PySurfaceDownloader (minor) issue: too fast!

[Yves33]

Hi again!
On my machine (core i7 12700 / GeForce RTX 3060), I find that running successfully.

PyDecoder.DecodeSingleSurface(nv12_surface)
PySurfaceDownloader.Run(nv12_surface, nv12_cpu_buffer)

may result in altered nv12 buffer ( top of image is cropped - see reconstituted and resized image from buffer below)

the problem

• only appears with hi res videos (here 5760 x 2880)
• does not appear when the same code is run in jupyter notebook (or at least appears less frequently...)
• does not appear if pycuda is used to copy to cpu (required GpuMem branch)
• does not appear if frame is converted to rgb on gpu then downloaded.
• can be solved by introducing 2ms delay between both operations (time.sleep(0.002))

minimal code to reproduce:
https://github.com/Yves33/Vali_luma_chroma_shift/blob/main/Vali_nv12_download.py

it seems to me that the download starts before decoded frame is ready.

[RomanArzumanyan]

Hi @Yves33

Thank you for the detailed analysis.
I suspect the actual source of problem is here:

VALI/src/TC/src/TaskDecodeFrame.cpp

Lines 518 to 534 in 3236822

	static void CopyToSurface(AVFrame& src, Surface& dst) {
	CUDA_MEMCPY2D m = {0};
	m.srcMemoryType = CU_MEMORYTYPE_DEVICE;
	m.dstMemoryType = CU_MEMORYTYPE_DEVICE;
	for (auto i = 0U; src.data[i]; i++) {
	m.srcDevice = (CUdeviceptr)src.data[i];
	m.srcPitch = src.linesize[i];
	m.dstDevice = dst.PixelPtr(i);
	m.dstPitch = dst.Pitch(i);
	m.WidthInBytes = dst.Width(i) * dst.ElemSize();
	m.Height = dst.Height(i);
	CudaCtxPush push_ctx(GetContextByDptr(m.dstDevice));
	ThrowOnCudaError(cuMemcpy2D(&m), __LINE__);
	}
	}

After I've switched to default CUDA stream and started to push context created by FFMpeg, the reproduction ratio fell but still some decoder unit tests fail to pass from time to time.

I'll continue investigation on this. Maybe if I make FFMpeg use the same CUDA context as VALI, it will be solved. Anyway, as I find something I'll come back to you.

[RomanArzumanyan]

Hi @Yves33

Please check out latest version 3.2.10. Looks like it solves the issue, at least 2 previously unstable hw decoder unit tests are now passing.

[Yves33]

version 3.2.10 sovles the issue (at least with my test videos...)