[Mlir-commits] [mlir] [mlir][tosa] Preserve all-NaN max-pool windows (PR #225744)

[Jack Frankland]

FranklandJack wrote:

> Found this while working on the related non-finites issue in #222140 . Noting that `max_pool2d`'s accumulator init has the other half of the same problem: the spec initializes it to `minimum_s<in_out_t>()`, which Table 5 gives as `-infinity` for floats, while the lowering uses `APFloat::getLargest(…, /*negative=*/true)`. Your patch covers the `nan_mode == IGNORE` arm of that initializer; the PROPAGATE arm is still the finite bound. Happy to do that as a follow-up once #222140 lands.

Cool, yeah perhaps that could be done in a separate commit :)

https://github.com/llvm/llvm-project/pull/225744