Skip to content

COMMUNITY.md

Latest commit

 

History

History
45 lines (35 loc) · 5.95 KB

COMMUNITY.md

File metadata and controls

45 lines (35 loc) · 5.95 KB
Raw

What SkillFishOS takes from the BC-250 community, and what it does not

The BC-250 scene moves fast. This page is the ledger: every community item we looked at in September 2026, where it comes from, whether it is in the distribution, and why not when it is not. Dates are when we checked.

In the distribution

Item Source License Where it lives Checked
Compute queues (4 ACE) in kernel and RADV DryhoppedIPA/bc250-gfx1013-fix MIT kernel patches, skillfish-mesa-gfx1013 +4.2% Cyberpunk 1080p (05/09)
8-core SMU telemetry layout community research, own patch GPL-2.0 kernel patch 0006 8 cores, correct clocks
SDMA microcode Debian's firmware-amd-graphics already loads the working 0x34 build kernel SDMA0 firmware version 0x34
C-states and P-states for all 16 threads bc250-collective/bc250-acpi-fix, rebuilt by e-tho MIT /usr/share/skillfish/acpi, early initrd via GRUB POLL C1 C2 C3, 8 steps 800–3200 MHz
ACPI stubs (APTS, AWAK, AFN7 as no-ops) e-tho/bc250-acpi-fix MIT SSDT-STUBS in the same cpio added 11/09
Unified memory heap in RADV Necrosiak/bc250-tweaks /usr/share/drirc.d/50-skillfish.conf no fps change, fewer OOM
Shader cache cap 10 GB same /etc/environment.d/50-skillfish-mesa.conf
PipeWire quantum 512 @ 48 kHz same /etc/pipewire/pipewire.conf.d/50-skillfish-latency.conf
8-core unlock before the bootloader Hexxeh/bc250-efi-core-unlock MIT /usr/share/skillfish/coreunlock, skillfish-coreunlock-efi cold boot on the dev board (BIOS P3.00 with chipset menu, smart-plug power cycle, 11/09): one OS boot, 16 threads, mask 0xFF already there when the in-OS service looked; ssh usable 52 s after power-on against 154 s + a second boot before
The program is unsigned: with Secure Boot enforcing, skillfish-coreunlock-efi refuses to install it and the in-OS service keeps doing the unlock (two boots). The BC-250 firmware has no Secure Boot. LUKS is untouched: the program runs before GRUB and never reads a disk; TPM-bound unlocking (which we do not use) would just see one more measured binary, the same one on every boot
DualSense behind a DS5 bridge rpf16rj/bc250-steamos-real-toolkit GPL-2.0 (kernel) kernel patch 0017 untested: no bridge here
YCbCr 4:4:4 + deep color + HDMI 2.1 FRL on PCON dongles same GPL-2.0 (kernel) kernel patch 0018, off by default (amdgpu.force_ycbcr444=1 force_min_bpc=10 dcfeaturemask=0x402) ported to 7.2, untested: no CH7218 dongle here
GE-Proton GloriousEggroll Control Center, Games
Dolby Digital 5.1 over HDMI/DP (live AC-3 encode) MastaG/bc250-dual-audio none in the repository, asked in bc250-dual-audio#1; shipped at the maintainer's decision with credit skillfish-audio-dolby, pinned to WirePlumber 0.5.17 the first install took over the only audio device: the monitor was in standby, no native sink, WirePlumber fell back to the AC-3 frontend and the monitor, once awake, played the bitstream as PCM (loud noise). Fixed with two changes of ours (see CHANGES.SkillFishOS in the package): AC-3 only on an explicit choice, encoded sinks below any real one. Proven with the native sink removed and restored
Gamepad navigation, CSV export ideas from movacx/bc250-control-center Control Center

| FSR 4 INT8 lowering for gfx1013 (V3) | dmorazasanchez/bc250-fsr4 | MIT (added 2026-09-11) | skillfish-mesa-gfx1013 26.09.2 | +12% with FSR 4 on, nothing changes with it off |

Tested here, not shipped

Item Source Why not Result
Native mesh shaders on gfx1013 lonewolf0622/BC250-Native-Mesh-Shaders- written for Mesa main, does not apply to 26.2.2; and it hangs ported to Mesa main (27 hunks by hand) on top of our compute-queue fix: meshShader = true, vkcube fine, but the first vkCmdDrawMeshTasksEXT hangs the gfx ring beyond recovery (MODE1 reset fails, board rebooted). Whether the culprit is the patch or its combination with the open compute queues is not yet separated

Measured, no change needed

  • XeSS instead of FSR 3 (community: XeSS Balanced 78 > FSR2/3 75 > FSR4 73 fps in Cyberpunk with ray tracing). Cyberpunk ships XeSS, so we measured the game's own XeSS on RADV, 1080p, Balanced against Balanced, alternating pairs on the dev board (11/09): FSR 3 97.6 / 97.9 fps, XeSS 97.7 / 97.9 fps. Identical at our settings (the frame time is not in the upscaler at 1080p without RT); the community gain shows with ray tracing, where the base frame rate is far lower. The Games page says so in the FSR 4 card.

Looked at, nothing to do

  • VCN (daveconde/bc250-vcn-enable): powered on through the SMU, but the register file reads all ones and the firmware must be signed. Same wall we hit.
  • ROCm (akandr/bc250-rocm): runs, Vulkan stays twice as fast at prompt processing. Unsloth Studio on Vulkan remains our AI engine.
  • llama.cpp Vulkan tuning for gfx1013 (TechMakesArt/llama.cpp-bc250): +48% tokens/s. Parked until we move the AI section off Unsloth.
  • Secondary power limit above 1850–2200 MHz that locks the board hard: hardware (shunt mod), not software. It is the most likely cause of our hangs at 2230 MHz.
  • fix-freq / fix-metrics governor options: our governor reads the clock from the SMN directly.