Method AsynchronousMemcpyD2D

tensorflow/stream_executor/cuda/cuda_driver.cc:1140–1197

static */

}

/* static */ bool GpuDriver::AsynchronousMemcpyD2D(GpuContext* context,
                                                   CUdeviceptr gpu_dst,
                                                   CUdeviceptr gpu_src,
                                                   uint64 size,
                                                   CUstream stream) {
  ScopedActivateContext activation(context);
  if (size > 0) {
    CheckPointerIsValid(gpu_src, "src");
    CheckPointerIsValid(gpu_dst, "dst");
  }

  CUresult result;
  // CreatedContexts::GetAnyContext() doesn't works when ptr == 0.
  // This happens when the size is 0.
  if(gpu_dst == 0 || gpu_src == 0 || !UseCudaMallocAsyncAllocator()){
    result = cuMemcpyDtoDAsync(gpu_dst, gpu_src, size, stream);
  } else {
    // Any context work here.
    CUcontext dstContext = CreatedContexts::GetAnyContext(
                                                          absl::bit_cast<void*>(gpu_dst));
    CUcontext srcContext = CreatedContexts::GetAnyContext(
                                                          absl::bit_cast<void*>(gpu_src));

    if ((void*)dstContext == nullptr) {
      port::StatusOr<GpuContext*> context = GetPointerContext(gpu_dst);
      if (context.ok()) {
        dstContext = context.ValueOrDie()->context();
      }
    }

    if ((void*)srcContext == nullptr) {
      port::StatusOr<GpuContext*> context = GetPointerContext(gpu_src);
      if (context.ok()) {
        srcContext = context.ValueOrDie()->context();
      }
    }

    result = cuMemcpyPeerAsync(gpu_dst, dstContext, gpu_src, srcContext, size, stream);
  }

  if (result != CUDA_SUCCESS) {
    LOG(ERROR) << absl::StrFormat(
        "failed to enqueue async memcpy from device to device: %s"
        "; GPU dst: %p on %s %s"
        "; GPU src: %p on %s %s"
        "; can access? %s; size: %u=0x%x",
        ToString(result), absl::bit_cast<void*>(gpu_dst),
        CUDAPointerToMemorySpaceString(gpu_dst),
        CUDAPointerToDeviceString(gpu_dst), absl::bit_cast<void*>(gpu_src),
        CUDAPointerToMemorySpaceString(gpu_src),
        CUDAPointerToDeviceString(gpu_src),
        CUDAPointersToCanAccessString(gpu_src, gpu_dst), size, size);

    return false;
  }
  VLOG(2) << "successfully enqueued async memcpy d2d of " << size << " bytes";
  return true;
}