FrogPilot 0.9.7

selfdrive/classic_modeld/SConscript

@@ -0,0 +1,74 @@
+import glob
+
+Import('env', 'envCython', 'arch', 'cereal', 'messaging', 'common', 'gpucommon', 'visionipc', 'transformations')
+lenv = env.Clone()
+lenvCython = envCython.Clone()
+
+libs = [cereal, messaging, visionipc, gpucommon, common, 'capnp', 'zmq', 'kj', 'pthread']
+frameworks = []
+
+common_src = [
+  "models/commonmodel.cc",
+  "transforms/loadyuv.cc",
+  "transforms/transform.cc",
+]
+
+thneed_src_common = [
+  "thneed/thneed_common.cc",
+  "thneed/serialize.cc",
+]
+
+thneed_src_qcom = thneed_src_common + ["thneed/thneed_qcom2.cc"]
+thneed_src_pc = thneed_src_common + ["thneed/thneed_pc.cc"]
+thneed_src = thneed_src_qcom if arch == "larch64" else thneed_src_pc
+
+# SNPE except on Mac and ARM Linux
+snpe_lib = []
+if arch != "Darwin" and arch != "aarch64":
+  common_src += ['runners/snpemodel.cc']
+  snpe_lib += ['SNPE']
+
+# OpenCL is a framework on Mac
+if arch == "Darwin":
+  frameworks += ['OpenCL']
+else:
+  libs += ['OpenCL']
+
+# Set path definitions
+for pathdef, fn in {'TRANSFORM': 'transforms/transform.cl', 'LOADYUV': 'transforms/loadyuv.cl'}.items():
+  for xenv in (lenv, lenvCython):
+    xenv['CXXFLAGS'].append(f'-D{pathdef}_PATH=\\"{File(fn).abspath}\\"')
+
+# Compile cython
+snpe_rpath_qcom = "/data/pythonpath/third_party/snpe/larch64"
+snpe_rpath_pc = f"{Dir('#').abspath}/third_party/snpe/x86_64-linux-clang"
+snpe_rpath = lenvCython['RPATH'] + [snpe_rpath_qcom if arch == "larch64" else snpe_rpath_pc]
+
+cython_libs = envCython["LIBS"] + libs
+snpemodel_lib = lenv.Library('snpemodel', ['runners/snpemodel.cc'])
+commonmodel_lib = lenv.Library('commonmodel', common_src)
+
+lenvCython.Program('runners/runmodel_pyx.so', 'runners/runmodel_pyx.pyx', LIBS=cython_libs, FRAMEWORKS=frameworks)
+lenvCython.Program('runners/snpemodel_pyx.so', 'runners/snpemodel_pyx.pyx', LIBS=[snpemodel_lib, snpe_lib, *cython_libs], FRAMEWORKS=frameworks, RPATH=snpe_rpath)
+lenvCython.Program('models/commonmodel_pyx.so', 'models/commonmodel_pyx.pyx', LIBS=[commonmodel_lib, *cython_libs], FRAMEWORKS=frameworks)
+
+tinygrad_files = ["#"+x for x in glob.glob(env.Dir("#tinygrad_repo").relpath + "/**", recursive=True, root_dir=env.Dir("#").abspath)]
+
+# Get model metadata
+fn = File("models/supercombo").abspath
+cmd = f'python3 {Dir("#selfdrive/classic_modeld").abspath}/get_model_metadata.py {fn}.onnx'
+lenv.Command(fn + "_metadata.pkl", [fn + ".onnx"] + tinygrad_files, cmd)
+
+# Build thneed model
+if arch == "larch64" or GetOption('pc_thneed'):
+  tinygrad_opts = []
+  if not GetOption('pc_thneed'):
+    # use FLOAT16 on device for speed + don't cache the CL kernels for space
+    tinygrad_opts += ["FLOAT16=1", "PYOPENCL_NO_CACHE=1"]
+  cmd = f"cd {Dir('#').abspath}/tinygrad_repo && " + ' '.join(tinygrad_opts) + f" python3 openpilot/compile2.py {fn}.onnx {fn}.thneed"
+
+  lenv.Command(fn + ".thneed", [fn + ".onnx"] + tinygrad_files, cmd)
+
+  thneed_lib = env.SharedLibrary('thneed', thneed_src, LIBS=[gpucommon, common, 'zmq', 'OpenCL', 'dl'])
+  thneedmodel_lib = env.Library('thneedmodel', ['runners/thneedmodel.cc'])
+  lenvCython.Program('runners/thneedmodel_pyx.so', 'runners/thneedmodel_pyx.pyx', LIBS=envCython["LIBS"]+[thneedmodel_lib, thneed_lib, gpucommon, common, 'dl', 'zmq', 'OpenCL'])

selfdrive/classic_modeld/classic_modeld

@@ -0,0 +1,10 @@
+#!/usr/bin/env bash
+
+DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" >/dev/null && pwd)"
+cd "$DIR/../../"
+
+if [ -f "$DIR/libthneed.so" ]; then
+  export LD_PRELOAD="$DIR/libthneed.so"
+fi
+
+exec "$DIR/classic_modeld.py" "$@"

selfdrive/classic_modeld/classic_modeld.py

@@ -0,0 +1,355 @@
+#!/usr/bin/env python3
+import os
+import time
+import pickle
+import numpy as np
+import cereal.messaging as messaging
+from cereal import car, log
+from pathlib import Path
+from setproctitle import setproctitle
+from cereal.messaging import PubMaster, SubMaster
+from msgq.visionipc import VisionIpcClient, VisionStreamType, VisionBuf
+from openpilot.common.swaglog import cloudlog
+from openpilot.common.params import Params
+from openpilot.common.filter_simple import FirstOrderFilter
+from openpilot.common.realtime import config_realtime_process
+from openpilot.common.transformations.camera import DEVICE_CAMERAS
+from openpilot.common.transformations.model import get_warp_matrix
+from openpilot.system import sentry
+from openpilot.selfdrive.car.car_helpers import get_demo_car_params
+from openpilot.selfdrive.controls.lib.desire_helper import DesireHelper
+from openpilot.selfdrive.classic_modeld.runners import ModelRunner, Runtime
+from openpilot.selfdrive.classic_modeld.parse_model_outputs import Parser
+from openpilot.selfdrive.classic_modeld.fill_model_msg import fill_model_msg, fill_pose_msg, PublishState
+from openpilot.selfdrive.classic_modeld.constants import ModelConstants
+from openpilot.selfdrive.classic_modeld.models.commonmodel_pyx import ModelFrame, CLContext
+
+from openpilot.selfdrive.frogpilot.frogpilot_variables import METADATAS_PATH, MODELS_PATH, get_frogpilot_toggles
+
+PROCESS_NAME = "selfdrive.classic_modeld.classic_modeld"
+SEND_RAW_PRED = os.getenv('SEND_RAW_PRED')
+
+MODEL_PATHS = {
+  ModelRunner.THNEED: Path(__file__).parent / 'models/supercombo.thneed',
+  ModelRunner.ONNX: Path(__file__).parent / 'models/supercombo.onnx'}
+
+class FrameMeta:
+  frame_id: int = 0
+  timestamp_sof: int = 0
+  timestamp_eof: int = 0
+
+  def __init__(self, vipc=None):
+    if vipc is not None:
+      self.frame_id, self.timestamp_sof, self.timestamp_eof = vipc.frame_id, vipc.timestamp_sof, vipc.timestamp_eof
+
+class ModelState:
+  frame: ModelFrame
+  wide_frame: ModelFrame
+  inputs: dict[str, np.ndarray]
+  output: np.ndarray
+  prev_desire: np.ndarray  # for tracking the rising edge of the pulse
+  model: ModelRunner
+
+  def __init__(self, context: CLContext, model: str, model_version: str):
+    # FrogPilot variables
+    MODEL_PATHS[ModelRunner.THNEED] = MODELS_PATH / f'{model}.thneed'
+
+    with open(METADATAS_PATH / f'supercombo_metadata_{model_version}.pkl', 'rb') as f:
+      model_metadata = pickle.load(f)
+
+    input_shapes = model_metadata.get('input_shapes')
+    self.navigation = 'nav_features' in input_shapes and 'nav_instructions' in input_shapes
+    self.radarless = 'radar_tracks' in input_shapes
+
+    self.frame = ModelFrame(context)
+    self.wide_frame = ModelFrame(context)
+    self.prev_desire = np.zeros(ModelConstants.DESIRE_LEN, dtype=np.float32)
+    self.inputs = {
+      'desire': np.zeros(ModelConstants.DESIRE_LEN * (ModelConstants.HISTORY_BUFFER_LEN+1), dtype=np.float32),
+      'traffic_convention': np.zeros(ModelConstants.TRAFFIC_CONVENTION_LEN, dtype=np.float32),
+      'lateral_control_params': np.zeros(ModelConstants.LATERAL_CONTROL_PARAMS_LEN, dtype=np.float32),
+      'prev_desired_curv': np.zeros(ModelConstants.PREV_DESIRED_CURV_LEN * (ModelConstants.HISTORY_BUFFER_LEN+1), dtype=np.float32),
+      **({'nav_features': np.zeros(ModelConstants.NAV_FEATURE_LEN, dtype=np.float32),
+          'nav_instructions': np.zeros(ModelConstants.NAV_INSTRUCTION_LEN, dtype=np.float32)} if self.navigation else {}),
+      'features_buffer': np.zeros(ModelConstants.HISTORY_BUFFER_LEN * ModelConstants.FEATURE_LEN, dtype=np.float32),
+      **({'radar_tracks': np.zeros(ModelConstants.RADAR_TRACKS_LEN * ModelConstants.RADAR_TRACKS_WIDTH, dtype=np.float32)} if self.radarless else {}),
+    }
+
+    self.output_slices = model_metadata['output_slices']
+    net_output_size = model_metadata['output_shapes']['outputs'][1]
+    self.output = np.zeros(net_output_size, dtype=np.float32)
+    self.parser = Parser()
+
+    self.model = ModelRunner(MODEL_PATHS, self.output, Runtime.GPU, False, context)
+    self.model.addInput("input_imgs", None)
+    self.model.addInput("big_input_imgs", None)
+    for k,v in self.inputs.items():
+      self.model.addInput(k, v)
+
+  def slice_outputs(self, model_outputs: np.ndarray) -> dict[str, np.ndarray]:
+    parsed_model_outputs = {k: model_outputs[np.newaxis, v] for k,v in self.output_slices.items()}
+    if SEND_RAW_PRED:
+      parsed_model_outputs['raw_pred'] = model_outputs.copy()
+    return parsed_model_outputs
+
+  def run(self, buf: VisionBuf, wbuf: VisionBuf, transform: np.ndarray, transform_wide: np.ndarray,
+                inputs: dict[str, np.ndarray], prepare_only: bool) -> dict[str, np.ndarray] | None:
+    # Model decides when action is completed, so desire input is just a pulse triggered on rising edge
+    inputs['desire'][0] = 0
+    self.inputs['desire'][:-ModelConstants.DESIRE_LEN] = self.inputs['desire'][ModelConstants.DESIRE_LEN:]
+    self.inputs['desire'][-ModelConstants.DESIRE_LEN:] = np.where(inputs['desire'] - self.prev_desire > .99, inputs['desire'], 0)
+    self.prev_desire[:] = inputs['desire']
+
+    self.inputs['traffic_convention'][:] = inputs['traffic_convention']
+    self.inputs['lateral_control_params'][:] = inputs['lateral_control_params']
+
+    if self.navigation:
+      self.inputs['nav_features'][:] = inputs['nav_features']
+      self.inputs['nav_instructions'][:] = inputs['nav_instructions']
+
+    if self.radarless:
+      self.inputs['radar_tracks'][:] = inputs['radar_tracks']
+
+    # if getCLBuffer is not None, frame will be None
+    self.model.setInputBuffer("input_imgs", self.frame.prepare(buf, transform.flatten(), self.model.getCLBuffer("input_imgs")))
+    if wbuf is not None:
+      self.model.setInputBuffer("big_input_imgs", self.wide_frame.prepare(wbuf, transform_wide.flatten(), self.model.getCLBuffer("big_input_imgs")))
+
+    if prepare_only:
+      return None
+
+    self.model.execute()
+    outputs = self.parser.parse_outputs(self.slice_outputs(self.output))
+
+    self.inputs['features_buffer'][:-ModelConstants.FEATURE_LEN] = self.inputs['features_buffer'][ModelConstants.FEATURE_LEN:]
+    self.inputs['features_buffer'][-ModelConstants.FEATURE_LEN:] = outputs['hidden_state'][0, :]
+    self.inputs['prev_desired_curv'][:-ModelConstants.PREV_DESIRED_CURV_LEN] = self.inputs['prev_desired_curv'][ModelConstants.PREV_DESIRED_CURV_LEN:]
+    self.inputs['prev_desired_curv'][-ModelConstants.PREV_DESIRED_CURV_LEN:] = outputs['desired_curvature'][0, :]
+    return outputs
+
+
+def main(demo=False):
+  # FrogPilot variables
+  frogpilot_toggles = get_frogpilot_toggles()
+
+  model_name = frogpilot_toggles.model
+  model_version = frogpilot_toggles.model_version
+
+  cloudlog.warning("classic_modeld init")
+
+  sentry.set_tag("daemon", PROCESS_NAME)
+  cloudlog.bind(daemon=PROCESS_NAME)
+  setproctitle(PROCESS_NAME)
+  config_realtime_process(7, 54)
+
+  cloudlog.warning("setting up CL context")
+  cl_context = CLContext()
+  cloudlog.warning("CL context ready; loading model")
+  model = ModelState(cl_context, model_name, model_version)
+  cloudlog.warning("models loaded, classic_modeld starting")
+
+  # visionipc clients
+  while True:
+    available_streams = VisionIpcClient.available_streams("camerad", block=False)
+    if available_streams:
+      use_extra_client = VisionStreamType.VISION_STREAM_WIDE_ROAD in available_streams and VisionStreamType.VISION_STREAM_ROAD in available_streams
+      main_wide_camera = VisionStreamType.VISION_STREAM_ROAD not in available_streams
+      break
+    time.sleep(.1)
+
+  vipc_client_main_stream = VisionStreamType.VISION_STREAM_WIDE_ROAD if main_wide_camera else VisionStreamType.VISION_STREAM_ROAD
+  vipc_client_main = VisionIpcClient("camerad", vipc_client_main_stream, True, cl_context)
+  vipc_client_extra = VisionIpcClient("camerad", VisionStreamType.VISION_STREAM_WIDE_ROAD, False, cl_context)
+  cloudlog.warning(f"vision stream set up, main_wide_camera: {main_wide_camera}, use_extra_client: {use_extra_client}")
+
+  while not vipc_client_main.connect(False):
+    time.sleep(0.1)
+  while use_extra_client and not vipc_client_extra.connect(False):
+    time.sleep(0.1)
+
+  cloudlog.warning(f"connected main cam with buffer size: {vipc_client_main.buffer_len} ({vipc_client_main.width} x {vipc_client_main.height})")
+  if use_extra_client:
+    cloudlog.warning(f"connected extra cam with buffer size: {vipc_client_extra.buffer_len} ({vipc_client_extra.width} x {vipc_client_extra.height})")
+
+  # messaging
+  pm = PubMaster(["modelV2", "cameraOdometry"])
+  sm = SubMaster(["deviceState", "carState", "roadCameraState", "liveCalibration", "driverMonitoringState", "navModel", "navInstruction", "carControl", "liveTracks", "frogpilotPlan"])
+
+  publish_state = PublishState()
+  params = Params()
+
+  # setup filter to track dropped frames
+  frame_dropped_filter = FirstOrderFilter(0., 10., 1. / ModelConstants.MODEL_FREQ)
+  frame_id = 0
+  last_vipc_frame_id = 0
+  run_count = 0
+
+  model_transform_main = np.zeros((3, 3), dtype=np.float32)
+  model_transform_extra = np.zeros((3, 3), dtype=np.float32)
+  live_calib_seen = False
+  nav_features = np.zeros(ModelConstants.NAV_FEATURE_LEN, dtype=np.float32)
+  nav_instructions = np.zeros(ModelConstants.NAV_INSTRUCTION_LEN, dtype=np.float32)
+  buf_main, buf_extra = None, None
+  meta_main = FrameMeta()
+  meta_extra = FrameMeta()
+
+
+  if demo:
+    CP = get_demo_car_params()
+  else:
+    with car.CarParams.from_bytes(params.get("CarParams", block=True)) as msg:
+      CP = msg
+  cloudlog.info("classic_modeld got CarParams: %s", CP.carName)
+
+  # TODO this needs more thought, use .2s extra for now to estimate other delays
+  steer_delay = CP.steerActuatorDelay + .2
+
+  DH = DesireHelper()
+
+  while True:
+    # Keep receiving frames until we are at least 1 frame ahead of previous extra frame
+    while meta_main.timestamp_sof < meta_extra.timestamp_sof + 25000000:
+      buf_main = vipc_client_main.recv()
+      meta_main = FrameMeta(vipc_client_main)
+      if buf_main is None:
+        break
+
+    if buf_main is None:
+      cloudlog.debug("vipc_client_main no frame")
+      continue
+
+    if use_extra_client:
+      # Keep receiving extra frames until frame id matches main camera
+      while True:
+        buf_extra = vipc_client_extra.recv()
+        meta_extra = FrameMeta(vipc_client_extra)
+        if buf_extra is None or meta_main.timestamp_sof < meta_extra.timestamp_sof + 25000000:
+          break
+
+      if buf_extra is None:
+        cloudlog.debug("vipc_client_extra no frame")
+        continue
+
+      if abs(meta_main.timestamp_sof - meta_extra.timestamp_sof) > 10000000:
+        cloudlog.error(f"frames out of sync! main: {meta_main.frame_id} ({meta_main.timestamp_sof / 1e9:.5f}),\
+                         extra: {meta_extra.frame_id} ({meta_extra.timestamp_sof / 1e9:.5f})")
+
+    else:
+      # Use single camera
+      buf_extra = buf_main
+      meta_extra = meta_main
+
+    sm.update(0)
+    desire = DH.desire
+    is_rhd = sm["driverMonitoringState"].isRHD
+    frame_id = sm["roadCameraState"].frameId
+    lateral_control_params = np.array([sm["carState"].vEgo, steer_delay], dtype=np.float32)
+    if sm.updated["liveCalibration"] and sm.seen['roadCameraState'] and sm.seen['deviceState']:
+      device_from_calib_euler = np.array(sm["liveCalibration"].rpyCalib, dtype=np.float32)
+      dc = DEVICE_CAMERAS[(str(sm['deviceState'].deviceType), str(sm['roadCameraState'].sensor))]
+      model_transform_main = get_warp_matrix(device_from_calib_euler, dc.ecam.intrinsics if main_wide_camera else dc.fcam.intrinsics, False).astype(np.float32)
+      model_transform_extra = get_warp_matrix(device_from_calib_euler, dc.ecam.intrinsics, True).astype(np.float32)
+      live_calib_seen = True
+
+    traffic_convention = np.zeros(2)
+    traffic_convention[int(is_rhd)] = 1
+
+    vec_desire = np.zeros(ModelConstants.DESIRE_LEN, dtype=np.float32)
+    if desire >= 0 and desire < ModelConstants.DESIRE_LEN:
+      vec_desire[desire] = 1
+
+    # Enable/disable nav features
+    timestamp_llk = sm["navModel"].locationMonoTime
+    nav_valid = sm.valid["navModel"] # and (nanos_since_boot() - timestamp_llk < 1e9)
+    nav_enabled = nav_valid and model.navigation
+
+    if not nav_enabled:
+      nav_features[:] = 0
+      nav_instructions[:] = 0
+
+    if nav_enabled and sm.updated["navModel"]:
+      nav_features = np.array(sm["navModel"].features)
+
+    if nav_enabled and sm.updated["navInstruction"]:
+      nav_instructions[:] = 0
+      for maneuver in sm["navInstruction"].allManeuvers:
+        distance_idx = 25 + int(maneuver.distance / 20)
+        direction_idx = 0
+        if maneuver.modifier in ("left", "slight left", "sharp left"):
+          direction_idx = 1
+        if maneuver.modifier in ("right", "slight right", "sharp right"):
+          direction_idx = 2
+        if 0 <= distance_idx < 50:
+          nav_instructions[distance_idx*3 + direction_idx] = 1
+
+    radar_tracks = np.zeros(ModelConstants.RADAR_TRACKS_LEN * ModelConstants.RADAR_TRACKS_WIDTH, dtype=np.float32)
+    if sm.updated["liveTracks"]:
+      for i, track in enumerate(sm["liveTracks"]):
+        if i >= ModelConstants.RADAR_TRACKS_LEN:
+          break
+        vec_index = i * ModelConstants.RADAR_TRACKS_WIDTH
+        radar_tracks[vec_index:vec_index+ModelConstants.RADAR_TRACKS_WIDTH] = [track.dRel, track.yRel, track.vRel]
+
+    # tracked dropped frames
+    vipc_dropped_frames = max(0, meta_main.frame_id - last_vipc_frame_id - 1)
+    frames_dropped = frame_dropped_filter.update(min(vipc_dropped_frames, 10))
+    if run_count < 10: # let frame drops warm up
+      frame_dropped_filter.x = 0.
+      frames_dropped = 0.
+    run_count = run_count + 1
+
+    frame_drop_ratio = frames_dropped / (1 + frames_dropped)
+    prepare_only = vipc_dropped_frames > 0
+    if prepare_only:
+      cloudlog.error(f"skipping model eval. Dropped {vipc_dropped_frames} frames")
+
+    inputs:dict[str, np.ndarray] = {
+      'desire': vec_desire,
+      'traffic_convention': traffic_convention,
+      'lateral_control_params': lateral_control_params,
+      **({'nav_features': nav_features, 'nav_instructions': nav_instructions} if model.navigation else {}),
+      **({'radar_tracks': radar_tracks} if model.radarless else {}),
+    }
+
+    mt1 = time.perf_counter()
+    model_output = model.run(buf_main, buf_extra, model_transform_main, model_transform_extra, inputs, prepare_only)
+    mt2 = time.perf_counter()
+    model_execution_time = mt2 - mt1
+
+    if model_output is not None:
+      modelv2_send = messaging.new_message('modelV2')
+      posenet_send = messaging.new_message('cameraOdometry')
+      fill_model_msg(modelv2_send, model_output, publish_state, meta_main.frame_id, meta_extra.frame_id, frame_id, frame_drop_ratio,
+                      meta_main.timestamp_eof, timestamp_llk, model_execution_time, live_calib_seen, nav_enabled)
+
+      desire_state = modelv2_send.modelV2.meta.desireState
+      l_lane_change_prob = desire_state[log.Desire.laneChangeLeft]
+      r_lane_change_prob = desire_state[log.Desire.laneChangeRight]
+      lane_change_prob = l_lane_change_prob + r_lane_change_prob
+      DH.update(sm['carState'], sm['carControl'].latActive, lane_change_prob, sm['frogpilotPlan'], frogpilot_toggles)
+      modelv2_send.modelV2.meta.laneChangeState = DH.lane_change_state
+      modelv2_send.modelV2.meta.laneChangeDirection = DH.lane_change_direction
+      modelv2_send.modelV2.meta.turnDirection = DH.turn_direction
+
+      fill_pose_msg(posenet_send, model_output, meta_main.frame_id, vipc_dropped_frames, meta_main.timestamp_eof, live_calib_seen)
+      pm.send('modelV2', modelv2_send)
+      pm.send('cameraOdometry', posenet_send)
+
+    last_vipc_frame_id = meta_main.frame_id
+
+    # Update FrogPilot parameters
+    if sm['frogpilotPlan'].togglesUpdated:
+      frogpilot_toggles = get_frogpilot_toggles()
+
+if __name__ == "__main__":
+  try:
+    import argparse
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--demo', action='store_true', help='A boolean for demo mode.')
+    args = parser.parse_args()
+    main(demo=args.demo)
+  except KeyboardInterrupt:
+    cloudlog.warning(f"child {PROCESS_NAME} got SIGINT")
+  except Exception:
+    sentry.capture_exception()
+    raise

selfdrive/classic_modeld/constants.py

@@ -0,0 +1,85 @@
+import numpy as np
+
+def index_function(idx, max_val=192, max_idx=32):
+  return (max_val) * ((idx/max_idx)**2)
+
+class ModelConstants:
+  # time and distance indices
+  IDX_N = 33
+  T_IDXS = [index_function(idx, max_val=10.0) for idx in range(IDX_N)]
+  X_IDXS = [index_function(idx, max_val=192.0) for idx in range(IDX_N)]
+  LEAD_T_IDXS = [0., 2., 4., 6., 8., 10.]
+  LEAD_T_OFFSETS = [0., 2., 4.]
+  META_T_IDXS = [2., 4., 6., 8., 10.]
+
+  # model inputs constants
+  MODEL_FREQ = 20
+  FEATURE_LEN = 512
+  HISTORY_BUFFER_LEN = 99
+  DESIRE_LEN = 8
+  TRAFFIC_CONVENTION_LEN = 2
+  NAV_FEATURE_LEN = 256
+  NAV_INSTRUCTION_LEN = 150
+  LAT_PLANNER_STATE_LEN = 4
+  LATERAL_CONTROL_PARAMS_LEN = 2
+  PREV_DESIRED_CURV_LEN = 1
+  RADAR_TRACKS_LEN = 64
+
+  # model outputs constants
+  FCW_THRESHOLDS_5MS2 = np.array([.05, .05, .15, .15, .15], dtype=np.float32)
+  FCW_THRESHOLDS_3MS2 = np.array([.7, .7], dtype=np.float32)
+  FCW_5MS2_PROBS_WIDTH = 5
+  FCW_3MS2_PROBS_WIDTH = 2
+
+  DISENGAGE_WIDTH = 5
+  POSE_WIDTH = 6
+  WIDE_FROM_DEVICE_WIDTH = 3
+  LEAD_WIDTH = 4
+  LANE_LINES_WIDTH = 2
+  ROAD_EDGES_WIDTH = 2
+  PLAN_WIDTH = 15
+  DESIRE_PRED_WIDTH = 8
+  LAT_PLANNER_SOLUTION_WIDTH = 4
+  DESIRED_CURV_WIDTH = 1
+  RADAR_TRACKS_WIDTH = 3
+
+  NUM_LANE_LINES = 4
+  NUM_ROAD_EDGES = 2
+
+  LEAD_TRAJ_LEN = 6
+  DESIRE_PRED_LEN = 4
+
+  PLAN_MHP_N = 5
+  LEAD_MHP_N = 2
+  PLAN_MHP_SELECTION = 1
+  LEAD_MHP_SELECTION = 3
+
+  FCW_THRESHOLD_5MS2_HIGH = 0.15
+  FCW_THRESHOLD_5MS2_LOW = 0.05
+  FCW_THRESHOLD_3MS2 = 0.7
+
+  CONFIDENCE_BUFFER_LEN = 5
+  RYG_GREEN = 0.01165
+  RYG_YELLOW = 0.06157
+
+# model outputs slices
+class Plan:
+  POSITION = slice(0, 3)
+  VELOCITY = slice(3, 6)
+  ACCELERATION = slice(6, 9)
+  T_FROM_CURRENT_EULER = slice(9, 12)
+  ORIENTATION_RATE = slice(12, 15)
+
+class Meta:
+  ENGAGED = slice(0, 1)
+  # next 2, 4, 6, 8, 10 seconds
+  GAS_DISENGAGE = slice(1, 36, 7)
+  BRAKE_DISENGAGE = slice(2, 36, 7)
+  STEER_OVERRIDE = slice(3, 36, 7)
+  HARD_BRAKE_3 = slice(4, 36, 7)
+  HARD_BRAKE_4 = slice(5, 36, 7)
+  HARD_BRAKE_5 = slice(6, 36, 7)
+  GAS_PRESS = slice(7, 36, 7)
+  # next 0, 2, 4, 6, 8, 10 seconds
+  LEFT_BLINKER = slice(36, 48, 2)
+  RIGHT_BLINKER = slice(37, 48, 2)

selfdrive/classic_modeld/dmonitoringmodeld.py

@@ -0,0 +1,156 @@
+#!/usr/bin/env python3
+import os
+import gc
+import math
+import time
+import ctypes
+import numpy as np
+from pathlib import Path
+
+from cereal import messaging
+from cereal.messaging import PubMaster, SubMaster
+from msgq.visionipc import VisionIpcClient, VisionStreamType, VisionBuf
+from openpilot.common.swaglog import cloudlog
+from openpilot.common.params import Params
+from openpilot.common.realtime import set_realtime_priority
+from openpilot.selfdrive.classic_modeld.runners import ModelRunner, Runtime
+from openpilot.selfdrive.classic_modeld.models.commonmodel_pyx import sigmoid
+
+CALIB_LEN = 3
+REG_SCALE = 0.25
+MODEL_WIDTH = 1440
+MODEL_HEIGHT = 960
+OUTPUT_SIZE = 84
+SEND_RAW_PRED = os.getenv('SEND_RAW_PRED')
+MODEL_PATHS = {
+  ModelRunner.SNPE: Path(__file__).parent / 'models/dmonitoring_model_q.dlc',
+  ModelRunner.ONNX: Path(__file__).parent / 'models/dmonitoring_model.onnx'}
+
+class DriverStateResult(ctypes.Structure):
+  _fields_ = [
+    ("face_orientation", ctypes.c_float*3),
+    ("face_position", ctypes.c_float*3),
+    ("face_orientation_std", ctypes.c_float*3),
+    ("face_position_std", ctypes.c_float*3),
+    ("face_prob", ctypes.c_float),
+    ("_unused_a", ctypes.c_float*8),
+    ("left_eye_prob", ctypes.c_float),
+    ("_unused_b", ctypes.c_float*8),
+    ("right_eye_prob", ctypes.c_float),
+    ("left_blink_prob", ctypes.c_float),
+    ("right_blink_prob", ctypes.c_float),
+    ("sunglasses_prob", ctypes.c_float),
+    ("occluded_prob", ctypes.c_float),
+    ("ready_prob", ctypes.c_float*4),
+    ("not_ready_prob", ctypes.c_float*2)]
+
+class DMonitoringModelResult(ctypes.Structure):
+  _fields_ = [
+    ("driver_state_lhd", DriverStateResult),
+    ("driver_state_rhd", DriverStateResult),
+    ("poor_vision_prob", ctypes.c_float),
+    ("wheel_on_right_prob", ctypes.c_float)]
+
+class ModelState:
+  inputs: dict[str, np.ndarray]
+  output: np.ndarray
+  model: ModelRunner
+
+  def __init__(self):
+    assert ctypes.sizeof(DMonitoringModelResult) == OUTPUT_SIZE * ctypes.sizeof(ctypes.c_float)
+    self.output = np.zeros(OUTPUT_SIZE, dtype=np.float32)
+    self.inputs = {
+      'input_img': np.zeros(MODEL_HEIGHT * MODEL_WIDTH, dtype=np.uint8),
+      'calib': np.zeros(CALIB_LEN, dtype=np.float32)}
+
+    self.model = ModelRunner(MODEL_PATHS, self.output, Runtime.DSP, True, None)
+    self.model.addInput("input_img", None)
+    self.model.addInput("calib", self.inputs['calib'])
+
+  def run(self, buf:VisionBuf, calib:np.ndarray) -> tuple[np.ndarray, float]:
+    self.inputs['calib'][:] = calib
+
+    v_offset = buf.height - MODEL_HEIGHT
+    h_offset = (buf.width - MODEL_WIDTH) // 2
+    buf_data = buf.data.reshape(-1, buf.stride)
+    input_data = self.inputs['input_img'].reshape(MODEL_HEIGHT, MODEL_WIDTH)
+    input_data[:] = buf_data[v_offset:v_offset+MODEL_HEIGHT, h_offset:h_offset+MODEL_WIDTH]
+
+    t1 = time.perf_counter()
+    self.model.setInputBuffer("input_img", self.inputs['input_img'].view(np.float32))
+    self.model.execute()
+    t2 = time.perf_counter()
+    return self.output, t2 - t1
+
+
+def fill_driver_state(msg, ds_result: DriverStateResult):
+  msg.faceOrientation = [x * REG_SCALE for x in ds_result.face_orientation]
+  msg.faceOrientationStd = [math.exp(x) for x in ds_result.face_orientation_std]
+  msg.facePosition = [x * REG_SCALE for x in ds_result.face_position[:2]]
+  msg.facePositionStd = [math.exp(x) for x in ds_result.face_position_std[:2]]
+  msg.faceProb = sigmoid(ds_result.face_prob)
+  msg.leftEyeProb = sigmoid(ds_result.left_eye_prob)
+  msg.rightEyeProb = sigmoid(ds_result.right_eye_prob)
+  msg.leftBlinkProb = sigmoid(ds_result.left_blink_prob)
+  msg.rightBlinkProb = sigmoid(ds_result.right_blink_prob)
+  msg.sunglassesProb = sigmoid(ds_result.sunglasses_prob)
+  msg.occludedProb = sigmoid(ds_result.occluded_prob)
+  msg.readyProb = [sigmoid(x) for x in ds_result.ready_prob]
+  msg.notReadyProb = [sigmoid(x) for x in ds_result.not_ready_prob]
+
+def get_driverstate_packet(model_output: np.ndarray, frame_id: int, location_ts: int, execution_time: float, dsp_execution_time: float):
+  model_result = ctypes.cast(model_output.ctypes.data, ctypes.POINTER(DMonitoringModelResult)).contents
+  msg = messaging.new_message('driverStateV2', valid=True)
+  ds = msg.driverStateV2
+  ds.frameId = frame_id
+  ds.modelExecutionTime = execution_time
+  ds.dspExecutionTime = dsp_execution_time
+  ds.poorVisionProb = sigmoid(model_result.poor_vision_prob)
+  ds.wheelOnRightProb = sigmoid(model_result.wheel_on_right_prob)
+  ds.rawPredictions = model_output.tobytes() if SEND_RAW_PRED else b''
+  fill_driver_state(ds.leftDriverData, model_result.driver_state_lhd)
+  fill_driver_state(ds.rightDriverData, model_result.driver_state_rhd)
+  return msg
+
+
+def main():
+  gc.disable()
+  set_realtime_priority(1)
+
+  model = ModelState()
+  cloudlog.warning("models loaded, dmonitoringmodeld starting")
+  Params().put_bool("DmModelInitialized", True)
+
+  cloudlog.warning("connecting to driver stream")
+  vipc_client = VisionIpcClient("camerad", VisionStreamType.VISION_STREAM_DRIVER, True)
+  while not vipc_client.connect(False):
+    time.sleep(0.1)
+  assert vipc_client.is_connected()
+  cloudlog.warning(f"connected with buffer size: {vipc_client.buffer_len}")
+
+  sm = SubMaster(["liveCalibration"])
+  pm = PubMaster(["driverStateV2"])
+
+  calib = np.zeros(CALIB_LEN, dtype=np.float32)
+  # last = 0
+
+  while True:
+    buf = vipc_client.recv()
+    if buf is None:
+      continue
+
+    sm.update(0)
+    if sm.updated["liveCalibration"]:
+      calib[:] = np.array(sm["liveCalibration"].rpyCalib)
+
+    t1 = time.perf_counter()
+    model_output, dsp_execution_time = model.run(buf, calib)
+    t2 = time.perf_counter()
+
+    pm.send("driverStateV2", get_driverstate_packet(model_output, vipc_client.frame_id, vipc_client.timestamp_sof, t2 - t1, dsp_execution_time))
+    # print("dmonitoring process: %.2fms, from last %.2fms\n" % (t2 - t1, t1 - last))
+    # last = t1
+
+
+if __name__ == "__main__":
+  main()

selfdrive/classic_modeld/fill_model_msg.py

@@ -0,0 +1,205 @@
+import os
+import capnp
+import numpy as np
+from cereal import log
+from openpilot.selfdrive.classic_modeld.constants import ModelConstants, Plan, Meta
+
+SEND_RAW_PRED = os.getenv('SEND_RAW_PRED')
+
+ConfidenceClass = log.ModelDataV2.ConfidenceClass
+
+class PublishState:
+  def __init__(self):
+    self.disengage_buffer = np.zeros(ModelConstants.CONFIDENCE_BUFFER_LEN*ModelConstants.DISENGAGE_WIDTH, dtype=np.float32)
+    self.prev_brake_5ms2_probs = np.zeros(ModelConstants.FCW_5MS2_PROBS_WIDTH, dtype=np.float32)
+    self.prev_brake_3ms2_probs = np.zeros(ModelConstants.FCW_3MS2_PROBS_WIDTH, dtype=np.float32)
+
+def fill_xyzt(builder, t, x, y, z, x_std=None, y_std=None, z_std=None):
+  builder.t = t
+  builder.x = x.tolist()
+  builder.y = y.tolist()
+  builder.z = z.tolist()
+  if x_std is not None:
+    builder.xStd = x_std.tolist()
+  if y_std is not None:
+    builder.yStd = y_std.tolist()
+  if z_std is not None:
+    builder.zStd = z_std.tolist()
+
+def fill_xyvat(builder, t, x, y, v, a, x_std=None, y_std=None, v_std=None, a_std=None):
+  builder.t = t
+  builder.x = x.tolist()
+  builder.y = y.tolist()
+  builder.v = v.tolist()
+  builder.a = a.tolist()
+  if x_std is not None:
+    builder.xStd = x_std.tolist()
+  if y_std is not None:
+    builder.yStd = y_std.tolist()
+  if v_std is not None:
+    builder.vStd = v_std.tolist()
+  if a_std is not None:
+    builder.aStd = a_std.tolist()
+
+def fill_model_msg(msg: capnp._DynamicStructBuilder, net_output_data: dict[str, np.ndarray], publish_state: PublishState,
+                   vipc_frame_id: int, vipc_frame_id_extra: int, frame_id: int, frame_drop: float,
+                   timestamp_eof: int, timestamp_llk: int, model_execution_time: float, valid: bool, nav_enabled: bool) -> None:
+  frame_age = frame_id - vipc_frame_id if frame_id > vipc_frame_id else 0
+  msg.valid = valid
+
+  modelV2 = msg.modelV2
+  modelV2.frameId = vipc_frame_id
+  modelV2.frameIdExtra = vipc_frame_id_extra
+  modelV2.frameAge = frame_age
+  modelV2.frameDropPerc = frame_drop * 100
+  modelV2.timestampEof = timestamp_eof
+  modelV2.modelExecutionTime = model_execution_time
+  modelV2.navEnabled = nav_enabled
+
+  # plan
+  position = modelV2.position
+  fill_xyzt(position, ModelConstants.T_IDXS, *net_output_data['plan'][0,:,Plan.POSITION].T, *net_output_data['plan_stds'][0,:,Plan.POSITION].T)
+  velocity = modelV2.velocity
+  fill_xyzt(velocity, ModelConstants.T_IDXS, *net_output_data['plan'][0,:,Plan.VELOCITY].T)
+  acceleration = modelV2.acceleration
+  fill_xyzt(acceleration, ModelConstants.T_IDXS, *net_output_data['plan'][0,:,Plan.ACCELERATION].T)
+  orientation = modelV2.orientation
+  fill_xyzt(orientation, ModelConstants.T_IDXS, *net_output_data['plan'][0,:,Plan.T_FROM_CURRENT_EULER].T)
+  orientation_rate = modelV2.orientationRate
+  fill_xyzt(orientation_rate, ModelConstants.T_IDXS, *net_output_data['plan'][0,:,Plan.ORIENTATION_RATE].T)
+
+  # lateral planning
+  action = modelV2.action
+  action.desiredCurvature = float(net_output_data['desired_curvature'][0,0])
+
+  # times at X_IDXS according to model plan
+  PLAN_T_IDXS = [np.nan] * ModelConstants.IDX_N
+  PLAN_T_IDXS[0] = 0.0
+  plan_x = net_output_data['plan'][0,:,Plan.POSITION][:,0].tolist()
+  for xidx in range(1, ModelConstants.IDX_N):
+    tidx = 0
+    # increment tidx until we find an element that's further away than the current xidx
+    while tidx < ModelConstants.IDX_N - 1 and plan_x[tidx+1] < ModelConstants.X_IDXS[xidx]:
+      tidx += 1
+    if tidx == ModelConstants.IDX_N - 1:
+      # if the Plan doesn't extend far enough, set plan_t to the max value (10s), then break
+      PLAN_T_IDXS[xidx] = ModelConstants.T_IDXS[ModelConstants.IDX_N - 1]
+      break
+    # interpolate to find `t` for the current xidx
+    current_x_val = plan_x[tidx]
+    next_x_val = plan_x[tidx+1]
+    p = (ModelConstants.X_IDXS[xidx] - current_x_val) / (next_x_val - current_x_val) if abs(next_x_val - current_x_val) > 1e-9 else float('nan')
+    PLAN_T_IDXS[xidx] = p * ModelConstants.T_IDXS[tidx+1] + (1 - p) * ModelConstants.T_IDXS[tidx]
+
+  # lane lines
+  modelV2.init('laneLines', 6)
+  lane_probs = net_output_data['lane_lines_prob'][0,1::2].tolist()
+  for i in range(6):
+    lane_line = modelV2.laneLines[i]
+    if i < 4:
+      fill_xyzt(lane_line, PLAN_T_IDXS, np.array(ModelConstants.X_IDXS), net_output_data['lane_lines'][0,i,:,0], net_output_data['lane_lines'][0,i,:,1])
+    elif i == 4:
+      if lane_probs[0] > 0:
+        leftLane_x = 0.5 * (net_output_data['lane_lines'][0,0,:,0] + net_output_data['lane_lines'][0,1,:,0])
+        leftLane_y = 0.5 * (net_output_data['lane_lines'][0,0,:,1] + net_output_data['lane_lines'][0,1,:,1])
+        fill_xyzt(lane_line, PLAN_T_IDXS, np.array(ModelConstants.X_IDXS), leftLane_x, leftLane_y)
+      else:
+        fill_xyzt(lane_line, PLAN_T_IDXS, np.empty((0,)), np.empty((0,)), np.empty((0,)))
+    elif i == 5:
+      if lane_probs[3] > 0:
+        rightLane_x = 0.5 * (net_output_data['lane_lines'][0,2,:,0] + net_output_data['lane_lines'][0,3,:,0])
+        rightLane_y = 0.5 * (net_output_data['lane_lines'][0,2,:,1] + net_output_data['lane_lines'][0,3,:,1])
+        fill_xyzt(lane_line, PLAN_T_IDXS, np.array(ModelConstants.X_IDXS), rightLane_x, rightLane_y)
+      else:
+        fill_xyzt(lane_line, PLAN_T_IDXS, np.empty((0,)), np.empty((0,)), np.empty((0,)))
+  modelV2.laneLineStds = net_output_data['lane_lines_stds'][0,:,0,0].tolist()
+  modelV2.laneLineProbs = lane_probs
+
+  # road edges
+  modelV2.init('roadEdges', 2)
+  for i in range(2):
+    road_edge = modelV2.roadEdges[i]
+    fill_xyzt(road_edge, PLAN_T_IDXS, np.array(ModelConstants.X_IDXS), net_output_data['road_edges'][0,i,:,0], net_output_data['road_edges'][0,i,:,1])
+  modelV2.roadEdgeStds = net_output_data['road_edges_stds'][0,:,0,0].tolist()
+
+  # leads
+  modelV2.init('leadsV3', 3)
+  for i in range(3):
+    lead = modelV2.leadsV3[i]
+    fill_xyvat(lead, ModelConstants.LEAD_T_IDXS, *net_output_data['lead'][0,i].T, *net_output_data['lead_stds'][0,i].T)
+    lead.prob = net_output_data['lead_prob'][0,i].tolist()
+    lead.probTime = ModelConstants.LEAD_T_OFFSETS[i]
+
+  # meta
+  meta = modelV2.meta
+  meta.desireState = net_output_data['desire_state'][0].reshape(-1).tolist()
+  meta.desirePrediction = net_output_data['desire_pred'][0].reshape(-1).tolist()
+  meta.engagedProb = net_output_data['meta'][0,Meta.ENGAGED].item()
+  meta.init('disengagePredictions')
+  disengage_predictions = meta.disengagePredictions
+  disengage_predictions.t = ModelConstants.META_T_IDXS
+  disengage_predictions.brakeDisengageProbs = net_output_data['meta'][0,Meta.BRAKE_DISENGAGE].tolist()
+  disengage_predictions.gasDisengageProbs = net_output_data['meta'][0,Meta.GAS_DISENGAGE].tolist()
+  disengage_predictions.steerOverrideProbs = net_output_data['meta'][0,Meta.STEER_OVERRIDE].tolist()
+  disengage_predictions.brake3MetersPerSecondSquaredProbs = net_output_data['meta'][0,Meta.HARD_BRAKE_3].tolist()
+  disengage_predictions.brake4MetersPerSecondSquaredProbs = net_output_data['meta'][0,Meta.HARD_BRAKE_4].tolist()
+  disengage_predictions.brake5MetersPerSecondSquaredProbs = net_output_data['meta'][0,Meta.HARD_BRAKE_5].tolist()
+
+  publish_state.prev_brake_5ms2_probs[:-1] = publish_state.prev_brake_5ms2_probs[1:]
+  publish_state.prev_brake_5ms2_probs[-1] = net_output_data['meta'][0,Meta.HARD_BRAKE_5][0]
+  publish_state.prev_brake_3ms2_probs[:-1] = publish_state.prev_brake_3ms2_probs[1:]
+  publish_state.prev_brake_3ms2_probs[-1] = net_output_data['meta'][0,Meta.HARD_BRAKE_3][0]
+  hard_brake_predicted = (publish_state.prev_brake_5ms2_probs > ModelConstants.FCW_THRESHOLDS_5MS2).all() and \
+    (publish_state.prev_brake_3ms2_probs > ModelConstants.FCW_THRESHOLDS_3MS2).all()
+  meta.hardBrakePredicted = hard_brake_predicted.item()
+
+  # temporal pose
+  temporal_pose = modelV2.temporalPose
+  temporal_pose.trans = net_output_data['sim_pose'][0,:ModelConstants.POSE_WIDTH//2].tolist()
+  temporal_pose.transStd = net_output_data['sim_pose_stds'][0,:ModelConstants.POSE_WIDTH//2].tolist()
+  temporal_pose.rot = net_output_data['sim_pose'][0,ModelConstants.POSE_WIDTH//2:].tolist()
+  temporal_pose.rotStd = net_output_data['sim_pose_stds'][0,ModelConstants.POSE_WIDTH//2:].tolist()
+
+  # confidence
+  if vipc_frame_id % (2*ModelConstants.MODEL_FREQ) == 0:
+    # any disengage prob
+    brake_disengage_probs = net_output_data['meta'][0,Meta.BRAKE_DISENGAGE]
+    gas_disengage_probs = net_output_data['meta'][0,Meta.GAS_DISENGAGE]
+    steer_override_probs = net_output_data['meta'][0,Meta.STEER_OVERRIDE]
+    any_disengage_probs = 1-((1-brake_disengage_probs)*(1-gas_disengage_probs)*(1-steer_override_probs))
+    # independent disengage prob for each 2s slice
+    ind_disengage_probs = np.r_[any_disengage_probs[0], np.diff(any_disengage_probs) / (1 - any_disengage_probs[:-1])]
+    # rolling buf for 2, 4, 6, 8, 10s
+    publish_state.disengage_buffer[:-ModelConstants.DISENGAGE_WIDTH] = publish_state.disengage_buffer[ModelConstants.DISENGAGE_WIDTH:]
+    publish_state.disengage_buffer[-ModelConstants.DISENGAGE_WIDTH:] = ind_disengage_probs
+
+  score = 0.
+  for i in range(ModelConstants.DISENGAGE_WIDTH):
+    score += publish_state.disengage_buffer[i*ModelConstants.DISENGAGE_WIDTH+ModelConstants.DISENGAGE_WIDTH-1-i].item() / ModelConstants.DISENGAGE_WIDTH
+  if score < ModelConstants.RYG_GREEN:
+    modelV2.confidence = ConfidenceClass.green
+  elif score < ModelConstants.RYG_YELLOW:
+    modelV2.confidence = ConfidenceClass.yellow
+  else:
+    modelV2.confidence = ConfidenceClass.red
+
+  # raw prediction if enabled
+  if SEND_RAW_PRED:
+    modelV2.rawPredictions = net_output_data['raw_pred'].tobytes()
+
+def fill_pose_msg(msg: capnp._DynamicStructBuilder, net_output_data: dict[str, np.ndarray],
+                  vipc_frame_id: int, vipc_dropped_frames: int, timestamp_eof: int, live_calib_seen: bool) -> None:
+  msg.valid = live_calib_seen & (vipc_dropped_frames < 1)
+  cameraOdometry = msg.cameraOdometry
+
+  cameraOdometry.frameId = vipc_frame_id
+  cameraOdometry.timestampEof = timestamp_eof
+
+  cameraOdometry.trans = net_output_data['pose'][0,:3].tolist()
+  cameraOdometry.rot = net_output_data['pose'][0,3:].tolist()
+  cameraOdometry.wideFromDeviceEuler = net_output_data['wide_from_device_euler'][0,:].tolist()
+  cameraOdometry.roadTransformTrans = net_output_data['road_transform'][0,:3].tolist()
+  cameraOdometry.transStd = net_output_data['pose_stds'][0,:3].tolist()
+  cameraOdometry.rotStd = net_output_data['pose_stds'][0,3:].tolist()
+  cameraOdometry.wideFromDeviceEulerStd = net_output_data['wide_from_device_euler_stds'][0,:].tolist()
+  cameraOdometry.roadTransformTransStd = net_output_data['road_transform_stds'][0,:3].tolist()

selfdrive/classic_modeld/get_model_metadata.py

@@ -0,0 +1,28 @@
+#!/usr/bin/env python3
+import sys
+import pathlib
+import onnx
+import codecs
+import pickle
+
+def get_name_and_shape(value_info:onnx.ValueInfoProto) -> tuple[str, tuple[int,...]]:
+  shape = tuple([int(dim.dim_value) for dim in value_info.type.tensor_type.shape.dim])
+  name = value_info.name
+  return name, shape
+
+if __name__ == "__main__":
+  model_path = pathlib.Path(sys.argv[1])
+  model = onnx.load(str(model_path))
+  i = [x.key for x in model.metadata_props].index('output_slices')
+  output_slices = model.metadata_props[i].value
+
+  metadata = {}
+  metadata['output_slices'] = pickle.loads(codecs.decode(output_slices.encode(), "base64"))
+  metadata['input_shapes'] = dict([get_name_and_shape(x) for x in model.graph.input])
+  metadata['output_shapes'] = dict([get_name_and_shape(x) for x in model.graph.output])
+
+  metadata_path = model_path.parent / (model_path.stem + '_metadata.pkl')
+  with open(metadata_path, 'wb') as f:
+    pickle.dump(metadata, f)
+
+  print(f'saved metadata to {metadata_path}')

selfdrive/classic_modeld/models/README.md

@@ -0,0 +1,62 @@
+## Neural networks in openpilot
+To view the architecture of the ONNX networks, you can use [netron](https://netron.app/)
+
+## Supercombo
+### Supercombo input format (Full size: 799906 x float32)
+* **image stream**
+  * Two consecutive images (256 * 512 * 3 in RGB) recorded at 20 Hz : 393216 = 2 * 6 * 128 * 256
+    * Each 256 * 512 image is represented in YUV420 with 6 channels : 6 * 128 * 256
+      * Channels 0,1,2,3 represent the full-res Y channel and are represented in numpy as Y[::2, ::2], Y[::2, 1::2], Y[1::2, ::2], and Y[1::2, 1::2]
+      * Channel 4 represents the half-res U channel
+      * Channel 5 represents the half-res V channel
+* **wide image stream**
+  * Two consecutive images (256 * 512 * 3 in RGB) recorded at 20 Hz : 393216 = 2 * 6 * 128 * 256
+    * Each 256 * 512 image is represented in YUV420 with 6 channels : 6 * 128 * 256
+      * Channels 0,1,2,3 represent the full-res Y channel and are represented in numpy as Y[::2, ::2], Y[::2, 1::2], Y[1::2, ::2], and Y[1::2, 1::2]
+      * Channel 4 represents the half-res U channel
+      * Channel 5 represents the half-res V channel
+* **desire**
+  * one-hot encoded buffer to command model to execute certain actions, bit needs to be sent for the past 5 seconds (at 20FPS) : 100 * 8
+* **traffic convention**
+  * one-hot encoded vector to tell model whether traffic is right-hand or left-hand traffic : 2
+* **feature buffer**
+  * A buffer of intermediate features that gets appended to the current feature to form a 5 seconds temporal context (at 20FPS) : 99 * 512
+
+
+### Supercombo output format (Full size: XXX x float32)
+Read [here](https://github.com/commaai/openpilot/blob/90af436a121164a51da9fa48d093c29f738adf6a/selfdrive/classic_modeld/models/driving.h#L236) for more.
+
+
+## Driver Monitoring Model
+* .onnx model can be run with onnx runtimes
+* .dlc file is a pre-quantized model and only runs on qualcomm DSPs
+
+### input format
+* single image W = 1440 H = 960 luminance channel (Y) from the planar YUV420 format:
+  * full input size is 1440 * 960 = 1382400
+  * normalized ranging from 0.0 to 1.0 in float32 (onnx runner) or ranging from 0 to 255 in uint8 (snpe runner)
+* camera calibration angles (roll, pitch, yaw) from liveCalibration: 3 x float32 inputs
+
+### output format
+* 84 x float32 outputs = 2 + 41 * 2 ([parsing example](https://github.com/commaai/openpilot/blob/22ce4e17ba0d3bfcf37f8255a4dd1dc683fe0c38/selfdrive/classic_modeld/models/dmonitoring.cc#L33))
+  * for each person in the front seats (2 * 41)
+    * face pose: 12 = 6 + 6
+      * face orientation [pitch, yaw, roll] in camera frame: 3
+      * face position [dx, dy] relative to image center: 2
+      * normalized face size: 1
+      * standard deviations for above outputs: 6
+    * face visible probability: 1
+    * eyes: 20 = (8 + 1) + (8 + 1) + 1 + 1
+      * eye position and size, and their standard deviations: 8
+      * eye visible probability: 1
+      * eye closed probability: 1
+    * wearing sunglasses probability: 1
+    * face occluded probability: 1
+    * touching wheel probability: 1
+    * paying attention probability: 1
+    * (deprecated) distracted probabilities: 2
+    * using phone probability: 1
+    * distracted probability: 1
+  * common outputs 2
+    * poor camera vision probability: 1
+    * left hand drive probability: 1

selfdrive/classic_modeld/models/commonmodel.pxd

@@ -0,0 +1,20 @@
+# distutils: language = c++
+
+from msgq.visionipc.visionipc cimport cl_device_id, cl_context, cl_mem
+
+cdef extern from "common/mat.h":
+  cdef struct mat3:
+    float v[9]
+
+cdef extern from "common/clutil.h":
+  cdef unsigned long CL_DEVICE_TYPE_DEFAULT
+  cl_device_id cl_get_device_id(unsigned long)
+  cl_context cl_create_context(cl_device_id)
+
+cdef extern from "selfdrive/classic_modeld/models/commonmodel.h":
+  float sigmoid(float)
+
+  cppclass ModelFrame:
+    int buf_size
+    ModelFrame(cl_device_id, cl_context)
+    float * prepare(cl_mem, int, int, int, int, mat3, cl_mem*)

selfdrive/classic_modeld/models/commonmodel_pyx.pxd

@@ -0,0 +1,13 @@
+# distutils: language = c++
+
+from msgq.visionipc.visionipc cimport cl_mem
+from msgq.visionipc.visionipc_pyx cimport CLContext as BaseCLContext
+
+cdef class CLContext(BaseCLContext):
+  pass
+
+cdef class CLMem:
+  cdef cl_mem * mem
+
+  @staticmethod
+  cdef create(void*)

selfdrive/classic_modeld/models/commonmodel_pyx.pyx

@@ -0,0 +1,47 @@
+# distutils: language = c++
+# cython: c_string_encoding=ascii
+
+import numpy as np
+cimport numpy as cnp
+from libc.string cimport memcpy
+
+from msgq.visionipc.visionipc cimport cl_mem
+from msgq.visionipc.visionipc_pyx cimport VisionBuf, CLContext as BaseCLContext
+from .commonmodel cimport CL_DEVICE_TYPE_DEFAULT, cl_get_device_id, cl_create_context
+from .commonmodel cimport mat3, sigmoid as cppSigmoid, ModelFrame as cppModelFrame
+
+def sigmoid(x):
+  return cppSigmoid(x)
+
+cdef class CLContext(BaseCLContext):
+  def __cinit__(self):
+    self.device_id = cl_get_device_id(CL_DEVICE_TYPE_DEFAULT)
+    self.context = cl_create_context(self.device_id)
+
+cdef class CLMem:
+  @staticmethod
+  cdef create(void * cmem):
+    mem = CLMem()
+    mem.mem = <cl_mem*> cmem
+    return mem
+
+cdef class ModelFrame:
+  cdef cppModelFrame * frame
+
+  def __cinit__(self, CLContext context):
+    self.frame = new cppModelFrame(context.device_id, context.context)
+
+  def __dealloc__(self):
+    del self.frame
+
+  def prepare(self, VisionBuf buf, float[:] projection, CLMem output):
+    cdef mat3 cprojection
+    memcpy(cprojection.v, &projection[0], 9*sizeof(float))
+    cdef float * data
+    if output is None:
+      data = self.frame.prepare(buf.buf.buf_cl, buf.width, buf.height, buf.stride, buf.uv_offset, cprojection, NULL)
+    else:
+      data = self.frame.prepare(buf.buf.buf_cl, buf.width, buf.height, buf.stride, buf.uv_offset, cprojection, output.mem)
+    if not data:
+      return None
+    return np.asarray(<cnp.float32_t[:self.frame.buf_size]> data)

selfdrive/classic_modeld/models/dmonitoring_model.current

@@ -0,0 +1,2 @@
+5ec97a39-0095-4cea-adfa-6d72b1966cc1
+26cac7a9757a27c783a365403040a1bd27ccdaea

selfdrive/classic_modeld/navmodeld.py

@@ -0,0 +1,111 @@
+#!/usr/bin/env python3
+import gc
+import math
+import time
+import ctypes
+import numpy as np
+from pathlib import Path
+
+from cereal import messaging
+from cereal.messaging import PubMaster, SubMaster
+from msgq.visionipc import VisionIpcClient, VisionStreamType
+from openpilot.common.swaglog import cloudlog
+from openpilot.common.params import Params
+from openpilot.common.realtime import set_realtime_priority
+from openpilot.selfdrive.classic_modeld.constants import ModelConstants
+from openpilot.selfdrive.classic_modeld.runners import ModelRunner, Runtime
+
+NAV_INPUT_SIZE = 256*256
+NAV_FEATURE_LEN = 256
+NAV_DESIRE_LEN = 32
+NAV_OUTPUT_SIZE = 2*2*ModelConstants.IDX_N + NAV_DESIRE_LEN + NAV_FEATURE_LEN
+MODEL_PATHS = {
+  ModelRunner.SNPE: Path(__file__).parent / 'models/navmodel_q.dlc',
+  ModelRunner.ONNX: Path(__file__).parent / 'models/navmodel.onnx'}
+
+class NavModelOutputXY(ctypes.Structure):
+  _fields_ = [
+    ("x", ctypes.c_float),
+    ("y", ctypes.c_float)]
+
+class NavModelOutputPlan(ctypes.Structure):
+  _fields_ = [
+    ("mean", NavModelOutputXY*ModelConstants.IDX_N),
+    ("std", NavModelOutputXY*ModelConstants.IDX_N)]
+
+class NavModelResult(ctypes.Structure):
+  _fields_ = [
+    ("plan", NavModelOutputPlan),
+    ("desire_pred", ctypes.c_float*NAV_DESIRE_LEN),
+    ("features", ctypes.c_float*NAV_FEATURE_LEN)]
+
+class ModelState:
+  inputs: dict[str, np.ndarray]
+  output: np.ndarray
+  model: ModelRunner
+
+  def __init__(self):
+    assert ctypes.sizeof(NavModelResult) == NAV_OUTPUT_SIZE * ctypes.sizeof(ctypes.c_float)
+    self.output = np.zeros(NAV_OUTPUT_SIZE, dtype=np.float32)
+    self.inputs = {'input_img': np.zeros(NAV_INPUT_SIZE, dtype=np.uint8)}
+    self.model = ModelRunner(MODEL_PATHS, self.output, Runtime.DSP, True, None)
+    self.model.addInput("input_img", None)
+
+  def run(self, buf:np.ndarray) -> tuple[np.ndarray, float]:
+    self.inputs['input_img'][:] = buf
+
+    t1 = time.perf_counter()
+    self.model.setInputBuffer("input_img", self.inputs['input_img'].view(np.float32))
+    self.model.execute()
+    t2 = time.perf_counter()
+    return self.output, t2 - t1
+
+def get_navmodel_packet(model_output: np.ndarray, valid: bool, frame_id: int, location_ts: int, execution_time: float, dsp_execution_time: float):
+  model_result = ctypes.cast(model_output.ctypes.data, ctypes.POINTER(NavModelResult)).contents
+  msg = messaging.new_message('navModel')
+  msg.valid = valid
+  msg.navModel.frameId = frame_id
+  msg.navModel.locationMonoTime = location_ts
+  msg.navModel.modelExecutionTime = execution_time
+  msg.navModel.dspExecutionTime = dsp_execution_time
+  msg.navModel.features = model_result.features[:]
+  msg.navModel.desirePrediction = model_result.desire_pred[:]
+  msg.navModel.position.x = [p.x for p in model_result.plan.mean]
+  msg.navModel.position.y = [p.y for p in model_result.plan.mean]
+  msg.navModel.position.xStd = [math.exp(p.x) for p in model_result.plan.std]
+  msg.navModel.position.yStd = [math.exp(p.y) for p in model_result.plan.std]
+  return msg
+
+
+def main():
+  gc.disable()
+  set_realtime_priority(1)
+
+  model = ModelState()
+  cloudlog.warning("models loaded, navmodeld starting")
+
+  vipc_client = VisionIpcClient("navd", VisionStreamType.VISION_STREAM_MAP, True)
+  while not vipc_client.connect(False):
+    time.sleep(0.1)
+  assert vipc_client.is_connected()
+  cloudlog.warning(f"connected with buffer size: {vipc_client.buffer_len}")
+
+  sm = SubMaster(["navInstruction"])
+  pm = PubMaster(["navModel"])
+
+  while True:
+    buf = vipc_client.recv()
+    if buf is None:
+      continue
+
+    sm.update(0)
+    t1 = time.perf_counter()
+    model_output, dsp_execution_time = model.run(buf.data[:buf.uv_offset])
+    t2 = time.perf_counter()
+
+    valid = vipc_client.valid and sm.valid["navInstruction"]
+    pm.send("navModel", get_navmodel_packet(model_output, valid, vipc_client.frame_id, vipc_client.timestamp_sof, t2 - t1, dsp_execution_time))
+
+
+if __name__ == "__main__":
+  main()

selfdrive/classic_modeld/parse_model_outputs.py

@@ -0,0 +1,103 @@
+import numpy as np
+from openpilot.selfdrive.classic_modeld.constants import ModelConstants
+
+def sigmoid(x):
+  return 1. / (1. + np.exp(-x))
+
+def softmax(x, axis=-1):
+  x -= np.max(x, axis=axis, keepdims=True)
+  if x.dtype == np.float32 or x.dtype == np.float64:
+    np.exp(x, out=x)
+  else:
+    x = np.exp(x)
+  x /= np.sum(x, axis=axis, keepdims=True)
+  return x
+
+class Parser:
+  def __init__(self, ignore_missing=False):
+    self.ignore_missing = ignore_missing
+
+  def check_missing(self, outs, name):
+    if name not in outs and not self.ignore_missing:
+      raise ValueError(f"Missing output {name}")
+    return name not in outs
+
+  def parse_categorical_crossentropy(self, name, outs, out_shape=None):
+    if self.check_missing(outs, name):
+      return
+    raw = outs[name]
+    if out_shape is not None:
+      raw = raw.reshape((raw.shape[0],) + out_shape)
+    outs[name] = softmax(raw, axis=-1)
+
+  def parse_binary_crossentropy(self, name, outs):
+    if self.check_missing(outs, name):
+      return
+    raw = outs[name]
+    outs[name] = sigmoid(raw)
+
+  def parse_mdn(self, name, outs, in_N=0, out_N=1, out_shape=None):
+    if self.check_missing(outs, name):
+      return
+    raw = outs[name]
+    raw = raw.reshape((raw.shape[0], max(in_N, 1), -1))
+
+    pred_mu = raw[:,:,:(raw.shape[2] - out_N)//2]
+    n_values = (raw.shape[2] - out_N)//2
+    pred_mu = raw[:,:,:n_values]
+    pred_std = np.exp(raw[:,:,n_values: 2*n_values])
+
+    if in_N > 1:
+      weights = np.zeros((raw.shape[0], in_N, out_N), dtype=raw.dtype)
+      for i in range(out_N):
+        weights[:,:,i - out_N] = softmax(raw[:,:,i - out_N], axis=-1)
+
+      if out_N == 1:
+        for fidx in range(weights.shape[0]):
+          idxs = np.argsort(weights[fidx][:,0])[::-1]
+          weights[fidx] = weights[fidx][idxs]
+          pred_mu[fidx] = pred_mu[fidx][idxs]
+          pred_std[fidx] = pred_std[fidx][idxs]
+      full_shape = tuple([raw.shape[0], in_N] + list(out_shape))
+      outs[name + '_weights'] = weights
+      outs[name + '_hypotheses'] = pred_mu.reshape(full_shape)
+      outs[name + '_stds_hypotheses'] = pred_std.reshape(full_shape)
+
+      pred_mu_final = np.zeros((raw.shape[0], out_N, n_values), dtype=raw.dtype)
+      pred_std_final = np.zeros((raw.shape[0], out_N, n_values), dtype=raw.dtype)
+      for fidx in range(weights.shape[0]):
+        for hidx in range(out_N):
+          idxs = np.argsort(weights[fidx,:,hidx])[::-1]
+          pred_mu_final[fidx, hidx] = pred_mu[fidx, idxs[0]]
+          pred_std_final[fidx, hidx] = pred_std[fidx, idxs[0]]
+    else:
+      pred_mu_final = pred_mu
+      pred_std_final = pred_std
+
+    if out_N > 1:
+      final_shape = tuple([raw.shape[0], out_N] + list(out_shape))
+    else:
+      final_shape = tuple([raw.shape[0],] + list(out_shape))
+    outs[name] = pred_mu_final.reshape(final_shape)
+    outs[name + '_stds'] = pred_std_final.reshape(final_shape)
+
+  def parse_outputs(self, outs: dict[str, np.ndarray]) -> dict[str, np.ndarray]:
+    self.parse_mdn('plan', outs, in_N=ModelConstants.PLAN_MHP_N, out_N=ModelConstants.PLAN_MHP_SELECTION,
+                   out_shape=(ModelConstants.IDX_N,ModelConstants.PLAN_WIDTH))
+    self.parse_mdn('lane_lines', outs, in_N=0, out_N=0, out_shape=(ModelConstants.NUM_LANE_LINES,ModelConstants.IDX_N,ModelConstants.LANE_LINES_WIDTH))
+    self.parse_mdn('road_edges', outs, in_N=0, out_N=0, out_shape=(ModelConstants.NUM_ROAD_EDGES,ModelConstants.IDX_N,ModelConstants.LANE_LINES_WIDTH))
+    self.parse_mdn('pose', outs, in_N=0, out_N=0, out_shape=(ModelConstants.POSE_WIDTH,))
+    self.parse_mdn('road_transform', outs, in_N=0, out_N=0, out_shape=(ModelConstants.POSE_WIDTH,))
+    self.parse_mdn('sim_pose', outs, in_N=0, out_N=0, out_shape=(ModelConstants.POSE_WIDTH,))
+    self.parse_mdn('wide_from_device_euler', outs, in_N=0, out_N=0, out_shape=(ModelConstants.WIDE_FROM_DEVICE_WIDTH,))
+    self.parse_mdn('lead', outs, in_N=ModelConstants.LEAD_MHP_N, out_N=ModelConstants.LEAD_MHP_SELECTION,
+                   out_shape=(ModelConstants.LEAD_TRAJ_LEN,ModelConstants.LEAD_WIDTH))
+    if 'lat_planner_solution' in outs:
+      self.parse_mdn('lat_planner_solution', outs, in_N=0, out_N=0, out_shape=(ModelConstants.IDX_N,ModelConstants.LAT_PLANNER_SOLUTION_WIDTH))
+    if 'desired_curvature' in outs:
+      self.parse_mdn('desired_curvature', outs, in_N=0, out_N=0, out_shape=(ModelConstants.DESIRED_CURV_WIDTH,))
+    for k in ['lead_prob', 'lane_lines_prob', 'meta']:
+      self.parse_binary_crossentropy(k, outs)
+    self.parse_categorical_crossentropy('desire_state', outs, out_shape=(ModelConstants.DESIRE_PRED_WIDTH,))
+    self.parse_categorical_crossentropy('desire_pred', outs, out_shape=(ModelConstants.DESIRE_PRED_LEN,ModelConstants.DESIRE_PRED_WIDTH))
+    return outs

selfdrive/classic_modeld/runners/__init__.py

@@ -0,0 +1,27 @@
+import os
+from openpilot.system.hardware import TICI
+from openpilot.selfdrive.classic_modeld.runners.runmodel_pyx import RunModel, Runtime
+assert Runtime
+
+USE_THNEED = int(os.getenv('USE_THNEED', str(int(TICI))))
+USE_SNPE = int(os.getenv('USE_SNPE', str(int(TICI))))
+
+class ModelRunner(RunModel):
+  THNEED = 'THNEED'
+  SNPE = 'SNPE'
+  ONNX = 'ONNX'
+
+  def __new__(cls, paths, *args, **kwargs):
+    if ModelRunner.THNEED in paths and USE_THNEED:
+      from openpilot.selfdrive.classic_modeld.runners.thneedmodel_pyx import ThneedModel as Runner
+      runner_type = ModelRunner.THNEED
+    elif ModelRunner.SNPE in paths and USE_SNPE:
+      from openpilot.selfdrive.classic_modeld.runners.snpemodel_pyx import SNPEModel as Runner
+      runner_type = ModelRunner.SNPE
+    elif ModelRunner.ONNX in paths:
+      from openpilot.selfdrive.classic_modeld.runners.onnxmodel import ONNXModel as Runner
+      runner_type = ModelRunner.ONNX
+    else:
+      raise Exception("Couldn't select a model runner, make sure to pass at least one valid model path")
+
+    return Runner(str(paths[runner_type]), *args, **kwargs)

selfdrive/classic_modeld/runners/onnxmodel.py

@@ -0,0 +1,93 @@
+import onnx
+import itertools
+import os
+import sys
+import numpy as np
+from typing import Any
+
+from openpilot.selfdrive.classic_modeld.runners.runmodel_pyx import RunModel
+
+ORT_TYPES_TO_NP_TYPES = {'tensor(float16)': np.float16, 'tensor(float)': np.float32, 'tensor(uint8)': np.uint8}
+
+def attributeproto_fp16_to_fp32(attr):
+  float32_list = np.frombuffer(attr.raw_data, dtype=np.float16)
+  attr.data_type = 1
+  attr.raw_data = float32_list.astype(np.float32).tobytes()
+
+def convert_fp16_to_fp32(path):
+  model = onnx.load(path)
+  for i in model.graph.initializer:
+    if i.data_type == 10:
+      attributeproto_fp16_to_fp32(i)
+  for i in itertools.chain(model.graph.input, model.graph.output):
+    if i.type.tensor_type.elem_type == 10:
+      i.type.tensor_type.elem_type = 1
+  for i in model.graph.node:
+    for a in i.attribute:
+      if hasattr(a, 't'):
+        if a.t.data_type == 10:
+          attributeproto_fp16_to_fp32(a.t)
+  return model.SerializeToString()
+
+def create_ort_session(path, fp16_to_fp32):
+  os.environ["OMP_NUM_THREADS"] = "4"
+  os.environ["OMP_WAIT_POLICY"] = "PASSIVE"
+
+  import onnxruntime as ort
+  print("Onnx available providers: ", ort.get_available_providers(), file=sys.stderr)
+  options = ort.SessionOptions()
+  options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_DISABLE_ALL
+
+  provider: str | tuple[str, dict[Any, Any]]
+  if 'OpenVINOExecutionProvider' in ort.get_available_providers() and 'ONNXCPU' not in os.environ:
+    provider = 'OpenVINOExecutionProvider'
+  elif 'CUDAExecutionProvider' in ort.get_available_providers() and 'ONNXCPU' not in os.environ:
+    options.intra_op_num_threads = 2
+    provider = ('CUDAExecutionProvider', {'cudnn_conv_algo_search': 'DEFAULT'})
+  else:
+    options.intra_op_num_threads = 2
+    options.execution_mode = ort.ExecutionMode.ORT_SEQUENTIAL
+    options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
+    provider = 'CPUExecutionProvider'
+
+  model_data = convert_fp16_to_fp32(path) if fp16_to_fp32 else path
+  print("Onnx selected provider: ", [provider], file=sys.stderr)
+  ort_session = ort.InferenceSession(model_data, options, providers=[provider])
+  print("Onnx using ", ort_session.get_providers(), file=sys.stderr)
+  return ort_session
+
+
+class ONNXModel(RunModel):
+  def __init__(self, path, output, runtime, use_tf8, cl_context):
+    self.inputs = {}
+    self.output = output
+    self.use_tf8 = use_tf8
+
+    self.session = create_ort_session(path, fp16_to_fp32=True)
+    self.input_names = [x.name for x in self.session.get_inputs()]
+    self.input_shapes = {x.name: [1, *x.shape[1:]] for x in self.session.get_inputs()}
+    self.input_dtypes = {x.name: ORT_TYPES_TO_NP_TYPES[x.type] for x in self.session.get_inputs()}
+
+    # run once to initialize CUDA provider
+    if "CUDAExecutionProvider" in self.session.get_providers():
+      self.session.run(None, {k: np.zeros(self.input_shapes[k], dtype=self.input_dtypes[k]) for k in self.input_names})
+    print("ready to run onnx model", self.input_shapes, file=sys.stderr)
+
+  def addInput(self, name, buffer):
+    assert name in self.input_names
+    self.inputs[name] = buffer
+
+  def setInputBuffer(self, name, buffer):
+    assert name in self.inputs
+    self.inputs[name] = buffer
+
+  def getCLBuffer(self, name):
+    return None
+
+  def execute(self):
+    inputs = {k: (v.view(np.uint8) / 255. if self.use_tf8 and k == 'input_img' else v) for k,v in self.inputs.items()}
+    inputs = {k: v.reshape(self.input_shapes[k]).astype(self.input_dtypes[k]) for k,v in inputs.items()}
+    outputs = self.session.run(None, inputs)
+    assert len(outputs) == 1, "Only single model outputs are supported"
+    self.output[:] = outputs[0]
+    return self.output

selfdrive/classic_modeld/runners/runmodel.pxd

@@ -0,0 +1,14 @@
+# distutils: language = c++
+
+from libcpp.string cimport string
+
+cdef extern from "selfdrive/classic_modeld/runners/runmodel.h":
+  cdef int USE_CPU_RUNTIME
+  cdef int USE_GPU_RUNTIME
+  cdef int USE_DSP_RUNTIME
+
+  cdef cppclass RunModel:
+    void addInput(string, float*, int)
+    void setInputBuffer(string, float*, int)
+    void * getCLBuffer(string)
+    void execute()

selfdrive/classic_modeld/runners/runmodel_pyx.pxd

@@ -0,0 +1,6 @@
+# distutils: language = c++
+
+from .runmodel cimport RunModel as cppRunModel
+
+cdef class RunModel:
+  cdef cppRunModel * model

selfdrive/classic_modeld/runners/runmodel_pyx.pyx

@@ -0,0 +1,37 @@
+# distutils: language = c++
+# cython: c_string_encoding=ascii
+
+from libcpp.string cimport string
+
+from .runmodel cimport USE_CPU_RUNTIME, USE_GPU_RUNTIME, USE_DSP_RUNTIME
+from selfdrive.classic_modeld.models.commonmodel_pyx cimport CLMem
+
+class Runtime:
+  CPU = USE_CPU_RUNTIME
+  GPU = USE_GPU_RUNTIME
+  DSP = USE_DSP_RUNTIME
+
+cdef class RunModel:
+  def __dealloc__(self):
+    del self.model
+
+  def addInput(self, string name, float[:] buffer):
+    if buffer is not None:
+      self.model.addInput(name, &buffer[0], len(buffer))
+    else:
+      self.model.addInput(name, NULL, 0)
+
+  def setInputBuffer(self, string name, float[:] buffer):
+    if buffer is not None:
+      self.model.setInputBuffer(name, &buffer[0], len(buffer))
+    else:
+      self.model.setInputBuffer(name, NULL, 0)
+
+  def getCLBuffer(self, string name):
+    cdef void * cl_buf = self.model.getCLBuffer(name)
+    if not cl_buf:
+      return None
+    return CLMem.create(cl_buf)
+
+  def execute(self):
+    self.model.execute()

selfdrive/classic_modeld/runners/snpemodel.pxd

@@ -0,0 +1,9 @@
+# distutils: language = c++
+
+from libcpp.string cimport string
+
+from msgq.visionipc.visionipc cimport cl_context
+
+cdef extern from "selfdrive/classic_modeld/runners/snpemodel.h":
+  cdef cppclass SNPEModel:
+    SNPEModel(string, float*, size_t, int, bool, cl_context)

selfdrive/classic_modeld/runners/snpemodel_pyx.pyx

@@ -0,0 +1,17 @@
+# distutils: language = c++
+# cython: c_string_encoding=ascii
+
+import os
+from libcpp cimport bool
+from libcpp.string cimport string
+
+from .snpemodel cimport SNPEModel as cppSNPEModel
+from selfdrive.classic_modeld.models.commonmodel_pyx cimport CLContext
+from selfdrive.classic_modeld.runners.runmodel_pyx cimport RunModel
+from selfdrive.classic_modeld.runners.runmodel cimport RunModel as cppRunModel
+
+os.environ['ADSP_LIBRARY_PATH'] = "/data/pythonpath/third_party/snpe/dsp/"
+
+cdef class SNPEModel(RunModel):
+  def __cinit__(self, string path, float[:] output, int runtime, bool use_tf8, CLContext context):
+    self.model = <cppRunModel *> new cppSNPEModel(path, &output[0], len(output), runtime, use_tf8, context.context)

selfdrive/classic_modeld/runners/thneedmodel.pxd

@@ -0,0 +1,9 @@
+# distutils: language = c++
+
+from libcpp.string cimport string
+
+from msgq.visionipc.visionipc cimport cl_context
+
+cdef extern from "selfdrive/classic_modeld/runners/thneedmodel.h":
+  cdef cppclass ThneedModel:
+    ThneedModel(string, float*, size_t, int, bool, cl_context)

selfdrive/classic_modeld/runners/thneedmodel_pyx.pyx

@@ -0,0 +1,14 @@
+# distutils: language = c++
+# cython: c_string_encoding=ascii
+
+from libcpp cimport bool
+from libcpp.string cimport string
+
+from .thneedmodel cimport ThneedModel as cppThneedModel
+from selfdrive.classic_modeld.models.commonmodel_pyx cimport CLContext
+from selfdrive.classic_modeld.runners.runmodel_pyx cimport RunModel
+from selfdrive.classic_modeld.runners.runmodel cimport RunModel as cppRunModel
+
+cdef class ThneedModel(RunModel):
+  def __cinit__(self, string path, float[:] output, int runtime, bool use_tf8, CLContext context):
+    self.model = <cppRunModel *> new cppThneedModel(path, &output[0], len(output), runtime, use_tf8, context.context)

selfdrive/classic_modeld/transforms/loadyuv.cl

@@ -0,0 +1,47 @@
+#define UV_SIZE ((TRANSFORMED_WIDTH/2)*(TRANSFORMED_HEIGHT/2))
+
+__kernel void loadys(__global uchar8 const * const Y,
+                     __global float * out,
+                     int out_offset)
+{
+    const int gid = get_global_id(0);
+    const int ois = gid * 8;
+    const int oy = ois / TRANSFORMED_WIDTH;
+    const int ox = ois % TRANSFORMED_WIDTH;
+
+    const uchar8 ys = Y[gid];
+    const float8 ysf = convert_float8(ys);
+
+    // 02
+    // 13
+
+    __global float* outy0;
+    __global float* outy1;
+    if ((oy & 1) == 0) {
+      outy0 = out + out_offset; //y0
+      outy1 = out + out_offset + UV_SIZE*2; //y2
+    } else {
+      outy0 = out + out_offset + UV_SIZE; //y1
+      outy1 = out + out_offset + UV_SIZE*3; //y3
+    }
+
+    vstore4(ysf.s0246, 0, outy0 + (oy/2) * (TRANSFORMED_WIDTH/2) + ox/2);
+    vstore4(ysf.s1357, 0, outy1 + (oy/2) * (TRANSFORMED_WIDTH/2) + ox/2);
+}
+
+__kernel void loaduv(__global uchar8 const * const in,
+                     __global float8 * out,
+                     int out_offset)
+{
+  const int gid = get_global_id(0);
+  const uchar8 inv = in[gid];
+  const float8 outv  = convert_float8(inv);
+  out[gid + out_offset / 8] = outv;
+}
+
+__kernel void copy(__global float8 * inout,
+                   int in_offset)
+{
+  const int gid = get_global_id(0);
+  inout[gid] = inout[gid + in_offset / 8];
+}

selfdrive/classic_modeld/transforms/transform.cc

@@ -0,0 +1,97 @@
+#include "selfdrive/classic_modeld/transforms/transform.h"
+
+#include <cassert>
+#include <cstring>
+
+#include "common/clutil.h"
+
+void transform_init(Transform* s, cl_context ctx, cl_device_id device_id) {
+  memset(s, 0, sizeof(*s));
+
+  cl_program prg = cl_program_from_file(ctx, device_id, TRANSFORM_PATH, "");
+  s->krnl = CL_CHECK_ERR(clCreateKernel(prg, "warpPerspective", &err));
+  // done with this
+  CL_CHECK(clReleaseProgram(prg));
+
+  s->m_y_cl = CL_CHECK_ERR(clCreateBuffer(ctx, CL_MEM_READ_WRITE, 3*3*sizeof(float), NULL, &err));
+  s->m_uv_cl = CL_CHECK_ERR(clCreateBuffer(ctx, CL_MEM_READ_WRITE, 3*3*sizeof(float), NULL, &err));
+}
+
+void transform_destroy(Transform* s) {
+  CL_CHECK(clReleaseMemObject(s->m_y_cl));
+  CL_CHECK(clReleaseMemObject(s->m_uv_cl));
+  CL_CHECK(clReleaseKernel(s->krnl));
+}
+
+void transform_queue(Transform* s,
+                     cl_command_queue q,
+                     cl_mem in_yuv, int in_width, int in_height, int in_stride, int in_uv_offset,
+                     cl_mem out_y, cl_mem out_u, cl_mem out_v,
+                     int out_width, int out_height,
+                     const mat3& projection) {
+  const int zero = 0;
+
+  // sampled using pixel center origin
+  // (because that's how fastcv and opencv does it)
+
+  mat3 projection_y = projection;
+
+  // in and out uv is half the size of y.
+  mat3 projection_uv = transform_scale_buffer(projection, 0.5);
+
+  CL_CHECK(clEnqueueWriteBuffer(q, s->m_y_cl, CL_TRUE, 0, 3*3*sizeof(float), (void*)projection_y.v, 0, NULL, NULL));
+  CL_CHECK(clEnqueueWriteBuffer(q, s->m_uv_cl, CL_TRUE, 0, 3*3*sizeof(float), (void*)projection_uv.v, 0, NULL, NULL));
+
+  const int in_y_width = in_width;
+  const int in_y_height = in_height;
+  const int in_y_px_stride = 1;
+  const int in_uv_width = in_width/2;
+  const int in_uv_height = in_height/2;
+  const int in_uv_px_stride = 2;
+  const int in_u_offset = in_uv_offset;
+  const int in_v_offset = in_uv_offset + 1;
+
+  const int out_y_width = out_width;
+  const int out_y_height = out_height;
+  const int out_uv_width = out_width/2;
+  const int out_uv_height = out_height/2;
+
+  CL_CHECK(clSetKernelArg(s->krnl, 0, sizeof(cl_mem), &in_yuv));  // src
+  CL_CHECK(clSetKernelArg(s->krnl, 1, sizeof(cl_int), &in_stride));  // src_row_stride
+  CL_CHECK(clSetKernelArg(s->krnl, 2, sizeof(cl_int), &in_y_px_stride));  // src_px_stride
+  CL_CHECK(clSetKernelArg(s->krnl, 3, sizeof(cl_int), &zero));  // src_offset
+  CL_CHECK(clSetKernelArg(s->krnl, 4, sizeof(cl_int), &in_y_height));  // src_rows
+  CL_CHECK(clSetKernelArg(s->krnl, 5, sizeof(cl_int), &in_y_width));  // src_cols
+  CL_CHECK(clSetKernelArg(s->krnl, 6, sizeof(cl_mem), &out_y));  // dst
+  CL_CHECK(clSetKernelArg(s->krnl, 7, sizeof(cl_int), &out_y_width));  // dst_row_stride
+  CL_CHECK(clSetKernelArg(s->krnl, 8, sizeof(cl_int), &zero));  // dst_offset
+  CL_CHECK(clSetKernelArg(s->krnl, 9, sizeof(cl_int), &out_y_height));  // dst_rows
+  CL_CHECK(clSetKernelArg(s->krnl, 10, sizeof(cl_int), &out_y_width));  // dst_cols
+  CL_CHECK(clSetKernelArg(s->krnl, 11, sizeof(cl_mem), &s->m_y_cl));  // M
+
+  const size_t work_size_y[2] = {(size_t)out_y_width, (size_t)out_y_height};
+
+  CL_CHECK(clEnqueueNDRangeKernel(q, s->krnl, 2, NULL,
+                              (const size_t*)&work_size_y, NULL, 0, 0, NULL));
+
+  const size_t work_size_uv[2] = {(size_t)out_uv_width, (size_t)out_uv_height};
+
+  CL_CHECK(clSetKernelArg(s->krnl, 2, sizeof(cl_int), &in_uv_px_stride));  // src_px_stride
+  CL_CHECK(clSetKernelArg(s->krnl, 3, sizeof(cl_int), &in_u_offset));  // src_offset
+  CL_CHECK(clSetKernelArg(s->krnl, 4, sizeof(cl_int), &in_uv_height));  // src_rows
+  CL_CHECK(clSetKernelArg(s->krnl, 5, sizeof(cl_int), &in_uv_width));  // src_cols
+  CL_CHECK(clSetKernelArg(s->krnl, 6, sizeof(cl_mem), &out_u));  // dst
+  CL_CHECK(clSetKernelArg(s->krnl, 7, sizeof(cl_int), &out_uv_width));  // dst_row_stride
+  CL_CHECK(clSetKernelArg(s->krnl, 8, sizeof(cl_int), &zero));  // dst_offset
+  CL_CHECK(clSetKernelArg(s->krnl, 9, sizeof(cl_int), &out_uv_height));  // dst_rows
+  CL_CHECK(clSetKernelArg(s->krnl, 10, sizeof(cl_int), &out_uv_width));  // dst_cols
+  CL_CHECK(clSetKernelArg(s->krnl, 11, sizeof(cl_mem), &s->m_uv_cl));  // M
+
+  CL_CHECK(clEnqueueNDRangeKernel(q, s->krnl, 2, NULL,
+                              (const size_t*)&work_size_uv, NULL, 0, 0, NULL));
+  CL_CHECK(clSetKernelArg(s->krnl, 3, sizeof(cl_int), &in_v_offset));  // src_ofset
+  CL_CHECK(clSetKernelArg(s->krnl, 6, sizeof(cl_mem), &out_v));  // dst
+
+  CL_CHECK(clEnqueueNDRangeKernel(q, s->krnl, 2, NULL,
+                              (const size_t*)&work_size_uv, NULL, 0, 0, NULL));
+}

selfdrive/classic_modeld/transforms/transform.cl

@@ -0,0 +1,54 @@
+#define INTER_BITS 5
+#define INTER_TAB_SIZE (1 << INTER_BITS)
+#define INTER_SCALE 1.f / INTER_TAB_SIZE
+
+#define INTER_REMAP_COEF_BITS 15
+#define INTER_REMAP_COEF_SCALE (1 << INTER_REMAP_COEF_BITS)
+
+__kernel void warpPerspective(__global const uchar * src,
+                              int src_row_stride, int src_px_stride, int src_offset, int src_rows, int src_cols,
+                              __global uchar * dst,
+                              int dst_row_stride, int dst_offset, int dst_rows, int dst_cols,
+                              __constant float * M)
+{
+    int dx = get_global_id(0);
+    int dy = get_global_id(1);
+
+    if (dx < dst_cols && dy < dst_rows)
+    {
+        float X0 = M[0] * dx + M[1] * dy + M[2];
+        float Y0 = M[3] * dx + M[4] * dy + M[5];
+        float W = M[6] * dx + M[7] * dy + M[8];
+        W = W != 0.0f ? INTER_TAB_SIZE / W : 0.0f;
+        int X = rint(X0 * W), Y = rint(Y0 * W);
+
+        int sx = convert_short_sat(X >> INTER_BITS);
+        int sy = convert_short_sat(Y >> INTER_BITS);
+
+        short sx_clamp = clamp(sx, 0, src_cols - 1);
+        short sx_p1_clamp = clamp(sx + 1, 0, src_cols - 1);
+        short sy_clamp = clamp(sy, 0, src_rows - 1);
+        short sy_p1_clamp = clamp(sy + 1, 0, src_rows - 1);
+        int v0 = convert_int(src[mad24(sy_clamp, src_row_stride, src_offset + sx_clamp*src_px_stride)]);
+        int v1 = convert_int(src[mad24(sy_clamp, src_row_stride, src_offset + sx_p1_clamp*src_px_stride)]);
+        int v2 = convert_int(src[mad24(sy_p1_clamp, src_row_stride, src_offset + sx_clamp*src_px_stride)]);
+        int v3 = convert_int(src[mad24(sy_p1_clamp, src_row_stride, src_offset + sx_p1_clamp*src_px_stride)]);
+
+        short ay = (short)(Y & (INTER_TAB_SIZE - 1));
+        short ax = (short)(X & (INTER_TAB_SIZE - 1));
+        float taby = 1.f/INTER_TAB_SIZE*ay;
+        float tabx = 1.f/INTER_TAB_SIZE*ax;
+
+        int dst_index = mad24(dy, dst_row_stride, dst_offset + dx);
+
+        int itab0 = convert_short_sat_rte( (1.0f-taby)*(1.0f-tabx) * INTER_REMAP_COEF_SCALE );
+        int itab1 = convert_short_sat_rte( (1.0f-taby)*tabx * INTER_REMAP_COEF_SCALE );
+        int itab2 = convert_short_sat_rte( taby*(1.0f-tabx) * INTER_REMAP_COEF_SCALE );
+        int itab3 = convert_short_sat_rte( taby*tabx * INTER_REMAP_COEF_SCALE );
+
+        int val = v0 * itab0 +  v1 * itab1 + v2 * itab2 + v3 * itab3;
+
+        uchar pix = convert_uchar_sat((val + (1 << (INTER_REMAP_COEF_BITS-1))) >> INTER_REMAP_COEF_BITS);
+        dst[dst_index] = pix;
+    }
+}