Introduce arith.scaling_extf and arith.scaling_truncf

mlir/include/mlir/Dialect/Arith/IR/ArithOps.td

@@ -1215,6 +1215,58 @@ def Arith_ExtFOp : Arith_FToFCastOp<"extf", [DeclareOpInterfaceMethods<ArithFast
                           attr-dict `:` type($in) `to` type($out) }];
 }
 
+//===----------------------------------------------------------------------===//
+// Scaling ExtFOp
+//===----------------------------------------------------------------------===//
+def Arith_ScalingExtFOp
+    : Arith_Op<
+          "scaling_extf", [Pure, SameInputOutputTensorDims,
+                           DeclareOpInterfaceMethods<ArithFastMathInterface>,
+                           DeclareOpInterfaceMethods<CastOpInterface>]>,
+      Arguments<(ins FloatLike:$in, FloatLike:$scale,
+          OptionalAttr<Arith_FastMathAttr>:$fastmath)>,
+      Results<(outs FloatLike:$out)> {
+  let summary = "Upcasts input floats using provided scales values following "
+                "OCP MXFP Spec";
+  let description = [{
+  This operation upcasts input floating-point values using provided scale 
+  values. It expects both scales and the input operand to be of the same shape, 
+  making the operation elementwise. Scales are usually calculated per block 
+  following the OCP MXFP spec as described in https://arxiv.org/abs/2310.10537.
+
+  If scales are calculated per block where blockSize != 1, then scales may 
+  require broadcasting to make this operation elementwise. For example, let's 
+  say the input is of shape `<dim1 x dim2 x ... dimN>`. Given blockSize != 1 and 
+  assuming quantization happens on the last axis, the input can be reshaped to 
+  `<dim1 x dim2 x ... (dimN/blockSize) x blockSize>`. Scales will be calculated 
+  per block on the last axis. Therefore, scales will be of shape 
+  `<dim1 x dim2 x ... (dimN/blockSize) x 1>`. Scales could also be of some other 
+  shape as long as it is broadcast compatible with the input, e.g., 
+  `<1 x 1 x ... (dimN/blockSize) x 1>`.
+
+  In this example, before calling into `arith.scaling_extf`, scales must be 
+  broadcasted to `<dim1 x dim2 x dim3 ... (dimN/blockSize) x blockSize>`. Note 
+  that there could be multiple quantization axes. Internally, 
+  `arith.scaling_extf` would perform the following:
+
+    ```
+    resultTy = get_type(result) 
+    scaleTy  = get_type(scale)
+    inputTy = get_type(input)
+    scale.exponent = arith.truncf(scale) : scaleTy to f8E8M0
+    scale.extf = arith.extf(scale.exponent) : f8E8M0 to resultTy
+    input.extf = arith.extf(input) : inputTy to resultTy
+    result = arith.mulf(scale.extf, input.extf)
+    ```
+    It propagates NaN values. Therefore, if either scale or the input element 
+    contains NaN, then the output element value will also be a NaN.
+  }];
+  let hasVerifier = 1;
+  let assemblyFormat =
+      [{ $in `,` $scale (`fastmath` `` $fastmath^)? attr-dict `:` 
+      type($in) `,` type($scale) `to` type($out)}];
+}
+
 //===----------------------------------------------------------------------===//
 // TruncIOp
 //===----------------------------------------------------------------------===//
@@ -1280,6 +1332,63 @@ 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 = "Downcasts input floating point values using provided scales "
+                "values following OCP MXFP Spec";
+  let description = [{
+    This operation downcasts input using the provided scale values. It expects 
+    both scales and the input operand to be of the same shape and, therefore, 
+    makes the operation elementwise. Scales are usually calculated per block 
+    following the OCP MXFP spec as described in https://arxiv.org/abs/2310.10537.
+    Users are required to normalize and clamp the scales as necessary before calling
+    passing them to this operation.  OCP MXFP spec also does the flushing of denorms
+    on the input operand, which should be handled during lowering by passing appropriate 
+    fastMath flag to this operation. 
+
+    If scales are calculated per block where blockSize != 1, scales may require 
+    broadcasting to make this operation elementwise. For example, let's say the 
+    input is of shape `<dim1 x dim2 x ... dimN>`. Given blockSize != 1 and 
+    assuming quantization happens on the last axis, the input can be reshaped to 
+    `<dim1 x dim2 x ... (dimN/blockSize) x blockSize>`. Scales will be calculated 
+    per block on the last axis. Therefore, scales will be of shape 
+    `<dim1 x dim2 x ... (dimN/blockSize) x 1>`. Scales could also be of some other 
+    shape as long as it is broadcast compatible with the input, e.g., 
+    `<1 x 1 x ... (dimN/blockSize) x 1>`.
+
+    In this example, before calling into `arith.scaling_truncf`, scales must be 
+    broadcasted to `<dim1 x dim2 x dim3 ... (dimN/blockSize) x blockSize>`. Note 
+    that there could be multiple quantization axes. Internally, 
+    `arith.scaling_truncf` would perform the following:
+
+    ```
+    scaleTy = get_type(scale)
+    inputTy = get_type(input)
+    resultTy = get_type(result)
+    scale.exponent = arith.truncf(scale) : scaleTy to f8E8M0
+    scale.extf = arith.extf(scale.exponent) : f8E8M0 to inputTy
+    result = arith.divf(input, scale.extf)
+    result.cast = arith.truncf(result, resultTy)
+    ```
+  }];
+  let hasVerifier = 1;
+  let assemblyFormat =
+      [{ $in `,` $scale ($roundingmode^)? (`fastmath` `` $fastmath^)? attr-dict `:` 
+      type($in) `,` type($scale) `to` type($out)}];
+}
+
 //===----------------------------------------------------------------------===//
 // UIToFPOp
 //===----------------------------------------------------------------------===//

mlir/include/mlir/Dialect/Arith/Transforms/Passes.h

@@ -62,6 +62,9 @@ void populateExpandBFloat16Patterns(RewritePatternSet &patterns);
 /// Add patterns to expand Arith f8e8m0 patterns to lower level bitcasts/shifts.
 void populateExpandF8E8M0Patterns(RewritePatternSet &patterns);
 
+/// Add patterns to expand scaling ExtF/TruncF ops to equivalent arith ops
+void populateExpandScalingExtTruncPatterns(RewritePatternSet &patterns);
+
 /// Add patterns to expand Arith ops.
 void populateArithExpandOpsPatterns(RewritePatternSet &patterns);
 

mlir/include/mlir/IR/Builders.h

@@ -60,6 +60,7 @@ class Builder {
                        Attribute metadata = Attribute());
 
   // Types.
+  FloatType getF8E8M0Type();
   FloatType getBF16Type();
   FloatType getF16Type();
   FloatType getTF32Type();

mlir/lib/Dialect/Arith/IR/ArithOps.cpp

@@ -1451,6 +1451,19 @@ bool arith::ExtFOp::areCastCompatible(TypeRange inputs, TypeRange outputs) {
 
 LogicalResult arith::ExtFOp::verify() { return verifyExtOp<FloatType>(*this); }
 
+//===----------------------------------------------------------------------===//
+// ScalingExtFOp
+//===----------------------------------------------------------------------===//
+
+bool arith::ScalingExtFOp::areCastCompatible(TypeRange inputs,
+                                             TypeRange outputs) {
+  return checkWidthChangeCast<std::greater, FloatType>(inputs.front(), outputs);
+}
+
+LogicalResult arith::ScalingExtFOp::verify() {
+  return verifyExtOp<FloatType>(*this);
+}
+
 //===----------------------------------------------------------------------===//
 // TruncIOp
 //===----------------------------------------------------------------------===//
@@ -1565,6 +1578,19 @@ LogicalResult arith::TruncFOp::verify() {
   return verifyTruncateOp<FloatType>(*this);
 }
 
+//===----------------------------------------------------------------------===//
+// ScalingTruncFOp
+//===----------------------------------------------------------------------===//
+
+bool arith::ScalingTruncFOp::areCastCompatible(TypeRange inputs,
+                                               TypeRange outputs) {
+  return checkWidthChangeCast<std::less, FloatType>(inputs.front(), outputs);
+}
+
+LogicalResult arith::ScalingTruncFOp::verify() {
+  return verifyTruncateOp<FloatType>(*this);
+}
+
 //===----------------------------------------------------------------------===//
 // AndIOp
 //===----------------------------------------------------------------------===//

mlir/lib/Dialect/Arith/Transforms/ExpandOps.cpp

@@ -6,10 +6,10 @@
 //
 //===----------------------------------------------------------------------===//
 
-#include "mlir/Dialect/Arith/Transforms/Passes.h"
-
 #include "mlir/Dialect/Arith/IR/Arith.h"
+#include "mlir/Dialect/Arith/Transforms/Passes.h"
 #include "mlir/Dialect/Vector/IR/VectorOps.h"
+#include "mlir/IR/BuiltinTypeInterfaces.h"
 #include "mlir/IR/ImplicitLocOpBuilder.h"
 #include "mlir/IR/TypeUtilities.h"
 #include "mlir/Transforms/DialectConversion.h"
@@ -31,7 +31,6 @@ static Value createConst(Location loc, Type type, int value,
     return rewriter.create<arith::ConstantOp>(
         loc, DenseElementsAttr::get(shapedTy, attr));
   }
-
   return rewriter.create<arith::ConstantOp>(loc, attr);
 }
 
@@ -357,9 +356,10 @@ struct F8E8M0ExtFOpConverter : public OpRewritePattern<arith::ExtFOp> {
     f32Bits = b.create<arith::SelectOp>(isNan, cF32NaN, f32Bits);
     Value result = b.create<arith::BitcastOp>(f32Ty, f32Bits);
     if (resultETy.getIntOrFloatBitWidth() < 32) {
-      result = b.create<arith::TruncFOp>(resultTy, result);
+      result = b.create<arith::TruncFOp>(resultTy, result, nullptr,
+                                         op.getFastmathAttr());
     } else if (resultETy.getIntOrFloatBitWidth() > 32) {
-      result = b.create<arith::ExtFOp>(resultTy, result);
+      result = b.create<arith::ExtFOp>(resultTy, result, op.getFastmathAttr());
     }
     rewriter.replaceOp(op, result);
     return success();
@@ -395,9 +395,10 @@ struct F8E8M0TruncFOpConverter : public OpRewritePattern<arith::TruncFOp> {
     Type f32Ty = cloneToShapedType(operandTy, b.getF32Type());
 
     if (operandETy.getIntOrFloatBitWidth() < 32) {
-      operand = b.create<arith::ExtFOp>(f32Ty, operand);
+      operand = b.create<arith::ExtFOp>(f32Ty, operand, op.getFastmathAttr());
     } else if (operandETy.getIntOrFloatBitWidth() > 32) {
-      operand = b.create<arith::TruncFOp>(f32Ty, operand);
+      operand = b.create<arith::TruncFOp>(
+          f32Ty, operand, op.getRoundingmodeAttr(), op.getFastmathAttr());
     }
     Value f32Bits = b.create<arith::BitcastOp>(i32Ty, operand);
     Value cF32MantissaWidth = createConst(op->getLoc(), i32Ty, 23, rewriter);
@@ -409,6 +410,83 @@ struct F8E8M0TruncFOpConverter : public OpRewritePattern<arith::TruncFOp> {
   }
 };
 
+struct ScalingExtFOpConverter : public OpRewritePattern<arith::ScalingExtFOp> {
+  using OpRewritePattern::OpRewritePattern;
+  LogicalResult matchAndRewrite(arith::ScalingExtFOp op,
+                                PatternRewriter &rewriter) const final {
+    ImplicitLocOpBuilder b(op.getLoc(), rewriter);
+    Value inputOperand = op.getIn();
+    Value scaleOperand = op.getScale();
+    Type scaleTy = scaleOperand.getType();
+    Type scaleETy = getElementTypeOrSelf(scaleOperand);
+    // allow implicit exponent extraction from 16/32 bits floats
+    if (scaleETy.getIntOrFloatBitWidth() >= 16) {
+      scaleETy = b.getF8E8M0Type();
+      scaleTy = cloneToShapedType(scaleTy, scaleETy);
+      scaleOperand = b.create<arith::TruncFOp>(scaleTy, scaleOperand, nullptr,
+                                               op.getFastmathAttr());
+    }
+    if (!llvm::isa<Float8E8M0FNUType>(scaleETy)) {
+      return rewriter.notifyMatchFailure(
+          op, "scaling_extf is using scales of type which can not be converted "
+              "to f8E8M0FNU");
+    }
+    Type resultTy = op.getType();
+    // extf on scale will essentially create floating point number
+    // of type resulTy that is 2^scale and will also propagate NaNs
+    Value scaleExt =
+        b.create<arith::ExtFOp>(resultTy, scaleOperand, op.getFastmathAttr());
+    Value inputExt =
+        b.create<arith::ExtFOp>(resultTy, inputOperand, op.getFastmathAttr());
+    Value result =
+        b.create<arith::MulFOp>(inputExt, scaleExt, op.getFastmathAttr());
+    rewriter.replaceOp(op, result);
+    return success();
+  }
+};
+
+/*
+Expands arith.ScalingTruncFOp(in, scale) into
+  scale = arith.truncf(scale) : scaleTy -> f8E8M0FNU
+  result = arith.truncf(in / (2^scale))
+ */
+struct ScalingTruncFOpConverter
+    : public OpRewritePattern<arith::ScalingTruncFOp> {
+  using OpRewritePattern::OpRewritePattern;
+  LogicalResult matchAndRewrite(arith::ScalingTruncFOp op,
+                                PatternRewriter &rewriter) const final {
+    ImplicitLocOpBuilder b(op.getLoc(), rewriter);
+    Value inputOperand = op.getIn();
+    Value scaleOperand = op.getScale();
+    Type scaleTy = scaleOperand.getType();
+    Type scaleETy = getElementTypeOrSelf(scaleOperand);
+    // allow implicit exponent extraction from 16/32 bits floats
+    if (scaleETy.getIntOrFloatBitWidth() >= 16) {
+      scaleETy = b.getF8E8M0Type();
+      scaleTy = cloneToShapedType(scaleTy, scaleETy);
+      scaleOperand = b.create<arith::TruncFOp>(scaleTy, scaleOperand, nullptr,
+                                               op.getFastmathAttr());
+    }
+    if (!llvm::isa<Float8E8M0FNUType>(scaleETy)) {
+      return rewriter.notifyMatchFailure(
+          op, "scaling_truncf is using scales type which can not be converted "
+              "to f8E8M0FNU");
+    }
+    Type resultTy = op.getType();
+    Type inputTy = inputOperand.getType();
+    // this will create a floating point number of type
+    // inputTy that is 2^scale and will also propagate NaNs
+    scaleOperand =
+        b.create<arith::ExtFOp>(inputTy, scaleOperand, op.getFastmathAttr());
+    Value result = b.create<arith::DivFOp>(inputOperand, scaleOperand,
+                                           op.getFastmathAttr());
+    Value resultCast = b.create<arith::TruncFOp>(
+        resultTy, result, op.getRoundingmodeAttr(), op.getFastmathAttr());
+    rewriter.replaceOp(op, resultCast);
+    return success();
+  }
+};
+
 struct ArithExpandOpsPass
     : public arith::impl::ArithExpandOpsPassBase<ArithExpandOpsPass> {
   using ArithExpandOpsPassBase::ArithExpandOpsPassBase;
@@ -432,7 +510,9 @@ struct ArithExpandOpsPass
       arith::MaximumFOp,
       arith::MinimumFOp,
       arith::MaxNumFOp,
-      arith::MinNumFOp
+      arith::MinNumFOp,
+      arith::ScalingExtFOp,
+      arith::ScalingTruncFOp
     >();
 
     if (includeBf16) {
@@ -492,8 +572,15 @@ void mlir::arith::populateExpandF8E8M0Patterns(RewritePatternSet &patterns) {
       patterns.getContext());
 }
 
+void mlir::arith::populateExpandScalingExtTruncPatterns(
+    RewritePatternSet &patterns) {
+  patterns.add<ScalingExtFOpConverter, ScalingTruncFOpConverter>(
+      patterns.getContext());
+}
+
 void mlir::arith::populateArithExpandOpsPatterns(RewritePatternSet &patterns) {
   populateCeilFloorDivExpandOpsPatterns(patterns);
+  populateExpandScalingExtTruncPatterns(patterns);
   // clang-format off
   patterns.add<
     MaxMinIOpConverter<MaxSIOp, arith::CmpIPredicate::sgt>,
@@ -503,7 +590,7 @@ void mlir::arith::populateArithExpandOpsPatterns(RewritePatternSet &patterns) {
     MaximumMinimumFOpConverter<MaximumFOp, arith::CmpFPredicate::UGT>,
     MaximumMinimumFOpConverter<MinimumFOp, arith::CmpFPredicate::ULT>,
     MaxNumMinNumFOpConverter<MaxNumFOp, arith::CmpFPredicate::UGT>,
-    MaxNumMinNumFOpConverter<MinNumFOp, arith::CmpFPredicate::ULT>
+    MaxNumMinNumFOpConverter<MinNumFOp, arith::CmpFPredicate::ULT> 
    >(patterns.getContext());
   // clang-format on
 }

mlir/lib/IR/Builders.cpp

@@ -34,6 +34,8 @@ Location Builder::getFusedLoc(ArrayRef<Location> locs, Attribute metadata) {
 // Types.
 //===----------------------------------------------------------------------===//
 
+FloatType Builder::getF8E8M0Type() { return Float8E8M0FNUType::get(context); }
+
 FloatType Builder::getBF16Type() { return BFloat16Type::get(context); }
 
 FloatType Builder::getF16Type() { return Float16Type::get(context); }