I’v used CUDA profiler with cuda 2.0 and all was fine, now i’v upgraded to cuda 2.1 and the new profiler shows only CPU time, GPU time and ocupancy
all other counters are unavail (I’v checked for example ‘branch’ and ‘instructions’ in Session settings, but none of them shows in a final table).
I have defined those two in environment
CUDA_PROFILE = 1
There is something else I must set to enable profiling ?
I’m profiling 32 bit app, compiled with cuda x86 but all this on Vista x64, is there any problem with this ?
As far as I know, those performance counter have never been available in Vista. Check the release notes, it is a known issue
The old version of the visual profiler would allow you to select the checkboxes even if they did nothing. It would appear that the current version queries the hardware and disables the chechboxes for unsupported counters (another person posting to the forums found that gld_incoherenet and related counters are grayed out on G200 chips.)
So, there is no way to profile app that uses DX10 and CUDA (and cuda interacts with DX10) ? damn
I’v checked both on vista x86 and vista x64 (latest drivers, cuda2.1, GF 280GTX) and in both cases the profiler does not show
for example warp serialization counter (the checkbox is not greyed out, i can freelly check it, but it changes nothing in profiler output )
One can only hope that NVIDIA adds the support in CUDA 2.2…