gstreamer/gst/goom/xmmx.c
Brian Cameron f3c032e6ac goom: add edx to clobber list in inline assembly code
mull modifies %edx, so should be mentioned in clobber list.
Fixes crash on Solaris ().
2010-04-17 10:26:25 +01:00

402 lines
10 KiB
C
Raw Blame History

/* xmmx.c
eXtended MultiMedia eXtensions GCC interface library for IA32.
To use this library, simply include this header file
and compile with GCC. You MUST have inlining enabled
in order for xmmx_ok() to work; this can be done by
simply using -O on the GCC command line.
Compiling with -DXMMX_TRACE will cause detailed trace
output to be sent to stderr for each mmx operation.
This adds lots of code, and obviously slows execution to
a crawl, but can be very useful for debugging.
THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT ANY
EXPRESS OR IMPLIED WARRANTIES, INCLUDING, WITHOUT
LIMITATION, THE IMPLIED WARRANTIES OF MERCHANTABILITY
AND FITNESS FOR ANY PARTICULAR PURPOSE.
1999 by R. Fisher
Based on libmmx, 1997-99 by H. Dietz and R. Fisher
Notes:
It appears that the latest gas has the pand problem fixed, therefore
I'll undefine BROKEN_PAND by default.
*/
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#include "goom_config.h"
#ifdef HAVE_MMX
/* a definir pour avoir exactement le meme resultat que la fonction C
* (un chouillat plus lent).. mais la difference est assez peu notable.
*/
// #define STRICT_COMPAT
#define BUFFPOINTNB 16
#define BUFFPOINTMASK 0xffff
#define BUFFINCR 0xff
#define sqrtperte 16
/* faire : a % sqrtperte <=> a & pertemask*/
#define PERTEMASK 0xf
/* faire : a / sqrtperte <=> a >> PERTEDEC*/
#define PERTEDEC 4
/*#define MMX_TRACE*/
#include "mmx.h"
/*#include "xmmx.h"*/
#include "goom_graphic.h"
int
xmmx_supported (void)
{
return (mm_support () & 0x8) >> 3;
}
void
zoom_filter_xmmx (int prevX, int prevY,
Pixel * expix1, Pixel * expix2,
int *lbruS, int *lbruD, int buffratio, int precalCoef[16][16])
{
int bufsize = prevX * prevY; /* taille du buffer */
volatile int loop; /* variable de boucle */
mmx_t *brutS = (mmx_t *) lbruS; /* buffer de transformation source */
mmx_t *brutD = (mmx_t *) lbruD; /* buffer de transformation dest */
volatile mmx_t prevXY;
volatile mmx_t ratiox;
/* volatile mmx_t interpix; */
expix1[0].val = expix1[prevX - 1].val = expix1[prevX * prevY - 1].val =
expix1[prevX * prevY - prevX].val = 0;
prevXY.ud[0] = (prevX - 1) << PERTEDEC;
prevXY.ud[1] = (prevY - 1) << PERTEDEC;
ratiox.d[0] = buffratio;
ratiox.d[1] = buffratio;
asm volatile ("\n\t movq %[ratio], %%mm6" "\n\t pslld $16, %%mm6" /* mm6 = [rat16=buffratio<<16 | rat16=buffratio<<16] */
"\n\t pxor %%mm7, %%mm7" /* mm7 = 0 */
::[ratio] "m" (ratiox));
loop = 0;
/*
* NOTE : mm6 et mm7 ne sont pas modifies dans la boucle.
*/
while (loop < bufsize) {
/* Thread #1
* pre : mm6 = [rat16|rat16]
* post : mm0 = S + ((D-S)*rat16 format [X|Y]
* modified = mm0,mm1,mm2
*/
asm volatile ("#1 \n\t movq 0(%[brutS]), %%mm0" "#1 \n\t movq 0(%[brutD]), %%mm1" "#1 \n\t psubd %%mm0, %%mm1" /* mm1 = D - S */
"#1 \n\t movq %%mm1, %%mm2" /* mm2 = D - S */
"#1 \n\t pslld $16, %%mm1" "#1 \n\t pmullw %%mm6, %%mm2" "#1 \n\t pmulhuw %%mm6, %%mm1" "#1 \n\t pslld $16, %%mm0" "#1 \n\t paddd %%mm2, %%mm1" /* mm1 = (D - S) * buffratio >> 16 */
"#1 \n\t paddd %%mm1, %%mm0" /* mm0 = S + mm1 */
"#1 \n\t psrld $16, %%mm0"::[brutS] "r" (&brutS[loop]),
[brutD] "r" (&brutD[loop])
); /* mm0 = S */
/*
* pre : mm0 : position vector on screen
* prevXY : coordinate of the lower-right point on screen
* post : clipped mm0
* modified : mm0,mm1,mm2
*/
asm volatile
("#1 \n\t movq %[prevXY], %%mm1" "#1 \n\t pcmpgtd %%mm0, %%mm1"
/* mm0 en X contient (idem pour Y) :
* 1111 si prevXY > px
* 0000 si prevXY <= px */
#ifdef STRICT_COMPAT
"#1 \n\t movq %%mm1, %%mm2"
"#1 \n\t punpckhdq %%mm2, %%mm2"
"#1 \n\t punpckldq %%mm1, %%mm1" "#1 \n\t pand %%mm2, %%mm0"
#endif
"#1 \n\t pand %%mm1, %%mm0" /* on met a zero la partie qui deborde */
::[prevXY] "m" (prevXY));
/* Thread #2
* pre : mm0 : clipped position on screen
*
* post : mm3 : coefs for this position
* mm1 : X vector [0|X]
*
* modif : eax,esi
*/
__asm__ __volatile__ ("#2 \n\t movd %%mm0,%%esi"
"#2 \n\t movq %%mm0,%%mm1"
"#2 \n\t andl $15,%%esi"
"#2 \n\t psrlq $32,%%mm1"
"#2 \n\t shll $6,%%esi"
"#2 \n\t movd %%mm1,%%eax"
"#2 \n\t addl %[precalCoef],%%esi"
"#2 \n\t andl $15,%%eax"
"#2 \n\t movd (%%esi,%%eax,4),%%mm3"::[precalCoef]
"g" (precalCoef):"eax", "esi");
/*
* extraction des coefficients... (Thread #3)
*
* pre : coef dans mm3
*
* post : coef extraits dans mm3 (c1 & c2)
* et mm4 (c3 & c4)
*
* modif : mm5
*/
/* (Thread #4)
* pre : mm0 : Y pos [*|Y]
* mm1 : X pos [*|X]
*
* post : mm0 : expix1[position]
* mm2 : expix1[position+largeur]
*
* modif : eax, esi
*/
__asm__ __volatile__ ("#2 \n\t psrld $4, %%mm0" "#2 \n\t psrld $4, %%mm1" /* PERTEDEC = $4 */
"#4 \n\t movd %%mm1,%%eax"
"#3 \n\t movq %%mm3,%%mm5"
"#4 \n\t mull %[prevX]"
"#4 \n\t movd %%mm0,%%esi"
"#3 \n\t punpcklbw %%mm5, %%mm3"
"#4 \n\t addl %%esi, %%eax"
"#3 \n\t movq %%mm3, %%mm4"
"#3 \n\t movq %%mm3, %%mm5"
"#4 \n\t movl %[expix1], %%esi"
"#3 \n\t punpcklbw %%mm5, %%mm3"
"#4 \n\t movq (%%esi,%%eax,4),%%mm0"
"#3 \n\t punpckhbw %%mm5, %%mm4"
"#4 \n\t addl %[prevX],%%eax"
"#4 \n\t movq (%%esi,%%eax,4),%%mm2"::[expix1] "g" (expix1)
,[prevX] "g" (prevX)
:"eax", "esi", "edx");
/*
* pre : mm0 : expix1[position]
* mm2 : expix1[position+largeur]
* mm3 & mm4 : coefs
*/
/* recopie des deux premiers pixels dans mm0 et mm1 */
movq_r2r (mm0, mm1); /* b1-v1-r1-a1-b2-v2-r2-a2 */
/* depackage du premier pixel */
punpcklbw_r2r (mm7, mm0); /* 00-b2-00-v2-00-r2-00-a2 */
/* extraction des coefficients... */
movq_r2r (mm3, mm5); /* c2-c2-c2-c2-c1-c1-c1-c1 */
/*^en parrallele^ *//* depackage du 2ieme pixel */
/*^ */ punpckhbw_r2r (mm7, mm1);
/* 00-b1-00-v1-00-r1-00-a1 */
punpcklbw_r2r (mm7, mm5); /* 00-c1-00-c1-00-c1-00-c1 */
punpckhbw_r2r (mm7, mm3); /* 00-c2-00-c2-00-c2-00-c2 */
/* multiplication des pixels par les coefficients */
pmullw_r2r (mm5, mm0); /* c1*b2-c1*v2-c1*r2-c1*a2 */
pmullw_r2r (mm3, mm1); /* c2*b1-c2*v1-c2*r1-c2*a1 */
paddw_r2r (mm1, mm0);
/* ...extraction des 2 derniers coefficients */
movq_r2r (mm4, mm5); /* c4-c4-c4-c4-c3-c3-c3-c3 */
punpcklbw_r2r (mm7, mm4); /* 00-c3-00-c3-00-c3-00-c3 */
punpckhbw_r2r (mm7, mm5); /* 00-c4-00-c4-00-c4-00-c4 */
/* recuperation des 2 derniers pixels */
movq_r2r (mm2, mm1);
/* depackage des pixels */
punpcklbw_r2r (mm7, mm1);
punpckhbw_r2r (mm7, mm2);
/* multiplication pas les coeffs */
pmullw_r2r (mm4, mm1);
pmullw_r2r (mm5, mm2);
/* ajout des valeurs obtenues <20> la valeur finale */
paddw_r2r (mm1, mm0);
paddw_r2r (mm2, mm0);
/* division par 256 = 16+16+16+16, puis repackage du pixel final */
psrlw_i2r (8, mm0);
packuswb_r2r (mm7, mm0);
movd_r2m (mm0, expix2[loop]);
++loop;
}
/* this was femms, which is AMD 3dnow */
__asm__ __volatile__ ("emms\n");
}
#define DRAWMETHOD_PLUS_XMMX(_out,_backbuf,_col) \
{ \
movd_m2r(_backbuf, mm0); \
paddusb_m2r(_col, mm0); \
movd_r2m(mm0, _out); \
}
#define DRAWMETHOD DRAWMETHOD_PLUS_XMMX(*p,*p,col)
void
draw_line_xmmx (Pixel * data, int x1, int y1, int x2, int y2, int col,
int screenx, int screeny)
{
int x, y, dx, dy, yy, xx;
Pixel *p;
if ((y1 < 0) || (y2 < 0) || (x1 < 0) || (x2 < 0) || (y1 >= screeny)
|| (y2 >= screeny) || (x1 >= screenx) || (x2 >= screenx))
goto end_of_line;
dx = x2 - x1;
dy = y2 - y1;
if (x1 >= x2) {
int tmp;
tmp = x1;
x1 = x2;
x2 = tmp;
tmp = y1;
y1 = y2;
y2 = tmp;
dx = x2 - x1;
dy = y2 - y1;
}
/* vertical line */
if (dx == 0) {
if (y1 < y2) {
p = &(data[(screenx * y1) + x1]);
for (y = y1; y <= y2; y++) {
DRAWMETHOD;
p += screenx;
}
} else {
p = &(data[(screenx * y2) + x1]);
for (y = y2; y <= y1; y++) {
DRAWMETHOD;
p += screenx;
}
}
goto end_of_line;
}
/* horizontal line */
if (dy == 0) {
if (x1 < x2) {
p = &(data[(screenx * y1) + x1]);
for (x = x1; x <= x2; x++) {
DRAWMETHOD;
p++;
}
goto end_of_line;
} else {
p = &(data[(screenx * y1) + x2]);
for (x = x2; x <= x1; x++) {
DRAWMETHOD;
p++;
}
goto end_of_line;
}
}
/* 1 */
/* \ */
/* \ */
/* 2 */
if (y2 > y1) {
/* steep */
if (dy > dx) {
dx = ((dx << 16) / dy);
x = x1 << 16;
for (y = y1; y <= y2; y++) {
xx = x >> 16;
p = &(data[(screenx * y) + xx]);
DRAWMETHOD;
if (xx < (screenx - 1)) {
p++;
/* DRAWMETHOD; */
}
x += dx;
}
goto end_of_line;
}
/* shallow */
else {
dy = ((dy << 16) / dx);
y = y1 << 16;
for (x = x1; x <= x2; x++) {
yy = y >> 16;
p = &(data[(screenx * yy) + x]);
DRAWMETHOD;
if (yy < (screeny - 1)) {
p += screeny;
/* DRAWMETHOD; */
}
y += dy;
}
}
}
/* 2 */
/* / */
/* / */
/* 1 */
else {
/* steep */
if (-dy > dx) {
dx = ((dx << 16) / -dy);
x = (x1 + 1) << 16;
for (y = y1; y >= y2; y--) {
xx = x >> 16;
p = &(data[(screenx * y) + xx]);
DRAWMETHOD;
if (xx < (screenx - 1)) {
p--;
/* DRAWMETHOD; */
}
x += dx;
}
goto end_of_line;
}
/* shallow */
else {
dy = ((dy << 16) / dx);
y = y1 << 16;
for (x = x1; x <= x2; x++) {
yy = y >> 16;
p = &(data[(screenx * yy) + x]);
DRAWMETHOD;
if (yy < (screeny - 1)) {
p += screeny;
/* DRAWMETHOD; */
}
y += dy;
}
goto end_of_line;
}
}
end_of_line:
/* this was femms, which is AMD 3dnow */
__asm__ __volatile__ ("emms\n");
}
#else
int
xmmx_supported (void)
{
return (0);
}
#endif