Arm backend: Update Ethos-U compiler vela to 4.1.0

backends/arm/test/ops/test_bmm.py

@@ -22,8 +22,8 @@ class TestBMM(unittest.TestCase):
 
     class BMM(torch.nn.Module):
         test_parameters = [
-            (torch.rand(5, 3, 5), torch.rand(5, 5, 2)),
             (torch.rand(2, 1, 1), torch.rand(2, 1, 1)),
+            (torch.rand(5, 3, 5), torch.rand(5, 5, 2)),
             (torch.ones(1, 55, 3), torch.ones(1, 3, 44)),
             (10000 * torch.randn(10, 1, 10), torch.randn(10, 10, 5)),
             (-10 * torch.randn(2, 32, 64), 5 + 5 * torch.randn(2, 64, 32)),
@@ -147,32 +147,37 @@ def test_bmm_single_input_tosa_BI(self, operand1: torch.Tensor):
 
     @parameterized.expand(BMM.test_parameters)
     @unittest.expectedFailure
-    def test_bmm_u55_BI(self, operand1: torch.Tensor, operand2: torch.Tensor):
+    def test_bmm_u55_BI_xfails(self, operand1: torch.Tensor, operand2: torch.Tensor):
         test_data = (operand1, operand2)
         self._test_bmm_ethosu_BI_pipeline(
             self.BMM(), common.get_u55_compile_spec(), test_data
         )
 
-    @parameterized.expand(BMM.test_parameters)
-    @common.expectedFailureOnFVP
+    @parameterized.expand(BMM.test_parameters[:1])
     def test_bmm_u85_BI(self, operand1: torch.Tensor, operand2: torch.Tensor):
         test_data = (operand1, operand2)
         self._test_bmm_ethosu_BI_pipeline(
             self.BMM(), common.get_u85_compile_spec(), test_data
         )
 
+    @parameterized.expand(BMM.test_parameters[1:])
+    @common.expectedFailureOnFVP
+    def test_bmm_u85_BI_xfails(self, operand1: torch.Tensor, operand2: torch.Tensor):
+        test_data = (operand1, operand2)
+        self._test_bmm_ethosu_BI_pipeline(
+            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
     @parameterized.expand(BMMSingleInput.test_parameters)
     @unittest.expectedFailure
-    def test_bmm_single_input_u55_BI(self, operand1: torch.Tensor):
+    def test_bmm_single_input_u55_BI_xfails(self, operand1: torch.Tensor):
         test_data = (operand1,)
         self._test_bmm_ethosu_BI_pipeline(
             self.BMMSingleInput(), common.get_u55_compile_spec(), test_data
         )
 
-    # Numerical issues on FVP, MLETORCH 534
     @parameterized.expand(BMMSingleInput.test_parameters)
-    @common.expectedFailureOnFVP
     def test_bmm_single_input_u85_BI(self, operand1: torch.Tensor):
         test_data = (operand1,)
         self._test_bmm_ethosu_BI_pipeline(

backends/arm/test/ops/test_conv_combos.py

@@ -275,8 +275,6 @@ def test_conv_meandim_u55_BI(self):
             model.get_inputs(),
         )
 
-    # Numerical Issues on FVP, MLETORCH-520
-    @common.expectedFailureOnFVP
     def test_conv_meandim_u85_BI(self):
         model = ComboConv2dMeandim()
         self._test_conv_combo_ethos_BI_pipeline(

backends/arm/test/ops/test_depthwise_conv.py

@@ -156,6 +156,19 @@
     ("two_dw_conv2d", two_dw_conv2d),
 ]
 
+testsuite_conv2d_u85 = [
+    ("2x2_1x6x4x4_gp6_st1", dw_conv2d_2x2_1x6x4x4_gp6_st1),
+    ("3x3_1x3x256x256_gp3_st1", dw_conv2d_3x3_1x3x256x256_gp3_st1),
+    ("3x3_1x4x256x256_gp4_st1", dw_conv2d_3x3_1x4x256x256_gp4_st1),
+    ("3x3_1x4x256x256_gp4_nobias", dw_conv2d_3x3_1x4x256x256_gp4_nobias),
+]
+
+testsuite_conv2d_u85_xfails = [
+    ("3x3_2x8x198x198_gp8_st3", dw_conv2d_3x3_2x8x198x198_gp8_st3),
+    ("two_dw_conv2d", two_dw_conv2d),
+]
+
+
 testsuite_conv1d = [
     ("2_1x6x4_gp6_st1", dw_conv1d_2_1x6x4_gp6_st1),
     ("two_dw_conv1d", two_dw_conv1d),
@@ -247,7 +260,7 @@ def test_dw_conv_tosa_BI(self, test_name: str, model: torch.nn.Module):
     )  # Works
 
     @parameterized.expand(testsuite_conv2d, skip_on_empty=True)
-    @common.expectedFailureOnFVP
+    @unittest.expectedFailure
     def test_dw_conv2d_u55_BI(
         self, test_name: str, model: torch.nn.Module, set_quantize_io: bool = False
     ):
@@ -274,10 +287,8 @@ def test_dw_conv1d_u55_BI(
             model.get_inputs(),
         )
 
-    # All test cases except 3x3_1x3x256x256_gp3_st1 have numerical issues on FVP. MLETORCH-520
-    @parameterized.expand(testsuite_conv1d[:-2] + testsuite_conv2d)
-    @common.expectedFailureOnFVP
-    def test_dw_conv_u85_BI_xfails(
+    @parameterized.expand(testsuite_conv1d + testsuite_conv2d_u85)
+    def test_dw_conv_u85_BI(
         self, test_name: str, model: torch.nn.Module, set_quantize_io: bool = False
     ):
         self._test_dw_conv_ethos_BI_pipeline(
@@ -288,8 +299,10 @@ def test_dw_conv_u85_BI_xfails(
             model.get_inputs(),
         )
 
-    @parameterized.expand(testsuite_conv1d[-2:])
-    def test_dw_conv_u85_BI(
+    # All test cases except 3x3_1x3x256x256_gp3_st1 have numerical issues on FVP. MLETORCH-520
+    @parameterized.expand(testsuite_conv2d_u85_xfails)
+    @common.expectedFailureOnFVP
+    def test_dw_conv_u85_BI_xfails(
         self, test_name: str, model: torch.nn.Module, set_quantize_io: bool = False
     ):
         self._test_dw_conv_ethos_BI_pipeline(

backends/arm/test/ops/test_div.py

@@ -26,18 +26,18 @@
         torch.ones(5),
         None,
     ),
-    (
-        "op_div_rank1_rand",
-        torch.rand(5) * 5,
-        torch.rand(5) * 5,
-        None,
-    ),
     (
         "op_div_rank1_negative_ones",
         torch.ones(5) * (-1),
         torch.ones(5) * (-1),
         None,
     ),
+    (
+        "op_div_rank1_rand",
+        torch.rand(5) * 5,
+        torch.rand(5) * 5,
+        None,
+    ),
     (
         "op_div_rank4_ones",
         torch.ones(5, 10, 25, 20),
@@ -183,9 +183,7 @@ def test_div_tosa_BI(
         test_data = (input_, other_)
         self._test_div_tosa_BI_pipeline(self.Div(), test_data)
 
-    # Numerical issues on FVP likely due to mul op, MLETORCH-521
-    @parameterized.expand(test_data_suite)
-    @common.expectedFailureOnFVP
+    @parameterized.expand(test_data_suite[:2])
     def test_div_u55_BI(
         self,
         test_name: str,
@@ -199,8 +197,21 @@ def test_div_u55_BI(
         )
 
     # Numerical issues on FVP likely due to mul op, MLETORCH-521
-    @parameterized.expand(test_data_suite)
+    @parameterized.expand(test_data_suite[2:])
     @common.expectedFailureOnFVP
+    def test_div_u55_BI_xfails(
+        self,
+        test_name: str,
+        input_: Union[torch.Tensor, torch.types.Number],
+        other_: Union[torch.Tensor, torch.types.Number],
+        rounding_mode: Optional[str] = None,
+    ):
+        test_data = (input_, other_)
+        self._test_div_ethos_BI_pipeline(
+            self.Div(), common.get_u55_compile_spec(), test_data
+        )
+
+    @parameterized.expand(test_data_suite[:2])
     def test_div_u85_BI(
         self,
         test_name: str,
@@ -212,3 +223,18 @@ def test_div_u85_BI(
         self._test_div_ethos_BI_pipeline(
             self.Div(), common.get_u85_compile_spec(), test_data
         )
+
+    # Numerical issues on FVP likely due to mul op, MLETORCH-521
+    @parameterized.expand(test_data_suite[2:])
+    @common.expectedFailureOnFVP
+    def test_div_u85_BI_xfails(
+        self,
+        test_name: str,
+        input_: Union[torch.Tensor, torch.types.Number],
+        other_: Union[torch.Tensor, torch.types.Number],
+        rounding_mode: Optional[str] = None,
+    ):
+        test_data = (input_, other_)
+        self._test_div_ethos_BI_pipeline(
+            self.Div(), common.get_u85_compile_spec(), test_data
+        )

backends/arm/test/ops/test_layer_norm.py

@@ -158,7 +158,7 @@ def test_layer_norm_tosa_BI(
     # Numerical issues on FVP likely due to mul op, MLETORCH-521
     # Skip tests that require transposes.
     @parameterized.expand(test_data_suite[:-2])
-    @common.expectedFailureOnFVP
+    @unittest.expectedFailure
     def test_layer_norm_u55_BI(
         self,
         test_name: str,
@@ -170,9 +170,8 @@ def test_layer_norm_u55_BI(
         )
 
     # Numerical issues on FVP likely due to mul op, MLETORCH-521
-    @parameterized.expand(test_data_suite[:-1])
-    @common.expectedFailureOnFVP
-    def test_layer_norm_u85_BI_fvp_xfails(
+    @parameterized.expand(test_data_suite[:-2])
+    def test_layer_norm_u85_BI_fvp(
         self,
         test_name: str,
         test_data: torch.Tensor,
@@ -182,7 +181,7 @@ def test_layer_norm_u85_BI_fvp_xfails(
             self.LayerNorm(*model_params), common.get_u85_compile_spec(), (test_data,)
         )
 
-    @parameterized.expand(test_data_suite[-1:])
+    @parameterized.expand(test_data_suite[-2:])
     @unittest.skip  # Flaky
     def test_layer_norm_u85_BI(
         self,

backends/arm/test/ops/test_logsoftmax.py

@@ -17,14 +17,29 @@
 
 test_data_suite = [
     # (test_name, test_data, dim)
-    ("zeros", torch.zeros(10, 10, 10, 10), 0),
-    ("zeros_neg_dim", torch.zeros(10, 10, 10, 10), -4),
+    ("zeros", torch.zeros(10, 8, 5, 2), 0),
+    ("zeros_neg_dim", torch.zeros(10, 7, 8, 9), -4),
     ("ones", torch.ones(10, 10), 1),
-    ("rand_neg_dim", torch.rand(10, 10, 10), -1),
-    ("rand", torch.rand(10, 10, 10, 10), 2),
-    ("rand_neg_dim", torch.rand(10, 10, 2, 3), -2),
-    ("randn", torch.randn(10, 10, 5, 10), 3),
-    ("randn_neg_dim", torch.randn(1, 10, 10, 10), -3),
+    ("ones_neg_dim", torch.ones(10, 3, 4), -1),
+    ("rand", torch.rand(1, 2, 5, 8), 2),
+    ("rand_neg_dim", torch.rand(2, 10, 8, 10), -2),
+    ("randn", torch.randn(10, 10, 10, 10), 3),
+    ("randn_neg_dim", torch.randn(10, 5, 8, 7), -3),
+]
+test_data_suite_u55 = [
+    # (test_name, test_data, dim)
+    ("ones", torch.ones(10, 10), 1),
+    ("ones_neg_dim", torch.ones(10, 3, 4), -1),
+    ("randn_neg_dim", torch.randn(10, 5, 8, 7), -3),
+]
+
+test_data_suite_u55_xfails = [
+    # (test_name, test_data, dim)
+    ("zeros", torch.zeros(10, 8, 5, 2), 0),
+    ("zeros_neg_dim", torch.zeros(10, 7, 8, 9), -4),
+    ("rand", torch.rand(1, 2, 5, 8), 2),
+    ("rand_neg_dim", torch.rand(2, 10, 8, 10), -2),
+    ("randn", torch.randn(10, 10, 10, 10), 3),
 ]
 
 
@@ -135,7 +150,7 @@ def test_logsoftmax_tosa_BI(
     ):
         self._test_logsoftmax_tosa_BI_pipeline(self.LogSoftmax(dim=dim), (test_data,))
 
-    @parameterized.expand(test_data_suite)
+    @parameterized.expand(test_data_suite_u55)
     def test_logsoftmax_tosa_u55_BI(
         self,
         test_name: str,
@@ -146,13 +161,26 @@ def test_logsoftmax_tosa_u55_BI(
             self.LogSoftmax(dim=dim), (test_data,)
         )
 
+    # Expected to fail as this is not supported on u55.
+    @parameterized.expand(test_data_suite_u55_xfails)
+    @unittest.expectedFailure
+    def test_logsoftmax_tosa_u55_BI_xfails(
+        self,
+        test_name: str,
+        test_data: torch.Tensor,
+        dim: int,
+    ):
+        self._test_logsoftmax_tosa_u55_BI_pipeline(
+            self.LogSoftmax(dim=dim), (test_data,)
+        )
+
     @parameterized.expand(test_data_suite)
     def test_logsoftmax_tosa_u85_BI(
         self,
         test_name: str,
         test_data: torch.Tensor,
         dim: int,
     ):
-        self._test_logsoftmax_tosa_u55_BI_pipeline(
+        self._test_logsoftmax_tosa_u85_BI_pipeline(
             self.LogSoftmax(dim=dim), (test_data,)
         )

backends/arm/test/ops/test_mean_dim.py

@@ -269,8 +269,10 @@ def test_meandim_tosa_BI(
     ):
         self._test_meandim_tosa_BI_pipeline(self.MeanDim(dim, keepdim), (test_data,))
 
+    # Expected to fail as this is not supported on u55.
     @parameterized.expand(MeanDim.test_data_suite)
-    def test_meandim_tosa_u55_BI(
+    @unittest.expectedFailure
+    def test_meandim_tosa_u55_BI_xfails(
         self,
         test_name: str,
         test_data: torch.Tensor,

backends/arm/test/ops/test_mul.py

@@ -152,9 +152,7 @@ def test_mul_tosa_BI(
         test_data = (input_, other_)
         self._test_mul_tosa_BI_pipeline(self.Mul(), test_data)
 
-    # Numerical issues on FVP, MLETORCH-521
     @parameterized.expand(test_data_sute)
-    @common.expectedFailureOnFVP
     def test_mul_u55_BI(
         self,
         test_name: str,
@@ -166,10 +164,7 @@ def test_mul_u55_BI(
             common.get_u55_compile_spec(), self.Mul(), test_data
         )
 
-    # Numerical issues on FVP, MLETORCH-521
-    # test_data_sute[0] works on U85
-    @parameterized.expand(test_data_sute[1:])
-    @common.expectedFailureOnFVP
+    @parameterized.expand(test_data_sute)
     def test_mul_u85_BI(
         self,
         test_name: str,

backends/arm/test/ops/test_softmax.py

@@ -28,6 +28,22 @@
     ("randn_neg_dim", torch.randn(10, 5, 8, 7), -3),
 ]
 
+test_data_suite_u55 = [
+    # (test_name, test_data, dim)
+    ("ones", torch.ones(10, 10), 1),
+    ("ones_neg_dim", torch.ones(10, 3, 4), -1),
+    ("randn_neg_dim", torch.randn(10, 5, 8, 7), -3),
+]
+
+test_data_suite_u55_xfails = [
+    # (test_name, test_data, dim)
+    ("zeros", torch.zeros(10, 8, 5, 2), 0),
+    ("zeros_neg_dim", torch.zeros(10, 7, 8, 9), -4),
+    ("rand", torch.rand(1, 2, 5, 8), 2),
+    ("rand_neg_dim", torch.rand(2, 10, 8, 10), -2),
+    ("randn", torch.randn(10, 10, 10, 10), 3),
+]
+
 
 class TestSoftmax(unittest.TestCase):
     """Tests softmax."""
@@ -136,7 +152,7 @@ def test_softmax_tosa_BI(
     ):
         self._test_softmax_tosa_BI_pipeline(self.Softmax(dim=dim), (test_data,))
 
-    @parameterized.expand(test_data_suite)
+    @parameterized.expand(test_data_suite_u55)
     def test_softmax_tosa_u55_BI(
         self,
         test_name: str,
@@ -145,6 +161,17 @@ def test_softmax_tosa_u55_BI(
     ):
         self._test_softmax_tosa_u55_BI_pipeline(self.Softmax(dim=dim), (test_data,))
 
+    # Expected to fail as this is not supported on u55.
+    @parameterized.expand(test_data_suite_u55_xfails)
+    @unittest.expectedFailure
+    def test_softmax_tosa_u55_BI_xfails(
+        self,
+        test_name: str,
+        test_data: torch.Tensor,
+        dim: int,
+    ):
+        self._test_softmax_tosa_u55_BI_pipeline(self.Softmax(dim=dim), (test_data,))
+
     @parameterized.expand(test_data_suite)
     def test_softmax_tosa_u85_BI(
         self,