Skip to content

[mlir][vector] Improve makeArithReduction expansion #75846

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

Merged
merged 1 commit into from
Dec 18, 2023
Merged

[mlir][vector] Improve makeArithReduction expansion #75846

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
6 changes: 4 additions & 2 deletions mlir/include/mlir/Dialect/Vector/IR/VectorOps.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -123,10 +123,12 @@ bool isDisjointTransferSet(VectorTransferOpInterface transferA,
VectorTransferOpInterface transferB,
bool testDynamicValueUsingBounds = false);

/// Return the result value of reducing two scalar/vector values with the
/// Returns the result value of reducing two scalar/vector values with the
/// corresponding arith operation.
Value makeArithReduction(OpBuilder &b, Location loc, CombiningKind kind,
Value v1, Value acc, Value mask = Value());
Value v1, Value acc,
arith::FastMathFlagsAttr fastmath = nullptr,
Value mask = nullptr);

/// Returns true if `attr` has "parallel" iterator type semantics.
inline bool isParallelIterator(Attribute attr) {
Expand Down
25 changes: 18 additions & 7 deletions mlir/lib/Dialect/Vector/IR/VectorOps.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -507,8 +507,9 @@ struct ElideUnitDimsInMultiDimReduction
zeroIdx);
}

Value result = vector::makeArithReduction(
rewriter, loc, reductionOp.getKind(), acc, cast, mask);
Value result =
vector::makeArithReduction(rewriter, loc, reductionOp.getKind(), acc,
cast, /*fastmath=*/nullptr, mask);
rewriter.replaceOp(rootOp, result);
return success();
}
Expand Down Expand Up @@ -650,7 +651,8 @@ struct ElideSingleElementReduction : public OpRewritePattern<ReductionOp> {

if (Value acc = reductionOp.getAcc())
result = vector::makeArithReduction(rewriter, loc, reductionOp.getKind(),
result, acc, mask);
result, acc,
reductionOp.getFastmathAttr(), mask);

rewriter.replaceOp(rootOp, result);
return success();
Expand Down Expand Up @@ -6212,6 +6214,7 @@ bool WarpExecuteOnLane0Op::areTypesCompatible(Type lhs, Type rhs) {

Value mlir::vector::makeArithReduction(OpBuilder &b, Location loc,
CombiningKind kind, Value v1, Value acc,
arith::FastMathFlagsAttr fastmath,
Value mask) {
Type t1 = getElementTypeOrSelf(v1.getType());
Type tAcc = getElementTypeOrSelf(acc.getType());
Expand All @@ -6222,7 +6225,7 @@ Value mlir::vector::makeArithReduction(OpBuilder &b, Location loc,
if (t1.isIntOrIndex() && tAcc.isIntOrIndex())
result = b.createOrFold<arith::AddIOp>(loc, v1, acc);
else if (llvm::isa<FloatType>(t1) && llvm::isa<FloatType>(tAcc))
result = b.createOrFold<arith::AddFOp>(loc, v1, acc);
result = b.createOrFold<arith::AddFOp>(loc, v1, acc, fastmath);
else
llvm_unreachable("invalid value types for ADD reduction");
break;
Expand All @@ -6231,16 +6234,24 @@ Value mlir::vector::makeArithReduction(OpBuilder &b, Location loc,
result = b.createOrFold<arith::AndIOp>(loc, v1, acc);
break;
case CombiningKind::MAXF:
assert(llvm::isa<FloatType>(t1) && llvm::isa<FloatType>(tAcc) &&
"expected float values");
result = b.createOrFold<arith::MaxNumFOp>(loc, v1, acc, fastmath);
break;
case CombiningKind::MAXIMUMF:
assert(llvm::isa<FloatType>(t1) && llvm::isa<FloatType>(tAcc) &&
"expected float values");
result = b.createOrFold<arith::MaximumFOp>(loc, v1, acc);
result = b.createOrFold<arith::MaximumFOp>(loc, v1, acc, fastmath);
break;
case CombiningKind::MINF:
assert(llvm::isa<FloatType>(t1) && llvm::isa<FloatType>(tAcc) &&
"expected float values");
result = b.createOrFold<arith::MinNumFOp>(loc, v1, acc, fastmath);
break;
case CombiningKind::MINIMUMF:
assert(llvm::isa<FloatType>(t1) && llvm::isa<FloatType>(tAcc) &&
"expected float values");
result = b.createOrFold<arith::MinimumFOp>(loc, v1, acc);
result = b.createOrFold<arith::MinimumFOp>(loc, v1, acc, fastmath);
break;
case CombiningKind::MAXSI:
assert(t1.isIntOrIndex() && tAcc.isIntOrIndex() && "expected int values");
Expand All @@ -6262,7 +6273,7 @@ Value mlir::vector::makeArithReduction(OpBuilder &b, Location loc,
if (t1.isIntOrIndex() && tAcc.isIntOrIndex())
result = b.createOrFold<arith::MulIOp>(loc, v1, acc);
else if (llvm::isa<FloatType>(t1) && llvm::isa<FloatType>(tAcc))
result = b.createOrFold<arith::MulFOp>(loc, v1, acc);
result = b.createOrFold<arith::MulFOp>(loc, v1, acc, fastmath);
else
llvm_unreachable("invalid value types for MUL reduction");
break;
Expand Down
3 changes: 2 additions & 1 deletion mlir/lib/Dialect/Vector/Transforms/LowerVectorContract.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -167,7 +167,8 @@ createContractArithOp(Location loc, Value x, Value y, Value acc,
if (!acc)
return std::optional<Value>(mul);

return makeArithReduction(rewriter, loc, kind, mul, acc, mask);
return makeArithReduction(rewriter, loc, kind, mul, acc,
/*fastmath=*/nullptr, mask);
}

/// Return the positions of the reductions in the given map.
Expand Down
4 changes: 2 additions & 2 deletions mlir/test/Conversion/VectorToLLVM/vector-to-llvm.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -450,7 +450,7 @@ func.func @masked_float_max_outerprod(%arg0: vector<2xf32>, %arg1: f32, %arg2: v
// CHECK-LABEL: func.func @masked_float_max_outerprod(
// CHECK-SAME: %[[VAL_0:.*]]: vector<2xf32>, %[[VAL_1:.*]]: f32, %[[VAL_2:.*]]: vector<2xf32>, %[[VAL_3:.*]]: vector<2xi1>) -> vector<2xf32> {
// CHECK: %[[VAL_8:.*]] = arith.mulf %[[VAL_0]], %{{.*}} : vector<2xf32>
// CHECK: %[[VAL_9:.*]] = arith.maximumf %[[VAL_8]], %[[VAL_2]] : vector<2xf32>
// CHECK: %[[VAL_9:.*]] = arith.maxnumf %[[VAL_8]], %[[VAL_2]] : vector<2xf32>
// CHECK: %[[VAL_10:.*]] = arith.select %[[VAL_3]], %[[VAL_9]], %[[VAL_2]] : vector<2xi1>, vector<2xf32>

// -----
Expand All @@ -463,7 +463,7 @@ func.func @masked_float_min_outerprod(%arg0: vector<2xf32>, %arg1: f32, %arg2: v
// CHECK-LABEL: func.func @masked_float_min_outerprod(
// CHECK-SAME: %[[VAL_0:.*]]: vector<2xf32>, %[[VAL_1:.*]]: f32, %[[VAL_2:.*]]: vector<2xf32>, %[[VAL_3:.*]]: vector<2xi1>) -> vector<2xf32> {
// CHECK: %[[VAL_8:.*]] = arith.mulf %[[VAL_0]], %{{.*}} : vector<2xf32>
// CHECK: %[[VAL_9:.*]] = arith.minimumf %[[VAL_8]], %[[VAL_2]] : vector<2xf32>
// CHECK: %[[VAL_9:.*]] = arith.minnumf %[[VAL_8]], %[[VAL_2]] : vector<2xf32>
// CHECK: %[[VAL_10:.*]] = arith.select %[[VAL_3]], %[[VAL_9]], %[[VAL_2]] : vector<2xi1>, vector<2xf32>

// -----
Expand Down
12 changes: 12 additions & 0 deletions mlir/test/Dialect/Vector/canonicalize.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2172,6 +2172,18 @@ func.func @reduce_one_element_vector_addf(%a : vector<1xf32>, %b: f32) -> f32 {

// -----

// CHECK-LABEL: func @reduce_one_element_vector_addf_fastmath
// CHECK-SAME: (%[[V:.+]]: vector<1xf32>, %[[B:.+]]: f32)
// CHECK: %[[A:.+]] = vector.extract %[[V]][0] : f32 from vector<1xf32>
// CHECK: %[[S:.+]] = arith.addf %[[A]], %arg1 fastmath<nnan,ninf> : f32
// CHECK: return %[[S]]
func.func @reduce_one_element_vector_addf_fastmath(%a : vector<1xf32>, %b: f32) -> f32 {
%s = vector.reduction <add>, %a, %b fastmath<nnan,ninf> : vector<1xf32> into f32
return %s : f32
}

// -----

// CHECK-LABEL: func @masked_reduce_one_element_vector_addf
// CHECK-SAME: %[[VAL_0:.*]]: vector<1xf32>, %[[VAL_1:.*]]: f32,
// CHECK-SAME: %[[VAL_2:.*]]: vector<1xi1>)
Expand Down
16 changes: 8 additions & 8 deletions mlir/test/Dialect/Vector/vector-multi-reduction-outer-lowering.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,13 +27,13 @@ func.func @vector_multi_reduction_min(%arg0: vector<2x4xf32>, %acc: vector<2xf32
// CHECK-SAME: %[[INPUT:.+]]: vector<2x4xf32>, %[[ACC:.*]]: vector<2xf32>
// CHECK: %[[TRANSPOSED:.+]] = vector.transpose %[[INPUT]], [1, 0] : vector<2x4xf32> to vector<4x2xf32>
// CHECK: %[[V0:.+]] = vector.extract %[[TRANSPOSED]][0] : vector<2xf32> from vector<4x2xf32>
// CHECK: %[[RV0:.+]] = arith.minimumf %[[V0]], %[[ACC]] : vector<2xf32>
// CHECK: %[[RV0:.+]] = arith.minnumf %[[V0]], %[[ACC]] : vector<2xf32>
// CHECK: %[[V1:.+]] = vector.extract %[[TRANSPOSED]][1] : vector<2xf32> from vector<4x2xf32>
// CHECK: %[[RV01:.+]] = arith.minimumf %[[V1]], %[[RV0]] : vector<2xf32>
// CHECK: %[[RV01:.+]] = arith.minnumf %[[V1]], %[[RV0]] : vector<2xf32>
// CHECK: %[[V2:.+]] = vector.extract %[[TRANSPOSED]][2] : vector<2xf32> from vector<4x2xf32>
// CHECK: %[[RV012:.+]] = arith.minimumf %[[V2]], %[[RV01]] : vector<2xf32>
// CHECK: %[[RV012:.+]] = arith.minnumf %[[V2]], %[[RV01]] : vector<2xf32>
// CHECK: %[[V3:.+]] = vector.extract %[[TRANSPOSED]][3] : vector<2xf32> from vector<4x2xf32>
// CHECK: %[[RESULT_VEC:.+]] = arith.minimumf %[[V3]], %[[RV012]] : vector<2xf32>
// CHECK: %[[RESULT_VEC:.+]] = arith.minnumf %[[V3]], %[[RV012]] : vector<2xf32>
Comment on lines +30 to +36
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I was confused because the RFC says that we should map minf/maxf to minimumf/minimumf.

https://discourse.llvm.org/t/rfc-fix-floating-point-max-and-min-operations-in-mlir/72671#proposed-solution-2

However, this is not the case in vector dialect. The execution introduces two combiner kinds minimumf/minimumf and keeps minf/maxf in https://reviews.llvm.org/D158618

Not sure who is working on it, but the current mapping is

<minimumf> --> arith.minimumf
<maximumf> --> arith.maximumf
<maxf> --> arith.maxnumf
<minf> --> arith.minnumf

In this context, the change looks okay to me.

// CHECK: return %[[RESULT_VEC]] : vector<2xf32>

func.func @vector_multi_reduction_max(%arg0: vector<2x4xf32>, %acc: vector<2xf32>) -> vector<2xf32> {
Expand All @@ -45,13 +45,13 @@ func.func @vector_multi_reduction_max(%arg0: vector<2x4xf32>, %acc: vector<2xf32
// CHECK-SAME: %[[INPUT:.+]]: vector<2x4xf32>, %[[ACC:.*]]: vector<2xf32>
// CHECK: %[[TRANSPOSED:.+]] = vector.transpose %[[INPUT]], [1, 0] : vector<2x4xf32> to vector<4x2xf32>
// CHECK: %[[V0:.+]] = vector.extract %[[TRANSPOSED]][0] : vector<2xf32> from vector<4x2xf32>
// CHECK: %[[RV0:.+]] = arith.maximumf %[[V0]], %[[ACC]] : vector<2xf32>
// CHECK: %[[RV0:.+]] = arith.maxnumf %[[V0]], %[[ACC]] : vector<2xf32>
// CHECK: %[[V1:.+]] = vector.extract %[[TRANSPOSED]][1] : vector<2xf32> from vector<4x2xf32>
// CHECK: %[[RV01:.+]] = arith.maximumf %[[V1]], %[[RV0]] : vector<2xf32>
// CHECK: %[[RV01:.+]] = arith.maxnumf %[[V1]], %[[RV0]] : vector<2xf32>
// CHECK: %[[V2:.+]] = vector.extract %[[TRANSPOSED]][2] : vector<2xf32> from vector<4x2xf32>
// CHECK: %[[RV012:.+]] = arith.maximumf %[[V2]], %[[RV01]] : vector<2xf32>
// CHECK: %[[RV012:.+]] = arith.maxnumf %[[V2]], %[[RV01]] : vector<2xf32>
// CHECK: %[[V3:.+]] = vector.extract %[[TRANSPOSED]][3] : vector<2xf32> from vector<4x2xf32>
// CHECK: %[[RESULT_VEC:.+]] = arith.maximumf %[[V3]], %[[RV012]] : vector<2xf32>
// CHECK: %[[RESULT_VEC:.+]] = arith.maxnumf %[[V3]], %[[RV012]] : vector<2xf32>
// CHECK: return %[[RESULT_VEC]] : vector<2xf32>

func.func @vector_multi_reduction_and(%arg0: vector<2x4xi32>, %acc: vector<2xi32>) -> vector<2xi32> {
Expand Down