Arm backend: Add support for sigmoid and tanh int16x8 (#15101)

Adds support for sigmoid and tanh to arm backend for int16x8 support 
Removes unnecessary test_sigmoid_16bit as these are now covered by
test_sigmoid

Author: SaoirseARM

backends/arm/operator_support/ethos_u55_support.py

@@ -114,9 +114,9 @@ def is_node_supported(  # noqa: C901
                 return False
 
         if node.target in self.target_ops_i8:
-            if dtype not in (torch.int8,):
+            if dtype not in (torch.int8, torch.int16):
                 self.reporter.report_reject(
-                    node, f"Unsupported dtype {dtype} (Supports i8)."
+                    node, f"Unsupported dtype {dtype} (Supports i8, i16)."
                 )
                 return False
 

backends/arm/quantizer/arm_quantizer.py

@@ -161,6 +161,7 @@ def get_symmetric_a16w8_quantization_config(
     is_dynamic: bool = False,
     weight_qmin: int = -127,
     weight_qmax: int = 127,
+    epsilon: float = 2**-12,
 ):
     """
     16A8W quantization config: 16-bit activations, 8-bit weights.
@@ -174,11 +175,12 @@ def get_symmetric_a16w8_quantization_config(
         is_dynamic: Whether to use dynamic quantization
         weight_qmin: Minimum quantization value for weights
         weight_qmax: Maximum quantization value for weights
+        epsilon: Value used to pad observed [qmin, qmax] before initial zero point and scale calculation
 
     Returns:
         QuantizationConfig with 16-bit activations and 8-bit weights
     """
-    extra_args: Dict[str, Any] = {"eps": 2**-12}
+    extra_args: Dict[str, Any] = {"eps": epsilon}
 
     # Setup observer/fake-quant for 16-bit activations
     if is_qat:

backends/arm/test/ops/test_sigmoid.py

@@ -34,6 +34,7 @@
     "zeros": lambda: torch.zeros(10, 10, 10, 10),
     "ones": lambda: torch.ones(10, 10, 10),
     "rand": lambda: torch.rand(10, 10) - 0.5,
+    "rand_4d": lambda: torch.rand(1, 1, 5, 10),
     "randn_pos": lambda: torch.randn(10) + 10,
     "randn_neg": lambda: torch.randn(10) - 10,
     "ramp": lambda: torch.arange(-16, 16, 0.2),
@@ -269,22 +270,23 @@ def get_symmetric_a16w8_sigmoid_quantizer(per_channel_quantization=False):
     }
 
     quantizer = TOSAQuantizer(tosa_profiles[tosa_version])
+
+    # Use a smaller episilon value to not greatly inflate [qmin, qmax]
     quantizer.set_global(
-        get_symmetric_a16w8_quantization_config(is_per_channel=per_channel_quantization)
+        get_symmetric_a16w8_quantization_config(
+            is_per_channel=per_channel_quantization, epsilon=2**-16
+        )
     )
 
     return Quantize(
         quantizer,
         get_symmetric_a16w8_quantization_config(
-            is_per_channel=per_channel_quantization
+            is_per_channel=per_channel_quantization, epsilon=2**-16
         ),
     )
 
 
 @common.parametrize("test_data", test_data_suite)
-@pytest.mark.xfail(
-    reason="missing int16 sigmoid ops support; fails at TOSA reference model with Unsupported operation type or rank. See: https://github.com/pytorch/executorch/issues/13974"
-)
 def test_sigmoid_16a8w_tosa_INT(test_data: torch.Tensor):
     """Test sigmoid operation with 16A8W quantization (16-bit activations, 8-bit weights)"""
     per_channel_quantization = False
@@ -311,7 +313,7 @@ def test_sigmoid_16a8w_tosa_INT(test_data: torch.Tensor):
 @common.parametrize("test_data", test_data_suite)
 @common.XfailIfNoCorstone300
 @pytest.mark.xfail(
-    reason="Vela compilation fails with 'Invalid arguments' for int16 sigmoid operations"
+    reason="MLETORCH-707: AssertionError: Output 0 does not match reference output."
 )
 def test_sigmoid_16a8w_u55_INT16(test_data: torch.Tensor):
     """Test sigmoid operation with 16A8W quantization on U55 (16-bit activations, 8-bit weights)"""
@@ -337,9 +339,6 @@ def test_sigmoid_16a8w_u55_INT16(test_data: torch.Tensor):
 
 @common.parametrize("test_data", test_data_suite)
 @common.XfailIfNoCorstone320
-@pytest.mark.xfail(
-    reason="Vela compilation fails with 'Invalid arguments' for int16 sigmoid operations"
-)
 def test_sigmoid_16a8w_u85_INT16(test_data: torch.Tensor):
     """Test sigmoid operation with 16A8W quantization on U85 (16-bit activations, 8-bit weights)"""
     per_channel_quantization = False

backends/arm/test/ops/test_sigmoid_16bit.py

@@ -121,22 +121,23 @@ def get_symmetric_a16w8_tanh_quantizer(per_channel_quantization=False):
     }
 
     quantizer = TOSAQuantizer(tosa_profiles[tosa_version])
+
+    # Use a smaller episilon value to not greatly inflate [qmin, qmax]
     quantizer.set_global(
-        get_symmetric_a16w8_quantization_config(is_per_channel=per_channel_quantization)
+        get_symmetric_a16w8_quantization_config(
+            is_per_channel=per_channel_quantization, epsilon=2**-16
+        )
     )
 
     return Quantize(
         quantizer,
         get_symmetric_a16w8_quantization_config(
-            is_per_channel=per_channel_quantization
+            is_per_channel=per_channel_quantization, epsilon=2**-16
         ),
     )
 
 
 @common.parametrize("test_data", test_data_suite)
-@pytest.mark.xfail(
-    reason="missing int16 tanh ops support; fails at TOSA reference model with Unsupported operation type or rank. See: https://github.com/pytorch/executorch/issues/13975"
-)
 def test_tanh_16a8w_tosa_INT(test_data: torch.Tensor):
     """Test tanh operation with 16A8W quantization (16-bit activations, 8-bit weights)"""
     per_channel_quantization = False
@@ -163,7 +164,7 @@ def test_tanh_16a8w_tosa_INT(test_data: torch.Tensor):
 @common.parametrize("test_data", test_data_suite)
 @common.XfailIfNoCorstone300
 @pytest.mark.xfail(
-    reason="Vela compilation fails with 'Invalid arguments' for int16 tanh operations"
+    reason="MLETORCH-707: AssertionError: Output 0 does not match reference output."
 )
 def test_tanh_16a8w_u55_INT16(test_data: torch.Tensor):
     """Test tanh operation with 16A8W quantization on U55 (16-bit activations, 8-bit weights)"""
@@ -189,9 +190,6 @@ def test_tanh_16a8w_u55_INT16(test_data: torch.Tensor):
 
 @common.parametrize("test_data", test_data_suite)
 @common.XfailIfNoCorstone320
-@pytest.mark.xfail(
-    reason="Vela compilation fails with 'Invalid arguments' for int16 tanh operations"
-)
 def test_tanh_16a8w_u85_INT16(test_data: torch.Tensor):
     """Test tanh operation with 16A8W quantization on U85 (16-bit activations, 8-bit weights)"""
     per_channel_quantization = False