Most important addition to this release are mixed precision interfaces. This
allows one to, e.g., put in single precision data but carry the computation
out in double precision accuracy.
Note that flame/blis is the upstream for the BLIS library. AMD has forked
BLIS and releases independently from the FLAME project but is strightly
downstream.
This release incorporates the changes needed for us which were previously only
available through a snapshot. It also allows us to drop some install hacks and
brings general improvements to blis itself.
Switch to llvm6 as the compiler.
Reviewed by: zeising (mentor)
Approved by: zeising (mentor)
Differential Revision: https://reviews.freebsd.org/D16522
Among other changes, this improves performance on recent AMD architectures and contains the official runtime kernel selection we asked them for.
Reviewed by: zeising (mentor)
Approved by: zeising (mentor)
Differential Revision: https://reviews.freebsd.org/D15334
Due to this, remove the old explicit kernel selection. Also switch to use devel/llvm50 as a compiler to actually support all the architectures (and upstream compile flags) properly. This hopefully will also help with getting blis to work on other architectures out of the box.
Reviewed by: swills (mentor)
Approved by: swills (mentor)
Differential Revision: https://reviews.freebsd.org/D13264
BLIS is a portable software framework for instantiating high-performance BLAS-like dense linear algebra libraries. The framework was designed to isolate essential kernels of computation that, when optimized, immediately enable optimized implementations of most of its commonly used and computationally intensive operations. BLIS is written in ISO C99 and available under a BSD3CLAUSE. While BLIS exports a new BLAS-like API, it also includes a BLAS compatibility layer which gives application developers access to BLIS implementations via traditional BLAS routine calls. (from their github)
Reviewed by: swills (mentor)
Approved by: swills (mentor)
Differential Revision: https://reviews.freebsd.org/D10811
