[Torch] support float_power and threshold ops (#3854)

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

@@ -5242,6 +5242,30 @@ def Torch_AtenPowScalarOp : Torch_Op<"aten.pow.Scalar", [
   }];
 }
 
+def Torch_AtenFloatPowerTensorTensorOp : Torch_Op<"aten.float_power.Tensor_Tensor", [
+    AllowsTypeRefinement,
+    HasValueSemantics,
+    ReadOnly
+  ]> {
+  let summary = "Generated op for `aten::float_power.Tensor_Tensor : (Tensor, Tensor) -> (Tensor)`";
+  let arguments = (ins
+    AnyTorchTensorType:$self,
+    AnyTorchTensorType:$exponent
+  );
+  let results = (outs
+    AnyTorchOptionalTensorType:$result
+  );
+  let hasCustomAssemblyFormat = 1;
+  let extraClassDefinition = [{
+    ParseResult AtenFloatPowerTensorTensorOp::parse(OpAsmParser &parser, OperationState &result) {
+      return parseDefaultTorchOp(parser, result, 2, 1);
+    }
+    void AtenFloatPowerTensorTensorOp::print(OpAsmPrinter &printer) {
+      printDefaultTorchOp(printer, *this, 2, 1);
+    }
+  }];
+}
+
 def Torch_AtenThresholdBackwardOp : Torch_Op<"aten.threshold_backward", [
     AllowsTypeRefinement,
     HasValueSemantics,

lib/Dialect/Torch/Transforms/DecomposeComplexOps.cpp

@@ -10439,6 +10439,63 @@ public:
 };
 } // namespace
 
+namespace {
+class DecomposeAtenThresholdOp : public OpRewritePattern<AtenThresholdOp> {
+public:
+  using OpRewritePattern<AtenThresholdOp>::OpRewritePattern;
+  LogicalResult matchAndRewrite(AtenThresholdOp op,
+                                PatternRewriter &rewriter) const override {
+    Location loc = op.getLoc();
+    Value self = op.getSelf();
+    auto selfType = dyn_cast<BaseTensorType>(self.getType());
+    if (!selfType || !selfType.hasSizes()) {
+      return rewriter.notifyMatchFailure(op,
+                                         "requires input is tensor with sizes");
+    }
+
+    Value threshold = op.getThreshold();
+    Value value = op.getValue();
+
+    auto comOp = rewriter.create<AtenGtScalarOp>(
+        loc,
+        selfType.getWithSizesAndDtype(selfType.getSizes(),
+                                      rewriter.getI1Type()),
+        self, threshold);
+
+    rewriter.replaceOpWithNewOp<AtenWhereScalarOtherOp>(op, op.getType(), comOp,
+                                                        self, value);
+    return success();
+  }
+};
+} // namespace
+
+namespace {
+class DecomposeAtenFloatPowerTensorTensorOp
+    : public OpRewritePattern<AtenFloatPowerTensorTensorOp> {
+public:
+  using OpRewritePattern<AtenFloatPowerTensorTensorOp>::OpRewritePattern;
+  LogicalResult matchAndRewrite(AtenFloatPowerTensorTensorOp op,
+                                PatternRewriter &rewriter) const override {
+    Location loc = op.getLoc();
+    Value self = op.getSelf();
+    Value exp = op.getExponent();
+
+    auto selfTy = dyn_cast<BaseTensorType>(self.getType());
+    if (!selfTy || !selfTy.hasDtype() || !selfTy.hasSizes()) {
+      return rewriter.notifyMatchFailure(
+          op, "requires input is tensor with dtype and sizes");
+    }
+
+    Value selfF64 =
+        convertTensorToDtype(rewriter, loc, self, rewriter.getF64Type());
+    rewriter.replaceOpWithNewOp<AtenPowTensorTensorOp>(op, op.getType(),
+                                                       selfF64, exp);
+
+    return success();
+  }
+};
+} // namespace
+
 namespace {
 class DecomposeComplexOpsPass
     : public DecomposeComplexOpsBase<DecomposeComplexOpsPass> {
@@ -10711,6 +10768,9 @@ public:
     addPatternIfTargetOpIsIllegal<DecomposeAtenConv1dOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenConv2dOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenConv3dOp>(patterns);
+    addPatternIfTargetOpIsIllegal<DecomposeAtenThresholdOp>(patterns);
+    addPatternIfTargetOpIsIllegal<DecomposeAtenFloatPowerTensorTensorOp>(
+        patterns);
 
     addPatternIfTargetOpIsIllegal<
         DecomposeAtenFMaxMinOp<AtenFmaxOp, AtenMaximumOp>>(patterns);

projects/pt1/e2e_testing/xfail_sets.py

@@ -886,13 +886,6 @@ FX_IMPORTER_STABLEHLO_XFAIL_SET = {
     "TensorToFloatZeroRank_basic",
     "TensorToFloat_basic",
     "TensorToInt_basic",
-    "Threshold1dFloatModule_basic",
-    "Threshold1dIntI32Module_basic",
-    "Threshold1dIntModule_basic",
-    "Threshold2dFloatModule_basic",
-    "Threshold2dIntModule_basic",
-    "Threshold3dFloatModule_basic",
-    "Threshold3dIntModule_basic",
     "ThresholdBackward1dFloatModule_basic",
     "ThresholdBackward1dIntModule_basic",
     "ThresholdBackward1dMixedModule_basic",
@@ -2717,6 +2710,7 @@ ONNX_XFAIL_SET = {
     "ElementwiseFminModule_basic",
     "ElementwiseFmaxModule_basic",
     "Exp2StaticModule_basic",
+    "FloatPowerTensorTensorStaticModule_basic",
     "MultinomialModule2D_basic",
     "MultinomialModule2D_F32",
     "PixelShuffleModuleStaticRank4Float32_basic",
@@ -2727,6 +2721,7 @@ ONNX_XFAIL_SET = {
     "SliceStaticComplexInputModule_basic",
     "StdCorrectionLargeInputModule_basic",
     "TupleModule_basic",
+    "ThresholdStaticModule_basic",
     "VarCorrectionLargeInputModule_basic",
     # Failure - incorrect shape
     "ArangeStartOutDtypeModule_basic",

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

@@ -499,6 +499,7 @@ def emit_ops(emitter_td: TextEmitter, registry: Registry):
     emit("aten::pow.Tensor_Scalar : (Tensor, Scalar) -> (Tensor)")
     emit("aten::pow.Tensor_Tensor : (Tensor, Tensor) -> (Tensor)")
     emit("aten::pow.Scalar : (Scalar, Tensor) -> (Tensor)")
+    emit("aten::float_power.Tensor_Tensor : (Tensor, Tensor) -> (Tensor)")
     emit("aten::threshold_backward : (Tensor, Tensor, Scalar) -> (Tensor)")
     emit("aten::floor_divide : (Tensor, Tensor) -> (Tensor)")
     emit("aten::softplus : (Tensor, Scalar, Scalar) -> (Tensor)")

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

@@ -491,6 +491,29 @@ def ElementwiseWhereSelfModule_basic(module, tu: TestUtils):
 # ==============================================================================
 
 
+class FloatPowerTensorTensorStaticModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args(
+        [
+            None,
+            ([3, 4], torch.float32, True),
+        ]
+    )
+    def forward(self, x):
+        return torch.ops.aten.float_power(x, torch.tensor(2))
+
+
+@register_test_case(module_factory=lambda: FloatPowerTensorTensorStaticModule())
+def FloatPowerTensorTensorStaticModule_basic(module, tu: TestUtils):
+    module.forward(tu.rand(3, 4))
+
+
+# ==============================================================================
+
+
 class ElementwiseWhereScalarModule(torch.nn.Module):
     def __init__(self):
         super().__init__()