Skip to content

Production I: 117 fps at 3x, three simulators at once; Coherent on by default on Arrow Lake - #7

Merged
0xdaniell merged 1 commit into
mainfrom
prod-i-build
Nov 10, 2026
Merged

0xdaniell merged 1 commit into
mainfrom
prod-i-build

Conversation

@0xdaniell

Copy link
Copy Markdown
Owner

macbox build now makes the build that runs every day on the reference Arrow Lake iGPU ("production I", since 2026-11-09).

Numbers (iPhone 17 Pro at 3x, GlassBench, Intel Core Ultra 7 265)

before now
one 3x simulator 50.6 fps (70.0 with Coherent, 2026-11-05) 117 fps (114–121; the screen's limit is 120)
two at once n/m 51–54 each
three at once n/m 31–32 each
desktop drag at 3440x1440 240 ticks/s alone, 126 with a simulator

What changed

  • Patches: metal2vulkan 34 → 92, Reims 25 → 33, QEMU 9 → 13. Applied the way container/build-reims.sh applies them, on fresh checkouts of the pinned commits, all three series match the production sources with zero diff (Reims without its local build wiring).
  • macbox: MACBOX_COHERENT=auto (new default) turns Coherent on with an Arrow Lake iGPU, together with the switches measured alongside it: grouped completion, staged page-ups, inline_checks, inline_lazy and the surface fixes. macbox rewrites only its own live.conf lines.
  • README, docs, site, llms.txt: a sharper pitch up front, the new numbers and chart, a 3x frame split into real rendering vs translation overhead, what was tried without gain, and the one GPU hang still open (repeated 3-simulator boots).

Checks

  • Patch series reproduce production: m2v 7fee47f5, QEMU 3999026c68 (exact), Reims 8dea1e59 (except the local Cargo path and committed core dumps, as before).
  • bash -n macbox; apply_tuning tested in a sandbox for auto / on / off / tuning off.
  • python3 site/build.py passes.
  • Not run: a full macbox build from scratch (20–60 min). The sources are identical to the running production build.

… default on Arrow Lake

macbox build now makes the build that has run day to day since 2026-11-09 ("production I"):
- patches/metal2vulkan: 92 patches on 0db5825a (was 34), rebuilding m2v 7fee47f5 exactly:
  grouped completion, inline_lazy, inline_checks, the surface fixes the lanes and the census
  found (make_once, reader_first, own_back, open_ids, max_texels), lab switches (all off) and
  diagnostics (off unless set).
- patches/reims-vgpu: 33 patches on 7198fe6 (was 25), Reims 8dea1e59 without its local build
  wiring: staged page-ups, REIMS_VGPU_GROUPED, FIFO exec pipelining, interrupt moderation.
- patches/qemu: 13 patches on bd95509 (was 9), QEMU 3999026c68: KVM irqfd, ioeventfd and
  halt-poll properties for the GPU device (off unless set).
Each series was applied as container/build-reims.sh applies it, on fresh checkouts of the
pinned commits, and compared with the production sources: no difference.

macbox: MACBOX_COHERENT defaults to auto, on with an Arrow Lake iGPU (where it runs every day),
and turns on the switches measured with Coherent: the live.conf lines, REIMS_VGPU_GROUPED and the
file-named Reims switches in $MACBOX_HOME/run/switches. macbox rewrites only its own live.conf
lines, so a mode change applies and the user's lines stay.

README, docs, website and llms.txt: a clearer pitch up front, the new numbers (3x simulator
17.2 upstream -> 117 fps; 2 at once 51-54 each, 3 at once 31-32 each; desktop drag 240 ticks/s
alone, 126 with a simulator), a 3x frame broken into real rendering and what translation adds,
what was tried without gain since, and the open GPU hang under repeated 3-simulator boots.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Nov 10, 2026 •

Copy link
Copy Markdown
Codex Review Summary: ✅ Completed

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ✅ Completed 2026-11-10T08:00:05.085785Z 2b4f608 PR opened
ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

@0xdaniell
0xdaniell merged commit b4574e9 into main Nov 10, 2026
2 checks passed
@0xdaniell
0xdaniell deleted the prod-i-build branch November 10, 2026 07:57

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 2b4f60867b

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Comment on lines +81 to +83
+/// LOCAL (Atlas): the command buffer `id` went without a commit: it holds no standing open.
+pub(crate) fn dropped(id: u64) {
+ open_ids().lock().unwrap().remove(&id);

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Badge Invalidate standings when an uncommitted command buffer is dropped

With surface_lazy_open_ids=1 (now enabled by macbox with Coherent), dropping an uncommitted command buffer removes its ID from open_ids, which makes was_committed(id) return true. However, dropped does not invalidate the standing that its encoder recorded in Lazy::stand. A subsequent command buffer using the same surface can therefore accept that standing in take_standing and skip the required copy-in, even though the abandoned buffer's commands never populated the image. This can produce stale or incorrect pixels. Invalidate standings belonging to abandoned buffers, or distinguish abandonment from successful commitment rather than treating both as absence from the open set.

Useful? React with 👍 / 👎.

0xdaniell added a commit that referenced this pull request Nov 10, 2026
…ndings with it (#10)

With surface_lazy_open_ids (on with Coherent), a command buffer dropped
without a commit left the open set, so the standing its encoders left read
as committed, and a later encoder could skip a copy in it needed and draw
stale pixels. Found by the Codex review of #7.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant