@szha - I agree, but, we need to expose the control for users as they have
preference on these choices of underlying algorithm.
I am thinking of standardizing the API with one parameter for cudnn controls.
```
conv( channels, ..., cudnn={‘off’, 'default', ‘fastest’, ‘limited_workspace’,
‘global’})
default: limited_workspace
off: Do not use cudnn for this layer.
default: Use Cudnn. Do not tune.
limited_workspace: Use Cudnn. Run the test and pick the fastest algorithm that
doesn’t exceed workspace limit.
fastest: Use Cudnn. Pick the fastest algorithm and ignore workspace limit.
global: Use Cuddn. Use global Cudnn behavior determined by the environment
variable MXNET_CUDNN_AUTOTUNE_DEFAULT. 0 for off, 1 for limited workspace
(default), 2 for fastest.
```
@nswamy @szha @sbodenstein - Let me know what you guys think. I will make
corresponding changes.
[ Full content available at:
https://github.com/apache/incubator-mxnet/pull/12691 ]
This message was relayed via gitbox.apache.org for [email protected]
