0

I'm in a situation where I have a (n+1)x(n+1) texture attached to a framebuffer. My idea is to update this texture in 2 passes.

  1. Draw a full screen quad and use the scissor test to mask out the outermost 1 pixels, so I write to the n x n 'inner' of the texture.

  2. Second pass where I draw line primitives to write to the outermost pixels, and update the same texture, using the result of pass 1 as input. The computation on this step depend on the state of the inner (n x n) grid which was computed in the first pass.

In the second pass I would bind the result of the first pass for both reading and writing. I know that according to OpenGL this would yield undefined behaviour, but since I am never reading and writing simultaneously from the same texels, I think it could work.

Is it a good idea to do it like this, or is it perhaps better to draw a full screen quad, and in my glsl shader do a check like:

if (gl_FragCoord.x == 0.5 || gl_FragCoord.x == n + 0.5){
   ...
}
genpfault
  • 51,148
  • 11
  • 85
  • 139
user3042966
  • 55
  • 1
  • 6

1 Answers1

2

As you say, OpenGL says it's undefined behaviour, so it won't work. Your best bet for things like this is to ping-pong between two render targets.

Since I am never reading and writing simultaneously from the same texels, I think it could work.

Unfortunately this doesn't help. Actually, there are extensions common on mobile (particularly Mali and PowerVR GPUs) that allow you to read and write simultaneously, but only if it's the same texels. Google 'Pixel Local Storage' for details.

Columbo
  • 6,648
  • 4
  • 19
  • 30