Add YUV420P Support to CudaDeviceInterface in TorchCodec

[mrezabayat]

Currently, CudaDeviceInterface::convertAVFrameToFrameOutput only supports CUDA AVFrames in NV12 format for GPU color conversion using NPP (nppiNV12ToRGB_* functions). If a decoded frame is in planar YUV420P format (e.g., from certain hardware decoders or non-8-bit content), the function raises an error due to the AV_PIX_FMT_NV12 check.

This limitation prevents GPU-accelerated color conversion for common formats like YUV420P, even though CUDA/NPP supports planar formats through functions like nppiYUV420ToRGB_*.

Proposed Change:

Extend convertAVFrameToFrameOutput to detect and handle AV_PIX_FMT_YUV420P in addition to NV12.

Implement the YUV420P → RGB conversion path using NPP (nppiYUV420ToRGB_8u_P3C3R_Ctx or _709CSC variants for BT.709).

Maintain the same caching logic for NppStreamContext and memory handling as in the NV12 path.

Benefits:

Broader format support for CUDA-based decoding pipelines.

Reduced need to fall back to CPU for planar YUV420P → RGB conversion, improving performance.

Improved interoperability with decoders that output YUV420P by default.

Motivation, pitch

No response

[NicolasHug]

Hi @mrezabayat , thanks for the feature request.

This is in scope and something we'll want to address eventually. Right now, the CUDA color-conversion logic is a bit messy and I'm working toward fixing it. I think this issue should largely be addressed by #776 (which is about 10bit support, but I think fixing 10-bit support overlaps with this).

I also suspect that #792 is already a step towards supporting YUV420P formats. Right now, we fallback to the CPU, but eventually we'll try to run the color-conversion step on the GPU (this is left as a TODO ATM).