From e80cf125a7df1c19f0228fb4a520ec577ae0be99 Mon Sep 17 00:00:00 2001 From: Michael Niedermayer Date: Fri, 3 Nov 2006 15:40:57 +0000 Subject: 2 instructions less (same speed) Originally committed as revision 6888 to svn://svn.ffmpeg.org/ffmpeg/trunk --- libavcodec/i386/h264dsp_mmx.c | 19 +++++++++++++------ 1 file changed, 13 insertions(+), 6 deletions(-) (limited to 'libavcodec') diff --git a/libavcodec/i386/h264dsp_mmx.c b/libavcodec/i386/h264dsp_mmx.c index a8fc46b2d1..e73aa9c949 100644 --- a/libavcodec/i386/h264dsp_mmx.c +++ b/libavcodec/i386/h264dsp_mmx.c @@ -317,6 +317,17 @@ static void ff_h264_idct8_dc_add_mmx2(uint8_t *dst, int16_t *block, int stride) "por "#t", "#o" \n\t"\ "psubusb "#a", "#o" \n\t" +// out: o = |x-y|>a +// clobbers: t +#define DIFF_GT2_MMX(x,y,a,o,t)\ + "movq "#y", "#t" \n\t"\ + "movq "#x", "#o" \n\t"\ + "psubusb "#x", "#t" \n\t"\ + "psubusb "#y", "#o" \n\t"\ + "psubusb "#a", "#t" \n\t"\ + "psubusb "#a", "#o" \n\t"\ + "pcmpeqb "#t", "#o" \n\t"\ + // in: mm0=p1 mm1=p0 mm2=q0 mm3=q1 // out: mm5=beta-1, mm7=mask // clobbers: mm4,mm6 @@ -398,9 +409,7 @@ static inline void h264_loop_filter_luma_mmx2(uint8_t *pix, int stride, int alph /* filter p1 */ "movq (%1), %%mm3 \n\t" //p2 - DIFF_GT_MMX(%%mm1, %%mm3, %%mm5, %%mm6, %%mm4) // |p2-p0|>beta-1 - "pandn %%mm7, %%mm6 \n\t" - "pcmpeqb %%mm7, %%mm6 \n\t" + DIFF_GT2_MMX(%%mm1, %%mm3, %%mm5, %%mm6, %%mm4) // |p2-p0|>beta-1 "pand %%mm7, %%mm6 \n\t" // mask & |p2-p0|beta-1 - "pandn %0, %%mm6 \n\t" - "pcmpeqb %0, %%mm6 \n\t" + DIFF_GT2_MMX(%%mm2, %%mm4, %%mm5, %%mm6, %%mm3) // |q2-q0|>beta-1 "pand %0, %%mm6 \n\t" "movq 8+%0, %%mm5 \n\t" // can be merged with the and below but is slower then "pand %%mm6, %%mm5 \n\t" -- cgit v1.2.3