Arm backend: Avoid subtraction error on boolean tensors in error diff logging

backends/arm/test/tester/analyze_output_utils.py

@@ -154,6 +154,13 @@ def print_error_diffs(
         output_str += f"BATCH {n}\n"
         result_batch = result[n, :, :, :]
         reference_batch = reference[n, :, :, :]
+
+        if reference_batch.dtype == torch.bool or result_batch.dtype == torch.bool:
+            mismatches = (reference_batch != result_batch).sum().item()
+            total = reference_batch.numel()
+            output_str += f"(BOOLEAN tensor) {mismatches} / {total} elements differ ({mismatches / total:.2%})\n"
+            continue
+
         is_close = torch.allclose(result_batch, reference_batch, rtol, atol)
         if is_close:
             output_str += ".\n"
@@ -180,14 +187,15 @@ def print_error_diffs(
                 output_str += _print_elements(
                     result[n, :, :, :], reference[n, :, :, :], C, H, W, rtol, atol
                 )
-
-    reference_range = torch.max(reference) - torch.min(reference)
-    diff = torch.abs(reference - result).flatten()
-    diff = diff[diff.nonzero()]
-    if not len(diff) == 0:
-        diff_percent = diff / reference_range
-        output_str += "\nMEAN      MEDIAN    MAX       MIN    (error as % of reference output range)\n"
-        output_str += f"{torch.mean(diff_percent):<8.2%}  {torch.median(diff_percent):<8.2%}  {torch.max(diff_percent):<8.2%}  {torch.min(diff_percent):<8.2%}\n"
+    # Only compute numeric error metrics if tensor is not boolean
+    if reference.dtype != torch.bool and result.dtype != torch.bool:
+        reference_range = torch.max(reference) - torch.min(reference)
+        diff = torch.abs(reference - result).flatten()
+        diff = diff[diff.nonzero()]
+        if not len(diff) == 0:
+            diff_percent = diff / reference_range
+            output_str += "\nMEAN      MEDIAN    MAX       MIN    (error as % of reference output range)\n"
+            output_str += f"{torch.mean(diff_percent):<8.2%}  {torch.median(diff_percent):<8.2%}  {torch.max(diff_percent):<8.2%}  {torch.min(diff_percent):<8.2%}\n"
 
     # Over-engineer separators to match output width
     lines = output_str.split("\n")

backends/xnnpack/test/tester/tester.py

@@ -1,6 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
-# Copyright 2024-2025 Arm Limited and/or its affiliates.
 # All rights reserved.
+# Copyright 2024-2025 Arm Limited and/or its affiliates.
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
@@ -714,23 +714,30 @@ def _assert_outputs_equal(model_output, ref_output, atol=1e-03, rtol=1e-03):
             assert (
                 ref.shape == model.shape
             ), f"Output {i} shape {model.shape} does not match reference output shape {ref.shape}"
-            assert torch.allclose(
-                model,
-                ref,
-                atol=atol,
-                rtol=rtol,
-            ), (
-                f"Output {i} does not match reference output.\n"
-                f"\tGiven atol: {atol}, rtol: {rtol}.\n"
-                f"\tOutput tensor shape: {model.shape}, dtype: {model.dtype}\n"
-                f"\tDifference: max: {torch.max(model-ref)}, abs: {torch.max(torch.abs(model-ref))}, mean abs error: {torch.mean(torch.abs(model-ref))}.\n"
-                f"\t-- Model vs. Reference --\n"
-                f"\t Numel: {model.numel()}, {ref.numel()}\n"
-                f"\tMedian: {model.median()}, {ref.median()}\n"
-                f"\t  Mean: {model.mean()}, {ref.mean()}\n"
-                f"\t   Max: {model.max()}, {ref.max()}\n"
-                f"\t   Min: {model.min()}, {ref.min()}\n"
-            )
+            if model.dtype == torch.bool:
+                assert torch.equal(model, ref), (
+                    f"Output {i} (bool tensor) does not match reference output.\n"
+                    f"\tShape: {model.shape}\n"
+                    f"\tMismatched count: {(model != ref).sum().item()} / {model.numel()}\n"
+                )
+            else:
+                assert torch.allclose(
+                    model,
+                    ref,
+                    atol=atol,
+                    rtol=rtol,
+                ), (
+                    f"Output {i} does not match reference output.\n"
+                    f"\tGiven atol: {atol}, rtol: {rtol}.\n"
+                    f"\tOutput tensor shape: {model.shape}, dtype: {model.dtype}\n"
+                    f"\tDifference: max: {torch.max(model-ref)}, abs: {torch.max(torch.abs(model-ref))}, mean abs error: {torch.mean(torch.abs(model-ref))}.\n"
+                    f"\t-- Model vs. Reference --\n"
+                    f"\t Numel: {model.numel()}, {ref.numel()}\n"
+                    f"\tMedian: {model.median()}, {ref.median()}\n"
+                    f"\t  Mean: {model.mean()}, {ref.mean()}\n"
+                    f"\t   Max: {model.max()}, {ref.max()}\n"
+                    f"\t   Min: {model.min()}, {ref.min()}\n"
+                )
 
     @staticmethod
     def _compare_outputs(