4x4 SSE compare function

wavelet based compare functions
make epzs_motion_search() more flexible so it can be used for a wider range of block sizes
make get_penalty_factor() independant of MpegEncContext

Originally committed as revision 3410 to svn://svn.ffmpeg.org/ffmpeg/trunk

5 files changed, 175 insertions, 44 deletions

diff --git a/libavcodec/avcodec.h b/libavcodec/avcodec.h
index 64218a099c..e2da0f85a2 100644
--- a/libavcodec/avcodec.h
+++ b/libavcodec/avcodec.h
@@ -1262,6 +1262,8 @@ typedef struct AVCodecContext {
 #define FF_CMP_VSAD 8
 #define FF_CMP_VSSE 9
 #define FF_CMP_NSSE 10
+#define FF_CMP_W53  11
+#define FF_CMP_W97  12
 #define FF_CMP_CHROMA 256
     
     /**
diff --git a/libavcodec/dsputil.c b/libavcodec/dsputil.c
index c4b627b683..0fa95fffe8 100644
--- a/libavcodec/dsputil.c
+++ b/libavcodec/dsputil.c
@@ -219,6 +219,23 @@ static void bswap_buf(uint32_t *dst, uint32_t *src, int w){
     }
 }
 
+static int sse4_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h)
+{
+    int s, i;
+    uint32_t *sq = squareTbl + 256;
+
+    s = 0;
+    for (i = 0; i < h; i++) {
+        s += sq[pix1[0] - pix2[0]];
+        s += sq[pix1[1] - pix2[1]];
+        s += sq[pix1[2] - pix2[2]];
+        s += sq[pix1[3] - pix2[3]];
+        pix1 += line_size;
+        pix2 += line_size;
+    }
+    return s;
+}
+
 static int sse8_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h)
 {
     int s, i;
@@ -270,6 +287,103 @@ static int sse16_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
     return s;
 }
 
+
+static inline int w_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int w, int h, int type){
+    int s, i, j;
+    const int dec_count= w==8 ? 3 : 4;
+    int tmp[16*16];
+#if 0
+    int level, ori;
+    static const int scale[2][2][4][4]={ 
+      {
+        {
+            //8x8 dec=3
+            {268, 239, 239, 213},
+            {  0, 224, 224, 152},
+            {  0, 135, 135, 110},
+        },{
+            //16x16 dec=4
+            {344, 310, 310, 280},
+            {  0, 320, 320, 228},
+            {  0, 175, 175, 136},
+            {  0, 129, 129, 102},
+        }
+      },{
+        {//FIXME 5/3
+            //8x8 dec=3
+            {275, 245, 245, 218},
+            {  0, 230, 230, 156},
+            {  0, 138, 138, 113},
+        },{
+            //16x16 dec=4
+            {352, 317, 317, 286},
+            {  0, 328, 328, 233},
+            {  0, 180, 180, 140},
+            {  0, 132, 132, 105},
+        }
+      }
+    };
+#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;
+        }
+        pix1 += line_size;
+        pix2 += line_size;
+    }
+    ff_spatial_dwt(tmp, w, h, 16, type, dec_count);
+
+    s=0;
+#if 0
+    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 sy= (ori&2) ? stride>>1 : 0;
+            int size= 1<<level;
+            
+            for(i=0; i<size; i++){
+                for(j=0; j<size; j++){
+                    int v= tmp[sx + sy + i*stride + j] * scale[type][dec_count-3][level][ori];
+                    s += ABS(v);
+                }
+            }
+        }
+    }
+#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;
+}
+
+static int w53_8_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h){
+    return w_c(v, pix1, pix2, line_size,  8, h, 1);
+}
+
+static int w97_8_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h){
+    return w_c(v, pix1, pix2, line_size,  8, h, 0);
+}
+
+static int w53_16_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h){
+    return w_c(v, pix1, pix2, line_size, 16, h, 1);
+}
+
+static int w97_16_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h){
+    return w_c(v, pix1, pix2, line_size, 16, h, 0);
+}
+
 static void get_pixels_c(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
 {
     int i;
@@ -2733,6 +2847,12 @@ void ff_set_cmp(DSPContext* c, me_cmp_func *cmp, int type){
         case FF_CMP_NSSE:
             cmp[i]= c->nsse[i];
             break;
+        case FF_CMP_W53:
+            cmp[i]= c->w53[i];
+            break;
+        case FF_CMP_W97:
+            cmp[i]= c->w97[i];
+            break;
         default:
             av_log(NULL, AV_LOG_ERROR,"internal error in cmp function selection\n");
         }
@@ -3359,6 +3479,7 @@ void dsputil_init(DSPContext* c, AVCodecContext *avctx)
     c->sad[1]= pix_abs8_c;
     c->sse[0]= sse16_c;
     c->sse[1]= sse8_c;
+    c->sse[2]= sse4_c;
     SET_CMP_FUNC(quant_psnr)
     SET_CMP_FUNC(rd)
     SET_CMP_FUNC(bit)
@@ -3368,7 +3489,11 @@ void dsputil_init(DSPContext* c, AVCodecContext *avctx)
     c->vsse[4]= vsse_intra16_c;
     c->nsse[0]= nsse16_c;
     c->nsse[1]= nsse8_c;
-        
+    c->w53[0]= w53_16_c;
+    c->w53[1]= w53_8_c;
+    c->w97[0]= w97_16_c;
+    c->w97[1]= w97_8_c;
+
     c->add_bytes= add_bytes_c;
     c->diff_bytes= diff_bytes_c;
     c->sub_hfyu_median_prediction= sub_hfyu_median_prediction_c;
diff --git a/libavcodec/dsputil.h b/libavcodec/dsputil.h
index 2ec3dc50c1..494c6f61c5 100644
--- a/libavcodec/dsputil.h
+++ b/libavcodec/dsputil.h
@@ -163,6 +163,8 @@ typedef struct DSPContext {
     me_cmp_func vsad[5];
     me_cmp_func vsse[5];
     me_cmp_func nsse[5];
+    me_cmp_func w53[5];
+    me_cmp_func w97[5];
 
     me_cmp_func me_pre_cmp[5];
     me_cmp_func me_cmp[5];
@@ -351,6 +353,29 @@ static inline uint32_t no_rnd_avg32(uint32_t a, uint32_t b)
     return (a & b) + (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
 }
 
+static inline int get_penalty_factor(int lambda, int lambda2, int type){
+    switch(type&0xFF){
+    default:
+    case FF_CMP_SAD:
+        return lambda>>FF_LAMBDA_SHIFT;
+    case FF_CMP_DCT:
+        return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
+    case FF_CMP_W53:
+        return (4*lambda)>>(FF_LAMBDA_SHIFT);
+    case FF_CMP_W97:
+        return (2*lambda)>>(FF_LAMBDA_SHIFT);
+    case FF_CMP_SATD:
+        return (2*lambda)>>FF_LAMBDA_SHIFT;
+    case FF_CMP_RD:
+    case FF_CMP_PSNR:
+    case FF_CMP_SSE:
+    case FF_CMP_NSSE:
+        return lambda2>>FF_LAMBDA_SHIFT;
+    case FF_CMP_BIT:
+        return 1;
+    }
+}
+
 /**
  * Empty mmx state.
  * this must be called between any dsp function and float/double code.
diff --git a/libavcodec/motion_est.c b/libavcodec/motion_est.c
index 13f7deb8d8..9239503373 100644
--- a/libavcodec/motion_est.c
+++ b/libavcodec/motion_est.c
@@ -219,25 +219,6 @@ static always_inline int cmp(MpegEncContext *s, const int x, const int y, const
 
 #include "motion_est_template.c"
 
-static inline int get_penalty_factor(MpegEncContext *s, int type){
-    switch(type&0xFF){
-    default:
-    case FF_CMP_SAD:
-        return s->lambda>>FF_LAMBDA_SHIFT;
-    case FF_CMP_DCT:
-        return (3*s->lambda)>>(FF_LAMBDA_SHIFT+1);
-    case FF_CMP_SATD:
-        return (2*s->lambda)>>FF_LAMBDA_SHIFT;
-    case FF_CMP_RD:
-    case FF_CMP_PSNR:
-    case FF_CMP_SSE:
-    case FF_CMP_NSSE:
-        return s->lambda2>>FF_LAMBDA_SHIFT;
-    case FF_CMP_BIT:
-        return 1;
-    }
-}
-
 static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride, int h){
     return 0;
 }
@@ -1161,9 +1142,9 @@ void ff_estimate_p_frame_motion(MpegEncContext * s,
     assert(s->linesize == c->stride);
     assert(s->uvlinesize == c->uvstride);
 
-    c->penalty_factor    = get_penalty_factor(s, c->avctx->me_cmp);
-    c->sub_penalty_factor= get_penalty_factor(s, c->avctx->me_sub_cmp);
-    c->mb_penalty_factor = get_penalty_factor(s, c->avctx->mb_cmp);
+    c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
+    c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
+    c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
     c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
 
     get_limits(s, 16*mb_x, 16*mb_y);
@@ -1256,7 +1237,7 @@ void ff_estimate_p_frame_motion(MpegEncContext * s,
             }
 
         }
-        dmin = epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift);       
+        dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);       
 
         break;
     }
@@ -1424,7 +1405,7 @@ int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
     
     assert(s->quarter_sample==0 || s->quarter_sample==1);
 
-    c->pre_penalty_factor    = get_penalty_factor(s, c->avctx->me_pre_cmp);
+    c->pre_penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
     c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
 
     get_limits(s, 16*mb_x, 16*mb_y);
@@ -1457,7 +1438,7 @@ int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
         c->pred_y = P_MEDIAN[1];
     }
 
-    dmin = epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift);       
+    dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);       
 
     s->p_mv_table[xy][0] = mx<<shift;
     s->p_mv_table[xy][1] = my<<shift;
@@ -1477,9 +1458,9 @@ static int ff_estimate_motion_b(MpegEncContext * s,
     uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
     int mv_scale;
         
-    c->penalty_factor    = get_penalty_factor(s, c->avctx->me_cmp);
-    c->sub_penalty_factor= get_penalty_factor(s, c->avctx->me_sub_cmp);
-    c->mb_penalty_factor = get_penalty_factor(s, c->avctx->mb_cmp);
+    c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
+    c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
+    c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
     c->current_mv_penalty= mv_penalty;
 
     get_limits(s, 16*mb_x, 16*mb_y);
@@ -1540,7 +1521,7 @@ static int ff_estimate_motion_b(MpegEncContext * s,
             mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
         }
         
-        dmin = epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale);
+        dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
  
         break;
     }
@@ -1731,7 +1712,7 @@ static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
         P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
     }
  
-    dmin = epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift));
+    dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
     if(c->sub_flags&FLAG_QPEL) 
         dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
     else
diff --git a/libavcodec/motion_est_template.c b/libavcodec/motion_est_template.c
index b5ee885730..db7c2a4cbc 100644
--- a/libavcodec/motion_est_template.c
+++ b/libavcodec/motion_est_template.c
@@ -851,15 +851,13 @@ static always_inline int diamond_search(MpegEncContext * s, int *best, int dmin,
 
 static always_inline int epzs_motion_search_internal(MpegEncContext * s, int *mx_ptr, int *my_ptr,
                              int P[10][2], int src_index, int ref_index, int16_t (*last_mv)[2], 
-                             int ref_mv_scale, int flags)
+                             int ref_mv_scale, int flags, int size, int h)
 {
     MotionEstContext * const c= &s->me;
     int best[2]={0, 0};
     int d, dmin;
     int map_generation;
     const int penalty_factor= c->penalty_factor;
-    const int size=0;
-    const int h=16;
     const int ref_mv_stride= s->mb_stride; //pass as arg  FIXME
     const int ref_mv_xy= s->mb_x + s->mb_y*ref_mv_stride; //add to last_mv beforepassing FIXME
     me_cmp_func cmpf, chroma_cmpf;
@@ -872,6 +870,7 @@ static always_inline int epzs_motion_search_internal(MpegEncContext * s, int *mx
     
     map_generation= update_map_generation(c);
 
+    assert(cmpf);
     dmin= cmp(s, 0, 0, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);
     map[0]= map_generation;
     score_map[0]= dmin;
@@ -882,7 +881,7 @@ static always_inline int epzs_motion_search_internal(MpegEncContext * s, int *mx
         CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, 
                         (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
     }else{
-        if(dmin<256 && ( P_LEFT[0]    |P_LEFT[1]
+        if(dmin<h*h && ( P_LEFT[0]    |P_LEFT[1]
                         |P_TOP[0]     |P_TOP[1]
                         |P_TOPRIGHT[0]|P_TOPRIGHT[1])==0){
             *mx_ptr= 0;
@@ -891,7 +890,7 @@ static always_inline int epzs_motion_search_internal(MpegEncContext * s, int *mx
             return dmin;
         }
         CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
-        if(dmin>256*2){
+        if(dmin>h*h*2){
             CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, 
                             (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
             CHECK_MV(P_LEFT[0]    >>shift, P_LEFT[1]    >>shift)
@@ -899,7 +898,7 @@ static always_inline int epzs_motion_search_internal(MpegEncContext * s, int *mx
             CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
         }
     }
-    if(dmin>256*4){
+    if(dmin>h*h*4){
         if(c->pre_pass){
             CHECK_CLIPED_MV((last_mv[ref_mv_xy-1][0]*ref_mv_scale + (1<<15))>>16, 
                             (last_mv[ref_mv_xy-1][1]*ref_mv_scale + (1<<15))>>16)
@@ -948,19 +947,18 @@ static always_inline int epzs_motion_search_internal(MpegEncContext * s, int *mx
 }
 
 //this function is dedicated to the braindamaged gcc
-static inline int epzs_motion_search(MpegEncContext * s, int *mx_ptr, int *my_ptr,
+inline int ff_epzs_motion_search(MpegEncContext * s, int *mx_ptr, int *my_ptr,
                              int P[10][2], int src_index, int ref_index, int16_t (*last_mv)[2], 
-                             int ref_mv_scale)
+                             int ref_mv_scale, int size, int h)
 {
     MotionEstContext * const c= &s->me;
 //FIXME convert other functions in the same way if faster
-    switch(c->flags){
-    case 0:
-        return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, 0);
+    if(c->flags==0 && h==16 && size==0){
+        return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, 0, 0, 16);
 //    case FLAG_QPEL:
 //        return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, FLAG_QPEL);
-    default:
-        return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, c->flags);
+    }else{
+        return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, c->flags, size, h);
     }
 }