I am looking into some profiling issue of XLA. I noticed that in tf2.5 and newer SafeH2DMemcpy part is remove from convolution_thunk.cc/custom_call_thunk/convolution_thunk.

By looking into old code of convlution_thunk's memcpy part

  // Write the output tuple.
  const int kNumOutputs = 2;
  auto ptrs = absl::make_unique<void*[]>(kNumOutputs);
  ptrs[0] = result_buffer.opaque();
  ptrs[1] = scratch.opaque();
  se::DeviceMemory<void*> tuple_addr(
      buffer_allocations.GetDeviceAddress(tuple_result_buffer_));
  SafeH2DMemcpy(tuple_addr, std::move(ptrs), kNumOutputs, params.stream,
                params.deferred_host_callbacks);

I am wondering

• What are result_buffer, scratch_buffer and tuple_result_buffer here? Are they cpu buffer or gpu buffer?
• Why it does a H2D copy instead of D2H copy? Why result is not from gpu to cpu?
• Why this memcpy is now removed? What is the original purpose of this?

Would really appreciate any ideas on this.