Adjust tolerance for fp16 exp op to handle reasonable calculation discrepancies

Author: leafs1

backends/xnnpack/test/ops/test_exp.py

@@ -10,6 +10,23 @@
 from executorch.backends.xnnpack.test.tester import Tester
 
 
+def calculate_fp16_exp_tolerance(ref_output_tensor):
+    # Calculate mixed tolerance for float16 used in XNNPACK's float16 policy
+    fp16_epsilon = 9.77e-4
+    abs_tol = 2 * fp16_epsilon
+    rel_tol = 6 * fp16_epsilon
+
+    ref_abs = ref_output_tensor.abs()
+    mixed_tol = torch.maximum(
+        torch.full_like(ref_abs, abs_tol),
+        ref_abs * rel_tol,
+    )
+
+    final_atol = mixed_tol.max().item()
+
+    return final_atol, rel_tol
+
+
 class TestExp(unittest.TestCase):
     def setUp(self):
         torch._dynamo.reset()
@@ -22,6 +39,16 @@ def forward(self, x):
             return torch.exp(x)
 
     def run_exp_test(self, inputs):
+        input_tensor = inputs[0]
+
+        if input_tensor.dtype == torch.float16:
+            with torch.no_grad():
+                ref_output = torch.exp(input_tensor.to(torch.float32)).to(torch.float16)
+            atol, rtol = calculate_fp16_exp_tolerance(ref_output)
+        else:
+            atol = 1e-03
+            rtol = 1e-03
+
         (
             Tester(self.Exp(), inputs)
             .export()
@@ -31,12 +58,9 @@ def run_exp_test(self, inputs):
             .check_not(["executorch_exir_dialects_edge__ops_aten_exp_default"])
             .to_executorch()
             .serialize()
-            .run_method_and_compare_outputs()
+            .run_method_and_compare_outputs(atol=atol, rtol=rtol)
         )
 
-    # TODO (leafs1): Fix flaky tests. Land fix asap
-    # and cherry-pick onto release/0.7 branch
-    @unittest.skip(reason="For float16, numerical discepancies are too high")
     def test_fp16_exp(self):
         inputs = (torch.randn(20).to(torch.float16),)
         self.run_exp_test(inputs)

backends/xnnpack/test/ops/test_gelu.py

@@ -1,14 +1,20 @@
-# Copyright (c) Meta Platforms, Inc. and affiliates.
-# All rights reserved.
-#
-# This source code is licensed under the BSD-style license found in the
-# LICENSE file in the root directory of this source tree.
-
 import unittest
-
 import torch
 from executorch.backends.xnnpack.test.tester import Tester
 
+def calculate_fp16_gelu_tolerance(ref_output_tensor):
+    fp16_epsilon = 9.77e-4
+    abs_tol = 2 * fp16_epsilon
+    rel_tol = 6 * fp16_epsilon
+
+    ref_abs = ref_output_tensor.abs()
+    mixed_tol = torch.maximum(
+        torch.full_like(ref_abs, abs_tol),
+        ref_abs * rel_tol,
+    )
+
+    final_atol = mixed_tol.max().item()
+    return final_atol, rel_tol
 
 class TestGelu(unittest.TestCase):
     def setUp(self):
@@ -23,6 +29,16 @@ def forward(self, x):
             return self.gelu(x)
 
     def run_gelu_test(self, inputs):
+        input_tensor = inputs[0]
+
+        if input_tensor.dtype == torch.float16:
+            with torch.no_grad():
+                ref_output = torch.nn.functional.gelu(input_tensor.to(torch.float32)).to(torch.float16)
+            atol, rtol = calculate_fp16_gelu_tolerance(ref_output)
+        else:
+            atol = 1e-03
+            rtol = 1e-03
+
         (
             Tester(self.Gelu(), inputs)
             .export()
@@ -32,7 +48,7 @@ def run_gelu_test(self, inputs):
             .check_not(["executorch_exir_dialects_edge__ops_aten_gelu_default"])
             .to_executorch()
             .serialize()
-            .run_method_and_compare_outputs()
+            .run_method_and_compare_outputs(atol=atol, rtol=rtol)
         )
 
     def test_fp16_gelu(self):