Re: ROCm progress - high-level libraries going smooth
Thanks for the updates!
On Tue, 2022-01-18 at 19:36 +0100, Maxime Chambonnet wrote:
> I uploaded 5 initial packages in the afternoon [1], I have good hopes
> of
> finishing
> high-level libraries tonight, not facing special hurdles.
> Cordell if you are here I have these wonders, what are those shared
> objects?
> libhipblas_fortran.so
> libhipsolver_fortran.so
I did not look into the code, but let me guess on the two shlibs.
I'm a co-maintainer of the blas/lapack packages. The reference
BLAS is written in Fortran, which uses a different memory layout
compared to C/C++ and most other languages:
C/C++ multi-dimensional array is row-major, while
Fortran multi-dimensional array is column major
So when passing pointers between the C/Fortran code, there
has to be some translations. For example, transposing the
matrices by flipping several arguments in the function call.
It won't surprise me if the libhipblas_fortran.so is
used to do such thing.
That said. Just my bold guess.
> librocfft-device-1.so, librocfft-device-2.so, ...
no idea.
> I also added a list of TODO passes in the same team repo [2].
> Thank you Mo for initiating a formal dependency tree! [3]
When we are really ging to upload the packages, or when we
are upgrading ROCm to 5.0.X (imagined). Our eventual progress
will follow the tree.
> [1]
>
> https://salsa.debian.org/rocm-team/community/team-project/-/blob/master/README.md#packaging-progress
> [2]
>
> https://salsa.debian.org/rocm-team/community/team-project/-/blob/master/README.md#simplified-todo-passes
> [3]
>
> https://salsa.debian.org/rocm-team/community/team-project/-/blob/master/README.md#dependency-tree
>
Reply to: