[WIP] Make conda builds for all platforms and Python versions

[kohr-h]

I added the meta.yaml file and a build.sh build script for Linux and Mac. For windows builds, we need a bld.bat file like this one, just with proper build instructions.

What's currently known to work:

Linux:

• build
• install
• import
• tests succeed

EDIT: I took the package built on my office computer and installed it on my laptop, where CUDA is located in a different directory. Works nicely.

Mac:

• build
• install
• import
• tests succeed

Windows:

• build
• install
• import
• tests succeed

Instruction on how to build a conda package are here.

Especially environment variables that can be used during build are listed here.

EDIT: Forgot to mention that the build expects the environment variables CUDA_ROOT and CUDA_COMPUTE to be defined. In Linux, you would run

CUDA_ROOT=/opt/cuda CUDA_COMPUTE=30 conda build .

in the conda directory.

[kohr-h]

not updated obviously

[adler-j]

0.4.0

[kohr-h]

I don't think we want to keep the version perfectly in sync, so probably a version after the split is sufficient. It will anyway pick the most recent one.

[adler-j]

Very good initiative, I'll try to get windows settled.

[adler-j]

Should use

$ python -c "import numpy; print(numpy.get_include())"

[kohr-h]

Now everything works as expected with conda's version of boost and CUDA-toolkit. That should also make it easier to get the Windows build working since the paths are more predictable. I've uploaded the last build to Anaconda cloud, users can smoothly install from there.

[adler-j]

I'll see if i can get this working on windows as well

[kohr-h]

I'll see if i can get this working on windows as well

Would be great! I can summarize some of my experiences:

• Despite of what I wrote above, the CUDA toolkit from conda didn't work as expected, but boost is fine. I'll keep trying to make also work, though. Alternatively, one could use the conda CUDA for the Windows build and a distribution CUDA for Unix builds.
• Something fails during the make pyinstall step in the build when using Python 2.7. It tries to download ODL during that step although it's installed as build dependency.
• Don't use conda build from within a conda environment, it can put you into an infinite loop (with conda-build >= 2.0.0. Use the root environment and give --python <version> as option to build for a certain Python version.
• Set CUDA_COMPUTE to 50: lower values fail the package to be usable for some reason, higher values won't work with machines without that compute capability.

[adler-j]

Proposal:

Go ahead with Linux, I'll catch up with windows "sometime".

[kohr-h]

When installing on a new machine using conda, I got a spurious "invalid device function" error. I definitely have a recent enough GPU, so too low compute capability is not the problem. For some reason, I experienced the same when building with CUDA_COMPUTE=30 and installing on a machine with "better" GPU. Any idea what happens @adler-j? I hope it's no problem that I have CUDA 8 here, that would make the build complexity explode.