Game: Stutter in Market - Caused by - Replay thread saturated in Castle Town market with NPCs (~8 ms of 16.7) #1925

Open
opened 2026-11-20 22:00:46 +00:00 by ShbellyJean120 · 1 comment

Longshot version

v3.0.0 (source 87cb36aa)

Channel

Release

What kind of problem

Low fps or stutter

What happened

In the Castle Town market, with NPCs around the fountain, the game averages 57.6–58.9 fps against its 60 fps target, and about 1 in 8 frames presents late. With the diagnostic switches on, the limit looks like the Vulkan replay thread (78–90% of a core). The GPU is roughly half idle, there's no pipeline compilation during play, and the game's own threads have headroom. Closed-off dungeons run almost perfectly; dips happen in open areas with many NPCs

Steps to reproduce

  1. Load a save in the Castle Town market as child Link, by day, with NPCs around the fountain.
  2. Run in circles near the fountain for a few minutes.
  3. Watch the frame rate and frame pacing, or check the log with the switches above.

Expected

A steady 60 fps. The GPU has plenty of headroom and the game's threads aren't saturated.

Actual

  • 57.6–58.9 fps average per 5 s window.
  • About 1 in 8 frames late: mostly 17.8–30 ms, an occasional 30–40 ms, and >40 ms roughly once every 5–10 s.

Measurements (per 5 s / 600-present window, while circling)

Renderer

  • Replay thread: 78–90% of a core for the whole period. Recorder thread: 62–74%.
  • Replay time per frame: submit 10.0–12.9 ms, of which draw 6.0–7.6, descriptors 2.1–2.8, commands 2.0–2.6, programs ~0.9–1.1.
  • The render thread line shows the game waiting for room at ~400–515 of 600 presents (about 2.2–2.6 ms a frame). The game is waiting on the renderer.
  • Host time in GPU resource waits: 3.0–6.0 ms a frame, almost all timelines (summed across threads, so not directly frame time).

GPU

  • GPU time: ~7.2–8.4 ms a frame.

Pipelines

  • No compilation during play: 0 compiled, all pipelines from cache.

Game threads

  • sead::MainThread 21–28%, BaseProcExecutor0–5 22–36%.

Loading

  • ARes Loading 0–8%, OptimizeChangeBanc 5–17% (compared with 43% during the area load itself).
  • Roughly every 5 s, a 100–220 ms vcore wait on tera::FileResourceMgr / PreActorModelCompose. These may explain the rare >40 ms hitches, but they aren't the main cost.

Reading

On this hardware the replay critical path is the limit in busy scenes. Its CPU cost scales with draw count, which matches dips in NPC-heavy open areas while dungeons stay smooth. This isn't GPU, pipeline compilation or the game's own threads, and loading (as in #1831) looks like a minor contributor at most.

Where in the game

Market_Child (child) - Running in a circle around the fountain for roughly 90 seconds

How often

Every time

Graphics card and driver version

NVIDIA GeForce RTX 5070 Ti, NVIDIA 621.01

CPU, RAM and operating system

AMD Ryzen 7 7800X3D 8-Core Processor (16 threads), 31 GB, Windows 11 (build 26200)

Changes from the defaults

DLAA before the bloom: On (default: Off)
Game frame rate cap: Uncapped (default: 60 (the console's))
Fullscreen: On (default: Off)
V-Sync: Off (default: Auto)
Speaker layout: Stereo (default: Auto)
Layout preset: Xbox / PC (default: Nintendo (Def))
Depends on this PC (the values now, not counted as changes):
Quality preset: Custom (default depends on this PC)
Output resolution: 2560x1440 (native) (default depends on this PC)
Profile: Docked (default depends on this PC)
Upscaler: NVIDIA DLSS (default depends on this PC)
Render resolution: Native (DLAA, 2560x1440) (default depends on this PC)
Sharpness: 80% (default depends on this PC)
Window scaling filter: Not needed (window = output) (default depends on this PC)
Game speed: Real time (default depends on this PC)
Frame pacing: On (default depends on this PC)
Switches set (their values are in the attached report.txt): HOS_FPS_LOG, HOS_THREAD_TIMES, LONGSHOT_NO_UPDATE_CHECK, RX_VK_PASS_TIMING, RX_VK_SLOW_LOG.

longshot_log.txt

Log attached

Screenshots or video

Test location image attached

Before you send it

  • The problem is still there on the newest release or nightly.
  • I searched the open issues for it.
  • I'm not attaching game files, keys, the game's executable, its .bin caches or saves.
### Longshot version v3.0.0 (source 87cb36aa) ### Channel Release ### What kind of problem Low fps or stutter ### What happened In the Castle Town market, with NPCs around the fountain, the game averages 57.6–58.9 fps against its 60 fps target, and about 1 in 8 frames presents late. With the diagnostic switches on, the limit looks like the Vulkan replay thread (78–90% of a core). The GPU is roughly half idle, there's no pipeline compilation during play, and the game's own threads have headroom. Closed-off dungeons run almost perfectly; dips happen in open areas with many NPCs ## Steps to reproduce 1. Load a save in the Castle Town market as child Link, by day, with NPCs around the fountain. 2. Run in circles near the fountain for a few minutes. 3. Watch the frame rate and frame pacing, or check the log with the switches above. ## Expected A steady 60 fps. The GPU has plenty of headroom and the game's threads aren't saturated. ## Actual - 57.6–58.9 fps average per 5 s window. - About 1 in 8 frames late: mostly 17.8–30 ms, an occasional 30–40 ms, and >40 ms roughly once every 5–10 s. ## Measurements (per 5 s / 600-present window, while circling) **Renderer** - Replay thread: **78–90% of a core** for the whole period. Recorder thread: 62–74%. - Replay time per frame: submit 10.0–12.9 ms, of which draw 6.0–7.6, descriptors 2.1–2.8, commands 2.0–2.6, programs ~0.9–1.1. - The render thread line shows the game waiting for room at ~400–515 of 600 presents (about 2.2–2.6 ms a frame). The game is waiting on the renderer. - Host time in GPU resource waits: 3.0–6.0 ms a frame, almost all timelines (summed across threads, so not directly frame time). **GPU** - GPU time: **~7.2–8.4 ms a frame**. **Pipelines** - No compilation during play: 0 compiled, all pipelines from cache. **Game threads** - sead::MainThread 21–28%, BaseProcExecutor0–5 22–36%. **Loading** - ARes Loading 0–8%, OptimizeChangeBanc 5–17% (compared with 43% during the area load itself). - Roughly every 5 s, a 100–220 ms vcore wait on tera::FileResourceMgr / PreActorModelCompose. These may explain the rare >40 ms hitches, but they aren't the main cost. ## Reading On this hardware the replay critical path is the limit in busy scenes. Its CPU cost scales with draw count, which matches dips in NPC-heavy open areas while dungeons stay smooth. This isn't GPU, pipeline compilation or the game's own threads, and loading (as in #1831) looks like a minor contributor at most. ### Where in the game Market_Child (child) - Running in a circle around the fountain for roughly 90 seconds ### How often Every time ### Graphics card and driver version NVIDIA GeForce RTX 5070 Ti, NVIDIA 621.01 ### CPU, RAM and operating system AMD Ryzen 7 7800X3D 8-Core Processor (16 threads), 31 GB, Windows 11 (build 26200) ### Changes from the defaults DLAA before the bloom: On (default: Off) Game frame rate cap: Uncapped (default: 60 (the console's)) Fullscreen: On (default: Off) V-Sync: Off (default: Auto) Speaker layout: Stereo (default: Auto) Layout preset: Xbox / PC (default: Nintendo (Def)) Depends on this PC (the values now, not counted as changes): Quality preset: Custom (default depends on this PC) Output resolution: 2560x1440 (native) (default depends on this PC) Profile: Docked (default depends on this PC) Upscaler: NVIDIA DLSS (default depends on this PC) Render resolution: Native (DLAA, 2560x1440) (default depends on this PC) Sharpness: 80% (default depends on this PC) Window scaling filter: Not needed (window = output) (default depends on this PC) Game speed: Real time (default depends on this PC) Frame pacing: On (default depends on this PC) Switches set (their values are in the attached report.txt): HOS_FPS_LOG, HOS_THREAD_TIMES, LONGSHOT_NO_UPDATE_CHECK, RX_VK_PASS_TIMING, RX_VK_SLOW_LOG. ### longshot_log.txt Log attached ### Screenshots or video Test location image attached ### Before you send it - [x] The problem is still there on the newest release or nightly. - [x] I searched the open issues for it. - [x] I'm not attaching game files, keys, the game's executable, its .bin caches or saves.
Member

Possibly related, for context:

  • issue #1639 Game: Inconsistent performance and severe stuttering in Hyrule Castle Town / Temple of Time — Same market-area frame pacing issue; this report isolates replay-thread CPU as a distinct limit.
  • issue #996 Game: Huge performance hit just running short circles in specific spot of Hyrule Town Market — Market NPC-area performance loss during circling; different spot and hardware.
  • issue #1831 Game: stutter at the Castle Town fountain on quick direction changes: Hyrule Field and Hyrule Castle are loaded ahead and dropped again at every turn (related: #996, #1639) — Same fountain area but linked to repeated neighboring-area loading, unlike this replay-bound symptom.
  • PR #590 perf(vk): three small costs of the replay thread removed: unread bookkeeping without mirrors, 16 vertex bindings a change, a lock a sampled texture — Merged replay-thread CPU optimizations; prior general costs, not this specific measured scene.
  • PR #794 WIP: perf(vk): a stage's descriptors are kept across draws in a recording part while what they were resolved from stands — Open experimental descriptor caching aimed at reducing replay-thread draw costs.
Possibly related, for context: - issue #1639 Game: Inconsistent performance and severe stuttering in Hyrule Castle Town / Temple of Time — Same market-area frame pacing issue; this report isolates replay-thread CPU as a distinct limit. - issue #996 Game: Huge performance hit just running short circles in specific spot of Hyrule Town Market — Market NPC-area performance loss during circling; different spot and hardware. - issue #1831 Game: stutter at the Castle Town fountain on quick direction changes: Hyrule Field and Hyrule Castle are loaded ahead and dropped again at every turn (related: #996, #1639) — Same fountain area but linked to repeated neighboring-area loading, unlike this replay-bound symptom. - PR #590 perf(vk): three small costs of the replay thread removed: unread bookkeeping without mirrors, 16 vertex bindings a change, a lock a sampled texture — Merged replay-thread CPU optimizations; prior general costs, not this specific measured scene. - PR #794 WIP: perf(vk): a stage's descriptors are kept across draws in a recording part while what they were resolved from stands — Open experimental descriptor caching aimed at reducing replay-thread draw costs.
Sign in to join this conversation.
No milestone
No project
No assignees
2 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
MagicBeanDev/Longshot#1925
No description provided.