From 665eccc54386a10d44a0621f57ef3f2dc4e7bc30 Mon Sep 17 00:00:00 2001
From: Oleksandr Pavlyk <21087696+oleksandr-pavlyk@users.noreply.github.com>
Date: Thu, 25 Jun 2026 16:08:31 -0500
Subject: [PATCH] Reject negative values for noise

---
 python/scripts/nvbench_compare.py   | 14 +++++++-------
 python/test/test_nvbench_compare.py | 18 ++++++++++++++++++
 2 files changed, 25 insertions(+), 7 deletions(-)

diff --git a/python/scripts/nvbench_compare.py b/python/scripts/nvbench_compare.py
index 1217bdc..ccaec52 100644
--- a/python/scripts/nvbench_compare.py
+++ b/python/scripts/nvbench_compare.py
@@ -1676,11 +1676,11 @@ def compare_timings_for_bulk_same(ref_timing, cmp_timing, thresholds):
 
 
 def compare_timings_for_same(ref_timing, cmp_timing, ref_noise, cmp_noise, thresholds):
-    if not has_finite_noise(ref_noise) or not has_finite_noise(cmp_noise):
+    if not is_usable_noise(ref_noise) or not is_usable_noise(cmp_noise):
         return make_decision(
             ComparisonStatus.UNDECIDED,
             "noise_unavailable",
-            "relative dispersion is unavailable or non-finite",
+            "relative dispersion is unavailable, negative, or non-finite",
         )
     if max(ref_noise, cmp_noise) > thresholds.same_relative_dispersion_ceiling:
         return make_decision(
@@ -1850,7 +1850,7 @@ def compare_gpu_timings(ref_timing, cmp_timing, comparison_thresholds=None):
         diff_interval = compute_diff_interval(ref_interval, cmp_interval)
         frac_diff_interval = compute_frac_diff_interval(ref_interval, cmp_interval)
 
-    if not has_finite_noise(ref_noise) or not has_finite_noise(cmp_noise):
+    if not is_usable_noise(ref_noise) or not is_usable_noise(cmp_noise):
         max_noise = None
     else:
         max_noise = max(ref_noise, cmp_noise)
@@ -2488,8 +2488,8 @@ def align_timing_interval_columns(rows, comparisons, axis_count):
         )
 
 
-def has_finite_noise(noise):
-    return noise is not None and math.isfinite(noise)
+def is_usable_noise(noise):
+    return noise is not None and math.isfinite(noise) and noise >= 0.0
 
 
 def colorize_comparison_status(status, no_color):
@@ -2918,9 +2918,9 @@ def compare_benches(
 
                         start = None
                         for i, noise_value in enumerate(noise):
-                            if has_finite_noise(noise_value) and start is None:
+                            if is_usable_noise(noise_value) and start is None:
                                 start = i
-                            if not has_finite_noise(noise_value) and start is not None:
+                            if not is_usable_noise(noise_value) and start is not None:
                                 plot_confidence_band(start, i)
                                 start = None
 
diff --git a/python/test/test_nvbench_compare.py b/python/test/test_nvbench_compare.py
index 96e4a67..a3861e1 100644
--- a/python/test/test_nvbench_compare.py
+++ b/python/test/test_nvbench_compare.py
@@ -732,6 +732,24 @@ def test_compare_gpu_timings_classifies_common_cases(tmp_path, nvbench_compare):
     assert same.diff_interval == pytest.approx((-0.28, 0.28))
     assert same.frac_diff_interval == pytest.approx((-0.2153846154, 0.28))
 
+    negative_noise = nvbench_compare.compare_gpu_timings(
+        ref_interval_timing,
+        make_gpu_timing_data(
+            nvbench_compare,
+            minimum=1.02,
+            first_quartile=1.1,
+            median=1.204,
+            third_quartile=1.28,
+            mean=1.204,
+            interquartile_range_relative=-0.01,
+            sm_clock_rate_mean=100.0,
+        ),
+    )
+    assert negative_noise is not None
+    assert negative_noise.status == nvbench_compare.ComparisonStatus.UNDECIDED
+    assert negative_noise.max_noise is None
+    assert negative_noise.reason.code == "noise_unavailable"
+
     weak_overlap = nvbench_compare.compare_gpu_timings(
         make_gpu_timing_data(
             nvbench_compare,