Function CopyDataToFrom

src/core/tensor/tensor.cc:606–644

====================Tensor Operations=======================================

// ====================Tensor Operations=======================================
void CopyDataToFrom(Tensor *dst, const Tensor &src, const size_t num,
                    const size_t dst_offset, const size_t src_offset) {
  auto width = SizeOf(src.data_type());
  CHECK_EQ(width, SizeOf(dst->data_type()));
  size_t nBytes = num * width;
  auto d_offset = dst_offset * width;
  auto s_offset = src_offset * width;
  CHECK_GE(src.MemSize(), s_offset + nBytes);
  CHECK_GE(dst->MemSize(), d_offset + nBytes);

  Device *dev = nullptr;
  CopyDirection direct;
  std::shared_ptr<Device> src_dev = src.device(), dst_dev = dst->device();
  if (dst_dev->lang() != src_dev->lang()) {
    // let the none cpp device conduct copy op
    if (dst_dev->lang() == kCpp) {
      dev = src_dev.get();
      direct = kDeviceToHost;
    } else if (src_dev->lang() == kCpp) {
      dev = dst_dev.get();
      direct = kHostToDevice;
    } else {
      LOG(FATAL) << "Not support mem copy between Cuda and OpenCL device";
    }
  } else {
    dev = src_dev.get();
    direct = src_dev->lang() == kCpp ? kHostToHost : kDeviceToDevice;
  }

  Tensor &dstRef = *dst;
  dev->Exec(
      [dev, dstRef, src, nBytes, direct, d_offset,
       s_offset](Context *ctx) mutable {
        Block *from = src.block(), *to = dstRef.block();
        dev->CopyDataToFrom(to, from, nBytes, direct, (int)d_offset,
                            (int)s_offset, ctx);
      },
      {src.block()}, {dst->block()}, "CopyDataToFrom");
}

void RepeatDataToFrom(bool broadcast_flag, const vector<size_t> &repeats,
                      int axis, Tensor *dst, const Tensor &src,