commit 4662c1b50bf6047274c555ef158ae0643cdc73e2
author bjacob <benoitjacob@google.com> Mon Feb 28 11:24:22 2022 -0500
committer GitHub <noreply@github.com> Mon Feb 28 11:24:22 2022 -0500
tree 782e11ec27b8cc1d82e243470a96b12672079189
parent 5e3d0b8eee2a0635500a07ec400fd18818989d11

aarch64 +i8mm kernels, plus necessary infra. (#8316)

* Add aarch64 +i8mm kernels, plus infra work needed.

With +i8mm for the first time we have a feature that we can't assume all
devices to support, not even our own testing devices.

This forced implementing cpu feature detection (could have used cpuinfo but IIUC
there's interest in not entrenching it in IREE and this was a good
example of to do modern CPUID work on linux/arm). See the cpu_features
library.

There is a new 'requirements' node in the YAML traces for e2e matmul
tests, and when iree-e2e-matmul-test finds that a requirement is not
met, it terminates early, recording a test success but logging a message
about that. This will allow tests to run on CI machines where some
machines support certain features and certain don't, while being green
everywhere and giving as much test coverage as possible. In particular,
only the runtime-testing coverage is dependent on the CI machine, while
the compilation aspect of the tests are always exercised (meaning we
unconditionally compile modules with +i8mm, regardless of whether we
will be able to run them).

The actual asm kernel with i8mm is not good at the moment. That's
because the way we are abstracting a matmul kernel as a loop around a
single-shape 3-iterators vector.contract is proving a bit too limited.
For now we're accumulating experimental data on this topic, not yet
trying to be fast / resolve this.

* clang-format

* build

* buildifier