Skip to content

[MLIR][Vector] Support vector reduction for xor/maxnumf/minnumf #163310

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

Open
wants to merge 3 commits into
base: main
Choose a base branch
from
Open

[MLIR][Vector] Support vector reduction for xor/maxnumf/minnumf #163310

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
6300dae
Support vector reduction for xor/maxnumf/minnumf
leewei05 Oct 14, 2025
cf39b25
Update variable names
leewei05 Oct 23, 2025
08a1706
Add minnumf maxnumf conversion test
leewei05 Oct 23, 2025
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
5 changes: 3 additions & 2 deletions mlir/lib/Dialect/Affine/Analysis/AffineAnalysis.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,9 +66,10 @@ static Value getSupportedReduction(AffineForOp forOp, unsigned pos,
.Case([](arith::MaxSIOp) { return arith::AtomicRMWKind::maxs; })
.Case([](arith::MinUIOp) { return arith::AtomicRMWKind::minu; })
.Case([](arith::MaxUIOp) { return arith::AtomicRMWKind::maxu; })
.Case([](arith::XOrIOp) { return arith::AtomicRMWKind::xori; })
.Case([](arith::MaxNumFOp) { return arith::AtomicRMWKind::maxnumf; })
.Case([](arith::MinNumFOp) { return arith::AtomicRMWKind::minnumf; })
.Default([](Operation *) -> std::optional<arith::AtomicRMWKind> {
// TODO: AtomicRMW supports other kinds of reductions this is
// currently not detecting, add those when the need arises.
return std::nullopt;
});
if (!maybeKind)
Expand Down
10 changes: 9 additions & 1 deletion mlir/lib/Dialect/Vector/IR/VectorOps.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -717,7 +717,15 @@ Value mlir::vector::getVectorReductionOp(arith::AtomicRMWKind op,
case arith::AtomicRMWKind::ori:
return vector::ReductionOp::create(builder, vector.getLoc(),
CombiningKind::OR, vector);
// TODO: Add remaining reduction operations.
case arith::AtomicRMWKind::minnumf:
return vector::ReductionOp::create(builder, vector.getLoc(),
CombiningKind::MINNUMF, vector);
case arith::AtomicRMWKind::maxnumf:
return vector::ReductionOp::create(builder, vector.getLoc(),
CombiningKind::MAXNUMF, vector);
case arith::AtomicRMWKind::xori:
return vector::ReductionOp::create(builder, vector.getLoc(),
CombiningKind::XOR, vector);
default:
(void)emitOptionalError(loc, "Reduction operation type not supported");
break;
Expand Down
100 changes: 100 additions & 0 deletions mlir/test/Dialect/Affine/SuperVectorize/vectorize_reduction.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -243,6 +243,106 @@ func.func @vecdim_reduction_ori(%in: memref<256x512xi32>, %out: memref<256xi32>)
// CHECK: affine.store %[[final_red]], %{{.*}} : memref<256xi32>
// CHECK: }

// -----

func.func @vecdim_reduction_xori(%in: memref<256x512xi32>, %out: memref<256xi32>) {
%cst = arith.constant 0 : i32
affine.for %i = 0 to 256 {
%final_red = affine.for %j = 0 to 512 iter_args(%red_iter = %cst) -> (i32) {
%ld = affine.load %in[%i, %j] : memref<256x512xi32>
%xor = arith.xori %red_iter, %ld : i32
affine.yield %xor : i32
}
affine.store %final_red, %out[%i] : memref<256xi32>
}
return
}

// CHECK-LABEL: func.func @vecdim_reduction_xori(
// CHECK-SAME: %[[ARG0:.*]]: memref<256x512xi32>,
// CHECK-SAME: %[[ARG1:.*]]: memref<256xi32>) {
// CHECK: %[[VAL_0:.*]] = arith.constant 0 : i32
// CHECK: affine.for %[[VAL_1:.*]] = 0 to 256 {
// CHECK: %[[VAL_2:.*]] = arith.constant dense<0> : vector<128xi32>
// CHECK: %[[VAL_3:.*]] = affine.for %[[VAL_4:.*]] = 0 to 512 step 128 iter_args(%[[VAL_5:.*]] = %[[VAL_2]]) -> (vector<128xi32>) {
// CHECK: %[[VAL_6:.*]] = ub.poison : i32
// CHECK: %[[VAL_7:.*]] = vector.transfer_read %[[ARG0]]{{\[}}%[[VAL_1]], %[[VAL_4]]], %[[VAL_6]] : memref<256x512xi32>, vector<128xi32>
// CHECK: %[[VAL_8:.*]] = arith.xori %[[VAL_5]], %[[VAL_7]] : vector<128xi32>
// CHECK: affine.yield %[[VAL_8]] : vector<128xi32>
// CHECK: }
// CHECK: %[[VAL_9:.*]] = vector.reduction <xor>, %[[VAL_3]] : vector<128xi32> into i32
// CHECK: affine.store %[[VAL_9]], %[[ARG1]]{{\[}}%[[VAL_1]]] : memref<256xi32>
// CHECK: }
// CHECK: return
// CHECK: }
Comment on lines 261 to 277
Copy link
Contributor

Choose a reason for hiding this comment

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

Please use meaningful FileCheck variable names. See comprehensive guidelines here: https://mlir.llvm.org/getting_started/TestingGuide/


// -----

func.func @vecdim_reduction_minnumf(%in: memref<256x512xf32>, %out: memref<256xf32>) {
%cst = arith.constant 0xFF800000 : f32
affine.for %i = 0 to 256 {
%final_red = affine.for %j = 0 to 512 iter_args(%red_iter = %cst) -> (f32) {
%ld = affine.load %in[%i, %j] : memref<256x512xf32>
%min = arith.minnumf %red_iter, %ld : f32
affine.yield %min : f32
}
affine.store %final_red, %out[%i] : memref<256xf32>
}
return
}

// CHECK-LABEL: func.func @vecdim_reduction_minnumf(
// CHECK-SAME: %[[ARG0:.*]]: memref<256x512xf32>,
// CHECK-SAME: %[[ARG1:.*]]: memref<256xf32>) {
// CHECK: %[[VAL_0:.*]] = arith.constant 0xFF800000 : f32
// CHECK: affine.for %[[VAL_1:.*]] = 0 to 256 {
// CHECK: %[[VAL_2:.*]] = arith.constant dense<0x7FC00000> : vector<128xf32>
// CHECK: %[[VAL_3:.*]] = affine.for %[[VAL_4:.*]] = 0 to 512 step 128 iter_args(%[[VAL_5:.*]] = %[[VAL_2]]) -> (vector<128xf32>) {
// CHECK: %[[VAL_6:.*]] = ub.poison : f32
// CHECK: %[[VAL_7:.*]] = vector.transfer_read %[[ARG0]]{{\[}}%[[VAL_1]], %[[VAL_4]]], %[[VAL_6]] : memref<256x512xf32>, vector<128xf32>
// CHECK: %[[VAL_8:.*]] = arith.minnumf %[[VAL_5]], %[[VAL_7]] : vector<128xf32>
// CHECK: affine.yield %[[VAL_8]] : vector<128xf32>
// CHECK: }
// CHECK: %[[VAL_9:.*]] = vector.reduction <minnumf>, %[[VAL_3]] : vector<128xf32> into f32
// CHECK: %[[VAL_10:.*]] = arith.minnumf %[[VAL_9]], %[[VAL_0]] : f32
// CHECK: affine.store %[[VAL_10]], %[[ARG1]]{{\[}}%[[VAL_1]]] : memref<256xf32>
// CHECK: }
// CHECK: return
// CHECK: }

// -----

func.func @vecdim_reduction_maxnumf(%in: memref<256x512xf32>, %out: memref<256xf32>) {
%cst = arith.constant 0xFF800000 : f32
affine.for %i = 0 to 256 {
%final_red = affine.for %j = 0 to 512 iter_args(%red_iter = %cst) -> (f32) {
%ld = affine.load %in[%i, %j] : memref<256x512xf32>
%max = arith.maxnumf %red_iter, %ld : f32
affine.yield %max : f32
}
affine.store %final_red, %out[%i] : memref<256xf32>
}
return
}

// CHECK-LABEL: func.func @vecdim_reduction_maxnumf(
// CHECK-SAME: %[[ARG0:.*]]: memref<256x512xf32>,
// CHECK-SAME: %[[ARG1:.*]]: memref<256xf32>) {
// CHECK: %[[VAL_0:.*]] = arith.constant 0xFF800000 : f32
// CHECK: affine.for %[[VAL_1:.*]] = 0 to 256 {
// CHECK: %[[VAL_2:.*]] = arith.constant dense<0xFFC00000> : vector<128xf32>
// CHECK: %[[VAL_3:.*]] = affine.for %[[VAL_4:.*]] = 0 to 512 step 128 iter_args(%[[VAL_5:.*]] = %[[VAL_2]]) -> (vector<128xf32>) {
// CHECK: %[[VAL_6:.*]] = ub.poison : f32
// CHECK: %[[VAL_7:.*]] = vector.transfer_read %[[ARG0]]{{\[}}%[[VAL_1]], %[[VAL_4]]], %[[VAL_6]] : memref<256x512xf32>, vector<128xf32>
// CHECK: %[[VAL_8:.*]] = arith.maxnumf %[[VAL_5]], %[[VAL_7]] : vector<128xf32>
// CHECK: affine.yield %[[VAL_8]] : vector<128xf32>
// CHECK: }
// CHECK: %[[VAL_9:.*]] = vector.reduction <maxnumf>, %[[VAL_3]] : vector<128xf32> into f32
// CHECK: %[[VAL_10:.*]] = arith.maxnumf %[[VAL_9]], %[[VAL_0]] : f32
// CHECK: affine.store %[[VAL_10]], %[[ARG1]]{{\[}}%[[VAL_1]]] : memref<256xf32>
// CHECK: }
// CHECK: return
// CHECK: }

// -----

Expand Down