file(GLOB_RECURSE DEVICE_SRC_LIST RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} "common/*.cc" "kernel_info.cc" "kernel_runtime.cc" "memory_manager.cc" "kernel_runtime_manager.cc" "convert_tensor_utils.cc" ) if (ENABLE_GPU) list(APPEND DEVICE_SRC_LIST "gpu/distribution/collective_init.cc") else () list(APPEND DEVICE_SRC_LIST "gpu/distribution/collective_fake_init.cc") endif () if (ENABLE_D) file(GLOB_RECURSE D_SRC_LIST RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} "ascend/*.cc" "kernel_adjust.cc") endif () if (ENABLE_CPU) file(GLOB_RECURSE CPU_SRC_LIST RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} "cpu/*.cc") endif () # gpu if (ENABLE_GPU) file(GLOB_RECURSE CUDA_SRC_LIST RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} "gpu/*.cc" "gpu/*.cu") # gpu_queue list(REMOVE_ITEM CUDA_SRC_LIST "gpu/blocking_queue.cc" "gpu/gpu_buffer_mgr.cc") add_library(gpu_queue SHARED "gpu/blocking_queue.cc" "gpu/gpu_buffer_mgr.cc") target_link_libraries(gpu_queue ${CMAKE_THREAD_LIBS_INIT} ${CUDA_PATH}/lib64/libcudart.so) list(REMOVE_ITEM CUDA_SRC_LIST "gpu/mpi/mpi_initializer.cc" "gpu/distribution/collective_wrapper.cc" "gpu/distribution/mpi_wrapper.cc" "gpu/distribution/nccl_wrapper.cc" ) if (ENABLE_MPI) include(ExternalProject) # gpu_collective add_library(gpu_collective SHARED "gpu/distribution/collective_wrapper.cc" "gpu/distribution/mpi_wrapper.cc" "gpu/distribution/nccl_wrapper.cc" ) # _ms_mpi pybind11_add_module(_ms_mpi "gpu/mpi/mpi_initializer.cc") target_link_libraries(_ms_mpi PRIVATE mindspore::pybind11_module mindspore::ompi) target_link_libraries(gpu_collective PRIVATE mindspore::ompi mindspore::nccl) endif () # add_library(_mindspore_device_cuda_obj OBJECT ${CUDA_SRC_LIST}) endif () add_library(_mindspore_device_obj OBJECT ${DEVICE_SRC_LIST} ${D_SRC_LIST} ${CPU_SRC_LIST})