execution time is so long about 70 seconds..

[Si-Dong Yang]

i read about several topic's which have common with execution time and execution file size..

It's similar with my problem but i can't fix it..

anyway my program stops at cudppPlan() and it waits about 70 seconds.

my execution file is about 70MB.. I think it's strongly have something with my problem..

i'm using ubuntu 12.04, GTX780, CUDA 5.5

i compiled cuddp 2.1 with sm_35 and my application too

and how can i change CUDA_CACHE_MAXSIZE? i use export CUDA_CACHE_MAXSIZE=1024 but it isn't work..

[John Owens]

Almost certainly the issue is that you’re compiling the library with a different compute capability from your GPU. This sort of delay has always been caused by this exact issue, and every time one of us has had this problem, even though we think we’ve done the right thing, it always turns out that the 70 seconds is runtime compilation of CUDPP to our actual compute version. This might be a CUDA_CACHE_MAXSIZE issue, as you note, see below

CUDPP is a *big* library. 70 MB does not surprise me. We do not currently have a good solution for making it smaller.

If you’re using 1024, that might be the problem, actually. Try something much much larger than 1024 bytes: 268435456 (that’s 256 MB) for starters.

JDO

[Mark Harris]

If you get the compilation to the correct SM version right, then the JIT cache won't come into play.  Do you have multiple GPUs installed?

--
You received this message because you are subscribed to the Google Groups "CUDPP" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cudpp+un...@googlegroups.com.
To post to this group, send email to cu...@googlegroups.com.
Visit this group at http://groups.google.com/group/cudpp.
For more options, visit https://groups.google.com/groups/opt_out.