IQ.Pilot/selfdrive/modeld/SConscript

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 ['driving_vision', 'driving_policy', '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 verify_prebuilt, outputs_match
    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)

# Compile BIG model if USB GPU is available
if "USBGPU" in os.environ:
  import subprocess
  # because tg doesn't support multi-process
  devs = subprocess.check_output('python3 -c "from tinygrad import Device; print(list(Device.get_available_devices()))"', shell=True, cwd=env.Dir('#').abspath)
  if b"AMD" in devs:
    print("USB GPU detected... building")
    flags = "DEV=AMD AMD_IFACE=USB AMD_LLVM=1 NOLOCALS=0 IMAGE=0"
    bp = tg_compile(flags, "big_driving_policy")
    bv = tg_compile(flags, "big_driving_vision")
    lenv.SideEffect('lock', [bp, bv])  # tg doesn't support multi-process so build serially
  else:
    print("USB GPU not detected... skipping")