Sunday, August 1, 2010

[From Jeffrey] cuda-hydro speedup

This is going to be a very short update.

I improved the CUDA code in nearly every way I know, so now is the moment of truth. How much faster is it?

16 times.

This is on k3, comparing the speed of my c code to my CUDA c code under the exact same conditions.

This is slightly slower than I have hoped. Probably I'm still doing something wrong.

Another disappointing thing is the maximum resolution I can get is still just 732 by 732. If I push it to 1k by 1k, I get a segmentation fault that indicates memory shortage, which is confusing since the amount of memory the program needs should not be that much.

The next thing I plan to do is to run it on k4/k5. Sounds like a simple thing to do, but unfortunately k4 & k5 don't have mathgl installed yet. Guess I need to do that first.