[Mlir-commits] [mlir] [MLIR][Math] add canonicalize-f32-promotion pass (PR #92482)

[Krzysztof Drewniak]

krzysz00 wrote:

Re your most recent comment, I'd want both:

And as to the truncf/extf pairs, the reason I don't want f32 legalization to delete existing ones is that I have exactly one truncf/extf pair I want to keep around for its effect. The code conceptually goes like this)

```mlir
func @test_kernel(%a: tensor<...xbf16>, %b: tensor<...xbf16>) -> tensor<...xbf16>
func @ref_kernel(%a: tensor<...xf32>, %: tensor<...xf32>) -> tensor<...xf32>

func @main() {
  %cTest = launch @test_kernel(...) : tensor<... x bf16>
  %cRef = launch @ref_kernel(...) : tensor<... x f32>
  %cTestF32 = arith.extf %cTest : tensor<... x bf16> to tensor<... x f32>
  // This is the truncf/extf pair I want to preserve. Note that, though I write it on tensors, by the time any sort of
  // math legalization is running, this is all scalar elementwise code.
  // The truncate/extend replicates the fact that @test_kernel accumulates in f32 and rounds down to bf16
  %cRefMatchingTestRounding = arith.truncf %cRef : tensor<...xf32> to tensor<...xbf16>
  %cRefF32 = arith.extf %cRefMatchingTestRounding : tensor<...xbf16> to tensor<...xf32>
   call @verify_nearness(%cTestF32, %cRefF32)
}
```

https://github.com/llvm/llvm-project/pull/92482