ComputeShader

[jimbok8]

Very interesting to see what speeds you get using the GPU in the PixelBufferGPUTest.
I am interested in display point clouds (from Lidar)
I have been trying to understand the techniques use by Markus Schuetz:

"Software Rasterization of 2 Billion Points in Real-Time"
In branch compute_rasterizer_2022
paper - video

"Rendering Point Clouds with Compute Shaders and Vertex Order Optimization"
In branch compute_rasterizer_2021

His results look very impressive - if only I did not dislike/not-understand C++ so much!
The key to his results seems to be Compute Shaders - not that I really understand them yet!
I see you can use compute shaders in Java/Jogl (see https://github.com/perses-games/jogl-compute-shaders-fireworks)

Is what you are doing with Aparapi roughly equivalent to the Compute Shader approach?
If not, do you think there would be any benefit in combining the two approaches?

Thanks
Jim

[AlmasB]

Hi Jim,

GPGPU is a bit outside of my domain. Using the Aparapi library was the quickest solution I could put together to leverage GPGPU in Java.

Generally speaking, both approaches sound similar / identical in theory and provide the ability to execute some arbitrary code via GPU cores. It is worth noting that neither are a quick way of gaining performance. A lot depends on the actual bottleneck that your use case has. If the bottleneck is an algorithm, then there needs to be a way to rephrase the algorithm in a way that aligns with parallel processing: perform a lot of simple operations in parallel. Otherwise, there are no performance gains. If the bottleneck is elsewhere (like the speed with which you can draw to the screen), then again GPGPU is not as helpful.