Game: Inconsistent performance and severe stuttering in Hyrule Castle Town / Temple of Time #1639
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
3 participants
Notifications
Due date
No due date set.
Dependencies
No dependencies set
Reference
MagicBeanDev/Longshot#1639
Loading…
Reference in a new issue
No description provided.
Delete branch "%!s()"
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?
Reported on Discord
Longshot version and channel
Nightly 486-09a20cd6 (source ce68e351), nightly. Additional comparison: Longshot v3.0.0-rc.8 (source a74a6465), RC track. Subsequent report: Nightly 665-6e1521a7 (source a8ab8a56), nightly.
What happened and steps
Running as child Link in Hyrule Castle Town Market and outside the Temple of Time produces significant FPS fluctuations and uneven frame pacing, even when FPS is relatively high. Reporter says it persists at lower resolutions and quality settings. Expected smooth movement with stable frame times; actual performance is stuttery in these areas.
Steps: Enter Market_Child or HyruleField_Child as child Link, run around and toward/back from nearby areas, aim and shoot the slingshot, and observe abrupt drops. The behavior recurs but not with every action. The latest RC comparison reports stutters in Hyrule Field, Hyrule Market, and other areas.
Where in the game
Hyrule Castle Town Market and Temple of Time exterior (child Link). Reporter also notes Hyrule Field and Kokiri Forest.
How often
Consistently noticeable, with intermittent severe drops. Later, after about an hour or more of play, a sudden FPS-cap-related stutter returned when entering a new zone. The reporter cannot predict the onset.
Graphics card and driver version
NVIDIA GeForce RTX 3080 Ti 12 GB, driver 621.24.
CPU, RAM and operating system
AMD Ryzen 9 5900X (24 threads), 64 GB RAM, Windows 11 build 26300.
Settings
Nightly tests included Vulkan at 4K (3840×2160, 100%), 1080p (1920×1080), and 1440p (2560×1440), with no upscaling in the latter runs; fullscreen, V-Sync on, native HDR/PsychoV35, highest shadows, frame pacing on, threaded rendering and sampler cache on, robust buffers off. RC 4K settings: fullscreen, V-Sync on, 100% render scale, no upscaling or sharpening, native HDR/PsychoV35, highest shadows, frame pacing on, Reflex off, threaded rendering and sampler cache on, robust buffers off. Other settings and exact RC report details are in the attached report.txt. External overlays/injectors are unspecified.
Performance observations
Initial report: around 46–60 FPS in sampled intervals, with frame-time spikes over 40–60 ms. Logs reportedly include slow Vulkan submissions and texture uploads; their relationship to the stuttering is not established.
Profiled Nightly 1080p and 1440p runs included intervals around 44–61 FPS. The reporter observed similar performance at the two lower resolutions, somewhat worse at higher resolution. The profiled 1440p route went from Temple of Time around town, through Hyrule Field to Kokiri Forest and Link's house, where they saved. Reported profiling included CPU-side replay submission around 12.98 ms in one 44.2 FPS interval (2,561 draws); thread activity and GPU pass timing do not establish the cause.
RC v3.0.0-rc.8 comparison: reporter says general performance is noticeably improved versus the prior Nightly and often stays near the 60 FPS cap, but intermittent severe drops remain. GPU utilization reportedly falls during slowdowns; lowering resolution does not remove spikes and some seem worse. The 4K log has 120-present intervals averaging about 45–47 FPS and individual slow frames over 40 ms; the preceding 1440p run has larger isolated spikes. Reporter cannot match individual spikes to actions. CPU-thread/GPU-pass profiling is enabled in the RC 4K run.
Later intermittent cap-related episode (Nightly 665): during play, removing the 60 FPS cap and setting 90 made the stutter disappear; cycling cap options down to 30 made it worse, then restoring 60 and saving a Bug report. The reporter says this kind of sudden frame-time oscillation happens after extended play (about an hour or more), apparently on entering a new zone. Their earlier Lost Woods and Dodongo's Cavern occurrences resolved after saving and reloading without restarting Longshot. The latest report identifies LostWood_Child as the last named map, 4K output/render, with fullscreen, V-Sync, HDR Native, frame pacing on, threaded rendering on, sampler cache on, robust buffers off. Current report records HOS_FPS_LOG=1, HOS_THREAD_TIMES=5, RX_VK_PASS_TIMING=1 and RX_VK_SLOW_LOG=50; the main log confirms HOS_THREAD_TIMES=5 and GPU pass timing is active. No precise frame-event capture was available because onset is intermittent. No specific action could be matched to the onset.
Logs and screenshot
Original Bug report folder: longshot_log.txt, longshot_log.1.txt, report.txt and picture.png.
Follow-up Bug report folders from 2026-11-18: profiled 1080p and 1440p reports, each with report.txt, longshot_log.txt and longshot_log.1.txt. Profiling switches were recognized in these reports.
Latest RC Bug report folder: report.txt, longshot_log.txt and longshot_log.1.txt; profiling enabled. The reporter supplied a preceding 1440p run as well. These attachments are from the follow-up message in the Discord thread.
Nightly 665 report from 2026-11-21: report.txt, longshot_log.txt and longshot_log.1.txt (the prior log is truncated in the report copy), plus picture.png. Profiling settings are present in report.txt and HOS_THREAD_TIMES appears in the log. These files were attached in the Discord follow-up.
Related reports
Related but not established as the same cause or condition: Longshot #996 tracks a Market performance hit at one specific spot while running circles on an RTX 5080/Linux; PR #994 addresses recurring render-target allocation work in Castle Town and is included since nightly-319-9f8958a5; PR #1478 limits texture heap compaction to at most once a second and is included since nightly-455-b8f3d547. Neither PR establishes a fix for this report.
Profiling follow-up
Reporter initially put
longshot_env.txtin the wrong similarly named location; after placing it in the actualbindirectory, later Nightly profiles confirmed profiling switches took effect. The new RC 4K report also has CPU-thread/GPU-pass profiling enabled. The exact logged intervals cannot be tied to particular actions; profile data and logs are provided to investigate.Profiling instructions: https://longshot.magicbean.dev/help/performance/ (enable the four settings in
longshot_env.txt, restart, reproduce for 60–90 seconds, then make a Bug report while the stutter is occurring; remove the profiling lines afterward). Because the intermittent episode may occur only after extended play and its onset is unpredictable, the reporter was unable to obtain a targeted capture at the onset; the latest report nevertheless includes profiling switches.Related merged changes
PR #1875 changes the default zstd decode behavior to give up its virtual core on most PCs. Its measurements include fewer long Market frames on a Ryzen desktop and a large Kokiri Forest gain on a Steam Deck. It is included since nightly-649-846e5399, so the reporter's Nightly 665 contains it; this is not evidence that it fixes this reporter's remaining cap-dependent, intermittent episode. PR #157 fixes specific HLE condition-variable and timed-semaphore paths, but its description says the tested gameplay scenes do not call those paths; it does not establish a fix for this stutter.
Additional independent report
An independent v3.0.0 stable RTX 5070 report records Market hitches and long virtual-core waits (up to 326 ms), while Kakariko and Hyrule Field were smoother. It is another system/build data point; the cause is unconfirmed and not established as the same symptom mechanism.
Further reports
After the initial market issue improved substantially, reporter also described severe cap-dependent frametime oscillation after playing ocarina with Skull Kid in Lost Woods and on entering Dodongo's Cavern near a lava pit. It worsened with a 60 cap, disappeared uncapped and worsened at 30; saving/reloading resolved it without restarting the build. That episode was not profiled at the time and no build version was given then. The Nightly 665 report described above adds a recurrence after extended play in a new zone and a profiled log, but does not identify the specific cause.
Files from the Discord report:
From Discord:
Reporter says they placed
longshot_env.txtwith HOS_FPS_LOG=1, HOS_THREAD_TIMES=5, RX_VK_PASS_TIMING=1 and RX_VK_SLOW_LOG=50 inC:\Games\Longshot build\nightly\nightly-486-09a20cd6\binand asks whether this is the correct location when launching from the launcher. They have not yet supplied a profiled Bug report folder.From Discord:
The reporter supplied another Bug report folder from Nightly 486-09a20cd6 (source ce68e351), showing Market_Child and the RTX 3080 Ti / driver 621.24 system.
report.txtlists onlyLONGSHOT_NO_UPDATE_CHECK=1under environment switches; neither attachedlongshot_log.txtnorlongshot_log.1.txtcontainsHOS_THREAD_TIMESorRX_VK_PASS_TIMING. So the new report still does not include the requested profiling data. The profiling env settings do not appear to have taken effect in this run.From Discord:
The reporter now clarifies this Bug report was made after launching Longshot.exe directly from the build location, bypassing the launcher. It is Nightly 486-09a20cd6 (source ce68e351), at Market_Child on the same RTX 3080 Ti / driver 621.24 system. However, report.txt still lists only LONGSHOT_NO_UPDATE_CHECK=1 among environment switches, and neither log contains HOS_THREAD_TIMES or RX_VK_PASS_TIMING, so this is not yet the requested profiling report.
Updated with more from Discord: Added the latest Bug report, which confirms profiling environment switches were recognized, and the reporter's clarification that the file was placed in the bin directory rather than beside image.bin.
Files:
Updated with more from Discord: Added 1080p profiled follow-up report with thread/pass profiling recognized and observed low-FPS interval; noted stutter also occurs in Kokiri Forest and planned 1440p comparison.
Files:
Updated with more from Discord: Added the 1440p profiled run, route through town, Hyrule Field and Kokiri Forest to Link's house, and reported resolution comparison; attached its report and logs.
Files:
Updated with more from Discord: Added RC v3.0.0-rc.8 comparison and its profiled logs: improvement over Nightly but persistent intermittent stutters in multiple areas, additional reproduction and resolution observations.
Files:
A data point from a low-end laptop that goes the other way from the RTX 3080 Ti comparison above: here, rc.8 hitches about 5× more often in the Market than nightly-486, at the same real frame rate.
Machine: Dell XPS 15 9520 laptop, i9-12900HK, RTX 3050 Ti Laptop 4 GB, NVIDIA 580.196.07, 16 GB RAM, Fedora 43 (kernel 7.2.15), KDE Wayland, internal panel 1920×1200 at 60 Hz.
Builds:
nightly-486-09a20cd6andv3.0.0-rc.8(2f1c0c9c), both from the signed AppImages, built locally with the bundled builder (--cli --update). Both builds carry the same small local audio-only patch (a gain on the Opus decoder output); the runs were muted with SDL's dummy audio driver. v3.0.0 is rc.8 plus 7 commits (icons, FSR HUD-less capture, DLAA, CI and docs), so I have not run it separately.Settings, identical for both: Custom preset, handheld profile, Render resolution 75 %, Anisotropic filtering 3, Sharpness 75, DLSS off,
frame_generation=2,frame_generation_by=1withHOS_FPS_MAX=30(30 real + 1 made frame), V-Sync Auto, Reflex 2, Frame pacing on, fullscreen, Threaded rendering Auto (on in both, 6 cores / 12 logical processors).longshot_env.txt:RX_VK_NO_READBACK_WAIT=1 RX_VK_COMPACT=1 RX_VK_TEX_MAX_DIM=1024 RX_VK_AUTO_TUNE=0 RX_VK_MOVIE_UPSCALE=linear RX_VK_BC_IDLE_S=0 RX_VK_REPLAY_PIN=0 RX_VK_MIRROR_SPARSE=0 RX_VK_MIRROR_MB=512 RX_VK_WB_MAX_KB=256 RX_VK_IMAGE_BLOCK_MB=32 HOS_RESERVE_CORE=0. GPU offsets +150 MHz core / +1000 MHz memory (NVML), CPU PL1 held at 36 W,intel_pstatemax_perf_pct 80. These are the settings I play with; I have not retested with the defaults.Method: the same save, child Link standing in the Market. After the load, 30 s of settling, then 60 s measured, with a scripted input: the camera circles left and right the whole time, and Link walks 1.5 s forward and 1.5 s back every 8 s. Three runs per build, the CPU cooled below 52 °C before each. Present times are from
RX_VK_FRAME_LOG; real fps, GPU time and replay CPU time are from the game's log.The average frame rate and the GPU and CPU time per frame are about the same; the difference is in the long frames. Every rc.8 run had more than every 486 run. For reference, nightly-459 (before #1553) gave 21.7 over 50 ms per minute in the same test, with slightly different settings (anisotropic filtering off, no GPU offsets). So 486 was the best of the three here.
The rc.8 changes that touch this path and that I'd suspect first, without having bisected: #1643 (temporary pipelines for required draws by default), #1596 (reusing live shader stages), #1605 (mid-size captures under tight spare budgets: this card is always near full VRAM), #1662 / #1663 (job barriers and YieldThread), #1664 (threaded rendering tiers). If one of these has a switch to turn it back, I can run the same three-run test with it off. I can also run with the profiling switches you asked the other reporter for.
Logs and the per-run frame logs are available on request. I'm not attaching them because they contain local paths.
The runs, the analysis and this write-up were done with an AI assistant (Claude Code) on my machine; I've reviewed the numbers.
From Discord:
Additional independent data point from Discord (2026-11-19): Longshot v3.0.0 stable, source 87cb36aa, on Windows 11 with i7-12700KF, RTX 5070 12 GB, NVIDIA 621.24, 32 GB RAM, 1440p/180 Hz G-Sync. Using DLSS Quality, DLSS FG 2x, 165 FPS cap, Reflex On, V-Sync Off, Threaded rendering Auto, the reporter measured the Market at night and day during route Kakariko → Hyrule Field → Market → Field → Kakariko. Kakariko was about 79 FPS with one >40 ms frame in 4m40s; Field about 82 FPS. Market night: median 69 FPS, approximately one >40 ms frame every 5s; day: about 2,000 draws vs 1,500 at night, median 56 FPS and around 20 frames >40 ms in 30s. VRAM 8.5/10.9 GB; no pipeline waits or translation waits. The log reports long core waits during Market hitches (up to 326 ms), with the reporter interpreting the short core-holder durations and core-1-only eligibility as a possible wake/handoff delay; this is an observation, not an established cause. Full log is attached to the Discord post but not included here. This is an additional system and build data point, not evidence that the underlying cause matches or is known.
From Discord:
Additional report from Discord (2026-11-20): Longshot v3.0.0 stable, source 87cb36aa, Windows 11 build 26200, Ryzen 7 5800X3D, 32 GB RAM, RTX 3080 Ti with NVIDIA 610.88. Reporter describes fence spikes in the Temple of Time as they move through the light-beam texture toward the altar. Settings include built-in frame generation 2x, DLSS Balanced at 1120x630 for 1080p output, fullscreen. Attached screenshot shows the Temple interior with bright beam-like light patches on the floor. Log reports 300 FPS samples, median 60, minimum 43, with sustained 54–60 FPS intervals; no HOS_THREAD_TIMES or RX_VK_PASS_TIMING profiling measurements are shown. Attach the existing Discord
longshot_log.txtandpicture.pngwhen updating the issue.Adding my v3.0.0 stable report details here (posted on Discord by SaKenLpB). I will attach the full log file (
2026-11-19 v3.0.0 session 2.txt) directly as soon as I'm back at my PC.System & Settings:
87cb36aa): DLSS Quality, DLSS FG 2x, 165 FPS cap, Reflex On, V-Sync Off, Threaded rendering Auto (on),RX_VK_FSR_FG=0.Route & Results (2 sessions):
Kakariko → Hyrule Field → Market (night, then Sun's Song to day) → Hyrule Field → KakarikoLog Observations:
PreActorModelCompose (26) waited 326.3 ms for a core: given by tera::FileResourceMgr (25), which held it 0.1 ms and was never the cores it may use now: core 1tera::FileResourceMgr (25) waited 310.8 ms: given by BaseProcCreateAndDeleteThread (23), which held it 0.0 msFrom Discord:
The reporter says they have now posted the v3.0.0 stable (source 87cb36aa) report details on #1639 and will attach the full
2026-11-19 v3.0.0 session 2.txtlog directly to the issue when back at their PC. Please attach that log there; no new symptoms or measurements are included in this Discord message.From Discord:
Additional report from the reporter (2026-11-20): after the main stutter improved substantially, a distinct severe frametime oscillation appeared immediately after playing the ocarina with Skull Kid in Lost Woods and again after entering Dodongo's Cavern near a lava pit. It is worse with FPS caps: removing the 60 FPS cap smooths frametimes; restoring 60 FPS triggers the issue, and a 30 FPS cap makes it worse. Saving and reloading the save clears it in both cases without restarting the build. The reporter could not capture a profile because the onset is unpredictable. No version/build or settings details for these observations were supplied. This may be distinct from the tracked general area-dependent stutter; please assess whether it belongs here or should be split.
From Discord:
Additional report from Discord (2026-11-21): The intermittent FPS-cap-related stutter has returned after about an hour or more of play, apparently on entering a new zone. The attached Bug report is Nightly 665-6e1521a7 (source a8ab8a56), LostWood_Child, RTX 3080 Ti with driver 621.24, Ryzen 9 5900X, 64 GB RAM, Windows 11 build 26300, 4K output/render resolution. The reporter says there was no deeper frame recording because onset is intermittent. report.txt lists only LONGSHOT_NO_UPDATE_CHECK=1; neither log contains HOS_THREAD_TIMES profiling. Attachments from the Discord post: report.txt, longshot_log.txt, longshot_log.1.txt, picture.png (from the report folder).
Updated with more from Discord: Added the Nightly 665 intermittent cap-dependent stutter report, profiling confirmation, and latest logs; noted relevant merged changes without claiming a fix.
Files: