Build CuPy against ROCm?

[leofang]

This issue won't be resolvable in the foreseeable future (at least I don't think so -- lots of extra work and time investment are needed), but it's still worth being kept track.

CuPy recently adds experimental support for ROCm (targeting AMD GPUs): https://docs-cupy.chainer.org/en/latest/install_rocm.html.

Now, there's a roctools from Numba: https://anaconda.org/numba/roctools, but based on the Numba issue tracker and the download times, I doubt it's been extensively tested. Moreover, it doesn't provide the full ROCm software stack, just some LLVM bindings. Users still need to install the ROCm environment themselves: http://numba.pydata.org/numba-doc/latest/roc/overview.html#installation.

For CuPy + ROCm to work in conda, I imagine we need the following things to work:

• a ROCm compiler stack supported by Anaconda (as NVCC is now)
• a ROCm counterpart for cudatoolkit (can roctools be reused here?)
• some docker images that support ROCm
• some ROCm-enabled CI infrastructure for building, testing and shipping the recipe

Disclaimer: I know nothing about ROCm. 🙂

[isuruf]

ROCM compiler stack is on conda-forge, https://github.com/conda-forge/staged-recipes/issues/10123

[leofang]

Cool, thanks for the info @isuruf! Any estimation for when the ROCm support will be up and running?

[isuruf]

According to cupy docs, we need hipblas hipsparse rocrand. I'm waiting on https://github.com/RadeonOpenCompute/clang-ocl/issues/18 to go further.

[leofang]

According to cupy docs, we need hipblas hipsparse rocrand.

A quick update: rocthrust is also needed now.

[leofang]

Due to 1. lack of interest from AMD, 2. lack of bandwidth, 3. potential conflict of interests, I am closing this issue as "won't fix".