General purpose GPU accelerated parallel programming

[spamegg1]

I looked around but could not find an easy way to take advantage of the GPU in Scala. There is the Parallel Collections library but that runs on the CPU as far as I know.

Is it possible to use Storch for non-deep learning, non-numerical, more general parallel computation purposes? Like using the GPU to run a parallel blur filter that can run thousands of times faster than on the CPU, for example. This would be fantastic for educational purposes.

Are there any other "general parallel programming on the GPU" libraries for Scala that are easy to use and "teachable"?

[hmf]

I will defer this to @sbrunk, but to my knowledge I would say no because storch "just" acts as a wrapper to the JavaCCP wrapper of the Pytorch library. I am assuming here that one would need to program the kernels directly for such an exercise. A quick search also shows image blur is only available in PytorchVision. I also did not find any JavaCPP preset for torchvision, but I may have missed it.

[spamegg1]

Immediately hit a bug! 🤣 It is very much a work in progress indeed.

I think it won't work anyway; it requires OpenCL >= 2.1 which needs different hardware than mine (1.1).

Well, at least it built successfully. I'll try some Java code. Then figure out how to do it in Scala!

[spamegg1]

Maybe it will work? 🤔

╰─ tornado --devices                                                                                                 ─╯
WARNING: Using incubator modules: jdk.incubator.vector

Number of Tornado drivers: 1
Driver: OpenCL
  Total number of OpenCL devices  : 2
  Tornado device=0:0  (DEFAULT)
	OPENCL --  [Clover] -- AMD Radeon Vega 11 Graphics (raven, LLVM 15.0.7, DRM 3.54, 6.5.0-28-generic)
		Global Memory Size: 6.8 GB
		Local Memory Size: 64.0 KB
		Workgroup Dimensions: 3
		Total Number of Block Threads: [256]
		Max WorkGroup Configuration: [256, 256, 256]
		Device OpenCL C version: OpenCL C 1.1

  Tornado device=0:1
	OPENCL --  [Portable Computing Language] -- pthread-AMD Ryzen 5 3400G with Radeon Vega Graphics
		Global Memory Size: 11.6 GB
		Local Memory Size: 512.0 KB
		Workgroup Dimensions: 3
		Total Number of Block Threads: [4096]
		Max WorkGroup Configuration: [4096, 4096, 4096]
		Device OpenCL C version: OpenCL C 1.2 pocl

[sbrunk]

PyTorch also supports AMD GPUs via rocm BTW: https://pytorch.org/blog/pytorch-for-amd-rocm-platform-now-available-as-python-package/

I don't know how well APUs are supported though. Might be worth a shot trying to install the upstream PyTorch rocm variant via pip and test if it's working on your hardware.

[spamegg1]

Tornado doesn't work unfortunately 😢
Running the examples just stalls with 0% CPU usage.
Needs newer hardware I think!
(Although the guy in the video runs it on a laptop. It's Nvidia... 😠 )

@sbrunk rocm is only supported on Instinct GPUs.

I think I give up on this dream 😢
GPU compute is reserved for the rich, no low-end peasants allowed.

[hmf]

Wouldn't hurt to ask in the TVM slack channel if the APU is supported. Quick search didn't not show anything.

It just occurred to me that TVM may also be a way for storch to use GPU for example when loading and augmenting images.

[spamegg1]

I found this! https://github.com/ComputeNode/scalag

Did any of you know about this? Seems promising!

Says SPIR-V and run through Vulkan, not OpenCL or CUDA. That can work on APUs, hopefully!

It's pre-alpha so I'll wait.

[hmf]

Not me. Seems interesting. And that image does not seem to be just ray-tracing. Sift shadows reminds me of the [Radiosity](https://en.wikipedia.org/wiki/Radiosity_(computer_graphics).