DRAFT circle-mlir: Introduce pass for Sign operation

circle-mlir/circle-mlir/lib/dialect/mlir/CircleOps.td

@@ -1946,6 +1946,23 @@ def CIR_ShapeOp: CIR_Op<"shape", [
   let hasFolder = 1;
 }
 
+def CIR_SignOp: CIR_Op<"sign", [
+    Pure,
+    SameOperandsAndResultType]> {
+  let summary = "Sign operation";
+  let description = [{
+    Returns NaN if x is NaN, 0 if x is 0, -1 if x < 0 and 1 if x > 0.
+  }];
+
+  let arguments = (ins
+    CIR_TensorOf<[F32, F64, I32]>:$x
+  );
+
+  let results = (outs
+    CIR_TensorOf<[F32, F64, I32]>:$output
+  );
+}
+
 def CIR_SinOp: CIR_Op<"sin", [
     Pure,
     /*TF_SameOperandsAndResultTypeResolveRef*/

circle-mlir/circle-mlir/lib/pass/src/ConvertONNXToCirclePass.cpp

@@ -75,6 +75,7 @@
 #include "ops/ResizeOp.h"
 #include "ops/ShapeOp.h"
 #include "ops/SigmoidOp.h"
+#include "ops/SignOp.h"
 #include "ops/SinOp.h"
 #include "ops/SliceOp.h"
 #include "ops/SoftmaxOp.h"
@@ -263,6 +264,7 @@ void ConvertONNXToCirclePass::runOnOperation()
   patterns.insert<ConvResizeV13>(typeConverter, context);
   patterns.insert<ConvShape>(typeConverter, context);
   patterns.insert<ConvSigmoid>(typeConverter, context);
+  patterns.insert<ConvSign>(typeConverter, context);
   patterns.insert<ConvSin>(typeConverter, context);
   patterns.insert<ConvSlice>(typeConverter, context);
   patterns.insert<ConvSoftmax>(typeConverter, context);

circle-mlir/circle-mlir/lib/pass/src/ops/SignOp.h

@@ -0,0 +1,53 @@
+/*
+ * Copyright (c) 2026 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef __CIRCLE_MLIR_PASS_OPS_SIGN_OP_H__
+#define __CIRCLE_MLIR_PASS_OPS_SIGN_OP_H__
+
+#include <circle-mlir/dialect/CircleDialect.h>
+
+#include "ConvertHelper.h"
+
+#include <mlir/Transforms/DialectConversion.h>
+
+#include <src/Dialect/ONNX/ONNXOps.hpp>
+
+namespace mlir
+{
+namespace Circle
+{
+
+class ConvSign : public mlir::OpConversionPattern<mlir::ONNXSignOp>
+{
+public:
+  using mlir::OpConversionPattern<mlir::ONNXSignOp>::OpConversionPattern;
+  using OpAdaptor = typename mlir::ONNXSignOp::Adaptor;
+
+  mlir::LogicalResult matchAndRewrite(mlir::ONNXSignOp op, OpAdaptor adaptor,
+                                      mlir::ConversionPatternRewriter &rewriter) const override
+  {
+    mlir::Value input = adaptor.getInput();
+
+    rewriter.replaceOpWithNewOp<SignOp>(op, op.getType(), input);
+
+    return mlir::success();
+  }
+};
+
+} // namespace Circle
+} // namespace mlir
+
+#endif // __CIRCLE_MLIR_PASS_OPS_SIGN_OP_H__

circle-mlir/circle-mlir/tools-test/circle-impexp-test/test.lst

@@ -129,6 +129,7 @@ AddModel(Pow_F32_R2)
 AddModel(Pow_F32_R2_3)
 AddModel(Pow_F32_R4)
 AddModel(PReLU_F32_R4)
+AddModel(Range_F32_R0_1)
 AddModel(Reciprocal_F32_R4)
 AddModel(ReduceMax_F32_R2)
 AddModel(ReduceMax_F32_R2_d0)
@@ -166,6 +167,7 @@ AddModel(Resize_F32_R4_nearest)
 AddModel(Shape_F32_R4)
 AddModel(Sigmoid_F32_R2)
 AddModel(Sigmoid_F32_R4)
+AddModel(Sign_F32_R4)
 AddModel(Sin_F32_R4)
 AddModel(Slice_F32_R2_4) # Rank2 of Slice_F32_R4_4
 AddModel(Slice_F32_R3_4) # Rank3 of Slice_F32_R4_4

circle-mlir/circle-mlir/tools-test/onnx2circle-models/test.lst

@@ -137,6 +137,7 @@ AddModel(QuantizeLinear_F32_R3_ui8)
 AddModel(QuantizeLinear_F32_R3_ui8_fq)
 AddModel(QuantizeLinear_F32_R4_i16_cw)
 AddModel(QuantizeLinear_F32_R4_ui8_cw)
+AddModel(Range_F32_R0_1)
 AddModel(Reciprocal_F32_R4)
 AddModel(ReduceMax_F32_R2)
 AddModel(ReduceMax_F32_R2_d0)
@@ -175,6 +176,7 @@ AddModel(Rsqrt_F32_R4)
 AddModel(Shape_F32_R4)
 AddModel(Sigmoid_F32_R2)
 AddModel(Sigmoid_F32_R4)
+AddModel(Sign_F32_R4)
 AddModel(Sin_F32_R4)
 AddModel(Slice_F32_R2_4) # Rank2 of Slice_F32_R4_4
 AddModel(Slice_F32_R3_4) # Rank3 of Slice_F32_R4_4

circle-mlir/circle-mlir/tools-test/onnx2circle-value-test/test.lst

@@ -137,6 +137,7 @@ AddModel(PReLU_F32_R4)
 # AddModel(QuantizeLinear_F32_R3_ui8_fq)
 # AddModel(QuantizeLinear_F32_R4_i16_cw)
 # AddModel(QuantizeLinear_F32_R4_ui8_cw)
+AddModel(Range_F32_R0_1)
 AddModel(Reciprocal_F32_R4)
 AddModel(ReduceMax_F32_R2)
 AddModel(ReduceMax_F32_R2_d0)
@@ -175,6 +176,7 @@ AddModel(Rsqrt_F32_R4)
 AddModel(Shape_F32_R4)
 AddModel(Sigmoid_F32_R2)
 AddModel(Sigmoid_F32_R4)
+AddModel(Sign_F32_R4)
 AddModel(Sin_F32_R4)
 AddModel(Slice_F32_R2_4) # Rank2 of Slice_F32_R4_4
 AddModel(Slice_F32_R3_4) # Rank3 of Slice_F32_R4_4

circle-mlir/models/unit/Range_F32_R0_1/__init__.py

@@ -0,0 +1,20 @@
+import torch
+
+
+# Generate Range operator with Float32, scalar
+class net_Range(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, input):
+        limit = input
+        return torch.arange(0, limit, 1, dtype=torch.float32)
+
+    def onnx_opset_version(self):
+        return 11
+
+
+_model_ = net_Range()
+
+# produce float32 scalar with fixed number
+_inputs_ = torch.tensor(10, dtype=torch.float32)

circle-mlir/models/unit/Range_F32_R0_2/__init__.py

@@ -0,0 +1,19 @@
+import torch
+
+
+# Generate Range operator with Float32, scalar
+class net_Range(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, limit, delta):
+        return torch.arange(0, limit, delta, dtype=torch.float32)
+
+    def onnx_opset_version(self):
+        return 11
+
+
+_model_ = net_Range()
+
+# produce float32 scalar with fixed number
+_inputs_ = (torch.tensor(10, dtype=torch.float32), torch.tensor(1, dtype=torch.float32))

circle-mlir/models/unit/Range_F32_R0_3/__init__.py

@@ -0,0 +1,19 @@
+import torch
+
+
+# Generate Range operator with Float32, scalar
+class net_Range(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, start, limit, delta):
+        return torch.arange(start, limit, delta, dtype=torch.float32)
+
+    def onnx_opset_version(self):
+        return 11
+
+
+_model_ = net_Range()
+
+# produce float32 scalar with fixed number
+_inputs_ = (torch.tensor(0, dtype=torch.float32), torch.tensor(10, dtype=torch.float32), torch.tensor(1, dtype=torch.float32))

circle-mlir/models/unit/Sign_F32_R4/__init__.py

@@ -0,0 +1,19 @@
+import torch
+
+
+# Generate Sign operator with Float32, Rank-4
+class net_Sign(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, input):
+        return torch.sign(input)
+
+    def onnx_opset_version(self):
+        # TODO set to appropriate value
+        return 14
+
+
+_model_ = net_Sign()
+
+_inputs_ = torch.randn(1, 2, 3, 3)

circle-mlir/models/unit/Unsqueeze_F32_R0/__init__.py

@@ -16,5 +16,5 @@ def onnx_opset_version(self):
 
 _model_ = net_Unsqueeze()
 
-# produuce float32 scalar
+# produce float32 scalar
 _inputs_ = torch.randn(1)[0]

circle-mlir/models/unit/Unsqueeze_F32_R0_v11/__init__.py

@@ -15,5 +15,5 @@ def onnx_opset_version(self):
 
 _model_ = net_Unsqueeze()
 
-# produuce float32 scalar
+# produce float32 scalar
 _inputs_ = torch.randn(1)[0]