Skip to content

L
libswscale
Commit 87d2542e authored by michael's avatar michael
Browse files
Options

horizontal scaling bugs fixed, should be mostly bugfree now 


git-svn-id: file:///var/local/repositories/mplayer/trunk/postproc@2280 b3059339-0415-0410-9bf9-f77b7e298cf2
parent e3240911
Hide whitespace changes
Inline Side-by-side
Showing with 56 additions and 50 deletions
swscale.c
View file @ 87d2542e
...@@ -11,14 +11,13 @@ ...@@ -11,14 +11,13 @@
//#undef HAVE_MMX //#undef HAVE_MMX
//#undef ARCH_X86 //#undef ARCH_X86
#define DITHER16BPP #define DITHER16BPP
#define ALT_ERROR //#define ALT_ERROR
#define RET 0xC3 //near return opcode #define RET 0xC3 //near return opcode
/* /*
NOTES NOTES
known BUGS with known cause (no bugreports please!) known BUGS with known cause (no bugreports please!)
line at the right (c,asm and mmx2)
code reads 1 sample too much (might cause a sig11) code reads 1 sample too much (might cause a sig11)
TODO TODO
...@@ -97,25 +96,24 @@ static int static_dstw; ...@@ -97,25 +96,24 @@ static int static_dstw;
static int old_dstw= -1; static int old_dstw= -1;
static int old_s_xinc= -1; static int old_s_xinc= -1;
// difference between the requested xinc and the required one for the mmx2 routine
static int s_xinc_diff=0;
static int s_xinc2_diff=0;
#endif #endif
int canMMX2BeUsed; int canMMX2BeUsed=0;
static int test=0;
// we need that precission at least for the mmx2 code int srcWidth= (dstw*s_xinc + 0x8000)>>16;
//s_xinc*= 256;
s_xinc2=s_xinc>>1;
canMMX2BeUsed= (s_xinc <= 0x10000 && (dstw&31)==0) ? 1 : 0;
#ifdef HAVE_MMX2 #ifdef HAVE_MMX2
if(canMMX2BeUsed) canMMX2BeUsed= (s_xinc <= 0x10000 && (dstw&31)==0 && (srcWidth&15)==0) ? 1 : 0;
{
s_xinc+= s_xinc_diff;
s_xinc2+= s_xinc2_diff;
}
#endif #endif
// match pixel 0 of the src to pixel 0 of dst and match pixel n-2 of src to pixel n-2 of dst
// n-2 is the last chrominance sample available
// FIXME this is not perfect, but noone shuld notice the difference, the more correct variant
// would be like the vertical one, but that would require some special code for the
// first and last pixel
if(canMMX2BeUsed) s_xinc+= 20;
else s_xinc = ((srcWidth-2)<<16)/(dstw-2) - 20;
s_xinc2=s_xinc>>1;
// force calculation of the horizontal interpolation of the first line // force calculation of the horizontal interpolation of the first line
s_last_ypos=-99; s_last_ypos=-99;
s_last_y1pos=-99; s_last_y1pos=-99;
...@@ -185,21 +183,21 @@ canMMX2BeUsed= (s_xinc <= 0x10000 && (dstw&31)==0) ? 1 : 0; ...@@ -185,21 +183,21 @@ canMMX2BeUsed= (s_xinc <= 0x10000 && (dstw&31)==0) ? 1 : 0;
"=r" (fragmentLength) "=r" (fragmentLength)
); );
xpos= xx=xalpha= 0; xpos= 0; //s_xinc/2 - 0x8000; // difference between pixel centers
/* choose xinc so that all 8 parts fit exactly /* choose xinc so that all 8 parts fit exactly
Note: we cannot use just 1 part because it would not fit in the code cache */ Note: we cannot use just 1 part because it would not fit in the code cache */
s_xinc2_diff= -((((s_xinc2*(dstw/8))&0xFFFF))/(dstw/8))+10; // s_xinc2_diff= -((((s_xinc2*(dstw/8))&0xFFFF))/(dstw/8))-10;
// s_xinc_diff= -((((s_xinc*(dstw/8))&0xFFFF))/(dstw/8)); // s_xinc_diff= -((((s_xinc*(dstw/8))&0xFFFF))/(dstw/8));
#ifdef ALT_ERROR #ifdef ALT_ERROR
s_xinc2_diff+= ((0x10000/(dstw/8))); // s_xinc2_diff+= ((0x10000/(dstw/8)));
#endif #endif
s_xinc_diff= s_xinc2_diff*2; // s_xinc_diff= s_xinc2_diff*2;
s_xinc2+= s_xinc2_diff; // s_xinc2+= s_xinc2_diff;
s_xinc+= s_xinc_diff; // s_xinc+= s_xinc_diff;
old_s_xinc= s_xinc; // old_s_xinc= s_xinc;
for(i=0; i<dstw/8; i++) for(i=0; i<dstw/8; i++)
{ {
...@@ -223,8 +221,7 @@ canMMX2BeUsed= (s_xinc <= 0x10000 && (dstw&31)==0) ? 1 : 0; ...@@ -223,8 +221,7 @@ canMMX2BeUsed= (s_xinc <= 0x10000 && (dstw&31)==0) ? 1 : 0;
xpos+=s_xinc; xpos+=s_xinc;
} }
xpos= xx=xalpha= 0; xpos= 0; //s_xinc2/2 - 0x10000; // difference between centers of chrom samples
//FIXME choose size and or xinc so that they fit exactly
for(i=0; i<dstw/8; i++) for(i=0; i<dstw/8; i++)
{ {
int xx=xpos>>16; int xx=xpos>>16;
...@@ -351,6 +348,7 @@ canMMX2BeUsed= (s_xinc <= 0x10000 && (dstw&31)==0) ? 1 : 0; ...@@ -351,6 +348,7 @@ canMMX2BeUsed= (s_xinc <= 0x10000 && (dstw&31)==0) ? 1 : 0;
"m" ((s_xinc*4)&0xFFFF), "m" (s_xinc&0xFFFF) "m" ((s_xinc*4)&0xFFFF), "m" (s_xinc&0xFFFF)
: "%eax", "%ebx", "%ecx", "%edx", "%esi", "%edi" : "%eax", "%ebx", "%ecx", "%edx", "%esi", "%edi"
); );
for(i=dstw-1; (i*s_xinc)>>16 >=srcWidth-1; i--) buf1[i] = src[srcWidth-1]*128;
} }
else else
{ {
...@@ -478,6 +476,11 @@ FUNNYUVCODE ...@@ -478,6 +476,11 @@ FUNNYUVCODE
"m" ((s_xinc2*4)&0xFFFF), "m" (s_xinc2&0xFFFF), "m" (src2) "m" ((s_xinc2*4)&0xFFFF), "m" (s_xinc2&0xFFFF), "m" (src2)
: "%eax", "%ebx", "%ecx", "%edx", "%esi", "%edi" : "%eax", "%ebx", "%ecx", "%edx", "%esi", "%edi"
); );
for(i=dstw-1; (i*s_xinc2)>>16 >=srcWidth/2-1; i--)
{
uvbuf1[i] = src1[srcWidth/2-1]*128;
uvbuf1[i+2048] = src2[srcWidth/2-1]*128;
}
} }
else else
{ {
......
swscale_template.c
View file @ 87d2542e
...@@ -11,14 +11,13 @@ ...@@ -11,14 +11,13 @@
//#undef HAVE_MMX //#undef HAVE_MMX
//#undef ARCH_X86 //#undef ARCH_X86
#define DITHER16BPP #define DITHER16BPP
#define ALT_ERROR //#define ALT_ERROR
#define RET 0xC3 //near return opcode #define RET 0xC3 //near return opcode
/* /*
NOTES NOTES
known BUGS with known cause (no bugreports please!) known BUGS with known cause (no bugreports please!)
line at the right (c,asm and mmx2)
code reads 1 sample too much (might cause a sig11) code reads 1 sample too much (might cause a sig11)
TODO TODO
...@@ -97,25 +96,24 @@ static int static_dstw; ...@@ -97,25 +96,24 @@ static int static_dstw;
static int old_dstw= -1; static int old_dstw= -1;
static int old_s_xinc= -1; static int old_s_xinc= -1;
// difference between the requested xinc and the required one for the mmx2 routine
static int s_xinc_diff=0;
static int s_xinc2_diff=0;
#endif #endif
int canMMX2BeUsed; int canMMX2BeUsed=0;
static int test=0;
// we need that precission at least for the mmx2 code int srcWidth= (dstw*s_xinc + 0x8000)>>16;
//s_xinc*= 256;
s_xinc2=s_xinc>>1;
canMMX2BeUsed= (s_xinc <= 0x10000 && (dstw&31)==0) ? 1 : 0;
#ifdef HAVE_MMX2 #ifdef HAVE_MMX2
if(canMMX2BeUsed) canMMX2BeUsed= (s_xinc <= 0x10000 && (dstw&31)==0 && (srcWidth&15)==0) ? 1 : 0;
{
s_xinc+= s_xinc_diff;
s_xinc2+= s_xinc2_diff;
}
#endif #endif
// match pixel 0 of the src to pixel 0 of dst and match pixel n-2 of src to pixel n-2 of dst
// n-2 is the last chrominance sample available
// FIXME this is not perfect, but noone shuld notice the difference, the more correct variant
// would be like the vertical one, but that would require some special code for the
// first and last pixel
if(canMMX2BeUsed) s_xinc+= 20;
else s_xinc = ((srcWidth-2)<<16)/(dstw-2) - 20;
s_xinc2=s_xinc>>1;
// force calculation of the horizontal interpolation of the first line // force calculation of the horizontal interpolation of the first line
s_last_ypos=-99; s_last_ypos=-99;
s_last_y1pos=-99; s_last_y1pos=-99;
...@@ -185,21 +183,21 @@ canMMX2BeUsed= (s_xinc <= 0x10000 && (dstw&31)==0) ? 1 : 0; ...@@ -185,21 +183,21 @@ canMMX2BeUsed= (s_xinc <= 0x10000 && (dstw&31)==0) ? 1 : 0;
"=r" (fragmentLength) "=r" (fragmentLength)
); );
xpos= xx=xalpha= 0; xpos= 0; //s_xinc/2 - 0x8000; // difference between pixel centers
/* choose xinc so that all 8 parts fit exactly /* choose xinc so that all 8 parts fit exactly
Note: we cannot use just 1 part because it would not fit in the code cache */ Note: we cannot use just 1 part because it would not fit in the code cache */
s_xinc2_diff= -((((s_xinc2*(dstw/8))&0xFFFF))/(dstw/8))+10; // s_xinc2_diff= -((((s_xinc2*(dstw/8))&0xFFFF))/(dstw/8))-10;
// s_xinc_diff= -((((s_xinc*(dstw/8))&0xFFFF))/(dstw/8)); // s_xinc_diff= -((((s_xinc*(dstw/8))&0xFFFF))/(dstw/8));
#ifdef ALT_ERROR #ifdef ALT_ERROR
s_xinc2_diff+= ((0x10000/(dstw/8))); // s_xinc2_diff+= ((0x10000/(dstw/8)));
#endif #endif
s_xinc_diff= s_xinc2_diff*2; // s_xinc_diff= s_xinc2_diff*2;
s_xinc2+= s_xinc2_diff; // s_xinc2+= s_xinc2_diff;
s_xinc+= s_xinc_diff; // s_xinc+= s_xinc_diff;
old_s_xinc= s_xinc; // old_s_xinc= s_xinc;
for(i=0; i<dstw/8; i++) for(i=0; i<dstw/8; i++)
{ {
...@@ -223,8 +221,7 @@ canMMX2BeUsed= (s_xinc <= 0x10000 && (dstw&31)==0) ? 1 : 0; ...@@ -223,8 +221,7 @@ canMMX2BeUsed= (s_xinc <= 0x10000 && (dstw&31)==0) ? 1 : 0;
xpos+=s_xinc; xpos+=s_xinc;
} }
xpos= xx=xalpha= 0; xpos= 0; //s_xinc2/2 - 0x10000; // difference between centers of chrom samples
//FIXME choose size and or xinc so that they fit exactly
for(i=0; i<dstw/8; i++) for(i=0; i<dstw/8; i++)
{ {
int xx=xpos>>16; int xx=xpos>>16;
...@@ -351,6 +348,7 @@ canMMX2BeUsed= (s_xinc <= 0x10000 && (dstw&31)==0) ? 1 : 0; ...@@ -351,6 +348,7 @@ canMMX2BeUsed= (s_xinc <= 0x10000 && (dstw&31)==0) ? 1 : 0;
"m" ((s_xinc*4)&0xFFFF), "m" (s_xinc&0xFFFF) "m" ((s_xinc*4)&0xFFFF), "m" (s_xinc&0xFFFF)
: "%eax", "%ebx", "%ecx", "%edx", "%esi", "%edi" : "%eax", "%ebx", "%ecx", "%edx", "%esi", "%edi"
); );
for(i=dstw-1; (i*s_xinc)>>16 >=srcWidth-1; i--) buf1[i] = src[srcWidth-1]*128;
} }
else else
{ {
...@@ -478,6 +476,11 @@ FUNNYUVCODE ...@@ -478,6 +476,11 @@ FUNNYUVCODE
"m" ((s_xinc2*4)&0xFFFF), "m" (s_xinc2&0xFFFF), "m" (src2) "m" ((s_xinc2*4)&0xFFFF), "m" (s_xinc2&0xFFFF), "m" (src2)
: "%eax", "%ebx", "%ecx", "%edx", "%esi", "%edi" : "%eax", "%ebx", "%ecx", "%edx", "%esi", "%edi"
); );
for(i=dstw-1; (i*s_xinc2)>>16 >=srcWidth/2-1; i--)
{
uvbuf1[i] = src1[srcWidth/2-1]*128;
uvbuf1[i+2048] = src2[srcWidth/2-1]*128;
}
} }
else else
{ {
......
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