[Mlir-commits] [mlir] [mlir][tosa][tosa-to-linalg] Fix rescale with double rounding failing validation (PR #184787)

[Luke Hutton]

================
@@ -114,11 +118,15 @@ void mlir::tosa::registerTosaToLinalgPipelines() {
       [](OpPassManager &pm) {
         TosaToLinalgOptions tosaToLinalgOptions;
         TosaToLinalgNamedOptions tosaToLinalgNamedOptions;
+        TosaAttachTargetOptions tosaAttachTargetOptions;
+        tosaAttachTargetOptions.profiles = {"pro_int", "pro_fp"};
+        // TODO: Populate all the extensions that this conversion supports
+        tosaAttachTargetOptions.extensions = {"doubleround"};
         TosaValidationOptions validationOptions;
         validationOptions.strictOpSpecAlignment = false;
         validationOptions.allowInvalidOpDatatypeCombinations = false;
----------------
lhutton1 wrote:

I'm not too familiar with the conversion logic, and in particular where `addTosaToLinalgPasses` is being used externally. I wanted to make sure this change didn't disrupt those uses by only adding target options to `registerTosaToLinalgPipelines` which seems to me to be a test/demo pipeline. If you think this is more widely useful, I'm happy to set the defaults in `addTosaToLinalgPasses` instead?

https://github.com/llvm/llvm-project/pull/184787