Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[compiler] fold mhlo.reduce with splat input #367

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

[compiler] fold mhlo.reduce with splat input #367

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
120 changes: 120 additions & 0 deletions compiler/lib/Dialect/mhlo/Transforms/CanonicalizeExt.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2182,6 +2182,121 @@ struct CanonicalizeBroadcastToBroadcastInDim
}
};

template <typename Op, typename ElementType = Type, typename ValType,
typename FuncType>
static Attribute ReduceConstFolder(mhlo::ReduceOp *op,
ArrayRef<SplatElementsAttr> attrs,
ValType reduceCnt) {
if (!attrs[0] || !attrs[1])
return {};
auto splatInput = attrs[0];
ShapedType type = cast<ShapedType>(op->getResults()[0].getType());
Type etype = type.getElementType();
auto signedInput = addSign(splatInput.getSplatValue<ValType>(), etype);
auto signedReduceCnt = addSign(reduceCnt, etype);
FailureOr<decltype(signedInput)> result;
if (std::is_same_v<Op, mhlo::AddOp>) {
result = FailureOr<decltype(signedInput)>(
std::multiplies<FuncType>()(signedInput, signedReduceCnt));
} else if (std::is_same_v<Op, mhlo::MulOp>) {
result = FailureOr<decltype(signedInput)>(
Pow<FuncType>()(signedInput, signedReduceCnt));
} else if (std::is_same_v<Op, mhlo::MaxOp> ||
std::is_same_v<Op, mhlo::MinOp>) {
result = FailureOr<decltype(signedInput)>(signedInput);
} else {
return {};
}
return succeeded(result) ? SplatElementsAttr::get(type, *result)
: Attribute();
}

template <typename RegionOp>
struct FoldReduceOp : public OpRewritePattern<mhlo::ReduceOp> {
using OpRewritePattern<mhlo::ReduceOp>::OpRewritePattern;
LogicalResult matchAndRewrite(mhlo::ReduceOp op,
PatternRewriter &rewriter) const override {
if (isRegularReduceOp<RegionOp>(op)) {
auto input = op.getInputs()[0].getDefiningOp<mhlo::ConstantOp>();
auto initValue = op.getInitValues()[0].getDefiningOp<mhlo::ConstantOp>();
if (!input || !initValue) {
return failure();
}
// Only covers the case of both attrs being splats
SplatElementsAttr splatInput =
dyn_cast<SplatElementsAttr>(input.getValue());
SplatElementsAttr splaInitValue =
dyn_cast<SplatElementsAttr>(initValue.getValue());
auto type = cast<ShapedType>(op.getResults()[0].getType());
if (!splatInput || !splaInitValue || !type || !type.hasStaticShape()) {
return failure();
}
auto inputShape = cast<ShapedType>(splatInput.getType()).getShape();
auto reduceDims =
llvm::to_vector(op.getDimensions().getValues<int64_t>());
if (!reduceDims.size()) {
return failure();
}
int64_t reduceCntInt = 1;
for (const auto &dim : reduceDims) {
reduceCntInt *= inputShape[dim];
}
Attribute result;
if (isa<FloatType>(type.getElementType())) {
APFloat reduceCnt(static_cast<float>(reduceCntInt));
bool loses_info;
auto status = reduceCnt.convert(
dyn_cast<FloatType>(type.getElementType()).getFloatSemantics(),
APFloat::rmNearestTiesToEven, &loses_info);
if ((status & (~APFloat::opInexact)) != APFloat::opOK) {
op->emitWarning() << "Could not convert reduceCnt to target fp "
"type: opStatus = "
<< static_cast<int>(status);
return failure();
}
result = ReduceConstFolder<RegionOp, FloatType, APFloat, APFloat>(
&op, ArrayRef<SplatElementsAttr>{splatInput, splaInitValue},
reduceCnt);
} else if (isa<IntegerType>(type.getElementType())) {
APInt reduceCnt(splatInput.getSplatValue<APInt>().getBitWidth(),
static_cast<uint64_t>(reduceCntInt));
result = ReduceConstFolder<RegionOp, IntegerType, APInt, APSInt>(
&op, ArrayRef<SplatElementsAttr>{splatInput, splaInitValue},
reduceCnt);
}
if (!result) {
return failure();
}
rewriter.replaceOpWithNewOp<mhlo::ConstantOp>(op, result);
return success();
}
return failure();
}
};

struct DotGeneralZero : public OpRewritePattern<mhlo::DotGeneralOp> {
using OpRewritePattern<mhlo::DotGeneralOp>::OpRewritePattern;
LogicalResult matchAndRewrite(mhlo::DotGeneralOp op,
PatternRewriter &rewriter) const override {
auto lhs = op.getLhs().getDefiningOp<mhlo::ConstantOp>();
auto rhs = op.getRhs().getDefiningOp<mhlo::ConstantOp>();
auto type = cast<ShapedType>(op.getType());
if (lhs && isZeroAttribute(lhs.getValue())) {
auto resizeSplat =
cast<SplatElementsAttr>(lhs.getValue()).resizeSplat(type);
rewriter.replaceOpWithNewOp<mhlo::ConstantOp>(op, resizeSplat);
return success();
}
if (rhs && isZeroAttribute(rhs.getValue())) {
auto resizeSplat =
cast<SplatElementsAttr>(rhs.getValue()).resizeSplat(type);
rewriter.replaceOpWithNewOp<mhlo::ConstantOp>(op, resizeSplat);
return success();
}
return failure();
}
};

} // namespace

void mlir::mhlo::populateFoldMultiplyZeroPattern(RewritePatternSet &patterns) {
Expand All @@ -2201,6 +2316,11 @@ void mlir::mhlo::populateFoldLargeBinaryOpPatterns(
patterns.add<FoldLargeBinaryOp<mhlo::PowOp, Pow>>(ctx);
patterns.add<FoldLargeCompareOp>(ctx);
patterns.add<FoldClampOp>(ctx);
patterns.add<FoldReduceOp<mhlo::AddOp>>(ctx);
patterns.add<FoldReduceOp<mhlo::MulOp>>(ctx);
patterns.add<FoldReduceOp<mhlo::MaxOp>>(ctx);
patterns.add<FoldReduceOp<mhlo::MinOp>>(ctx);
patterns.add<DotGeneralZero>(ctx);
}

void mlir::mhlo::populateConvertOpPattern(RewritePatternSet &patterns,
Expand Down
71 changes: 71 additions & 0 deletions compiler/test/Transforms/CanonicalizeExt/reduce_const.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,71 @@
// RUN: byteir-opt %s -canonicalize-ext | FileCheck %s

func.func private @fold_reduce_add_f() -> tensor<16xf32> {
%0 = mhlo.constant dense<0.000000e+00> : tensor<f32>
%1 = mhlo.constant dense<5.000000e+00> : tensor<1024x16x512xf32>
%2 = mhlo.reduce(%1 init: %0) applies mhlo.add across dimensions = [0, 2] : (tensor<1024x16x512xf32>, tensor<f32>) -> tensor<16xf32>
return %2 : tensor<16xf32>
}
// CHECK-LABEL: fold_reduce_add_f
// CHECK: mhlo.constant dense<2.621440e+06>
// CHECK-NOT: mhlo.reduce

func.func private @fold_reduce_add_d() -> tensor<16xf64> {
%0 = mhlo.constant dense<0.000000e+00> : tensor<f64>
%1 = mhlo.constant dense<5.000000e+00> : tensor<1024x16x512xf64>
%2 = mhlo.reduce(%1 init: %0) applies mhlo.add across dimensions = [0, 2] : (tensor<1024x16x512xf64>, tensor<f64>) -> tensor<16xf64>
return %2 : tensor<16xf64>
}
// CHECK-LABEL: fold_reduce_add_d
// CHECK: mhlo.constant dense<2.621440e+06>
// CHECK-NOT: mhlo.reduce

func.func private @fold_reduce_add_i() -> tensor<16xi32> {
%0 = mhlo.constant dense<0> : tensor<i32>
%1 = mhlo.constant dense<5> : tensor<16x1024xi32>
%2 = mhlo.reduce(%1 init: %0) applies mhlo.add across dimensions = [1] : (tensor<16x1024xi32>, tensor<i32>) -> tensor<16xi32>
return %2 : tensor<16xi32>
}
// CHECK-LABEL: fold_reduce_add_i
// CHECK: mhlo.constant dense<5120>
// CHECK-NOT: mhlo.reduce

func.func private @fold_reduce_mul_f() -> tensor<16xf32> {
%0 = mhlo.constant dense<1.000000e+00> : tensor<f32>
%1 = mhlo.constant dense<2.000000e+00> : tensor<16x2x4xf32>
%2 = mhlo.reduce(%1 init: %0) applies mhlo.multiply across dimensions = [1, 2] : (tensor<16x2x4xf32>, tensor<f32>) -> tensor<16xf32>
return %2 : tensor<16xf32>
}
// CHECK-LABEL: fold_reduce_mul_f
// CHECK: mhlo.constant dense<2.560000e+02>
// CHECK-NOT: mhlo.reduce

func.func private @fold_reduce_mul_i() -> tensor<16xi32> {
%0 = mhlo.constant dense<1> : tensor<i32>
%1 = mhlo.constant dense<2> : tensor<16x16xi32>
%2 = mhlo.reduce(%1 init: %0) applies mhlo.multiply across dimensions = [0] : (tensor<16x16xi32>, tensor<i32>) -> tensor<16xi32>
return %2 : tensor<16xi32>
}
// CHECK-LABEL: fold_reduce_mul_i
// CHECK: mhlo.constant dense<65536>
// CHECK-NOT: mhlo.reduce

func.func private @fold_reduce_min_f() -> tensor<16xf32> {
%0 = mhlo.constant dense<0x7F800000> : tensor<f32>
%1 = mhlo.constant dense<5.000000e+00> : tensor<1024x16xf32>
%2 = mhlo.reduce(%1 init: %0) applies mhlo.minimum across dimensions = [0] : (tensor<1024x16xf32>, tensor<f32>) -> tensor<16xf32>
return %2 : tensor<16xf32>
}
// CHECK-LABEL: fold_reduce_min_f
// CHECK: mhlo.constant dense<5.000000e+00>
// CHECK-NOT: mhlo.reduce

func.func private @fold_reduce_max_i() -> tensor<16xi32> {
%0 = mhlo.constant dense<-2147483648> : tensor<i32>
%1 = mhlo.constant dense<5> : tensor<1024x512x16xi32>
%2 = mhlo.reduce(%1 init: %0) applies mhlo.maximum across dimensions = [0, 1] : (tensor<1024x512x16xi32>, tensor<i32>) -> tensor<16xi32>
return %2 : tensor<16xi32>
}
// CHECK-LABEL: fold_reduce_max_i
// CHECK: mhlo.constant dense<5>
// CHECK-NOT: mhlo.reduce
Loading