Skip to content

CONTRIBUTING.md

Latest commit

 

History

History
50 lines (43 loc) · 3.03 KB

CONTRIBUTING.md

File metadata and controls

50 lines (43 loc) · 3.03 KB
Raw

Contributing

This repo is meant to be a living, reproducible baseline for the Allwinner A733 PowerVR + FEX stack. Test it, benchmark it, push the walls, fix bugs — and send the findings back so the next person starts further ahead.

Ways to contribute (all welcome)

  • Run the benchmarks (bench/) on your board and add your numbers to RESULTS.md — cross-board data is the whole point. Even "same board, different kernel/DDK" rows are valuable.
  • Reproduce or challenge a finding in docs/FINDINGS.md. If something that's marked a "wall" works for you (newer DDK, mainline, a flag we missed) — that's a great PR/issue. Findings are dated observations, not gospel.
  • Optimize: faster FEX (AOT/code cache), better Zink coverage, a working Wayland-client path, GPU desktop improvements, encode work.
  • Fix bugs in the scripts/patch, or port the kernel patch to a new BSP/kernel.
  • Extend: new device-tree overlays, other A733 boards (A7S/A7Z), mainline tracking.

Benchmark discipline (required for perf changes)

Any change that could affect performance must carry a before→after bench/run.sh diff in its commit message, and bench/baseline.txt is updated only for verified, intended improvements. Compare the ratio.* + gpu.* lines (reliable); raw absolutes are ±~20% load-noise on an always-on box. See bench/WORKFLOW.md.

How to submit

  1. Fork → branch → PR. Keep PRs focused; describe what you tested and on what (board, kernel uname -r, DDK strings /usr/lib/libVK_IMG.so* | grep -m1 24.).
  2. For results, edit RESULTS.md or open a Benchmark result issue.
  3. For "works/doesn't on my board", open a Board test report issue.

Ground rules

  • No proprietary blobs in commits — never commit the PowerVR userspace/firmware, the FEX rootfs, or third-party app binaries. The .gitignore blocks the obvious ones; the install/recipe model fetches those from the vendor (see README).
  • No secrets — keys, tokens, host identifiers.
  • Patches against vendor sources should be diffs, applied on top of the vendor's own distribution (don't paste their source in).
  • Be honest about methodology: single-board numbers are indicative; say how you measured.

Open problems / help wanted

These are the live walls from docs/FINDINGS.md — concrete places to dig:

  • Wayland GPU clients crash (zink kopper SIGSEGV on the img Vulkan ICD). The compositor GPU-composites; client apps presenting via wl_egl_window don't.
  • Transparent EGL→GPU (no PowerVR GLVND vendor; vendor EGL is X11-only fork).
  • FEX cold-start is JIT-compile-bound — an AOT/persistent code cache is the lever.
  • Steam under FEX — CEF steamwebhelper bwrap/pressure-vessel blocker.
  • Mainline tracking — drm/imagination + Mesa pvr for A733 (sun60iw2); the upstream DTS effort is the path that lifts the hard walls. Test/track it.
  • HEVC hardware encode — vendor-lib limit (no IDR); H.264 encode works.