Edit Info Other
Login

CUDA"

Differences between revisions 29 and 57 (spanning 28 versions)
Revision 29 as of 2019-04-06 11:57:18
Size: 5400
Comment:
Revision 57 as of 2021-03-12 10:04:40
Size: 7597
Comment:
Deletions are marked like this. Additions are marked like this.
Line 13: Line 13:
 * Fedora 33 and later (if using a compatible compiler, see also {{{
sudo dnf config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/fedora33/x86_64/cuda-fedora33.repo
sudo dnf clean all
sudo dnf module disable nvidia-driver
sudo dnf -y install cuda
}}}

 * RHEL/CentOS 8 {{{
sudo dnf config-manager --add-repo http://developer.download.nvidia.com/compute/cuda/repos/rhel8/x86_64/cuda-rhel8.repo
sudo dnf clean all
sudo dnf module disable nvidia-driver
sudo dnf -y install cuda
}}}
Line 14: Line 27:
sudo yum install https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-repo-rhel7-10.1.105-1.x86_64.rpm sudo yum-config-manager --add-repo http://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo
sudo yu
m clean all
Line 18: Line 32:
sudo yum install https://developer.download.nvidia.com/compute/cuda/repos/rhel6/x86_64/cuda-repo-rhel6-10.1.105-1.x86_64.rpm sudo yum-config-manager --add-repo http://developer.download.nvidia.com/compute/cuda/repos/rhel6/x86_64/cuda-rhel6.repo
sudo yu
m clean all
Line 20: Line 35:
}}}
 * Fedora 29 (and later) {{{
sudo dnf install https://developer.download.nvidia.com/compute/cuda/repos/fedora29/x86_64/cuda-repo-fedora29-10.1.105-1.x86_64.rpm
sudo dnf install cuda
Line 31: Line 42:

 * RHEL/CentOS 8 / Fedora {{{
sudo dnf install https://developer.download.nvidia.com/compute/machine-learning/repos/rhel8/x86_64/nvidia-machine-learning-repo-rhel8-1.0.0-1.x86_64.rpm
sudo dnf install libcudnn7 libcudnn7-devel libnccl libnccl-devel
}}}
Line 33: Line 50:
sudo yum install libcudnn7 libcudnn7-dev
}}}

 * Fedora {{{
sudo dnf install https://developer.download.nvidia.com/compute/machine-learning/repos/rhel7/x86_64/nvidia-machine-learning-repo-rhel7-1.0.0-1.x86_64.rpm
sudo dnf install libcudnn7 libcudnn7-dev
sudo yum install libcudnn7 libcudnn7-devel libnccl libnccl-devel
Line 42: Line 54:
== Legacy NVIDIA-340XX/CUDA ==
This repository contains a legacy version of CUDA 6.5 that will works with the NVIDIA 340xx serie
=== TensorRT repository ===

You can download the TensorRT component using the appropriate version from https://developer.nvidia.com/nvidia-tensorrt-download

This requires to login with the NVIDIA CUDA program subscription.


=== Legacy NVIDIA 340xx/CUDA 6.5 ===
This repository contains a legacy version of CUDA 6.5 that will works with the NVIDIA 340xx series
Line 63: Line 82:
 * Tweak the /usr/local/cuda-9.2/targets/x86_64-linux/include/host_defines.h to accept the Fedora default compiler. (Not recommended). Install an older gcc for dedicated for CUDA from COPR (Recommended on Fedora).
Line 65: Line 84:
 * Install the appropriate gcc version from developper toolset. It will install in parallel. Please see https://www.softwarecollections.org/en/scls/rhscl/devtoolset-7/
* GCC8 Works up to Fedora 32 for cuda-10.1 and later (up to CUDA 11)
{{{
dnf copr enable kwizart/cuda-gcc-10.1 -y
dnf install cuda-gcc cuda-gcc-c++ -y
}}}

You will need to tell CUDA to use it instead of using the default g++ this can be done for the cuda-samples with:
{{{
export HOST_COMP
ILER=cuda-g++
}}}

 * I
nstall the appropriate gcc version from developer toolset. It will install in parallel. Please see https://www.softwarecollections.org/en/scls/rhscl/devtoolset-8/
Line 68: Line 99:
sudo dnf install http://centos.mirrors.ovh.net/ftp.centos.org/7/extras/x86_64/Packages/centos-release-scl-rh-2-2.el7.centos.noarch.rpm
sudo dnf install http://centos.mirrors.ovh.net/ftp.centos.org/7/updates/x86_64/Packages/libgfortran4-7.2.1-1.2.1.el7_5.x86_64.rpm
sudo dnf install
devtoolset-7-toolchain
sudo dnf install https://rpmfind.net/linux/centos/7/extras/x86_64/Packages/centos-release-scl-rh-2-3.el7.centos.noarch.rpm
sudo dnf install devtoolset-8-toolchain
Line 72: Line 102:
You cannot install the whole devtoolset-7 collection, but the toolchain is enough , then each time you need to build using cuda, you start by You cannot install the whole devtoolset-8 collection, but the toolchain is enough , then each time you need to build using cuda, you start by
Line 74: Line 104:
scl run devtoolset-7 bash scl run devtoolset-8 bash
Line 76: Line 106:
gcc (GCC) 7.3.1 20180303 (Red Hat 7.3.1-5)
Copyright (C) 2017 Free Software Foundation, Inc.
gcc (GCC) 8.3.1 20190311 (Red Hat 8.3.1-3)
Copyright (C) 2018 Free Software Foundation, Inc.
Line 82: Line 112:
gcc (GCC) 8.1.1 20180712 (Red Hat 8.1.1-5)
Copyright (C) 2018 Free Software Foundation, Inc.
gcc (GCC) 9.2.1 20190827 (Red Hat 9.2.1-1)
Copyright (C) 2019 Free Software Foundation, Inc.
Line 88: Line 118:
 * Tweak the /usr/local/cuda*/targets/x86_64-linux/include/crt/host_defines.h to accept the Fedora default compiler. (Not recommended - not always working).
Line 89: Line 121:
Both "CUDA" and "RPM Fusion" repositories provide the nvidia driver packages. Usually, the package provided by RPM Fusion is higher. But in case you want to avoid the risk, add this: Both "CUDA" and "RPM Fusion" repositories provide the nvidia driver packages. Unfortunately, the packaging method is way too different and can conflicts. We recommends to use the publicly and community based packaging method (RPM Fusion) and avoid the NVIDIA packaged nvidia-driver. From time to time, NVIDIA uses non-publictly released driver, so you will have to wait for a public driver for the RPM Fusion counterpart...

With current RHEL8 repositories, the nvidia-driver is packaged as a module. So it's easy to disable with:
Line 91: Line 125:
#/etc/yum.repos.d/cuda.repo
[cuda]
name=cuda
...
exclude=xorg-x11-drv-nvidia*,akmod-nvidia*,kmod-nvidia*,nvidia-driver*,nvidia-settings,nvidia-xconfig,nvidia-persistenced,cuda-nvidia-kmod-common,dkms-nvidia,nvidia-libXNVCtrl
sudo dnf module disable nvidia-driver
Line 97: Line 127:


=== NVIDIA driver higher in CUDA repo ===
Often when NVIDIA release a newer CUDA version or even in the case of pre-release software the NVIDIA driver is at a higher version than the driver provided by RPM Fusion. There is no way for us to provide a version that will match the newer CUDA requirement "ahead" of any NVIDIA public driver release. With that said, the dependencies can sometime be faked at the RPM level with:
{{{
dnf module enable nvidia-driver -y && dnf download cuda-drivers && dnf module disable nvidia-driver -y
rpm -Uvh cuda-drivers*.rpm --nodeps
dnf update
}}}

Please remind to remove the cuda-drivers package when the RPM Fusion provided driver is high enough.
Complain to NVIDIA for this bad behaviour, not to us.

Once a newer version of the driver is available publicly, it will likely be available on the RPM Fusion rawhide repository in the first step, please follow this guide on how to upgrade to the newer driver (This is currently the case with CUDA 11 and 450xx driver serie) :
https://rpmfusion.org/Howto/NVIDIA#Latest.2FBeta_driver

Installation

This Howto provides a way to install the official NVIDIA packages for CUDA.

NVIDIA official repositories

These repositories contain versions of CUDA that are parallel installables along with another version.

CUDA Toolkit

Please use the Official link: https://developer.nvidia.com/cuda-downloads

  • Fedora 33 and later (if using a compatible compiler, see also

    sudo dnf config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/fedora33/x86_64/cuda-fedora33.repo
    sudo dnf clean all
    sudo dnf module disable nvidia-driver
    sudo dnf -y install cuda
  • RHEL/CentOS 8

    sudo dnf config-manager --add-repo http://developer.download.nvidia.com/compute/cuda/repos/rhel8/x86_64/cuda-rhel8.repo
    sudo dnf clean all
    sudo dnf module disable nvidia-driver
    sudo dnf -y install cuda
  • RHEL/CentOS 7

    sudo yum-config-manager --add-repo http://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo
    sudo yum clean all
    sudo yum install cuda
  • RHEL/CentOS 6

    sudo yum-config-manager --add-repo http://developer.download.nvidia.com/compute/cuda/repos/rhel6/x86_64/cuda-rhel6.repo
    sudo yum clean all
    sudo yum install cuda

Machine Learning repository

Please use the official link: https://developer.nvidia.com/nccl/nccl-download

  • RHEL/CentOS 8 / Fedora

    sudo dnf install https://developer.download.nvidia.com/compute/machine-learning/repos/rhel8/x86_64/nvidia-machine-learning-repo-rhel8-1.0.0-1.x86_64.rpm
    sudo dnf install libcudnn7 libcudnn7-devel libnccl libnccl-devel
  • RHEL/CentOS 7

    sudo yum install https://developer.download.nvidia.com/compute/machine-learning/repos/rhel7/x86_64/nvidia-machine-learning-repo-rhel7-1.0.0-1.x86_64.rpm
    sudo yum install libcudnn7 libcudnn7-devel libnccl libnccl-devel

TensorRT repository

You can download the TensorRT component using the appropriate version from https://developer.nvidia.com/nvidia-tensorrt-download

This requires to login with the NVIDIA CUDA program subscription.

Legacy NVIDIA 340xx/CUDA 6.5

This repository contains a legacy version of CUDA 6.5 that will works with the NVIDIA 340xx series

Please use the Official link: https://developer.nvidia.com/cuda-toolkit-65

  • RHEL/CentOS 6

    sudo yum install http://developer.download.nvidia.com/compute/cuda/repos/rhel6/x86_64/cuda-repo-rhel6-6.5-14.x86_64.rpm
    sudo yum install cuda
  • Fedora 20 (and later)

    sudo yum install install http://developer.download.nvidia.com/compute/cuda/repos/fedora20/x86_64/cuda-repo-fedora20-6.5-14.x86_64.rpm
    sudo yum install cuda

Please verify to have a compatible compiler.

Known issues

GCC version

When using a later version of Fedora than what is supported by the NVIDIA CUDA Official repository, you might be unable to compile. You can either:

Install an older gcc for dedicated for CUDA from COPR (Recommended on Fedora).

  • GCC8 Works up to Fedora 32 for cuda-10.1 and later (up to CUDA 11)

dnf copr enable kwizart/cuda-gcc-10.1 -y
dnf install cuda-gcc cuda-gcc-c++ -y

You will need to tell CUDA to use it instead of using the default g++ this can be done for the cuda-samples with:

export HOST_COMPILER=cuda-g++

sudo dnf install https://rpmfind.net/linux/centos/7/extras/x86_64/Packages/centos-release-scl-rh-2-3.el7.centos.noarch.rpm
sudo dnf install devtoolset-8-toolchain

You cannot install the whole devtoolset-8 collection, but the toolchain is enough , then each time you need to build using cuda, you start by

scl run devtoolset-8 bash
gcc --version
gcc (GCC) 8.3.1 20190311 (Red Hat 8.3.1-3)
Copyright (C) 2018 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
exit
gcc --version
gcc (GCC) 9.2.1 20190827 (Red Hat 9.2.1-1)
Copyright (C) 2019 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  • Tweak the /usr/local/cuda*/targets/x86_64-linux/include/crt/host_defines.h to accept the Fedora default compiler. (Not recommended - not always working).

Which driver Package

Both "CUDA" and "RPM Fusion" repositories provide the nvidia driver packages. Unfortunately, the packaging method is way too different and can conflicts. We recommends to use the publicly and community based packaging method (RPM Fusion) and avoid the NVIDIA packaged nvidia-driver. From time to time, NVIDIA uses non-publictly released driver, so you will have to wait for a public driver for the RPM Fusion counterpart...

With current RHEL8 repositories, the nvidia-driver is packaged as a module. So it's easy to disable with:

sudo dnf module disable nvidia-driver

NVIDIA driver higher in CUDA repo

Often when NVIDIA release a newer CUDA version or even in the case of pre-release software the NVIDIA driver is at a higher version than the driver provided by RPM Fusion. There is no way for us to provide a version that will match the newer CUDA requirement "ahead" of any NVIDIA public driver release. With that said, the dependencies can sometime be faked at the RPM level with:

dnf module enable nvidia-driver -y && dnf download cuda-drivers && dnf module disable nvidia-driver -y
rpm -Uvh cuda-drivers*.rpm --nodeps
dnf update

Please remind to remove the cuda-drivers package when the RPM Fusion provided driver is high enough. Complain to NVIDIA for this bad behaviour, not to us.

Once a newer version of the driver is available publicly, it will likely be available on the RPM Fusion rawhide repository in the first step, please follow this guide on how to upgrade to the newer driver (This is currently the case with CUDA 11 and 450xx driver serie) : https://rpmfusion.org/Howto/NVIDIA#Latest.2FBeta_driver

NVIDIA provided libOpenCL

NVIDIA only advertise OpenCL 1.2 with the binary driver at this time. As a consequence, they provide an old version of libOpenCL.so.1 which works fine with their binary driver. As most software in Fedora and RPM Fusion are built using a newer libOpenCL, the system linker detects that and issues the following message:

 /usr/local/cuda-9.2/targets/x86_64-linux/lib/libOpenCL.so.1: no version information available (required by ffmpeg)

You can either ignore the message or manually delete the libOpenCL.so.1 provided by NVIDIA (run sudo ldconfig once deleted). Please verify to not have other OpenCL providers that might interfere with NVIDIA OpenCL usage. (looking at /etc/OpenCL/vendors ).

Running blender

Even when only running blender, you need a CUDA compatible compiler as described above. This is because blender will compile the "CUDA Kernels" optimized for your own GPU. You can run blender with:

 scl run devtoolset-7 blender

Once the "CUDA kernels" are compiled, you can run blender normally

References


CategoryHowto

Howto/CUDA (last edited 2023-12-04 18:29:33 by NicolasChauvet)