Hi Frederic, I installed numba and then called cuda.profile_stop() to flush the profile data. Then I could use nvprof and import the profile data into nvvp too! Thanks, Jo
On Tuesday, March 14, 2017 at 12:32:02 AM UTC, nouiz wrote: > > On linux, this was working in the past: > > nvprof python my_python_script.py > > I just double checked and it work with the new back-end and the old > back-end with this as an example: > > export THEANO_FLAGS=device=cuda ; nvprof python -c "import > theano;v=theano.tensor.fvector();f=theano.function([v], v+1); f([1])" > > So the problem seem to be related that windows have different need for > this. Maybe you need to modify Theano to call the function you named. But > maybe it is simpler to install Linux? > > If you try to modify Theano, make sure to modify the new back-end (with > device=cuda). We will get rid of the old one soon and it would be a waste > of time. > > Frédéric > > On Mon, Mar 13, 2017 at 6:14 PM <[email protected] <javascript:>> wrote: > >> Hi there, >> >> I was wanting to use nvprof with theano code on windows command line. >> >> With CUDA code, I compile with: >> >> nvcc Test.cu -ccbin "C:\Program Files (x86)\Microsoft Visual Studio >> 10.0\VC" -o Test1 "C:\Program Files\NVIDIA GPU Computing >> Toolkit\CUDA\v7.5\lib\x64\cuda.lib" "C:\Program Files\NVIDIA GPU Computing >> Toolkit\CUDA\v7.5\lib\x64\cudart.lib" >> >> And then nvprof can be used: >> >> "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5\bin\nvprof.exe" >> Test1.exe >> >> Test1.cu has to have cudaProfilerStop() at the end of main, and have >> #include cuda.h and #include cuda_profiler_api.h, and the linked libraries >> in the compile stage. If it doesn't, nvprof doesn't output any profiling >> information and gives a warning: Some profiling data are not recorded. Make >> sure cudaProfilerStop() or cuProfilerStop() is called before application >> exit to flush profile data. >> >> For a theano python code, I have a .theanorc file with: >> [nvcc] >> compiler_bindir=C:\Program Files (x86)\Microsoft Visual Studio 10.0\VC\bin >> >> [cuda] >> root=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5. >> >> When I do: >> >> nvprof python theano_code.py >> >> The code runs but I don't get any profiling information and the same >> warning about requiring cudaProfilerStop() to flush. So my question is what >> function can I use in the theano code to achieve the same as >> cudaProfilerStop() and what libraries would I need to link in the theanorc >> file for the nvcc compile phase? From what I've seen/googled it seems like >> nvprof should just work with the python/theano code and I've played around >> a bit with editing the theanorc file but I can't get it to work! >> >> Anyone got any tips please? Would be much appreciated! >> Thanks, >> Jo >> >> -- >> >> --- >> You received this message because you are subscribed to the Google Groups >> "theano-users" group. >> To unsubscribe from this group and stop receiving emails from it, send an >> email to [email protected] <javascript:>. >> For more options, visit https://groups.google.com/d/optout. >> > -- --- You received this message because you are subscribed to the Google Groups "theano-users" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. For more options, visit https://groups.google.com/d/optout.
