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 208 forks source link

Any updates if it works with new versions of cuda 11.8? #304

Open yemuzi opened 10 months ago

yemuzi commented 10 months ago

on 2070 gpu card, image

Cuda failure: out of memory /root/trt8/src/Conv2d.cpp:177 Aborting...

ckurtz22 commented 8 months ago

I've tested it up to Cuda 12.3, works fine. The memory error is because the tests are written around running it on a Jetson Xavier with unified memory. For normal GPUs you need to comment out this line so when it can just serialize the network without running it: https://github.com/ceccocats/tkDNN/blob/master/src/Conv2d.cpp#L176 And then you can load the network.rt file and run it later on.

vtyw commented 7 months ago

@ckurtz22 What version of OpenCV did you have to use with CUDA 12?

ckurtz22 commented 7 months ago

@ckurtz22 What version of OpenCV did you have to use with CUDA 12?

I used the default opencv installation for Ubuntu 20.04, I believe it was 4.1. I also was able to build 4.8 from source and get it working without issue. I have not been using OpenCV with CUDA enabled,