Add metric sorters for RCCL tests and rocHPCG

superbench/analyzer/metric_sorter.py

+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+"""Metric sort helpers for analyzer outputs.
+
+This module keeps benchmark-specific metric ordering isolated from the generic
+summary generation flow. Benchmarks without a registered sorter fall back to
+plain string ordering.
+"""
+
+import re
+
+
+_RCCL_PATTERN = re.compile(r'^(?P<bench>rccl-bw(?::[^/]+)?)/(?P<op>[^_]+)_(?P<size>\d+)_(?P<suffix>.+?)(?::\d+)?$')
+_HPCG_PATTERN = re.compile(r'^(?P<bench>gpu-hpcg(?::[^/]+)?)/(?P<metric>.+?)(?::\d+)?$')
+
+_HPCG_METRIC_ORDER = {
+    'local_domain_x': 0,
+    'local_domain_y': 1,
+    'local_domain_z': 2,
+    'global_domain_x': 3,
+    'global_domain_y': 4,
+    'global_domain_z': 5,
+    'process_domain_x': 6,
+    'process_domain_y': 7,
+    'process_domain_z': 8,
+    'total_time': 9,
+    'setup_time': 10,
+    'optimization_time': 11,
+    'ddot_gflops': 12,
+    'ddot_bandwidth': 13,
+    'ddot_gflops_per_process': 14,
+    'ddot_bandwidth_per_process': 15,
+    'waxpby_gflops': 16,
+    'waxpby_bandwidth': 17,
+    'waxpby_gflops_per_process': 18,
+    'waxpby_bandwidth_per_process': 19,
+    'spmv_gflops': 20,
+    'spmv_bandwidth': 21,
+    'spmv_gflops_per_process': 22,
+    'spmv_bandwidth_per_process': 23,
+    'mg_gflops': 24,
+    'mg_bandwidth': 25,
+    'mg_gflops_per_process': 26,
+    'mg_bandwidth_per_process': 27,
+    'total_gflops': 28,
+    'total_bandwidth': 29,
+    'total_gflops_per_process': 30,
+    'total_bandwidth_per_process': 31,
+    'final_gflops': 32,
+    'final_bandwidth': 33,
+    'final_gflops_per_process': 34,
+    'final_bandwidth_per_process': 35,
+    'is_valid': 36,
+}
+
+
+def _rccl_sort_key(metric_name):
+    """Sort RCCL metrics by benchmark, operation, then numeric message size."""
+    match = _RCCL_PATTERN.match(metric_name)
+    if not match:
+        return None
+
+    return (
+        0,
+        match.group('bench'),
+        match.group('op'),
+        int(match.group('size')),
+        match.group('suffix'),
+        metric_name,
+    )
+
+
+def _hpcg_sort_key(metric_name):
+    """Sort HPCG metrics roughly in the order they appear in rocHPCG logs."""
+    match = _HPCG_PATTERN.match(metric_name)
+    if not match:
+        return None
+
+    metric = match.group('metric')
+    return (
+        1,
+        match.group('bench'),
+        _HPCG_METRIC_ORDER.get(metric, 999),
+        metric,
+        metric_name,
+    )
+
+
+_SORTERS = (
+    _rccl_sort_key,
+    _hpcg_sort_key,
+)
+
+
+def sort_metrics(metrics):
+    """Sort metrics with benchmark-specific sorters and a stable default fallback."""
+
+    def sort_key(metric_name):
+        for sorter in _SORTERS:
+            key = sorter(metric_name)
+            if key is not None:
+                return key
+        return (999, metric_name)
+
+    return sorted(metrics, key=sort_key)

superbench/analyzer/result_summary.py

@@ -11,6 +11,7 @@
 from superbench.common.utils import logger
 from superbench.analyzer import file_handler
 from superbench.analyzer.summary_op import SummaryOp, SummaryType
+from superbench.analyzer.metric_sorter import sort_metrics
 from superbench.analyzer import RuleBase
 from superbench.analyzer import data_analysis
 
@@ -148,7 +149,7 @@ def _generate_summary(self, round):
                     # if aggregate is not empty and is a pattern in regex, aggregate according to pattern
                     else:
                         data_df_of_rule = data_analysis.aggregate(data_df_of_rule, self._sb_rules[rule]['aggregate'])
-                summary_df_of_rule = pd.DataFrame(columns=sorted(data_df_of_rule.columns))
+                summary_df_of_rule = pd.DataFrame(columns=sort_metrics(list(data_df_of_rule.columns)))
                 for statistic_name in statistics:
                     # get SummaryOp and calculate statistics
                     # if statistic_name is 'p\d\d?', SummaryOp should be pencentile