Convert RoQ decoder to use YUV 4:4:4 unpacked macroblocks

instead of packed 4:2:0 clusters
patch by Eric Lasota: \ riot icculus org /
original threads: [FFmpeg-devel]   [PATCH] RoQ decoder 4:4:4 update
date: 06/04/2007 03:52 AM and 06/05/2007 01:10 AM

Originally committed as revision 9223 to svn://svn.ffmpeg.org/ffmpeg/trunk

1 files changed, 73 insertions, 173 deletions

diff --git a/libavcodec/roqvideo.c b/libavcodec/roqvideo.c
index 5725757834..f6b0184360 100644
--- a/libavcodec/roqvideo.c
+++ b/libavcodec/roqvideo.c
@@ -27,203 +27,103 @@
 #include "avcodec.h"
 #include "roqvideo.h"
 
-#define avg2(a,b) av_clip_uint8(((int)(a)+(int)(b)+1)>>1)
-#define avg4(a,b,c,d) av_clip_uint8(((int)(a)+(int)(b)+(int)(c)+(int)(d)+2)>>2)
+static inline void block_copy(unsigned char *out, unsigned char *in,
+                              int outstride, int instride, int sz)
+{
+    int rows = sz;
+    while(rows--) {
+        memcpy(out, in, sz);
+        out += outstride;
+        in += instride;
+    }
+}
 
 void ff_apply_vector_2x2(RoqContext *ri, int x, int y, roq_cell *cell)
 {
-    unsigned char *yptr;
-
-    yptr = ri->current_frame->data[0] + (y * ri->y_stride) + x;
-    *yptr++ = cell->y[0];
-    *yptr++ = cell->y[1];
-    yptr += (ri->y_stride - 2);
-    *yptr++ = cell->y[2];
-    *yptr++ = cell->y[3];
-    ri->current_frame->data[1][(y/2) * (ri->c_stride) + x/2] = cell->u;
-    ri->current_frame->data[2][(y/2) * (ri->c_stride) + x/2] = cell->v;
+    unsigned char *bptr;
+    int boffs,stride;
+
+    stride = ri->y_stride;
+    boffs = (y * stride) + x;
+
+    bptr = ri->current_frame->data[0] + boffs;
+    bptr[0       ] = cell->y[0];
+    bptr[1       ] = cell->y[1];
+    bptr[stride  ] = cell->y[2];
+    bptr[stride+1] = cell->y[3];
+
+    bptr = ri->current_frame->data[1] + boffs;
+    bptr[0       ] =
+    bptr[1       ] =
+    bptr[stride  ] =
+    bptr[stride+1] = cell->u;
+
+    bptr = ri->current_frame->data[2] + boffs;
+    bptr[0       ] =
+    bptr[1       ] =
+    bptr[stride  ] =
+    bptr[stride+1] = cell->v;
 }
 
 void ff_apply_vector_4x4(RoqContext *ri, int x, int y, roq_cell *cell)
 {
-    unsigned long row_inc, c_row_inc;
-    register unsigned char y0, y1, u, v;
-    unsigned char *yptr, *uptr, *vptr;
-
-    yptr = ri->current_frame->data[0] + (y * ri->y_stride) + x;
-    uptr = ri->current_frame->data[1] + (y/2) * (ri->c_stride) + x/2;
-    vptr = ri->current_frame->data[2] + (y/2) * (ri->c_stride) + x/2;
-
-    row_inc = ri->y_stride - 4;
-    c_row_inc = (ri->c_stride) - 2;
-    *yptr++ = y0 = cell->y[0]; *uptr++ = u = cell->u; *vptr++ = v = cell->v;
-    *yptr++ = y0;
-    *yptr++ = y1 = cell->y[1]; *uptr++ = u; *vptr++ = v;
-    *yptr++ = y1;
-
-    yptr += row_inc;
-
-    *yptr++ = y0;
-    *yptr++ = y0;
-    *yptr++ = y1;
-    *yptr++ = y1;
-
-    yptr += row_inc; uptr += c_row_inc; vptr += c_row_inc;
-
-    *yptr++ = y0 = cell->y[2]; *uptr++ = u; *vptr++ = v;
-    *yptr++ = y0;
-    *yptr++ = y1 = cell->y[3]; *uptr++ = u; *vptr++ = v;
-    *yptr++ = y1;
-
-    yptr += row_inc;
-
-    *yptr++ = y0;
-    *yptr++ = y0;
-    *yptr++ = y1;
-    *yptr++ = y1;
+    unsigned char *bptr;
+    int boffs,stride;
+
+    stride = ri->y_stride;
+    boffs = (y * stride) + x;
+
+    bptr = ri->current_frame->data[0] + boffs;
+    bptr[         0] = bptr[         1] = bptr[stride    ] = bptr[stride  +1] = cell->y[0];
+    bptr[         2] = bptr[         3] = bptr[stride  +2] = bptr[stride  +3] = cell->y[1];
+    bptr[stride*2  ] = bptr[stride*2+1] = bptr[stride*3  ] = bptr[stride*3+1] = cell->y[2];
+    bptr[stride*2+2] = bptr[stride*2+3] = bptr[stride*3+2] = bptr[stride*3+3] = cell->y[2];
+
+    bptr = ri->current_frame->data[1] + boffs;
+    bptr[         0] = bptr[         1] = bptr[stride    ] = bptr[stride  +1] =
+    bptr[         2] = bptr[         3] = bptr[stride  +2] = bptr[stride  +3] =
+    bptr[stride*2  ] = bptr[stride*2+1] = bptr[stride*3  ] = bptr[stride*3+1] =
+    bptr[stride*2+2] = bptr[stride*2+3] = bptr[stride*3+2] = bptr[stride*3+3] = cell->u;
+
+    bptr = ri->current_frame->data[2] + boffs;
+    bptr[         0] = bptr[         1] = bptr[stride    ] = bptr[stride  +1] =
+    bptr[         2] = bptr[         3] = bptr[stride  +2] = bptr[stride  +3] =
+    bptr[stride*2  ] = bptr[stride*2+1] = bptr[stride*3  ] = bptr[stride*3+1] =
+    bptr[stride*2+2] = bptr[stride*2+3] = bptr[stride*3+2] = bptr[stride*3+3] = cell->v;
 }
 
-void ff_apply_motion_4x4(RoqContext *ri, int x, int y,
-                             int deltax, int deltay)
+
+static inline void apply_motion_generic(RoqContext *ri, int x, int y, int deltax,
+                                        int deltay, int sz)
 {
-    int i, hw, mx, my;
-    unsigned char *pa, *pb;
+    int mx, my, cp;
 
     mx = x + deltax;
     my = y + deltay;
 
     /* check MV against frame boundaries */
-    if ((mx < 0) || (mx > ri->avctx->width - 4) ||
-        (my < 0) || (my > ri->avctx->height - 4)) {
+    if ((mx < 0) || (mx > ri->avctx->width - sz) ||
+        (my < 0) || (my > ri->avctx->height - sz)) {
         av_log(ri->avctx, AV_LOG_ERROR, "motion vector out of bounds: MV = (%d, %d), boundaries = (0, 0, %d, %d)\n",
             mx, my, ri->avctx->width, ri->avctx->height);
         return;
     }
 
-    pa = ri->current_frame->data[0] + (y * ri->y_stride) + x;
-    pb = ri->last_frame->data[0] + (my * ri->y_stride) + mx;
-    for(i = 0; i < 4; i++) {
-        pa[0] = pb[0];
-        pa[1] = pb[1];
-        pa[2] = pb[2];
-        pa[3] = pb[3];
-        pa += ri->y_stride;
-        pb += ri->y_stride;
-    }
+    for(cp = 0; cp < 3; cp++)
+        block_copy(ri->current_frame->data[cp] + (y * ri->y_stride) + x,
+                   ri->last_frame->data[cp] + (my * ri->y_stride) + mx,
+                   ri->y_stride, ri->y_stride, sz);
+}
 
-    hw = ri->y_stride/2;
-    pa = ri->current_frame->data[1] + (y * ri->y_stride)/4 + x/2;
-    pb = ri->last_frame->data[1] + (my/2) * (ri->y_stride/2) + (mx + 1)/2;
-
-    for(i = 0; i < 2; i++) {
-        switch(((my & 0x01) << 1) | (mx & 0x01)) {
-
-        case 0:
-            pa[0] = pb[0];
-            pa[1] = pb[1];
-            pa[hw] = pb[hw];
-            pa[hw+1] = pb[hw+1];
-            break;
-
-        case 1:
-            pa[0] = avg2(pb[0], pb[1]);
-            pa[1] = avg2(pb[1], pb[2]);
-            pa[hw] = avg2(pb[hw], pb[hw+1]);
-            pa[hw+1] = avg2(pb[hw+1], pb[hw+2]);
-            break;
-
-        case 2:
-            pa[0] = avg2(pb[0], pb[hw]);
-            pa[1] = avg2(pb[1], pb[hw+1]);
-            pa[hw] = avg2(pb[hw], pb[hw*2]);
-            pa[hw+1] = avg2(pb[hw+1], pb[(hw*2)+1]);
-            break;
-
-        case 3:
-            pa[0] = avg4(pb[0], pb[1], pb[hw], pb[hw+1]);
-            pa[1] = avg4(pb[1], pb[2], pb[hw+1], pb[hw+2]);
-            pa[hw] = avg4(pb[hw], pb[hw+1], pb[hw*2], pb[(hw*2)+1]);
-            pa[hw+1] = avg4(pb[hw+1], pb[hw+2], pb[(hw*2)+1], pb[(hw*2)+1]);
-            break;
-        }
-
-        pa = ri->current_frame->data[2] + (y * ri->y_stride)/4 + x/2;
-        pb = ri->last_frame->data[2] + (my/2) * (ri->y_stride/2) + (mx + 1)/2;
-    }
+
+void ff_apply_motion_4x4(RoqContext *ri, int x, int y,
+                             int deltax, int deltay)
+{
+    apply_motion_generic(ri, x, y, deltax, deltay, 4);
 }
 
 void ff_apply_motion_8x8(RoqContext *ri, int x, int y,
                              int deltax, int deltay)
 {
-    int mx, my, i, j, hw;
-    unsigned char *pa, *pb;
-
-    mx = x + deltax;
-    my = y + deltay;
-
-    /* check MV against frame boundaries */
-    if ((mx < 0) || (mx > ri->avctx->width - 8) ||
-        (my < 0) || (my > ri->avctx->height - 8)) {
-        av_log(ri->avctx, AV_LOG_ERROR, "motion vector out of bounds: MV = (%d, %d), boundaries = (0, 0, %d, %d)\n",
-            mx, my, ri->avctx->width, ri->avctx->height);
-        return;
-    }
-
-    pa = ri->current_frame->data[0] + (y * ri->y_stride) + x;
-    pb = ri->last_frame->data[0] + (my * ri->y_stride) + mx;
-    for(i = 0; i < 8; i++) {
-        pa[0] = pb[0];
-        pa[1] = pb[1];
-        pa[2] = pb[2];
-        pa[3] = pb[3];
-        pa[4] = pb[4];
-        pa[5] = pb[5];
-        pa[6] = pb[6];
-        pa[7] = pb[7];
-        pa += ri->y_stride;
-        pb += ri->y_stride;
-    }
-
-    hw = ri->c_stride;
-    pa = ri->current_frame->data[1] + (y * ri->y_stride)/4 + x/2;
-    pb = ri->last_frame->data[1] + (my/2) * (ri->y_stride/2) + (mx + 1)/2;
-    for(j = 0; j < 2; j++) {
-        for(i = 0; i < 4; i++) {
-            switch(((my & 0x01) << 1) | (mx & 0x01)) {
-
-            case 0:
-                pa[0] = pb[0];
-                pa[1] = pb[1];
-                pa[2] = pb[2];
-                pa[3] = pb[3];
-                break;
-
-            case 1:
-                pa[0] = avg2(pb[0], pb[1]);
-                pa[1] = avg2(pb[1], pb[2]);
-                pa[2] = avg2(pb[2], pb[3]);
-                pa[3] = avg2(pb[3], pb[4]);
-                break;
-
-            case 2:
-                pa[0] = avg2(pb[0], pb[hw]);
-                pa[1] = avg2(pb[1], pb[hw+1]);
-                pa[2] = avg2(pb[2], pb[hw+2]);
-                pa[3] = avg2(pb[3], pb[hw+3]);
-                break;
-
-            case 3:
-                pa[0] = avg4(pb[0], pb[1], pb[hw], pb[hw+1]);
-                pa[1] = avg4(pb[1], pb[2], pb[hw+1], pb[hw+2]);
-                pa[2] = avg4(pb[2], pb[3], pb[hw+2], pb[hw+3]);
-                pa[3] = avg4(pb[3], pb[4], pb[hw+3], pb[hw+4]);
-                break;
-            }
-            pa += ri->c_stride;
-            pb += ri->c_stride;
-        }
-
-        pa = ri->current_frame->data[2] + (y * ri->y_stride)/4 + x/2;
-        pb = ri->last_frame->data[2] + (my/2) * (ri->y_stride/2) + (mx + 1)/2;
-    }
+    apply_motion_generic(ri, x, y, deltax, deltay, 8);
 }