Release: Community port of tiny-cuda-nn and gsplat for RDNA4 (gfx1201) / ROCm 7.2 and more...

[Painter3000]

Hello ROCm Community,

I wanted to share a milestone regarding neural graphics pipeline support on RDNA4 hardware that I've been working on.
I have published a complete standalone porting effort, validating a fused MLP and rasterizer pipeline on the AMD Radeon AI PRO R9700 under Ubuntu 24.04 and ROCm 7.2.
My stack addresses several known hurdles for RDNA4:

• tiny-rdna4-nn: Implements a dedicated rocWMMA Width-64 fused MLP path to bypass generic upstream configurations.
• amd-gsplat: Resolves Wave32 execution requirements inside a custom port (based on 1.5.3b2).
• nvdiffrast & SuGaR: Fixes runtime compatibility and validation on gfx1201 for differentiable rendering and mesh extraction.

The Phase-4A2 milestone validates real forward, backward, optimization, replay, resume, and 100-step execution purely on the AMD ROCm image path.
For anyone looking to benchmark or deploy 3DGS/NeRF workloads natively on newer AMD architectures, the codebases and reproducibility reports are fully public on my profile:
https://github.com/Painter3000
Feedback, issue reports, or contributions from fellow RDNA4 owners are highly welcome!