[Mlir-commits] [mlir] [mlir][AMDGPU] Add scaled wmma ops for gfx1250 (PR #169854)

[Jakub Kuderski]

================
@@ -1363,6 +1373,136 @@ struct WMMAOpLowering : public ConvertOpToLLVMPattern<WMMAOp> {
   }
 };
 
+struct ScaledWMMAOpLowering : public ConvertOpToLLVMPattern<ScaledWMMAOp> {
+  ScaledWMMAOpLowering(const LLVMTypeConverter &converter, Chipset chipset)
+      : ConvertOpToLLVMPattern<ScaledWMMAOp>(converter), chipset(chipset) {}
+
+  Chipset chipset;
+
+  LogicalResult
+  matchAndRewrite(ScaledWMMAOp op, ScaledWMMAOpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    Location loc = op.getLoc();
+    auto outType =
+        typeConverter->convertType<VectorType>(op.getDestD().getType());
+    if (!outType)
+      return rewriter.notifyMatchFailure(op, "type conversion failed");
+
+    if (chipset < Chipset(12, 5, 0))
+      return op->emitOpError("WMMA scale only supported on gfx1250+");
+
+    int64_t m = op.getM();
+    int64_t n = op.getN();
+    int64_t k = op.getK();
+
+    Type aElemType = getElementTypeOrSelf(op.getSourceA().getType());
+    Type bElemType = getElementTypeOrSelf(op.getSourceB().getType());
+
+    std::optional<uint32_t> aFmtCode = smallFloatTypeToFormatCode(aElemType);
+    std::optional<uint32_t> bFmtCode = smallFloatTypeToFormatCode(bElemType);
+
+    if (!aFmtCode || !bFmtCode)
+      return op.emitOpError("unsupported element types for scaled_wmma");
+
+    // Get scale vector types and determine variant (scale vs scale16)
----------------
kuhar wrote:

```suggestion
    // Get scale vector types and determine variant (scale vs scale16).
```

https://github.com/llvm/llvm-project/pull/169854