Hi!
I’m working on a Jetson Xavier, using the nvcr.io/nvidia/l4t-pytorch:r32.5.0-pth1.7-py3 docker image. (I’ve added pytorch-lightning and jupyter to it).
As the title says: when I’m trying to use half precision, I get an “Unable to find a valid cuDNN algorithm to run convolution” error.
I have no problem using single precision.
I have no problem using half precision on my PC with a GeForce RTX card.
Is it something fixable?
(More details…)
Using jetsonUtilities from jetsonhacks.com oustide the container, I get:
NVIDIA Jetson AGX Xavier [16GB]
L4T 32.5.1 [ JetPack 4.5.1 ]
Ubuntu 18.04.5 LTS
Kernel Version: 4.9.201-tegra
X Error of failed request: BadLength (poly request too large or internal Xlib length error)
Major opcode of failed request: 156 (NV-GLX)
Minor opcode of failed request: 1 ()
Serial number of failed request: 19
Current serial number in output stream: 19
CUDA 10.2.89
CUDA Architecture: 7.2
OpenCV version: 4.1.1
OpenCV Cuda: NO
CUDNN: 8.0.0.180
TensorRT: 7.1.3.0
Vision Works: 1.6.0.501
VPI: ii libnvvpi1 1.0.15 arm64 NVIDIA Vision Programming Interface library
Vulcan: 1.2.70
Using it inside the container, I get:
NVIDIA Jetson UNKNOWN
L4T N.N.N [ JetPack UNKNOWN ]
Ubuntu 18.04.5 LTS
Kernel Version: 4.9.201-tegra
CUDA 10.2.89
CUDA Architecture: NONE
OpenCV version: NOT_INSTALLED
OpenCV Cuda: NO
CUDNN: NOT_INSTALLED
TensorRT: NOT_INSTALLED
Vision Works: NOT_INSTALLED
VPI: NOT_INSTALLED
Vulcan: 1.2.70