• Home
  • Features
  • Pricing
  • Docs
  • Announcements
  • Sign In

super3 / llmjob / 43658527618
100%

Build:
DEFAULT BRANCH: main
Ran 15 Nov 2026 12:31AM UTC
Jobs 1
Files 27
Run time 1min
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst

15 Nov 2026 12:30AM UTC coverage: 100.0%. Remained the same
43658527618

push

github

web-flow
Release v0.7.2 (#296)

* RTX 40: build the 96 x 512 block in the probe and race it; it loses to the shipped loop

The last round priced a block where all eight warps share A as the one layout with fewer shared-memory steps.
This builds it in the probe (gprobe -DGP_W8=1, rfold.py w8=1) and it is correct, but each barrier of all eight
warps stops every scheduler and costs about 0.8 points of pipe. At 810 MHz memory the card already sits near its
top clock, so the 2.4% energy it saves buys no clock: 97.3% pipe x 2685 MHz against the shipped loop's 99.1% x
2670. Group barriers (w8g) get to 98.8% but are still slower, and the fastest form still gets a few tiles wrong.

All new keys are off by default; the shipped cubin rebuilds to its listed hash (3a354cac). The miner side of the
new tile shape is left out of the repo since nothing ships. benchmark.md records the numbers and two barrier
findings (only barriers 0-7 work in a patched cubin; set a barrier's predicate well ahead of it).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018A6hR4v2YEpsJuAoZzfi3r

* Start the work after v0.7.0

An empty commit, so the branch can open its PR. The work continues the
goal of beating the best competing miner on every AMD and RTX 20-50 card.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018A6hR4v2YEpsJuAoZzfi3r

* RTX 50 w8m: timing keys to find its lost point of pipe, and the miner's side of the 96 x 512 block

w8m is correct on a 575 W RTX 5090 (16,384 of 16,384 at k 2048, 8,192 at 8192, 512 at 65536) but holds 98.3% of
the pipe against the shipped loop's 99.25%, so its +1% of clock only makes it level. w8mx leaves out the waits
or arrives one kind at a time (timing only); w8mpl, w8mbs and w8mt move or shorten the waits.

The miner can now carry the block: goal-nvk22's host patch (the 96 x 512 tile read from the cubin's rot_w8
symbol), pearl_rot.cu ROT_W8M... (continued)

1023 of 1023 branches covered (100.0%)

Branch coverage included in aggregate %.

1590 of 1590 relevant lines covered (100.0%)

31.3 hits per line

Jobs
ID Job ID Ran Files Coverage
1 43658527618.1 15 Nov 2026 12:31AM UTC 27
100.0
GitHub Action Run
Source Files on build 43658527618
  • Tree
  • List 27
  • Changed 0
  • Source Changed 0
  • Coverage Changed 0
Coverage ∆ File Lines Relevant Covered Missed Hits/Line Branch Hits Branch Misses
  • Back to Repo
  • Github Actions Build #43658527618
  • 9c1adc6b on github
  • Prev Build on main (#43537176152)
  • Next Build on main (#43811315449)
STATUS · Troubleshooting · Open an Issue · Sales · Support · CAREERS · ENTERPRISE · START FREE TRIAL · SCHEDULE DEMO
ANNOUNCEMENTS · TWITTER · TOS & SLA · Supported CI Services · What's a CI service? · Automated Testing

© 2026 Coveralls, Inc