[Mlir-commits] [mlir] [Arith][Transforms] Adds Truncf f32 to f4e2m1 (PR #144157)

[Jakub Kuderski]

================
@@ -366,6 +378,120 @@ struct F8E8M0ExtFOpConverter : public OpRewritePattern<arith::ExtFOp> {
   }
 };
 
+/// Conversion from F32 to F4E2M1 according to the OCP Spec:
+/// www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf
+///
+/// The spec requiers us to perform Round to Nearest, Ties to Even.
+///
+/// This means that after rounding, we should break ties by choosing the option
+/// which results in a mantissa of 0 in the least significant digit.
+///
+/// Table of representable values in F4E2M1:
+///
+/// Note: x is sign bit
+/// | Binary | Value ( + / - ) 
+/// | x000   | 0.0
+/// | x001   | 0.5
+/// | x010   | 1.0
+/// | x011   | 1.5
+/// | x100   | 2.0
+/// | x101   | 3.0
+/// | x110   | 4.0
+/// | x111   | 6.0
+///
+/// Conversion procedure: 
+///   Step 1: Clamp to representable bounds.
+///   Step 2: Convert exponent by adjusting bias.
+///   Step 3: Set mantissa to first bit.
+///   Step 4: Special consideration for subnormal and zero exponent.
+///   Step 5: Round up if necessary, if mantissa[1:] greater than 1000000 or subnormal.
+struct F4E2M1TruncFOpConverter : public OpRewritePattern<arith::TruncFOp> {
+  using OpRewritePattern::OpRewritePattern;
+  LogicalResult matchAndRewrite(arith::TruncFOp op,
+                                PatternRewriter &rewriter) const final {
+    ImplicitLocOpBuilder b(op.getLoc(), rewriter);
+    Value operand = op.getOperand();
+    Type operandTy = operand.getType();
+    Type resultTy = op.getType();
+    Type operandETy = getElementTypeOrSelf(operandTy);
+    Type resultETy = getElementTypeOrSelf(resultTy);
+
+    if (!llvm::isa<Float4E2M1FNType>(resultETy)) {
----------------
kuhar wrote:

```suggestion
    if (!isa<Float4E2M1FNType>(resultETy)) {
```

https://github.com/llvm/llvm-project/pull/144157