Re: state of clblas/clfft/clsparse
I haven't found anything officially saying they're abandoned upstream,
but none of them have any commits since August 2017, despite open pull
requests.
clblas and clfft currently appear to be mostly OK in Debian (and most of
the bugs that are known have patches; clsparse was removed because of a
preference not to have abandoned software, its actual bug was fixable),
so this is not an urgent problem. However, they don't run their tests,
so this doesn't actually prove that they work.
Other projects with similar (but not directly API-compatible) functionality:
- rocBLAS/rocFFT/rocSPARSE - https://github.com/ROCmSoftwarePlatform
Also by AMD, with no obvious comment on their relationship; I don't
currently know whether they are limited to Radeon hardware, or how
similar the API is
- Apple clfft - supposedly
https://developer.apple.com/mac/library/samplecode/OpenCL_FFT/index.html
but this now redirects to an archive page; an embedded copy is found at
https://github.com/ufo-kit/ufo-filters/tree/master/deps/oclfft
- reikna - https://pypi.org/project/reikna/ - FFT and matrix
multiplication; in Python
- clblast - https://github.com/CNugteren/CLBlast - says it is trying to
be similar to the clblas API, but unclear whether this means switching
is just a search-and-replace to change the name prefix
Reverse dependencies in Debian:
- arrayfire (clblas+clfft) - can use clblast instead of clblas (upstream
discussion of whether to switch to it by default, doesn't mention clblas
abandonment: https://github.com/arrayfire/arrayfire/issues/1956); have
their own fork of clfft, but apparently only to get rid of some warnings
they don't want
- gpyfft (clfft) - Python interface to clfft (not just a 1:1 wrapper,
but if clfft goes maybe this should too); no obvious upstream comment
- libgpuarray (clblas) - can instead use clblast
- ufo-filters (clfft) - can instead use (and upstream embeds) Apple clfft
Reply to: