Page 1 of 1

AtomicAdd 300% faster than simple store on Pascal GPUs!

Posted: Tue Jul 03, 2018 11:53 pm
by devsh
I have a very peculiar request as I'm getting very peculiar results (atomic to the same address by all threads faster than a simple write), could anyone on Linux with a GPU supporting OpenGL 4.3+ help me with a benchmark?

It's already compiled for Ubuntu 17.04 (will run on higher Ubu or Debian).
I just need it ran in the terminal (command-line) with parameters 1920,1080,MODE,n when std::cin asks for them.

MODE will be a number from 0 to 7 inclusive

Press on the window (focus on it) and hit Q to stop the benchmark.

If you could let me know what numbers you get I would be much obliged!
(some tests won't run if you don't have NVidia)

https://github.com/buildaworldnet/Irrli ... Benchmark/ (everything in one folder)

P.S. I'm working on a Windows Visual Studio project and a compute shader-based benchmark.

Re: Help Me with a Benchmark Please...

Posted: Wed Jul 04, 2018 3:53 pm
by devsh
I actually figured out why this strange behaviour happens.

I wrote a wiki page on it
https://github.com/buildaworldnet/Irrli ... hink-Again!

If you are interested in atomics on GPUs work, how shaders get scheduled and how this all works in hardware and what the NVIDIA shader compiler does behind the scenes... give it a read.