diff --git a/src/blur.cc b/src/blur.cc
index cb06b3d..73c57f1 100644
--- a/src/blur.cc
+++ b/src/blur.cc
@@ -23,118 +23,73 @@
  */
 
 #include <math.h>
+#include <memory.h>
 #include <stdint.h>
+#include <stdlib.h>
 #include <cairo.h>
 #include "canvas.h"
 
-#define ARRAY_LENGTH(a) (sizeof (a) / sizeof (a)[0])
-
 void
 Canvas::blur(cairo_surface_t *surface, int radius) {
-    cairo_surface_t *tmp;
-    int width, height;
-    int src_stride, dst_stride;
-    int x, y, z, w;
-    uint8_t *src, *dst;
-    uint32_t *s, *d, a, p;
-    int i, j, k;
-    uint8_t kernel[17];
-    const int size = ARRAY_LENGTH (kernel);
-    const int half = size / 2;
-
-    if (cairo_surface_status (surface))
-	return;
-
-    width = cairo_image_surface_get_width (surface);
-    height = cairo_image_surface_get_height (surface);
-
-    switch (cairo_image_surface_get_format (surface)) {
-    case CAIRO_FORMAT_A1:
-    default:
-	/* Don't even think about it! */
-	return;
-
-    case CAIRO_FORMAT_A8:
-	/* Handle a8 surfaces by effectively unrolling the loops by a
-	 * factor of 4 - this is safe since we know that stride has to be a
-	 * multiple of uint32_t. */
-	width /= 4;
-	break;
-
-    case CAIRO_FORMAT_RGB24:
-    case CAIRO_FORMAT_ARGB32:
-	break;
-    }
-
-    tmp = cairo_image_surface_create (CAIRO_FORMAT_ARGB32, width, height);
-    if (cairo_surface_status (tmp))
-	return;
-
-    src = cairo_image_surface_get_data (surface);
-    src_stride = cairo_image_surface_get_stride (surface);
-
-    dst = cairo_image_surface_get_data (tmp);
-    dst_stride = cairo_image_surface_get_stride (tmp);
-
-    a = 0;
-    for (i = 0; i < size; i++) {
-	double f = i - half;
-	a += kernel[i] = exp (- f * f / 30.0) * 80;
-    }
-
-    /* Horizontally blur from surface -> tmp */
-    for (i = 0; i < height; i++) {
-	s = (uint32_t *) (src + i * src_stride);
-	d = (uint32_t *) (dst + i * dst_stride);
-	for (j = 0; j < width; j++) {
-	    if (radius < j && j < width - radius) {
-		d[j] = s[j];
-		continue;
-	    }
-
-	    x = y = z = w = 0;
-	    for (k = 0; k < size; k++) {
-		if (j - half + k < 0 || j - half + k >= width)
-		    continue;
-
-		p = s[j - half + k];
-
-		x += ((p >> 24) & 0xff) * kernel[k];
-		y += ((p >> 16) & 0xff) * kernel[k];
-		z += ((p >>  8) & 0xff) * kernel[k];
-		w += ((p >>  0) & 0xff) * kernel[k];
-	    }
-	    d[j] = (x / a << 24) | (y / a << 16) | (z / a << 8) | w / a;
-	}
-    }
-
-    /* Then vertically blur from tmp -> surface */
-    for (i = 0; i < height; i++) {
-	s = (uint32_t *) (dst + i * dst_stride);
-	d = (uint32_t *) (src + i * src_stride);
-	for (j = 0; j < width; j++) {
-	    if (radius <= i && i < height - radius) {
-		d[j] = s[j];
-		continue;
-	    }
-
-	    x = y = z = w = 0;
-	    for (k = 0; k < size; k++) {
-		if (i - half + k < 0 || i - half + k >= height)
-		    continue;
-
-		s = (uint32_t *) (dst + (i - half + k) * dst_stride);
-		p = s[j];
-
-		x += ((p >> 24) & 0xff) * kernel[k];
-		y += ((p >> 16) & 0xff) * kernel[k];
-		z += ((p >>  8) & 0xff) * kernel[k];
-		w += ((p >>  0) & 0xff) * kernel[k];
-	    }
-	    d[j] = (x / a << 24) | (y / a << 16) | (z / a << 8) | w / a;
-	}
-    }
-
-  cairo_surface_destroy (tmp);
-  cairo_surface_mark_dirty (surface);
+  // Steve Hanov, 2009
+  // Released into the public domain.
+  
+  // get width, height
+  int width = cairo_image_surface_get_width( surface );
+  int height = cairo_image_surface_get_height( surface );
+  unsigned char* dst = (unsigned char*)malloc(width*height*4);
+  unsigned* precalc = 
+      (unsigned*)malloc(width*height*sizeof(unsigned));
+  unsigned char* src = cairo_image_surface_get_data( surface );
+  double mul=1.f/((radius*2)*(radius*2));
+  int channel;
+  
+  // The number of times to perform the averaging. According to wikipedia,
+  // three iterations is good enough to pass for a gaussian.
+  const int MAX_ITERATIONS = 3; 
+  int iteration;
+
+  memcpy( dst, src, width*height*4 );
+
+  for ( iteration = 0; iteration < MAX_ITERATIONS; iteration++ ) {
+      for( channel = 0; channel < 4; channel++ ) {
+          int x,y;
+
+          // precomputation step.
+          unsigned char* pix = src;
+          unsigned* pre = precalc;
+
+          pix += channel;
+          for (y=0;y<height;y++) {
+              for (x=0;x<width;x++) {
+                  int tot=pix[0];
+                  if (x>0) tot+=pre[-1];
+                  if (y>0) tot+=pre[-width];
+                  if (x>0 && y>0) tot-=pre[-width-1];
+                  *pre++=tot;
+                  pix += 4;
+              }
+          }
+
+          // blur step.
+          pix = dst + (int)radius * width * 4 + (int)radius * 4 + channel;
+          for (y=radius;y<height-radius;y++) {
+              for (x=radius;x<width-radius;x++) {
+                  int l = x < radius ? 0 : x - radius;
+                  int t = y < radius ? 0 : y - radius;
+                  int r = x + radius >= width ? width - 1 : x + radius;
+                  int b = y + radius >= height ? height - 1 : y + radius;
+                  int tot = precalc[r+b*width] + precalc[l+t*width] - 
+                      precalc[l+b*width] - precalc[r+t*width];
+                  *pix=(unsigned char)(tot*mul);
+                  pix += 4;
+              }
+              pix += (int)radius * 2 * 4;
+          }
+      }
+      memcpy( src, dst, width*height*4 );
+  }
+
+  free( dst );
+  free( precalc );
 }
\ No newline at end of file