[mlir][arith] Support wide int cast emulation

[kuhar]

Summary

Add support for arith.extsi, arith.extui, and arith.trunci ops.

Tested by checking the results for all 16-bit inputs when emulating i16 with i8.

[kuhar]

Rebased.

[kuhar]

Removed accidentally added includes.

[Mogball]

the implementation lgtm

mlir/lib/Dialect/Arithmetic/Transforms/EmulateWideInt.cpp

277

This is unnecessary. The operands to adaptor will always be materialized as legal for the target.

281

Type conversion can fail. And the pattern should if it does.

318–323

same here

[kuhar]

Fixed type checks for ExtSI and ExtUI. Rebased.

[kuhar]

Do you have any other suggestions, @Mogball?

[Mogball]

LGTM

mlir/lib/Dialect/Arithmetic/Transforms/EmulateWideInt.cpp

 // Extracts the `input` vector slice with elements at the last dimension offset
 // by `lastOffset`. Returns a value of vector type with the last dimension
 // reduced to x1 or fully scalarized, e.g.:
 //   - vector<3x2xi16> --> vector<3x1xi16>
 //   - vector<2xi16>   --> i16
 static Value extractLastDimSlice(ConversionPatternRewriter &rewriter,
                                  Location loc, Value input,
                                  int64_t lastOffset) {
-  llvm::ArrayRef<int64_t> shape = input.getType().cast<VectorType>().getShape();
+  ArrayRef<int64_t> shape = input.getType().cast<VectorType>().getShape();
   assert(lastOffset < shape.back() && "Offset out of bounds");
 
   // Scalarize the result in case of 1D vectors.
   if (shape.size() == 1)
     return rewriter.create<vector::ExtractOp>(loc, input, lastOffset);
 
   SmallVector<int64_t> offsets(shape.size(), 0);
   offsets.back() = lastOffset;

 // with the first element at offset 0 and the second element at offset 1.
 static std::pair<Value, Value>
 extractLastDimHalves(ConversionPatternRewriter &rewriter, Location loc,
                      Value input) {
   return {extractLastDimSlice(rewriter, loc, input, 0),
           extractLastDimSlice(rewriter, loc, input, 1)};
 }
 
+// Performs a vector shape cast to drop the trailing x1 dimension. If the
+// `input` is a scalar, this is a noop.
+static Value dropTrailingX1Dim(ConversionPatternRewriter &rewriter,
+                               Location loc, Value input) {
+  auto vecTy = input.getType().dyn_cast<VectorType>();
+  if (!vecTy)
+    return input;
+
+  // Shape cast to drop the last x1 dimention.
+  ArrayRef<int64_t> shape = vecTy.getShape();
+  assert(shape.size() >= 2 && "Expected vector with at list two dims");
+  assert(shape.back() == 1 && "Expected the last vector dim to be x1");
+
+  auto newVecTy = VectorType::get(shape.drop_back(), vecTy.getElementType());
+  return rewriter.create<vector::ShapeCastOp>(loc, newVecTy, input);
+}
+
+// Performs a vector shape cast to append an x1 dimension. If the
+// `input` is a scalar, this is a noop.
+static Value appendX1Dim(ConversionPatternRewriter &rewriter, Location loc,
+                         Value input) {
+  auto vecTy = input.getType().dyn_cast<VectorType>();
+  if (!vecTy)
+    return input;
+
+  // Add a trailing x1 dim.
+  auto newShape = llvm::to_vector(vecTy.getShape());
+  newShape.push_back(1);
+  auto newTy = VectorType::get(newShape, vecTy.getElementType());
+  return rewriter.create<vector::ShapeCastOp>(loc, newTy, input);
+}
+
 // Inserts the `source` vector slice into the `dest` vector at offset
 // `lastOffset` in the last dimension. `source` can be a scalar when `dest` is a
 // 1D vector.
 static Value insertLastDimSlice(ConversionPatternRewriter &rewriter,
                                 Location loc, Value source, Value dest,
                                 int64_t lastOffset) {
-  llvm::ArrayRef<int64_t> shape = dest.getType().cast<VectorType>().getShape();
+  ArrayRef<int64_t> shape = dest.getType().cast<VectorType>().getShape();
   assert(lastOffset < shape.back() && "Offset out of bounds");
 
   // Handle scalar source.
   if (source.getType().isa<IntegerType>())
     return rewriter.create<vector::InsertOp>(loc, source, dest, lastOffset);
 
   SmallVector<int64_t> offsets(shape.size(), 0);
   offsets.back() = lastOffset;
matchAndRewrite(arith::AddIOp op, OpAdaptor adaptor,
     Value resultVec =
         constructResultVector(rewriter, loc, newTy, {lowSum.getSum(), high});
     rewriter.replaceOp(op, resultVec);
     return success();
   }
 };
 
 //===----------------------------------------------------------------------===//
+// ConvertExtSI
+//===----------------------------------------------------------------------===//
+
+struct ConvertExtSI final : OpConversionPattern<arith::ExtSIOp> {
+  using OpConversionPattern::OpConversionPattern;
+
+  LogicalResult
+  matchAndRewrite(arith::ExtSIOp op, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    Location loc = op->getLoc();
+    auto newTy = getTypeConverter()
+                     ->convertType(op.getType())
+                     .dyn_cast_or_null<VectorType>();
+    if (!newTy)
+      return rewriter.notifyMatchFailure(loc, "unsupported type");
+
+    Type newResultComponentTy = reduceInnermostDim(newTy);
+
+    // Sign-extend the input value to determine the low half of the result.
+    // Then, check if the low half is negative, and sign-extend the comparison
+    // result to get the high half.
+    Value newOperand = appendX1Dim(rewriter, loc, adaptor.getIn());
+    Value extended = rewriter.createOrFold<arith::ExtSIOp>(
+        loc, newResultComponentTy, newOperand);
+    Value operandZeroCst = rewriter.create<arith::ConstantOp>(
+        loc, rewriter.getZeroAttr(newResultComponentTy));
+    Value signBit = rewriter.create<arith::CmpIOp>(
+        loc, arith::CmpIPredicate::slt, extended, operandZeroCst);
+    Value signValue =
+        rewriter.create<arith::ExtSIOp>(loc, newResultComponentTy, signBit);
+
+    Value resultVec =
+        constructResultVector(rewriter, loc, newTy, {extended, signValue});
+    rewriter.replaceOp(op, resultVec);
+    return success();
+  }
+};
+
+//===----------------------------------------------------------------------===//
+// ConvertExtUI
+//===----------------------------------------------------------------------===//
+
+struct ConvertExtUI final : OpConversionPattern<arith::ExtUIOp> {
+  using OpConversionPattern::OpConversionPattern;
+
+  LogicalResult
+  matchAndRewrite(arith::ExtUIOp op, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    Location loc = op->getLoc();
+    auto newTy = getTypeConverter()
+                     ->convertType(op.getType())
+                     .dyn_cast_or_null<VectorType>();
+    if (!newTy)
+      return rewriter.notifyMatchFailure(loc, "unsupported type");
+
+    Type newResultComponentTy = reduceInnermostDim(newTy);
+
+    // Zero-extend the input value to determine the low half of the result.
+    // The high half is always zero.
+    Value newOperand = appendX1Dim(rewriter, loc, adaptor.getIn());
+    Value extended = rewriter.createOrFold<arith::ExtUIOp>(
+        loc, newResultComponentTy, newOperand);
+    Value zeroCst = rewriter.create<arith::ConstantOp>(
+        op->getLoc(), rewriter.getZeroAttr(newTy));
+    Value newRes = insertLastDimSlice(rewriter, loc, extended, zeroCst, 0);
+    rewriter.replaceOp(op, newRes);
+    return success();
+  }
+};
+
+//===----------------------------------------------------------------------===//
+// ConvertTruncI
+//===----------------------------------------------------------------------===//
+
+struct ConvertTruncI final : OpConversionPattern<arith::TruncIOp> {
+  using OpConversionPattern::OpConversionPattern;
+
+  LogicalResult
+  matchAndRewrite(arith::TruncIOp op, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    Location loc = op.getLoc();
+    // Check if the result type is legal for this target. Currently, we do not
+    // support truncation to types wider than supported by the target.
+    if (!getTypeConverter()->isLegal(op.getType()))
+      return rewriter.notifyMatchFailure(loc,
+                                         "unsupported truncation result type");
+
+    // Discard the high half of the input. Truncate the low half, if necessary.
+    Value extracted = extractLastDimSlice(rewriter, loc, adaptor.getIn(), 0);
+    extracted = dropTrailingX1Dim(rewriter, loc, extracted);
+    Value truncated =
+        rewriter.createOrFold<arith::TruncIOp>(loc, op.getType(), extracted);
+    rewriter.replaceOp(op, truncated);
+    return success();
+  }
+};
+
+//===----------------------------------------------------------------------===//
 // Pass Definition
 //===----------------------------------------------------------------------===//
 
 struct EmulateWideIntPass final
     : arith::impl::ArithmeticEmulateWideIntBase<EmulateWideIntPass> {
   using ArithmeticEmulateWideIntBase::ArithmeticEmulateWideIntBase;
 
   void runOnOperation() override {
void arith::populateWideIntEmulationPatterns(
     WideIntEmulationConverter &typeConverter, RewritePatternSet &patterns) {
   // Populate `func.*` conversion patterns.
   populateFunctionOpInterfaceTypeConversionPattern<func::FuncOp>(patterns,
                                                                  typeConverter);
   populateCallOpTypeConversionPattern(patterns, typeConverter);
   populateReturnOpTypeConversionPattern(patterns, typeConverter);
 
   // Populate `arith.*` conversion patterns.
-  patterns.add<ConvertConstant, ConvertAddI>(typeConverter,
+  patterns.add<
+      // Misc ops.
+      ConvertConstant,
+      // Binary ops.
+      ConvertAddI,
+      // Extension and truncation ops.
+      ConvertExtSI, ConvertExtUI, ConvertTruncI>(typeConverter,
                                                  patterns.getContext());
 }

mlir/test/Dialect/Arithmetic/emulate-wide-int.mlir

 // CHECK-NEXT:    [[SUM_H1:%.+]] = arith.addi [[SUM_H0]], [[HIGH1]] : vector<4x1xi32>
 // CHECK:         [[INS0:%.+]]   = vector.insert_strided_slice [[SUM_L]], {{%.+}} {offsets = [0, 0], strides = [1, 1]} : vector<4x1xi32> into vector<4x2xi32>
 // CHECK-NEXT:    [[INS1:%.+]]   = vector.insert_strided_slice [[SUM_H1]], [[INS0]] {offsets = [0, 1], strides = [1, 1]} : vector<4x1xi32> into vector<4x2xi32>
 // CHECK-NEXT:    return [[INS1]] : vector<4x2xi32>
 func.func @addi_vector_a_b(%a : vector<4xi64>, %b : vector<4xi64>) -> vector<4xi64> {
     %x = arith.addi %a, %b : vector<4xi64>
     return %x : vector<4xi64>
 }
+
+// CHECK-LABEL: func @extsi_scalar
+// CHECK-SAME:    ([[ARG:%.+]]: i16) -> vector<2xi32>
+// CHECK-NEXT:    [[EXT:%.+]]  = arith.extsi [[ARG]] : i16 to i32
+// CHECK-NEXT:    [[SZ:%.+]]   = arith.constant 0 : i32
+// CHECK-NEXT:    [[SB:%.+]]   = arith.cmpi slt, [[EXT]], [[SZ]] : i32
+// CHECK-NEXT:    [[SV:%.+]]   = arith.extsi [[SB]] : i1 to i32
+// CHECK-NEXT:    [[VZ:%.+]]   = arith.constant dense<0> : vector<2xi32>
+// CHECK-NEXT:    [[INS0:%.+]] = vector.insert [[EXT]], [[VZ]] [0] : i32 into vector<2xi32>
+// CHECK-NEXT:    [[INS1:%.+]] = vector.insert [[SV]], [[INS0]] [1] : i32 into vector<2xi32>
+// CHECK:         return [[INS1]] : vector<2xi32>
+func.func @extsi_scalar(%a : i16) -> i64 {
+    %r = arith.extsi %a : i16 to i64
+    return %r : i64
+}
+
+// CHECK-LABEL: func @extsi_vector
+// CHECK-SAME:    ([[ARG:%.+]]: vector<3xi16>) -> vector<3x2xi32>
+// CHECK-NEXT:    [[SHAPE:%.+]] = vector.shape_cast [[ARG]] : vector<3xi16> to vector<3x1xi16>
+// CHECK-NEXT:    [[EXT:%.+]]   = arith.extsi [[SHAPE]] : vector<3x1xi16> to vector<3x1xi32>
+// CHECK-NEXT:    [[CSTE:%.+]]  = arith.constant dense<0> : vector<3x1xi32>
+// CHECK-NEXT:    [[CMP:%.+]]   = arith.cmpi slt, [[EXT]], [[CSTE]] : vector<3x1xi32>
+// CHECK-NEXT:    [[HIGH:%.+]]  = arith.extsi [[CMP]] : vector<3x1xi1> to vector<3x1xi32>
+// CHECK-NEXT:    [[CSTZ:%.+]]  = arith.constant dense<0> : vector<3x2xi32>
+// CHECK-NEXT:    [[INS0:%.+]]  = vector.insert_strided_slice [[EXT]], [[CSTZ]] {offsets = [0, 0], strides = [1, 1]} : vector<3x1xi32> into vector<3x2xi32>
+// CHECK-NEXT:    [[INS1:%.+]]  = vector.insert_strided_slice [[HIGH]], [[INS0]] {offsets = [0, 1], strides = [1, 1]} : vector<3x1xi32> into vector<3x2xi32>
+// CHECK-NEXT:    return [[INS1]] : vector<3x2xi32>
+func.func @extsi_vector(%a : vector<3xi16>) -> vector<3xi64> {
+    %r = arith.extsi %a : vector<3xi16> to vector<3xi64>
+    return %r : vector<3xi64>
+}
+
+// CHECK-LABEL: func @extui_scalar1
+// CHECK-SAME:    ([[ARG:%.+]]: i16) -> vector<2xi32>
+// CHECK-NEXT:    [[EXT:%.+]]  = arith.extui [[ARG]] : i16 to i32
+// CHECK-NEXT:    [[VZ:%.+]]   = arith.constant dense<0> : vector<2xi32>
+// CHECK-NEXT:    [[INS0:%.+]] = vector.insert [[EXT]], [[VZ]] [0] : i32 into vector<2xi32>
+// CHECK:         return [[INS0]] : vector<2xi32>
+func.func @extui_scalar1(%a : i16) -> i64 {
+    %r = arith.extui %a : i16 to i64
+    return %r : i64
+}
+
+// CHECK-LABEL: func @extui_scalar2
+// CHECK-SAME:    ([[ARG:%.+]]: i32) -> vector<2xi32>
+// CHECK-NEXT:    [[VZ:%.+]]   = arith.constant dense<0> : vector<2xi32>
+// CHECK-NEXT:    [[INS0:%.+]] = vector.insert [[ARG]], [[VZ]] [0] : i32 into vector<2xi32>
+// CHECK:         return [[INS0]] : vector<2xi32>
+func.func @extui_scalar2(%a : i32) -> i64 {
+    %r = arith.extui %a : i32 to i64
+    return %r : i64
+}
+
+// CHECK-LABEL: func @extui_vector
+// CHECK-SAME:    ([[ARG:%.+]]: vector<3xi16>) -> vector<3x2xi32>
+// CHECK-NEXT:    [[SHAPE:%.+]] = vector.shape_cast [[ARG]] : vector<3xi16> to vector<3x1xi16>
+// CHECK-NEXT:    [[EXT:%.+]]   = arith.extui [[SHAPE]] : vector<3x1xi16> to vector<3x1xi32>
+// CHECK-NEXT:    [[CST:%.+]]   = arith.constant dense<0> : vector<3x2xi32>
+// CHECK-NEXT:    [[INS0:%.+]]  = vector.insert_strided_slice [[EXT]], [[CST]] {offsets = [0, 0], strides = [1, 1]} : vector<3x1xi32> into vector<3x2xi32>
+// CHECK:         return [[INS0]] : vector<3x2xi32>
+func.func @extui_vector(%a : vector<3xi16>) -> vector<3xi64> {
+    %r = arith.extui %a : vector<3xi16> to vector<3xi64>
+    return %r : vector<3xi64>
+}
+
+// CHECK-LABEL: func @trunci_scalar1
+// CHECK-SAME:    ([[ARG:%.+]]: vector<2xi32>) -> i32
+// CHECK-NEXT:    [[EXT:%.+]] = vector.extract [[ARG]][0] : vector<2xi32>
+// CHECK-NEXT:    return [[EXT]] : i32
+func.func @trunci_scalar1(%a : i64) -> i32 {
+    %b = arith.trunci %a : i64 to i32
+    return %b : i32
+}
+
+// CHECK-LABEL: func @trunci_scalar2
+// CHECK-SAME:    ([[ARG:%.+]]: vector<2xi32>) -> i16
+// CHECK-NEXT:    [[EXTR:%.+]] = vector.extract [[ARG]][0] : vector<2xi32>
+// CHECK-NEXT:    [[TRNC:%.+]] = arith.trunci [[EXTR]] : i32 to i16
+// CHECK-NEXT:    return [[TRNC]] : i16
+func.func @trunci_scalar2(%a : i64) -> i16 {
+    %b = arith.trunci %a : i64 to i16
+    return %b : i16
+}
+
+// CHECK-LABEL: func @trunci_vector
+// CHECK-SAME:    ([[ARG:%.+]]: vector<3x2xi32>) -> vector<3xi16>
+// CHECK-NEXT:    [[EXTR:%.+]]  = vector.extract_strided_slice [[ARG]] {offsets = [0, 0], sizes = [3, 1], strides = [1, 1]} : vector<3x2xi32> to vector<3x1xi32>
+// CHECK-NEXT:    [[SHAPE:%.+]] = vector.shape_cast [[EXTR]] : vector<3x1xi32> to vector<3xi32>
+// CHECK-NEXT:    [[TRNC:%.+]]  = arith.trunci [[SHAPE]] : vector<3xi32> to vector<3xi16>
+// CHECK-NEXT:    return [[TRNC]] : vector<3xi16>
+func.func @trunci_vector(%a : vector<3xi64>) -> vector<3xi16> {
+    %b = arith.trunci %a : vector<3xi64> to vector<3xi16>
+    return %b : vector<3xi16>
+}