Revert "modeld: single jit (#37758)"

This reverts commit d81d66193f.

selfdrive/modeld/SConscript

@@ -16,6 +16,7 @@ tinygrad_files = ["#"+x for x in glob.glob(env.Dir("#tinygrad_repo").relpath + "
 def estimate_pickle_max_size(onnx_size):
   return 1.2 * onnx_size  + 10 * 1024 * 1024  # 20% + 10MB is plenty
 
+# THREADS=0 is need to prevent bug: https://github.com/tinygrad/tinygrad/issues/14689
 # get fastest TG config
 available = set(Device.get_available_devices())
 if 'CUDA' in available:
@@ -23,10 +24,9 @@ if 'CUDA' in available:
   tg_flags = f'DEV={tg_backend}'
 elif 'QCOM' in available:
   tg_backend = 'QCOM'
-  tg_flags = f'DEV={tg_backend} FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0 OPENPILOT_HACKS=1'
+  tg_flags = f'DEV={tg_backend} FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0'
 else:
   tg_backend = 'CPU' if arch == 'Darwin' else 'CPU:LLVM'
-  # THREADS=0 is need to prevent bug: https://github.com/tinygrad/tinygrad/issues/14689
   tg_flags = f'DEV={tg_backend} THREADS=0'
 
 def write_tg_compiled_flags(target, source, env):
@@ -53,23 +53,14 @@ for model_name in ['driving_vision', 'driving_policy', 'dmonitoring_model']:
 image_flag = {
      'larch64': 'IMAGE=2',
 }.get(arch, 'IMAGE=0')
-script_files = [File(Dir("#selfdrive/modeld").File("compile_modeld.py").abspath)]
-compile_modeld_cmd = f'{tg_flags} {mac_brew_string} {image_flag} python3 {Dir("#selfdrive/modeld").abspath}/compile_modeld.py '
-driving_onnx_deps = [File(f"models/{m}.onnx").abspath for m in ['driving_vision', 'driving_policy']]
-driving_metadata_deps = [File(f"models/{m}_metadata.pkl").abspath for m in ['driving_vision', 'driving_policy']]
-
-from openpilot.selfdrive.modeld.compile_modeld import MODELD_CONFIGS, DM_WARP_CONFIGS
-policy_pkls = [File(cfg.pkl_path).abspath for cfg in MODELD_CONFIGS]
-modeld_targets = policy_pkls + [File(cfg.pkl_path).abspath for cfg in DM_WARP_CONFIGS]
-compile_node = lenv.Command(modeld_targets, tinygrad_files + script_files + driving_onnx_deps + driving_metadata_deps + [chunker_file, compiled_flags_node], compile_modeld_cmd)
-
-# chunk the combined policy pkls
-for policy_pkl in policy_pkls:
-  onnx_sizes_sum = sum(os.path.getsize(f) for f in driving_onnx_deps)
-  chunk_targets = get_chunk_paths(policy_pkl, estimate_pickle_max_size(onnx_sizes_sum))
-  def do_chunk(target, source, env, pkl=policy_pkl, chunks=chunk_targets):
-    chunk_file(pkl, chunks)
-  lenv.Command(chunk_targets, compile_node, do_chunk)
+script_files = [File(Dir("#selfdrive/modeld").File("compile_warp.py").abspath)]
+compile_warp_cmd = f'{tg_flags} {mac_brew_string} python3 {Dir("#selfdrive/modeld").abspath}/compile_warp.py '
+from openpilot.common.transformations.camera import _ar_ox_fisheye, _os_fisheye
+warp_targets = []
+for cam in [_ar_ox_fisheye, _os_fisheye]:
+  w, h = cam.width, cam.height
+  warp_targets += [File(f"models/warp_{w}x{h}_tinygrad.pkl").abspath, File(f"models/dm_warp_{w}x{h}_tinygrad.pkl").abspath]
+lenv.Command(warp_targets, tinygrad_files + script_files + [compiled_flags_node], compile_warp_cmd)
 
 def tg_compile(flags, model_name):
   pythonpath_string = 'PYTHONPATH="${PYTHONPATH}:' + env.Dir("#tinygrad_repo").abspath + '"'
@@ -90,4 +81,7 @@ def tg_compile(flags, model_name):
     do_chunk,
   )
 
-tg_compile(tg_flags, 'dmonitoring_model')
+# Compile small models
+for model_name in ['driving_vision', 'driving_policy', 'dmonitoring_model']:
+  tg_compile(tg_flags, model_name)
+

selfdrive/modeld/compile_modeld.py

@@ -1,289 +0,0 @@
-#!/usr/bin/env python3
-import time
-import pickle
-from dataclasses import dataclass
-from itertools import product
-from functools import partial
-
-import numpy as np
-from tinygrad.tensor import Tensor
-from tinygrad.helpers import Context
-from tinygrad.device import Device
-from tinygrad.engine.jit import TinyJit
-
-from openpilot.selfdrive.modeld.tinygrad_helpers import MODELS_DIR
-from openpilot.system.camerad.cameras.nv12_info import get_nv12_info
-from openpilot.common.transformations.model import MEDMODEL_INPUT_SIZE, DM_INPUT_SIZE
-from openpilot.common.transformations.camera import _ar_ox_fisheye, _os_fisheye
-
-# https://github.com/tinygrad/tinygrad/issues/15682
-from tinygrad.uop.ops import UOp, Ops
-_orig = UOp.__reduce__
-UOp.__reduce__ = lambda self: (UOp.unique, ()) if self.op is Ops.UNIQUE else _orig(self)
-
-
-@dataclass
-class CompileConfig:
-  cam_w: int
-  cam_h: int
-  prepare_only: bool
-  prefix: str
-
-  @property
-  def pkl_path(self):
-    return str(MODELS_DIR / f'{self.prefix}{"warp_" if self.prepare_only else ""}{self.cam_w}x{self.cam_h}_tinygrad.pkl')
-
-CAMERA_CONFIGS = [
-  (_ar_ox_fisheye.width, _ar_ox_fisheye.height),  # tici: 1928x1208
-  (_os_fisheye.width, _os_fisheye.height),        # mici: 1344x760
-]
-MODELD_CONFIGS = [CompileConfig(cam_w, cam_h, prepare_only, 'driving_') for (cam_w, cam_h), prepare_only in product(CAMERA_CONFIGS, [True, False])]
-DM_WARP_CONFIGS = [CompileConfig(cam_w, cam_h, True, 'dm_') for cam_w, cam_h in CAMERA_CONFIGS]
-
-
-UV_SCALE_MATRIX = np.array([[0.5, 0, 0], [0, 0.5, 0], [0, 0, 1]], dtype=np.float32)
-UV_SCALE_MATRIX_INV = np.linalg.inv(UV_SCALE_MATRIX)
-
-
-def warp_perspective_tinygrad(src_flat, M_inv, dst_shape, src_shape, stride_pad):
-  w_dst, h_dst = dst_shape
-  h_src, w_src = src_shape
-
-  x = Tensor.arange(w_dst).reshape(1, w_dst).expand(h_dst, w_dst).reshape(-1)
-  y = Tensor.arange(h_dst).reshape(h_dst, 1).expand(h_dst, w_dst).reshape(-1)
-
-  # inline 3x3 matmul as elementwise to avoid reduce op (enables fusion with gather)
-  src_x = M_inv[0, 0] * x + M_inv[0, 1] * y + M_inv[0, 2]
-  src_y = M_inv[1, 0] * x + M_inv[1, 1] * y + M_inv[1, 2]
-  src_w = M_inv[2, 0] * x + M_inv[2, 1] * y + M_inv[2, 2]
-
-  src_x = src_x / src_w
-  src_y = src_y / src_w
-
-  x_nn_clipped = Tensor.round(src_x).clip(0, w_src - 1).cast('int')
-  y_nn_clipped = Tensor.round(src_y).clip(0, h_src - 1).cast('int')
-  idx = y_nn_clipped * (w_src + stride_pad) + x_nn_clipped
-
-  return src_flat[idx]
-
-
-def frames_to_tensor(frames):
-  H = (frames.shape[0] * 2) // 3
-  W = frames.shape[1]
-  in_img1 = Tensor.cat(frames[0:H:2, 0::2],
-                       frames[1:H:2, 0::2],
-                       frames[0:H:2, 1::2],
-                       frames[1:H:2, 1::2],
-                       frames[H:H+H//4].reshape((H//2, W//2)),
-                       frames[H+H//4:H+H//2].reshape((H//2, W//2)), dim=0).reshape((6, H//2, W//2))
-  return in_img1
-
-
-def make_frame_prepare(cam_w, cam_h, model_w, model_h):
-  stride, y_height, uv_height, _ = get_nv12_info(cam_w, cam_h)
-  uv_offset = stride * y_height
-  stride_pad = stride - cam_w
-
-  def frame_prepare_tinygrad(input_frame, M_inv):
-    # UV_SCALE @ M_inv @ UV_SCALE_INV simplifies to elementwise scaling
-    M_inv_uv = M_inv * Tensor([[1.0, 1.0, 0.5], [1.0, 1.0, 0.5], [2.0, 2.0, 1.0]])
-    # deinterleave NV12 UV plane (UVUV... -> separate U, V)
-    uv = input_frame[uv_offset:uv_offset + uv_height * stride].reshape(uv_height, stride)
-    with Context(SPLIT_REDUCEOP=0):
-      y = warp_perspective_tinygrad(input_frame[:cam_h*stride],
-                                    M_inv, (model_w, model_h),
-                                    (cam_h, cam_w), stride_pad).realize()
-      u = warp_perspective_tinygrad(uv[:cam_h//2, :cam_w:2].flatten(),
-                                    M_inv_uv, (model_w//2, model_h//2),
-                                    (cam_h//2, cam_w//2), 0).realize()
-      v = warp_perspective_tinygrad(uv[:cam_h//2, 1:cam_w:2].flatten(),
-                                    M_inv_uv, (model_w//2, model_h//2),
-                                    (cam_h//2, cam_w//2), 0).realize()
-    yuv = y.cat(u).cat(v).reshape((model_h * 3 // 2, model_w))
-    tensor = frames_to_tensor(yuv)
-    return tensor
-  return frame_prepare_tinygrad
-
-
-def make_input_queues(vision_input_shapes, policy_input_shapes, frame_skip):
-  img = vision_input_shapes['img']  # (1, 12, 128, 256)
-  n_frames = img[1] // 6
-  img_buf_shape = (frame_skip * (n_frames - 1) + 1, 6, img[2], img[3])
-
-  fb = policy_input_shapes['features_buffer']  # (1, 25, 512)
-  dp = policy_input_shapes['desire_pulse']  # (1, 25, 8)
-  tc = policy_input_shapes['traffic_convention']  # (1, 2)
-
-  npy = {
-    'desire': np.zeros(dp[2], dtype=np.float32),
-    'traffic_convention': np.zeros(tc, dtype=np.float32),
-    'tfm': np.zeros((3, 3), dtype=np.float32),
-    'big_tfm': np.zeros((3, 3), dtype=np.float32),
-  }
-  input_queues = {
-    'img_q': Tensor.zeros(img_buf_shape, dtype='uint8').contiguous().realize(),
-    'big_img_q': Tensor.zeros(img_buf_shape, dtype='uint8').contiguous().realize(),
-    'feat_q': Tensor.zeros(frame_skip * (fb[1] - 1) + 1, fb[0], fb[2]).contiguous().realize(),
-    'desire_q': Tensor.zeros(frame_skip * dp[1], dp[0], dp[2]).contiguous().realize(),
-    **{k: Tensor(v, device='NPY').realize() for k, v in npy.items()},
-  }
-  return input_queues, npy
-
-
-def shift_and_sample(buf, new_val, sample_fn):
-  buf.assign(buf[1:].cat(new_val, dim=0).contiguous())
-  return sample_fn(buf)
-
-
-def sample_skip(buf, frame_skip):
-  return buf[::frame_skip].contiguous().flatten(0, 1).unsqueeze(0)
-
-
-def sample_desire(buf, frame_skip):
-  return buf.reshape(-1, frame_skip, *buf.shape[1:]).max(1).flatten(0, 1).unsqueeze(0)
-
-
-def make_warp_dm(cam_w, cam_h, dm_w, dm_h):
-  stride, y_height, _, _ = get_nv12_info(cam_w, cam_h)
-  stride_pad = stride - cam_w
-
-  def warp_dm(input_frame, M_inv):
-    M_inv = M_inv.to(Device.DEFAULT)
-    result = warp_perspective_tinygrad(input_frame[:cam_h*stride], M_inv, (dm_w, dm_h), (cam_h, cam_w), stride_pad).reshape(-1, dm_h * dm_w)
-    return result
-  return warp_dm
-
-
-def make_run_policy(vision_runner, policy_runner, cam_w, cam_h,
-                    vision_features_slice, frame_skip, prepare_only=False):
-  model_w, model_h = MEDMODEL_INPUT_SIZE
-  frame_prepare = make_frame_prepare(cam_w, cam_h, model_w, model_h)
-  sample_skip_fn = partial(sample_skip, frame_skip=frame_skip)
-  sample_desire_fn = partial(sample_desire, frame_skip=frame_skip)
-
-  def run_policy(img_q, big_img_q, feat_q, desire_q, desire, traffic_convention, tfm, big_tfm, frame, big_frame):
-    img = shift_and_sample(img_q, frame_prepare(frame, tfm.to(Device.DEFAULT)).unsqueeze(0), sample_skip_fn)
-    big_img = shift_and_sample(big_img_q, frame_prepare(big_frame, big_tfm.to(Device.DEFAULT)).unsqueeze(0), sample_skip_fn)
-
-    if prepare_only:
-      return img, big_img
-
-    vision_out = next(iter(vision_runner({'img': img, 'big_img': big_img}).values())).cast('float32')
-
-    new_feat = vision_out[:, vision_features_slice].reshape(1, -1).unsqueeze(0)
-    feat_buf = shift_and_sample(feat_q, new_feat, sample_skip_fn)
-    desire_buf = shift_and_sample(desire_q, desire.to(Device.DEFAULT).reshape(1, 1, -1), sample_desire_fn)
-
-    inputs = {'features_buffer': feat_buf, 'desire_pulse': desire_buf, 'traffic_convention': traffic_convention.to(Device.DEFAULT)}
-    policy_out = next(iter(policy_runner(inputs).values())).cast('float32')
-
-    return vision_out, policy_out
-  return run_policy
-
-
-def compile_modeld(cam_w, cam_h, prepare_only, pkl_path):
-  from tinygrad.nn.onnx import OnnxRunner
-  from openpilot.selfdrive.modeld.constants import ModelConstants
-
-  _, _, _, yuv_size = get_nv12_info(cam_w, cam_h)
-  print(f"Compiling combined policy JIT for {cam_w}x{cam_h}...")
-
-  vision_runner = OnnxRunner(MODELS_DIR / 'driving_vision.onnx')
-  policy_runner = OnnxRunner(MODELS_DIR / 'driving_policy.onnx')
-
-  with open(MODELS_DIR / 'driving_vision_metadata.pkl', 'rb') as f:
-    vision_metadata = pickle.load(f)
-    vision_features_slice = vision_metadata['output_slices']['hidden_state']
-    vision_input_shapes = vision_metadata['input_shapes']
-  with open(MODELS_DIR / 'driving_policy_metadata.pkl', 'rb') as f:
-    policy_input_shapes = pickle.load(f)['input_shapes']
-
-  frame_skip = ModelConstants.MODEL_RUN_FREQ // ModelConstants.MODEL_CONTEXT_FREQ
-
-  _run = make_run_policy(vision_runner, policy_runner,
-                         cam_w, cam_h, vision_features_slice, frame_skip, prepare_only)
-  run_policy_jit = TinyJit(_run, prune=True)
-
-  N_RUNS = 3
-  SEED = 42
-
-  def random_inputs_run_fn(fn, seed, test_val=None, test_buffers=None, expect_match=True):
-    input_queues, npy = make_input_queues(vision_input_shapes, policy_input_shapes, frame_skip)
-    np.random.seed(seed)
-
-    for i in range(N_RUNS):
-      frame = Tensor(np.random.randint(0, 256, yuv_size, dtype=np.uint8)).realize()
-      big_frame = Tensor(np.random.randint(0, 256, yuv_size, dtype=np.uint8)).realize()
-      for v in npy.values():
-        v[:] = np.random.randn(*v.shape).astype(v.dtype)
-      Device.default.synchronize()
-      st = time.perf_counter()
-      outs = fn(**input_queues, frame=frame, big_frame=big_frame)
-      mt = time.perf_counter()
-      for o in outs:
-        # .realize() not needed once jitted, but needed for unjitted fn
-        o.realize()
-      Device.default.synchronize()
-      et = time.perf_counter()
-      print(f"  [{i+1}/{N_RUNS}] enqueue {(mt-st)*1e3:6.2f} ms -- total {(et-st)*1e3:6.2f} ms")
-
-    val = [np.copy(v.numpy()) for v in outs]
-    buffers = [np.copy(v.numpy().copy()) for v in input_queues.values()]
-
-    if test_val is not None:
-      match = all(np.array_equal(a, b) for a, b in zip(val, test_val, strict=True))
-      assert match == expect_match, f"outputs {'differ from' if expect_match else 'match'} baseline (seed={seed})"
-    if test_buffers is not None:
-      match = all(np.array_equal(a, b) for a, b in zip(buffers, test_buffers, strict=True))
-      assert match == expect_match, f"buffers {'differ from' if expect_match else 'match'} baseline (seed={seed})"
-    return fn, val, buffers
-
-  print('run unjitted')
-  _, test_val, test_buffers = random_inputs_run_fn(_run, seed=SEED)
-  print('capture + replay')
-  run_policy_jit, _, _ = random_inputs_run_fn(run_policy_jit, SEED, test_val, test_buffers)
-
-  print('pickle round trip')
-  with open(pkl_path, "wb") as f:
-    pickle.dump(run_policy_jit, f)
-    print(f"  Saved to {pkl_path}")
-  with open(pkl_path, "rb") as f:
-    run_policy_jit = pickle.load(f)
-  random_inputs_run_fn(run_policy_jit, SEED, test_val, test_buffers, expect_match=True)
-  random_inputs_run_fn(run_policy_jit, SEED+1, test_val, test_buffers, expect_match=False)
-
-
-def compile_dm_warp(cam_w, cam_h, pkl_path):
-  dm_w, dm_h = DM_INPUT_SIZE
-  _, _, _, yuv_size = get_nv12_info(cam_w, cam_h)
-
-  print(f"Compiling DM warp for {cam_w}x{cam_h}...")
-
-  warp_dm = make_warp_dm(cam_w, cam_h, dm_w, dm_h)
-  warp_dm_jit = TinyJit(warp_dm, prune=True)
-
-  for i in range(10):
-    inputs = [Tensor(np.random.randint(0, 256, yuv_size, dtype=np.uint8)).realize(),
-              Tensor(Tensor.randn(3, 3).mul(8).realize().numpy(), device='NPY')]
-    Device.default.synchronize()
-    st = time.perf_counter()
-    warp_dm_jit(*inputs)
-    mt = time.perf_counter()
-    Device.default.synchronize()
-    et = time.perf_counter()
-    print(f"  [{i+1}/10] enqueue {(mt-st)*1e3:6.2f} ms -- total {(et-st)*1e3:6.2f} ms")
-
-  with open(pkl_path, "wb") as f:
-    pickle.dump(warp_dm_jit, f)
-  print(f"  Saved to {pkl_path}")
-
-
-def run_and_save_pickle():
-  for cfg in MODELD_CONFIGS:
-    compile_modeld(cfg.cam_w, cfg.cam_h, cfg.prepare_only, cfg.pkl_path)
-  for cfg in DM_WARP_CONFIGS:
-    compile_dm_warp(cfg.cam_w, cfg.cam_h, cfg.pkl_path)
-
-
-if __name__ == "__main__":
-  run_and_save_pickle()

selfdrive/modeld/compile_warp.py

@@ -0,0 +1,198 @@
+#!/usr/bin/env python3
+import time
+import pickle
+import numpy as np
+from pathlib import Path
+from tinygrad.tensor import Tensor
+from tinygrad.helpers import Context
+from tinygrad.device import Device
+from tinygrad.engine.jit import TinyJit
+
+from openpilot.system.camerad.cameras.nv12_info import get_nv12_info
+from openpilot.common.transformations.model import MEDMODEL_INPUT_SIZE, DM_INPUT_SIZE
+from openpilot.common.transformations.camera import _ar_ox_fisheye, _os_fisheye
+
+MODELS_DIR = Path(__file__).parent / 'models'
+
+CAMERA_CONFIGS = [
+  (_ar_ox_fisheye.width, _ar_ox_fisheye.height),  # tici: 1928x1208
+  (_os_fisheye.width, _os_fisheye.height),        # mici: 1344x760
+]
+
+UV_SCALE_MATRIX = np.array([[0.5, 0, 0], [0, 0.5, 0], [0, 0, 1]], dtype=np.float32)
+UV_SCALE_MATRIX_INV = np.linalg.inv(UV_SCALE_MATRIX)
+
+IMG_BUFFER_SHAPE = (30, MEDMODEL_INPUT_SIZE[1] // 2, MEDMODEL_INPUT_SIZE[0] // 2)
+
+
+def warp_pkl_path(w, h):
+  return MODELS_DIR / f'warp_{w}x{h}_tinygrad.pkl'
+
+
+def dm_warp_pkl_path(w, h):
+  return MODELS_DIR / f'dm_warp_{w}x{h}_tinygrad.pkl'
+
+
+def warp_perspective_tinygrad(src_flat, M_inv, dst_shape, src_shape, stride_pad):
+  w_dst, h_dst = dst_shape
+  h_src, w_src = src_shape
+
+  x = Tensor.arange(w_dst).reshape(1, w_dst).expand(h_dst, w_dst).reshape(-1)
+  y = Tensor.arange(h_dst).reshape(h_dst, 1).expand(h_dst, w_dst).reshape(-1)
+
+  # inline 3x3 matmul as elementwise to avoid reduce op (enables fusion with gather)
+  src_x = M_inv[0, 0] * x + M_inv[0, 1] * y + M_inv[0, 2]
+  src_y = M_inv[1, 0] * x + M_inv[1, 1] * y + M_inv[1, 2]
+  src_w = M_inv[2, 0] * x + M_inv[2, 1] * y + M_inv[2, 2]
+
+  src_x = src_x / src_w
+  src_y = src_y / src_w
+
+  x_nn_clipped = Tensor.round(src_x).clip(0, w_src - 1).cast('int')
+  y_nn_clipped = Tensor.round(src_y).clip(0, h_src - 1).cast('int')
+  idx = y_nn_clipped * (w_src + stride_pad) + x_nn_clipped
+
+  return src_flat[idx]
+
+
+def frames_to_tensor(frames, model_w, model_h):
+  H = (frames.shape[0] * 2) // 3
+  W = frames.shape[1]
+  in_img1 = Tensor.cat(frames[0:H:2, 0::2],
+                       frames[1:H:2, 0::2],
+                       frames[0:H:2, 1::2],
+                       frames[1:H:2, 1::2],
+                       frames[H:H+H//4].reshape((H//2, W//2)),
+                       frames[H+H//4:H+H//2].reshape((H//2, W//2)), dim=0).reshape((6, H//2, W//2))
+  return in_img1
+
+
+def make_frame_prepare(cam_w, cam_h, model_w, model_h):
+  stride, y_height, uv_height, _ = get_nv12_info(cam_w, cam_h)
+  uv_offset = stride * y_height
+  stride_pad = stride - cam_w
+
+  def frame_prepare_tinygrad(input_frame, M_inv):
+    # UV_SCALE @ M_inv @ UV_SCALE_INV simplifies to elementwise scaling
+    M_inv_uv = M_inv * Tensor([[1.0, 1.0, 0.5], [1.0, 1.0, 0.5], [2.0, 2.0, 1.0]])
+    # deinterleave NV12 UV plane (UVUV... -> separate U, V)
+    uv = input_frame[uv_offset:uv_offset + uv_height * stride].reshape(uv_height, stride)
+    with Context(SPLIT_REDUCEOP=0):
+      y = warp_perspective_tinygrad(input_frame[:cam_h*stride],
+                                    M_inv, (model_w, model_h),
+                                    (cam_h, cam_w), stride_pad).realize()
+      u = warp_perspective_tinygrad(uv[:cam_h//2, :cam_w:2].flatten(),
+                                    M_inv_uv, (model_w//2, model_h//2),
+                                    (cam_h//2, cam_w//2), 0).realize()
+      v = warp_perspective_tinygrad(uv[:cam_h//2, 1:cam_w:2].flatten(),
+                                    M_inv_uv, (model_w//2, model_h//2),
+                                    (cam_h//2, cam_w//2), 0).realize()
+    yuv = y.cat(u).cat(v).reshape((model_h * 3 // 2, model_w))
+    tensor = frames_to_tensor(yuv, model_w, model_h)
+    return tensor
+  return frame_prepare_tinygrad
+
+
+def make_update_img_input(frame_prepare, model_w, model_h):
+  def update_img_input_tinygrad(frame_buffer, frame, M_inv):
+    M_inv = M_inv.to(Device.DEFAULT)
+    new_img = frame_prepare(frame, M_inv)
+    frame_buffer.assign(frame_buffer[6:].cat(new_img, dim=0).contiguous())
+    return Tensor.cat(frame_buffer[:6], frame_buffer[-6:], dim=0).contiguous().reshape(1, 12, model_h//2, model_w//2)
+  return update_img_input_tinygrad
+
+
+def make_update_both_imgs(frame_prepare, model_w, model_h):
+  update_img = make_update_img_input(frame_prepare, model_w, model_h)
+
+  def update_both_imgs_tinygrad(calib_img_buffer, new_img, M_inv,
+                                calib_big_img_buffer, new_big_img, M_inv_big):
+    calib_img_pair = update_img(calib_img_buffer, new_img, M_inv)
+    calib_big_img_pair = update_img(calib_big_img_buffer, new_big_img, M_inv_big)
+    return calib_img_pair, calib_big_img_pair
+  return update_both_imgs_tinygrad
+
+
+def make_warp_dm(cam_w, cam_h, dm_w, dm_h):
+  stride, y_height, _, _ = get_nv12_info(cam_w, cam_h)
+  stride_pad = stride - cam_w
+
+  def warp_dm(input_frame, M_inv):
+    M_inv = M_inv.to(Device.DEFAULT)
+    result = warp_perspective_tinygrad(input_frame[:cam_h*stride], M_inv, (dm_w, dm_h), (cam_h, cam_w), stride_pad).reshape(-1, dm_h * dm_w)
+    return result
+  return warp_dm
+
+
+def compile_modeld_warp(cam_w, cam_h):
+  model_w, model_h = MEDMODEL_INPUT_SIZE
+  _, _, _, yuv_size = get_nv12_info(cam_w, cam_h)
+
+  print(f"Compiling modeld warp for {cam_w}x{cam_h}...")
+
+  frame_prepare = make_frame_prepare(cam_w, cam_h, model_w, model_h)
+  update_both_imgs = make_update_both_imgs(frame_prepare, model_w, model_h)
+  update_img_jit = TinyJit(update_both_imgs, prune=True)
+
+  full_buffer = Tensor.zeros(IMG_BUFFER_SHAPE, dtype='uint8').contiguous().realize()
+  big_full_buffer = Tensor.zeros(IMG_BUFFER_SHAPE, dtype='uint8').contiguous().realize()
+  for i in range(10):
+    img_inputs = [full_buffer,
+                  Tensor(np.random.randint(0, 256, yuv_size, dtype=np.uint8)).realize(),
+                  Tensor(Tensor.randn(3, 3).mul(8).realize().numpy(), device='NPY')]
+    big_img_inputs = [big_full_buffer,
+                      Tensor(np.random.randint(0, 256, yuv_size, dtype=np.uint8)).realize(),
+                      Tensor(Tensor.randn(3, 3).mul(8).realize().numpy(), device='NPY')]
+    inputs = img_inputs + big_img_inputs
+    Device.default.synchronize()
+
+    st = time.perf_counter()
+    _ = update_img_jit(*inputs)
+    mt = time.perf_counter()
+    Device.default.synchronize()
+    et = time.perf_counter()
+    print(f"  [{i+1}/10] enqueue {(mt-st)*1e3:6.2f} ms -- total {(et-st)*1e3:6.2f} ms")
+
+  pkl_path = warp_pkl_path(cam_w, cam_h)
+  with open(pkl_path, "wb") as f:
+    pickle.dump(update_img_jit, f)
+  print(f"  Saved to {pkl_path}")
+
+  jit = pickle.load(open(pkl_path, "rb"))
+  jit(*inputs)
+
+
+def compile_dm_warp(cam_w, cam_h):
+  dm_w, dm_h = DM_INPUT_SIZE
+  _, _, _, yuv_size = get_nv12_info(cam_w, cam_h)
+
+  print(f"Compiling DM warp for {cam_w}x{cam_h}...")
+
+  warp_dm = make_warp_dm(cam_w, cam_h, dm_w, dm_h)
+  warp_dm_jit = TinyJit(warp_dm, prune=True)
+
+  for i in range(10):
+    inputs = [Tensor(np.random.randint(0, 256, yuv_size, dtype=np.uint8)).realize(),
+              Tensor(Tensor.randn(3, 3).mul(8).realize().numpy(), device='NPY')]
+    Device.default.synchronize()
+    st = time.perf_counter()
+    warp_dm_jit(*inputs)
+    mt = time.perf_counter()
+    Device.default.synchronize()
+    et = time.perf_counter()
+    print(f"  [{i+1}/10] enqueue {(mt-st)*1e3:6.2f} ms -- total {(et-st)*1e3:6.2f} ms")
+
+  pkl_path = dm_warp_pkl_path(cam_w, cam_h)
+  with open(pkl_path, "wb") as f:
+    pickle.dump(warp_dm_jit, f)
+  print(f"  Saved to {pkl_path}")
+
+
+def run_and_save_pickle():
+  for cam_w, cam_h in CAMERA_CONFIGS:
+    compile_modeld_warp(cam_w, cam_h)
+    compile_dm_warp(cam_w, cam_h)
+
+
+if __name__ == "__main__":
+  run_and_save_pickle()

selfdrive/modeld/modeld.py

@@ -26,7 +26,6 @@ from openpilot.common.transformations.model import get_warp_matrix
 from openpilot.selfdrive.controls.lib.desire_helper import DesireHelper
 from openpilot.selfdrive.controls.lib.drive_helpers import get_accel_from_plan, smooth_value, get_curvature_from_plan
 from openpilot.selfdrive.modeld.parse_model_outputs import Parser
-from openpilot.selfdrive.modeld.compile_modeld import CompileConfig, make_input_queues
 from openpilot.selfdrive.modeld.fill_model_msg import fill_model_msg, fill_pose_msg, PublishState
 from openpilot.common.file_chunker import read_file_chunked
 from openpilot.selfdrive.modeld.constants import ModelConstants, Plan
@@ -35,13 +34,17 @@ from openpilot.selfdrive.modeld.constants import ModelConstants, Plan
 PROCESS_NAME = "selfdrive.modeld.modeld"
 SEND_RAW_PRED = os.getenv('SEND_RAW_PRED')
 
+VISION_PKL_PATH = MODELS_DIR / 'driving_vision_tinygrad.pkl'
 VISION_METADATA_PATH = MODELS_DIR / 'driving_vision_metadata.pkl'
+POLICY_PKL_PATH = MODELS_DIR / 'driving_policy_tinygrad.pkl'
 POLICY_METADATA_PATH = MODELS_DIR / 'driving_policy_metadata.pkl'
 
 LAT_SMOOTH_SECONDS = 0.0
 LONG_SMOOTH_SECONDS = 0.3
 MIN_LAT_CONTROL_SPEED = 0.3
 
+IMG_QUEUE_SHAPE = (6*(ModelConstants.MODEL_RUN_FREQ//ModelConstants.MODEL_CONTEXT_FREQ + 1), 128, 256)
+assert IMG_QUEUE_SHAPE[0] == 30
 
 
 def get_action_from_model(model_output: dict[str, np.ndarray], prev_action: log.ModelDataV2.Action,
@@ -76,36 +79,106 @@ class FrameMeta:
     if vipc is not None:
       self.frame_id, self.timestamp_sof, self.timestamp_eof = vipc.frame_id, vipc.timestamp_sof, vipc.timestamp_eof
 
+class InputQueues:
+  def __init__ (self, model_fps, env_fps, n_frames_input):
+    assert env_fps % model_fps == 0
+    assert env_fps >= model_fps
+    self.model_fps = model_fps
+    self.env_fps = env_fps
+    self.n_frames_input = n_frames_input
+
+    self.dtypes = {}
+    self.shapes = {}
+    self.q = {}
+
+  def update_dtypes_and_shapes(self, input_dtypes, input_shapes) -> None:
+    self.dtypes.update(input_dtypes)
+    if self.env_fps == self.model_fps:
+      self.shapes.update(input_shapes)
+    else:
+      for k in input_shapes:
+        shape = list(input_shapes[k])
+        if 'img' in k:
+          n_channels = shape[1] // self.n_frames_input
+          shape[1] = (self.env_fps // self.model_fps + (self.n_frames_input - 1)) * n_channels
+        else:
+          shape[1] = (self.env_fps // self.model_fps) * shape[1]
+        self.shapes[k] = tuple(shape)
+
+  def reset(self) -> None:
+    self.q = {k: np.zeros(self.shapes[k], dtype=self.dtypes[k]) for k in self.dtypes.keys()}
+
+  def enqueue(self, inputs:dict[str, np.ndarray]) -> None:
+    for k in inputs.keys():
+      if inputs[k].dtype != self.dtypes[k]:
+        raise ValueError(f'supplied input <{k}({inputs[k].dtype})> has wrong dtype, expected {self.dtypes[k]}')
+      input_shape = list(self.shapes[k])
+      input_shape[1] = -1
+      single_input = inputs[k].reshape(tuple(input_shape))
+      sz = single_input.shape[1]
+      self.q[k][:,:-sz] = self.q[k][:,sz:]
+      self.q[k][:,-sz:] = single_input
+
+  def get(self, *names) -> dict[str, np.ndarray]:
+    if self.env_fps == self.model_fps:
+      return {k: self.q[k] for k in names}
+    else:
+      out = {}
+      for k in names:
+        shape = self.shapes[k]
+        if 'img' in k:
+          n_channels = shape[1] // (self.env_fps // self.model_fps + (self.n_frames_input - 1))
+          out[k] = np.concatenate([self.q[k][:, s:s+n_channels] for s in np.linspace(0, shape[1] - n_channels, self.n_frames_input, dtype=int)], axis=1)
+        elif 'pulse' in k:
+          # any pulse within interval counts
+          out[k] = self.q[k].reshape((shape[0], shape[1] * self.model_fps // self.env_fps, self.env_fps // self.model_fps, -1)).max(axis=2)
+        else:
+          idxs = np.arange(-1, -shape[1], -self.env_fps // self.model_fps)[::-1]
+          out[k] = self.q[k][:, idxs]
+      return out
 
 class ModelState:
+  inputs: dict[str, np.ndarray]
+  output: np.ndarray
   prev_desire: np.ndarray  # for tracking the rising edge of the pulse
 
-  def __init__(self, cam_w: int, cam_h: int):
+  def __init__(self):
     with open(VISION_METADATA_PATH, 'rb') as f:
       vision_metadata = pickle.load(f)
       self.vision_input_shapes =  vision_metadata['input_shapes']
       self.vision_input_names = list(self.vision_input_shapes.keys())
       self.vision_output_slices = vision_metadata['output_slices']
+      vision_output_size = vision_metadata['output_shapes']['outputs'][1]
 
     with open(POLICY_METADATA_PATH, 'rb') as f:
       policy_metadata = pickle.load(f)
       self.policy_input_shapes =  policy_metadata['input_shapes']
       self.policy_output_slices = policy_metadata['output_slices']
+      policy_output_size = policy_metadata['output_shapes']['outputs'][1]
 
     self.prev_desire = np.zeros(ModelConstants.DESIRE_LEN, dtype=np.float32)
 
-    self.frame_skip = ModelConstants.MODEL_RUN_FREQ // ModelConstants.MODEL_CONTEXT_FREQ
-    self.input_queues, self.npy = make_input_queues(self.vision_input_shapes, self.policy_input_shapes, self.frame_skip)
+    # policy inputs
+    self.numpy_inputs = {k: np.zeros(self.policy_input_shapes[k], dtype=np.float32) for k in self.policy_input_shapes}
+    self.full_input_queues = InputQueues(ModelConstants.MODEL_CONTEXT_FREQ, ModelConstants.MODEL_RUN_FREQ, ModelConstants.N_FRAMES)
+    for k in ['desire_pulse', 'features_buffer']:
+      self.full_input_queues.update_dtypes_and_shapes({k: self.numpy_inputs[k].dtype}, {k: self.numpy_inputs[k].shape})
+    self.full_input_queues.reset()
+
+    self.img_queues = {'img': Tensor.zeros(IMG_QUEUE_SHAPE, dtype='uint8').contiguous().realize(),
+                       'big_img': Tensor.zeros(IMG_QUEUE_SHAPE, dtype='uint8').contiguous().realize()}
     self.full_frames : dict[str, Tensor] = {}
     self._blob_cache : dict[int, Tensor] = {}
+    self.transforms_np = {k: np.zeros((3,3), dtype=np.float32) for k in self.img_queues}
+    self.transforms = {k: Tensor(v, device='NPY').realize() for k, v in self.transforms_np.items()}
+    self.vision_output = np.zeros(vision_output_size, dtype=np.float32)
+    self.policy_inputs = {k: Tensor(v, device='NPY').realize() for k,v in self.numpy_inputs.items()}
+    self.policy_output = np.zeros(policy_output_size, dtype=np.float32)
     self.parser = Parser()
-    self.frame_buf_params = {k: get_nv12_info(cam_w, cam_h) for k in ('img', 'big_img')}
-    self.run_policy = pickle.loads(read_file_chunked(CompileConfig(cam_w, cam_h, prefix='driving_', prepare_only=False).pkl_path))
-    self.warp_enqueue = pickle.loads(read_file_chunked(CompileConfig(cam_w, cam_h, prefix='driving_', prepare_only=True).pkl_path))
-    self.warp_enqueue(
-      **self.input_queues,
-      frame=Tensor.zeros(self.frame_buf_params['img'][3], dtype='uint8').contiguous().realize(),
-      big_frame=Tensor.zeros(self.frame_buf_params['big_img'][3], dtype='uint8').contiguous().realize())
+    self.frame_buf_params : dict[str, tuple[int, int, int, int]] = {}
+    self.update_imgs = None
+    self.vision_run = pickle.loads(read_file_chunked(str(VISION_PKL_PATH)))
+    self.policy_run = pickle.loads(read_file_chunked(str(POLICY_PKL_PATH)))
 
   def slice_outputs(self, model_outputs: np.ndarray, output_slices: dict[str, slice]) -> dict[str, np.ndarray]:
     parsed_model_outputs = {k: model_outputs[np.newaxis, v] for k,v in output_slices.items()}
@@ -113,6 +186,18 @@ class ModelState:
 
   def run(self, bufs: dict[str, VisionBuf], transforms: dict[str, 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_pulse'][0] = 0
+    new_desire = np.where(inputs['desire_pulse'] - self.prev_desire > .99, inputs['desire_pulse'], 0)
+    self.prev_desire[:] = inputs['desire_pulse']
+    if self.update_imgs is None:
+      for key in bufs.keys():
+        w, h = bufs[key].width, bufs[key].height
+        self.frame_buf_params[key] = get_nv12_info(w, h)
+      warp_path = MODELS_DIR / f'warp_{w}x{h}_tinygrad.pkl'
+      with open(warp_path, "rb") as f:
+        self.update_imgs = pickle.load(f)
+
     for key in bufs.keys():
       ptr = bufs[key].data.ctypes.data
       yuv_size = self.frame_buf_params[key][3]
@@ -121,31 +206,30 @@ class ModelState:
       if cache_key not in self._blob_cache:
         self._blob_cache[cache_key] = Tensor.from_blob(ptr, (yuv_size,), dtype='uint8')
       self.full_frames[key] = self._blob_cache[cache_key]
+    for key in bufs.keys():
+      self.transforms_np[key][:,:] = transforms[key][:,:]
 
-    # Model decides when action is completed, so desire input is just a pulse triggered on rising edge
-    inputs['desire_pulse'][0] = 0
-    self.npy['desire'][:] = np.where(inputs['desire_pulse'] - self.prev_desire > .99, inputs['desire_pulse'], 0)
-    self.prev_desire[:] = inputs['desire_pulse']
-    self.npy['traffic_convention'][:] = inputs['traffic_convention']
-    self.npy['tfm'][:,:] = transforms['img'][:,:]
-    self.npy['big_tfm'][:,:] = transforms['big_img'][:,:]
+    out = self.update_imgs(self.img_queues['img'], self.full_frames['img'], self.transforms['img'],
+                           self.img_queues['big_img'], self.full_frames['big_img'], self.transforms['big_img'])
+    vision_inputs = {'img': out[0], 'big_img': out[1]}
 
     if prepare_only:
-      self.warp_enqueue(**self.input_queues, frame=self.full_frames['img'], big_frame=self.full_frames['big_img'])
       return None
 
-    vision_output, policy_output = self.run_policy(
-      **self.input_queues, frame=self.full_frames['img'], big_frame=self.full_frames['big_img']
-    )
+    self.vision_output = self.vision_run(**vision_inputs).contiguous().realize().uop.base.buffer.numpy().flatten()
+    vision_outputs_dict = self.parser.parse_vision_outputs(self.slice_outputs(self.vision_output, self.vision_output_slices))
 
-    vision_output = vision_output.numpy().flatten()
-    policy_output = policy_output.numpy().flatten()
-    vision_outputs_dict = self.parser.parse_vision_outputs(self.slice_outputs(vision_output, self.vision_output_slices))
-    policy_outputs_dict = self.parser.parse_policy_outputs(self.slice_outputs(policy_output, self.policy_output_slices))
+    self.full_input_queues.enqueue({'features_buffer': vision_outputs_dict['hidden_state'], 'desire_pulse': new_desire})
+    for k in ['desire_pulse', 'features_buffer']:
+      self.numpy_inputs[k][:] = self.full_input_queues.get(k)[k]
+    self.numpy_inputs['traffic_convention'][:] = inputs['traffic_convention']
+
+    self.policy_output = self.policy_run(**self.policy_inputs).contiguous().realize().uop.base.buffer.numpy().flatten()
+    policy_outputs_dict = self.parser.parse_policy_outputs(self.slice_outputs(self.policy_output, self.policy_output_slices))
     combined_outputs_dict = {**vision_outputs_dict, **policy_outputs_dict}
-
     if SEND_RAW_PRED:
-      combined_outputs_dict['raw_pred'] = np.concatenate([vision_output.copy(), policy_output.copy()])
+      combined_outputs_dict['raw_pred'] = np.concatenate([self.vision_output.copy(), self.policy_output.copy()])
+
     return combined_outputs_dict
 
 
@@ -157,6 +241,11 @@ def main(demo=False):
     # also need to move the aux USB interrupts for good timings
     config_realtime_process(7, 54)
 
+  st = time.monotonic()
+  cloudlog.warning("loading model")
+  model = ModelState()
+  cloudlog.warning(f"models loaded in {time.monotonic() - st:.1f}s, modeld starting")
+
   # visionipc clients
   while True:
     available_streams = VisionIpcClient.available_streams("camerad", block=False)
@@ -180,11 +269,6 @@ def main(demo=False):
   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})")
 
-  st = time.monotonic()
-  cloudlog.warning("loading model")
-  model = ModelState(vipc_client_main.width, vipc_client_main.height)
-  cloudlog.warning(f"models loaded in {time.monotonic() - st:.1f}s, modeld starting")
-
   # messaging
   pm = PubMaster(["modelV2", "drivingModelData", "cameraOdometry"])
   sm = SubMaster(["deviceState", "carState", "roadCameraState", "liveCalibration", "driverMonitoringState", "carControl", "liveDelay"])

selfdrive/modeld/tinygrad_helpers.py

@@ -2,7 +2,7 @@ import json
 import os
 from pathlib import Path
 
-MODELS_DIR = Path(__file__).resolve().parent / 'models'
+MODELS_DIR = Path(__file__).parent / 'models'
 COMPILED_FLAGS_PATH = MODELS_DIR / 'tg_compiled_flags.json'