Compiling OpenCV with CUDA Support on Windows

-

I have actively avoided doing this for a while, but now I need to run CUDA-dependent OpenCV Code on a Windows Machine.  So it's time to bite the bullet and get it done.

Want to go for a ride?  It will be fun, I promise!

Background

This is being run on a Gigabyte Motherboard with a 10th-Generation i7 CPU, 64GB RAM and Windows 11 (Release).  You want at least 4GB of RAM and an i5 to do the build with any sort of efficiency.

First, get your pre-requisites

You will need:

Install CUDA, Python, Visual Studio and CMake.  Make sure they are in the user path, and that a handler is installed for "sln" files.  This should be the case by default.

As a sanity check, run "nvidia-smi" to make sure you can talk to your GPU.  On Windows, installing CuDNN involves manually copying files into the CUDA directories, so follow the installation instructions carefully.

If possible, do your build on an SSD, because parts of the process are very I/O heavy.

Configure Everything

Yes, you can use the CMake GUI, but you could also put lit bamboo skewers under your toenails.  The command-line is actually easier.

Firstly extract opencv-x.x.x into a directory (E:\src\opencv\opencv-4.5.4 in my case) and then extract opencv_contrib-x.x.x at the same level.  You should have two directories thus:


Now, change into the opencv-x.x.x directory and create a build directory under that.  Call it something useful like "build".  Change  into that directory.

Now here's the magic.  We're going to get CMake to do all the hard work for us with one simple command:

cmake -DWITH_CUDA=YES -DOPENCV_DNN_CUDA=YES -DENABLE_FAST_MATH=YES -DOPENCV_EXTRA_MODULES_PATH=E:\src\opencv\opencv_contrib-x.x.x\modules -DCUDA_FAST_MATH=YES -DCUDA_ARCH_BIN=7.5 ..

All of the above is a single line, so be careful with the copy and paste.  Also note that the CUDA_ARCH_BIN value needs to be set according to the GPU in use.  You can find a list here: https://en.wikipedia.org/wiki/CUDA.

There are also a load more options you could specify depending on your needs.  Please refer to the OpenCV documentation, but the above will get you started.

That will grind away for a while.  Take a few sips of your beverage.  If all goes to plan, you will have a file "OpenCV.sln" in your build directory.

The Build

Now, open the OpenCV.sln file in Visual Studio".  You should simply be able to type "OpenCV.sln" in the command prompt, and Windows will work stuff out for you.  You'll have a solution view resembling the following:

Probably a good idea to select "Release" for your build type since there is not too much benefit to building debug binaries in this case


Now, expand "CMakeTargets in the Solution Explorer and choose "ALL_BUILD"



Click "Build" from the right-click menu and take a long draught from your beverage.  This is going to take quite a while even on a fast machine.  My system took just on 30 minutes to complete.

There's a lot of warnings, but it should build without error.  Progress will look like this:

Installation

You're probably onto your second pot of coffee by now, but we're approaching the finish line, I promise...

Select the INSTALL project and choose "build" from the context menu.  This will take a couple of minutes to run.

Now try importing "cv2" from within the Python interpreter.  If it succeeds, you're done.  If this fails, go back and check that CMake detected Python in the first instance.









Comments

Post a Comment

Popular posts from this blog

The QYT KT-7900D

-
Image
But is it any good? As you might have noticed, I am now a freshly licensed radio amateur .  As such I do need a transceiver or two in order to get on the air.  This is the story of one of my new radios. I have a limited discretionary budget to use for my new hobby, and I have to say from the outset that my lovely wife has been very supportive.  As such, I want to try and get the best value for our money.  One evening I stumbled upon this little item on Banggood for a little over $A150 including shipping.  Covering both 2m and 70cm bands with a claimed 20W output it seemed almost too good to be true.  I bought one. About ten days later a tiny package arrived for me.  At just 98mm x 43mm x 126mm the photos don't convey quite how small it actually is.  The entire case serves as a heatsink so it is heavier than it looks, but certainly not bulky. The standard power cable comes with a cigarette lighter plug.  As I intended to run it from home using a regulated PSU, I just cut the plug o
Read more

Life with Deep Racer - Part 2

-
Image
 An Update on AWS Deep Racer OK, so AWS Summit Sydney has been run and won.  Not by us however, but there are few interesting tidbits to report: The console is now widely available This is actually pretty big news, because: The console ties together Robomaker, SageMaker and everything else you need in a convenient package There are actual examples of functions included It seems to stay running now There is new firmware for the car Also important; this includes a rolling update for the Ubuntu base as well as the Deep Racer specific bits.  WiFi connection seems more reliable now as well. Our car has a new look Since my employer (Versent) was generous enough to send me to Re:Invent last year (where I got the Deep Racer), I thought it appropriate to deck it out in the company colours, and it was well-received at Summit last week.
Read more

DIY Self-Driving - Part 9 - New Imaging

-
Image
Please Note:    Unlike most of my projects which are fully complete before they are published, this project is as much a build diary as anything else.  Whilst I am attempting to clearly document the processes, this is not a step-by-step guide, and later articles may well contradict earlier instructions. If you decide you want to duplicate my build, please read through all the relevant articles before you start. We have a lot of work to do in pretty short space of time now, with new competitive opportunities in the very near future. To be honest, the line detection has simply not been good enough, and this is leading to some of the poor performance I have been experiencing with the car.  Whilst it really does work, it is not accurate enough.  There are a couple of issues here which need to be dealt with: The core processor is just not powerful enough; and The optics aren't good enough Processing As I noted in my "Next Iteration" article I am looking at othe
Read more