search

ceccocats / tkDNN

Deep neural network library and toolkit to do high performace inference on NVIDIA jetson platforms
GNU General Public License v2.0
718 stars 209 forks source link

Support for cuDNN 8.1.1 #214

Closed syjeon121 closed 2 years ago

syjeon121 commented 3 years ago

hi @mive93 according to cudnn documentation, the performance of RTX 3090 has improved in 8.1.1 so i am currently testing on cuDNN 8.1.1, TensorRT 7.2.3.4 using RTX 3090

can i see these improvement in current tkDNN code?

mive93 commented 3 years ago

Hi @syjeon121, sorry but I don't get the question. Are you asking if CuDNN 8.1.1 is supported?

syjeon121 commented 3 years ago

thanks @mive93 yes, and TensorRT 7.2.3.4 too

perseusdg commented 3 years ago

@syjeon121 I was able to use tkdnn with cudnn 8.1.1 and tensorrt 7.2.3 on ubuntu 20.04

mive93 commented 3 years ago

Thanks @perseusdg :)

syjeon121 commented 3 years ago

Thanks!

syjeon121 commented 3 years ago

@mive93 @perseusdg hi, sorry to reopen this issue one thing i'm curious about is how to use internal function according to the cuDNN version up for example convolution forward function

or is it automatically applied by simply upgrading the version?

perseusdg commented 3 years ago

@syjeon121 are you asking how tkdnn is able to work with different versions of cudnn ?

syjeon121 commented 3 years ago

@perseusdg yes. so i want to apply cuDNN 8.1.1 function if there is something new

perseusdg commented 3 years ago

you can check this commit out to see the changes made to make the library work with cudnn 8.x and going by the improvements/new features mentioned in the release notes of cudnn 8.1.1 i dont think there is anything new that is pertinent to this library

syjeon121 commented 3 years ago

@perseusdg ok, i will check the commit thanks!