sse2 16x16 sum squared diff (306=>268 cycles on a K8)

faster 8x8 mmx ssd (77=>70 cycles)

Originally committed as revision 4623 to svn://svn.ffmpeg.org/ffmpeg/trunk

diff --git a/libavcodec/i386/dsputil_mmx.c b/libavcodec/i386/dsputil_mmx.c
index 51ed07cd1a22f37cbf1a2e2340919d4b23925362..12d9900f66cdc738337bc5bf3ffc688f05320b03 100644 (file)
--- a/libavcodec/i386/dsputil_mmx.c
+++ b/libavcodec/i386/dsputil_mmx.c
@@ -744,31 +744,49 @@ static int sse8_mmx(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int
     int tmp;
   asm volatile (
       "movl %4,%%ecx\n"
+      "shr $1,%%ecx\n"
       "pxor %%mm0,%%mm0\n"     /* mm0 = 0 */
       "pxor %%mm7,%%mm7\n"     /* mm7 holds the sum */
       "1:\n"
-      "movq (%0),%%mm1\n"      /* mm1 = pix1[0-7] */
-      "movq (%1),%%mm2\n"      /* mm2 = pix2[0-7] */
+      "movq (%0),%%mm1\n"      /* mm1 = pix1[0][0-7] */
+      "movq (%1),%%mm2\n"      /* mm2 = pix2[0][0-7] */
+      "movq (%0,%3),%%mm3\n"   /* mm3 = pix1[1][0-7] */
+      "movq (%1,%3),%%mm4\n"   /* mm4 = pix2[1][0-7] */
 
+      /* todo: mm1-mm2, mm3-mm4 */
+      /* algo: substract mm1 from mm2 with saturation and vice versa */
+      /*       OR the results to get absolute difference */
       "movq %%mm1,%%mm5\n"
+      "movq %%mm3,%%mm6\n"
       "psubusb %%mm2,%%mm1\n"
+      "psubusb %%mm4,%%mm3\n"
       "psubusb %%mm5,%%mm2\n"
+      "psubusb %%mm6,%%mm4\n"
 
       "por %%mm1,%%mm2\n"
+      "por %%mm3,%%mm4\n"
 
+      /* now convert to 16-bit vectors so we can square them */
       "movq %%mm2,%%mm1\n"
+      "movq %%mm4,%%mm3\n"
 
       "punpckhbw %%mm0,%%mm2\n"
+      "punpckhbw %%mm0,%%mm4\n"
       "punpcklbw %%mm0,%%mm1\n"        /* mm1 now spread over (mm1,mm2) */
+      "punpcklbw %%mm0,%%mm3\n"        /* mm4 now spread over (mm3,mm4) */
 
       "pmaddwd %%mm2,%%mm2\n"
+      "pmaddwd %%mm4,%%mm4\n"
       "pmaddwd %%mm1,%%mm1\n"
+      "pmaddwd %%mm3,%%mm3\n"
 
-      "add %3,%0\n"
-      "add %3,%1\n"
+      "lea (%0,%3,2), %0\n"    /* pix1 += 2*line_size */
+      "lea (%1,%3,2), %1\n"    /* pix2 += 2*line_size */
 
       "paddd %%mm2,%%mm1\n"
+      "paddd %%mm4,%%mm3\n"
       "paddd %%mm1,%%mm7\n"
+      "paddd %%mm3,%%mm7\n"
 
       "decl %%ecx\n"
       "jnz 1b\n"
@@ -843,6 +861,68 @@ static int sse16_mmx(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int
     return tmp;
 }
 
+static int sse16_sse2(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) {
+    int tmp;
+  asm volatile (
+      "shr $1,%2\n"
+      "pxor %%xmm0,%%xmm0\n"   /* mm0 = 0 */
+      "pxor %%xmm7,%%xmm7\n"   /* mm7 holds the sum */
+      "1:\n"
+      "movdqu (%0),%%xmm1\n"   /* mm1 = pix1[0][0-15] */
+      "movdqu (%1),%%xmm2\n"   /* mm2 = pix2[0][0-15] */
+      "movdqu (%0,%4),%%xmm3\n"        /* mm3 = pix1[1][0-15] */
+      "movdqu (%1,%4),%%xmm4\n"        /* mm4 = pix2[1][0-15] */
+
+      /* todo: mm1-mm2, mm3-mm4 */
+      /* algo: substract mm1 from mm2 with saturation and vice versa */
+      /*       OR the results to get absolute difference */
+      "movdqa %%xmm1,%%xmm5\n"
+      "movdqa %%xmm3,%%xmm6\n"
+      "psubusb %%xmm2,%%xmm1\n"
+      "psubusb %%xmm4,%%xmm3\n"
+      "psubusb %%xmm5,%%xmm2\n"
+      "psubusb %%xmm6,%%xmm4\n"
+
+      "por %%xmm1,%%xmm2\n"
+      "por %%xmm3,%%xmm4\n"
+
+      /* now convert to 16-bit vectors so we can square them */
+      "movdqa %%xmm2,%%xmm1\n"
+      "movdqa %%xmm4,%%xmm3\n"
+
+      "punpckhbw %%xmm0,%%xmm2\n"
+      "punpckhbw %%xmm0,%%xmm4\n"
+      "punpcklbw %%xmm0,%%xmm1\n"      /* mm1 now spread over (mm1,mm2) */
+      "punpcklbw %%xmm0,%%xmm3\n"      /* mm4 now spread over (mm3,mm4) */
+
+      "pmaddwd %%xmm2,%%xmm2\n"
+      "pmaddwd %%xmm4,%%xmm4\n"
+      "pmaddwd %%xmm1,%%xmm1\n"
+      "pmaddwd %%xmm3,%%xmm3\n"
+
+      "lea (%0,%4,2), %0\n"    /* pix1 += 2*line_size */
+      "lea (%1,%4,2), %1\n"    /* pix2 += 2*line_size */
+
+      "paddd %%xmm2,%%xmm1\n"
+      "paddd %%xmm4,%%xmm3\n"
+      "paddd %%xmm1,%%xmm7\n"
+      "paddd %%xmm3,%%xmm7\n"
+
+      "decl %2\n"
+      "jnz 1b\n"
+
+      "movdqa %%xmm7,%%xmm1\n"
+      "psrldq $8, %%xmm7\n"    /* shift hi qword to lo */
+      "paddd %%xmm1,%%xmm7\n"
+      "movdqa %%xmm7,%%xmm1\n"
+      "psrldq $4, %%xmm7\n"    /* shift hi dword to lo */
+      "paddd %%xmm1,%%xmm7\n"
+      "movd %%xmm7,%3\n"
+      : "+r" (pix1), "+r" (pix2), "+r"(h), "=r"(tmp) 
+      : "r" ((long)line_size));
+    return tmp;
+}
+
 static int hf_noise8_mmx(uint8_t * pix1, int line_size, int h) {
     int tmp;
   asm volatile (
@@ -2626,7 +2706,7 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
         c->hadamard8_diff[1]= hadamard8_diff_mmx;
         
        c->pix_norm1 = pix_norm1_mmx;
-       c->sse[0] = sse16_mmx;
+       c->sse[0] = (mm_flags & MM_SSE2) ? sse16_sse2 : sse16_mmx;
        c->sse[1] = sse8_mmx;
         c->vsad[4]= vsad_intra16_mmx;