Merge branch 'main' into task/replayq

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Valiantsin DzerakhandClaude Opus 5.5 committed 2026-10-26 08:20:17 +03:00
commit 792cc378f5
6 files changed
+198 -53

No files matched your search

+20 -4
View File
@@ -34,7 +34,8 @@ other `totk.exe` or `ninja.exe` runs. The log's first and last lines record the
frame 600. `mem: class` (once at start) says whether the GPU counts as uma or discrete and lists its heaps and types.
- With `perf`, `mem: dirty` every 60 frames: what a VRAM mirror of the buffer-bound pools would upload (pages the CPU
wrote per frame and per submit, eager and lazy, by pool class), FlushMappedRange calls and bytes, and the mirror
size. The census costs about 1.1 ms of CPU per frame on the submitting thread; `perf_summary.py` summarises it.
size. The dirty count costs about 1.1 ms of CPU per frame on the submitting thread here (3 ms on an 8th-gen i7);
`perf_summary.py` summarises it.
- `TOTK_PROFILE=<path> TOTK_PROFILE_FRAMES=lo-hi`: sampling profiler (`runtime/prof_win.c`), read with
`tools/prof_report.py <profile> <totk.map>`. With `TOTK_PROFILE_RENDER=1` it samples the render thread only and keeps
call chains: the report adds inclusive time per function and the exe callers of the driver and system DLLs.
@@ -66,11 +67,26 @@ RenderDoc cannot capture: the driver hides `VK_EXT_external_memory_host`.
## Other GPUs
The RX 9070 is the only GPU here; the build also has to run on NVIDIA and older AMD (Vega, wave64 only). New Vulkan code
queries the feature, extension, format or limit it uses and keeps a working path or a logged fallback when it is
missing (`vk_pipeline.c` provoking vertex and subgroup size pinning are examples). Never hard-code what this GPU
The RX 9070 is the only GPU agents can run; the build also has to run on NVIDIA and older AMD (Vega, wave64 only). New
Vulkan code queries the feature, extension, format or limit it uses and keeps a working path or a logged fallback when
it is missing (`vk_pipeline.c` provoking vertex and subgroup size pinning are examples). Never hard-code what this GPU
reports: alignments, subgroup size, depth formats, `maxPushDescriptors`.
Numbers from this machine do not carry over. Measured on the user's other machines (2026-10-24/25, logs and analysis
in `build/mem-plan.md`):
- GTX 1050 Ti (4 GiB, 214 MiB BAR, no ReBAR; i7 8th-10th gen) and RTX 3090 Ti: GPU time 110-120 ms per frame in the
world against 4.3 ms here, because the GPU reads imported host pools over PCIe. The VRAM mirror (cd6f490) brings
the 1050 Ti to 10-11 ms; that machine is now CPU-bound at 25-45 fps (replay 25-30 ms per frame, backlog 0b).
- The same write-watch scan costs about 3x the CPU time it costs here.
- NVIDIA has no ASTC: first texture uploads decode on the CPU, 0.4-0.7 s load frames (backlog 5a.2).
- `cannot alias its pool memory` fills its 16-line log limit on NVIDIA and on the RX 9070 alike (1920x1080, 1600x900,
800x450 render targets); a suspect for the black patches of backlog 0a (`build/alias-plan.md`).
A run on another machine goes through the user: copy `build/release/TotK-perf-log.vbs` next to the `totk.exe` you
hand over. It starts the game with `TOTK_TRACE=perf` and writes `logs\perf-<date>-<time>.log`. Ask for the log; the
user cannot run scripted input there.
## Checks before handing back
`totk.exe --probe-kernel`, `--probe-vkexec`, `--probe-nvn-tex` and `--probe-shader` run from the build directory and
+149 -45
View File
@@ -4,6 +4,83 @@ Priorities: no artifacts and no crashes first, then 60+ fps without dips, then o
names its evidence; Maxwell item numbers refer to [research/maxwell.md](research/maxwell.md). Remove an item when it is
merged or dropped.
## Open work (2026-10-25)
Everything still to do, in one list. Items below this section give older detail. Merged into main on 2026-10-25:
shader memo and push skip, page-indexed pool resolve cache, region-only Read/WriteTexels, GOB-wide swizzle, Intel
legacy vertex-input fallback, all-thread profile chains (04fb67b); descriptor build skip and update templates
(283bc70; field replay 8.87 to 7.99 ms); ASTC decode in a compute shader and native ASTC where sampled (0a0584e;
monsters2 worst frame 340 to 139 ms); waiting draws' stage libraries on idle workers and skipping draws into passes
with a per-frame target (3753939; cold monsters2 worst 257 to 131 ms). Dropped after measuring: narrower guest-written
marks for WriteTexels and box-only image updates (re-uploads after WriteTexels are 0.003-0.05% of frame time;
`build/v121-astcgpu/texels-narrow.patch`, `build/v121-texbox/texbox-wip.patch`).
Measure before building: every item states the number that decides it.
**Freezes and frame time**
1. Replay thread (5a.1): `task/replayq`, `build/v121-replayq/HANDOFF.md`. Ported onto current main, probes pass; game
checks and async against sync timings unfinished. The laptop is replay-bound (25-30 ms per frame, item 0b).
2. Texture uploads in gameplay, now the largest cause of long frames: `task/texasync`,
`build/v121-texasync/HANDOFF.md`. Step 1 splits the cost (image and memory creation, CPU deswizzle, staging,
copies, locks) before choosing a worker-thread upload, GPU deswizzle, image pooling or a per-frame budget.
3. Virtual-core preemption as Horizon does (MacNativeGames 8bf0893 / e4b4601, PR #488: asynchronous preemption of an
owner in game code, priority inheritance, the SDK's preemption-disable protocol; Mac worst-1% frame 18.4 to 10.6
ms): `task/preempt`, `build/v121-preempt/HANDOFF.md`. Port only if the measured waits behind an owner in game code
are significant; our checkpoints at indirect calls may already cover most.
4. Start-up re-warm when the driver cache does not match the exe path or device (5a.3): `task/rewarm` (main merged
in, 0b3bc73), untested in game. Needed before the laptop package.
5. Shared shader list: the newer MacNativeGames lists collapse variants that differ only in texture kind into `"*"`
rows, which our runtime cannot match (it translates for the exact kind). `tools/prepare_shaders.py` reads `"*"`
as 2D (uncommitted), and `build/shader-lists/union.tsv` unions every list revision since b75be8c. Option: match
`"*"` rows at run time as MacNativeGames does (`prepare_shader_seed.py:60-76`).
6. Freeze census: a 10-15 minute user session on current main with `TOTK_FLIGHT=1`, `TOTK_TRACE=perf` and
`TOTK_INPUT_RECORD`, ranking every frame over 50 ms by cause. Decides items 7-9.
7. `TOTK_PIPE_SKIP=fullscreen` as default (cold monsters2: 0 frames over 100 ms, 1% low 33.9 to 41.7 fps; risk: one
frame without a full-screen pass the first time its pipeline appears). Decide from item 6.
8. ASTC per-frame decode budget (4b): the largest batch, the 169 M texel 8x8 array at boot, scales to about 375 ms of
GPU time on a GTX 1050 Ti. Split by mips with `minLod` if laptop runs (`TOTK_ASTC_TIME=1`) show decodes in play.
9. ASTC to BC7 re-encode on the GPU for 4 GB cards: the same array is 677 MB as RGBA8 against 42 MB of blocks.
10. Spin-wait gates (MacNativeGames a038e6b, from nxnative 066f2af / 6d34984, MIT): loops polling memory hand the
core to a queued thread after 50 us. Survey TotK 1.2.1 for such loops first.
11. Over the GPU memory budget, wait for running work and drain deferred releases before failing (MacNativeGames
4de7b4e); 4 GB cards.
12. Later: recording on several threads, descriptor buffers or bindless (touches the translator), CPU/GPU frame
overlap as LAS Recomp 13398126d (only with item 1; keeps the phantom-object guard), an x86-64-v2 build for CPUs
without AVX2/FMA (SavageWindRecomp ae42014b4).
13. Laptop package and test after items 1-5 (GTX 1050 Ti and UHD 630). Untested there: native ASTC on Intel, the
Intel `minStorageBufferOffsetAlignment` padding of the ASTC path, the legacy-driver vertex-input fallback.
**Bugs**
14. Yunobo fades out and vanishes from the main view about 45 frames after the user presses A (ball fades at
1335-1350, gone at 1352; his shadow-map draws stay). Repro: `run_game.py build/v121-yunobo --save
user-save-2026-10-24-yunobo1 --script build/recordings/yunobo-cast2.txt --frames 1360`. Ruled out: occlusion
counts, CPU readback of GPU targets, missing NVN entry points, GPU-side resources: the game itself stops issuing
his main-view draws. Leading hypothesis: game logic fades the model alpha (material c11[33].w) and drops it from
the queues on a CPU input (camera-near fade, GfxColFade, CancelDraw distance or an AI fade). Next: find his c11
buffer, write-watch c11[33].w to its writer, walk up in Ghidra; cheap check first: fps60 and real-time patches
off from present 1316. Debug tools uncommitted in `.worktrees/yunobo` (`build/v121-yunobo/debug-tools.diff`);
detail in `build/v121-yunobo/HANDOFF.md`.
15. Black patches for a moment on the GTX 1050 Ti (item 0a below, `task/alias`, `build/alias-plan.md`).
16. Lantern texture bug (item 6 below): waiting for the user's description or a cropped shot.
17. Validation: `cp_t_tcb_20` R32F view sampled with depth compare in `vkCmdDispatch` (correctness item 8), "VkDescriptorSet
0x0 does not have a binding for vp_c1" in `vkCmdDrawIndexed` (also in `build/v121/logs/val-base.log`), a fragment
write to Location 0 with no colour attachment.
18. The correctness queue below.
**Machine and housekeeping**
19. Faulty RAM: about 25 bugchecks in three days, also at 4800 without EXPO; each stick alone passed MemTest86 test
5. Since 2026-10-25 21:08 one stick (BD100C33) runs in channel A, single-channel, so timings from then on are not
comparable with dual-channel ones. Next: one stick in slot B; if clean, the pair is the fault (RMA the kit).
After the fix: clean `gen.py` and `build_game.py`, all diff tests and probes. Zero-filled cache files from the
2026-10-22 crash are in `build/quarantine-zero-cache-20261025`.
20. Waiting for the user's approval to commit: `docs/backlog.md`, `docs/agent-playbook.md`, `totk.ini.example`,
`tools/prepare_shaders.py`.
21. Clean up merged worktrees and build dirs (bindskip, astcgpu, pipewait, texbox): notes to `build/archive/`.
22. Paused branches: `task/menu` (in-game menu, 2 commits), `task/v143` (game version 1.4.3, item 22).
## Plan (agreed 2026-10-23)
Work the stages in order. At most three agents at once, at most two of them running the game. Done 2026-10-23 and
@@ -12,13 +89,32 @@ removed: validation-layer crash (`build/archive/v121-val`), frame-time summary i
and README, settings file `totk.ini` (`runtime/cfg.c`, `--probe-cfg`; a new variable needs a row in `known[]`;
`build/archive/v121-ini`), portable folder (`tools/package_build.py`, README build step 7; `build/archive/v121-port`),
freezes on new pipelines in fights (cold fight 53.9 to 59.7 fps mean, frames over 50 ms 10 to 0;
`TOTK_GPL_DRAW_BUILD=0` turns it off; `build/archive/v121-pfh`).
`TOTK_GPL_DRAW_BUILD=0` turns it off; `build/archive/v121-pfh`), agent tasks in git worktrees (`tools/worktree.py`,
`private-build` skill), rumble strength and curve (`rumble_strength`, `rumble_curve`, `rumble_triggers`), even
present cadence by back-edge pacing (fight intervals outside 16.7 +- 1 ms 24.5% to 2.4%, +3.3 ms latency;
`TOTK_TRACE=present`; `build/archive/v121-pace`), PC keyboard and mouse layout with the mouse camera (`input_map.c`,
`mouse_sens`; `build/archive/v121-kbm`; the game's button actions and the D-pad order 1-4 still to check in game).
**Stage 1: before the tester weekend**
0a. Black screen patches for a moment, seen by the user on the GTX 1050 Ti in the field (2026-10-25). Not the VRAM
mirror: a verify run that caught the glitch compared 4.45 million mirror pages with 0 mismatches. Leading
hypothesis: a new texture descriptor over a GPU-rendered texture fails `alias_compatible` and starts as a
zero-cleared own-memory image (`vk_texture.c` `entry_of`, `create_tex`, `init_image`). The run's two largest
`left uninitialised` counts (33, 41) fall in its last seconds, where the glitch was. Repro every time: open the
Purah Pad with a right stick click (keyboard C). Next: `task/alias`, per-alias
diagnostic lines and a `TOTK_TEX_FILL=magenta` knob, then a user run. Details: `build/alias-plan.md`.
0b. CPU-bound replay on weaker CPUs: with the mirror the 1050 Ti machine (8th-10th gen i7) renders the GPU part in
10-11 ms, but replay takes 25-30 ms per frame (5.2 ms on the RX 9070 machine), so 25-45 fps. The same machine
replayed in 7-13 ms at 8 fps in import mode. Mirror bookkeeping explains 2-3 ms. A `sched` run rules out core
contention: the render thread spends 79-88% of the wall time on the CPU inside `nvnQueueSubmitCommands`; guest threads
never preempt it and the OS takes about 5%. Suspects: NVIDIA driver cost per command, mirror and texture checks inside replay. Next: a
render-thread profile from the 1050 Ti (7f2f332 writes it on window close). Details: `build/cpu-plan.md`.
1. User play test of the prepared build: F11, menus, movies, rumble and gyro still to check. First report
(2026-10-23): flying smooth, short freezes in fights with monsters.
5a. Fight hitches, the user's main complaint. In a warm fight (`fight.txt`, frames 1500+) the game runs 58.3 fps mean,
5a. Fight hitches and stutters on new content, the user's main complaint. GPU memory mirror merged (cd6f490,
`build/mem-plan.md`; 1050 Ti GPU time 110-120 to 10-11 ms). Order now: 2 (ASTC), 3 (re-warm), 1 (replay thread,
also a candidate fix for 0b). In a warm fight (`fight.txt`, frames 1500+) the game runs 58.3 fps mean,
52.9% of frames over 16.7 ms, 6.3% over 20 ms; frames over 20 ms are 87% replay (19 ms) with no pipeline or
shader waits. The user's play added 0 shader translations and 29 full pipelines.
1. Uneven frames: replay off the critical path. Translate command buffers at `EndRecording` or on a queued replay
@@ -26,10 +122,23 @@ freezes on new pipelines in fights (cold fight 53.9 to 59.7 fps mean, frames ove
the new path off by default if the play test leaves any doubt. Code and probes done in `build/v121-rq`
(0 validation errors); no valid timings yet (its runs fell in the 21:00 file-cache corruption). Next: rebase on
runtime/, measure `fight.txt` warm with `--quiet`.
2. Streaming: the slow loading frames are first texture uploads (ASTC, 367 ms at frame 219 of `fight.txt`).
Measure where the time goes, then decode on the GPU or ahead of use.
2. Hitches when something new appears: first texture uploads decode ASTC on the CPU inside the frame (367 ms at
frame 219 of `fight.txt` on the RX 9070; 402-714 ms load frames on the GTX 1050 Ti, `build/mem-plan.md`).
Desktop AMD and NVIDIA have no ASTC; Intel Gen9-11 iGPUs do (check that `vk_texture.c` uses it when the
device reports `textureCompressionASTC_LDR`). Plan, after the VRAM mirror:
1. Measure per load: ASTC bytes, decode time, block-linear deswizzle time (`PERF_ASTC` already times decode).
2. Licence check: Granite (Themaister, MIT) is believed to ship an ASTC compute decoder; yuzu's
`astc_decoder.comp` is GPL, read only.
3. Compute-shader decoder (plain Vulkan 1.1 compute, no extension, so every GPU we run on). Oracle: decode
every ASTC texture of the 1.2.1 RomFS with it and with `astc_decode_ldr_mt`, bit-exact match required.
4. In `vk_texture.c`: raw blocks go to staging (4-16x less PCIe traffic), a dispatch writes RGBA8 into the
image; `TOTK_ASTC=cpu` keeps the CPU path. VRAM use stays as today (decoded RGBA8); GPU re-encode to BC7
(yuzu's ASTC recompression) only if 2-4 GB cards run out.
3. Driver pipeline cache after a copy or driver update: the AMD driver keys its cache by exe path, so a moved folder
stutters until `--prepare-shaders` runs again. `task/rewarm` (1f3cd49, WIP) stamps `shader-prepared.bin` with
device, driver and exe path and re-warms at start on a mismatch; builds, probes pass, never run in game. Next
steps in `build/v121-rewarm/HANDOFF.md`. Whether NVIDIA caches pipeline libraries the same way: item 8.
6. Lantern texture bug: waiting for the user's description or a cropped shot.
7a. Keyboard and mouse: the PC layout and mouse camera agreed with the user (details below). Sonnet.
**Stage 2: tester weekend (2026-10-24/25)**
@@ -54,7 +163,14 @@ freezes on new pipelines in fights (cold fight 53.9 to 59.7 fps mean, frames ove
17. Platform layer: no Win32 outside `runtime/platform/` (`tools/check_platform.py`: 1195 uses in 41 files; FAIL in
`motion_fusion.c`, `movie_probe.c`, `ov_movie.c`, `swkbd_win.c`, `vk_exec_draw.c`).
18. Linux: `runtime/platform/linux.c`, as SavageWindRecomp splits windows, linux and macos files.
18. Linux: `runtime/platform/linux.c`, as SavageWindRecomp splits windows, linux and macos files. Write tracking
(`host.h` `mem_written_pages`, `mem_freeze`/`mem_thaw`/`mem_frozen`, used by the VRAM mirror and texture checks):
freezing maps to `mprotect` + a `SIGSEGV` handler; a kernel write (`read()` into guest memory) returns `EFAULT`
instead of faulting, so the same thaw-before-read sites apply. Written pages: `PAGEMAP_SCAN` ioctl with
write-protect matching (kernel 6.7+, made for Wine's `GetWriteWatch`); older kernels: write-protect the mirrored
and texture-bound pages after each sync and mark them in the fault handler (Ryujinx), uploading pools the game
rewrites every frame whole instead of tracking them. Soft-dirty (`/proc/self/clear_refs`) is process-wide and
slow, last resort. Under Proton the Windows build already works: Wine emulates both.
19. Steam Deck: RADV offers Vulkan 1.3 and `VK_EXT_external_memory_host`; 4 Zen 2 cores, so expect 30-40 fps.
20. Launcher (details below).
@@ -117,48 +233,36 @@ freezes on new pipelines in fights (cold fight 53.9 to 59.7 fps mean, frames ove
## Details for later items
- Keyboard and mouse (plan item 7a, agreed 2026-10-23). Today the keyboard map is `input_win.c:34-36` (camera on
I J K L, attack on F) and there is no mouse input at all.
- Layout:
- In-game menu, agreed with the user 2026-10-23. Base: LAS Recomp's menu (MIT, same Vulkan base as ours):
`refs/LASRecomp/runtime/nvn/vulkan/imgui_glue.cpp` (Dear ImGui on our device and dynamic rendering),
`runtime/ui/menu.mm`, `option_rows.mm`, `controls_ui.mm`; SavageWindRecomp has the same menu. Opens with F1/Esc or
Minus+Plus. While it is open, the game gets no input and keeps running (no pause). Every setting is saved to
`totk.ini`. Drawn after the blit of the game image into the swapchain image (`gfx_window.c` around line 536).
Linux comes later (SDL3 window, item 17): the menu and settings code stays free of Win32; only a thin adapter feeds
window and input events to ImGui. Menu text goes through a translation catalog (SavageWindRecomp
`runtime/ui/i18n.mm`), English and Russian. Stages, one branch each:
1. Frame: ImGui, open and close, input, live settings with `cfg_save`, catalog; pages video (fullscreen and window
size, FPS counter, frame limit 30 or 60), controls (rumble strength and curve, gyro on or off), sound (volume),
quit.
2. The remapping page (`controls_ui.mm`; the bindings of `input_map.c` and
[controls.md](controls.md) become the defaults) and mouse sensitivity.
| Switch | Action in the world | Keys |
|---|---|---|
| left stick | move | W A S D |
| right stick | camera | mouse; arrow keys as a second source |
| Y | attack | left mouse button |
| ZL | shield, lock on | right mouse button, also Z |
| ZR | bow | F, also X |
| X | jump | Space |
| B | run, cancel | Left Shift |
| A | interact, pick up | E |
| L | abilities | Q |
| R | throw weapon | R |
| left stick click | crouch | Left Ctrl |
| right stick click | | C |
| D-pad | | 1 2 3 4 |
| Plus | inventory | Tab |
| Minus | map | M |
Later pages: AMD FSR 1 or bilinear scaling of the 900p image to the window (LAS Recomp's Windows upscaler); a
frame-time graph with the wait breakdown of `TOTK_TRACE=perf` (also on F10); VSync (FIFO) for 60 Hz displays
without VRR, if testers report double frames under back-edge pacing; gyro sensitivity.
Esc stays free for the in-game menu. Z and X duplicate ZL and ZR so menu pages turn with two neighbouring keys;
menu tabs (L, R) are Q and R. The mouse wheel stays unbound: L and R are abilities and weapon throw in the world.
- Before coding, check in the game what L/R, ZL/ZR, X/Y, the D-pad and the right stick click do in the inventory
and in the world, and correct the action column.
- Mouse camera: raw mouse deltas per frame become right-stick deflection, with sensitivity
(`TOTK_MOUSE_SENS`) and the larger-deflection rule against the gamepad. The cursor is hidden and confined to the
client area while the window is focused and released on focus loss and while F11 changes the window. Not true
mouselook: the game's stick curve and dead zone still apply.
- Update [controls.md](controls.md). Remapping goes into `totk.ini` (plan item 21).
Live settings: today each module reads its `cfg_get` string once and keeps the value (`rumble_cfg` in
`ov_input.c`). The menu needs `cfg_set` and a version counter that consumers check (cheap; per-frame readers
re-read only when it changes), and `cfg_save` that rewrites only the changed lines of `totk.ini`, keeps the user's
comments and adds a missing key under its section (LAS Recomp's `settings_save` rewrites the whole file). A key
set in the environment shows greyed out in the menu. Settings read only at start (RomFS mods, shader preparation,
operation mode) are marked "after restart".
- In-game menu (Dear ImGui over our Vulkan output, as SavageWindRecomp `runtime/ui/menu.mm`). Opens with F1/Esc or
Minus+Plus. While it is open, the game gets no input. Pages:
- fullscreen and window size;
- performance overlay on F10 (fps, or a frame-time graph with the wait breakdown of `TOTK_TRACE=perf`);
- volume;
- async shader builds;
- quit.
Settings live in `totk.ini`. Leave room for later pages: upscaler, frame-rate choices, remapping,
and the software keyboard moved onto SavageWindRecomp's `osk.mm`.
Later: async shader builds, the software keyboard moved onto SavageWindRecomp's `osk.mm`. Rendering above the
game's 900p is a separate study (a game-side resolution patch, as Switch mods do, or scaling every render target
ourselves); MacNativeGames has neither (render scale fixed at 1).
- Gyro: the user's logs show only Xbox One and Series X pads, which have no gyro. With a DualSense or Pro Controller,
check Steam Input (it presents pads as a virtual Xbox pad without sensors) and the game's own motion-aiming option.
- Launcher (plan item 20): one exe for Windows and Linux, SDL3 and Dear ImGui (SavageWindRecomp
`runtime/ui/launcher.mm`, `tools/fetch_zig.sh`).
- The player picks the base game (XCI or NSP), the 1.2.1 update and keys. The launcher checks the version,
+1
View File
@@ -52,6 +52,7 @@ A one-step ordering difference between two engine threads can show as one frame
| S4 | rare hang or lost wake | `SignalToAddress` type 2, count <= 0 with waiters: Ryujinx, yuzu and nxnative set `value - 2` | done: `kernel.c:1017` writes `value - 1` and the count > 0 branch writes `nw <= count ? value - 1 : value`, both as mesosphere master `kern_k_address_arbiter.cpp`; 1.2.1 calls type 2 only with count 1, and `value - 2` did not change the phantom rocks (2026-10-23, `build/archive/v121-lock/NOTES.md`) | none |
| S5 | the main thread holds a core a worker needs | yield types 0 / -1 / -2 differ on Horizon; TotK's render-graph wait is a poll plus yield | unknown: all three hand the core only to an equal or better waiter (`kernel.c:835-846`) | trace which yield type 1.2.1's frame-end poll uses |
| S6 | lost or extra wakes | MacNativeGames `dab728a`: `SignalToAddress` wake-all, `SystemEvent` rebinding, targeted wakes | unknown | read the commit, compare with `kernel.c` |
| S7 | stutter: a better thread waits milliseconds for a core whose owner runs game code | MacNativeGames 8bf0893 (PR #488, on by default e4b4601): the waiter signals the owner, which gives the core up when interrupted in top-level game code; priority inheritance through guest mutexes; the SDK's preemption disable (TLR+0x100 / +0x102, SVC 0x36); without the last two their game hung at "Continue". 1.1.2: 58% of long core waits were this class; Mac worst-1% frame 18.4 to 10.6 ms | ported on `task/preempt` (`TOTK_ASYNC_PREEMPT`, SuspendThread and a register-saving stub instead of a signal; `build/v121-preempt/NOTES.md`). `TOTK_CORE_DIAG=1` measured this class at 0.47 / 0.83 / 1.06 ms of thread-time per frame (field / fight / monsters2) with our indirect-call checkpoints, waits to 16+ ms; with the port 0.29 / 0.34 / 0.84 and the waits of 1 ms or more 53 / 41 / 27 to 4 / 10 / 18. 1.2.1's SDK has the disable protocol (sdk f_70051242a8, f_7005124360). A first stub clobbered the flags before saving them (`sub` before `pushfq`) and crashed at boot | see the backlog item for the A/B timings |
## Open: graphics
+24 -3
View File
@@ -55,6 +55,7 @@ typedef struct Tex {
uint64_t wr_fence; // any entry: newest submission recorded to write its image (render, copy, clear,
// storage binding); ReadTexels waits for it instead of draining the replay queue
uint8_t listed; // in owned[]
uint32_t wt_n; // WriteTexels into the storage since the last upload
} Tex;
// A slot keeps what vk_texture_by_id resolved for it, so a draw binding skips the key lookup and the address chain: the
@@ -82,6 +83,8 @@ static int nsslots;
static nvn_hmap samplers = {.klen = sizeof(samp_par)};
static int nreleased;
static uint64_t n_uploads, n_reuploads;
// Re-uploads since the last perf line (tex_lock): all, and those of textures a WriteTexels wrote since their last upload.
static uint64_t ru_n, ru_bytes, ru_ns, ru_wt_n, ru_wt_bytes, ru_wt_ns;
// TOTK_TEX_ALIAS (default on; =0: a new image over live storage starts empty): a new descriptor over the storage of a
// live image that the GPU wrote gets that image's content (alias_copy). Counts since the last log line.
static int tex_alias = -1;
@@ -516,7 +519,7 @@ static void upload(Tex *e) {
const nvn_texdesc *d = &t->d;
int bw, bh, bpb = nvn_fmt_bpb(d->format, &bw, &bh);
if (!bpb || !t->mem) return;
uint64_t p0 = perf_begin();
uint64_t p0 = perf_begin(), u0 = kern_now_ns();
int linear = nvn_tex_is_linear(d), is3d = d->target == NVN_TEXTURE_TARGET_3D, one_d = e->pub.type == VK_IMAGE_TYPE_1D;
uint32_t levels = linear ? 1 : e->pub.levels, layers = linear ? 1 : e->pub.layers;
int astc = e->f.astc_bw != 0, rgb = e->f.expand_rgb;
@@ -596,8 +599,21 @@ static void upload(Tex *e) {
free(raw_copy);
free(ar);
free(rg);
if (e->uploaded) n_reuploads++, perf_count(PERF_CNT_TEX_REUP);
else n_uploads++, perf_count(PERF_CNT_TEX_UP);
uint64_t dt = kern_now_ns() - u0;
if (e->uploaded) {
n_reuploads++, perf_count(PERF_CNT_TEX_REUP);
ru_n++, ru_bytes += t->L.size, ru_ns += dt;
if (e->wt_n) {
ru_wt_n++, ru_wt_bytes += t->L.size, ru_wt_ns += dt;
if (rt_trace & RT_TRACE_PERF)
rt_log("perf: frame %llu: re-upload after %u WriteTexels: texture %#llx fmt %#x %ux%u, %.1f KiB, %.3f ms",
(unsigned long long)__atomic_load_n(&gfx_counters.frames, __ATOMIC_RELAXED), e->wt_n,
(unsigned long long)e->pub.addr, d->format, d->width, d->height, t->L.size / 1024.0, dt / 1e6);
}
} else {
n_uploads++, perf_count(PERF_CNT_TEX_UP);
}
e->wt_n = 0;
e->uploaded = 1;
e->n_up++;
perf_end(PERF_TEX, p0);
@@ -976,6 +992,10 @@ void vk_texture_new_submit(void) {
(unsigned long long)f * 60, (unsigned long long)n_alias_copied, (unsigned long long)n_alias_upload,
(unsigned long long)n_alias_uninit);
n_alias_copied = n_alias_upload = n_alias_uninit = 0;
rt_log("perf: frame %llu: texture re-uploads %llu (%.1f MiB, %.2f ms), after WriteTexels %llu (%.1f MiB, %.2f ms)",
(unsigned long long)f * 60, (unsigned long long)ru_n, ru_bytes / 1048576.0, ru_ns / 1e6,
(unsigned long long)ru_wt_n, ru_wt_bytes / 1048576.0, ru_wt_ns / 1e6);
ru_n = ru_bytes = ru_ns = ru_wt_n = ru_wt_bytes = ru_wt_ns = 0;
uint64_t ck, calls, untr;
vk_mem_snap_stats(&ck, &calls, &untr);
rt_log("perf: frame %llu: binding storage checks %llu (texture %llu), write-watch snapshot queries %llu, untracked %llu",
@@ -1180,6 +1200,7 @@ void vk_write_texels(const tex_t *t, const view_t *v, const NVNcopyRegion *r, co
log_limited(&n_bad, "WriteTexels: region outside level %u of texture %#llx", g.level, (unsigned long long)e->pub.addr);
vk_guest_written((uintptr_t)t->mem, t->L.size);
if (e->pub.gpu_owned) image_texels(e, &g, r, (uint8_t *)src, 1);
else if (e->uploaded) e->wt_n++;
} else if (e) {
image_texels(e, &g, r, (uint8_t *)src, 1);
}
+2 -1
View File
@@ -94,7 +94,8 @@ def answers(deps, stage, control):
"""GfxShaderAnswers in gfx_shader.c pack_answers layout (147 bytes) from the list's deps JSON."""
tex, img, fmt, attr, fout = bytearray(32), bytearray(32), bytearray(32), bytearray(32), bytearray(8)
for slot, t in deps.get('tex', {}).items():
tex[int(slot)] = TEX.index(t.removesuffix('_depth'))
t = t.removesuffix('_depth')
tex[int(slot)] = TEX.index('2d' if t == '*' else t) # "*": any kind, translated as the plain 2D it stands for
for slot, o in deps.get('img', {}).items():
img[int(slot)] = TEX.index(o['type'])
fmt[int(slot)] = FMT.index(o['fmt']) if o.get('fmt') in FMT else 0
+2
View File
@@ -41,6 +41,8 @@
# rumble_strength = 130
# rumble_curve = 0.6
# rumble_triggers = 0
# Mouse camera sensitivity (0 to 10; 0 = no mouse camera and no cursor capture).
# mouse_sens = 1
# Remap buttons, one line per Switch button: bind_<A B X Y L R ZL ZR PLUS MINUS LEFT UP RIGHT DOWN LCLICK RCLICK> =
# comma separated pad:<name> and key:<name> entries, or none. A line replaces that button's defaults (pad and keyboard);
# buttons without a line keep them. Invalid entries are logged and skipped.