BMRLNAP (#38681)

openpilot/selfdrive/modeld/SConscript

@@ -24,7 +24,9 @@ tinygrad_files = ["#"+x for x in glob.glob(env.Dir("#tinygrad_repo").relpath + "
                   if 'pycache' not in x and os.path.isfile(os.path.join(tinygrad_root, x))]
 
 def estimate_pickle_max_size(onnx_size):
-  return 1.2 * onnx_size  + 10 * 1024 * 1024  # 20% + 10MB is plenty
+  # QCOM programs for models with spatial recurrent features can approach 2x
+  # the ONNX size. Overestimating only adds an empty trailing chunk.
+  return 2.0 * onnx_size + 10 * 1024 * 1024
 
 if arch == 'comma_arm64':
   tg_backend = 'QCOM'
@@ -45,7 +47,7 @@ tg_devices = { # which device to put jit inputs to at runtime
 
 CHESTNUT = chestnut_present()
 if CHESTNUT:
-  chestnut_tg_flags = f'DEBUG=2 DEV=USB+AMD:LLVM WARP_DEV={tg_backend} FLOAT16=1 JIT_BATCH_SIZE=0 GMMU=0 TC_OPT=2'
+  chestnut_tg_flags = f'DEBUG=1 DEV=USB+AMD:LLVM WARP_DEV={tg_backend} FLOAT16=1 JIT_BATCH_SIZE=0 GMMU=0 TC_OPT=2'
   # the USB+AMD GPU takes an exclusive flock; serialize all targets that touch it
   chestnut_lock = File("models/.chestnut.lock").abspath
 

openpilot/selfdrive/modeld/compile_modeld.py

@@ -139,16 +139,18 @@ def get_policy_npy_shapes(input_shapes):
   dp = input_shapes['desire_pulse']  # (1, 25, 8)
   tc = input_shapes['traffic_convention']  # (1, 2)
   at = input_shapes['action_t']  # (1, 2)
-  fb = input_shapes['features_buffer']  # (1, 24, 512)
+  fb = input_shapes['features_buffer']  # (1, T-1, ...) e.g. (1, 24, 32, 512) with spatial features
+  feat_dim = math.prod(fb[2:])
   # TODO prev_feat shouldn't exist and be handled inside the JIT, but corrupt on QCOM for now
-  shapes = {'desire': (dp[2],), 'traffic_convention': tuple(tc), 'action_t': tuple(at), 'prev_feat': (fb[0], fb[2])}
+  shapes = {'desire': (dp[2],), 'traffic_convention': tuple(tc), 'action_t': tuple(at), 'prev_feat': (fb[0], feat_dim)}
   return shapes, [math.prod(s) for s in shapes.values()]
 
 
 def make_input_queues(input_shapes, frame_skip, device):
   input_queues, npy = make_warp_input_queues(input_shapes, frame_skip, device)
 
-  fb = input_shapes['features_buffer']  # (1, 24, 512), past features only; the model appends the current frame's feature
+  fb = input_shapes['features_buffer']  # (1, T-1, ...), past features only; the model appends the current frame's feature
+  feat_dim = math.prod(fb[2:])
   dp = input_shapes['desire_pulse']  # (1, 25, 8)
 
   shapes, sizes = get_policy_npy_shapes(input_shapes)
@@ -156,7 +158,7 @@ def make_input_queues(input_shapes, frame_skip, device):
   # views into the packed inputs, to be refilled at runtime
   npy.update({k: v.reshape(s) for (k, s), v in zip(shapes.items(), np.split(packed_npy_inputs, np.cumsum(sizes[:-1])), strict=True)})
   input_queues.update({
-    'feat_q': Tensor(np.zeros((frame_skip * fb[1], fb[0], fb[2]), dtype=np.float32), device=device).contiguous().realize(),
+    'feat_q': Tensor(np.zeros((frame_skip * fb[1], fb[0], feat_dim), dtype=np.float32), device=device).contiguous().realize(),
     'desire_q': Tensor(np.zeros((frame_skip * dp[1], dp[0], dp[2]), dtype=np.float32), device=device).contiguous().realize(),
     'packed_npy_inputs': Tensor(packed_npy_inputs, device='NPY').realize(),
   })
@@ -211,7 +213,7 @@ def make_run_policy(model_runner, model_metadata, frame_skip):
     inputs = {
       'img': img,
       'big_img': big_img,
-      'features_buffer': feat_buf,
+      'features_buffer': feat_buf.reshape(model_metadata['input_shapes']['features_buffer']),
       'desire_pulse': desire_buf,
       'traffic_convention': traffic_convention,
       'action_t': action_t,

openpilot/selfdrive/modeld/models/big_driving_supercombo.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a501760a9d1d5fef0eab2b8c5d122d06124fc26dc8e0782e0aa94b82a208f0ff
-size 1757355221
+oid sha256:a086d5249fc308bb73993d1e64630c669d4c7df5bde85f42ad61902543648525
+size 765953504