From 9539ba2a781f2a97ae328f62bd0e9d4d0df89336 Mon Sep 17 00:00:00 2001
From: Oleksandr Pavlyk <21087696+oleksandr-pavlyk@users.noreply.github.com>
Date: Tue, 2 Jun 2026 16:21:39 -0500
Subject: [PATCH] Add "nv/cold/sm_clock_rate/mean" to GPU time summary data

Its intent is to be cheaply retrievable metric of average
SM clock frequence over entire sample
---
 python/scripts/nvbench_compare.py   |  3 +++
 python/test/test_nvbench_compare.py | 14 ++++++++++++++
 2 files changed, 17 insertions(+)

diff --git a/python/scripts/nvbench_compare.py b/python/scripts/nvbench_compare.py
index 276643d..3a80d24 100644
--- a/python/scripts/nvbench_compare.py
+++ b/python/scripts/nvbench_compare.py
@@ -40,6 +40,7 @@ GPU_TIME_STDEV_RELATIVE_TAG = "nv/cold/time/gpu/stdev/relative"
 GPU_TIME_MEDIAN_TAG = "nv/cold/time/gpu/median"
 GPU_TIME_IR_TAG = "nv/cold/time/gpu/ir/absolute"
 GPU_TIME_IR_RELATIVE_TAG = "nv/cold/time/gpu/ir/relative"
+GPU_SM_CLOCK_RATE_MEAN_TAG = "nv/cold/sm_clock_rate/mean"
 SAMPLE_TIMES_TAG = "nv/json/bin:nv/cold/sample_times"
 SAMPLE_FREQUENCIES_TAG = "nv/json/freqs-bin:nv/cold/sample_freqs"
 
@@ -81,6 +82,7 @@ class GpuTimingData:
     median: float | None
     interquartile_range: float | None
     interquartile_range_relative: float | None
+    sm_clock_rate_mean: float | None = None
     sample_source: Float32BinarySource | None = None
     frequency_source: Float32BinarySource | None = None
 
@@ -466,6 +468,7 @@ def extract_gpu_timing_data(summaries, json_dir=None, float32_reader=read_float3
         interquartile_range_relative=extract_summary_float(
             summaries, GPU_TIME_IR_RELATIVE_TAG, null_value=math.inf
         ),
+        sm_clock_rate_mean=extract_summary_float(summaries, GPU_SM_CLOCK_RATE_MEAN_TAG),
         sample_source=sample_source,
         frequency_source=frequency_source,
     )
diff --git a/python/test/test_nvbench_compare.py b/python/test/test_nvbench_compare.py
index b9cb478..be20141 100644
--- a/python/test/test_nvbench_compare.py
+++ b/python/test/test_nvbench_compare.py
@@ -119,6 +119,7 @@ def make_gpu_timing_data(
     median=None,
     interquartile_range=None,
     interquartile_range_relative=None,
+    sm_clock_rate_mean=None,
 ):
     return nvbench_compare.GpuTimingData(
         minimum=None,
@@ -129,6 +130,7 @@ def make_gpu_timing_data(
         median=median,
         interquartile_range=interquartile_range,
         interquartile_range_relative=interquartile_range_relative,
+        sm_clock_rate_mean=sm_clock_rate_mean,
     )
 
 
@@ -385,6 +387,18 @@ def test_gpu_timing_data_loads_samples_and_frequencies_lazily(
     assert reader_calls == [str(samples_file), str(freqs_file)]
 
 
+def test_gpu_timing_data_parses_sm_clock_rate_mean(nvbench_compare):
+    timing = nvbench_compare.extract_gpu_timing_data(
+        [
+            make_summary(nvbench_compare, "GPU_TIME_MEAN_TAG", "2.0"),
+            make_summary(nvbench_compare, "GPU_SM_CLOCK_RATE_MEAN_TAG", "1.5e9"),
+        ],
+    )
+
+    assert timing.sm_clock_rate_mean == pytest.approx(1.5e9)
+    assert timing.frequencies is None
+
+
 def test_gpu_timing_data_treats_mismatched_sample_and_frequency_counts_as_unavailable(
     tmp_path, nvbench_compare
 ):