prefetch pixels for future motion compensation. 2-5% faster h264.

Originally committed as revision 5203 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/dsputil.c

@@ -3773,6 +3773,8 @@ static void ff_jref_idct1_add(uint8_t *dest, int line_size, DCTELEM *block)
     dest[0] = cm[dest[0] + ((block[0] + 4)>>3)];
 }
 
+static void just_return() { return; }
+
 /* init static data */
 void dsputil_static_init(void)
 {
@@ -4054,6 +4056,8 @@ void dsputil_init(DSPContext* c, AVCodecContext *avctx)
     c->inner_add_yblock = ff_snow_inner_add_yblock;
 #endif
 
+    c->prefetch= just_return;
+
 #ifdef HAVE_MMX
     dsputil_init_mmx(c, avctx);
 #endif

libavcodec/dsputil.h

@@ -343,6 +343,8 @@ typedef struct DSPContext {
     void (*vertical_compose97i)(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTELEM *b3, DWTELEM *b4, DWTELEM *b5, int width);
     void (*horizontal_compose97i)(DWTELEM *b, int width);
     void (*inner_add_yblock)(uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h, int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8);
+
+    void (*prefetch)(void *mem, int stride, int h);
 } DSPContext;
 
 void dsputil_static_init(void);

libavcodec/h264.c

@@ -2752,6 +2752,22 @@ static inline void mc_part(H264Context *h, int n, int square, int chroma_height,
                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg, chroma_avg, list0, list1);
 }
 
+static inline void prefetch_motion(H264Context *h, int list){
+    /* fetch pixels for estimated mv 4 macroblocks ahead
+     * optimized for 64byte cache lines */
+    MpegEncContext * const s = &h->s;
+    const int refn = h->ref_cache[list][scan8[0]];
+    if(refn >= 0){
+        const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
+        const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
+        uint8_t **src= h->ref_list[list][refn].data;
+        int off= mx + (my + (s->mb_x&3)*4)*s->linesize + 64;
+        s->dsp.prefetch(src[0]+off, s->linesize, 4);
+        off= (mx>>1) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + 64;
+        s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
+    }
+}
+
 static void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
@@ -2762,6 +2778,8 @@ static void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t
 
     assert(IS_INTER(mb_type));
 
+    prefetch_motion(h, 0);
+
     if(IS_16X16(mb_type)){
         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
@@ -2833,6 +2851,8 @@ static void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t
             }
         }
     }
+
+    prefetch_motion(h, 1);
 }
 
 static void decode_init_vlc(H264Context *h){

libavcodec/i386/dsputil_mmx.c

@@ -2489,6 +2489,18 @@ static void add_8x8basis_mmx(int16_t rem[64], int16_t basis[64], int scale){
     }
 }
 
+#define PREFETCH(name, op) \
+void name(void *mem, int stride, int h){\
+    const uint8_t *p= mem;\
+    do{\
+        asm volatile(#op" %0" :: "m"(*p));\
+        p+= stride;\
+    }while(--h);\
+}
+PREFETCH(prefetch_mmx2,  prefetcht0)
+PREFETCH(prefetch_3dnow, prefetch)
+#undef PREFETCH
+
 #include "h264dsp_mmx.c"
 
 /* external functions, from idct_mmx.c */
@@ -2749,6 +2761,8 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
         c->h264_idct8_add= ff_h264_idct8_add_mmx;
 
         if (mm_flags & MM_MMXEXT) {
+            c->prefetch = prefetch_mmx2;
+
             c->put_pixels_tab[0][1] = put_pixels16_x2_mmx2;
             c->put_pixels_tab[0][2] = put_pixels16_y2_mmx2;
 
@@ -2879,6 +2893,8 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
             c->sub_hfyu_median_prediction= sub_hfyu_median_prediction_mmx2;
 #endif //CONFIG_ENCODERS
         } else if (mm_flags & MM_3DNOW) {
+            c->prefetch = prefetch_3dnow;
+
             c->put_pixels_tab[0][1] = put_pixels16_x2_3dnow;
             c->put_pixels_tab[0][2] = put_pixels16_y2_3dnow;