Skip to content

P
Paddle

PaddlePaddle / Paddle
大约 2 年 前同步成功

通知 2325
Star 20933
Fork 5424
P
Paddle
提交 46894436 编写于 作者: P phlrain
浏览文件
操作

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into move_embedding_to_phi

上级 e81773c9 a0cb3203
展开全部 显示空白变更内容
内联 并排
Showing with 14889 addition and 5098 deletion
.gitignore
浏览文件 @ 46894436
......@@ -20,6 +20,7 @@ build/
build_doc/
*.user
*.tmp
*.pyc
.vscode
.idea
......
CMakeLists.txt
浏览文件 @ 46894436
......@@ -330,6 +330,7 @@ if(WITH_BRPC_RDMA)
endif()
endif()
if(WITH_GPU)
include(cuda)
# lite subgraph compilation depends on CUDNN_ROOT,
......
cmake/external/llvm.cmake
浏览文件 @ 46894436
......@@ -99,7 +99,7 @@ endfunction()
function(mlir_add_rewriter td_base)
set(LLVM_TARGET_DEFINITIONS ${td_base}.td)
mlir_tablegen(${td_base}.hpp.inc -gen-rewriters "-I${CMAKE_SOURCE_DIR}/infrt/dialect/pass")
mlir_tablegen(${td_base}.cpp.inc -gen-rewriters "-I${CMAKE_SOURCE_DIR}/infrt/dialect/pass")
add_public_tablegen_target(${td_base}_IncGen)
add_custom_target(${td_base}_inc DEPENDS ${td_base}_IncGen)
endfunction()
......
cmake/generic.cmake
浏览文件 @ 46894436
......@@ -116,19 +116,19 @@ function(find_fluid_modules TARGET_NAME)
endif()
endfunction(find_fluid_modules)
set_property(GLOBAL PROPERTY PTEN_MODULES "")
# find all pten modules is used for paddle static library
set_property(GLOBAL PROPERTY PHI_MODULES "")
# find all phi modules is used for paddle static library
# for building inference libs
function(find_pten_modules TARGET_NAME)
function(find_phi_modules TARGET_NAME)
get_filename_component(__target_path ${TARGET_NAME} ABSOLUTE)
string(REGEX REPLACE "^${PADDLE_SOURCE_DIR}/" "" __target_path ${__target_path})
string(FIND "${__target_path}" "phi" pos)
if(pos GREATER 1)
get_property(pten_modules GLOBAL PROPERTY PTEN_MODULES)
set(pten_modules ${pten_modules} ${TARGET_NAME})
set_property(GLOBAL PROPERTY PTEN_MODULES "${pten_modules}")
get_property(phi_modules GLOBAL PROPERTY PHI_MODULES)
set(phi_modules ${phi_modules} ${TARGET_NAME})
set_property(GLOBAL PROPERTY PHI_MODULES "${phi_modules}")
endif()
endfunction(find_pten_modules)
endfunction(find_phi_modules)
function(common_link TARGET_NAME)
if (WITH_PROFILER)
......@@ -324,7 +324,7 @@ function(cc_library TARGET_NAME)
else()
add_library(${TARGET_NAME} STATIC ${cc_library_SRCS})
find_fluid_modules(${TARGET_NAME})
find_pten_modules(${TARGET_NAME})
find_phi_modules(${TARGET_NAME})
endif()
if(cc_library_DEPS)
# Don't need link libwarpctc.so
......@@ -497,7 +497,7 @@ function(nv_library TARGET_NAME)
else()
add_library(${TARGET_NAME} STATIC ${nv_library_SRCS})
find_fluid_modules(${TARGET_NAME})
find_pten_modules(${TARGET_NAME})
find_phi_modules(${TARGET_NAME})
endif()
if (nv_library_DEPS)
add_dependencies(${TARGET_NAME} ${nv_library_DEPS})
......@@ -588,7 +588,7 @@ function(hip_library TARGET_NAME)
else()
hip_add_library(${TARGET_NAME} STATIC ${hip_library_SRCS})
find_fluid_modules(${TARGET_NAME})
find_pten_modules(${TARGET_NAME})
find_phi_modules(${TARGET_NAME})
endif()
if (hip_library_DEPS)
add_dependencies(${TARGET_NAME} ${hip_library_DEPS})
......
cmake/inference_lib.cmake
浏览文件 @ 46894436
......@@ -224,7 +224,7 @@ copy(inference_lib_dist
DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/crypto/)
include_directories(${CMAKE_BINARY_DIR}/../paddle/fluid/framework/io)
# copy api headers for pten & custom op
# copy api headers for phi & custom op
copy(inference_lib_dist
SRCS ${PADDLE_SOURCE_DIR}/paddle/phi/api/ext/*.h
DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/experimental/phi/api/ext/)
......@@ -244,11 +244,11 @@ copy(inference_lib_dist
SRCS ${PADDLE_SOURCE_DIR}/paddle/extension.h
DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/experimental/)
# the header file of pten is copied to the experimental directory,
# the include path of pten needs to be changed to adapt to inference api path
# the header file of phi is copied to the experimental directory,
# the include path of phi needs to be changed to adapt to inference api path
add_custom_command(TARGET inference_lib_dist POST_BUILD
COMMAND ${CMAKE_COMMAND} -P "${PADDLE_SOURCE_DIR}/cmake/pten_header.cmake"
COMMENT "Change pten header include path to adapt to inference api path")
COMMAND ${CMAKE_COMMAND} -P "${PADDLE_SOURCE_DIR}/cmake/phi_header.cmake"
COMMENT "Change phi header include path to adapt to inference api path")
# CAPI inference library for only inference
set(PADDLE_INFERENCE_C_INSTALL_DIR "${CMAKE_BINARY_DIR}/paddle_inference_c_install_dir" CACHE STRING
......
cmake/operators.cmake
浏览文件 @ 46894436
......@@ -73,6 +73,12 @@ function(op_library TARGET)
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu)
list(APPEND cu_srcs ${TARGET}.cu)
endif()
# rename in KP: .kps -> .cu
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.kps)
file(COPY ${TARGET}.kps DESTINATION ${CMAKE_CURRENT_BINARY_DIR})
file(RENAME ${CMAKE_CURRENT_BINARY_DIR}/${TARGET}.kps ${CMAKE_CURRENT_BINARY_DIR}/${TARGET}.cu)
list(APPEND cu_srcs ${CMAKE_CURRENT_BINARY_DIR}/${TARGET}.cu)
endif()
if (WITH_NV_JETSON)
list(REMOVE_ITEM cu_srcs "decode_jpeg_op.cu")
endif()
......@@ -96,6 +102,12 @@ function(op_library TARGET)
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu)
list(APPEND hip_srcs ${TARGET}.cu)
endif()
# rename in KP: .kps -> .cu
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.kps)
file(COPY ${TARGET}.kps DESTINATION ${CMAKE_CURRENT_BINARY_DIR})
file(RENAME ${CMAKE_CURRENT_BINARY_DIR}/${TARGET}.kps ${CMAKE_CURRENT_BINARY_DIR}/${TARGET}.cu)
list(APPEND hip_srcs ${CMAKE_CURRENT_BINARY_DIR}/${TARGET}.cu)
endif()
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.part.cu)
set(PART_CUDA_KERNEL_FILES ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.part.cu
${PART_CUDA_KERNEL_FILES} PARENT_SCOPE)
......
cmake/pten.cmake → cmake/phi.cmake
浏览文件 @ 46894436
......@@ -51,7 +51,7 @@ function(generate_unify_header DIR_NAME)
endforeach()
# append header into extension.h
string(REPLACE "${PADDLE_SOURCE_DIR}\/" "" header_file "${header_file}")
file(APPEND ${pten_extension_header_file} "#include \"${header_file}\"\n")
file(APPEND ${phi_extension_header_file} "#include \"${header_file}\"\n")
endfunction()
# call kernel_declare need to make sure whether the target of input exists
......@@ -81,6 +81,8 @@ function(kernel_declare TARGET_LIST)
file(APPEND ${kernel_declare_file} "PD_DECLARE_KERNEL(${kernel_name}, GPU, ALL_LAYOUT);\n")
elseif (${kernel_path} MATCHES "./xpu\/")
file(APPEND ${kernel_declare_file} "PD_DECLARE_KERNEL(${kernel_name}, XPU, ALL_LAYOUT);\n")
elseif (${kernel_path} MATCHES "./gpudnn\/")
file(APPEND ${kernel_declare_file} "PD_DECLARE_KERNEL(${kernel_name}, GPUDNN, ALL_LAYOUT);\n")
else ()
# deal with device independent kernel, now we use CPU temporaary
file(APPEND ${kernel_declare_file} "PD_DECLARE_KERNEL(${kernel_name}, CPU, ALL_LAYOUT);\n")
......@@ -94,6 +96,7 @@ function(kernel_library TARGET)
set(cpu_srcs)
set(gpu_srcs)
set(xpu_srcs)
set(gpudnn_srcs)
set(selected_rows_srcs)
# parse and save the deps kerenl targets
set(all_srcs)
......@@ -101,6 +104,8 @@ function(kernel_library TARGET)
set(oneValueArgs SUB_DIR)
set(multiValueArgs SRCS DEPS)
set(target_build_flag 1)
cmake_parse_arguments(kernel_library "${options}" "${oneValueArgs}"
"${multiValueArgs}" ${ARGN})
......@@ -123,6 +128,9 @@ function(kernel_library TARGET)
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/gpu/${TARGET}.cu.cc)
list(APPEND gpu_srcs ${CMAKE_CURRENT_SOURCE_DIR}/gpu/${TARGET}.cu.cc)
endif()
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/gpudnn/${TARGET}_gpudnn.cu)
list(APPEND gpudnn_srcs ${CMAKE_CURRENT_SOURCE_DIR}/gpudnn/${TARGET}_gpudnn.cu)
endif()
endif()
if (WITH_XPU)
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/xpu/${TARGET}.cc)
......@@ -141,6 +149,7 @@ function(kernel_library TARGET)
list(APPEND all_srcs ${cpu_srcs})
list(APPEND all_srcs ${gpu_srcs})
list(APPEND all_srcs ${xpu_srcs})
list(APPEND all_srcs ${gpudnn_srcs})
foreach(src ${all_srcs})
file(READ ${src} target_content)
string(REGEX MATCHALL "#include \"paddle\/phi\/kernels\/[a-z0-9_]+_kernel.h\"" include_kernels ${target_content})
......@@ -166,21 +175,22 @@ function(kernel_library TARGET)
list(LENGTH cpu_srcs cpu_srcs_len)
list(LENGTH gpu_srcs gpu_srcs_len)
list(LENGTH xpu_srcs xpu_srcs_len)
list(LENGTH gpudnn_srcs gpudnn_srcs_len)
list(LENGTH selected_rows_srcs selected_rows_srcs_len)
# Build Target according different src organization
if((${cpu_srcs_len} GREATER 0 OR ${gpu_srcs_len} GREATER 0 OR
${xpu_srcs_len} GREATER 0) AND (${common_srcs_len} GREATER 0 OR
${selected_rows_srcs_len} GREATER 0))
${xpu_srcs_len} GREATER 0 OR ${gpudnn_srcs_len} GREATER 0) AND
(${common_srcs_len} GREATER 0 OR ${selected_rows_srcs_len} GREATER 0))
# If the common_srcs/selected_rows_srcs depends on specific device srcs, build target using this rule.
if (WITH_GPU)
if (${cpu_srcs_len} GREATER 0 OR ${gpu_srcs_len} GREATER 0)
nv_library(${TARGET}_part SRCS ${cpu_srcs} ${gpu_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
if (${cpu_srcs_len} GREATER 0 OR ${gpu_srcs_len} GREATER 0 OR ${gpudnn_srcs_len} GREATER 0)
nv_library(${TARGET}_part SRCS ${cpu_srcs} ${gpu_srcs} ${gpudnn_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
nv_library(${TARGET} SRCS ${common_srcs} ${selected_rows_srcs} DEPS ${TARGET}_part)
endif()
elseif (WITH_ROCM)
if (${cpu_srcs_len} GREATER 0 OR ${gpu_srcs_len} GREATER 0)
hip_library(${TARGET}_part SRCS ${cpu_srcs} ${gpu_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
if (${cpu_srcs_len} GREATER 0 OR ${gpu_srcs_len} GREATER 0 OR ${gpudnn_srcs_len} GREATER 0)
hip_library(${TARGET}_part SRCS ${cpu_srcs} ${gpu_srcs} ${gpudnn_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
hip_library(${TARGET} SRCS ${common_srcs} ${selected_rows_srcs} DEPS ${TARGET}_part)
endif()
else()
......@@ -190,14 +200,14 @@ function(kernel_library TARGET)
endif()
endif()
# If there are only specific device srcs, build target using this rule.
elseif (${cpu_srcs_len} GREATER 0 OR ${gpu_srcs_len} GREATER 0 OR ${xpu_srcs_len} GREATER 0)
elseif (${cpu_srcs_len} GREATER 0 OR ${gpu_srcs_len} GREATER 0 OR ${xpu_srcs_len} GREATER 0 OR ${gpudnn_srcs_len} GREATER 0)
if (WITH_GPU)
if (${cpu_srcs_len} GREATER 0 OR ${gpu_srcs_len} GREATER 0)
nv_library(${TARGET} SRCS ${cpu_srcs} ${gpu_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
if (${cpu_srcs_len} GREATER 0 OR ${gpu_srcs_len} GREATER 0 OR ${gpudnn_srcs_len} GREATER 0)
nv_library(${TARGET} SRCS ${cpu_srcs} ${gpu_srcs} ${gpudnn_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
endif()
elseif (WITH_ROCM)
if (${cpu_srcs_len} GREATER 0 OR ${gpu_srcs_len} GREATER 0)
hip_library(${TARGET} SRCS ${cpu_srcs} ${gpu_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
if (${cpu_srcs_len} GREATER 0 OR ${gpu_srcs_len} GREATER 0 OR ${gpudnn_srcs_len} GREATER 0)
hip_library(${TARGET} SRCS ${cpu_srcs} ${gpu_srcs} ${gpudnn_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
endif()
else()
if (${cpu_srcs_len} GREATER 0 OR ${xpu_srcs_len} GREATER 0)
......@@ -234,16 +244,17 @@ function(kernel_library TARGET)
cc_library(${TARGET} SRCS ${selected_rows_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
endif()
else()
message(FATAL_ERROR "Cannot find any implementation for ${TARGET}")
set(target_build_flag 0)
endif()
if (${target_build_flag} EQUAL 1)
if (${common_srcs_len} GREATER 0 OR ${cpu_srcs_len} GREATER 0 OR
${gpu_srcs_len} GREATER 0 OR ${xpu_srcs_len} GREATER 0 OR
${selected_rows_srcs_len} GREATER 0)
# append target into PTEN_KERNELS property
get_property(pten_kernels GLOBAL PROPERTY PTEN_KERNELS)
set(pten_kernels ${pten_kernels} ${TARGET})
set_property(GLOBAL PROPERTY PTEN_KERNELS ${pten_kernels})
${gpudnn_srcs_len} GREATER 0 OR ${selected_rows_srcs_len} GREATER 0)
# append target into PHI_KERNELS property
get_property(phi_kernels GLOBAL PROPERTY PHI_KERNELS)
set(phi_kernels ${phi_kernels} ${TARGET})
set_property(GLOBAL PROPERTY PHI_KERNELS ${phi_kernels})
endif()
# parse kernel name and auto generate kernel declaration
......@@ -261,9 +272,13 @@ function(kernel_library TARGET)
if (${xpu_srcs_len} GREATER 0)
kernel_declare(${xpu_srcs})
endif()
if (${gpudnn_srcs_len} GREATER 0)
kernel_declare(${gpudnn_srcs})
endif()
if (${selected_rows_srcs_len} GREATER 0)
kernel_declare(${selected_rows_srcs})
endif()
endif()
endfunction()
function(register_kernels)
......
cmake/pten_header.cmake → cmake/phi_header.cmake
浏览文件 @ 46894436
......@@ -14,8 +14,8 @@
set(PADDLE_INFERENCE_INSTALL_DIR "${CMAKE_BINARY_DIR}/paddle_inference_install_dir")
function(pten_header_path_compat TARGET_PATH)
message(STATUS "pten header path compat processing: ${TARGET_PATH}")
function(phi_header_path_compat TARGET_PATH)
message(STATUS "phi header path compat processing: ${TARGET_PATH}")
string(FIND ${TARGET_PATH} "experimental" pos)
if (pos GREATER 1)
file(GLOB HEADERS "${TARGET_PATH}/*" "*.h")
......@@ -25,17 +25,17 @@ if (pos GREATER 1)
string(REPLACE "paddle/phi/" "paddle/include/experimental/phi/" HEADER_CONTENT "${HEADER_CONTENT}")
string(REPLACE "paddle/utils/" "paddle/include/experimental/utils/" HEADER_CONTENT "${HEADER_CONTENT}")
file(WRITE ${header} "${HEADER_CONTENT}")
message(STATUS "pten header path compat processing complete: ${header}")
message(STATUS "phi header path compat processing complete: ${header}")
endif()
endforeach()
endif()
endfunction()
pten_header_path_compat(${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/experimental)
pten_header_path_compat(${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/experimental/phi/api)
pten_header_path_compat(${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/experimental/phi/api/ext)
pten_header_path_compat(${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/experimental/phi/api/include)
pten_header_path_compat(${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/experimental/phi/common)
phi_header_path_compat(${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/experimental)
phi_header_path_compat(${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/experimental/phi/api)
phi_header_path_compat(${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/experimental/phi/api/ext)
phi_header_path_compat(${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/experimental/phi/api/include)
phi_header_path_compat(${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/experimental/phi/common)
# In order to be compatible with the original behavior, the header file name needs to be changed
file(RENAME ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/experimental/extension.h
......
paddle/fluid/distributed/collective/CMakeLists.txt
浏览文件 @ 46894436
cc_library(processgroup SRCS ProcessGroup.cc DEPS pten pten_api eager_api)
cc_library(processgroup SRCS ProcessGroup.cc DEPS phi phi_api eager_api)
if(WITH_NCCL)
cc_library(processgroup_nccl SRCS ProcessGroupNCCL.cc DEPS place cuda_stream enforce collective_helper device_context pten pten_api eager_api)
cc_library(processgroup_nccl SRCS ProcessGroupNCCL.cc DEPS place cuda_stream enforce collective_helper device_context phi phi_api eager_api)
endif()
paddle/fluid/distributed/ps/service/brpc_utils.cc
浏览文件 @ 46894436
......@@ -238,7 +238,7 @@ void DeserializeLodTensor(framework::Variable* var, const VarMsg& msg,
void* tensor_data = tensor->mutable_data(
place,
framework::TransToPtenDataType(VarMessageToVarType(msg.data_type())));
framework::TransToPhiDataType(VarMessageToVarType(msg.data_type())));
// IO Buffer
if (platform::is_cpu_place(place)) {
......@@ -281,7 +281,7 @@ void DeserializeSelectedRows(
tensor->Resize(phi::make_ddim(vec_dim));
void* tensor_data = tensor->mutable_data(
place,
framework::TransToPtenDataType(VarMessageToVarType(msg.data_type())));
framework::TransToPhiDataType(VarMessageToVarType(msg.data_type())));
// IO Buffer
if (platform::is_cpu_place(place)) {
unsigned long data_len; // NOLINT
......
paddle/fluid/eager/CMakeLists.txt
浏览文件 @ 46894436
set(eager_deps pten pten_api hook_utils tensor_utils utils global_utils backward pten_tensor tracer layer autograd_meta grad_node_info grad_tensor_holder accumulation_node)
set(eager_deps phi phi_api hook_utils tensor_utils utils global_utils backward phi_tensor tracer layer autograd_meta grad_node_info grad_tensor_holder accumulation_node)
set(fluid_deps tracer layer proto_desc operator op_registry variable_helper memcpy)
set(generated_deps dygraph_function dygraph_node)
......@@ -10,11 +10,11 @@ endif()
add_subdirectory(api)
add_subdirectory(accumulation)
cc_library(grad_node_info SRCS grad_node_info.cc DEPS pten pten_api)
cc_library(grad_node_info SRCS grad_node_info.cc DEPS phi phi_api)
cc_library(grad_tensor_holder SRCS grad_tensor_holder.cc DEPS grad_node_info gradient_accumulator)
cc_library(autograd_meta SRCS autograd_meta.cc DEPS pten pten_api)
cc_library(utils SRCS utils.cc DEPS pten pten_api global_utils layer proto_desc operator op_registry variable_helper memcpy scale_op autograd_meta hook_utils)
cc_library(autograd_meta SRCS autograd_meta.cc DEPS phi phi_api)
cc_library(utils SRCS utils.cc DEPS phi phi_api global_utils layer proto_desc operator op_registry variable_helper memcpy scale_op autograd_meta hook_utils)
cc_library(backward SRCS backward.cc DEPS grad_tensor_holder utils autograd_meta grad_node_info)
add_subdirectory(tests)
paddle/fluid/eager/accumulation/CMakeLists.txt
浏览文件 @ 46894436
cc_library(accumulation_node SRCS accumulation_node.cc DEPS gradient_accumulator pten pten_api grad_node_info)
cc_library(accumulation_node SRCS accumulation_node.cc DEPS gradient_accumulator phi phi_api grad_node_info)
paddle/fluid/eager/accumulation/accumulation_node.cc
浏览文件 @ 46894436
......@@ -76,13 +76,13 @@ operator()(
}
void GradNodeAccumulation::RegisterReduceHook(
const std::function<void(void)>& hook) {
reduce_hooks_.emplace_back(hook);
std::shared_ptr<TensorVoidHook>&& hook) {
reduce_hooks_.emplace_back(std::move(hook));
}
void GradNodeAccumulation::ApplyReduceHooks() {
for (auto& hook : reduce_hooks_) {
hook();
(*hook)();
}
}
} // namespace egr
paddle/fluid/eager/accumulation/accumulation_node.h
浏览文件 @ 46894436
......@@ -16,6 +16,7 @@
#include "paddle/fluid/eager/autograd_meta.h"
#include "paddle/fluid/eager/grad_node_info.h"
#include "paddle/fluid/eager/hooks.h"
namespace egr {
......@@ -39,7 +40,7 @@ class GradNodeAccumulation : public GradNodeBase {
/**
* Register ReduceHook
* **/
void RegisterReduceHook(const std::function<void(void)>& hook);
void RegisterReduceHook(std::shared_ptr<TensorVoidHook>&& hook);
/**
* Apply ReduceHook here
......@@ -54,7 +55,7 @@ class GradNodeAccumulation : public GradNodeBase {
const paddle::experimental::Tensor&)>
retain_grad_hook_;
std::vector<std::function<void(void)>> reduce_hooks_;
std::vector<std::shared_ptr<TensorVoidHook>> reduce_hooks_;
};
} // namespace egr
paddle/fluid/eager/api/generated/eager_generated/backwards/CMakeLists.txt
浏览文件 @ 46894436
cc_library(scale_node SRCS scale_node.cc DEPS global_utils pten pten_api grad_node_info)
cc_library(scale_node SRCS scale_node.cc DEPS global_utils phi phi_api grad_node_info)
if(NOT ON_INFER)
cc_library(final_dygraph_node SRCS nodes.cc DEPS ${eager_deps})
......
paddle/fluid/eager/api/generated/eager_generated/backwards/scale_node.cc
浏览文件 @ 46894436
......@@ -33,36 +33,36 @@ static void ScaleDeviceDispatch(const phi::DenseTensor& dense_tensor,
phi::DenseTensor* dense_out) {
switch (dense_tensor.dtype()) {
case phi::DataType::FLOAT64: {
phi::ScaleKernel<double, typename paddle::framework::ConvertToPtenContext<
phi::ScaleKernel<double, typename paddle::framework::ConvertToPhiContext<
DeviceContext>::TYPE>(
static_cast<const typename paddle::framework::ConvertToPtenContext<
static_cast<const typename paddle::framework::ConvertToPhiContext<
DeviceContext>::TYPE&>(dev_ctx),
dense_tensor /* tensor */, scale /* scale */, bias /* bias */,
bias_after_scale /* bias_after_scale */, dense_out /* out tensor */);
break;
}
case phi::DataType::FLOAT32: {
phi::ScaleKernel<float, typename paddle::framework::ConvertToPtenContext<
phi::ScaleKernel<float, typename paddle::framework::ConvertToPhiContext<
DeviceContext>::TYPE>(
static_cast<const typename paddle::framework::ConvertToPtenContext<
static_cast<const typename paddle::framework::ConvertToPhiContext<
DeviceContext>::TYPE&>(dev_ctx),
dense_tensor /* tensor */, scale /* scale */, bias /* bias */,
bias_after_scale /* bias_after_scale */, dense_out /* out tensor */);
break;
}
case phi::DataType::INT64: {
phi::ScaleKernel<int64_t, typename paddle::framework::
ConvertToPtenContext<DeviceContext>::TYPE>(
static_cast<const typename paddle::framework::ConvertToPtenContext<
phi::ScaleKernel<int64_t, typename paddle::framework::ConvertToPhiContext<
DeviceContext>::TYPE>(
static_cast<const typename paddle::framework::ConvertToPhiContext<
DeviceContext>::TYPE&>(dev_ctx),
dense_tensor /* tensor */, scale /* scale */, bias /* bias */,
bias_after_scale /* bias_after_scale */, dense_out /* out tensor */);
break;
}
case phi::DataType::INT32: {
phi::ScaleKernel<int32_t, typename paddle::framework::
ConvertToPtenContext<DeviceContext>::TYPE>(
static_cast<const typename paddle::framework::ConvertToPtenContext<
phi::ScaleKernel<int32_t, typename paddle::framework::ConvertToPhiContext<
DeviceContext>::TYPE>(
static_cast<const typename paddle::framework::ConvertToPhiContext<
DeviceContext>::TYPE&>(dev_ctx),
dense_tensor /* tensor */, scale /* scale */, bias /* bias */,
bias_after_scale /* bias_after_scale */, dense_out /* out tensor */);
......
paddle/fluid/eager/api/generated/eager_generated/forwards/CMakeLists.txt
浏览文件 @ 46894436
cc_library(eager_scale SRCS scale.cc DEPS pten_api pten autograd_meta scale_node)
cc_library(eager_scale SRCS scale.cc DEPS phi_api phi autograd_meta scale_node)
if(NOT ON_INFER)
cc_library(final_dygraph_function SRCS dygraph_functions.cc DEPS ${eager_deps})
......
paddle/fluid/eager/api/utils/CMakeLists.txt
浏览文件 @ 46894436
cc_library(tensor_utils SRCS tensor_utils.cc DEPS pten pten_api autograd_meta grad_node_info accumulation_node)
cc_library(hook_utils SRCS hook_utils.cc DEPS pten tensor_utils autograd_meta grad_node_info utils accumulation_node)
cc_library(tensor_utils SRCS tensor_utils.cc DEPS phi phi_api autograd_meta grad_node_info accumulation_node)
cc_library(hook_utils SRCS hook_utils.cc DEPS phi tensor_utils autograd_meta grad_node_info utils accumulation_node)
cc_library(global_utils SRCS global_utils.cc DEPS place tracer)
paddle/fluid/eager/api/utils/hook_utils.cc
浏览文件 @ 46894436
......@@ -22,19 +22,19 @@
namespace egr {
namespace egr_utils_api {
void RegisterGradientHookForTensor(
int64_t RegisterGradientHookForTensor(
const paddle::experimental::Tensor& tensor,
std::function<paddle::experimental::Tensor(
const paddle::experimental::Tensor&)>& hook) {
std::shared_ptr<egr::TensorHook>&& hook) {
// Find grad_node and out_rank from AutogradMeta
std::shared_ptr<GradNodeBase> grad_node = EagerUtils::grad_node(tensor);
auto rank_info = EagerUtils::unsafe_autograd_meta(tensor)->OutRankInfo();
grad_node->RegisterGradientHook(rank_info.first, rank_info.second, hook);
return grad_node->RegisterGradientHook(rank_info.first, rank_info.second,
std::move(hook));
}
void RegisterReduceHookForTensor(const paddle::experimental::Tensor& tensor,
const std::function<void(void)>& hook) {
std::shared_ptr<egr::TensorVoidHook>&& hook) {
if (IsLeafTensor(tensor)) {
VLOG(6) << "Register ReduceHook for leaf tensor";
std::shared_ptr<GradNodeBase> grad_node = EagerUtils::grad_node(tensor);
......@@ -45,7 +45,7 @@ void RegisterReduceHookForTensor(const paddle::experimental::Tensor& tensor,
"with type: GradNodeAccumulation"));
auto accumulation_grad_node =
std::dynamic_pointer_cast<GradNodeAccumulation>(grad_node);
accumulation_grad_node->RegisterReduceHook(hook);
accumulation_grad_node->RegisterReduceHook(std::move(hook));
} else {
PADDLE_THROW(paddle::platform::errors::Fatal(
"Only can register reduce hook for leaf Tensor."));
......@@ -65,9 +65,7 @@ static void RetainGradForRegularNode(
meta->WeakGrad();
// Define Hook
std::function<paddle::experimental::Tensor(
const paddle::experimental::Tensor&)>
hook = [weak_grad_tensor](const paddle::experimental::Tensor& t) {
auto hook = [weak_grad_tensor](const paddle::experimental::Tensor& t) {
if (!weak_grad_tensor.expired()) {
auto grad_tensor = weak_grad_tensor.lock();
if (t.defined()) {
......@@ -86,7 +84,8 @@ static void RetainGradForRegularNode(
};
// Append to GradientHooks
RegisterGradientHookForTensor(tensor, hook);
RegisterGradientHookForTensor(tensor,
std::make_shared<egr::CppTensorHook>(hook));
}
void RetainGradForTensor(const paddle::experimental::Tensor& tensor) {
......
paddle/fluid/eager/api/utils/hook_utils.h
浏览文件 @ 46894436
......@@ -16,17 +16,17 @@
#include "paddle/fluid/eager/eager_tensor.h"
#include "paddle/fluid/eager/grad_node_info.h"
#include "paddle/fluid/eager/hooks.h"
#include "paddle/phi/api/all.h"
namespace egr {
namespace egr_utils_api {
void RegisterGradientHookForTensor(
int64_t RegisterGradientHookForTensor(
const paddle::experimental::Tensor& tensor,
std::function<paddle::experimental::Tensor(
const paddle::experimental::Tensor&)>& hook);
std::shared_ptr<egr::TensorHook>&& hook);
void RegisterReduceHookForTensor(const paddle::experimental::Tensor& tensor,
const std::function<void(void)>& hook);
std::shared_ptr<egr::TensorVoidHook>&& hook);
void RetainGradForTensor(const paddle::experimental::Tensor& tensor);
} // namespace egr_utils_api
......
paddle/fluid/eager/api/utils/tensor_utils.cc
浏览文件 @ 46894436
......@@ -22,7 +22,7 @@
#include "paddle/phi/api/all.h"
#include "paddle/fluid/framework/data_layout.h"
#include "paddle/fluid/framework/pten_utils.h"
#include "paddle/fluid/framework/phi_utils.h"
#include "paddle/fluid/framework/variable.h"
namespace egr {
......@@ -43,7 +43,7 @@ paddle::experimental::Tensor CreateTensorWithValue(
bool is_leaf) {
paddle::experimental::Tensor out = paddle::experimental::full(
phi::vectorize(ddim), paddle::experimental::Scalar(value), dtype,
phi::TransToPtenBackend(place));
phi::TransToPhiBackend(place));
auto meta = EagerUtils::autograd_meta(&out);
if (is_leaf) {
......
paddle/fluid/eager/auto_code_generator/eager_generator.cc
浏览文件 @ 46894436
......@@ -27,7 +27,7 @@
#include "paddle/fluid/pybind/pybind.h"
#include "paddle/fluid/string/string_helper.h"
// pten
// phi
#include "paddle/phi/kernels/declarations.h"
#define NUM_CREATED_DUP_INPUTS 4
......@@ -544,7 +544,7 @@ static bool CheckOpProto(proto::OpProto* op_proto) {
// since only OperatorWithKernel can run in dygraph mode.
auto& all_kernels = paddle::framework::OperatorWithKernel::AllOpKernels();
if (!all_kernels.count(op_type) &&
!phi::KernelFactory::Instance().HasCompatiblePtenKernel(op_type)) {
!phi::KernelFactory::Instance().HasCompatiblePhiKernel(op_type)) {
return false;
}
......@@ -2040,12 +2040,13 @@ static std::string GenerateGradNodeCCContents(
const char* BWD_RETURN_TEMPLATE =
" std::vector<std::vector<paddle::experimental::Tensor>> hooked_grads = "
"egr::GradNodeBase::ApplyGradientHooks(grads);\n"
"GradNode%s::ApplyGradientHooks(grads);\n"
" std::vector<std::vector<paddle::experimental::Tensor>> outputs(%d);\n"
" %s\n"
" return outputs;\n";
generated_grad_function_body = paddle::string::Sprintf(
BWD_RETURN_TEMPLATE, in_vars.size(), generated_grad_function_body);
generated_grad_function_body =
paddle::string::Sprintf(BWD_RETURN_TEMPLATE, fwd_op_type, in_vars.size(),
generated_grad_function_body);
// [Generation] Get Full Grad Function
const char* GRAD_FUNCTION_TEMPLATE =
......
paddle/fluid/eager/auto_code_generator/final_state_generator/python_c_gen.py
浏览文件 @ 46894436
......@@ -143,7 +143,7 @@ static PyObject * eager_final_state_api_{}(PyObject *self, PyObject *args, PyObj
fwd_api_name, fwd_api_name, get_eager_tensor_str, parse_attributes_str,
GetForwardFunctionName(fwd_api_name), dygraph_function_call_str)
python_c_function_reg_str = f"{{\"final_state_{fwd_api_name}\", (PyCFunction)(void(*)(void))eager_final_state_api_{fwd_api_name}, METH_VARARGS | METH_KEYWORDS, \"C++ interface function for {fwd_api_name} in dygraph.\"}},\n"
python_c_function_reg_str = f"{{\"final_state_{fwd_api_name}\", (PyCFunction)(void(*)(void))eager_final_state_api_{fwd_api_name}, METH_VARARGS | METH_KEYWORDS, \"C++ interface function for {fwd_api_name} in dygraph.\"}}\n"
return python_c_function_str, python_c_function_reg_str
......@@ -197,7 +197,7 @@ static PyObject * eager_get_final_state_core_ops_returns_info(PyObject *self) {
"""
core_ops_infos_registry = """
{\"get_final_state_core_ops_args_info\",
,{\"get_final_state_core_ops_args_info\",
(PyCFunction)(void(*)(void))eager_get_final_state_core_ops_args_info, METH_NOARGS,
\"C++ interface function for eager_get_final_state_core_ops_args_info.\"},
{\"get_final_state_core_ops_args_type_info\",
......
paddle/fluid/eager/eager_tensor.h
浏览文件 @ 46894436
......@@ -14,10 +14,10 @@
#pragma once
// framework deps
#include "paddle/fluid/framework/pten_utils.h"
#include "paddle/fluid/framework/phi_utils.h"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/framework/variable.h"
// pten deps
// Phi deps
#include "paddle/phi/api/include/tensor.h"
#include "paddle/phi/api/lib/api_declare.h"
#include "paddle/phi/api/lib/utils/tensor_utils.h"
......@@ -31,7 +31,7 @@
* provide variable in
* paddle::framework::ExecutionContext to support it. We should remove this as
* soon as we finish our latest
* Pten Lib, and use paddle::experimental::Tensor instead.
* Phi Lib, and use paddle::experimental::Tensor instead.
*
* Note: Keep this class as clean as possible.
* This class should only support method declared in
......
paddle/fluid/eager/grad_node_info.cc
浏览文件 @ 46894436
......@@ -210,22 +210,22 @@ const std::vector<std::vector<Edge>>& GradNodeBase::GetEdges() const {
return adj_edges_;
}
void GradNodeBase::RegisterGradientHook(
size_t slot_id, size_t rank,
const std::function<paddle::experimental::Tensor(
const paddle::experimental::Tensor&)>& hook) {
gradient_hooks_.emplace_back(std::make_tuple(slot_id, rank, hook));
int64_t GradNodeBase::RegisterGradientHook(
size_t slot_id, size_t rank, std::shared_ptr<egr::TensorHook>&& hook) {
gradient_hooks_.emplace(next_hook_id_,
std::make_tuple(slot_id, rank, std::move(hook)));
return next_hook_id_++;
}
std::vector<std::vector<paddle::experimental::Tensor>>
GradNodeBase::ApplyGradientHooks(
const std::vector<std::vector<paddle::experimental::Tensor>>& tensors) {
std::vector<std::vector<paddle::experimental::Tensor>> outs(tensors.size());
for (auto& tuple : gradient_hooks_) {
size_t slot_id = std::get<0>(tuple);
size_t rank = std::get<1>(tuple);
std::function<paddle::experimental::Tensor(
const paddle::experimental::Tensor&)>& hook = std::get<2>(tuple);
for (auto& hook_pair : gradient_hooks_) {
size_t slot_id = std::get<0>(hook_pair.second);
size_t rank = std::get<1>(hook_pair.second);
auto hook = std::get<2>(hook_pair.second);
PADDLE_ENFORCE(slot_id < tensors.size(),
paddle::platform::errors::Fatal(
......@@ -242,12 +242,11 @@ GradNodeBase::ApplyGradientHooks(
slot_out.resize(tensors[slot_id].size());
paddle::experimental::Tensor& out = slot_out[rank];
if (!out.defined() || !out.initialized()) {
VLOG(8) << "Run Hook for tensor: " << tensors[slot_id][rank].name();
out = hook(tensors[slot_id][rank]);
out = (*hook)(tensors[slot_id][rank]);
} else {
// If more than one hook is registered, the input to the next hook func
// should be the output of the previous hook
out = hook(out);
out = (*hook)(out);
}
}
......
paddle/fluid/eager/grad_node_info.h
浏览文件 @ 46894436
......@@ -15,6 +15,7 @@
#pragma once
#include "paddle/fluid/eager/eager_tensor.h"
#include "paddle/fluid/eager/hooks.h"
#include "paddle/phi/api/all.h"
namespace egr {
......@@ -135,14 +136,24 @@ class GradNodeBase {
/**
* Register GradientHook
* **/
void RegisterGradientHook(size_t slot_id, size_t rank,
const std::function<paddle::experimental::Tensor(
const paddle::experimental::Tensor&)>& hook);
int64_t RegisterGradientHook(size_t slot_id, size_t rank,
std::shared_ptr<egr::TensorHook>&& hook);
/**
* Remove GradientHook
* **/
bool RemoveGradientHook(const int64_t& hook_id) {
auto remove_cnt = gradient_hooks_.erase(hook_id);
if (remove_cnt == 0) {
return false;
}
return true;
}
/**
* Apply GradientHook
* **/
inline bool GradientHooksRegistered() { return gradient_hooks_.size() != 0; }
inline bool GradientHooksRegistered() { return !gradient_hooks_.empty(); }
std::vector<std::vector<paddle::experimental::Tensor>> ApplyGradientHooks(
const std::vector<std::vector<paddle::experimental::Tensor>>& tensors);
......@@ -166,12 +177,14 @@ class GradNodeBase {
// Gradient Hooks
// Customer may register a list of hooks which will be called in order during
// backward
// Each entry consists one pair of <out_rank, std::function>
std::vector<std::tuple<
// Each entry consists one pair of
// <hook_id, <out_rank, std::shared_ptr<TensorHook>>>
std::map<int64_t, std::tuple<
/* slot id */ size_t, /* rank */ size_t,
/* hook */ std::function<paddle::experimental::Tensor(
const paddle::experimental::Tensor&)>>>
/* hook */ std::shared_ptr<TensorHook>>>
gradient_hooks_;
int64_t next_hook_id_{0};
};
class Edge {
......
paddle/fluid/eager/hooks.h 0 → 100644
浏览文件 @ 46894436
// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed 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.
#pragma once
#include <functional>
#include <memory>
#include <utility>
#include <vector>
#include "paddle/phi/api/include/tensor.h"
namespace egr {
class TensorHook {
public:
virtual ~TensorHook() = default;
virtual paddle::experimental::Tensor operator()(
const paddle::experimental::Tensor& var) = 0;
};
class TensorVoidHook {
public:
virtual ~TensorVoidHook() = default;
virtual void operator()() = 0;
};
class CppTensorHook : public TensorHook {
public:
explicit CppTensorHook(std::function<paddle::experimental::Tensor(
const paddle::experimental::Tensor&)>&& fn)
: fn_(std::move(fn)) {}
paddle::experimental::Tensor operator()(
const paddle::experimental::Tensor& var) override {
return fn_(var);
}
private:
std::function<paddle::experimental::Tensor(
const paddle::experimental::Tensor&)>
fn_;
};
class CppTensorVoidHook : public TensorVoidHook {
public:
explicit CppTensorVoidHook(std::function<void()>&& fn) : fn_(std::move(fn)) {}
void operator()() override { return fn_(); }
private:
std::function<void()> fn_;
};
} // namespace egr
paddle/fluid/eager/tests/data_structure_tests/accumulation_node_test.cc
浏览文件 @ 46894436
......@@ -23,6 +23,7 @@
#include "paddle/fluid/eager/grad_tensor_holder.h"
#include "paddle/fluid/eager/utils.h"
#include "paddle/fluid/eager/hooks.h"
#include "paddle/phi/api/lib/utils/allocator.h"
#include "paddle/phi/core/kernel_registry.h"
......@@ -116,7 +117,8 @@ TEST(AccumulationNode, Tensor) {
VLOG(6) << "Running Reduce Hook";
};
node->RegisterReduceHook(reduce_hook_1);
node->RegisterReduceHook(
std::make_shared<egr::CppTensorVoidHook>(reduce_hook_1));
// operator()
paddle::experimental::Tensor _ret = node->operator()({{et0}})[0][0];
......@@ -141,7 +143,8 @@ TEST(AccumulationNode, Tensor) {
ret_et0_ptr[0] = 100.0; // set to 100.0
VLOG(6) << "Running Reduce Hook";
};
node->RegisterReduceHook(reduce_hook_2);
node->RegisterReduceHook(
std::make_shared<egr::CppTensorVoidHook>(reduce_hook_2));
node->ApplyReduceHooks();
// Check ApplyReduceHooks result
......
paddle/fluid/eager/tests/data_structure_tests/grad_node_info_test.cc
浏览文件 @ 46894436
......@@ -17,6 +17,7 @@
#include "paddle/fluid/eager/autograd_meta.h"
#include "paddle/fluid/eager/eager_tensor.h"
#include "paddle/fluid/eager/grad_node_info.h"
#include "paddle/fluid/eager/hooks.h"
#include "paddle/fluid/eager/tests/data_structure_tests/grad_node_test.h"
#include "paddle/phi/api/lib/utils/allocator.h"
......@@ -32,7 +33,7 @@ TEST(GradNodeInfo, GradSlotMeta) {
CHECK_EQ(grad_slot.Size(), 2);
}
TEST(GradNodeInfo, GradNodeBase) {
void TestGradNodeBase(bool is_remove_gradient_hook) {
VLOG(6) << "Construct Grad Node";
auto grad_test_node0 = std::make_shared<eager_test::GradTestNode>(
/* val */ 5.0, /* in_num */ 2, /* out_num */ 2);
......@@ -112,13 +113,25 @@ TEST(GradNodeInfo, GradNodeBase) {
VLOG(6) << "Running Gradient Hook";
return res;
};
grad_test_node0->RegisterGradientHook(0, 0, gradient_hook);
// 5 + 6
int64_t hook_id = grad_test_node0->RegisterGradientHook(
0, 0, std::make_shared<egr::CppTensorHook>(gradient_hook));
if (is_remove_gradient_hook) {
// Remove GradientHook
grad_test_node0->RemoveGradientHook(hook_id);
}
// Check results
auto grad_hook_res = grad_test_node0->ApplyGradientHooks(grads);
CHECK_EQ(
std::dynamic_pointer_cast<phi::DenseTensor>(grad_hook_res[0][0].impl())
->data<float>()[0],
11.0);
is_remove_gradient_hook ? 5.0 : 11.0);
}
TEST(GradNodeInfo, GradNodeBase) {
TestGradNodeBase(true);
TestGradNodeBase(false);
}
TEST(GradNodeInfo, Edge) {
......
paddle/fluid/eager/tests/task_tests/fwd_bwd_joint_test.cc
浏览文件 @ 46894436
......@@ -27,6 +27,7 @@
#include "paddle/phi/core/dense_tensor.h"
#include "paddle/phi/core/tensor_meta.h"
#include "paddle/fluid/eager/hooks.h"
#include "paddle/fluid/eager/tests/test_utils.h"
namespace egr {
......@@ -221,10 +222,6 @@ TEST(FwdBwdJoint, GradientHook) {
phi::DataLayout::NCHW, 5.0 /*value*/, true /*is_leaf*/);
egr_utils_api::RetainGradForTensor(tensor);
std::function<paddle::experimental::Tensor(
const paddle::experimental::Tensor&)>
hook = &hook_function;
// 3. Run Forward
// Run Forward Node 0
float scale0 = 2.0;
......@@ -233,7 +230,8 @@ TEST(FwdBwdJoint, GradientHook) {
egr::scale(tensor, scale0, bias0, true /*bias_after_scale*/,
true /*trace_backward*/);
egr_utils_api::RetainGradForTensor(out0); // hook: +5
egr_utils_api::RegisterGradientHookForTensor(out0, hook); // hook: +5
egr_utils_api::RegisterGradientHookForTensor(
out0, std::make_shared<egr::CppTensorHook>(hook_function)); // hook: +5
// Run Forward Node 1
float scale1 = 5.0;
......@@ -241,7 +239,8 @@ TEST(FwdBwdJoint, GradientHook) {
paddle::experimental::Tensor out1 = egr::scale(
out0, scale1, bias1, true /*bias_after_scale*/, true /*trace_backward*/);
egr_utils_api::RetainGradForTensor(out1); // hook: +5
egr_utils_api::RegisterGradientHookForTensor(out1, hook); // hook: +5
egr_utils_api::RegisterGradientHookForTensor(
out1, std::make_shared<egr::CppTensorHook>(hook_function)); // hook: +5
// Run Forward Node 2
float scale2 = 10.0;
......@@ -249,7 +248,8 @@ TEST(FwdBwdJoint, GradientHook) {
paddle::experimental::Tensor out2 = egr::scale(
out0, scale2, bias2, true /*bias_after_scale*/, true /*trace_backward*/);
egr_utils_api::RetainGradForTensor(out2); // hook: +5
egr_utils_api::RegisterGradientHookForTensor(out2, hook); // hook: +5
egr_utils_api::RegisterGradientHookForTensor(
out2, std::make_shared<egr::CppTensorHook>(hook_function)); // hook: +5
// 4. Run Backward
std::vector<paddle::experimental::Tensor> outs = {out1, out2};
......
paddle/fluid/eager/tests/task_tests/hook_test.cc
浏览文件 @ 46894436
......@@ -28,6 +28,7 @@
#include "paddle/phi/core/dense_tensor.h"
#include "paddle/phi/core/tensor_meta.h"
#include "paddle/fluid/eager/hooks.h"
#include "paddle/fluid/eager/tests/test_utils.h"
namespace egr {
......@@ -83,9 +84,6 @@ TEST(RetainGrad, HookBeforeRetainGrad) {
// Apply RetainGrad
{
// ScaleNode Hook: +3
std::function<paddle::experimental::Tensor(
const paddle::experimental::Tensor&)>
hook = &hook_function;
auto auto_grad_meta = std::make_shared<AutogradMeta>();
auto_grad_meta->SetGradNode(
......@@ -96,7 +94,8 @@ TEST(RetainGrad, HookBeforeRetainGrad) {
std::dynamic_pointer_cast<paddle::experimental::AbstractAutogradMeta>(
auto_grad_meta));
egr_utils_api::RegisterGradientHookForTensor(target_tensor, hook);
egr_utils_api::RegisterGradientHookForTensor(
target_tensor, std::make_shared<egr::CppTensorHook>(hook_function));
egr_utils_api::RetainGradForTensor(
target_tensor); // result: 1.0 + 3.0 = 4.0
egr_utils_api::RetainGradForTensor(
......@@ -107,9 +106,6 @@ TEST(RetainGrad, HookBeforeRetainGrad) {
paddle::experimental::Tensor leaf_tensor = paddle::experimental::Tensor();
{
// AccumulationNode Hook: +3
std::function<paddle::experimental::Tensor(
const paddle::experimental::Tensor&)>
hook = &hook_function;
auto auto_grad_meta = std::make_shared<AutogradMeta>();
......@@ -126,7 +122,8 @@ TEST(RetainGrad, HookBeforeRetainGrad) {
std::dynamic_pointer_cast<paddle::experimental::AbstractAutogradMeta>(
auto_grad_meta));
egr_utils_api::RegisterGradientHookForTensor(leaf_tensor, hook);
egr_utils_api::RegisterGradientHookForTensor(
leaf_tensor, std::make_shared<egr::CppTensorHook>(hook_function));
egr_utils_api::RetainGradForTensor(
leaf_tensor); // result: 4.0*5.0 + 3.0 = 23.0
}
......@@ -161,9 +158,6 @@ TEST(RetainGrad, HookAfterRetainGrad) {
// Apply RetainGrad
{
// ScaleNode Hook: +3
std::function<paddle::experimental::Tensor(
const paddle::experimental::Tensor&)>
hook = &hook_function;
auto auto_grad_meta = std::make_shared<AutogradMeta>();
auto_grad_meta->SetGradNode(
......@@ -175,16 +169,14 @@ TEST(RetainGrad, HookAfterRetainGrad) {
auto_grad_meta));
egr_utils_api::RetainGradForTensor(target_tensor); // result: 1.0
egr_utils_api::RegisterGradientHookForTensor(target_tensor, hook);
egr_utils_api::RegisterGradientHookForTensor(
target_tensor, std::make_shared<egr::CppTensorHook>(hook_function));
}
// Retain Grad for leaf tensor1
paddle::experimental::Tensor leaf_tensor = paddle::experimental::Tensor();
{
// AccumulationNode Hook: +3
std::function<paddle::experimental::Tensor(
const paddle::experimental::Tensor&)>
hook = &hook_function;
auto auto_grad_meta = std::make_shared<AutogradMeta>();
auto acc_node_ptr =
......@@ -199,7 +191,8 @@ TEST(RetainGrad, HookAfterRetainGrad) {
std::dynamic_pointer_cast<paddle::experimental::AbstractAutogradMeta>(
auto_grad_meta));
egr_utils_api::RegisterGradientHookForTensor(leaf_tensor, hook);
egr_utils_api::RegisterGradientHookForTensor(
leaf_tensor, std::make_shared<egr::CppTensorHook>(hook_function));
}
RunBackward(target_tensors, {});
......
paddle/fluid/eager/tests/task_tests/hook_test_intermidiate.cc
浏览文件 @ 46894436
......@@ -24,6 +24,7 @@
#include "paddle/phi/core/dense_tensor.h"
#include "paddle/fluid/eager/api/generated/fluid_generated/dygraph_forward_api.h"
#include "paddle/fluid/eager/hooks.h"
#include "paddle/phi/core/kernel_registry.h"
namespace egr {
......@@ -54,7 +55,7 @@ paddle::experimental::Tensor hook_function(
return ret;
}
TEST(Hook_intermidiate, Sigmoid) {
void test_sigmoid(bool is_remove_gradient_hook) {
// Prepare Device Contexts
VLOG(6) << "Init Env";
eager_test::InitEnv(paddle::platform::CPUPlace());
......@@ -67,11 +68,6 @@ TEST(Hook_intermidiate, Sigmoid) {
ddim, paddle::platform::CPUPlace(), phi::DataType::FLOAT32,
phi::DataLayout::NCHW, 0.0, true);
VLOG(6) << "Make Hook function";
std::function<paddle::experimental::Tensor(
const paddle::experimental::Tensor&)>
hook = &hook_function;
VLOG(6) << "Make ReduceHook function";
auto reduce_hook = [&](void) -> void {
auto* t_ptr = std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl())
......@@ -85,10 +81,12 @@ TEST(Hook_intermidiate, Sigmoid) {
egr_utils_api::RetainGradForTensor(tensor);
VLOG(6) << "Register GradientHook for Tensor";
egr_utils_api::RegisterGradientHookForTensor(tensor, hook);
int64_t hook_id = egr_utils_api::RegisterGradientHookForTensor(
tensor, std::make_shared<CppTensorHook>(hook_function));
VLOG(6) << "Register ReduceHook for Tensor";
egr_utils_api::RegisterReduceHookForTensor(tensor, reduce_hook);
egr_utils_api::RegisterReduceHookForTensor(
tensor, std::make_shared<CppTensorVoidHook>(reduce_hook));
VLOG(6) << "Runing Forward";
auto output_tensor = sigmoid_dygraph_function(tensor, {});
......@@ -98,11 +96,17 @@ TEST(Hook_intermidiate, Sigmoid) {
std::vector<paddle::experimental::Tensor> target_tensors = {output_tensor};
if (is_remove_gradient_hook) {
std::shared_ptr<GradNodeBase> grad_node_tmp = EagerUtils::grad_node(tensor);
grad_node_tmp->RemoveGradientHook(hook_id);
}
VLOG(6) << "Runing Backward";
RunBackward(target_tensors, {});
VLOG(6) << "Finish Backward";
eager_test::CompareGradTensorWithValue<float>(tensor, 0.25 + 3);
eager_test::CompareGradTensorWithValue<float>(
tensor, is_remove_gradient_hook ? 0.25 : 0.25 + 3.0);
VLOG(6) << "Checking ReduceHook results";
for (int i = 0; i < tensor.numel(); i++) {
......@@ -113,7 +117,7 @@ TEST(Hook_intermidiate, Sigmoid) {
VLOG(6) << "After Tests";
}
TEST(Hook_intermidiate, ElementwiseAdd) {
void test_elementwiseAdd(bool is_remove_gradient_hook) {
// Prepare Device Contexts
eager_test::InitEnv(paddle::platform::CPUPlace());
......@@ -132,11 +136,7 @@ TEST(Hook_intermidiate, ElementwiseAdd) {
ddimY, paddle::platform::CPUPlace(), phi::DataType::FLOAT32,
phi::DataLayout::NCHW, 2.0, true);
std::function<paddle::experimental::Tensor(
const paddle::experimental::Tensor&)>
hook = &hook_function;
auto reduce_hook = [&](void) -> void {
auto reduce_hook = [&]() -> void {
auto* t_ptr =
std::dynamic_pointer_cast<phi::DenseTensor>(Y.impl())->data<float>();
for (int i = 0; i < Y.numel(); i++) {
......@@ -145,18 +145,26 @@ TEST(Hook_intermidiate, ElementwiseAdd) {
};
egr_utils_api::RetainGradForTensor(Y);
egr_utils_api::RegisterGradientHookForTensor(Y, hook);
egr_utils_api::RegisterReduceHookForTensor(Y, reduce_hook);
int64_t hook_id = egr_utils_api::RegisterGradientHookForTensor(
Y, std::make_shared<CppTensorHook>(hook_function));
egr_utils_api::RegisterReduceHookForTensor(
Y, std::make_shared<CppTensorVoidHook>(reduce_hook));
auto output_tensor = elementwise_add_dygraph_function(X, Y, {});
eager_test::CompareTensorWithValue<float>(output_tensor, 5);
std::vector<paddle::experimental::Tensor> target_tensors = {output_tensor};
if (is_remove_gradient_hook) {
std::shared_ptr<GradNodeBase> grad_node_tmp = EagerUtils::grad_node(Y);
grad_node_tmp->RemoveGradientHook(hook_id);
}
RunBackward(target_tensors, {});
eager_test::CompareGradTensorWithValue<float>(X, 1.0);
eager_test::CompareGradTensorWithValue<float>(Y, 4.0);
eager_test::CompareGradTensorWithValue<float>(
Y, is_remove_gradient_hook ? 1.0 : 1.0 + 3.0);
// Checking ReduceHook results
for (int i = 0; i < Y.numel(); i++) {
......@@ -166,7 +174,7 @@ TEST(Hook_intermidiate, ElementwiseAdd) {
}
}
TEST(Hook_intermidiate, Matmul_v2) {
void test_matmul(bool is_remove_gradient_hook) {
// Prepare Device Contexts
eager_test::InitEnv(paddle::platform::CPUPlace());
......@@ -185,10 +193,6 @@ TEST(Hook_intermidiate, Matmul_v2) {
ddimY, paddle::platform::CPUPlace(), phi::DataType::FLOAT32,
phi::DataLayout::NCHW, 2.0, true);
std::function<paddle::experimental::Tensor(
const paddle::experimental::Tensor&)>
hook = &hook_function;
auto reduce_hook = [&](void) -> void {
auto* t_ptr =
std::dynamic_pointer_cast<phi::DenseTensor>(Y.impl())->data<float>();
......@@ -198,19 +202,27 @@ TEST(Hook_intermidiate, Matmul_v2) {
};
egr_utils_api::RetainGradForTensor(Y);
egr_utils_api::RegisterGradientHookForTensor(Y, hook);
egr_utils_api::RegisterReduceHookForTensor(Y, reduce_hook);
int64_t hook_id = egr_utils_api::RegisterGradientHookForTensor(
Y, std::make_shared<CppTensorHook>(hook_function));
egr_utils_api::RegisterReduceHookForTensor(
Y, std::make_shared<CppTensorVoidHook>(reduce_hook));
auto output_tensor = matmul_v2_dygraph_function(
X, Y, {{"trans_x", false}, {"trans_y", false}});
eager_test::CompareTensorWithValue<float>(output_tensor, 96);
std::vector<paddle::experimental::Tensor> target_tensors = {output_tensor};
if (is_remove_gradient_hook) {
std::shared_ptr<GradNodeBase> grad_node_tmp = EagerUtils::grad_node(Y);
grad_node_tmp->RemoveGradientHook(hook_id);
}
RunBackward(target_tensors, {});
eager_test::CompareGradTensorWithValue<float>(X, 2.0 * 20);
eager_test::CompareGradTensorWithValue<float>(Y, 3.0 * 4 + 3);
eager_test::CompareGradTensorWithValue<float>(
Y, is_remove_gradient_hook ? 3.0 * 4 : 3.0 * 4 + 3);
// Checking ReduceHook results
for (int i = 0; i < Y.numel(); i++) {
......@@ -219,6 +231,22 @@ TEST(Hook_intermidiate, Matmul_v2) {
static_cast<float>(100.0f));
}
}
TEST(Hook_intermidiate, Sigmoid) {
// True or false represents whether to call RemoveGradientHook
test_sigmoid(true);
test_sigmoid(false);
}
TEST(Hook_intermidiate, ElementwiseAdd) {
test_elementwiseAdd(true);
test_elementwiseAdd(false);
}
TEST(Hook_intermidiate, Matmul_v2) {
test_matmul(true);
test_matmul(false);
}
} // namespace egr
USE_OP(sigmoid);
......
paddle/fluid/eager/utils.cc
浏览文件 @ 46894436
......@@ -23,7 +23,7 @@
#include "paddle/phi/core/tensor_meta.h"
#include "paddle/fluid/framework/data_layout.h"
#include "paddle/fluid/framework/pten_utils.h"
#include "paddle/fluid/framework/phi_utils.h"
#include "paddle/fluid/framework/variable.h"
PADDLE_DEFINE_EXPORTED_bool(retain_grad_for_all_tensor, true,
......
paddle/fluid/framework/CMakeLists.txt
浏览文件 @ 46894436
......@@ -193,19 +193,19 @@ cc_library(unused_var_check SRCS unused_var_check.cc DEPS glog no_need_buffer_va
cc_library(op_kernel_type SRCS op_kernel_type.cc DEPS device_context place)
IF(WITH_XPU)
cc_library(pten_utils SRCS pten_utils.cc DEPS lod_tensor selected_rows_utils place pten var_type_traits pten_api_utils op_info xpu_op_list)
cc_library(phi_utils SRCS phi_utils.cc DEPS lod_tensor selected_rows_utils place phi var_type_traits phi_api_utils op_info xpu_op_list)
ELSE()
cc_library(pten_utils SRCS pten_utils.cc DEPS lod_tensor selected_rows_utils place pten var_type_traits pten_api_utils op_info)
cc_library(phi_utils SRCS phi_utils.cc DEPS lod_tensor selected_rows_utils place phi var_type_traits phi_api_utils op_info)
ENDIF()
IF(WITH_XPU)
cc_library(operator SRCS operator.cc DEPS xpu_op_list op_info device_context tensor scope glog trainer_desc_proto data_feed_proto
shape_inference data_transform lod_tensor profiler transfer_scope_cache op_kernel_type op_call_stack unused_var_check nan_inf_utils
pten pten_utils kernel_factory infershape_utils op_utils)
phi phi_utils kernel_factory infershape_utils op_utils)
ELSE()
cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog trainer_desc_proto data_feed_proto
shape_inference data_transform lod_tensor profiler transfer_scope_cache op_kernel_type op_call_stack unused_var_check nan_inf_utils
pten pten_utils kernel_factory infershape_utils op_utils)
phi phi_utils kernel_factory infershape_utils op_utils)
ENDIF()
cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry device_context)
......@@ -412,7 +412,7 @@ cc_library(save_load_util SRCS save_load_util.cc DEPS tensor scope layer)
cc_test(save_load_util_test SRCS save_load_util_test.cc DEPS save_load_util tensor scope layer)
cc_library(generator SRCS generator.cc DEPS enforce place)
cc_library(infershape_utils SRCS infershape_utils.cc DEPS lod_tensor selected_rows_utils attribute place pten var_type_traits pten pten_api_utils op_info shape_inference)
cc_library(infershape_utils SRCS infershape_utils.cc DEPS lod_tensor selected_rows_utils attribute place phi var_type_traits phi phi_api_utils op_info shape_inference)
cc_test(infershape_utils_test SRCS infershape_utils_test.cc DEPS infershape_utils infermeta_utils meta_tensor)
# Get the current working branch
......@@ -436,8 +436,8 @@ message(STATUS "branch: ${PADDLE_BRANCH}")
configure_file(commit.h.in commit.h)
cc_library(custom_operator SRCS custom_operator.cc DEPS tensor attribute framework_proto op_registry operator dynamic_loader string_helper pten_tensor op_meta_info pten_api)
cc_library(custom_kernel SRCS custom_kernel.cc DEPS op_registry pten_custom_kernel pten_tensor_raw)
cc_library(custom_operator SRCS custom_operator.cc DEPS tensor attribute framework_proto op_registry operator dynamic_loader string_helper phi_tensor op_meta_info phi_api)
cc_library(custom_kernel SRCS custom_kernel.cc DEPS op_registry phi_custom_kernel phi_tensor_raw)
#cc_binary(test_executor SRCS test_executor.cc DEPS executor op_registry ${GLOB_OP_LIB} ${GLOB_OPERATOR_DEPS} )
#cc_binary(new_executor SRCS new_exec_test.cc DEPS operator op_registry executor ${GLOB_OP_LIB} ${GLOB_OPERATOR_DEPS} profiler)
......@@ -450,7 +450,7 @@ if(WITH_TESTING AND TEST selected_rows_utils_test)
endif()
cc_test(scope_guard_test SRCS scope_guard_test.cc)
cc_test(pten_utils_test SRCS pten_utils_test.cc DEPS pten_utils)
cc_test(phi_utils_test SRCS phi_utils_test.cc DEPS phi_utils)
if(WITH_GPU OR WITH_ROCM)
cc_library(fluid_convert_utils SRCS convert_utils.cc DEPS data_type place gpu_info)
......
paddle/fluid/framework/async_executor.cc
浏览文件 @ 46894436
......@@ -33,7 +33,7 @@ limitations under the License. */
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/pybind/pybind.h"
// pten
// phi
#include "paddle/phi/kernels/declarations.h"
namespace paddle {
......
paddle/fluid/framework/convert_utils.cc
浏览文件 @ 46894436
......@@ -18,7 +18,7 @@ limitations under the License. */
namespace paddle {
namespace framework {
paddle::experimental::DataType TransToPtenDataType(
paddle::experimental::DataType TransToPhiDataType(
const paddle::framework::proto::VarType::Type& dtype) {
// Set the order of case branches according to the frequency with
// the data type is used
......
paddle/fluid/framework/convert_utils.h
浏览文件 @ 46894436
......@@ -32,7 +32,7 @@ namespace framework {
using DataType = paddle::experimental::DataType;
using DataLayout = paddle::experimental::DataLayout;
DataType TransToPtenDataType(
DataType TransToPhiDataType(
const paddle::framework::proto::VarType::Type& dtype);
paddle::framework::proto::VarType::Type TransToProtoVarType(
......
paddle/fluid/framework/convert_utils_test.cc
浏览文件 @ 46894436
......@@ -43,35 +43,35 @@ TEST(ConvertUtils, DataType) {
CHECK(paddle::framework::TransToProtoVarType(paddle::DataType::FLOAT16) ==
paddle::framework::proto::VarType::FP16);
// proto -> enum
CHECK(paddle::framework::TransToPtenDataType(
CHECK(paddle::framework::TransToPhiDataType(
paddle::framework::proto::VarType::FP64) ==
paddle::DataType::FLOAT64);
CHECK(paddle::framework::TransToPtenDataType(
CHECK(paddle::framework::TransToPhiDataType(
paddle::framework::proto::VarType::FP32) ==
paddle::DataType::FLOAT32);
CHECK(paddle::framework::TransToPtenDataType(
CHECK(paddle::framework::TransToPhiDataType(
paddle::framework::proto::VarType::INT64) ==
paddle::DataType::INT64);
CHECK(paddle::framework::TransToPtenDataType(
CHECK(paddle::framework::TransToPhiDataType(
paddle::framework::proto::VarType::INT32) ==
paddle::DataType::INT32);
CHECK(paddle::framework::TransToPtenDataType(
CHECK(paddle::framework::TransToPhiDataType(
paddle::framework::proto::VarType::INT8) == paddle::DataType::INT8);
CHECK(paddle::framework::TransToPtenDataType(
CHECK(paddle::framework::TransToPhiDataType(
paddle::framework::proto::VarType::UINT8) ==
paddle::DataType::UINT8);
CHECK(paddle::framework::TransToPtenDataType(
CHECK(paddle::framework::TransToPhiDataType(
paddle::framework::proto::VarType::INT16) ==
paddle::DataType::INT16);
CHECK(paddle::framework::TransToPtenDataType(
CHECK(paddle::framework::TransToPhiDataType(
paddle::framework::proto::VarType::BOOL) == paddle::DataType::BOOL);
CHECK(paddle::framework::TransToPtenDataType(
CHECK(paddle::framework::TransToPhiDataType(
paddle::framework::proto::VarType::COMPLEX64) ==
paddle::DataType::COMPLEX64);
CHECK(paddle::framework::TransToPtenDataType(
CHECK(paddle::framework::TransToPhiDataType(
paddle::framework::proto::VarType::COMPLEX128) ==
paddle::DataType::COMPLEX128);
CHECK(paddle::framework::TransToPtenDataType(
CHECK(paddle::framework::TransToPhiDataType(
paddle::framework::proto::VarType::FP16) ==
paddle::DataType::FLOAT16);
}
......
paddle/fluid/framework/custom_operator.cc
浏览文件 @ 46894436
......@@ -30,7 +30,7 @@ limitations under the License. */
#include "paddle/fluid/framework/op_meta_info_helper.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/pten_utils.h"
#include "paddle/fluid/framework/phi_utils.h"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/platform/dynload/dynamic_loader.h"
#include "paddle/fluid/string/string_helper.h"
......@@ -779,13 +779,13 @@ void RegisterOperatorWithMetaInfo(const std::vector<OpMetaInfo>& op_meta_infos,
for (size_t i = 0; i < ctx->InputSize(in_name); ++i) {
auto dtype = ctx->GetInputDataType(in_name, i);
vec_custom_dtype.emplace_back(
paddle::framework::TransToPtenDataType(dtype));
paddle::framework::TransToPhiDataType(dtype));
}
vec_input_dtypes.emplace_back(vec_custom_dtype);
} else {
auto dtype = ctx->GetInputDataType(in_name);
input_dtypes.emplace_back(
paddle::framework::TransToPtenDataType(dtype));
paddle::framework::TransToPhiDataType(dtype));
}
}
......
paddle/fluid/framework/data_device_transform_test.cu
浏览文件 @ 46894436
......@@ -23,7 +23,7 @@ limitations under the License. */
#include "paddle/fluid/platform/init.h"
#include "paddle/phi/kernels/funcs/math_function.h"
#include "paddle/fluid/framework/pten_utils.h"
#include "paddle/fluid/framework/phi_utils.h"
namespace paddle {
namespace framework {
......
paddle/fluid/framework/data_type_test.cc
浏览文件 @ 46894436
......@@ -28,7 +28,7 @@ TEST(DataType, float16) {
Tensor tensor;
CPUPlace cpu;
tensor.mutable_data(cpu, f::TransToPtenDataType(dtype));
tensor.mutable_data(cpu, f::TransToPhiDataType(dtype));
// test fp16 tensor
EXPECT_EQ(f::TransToProtoVarType(tensor.dtype()),
......@@ -51,7 +51,7 @@ TEST(DataType, bfloat16) {
Tensor tensor;
CPUPlace cpu;
tensor.mutable_data(cpu, f::TransToPtenDataType(dtype));
tensor.mutable_data(cpu, f::TransToPhiDataType(dtype));
// test bf16 tensor
EXPECT_EQ(f::TransToProtoVarType(tensor.dtype()),
......
paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc
浏览文件 @ 46894436
......@@ -231,6 +231,8 @@ void FastThreadedSSAGraphExecutor::RunOpAsync(
OpHandleBase *op,
const std::shared_ptr<BlockingQueue<size_t>> &complete_q) {
++remaining_;
platform::RecordEvent("WorkQueue::AddTask",
platform::TracerEventType::UserDefined, 10 /*level*/);
this->pool_->enqueue([=] {
std::deque<OpHandleBase *> op_queue;
op_queue.push_front(op);
......
paddle/fluid/framework/executor_thread_worker.cc
浏览文件 @ 46894436
......@@ -34,7 +34,7 @@ limitations under the License. */
#include "paddle/fluid/platform/timer.h"
#include "paddle/fluid/pybind/pybind.h"
// pten
// phi
#include "paddle/phi/kernels/declarations.h"
namespace paddle {
namespace framework {
......
paddle/fluid/framework/fleet/heter_wrapper.cc
浏览文件 @ 46894436
......@@ -161,7 +161,7 @@ void HeterWrapper::DeSerializeToTensor(Scope* scope,
tensor->set_lod(lod);
void* tensor_data = tensor->mutable_data(
place, framework::TransToPtenDataType(ToVarType(req_var.data_type())));
place, framework::TransToPhiDataType(ToVarType(req_var.data_type())));
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
memory::Copy(place, tensor_data, platform::CPUPlace(), req_var.data().data(),
......@@ -202,7 +202,7 @@ void HeterWrapper::DeSerializeToTensor(Scope* scope,
tensor->set_lod(lod);
void* tensor_data = tensor->mutable_data(
place, framework::TransToPtenDataType(ToVarType(req_var.data_type())));
place, framework::TransToPhiDataType(ToVarType(req_var.data_type())));
#ifdef PADDLE_WITH_XPU
memory::Copy(place, tensor_data, platform::CPUPlace(), req_var.data().data(),
......
paddle/fluid/framework/heter_section_worker.cc
浏览文件 @ 46894436
......@@ -38,7 +38,7 @@ void SetMicroId(paddle::framework::Scope* scope,
std::vector<int> dims{1};
tensor->Resize(phi::make_ddim(dims));
void* tensor_data = tensor->mutable_data(
place, framework::TransToPtenDataType(framework::proto::VarType::FP32));
place, framework::TransToPhiDataType(framework::proto::VarType::FP32));
if (platform::is_gpu_place(place)) {
#ifdef PADDLE_WITH_CUDA
std::vector<char> temp;
......
paddle/fluid/framework/infershape_utils.cc
浏览文件 @ 46894436
......@@ -18,7 +18,7 @@ limitations under the License. */
#include "paddle/fluid/framework/convert_utils.h"
#include "paddle/fluid/framework/framework.pb.h"
#include "paddle/fluid/framework/pten_utils.h"
#include "paddle/fluid/framework/phi_utils.h"
#include "paddle/fluid/platform/enforce.h"
#include "paddle/phi/common/scalar.h"
#include "paddle/phi/common/scalar_array.h"
......@@ -144,7 +144,7 @@ class CompatMetaTensor : public phi::MetaTensor {
}
} else {
auto* var = BOOST_GET_CONST(VarDesc*, var_);
return paddle::framework::TransToPtenDataType(var->GetDataType());
return paddle::framework::TransToPhiDataType(var->GetDataType());
}
}
......@@ -341,24 +341,37 @@ phi::InferMetaContext BuildInferMetaContext(InferShapeContext* ctx,
}
if (infershape_inputs.size() != 1) {
infer_meta_context.EmplaceBackAttr(
std::move(experimental::MakePtenScalarArrayFromVarList(vars)));
std::move(experimental::MakePhiScalarArrayFromVarList(vars)));
} else {
infer_meta_context.EmplaceBackAttr(
std::move(experimental::MakePtenScalarArrayFromVar(*vars[0])));
std::move(experimental::MakePhiScalarArrayFromVar(*vars[0])));
}
} else {
// If is not in runtime, we will set default value(-1) for ScalarArray
int64_t num_ele = 1;
int64_t num_ele = 0;
std::vector<VarDesc*> vars;
vars.reserve(infershape_inputs.size());
for (size_t i = 0; i < infershape_inputs.size(); i++) {
vars.push_back(BOOST_GET_CONST(VarDesc*, infershape_inputs[i]));
}
for (auto& var : vars) {
const auto& tensor_dims = var->GetShape();
if (vars.size() == 1) {
num_ele = 1;
const auto& tensor_dims = vars[0]->GetShape();
for (size_t i = 0; i < tensor_dims.size(); ++i) {
num_ele *= tensor_dims[i];
}
} else {
for (auto& var : vars) {
const auto& tensor_dims = var->GetShape();
PADDLE_ENFORCE_EQ(tensor_dims.size(), 1,
platform::errors::InvalidArgument(
"The shape is constructed by multi-tensor, "
"every tensor's dims should be 1. But your "
"shape has tensor that dims is %s.",
tensor_dims.size()));
num_ele += tensor_dims[0];
}
}
phi::ScalarArray tensor_attr(std::vector<int32_t>(num_ele, -1));
tensor_attr.SetFromTensor(true);
......@@ -406,7 +419,7 @@ phi::InferMetaContext BuildInferMetaContext(InferShapeContext* ctx,
if (ctx->IsRuntime()) {
Variable* var = BOOST_GET_CONST(Variable*, infershape_input[0]);
infer_meta_context.EmplaceBackAttr(
std::move(experimental::MakePtenScalarFromVar(*var)));
std::move(experimental::MakePhiScalarFromVar(*var)));
} else {
phi::Scalar tensor_scalar(-1);
tensor_scalar.SetFromTensor(true);
......@@ -468,7 +481,7 @@ phi::InferMetaContext BuildInferMetaContext(InferShapeContext* ctx,
BOOST_GET_CONST(std::vector<std::string>, attr));
} else if (attr_defs[i].type_index ==
std::type_index(typeid(phi::DataType))) {
auto data_type = paddle::framework::TransToPtenDataType(
auto data_type = paddle::framework::TransToPhiDataType(
static_cast<framework::proto::VarType::Type>(
BOOST_GET_CONST(int, attr)));
infer_meta_context.EmplaceBackAttr(data_type);
......
paddle/fluid/framework/ir/fuse_optimizer_ops_pass/fuse_optimizer_op_pass.cc
浏览文件 @ 46894436
......@@ -276,13 +276,13 @@ bool FuseOptimizerOpPass::OpWithKernelSupportCPUAndGPU(
bool support_gpu = false;
auto &kernel_factory = phi::KernelFactory::Instance();
auto kernel_key_map =
kernel_factory.SelectKernelMap(phi::TransToPtenKernelName(op_type));
kernel_factory.SelectKernelMap(phi::TransToPhiKernelName(op_type));
bool has_op_kernel = kernel_key_map.size() > 0 ? true : false;
for (auto &kernel : kernel_key_map) {
if (platform::is_gpu_place(phi::TransToPtenPlace(kernel.first.backend()))) {
if (platform::is_gpu_place(phi::TransToPhiPlace(kernel.first.backend()))) {
support_gpu = true;
} else if (platform::is_cpu_place(
phi::TransToPtenPlace(kernel.first.backend()))) {
phi::TransToPhiPlace(kernel.first.backend()))) {
support_cpu = true;
}
}
......
paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass_tester.cc
浏览文件 @ 46894436
......@@ -96,7 +96,7 @@ void InitTensorHolder(Scope* scope, const paddle::platform::Place& place,
auto x = scope->Var(var_name);
auto tensor = x->GetMutable<LoDTensor>();
tensor->mutable_data(place,
framework::TransToPtenDataType(proto::VarType::FP32), 1);
framework::TransToPhiDataType(proto::VarType::FP32), 1);
}
void MainTest(bool convWithExistingBias) {
......
paddle/fluid/framework/ir/mkldnn/cpu_quantize_pass_tester.cc
浏览文件 @ 46894436
......@@ -126,7 +126,7 @@ void InitTensorHolder(Scope* scope, const paddle::platform::Place& place,
auto x = scope->Var(var_name);
auto tensor = x->GetMutable<LoDTensor>();
tensor->mutable_data(place,
framework::TransToPtenDataType(proto::VarType::FP32), 1);
framework::TransToPhiDataType(proto::VarType::FP32), 1);
}
void PreparePass(std::unique_ptr<ir::Graph>* graph, const ProgramDesc& prog,
......
paddle/fluid/framework/ir/mkldnn/cpu_quantize_squash_pass_tester.cc
浏览文件 @ 46894436
......@@ -526,7 +526,7 @@ void InitTensorHolder(Scope* scope, const paddle::platform::Place& place,
auto x = scope->Var(var_name);
auto tensor = x->GetMutable<LoDTensor>();
tensor->mutable_data(place,
framework::TransToPtenDataType(proto::VarType::FP32), 1);
framework::TransToPhiDataType(proto::VarType::FP32), 1);
}
void PrepareGraph(std::unique_ptr<ir::Graph>* graph, const ProgramDesc& prog) {
......
paddle/fluid/framework/ir/mkldnn/mkldnn_inplace_pass_tester.cc
浏览文件 @ 46894436
......@@ -20,7 +20,7 @@
#include "paddle/fluid/framework/ir/pass_tester_helper.h"
#include "paddle/fluid/framework/op_registry.h"
USE_OP(softmax);
USE_OP_ITSELF(softmax);
USE_OP_DEVICE_KERNEL(softmax, MKLDNN);
USE_OP_ITSELF(elementwise_add);
USE_OP_DEVICE_KERNEL(elementwise_add, MKLDNN);
......
paddle/fluid/framework/lod_tensor.cc
浏览文件 @ 46894436
......@@ -447,7 +447,7 @@ void MergeLoDTensor(LoDTensor *target,
target->set_layout(new_layout);
target->set_lod(new_lod);
target->mutable_data(dst_place,
paddle::framework::TransToPtenDataType(new_type));
paddle::framework::TransToPhiDataType(new_type));
int begin = 0;
for (auto *src : lod_tensors) {
......
paddle/fluid/framework/new_executor/interpretercore.cc
浏览文件 @ 46894436
......@@ -389,7 +389,7 @@ void InterpreterCore::RunInstruction(const Instruction& instr_node) {
auto op_with_kernel = dynamic_cast<const framework::OperatorWithKernel*>(op);
{
platform::RecordEvent infershape_event(
"InferShape", platform::TracerEventType::OperatorInner, 1,
"infer_shape", platform::TracerEventType::OperatorInner, 1,
platform::EventRole::kInnerOp);
// If it is OperatorBase, InferShape do nothing.
if (op_with_kernel != nullptr)
......@@ -411,23 +411,23 @@ void InterpreterCore::RunInstruction(const Instruction& instr_node) {
}
{
platform::RecordEvent compute_event(
"Compute", platform::TracerEventType::OperatorInner, 1,
"compute", platform::TracerEventType::OperatorInner, 1,
platform::EventRole::kInnerOp);
if (op_with_kernel == nullptr) {
instr_node.OpBase()->Run(*local_scope, place_);
} else {
// fit for pten
if (instr_node.PtenKernel() && instr_node.PtenKernel()->IsValid()) {
VLOG(4) << "Run pten kernel: " << op->Type();
// fit for phi
if (instr_node.PhiKernel() && instr_node.PhiKernel()->IsValid()) {
VLOG(4) << "Run phi kernel: " << op->Type();
VLOG(4) << instr_node.InnerRuntimeContext().get() << " "
<< &instr_node.DeviceContext();
phi::KernelContext pt_kernel_context;
op_with_kernel->BuildPtenKernelContext(
op_with_kernel->BuildPhiKernelContext(
*instr_node.InnerRuntimeContext().get(),
const_cast<platform::DeviceContext*>(&instr_node.DeviceContext()),
&pt_kernel_context);
(*instr_node.PtenKernel())(&pt_kernel_context);
(*instr_node.PhiKernel())(&pt_kernel_context);
} else {
instr_node.KernelFunc()(*instr_node.InnerExecutionContext().get());
......@@ -561,7 +561,8 @@ void InterpreterCore::RunInstructionAsync(size_t instr_id) {
<< " runs on " << platform::GetCurrentThreadName();
auto* op = instr_node.OpBase();
platform::RecordEvent instruction_event(op->Type().c_str());
platform::RecordEvent instruction_event(
op->Type(), platform::TracerEventType::Operator, 1);
interpreter::WaitEvent(instr_node, place_);
try {
......
paddle/fluid/framework/new_executor/interpretercore_util.cc
浏览文件 @ 46894436
......@@ -407,14 +407,14 @@ void build_op_func_list(const platform::Place& place,
auto exec_ctx =
ExecutionContext(*op_with_kernel, scope, *dev_ctx, runtime_context);
auto run_pten_kernel = false;
if (phi::KernelFactory::Instance().HasCompatiblePtenKernel(
auto run_phi_kernel = false;
if (phi::KernelFactory::Instance().HasCompatiblePhiKernel(
op_with_kernel->Type())) {
auto pt_kernel_key = op_with_kernel->ChoosePtenKernel(exec_ctx);
auto pt_kernel_name = op_with_kernel->PtenKernelSignature()->name;
auto pt_kernel_key = op_with_kernel->ChoosePhiKernel(exec_ctx);
auto pt_kernel_name = op_with_kernel->PhiKernelSignature()->name;
if (op_with_kernel->PtenKernel()->IsValid()) {
run_pten_kernel = true;
if (op_with_kernel->PhiKernel()->IsValid()) {
run_phi_kernel = true;
} else {
auto kernels_iter = all_op_kernels.find(op_with_kernel->Type());
if (kernels_iter == all_op_kernels.end() ||
......@@ -422,26 +422,26 @@ void build_op_func_list(const platform::Place& place,
kernels_iter->second.end()) {
auto pt_cpu_kernel_key = FallBackToCpu(
expected_kernel_key, pt_kernel_key, *op_with_kernel);
op_with_kernel->ResetPtenKernel(
op_with_kernel->ResetPhiKernel(
new phi::Kernel(phi::KernelFactory::Instance().SelectKernel(
pt_kernel_name, pt_cpu_kernel_key)));
if (op_with_kernel->PtenKernel()->IsValid()) {
if (op_with_kernel->PhiKernel()->IsValid()) {
VLOG(6) << "Static mode PrepareImpl - kernel name: "
<< pt_kernel_name
<< " | kernel key: " << pt_cpu_kernel_key
<< " | kernel: " << *(op_with_kernel->PtenKernel());
run_pten_kernel = true;
<< " | kernel: " << *(op_with_kernel->PhiKernel());
run_phi_kernel = true;
}
}
}
}
VLOG(3) << op_with_kernel->Type()
<< " : expected_kernel_key : " << expected_kernel_key;
if (run_pten_kernel) {
if (run_phi_kernel) {
phi::KernelContext pt_kernel_context;
op_with_kernel->BuildPtenKernelContext(runtime_context, dev_ctx,
op_with_kernel->BuildPhiKernelContext(runtime_context, dev_ctx,
&pt_kernel_context);
op_func_node.pt_kernel_ = op_with_kernel->PtenKernel();
op_func_node.pt_kernel_ = op_with_kernel->PhiKernel();
(*op_func_node.pt_kernel_)(&pt_kernel_context);
} else {
......
paddle/fluid/framework/new_executor/new_executor_defs.cc
浏览文件 @ 46894436
......@@ -688,9 +688,7 @@ OpKernelComputeFunc Instruction::KernelFunc() const {
return op_func_node_.kernel_func_;
}
phi::Kernel* Instruction::PtenKernel() const {
return op_func_node_.pt_kernel_;
}
phi::Kernel* Instruction::PhiKernel() const { return op_func_node_.pt_kernel_; }
OpFuncType Instruction::KernelType() const { return op_func_node_.type_; }
......
paddle/fluid/framework/new_executor/new_executor_defs.h
浏览文件 @ 46894436
......@@ -300,7 +300,7 @@ struct OpFuncNode {
OpKernelComputeFunc kernel_func_;
platform::DeviceContext* dev_ctx_; // not owned
// fit for pten kernel
// fit for phi kernel
phi::Kernel* pt_kernel_{nullptr}; // not owned
OpFuncType type_;
......@@ -321,7 +321,7 @@ class Instruction {
OpKernelComputeFunc KernelFunc() const;
phi::Kernel* PtenKernel() const;
phi::Kernel* PhiKernel() const;
OpFuncType KernelType() const;
......
paddle/fluid/framework/new_executor/workqueue/thread_data_registry.h
浏览文件 @ 46894436
......@@ -44,7 +44,6 @@ class ThreadDataRegistry {
template <typename Alias = T,
typename = std::enable_if_t<std::is_copy_assignable<Alias>::value>>
void SetCurrentThreadData(const T& val) {
std::lock_guard<std::mutex> lock(lock_);
CurrentThreadData() = val;
}
......
paddle/fluid/framework/new_executor/workqueue/workqueue.cc
浏览文件 @ 46894436
......@@ -8,6 +8,7 @@
#include "paddle/fluid/framework/new_executor/workqueue/nonblocking_threadpool.h"
#include "paddle/fluid/framework/new_executor/workqueue/workqueue_utils.h"
#include "paddle/fluid/platform/enforce.h"
#include "paddle/fluid/platform/profiler/event_tracing.h"
namespace paddle {
namespace framework {
......@@ -61,6 +62,8 @@ class WorkQueueImpl : public WorkQueue {
}
void AddTask(std::function<void()> fn) override {
platform::RecordEvent("WorkQueue::AddTask",
platform::TracerEventType::UserDefined, 10 /*level*/);
if (tracker_ != nullptr) {
fn = [
task = std::move(fn), raii = CounterGuard<TaskTracker>(tracker_)
......@@ -156,6 +159,8 @@ WorkQueueGroupImpl::~WorkQueueGroupImpl() {
}
void WorkQueueGroupImpl::AddTask(size_t queue_idx, std::function<void()> fn) {
platform::RecordEvent("WorkQueue::AddTask",
platform::TracerEventType::UserDefined, 10 /*level*/);
assert(queue_idx < queues_.size());
if (queues_options_.at(queue_idx).track_task) {
fn = [
......
paddle/fluid/framework/operator.cc
浏览文件 @ 46894436
......@@ -24,7 +24,7 @@ limitations under the License. */
#include "paddle/fluid/framework/data_type_transform.h"
#include "paddle/fluid/framework/details/nan_inf_utils.h"
#include "paddle/fluid/framework/op_call_stack.h"
#include "paddle/fluid/framework/pten_utils.h"
#include "paddle/fluid/framework/phi_utils.h"
#include "paddle/fluid/framework/shape_inference.h"
#include "paddle/fluid/framework/transfer_scope_cache.h"
#include "paddle/fluid/framework/unused_var_check.h"
......@@ -263,11 +263,11 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
// in order to record different op type cost time
// and different op name cost time,we set two event.
platform::RecordEvent op_type_record_event(
Type().c_str(), platform::TracerEventType::Operator, 1);
auto op_name = platform::OpName(outputs_, Type());
platform::RecordEvent op_name_record_event(
op_name, platform::TracerEventType::Operator, 1,
platform::EventRole::kUniqueOp);
Type(), platform::TracerEventType::Operator, 1);
// auto op_name = platform::OpName(outputs_, Type());
// platform::RecordEvent op_name_record_event(
// op_name, platform::TracerEventType::Operator, 1,
// platform::EventRole::kUniqueOp);
RunImpl(scope, place);
}
......@@ -616,9 +616,9 @@ bool OpSupportGPU(const std::string& op_type) {
// check in new Function kernel first
auto& kernel_factory = phi::KernelFactory::Instance();
auto kernel_key_map =
kernel_factory.SelectKernelMap(phi::TransToPtenKernelName(op_type));
kernel_factory.SelectKernelMap(phi::TransToPhiKernelName(op_type));
for (auto& kernel : kernel_key_map) {
if (platform::is_gpu_place(phi::TransToPtenPlace(kernel.first.backend()))) {
if (platform::is_gpu_place(phi::TransToPhiPlace(kernel.first.backend()))) {
return true;
}
}
......@@ -1186,10 +1186,10 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
// phase
phi::KernelKey pt_kernel_key;
std::string pt_kernel_name;
if (phi::KernelFactory::Instance().HasCompatiblePtenKernel(type_)) {
if (phi::KernelFactory::Instance().HasCompatiblePhiKernel(type_)) {
if (pt_kernel_signature_ == nullptr || pt_kernel_ == nullptr) {
pt_kernel_signature_.reset(
new KernelSignature(std::move(GetExpectedPtenKernelArgs(exe_ctx))));
new KernelSignature(std::move(GetExpectedPhiKernelArgs(exe_ctx))));
VLOG(6) << *pt_kernel_signature_.get();
kernel_type_.reset(
......@@ -1197,17 +1197,17 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
dev_ctx = pool.Get(kernel_type_->place_);
pt_kernel_name = pt_kernel_signature_->name;
pt_kernel_key = TransOpKernelTypeToPtenKernelKey(*kernel_type_.get());
pt_kernel_key = TransOpKernelTypeToPhiKernelKey(*kernel_type_.get());
pt_kernel_.reset(
new phi::Kernel(phi::KernelFactory::Instance().SelectKernel(
pt_kernel_name, pt_kernel_key)));
if (pt_kernel_->IsValid()) {
VLOG(6) << "Static mode ChoosePtenKernel - kernel name: "
VLOG(6) << "Static mode ChoosePhiKernel - kernel name: "
<< pt_kernel_name << " | kernel key: " << pt_kernel_key
<< " | kernel: " << *pt_kernel_;
} else {
VLOG(6) << "Static mode ChoosePtenKernel - kernel `" << pt_kernel_name
VLOG(6) << "Static mode ChoosePhiKernel - kernel `" << pt_kernel_name
<< "` not found.";
}
}
......@@ -1222,7 +1222,7 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
&& !is_xpu_unsupport
#endif
) {
run_pten_kernel_ = true;
run_phi_kernel_ = true;
} else {
auto& all_op_kernels = AllOpKernels();
auto kernels_iter = all_op_kernels.find(type_);
......@@ -1244,12 +1244,12 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
VLOG(6) << "Static mode PrepareImpl - kernel name: " << pt_kernel_name
<< " | kernel key: " << pt_cpu_kernel_key
<< " | kernel: " << *pt_kernel_;
run_pten_kernel_ = true;
run_phi_kernel_ = true;
}
}
}
}
if (!run_pten_kernel_) {
if (!run_phi_kernel_) {
if (kernel_type_.get() == nullptr || kernel_func_.get() == nullptr) {
ChooseKernel(exe_ctx);
dev_ctx = pool.Get(kernel_type_->place_);
......@@ -1290,13 +1290,13 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
platform::RecordEvent record_event("compute",
platform::TracerEventType::OperatorInner,
1, platform::EventRole::kInnerOp);
if (run_pten_kernel_) {
if (run_phi_kernel_) {
phi::KernelContext pt_kernel_context;
// Do data transform before building KernelContext
// TODO(zhiqiu): support TransferInplaceVarsBack
PreparePtenData(exec_scope, *pt_kernel_, *pt_kernel_signature_,
PreparePhiData(exec_scope, *pt_kernel_, *pt_kernel_signature_,
runtime_ctx);
BuildPtenKernelContext(*runtime_ctx, dev_ctx, &pt_kernel_context);
BuildPhiKernelContext(*runtime_ctx, dev_ctx, &pt_kernel_context);
(*pt_kernel_)(&pt_kernel_context);
} else {
(*kernel_func_)(
......@@ -1388,26 +1388,26 @@ OpKernelType OperatorWithKernel::InnerGetExpectedKernelType(
return expected_kernel_key;
}
phi::KernelKey OperatorWithKernel::ChoosePtenKernel(
phi::KernelKey OperatorWithKernel::ChoosePhiKernel(
const ExecutionContext& ctx) const {
pt_kernel_signature_.reset(
new KernelSignature(std::move(GetExpectedPtenKernelArgs(ctx))));
new KernelSignature(std::move(GetExpectedPhiKernelArgs(ctx))));
VLOG(6) << *pt_kernel_signature_.get();
kernel_type_.reset(
new OpKernelType(std::move(InnerGetExpectedKernelType(ctx))));
auto pt_kernel_name = pt_kernel_signature_->name;
auto pt_kernel_key = TransOpKernelTypeToPtenKernelKey(*kernel_type_.get());
auto pt_kernel_key = TransOpKernelTypeToPhiKernelKey(*kernel_type_.get());
pt_kernel_.reset(new phi::Kernel(phi::KernelFactory::Instance().SelectKernel(
pt_kernel_name, pt_kernel_key)));
if (pt_kernel_->IsValid()) {
VLOG(6) << "Static mode ChoosePtenKernel - kernel name: " << pt_kernel_name
VLOG(6) << "Static mode ChoosePhiKernel - kernel name: " << pt_kernel_name
<< " | kernel key: " << pt_kernel_key
<< " | kernel: " << *pt_kernel_;
} else {
VLOG(6) << "Static mode ChoosePtenKernel - kernel `" << pt_kernel_name
VLOG(6) << "Static mode ChoosePhiKernel - kernel `" << pt_kernel_name
<< "` not found.";
}
return pt_kernel_key;
......@@ -1918,7 +1918,7 @@ OpKernelType OperatorWithKernel::GetKernelTypeForVar(
tensor.layout());
}
KernelSignature OperatorWithKernel::GetExpectedPtenKernelArgs(
KernelSignature OperatorWithKernel::GetExpectedPhiKernelArgs(
const ExecutionContext& ctx) const {
InitDefaultKernelSignatureMap();
ExecutionArgumentMappingContext arg_mapping_ctx(ctx);
......@@ -1926,7 +1926,7 @@ KernelSignature OperatorWithKernel::GetExpectedPtenKernelArgs(
arg_mapping_ctx);
}
Scope* OperatorWithKernel::PreparePtenData(
Scope* OperatorWithKernel::PreparePhiData(
const Scope& scope, const phi::Kernel& pt_kernel,
const KernelSignature& pt_kernel_signature, RuntimeContext* ctx) const {
auto& input_names = std::get<0>(pt_kernel_signature.args);
......@@ -1981,12 +1981,12 @@ Scope* OperatorWithKernel::PreparePtenData(
if (in_def.backend == phi::Backend::ALL_BACKEND) {
continue;
}
auto expected_place = phi::TransToPtenPlace(in_def.backend);
auto expected_place = phi::TransToPhiPlace(in_def.backend);
if (platform::is_same_place(tensor_in->place(), expected_place)) {
continue;
}
VLOG(3) << "PTen Transform Variable " << input_names[i] << " from "
VLOG(3) << "phi Transform Variable " << input_names[i] << " from "
<< tensor_in->place() << " to " << expected_place;
if (!new_scope) {
......@@ -2007,7 +2007,7 @@ Scope* OperatorWithKernel::PreparePtenData(
return new_scope;
}
void OperatorWithKernel::BuildPtenKernelContext(
void OperatorWithKernel::BuildPhiKernelContext(
const RuntimeContext& ctx, platform::DeviceContext* dev_ctx,
phi::KernelContext* pt_kernel_context) const {
pt_kernel_context->SetDeviceContext(dev_ctx);
......@@ -2111,7 +2111,7 @@ void OperatorWithKernel::BuildPtenKernelContext(
experimental::ResetTensorDtypeAndLayoutByArgDef(tensor_out,
output_defs.at(i));
SetAllocationForOutputTenosr(
tensor_out, phi::TransToPtenPlace(output_defs.at(i).backend));
tensor_out, phi::TransToPhiPlace(output_defs.at(i).backend));
pt_kernel_context->EmplaceBackOutputWithoutSetRange(tensor_out);
}
......@@ -2145,10 +2145,10 @@ void OperatorWithKernel::BuildPtenKernelContext(
auto& ins_vector = ctx.inputs.at(attr_names[i]);
if (ins_vector.size() == 1) { // ShapeTensor
pt_kernel_context->EmplaceBackAttr(std::move(
experimental::MakePtenScalarArrayFromVar(*ins_vector.front())));
experimental::MakePhiScalarArrayFromVar(*ins_vector.front())));
} else { // ShapeTensorList
pt_kernel_context->EmplaceBackAttr(std::move(
experimental::MakePtenScalarArrayFromVarList(ins_vector)));
experimental::MakePhiScalarArrayFromVarList(ins_vector)));
}
}
} else if (attr_defs[i].type_index ==
......@@ -2178,8 +2178,8 @@ void OperatorWithKernel::BuildPtenKernelContext(
}
} else {
auto& ins_vector = ctx.inputs.at(attr_names[i]);
pt_kernel_context->EmplaceBackAttr(std::move(
experimental::MakePtenScalarFromVar(*ins_vector.front())));
pt_kernel_context->EmplaceBackAttr(
std::move(experimental::MakePhiScalarFromVar(*ins_vector.front())));
}
} else {
......@@ -2198,7 +2198,7 @@ void OperatorWithKernel::BuildPtenKernelContext(
pt_kernel_context->EmplaceBackAttr(BOOST_GET_CONST(std::string, attr));
} else if (attr_defs[i].type_index ==
std::type_index(typeid(phi::DataType))) {
auto data_type = paddle::framework::TransToPtenDataType(
auto data_type = paddle::framework::TransToPhiDataType(
static_cast<framework::proto::VarType::Type>(
BOOST_GET_CONST(int, attr)));
pt_kernel_context->EmplaceBackAttr(data_type);
......@@ -2206,7 +2206,7 @@ void OperatorWithKernel::BuildPtenKernelContext(
std::type_index(typeid(std::vector<int64_t>))) {
if (std::type_index(attr.type()) ==
std::type_index(typeid(std::vector<int>))) {
// Emplace Back Attr according to the type of Pten_Kernel args.
// Emplace Back Attr according to the type of Phi_Kernel args.
const auto& vector_int_attr = BOOST_GET_CONST(std::vector<int>, attr);
const std::vector<int64_t> vector_int64_attr(vector_int_attr.begin(),
vector_int_attr.end());
......
paddle/fluid/framework/operator.h
浏览文件 @ 46894436
......@@ -30,7 +30,7 @@ limitations under the License. */
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/op_info.h"
#include "paddle/fluid/framework/op_kernel_type.h"
#include "paddle/fluid/framework/pten_utils.h"
#include "paddle/fluid/framework/phi_utils.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/selected_rows_utils.h"
#include "paddle/fluid/framework/tensor.h"
......@@ -423,7 +423,7 @@ class ExecutionContext {
"size(%d).",
allocation_ptr->size(), phi::product(dim) * sizeof(T)));
paddle::framework::Tensor temp_tensor(framework::TransToPtenDataType(
paddle::framework::Tensor temp_tensor(framework::TransToPhiDataType(
framework::ToDataType(std::type_index(typeid(T)))));
temp_tensor.Resize(dim);
temp_tensor.ResetHolder(std::move(shared_allocation));
......@@ -538,14 +538,14 @@ class OperatorWithKernel : public OperatorBase {
}
bool SupportGPU() const override {
auto pten_kernels = phi::KernelFactory::Instance().SelectKernelMap(
phi::TransToPtenKernelName(type_));
auto has_pten_kernel =
std::any_of(pten_kernels.begin(), pten_kernels.end(),
auto phi_kernels = phi::KernelFactory::Instance().SelectKernelMap(
phi::TransToPhiKernelName(type_));
auto has_phi_kernel =
std::any_of(phi_kernels.begin(), phi_kernels.end(),
[](phi::KernelKeyMap::const_reference kern_pair) {
return kern_pair.first.backend() == phi::Backend::GPU;
});
if (has_pten_kernel) {
if (has_phi_kernel) {
return true;
} else {
auto& op_kernels = OperatorWithKernel::AllOpKernels().at(type_);
......@@ -558,7 +558,7 @@ class OperatorWithKernel : public OperatorBase {
}
bool SupportNPU() const override {
// TODO(zhiqiu): support pten if needed?
// TODO(zhiqiu): support phi if needed?
auto& op_kernels = OperatorWithKernel::AllOpKernels().at(type_);
return std::any_of(op_kernels.begin(), op_kernels.end(),
[](OpKernelMap::const_reference kern_pair) {
......@@ -566,7 +566,7 @@ class OperatorWithKernel : public OperatorBase {
});
}
bool SupportMLU() const override {
// TODO(zhiqiu): support pten if needed?
// TODO(zhiqiu): support phi if needed?
auto& op_kernels = OperatorWithKernel::AllOpKernels().at(type_);
return std::any_of(op_kernels.begin(), op_kernels.end(),
[](OpKernelMap::const_reference kern_pair) {
......@@ -603,39 +603,39 @@ class OperatorWithKernel : public OperatorBase {
return kernel_type_->place_;
}
/* member functions for adapting to pten lib */
/* member functions for adapting to phi lib */
/** In the Tensor calculation library, the new Kernel adopts a clearer and
* more streamlined design. The arguments of the Kernel and the input and
* output arguments registered in the original OpMaker do not match in some
* cases, so we use map to record the arguments required by the kernel.
* When selecting Kernel during Op execution, select the arguments of the
* original Op according to the GetExpectedPtenKernelArgs returned arguments.
* original Op according to the GetExpectedPhiKernelArgs returned arguments.
*/
phi::KernelSignature GetExpectedPtenKernelArgs(
phi::KernelSignature GetExpectedPhiKernelArgs(
const ExecutionContext& ctx) const;
/* member functions for adapting to pten lib */
phi::KernelKey ChoosePtenKernel(const ExecutionContext& ctx) const;
/* member functions for adapting to phi lib */
phi::KernelKey ChoosePhiKernel(const ExecutionContext& ctx) const;
/**
* Transfer data place for pten kernel
* Transfer data place for phi kernel
* Is this really needed?
*/
Scope* PreparePtenData(const Scope& scope, const phi::Kernel& pt_kernel,
Scope* PreparePhiData(const Scope& scope, const phi::Kernel& pt_kernel,
const phi::KernelSignature& pt_kernel_signature,
RuntimeContext* ctx) const;
void BuildPtenKernelContext(const RuntimeContext& ctx,
void BuildPhiKernelContext(const RuntimeContext& ctx,
platform::DeviceContext* dev_ctx,
phi::KernelContext* pt_kernel_context) const;
phi::KernelSignature* PtenKernelSignature() const {
phi::KernelSignature* PhiKernelSignature() const {
return pt_kernel_signature_.get();
}
phi::Kernel* PtenKernel() const { return pt_kernel_.get(); }
phi::Kernel* PhiKernel() const { return pt_kernel_.get(); }
void ResetPtenKernel(phi::Kernel* kernel) const {
void ResetPhiKernel(phi::Kernel* kernel) const {
return pt_kernel_.reset(kernel);
}
......@@ -692,9 +692,9 @@ class OperatorWithKernel : public OperatorBase {
mutable std::mutex cache_update_mutex_;
mutable bool enable_cache_transfer_scope_ = false;
// NOTE(chenweihang): Similar op members are used to adapt to
// new pten kernel, if there is a better design in the future,
// new phi kernel, if there is a better design in the future,
// we may polish the implementation here
mutable bool run_pten_kernel_ = false;
mutable bool run_phi_kernel_ = false;
mutable bool run_kp_kernel = false;
mutable std::unique_ptr<phi::KernelSignature> pt_kernel_signature_;
mutable std::unique_ptr<phi::Kernel> pt_kernel_;
......
paddle/fluid/framework/paddle2cinn/build_cinn_pass.cc
浏览文件 @ 46894436
......@@ -44,11 +44,6 @@ DECLARE_string(deny_cinn_ops);
namespace paddle {
namespace framework {
namespace ir {
class MemOptVarInfo;
} // namespace ir
namespace paddle2cinn {
using framework::ir::Graph;
......@@ -398,9 +393,7 @@ std::unique_ptr<Graph> CreateNewSubGraph(const GraphNodeSet& cluster,
kNoNeedBufferFeeds, no_need_buffer_feeds.release());
// initialize empty map for kMemOptVarInfoFromMainGraph attribute,
// it will be filled on the share_mem_opt_info_to_subgraph pass
subgraph->GetOrInit<std::unordered_map<
std::string, std::shared_ptr<framework::ir::MemOptVarInfo>>>(
kMemOptVarInfoFromMainGraph);
subgraph->GetOrInit<Name2VarInfoMap>(kMemOptVarInfoFromMainGraph);
return subgraph;
}
......
paddle/fluid/framework/paddle2cinn/build_cinn_pass.h
浏览文件 @ 46894436
......@@ -18,6 +18,10 @@ limitations under the License. */
namespace paddle {
namespace framework {
namespace ir {
class MemOptVarInfo;
} // namespace ir
namespace paddle2cinn {
constexpr char kCinnLaunchOp[] = "cinn_launch";
......@@ -27,6 +31,9 @@ constexpr char kInternalVars[] = "InternalVars";
constexpr char kOutputVars[] = "OutputVars";
constexpr char kMemOptVarInfoFromMainGraph[] =
"mem_opt_var_info_from_main_graph";
using Name2VarInfoMap =
std::unordered_map<std::string,
std::shared_ptr<framework::ir::MemOptVarInfo>>;
// A pass named BuildCinnPass, the function of this pass is:
//
......
paddle/fluid/framework/paddle2cinn/build_cinn_pass_test.cc
浏览文件 @ 46894436
......@@ -255,7 +255,9 @@ TEST(BuildCinnPassTest, AllOpSupportCinn) {
ASSERT_EQ(
std::unordered_set<Node*>(cinn_op->inputs.begin(), cinn_op->inputs.end()),
std::unordered_set<Node*>({v0, v1, v2, v4}));
ASSERT_EQ(cinn_op->outputs, std::vector<Node*>({v6, v7}));
ASSERT_EQ(std::unordered_set<Node*>(cinn_op->outputs.begin(),
cinn_op->outputs.end()),
std::unordered_set<Node*>({v6, v7}));
ASSERT_EQ(v1->outputs, std::vector<Node*>({cinn_op}));
ASSERT_EQ(v6->inputs, std::vector<Node*>({cinn_op}));
......
paddle/fluid/framework/paddle2cinn/cinn_compiler.cc
浏览文件 @ 46894436
......@@ -248,10 +248,10 @@ std::unique_ptr<CinnCompiledObject> CinnCompiler::CompileGraph(
*compiled_obj = {std::move(graph_compiler),
std::move(compiled_res.runtime_program), scope,
symbol.var_model_to_program_map()};
compiled_obj->launch_context =
std::make_unique<operators::details::CinnLaunchContext>(
compiled_obj->paddle2cinn_varmap, compiled_obj->scope);
compiled_obj->cached_index = compiled_num;
compiled_obj->launch_context =
std::make_unique<operators::details::CinnLaunchContext>(graph,
*compiled_obj);
return compiled_obj;
}
......
paddle/fluid/framework/paddle2cinn/cinn_graph_symbolization_test.cc
浏览文件 @ 46894436
......@@ -209,7 +209,7 @@ class CinnGraphSymbolizationTest : public ::testing::Test {
tensor.Resize(dims);
tensor.mutable_data(
platform::CPUPlace(),
framework::TransToPtenDataType(framework::proto::VarType::FP32));
framework::TransToPhiDataType(framework::proto::VarType::FP32));
return tensor;
};
#define FillFeedList(Name) feed_targets[#Name] = create_tensor();
......
paddle/fluid/framework/pten_utils.cc → paddle/fluid/framework/phi_utils.cc
浏览文件 @ 46894436
......@@ -15,7 +15,7 @@ limitations under the License. */
#include <sstream>
#include "paddle/fluid/framework/convert_utils.h"
#include "paddle/fluid/framework/pten_utils.h"
#include "paddle/fluid/framework/phi_utils.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/op_info.h"
......@@ -57,17 +57,16 @@ class KernelArgsNameMakerByOpProto : public KernelArgsNameMaker {
paddle::SmallVector<std::string> attr_names_;
};
OpKernelType TransPtenKernelKeyToOpKernelType(
const phi::KernelKey& kernel_key) {
OpKernelType TransPhiKernelKeyToOpKernelType(const phi::KernelKey& kernel_key) {
proto::VarType::Type data_type =
paddle::framework::TransToProtoVarType(kernel_key.dtype());
// no need to set current device id here
platform::Place place = phi::TransToPtenPlace(kernel_key.backend(), false);
platform::Place place = phi::TransToPhiPlace(kernel_key.backend(), false);
DataLayout data_layout = kernel_key.layout();
LibraryType library_type = LibraryType::kPlain;
if (kernel_key.backend() == phi::Backend::MKLDNN) {
library_type = LibraryType::kMKLDNN;
} else if (kernel_key.backend() == phi::Backend::CUDNN) {
} else if (kernel_key.backend() == phi::Backend::GPUDNN) {
library_type = LibraryType::kCUDNN;
} else {
// do nothing
......@@ -76,19 +75,19 @@ OpKernelType TransPtenKernelKeyToOpKernelType(
return OpKernelType(data_type, place, data_layout, library_type);
}
phi::KernelKey TransOpKernelTypeToPtenKernelKey(
phi::KernelKey TransOpKernelTypeToPhiKernelKey(
const OpKernelType& kernel_type) {
phi::Backend backend = phi::TransToPtenBackend(kernel_type.place_);
phi::Backend backend = phi::TransToPhiBackend(kernel_type.place_);
if (kernel_type.library_type_ == LibraryType::kMKLDNN) {
backend = phi::Backend::MKLDNN;
} else if (kernel_type.library_type_ == LibraryType::kCUDNN) {
backend = phi::Backend::CUDNN;
backend = phi::Backend::GPUDNN;
} else {
// do
}
paddle::experimental::DataLayout layout = kernel_type.data_layout_;
paddle::experimental::DataType dtype =
paddle::framework::TransToPtenDataType(kernel_type.data_type_);
paddle::framework::TransToPhiDataType(kernel_type.data_type_);
return phi::KernelKey(backend, layout, dtype);
}
......@@ -98,8 +97,8 @@ phi::KernelKey FallBackToCpu(const OpKernelType& expected_kernel_key,
#ifdef PADDLE_WITH_XPU
if (platform::is_xpu_place(expected_kernel_key.place_) ||
paddle::platform::is_in_xpu_black_list(op.Type())) {
VLOG(3) << "pten missing XPU kernel: " << op.Type()
<< ", expected_kernel_key:" << expected_kernel_key
VLOG(3) << "phi missing XPU kernel: " << op.Type()
<< "phipected_kernel_key:" << expected_kernel_key
<< ", fallbacking to CPU one!";
return phi::KernelKey(phi::Backend::CPU, kernel_key.layout(),
kernel_key.dtype());
......@@ -107,8 +106,8 @@ phi::KernelKey FallBackToCpu(const OpKernelType& expected_kernel_key,
#endif
#ifdef PADDLE_WITH_ASCEND_CL
if (platform::is_npu_place(expected_kernel_key.place_)) {
VLOG(3) << "pten missing NPU kernel: " << op.Type()
<< ", expected_kernel_key:" << expected_kernel_key
VLOG(3) << "phi missing NPU kernel: " << op.Type()
<< "phipected_kernel_key:" << expected_kernel_key
<< ", fallbacking to CPU one!";
return phi::KernelKey(phi::Backend::CPU, kernel_key.layout(),
kernel_key.dtype());
......@@ -116,8 +115,8 @@ phi::KernelKey FallBackToCpu(const OpKernelType& expected_kernel_key,
#endif
#ifdef PADDLE_WITH_MLU
if (platform::is_mlu_place(expected_kernel_key.place_)) {
VLOG(3) << "pten missing MLU kernel: " << op.Type()
<< ", expected_kernel_key:" << expected_kernel_key
VLOG(3) << "phi missing MLU kernel: " << op.Type()
<< "phipected_kernel_key:" << expected_kernel_key
<< ", fallbacking to CPU one!";
return phi::KernelKey(phi::Backend::CPU, kernel_key.layout(),
kernel_key.dtype());
......@@ -132,17 +131,17 @@ KernelArgsNameMakerByOpProto::GetInputArgsNames() {
auto& in = op_proto_->inputs()[i];
auto& in_name = in.name();
if ((in.has_extra() && in.extra()) || (in.has_quant() && in.quant())) {
VLOG(6) << "Parse PtenKernel input: skip extra & quant input - "
VLOG(6) << "Parse PhiKernel input: skip extra & quant input - "
<< in_name;
continue;
}
// If contains dispensable input, we should override the
// OpArgumentMapping method self in phi/ops/compat dir
if (in.has_dispensable() && in.dispensable()) {
VLOG(6) << "Parse PtenKernel input: skip dispensable input - " << in_name;
VLOG(6) << "Parse PhiKernel input: skip dispensable input - " << in_name;
continue;
}
VLOG(6) << "Parse PtenKernel input: " << in_name;
VLOG(6) << "Parse PhiKernel input: " << in_name;
input_names_.emplace_back(in_name);
}
return input_names_;
......@@ -154,11 +153,11 @@ KernelArgsNameMakerByOpProto::GetOutputArgsNames() {
auto& out = op_proto_->outputs()[i];
auto& out_name = out.name();
if ((out.has_extra() && out.extra()) || (out.has_quant() && out.quant())) {
VLOG(6) << "Parse PtenKernel output: skip extra & quant output - "
VLOG(6) << "Parse PhiKernel output: skip extra & quant output - "
<< out_name;
continue;
}
VLOG(6) << "Parse PtenKernel output: " << out_name;
VLOG(6) << "Parse PhiKernel output: " << out_name;
output_names_.emplace_back(out_name);
}
return output_names_;
......@@ -173,17 +172,17 @@ KernelArgsNameMakerByOpProto::GetAttrsArgsNames() {
attr_name == "op_role" || attr_name == "op_role_var" ||
attr_name == "op_namescope" || attr_name == "op_callstack" ||
attr_name == "op_device") {
VLOG(6) << "Parse PtenKernel attribute: skip needless attr - "
VLOG(6) << "Parse PhiKernel attribute: skip needless attr - "
<< attr_name;
continue;
}
if ((attr.has_extra() && attr.extra()) ||
(attr.has_quant() && attr.quant())) {
VLOG(6) << "Parse PtenKernel attribute: skip extra & quant attr - "
VLOG(6) << "Parse PhiKernel attribute: skip extra & quant attr - "
<< attr_name;
continue;
}
VLOG(6) << "Parse PtenKernel attribute: " << attr_name;
VLOG(6) << "Parse PhiKernel attribute: " << attr_name;
attr_names_.emplace_back(attr_name);
}
......@@ -191,7 +190,7 @@ KernelArgsNameMakerByOpProto::GetAttrsArgsNames() {
}
KernelSignature KernelArgsNameMakerByOpProto::GetKernelSignature() {
return KernelSignature(phi::TransToPtenKernelName(op_proto_->type()),
return KernelSignature(phi::TransToPhiKernelName(op_proto_->type()),
GetInputArgsNames(), GetAttrsArgsNames(),
GetOutputArgsNames());
}
......@@ -203,7 +202,7 @@ void InitDefaultKernelSignatureMap() {
for (const auto& pair : paddle::framework::OpInfoMap::Instance().map()) {
const auto& op_type = pair.first;
const auto* op_proto = pair.second.proto_;
if (phi::KernelFactory::Instance().HasCompatiblePtenKernel(op_type) &&
if (phi::KernelFactory::Instance().HasCompatiblePhiKernel(op_type) &&
op_proto) {
paddle::framework::KernelArgsNameMakerByOpProto maker(op_proto);
VLOG(10) << "Register kernel signature for " << op_type;
......
paddle/fluid/framework/pten_utils.h → paddle/fluid/framework/phi_utils.h
浏览文件 @ 46894436
......@@ -44,9 +44,8 @@ using KernelSignature = phi::KernelSignature;
/* Kernel Key translate */
OpKernelType TransPtenKernelKeyToOpKernelType(const phi::KernelKey& kernel_key);
phi::KernelKey TransOpKernelTypeToPtenKernelKey(
const OpKernelType& kernel_type);
OpKernelType TransPhiKernelKeyToOpKernelType(const phi::KernelKey& kernel_key);
phi::KernelKey TransOpKernelTypeToPhiKernelKey(const OpKernelType& kernel_type);
phi::KernelKey FallBackToCpu(const OpKernelType& expected_kernel_key,
const phi::KernelKey& kernel_key,
const framework::OperatorBase& op);
......@@ -68,25 +67,25 @@ void SetAllocationForOutputTenosr(phi::TensorBase* tensor,
// TODO(Wilber): support others device context.
template <typename T>
struct ConvertToPtenContext {
struct ConvertToPhiContext {
using TYPE = T;
};
template <>
struct ConvertToPtenContext<platform::CPUDeviceContext> {
struct ConvertToPhiContext<platform::CPUDeviceContext> {
using TYPE = phi::CPUContext;
};
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
template <>
struct ConvertToPtenContext<platform::CUDADeviceContext> {
struct ConvertToPhiContext<platform::CUDADeviceContext> {
using TYPE = phi::GPUContext;
};
#endif
#ifdef PADDLE_WITH_XPU
template <>
struct ConvertToPtenContext<platform::XPUDeviceContext> {
struct ConvertToPhiContext<platform::XPUDeviceContext> {
using TYPE = phi::XPUContext;
};
#endif
......
paddle/fluid/framework/pten_utils_test.cc → paddle/fluid/framework/phi_utils_test.cc
浏览文件 @ 46894436
......@@ -12,17 +12,17 @@ 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. */
#include "paddle/fluid/framework/pten_utils.h"
#include "paddle/fluid/framework/phi_utils.h"
#include "gtest/gtest.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/selected_rows_utils.h"
#include "paddle/fluid/framework/variable.h"
TEST(PtenUtils, TransPtenKernelKeyToOpKernelType) {
TEST(PhiUtils, TransPhiKernelKeyToOpKernelType) {
phi::KernelKey kernel_key(phi::Backend::CPU, phi::DataLayout::NCHW,
phi::DataType::FLOAT32);
auto op_kernel_type =
paddle::framework::TransPtenKernelKeyToOpKernelType(kernel_key);
paddle::framework::TransPhiKernelKeyToOpKernelType(kernel_key);
ASSERT_EQ(op_kernel_type.data_type_, paddle::framework::proto::VarType::FP32);
ASSERT_EQ(op_kernel_type.data_layout_, paddle::framework::DataLayout::kNCHW);
ASSERT_TRUE(paddle::platform::is_cpu_place(op_kernel_type.place_));
......@@ -33,7 +33,7 @@ TEST(PtenUtils, TransPtenKernelKeyToOpKernelType) {
phi::KernelKey kernel_key_mkldnn(phi::Backend::MKLDNN, phi::DataLayout::NCHW,
phi::DataType::FLOAT32);
op_kernel_type =
paddle::framework::TransPtenKernelKeyToOpKernelType(kernel_key_mkldnn);
paddle::framework::TransPhiKernelKeyToOpKernelType(kernel_key_mkldnn);
ASSERT_EQ(op_kernel_type.data_type_, paddle::framework::proto::VarType::FP32);
ASSERT_EQ(op_kernel_type.data_layout_, paddle::framework::DataLayout::kNCHW);
ASSERT_TRUE(paddle::platform::is_cpu_place(op_kernel_type.place_));
......@@ -42,10 +42,10 @@ TEST(PtenUtils, TransPtenKernelKeyToOpKernelType) {
#endif
#ifdef PADDLE_WITH_CUDA
phi::KernelKey kernel_key_cudnn(phi::Backend::CUDNN, phi::DataLayout::NCHW,
phi::KernelKey kernel_key_cudnn(phi::Backend::GPUDNN, phi::DataLayout::NCHW,
phi::DataType::FLOAT32);
op_kernel_type =
paddle::framework::TransPtenKernelKeyToOpKernelType(kernel_key_cudnn);
paddle::framework::TransPhiKernelKeyToOpKernelType(kernel_key_cudnn);
ASSERT_EQ(op_kernel_type.data_type_, paddle::framework::proto::VarType::FP32);
ASSERT_EQ(op_kernel_type.data_layout_, paddle::framework::DataLayout::kNCHW);
ASSERT_TRUE(paddle::platform::is_gpu_place(op_kernel_type.place_));
......@@ -53,3 +53,38 @@ TEST(PtenUtils, TransPtenKernelKeyToOpKernelType) {
paddle::framework::LibraryType::kCUDNN);
#endif
}
TEST(PhiUtils, TransOpKernelTypeToPhiKernelKey) {
paddle::framework::OpKernelType op_kernel_type(
paddle::framework::proto::VarType::FP32, paddle::platform::CPUPlace(),
paddle::framework::DataLayout::kNCHW);
auto kernel_key =
paddle::framework::TransOpKernelTypeToPhiKernelKey(op_kernel_type);
ASSERT_EQ(kernel_key.dtype(), phi::DataType::FLOAT32);
ASSERT_EQ(kernel_key.layout(), phi::DataLayout::NCHW);
ASSERT_EQ(kernel_key.backend(), phi::Backend::CPU);
#ifdef PADDLE_WITH_MKLDNN
paddle::framework::OpKernelType op_kernel_type_mkldnn(
paddle::framework::proto::VarType::FP32, paddle::platform::CPUPlace(),
paddle::framework::DataLayout::kMKLDNN,
paddle::framework::LibraryType::kMKLDNN);
auto kernel_key_mkldnn =
paddle::framework::TransOpKernelTypeToPhiKernelKey(op_kernel_type_mkldnn);
ASSERT_EQ(kernel_key_mkldnn.dtype(), phi::DataType::FLOAT32);
ASSERT_EQ(kernel_key_mkldnn.layout(), phi::DataLayout::MKLDNN);
ASSERT_EQ(kernel_key_mkldnn.backend(), phi::Backend::MKLDNN);
#endif
#ifdef PADDLE_WITH_CUDA
paddle::framework::OpKernelType op_kernel_type_cudnn(
paddle::framework::proto::VarType::FP32, paddle::platform::CPUPlace(),
paddle::framework::DataLayout::kNCHW,
paddle::framework::LibraryType::kCUDNN);
auto kernel_key_cudnn =
paddle::framework::TransOpKernelTypeToPhiKernelKey(op_kernel_type_cudnn);
ASSERT_EQ(kernel_key_cudnn.dtype(), phi::DataType::FLOAT32);
ASSERT_EQ(kernel_key_cudnn.layout(), phi::DataLayout::NCHW);
ASSERT_EQ(kernel_key_cudnn.backend(), phi::Backend::GPUDNN);
#endif
}
paddle/fluid/framework/tensor_util.cc
浏览文件 @ 46894436
......@@ -1457,7 +1457,7 @@ std::ostream& print_tensor<paddle::platform::complex<double>>(
std::ostream& operator<<(std::ostream& os, const LoD& lod) {
// NOTE(xiongkun):
// https://stackoverflow.com/questions/5195512/namespaces-and-operator-resolution
// if we don't redefine, the operator << of pten / framework LoD is not found.
// if we don't redefine, the operator << of phi / framework LoD is not found.
paddle::string::operator<<(os, lod);
return os;
}
......
paddle/fluid/imperative/CMakeLists.txt
浏览文件 @ 46894436
cc_library(imperative_flag SRCS flags.cc DEPS gflags flags)
cc_library(var_helper SRCS var_helper.cc DEPS tensor pten_api)
cc_library(var_helper SRCS var_helper.cc DEPS tensor phi_api)
IF(WITH_XPU)
cc_library(prepared_operator SRCS prepared_operator.cc DEPS xpu_op_list proto_desc operator device_context lod_tensor selected_rows_utils var_type_traits op_kernel_type data_transform nan_inf_utils pten_api pten pten_utils var_helper)
cc_library(prepared_operator SRCS prepared_operator.cc DEPS xpu_op_list proto_desc operator device_context lod_tensor selected_rows_utils var_type_traits op_kernel_type data_transform nan_inf_utils phi_api phi phi_utils var_helper)
ELSE()
cc_library(prepared_operator SRCS prepared_operator.cc DEPS proto_desc operator device_context lod_tensor selected_rows_utils var_type_traits op_kernel_type data_transform nan_inf_utils pten_api pten pten_utils var_helper)
cc_library(prepared_operator SRCS prepared_operator.cc DEPS proto_desc operator device_context lod_tensor selected_rows_utils var_type_traits op_kernel_type data_transform nan_inf_utils phi_api phi phi_utils var_helper)
ENDIF()
cc_library(layer SRCS layer.cc DEPS prepared_operator math_function imperative_flag variable_helper op_registry var_helper pten_api)
cc_library(layer SRCS layer.cc DEPS prepared_operator math_function imperative_flag variable_helper op_registry var_helper phi_api)
add_subdirectory(jit)
cc_library(amp SRCS amp_auto_cast.cc DEPS layer var_helper)
cc_library(tracer SRCS tracer.cc DEPS layer engine program_desc_tracer amp denormal garbage_collector var_helper)
......@@ -47,9 +47,9 @@ if(WITH_GLOO)
endif()
if(NOT WITH_ASCEND_CL)
cc_library(gradient_accumulator SRCS gradient_accumulator.cc DEPS blas operator lod_tensor selected_rows_utils selected_rows_functor var_type_traits layer math_function pten_tensor)
cc_library(gradient_accumulator SRCS gradient_accumulator.cc DEPS blas operator lod_tensor selected_rows_utils selected_rows_functor var_type_traits layer math_function phi_tensor)
else()
cc_library(gradient_accumulator SRCS gradient_accumulator.cc DEPS blas operator lod_tensor selected_rows_utils selected_rows_functor var_type_traits layer math_function npu_op_runner pten_tensor)
cc_library(gradient_accumulator SRCS gradient_accumulator.cc DEPS blas operator lod_tensor selected_rows_utils selected_rows_functor var_type_traits layer math_function npu_op_runner phi_tensor)
endif()
add_subdirectory(tests)
paddle/fluid/imperative/amp_auto_cast.cc
浏览文件 @ 46894436
......@@ -70,12 +70,12 @@ OpSupportedInfos(const std::string& place,
}
}
auto pten_kernels = phi::KernelFactory::Instance().kernels();
for (auto& kernel_pair : pten_kernels) {
auto phi_kernels = phi::KernelFactory::Instance().kernels();
for (auto& kernel_pair : phi_kernels) {
auto op_type = phi::TransToFluidOpName(kernel_pair.first);
for (auto& info_pair : kernel_pair.second) {
framework::OpKernelType kernel_type =
framework::TransPtenKernelKeyToOpKernelType(info_pair.first);
framework::TransPhiKernelKeyToOpKernelType(info_pair.first);
if (is_target_place[query_place](kernel_type.place_) &&
kernel_type.data_type_ == dtype && all_ops.count(op_type)) {
VLOG(4) << op_type << " " << supported_ops.size();
......@@ -273,8 +273,9 @@ static inline std::shared_ptr<VarType> CastToBF16(
template <typename VarType>
static inline framework::proto::VarType::Type GetPromoteType(
const std::string& op_type, const NameVarMap<VarType>& ins) {
auto dst_type = framework::proto::VarType::FP16;
const std::string& op_type, const NameVarMap<VarType>& ins,
const framework::proto::VarType::Type amp_dtype) {
auto dst_type = amp_dtype;
for (const auto& pair : ins) {
for (const auto& var : pair.second) {
if (GetDataType<VarType>(var) == framework::proto::VarType::FP32) {
......@@ -337,7 +338,8 @@ NameVarMap<VarType> AutoCastInputs(const std::string& op_type,
}
return new_ins;
} else {
auto dst_type = GetPromoteType<VarType>(op_type, ins);
auto dst_type =
GetPromoteType<VarType>(op_type, ins, framework::proto::VarType::FP16);
// NOTE(zhiqiu): if the op has op fp16 kernel, fall back to fp32.
if (dst_type == framework::proto::VarType::FP16 &&
......@@ -435,7 +437,7 @@ NameVarMap<VarType> AutoCastBF16Inputs(const std::string& op_type,
}
}
return new_ins;
} else {
} else if (AmpOperators::Instance().GetMutableBlockOps()->count(op_type)) {
for (auto& pair : new_ins) {
VLOG(5) << "Op(" << op_type << "): Cast " << pair.first << " from "
<< GetDtypeStr(*pair.second.cbegin()) << " to float";
......@@ -444,6 +446,26 @@ NameVarMap<VarType> AutoCastBF16Inputs(const std::string& op_type,
}
}
return new_ins;
} else {
auto dst_type =
GetPromoteType<VarType>(op_type, ins, framework::proto::VarType::BF16);
// NOTE(zhangbo): if the op has op fp16 kernel, fall back to fp32.
if (dst_type == framework::proto::VarType::BF16 &&
AmpOperators::Instance().GetMutableUnsupportedBf16Ops()->count(
op_type)) {
dst_type = framework::proto::VarType::FP32;
}
for (auto& pair : new_ins) {
VLOG(5) << "Op(" << op_type << "): Cast " << pair.first << " from "
<< GetDtypeStr(*pair.second.cbegin()) << " to "
<< framework::DataTypeToString(dst_type);
for (auto& var : pair.second) {
var = (dst_type == framework::proto::VarType::FP32
? CastToFP32<VarType>(var)
: CastToBF16<VarType>(var));
}
}
return new_ins;
}
return new_ins;
}
......
paddle/fluid/imperative/basic_engine.cc
浏览文件 @ 46894436
......@@ -154,7 +154,7 @@ void BasicEngine::CheckBackwardInputs(const OpBase& op) {
// Here, we use the type of the corresponding forward datatype.
tensor->mutable_data(
op.place(), framework::TransToPtenDataType(var->ForwardDataType()));
op.place(), framework::TransToPhiDataType(var->ForwardDataType()));
VLOG(6) << "Set ungenerated Grad: " << var->Name()
<< " as zero with dtype "
<< framework::DataTypeToString(var->ForwardDataType());
......
paddle/fluid/imperative/gradient_accumulator.cc
浏览文件 @ 46894436
......@@ -791,13 +791,13 @@ void EagerGradientAccumulator::SumGrad(std::shared_ptr<VariableWrapper> var,
<< var->Var().Get<framework::LoDTensor>().dims();
tensor->Resize(var->Var().Get<framework::LoDTensor>().dims());
tensor->mutable_data(place,
framework::TransToPtenDataType(var->DataType()));
framework::TransToPhiDataType(var->DataType()));
phi::funcs::set_constant(*dev_ctx, tensor, 0.0);
} else {
auto* tensor =
dst_var->MutableVar()->GetMutable<framework::LoDTensor>();
tensor->mutable_data(place,
framework::TransToPtenDataType(var->DataType()));
framework::TransToPhiDataType(var->DataType()));
phi::funcs::set_constant(*dev_ctx, tensor, 0.0);
}
}
......@@ -925,13 +925,13 @@ void SortedGradientAccumulator::SumGrad(std::shared_ptr<VariableWrapper> var,
<< var->Var().Get<framework::LoDTensor>().dims();
tensor->Resize(var->Var().Get<framework::LoDTensor>().dims());
tensor->mutable_data(place,
framework::TransToPtenDataType(var->DataType()));
framework::TransToPhiDataType(var->DataType()));
phi::funcs::set_constant(*dev_ctx, tensor, 0.0);
} else {
auto* tensor =
dst_var->MutableVar()->GetMutable<framework::LoDTensor>();
tensor->mutable_data(place,
framework::TransToPtenDataType(var->DataType()));
framework::TransToPhiDataType(var->DataType()));
phi::funcs::set_constant(*dev_ctx, tensor, 0.0);
}
}
......
paddle/fluid/imperative/partial_grad_engine.cc
浏览文件 @ 46894436
......@@ -314,10 +314,10 @@ static void FillConstantLike(const VariableWrapper &ref_var,
// default data_type for now.
if (ref_var.ForwardDataType() != -1) {
dst_tensor->mutable_data(
place, framework::TransToPtenDataType(ref_var.ForwardDataType()));
place, framework::TransToPhiDataType(ref_var.ForwardDataType()));
} else {
dst_tensor->mutable_data(
place, framework::TransToPtenDataType(ref_var.DataType()));
dst_tensor->mutable_data(place,
framework::TransToPhiDataType(ref_var.DataType()));
}
phi::funcs::set_constant(*dev_ctx, dst_tensor, value);
}
......
paddle/fluid/imperative/prepared_operator.cc
浏览文件 @ 46894436
......@@ -121,7 +121,7 @@ PreparedOp::PreparedOp(const framework::OperatorBase& op,
kernel_type_(kernel_type),
func_(nullptr),
dev_ctx_(dev_ctx),
run_pten_kernel_(true),
run_phi_kernel_(true),
pt_kernel_signature_(kernel_signature),
pt_kernel_(pt_kernel) {}
......@@ -151,7 +151,7 @@ PreparedOp PrepareImpl(const NameVarMap<VarType>& ins,
#endif
// NOTE(zhiqiu): for kernels on given device, for example NPU, the order to
// choose is:
// pten npu kernel > fluid npu kernel > pten cpu kernel > fluid cpu kernel
// phi npu kernel > fluid npu kernel > phi cpu kernel > fluid cpu kernel
// 1. get expected kernel key
auto dygraph_exe_ctx = DygraphExecutionContext<VarType>(
......@@ -168,12 +168,12 @@ PreparedOp PrepareImpl(const NameVarMap<VarType>& ins,
expected_kernel_key) ||
paddle::platform::is_in_xpu_black_list(op.Type());
#endif
if (phi::KernelFactory::Instance().HasCompatiblePtenKernel(op.Type())) {
pt_kernel_signature = op.GetExpectedPtenKernelArgs(dygraph_exe_ctx);
if (phi::KernelFactory::Instance().HasCompatiblePhiKernel(op.Type())) {
pt_kernel_signature = op.GetExpectedPhiKernelArgs(dygraph_exe_ctx);
VLOG(6) << pt_kernel_signature;
pt_kernel_name = pt_kernel_signature.name;
pt_kernel_key = TransOpKernelTypeToPtenKernelKey(expected_kernel_key);
pt_kernel_key = TransOpKernelTypeToPhiKernelKey(expected_kernel_key);
auto pt_kernel = phi::KernelFactory::Instance().SelectKernel(pt_kernel_name,
pt_kernel_key);
......@@ -195,7 +195,7 @@ PreparedOp PrepareImpl(const NameVarMap<VarType>& ins,
return PreparedOp(op, ctx, expected_kernel_key, pt_kernel_signature,
pt_kernel, dev_ctx);
} else {
VLOG(6) << "Dynamic mode ChoosePtenKernel - kernel `" << pt_kernel_name
VLOG(6) << "Dynamic mode ChoosePhiKernel - kernel `" << pt_kernel_name
<< "` not found.";
}
}
......@@ -211,7 +211,7 @@ PreparedOp PrepareImpl(const NameVarMap<VarType>& ins,
|| is_xpu_unsupport
#endif
) {
if (phi::KernelFactory::Instance().HasCompatiblePtenKernel(op.Type())) {
if (phi::KernelFactory::Instance().HasCompatiblePhiKernel(op.Type())) {
auto pt_cpu_kernel_key =
FallBackToCpu(expected_kernel_key, pt_kernel_key, op);
auto pt_cpu_kernel = phi::KernelFactory::Instance().SelectKernel(
......@@ -423,10 +423,10 @@ static void PreparedOpRunPtImpl(
platform::TracerEventType::OperatorInner,
1, platform::EventRole::kInnerOp);
PreparePtenData<VarType>(pt_kernel, pt_kernel_signature, ins);
PreparePhiData<VarType>(pt_kernel, pt_kernel_signature, ins);
phi::KernelContext pt_kernel_context;
BuildDygraphPtenKernelContext<VarType>(pt_kernel_signature, pt_kernel, ins,
BuildDygraphPhiKernelContext<VarType>(pt_kernel_signature, pt_kernel, ins,
outs, attrs, default_attrs, dev_ctx,
&pt_kernel_context);
......@@ -451,7 +451,7 @@ void PreparedOp::Run(const NameVarMap<VarBase>& ins,
const NameVarMap<VarBase>& outs,
const framework::AttributeMap& attrs,
const framework::AttributeMap& default_attrs) {
if (run_pten_kernel_) {
if (run_phi_kernel_) {
PreparedOpRunPtImpl<VarBase>(op_, kernel_type_, pt_kernel_signature_,
pt_kernel_, dev_ctx_, ins, outs, attrs,
default_attrs);
......@@ -465,7 +465,7 @@ void PreparedOp::Run(const NameVarMap<VariableWrapper>& ins,
const NameVarMap<VariableWrapper>& outs,
const framework::AttributeMap& attrs,
const framework::AttributeMap& default_attrs) {
if (run_pten_kernel_) {
if (run_phi_kernel_) {
PreparedOpRunPtImpl<VariableWrapper>(
op_, kernel_type_, pt_kernel_signature_, pt_kernel_, dev_ctx_, ins,
outs, attrs, default_attrs);
......@@ -479,7 +479,7 @@ void PreparedOp::Run(const NameVarMap<egr::EagerVariable>& ins,
const NameVarMap<egr::EagerVariable>& outs,
const framework::AttributeMap& attrs,
const framework::AttributeMap& default_attrs) {
if (run_pten_kernel_) {
if (run_phi_kernel_) {
PreparedOpRunPtImpl<egr::EagerVariable>(
op_, kernel_type_, pt_kernel_signature_, pt_kernel_, dev_ctx_, ins,
outs, attrs, default_attrs);
......
paddle/fluid/imperative/prepared_operator.h
浏览文件 @ 46894436
......@@ -22,7 +22,7 @@
#include "paddle/fluid/framework/data_transform.h"
#include "paddle/fluid/framework/op_kernel_type.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/pten_utils.h"
#include "paddle/fluid/framework/phi_utils.h"
#include "paddle/fluid/framework/type_defs.h"
#include "paddle/fluid/imperative/execution_context.h"
#include "paddle/fluid/imperative/layer.h"
......@@ -201,9 +201,9 @@ class PreparedOp {
framework::OperatorWithKernel::OpKernelFunc func_;
platform::DeviceContext* dev_ctx_;
// NOTE(chenweihang): Similar op members are used to adapt to
// new pten kernel, if there is a better design in the future,
// new phi kernel, if there is a better design in the future,
// we may polish the implementation here
bool run_pten_kernel_{false};
bool run_phi_kernel_{false};
bool run_kp_kernel_{false};
framework::KernelSignature pt_kernel_signature_;
phi::Kernel pt_kernel_;
......@@ -225,7 +225,7 @@ const inline framework::Attribute& GetAttr(
}
template <typename VarType>
void BuildDygraphPtenKernelContext(
void BuildDygraphPhiKernelContext(
const framework::KernelSignature& pt_kernel_signature,
const phi::Kernel& pt_kernel, const NameVarMap<VarType>& ins,
const NameVarMap<VarType>& outs, const framework::AttributeMap& attrs,
......@@ -327,7 +327,7 @@ void BuildDygraphPtenKernelContext(
experimental::ResetTensorDtypeAndLayoutByArgDef(tensor_out,
output_defs.at(i));
framework::SetAllocationForOutputTenosr(
tensor_out, phi::TransToPtenPlace(output_defs.at(i).backend));
tensor_out, phi::TransToPhiPlace(output_defs.at(i).backend));
kernel_ctx->EmplaceBackOutputWithoutSetRange(tensor_out);
}
......@@ -369,7 +369,7 @@ void BuildDygraphPtenKernelContext(
auto& ins_vector = ins.at(attr_names[i]);
if (ins_vector.size() == 1) { // ShapeTensor
kernel_ctx->EmplaceBackAttr(std::move(
experimental::MakePtenScalarArrayFromVar(ins_vector[0]->Var())));
experimental::MakePhiScalarArrayFromVar(ins_vector[0]->Var())));
} else { // ShapeTensorList
std::vector<framework::Variable*> variables;
variables.reserve(ins_vector.size());
......@@ -377,7 +377,7 @@ void BuildDygraphPtenKernelContext(
variables.push_back(var_base->MutableVar());
}
kernel_ctx->EmplaceBackAttr(std::move(
experimental::MakePtenScalarArrayFromVarList(variables)));
experimental::MakePhiScalarArrayFromVarList(variables)));
}
}
} else if (attr_defs[i].type_index ==
......@@ -409,7 +409,7 @@ void BuildDygraphPtenKernelContext(
} else { // scalar is in the input
auto& ins_vector = ins.at(attr_names[i]);
kernel_ctx->EmplaceBackAttr(std::move(
experimental::MakePtenScalarFromVar(ins_vector[0]->Var())));
experimental::MakePhiScalarFromVar(ins_vector[0]->Var())));
}
} else {
......@@ -428,7 +428,7 @@ void BuildDygraphPtenKernelContext(
kernel_ctx->EmplaceBackAttr(BOOST_GET_CONST(std::string, attr));
} else if (attr_defs[i].type_index ==
std::type_index(typeid(phi::DataType))) {
auto data_type = framework::TransToPtenDataType(
auto data_type = framework::TransToPhiDataType(
static_cast<framework::proto::VarType::Type>(
BOOST_GET_CONST(int, attr)));
kernel_ctx->EmplaceBackAttr(data_type);
......@@ -436,7 +436,7 @@ void BuildDygraphPtenKernelContext(
std::type_index(typeid(std::vector<int64_t>))) {
if (std::type_index(attr.type()) ==
std::type_index(typeid(std::vector<int>))) {
// Emplace Back Attr according to the type of Pten_Kernel args.
// Emplace Back Attr according to the type of Phi_Kernel args.
const auto& vector_int_attr = BOOST_GET_CONST(std::vector<int>, attr);
const std::vector<int64_t> vector_int64_attr(vector_int_attr.begin(),
vector_int_attr.end());
......@@ -456,7 +456,7 @@ void BuildDygraphPtenKernelContext(
}
template <typename VarType>
void PreparePtenData(const phi::Kernel& pt_kernel,
void PreparePhiData(const phi::Kernel& pt_kernel,
const framework::KernelSignature& pt_kernel_signature,
const NameVarMap<VarType>& ins) {
auto& input_names = std::get<0>(pt_kernel_signature.args);
......@@ -482,12 +482,12 @@ void PreparePtenData(const phi::Kernel& pt_kernel,
if (in_def.backend == phi::Backend::ALL_BACKEND) {
continue;
}
auto expected_place = phi::TransToPtenPlace(in_def.backend);
auto expected_place = phi::TransToPhiPlace(in_def.backend);
if (platform::is_same_place(tensor_in->place(), expected_place)) {
continue;
}
VLOG(3) << "Pten Transform Variable " << input_names[i] << " from "
VLOG(3) << "Phi Transform Variable " << input_names[i] << " from "
<< tensor_in->place() << " to " << expected_place;
framework::Tensor tmp_tensor;
......
paddle/fluid/imperative/reducer.cc
浏览文件 @ 46894436
......@@ -446,7 +446,7 @@ void Reducer::InitializeGroups(
InitializeDenseGroups(variable_indices_, &group);
auto tensor = group.dense_contents_.GetMutable<framework::LoDTensor>();
tensor->Resize(phi::make_ddim({group.all_length_}))
.mutable_data(place_, framework::TransToPtenDataType(group.dtype_));
.mutable_data(place_, framework::TransToPhiDataType(group.dtype_));
}
// map variables to this group by VariableLocator
......@@ -738,7 +738,7 @@ void Reducer::MarkVarReady(const size_t var_index, const bool is_used_var) {
if (!group_tensor.IsInitialized()) {
group_tensor.Resize({static_cast<int64_t>(length)});
group_tensor.mutable_data(place_,
framework::TransToPtenDataType(group.dtype_));
framework::TransToPhiDataType(group.dtype_));
}
#ifdef PADDLE_WITH_XPU_BKCL
......
paddle/fluid/imperative/tests/CMakeLists.txt
浏览文件 @ 46894436
......@@ -15,7 +15,7 @@ else()
endif(WIN32)
cc_test(test_gradient_accmulator SRCS test_gradient_accmulator.cc DEPS memcpy selected_rows_utils selected_rows_functor gradient_accumulator math_function pten_tensor pten_api pten_api_utils)
cc_test(test_gradient_accmulator SRCS test_gradient_accmulator.cc DEPS memcpy selected_rows_utils selected_rows_functor gradient_accumulator math_function phi_tensor phi_api phi_api_utils)
cc_test(test_layer SRCS test_layer.cc DEPS layer proto_desc operator op_registry variable_helper mul_op memcpy)
cc_test(test_prepare_op SRCS test_prepare_op.cc DEPS prepared_operator op_info split_op layer concat_and_split activation_op place)
cc_test(test_tracer SRCS test_tracer.cc DEPS tracer layer proto_desc operator op_registry variable_helper mul_op reduce_sum_op elementwise_add_op memcpy)
......
paddle/fluid/imperative/tests/test_group.cc
浏览文件 @ 46894436
......@@ -96,7 +96,7 @@ void GroupConcatSplit(Place place, size_t size) {
{ // concat
auto* tensor = group.dense_contents_.GetMutable<framework::LoDTensor>();
tensor->Resize(phi::make_ddim({group.all_length_}))
.mutable_data(place, framework::TransToPtenDataType(group.dtype_));
.mutable_data(place, framework::TransToPhiDataType(group.dtype_));
group.ConcatTensors(*dev_ctx);
group.DivNRanks(*dev_ctx, 1);
......
paddle/fluid/imperative/tracer.cc
浏览文件 @ 46894436
......@@ -175,7 +175,7 @@ void Tracer::TraceOp(const std::string& type, const NameVarMap<VarType>& ins,
paddle::framework::AttributeMap* passed_default_attrs_,
bool use_default_attr_map) {
platform::RecordEvent op_type_record_event(
type, platform::TracerEventType::Operator, 2);
type, platform::TracerEventType::Operator, 1);
platform::ScopedFlushDenormal flush;
VLOG(1) << "Trace Op: " << type;
if (FLAGS_use_mkldnn) {
......@@ -205,17 +205,19 @@ void Tracer::TraceOp(const std::string& type, const NameVarMap<VarType>& ins,
NameVarMap<VarType> new_ins = ins;
if (amp_level_ == AmpLevel::O1) {
VLOG(5) << "Auto mixed precision run operator: " << type;
if (amp_dtype_ == phi::DataType::FLOAT16) {
VLOG(5) << "Float16 Auto Mixed Precision O1 run operator: " << type;
new_ins = AutoCastInputs<VarType>(type, ins);
} else if (amp_dtype_ == phi::DataType::BFLOAT16) {
VLOG(5) << "BFloat16 Auto Mixed Precision O1 run operator: " << type;
new_ins = AutoCastBF16Inputs<VarType>(type, ins);
}
} else if (amp_level_ == AmpLevel::O2) {
VLOG(5) << "Pure fp16 run operator: " << type;
if (amp_dtype_ == phi::DataType::FLOAT16) {
VLOG(5) << "Float16 Auto Mixed Precision O2 run operator: " << type;
new_ins = CastPureFp16Inputs<VarType>(type, ins);
} else if (amp_dtype_ == phi::DataType::BFLOAT16) {
VLOG(5) << "BFloat16 Auto Mixed Precision O2 run operator: " << type;
new_ins = CastPureBf16Inputs<VarType>(type, ins);
}
}
......
paddle/fluid/inference/CMakeLists.txt
浏览文件 @ 46894436
......@@ -35,7 +35,7 @@ endif()
# fluid_modules exclude API-interface of inference/api and inference/capi_exp
get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES)
get_property(pten_modules GLOBAL PROPERTY PTEN_MODULES)
get_property(phi_modules GLOBAL PROPERTY PHI_MODULES)
set(utils_modules stringpiece pretty_log string_helper)
add_subdirectory(api)
......@@ -47,11 +47,11 @@ set(STATIC_INFERENCE_API paddle_inference_api analysis_predictor
analysis_config paddle_pass_builder activation_functions ${mkldnn_quantizer_cfg})
#TODO(wilber, T8T9): Do we still need to support windows gpu static library?
if(WIN32 AND WITH_GPU)
cc_library(paddle_inference DEPS ${fluid_modules} ${pten_modules} ${STATIC_INFERENCE_API} ${utils_modules})
cc_library(paddle_inference DEPS ${fluid_modules} ${phi_modules} ${STATIC_INFERENCE_API} ${utils_modules})
elseif(WITH_IPU)
cc_library(paddle_inference DEPS ${fluid_modules} ${pten_modules} ${STATIC_INFERENCE_API} ${utils_modules} paddle_ipu)
cc_library(paddle_inference DEPS ${fluid_modules} ${phi_modules} ${STATIC_INFERENCE_API} ${utils_modules} paddle_ipu)
else()
create_static_lib(paddle_inference ${fluid_modules} ${pten_modules} ${STATIC_INFERENCE_API} ${utils_modules})
create_static_lib(paddle_inference ${fluid_modules} ${phi_modules} ${STATIC_INFERENCE_API} ${utils_modules})
endif()
if(NOT APPLE)
......@@ -81,7 +81,7 @@ set(SHARED_INFERENCE_SRCS
${PADDLE_CUSTOM_OP_SRCS})
# shared inference library deps
set(SHARED_INFERENCE_DEPS ${fluid_modules} ${pten_modules} analysis_predictor)
set(SHARED_INFERENCE_DEPS ${fluid_modules} ${phi_modules} analysis_predictor)
if (WITH_CRYPTO)
set(SHARED_INFERENCE_DEPS ${SHARED_INFERENCE_DEPS} paddle_crypto)
......
paddle/fluid/inference/api/CMakeLists.txt
浏览文件 @ 46894436
......@@ -56,8 +56,10 @@ cc_test(test_paddle_inference_api SRCS api_tester.cc DEPS paddle_inference_api)
if(WITH_TESTING)
if (NOT APPLE AND NOT WIN32)
if (WITH_GPU)
inference_base_test(test_api_impl SRCS api_impl_tester.cc DEPS paddle_inference_shared
ARGS --word2vec_dirname=${WORD2VEC_MODEL_DIR} --book_dirname=${IMG_CLS_RESNET_INSTALL_DIR})
endif()
elseif(WIN32)
inference_base_test(test_api_impl SRCS api_impl_tester.cc DEPS ${inference_deps}
ARGS --word2vec_dirname=${WORD2VEC_MODEL_DIR} --book_dirname=${IMG_CLS_RESNET_INSTALL_DIR})
......
paddle/fluid/inference/io.cc
浏览文件 @ 46894436
......@@ -26,7 +26,7 @@ limitations under the License. */
#include "paddle/fluid/platform/enforce.h"
#include "paddle/fluid/pybind/pybind.h"
// pten
// phi
#include "paddle/phi/kernels/declarations.h"
DEFINE_string(devices, "", "The devices to be used which is joined by comma.");
......
paddle/fluid/inference/lite/tensor_utils.cc
浏览文件 @ 46894436
......@@ -198,7 +198,7 @@ void InitDstTensor(framework::LoDTensor* dst,
const paddle::lite_api::Tensor& src) {
dst->mutable_data(
inference::lite::utils::GetNativePlace(src.target()),
framework::TransToPtenDataType(GetNativePrecisionType(src.precision())));
framework::TransToPhiDataType(GetNativePrecisionType(src.precision())));
SetLoD(dst->mutable_lod(), src.lod());
}
......@@ -269,7 +269,7 @@ void TensorDataShare(framework::LoDTensor* dst, paddle::lite_api::Tensor* src) {
SetLoD(dst->mutable_lod(), src->lod());
dst->ResetHolderWithType(
holder,
framework::TransToPtenDataType(GetNativePrecisionType(src->precision())));
framework::TransToPhiDataType(GetNativePrecisionType(src->precision())));
}
} // namespace utils
......
paddle/fluid/inference/tensorrt/convert/softmax_op.cc
浏览文件 @ 46894436
......@@ -88,5 +88,5 @@ class SoftMaxOpConverter : public OpConverter {
} // namespace inference
} // namespace paddle
USE_OP(softmax);
USE_OP_ITSELF(softmax);
REGISTER_TRT_OP_CONVERTER(softmax, SoftMaxOpConverter);
paddle/fluid/inference/tensorrt/convert/test_softmax_op.cc
浏览文件 @ 46894436
......@@ -45,4 +45,4 @@ TEST(SoftMaxOpConverter, main) {
} // namespace inference
} // namespace paddle
USE_OP(softmax);
USE_OP_ITSELF(softmax);
paddle/fluid/inference/tensorrt/plugin/special_slice_plugin.cu
浏览文件 @ 46894436
......@@ -113,12 +113,12 @@ nvinfer1::DataType SpecialSlicePluginDynamic::getOutputDataType(
template <typename T>
__global__ void SpecialSliceKernel(const T* slice_input,
const int32_t* cu_seqlens, T* output) {
const int hidden = blockDim.x * gridDim.y;
const int batch = blockIdx.x;
const int local_idx = blockIdx.y * blockDim.y + threadIdx.x;
const int hidden = blockDim.x * gridDim.x;
const int hidden_id = blockIdx.x * blockDim.x + threadIdx.x;
const int batch_id = blockIdx.y;
output[batch * hidden + local_idx] =
slice_input[cu_seqlens[batch] * hidden + local_idx];
output[batch_id * hidden + hidden_id] =
slice_input[cu_seqlens[batch_id] * hidden + hidden_id];
}
int SpecialSlicePluginDynamic::enqueue(
......@@ -137,15 +137,16 @@ int SpecialSlicePluginDynamic::enqueue(
"hidden should be multiple of 128."));
constexpr int num_threads = 128;
const dim3 blocks(out_dims.d[0], hidden / num_threads);
const half* slice_input = static_cast<const half*>(inputs[0]);
const int32_t* cu_seqlens = static_cast<const int32_t*>(inputs[1]);
half* output = static_cast<half*>(outputs[0]);
SpecialSliceKernel<<<blocks, num_threads, 0, stream>>>(slice_input,
cu_seqlens, output);
const int32_t num_blocks_x = hidden / num_threads;
const int32_t num_blocks_y = out_dims.d[0]; // batchs
const dim3 num_blocks(num_blocks_x, num_blocks_y); // blocks
SpecialSliceKernel<<<num_blocks, num_threads, 0, stream>>>(
slice_input, cu_seqlens, output);
return cudaGetLastError() != cudaSuccess;
}
......
paddle/fluid/inference/tests/api/CMakeLists.txt
浏览文件 @ 46894436
......@@ -299,7 +299,9 @@ inference_analysis_api_test(test_analyzer_pyramid_dnn ${PYRAMID_DNN_INSTALL_DIR}
set(ERNIE_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/Ernie")
download_model_and_data(${ERNIE_INSTALL_DIR} "Ernie_model.tar.gz" aa59192dd41ed377f9f168e3a1309fa6 "Ernie_data.txt.tar.gz" 5396e63548edad7ca561e7e26a9476d1)
download_result(${ERNIE_INSTALL_DIR} "Ernie_result.txt.tar.gz" 73beea65abda2edb61c1662cd3180c62)
inference_analysis_api_test(test_analyzer_ernie ${ERNIE_INSTALL_DIR} analyzer_ernie_tester.cc)
if (WITH_GPU)
inference_analysis_api_test(test_analyzer_ernie ${ERNIE_INSTALL_DIR} analyzer_ernie_tester.cc)
endif()
inference_analysis_api_int8_test(test_analyzer_ernie_int8 ${ERNIE_INSTALL_DIR} analyzer_ernie_int8_tester.cc)
# Ernie large
......@@ -551,7 +553,9 @@ endif()
# bert, max_len=20, embedding_dim=128
set(BERT_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/bert_emb128")
download_model_and_data_without_verify(${BERT_INSTALL_DIR} "bert_emb128_model.tar.gz" "bert_data_len20.txt.tar.gz")
inference_analysis_api_test(test_analyzer_bert ${BERT_INSTALL_DIR} analyzer_bert_tester.cc)
if (WITH_GPU)
inference_analysis_api_test(test_analyzer_bert ${BERT_INSTALL_DIR} analyzer_bert_tester.cc)
endif()
# multiple models prediction
set(MMP_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/multi_model_prediction")
......@@ -741,13 +745,15 @@ set_tests_properties(lite_resnet50_test PROPERTIES TIMEOUT 120)
set_tests_properties(test_analyzer_mobilenet_transpose PROPERTIES TIMEOUT 120)
set_tests_properties(test_analyzer_resnet50 PROPERTIES TIMEOUT 120)
set_tests_properties(test_analyzer_ner PROPERTIES TIMEOUT 120)
set_tests_properties(test_analyzer_ernie PROPERTIES TIMEOUT 120)
set_tests_properties(test_analyzer_ernie_int8 PROPERTIES TIMEOUT 120)
set_tests_properties(test_analyzer_googlenet PROPERTIES TIMEOUT 120)
set_tests_properties(test_analyzer_small_dam PROPERTIES TIMEOUT 120)
set_tests_properties(test_analyzer_transformer PROPERTIES TIMEOUT 120)
set_tests_properties(test_analyzer_bert PROPERTIES TIMEOUT 120)
set_tests_properties(test_analyzer_mobilenet_depthwise_conv PROPERTIES TIMEOUT 120)
if (WITH_GPU)
set_tests_properties(test_analyzer_bert PROPERTIES TIMEOUT 120)
set_tests_properties(test_analyzer_ernie PROPERTIES TIMEOUT 120)
endif()
if(WITH_GPU AND TENSORRT_FOUND)
set_tests_properties(trt_mobilenet_test PROPERTIES TIMEOUT 120)
if(WITH_MKLDNN)
......
paddle/fluid/memory/allocation/allocator_facade.cc
浏览文件 @ 46894436
......@@ -493,7 +493,8 @@ class AllocatorFacadePrivate {
"support allocating managed memory.\n"
"If you don't actually need to use managed memory, please disable "
"it with command `export FLAGS_use_cuda_managed_memory=false`.\n"
"Or you must use the gpu device that supports managed memory."));
"Or you must use the gpu device that supports managed memory.",
p.device));
}
return std::make_shared<CUDAManagedAllocator>(p);
}
......
paddle/fluid/memory/allocation/auto_growth_best_fit_allocator.cc
浏览文件 @ 46894436
......@@ -18,6 +18,7 @@
#include <mutex> // NOLINT
#include "paddle/fluid/memory/allocation/aligned_allocator.h"
#include "paddle/fluid/platform/flags.h"
#include "paddle/fluid/platform/profiler/event_tracing.h"
PADDLE_DEFINE_EXPORTED_READONLY_bool(
free_idle_chunk, false,
......@@ -47,6 +48,8 @@ AutoGrowthBestFitAllocator::AutoGrowthBestFitAllocator(
phi::Allocation *AutoGrowthBestFitAllocator::AllocateImpl(
size_t unaligned_size) {
platform::RecordEvent("AutoGrowthBestFitAllocator::Allocate",
platform::TracerEventType::UserDefined, 9 /*level*/);
size_t size = AlignedSize(unaligned_size, alignment_);
VLOG(10) << "Allocate " << unaligned_size << " bytes, aligned to " << size;
......@@ -108,6 +111,8 @@ phi::Allocation *AutoGrowthBestFitAllocator::AllocateImpl(
}
void AutoGrowthBestFitAllocator::FreeImpl(phi::Allocation *allocation) {
platform::RecordEvent("AutoGrowthBestFitAllocator::Free",
platform::TracerEventType::UserDefined, 9 /*level*/);
VLOG(10) << "Free " << allocation->size()
<< " bytes, ptr = " << allocation->ptr();
std::lock_guard<SpinLock> guard(spinlock_);
......
paddle/fluid/memory/allocation/stream_safe_cuda_allocator.cc
浏览文件 @ 46894436
......@@ -13,6 +13,7 @@
// limitations under the License.
#include "paddle/fluid/memory/allocation/stream_safe_cuda_allocator.h"
#include "paddle/fluid/platform/profiler/event_tracing.h"
namespace paddle {
namespace memory {
......@@ -117,6 +118,8 @@ StreamSafeCUDAAllocator::~StreamSafeCUDAAllocator() {
bool StreamSafeCUDAAllocator::IsAllocThreadSafe() const { return true; }
phi::Allocation* StreamSafeCUDAAllocator::AllocateImpl(size_t size) {
platform::RecordEvent("StreamSafeCUDAAllocator::Allocate",
platform::TracerEventType::UserDefined, 9 /*level*/);
ProcessUnfreedAllocations();
VLOG(8) << "Try allocate " << size << " bytes";
AllocationPtr underlying_allocation;
......@@ -144,6 +147,8 @@ phi::Allocation* StreamSafeCUDAAllocator::AllocateImpl(size_t size) {
}
void StreamSafeCUDAAllocator::FreeImpl(phi::Allocation* allocation) {
platform::RecordEvent("StreamSafeCUDAAllocator::Free",
platform::TracerEventType::UserDefined, 9 /*level*/);
StreamSafeCUDAAllocation* stream_safe_cuda_allocation =
dynamic_cast<StreamSafeCUDAAllocation*>(allocation);
PADDLE_ENFORCE_NOT_NULL(stream_safe_cuda_allocation,
......
paddle/fluid/memory/cuda_managed_memory_test.cu
浏览文件 @ 46894436
......@@ -128,6 +128,9 @@ TEST(ManagedMemoryTest, OversubscribeGPUMemoryTest) {
}
TEST(ManagedMemoryTest, OOMExceptionTest) {
if (!platform::IsGPUManagedMemorySupported(0)) {
return;
}
EXPECT_THROW(Alloc(platform::CUDAPlace(0), size_t(1) << 60),
memory::allocation::BadAlloc);
}
......
paddle/fluid/operators/CMakeLists.txt
浏览文件 @ 46894436
......@@ -100,7 +100,7 @@ else()
cc_library(gather_scatter_kernel SRCS gather_scatter_kernel.cc gather_scatter_kernel.cu DEPS tensor)
endif()
set(OP_HEADER_DEPS ${OP_HEADER_DEPS} pten pten_api_utils gather_scatter_kernel)
set(OP_HEADER_DEPS ${OP_HEADER_DEPS} phi phi_api_utils gather_scatter_kernel)
register_operators(EXCLUDES py_layer_op py_func_op warpctc_op dgc_op load_combine_op lstm_op run_program_op eye_op
recurrent_op save_combine_op sparse_attention_op sync_batch_norm_op spectral_op ${OP_MKL_DEPS} DEPS ${OP_HEADER_DEPS})
......
paddle/fluid/operators/addmm_op.cc
浏览文件 @ 46894436
......@@ -12,11 +12,14 @@ 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. */
#include "paddle/fluid/operators/addmm_op.h"
#include <memory>
#include <string>
#include <unordered_map>
#include <vector>
#include "paddle/fluid/framework/infershape_utils.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/phi/core/infermeta_utils.h"
#include "paddle/phi/infermeta/ternary.h"
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#endif
......@@ -24,6 +27,8 @@ limitations under the License. */
namespace paddle {
namespace operators {
constexpr int kMULMKLDNNINT8 = 1;
using framework::OpKernelType;
using framework::Tensor;
......@@ -31,85 +36,6 @@ class AddMMOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE_EQ(ctx->HasInput("Input"), true,
platform::errors::NotFound(
"Input(Input) of AddMMOp should not be null."));
PADDLE_ENFORCE_EQ(
ctx->HasInput("X"), true,
platform::errors::NotFound("Input(X) of AddMMOp should not be null."));
PADDLE_ENFORCE_EQ(
ctx->HasInput("Y"), true,
platform::errors::NotFound("Input(Y) of AddMMOp should not be null."));
PADDLE_ENFORCE_EQ(ctx->HasOutput("Out"), true,
platform::errors::NotFound(
"Output(Out) of AddMMOp should not be null."));
auto input_dims = ctx->GetInputDim("Input");
auto x_dims = ctx->GetInputDim("X");
auto y_dims = ctx->GetInputDim("Y");
auto ndim_input = input_dims.size();
auto ndim_x = x_dims.size();
auto ndim_y = y_dims.size();
float alpha = ctx->Attrs().Get<float>("Alpha");
float beta = ctx->Attrs().Get<float>("Beta");
VLOG(3) << "addmm operator input.shape=" << input_dims
<< " x.shape=" << x_dims << " y.shape=" << y_dims
<< " beta=" << beta << " alpha=" << alpha
<< " ndim_input=" << ndim_input << " ndim_x=" << ndim_x
<< " ndim_y=" << ndim_y;
PADDLE_ENFORCE_NE(phi::product(input_dims), 0,
platform::errors::PreconditionNotMet(
"The Input variable Input(%s) has not "
"been initialized. You may need to confirm "
"if you put exe.run(startup_program) "
"after optimizer.minimize function.",
ctx->Inputs("Input").front()));
PADDLE_ENFORCE_NE(phi::product(x_dims), 0,
platform::errors::PreconditionNotMet(
"The Input variable X(%s) has not "
"been initialized. You may need to confirm "
"if you put exe.run(startup_program) "
"after optimizer.minimize function.",
ctx->Inputs("X").front()));
PADDLE_ENFORCE_NE(phi::product(y_dims), 0,
platform::errors::PreconditionNotMet(
"The Input variable Y(%s) has not "
"been initialized. You may need to confirm "
"if you put exe.run(startup_program) "
"after optimizer.minimize function.",
ctx->Inputs("Y").front()));
// dim check
PADDLE_ENFORCE_EQ(ndim_input, 2,
platform::errors::InvalidArgument(
"The input tensor input's dimension must be 2. "
"But received input's dimension = [%s].",
ndim_input));
PADDLE_ENFORCE_EQ(ndim_x, 2,
platform::errors::InvalidArgument(
"The input tensor x's dimension must be 2. "
"But received x's dimension = [%s].",
ndim_x));
PADDLE_ENFORCE_EQ(ndim_y, 2,
platform::errors::InvalidArgument(
"The input tensor y's dimension must be 2. "
"But received y's dimension = [%s].",
ndim_y));
std::vector<int64_t> output_dims;
output_dims.push_back(x_dims[0]);
output_dims.push_back(y_dims[1]);
ctx->SetOutputDim("Out", phi::make_ddim(output_dims));
ctx->ShareLoD("Input", /*->*/ "Out");
}
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const {
framework::LibraryType library = framework::LibraryType::kPlain;
......@@ -221,17 +147,11 @@ class AddMMOpGradMaker : public framework::SingleGradOpMaker<T> {
} // namespace paddle
namespace ops = paddle::operators;
DELCARE_INFER_SHAPE_FUNCTOR(addmm, AddmmInferShapeFunctor,
PT_INFER_META(phi::AddmmInferMeta));
REGISTER_OPERATOR(addmm, ops::AddMMOp, ops::AddMMOpMaker,
ops::AddMMOpGradMaker<paddle::framework::OpDesc>,
ops::AddMMOpGradMaker<paddle::imperative::OpBase>);
ops::AddMMOpGradMaker<paddle::imperative::OpBase>,
AddmmInferShapeFunctor);
REGISTER_OPERATOR(addmm_grad, ops::AddMMGradOp);
REGISTER_OP_CPU_KERNEL(
addmm, ops::AddMMKernel<paddle::platform::CPUDeviceContext, float>,
ops::AddMMKernel<paddle::platform::CPUDeviceContext, double>);
REGISTER_OP_CPU_KERNEL(
addmm_grad, ops::AddMMGradKernel<paddle::platform::CPUDeviceContext, float>,
ops::AddMMGradKernel<paddle::platform::CPUDeviceContext, double>);
paddle/fluid/operators/addmm_op.h 已删除 100644 → 0
浏览文件 @ e81773c9
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed 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. */
#pragma once
#include <boost/preprocessor/repetition/repeat.hpp>
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/operators/eigen/eigen_function.h"
#include "paddle/phi/kernels/funcs/blas/blas.h"
#include "paddle/phi/kernels/funcs/math_function.h"
namespace ops = paddle::operators;
namespace plat = paddle::platform;
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
template <typename T, size_t D, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
using Array1 = Eigen::DSizes<Eigen::DenseIndex, 1>;
using Array2 = Eigen::DSizes<Eigen::DenseIndex, 2>;
using Tensor = framework::Tensor;
constexpr int kMULMKLDNNINT8 = 1;
template <typename DeviceContext, typename T>
class AddMMKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
const Tensor* input = context.Input<Tensor>("Input");
const Tensor* x = context.Input<Tensor>("X");
const Tensor* y = context.Input<Tensor>("Y");
auto input_dims = input->dims();
auto x_dims = x->dims();
auto y_dims = y->dims();
// broadcast mode check
if (x_dims[0] != input_dims[0]) {
PADDLE_ENFORCE_EQ(input_dims[0], 1,
platform::errors::InvalidArgument(
"When x_dims[0] is not equal with input_dims[0], "
"input_dims[0] must be 1 but got %s",
input_dims[0]));
PADDLE_ENFORCE_EQ(
y_dims[1] == input_dims[1] || input_dims[1] == 1, true,
platform::errors::InvalidArgument(
"The input tensor shape mismatch, input shape=[%s], "
"x shape=[%s], y shape=[%s]",
input_dims, x_dims, y_dims));
}
// broadcast mode check
if (y_dims[1] != input_dims[1]) {
PADDLE_ENFORCE_EQ(input_dims[1], 1,
platform::errors::InvalidArgument(
"When y_dims[1] is not equal with input_dims[0], "
"input_dims[0] must be 1 but got %s",
input_dims[1]));
PADDLE_ENFORCE_EQ(
x_dims[0] == input_dims[0] || input_dims[0] == 1, true,
platform::errors::InvalidArgument(
"The input tensor shape mismatch, input shape=[%s], "
"x shape=[%s], y shape=[%s]",
input_dims, x_dims, y_dims));
}
// broadcast mode check
PADDLE_ENFORCE_EQ(
x_dims[1], y_dims[0],
platform::errors::InvalidArgument(
"The input tensor X's width must be equal with matrix Y' height. "
"But received X's shape = [%s], Y's shape = [%s].",
x_dims[1], y_dims[0]));
auto* out = context.Output<Tensor>("Out");
out->mutable_data<T>({x_dims[0], y_dims[1]}, context.GetPlace());
float alpha = context.template Attr<float>("Alpha");
float beta = context.template Attr<float>("Beta");
auto blas = phi::funcs::GetBlas<DeviceContext, T>(context);
// calc broadcast dim
Array2 bcast_dims;
bcast_dims[0] = x_dims[0] / input_dims[0];
bcast_dims[1] = y_dims[1] / input_dims[1];
VLOG(3) << "bcast_dims=[" << bcast_dims[0] << "," << bcast_dims[1] << "]";
// broadcast using eigen
auto eigen_input = EigenTensor<T, 2>::From(*input);
auto eigen_out = EigenTensor<T, 2>::From(*out);
auto& place =
*context.template device_context<DeviceContext>().eigen_device();
EigenBroadcast<std::decay_t<decltype(place)>, T, 2>::Eval(
place, eigen_out, eigen_input, bcast_dims);
blas.GEMM(false, false, x_dims[0], y_dims[1], x_dims[1], alpha,
x->data<T>(), x_dims[1], y->data<T>(), y_dims[1], beta,
out->data<T>(), y_dims[1]);
}
};
template <typename DeviceContext, typename T>
class AddMMGradKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* x = ctx.Input<framework::LoDTensor>("X");
auto* y = ctx.Input<framework::LoDTensor>("Y");
auto* dout = ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"));
auto in_dims = ctx.Input<framework::LoDTensor>("Input")->dims();
auto* dinput =
ctx.Output<framework::LoDTensor>(framework::GradVarName("Input"));
auto* dx = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
auto* dy = ctx.Output<framework::LoDTensor>(framework::GradVarName("Y"));
float alpha = ctx.Attr<float>("Alpha");
float beta = ctx.Attr<float>("Beta");
int total_elems = 0;
VLOG(3) << "alpha: " << alpha << " beta: " << beta;
if (dinput != nullptr) {
dinput->set_lod(dout->lod());
}
if (dx != nullptr) {
dx->set_lod(x->lod());
}
if (dy != nullptr) {
dy->set_lod(y->lod());
}
auto& dev_ctx = ctx.template device_context<DeviceContext>();
auto blas = phi::funcs::GetBlas<DeviceContext, T>(dev_ctx);
if (dinput) {
dinput->mutable_data<T>(ctx.GetPlace());
total_elems = in_dims[0] * in_dims[1];
auto& place =
*ctx.template device_context<DeviceContext>().eigen_device();
auto eigen_dout = EigenTensor<T, 2>::From(*dout);
auto eigen_dinput = EigenTensor<T, 2>::From(*dinput);
bool row_compress = in_dims[0] != dout->dims()[0];
bool col_compress = in_dims[1] != dout->dims()[1];
auto eigen_dinput_shape = Array2(dinput->dims()[0], dinput->dims()[1]);
if (row_compress && col_compress) {
eigen_dinput.device(place) =
eigen_dout.sum().eval().reshape(eigen_dinput_shape);
} else if (row_compress) {
eigen_dinput.device(place) =
eigen_dout.sum(Array1(0)).eval().reshape(eigen_dinput_shape);
} else if (col_compress) {
eigen_dinput.device(place) =
eigen_dout.sum(Array1(1)).eval().reshape(eigen_dinput_shape);
} else {
blas.VCOPY(total_elems, dout->data<T>(), dinput->data<T>());
}
blas.SCAL(total_elems, beta, dinput->data<T>());
}
if (dx) {
dx->mutable_data<T>(ctx.GetPlace());
total_elems = x->dims()[0] * x->dims()[1];
// dx = dout * y'. dx: M x K, dout : M x N, y : K x N
blas.MatMul(*dout, false, *y, true, dx);
blas.SCAL(total_elems, alpha, dx->data<T>());
}
if (dy) {
dy->mutable_data<T>(ctx.GetPlace());
total_elems = x->dims()[1] * y->dims()[1];
// dy = x' * dout. dy K x N, dout : M x N, x : M x K
blas.MatMul(*x, true, *dout, false, dy);
blas.SCAL(total_elems, alpha, dy->data<T>());
}
}
};
} // namespace operators
} // namespace paddle
paddle/fluid/operators/benchmark/op_tester.cc
浏览文件 @ 46894436
......@@ -24,7 +24,7 @@ limitations under the License. */
#include "paddle/fluid/platform/timer.h"
#include "paddle/fluid/pybind/pybind.h"
// pten
// phi
#include "paddle/phi/kernels/declarations.h"
namespace paddle {
......
paddle/fluid/operators/bilinear_tensor_product_op.cc
浏览文件 @ 46894436
......@@ -12,84 +12,18 @@ 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. */
#include "paddle/fluid/operators/bilinear_tensor_product_op.h"
#include <memory>
#include <string>
#include <vector>
#include "paddle/fluid/framework/infershape_utils.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/phi/core/infermeta_utils.h"
#include "paddle/phi/infermeta/backward.h"
#include "paddle/phi/infermeta/multiary.h"
namespace paddle {
namespace operators {
using framework::Tensor;
class BilinearTensorProductOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE_EQ(
ctx->HasInput("X"), true,
platform::errors::InvalidArgument("Input(X) should not be null."));
PADDLE_ENFORCE_EQ(
ctx->HasInput("Y"), true,
platform::errors::InvalidArgument("Input(Y) should not be null."));
PADDLE_ENFORCE_EQ(
ctx->HasInput("Weight"), true,
platform::errors::InvalidArgument("Input(Weight) should not be null."));
PADDLE_ENFORCE_EQ(
ctx->HasOutput("Out"), true,
platform::errors::InvalidArgument("Output(Out) should not be null."));
auto x_dims = ctx->GetInputDim("X");
auto y_dims = ctx->GetInputDim("Y");
auto weight_dims = ctx->GetInputDim("Weight");
PADDLE_ENFORCE_EQ(
x_dims.size(), 2UL,
platform::errors::InvalidArgument("The input(X) must be a 2D Tensor."));
PADDLE_ENFORCE_EQ(
y_dims.size(), 2UL,
platform::errors::InvalidArgument("The input(Y) must be a 2D Tensor."));
PADDLE_ENFORCE_EQ(
weight_dims.size(), 3UL,
platform::errors::InvalidArgument("Expected the input(Weight) is a 3D "
"tensor. But received %dD tensor.",
weight_dims.size()));
if (ctx->IsRuntime() || (x_dims[0] > 0 && y_dims[0] > 0)) {
PADDLE_ENFORCE_EQ(
x_dims[0], y_dims[0],
platform::errors::InvalidArgument(
"The first dimension(batch_size) of input(X) must be "
"equal to the first dimension of the input(Y)."));
}
PADDLE_ENFORCE_EQ(x_dims[1], weight_dims[1],
platform::errors::InvalidArgument(
"The second dimension of input(X) must be equal to "
"the second dimension of the input(Weight)."));
PADDLE_ENFORCE_EQ(y_dims[1], weight_dims[2],
platform::errors::InvalidArgument(
"The second dimension of input(Y) must be equal to "
"the third dimension of the input(Weight)."));
if (ctx->HasInput("Bias")) {
auto bias_dims = ctx->GetInputDim("Bias");
PADDLE_ENFORCE_EQ(bias_dims.size(), 2UL,
platform::errors::InvalidArgument(
"The Input(Bias) must be a 2-D tensor with "
"the 2nd dimension fixed to 1 (a row vector)."));
PADDLE_ENFORCE_EQ(bias_dims[0], 1UL,
platform::errors::InvalidArgument(
"The Input(Bias) must be a 2-D tensor with "
"the 2nd dimension fixed to 1 (a row vector)."));
PADDLE_ENFORCE_EQ(bias_dims[1], weight_dims[0],
platform::errors::InvalidArgument(
"The second dimension of input(Bias) must be equal "
"to the first dimension of the input(Weight)."));
}
ctx->SetOutputDim("Out", {x_dims[0], weight_dims[0]});
ctx->ShareLoD("X", /*->*/ "Out");
}
};
class BilinearTensorProductOpMaker : public framework::OpProtoAndCheckerMaker {
......@@ -125,59 +59,6 @@ Where $W_i$ is the $i$-th slice of Input(Weight);
class BilinearTensorProductOpGrad : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE_EQ(
ctx->HasInput("X"), true,
platform::errors::InvalidArgument("Input(X) should not be null."));
PADDLE_ENFORCE_EQ(
ctx->HasInput("Y"), true,
platform::errors::InvalidArgument("Input(Y) should not be null."));
PADDLE_ENFORCE_EQ(
ctx->HasInput("Weight"), true,
platform::errors::InvalidArgument("Input(Weight) should not be null."));
PADDLE_ENFORCE_EQ(ctx->HasInput(framework::GradVarName("Out")), true,
platform::errors::InvalidArgument(
"Input(Out@GRAD) should not be null."));
auto x_dims = ctx->GetInputDim("X");
auto y_dims = ctx->GetInputDim("Y");
auto weight_dims = ctx->GetInputDim("Weight");
auto out_dims = ctx->GetInputDim(framework::GradVarName("Out"));
PADDLE_ENFORCE_EQ(out_dims.size(), 2UL,
platform::errors::InvalidArgument(
"The input(Out@GRAD) must be a 2D Tensor."));
PADDLE_ENFORCE_EQ(
x_dims[0], out_dims[0],
platform::errors::InvalidArgument(
"The first dimension(batch_size) of input(Out@GRAD) must be "
"equal to the first dimension of the Input(X)."));
PADDLE_ENFORCE_EQ(
weight_dims[0], out_dims[1],
platform::errors::InvalidArgument(
"The second dimension of input(Out@GRAD) must be equal to "
"the third dimension of the Input(Weight)."));
auto bias_grad_name = framework::GradVarName("Bias");
if (ctx->HasOutput(bias_grad_name)) {
ctx->SetOutputDim(bias_grad_name, {1, out_dims[1]});
}
auto x_grad_name = framework::GradVarName("X");
auto y_grad_name = framework::GradVarName("Y");
auto weight_grad_name = framework::GradVarName("Weight");
if (ctx->HasOutput(x_grad_name)) {
ctx->SetOutputDim(x_grad_name, x_dims);
}
if (ctx->HasOutput(y_grad_name)) {
ctx->SetOutputDim(y_grad_name, y_dims);
}
if (ctx->HasOutput(weight_grad_name)) {
ctx->SetOutputDim(weight_grad_name, weight_dims);
}
}
};
template <typename T>
......@@ -208,21 +89,20 @@ class BilinearTensorProductGradOpMaker
} // namespace paddle
namespace ops = paddle::operators;
DELCARE_INFER_SHAPE_FUNCTOR(bilinear_tensor_product,
BilinearTensorProductInferShapeFunctor,
PT_INFER_META(phi::BilinearTensorProductInferMeta));
DELCARE_INFER_SHAPE_FUNCTOR(
bilinear_tensor_product_grad, BilinearTensorProductGradInferShapeFunctor,
PT_INFER_META(phi::BilinearTensorProductGradInferMeta));
REGISTER_OPERATOR(
bilinear_tensor_product, ops::BilinearTensorProductOp,
ops::BilinearTensorProductOpMaker,
ops::BilinearTensorProductGradOpMaker<paddle::framework::OpDesc>,
ops::BilinearTensorProductGradOpMaker<paddle::imperative::OpBase>);
ops::BilinearTensorProductGradOpMaker<paddle::imperative::OpBase>,
BilinearTensorProductInferShapeFunctor);
REGISTER_OPERATOR(bilinear_tensor_product_grad,
ops::BilinearTensorProductOpGrad);
REGISTER_OP_CPU_KERNEL(
bilinear_tensor_product,
ops::BilinearTensorProductKernel<paddle::platform::CPUDeviceContext, float>,
ops::BilinearTensorProductKernel<paddle::platform::CPUDeviceContext,
double>);
REGISTER_OP_CPU_KERNEL(
bilinear_tensor_product_grad,
ops::BilinearTensorProductGradKernel<paddle::platform::CPUDeviceContext,
float>,
ops::BilinearTensorProductGradKernel<paddle::platform::CPUDeviceContext,
double>);
ops::BilinearTensorProductOpGrad,
BilinearTensorProductGradInferShapeFunctor);
paddle/fluid/operators/bilinear_tensor_product_op.cu 已删除 100644 → 0
浏览文件 @ e81773c9
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed 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. */
#include "paddle/fluid/operators/bilinear_tensor_product_op.h"
namespace ops = paddle::operators;
REGISTER_OP_CUDA_KERNEL(
bilinear_tensor_product,
ops::BilinearTensorProductKernel<paddle::platform::CUDADeviceContext,
float>,
ops::BilinearTensorProductKernel<paddle::platform::CUDADeviceContext,
double>);
REGISTER_OP_CUDA_KERNEL(
bilinear_tensor_product_grad,
ops::BilinearTensorProductGradKernel<paddle::platform::CUDADeviceContext,
float>,
ops::BilinearTensorProductGradKernel<paddle::platform::CUDADeviceContext,
double>);
paddle/fluid/operators/bilinear_tensor_product_op.h 已删除 100644 → 0
浏览文件 @ e81773c9
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed 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. */
#pragma once
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/phi/kernels/funcs/blas/blas.h"
namespace paddle {
namespace operators {
using framework::Tensor;
template <typename T, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
template <typename DeviceContext, typename T>
class BilinearTensorProductKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* x = ctx.Input<Tensor>("X");
auto* y = ctx.Input<Tensor>("Y");
auto* weight = ctx.Input<Tensor>("Weight");
auto* bias = ctx.Input<Tensor>("Bias");
auto* out = ctx.Output<Tensor>("Out");
out->mutable_data<T>(ctx.GetPlace());
auto y_mat = EigenMatrix<T>::From(*y);
auto output_mat = EigenMatrix<T>::From(*out);
auto batch_size = x->dims()[0];
auto weight_dims = weight->dims();
int out_dim = weight_dims[0];
auto x_dim = weight_dims[1];
auto y_dim = weight_dims[2];
auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
auto& dev_ctx = ctx.template device_context<DeviceContext>();
// Create the intermediate variable to calculate the result of
// Input(X) multiplied by Input(Weight_i), the formula is:
// left_mul = X Weight_i.
Tensor left_mul;
left_mul.mutable_data<T>(phi::make_ddim({batch_size, y_dim}),
ctx.GetPlace());
auto left_mul_mat = EigenMatrix<T>::From(left_mul);
for (int i = 0; i < out_dim; ++i) {
auto output_col_vec = output_mat.chip(i, 1);
Tensor weight_mat =
weight->Slice(i, i + 1).Resize(phi::make_ddim({x_dim, y_dim}));
phi::funcs::GetBlas<DeviceContext, T>(dev_ctx).GEMM(
CblasNoTrans, CblasNoTrans, batch_size, y_dim, x_dim, 1, x->data<T>(),
weight_mat.data<T>(), 0, left_mul.data<T>());
output_col_vec.device(place) =
(left_mul_mat * y_mat).sum(Eigen::DSizes<int, 1>(1));
}
if (bias) {
auto bias_vec = EigenMatrix<T>::From(*bias);
Eigen::DSizes<int, 2> bcast(batch_size, 1);
output_mat.device(place) = bias_vec.broadcast(bcast) + output_mat;
}
}
};
template <typename DeviceContext, typename T>
class BilinearTensorProductGradKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
const Tensor* x = ctx.Input<Tensor>("X");
const Tensor* y = ctx.Input<Tensor>("Y");
const Tensor* weight = ctx.Input<Tensor>("Weight");
Tensor* d_x = ctx.Output<Tensor>(framework::GradVarName("X"));
Tensor* d_y = ctx.Output<Tensor>(framework::GradVarName("Y"));
Tensor* d_weight = ctx.Output<Tensor>(framework::GradVarName("Weight"));
Tensor* d_bias = ctx.Output<Tensor>(framework::GradVarName("Bias"));
const Tensor* d_out = ctx.Input<Tensor>(framework::GradVarName("Out"));
auto batch_size = x->dims()[0];
auto weight_dims = weight->dims();
int out_dim = weight_dims[0];
auto x_dim = weight_dims[1];
auto y_dim = weight_dims[2];
auto x_mat = EigenMatrix<T>::From(*x);
auto y_mat = EigenMatrix<T>::From(*y);
auto d_out_mat = EigenMatrix<T>::From(*d_out);
auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
auto& dev_ctx = ctx.template device_context<DeviceContext>();
// Create the intermediate variable to calculate the Output(Y@Grad).
Tensor x_scale;
x_scale.mutable_data<T>(phi::make_ddim({batch_size, x_dim}),
ctx.GetPlace());
auto x_scale_mat = EigenMatrix<T>::From(x_scale);
// Create the intermediate variable to calculate the Output(X@Grad).
Tensor y_scale;
y_scale.mutable_data<T>(phi::make_ddim({batch_size, y_dim}),
ctx.GetPlace());
auto y_scale_mat = EigenMatrix<T>::From(y_scale);
phi::funcs::SetConstant<DeviceContext, T> set_zero;
if (d_x) {
d_x->mutable_data<T>(ctx.GetPlace());
set_zero(dev_ctx, d_x, static_cast<T>(0));
}
if (d_y) {
d_y->mutable_data<T>(ctx.GetPlace());
set_zero(dev_ctx, d_y, static_cast<T>(0));
}
if (d_weight) {
d_weight->mutable_data<T>(ctx.GetPlace());
}
auto blas = phi::funcs::GetBlas<DeviceContext, T>(ctx);
// Caculate the Output(X@Grad) and Output(Y@Grad).
if (d_x || d_y || d_weight) {
Eigen::DSizes<int, 2> bcast_for_x(1, y_dim);
Eigen::DSizes<int, 2> bcast_for_y(1, x_dim);
Eigen::DSizes<int, 2> bcast_for_weight(1, x_dim);
for (int i = 0; i < out_dim; ++i) {
Tensor weight_i =
weight->Slice(i, i + 1).Resize(phi::make_ddim({x_dim, y_dim}));
auto output_vec = d_out_mat.chip(i, 1);
if (d_x) {
y_scale_mat.device(place) =
output_vec.reshape(Eigen::DSizes<int, 2>(batch_size, 1))
.broadcast(bcast_for_x) *
y_mat;
blas.GEMM(CblasNoTrans, CblasTrans, batch_size, x_dim, y_dim, 1,
y_scale.data<T>(), weight_i.data<T>(), 1, d_x->data<T>());
}
if (d_y || d_weight) {
auto output_vec_y =
output_vec.reshape(Eigen::DSizes<int, 2>(batch_size, 1))
.broadcast(bcast_for_y);
x_scale_mat.device(place) = output_vec_y * x_mat;
if (d_y) {
blas.GEMM(CblasNoTrans, CblasNoTrans, batch_size, y_dim, x_dim, 1,
x_scale.data<T>(), weight_i.data<T>(), 1, d_y->data<T>());
}
if (d_weight) {
Tensor d_weight_i = d_weight->Slice(i, i + 1).Resize(
phi::make_ddim({x_dim, y_dim}));
blas.GEMM(CblasTrans, CblasNoTrans, x_dim, y_dim, batch_size, 1,
x_scale.data<T>(), y->data<T>(), 0, d_weight_i.data<T>());
}
}
}
}
// calculate the gradient of Input(Bias).
if (d_bias) {
d_bias->mutable_data<T>(ctx.GetPlace());
auto d_bias_mat = framework::EigenVector<T>::Flatten(*d_bias);
d_bias_mat.device(place) = d_out_mat.sum(Eigen::DSizes<int, 1>(0));
}
}
};
} // namespace operators
} // namespace paddle
paddle/fluid/operators/cast_op.cc
浏览文件 @ 46894436
......@@ -138,7 +138,7 @@ class CastOp : public framework::OperatorWithKernel {
namespace ops = paddle::operators;
using CPU = paddle::platform::CPUDeviceContext;
// cast use pten kernel, so no need to REGISTER_OP_CPU_KERNEL here.
// cast use phi kernel, so no need to REGISTER_OP_CPU_KERNEL here.
REGISTER_OPERATOR(cast, ops::CastOp,
ops::CastOpGradMaker<paddle::framework::OpDesc>,
ops::CastOpGradMaker<paddle::imperative::OpBase>,
......
paddle/fluid/operators/cast_op.h
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/operators/cholesky_op.cu 已删除 100644 → 0
浏览文件 @ e81773c9
此差异已折叠。
paddle/fluid/operators/conj_op.h
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/operators/diag_v2_op.cu 已删除 100644 → 0
浏览文件 @ e81773c9
此差异已折叠。
paddle/fluid/operators/dot_op.h
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/operators/empty_op.h
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/operators/erfinv_op.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/operators/erfinv_op.h 已删除 100644 → 0
浏览文件 @ e81773c9
此差异已折叠。
paddle/fluid/operators/eye_op.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/operators/eye_op.h 已删除 100644 → 0
浏览文件 @ e81773c9
此差异已折叠。
paddle/fluid/operators/fill_op.h
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/operators/flatten_op.h
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/operators/gumbel_softmax_op.h 已删除 100644 → 0
浏览文件 @ e81773c9
此差异已折叠。
paddle/fluid/operators/lu_op.h
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/operators/mkldnn/log_softmax_mkldnn_op.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/operators/mv_op.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/operators/mv_op.cu 已删除 100644 → 0
浏览文件 @ e81773c9
此差异已折叠。
paddle/fluid/operators/mv_op.h 已删除 100644 → 0
浏览文件 @ e81773c9
此差异已折叠。
paddle/fluid/operators/optimizers/multi_tensor_apply.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/operators/pixel_shuffle_op.h 已删除 100644 → 0
浏览文件 @ e81773c9
此差异已折叠。
paddle/fluid/operators/randint_op.cu 已删除 100644 → 0
浏览文件 @ e81773c9
此差异已折叠。
paddle/fluid/operators/reduce_ops/reduce_max_op_mlu.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/operators/reduce_ops/reduce_min_op_mlu.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/operators/size_op.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/operators/softmax_cudnn_op.cu 已删除 100644 → 0
浏览文件 @ e81773c9
此差异已折叠。
paddle/fluid/operators/softmax_op.h 已删除 100644 → 0
浏览文件 @ e81773c9
此差异已折叠。
paddle/fluid/platform/for_range.h
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/platform/os_info.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/platform/os_info.h
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/platform/profiler/CMakeLists.txt 100644 → 100755
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/platform/profiler/cpu_utilization.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/platform/profiler/cpu_utilization.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/platform/profiler/test_extra_info.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/platform/profiler/utils.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/platform/transform.h
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/pybind/CMakeLists.txt
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/pybind/eager.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/pybind/eager_method.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/pybind/pybind.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/fluid/pybind/tensor_py.h
浏览文件 @ 46894436
此差异已折叠。
paddle/infrt/CMakeLists.txt
浏览文件 @ 46894436
此差异已折叠。
paddle/infrt/backends/CMakeLists.txt 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/infrt/backends/tensorrt/CMakeLists.txt 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/infrt/backends/tensorrt/test_trt_engine.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/infrt/backends/tensorrt/trt_engine.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/infrt/backends/tensorrt/trt_engine.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/infrt/backends/tensorrt/trt_options.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/infrt/backends/tensorrt/trt_utils.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/infrt/dialect/infrt_base.h
浏览文件 @ 46894436
此差异已折叠。
paddle/infrt/dialect/infrt_base.td
浏览文件 @ 46894436
此差异已折叠。
paddle/infrt/dialect/pd_ops.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/infrt/dialect/tensorrt/pd_lower_to_trt.td 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/infrt/dialect/tensorrt/trt_op_converter_pass.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/infrt/dialect/tensorrt/trt_op_converter_pass.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/CMakeLists.txt
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/api/CMakeLists.txt
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/api/all.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/api/include/tensor.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/api/lib/CMakeLists.txt
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/api/lib/api_declare.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/api/lib/api_registry.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/api/lib/api_utils.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/api/lib/data_transform.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/api/lib/sparse_api.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/api/lib/tensor.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/api/lib/tensor_method.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/backends/CMakeLists.txt
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/backends/all_context.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/backends/dynload/cupti.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/backends/xpu/xpu_info.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/common/CMakeLists.txt
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/common/backend.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/common/float16.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/common/layout.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/common/place.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/core/CMakeLists.txt
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/core/compat/op_utils.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/core/custom_kernel.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/core/dense_tensor.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/core/device_context.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/core/kernel_factory.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/core/kernel_utils.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/core/utils/data_type.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/infermeta/CMakeLists.txt
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/infermeta/backward.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/infermeta/backward.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/infermeta/binary.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/infermeta/multiary.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/infermeta/multiary.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/infermeta/ternary.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/infermeta/unary.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/infermeta/unary.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/CMakeLists.txt
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/addmm_grad_kernel.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/bilinear_tensor_product_kernel.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/copy_kernel.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/cpu/addmm_kernel.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/cpu/cholesky_grad_kernel.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/cpu/cholesky_kernel.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/cpu/diag_kernel.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/cpu/erfinv_grad_kernel.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/cpu/erfinv_kernel.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/cpu/eye_kernel.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/cpu/gumbel_softmax_grad_kernel.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/cpu/gumbel_softmax_kernel.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/cpu/increment_kernel.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/cpu/multinomial_kernel.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/cpu/mv_grad_kernel.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/cpu/mv_kernel.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/cpu/pixel_shuffle_grad_kernel.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/cpu/pixel_shuffle_kernel.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/cpu/randint_kernel.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/cpu/size_kernel.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/cpu/softmax_grad_kernel.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/cpu/softmax_kernel.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/diag_kernel.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/digamma_kernel.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/empty_kernel.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/empty_kernel.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/erfinv_grad_kernel.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/erfinv_kernel.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/expand_kernel.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/eye_kernel.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/full_kernel.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/funcs/diag_functor.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/funcs/diagonal.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/funcs/for_range.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/funcs/functors.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/gpu/addmm_grad_kernel.cu 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/gpu/addmm_kernel.cu 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/gpu/cholesky_grad_kernel.cu 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/gpu/cholesky_kernel.cu 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/gpu/diag_kernel.cu 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/gpu/erfinv_grad_kernel.cu 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/gpu/erfinv_kernel.cu 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/gpu/gumbel_softmax_grad_kernel.cu 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/gpu/increment_kernel.cu 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/gpu/mv_grad_kernel.cu 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/gpu/mv_kernel.cu 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/gpu/pixel_shuffle_grad_kernel.cu 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/gpu/pixel_shuffle_kernel.cu 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/gpu/randint_kernel.cu 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/gpu/size_kernel.cu 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/gpu/softmax_kernel.cu 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/gpudnn/softmax_grad_kernel_gpudnn.cu 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/gpudnn/softmax_kernel_gpudnn.cu 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/gumbel_softmax_grad_kernel.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/gumbel_softmax_kernel.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/impl/addmm_grad_kernel_impl.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/impl/addmm_kernel_impl.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/impl/erfinv_grad_kernel_impl.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/impl/erfinv_kernel_impl.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/impl/eye_kernel_impl.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/impl/full_kernel_impl.h 已删除 100644 → 0
浏览文件 @ e81773c9
此差异已折叠。
paddle/phi/kernels/impl/gumbel_softmax_kernel_impl.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/impl/increment_kernel_impl.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/impl/mv_kernel_impl.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/impl/pixel_shuffle_kernel_impl.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/impl/size_kernel_impl.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/impl/softmax_grad_kernel_impl.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/impl/softmax_kernel_impl.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/increment_kernel.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/math_kernel.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/multinomial_kernel.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/mv_grad_kernel.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/mv_kernel.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/pixel_shuffle_grad_kernel.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/pixel_shuffle_kernel.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/randint_kernel.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/size_kernel.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/softmax_grad_kernel.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/softmax_kernel.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/sparse/convolution_kernel.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/sparse/cpu/convolution.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/sparse/cpu/convolution_kernel.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/kernels/truncated_gaussian_random_kernel.h 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/ops/compat/addmm_sig.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/ops/compat/bilinear_tensor_product_sig.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/ops/compat/cholesky_sig.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/ops/compat/diag_sig.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/ops/compat/erfinv_sig.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/ops/compat/eye_sig.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/ops/compat/gumbel_softmax_sig.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/ops/compat/mv_sig.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/ops/compat/pixel_shuffle_sig.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/ops/compat/randint_sig.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/ops/compat/scale_sig.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/ops/compat/softmax_sig.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/ops/compat/truncated_gaussian_random_sig.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/tests/CMakeLists.txt
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/tests/api/CMakeLists.txt
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/tests/api/scale_api.h
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/tests/api/test_to_api.cc
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/tests/kernels/test_sparse_conv3d_dev_api.cc 0 → 100644
浏览文件 @ 46894436
此差异已折叠。
paddle/phi/tools/CMakeLists.txt
浏览文件 @ 46894436
此差异已折叠。
paddle/scripts/paddle_build.sh
浏览文件 @ 46894436
此差异已折叠。
paddle/utils/optional.h
浏览文件 @ 46894436
此差异已折叠。
python/paddle/autograd/py_layer.py
浏览文件 @ 46894436
此差异已折叠。
python/paddle/fluid/compiler.py
浏览文件 @ 46894436
此差异已折叠。
python/paddle/fluid/executor.py
浏览文件 @ 46894436
此差异已折叠。
python/paddle/tensor/search.py
浏览文件 @ 46894436
此差异已折叠。
python/setup.py.in
浏览文件 @ 46894436
此差异已折叠。
tools/infrt/get_phi_kernel_info.py
浏览文件 @ 46894436
此差异已折叠。
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或