-
Notifications
You must be signed in to change notification settings - Fork 31
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
cudnn 7 #10
Comments
Can you install distro? I believe this comes with all setups for torch. I will not suggest to compile CUDNN or CUDA. And I do believe there should be torch release supporting the newest CUDA or CUDNN even though it’s no longer under development. |
The problem is that the default cudnn bindings are for cudnn5. But you can get the R7 branch and then it compiles fine for cudnn 7. I'm not finding it much faster than my cpu, is that normal? :) |
Oh, I was about to tell you this: torch/torch7#1133 (comment) So generating optical flow is pretty slow and it's on CPU. but for applying styles you should see improvements on GPU. (run nvidia-smi to see if GPU is actually used) |
Thanks. Is there a way to compute optical flow on the GPU? Would it be faster? :) |
I actually tried OpenACC for deepflow and it does not work well. The reason is there're very limited parallelizations you can do given your image is not so large and the optimization steps shoud be run epoch by epoch which are serial. I would suggest using OpenMP to parallelize the optical flow algorithm if it hasn't been multi-threading. |
Is it possible to recompile this for cudnn 7? I could attempt to do it if I could get some pointers.
The text was updated successfully, but these errors were encountered: