llvm-project

faea5734 - [mlir][tosa] Extend allow-non-finites to max_pool2d (#229869)

[mlir][tosa] Extend allow-non-finites to max_pool2d (#229869) `TosaToLinalgNamed` seeds float max pooling with `APFloat::getLargest`, i.e. `-3.40282347E+38` for f32. The spec asks for `-infinity`: MAX_POOL2D (§2.3.8) initializes its accumulator with `minimum_s<in_out_t>()`, and Table 5 (§1.9) defines that as `-infinity` for the floating-point types. This adds an `allow-non-finites` option to `tosa-to-linalg-named` that selects the conforming seed, mirroring the option added to `tosa-to-linalg` in #222140. It is off by default, so nothing changes unless you ask for it — the finite seed stays available for targets that can't represent infinities, which includes TOSA's own `fp8e4m3_t`. This is the follow-up I mentioned on #222140, where `max_pool2d` was left out to keep that review small. The `nan_mode = IGNORE` arm is deliberately untouched. The spec requires a NaN seed there and #225744 already implements it, so between the two patches both arms of that one spec line are now covered. Tests assert the NaN seed under *both* settings to pin that down. One refactor came along for the ride: `getFloatMinMaxIdentity` moves out of `TosaToLinalg.cpp` into `Tosa/Utils/ConversionUtils`, which both conversions already include. The `semanticsHasInf` guard inside it is load-bearing — `APFloat::getInf` silently returns a NaN under NanOnly semantics — so one shared copy is better than two that can drift. Testing: `tosa-to-linalg-named.mlir` gains a fourth RUN line with `allow-non-finites=true`, so every existing case in the file is now checked under both settings. Only the four seed constants differ, split `FINITE`/`NONFINITE`. Added PROPAGATE-mode bf16 and f16 cases, which had no coverage before and are what pin the non-f32 encodings. `check-mlir` is clean. --- Parts of this patch were drafted with AI assistance (Claude). I reviewed, tested, and edited the result, and I'm accountable for it. Co-authored-by: Hariprasad Ravishankar <hravisha@ls-01-1c-a0-b8-77-e7-5d.dhcp.mathworks.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>