fix motion compensation with (x+1/2,y+1/2) MVs

Originally committed as revision 2849 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/dsputil.c

@@ -874,6 +874,13 @@ PIXOP2(put, op_put)
 #define avg2(a,b) ((a+b+1)>>1)
 #define avg4(a,b,c,d) ((a+b+c+d+2)>>2)
 
+static void put_no_rnd_pixels16_l2_c(uint8_t *dst, const uint8_t *a, const uint8_t *b, int stride, int h){
+    put_no_rnd_pixels16_l2(dst, a, b, stride, stride, stride, h);
+}
+
+static void put_no_rnd_pixels8_l2_c(uint8_t *dst, const uint8_t *a, const uint8_t *b, int stride, int h){
+    put_no_rnd_pixels8_l2(dst, a, b, stride, stride, stride, h);
+}
 
 static void gmc1_c(uint8_t *dst, uint8_t *src, int stride, int h, int x16, int y16, int rounder)
 {
@@ -3158,6 +3165,9 @@ void dsputil_init(DSPContext* c, AVCodecContext *avctx)
     dspfunc(avg, 3, 2);
 #undef dspfunc
 
+    c->put_no_rnd_pixels_l2[0]= put_no_rnd_pixels16_l2_c;
+    c->put_no_rnd_pixels_l2[1]= put_no_rnd_pixels8_l2_c;
+
     c->put_tpel_pixels_tab[ 0] = put_tpel_pixels_mc00_c;
     c->put_tpel_pixels_tab[ 1] = put_tpel_pixels_mc10_c;
     c->put_tpel_pixels_tab[ 2] = put_tpel_pixels_mc20_c;

libavcodec/dsputil.h

@@ -204,6 +204,8 @@ typedef struct DSPContext {
      */
     op_pixels_func avg_no_rnd_pixels_tab[2][4];
     
+    void (*put_no_rnd_pixels_l2[2])(uint8_t *block/*align width (8 or 16)*/, const uint8_t *a/*align 1*/, const uint8_t *b/*align 1*/, int line_size, int h);
+    
     /**
      * Thirdpel motion compensation with rounding (a+b+1)>>1.
      * this is an array[12] of motion compensation funcions for the 9 thirdpel positions<br>

libavcodec/vp3.c

@@ -2452,15 +2452,24 @@ av_log(s->avctx, AV_LOG_ERROR, " help! got beefy vector! (%X, %X)\n", motion_x,
                     }
                 }
                 
+
                 /* first, take care of copying a block from either the
                  * previous or the golden frame */
                 if (s->all_fragments[i].coding_method != MODE_INTRA) {
-
+                    //Note, it is possible to implement all MC cases with put_no_rnd_pixels_l2 which would look more like the VP3 source but this would be slower as put_no_rnd_pixels_tab is better optimzed
+                    if(motion_halfpel_index != 3){
                         s->dsp.put_no_rnd_pixels_tab[1][motion_halfpel_index](
                             output_plane + s->all_fragments[i].first_pixel,
-                        motion_source,
+                            motion_source, stride, 8);
+                    }else{
+                        int d= (motion_x ^ motion_y)>>31; // d is 0 if motion_x and _y have the same sign, else -1
+                        s->dsp.put_no_rnd_pixels_l2[1](
+                            output_plane + s->all_fragments[i].first_pixel,
+                            motion_source - d, 
+                            motion_source + stride + 1 + d, 
                             stride, 8);
                     }
+                }
 
                 /* dequantize the DCT coefficients */
                 debug_idct("fragment %d, coding mode %d, DC = %d, dequant = %d:\n",