Tinygrad runner (#34171)

* squash

* bump tg

* bump tg

* debump tinygrad

* bump tinygrad

* bump tg

* Skip init iteration

* fixes

* cleanups

* skip first test sample

* typos

* linter unhappy

* update cpu usage

* OPENCL just zeros for now

* imports

* Try printing

* Runs again, but slower

* unused import

* Allow more buffer with tg and all on gpu

* bump tinygrad

* seems ok

* stricter timings for driving looser for dm

* try llvm

* check nvidia

* More timeout for now

* make test pass

* Revert "try llvm"

This reverts commit ef136e478320101fea262bae3579e558da991902.

* small fixes

* whitespace

* revert test timeout

* No model runners

* Always CPU always fast

* No onnx runtime GPU

* more cores

* cleanup

* Is this faster

* Is this faster

* at least runs

* FP32 is faster than 16

* fix deps

* whitespace

* comment

---------

Co-authored-by: Adeeb Shihadeh <adeebshihadeh@gmail.com>

selfdrive/modeld/SConscript

@@ -13,20 +13,6 @@ common_src = [
   "transforms/transform.cc",
 ]
 
-thneed_src_common = [
-  "thneed/thneed_common.cc",
-  "thneed/serialize.cc",
-]
-
-thneed_src_qcom = thneed_src_common + ["thneed/thneed_qcom2.cc"]
-thneed_src_pc = thneed_src_common + ["thneed/thneed_pc.cc"]
-thneed_src = thneed_src_qcom if arch == "larch64" else thneed_src_pc
-
-# SNPE except on Mac and ARM Linux
-snpe_lib = []
-if arch != "Darwin" and arch != "aarch64":
-  common_src += ['runners/snpemodel.cc']
-  snpe_lib += ['SNPE']
 
 # OpenCL is a framework on Mac
 if arch == "Darwin":
@@ -45,11 +31,7 @@ snpe_rpath_pc = f"{Dir('#').abspath}/third_party/snpe/x86_64-linux-clang"
 snpe_rpath = lenvCython['RPATH'] + [snpe_rpath_qcom if arch == "larch64" else snpe_rpath_pc]
 
 cython_libs = envCython["LIBS"] + libs
-snpemodel_lib = lenv.Library('snpemodel', ['runners/snpemodel.cc'])
 commonmodel_lib = lenv.Library('commonmodel', common_src)
-
-lenvCython.Program('runners/runmodel_pyx.so', 'runners/runmodel_pyx.pyx', LIBS=cython_libs, FRAMEWORKS=frameworks)
-lenvCython.Program('runners/snpemodel_pyx.so', 'runners/snpemodel_pyx.pyx', LIBS=[snpemodel_lib, snpe_lib, *cython_libs], FRAMEWORKS=frameworks, RPATH=snpe_rpath)
 lenvCython.Program('models/commonmodel_pyx.so', 'models/commonmodel_pyx.pyx', LIBS=[commonmodel_lib, *cython_libs], FRAMEWORKS=frameworks)
 
 tinygrad_files = ["#"+x for x in glob.glob(env.Dir("#tinygrad_repo").relpath + "/**", recursive=True, root_dir=env.Dir("#").abspath)]
@@ -59,20 +41,17 @@ fn = File("models/supercombo").abspath
 cmd = f'python3 {Dir("#selfdrive/modeld").abspath}/get_model_metadata.py {fn}.onnx'
 lenv.Command(fn + "_metadata.pkl", [fn + ".onnx"] + tinygrad_files, cmd)
 
-# Build thneed model
-if arch == "larch64" or GetOption('pc_thneed'):
-  tinygrad_opts = []
-  if not GetOption('pc_thneed'):
-    # use FLOAT16 on device for speed + don't cache the CL kernels for space
-    tinygrad_opts += ["FLOAT16=1", "PYOPENCL_NO_CACHE=1"]
-  cmd = f"cd {Dir('#').abspath}/tinygrad_repo && " + ' '.join(tinygrad_opts) + f" python3 openpilot/compile2.py {fn}.onnx {fn}.thneed"
+# Compile tinygrad model
+# TODO this is all super hacky
 
-  lenv.Command(fn + ".thneed", [fn + ".onnx"] + tinygrad_files, cmd)
+pythonpath_string = 'PYTHONPATH="${PYTHONPATH}:' + env.Dir("#tinygrad_repo").abspath + '"'
+if arch == 'larch64':
+  device_string = 'QCOM=1'
+else:
+  device_string = 'CLANG=1 IMAGE=0'
 
-  fn_dm = File("models/dmonitoring_model").abspath
-  cmd = f"cd {Dir('#').abspath}/tinygrad_repo && " + ' '.join(tinygrad_opts) + f" python3 openpilot/compile2.py {fn_dm}.onnx {fn_dm}.thneed"
-  lenv.Command(fn_dm + ".thneed", [fn_dm + ".onnx"] + tinygrad_files, cmd)
+for model_name in ['supercombo', 'dmonitoring_model']:
+  fn = File(f"models/{model_name}").abspath
+  cmd = f'{pythonpath_string} {device_string} python3 {Dir("#tinygrad_repo").abspath}/examples/openpilot/compile3.py {fn}.onnx {fn}_tinygrad.pkl'
+  lenv.Command(fn + "_tinygrad.pkl", [fn + ".onnx"] + tinygrad_files, cmd)
 
-  thneed_lib = env.SharedLibrary('thneed', thneed_src, LIBS=[gpucommon, common, 'OpenCL', 'dl'])
-  thneedmodel_lib = env.Library('thneedmodel', ['runners/thneedmodel.cc'])
-  lenvCython.Program('runners/thneedmodel_pyx.so', 'runners/thneedmodel_pyx.pyx', LIBS=envCython["LIBS"]+[thneedmodel_lib, thneed_lib, gpucommon, common, 'dl', 'OpenCL'])