This is an archive of the discontinued LLVM Phabricator instance.

[SE] Let users specify CUDA path
ClosedPublic

Authored by jhen on Sep 14 2016, 1:28 PM.

Details

Summary

Add logic to allow users to specify the CUDA path at configuration time.

Diff Detail

Repository
rL LLVM

Event Timeline

jhen updated this revision to Diff 71420.Sep 14 2016, 1:28 PM
jhen retitled this revision from to [SE] Let users specify CUDA path.
jhen updated this object.
jhen added a reviewer: jlebar.
jhen added subscribers: parallel_libs-commits, jprice.
jlebar accepted this revision.Sep 14 2016, 1:31 PM
jlebar edited edge metadata.

Seems simple enough when someone else writes it. :)

This revision is now accepted and ready to land.Sep 14 2016, 1:31 PM
jhen updated this revision to Diff 71429.Sep 14 2016, 2:07 PM
jhen edited edge metadata.
  • Use CMake's standard FindCUDA
jhen added a comment.Sep 14 2016, 2:09 PM

Seems simple enough when someone else writes it. :)

Speaking of which, it turns out we can use CMake's standard FindCUDA module to do everything except find the driver library. :) So I switched to doing things that way. It also allows you to control where it looks for CUDA by using the CUDA_TOOLKIT_ROOT_DIR variable.

jlebar added inline comments.Sep 14 2016, 2:12 PM
streamexecutor/lib/platforms/cuda/cmake/modules/FindLIBCUDA.cmake
9 ↗(On Diff #71420)

Do you still want this?

Does this also need the CUDA library to be added to streamexecutor-config output for --libs?

At the moment (with this patch), if I try and link against SE using streamexecutor-config, I get undefined references to all of the CUDA driver functions. If I manually add -framework cuda (OS X), it all works. The CUDA SAXPY example works for me, but I can see from the verbose ninja output that the example is building with -framework cuda already.

jhen updated this revision to Diff 71453.Sep 14 2016, 3:32 PM
  • streamexecutor-config report CUDA lib
jhen added a comment.Sep 14 2016, 3:32 PM

Does this also need the CUDA library to be added to streamexecutor-config output for --libs?

At the moment (with this patch), if I try and link against SE using streamexecutor-config, I get undefined references to all of the CUDA driver functions. If I manually add -framework cuda (OS X), it all works. The CUDA SAXPY example works for me, but I can see from the verbose ninja output that the example is building with -framework cuda already.

Thanks for bringing this up! I added logic to streamexecutor-config to add in the CUDA library for the --libs output if CUDA is enabled. It seems to work on Ubuntu. I'd definitely be glad to know if it works on OS X as well.

streamexecutor/lib/platforms/cuda/cmake/modules/FindLibcuda.cmake
1 ↗(On Diff #71420)

I think this was a phabricator bug. This whole file is gone from my standpoint (and I'm glad it's gone :)).

In D24580#543239, @jhen wrote:

Thanks for bringing this up! I added logic to streamexecutor-config to add in the CUDA library for the --libs output if CUDA is enabled. It seems to work on Ubuntu. I'd definitely be glad to know if it works on OS X as well.

Alas, this still doesn't work for OS X. It seems that CUDA_DRIVER_LIBRARY is set to the full path to the framework (e.g. /Library/Frameworks/cuda.framework) whereas the flag that needs to be passed is -framework cuda. CMake automagically sorts this out when passing the framework path to target_link_libraries, but I'm not sure there's a simple way of getting it to spit out the correct flag for streamexecutor-config in a platform independent manner.

This doesn't have to hold up this patch BTW, I just thought I'd mention it here since I noticed it.

jhen added a comment.Sep 14 2016, 5:26 PM

Alas, this still doesn't work for OS X. It seems that CUDA_DRIVER_LIBRARY is set to the full path to the framework (e.g. /Library/Frameworks/cuda.framework) whereas the flag that needs to be passed is -framework cuda. CMake automagically sorts this out when passing the framework path to target_link_libraries, but I'm not sure there's a simple way of getting it to spit out the correct flag for streamexecutor-config in a platform independent manner.

What do you think of the idea of doing a check in Python to see if the library base name is *.framework and then emitting the flag -lframework * instead of the raw path, if so? It's hacky, but it seems like it might work pretty much all the time.

This doesn't have to hold up this patch BTW, I just thought I'd mention it here since I noticed it.

I can't tell you how much I appreciate the work you are doing checking StreamExecutor on different platforms.

In D24580#543354, @jhen wrote:

What do you think of the idea of doing a check in Python to see if the library base name is *.framework and then emitting the flag -lframework * instead of the raw path, if so? It's hacky, but it seems like it might work pretty much all the time.

Sounds good to me. It sounds hacky, but I've just looked at the CMake source code for this and that's almost exactly what they do :-)

jhen updated this revision to Diff 71516.Sep 15 2016, 9:27 AM
  • Convert framework library names
jhen added a comment.Sep 15 2016, 9:29 AM

I've just looked at the CMake source code for this and that's almost exactly what they do :-)

I'll take a positive view of that and say that great minds hack alike. :)

The hack is in now. Please let me know if it works for you. Thanks!

In D24580#543785, @jhen wrote:

The hack is in now. Please let me know if it works for you. Thanks!

Great, it now works perfectly :-)

This revision was automatically updated to reflect the committed changes.
parallel-libs/trunk/streamexecutor/lib/CMakeLists.txt