[Starpu-devel] Finding Nvidia capabilities when configuring StarPU.

[Cyril Roelandt <cyril.roelandt@inria.fr>]

Hi !

On Hydra, some tests fail with CUDA on Darwin. The Nvidia GPU is a 1.1 
card, and the examples are compiled with "-arch sm_13" (cf. 
examples/Makefile.am). An easy fix would be to either use "-arch sm_11" 
or no option at all. The thing is, using "-arch sm_xx" is  probably good 
for performance.

On a related topic, I had to find the card capabilities for r.7325, but 
this patch still isn't perfect. What we should do is:
- if there is no card >= 1.3, disable the cuda_funcs in the example;
- if there is no card < 1.3, do not use the .can_execute field;
- otherwise, use both fields.

We could do something similar for the NVCCFLAGS in examples/Makefile.am: 
add "-sm_$(min_capabilities)".

So we basically need to find the minimum and maximum capabilities when 
running configure; in order to do that, we need to run NVCC. I'm not 
sure this can be done cleanly with AC_LANG, though. Does anyone see how 
to do this in a non-hackish way ?

Cyril.