factor sum_abs_dctelem out of dct_sad, and simd it.

 sum_abs_dctelem_* alone:
core2: c=186 mmx2=39 sse2=21 ssse3=13 (cycles)
k8: c=163 mmx2=33 sse2=31
p4: c=370 mmx2=60 sse2=60
 dct_sad including sum_abs_dctelem_*:
core2: c=405 mmx2=258 sse2=240 ssse3=232
k8: c=624 mmx2=394 sse2=392
p4: c=849 mmx2=556 sse2=556

Originally committed as revision 9001 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/dsputil.c

@@ -592,6 +592,14 @@ static void add_pixels4_c(uint8_t *restrict pixels, DCTELEM *block, int line_siz
     }
 }
 
+static int sum_abs_dctelem_c(DCTELEM *block)
+{
+    int sum=0, i;
+    for(i=0; i<64; i++)
+        sum+= FFABS(block[i]);
+    return sum;
+}
+
 #if 0
 
 #define PIXOP2(OPNAME, OP) \
@@ -3385,19 +3393,14 @@ static int hadamard8_intra8x8_c(/*MpegEncContext*/ void *s, uint8_t *src, uint8_
 
 static int dct_sad8x8_c(/*MpegEncContext*/ void *c, uint8_t *src1, uint8_t *src2, int stride, int h){
     MpegEncContext * const s= (MpegEncContext *)c;
-    DECLARE_ALIGNED_8(uint64_t, aligned_temp[sizeof(DCTELEM)*64/8]);
+    DECLARE_ALIGNED_16(uint64_t, aligned_temp[sizeof(DCTELEM)*64/8]);
     DCTELEM * const temp= (DCTELEM*)aligned_temp;
-    int sum=0, i;
 
     assert(h==8);
 
     s->dsp.diff_pixels(temp, src1, src2, stride);
     s->dsp.fdct(temp);
-
-    for(i=0; i<64; i++)
-        sum+= FFABS(temp[i]);
-
-    return sum;
+    return s->dsp.sum_abs_dctelem(temp);
 }
 
 #ifdef CONFIG_GPL
@@ -3905,6 +3908,7 @@ void dsputil_init(DSPContext* c, AVCodecContext *avctx)
     c->add_pixels_clamped = add_pixels_clamped_c;
     c->add_pixels8 = add_pixels8_c;
     c->add_pixels4 = add_pixels4_c;
+    c->sum_abs_dctelem = sum_abs_dctelem_c;
     c->gmc1 = gmc1_c;
     c->gmc = ff_gmc_c;
     c->clear_blocks = clear_blocks_c;

libavcodec/dsputil.h

@@ -163,6 +163,7 @@ typedef struct DSPContext {
     void (*add_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
     void (*add_pixels8)(uint8_t *pixels, DCTELEM *block, int line_size);
     void (*add_pixels4)(uint8_t *pixels, DCTELEM *block, int line_size);
+    int (*sum_abs_dctelem)(DCTELEM *block/*align 16*/);
     /**
      * translational global motion compensation.
      */

libavcodec/i386/dsputil_mmx.c

@@ -1649,6 +1649,9 @@ static void sub_hfyu_median_prediction_mmx2(uint8_t *dst, uint8_t *src1, uint8_t
     "movq "#c", "#o"+16(%1)           \n\t"\
     "movq "#d", "#o"+24(%1)           \n\t"\
 
+/* FIXME: HSUM_* saturates at 64k, while an 8x8 hadamard or dct block can get up to
+ * about 100k on extreme inputs. But that's very unlikely to occur in natural video,
+ * and it's even more unlikely to not have any alternative mvs/modes with lower cost. */
 #define HSUM_MMX(a, t, dst)\
     "movq "#a", "#t"                  \n\t"\
     "psrlq $32, "#a"                  \n\t"\
@@ -1802,6 +1805,71 @@ HADAMARD8_DIFF_SSE2(ssse3)
 #undef MMABS_SUM_8x8
 #endif
 
+#define DCT_SAD4(m,mm,o)\
+    "mov"#m" "#o"+ 0(%1), "#mm"2      \n\t"\
+    "mov"#m" "#o"+16(%1), "#mm"3      \n\t"\
+    "mov"#m" "#o"+32(%1), "#mm"4      \n\t"\
+    "mov"#m" "#o"+48(%1), "#mm"5      \n\t"\
+    MMABS_SUM(mm##2, mm##6, mm##0)\
+    MMABS_SUM(mm##3, mm##7, mm##1)\
+    MMABS_SUM(mm##4, mm##6, mm##0)\
+    MMABS_SUM(mm##5, mm##7, mm##1)\
+
+#define DCT_SAD_MMX\
+    "pxor %%mm0, %%mm0                \n\t"\
+    "pxor %%mm1, %%mm1                \n\t"\
+    DCT_SAD4(q, %%mm, 0)\
+    DCT_SAD4(q, %%mm, 8)\
+    DCT_SAD4(q, %%mm, 64)\
+    DCT_SAD4(q, %%mm, 72)\
+    "paddusw %%mm1, %%mm0             \n\t"\
+    HSUM(%%mm0, %%mm1, %0)
+
+#define DCT_SAD_SSE2\
+    "pxor %%xmm0, %%xmm0              \n\t"\
+    "pxor %%xmm1, %%xmm1              \n\t"\
+    DCT_SAD4(dqa, %%xmm, 0)\
+    DCT_SAD4(dqa, %%xmm, 64)\
+    "paddusw %%xmm1, %%xmm0           \n\t"\
+    HSUM(%%xmm0, %%xmm1, %0)
+
+#define DCT_SAD_FUNC(cpu) \
+static int sum_abs_dctelem_##cpu(DCTELEM *block){\
+    int sum;\
+    asm volatile(\
+        DCT_SAD\
+        :"=r"(sum)\
+        :"r"(block)\
+    );\
+    return sum&0xFFFF;\
+}
+
+#define DCT_SAD       DCT_SAD_MMX
+#define HSUM(a,t,dst) HSUM_MMX(a,t,dst)
+#define MMABS(a,z)    MMABS_MMX(a,z)
+DCT_SAD_FUNC(mmx)
+#undef MMABS
+#undef HSUM
+
+#define HSUM(a,t,dst) HSUM_MMX2(a,t,dst)
+#define MMABS(a,z)    MMABS_MMX2(a,z)
+DCT_SAD_FUNC(mmx2)
+#undef HSUM
+#undef DCT_SAD
+
+#define DCT_SAD       DCT_SAD_SSE2
+#define HSUM(a,t,dst) HSUM_SSE2(a,t,dst)
+DCT_SAD_FUNC(sse2)
+#undef MMABS
+
+#ifdef HAVE_SSSE3
+#define MMABS(a,z)    MMABS_SSSE3(a,z)
+DCT_SAD_FUNC(ssse3)
+#undef MMABS
+#endif
+#undef HSUM
+#undef DCT_SAD
+
 static int ssd_int8_vs_int16_mmx(int8_t *pix1, int16_t *pix2, int size){
     int sum;
     long i=size;
@@ -3298,6 +3366,7 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
         c->add_bytes= add_bytes_mmx;
 #ifdef CONFIG_ENCODERS
         c->diff_bytes= diff_bytes_mmx;
+        c->sum_abs_dctelem= sum_abs_dctelem_mmx;
 
         c->hadamard8_diff[0]= hadamard8_diff16_mmx;
         c->hadamard8_diff[1]= hadamard8_diff_mmx;
@@ -3350,6 +3419,7 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
             c->avg_pixels_tab[1][2] = avg_pixels8_y2_mmx2;
 
 #ifdef CONFIG_ENCODERS
+            c->sum_abs_dctelem= sum_abs_dctelem_mmx2;
             c->hadamard8_diff[0]= hadamard8_diff16_mmx2;
             c->hadamard8_diff[1]= hadamard8_diff_mmx2;
             c->vsad[4]= vsad_intra16_mmx2;
@@ -3569,12 +3639,14 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
 
 #ifdef CONFIG_ENCODERS
         if(mm_flags & MM_SSE2){
+            c->sum_abs_dctelem= sum_abs_dctelem_sse2;
             c->hadamard8_diff[0]= hadamard8_diff16_sse2;
             c->hadamard8_diff[1]= hadamard8_diff_sse2;
         }
 
 #ifdef HAVE_SSSE3
         if(mm_flags & MM_SSSE3){
+            c->sum_abs_dctelem= sum_abs_dctelem_ssse3;
             c->hadamard8_diff[0]= hadamard8_diff16_ssse3;
             c->hadamard8_diff[1]= hadamard8_diff_ssse3;
         }