WIP: perf(vk): RX_VK_RT_HOME=1 keeps the render thread on its own core on Windows, with a guard for another program holding it (off by default) #2240

Draft
klimovich008 wants to merge 14 commits from klimovich008/Longshot:perf/rt-home-v3 into main
Member

Summary

This brings back a switch that keeps the render thread on one core, after #1309 removed RX_VK_RT_STRICT_PIN. This time
it comes with a guard for #838's case (another program holding that core) and a farm plan. It is off by default, and
this PR proposes no default:
that is the team's call once the farm has run it.

What it does, in player words. On Windows, the renderer keeps one core free of the game's own threads for the render
thread (the "reserved core"). Today the thread prefers that core, but Windows may move it to any fast core. When it moves,
it often lands beside a busy game thread on the same physical core, and both run slower. RX_VK_RT_HOME=1 keeps the
render thread on its own core. When something else holds that core off it (#838's case: another program at high
priority on the same core), it lets the thread move again for a while:

  • 2 s at first, doubled for each hold within a minute, 30 s at most;
  • then it goes home again.

It does nothing on Linux, the Steam Deck or macOS, and on a CPU without SMT (two logical processors a core) it changes
nothing that matters.

The switch. RX_VK_RT_HOME (kind attempt):

  • unset (or 0): main's placement, exactly;
  • 1: home, with the guard;
  • strict: home with no guard. It is a measuring arm and the tests' negative control, never meant to be used.

RX_VK_RT_HOME_MS (kind tune, default 4) is the guard's threshold.

How the guard decides it is being held. Between two presents 20-250 ms apart, it looks at two things:

  1. whether the render thread lost RX_VK_RT_HOME_MS or more off the CPU outside its own waits. The time it spent ready
    but not woken from its job wait counts as lost;
  2. whether every processor of its core was taken by other threads for that long.

It acts only once it has seen a hold happen, so about one stall frame gets through each time the thread goes back
home
(measured below).

Read first: gpu/vk/rt_pin.h (the Guard), then in gpu/vk/vk_replay.cpp: rt_place, rt_home_frame, job_wait and
rt_ready_note. Also changed:

  • gpu/vk/vk_fps.cpp/.h: the render thread's waits by kind;
  • gpu/vk/switches.def: the two rows;
  • docs/dev/reference/threads.md;
  • the tests: tests/test_rt_pin.py (new) and tests/test_reserved_core_mask.py.

Testing

Tested: Windows 11, Ryzen 7 5700X3D (8 cores, 16 threads), RX 9070 XT, AMD 26.11.2. The PR's renderer DLL went over
main dd5ffe6a's build (DLL b627b1f4, exe 302120c6):

  • The standard test (Hyrule Field turning), RX_VK_RT_HOME=1 against unset, 3 alternating pairs: see Measurements.

    • In every 1 run the guard's lines show the thread at home through the measured window, with 0 holds.
    • The live affinity mask, read once per arm inside the window, was 0xc000 with 1 and 0xffff unset.
    • The longest wake from a job wait was 0.66-0.68 ms, against the 4 ms threshold, so ordinary play is never taken for a
      hold.
  • A stall made on purpose: a test program held both processors of the reserved core at priority 15 for 35 ms in every
    250 ms, beside the game, for three rounds of strict, 1 and unset in turn. Frames over 40 ms in the window:

    round strict 1 unset
    1 80 2 rejected (clock)
    2 81 2 0
    3 81 2 1
    • Every run of every arm has one long frame right at the end of the load. Unset's was 39.96 ms in round 2, just
      under 40, which is why that round counts 0.
    • 1's second frame in each run is the first stall after a return home, caught and held at that very present. The
      guard can only see a hold after it has happened, so one stall frame gets through per return home.
    • strict (no guard) let through 80-81, so the injection worked.
  • Unset against main's own build: the same mask and ideal processor, the same two placement lines, and no new error
    lines.

  • tests/test_rt_pin.py:

    • The placement for every layout, with unset checked against main's mask, ideal processor and lines.
    • The guard through the real queue code, cut from the source, with a fake clock and a fake mutex.
    • Going home, and the stall cases.
    • 47 broken variants, each of which must fail and does.
    • It passes with clang++ and g++ after merging main a2dc5385. One stub was added for main's new spawn_needed (#2170),
      with the PR's code unchanged.
  • Other tests:

    • test_reserved_core_mask, test_reserved_core, test_fps_stats, test_perf_pipe_waits and test_threaded_rendering
      pass with both compilers.
    • test_switch_table fails on RX_VK_GRASS_DISTANCE's order in switches.def, as on main.
    • The full run_all.py on dd5ffe6a, with both compilers: nothing fails on the branch that doesn't fail on main.
  • Linux syntax (clang 21, gcc 15): ok on vk_replay.cpp and vk_fps.cpp, after the merge.

  • The Windows build after the merge (llvm-mingw, the renderer and the game over main's base): it builds clean. It was
    not run in the game: the numbers below are from the PR's code before the merge.

  • The cost off: one bool test at each job hand-over (1.5 ns measured). With 1, a clock read on each side of a
    hand-over to the waiting thread, about 55 ns.

Still to test:

  • Windows PCs with SMT CPUs, 4, 6 and 8 cores, Intel with Hyper-Threading and AMD, hybrid Intel too: RX_VK_RT_HOME=1
    against unset on this PR's build, uncapped Hyrule Field. Expected: fps up or level, and no more long frames
    (@Beanstalk, request below).
  • Real 4-core, 8-thread CPUs in particular: on this PC held to 4 cores the gain was much larger (report-only, below)
    (@Beanstalk).
  • #838's real case (Intel's PresentMon service on an i5-13600K holding the core): with 1, about one stall frame per
    return home, not every stall (by hand).
  • SMT off: expected to change nothing (by hand).

Measurements

Verdict: on this PC, +6.7 % fps with RX_VK_RT_HOME=1 in the standard test, CPU 43.86 -> 41.31 ms and p99 15.9 ->
14.4 ms. Every pair gained (+8.4, +6.5, +6.8 %). A second session (the implementer's) gave +6.2 %.

Baseline: this PR's build with RX_VK_RT_HOME unset (main's placement), on main dd5ffe6a's build.

Setup:

  • Windows 11, Ryzen 7 5700X3D, RX 9070 XT, AMD 26.11.2. Hyrule Field turning, 16:9, RX_VK_ASYNC=1.
  • Uncapped: HOS_FPS_MAX=240 with the options file's cap 0 and vsync 0. At about 98 fps that is in effect the farm's
    frame_rate_cap=6, vsync=2.
  • A fixed light GPU load beside the game. 3 alternating pairs, no warm-ups, no replacements, no run rejected, medians.
Before: this PR @ 608775b2, RX_VK_RT_HOME unset After: this PR @ 608775b2, RX_VK_RT_HOME=1
fps, median (runs) 91.64 (89.99, 91.82, 91.64) 97.79 (97.59, 97.79, 97.84)
CPU ms 43.86 41.31
p99 frame time, ms 15.87 14.36
worst frame, ms (median run) 59.8 48.7
frames over 50 / 100 / 250 ms, per run 1/0/0, 1/0/0, 0/0/0 0/0/0, 0/0/0, 1/0/0
clock 1.000, 1.001, 1.000 1.000, 1.000, 1.000
the replay's submit, ms a frame 8.36 7.26
where the render thread ran 0xffff (may move) 0xc000 (its core), 0 holds
  • Report only, not the claim: one pair held to 4 cores with SMT (--affinity 0xff), which is a mask of an 8-core
    CPU, not a 4-core CPU. It gave 65.04 -> 82.34 fps (+26.6 %), p99 23.1 -> 19.1 ms, one run a side.
  • Measured on main dd5ffe6a. The PR is now on main a2dc5385. Main's 7 commits since then don't move the render thread
    or its queue wait. #2170 starts the thread through spawn_needed, and #2174 changes Hyrule Field's simulated turf on
    the way in, so the absolute numbers may move. Not re-measured here.

Risks

  • About one stall frame per return home gets through where another program holds the render thread's core (#838's
    case). The guard sees a hold only once it has happened. Unset doesn't have this.
  • A false hold costs one stay (2 s, at most 30 s, on main's placement), never the session. These limits can cause
    one:
    • Untracked GPU waits on the render thread: the present's frame-slot fence wait and the acquire after it are not
      marked as waits, so they count as lost time. Small wakes that add up to 4 ms in a window, with others busy on the
      core during such a fence wait, could make a false hold. Follow-up: wrap them in FpsRtWait(kRtGpu).
    • Two clocks: the wake is timed on steady_clock (QPC), the waits on fps_clock (the TSC, scaled once at its
      first call). A calibration preempted during its loop mis-scales the waits. Follow-up: compare fps_clock's rate with
      the guard's own 1 s calibration, and give up (rt_home_off) when they disagree by more than 2 %.
    • The time read: a delay (a DPC burst) between the idle-count read and __rdtsc() makes phantom busy time.
    • A notifier preempted while it still holds the queue's lock after stamping the wake, or other threads taking the
      core's two processors one after the other while the thread is held off for 4 ms or more. Each needs a wake of 4 ms
      or more. In ordinary play the longest was 0.68 ms.
  • Where it could cost: on 4 cores the reserved core is a quarter of the CPU. What keeping the render thread there does
    to the game's other three cores is measured only on the mask above.
  • Not touched: RX_VK_ASYNC=0's pin (ReplayPin), host/ and the shader-cache stamp files (no re-prepare). Hybrid
    CPUs never keep the thread on an efficiency core. A process affinity mask is respected (#1776).
  • An explicit RX_VK_RT_HOME=0 adds one start-up line naming the value. Unset gives exactly main's lines.
  • To turn it off: unset the switch. It is off by default, so a player notices nothing after updating.

Checklist

Ready for review

  • One topic, on current main; the title is a Conventional Commit.
  • Tested on my own PC, in the game for a change the game runs, as Testing says.
  • mise run test passes (any failure that also fails on main is named).
  • Measured against main, or Measurements says why there is nothing to measure.
  • No game files, build output or hand-edited generated files; third_party/ untouched; nothing GPL-incompatible.

When it applies

  • A switch added or changed: it ships at the default the measurements chose, with its row in gpu/vk/switches.def.
  • A check under tests/ that reads changed source is updated, not weakened.
  • Docs updated.
  • AI assistance recorded with Assisted-by: / Co-Authored-By: trailers.

Ready to merge

  • Everything in Still to test has run, its results in Tested.
  • Tier 1 hardware it can affect is tested (docs/dev/standards.md, "Hardware tiers").
  • Approved in review, every comment answered.
  • CI passes on the head; on current main, measured again or a comment says why the numbers still hold.

🤖 Generated with Claude Code

## Summary This brings back a switch that keeps the render thread on one core, after #1309 removed `RX_VK_RT_STRICT_PIN`. This time it comes with a guard for #838's case (another program holding that core) and a farm plan. **It is off by default, and this PR proposes no default:** that is the team's call once the farm has run it. **What it does, in player words.** On Windows, the renderer keeps one core free of the game's own threads for the render thread (the "reserved core"). Today the thread prefers that core, but Windows may move it to any fast core. When it moves, it often lands beside a busy game thread on the same physical core, and both run slower. `RX_VK_RT_HOME=1` keeps the render thread on its own core. When something else holds that core off it (#838's case: another program at high priority on the same core), it lets the thread move again for a while: - 2 s at first, doubled for each hold within a minute, 30 s at most; - then it goes home again. It does nothing on Linux, the Steam Deck or macOS, and on a CPU without SMT (two logical processors a core) it changes nothing that matters. **The switch.** `RX_VK_RT_HOME` (kind `attempt`): - unset (or `0`): main's placement, exactly; - `1`: home, with the guard; - `strict`: home with no guard. It is a measuring arm and the tests' negative control, never meant to be used. `RX_VK_RT_HOME_MS` (kind `tune`, default 4) is the guard's threshold. **How the guard decides it is being held.** Between two presents 20-250 ms apart, it looks at two things: 1. whether the render thread lost `RX_VK_RT_HOME_MS` or more off the CPU outside its own waits. The time it spent ready but not woken from its job wait counts as lost; 2. whether every processor of its core was taken by other threads for that long. It acts only once it has seen a hold happen, so **about one stall frame gets through each time the thread goes back home** (measured below). Read first: `gpu/vk/rt_pin.h` (the Guard), then in `gpu/vk/vk_replay.cpp`: `rt_place`, `rt_home_frame`, `job_wait` and `rt_ready_note`. Also changed: - `gpu/vk/vk_fps.cpp`/`.h`: the render thread's waits by kind; - `gpu/vk/switches.def`: the two rows; - `docs/dev/reference/threads.md`; - the tests: `tests/test_rt_pin.py` (new) and `tests/test_reserved_core_mask.py`. ## Testing **Tested:** Windows 11, Ryzen 7 5700X3D (8 cores, 16 threads), RX 9070 XT, AMD 26.11.2. The PR's renderer DLL went over main dd5ffe6a's build (DLL b627b1f4, exe 302120c6): - **The standard test** (Hyrule Field turning), `RX_VK_RT_HOME=1` against unset, 3 alternating pairs: see Measurements. - In every `1` run the guard's lines show the thread at home through the measured window, with 0 holds. - The live affinity mask, read once per arm inside the window, was 0xc000 with `1` and 0xffff unset. - The longest wake from a job wait was 0.66-0.68 ms, against the 4 ms threshold, so ordinary play is never taken for a hold. - **A stall made on purpose:** a test program held both processors of the reserved core at priority 15 for 35 ms in every 250 ms, beside the game, for three rounds of `strict`, `1` and unset in turn. Frames over 40 ms in the window: | round | strict | `1` | unset | |---|---|---|---| | 1 | 80 | 2 | rejected (clock) | | 2 | 81 | 2 | 0 | | 3 | 81 | 2 | 1 | - **Every run of every arm has one long frame right at the end of the load.** Unset's was 39.96 ms in round 2, just under 40, which is why that round counts 0. - **`1`'s second frame in each run is the first stall after a return home,** caught and held at that very present. The guard can only see a hold after it has happened, so one stall frame gets through per return home. - `strict` (no guard) let through 80-81, so the injection worked. - **Unset against main's own build:** the same mask and ideal processor, the same two placement lines, and no new error lines. - **`tests/test_rt_pin.py`:** - The placement for every layout, with unset checked against main's mask, ideal processor and lines. - The guard through the real queue code, cut from the source, with a fake clock and a fake mutex. - Going home, and the stall cases. - **47 broken variants, each of which must fail and does.** - It passes with clang++ and g++ after merging main a2dc5385. One stub was added for main's new `spawn_needed` (#2170), with the PR's code unchanged. - **Other tests:** - `test_reserved_core_mask`, `test_reserved_core`, `test_fps_stats`, `test_perf_pipe_waits` and `test_threaded_rendering` pass with both compilers. - `test_switch_table` fails on `RX_VK_GRASS_DISTANCE`'s order in `switches.def`, as on main. - The full `run_all.py` on dd5ffe6a, with both compilers: nothing fails on the branch that doesn't fail on main. - **Linux syntax** (clang 21, gcc 15): ok on `vk_replay.cpp` and `vk_fps.cpp`, after the merge. - **The Windows build after the merge** (llvm-mingw, the renderer and the game over main's base): it builds clean. It was not run in the game: the numbers below are from the PR's code before the merge. - **The cost off:** one bool test at each job hand-over (1.5 ns measured). With `1`, a clock read on each side of a hand-over to the waiting thread, about 55 ns. **Still to test:** - Windows PCs with SMT CPUs, 4, 6 and 8 cores, Intel with Hyper-Threading and AMD, hybrid Intel too: `RX_VK_RT_HOME=1` against unset on this PR's build, uncapped Hyrule Field. Expected: fps up or level, and no more long frames (`@Beanstalk`, request below). - Real 4-core, 8-thread CPUs in particular: on this PC held to 4 cores the gain was much larger (report-only, below) (`@Beanstalk`). - #838's real case (Intel's PresentMon service on an i5-13600K holding the core): with `1`, about one stall frame per return home, not every stall (by hand). - SMT off: expected to change nothing (by hand). ## Measurements **Verdict:** on this PC, +6.7 % fps with `RX_VK_RT_HOME=1` in the standard test, CPU 43.86 -> 41.31 ms and p99 15.9 -> 14.4 ms. Every pair gained (+8.4, +6.5, +6.8 %). A second session (the implementer's) gave +6.2 %. Baseline: this PR's build with `RX_VK_RT_HOME` unset (main's placement), on main dd5ffe6a's build. Setup: - Windows 11, Ryzen 7 5700X3D, RX 9070 XT, AMD 26.11.2. Hyrule Field turning, 16:9, `RX_VK_ASYNC=1`. - Uncapped: `HOS_FPS_MAX=240` with the options file's cap 0 and vsync 0. At about 98 fps that is in effect the farm's `frame_rate_cap=6`, `vsync=2`. - A fixed light GPU load beside the game. 3 alternating pairs, no warm-ups, no replacements, no run rejected, medians. | | Before: this PR @ 608775b2, `RX_VK_RT_HOME` unset | After: this PR @ 608775b2, `RX_VK_RT_HOME=1` | |---|---|---| | fps, median (runs) | 91.64 (89.99, 91.82, 91.64) | 97.79 (97.59, 97.79, 97.84) | | CPU ms | 43.86 | 41.31 | | p99 frame time, ms | 15.87 | 14.36 | | worst frame, ms (median run) | 59.8 | 48.7 | | frames over 50 / 100 / 250 ms, per run | 1/0/0, 1/0/0, 0/0/0 | 0/0/0, 0/0/0, 1/0/0 | | clock | 1.000, 1.001, 1.000 | 1.000, 1.000, 1.000 | | the replay's submit, ms a frame | 8.36 | 7.26 | | where the render thread ran | 0xffff (may move) | 0xc000 (its core), 0 holds | - **Report only, not the claim:** one pair held to 4 cores with SMT (`--affinity 0xff`), which is a mask of an 8-core CPU, not a 4-core CPU. It gave 65.04 -> 82.34 fps (+26.6 %), p99 23.1 -> 19.1 ms, one run a side. - **Measured on main dd5ffe6a.** The PR is now on main a2dc5385. Main's 7 commits since then don't move the render thread or its queue wait. #2170 starts the thread through `spawn_needed`, and #2174 changes Hyrule Field's simulated turf on the way in, so the absolute numbers may move. Not re-measured here. ## Risks - **About one stall frame per return home gets through** where another program holds the render thread's core (#838's case). The guard sees a hold only once it has happened. Unset doesn't have this. - **A false hold costs one stay** (2 s, at most 30 s, on main's placement), never the session. These limits can cause one: - **Untracked GPU waits on the render thread:** the present's frame-slot fence wait and the acquire after it are not marked as waits, so they count as lost time. Small wakes that add up to 4 ms in a window, with others busy on the core during such a fence wait, could make a false hold. Follow-up: wrap them in `FpsRtWait(kRtGpu)`. - **Two clocks:** the wake is timed on `steady_clock` (QPC), the waits on `fps_clock` (the TSC, scaled once at its first call). A calibration preempted during its loop mis-scales the waits. Follow-up: compare `fps_clock`'s rate with the guard's own 1 s calibration, and give up (`rt_home_off`) when they disagree by more than 2 %. - **The time read:** a delay (a DPC burst) between the idle-count read and `__rdtsc()` makes phantom busy time. - **A notifier preempted while it still holds the queue's lock** after stamping the wake, or other threads taking the core's two processors one after the other while the thread is held off for 4 ms or more. Each needs a wake of 4 ms or more. In ordinary play the longest was 0.68 ms. - **Where it could cost:** on 4 cores the reserved core is a quarter of the CPU. What keeping the render thread there does to the game's other three cores is measured only on the mask above. - **Not touched:** `RX_VK_ASYNC=0`'s pin (`ReplayPin`), host/ and the shader-cache stamp files (no re-prepare). Hybrid CPUs never keep the thread on an efficiency core. A process affinity mask is respected (#1776). - **An explicit `RX_VK_RT_HOME=0`** adds one start-up line naming the value. Unset gives exactly main's lines. - **To turn it off:** unset the switch. It is off by default, so a player notices nothing after updating. ## Checklist ### Ready for review - [x] One topic, on current `main`; the title is a Conventional Commit. - [x] Tested on my own PC, in the game for a change the game runs, as Testing says. - [x] `mise run test` passes (any failure that also fails on `main` is named). - [x] Measured against `main`, or Measurements says why there is nothing to measure. - [x] No game files, build output or hand-edited generated files; `third_party/` untouched; nothing GPL-incompatible. ### When it applies - [x] A switch added or changed: it ships at the default the measurements chose, with its row in `gpu/vk/switches.def`. - [x] A check under `tests/` that reads changed source is updated, not weakened. - [x] Docs updated. - [x] AI assistance recorded with `Assisted-by:` / `Co-Authored-By:` trailers. ### Ready to merge - [ ] Everything in Still to test has run, its results in Tested. - [ ] Tier 1 hardware it can affect is tested (docs/dev/standards.md, "Hardware tiers"). - [ ] Approved in review, every comment answered. - [ ] CI passes on the head; on current `main`, measured again or a comment says why the numbers still hold. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
core of the fastest kind, because kept to the reserved core alone another program's priority-15 thread held it off for
30-40 ms at a time. RX_VK_RT_HOME=1 keeps it home with a guard read at its presents, at most every 20 ms: when, since the
last reading (250 ms at most), it lost RX_VK_RT_HOME_MS (4) off the CPU outside its own waits (QueryThreadCycleTime,
FpsRtWait) and other threads used its core's logical processors at least as long (QueryIdleProcessorCycleTimeEx), it was
held off, and it moves as unset for 2 s, doubled for each hold within a minute at home, 30 s at most. Blocked in the driver
or on a lock, its core is idle: not a hold. strict keeps it there with no guard, for measuring. Unset or 0: exactly as
before (the masks and the log lines). Elsewhere than on Windows nothing changes.

The placement and the guard's rule are in gpu/vk/rt_pin.h (rt_place and rt_home_frame in vk_replay.cpp, the guard run from
bh_shown at the render thread's presents), tested game-free in tests/test_rt_pin.py over made-up CPUs with a fake clock,
cycle counts and idle counts, with twelve broken variants that must fail.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
From the gpt-6.2-sol review of the commit before. The render thread waiting for the recorder thread while the recorder ran on
its core's second processor read as held: the processors' busy time summed was as long as what it lost. Fact 2 is now the
least time every processor of its core it could run on was taken by other threads together (busy - cpu - (n - 1) x gone),
over the processors Windows has not parked (GetSystemCpuSetInformation, read only when fact 1 holds), and at least
RX_VK_RT_HOME_MS. The window's length is taken from the same TSC readings as the idle counts. An affinity call that fails
(the process's affinity narrowed since the start) gives the guard up, free to move as unset. Kept to a core alone (1 or
strict) refuses a reserved core that is not all of the fastest kind (an efficiency core the host may reserve under a
launcher's mask).

tests/test_rt_pin.py: the recorder's wait, a parked processor, an efficiency core reserved, a refused affinity call, and the
real FpsRtWait, fps_rt_wait_sum and fps_rt_waited cut from vk_fps.cpp; seven more broken variants that must fail.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
From the gpt-6.2-sol review's second round. Fact 2's bound now takes the render thread's own waits out (its core taken during
a wait was no hold), and a processor Windows reports parked counts as parked only when it was idle the whole window. When an
affinity call is refused (the process's affinity narrowed since the start), the guard lets the thread move within the
process's affinity now, or says it could not. A refused placement says so instead of claiming a pin. The CPU-set buffer grows
to what Windows asks for.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
From the gpt-6.2-sol review's third round: a process affinity narrowed at run time to the core and processors outside the
wide mask left the thread at home; an unreadable process affinity (threads in several groups) skipped the widening. The start
line names the rule as it is.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
The job's step 3 (a priority-15 stall on the reserved core, 35 ms in every 250 ms) let several stalls through after each
return home: windows of a normal frame and the held one (about 55 ms) added the normal frame's idle sibling to fact 2's bound
(35 ms lost, the bound at -2.5 and 1.7 ms). The held frame judged alone reads about 25 ms. A sample at every present costs
~15 us; it is taken only for Guard::kFineNs after a window that lost RX_VK_RT_HOME_MS and after going home from a hold, so a
PC with no holds keeps the 20 ms samples outside its loads.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
A hold that starts while the render thread waits for a job was counted as waiting (FpsRtWait), so the guard missed it: the
thread is ready from the moment a job it may take is handed over, but the wait's clock runs until it gets the CPU back.
job_wait (the render thread's job wait, moved out of replay_thread_main unchanged) now, with the guard on, takes the moment
the job could go while it waited (stamped once under q_mu by enqueue, gate_need_value or wait_queue_empty, or by the frame
gate's deadline, never before its own timed wait let it wake) and sums the time from then to the job's start; the guard
takes that out of its waits, in fact 1's lost time and fact 2's bound. A job handed over while the thread ran counts
nothing. Going home now takes the readings, so the first window at home is judged (11-21's stalls came one after each
return home). The log says, for the first 16 presents at home more than 40 ms after the one before, the window's time on
the CPU, its waits by kind, its wake delay, the facts and what the guard did, and the longest wake in the 600-presents line.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
The gpt-6.2-sol review of the range (round 1): a stopped queue's wait whose deadline had gone by while the thread ran
backdated the gate deadline's stamp (the floor is now never before it went to sleep); the three places stamp as the last
thing before they let q_mu go, after their notify; going home takes the readings before the affinity call, so a hold the
move itself meets falls in the first window; the long present's line is said after the guard's affinity call, and says when
the move was refused. The fixture's mutex and queue now fail when the queue is touched without q_mu, and the tests cover a
stay longer than the fine sampling's 10 s and a hold met by the move home. The long-present lines are 16 a 600 presents.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Co-Authored-By: Claude Opus 5.5 <[email protected]>
In 11-22's step 3 (a priority-15 stall on both reserved processors, 35 ms in every 250 ms), the stall after a return home
held the render thread in its job wait: it woke 32.9 ms late, fact 1 counted 34.1 ms lost, but fact 2's bound, the overlap
of the core's processors over a 62 ms window with 16.5 ms of waits of its own, came to -7.2 ms and the frame got through.
With a wake delay that long the thread was known to be ready and not running; what fact 2 must still rule out is Windows
being slow with the core free, which each processor's own busy time (less the thread's CPU time and its other waits) shows.
Without such a wake the rule is as before (a recorder on one processor and another thread on the other, one after the
other, is still no hold).

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Six broken copies of cd0558d3 passed the test: any wake (or 1 ms) enabling the rule, a 10 ms wake needed, the most busy
processor or the processors' sum in place of the least busy, and the thread's CPU time or its other waits left in. Four
cases in the fine sampling now settle them: a 3.9 ms wake with each processor busy in turn (no hold); a 35 ms wake with one
processor taken and the other free (no hold); a 5 ms wake with the core taken only while it waited for work (no hold); and
11-22's step-3 window with a 5 ms wake (a hold). Each copy is a broken variant of the test that must fail.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
vk_replay.cpp's include block: main moved vk_work.h after its new
mma_layout.h; rt_pin.h and the guard's clock block stay before them.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
test(vk): test_rt_pin's queue fixture stubs cxx_escape::spawn_needed, which main's #2170 made enqueue's way to start the render thread
Some checks failed
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been skipped
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been skipped
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been skipped
Third-party parts / The same as upstream (pull_request) Has been skipped
Website / Website on Cloudflare (pull_request) Has been skipped
CI / Conventional Commit title (pull_request) Successful in 7s
PR description / Description follows the template (pull_request_target) Successful in 10s
Guide screenshots / Guide images and local UI source review (pull_request) Failing after 11s
CI / Workflow lint (actionlint) (pull_request) Successful in 12s
Website / Website on Cloudflare (pull_request_target) Has been skipped
Third-party parts / As their notes say (pull_request) Successful in 14s
CI / Website build and checks (pull_request) Failing after 27s
beanstalk/approval approved by klimovich008 (reviewed)
beanstalk/windows+cpu-3300x_i3-10100_i3-12100_i3-13100_i7-6700_i7-6700k_i7-7700_i7-7700k 0/1 pass, 0 running, 2 queued
beanstalk/windows+cpu-3600_5600_5600x_7600_7600x_i5-10400_i5-11400_i7-8700k 0/1 pass, 0 running, 2 queued
beanstalk/windows+cpu-3700x_5700x_5800x_5800x3d_7700x_7800x3d_i7-10700k_i9-9900k 0/1 pass, 2 running, 0 queued
beanstalk/windows+cpu-i5-12400_i5-12600k_i5-13600k_i5-14600k_i7-12700k_i7-13700k 0/1 pass, 0 running, 2 queued
5867310bc8
The fixture cuts enqueue from vk_replay.cpp; since #2170 it starts the
render thread with cxx_escape::spawn_needed instead of std::thread. The
stub starts the fixture's empty replay_thread_main as std::thread did.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Author
Member

Tested in game: Windows 11, Ryzen 7 5700X3D (8C/16T), RX 9070 XT, AMD 26.11.2, the PR's renderer over main dd5ffe6a's build: Hyrule Field turning, uncapped (HOS_FPS_MAX=240), 3 alternating pairs of RX_VK_RT_HOME=1 against unset: 91.64 -> 97.79 fps, the thread at home through every window with 0 holds, live mask 0xc000 against 0xffff, clock 1.000. A stall made on purpose (both reserved processors at priority 15, 35 ms in every 250 ms), 3 rounds: frames over 40 ms strict 80/81/81, =1 2/2/2, unset -/0/1 (about one stall frame per return home gets through). Unset against main's own build: the same mask, the same placement lines, no new error lines. The merge with main a2dc5385 was built, not run. Not run: Intel, other AMD CPUs, 4- and 6-core CPUs, SMT off, #838's real stall.

Tested in game: Windows 11, Ryzen 7 5700X3D (8C/16T), RX 9070 XT, AMD 26.11.2, the PR's renderer over main dd5ffe6a's build: Hyrule Field turning, uncapped (HOS_FPS_MAX=240), 3 alternating pairs of RX_VK_RT_HOME=1 against unset: 91.64 -> 97.79 fps, the thread at home through every window with 0 holds, live mask 0xc000 against 0xffff, clock 1.000. A stall made on purpose (both reserved processors at priority 15, 35 ms in every 250 ms), 3 rounds: frames over 40 ms strict 80/81/81, =1 2/2/2, unset -/0/1 (about one stall frame per return home gets through). Unset against main's own build: the same mask, the same placement lines, no new error lines. The merge with main a2dc5385 was built, not run. Not run: Intel, other AMD CPUs, 4- and 6-core CPUs, SMT off, #838's real stall.
Author
Member

@Beanstalk the PR's own build only, its switch against itself: baseline ref = pr, [matrix.pr_only] RX_VK_RT_HOME = ["0", "1"], and pass.compare RX_VK_RT_HOME = "1" against "0" on the pr side (unset is the same as 0: main's placement). It is off by default and does anything only on Windows CPUs with SMT, so every target is Windows.

Targets, each with 1-2 reports. Windows with a CPU that has two logical processors a core, by model:

  • 4 cores / 8 threads: i7-6700K, i7-7700K, i7-6700, i7-7700, i3-10100, i3-12100, i3-13100, Ryzen 3 3300X;
  • 6 / 12: Ryzen 5 3600, 5600X, 5600, 7600X, Ryzen 5 7600, i5-10400, i5-11400, i7-8700K;
  • 8 / 16: Ryzen 7 3700X, 5700X, 5800X, 5800X3D, 7700X, 7800X3D, i9-9900K, i7-10700K;
  • hybrid Intel: i5-12600K, i5-13600K, i5-14600K, i7-12700K, i7-13700K, i5-12400.
    The 4-core ones matter most (on this PC a 4-core mask gained far more, report-only), but don't let the plan wait on them alone.

Settings: frame_rate_cap = 6 (Uncapped), vsync = 2 (Off), frame_generation = 0; headless is fine.

Scenario: Hyrule Field (HyruleField_Child), measured 30 s, 3 game starts a cell (2 at least). Compare rules, pr side, RX_VK_RT_HOME=1 vs 0: fps_mean min_delta_pct = 3 (fails on a loss of more than 3 %) and frametime_p99_ms max_delta_pct = 10; crash = false. Hitch counts (run_hitch_count, run_frametime_max_ms) report-only.

Log on the =1 cells: "vk: RX_VK_RT_HOME=1: the render thread is kept to logical processors" expected (the thread really went home: on a PC where the placement does not apply it says "free to move, as unset" instead, and that cell compares nothing).

@Beanstalk the PR's own build only, its switch against itself: baseline ref = pr, [matrix.pr_only] RX_VK_RT_HOME = ["0", "1"], and [[pass.compare]] RX_VK_RT_HOME = "1" against "0" on the pr side (unset is the same as 0: main's placement). It is off by default and does anything only on Windows CPUs with SMT, so every target is Windows. Targets, each with 1-2 reports. Windows with a CPU that has two logical processors a core, by model: - 4 cores / 8 threads: i7-6700K, i7-7700K, i7-6700, i7-7700, i3-10100, i3-12100, i3-13100, Ryzen 3 3300X; - 6 / 12: Ryzen 5 3600, 5600X, 5600, 7600X, Ryzen 5 7600, i5-10400, i5-11400, i7-8700K; - 8 / 16: Ryzen 7 3700X, 5700X, 5800X, 5800X3D, 7700X, 7800X3D, i9-9900K, i7-10700K; - hybrid Intel: i5-12600K, i5-13600K, i5-14600K, i7-12700K, i7-13700K, i5-12400. The 4-core ones matter most (on this PC a 4-core mask gained far more, report-only), but don't let the plan wait on them alone. Settings: frame_rate_cap = 6 (Uncapped), vsync = 2 (Off), frame_generation = 0; headless is fine. Scenario: Hyrule Field (HyruleField_Child), measured 30 s, 3 game starts a cell (2 at least). Compare rules, pr side, RX_VK_RT_HOME=1 vs 0: fps_mean min_delta_pct = 3 (fails on a loss of more than 3 %) and frametime_p99_ms max_delta_pct = 10; crash = false. Hitch counts (run_hitch_count, run_frametime_max_ms) report-only. Log on the =1 cells: "vk: RX_VK_RT_HOME=1: the render thread is kept to logical processors" expected (the thread really went home: on a PC where the placement does not apply it says "free to move, as unset" instead, and that cell compares nothing).
Member

Beanstalk test farm · 🔄 testing (0 of 4 reports so far)

Commit 5867310 · approved by klimovich008 (reviewed) · dashboard

⚠️ Note: Two starts per setting keep the plan within the farm’s time limit, but leave no spare usable start for the frame-time comparison.
⚠️ Note: The placement log is unjudged when absent, including in the 0 cell; the farm cannot suppress the performance comparison when placement is refused, so such results may compare identical placement.
⚠️ Note: The farm cannot inject the competing high-priority workload needed to test guard recovery under a deliberate core hold.

Started at 10:54 UTC.

Estimate: done in about 1 h 15 min, from the queue and the testers online now; some jobs can't start yet.

Target Result Reports VRAM peak (MB) Crashes
windows+cpu-3300x_i3-10100_i3-12100_i3-13100_i7-6700_i7-6700k_i7-7700_i7-7700k (CPU "3300x" or "i3 10100" or "i3 12100" or "i3 13100" or "i7 6700" or "i7 6700k" or "i7 7700" or "i7 7700k") ⏳ queued 0 of 1 — —
windows+cpu-3600_5600_5600x_7600_7600x_i5-10400_i5-11400_i7-8700k (CPU "3600" or "5600" or "5600x" or "7600" or "7600x" or "i5 10400" or "i5 11400" or "i7 8700k") ⏳ queued 0 of 1 — —
windows+cpu-3700x_5700x_5800x_5800x3d_7700x_7800x3d_i7-10700k_i9-9900k (CPU "3700x" or "5700x" or "5800x" or "5800x3d" or "7700x" or "7800x3d" or "i7 10700k" or "i9 9900k") 🔄 running 0 of 1 — —
windows+cpu-i5-12400_i5-12600k_i5-13600k_i5-14600k_i7-12700k_i7-13700k (CPU "i5 12400" or "i5 12600k" or "i5 13600k" or "i5 14600k" or "i7 12700k" or "i7 13700k") ⏳ queued 0 of 1 — —

hyrule-field · frame time p99 (ms)

Target switch off RX_VK_RT_HOME=0 (reference) RX_VK_RT_HOME=1
windows+cpu-3300x_i3-10100_i3-12100_i3-13100_i7-6700_i7-6700k_i7-7700_i7-7700k (CPU "3300x" or "i3 10100" or "i3 12100" or "i3 13100" or "i7 6700" or "i7 6700k" or "i7 7700" or "i7 7700k") — — —
windows+cpu-3600_5600_5600x_7600_7600x_i5-10400_i5-11400_i7-8700k (CPU "3600" or "5600" or "5600x" or "7600" or "7600x" or "i5 10400" or "i5 11400" or "i7 8700k") — — —
windows+cpu-3700x_5700x_5800x_5800x3d_7700x_7800x3d_i7-10700k_i9-9900k (CPU "3700x" or "5700x" or "5800x" or "5800x3d" or "7700x" or "7800x3d" or "i7 10700k" or "i9 9900k") — — —
windows+cpu-i5-12400_i5-12600k_i5-13600k_i5-14600k_i7-12700k_i7-13700k (CPU "i5 12400" or "i5 12600k" or "i5 13600k" or "i5 14600k" or "i7 12700k" or "i7 13700k") — — —
What was tested

What will be tested · AI plan (gpt-6.1-terra)

On selected Windows SMT CPUs, this compares the PR’s render-thread placement switch with its off setting in Hyrule Field. Two starts per setting measure 30 seconds of uncapped gameplay, taking about 45 minutes per tester. A crash, an FPS loss over 3%, or a p99 frame-time increase over 10% fails the comparison; hitches are reported without a limit. Some requested CPU targets may wait for matching Windows testers. No tester online now matches windows+cpu-3300x_i3-10100_i3-12100_i3-13100_i7-6700_i7-6700k_i7-7700_i7-7700k, windows+cpu-3600_5600_5600x_7600_7600x_i5-10400_i5-11400_i7-8700k, windows+cpu-i5-12400_i5-12600k_i5-13600k_i5-14600k_i7-12700k_i7-13700k: those targets wait for one.

  • Where: windows+cpu-3300x_i3-10100_i3-12100_i3-13100_i7-6700_i7-6700k_i7-7700_i7-7700k (CPU "3300x" or "i3 10100" or "i3 12100" or "i3 13100" or "i7 6700" or "i7 6700k" or "i7 7700" or "i7 7700k", 1-2 reports), windows+cpu-3600_5600_5600x_7600_7600x_i5-10400_i5-11400_i7-8700k (CPU "3600" or "5600" or "5600x" or "7600" or "7600x" or "i5 10400" or "i5 11400" or "i7 8700k", 1-2 reports), windows+cpu-3700x_5700x_5800x_5800x3d_7700x_7800x3d_i7-10700k_i9-9900k (CPU "3700x" or "5700x" or "5800x" or "5800x3d" or "7700x" or "7800x3d" or "i7 10700k" or "i9 9900k", 1-2 reports), windows+cpu-i5-12400_i5-12600k_i5-13600k_i5-14600k_i7-12700k_i7-13700k (CPU "i5 12400" or "i5 12600k" or "i5 13600k" or "i5 14600k" or "i7 12700k" or "i7 13700k", 1-2 reports).
  • Against: the PR itself, switch on (RX_VK_RT_HOME) and off; main is not built or run.
  • Matrix: 2 cells, each with the switch on and off: RX_VK_RT_HOME 0 or 1 (PR only).
  • hyrule-field: from the repository's save castle-gate (settings vsync=2, frame_rate_cap=6, frame_generation=0), then teleport to Hyrule Field, measure field for 30 s (after 5 s warm-up); 2 runs on each side for each of 2 cells (2 game starts a run).
  • Measured: crashes, frame times, stutters and hitches, draws and frame rate, the game's log (an expected line that never appears: could not be judged).
  • The PR fails if: it crashes; or the PR's frame rate with RX_VK_RT_HOME=1 is more than 3% below RX_VK_RT_HOME=0; or the PR's frame time (p99) with RX_VK_RT_HOME=1 is more than 10% above RX_VK_RT_HOME=0; or expected log lines missing is above 0.
  • Time: about 45 min of game runs per tester (8 game starts), plus the two builds.
<!-- beanstalk:sticky --> ### Beanstalk test farm · 🔄 testing (0 of 4 reports so far) Commit `5867310` · approved by klimovich008 (reviewed) · [dashboard](https://beanstalk.magicbean.dev/prs/2240) ⚠️ **Note:** Two starts per setting keep the plan within the farm’s time limit, but leave no spare usable start for the frame-time comparison. ⚠️ **Note:** The placement log is unjudged when absent, including in the 0 cell; the farm cannot suppress the performance comparison when placement is refused, so such results may compare identical placement. ⚠️ **Note:** The farm cannot inject the competing high-priority workload needed to test guard recovery under a deliberate core hold. Started at 10:54 UTC. **Estimate:** done in about 1 h 15 min, from the queue and the testers online now; some jobs can't start yet. | Target | Result | Reports | VRAM peak (MB) | Crashes | | --- | --- | --- | --- | --- | | `windows+cpu-3300x_i3-10100_i3-12100_i3-13100_i7-6700_i7-6700k_i7-7700_i7-7700k` (CPU "3300x" or "i3 10100" or "i3 12100" or "i3 13100" or "i7 6700" or "i7 6700k" or "i7 7700" or "i7 7700k") | ⏳ queued | 0 of 1 | — | — | | `windows+cpu-3600_5600_5600x_7600_7600x_i5-10400_i5-11400_i7-8700k` (CPU "3600" or "5600" or "5600x" or "7600" or "7600x" or "i5 10400" or "i5 11400" or "i7 8700k") | ⏳ queued | 0 of 1 | — | — | | `windows+cpu-3700x_5700x_5800x_5800x3d_7700x_7800x3d_i7-10700k_i9-9900k` (CPU "3700x" or "5700x" or "5800x" or "5800x3d" or "7700x" or "7800x3d" or "i7 10700k" or "i9 9900k") | 🔄 running | 0 of 1 | — | — | | `windows+cpu-i5-12400_i5-12600k_i5-13600k_i5-14600k_i7-12700k_i7-13700k` (CPU "i5 12400" or "i5 12600k" or "i5 13600k" or "i5 14600k" or "i7 12700k" or "i7 13700k") | ⏳ queued | 0 of 1 | — | — | **`hyrule-field`** · frame time p99 (ms) | Target | switch off | `RX_VK_RT_HOME=0` (reference) | `RX_VK_RT_HOME=1` | | --- | --- | --- | --- | | `windows+cpu-3300x_i3-10100_i3-12100_i3-13100_i7-6700_i7-6700k_i7-7700_i7-7700k` (CPU "3300x" or "i3 10100" or "i3 12100" or "i3 13100" or "i7 6700" or "i7 6700k" or "i7 7700" or "i7 7700k") | — | — | — | | `windows+cpu-3600_5600_5600x_7600_7600x_i5-10400_i5-11400_i7-8700k` (CPU "3600" or "5600" or "5600x" or "7600" or "7600x" or "i5 10400" or "i5 11400" or "i7 8700k") | — | — | — | | `windows+cpu-3700x_5700x_5800x_5800x3d_7700x_7800x3d_i7-10700k_i9-9900k` (CPU "3700x" or "5700x" or "5800x" or "5800x3d" or "7700x" or "7800x3d" or "i7 10700k" or "i9 9900k") | — | — | — | | `windows+cpu-i5-12400_i5-12600k_i5-13600k_i5-14600k_i7-12700k_i7-13700k` (CPU "i5 12400" or "i5 12600k" or "i5 13600k" or "i5 14600k" or "i7 12700k" or "i7 13700k") | — | — | — | <details><summary>What was tested</summary> **What will be tested** · AI plan (`gpt-6.1-terra`) On selected Windows SMT CPUs, this compares the PR’s render-thread placement switch with its off setting in Hyrule Field. Two starts per setting measure 30 seconds of uncapped gameplay, taking about 45 minutes per tester. A crash, an FPS loss over 3%, or a p99 frame-time increase over 10% fails the comparison; hitches are reported without a limit. Some requested CPU targets may wait for matching Windows testers. No tester online now matches windows+cpu-3300x_i3-10100_i3-12100_i3-13100_i7-6700_i7-6700k_i7-7700_i7-7700k, windows+cpu-3600_5600_5600x_7600_7600x_i5-10400_i5-11400_i7-8700k, windows+cpu-i5-12400_i5-12600k_i5-13600k_i5-14600k_i7-12700k_i7-13700k: those targets wait for one. - **Where:** windows+cpu-3300x_i3-10100_i3-12100_i3-13100_i7-6700_i7-6700k_i7-7700_i7-7700k (CPU "3300x" or "i3 10100" or "i3 12100" or "i3 13100" or "i7 6700" or "i7 6700k" or "i7 7700" or "i7 7700k", 1-2 reports), windows+cpu-3600_5600_5600x_7600_7600x_i5-10400_i5-11400_i7-8700k (CPU "3600" or "5600" or "5600x" or "7600" or "7600x" or "i5 10400" or "i5 11400" or "i7 8700k", 1-2 reports), windows+cpu-3700x_5700x_5800x_5800x3d_7700x_7800x3d_i7-10700k_i9-9900k (CPU "3700x" or "5700x" or "5800x" or "5800x3d" or "7700x" or "7800x3d" or "i7 10700k" or "i9 9900k", 1-2 reports), windows+cpu-i5-12400_i5-12600k_i5-13600k_i5-14600k_i7-12700k_i7-13700k (CPU "i5 12400" or "i5 12600k" or "i5 13600k" or "i5 14600k" or "i7 12700k" or "i7 13700k", 1-2 reports). - **Against:** the PR itself, switch on (`RX_VK_RT_HOME`) and off; main is not built or run. - **Matrix:** 2 cells, each with the switch on and off: RX_VK_RT_HOME `0` or `1` (PR only). - **hyrule-field:** from the repository's save `castle-gate` (settings vsync=2, frame_rate_cap=6, frame_generation=0), then teleport to Hyrule Field, measure field for 30 s (after 5 s warm-up); 2 runs on each side for each of 2 cells (2 game starts a run). - **Measured:** crashes, frame times, stutters and hitches, draws and frame rate, the game's log (an expected line that never appears: could not be judged). - **The PR fails if:** it crashes; or the PR's frame rate with RX_VK_RT_HOME=1 is more than 3% below RX_VK_RT_HOME=0; or the PR's frame time (p99) with RX_VK_RT_HOME=1 is more than 10% above RX_VK_RT_HOME=0; or expected log lines missing is above 0. - **Time:** about 45 min of game runs per tester (8 game starts), plus the two builds. </details>
Some checks failed
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been skipped
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been skipped
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been skipped
Third-party parts / The same as upstream (pull_request) Has been skipped
Website / Website on Cloudflare (pull_request) Has been skipped
CI / Conventional Commit title (pull_request) Successful in 7s
PR description / Description follows the template (pull_request_target) Successful in 10s
Guide screenshots / Guide images and local UI source review (pull_request) Failing after 11s
CI / Workflow lint (actionlint) (pull_request) Successful in 12s
Website / Website on Cloudflare (pull_request_target) Has been skipped
Third-party parts / As their notes say (pull_request) Successful in 14s
CI / Website build and checks (pull_request) Failing after 27s
beanstalk/approval approved by klimovich008 (reviewed)
beanstalk/windows+cpu-3300x_i3-10100_i3-12100_i3-13100_i7-6700_i7-6700k_i7-7700_i7-7700k 0/1 pass, 0 running, 2 queued
beanstalk/windows+cpu-3600_5600_5600x_7600_7600x_i5-10400_i5-11400_i7-8700k 0/1 pass, 0 running, 2 queued
beanstalk/windows+cpu-3700x_5700x_5800x_5800x3d_7700x_7800x3d_i7-10700k_i9-9900k 0/1 pass, 2 running, 0 queued
beanstalk/windows+cpu-i5-12400_i5-12600k_i5-13600k_i5-14600k_i7-12700k_i7-13700k 0/1 pass, 0 running, 2 queued
This pull request is marked as a work in progress.
View command line instructions

Checkout

From your project repository, check out a new branch and test the changes.
git fetch -u 'perf/rt-home-v3:klimovich008-perf/rt-home-v3'
git switch 'klimovich008-perf/rt-home-v3'
Sign in to join this conversation.
No reviewers
No milestone
No project
No assignees
2 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
MagicBeanDev/Longshot!2240
No description provided.