Hi On 18 June 2014 13:22, Xiang, Haihao <[email protected]> wrote: > > Yes, The buffer for vaGetImage() will be mapped to > USWC memory too after applying the patch. >
Ok.. here are my results. I've timed the following procedure: vaSyncSurface(); vaGetImage(); vaMapBuffer(); copy(dst, src); vaUnMapBuffer(); vaDestroyImage() This is run within MythTV, so in actual use, with an actual video. This is how long it takes for doing so on 1000 frames, in nanoseconds: h264 592x336 video. original: 854922255ns (0.85ms/frame) patched: 4389699017ns (4.38ms/frame) 5.13x slower mpeg2 1440x1080 original: 3936736940ns (3.93ms/frame) patched: 13202583358ns (13.20ms/frame) 3.35x slower Without the patch, the copy routine detects that using memcpy is faster than using SSE4 code and as such copy uses memcpy. With the patch, copy uses SSE4 code as it's faster than memcpy. So this indeed confirms it's USWC memory So in all... big regression speed-wise JY _______________________________________________ Libva mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/libva
