GPU programming

I posted this in /g/ a few minutes ago.

Is someone here familiar with GPU programming?
To make this clear in prior: I know Assembly better than C and I intend to use it.

How (from the CPU software in the main memory's point of view) does one initiate contact with the GPU? Hand it kernels and otherwise any communication between. Is the GPU already running in a loop from the very beginning and all we have to do is edit the video memory to let it "escape"?

On second thought: This question might apply just as much to any other CPU cores that might be running in parallel. Although I suspect that is way simpler.

>>4
The GPU doesn't have superiority with mathematical operations.  It has superiority in certain types of SIMD operations.  Learn to tell the difference.  Also, if you're not using CUDA or OpenGL, then what you're doing is writing your own GPU drivers.  That's an abyssal pit of suck from which you'll never recover.

Also, if the words "cache coherency" don't ring any bells (and be honest... you had to look it up, didn't you?) then you are not fit to write GPU drivers.