GPU not using gpu on Windows 10

[lucydjo]

Hi !

I compiled darknet with vcpkg and build.ps1.
I have modified the Makefile like this before:

GPU=1
CUDNN=1
CUDNN_HALF=0
OPENCV=1
AVX=0
OPENMP=0
LIBSO=0
ZED_CAMERA=0

Unfortunately my CPU is used instead of the GPU.

darknet.exe detector demo cfg/coco.data cfg/yolov3.cfg yolov3.weights road.mp4 -gpus 0

I have the impression that the modification of the Makefile didn't change anything. Do you have a solution? Have I forgotten something?

Thx !

[sholalkere]

I'm having the same issue, torch can use cuda and everything but for some reason it darknet doesnt use it

[cenit]

build.ps1 triggers the CMake build, which should detect gpu and cuda automatically without requiring any modification at all. Can you please post some configuration and some build logs? Please restart from a clean repository if possible, without any modification.

[lucydjo]

I didn't know what files you were talking about, so I created a git with all the files after compilation.

https://github.com/lucaspojo/darknet-yolo-again

EDIT :

Console log : https://gist.github.com/lucaspojo/d207a2743bae3b59334f530c2e5f1f82
CmakeCache : https://gist.github.com/lucaspojo/4f577e509e9ce5433c4f92e246f78229

Thx !

[cenit]

our .gitignore is made so no build artifacts can be pushed to a repository. From a quick glance there's nothing more in your repo than in a clean repo.

What I mean is just copy-paste the messages that appear on console after you launch build.ps1, and also the CMakeCache.txt file that is created inside the build_win_release folder.

Again, please launch build from a clean repository folder and do not modify any code

[lucydjo]

Console log : https://gist.github.com/lucaspojo/d207a2743bae3b59334f530c2e5f1f82
CmakeCache : https://gist.github.com/lucaspojo/4f577e509e9ce5433c4f92e246f78229

I did start the build again with a new repo. I started from scratch.

[cenit]

Your compiler version requires at least CUDA 10.1, if I remember exactly.
Could you please update CUDA to 10.1 and retry?

[cenit]

-- Looking for a CUDA compiler
-- Looking for a CUDA compiler - NOTFOUND

these lines are key to understand. Visual Studio reports back that there's no CUDA integration installed. Since you have cuda installed (-- Found CUDA: C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v10.0 (found suitable exact version "10.0")), this just means that your cuda version is not properly installed or just "too old"

[lucydjo]

I changed the version of cuda to 10.1. I have a new error. (compilation of zero at each attempt).

CMake Error at C:/Users/lpjo/Desktop/vcpkg-master/installed/x64-windows/share/opencv/OpenCVConfig.cmake:108 (message):
OpenCV static library was compiled with CUDA 10.0 support. Please, use the
same version or rebuild OpenCV with CUDA 10.1

I tried to upgrade from version 3.0 to 4.1 of openCV but the error remains the same.
I'll try to compile opencv myself

[lucydjo]

Idk how build opencv.

After several attempts by following instructions find on the opencv website and other forum. I can't do it.

I think I'm going to have to go back to linux! Too bad

[cenit]

nope, don't worry
to build opencv for windows, please follow the vcpkg section of our readme.

[lucydjo]

I made another attempt with Cuda 10.1, but the GPU still hasn't used.

CMakeCache : https://gist.github.com/lucaspojo/d5068fd6552c13331f9f9ce82edfb94b

I have the same :

-- Looking for a CUDA compiler
-- Looking for a CUDA compiler - NOTFOUND

[cenit]

Is the nvidia installer giving you any problem? Something bad must be going on there. The message indicates that for VS there is no cuda installed!
If you open visual studio, can you create a new empty cuda project?

[lucydjo]

Nvidia Cuda does not appear in the list when I want to create a new project.

Nvcc --version return :

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2019 NVIDIA Corporation
Built on Fri_Feb__8_19:08:26_Pacific_Standard_Time_2019
Cuda compilation tools, release 10.1, V10.1.105

did I forget something?

[lucydjo]

I found a solution by accident. I installed Cuda 8, restarted my PC, reinstalled cuda 10.1, launched GeforceExperience to update the graphics driver, restarted the compilation, and everything works!

I hope it will help someone else.

[cenit]

Ok good. Glad that you were able to fix your CUDA setup :)

[cenit]

In case you find this problem again, we now have also another workaround that might help you

1. pull the latest master
2. edit build.ps1 and write "true" next to $use_ninja at the very beginning
3. make sure you have ninja-build installed
4. try to launch .\build.ps1 again.

This procedure does not need VS at all, just the compiler (Build Tools are enough). It might work also in your previous "broken" CUDA status

[Monkey1GIt]

My environment: Win10 + cuda 10.1 + cuddn + Visual Studio 2019.

I think I found the reason, this is problem between the cuda and Visual Studio. Because the cuda should be installed after the Visual Studio. When cuda is installed there will be a item for finding whether you have installed Visual Studio before, if you did then cuda adapting the environment, if you didn't you can ignore. Like #2971 (comment)

You can reinstall the cuda to solve the problem. There is a way to test if cuda is satisfied to Visual Studio:

open Visual Studio 2019 > New > New item

you will find a item named cuda 10.1 Runtime, like this:

The cmake will like this: