Skip to content

F
ffmpeg-mt
Commit 6b86daa0 authored by michaelni's avatar michaelni
Browse files
Options 

hopefully faster mmx2&3dnow MC


git-svn-id: file:///var/local/repositories/ffmpeg/trunk@506 9553f0bf-9b14-0410-a0b8-cfaf0461ba5b
parent 9be7ca5c
Expand all Show whitespace changes
Inline Side-by-side
Showing with 362 additions and 323 deletions
libavcodec/i386/dsputil_mmx.c
View file @ 6b86daa0
...@@ -21,6 +21,7 @@ ...@@ -21,6 +21,7 @@
#include "../dsputil.h" #include "../dsputil.h"
#include "../simple_idct.h" #include "../simple_idct.h"
#include "../mangle.h"
int mm_flags; /* multimedia extension flags */ int mm_flags; /* multimedia extension flags */
...@@ -49,6 +50,7 @@ void ff_mmx_idct(DCTELEM *block); ...@@ -49,6 +50,7 @@ void ff_mmx_idct(DCTELEM *block);
void ff_mmxext_idct(DCTELEM *block); void ff_mmxext_idct(DCTELEM *block);
/* pixel operations */ /* pixel operations */
static const unsigned long long int mm_bone __attribute__ ((aligned(8))) = 0x0101010101010101LL;
static const unsigned long long int mm_wone __attribute__ ((aligned(8))) = 0x0001000100010001LL; static const unsigned long long int mm_wone __attribute__ ((aligned(8))) = 0x0001000100010001LL;
static const unsigned long long int mm_wtwo __attribute__ ((aligned(8))) = 0x0002000200020002LL; static const unsigned long long int mm_wtwo __attribute__ ((aligned(8))) = 0x0002000200020002LL;
//static const unsigned short mm_wone[4] __attribute__ ((aligned(8))) = { 0x1, 0x1, 0x1, 0x1 }; //static const unsigned short mm_wone[4] __attribute__ ((aligned(8))) = { 0x1, 0x1, 0x1, 0x1 };
...@@ -90,7 +92,7 @@ static const unsigned long long int mm_wtwo __attribute__ ((aligned(8))) = 0x000 ...@@ -90,7 +92,7 @@ static const unsigned long long int mm_wtwo __attribute__ ((aligned(8))) = 0x000
/***********************************/ /***********************************/
/* MMX2 specific */ /* MMX2 specific */
#define DEF(x) x ## _sse #define DEF(x) x ## _mmx2
/* Introduced only in MMX2 set */ /* Introduced only in MMX2 set */
#define PAVGB "pavgb" #define PAVGB "pavgb"
...@@ -105,41 +107,38 @@ static const unsigned long long int mm_wtwo __attribute__ ((aligned(8))) = 0x000 ...@@ -105,41 +107,38 @@ static const unsigned long long int mm_wtwo __attribute__ ((aligned(8))) = 0x000
static void get_pixels_mmx(DCTELEM *block, const UINT8 *pixels, int line_size) static void get_pixels_mmx(DCTELEM *block, const UINT8 *pixels, int line_size)
{ {
DCTELEM *p; asm volatile(
const UINT8 *pix; "movl $-128, %%eax \n\t"
int i; "pxor %%mm7, %%mm7 \n\t"
".balign 16 \n\t"
/* read the pixels */ "1: \n\t"
p = block; "movq (%0), %%mm0 \n\t"
pix = pixels; "movq (%0, %2), %%mm2 \n\t"
MOVQ_ZERO(mm7); "movq %%mm0, %%mm1 \n\t"
for(i=0;i<4;i++) { "movq %%mm2, %%mm3 \n\t"
__asm __volatile( "punpcklbw %%mm7, %%mm0 \n\t"
"movq %1, %%mm0\n\t" "punpckhbw %%mm7, %%mm1 \n\t"
"movq %2, %%mm1\n\t" "punpcklbw %%mm7, %%mm2 \n\t"
"movq %%mm0, %%mm2\n\t" "punpckhbw %%mm7, %%mm3 \n\t"
"movq %%mm1, %%mm3\n\t" "movq %%mm0, (%1, %%eax)\n\t"
"punpcklbw %%mm7, %%mm0\n\t" "movq %%mm1, 8(%1, %%eax)\n\t"
"punpckhbw %%mm7, %%mm2\n\t" "movq %%mm2, 16(%1, %%eax)\n\t"
"punpcklbw %%mm7, %%mm1\n\t" "movq %%mm3, 24(%1, %%eax)\n\t"
"punpckhbw %%mm7, %%mm3\n\t" "addl %3, %0 \n\t"
"movq %%mm0, %0\n\t" "addl $32, %%eax \n\t"
"movq %%mm2, 8%0\n\t" "js 1b \n\t"
"movq %%mm1, 16%0\n\t" : "+r" (pixels)
"movq %%mm3, 24%0\n\t" : "r" (block+64), "r" (line_size), "r" (line_size*2)
:"=m"(*p) : "%eax"
:"m"(*pix), "m"(*(pix+line_size)) );
:"memory");
pix += line_size*2;
p += 16;
}
} }
static void diff_pixels_mmx(DCTELEM *block, const UINT8 *s1, const UINT8 *s2, int stride) static void diff_pixels_mmx(DCTELEM *block, const UINT8 *s1, const UINT8 *s2, int stride)
{ {
asm volatile( asm volatile(
".balign 16 \n\t" "pxor %%mm7, %%mm7 \n\t"
"movl $-128, %%eax \n\t" "movl $-128, %%eax \n\t"
".balign 16 \n\t"
"1: \n\t" "1: \n\t"
"movq (%0), %%mm0 \n\t" "movq (%0), %%mm0 \n\t"
"movq (%1), %%mm2 \n\t" "movq (%1), %%mm2 \n\t"
...@@ -261,56 +260,62 @@ static void add_pixels_clamped_mmx(const DCTELEM *block, UINT8 *pixels, int line ...@@ -261,56 +260,62 @@ static void add_pixels_clamped_mmx(const DCTELEM *block, UINT8 *pixels, int line
static void put_pixels_mmx(UINT8 *block, const UINT8 *pixels, int line_size, int h) static void put_pixels_mmx(UINT8 *block, const UINT8 *pixels, int line_size, int h)
{ {
int hh; #if 0 //FIXME h==4 case
UINT8 *p; asm volatile(
const UINT8 *pix; "xorl %%eax, %%eax \n\t"
"movl %3, %%esi \n\t"
p = block; "1: \n\t"
pix = pixels; // 2s "movq (%1, %%eax), %%mm0 \n\t"
#if 0 "movq %%mm0, (%0, %%eax) \n\t"
do { "addl %2, %%eax \n\t"
__asm __volatile( "movq (%1, %%eax), %%mm0 \n\t"
"movq %1, %%mm0\n\t" "movq %%mm0, (%0, %%eax) \n\t"
"movq %%mm0, %0\n\t" "addl %2, %%eax \n\t"
:"=m"(*p) "movq (%1, %%eax), %%mm0 \n\t"
:"m"(*pix) "movq %%mm0, (%0, %%eax) \n\t"
:"memory"); "addl %2, %%eax \n\t"
pix += line_size; "movq (%1, %%eax), %%mm0 \n\t"
p += line_size; "movq %%mm0, (%0, %%eax) \n\t"
} while (--h); "addl %2, %%eax \n\t"
"movq (%1, %%eax), %%mm0 \n\t"
"movq %%mm0, (%0, %%eax) \n\t"
"addl %2, %%eax \n\t"
"movq (%1, %%eax), %%mm0 \n\t"
"movq %%mm0, (%0, %%eax) \n\t"
"addl %2, %%eax \n\t"
"movq (%1, %%eax), %%mm0 \n\t"
"movq %%mm0, (%0, %%eax) \n\t"
"addl %2, %%eax \n\t"
"movq (%1, %%eax), %%mm0 \n\t"
"movq %%mm0, (%0, %%eax) \n\t"
"addl %2, %%eax \n\t"
"subl $8, %%esi \n\t"
" jnz 1b \n\t"
:: "r" (block), "r" (pixels), "r"(line_size), "m"(h)
: "%eax", "%esi", "memory"
);
#else #else
// this optimized code is not very usefull asm volatile(
// the above loop is definitely faster "xorl %%eax, %%eax \n\t"
// at least on Celeron 500MHz "movl %3, %%esi \n\t"
hh = h & 3; "1: \n\t"
while (hh) { "movq (%1, %%eax), %%mm0 \n\t"
__asm __volatile( "movq %%mm0, (%0, %%eax) \n\t"
"movq %1, %%mm0\n\t" "addl %2, %%eax \n\t"
"movq %%mm0, %0\n\t" "movq (%1, %%eax), %%mm0 \n\t"
:"=m"(*p) "movq %%mm0, (%0, %%eax) \n\t"
:"m"(*pix) "addl %2, %%eax \n\t"
:"memory"); "movq (%1, %%eax), %%mm0 \n\t"
pix += line_size; "movq %%mm0, (%0, %%eax) \n\t"
p += line_size; "addl %2, %%eax \n\t"
hh--; "movq (%1, %%eax), %%mm0 \n\t"
} "movq %%mm0, (%0, %%eax) \n\t"
hh=h>>2; "addl %2, %%eax \n\t"
while (hh) { "subl $4, %%esi \n\t"
__asm __volatile( " jnz 1b \n\t"
"movq (%1), %%mm0 \n\t" :: "r" (block), "r" (pixels), "r"(line_size), "m"(h)
"movq (%1, %2), %%mm1 \n\t" : "%eax", "%esi", "memory"
"movq (%1, %2, 2), %%mm2 \n\t" );
"movq (%1, %3), %%mm3 \n\t"
"movq %%mm0, (%0) \n\t"
"movq %%mm1, (%0, %2) \n\t"
"movq %%mm2, (%0, %2, 2) \n\t"
"movq %%mm3, (%0, %3) \n\t"
::"r"(p), "r"(pix), "r"(line_size), "r"(line_size*3)
:"memory");
pix += line_size*4;
p += line_size*4;
hh--;
}
#endif #endif
} }
...@@ -1141,19 +1146,23 @@ void dsputil_init_mmx(void) ...@@ -1141,19 +1146,23 @@ void dsputil_init_mmx(void)
pix_abs8x8_y2 = pix_abs8x8_y2_mmx2; pix_abs8x8_y2 = pix_abs8x8_y2_mmx2;
pix_abs8x8_xy2= pix_abs8x8_xy2_mmx2; pix_abs8x8_xy2= pix_abs8x8_xy2_mmx2;
put_pixels_tab[1] = put_pixels_x2_sse; put_pixels_tab[1] = put_pixels_x2_mmx2;
put_pixels_tab[2] = put_pixels_y2_sse; put_pixels_tab[2] = put_pixels_y2_mmx2;
put_no_rnd_pixels_tab[1] = put_no_rnd_pixels_x2_mmx2;
put_no_rnd_pixels_tab[2] = put_no_rnd_pixels_y2_mmx2;
avg_pixels_tab[0] = avg_pixels_sse; avg_pixels_tab[0] = avg_pixels_mmx2;
avg_pixels_tab[1] = avg_pixels_x2_sse; avg_pixels_tab[1] = avg_pixels_x2_mmx2;
avg_pixels_tab[2] = avg_pixels_y2_sse; avg_pixels_tab[2] = avg_pixels_y2_mmx2;
avg_pixels_tab[3] = avg_pixels_xy2_sse; avg_pixels_tab[3] = avg_pixels_xy2_mmx2;
sub_pixels_tab[1] = sub_pixels_x2_sse; sub_pixels_tab[1] = sub_pixels_x2_mmx2;
sub_pixels_tab[2] = sub_pixels_y2_sse; sub_pixels_tab[2] = sub_pixels_y2_mmx2;
} else if (mm_flags & MM_3DNOW) { } else if (mm_flags & MM_3DNOW) {
put_pixels_tab[1] = put_pixels_x2_3dnow; put_pixels_tab[1] = put_pixels_x2_3dnow;
put_pixels_tab[2] = put_pixels_y2_3dnow; put_pixels_tab[2] = put_pixels_y2_3dnow;
put_no_rnd_pixels_tab[1] = put_no_rnd_pixels_x2_3dnow;
put_no_rnd_pixels_tab[2] = put_no_rnd_pixels_y2_3dnow;
avg_pixels_tab[0] = avg_pixels_3dnow; avg_pixels_tab[0] = avg_pixels_3dnow;
avg_pixels_tab[1] = avg_pixels_x2_3dnow; avg_pixels_tab[1] = avg_pixels_x2_3dnow;
......
libavcodec/i386/dsputil_mmx_avg.h
View file @ 6b86daa0
This diff is collapsed.
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment