CMakeLists.txt

cc_library(var_handle SRCS var_handle.cc DEPS place framework_proto node)
cc_library(op_handle_base SRCS op_handle_base.cc DEPS var_handle device_context lod_tensor)

cc_library(scale_loss_grad_op_handle SRCS scale_loss_grad_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory)
cc_library(fetch_op_handle SRCS fetch_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory)
cc_library(computation_op_handle SRCS computation_op_handle.cc DEPS framework_proto scope place operator op_registry)
cc_library(rpc_op_handle SRCS rpc_op_handle.cc DEPS framework_proto scope place operator op_registry)
cc_library(fetch_barrier_op_handle SRCS fetch_barrier_op_handle.cc DEPS framework_proto scope place operator op_registry)
cc_library(multi_devices_helper SRCS multi_devices_helper.cc DEPS graph graph_helper)

cc_library(variable_visitor SRCS variable_visitor.cc DEPS lod_tensor selected_rows)

if(WITH_DISTRIBUTE)
    if(NOT WITH_GRPC)
        set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
        set_source_files_properties(reduce_op_handle.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
    endif()
endif()


if(WITH_GPU)
    nv_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
            dynload_cuda variable_visitor)
    nv_library(fused_all_reduce_op_handle SRCS fused_all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
            dynload_cuda variable_visitor)

    if(WITH_DGC)
        nv_library(sparse_all_reduce_op_handle SRCS sparse_all_reduce_op_handle.cc DEPS op_handle_base scope 
            lod_tensor ddim memory dynload_cuda variable_visitor dgc all_reduce_op_handle)
    endif()

    if(WITH_DISTRIBUTE)
        nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
            ddim dynload_cuda selected_rows_functor sendrecvop_rpc)
    else()
        nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
            ddim dynload_cuda selected_rows_functor)
    endif()
    nv_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor dynload_cuda)
    nv_library(fused_broadcast_op_handle SRCS fused_broadcast_op_handle.cc DEPS broadcast_op_handle)

else()
    cc_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
             variable_visitor)
    cc_library(fused_all_reduce_op_handle SRCS fused_all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
            variable_visitor)
    if(WITH_DISTRIBUTE)
        cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
            ddim selected_rows_functor sendrecvop_rpc)
    else()
        cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
            ddim selected_rows_functor)
    endif()
    cc_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
    cc_library(fused_broadcast_op_handle SRCS fused_broadcast_op_handle.cc DEPS broadcast_op_handle)
endif()

cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)

cc_library(eager_deletion_op_handle SRCS eager_deletion_op_handle.cc DEPS lod_tensor selected_rows reference_count_pass_helper)

set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass reference_count_pass eager_deletion_pass memory_optimize_pass inplace_op_pass)
if (WITH_GPU)
  list(APPEND SSA_GRAPH_EXECUTOR_DEPS reference_count_pass)
endif()
cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ${SSA_GRAPH_EXECUTOR_DEPS})

cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope
        simple_threadpool device_context)

cc_library(parallel_ssa_graph_executor SRCS parallel_ssa_graph_executor.cc DEPS threaded_ssa_graph_executor)

set(ASYNC_SSA_GRAPH_EXECUTOR_DEPS threaded_ssa_graph_executor)
if(WITH_DISTRIBUTE)
    list(APPEND ASYNC_SSA_GRAPH_EXECUTOR_DEPS communicator)
endif()
cc_library(async_ssa_graph_executor SRCS async_ssa_graph_executor.cc DEPS ${ASYNC_SSA_GRAPH_EXECUTOR_DEPS})

cc_test(broadcast_op_test SRCS broadcast_op_handle_test.cc DEPS var_handle op_handle_base scope ddim memory
        device_context broadcast_op_handle)
cc_test(gather_op_test SRCS gather_op_handle_test.cc DEPS var_handle op_handle_base scope ddim memory
        device_context gather_op_handle)
cc_library(scope_buffered_ssa_graph_executor SRCS scope_buffered_ssa_graph_executor.cc DEPS ssa_graph_executor)
#cc_test(reduce_op_handle_test SRCS reduce_op_handle_test.cc DEPS var_handle op_handle_base scope ddim memory
#        device_context reduce_op_handle )
cc_library(fast_threaded_ssa_graph_executor SRCS fast_threaded_ssa_graph_executor.cc
        DEPS fetch_op_handle ssa_graph_executor scope simple_threadpool device_context)
cc_test(fused_broadcast_op_test SRCS fused_broadcast_op_handle_test.cc DEPS fused_broadcast_op_handle)

cc_library(build_strategy SRCS build_strategy.cc DEPS
        graph_viz_pass multi_devices_graph_pass
        multi_devices_graph_print_pass multi_devices_graph_check_pass
        fuse_elewise_add_act_pass multi_batch_merge_pass 
        fuse_relu_depthwise_conv_pass
        memory_optimize_pass lock_free_optimize_pass
        alloc_continuous_space_for_grad_pass fuse_all_reduce_op_pass backward_optimizer_op_deps_pass
        fuse_adam_op_pass fuse_sgd_op_pass fuse_momentum_op_pass record_skip_memory_opt_vars_pass)