Am Mi., 20. Mai 2020 um 20:36 Uhr schrieb Alex <[email protected]>: > > I've been working on implementing hardware decoder integration. While > things generally are working, there's not much gain: the color > conversion from NV12 seems to be dramatically more expensive than from > the YUV420P in case of software decoder. > > When profiling, I'm seeing YUV420P->RGB24 using yuv420_rgb24_mmxext > (which barely registers in profiler), while NV12->RGB24 using > yuv2rgb24_X_c, which takes a fairly dramatic amount of time. > > Is that something that is expected? A known problem? Or perhaps > something is wrong on my side of things?
Very, very generally, hardware decoding is very useful if you want to display on the same gpu where the decoding takes place and if you want to encode on the same gpu where decoding happens. If you have to download the decoded frame from graphics memory to cpu memory, some of the possible gains are always lost. If you - in addition - need a colour space conversion (in your case, a conversion that rarely happens when using a software decoder which is therefore not so well accelerated), even more time is lost. That being said: Did you try to first convert nv12->yuv420p using ff_nv12ToUV_avx, then using above yuv420_rgb24_mmxext ? I would have expected nv12ToUV to be much faster than any yuv2rgb conversion. Carl Eugen _______________________________________________ Libav-user mailing list [email protected] https://ffmpeg.org/mailman/listinfo/libav-user To unsubscribe, visit link above, or email [email protected] with subject "unsubscribe".
