Skip to content

[mlir][linalg] Reject unsigned pooling on non-integer element types #166070

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 4 commits into
base: main
Choose a base branch
from
+136 −38
Open

[mlir][linalg] Reject unsigned pooling on non-integer element types #166070

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
03ef5fc
[mlir][linalg] Reject unsigned pooling on non-integer element types
Men-cotton Nov 2, 2025
9e394ab
[mlir][linalg][python] Reject unsigned pooling on non-integer element…
Men-cotton Nov 3, 2025
13f41de
[mlir][linalg][test] Drop float min_unsigned roundtrip
Men-cotton Nov 15, 2025
400406e
Merge branch 'main' into users/Men-cotton/mlir/164800
Men-cotton Nov 15, 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
18 changes: 14 additions & 4 deletions mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -579,13 +579,23 @@ class RegionBuilderHelper {
return arith::MinSIOp::create(builder, arg0.getLoc(), arg0, arg1);
case BinaryFn::max_unsigned:
assert(!allComplex);
if (allFloatingPoint)
return arith::MaximumFOp::create(builder, arg0.getLoc(), arg0, arg1);
if (!allInteger || allBool) {
if (emitError) {
emitError() << "unsupported operation: unsigned max not on uint";
return nullptr;
}
llvm_unreachable("unsupported operation: unsigned max not on uint");
}
Comment on lines +582 to +588
Copy link
Contributor

@Abhishek-Varma Abhishek-Varma Nov 3, 2025

Choose a reason for hiding this comment

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

Hi @Men-cotton - thanks for taking this up.
Although this addresses the immediate issue of linalg.*_(max|min)_unsigned_* - a better solution might be to indeed fix :-

llvm-project/mlir/python/mlir/dialects/linalg/opdsl/lang/emitter.py

Lines 534 to 539 in f5885de

def _binary_max_unsigned(self, lhs: Value, rhs: Value) -> Value:
if _is_floating_point_type(lhs.type):
return arith.MaximumFOp(lhs, rhs).result
if _is_integer_type(lhs.type) or _is_index_type(lhs.type):
return arith.MaxUIOp(lhs, rhs).result
raise NotImplementedError("Unsupported 'max_unsigned' operands: {lhs}, {rhs}")

Because that'd ensure that in future some other op implementing a (max|min)_unsigned also doesn't warrant a similar fix in their verifier as the one above.

I also see something similar already done in this file for a non-Convolution op : linalg.div_unsigned :-

llvm-project/mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp

Lines 561 to 568 in 03ef5fc

case BinaryFn::div_unsigned:
if (!allInteger || allBool) {
if (emitError) {
emitError() << "unsupported operation: unsigned div not on uint";
return nullptr;
}
llvm_unreachable("unsupported operation: unsigned div not on uint");
}

Again, these are just my opinion - I'll let @banach-space take a closer look. :)

Copy link
Contributor

@banach-space banach-space Nov 10, 2025

Choose a reason for hiding this comment

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

I'm not really familiar with OPDSL, but in ideal world we should be able to write verification tests and add them in https://github.com/llvm/llvm-project/blob/main/mlir/test/Dialect/Linalg/invalid.mlir - could you try @Men-cotton ?

Copy link
Contributor Author

@Men-cotton Men-cotton Nov 21, 2025

Choose a reason for hiding this comment

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

Hi @banach-space , thank you for the suggestion.
I actually already added a verification test in mlir/test/Dialect/Linalg/named-ops-fail.mlir. If you prefer invalid.mlir or find the current test insufficient, please let me know.

return arith::MaxUIOp::create(builder, arg0.getLoc(), arg0, arg1);
case BinaryFn::min_unsigned:
assert(!allComplex);
if (allFloatingPoint)
return arith::MinimumFOp::create(builder, arg0.getLoc(), arg0, arg1);
if (!allInteger || allBool) {
if (emitError) {
emitError() << "unsupported operation: unsigned min not on uint";
return nullptr;
}
llvm_unreachable("unsupported operation: unsigned min not on uint");
}
return arith::MinUIOp::create(builder, arg0.getLoc(), arg0, arg1);
case BinaryFn::powf:
assert(allFloatingPoint);
Expand Down
18 changes: 12 additions & 6 deletions mlir/python/mlir/dialects/linalg/opdsl/lang/emitter.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -532,9 +532,9 @@ def _binary_max_signed(self, lhs: Value, rhs: Value) -> Value:
raise NotImplementedError("Unsupported 'max' operands: {lhs}, {rhs}")

def _binary_max_unsigned(self, lhs: Value, rhs: Value) -> Value:
if _is_floating_point_type(lhs.type):
return arith.MaximumFOp(lhs, rhs).result
if _is_integer_type(lhs.type) or _is_index_type(lhs.type):
if (
_is_integer_type(lhs.type) and not _is_bool_type(lhs.type)
) or _is_index_type(lhs.type):
return arith.MaxUIOp(lhs, rhs).result
raise NotImplementedError("Unsupported 'max_unsigned' operands: {lhs}, {rhs}")

Expand All @@ -546,9 +546,9 @@ def _binary_min_signed(self, lhs: Value, rhs: Value) -> Value:
raise NotImplementedError("Unsupported 'min' operands: {lhs}, {rhs}")

def _binary_min_unsigned(self, lhs: Value, rhs: Value) -> Value:
if _is_floating_point_type(lhs.type):
return arith.MinimumFOp(lhs, rhs).result
if _is_integer_type(lhs.type) or _is_index_type(lhs.type):
if (
_is_integer_type(lhs.type) and not _is_bool_type(lhs.type)
) or _is_index_type(lhs.type):
return arith.MinUIOp(lhs, rhs).result
raise NotImplementedError("Unsupported 'min_unsigned' operands: {lhs}, {rhs}")

Expand Down Expand Up @@ -634,6 +634,12 @@ def _is_index_type(t: Type) -> bool:
return IndexType.isinstance(t)


def _is_bool_type(t: Type) -> bool:
if not IntegerType.isinstance(t):
return False
return IntegerType(t).width == 1


def _get_floating_point_width(t: Type) -> int:
# TODO: Create a FloatType in the Python API and implement the switch
# there.
Expand Down
13 changes: 0 additions & 13 deletions mlir/test/Dialect/Linalg/convolution/roundtrip-convolution.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -104,16 +104,3 @@ func.func @pooling_nhwc_min_unsigned_integer(%input: tensor<?x?x?x?xi32>, %filte
// CHECK: @pooling_nhwc_min_unsigned_integer
// CHECK: linalg.pooling_nhwc_min_unsigned
// CHECK-SAME: dilations = dense<1> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>

// -----

func.func @pooling_nhwc_min_unsigned_float(%input: tensor<?x?x?x?xf32>, %filter: tensor<?x?xf32>, %output: tensor<?x?x?x?xf32>) -> tensor<?x?x?x?xf32> {
%0 = linalg.pooling_nhwc_min_unsigned
{dilations = dense<1> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>}
ins (%input, %filter: tensor<?x?x?x?xf32>, tensor<?x?xf32>)
outs (%output: tensor<?x?x?x?xf32>) -> tensor<?x?x?x?xf32>
return %0 : tensor<?x?x?x?xf32>
}
// CHECK: @pooling_nhwc_min_unsigned_float
// CHECK: linalg.pooling_nhwc_min
// CHECK-SAME: dilations = dense<1> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>
15 changes: 14 additions & 1 deletion mlir/test/Dialect/Linalg/named-ops-fail.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -80,6 +80,20 @@ func.func @divu_broadcast(%arg0: memref<8x16xi32>, %arg1: memref<4x8x16xi32>, %a

// -----

func.func @pooling_nhwc_max_unsigned_float(
%input: tensor<?x?x?x?xf32>,
%filter: tensor<?x?xf32>,
%init_val: tensor<?x?x?x?xf32>) -> tensor<?x?x?x?xf32> {
// CHECK: unsupported operation: unsigned max not on uint
linalg.pooling_nhwc_max_unsigned {dilations = dense<1> : tensor<2xi64>,
strides = dense<1> : tensor<2xi64>}
ins (%input, %filter: tensor<?x?x?x?xf32>, tensor<?x?xf32>)
outs (%init_val: tensor<?x?x?x?xf32>) -> tensor<?x?x?x?xf32>
return %init_val : tensor<?x?x?x?xf32>
}

// -----

func.func @exp_type_cast(%arg: memref<4x8x16xf16>, %out: memref<4x8x16xf32>) {
// CHECK: operand 1 ('f16') doesn't match the element type of the enclosing linalg.generic op ('f32')
linalg.exp ins(%arg : memref<4x8x16xf16>) outs(%out: memref<4x8x16xf32>)
Expand Down Expand Up @@ -349,4 +363,3 @@ func.func @select_wrong_condition_type(%arg0: memref<4x8x16xf32>, %arg1: memref<
linalg.select ins(%arg0, %arg1, %arg2 : memref<4x8x16xf32>, memref<4x8x16xf32>, memref<4x8x16xf32>) outs(%arg3: memref<4x8x16xf32>)
return
}

34 changes: 34 additions & 0 deletions mlir/test/Dialect/Linalg/named-ops.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -705,6 +705,23 @@ func.func @pooling_nhwc_max_tensor(%input: tensor<1x4x4x1xf32>) -> tensor<1x2x2x
return %res : tensor<1x2x2x1xf32>
}

// -----

// CHECK-LABEL: func @pooling_nhwc_max_unsigned_tensor
// CHECK: %{{.+}} = linalg.pooling_nhwc_max_unsigned
// CHECK-SAME: ins(%{{.+}}, %{{.+}} : tensor<1x4x4x1xi32>, tensor<3x3xi32>)
// CHECK-SAME: outs(%{{.+}} : tensor<1x2x2x1xi32>) -> tensor<1x2x2x1xi32>
func.func @pooling_nhwc_max_unsigned_tensor(%input: tensor<1x4x4x1xi32>) -> tensor<1x2x2x1xi32> {
%fake = tensor.empty() : tensor<3x3xi32>
%init = tensor.empty() : tensor<1x2x2x1xi32>
%cst = arith.constant 0 : i32
%fill = linalg.fill ins(%cst : i32) outs(%init : tensor<1x2x2x1xi32>) -> tensor<1x2x2x1xi32>
%res = linalg.pooling_nhwc_max_unsigned {dilations = dense<1> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>}
ins(%input, %fake: tensor<1x4x4x1xi32>, tensor<3x3xi32>)
outs(%fill: tensor<1x2x2x1xi32>) -> tensor<1x2x2x1xi32>
return %res : tensor<1x2x2x1xi32>
}

// -----
// CHECK-LABEL: func @pooling_nwc_max_tensor
// CHECK: %{{.+}} = linalg.pooling_nwc_max
Expand Down Expand Up @@ -1017,6 +1034,23 @@ func.func @pooling_nhwc_min_tensor(%input: tensor<1x4x4x1xf32>) -> tensor<1x2x2x

// -----

// CHECK-LABEL: func @pooling_nhwc_min_unsigned_tensor
// CHECK: %{{.+}} = linalg.pooling_nhwc_min_unsigned
// CHECK-SAME: ins(%{{.+}}, %{{.+}} : tensor<1x4x4x1xi32>, tensor<3x3xi32>)
// CHECK-SAME: outs(%{{.+}} : tensor<1x2x2x1xi32>) -> tensor<1x2x2x1xi32>
func.func @pooling_nhwc_min_unsigned_tensor(%input: tensor<1x4x4x1xi32>) -> tensor<1x2x2x1xi32> {
%fake = tensor.empty() : tensor<3x3xi32>
%init = tensor.empty() : tensor<1x2x2x1xi32>
%cst = arith.constant 0 : i32
%fill = linalg.fill ins(%cst : i32) outs(%init : tensor<1x2x2x1xi32>) -> tensor<1x2x2x1xi32>
%res = linalg.pooling_nhwc_min_unsigned {dilations = dense<1> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>}
ins(%input, %fake: tensor<1x4x4x1xi32>, tensor<3x3xi32>)
outs(%fill: tensor<1x2x2x1xi32>) -> tensor<1x2x2x1xi32>
return %res : tensor<1x2x2x1xi32>
}

// -----

// CHECK-LABEL: func @pooling_nwc_min_tensor
// CHECK: %{{.+}} = linalg.pooling_nwc_min
// CHECK-SAME: dilations = dense<1> : tensor<1xi64>
Expand Down
28 changes: 14 additions & 14 deletions mlir/test/Dialect/Linalg/transform-op-decompose.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -131,21 +131,21 @@ func.func @pooling_nhwc_max(%input: tensor<?x1x?x?xf32>, %filter: tensor<1x?xf32
}

// CHECK-LABEL: @pooling_nhwc_max_unsigned
// CHECK-SAME: %[[ARG0:.+]]: tensor<?x1x?x?xf32>,
// CHECK-SAME: %[[ARG1:.+]]: tensor<1x?xf32>
// CHECK-SAME: %[[ARG2:.+]]: tensor<?x1x?x?xf32>
func.func @pooling_nhwc_max_unsigned(%input: tensor<?x1x?x?xf32>, %filter: tensor<1x?xf32>, %init: tensor<?x1x?x?xf32>) -> tensor<?x1x?x?xf32> {
// CHECK-SAME: %[[ARG0:.+]]: tensor<?x1x?x?xi32>,
// CHECK-SAME: %[[ARG1:.+]]: tensor<1x?xi32>
// CHECK-SAME: %[[ARG2:.+]]: tensor<?x1x?x?xi32>
func.func @pooling_nhwc_max_unsigned(%input: tensor<?x1x?x?xi32>, %filter: tensor<1x?xi32>, %init: tensor<?x1x?x?xi32>) -> tensor<?x1x?x?xi32> {
// CHECK: %[[SLICE0:.+]] = tensor.extract_slice %[[ARG0]]
// CHECK: %[[SLICE1:.+]] = tensor.extract_slice %[[ARG1]]
// CHECK: %[[SLICE2:.+]] = tensor.extract_slice %[[ARG2]]
// CHECK: %[[SLICERES:.+]] = linalg.pooling_nwc_max_unsigned
// CHECK: %[[RES:.+]] = tensor.insert_slice %[[SLICERES]] into %[[ARG2]]
%0 = linalg.pooling_nhwc_max_unsigned {dilations = dense<1> : tensor<2xi64>,
strides = dense<1> : tensor<2xi64>}
ins (%input, %filter: tensor<?x1x?x?xf32>, tensor<1x?xf32>)
outs (%init: tensor<?x1x?x?xf32>) -> tensor<?x1x?x?xf32>
ins (%input, %filter: tensor<?x1x?x?xi32>, tensor<1x?xi32>)
outs (%init: tensor<?x1x?x?xi32>) -> tensor<?x1x?x?xi32>
// CHECK: return %[[RES]]
return %0 : tensor<?x1x?x?xf32>
return %0 : tensor<?x1x?x?xi32>
}

// CHECK-LABEL: @pooling_nhwc_min
Expand All @@ -167,21 +167,21 @@ func.func @pooling_nhwc_min(%input: tensor<?x1x?x?xf32>, %filter: tensor<1x?xf32
}

// CHECK-LABEL: @pooling_nhwc_min_unsigned
// CHECK-SAME: %[[ARG0:.+]]: tensor<?x1x?x?xf32>,
// CHECK-SAME: %[[ARG1:.+]]: tensor<1x?xf32>
// CHECK-SAME: %[[ARG2:.+]]: tensor<?x1x?x?xf32>
func.func @pooling_nhwc_min_unsigned(%input: tensor<?x1x?x?xf32>, %filter: tensor<1x?xf32>, %init: tensor<?x1x?x?xf32>) -> tensor<?x1x?x?xf32> {
// CHECK-SAME: %[[ARG0:.+]]: tensor<?x1x?x?xi32>,
// CHECK-SAME: %[[ARG1:.+]]: tensor<1x?xi32>
// CHECK-SAME: %[[ARG2:.+]]: tensor<?x1x?x?xi32>
func.func @pooling_nhwc_min_unsigned(%input: tensor<?x1x?x?xi32>, %filter: tensor<1x?xi32>, %init: tensor<?x1x?x?xi32>) -> tensor<?x1x?x?xi32> {
// CHECK: %[[SLICE0:.+]] = tensor.extract_slice %[[ARG0]]
// CHECK: %[[SLICE1:.+]] = tensor.extract_slice %[[ARG1]]
// CHECK: %[[SLICE2:.+]] = tensor.extract_slice %[[ARG2]]
// CHECK: %[[SLICERES:.+]] = linalg.pooling_nwc_min_unsigned
// CHECK: %[[RES:.+]] = tensor.insert_slice %[[SLICERES]] into %[[ARG2]]
%0 = linalg.pooling_nhwc_min_unsigned {dilations = dense<1> : tensor<2xi64>,
strides = dense<1> : tensor<2xi64>}
ins (%input, %filter: tensor<?x1x?x?xf32>, tensor<1x?xf32>)
outs (%init: tensor<?x1x?x?xf32>) -> tensor<?x1x?x?xf32>
ins (%input, %filter: tensor<?x1x?x?xi32>, tensor<1x?xi32>)
outs (%init: tensor<?x1x?x?xi32>) -> tensor<?x1x?x?xi32>
// CHECK: return %[[RES]]
return %0 : tensor<?x1x?x?xf32>
return %0 : tensor<?x1x?x?xi32>
}

// CHECK-LABEL: @pooling_nchw_max
Expand Down
48 changes: 48 additions & 0 deletions mlir/test/python/dialects/linalg/opdsl/emit_pooling.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -150,3 +150,51 @@ def test_f32f32_min_pooling(input, shape, init_result):


print(module)

with Context() as ctx, Location.unknown():
module = Module.create()
with InsertionPoint(module.body):
f32 = F32Type.get()
bool_t = IntegerType.get_signless(1)

# CHECK: bool_max_unsigned_error: Unsupported 'max_unsigned' operands
@func.FuncOp.from_py_func(
RankedTensorType.get((1, 4, 16, 1), f32),
RankedTensorType.get((2, 2), f32),
RankedTensorType.get((1, 2, 4, 1), bool_t),
)
def test_bool_i1_max_unsigned_pooling_error(input, shape, init_result):
try:
pooling_poly(
input,
shape,
outs=[init_result],
reduce=BinaryFn.max_unsigned,
cast=TypeFn.cast_unsigned,
strides=[2, 4],
dilations=[1, 2],
)
except NotImplementedError as e:
print(f"bool_max_unsigned_error: {e}")
return init_result

# CHECK: float_max_unsigned_error: Unsupported 'max_unsigned' operands
@func.FuncOp.from_py_func(
RankedTensorType.get((1, 4, 16, 1), f32),
RankedTensorType.get((2, 2), f32),
RankedTensorType.get((1, 2, 4, 1), f32),
)
def test_f32f32_max_unsigned_pooling_error(input, shape, init_result):
try:
pooling_poly(
input,
shape,
outs=[init_result],
reduce=BinaryFn.max_unsigned,
cast=TypeFn.cast_unsigned,
strides=[2, 4],
dilations=[1, 2],
)
except NotImplementedError as e:
print(f"float_max_unsigned_error: {e}")
return init_result
Loading