
"SimplifiedLayerNormalization ai.onnx CPUExecutionProvider",
16349480652468900704
],
[
"SimplifiedLayerNormalization ai.onnx CPUExecutionProvider",
418129161279605176
],
[
"ThresholdedRelu ai.onnx CPUExecutionProvider",
17820769706565099200
],
[
"Attention com.microsoft CPUExecutionProvider",
16464502426529915864
],
[
"AttnLSTM com.microsoft CPUExecutionProvider",
15421184737689665128
],
[
"BiasGelu com.microsoft CPUExecutionProvider",
12457646955212583504
],
[
"CDist com.microsoft CPUExecutionProvider",
889036143745127232
],
[
"CDist com.microsoft CPUExecutionProvider",
6280134801002897280
],
[
"ConvTransposeWithDynamicPads com.microsoft CPUExecutionProvider",
1596732273609633752
],
[
"CropAndResize com.microsoft CPUExecutionProvider",
7642430665819070720
],
[
"DequantizeLinear com.microsoft CPUExecutionProvider",
9034670407031092344
],
[
"DequantizeLinear com.microsoft CPUExecutionProvider",
12760451019866331016
],
[
"DynamicQuantizeLSTM com.microsoft CPUExecutionProvider",
640700714499684624
],
[
"DynamicQuantizeMatMul com.microsoft CPUExecutionProvider",
9591826880179639184
],
[
"EmbedLayerNormalization com.microsoft CPUExecutionProvider",
14614049725238705256
],
[
"ExpandDims com.microsoft CPUExecutionProvider",
5671892069881567792
],
[
"FastGelu com.microsoft CPUExecutionProvider",
18210983793195477200
],
[
"FusedConv com.microsoft CPUExecutionProvider",
11366858116389652832
],
[
"FusedGemm com.microsoft CPUExecutionProvider",
1341171831223136792
],
[
"FusedMatMul com.microsoft CPUExecutionProvider",
665364151288353496
],
[
"GatherND com.microsoft CPUExecutionProvider",
8466578404783779600
],
[
"Gelu com.microsoft CPUExecutionProvider",
4658746266161736328
],
[
"Inverse com.microsoft CPUExecutionProvider",
1037755270231788608
],
[
"MatMulInteger16 com.microsoft CPUExecutionProvider",
5265636774129358144
],
[
"MatMulIntegerToFloat com.microsoft CPUExecutionProvider",
7172777464471435800
],
[
"MaxpoolWithMask com.microsoft CPUExecutionProvider",
3144686615632467360
],
[
"MurmurHash3 com.microsoft CPUExecutionProvider",
2533733396673225096
],
[
"NhwcMaxPool com.microsoft CPUExecutionProvider",
8512357837341844248
],
[
"Pad com.microsoft CPUExecutionProvider",
15076596470814458544
],
[
"QAttention com.microsoft CPUExecutionProvider",
9844377440996919912
],
[
"QLinearAdd com.microsoft CPUExecutionProvider",
9958112514164905192
],
[
"QLinearAdd com.microsoft CPUExecutionProvider",
16322459350118343880
],
[
"QLinearAveragePool com.microsoft CPUExecutionProvider",
9152647959212466896
],
[
"QLinearConv com.microsoft CPUExecutionProvider",
16835965565578160400
],
[
"QLinearGlobalAveragePool com.microsoft CPUExecutionProvider",
8729391959357542728
],
[
"QLinearLeakyRelu com.microsoft CPUExecutionProvider",
3677670974923917280
],
[
"QLinearLeakyRelu com.microsoft CPUExecutionProvider",
17073324515720209136
],
[
"QLinearMul com.microsoft CPUExecutionProvider",
2406593953080780408
],
[
"QLinearMul com.microsoft CPUExecutionProvider",
17403503869116794888
],
[
"QLinearSigmoid com.microsoft CPUExecutionProvider",
17020165931626188400
],
[
"QLinearSigmoid com.microsoft CPUExecutionProvider",
17315947486917903320
],
[
"QuantizeLinear com.microsoft CPUExecutionProvider",
616915237400456368
],
[
"QuantizeLinear com.microsoft CPUExecutionProvider",
13556449850953958792
],
[
"Range com.microsoft CPUExecutionProvider",
9333951582187402912
],
[
"SampleOp com.microsoft CPUExecutionProvider",
11028204786545834016
],
[
"SkipLayerNormalization com.microsoft CPUExecutionProvider",
1829676129267529920
],
[
"SkipLayerNormalization com.microsoft CPUExecutionProvider",
15124962608939318760
],
[
"Tokenizer com.microsoft CPUExecutionProvider",
12821105347567077024
],
[
"TransposeMatMul com.microsoft CPUExecutionProvider",
3696625852111461496
],
[
"Trilu com.microsoft CPUExecutionProvider",
1828108687906670152
],
[
"Unique com.microsoft CPUExecutionProvider",
17512097873619224240
],
[
"WordConvEmbedding com.microsoft CPUExecutionProvider",
7416606351345164776
]
]

34
onnxruntime/test/testdata/kernel_def_hashes/contrib.nchwc.cpu.json vendored Normal file
View file

@ -0,0 +1,34 @@
[
[
"AveragePool com.microsoft.nchwc CPUExecutionProvider",
12528194512485261552
],
[
"Conv com.microsoft.nchwc CPUExecutionProvider",
10643058043438608528
],
[
"GlobalAveragePool com.microsoft.nchwc CPUExecutionProvider",
9401543287182687288
],
[
"GlobalMaxPool com.microsoft.nchwc CPUExecutionProvider",
17341568537930161320
],
[
"MaxPool com.microsoft.nchwc CPUExecutionProvider",
14527249939908647936
],
[
"ReorderInput com.microsoft.nchwc CPUExecutionProvider",
14330795113746035424
],
[
"ReorderOutput com.microsoft.nchwc CPUExecutionProvider",
13428915370009679360
],
[
"Upsample com.microsoft.nchwc CPUExecutionProvider",
16347985363638744760
]
]

2402
onnxruntime/test/testdata/kernel_def_hashes/onnx.cpu.json vendored Normal file
View file

File diff suppressed because it is too large Load diff

178
onnxruntime/test/testdata/kernel_def_hashes/onnx.ml.cpu.json vendored Normal file
View file

@ -0,0 +1,178 @@
[
[
"ArrayFeatureExtractor ai.onnx.ml CPUExecutionProvider",
2350526174341531704
],
[
"ArrayFeatureExtractor ai.onnx.ml CPUExecutionProvider",
6530469673967836712
],
[
"ArrayFeatureExtractor ai.onnx.ml CPUExecutionProvider",
6756475889913415064
],
[
"ArrayFeatureExtractor ai.onnx.ml CPUExecutionProvider",
11064940903354222584
],
[
"ArrayFeatureExtractor ai.onnx.ml CPUExecutionProvider",
16432528150559367472
],
[
"Binarizer ai.onnx.ml CPUExecutionProvider",
5046510072939685048
],
[
"CastMap ai.onnx.ml CPUExecutionProvider",
8048827990321743320
],
[
"CategoryMapper ai.onnx.ml CPUExecutionProvider",
2303612724843538232
],
[
"DictVectorizer ai.onnx.ml CPUExecutionProvider",
6689651410341671832
],
[
"DictVectorizer ai.onnx.ml CPUExecutionProvider",
7005642460916971440
],
[
"DictVectorizer ai.onnx.ml CPUExecutionProvider",
9815805762175478424
],
[
"DictVectorizer ai.onnx.ml CPUExecutionProvider",
10888259614400691552
],
[
"DictVectorizer ai.onnx.ml CPUExecutionProvider",
12166013579191601224
],
[
"DictVectorizer ai.onnx.ml CPUExecutionProvider",
14175801436008296008
],
[
"FeatureVectorizer ai.onnx.ml CPUExecutionProvider",
13725963185932500216
],
[
"Imputer ai.onnx.ml CPUExecutionProvider",
1198484622450883960
],
[
"LabelEncoder ai.onnx.ml CPUExecutionProvider",
72584427791628640
],
[
"LabelEncoder ai.onnx.ml CPUExecutionProvider",
643259145804007712
],
[
"LabelEncoder ai.onnx.ml CPUExecutionProvider",
2667163092143996224
],
[
"LabelEncoder ai.onnx.ml CPUExecutionProvider",
5922496927153994000
],
[
"LabelEncoder ai.onnx.ml CPUExecutionProvider",
11838196408428523536
],
[
"LabelEncoder ai.onnx.ml CPUExecutionProvider",
12429922703862798200
],
[
"LabelEncoder ai.onnx.ml CPUExecutionProvider",
12910259746827434448
],
[
"LabelEncoder ai.onnx.ml CPUExecutionProvider",
13358273281151927664
],
[
"LinearClassifier ai.onnx.ml CPUExecutionProvider",
2370253020407350728
],
[
"LinearRegressor ai.onnx.ml CPUExecutionProvider",
16822518791234245504
],
[
"Normalizer ai.onnx.ml CPUExecutionProvider",
4575839837655511352
],
[
"OneHotEncoder ai.onnx.ml CPUExecutionProvider",
566681355974248448
],
[
"OneHotEncoder ai.onnx.ml CPUExecutionProvider",
5824294864440328760
],
[
"OneHotEncoder ai.onnx.ml CPUExecutionProvider",
10362613019971997872
],
[
"OneHotEncoder ai.onnx.ml CPUExecutionProvider",
10397925769639456800
],
[
"Scaler ai.onnx.ml CPUExecutionProvider",
2957692969286147560
],
[
"Scaler ai.onnx.ml CPUExecutionProvider",
4622934914100034264
],
[
"Scaler ai.onnx.ml CPUExecutionProvider",
15151724670894748016
],
[
"Scaler ai.onnx.ml CPUExecutionProvider",
17146108806137553320
],
[
"SVMClassifier ai.onnx.ml CPUExecutionProvider",
16410276550989508464
],
[
"SVMRegressor ai.onnx.ml CPUExecutionProvider",
6414907892091941152
],
[
"TreeEnsembleClassifier ai.onnx.ml CPUExecutionProvider",
486797053154945264
],
[
"TreeEnsembleClassifier ai.onnx.ml CPUExecutionProvider",
7696156941733005088
],
[
"TreeEnsembleClassifier ai.onnx.ml CPUExecutionProvider",
7775841366762359432
],
[
"TreeEnsembleClassifier ai.onnx.ml CPUExecutionProvider",
18326787492602399840
],
[
"TreeEnsembleRegressor ai.onnx.ml CPUExecutionProvider",
1006399804521896912
],
[
"TreeEnsembleRegressor ai.onnx.ml CPUExecutionProvider",
12993125630596348064
],
[
"ZipMap ai.onnx.ml CPUExecutionProvider",
868519487849210656
]
]

238
onnxruntime/test/testdata/kernel_def_hashes/training_ops.cpu.json vendored Normal file
View file

@ -0,0 +1,238 @@
[
[
"AdamOptimizer com.microsoft CPUExecutionProvider",
13275719513674142848
],
[
"AveragePoolGrad ai.onnx CPUExecutionProvider",
5748823370585834408
],
[
"BiasFastGeluGrad_dX com.microsoft CPUExecutionProvider",
18012658855595136536
],
[
"BiasGeluGrad_dX com.microsoft CPUExecutionProvider",
15594101660509653368
],
[
"BroadcastGradientArgs com.microsoft CPUExecutionProvider",
11924624129611280440
],
[
"ConcatTraining com.microsoft CPUExecutionProvider",
407435603592769928
],
[
"ConvGrad ai.onnx CPUExecutionProvider",
551027277226613536
],
[
"DropoutGrad com.microsoft CPUExecutionProvider",
5281827689086376112
],
[
"DropoutGrad com.microsoft CPUExecutionProvider",
5974036940246406232
],
[
"DropoutGrad com.microsoft CPUExecutionProvider",
6251139593746398664
],
[
"DropoutGrad com.microsoft CPUExecutionProvider",
11134433709210415000
],
[
"DropoutGrad com.microsoft CPUExecutionProvider",
14442689431073529904
],
[
"DropoutGrad com.microsoft CPUExecutionProvider",
14487527510076876072
],
[
"FastGeluGrad com.microsoft CPUExecutionProvider",
15449034010577224840
],
[
"GatherElementsGrad com.microsoft CPUExecutionProvider",
4371565130613539784
],
[
"GatherGrad com.microsoft CPUExecutionProvider",
16767260238333903216
],
[
"GatherNDGrad com.microsoft CPUExecutionProvider",
14458526033602252960
],
[
"GeluGrad com.microsoft CPUExecutionProvider",
3065626784130845072
],
[
"Group com.microsoft CPUExecutionProvider",
488667512000820344
],
[
"InPlaceAccumulator com.microsoft CPUExecutionProvider",
10152728201494720480
],
[
"InvertibleLayerNormalizationGrad com.microsoft CPUExecutionProvider",
678361860206315480
],
[
"InvertibleLayerNormalizationGrad com.microsoft CPUExecutionProvider",
3647079615262464056
],
[
"LayerNormalizationGrad com.microsoft CPUExecutionProvider",
9571863035992961528
],
[
"LayerNormalizationGrad com.microsoft CPUExecutionProvider",
15120500842480246904
],
[
"LogSoftmaxGrad com.microsoft CPUExecutionProvider",
2657523710083167200
],
[
"MaxPoolGrad ai.onnx CPUExecutionProvider",
17526822836083413768
],
[
"PassThrough com.microsoft CPUExecutionProvider",
15753758832962034552
],
[
"ReduceAllL2 com.microsoft CPUExecutionProvider",
10206006236140935288
],
[
"ReduceSumTraining com.microsoft CPUExecutionProvider",
4143442227179196240
],
[
"ReduceSumTraining com.microsoft CPUExecutionProvider",
8270932107741781432
],
[
"ReduceSumTraining com.microsoft CPUExecutionProvider",
10014758284200613760
],
[
"ReduceSumTraining com.microsoft CPUExecutionProvider",
11970256914235931792
],
[
"ReluGrad com.microsoft CPUExecutionProvider",
6194712211707544696
],
[
"SGDOptimizer com.microsoft CPUExecutionProvider",
6413752339355984752
],
[
"Scale com.microsoft CPUExecutionProvider",
4626702086191057400
],
[
"Scale com.microsoft CPUExecutionProvider",
7066774539152819712
],
[
"Scale com.microsoft CPUExecutionProvider",
7086129391615471904
],
[
"Scale com.microsoft CPUExecutionProvider",
8529929466096310624
],
[
"Scale com.microsoft CPUExecutionProvider",
9194708927698241584
],
[
"Scale com.microsoft CPUExecutionProvider",
15004008446550052608
],
[
"Scale com.microsoft CPUExecutionProvider",
16665599877841166880
],
[
"Scale com.microsoft CPUExecutionProvider",
17615601449933173216
],
[
"SimplifiedLayerNormalizationGrad com.microsoft CPUExecutionProvider",
9823114592183077488
],
[
"SimplifiedLayerNormalizationGrad com.microsoft CPUExecutionProvider",
17704727407887118880
],
[
"SinGrad ai.onnx CPUExecutionProvider",
6971400065913943256
],
[
"SliceGrad com.microsoft CPUExecutionProvider",
18003932513454931536
],
[
"SoftmaxCrossEntropy com.microsoft CPUExecutionProvider",
5470752940282787040
],
[
"SoftmaxCrossEntropyGrad com.microsoft CPUExecutionProvider",
17579254067477868408
],
[
"SoftmaxCrossEntropyLoss ai.onnx CPUExecutionProvider",
379111282162709688
],
[
"SoftmaxCrossEntropyLoss ai.onnx CPUExecutionProvider",
5256121368123320104
],
[
"SoftmaxCrossEntropyLoss ai.onnx CPUExecutionProvider",
14833827121724789864
],
[
"SoftmaxCrossEntropyLoss ai.onnx CPUExecutionProvider",
15405100773745075656
],
[
"SoftmaxCrossEntropyLossGrad com.microsoft CPUExecutionProvider",
8253282220433537112
],
[
"SoftmaxCrossEntropyLossGrad com.microsoft CPUExecutionProvider",
16283082098560169504
],
[
"SoftmaxGrad com.microsoft CPUExecutionProvider",
4483165757863027152
],
[
"SparseSoftmaxCrossEntropy ai.onnx CPUExecutionProvider",
10638058507241762520
],
[
"SparseSoftmaxCrossEntropyGrad ai.onnx CPUExecutionProvider",
17612183648106847376
],
[
"SplitTraining com.microsoft CPUExecutionProvider",
12689204749897364688
],
[
"ZeroGradient com.microsoft CPUExecutionProvider",
3284255990062374928
]
]

2
orttraining/orttraining/models/runner/training_runner.cc
View file

@ -24,7 +24,7 @@
#if defined(USE_CUDA) && defined(ORT_USE_NCCL) && defined(USE_NCCL_P2P)
#include "orttraining/training_ops/cuda/communication/nccl_service.h"
#endif
#include "single_include/nlohmann/json.hpp"
#include "nlohmann/json.hpp"
#include "test/perftest/utils.h"
using json = nlohmann::json;

57
orttraining/orttraining/training_ops/cpu/cpu_training_kernels.cc
View file

@ -7,10 +7,6 @@
namespace onnxruntime {
namespace contrib {
#ifdef USE_MPI
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, AdasumAllReduce);
#endif
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, SGDOptimizer);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, AdamOptimizer);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, InPlaceAccumulator);
@ -58,14 +54,6 @@ class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1,
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, double_float, DropoutGrad);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, double_double, DropoutGrad);
#ifdef ENABLE_TRAINING
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, SummaryScalar);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, SummaryHistogram);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, SummaryMerge);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, SummaryText);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, GistBinarizeEncoder);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, GistBinarizeDecoder);
#endif
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, float, LayerNormalizationGrad);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, double, LayerNormalizationGrad);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, float, SimplifiedLayerNormalizationGrad);
@ -88,7 +76,12 @@ class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1,
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, float_float, ReduceAllL2);
// the kernels within the following ifdef are not included in a build with
// --enable_training_ops but without --enable_training
#ifdef ENABLE_TRAINING
#ifdef USE_MPI
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, AdasumAllReduce);
// Pipeline communication operators.
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, Send);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, Recv);
@ -98,12 +91,18 @@ class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, Recor
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, WaitEvent);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, YieldOp);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, SummaryScalar);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, SummaryHistogram);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, SummaryMerge);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, SummaryText);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, GistBinarizeEncoder);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, GistBinarizeDecoder);
#endif
Status RegisterCpuTrainingKernels(KernelRegistry& kernel_registry) {
static const BuildKernelCreateInfoFn function_table[] = {
#ifdef USE_MPI
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, AdasumAllReduce)>,
#endif
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, SGDOptimizer)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, AdamOptimizer)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, InPlaceAccumulator)>,
@ -150,22 +149,13 @@ Status RegisterCpuTrainingKernels(KernelRegistry& kernel_registry) {
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, double_float, DropoutGrad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, double_double, DropoutGrad)>,
#ifdef ENABLE_TRAINING
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, SummaryScalar)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, SummaryHistogram)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, SummaryMerge)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, SummaryText)>,
#endif
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, float, LayerNormalizationGrad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, double, LayerNormalizationGrad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, float, SimplifiedLayerNormalizationGrad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, double, SimplifiedLayerNormalizationGrad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, float, InvertibleLayerNormalizationGrad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, double, InvertibleLayerNormalizationGrad)>,
#ifdef ENABLE_TRAINING
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, GistBinarizeEncoder)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, GistBinarizeDecoder)>,
#endif
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, SliceGrad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, FastGeluGrad)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, BiasGeluGrad_dX)>,
@ -182,14 +172,29 @@ Status RegisterCpuTrainingKernels(KernelRegistry& kernel_registry) {
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, float_float, ReduceAllL2)>,
// the kernels within the following ifdef are not included in a build with
// --enable_training_ops but without --enable_training
#ifdef ENABLE_TRAINING
#ifdef USE_MPI
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, AdasumAllReduce)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, Send)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, Recv)>,
#endif
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, RecordEvent)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, WaitEvent)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, YieldOp)>};
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, YieldOp)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, SummaryScalar)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, SummaryHistogram)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, SummaryMerge)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, SummaryText)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, GistBinarizeEncoder)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kMSDomain, 1, GistBinarizeDecoder)>,
#endif
};
for (auto& function_table_entry : function_table) {
ORT_RETURN_IF_ERROR(kernel_registry.Register(function_table_entry()));

7
tools/ci_build/github/azure-pipelines/linux-cpu-minimal-build-ci-pipeline.yml
View file

@ -59,10 +59,10 @@ jobs:
workingDirectory: $(Build.SourcesDirectory)
- task: CmdLine@2
displayName: Build minimal onnxruntime [exceptions DISABLED, type reduction DISABLED]
displayName: Build minimal onnxruntime [exceptions DISABLED, type reduction DISABLED, training ops ENABLED]
inputs:
script: |
# We will try to build minimal ORT with exceptions disabled
# We will try to build minimal ORT with exceptions disabled and training ops enabled
# Only the building process is verified here, no test will be performed
docker run --rm \
--volume $(Build.SourcesDirectory):/onnxruntime_src \
@ -79,7 +79,8 @@ jobs:
--parallel \
--skip_tests \
--minimal_build \
--disable_exceptions
--disable_exceptions \
--enable_training_ops
workingDirectory: $(Build.SourcesDirectory)
- task: CmdLine@2

8
tools/ci_build/github/linux/ort_minimal/build_full_ort_and_create_ort_files.sh
View file

@ -14,7 +14,7 @@ python3 /onnxruntime_src/tools/ci_build/op_registration_validator.py
# Run a full build of ORT.
# We need the ORT python package to generate the ORT format files and the required ops config files.
# We do not run tests since those are covered by other CIs
# We do not run tests in this command since those are covered by other CIs.
python3 /onnxruntime_src/tools/ci_build/build.py \
--build_dir /build --cmake_generator Ninja \
--config Debug \
@ -22,8 +22,14 @@ python3 /onnxruntime_src/tools/ci_build/build.py \
--parallel \
--build_wheel \
--skip_tests \
--enable_training_ops \
--enable_pybind --cmake_extra_defines PYTHON_INCLUDE_DIR=/opt/python/cp37-cp37m/include/python3.7m PYTHON_LIBRARY=/usr/lib64/librt.so
# Run kernel def hash verification test
pushd /build/Debug
ORT_TEST_STRICT_KERNEL_DEF_HASH_CHECK=1 ./onnxruntime_test_all --gtest_filter="KernelDefHashTest.ExpectedCpuKernelDefHashes"
popd
# Install the ORT python wheel
python3 -m pip install --user /build/Debug/dist/*