Chiebot-Mirror
/
FFmpeg
mirror of https://github.com/FFmpeg/FFmpeg.git
8
0
Fork 0
Code Issues Projects Releases Wiki Activity

rewrite H264_CHROMA_MC4_TMPL (20% faster)

Browse Source 
Originally committed as revision 8012 to svn://svn.ffmpeg.org/ffmpeg/trunk
pull/126/head
Michael Niedermayer 18 years ago
parent aeb292227e
commit b21e0b6dfc
1 changed files with 61 additions and 73 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 134
      libavcodec/i386/dsputil_h264_template_mmx.c

134
libavcodec/i386/dsputil_h264_template_mmx.c
Unescape View File

@ -184,82 +184,70 @@ static void H264_CHROMA_MC8_TMPL(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*
static void H264_CHROMA_MC4_TMPL(uint8_t *dst/*align 4*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y) static void H264_CHROMA_MC4_TMPL(uint8_t *dst/*align 4*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y)
{ {
DECLARE_ALIGNED_8(uint64_t, AA);
DECLARE_ALIGNED_8(uint64_t, DD);
int i;
/* no special case for mv=(0,0) in 4x*, since it's much less common than in 8x*.
* could still save a few cycles, but maybe not worth the complexity. */
assert(x<8 && y<8 && x>=0 && y>=0);
asm volatile("movd %2, %%mm4\n\t"
"movd %3, %%mm6\n\t"
"punpcklwd %%mm4, %%mm4\n\t"
"punpcklwd %%mm6, %%mm6\n\t"
"punpckldq %%mm4, %%mm4\n\t" /* mm4 = x words */
"punpckldq %%mm6, %%mm6\n\t" /* mm6 = y words */
"movq %%mm4, %%mm5\n\t"
"pmullw %%mm6, %%mm4\n\t" /* mm4 = x * y */
"psllw $3, %%mm5\n\t"
"psllw $3, %%mm6\n\t"
"movq %%mm5, %%mm7\n\t"
"paddw %%mm6, %%mm7\n\t"
"movq %%mm4, %1\n\t" /* DD = x * y */
"psubw %%mm4, %%mm5\n\t" /* mm5 = B = 8x - xy */
"psubw %%mm4, %%mm6\n\t" /* mm6 = C = 8y - xy */
"paddw %4, %%mm4\n\t"
"psubw %%mm7, %%mm4\n\t" /* mm4 = A = xy - (8x+8y) + 64 */
"pxor %%mm7, %%mm7\n\t"
"movq %%mm4, %0\n\t"
: "=m" (AA), "=m" (DD) : "rm" (x), "rm" (y), "m" (ff_pw_64));
asm volatile( asm volatile(
/* mm0 = src[0..3], mm1 = src[1..4] */ "pxor %%mm7, %%mm7 \n\t"
"movd %0, %%mm0\n\t" "movd %5, %%mm2 \n\t"
"movd %1, %%mm1\n\t" "movd %6, %%mm3 \n\t"
"punpcklbw %%mm7, %%mm0\n\t" "movq %7, %%mm4 \n\t"
"punpcklbw %%mm7, %%mm1\n\t" "movq %7, %%mm5 \n\t"
: : "m" (src[0]), "m" (src[1])); "punpcklwd %%mm2, %%mm2 \n\t"
"punpcklwd %%mm3, %%mm3 \n\t"
"punpcklwd %%mm2, %%mm2 \n\t"
"punpcklwd %%mm3, %%mm3 \n\t"
"psubw %%mm2, %%mm4 \n\t"
"psubw %%mm3, %%mm5 \n\t"
for(i=0; i<h; i++) { "movd (%1), %%mm0 \n\t"
asm volatile( "movd 1(%1), %%mm6 \n\t"
/* mm2 = A * src[0..3] + B * src[1..4] */ "add %3, %1 \n\t"
"movq %%mm0, %%mm2\n\t" "punpcklbw %%mm7, %%mm0 \n\t"
"pmullw %0, %%mm2\n\t" "punpcklbw %%mm7, %%mm6 \n\t"
"pmullw %%mm5, %%mm1\n\t" "pmullw %%mm4, %%mm0 \n\t"
"paddw %%mm1, %%mm2\n\t" "pmullw %%mm2, %%mm6 \n\t"
: : "m" (AA)); "paddw %%mm0, %%mm6 \n\t"
src += stride; "1: \n\t"
asm volatile( "movd (%1), %%mm0 \n\t"
/* mm0 = src[0..3], mm1 = src[1..4] */ "movd 1(%1), %%mm1 \n\t"
"movd %0, %%mm0\n\t" "add %3, %1 \n\t"
"movd %1, %%mm1\n\t" "punpcklbw %%mm7, %%mm0 \n\t"
"punpcklbw %%mm7, %%mm0\n\t" "punpcklbw %%mm7, %%mm1 \n\t"
"punpcklbw %%mm7, %%mm1\n\t" "pmullw %%mm4, %%mm0 \n\t"
: : "m" (src[0]), "m" (src[1])); "pmullw %%mm2, %%mm1 \n\t"
"paddw %%mm0, %%mm1 \n\t"
asm volatile( "movq %%mm1, %%mm0 \n\t"
/* mm2 += C * src[0..3] + D * src[1..4] */ "pmullw %%mm5, %%mm6 \n\t"
"movq %%mm0, %%mm3\n\t" "pmullw %%mm3, %%mm1 \n\t"
"movq %%mm1, %%mm4\n\t" "paddw %%mm6, %%mm1 \n\t"
"pmullw %%mm6, %%mm3\n\t" "paddw %4, %%mm1 \n\t"
"pmullw %0, %%mm4\n\t" "psrlw $6, %%mm1 \n\t"
"paddw %%mm3, %%mm2\n\t" "packuswb %%mm1, %%mm1 \n\t"
"paddw %%mm4, %%mm2\n\t" H264_CHROMA_OP4((%0), %%mm1, %%mm6)
: : "m" (DD)); "movd %%mm1, (%0) \n\t"
"add %3, %0 \n\t"
asm volatile( "movd (%1), %%mm6 \n\t"
/* dst[0..3] = pack((mm2 + 32) >> 6) */ "movd 1(%1), %%mm1 \n\t"
"paddw %1, %%mm2\n\t" "add %3, %1 \n\t"
"psrlw $6, %%mm2\n\t" "punpcklbw %%mm7, %%mm6 \n\t"
"packuswb %%mm7, %%mm2\n\t" "punpcklbw %%mm7, %%mm1 \n\t"
H264_CHROMA_OP4(%0, %%mm2, %%mm3) "pmullw %%mm4, %%mm6 \n\t"
"movd %%mm2, %0\n\t" "pmullw %%mm2, %%mm1 \n\t"
: "=m" (dst[0]) : "m" (ff_pw_32)); "paddw %%mm6, %%mm1 \n\t"
dst += stride; "movq %%mm1, %%mm6 \n\t"
} "pmullw %%mm5, %%mm0 \n\t"
"pmullw %%mm3, %%mm1 \n\t"
"paddw %%mm0, %%mm1 \n\t"
"paddw %4, %%mm1 \n\t"
"psrlw $6, %%mm1 \n\t"
"packuswb %%mm1, %%mm1 \n\t"
H264_CHROMA_OP4((%0), %%mm1, %%mm0)
"movd %%mm1, (%0) \n\t"
"add %3, %0 \n\t"
"sub $2, %2 \n\t"
"jnz 1b \n\t"
: "+r"(dst), "+r"(src), "+r"(h)
: "r"(stride), "m"(ff_pw_32), "m"(x), "m"(y), "m"(ff_pw_8)
);
} }
#ifdef H264_CHROMA_MC2_TMPL #ifdef H264_CHROMA_MC2_TMPL

Write Preview
Loading…
Cancel
Save