From what I remember from the CUDA Toolkit 3.2 Readiness TechBrief that was recently posted on the developer site, it stated that a compatibility layer allows you to still use 32 bit kernels on 64 bit architectures. The new API features from SDK versions 3.2 and later will not be added to this layer. Which means if you insist on using 32 bit kernels in 64 bit apps, you won’t be getting all the goodies from CUDA SDK 3.2 and above.
Old 64 bit apps using 32 bit kernels will continue to run.
Just out of curiosity: what field of science or technology requires hand coded cubins? ;) I usually beat nvcc with a stick until it gives me the low register count I need.
Yeah, I saw that compatibility option and its probably the way I’ll end up going. Either that or stick with CUDA 2.3 for a while longer. Or maybe when CUDA 3.2 actually comes out I’ll figure out some nasty hack.
Yeah, I saw that compatibility option and its probably the way I’ll end up going. Either that or stick with CUDA 2.3 for a while longer. Or maybe when CUDA 3.2 actually comes out I’ll figure out some nasty hack.