model timing benchmark script

old-commit-hash: 5ba7276118

selfdrive/modeld/test/timing/benchmark.py

@@ -0,0 +1,36 @@
+#!/usr/bin/env python3
+import os
+import time
+import numpy as np
+import cereal.messaging as messaging
+import selfdrive.manager as manager
+
+
+N = int(os.getenv("N", "5"))
+TIME = int(os.getenv("TIME", "30"))
+
+if __name__ == "__main__":
+  sock = messaging.sub_sock('modelV2', conflate=False, timeout=1000)
+
+  execution_times = []
+
+  for _ in range(N):
+    os.environ['LOGPRINT'] = 'debug'
+    manager.start_managed_process('modeld')
+    time.sleep(5)
+
+    t = []
+    start = time.monotonic()
+    while time.monotonic() - start < TIME:
+      msgs = messaging.drain_sock(sock, wait_for_one=True)
+      for m in msgs:
+        t.append(m.modelV2.modelExecutionTime)
+
+    execution_times.append(np.array(t[10:]) * 1000)
+    manager.kill_managed_process('modeld')
+
+  print("\n\n")
+  print(f"ran modeld {N} times for {TIME}s each")
+  for n, t in enumerate(execution_times):
+    print(f"\tavg: {sum(t)/len(t):0.2f}ms, min: {min(t):0.2f}ms, max: {max(t):0.2f}ms")
+  print("\n\n")