Re: RFS: hipfft/5.5.1-1 [ITP] -- portable interface for Fast Fourier Transforms on the GPU
Hi Christian,
On 2023-06-17 06:11, Christian Kastner wrote:
The source of rocFFT is only ~900KB and I think this would justify an
exception from the guidelines to not embed copies of other sources.
Just from the gut, I'd say the added benefit of having tests outweigh
the minor inconvenience of this code copy, and the typical other
argument "security" [1] doesn't really apply here either.
Ok. I will look into how to do this.
The only issue I found on a first examination was that a B-D on hipcc
was missing, causing a FTBFS. Added + pushed it.
That's embarrassing. I will have to adjust my process to ensure that
mistakes like that are caught earlier in the process. As is often the
case, the error was introduced in a revision to the original design. I
went back and forth on whether to use hipcc, as it's only really
required for the tests.
I should probably make a checklist for the various things that cannot be
verified automatically in my packaging process. An automatic check of
the build dependencies would be better, but I've been using gbp with
debuild and I haven't found a way to do that yet. I'll put learning
sbuild on my todo list, so I can use a proper chroot.
You also reminded me to review the dev package dependencies and verify
them by building the examples. I've updated the package with the
necessary changes.
Tangential: B-D contains cmake and rocm-cmake. I guess we could have
rocm-cmake depend on cmake? Or is there a use case where a user might
want to install rocm-cmake, without cmake?
The rocm-cmake package provides a library of cmake scripts. I'm not sure
what the right answer is. Similar packages, such as cmake-extras or
extra-cmake-modules don't seem to be consistent about this. That said, I
can't think of any way to use rocm-cmake without cmake.
Thanks,
Cory Bloor
Reply to: