A flexible and efficient deep neural network (DNN) compiler that generates high-performance executable from a DNN model description.
948
stars
158
forks
source link
[Feat] accelerate fp16 inference with cudnn library #444
Closed
LeiWang1999 closed 2 years ago
The cuda code nnfusion generated doesn't use tensor core.
This PR handles it, and takes about 2x speed up on fp16 models' inference.