Use more accurate conversion for rgb15/16 to rgb24/32 (C/MMX).

Fate update by michael.

Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

4 files changed, 82 insertions, 112 deletions

diff --git a/libswscale/rgb2rgb.c b/libswscale/rgb2rgb.c
index 9a7f698935..adb3005753 100644
--- a/libswscale/rgb2rgb.c
+++ b/libswscale/rgb2rgb.c
@@ -171,13 +171,13 @@ void rgb16tobgr32(const uint8_t *src, uint8_t *dst, int src_size)
         bgr = *s++;
 #if HAVE_BIGENDIAN
         *d++ = 255;
-        *d++ = (bgr&0x1F)<<3;
-        *d++ = (bgr&0x7E0)>>3;
-        *d++ = (bgr&0xF800)>>8;
+        *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
+        *d++ = ((bgr&0x7E0)>>3) | ((bgr&0x7E0)>>9);
+        *d++ = ((bgr&0xF800)>>8) | ((bgr&0xF800)>>13);
 #else
-        *d++ = (bgr&0xF800)>>8;
-        *d++ = (bgr&0x7E0)>>3;
-        *d++ = (bgr&0x1F)<<3;
+        *d++ = ((bgr&0xF800)>>8) | ((bgr&0xF800)>>13);
+        *d++ = ((bgr&0x7E0)>>3) | ((bgr&0x7E0)>>9);
+        *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
         *d++ = 255;
 #endif
     }
@@ -192,9 +192,9 @@ void rgb16to24(const uint8_t *src, uint8_t *dst, int src_size)
     while (s < end) {
         register uint16_t bgr;
         bgr = *s++;
-        *d++ = (bgr&0xF800)>>8;
-        *d++ = (bgr&0x7E0)>>3;
-        *d++ = (bgr&0x1F)<<3;
+        *d++ = ((bgr&0xF800)>>8) | ((bgr&0xF800)>>13);
+        *d++ = ((bgr&0x7E0)>>3) | ((bgr&0x7E0)>>9);
+        *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
     }
 }
 
@@ -231,13 +231,13 @@ void rgb15tobgr32(const uint8_t *src, uint8_t *dst, int src_size)
         bgr = *s++;
 #if HAVE_BIGENDIAN
         *d++ = 255;
-        *d++ = (bgr&0x1F)<<3;
-        *d++ = (bgr&0x3E0)>>2;
-        *d++ = (bgr&0x7C00)>>7;
+        *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
+        *d++ = ((bgr&0x3E0)>>2) | ((bgr&0x3E0)>>7);
+        *d++ = ((bgr&0x7C00)>>7) | ((bgr&0x7C00)>>12);
 #else
-        *d++ = (bgr&0x7C00)>>7;
-        *d++ = (bgr&0x3E0)>>2;
-        *d++ = (bgr&0x1F)<<3;
+        *d++ = ((bgr&0x7C00)>>7) | ((bgr&0x7C00)>>12);
+        *d++ = ((bgr&0x3E0)>>2) | ((bgr&0x3E0)>>7);
+        *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
         *d++ = 255;
 #endif
     }
@@ -252,9 +252,9 @@ void rgb15to24(const uint8_t *src, uint8_t *dst, int src_size)
     while (s < end) {
         register uint16_t bgr;
         bgr = *s++;
-        *d++ = (bgr&0x7C00)>>7;
-        *d++ = (bgr&0x3E0)>>2;
-        *d++ = (bgr&0x1F)<<3;
+        *d++ = ((bgr&0x7C00)>>7) | ((bgr&0x7C00)>>12);
+        *d++ = ((bgr&0x3E0)>>2) | ((bgr&0x3E0)>>7);
+        *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
     }
 }
 
diff --git a/libswscale/rgb2rgb_template.c b/libswscale/rgb2rgb_template.c
index 0734e8891b..6363bc19aa 100644
--- a/libswscale/rgb2rgb_template.c
+++ b/libswscale/rgb2rgb_template.c
@@ -227,27 +227,6 @@ static inline void rgb24to15_c(const uint8_t *src, uint8_t *dst, int src_size)
     }
 }
 
-/*
-  I use less accurate approximation here by simply left-shifting the input
-  value and filling the low order bits with zeroes. This method improves PNG
-  compression but this scheme cannot reproduce white exactly, since it does
-  not generate an all-ones maximum value; the net effect is to darken the
-  image slightly.
-
-  The better method should be "left bit replication":
-
-   4 3 2 1 0
-   ---------
-   1 1 0 1 1
-
-   7 6 5 4 3  2 1 0
-   ----------------
-   1 1 0 1 1  1 1 0
-   |=======|  |===|
-       |      leftmost bits repeated to fill open bits
-       |
-   original bits
-*/
 static inline void rgb15tobgr24_c(const uint8_t *src, uint8_t *dst, int src_size)
 {
     const uint16_t *end;
@@ -257,9 +236,9 @@ static inline void rgb15tobgr24_c(const uint8_t *src, uint8_t *dst, int src_size
     while (s < end) {
         register uint16_t bgr;
         bgr = *s++;
-        *d++ = (bgr&0x1F)<<3;
-        *d++ = (bgr&0x3E0)>>2;
-        *d++ = (bgr&0x7C00)>>7;
+        *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
+        *d++ = ((bgr&0x3E0)>>2) | ((bgr&0x3E0)>>7);
+        *d++ = ((bgr&0x7C00)>>7) | ((bgr&0x7C00)>>12);
     }
 }
 
@@ -272,9 +251,9 @@ static inline void rgb16tobgr24_c(const uint8_t *src, uint8_t *dst, int src_size
     while (s < end) {
         register uint16_t bgr;
         bgr = *s++;
-        *d++ = (bgr&0x1F)<<3;
-        *d++ = (bgr&0x7E0)>>3;
-        *d++ = (bgr&0xF800)>>8;
+        *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
+        *d++ = ((bgr&0x7E0)>>3) | ((bgr&0x7E0)>>9);
+        *d++ = ((bgr&0xF800)>>8) | ((bgr&0xF800)>>13);
     }
 }
 
@@ -289,13 +268,13 @@ static inline void rgb15to32_c(const uint8_t *src, uint8_t *dst, int src_size)
         bgr = *s++;
 #if HAVE_BIGENDIAN
         *d++ = 255;
-        *d++ = (bgr&0x7C00)>>7;
-        *d++ = (bgr&0x3E0)>>2;
-        *d++ = (bgr&0x1F)<<3;
+        *d++ = ((bgr&0x7C00)>>7) | ((bgr&0x7C00)>>12);
+        *d++ = ((bgr&0x3E0)>>2) | ((bgr&0x3E0)>>7);
+        *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
 #else
-        *d++ = (bgr&0x1F)<<3;
-        *d++ = (bgr&0x3E0)>>2;
-        *d++ = (bgr&0x7C00)>>7;
+        *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
+        *d++ = ((bgr&0x3E0)>>2) | ((bgr&0x3E0)>>7);
+        *d++ = ((bgr&0x7C00)>>7) | ((bgr&0x7C00)>>12);
         *d++ = 255;
 #endif
     }
@@ -312,13 +291,13 @@ static inline void rgb16to32_c(const uint8_t *src, uint8_t *dst, int src_size)
         bgr = *s++;
 #if HAVE_BIGENDIAN
         *d++ = 255;
-        *d++ = (bgr&0xF800)>>8;
-        *d++ = (bgr&0x7E0)>>3;
-        *d++ = (bgr&0x1F)<<3;
+        *d++ = ((bgr&0xF800)>>8) | ((bgr&0xF800)>>13);
+        *d++ = ((bgr&0x7E0)>>3) | ((bgr&0x7E0)>>9);
+        *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
 #else
-        *d++ = (bgr&0x1F)<<3;
-        *d++ = (bgr&0x7E0)>>3;
-        *d++ = (bgr&0xF800)>>8;
+        *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
+        *d++ = ((bgr&0x7E0)>>3) | ((bgr&0x7E0)>>9);
+        *d++ = ((bgr&0xF800)>>8) | ((bgr&0xF800)>>13);
         *d++ = 255;
 #endif
     }
diff --git a/libswscale/x86/rgb2rgb.c b/libswscale/x86/rgb2rgb.c
index ed7f5adb74..9359f0b1f3 100644
--- a/libswscale/x86/rgb2rgb.c
+++ b/libswscale/x86/rgb2rgb.c
@@ -68,6 +68,9 @@ DECLARE_ASM_CONST(8, uint64_t, blue_16mask)  = 0x0000001f0000001fULL;
 DECLARE_ASM_CONST(8, uint64_t, red_15mask)   = 0x00007c0000007c00ULL;
 DECLARE_ASM_CONST(8, uint64_t, green_15mask) = 0x000003e0000003e0ULL;
 DECLARE_ASM_CONST(8, uint64_t, blue_15mask)  = 0x0000001f0000001fULL;
+DECLARE_ASM_CONST(8, uint64_t, mul15_mid)    = 0x4200420042004200ULL;
+DECLARE_ASM_CONST(8, uint64_t, mul15_hi)     = 0x0210021002100210ULL;
+DECLARE_ASM_CONST(8, uint64_t, mul16_mid)    = 0x2080208020802080ULL;
 
 #define RGB2YUV_SHIFT 8
 #define BY ((int)( 0.098*(1<<RGB2YUV_SHIFT)+0.5))
diff --git a/libswscale/x86/rgb2rgb_template.c b/libswscale/x86/rgb2rgb_template.c
index baef3f8ae5..3bca43c42e 100644
--- a/libswscale/x86/rgb2rgb_template.c
+++ b/libswscale/x86/rgb2rgb_template.c
@@ -801,27 +801,6 @@ static inline void RENAME(rgb24to15)(const uint8_t *src, uint8_t *dst, int src_s
     }
 }
 
-/*
-  I use less accurate approximation here by simply left-shifting the input
-  value and filling the low order bits with zeroes. This method improves PNG
-  compression but this scheme cannot reproduce white exactly, since it does
-  not generate an all-ones maximum value; the net effect is to darken the
-  image slightly.
-
-  The better method should be "left bit replication":
-
-   4 3 2 1 0
-   ---------
-   1 1 0 1 1
-
-   7 6 5 4 3  2 1 0
-   ----------------
-   1 1 0 1 1  1 1 0
-   |=======|  |===|
-       |      leftmost bits repeated to fill open bits
-       |
-   original bits
-*/
 static inline void RENAME(rgb15tobgr24)(const uint8_t *src, uint8_t *dst, int src_size)
 {
     const uint16_t *end;
@@ -840,9 +819,10 @@ static inline void RENAME(rgb15tobgr24)(const uint8_t *src, uint8_t *dst, int sr
             "pand          %2, %%mm0    \n\t"
             "pand          %3, %%mm1    \n\t"
             "pand          %4, %%mm2    \n\t"
-            "psllq         $3, %%mm0    \n\t"
-            "psrlq         $2, %%mm1    \n\t"
-            "psrlq         $7, %%mm2    \n\t"
+            "psllq         $5, %%mm0    \n\t"
+            "pmulhw        %6, %%mm0    \n\t"
+            "pmulhw        %6, %%mm1    \n\t"
+            "pmulhw        %7, %%mm2    \n\t"
             "movq       %%mm0, %%mm3    \n\t"
             "movq       %%mm1, %%mm4    \n\t"
             "movq       %%mm2, %%mm5    \n\t"
@@ -870,9 +850,10 @@ static inline void RENAME(rgb15tobgr24)(const uint8_t *src, uint8_t *dst, int sr
             "pand          %2, %%mm0    \n\t"
             "pand          %3, %%mm1    \n\t"
             "pand          %4, %%mm2    \n\t"
-            "psllq         $3, %%mm0    \n\t"
-            "psrlq         $2, %%mm1    \n\t"
-            "psrlq         $7, %%mm2    \n\t"
+            "psllq         $5, %%mm0    \n\t"
+            "pmulhw        %6, %%mm0    \n\t"
+            "pmulhw        %6, %%mm1    \n\t"
+            "pmulhw        %7, %%mm2    \n\t"
             "movq       %%mm0, %%mm3    \n\t"
             "movq       %%mm1, %%mm4    \n\t"
             "movq       %%mm2, %%mm5    \n\t"
@@ -892,7 +873,7 @@ static inline void RENAME(rgb15tobgr24)(const uint8_t *src, uint8_t *dst, int sr
             "por        %%mm5, %%mm3    \n\t"
 
             :"=m"(*d)
-            :"m"(*s),"m"(mask15b),"m"(mask15g),"m"(mask15r), "m"(mmx_null)
+            :"m"(*s),"m"(mask15b),"m"(mask15g),"m"(mask15r),"m"(mmx_null),"m"(mul15_mid),"m"(mul15_hi)
             :"memory");
         /* borrowed 32 to 24 */
         __asm__ volatile(
@@ -919,9 +900,9 @@ static inline void RENAME(rgb15tobgr24)(const uint8_t *src, uint8_t *dst, int sr
     while (s < end) {
         register uint16_t bgr;
         bgr = *s++;
-        *d++ = (bgr&0x1F)<<3;
-        *d++ = (bgr&0x3E0)>>2;
-        *d++ = (bgr&0x7C00)>>7;
+        *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
+        *d++ = ((bgr&0x3E0)>>2) | ((bgr&0x3E0)>>7);
+        *d++ = ((bgr&0x7C00)>>7) | ((bgr&0x7C00)>>12);
     }
 }
 
@@ -943,9 +924,11 @@ static inline void RENAME(rgb16tobgr24)(const uint8_t *src, uint8_t *dst, int sr
             "pand          %2, %%mm0    \n\t"
             "pand          %3, %%mm1    \n\t"
             "pand          %4, %%mm2    \n\t"
-            "psllq         $3, %%mm0    \n\t"
-            "psrlq         $3, %%mm1    \n\t"
-            "psrlq         $8, %%mm2    \n\t"
+            "psllq         $5, %%mm0    \n\t"
+            "psrlq         $1, %%mm2    \n\t"
+            "pmulhw        %6, %%mm0    \n\t"
+            "pmulhw        %8, %%mm1    \n\t"
+            "pmulhw        %7, %%mm2    \n\t"
             "movq       %%mm0, %%mm3    \n\t"
             "movq       %%mm1, %%mm4    \n\t"
             "movq       %%mm2, %%mm5    \n\t"
@@ -973,9 +956,11 @@ static inline void RENAME(rgb16tobgr24)(const uint8_t *src, uint8_t *dst, int sr
             "pand          %2, %%mm0    \n\t"
             "pand          %3, %%mm1    \n\t"
             "pand          %4, %%mm2    \n\t"
-            "psllq         $3, %%mm0    \n\t"
-            "psrlq         $3, %%mm1    \n\t"
-            "psrlq         $8, %%mm2    \n\t"
+            "psllq         $5, %%mm0    \n\t"
+            "psrlq         $1, %%mm2    \n\t"
+            "pmulhw        %6, %%mm0    \n\t"
+            "pmulhw        %8, %%mm1    \n\t"
+            "pmulhw        %7, %%mm2    \n\t"
             "movq       %%mm0, %%mm3    \n\t"
             "movq       %%mm1, %%mm4    \n\t"
             "movq       %%mm2, %%mm5    \n\t"
@@ -994,7 +979,7 @@ static inline void RENAME(rgb16tobgr24)(const uint8_t *src, uint8_t *dst, int sr
             "por        %%mm4, %%mm3    \n\t"
             "por        %%mm5, %%mm3    \n\t"
             :"=m"(*d)
-            :"m"(*s),"m"(mask16b),"m"(mask16g),"m"(mask16r),"m"(mmx_null)
+            :"m"(*s),"m"(mask16b),"m"(mask16g),"m"(mask16r),"m"(mmx_null),"m"(mul15_mid),"m"(mul15_hi),"m"(mul16_mid)
             :"memory");
         /* borrowed 32 to 24 */
         __asm__ volatile(
@@ -1021,9 +1006,9 @@ static inline void RENAME(rgb16tobgr24)(const uint8_t *src, uint8_t *dst, int sr
     while (s < end) {
         register uint16_t bgr;
         bgr = *s++;
-        *d++ = (bgr&0x1F)<<3;
-        *d++ = (bgr&0x7E0)>>3;
-        *d++ = (bgr&0xF800)>>8;
+        *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
+        *d++ = ((bgr&0x7E0)>>3) | ((bgr&0x7E0)>>9);
+        *d++ = ((bgr&0xF800)>>8) | ((bgr&0xF800)>>13);
     }
 }
 
@@ -1066,12 +1051,13 @@ static inline void RENAME(rgb15to32)(const uint8_t *src, uint8_t *dst, int src_s
             "pand          %2, %%mm0    \n\t"
             "pand          %3, %%mm1    \n\t"
             "pand          %4, %%mm2    \n\t"
-            "psllq         $3, %%mm0    \n\t"
-            "psrlq         $2, %%mm1    \n\t"
-            "psrlq         $7, %%mm2    \n\t"
+            "psllq         $5, %%mm0    \n\t"
+            "pmulhw        %5, %%mm0    \n\t"
+            "pmulhw        %5, %%mm1    \n\t"
+            "pmulhw        %6, %%mm2    \n\t"
             PACK_RGB32
             :"=m"(*d)
-            :"m"(*s),"m"(mask15b),"m"(mask15g),"m"(mask15r)
+            :"m"(*s),"m"(mask15b),"m"(mask15g),"m"(mask15r),"m"(mul15_mid),"m"(mul15_hi)
             :"memory");
         d += 16;
         s += 4;
@@ -1081,9 +1067,9 @@ static inline void RENAME(rgb15to32)(const uint8_t *src, uint8_t *dst, int src_s
     while (s < end) {
         register uint16_t bgr;
         bgr = *s++;
-        *d++ = (bgr&0x1F)<<3;
-        *d++ = (bgr&0x3E0)>>2;
-        *d++ = (bgr&0x7C00)>>7;
+        *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
+        *d++ = ((bgr&0x3E0)>>2) | ((bgr&0x3E0)>>7);
+        *d++ = ((bgr&0x7C00)>>7) | ((bgr&0x7C00)>>12);
         *d++ = 255;
     }
 }
@@ -1108,12 +1094,14 @@ static inline void RENAME(rgb16to32)(const uint8_t *src, uint8_t *dst, int src_s
             "pand          %2, %%mm0    \n\t"
             "pand          %3, %%mm1    \n\t"
             "pand          %4, %%mm2    \n\t"
-            "psllq         $3, %%mm0    \n\t"
-            "psrlq         $3, %%mm1    \n\t"
-            "psrlq         $8, %%mm2    \n\t"
+            "psllq         $5, %%mm0    \n\t"
+            "psrlq         $1, %%mm2    \n\t"
+            "pmulhw        %5, %%mm0    \n\t"
+            "pmulhw        %7, %%mm1    \n\t"
+            "pmulhw        %6, %%mm2    \n\t"
             PACK_RGB32
             :"=m"(*d)
-            :"m"(*s),"m"(mask16b),"m"(mask16g),"m"(mask16r)
+            :"m"(*s),"m"(mask16b),"m"(mask16g),"m"(mask16r),"m"(mul15_mid),"m"(mul15_hi),"m"(mul16_mid)
             :"memory");
         d += 16;
         s += 4;
@@ -1123,9 +1111,9 @@ static inline void RENAME(rgb16to32)(const uint8_t *src, uint8_t *dst, int src_s
     while (s < end) {
         register uint16_t bgr;
         bgr = *s++;
-        *d++ = (bgr&0x1F)<<3;
-        *d++ = (bgr&0x7E0)>>3;
-        *d++ = (bgr&0xF800)>>8;
+        *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
+        *d++ = ((bgr&0x7E0)>>3) | ((bgr&0x7E0)>>9);
+        *d++ = ((bgr&0xF800)>>8) | ((bgr&0xF800)>>13);
         *d++ = 255;
     }
 }