Block filters using fragment shaders

Question

I was following this tutorial using Apple's OpenGL Shader Builder (tool similar to Nvidia's fx composer, but simpler).

I could easily apply the filters, but I don't understand if they worked correct (and if so how can I improve the output). For example the blur filter: OpenGL itself does some image processing on the textures, so if they are displayed in a higher resolution than the original image, they are blurred already by OpenGL. Second the blurred part is brighter then the part not processed, I think this does not make sense, since it just takes pixels from the direct neighborhood. This is defined by

float step_w = (1.0/width);

Which I don't quite understand: The pixels are indexed using floating point values??

Edit: I forgot to attach the exact code I used:

Fragment Shader

// Originally taken from: http://www.ozone3d.net/tutorials/image_filtering_p2.php#part_2

#define KERNEL_SIZE 9

float kernel[KERNEL_SIZE];

uniform sampler2D colorMap;
uniform float width;
uniform float height;

float step_w = (1.0/width);
float step_h = (1.0/height);

// float step_w = 20.0;
// float step_h = 20.0;

vec2 offset[KERNEL_SIZE];

void main(void)
{
   int i = 0;
   vec4 sum = vec4(0.0);

   offset[0] = vec2(-step_w, -step_h);  // south west
   offset[1] = vec2(0.0, -step_h);      // south
   offset[2] = vec2(step_w, -step_h);       // south east

   offset[3] = vec2(-step_w, 0.0);      // west
   offset[4] = vec2(0.0, 0.0);          // center
   offset[5] = vec2(step_w, 0.0);       // east

   offset[6] = vec2(-step_w, step_h);       // north west
   offset[7] = vec2(0.0, step_h);       // north
   offset[8] = vec2(step_w, step_h);        // north east


// Gaussian kernel
// 1 2 1
// 2 4 2
// 1 2 1


   kernel[0] = 1.0;     kernel[1] = 2.0;    kernel[2] = 1.0;
   kernel[3] = 2.0; kernel[4] = 4.0;    kernel[5] = 2.0;
   kernel[6] = 1.0;     kernel[7] = 2.0;    kernel[8] = 1.0;


// TODO make grayscale first
// Laplacian Filter
// 0   1   0
// 1  -4   1
// 0   1   0

/*
kernel[0] = 0.0;    kernel[1] = 1.0;    kernel[2] = 0.0;
kernel[3] = 1.0;    kernel[4] = -4.0;   kernel[5] = 1.0;
kernel[6] = 0.0;   kernel[7] = 2.0; kernel[8] = 0.0;
*/

// Mean Filter
// 1  1  1
// 1  1  1
// 1  1  1

/*
kernel[0] = 1.0;    kernel[1] = 1.0;    kernel[2] = 1.0;
kernel[3] = 1.0;    kernel[4] = 1.0;    kernel[5] = 1.0;
kernel[6] = 1.0;   kernel[7] = 1.0; kernel[8] = 1.0;
*/

   if(gl_TexCoord[0].s<0.5)
   {
       // For every pixel sample the neighbor pixels and sum up
       for( i=0; i<KERNEL_SIZE; i++ )
       {
            // select the pixel with the concerning offset
            vec4 tmp = texture2D(colorMap, gl_TexCoord[0].st + offset[i]);
            sum += tmp * kernel[i];
       }

        sum /= 16.0;
   }
   else if( gl_TexCoord[0].s>0.51 )
   {
        sum = texture2D(colorMap, gl_TexCoord[0].xy);
   }
   else // Draw a red line
   {
        sum = vec4(1.0, 0.0, 0.0, 1.0);
   }

   gl_FragColor = sum;
}

Vertex Shader

void main(void)
{
    gl_TexCoord[0] = gl_MultiTexCoord0;
    gl_Position = ftransform();
}

Answer 1

Texture coordinates conventionally reach from (0,0) (bottom left) to (1,1) (top right), so in fact, they are floats.

So if you have texturecoordinates (u,v), the "original" coordinates are computed by (u*textureWidth, v*textureHeight).

If the resulting values are not integral numbers, there may be different ways to handle that:

• just take floor or ceil of the result in order to make the number integral
• interpolate between the neighbouring texels

However I think every shading language has a method to access a texture by their "original", i.e. integral index.