set_cudnn_flags¶

dragon.cuda.set_cudnn_flags(
  enabled=None,
  benchmark=None,
  deterministic=None,
  allow_tf32=None
)[source]¶

Set the flags of cuDNN library.

Parameters:

• enabled (bool, optional, default=True) – Use cuDNN library or not.
• benchmark (bool, optional, default=False) – Select fastest algorithms via benchmark or heuristics.
• deterministic (bool, optional, default=False) – Select deterministic algorithms instead of fastest.
• allow_tf32 (bool, optional, default=False) – Allow TF32 tensor core operation or not.