Question

I am enrolled in shaders course and interested in computer vision and image processing. I was wondering how can I mix GLSL shaders knowledge with image processing? What do I gain if I implement image processing algorithms with GLSL?

Answer 1

The first obvious answer is that you gain parallelism. Now, why using GLSL rather than, say CUDA which is more flexible ? GLSL doesn’t require you to have an NVIDIA graphics card, so it’s a much more portable solution (you’d still have the option of OpenCL though).

What can you gain with parallelism ? Most of the time, you can treat pixels independantly. For instance, increasing the contrast of an image usually requires you to loop over all pixels and apply an affine transform of the pixel values. If each pixel is handled by a separate thread, then you don’t need to do this loop anymore : you just raterize a quad, and apply a pixel shader that reads a texture at the current rasterized point, and ouput to the render target (or the screen) the transformed pixel value.

The drawback is that your data need to reside on the GPU : you’ll need to transfer all your images to the GPU which can take some time, and can make the speedup gained with the parallelization useless. As such, GPU implementations are often done either when the operations to be made are compute intensive, or when the whole pipeline can remain on the GPU (for instance, if the goal is to only display the modified image on screen, you save the need to transfer back the image on the CPU).