Add dual-conv tests for add/sub dual-input ifm scaling (#18755)

Differential Revision: D99934624

Pull Request resolved: #18755

Author: 3l1

backends/arm/test/ops/test_add.py

@@ -339,6 +339,46 @@ def test_add_conv_residual_u85_INT(test_data: input_t1):
     pipeline.run()
 
 
+class AddDualConv(torch.nn.Module):
+    """Conv1(x) + conv2(x) — both inputs have Rescale producers."""
+
+    def __init__(self):
+        super().__init__()
+        self.conv1 = torch.nn.Conv2d(3, 3, 1, bias=False)
+        self.conv2 = torch.nn.Conv2d(3, 3, 1, bias=False)
+
+    def forward(self, x):
+        return self.conv1(x) + self.conv2(x)
+
+    test_data = {
+        "4d_randn": lambda: (torch.randn(1, 3, 4, 4),),
+    }
+
+
+@common.parametrize("test_data", AddDualConv.test_data)
+def test_add_dual_conv_tosa_INT(test_data: input_t1):
+    pipeline = TosaPipelineINT[input_t1](AddDualConv(), test_data(), aten_op, exir_op)
+    pipeline.run()
+
+
+@common.parametrize("test_data", AddDualConv.test_data)
+@common.XfailIfNoCorstone300
+def test_add_dual_conv_u55_INT(test_data: input_t1):
+    pipeline = EthosU55PipelineINT[input_t1](
+        AddDualConv(), test_data(), aten_op, exir_op
+    )
+    pipeline.run()
+
+
+@common.parametrize("test_data", AddDualConv.test_data)
+@common.XfailIfNoCorstone320
+def test_add_dual_conv_u85_INT(test_data: input_t1):
+    pipeline = EthosU85PipelineINT[input_t1](
+        AddDualConv(), test_data(), aten_op, exir_op
+    )
+    pipeline.run()
+
+
 @common.parametrize("test_data", Add.test_data)
 def test_add_tensor_tosa_INT_16a8w(test_data: input_t1):
     """Test add operation with 16A8W quantization (16-bit activations, 8-bit

backends/arm/test/ops/test_sub.py

@@ -354,6 +354,46 @@ def test_sub_conv_residual_u85_INT(test_data: input_t1):
     pipeline.run()
 
 
+class SubDualConv(torch.nn.Module):
+    """conv1(x) - conv2(x) — both inputs have Rescale producers."""
+
+    def __init__(self):
+        super().__init__()
+        self.conv1 = torch.nn.Conv2d(3, 3, 1, bias=False)
+        self.conv2 = torch.nn.Conv2d(3, 3, 1, bias=False)
+
+    def forward(self, x):
+        return self.conv1(x) - self.conv2(x)
+
+    test_data = {
+        "4d_randn": lambda: (torch.randn(1, 3, 4, 4),),
+    }
+
+
+@common.parametrize("test_data", SubDualConv.test_data)
+def test_sub_dual_conv_tosa_INT(test_data: input_t1):
+    pipeline = TosaPipelineINT[input_t1](SubDualConv(), test_data(), aten_op, exir_op)
+    pipeline.run()
+
+
+@common.parametrize("test_data", SubDualConv.test_data)
+@common.XfailIfNoCorstone300
+def test_sub_dual_conv_u55_INT(test_data: input_t1):
+    pipeline = EthosU55PipelineINT[input_t1](
+        SubDualConv(), test_data(), aten_op, exir_op
+    )
+    pipeline.run()
+
+
+@common.parametrize("test_data", SubDualConv.test_data)
+@common.XfailIfNoCorstone320
+def test_sub_dual_conv_u85_INT(test_data: input_t1):
+    pipeline = EthosU85PipelineINT[input_t1](
+        SubDualConv(), test_data(), aten_op, exir_op
+    )
+    pipeline.run()
+
+
 @common.parametrize("test_data", sub_test_data)
 def test_sub_tensor_16a8w_tosa_INT(test_data: input_t1):
     """Test sub operation with 16A8W quantization (16-bit activations, 8-bit