Game: Camera reversal and combat cause repeatable FPS drops in Hyrule Field on RTX 4070 Ti #1720

Open
opened 2026-11-19 07:15:19 +00:00 by BeanBot · 12 comments
Member

Reported on Discord

Longshot version

v3.0.0 (source 87cb36aa)

Channel

Release (stable)

What happened and steps

The reporter sees a major FPS drop whenever using a button to switch the camera to the opposite side (a turn greater than 180°). They also report major FPS drops when starting fights with monsters in Hyrule Field; this was already present on the previous RC7.

Where in the game

Hyrule Field, child map (HyruleField_Child in the bug report).

How often

Camera drops occur each time they switch sides; combat drops when engaging monsters.

Graphics card and driver version

NVIDIA GeForce RTX 4070 Ti, NVIDIA driver 617.42.

CPU, RAM and operating system

Intel Core i7-14700KF (28 threads), 32 GB RAM, Windows 11 build 26300 (26H2).

Graphics settings

1920×1080 output, Docked profile, Custom preset, NVIDIA DLSS, Auto (dynamic) render resolution; frame generation 2x, game frame-rate cap 90, Reflex On, fullscreen On, V-Sync On. See report for other settings.

Logs and measurements

The bug-report log is from v3.0.0 (source 87cb36aa). It records 78 FPS samples: minimum 47, median 90, maximum 91, with isolated samples at 47 and 59–62 FPS. The report lists HOS_THREAD_TIMES=5 and RX_VK_SLOW_LOG=50, but does not confirm RX_VK_PASS_TIMING=1; available frame summaries do not provide the complete profile captured during the reported actions. Please follow https://longshot.magicbean.dev/help/performance/ and send a profile while reproducing each symptom, noting which action/place corresponds to each drop.

Still needed

Exact steps and location for the camera reversal and combat (including enemies involved), whether each happens on every repeat, and measured profile data while each symptom occurs. Please clarify whether “latest version” means v3.0.0 stable.

  • Longshot #1213: repeatable enemy-death spikes in Hyrule Field (RTX 5080, Linux); this report describes drops on engaging fights, so the symptoms may differ.
  • Longshot #1456: camera-rotation drops in Lake Hylia, a different place/condition.
  • Longshot #1406: direction-change stutter in Hyrule Field on an RTX 5060 Ti; similar, but not confirmed to be the same.

Steps taken and expected result

The reporter plays in Hyrule Field, reverses camera direction using a button and engages monsters; expected behavior is that these actions do not cause major FPS drops.

Attachments

Video: longshot-2026-11-19-07-49-04_Cv9z7GqX.mp4.
The bug-report ZIP is not copied; its useful contents are 2026-11-19_08-03-08/report.txt, longshot_log.txt, longshot_log.1.txt, picture.png, graphics.png and graphics_bis.png.

Reported on [Discord](https://discord.com/channels/1569849200047427645/1572872611967402085/1572872611967402085) ### Longshot version v3.0.0 (source 87cb36aa) ### Channel Release (stable) ### What happened and steps The reporter sees a major FPS drop whenever using a button to switch the camera to the opposite side (a turn greater than 180°). They also report major FPS drops when starting fights with monsters in Hyrule Field; this was already present on the previous RC7. ### Where in the game Hyrule Field, child map (`HyruleField_Child` in the bug report). ### How often Camera drops occur each time they switch sides; combat drops when engaging monsters. ### Graphics card and driver version NVIDIA GeForce RTX 4070 Ti, NVIDIA driver 617.42. ### CPU, RAM and operating system Intel Core i7-14700KF (28 threads), 32 GB RAM, Windows 11 build 26300 (26H2). ### Graphics settings 1920×1080 output, Docked profile, Custom preset, NVIDIA DLSS, Auto (dynamic) render resolution; frame generation 2x, game frame-rate cap 90, Reflex On, fullscreen On, V-Sync On. See report for other settings. ### Logs and measurements The bug-report log is from v3.0.0 (source 87cb36aa). It records 78 FPS samples: minimum 47, median 90, maximum 91, with isolated samples at 47 and 59–62 FPS. The report lists `HOS_THREAD_TIMES=5` and `RX_VK_SLOW_LOG=50`, but does not confirm `RX_VK_PASS_TIMING=1`; available frame summaries do not provide the complete profile captured during the reported actions. Please follow https://longshot.magicbean.dev/help/performance/ and send a profile while reproducing each symptom, noting which action/place corresponds to each drop. ### Still needed Exact steps and location for the camera reversal and combat (including enemies involved), whether each happens on every repeat, and measured profile data while each symptom occurs. Please clarify whether “latest version” means v3.0.0 stable. ### Related reports - Longshot #1213: repeatable enemy-death spikes in Hyrule Field (RTX 5080, Linux); this report describes drops on engaging fights, so the symptoms may differ. - Longshot #1456: camera-rotation drops in Lake Hylia, a different place/condition. - Longshot #1406: direction-change stutter in Hyrule Field on an RTX 5060 Ti; similar, but not confirmed to be the same. ### Steps taken and expected result The reporter plays in Hyrule Field, reverses camera direction using a button and engages monsters; expected behavior is that these actions do not cause major FPS drops. ### Attachments Video: `longshot-2026-11-19-07-49-04_Cv9z7GqX.mp4`. The bug-report ZIP is not copied; its useful contents are `2026-11-19_08-03-08/report.txt`, `longshot_log.txt`, `longshot_log.1.txt`, `picture.png`, `graphics.png` and `graphics_bis.png`.
Author
Member
Files from the Discord report: - [longshot_log.txt](https://git.magicbean.dev/attachments/34151289-159e-4569-9066-7b9748aa3ca1) - [longshot_log.1.txt](https://git.magicbean.dev/attachments/0f0cf01d-3c0b-43cf-9e5e-9147a1eda7a9) - [report.txt](https://git.magicbean.dev/attachments/ca1329ef-5aaa-44bc-8d73-26aa352c179c) - [picture.png](https://git.magicbean.dev/attachments/334960e5-8e50-44f3-a1ca-712a607b4e39)
Contributor

Claimed: I'm reproducing this on main on Windows 11 with an RTX 4070 and NVIDIA 616.56, using the report's settings (DLSS, Auto render resolution, frame generation 2x, cap 90, Reflex and V-Sync on). Each drop will be profiled (camera swung to the other side, then a fight starting) and compared with the same scene without frame generation. Prior work: none found. #1406 (direction changes, RTX 5060 Ti) and #1213 (enemy deaths, RTX 5080 Linux) may be the same cause; I'll say there if what I find covers them.

Claimed: I'm reproducing this on main on Windows 11 with an RTX 4070 and NVIDIA 616.56, using the report's settings (DLSS, Auto render resolution, frame generation 2x, cap 90, Reflex and V-Sync on). Each drop will be profiled (camera swung to the other side, then a fight starting) and compared with the same scene without frame generation. Prior work: none found. #1406 (direction changes, RTX 5060 Ti) and #1213 (enemy deaths, RTX 5080 Linux) may be the same cause; I'll say there if what I find covers them.
Contributor

Reproduced on Windows 11 with an RTX 4070 and NVIDIA 616.56, on main 868f6be8 with #1729. Adult Link in Hyrule Field, 2560x1440 output, DLSS, cap 90, Reflex and V-Sync on. A person played; the log had RX_VK_PASS_TIMING=1, RX_VK_SLOW_LOG=50 and HOS_THREAD_TIMES=5. The display is 240 Hz, and FIFO held at most 4 frames, so the refresh rate is not the limit.

Camera reversals drop frames with frame generation on, not off. Each line is one 120-present window of the log.

cap 90 standing reversals render thread waiting on the GPU
frame generation off 88-90 fps 70-88 fps ~1.5 ms a frame
frame generation 2x (Quality, and Auto) ~78 fps 56-64 fps ~7.9 ms a frame

Render resolution made no difference: Quality and Auto dropped the same way.

Where the time goes. Frame generation's copies and presents go to the game's own graphics queue, so they run in turns with the game's frames (docs/design.md section 13, rule 5). In the pass timing, the GPU time between the game's own work within a frame rises from ~1.37 ms with frame generation off to 3.4-4.2 ms with it on, while the game's work stays at 6.2-8.6 ms. At a 90 cap (11.1 ms) the GPU is full even standing still. A reversal adds draws (2,000-2,500 against ~1,250) and replay work (8-11 ms), which takes it to the drops reported. The readback waits are too small to explain it (0.07 ms each, about 0.2 ms a frame).

Fights drop frames even without frame generation: 63-67 fps at cap 90 with it off, and 46-50 at cap 60 with it off. Compute dispatches go from ~155 to 270-370 a frame and draws to ~3,500, both GPU (11-12 ms) and replay (9-13 ms) pass 90's budget, and the game's own limiter drops its nominal rate to 31-32 at times. In the same windows, the game's loading threads (tera::FileResourceMgr, PreActorModelCompose, aglTextureDataMgr::Compaction) wait 50-115 ms for a virtual core, as in #1610 / #1646. Whether that delays frames or only loading is not established.

Also seen: switching frame generation on freezes the game for one frame of 827 ms.

Next: an A/B of RX_VK_PRESENT_QUEUE=1 (presents on a compute queue, off by default since it lost the GPU on an RTX 5080 with 621.24) against the default, the same settings and place, on this RTX 4070, the reporter's generation. Results here.

Reproduced on Windows 11 with an RTX 4070 and NVIDIA 616.56, on main `868f6be8` with #1729. Adult Link in Hyrule Field, 2560x1440 output, DLSS, cap 90, Reflex and V-Sync on. A person played; the log had `RX_VK_PASS_TIMING=1`, `RX_VK_SLOW_LOG=50` and `HOS_THREAD_TIMES=5`. The display is 240 Hz, and FIFO held at most 4 frames, so the refresh rate is not the limit. **Camera reversals drop frames with frame generation on, not off.** Each line is one 120-present window of the log. | cap 90 | standing | reversals | render thread waiting on the GPU | |---|---|---|---| | frame generation off | 88-90 fps | 70-88 fps | ~1.5 ms a frame | | frame generation 2x (Quality, and Auto) | **~78 fps** | **56-64 fps** | **~7.9 ms a frame** | Render resolution made no difference: Quality and Auto dropped the same way. **Where the time goes.** Frame generation's copies and presents go to the game's own graphics queue, so they run in turns with the game's frames (`docs/design.md` section 13, rule 5). In the pass timing, the GPU time *between* the game's own work within a frame rises from ~1.37 ms with frame generation off to **3.4-4.2 ms** with it on, while the game's work stays at 6.2-8.6 ms. At a 90 cap (11.1 ms) the GPU is full even standing still. A reversal adds draws (2,000-2,500 against ~1,250) and replay work (8-11 ms), which takes it to the drops reported. The readback waits are too small to explain it (0.07 ms each, about 0.2 ms a frame). **Fights drop frames even without frame generation:** 63-67 fps at cap 90 with it off, and 46-50 at cap 60 with it off. Compute dispatches go from ~155 to 270-370 a frame and draws to ~3,500, both GPU (11-12 ms) and replay (9-13 ms) pass 90's budget, and the game's own limiter drops its nominal rate to 31-32 at times. In the same windows, the game's loading threads (`tera::FileResourceMgr`, `PreActorModelCompose`, `aglTextureDataMgr::Compaction`) wait 50-115 ms for a virtual core, as in #1610 / #1646. Whether that delays frames or only loading is not established. **Also seen:** switching frame generation on freezes the game for one frame of 827 ms. **Next:** an A/B of `RX_VK_PRESENT_QUEUE=1` (presents on a compute queue, off by default since it lost the GPU on an RTX 5080 with 621.24) against the default, the same settings and place, on this RTX 4070, the reporter's generation. Results here.
Contributor

The A/B of RX_VK_PRESENT_QUEUE=1: the same PC, settings and place as above (frame generation 2x, cap 90, Quality), with reversals played by hand. Each row is 120-present windows of the log.

frame generation 2x, cap 90 standing reversals GPU time between the game's work
presents on the game's queue (default) ~78 fps 56-64 fps 3.4-4.2 ms
RX_VK_PRESENT_QUEUE=1 (presents on compute family 2) 76-79 fps 61-72 fps 2.7-4.5 ms

It doesn't help: the extra GPU time on the game's queue stays. So it isn't the presents but frame generation's own GPU work (DLSS-G and its copies), about 3-4 ms a game frame on an RTX 4070 at 2560x1440. With frame generation on, about 4.6 ms a frame of the render thread's wait is on fences: it waits for those copies (design.md section 13, rule 4's frame slots).

What this means for the report: the camera-reversal drops are frame generation's GPU cost at a 90 cap. A 4070-class card can't make 90 game frames plus the generated ones, so the game sits near 78 standing, and a reversal's extra geometry takes it into the 60s. The FPS counter shows the game's frames, so it reads as a drop. Workarounds for now: frame generation with a 60 cap, or frame generation off at 90. Neither switch I tried removes the cost.

The fight drops are separate and happen without frame generation (above). I'm testing #1646 against them next, since the loading threads' core waits in those fights look like #1610.

For design.md section 13, rule 5: RX_VK_PRESENT_QUEUE=1 with frame generation 2x ran a whole session on an RTX 4070 (Ada) with NVIDIA 616.56 without losing the GPU, against the RTX 5080 with 621.24 that it is off for.

The A/B of `RX_VK_PRESENT_QUEUE=1`: the same PC, settings and place as above (frame generation 2x, cap 90, Quality), with reversals played by hand. Each row is 120-present windows of the log. | frame generation 2x, cap 90 | standing | reversals | GPU time between the game's work | |---|---|---|---| | presents on the game's queue (default) | ~78 fps | 56-64 fps | 3.4-4.2 ms | | `RX_VK_PRESENT_QUEUE=1` (presents on compute family 2) | 76-79 fps | 61-72 fps | 2.7-4.5 ms | It doesn't help: the extra GPU time on the game's queue stays. So it isn't the presents but frame generation's own GPU work (DLSS-G and its copies), about 3-4 ms a game frame on an RTX 4070 at 2560x1440. With frame generation on, about 4.6 ms a frame of the render thread's wait is on fences: it waits for those copies (design.md section 13, rule 4's frame slots). **What this means for the report:** the camera-reversal drops are frame generation's GPU cost at a 90 cap. A 4070-class card can't make 90 game frames plus the generated ones, so the game sits near 78 standing, and a reversal's extra geometry takes it into the 60s. The FPS counter shows the game's frames, so it reads as a drop. Workarounds for now: frame generation with a 60 cap, or frame generation off at 90. Neither switch I tried removes the cost. **The fight drops are separate** and happen without frame generation (above). I'm testing #1646 against them next, since the loading threads' core waits in those fights look like #1610. **For design.md section 13, rule 5:** `RX_VK_PRESENT_QUEUE=1` with frame generation 2x ran a whole session on an RTX 4070 (Ada) with NVIDIA 616.56 without losing the GPU, against the RTX 5080 with 621.24 that it is off for.
Contributor

Data from Linux, for whoever is on this (enum's claim stands): Fedora 45 (kernel 7.3.4), AMD Radeon RX 9070 XT (RADV GFX1201, Mesa 26.2.7), Ryzen 5 5600X (12 logical processors), 32 GB RAM, frame generation off, no frame cap. So this may be a different drop from the frame-generation one above.

On current main (819a5967, after #1121) with #1667's and #1666's probes cherry-picked, a scripted route out of the Castle Town gate into Hyrule Field with a 180° camera turn every 10 s, 8 runs:

  • Out in the field, the long frames are mostly the renderer's capture copy, bound by page faults. With RX_VK_CAPTURE_NT=1 no run had a frame of 100 ms or more (worst 50-68 ms, against 313-1444 ms), and the 1% low was 27.7-30.1 fps against 9.6-19.2. Details: #1667 (comment).
  • At the field entry, the drops are the game waiting on a GPU sync (nvnSyncWait at lr 0x8042cde0, the frame wait madebyjordan found), (on an older build the renderer was 4 submissions behind at those waits): 8-12 waits of 100 ms or more a run, up to 496 ms. They're there with #1121's gate off and on, and with NT off and on.

Every run is on record in #1749. Happy to run anything on Linux AMD.

Data from Linux, for whoever is on this (enum's claim stands): Fedora 45 (kernel 7.3.4), AMD Radeon RX 9070 XT (RADV GFX1201, Mesa 26.2.7), Ryzen 5 5600X (12 logical processors), 32 GB RAM, frame generation **off**, no frame cap. So this may be a different drop from the frame-generation one above. On current main (`819a5967`, after #1121) with #1667's and #1666's probes cherry-picked, a scripted route out of the Castle Town gate into Hyrule Field with a 180° camera turn every 10 s, 8 runs: - **Out in the field, the long frames are mostly the renderer's capture copy**, bound by page faults. With `RX_VK_CAPTURE_NT=1` no run had a frame of 100 ms or more (worst 50-68 ms, against 313-1444 ms), and the 1% low was 27.7-30.1 fps against 9.6-19.2. Details: https://git.magicbean.dev/MagicBeanDev/Longshot/pulls/1667#issuecomment-54713. - **At the field entry, the drops are the game waiting on a GPU sync** (nvnSyncWait at lr 0x8042cde0, the frame wait madebyjordan found), (on an older build the renderer was 4 submissions behind at those waits): 8-12 waits of 100 ms or more a run, up to 496 ms. They're there with #1121's gate off and on, and with NT off and on. Every run is on record in #1749. Happy to run anything on Linux AMD.
Contributor

Follow-up on the field-entry drops on Linux (RX 9070 XT, RADV GFX1201, Mesa 26.2.7, same setup and route as above). The longest ones look like the GPU's ASTC work at the gate: the GPU is at 100% for ~450 ms there, and skipping the BC3 re-encode takes about a third off. Not the whole story: the decode's share is inferred, and a few long waits around the same moment are slow captures instead (below).

How it was found, on main 819a5967 (after #1121) with #1667's and #1666's probes, 2 runs a side unless stated:

  • The GPU really is behind, the CPU side is fine. In the worst wait (~460 ms) the renderer waits 250 ms for a submission's write-back and times out. A local probe at that timeout showed the submission recorded and sent, nothing held in a batch, and the GPU 3 recordings short of it. Turning off sparse mirrors (RX_VK_MIRROR_SPARSE=0), batched submits (RX_VK_BATCH=0) or the recorder thread (RX_VK_RECORDER=off, 1 run: the second was ended by the kernel's OOM killer, a memory jump I haven't looked into) changed nothing.
  • The GPU is at 100% through those waits. The card's busy percent (amdgpu gpu_busy_percent, every 50 ms) reads 100 for nearly every sample of the ~455 ms waits, against a median of 34-42% through play. The log's "GPU time" line (longest submission about 20 ms) doesn't show this work.
  • The ASTC work is the likely part. This card doesn't sample ASTC, so the log says it's decoded on the GPU, and the big textures are re-encoded to BC3 on the GPU too:
default RX_VK_ASTC_RGBA8=1 (no BC3 re-encode) RX_VK_ASTC_CPU=1
main-thread GPU-sync waits ≥100 ms 8, 7 5, 6 74, 73
worst 460, 463 ms 311, 326 ms 20.6, 20.7 s (GPU idle)
250 ms write-back timeouts 1, 1 0, 0 0, 0
measured part after entry (median of 2) 57.8 fps, 1% low 28.9 57.2 fps, 1% low 29.7 39.9 fps, 1% low 3.4

Measured: skipping the BC3 re-encode takes about 150 ms off the worst wait and removes the timeouts (2 runs a side, one PC). Not measured: what the GPU does in the remaining ~300 ms, still at 100%. The decode is the likely candidate, but no run took it off the GPU without moving it to the CPU, which is far worse here: that path is meant for iGPUs. Right after the worst wait, the upload budget still put off ~410 new ASTC textures to later frames (one run's log).

A guess at why the upload budget doesn't spread this out (from reading upload_defer on main, not tested): it bounds the CPU time spent in upload() (12 ms a frame), not the GPU time of the decode and encode, and it never defers terrain textures (layer arrays, or over 1024x1024). Capping the ASTC decode/encode GPU time per frame, or at least counting it in the GPU time line, might be where a fix goes.

A few other long waits have the GPU idle, and those are the capture again: about 2.5 s before the field's measured part, stepping out of the gate, the main thread sits 205-784 ms in nvnQueueSubmitCommands on a slow capture copy (36-85 MB at about 110-180 MB/s, with RX_VK_CAPTURE_NT=1 on), or waits on the Presentation Thread's. These copies have almost no minor faults (21-102) and spend half or more of their time off the CPU, so they aren't the page-fault cost NT fixed. My guess is the guest memory's pages being read back in (major faults, maybe from zram swap), but the capture probe counts only minor faults, so that's untested. I'm running that next.

Every run is on record in #1749 (run20-run39). Happy to run a fix on this card.

Follow-up on the field-entry drops on Linux (RX 9070 XT, RADV GFX1201, Mesa 26.2.7, same setup and route as above). The longest ones look like the GPU's ASTC work at the gate: the GPU is at 100% for ~450 ms there, and skipping the BC3 re-encode takes about a third off. Not the whole story: the decode's share is inferred, and a few long waits around the same moment are slow captures instead (below). How it was found, on main `819a5967` (after #1121) with #1667's and #1666's probes, 2 runs a side unless stated: - **The GPU really is behind, the CPU side is fine.** In the worst wait (~460 ms) the renderer waits 250 ms for a submission's write-back and times out. A local probe at that timeout showed the submission recorded and sent, nothing held in a batch, and the GPU 3 recordings short of it. Turning off sparse mirrors (`RX_VK_MIRROR_SPARSE=0`), batched submits (`RX_VK_BATCH=0`) or the recorder thread (`RX_VK_RECORDER=off`, 1 run: the second was ended by the kernel's OOM killer, a memory jump I haven't looked into) changed nothing. - **The GPU is at 100% through those waits.** The card's busy percent (amdgpu `gpu_busy_percent`, every 50 ms) reads 100 for nearly every sample of the ~455 ms waits, against a median of 34-42% through play. The log's "GPU time" line (longest submission about 20 ms) doesn't show this work. - **The ASTC work is the likely part.** This card doesn't sample ASTC, so the log says it's decoded on the GPU, and the big textures are re-encoded to BC3 on the GPU too: | | default | `RX_VK_ASTC_RGBA8=1` (no BC3 re-encode) | `RX_VK_ASTC_CPU=1` | |---|---|---|---| | main-thread GPU-sync waits ≥100 ms | 8, 7 | 5, 6 | 74, 73 | | worst | 460, 463 ms | 311, 326 ms | 20.6, 20.7 s (GPU idle) | | 250 ms write-back timeouts | 1, 1 | 0, 0 | 0, 0 | | measured part after entry (median of 2) | 57.8 fps, 1% low 28.9 | 57.2 fps, 1% low 29.7 | 39.9 fps, 1% low 3.4 | Measured: skipping the BC3 re-encode takes about 150 ms off the worst wait and removes the timeouts (2 runs a side, one PC). Not measured: what the GPU does in the remaining ~300 ms, still at 100%. The decode is the likely candidate, but no run took it off the GPU without moving it to the CPU, which is far worse here: that path is meant for iGPUs. Right after the worst wait, the upload budget still put off ~410 new ASTC textures to later frames (one run's log). A guess at why the upload budget doesn't spread this out (from reading `upload_defer` on main, not tested): it bounds the CPU time spent in `upload()` (12 ms a frame), not the GPU time of the decode and encode, and it never defers terrain textures (layer arrays, or over 1024x1024). Capping the ASTC decode/encode GPU time per frame, or at least counting it in the GPU time line, might be where a fix goes. A few other long waits have the GPU idle, and those are the capture again: about 2.5 s before the field's measured part, stepping out of the gate, the main thread sits 205-784 ms in `nvnQueueSubmitCommands` on a slow capture copy (36-85 MB at about 110-180 MB/s, with `RX_VK_CAPTURE_NT=1` on), or waits on the Presentation Thread's. These copies have almost no minor faults (21-102) and spend half or more of their time off the CPU, so they aren't the page-fault cost NT fixed. My guess is the guest memory's pages being read back in (major faults, maybe from zram swap), but the capture probe counts only minor faults, so that's untested. I'm running that next. Every run is on record in #1749 (run20-run39). Happy to run a fix on this card.
Contributor

@GrimVisage, thanks: the field-entry drop shows on NVIDIA too. This RTX 4070 (NVIDIA 616.56, Windows 11) also decodes ASTC on the GPU and keeps big textures as BC3. The log says ASTC is decoded on the GPU (Eden's decoder, Switch 2 tiling) and big ASTC textures are kept as BC3 (re-encoded on the GPU ...).

From one run on main 868f6be8 + #1729, frame generation off, cap 90, adult Link in Hyrule Field. Each row is 120-present windows of the log:

  • Field entry: the window right after arriving ran at 42 fps with 140 ASTC textures put off to later frames. That fits your entry drop.
  • Fights: 0 ASTC textures put off in every fight window (200+ dispatches a frame), and 0.5-5 new textures a frame. Yet the fights run 51-68 fps for several seconds. That's the steady cost of twice the draws (~3,500) and dispatches (270-370), plus the loading threads' core waits, not texture work.

So this issue now holds three drops:

Drop Cause so far Seen on
Camera reversals frame generation's GPU work on the game's queue (above) RTX 4070, frame generation 2x
Field entry, one long hitch ASTC decode + BC3 re-encode filling the GPU; slow capture copies RX 9070 XT Linux; RTX 4070 (140 put off)
Fights, seconds of low fps draws and dispatches doubled; loading threads wait 50-115 ms for a core RTX 4070, frame generation off

Your ASTC lead looks worth a fix of its own, with the decode and encode's GPU time counted in the upload budget, or at least in the GPU time line. I'm on the fights: an A/B of #1646 against main is under way, and I'll post it here. If you take the ASTC side, a run of it on this NVIDIA card is easy from here.

@GrimVisage, thanks: the field-entry drop shows on NVIDIA too. This RTX 4070 (NVIDIA 616.56, Windows 11) also decodes ASTC on the GPU and keeps big textures as BC3. The log says `ASTC is decoded on the GPU (Eden's decoder, Switch 2 tiling)` and `big ASTC textures are kept as BC3 (re-encoded on the GPU ...)`. From one run on main `868f6be8` + #1729, frame generation off, cap 90, adult Link in Hyrule Field. Each row is 120-present windows of the log: - **Field entry:** the window right after arriving ran at 42 fps with **140 ASTC textures put off** to later frames. That fits your entry drop. - **Fights:** **0 ASTC textures put off** in every fight window (200+ dispatches a frame), and 0.5-5 new textures a frame. Yet the fights run 51-68 fps for several seconds. That's the steady cost of twice the draws (~3,500) and dispatches (270-370), plus the loading threads' core waits, not texture work. So this issue now holds three drops: | Drop | Cause so far | Seen on | |---|---|---| | Camera reversals | frame generation's GPU work on the game's queue (above) | RTX 4070, frame generation 2x | | Field entry, one long hitch | ASTC decode + BC3 re-encode filling the GPU; slow capture copies | RX 9070 XT Linux; RTX 4070 (140 put off) | | Fights, seconds of low fps | draws and dispatches doubled; loading threads wait 50-115 ms for a core | RTX 4070, frame generation off | Your ASTC lead looks worth a fix of its own, with the decode and encode's GPU time counted in the upload budget, or at least in the GPU time line. I'm on the fights: an A/B of #1646 against main is under way, and I'll post it here. If you take the ASTC side, a run of it on this NVIDIA card is easy from here.
Contributor

The fight drops are not the virtual-core waits. A child-at-night Hyrule Field A/B of #1646 against main (Stalchildren fights, cap 90, frame generation off, the same PC; details on #1646 (comment)) took the loading threads' waits over 50 ms from 17 to 0-1, and fight fps stayed the same: 68-69 median in all three runs. So in a fight the frames are lost to the render cost of the extra objects. Draws go from ~2,000 to ~2,300-3,500, and the replay's CPU work rises from ~7 to 9-13 ms a frame, about 3.5 us a draw. The GPU adds about 1.5-2 ms, spread over many small dispatches and barriers, with no single pass to blame. At a 90 cap that is past the 11.1 ms budget. Summary for this issue: camera reversals are frame generation's GPU cost (above), and fights are the replay's per-draw CPU cost plus GPU growth. Neither is a single fault I can fix in one change; the per-draw cost of the replay is the larger lever.

The fight drops are not the virtual-core waits. A child-at-night Hyrule Field A/B of #1646 against main (Stalchildren fights, cap 90, frame generation off, the same PC; details on https://git.magicbean.dev/MagicBeanDev/Longshot/pulls/1646#issuecomment-55836) took the loading threads' waits over 50 ms from 17 to 0-1, and fight fps stayed the same: 68-69 median in all three runs. So in a fight the frames are lost to the render cost of the extra objects. Draws go from ~2,000 to ~2,300-3,500, and the replay's CPU work rises from ~7 to 9-13 ms a frame, about 3.5 us a draw. The GPU adds about 1.5-2 ms, spread over many small dispatches and barriers, with no single pass to blame. At a 90 cap that is past the 11.1 ms budget. Summary for this issue: camera reversals are frame generation's GPU cost (above), and fights are the replay's per-draw CPU cost plus GPU growth. Neither is a single fault I can fix in one change; the per-draw cost of the replay is the larger lever.
Contributor

Summary, and the claim released. The drops are reproduced on an RTX 4070 / Ryzen 9 7900X on Windows (details in the comments above). They have three causes, none of them one bug to fix here:

  1. Camera reversals: frame generation's GPU cost. DLSS frame generation's own work runs on the game's GPU queue, about 3-4 ms per game frame here. At a 90 cap a 4070-class card then can't reach 90 even standing still (~78 fps), and a reversal's extra geometry takes it into the 60s. Without frame generation the same reversals stay at 70-88 fps. Moving the presents to their own queue (RX_VK_PRESENT_QUEUE=1) doesn't change it.
  2. Fights: the replay's per-draw CPU cost. A fight roughly doubles draws and dispatches, and the replay (about 3.5 µs a draw here) runs past the 11.1 ms a 90 fps frame allows: 68-69 fps in fights with frame generation off. It is not the loading threads' core waits: #1646 removes those, and fight fps doesn't move.
  3. Entering the field: ASTC texture work on the GPU, a one-off hitch. GrimVisage's finding above, which they are following up.

For now (to the reporter): keep frame generation with a 60 cap, or turn it off at 90. Fights will still dip a little until the replay gets cheaper. No more profiles are needed from you.

Where the fixes are: the per-draw cost is the replay and draw path of the renderer rework, #1493 (our numbers: #1493 (comment)), and the thread redesign in #1763. On this PC #1763 currently makes fights 4-5 fps slower, probably from removing the recorder thread (#1763 (comment)). Whether this issue stays open as a pointer to those or closes as explained is for the maintainers.

I'm releasing my claim: I won't build a fix in this issue.

**Summary, and the claim released.** The drops are reproduced on an RTX 4070 / Ryzen 9 7900X on Windows (details in the comments above). They have three causes, none of them one bug to fix here: 1. **Camera reversals: frame generation's GPU cost.** DLSS frame generation's own work runs on the game's GPU queue, about 3-4 ms per game frame here. At a 90 cap a 4070-class card then can't reach 90 even standing still (~78 fps), and a reversal's extra geometry takes it into the 60s. Without frame generation the same reversals stay at 70-88 fps. Moving the presents to their own queue (`RX_VK_PRESENT_QUEUE=1`) doesn't change it. 2. **Fights: the replay's per-draw CPU cost.** A fight roughly doubles draws and dispatches, and the replay (about 3.5 µs a draw here) runs past the 11.1 ms a 90 fps frame allows: 68-69 fps in fights with frame generation off. It is not the loading threads' core waits: #1646 removes those, and fight fps doesn't move. 3. **Entering the field: ASTC texture work on the GPU**, a one-off hitch. GrimVisage's finding above, which they are following up. **For now (to the reporter):** keep frame generation with a **60 cap**, or turn it **off at 90**. Fights will still dip a little until the replay gets cheaper. No more profiles are needed from you. **Where the fixes are:** the per-draw cost is the replay and draw path of the renderer rework, #1493 (our numbers: https://git.magicbean.dev/MagicBeanDev/Longshot/issues/1493#issuecomment-56003), and the thread redesign in #1763. On this PC #1763 currently makes fights 4-5 fps slower, probably from removing the recorder thread (https://git.magicbean.dev/MagicBeanDev/Longshot/pulls/1763#issuecomment-56028). Whether this issue stays open as a pointer to those or closes as explained is for the maintainers. I'm releasing my claim: I won't build a fix in this issue.
Contributor

A correction to my last comment: the slow capture copies with the GPU idle are not zram swap. The runs that showed them (the ones with the GPU's busy percent sampled) ran in a cgroup with swap off (MemorySwapMax=0, its 20 GB cap never reached), so none of the game's memory could be swapped out; in the earlier runs, without that limit, swap use grew by at most 23 MB through a run (more only in the run the kernel's OOM killer ended). Two runs that counted the game's major faults didn't show those slow copies at all, so what they wait on is still open.

@enum thanks for the RTX 4070 run, that it shows there too helps. I'll take the ASTC side. One thing I found looking at it: the upload budget (upload_defer) can't spread all of it even with the GPU's time counted, because it never puts off textures over 1024x1024 or layer arrays (the ground flickered when they waited a frame, the comment there says). This log doesn't say how much of the entry frame's 70 GPU-decoded ASTC textures (155 MB) are of that kind, so the first step is counting it: the decode and re-encode's GPU time per frame, in the GPU time line, split by what the budget can and can't put off. Then the fix goes where that points. I'll post the PR here, and a run on your card would be welcome then.

A correction to my last comment: the slow capture copies with the GPU idle are not zram swap. The runs that showed them (the ones with the GPU's busy percent sampled) ran in a cgroup with swap off (`MemorySwapMax=0`, its 20 GB cap never reached), so none of the game's memory could be swapped out; in the earlier runs, without that limit, swap use grew by at most 23 MB through a run (more only in the run the kernel's OOM killer ended). Two runs that counted the game's major faults didn't show those slow copies at all, so what they wait on is still open. @enum thanks for the RTX 4070 run, that it shows there too helps. I'll take the ASTC side. One thing I found looking at it: the upload budget (`upload_defer`) can't spread all of it even with the GPU's time counted, because it never puts off textures over 1024x1024 or layer arrays (the ground flickered when they waited a frame, the comment there says). This log doesn't say how much of the entry frame's 70 GPU-decoded ASTC textures (155 MB) are of that kind, so the first step is counting it: the decode and re-encode's GPU time per frame, in the GPU time line, split by what the budget can and can't put off. Then the fix goes where that points. I'll post the PR here, and a run on your card would be welcome then.
Contributor

I've opened #1843 for this. ASTC mip levels of 1024 x 1024 or less now re-encode to BC eight layers a run, through a second, small scratch. The 2048 scratch and RX_VK_BC_LAYERS are unchanged. On an RX 9070 XT (RADV), each of Hyrule Field's never-put-off terrain arrays now takes 16-17.5 ms of GPU time instead of ~50 ms. The first commit adds an astc span to RX_VK_PASS_TIMING, which is how this was measured. There were no grid lines in 4 runs. Other hardware is listed in the PR's Still to test.

I've opened #1843 for this. ASTC mip levels of 1024 x 1024 or less now re-encode to BC eight layers a run, through a second, small scratch. The 2048 scratch and `RX_VK_BC_LAYERS` are unchanged. On an RX 9070 XT (RADV), each of Hyrule Field's never-put-off terrain arrays now takes 16-17.5 ms of GPU time instead of ~50 ms. The first commit adds an `astc` span to `RX_VK_PASS_TIMING`, which is how this was measured. There were no grid lines in 4 runs. Other hardware is listed in the PR's Still to test.
Author
Member

From Discord:

Discord thread adds corroborating reports from other players: one says Hyrule Field then Death Mountain Trail drops from 60 to about 20 FPS while running back/forth, slashing and fighting (RTX 3070, Ryzen 5 5600X, stable v3.0.0, default settings; message 1572990936563196006); another reports the same on RX 6600/CachyOS (1573010980852596918); another reports it on RTX 5070 Ti (1573032413867544699). These are useful reports of similar gameplay drops, but they do not provide profiling data or establish that all share the same cause as this issue's camera reversal and fight symptoms.

From [Discord](https://discord.com/channels/1569849200047427645/1572872611967402085/1572872611967402085): Discord thread adds corroborating reports from other players: one says Hyrule Field then Death Mountain Trail drops from 60 to about 20 FPS while running back/forth, slashing and fighting (RTX 3070, Ryzen 5 5600X, stable v3.0.0, default settings; message 1572990936563196006); another reports the same on RX 6600/CachyOS (1573010980852596918); another reports it on RTX 5070 Ti (1573032413867544699). These are useful reports of similar gameplay drops, but they do not provide profiling data or establish that all share the same cause as this issue's camera reversal and fight symptoms.
Sign in to join this conversation.
No milestone
No project
No assignees
3 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#1720
No description provided.