Accelerate Chestnut model loading

openpilot/selfdrive/modeld/modeld.py

@@ -35,7 +35,7 @@ from openpilot.common.swaglog import cloudlog
 from openpilot.common.params import Params
 from openpilot.common.hardware.usb import cable_connected
 from openpilot.common.filter_simple import FirstOrderFilter
-from openpilot.common.realtime import config_realtime_process, DT_MDL
+from openpilot.common.realtime import config_realtime_process, set_core_affinity, DT_MDL
 from openpilot.common.transformations.camera import DEVICE_CAMERAS
 from openpilot.system.camerad.cameras.nv12_info import get_nv12_info
 from openpilot.common.transformations.model import get_warp_matrix
@@ -374,6 +374,7 @@ def main(demo=False):
         wait_for_chestnut()
         if not demo:
           wait_for_chestnut_power(CP, timeout=min(30., BIG_MODEL_TIMEOUT / 2))
+        set_core_affinity([7])
         if selection.big_path is not None:
           m, prepared = load_verified_model(vipc_client_main.width, vipc_client_main.height,
                                             big_path, selection.big_version, True, selection.big_id, selection.big_sha256)

openpilot/starpilot/flm/producer_clock_equivalence.py

@@ -18,7 +18,7 @@ PRODUCER_SOURCE_HASHES = (
   ('openpilot', 'system/camerad/cameras/spectra.cc', '4fa26c17f0fb792a4ab6815657beccddee1d641faa2fd02e5aa7393f717a2f3a'),
   ('openpilot', 'system/camerad/cameras/camera_qcom2.cc', 'b34c24c9984a0ac086f7b961cb7ff09d0501231da50aaad0e676bb3dc3f2edd9'),
   ('openpilot', 'system/camerad/cameras/camera_common.cc', 'bd9a416a6ee21047b4deec6df04276d8717cd92fe14de43f3e9daff86b62bb8f'),
-  ('openpilot', 'selfdrive/modeld/modeld.py', '3cbd9a9843716cf44832d6dd1cf4bef7fbc1facd02b9583b9b7db7e9abeb4c7b'),
+  ('openpilot', 'selfdrive/modeld/modeld.py', '75f050328e073bdc2476da0172645093a8aa74179b7ec24990414fc4ec98fe69'),
   ('openpilot', 'selfdrive/modeld/fill_model_msg.py', 'b3916c34d1073d0c26a7a8adc51e7b148f63a1ac68b388226170fd0f8b8b56b4'),
   ('openpilot', 'system/sensord/sensord.py', '7c79ab07d775622a43001e5a9b9ccc9a7bafb065940185849b33eefc2fd26420'),
   ('openpilot', 'system/sensord/sensors/lsm6ds3_gyro.py', '3e561f3d8572e1477c9b4b810c7570837d51e81c5672fe21ec67a68cae00dd92'),

openpilot/starpilot/flm/tests/test_upstream_producer_transport.py

@@ -8,7 +8,7 @@ from openpilot.starpilot.flm import producer_clock_equivalence as clock
 
 
 APPROVED = {
-  'selfdrive/modeld/modeld.py': '3cbd9a9843716cf44832d6dd1cf4bef7fbc1facd02b9583b9b7db7e9abeb4c7b',
+  'selfdrive/modeld/modeld.py': '75f050328e073bdc2476da0172645093a8aa74179b7ec24990414fc4ec98fe69',
   'visionipc/visionipc_server.cc': 'a4325a5171eacb931eb271d06a992d2f27216040df113d92099f62cdfa67d81a',
   'visionipc/visionipc_client.cc': 'd142cd60eedabfc581e9770fcd6f0ca81b762ba9dbc608b54dd20a27a317a52d',
 }

openpilot/starpilot/models/tests/test_loader_cpu_affinity.py

@@ -0,0 +1,88 @@
+"""Exercise the actual nested GPU loader without constructing a GPU context."""
+import ast
+from pathlib import Path
+from types import SimpleNamespace
+import threading
+
+import pytest
+
+
+SOURCE = Path(__file__).resolve().parents[3] / 'selfdrive/modeld/modeld.py'
+
+
+def loader(source, *, demo=False, bundled=False, failure=None):
+  tree = ast.parse(source.read_text())
+  main = next(n for n in tree.body if isinstance(n, ast.FunctionDef) and n.name == 'main')
+  function = next(n for n in ast.walk(main) if isinstance(n, ast.FunctionDef) and n.name == 'load_big')
+  calls = []
+
+  def record(name, result=None):
+    def run(*args, **kwargs):
+      calls.append((name, args, threading.get_ident()))
+      if name == failure:
+        raise OSError(name)
+      return result
+    return run
+
+  model = SimpleNamespace(warmup=record('warmup'))
+  namespace = {'demo': demo, 'CP': object(), 'BIG_MODEL_TIMEOUT': 120, 'wait_for_chestnut': record('usb'),
+                   'wait_for_chestnut_power': record('power'), 'set_core_affinity': record('affinity'),
+                   'selection': SimpleNamespace(big_path=None if bundled else Path('/verified.pkl'), big_version='v16',
+                                             big_id='mountaindewv3', big_sha256='a' * 64),
+                   'vipc_client_main': SimpleNamespace(width=1928, height=1208), 'big_path': Path('/verified.pkl'),
+                   'load_verified_model': record('verified', (model, 'prepared')), 'ModelState': record('bundled', model),
+                   'receipt_owner': SimpleNamespace(prepare=record('receipt', 'prepared')),
+                   'big_artifact_identity': object(), 'cloudlog': SimpleNamespace(exception=record('error'))}
+  outer = ast.parse('def outer():\n big_model = None\n big_prepared = None\n return load_big, lambda: (big_model, big_prepared)\n').body[0]
+  outer.body.insert(2, function)
+  module = ast.fix_missing_locations(ast.Module(body=[outer], type_ignores=[]))
+  exec(compile(module, str(source), 'exec'), namespace)
+  callback, state = namespace['outer']()
+  return callback, state, calls
+
+
+@pytest.mark.parametrize('demo,expected', [(False, ['usb', 'power', 'affinity', 'verified']),
+                                          (True, ['usb', 'affinity', 'verified'])])
+def test_actual_gpu_loader_pins_after_readiness_before_verified_load(demo, expected):
+  callback, state, calls = loader(SOURCE, demo=demo)
+  callback()
+  assert [row[0] for row in calls] == expected
+  assert next(row[1] for row in calls if row[0] == 'affinity') == ([7],)
+  assert state()[1] == 'prepared'
+
+
+@pytest.mark.parametrize('failure,expected', [('power', ['usb', 'power', 'error']),
+                                             ('affinity', ['usb', 'power', 'affinity', 'error'])])
+def test_failed_readiness_or_affinity_never_loads_a_model(failure, expected):
+  callback, state, calls = loader(SOURCE, failure=failure)
+  callback()
+  assert [row[0] for row in calls] == expected
+  assert state() == (None, None)
+
+
+def test_bundled_gpu_loader_keeps_warmup_and_receipt():
+  callback, state, calls = loader(SOURCE, bundled=True)
+  callback()
+  assert [row[0] for row in calls] == ['usb', 'power', 'affinity', 'bundled', 'warmup', 'receipt']
+  assert state()[1] == 'prepared'
+
+
+def test_affinity_is_applied_to_the_loader_thread_only():
+  callback, state, calls = loader(SOURCE)
+  main_thread = threading.get_ident()
+  worker = threading.Thread(target=callback)
+  worker.start()
+  worker.join(timeout=2)
+  assert not worker.is_alive()
+  assert state()[1] == 'prepared'
+  assert all(row[2] == worker.ident and row[2] != main_thread for row in calls)
+
+
+def test_no_fifo_priority_change_inside_gpu_loader():
+  tree = ast.parse(SOURCE.read_text())
+  main = next(n for n in tree.body if isinstance(n, ast.FunctionDef) and n.name == 'main')
+  callback = next(n for n in ast.walk(main) if isinstance(n, ast.FunctionDef) and n.name == 'load_big')
+  assert not any(isinstance(n, ast.Call) and isinstance(n.func, ast.Name) and n.func.id == 'config_realtime_process'
+                 for n in ast.walk(callback))
+  assert sum(isinstance(n, ast.Call) and isinstance(n.func, ast.Name) and n.func.id == 'config_realtime_process'
+             for n in ast.walk(main)) == 1

prebuilt.json

@@ -1 +1 @@
-{"protected_inventory_sha256": "0fb8cadb3c402848011c086cbf34c73c3325c804fa3de31309bc2cd0ae7b1b11", "schema_version": 1}
+{"protected_inventory_sha256": "2783830363fab32346bf9a66afcf51e6840c7116f139bc41d07230ca52723763", "schema_version": 1}