OT: Any Linux users here played with gpuOcelot?

mtl1

I stumbled onto gpuOcelot the other day. It's essentially a Linux JIT emulator -- that is, runtime -- for CUDA function calls so that GPGPU programs could be run on different platforms.

gpuOcelot hasn't been actively maintained for over a year and it only supports an older version of CUDA. Still, has anyone tinkered with this before?

Oh and before anyone asks, no it doesn't seem to build/compile on Windows -- at least, not without a lot of hoops and tweaking :P

 

Paging @jimhug too (this sends him a notification, perhaps?) since he originally developed his dynamics simulation to understand GPGPU programming...

Kendall Sears

gpuOcelot is a cool little toy, but that is pretty much it.  It allows one to play with CUDA instructions but none of what it allows one to do is of use for actual coding.

Kendall

mtl1

Yeah, I was primarily wondering how on earth gpuOcelot is able to do JIT compilation on CUDA instructions... o_o

Kendall Sears

mtl1 said:

Yeah, I was primarily wondering how on earth gpuOcelot is able to do JIT compilation on CUDA instructions... o_o

Just like any embedded interpreter in a compiled exe.  The GPCPU commands are accepted in strings, then interpreted into the necessary machine code, attached to the correct initialization routine and then submitted to the GPUs as a runnable routine.  JITs tend to be very minimalistic and have poor performance, Ocelot is no different.  Most code is sent to the GPUs to be run sequentially.  This is why I said that it isn't much help for actual coding.  It can help to determine if a specific set of instructions will indeed succeed, or if they'll abend, but it's not going to help (much) in creating optimized parallel code.  Where it helps is finding whether a piece of code will end up creating stack over/under-runs, invalid or out-of-order memory accesses, register collisions, etc.

Kendall