Feature Request: ROCm/Radeon Open Compute Support for AMD GPUs

[patmansk]

Problem

Meshroom currently only supports NVIDIA GPUs via CUDA, which excludes users with AMD graphics cards. While AMD GPUs offer excellent performance-to-price ratios and Open Source alternatives like ROCm (Radeon Open Compute), there is no official support.

Current Status

• Meshroom is built on AliceVision, which uses CUDA exclusively
• Community workarounds like ZLUDA exist but are unmaintained and incompatible with recent versions
• MeshroomCL exists as an OpenCL alternative but is not officially supported
• The core team has indicated that porting from CUDA would require significant rework [1]

Request

Add official support for AMD GPUs through ROCm/HIP (Heterogeneous-Compute Interface for Portability). HIP allows code porting from CUDA to AMD GPUs with minimal changes, making it a practical solution.

Benefits

1. Broader user base - Support AMD users who prefer open-source solutions or have budget constraints
2. Hardware diversity - Reduce dependency on NVIDIA hardware
3. Future-proofing - ROCm is actively developed by AMD
4. Performance - Modern AMD GPUs (RX 7000 series, MI series) offer competitive performance

Technical Approach

• Use AMD's HIP toolkit to port CUDA kernels
• HIP provides source-to-source translation tools (hipify) for CUDA code
• This would be less invasive than a full OpenCL rewrite
• Could potentially maintain compatibility with CUDA builds

Reference Implementation

Similar projects have successfully ported CUDA applications to HIP:

• [Request] Remove CUDA dependency AliceVision#439 discusses CUDA vs OpenCL/ROCm
• Discussion AMD vs Nvidia #2571 mentions ZLUDA workarounds

Additional Context

From the Meshroom community:

"I don't understand why Meshroom chooses to be open source, but use a closed/proprietary technology like CUDA"

"There are many of us in the world with AMD cards who want to use Meshroom"

Suggested Next Steps

1. Assess feasibility of HIP port for AliceVision core
2. Create proof-of-concept for key CUDA kernels
3. Document technical requirements and estimated effort
4. Consider community collaboration for development

Related Issues

• Issue No AMD GPU support? [question] #2140: No AMD GPU support
• Discussion AMD vs Nvidia #2571: AMD vs Nvidia
• Discussion Meshroom CUDA alternative #1283: Meshroom CUDA alternative

---

This feature would significantly expand Meshroom's accessibility and align better with its open-source philosophy.

[philippremy]

There is already an official SYCL port upstream (implemented with AdaptiveCpp). It targets the generic offloading JIT backend, which supports ROCm natively. I highly doubt that AliceVision will get another DepthMap backend if it can be avoided.
The reason most users are not able to use it is because there has not been an official release since upstreaming the SYCL backend.