file(GLOB GENERAL_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*_op.cc")
string(REPLACE ".cc" "" GENERAL_OPS "${GENERAL_OPS}")
set(DEPS_OPS "")
set(pybind_file ${PADDLE_SOURCE_DIR}/paddle/pybind/pybind.h)
file(WRITE ${pybind_file} "// Generated by the paddle/operator/CMakeLists.txt.  DO NOT EDIT!\n\n")
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(OP_LIBRARY ${TARGET} ${OP_LIBRARY} PARENT_SCOPE)
    set(cc_srcs)
    set(cu_srcs)
    set(cu_cc_srcs)
    set(op_common_deps operator op_registry math_function)
    set(options "")
    set(oneValueArgs "")
    set(multiValueArgs SRCS DEPS)
    set(pybind_flag 0)
    cmake_parse_arguments(op_library "${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()
    else()
        foreach(src ${op_library_SRCS})
            if (${src} MATCHES ".*\\.cu$")
                list(APPEND cu_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()

    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} ${cu_srcs} DEPS ${op_library_DEPS}
                ${op_common_deps})
    else()
        cc_library(${TARGET} SRCS ${cc_srcs} DEPS ${op_library_DEPS}
                ${op_common_deps})
    endif()

    # Define operators that don't need pybind here.
    foreach(manual_pybind_op "net_op" "compare_op" "logical_op" "nccl_op" "tensor_array_read_write_op")
        if ("${TARGET}" STREQUAL "${manual_pybind_op}")
            set(pybind_flag 1)
        endif()
    endforeach()

    # The registration of USE_OP, please refer to paddle/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 ${TARGET}.cc TARGET_CONTENT)
    string(REGEX MATCH "REGISTER_OP\\(.*REGISTER_OP\\(" multi_register "${TARGET_CONTENT}")
    string(REGEX MATCH "REGISTER_OP\\([a-z0-9_]*," one_register "${multi_register}")
    if (one_register STREQUAL "")
        string(REPLACE "_op" "" TARGET "${TARGET}")
    else ()
        string(REPLACE "REGISTER_OP(" "" 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)
    if (${pybind_flag} EQUAL 0 AND ${cu_srcs_len} EQUAL 0 AND ${cu_cc_srcs_len} EQUAL 0)
        file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n")
        set(pybind_flag 1)
    endif()

    # pybind USE_OP
    if (${pybind_flag} EQUAL 0)
        file(APPEND ${pybind_file} "USE_OP(${TARGET});\n")
    endif()
endfunction()

add_subdirectory(math)
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()

if(WITH_DISTRIBUTE)
    add_subdirectory(detail)
    set(DISTRIBUTE_DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib_target protobuf)
    set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
    op_library(send_op DEPS ${DISTRIBUTE_DEPS})
    set_source_files_properties(send_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
    op_library(recv_op DEPS ${DISTRIBUTE_DEPS})
    set_source_files_properties(recv_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
    cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS send_op recv_op sum_op executor)
else()
    set(DEPS_OPS ${DEPS_OPS} send_op recv_op)
endif()

op_library(cond_op DEPS framework_proto tensor net_op)
op_library(cross_entropy_op DEPS cross_entropy)
op_library(softmax_with_cross_entropy_op DEPS cross_entropy softmax)
op_library(softmax_op DEPS softmax)
op_library(detection_output_op DEPS softmax)
op_library(sequence_softmax_op DEPS softmax)
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_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)
op_library(lstm_op DEPS sequence2batch lstm_compute)
op_library(gru_op DEPS sequence2batch gru_compute)
op_library(recurrent_op DEPS executor)
op_library(warpctc_op DEPS dynload_warpctc sequence_padding math_function)
op_library(cos_sim_op DEPS cos_sim_functor)
op_library(parallel_do_op DEPS executor)

# Regist multiple Kernel to pybind
if (WITH_GPU)
op_library(conv_op SRCS conv_op.cc conv_op.cu.cc conv_cudnn_op.cu.cc DEPS vol2col)
op_library(pool_op SRCS pool_op.cc pool_op.cu.cc pool_cudnn_op.cu.cc DEPS pooling)
op_library(conv_transpose_op SRCS conv_transpose_op.cc conv_transpose_op.cu.cc
  conv_transpose_cudnn_op.cu.cc DEPS vol2col)
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(conv2d, CUDNN);\n")
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(pool2d, CUDNN);\n")
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(conv2d_transpose, CUDNN);\n")
else()
op_library(conv_op SRCS conv_op.cc DEPS vol2col)
op_library(pool_op SRCS pool_op.cc DEPS pooling)
op_library(conv_transpose_op SRCS conv_transpose_op.cc DEPS vol2col)
endif()

# FIXME(typhoonzero): save/load depends lodtensor serialization functions
op_library(save_op DEPS lod_tensor)
op_library(load_op DEPS lod_tensor)

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(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library")


cc_test(gather_test SRCS gather_test.cc DEPS tensor)
cc_test(net_op_test SRCS net_op_test.cc DEPS net_op)
cc_test(scatter_test SRCS scatter_test.cc DEPS tensor)
cc_test(beam_search_decode_op_test SRCS beam_search_decode_op_test.cc DEPS lod_tensor)
cc_test(strided_memcpy_test SRCS strided_memcpy_test.cc DEPS tensor paddle_memory)
if(WITH_GPU)
    cc_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context)
endif()
cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op)