[onnx] Lowerings from onnx.tan (#2642)

rsuderman · web-flow · commit 11cc92d4ab41 · 2023-12-20T10:09:39.000-08:00
Started work on the `tan` lowerings for ONNX to Torch. Uses `sin` and
`cos` to represent a `tan`.
diff --git a/include/torch-mlir/Dialect/Torch/IR/GeneratedTorchOps.td b/include/torch-mlir/Dialect/Torch/IR/GeneratedTorchOps.td
@@ -1066,6 +1066,51 @@ def Torch_AtenAcos_Op : Torch_Op<"aten.acos_", [
   }];
 }
 
+def Torch_AtenTanOp : Torch_Op<"aten.tan", [
+    AllowsTypeRefinement,
+    HasValueSemantics,
+    ReadOnly
+  ]> {
+  let summary = "Generated op for `aten::tan : (Tensor) -> (Tensor)`";
+  let arguments = (ins
+    AnyTorchTensorType:$self
+  );
+  let results = (outs
+    AnyTorchTensorType:$result
+  );
+  let hasCustomAssemblyFormat = 1;
+  let extraClassDefinition = [{
+    ParseResult AtenTanOp::parse(OpAsmParser &parser, OperationState &result) {
+      return parseDefaultTorchOp(parser, result, 1, 1);
+    }
+    void AtenTanOp::print(OpAsmPrinter &printer) {
+      printDefaultTorchOp(printer, *this, 1, 1);
+    }
+  }];
+}
+
+def Torch_AtenTan_Op : Torch_Op<"aten.tan_", [
+    IsTrailingUnderscoreInplaceVariant,
+    AllowsTypeRefinement
+  ]> {
+  let summary = "Generated op for `aten::tan_ : (Tensor) -> (Tensor)`";
+  let arguments = (ins
+    Torch_NonValueTensorType:$self
+  );
+  let results = (outs
+    Torch_NonValueTensorType:$result
+  );
+  let hasCustomAssemblyFormat = 1;
+  let extraClassDefinition = [{
+    ParseResult AtenTan_Op::parse(OpAsmParser &parser, OperationState &result) {
+      return parseDefaultTorchOp(parser, result, 1, 1);
+    }
+    void AtenTan_Op::print(OpAsmPrinter &printer) {
+      printDefaultTorchOp(printer, *this, 1, 1);
+    }
+  }];
+}
+
 def Torch_AtenAtanOp : Torch_Op<"aten.atan", [
     AllowsTypeRefinement,
     HasValueSemantics,
diff --git a/lib/Conversion/TorchOnnxToTorch/DefaultDomainQtoZ.cpp b/lib/Conversion/TorchOnnxToTorch/DefaultDomainQtoZ.cpp
@@ -794,7 +794,19 @@ void mlir::torch::onnx_c::populateDefaultDomainQtoZ(
                       binder.op, resultType, operand);
                   return success();
                 });
-  
+
+  patterns.onOp("Tan", 7,
+                [](OpBinder binder, ConversionPatternRewriter &rewriter) {
+                  Torch::ValueTensorType resultType;
+                  Value operand;
+                  if (binder.tensorOperand(operand) ||
+                      binder.tensorResultType(resultType))
+                    return failure();
+                  rewriter.replaceOpWithNewOp<Torch::AtenTanOp>(
+                      binder.op, resultType, operand);
+                  return success();
+                });
+
   patterns.onOp(
       "Transpose", 13,
       [](OpBinder binder, ConversionPatternRewriter &rewriter) {
diff --git a/lib/Conversion/TorchToLinalg/Uncategorized.cpp b/lib/Conversion/TorchToLinalg/Uncategorized.cpp
@@ -216,6 +216,10 @@ static Value createLinalgPayloadCalculationForElementwiseOp(
     return b.create<math::FloorOp>(loc, payloadArgs[0]);
   if (isa<AtenCeilOp>(op))
     return b.create<math::CeilOp>(loc, payloadArgs[0]);
+  if (isa<AtenTanOp>(op)) {
+    return createCalculationForMathOpWithDtypeConversion<math::TanOp>(
+        b, converter, payloadArgs[0], op);
+  }
   if (isa<AtenTanhOp>(op)) {
     return createCalculationForMathOpWithDtypeConversion<math::TanhOp>(
         b, converter, payloadArgs[0], op);
@@ -1319,15 +1323,15 @@ class ConvertElementwiseOp : public ConversionPattern {
   LogicalResult
   matchAndRewrite(Operation *op, ArrayRef<Value> operands,
                   ConversionPatternRewriter &rewriter) const override {
-    if (!isa<AtenTanhOp, AtenSinhOp, AtenCoshOp, AtenReluOp, AtenPreluOp,
-             AtenGeluOp, AtenGeluBackwardOp, AtenAddTensorOp, AtenMulTensorOp,
-             AtenDivTensorOp, AtenDivTensorModeOp, AtenSubTensorOp, AtenAtan2Op,
-             AtenLerpTensorOp, AtenSigmoidOp, AtenExpOp, AtenExpm1Op,
-             AtenMinimumOp, AtenMaximumOp, AtenToDtypeOp, AtenClampOp,
-             AtenClampTensorOp, AtenRsubScalarOp, AtenMulScalarOp, AtenLogOp,
-             AtenErfOp, AtenSqrtOp, AtenFloorOp, AtenPowScalarOp,
-             AtenPowTensorScalarOp, AtenPowTensorTensorOp, AtenLog2Op,
-             AtenLog10Op, AtenLog1pOp, AtenRsqrtOp, AtenDivScalarOp,
+    if (!isa<AtenTanOp, AtenTanhOp, AtenSinhOp, AtenCoshOp, AtenReluOp,
+             AtenPreluOp, AtenGeluOp, AtenGeluBackwardOp, AtenAddTensorOp,
+             AtenMulTensorOp, AtenDivTensorOp, AtenDivTensorModeOp,
+             AtenSubTensorOp, AtenAtan2Op, AtenLerpTensorOp, AtenSigmoidOp,
+             AtenExpOp, AtenExpm1Op, AtenMinimumOp, AtenMaximumOp,
+             AtenToDtypeOp, AtenClampOp, AtenClampTensorOp, AtenRsubScalarOp,
+             AtenMulScalarOp, AtenLogOp, AtenErfOp, AtenSqrtOp, AtenFloorOp,
+             AtenPowScalarOp, AtenPowTensorScalarOp, AtenPowTensorTensorOp,
+             AtenLog2Op, AtenLog10Op, AtenLog1pOp, AtenRsqrtOp, AtenDivScalarOp,
              AtenRemainderScalarOp, AtenAbsOp, AtenReciprocalOp,
              AtenBitwiseAndTensorOp, AtenBitwiseAndScalarOp,
              AtenBitwiseOrTensorOp, AtenBitwiseXorTensorOp,
@@ -1972,7 +1976,7 @@ void mlir::torch::torch_to_linalg::populateUncategorizedPatternsAndLegality(
     ConversionTarget &target) {
   MLIRContext *context = patterns.getContext();
   target.addIllegalOp<
-      AtenTanhOp, AtenSinhOp, AtenCoshOp, AtenReluOp, AtenGeluOp,
+      AtenTanOp, AtenTanhOp, AtenSinhOp, AtenCoshOp, AtenReluOp, AtenGeluOp,
       AtenGeluBackwardOp, AtenAddTensorOp, AtenMulTensorOp, AtenDivTensorOp,
       AtenDivTensorModeOp, AtenSubTensorOp, AtenLerpTensorOp, AtenSigmoidOp,
       AtenMinimumOp, AtenAtan2Op, AtenMaximumOp, AtenToDtypeOp, AtenClampOp,
diff --git a/lib/Dialect/Torch/Transforms/AbstractInterpLibrary.cpp b/lib/Dialect/Torch/Transforms/AbstractInterpLibrary.cpp
@@ -6238,6 +6238,10 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    %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.tan\"(%arg0: !torch.list<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"
+"  }\n"
 "  func.func @\"__torch_mlir_shape_fn.aten.atan\"(%arg0: !torch.list<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"
@@ -11396,6 +11400,17 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    }\n"
 "    return %4 : !torch.tuple<int, int>\n"
 "  }\n"
+"  func.func @\"__torch_mlir_dtype_fn.aten.tan\"(%arg0: !torch.tuple<int, int>) -> !torch.int {\n"
+"    %int6 = torch.constant.int 6\n"
+"    %0:2 = torch.prim.TupleUnpack %arg0 : !torch.tuple<int, int> -> !torch.int, !torch.int\n"
+"    %1 = call @__torch__.torch_mlir.jit_ir_importer.build_tools.library_generator.is_integer_dtype(%0#1) : (!torch.int) -> !torch.bool\n"
+"    %2 = torch.prim.If %1 -> (!torch.int) {\n"
+"      torch.prim.If.yield %int6 : !torch.int\n"
+"    } else {\n"
+"      torch.prim.If.yield %0#1 : !torch.int\n"
+"    }\n"
+"    return %2 : !torch.int\n"
+"  }\n"
 "  func.func @\"__torch_mlir_dtype_fn.aten.atan2\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.tuple<int, int>) -> !torch.int {\n"
 "    %int6 = torch.constant.int 6\n"
 "    %0:2 = torch.prim.TupleUnpack %arg0 : !torch.tuple<int, int> -> !torch.int, !torch.int\n"
diff --git a/projects/pt1/python/torch_mlir/jit_ir_importer/build_tools/abstract_interp_lib_gen.py b/projects/pt1/python/torch_mlir/jit_ir_importer/build_tools/abstract_interp_lib_gen.py
@@ -59,6 +59,9 @@ def aten〇triu〡shape(self: List[int], diagonal: int = 0) -> List[int]:
 def aten〇tril〡shape(self: List[int], diagonal: int = 0) -> List[int]:
     return upstream_shape_functions.unary(self)
 
+def aten〇tan〡shape(self: List[int]) -> List[int]:
+    return upstream_shape_functions.unary(self)
+
 def aten〇atan〡shape(self: List[int]) -> List[int]:
     return upstream_shape_functions.unary(self)
 
@@ -3721,6 +3724,13 @@ def aten〇var_mean〡dtype(self_rank_dtype: Tuple[int, int], unbiased: bool = T
         return torch.float64, self_dtype
     return self_dtype, self_dtype
 
+@check_dtype_function(_check_tensors_with_the_same_dtype(num_of_tensors=1))
+def aten〇tan〡dtype(self_rank_dtype: Tuple[int, int]) -> int:
+    self_rank, self_dtype = self_rank_dtype
+    if is_integer_dtype(self_dtype):
+        return torch.float32
+    return self_dtype
+
 @check_dtype_function(_check_two_tensor_op())
 def aten〇atan2〡dtype(self_rank_dtype: Tuple[int, int], other_rank_dtype: Tuple[int, int]) -> int:
     self_rank, self_dtype = self_rank_dtype
diff --git a/projects/pt1/python/torch_mlir/jit_ir_importer/build_tools/torch_ods_gen.py b/projects/pt1/python/torch_mlir/jit_ir_importer/build_tools/torch_ods_gen.py
@@ -278,6 +278,7 @@ def emit_with_mutating_variants(key, **kwargs):
             "aten::expm1 : (Tensor) -> (Tensor)",
             "aten::cos : (Tensor) -> (Tensor)",
             "aten::acos : (Tensor) -> (Tensor)",
+            "aten::tan : (Tensor) -> (Tensor)",
             "aten::atan : (Tensor) -> (Tensor)",
             "aten::atan2 : (Tensor, Tensor) -> (Tensor)",
             "aten::neg : (Tensor) -> (Tensor)",
diff --git a/projects/pt1/python/torch_mlir_e2e_test/test_suite/elementwise.py b/projects/pt1/python/torch_mlir_e2e_test/test_suite/elementwise.py
@@ -3009,6 +3009,46 @@ def ElementwiseAcosIntModule_basic(module, tu: TestUtils):
 
 # ==============================================================================
 
+class ElementwiseTanModule(torch.nn.Module):
+
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args([
+        None,
+        ([-1, -1], torch.float32, True),
+    ])
+    def forward(self, a):
+        return torch.tan(a)
+
+
+@register_test_case(module_factory=lambda: ElementwiseTanModule())
+def ElementwiseTanModule_basic(module, tu: TestUtils):
+    module.forward(tu.rand(3, 4))
+
+# ==============================================================================
+
+class ElementwiseTanIntModule(torch.nn.Module):
+
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args([
+        None,
+        ([-1, -1], torch.int32, True),
+    ])
+    def forward(self, a):
+        return torch.tan(a)
+
+
+@register_test_case(module_factory=lambda: ElementwiseTanIntModule())
+def ElementwiseTanIntModule_basic(module, tu: TestUtils):
+    module.forward(tu.randint(3, 4, low=1, high=10).to(torch.int32))
+
+# ==============================================================================
+
 class ElementwiseNegModule(torch.nn.Module):
 
     def __init__(self):
diff --git a/test/Conversion/TorchOnnxToTorch/simple_ops_q_to_z.mlir b/test/Conversion/TorchOnnxToTorch/simple_ops_q_to_z.mlir
@@ -795,6 +795,15 @@ func.func @test_sinh_example(%arg0: !torch.vtensor<[3],f32>) -> !torch.vtensor<[
 
 // -----
 
+// CHECK-LABEL: func.func @test_tan
+func.func @test_tan(%arg0: !torch.vtensor<[3,4,5],f32>) -> !torch.vtensor<[3,4,5],f32> attributes {torch.onnx_meta.ir_version = 3 : si64, torch.onnx_meta.opset_version = 7 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
+  // CHECK: %[[TAN:.+]] = torch.aten.tan %arg0
+  %0 = torch.operator "onnx.Tan"(%arg0) : (!torch.vtensor<[3,4,5],f32>) -> !torch.vtensor<[3,4,5],f32>
+  return %0 : !torch.vtensor<[3,4,5],f32>
+}
+
+// -----
+
 // CHECK-LABEL: func.func @test_transpose_default
 func.func @test_transpose_default(%arg0: !torch.vtensor<[2,3,4],f32>) -> !torch.vtensor<[4,3,2],f32> attributes {torch.onnx_meta.ir_version = 7 : si64, torch.onnx_meta.opset_version = 13 : si64} {
   // CHECK-DAG: %[[I0:.+]] = torch.constant.int 0
