Add decomposition for aten.miopen_batch_norm to aten.native_batch_norm

On ROCm 7.1+ with MIOpen 3.5+, PyTorch decomposes BatchNorm2d with
channels-last layout to aten.miopen_batch_norm instead of
aten._native_batch_norm_legit_functional. torch-mlir didn't support
this op, causing legalization to fail.

Fixes: #4476

- Register aten.miopen_batch_norm in torch_ods_gen.py (signature:
  (Tensor, Tensor, Tensor?, Tensor?, Tensor?, bool, float, float) -> (Tensor, Tensor, Tensor))
- Add decomposition pattern to DecomposeComplexOps.cpp that rewrites
  miopen_batch_norm to native_batch_norm by passing weight directly
  (type-compatible since native_batch_norm accepts AnyTorchOptionalTensorType)
- Add lit test verifying the op is decomposed away

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: rkayaith

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

@@ -7353,6 +7353,38 @@ def Torch_AtenNativeBatchNormOp : Torch_Op<"aten.native_batch_norm", [
   }];
 }
 
+def Torch_AtenMiopenBatchNormOp : Torch_Op<"aten.miopen_batch_norm", [
+    AllowsTypeRefinement,
+    HasValueSemantics,
+    ReadOnly
+  ]> {
+  let summary = "Generated op for `aten::miopen_batch_norm : (Tensor, Tensor, Tensor?, Tensor?, Tensor?, bool, float, float) -> (Tensor, Tensor, Tensor)`";
+  let arguments = (ins
+    AnyTorchTensorType:$input,
+    AnyTorchTensorType:$weight,
+    AnyTorchOptionalTensorType:$bias,
+    AnyTorchOptionalTensorType:$running_mean,
+    AnyTorchOptionalTensorType:$running_var,
+    Torch_BoolType:$training,
+    Torch_FloatType:$exponential_average_factor,
+    Torch_FloatType:$epsilon
+  );
+  let results = (outs
+    AnyTorchOptionalTensorType:$result0,
+    AnyTorchOptionalTensorType:$result1,
+    AnyTorchOptionalTensorType:$result2
+  );
+  let hasCustomAssemblyFormat = 1;
+  let extraClassDefinition = [{
+    ParseResult AtenMiopenBatchNormOp::parse(OpAsmParser &parser, OperationState &result) {
+      return parseDefaultTorchOp(parser, result, 8, 3);
+    }
+    void AtenMiopenBatchNormOp::print(OpAsmPrinter &printer) {
+      printDefaultTorchOp(printer, *this, 8, 3);
+    }
+  }];
+}
+
 def Torch_AtenBatchNormOp : Torch_Op<"aten.batch_norm", [
     AllowsTypeRefinement,
     HasValueSemantics,

lib/Dialect/Torch/Transforms/DecomposeComplexOps.cpp

@@ -8526,6 +8526,28 @@ class DecomposeAtenNativeGroupNormOp
 };
 } // namespace
 
+// Decompose aten.miopen_batch_norm to aten.native_batch_norm.
+// PyTorch dispatches to miopen_batch_norm on ROCm in some cases:
+// https://github.com/pytorch/pytorch/blob/2f0a6bd3f5efddf0eadf6252924ac4bd0276ee71/aten/src/ATen/native/Normalization.cpp#L626
+// We copy the inductor decomposition here which just lowers to
+// `aten.native_batch_norm` (inference mode returning empty tensors is an
+// optimization we can skip here):
+// https://github.com/pytorch/pytorch/blob/2f0a6bd3f5efddf0eadf6252924ac4bd0276ee71/torch/_inductor/decomposition.py#L888-L917
+namespace {
+class DecomposeAtenMiopenBatchNormOp
+    : public OpRewritePattern<AtenMiopenBatchNormOp> {
+  using OpRewritePattern::OpRewritePattern;
+  LogicalResult matchAndRewrite(AtenMiopenBatchNormOp op,
+                                PatternRewriter &rewriter) const override {
+    rewriter.replaceOpWithNewOp<AtenNativeBatchNormOp>(
+        op, op->getResultTypes(), op.getInput(), op.getWeight(), op.getBias(),
+        op.getRunningMean(), op.getRunningVar(), op.getTraining(),
+        op.getExponentialAverageFactor(), op.getEpsilon());
+    return success();
+  }
+};
+} // namespace
+
 namespace {
 class DecomposeAtenNativeBatchNormOp
     : public OpRewritePattern<AtenNativeBatchNormOp> {
@@ -13380,6 +13402,7 @@ class DecomposeComplexOpsPass
     addPatternIfTargetOpIsIllegal<DecomposeAtenGroupNormOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenNativeGroupNormOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenNativeBatchNormOp>(patterns);
+    addPatternIfTargetOpIsIllegal<DecomposeAtenMiopenBatchNormOp>(patterns);
     addPatternIfTargetOpIsIllegal<
         DecomposeAten_ConvolutionLikeOp<Aten_ConvolutionOp>>(patterns);
     addPatternIfTargetOpIsIllegal<

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

@@ -631,6 +631,9 @@ def emit_with_mutating_variants(key, **kwargs):
     emit(
         "aten::native_batch_norm : (Tensor, Tensor?, Tensor?, Tensor?, Tensor?, bool, float, float) -> (Tensor, Tensor, Tensor)"
     )
+    emit(
+        "aten::miopen_batch_norm : (Tensor, Tensor, Tensor?, Tensor?, Tensor?, bool, float, float) -> (Tensor, Tensor, Tensor)"
+    )
     emit(
         "aten::batch_norm : (Tensor, Tensor?, Tensor?, Tensor?, Tensor?, bool, float, float, bool) -> (Tensor)"
     )

test/Dialect/Torch/decompose-complex-ops.mlir

@@ -1013,3 +1013,15 @@ func.func @channel_shuffle(%arg0: !torch.vtensor<[1,8,4,4],f32>) -> !torch.vtens
   %0 = torch.aten.channel_shuffle %arg0, %int4 : !torch.vtensor<[1,8,4,4],f32>, !torch.int -> !torch.vtensor<[1,8,4,4],f32>
   return %0 : !torch.vtensor<[1,8,4,4],f32>
 }
+
+// -----
+
+// CHECK-LABEL: func.func @miopen_batch_norm(
+// CHECK-NOT: torch.aten.miopen_batch_norm
+func.func @miopen_batch_norm(%input: !torch.vtensor<[1,3,4,4],f32>, %weight: !torch.vtensor<[3],f32>, %bias: !torch.vtensor<[3],f32>, %mean: !torch.vtensor<[3],f32>, %var: !torch.vtensor<[3],f32>) -> (!torch.vtensor<[1,3,4,4],f32>, !torch.vtensor<[3],f32>, !torch.vtensor<[3],f32>) {
+  %false = torch.constant.bool false
+  %momentum = torch.constant.float 1.000000e-01
+  %eps = torch.constant.float 1.000000e-05
+  %0:3 = torch.aten.miopen_batch_norm %input, %weight, %bias, %mean, %var, %false, %momentum, %eps : !torch.vtensor<[1,3,4,4],f32>, !torch.vtensor<[3],f32>, !torch.vtensor<[3],f32>, !torch.vtensor<[3],f32>, !torch.vtensor<[3],f32>, !torch.bool, !torch.float, !torch.float -> !torch.vtensor<[1,3,4,4],f32>, !torch.vtensor<[3],f32>, !torch.vtensor<[3],f32>
+  return %0#0, %0#1, %0#2 : !torch.vtensor<[1,3,4,4],f32>, !torch.vtensor<[3],f32>, !torch.vtensor<[3],f32>
+}