[PATCH] D70951: [Matrix] Use fmuladd when lowering matrix.multiply.

[Adam Nemet via Phabricator via llvm-commits]

anemet added a comment.

> We should be able to use fmuladd instead of fmul/fadd, as this just
>  reduces the rounding error between operations.

I am concerned about the stability of the result in the presence of the FMA and non-FMA targets.  We ought to generate the same result in IEEE fashion but with this we wouldn't.  I think that this should still be driven by the fp-contract fp flag and perhaps we should have command-line option that enables this specifically for matmul?


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D70951/new/

https://reviews.llvm.org/D70951