Skip to content

GPU_support.md

Latest commit

 

History

History
59 lines (40 loc) · 3.58 KB

GPU_support.md

File metadata and controls

59 lines (40 loc) · 3.58 KB

GPU support with CUDA, cuDNN, TensorRT

I recommend the following tested configuration with CUDA ecosystem.

CUDA 11.8

  • CUDA 11.8 (or 11.6).
    • Install instructions here.
  • cuDNN 8.6.0.163-1+cuda11.8.
    • Install instructions here.
    • Once you downloaded the proper deb package, run the following command:
      sudo apt install -y libcudnn8=8.6.0.163-1+cuda11.8 libcudnn8-dev=8.6.0.163-1+cuda11.8
  • TensorRT 8.5.1.7 will be automatically installed by the main script build.sh (via install_local_tensorrt.sh) in thirdparty/TensorRT (once CUDA has been installed) and cmake will automatically use it by default.

CUDA 12.5 - WIP

Tested via rosdocker (container ubuntu24_cuda with CUDA 12.5). Currently work in progress.

  • CUDA 12.5
    • Install instructions here.
  • cuDNN 8.9.2.26~cuda12+3.
    • Install instructions here.
    • Once you downloaded the proper deb package, run the following command:
      sudo apt install -y nvidia-cudnn=8.9.2.26~cuda12+3
  • TensorRT 10.3.0.26 will be automatically installed by the main script build.sh (via install_local_tensorrt.sh)in thirdparty/TensorRT (once CUDA has been installed) and cmake will automatically use it by default.

Warning: Currently, there are some known issues with:

  • SAM under CUDA 12.5. The behaviour is not exactly the same we have under CUDA 11.8.
  • Superpointglue CUDA 12.5. The behaviour is not exactly the same we have under CUDA 11.8.

tensorflow_cc

As for tensorflow support, I recommend the usage of tensorflow_cc and the following tested configuration (default one for tensorflow_cc) under Ubuntu 20.04:

  • C++: 17
  • TENSORFLOW_VERSION: 2.9.0
  • BAZEL_VERSION: 5.1.1
  • CUDA: 11.6
  • CUDNN: 8.6.0.163-1+cuda11.8
    sudo apt install -y libcudnn8=8.6.0.163-1+cuda11.8 libcudnn8-dev=8.6.0.163-1+cuda11.8

As noted here, I successfully built and deployed other newer tensorflow configurations (see the list here). However, note that tensorflow does download and use its own custom versions of Eigen (and of other base libraries, according to the selected tensorflow version) and the used library versions may not be the same ones that are installed in your system. This fact may cause severe issues (undefined behaviors and uncontrolled crashes) in your final target projects (where you import the built and deployed Tensorflow C++): In fact, in such case, you may be mixing libraries built with different versions of Eigen (so with different data alignments and special effects)!

Feedback

Feel free to share any issue and suggestions. I'll be very glad to receive any feedback and improve the install procedure.