Hi everyone,
I've managed to get nvdiffrast working on AMD GPUs with ROCm 7.1 (specifically tested on gfx1201 / RDNA 3/4).
The main challenges were the Wave64 architecture requirement for 64-bit lane masks and porting NVIDIA-specific PTX assembly to HIP intrinsics. I've created a comprehensive patch script that automates all the necessary code changes (ASM porting, Wave64 masks, and PyTorch binding fixes).
You can find the patch and instructions here:
https://github.com/tashibi/nvdiffrast-rocm-patch.git
Hope this helps other AMD users running InstantMesh, TRELLIS, or other 3D Gaussian Splatting projects!
Hi everyone,
I've managed to get
nvdiffrastworking on AMD GPUs with ROCm 7.1 (specifically tested on gfx1201 / RDNA 3/4).The main challenges were the Wave64 architecture requirement for 64-bit lane masks and porting NVIDIA-specific PTX assembly to HIP intrinsics. I've created a comprehensive patch script that automates all the necessary code changes (ASM porting, Wave64 masks, and PyTorch binding fixes).
You can find the patch and instructions here:
https://github.com/tashibi/nvdiffrast-rocm-patch.git
Hope this helps other AMD users running InstantMesh, TRELLIS, or other 3D Gaussian Splatting projects!