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

[Zuxy Meng]

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:-)
).

-- 
Zuxy
Beauty is truth,
While truth is beauty.
PGP KeyID: E8555ED6
-------------- next part --------------
Index: libavcodec/i386/dsputil_mmx.c
===================================================================
--- libavcodec/i386/dsputil_mmx.c	?????? 8932??
+++ libavcodec/i386/dsputil_mmx.c	????????????
@@ -1537,6 +1537,13 @@
     "pmaxsw " #z ", " #a "            \n\t"\
     "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 TRANSPOSE4(a,b,c,d,t)\
     SBUTTERFLY(a,b,t,wd) /* a=aebf t=cgdh */\
     SBUTTERFLY(c,d,b,wd) /* c=imjn b=kolp */\
@@ -1727,9 +1734,90 @@
     return sum&0xFFFF;
 }
 
+static int hadamard8_diff_ssse3(void *s, uint8_t *src1, uint8_t *src2, int stride, int h){
+    DECLARE_ALIGNED_8(uint64_t, temp[16]);
+    int sum=0;
 
+    assert(h==8);
+
+    diff_pixels_mmx((DCTELEM*)temp, src1, src2, stride);
+
+    asm volatile(
+        LOAD4(0 , %%mm0, %%mm1, %%mm2, %%mm3)
+        LOAD4(64, %%mm4, %%mm5, %%mm6, %%mm7)
+
+        HADAMARD48
+
+        "movq %%mm7, 112(%1)            \n\t"
+
+        TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7)
+        STORE4(0 , %%mm0, %%mm3, %%mm7, %%mm2)
+
+        "movq 112(%1), %%mm7            \n\t"
+        TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0)
+        STORE4(64, %%mm4, %%mm7, %%mm0, %%mm6)
+
+        LOAD4(8 , %%mm0, %%mm1, %%mm2, %%mm3)
+        LOAD4(72, %%mm4, %%mm5, %%mm6, %%mm7)
+
+        HADAMARD48
+
+        "movq %%mm7, 120(%1)            \n\t"
+
+        TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7)
+        STORE4(8 , %%mm0, %%mm3, %%mm7, %%mm2)
+
+        "movq 120(%1), %%mm7            \n\t"
+        TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0)
+        "movq %%mm7, %%mm5              \n\t"//FIXME remove
+        "movq %%mm6, %%mm7              \n\t"
+        "movq %%mm0, %%mm6              \n\t"
+//        STORE4(72, %%mm4, %%mm7, %%mm0, %%mm6) //FIXME remove
+
+        LOAD4(64, %%mm0, %%mm1, %%mm2, %%mm3)
+//        LOAD4(72, %%mm4, %%mm5, %%mm6, %%mm7)
+
+        HADAMARD48
+        MMABS_SSSE3(%%mm0)
+        MMABS_SUM_SSSE3(%%mm1, %%mm0)
+        MMABS_SUM_SSSE3(%%mm2, %%mm0)
+        MMABS_SUM_SSSE3(%%mm3, %%mm0)
+        MMABS_SUM_SSSE3(%%mm4, %%mm0)
+        MMABS_SUM_SSSE3(%%mm5, %%mm0)
+        MMABS_SUM_SSSE3(%%mm6, %%mm0)
+        MMABS_SUM_SSSE3(%%mm7, %%mm0)
+        "movq %%mm0, 64(%1)             \n\t"
+
+        LOAD4(0 , %%mm0, %%mm1, %%mm2, %%mm3)
+        LOAD4(8 , %%mm4, %%mm5, %%mm6, %%mm7)
+
+        HADAMARD48
+        MMABS_SSSE3(%%mm0)
+        MMABS_SUM_SSSE3(%%mm1, %%mm0)
+        MMABS_SUM_SSSE3(%%mm2, %%mm0)
+        MMABS_SUM_SSSE3(%%mm3, %%mm0)
+        MMABS_SUM_SSSE3(%%mm4, %%mm0)
+        MMABS_SUM_SSSE3(%%mm5, %%mm0)
+        MMABS_SUM_SSSE3(%%mm6, %%mm0)
+        MMABS_SUM_SSSE3(%%mm7, %%mm0)
+        "pabsw 64(%1), %%mm1            \n\t"
+        "paddusw %%mm1, %%mm0           \n\t" // MMABS_SUM_SSSE3(64(%1), %%mm0)
+
+        "pshufw $0x0E, %%mm0, %%mm1     \n\t"
+        "paddusw %%mm1, %%mm0           \n\t"
+        "pshufw $0x01, %%mm0, %%mm1     \n\t"
+        "paddusw %%mm1, %%mm0           \n\t"
+        "movd %%mm0, %0                 \n\t"
+
+        : "=r" (sum)
+        : "r"(temp)
+    );
+    return sum&0xFFFF;
+}
+
 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;
@@ -3526,6 +3614,12 @@
         }
         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