wp: Add fps instrument analysis

Add new 'fps' metric to the processor for the fps instrument.
Add new 'frame_rendering_line' plotting function to the notebook
plotter.

Signed-off-by: Kajetan Puchalski <[email protected]>

wp/constants.py

@@ -5,8 +5,8 @@
 
 FULL_METRICS = [
     'power', 'idle', 'idle-miss', 'freq', 'overutil', 'pelt', 'capacity',
-    'uclamp', 'adpf', 'thermal', 'perf-trace-event', 'wakeup-latency',
-    'tasks-residency', 'tasks-activations',
+    'uclamp', 'adpf', 'thermal', 'fps', 'wakeup-latency',
+    'tasks-residency', 'tasks-activations', 'perf-trace-event',
     'cgroup-attach', 'wakeup-latency-cgroup', 'tasks-residency-cgroup',
     'energy-estimate',
 ]

wp/notebook.py

@@ -259,6 +259,7 @@ def analysis_to_loader(self, analysis: str) -> Callable:
             'task_activations_stats_cluster': self._load_task_activations_stats,
             'task_activations_stats_cluster_melt': self._load_task_activations_stats,
             'uclamp_updates': self._load_uclamp_updates,
+            'fps': self._load_fps,
         }
         return mapping[analysis]
 
@@ -1592,6 +1593,44 @@ def uclamp_per_task_line(self, height=600, width=1600, include_label=True,
         self.ana.hv_figures[self.ana._title_to_filename(title, '__line')] = layout
         return layout
 
+    # --- FPS instrument ---
+
+    def _load_fps(self):
+        self.ana.load_combined_analysis('fps.pqt', allow_missing=True)
+        log.info('Loaded fps into analysis')
+
+    @requires_analysis(['fps'])
+    def frame_rendering_line(self, height: int = 600, width: int = 1600, include_label: bool = True,
+                             target_fps: int = 30, title: str = 'Frame rendering times over time'):
+        """
+        Plot the frame rendering times over time overlaid between different runs, with a toggle to select the iteration.
+
+        :param target_fps: Target fps to draw the target rendering line from
+
+        .. note:: This plot uses HoloViews. The resulting figure will be saved to `WorkloadNotebookAnalysis.hv_figures`.
+        """
+        if include_label:
+            title = f"{self.ana.label} - {title}"
+
+        df = self.ana.analysis['fps']
+        df['target_render_time'] = 1000.0 / target_fps
+
+        ds = hv.Dataset(
+            df,
+            ['ts_iter', hv.Dimension('tag', values=self.ana.tags), 'iteration'],
+            ['render_time']
+        )
+
+        layout = ds.to(hv.Curve, 'ts_iter', 'render_time').overlay('tag').opts(
+            legend_position='bottom', title=title, framewise=True
+        )
+        layout *= hv.Curve(df, 'ts_iter', 'target_render_time').opts(color='orange')
+        layout.opts(
+            opts.Curve(height=height, width=width, interpolation='steps-pre', framewise=True),
+        )
+
+        return layout
+
     # -------- helper functions --------
 
     def gmean_bars(self, df: pd.DataFrame, x: str = 'stat', y: str = 'value', color: str = 'tag',

wp/processor.py

@@ -385,3 +385,31 @@ def capacity_analysis(self):
 
         df.write_parquet(os.path.join(self.analysis_path, 'capacity.pqt'))
         print(df)
+
+    def fps_analysis(self):
+        log.info('Collecting fps data')
+
+        def process_fps_df(df, iteration):
+            return df.filter(
+                pl.col('actual_present_time_us') != 0x7fffffffffffffff
+            ).with_columns(
+                pl.col('*') / 1000000.0
+            ).select(
+                pl.col('actual_present_time_us').alias('ts') / 1000.0,
+                (pl.col('actual_present_time_us') - pl.col('actual_present_time_us').first()).alias('ts_iter') / 1000.0,
+                pl.col('desired_present_time_us').alias('desired_ts') / 1000.0,
+                pl.col('frame_ready_time_us').alias('ready_ts') / 1000.0,
+                pl.col('actual_present_time_us').diff().alias('render_time'),
+                pl.col('frame_ready_time_us').diff().alias('render_ready_time'),
+                pl.lit(iteration).alias('iteration'),
+            )[1:]
+
+        fps = df_add_wa_output_tags(pl.concat([
+            process_fps_df(
+                pl.read_csv(job.get_artifact_path('frames')), job.iteration
+            )
+            for job in self.wa_output.jobs
+        ]), self.wa_output).sort('iteration')
+
+        fps.write_parquet(os.path.join(self.analysis_path, 'fps.pqt'))
+        print(fps)