hopefully faster mmx2&3dnow MC

Originally committed as revision 506 to svn://svn.ffmpeg.org/ffmpeg/trunk
This commit is contained in:
Michael Niedermayer 2002-05-17 01:04:14 +00:00
parent 59fe111e81
commit 607dce96c0
2 changed files with 363 additions and 324 deletions

View File

@ -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);
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
/***********************************/ /***********************************/
/* 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
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
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
} }
@ -1124,7 +1129,7 @@ void dsputil_init_mmx(void)
avg_no_rnd_pixels_tab[1] = avg_no_rnd_pixels_x2_mmx; avg_no_rnd_pixels_tab[1] = avg_no_rnd_pixels_x2_mmx;
avg_no_rnd_pixels_tab[2] = avg_no_rnd_pixels_y2_mmx; avg_no_rnd_pixels_tab[2] = avg_no_rnd_pixels_y2_mmx;
avg_no_rnd_pixels_tab[3] = avg_no_rnd_pixels_xy2_mmx; avg_no_rnd_pixels_tab[3] = avg_no_rnd_pixels_xy2_mmx;
sub_pixels_tab[0] = sub_pixels_mmx; sub_pixels_tab[0] = sub_pixels_mmx;
sub_pixels_tab[1] = sub_pixels_x2_mmx; sub_pixels_tab[1] = sub_pixels_x2_mmx;
sub_pixels_tab[2] = sub_pixels_y2_mmx; sub_pixels_tab[2] = sub_pixels_y2_mmx;
@ -1140,20 +1145,24 @@ void dsputil_init_mmx(void)
pix_abs8x8_x2 = pix_abs8x8_x2_mmx2; pix_abs8x8_x2 = pix_abs8x8_x2_mmx2;
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[2] = put_pixels_y2_sse;
avg_pixels_tab[0] = avg_pixels_sse;
avg_pixels_tab[1] = avg_pixels_x2_sse;
avg_pixels_tab[2] = avg_pixels_y2_sse;
avg_pixels_tab[3] = avg_pixels_xy2_sse;
sub_pixels_tab[1] = sub_pixels_x2_sse; put_pixels_tab[1] = put_pixels_x2_mmx2;
sub_pixels_tab[2] = sub_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_mmx2;
avg_pixels_tab[1] = avg_pixels_x2_mmx2;
avg_pixels_tab[2] = avg_pixels_y2_mmx2;
avg_pixels_tab[3] = avg_pixels_xy2_mmx2;
sub_pixels_tab[1] = sub_pixels_x2_mmx2;
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;

View File

@ -1,6 +1,7 @@
/* /*
* DSP utils : average functions are compiled twice for 3dnow/mmx2 * DSP utils : average functions are compiled twice for 3dnow/mmx2
* Copyright (c) 2000, 2001 Gerard Lantau. * Copyright (c) 2000, 2001 Gerard Lantau.
* Copyright (c) 2002 Michael Niedermayer
* *
* This program is free software; you can redistribute it and/or modify * This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by * it under the terms of the GNU General Public License as published by
@ -17,271 +18,300 @@
* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
* *
* MMX optimization by Nick Kurshev <nickols_k@mail.ru> * MMX optimization by Nick Kurshev <nickols_k@mail.ru>
* mostly rewritten by Michael Niedermayer <michaelni@gmx.at>
*/ */
static void DEF(put_pixels_x2)(UINT8 *block, const UINT8 *pixels, int line_size, int h) static void DEF(put_pixels_x2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
{ {
int dh, hh;
UINT8 *p;
const UINT8 *pix;
p = block;
pix = pixels;
hh=h>>2;
dh=h&3;
while(hh--) {
__asm __volatile( __asm __volatile(
"movq (%1), %%mm0\n\t" "xorl %%eax, %%eax \n\t"
"movq 1(%1), %%mm1\n\t" ".balign 16 \n\t"
"movq (%1, %2), %%mm2\n\t" "1: \n\t"
"movq 1(%1, %2), %%mm3\n\t" "movq (%1, %%eax), %%mm0 \n\t"
"movq (%1, %2, 2), %%mm4\n\t" "movq 1(%1, %%eax), %%mm1 \n\t"
"movq 1(%1, %2, 2), %%mm5\n\t" "movq (%2, %%eax), %%mm2 \n\t"
"movq (%1, %3), %%mm6\n\t" "movq 1(%2, %%eax), %%mm3 \n\t"
"movq 1(%1, %3), %%mm7\n\t" PAVGB" %%mm1, %%mm0 \n\t"
PAVGB" %%mm1, %%mm0\n\t" PAVGB" %%mm3, %%mm2 \n\t"
PAVGB" %%mm3, %%mm2\n\t" "movq %%mm0, (%3, %%eax) \n\t"
PAVGB" %%mm5, %%mm4\n\t" "movq %%mm2, (%4, %%eax) \n\t"
PAVGB" %%mm7, %%mm6\n\t" "addl %5, %%eax \n\t"
"movq %%mm0, (%0)\n\t" "movq (%1, %%eax), %%mm0 \n\t"
"movq %%mm2, (%0, %2)\n\t" "movq 1(%1, %%eax), %%mm1 \n\t"
"movq %%mm4, (%0, %2, 2)\n\t" "movq (%2, %%eax), %%mm2 \n\t"
"movq %%mm6, (%0, %3)\n\t" "movq 1(%2, %%eax), %%mm3 \n\t"
::"r"(p), "r"(pix), "r" (line_size), "r" (line_size*3) PAVGB" %%mm1, %%mm0 \n\t"
:"memory"); PAVGB" %%mm3, %%mm2 \n\t"
pix += line_size*4; p += line_size*4; "movq %%mm0, (%3, %%eax) \n\t"
} "movq %%mm2, (%4, %%eax) \n\t"
while(dh--) { "addl %5, %%eax \n\t"
"subl $4, %0 \n\t"
" jnz 1b \n\t"
:"+g"(h)
:"r"(pixels), "r"(pixels+line_size), "r" (block), "r" (block+line_size),
"r"(line_size<<1)
:"%eax", "memory");
}
static void DEF(put_no_rnd_pixels_x2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
{
__asm __volatile( __asm __volatile(
"movq %1, %%mm0\n\t" "xorl %%eax, %%eax \n\t"
"movq 1%1, %%mm1\n\t" "movq "MANGLE(mm_bone)", %%mm7 \n\t"
PAVGB" %%mm1, %%mm0\n\t" ".balign 16 \n\t"
"movq %%mm0, %0\n\t" "1: \n\t"
:"=m"(*p) "movq (%1, %%eax), %%mm0 \n\t"
:"m"(*pix) "movq 1(%1, %%eax), %%mm1 \n\t"
:"memory"); "movq (%2, %%eax), %%mm2 \n\t"
pix += line_size; p += line_size; "movq 1(%2, %%eax), %%mm3 \n\t"
} "psubusb %%mm7, %%mm0 \n\t"
"psubusb %%mm7, %%mm2 \n\t"
PAVGB" %%mm1, %%mm0 \n\t"
PAVGB" %%mm3, %%mm2 \n\t"
"movq %%mm0, (%3, %%eax) \n\t"
"movq %%mm2, (%4, %%eax) \n\t"
"addl %5, %%eax \n\t"
"movq (%1, %%eax), %%mm0 \n\t"
"movq 1(%1, %%eax), %%mm1 \n\t"
"movq (%2, %%eax), %%mm2 \n\t"
"movq 1(%2, %%eax), %%mm3 \n\t"
"psubusb %%mm7, %%mm0 \n\t"
"psubusb %%mm7, %%mm2 \n\t"
PAVGB" %%mm1, %%mm0 \n\t"
PAVGB" %%mm3, %%mm2 \n\t"
"movq %%mm0, (%3, %%eax) \n\t"
"movq %%mm2, (%4, %%eax) \n\t"
"addl %5, %%eax \n\t"
"subl $4, %0 \n\t"
" jnz 1b \n\t"
:"+g"(h)
:"r"(pixels), "r"(pixels+line_size), "r" (block), "r" (block+line_size),
"r"(line_size<<1)
:"%eax", "memory");
} }
static void DEF(put_pixels_y2)(UINT8 *block, const UINT8 *pixels, int line_size, int h) static void DEF(put_pixels_y2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
{ {
int dh, hh; __asm __volatile(
UINT8 *p; "xorl %%eax, %%eax \n\t"
const UINT8 *pix; "movq (%1), %%mm0 \n\t"
p = block; ".balign 16 \n\t"
pix = pixels; "1: \n\t"
"movq (%2, %%eax), %%mm1 \n\t"
"movq (%3, %%eax), %%mm2 \n\t"
PAVGB" %%mm1, %%mm0 \n\t"
PAVGB" %%mm2, %%mm1 \n\t"
"movq %%mm0, (%4, %%eax) \n\t"
"movq %%mm1, (%5, %%eax) \n\t"
"addl %6, %%eax \n\t"
"movq (%2, %%eax), %%mm1 \n\t"
"movq (%3, %%eax), %%mm0 \n\t"
PAVGB" %%mm1, %%mm2 \n\t"
PAVGB" %%mm0, %%mm1 \n\t"
"movq %%mm2, (%4, %%eax) \n\t"
"movq %%mm1, (%5, %%eax) \n\t"
"addl %6, %%eax \n\t"
"subl $4, %0 \n\t"
" jnz 1b \n\t"
:"+g"(h)
:"r"(pixels), "r"(pixels+line_size), "r"(pixels+line_size*2), "r" (block),
"r" (block+line_size), "g"(line_size<<1)
:"%eax", "memory");
}
hh=h>>1; static void DEF(put_no_rnd_pixels_y2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
dh=h&1; {
while(hh--) {
__asm __volatile( __asm __volatile(
"movq %2, %%mm0\n\t" "movq "MANGLE(mm_bone)", %%mm7 \n\t"
"movq %3, %%mm1\n\t" "xorl %%eax, %%eax \n\t"
"movq %4, %%mm2\n\t" "movq (%1), %%mm0 \n\t"
PAVGB" %%mm1, %%mm0\n\t" ".balign 16 \n\t"
PAVGB" %%mm2, %%mm1\n\t" "1: \n\t"
"movq %%mm0, %0\n\t" "movq (%2, %%eax), %%mm1 \n\t"
"movq %%mm1, %1\n\t" "movq (%3, %%eax), %%mm2 \n\t"
:"=m"(*p), "=m"(*(p+line_size)) "psubusb %%mm7, %%mm1 \n\t"
:"m"(*pix), "m"(*(pix+line_size)), PAVGB" %%mm1, %%mm0 \n\t"
"m"(*(pix+line_size*2)) PAVGB" %%mm2, %%mm1 \n\t"
:"memory"); "movq %%mm0, (%4, %%eax) \n\t"
pix += line_size*2; "movq %%mm1, (%5, %%eax) \n\t"
p += line_size*2; "addl %6, %%eax \n\t"
} "movq (%2, %%eax), %%mm1 \n\t"
if(dh) { "movq (%3, %%eax), %%mm0 \n\t"
__asm __volatile( "psubusb %%mm7, %%mm1 \n\t"
"movq %1, %%mm0\n\t" PAVGB" %%mm1, %%mm2 \n\t"
"movq %2, %%mm1\n\t" PAVGB" %%mm0, %%mm1 \n\t"
PAVGB" %%mm1, %%mm0\n\t" "movq %%mm2, (%4, %%eax) \n\t"
"movq %%mm0, %0\n\t" "movq %%mm1, (%5, %%eax) \n\t"
:"=m"(*p) "addl %6, %%eax \n\t"
:"m"(*pix), "subl $4, %0 \n\t"
"m"(*(pix+line_size)) " jnz 1b \n\t"
:"memory"); :"+g"(h)
} :"r"(pixels), "r"(pixels+line_size), "r"(pixels+line_size*2), "r" (block),
"r" (block+line_size), "g"(line_size<<1)
:"%eax", "memory");
} }
static void DEF(avg_pixels)(UINT8 *block, const UINT8 *pixels, int line_size, int h) static void DEF(avg_pixels)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
{ {
int dh, hh;
UINT8 *p;
const UINT8 *pix;
p = block;
pix = pixels;
hh=h>>2;
dh=h&3;
while(hh--) {
__asm __volatile( __asm __volatile(
"movq (%0), %%mm0\n\t" "xorl %%eax, %%eax \n\t"
"movq (%1), %%mm1\n\t" ".balign 16 \n\t"
"movq (%0, %2), %%mm2\n\t" "1: \n\t"
"movq (%1, %2), %%mm3\n\t" "movq (%1, %%eax), %%mm0 \n\t"
"movq (%0, %2, 2), %%mm4\n\t" "movq (%2, %%eax), %%mm2 \n\t"
"movq (%1, %2, 2), %%mm5\n\t" "movq (%3, %%eax), %%mm3 \n\t"
"movq (%0, %3), %%mm6\n\t" "movq (%4, %%eax), %%mm4 \n\t"
"movq (%1, %3), %%mm7\n\t" PAVGB" %%mm3, %%mm0 \n\t"
PAVGB" %%mm1, %%mm0\n\t" PAVGB" %%mm4, %%mm2 \n\t"
PAVGB" %%mm3, %%mm2\n\t" "movq %%mm0, (%3, %%eax) \n\t"
PAVGB" %%mm5, %%mm4\n\t" "movq %%mm2, (%4, %%eax) \n\t"
PAVGB" %%mm7, %%mm6\n\t" "addl %5, %%eax \n\t"
"movq %%mm0, (%0)\n\t" "movq (%1, %%eax), %%mm0 \n\t"
"movq %%mm2, (%0, %2)\n\t" "movq (%2, %%eax), %%mm2 \n\t"
"movq %%mm4, (%0, %2, 2)\n\t" "movq (%3, %%eax), %%mm3 \n\t"
"movq %%mm6, (%0, %3)\n\t" "movq (%4, %%eax), %%mm4 \n\t"
::"r"(p), "r"(pix), "r" (line_size), "r" (line_size*3) PAVGB" %%mm3, %%mm0 \n\t"
:"memory"); PAVGB" %%mm4, %%mm2 \n\t"
pix += line_size*4; p += line_size*4; "movq %%mm0, (%3, %%eax) \n\t"
} "movq %%mm2, (%4, %%eax) \n\t"
while(dh--) { "addl %5, %%eax \n\t"
__asm __volatile( "subl $4, %0 \n\t"
"movq %0, %%mm0\n\t" " jnz 1b \n\t"
"movq %1, %%mm1\n\t" :"+g"(h)
PAVGB" %%mm1, %%mm0\n\t" :"r"(pixels), "r"(pixels+line_size), "r" (block), "r" (block+line_size),
"movq %%mm0, %0\n\t" "r"(line_size<<1)
:"+m"(*p) :"%eax", "memory");
:"m"(*pix)
:"memory");
pix += line_size; p += line_size;
}
} }
static void DEF(avg_pixels_x2)( UINT8 *block, const UINT8 *pixels, int line_size, int h) static void DEF(avg_pixels_x2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
{ {
int dh, hh;
UINT8 *p;
const UINT8 *pix;
p = block;
pix = pixels;
hh=h>>1;
dh=h&1;
while(hh--) {
__asm __volatile( __asm __volatile(
"movq %2, %%mm2\n\t" "xorl %%eax, %%eax \n\t"
"movq 1%2, %%mm3\n\t" ".balign 16 \n\t"
"movq %3, %%mm4\n\t" "1: \n\t"
"movq 1%3, %%mm5\n\t" "movq (%1, %%eax), %%mm0 \n\t"
"movq %0, %%mm0\n\t" "movq 1(%1, %%eax), %%mm1 \n\t"
"movq %1, %%mm1\n\t" "movq (%2, %%eax), %%mm2 \n\t"
PAVGB" %%mm3, %%mm2\n\t" "movq 1(%2, %%eax), %%mm3 \n\t"
PAVGB" %%mm2, %%mm0\n\t" PAVGB" %%mm1, %%mm0 \n\t"
PAVGB" %%mm5, %%mm4\n\t" PAVGB" %%mm3, %%mm2 \n\t"
PAVGB" %%mm4, %%mm1\n\t" "movq (%3, %%eax), %%mm3 \n\t"
"movq %%mm0, %0\n\t" "movq (%4, %%eax), %%mm4 \n\t"
"movq %%mm1, %1\n\t" PAVGB" %%mm3, %%mm0 \n\t"
:"+m"(*p), "+m"(*(p+line_size)) PAVGB" %%mm4, %%mm2 \n\t"
:"m"(*pix), "m"(*(pix+line_size)) "movq %%mm0, (%3, %%eax) \n\t"
:"memory"); "movq %%mm2, (%4, %%eax) \n\t"
pix += line_size*2; "addl %5, %%eax \n\t"
p += line_size*2; "movq (%1, %%eax), %%mm0 \n\t"
} "movq 1(%1, %%eax), %%mm1 \n\t"
if(dh) { "movq (%2, %%eax), %%mm2 \n\t"
__asm __volatile( "movq 1(%2, %%eax), %%mm3 \n\t"
"movq %1, %%mm1\n\t" PAVGB" %%mm1, %%mm0 \n\t"
"movq 1%1, %%mm2\n\t" PAVGB" %%mm3, %%mm2 \n\t"
"movq %0, %%mm0\n\t" "movq (%3, %%eax), %%mm3 \n\t"
PAVGB" %%mm2, %%mm1\n\t" "movq (%4, %%eax), %%mm4 \n\t"
PAVGB" %%mm1, %%mm0\n\t" PAVGB" %%mm3, %%mm0 \n\t"
"movq %%mm0, %0\n\t" PAVGB" %%mm4, %%mm2 \n\t"
:"+m"(*p) "movq %%mm0, (%3, %%eax) \n\t"
:"m"(*pix) "movq %%mm2, (%4, %%eax) \n\t"
:"memory"); "addl %5, %%eax \n\t"
} "subl $4, %0 \n\t"
" jnz 1b \n\t"
:"+g"(h)
:"r"(pixels), "r"(pixels+line_size), "r" (block), "r" (block+line_size),
"r"(line_size<<1)
:"%eax", "memory");
} }
static void DEF(avg_pixels_y2)( UINT8 *block, const UINT8 *pixels, int line_size, int h) static void DEF(avg_pixels_y2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
{ {
int dh, hh;
UINT8 *p;
const UINT8 *pix;
p = block;
pix = pixels;
hh=h>>1;
dh=h&1;
while(hh--) {
__asm __volatile( __asm __volatile(
"movq %2, %%mm2\n\t" "xorl %%eax, %%eax \n\t"
"movq %3, %%mm3\n\t" "movq (%1), %%mm0 \n\t"
"movq %3, %%mm4\n\t" ".balign 16 \n\t"
"movq %4, %%mm5\n\t" "1: \n\t"
"movq %0, %%mm0\n\t" "movq (%2, %%eax), %%mm1 \n\t"
"movq %1, %%mm1\n\t" "movq (%3, %%eax), %%mm2 \n\t"
PAVGB" %%mm3, %%mm2\n\t" PAVGB" %%mm1, %%mm0 \n\t"
PAVGB" %%mm2, %%mm0\n\t" PAVGB" %%mm2, %%mm1 \n\t"
PAVGB" %%mm5, %%mm4\n\t" "movq (%4, %%eax), %%mm3 \n\t"
PAVGB" %%mm4, %%mm1\n\t" "movq (%5, %%eax), %%mm4 \n\t"
"movq %%mm0, %0\n\t" PAVGB" %%mm3, %%mm0 \n\t"
"movq %%mm1, %1\n\t" PAVGB" %%mm4, %%mm1 \n\t"
:"+m"(*p), "+m"(*(p+line_size)) "movq %%mm0, (%4, %%eax) \n\t"
:"m"(*pix), "m"(*(pix+line_size)), "m"(*(pix+line_size*2)) "movq %%mm1, (%5, %%eax) \n\t"
:"memory"); "addl %6, %%eax \n\t"
pix += line_size*2; "movq (%2, %%eax), %%mm1 \n\t"
p += line_size*2; "movq (%3, %%eax), %%mm0 \n\t"
} PAVGB" %%mm1, %%mm2 \n\t"
if(dh) { PAVGB" %%mm0, %%mm1 \n\t"
__asm __volatile( "movq (%4, %%eax), %%mm3 \n\t"
"movq %1, %%mm1\n\t" "movq (%5, %%eax), %%mm4 \n\t"
"movq %2, %%mm2\n\t" PAVGB" %%mm3, %%mm2 \n\t"
"movq %0, %%mm0\n\t" PAVGB" %%mm4, %%mm1 \n\t"
PAVGB" %%mm2, %%mm1\n\t" "movq %%mm2, (%4, %%eax) \n\t"
PAVGB" %%mm1, %%mm0\n\t" "movq %%mm1, (%5, %%eax) \n\t"
"movq %%mm0, %0\n\t" "addl %6, %%eax \n\t"
:"+m"(*p) "subl $4, %0 \n\t"
:"m"(*pix), "m"(*(pix+line_size)) " jnz 1b \n\t"
:"memory"); :"+g"(h)
} :"r"(pixels), "r"(pixels+line_size), "r"(pixels+line_size*2), "r" (block),
"r" (block+line_size), "g"(line_size<<1)
:"%eax", "memory");
} }
static void DEF(avg_pixels_xy2)( UINT8 *block, const UINT8 *pixels, int line_size, int h) // Note this is not correctly rounded, but this function is only used for b frames so it doesnt matter
static void DEF(avg_pixels_xy2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
{ {
UINT8 *p;
const UINT8 *pix;
p = block;
pix = pixels;
__asm __volatile(
"pxor %%mm7, %%mm7\n\t"
"movq %0, %%mm6\n\t"
::"m"(mm_wtwo));
do {
__asm __volatile( __asm __volatile(
"movq %1, %%mm0\n\t" "movq "MANGLE(mm_bone)", %%mm7 \n\t"
"movq %2, %%mm1\n\t" "xorl %%eax, %%eax \n\t"
"movq 1%1, %%mm4\n\t" "movq (%1), %%mm0 \n\t"
"movq 1%2, %%mm5\n\t" "movq 1(%1), %%mm1 \n\t"
"movq %%mm0, %%mm2\n\t" PAVGB" %%mm1, %%mm0 \n\t"
"movq %%mm1, %%mm3\n\t" ".balign 16 \n\t"
"punpcklbw %%mm7, %%mm0\n\t" "1: \n\t"
"punpcklbw %%mm7, %%mm1\n\t" "movq (%2, %%eax), %%mm1 \n\t"
"punpckhbw %%mm7, %%mm2\n\t" "movq (%3, %%eax), %%mm2 \n\t"
"punpckhbw %%mm7, %%mm3\n\t" "movq 1(%2, %%eax), %%mm3 \n\t"
"paddusw %%mm1, %%mm0\n\t" "movq 1(%3, %%eax), %%mm4 \n\t"
"paddusw %%mm3, %%mm2\n\t" "psubusb %%mm7, %%mm2 \n\t"
"movq %%mm4, %%mm1\n\t" PAVGB" %%mm3, %%mm1 \n\t"
"movq %%mm5, %%mm3\n\t" PAVGB" %%mm4, %%mm2 \n\t"
"punpcklbw %%mm7, %%mm4\n\t" PAVGB" %%mm1, %%mm0 \n\t"
"punpcklbw %%mm7, %%mm5\n\t" PAVGB" %%mm2, %%mm1 \n\t"
"punpckhbw %%mm7, %%mm1\n\t" "movq (%4, %%eax), %%mm3 \n\t"
"punpckhbw %%mm7, %%mm3\n\t" "movq (%5, %%eax), %%mm4 \n\t"
"paddusw %%mm5, %%mm4\n\t" PAVGB" %%mm3, %%mm0 \n\t"
"paddusw %%mm3, %%mm1\n\t" PAVGB" %%mm4, %%mm1 \n\t"
"paddusw %%mm6, %%mm4\n\t" "movq %%mm0, (%4, %%eax) \n\t"
"paddusw %%mm6, %%mm1\n\t" "movq %%mm1, (%5, %%eax) \n\t"
"paddusw %%mm4, %%mm0\n\t" "addl %6, %%eax \n\t"
"paddusw %%mm1, %%mm2\n\t" "movq (%2, %%eax), %%mm1 \n\t"
"psrlw $2, %%mm0\n\t" "movq (%3, %%eax), %%mm0 \n\t"
"psrlw $2, %%mm2\n\t" "movq 1(%2, %%eax), %%mm3 \n\t"
"packuswb %%mm2, %%mm0\n\t" "movq 1(%3, %%eax), %%mm4 \n\t"
PAVGB" %0, %%mm0\n\t" PAVGB" %%mm3, %%mm1 \n\t"
"movq %%mm0, %0\n\t" PAVGB" %%mm4, %%mm0 \n\t"
:"+m"(*p) PAVGB" %%mm1, %%mm2 \n\t"
:"m"(*pix), PAVGB" %%mm0, %%mm1 \n\t"
"m"(*(pix+line_size)) "movq (%4, %%eax), %%mm3 \n\t"
:"memory"); "movq (%5, %%eax), %%mm4 \n\t"
pix += line_size; PAVGB" %%mm3, %%mm2 \n\t"
p += line_size ; PAVGB" %%mm4, %%mm1 \n\t"
} while(--h); "movq %%mm2, (%4, %%eax) \n\t"
"movq %%mm1, (%5, %%eax) \n\t"
"addl %6, %%eax \n\t"
"subl $4, %0 \n\t"
" jnz 1b \n\t"
:"+g"(h)
:"r"(pixels), "r"(pixels+line_size), "r"(pixels+line_size*2), "r" (block),
"r" (block+line_size), "g"(line_size<<1)
:"%eax", "memory");
} }
//Note: the sub* functions are no used
static void DEF(sub_pixels_x2)( DCTELEM *block, const UINT8 *pixels, int line_size, int h) static void DEF(sub_pixels_x2)( DCTELEM *block, const UINT8 *pixels, int line_size, int h)
{ {
DCTELEM *p; DCTELEM *p;