[Mlir-commits] [mlir] [mlir][AMDGPU] Add canonicalization pattern to pack scales for ScaledMFMAOp (PR #155951)

[llvmlistbot at llvm.org]

================
@@ -631,6 +634,144 @@ LogicalResult TransposeLoadOp::verify() {
   return success();
 }
 
+//===----------------------------------------------------------------------===//
+// ScaledMFMAOp
+//===----------------------------------------------------------------------===//
+
+namespace {
+/// Check if the scales input is used in other scaled mfma's while they exist.
+/// If theyre unused then pack the scales.
+struct PackScales final : OpRewritePattern<ScaledMFMAOp> {
+  using OpRewritePattern::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(ScaledMFMAOp op,
+                                PatternRewriter &rewriter) const override {
+    Location loc = op.getLoc();
+    // If this use of a scale has a non zero opsel, packing has already been
+    // done.
+    auto checkIfUnpackable = [&](OpOperand &op) {
----------------
Muzammiluddin-Syed-ECE wrote:

Removed the function all together, see comment below about the check.

https://github.com/llvm/llvm-project/pull/155951