[mlir] Add `ComplexType` conversion support for `convertScalarToDtype`

[rsuderman]

Summary

Linalg operations can include complex types in the src/target types.
This should include conversion between arith and complex types when
constructing linalg operations.

[kuhar]

mlir/lib/Dialect/Arith/Utils/Utils.cpp

178

nit: this function seems to be getting long, I wonder if a type switch and dispatching to helper functions would help here

201

nit: use free cast functions: https://mlir.llvm.org/deprecation/

also below

[rsuderman]

Factored into helper functions and performed general var name cleanup.

[jpienaar]

mlir/lib/Dialect/Arith/Utils/Utils.cpp

127

Should operand be returned if width is equal or is it expected that canonicalized or fold would have handled that?

[rsuderman]

Add operand bailout

[rsuderman]

mlir/lib/Dialect/Arith/Utils/Utils.cpp

127

Added the return operand;. It was technically unnecessary due to the parent function checking quality, but best to handle defense in depth.

[jpienaar]

mlir/lib/Dialect/Arith/Utils/CMakeLists.txt

9

This feels off.

[rsuderman]

mlir/lib/Dialect/Arith/Utils/CMakeLists.txt

9

I agree. The question is where would be the right place to move the conversion?

[kuhar]

Do you plan to add some tests, or is this covered by some existing one?

[rsuderman]

In D154740#4494857, @kuhar wrote:

Do you plan to add some tests, or is this covered by some existing one?

Happy to add some. I am less familiar with this area in the codebase so thoughts on how / where to test is appreciate.

[mravishankar]

mlir/lib/Dialect/Arith/Utils/Utils.cpp

89

Nit: I think the static methods need some comments to say what each one of these are for? Also maybe call this convertScalarToIntDtype and so on to make it clearer that these are implementations of parts of convertScalarToDtype?

[kuhar]

In D154740#4495452, @rsuderman wrote:

In D154740#4494857, @kuhar wrote:

Do you plan to add some tests, or is this covered by some existing one?

Happy to add some. I am less familiar with this area in the codebase so thoughts on how / where to test is appreciate.

I checked and the code from this patch is not exercised by any lit test. The helper is used with the other types in these:

Dialect/Linalg/IR/LinalgOps.cpp
Dialect/Linalg/Transforms/ElementwiseOpFusion.cpp
Dialect/Linalg/Transforms/ConvertConv2DToImg2Col.cpp
Dialect/SparseTensor/Transforms/CodegenUtils.cpp

So I'd think we either add some linalg/sparse tensor tests, or a new test pass.

[rsuderman]

Added testing and added complex type support for ConvertConv2DtoImg2Col

[rsuderman]

In D154740#4495836, @kuhar wrote:

In D154740#4495452, @rsuderman wrote:

In D154740#4494857, @kuhar wrote:

Do you plan to add some tests, or is this covered by some existing one?

Happy to add some. I am less familiar with this area in the codebase so thoughts on how / where to test is appreciate.

I checked and the code from this patch is not exercised by any lit test. The helper is used with the other types in these:

Dialect/Linalg/IR/LinalgOps.cpp
Dialect/Linalg/Transforms/ElementwiseOpFusion.cpp
Dialect/Linalg/Transforms/ConvertConv2DToImg2Col.cpp
Dialect/SparseTensor/Transforms/CodegenUtils.cpp

So I'd think we either add some linalg/sparse tensor tests, or a new test pass.

Added tests to the Img2Col case. This included fixing the pass to work for complex types and expanding support for mixed types.

[kuhar]

LGTM

mlir/lib/Dialect/Arith/Utils/CMakeLists.txt

 add_mlir_dialect_library(MLIRArithUtils
   Utils.cpp
 
   ADDITIONAL_HEADER_DIRS
   ${MLIR_MAIN_INCLUDE_DIR}/mlir/Dialect/Arith/Utils
 
   LINK_LIBS PUBLIC
   MLIRArithDialect
+  MLIRComplexDialect
   MLIRDialect
   MLIRIR
   )

mlir/lib/Dialect/Arith/Utils/Utils.cpp

 //===- Utils.cpp - Utilities to support the Linalg dialect ----------------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
 // This file implements utilities for the Linalg dialect.
 //
 //===----------------------------------------------------------------------===//
 
 #include "mlir/Dialect/Arith/Utils/Utils.h"
 #include "mlir/Dialect/Arith/IR/Arith.h"
+#include "mlir/Dialect/Complex/IR/Complex.h"
+#include "mlir/IR/ImplicitLocOpBuilder.h"
 #include "llvm/ADT/SmallBitVector.h"
 
 using namespace mlir;
 
 /// Matches a ConstantIndexOp.
 /// TODO: This should probably just be a general matcher that uses matchConstant
 /// and checks the operation for an index type.
 detail::op_matcher<arith::ConstantIndexOp> mlir::matchConstantIndex() {
assert(targetIntegerType && valueIntegerType &&
          "unexpected cast between types other than integers and index");
   assert(targetIntegerType.getSignedness() == valueIntegerType.getSignedness());
 
   if (targetIntegerType.getWidth() > valueIntegerType.getWidth())
     return b.create<arith::ExtSIOp>(loc, targetIntegerType, value);
   return b.create<arith::TruncIOp>(loc, targetIntegerType, value);
 }
 
-Value mlir::convertScalarToDtype(OpBuilder &b, Location loc, Value operand,
-                                 Type toType, bool isUnsignedCast) {
-  if (operand.getType() == toType)
-    return operand;
-  if (auto toIntType = dyn_cast<IntegerType>(toType)) {
-    // If operand is floating point, cast directly to the int type.
-    if (isa<FloatType>(operand.getType())) {
-      if (isUnsignedCast)
-        return b.create<arith::FPToUIOp>(loc, toType, operand);
-      return b.create<arith::FPToSIOp>(loc, toType, operand);
-    }
-    // Cast index operands directly to the int type.
-    if (operand.getType().isIndex())
-      return b.create<arith::IndexCastOp>(loc, toType, operand);
-    if (auto fromIntType = dyn_cast<IntegerType>(operand.getType())) {
-      // Either extend or truncate.
-      if (toIntType.getWidth() > fromIntType.getWidth()) {
-        if (isUnsignedCast)
-          return b.create<arith::ExtUIOp>(loc, toType, operand);
-        return b.create<arith::ExtSIOp>(loc, toType, operand);
-      }
-      if (toIntType.getWidth() < fromIntType.getWidth())
-        return b.create<arith::TruncIOp>(loc, toType, operand);
-    }
-  } else if (auto toFloatType = dyn_cast<FloatType>(toType)) {
-    // If operand is integer, cast directly to the float type.
-    // Note that it is unclear how to cast from BF16<->FP16.
-    if (isa<IntegerType>(operand.getType())) {
-      if (isUnsignedCast)
-        return b.create<arith::UIToFPOp>(loc, toFloatType, operand);
-      return b.create<arith::SIToFPOp>(loc, toFloatType, operand);
-    }
-    if (auto fromFloatType = dyn_cast<FloatType>(operand.getType())) {
-      if (toFloatType.getWidth() > fromFloatType.getWidth())
-        return b.create<arith::ExtFOp>(loc, toFloatType, operand);
-      if (toFloatType.getWidth() < fromFloatType.getWidth())
-        return b.create<arith::TruncFOp>(loc, toFloatType, operand);
-    }
+static Value convertToTargetInt(ImplicitLocOpBuilder &b, Value operand,
+                                IntegerType toType, bool isUnsigned) {
+    if (isUnsigned)
+      return b.create<arith::FPToUIOp>(toType, operand);
+    return b.create<arith::FPToSIOp>(toType, operand);
+    return b.create<arith::IndexCastOp>(toType, operand);
+    if (toType.getWidth() > fromIntType.getWidth()) {
+      if (isUnsigned)
+        return b.create<arith::ExtUIOp>(toType, operand);
+      return b.create<arith::ExtSIOp>(toType, operand);
+    if (toType.getWidth() < fromIntType.getWidth())
+      return b.create<arith::TruncIOp>(toType, operand);
+    return operand;
+  }
+
+  return {};
+
+static Value convertToTargetFp(ImplicitLocOpBuilder &b, Value operand,
+                               FloatType toType, bool isUnsigned) {
+    if (isUnsigned)
+      return b.create<arith::UIToFPOp>(toType, operand);
+    return b.create<arith::SIToFPOp>(toType, operand);
+  if (auto fromFpTy = dyn_cast<FloatType>(operand.getType())) {
+    if (toType.getWidth() > fromFpTy.getWidth())
+      return b.create<arith::ExtFOp>(toType, operand);
+    if (toType.getWidth() < fromFpTy.getWidth())
+      return b.create<arith::TruncFOp>(toType, operand);
+    return operand;
+  }
+
+  return {};
+}
+
+static Value convertToTargetComplex(ImplicitLocOpBuilder &b, Value operand,
+                                    ComplexType targetType, bool isUnsigned) {
+  if (auto fromComplexType = dyn_cast<ComplexType>(operand.getType())) {
+    if (isa<FloatType>(targetType.getElementType()) &&
+        isa<FloatType>(fromComplexType.getElementType())) {
+      Value real = b.create<complex::ReOp>(operand);
+      Value imag = b.create<complex::ImOp>(operand);
+      Type targetETy = targetType.getElementType();
+      if (targetType.getElementType().getIntOrFloatBitWidth() <
+          fromComplexType.getElementType().getIntOrFloatBitWidth()) {
+        real = b.create<arith::TruncFOp>(targetETy, real);
+        imag = b.create<arith::TruncFOp>(targetETy, imag);
+      } else {
+        real = b.create<arith::ExtFOp>(targetETy, real);
+        imag = b.create<arith::ExtFOp>(targetETy, imag);
+      }
+      return b.create<complex::CreateOp>(targetType, real, imag);
+    }
+
+  if (auto fromFpType = dyn_cast<FloatType>(operand.getType())) {
+    FloatType toFpTy = cast<FloatType>(targetType.getElementType());
+    auto toBitwidth = toFpTy.getIntOrFloatBitWidth();
+    Value from = operand;
+    if (from.getType().getIntOrFloatBitWidth() < toBitwidth) {
+      from = b.create<arith::ExtFOp>(toFpTy, from);
+    }
+    if (from.getType().getIntOrFloatBitWidth() > toBitwidth) {
+      from = b.create<arith::TruncFOp>(toFpTy, from);
+    }
+    Value zero = b.create<mlir::arith::ConstantFloatOp>(
+        mlir::APFloat(toFpTy.getFloatSemantics(), 0), toFpTy);
+    return b.create<complex::CreateOp>(targetType, from, zero);
+  }
+
+  if (auto fromIntType = dyn_cast<IntegerType>(operand.getType())) {
+    FloatType toFpTy = cast<FloatType>(targetType.getElementType());
+    Value from = operand;
+    if (isUnsigned) {
+      from = b.create<arith::UIToFPOp>(toFpTy, from);
+    } else {
+      from = b.create<arith::SIToFPOp>(toFpTy, from);
+    }
+    Value zero = b.create<mlir::arith::ConstantFloatOp>(
+        mlir::APFloat(toFpTy.getFloatSemantics(), 0), toFpTy);
+    return b.create<complex::CreateOp>(targetType, from, zero);
+  }
+
+  return {};
+}
+
+Value mlir::convertScalarToDtype(OpBuilder &b, Location loc, Value operand,
+                                 Type toType, bool isUnsignedCast) {
+  if (operand.getType() == toType)
+    return operand;
+  ImplicitLocOpBuilder ib(loc, b);
+  Value result;
+  if (auto intTy = dyn_cast<IntegerType>(toType)) {
+    result = convertToTargetInt(ib, operand, intTy, isUnsignedCast);
+  } else if (auto floatTy = dyn_cast<FloatType>(toType)) {
+    result = convertToTargetFp(ib, operand, floatTy, isUnsignedCast);
+  } else if (auto complexTy = dyn_cast<ComplexType>(toType)) {
+    result = convertToTargetComplex(ib, operand, complexTy, isUnsignedCast);
   }
+
+  if (result)
+    return result;
+
   emitWarning(loc) << "could not cast operand of type " << operand.getType()
                    << " to " << toType;
   return operand;
 }
 
 SmallVector<Value>
 mlir::getValueOrCreateConstantIndexOp(OpBuilder &b, Location loc,
                                       ArrayRef<OpFoldResult> valueOrAttrVec) {