Browse code

vsrc_mandelbrot: spatial interpolation speeds the code up by a factor of about 2-3.

Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

Michael Niedermayer authored on 2011/11/20 03:41:28
Showing 1 changed files
... ...
@@ -206,6 +206,58 @@ static void fill_from_cache(AVFilterContext *ctx, uint32_t *color, int *in_cidx,
206 206
     }
207 207
 }
208 208
 
209
+static int interpol(MBContext *mb, uint32_t *color, int x, int y, int linesize)
210
+{
211
+    uint32_t a,b,c,d, i;
212
+    uint32_t ipol=0;
213
+    int dist;
214
+
215
+    if(!x || !y || x+1==mb->w || y+1==mb->h)
216
+        return 0;
217
+
218
+    dist= FFMAX(FFABS(x-(mb->w>>1))*mb->h, FFABS(y-(mb->h>>1))*mb->w);
219
+
220
+    if(dist<(mb->w*mb->h>>3))
221
+        return 0;
222
+
223
+    a=color[(x+1) + (y+0)*linesize];
224
+    b=color[(x-1) + (y+1)*linesize];
225
+    c=color[(x+0) + (y+1)*linesize];
226
+    d=color[(x+1) + (y+1)*linesize];
227
+
228
+    if(a&&c){
229
+        b= color[(x-1) + (y+0)*linesize];
230
+        d= color[(x+0) + (y-1)*linesize];
231
+    }else if(b&&d){
232
+        a= color[(x+1) + (y-1)*linesize];
233
+        c= color[(x-1) + (y-1)*linesize];
234
+    }else if(c){
235
+        d= color[(x+0) + (y-1)*linesize];
236
+        a= color[(x-1) + (y+0)*linesize];
237
+        b= color[(x+1) + (y-1)*linesize];
238
+    }else if(d){
239
+        c= color[(x-1) + (y-1)*linesize];
240
+        a= color[(x-1) + (y+0)*linesize];
241
+        b= color[(x+1) + (y-1)*linesize];
242
+    }else
243
+        return 0;
244
+
245
+    for(i=0; i<3; i++){
246
+        int s= 8*i;
247
+        uint8_t ac= a>>s;
248
+        uint8_t bc= b>>s;
249
+        uint8_t cc= c>>s;
250
+        uint8_t dc= d>>s;
251
+        int ipolab= (ac + bc);
252
+        int ipolcd= (cc + dc);
253
+        if(FFABS(ipolab - ipolcd) > 5)
254
+            return 0;
255
+        ipol |= ((ipolab + ipolcd + 2)/4)<<s;
256
+    }
257
+    color[x + y*linesize]= ipol | 0xFF000000;
258
+    return 1;
259
+}
260
+
209 261
 static void draw_mandelbrot(AVFilterContext *ctx, uint32_t *color, int linesize, int64_t pts)
210 262
 {
211 263
     MBContext *mb = ctx->priv;
... ...
@@ -213,12 +265,16 @@ static void draw_mandelbrot(AVFilterContext *ctx, uint32_t *color, int linesize,
213 213
     double scale= mb->start_scale*pow(mb->end_scale/mb->start_scale, pts/mb->end_pts);
214 214
     int use_zyklus=0;
215 215
     fill_from_cache(ctx, NULL, &in_cidx, NULL, mb->start_y+scale*(-mb->h/2-0.5), scale);
216
+    tmp_cidx= in_cidx;
217
+    memset(color, 0, sizeof(*color)*mb->w);
216 218
     for(y=0; y<mb->h; y++){
219
+        int y1= y+1;
217 220
         const double ci=mb->start_y+scale*(y-mb->h/2);
218
-        memset(color+linesize*y, 0, sizeof(*color)*mb->w);
219
-        fill_from_cache(ctx, color+linesize*y, &in_cidx, &next_cidx, ci, scale);
220
-        tmp_cidx= in_cidx;
221
-        fill_from_cache(ctx, color+linesize*y, &tmp_cidx, NULL, ci + scale/2, scale);
221
+        fill_from_cache(ctx, NULL, &in_cidx, &next_cidx, ci, scale);
222
+        if(y1<mb->h){
223
+            memset(color+linesize*y1, 0, sizeof(*color)*mb->w);
224
+            fill_from_cache(ctx, color+linesize*y1, &tmp_cidx, NULL, ci + 3*scale/2, scale);
225
+        }
222 226
 
223 227
         for(x=0; x<mb->w; x++){
224 228
             const double cr=mb->start_x+scale*(x-mb->w/2);
... ...
@@ -230,6 +286,14 @@ static void draw_mandelbrot(AVFilterContext *ctx, uint32_t *color, int linesize,
230 230
 
231 231
             if(color[x + y*linesize] & 0xFF000000)
232 232
                 continue;
233
+            if(interpol(mb, color, x, y, linesize)){
234
+                if(next_cidx < mb->cache_allocated){
235
+                    mb->next_cache[next_cidx  ].p[0]= cr;
236
+                    mb->next_cache[next_cidx  ].p[1]= ci;
237
+                    mb->next_cache[next_cidx++].val = color[x + y*linesize];
238
+                }
239
+                continue;
240
+            }
233 241
 
234 242
             use_zyklus= (x==0 || mb->inner!=BLACK ||color[x-1 + y*linesize] == 0xFF000000);
235 243