blis

Add linux-cortexa57 build files

• Resolve Windows instability by reverting back to the v0.7 version of the vendored Blis library we'd been using. This was surprisingly hard to do, because I had a lot of trouble getting it to compile due to conflicts between cpuid.h and intrin.h on Windows.
• Widen the runtime numpy compatibility range to allow usage alongside numpy 1 and numpy 2. We still build against numpy 2 though.

The 0.9 version of Blis vendored in this package's v1.0 had memory access errors on Windows.

I've had trouble rolling back to v0.7 to address this, as it isn't building on Github Actions (some environment/compiler toolchain version issue). I've therefore updated the vendored Blis to v1.0, which seems to address the crashes.

I've also widened the numpy compatibility across v1 and v2. I had thought we had to use the same numpy version we built against, but this isn't the case.

MacOS and Linux ARM wheels were missing from the v1.0.0 release, as we've migrated to a more streamlined process to build wheels using only Github-hosted Actions runners. This release restores support for MacOS ARM wheels. Support for Linux ARM runners is coming soon, but currently only available for private repos. Unfortunately QEMU emulation is too slow for the build.

Depend on numpy v2, instead of v1. This introduces binary incompatibilities, so libraries won't easily be able to depend across this version and previous. There's therefore no other changes.

Incrementing to 1.0 as we've been stable a while, so no reason to be 0.x.

• Make BLIS symbols visible when compiled with gcc/clang (#71)
• Use uninitialized array for gemm when beta is zero (#72)
• Fix OOB reads in certain sgemm Haswell kernels (#73, #74)

• Update to BLIS v0.9.0
• Add cdef nogil sgemm and saxpy functions (#70)

Comment out inclusion of cpuid.h header

• Package updates and binary wheels for python 3.12.

• Restrict build to Cython 0.29.x due to issues with Cython 3 for windows.

• Package updates and binary wheels for python 3.11.

• Add cdef nogil sgemm and saxpy functions (#70)
• Make BLIS symbols visible when compiled with gcc/clang (#71)
• Use uninitialized array for gemm when beta is zero (#72)

• Add support for Windows generic target.

Thanks to @nsait-linaro for this contribution!

• Fix issue #61: Support multi-token compiler options such as "ccache gcc".
• Fix issue #65: Check shape compatibility in gemm.

• Fall back to Zen 2 kernels on Zen 3 CPU

This version updates the underlying Blis library to v0.7.0, which brings a number of performance improvements, including better support for AMD architectures.

The new version requires a relatively recent compiler such as gcc v9 to compile the kernels for the more recent architectures. You can set BLIS_ARCH=generic or BLIS_ARCH=x86_64_no_skx to avoid the need for this.

I haven't yet added build logs for ARM systems, so you'll need to use BLIS_ARCH=generic for now to build for ARM hardware.

• Add support for ARM-64 cortexa53, used in e.g. Raspberry Pi. To build a non-x86_64 architecture, set the environment variable BLIS_ARCH=cortexa53 prior to installation.
• Support for read-only numpy arrays, via const declarations.
• Fix build system to allow building on conda-forge.

• Change license to BSD.
• Update vendored blis to v0.5.1.