Skip to content

Commit

Permalink
* Deleted unneeded library "cufft"
Browse files Browse the repository at this point in the history
  • Loading branch information
@tpapaz
tpapaz committed Oct 7, 2013
1 parent 16ad23b commit 9c2a156
Show file tree
Hide file tree
Showing 2 changed files with 0 additions and 0 deletions.
Binary file removed binaries/windows/x64/windows/cufft64_55.dll
View file
Binary file not shown.
Binary file removed binaries/windows/x86/windows/cufft32_55.dll
View file
Binary file not shown.

5 comments on commit 9c2a156

@cysin
Copy link

@cysin cysin commented on 9c2a156 Jun 17, 2014

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I am using svm API to write my own trainer but after looking at this code the cuda-accelerated part is only in cross validation. So I am wondering how can I benefit from cuda if I don't use cross validation?

@cysin
Copy link

@cysin cysin commented on 9c2a156 Jun 17, 2014

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

By the way, I can't find a place to open an issue...

@tpapaz
Copy link
Contributor Author

@tpapaz tpapaz commented on 9c2a156 Jun 17, 2014

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hello cysin,

You are correct, CUDA acceleration takes part only when using cross validation.
Currently there is no other way to benefit from the GPU.

Thanks for taking the time to play around with our implementation.

@barius
Copy link

@barius barius commented on 9c2a156 May 22, 2015

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hi @tpapaz , I have noticed the same thing as @cysin did, that the GPU only takes part in cross-validation.

After some poking around, it seems that the GPU version is slower even if I use the cross-validation, e.g. with a dataset of 200k samples and 16 dimension features, the GPU implementation (based on libsvm 3.1.7, ~25min) is slower than CPU version (latest libsvm 3.20, ~20min).

p.s. My hardware is Intel E5-2620 and Nvidia Tesla K10. The GPU reports only 3% of usage.

p.p.s Also if I use too many samples to train the SVM, say 2million samples, the GPU implementation will burst out a SIGSEGV and leave a HUGE core dump of 2.0TB. Strange.

Regards~

@Elkoumy
Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hello @tpapaz,
I am trying to use your implementation of the SVM on GPU with the same dataset but there is no acceleration,
Please advice.

Please sign in to comment.