Package: wnpp
Severity: wishlist
Owner: Michael Lehn <[email protected]>
X-Debbugs-Cc: [email protected], [email protected]

* Package name    : ulmblas
  Version         : 0.4.1
  Upstream Contact: Michael Lehn <[email protected]>
* URL             : https://github.com/michael-lehn/ulmBLAS
* License         : BSD-3-Clause
  Programming Lang: C++, C
  Description     : BLAS with micro kernels selected for the CPU at load time

ulmBLAS implements all routines of the reference BLAS (Basic Linear
Algebra Subprograms) for single, double, complex and double complex
precision with the Fortran 77 and the C interface (CBLAS).  On amd64
libulmblas.so.0 selects at load time the build of the kernels for the
CPU (central processing unit): AVX-512 (Advanced Vector Extensions
512), AVX2 with FMA (fused multiply-add), SSE3 (Streaming SIMD
Extensions 3; SIMD: single instruction, multiple data) or generic
x86-64; on arm64 it uses NEON kernels.  It registers itself as an
alternative for libblas.so.3 and libcblas.so.3, like OpenBLAS and BLIS
(BLAS-like Library Instantiation Software).

The level 3 operations (matrix-matrix products, rank-k updates,
triangular products and solves) run on one blocked algorithm with an
own thread pool; kernels and block sizes are selected by problem size
from tables measured per CPU.  Some micro kernels are taken from BLIS
and OpenBLAS (BSD-3-Clause, notices kept).  All internal buffers are
bounded by constants.

Why in Debian: another BLAS alternative next to OpenBLAS and BLIS,
with a small code base (release tarball of about 1 MB, megabyte).  I am
the upstream author and will maintain the package.

The package will be maintained in the Debian Science team
(https://salsa.debian.org/science-team/ulmblas); I need a sponsor.

Reply via email to