Support CODEC_FLAG_EMU_EDGE in VP8 decoder.

Originally committed as revision 26117 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/h264pred.c

@@ -80,6 +80,20 @@ static void pred4x4_128_dc_c(uint8_t *src, const uint8_t *topright, int stride){
     ((uint32_t*)(src+3*stride))[0]= 128U*0x01010101U;
 }
 
+static void pred4x4_127_dc_c(uint8_t *src, const uint8_t *topright, int stride){
+    ((uint32_t*)(src+0*stride))[0]=
+    ((uint32_t*)(src+1*stride))[0]=
+    ((uint32_t*)(src+2*stride))[0]=
+    ((uint32_t*)(src+3*stride))[0]= 127U*0x01010101U;
+}
+
+static void pred4x4_129_dc_c(uint8_t *src, const uint8_t *topright, int stride){
+    ((uint32_t*)(src+0*stride))[0]=
+    ((uint32_t*)(src+1*stride))[0]=
+    ((uint32_t*)(src+2*stride))[0]=
+    ((uint32_t*)(src+3*stride))[0]= 129U*0x01010101U;
+}
+
 
 #define LOAD_TOP_RIGHT_EDGE\
     const int av_unused t4= topright[0];\
@@ -547,6 +561,28 @@ static void pred16x16_128_dc_c(uint8_t *src, int stride){
     }
 }
 
+static void pred16x16_127_dc_c(uint8_t *src, int stride){
+    int i;
+
+    for(i=0; i<16; i++){
+        ((uint32_t*)(src+i*stride))[0]=
+        ((uint32_t*)(src+i*stride))[1]=
+        ((uint32_t*)(src+i*stride))[2]=
+        ((uint32_t*)(src+i*stride))[3]= 0x01010101U*127U;
+    }
+}
+
+static void pred16x16_129_dc_c(uint8_t *src, int stride){
+    int i;
+
+    for(i=0; i<16; i++){
+        ((uint32_t*)(src+i*stride))[0]=
+        ((uint32_t*)(src+i*stride))[1]=
+        ((uint32_t*)(src+i*stride))[2]=
+        ((uint32_t*)(src+i*stride))[3]= 0x01010101U*129U;
+    }
+}
+
 static inline void pred16x16_plane_compat_c(uint8_t *src, int stride, const int svq3, const int rv40){
   int i, j, k;
   int a;
@@ -658,6 +694,23 @@ static void pred8x8_128_dc_c(uint8_t *src, int stride){
     }
 }
 
+static void pred8x8_127_dc_c(uint8_t *src, int stride){
+    int i;
+
+    for(i=0; i<8; i++){
+        ((uint32_t*)(src+i*stride))[0]=
+        ((uint32_t*)(src+i*stride))[1]= 0x01010101U*127U;
+    }
+}
+static void pred8x8_129_dc_c(uint8_t *src, int stride){
+    int i;
+
+    for(i=0; i<8; i++){
+        ((uint32_t*)(src+i*stride))[0]=
+        ((uint32_t*)(src+i*stride))[1]= 0x01010101U*129U;
+    }
+}
+
 static void pred8x8_left_dc_c(uint8_t *src, int stride){
     int i;
     int dc0, dc2;
@@ -1210,11 +1263,17 @@ void ff_h264_pred_init(H264PredContext *h, int codec_id){
         } else
             h->pred4x4[VERT_LEFT_PRED  ]= pred4x4_vertical_left_c;
         h->pred4x4[HOR_UP_PRED         ]= pred4x4_horizontal_up_c;
+        if(codec_id != CODEC_ID_VP8) {
         h->pred4x4[LEFT_DC_PRED        ]= pred4x4_left_dc_c;
         h->pred4x4[TOP_DC_PRED         ]= pred4x4_top_dc_c;
         h->pred4x4[DC_128_PRED         ]= pred4x4_128_dc_c;
-        if(codec_id == CODEC_ID_VP8)
+        } else {
             h->pred4x4[TM_VP8_PRED     ]= pred4x4_tm_vp8_c;
+            h->pred4x4[DC_127_PRED     ]= pred4x4_127_dc_c;
+            h->pred4x4[DC_129_PRED     ]= pred4x4_129_dc_c;
+            h->pred4x4[VERT_VP8_PRED   ]= pred4x4_vertical_c;
+            h->pred4x4[HOR_VP8_PRED    ]= pred4x4_horizontal_c;
+        }
     }else{
         h->pred4x4[VERT_PRED           ]= pred4x4_vertical_c;
         h->pred4x4[HOR_PRED            ]= pred4x4_horizontal_c;
@@ -1264,13 +1323,16 @@ void ff_h264_pred_init(H264PredContext *h, int codec_id){
         h->pred8x8[DC_PRED8x8     ]= pred8x8_dc_rv40_c;
         h->pred8x8[LEFT_DC_PRED8x8]= pred8x8_left_dc_rv40_c;
         h->pred8x8[TOP_DC_PRED8x8 ]= pred8x8_top_dc_rv40_c;
+        if (codec_id == CODEC_ID_VP8) {
+            h->pred8x8[DC_127_PRED8x8]= pred8x8_127_dc_c;
+            h->pred8x8[DC_129_PRED8x8]= pred8x8_129_dc_c;
+        }
     }
     h->pred8x8[DC_128_PRED8x8 ]= pred8x8_128_dc_c;
 
     h->pred16x16[DC_PRED8x8     ]= pred16x16_dc_c;
     h->pred16x16[VERT_PRED8x8   ]= pred16x16_vertical_c;
     h->pred16x16[HOR_PRED8x8    ]= pred16x16_horizontal_c;
-    h->pred16x16[PLANE_PRED8x8  ]= pred16x16_plane_c;
     switch(codec_id){
     case CODEC_ID_SVQ3:
        h->pred16x16[PLANE_PRED8x8  ]= pred16x16_plane_svq3_c;
@@ -1280,9 +1342,12 @@ void ff_h264_pred_init(H264PredContext *h, int codec_id){
        break;
     case CODEC_ID_VP8:
        h->pred16x16[PLANE_PRED8x8  ]= pred16x16_tm_vp8_c;
+       h->pred16x16[DC_127_PRED8x8]= pred16x16_127_dc_c;
+       h->pred16x16[DC_129_PRED8x8]= pred16x16_129_dc_c;
        break;
     default:
        h->pred16x16[PLANE_PRED8x8  ]= pred16x16_plane_c;
+       break;
     }
     h->pred16x16[LEFT_DC_PRED8x8]= pred16x16_left_dc_c;
     h->pred16x16[TOP_DC_PRED8x8 ]= pred16x16_top_dc_c;

libavcodec/h264pred.h

@@ -45,29 +45,45 @@
 #define VERT_LEFT_PRED        7
 #define HOR_UP_PRED           8
 
+// DC edge (not for VP8)
 #define LEFT_DC_PRED          9
 #define TOP_DC_PRED           10
 #define DC_128_PRED           11
 
-#define TM_VP8_PRED           9     ///< "True Motion", used instead of plane
+// RV40 specific
-
 #define DIAG_DOWN_LEFT_PRED_RV40_NODOWN   12
 #define HOR_UP_PRED_RV40_NODOWN           13
 #define VERT_LEFT_PRED_RV40_NODOWN        14
 
+// VP8 specific
+#define TM_VP8_PRED           9     ///< "True Motion", used instead of plane
+#define VERT_VP8_PRED         10    ///< for VP8, #VERT_PRED is the average of
+                                    ///< (left col+cur col x2+right col) / 4;
+                                    ///< this is the "unaveraged" one
+#define HOR_VP8_PRED          11    ///< unaveraged version of #HOR_PRED, see
+                                    ///< #VERT_VP8_PRED for details
+#define DC_127_PRED           12
+#define DC_129_PRED           13
+
 #define DC_PRED8x8            0
 #define HOR_PRED8x8           1
 #define VERT_PRED8x8          2
 #define PLANE_PRED8x8         3
 
+// DC edge
 #define LEFT_DC_PRED8x8       4
 #define TOP_DC_PRED8x8        5
 #define DC_128_PRED8x8        6
 
+// H264/SVQ3 (8x8) specific
 #define ALZHEIMER_DC_L0T_PRED8x8 7
 #define ALZHEIMER_DC_0LT_PRED8x8 8
 #define ALZHEIMER_DC_L00_PRED8x8 9
 #define ALZHEIMER_DC_0L0_PRED8x8 10
+
+// VP8 specific
+#define DC_127_PRED8x8        7
+#define DC_129_PRED8x8        8
 //@}
 
 /**
@@ -77,7 +93,7 @@ typedef struct H264PredContext{
     void (*pred4x4  [9+3+3])(uint8_t *src, const uint8_t *topright, int stride);//FIXME move to dsp?
     void (*pred8x8l [9+3])(uint8_t *src, int topleft, int topright, int stride);
     void (*pred8x8  [4+3+4])(uint8_t *src, int stride);
-    void (*pred16x16[4+3])(uint8_t *src, int stride);
+    void (*pred16x16[4+3+2])(uint8_t *src, int stride);
 
     void (*pred4x4_add  [2])(uint8_t *pix/*align  4*/, const DCTELEM *block/*align 16*/, int stride);
     void (*pred8x8l_add [2])(uint8_t *pix/*align  8*/, const DCTELEM *block/*align 16*/, int stride);

libavcodec/vp8.c

@@ -983,14 +983,91 @@ void xchg_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_
 }
 
 static av_always_inline
-int check_intra_pred_mode(int mode, int mb_x, int mb_y)
+int check_dc_pred8x8_mode(int mode, int mb_x, int mb_y)
+{
+    if (!mb_x) {
+        return mb_y ? TOP_DC_PRED8x8 : DC_128_PRED8x8;
+    } else {
+        return mb_y ? mode : LEFT_DC_PRED8x8;
+    }
+}
+
+static av_always_inline
+int check_tm_pred8x8_mode(int mode, int mb_x, int mb_y)
+{
+    if (!mb_x) {
+        return mb_y ? VERT_PRED8x8 : DC_129_PRED8x8;
+    } else {
+        return mb_y ? mode : HOR_PRED8x8;
+    }
+}
+
+static av_always_inline
+int check_intra_pred8x8_mode(int mode, int mb_x, int mb_y)
 {
     if (mode == DC_PRED8x8) {
-        if (!mb_x) {
+        return check_dc_pred8x8_mode(mode, mb_x, mb_y);
-            mode = mb_y ? TOP_DC_PRED8x8 : DC_128_PRED8x8;
+    } else {
-        } else if (!mb_y) {
+        return mode;
-            mode = LEFT_DC_PRED8x8;
+    }
+}
+
+static av_always_inline
+int check_intra_pred8x8_mode_emuedge(int mode, int mb_x, int mb_y)
+{
+    switch (mode) {
+    case DC_PRED8x8:
+        return check_dc_pred8x8_mode(mode, mb_x, mb_y);
+    case VERT_PRED8x8:
+        return !mb_y ? DC_127_PRED8x8 : mode;
+    case HOR_PRED8x8:
+        return !mb_x ? DC_129_PRED8x8 : mode;
+    case PLANE_PRED8x8 /*TM*/:
+        return check_tm_pred8x8_mode(mode, mb_x, mb_y);
+    }
+    return mode;
+}
+
+static av_always_inline
+int check_tm_pred4x4_mode(int mode, int mb_x, int mb_y)
+{
+    if (!mb_x) {
+        return mb_y ? VERT_VP8_PRED : DC_129_PRED;
+    } else {
+        return mb_y ? mode : HOR_VP8_PRED;
+    }
+}
+
+static av_always_inline
+int check_intra_pred4x4_mode_emuedge(int mode, int mb_x, int mb_y, int *copy_buf)
+{
+    switch (mode) {
+    case VERT_PRED:
+        if (!mb_x && mb_y) {
+            *copy_buf = 1;
+            return mode;
         }
+        /* fall-through */
+    case DIAG_DOWN_LEFT_PRED:
+    case VERT_LEFT_PRED:
+        return !mb_y ? DC_127_PRED : mode;
+    case HOR_PRED:
+        if (!mb_y) {
+            *copy_buf = 1;
+            return mode;
+        }
+        /* fall-through */
+    case HOR_UP_PRED:
+        return !mb_x ? DC_129_PRED : mode;
+    case TM_VP8_PRED:
+        return check_tm_pred4x4_mode(mode, mb_x, mb_y);
+    case DC_PRED: // 4x4 DC doesn't use the same "H.264-style" exceptions as 16x16/8x8 DC
+    case DIAG_DOWN_RIGHT_PRED:
+    case VERT_RIGHT_PRED:
+    case HOR_DOWN_PRED:
+        if (!mb_y || !mb_x)
+            *copy_buf = 1;
+        return mode;
     }
     return mode;
 }
@@ -999,21 +1076,27 @@ static av_always_inline
 void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
                    int mb_x, int mb_y)
 {
+    AVCodecContext *avctx = s->avctx;
     int x, y, mode, nnz, tr;
 
     // for the first row, we need to run xchg_mb_border to init the top edge to 127
     // otherwise, skip it if we aren't going to deblock
-    if (s->deblock_filter || !mb_y)
+    if (!(avctx->flags & CODEC_FLAG_EMU_EDGE && !mb_y) && (s->deblock_filter || !mb_y))
         xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
                        s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
                        s->filter.simple, 1);
 
     if (mb->mode < MODE_I4x4) {
-        mode = check_intra_pred_mode(mb->mode, mb_x, mb_y);
+        if (avctx->flags & CODEC_FLAG_EMU_EDGE) { // tested
+            mode = check_intra_pred8x8_mode_emuedge(mb->mode, mb_x, mb_y);
+        } else {
+            mode = check_intra_pred8x8_mode(mb->mode, mb_x, mb_y);
+        }
         s->hpc.pred16x16[mode](dst[0], s->linesize);
     } else {
         uint8_t *ptr = dst[0];
         uint8_t *intra4x4 = s->intra4x4_pred_mode_mb;
+        uint8_t tr_top[4] = { 127, 127, 127, 127 };
 
         // all blocks on the right edge of the macroblock use bottom edge
         // the top macroblock for their topright edge
@@ -1032,10 +1115,53 @@ void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
         for (y = 0; y < 4; y++) {
             uint8_t *topright = ptr + 4 - s->linesize;
             for (x = 0; x < 4; x++) {
-                if (x == 3)
+                int copy = 0, linesize = s->linesize;
+                uint8_t *dst = ptr+4*x;
+                DECLARE_ALIGNED(4, uint8_t, copy_dst)[5*8];
+
+                if ((y == 0 || x == 3) && mb_y == 0 && avctx->flags & CODEC_FLAG_EMU_EDGE) {
+                    topright = tr_top;
+                } else if (x == 3)
                     topright = tr_right;
 
-                s->hpc.pred4x4[intra4x4[x]](ptr+4*x, topright, s->linesize);
+                if (avctx->flags & CODEC_FLAG_EMU_EDGE) { // mb_x+x or mb_y+y is a hack but works
+                    mode = check_intra_pred4x4_mode_emuedge(intra4x4[x], mb_x + x, mb_y + y, &copy);
+                    if (copy) {
+                        dst = copy_dst + 12;
+                        linesize = 8;
+                        if (!(mb_y + y)) {
+                            copy_dst[3] = 127U;
+                            * (uint32_t *) (copy_dst + 4) = 127U * 0x01010101U;
+                        } else {
+                            * (uint32_t *) (copy_dst + 4) = * (uint32_t *) (ptr+4*x-s->linesize);
+                            if (!(mb_x + x)) {
+                                copy_dst[3] = 129U;
+                            } else {
+                                copy_dst[3] = ptr[4*x-s->linesize-1];
+                            }
+                        }
+                        if (!(mb_x + x)) {
+                            copy_dst[11] =
+                            copy_dst[19] =
+                            copy_dst[27] =
+                            copy_dst[35] = 129U;
+                        } else {
+                            copy_dst[11] = ptr[4*x              -1];
+                            copy_dst[19] = ptr[4*x+s->linesize  -1];
+                            copy_dst[27] = ptr[4*x+s->linesize*2-1];
+                            copy_dst[35] = ptr[4*x+s->linesize*3-1];
+                        }
+                    }
+                } else {
+                    mode = intra4x4[x];
+                }
+                s->hpc.pred4x4[mode](dst, topright, linesize);
+                if (copy) {
+                    * (uint32_t *) (ptr+4*x)               = * (uint32_t *) (copy_dst + 12);
+                    * (uint32_t *) (ptr+4*x+s->linesize)   = * (uint32_t *) (copy_dst + 20);
+                    * (uint32_t *) (ptr+4*x+s->linesize*2) = * (uint32_t *) (copy_dst + 28);
+                    * (uint32_t *) (ptr+4*x+s->linesize*3) = * (uint32_t *) (copy_dst + 36);
+                }
 
                 nnz = s->non_zero_count_cache[y][x];
                 if (nnz) {
@@ -1052,11 +1178,15 @@ void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
         }
     }
 
-    mode = check_intra_pred_mode(s->chroma_pred_mode, mb_x, mb_y);
+    if (avctx->flags & CODEC_FLAG_EMU_EDGE) {
+        mode = check_intra_pred8x8_mode_emuedge(s->chroma_pred_mode, mb_x, mb_y);
+    } else {
+        mode = check_intra_pred8x8_mode(s->chroma_pred_mode, mb_x, mb_y);
+    }
     s->hpc.pred8x8[mode](dst[1], s->uvlinesize);
     s->hpc.pred8x8[mode](dst[2], s->uvlinesize);
 
-    if (s->deblock_filter || !mb_y)
+    if (!(avctx->flags & CODEC_FLAG_EMU_EDGE && !mb_y) && (s->deblock_filter || !mb_y))
         xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
                        s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
                        s->filter.simple, 0);
@@ -1533,7 +1663,10 @@ static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
     memset(s->macroblocks + s->mb_height*2 - 1, 0, (s->mb_width+1)*sizeof(*s->macroblocks));
 
     // top edge of 127 for intra prediction
-    memset(s->top_border, 127, (s->mb_width+1)*sizeof(*s->top_border));
+    if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) {
+        s->top_border[0][15] = s->top_border[0][23] = 127;
+        memset(s->top_border[1]-1, 127, s->mb_width*sizeof(*s->top_border)+1);
+    }
     memset(s->ref_count, 0, sizeof(s->ref_count));
     if (s->keyframe)
         memset(s->intra4x4_pred_mode_top, DC_PRED, s->mb_width*4);
@@ -1553,12 +1686,13 @@ static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
         AV_WN32A(s->intra4x4_pred_mode_left, DC_PRED*0x01010101);
 
         // left edge of 129 for intra prediction
-        if (!(avctx->flags & CODEC_FLAG_EMU_EDGE))
+        if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) {
             for (i = 0; i < 3; i++)
                 for (y = 0; y < 16>>!!i; y++)
                     dst[i][y*curframe->linesize[i]-1] = 129;
-        if (mb_y)
+            if (mb_y == 1) // top left edge is also 129
-            memset(s->top_border, 129, sizeof(*s->top_border));
+                s->top_border[0][15] = s->top_border[0][23] = s->top_border[0][31] = 129;
+        }
 
         for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) {
             /* Prefetch the current frame, 4 MBs ahead */
@@ -1658,12 +1792,6 @@ static av_cold int vp8_decode_init(AVCodecContext *avctx)
     ff_h264_pred_init(&s->hpc, CODEC_ID_VP8);
     ff_vp8dsp_init(&s->vp8dsp);
 
-    // intra pred needs edge emulation among other things
-    if (avctx->flags&CODEC_FLAG_EMU_EDGE) {
-        av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported\n");
-        return AVERROR_PATCHWELCOME;
-    }
-
     return 0;
 }