static */
| 331 | |
| 332 | /* static */ |
| 333 | void GPUUtil::CopyCPUTensorToGPU(const Tensor* cpu_tensor, |
| 334 | const DeviceContext* device_context, |
| 335 | Device* gpu_device, Tensor* gpu_tensor, |
| 336 | StatusCallback done, bool sync_dst_compute) { |
| 337 | VLOG(1) << "CopyCPUTensorToGPU"; |
| 338 | const DeviceBase::GpuDeviceInfo* dev_info = nullptr; |
| 339 | se::Stream* recv_stream = nullptr; |
| 340 | Status s = PrepareCopy(gpu_device, device_context, *cpu_tensor, gpu_tensor, |
| 341 | &dev_info, &recv_stream); |
| 342 | if (!s.ok()) { |
| 343 | done(s); |
| 344 | return; |
| 345 | } |
| 346 | |
| 347 | se::Stream* recv_host_to_device_stream = nullptr; |
| 348 | if (MergeComputeAndCopyStream()) { |
| 349 | recv_host_to_device_stream = recv_stream; |
| 350 | } else { |
| 351 | recv_host_to_device_stream = |
| 352 | static_cast<const GPUDeviceContext*>(device_context) |
| 353 | ->host_to_device_stream(); |
| 354 | if (recv_host_to_device_stream == nullptr) { |
| 355 | done(errors::Internal("No send gpu copy-out-stream is available.")); |
| 356 | return; |
| 357 | } |
| 358 | // Wait for the recv-stream to make sure the buffer is truly available. |
| 359 | if (sync_dst_compute) { |
| 360 | if (recv_host_to_device_stream != recv_stream) { |
| 361 | recv_host_to_device_stream->ThenWaitFor(recv_stream); |
| 362 | } |
| 363 | } |
| 364 | } |
| 365 | |
| 366 | const int64 total_bytes = cpu_tensor->TotalBytes(); |
| 367 | // Note that 0-size tensors have no backing buffer. |
| 368 | if (total_bytes > 0) { |
| 369 | void* src_ptr = GetBase(cpu_tensor); |
| 370 | void* dst_ptr = GetBase(gpu_tensor); |
| 371 | DeviceMemoryBase gpu_dst_ptr(dst_ptr, total_bytes); |
| 372 | recv_host_to_device_stream->ThenMemcpy(&gpu_dst_ptr, src_ptr, total_bytes); |
| 373 | } |
| 374 | |
| 375 | if (MergeComputeAndCopyStream()) { |
| 376 | done(Status::OK()); |
| 377 | } else { |
| 378 | // Use of cpu_tensor may outlive stack scope, so keep a ref. |
| 379 | TensorReference input_ref(*cpu_tensor); |
| 380 | dev_info->event_mgr->ThenExecute( |
| 381 | recv_host_to_device_stream, |
| 382 | [recv_host_to_device_stream, done, input_ref]() { |
| 383 | input_ref.Unref(); |
| 384 | if (!recv_host_to_device_stream->ok()) { |
| 385 | LOG(FATAL) << "CPU->GPU Memcpy failed"; |
| 386 | } |
| 387 | done(Status::OK()); |
| 388 | }); |
| 389 | } |
| 390 | } |
nothing calls this directly
no test coverage detected