Function TransDataPlace

paddle/phi/api/lib/data_transform.cc:223–255

}

inline phi::DenseTensor TransDataPlace(const phi::DenseTensor& tensor,
                                       Place dst_place) {
  VLOG(3) << "DeviceTransform in, src_place " << tensor.place()
          << " dst_place: " << dst_place;

  auto& pool = phi::DeviceContextPool::Instance();
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
  // NOTE(yy): TransDataPlace should wait for computation of input.
  if (tensor.place().GetType() != phi::AllocationType::GPUPINNED) {
    pool.Get(tensor.place())->Wait();
    pool.Get(dst_place)->Wait();
  }
#endif

  // FIXME(zcd): TransDataPlace is used to transform data from GPU to CPU and
  // the enforced checks have been done in GetDeviceContext, so the
  // `dev_ctx->Wait()` is necessary. But `dev_ctx->Wait()` will make the program
  // slow, especially when the number of elements is little, for example,
  // the elements of learning rate are one and it's CPU side.
  // One solution is to use a CUDA kernel to complete the copy operation when
  // the transforming is from CPU to GPU and the number of elements is little.
  // But the embarrassment is that this solution this solution makes training
  // slower.
  phi::DenseTensor out;
  phi::DeviceContext* dev_ctx = nullptr;
  if (dst_place.GetType() != AllocationType::CPU) {
    dev_ctx = pool.Get(dst_place);
  } else {
    dev_ctx = pool.Get(tensor.place());
  }
  phi::Copy(*dev_ctx, tensor, dst_place, true, &out);
  return out;
}

template <typename Context>
phi::DenseTensor TensorContiguous(const Context& dev_ctx,