Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.

Originally committed as revision 5427 to svn://svn.ffmpeg.org/ffmpeg/trunk

3 files changed, 45 insertions, 35 deletions

diff --git a/libavcodec/dsputil.c b/libavcodec/dsputil.c
index ad1bfd482f..a3b7cf3c64 100644
--- a/libavcodec/dsputil.c
+++ b/libavcodec/dsputil.c
@@ -293,34 +293,33 @@ static int sse16_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
 
 
 static inline int w_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int w, int h, int type){
-#ifdef CONFIG_SNOW_ENCODER //idwt is in snow.c
+#ifdef CONFIG_SNOW_ENCODER //dwt is in snow.c
     int s, i, j;
     const int dec_count= w==8 ? 3 : 4;
-    int tmp[16*16];
-#if 0
+    int tmp[32*32];
     int level, ori;
     static const int scale[2][2][4][4]={
       {
         {
-            //8x8 dec=3
+            // 9/7 8x8 dec=3
             {268, 239, 239, 213},
             {  0, 224, 224, 152},
             {  0, 135, 135, 110},
         },{
-            //16x16 dec=4
+            // 9/7 16x16 or 32x32 dec=4
             {344, 310, 310, 280},
             {  0, 320, 320, 228},
             {  0, 175, 175, 136},
             {  0, 129, 129, 102},
         }
       },{
-        {//FIXME 5/3
-            //8x8 dec=3
+        {
+            // 5/3 8x8 dec=3
             {275, 245, 245, 218},
             {  0, 230, 230, 156},
             {  0, 138, 138, 113},
         },{
-            //16x16 dec=4
+            // 5/3 16x16 or 32x32 dec=4
             {352, 317, 317, 286},
             {  0, 328, 328, 233},
             {  0, 180, 180, 140},
@@ -328,29 +327,28 @@ static inline int w_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, in
         }
       }
     };
-#endif
 
     for (i = 0; i < h; i++) {
         for (j = 0; j < w; j+=4) {
-            tmp[16*i+j+0] = (pix1[j+0] - pix2[j+0])<<4;
-            tmp[16*i+j+1] = (pix1[j+1] - pix2[j+1])<<4;
-            tmp[16*i+j+2] = (pix1[j+2] - pix2[j+2])<<4;
-            tmp[16*i+j+3] = (pix1[j+3] - pix2[j+3])<<4;
+            tmp[32*i+j+0] = (pix1[j+0] - pix2[j+0])<<4;
+            tmp[32*i+j+1] = (pix1[j+1] - pix2[j+1])<<4;
+            tmp[32*i+j+2] = (pix1[j+2] - pix2[j+2])<<4;
+            tmp[32*i+j+3] = (pix1[j+3] - pix2[j+3])<<4;
         }
         pix1 += line_size;
         pix2 += line_size;
     }
 
-    ff_spatial_dwt(tmp, w, h, 16, type, dec_count);
+    ff_spatial_dwt(tmp, w, h, 32, type, dec_count);
 
     s=0;
-#if 0
+    assert(w==h);
     for(level=0; level<dec_count; level++){
         for(ori= level ? 1 : 0; ori<4; ori++){
-            int sx= (ori&1) ? 1<<level: 0;
-            int stride= 16<<(dec_count-level);
+            int size= w>>(dec_count-level);
+            int sx= (ori&1) ? size : 0;
+            int stride= 32<<(dec_count-level);
             int sy= (ori&2) ? stride>>1 : 0;
-            int size= 1<<level;
 
             for(i=0; i<size; i++){
                 for(j=0; j<size; j++){
@@ -360,18 +358,8 @@ static inline int w_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, in
             }
         }
     }
-#endif
-    for (i = 0; i < h; i++) {
-        for (j = 0; j < w; j+=4) {
-            s+= ABS(tmp[16*i+j+0]);
-            s+= ABS(tmp[16*i+j+1]);
-            s+= ABS(tmp[16*i+j+2]);
-            s+= ABS(tmp[16*i+j+3]);
-        }
-    }
     assert(s>=0);
-
-    return s>>2;
+    return s>>9;
 #endif
 }
 
@@ -391,6 +379,14 @@ static int w97_16_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int
     return w_c(v, pix1, pix2, line_size, 16, h, 0);
 }
 
+int w53_32_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h){
+    return w_c(v, pix1, pix2, line_size, 32, h, 1);
+}
+
+int w97_32_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h){
+    return w_c(v, pix1, pix2, line_size, 32, h, 0);
+}
+
 static void get_pixels_c(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
 {
     int i;
diff --git a/libavcodec/snow.c b/libavcodec/snow.c
index daac0107e6..9dce5aea0b 100644
--- a/libavcodec/snow.c
+++ b/libavcodec/snow.c
@@ -3092,12 +3092,23 @@ static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, con
             memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
     }
 
-    //FIXME sad/ssd can be broken up, but wavelet cmp should be one 32x32 block
     if(block_w==16){
-        distortion = 0;
-        for(i=0; i<4; i++){
-            int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
-            distortion += s->dsp.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
+        /* FIXME rearrange dsputil to fit 32x32 cmp functions */
+        /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
+        /* FIXME cmps overlap but don't cover the wavelet's whole support,
+         * so improving the score of one block is not strictly guaranteed to
+         * improve the score of the whole frame, so iterative motion est
+         * doesn't always converge. */
+        if(s->avctx->me_cmp == FF_CMP_W97)
+            distortion = w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
+        else if(s->avctx->me_cmp == FF_CMP_W53)
+            distortion = w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
+        else{
+            distortion = 0;
+            for(i=0; i<4; i++){
+                int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
+                distortion += s->dsp.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
+            }
         }
     }else{
         assert(block_w==8);
@@ -3282,7 +3293,7 @@ static void iterative_me(SnowContext *s){
         memcpy(s->block_state, state, sizeof(s->block_state));
     }
 
-    for(pass=0; pass<50; pass++){
+    for(pass=0; pass<25; pass++){
         int change= 0;
 
         for(mb_y= 0; mb_y<b_height; mb_y++){
diff --git a/libavcodec/snow.h b/libavcodec/snow.h
index 99b04f7a46..4704177be9 100644
--- a/libavcodec/snow.h
+++ b/libavcodec/snow.h
@@ -121,6 +121,9 @@ extern void ff_snow_vertical_compose97i(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, D
 extern void ff_snow_horizontal_compose97i(DWTELEM *b, int width);
 extern void ff_snow_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);
 
+int w53_32_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h);
+int w97_32_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h);
+
 
 /* C bits used by mmx/sse2/altivec */