After run gdr_copy_to_mapping, how does the kernel function utilize the data transferred from host? In other words, can cudamalloc and cudaMemcpy be replaced?

[LiShuang-codes]

A sample code is here, plz :

#define DTYPE float

__global__ void sample(const float* A, const float* B, int64_t n) {
    int i = blockIdx.x * blockDim.x + threadIdx.x;
    printf("%f at %p\n", A[0], A);
    if (i < N)
    {
       B[i] = A[i] +1;
    }
}

int main(){
  gdr_t gdr_handle;
  void *gpu_buffer;
  gpu_mem_handle_t mhandle;
  void *gpu_map_addr = NULL;
  gdr_mh_t mh;
  cudaError_t cuda_status;
  CUdevice dev;
  CUcontext dev_ctx;
  ASSERTDRV(cuInit(0));
  ASSERTDRV(cuDeviceGet(&dev, 0));
  ASSERTDRV(cuDevicePrimaryCtxRetain(&dev_ctx, dev));
  ASSERTDRV(cuCtxSetCurrent(dev_ctx));
  ASSERTDRV(gpu_mem_alloc(&mhandle, SIZE, 1, 1));
  gdr_handle = gdr_open();
  if (!gdr_handle)
  {
      printf("RDMA initialize failed.\n");
      goto close_and_clear;
  }
  printf("RDMA initialize succeed.\n");
  if (gdr_pin_buffer(gdr_handle, mhandle.ptr, 2*size_bytes, 0, 0, &mh) != 0)
  {
      printf("*** ERROR: gdr_pin_buffer failed\n");
      goto close_and_clear;
  }


  if (gdr_map(gdr_handle, mh, &gpu_map_addr, 2*size_bytes) != 0)
  {
      printf("*** ERROR: gdr_map failed\n");
      gdr_unpin_buffer(gdr_handle, mh);
      goto close_and_clear;
  }
  DTYPE *a = (DTYPE *)malloc(size_bytes);
  DTYPE *b = (DTYPE *)malloc(size_bytes);
gdr_copy_to_mapping(mh, gpu_map_addr, a, size_bytes);
//***HOW TO CALL KERNAL FUNCTION?
//***cudaDeviceSynchronize() CAN BE CALLED TO WAITE?
gdr_copy_from_mapping(mh, gpu_map_addr+size_bytes, b, size_bytes);
close_and_clear:
    gdr_unmap(gdr_handle, mh, gpu_map_addr, SIZE);
    gdr_unpin_buffer(gdr_handle, mh);
    cudaFree(gpu_buffer);
    gdr_close(gdr_handle);
    free(buffer);
    return 0;
}