linalg: lower `aten.triu` op to `linalg.generic` (#965)

Prior to this patch, the torch dialect included `AtenTriuOp` for
computing the upper triangular part of the input matrix, but there was
no code for lowering the op to the linalg dialect.

This patch adds code to generate a `linalg.generic` operation that
compares indices (computed using `linalg.index`) to choose between zero
or the original value (using `arith.select`).  The lowering fails if the
number of dimensions are less than two.  This patch also adds a few
end-to-end tests.

lib/Conversion/TorchToLinalg/Uncategorized.cpp

@@ -857,6 +857,33 @@ static Value createLinalgPayloadCalculationForElementwiseOp(
     return b.create<arith::SelectOp>(loc, mask, fillValue, input);
   }
 
+  if (auto triu = dyn_cast<AtenTriuOp>(op)) {
+    // Check if the rank of the input tensor is valid.
+    AtenTriuOp::Adaptor adaptor(operands);
+    auto inputType = adaptor.self().getType().cast<RankedTensorType>();
+    uint64_t inputRank = inputType.getRank();
+    if (inputRank < 2) {
+      triu.emitError("too few dimensions to compute triangular part of matrix");
+      return nullptr;
+    }
+
+    // Use the indices of the two innermost dimensions.
+    auto rowIndex = b.create<linalg::IndexOp>(loc, inputRank - 2);
+    Value rowIndexI64 = castIndexToInt64(b, loc, rowIndex);
+    auto colIndex = b.create<linalg::IndexOp>(loc, inputRank - 1);
+    Value colIndexI64 = castIndexToInt64(b, loc, colIndex);
+
+    // columnIndex >= rowIndex + diagonal?
+    auto sum = b.create<arith::AddIOp>(loc, rowIndexI64, adaptor.diagonal());
+    auto pred = b.create<arith::CmpIOp>(loc, arith::CmpIPredicate::sge,
+                                        colIndexI64, sum);
+
+    Value scalar = payloadArgs[0];
+    Type elementType = inputType.getElementType();
+    Value zero = getConstant(b, loc, 0, elementType);
+    return b.create<arith::SelectOp>(loc, pred, scalar, zero);
+  }
+
   op->emitError("unimplemented lowering in "
                 "createLinalgPayloadCalculationForElementwiseOp");
   return nullptr;
@@ -902,7 +929,7 @@ public:
              AtenEqTensorOp, AtenLtTensorOp, AtenSubScalarOp, AtenAddScalarOp,
              AtenThresholdOp, AtenThresholdBackwardOp, AtenCloneOp, AtenSinOp,
              AtenCosOp, AtenNeScalarOp, AtenNegOp, AtenMaskedFillScalarOp,
-             AtenLogicalOrOp>(op))
+             AtenLogicalOrOp, AtenTriuOp>(op))
       return rewriter.notifyMatchFailure(op, "not a supported elementwise op");
 
     if (failed(verifyLinalgCompatibleTypes(op, rewriter)))
@@ -1640,7 +1667,7 @@ void mlir::torch::torch_to_linalg::populateUncategorizedPatternsAndLegality(
       AtenLtScalarOp, AtenLeScalarOp, AtenWhereSelfOp, AtenGtTensorOp,
       AtenEqTensorOp, AtenLtTensorOp, AtenThresholdOp, AtenThresholdBackwardOp,
       AtenCloneOp, AtenSinOp, AtenCosOp, AtenNeScalarOp, AtenMaskedFillScalarOp,
-      AtenLogicalOrOp>();
+      AtenLogicalOrOp, AtenTriuOp>();
   patterns.add<ConvertElementwiseOp>(typeConverter, context);
   target.addIllegalOp<AtenNllLossForwardOp>();
   patterns.add<ConvertAtenDetachOp>(typeConverter, context);

lib/Dialect/Torch/Transforms/RefineTypes.cpp

@@ -642,7 +642,7 @@ ChangeResult TypeAnalyzer::visitOperation(
           AtenZero_Op, AtenIndexTensorOp, ValsemVariantAtenIndexPutImplOp,
           AtenIndexPutOp, ValsemVariantAtenCopyOp, AtenZeroFunctionalOp,
           AtenIndexPutHackedTwinOp, AtenMaskedFillScalarOp, AtenFlipOp,
-          PrimAbsScalarOp, AtenNumpyTOp>(op)) {
+          PrimAbsScalarOp, AtenNumpyTOp, AtenTriuOp>(op)) {
     return incorporateKnowledge(op->getResult(0), operands[0]->getValue());
   }
 

lib/Dialect/Torch/Transforms/ShapeLibrary.cpp

@@ -5297,6 +5297,10 @@ module {
     %3 = torch.prim.ListConstruct %1, %2 : (!torch.int, !torch.int) -> !torch.list<int>
     return %3 : !torch.list<int>
   }
+  func.func @"__torch_mlir_shape_fn.aten.triu"(%arg0: !torch.list<int>, %arg1: !torch.int) -> !torch.list<int> {
+    %0 = call @__torch__.torch.jit._shape_functions.unary(%arg0) : (!torch.list<int>) -> !torch.list<int>
+    return %0 : !torch.list<int>
+  }
   func.func @"__torch_mlir_shape_fn.aten.tanh"(%arg0: !torch.list<int>) -> !torch.list<int> {
     %0 = call @__torch__.torch.jit._shape_functions.unary(%arg0) : (!torch.list<int>) -> !torch.list<int>
     return %0 : !torch.list<int>

python/torch_mlir/dialects/torch/importer/jit_ir/build_tools/shape_lib_gen.py

@@ -286,6 +286,9 @@ def not_present_in_registry(f):
 # Shape functions
 # ==============================================================================
 
+def aten〇triu(self: List[int], diagonal: int = 0) -> List[int]:
+    return upstream_shape_functions.unary(self)
+
 def aten〇tanh(self: List[int]) -> List[int]:
     return upstream_shape_functions.unary(self)
 

python/torch_mlir_e2e_test/test_suite/elementwise.py

@@ -1806,3 +1806,69 @@ class ElementwiseAtenFloorDivideBroadcastModule(torch.nn.Module):
     module_factory=lambda: ElementwiseAtenFloorDivideBroadcastModule())
 def ElementwiseAtenFloorDivideBroadcastModule_basic(module, tu: TestUtils):
     module.forward(tu.rand(3), tu.rand(4, 3))
+
+
+# ==============================================================================
+
+
+class AtenTriuModule(torch.nn.Module):
+
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args([
+        None,
+        ([-1, -1, -1, -1, -1], torch.float32, True),
+    ])
+    def forward(self, x):
+        return torch.triu(x)
+
+
+@register_test_case(module_factory=lambda: AtenTriuModule())
+def AtenTriuModule_basic(module, tu: TestUtils):
+    module.forward(tu.rand(5, 8, 3, 4, 3))
+
+
+# ==============================================================================
+
+
+class AtenTriuWithPosDiagonalModule(torch.nn.Module):
+
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args([
+        None,
+        ([-1, -1, -1], torch.float32, True),
+    ])
+    def forward(self, x):
+        return torch.triu(x, diagonal=2)
+
+
+@register_test_case(module_factory=lambda: AtenTriuWithPosDiagonalModule())
+def AtenTriuWithPosDiagonalModule_basic(module, tu: TestUtils):
+    module.forward(tu.rand(9, 4, 3))
+
+
+# ==============================================================================
+
+
+class AtenTriuWithNegDiagonalModule(torch.nn.Module):
+
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args([
+        None,
+        ([-1, -1, -1, -1], torch.float32, True),
+    ])
+    def forward(self, x):
+        return torch.triu(x, diagonal=-4)
+
+
+@register_test_case(module_factory=lambda: AtenTriuWithNegDiagonalModule())
+def AtenTriuWithNegDiagonalModule_basic(module, tu: TestUtils):
+    module.forward(tu.rand(3, 1, 5, 9))

test/Conversion/TorchToLinalg/elementwise.mlir

@@ -84,3 +84,13 @@ func.func @elementwise$static_1(%arg0: !torch.vtensor<[?],f32>, %arg1: !torch.vt
   %1 = torch.aten.mul.Tensor %arg0, %arg1 : !torch.vtensor<[?],f32>, !torch.vtensor<[1],f32> -> !torch.vtensor<[?],f32>
   return %1 : !torch.vtensor<[?],f32>
 }
+
+// -----
+
+func.func @insufficient_dims_for_triu(%arg0: !torch.vtensor<[?],f32>) -> !torch.vtensor<[?],f32> {
+  %int0 = torch.constant.int 0
+  // expected-error@+2 {{failed to legalize operation 'torch.aten.triu' that was explicitly marked illegal}}
+  // expected-error@+1 {{too few dimensions to compute triangular part of matrix}}
+  %0 = torch.aten.triu %arg0, %int0 : !torch.vtensor<[?],f32>, !torch.int -> !torch.vtensor<[?],f32>
+  return %0 : !torch.vtensor<[?],f32>
+}