Run DM model twice, prep for bigger models (#34894)

* twice

* swap

* adjust

selfdrive/modeld/dmonitoringmodeld.py

@@ -134,7 +134,7 @@ def get_driverstate_packet(model_output: np.ndarray, frame_id: int, location_ts:
 
 def main():
   setproctitle(PROCESS_NAME)
-  config_realtime_process([0, 1, 2, 3], 5)
+  config_realtime_process(7, 5)
 
   sentry.set_tag("daemon", PROCESS_NAME)
   cloudlog.bind(daemon=PROCESS_NAME)
@@ -173,6 +173,9 @@ def main():
     model_output, gpu_execution_time = model.run(buf, calib, model_transform)
     t2 = time.perf_counter()
 
+    # run one more time, just for the load
+    model.run(buf, calib, model_transform)
+
     pm.send("driverStateV2", get_driverstate_packet(model_output, vipc_client.frame_id, vipc_client.timestamp_sof, t2 - t1, gpu_execution_time))
 
 

selfdrive/test/test_onroad.py

@@ -391,8 +391,12 @@ class TestOnroad:
     result += "----------------- Model Timing -----------------\n"
     result += "------------------------------------------------\n"
     cfgs = [
-      ("modelV2", 0.045, 0.040),  # TODO: this should be stricter but it's hard to measure exactly
-      ("driverStateV2", 0.045, 0.035),
+      # since multiple processes use the GPU and can preempt each other,
+      # these numbers are not fully self-contained.
+      ("modelV2", 0.06, 0.040),
+
+      # can miss cycles here and there, just important the avg frequency is 20Hz
+      ("driverStateV2", 0.2, 0.05),
     ]
     for (s, instant_max, avg_max) in cfgs:
       ts = [getattr(m, s).modelExecutionTime for m in self.msgs[s]]

system/hardware/tici/tests/test_power_draw.py

@@ -33,7 +33,7 @@ class Proc:
 PROCS = [
   Proc(['camerad'], 1.75, msgs=['roadCameraState', 'wideRoadCameraState', 'driverCameraState']),
   Proc(['modeld'], 1.12, atol=0.2, msgs=['modelV2']),
-  Proc(['dmonitoringmodeld'], 0.6, msgs=['driverStateV2']),
+  Proc(['dmonitoringmodeld'], 1.4, msgs=['driverStateV2']),
   Proc(['encoderd'], 0.23, msgs=[]),
 ]