Do GLSL calculations in isolation speed up with the use of multiple GPUs or at least SLI?
If not, does it mean a proper OpenCL setup would beat GLSL (if it’s coded for SMP)?
This is the second time you’ve asked the question, “Would it be reasonable to use OpenCL to implement a graphics API?”
Is there some particular reason that you would even want to do this? If you really want to, then just do it. We’re not going to stop you.
In answer to your questions, it is highly unlikely that you would be able to write an OpenCL implementation of a renderer that would out-perform the hardware-based OpenGL renderer. Even on two (or more) GPUs.
Shaders (GLSL) are a part of all rendering. SLI accelerates rendering, so SLI accelerates GLSL. If it didn’t, nobody would use it.
Maybe the OP wants to know the other way around: is GeneralPurpose-GPU better implemented via OpenGL or OpenCL?
If you are doing traditional GP-GPU with OpenGL(drawing a fullscreen quad with calculation in fragment shader), you might get the speedup from SLI without having to change your program to support Multi-gpu(each GPU renders top or bottom half of screen).
However, render targets may have to be sent over the bus by the graphics driver, which might eat up much of the gained performance.
In OpenCL, you have to program for each GPU seperately and do any data exchange between the GPUs manually. => More control(and possibly better performance), but also more work to implement.