@pkhuong @fclc Briefly, to scratch the personal itch of insufficient documentation.

@mbr @fclc I should tighten up the description at some point, but “here be dragons” is the correct conclusion - the semantics are only easily describable when used as just-mul or just-add (though perhaps if I study it long enough, I’ll find a good description of the combined semantics).

@raph RE the Tenstorrent mention, there’s 32-wide SIMD in addition to the matrix unit, in case that’s useful for your workloads (see https://www.corsix.org/content/tt-wh-part6).

@steve Am skiing in Austria at the moment. Is good.

@fclc Crays come in a size other than large?

@fclc Puns will continue until morale improves?

@steve @dotstdy @pervognsen Or if you’re looking for an excuse to use avx512, VFIXUPIMMPS.

@moonchild @harold And then arm64 came along, with sane and tasteful flag behaviour, but the infinite monkeys wanted to efficiently emulate x86 on arm64, and so rmif was invented.

@moonchild @harold bt/btc/similar modify CF, preserve ZF, and undefined other flags, so a valid implementation could preserve all but CF

@steve So many sparse FLOP/s!