cc_library( phi_dynamic_loader SRCS dynamic_loader.cc port.cc DEPS enforce glog gflags) list( APPEND CUDA_SRCS cublas.cc cublasLt.cc cudnn.cc curand.cc cusolver.cc cusparse.cc nvtx.cc cufft.cc) if(NOT WITH_NV_JETSON) list(APPEND CUDA_SRCS nvjpeg.cc) endif() if(WITH_ROCM) list(APPEND HIP_SRCS rocblas.cc miopen.cc hiprand.cc hipfft.cc) endif() # There is no macOS version of NCCL. # Disable nvrtc and cuda_driver api on MacOS, and only do a early test on Linux and Windows. if(NOT APPLE) list(APPEND CUDA_SRCS nvrtc.cc cuda_driver.cc) if(WITH_NCCL) list(APPEND CUDA_SRCS nccl.cc) endif() if(WITH_ROCM) list(APPEND HIP_SRCS hiprtc.cc rocm_driver.cc) if(WITH_RCCL) list(APPEND HIP_SRCS rccl.cc) endif() if(CUPTI_FOUND) list(APPEND HIP_SRCS cupti.cc) endif() endif() endif() if(TENSORRT_FOUND) list(APPEND CUDA_SRCS tensorrt.cc) endif() if(CUSPARSELT_FOUND) list(APPEND CUDA_SRCS cusparseLt.cc) endif() configure_file(cupti_lib_path.h.in ${CMAKE_CURRENT_BINARY_DIR}/cupti_lib_path.h) if(CUPTI_FOUND) list(APPEND CUDA_SRCS cupti.cc) endif() if(WITH_ROCM) hip_library( phi_dynload_cuda SRCS ${HIP_SRCS} DEPS phi_dynamic_loader) cc_library( phi_dynload_warpctc SRCS warpctc.cc DEPS phi_dynamic_loader warpctc) elseif(WITH_ASCEND_CL) cc_library( phi_dynload_warpctc SRCS warpctc.cc DEPS phi_dynamic_loader warpctc npu_hccl) else() nv_library( phi_dynload_cuda SRCS ${CUDA_SRCS} DEPS phi_dynamic_loader) cc_library( phi_dynload_warpctc SRCS warpctc.cc DEPS phi_dynamic_loader warpctc) endif() if(WITH_MKLML) cc_library( phi_dynload_mklml SRCS mklml.cc DEPS phi_dynamic_loader mklml) endif() cc_library( phi_dynload_lapack SRCS lapack.cc DEPS phi_dynamic_loader) add_dependencies(phi_dynload_lapack extern_lapack) # TODO(TJ): add iomp, mkldnn? if(MKL_FOUND AND WITH_ONEMKL) message("ONEMKL INCLUDE directory is ${MKL_INCLUDE}") cc_library( phi_dynload_mklrt SRCS mklrt.cc DEPS phi_dynamic_loader) target_include_directories(phi_dynload_mklrt PRIVATE ${MKL_INCLUDE}) endif() if(WITH_CUDNN_FRONTEND) nv_test( cudnn_frontend_test SRCS cudnn_frontend_test.cc DEPS phi_dynload_cuda cudnn-frontend) endif()