[FFmpeg-devel] Anybody has a Core 2? [PATCH] Small SSSE3 optimization

[Zuxy Meng]

Hi,

2007/5/8, Zuxy Meng <zuxy.meng at gmail.com>:
> Hi,
>
> Attached patch makes use of SSSE3 instruction pabsw to calculate the
> absolute value of packed words. Just for fun. And I don't have a SSSE3
> capable CPU so hopefully someone with a Core 2 can help test it to
> ensure it doesn't break anything (better with benchmarks of course:-)
> ).


Updated patch against curren SVN HEAD. Full test passed on MMX2. Of
course it still needs testing under Core 2.

-- 
Zuxy
Beauty is truth,
While truth is beauty.
PGP KeyID: E8555ED6
-------------- next part --------------
Index: libavcodec/i386/dsputil_mmx.c
===================================================================
--- libavcodec/i386/dsputil_mmx.c	?????? 8946??
+++ libavcodec/i386/dsputil_mmx.c	????????????
@@ -1538,6 +1538,13 @@
     MMABS_MMX2(a,z)\
     "paddusw " #a ", " #sum "         \n\t"
 
+#define MMABS_SSSE3(a)\
+    "pabsw " #a ", " #a "             \n\t"
+
+#define MMABS_SUM_SSSE3(a, sum)\
+    "pabsw " #a ", " #a "             \n\t"\
+    "paddusw " #a ", " #sum "         \n\t"
+
 #define LOAD4(o, a, b, c, d)\
         "movq "#o"(%1), " #a "        \n\t"\
         "movq "#o"+16(%1), " #b "     \n\t"\
@@ -1609,7 +1616,7 @@
         LOAD4(64, %%mm0, %%mm1, %%mm2, %%mm3)\
 \
         HADAMARD48\
-        "movq %%mm7, 64(%1)             \n\t"\
+        SAVE_MM7(64)\
         MMABS(%%mm0, %%mm7)\
         MMABS_SUM(%%mm1, %%mm7, %%mm0)\
         MMABS_SUM(%%mm2, %%mm7, %%mm0)\
@@ -1617,15 +1624,14 @@
         MMABS_SUM(%%mm4, %%mm7, %%mm0)\
         MMABS_SUM(%%mm5, %%mm7, %%mm0)\
         MMABS_SUM(%%mm6, %%mm7, %%mm0)\
-        "movq 64(%1), %%mm1             \n\t"\
-        MMABS_SUM(%%mm1, %%mm7, %%mm0)\
+        RESTORE_MM7(64)\
         "movq %%mm0, 64(%1)             \n\t"\
 \
         LOAD4(0 , %%mm0, %%mm1, %%mm2, %%mm3)\
         LOAD4(8 , %%mm4, %%mm5, %%mm6, %%mm7)\
 \
         HADAMARD48\
-        "movq %%mm7, (%1)               \n\t"\
+        SAVE_MM7(0)\
         MMABS(%%mm0, %%mm7)\
         MMABS_SUM(%%mm1, %%mm7, %%mm0)\
         MMABS_SUM(%%mm2, %%mm7, %%mm0)\
@@ -1633,10 +1639,8 @@
         MMABS_SUM(%%mm4, %%mm7, %%mm0)\
         MMABS_SUM(%%mm5, %%mm7, %%mm0)\
         MMABS_SUM(%%mm6, %%mm7, %%mm0)\
-        "movq (%1), %%mm1               \n\t"\
-        MMABS_SUM(%%mm1, %%mm7, %%mm0)\
-        "movq 64(%1), %%mm1             \n\t"\
-        MMABS_SUM(%%mm1, %%mm7, %%mm0)\
+        RESTORE_MM7(0)\
+        "paddusw 64(%1), %%mm0          \n\t"\
 \
         HSUM(%%mm0, %%mm1, %0)\
 \
@@ -1646,6 +1650,11 @@
     return sum&0xFFFF;\
 }
 
+#define SAVE_MM7(x)\
+    "movq %%mm7, "#x"(%1)             \n\t"
+#define RESTORE_MM7(x)\
+    "movq "#x"(%1), %%mm1             \n\t"\
+    MMABS_SUM(%%mm1, %%mm7, %%mm0)
 #define MMABS(a,z)         MMABS_MMX(a,z)
 #define MMABS_SUM(a,z,sum) MMABS_SUM_MMX(a,z,sum)
 #define HSUM(a,t,dst)      HSUM_MMX(a,t,dst)
@@ -1661,9 +1670,24 @@
 #undef MMABS
 #undef MMABS_SUM
 #undef HSUM
+#undef SAVE_MM7
+#undef RESTORE_MM7
 
+#define SAVE_MM7(x)
+#define RESTORE_MM7(x)     MMABS_SUM_SSSE3(%%mm7, %%mm0)
+#define MMABS(a,z)         MMABS_SSSE3(a)
+#define MMABS_SUM(a,z,sum) MMABS_SUM_SSSE3(a,sum)
+#define HSUM(a,t,dst)      HSUM_MMX2(a,t,dst)
+HADAMARD8_DIFF_MMX(ssse3)
+#undef MMABS
+#undef MMABS_SUM
+#undef HSUM
+#undef SAVE_MM7
+#undef RESTORE_MM7
+
 WARPER8_16_SQ(hadamard8_diff_mmx, hadamard8_diff16_mmx)
 WARPER8_16_SQ(hadamard8_diff_mmx2, hadamard8_diff16_mmx2)
+WARPER8_16_SQ(hadamard8_diff_ssse3, hadamard8_diff16_ssse3)
 
 static int ssd_int8_vs_int16_mmx(int8_t *pix1, int16_t *pix2, int size){
     int sum;
@@ -3460,6 +3484,12 @@
         }
         if(mm_flags & MM_3DNOW)
             c->vector_fmul_add_add = vector_fmul_add_add_3dnow; // faster than sse
+#ifdef CONFIG_ENCODERS
+        if(mm_flags & MM_SSSE3){
+            c->hadamard8_diff[0]= hadamard8_diff16_ssse3;
+            c->hadamard8_diff[1]= hadamard8_diff_ssse3;
+        }
+#endif //CONFIG_ENCODERS
     }
 
 #ifdef CONFIG_ENCODERS