On 03.07.2025 23:05, Jorge Estrada wrote:
ping

Completely forgot about this, sorry.

Patch does not compile, cause vector_types.h does not exist.
If I remove it, it can't find uchar2, which is purely a CUDA side type.

Isn't that whole construct around there entirely unneeded?
You can always just pass &s->parsed_color[plane] in place of p_fill_val, and it'll always be correct.


That re-invented RGB to YUV conversion is also entirely unneccesary.
There's macros for that in colorspace.h


Also, there's A LOT of use of mixed declarations and code, though I think that is fine nowadays, and the warnings about it have since been disabled.
So it's probably fine to leave it as-is.


There's also a bit of potential to further generalize and micro-optimize the CUDA kernel and calling it. Primarily moving the one extra division out of the kernel, and doing it in advance for it. (There's also a tiny superfluous section that sets hsub/vsub to the same value again if it's NV12, which I removed).
See the second commit for what I mean.

I've pushed the necessary fixes and further optimization and generalization to GitHub:
https://github.com/BtbN/FFmpeg
https://github.com/BtbN/FFmpeg/commit/45e4baa3d82c6e6fffbd6c1ac8389e305a3f9137
https://github.com/BtbN/FFmpeg/commit/14cf522b51e66e65ca60e0447348f2bfbcf631ea


Works fine for me like that, if you're fine with those as well, I'll squash them into the main commit and merge it.
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".

Reply via email to