Andreas Rheinhardt b91081274f avcodec/x86/h264_qpel: Add and use ff_{avg,put}_pixels16x16_l2_sse2() ...

This avoids mmx (the size 16 h264qpel dsp now no longer uses any mmx)
and improves performance, particularly for the avg case:

Old benchmarks:
avg_h264_qpel_16_mc01_8_c:                             780.0 ( 1.00x)
avg_h264_qpel_16_mc01_8_sse2:                           91.2 ( 8.55x)
avg_h264_qpel_16_mc03_8_c:                             804.0 ( 1.00x)
avg_h264_qpel_16_mc03_8_sse2:                           91.2 ( 8.82x)
put_h264_qpel_16_mc01_8_c:                             779.5 ( 1.00x)
put_h264_qpel_16_mc01_8_sse2:                           82.8 ( 9.41x)
put_h264_qpel_16_mc03_8_c:                             770.1 ( 1.00x)
put_h264_qpel_16_mc03_8_sse2:                           82.5 ( 9.33x)

New benchmarks:
avg_h264_qpel_16_mc01_8_c:                             783.9 ( 1.00x)
avg_h264_qpel_16_mc01_8_sse2:                           84.1 ( 9.32x)
avg_h264_qpel_16_mc03_8_c:                             797.4 ( 1.00x)
avg_h264_qpel_16_mc03_8_sse2:                           83.9 ( 9.51x)
put_h264_qpel_16_mc01_8_c:                             767.4 ( 1.00x)
put_h264_qpel_16_mc01_8_sse2:                           80.5 ( 9.53x)
put_h264_qpel_16_mc03_8_c:                             779.9 ( 1.00x)
put_h264_qpel_16_mc03_8_sse2:                           80.3 ( 9.71x)

(qpeldsp will use these functions when it gets ported to SSE2;
then the mmxext functions will be removed as well.)

Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>

2025-11-01 15:17:05 +01:00

2.0 KiB

Raw Blame History

View Raw