Eliminate redundant compute_metrics() calls

TraceLens/TreePerf/gpu_event_analyser.py

@@ -226,9 +226,10 @@ def verify_dict_gpu_event_lists(dict_gpu_event_lists):
 
     @staticmethod
     def compute_metrics_dict(dict: dict, micro_idle_thresh_us=None):
-        dict_intervals = {}
-        for key, events in dict.items():
-            dict_intervals[key] = [(event["ts"], event["t_end"]) for event in events]
+        dict_intervals = {
+            key: GPUEventAnalyser.kernels_to_intervals(events)
+            for key, events in dict.items()
+        }
 
         # Merge intervals within each category.
         comp_union = GPUEventAnalyser.merge_intervals(dict_intervals["computation"])
@@ -321,6 +322,21 @@ def compute_metrics_dict(dict: dict, micro_idle_thresh_us=None):
                 "total_memcpy_time": total_memcpy_time,
             }
 
+    @staticmethod
+    def kernels_to_intervals(kernels, filter_func=None):
+        """Extract (ts, t_end) tuples from a kernel list, with an optional filter."""
+        return [
+            (k["ts"], k["t_end"])
+            for k in kernels
+            if filter_func is None or filter_func(k)
+        ]
+
+    @staticmethod
+    def compute_busy_time(intervals):
+        """Return the merged busy time (µs) for a list of (ts, t_end) tuples."""
+        merged = GPUEventAnalyser.merge_intervals(intervals)
+        return sum(end - start for start, end in merged)
+
     def compute_metrics(self, micro_idle_thresh_us=None):
         """
         Compute various metrics from the GPU event data.

TraceLens/TreePerf/tree_perf.py

@@ -391,22 +391,14 @@ def compute_perf_metrics(
         cpu_op_list = [self.tree.get_UID2event(uid) for uid in cpu_op_uids]
         _, list_kernelUIDS = self.loop_and_aggregate_kernels(cpu_op_list)
         list_kernels = [self.tree.events_by_uid[uid] for uid in list_kernelUIDS]
-        busy_kernel_time = 0
-        if len(list_kernels) > 0:
-            busy_kernel_time = self.GPUEventAnalyser(list_kernels).compute_metrics()[
-                "busy_time"
-            ]
-        _, list_non_data_mov_kernelUIDs = self.loop_and_aggregate_kernels(
-            cpu_op_list, filter_func=self.non_data_mov_filter
+        kernel_intervals = self.GPUEventAnalyser.kernels_to_intervals(list_kernels)
+        busy_kernel_time = self.GPUEventAnalyser.compute_busy_time(kernel_intervals)
+        non_data_mov_intervals = self.GPUEventAnalyser.kernels_to_intervals(
+            list_kernels, self.non_data_mov_filter
+        )
+        busy_non_data_mov_time = self.GPUEventAnalyser.compute_busy_time(
+            non_data_mov_intervals
         )
-        list_non_data_mov_kernels = [
-            self.tree.events_by_uid[uid] for uid in list_non_data_mov_kernelUIDs
-        ]
-        busy_non_data_mov_time = 0
-        if len(list_non_data_mov_kernels) > 0:
-            busy_non_data_mov_time = self.GPUEventAnalyser(
-                list_non_data_mov_kernels
-            ).compute_metrics()["busy_time"]
         event["kernel_details"] = [
             {
                 "name": kernel["name"],