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

[Jakub Kuderski]

================
@@ -631,6 +635,139 @@ 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();
+    auto setOpsel = [&](unsigned idx, int64_t val) {
+      switch (idx) {
+      case 3:
+        op.setScalesIdxA(val);
+        break;
+      case 4:
+        op.setScalesIdxB(val);
+        break;
+      default:
+        break;
+      }
+    };
+
+    // For every scale operand of this ScaledMFMAOp, if the scale is produced by
+    // the extraction of a single scale from some vector, then attempt to
+    // extract 4 values from that vector instead.
+    //
+    // Example: (f8 here means f8E8M0FNU)
+    // %unit = vector.extract %ScaleSrc[offsets] : f8 from vector<...>
+    // %scale = vector.insert %unit, ... : f8 into vector<4xf8>
+    // amdgpu.scaled_mfma(%scale[0] * ...
+    //
+    // rewrite to:
+    //
+    // %reshaped = vector.shape_cast %ScaleSrc : vector<...> to vector<?xf8>
+    // %scale = vector.extract %reshaped[?] : vector<4xf8> from vector<?xf8>
+    // amdgpu.scaled_mfma(%scale[0-3] * ...
+    //
+    // This creates duplicate shape_casts for every use but these will be
+    // removed in CSE.
+    for (auto opIdx : std::array<int64_t, 2>({3, 4})) {
+      auto insertOp = op.getOperand(opIdx).getDefiningOp<vector::InsertOp>();
+      if (!insertOp) {
+        return rewriter.notifyMatchFailure(op,
+                                           "defining op not a vector.insert");
+      }
+      // if the extracted value is not a single scalar, then it has been packed.
----------------
kuhar wrote:

```suggestion
      // If the extracted value is not a single scalar, then it has been packed.
```

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