[Mlir-commits] [mlir] Introduce `arith.scaling_extf` and `arith.scaling_truncf` (PR #141965)

[Umang Yadav]

================
@@ -1280,6 +1318,49 @@ def Arith_TruncFOp :
                           attr-dict `:` type($in) `to` type($out) }];
 }
 
+//===----------------------------------------------------------------------===//
+// Scaling TruncFOp
+//===----------------------------------------------------------------------===//
+
+def Arith_ScalingTruncFOp
+    : Arith_Op<"scaling_truncf",
+               [Pure, SameInputOutputTensorDims,
+                DeclareOpInterfaceMethods<ArithRoundingModeInterface>,
+                DeclareOpInterfaceMethods<ArithFastMathInterface>,
+                DeclareOpInterfaceMethods<CastOpInterface>]>,
+      Arguments<(ins FloatLike:$in, FloatLike:$scale,
+          OptionalAttr<Arith_RoundingModeAttr>:$roundingmode,
+          OptionalAttr<Arith_FastMathAttr>:$fastmath)>,
+      Results<(outs FloatLike:$out)> {
+  let summary =
+      "cast from floating-point to narrower floating-point with scales";
+  let description = [{
+    This operation implements micro-scaling (OCP MXFP) quantization of input using provided scale values.
+    This quantization usually happens over a block of values. All values in that block share same scale value for quantization purposes. 
+    Therefore original input of shape `<dim1 x dim2 ... dimN>` can be thought of as of shape `<dim1 x dim2 x ... (dimN / blockSize) x blockSize>`, 
+    assuming quantization axis is the last axis. 
+    Original scales values therefore would be of shape `<dim1 x dim2 x ... x dimN-1 x (dimN/blockSize)>`. 
+    `arith.scaling_truncf` operation is an elementwise operation. Therefore, before calling into `arith.scaling_truncf`, if `blockSize != 1` then 
+    Scales must be broadcast appropriately to ensure they are of the same shape as the input operand.
----------------
umangyadav wrote:

Done.

https://github.com/llvm/llvm-project/pull/141965