include(operators)

# 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.tmp CACHE INTERNAL "pybind.h file")
set(pybind_file_final ${PADDLE_BINARY_DIR}/paddle/fluid/pybind/pybind.h)
file(WRITE ${pybind_file} "// Generated by the paddle/fluid/operators/CMakeLists.txt.  DO NOT EDIT!\n\n")

add_subdirectory(math)
add_subdirectory(eigen)
add_subdirectory(controlflow)
add_subdirectory(detection)
add_subdirectory(elementwise)
add_subdirectory(fused)
add_subdirectory(metrics)
add_subdirectory(optimizers)
add_subdirectory(reduce_ops)
add_subdirectory(sequence_ops)
add_subdirectory(string)
add_subdirectory(jit)
if(WITH_MKLDNN)
    add_subdirectory(mkldnn)
endif()


if(WITH_DISTRIBUTE)
    add_subdirectory(collective)
endif()

if (WITH_PSCORE)
    add_subdirectory(pscore)
endif()

add_subdirectory(amp)

add_subdirectory(reader)

if (NOT WIN32)
    add_subdirectory(nccl)
endif()

if (WITH_GPU AND TENSORRT_FOUND)
    add_subdirectory(tensorrt)
endif()

if (WITH_DLNNE)
    add_subdirectory(dlnne)
endif()

if (WITH_LITE)
    add_subdirectory(lite)
endif()

SET(OP_HEADER_DEPS xxhash executor)

if (WITH_GPU)
    if (${CMAKE_CUDA_COMPILER_VERSION} LESS 11.0)
        SET(OP_HEADER_DEPS ${OP_HEADER_DEPS} cub)
    endif()
endif()

if (WITH_POCKETFFT)
    SET(OP_HEADER_DEPS ${OP_HEADER_DEPS} pocketfft)
endif()


SET(OP_MKL_DEPS "")
if (NOT WITH_MKL OR NOT WITH_AVX)
    SET(OP_MKL_DEPS ${OP_MKL_DEPS} match_matrix_tensor_op)
    SET(OP_MKL_DEPS ${OP_MKL_DEPS} var_conv_2d_op)
endif()
if(WITH_COVERAGE OR WIN32 OR WITH_NV_JETSON)
    SET(OP_MKL_DEPS ${OP_MKL_DEPS} pyramid_hash_op)
endif()

if(WITH_UNITY_BUILD)
    # Load Unity Build rules for operators in paddle/fluid/operators.
    include(unity_build_rule.cmake)
endif()

set(OP_HEADER_DEPS ${OP_HEADER_DEPS} pten)

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 cinn_launch_op ${OP_MKL_DEPS} DEPS ${OP_HEADER_DEPS})

op_library(run_program_op SRCS run_program_op.cc run_program_op.cu.cc DEPS executor_cache ${OP_HEADER_DEPS})
op_library(save_combine_op DEPS string_array)
op_library(load_combine_op DEPS string_array)

if (WITH_GPU OR WITH_ROCM)
    if(WITH_ROCM)
        op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale SRCS warpctc_op.cc warpctc_op.cu)
    # warpctc_op needs cudnn 7 above
    elseif(${CUDNN_MAJOR_VERSION} VERSION_LESS 7)
        op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale SRCS warpctc_op.cc warpctc_op.cu)
    else()
        op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale)
    endif()
    op_library(sync_batch_norm_op)
    file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(sync_batch_norm);\n")
    if ((NOT WIN32) AND (NOT WITH_ROCM) AND (NOT PADDLE_WITH_ARM) AND (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_LESS 11.2) )
        op_library(sparse_attention_op)
        file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(sparse_attention);\n")
    endif()
else()
    op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale)
endif()

if (WITH_GPU OR WITH_ROCM)
    if (MKL_FOUND AND WITH_ONEMKL)
        op_library(spectral_op SRCS spectral_op.cc spectral_op.cu DEPS dynload_cuda dynload_mklrt ${OP_HEADER_DEPS})
        target_include_directories(spectral_op PRIVATE ${MKL_INCLUDE})
    else()
        op_library(spectral_op SRCS spectral_op.cc spectral_op.cu DEPS dynload_cuda ${OP_HEADER_DEPS})
    endif()
else()
    if (MKL_FOUND AND WITH_ONEMKL)
        op_library(spectral_op SRCS spectral_op.cc DEPS dynload_mklrt ${OP_HEADER_DEPS})
        target_include_directories(spectral_op PRIVATE ${MKL_INCLUDE})
    else()
        op_library(spectral_op SRCS spectral_op.cc DEPS ${OP_HEADER_DEPS})
    endif()
endif()

if (WITH_ASCEND_CL)
  op_library(sync_batch_norm_op)
  file(APPEND ${pybind_file} "USE_NO_KERNEL_OP(sync_batch_norm);\n")
endif()

op_library(lstm_op DEPS ${OP_HEADER_DEPS}  lstm_compute)
op_library(eye_op DEPS ${OP_HEADER_DEPS})
op_library(recurrent_op DEPS ${OP_HEADER_DEPS})

set(COMMON_OP_DEPS ${OP_HEADER_DEPS})

if (WITH_DGC)
    op_library(dgc_op DEPS dgc)
    file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(dgc);\n")
    set(COMMON_OP_DEPS ${COMMON_OP_DEPS} dgc)
endif()

cc_library(common_infer_shape_functions SRCS common_infer_shape_functions.cc DEPS operator)

set(COMMON_OP_DEPS ${COMMON_OP_DEPS} selected_rows_functor selected_rows lapack_function
lod_tensor maxouting unpooling pooling lod_rank_table context_project
sequence_pooling segment_pooling executor device_memory_aligment generator)
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} dynload_warpctc)
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence_padding sequence_scale cos_sim_functor memory jit_kernel_helper concat_and_split cross_entropy softmax vol2col im2col sampler sample_prob tree2col)
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence2batch lstm_compute matrix_bit_code gru_compute activation_functions beam_search fc matrix_inverse matrix_solve)
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} box_wrapper boost ps_gpu_wrapper)
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} common_infer_shape_functions)
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} eigen_function)
if (WITH_GPU OR WITH_ROCM)
  set(COMMON_OP_DEPS ${COMMON_OP_DEPS} depthwise_conv prelu bert_encoder_functor)
endif()
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} device_memory_aligment)
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} layer)
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} tensor_formatter)
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} op_version_registry)
if (WITH_ASCEND)
  set(COMMON_OP_DEPS ${COMMON_OP_DEPS} ascend_wrapper)
endif()

if (WITH_ASCEND_CL)
  cc_test(assign_op_npu_test SRCS assign_op_npu_test.cc DEPS assign_op)
  cc_library(npu_op_runner SRCS npu_op_runner.cc DEPS operator npu_info)
  set(COMMON_OP_DEPS ${COMMON_OP_DEPS} npu_op_runner)
endif()

if (WITH_CINN)
  cc_library(cinn_launch_op_helper SRCS cinn_launch_op_helper.cc DEPS operator cinn)
  cc_test(cinn_launch_op_helper_test SRCS cinn_launch_op_helper_test.cc DEPS cinn_launch_op_helper)
  op_library(cinn_launch_op SRCS cinn_launch_op.cc cinn_launch_op.cu.cc DEPS cinn_compiler cinn_launch_op_helper cinn ${OP_HEADER_DEPS})
  if (WITH_GPU)
      nv_test(cinn_launch_op_test SRCS cinn_launch_op_test.cc DEPS cinn_compiler cinn_launch_op elementwise_add_op)
  endif()
endif()

# 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(OPERATOR_DEPS ${OPERATOR_DEPS} ${COMMON_OP_DEPS})
set(GLOB_OPERATOR_DEPS ${OPERATOR_DEPS} CACHE INTERNAL "Global Op dependencies")

cc_test(test_common_infer_shape_functions SRCS test_common_infer_shape_functions.cc DEPS common_infer_shape_functions ${COMMON_OP_DEPS} activation_op elementwise_add_op softmax_op softmax)
cc_test(gather_test SRCS gather_test.cc DEPS tensor)
cc_test(assign_op_test SRCS assign_op_test.cc DEPS assign_op)
cc_test(scatter_test SRCS scatter_test.cc DEPS tensor math_function)
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 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 (WITH_GPU)
    nv_test(dropout_op_test SRCS dropout_op_test.cc DEPS dropout_op tensor generator)
    nv_test(test_leaky_relu_grad_grad_functor SRCS test_leaky_relu_grad_grad_functor.cc test_leaky_relu_grad_grad_functor.cu DEPS tensor device_context eigen3)
    nv_test(feed_forward_test SRCS feed_forward_test.cu DEPS elementwise_add_op matmul_op tensor generator)
elseif(WITH_ROCM)
    hip_test(dropout_op_test SRCS dropout_op_test.cc DEPS dropout_op tensor generator)
    hip_test(test_leaky_relu_grad_grad_functor SRCS test_leaky_relu_grad_grad_functor.cc test_leaky_relu_grad_grad_functor.cu DEPS tensor device_context eigen3)
else()
    cc_test(test_leaky_relu_grad_grad_functor SRCS test_leaky_relu_grad_grad_functor.cc DEPS tensor device_context eigen3)
endif()

cc_library(tensor_formatter SRCS tensor_formatter.cc DEPS ${OP_HEADER_DEPS})
if (WITH_PYTHON)
  cc_library(py_func_op SRCS py_func_op.cc DEPS op_registry python pybind)
  cc_library(py_layer_op SRCS py_layer_op.cc DEPS op_registry python pybind)
endif()

if (WITH_ASCEND_CL)
  cc_test(range_op_npu_test SRCS range_op_npu_test.cc DEPS op_registry range_op scope device_context enforce executor)
  cc_test(expand_op_npu_test SRCS expand_op_npu_test.cc DEPS op_registry expand_op eigen_function scope device_context enforce executor compare_op)
endif()

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

cc_test(op_debug_string_test SRCS op_debug_string_test.cc DEPS elementwise_add_op)
if (WITH_ASCEND_CL)
    cc_test(transpose_op_npu_test SRCS transpose_op_npu_test.cc DEPS op_registry transpose_op scope device_context enforce executor)
endif()


if(WITH_MKLDNN)
include(mkldnn/inplace_op_tests.cmake)
include(mkldnn/caching_tests.cmake)
include(mkldnn/nhwc_op_tests.cmake)
endif()

if(WITH_UNITY_BUILD)
    # Using Unity Build to compile operators, `register_operator` will cause
    # the unity library to lose some symbols.
    # The specified link dependency needs to be displayed here.
    target_link_libraries(paddle_operators_unity ${OP_HEADER_DEPS} ${COMMON_OP_DEPS})
endif()

if(WITH_ASCEND_CL)
cc_test(gelu_op_npu_test SRCS gelu_op_npu_test.cc DEPS op_registry gelu_op scope device_context enforce executor)
endif()

if (WITH_GPU OR WITH_ASCEND_CL)
cc_test(copy_cross_scope_test SRCS copy_cross_scope_test.cc DEPS op_registry copy_cross_scope_op scope device_context enforce executor)
endif()

copy_if_different(${pybind_file} ${pybind_file_final})