Game: Severe intermittent FPS drops in Hyrule Field on RX 6700 XT (Linux) #1787
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#1787
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
nightly-589-a5fbd680 (source d85cf2cd), nightly.
What happened
The reporter sees severe intermittent FPS drops while playing in Hyrule Field, varying between roughly 9 and 45 FPS in the supplied video. The profiling overlay shows examples of 9 FPS / 360 ms, 12 FPS / 477 ms, 24 FPS / 63.3 ms and 45 FPS / 25.1 ms. The video shows movement and combat in the field. Exact route/actions during the capture and frequency over longer play were not provided.
Where in the game
HyruleField_Child (child).
Graphics card and driver version
AMD Radeon RX 6700 XT (RADV NAVI22), radv Mesa 26.2.6-arch3.1.
CPU, RAM and operating system
AMD Ryzen 5 5500 (12 threads), 15 GB RAM, CachyOS (Linux 7.3.4-1-cachyos), KDE (Wayland), Linux 7.3.4-1-cachyos.
Graphics settings
Custom preset; 1920x1080 native output and render resolution (100%); Docked profile; AMD FSR 3.1 upscaler; built-in frame generation (not default FSR 3.1 FG); frame pacing On; game's shadow resolution. Other settings are in attached report.txt.
Performance measurements
The reporter enabled HOS_FPS_LOG=1, HOS_THREAD_TIMES=5, RX_VK_PASS_TIMING (value omitted in report) and RX_VK_SLOW_LOG=50. The newly supplied bug-report folder contains another run log pair from nightly-589. The current
longshot_log.txt(564 KB; run_id 80a28f5c537bdda2012e337b65015306) records FPS samples 33, min 22, median 51, max 242, with drops below 36 FPS at lines 1632–1878 (minimum 22) and 2665 (33). Example replay timings at those drops are present in the log. Its end reports 3960 presents at 56.2 FPS and GPU time 10.2 ms/frame. The accompanyinglongshot_log.1.txtis a separate run (run_id 4784fab1fe804aa3227294372d722e37): 71 samples, min 31, median 59, max 243; end reports 8520 presents at 58.7 FPS and GPU time 10.5 ms/frame. The supplied video overlay shows more severe momentary readings; correlation between its readings and these log intervals is unknown. Neither log identifies the precise route/actions from the video.Attachments
longshot_log.txt— current run's profiling log from the newest supplied report folder.longshot_log.1.txt— preceding run's profiling log.report.txt— generated report with version, system, graphics configuration and enabled switches.picture.png— gameplay screenshot from the report.Still needed
Exact route/actions during the capture and how often the drops occur during ordinary play. The logs provide additional separate-run measurements but do not resolve the mismatch with the lowest video-overlay readings. The report provides version, system, graphics configuration and profiling-switch status.
Files from the Discord report:
Updated with more from Discord: Added the current-run longshot_log.txt and summarized its measured FPS range and drop intervals; it records GPU time 5.4 ms/frame at the end, distinct from the more severe video overlay readings.
Files:
Updated with more from Discord: Added the latest report folder's settings and two separate-run logs with their measured FPS and end-of-run GPU timings; exact route/actions and frequency remain unknown.
Files:
Claimed for the reproduction on RDNA 2 and the sparse-bind stalls: an RX 6700 XT, RADV (Mesa 26.2.7), Linux, the reporter's card. The ASTC side of the field's drops stays grimvisage's claim on #1720.
What the two attached logs show (nightly-589):
slow frame at present 2182: 536.0 ms; submit 354.8with ~25 ms of it itemized, in the 5 s window where the sparse mirrors' binds went from 257.9 to 595.2 ms (21 binds, ~16 ms each). Others arequeue submit 292.2and a 253.7 ms capture. Each window holds 2,000-15,000 blocks given memory and given back, on top of the ASTC decode and BC3 re-encode on the GPU at the field's entry (57-85 ms of uploads a frame, hundreds of ASTC textures left for later frames).sparse_bind_infosingpu/vk/vk_memory.cpp); RADV also offers a queue family for sparse binding alone.Prior work read: #1720, #1749 (grimvisage's RDNA 4 runs:
RX_VK_MIRROR_SPARSE=0changed nothing there), #1814, #1827, #1813, #1763. No PR on sparse binds or a sparse queue.Next: the same route here, with the #1749 switches (
RX_VK_CAPTURE_NT=1,RX_VK_ASTC_RGBA8=1,RX_VK_MIRROR_SPARSE=0) on RDNA 2, then the bind's time split into lock, flush and driver.#1843 may help with the Hyrule Field entry hitch in your logs. That hitch comes from the same two terrain arrays, which are decoded from ASTC and re-encoded to BC on the GPU. With the PR, that takes 16-17.5 ms a texture instead of ~50 ms on an RX 9070 XT. It doesn't change the drops in play. If you can, a run of the PR on your RX 6700 XT would help: enter the field from the Castle Town gate and look for grid lines on the far ground.