perf(vk): the renderer's threads rewritten: one render thread, the presenter on every PC, parts that end at the game's own pass ends #1763
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
5 participants
Notifications
Due date
No due date set.
Dependencies
No dependencies set
Reference
MagicBeanDev/Longshot!1763
Loading…
Reference in a new issue
No description provided.
Delete branch "HumanBean/Longshot:perf/thread-redesign"
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?
Summary
The renderer's threads, rewritten as one design for every PC, with no old path kept:
RX_VK_GPU_TIMELINE=1(each recording's submit and GPU span, the jobs, the game's fence waits, the presents),RX_VK_PASS_TIMING=3(each command outside a named span timed as glue), and the log's new textures by kind.Read first:
gpu/vk/vk_replay.cpp(maybe_split,presenter_loop,present_thread_on,readback_late_wanted),gpu/vk/threaded_rendering.h,gpu/vk/vk_texture.cpp(pass_start_barrier,full_barrier),docs/dev/reference/threads.md.Testing
Tested: CachyOS Linux, RTX 3090 (NVIDIA 620.70.06), Ryzen 9 7900X: the previous head (
a2cc7b5, the same code before this rebase onto main's docs and tools revamp) in the game through Kokiri Forest, Hyrule Field and Castle Town Market, 3 interleaved rounds against main plus a 15 s Market walk, no crash, no Vulkan errors, pictures as main's (the differences are fireflies, Navi and the upscaler's noise). The lab (RX 580 / i7-9700K, RADV, 3 rounds): even with main where the GPU is 85-94% busy, the same draws per frame. The rebased head (21b606e) builds on Linux (dev.py vk) and passes a Windows llvm-mingw syntax check of every changed renderer file; it has not run in the game yet here (main's MKWVK_API_VERSION 35 needs a new base).mise run test: not run on this PC (the maintainer's rule); CI runs it.Still to test:
RX_VK_PART_MIN=240 RX_VK_PART_MAX=240,RX_VK_PASS_BARRIER=full,RX_VK_READBACK_LATE=0), expecting one of them to recover it or none, which points at the presenter thread (@Beanstalk).a2cc7b5(by hand, once the base is built).Measurements
Verdict: faster on NVIDIA and in most scenes on fast AMD cards, even where the GPU is the limit, and slower in the Market standing on fast AMD cards (under investigation, above).
Setup: CachyOS Linux, RTX 3090 (NVIDIA 620.70.06), Ryzen 9 7900X; 2560x1440, DLSS Quality, uncapped, Reflex on; each area 15 s standing, 5 s walking, 20 s turning; the renderer built alone on one base (main at
9ea65c7and this PR ata2cc7b5), a warm-up each, 3 rounds interleaved on the PC's timing queue; fps from the renderer's present counts, the median of the rounds.9ea65c7a2cc7b5The Test Farm on
a2cc7b5(one report each, main against this PR): RX 7900 XTX (Linux) Kokiri +14% / +17%, Hyrule Field +14% / +30%, Market +8% turning and -24.5% standing; RX 9070 XT (Windows) Kokiri +11% / +14%, Hyrule Field -2% / -5%, Market +5% turning and -14.5% standing; Radeon 760M (Linux) within its noise. The switches this adds or changes ship at the defaults these runs chose (RX_VK_PART_MIN120,RX_VK_PART_MAX512: a fixed 30, 60, 120, 960 or no parts were each slower than these).HOS_PROFILE_DIRandRX_VK_ASYNCare not new: their rows changed (RX_VK_GPU_TIMELINEwrites intoHOS_PROFILE_DIR;RX_VK_ASYNC's default says the one 4-core tier), and their defaults are as before. The numbers were taken on main9ea65c7; this head is on main068fb8eb, whose changes to the renderer since are comments and the version files, nothing the frame runs, so they still hold (the Test Farm measures this head against main again).Risks
What could break: the Market standing on fast AMD cards (measured, above); a present problem on a driver not yet run with the presenter thread on (Intel); flashes on the Deck from the newest submission's readbacks read late. Turning it off:
RX_VK_PART_MIN=240 RX_VK_PART_MAX=240(the old parts),RX_VK_PASS_BARRIER=full,RX_VK_READBACK_LATE=0; the presenter thread and the removed threads have no switch, so a revert is the way back. What a player notices: nothing in their files;RX_VK_RECORDER*,RX_VK_PARALLEL*,RX_VK_PRESENT_THREADandRX_VK_CHUNKin alongshot_env.txtdo nothing now (their rows are GONE).Checklist
Ready for review
main; the title is a Conventional Commit.a2cc7b5, in the game; the rebased head not yet: it needs a new base for main's MKWVK_API_VERSION 35.)mise run testpasses (any failure that also fails onmainis named). (Not run on this PC; CI runs it.)main, or Measurements says why there is nothing to measure.third_party/untouched; nothing GPL-incompatible.When it applies
gpu/vk/switches.def.tests/that reads changed source is updated, not weakened.Assisted-by:/Co-Authored-By:trailers.Ready to merge
main, measured again or a comment says why the numbers still hold.🤖 Generated with Claude Code
https://claude.ai/code/session_017vw8e3aEPBsd7QfUkaatjP
Tested in game: CachyOS Linux, RTX 3090 (NVIDIA 620.70.06), Ryzen 9 7900X: Kokiri Forest, Hyrule Field and the Market with Auto's render thread and segment workers and the presenter thread; no error
@Beanstalk please test this PR against main on every platform: windows+amd, windows+nvidia, windows+intel, linux+amd, linux+nvidia, linux+intel, the steamdeck, and amd+igpu. Settings vsync=2 (Off) and frame_rate_cap=6 (Uncapped), defaults otherwise (Threaded rendering Auto). From the
kokiri-startsave: Kokiri Forest, Hyrule Field and Castle Town Market, in each 20 s standing then 15 s turning the camera, two game starts a side. The segment workers' count is measured in play during the first ~30 s of play (logged asvk: segment workers: N kept), so please run each area long enough after loading. Measure fps, p99 frame time and stutters, and compare pictures with main. Also run one extra cell on the PR side withRX_VK_PAR_CALIBRATE=0(the ceiling, no calibration) to separate the calibration from the rest. Expect fps at least main's everywhere, no crash, and no new picture differences beyond main's own noise. Logs to look for:vk: Threaded rendering:,vk: segment workers run on logical processors,vk: presenter thread:.Beanstalk test farm · an earlier run: commit
9259257(the newer results are in a comment below)Commit
9259257· tier reviewed · approved by HumanBean · dashboardWhat will be tested · AI plan (
gpt-6.1-terra)Two starts per side will visit Kokiri Forest, Hyrule Field and the Market, with V-Sync off and the frame-rate cap uncapped. Each area gets a 20-second standing measurement, a camera turn and a 15-second post-turn measurement; FPS, p99 frame time, stutters, worker-count logs and pictures are compared with main. An extra PR-only calibration-off cell separates the worker ceiling from the default calibration; the plan takes roughly 40 minutes per tester. Crashes, lower FPS, increased p99 or stutters, and picture differences above the comparison tolerance fail the plan. Windows and Linux Intel targets currently have no online tester and will wait for one. No tester online now matches windows+intel, linux+intel: those targets wait for one.
0(PR only).kokiri-start(settings vsync=2, frame_rate_cap=6), then teleport to Kokiri Forest, measure forest-standing for 20 s (after 0 s warm-up), take the picture forest-standing to compare with main, move the stick, measure forest-after-turn for 15 s (after 0 s warm-up), take the picture forest-after-turn to compare with main, teleport to Hyrule Field, measure field-standing for 20 s (after 0 s warm-up), take the picture field-standing to compare with main, measure field-after-turn for 15 s (after 0 s warm-up), take the picture field-after-turn to compare with main, teleport to Market, measure market-standing for 20 s (after 0 s warm-up), take the picture market-standing to compare with main, measure market-after-turn for 15 s (after 0 s warm-up), take the picture market-after-turn to compare with main; 2 runs on each side.Started at 17:35 UTC.
Estimate: done in about 2 h 30 min, from the queue and the testers online now; some jobs can't start yet.
âš No tester has a PC for
linux+intel(none seen in 14 days): it waits until one signs in.NVIDIA GeForce RTX 3050 Ti Laptop GPU, 621.24.0Measurements · windows+nvidia
Verdict: running on windows+nvidia (1/1 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 RAMScene: three-areas, 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); three-areas: 2 game starts a side, forest-standing 20 s after 0 s warm-up, forest-after-turn 15 s after 0 s warm-up, field-standing 20 s after 0 s warm-up, field-after-turn 15 s after 0 s warm-up, market-standing 20 s after 0 s warm-up, market-after-turn 15 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)Nothing was measured: no metric came back.
Flags and settings
RX_VK_PAR_CALIBRATE0Markdown for the PR description
Beanstalk test farm · an earlier run: commit
85c0d23(the newer results are in a comment below)Commit
85c0d23· tier reviewed · approved by HumanBean · dashboardWhat will be tested · AI plan (
gpt-6.1-terra)Two starts per side visit Kokiri Forest, Hyrule Field and Castle Town Market with V-Sync off and the frame rate uncapped. Each area is measured standing and after a camera turn; the turning input itself is not measured. An extra PR cell disables calibration to expose the worker ceiling, and the retained worker count is recorded when logged. The plan takes about 45 minutes per tester and fails on crashes, lower FPS, worse p99 or stutter counts, or picture differences above the allowed tolerance. Intel targets have no testers listed online and will wait for matching PCs; other combinations may also wait if their tags do not coincide on one PC. No tester online now matches windows+intel, linux+intel: those targets wait for one.
0(PR only).kokiri-start(settings vsync=2, frame_rate_cap=6), then teleport to Kokiri Forest, measure kokiri-standing for 20 s (after 15 s warm-up), take the picture kokiri-standing to compare with main, move the stick, measure kokiri-turned for 15 s (after 0 s warm-up), take the picture kokiri-turned to compare with main, teleport to Hyrule Field, measure field-standing for 20 s (after 15 s warm-up), take the picture field-standing to compare with main, measure field-turned for 15 s (after 0 s warm-up), take the picture field-turned to compare with main, teleport to Market, measure market-standing for 20 s (after 15 s warm-up), take the picture market-standing to compare with main, measure market-turned for 15 s (after 0 s warm-up), take the picture market-turned to compare with main; 2 runs on each side.Started at 17:48 UTC.
Estimate: done in about 2 h, from the queue and the testers online now; some jobs can't start yet.
âš No tester has a PC for
linux+intel(none seen in 14 days): it waits until one signs in.NVIDIA 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), 26.2.7; 2: linux+amd+radv+dgpuAMD Radeon RX 9070 XT (RADV GFX1201), 26.2.6Measurements · windows+nvidia
Verdict: fail on windows+nvidia (2/1 report):
three-areas crash;three-areas kokiri-standing.fps_mean.min_delta_pct (-31.6%);three-areas kokiri-turned.fps_mean.min_delta_pct (-35.3%)failed.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: three-areas, 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); three-areas: 2 game starts a side, kokiri-standing 20 s after 15 s warm-up, kokiri-turned 15 s after 0 s warm-up, field-standing 20 s after 15 s warm-up, field-turned 15 s after 0 s warm-up, market-standing 20 s after 15 s warm-up, market-turned 15 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)9ea65c785c0d23RX_VK_PAR_CALIBRATE=0 (PR only)· field-turned · frame rateRX_VK_PAR_CALIBRATE=0 (PR only)· kokiri-turned · frame rateRX_VK_PAR_CALIBRATE=0 (PR only)· market-turned · frame rateRX_VK_PAR_CALIBRATE=0 (PR only)· field-standing · frame rateRX_VK_PAR_CALIBRATE=0 (PR only)· kokiri-standing · frame rateRX_VK_PAR_CALIBRATE=0 (PR only)· market-standing · frame rateRX_VK_PAR_CALIBRATE=0 (PR only)· field-turned · stuttersRX_VK_PAR_CALIBRATE=0 (PR only)· kokiri-turned · stuttersRX_VK_PAR_CALIBRATE=0 (PR only)· market-turned · stuttersRX_VK_PAR_CALIBRATE=0 (PR only)· field-standing · stuttersRX_VK_PAR_CALIBRATE=0 (PR only)· kokiri-standing · stuttersRX_VK_PAR_CALIBRATE=0 (PR only)· market-standing · stuttersRX_VK_PAR_CALIBRATE=0 (PR only)· field-turned · frame time (p99) (ms)RX_VK_PAR_CALIBRATE=0 (PR only)· kokiri-turned · frame time (p99) (ms)RX_VK_PAR_CALIBRATE=0 (PR only)· market-turned · frame time (p99) (ms)RX_VK_PAR_CALIBRATE=0 (PR only)· field-standing · frame time (p99) (ms)RX_VK_PAR_CALIBRATE=0 (PR only)· kokiri-standing · frame time (p99) (ms)RX_VK_PAR_CALIBRATE=0 (PR only)· market-standing · frame time (p99) (ms)RX_VK_PAR_CALIBRATE=0 (PR only)· forbidden log linesFlags and settings
RX_VK_PAR_CALIBRATE0Markdown for the PR description
Measurements · linux+amd
Verdict: fail on linux+amd (2/1 report):
three-areas kokiri-standing.fps_mean.min_delta_pct (-31.8%);three-areas kokiri-turned.fps_mean.min_delta_pct (-34.1%);three-areas field-turned.fps_mean.min_delta_pct (-3.5%)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+dgpuAMD Radeon RX 9070 XT (RADV GFX1201), 26.2.611th Gen Intel(R) Core(TM) i9-11900KB @ 3.30GHz, 32 GB RAMScene: three-areas, 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); three-areas: 2 game starts a side, kokiri-standing 20 s after 15 s warm-up, kokiri-turned 15 s after 0 s warm-up, field-standing 20 s after 15 s warm-up, field-turned 15 s after 0 s warm-up, market-standing 20 s after 15 s warm-up, market-turned 15 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)9ea65c785c0d23RX_VK_PAR_CALIBRATE=0 (PR only)· field-turned · frame rateRX_VK_PAR_CALIBRATE=0 (PR only)· kokiri-turned · frame rateRX_VK_PAR_CALIBRATE=0 (PR only)· market-turned · frame rateRX_VK_PAR_CALIBRATE=0 (PR only)· field-standing · frame rateRX_VK_PAR_CALIBRATE=0 (PR only)· kokiri-standing · frame rateRX_VK_PAR_CALIBRATE=0 (PR only)· market-standing · frame rateRX_VK_PAR_CALIBRATE=0 (PR only)· field-turned · stuttersRX_VK_PAR_CALIBRATE=0 (PR only)· kokiri-turned · stuttersRX_VK_PAR_CALIBRATE=0 (PR only)· market-turned · stuttersRX_VK_PAR_CALIBRATE=0 (PR only)· field-standing · stuttersRX_VK_PAR_CALIBRATE=0 (PR only)· kokiri-standing · stuttersRX_VK_PAR_CALIBRATE=0 (PR only)· market-standing · stuttersRX_VK_PAR_CALIBRATE=0 (PR only)· field-turned · frame time (p99) (ms)RX_VK_PAR_CALIBRATE=0 (PR only)· kokiri-turned · frame time (p99) (ms)RX_VK_PAR_CALIBRATE=0 (PR only)· market-turned · frame time (p99) (ms)RX_VK_PAR_CALIBRATE=0 (PR only)· field-standing · frame time (p99) (ms)RX_VK_PAR_CALIBRATE=0 (PR only)· kokiri-standing · frame time (p99) (ms)RX_VK_PAR_CALIBRATE=0 (PR only)· market-standing · frame time (p99) (ms)RX_VK_PAR_CALIBRATE=0 (PR only)· forbidden log linesFlags and settings
RX_VK_PAR_CALIBRATE0Markdown for the PR description
@Beanstalk (new head
85c0d23: segment workers from 4 cores, the late fix kept with them, a faster calibration saved per CPU and GPU) please test this PR against main on every platform: windows+amd, windows+nvidia, windows+intel, linux+amd, linux+nvidia, linux+intel, the steamdeck, and amd+igpu. Settings vsync=2 (Off) and frame_rate_cap=6 (Uncapped), defaults otherwise (Threaded rendering Auto). From thekokiri-startsave: Kokiri Forest, Hyrule Field and Castle Town Market, in each 20 s standing then 15 s turning the camera, two game starts a side. The segment workers' count is measured in play during the first ~10-15 s of play and then saved (logged asvk: segment workers: N kept), so please run each area long enough after loading. Measure fps, p99 frame time and stutters, and compare pictures with main. Also run one extra cell on the PR side withRX_VK_PAR_CALIBRATE=0(the ceiling, no calibration) to separate the calibration from the rest. Expect fps at least main's everywhere, no crash, and no new picture differences beyond main's own noise. Logs to look for:vk: Threaded rendering:,vk: segment workers run on logical processors,vk: presenter thread:.@Beanstalk (new head: segment workers are now off by default; the last run's RX_VK_PAR_CALIBRATE=0 cell, every core's worth of workers, lost 3-35% and crashed on the RTX 3070, so that path is opt-in only.) Please test this PR with its defaults, no switches, against main, on every platform: windows+amd, windows+nvidia, windows+intel, linux+amd, linux+nvidia, linux+intel, the steamdeck, amd+igpu. Settings vsync=2 (Off) and frame_rate_cap=6 (Uncapped), Threaded rendering Auto. From the
kokiri-startsave: Kokiri Forest, Hyrule Field and Castle Town Market, in each 20 s standing then 15 s turning the camera, two game starts a side. Measure fps, p99 frame time, stutters and crashes, and compare pictures with main. No extra cells. Expect fps at least main's and no crash; the log'svk: Threaded rendering:line should say the render thread only.Beanstalk test farm · an earlier run: commit
a2cc7b5(the newer results are in a comment below)Commit
a2cc7b5· tier reviewed · approved by HumanBean · dashboardWhat will be tested · AI plan (
gpt-6.1-terra)On eight requested hardware targets, this compares the PR defaults with main across Kokiri Forest, Hyrule Field and Castle Town Market, using two starts per side and taking roughly 23 minutes per tester. Each area is confirmed before measurement; FPS, p99 frame time, stutters, crashes and pictures are compared with V-Sync off and the frame cap uncapped. Camera turns are exercised between measurements, but the plan cannot measure while holding the camera stick, so the 15-second follow-up measurements occur after turning. A crash, lower FPS in any segment, a p99 increase over 15%, or a picture difference above the threshold fails the PR; stutters are reported. Intel targets have no currently listed online tester and will wait for one. 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-still for 20 s (after 5 s warm-up), take the picture forest to compare with main, move the stick, measure forest-after-turn for 15 s (after 0 s warm-up), take the picture forest-turned to compare with main, teleport to Hyrule Field, measure field-still for 20 s (after 5 s warm-up), take the picture field to compare with main, measure field-after-turn for 15 s (after 0 s warm-up), take the picture field-turned to compare with main, teleport to Market, measure market-still for 20 s (after 5 s warm-up), take the picture market to compare with main, measure market-after-turn for 15 s (after 0 s warm-up), take the picture market-turned to compare with main; 2 runs on each side.Started at 2026-11-19 21:35 UTC.
Estimate: done in about 40 min, from the queue and the testers online now; some jobs can't start yet.
âš No tester has a PC for
linux+intel(none seen in 14 days): it waits until one signs in.AMD Radeon RX 9070 XT, 2.0.416NVIDIA 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+igpuAMD Radeon 760M Graphics (RADV PHOENIX), 26.0.8NVIDIA GeForce RTX 5070, 620.70.6AMD Radeon(TM) 880M Graphics, 2.0.416; 2: linux+amd+radv+igpuAMD Radeon 760M Graphics (RADV PHOENIX), 26.0.8Run by hand on PCs that don't match their target (not counted toward it):
NVIDIA GeForce RTX 4080, 621.14.0)Measurements · windows+amd
Verdict: fail on windows+amd (1/1 report):
three-areas field-still.fps_mean.min_delta_pct (-2.1%);three-areas field-after-turn.fps_mean.min_delta_pct (-4.6%);three-areas market-still.fps_mean.min_delta_pct (-14.5%)failed.Hardware: 1: windows+amd+amd-proprietary+dgpu
AMD Radeon RX 9070 XT, 2.0.416AMD Ryzen 7 5800X3D 8-Core Processor, 32 GB RAMScene: three-areas, 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); three-areas: 2 game starts a side, forest-still 20 s after 5 s warm-up, forest-after-turn 15 s after 0 s warm-up, field-still 20 s after 5 s warm-up, field-after-turn 15 s after 0 s warm-up, market-still 20 s after 5 s warm-up, market-after-turn 15 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)9ea65c7a2cc7b5Flags 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):
three-areas field-still.fps_mean.min_delta_pct (-1.7%);three-areas market-still.fps_mean.min_delta_pct (-6.2%);three-areas 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: three-areas, 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); three-areas: 2 game starts a side, forest-still 20 s after 5 s warm-up, forest-after-turn 15 s after 0 s warm-up, field-still 20 s after 5 s warm-up, field-after-turn 15 s after 0 s warm-up, market-still 20 s after 5 s warm-up, market-after-turn 15 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)9ea65c7a2cc7b5Flags and settings
None: the plan varied no setting, switch or build option.
Markdown for the PR description
Measurements · linux+amd
Verdict: fail on linux+amd (2/1 report):
three-areas market-still.fps_mean.min_delta_pct (-24.5%);three-areas screenshot.max_diff;three-areas forest-still.fps_mean.min_delta_pct (-17.7%)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+igpuAMD Radeon 760M Graphics (RADV PHOENIX), 26.0.8AMD Ryzen 5 230 w/ Radeon 760M Graphics, 16 GB RAMScene: three-areas, 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); three-areas: 2 game starts a side, forest-still 20 s after 5 s warm-up, forest-after-turn 15 s after 0 s warm-up, field-still 20 s after 5 s warm-up, field-after-turn 15 s after 0 s warm-up, market-still 20 s after 5 s warm-up, market-after-turn 15 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)9ea65c7a2cc7b5Flags and settings
None: the plan varied no setting, switch or build option.
Markdown for the PR description
Measurements · linux+nvidia
Verdict: queued on linux+nvidia (1/1 report).
Hardware: 1: linux+nvidia+nvidia-proprietary+dgpu
NVIDIA GeForce RTX 5070, 620.70.6AMD Ryzen 5 9600X 6-Core Processor, 32 GB RAMScene: three-areas, 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); three-areas: 2 game starts a side, forest-still 20 s after 5 s warm-up, forest-after-turn 15 s after 0 s warm-up, field-still 20 s after 5 s warm-up, field-after-turn 15 s after 0 s warm-up, market-still 20 s after 5 s warm-up, market-after-turn 15 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)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 · amd+igpu
Verdict: fail on amd+igpu (2/1 report):
three-areas market-after-turn.fps_mean.min_delta_pct (-8.6%);three-areas screenshot.max_difffailed.Hardware: 1: windows+amd+amd-proprietary+igpu
AMD Radeon(TM) 880M Graphics, 2.0.416AMD Ryzen AI 9 365 w/ Radeon 880M, 32 GB RAM; 2: linux+amd+radv+igpuAMD Radeon 760M Graphics (RADV PHOENIX), 26.0.8AMD Ryzen 5 230 w/ Radeon 760M Graphics, 16 GB RAMScene: three-areas, 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); three-areas: 2 game starts a side, forest-still 20 s after 5 s warm-up, forest-after-turn 15 s after 0 s warm-up, field-still 20 s after 5 s warm-up, field-after-turn 15 s after 0 s warm-up, market-still 20 s after 5 s warm-up, market-after-turn 15 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)9ea65c7a2cc7b5Flags and settings
None: the plan varied no setting, switch or build option.
Markdown for the PR description
@HumanBean the PR got a new push (
59286c3) after your comment. Review it and approve again.Early findings from Windows 11 + RTX 4070 SUPER: it runs cleanly here, frames are noticeably more even, and the segment size looks like a useful tuning knob.
Just notes from trying the WIP, in case they help. PC: Ryzen 9 9900X (12 cores, 24 threads), RTX 4070 SUPER, Windows 11. 3840x2160 output, DLSS Quality, 90 fps cap, V-Sync on. A walking route: the Market, Hyrule Castle's grounds, the Market again (60 s). One build for every row: main
3653c9c5with this PR merged (85c0d23d) and a few local test changes that are the same in every row. Only switches differ, one run each unless said.The segment size and the scene interact. Medians of the frame time, same build:
RX_VK_PASS_BARRIER=fullRX_VK_PARALLEL=0 RX_VK_PASS_BARRIER=fullRX_VK_PARALLEL_SEG=80(2 runs)RX_VK_PARALLEL_SEG=640RX_VK_PASS_TIMINGrecords serially (gprof_on()in the parallel check invk_replay.cpp), so a timed run doesn't show the segments.@Beanstalk (new head
a2cc7b5: the parallel replay is removed entirely, one render thread records every command buffer; and a submission's parts now end where the game has already ended its render pass, from 120 draws, at 512 at the latest, instead of every 240 wherever it fell.) Please test this PR with its defaults, no switches, against main, on every platform: windows+amd, windows+nvidia, windows+intel, linux+amd, linux+nvidia, linux+intel, the steamdeck, amd+igpu. Settings vsync=2 (Off) and frame_rate_cap=6 (Uncapped), Threaded rendering Auto. From thekokiri-startsave: Kokiri Forest, Hyrule Field and Castle Town Market, in each 20 s standing then 15 s turning the camera, two game starts a side. Make sure each side has reached the area before measuring it (an earlier run measured one side still at the start). Measure fps, p99 frame time, stutters and crashes, and compare pictures with main. No extra cells. Expect fps at least main's everywhere and no crash; GPU-limited machines (the Deck, integrated GPUs) may gain the most from the new part ends.@HumanBean the PR got a new push (
a2cc7b5) after your comment. Review it and approve again.@Beanstalk (new head
a2cc7b5: the parallel replay is removed entirely, one render thread records every command buffer; and a submission's parts now end where the game has already ended its render pass, from 120 draws, at 512 at the latest, instead of every 240 wherever it fell.) Please test this PR with its defaults, no switches, against main, on every platform: windows+amd, windows+nvidia, windows+intel, linux+amd, linux+nvidia, linux+intel, the steamdeck, amd+igpu. Settings vsync=2 (Off) and frame_rate_cap=6 (Uncapped), Threaded rendering Auto. From thekokiri-startsave: Kokiri Forest, Hyrule Field and Castle Town Market, in each 20 s standing then 15 s turning the camera, two game starts a side. Make sure each side has reached the area before measuring it (an earlier run measured one side still at the start). Measure fps, p99 frame time, stutters and crashes, and compare pictures with main. No extra cells. Expect fps at least main's everywhere and no crash; GPU-limited machines (the Deck, integrated GPUs) may gain the most from the new part ends.A first Windows run of this head
a2cc7b53against main868f6be8with #1729: Windows 11, RTX 4070, NVIDIA 616.56, Ryzen 9 7900X (the CPU from your measurements). The two builds differ only in the PR (packs on one base).The route: a save, Debug tab → Teleport to Hyrule Field, Switch age to child, Set time 21:00, then about 2 minutes running around the field fighting the Stalchildren that spawn, with camera turns (it gives fights on demand: #1720, #1493 (comment)). Settings: 2560x1440, DLSS Quality, cap 90, frame generation off, Reflex and V-Sync on. Switches:
RX_VK_PASS_TIMING=1RX_VK_SLOW_LOG=50HOS_THREAD_TIMES=5HOS_FPS_LOG=1on every run. A person played; numbers come from 120-present windows from the field's arrival, and a fight window has 200 or more dispatches a frame.a frame's replay time, submit, median)A guess, not checked: with
283fbde2the one render thread records every command buffer, and on this Windows NVIDIA PC main's parallel recording may have been doing real work (your Linux calibration had already picked 0 workers). Main was run once on this route; both of the PR's runs agree with each other.A correction to my guess above: on main the work wasn't parallel segments but the recorder thread. In main's run its
vk: recorder:lines show it busy 0.5-1.4 s in every 10 s (about 1-2 ms a frame of Vulkan calls taken off the replay), with the replay waiting for it only 50-100 ms in 10 s. This PR removes the recorder, which would put those calls back on the render thread, about the +2.4 ms of replay work measured here. So on this Windows NVIDIA PC the recorder was a net gain, not a loss as on the i7-9700K in #1745.a2cc7b53ec21b606ef63Beanstalk test farm · an earlier run: commit
21b606e(the newer results are in a comment below)Commit
21b606e· tier reviewed · approved by HumanBean · dashboardWhat will be tested · AI plan (
gpt-6.1-terra)Two starts per side compare the PR’s default renderer with main in Kokiri Forest, Hyrule Field and the Market, using uncapped frame rates and V-Sync off; the route takes about 23 minutes per tester. It measures standing FPS, p99 frame time and stutters and compares pictures; crashes, any FPS decrease, a p99 increase over 15%, a stutter increase over 50%, or excessive picture differences fail the plan. The requested Intel targets will wait for matching testers. Camera turning during measurement, the AMD-only diagnostic cells and Steam Deck flash counts cannot be expressed by this plan; the farm forbids the frame-log switch. 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 20 s (after 5 s warm-up), take the picture forest to compare with main, move the stick, teleport to Hyrule Field, measure field-standing for 20 s (after 5 s warm-up), take the picture field to compare with main, teleport to Market, measure market-standing for 20 s (after 5 s warm-up), take the picture market to compare with main; 2 runs on each side.Started at 00:43 UTC.
Estimate: done in a few minutes, from the queue and the testers online now; some jobs can't start yet.
âš No tester has a PC for
linux+intel(none seen in 14 days): it waits until one signs in.AMD Radeon RX 9060 XT, 2.0.416NVIDIA GeForce RTX 4080, 621.1.0; 2: windows+nvidia+nvidia-proprietary+dgpuNVIDIA GeForce RTX 3090, 621.24.0AMD Radeon RX 9070 XT (RADV GFX1201), 26.2.6NVIDIA GeForce RTX 5070, 620.70.6AMD Custom GPU 0405 (RADV VANGOGH), 26.1.99AMD Radeon 760M Graphics (RADV PHOENIX), 26.0.8; 2: linux+amd+radv+igpu+handheldAMD Custom GPU 0405 (RADV VANGOGH), 26.1.99Measurements · windows+amd
Verdict: fail on windows+amd (1/1 report):
three-areas field-standing.fps_mean.min_delta_pct (-0.5%);three-areas market-standing.fps_mean.min_delta_pct (-10.9%);three-areas screenshot.max_difffailed.Hardware: 1: windows+amd+amd-proprietary+dgpu
AMD Radeon RX 9060 XT, 2.0.416AMD Ryzen 5 5600 6-Core Processor, 32 GB RAMScene: three-areas, 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); three-areas: 2 game starts a side, forest-standing 20 s after 5 s warm-up, field-standing 20 s after 5 s warm-up, market-standing 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)068fb8e21b606eFlags 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):
three-areas field-standing.fps_mean.min_delta_pct (-0.1%);three-areas market-standing.fps_mean.min_delta_pct (-0.2%)failed.Hardware: 1: windows+nvidia+nvidia-proprietary+dgpu
NVIDIA GeForce RTX 4080, 621.1.0AMD Ryzen 7 9800X3D 8-Core Processor, 32 GB RAM; 2: windows+nvidia+nvidia-proprietary+dgpuNVIDIA GeForce RTX 3090, 621.24.0AMD Ryzen 9 5900X 12-Core Processor, 128 GB RAMScene: three-areas, 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); three-areas: 2 game starts a side, forest-standing 20 s after 5 s warm-up, field-standing 20 s after 5 s warm-up, market-standing 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)068fb8e21b606eFlags and settings
None: the plan varied no setting, switch or build option.
Markdown for the PR description
Measurements · linux+amd
Verdict: fail on linux+amd (1/1 report):
three-areas market-standing.fps_mean.min_delta_pct (-14.9%);three-areas forest-standing.stutter_count.max_delta_pct;three-areas field-standing.stutter_count.max_delta_pctfailed.Hardware: 1: linux+amd+radv+dgpu
AMD Radeon RX 9070 XT (RADV GFX1201), 26.2.611th Gen Intel(R) Core(TM) i9-11900KB @ 3.30GHz, 32 GB RAMScene: three-areas, 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); three-areas: 2 game starts a side, forest-standing 20 s after 5 s warm-up, field-standing 20 s after 5 s warm-up, market-standing 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)068fb8e21b606eFlags and settings
None: the plan varied no setting, switch or build option.
Markdown for the PR description
Measurements · linux+nvidia
Verdict: queued on linux+nvidia (1/1 report).
Hardware: 1: linux+nvidia+nvidia-proprietary+dgpu
NVIDIA GeForce RTX 5070, 620.70.6AMD Ryzen 5 9600X 6-Core Processor, 32 GB RAMScene: three-areas, 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); three-areas: 2 game starts a side, forest-standing 20 s after 5 s warm-up, field-standing 20 s after 5 s warm-up, market-standing 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)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 · steamdeck
Verdict: fail on steamdeck (1/1 report):
three-areas screenshot.max_diff;three-areas forest-standing.stutter_count.max_delta_pct (+200.0%)failed.Hardware: 1: linux+amd+radv+igpu+handheld
AMD Custom GPU 0405 (RADV VANGOGH), 26.1.99AMD Custom APU 0405, 12 GB RAMScene: three-areas, 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); three-areas: 2 game starts a side, forest-standing 20 s after 5 s warm-up, field-standing 20 s after 5 s warm-up, market-standing 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)068fb8e21b606eFlags and settings
None: the plan varied no setting, switch or build option.
Markdown for the PR description
Measurements · amd+igpu
Verdict: fail on amd+igpu (2/1 report):
three-areas screenshot.max_difffailed.Hardware: 1: linux+amd+radv+igpu
AMD Radeon 760M Graphics (RADV PHOENIX), 26.0.8AMD Ryzen 5 230 w/ Radeon 760M Graphics, 16 GB RAM; 2: linux+amd+radv+igpu+handheldAMD Custom GPU 0405 (RADV VANGOGH), 26.1.99AMD Custom APU 0405, 12 GB RAMScene: three-areas, 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); three-areas: 2 game starts a side, forest-standing 20 s after 5 s warm-up, field-standing 20 s after 5 s warm-up, market-standing 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)068fb8e21b606eFlags and settings
None: the plan varied no setting, switch or build option.
Markdown for the PR description
WIP: perf(vk): the renderer's threads redesigned: segment workers measured in play, the presenter on every PC, narrower pass barriersto WIP: perf(vk): the renderer's threads rewritten: one render thread, the presenter on every PC, parts that end at the game's own pass ends@Beanstalk (new head
21b606e: rebased onto main068fb8ebas one commit; it adds the Steam Deck's late readbacks from #1725.) Please test this PR against main on every platform: windows+amd, windows+nvidia, windows+intel, linux+amd, linux+nvidia, linux+intel, the steamdeck, amd+igpu. Settings vsync=2 (Off) and frame_rate_cap=6 (Uncapped), Threaded rendering Auto. From thekokiri-startsave: Kokiri Forest, Hyrule Field and Castle Town Market, in each 20 s standing then 15 s turning the camera, two game starts a side; make sure each side has reached the area before measuring it (twice the PR side was measured still at the save's start). Measure fps, p99 frame time, stutters and crashes, and compare pictures with main.Cells: the PR's defaults on every platform. On linux+amd and windows+amd only, three more PR-only cells, to find what makes the Market standing slower there (-24.5% on an RX 7900 XTX, -14.5% on an RX 9070 XT last time):
RX_VK_PART_MIN=240 RX_VK_PART_MAX=240;RX_VK_PASS_BARRIER=full;RX_VK_READBACK_LATE=0. On the steamdeck, count the frame log's flashes (RX_VK_FRAME_LOG) on both sides: the PR reads the newest submission's results late there, the exposure meter among them, and expects no more flashes than main.@Beanstalk plan only linux+amd and windows+amd (an RX 9070 XT, RX 9060 XT or RX 7900 XTX each), on this head
21b606e: the Market standing loses 11-15% against main there while every other platform gains, and the renderer's own timing points at the replay's CPU cost per draw on AMD's drivers. From thekokiri-startsave, teleport to Castle Town Market and make sure the game has reached it, wait 15 s, then measure 40 s standing still; vsync=2 (Off), frame_rate_cap=6 (Uncapped). Cells, each on main and the PR: the defaults;RX_VK_PROF_FINE=1(logs where each draw's replay time goes, every few seconds: keep the whole log of both sides); on the PR only,RX_VK_PART_MIN=240 RX_VK_PART_MAX=240,RX_VK_PASS_BARRIER=fullandRX_VK_READBACK_LATE=0. Two game starts a side. Keep Windows Intel and Linux Intel queued as they are.Beanstalk test farm · an earlier run: commit
21b606e, run 2 (the newer results are in a comment below)Commit
21b606e· tier reviewed · approved by HumanBean · dashboardWhat will be tested · AI plan (
gpt-6.1-terra)On Linux and Windows AMD PCs with one of the three requested GPUs, the plan compares standing in Castle Town Market on the PR and main, with V-Sync off and the frame rate uncapped. Two runs per side measure FPS and frame times after a 15-second wait, and compare a screenshot; the profiling axis also collects the renderer’s detailed replay logs. The run takes about 44 minutes per tester and fails on a crash, a large picture difference, a 5% FPS loss, or a 15% p99 frame-time increase. The three PR-only rollback experiments are not included, and existing Intel queues are not changed by this plan.
0or1.kokiri-start(settings vsync=2, frame_rate_cap=6), then teleport to Market, take the picture market to compare with main, measure standing for 40 s (after 0 s warm-up); 2 runs on each side for each of 2 cells (2 game starts a run).Started at 02:17 UTC. Run 2: re-planned by HumanBean.
AMD Radeon RX 7900 XTX (RADV NAVI31), 26.2.7AMD Radeon RX 9060 XT, 2.0.416RX_VK_PROF_FINE=0RX_VK_PROF_FINE=0RX_VK_PROF_FINE=1RX_VK_PROF_FINE=1Measurements · linux+amd+gpu-rx-7900-xtx_rx-9060-xt_rx-9070-xt
Verdict: fail on linux+amd+gpu-rx-7900-xtx_rx-9060-xt_rx-9070-xt (1/1 report):
market-standing standing.fps_mean.min_delta_pct (-18.9%);market-standing screenshot.max_diff;market-standing standing.frametime_p99_ms.max_delta_pct (+246.6%)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 RAMScene: market-standing, 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); market-standing: 2 game starts a side, standing 40 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)068fb8e21b606eRX_VK_PROF_FINE=0· standing · frame rateRX_VK_PROF_FINE=0· standing · frame time (p99) (ms)RX_VK_PROF_FINE=0· forbidden log linesRX_VK_PROF_FINE=1· standing · frame rateRX_VK_PROF_FINE=1· standing · frame time (p99) (ms)RX_VK_PROF_FINE=1· forbidden log linesFlags and settings
RX_VK_PROF_FINE0,1Markdown for the PR description
Measurements · windows+amd+gpu-rx-7900-xtx_rx-9060-xt_rx-9070-xt
Verdict: fail on windows+amd+gpu-rx-7900-xtx_rx-9060-xt_rx-9070-xt (1/1 report):
market-standing standing.fps_mean.min_delta_pct (-12.1%)failed.Hardware: 1: windows+amd+amd-proprietary+dgpu
AMD Radeon RX 9060 XT, 2.0.416AMD Ryzen 5 5600 6-Core Processor, 32 GB RAMScene: market-standing, 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); market-standing: 2 game starts a side, standing 40 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)068fb8e21b606eRX_VK_PROF_FINE=0· standing · frame rateRX_VK_PROF_FINE=0· standing · frame time (p99) (ms)RX_VK_PROF_FINE=0· forbidden log linesRX_VK_PROF_FINE=1· standing · frame rateRX_VK_PROF_FINE=1· standing · frame time (p99) (ms)RX_VK_PROF_FINE=1· forbidden log linesFlags and settings
RX_VK_PROF_FINE0,1Markdown for the PR description
Beanstalk test farm · an earlier run: commit
69263e4(the newer results are in a comment below)Commit
69263e4· tier reviewed · approved by HumanBean · dashboardWhat will be tested · AI plan (
gpt-6.1-terra)Two starts per side compare main and the PR in Kokiri Forest, Hyrule Field and Market with V-Sync off, the frame cap uncapped and Threaded rendering Auto. The plan measures FPS, p99 frame time and stutters while standing and after camera turns, and compares pictures; crashes, performance regressions or picture differences fail. The requested AMD-only diagnostic cells and measurements during turns cannot be represented by this plan schema, so they are omitted. Each target takes about 23 minutes; Intel targets may wait for a matching tester. 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 kokiri-standing for 20 s (after 5 s warm-up), take the picture kokiri-standing to compare with main, move the stick, measure kokiri-after-turn for 15 s (after 0 s warm-up), take the picture kokiri-after-turn to compare with main, teleport to Hyrule Field, measure field-standing for 20 s (after 5 s warm-up), take the picture field-standing to compare with main, measure field-after-turn for 15 s (after 0 s warm-up), take the picture field-after-turn to compare with main, teleport to Market, measure market-standing for 20 s (after 5 s warm-up), take the picture market-standing to compare with main, measure market-after-turn for 15 s (after 0 s warm-up), take the picture market-after-turn to compare with main; 2 runs on each side.Started at 02:48 UTC.
Estimate: done in about 5 h, from the queue and the testers online now; some jobs can't start yet.
âš No tester has a PC for
linux+intel(none seen in 14 days): it waits until one signs in.AMD Radeon RX 9060 XT, 2.0.416NVIDIA GeForce RTX 3050 Ti Laptop GPU, 621.24.0; 2: windows+nvidia+nvidia-proprietary+dgpuNVIDIA GeForce RTX 4070 Ti SUPER, 621.14.0AMD Radeon RX 7900 XTX (RADV NAVI31), 26.2.7; 2: linux+amd+radv+dgpuAMD Steam Machine (RADV NAVI33), 26.1.99NVIDIA GeForce RTX 5070, 620.70.6Measurements · windows+amd
Verdict: fail on windows+amd (1/1 report):
three-areas market-standing.fps_mean.min_delta_pct (-25.2%);three-areas screenshot.max_difffailed.Hardware: 1: windows+amd+amd-proprietary+dgpu
AMD Radeon RX 9060 XT, 2.0.416AMD Ryzen 5 5600 6-Core Processor, 32 GB RAMScene: three-areas, 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); three-areas: 2 game starts a side, kokiri-standing 20 s after 5 s warm-up, kokiri-after-turn 15 s after 0 s warm-up, field-standing 20 s after 5 s warm-up, field-after-turn 15 s after 0 s warm-up, market-standing 20 s after 5 s warm-up, market-after-turn 15 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)068fb8e69263e4Flags 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):
three-areas crash;three-areas market-standing.fps_mean.min_delta_pct (-9.1%);three-areas 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 4070 Ti SUPER, 621.14.0AMD Ryzen 7 7800X3D 8-Core Processor, 32 GB RAMScene: three-areas, 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); three-areas: 2 game starts a side, kokiri-standing 20 s after 5 s warm-up, kokiri-after-turn 15 s after 0 s warm-up, field-standing 20 s after 5 s warm-up, field-after-turn 15 s after 0 s warm-up, market-standing 20 s after 5 s warm-up, market-after-turn 15 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)068fb8e69263e4Flags and settings
None: the plan varied no setting, switch or build option.
Markdown for the PR description
Measurements · linux+amd
Verdict: fail on linux+amd (2/1 report):
three-areas market-standing.fps_mean.min_delta_pct (-15.6%);three-areas screenshot.max_diff;three-areas kokiri-standing.stutter_count.max_delta_pctfailed.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+dgpuAMD Steam Machine (RADV NAVI33), 26.1.99AMD Custom CPU 1772, 16 GB RAMScene: three-areas, 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); three-areas: 2 game starts a side, kokiri-standing 20 s after 5 s warm-up, kokiri-after-turn 15 s after 0 s warm-up, field-standing 20 s after 5 s warm-up, field-after-turn 15 s after 0 s warm-up, market-standing 20 s after 5 s warm-up, market-after-turn 15 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)068fb8e69263e4Flags and settings
None: the plan varied no setting, switch or build option.
Markdown for the PR description
Measurements · linux+nvidia
Verdict: running on linux+nvidia (1/1 report).
Hardware: 1: linux+nvidia+nvidia-proprietary+dgpu
NVIDIA GeForce RTX 5070, 620.70.6AMD Ryzen 5 9600X 6-Core Processor, 32 GB RAMScene: three-areas, 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); three-areas: 2 game starts a side, kokiri-standing 20 s after 5 s warm-up, kokiri-after-turn 15 s after 0 s warm-up, field-standing 20 s after 5 s warm-up, field-after-turn 15 s after 0 s warm-up, market-standing 20 s after 5 s warm-up, market-after-turn 15 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)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
@Beanstalk (new head
69263e4: two commits on21b606eagainst the Market loss on AMD discrete cards: an indexed draw's index buffer bound once per buffer with its place given as the first index (the fine profile on the RX 7900 XTX showed a bind for 0.97 of its 2,500 draws a frame), and a submission's first part ended sooner so the GPU starts sooner.) Please test this PR against main on every platform: windows+amd, windows+nvidia, windows+intel, linux+amd, linux+nvidia, linux+intel, the steamdeck, amd+igpu. Settings vsync=2 (Off) and frame_rate_cap=6 (Uncapped), Threaded rendering Auto. From thekokiri-startsave: Kokiri Forest, Hyrule Field and Castle Town Market, in each 20 s standing then 15 s turning the camera, two game starts a side; make sure each side has reached the area before measuring it. Measure fps, p99 frame time, stutters and crashes, and compare pictures with main (the index change must draw exactly as main does).Cells: the PR's defaults on every platform. On linux+amd and windows+amd only, three more PR-only cells:
RX_VK_INDEX_REBASE=0;RX_VK_PART_FIRST=0;RX_VK_PROF_FINE=1(keep its log).69263e4d668ba0d0281bBeanstalk test farm · an earlier run: commit
8ba0d02(the newer results are in a comment below)Commit
8ba0d02· tier reviewed · approved by HumanBean · dashboardWhat will be tested · AI plan (
gpt-6.1-terra)On eight requested hardware targets, the plan compares main and the PR over two starts per side, visiting Kokiri Forest, Hyrule Field and Market with V-Sync off and the frame-rate cap uncapped. It measures standing and post-turn performance, stutters and crashes, and compares pictures; the schema cannot measure while a camera-stick step is active. Regressions in FPS, p99 frame time, stutters or pictures fail the plan. The AMD-only switch cells cannot be restricted to those targets by the plan-wide matrix and are omitted; Intel targets will wait because no Intel tester is currently listed. 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 20 s (after 5 s warm-up), take the picture forest to compare with main, move the stick, measure forest-after-turn for 15 s (after 0 s warm-up), teleport to Hyrule Field, measure field-standing for 20 s (after 5 s warm-up), take the picture field to compare with main, measure field-after-turn for 15 s (after 0 s warm-up), teleport to Market, measure market-standing for 20 s (after 5 s warm-up), take the picture market to compare with main, measure market-after-turn for 15 s (after 0 s warm-up); 2 runs on each side.Started at 04:07 UTC.
Estimate: done in about 1 h, from the queue and the testers online now; some jobs can't start yet.
âš No tester has a PC for
linux+intel(none seen in 14 days): it waits until one signs in.AMD Radeon RX 9060 XT, 2.0.416NVIDIA GeForce RTX 3050 Ti Laptop GPU, 621.24.0; 2: windows+nvidia+nvidia-proprietary+dgpuNVIDIA GeForce RTX 3080 Ti, 621.24.0AMD Radeon RX 7900 XTX (RADV NAVI31), 26.2.7NVIDIA GeForce GTX 1070, 580.196.7; 2: linux+nvidia+nvidia-proprietary+dgpuNVIDIA GeForce RTX 4070, 615.71.9AMD Custom GPU 0405 (RADV VANGOGH), 26.1.99; 2: linux+amd+radv+igpuAMD Radeon 780M Graphics (RADV PHOENIX), 26.1.99Measurements · windows+amd
Verdict: fail on windows+amd (1/1 report):
three-areas crash;three-areas forest-after-turn.fps_mean.min_delta_pct (-22.3%);three-areas market-standing.fps_mean.min_delta_pct (-12.8%)failed.Hardware: 1: windows+amd+amd-proprietary+dgpu
AMD Radeon RX 9060 XT, 2.0.416AMD Ryzen 5 5600 6-Core Processor, 32 GB RAMScene: three-areas, 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); three-areas: 2 game starts a side, forest-standing 20 s after 5 s warm-up, forest-after-turn 15 s after 0 s warm-up, field-standing 20 s after 5 s warm-up, field-after-turn 15 s after 0 s warm-up, market-standing 20 s after 5 s warm-up, market-after-turn 15 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)08994a98ba0d02Flags 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):
three-areas market-after-turn.fps_mean.min_delta_pct (-5.3%);three-areas market-after-turn.stutter_count.max_delta_pctfailed.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 3080 Ti, 621.24.0AMD Ryzen 9 5900X 12-Core Processor, 64 GB RAMScene: three-areas, 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); three-areas: 2 game starts a side, forest-standing 20 s after 5 s warm-up, forest-after-turn 15 s after 0 s warm-up, field-standing 20 s after 5 s warm-up, field-after-turn 15 s after 0 s warm-up, market-standing 20 s after 5 s warm-up, market-after-turn 15 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)08994a98ba0d02Flags and settings
None: the plan varied no setting, switch or build option.
Markdown for the PR description
Measurements · linux+amd
Verdict: fail on linux+amd (1/1 report):
three-areas market-standing.fps_mean.min_delta_pct (-22.9%);three-areas screenshot.max_diff;three-areas field-after-turn.stutter_count.max_delta_pctfailed.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: three-areas, 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); three-areas: 2 game starts a side, forest-standing 20 s after 5 s warm-up, forest-after-turn 15 s after 0 s warm-up, field-standing 20 s after 5 s warm-up, field-after-turn 15 s after 0 s warm-up, market-standing 20 s after 5 s warm-up, market-after-turn 15 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)08994a98ba0d02Flags and settings
None: the plan varied no setting, switch or build option.
Markdown for the PR description
Measurements · linux+nvidia
Verdict: pass on linux+nvidia (2/1 report): no rule failed and no PR run crashed.
Hardware: 1: linux+nvidia+nvidia-proprietary+dgpu
NVIDIA GeForce GTX 1070, 580.196.7Intel(R) Core(TM) i7-8750H CPU @ 2.20GHz, 32 GB RAM; 2: linux+nvidia+nvidia-proprietary+dgpuNVIDIA GeForce RTX 4070, 615.71.9AMD Ryzen 7 5800X3D 8-Core Processor, 64 GB RAMScene: three-areas, 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); three-areas: 2 game starts a side, forest-standing 20 s after 5 s warm-up, forest-after-turn 15 s after 0 s warm-up, field-standing 20 s after 5 s warm-up, field-after-turn 15 s after 0 s warm-up, market-standing 20 s after 5 s warm-up, market-after-turn 15 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)08994a98ba0d02Flags and settings
None: the plan varied no setting, switch or build option.
Markdown for the PR description
Measurements · amd+igpu
Verdict: fail on amd+igpu (2/1 report):
three-areas screenshot.max_diff;three-areas market-after-turn.stutter_count.max_delta_pct (+550.0%);three-areas market-after-turn.frametime_p99_ms.max_delta_pct (+68.4%)failed.Hardware: 1: linux+amd+radv+igpu+handheld
AMD Custom GPU 0405 (RADV VANGOGH), 26.1.99AMD Custom APU 0405, 12 GB RAM; 2: linux+amd+radv+igpuAMD Radeon 780M Graphics (RADV PHOENIX), 26.1.99AMD Ryzen 7 7840HS w/ Radeon 780M Graphics, 24 GB RAMScene: three-areas, 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); three-areas: 2 game starts a side, forest-standing 20 s after 5 s warm-up, forest-after-turn 15 s after 0 s warm-up, field-standing 20 s after 5 s warm-up, field-after-turn 15 s after 0 s warm-up, market-standing 20 s after 5 s warm-up, market-after-turn 15 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)08994a98ba0d02Flags and settings
None: the plan varied no setting, switch or build option.
Markdown for the PR description
@Beanstalk (new head
8ba0d02: rebased on main with #1815, so Linux starts again; plus comment and doc cleanup.) Please test this PR against main on every platform: windows+amd, windows+nvidia, windows+intel, linux+amd, linux+nvidia, linux+intel, the steamdeck, amd+igpu. Settings vsync=2 (Off) and frame_rate_cap=6 (Uncapped), Threaded rendering Auto. From thekokiri-startsave: Kokiri Forest, Hyrule Field and Castle Town Market, in each 20 s standing then 15 s turning the camera, two game starts a side; make sure each side has reached the area before measuring it. Measure fps, p99 frame time, stutters and crashes, and compare pictures with main.Matrix, please plan all of these cells. Every platform: the PR's defaults. On linux+amd and windows+amd, three more cells, each on the PR side only against the same main: (a) env
RX_VK_INDEX_REBASE=0; (b) envRX_VK_PART_FIRST=0; (c) envRX_VK_INDEX_REBASE=0 RX_VK_PART_FIRST=0. They tell which of the two new changes moves the Market.This is the Steam Deck check that #1725 asked for, run on that PR's commit (
fec53338d) before it was folded into this one. It is not a test of this PR's head: it measures only the late-readback default on a handheld, on a build without the thread rewrite. I'm posting it here because 1725's closing comment points its Deck flash check at this PR's Still to test; I can run the same routes on this PR's head next if that is wanted.Steam Deck run of pull 1725's commit: faster where the readback wait was the cost, no new flashes
On a Steam Deck OLED the default that commit turns on is worth about +14% standing in Hyrule Field and +7% in the Market, nothing in Kokiri Forest, and +5% fps while the camera turns, with lower 1% lows there. The flash detector found no one-frame white or dark picture with that default that the same build without the rule doesn't also show.
What was built.
de22a6b9d: main's parent621dae505with pull 1725's commitfec53338dcherry-picked on it (clean). Not main's head at the time (0de25c3e4): it did not compile on Linux then (gpu/vk/menu/pipeline_list.cppusedSIGPIPE,sigtimedwait,waitpidandunamewithout their POSIX headers), since fixed by #1807. Main's head differed from its parent only by that menu change, which touches none of that commit's files. A dev-packs base,RX_MARCH=x86-64-v3, noRX_HOT_OPT, no Tracy. One binary for both sides: D = nothing set (that commit's default), Z =RX_VK_READBACK_LATE=0(its own "turning it off").Hardware. Steam Deck OLED, SteamOS 3.9.5, RADV VANGOGH (Mesa 26.2.0-devel), GPU clock manual 1600 MHz, headless (the harness, no one holding it).
Settings. Minimum preset, Render 50%, output 1280x800, frame rate uncapped, vsync off. Fresh options each run, the same warm shader cache for all runs.
Policy lines. Both sides print
readback_late_newest=1invk: policy: can and measured: ...: the line reports the device's default, and withRX_VK_READBACK_LATE=0it still prints=1. What tells the sides apart is the rule line, which D prints and Z doesn't:vk: readbacks read without their wait: the default on a handheld (newest 1, 0 sizes). Worth a word in the log, since=1on a run with the switch set to0reads as the rule being on.Method. Kokiri Forest standing 120 s, then a teleport to Hyrule Field: standing 60 s, 4 camera pans (about 46 s), standing again (about 84 s). The Market: standing, 120 s from 20 s after arrival. Two runs a side per route, alternated D, Z, Z, D. fps from the presents, p99 frame time, 1% low = 1000 / mean of the slowest 1% of frame times.
The readback lines (
vk: readbacks of GPU results: <n> of 3000 waited for their write-back, <ms> ms each, median per window, run 1 / run 2):Where the gain is, each wait is about a quarter as long with the rule (0.2 ms against 0.8 ms in the field). In Kokiri Forest the waits are already 0.03 ms, which fits nothing changing there. With the rule more of the 3000 reads count as waited, but each wait is short.
Flashes (
RX_VK_FLASH=1, every present compared with its neighbours,RX_VK_FLASH_REGION=40; one run a side on the main route, two a side on the Lon Lon Ranch song route). I looked at every flagged triplet (the frame before, the flagged one and the one after) and measured the whole picture's mean brightness of each.Limits.
Beanstalk test farm
Commit
4543d04· tier reviewed · approved by HumanBean · dashboardWhat will be tested · AI plan (
gpt-6.1-terra)The plan compares main with the PR’s default recorder behavior and with the recorder disabled, because the new head restores that thread on PCs with at least six physical cores. Two starts per side visit Kokiri Forest, Hyrule Field and Market with V-Sync off and the frame rate uncapped, measuring FPS, p99 frame time and stutters and comparing pictures. Camera turns are exercised, but performance is measured immediately after each turn rather than during it. Crashes, significant performance regressions or large picture differences fail the run; each tester’s game runs take roughly 45 minutes. The requested Intel targets have no testers shown online and will wait for matching PCs. No tester online now matches windows+intel, linux+intel: those targets wait for one.
off(PR only).kokiri-start(settings vsync=2, frame_rate_cap=6), then teleport to Kokiri Forest, measure forest-standing for 20 s (after 5 s warm-up), take the picture forest to compare with main, move the stick, measure forest-after-turn for 15 s (after 0 s warm-up), teleport to Hyrule Field, measure field-standing for 20 s (after 5 s warm-up), take the picture field to compare with main, measure field-after-turn for 15 s (after 0 s warm-up), teleport to Market, measure market-standing for 20 s (after 5 s warm-up), take the picture market to compare with main, measure market-after-turn for 15 s (after 0 s warm-up); 2 runs on each side.Done (started 05:33 UTC, last report 05:33 UTC). Comment
@Beanstalkto run it again.NVIDIA GeForce RTX 3050 Ti Laptop GPU, 621.24.0Measurements · windows+nvidia
Verdict: could not be judged on windows+nvidia (1/1 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 RAMScene: three-areas, 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); three-areas: 2 game starts a side, forest-standing 20 s after 5 s warm-up, forest-after-turn 15 s after 0 s warm-up, field-standing 20 s after 5 s warm-up, field-after-turn 15 s after 0 s warm-up, market-standing 20 s after 5 s warm-up, market-after-turn 15 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)Nothing was measured: no metric came back.
Flags and settings
RX_VK_RECORDERoffMarkdown for the PR description
@Beanstalk (new head
4543d04: the recorder thread is back, from 6 physical cores, without the segment workers. It moves the driver's calls off the render thread, which should recover the Market standing loss on fast GPUs, AMD most.) Please test this PR against main on every platform: windows+amd, windows+nvidia, windows+intel, linux+amd, linux+nvidia, linux+intel, the steamdeck, amd+igpu. Settings vsync=2 (Off) and frame_rate_cap=6 (Uncapped), Threaded rendering Auto. From thekokiri-startsave: Kokiri Forest, Hyrule Field and Castle Town Market, in each 20 s standing then 15 s turning the camera, two game starts a side; make sure each side has reached the area before measuring it. Measure fps, p99 frame time, stutters and crashes, and compare pictures with main.Matrix, on every platform, two cells: the PR's defaults; and the PR with env
RX_VK_RECORDER=off(the same build without the recorder thread), each against main.WIP: perf(vk): the renderer's threads rewritten: one render thread, the presenter on every PC, parts that end at the game's own pass endsto perf(vk): the renderer's threads rewritten: one render thread, the presenter on every PC, parts that end at the game's own pass endsPost-merge Steam Deck report (zeldaootr). This is the first Steam Deck report for this change. It compares current
mainafter the merge with stock v3.0.0, and checks the change's turn-back switches againstmain's defaults.Summary. On the Deck,
mainmatches v3.0.0 in Kokiri Forest and the Market. In the forest exit toward Hyrule Field it is clearly faster: +25% standing, +14% panning, +33% standing after the pans. This compares two builds 64 commits apart, so the Field gain is not this PR's alone. Against the PR's own turn-back switches on the same build,main's defaults win in the Field (+11% standing). The switches win in Kokiri (one pair only, see below), and the Market is a tie. The Deck runs without the render thread, by its own policy. These headless runs make no swapchain presents, so the presenter thread never started. There were no errors, device losses, validation lines or crashes. Pictures match v3.0.0. The flash detector's flags onmainare Navi and the camera passing through Link during the pans, not exposure flashes.Builds
d12ddf69b(#1833, one commit after this PR's merge commita3565ac5b), versionv3.0.0-64-gd12ddf69b. Built withdev.py baseanddev.py rt --empty,RX_MARCH=x86-64-v3, the release compile levels.b9d6fe244df8cef0. Neither side prepared or compiled pipelines during the runs: novk: preparation:lines, and both logged 47798 program keys kept.Hardware and settings
quality_preset=1),handheld_profile=0, Render 50% (resolution=4), frame pacing on, uncapped (frame_rate_cap=6), V-Sync off, frame generation off.kokiri-start. No switches set on either side (stock defaults), except in the O cell below.main against v3.0.0 (median of two runs; each run in brackets)
main's two Market runs spread widely (14.4 and 18.2). A thirdmainMarket run, from the pair below, gave 16.0. Read the Market as no difference.This PR's turn-back switches (O) against main's defaults (M), one pair a route
The rewrite has no single off switch: the recorder and parallel-replay switches are gone, and the presenter thread has none. So the comparison cell is the set this PR's Risks names for turning it back:
RX_VK_PART_MIN=240 RX_VK_PART_MAX=240 RX_VK_PASS_BARRIER=full RX_VK_READBACK_LATE=0.RX_VK_PART_FIRSTstays at its default. This measures the parts, the barriers and the late readbacks, not the thread design. It is one run a side, run O then M on each route.The O cell's log confirms the switches took effect:
The Kokiri gap for O comes from one run. All five other Kokiri runs (M and V) fell between 19.06 and 19.95 fps, so it needs repeating before anyone reads it as a cost of the new parts or barriers there.
What main chose on the Deck
Lines from main's log, each the same in every main run:
the replay runs on a thread of its ownline appears. The replay runs on the game's presenting thread.presenter thread: each present made on a thread of its own ...) and its per-600-presents lines never appear. Only the device-setup line about present-wait does. Whether the presenter thread costs or gains anything on the Deck needs a run with a real swapchain (Game Mode or a window).Threaded rendering: Auto: off (3 cores ...)as well. Itscan and measuredpolicy line haspresenter_queue=0and noreadback_late_newest, and it logs no rule line about reading without waiting.VK_ERROR, device-lost, validation or[error]line. No crash records, and every exit code was 0. Lines matching "fallback" were the same on both builds: the Deck's compute fallback for mesh shaders (#640), and the reserved core mask's fallback.Pictures
One screenshot per area, at the same spot after settling,
mainagainst v3.0.0. The difference measure is the share of pixels whose largest channel differs by more than 8, and the mean absolute difference.mainrun's shot differs from every other shot (40%, mean 10): Navi had flown into Link's house, lighting its inside, and the exposure followedNo difference in lighting, geometry, textures or post-processing was seen.
Flash detector
One
mainrun of the main route withRX_VK_FLASH=1,RX_VK_FLASH_REGION=40,RX_VK_SHOT_EVERY=1. This is the Deck item in Still to test, with the newest submission's readbacks read late.Limits
mainagainst v3.0.0 covers everything merged since v3.0.0, not only this PR. The O pair is the closest to this PR's own effect on the Deck.tests/run_all.pyond12ddf69b, run separately) have 5 failures that start with this PR: checks that readvk_replay.cppsource this PR reshaped. They aretest_split_before_draw.py,test_early_split_before_draw.py,test_hold_gpu.py,test_readback_early.pyandtest_readback_late.py. The other 46 failures already failed on this PR's parent.Correction to my report above: the link after "the Deck's compute fallback for mesh shaders" is wrong. "issue 640" is the wording of the game's own log line (
vk: mesh shaders ...: task and mesh programs run as compute shaders (issue 640 fallback)), not Longshot's number 640; ignore the link.