Beginner to GPU compute programming

[porina]

I have an idea for a GPU compute task. The difficulty lies in that I'm totally out of date with modern programming options, and how a GPU actually looks from a programming perspective. The only language I've used in anger is vanilla C. No ++. No #. I have looked looked at game engine programming and it feels like I'm missing a lot from modern languages, which if anything, should make things a lot easier once I get used to it.

 

So, what are the easiest options to get started? Any pointers welcome, as long as they're not like pointers in C.

• Windows environment
• Only needs to run on nvidia GPUs/CUDA.
• I don't mind learning another language beyond C. It looks like CUDA can work with C, C++, Python and more, but I think those will be the main possibilities for me. Learning Python sounds more attractive to me than remembering how to use C. 
• Application will be console/command line only. I don't need/want to worry about a GUI.

[igormp]

You could go with numba + cuda:

https://numba.readthedocs.io/en/stable/cuda/index.html

[Raytsou]

The limited experience I have around GPU programming is through opencv, which is a computer vision / linear algebra package. You can construct cv::cuda::GpuMat objects to create matrices in vram, and operations on these objects will occur on GPU.

[Franck]

CUDA is made for C or C++. There was another competition but it died a while ago called AleaGPU which was just the CUDA libs wrapped for C# so that it could be used in normal day to day business app, website, etc. It was hella fast specially for large data and looping functions. The team unfortunately stopped updating the core so it lost support for newer cards.

 

I never tried the C version but I have used successfully the C++ and C# quite a lot.

[igormp]

3 hours ago, Franck said:

CUDA is made for C or C++. There was another competition but it died a while ago called AleaGPU which was just the CUDA libs wrapped for C# so that it could be used in normal day to day business app, website, etc. It was hella fast specially for large data and looping functions. The team unfortunately stopped updating the core so it lost support for newer cards.

 

I never tried the C version but I have used successfully the C++ and C# quite a lot.

Nvidia has their own Python wrapper and endorse libraries such as Numba:

https://developer.nvidia.com/cuda-python

 

So using C, C++, Fortran or Python is pretty much a personal decision, CUDA itself should work with any of those without problems.

[Timme]

On 7/4/2023 at 6:06 PM, igormp said:

Nvidia has their own Python wrapper and endorse libraries such as Numba:

https://developer.nvidia.com/cuda-python

 

So using C, C++, Fortran or Python is pretty much a personal decision, CUDA itself should work with any of those without problems.

Python for low-level hardware programming? Isn't it a bit too abstract for proper learning and development in the nitty-gritty parts of the subject?

[riklaunim]

4 hours ago, Timme said:

Python for low-level hardware programming?

It's just a glue while you code CUDA kernels separately, the same with other aspects of GPU computing.

[igormp]

4 hours ago, Timme said:

Python for low-level hardware programming? Isn't it a bit too abstract for proper learning and development in the nitty-gritty parts of the subject?

Well, OP wants to accomplish a task, not necessarily learn how to do proper gpu programming nor do low level hardware programming (by at which point you could argue that only asm/shaders apply). 

25 minutes ago, riklaunim said:

It's just a glue while you code CUDA kernels separately, the same with other aspects of GPU computing.

You don't need to hand write those Cuda kernels tho. 

[Timme]

3 minutes ago, igormp said:

Well, OP wants to accomplish a task, not necessarily learn how to do proper gpu programming nor do low level hardware programming (by at which point you could argue that only asm/shaders apply).

You don't need to hand write those Cuda kernels tho.

Well, then I guess to learn game dev I can just learn how to Unity suite? Who cares about optimizations and actually understanding what you're doing, right? Why not just use ChatGPT as a side tool? Because why not, it'll work. Somehow.

I wasn't talking about drivers, but rather about all that stuff that happens under the hood. And C is quite enough.   

[porina]

Looks like I forgot to reply to this thread. I'm still reading up and getting used to the concepts involved, which there are more of than I thought initially. The understanding required to interface to the GPU doesn't seem any different if I use some version of C or Python, but my gut feeling is Python is more friendly as a modern language. Not having to waste so much time on low level stuff is a big plus.

[Johann-Tobias xn--schg-noa]

I would consider having a look at the Julia programming. It has quite decent support for NVIDIA, Intel and AMD GPUs and is used a lot for scientific computing. The tooling in language expose a bunch of different programming paradigms and is generally nice to work with. Managing software dependencies is easier then in other languages. One disadvantage is that in Julia if a new function is called for the first time a lot of code gets compiled. Moving the compile time so close to run time gives you a lot of flexibility but it is painful for doing something once and then closing the program.

[porina]

11 minutes ago, Johann-Tobias xn--schg-noa said:

I would consider having a look at the Julia programming.

I haven't even heard of that one. In a quick skim it sounds like it has interesting features. A reservation is I may be less likely to find help than for more well known languages.

[Johann-Tobias xn--schg-noa]

3 minutes ago, porina said:

A reservation is I may be less likely to find help than for more well known languages.

That is a good observation and valid concern. One thing i would hold against that is since you compile at run time and all the different GPU targets are just LLVM back-ends you actually get better insight what is going on. You can dump the assembly of the GPU function you just ran and ended up being slower in the REPL and compare. Julia is rough around the edges but all the cool stuff in Julia are build in Julia itself and since Julia doesn't ship as binary you get deep insights how stuff happens and works should you want to or need to.

[Franck]

On 7/29/2023 at 5:06 PM, porina said:

Python is more friendly as a modern language.

Python is not a modern language. I stopped using it about 25 years ago

[porina]

1 hour ago, Franck said:

Python is not a modern language. I stopped using it about 25 years ago

I'm coming from C. It's about 20 years younger.

[iggy12345]

On 7/31/2023 at 5:45 AM, Franck said:

Python is not a modern language. I stopped using it about 25 years ago

Python2 is not a modern language, Python3+ is definitely a modern language

[The Hope]

Obsidian: GPU Programming in Haskell

https://svenssonjoel.github.io/writing/dccpaper_obsidian.pdf

 

GPU programming in Haskell

https://bobkonf.de/2015/slides/thielemann.pdf

 

Parallel and Concurrent Programming in Haskell

https://www.oreilly.com/library/view/parallel-and-concurrent/9781449335939/

 

If you were to ask yourself the question why Haskell and not Python:

tornado=Python and warp=Haskell

 

Another fact is that good Haskell programmers are on average 2.5 times more productive than the best Python programmers.

 

Another great Python alternative:

https://juliagpu.org/

https://blogs.oracle.com/javamagazine/post/programming-the-gpu-in-java

https://github.com/takagi/cl-cuda