View Single Post
Old 2009-07-12, 22:15   #9
frmky's Avatar
Jul 2003
So Cal

219410 Posts

I've written just a few simple programs. Besides the obvious things like have lots of threads to mask latencies and access global memory as little as possible, getting the best performance really seems to be trial and error. I was involved in the integration of's RC5 CUDA core. I found that seemingly trivial changes like changing an always positive integer from signed to unsigned, and assigning blockDim.x, blockIdx.x, and threadIdx.x to an integer before doing ANY math made huge differences in the speed of the code. In fact, just the move to CUDA 2.2 has halved the speed of the RC5 code. I recently asked why on the nVidia forums but haven't gotten any replies.
frmky is offline   Reply With Quote