OnnxToTorch support for onnx.InstanceNormalization op (#2710)

https://github.com/nod-ai/SHARK-Turbine/issues/327

include/torch-mlir/Dialect/Torch/IR/GeneratedTorchOps.td

@@ -6203,6 +6203,37 @@ def Torch_AtenBatchNormOp : Torch_Op<"aten.batch_norm", [
   }];
 }
 
+def Torch_AtenInstanceNormOp : Torch_Op<"aten.instance_norm", [
+    AllowsTypeRefinement,
+    HasValueSemantics,
+    ReadOnly
+  ]> {
+  let summary = "Generated op for `aten::instance_norm : (Tensor, Tensor?, Tensor?, Tensor?, Tensor?, bool, float, float, bool) -> (Tensor)`";
+  let arguments = (ins
+    AnyTorchTensorType:$input,
+    AnyTorchOptionalTensorType:$weight,
+    AnyTorchOptionalTensorType:$bias,
+    AnyTorchOptionalTensorType:$running_mean,
+    AnyTorchOptionalTensorType:$running_var,
+    Torch_BoolType:$use_input_stats,
+    Torch_FloatType:$momentum,
+    Torch_FloatType:$eps,
+    Torch_BoolType:$cudnn_enabled
+  );
+  let results = (outs
+    AnyTorchTensorType:$result
+  );
+  let hasCustomAssemblyFormat = 1;
+  let extraClassDefinition = [{
+    ParseResult AtenInstanceNormOp::parse(OpAsmParser &parser, OperationState &result) {
+      return parseDefaultTorchOp(parser, result, 9, 1);
+    }
+    void AtenInstanceNormOp::print(OpAsmPrinter &printer) {
+      printDefaultTorchOp(printer, *this, 9, 1);
+    }
+  }];
+}
+
 def Torch_AtenNativeGroupNormOp : Torch_Op<"aten.native_group_norm", [
     AllowsTypeRefinement,
     HasValueSemantics,

lib/Conversion/TorchOnnxToTorch/DefaultDomainGtoP.cpp

@@ -392,6 +392,37 @@ void mlir::torch::onnx_c::populateDefaultDomainGtoP(
                       binder.op, resultType, lhs, rhs);
                   return success();
                 });
+  patterns.onOp(
+      "InstanceNormalization", 6,
+      [](OpBinder binder, ConversionPatternRewriter &rewriter) {
+        Torch::ValueTensorType resultType;
+        llvm::SmallVector<Value> operands;
+        float eps;
+
+        if (binder.tensorOperands(operands, 3) ||
+            binder.tensorResultType(resultType) || operands.size() != 3 ||
+            binder.f32FloatAttr(eps, "epsilon", 1e-05f)) {
+          return failure();
+        }
+        Value none = rewriter.create<Torch::ConstantNoneOp>(binder.getLoc());
+        Value boolTrue =
+            rewriter.create<Torch::ConstantBoolOp>(binder.getLoc(), true);
+        Value boolFalse =
+            rewriter.create<Torch::ConstantBoolOp>(binder.getLoc(), false);
+        auto epsValue = rewriter.create<Torch::ConstantFloatOp>(
+            binder.getLoc(), rewriter.getF64FloatAttr(eps));
+
+        auto momentum = rewriter.create<Torch::ConstantFloatOp>(
+            binder.getLoc(), rewriter.getF64FloatAttr(0.0f));
+        rewriter.replaceOpWithNewOp<Torch::AtenInstanceNormOp>(
+            binder.op, resultType, /* input */ operands[0],
+            /* weight */ operands[1],
+            /* bias */ operands[2], /* running mean */ none,
+            /* running var */ none,
+            /* use input stats */ boolTrue, momentum, epsValue,
+            /* cudnn enabled */ boolFalse);
+        return success();
+      });
   patterns.onOp(
       "Max", 1, [](OpBinder binder, ConversionPatternRewriter &rewriter) {
         Torch::ValueTensorType resultType;

lib/Dialect/Torch/Transforms/AbstractInterpLibrary.cpp

@@ -8784,6 +8784,10 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    %3 = torch.prim.TupleConstruct %0, %1, %2 : !torch.list<int>, !torch.list<int>, !torch.list<int> -> !torch.tuple<list<int>, list<int>, list<int>>\n"
 "    return %3 : !torch.tuple<list<int>, list<int>, list<int>>\n"
 "  }\n"
+"  func.func @\"__torch_mlir_shape_fn.aten.instance_norm\"(%arg0: !torch.list<int>, %arg1: !torch.optional<list<int>>, %arg2: !torch.optional<list<int>>, %arg3: !torch.optional<list<int>>, %arg4: !torch.optional<list<int>>, %arg5: !torch.bool, %arg6: !torch.float, %arg7: !torch.float, %arg8: !torch.bool) -> !torch.list<int> {\n"
+"    %0 = call @__torch__.torch.jit._shape_functions.unary(%arg0) : (!torch.list<int>) -> !torch.list<int>\n"
+"    return %0 : !torch.list<int>\n"
+"  }\n"
 "  func.func @\"__torch_mlir_shape_fn.aten.slice.Tensor\"(%arg0: !torch.list<int>, %arg1: !torch.int, %arg2: !torch.optional<int>, %arg3: !torch.optional<int>, %arg4: !torch.int) -> !torch.list<int> {\n"
 "    %0 = call @__torch__.torch.jit._shape_functions.slice(%arg0, %arg1, %arg2, %arg3, %arg4) : (!torch.list<int>, !torch.int, !torch.optional<int>, !torch.optional<int>, !torch.int) -> !torch.list<int>\n"
 "    return %0 : !torch.list<int>\n"
@@ -9643,6 +9647,10 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    %3 = torch.prim.TupleConstruct %0#1, %0#1, %0#1 : !torch.int, !torch.int, !torch.int -> !torch.tuple<int, int, int>\n"
 "    return %3 : !torch.tuple<int, int, int>\n"
 "  }\n"
+"  func.func @\"__torch_mlir_dtype_fn.aten.instance_norm\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.optional<tuple<int, int>>, %arg2: !torch.optional<tuple<int, int>>, %arg3: !torch.optional<tuple<int, int>>, %arg4: !torch.optional<tuple<int, int>>, %arg5: !torch.bool, %arg6: !torch.float, %arg7: !torch.float, %arg8: !torch.bool) -> !torch.int {\n"
+"    %0:2 = torch.prim.TupleUnpack %arg0 : !torch.tuple<int, int> -> !torch.int, !torch.int\n"
+"    return %0#1 : !torch.int\n"
+"  }\n"
 "  func.func @\"__torch_mlir_dtype_fn.aten.bernoulli_.float\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.float, %arg2: !torch.any) -> !torch.int {\n"
 "    %0:2 = torch.prim.TupleUnpack %arg0 : !torch.tuple<int, int> -> !torch.int, !torch.int\n"
 "    return %0#1 : !torch.int\n"

lib/Dialect/Torch/Transforms/DecomposeComplexOps.cpp

@@ -3962,6 +3962,151 @@ class DecomposeAtenLayerNormOp : public OpRewritePattern<AtenLayerNormOp> {
 };
 } // namespace
 
+namespace {
+class DecomposeAtenInstanceNormOp
+    : public OpRewritePattern<AtenInstanceNormOp> {
+  using OpRewritePattern<AtenInstanceNormOp>::OpRewritePattern;
+  LogicalResult matchAndRewrite(AtenInstanceNormOp op,
+                                PatternRewriter &rewriter) const override {
+
+    Location loc = op.getLoc();
+    auto context = op.getContext();
+
+    auto inputTy = op.getInput().getType().cast<BaseTensorType>();
+    int64_t inputRank = inputTy.getSizes().size();
+    auto reduceDimInts =
+        llvm::SmallVector<int64_t>({inputRank - 2, inputRank - 1});
+
+    SmallVector<int64_t> reducedShape(inputTy.getSizes());
+    reducedShape[inputRank - 1] = 1;
+    reducedShape[inputRank - 2] = 1;
+
+    Type dtype = inputTy.getOptionalDtype();
+    Type reducedTy = ValueTensorType::get(op.getContext(),
+                                          llvm::ArrayRef(reducedShape), dtype);
+
+    auto sizeListType = ListType::get(IntType::get(context));
+    SmallVector<Value> reduceDimVals;
+    reduceDimVals.reserve(reduceDimInts.size());
+    std::transform(reduceDimInts.begin(), reduceDimInts.end(),
+                   std::back_inserter(reduceDimVals), [&](int64_t d) {
+                     return rewriter.create<Torch::ConstantIntOp>(
+                         loc, rewriter.getI64IntegerAttr(d));
+                   });
+    Value reduceDimList =
+        rewriter.create<PrimListConstructOp>(loc, sizeListType, reduceDimVals);
+    Value cstTrue = rewriter.create<Torch::ConstantBoolOp>(loc, true);
+    Value none = rewriter.create<Torch::ConstantNoneOp>(loc);
+
+    Value one = rewriter.create<Torch::ConstantIntOp>(
+        loc, rewriter.getI64IntegerAttr(1));
+
+    // mean(x)
+    Value inputMean = rewriter.create<AtenMeanDimOp>(
+        loc, reducedTy, op.getInput(), reduceDimList, cstTrue, none);
+
+    // x - mean(x)
+    Value inputMeanExpanded =
+        rewriter.create<AtenExpandAsOp>(loc, inputTy, inputMean, op.getInput());
+    Value inputSubMean = rewriter.create<AtenSubTensorOp>(
+        loc, inputTy, op.getInput(), inputMeanExpanded, one);
+    // (x - mean(x))^2
+    Value inputSubMeanSquare = rewriter.create<AtenMulTensorOp>(
+        loc, inputTy, inputSubMean, inputSubMean);
+
+    Value variancesum = rewriter.create<AtenSumDimIntListOp>(
+        loc, reducedTy, inputSubMeanSquare, reduceDimList, cstTrue,
+        /*dtype=*/none);
+
+    Value hw = rewriter.create<Torch::ConstantIntOp>(
+        loc, rewriter.getI64IntegerAttr(inputTy.getSizes()[inputRank - 1] *
+                                        inputTy.getSizes()[inputRank - 2]));
+    Value inputVar =
+        rewriter.create<AtenDivScalarOp>(loc, reducedTy, variancesum, hw);
+
+    // rsqrt(var(x) + eps)
+    Value inputVarPlusEps = rewriter.create<AtenAddScalarOp>(
+        loc, reducedTy, inputVar, op.getEps(), one);
+    Value inputRsqrtVar =
+        rewriter.create<AtenRsqrtOp>(loc, reducedTy, inputVarPlusEps);
+
+    // (x - mean(x)) * rsqrt(var(x) + eps)
+    Value inputRsqrtVarExpanded = rewriter.create<AtenExpandAsOp>(
+        loc, inputTy, inputRsqrtVar, op.getInput());
+    Value inputNormalized = rewriter.create<AtenMulTensorOp>(
+        loc, inputTy, inputSubMean, inputRsqrtVarExpanded);
+    Value out = rewriter.create<TensorStaticInfoCastOp>(
+        loc, op.getResult().getType(), inputNormalized);
+
+    Value weight = op.getWeight();
+    auto weightTy = weight.getType().cast<BaseTensorType>();
+    dtype = weightTy.getOptionalDtype();
+
+    SmallVector<int64_t> weightShape(weightTy.getSizes());
+    SmallVector<int64_t> newWeightShape;
+    newWeightShape.push_back(1);
+    newWeightShape.append(weightShape);
+
+    Value zero = rewriter.create<Torch::ConstantIntOp>(
+        loc, rewriter.getI64IntegerAttr(0));
+    Type newWeightTy = ValueTensorType::get(
+        op.getContext(), llvm::ArrayRef(newWeightShape), dtype);
+    weight = rewriter.create<AtenUnsqueezeOp>(loc, newWeightTy, weight, zero);
+
+    Value two = rewriter.create<Torch::ConstantIntOp>(
+        loc, rewriter.getI64IntegerAttr(2));
+    newWeightShape.push_back(1);
+    newWeightTy = ValueTensorType::get(op.getContext(),
+                                       llvm::ArrayRef(newWeightShape), dtype);
+    weight = rewriter.create<AtenUnsqueezeOp>(loc, newWeightTy, weight, two);
+
+    Value three = rewriter.create<Torch::ConstantIntOp>(
+        loc, rewriter.getI64IntegerAttr(3));
+    newWeightShape.push_back(1);
+    newWeightTy = ValueTensorType::get(op.getContext(),
+                                       llvm::ArrayRef(newWeightShape), dtype);
+    weight = rewriter.create<AtenUnsqueezeOp>(loc, newWeightTy, weight, three);
+
+    Value weightExpanded =
+        rewriter.create<AtenExpandAsOp>(loc, inputTy, weight, op.getInput());
+
+    Value bias = op.getBias();
+    auto biasTy = bias.getType().cast<BaseTensorType>();
+    dtype = biasTy.getOptionalDtype();
+
+    SmallVector<int64_t> biasShape(biasTy.getSizes());
+    SmallVector<int64_t> newBiasShape;
+    newBiasShape.push_back(1);
+    newBiasShape.append(biasShape);
+
+    Type newBiasTy = ValueTensorType::get(op.getContext(),
+                                          llvm::ArrayRef(newBiasShape), dtype);
+    bias = rewriter.create<AtenUnsqueezeOp>(loc, newBiasTy, bias, zero);
+
+    newBiasShape.push_back(1);
+    newBiasTy = ValueTensorType::get(op.getContext(),
+                                     llvm::ArrayRef(newBiasShape), dtype);
+    bias = rewriter.create<AtenUnsqueezeOp>(loc, newBiasTy, bias, two);
+
+    newBiasShape.push_back(1);
+    newBiasTy = ValueTensorType::get(op.getContext(),
+                                     llvm::ArrayRef(newBiasShape), dtype);
+    bias = rewriter.create<AtenUnsqueezeOp>(loc, newBiasTy, bias, three);
+
+    Value biasExpanded =
+        rewriter.create<AtenExpandAsOp>(loc, inputTy, bias, op.getInput());
+
+    out = rewriter.create<AtenMulTensorOp>(loc, out.getType(), out,
+                                           weightExpanded);
+    out = rewriter.create<AtenAddTensorOp>(loc, out.getType(), out,
+                                           biasExpanded, one);
+
+    rewriter.replaceOp(op, out);
+    return success();
+  }
+};
+} // namespace
+
 namespace {
 class DecomposeAtenNativeLayerNormOp
     : public OpRewritePattern<AtenNativeLayerNormOp> {
@@ -6733,6 +6878,7 @@ public:
         DecomposeAtenAddCLikeOp<AtenAddcmulOp, AtenMulTensorOp>>(patterns);
     addPatternIfTargetOpIsIllegal<
         DecomposeAtenAddCLikeOp<AtenAddcdivOp, AtenDivTensorOp>>(patterns);
+    addPatternIfTargetOpIsIllegal<DecomposeAtenInstanceNormOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenLayerNormOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenNativeLayerNormOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenGroupNormOp>(patterns);

lib/Dialect/Torch/Transforms/LowerToBackendContract.cpp

@@ -409,6 +409,7 @@ static void markDecomposedOpsAsIllegal(MLIRContext *context,
   });
   target.addIllegalOp<AtenAddcmulOp>();
   target.addIllegalOp<AtenAddcdivOp>();
+  target.addIllegalOp<AtenInstanceNormOp>();
   target.addIllegalOp<AtenLayerNormOp>();
   target.addIllegalOp<AtenNativeLayerNormOp>();
   target.addIllegalOp<AtenGroupNormOp>();

projects/pt1/e2e_testing/xfail_sets.py

@@ -233,6 +233,7 @@ TORCHDYNAMO_XFAIL_SET = {
     # END tests failing due to: 'torch.aten.mul.Tensor' op operand #1 must be Any Torch tensor type, but got '!torch.float'
 
     # START tests failing due to: 'torch.aten.add.Tensor' op operand #1 must be Any Torch tensor type, but got '!torch.float'
+    "AtenInstanceNormModule_basic",
     "BatchNorm1DModule_basic",
     "BatchNorm1DWith2DInputModule_basic",
     "BatchNorm2DModule_basic",
@@ -898,6 +899,7 @@ TOSA_PASS_SET = {
     "AtenEyeModuleFalsePinMemory_basic",
     "AtenEyeModuleFloat2D_basic",
     "AtenRoundIntModule_basic",
+    "AtenInstanceNormModule_basic",
     "AtenToDeviceModule_basic",
     "BaddbmmBroadcast1DInputModule_basic",
     "BaddbmmBroadcast2DInputModule_basic",
@@ -1306,6 +1308,8 @@ MAKE_FX_TOSA_PASS_SET = (TOSA_PASS_SET | {
     "Conv2dNoPaddingModule_basic",
     "Conv2dWithPaddingDilationStrideModule_basic",
     "Conv2dWithPaddingModule_basic",
+
+    "AtenInstanceNormModule_basic",
 }
 
 LTC_CRASHING_SET = {

projects/pt1/python/torch_mlir/jit_ir_importer/build_tools/abstract_interp_lib_gen.py

@@ -1415,6 +1415,9 @@ def aten〇group_norm〡shape(input: List[int], num_groups: int, weight: Optiona
 def aten〇native_group_norm〡shape(input: List[int], weight: Optional[List[int]], bias: Optional[List[int]], N: int, C: int, HxW: int, group: int, eps: float) -> Tuple[List[int], List[int], List[int]]:
     return upstream_shape_functions.unary(input), [N, group], [N, group]
 
+def aten〇instance_norm〡shape(input: List[int], weight: Optional[List[int]], bias: Optional[List[int]], running_mean: Optional[List[int]], running_var: Optional[List[int]], use_input_stats: bool, momentum: float, eps: float, cudnn_enabled: bool) -> List[int]:
+    return upstream_shape_functions.unary(input)
+
 def aten〇slice〇Tensor〡shape(self: List[int], dim: int = 0, start: Optional[int] = None, end: Optional[int] = None, step: int = 1) -> List[int]:
     return upstream_shape_functions.slice(self, dim, start, end, step)
 
@@ -2048,6 +2051,11 @@ def aten〇native_group_norm〡dtype(input_rank_dtype: Tuple[int, int], weight_r
     assert not is_integer_dtype(input_dtype)
     return input_dtype, input_dtype, input_dtype
 
+# device is not supported hence unable to check the dtype function
+def aten〇instance_norm〡dtype(input_rank_dtype: Tuple[int, int], weight_rank_dtype: Optional[Tuple[int, int]], bias_rank_dtype: Optional[Tuple[int, int]], running_mean_rank_dtype: Optional[Tuple[int, int]], running_var_rank_dtype: Optional[Tuple[int, int]], use_input_stats: bool, momentum: float, eps: float, cudnn_enabled: bool) -> int:
+    input_rank, input_dtype = input_rank_dtype
+    return input_dtype
+
 @check_dtype_function(_check_tensors_with_the_same_dtype(num_of_tensors=1))
 def aten〇bernoulli_〇float〡dtype(self_rank_dtype: Tuple[int, int], p: float = 0.5, generator: Any = None) -> int:
     self_rank, self_dtype = self_rank_dtype

projects/pt1/python/torch_mlir/jit_ir_importer/build_tools/torch_ods_gen.py

@@ -437,6 +437,9 @@ def emit_ops(emitter_td: TextEmitter, registry: Registry):
     emit(
         "aten::batch_norm : (Tensor, Tensor?, Tensor?, Tensor?, Tensor?, bool, float, float, bool) -> (Tensor)"
     )
+    emit(
+        "aten::instance_norm : (Tensor, Tensor?, Tensor?, Tensor?, Tensor?, bool, float, float, bool) -> (Tensor)"
+    )
     emit(
         "aten::native_group_norm : (Tensor, Tensor?, Tensor?, int, int, int, int, float) -> (Tensor, Tensor, Tensor)"
     )

projects/pt1/python/torch_mlir_e2e_test/test_suite/norm_like.py

@@ -489,3 +489,21 @@ class LayerNormNormalizeOverAllDimsModule(torch.nn.Module):
 def LayerNormNormalizeOverAllDimsModule_basic(module, tu: TestUtils):
     module.forward(tu.rand(2, 2, 3))
 
+class AtenInstanceNormModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args([
+        None,
+        ([1, 2, 1, 3], torch.float32, True),
+        ([2], torch.float32, True),
+        ([2], torch.float32, True)
+    ])
+    def forward(self, x, w, b):
+        return torch.ops.aten.instance_norm(x, w, b, None,
+                None, True, 0.0, 1e-05, False)
+
+@register_test_case(module_factory=lambda: AtenInstanceNormModule())
+def AtenInstanceNormModule_basic(module, tu: TestUtils):
+    module.forward(tu.rand(1, 2, 1, 3), tu.rand(2), tu.rand(2))

test/Conversion/TorchOnnxToTorch/simple_ops_g_to_p.mlir

@@ -603,6 +603,15 @@ func.func @test_globalaveragepool_precomputed(%arg0: !torch.vtensor<[1,1,3,3],f3
 
 // -----
 
+// CHECK-LABEL: func.func @test_instancenorm
+  func.func @test_instancenorm(%arg0: !torch.vtensor<[1,2,1,3],f32>, %arg1: !torch.vtensor<[2],f32>, %arg2: !torch.vtensor<[2],f32>) -> !torch.vtensor<[1,2,1,3],f32> attributes {torch.onnx_meta.ir_version = 3 : si64, torch.onnx_meta.opset_version = 6 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
+    // CHECK: torch.aten.instance_norm %arg0, %arg1, %arg2, %none, %none, %true, %float0.000000e00, %float9.999990e-06, %false : !torch.vtensor<[1,2,1,3],f32>, !torch.vtensor<[2],f32>, !torch.vtensor<[2],f32>, !torch.none, !torch.none, !torch.bool, !torch.float, !torch.float, !torch.bool -> !torch.vtensor<[1,2,1,3],f32>
+    %0 = torch.operator "onnx.InstanceNormalization"(%arg0, %arg1, %arg2) : (!torch.vtensor<[1,2,1,3],f32>, !torch.vtensor<[2],f32>, !torch.vtensor<[2],f32>) -> !torch.vtensor<[1,2,1,3],f32>
+    return %0 : !torch.vtensor<[1,2,1,3],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>