Arm backend: Enable FVP tests for ops lowered to REDUCE_SUM

backends/arm/test/misc/test_multiple_outputs.py

@@ -76,23 +76,21 @@ def _test_ethosu_BI_pipeline(
             tester.run_method_and_compare_outputs(qtol=1, inputs=test_data)
 
     @pytest.mark.corstone_fvp
-    def test_u85_BI(self):
+    def test_u55_BI(self):
         module = self.MultipleOutputsModule()
         test_data = module.get_inputs()
         self._test_ethosu_BI_pipeline(
             module,
             test_data,
-            common.get_u85_compile_spec(),
+            common.get_u55_compile_spec(),
         )
 
     @pytest.mark.corstone_fvp
-    @conftest.expectedFailureOnFVP
-    # TODO MLETORCH-598
-    def test_u55_BI(self):
+    def test_u85_BI(self):
         module = self.MultipleOutputsModule()
         test_data = module.get_inputs()
         self._test_ethosu_BI_pipeline(
             module,
             test_data,
-            common.get_u55_compile_spec(),
+            common.get_u85_compile_spec(),
         )

backends/arm/test/ops/test_bmm.py

@@ -150,9 +150,10 @@ def test_bmm_single_input_tosa_BI(self, test_data_generator: Callable[[], Tuple]
         test_data = test_data_generator()
         self._test_bmm_tosa_BI_pipeline(self.BMMSingleInput(), test_data)
 
+    # Expected to fail on FVP as TOSA.MATMUL is not supported on U55
     @parameterized.expand(BMM.test_data_generators)
     @pytest.mark.corstone_fvp
-    @unittest.expectedFailure
+    @conftest.expectedFailureOnFVP
     def test_bmm_u55_BI_xfails(self, test_data_generator: Callable[[], Tuple]):
         test_data = test_data_generator()
         self._test_bmm_ethosu_BI_pipeline(
@@ -167,10 +168,10 @@ def test_bmm_u85_BI(self, test_data_generator: Callable[[], Tuple]):
             self.BMM(), common.get_u85_compile_spec(), test_data
         )
 
-    # Expected to fail with error: Warning, unsupported fusing of TOSA Rescale previous operator is of type: Memcpy
+    # Expected to fail on FVP as TOSA.MATMUL is not supported on U55
     @parameterized.expand(BMMSingleInput.test_data_generators)
     @pytest.mark.corstone_fvp
-    @unittest.expectedFailure
+    @conftest.expectedFailureOnFVP
     def test_bmm_single_input_u55_BI_xfails(
         self, test_data_generator: Callable[[], Tuple]
     ):

backends/arm/test/ops/test_layer_norm.py

@@ -158,7 +158,7 @@ def test_layer_norm_tosa_BI(
             self.LayerNorm(*model_params), (test_data,)
         )
 
-    @parameterized.expand(test_data_suite[4:])
+    @parameterized.expand(test_data_suite)
     @pytest.mark.corstone_fvp
     def test_layer_norm_u55_BI(
         self,
@@ -170,36 +170,7 @@ def test_layer_norm_u55_BI(
             self.LayerNorm(*model_params), common.get_u55_compile_spec(), (test_data,)
         )
 
-    # Numerical issues on FVP likely due to mul op, MLETORCH-521
-    # Skip tests that require transposes.
-    @parameterized.expand(test_data_suite[:4])
-    @pytest.mark.corstone_fvp
-    @conftest.expectedFailureOnFVP
-    def test_layer_norm_u55_BI_xfails(
-        self,
-        test_name: str,
-        test_data: torch.Tensor,
-        model_params,
-    ):
-        self._test_layernorm_ethosu_BI_pipeline(
-            self.LayerNorm(*model_params), common.get_u55_compile_spec(), (test_data,)
-        )
-
-    # Numerical issues on FVP likely due to mul op, MLETORCH-521
-    @parameterized.expand(test_data_suite[:-2])
-    @pytest.mark.corstone_fvp
-    @conftest.expectedFailureOnFVP
-    def test_layer_norm_u85_BI_xfails(
-        self,
-        test_name: str,
-        test_data: torch.Tensor,
-        model_params,
-    ):
-        self._test_layernorm_ethosu_BI_pipeline(
-            self.LayerNorm(*model_params), common.get_u85_compile_spec(), (test_data,)
-        )
-
-    @parameterized.expand(test_data_suite[-2:])
+    @parameterized.expand(test_data_suite)
     @pytest.mark.corstone_fvp
     def test_layer_norm_u85_BI(
         self,

backends/arm/test/ops/test_logsoftmax.py

@@ -11,7 +11,7 @@
 import pytest
 
 import torch
-from executorch.backends.arm.test import common
+from executorch.backends.arm.test import common, conftest
 from executorch.backends.arm.test.tester.arm_tester import ArmTester
 from executorch.exir.backend.compile_spec_schema import CompileSpec
 from parameterized import parameterized
@@ -28,16 +28,17 @@
     lambda: ("randn", torch.randn(10, 10, 10, 10), 3),
     lambda: ("randn_neg_dim", torch.randn(10, 5, 8, 7), -3),
 ]
-test_data_generators_u55 = [
+
+test_data_generators_FVP = [
     # (test_name, test_data, dim)
     lambda: ("ones", torch.ones(10, 10), 1),
     lambda: ("ones_neg_dim", torch.ones(10, 3, 4), -1),
-    lambda: ("randn_neg_dim", torch.randn(10, 5, 8, 7), -3),
-    lambda: ("zeros", torch.zeros(10, 8, 5, 2), 0),
-    lambda: ("zeros_neg_dim", torch.zeros(10, 7, 8, 9), -4),
+    lambda: ("randn_neg_dim", torch.randn(1, 5, 8, 7), -3),
+    lambda: ("zeros", torch.zeros(1, 8, 5, 2), 0),
+    lambda: ("zeros_neg_dim", torch.zeros(1, 7, 8, 9), -4),
     lambda: ("rand", torch.rand(1, 2, 5, 8), 2),
-    lambda: ("rand_neg_dim", torch.rand(2, 10, 8, 10), -2),
-    lambda: ("randn", torch.randn(10, 10, 10, 10), 3),
+    lambda: ("rand_neg_dim", torch.rand(1, 10, 8, 10), -2),
+    lambda: ("randn", torch.randn(1, 10, 10, 10), 3),
 ]
 
 
@@ -99,7 +100,7 @@ def _test_logsoftmax_tosa_ethos_BI_pipeline(
         module: torch.nn.Module,
         test_data: Tuple[torch.tensor],
     ):
-        (
+        tester = (
             ArmTester(
                 module,
                 example_inputs=test_data,
@@ -114,21 +115,10 @@ def _test_logsoftmax_tosa_ethos_BI_pipeline(
             .check_not(["executorch_exir_dialects_edge__ops_aten__logsoftmax_default"])
             .check_count({"torch.ops.higher_order.executorch_call_delegate": 1})
             .to_executorch()
+            .serialize()
         )
-
-    def _test_logsoftmax_tosa_u55_BI_pipeline(
-        self, module: torch.nn.Module, test_data: Tuple[torch.tensor]
-    ):
-        self._test_logsoftmax_tosa_ethos_BI_pipeline(
-            common.get_u55_compile_spec(), module, test_data
-        )
-
-    def _test_logsoftmax_tosa_u85_BI_pipeline(
-        self, module: torch.nn.Module, test_data: Tuple[torch.tensor]
-    ):
-        self._test_logsoftmax_tosa_ethos_BI_pipeline(
-            common.get_u85_compile_spec(), module, test_data
-        )
+        if conftest.is_option_enabled("corstone_fvp"):
+            tester.run_method_and_compare_outputs(inputs=test_data, qtol=1)
 
     @parameterized.expand(test_data_generators)
     def test_logsoftmax_tosa_MI(self, test_data_generator: Callable[[], Tuple]):
@@ -141,18 +131,18 @@ def test_logsoftmax_tosa_BI(self, test_data_generator: Callable[[], Tuple]):
         test_name, test_data, dim = test_data_generator()
         self._test_logsoftmax_tosa_BI_pipeline(self.LogSoftmax(dim=dim), (test_data,))
 
-    @parameterized.expand(test_data_generators_u55)
+    @parameterized.expand(test_data_generators_FVP)
     @pytest.mark.flaky  # TODO: MLETORCH-460 - Numerically stabler (log)softmax implementation
     def test_logsoftmax_tosa_u55_BI(self, test_data_generator: Callable[[], Tuple]):
         test_name, test_data, dim = test_data_generator()
-        self._test_logsoftmax_tosa_u55_BI_pipeline(
-            self.LogSoftmax(dim=dim), (test_data,)
+        self._test_logsoftmax_tosa_ethos_BI_pipeline(
+            common.get_u55_compile_spec(), self.LogSoftmax(dim=dim), (test_data,)
         )
 
-    @parameterized.expand(test_data_generators)
+    @parameterized.expand(test_data_generators_FVP)
     @pytest.mark.flaky  # TODO: MLETORCH-460 - Numerically stabler (log)softmax implementation
     def test_logsoftmax_tosa_u85_BI(self, test_data_generator: Callable[[], Tuple]):
         test_name, test_data, dim = test_data_generator()
-        self._test_logsoftmax_tosa_u85_BI_pipeline(
-            self.LogSoftmax(dim=dim), (test_data,)
+        self._test_logsoftmax_tosa_ethos_BI_pipeline(
+            common.get_u85_compile_spec(), self.LogSoftmax(dim=dim), (test_data,)
         )

backends/arm/test/ops/test_mean_dim.py

@@ -10,7 +10,7 @@
 from typing import Tuple
 
 import torch
-from executorch.backends.arm.test import common
+from executorch.backends.arm.test import common, conftest
 from executorch.backends.arm.test.tester.arm_tester import ArmTester
 from executorch.exir.backend.backend_details import CompileSpec
 from parameterized import parameterized
@@ -121,7 +121,7 @@ def _test_adaptive_avg_pool2d_tosa_ethosu_BI_pipeline(
         compile_spec: CompileSpec,
         test_data: Tuple[torch.tensor],
     ):
-        (
+        tester = (
             ArmTester(
                 module,
                 example_inputs=test_data,
@@ -141,7 +141,10 @@ def _test_adaptive_avg_pool2d_tosa_ethosu_BI_pipeline(
             )
             .check_count({"torch.ops.higher_order.executorch_call_delegate": 1})
             .to_executorch()
+            .serialize()
         )
+        if conftest.is_option_enabled("corstone_fvp"):
+            tester.run_method_and_compare_outputs(inputs=test_data)
 
     def _test_meandim_tosa_MI_pipeline(
         self, module: torch.nn.Module, test_data: Tuple[torch.tensor]
@@ -188,7 +191,7 @@ def _test_meandim_tosa_ethosu_BI_pipeline(
         compile_spec: CompileSpec,
         test_data: Tuple[torch.tensor],
     ):
-        (
+        tester = (
             ArmTester(
                 module,
                 example_inputs=test_data,
@@ -207,7 +210,10 @@ def _test_meandim_tosa_ethosu_BI_pipeline(
             )
             .check_count({"torch.ops.higher_order.executorch_call_delegate": 1})
             .to_executorch()
+            .serialize()
         )
+        if conftest.is_option_enabled("corstone_fvp"):
+            tester.run_method_and_compare_outputs(inputs=test_data, qtol=1)
 
     @parameterized.expand(AdaptiveAveragePool2d.test_data_suite)
     def test_adaptive_avg_pool2d_tosa_MI(

backends/arm/test/ops/test_mm.py

@@ -132,18 +132,16 @@ def test_mm_single_input_tosa_BI(self, test_data_generator: Callable[[], Tuple])
         test_data = test_data_generator()
         self._test_mm_tosa_BI_pipeline(self.MMSingleInput(), test_data)
 
-    # Expected to fail with error: CPU performance estimation for "MatMul" not implemented
+    # TODO: Enable numerical testing
     @parameterized.expand(MM.test_data_generators)
-    @unittest.expectedFailure
     def test_mm_u55_BI(self, test_data_generator: Callable[[], Tuple]):
         test_data = test_data_generator()
         self._test_mm_ethosu_BI_pipeline(
             common.get_u55_compile_spec(), self.MM(), test_data
         )
 
-    # Expected to fail with error: Warning, unsupported fusing of TOSA Rescale previous operator is of type: Memcpy
+    # TODO: Enable numerical testing
     @parameterized.expand(MMSingleInput.test_data_generators)
-    @unittest.expectedFailure
     def test_mm_single_input_u55_BI(self, test_data_generator: Callable[[], Tuple]):
         test_data = test_data_generator()
         self._test_mm_ethosu_BI_pipeline(

backends/arm/test/ops/test_softmax.py

@@ -12,7 +12,7 @@
 import pytest
 
 import torch
-from executorch.backends.arm.test import common
+from executorch.backends.arm.test import common, conftest
 from executorch.backends.arm.test.tester.arm_tester import ArmTester
 from executorch.exir.backend.compile_spec_schema import CompileSpec
 from parameterized import parameterized
@@ -30,16 +30,16 @@
     lambda: ("randn_neg_dim", torch.randn(10, 5, 8, 7), -3),
 ]
 
-test_data_generators_u55 = [
+test_data_generators_FVP = [
     # (test_name, test_data, dim)
     lambda: ("ones", torch.ones(10, 10), 1),
-    lambda: ("ones_neg_dim", torch.ones(10, 3, 4), -1),
-    lambda: ("randn_neg_dim", torch.randn(10, 5, 8, 7), -3),
-    lambda: ("zeros", torch.zeros(10, 8, 5, 2), 0),
-    lambda: ("zeros_neg_dim", torch.zeros(10, 7, 8, 9), -4),
+    lambda: ("ones_neg_dim", torch.ones(1, 3, 4), -1),
+    lambda: ("randn_neg_dim", torch.randn(1, 5, 8, 7), -3),
+    lambda: ("zeros", torch.zeros(1, 8, 5, 2), 0),
+    lambda: ("zeros_neg_dim", torch.zeros(1, 7, 8, 9), -4),
     lambda: ("rand", torch.rand(1, 2, 5, 8), 2),
-    lambda: ("rand_neg_dim", torch.rand(2, 10, 8, 10), -2),
-    lambda: ("randn", torch.randn(10, 10, 10, 10), 3),
+    lambda: ("rand_neg_dim", torch.rand(1, 10, 8, 10), -2),
+    lambda: ("randn", torch.randn(1, 10, 10, 10), 3),
 ]
 
 
@@ -95,13 +95,13 @@ def _test_softmax_tosa_BI_pipeline(
             .run_method_and_compare_outputs(inputs=test_data)
         )
 
-    def _test_softmax_tosa_ethos_BI_pipeline(
+    def _test_softmax_ethosu_BI_pipeline(
         self,
         compile_spec: list[CompileSpec],
         module: torch.nn.Module,
         test_data: Tuple[torch.tensor],
     ):
-        (
+        tester = (
             ArmTester(
                 module,
                 example_inputs=test_data,
@@ -116,21 +116,10 @@ def _test_softmax_tosa_ethos_BI_pipeline(
             .check_not(["executorch_exir_dialects_edge__ops_aten__softmax_default"])
             .check_count({"torch.ops.higher_order.executorch_call_delegate": 1})
             .to_executorch()
+            .serialize()
         )
-
-    def _test_softmax_tosa_u55_BI_pipeline(
-        self, module: torch.nn.Module, test_data: Tuple[torch.tensor]
-    ):
-        self._test_softmax_tosa_ethos_BI_pipeline(
-            common.get_u55_compile_spec(), module, test_data
-        )
-
-    def _test_softmax_tosa_u85_BI_pipeline(
-        self, module: torch.nn.Module, test_data: Tuple[torch.tensor]
-    ):
-        self._test_softmax_tosa_ethos_BI_pipeline(
-            common.get_u85_compile_spec(), module, test_data
-        )
+        if conftest.is_option_enabled("corstone_fvp"):
+            tester.run_method_and_compare_outputs(inputs=test_data, qtol=1)
 
     @parameterized.expand(test_data_generators)
     def test_softmax_tosa_MI(self, test_data_generator: Callable[[], Tuple]):
@@ -143,14 +132,18 @@ def test_softmax_tosa_BI(self, test_data_generator: Callable[[], Tuple]):
         test_name, test_data, dim = test_data_generator()
         self._test_softmax_tosa_BI_pipeline(self.Softmax(dim=dim), (test_data,))
 
-    @parameterized.expand(test_data_generators_u55)
+    @parameterized.expand(test_data_generators_FVP)
     @pytest.mark.flaky  # TODO: MLETORCH-460 - Numerically stabler (log)softmax implementation
-    def test_softmax_tosa_u55_BI(self, test_data_generator: Callable[[], Tuple]):
+    def test_softmax_u55_BI(self, test_data_generator: Callable[[], Tuple]):
         test_name, test_data, dim = test_data_generator()
-        self._test_softmax_tosa_u55_BI_pipeline(self.Softmax(dim=dim), (test_data,))
+        self._test_softmax_ethosu_BI_pipeline(
+            common.get_u55_compile_spec(), self.Softmax(dim=dim), (test_data,)
+        )
 
-    @parameterized.expand(test_data_generators)
+    @parameterized.expand(test_data_generators_FVP)
     @pytest.mark.flaky  # TODO: MLETORCH-460 - Numerically stabler (log)softmax implementation
-    def test_softmax_tosa_u85_BI(self, test_data_generator: Callable[[], Tuple]):
+    def test_softmax_u85_BI(self, test_data_generator: Callable[[], Tuple]):
         test_name, test_data, dim = test_data_generator()
-        self._test_softmax_tosa_u85_BI_pipeline(self.Softmax(dim=dim), (test_data,))
+        self._test_softmax_ethosu_BI_pipeline(
+            common.get_u85_compile_spec(), self.Softmax(dim=dim), (test_data,)
+        )