Refine operator cmake (#14413)

* wip simplify operator framework

* wip

* wip

* done test=develop

* clean test=develop

* fix test=develop

* fix deps test=develop

* fix cpu build test=develop

* fix tensorrt build test=develop

* fix tests test=develop

* fix test=develop

* fix cpu build test=develop

cmake/operators.cmake

+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()

paddle/fluid/framework/data_device_transform_test.cu

@@ -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"

paddle/fluid/inference/CMakeLists.txt

@@ -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)

paddle/fluid/inference/tensorrt/CMakeLists.txt

-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)

paddle/fluid/inference/tensorrt/convert/CMakeLists.txt

@@ -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)

paddle/fluid/operators/CMakeLists.txt

-file(GLOB GENERAL_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*_op.cc")
-string(REPLACE "_mkldnn" "" GENERAL_OPS "${GENERAL_OPS}")
-string(REPLACE ".cc" "" GENERAL_OPS "${GENERAL_OPS}")
-list(REMOVE_DUPLICATES GENERAL_OPS)
-set(DEPS_OPS "")
-set(pybind_file ${PADDLE_BINARY_DIR}/paddle/fluid/pybind/pybind.h)
-file(WRITE ${pybind_file} "// Generated by the paddle/fluid/operator/CMakeLists.txt.  DO NOT EDIT!\n\n")
-
-set(PART_CUDA_KERNEL_FILES)
-function(op_library TARGET)
-    # op_library is a function to create op library. The interface is same as
-    # cc_library. But it handle split GPU/CPU code and link some common library
-    # for ops.
-    set(cc_srcs)
-    set(cu_srcs)
-    set(hip_cu_srcs)
-    set(miopen_hip_cc_srcs)
-    set(cu_cc_srcs)
-    set(cudnn_cu_cc_srcs)
-    set(CUDNN_FILE)
-    set(mkldnn_cc_srcs)
-    set(MKLDNN_FILE)
-    set(op_common_deps operator op_registry math_function)
-    set(options "")
-    set(oneValueArgs "")
-    set(multiValueArgs SRCS DEPS)
-    set(pybind_flag 0)
-    cmake_parse_arguments(op_library "${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 ${TARGET} should contains at least one .cc file")
-    endif()
-    if (WIN32)
-    # remove windows unsupported op, because windows has no nccl, no warpctc such ops.
-    foreach(windows_unsupport_op "nccl_op" "gen_nccl_id_op" "warpctc_op" "hierarchical_sigmoid_op"
-     "crf_decoding_op" "select_op" "lstmp_op" "gru_op" "fusion_gru_op" "lstm_op" "fusion_lstm_op" "cumsum_op"
-      "fusion_seqconv_eltadd_relu_op" "channel_send_op" "channel_create_op" "channel_close_op" "channel_recv_op"
-            "fusion_seqexpand_concat_fc_op" "attention_lstm_op" "fused_embedding_fc_lstm_op" "fc_op")
-        if ("${TARGET}" STREQUAL "${windows_unsupport_op}")
-          return()
-        endif()
-    endforeach()
-    endif(WIN32)
-    set(OP_LIBRARY ${TARGET} ${OP_LIBRARY} PARENT_SCOPE)
+include(operators)
 
-    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()
+# clean cache and pybind_file content first when rebuild
+unset(GLOB_OP_LIB CACHE)
+unset(OP_LIBRARY CACHE)
+set(pybind_file ${PADDLE_BINARY_DIR}/paddle/fluid/pybind/pybind.h CACHE INTERNAL "pybind.h file")
+file(WRITE ${pybind_file} "// Generated by the paddle/fluid/operator/CMakeLists.txt.  DO NOT EDIT!\n\n")
 
 add_subdirectory(math)
-if (NOT WIN32)
-add_subdirectory(nccl)
-if(WITH_GPU)
-    op_library(nccl_op DEPS nccl_common)
-    file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(ncclAllReduce);\n")
-else()
-    set(DEPS_OPS ${DEPS_OPS} nccl_op)
-endif()
-endif() # NOT WIN32
+add_subdirectory(controlflow)
+add_subdirectory(csp)
+add_subdirectory(detection)
+add_subdirectory(elementwise)
+add_subdirectory(fused)
+add_subdirectory(metrics)
+add_subdirectory(optimizers)
+add_subdirectory(reduce_ops)
+add_subdirectory(sequence_ops)
 
-set(DISTRIBUTE_DEPS "")
 if(WITH_DISTRIBUTE)
     add_subdirectory(distributed)
-    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 ${DISTRIBUTE_DEPS} ibverbs rdmacm)
-        endif()
-    endif()
-
-    set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
-    foreach(dist_op "prefetch_op" "checkpoint_notify_op" "listen_and_serv_op" "send_op" "recv_op" "send_barrier_op" "fetch_barrier_op")
-        op_library(${dist_op} DEPS ${DISTRIBUTE_DEPS})
-        set_source_files_properties(${dist_op}.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
-    endforeach()
+    add_subdirectory(distributed_ops)
+endif()
 
-    #set_source_files_properties(send_recv_op_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
-    #cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS prefetch_op send_op
-    #        listen_and_serv_op sum_op executor SERIAL)
-    if(WITH_GPU AND NOT WIN32)
-        set_source_files_properties(test_send_nccl_id.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
-        cc_test(test_send_nccl_id SRCS test_send_nccl_id.cc DEPS listen_and_serv_op ${DISTRIBUTE_DEPS} executor SERIAL)
-        if(WITH_GRPC)
-            op_library(gen_nccl_id_op DEPS nccl_common sendrecvop_grpc)
-        else()
-            op_library(gen_nccl_id_op DEPS nccl_common sendrecvop_brpc)
-        endif()
-        set_source_files_properties(gen_nccl_id_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
-    else()
-        set(DEPS_OPS ${DEPS_OPS} gen_nccl_id_op)
-    endif() # WITH_GPU AND NOT WIN32
-else()
-    set(DEPS_OPS ${DEPS_OPS}  checkpoint_notify_op prefetch_op recv_op listen_and_serv_op send_op send_barrier_op fetch_barrier_op gen_nccl_id_op)
+if (NOT WIN32)
+    add_subdirectory(reader)
 endif()
 
-op_library(cross_entropy_op DEPS cross_entropy)
-if(WITH_GPU)
-  op_library(softmax_with_cross_entropy_op DEPS cross_entropy softmax cub)
-  op_library(sequence_softmax_op DEPS cub)
-else()
-  op_library(softmax_with_cross_entropy_op DEPS cross_entropy softmax)
+if (NOT WIN32)
+    add_subdirectory(nccl)
 endif()
 
-op_library(softmax_op DEPS softmax)
 if (WITH_GPU AND TENSORRT_FOUND)
-    op_library(tensorrt_engine_op DEPS tensorrt_engine tensorrt_converter)
-    file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(tensorrt_engine);\n")
-    nv_test(test_tensorrt_engine_op SRCS tensorrt_engine_op_test.cc
-      DEPS tensorrt_engine_op
-      analysis)
-else()
-    set(DEPS_OPS ${DEPS_OPS} tensorrt_engine_op)
+    add_subdirectory(tensorrt)
 endif()
-op_library(hash_op DEPS xxhash)
-op_library(clip_by_norm_op DEPS selected_rows_functor selected_rows)
-op_library(sum_op DEPS selected_rows_functor)
-op_library(sgd_op DEPS selected_rows_functor)
-op_library(print_op DEPS lod_tensor)
-op_library(adagrad_op DEPS selected_rows_functor)
-op_library(maxout_op DEPS maxouting)
-op_library(unpool_op DEPS unpooling)
-op_library(pool_op DEPS pooling)
-op_library(pool_with_index_op DEPS pooling)
-op_library(lod_rank_table_op DEPS lod_rank_table)
-op_library(lod_tensor_to_array_op DEPS lod_rank_table_op)
-op_library(array_to_lod_tensor_op DEPS lod_rank_table_op)
-op_library(max_sequence_len_op DEPS lod_rank_table)
-op_library(sequence_conv_op DEPS context_project)
-op_library(sequence_pool_op DEPS sequence_pooling)
-if (NOT WIN32)
-    op_library(lstm_op DEPS sequence2batch lstm_compute)
-    op_library(hierarchical_sigmoid_op DEPS matrix_bit_code)
-    op_library(lstmp_op DEPS sequence2batch lstm_compute)
-    op_library(gru_op DEPS sequence2batch gru_compute)
-endif(NOT WIN32)
-op_library(recurrent_op DEPS executor)
-op_library(cos_sim_op DEPS cos_sim_functor)
-op_library(parallel_do_op DEPS executor)
-op_library(unsqueeze_op DEPS reshape_op)
-op_library(squeeze_op DEPS reshape_op)
-op_library(flatten_op DEPS reshape_op)
-op_library(sequence_pad_op DEPS sequence_padding)
-op_library(unstack_op DEPS stack_op)
-op_library(fake_quantize_op DEPS memory)
-op_library(nce_op DEPS sampler)
-if (NOT WIN32)
-op_library(crf_decoding_op DEPS jit_kernel)
-op_library(fusion_lstm_op DEPS jit_kernel)
-endif(NOT WIN32)
-if (WITH_GPU)
-    op_library(conv_op DEPS vol2col depthwise_conv im2col)
-    op_library(layer_norm_op DEPS cub)
-    op_library(reduce_mean_op DEPS cub)
-    op_library(affine_channel_op DEPS cub)
-else()
-    op_library(conv_op DEPS vol2col im2col)
-endif()
-op_library(conv_transpose_op DEPS vol2col im2col)
 
-# FIXME(typhoonzero): save/load depends lodtensor serialization functions
-op_library(save_op DEPS lod_tensor)
-op_library(load_op DEPS lod_tensor)
-op_library(save_combine_op DEPS lod_tensor)
-op_library(load_combine_op DEPS lod_tensor)
-op_library(concat_op DEPS concat_and_split)
-op_library(tensor_array_to_tensor_op DEPS concat_op)
+register_operators(EXCLUDES warpctc_op)
 
-set(DEPS_OPS ${DEPS_OPS} warpctc_op)
+# warpctc_cudnn need cudnn 7 above
 if (WITH_GPU)
     if (${CUDNN_MAJOR_VERSION} VERSION_LESS 7)
         op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale SRCS warpctc_op.cc warpctc_op.cu.cc)
+    else()
+        op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale)
     endif()
+else()
+    op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale)
 endif()
-op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale)
-
-list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})
-
-foreach(src ${GENERAL_OPS})
-    op_library(${src})
-endforeach()
-
-file(APPEND ${pybind_file} "USE_OP(less_than);\nUSE_OP(logical_and);\nUSE_NO_KERNEL_OP(read_from_array);\n")
 
+set(COMMON_OP_DEPS "")
 
+set(COMMON_OP_DEPS ${COMMON_OP_DEPS} xxhash selected_rows_functor selected_rows lod_tensor maxouting unpooling pooling lod_rank_table context_project sequence_pooling executor dynload_warpctc sequence_padding sequence_scale cos_sim_functor memory jit_kernel concat_and_split cross_entropy softmax vol2col im2col sampler)
 if (NOT WIN32)
-add_subdirectory(reader)
-endif(NOT WIN32)
-foreach(src ${READER_LIBRARY})
-    set(OP_LIBRARY ${src} ${OP_LIBRARY})
-endforeach()
+  set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence2batch lstm_compute matrix_bit_code gru_compute activation_functions)
+endif()
+if (WITH_GPU)
+  set(COMMON_OP_DEPS ${COMMON_OP_DEPS} depthwise_conv cub)
+endif()
 
-add_subdirectory(detection)
-foreach(src ${DETECTION_LIBRARY})
-    set(OP_LIBRARY ${src} ${OP_LIBRARY})
-endforeach()
+# FIXME(typhoonzero): operator deps may not needed.
+# op_library(lod_tensor_to_array_op DEPS lod_rank_table_op)
+# op_library(array_to_lod_tensor_op DEPS lod_rank_table_op)
+# op_library(unsqueeze_op DEPS reshape_op)
+# op_library(squeeze_op DEPS reshape_op)
+# op_library(flatten_op DEPS reshape_op)
+# op_library(unstack_op DEPS stack_op)
+# op_library(tensor_array_to_tensor_op DEPS concat_op)
 
-set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library")
-set(GLOB_DISTRIBUTE_DEPS ${DISTRIBUTE_DEPS} CACHE INTERNAL "distributed dependency")
+set(OPERATOR_DEPS ${OPERATOR_DEPS} ${COMMON_OP_DEPS})
+set(GLOB_OPERATOR_DEPS ${OPERATOR_DEPS} CACHE INTERNAL "Global Op dependencies")
 
 cc_test(gather_test SRCS gather_test.cc DEPS tensor)
 cc_test(scatter_test SRCS scatter_test.cc DEPS tensor)
@@ -370,18 +76,6 @@ cc_test(beam_search_op_test SRCS beam_search_op_test.cc DEPS lod_tensor beam_sea
 cc_test(strided_memcpy_test SRCS strided_memcpy_test.cc DEPS tensor memory)
 cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op)
 cc_test(save_load_combine_op_test SRCS save_load_combine_op_test.cc DEPS save_combine_op load_combine_op)
-if(NOT WIN32)
-    nv_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context)
-endif()
 nv_test(dropout_op_test SRCS dropout_op_test.cc DEPS dropout_op tensor)
 
-if(WITH_GPU)
-    foreach(CUDA_KERNEL_FILE ${PART_CUDA_KERNEL_FILES})
-        file(READ ${CUDA_KERNEL_FILE} TARGET_CONTENT)
-        string(REGEX MATCH "REGISTER_OP_CUDA_KERNEL\\(\\n?([^,]+),.*" MATCHED ${TARGET_CONTENT})
-        if (MATCHED)
-            string(STRIP ${CMAKE_MATCH_1} MATCHED)
-            file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${MATCHED}, CUDA);\n")
-        endif()
-    endforeach()
-endif()
+set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library")

paddle/fluid/operators/controlflow/CMakeLists.txt

+include(operators)
+register_operators()
+
+file(APPEND ${pybind_file} "USE_OP(less_than);\nUSE_OP(logical_and);\nUSE_NO_KERNEL_OP(read_from_array);\n")

paddle/fluid/operators/compare_op.cc → paddle/fluid/operators/controlflow/compare_op.cc

@@ -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"
 

paddle/fluid/operators/compare_op.cu → paddle/fluid/operators/controlflow/compare_op.cu

@@ -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);

paddle/fluid/operators/compare_op.h → paddle/fluid/operators/controlflow/compare_op.h

@@ -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 {

paddle/fluid/operators/logical_op.cc → paddle/fluid/operators/controlflow/logical_op.cc

@@ -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"
 

paddle/fluid/operators/logical_op.cu → paddle/fluid/operators/controlflow/logical_op.cu

@@ -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);

paddle/fluid/operators/csp/CMakeLists.txt

+include(operators)
+register_operators()

paddle/fluid/operators/detection/CMakeLists.txt

@@ -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()

paddle/fluid/operators/distributed_ops/CMakeLists.txt

+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")

paddle/fluid/operators/checkpoint_notify_op.cc → paddle/fluid/operators/distributed_ops/checkpoint_notify_op.cc

@@ -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 {

paddle/fluid/operators/listen_and_serv_op.cc → paddle/fluid/operators/distributed_ops/listen_and_serv_op.cc

@@ -25,7 +25,7 @@ limitations under the License. */
 #include "paddle/fluid/operators/math/math_function.h"
 
 #include "paddle/fluid/operators/distributed/request_handler_impl.h"
-#include "paddle/fluid/operators/listen_and_serv_op.h"
+#include "paddle/fluid/operators/distributed_ops/listen_and_serv_op.h"
 
 DEFINE_int32(rpc_send_thread_num, 5, "number of threads for rpc send");
 DEFINE_int32(rpc_get_thread_num, 5, "number of threads for rpc get");

paddle/fluid/operators/merge_ids_op.cc → paddle/fluid/operators/distributed_ops/merge_ids_op.cc

@@ -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/merge_ids_op.h"
+#include "paddle/fluid/operators/distributed_ops/merge_ids_op.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/prefetch_op.cc → paddle/fluid/operators/distributed_ops/prefetch_op.cc

@@ -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"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/ref_by_trainer_id_op.cc → paddle/fluid/operators/distributed_ops/ref_by_trainer_id_op.cc

@@ -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/ref_by_trainer_id_op.h"
+#include "paddle/fluid/operators/distributed_ops/ref_by_trainer_id_op.h"
 #include <string>
 
 namespace paddle {

paddle/fluid/operators/ref_by_trainer_id_op.cu.cc → paddle/fluid/operators/distributed_ops/ref_by_trainer_id_op.cu.cc

@@ -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/ref_by_trainer_id_op.h"
+#include "paddle/fluid/operators/distributed_ops/ref_by_trainer_id_op.h"
 
 REGISTER_OP_CUDA_KERNEL(
     ref_by_trainer_id,

paddle/fluid/operators/send_op.cc → paddle/fluid/operators/distributed_ops/send_op.cc

@@ -20,7 +20,7 @@ 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/platform/profiler.h"
 
 namespace paddle {

paddle/fluid/operators/send_recv_op_test.cc → paddle/fluid/operators/distributed_ops/send_recv_op_test.cc

@@ -20,7 +20,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/framework/program_desc.h"
-#include "paddle/fluid/operators/listen_and_serv_op.h"
+#include "paddle/fluid/operators/distributed_ops/listen_and_serv_op.h"
 #include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/operators/math/selected_rows_functor.h"
 #include "paddle/fluid/string/printf.h"

paddle/fluid/operators/split_byref_op.cc → paddle/fluid/operators/distributed_ops/split_byref_op.cc

@@ -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/split_byref_op.h"
+#include "paddle/fluid/operators/distributed_ops/split_byref_op.h"
 #include "paddle/fluid/operators/split_op.h"
 
 namespace paddle {

paddle/fluid/operators/split_byref_op.cu.cc → paddle/fluid/operators/distributed_ops/split_byref_op.cu.cc

@@ -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/split_byref_op.h"
+#include "paddle/fluid/operators/distributed_ops/split_byref_op.h"
 namespace ops = paddle::operators;
 REGISTER_OP_CUDA_KERNEL(
     split_byref,

paddle/fluid/operators/split_ids_op.cc → paddle/fluid/operators/distributed_ops/split_ids_op.cc

@@ -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/split_ids_op.h"
+#include "paddle/fluid/operators/distributed_ops/split_ids_op.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/test_send_nccl_id.cc → paddle/fluid/operators/distributed_ops/test_send_nccl_id.cc

@@ -22,14 +22,14 @@ limitations under the License. */
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/operators/detail/macros.h"
 #include "paddle/fluid/operators/distributed/request_handler_impl.h"
-#include "paddle/fluid/operators/listen_and_serv_op.h"
+#include "paddle/fluid/operators/distributed_ops/listen_and_serv_op.h"
 #include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/operators/math/selected_rows_functor.h"
 #include "paddle/fluid/platform/nccl_helper.h"
 #include "paddle/fluid/string/printf.h"
 
 #ifdef PADDLE_WITH_GRPC
-#include "paddle/fluid/operators/send_recv_util.h"
+#include "paddle/fluid/operators/distributed_ops/send_recv_util.h"
 #endif
 
 USE_NO_KERNEL_OP(listen_and_serv);

paddle/fluid/operators/elementwise/CMakeLists.txt

+include(operators)
+register_operators()

paddle/fluid/operators/elementwise_add_mkldnn_op.cc → paddle/fluid/operators/elementwise/elementwise_add_mkldnn_op.cc

@@ -13,8 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/memory/memcpy.h"
-#include "paddle/fluid/operators/elementwise_add_op.h"
-#include "paddle/fluid/operators/elementwise_op_function.h"
+#include "paddle/fluid/operators/elementwise/elementwise_add_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
 
 #include "paddle/fluid/platform/mkldnn_helper.h"
 

paddle/fluid/operators/elementwise_add_op.cc → paddle/fluid/operators/elementwise/elementwise_add_op.cc

@@ -12,8 +12,8 @@ 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/elementwise_add_op.h"
-#include "paddle/fluid/operators/elementwise_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_add_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op.h"
 namespace ops = paddle::operators;
 REGISTER_ELEMWISE_GRAD_MAKER(elementwise_add, Add);
 REGISTER_ELEMWISE_EXPLICIT_OP(elementwise_add, "Add", "Out = X + Y", "Out",

paddle/fluid/operators/elementwise_add_op.cu → paddle/fluid/operators/elementwise/elementwise_add_op.cu

@@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #define EIGEN_USE_GPU
-#include "paddle/fluid/operators/elementwise_add_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_add_op.h"
 #include "paddle/fluid/platform/float16.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/elementwise_add_op.h → paddle/fluid/operators/elementwise/elementwise_add_op.h

@@ -15,8 +15,8 @@ limitations under the License. */
 #pragma once
 
 #include "paddle/fluid/framework/eigen.h"
-#include "paddle/fluid/operators/elementwise_op.h"
-#include "paddle/fluid/operators/elementwise_op_function.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
 #include "paddle/fluid/operators/math/blas.h"
 
 namespace paddle {

paddle/fluid/operators/elementwise_div_op.cc → paddle/fluid/operators/elementwise/elementwise_div_op.cc

@@ -12,8 +12,8 @@ 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/elementwise_div_op.h"
-#include "paddle/fluid/operators/elementwise_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_div_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op.h"
 namespace ops = paddle::operators;
 
 REGISTER_ELEMWISE_OP(elementwise_div, "Div", "Out = X / Y");

paddle/fluid/operators/elementwise_div_op.cu → paddle/fluid/operators/elementwise/elementwise_div_op.cu

@@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #define EIGEN_USE_GPU
-#include "paddle/fluid/operators/elementwise_div_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_div_op.h"
 
 namespace ops = paddle::operators;
 

paddle/fluid/operators/elementwise_div_op.h → paddle/fluid/operators/elementwise/elementwise_div_op.h

@@ -14,8 +14,8 @@ limitations under the License. */
 
 #pragma once
 
-#include "paddle/fluid/operators/elementwise_op.h"
-#include "paddle/fluid/operators/elementwise_op_function.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
 namespace paddle {
 namespace operators {
 

paddle/fluid/operators/elementwise_max_op.cc → paddle/fluid/operators/elementwise/elementwise_max_op.cc

@@ -12,8 +12,8 @@ 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/elementwise_max_op.h"
-#include "paddle/fluid/operators/elementwise_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_max_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op.h"
 namespace ops = paddle::operators;
 REGISTER_ELEMWISE_OP(elementwise_max, "Max", "Out = max(X, Y)");
 REGISTER_OP_CPU_KERNEL(

paddle/fluid/operators/elementwise_max_op.cu → paddle/fluid/operators/elementwise/elementwise_max_op.cu

@@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #define EIGEN_USE_GPU
-#include "paddle/fluid/operators/elementwise_max_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_max_op.h"
 
 namespace ops = paddle::operators;
 

paddle/fluid/operators/elementwise_max_op.h → paddle/fluid/operators/elementwise/elementwise_max_op.h

@@ -14,8 +14,8 @@ limitations under the License. */
 
 #pragma once
 
-#include "paddle/fluid/operators/elementwise_op.h"
-#include "paddle/fluid/operators/elementwise_op_function.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/elementwise_min_op.cc → paddle/fluid/operators/elementwise/elementwise_min_op.cc

@@ -12,8 +12,8 @@ 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/elementwise_min_op.h"
-#include "paddle/fluid/operators/elementwise_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_min_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op.h"
 namespace ops = paddle::operators;
 REGISTER_ELEMWISE_OP(elementwise_min, "Min", "Out = min(X, Y)");
 REGISTER_OP_CPU_KERNEL(

paddle/fluid/operators/elementwise_min_op.cu → paddle/fluid/operators/elementwise/elementwise_min_op.cu

@@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #define EIGEN_USE_GPU
-#include "paddle/fluid/operators/elementwise_min_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_min_op.h"
 
 namespace ops = paddle::operators;
 

paddle/fluid/operators/elementwise_min_op.h → paddle/fluid/operators/elementwise/elementwise_min_op.h

@@ -14,8 +14,8 @@ limitations under the License. */
 
 #pragma once
 
-#include "paddle/fluid/operators/elementwise_op.h"
-#include "paddle/fluid/operators/elementwise_op_function.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
 namespace paddle {
 namespace operators {
 

paddle/fluid/operators/elementwise_mul_op.cc → paddle/fluid/operators/elementwise/elementwise_mul_op.cc

@@ -12,9 +12,9 @@ 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/elementwise_mul_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_mul_op.h"
 #include <string>
-#include "paddle/fluid/operators/elementwise_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/elementwise_mul_op.cu → paddle/fluid/operators/elementwise/elementwise_mul_op.cu

@@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #define EIGEN_USE_GPU
-#include "paddle/fluid/operators/elementwise_mul_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_mul_op.h"
 
 namespace ops = paddle::operators;
 

paddle/fluid/operators/elementwise_mul_op.h → paddle/fluid/operators/elementwise/elementwise_mul_op.h

@@ -13,8 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
-#include "paddle/fluid/operators/elementwise_op.h"
-#include "paddle/fluid/operators/elementwise_op_function.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
 #include "paddle/fluid/operators/math/blas.h"
 
 namespace paddle {

paddle/fluid/operators/elementwise_pow_op.cc → paddle/fluid/operators/elementwise/elementwise_pow_op.cc

@@ -12,9 +12,9 @@ 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/elementwise_pow_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_pow_op.h"
 #include <string>
-#include "paddle/fluid/operators/elementwise_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/elementwise_pow_op.cu → paddle/fluid/operators/elementwise/elementwise_pow_op.cu

@@ -10,7 +10,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #define EIGEN_USE_GPU
-#include "paddle/fluid/operators/elementwise_pow_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_pow_op.h"
 
 namespace ops = paddle::operators;
 

paddle/fluid/operators/elementwise_pow_op.h → paddle/fluid/operators/elementwise/elementwise_pow_op.h

@@ -15,7 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include <cmath>
-#include "paddle/fluid/operators/elementwise_op_function.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/elementwise_sub_op.cc → paddle/fluid/operators/elementwise/elementwise_sub_op.cc

@@ -12,8 +12,8 @@ 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/elementwise_sub_op.h"
-#include "paddle/fluid/operators/elementwise_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_sub_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op.h"
 namespace ops = paddle::operators;
 REGISTER_ELEMWISE_GRAD_MAKER(elementwise_sub, Sub);
 REGISTER_ELEMWISE_EXPLICIT_OP(elementwise_sub, "Sub", "Out = X - Y", "Out",

paddle/fluid/operators/elementwise_sub_op.cu → paddle/fluid/operators/elementwise/elementwise_sub_op.cu

@@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #define EIGEN_USE_GPU
-#include "paddle/fluid/operators/elementwise_sub_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_sub_op.h"
 
 namespace ops = paddle::operators;
 

paddle/fluid/operators/elementwise_sub_op.h → paddle/fluid/operators/elementwise/elementwise_sub_op.h

@@ -13,8 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
-#include "paddle/fluid/operators/elementwise_op.h"
-#include "paddle/fluid/operators/elementwise_op_function.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/fused/CMakeLists.txt

+include(operators)
+register_operators()

paddle/fluid/operators/fused_elemwise_activation_op.cc → paddle/fluid/operators/fused/fused_elemwise_activation_op.cc

@@ -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/fused_elemwise_activation_op.h"
+#include "paddle/fluid/operators/fused/fused_elemwise_activation_op.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/fused_elemwise_activation_op.cu → paddle/fluid/operators/fused/fused_elemwise_activation_op.cu

@@ -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/fused_elemwise_activation_op.h"
+#include "paddle/fluid/operators/fused/fused_elemwise_activation_op.h"
 
 namespace ops = paddle::operators;
 REGISTER_OP_CUDA_KERNEL(

paddle/fluid/operators/fused_elemwise_activation_op.h → paddle/fluid/operators/fused/fused_elemwise_activation_op.h

@@ -19,7 +19,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/op_desc.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/detail/safe_ref.h"
-#include "paddle/fluid/operators/elementwise_op_function.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
 #include "paddle/fluid/operators/math/compound_functors.h"
 #include "paddle/fluid/operators/math/functors.h"
 

paddle/fluid/operators/fused_embedding_fc_lstm_op.cc → paddle/fluid/operators/fused/fused_embedding_fc_lstm_op.cc

@@ -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/fused_embedding_fc_lstm_op.h"
+#include "paddle/fluid/operators/fused/fused_embedding_fc_lstm_op.h"
 #include <string>
 #include "paddle/fluid/operators/math/blas.h"
 #include "paddle/fluid/operators/math/cpu_vec.h"

paddle/fluid/operators/fusion_gru_op.cc → paddle/fluid/operators/fused/fusion_gru_op.cc

@@ -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/fusion_gru_op.h"
+#include "paddle/fluid/operators/fused/fusion_gru_op.h"
 #include <cstring>  // for memcpy
 #include <string>
 #include "paddle/fluid/operators/math/blas.h"

paddle/fluid/operators/fusion_lstm_op.cc → paddle/fluid/operators/fused/fusion_lstm_op.cc

@@ -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/fusion_lstm_op.h"
+#include "paddle/fluid/operators/fused/fusion_lstm_op.h"
 #include <string>
 #include "paddle/fluid/operators/math/blas.h"
 #include "paddle/fluid/operators/math/fc_compute.h"

paddle/fluid/operators/fusion_seqconv_eltadd_relu_op.cc → paddle/fluid/operators/fused/fusion_seqconv_eltadd_relu_op.cc

@@ -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/fusion_seqconv_eltadd_relu_op.h"
+#include "paddle/fluid/operators/fused/fusion_seqconv_eltadd_relu_op.h"
 #include <algorithm>  // for min, max
 #include <string>
 #include "paddle/fluid/operators/math/blas.h"

paddle/fluid/operators/fusion_seqexpand_concat_fc_op.cc → paddle/fluid/operators/fused/fusion_seqexpand_concat_fc_op.cc

@@ -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/fusion_seqexpand_concat_fc_op.h"
+#include "paddle/fluid/operators/fused/fusion_seqexpand_concat_fc_op.h"
 #include <string>
 #include "paddle/fluid/operators/math/blas.h"
 #include "paddle/fluid/operators/math/cpu_vec.h"

paddle/fluid/operators/layer_norm_op.h

@@ -15,7 +15,7 @@ limitations under the License. */
 #pragma once
 #include "paddle/fluid/framework/eigen.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/blas.h"
 #include "paddle/fluid/operators/math/math_function.h"
 

paddle/fluid/operators/metrics/CMakeLists.txt

+include(operators)
+register_operators()

paddle/fluid/operators/accuracy_op.cc → paddle/fluid/operators/metrics/accuracy_op.cc

@@ -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/accuracy_op.h"
+#include "paddle/fluid/operators/metrics/accuracy_op.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/accuracy_op.cu → paddle/fluid/operators/metrics/accuracy_op.cu

@@ -14,7 +14,7 @@ limitations under the License. */
 
 #include <thrust/execution_policy.h>
 #include <thrust/reduce.h>
-#include "paddle/fluid/operators/accuracy_op.h"
+#include "paddle/fluid/operators/metrics/accuracy_op.h"
 #include "paddle/fluid/platform/cuda_primitives.h"
 #include "paddle/fluid/platform/gpu_info.h"
 

paddle/fluid/operators/auc_op.cc → paddle/fluid/operators/metrics/auc_op.cc

@@ -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/auc_op.h"
+#include "paddle/fluid/operators/metrics/auc_op.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/precision_recall_op.cc → paddle/fluid/operators/metrics/precision_recall_op.cc

@@ -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/precision_recall_op.h"
+#include "paddle/fluid/operators/metrics/precision_recall_op.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/nccl/CMakeLists.txt

 if(WITH_GPU AND NOT WIN32)
   nv_library(nccl_common SRCS nccl_gpu_common.cc DEPS device_context operator )
 endif()
+
+if(WITH_GPU)
+    op_library(nccl_op DEPS nccl_common)
+    file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(ncclAllReduce);\n")
+    set(OPERATOR_DEPS ${OPERATOR_DEPS} nccl_common PARENT_SCOPE)
+endif()
+
+if(NOT WIN32)
+    nv_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context)
+endif()

paddle/fluid/operators/optimizers/CMakeLists.txt

+include(operators)
+register_operators()