[Patch]: altivec fixes, was Re: [MPlayer-dev-eng] Re: altivec: green blur with mpeg4
Guido Guenther
agx at sigxcpu.org
Tue Aug 3 23:46:51 CEST 2004
Hi,
On Tue, Aug 03, 2004 at 02:31:50PM +0200, Marcin 'Morgoth' Kurek wrote:
> The gree blur is present only if I compile MPlayer using GCC 3.3.x, first
> time when I compiled MPlayer using 3.4.1 I was realy soprised because it
> works fine without this ugly effect.
Yeah that helps, however I need the attached fixes in postproc/ to get
current CVS to build. I need other fixes in other subdirs, but they're
not as clean and I'd rather send them separately.
I've moved a minimal gcc_fixes.h to osdep/ since this might be needed in
a couple of other places. Please apply.
Cheers,
-- Guido
-------------- next part --------------
--- postproc/yuv2rgb_altivec.c.old 2004-08-03 23:27:56.295321776 +0200
+++ postproc/yuv2rgb_altivec.c 2004-08-03 23:28:58.182913432 +0200
@@ -73,6 +73,7 @@
#include "swscale_internal.h"
#include "../mangle.h"
#include "../libvo/img_format.h" //FIXME try to reduce dependency of such stuff
+#include <osdep/gcc_fixes.h>
#undef PROFILE_THE_BEAST
#undef INC_SCALING
@@ -119,13 +120,13 @@
*/
static
const vector unsigned char
- perm_rgb_0 = (vector unsigned char)(0x00,0x01,0x10,0x02,0x03,0x11,0x04,0x05,
+ perm_rgb_0 = (const vector unsigned char)AVV(0x00,0x01,0x10,0x02,0x03,0x11,0x04,0x05,
0x12,0x06,0x07,0x13,0x08,0x09,0x14,0x0a),
- perm_rgb_1 = (vector unsigned char)(0x0b,0x15,0x0c,0x0d,0x16,0x0e,0x0f,0x17,
+ perm_rgb_1 = (const vector unsigned char)AVV(0x0b,0x15,0x0c,0x0d,0x16,0x0e,0x0f,0x17,
0x18,0x19,0x1a,0x1b,0x1c,0x1d,0x1e,0x1f),
- perm_rgb_2 = (vector unsigned char)(0x10,0x11,0x12,0x13,0x14,0x15,0x16,0x17,
+ perm_rgb_2 = (const vector unsigned char)AVV(0x10,0x11,0x12,0x13,0x14,0x15,0x16,0x17,
0x00,0x01,0x18,0x02,0x03,0x19,0x04,0x05),
- perm_rgb_3 = (vector unsigned char)(0x1a,0x06,0x07,0x1b,0x08,0x09,0x1c,0x0a,
+ perm_rgb_3 = (const vector unsigned char)AVV(0x1a,0x06,0x07,0x1b,0x08,0x09,0x1c,0x0a,
0x0b,0x1d,0x0c,0x0d,0x1e,0x0e,0x0f,0x1f);
#define vec_merge3(x2,x1,x0,y0,y1,y2) \
@@ -198,25 +199,25 @@
#define vec_unh(x) \
(vector signed short) \
- vec_perm(x,(typeof(x))(0),\
- (vector unsigned char)(0x10,0x00,0x10,0x01,0x10,0x02,0x10,0x03,\
+ vec_perm(x,(typeof(x))AVV(0),\
+ (vector unsigned char)AVV(0x10,0x00,0x10,0x01,0x10,0x02,0x10,0x03,\
0x10,0x04,0x10,0x05,0x10,0x06,0x10,0x07))
#define vec_unl(x) \
(vector signed short) \
- vec_perm(x,(typeof(x))(0),\
- (vector unsigned char)(0x10,0x08,0x10,0x09,0x10,0x0A,0x10,0x0B,\
+ vec_perm(x,(typeof(x))AVV(0),\
+ (vector unsigned char)AVV(0x10,0x08,0x10,0x09,0x10,0x0A,0x10,0x0B,\
0x10,0x0C,0x10,0x0D,0x10,0x0E,0x10,0x0F))
#define vec_clip(x) \
- vec_max (vec_min (x, (typeof(x))(255)), (typeof(x))(0))
+ vec_max (vec_min (x, (typeof(x))AVV(255)), (typeof(x))AVV(0))
#define vec_packclp_a(x,y) \
(vector unsigned char)vec_pack (vec_clip (x), vec_clip (y))
#define vec_packclp(x,y) \
(vector unsigned char)vec_packs \
- ((vector unsigned short)vec_max (x,(vector signed short) (0)), \
- (vector unsigned short)vec_max (y,(vector signed short) (0)))
+ ((vector unsigned short)vec_max (x,(vector signed short)AVV(0)), \
+ (vector unsigned short)vec_max (y,(vector signed short)AVV(0)))
//#define out_pixels(a,b,c,ptr) vec_mstrgb32(typeof(a),((typeof (a))(0)),a,a,a,ptr)
@@ -229,8 +230,8 @@
Y = vec_mradds (Y, c->CY, c->OY);
- U = vec_sub (U,(vector signed short)(128));
- V = vec_sub (V,(vector signed short)(128));
+ U = vec_sub (U,(vector signed short)AVV(128));
+ V = vec_sub (V,(vector signed short)AVV(128));
// ux = (CBU*(u<<c->CSHIFT)+0x4000)>>15;
ux = vec_sl (U, c->CSHIFT);
@@ -324,8 +325,8 @@
align_perm = vec_lvsl (0, vi); \
v = (vector signed char)vec_perm (vivP[0], vivP[1], align_perm); \
\
- u = (vector signed char)vec_sub (u, (vector signed char)(128)); \
- v = (vector signed char)vec_sub (v, (vector signed char)(128)); \
+ u = (vector signed char)vec_sub (u, (vector signed char)AVV(128)); \
+ v = (vector signed char)vec_sub (v, (vector signed char)AVV(128)); \
U = vec_unpackh (u); \
V = vec_unpackh (v); \
\
@@ -342,18 +343,18 @@
\
/* ux = (CBU*(u<<CSHIFT)+0x4000)>>15 */ \
ux = vec_sl (U, lCSHIFT); \
- ux = vec_mradds (ux, lCBU, (vector signed short)(0)); \
+ ux = vec_mradds (ux, lCBU, (vector signed short)AVV(0)); \
ux0 = vec_mergeh (ux,ux); \
ux1 = vec_mergel (ux,ux); \
\
/* vx = (CRV*(v<<CSHIFT)+0x4000)>>15; */ \
vx = vec_sl (V, lCSHIFT); \
- vx = vec_mradds (vx, lCRV, (vector signed short)(0)); \
+ vx = vec_mradds (vx, lCRV, (vector signed short)AVV(0)); \
vx0 = vec_mergeh (vx,vx); \
vx1 = vec_mergel (vx,vx); \
\
/* uvx = ((CGU*u) + (CGV*v))>>15 */ \
- uvx = vec_mradds (U, lCGU, (vector signed short)(0)); \
+ uvx = vec_mradds (U, lCGU, (vector signed short)AVV(0)); \
uvx = vec_mradds (V, lCGV, uvx); \
uvx0 = vec_mergeh (uvx,uvx); \
uvx1 = vec_mergel (uvx,uvx); \
@@ -403,10 +404,10 @@
}
-#define out_abgr(a,b,c,ptr) vec_mstrgb32(typeof(a),((typeof (a))(0)),c,b,a,ptr)
-#define out_bgra(a,b,c,ptr) vec_mstrgb32(typeof(a),c,b,a,((typeof (a))(0)),ptr)
-#define out_rgba(a,b,c,ptr) vec_mstrgb32(typeof(a),a,b,c,((typeof (a))(0)),ptr)
-#define out_argb(a,b,c,ptr) vec_mstrgb32(typeof(a),((typeof (a))(0)),a,b,c,ptr)
+#define out_abgr(a,b,c,ptr) vec_mstrgb32(typeof(a),((typeof (a))AVV(0)),c,b,a,ptr)
+#define out_bgra(a,b,c,ptr) vec_mstrgb32(typeof(a),c,b,a,((typeof (a))AVV(0)),ptr)
+#define out_rgba(a,b,c,ptr) vec_mstrgb32(typeof(a),a,b,c,((typeof (a))AVV(0)),ptr)
+#define out_argb(a,b,c,ptr) vec_mstrgb32(typeof(a),((typeof (a))AVV(0)),a,b,c,ptr)
#define out_rgb24(a,b,c,ptr) vec_mstrgb24(a,b,c,ptr)
#define out_bgr24(a,b,c,ptr) vec_mstrgb24(c,b,a,ptr)
@@ -422,15 +423,15 @@
// 0123 4567 89ab cdef
static
const vector unsigned char
- demux_u = (vector unsigned char)(0x10,0x00,0x10,0x00,
+ demux_u = (const vector unsigned char)AVV(0x10,0x00,0x10,0x00,
0x10,0x04,0x10,0x04,
0x10,0x08,0x10,0x08,
0x10,0x0c,0x10,0x0c),
- demux_v = (vector unsigned char)(0x10,0x02,0x10,0x02,
+ demux_v = (const vector unsigned char)AVV(0x10,0x02,0x10,0x02,
0x10,0x06,0x10,0x06,
0x10,0x0A,0x10,0x0A,
0x10,0x0E,0x10,0x0E),
- demux_y = (vector unsigned char)(0x10,0x01,0x10,0x03,
+ demux_y = (const vector unsigned char)AVV(0x10,0x01,0x10,0x03,
0x10,0x05,0x10,0x07,
0x10,0x09,0x10,0x0B,
0x10,0x0D,0x10,0x0F);
@@ -461,25 +462,25 @@
for (j=0;j<w/16;j++) {
uyvy = vec_ld (0, img);
U = (vector signed short)
- vec_perm (uyvy, (vector unsigned char)(0), demux_u);
+ vec_perm (uyvy, (vector unsigned char)AVV(0), demux_u);
V = (vector signed short)
- vec_perm (uyvy, (vector unsigned char)(0), demux_v);
+ vec_perm (uyvy, (vector unsigned char)AVV(0), demux_v);
Y = (vector signed short)
- vec_perm (uyvy, (vector unsigned char)(0), demux_y);
+ vec_perm (uyvy, (vector unsigned char)AVV(0), demux_y);
cvtyuvtoRGB (c, Y,U,V,&R0,&G0,&B0);
uyvy = vec_ld (16, img);
U = (vector signed short)
- vec_perm (uyvy, (vector unsigned char)(0), demux_u);
+ vec_perm (uyvy, (vector unsigned char)AVV(0), demux_u);
V = (vector signed short)
- vec_perm (uyvy, (vector unsigned char)(0), demux_v);
+ vec_perm (uyvy, (vector unsigned char)AVV(0), demux_v);
Y = (vector signed short)
- vec_perm (uyvy, (vector unsigned char)(0), demux_y);
+ vec_perm (uyvy, (vector unsigned char)AVV(0), demux_y);
cvtyuvtoRGB (c, Y,U,V,&R1,&G1,&B1);
@@ -608,7 +609,7 @@
CGV = vec_lde (0, &tmp);
CGV = vec_splat (CGV, 0);
- c->CSHIFT = (vector unsigned short)(2);
+ c->CSHIFT = (vector unsigned short)AVV(2);
c->CY = CY;
c->OY = OY;
c->CRV = CRV;
@@ -644,8 +645,8 @@
vector unsigned char R,G,B,pels[3];
vector unsigned char *out,*nout;
- vector signed short RND = (vector signed short)(1<<3);
- vector unsigned short SCL = (vector unsigned short)(4);
+ vector signed short RND = (vector signed short)AVV(1<<3);
+ vector unsigned short SCL = (vector unsigned short)AVV(4);
unsigned long scratch[16] __attribute__ ((aligned (16)));
vector signed short *vYCoeffsBank, *vCCoeffsBank;
--- /dev/null 2004-08-03 20:52:07.754821088 +0200
+++ osdep/gcc_fixes.h 2004-08-03 23:39:30.281819792 +0200
@@ -0,0 +1,24 @@
+/*
+ * gcc fixes for altivec.
+ * Used to workaround broken gcc (FSF gcc-3 pre gcc-3.3)
+ * and to stay somewhat compatible with Darwin.
+ */
+
+#ifndef _GCC_FIXES_
+#define _GCC_FIXES_
+
+#ifdef HAVE_ALTIVEC_H
+#include <altivec.h>
+#endif
+
+#ifdef CONFIG_DARWIN
+# ifndef __MWERKS__
+# define AVV(x...) (x)
+# else
+# define AVV
+# endif
+#else
+#define AVV(x...) {x}
+#endif
+
+#endif /* _GCC_FIXES_ */
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: Digital signature
URL: <http://lists.mplayerhq.hu/pipermail/mplayer-dev-eng/attachments/20040803/8a301d10/attachment.pgp>
More information about the MPlayer-dev-eng
mailing list