This documentation walks you through the installation of NVIDIA driver version 390.157 and CUDA-Toolkit-5.0 on Ubuntu 22.04 (kernel 6.2.0-39-generic). At the end of this guide you should be able to render the result of nbody sample from CUDA-Toolkit.
Note: During this documentation, feel free to man page every command you're not sure of
It all began when I upgraded my Ubuntu kernel version from 6.2.0-39-generic to 6.5.0-generic and my NVIDIA graphic card stopped working properly.
You can issue the kernel version via
$ uname -r6.2.0-39-genericIf you just want a short name of your system, normally your OS name you can do it via
$ uname -mLinuxA full system description can be issued via
$ uname -a
On Ubuntu when you click Power button\Settings\Info your Graphic Setting should give you detail like
GeForce 605/PCIe/SSE2and not
llvmpipeWhen the latter renderer is used which normally means the NVIDIA is not working properly it gives you very poor graphic quality which discourages you to use your machine a lot 🤣 Asking Microsoft Bing Copilot led me to this ubuntuforum thread The reason behind that was Linux kernel 6.5.0-generic was compiled by gcc-12 whereas the driver was compiled by gcc-11 and the incompatibility on the header files between two compiler versions led to the driver not being recognized by the new kernel.
You can check your gcc version by issuing
$ gcc --version
The Ubuntuforum thread offered two solutions for this particular issue: either to downgrade your gcc version and downgrade linux kernel altogether and reinstall NVIDIA driver or download a patched driver version from a launchpad site which I believe was compiled by gcc-12 and then reinstall it while the official fix from NVIDIA has not been released yet. I opted in the first option although I tested and could verify that the second option also worked. Details here
$ sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 11 --slave /usr/bin/g++ g++ /usr/bin/g++-11 --slave /usr/bin/gcov gcov /usr/bin/gcov-11
$ sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 12 --slave /usr/bin/g++ g++ /usr/bin/g++-12 --slave /usr/bin/gcov gcov /usr/bin/gcov-12
$ sudo update-alternatives --config gcc
$ sudo apt remove --purge --yes linux-headers-6.5.0-14-generic linux-image-6.5.0-14-generic linux-modules-6.5.0-14-generic linux-modules-extra-6.5.0-14-generic
$ sudo apt install --reinstall --yes linux-headers-6.2.0-39-generic linux-image-6.2.0-39-generic linux-modules-6.2.0-39-generic linux-modules-extra-6.2.0-39-generic
$ sudo apt-mark hold linux-headers-6.2.0-39-generic linux-image-6.2.0-39-generic linux-modules-6.2.0-39-generic linux-modules-extra-6.2.0-39-generic
# When they come up with a fix to compile NVidia Legacy drivers with Linux kernel 6.5 series kernels, then you can re-add that kernel series by unpinning the kernels via
# sudo apt-mark unhold linux-headers-6.2.0-39-generic linux-image-6.2.0-39-generic linux-modules-6.2.0-39-generic linux-modules-extra-6.2.0-39-generic
# sudo apt update && sudo apt upgrade
# Reboot into 6.2.0-39-generic kernel from GRUB options and reinstall nvidia-driver-390One question crossed my mind which as to how we could verify that the graphic driver had been properly installed.
Firstly one can check if your graphic card has been connected to PCI port. This command should give you similar result independent of the driver being properly installed
$ lspci | grep -i nvidia
01:00.0 VGA compatible controller: NVIDIA Corporation GF119 [GeForce 605] (rev a1)
01:00.1 Audio device: NVIDIA Corporation GF119 HDMI Audio Controller (rev a1)One may as well test the graphic driver using NVIDIA util like
$ nvidia-smi
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 390.157 Driver Version: 390.157 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 GeForce 605 Off | 00000000:01:00.0 N/A | N/A |
| 47% 64C P0 N/A / N/A | 472MiB / 961MiB | N/A Default |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: GPU Memory |
| GPU PID Type Process name Usage |
|=============================================================================|
| 0 Not Supported |
+-----------------------------------------------------------------------------+Actually things didn't work out flawlessly for me, I stumbled across different Ubuntuforums and ended up installing NVIDIA driver on 6.5.0-generic kernel and then mistakenly remove xserver-xorg which is a very BIG issue you might be yelling at me. But I was stupid and I commited the crime anyway. If you run into the problem that the desktop doesn't load and leaves you staring at the black screen, don't worry 'cause I have a fix for you.
Reboot into recovery mode, select something like Linux kernel 6.5.0-generic (recovery mode) from GRUB options. Or if you get stuck at the middle of the installation you can hold Alt + F[1-9] key. In my case I have FUJITSU ESPRIMO P910 model desktop at this time of writing it was F2 key which brings you to recovery mode. There you still have full access to file system and services. You're gonna start with reinstalling xserver-xorg and ubuntu-gnome-desktop
$ apt-cache search ubuntu-gnome-desktop
$ sudo apt install xserver-xorg ubuntu-gnome-desktop
$ sudo reboot
I recall that I even went too far by running
$ sudo telinit 3which I thought that might add to the problem that the GUI disappeared later. But as you find it out later in the official NVIDIA installation guide it was actually a good practice during installation. Basically it sets the runlevel to 3 which tells the system to disable GUI and that prevents the system from looping the unfinished GUI
If you'd rather write a rudimentary opengl program then here you go (Again Microsoft Bing Copilot did a great job except that it forgot to tell me to link all the necessary libs and I had to iteratively ask for the missing ones 😄)
#include <GL/glew.h>
#include <GLFW/glfw3.h>
int main() {
// Initialize the library
if (!glfwInit())
return -1;
// Create a windowed mode window and its OpenGL context
GLFWwindow* window;
window = glfwCreateWindow(640, 480, "Hello World", NULL, NULL);
if (!window) {
glfwTerminate();
return -1;
}
// Make the window's context current
glfwMakeContextCurrent(window);
// Initialize GLEW
if (glewInit() != GLEW_OK)
return -1;
// Loop until the user closes the window
while (!glfwWindowShouldClose(window)) {
// Render here
glClear(GL_COLOR_BUFFER_BIT);
// Swap front and back buffers
glfwSwapBuffers(window);
// Poll for and process events
glfwPollEvents();
}
glfwTerminate();
return 0;
}After saving the code as GLTest.cpp, compile it via
$ g++ GLTest.cpp -o GLTest -lglfw -lGLEW -lGL
$ ./GLTestAt the end you should see a black screen which confirms that OpenGL+NVIDIA Driver stack is working.
Well it's all good and gives you a pleasure of writing OpenGL code, does it? If you are an admin and prefer looking up the installation guide you may consult https://us.download.nvidia.com/XFree86/Linux-x86_64/390.157/README/index.html
This is indeed a very very good resource which you can dig in. To be honest I learned quite a bunch of basic Linux from it.
If lspci and nvidia-smi both give you normal result we can still verify your OpenGL/NVIDIA stack by means of glxinfo
$ sudo glxinfoIf it gives you the following you're good to go 😉
name of display: :1
display: :1 screen: 0
direct rendering: Yes
server glx vendor string: NVIDIA Corporation
server glx version string: 1.4
server glx extensions:
GLX_ARB_context_flush_control, GLX_ARB_create_context,
GLX_ARB_create_context_no_error, GLX_ARB_create_context_profile,
GLX_ARB_create_context_robustness, GLX_ARB_fbconfig_float,
GLX_ARB_multisample, GLX_EXT_buffer_age,
GLX_EXT_create_context_es2_profile, GLX_EXT_create_context_es_profile,
GLX_EXT_framebuffer_sRGB, GLX_EXT_import_context, GLX_EXT_libglvnd,
GLX_EXT_stereo_tree, GLX_EXT_swap_control, GLX_EXT_swap_control_tear,
GLX_EXT_texture_from_pixmap, GLX_EXT_visual_info, GLX_EXT_visual_rating,
GLX_NV_copy_image, GLX_NV_delay_before_swap, GLX_NV_float_buffer,
GLX_NV_multisample_coverage, GLX_NV_robustness_video_memory_purge,
GLX_SGIX_fbconfig, GLX_SGIX_pbuffer, GLX_SGI_swap_control,
GLX_SGI_video_sync
client glx vendor string: NVIDIA CorporationOn the contrary, if you get some errornous result like
name of display: :0
Error: couldn't find RGB GLX visual or fbconfigor indirectly from compile log freeglut (./nbody): OpenGL GLX extension not supported by display ':1' something is still off. If you run into trouble and does not know what's happening, you can find logs for most of the cases in /var/log/
Inspecting Xorg.0.log grep -i nvidia /var/log/Xorg.0.log as suggested from https://discussion.fedoraproject.org/t/error-couldnt-find-rgb-glx-visual-or-fbconfig/57646/8 showed me that the nvidia driver cannot be loaded.
[ 28.204] (II) LoadModule: "glx"
[ 28.204] (II) Loading /usr/lib/x86_64-linux-gnu/xorg/extra-modules/libglx.so
[ 28.279] (EE) Failed to load /usr/lib/x86_64-linux-gnu/xorg/extra-modules /libglx.so: libnvidia-tls.so.390.157: cannot open shared object file: No such file or directory
[ 28.279] (II) UnloadModule: "glx"
[ 28.279] (II) Unloading glx
[ 28.279] (EE) Failed to load module "glx" (loader failed, 7)More specifically freeglut cannot dynamically link to libnvidia-tls, libnvidia-glcore and libnvidia-ml
$ ldd /usr/lib/x86_64-linux-gnu/nvidia/xorg/libglx.so
...
libnvidia-tls.so.390.157 => not found
libnvidia-glcore.so.390.157 => not found
...Quick explanation from Stackoverflow:
GLX is the X11 protocol extension used to setup OpenGL contexts on X11 drawables. However this is an extension provided by the device driver. You are using a NVidia card. My guess is, that this is a vanilla installation of a system that doesn't automatically install the proprietary nvidia drivers and neither configures the open nouveau drivers.
So the X11 server probably uses either the nv or the fbdev or the vesa driver; none of those support OpenGL/GLX.
Solution: Install and configure the proper driver. Either nouveau or the drivers you can download from http://www.nvidia.com/object/unix.html and install that
My initial plan was to install Cuda Toolkit in order to run Tensorflow code with GPU accelarated on my Ubuntu machine which I later realized that Tensorflow requires Compute Capability 3.0 :sick. In simple words, CUDA is an abstract layer for developer to write General Purpose GPU code on your graphic card as a hardware. By means of CUDA SDK you can program your Processing Units, make them parallel and able to communicate with CPU without the need to involve with low-level GL operations. Of course CUDA requires that NVIDIA or other provider driver has to be installed beforehand.
One difficulty is that you must determine the right CUDA version for your driver (my old and weak buddy). It is recommended to check the documentation carefully before installation or you'll end up going through vicious circle of installation/removal of CUDA Toolkit. You may need to deep dive into this list
https://docs.nvidia.com/deploy/cuda-compatibility/index.html#binary-compatibility__table-toolkit-driver (Compatibility concept)
https://docs.nvidia.com/cuda/archive/11.0/cuda-toolkit-release-notes/index.html#title-new-features (Release Note)
https://developer.nvidia.com/cuda-gpus (Compute Capability Table)
https://developer.nvidia.com/cuda-toolkit-archive (Older Toolkit download page)
Now I'm about to note down my installation steps for future reference, it's a reward for your hardwork to keep your old man up and running 😃
#if you're installing cuda toolkit 9
$ wget https://developer.download.nvidia.com/compute/cuda/9.0/secure/Prod/local_installers/cuda-repo-ubuntu1704-9-0-local_9.0.176-1_amd64.deb
$ sudo dpkg -i ./cuda-repo-ubuntu1704-9-0-local_9.0.176-1_amd64.deb # if you're installing cuda toolkit 9
$ sudo apt update
$ sudo apt install cuda-toolkit-9-0
#if you're installing cuda toolkit 5
$ wget https://developer.download.nvidia.com/compute/cuda/5_0/rel-update-1/installers/cuda_5.0.35_linux_64_ubuntu11.10-1.run
$ sudo ./cuda_5.0.35_linux_64_ubuntu11.10-1.run
# set environment variables
$ export PATH=/usr/local/cuda/bin:$PATH
$ export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
$ nvcc --version # it should print out the right version on successNow that you have your CUDA Toolkit installed, the next thing you'll want to do is to navigate to sample folder and then compile it and run the sample application. With CUDA version 9.0 you can find its documentation online at https://docs.nvidia.com/cuda/archive/9.0/cuda-samples/index.html#simple. With version 5.0 the documentation can be found in sample folder. By the time of this writing there was already a github repository for CUDA samples which can be found at https://github.com/NVIDIA/cuda-samples.git. Unfortunately it doesn't support my old buddies 😦 and I have to resort to using sample code shipped with CUDA package. In the following subsection I mean $CUDA_9_INSTALL_DIR=/usr/local/cuda-9.0 and $CUDA_5_INSTALL_DIR=/usr/local/cuda-5.0 by convention. Because cuda-5.0 installation kit installs samples in a separate directory as CUDA_5_INSTALL_DIR, I use CUDA_5_SAMPLE_DIR to denote the sample directory for cuda-5.0
The very basic application you will want to build is the deviceQuery or deviceQueryDrv which is located at $CUDA_9_INSTALL_DIR/samples/1_Utilities/deviceQuery[Drv] with CUDA Toolkit 9.0 and $CUDA_5_SAMPLE_DIR/1_Utilities/deviceQuery[Drv] with CUDA Toolkit 5.0 respectively. Now cd into the source directory, create objdir directory (it's always a good practice to create this dir to hold our built artifacts) then cd to it. After running make (cmake) you might run into the following error/warning message
MapSMtoCores for SM 2.1 is undefined. Default to use 128 Cores/SM
# or
CUDA error at bodysystemcuda_impl.h:302 code=13(cudaErrorInvalidSymbol) "setSofteningSquared(softeningSq)"As I mentioned in previous section, you should carefully check your CUDA Compute Capabilities to avoid unnecessary incompatibility problems and now this error message rings a bell that I have too new CUDA version which no longer supports my old graphic driver. My Fermi graphic card has SM 2.1 architecture and it cannot be found by CUDA 9.0 and ironically it defaults to Hopper with 128 Cores which is the latest at this time being. I indeed overlooked the minimum requirement for CUDA-9.0 and installed it despite the requirement. Another tale tell sign to be sure your graphic driver (capability) is support is to look into the Makefile or CMakeLists.txt and search for supported architectures. For example:
in $CUDA_5_SAMPLE_DIR/1_Utilities/deviceQueryDrv/Makefile
# CUDA code generation flags
GENCODE_SM10 := -gencode arch=compute_10,code=sm_10
GENCODE_SM20 := -gencode arch=compute_20,code=sm_20
GENCODE_SM30 := -gencode arch=compute_30,code=sm_30 -gencode arch=compute_35,code=sm_35
GENCODE_FLAGS := $(GENCODE_SM10) $(GENCODE_SM20) $(GENCODE_SM30)in $CUDA_5_SAMPLE_DIR/5_Simulations/nbody/CMakeLists.txt
set(GENCODE -gencode=arch=compute_30,code=sm_30 -gencode=arch=compute_30,code=compute_30)
set(GENCODE ${GENCODE} -gencode=arch=compute_20,code=sm_20 -gencode=arch=compute_20,code=compute_20)Note the difference between archtype and code, they don't always map to the same figure. Details can be found here https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#application-compatibility.
In my case the variables should be passed as arch=compute_20,code=sm_20 so that nvcc will later embed binary code compatible with compute capability 2.0. As in https://developer.nvidia.com/cuda-gpus (Compute Capability Table) my graphic card maps to compute capability 2.1. So host code is generated to automatically select at runtime 2.0 binary code (most appropriate code) to load and execute for my device.
in $CUDA_9_INSTALL_DIR/samples/1_Utilities/deviceQuery/Makefile
# Gencode arguments
SMS ?= 30 35 37 50 52 60 70In any case, you should read the Release Notes to get informed of the new features after each CUDA Toolkit comes around.
Once you're done with that basic application you can go ahead building the n-body sample application. One problem I encountered was that one particular CUDA header file purposely refuses to accept higher version of gcc. In particular, after installing CUDA Toolkit 9.0 (or 5.0) and compiled the n-body source code an error was thrown due to this assertion in $CUDA_9_INSTALL_DIR/include/crt/host_config.h
#if __GNUC__ > 6
#error -- unsupported GNU version! gcc versions later than 6 are not supported!
#endif /* __GNUC__ > 6 */and in CUDA_5_INSTALL_DIR/include/host_config.h
#if defined(__GNUC__)
#if __GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ > 6)
#error -- unsupported GNU version! gcc 4.7 and up are not supported!
#endif /* __GNUC__> 4 || (__GNUC__ == 4 && __GNUC_MINOR__ > 6) */
#endif /* __GNUC__ */I decided to make this section short and to the point. In general you have two options: build gcc from source code (I'm bad at this although I followed instructions at https://gcc.gnu.org/wiki/InstallingGCC all along) or from downloaded package which is done after finding an appropriate ppa or launchpad repository which covers your old gcc. Well the old guys oftentimes team up together 🤷 I resorted to build it the easier way because first of all I'm not confident with figuring out what could go wrong if make gets stuck at the middle of the build and I'm pretty sure you won't want to wait for an hour compiling gcc from source unless you have a good reason to do so.
# The following is to install gcc-6.3
$ wget https://drive.google.com/drive/folders/1xVEATaYAwqvseBzYxKDzJoZ4-Hc_XOJm/gcc63-c++-6.3.0-1_amd64.deb
$ sudo dpkg -i gcc63-c++-6.3.0-1_amd64.deb
$ sudo apt update
$ sudo apt install gcc63
# Not necessary, only if you want to add repository
$ sudo vi /etc/apt/sources.list
# add one entry into the list which contains URL and distribution of the repo
$ sudo apt update
# The following is to install gcc-4.0
$ wget https://mirrors.edge.kernel.org/ubuntu/pool/universe/g/gcc-4.6/libstdc++6-4.6-dev_4.6.4-6ubuntu2_amd64.deb
$ wget https://mirrors.edge.kernel.org/ubuntu/pool/universe/g/gcc-4.6/gcc-4.6_4.6.4-6ubuntu2_amd64.deb
$ wget https://mirrors.edge.kernel.org/ubuntu/pool/universe/g/gcc-4.6/g++-4.6_4.6.4-6ubuntu2_amd64.deb
$ wget https://mirrors.edge.kernel.org/ubuntu/pool/universe/g/gcc-4.6/gcc-4.6-base_4.6.4-6ubuntu2_amd64.deb
$ wget https://mirrors.edge.kernel.org/ubuntu/pool/universe/g/gcc-4.6/cpp-4.6_4.6.4-6ubuntu2_amd64.deb
$ wget http://launchpadlibrarian.net/336269522/libmpfr4_3.1.6-1_amd64.deb
$ sudo dpkg -i libmpfr4_3.1.6-1_amd64.deb
$ sudo apt update
$ sudo apt install libmpfr4
$ sudo apt install ./gcc-4.6_4.6.4-6ubuntu2_amd64.deb ./gcc-4.6-base_4.6.4-6ubuntu2_amd64.deb ./libstdc++6-4.6-dev_4.6.4-6ubuntu2_amd64.deb ./cpp-4.6_4.6.4-6ubuntu2_amd64.deb ./g++-4.6_4.6.4-6ubuntu2_amd64.deb
$ gcc-4.6 --version
# Configure default gcc when you have multiple gcc versions installed
$ sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc63 90 --slave /usr/bin/g++ g++ /usr/bin/g++63
$ sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 80 --slave /usr/bin/g++ g++ /usr/bin/g++-11 --slave /usr/bin/gcov gcov /usr/bin/gcov-11
$ sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 70 --slave /usr/bin/g++ g++ /usr/bin/g++-12 --slave /usr/bin/gcov gcov /usr/bin/gcov-12
$ sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-4.6 110 --slave /usr/bin/g++ g++ /usr/bin/g++-4.6Another problem I encountered was that since the release of glibc-xx it seemed to tamper with part of the system which supports large complex number, as we name it float128
To check glibc version, do this following command
$ ldd --version
ldd (Ubuntu GLIBC 2.35-0ubuntu3.7) 2.35
Copyright © 2022 Free Software Foundation, Inc.
Dies ist freie Software; in den Quellen befinden sich die Lizenzbedingungen.
Es gibt KEINERLEI Garantie; nicht einmal für die TAUGLICHKEIT oder
VERWENDBARKEIT FÜR EINEN ANGEGEBENEN ZWECK.
Implementiert von Roland McGrath und Ulrich Drepper.Another command give you a slightly different result
GNU C Library (Ubuntu GLIBC 2.35-0ubuntu3.7) stable release version 2.35.
Copyright (C) 2022 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.
Compiled by GNU CC version 11.4.0.
libc ABIs: UNIQUE IFUNC ABSOLUTE
For bug reporting instructions, please see:
<https://bugs.launchpad.net/ubuntu/+source/glibc/+bugs>.A short background about ldd: It is a standard dynamic linker which can inspect the program's dynamic dependencies and load the objects that satisfy dependencies.
To fix this, firstly cd into /usr/include/x86_64-linux-gnu/bits (In other Linux system it may be different than this, for example /usr/include/bits/floatn.h), create a new file named floatn.h.patch and paste the following content into it. Don't forget to do it in sudo mode 😉
--- a/usr/include/bits/floatn.h 2017-09-13 16:00:37.079654880 +0200
+++ b/usr/include/bits/floatn.h 2017-09-13 16:00:00.999104815 +0200
@@ -31,7 +31,8 @@
support, for x86_64 and x86. */
#if (defined __x86_64__ \
? __GNUC_PREREQ (4, 3) \
- : (defined __GNU__ ? __GNUC_PREREQ (4, 5) : __GNUC_PREREQ (4, 4)))
+ : (defined __GNU__ ? __GNUC_PREREQ (4, 5) : __GNUC_PREREQ (4, 4))) \
+ && !defined(__CUDACC__)
# define __HAVE_FLOAT128 1
#else
# define __HAVE_FLOAT128 0After that apply the patch
$ sudo patch floatn.h floatn.h.patch
If it logs out this, you're one step further 😙
patching file floatn.h
Hunk #1 succeeded at 28 (offset -3 lines).The cleaner way would be to download glibc-source, create a patch from that source and then apply the patch on your system. But I was too lazy to try it out and was ready to take risk of breaking other packages
To fix 'dso-missing' linking error we can add customized option to gcc:
export LDFLAGS="-Wl,--copy-dt-needed-entries"See https://linuxpip.org/how-to-fix-dso-missing-from-command-line?utm_content=cmp-true
In CMakeLists.txt we add the following line in CMakeLists to set option for target link
target_link_options(nbody PRIVATE "LINKER:--copy-dt-needed-entries")And adding two more OpenGl libraries to the original CMakeLists.txt file seems to fix the last hurdle!
if (WIN32)
if( CMAKE_SIZEOF_VOID_P EQUAL 8 )
set( LIB_PATH ${CUDA_SDK_ROOT_DIR}/common/lib/x64/ )
set( GLEW_NAME glew64 )
else( CMAKE_SIZEOF_VOID_P EQUAL 8 )
set( LIB_PATH ${CUDA_SDK_ROOT_DIR}/common/lib/win32/ )
set( GLEW_NAME glew32 )
endif( CMAKE_SIZEOF_VOID_P EQUAL 8 )
else (WIN32)
if( CMAKE_SIZEOF_VOID_P EQUAL 8 )
set( LIB_PATH ${CUDA_SDK_ROOT_DIR}/common/lib/linux/x86_64/ )
set( GLEW_NAME "libGLEW.so" ) # this was hardcoded
set(GLU_NAME "libGLU.so") # this was also hardcoded
else( CMAKE_SIZEOF_VOID_P EQUAL 8 )
set( LIB_PATH ${CUDA_SDK_ROOT_DIR}/common/lib/linux/i686/ )
set( GLEW_NAME GLEW )
endif( CMAKE_SIZEOF_VOID_P EQUAL 8 )
endif (WIN32)
if (WIN32)
FIND_LIBRARY(GLEW_LIBRARY NAMES ${GLEW_NAME} PATHS ${LIB_PATH})
else (WIN32)
message("GLEW_NAME=" ${GLEW_NAME})
FIND_LIBRARY(GLEW_LIBRARY NAMES ${GLEW_NAME} PATHS ${LIB_PATH})
FIND_LIBRARY(GLU_LIBRARY NAMES ${GLU_NAME} PATHS ${LIB_PATH})
endif (WIN32)Rerun cmake and make and finally
Bildschirmaufzeichnung.vom.22.04.2024.12.22.27.webm
Voila! We made it! Although not 100% as my original plan because Tensorflow supports SM 3.0 at the minimum but still we've been through a lot and learned a lot from this small experiment!
The whitepaper was uploaded to this repository in case you want to read it.
The reason why Tensorflow requires NVIDIA graphic card of compute capability >= 3.0 is that https://www.tensorflow.org/install/source#configuration_options
GPU support For GPU support, set cuda=Y during configuration and specify the versions of CUDA and cuDNN.
This implicitly means Tensorflow uses cuDNN as dependency. Yet you can argue to install the very old version of tensorflow which is for example accessible via https://github.com/tensorflow/tensorflow/tree/v1.9.0 and get the very first version of cuDNN https://developer.nvidia.com/rdp/assets/cudnn-65-linux-r1 as suggested in the last answer in https://stackoverflow.com/questions/38542763/how-can-i-make-tensorflow-run-on-a-gpu-with-capability-2-x. I have downloaded cuDNN 6.5 R1 and when I opened User's Guide pdf file it read
2.4. Requirements cuDNN supports NVIDIA GPUs of compute capability 3.0 and higher and requires an NVIDIA Driver compatible with CUDA Toolkit 6.5.
Not tested but worth trying: tensorflow/tensorflow#1498. In the discussion it was figured out the main problem was:
in this thread wants to support compute_20 device. However, since Eigen starts to use warp shuffling instructions, which is only available for compute_30 devices and later, it won't work.
If you've already got NVIDIA GPU of compute capability 3.0 and still enjoy the thrill of building Tensorflow from source you can get support at yaroslavvb/tensorflow-community-wheels#107