perf(vk): the presenter thread on by default with NVIDIA on Linux and AMD on Windows, with NVIDIA Reflex; after #1121 #1702
No reviewers
Labels
No labels
area/accessibility
area/audio
area/input
area/install
area/menu
area/renderer
area/saves
area/tooling
area/upscaling
beanstalk: approved
beanstalk: failed
beanstalk: needs approval
beanstalk: running
beanstalk: tested
ci: approved
good first issue
gpu/amd
gpu/igpu
gpu/intel
gpu/nvidia
help wanted
kind
bug
kind
chore
kind
docs
kind
feature
kind
performance
kind
question
needs measurements
platform/linux
platform/macos
platform/steam-deck
platform/windows
preview
priority
critical
priority
high
priority
low
regression
risk
high
risk
low
risk
medium
status
blocked
status
confirmed
status
duplicate
status
invalid
status
needs-info
status
wontfix
ci: approved
No milestone
No project
No assignees
2 participants
Notifications
Due date
No due date set.
Dependencies
No dependencies set
Reference
MagicBeanDev/Longshot!1702
Loading…
Reference in a new issue
No description provided.
Delete branch "HumanBean/Longshot:perf/presenter-thread-nvidia-linux"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Type: perf · Risk: medium · Tested on: CachyOS Linux, RTX 3090; the Test Farm: Windows RX 9070 XT (2), Windows RTX 3050 Ti and 3090, Linux RX 7900 XTX, Steam Deck · Needs: Intel (Windows and Linux; queued on the farm, no Linux Intel tester yet)
TL;DR
On Linux with NVIDIA's driver, a present blocks until the frame's copy to the screen is done, about 4.5 ms, and the render thread can't start the next frame meanwhile, so the GPU sits idle at each frame's start. This PR presents from a thread of its own by default with NVIDIA's driver on Linux and with AMD's driver on Windows (discrete cards), where the Test Farm measured it faster, and lets that thread work with NVIDIA Reflex on (until now Reflex kept presents on the game's thread). It builds on #1121, which replays the game's work as it comes instead of waiting for each frame's present. Together, on an RTX 3090 in Castle Town Market: about +7% fps walking and +4% turning (+8% for both with Reflex on), with no one-frame flash more than main in the frame log. On the Test Farm, uncapped: Windows RX 9070 XT Kokiri Forest +11%, Market +6% standing and +2% walking (presenter thread on); Linux RX 7900 XTX Kokiri +41%, Market 9 → 72 fps (main stalls ~1 s at a time there with V-Sync off, the PR doesn't); Windows NVIDIA even in fps with a lower p99; the Steam Deck +2% to +6%.
Breakdown
vkQueuePresentKHRuntil the frame's copy is done (~4.5 ms), and with Reflex on, the presenter thread that avoids that never ran.RX_VK_FRAME_GATE=1andRX_VK_PRESENT_THREAD=0turn each part off.Why
A profile of the Market on an RTX 3090 (DLSS Quality, 2560x1440) showed the render thread busy all frame and the GPU idle about 4 ms at each frame's start. NVIDIA's driver on Linux returns from
vkQueuePresentKHRonly once the frame's copy to the swapchain is done (the presenter thread's log: the copy's fence 4.5-4.6 ms in play, the present itself 0.23-0.24 ms), and the render thread was waiting on it. The presenter thread (#1043's work) makes the present elsewhere, butpresent_outhanded nothing to it when the present carried Reflex's present id, and DLSS players have Reflex on. Without #1121, the render thread also waits for each frame's present before replaying its work, which keeps the same gap.See #1121 (now its own PR, ready: this branch sits on it, so its commits show here until it merges; review only the top three, HumanBean's presenter commits and their docs), #1043.
What changed
The presenter thread is on by default with NVIDIA's driver on Linux and AMD's Windows driver on a discrete card, and takes Reflex's present id with each present. It sits on #1121 (the render thread replays jobs as they come), which merges first; the presenter thread's gain on Linux NVIDIA was measured with it.
gate_on()ingpu/vk/vk_replay.cppis off unlessRX_VK_FRAME_GATE=1; the present still waits for the main thread's signal (RX_VK_PRESENT_AFTER_SIGNAL).present_thread_on()defaults topresent_thread_default()(gpu/vk/presenter_queue.h: NVIDIA not on Windows, or AMD's proprietary driver on Windows on a discrete card), andcreate_deviceasks for the presenter's queue there too;RX_VK_PRESENT_THREAD=0/1chooses over it.tests/presenter_queue_test.cppchecks the default on each platform.present_outhands a present with Reflex's id over too:present_hand_off(index, fence, reflex_id), andpresenter_looppresents with that id, then marksPRESENT_END.reflex_present_handed_off()andreflex_present_end(id)ingpu/vk/vk_reflex.cpp: the game's thread clearspresentingat the hand-off; the presenter thread marks the present's end.RX_VK_PRESENT_THREADandRX_VK_FRAME_GATEingpu/vk/switches.defanddocs/development.md;docs/development/frame-pipeline.md;device_policy.h's presenter queue comment.Measurements
Verdict: Faster on an RTX 3090 on Linux: Castle Town Market walking 78.2 → 84.0 fps (+7%) and turning 97.2 → 100.8 (+3.7%), medians of 3 and 3 sessions (the sides' runs don't overlap in either) with the defaults; with Reflex on, one session each, walking 76.9 → 83.1 and turning 94.8 → 102.5. The GPU's work a frame is the same (6.8-6.95 ms turning, 8.4-8.5 ms walking): it waits less between frames. The frame log: no one-frame dark or bright picture on either side (0 in 25,400 frames on main, 0 in 27,486 with this PR, Reflex on), and as many small one-frame block changes as main, all of them where the turning camera brushes a wall.
Hardware: CachyOS Linux (kernel 7.3.3), RTX 3090 24 GB, NVIDIA 620.70.06, Ryzen 9 7900X (12 cores, 24 threads; Threaded rendering's Auto: both threads), 32 GB
Scene: Castle Town Market (child, debug teleport from the Kokiri save), day; 5 s walking forward, then 20 s walking with the camera turning; 2560x1440 output, DLSS Quality (Preset K), no frame generation,
HOS_FPS_MAX=1000, a private headless sway at 240 Hz, sound offMethod: the renderer built alone on one base (only
longshot_vk.sodiffers), each side with its own NVIDIA shader cache. Warm-up then 2 timed sessions each, interleaved (main, PR, main, PR), on the PC's timing queue (one core complex to itself); the fps from the renderer'svk: <n> presents, <fps> fpslines inside each phase, frame times fromRX_VK_FRAME_LOG, GPU time from the log'svk: GPU time <n> ms a framelines. Reflex is off by default in a headless session: one more pair withRX_VK_REFLEX=1. Flashes: 3 more sessions each withRX_VK_REFLEX=1and the frame log (RX_VK_FRAMELOG) on over Kokiri Forest and then the Market, in each 30 s standing, 5 s walking and 20 s walking with the camera turning (interleaved after a warm-up: main, PR, main, PR, on the shared queue, so their fps are not reported);scripts/framelog.pyon each, and a whole-picture check: a frame whose mean brightness is 12+ levels below (or above) both neighbours that are within 6 of each other, and one where 3+ of the picture's 16x9 blocks are 20+ levels below (or above) both neighbours by more than those differHow to reproduce:
longshot_ctl.py teleport Market_Child, then hold forward 5 s, then forward and the right stick right 20 s; the log'svk: <n> presents, <fps> fpslines.Baseline: both columns are renderers built alone on a base at main
da91e770, measured before the rebase onto09cbdd97. The rebase met main's present-wait bookkeeping inpresenter_loop(the frame trace's present-to-shown times) and Threaded rendering's new Auto (#1664), which still turns both threads on for this CPU; it merged without a conflict in the code. The After column also had #1701's commit (the texture pipelines before the game), which doesn't run in play once the pipelines exist.da91e770da91e770)RX_VK_REFLEX=1, one session each: walking / turning fpsscripts/framelog.pyevents (all windows)The Test Farm
Runs 1-3 of
@Beanstalkon this PR, main and the PR built and run on the same PC, ABBA, two game starts a side: Kokiri Forest and Castle Town Market, 15 s standing and 10 s after moving, V-Sync off and the frame rate uncapped (run 3; run 2 was at the 60 cap). Run 3's PR side was the head before the last commit, so its forced cell (RX_VK_PRESENT_THREAD=1) is what Windows AMD gets by default now. Medians of the reports, fps (p99 frame time in ms):(The forced column's main differs a little from the default column's: each cell has its own main runs; the Steam Deck's forced cell against its own main was even, p99 +4% to +14% in the Market. The farm's second Linux AMD report was a Steam Deck's, left out of the Linux row. The picture checks "fail" on every target as main does against itself in these scenes: the fails are the farm's noise, not a picture change.)
Main on the RX 7900 XTX stalls about 1 s at a time with V-Sync off (frame max 1,000-1,040 ms in every scene, Market p99 1,010 ms); with #1121's part it doesn't (max 46-70 ms). Run 2 (60 cap) had the same: main 45.9 fps, the PR 60.
Per platform: #1121's part is on for everyone (faster or even everywhere). The presenter thread: on for Windows AMD (Market +8% standing and +10% after moving over the PR's default, p99 lower in all four), off for Windows NVIDIA (fps even, p99 +17% in Kokiri), off for RADV (even on the RX 7900 XTX; p99 up to +14% on the Steam Deck).
Flags and settings
One switch changes its default on NVIDIA Linux and Windows AMD; #1121's commit changes another for everyone.
RX_VK_PRESENT_THREAD=1on)=0/=1chooseRX_VK_FRAME_GATE(#1121)=0off)=1on)Testing
Tested: CachyOS Linux, RTX 3090, NVIDIA 620.70.06: the renderer built alone on a base of main; 8 sessions to Castle Town Market walking and turning (above), and 6 sessions (3 each side, Reflex on) with the frame log over Kokiri Forest and the Market; no device loss, no failed present, the presenter thread on in every session of this PR. The rebased head (on main
c4e8e2e) compiles (scripts/dev_build.py vk). The Test Farm, runs 1-3 (above): Windows AMD and NVIDIA, Linux AMD, the Steam Deck; no crash on any.Not tested: Intel (Windows and Linux; queued on the farm); the last commit's Windows AMD default as a default (the farm ran it as
RX_VK_PRESENT_THREAD=1, the same code path; run 4 checks it); Reflex's own latency numbers with the presenter thread (its markers are sent; the latency wasn't measured); a real display (the runs are in a headless compositor).mise run test: not run on this PC.Must be tested:
presenter thread: each present made on a threadin the log, no tearing or stutter, and fps at least main's (Test Farm: @Beanstalk)vk: presenter thread: each present madein the log, fps at least main's (Test Farm: @Beanstalk, run 4)RX_VK_FRAMELOG=<file>, thenpython -I scripts/framelog.py <file>, expecting no one-frame dark patch more than on main (Test Farm: @Beanstalk)Risks
What could break: a one-frame dark picture where the replay catches up with the game (#1121); a present out of order or a stall on an NVIDIA Linux driver not run here; Reflex's latency readings with presents on another thread.
Turning it off:
RX_VK_PRESENT_THREAD=0(presents on the game's thread),RX_VK_FRAME_GATE=1(whole frames); or revert.Compatibility: Nothing changes: no option, cache or API.
Checklist
Everyone
type(scope): description) of the status line's type.can't be run by hand is in Testing's Not testable by hand, with why.
in Risks. Not ticked: Reflex's latency with the presenter thread is not measured.
main.python -I -B tests/run_all.pypasses (any failure that also fails onmainis named). Not run on this PC.third_party/untouched (a new folder with aPROVENANCE.mdinstead).When it applies
Assisted-by:/Co-Authored-By:trailers.docs/tested/for each new measured run (format indocs/tested-configurations.md): 14 (tests/test_tested_configurations.pynot run on this PC).🤖 Generated with Claude Code
https://claude.ai/code/session_016o1k1PazAdTJxQQWQBksW2
Tested in game: RTX 3090, NVIDIA 620.70.06, CachyOS Linux (headless sway), 2560x1440 DLSS Quality: 8 sessions in Castle Town Market walking and turning (78.2 -> 84.0 fps walking, 97.2 -> 100.8 turning, medians) and 6 with Reflex on and the frame log over Kokiri Forest and the Market, standing, walking and turning: 0 one-frame dark or bright pictures on either side; the presenter thread on in every session of this PR; no device loss or failed present
@Beanstalk test if this helps on windows/amd too
Beanstalk test farm · an earlier run: commit
3569b8d(the newer results are in a comment below)Commit
3569b8d· tier reviewed · approved by HumanBean · dashboardWhat will be tested · AI plan (
gpt-6.1-terra)On Windows AMD, compare main and the PR in Kokiri Forest and the Market over about 23 minutes per tester. The plan measures frame times and FPS with threaded replay enabled, and compares sampled images for rendering regressions. Crashes, device loss, substantial frame-time or FPS regressions, or sufficiently different sampled images fail the plan; brief flashes between samples may be missed.
kokiri-start(settings vsync=2, frame_rate_cap=6, frame_generation=0; switchesRX_VK_ASYNC=1), then teleport to Kokiri Forest, measure forest for 20 s (after 5 s warm-up), take the picture forest to compare with main, teleport to Market, move the stick, measure market for 20 s (after 5 s warm-up), take the picture market to compare with main; 2 runs on each side.Started at 06:45 UTC.
AMD Radeon RX 9070 XT, 2.0.416Measurements · windows+amd
Verdict: fail on windows+amd (1/1 report):
forest-market screenshot.max_difffailed.Hardware: 1: windows+amd+amd-proprietary+dgpu
AMD Radeon RX 9070 XT, 2.0.416AMD Ryzen 9 5900X 12-Core Processor, 32 GB RAMScene: forest-market, from the repository's save kokiri-start (vsync=2, frame_rate_cap=6, frame_generation=0, RX_VK_ASYNC=1); each cell's settings and switches as the table names them
Method: the Beanstalk test farm: main and this PR built and run on the same PC in one session, game starts alternating (ABBA); forest-market: 2 game starts a side, forest 20 s after 5 s warm-up, market 20 s after 5 s warm-up; each number the median of the runs, then of 1 report
How to reproduce: the farm's plan above (
@Beanstalkon this PR runs it again)09cbdd93569b8dFlags and settings
None: the plan varied no setting, switch or build option.
Markdown for the PR description
WIP: perf(vk): the presenter thread on by default with NVIDIA's driver on Linux, with NVIDIA Reflex; carries #1121to perf(vk): the presenter thread on by default with NVIDIA's driver on Linux, with NVIDIA Reflex; carries #1121On the Windows AMD run (RX 9070 XT): the presenter thread was off there (this PR turns it on only with NVIDIA's driver on Linux), so what it measured is #1121's part, the replay taking a frame's jobs as they come. It helps there: Kokiri Forest 73.8 → 80.5 fps (+9%), the Market 66.4 → 73.5 (+11%), the Market's p99 19.9 → 17.5 ms.
The
screenshot.max_difffail is the Market picture rule, not a rendering difference: the two pictures of the failure have the same lighting and nothing missing or dark; Link and the camera are a little further along in one (the stick is held for a set time and the frames fall at other moments). Main against itself already differs by 0.37–0.39 in the Market, over the rule's 0.1, while Kokiri Forest's pictures differ from main no more than main from itself (0.015–0.035 against 0.026–0.040).Next: the presenter thread for everyone. A farm run with
RX_VK_PRESENT_THREAD=1against this PR's default on the setups where it is off, below; if it helps or changes nothing there, the default becomes on for all.🤖 Generated with Claude Code
https://claude.ai/code/session_016o1k1PazAdTJxQQWQBksW2
@Beanstalk compare this PR with
RX_VK_PRESENT_THREAD=1against this PR with its default (the presenter thread off on these) on windows+amd, windows+nvidia and linux+amd: Kokiri Forest and the Market, fps, p99 frame time and the picturesBeanstalk test farm · an earlier run: commit
3569b8d, run 2 (the newer results are in a comment below)Commit
3569b8d· tier reviewed · approved by HumanBean · dashboardWhat will be tested · AI plan (
gpt-6.1-terra)On Windows AMD, Windows NVIDIA and Linux AMD, this plan measures FPS and p99 frame time and compares pictures in Kokiri Forest and the Market. It takes about 21 minutes per tester and tests the presenter thread both off and forced on. Each setting is compared with main, rather than with the PR’s default, because the PR also changes replay when the presenter thread is off. A crash, a frame-time increase over 15%, an FPS drop over 10%, or a picture difference over the threshold fails the plan; sampled pictures cannot reliably detect one-frame flashes.
0or1.kokiri-start, then teleport to Kokiri Forest, measure forest for 20 s (after 5 s warm-up), take the picture forest to compare with main, teleport to Market, measure market for 20 s (after 5 s warm-up), take the picture market to compare with main; 1 run on each side for each of 2 cells (2 game starts a run).Started at 07:21 UTC. Run 2: re-planned by HumanBean.
Estimate: done in about 1 h, from the queue and the testers online now.
AMD Radeon RX 9070 XT, 2.0.416; 2: windows+amd+amd-proprietary+dgpuAMD Radeon RX 9070 XT, 2.0.416NVIDIA GeForce RTX 3050 Ti Laptop GPU, 621.24.0; 2: windows+nvidia+nvidia-proprietary+dgpuNVIDIA GeForce RTX 3070, 621.24.0AMD Radeon RX 7900 XTX (RADV NAVI31), 26.2.7RX_VK_PRESENT_THREAD=0RX_VK_PRESENT_THREAD=0RX_VK_PRESENT_THREAD=0RX_VK_PRESENT_THREAD=1RX_VK_PRESENT_THREAD=1RX_VK_PRESENT_THREAD=1Measurements · windows+amd
Verdict: fail on windows+amd (2/1 report):
forest-market screenshot.max_difffailed.Hardware: 1: windows+amd+amd-proprietary+dgpu
AMD Radeon RX 9070 XT, 2.0.416AMD Ryzen 9 5900X 12-Core Processor, 32 GB RAM; 2: windows+amd+amd-proprietary+dgpuAMD Radeon RX 9070 XT, 2.0.416AMD Ryzen 7 5800X3D 8-Core Processor, 32 GB RAMScene: forest-market, from the repository's save kokiri-start; each cell's settings and switches as the table names them
Method: the Beanstalk test farm: main and this PR built and run on the same PC in one session, game starts alternating (ABBA); forest-market: 1 game start a side, forest 20 s after 5 s warm-up, market 20 s after 5 s warm-up; each number the median of the runs, then of 2 reports
How to reproduce: the farm's plan above (
@Beanstalkon this PR runs it again)09cbdd93569b8dRX_VK_PRESENT_THREAD=0· forest · frame rateRX_VK_PRESENT_THREAD=0· market · frame rateRX_VK_PRESENT_THREAD=0· forest · frame time (p99) (ms)RX_VK_PRESENT_THREAD=0· market · frame time (p99) (ms)RX_VK_PRESENT_THREAD=1· forest · frame rateRX_VK_PRESENT_THREAD=1· market · frame rateRX_VK_PRESENT_THREAD=1· forest · frame time (p99) (ms)RX_VK_PRESENT_THREAD=1· market · frame time (p99) (ms)Flags and settings
RX_VK_PRESENT_THREAD0,1Markdown for the PR description
Measurements · windows+nvidia
Verdict: fail on windows+nvidia (2/1 report):
forest-market screenshot.max_difffailed.Hardware: 1: windows+nvidia+nvidia-proprietary+dgpu
NVIDIA GeForce RTX 3050 Ti Laptop GPU, 621.24.0AMD Ryzen 7 5800H with Radeon Graphics, 16 GB RAM; 2: windows+nvidia+nvidia-proprietary+dgpuNVIDIA GeForce RTX 3070, 621.24.0AMD Ryzen 5 5600X3D 6-Core Processor, 32 GB RAMScene: forest-market, from the repository's save kokiri-start; each cell's settings and switches as the table names them
Method: the Beanstalk test farm: main and this PR built and run on the same PC in one session, game starts alternating (ABBA); forest-market: 1 game start a side, forest 20 s after 5 s warm-up, market 20 s after 5 s warm-up; each number the median of the runs, then of 2 reports
How to reproduce: the farm's plan above (
@Beanstalkon this PR runs it again)09cbdd93569b8dRX_VK_PRESENT_THREAD=0· forest · frame rateRX_VK_PRESENT_THREAD=0· market · frame rateRX_VK_PRESENT_THREAD=0· forest · frame time (p99) (ms)RX_VK_PRESENT_THREAD=0· market · frame time (p99) (ms)RX_VK_PRESENT_THREAD=1· forest · frame rateRX_VK_PRESENT_THREAD=1· market · frame rateRX_VK_PRESENT_THREAD=1· forest · frame time (p99) (ms)RX_VK_PRESENT_THREAD=1· market · frame time (p99) (ms)Flags and settings
RX_VK_PRESENT_THREAD0,1Markdown for the PR description
Measurements · linux+amd
Verdict: fail on linux+amd (1/1 report):
forest-market screenshot.max_difffailed.Hardware: 1: linux+amd+radv+dgpu
AMD Radeon RX 7900 XTX (RADV NAVI31), 26.2.7AMD Ryzen 7 5800X3D 8-Core Processor, 64 GB RAMScene: forest-market, from the repository's save kokiri-start; each cell's settings and switches as the table names them
Method: the Beanstalk test farm: main and this PR built and run on the same PC in one session, game starts alternating (ABBA); forest-market: 1 game start a side, forest 20 s after 5 s warm-up, market 20 s after 5 s warm-up; each number the median of the runs, then of 1 report
How to reproduce: the farm's plan above (
@Beanstalkon this PR runs it again)09cbdd93569b8dRX_VK_PRESENT_THREAD=0· forest · frame rateRX_VK_PRESENT_THREAD=0· market · frame rateRX_VK_PRESENT_THREAD=0· forest · frame time (p99) (ms)RX_VK_PRESENT_THREAD=0· market · frame time (p99) (ms)RX_VK_PRESENT_THREAD=1· forest · frame rateRX_VK_PRESENT_THREAD=1· market · frame rateRX_VK_PRESENT_THREAD=1· forest · frame time (p99) (ms)RX_VK_PRESENT_THREAD=1· market · frame time (p99) (ms)Flags and settings
RX_VK_PRESENT_THREAD0,1Markdown for the PR description
@Beanstalk compare main, this PR with its default, and this PR with
RX_VK_PRESENT_THREAD=1, on windows+amd, windows+nvidia, linux+amd, steamdeck, windows+intel and linux+intel: Kokiri Forest and the Market, standing and walking with the camera turning; fps, p99 frame time, the pictures against main, and the frame log (RX_VK_FRAMELOG) for one-frame dark pictures and pictures that go dark and stay dark (this replaces the request above: Steam Deck and Intel added, Intel being where #1121 saw dark pictures)@Beanstalk with the frame rate uncapped (settings frame_rate_cap=0, vsync=0: the last run sat at the 60 cap on every side), compare main, this PR with its default, and this PR with
RX_VK_PRESENT_THREAD=1, on windows+amd, windows+nvidia, linux+amd, steamdeck, windows+intel and linux+intel: Kokiri Forest and the Market, standing and walking with the camera turning, 2 runs a side; fps, p99 frame time, the pictures against main, and the frame log (RX_VK_FRAMELOG) for one-frame dark pictures and pictures that go dark and stay darkBeanstalk test farm · an earlier run: commit
3569b8d, run 3 (the newer results are in a comment below)Commit
3569b8d· tier reviewed · approved by HumanBean · dashboardWhat will be tested · AI plan (
gpt-6.1-terra)Two runs per side visit Kokiri Forest and the Market on the six requested hardware targets, comparing main with the PR default and with the presenter thread forced on; Intel targets wait for matching testers. The presenter-thread axis isolates the forced-on configuration, and the settings use the actual uncapped frame-rate and V-Sync-off values rather than the requested numeric values, which select 60 FPS and Auto. The runs take about 45 minutes per tester and measure FPS and p99 frame time, with pictures compared against main; crashes, substantial performance regressions, or picture differences fail. The farm cannot use the requested frame log, so sampled pictures cannot reliably detect single-frame flashes or every persistent darkening; walking and camera turning are exercised separately. No tester online now matches windows+intel, linux+intel: those targets wait for one.
1(PR only).kokiri-start(settings vsync=2, frame_rate_cap=6), then teleport to Kokiri Forest, measure kokiri-standing for 15 s (after 5 s warm-up), take the picture kokiri-standing to compare with main, move the stick, measure kokiri-after-motion for 10 s (after 0 s warm-up), take the picture kokiri-after-motion to compare with main, teleport to Market, measure market-standing for 15 s (after 5 s warm-up), take the picture market-standing to compare with main, measure market-after-motion for 10 s (after 0 s warm-up), take the picture market-after-motion to compare with main; 2 runs on each side.kokiri-start(settings vsync=2, frame_rate_cap=6), then teleport to Kokiri Forest, measure kokiri-standing for 15 s (after 5 s warm-up), take the picture kokiri-standing to compare with main, move the stick, measure kokiri-after-motion for 10 s (after 0 s warm-up), take the picture kokiri-after-motion to compare with main, teleport to Market, measure market-standing for 15 s (after 5 s warm-up), take the picture market-standing to compare with main, measure market-after-motion for 10 s (after 0 s warm-up), take the picture market-after-motion to compare with main; 2 runs on each side.Started at 08:41 UTC. Run 3: re-planned by HumanBean.
⚠ No tester has a PC for
linux+intel(none seen in 14 days): it waits until one signs in.AMD Radeon RX 9070 XT, 2.0.416; 2: windows+amd+amd-proprietary+dgpuAMD Radeon RX 9070 XT, 2.0.416NVIDIA GeForce RTX 3050 Ti Laptop GPU, 621.24.0; 2: windows+nvidia+nvidia-proprietary+dgpuNVIDIA GeForce RTX 3090, 621.24.0AMD Radeon RX 7900 XTX (RADV NAVI31), 26.2.7; 2: linux+amd+radv+igpu+handheldAMD Custom GPU 0405 (RADV VANGOGH), 26.1.99AMD Custom GPU 0405 (RADV VANGOGH), 26.1.99Measurements · windows+amd
Verdict: fail on windows+amd (2/1 report):
forced-presenter screenshot.max_diff;default-presenter screenshot.max_difffailed.Hardware: 1: windows+amd+amd-proprietary+dgpu
AMD Radeon RX 9070 XT, 2.0.416AMD Ryzen 7 5800X3D 8-Core Processor, 32 GB RAM; 2: windows+amd+amd-proprietary+dgpuAMD Radeon RX 9070 XT, 2.0.416AMD Ryzen 9 5900X 12-Core Processor, 32 GB RAMScene: default-presenter, from the repository's save kokiri-start (vsync=2, frame_rate_cap=6); forced-presenter, from the repository's save kokiri-start (vsync=2, frame_rate_cap=6); each cell's settings and switches as the table names them
Method: the Beanstalk test farm: main and this PR built and run on the same PC in one session, game starts alternating (ABBA); default-presenter: 2 game starts a side, kokiri-standing 15 s after 5 s warm-up, kokiri-after-motion 10 s after 0 s warm-up, market-standing 15 s after 5 s warm-up, market-after-motion 10 s after 0 s warm-up; forced-presenter: 2 game starts a side, kokiri-standing 15 s after 5 s warm-up, kokiri-after-motion 10 s after 0 s warm-up, market-standing 15 s after 5 s warm-up, market-after-motion 10 s after 0 s warm-up; each number the median of the runs, then of 2 reports
How to reproduce: the farm's plan above (
@Beanstalkon this PR runs it again)09cbdd93569b8dRX_VK_PRESENT_THREAD=1 (PR only)· kokiri-standing · frame rateRX_VK_PRESENT_THREAD=1 (PR only)· market-standing · frame rateRX_VK_PRESENT_THREAD=1 (PR only)· kokiri-after-motion · frame rateRX_VK_PRESENT_THREAD=1 (PR only)· market-after-motion · frame rateRX_VK_PRESENT_THREAD=1 (PR only)· kokiri-standing · frame time (p99) (ms)RX_VK_PRESENT_THREAD=1 (PR only)· market-standing · frame time (p99) (ms)RX_VK_PRESENT_THREAD=1 (PR only)· kokiri-after-motion · frame time (p99) (ms)RX_VK_PRESENT_THREAD=1 (PR only)· market-after-motion · frame time (p99) (ms)RX_VK_PRESENT_THREAD=1 (PR only)· forbidden log linesFlags and settings
RX_VK_PRESENT_THREAD1Markdown for the PR description
Measurements · windows+nvidia
Verdict: fail on windows+nvidia (2/1 report):
default-presenter screenshot.max_diff;forced-presenter screenshot.max_difffailed.Hardware: 1: windows+nvidia+nvidia-proprietary+dgpu
NVIDIA GeForce RTX 3050 Ti Laptop GPU, 621.24.0AMD Ryzen 7 5800H with Radeon Graphics, 16 GB RAM; 2: windows+nvidia+nvidia-proprietary+dgpuNVIDIA GeForce RTX 3090, 621.24.0AMD Ryzen 9 5900X 12-Core Processor, 128 GB RAMScene: default-presenter, from the repository's save kokiri-start (vsync=2, frame_rate_cap=6); forced-presenter, from the repository's save kokiri-start (vsync=2, frame_rate_cap=6); each cell's settings and switches as the table names them
Method: the Beanstalk test farm: main and this PR built and run on the same PC in one session, game starts alternating (ABBA); default-presenter: 2 game starts a side, kokiri-standing 15 s after 5 s warm-up, kokiri-after-motion 10 s after 0 s warm-up, market-standing 15 s after 5 s warm-up, market-after-motion 10 s after 0 s warm-up; forced-presenter: 2 game starts a side, kokiri-standing 15 s after 5 s warm-up, kokiri-after-motion 10 s after 0 s warm-up, market-standing 15 s after 5 s warm-up, market-after-motion 10 s after 0 s warm-up; each number the median of the runs, then of 2 reports
How to reproduce: the farm's plan above (
@Beanstalkon this PR runs it again)09cbdd93569b8dRX_VK_PRESENT_THREAD=1 (PR only)· kokiri-standing · frame rateRX_VK_PRESENT_THREAD=1 (PR only)· market-standing · frame rateRX_VK_PRESENT_THREAD=1 (PR only)· kokiri-after-motion · frame rateRX_VK_PRESENT_THREAD=1 (PR only)· market-after-motion · frame rateRX_VK_PRESENT_THREAD=1 (PR only)· kokiri-standing · frame time (p99) (ms)RX_VK_PRESENT_THREAD=1 (PR only)· market-standing · frame time (p99) (ms)RX_VK_PRESENT_THREAD=1 (PR only)· kokiri-after-motion · frame time (p99) (ms)RX_VK_PRESENT_THREAD=1 (PR only)· market-after-motion · frame time (p99) (ms)RX_VK_PRESENT_THREAD=1 (PR only)· forbidden log linesFlags and settings
RX_VK_PRESENT_THREAD1Markdown for the PR description
Measurements · linux+amd
Verdict: fail on linux+amd (2/1 report):
default-presenter screenshot.max_diff;forced-presenter screenshot.max_diff;forced-presenter kokiri-standing.frametime_p99_ms.max_delta_pct (+50.2%)failed.Hardware: 1: linux+amd+radv+dgpu
AMD Radeon RX 7900 XTX (RADV NAVI31), 26.2.7AMD Ryzen 7 5800X3D 8-Core Processor, 64 GB RAM; 2: linux+amd+radv+igpu+handheldAMD Custom GPU 0405 (RADV VANGOGH), 26.1.99AMD Custom APU 0405, 12 GB RAMScene: default-presenter, from the repository's save kokiri-start (vsync=2, frame_rate_cap=6); forced-presenter, from the repository's save kokiri-start (vsync=2, frame_rate_cap=6); each cell's settings and switches as the table names them
Method: the Beanstalk test farm: main and this PR built and run on the same PC in one session, game starts alternating (ABBA); default-presenter: 2 game starts a side, kokiri-standing 15 s after 5 s warm-up, kokiri-after-motion 10 s after 0 s warm-up, market-standing 15 s after 5 s warm-up, market-after-motion 10 s after 0 s warm-up; forced-presenter: 2 game starts a side, kokiri-standing 15 s after 5 s warm-up, kokiri-after-motion 10 s after 0 s warm-up, market-standing 15 s after 5 s warm-up, market-after-motion 10 s after 0 s warm-up; each number the median of the runs, then of 2 reports
How to reproduce: the farm's plan above (
@Beanstalkon this PR runs it again)09cbdd93569b8dRX_VK_PRESENT_THREAD=1 (PR only)· kokiri-standing · frame rateRX_VK_PRESENT_THREAD=1 (PR only)· market-standing · frame rateRX_VK_PRESENT_THREAD=1 (PR only)· kokiri-after-motion · frame rateRX_VK_PRESENT_THREAD=1 (PR only)· market-after-motion · frame rateRX_VK_PRESENT_THREAD=1 (PR only)· kokiri-standing · frame time (p99) (ms)RX_VK_PRESENT_THREAD=1 (PR only)· market-standing · frame time (p99) (ms)RX_VK_PRESENT_THREAD=1 (PR only)· kokiri-after-motion · frame time (p99) (ms)RX_VK_PRESENT_THREAD=1 (PR only)· market-after-motion · frame time (p99) (ms)RX_VK_PRESENT_THREAD=1 (PR only)· forbidden log linesFlags and settings
RX_VK_PRESENT_THREAD1Markdown for the PR description
Measurements · steamdeck
Verdict: fail on steamdeck (1/1 report):
default-presenter screenshot.max_difffailed.Hardware: 1: linux+amd+radv+igpu+handheld
AMD Custom GPU 0405 (RADV VANGOGH), 26.1.99AMD Custom APU 0405, 12 GB RAMScene: default-presenter, from the repository's save kokiri-start (vsync=2, frame_rate_cap=6); forced-presenter, from the repository's save kokiri-start (vsync=2, frame_rate_cap=6); each cell's settings and switches as the table names them
Method: the Beanstalk test farm: main and this PR built and run on the same PC in one session, game starts alternating (ABBA); default-presenter: 2 game starts a side, kokiri-standing 15 s after 5 s warm-up, kokiri-after-motion 10 s after 0 s warm-up, market-standing 15 s after 5 s warm-up, market-after-motion 10 s after 0 s warm-up; forced-presenter: 2 game starts a side, kokiri-standing 15 s after 5 s warm-up, kokiri-after-motion 10 s after 0 s warm-up, market-standing 15 s after 5 s warm-up, market-after-motion 10 s after 0 s warm-up; each number the median of the runs, then of 1 report
How to reproduce: the farm's plan above (
@Beanstalkon this PR runs it again)09cbdd93569b8dRX_VK_PRESENT_THREAD=1 (PR only)· kokiri-standing · frame rateRX_VK_PRESENT_THREAD=1 (PR only)· market-standing · frame rateRX_VK_PRESENT_THREAD=1 (PR only)· kokiri-after-motion · frame rateRX_VK_PRESENT_THREAD=1 (PR only)· market-after-motion · frame rateRX_VK_PRESENT_THREAD=1 (PR only)· kokiri-standing · frame time (p99) (ms)RX_VK_PRESENT_THREAD=1 (PR only)· market-standing · frame time (p99) (ms)RX_VK_PRESENT_THREAD=1 (PR only)· kokiri-after-motion · frame time (p99) (ms)RX_VK_PRESENT_THREAD=1 (PR only)· market-after-motion · frame time (p99) (ms)RX_VK_PRESENT_THREAD=1 (PR only)· forbidden log linesFlags and settings
RX_VK_PRESENT_THREAD1Markdown for the PR description
3569b8da6a0c3808483cBeanstalk test farm · an earlier run: commit
0c38084(the newer results are in a comment below)Commit
0c38084· tier reviewed · approved by HumanBean · dashboardWhat will be tested · AI plan (
gpt-6.1-terra)Compare the PR with main on Windows AMD and NVIDIA PCs, with Windows and Linux Intel targets also queued; no Intel tester is currently online. Two starts per side visit Kokiri Forest and Castle Town Market with V-Sync off and the frame rate uncapped, measuring standing and post-movement performance over roughly 22 minutes per tester. Crashes, device loss, missing presenter-mode logs, substantial FPS or frame-time regressions, and large picture differences fail the plan. The shared log rule accepts either presenter mode, so it cannot independently enforce AMD-on and NVIDIA-off; the reported logs show which mode was selected. Snapshot comparisons also cannot reliably catch one-frame flashes. No tester online now matches windows+intel, linux+intel: those targets wait for one.
kokiri-start(settings vsync=2, frame_rate_cap=6), then teleport to Kokiri Forest, measure forest-standing for 15 s (after 5 s warm-up), take the picture forest-standing to compare with main, move the stick, measure forest-after-moving for 10 s (after 0 s warm-up), take the picture forest-after-moving to compare with main, teleport to Market, measure market-standing for 15 s (after 5 s warm-up), take the picture market-standing to compare with main, measure market-after-moving for 10 s (after 0 s warm-up), take the picture market-after-moving to compare with main; 2 runs on each side.Started at 11:56 UTC.
⚠ No tester has a PC for
linux+intel(none seen in 14 days): it waits until one signs in.AMD Radeon RX 9070 XT, 2.0.416; 2: windows+amd+amd-proprietary+dgpuAMD Radeon RX 9070 XT, 2.0.416NVIDIA GeForce RTX 3050 Ti Laptop GPU, 621.24.0; 2: windows+nvidia+nvidia-proprietary+dgpuNVIDIA GeForce RTX 3090, 621.24.0Measurements · windows+amd
Verdict: could not be judged on windows+amd (2/1 report).
Hardware: 1: windows+amd+amd-proprietary+dgpu
AMD Radeon RX 9070 XT, 2.0.416AMD Ryzen 9 5900X 12-Core Processor, 32 GB RAM; 2: windows+amd+amd-proprietary+dgpuAMD Radeon RX 9070 XT, 2.0.416AMD Ryzen 7 5800X3D 8-Core Processor, 32 GB RAMScene: forest-market, from the repository's save kokiri-start (vsync=2, frame_rate_cap=6); each cell's settings and switches as the table names them
Method: the Beanstalk test farm: main and this PR built and run on the same PC in one session, game starts alternating (ABBA); forest-market: 2 game starts a side, forest-standing 15 s after 5 s warm-up, forest-after-moving 10 s after 0 s warm-up, market-standing 15 s after 5 s warm-up, market-after-moving 10 s after 0 s warm-up; each number the median of the runs, then of 2 reports
How to reproduce: the farm's plan above (
@Beanstalkon this PR runs it again)Nothing was measured: no metric came back.
Flags and settings
None: the plan varied no setting, switch or build option.
Markdown for the PR description
Measurements · windows+nvidia
Verdict: could not be judged on windows+nvidia (2/1 report).
Hardware: 1: windows+nvidia+nvidia-proprietary+dgpu
NVIDIA GeForce RTX 3050 Ti Laptop GPU, 621.24.0AMD Ryzen 7 5800H with Radeon Graphics, 16 GB RAM; 2: windows+nvidia+nvidia-proprietary+dgpuNVIDIA GeForce RTX 3090, 621.24.0AMD Ryzen 9 5900X 12-Core Processor, 128 GB RAMScene: forest-market, from the repository's save kokiri-start (vsync=2, frame_rate_cap=6); each cell's settings and switches as the table names them
Method: the Beanstalk test farm: main and this PR built and run on the same PC in one session, game starts alternating (ABBA); forest-market: 2 game starts a side, forest-standing 15 s after 5 s warm-up, forest-after-moving 10 s after 0 s warm-up, market-standing 15 s after 5 s warm-up, market-after-moving 10 s after 0 s warm-up; each number the median of the runs, then of 2 reports
How to reproduce: the farm's plan above (
@Beanstalkon this PR runs it again)Nothing was measured: no metric came back.
Flags and settings
None: the plan varied no setting, switch or build option.
Markdown for the PR description
perf(vk): the presenter thread on by default with NVIDIA's driver on Linux, with NVIDIA Reflex; carries #1121to perf(vk): the presenter thread on by default with NVIDIA on Linux and AMD on Windows, with NVIDIA Reflex; carries #1121@Beanstalk run 4: the last commit turns the presenter thread on by default with AMD's Windows driver on a discrete card (run 3's forced cell). Please test this PR against main on windows+amd and windows+nvidia (NVIDIA should be unchanged from run 3: presenter thread off there), with the settings vsync=2 (Off) and frame_rate_cap=6 (Uncapped): Kokiri Forest and Castle Town Market, standing 15 s and 10 s after moving, two game starts a side. No switches: the defaults only. Expect
vk: presenter thread: each present made on a threadin the log on AMD andoff (RX_VK_PRESENT_THREAD=1: on)on NVIDIA. Intel targets from run 3 can stay queued.0c3808483ca93a983fbaBeanstalk test farm
Commit
28f5fb2· tier reviewed · approved by HumanBean · dashboardWhat will be tested · AI plan (
gpt-6.1-terra)Compare the PR with main on Windows AMD and NVIDIA PCs, with V-Sync off and the frame rate uncapped. Each of two starts per side visits Kokiri Forest and Castle Town Market, measuring 15 seconds standing and 10 seconds after moving at each location; the run takes about 22 minutes per tester. The PR fails on a crash, device loss, a sampled picture difference, or FPS below main in any segment. The log check accepts either requested presenter-thread message, but cannot enforce the AMD and NVIDIA expectations separately in one shared plan.
kokiri-start(settings vsync=2, frame_rate_cap=6), then teleport to Kokiri Forest, measure forest-standing for 15 s (after 5 s warm-up), take the picture forest to compare with main, move the stick, measure forest-after-moving for 10 s (after 0 s warm-up), teleport to Market, measure market-standing for 15 s (after 5 s warm-up), take the picture market to compare with main, measure market-after-moving for 10 s (after 0 s warm-up); 2 runs on each side.Done (started 16:18 UTC, last report 18:55 UTC). Comment
@Beanstalkto run it again.AMD Radeon RX 7900 XTX, 2.0.416; 2: windows+amd+amd-proprietary+igpuAMD Radeon(TM) 880M Graphics, 2.0.416NVIDIA GeForce RTX 3050 Ti Laptop GPU, 621.24.0; 2: windows+nvidia+nvidia-proprietary+dgpuNVIDIA GeForce RTX 3090, 621.24.0Measurements · windows+amd
Verdict: could not be judged on windows+amd (2/1 report).
Hardware: 1: windows+amd+amd-proprietary+dgpu
AMD Radeon RX 7900 XTX, 2.0.416AMD Ryzen 7 5800X3D 8-Core Processor, 32 GB RAM; 2: windows+amd+amd-proprietary+igpuAMD Radeon(TM) 880M Graphics, 2.0.416AMD Ryzen AI 9 365 w/ Radeon 880M, 32 GB RAMScene: forest-and-market, from the repository's save kokiri-start (vsync=2, frame_rate_cap=6); each cell's settings and switches as the table names them
Method: the Beanstalk test farm: main and this PR built and run on the same PC in one session, game starts alternating (ABBA); forest-and-market: 2 game starts a side, forest-standing 15 s after 5 s warm-up, forest-after-moving 10 s after 0 s warm-up, market-standing 15 s after 5 s warm-up, market-after-moving 10 s after 0 s warm-up; each number the median of the runs, then of 2 reports
How to reproduce: the farm's plan above (
@Beanstalkon this PR runs it again)Nothing was measured: no metric came back.
Flags and settings
None: the plan varied no setting, switch or build option.
Markdown for the PR description
Measurements · windows+nvidia
Verdict: fail on windows+nvidia (2/1 report):
forest-and-market log_expected_missing.maxfailed.Hardware: 1: windows+nvidia+nvidia-proprietary+dgpu
NVIDIA GeForce RTX 3050 Ti Laptop GPU, 621.24.0AMD Ryzen 7 5800H with Radeon Graphics, 16 GB RAM; 2: windows+nvidia+nvidia-proprietary+dgpuNVIDIA GeForce RTX 3090, 621.24.0AMD Ryzen 9 5900X 12-Core Processor, 128 GB RAMScene: forest-and-market, from the repository's save kokiri-start (vsync=2, frame_rate_cap=6); each cell's settings and switches as the table names them
Method: the Beanstalk test farm: main and this PR built and run on the same PC in one session, game starts alternating (ABBA); forest-and-market: 2 game starts a side, forest-standing 15 s after 5 s warm-up, forest-after-moving 10 s after 0 s warm-up, market-standing 15 s after 5 s warm-up, market-after-moving 10 s after 0 s warm-up; each number the median of the runs, then of 2 reports
How to reproduce: the farm's plan above (
@Beanstalkon this PR runs it again)d5794c828f5fb2Flags and settings
None: the plan varied no setting, switch or build option.
Markdown for the PR description
Run 4 measured nothing: main's base failed to build on Windows on every tester (
gpu/vk/descriptor_array_census.h: a member namednear, which the Windows headers define away, since #475). #1729 fixes it. This PR is rebased on main868f6be(still without that fix); the Windows AMD default goes back to the farm once #1729 is in.@Beanstalk run 5 (run 4 failed: main didn't build on Windows; #1729 is merged and this PR is rebased on it). The last commit turns the presenter thread on by default with AMD's Windows driver on a discrete card. Please test this PR against main on windows+amd and windows+nvidia, settings vsync=2 (Off) and frame_rate_cap=6 (Uncapped), no switches: Kokiri Forest and Castle Town Market, standing 15 s and 10 s after moving, two game starts a side. Expect
vk: presenter thread: each present made on a threadin the log on AMD andoff (RX_VK_PRESENT_THREAD=1: on)on NVIDIA, and fps at least main's on both.a93a983fba08ef67ccd308ef67ccd328f5fb2b1eperf(vk): the presenter thread on by default with NVIDIA on Linux and AMD on Windows, with NVIDIA Reflex; carries #1121to perf(vk): the presenter thread on by default with NVIDIA on Linux and AMD on Windows, with NVIDIA Reflex; after #1121@Beanstalk run 5, again for the new head
28f5fb2(the same code: #1121 now sits under it as its own PR; run 4 failed because main did not build on Windows, fixed by #1729). The last commit turns the presenter thread on by default with AMD's Windows driver on a discrete card. Please test this PR against main on windows+amd and windows+nvidia, settings vsync=2 (Off) and frame_rate_cap=6 (Uncapped), no switches: Kokiri Forest and Castle Town Market, standing 15 s and 10 s after moving, two game starts a side. Expectvk: presenter thread: each present made on a threadin the log on AMD andoff (RX_VK_PRESENT_THREAD=1: on)on NVIDIA, and fps at least main's on both.Replaced by #1763, which carries this PR's presenter thread (and its Reflex present id) as part of one rewrite of the renderer's threads: the presenter on every PC, on the device's own present queue, with the recorder thread and the parallel replay removed. Its farm run decides the presenter per platform from there; closing this one so there is a single PR.
Pull request closed