Merge branch 'windows/build' into windows/online

test=develop
local_add_cudnn_lstm
peizhilin 6 years ago
commit 1f0449aded

@ -0,0 +1,214 @@
set(PART_CUDA_KERNEL_FILES)
function(op_library TARGET)
# op_library is a function to create op library. The interface is same as
# cc_library. But it handle split GPU/CPU code and link some common library
# for ops.
set(cc_srcs)
set(cu_srcs)
set(hip_cu_srcs)
set(miopen_hip_cc_srcs)
set(cu_cc_srcs)
set(cudnn_cu_cc_srcs)
set(CUDNN_FILE)
set(mkldnn_cc_srcs)
set(MKLDNN_FILE)
set(op_common_deps operator op_registry math_function)
set(options "")
set(oneValueArgs "")
set(multiValueArgs SRCS DEPS)
set(pybind_flag 0)
cmake_parse_arguments(op_library "${options}" "${oneValueArgs}"
"${multiValueArgs}" ${ARGN})
list(LENGTH op_library_SRCS op_library_SRCS_len)
if (${op_library_SRCS_len} EQUAL 0)
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cc)
list(APPEND cc_srcs ${TARGET}.cc)
endif()
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu.cc)
list(APPEND cu_cc_srcs ${TARGET}.cu.cc)
endif()
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu)
list(APPEND cu_srcs ${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)
list(APPEND cu_srcs ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.part.cu)
endif()
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.hip.cu)
list(APPEND hip_cu_srcs ${TARGET}.hip.cu)
endif()
string(REPLACE "_op" "_cudnn_op" CUDNN_FILE "${TARGET}")
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${CUDNN_FILE}.cu.cc)
list(APPEND cudnn_cu_cc_srcs ${CUDNN_FILE}.cu.cc)
endif()
if(WITH_AMD_GPU)
string(REPLACE "_op" "_miopen_op" MIOPEN_FILE "${TARGET}")
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${MIOPEN_FILE}.hip.cc)
list(APPEND miopen_hip_cc_srcs ${MIOPEN_FILE}.hip.cc)
endif()
endif()
if(WITH_MKLDNN)
string(REPLACE "_op" "_mkldnn_op" MKLDNN_FILE "${TARGET}")
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${MKLDNN_FILE}.cc)
list(APPEND mkldnn_cc_srcs ${MKLDNN_FILE}.cc)
endif()
endif()
else()
foreach(src ${op_library_SRCS})
if (${src} MATCHES ".*\\.hip.cu$")
list(APPEND hip_cu_srcs ${src})
elseif (${src} MATCHES ".*\\.cu$")
list(APPEND cu_srcs ${src})
elseif(${src} MATCHES ".*_cudnn_op.cu.cc$")
list(APPEND cudnn_cu_cc_srcs ${src})
elseif(WITH_AMD_GPU AND ${src} MATCHES ".*_miopen_op.hip.cc$")
list(APPEND miopen_hip_cc_srcs ${src})
elseif(WITH_MKLDNN AND ${src} MATCHES ".*_mkldnn_op.cc$")
list(APPEND mkldnn_cc_srcs ${src})
elseif(${src} MATCHES ".*\\.cu.cc$")
list(APPEND cu_cc_srcs ${src})
elseif(${src} MATCHES ".*\\.cc$")
list(APPEND cc_srcs ${src})
else()
message(FATAL_ERROR "${TARGET} Source file ${src} should only be .cc or .cu")
endif()
endforeach()
endif()
list(LENGTH cc_srcs cc_srcs_len)
if (${cc_srcs_len} EQUAL 0)
message(FATAL_ERROR "The op library ${TARGET} should contains at least one .cc file")
endif()
if (WIN32)
# remove windows unsupported op, because windows has no nccl, no warpctc such ops.
foreach(windows_unsupport_op "nccl_op" "gen_nccl_id_op" "warpctc_op" "hierarchical_sigmoid_op"
"crf_decoding_op" "select_op" "lstmp_op" "gru_op" "fusion_gru_op" "lstm_op" "fusion_lstm_op" "cumsum_op"
"fusion_seqconv_eltadd_relu_op" "channel_send_op" "channel_create_op" "channel_close_op" "channel_recv_op")
if ("${TARGET}" STREQUAL "${windows_unsupport_op}")
return()
endif()
endforeach()
endif(WIN32)
set(OP_LIBRARY ${TARGET} ${OP_LIBRARY} CACHE INTERNAL "op libs")
list(LENGTH op_library_DEPS op_library_DEPS_len)
if (${op_library_DEPS_len} GREATER 0)
set(DEPS_OPS ${TARGET} ${DEPS_OPS} PARENT_SCOPE)
endif()
if (WITH_GPU)
nv_library(${TARGET} SRCS ${cc_srcs} ${cu_cc_srcs} ${cudnn_cu_cc_srcs} ${mkldnn_cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS}
${op_common_deps})
elseif (WITH_AMD_GPU)
hip_library(${TARGET} SRCS ${cc_srcs} ${hip_cu_srcs} ${miopen_hip_cc_srcs} ${mkldnn_cc_srcs} DEPS ${op_library_DEPS}
${op_common_deps})
else()
cc_library(${TARGET} SRCS ${cc_srcs} ${mkldnn_cc_srcs} DEPS ${op_library_DEPS}
${op_common_deps})
endif()
# Define operators that don't need pybind here.
foreach(manual_pybind_op "compare_op" "logical_op" "nccl_op"
"tensor_array_read_write_op" "tensorrt_engine_op")
if ("${TARGET}" STREQUAL "${manual_pybind_op}")
set(pybind_flag 1)
endif()
endforeach()
# The registration of USE_OP, please refer to paddle/fluid/framework/op_registry.h.
# Note that it's enough to just adding one operator to pybind in a *_op.cc file.
# And for detail pybind information, please see generated paddle/pybind/pybind.h.
file(READ ${TARGET}.cc TARGET_CONTENT)
string(REGEX MATCH "REGISTER_OPERATOR\\(.*REGISTER_OPERATOR\\(" multi_register "${TARGET_CONTENT}")
string(REGEX MATCH "REGISTER_OPERATOR\\([a-z0-9_]*," one_register "${multi_register}")
if (one_register STREQUAL "")
string(REPLACE "_op" "" TARGET "${TARGET}")
else ()
string(REPLACE "REGISTER_OPERATOR(" "" TARGET "${one_register}")
string(REPLACE "," "" TARGET "${TARGET}")
endif()
# pybind USE_NO_KERNEL_OP
# HACK: if REGISTER_OP_CPU_KERNEL presents the operator must have kernel
string(REGEX MATCH "REGISTER_OP_CPU_KERNEL" regex_result "${TARGET_CONTENT}")
string(REPLACE "_op" "" TARGET "${TARGET}")
if (${pybind_flag} EQUAL 0 AND regex_result STREQUAL "")
file(APPEND ${pybind_file} "USE_NO_KERNEL_OP(${TARGET});\n")
set(pybind_flag 1)
endif()
# pybind USE_CPU_ONLY_OP
list(LENGTH cu_srcs cu_srcs_len)
list(LENGTH cu_cc_srcs cu_cc_srcs_len)
list(LENGTH mkldnn_cc_srcs mkldnn_cc_srcs_len)
list(LENGTH hip_cu_srcs hip_cu_srcs_len)
list(LENGTH miopen_hip_cc_srcs miopen_hip_cc_srcs_len)
if (${pybind_flag} EQUAL 0 AND ${mkldnn_cc_srcs_len} EQUAL 0 AND ${cu_srcs_len} EQUAL 0 AND ${cu_cc_srcs_len} EQUAL 0 AND
${hip_cu_srcs_len} EQUAL 0 AND ${miopen_hip_cc_srcs_len} EQUAL 0)
file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n")
set(pybind_flag 1)
endif()
# pybind USE_OP_DEVICE_KERNEL for CUDNN
list(LENGTH cudnn_cu_cc_srcs cudnn_cu_cc_srcs_len)
if (WITH_GPU AND ${cudnn_cu_cc_srcs_len} GREATER 0)
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, CUDNN);\n")
endif()
# pybind USE_OP_DEVICE_KERNEL for MIOPEN
if (WITH_AMD_GPU AND ${miopen_hip_cc_srcs_len} GREATER 0)
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MIOPEN);\n")
endif()
# pybind USE_OP_DEVICE_KERNEL for MKLDNN
if (WITH_MKLDNN AND ${mkldnn_cc_srcs_len} GREATER 0)
# Append first implemented MKLDNN activation operator
if (${MKLDNN_FILE} STREQUAL "activation_mkldnn_op")
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(relu, MKLDNN);\n")
else()
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MKLDNN);\n")
endif()
endif()
# pybind USE_OP
if (${pybind_flag} EQUAL 0)
# NOTE(*): activation use macro to regist the kernels, set use_op manually.
if(${TARGET} STREQUAL "activation")
file(APPEND ${pybind_file} "USE_OP(relu);\n")
elseif(${TARGET} STREQUAL "fake_dequantize")
file(APPEND ${pybind_file} "USE_OP(fake_dequantize_max_abs);\n")
elseif(${TARGET} STREQUAL "fake_quantize")
file(APPEND ${pybind_file} "USE_OP(fake_quantize_abs_max);\n")
elseif(${TARGET} STREQUAL "tensorrt_engine_op")
message(STATUS "Pybind skips [tensorrt_engine_op], for this OP is only used in inference")
elseif(${TARGET} STREQUAL "fc")
# HACK: fc only have mkldnn and cpu, which would mismatch the cpu only condition
file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n")
else()
file(APPEND ${pybind_file} "USE_OP(${TARGET});\n")
endif()
endif()
endfunction()
function(register_operators)
set(options "")
set(oneValueArgs "")
set(multiValueArgs EXCLUDES)
cmake_parse_arguments(register_operators "${options}" "${oneValueArgs}"
"${multiValueArgs}" ${ARGN})
file(GLOB OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*_op.cc")
string(REPLACE "_mkldnn" "" OPS "${OPS}")
string(REPLACE ".cc" "" OPS "${OPS}")
list(REMOVE_DUPLICATES OPS)
foreach(src ${OPS})
list(FIND register_operators_EXCLUDES ${src} _index)
if (${_index} EQUAL -1)
op_library(${src})
endif()
endforeach()
endfunction()

@ -17,7 +17,7 @@ limitations under the License. */
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/op_info.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/elementwise_op_function.h"
#include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/init.h"

@ -13,7 +13,7 @@ set(FLUID_CORE_MODULES proto_desc memory lod_tensor executor)
# TODO(panyx0718): Should this be called paddle_fluid_inference_api_internal?
cc_library(paddle_fluid_api
SRCS io.cc
DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB} ${GLOB_OPERATOR_DEPS})
get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES)
get_property(cuda_modules GLOBAL PROPERTY CUDA_MODULES)

@ -1,4 +1,4 @@
nv_library(tensorrt_engine SRCS engine.cc DEPS framework_proto device_context)
nv_library(tensorrt_engine SRCS engine.cc DEPS ${GLOB_OPERATOR_DEPS} framework_proto device_context)
nv_test(test_tensorrt SRCS test_tensorrt.cc DEPS dynload_cuda device_context dynamic_loader)
nv_test(test_tensorrt_engine SRCS test_engine.cc DEPS dynload_cuda tensorrt_engine)
add_subdirectory(plugin)

@ -6,34 +6,34 @@ pad_op.cc split_op.cc prelu_op.cc
DEPS tensorrt_engine tensorrt_plugin operator scope framework_proto op_registry)
nv_test(test_op_converter SRCS test_op_converter.cc DEPS
${FLUID_CORE_MODULES} tensorrt_engine tensorrt_converter)
${FLUID_CORE_MODULES} ${GLOB_OPERATOR_DEPS} tensorrt_engine tensorrt_converter)
nv_test(test_io_converter SRCS test_io_converter.cc io_converter.cc DEPS dynload_cuda dynamic_loader lod_tensor)
nv_test(test_trt_mul_op SRCS test_mul_op.cc mul_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine mul_op SERIAL)
DEPS ${FLUID_CORE_MODULES} ${GLOB_OPERATOR_DEPS} tensorrt_engine mul_op SERIAL)
nv_test(test_trt_fc_op SRCS test_fc_op.cc fc_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine mul_op SERIAL)
DEPS ${FLUID_CORE_MODULES} ${GLOB_OPERATOR_DEPS} tensorrt_engine mul_op SERIAL)
nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine activation_op SERIAL)
DEPS ${FLUID_CORE_MODULES} ${GLOB_OPERATOR_DEPS} tensorrt_engine activation_op SERIAL)
nv_test(test_trt_conv_op SRCS test_conv2d_op.cc conv2d_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine conv_op conv_transpose_op SERIAL)
DEPS ${FLUID_CORE_MODULES} ${GLOB_OPERATOR_DEPS} tensorrt_engine conv_op conv_transpose_op SERIAL)
nv_test(test_trt_pool2d_op SRCS test_pool2d_op.cc pool2d_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine pool_op SERIAL)
DEPS ${FLUID_CORE_MODULES} ${GLOB_OPERATOR_DEPS} tensorrt_engine pool_op SERIAL)
nv_test(test_trt_elementwise_op SRCS test_elementwise_op.cc elementwise_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine elementwise_add_op SERIAL)
DEPS ${FLUID_CORE_MODULES} ${GLOB_OPERATOR_DEPS} tensorrt_engine elementwise_add_op SERIAL)
nv_test(test_trt_softmax_op SRCS test_softmax_op.cc softmax_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine softmax_op SERIAL)
DEPS ${FLUID_CORE_MODULES} ${GLOB_OPERATOR_DEPS} tensorrt_engine softmax_op SERIAL)
nv_test(test_trt_batch_norm_op SRCS test_batch_norm_op.cc batch_norm_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine batch_norm_op SERIAL)
DEPS ${FLUID_CORE_MODULES} ${GLOB_OPERATOR_DEPS} tensorrt_engine batch_norm_op SERIAL)
nv_test(test_trt_concat_op SRCS test_concat_op.cc concat_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine concat_op SERIAL)
DEPS ${FLUID_CORE_MODULES} ${GLOB_OPERATOR_DEPS} tensorrt_engine concat_op SERIAL)
nv_test(test_trt_dropout_op SRCS test_dropout_op.cc dropout_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine dropout_op SERIAL)
DEPS ${FLUID_CORE_MODULES} ${GLOB_OPERATOR_DEPS} tensorrt_engine dropout_op SERIAL)
nv_test(test_trt_pad_op SRCS test_pad_op.cc pad_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine pad_op SERIAL)
DEPS ${FLUID_CORE_MODULES} ${GLOB_OPERATOR_DEPS} tensorrt_engine pad_op SERIAL)
nv_test(test_trt_split_op SRCS test_split_op.cc split_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine tensorrt_plugin
DEPS ${FLUID_CORE_MODULES} ${GLOB_OPERATOR_DEPS} tensorrt_engine tensorrt_plugin
split_op concat_op SERIAL)
nv_test(test_trt_prelu_op SRCS test_prelu_op.cc prelu_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine tensorrt_plugin
DEPS ${FLUID_CORE_MODULES} ${GLOB_OPERATOR_DEPS} tensorrt_engine tensorrt_plugin
prelu_op SERIAL)

File diff suppressed because it is too large Load Diff

@ -0,0 +1,4 @@
include(operators)
register_operators()
file(APPEND ${pybind_file} "USE_OP(less_than);\nUSE_OP(logical_and);\nUSE_NO_KERNEL_OP(read_from_array);\n")

@ -12,7 +12,7 @@ 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/compare_op.h"
#include "paddle/fluid/operators/controlflow/compare_op.h"
#include <string>
#include "paddle/fluid/framework/op_registry.h"

@ -12,7 +12,7 @@ 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/compare_op.h"
#include "paddle/fluid/operators/controlflow/compare_op.h"
REGISTER_COMPARE_KERNEL(less_than, CUDA, paddle::operators::LessThanFunctor);
REGISTER_COMPARE_KERNEL(less_equal, CUDA, paddle::operators::LessEqualFunctor);

@ -16,7 +16,7 @@ limitations under the License. */
#include <math.h>
#include <type_traits>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/elementwise_op_function.h"
#include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
#include "paddle/fluid/platform/transform.h"
namespace paddle {

@ -12,7 +12,7 @@ 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/logical_op.h"
#include "paddle/fluid/operators/controlflow/logical_op.h"
#include <string>
#include "paddle/fluid/framework/op_registry.h"

@ -12,7 +12,7 @@ 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/logical_op.h"
#include "paddle/fluid/operators/controlflow/logical_op.h"
REGISTER_BINARY_LOGICAL_KERNEL(logical_and, CUDA,
paddle::operators::LogicalAndFunctor);

@ -0,0 +1,2 @@
include(operators)
register_operators()

@ -40,4 +40,8 @@ endif()
detection_library(roi_perspective_transform_op SRCS roi_perspective_transform_op.cc roi_perspective_transform_op.cu)
#Export local libraries to parent
set(DETECTION_LIBRARY ${LOCAL_DETECTION_LIBS} PARENT_SCOPE)
# set(DETECTION_LIBRARY ${LOCAL_DETECTION_LIBS} PARENT_SCOPE)
foreach(src ${LOCAL_DETECTION_LIBS})
set(OP_LIBRARY ${src} ${OP_LIBRARY} CACHE INTERNAL "op libs")
endforeach()

@ -0,0 +1,40 @@
include(operators)
set(DISTRIBUTE_DEPS "")
if(WITH_GRPC)
set(DISTRIBUTE_DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf node)
else()
set(DISTRIBUTE_DEPS sendrecvop_brpc brpc leveldb snappystream snappy protobuf ssl crypto zlib node)
if(WITH_BRPC_RDMA)
find_library(IBVERBS_LIBRARY NAMES ibverbs)
ADD_LIBRARY(ibverbs SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET ibverbs PROPERTY IMPORTED_LOCATION ${IBVERBS_LIBRARY})
find_library(RDMACM_LIBRARY NAMES rdmacm)
ADD_LIBRARY(rdmacm SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET rdmacm PROPERTY IMPORTED_LOCATION ${RDMACM_LIBRARY})
set(DISTRIBUTE_DEPS ${DISTRIBUTE_DEPS} ibverbs rdmacm)
endif()
endif()
set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
file(GLOB OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*_op.cc")
list(REMOVE_DUPLICATES OPS)
foreach(src ${OPS})
set_source_files_properties(${src} PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
endforeach()
register_operators(EXCLUDES gen_nccl_id_op)
if(WITH_GPU AND NOT WIN32)
set(DISTRIBUTE_DEPS ${DISTRIBUTE_DEPS} nccl_common)
op_library(gen_nccl_id_op)
endif()
set(OPERATOR_DEPS ${OPERATOR_DEPS} ${DISTRIBUTE_DEPS} PARENT_SCOPE)
set(GLOB_DISTRIBUTE_DEPS ${DISTRIBUTE_DEPS} CACHE INTERNAL "distributed dependency")

@ -19,7 +19,7 @@ limitations under the License. */
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/detail/macros.h"
#include "paddle/fluid/operators/send_recv_util.h"
#include "paddle/fluid/operators/distributed_ops/send_recv_util.h"
#include "paddle/fluid/string/printf.h"
namespace paddle {

Some files were not shown because too many files have changed in this diff Show More

Loading…
Cancel
Save