Error when generating CUDA code

26 views
Skip to first unread message

Yarco Hayduk

unread,
Jul 30, 2014, 7:25:14 AM7/30/14
to opt...@googlegroups.com
Hi,

I tried generating CUDA code for a few OptiML applications and failed. I get the following exception:

https://gist.github.com/anonymous/b01204f42734ea325187

I checked out the latest version of https://github.com/stanford-ppl/hyperdsl and followed the instructions at https://stanford-ppl.github.io/Delite/optiml/getting_started.html

I get this error when staging kMeans and other applications as well. Only the basic HelloWorld app compiles to CUDA without throwing exceptions.

Please advise.

Thank you,
Yarco

HyoukJoong Lee

unread,
Jul 31, 2014, 5:00:57 AM7/31/14
to opt...@googlegroups.com
Hi Yarco,

Would you try again without --debug option?
The exception you see is because some of the IR nodes are not generated for CUDA,
which should be fine in normal mode since they will be executed by Scala kernels.

Cheers,
HyoukJoong.


--
You received this message because you are subscribed to the Google Groups "OptiML" group.
To unsubscribe from this group and stop receiving emails from it, send an email to optiml+un...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

Yarco Hayduk

unread,
Jul 31, 2014, 9:09:36 AM7/31/14
to opt...@googlegroups.com
Hello HyoukJoong,

Your suggestion did help. Thank you.
Now I am having a problem with executing the generated CUDA application.

https://gist.github.com/anonymous/60723f6faa50ef4abaa3

I don't have problems when executing the same thing with the --cpp option or with no options (Scala).

Thank you for your help,
Yarco

Yarco Hayduk

unread,
Aug 10, 2014, 4:25:27 PM8/10/14
to opt...@googlegroups.com
Hello again,

Do you have any updates on the issue that I've encountered?

I tried generating CUDA code on OSX; should I perhaps try to run on Linux?

Thank you,
Yarco

HyoukJoong Lee

unread,
Aug 10, 2014, 8:21:28 PM8/10/14
to opt...@googlegroups.com
Hi.

Sorry for the late response.
I just tried running the k-means app using the same input data (mandrill-small), and it's running correctly on the hyperdsl master branch.
I'm running on Ubuntu 12.04, but I assume it should also work for Mac.
I think the issue could be related to your CUDA configuration (hyperdsl/delite/config/delite/CUDA.xml).
If you are using kepler device, then you need to use <arch> 3.0 </arch>.

Cheers,
HyoukJoong.

Yarco Hayduk

unread,
Aug 13, 2014, 8:55:39 AM8/13/14
to opt...@googlegroups.com
Thank you, HyoukJoong!
It did resolve the problem.
Reply all
Reply to author
Forward
0 new messages