#566 Proposal: extend existing calibration to Linux HIP

描述

Would you be open to a small contribution extending the existing calibration flow to Linux HIP?

In the revision I reviewed (db4f91a), setup excludes HIP from the calibration offer and saved-profile reuse, while the calibration hardware key uses NVIDIA-oriented GPU discovery.

I propose starting with:

Correct HIP device identification and calibration profile matching.
Would you be open to a small contribution extending the existing calibration flow to Linux HIP?

In the revision I reviewed (db4f91a), setup excludes HIP from the calibration offer and saved-profile reuse, while the calibration hardware key uses NVIDIA-oriented GPU discovery.

I propose starting with:

- Correct HIP device identification and calibration profile matching.
- The existing CPU worker count, PCIe work-share and speculative-confidence controls, after verifying each works on HIP
- Preserving the current baseline and acceptance conventions, with configuration recovery if calibration fails.
- Integration tests and repeated measurements on an RX 7900 XTX, documenting the actual hardware, software and workload.

The goal is to let AMD users tune the existing execution settings for their CPU/GPU combination. I would build on tools/calibrate.py and the current setup flow.

Broader workload tuning and RDNA3 matrix-core attention could follow as separate contributions once there is supporting evidence.

Does this scope fit your direction, and are there existing changes or interface preferences I should coordinate with before implementing it?