Skip to content

BETA CUDA interface: clarify 10->8 bit conversions #934

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
NicolasHug merged 2 commits into from
Oct 6, 2025
Merged

BETA CUDA interface: clarify 10->8 bit conversions #934

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
33ae295
Clarify stuff
NicolasHug Oct 5, 2025
001bbc8
Clearer bitwise logic
NicolasHug Oct 5, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
20 changes: 20 additions & 0 deletions src/torchcodec/_core/BetaCudaDeviceInterface.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -108,10 +108,30 @@ static UniqueCUvideodecoder createDecoder(CUVIDEOFORMAT* videoFormat) {
" vs supported:",
caps.nMaxMBCount);

// Below we'll set the decoderParams.OutputFormat to NV12, so we need to make
// sure it's actually supported.
TORCH_CHECK(
caps.nOutputFormatMask & (1 << cudaVideoSurfaceFormat_NV12),
"NV12 output format is not supported for this configuration. ",
"Codec: ",
static_cast<int>(videoFormat->codec),
", chroma format: ",
static_cast<int>(videoFormat->chroma_format),
", bit depth: ",
videoFormat->bit_depth_luma_minus8 + 8);

// Decoder creation parameters, most are taken from DALI
CUVIDDECODECREATEINFO decoderParams = {};
decoderParams.bitDepthMinus8 = videoFormat->bit_depth_luma_minus8;
decoderParams.ChromaFormat = videoFormat->chroma_format;
// We explicitly request NV12 format, which means 10bit videos will be
// automatically converted to 8bits by NVDEC itself. That is, the raw frames
// we get back from cuvidMapVideoFrame will already be in 8bit format. We
// won't need to do the conversion ourselves, so that's a lot easier.
// In the default interface, we have to do the 10 -> 8bits conversion
// ourselves later in convertAVFrameToFrameOutput(), because FFmpeg explicitly
// requests 10 or 16bits output formats for >8-bit videos!
// https://github.com/FFmpeg/FFmpeg/blob/e05f8acabff468c1382277c1f31fa8e9d90c3202/libavcodec/nvdec.c#L376-L403
decoderParams.OutputFormat = cudaVideoSurfaceFormat_NV12;
decoderParams.ulCreationFlags = cudaVideoCreate_Default;
decoderParams.CodecType = videoFormat->codec;
Expand Down
20 changes: 8 additions & 12 deletions test/test_decoders.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1288,18 +1288,15 @@ def test_10bit_videos(self, device, asset):
# This just validates that we can decode 10-bit videos.
# TODO validate against the ref that the decoded frames are correct

if device == "cuda:0:beta":
# This fails on our BETA interface on asset 0 (only!) with:
if device == "cuda:0:beta" and asset is H264_10BITS:
# This fails on the BETA interface with:
#
# RuntimeError: Codec configuration not supported on this GPU.
# Codec: 4, chroma format: 1, bit depth: 10
#
# I don't remember but I suspect asset 0 is actually the one that
# fallsback to the CPU path on the default CUDA interface (that
# would make sense)
# We should investigate if and how we could make that fallback
# happen for the BETA interface.
pytest.skip("TODONVDEC P2 - investigate and unskip")
# It works on the default interface because FFmpeg fallsback to the
# CPU, while the BETA interface doesn't.
pytest.skip("Asset not supported by NVDEC")
Copy link
Contributor Author

@NicolasHug NicolasHug Oct 5, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Note above how we're now not skipping this test anymore for the H265_10BITS asset. Support for H265_10BITS isn't introduced by this PR, it was already working before, but now we know we can safely unskip it.

Also, this isn't visible in this PR and not in scope anyway, but there's something interesting: on H265_10BITS, the frame we decode with the BETA interface looks good, but it is slightly different than the on from the default CUDA interface. The difference is due to the fact that in the BETA interface, the 10 -> 8 bit conversion is done by NVDEC itself, while in the default interface it's done later via filtergraph in convertAVFrameToFrameOutput().

So overall we should expect small differences for 10bit videos between the BETA and default interface. I will follow-up on that by creating an issue. Generally, the BETA interfaces allows for drastically simpler code (no explicit conversion needed!)


decoder = VideoDecoder(asset.path, device=device)
decoder.get_frame_at(10)
Expand Down Expand Up @@ -1674,12 +1671,11 @@ def test_beta_cuda_interface_backwards(self, asset, seek_mode):

@needs_cuda
def test_beta_cuda_interface_small_h265(self):
# TODONVDEC P2 investigate why/how the default interface can decode this
# video.
# Test to illustrate current difference in behavior between the BETA and
# the default interface: this video isn't supported by NVDEC, but in the
# default interface, FFMPEG fallsback to the CPU while we don't.
Copy link
Contributor Author

@NicolasHug NicolasHug Oct 5, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Above is a drive-by, I was able to confirm that the default interface can decode this video because it falls back to the CPU. I will open a separate issue to discuss the CPU fallback behavior - I think we'll want to do a few changes eventually.


# This is fine on the default interface - why?
VideoDecoder(H265_VIDEO.path, device="cuda").get_frame_at(0)
# But it fails on the beta interface due to input validation checks, which we took from DALI!
with pytest.raises(
RuntimeError,
match="Video is too small in at least one dimension. Provided: 128x128 vs supported:144x144",
Expand Down
Loading