5
$\begingroup$

Asking it here and not on SO as it seems to be appropriate question for CG. I am learning NVIDIA NVENC API.The SDK supplies a sampled called "NvEncoderCudaInterop" .There is a chunk of code which copies YUV plane arrays from CPU to GPU buffers. This is the code:

 // copy luma
 CUDA_MEMCPY2D copyParam;
memset(&copyParam, 0, sizeof(copyParam));
copyParam.dstMemoryType = CU_MEMORYTYPE_DEVICE;
copyParam.dstDevice = pEncodeBuffer->stInputBfr.pNV12devPtr;
copyParam.dstPitch = pEncodeBuffer->stInputBfr.uNV12Stride;
copyParam.srcMemoryType = CU_MEMORYTYPE_HOST;
copyParam.srcHost = yuv[0];
copyParam.srcPitch = width;
copyParam.WidthInBytes = width;
copyParam.Height = height;
__cu(cuMemcpy2D(&copyParam));

// copy chroma

__cu(cuMemcpyHtoD(m_ChromaDevPtr[0], yuv[1], width*height / 4));
__cu(cuMemcpyHtoD(m_ChromaDevPtr[1], yuv[2], width*height / 4));

I do understand the rationale behind the procedure.The memory is copied to GPU for the kernel to process it.What I don't understand is why,in order to copy Y plane, cuMemcpy2D is used and for UV cuMemcpyHtoD?Why Y can't be copied using cuMemcpyHtoD as well?As far as I understand,YUV planes have the same linear memory layout.The only difference is their size.

$\endgroup$
1
  • $\begingroup$ I'm not too familiar with NVENC, but it looks as if the dstPitch for the Y array is being supplied by the NVENC API, so it might conceivably differ from srcPitch, thus requiring the 2D copy instead of a straight memory copy. Not sure why the Y channel would require this where Uand V don't, though. $\endgroup$ Commented Oct 1, 2015 at 4:45

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.