[Torch] Add aten.diag decomposition

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

@@ -10251,6 +10251,30 @@ def Torch_AtenDiagEmbedOp : Torch_Op<"aten.diag_embed", [
   }];
 }
 
+def Torch_AtenDiagOp : Torch_Op<"aten.diag", [
+    AllowsTypeRefinement,
+    HasValueSemantics,
+    ReadOnly
+  ]> {
+  let summary = "Generated op for `aten::diag : (Tensor, int) -> (Tensor)`";
+  let arguments = (ins
+    AnyTorchTensorType:$self,
+    Torch_IntType:$diagonal
+  );
+  let results = (outs
+    AnyTorchOptionalTensorType:$result
+  );
+  let hasCustomAssemblyFormat = 1;
+  let extraClassDefinition = [{
+    ParseResult AtenDiagOp::parse(OpAsmParser &parser, OperationState &result) {
+      return parseDefaultTorchOp(parser, result, 2, 1);
+    }
+    void AtenDiagOp::print(OpAsmPrinter &printer) {
+      printDefaultTorchOp(printer, *this, 2, 1);
+    }
+  }];
+}
+
 def Torch_Aten_WeightNormInterfaceOp : Torch_Op<"aten._weight_norm_interface", [
     AllowsTypeRefinement,
     HasValueSemantics,

lib/Dialect/Torch/Transforms/AbstractInterpLibrary.cpp

@@ -7025,6 +7025,58 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    %18 = torch.aten.append.t %7, %17 : !torch.list<int>, !torch.int -> !torch.list<int>\n"
 "    return %7 : !torch.list<int>\n"
 "  }\n"
+"  func.func @\"__torch_mlir_shape_fn.aten.diag\"(%arg0: !torch.list<int>, %arg1: !torch.int) -> !torch.list<int> {\n"
+"    %none = torch.constant.none\n"
+"    %str = torch.constant.str \"AssertionError: input must be 1D or 2D\"\n"
+"    %true = torch.constant.bool true\n"
+"    %int1 = torch.constant.int 1\n"
+"    %int2 = torch.constant.int 2\n"
+"    %int0 = torch.constant.int 0\n"
+"    %0 = torch.aten.len.t %arg0 : !torch.list<int> -> !torch.int\n"
+"    %1 = torch.aten.eq.int %0, %int1 : !torch.int, !torch.int -> !torch.bool\n"
+"    %2 = torch.prim.If %1 -> (!torch.bool) {\n"
+"      torch.prim.If.yield %true : !torch.bool\n"
+"    } else {\n"
+"      %6 = torch.aten.len.t %arg0 : !torch.list<int> -> !torch.int\n"
+"      %7 = torch.aten.eq.int %6, %int2 : !torch.int, !torch.int -> !torch.bool\n"
+"      torch.prim.If.yield %7 : !torch.bool\n"
+"    }\n"
+"    torch.prim.If %2 -> () {\n"
+"      torch.prim.If.yield\n"
+"    } else {\n"
+"      torch.prim.RaiseException %str, %none : !torch.str, !torch.none\n"
+"      torch.prim.If.yield\n"
+"    }\n"
+"    %3 = torch.aten.len.t %arg0 : !torch.list<int> -> !torch.int\n"
+"    %4 = torch.aten.eq.int %3, %int1 : !torch.int, !torch.int -> !torch.bool\n"
+"    %5 = torch.prim.If %4 -> (!torch.list<int>) {\n"
+"      %6 = torch.aten.__getitem__.t %arg0, %int0 : !torch.list<int>, !torch.int -> !torch.int\n"
+"      %7 = torch.operator \"prim.abs.int\"(%arg1) : (!torch.int) -> !torch.int \n"
+"      %8 = torch.aten.add.int %6, %7 : !torch.int, !torch.int -> !torch.int\n"
+"      %9 = torch.prim.ListConstruct %8, %8 : (!torch.int, !torch.int) -> !torch.list<int>\n"
+"      torch.prim.If.yield %9 : !torch.list<int>\n"
+"    } else {\n"
+"      %6 = torch.aten.__getitem__.t %arg0, %int0 : !torch.list<int>, !torch.int -> !torch.int\n"
+"      %7 = torch.aten.__getitem__.t %arg0, %int1 : !torch.list<int>, !torch.int -> !torch.int\n"
+"      %8 = torch.aten.sub.int %7, %arg1 : !torch.int, !torch.int -> !torch.int\n"
+"      %9 = torch.prim.min.int %6, %8 : !torch.int, !torch.int -> !torch.int\n"
+"      %10 = torch.prim.max.int %9, %int0 : !torch.int, !torch.int -> !torch.int\n"
+"      %11 = torch.aten.lt.int %arg1, %int0 : !torch.int, !torch.int -> !torch.bool\n"
+"      %12 = torch.prim.If %11 -> (!torch.int) {\n"
+"        %14 = torch.aten.__getitem__.t %arg0, %int0 : !torch.list<int>, !torch.int -> !torch.int\n"
+"        %15 = torch.aten.add.int %14, %arg1 : !torch.int, !torch.int -> !torch.int\n"
+"        %16 = torch.aten.__getitem__.t %arg0, %int1 : !torch.list<int>, !torch.int -> !torch.int\n"
+"        %17 = torch.prim.min.int %15, %16 : !torch.int, !torch.int -> !torch.int\n"
+"        %18 = torch.prim.max.int %17, %int0 : !torch.int, !torch.int -> !torch.int\n"
+"        torch.prim.If.yield %18 : !torch.int\n"
+"      } else {\n"
+"        torch.prim.If.yield %10 : !torch.int\n"
+"      }\n"
+"      %13 = torch.prim.ListConstruct %12 : (!torch.int) -> !torch.list<int>\n"
+"      torch.prim.If.yield %13 : !torch.list<int>\n"
+"    }\n"
+"    return %5 : !torch.list<int>\n"
+"  }\n"
 "  func.func @\"__torch_mlir_shape_fn.aten.fake_quantize_per_tensor_affine\"(%arg0: !torch.list<int>, %arg1: !torch.float, %arg2: !torch.int, %arg3: !torch.int, %arg4: !torch.int) -> !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"
@@ -14301,6 +14353,10 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    %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.diag\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.int) -> !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.uniform\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.float, %arg2: !torch.float, %arg3: !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

@@ -10354,6 +10354,48 @@ class DecomposeAtenL1LossOp : public OpRewritePattern<AtenL1LossOp> {
 };
 } // namespace
 
+namespace {
+// `aten.diag` is a rank-dependent alias (aten/src/ATen/native/TensorShape.cpp):
+// a 1-D input builds a 2-D matrix with the input on the `diagonal`-th diagonal
+// (aten.diag_embed); a 2-D input extracts the `diagonal`-th diagonal as a 1-D
+// tensor (aten.diagonal).
+class DecomposeAtenDiagOp : public OpRewritePattern<AtenDiagOp> {
+public:
+  using OpRewritePattern::OpRewritePattern;
+  LogicalResult matchAndRewrite(AtenDiagOp op,
+                                PatternRewriter &rewriter) const override {
+    Location loc = op.getLoc();
+    Value self = op.getSelf();
+    std::optional<unsigned> maybeRank = getTensorRank(self);
+    if (!maybeRank)
+      return rewriter.notifyMatchFailure(op, "expected input to have a rank");
+    unsigned rank = *maybeRank;
+    if (rank != 1 && rank != 2)
+      return rewriter.notifyMatchFailure(op,
+                                         "expected input to be rank 1 or 2");
+
+    Value zero =
+        ConstantIntOp::create(rewriter, loc, rewriter.getI64IntegerAttr(0));
+    Value one =
+        ConstantIntOp::create(rewriter, loc, rewriter.getI64IntegerAttr(1));
+    Value diagonal = op.getDiagonal();
+
+    if (rank == 1) {
+      Value embed = AtenDiagEmbedOp::create(rewriter, loc, op.getType(), self,
+                                            /*offset=*/diagonal, /*dim1=*/zero,
+                                            /*dim2=*/one);
+      rewriter.replaceOp(op, embed);
+      return success();
+    }
+    Value diag = AtenDiagonalOp::create(rewriter, loc, op.getType(), self,
+                                        /*offset=*/diagonal, /*dim1=*/zero,
+                                        /*dim2=*/one);
+    rewriter.replaceOp(op, diag);
+    return success();
+  }
+};
+} // namespace
+
 namespace {
 // Decompose `aten.norm.ScalarOpt_dim` op to `aten.linalg_vector_norm` op
 class DecomposeAtenNormScalarOptDimOp
@@ -13620,6 +13662,7 @@ class DecomposeComplexOpsPass
     addPatternIfTargetOpIsIllegal<DecomposeAtenLiftFreshCopyOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenMseLossOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenL1LossOp>(patterns);
+    addPatternIfTargetOpIsIllegal<DecomposeAtenDiagOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenNormScalarOptDimOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenRandintOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenRandintLowOp>(patterns);

lib/Dialect/Torch/Transforms/LowerToBackendContract.cpp

@@ -544,6 +544,7 @@ static void markDecomposedOpsAsIllegal(MLIRContext *context,
   target.addIllegalOp<AtenLerpTensorOp>();
   target.addIllegalOp<AtenMseLossOp>();
   target.addIllegalOp<AtenL1LossOp>();
+  target.addIllegalOp<AtenDiagOp>();
   target.addIllegalOp<AtenPoissonNllLossOp>();
   target.addIllegalOp<AtenRandintLowOp>();
   target.addIllegalOp<AtenRandintOp>();

projects/pt1/e2e_testing/xfail_sets.py

@@ -67,6 +67,9 @@
     }
 
 LINALG_CRASHING_SET = {
+    # aten.diag with a nonzero offset lowers to aten.diag_embed, whose linalg
+    # lowering does an out-of-bounds extract (same crash as AtenDiagEmbedOffsetDiag).
+    "Diag1DOffsetModule_basic",
     # Runtime op verification: Out of bounds access
     "AtenDiagEmbedNegOffsetDiag_basic",
     "AtenDiagEmbedNonDefault4DDiag_basic",
@@ -550,6 +553,12 @@
 }
 
 FX_IMPORTER_STABLEHLO_XFAIL_SET = {
+    # aten.diag decomposes to aten.diag_embed / aten.diagonal, which are xfail
+    # on the stablehlo backend (same as AtenDiagEmbed*/DiagonalModule).
+    "Diag1DModule_basic",
+    "Diag1DOffsetModule_basic",
+    "Diag2DModule_basic",
+    "Diag2DNegativeOffsetModule_basic",
     "ArgsortTensor_basic",
     "ArgsortTensorInteger_basic",
     "AddFloatIntModule_basic",
@@ -2779,6 +2788,11 @@
 }
 
 ONNX_XFAIL_SET = {
+    # aten.diag has no ONNX import path, so these fail at import time.
+    "Diag1DModule_basic",
+    "Diag1DOffsetModule_basic",
+    "Diag2DModule_basic",
+    "Diag2DNegativeOffsetModule_basic",
     "ToDtypeIntFromFloatModule_basic",
     # This test is expected to time out
     "TimeOutModule_basic",
@@ -4101,6 +4115,11 @@
 }
 
 ONNX_TOSA_XFAIL_SET = {
+    # aten.diag has no ONNX import path, so these fail at import time.
+    "Diag1DModule_basic",
+    "Diag1DOffsetModule_basic",
+    "Diag2DModule_basic",
+    "Diag2DNegativeOffsetModule_basic",
     "AtenFftRfft2DLastDim_basic",
     "AtenFftRfft2DMiddleDim_basic",
     "AtenStftCenter1D_basic",

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

@@ -135,6 +135,26 @@ def aten〇diagonal〡shape(self: List[int], offset: int = 0, dim1: int = 0, dim
 
     return diagonal
 
+@check_shape_function([
+    Invocation(TensorOfShape(3)), # 1-D: builds a square matrix.
+    Invocation(TensorOfShape(3), diagonal=2), # 1-D positive offset.
+    Invocation(TensorOfShape(3), diagonal=-2), # 1-D negative offset.
+    Invocation(TensorOfShape(3, 5)), # 2-D: extracts the main diagonal.
+    Invocation(TensorOfShape(3, 5), diagonal=1), # 2-D positive offset.
+    Invocation(TensorOfShape(3, 5), diagonal=-1), # 2-D negative offset.
+    Invocation(TensorOfShape(5, 3), diagonal=3), # 2-D empty result (offset past last column).
+    ErrorInvocation(TensorOfShape(2, 3, 4)), # Rank > 2 unsupported.
+])
+def aten〇diag〡shape(self: List[int], diagonal: int = 0) -> List[int]:
+    assert len(self) == 1 or len(self) == 2, "input must be 1D or 2D"
+    if len(self) == 1:
+        side = self[0] + abs(diagonal)
+        return [side, side]
+    diag_size = max(min(self[0], self[1] - diagonal), 0)
+    if diagonal < 0:
+        diag_size = max(min(self[0] + diagonal, self[1]), 0)
+    return [diag_size]
+
 def aten〇fake_quantize_per_tensor_affine〡shape(self: List[int], scale: float, zero_point: int, quant_min: int, quant_max: int) -> List[int]:
     return upstream_shape_functions.unary(self)
 
@@ -4065,6 +4085,11 @@ def aten〇diagonal〡dtype(self_rank_dtype: Tuple[int, int], offset: int = 0, d
     self_rank, self_dtype = self_rank_dtype
     return self_dtype
 
+@check_dtype_function(_check_tensors_with_the_same_dtype(num_of_tensors=1))
+def aten〇diag〡dtype(self_rank_dtype: Tuple[int, int], diagonal: int = 0) -> int:
+    self_rank, self_dtype = self_rank_dtype
+    return self_dtype
+
 @check_dtype_function(_check_tensors_with_the_same_dtype(num_of_tensors=1))
 def aten〇uniform〡dtype(self_rank_dtype: Tuple[int, int], from_: float = 0., to: float = 1., 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

@@ -815,6 +815,7 @@ def emit_with_mutating_variants(key, **kwargs):
         "aten::cosine_embedding_loss : (Tensor, Tensor, Tensor, float, int) -> (Tensor)"
     )
     emit("aten::diag_embed : (Tensor, int, int, int) -> (Tensor)")
+    emit("aten::diag : (Tensor, int) -> (Tensor)")
     emit("aten::_weight_norm_interface : (Tensor, Tensor, int) -> (Tensor, Tensor)")
     emit("aten::rot90 : (Tensor, int, int[]) -> (Tensor)", has_verifier=True)
     emit("aten::count_nonzero : (Tensor, int?) -> (Tensor)", has_verifier=True)

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

@@ -180,3 +180,95 @@ def forward(self, a):
 @register_test_case(module_factory=lambda: DiagonalWithDimsOffsetModule())
 def DiagonalModule_with_dims_and_offset(module, tu: TestUtils):
     module.forward(tu.rand(3, 4, 5))
+
+
+# ==============================================================================
+
+
+class Diag1DModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args(
+        [
+            None,
+            ([-1], torch.float32, True),
+        ]
+    )
+    def forward(self, a):
+        return torch.ops.aten.diag(a)
+
+
+@register_test_case(module_factory=lambda: Diag1DModule())
+def Diag1DModule_basic(module, tu: TestUtils):
+    module.forward(tu.rand(5))
+
+
+# ==============================================================================
+
+
+class Diag1DOffsetModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args(
+        [
+            None,
+            ([-1], torch.float32, True),
+        ]
+    )
+    def forward(self, a):
+        return torch.ops.aten.diag(a, diagonal=2)
+
+
+@register_test_case(module_factory=lambda: Diag1DOffsetModule())
+def Diag1DOffsetModule_basic(module, tu: TestUtils):
+    module.forward(tu.rand(5))
+
+
+# ==============================================================================
+
+
+class Diag2DModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args(
+        [
+            None,
+            ([-1, -1], torch.float32, True),
+        ]
+    )
+    def forward(self, a):
+        return torch.ops.aten.diag(a)
+
+
+@register_test_case(module_factory=lambda: Diag2DModule())
+def Diag2DModule_basic(module, tu: TestUtils):
+    module.forward(tu.rand(4, 6))
+
+
+# ==============================================================================
+
+
+class Diag2DNegativeOffsetModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args(
+        [
+            None,
+            ([-1, -1], torch.float32, True),
+        ]
+    )
+    def forward(self, a):
+        return torch.ops.aten.diag(a, diagonal=-1)
+
+
+@register_test_case(module_factory=lambda: Diag2DNegativeOffsetModule())
+def Diag2DNegativeOffsetModule_basic(module, tu: TestUtils):
+    module.forward(tu.rand(4, 6))