tweak mmx2 sad.

40% faster on core2, 18% faster on k8, 5% faster on p4.

Originally committed as revision 8991 to svn://svn.ffmpeg.org/ffmpeg/trunk

1 files changed, 121 insertions, 90 deletions

diff --git a/libavcodec/i386/motion_est_mmx.c b/libavcodec/i386/motion_est_mmx.c
index 3abbbe8cbc..af8db79b6b 100644
--- a/libavcodec/i386/motion_est_mmx.c
+++ b/libavcodec/i386/motion_est_mmx.c
@@ -70,74 +70,98 @@ static inline void sad8_1_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
 
 static inline void sad8_1_mmx2(uint8_t *blk1, uint8_t *blk2, int stride, int h)
 {
-    long len= -(stride*h);
     asm volatile(
         ASMALIGN(4)
         "1:                             \n\t"
-        "movq (%1, %%"REG_a"), %%mm0    \n\t"
-        "psadbw (%2, %%"REG_a"), %%mm0  \n\t"
-        "add %3, %%"REG_a"              \n\t"
-        "movq (%1, %%"REG_a"), %%mm1    \n\t"
-        "psadbw (%2, %%"REG_a"), %%mm1  \n\t"
-        "paddw %%mm1, %%mm0             \n\t"
+        "movq (%1), %%mm0               \n\t"
+        "movq (%1, %3), %%mm1           \n\t"
+        "psadbw (%2), %%mm0             \n\t"
+        "psadbw (%2, %3), %%mm1         \n\t"
         "paddw %%mm0, %%mm6             \n\t"
-        "add %3, %%"REG_a"              \n\t"
-        " js 1b                         \n\t"
-        : "+a" (len)
-        : "r" (blk1 - len), "r" (blk2 - len), "r" ((long)stride)
+        "paddw %%mm1, %%mm6             \n\t"
+        "lea (%1,%3,2), %1              \n\t"
+        "lea (%2,%3,2), %2              \n\t"
+        "sub $2, %0                     \n\t"
+        " jg 1b                         \n\t"
+        : "+r" (h), "+r" (blk1), "+r" (blk2)
+        : "r" ((long)stride)
     );
 }
 
-static inline void sad8_2_mmx2(uint8_t *blk1a, uint8_t *blk1b, uint8_t *blk2, int stride, int h)
+static inline void sad8_x2a_mmx2(uint8_t *blk1, uint8_t *blk2, int stride, int h)
 {
-    long len= -(stride*h);
     asm volatile(
         ASMALIGN(4)
         "1:                             \n\t"
-        "movq (%1, %%"REG_a"), %%mm0    \n\t"
-        "pavgb (%2, %%"REG_a"), %%mm0   \n\t"
-        "psadbw (%3, %%"REG_a"), %%mm0  \n\t"
-        "add %4, %%"REG_a"              \n\t"
-        "movq (%1, %%"REG_a"), %%mm1    \n\t"
-        "pavgb (%2, %%"REG_a"), %%mm1   \n\t"
-        "psadbw (%3, %%"REG_a"), %%mm1  \n\t"
-        "paddw %%mm1, %%mm0             \n\t"
+        "movq (%1), %%mm0               \n\t"
+        "movq (%1, %3), %%mm1           \n\t"
+        "pavgb 1(%1), %%mm0             \n\t"
+        "pavgb 1(%1, %3), %%mm1         \n\t"
+        "psadbw (%2), %%mm0             \n\t"
+        "psadbw (%2, %3), %%mm1         \n\t"
         "paddw %%mm0, %%mm6             \n\t"
-        "add %4, %%"REG_a"              \n\t"
-        " js 1b                         \n\t"
-        : "+a" (len)
-        : "r" (blk1a - len), "r" (blk1b -len), "r" (blk2 - len), "r" ((long)stride)
+        "paddw %%mm1, %%mm6             \n\t"
+        "lea (%1,%3,2), %1              \n\t"
+        "lea (%2,%3,2), %2              \n\t"
+        "sub $2, %0                     \n\t"
+        " jg 1b                         \n\t"
+        : "+r" (h), "+r" (blk1), "+r" (blk2)
+        : "r" ((long)stride)
+    );
+}
+
+static inline void sad8_y2a_mmx2(uint8_t *blk1, uint8_t *blk2, int stride, int h)
+{
+    asm volatile(
+        "movq (%1), %%mm0               \n\t"
+        "add %3, %1                     \n\t"
+        ASMALIGN(4)
+        "1:                             \n\t"
+        "movq (%1), %%mm1               \n\t"
+        "movq (%1, %3), %%mm2           \n\t"
+        "pavgb %%mm1, %%mm0             \n\t"
+        "pavgb %%mm2, %%mm1             \n\t"
+        "psadbw (%2), %%mm0             \n\t"
+        "psadbw (%2, %3), %%mm1         \n\t"
+        "paddw %%mm0, %%mm6             \n\t"
+        "paddw %%mm1, %%mm6             \n\t"
+        "movq %%mm2, %%mm0              \n\t"
+        "lea (%1,%3,2), %1              \n\t"
+        "lea (%2,%3,2), %2              \n\t"
+        "sub $2, %0                     \n\t"
+        " jg 1b                         \n\t"
+        : "+r" (h), "+r" (blk1), "+r" (blk2)
+        : "r" ((long)stride)
     );
 }
 
 static inline void sad8_4_mmx2(uint8_t *blk1, uint8_t *blk2, int stride, int h)
-{ //FIXME reuse src
-    long len= -(stride*h);
+{
     asm volatile(
         "movq "MANGLE(bone)", %%mm5     \n\t"
+        "movq (%1), %%mm0               \n\t"
+        "pavgb 1(%1), %%mm0             \n\t"
+        "add %3, %1                     \n\t"
         ASMALIGN(4)
         "1:                             \n\t"
-        "movq (%1, %%"REG_a"), %%mm0    \n\t"
-        "movq 1(%1, %%"REG_a"), %%mm1   \n\t"
-        "pavgb (%2, %%"REG_a"), %%mm0   \n\t"
-        "pavgb 1(%2, %%"REG_a"), %%mm1  \n\t"
+        "movq (%1), %%mm1               \n\t"
+        "movq (%1,%3), %%mm2            \n\t"
+        "pavgb 1(%1), %%mm1             \n\t"
+        "pavgb 1(%1,%3), %%mm2          \n\t"
         "psubusb %%mm5, %%mm1           \n\t"
         "pavgb %%mm1, %%mm0             \n\t"
-        "psadbw (%3, %%"REG_a"), %%mm0  \n\t"
-        "add %4, %%"REG_a"              \n\t"
-        "movq (%1, %%"REG_a"), %%mm2    \n\t"
-        "movq 1(%1, %%"REG_a"), %%mm3   \n\t"
-        "pavgb (%2, %%"REG_a"), %%mm2   \n\t"
-        "pavgb 1(%2, %%"REG_a"), %%mm3  \n\t"
-        "psubusb %%mm5, %%mm3           \n\t"
-        "pavgb %%mm3, %%mm2             \n\t"
-        "psadbw (%3, %%"REG_a"), %%mm2  \n\t"
-        "paddw %%mm2, %%mm0             \n\t"
+        "pavgb %%mm2, %%mm1             \n\t"
+        "psadbw (%2), %%mm0             \n\t"
+        "psadbw (%2,%3), %%mm1          \n\t"
         "paddw %%mm0, %%mm6             \n\t"
-        "add %4, %%"REG_a"              \n\t"
-        " js 1b                         \n\t"
-        : "+a" (len)
-        : "r" (blk1 - len), "r" (blk1 - len + stride), "r" (blk2 - len), "r" ((long)stride)
+        "paddw %%mm1, %%mm6             \n\t"
+        "movq %%mm2, %%mm0              \n\t"
+        "lea (%1,%3,2), %1              \n\t"
+        "lea (%2,%3,2), %2              \n\t"
+        "sub $2, %0                     \n\t"
+        " jg 1b                         \n\t"
+        : "+r" (h), "+r" (blk1), "+r" (blk2)
+        : "r" ((long)stride)
     );
 }
 
@@ -183,45 +207,48 @@ static inline void sad8_4_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
 {
     long len= -(stride*h);
     asm volatile(
-        ASMALIGN(4)
-        "1:                             \n\t"
         "movq (%1, %%"REG_a"), %%mm0    \n\t"
-        "movq (%2, %%"REG_a"), %%mm1    \n\t"
-        "movq %%mm0, %%mm4              \n\t"
-        "movq %%mm1, %%mm2              \n\t"
-        "punpcklbw %%mm7, %%mm0         \n\t"
-        "punpcklbw %%mm7, %%mm1         \n\t"
-        "punpckhbw %%mm7, %%mm4         \n\t"
-        "punpckhbw %%mm7, %%mm2         \n\t"
-        "paddw %%mm1, %%mm0             \n\t"
-        "paddw %%mm2, %%mm4             \n\t"
         "movq 1(%1, %%"REG_a"), %%mm2   \n\t"
-        "movq 1(%2, %%"REG_a"), %%mm3   \n\t"
-        "movq %%mm2, %%mm1              \n\t"
-        "punpcklbw %%mm7, %%mm2         \n\t"
+        "movq %%mm0, %%mm1              \n\t"
+        "movq %%mm2, %%mm3              \n\t"
+        "punpcklbw %%mm7, %%mm0         \n\t"
         "punpckhbw %%mm7, %%mm1         \n\t"
-        "paddw %%mm0, %%mm2             \n\t"
-        "paddw %%mm4, %%mm1             \n\t"
-        "movq %%mm3, %%mm4              \n\t"
-        "punpcklbw %%mm7, %%mm3         \n\t"
-        "punpckhbw %%mm7, %%mm4         \n\t"
-        "paddw %%mm3, %%mm2             \n\t"
-        "paddw %%mm4, %%mm1             \n\t"
-        "movq (%3, %%"REG_a"), %%mm3    \n\t"
-        "movq (%3, %%"REG_a"), %%mm4    \n\t"
-        "paddw %%mm5, %%mm2             \n\t"
+        "punpcklbw %%mm7, %%mm2         \n\t"
+        "punpckhbw %%mm7, %%mm3         \n\t"
+        "paddw %%mm2, %%mm0             \n\t"
+        "paddw %%mm3, %%mm1             \n\t"
+        ASMALIGN(4)
+        "1:                             \n\t"
+        "movq (%2, %%"REG_a"), %%mm2    \n\t"
+        "movq 1(%2, %%"REG_a"), %%mm4   \n\t"
+        "movq %%mm2, %%mm3              \n\t"
+        "movq %%mm4, %%mm5              \n\t"
+        "punpcklbw %%mm7, %%mm2         \n\t"
+        "punpckhbw %%mm7, %%mm3         \n\t"
+        "punpcklbw %%mm7, %%mm4         \n\t"
+        "punpckhbw %%mm7, %%mm5         \n\t"
+        "paddw %%mm4, %%mm2             \n\t"
+        "paddw %%mm5, %%mm3             \n\t"
+        "movq 16+"MANGLE(round_tab)", %%mm5 \n\t"
+        "paddw %%mm2, %%mm0             \n\t"
+        "paddw %%mm3, %%mm1             \n\t"
+        "paddw %%mm5, %%mm0             \n\t"
         "paddw %%mm5, %%mm1             \n\t"
-        "psrlw $2, %%mm2                \n\t"
+        "movq (%3, %%"REG_a"), %%mm4    \n\t"
+        "movq (%3, %%"REG_a"), %%mm5    \n\t"
+        "psrlw $2, %%mm0                \n\t"
         "psrlw $2, %%mm1                \n\t"
-        "packuswb %%mm1, %%mm2          \n\t"
-        "psubusb %%mm2, %%mm3           \n\t"
-        "psubusb %%mm4, %%mm2           \n\t"
-        "por %%mm3, %%mm2               \n\t"
-        "movq %%mm2, %%mm0              \n\t"
+        "packuswb %%mm1, %%mm0          \n\t"
+        "psubusb %%mm0, %%mm4           \n\t"
+        "psubusb %%mm5, %%mm0           \n\t"
+        "por %%mm4, %%mm0               \n\t"
+        "movq %%mm0, %%mm4              \n\t"
         "punpcklbw %%mm7, %%mm0         \n\t"
-        "punpckhbw %%mm7, %%mm2         \n\t"
-        "paddw %%mm2, %%mm0             \n\t"
+        "punpckhbw %%mm7, %%mm4         \n\t"
         "paddw %%mm0, %%mm6             \n\t"
+        "paddw %%mm4, %%mm6             \n\t"
+        "movq  %%mm2, %%mm0             \n\t"
+        "movq  %%mm3, %%mm1             \n\t"
         "add %4, %%"REG_a"              \n\t"
         " js 1b                         \n\t"
         : "+a" (len)
@@ -255,6 +282,15 @@ static inline int sum_mmx2(void)
     return ret;
 }
 
+static inline void sad8_x2a_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
+{
+    sad8_2_mmx(blk1, blk1+1, blk2, stride, h);
+}
+static inline void sad8_y2a_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
+{
+    sad8_2_mmx(blk1, blk1+stride, blk2, stride, h);
+}
+
 
 #define PIX_SAD(suf)\
 static int sad8_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
@@ -272,11 +308,10 @@ static int sad8_x2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, in
     assert(h==8);\
     asm volatile("pxor %%mm7, %%mm7     \n\t"\
                  "pxor %%mm6, %%mm6     \n\t"\
-                 "movq %0, %%mm5        \n\t"\
                  :: "m"(round_tab[1]) \
                  );\
 \
-    sad8_2_ ## suf(blk1, blk1+1, blk2, stride, 8);\
+    sad8_x2a_ ## suf(blk1, blk2, stride, 8);\
 \
     return sum_ ## suf();\
 }\
@@ -290,7 +325,7 @@ static int sad8_y2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, in
                  :: "m"(round_tab[1]) \
                  );\
 \
-    sad8_2_ ## suf(blk1, blk1+stride, blk2, stride, 8);\
+    sad8_y2a_ ## suf(blk1, blk2, stride, 8);\
 \
     return sum_ ## suf();\
 }\
@@ -300,9 +335,7 @@ static int sad8_xy2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, i
     assert(h==8);\
     asm volatile("pxor %%mm7, %%mm7     \n\t"\
                  "pxor %%mm6, %%mm6     \n\t"\
-                 "movq %0, %%mm5        \n\t"\
-                 :: "m"(round_tab[2]) \
-                 );\
+                 ::);\
 \
     sad8_4_ ## suf(blk1, blk2, stride, 8);\
 \
@@ -327,8 +360,8 @@ static int sad16_x2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, i
                  :: "m"(round_tab[1]) \
                  );\
 \
-    sad8_2_ ## suf(blk1  , blk1+1, blk2  , stride, h);\
-    sad8_2_ ## suf(blk1+8, blk1+9, blk2+8, stride, h);\
+    sad8_x2a_ ## suf(blk1  , blk2  , stride, h);\
+    sad8_x2a_ ## suf(blk1+8, blk2+8, stride, h);\
 \
     return sum_ ## suf();\
 }\
@@ -340,8 +373,8 @@ static int sad16_y2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, i
                  :: "m"(round_tab[1]) \
                  );\
 \
-    sad8_2_ ## suf(blk1  , blk1+stride,  blk2  , stride, h);\
-    sad8_2_ ## suf(blk1+8, blk1+stride+8,blk2+8, stride, h);\
+    sad8_y2a_ ## suf(blk1  , blk2  , stride, h);\
+    sad8_y2a_ ## suf(blk1+8, blk2+8, stride, h);\
 \
     return sum_ ## suf();\
 }\
@@ -349,9 +382,7 @@ static int sad16_xy2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride,
 {\
     asm volatile("pxor %%mm7, %%mm7     \n\t"\
                  "pxor %%mm6, %%mm6     \n\t"\
-                 "movq %0, %%mm5        \n\t"\
-                 :: "m"(round_tab[2]) \
-                 );\
+                 ::);\
 \
     sad8_4_ ## suf(blk1  , blk2  , stride, h);\
     sad8_4_ ## suf(blk1+8, blk2+8, stride, h);\