* blur.c: separated blur; it's now incredibly faster.

git-svn-id: file:///srv/caca.zoy.org/var/lib/svn/libpipi/trunk@2608 92316355-f0b4-4df1-b90c-862c8a59935f

pipi/filter/blur.c

@@ -32,9 +32,9 @@ pipi_image_t *pipi_gaussian_blur(pipi_image_t *src, float radius)
     pipi_image_t *dst;
     pipi_pixels_t *srcp, *dstp;
     float *srcdata, *dstdata;
-    double *kernel;
+    double *kernel, *buffer;
     double K, L;
-    int x, y, i, j, w, h, kr, kw;
+    int x, y, i, w, h, kr, kw;
 
     w = src->w;
     h = src->h;
@@ -48,40 +48,61 @@ pipi_image_t *pipi_gaussian_blur(pipi_image_t *src, float radius)
 
     kr = (int)(3. * radius + 0.99999);
     kw = 2 * kr + 1;
-    K = 1. / (2. * M_PI * radius * radius);
+    K = 1. / (sqrt(2. * M_PI) * radius);
     L = -1. / (2. * radius * radius);
 
-    kernel = malloc(kw * kw * sizeof(double));
-    for(j = -kr; j <= kr; j++)
-        for(i = -kr; i <= kr; i++)
-            kernel[(j + kr) * kw + (i + kr)] = exp(L * (i * i + j * j)) * K;
+    kernel = malloc(kw * sizeof(double));
+    for(i = -kr; i <= kr; i++)
+        kernel[i + kr] = exp(L * i * i) * K;
+
+    buffer = malloc(w * h * 4 * sizeof(double));
 
     for(y = 0; y < h; y++)
     {
         for(x = 0; x < w; x++)
         {
             double R = 0., G = 0., B = 0., t = 0.;
-            int x2, y2;
+            int x2;
 
-            for(j = -kr; j <= kr; j++)
+            for(i = -kr; i <= kr; i++)
             {
-                y2 = y + j;
-                if(y2 < 0) y2 = 0;
-                else if(y2 >= h) y2 = h - 1;
+                double f = kernel[i + kr];
 
-                for(i = -kr; i <= kr; i++)
-                {
-                    double f = kernel[(j + kr) * kw + (i + kr)];
+                x2 = x + i;
+                if(x2 < 0) x2 = 0;
+                else if(x2 >= w) x2 = w - 1;
+
+                R += f * srcdata[(y * w + x2) * 4];
+                G += f * srcdata[(y * w + x2) * 4 + 1];
+                B += f * srcdata[(y * w + x2) * 4 + 2];
+                t += f;
+            }
 
-                    x2 = x + i;
-                    if(x2 < 0) x2 = 0;
-                    else if(x2 >= w) x2 = w - 1;
+            buffer[(y * w + x) * 4] = R / t;
+            buffer[(y * w + x) * 4 + 1] = G / t;
+            buffer[(y * w + x) * 4 + 2] = B / t;
+        }
+    }
 
-                    R += f * srcdata[(y2 * w + x2) * 4];
-                    G += f * srcdata[(y2 * w + x2) * 4 + 1];
-                    B += f * srcdata[(y2 * w + x2) * 4 + 2];
-                    t += f;
-                }
+    for(y = 0; y < h; y++)
+    {
+        for(x = 0; x < w; x++)
+        {
+            double R = 0., G = 0., B = 0., t = 0.;
+            int y2;
+
+            for(i = -kr; i <= kr; i++)
+            {
+                double f = kernel[i + kr];
+
+                y2 = y + i;
+                if(y2 < 0) y2 = 0;
+                else if(y2 >= h) y2 = h - 1;
+
+                R += f * buffer[(y2 * w + x) * 4];
+                G += f * buffer[(y2 * w + x) * 4 + 1];
+                B += f * buffer[(y2 * w + x) * 4 + 2];
+                t += f;
             }
 
             dstdata[(y * w + x) * 4] = R / t;
@@ -90,6 +111,9 @@ pipi_image_t *pipi_gaussian_blur(pipi_image_t *src, float radius)
         }
     }
 
+    free(buffer);
+    free(kernel);
+
     return dst;
 }