Skip to content

[OnnxToTorch] Lower onnx.MeanVarianceNorm to torch dialect without expansion #4219

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 5 commits into from
Aug 6, 2025
Merged

[OnnxToTorch] Lower onnx.MeanVarianceNorm to torch dialect without expansion #4219

Show file tree
Hide file tree
Changes from 4 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
e8a7ddf
Lower to torch dialect without expansion
zahidwx Jun 2, 2025
75bbd1c
Remove MeanVarNorm from allowlists to avoid expansion
zahidwx Jun 2, 2025
fb80dc1
Add MeanVarNorm lit tests
zahidwx Jun 2, 2025
c9fe057
Add type checks & allow MVN expansion by default
zahidwx Jul 1, 2025
801ba0d
Minor fix for negative axes
zahidwx Jul 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
68 changes: 68 additions & 0 deletions lib/Conversion/TorchOnnxToTorch/DefaultDomainGtoP.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1606,6 +1606,74 @@ void mlir::torch::onnx_c::populateDefaultDomainGtoP(
/* cudnn enabled */ boolFalse);
return success();
});
patterns.onOp(
"MeanVarianceNormalization", 13,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value input;
SmallVector<int64_t> axes;

if (binder.tensorOperand(input) ||
binder.s64IntegerArrayAttr(axes, "axes",
llvm::SmallVector<int64_t>({0, 2, 3})) ||
binder.tensorResultType(resultType)) {
return failure();
}
if (!resultType.hasSizes() || !resultType.hasDtype()) {
return failure();
}
auto inputTy = cast<Torch::ValueTensorType>(input.getType());
if (!inputTy || !inputTy.hasSizes()) {
return failure();
}
int64_t inputRank = inputTy.getSizes().size();

Location loc = binder.getLoc();
Value keepDim = rewriter.create<Torch::ConstantBoolOp>(loc, true);
Value unBiased = rewriter.create<Torch::ConstantBoolOp>(loc, false);
Value none = rewriter.create<Torch::ConstantNoneOp>(loc);

ArrayRef<int64_t> output_shape = resultType.getSizes();
SmallVector<int64_t> reduced_shape(output_shape);

for (int64_t i : axes) {
int64_t dim = Torch::toPositiveDim(i, inputRank);
if (!Torch::isValidDim(dim, inputRank)) {
return failure();
}
reduced_shape[i] = 1;
}
Torch::ValueTensorType reducedOutTy = Torch::ValueTensorType::get(
resultType.getContext(), reduced_shape, resultType.getDtype());
SmallVector<Value> cstAxes;
for (int64_t i : axes) {
cstAxes.push_back(rewriter.create<Torch::ConstantIntOp>(
loc, rewriter.getI64IntegerAttr(i)));
}
Value axes_list = rewriter.create<Torch::PrimListConstructOp>(
loc,
Torch::ListType::get(Torch::IntType::get(binder.op->getContext())),
cstAxes);
Value mean = rewriter.create<Torch::AtenMeanDimOp>(
loc, reducedOutTy, input, axes_list, keepDim, none);
Value variance = rewriter.create<Torch::AtenVarDimOp>(
loc, reducedOutTy, input, axes_list, unBiased, keepDim);
Value cstOne = rewriter.create<Torch::ConstantIntOp>(
loc, rewriter.getI64IntegerAttr(1));
Value cstEps = rewriter.create<Torch::ConstantFloatOp>(
loc, rewriter.getF64FloatAttr(1e-9));
variance = rewriter.create<Torch::AtenAddScalarOp>(
loc, reducedOutTy, variance, cstEps, cstOne);
Value sqrtVar =
rewriter.create<Torch::AtenSqrtOp>(loc, reducedOutTy, variance);
Value inputMinusMean = rewriter.create<Torch::AtenSubTensorOp>(
loc, resultType, input, mean, cstOne);
Value meanVarNorm = rewriter.create<Torch::AtenDivTensorOp>(
loc, resultType, inputMinusMean, sqrtVar);

rewriter.replaceOp(binder.op, meanVarNorm);
return success();
});
patterns.onOp(
"Max", 1, [](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Expand Down
51 changes: 51 additions & 0 deletions test/Conversion/TorchOnnxToTorch/simple_ops_g_to_p.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1595,6 +1595,57 @@ func.func @test_mod_int64_no_fmod(%arg0: !torch.vtensor<[6],si64>, %arg1: !torch

// -----

// CHECK-LABEL: func.func @test_meanvarnorm(
func.func @test_meanvarnorm(%arg0: !torch.vtensor<[3,5,2,2],f32>) -> !torch.vtensor<[3,5,2,2],f32> attributes {torch.onnx_meta.ir_version = 3 : si64, torch.onnx_meta.opset_version = 13 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
// CHECK-SAME: %[[ARG0:.*]]: !torch.vtensor<[3,5,2,2],f32>) -> !torch.vtensor<[3,5,2,2],f32> attributes {torch.onnx_meta.ir_version = 3 : si64, torch.onnx_meta.opset_version = 13 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
// CHECK: %[[VAL_0:.*]] = torch.constant.bool true
// CHECK: %[[VAL_1:.*]] = torch.constant.bool false
// CHECK: %[[VAL_2:.*]] = torch.constant.none
// CHECK: %[[VAL_3:.*]] = torch.constant.int 0
// CHECK: %[[VAL_4:.*]] = torch.constant.int 2
// CHECK: %[[VAL_5:.*]] = torch.constant.int 3
// CHECK: %[[VAL_6:.*]] = torch.prim.ListConstruct %[[VAL_3]], %[[VAL_4]], %[[VAL_5]] : (!torch.int, !torch.int, !torch.int) -> !torch.list<int>
// CHECK: %[[VAL_7:.*]] = torch.aten.mean.dim %[[ARG0]], %[[VAL_6]], %[[VAL_0]], %[[VAL_2]] : !torch.vtensor<[3,5,2,2],f32>, !torch.list<int>, !torch.bool, !torch.none -> !torch.vtensor<[1,5,1,1],f32>
// CHECK: %[[VAL_8:.*]] = torch.aten.var.dim %[[ARG0]], %[[VAL_6]], %[[VAL_1]], %[[VAL_0]] : !torch.vtensor<[3,5,2,2],f32>, !torch.list<int>, !torch.bool, !torch.bool -> !torch.vtensor<[1,5,1,1],f32>
// CHECK: %[[VAL_9:.*]] = torch.constant.int 1
// CHECK: %[[VAL_10:.*]] = torch.constant.float 1.000000e-09
// CHECK: %[[VAL_11:.*]] = torch.aten.add.Scalar %[[VAL_8]], %[[VAL_10]], %[[VAL_9]] : !torch.vtensor<[1,5,1,1],f32>, !torch.float, !torch.int -> !torch.vtensor<[1,5,1,1],f32>
// CHECK: %[[VAL_12:.*]] = torch.aten.sqrt %[[VAL_11]] : !torch.vtensor<[1,5,1,1],f32> -> !torch.vtensor<[1,5,1,1],f32>
// CHECK: %[[VAL_13:.*]] = torch.aten.sub.Tensor %[[ARG0]], %[[VAL_7]], %[[VAL_9]] : !torch.vtensor<[3,5,2,2],f32>, !torch.vtensor<[1,5,1,1],f32>, !torch.int -> !torch.vtensor<[3,5,2,2],f32>
// CHECK: %[[VAL_14:.*]] = torch.aten.div.Tensor %[[VAL_13]], %[[VAL_12]] : !torch.vtensor<[3,5,2,2],f32>, !torch.vtensor<[1,5,1,1],f32> -> !torch.vtensor<[3,5,2,2],f32>
// CHECK: return %[[VAL_14]] : !torch.vtensor<[3,5,2,2],f32>
// CHECK: }
%0 = torch.operator "onnx.MeanVarianceNormalization"(%arg0) : (!torch.vtensor<[3,5,2,2],f32>) -> !torch.vtensor<[3,5,2,2],f32>
return %0 : !torch.vtensor<[3,5,2,2],f32>
}

// -----

// CHECK-LABEL: func.func @test_meanvarnorm_axes(
func.func @test_meanvarnorm_axes(%arg0: !torch.vtensor<[3,5,2,2],f32>) -> !torch.vtensor<[3,5,2,2],f32> attributes {torch.onnx_meta.ir_version = 3 : si64, torch.onnx_meta.opset_version = 13 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
// CHECK-SAME: %[[ARG0:.*]]: !torch.vtensor<[3,5,2,2],f32>) -> !torch.vtensor<[3,5,2,2],f32> attributes {torch.onnx_meta.ir_version = 3 : si64, torch.onnx_meta.opset_version = 13 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
// CHECK: %[[VAL_0:.*]] = torch.constant.bool true
// CHECK: %[[VAL_1:.*]] = torch.constant.bool false
// CHECK: %[[VAL_2:.*]] = torch.constant.none
// CHECK: %[[VAL_3:.*]] = torch.constant.int 1
// CHECK: %[[VAL_4:.*]] = torch.constant.int 3
// CHECK: %[[VAL_5:.*]] = torch.prim.ListConstruct %[[VAL_3]], %[[VAL_4]] : (!torch.int, !torch.int) -> !torch.list<int>
// CHECK: %[[VAL_6:.*]] = torch.aten.mean.dim %[[ARG0]], %[[VAL_5]], %[[VAL_0]], %[[VAL_2]] : !torch.vtensor<[3,5,2,2],f32>, !torch.list<int>, !torch.bool, !torch.none -> !torch.vtensor<[3,1,2,1],f32>
// CHECK: %[[VAL_7:.*]] = torch.aten.var.dim %[[ARG0]], %[[VAL_5]], %[[VAL_1]], %[[VAL_0]] : !torch.vtensor<[3,5,2,2],f32>, !torch.list<int>, !torch.bool, !torch.bool -> !torch.vtensor<[3,1,2,1],f32>
// CHECK: %[[VAL_8:.*]] = torch.constant.int 1
// CHECK: %[[VAL_9:.*]] = torch.constant.float 1.000000e-09
// CHECK: %[[VAL_10:.*]] = torch.aten.add.Scalar %[[VAL_7]], %[[VAL_9]], %[[VAL_8]] : !torch.vtensor<[3,1,2,1],f32>, !torch.float, !torch.int -> !torch.vtensor<[3,1,2,1],f32>
// CHECK: %[[VAL_11:.*]] = torch.aten.sqrt %[[VAL_10]] : !torch.vtensor<[3,1,2,1],f32> -> !torch.vtensor<[3,1,2,1],f32>
// CHECK: %[[VAL_12:.*]] = torch.aten.sub.Tensor %[[ARG0]], %[[VAL_6]], %[[VAL_8]] : !torch.vtensor<[3,5,2,2],f32>, !torch.vtensor<[3,1,2,1],f32>, !torch.int -> !torch.vtensor<[3,5,2,2],f32>
// CHECK: %[[VAL_13:.*]] = torch.aten.div.Tensor %[[VAL_12]], %[[VAL_11]] : !torch.vtensor<[3,5,2,2],f32>, !torch.vtensor<[3,1,2,1],f32> -> !torch.vtensor<[3,5,2,2],f32>
// CHECK: return %[[VAL_13]] : !torch.vtensor<[3,5,2,2],f32>
// CHECK: }
%0 = torch.operator "onnx.MeanVarianceNormalization"(%arg0) {torch.onnx.axes = [1 : si64, 3 : si64]} : (!torch.vtensor<[3,5,2,2],f32>) -> !torch.vtensor<[3,5,2,2],f32>
return %0 : !torch.vtensor<[3,5,2,2],f32>
}

// -----

// CHECK-LABEL: func.func @test_not_2d
func.func @test_not_2d(%arg0: !torch.vtensor<[3,4],i1>) -> !torch.vtensor<[3,4],i1> attributes {torch.onnx_meta.ir_version = 3 : si64, torch.onnx_meta.opset_version = 1 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
// CHECK: torch.aten.bitwise_not %arg0 : !torch.vtensor<[3,4],i1> -> !torch.vtensor<[3,4],i1>
Expand Down
Loading