IQ.Pilot Release Commit @ 5bc9cd3

2026-08-22 23:42:42 -05:00
commit d037016281
4504 changed files with 1129827 additions and 0 deletions
--- a/selfdrive/modeld/SConscript
+++ b/selfdrive/modeld/SConscript
@@ -0,0 +1,103 @@
+import os
+import glob
+
+Import('env', 'envCython', 'arch', 'cereal', 'messaging', 'common', 'visionipc')
+lenv = env.Clone()
+lenvCython = envCython.Clone()
+
+libs = [cereal, messaging, visionipc, common, 'capnp', 'kj', 'pthread']
+frameworks = []
+
+common_src = [
+  "models/commonmodel.cc",
+  "transforms/loadyuv.cc",
+  "transforms/transform.cc",
+]
+
+# 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
+cython_libs = envCython["LIBS"] + libs
+commonmodel_lib = lenv.Library('commonmodel', common_src)
+lenvCython.Program('models/commonmodel_pyx.so', 'models/commonmodel_pyx.pyx', LIBS=[commonmodel_lib, *cython_libs], FRAMEWORKS=frameworks)
+tinygrad_files = sorted(["#"+x for x in glob.glob(env.Dir("#tinygrad_repo").relpath + "/**", recursive=True, root_dir=env.Dir("#").abspath) if 'pycache' not in x])
+
+def tg_compile(flags, model_name):
+  pythonpath_string = 'PYTHONPATH="${PYTHONPATH}:' + env.Dir("#tinygrad_repo").abspath + '"'
+  fn = File(f"models/{model_name}").abspath
+  cmd = lenv.Command(
+    fn + "_tinygrad.pkl",
+    [fn + ".onnx"] + tinygrad_files,
+    lenv.PrettyAction(
+      f'${{PYWARN}} {pythonpath_string} {flags} python3 {Dir("#tinygrad_repo").abspath}/examples/openpilot/compile3.py {fn}.onnx {fn}_tinygrad.pkl',
+      'MODEL', logfile='${TARGET}.log')
+  )
+  # committed pkls must survive a failed rebuild (Precious: no pre-build
+  # delete) and scons -c (NoClean); a failed compile must not brick modeld
+  lenv.Precious(cmd)
+  lenv.NoClean(cmd)
+  return cmd
+
+def host_tinygrad_flags(*, float16=False):
+  if arch == "larch64":
+    base = "DEV=QCOM IMAGE=2 FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0 OPENPILOT_HACKS=1"
+    return base
+  if arch == "Darwin":
+    base = f'DEV=CPU HOME={os.path.expanduser("~")} IMAGE=0'
+  elif arch == "x86_64":
+    base = "DEV=CPU:LLVM IMAGE=0"
+  else:
+    base = "DEV=CPU:LLVM IMAGE=0"
+  return f"{base} FLOAT16=1" if float16 else base
+
+# Compile small models
+for model_name in ['dmonitoring_model']:
+  # The optimization flags are mandatory on QCOM: without FLOAT16/NOLOCALS/JIT_BATCH_SIZE/OPENPILOT_HACKS these
+  # models compile to unoptimized QCOM kernels and run ~20x slower (dmonitoring_model: ~300ms -> ~14ms),
+  # which starves the driving model on the shared Adreno. IMAGE=2 (not upstream's IMAGE=1) because the
+  # pinned tinygrad (fd992d66) hits an IMAGE=1 codegen bug on this Adreno; IMAGE=2 is correct and fast here.
+  flags = host_tinygrad_flags()
+
+  # Shipped prebuilt pkls: on device, a fresh install has no .sconsign, so scons would recompile these
+  # from onnx (5-10 min) even though up-to-date pkls are committed. The check file pins the exact
+  # inputs (onnx + tinygrad_repo + flags + metadata script) and output hashes; if it matches, skip
+  # declaring the targets entirely. Any mismatch falls back to a normal on-device compile.
+  if arch == "larch64":
+    from openpilot.selfdrive.modeld.prebuilt_models import packaged_prebuilt_matches, verify_prebuilt, outputs_match
+    if packaged_prebuilt_matches(model_name):
+      print(lenv.PrettyNote('SKIP', f"{model_name} — packaged prebuilt pkl"))
+      continue
+    if verify_prebuilt(model_name, flags):
+      print(lenv.PrettyNote('SKIP', f"{model_name} — prebuilt pkl"))
+      continue
+    # Input digest mismatch but the committed artifacts are intact: this is a
+    # device with a modified/partial tinygrad_repo (e.g. failed submodule fetch
+    # on an install without konn3kt auth). Recompiling here would DELETE the
+    # known-good pkl and then fail (compile3.py may not even exist), bricking
+    # modeld. Keep the shipped artifacts and say so.
+    if outputs_match(model_name):
+      print(lenv.PrettyNote('WARN', f"{model_name} — input digest mismatch (tinygrad_repo incomplete/modified?), keeping committed pkl"))
+      continue
+  elif not os.environ.get("COMPILE_MODELS"):
+    # The committed pkls ARE the device (QCOM) artifacts. A host build would
+    # overwrite them with host-flavor pkls (and `scons -c` deletes them), which
+    # then show up as staged changes and brick devices if committed. Host pkls
+    # only on explicit request: COMPILE_MODELS=1 scons ...
+    print(lenv.PrettyNote('SKIP', f"{model_name} — QCOM pkl kept (COMPILE_MODELS=1 to build host)"))
+    continue
+
+  fn = File(f"models/{model_name}").abspath
+  script_files = [File(Dir("#selfdrive/modeld").File("get_model_metadata.py").abspath)]
+  metadata_cmd = f'${{PYWARN}} python3 {Dir("#selfdrive/modeld").abspath}/get_model_metadata.py {fn}.onnx'
+  lenv.Command(fn + "_metadata.pkl", [fn + ".onnx"] + tinygrad_files + script_files,
+               lenv.PrettyAction(metadata_cmd, 'META', logfile='${TARGET}.log'))
+  tg_compile(flags, model_name)
--- a/selfdrive/modeld/init.py
+++ b/selfdrive/modeld/init.py
--- a/selfdrive/modeld/constants.py
+++ b/selfdrive/modeld/constants.py
@@ -0,0 +1,87 @@
+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
+  N_FRAMES = 2
+  MODEL_RUN_FREQ = 20
+  MODEL_CONTEXT_FREQ = 5 # "model_trained_fps"
+
+  FEATURE_LEN = 512
+
+  DESIRE_LEN = 8
+  TRAFFIC_CONVENTION_LEN = 2
+  LAT_PLANNER_STATE_LEN = 4
+  LATERAL_CONTROL_PARAMS_LEN = 2
+  PREV_DESIRED_CURV_LEN = 1
+
+  # 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
+
+  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
+
+  POLY_PATH_DEGREE = 4
+
+# 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, 31, 6)
+  BRAKE_DISENGAGE = slice(2, 31, 6)
+  STEER_OVERRIDE = slice(3, 31, 6)
+  HARD_BRAKE_3 = slice(4, 31, 6)
+  HARD_BRAKE_4 = slice(5, 31, 6)
+  HARD_BRAKE_5 = slice(6, 31, 6)
+  # next 0, 2, 4, 6, 8, 10 seconds
+  GAS_PRESS = slice(31, 55, 4)
+  BRAKE_PRESS = slice(32, 55, 4)
+  LEFT_BLINKER = slice(33, 55, 4)
+  RIGHT_BLINKER = slice(34, 55, 4)
--- a/selfdrive/modeld/dmonitoringmodeld.py
+++ b/selfdrive/modeld/dmonitoringmodeld.py
@@ -0,0 +1,157 @@
+#!/usr/bin/env python3
+import os
+from openpilot.system.hardware import TICI
+os.environ['DEV'] = 'QCOM' if TICI else 'CPU'
+from tinygrad.tensor import Tensor
+from tinygrad.dtype import dtypes
+import time
+import pickle
+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.realtime import config_realtime_process
+from openpilot.common.transformations.model import dmonitoringmodel_intrinsics
+from openpilot.common.transformations.camera import _ar_ox_fisheye, _os_fisheye
+from openpilot.selfdrive.locationd.calibration_helpers import get_calibrated_rpy
+from openpilot.selfdrive.modeld.models.commonmodel_pyx import CLContext, MonitoringModelFrame
+from openpilot.selfdrive.modeld.parse_model_outputs import sigmoid, safe_exp
+from openpilot.selfdrive.modeld.runners.tinygrad_helpers import qcom_tensor_from_opencl_address
+
+PROCESS_NAME = "selfdrive.modeld.dmonitoringmodeld"
+SEND_RAW_PRED = os.getenv('SEND_RAW_PRED')
+MODEL_PKL_PATH = Path(__file__).parent / 'models/dmonitoring_model_tinygrad.pkl'
+METADATA_PATH = Path(__file__).parent / 'models/dmonitoring_model_metadata.pkl'
+
+
+class ModelState:
+  inputs: dict[str, np.ndarray]
+  output: np.ndarray
+
+  def __init__(self, cl_ctx):
+    with open(METADATA_PATH, 'rb') as f:
+      model_metadata = pickle.load(f)
+      self.input_shapes = model_metadata['input_shapes']
+      self.output_slices = model_metadata['output_slices']
+
+    self.frame = MonitoringModelFrame(cl_ctx)
+    self.numpy_inputs = {
+      'calib': np.zeros(self.input_shapes['calib'], dtype=np.float32),
+    }
+
+    self.tensor_inputs = {k: Tensor(v, device='NPY').realize() for k,v in self.numpy_inputs.items()}
+    with open(MODEL_PKL_PATH, "rb") as f:
+      self.model_run = pickle.load(f)
+
+  def run(self, buf: VisionBuf, calib: np.ndarray, transform: np.ndarray) -> tuple[np.ndarray, float]:
+    self.numpy_inputs['calib'][0,:] = calib
+
+    t1 = time.perf_counter()
+
+    input_img_cl = self.frame.prepare(buf, transform.flatten())
+    if TICI:
+      # The imgs tensors are backed by opencl memory, only need init once
+      if 'input_img' not in self.tensor_inputs:
+        self.tensor_inputs['input_img'] = qcom_tensor_from_opencl_address(input_img_cl.mem_address, self.input_shapes['input_img'], dtype=dtypes.uint8)
+    else:
+      self.tensor_inputs['input_img'] = Tensor(self.frame.buffer_from_cl(input_img_cl).reshape(self.input_shapes['input_img']), dtype=dtypes.uint8).realize()
+
+
+    output = self.model_run(**self.tensor_inputs).contiguous().realize().uop.base.buffer.numpy()
+
+    t2 = time.perf_counter()
+    return output, t2 - t1
+
+def slice_outputs(model_outputs, output_slices):
+  return  {k: model_outputs[np.newaxis, v] for k,v in output_slices.items()}
+
+def parse_model_output(model_output):
+  parsed = {}
+  parsed['wheel_on_right'] = sigmoid(model_output['wheel_on_right'])
+  for ds_suffix in ['lhd', 'rhd']:
+    face_descs = model_output[f'face_descs_{ds_suffix}']
+    parsed[f'face_descs_{ds_suffix}'] = face_descs[:, :-6]
+    parsed[f'face_descs_{ds_suffix}_std'] = safe_exp(face_descs[:, -6:])
+    for key in ['face_prob', 'left_eye_prob', 'right_eye_prob','left_blink_prob', 'right_blink_prob', 'sunglasses_prob', 'using_phone_prob']:
+      parsed[f'{key}_{ds_suffix}'] = sigmoid(model_output[f'{key}_{ds_suffix}'])
+  return parsed
+
+def fill_driver_data(msg, model_output, ds_suffix):
+  msg.faceOrientation = model_output[f'face_descs_{ds_suffix}'][0, :3].tolist()
+  msg.faceOrientationStd = model_output[f'face_descs_{ds_suffix}_std'][0, :3].tolist()
+  msg.facePosition = model_output[f'face_descs_{ds_suffix}'][0, 3:5].tolist()
+  msg.facePositionStd = model_output[f'face_descs_{ds_suffix}_std'][0, 3:5].tolist()
+  msg.faceProb = model_output[f'face_prob_{ds_suffix}'][0, 0].item()
+  msg.leftEyeProb = model_output[f'left_eye_prob_{ds_suffix}'][0, 0].item()
+  msg.rightEyeProb = model_output[f'right_eye_prob_{ds_suffix}'][0, 0].item()
+  msg.leftBlinkProb = model_output[f'left_blink_prob_{ds_suffix}'][0, 0].item()
+  msg.rightBlinkProb = model_output[f'right_blink_prob_{ds_suffix}'][0, 0].item()
+  msg.sunglassesProb = model_output[f'sunglasses_prob_{ds_suffix}'][0, 0].item()
+  msg.phoneProb = model_output[f'using_phone_prob_{ds_suffix}'][0, 0].item()
+
+def get_driverstate_packet(model_output, frame_id: int, location_ts: int, exec_time: float, gpu_exec_time: float):
+  msg = messaging.new_message('driverStateV2', valid=True)
+  ds = msg.driverStateV2
+  ds.frameId = frame_id
+  ds.modelExecutionTime = exec_time
+  ds.gpuExecutionTime = gpu_exec_time
+  ds.rawPredictions = model_output['raw_pred']
+  ds.wheelOnRightProb = model_output['wheel_on_right'][0, 0].item()
+  fill_driver_data(ds.leftDriverData, model_output, 'lhd')
+  fill_driver_data(ds.rightDriverData, model_output, 'rhd')
+  return msg
+
+
+def main():
+  config_realtime_process(7, 5)
+
+  cl_context = CLContext()
+  model = ModelState(cl_context)
+  cloudlog.warning("models loaded, dmonitoringmodeld starting")
+
+  cloudlog.warning("connecting to driver stream")
+  vipc_client = VisionIpcClient("camerad", VisionStreamType.VISION_STREAM_DRIVER, True, cl_context)
+  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(model.numpy_inputs['calib'].size, dtype=np.float32)
+  model_transform = None
+
+  while True:
+    buf = vipc_client.recv()
+    if buf is None:
+      continue
+
+    if model_transform is None:
+      cam = _os_fisheye if buf.width == _os_fisheye.width else _ar_ox_fisheye
+      model_transform = np.linalg.inv(np.dot(dmonitoringmodel_intrinsics, np.linalg.inv(cam.intrinsics))).astype(np.float32)
+
+    sm.update(0)
+    if sm.updated["liveCalibration"]:
+      calib_rpy = get_calibrated_rpy(sm["liveCalibration"])
+      calib[:] = calib_rpy if calib_rpy is not None else np.zeros_like(calib)
+
+    t1 = time.perf_counter()
+    model_output, gpu_execution_time = model.run(buf, calib, model_transform)
+    t2 = time.perf_counter()
+    raw_pred = model_output.tobytes() if SEND_RAW_PRED else b''
+    model_output = slice_outputs(model_output, model.output_slices)
+    model_output = parse_model_output(model_output)
+    model_output['raw_pred'] = raw_pred
+    msg = get_driverstate_packet(model_output, vipc_client.frame_id, vipc_client.timestamp_sof, t2 - t1, gpu_execution_time)
+    pm.send("driverStateV2", msg)
+
+
+if __name__ == "__main__":
+  try:
+    main()
+  except KeyboardInterrupt:
+    cloudlog.warning("got SIGINT")
--- a/selfdrive/modeld/fill_model_msg.py
+++ b/selfdrive/modeld/fill_model_msg.py
@@ -0,0 +1,194 @@
+import os
+import capnp
+import numpy as np
+from cereal import log
+from openpilot.selfdrive.modeld.constants import ModelConstants, Plan, Meta
+from openpilot.iqpilot.selfdrive.iqmodeld.models.helpers import plan_x_idxs_helper
+
+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_xyz_poly(builder, degree, x, y, z):
+  xyz = np.stack([x, y, z], axis=1)
+  coeffs = np.polynomial.polynomial.polyfit(ModelConstants.T_IDXS, xyz, deg=degree)
+  builder.xCoefficients = coeffs[:, 0].tolist()
+  builder.yCoefficients = coeffs[:, 1].tolist()
+  builder.zCoefficients = coeffs[:, 2].tolist()
+
+def fill_lane_line_meta(builder, lane_lines, lane_line_probs):
+  builder.leftY = lane_lines[1].y[0]
+  builder.leftProb = lane_line_probs[1]
+  builder.rightY = lane_lines[2].y[0]
+  builder.rightProb = lane_line_probs[2]
+
+def fill_model_msg(base_msg: capnp._DynamicStructBuilder, extended_msg: capnp._DynamicStructBuilder,
+                   net_output_data: dict[str, np.ndarray], action: log.ModelDataV2.Action,
+                   publish_state: PublishState, vipc_frame_id: int, vipc_frame_id_extra: int,
+                   frame_id: int, frame_drop: float, timestamp_eof: int, model_execution_time: float,
+                   valid: bool) -> None:
+  frame_age = frame_id - vipc_frame_id if frame_id > vipc_frame_id else 0
+  frame_drop_perc = frame_drop * 100
+  extended_msg.valid = valid
+  base_msg.valid = valid
+
+  driving_model_data = base_msg.drivingModelData
+
+  driving_model_data.frameId = vipc_frame_id
+  driving_model_data.frameIdExtra = vipc_frame_id_extra
+  driving_model_data.frameDropPerc = frame_drop_perc
+  driving_model_data.modelExecutionTime = model_execution_time
+
+  driving_model_data.action = action
+
+  modelV2 = extended_msg.modelV2
+  modelV2.frameId = vipc_frame_id
+  modelV2.frameIdExtra = vipc_frame_id_extra
+  modelV2.frameAge = frame_age
+  modelV2.frameDropPerc = frame_drop_perc
+  modelV2.timestampEof = timestamp_eof
+  modelV2.modelExecutionTime = model_execution_time
+
+  # plan
+  fill_xyzt(modelV2.position, ModelConstants.T_IDXS, *net_output_data['plan'][0,:,Plan.POSITION].T, *net_output_data['plan_stds'][0,:,Plan.POSITION].T)
+  fill_xyzt(modelV2.velocity, ModelConstants.T_IDXS, *net_output_data['plan'][0,:,Plan.VELOCITY].T)
+  fill_xyzt(modelV2.acceleration, ModelConstants.T_IDXS, *net_output_data['plan'][0,:,Plan.ACCELERATION].T)
+  fill_xyzt(modelV2.orientation, ModelConstants.T_IDXS, *net_output_data['plan'][0,:,Plan.T_FROM_CURRENT_EULER].T)
+  fill_xyzt(modelV2.orientationRate, ModelConstants.T_IDXS, *net_output_data['plan'][0,:,Plan.ORIENTATION_RATE].T)
+
+  # poly path
+  fill_xyz_poly(driving_model_data.path, ModelConstants.POLY_PATH_DEGREE, *net_output_data['plan'][0,:,Plan.POSITION].T)
+
+  # action
+  modelV2.action = action
+
+  # times at X_IDXS of edges and lines
+  LINE_T_IDXS: list[float] = plan_x_idxs_helper(ModelConstants, Plan, net_output_data)
+
+  # lane lines
+  modelV2.init('laneLines', 4)
+  for i in range(4):
+    lane_line = modelV2.laneLines[i]
+    fill_xyzt(lane_line, LINE_T_IDXS, np.array(ModelConstants.X_IDXS), net_output_data['lane_lines'][0,i,:,0], net_output_data['lane_lines'][0,i,:,1])
+  modelV2.laneLineStds = net_output_data['lane_lines_stds'][0,:,0,0].tolist()
+  modelV2.laneLineProbs = net_output_data['lane_lines_prob'][0,1::2].tolist()
+
+  fill_lane_line_meta(driving_model_data.laneLineMeta, modelV2.laneLines, modelV2.laneLineProbs)
+
+  # road edges
+  modelV2.init('roadEdges', 2)
+  for i in range(2):
+    road_edge = modelV2.roadEdges[i]
+    fill_xyzt(road_edge, LINE_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()
+  disengage_predictions.gasPressProbs = net_output_data['meta'][0,Meta.GAS_PRESS].tolist()
+  disengage_predictions.brakePressProbs = net_output_data['meta'][0,Meta.BRAKE_PRESS].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()
+
+  # confidence
+  if vipc_frame_id % (2*ModelConstants.MODEL_RUN_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()
--- a/selfdrive/modeld/get_model_metadata.py
+++ b/selfdrive/modeld/get_model_metadata.py
@@ -0,0 +1,37 @@
+#!/usr/bin/env python3
+import sys
+import pathlib
+import onnx
+import codecs
+import pickle
+from typing import Any
+
+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
+
+def get_metadata_value_by_name(model:onnx.ModelProto, name:str) -> str | Any:
+  for prop in model.metadata_props:
+    if prop.key == name:
+      return prop.value
+  return None
+
+if __name__ == "__main__":
+  model_path = pathlib.Path(sys.argv[1])
+  model = onnx.load(str(model_path))
+  output_slices = get_metadata_value_by_name(model, 'output_slices')
+  assert output_slices is not None, 'output_slices not found in metadata'
+
+  metadata = {
+    'model_checkpoint': get_metadata_value_by_name(model, 'model_checkpoint'),
+    'output_slices': pickle.loads(codecs.decode(output_slices.encode(), "base64")),
+    'input_shapes': dict([get_name_and_shape(x) for x in model.graph.input]),
+    '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}')
--- a/selfdrive/modeld/modeld.py
+++ b/selfdrive/modeld/modeld.py
@@ -0,0 +1 @@
+LAT_SMOOTH_SECONDS = 0.0
--- a/selfdrive/modeld/models/README.md
+++ b/selfdrive/modeld/models/README.md
@@ -0,0 +1,66 @@
+## Neural networks in openpilot
+To view the architecture of the ONNX networks, you can use [netron](https://netron.app/)
+
+## Driving Model (vision model + temporal policy model)
+### Vision inputs (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
+### Policy inputs
+* **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
+* **lateral control params**
+  * speed and steering delay for predicting the desired curvature: 2
+* **previous desired curvatures**
+  * vector of previously predicted desired curvatures: 100 * 1
+* **feature buffer**
+  * a buffer of intermediate features including the current feature to form a 5 seconds temporal context (at 20FPS) : 100 * 512
+
+
+### Driving Model output format (Full size: XXX x float32)
+Refer to **slice_outputs** and **parse_vision_outputs/parse_policy_outputs** in modeld.
+
+
+## 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/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 1
+    * left hand drive probability: 1
--- a/selfdrive/modeld/models/init.py
+++ b/selfdrive/modeld/models/init.py
--- a/selfdrive/modeld/models/commonmodel.cc
+++ b/selfdrive/modeld/models/commonmodel.cc
@@ -0,0 +1,64 @@
+#include "selfdrive/modeld/models/commonmodel.h"
+
+#include <cmath>
+#include <cstring>
+
+#include "common/clutil.h"
+
+DrivingModelFrame::DrivingModelFrame(cl_device_id device_id, cl_context context, int _temporal_skip) : ModelFrame(device_id, context) {
+  input_frames = std::make_unique<uint8_t[]>(buf_size);
+  temporal_skip = _temporal_skip;
+  input_frames_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE, buf_size, NULL, &err));
+  img_buffer_20hz_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE, (temporal_skip+1)*frame_size_bytes, NULL, &err));
+  region.origin = temporal_skip * frame_size_bytes;
+  region.size = frame_size_bytes;
+  last_img_cl = CL_CHECK_ERR(clCreateSubBuffer(img_buffer_20hz_cl, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, &region, &err));
+
+  loadyuv_init(&loadyuv, context, device_id, MODEL_WIDTH, MODEL_HEIGHT);
+  init_transform(device_id, context, MODEL_WIDTH, MODEL_HEIGHT);
+}
+
+cl_mem* DrivingModelFrame::prepare(cl_mem yuv_cl, int frame_width, int frame_height, int frame_stride, int frame_uv_offset, const mat3& projection) {
+  run_transform(yuv_cl, MODEL_WIDTH, MODEL_HEIGHT, frame_width, frame_height, frame_stride, frame_uv_offset, projection);
+
+  for (int i = 0; i < temporal_skip; i++) {
+    CL_CHECK(clEnqueueCopyBuffer(q, img_buffer_20hz_cl, img_buffer_20hz_cl, (i+1)*frame_size_bytes, i*frame_size_bytes, frame_size_bytes, 0, nullptr, nullptr));
+  }
+  loadyuv_queue(&loadyuv, q, y_cl, u_cl, v_cl, last_img_cl);
+
+  copy_queue(&loadyuv, q, img_buffer_20hz_cl, input_frames_cl, 0, 0, frame_size_bytes);
+  copy_queue(&loadyuv, q, last_img_cl, input_frames_cl, 0, frame_size_bytes, frame_size_bytes);
+
+  // NOTE: Since thneed is using a different command queue, this clFinish is needed to ensure the image is ready.
+  clFinish(q);
+  return &input_frames_cl;
+}
+
+DrivingModelFrame::~DrivingModelFrame() {
+  deinit_transform();
+  loadyuv_destroy(&loadyuv);
+  CL_CHECK(clReleaseMemObject(input_frames_cl));
+  CL_CHECK(clReleaseMemObject(img_buffer_20hz_cl));
+  CL_CHECK(clReleaseMemObject(last_img_cl));
+  CL_CHECK(clReleaseCommandQueue(q));
+}
+
+
+MonitoringModelFrame::MonitoringModelFrame(cl_device_id device_id, cl_context context) : ModelFrame(device_id, context) {
+  input_frames = std::make_unique<uint8_t[]>(buf_size);
+  input_frame_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE, buf_size, NULL, &err));
+
+  init_transform(device_id, context, MODEL_WIDTH, MODEL_HEIGHT);
+}
+
+cl_mem* MonitoringModelFrame::prepare(cl_mem yuv_cl, int frame_width, int frame_height, int frame_stride, int frame_uv_offset, const mat3& projection) {
+  run_transform(yuv_cl, MODEL_WIDTH, MODEL_HEIGHT, frame_width, frame_height, frame_stride, frame_uv_offset, projection);
+  clFinish(q);
+  return &y_cl;
+}
+
+MonitoringModelFrame::~MonitoringModelFrame() {
+  deinit_transform();
+  CL_CHECK(clReleaseMemObject(input_frame_cl));
+  CL_CHECK(clReleaseCommandQueue(q));
+}
--- a/selfdrive/modeld/models/commonmodel.h
+++ b/selfdrive/modeld/models/commonmodel.h
@@ -0,0 +1,97 @@
+#pragma once
+
+#include <cfloat>
+#include <cstdlib>
+#include <cassert>
+
+#include <memory>
+
+#define CL_USE_DEPRECATED_OPENCL_1_2_APIS
+#ifdef __APPLE__
+#include <OpenCL/cl.h>
+#else
+#include <CL/cl.h>
+#endif
+
+#include "common/mat.h"
+#include "selfdrive/modeld/transforms/loadyuv.h"
+#include "selfdrive/modeld/transforms/transform.h"
+
+class ModelFrame {
+public:
+  ModelFrame(cl_device_id device_id, cl_context context) {
+    q = CL_CHECK_ERR(clCreateCommandQueue(context, device_id, 0, &err));
+  }
+  virtual ~ModelFrame() {}
+  virtual cl_mem* prepare(cl_mem yuv_cl, int frame_width, int frame_height, int frame_stride, int frame_uv_offset, const mat3& projection) { return NULL; }
+  uint8_t* buffer_from_cl(cl_mem *in_frames, int buffer_size) {
+    CL_CHECK(clEnqueueReadBuffer(q, *in_frames, CL_TRUE, 0, buffer_size, input_frames.get(), 0, nullptr, nullptr));
+    clFinish(q);
+    return &input_frames[0];
+  }
+
+  int MODEL_WIDTH;
+  int MODEL_HEIGHT;
+  int MODEL_FRAME_SIZE;
+  int buf_size;
+
+protected:
+  cl_mem y_cl, u_cl, v_cl;
+  Transform transform;
+  cl_command_queue q;
+  std::unique_ptr<uint8_t[]> input_frames;
+
+  void init_transform(cl_device_id device_id, cl_context context, int model_width, int model_height) {
+    y_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE, model_width * model_height, NULL, &err));
+    u_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE, (model_width / 2) * (model_height / 2), NULL, &err));
+    v_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE, (model_width / 2) * (model_height / 2), NULL, &err));
+    transform_init(&transform, context, device_id);
+  }
+
+  void deinit_transform() {
+    transform_destroy(&transform);
+    CL_CHECK(clReleaseMemObject(v_cl));
+    CL_CHECK(clReleaseMemObject(u_cl));
+    CL_CHECK(clReleaseMemObject(y_cl));
+  }
+
+  void run_transform(cl_mem yuv_cl, int model_width, int model_height, int frame_width, int frame_height, int frame_stride, int frame_uv_offset, const mat3& projection) {
+    transform_queue(&transform, q,
+        yuv_cl, frame_width, frame_height, frame_stride, frame_uv_offset,
+        y_cl, u_cl, v_cl, model_width, model_height, projection);
+  }
+};
+
+class DrivingModelFrame : public ModelFrame {
+public:
+  DrivingModelFrame(cl_device_id device_id, cl_context context, int _temporal_skip);
+  ~DrivingModelFrame();
+  cl_mem* prepare(cl_mem yuv_cl, int frame_width, int frame_height, int frame_stride, int frame_uv_offset, const mat3& projection);
+
+  const int MODEL_WIDTH = 512;
+  const int MODEL_HEIGHT = 256;
+  const int MODEL_FRAME_SIZE = MODEL_WIDTH * MODEL_HEIGHT * 3 / 2;
+  const int buf_size = MODEL_FRAME_SIZE * 2; // 2 frames are temporal_skip frames apart
+  const size_t frame_size_bytes = MODEL_FRAME_SIZE * sizeof(uint8_t);
+
+private:
+  LoadYUVState loadyuv;
+  cl_mem img_buffer_20hz_cl, last_img_cl, input_frames_cl;
+  cl_buffer_region region;
+  int temporal_skip;
+};
+
+class MonitoringModelFrame : public ModelFrame {
+public:
+  MonitoringModelFrame(cl_device_id device_id, cl_context context);
+  ~MonitoringModelFrame();
+  cl_mem* prepare(cl_mem yuv_cl, int frame_width, int frame_height, int frame_stride, int frame_uv_offset, const mat3& projection);
+
+  const int MODEL_WIDTH = 1440;
+  const int MODEL_HEIGHT = 960;
+  const int MODEL_FRAME_SIZE = MODEL_WIDTH * MODEL_HEIGHT;
+  const int buf_size = MODEL_FRAME_SIZE;
+
+private:
+  cl_mem input_frame_cl;
+};
--- a/selfdrive/modeld/models/commonmodel.pxd
+++ b/selfdrive/modeld/models/commonmodel.pxd
@@ -0,0 +1,27 @@
+# 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)
+  void cl_release_context(cl_context)
+
+cdef extern from "selfdrive/modeld/models/commonmodel.h":
+  cppclass ModelFrame:
+    int buf_size
+    unsigned char * buffer_from_cl(cl_mem*, int);
+    cl_mem * prepare(cl_mem, int, int, int, int, mat3)
+
+  cppclass DrivingModelFrame:
+    int buf_size
+    DrivingModelFrame(cl_device_id, cl_context, int)
+
+  cppclass MonitoringModelFrame:
+    int buf_size
+    MonitoringModelFrame(cl_device_id, cl_context)
--- a/selfdrive/modeld/models/commonmodel_pyx.pxd
+++ b/selfdrive/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*)
--- a/selfdrive/modeld/models/commonmodel_pyx.pyx
+++ b/selfdrive/modeld/models/commonmodel_pyx.pyx
@@ -0,0 +1,74 @@
+# distutils: language = c++
+# cython: c_string_encoding=ascii, language_level=3
+
+import numpy as np
+cimport numpy as cnp
+from libc.string cimport memcpy
+from libc.stdint cimport uintptr_t
+
+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, cl_release_context
+from .commonmodel cimport mat3, ModelFrame as cppModelFrame, DrivingModelFrame as cppDrivingModelFrame, MonitoringModelFrame as cppMonitoringModelFrame
+
+
+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)
+
+  def __dealloc__(self):
+    if self.context:
+      cl_release_context(self.context)
+
+cdef class CLMem:
+  @staticmethod
+  cdef create(void * cmem):
+    mem = CLMem()
+    mem.mem = <cl_mem*> cmem
+    return mem
+
+  @property
+  def mem_address(self):
+    return <uintptr_t>(self.mem)
+
+def cl_from_visionbuf(VisionBuf buf):
+  return CLMem.create(<void*>&buf.buf.buf_cl)
+
+
+cdef class ModelFrame:
+  cdef cppModelFrame * frame
+  cdef int buf_size
+
+  def __dealloc__(self):
+    del self.frame
+
+  def prepare(self, VisionBuf buf, float[:] projection):
+    cdef mat3 cprojection
+    memcpy(cprojection.v, &projection[0], 9*sizeof(float))
+    cdef cl_mem * data
+    data = self.frame.prepare(buf.buf.buf_cl, buf.width, buf.height, buf.stride, buf.uv_offset, cprojection)
+    return CLMem.create(data)
+
+  def buffer_from_cl(self, CLMem in_frames):
+    cdef unsigned char * data2
+    data2 = self.frame.buffer_from_cl(in_frames.mem, self.buf_size)
+    return np.asarray(<cnp.uint8_t[:self.buf_size]> data2)
+
+
+cdef class DrivingModelFrame(ModelFrame):
+  cdef cppDrivingModelFrame * _frame
+
+  def __cinit__(self, CLContext context, int temporal_skip):
+    self._frame = new cppDrivingModelFrame(context.device_id, context.context, temporal_skip)
+    self.frame = <cppModelFrame*>(self._frame)
+    self.buf_size = self._frame.buf_size
+
+cdef class MonitoringModelFrame(ModelFrame):
+  cdef cppMonitoringModelFrame * _frame
+
+  def __cinit__(self, CLContext context):
+    self._frame = new cppMonitoringModelFrame(context.device_id, context.context)
+    self.frame = <cppModelFrame*>(self._frame)
+    self.buf_size = self._frame.buf_size
+
--- a/selfdrive/modeld/models/dmonitoring_model.onnx
+++ b/selfdrive/modeld/models/dmonitoring_model.onnx
--- a/selfdrive/modeld/models/dmonitoring_model_metadata.pkl
+++ b/selfdrive/modeld/models/dmonitoring_model_metadata.pkl
--- a/selfdrive/modeld/models/dmonitoring_model_tinygrad.pkl
+++ b/selfdrive/modeld/models/dmonitoring_model_tinygrad.pkl
--- a/selfdrive/modeld/models/prebuilt_check.json
+++ b/selfdrive/modeld/models/prebuilt_check.json
@@ -0,0 +1,9 @@
+{
+  "dmonitoring_model": {
+    "outputs": {
+      "dmonitoring_model_metadata.pkl": "31a86ab7a92dc0af088b15787a440dd3b210aa662e445a15145900e559a1b5c3",
+      "dmonitoring_model_tinygrad.pkl": "806c0ea75df6bf6dfeb81b832314c68e31df5865a52d0359e6eeb76d93ad2b52"
+    },
+    "signature": "e1eeb5ce45774a816c8da2394e6ee35ebf700b71dff345e0141dbab8ff592349"
+  }
+}
--- a/selfdrive/modeld/parse_model_outputs.py
+++ b/selfdrive/modeld/parse_model_outputs.py
@@ -0,0 +1,122 @@
+import numpy as np
+from openpilot.selfdrive.modeld.constants import ModelConstants
+
+def safe_exp(x, out=None):
+  # -11 is around 10**14, more causes float16 overflow
+  return np.exp(np.clip(x, -np.inf, 11), out=out)
+
+def sigmoid(x):
+  return 1. / (1. + safe_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:
+    safe_exp(x, out=x)
+  else:
+    x = safe_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):
+    missing = name not in outs
+    if missing and not self.ignore_missing:
+      raise ValueError(f"Missing output {name}")
+    return missing
+
+  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))
+
+    n_values = (raw.shape[2] - out_N)//2
+    pred_mu = raw[:,:,:n_values]
+    pred_std = safe_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 is_mhp(self, outs, name, shape):
+    if self.check_missing(outs, name):
+      return False
+    if outs[name].shape[1] == 2 * shape:
+      return False
+    return True
+
+  def parse_vision_outputs(self, outs: dict[str, np.ndarray]) -> dict[str, np.ndarray]:
+    self.parse_mdn('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('road_transform', outs, in_N=0, out_N=0, out_shape=(ModelConstants.POSE_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_binary_crossentropy('lane_lines_prob', outs)
+    self.parse_categorical_crossentropy('desire_pred', outs, out_shape=(ModelConstants.DESIRE_PRED_LEN,ModelConstants.DESIRE_PRED_WIDTH))
+    self.parse_binary_crossentropy('meta', outs)
+    self.parse_binary_crossentropy('lead_prob', outs)
+    lead_mhp = self.is_mhp(outs, 'lead', ModelConstants.LEAD_MHP_SELECTION * ModelConstants.LEAD_TRAJ_LEN * ModelConstants.LEAD_WIDTH)
+    lead_in_N, lead_out_N = (ModelConstants.LEAD_MHP_N, ModelConstants.LEAD_MHP_SELECTION) if lead_mhp else (0, 0)
+    lead_out_shape = (ModelConstants.LEAD_TRAJ_LEN, ModelConstants.LEAD_WIDTH) if lead_mhp else \
+        (ModelConstants.LEAD_MHP_SELECTION, ModelConstants.LEAD_TRAJ_LEN, ModelConstants.LEAD_WIDTH)
+    self.parse_mdn('lead', outs, in_N=lead_in_N, out_N=lead_out_N, out_shape=lead_out_shape)
+    return outs
+
+  def parse_policy_outputs(self, outs: dict[str, np.ndarray]) -> dict[str, np.ndarray]:
+    plan_mhp = self.is_mhp(outs, 'plan',  ModelConstants.IDX_N * ModelConstants.PLAN_WIDTH)
+    plan_in_N, plan_out_N = (ModelConstants.PLAN_MHP_N, ModelConstants.PLAN_MHP_SELECTION) if plan_mhp else (0, 0)
+    self.parse_mdn('plan', outs, in_N=plan_in_N, out_N=plan_out_N, out_shape=(ModelConstants.IDX_N, ModelConstants.PLAN_WIDTH))
+    self.parse_categorical_crossentropy('desire_state', outs, out_shape=(ModelConstants.DESIRE_PRED_WIDTH,))
+    return outs
+
+  def parse_outputs(self, outs: dict[str, np.ndarray]) -> dict[str, np.ndarray]:
+    outs = self.parse_vision_outputs(outs)
+    outs = self.parse_policy_outputs(outs)
+    return outs
--- a/selfdrive/modeld/prebuilt_models.py
+++ b/selfdrive/modeld/prebuilt_models.py
@@ -0,0 +1,117 @@
+#!/usr/bin/env python3
+import functools
+import hashlib
+import json
+import sys
+from pathlib import Path
+
+MODELD_DIR = Path(__file__).resolve().parent
+MODELS_DIR = MODELD_DIR / 'models'
+BASEDIR = MODELD_DIR.parents[1]
+TINYGRAD_DIR = BASEDIR / 'tinygrad_repo'
+METADATA_SCRIPT = MODELD_DIR / 'get_model_metadata.py'
+
+MODEL_NAMES = ['dmonitoring_model']
+
+
+def _hash_file(h, path: Path) -> None:
+  with open(path, 'rb') as f:
+    for chunk in iter(lambda: f.read(1024 * 1024), b''):
+      h.update(chunk)
+
+
+def _file_sha256(path: Path) -> str:
+  h = hashlib.sha256()
+  _hash_file(h, path)
+  return h.hexdigest()
+
+
+@functools.lru_cache(maxsize=1)
+def _tinygrad_digest() -> str:
+  h = hashlib.sha256()
+  # mirror the SConscript glob: no hidden files/dirs (also excludes .git, which differs per clone), no pycache
+  def _included(p: Path) -> bool:
+    rel = p.relative_to(TINYGRAD_DIR).parts
+    return p.is_file() and '__pycache__' not in rel and not any(part.startswith('.') for part in rel)
+  files = sorted(p for p in TINYGRAD_DIR.rglob('*') if _included(p))
+  for p in files:
+    h.update(str(p.relative_to(BASEDIR)).encode())
+    _hash_file(h, p)
+  return h.hexdigest()
+
+
+CHECK_PATH = MODELS_DIR / 'prebuilt_check.json'
+
+
+def _load_checks() -> dict:
+  try:
+    data = json.loads(CHECK_PATH.read_text())
+  except (json.JSONDecodeError, OSError):
+    return {}
+  return data if isinstance(data, dict) else {}
+
+
+def _output_names(model_name: str) -> list[str]:
+  return [f'{model_name}_tinygrad.pkl', f'{model_name}_metadata.pkl']
+
+
+def compute_signature(model_name: str, flags: str) -> str:
+  h = hashlib.sha256()
+  h.update(flags.encode())
+  h.update(_tinygrad_digest().encode())
+  _hash_file(h, METADATA_SCRIPT)
+  _hash_file(h, MODELS_DIR / f'{model_name}.onnx')
+  return h.hexdigest()
+
+
+def outputs_match(model_name: str) -> bool:
+  """The committed artifacts on disk are exactly the ones the check file pins."""
+  data = _load_checks().get(model_name, {})
+  outputs = data.get('outputs', {})
+  if set(outputs) != set(_output_names(model_name)):
+    return False
+  for fn, expected in outputs.items():
+    p = MODELS_DIR / fn
+    if not p.is_file() or _file_sha256(p) != expected:
+      return False
+  return True
+
+
+def packaged_prebuilt_matches(model_name: str) -> bool:
+  return not (MODELS_DIR / f'{model_name}.onnx').is_file() and outputs_match(model_name)
+
+
+def verify_prebuilt(model_name: str, flags: str) -> bool:
+  if not (MODELS_DIR / f'{model_name}.onnx').is_file():
+    return False
+  data = _load_checks().get(model_name, {})
+  if data.get('signature') != compute_signature(model_name, flags):
+    return False
+  return outputs_match(model_name)
+
+
+def write_check(model_name: str, flags: str) -> None:
+  outputs = {}
+  for fn in _output_names(model_name):
+    p = MODELS_DIR / fn
+    if not p.is_file():
+      raise FileNotFoundError(f'missing build output: {p}')
+    outputs[fn] = _file_sha256(p)
+  checks = _load_checks()
+  checks[model_name] = {'signature': compute_signature(model_name, flags), 'outputs': outputs}
+  CHECK_PATH.write_text(json.dumps(checks, indent=2, sort_keys=True) + '\n')
+
+
+def _larch64_flags() -> str:
+  return "DEV=QCOM IMAGE=2 FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0 OPENPILOT_HACKS=1"
+
+
+if __name__ == '__main__':
+  mode = sys.argv[1] if len(sys.argv) > 1 else 'verify'
+  flags = _larch64_flags()
+  for name in MODEL_NAMES:
+    if mode == 'write':
+      write_check(name, flags)
+      print(f'{name}: check written')
+    else:
+      print(f'{name}: {"OK" if verify_prebuilt(name, flags) else "STALE"}')
--- a/selfdrive/modeld/runners/tinygrad_helpers.py
+++ b/selfdrive/modeld/runners/tinygrad_helpers.py
@@ -0,0 +1,8 @@
+
+from tinygrad.tensor import Tensor
+from tinygrad.helpers import to_mv
+
+def qcom_tensor_from_opencl_address(opencl_address, shape, dtype):
+  cl_buf_desc_ptr = to_mv(opencl_address, 8).cast('Q')[0]
+  rawbuf_ptr = to_mv(cl_buf_desc_ptr, 0x100).cast('Q')[20] # offset 0xA0 is a raw gpu pointer.
+  return Tensor.from_blob(rawbuf_ptr, shape, dtype=dtype, device='QCOM')
--- a/selfdrive/modeld/test_prebuilt_models.py
+++ b/selfdrive/modeld/test_prebuilt_models.py
@@ -0,0 +1,51 @@
+import hashlib
+import json
+
+from openpilot.selfdrive.modeld import prebuilt_models
+
+
+def write_outputs(models_dir, check_path):
+  outputs = {}
+  for name, contents in {
+    'dmonitoring_model_tinygrad.pkl': b'tinygrad',
+    'dmonitoring_model_metadata.pkl': b'metadata',
+  }.items():
+    (models_dir / name).write_bytes(contents)
+    outputs[name] = hashlib.sha256(contents).hexdigest()
+  check_path.write_text(json.dumps({'dmonitoring_model': {'outputs': outputs}}))
+
+
+def test_packaged_prebuilt_without_onnx(tmp_path, monkeypatch):
+  models_dir = tmp_path / 'models'
+  models_dir.mkdir()
+  check_path = models_dir / 'prebuilt_check.json'
+  write_outputs(models_dir, check_path)
+  monkeypatch.setattr(prebuilt_models, 'MODELS_DIR', models_dir)
+  monkeypatch.setattr(prebuilt_models, 'CHECK_PATH', check_path)
+
+  assert prebuilt_models.packaged_prebuilt_matches('dmonitoring_model')
+  assert not prebuilt_models.verify_prebuilt('dmonitoring_model', 'flags')
+
+
+def test_packaged_prebuilt_rejects_corrupt_output(tmp_path, monkeypatch):
+  models_dir = tmp_path / 'models'
+  models_dir.mkdir()
+  check_path = models_dir / 'prebuilt_check.json'
+  write_outputs(models_dir, check_path)
+  (models_dir / 'dmonitoring_model_tinygrad.pkl').write_bytes(b'corrupt')
+  monkeypatch.setattr(prebuilt_models, 'MODELS_DIR', models_dir)
+  monkeypatch.setattr(prebuilt_models, 'CHECK_PATH', check_path)
+
+  assert not prebuilt_models.packaged_prebuilt_matches('dmonitoring_model')
+
+
+def test_source_checkout_is_not_packaged_prebuilt(tmp_path, monkeypatch):
+  models_dir = tmp_path / 'models'
+  models_dir.mkdir()
+  check_path = models_dir / 'prebuilt_check.json'
+  write_outputs(models_dir, check_path)
+  (models_dir / 'dmonitoring_model.onnx').write_bytes(b'onnx')
+  monkeypatch.setattr(prebuilt_models, 'MODELS_DIR', models_dir)
+  monkeypatch.setattr(prebuilt_models, 'CHECK_PATH', check_path)
+
+  assert not prebuilt_models.packaged_prebuilt_matches('dmonitoring_model')
--- a/selfdrive/modeld/transforms/loadyuv.cc
+++ b/selfdrive/modeld/transforms/loadyuv.cc
@@ -0,0 +1,76 @@
+#include "selfdrive/modeld/transforms/loadyuv.h"
+
+#include <cassert>
+#include <cstdio>
+#include <cstring>
+
+void loadyuv_init(LoadYUVState* s, cl_context ctx, cl_device_id device_id, int width, int height) {
+  memset(s, 0, sizeof(*s));
+
+  s->width = width;
+  s->height = height;
+
+  char args[1024];
+  snprintf(args, sizeof(args),
+           "-cl-fast-relaxed-math -cl-denorms-are-zero "
+           "-DTRANSFORMED_WIDTH=%d -DTRANSFORMED_HEIGHT=%d",
+           width, height);
+  cl_program prg = cl_program_from_file(ctx, device_id, LOADYUV_PATH, args);
+
+  s->loadys_krnl = CL_CHECK_ERR(clCreateKernel(prg, "loadys", &err));
+  s->loaduv_krnl = CL_CHECK_ERR(clCreateKernel(prg, "loaduv", &err));
+  s->copy_krnl = CL_CHECK_ERR(clCreateKernel(prg, "copy", &err));
+
+  // done with this
+  CL_CHECK(clReleaseProgram(prg));
+}
+
+void loadyuv_destroy(LoadYUVState* s) {
+  CL_CHECK(clReleaseKernel(s->loadys_krnl));
+  CL_CHECK(clReleaseKernel(s->loaduv_krnl));
+  CL_CHECK(clReleaseKernel(s->copy_krnl));
+}
+
+void loadyuv_queue(LoadYUVState* s, cl_command_queue q,
+                   cl_mem y_cl, cl_mem u_cl, cl_mem v_cl,
+                   cl_mem out_cl) {
+  cl_int global_out_off = 0;
+
+  CL_CHECK(clSetKernelArg(s->loadys_krnl, 0, sizeof(cl_mem), &y_cl));
+  CL_CHECK(clSetKernelArg(s->loadys_krnl, 1, sizeof(cl_mem), &out_cl));
+  CL_CHECK(clSetKernelArg(s->loadys_krnl, 2, sizeof(cl_int), &global_out_off));
+
+  const size_t loadys_work_size = (s->width*s->height)/8;
+  CL_CHECK(clEnqueueNDRangeKernel(q, s->loadys_krnl, 1, NULL,
+                               &loadys_work_size, NULL, 0, 0, NULL));
+
+  const size_t loaduv_work_size = ((s->width/2)*(s->height/2))/8;
+  global_out_off += (s->width*s->height);
+
+  CL_CHECK(clSetKernelArg(s->loaduv_krnl, 0, sizeof(cl_mem), &u_cl));
+  CL_CHECK(clSetKernelArg(s->loaduv_krnl, 1, sizeof(cl_mem), &out_cl));
+  CL_CHECK(clSetKernelArg(s->loaduv_krnl, 2, sizeof(cl_int), &global_out_off));
+
+  CL_CHECK(clEnqueueNDRangeKernel(q, s->loaduv_krnl, 1, NULL,
+                               &loaduv_work_size, NULL, 0, 0, NULL));
+
+  global_out_off += (s->width/2)*(s->height/2);
+
+  CL_CHECK(clSetKernelArg(s->loaduv_krnl, 0, sizeof(cl_mem), &v_cl));
+  CL_CHECK(clSetKernelArg(s->loaduv_krnl, 1, sizeof(cl_mem), &out_cl));
+  CL_CHECK(clSetKernelArg(s->loaduv_krnl, 2, sizeof(cl_int), &global_out_off));
+
+  CL_CHECK(clEnqueueNDRangeKernel(q, s->loaduv_krnl, 1, NULL,
+                               &loaduv_work_size, NULL, 0, 0, NULL));
+}
+
+void copy_queue(LoadYUVState* s, cl_command_queue q, cl_mem src, cl_mem dst,
+                 size_t src_offset, size_t dst_offset, size_t size) {
+  CL_CHECK(clSetKernelArg(s->copy_krnl, 0, sizeof(cl_mem), &src));
+  CL_CHECK(clSetKernelArg(s->copy_krnl, 1, sizeof(cl_mem), &dst));
+  CL_CHECK(clSetKernelArg(s->copy_krnl, 2, sizeof(cl_int), &src_offset));
+  CL_CHECK(clSetKernelArg(s->copy_krnl, 3, sizeof(cl_int), &dst_offset));
+  const size_t copy_work_size = size/8;
+  CL_CHECK(clEnqueueNDRangeKernel(q, s->copy_krnl, 1, NULL,
+                              &copy_work_size, NULL, 0, 0, NULL));
+}
--- a/selfdrive/modeld/transforms/loadyuv.cl
+++ b/selfdrive/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 uchar * 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];
+
+    // 02
+    // 13
+
+    __global uchar* outy0;
+    __global uchar* 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(ys.s0246, 0, outy0 + (oy/2) * (TRANSFORMED_WIDTH/2) + ox/2);
+    vstore4(ys.s1357, 0, outy1 + (oy/2) * (TRANSFORMED_WIDTH/2) + ox/2);
+}
+
+__kernel void loaduv(__global uchar8 const * const in,
+                     __global uchar8 * out,
+                     int out_offset)
+{
+  const int gid = get_global_id(0);
+  const uchar8 inv = in[gid];
+  out[gid + out_offset / 8] = inv;
+}
+
+__kernel void copy(__global uchar8 * in,
+                   __global uchar8 * out,
+                   int in_offset,
+                   int out_offset)
+{
+  const int gid = get_global_id(0);
+  out[gid + out_offset / 8] = in[gid + in_offset / 8];
+}
--- a/selfdrive/modeld/transforms/loadyuv.h
+++ b/selfdrive/modeld/transforms/loadyuv.h
@@ -0,0 +1,20 @@
+#pragma once
+
+#include "common/clutil.h"
+
+typedef struct {
+  int width, height;
+  cl_kernel loadys_krnl, loaduv_krnl, copy_krnl;
+} LoadYUVState;
+
+void loadyuv_init(LoadYUVState* s, cl_context ctx, cl_device_id device_id, int width, int height);
+
+void loadyuv_destroy(LoadYUVState* s);
+
+void loadyuv_queue(LoadYUVState* s, cl_command_queue q,
+                   cl_mem y_cl, cl_mem u_cl, cl_mem v_cl,
+                   cl_mem out_cl);
+
+
+void copy_queue(LoadYUVState* s, cl_command_queue q, cl_mem src, cl_mem dst,
+                 size_t src_offset, size_t dst_offset, size_t size);
--- a/selfdrive/modeld/transforms/transform.cc
+++ b/selfdrive/modeld/transforms/transform.cc
@@ -0,0 +1,97 @@
+#include "selfdrive/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));
+}
--- a/selfdrive/modeld/transforms/transform.cl
+++ b/selfdrive/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;
+    }
+}
--- a/selfdrive/modeld/transforms/transform.h
+++ b/selfdrive/modeld/transforms/transform.h
@@ -0,0 +1,25 @@
+#pragma once
+
+#define CL_USE_DEPRECATED_OPENCL_1_2_APIS
+#ifdef __APPLE__
+#include <OpenCL/cl.h>
+#else
+#include <CL/cl.h>
+#endif
+
+#include "common/mat.h"
+
+typedef struct {
+  cl_kernel krnl;
+  cl_mem m_y_cl, m_uv_cl;
+} Transform;
+
+void transform_init(Transform* s, cl_context ctx, cl_device_id device_id);
+
+void transform_destroy(Transform* transform);
+
+void transform_queue(Transform* s, cl_command_queue q,
+                     cl_mem 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);