[mlir][tosa] Add a pass to narrow i64 to i32 (#165581)

This pass aims to narrow i64 types on TOSA operations to i32. It can be
useful for legalizations from various frameworks. It comes with the
following options:
- "aggressive-rewrite" - This option is typically able to narrow more
values, but may impact numerical behaviour if not used carefully.
- "convert-function-boundaries" - If enabled, parameters/ results
to/from a function may be narrowed. Otherwise, casts are inserted to
preserve the I/O of the function.

Currently the non aggressive mode is very limited, targeting an argmax
-> cast sequence that has been observed during legalization as well as
some data layout operations that can always narrow. Support for more
operations will be added in the future.

Co-authored-by: Vitalii Shutov <[email protected]>
Co-authored-by: Shubham <[email protected]>
Co-authored-by: Declan Flavin <[email protected]>

Signed-off-by: Luke Hutton <[email protected]>
Co-authored-by: Vitalii Shutov <[email protected]>
Co-authored-by: Shubham <[email protected]>
Co-authored-by: Declan Flavin <[email protected]>

Author: 4 people

mlir/include/mlir/Dialect/Tosa/Transforms/Passes.td

@@ -166,4 +166,27 @@ def TosaAttachTarget : Pass<"tosa-attach-target", "ModuleOp"> {
   ];
 }
 
+def TosaNarrowI64ToI32Pass : Pass<"tosa-narrow-i64-to-i32", "func::FuncOp"> {
+  let summary = "Narrow I64 TOSA operations to I32";
+  let description = [{
+    This pass narrows TOSA operations with 64-bit integer tensor types to
+    32-bit integer tensor types. This can be useful for backends that do not
+    support the EXT-INT64 extension of TOSA.
+  }];
+
+  let options = [
+    Option<"aggressiveRewrite", "aggressive-rewrite", "bool", "false",
+      "If enabled, all TOSA operations are rewritten, regardless or whether the narrowing"
+      "is safe. This option may lead to data loss if not used carefully.">,
+    Option<"convertFunctionBoundaries", "convert-function-boundaries", "bool", "false",
+      "If enabled, the pass will convert function I/O types as well. Otherwise casts will"
+      "be inserted at the I/O boundaries.">
+  ];
+
+  let dependentDialects = [
+    "func::FuncDialect",
+    "tosa::TosaDialect",
+  ];
+}
+
 #endif // MLIR_DIALECT_TOSA_TRANSFORMS_PASSES

mlir/lib/Dialect/Tosa/Transforms/CMakeLists.txt

@@ -12,6 +12,7 @@ add_mlir_dialect_library(MLIRTosaTransforms
   TosaTypeConverters.cpp
   TosaProfileCompliance.cpp
   TosaValidation.cpp
+  TosaNarrowI64ToI32.cpp
 
   ADDITIONAL_HEADER_DIRS
   ${MLIR_MAIN_INCLUDE_DIR}/mlir/Dialect/Tosa/Transforms

mlir/lib/Dialect/Tosa/Transforms/TosaNarrowI64ToI32.cpp

@@ -0,0 +1,310 @@
+//===- TosaNarrowI64ToI32.cpp ---------------------------------------------===//
+//
+// 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 pass narrows TOSA operations with 64-bit integer tensor types to
+// 32-bit integer tensor types. This can be useful for backends that do not
+// support the EXT-INT64 extension of TOSA. The pass has two options:
+//
+// - aggressive-rewrite - If enabled, all TOSA operations are rewritten,
+//     regardless or whether the narrowing is safe. This option may lead to
+//     data loss if not used carefully.
+// - convert-function-boundaries - If enabled, the pass will convert function
+//     I/O types as well. Otherwise casts will be inserted at the I/O
+//     boundaries.
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Dialect/Tosa/Transforms/Passes.h"
+
+#include "mlir/Dialect/Func/IR/FuncOps.h"
+#include "mlir/Dialect/Func/Transforms/FuncConversions.h"
+#include "mlir/IR/Verifier.h"
+#include "mlir/Pass/Pass.h"
+
+namespace mlir {
+namespace tosa {
+#define GEN_PASS_DEF_TOSANARROWI64TOI32PASS
+#include "mlir/Dialect/Tosa/Transforms/Passes.h.inc"
+} // namespace tosa
+} // namespace mlir
+
+using namespace mlir;
+using namespace mlir::tosa;
+
+namespace {
+
+LogicalResult convertGenericOp(Operation *op, ValueRange operands,
+                               ConversionPatternRewriter &rewriter,
+                               const TypeConverter *typeConverter) {
+  // Convert types of results
+  SmallVector<Type, 4> newResults;
+  if (failed(typeConverter->convertTypes(op->getResultTypes(), newResults)))
+    return failure();
+
+  // Create a new operation state
+  OperationState state(op->getLoc(), op->getName().getStringRef(), operands,
+                       newResults, {}, op->getSuccessors());
+
+  for (const NamedAttribute &namedAttribute : op->getAttrs()) {
+    const Attribute attribute = namedAttribute.getValue();
+
+    // Convert integer attribute type
+    if (const auto intAttr = dyn_cast<IntegerAttr>(attribute)) {
+      const std::optional<Attribute> convertedAttribute =
+          typeConverter->convertTypeAttribute(intAttr.getType(), attribute);
+      state.addAttribute(namedAttribute.getName(), convertedAttribute.value());
+      continue;
+    }
+
+    if (const auto typeAttr = dyn_cast<TypeAttr>(attribute)) {
+      Type type = typeAttr.getValue();
+      const std::optional<Attribute> convertedAttribute =
+          typeConverter->convertTypeAttribute(type, attribute);
+      if (!convertedAttribute)
+        return rewriter.notifyMatchFailure(op,
+                                           "Failed to convert type attribute.");
+      state.addAttribute(namedAttribute.getName(), convertedAttribute.value());
+      continue;
+    }
+
+    if (const auto denseElementsAttr = dyn_cast<DenseElementsAttr>(attribute)) {
+      const Type type = denseElementsAttr.getType();
+      const std::optional<Attribute> convertedAttribute =
+          typeConverter->convertTypeAttribute(type, denseElementsAttr);
+      if (!convertedAttribute)
+        return rewriter.notifyMatchFailure(
+            op, "Failed to convert dense elements attribute.");
+      state.addAttribute(namedAttribute.getName(), convertedAttribute.value());
+      continue;
+    }
+
+    state.addAttribute(namedAttribute.getName(), attribute);
+  }
+
+  for (Region &region : op->getRegions()) {
+    Region *newRegion = state.addRegion();
+    rewriter.inlineRegionBefore(region, *newRegion, newRegion->begin());
+    if (failed(rewriter.convertRegionTypes(newRegion, *typeConverter)))
+      return failure();
+  }
+
+  Operation *newOp = rewriter.create(state);
+  rewriter.replaceOp(op, newOp->getResults());
+  return success();
+}
+
+// ===========================
+// Aggressive rewrite patterns
+// ===========================
+
+class ConvertGenericOp : public ConversionPattern {
+public:
+  ConvertGenericOp(TypeConverter &typeConverter, MLIRContext *context)
+      : ConversionPattern(typeConverter, MatchAnyOpTypeTag{}, 0, context) {}
+
+  LogicalResult
+  matchAndRewrite(Operation *op, ArrayRef<Value> operands,
+                  ConversionPatternRewriter &rewriter) const final {
+    if (!isa<tosa::TosaOp>(op))
+      return rewriter.notifyMatchFailure(
+          op,
+          "Support for operations other than TOSA has not been implemented.");
+
+    return convertGenericOp(op, operands, rewriter, typeConverter);
+  }
+};
+
+// ===============================
+// Bounds checked rewrite patterns
+// ===============================
+
+class ConvertArgMaxOpWithBoundsChecking
+    : public OpConversionPattern<tosa::ArgMaxOp> {
+  using OpConversionPattern::OpConversionPattern;
+
+  LogicalResult
+  matchAndRewrite(tosa::ArgMaxOp op, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const final {
+    // Output type can be narrowed based on the size of the axis dimension
+    const int32_t axis = op.getAxis();
+    const auto inputType = dyn_cast<ShapedType>(adaptor.getInput().getType());
+    if (!inputType || !inputType.isStaticDim(axis))
+      return rewriter.notifyMatchFailure(
+          op, "Requires a static axis dimension for bounds checking.");
+    const int64_t axisDim = inputType.getDimSize(axis);
+    if (axisDim >= std::numeric_limits<int32_t>::max())
+      return rewriter.notifyMatchFailure(
+          op, "Axis dimension is too large to narrow safely.");
+
+    const Type resultType = op.getOutput().getType();
+    const Type newResultType = typeConverter->convertType(resultType);
+    rewriter.replaceOpWithNewOp<tosa::ArgMaxOp>(op, newResultType,
+                                                adaptor.getInput(), axis);
+    return success();
+  }
+};
+
+class ConvertCastOpWithBoundsChecking
+    : public OpConversionPattern<tosa::CastOp> {
+  using OpConversionPattern::OpConversionPattern;
+
+  LogicalResult
+  matchAndRewrite(tosa::CastOp op, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const final {
+    const auto inputType = dyn_cast<ShapedType>(adaptor.getInput().getType());
+    const auto resultType = dyn_cast<ShapedType>(op.getResult().getType());
+    if (!inputType || !resultType)
+      return failure();
+
+    const auto elementInputIntType =
+        dyn_cast<IntegerType>(inputType.getElementType());
+    const auto elementResultIntType =
+        dyn_cast<IntegerType>(resultType.getElementType());
+    if (elementInputIntType && elementResultIntType &&
+        elementInputIntType.getWidth() > elementResultIntType.getWidth())
+      return rewriter.notifyMatchFailure(
+          op, "Narrowing cast may lead to data loss.");
+
+    rewriter.replaceOpWithNewOp<tosa::CastOp>(
+        op, typeConverter->convertType(resultType), adaptor.getInput());
+    return success();
+  }
+};
+
+template <typename OpTy>
+class ConvertTypedOp : public OpConversionPattern<OpTy> {
+  using OpConversionPattern<OpTy>::OpConversionPattern;
+
+  LogicalResult
+  matchAndRewrite(OpTy op, typename OpTy::Adaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const final {
+    return convertGenericOp(op, adaptor.getOperands(), rewriter,
+                            this->getTypeConverter());
+  }
+};
+
+struct TosaNarrowI64ToI32
+    : public tosa::impl::TosaNarrowI64ToI32PassBase<TosaNarrowI64ToI32> {
+public:
+  explicit TosaNarrowI64ToI32() = default;
+  explicit TosaNarrowI64ToI32(const TosaNarrowI64ToI32PassOptions &options)
+      : TosaNarrowI64ToI32() {
+    this->aggressiveRewrite = options.aggressiveRewrite;
+    this->convertFunctionBoundaries = options.convertFunctionBoundaries;
+  }
+
+  void runOnOperation() override {
+    MLIRContext *context = &getContext();
+
+    TypeConverter typeConverter;
+    typeConverter.addConversion([](Type type) -> Type { return type; });
+    typeConverter.addConversion([](IntegerType type) -> Type {
+      if (!type.isInteger(64))
+        return type;
+      return IntegerType::get(type.getContext(), 32);
+    });
+    typeConverter.addConversion(
+        [&typeConverter](RankedTensorType type) -> Type {
+          const Type elementType = type.getElementType();
+          if (!elementType.isInteger(64))
+            return type;
+          return RankedTensorType::get(type.getShape(),
+                                       typeConverter.convertType(elementType));
+        });
+
+    const auto materializeCast = [](OpBuilder &builder, Type resultType,
+                                    ValueRange inputs, Location loc) -> Value {
+      if (inputs.size() != 1)
+        return Value();
+      return tosa::CastOp::create(builder, loc, resultType, inputs.front());
+    };
+    typeConverter.addSourceMaterialization(materializeCast);
+    typeConverter.addTargetMaterialization(materializeCast);
+
+    typeConverter.addTypeAttributeConversion(
+        [](IntegerType type, IntegerAttr attribute) -> Attribute {
+          const APInt value = attribute.getValue().truncSSat(32);
+          return IntegerAttr::get(IntegerType::get(type.getContext(), 32),
+                                  value);
+        });
+    typeConverter.addTypeAttributeConversion(
+        [&typeConverter](ShapedType type,
+                         DenseIntElementsAttr attr) -> Attribute {
+          const ShapedType newType =
+              cast<ShapedType>(typeConverter.convertType(type));
+          const auto oldElementType = cast<IntegerType>(type.getElementType());
+          const auto newElementType =
+              cast<IntegerType>(newType.getElementType());
+          if (oldElementType.getWidth() == newElementType.getWidth())
+            return attr;
+
+          DenseElementsAttr mapped =
+              attr.mapValues(newElementType, [&](const APInt &v) {
+                return v.truncSSat(newElementType.getWidth());
+              });
+          return mapped;
+        });
+
+    ConversionTarget target(*context);
+    target.addDynamicallyLegalDialect<tosa::TosaDialect>(
+        [&typeConverter](Operation *op) {
+          return typeConverter.isLegal(op->getResultTypes()) &&
+                 typeConverter.isLegal(op->getOperandTypes());
+        });
+    if (convertFunctionBoundaries) {
+      target.addDynamicallyLegalOp<func::FuncOp>(
+          [&typeConverter](func::FuncOp op) {
+            return typeConverter.isSignatureLegal(op.getFunctionType()) &&
+                   typeConverter.isLegal(&op.getBody());
+          });
+      target.addDynamicallyLegalOp<func::ReturnOp>([](func::ReturnOp op) {
+        const FunctionType funcType =
+            op->getParentOfType<func::FuncOp>().getFunctionType();
+        return llvm::equal(op.getOperandTypes(), funcType.getResults());
+      });
+    } else {
+      target.addDynamicallyLegalOp<func::FuncOp>(
+          [](func::FuncOp op) { return true; });
+      target.addDynamicallyLegalOp<func::ReturnOp>(
+          [](func::ReturnOp op) { return true; });
+    }
+
+    RewritePatternSet patterns(context);
+    if (convertFunctionBoundaries) {
+      populateFunctionOpInterfaceTypeConversionPattern<func::FuncOp>(
+          patterns, typeConverter);
+      populateReturnOpTypeConversionPattern(patterns, typeConverter);
+    }
+    if (aggressiveRewrite) {
+      patterns.add<ConvertGenericOp>(typeConverter, context);
+    } else {
+      // Tensor
+      patterns.add<ConvertArgMaxOpWithBoundsChecking>(typeConverter, context);
+      // Data layout
+      patterns.add<ConvertTypedOp<tosa::ConcatOp>>(typeConverter, context);
+      patterns.add<ConvertTypedOp<tosa::PadOp>>(typeConverter, context);
+      patterns.add<ConvertTypedOp<tosa::ReshapeOp>>(typeConverter, context);
+      patterns.add<ConvertTypedOp<tosa::ReverseOp>>(typeConverter, context);
+      patterns.add<ConvertTypedOp<tosa::SliceOp>>(typeConverter, context);
+      patterns.add<ConvertTypedOp<tosa::TileOp>>(typeConverter, context);
+      patterns.add<ConvertTypedOp<tosa::TransposeOp>>(typeConverter, context);
+      patterns.add<ConvertTypedOp<tosa::IdentityOp>>(typeConverter, context);
+      // Type conversion
+      patterns.add<ConvertCastOpWithBoundsChecking>(typeConverter, context);
+      // Controlflow
+      patterns.add<ConvertTypedOp<tosa::IfOp>>(typeConverter, context);
+      patterns.add<ConvertTypedOp<tosa::WhileOp>>(typeConverter, context);
+    }
+
+    if (failed(
+            applyFullConversion(getOperation(), target, std::move(patterns))))
+      signalPassFailure();
+  }
+};
+
+} // namespace