PC: fix onnxruntime excessive CPU (#23051)

* onnxruntime-gpu two threads

* relock pipfile

Co-authored-by: jwolffe <wolffja@gmail.com>
Co-authored-by: Willem Melching <willem.melching@gmail.com>
old-commit-hash: db9ea45cd5

Pipfile

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:52b8d2b00d9d1d8b3430c2bc74e4ff51ab460317f931ed35cc74c6b9cc90f4d0
-size 1211
+oid sha256:46258ab149de8606b809e271264382d5450c11e914bb72c37697345290bd39f1
+size 1215

Pipfile.lock

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3a3f3fe1935735e8ad7b9f24c39e200e5d462704d97276520227043bab07890e
-size 145075
+oid sha256:871bf11488728e41699192b10314fde73cd8ca57acc803d452de76ac756bf3af
+size 144757

selfdrive/modeld/runners/onnx_runner.py

@@ -5,8 +5,9 @@ import sys
 import numpy as np
 
 os.environ["OMP_NUM_THREADS"] = "4"
+os.environ["OMP_WAIT_POLICY"] = "PASSIVE"
 
-import onnxruntime as ort
+import onnxruntime as ort # pylint: disable=import-error
 
 def read(sz):
   dd = []
@@ -38,22 +39,22 @@ def run_loop(m):
 
 
 if __name__ == "__main__":
-  print(ort.get_available_providers(), file=sys.stderr)
+  print("Onnx available providers: ", ort.get_available_providers(), file=sys.stderr)
+  options = ort.SessionOptions()
+  options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_DISABLE_ALL
   if 'OpenVINOExecutionProvider' in ort.get_available_providers() and 'ONNXCPU' not in os.environ:
-    print("OnnxJit is using openvino", file=sys.stderr)
-    options = ort.SessionOptions()
-    options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_DISABLE_ALL
     provider = 'OpenVINOExecutionProvider'
+  elif 'CUDAExecutionProvider' in ort.get_available_providers() and 'ONNXCPU' not in os.environ:
+    options.intra_op_num_threads = 2
+    provider = 'CUDAExecutionProvider'
   else:
-    print("OnnxJit is using CPU", file=sys.stderr)
-    options = ort.SessionOptions()
-    options.intra_op_num_threads = 4
+    options.intra_op_num_threads = 2
     options.inter_op_num_threads = 8
     options.execution_mode = ort.ExecutionMode.ORT_SEQUENTIAL
     options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
-
     provider = 'CPUExecutionProvider'
 
-  ort_session = ort.InferenceSession(sys.argv[1], options)
-  ort_session.set_providers([provider], None)
+  print("Onnx selected provider: ", [provider], file=sys.stderr)
+  ort_session = ort.InferenceSession(sys.argv[1], options, providers=[provider])
+  print("Onnx using ", ort_session.get_providers(), file=sys.stderr)
   run_loop(ort_session)