[libcxx-commits] [libcxx] ab2ece0 - [libc++] Compute geomean in compare-benchmarks

[Louis Dionne via libcxx-commits]

Author: Louis Dionne
Date: 2025-10-17T17:05:58-04:00
New Revision: ab2ece03ae49877f0a8a9149b19fee3ee7326791

URL: https://github.com/llvm/llvm-project/commit/ab2ece03ae49877f0a8a9149b19fee3ee7326791
DIFF: https://github.com/llvm/llvm-project/commit/ab2ece03ae49877f0a8a9149b19fee3ee7326791.diff

LOG: [libc++] Compute geomean in compare-benchmarks

Added: 
    

Modified: 
    libcxx/utils/compare-benchmarks

Removed: 
    


################################################################################
diff  --git a/libcxx/utils/compare-benchmarks b/libcxx/utils/compare-benchmarks
index 988e2439e6269..88b03dcb97c3d 100755
--- a/libcxx/utils/compare-benchmarks
+++ b/libcxx/utils/compare-benchmarks
@@ -65,9 +65,16 @@ def plain_text_comparison(data, metric, baseline_name=None, candidate_name=None)
     """
     data = data.replace(numpy.nan, None) # avoid NaNs in tabulate output
     headers = ['Benchmark', baseline_name, candidate_name, 'Difference', '% Difference']
-    fmt = (None, '.2f', '.2f', '.2f', '.2f')
-    table = data[['benchmark', f'{metric}_0', f'{metric}_1', '
diff erence', 'percent']].set_index('benchmark')
-    return tabulate.tabulate(table, headers=headers, floatfmt=fmt, numalign='right')
+    fmt = (None, '.2f', '.2f', '.2f', '.2%')
+    table = data[['benchmark', f'{metric}_0', f'{metric}_1', '
diff erence', 'percent']]
+
+    # Compute the geomean and report on their 
diff erence
+    geomean_0 = statistics.geometric_mean(data[f'{metric}_0'])
+    geomean_1 = statistics.geometric_mean(data[f'{metric}_1'])
+    geomean_row = ['Geomean', geomean_0, geomean_1, (geomean_1 - geomean_0), (geomean_1 - geomean_0) / geomean_0]
+    table.loc[len(table)] = geomean_row
+
+    return tabulate.tabulate(table.set_index('benchmark'), headers=headers, floatfmt=fmt, numalign='right')
 
 def create_chart(data, metric, subtitle=None, series_names=None):
     """
@@ -154,7 +161,7 @@ def main(argv):
     # If we have exactly two data sets, compute additional info in new columns
     if len(lnt_inputs) == 2:
         data['
diff erence'] = data[f'{args.metric}_1'] - data[f'{args.metric}_0']
-        data['percent'] = 100 * (data['
diff erence'] / data[f'{args.metric}_0'])
+        data['percent'] = data['
diff erence'] / data[f'{args.metric}_0']
 
     if args.filter is not None:
         keeplist = [b for b in data['benchmark'] if re.search(args.filter, b) is not None]