[Mlir-commits] [mlir] [mlir][gpu] Support `minimumnumf` and `maximumnumf` in GPU dialect (PR #226903)

[Artem Gindinson]

================
@@ -0,0 +1,21 @@
+// RUN: mlir-opt --convert-gpu-to-nvvm %s | FileCheck %s
----------------
AGindinson wrote:

Thanks for the suggestion. I've reworked it into a matrix of [`subgroup_reduce` + `all_reduce] * [all `min` + `max` variants'], using the autogen approach from the pre-existing test.

One alternative would be to move all `subgroup_reduce` cases into `subgroup-reduce-lowering.mlir` at the same level, which would be a bit less verbose but would also congest that test file. Do you folks have a preference here, @LewisCrawford, @chelini?

https://github.com/llvm/llvm-project/pull/226903