The part command buffers and their pool were used by two threads. The submitting thread still waits for the
submission, so the queue order is unchanged. TOTK_REC_END=0 brings the old split back.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Storage-buffer stores, copies and fills land in imported guest memory without a fault or a write-watch bit, so a
region only the GPU wrote looked idle and was frozen while the GPU kept writing it. LAS Recomp never changes the
protection of imported memory because of NVIDIA. A possible cause of the NVIDIA "write invalid" device faults
(#788, #802), not proven.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A game resolution that is not 16:9 widened the camera while the HUD stayed at 16:9 stretched across the picture whenever romfs-patches-<w>x<h> was not next to romfs-index.bin, which is always so for a build from source. resolution.c now falls back to the largest 16:9 size inside the asked one and keeps the camera aspect patch off, logging the command that makes the layout. build_game.py copies every romfs-patches* folder with a manifest.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Keep 1920x1080 as the default with prompts = xbox or pc: the setup builds those prompts' folders for 16:9 only, so the
shaped default would have swapped the chosen prompts for the Switch ones.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The 72 mouse_direct sites of patches-v143.txt mapped to the 1.2.1 main by enclosing function, each original word
checked: the camera input update, free orbit, Ultrahand and the 16 player camera modes. 1.2.1 keeps some values in
other registers and stack slots (camera input x in s11, orbit angles at sp+0x3c/0x40, Ultrahand x in s9), so the
hooks read a per-version MouseLayout; the 1.4.3 values and behaviour are unchanged.
Where 1.2.1 adds a turn in a shared helper, the degrees go to the store of the helper's result (Aiming's pitch
target, new ROLE_BASE_STORE) or to the turn it is given (Horse yaw). Wakeboard, ShieldSurfing and Tail read the stick
twice (REPEAT on the first read), the dialogue cameras' angle loads are pre-indexed (mov -> add, as ClimbObj).
The player state (ability, bow in hand) is not known on 1.2.1 and reads as not found.
tests/mouse_direct_check.py runs the hooks with both layouts and checks each load-site table against its patch file.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PR #753 unified the cameras at 70/1500 = 0.0467 degrees a count and doubled Ultrahand's rate, which made every
existing mouse_sens turn the free camera 42% slower. The free camera keeps 0.08 degrees a count at mouse_sens 1, every
camera #753 unifies takes the same 0.08, and the held object keeps 0.01 stick units a count over the frame ratio (with
#753's cap of three full tilts a step). A mouse shake now needs 125 counts a swing at sensitivity 1 (10 degrees).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The stick loads returned the sample's pad stick even in a step without mouse motion, which drops anything the game
writes to the camera input's stick after the input update (main has 426 stores to +0x420/+0x424). The base loads added
0 degrees, which turns a -0 into +0. Both now return the loaded value untouched when the step has no mouse motion and no
mouse shake.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Resolved the conflict with #760 in runtime/input_sdl.c by keeping both the window enter/leave cases and the new wheel case.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The game loads each member of a pack (Pack/**/*.pack.zs) as a resource of its own, in a heap sized by that member's
own entry in the resource size table. The port raised the entries of the mods' loose files only, so a mod that changed
a file inside a pack and shipped no table for it left the game's entry. A crash when heavy snow started in Hebra was
that: an .ainb member of about 13 KB loaded into a heap sized for the game's smaller copy, 0x48 bytes stayed free after
the file, the frame heap gave null for the 0x68-byte object the resource builds next, and main+0x21794c stored that
object's vtable at address 0.
restbl_overlay now unpacks each pack a mod serves and the game's copy of it with the dictionaries of
Pack/ZsDic.pack.zs, and raises each member's entry as it does a loose file's: by the growth for a member the game's
copy has, by the estimate of its type for one it lacks. It never lowers an entry, a mod's own table included, and the
log names the first members it raised. --probe-mods checks restbl_pack_raise on archives of its own, and docs/mods.md
says that the files inside packs count too.
🤖 Generated with [ECA](https://eca.dev) (nubank-anthropic/opus-5.5 - max)
Co-Authored-By: eca-agent <git@eca.dev>
Under pressure only clean mirrored entries could go, so once those ran out the render targets stayed: 1.7-2.0 GiB of released images with the use still over the line. Most of them sit on storage the game has since given to another texture, and they are never taken back. The pressure path now drops those first, 30 s after their release and when no finalized slot has bound them for 600 submits. Clean entries wait 5 s, which ends the re-creation churn in fights, and both kinds stop once the use is back under the line.
TOTK_TRACE=perf gains mem: released lines that split released memory by kind, age, overlap and dedicated memory, and count take-backs and re-creations. TOTK_VRAM_LIMIT_MIB caps the budget the port sees, so the path can be driven on a large card.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The ClimbObj pitch hook replaced the pre-indexed load ldr s1, [x22, #0x88]! and wrote the write-back to x22 from the
host call. The lifter keeps X19-X29 in locals across a call and does not reload them, so the write-back was lost and
the following str s0, [x22] wrote the pitch over the camera object's vtable pointer; the next virtual call crashed
(seen climbing a Stone Talus). The mov x22, x19 before the load is now patched to add x22, x19, #0x88, the hook
performs a plain ldr s1, [x22], and load hooks no longer support write-back loads at all.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
146 of the 254 names in known[] (runtime/cfg.c) were described nowhere for a reader: diagnostics, test hooks, paths and
experiments. docs/settings.md gives each one what it does, the effect of changing it and a safety level. The rows come
from reading the code, and most effects have not been checked in a run.
tests/settings_doc_check.py fails when a known name is in neither file, when the page lists an unknown name, or when a
default on the page differs from metas[] or an rt_knob call.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On the map the mouse moves the cursor, so it had no way to zoom: the zoom is the right stick's y, which only the
keys and a pad reached. While the map is open, each wheel notch is now a full-tilt push of that stick for 100 ms with
50 ms of neutral after it (a turn away from the player zooms in). A fast scroll's notches wait their turn, at most four,
and a turn the other way drops them. The wheel stays unbound elsewhere.
The back ends count the notches while the cursor is captured (input_win.c from raw input's usButtonData, input_sdl.c
from SDL_EVENT_MOUSE_WHEEL), keeping a high-resolution wheel's part of a notch (ikb_wheel_take).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Only the free camera took the mouse as degrees. The other camera modes took it as stick input, which their own speeds,
frame time, slow motion and curves reshaped, with the on-foot camera's square-root pitch shaping on top: bow aiming was
far too sensitive vertically on slow movements, abilities and the Purah Pad camera eased their pitch, lock-on turned
27% slower, and climbing, shield surfing and the rest each turned at their own rate.
Each camera mode's update now loads the gamepad's part of the stick where it computes its turn and adds the mouse's
degrees where it adds that turn to its angle (runtime/mouse_direct.c camera_load, the load sites of
recomp/patches-v143.txt): aiming, scope, abilities (and Recall's cursor), horse, Purah Pad camera, lock-on, bow
lock-on, wall and object climbing, the waterfall's turn, shield surfing, wakeboard and the dialogue cameras. Recall's
cursor moves so that its aim turns as many degrees a count. The waterfall's height and the horse's empty lock-on are
not angles and stay stick input, now linear. A step whose mouse motion no hook took is logged with the camera's name.
Holding L to soothe the horse gave the mouse to the ability wheel's virtual cursor, a held stick for the camera: the
mouse goes to the wheel and the D-pad strips only while their screen has the focus (menu_cursor.c
menu_cursor_shortcut). A mouse shake no longer plays a 0.8 s swing after it: the right stick follows the mouse's own
swings and lets go when it rests, and the camera does not turn by that swing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
menu_cursor.c reads the pointer's place through input_wsys.h, which the
Win32 back end fills (GetCursorPos, SetCursorPos, raw-input travel) and
the SDL one stubbed, so on Linux a right click still went back (B is
injected before the place is read) but no item took the focus under the
pointer and a left click confirmed nothing.
input_sdl.c now keeps the pointer's place from the window's own mouse
events, as fractions of its size packed in one atomic word the game
thread reads; entering the window seeds it, and leaving or losing the
focus drops it. ikb_pointer_set warps through the window thread, as SDL
asks for its video thread, and the warp target stands as the place where
the compositor does not echo the warp. Mouse motion also counts towards
the device used last (ikb_mouse_travel), so moving the mouse alone
brings the pointer into the menus after a gamepad.
Co-Authored-By: GLM-5.2 <noreply@z.ai>
plat_fiber_entry is reached by a ret with rsp at top - 8 and called the fiber function without realigning, so the
function started with rsp = 0 mod 16 and every 16-byte local was off by 8. The VEX instructions of a v3 build
tolerate that; the SSE ones of a v2 build fault on it, and probe-kernel's fiber test crashed in same_regs on pcmpeqb
with a memory operand. plat_fiber_entry now aligns rsp to 16 bytes before the call.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
player_state follows about fifteen pointers from the player root to its action verbs, and each was checked with
plat_mem_readable, a VirtualQuery. That query costs tens of microseconds in this process, so the read stalled the
game thread while the mouse moved. mem_guest_readable answers from the layout for the committed heap and the loaded
modules, where the player and its components live, and asks the OS only for anything else. Together with the same
change in the menu pointer, a Ryzen 5 4500 runs the scene that fell to 50-54 fps at 58-60 fps again.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The map detection reads every open screen of eui::ScreenMgr every 100 ms, and each pointer it follows was checked
with plat_mem_readable, a VirtualQuery. That query costs tens of microseconds in this process, so each scan stalled
the game thread for milliseconds: on a Ryzen 5 4500 the mean frame rate fell from 59 to 50-54 fps and the 1% low from
44 to 23 fps. mem_guest_readable answers from the layout for the committed heap and the loaded modules, where the
game's objects and globals live, and asks the OS only for anything else. With it the same scene runs at 58-60 fps,
1% low 45-48 fps, as without the map detection.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The game detaches a fused part when the right stick swings left and right a few times. In direct mode the mouse never
reaches that stick, so shaking the mouse did nothing and only the shake key worked. A shake of the mouse while
Ultrahand holds an object (at least three changes of direction within 0.6 s, each swing worth 10 degrees of camera turn
or more) now plays the shake key's right-stick swing. Turning a held object back and forth at an ordinary pace takes
longer and does not count.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The free-orbit camera turned 0.08 degrees per count while the other cameras (the bow, the abilities) took the mouse as
0.01 stick units a count, about 0.023 degrees for yaw and through the camera's signed-square pitch for y, so the bow
turned at half the speed and small vertical moves barely moved it. Every camera now turns 70/1500 = 0.0467 degrees
per count: the stick path gets 0.02 a count for x and the signed square root for y, so pitch is linear and as fast as
yaw. While the player uses an ability (the action verb UseSpecialPower, read from the player actor with the 1.4.3
placement), the ability camera's own linear scales are undone instead. stick_sens is divided out, Ultrahand's held
object gets the camera's stick value a count (at most three full tilts a step), and mouse_aim_scale (0.1 to 3)
scales the mouse while the bow is drawn.
The game's lock-on, recenter and aiming override released the mouse's pitch hold and restored the target from before
it, so after a shot the camera sprang back to that pitch, or to the game's default target. The mouse now lets go
without restoring, and when the override ends (or the camera went 100 ms without a step) it holds the pitch the
camera has then. A gamepad's vertical input still hands the pitch back to the game.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A tab switch or a popup opening drops the game's focus box for a frame or two. The menu pointer read that as the menu
closing: the mouse camera captured the cursor again and, when the next tab took the focus, the pointer came back in the
middle of the window. A focus gap of up to 300 ms now keeps the pointer where it is.
The map has no focus box, so it stayed gameplay and the mouse drove the right stick, which zooms the map. The open
screens in eui::ScreenMgr's list now find the map (the Purah Pad's frame with AppMap_00 and the menu background; the
Scope opens AppMap_00 without the background), and there the mouse is the left stick, which moves the map's cursor.
The list is read every 100 ms, as each screen costs memory queries.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The level picker is the same function in 1.4.3 (main+0x1ee8424) with the same first words and an early return straight
to its ret, so the lod site takes the 1.2.1 words as they are. The graphics settings setup (main+0x5868c4) inlines the
getters that 1.2.1 calls, and their slow paths come back before the compare, so the steady site is the compare itself:
w10 is set to 9 there and the list walk is skipped to the epilogue, which does not read w10. The menu texts, README and
totk.ini.example now name both versions.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
nvnCommandBufferUpdateUniformBuffer between draws was replayed as a buffer copy at that point, which needs a barrier
against the draws that read the old bytes, and a barrier ends the open pass. The terrain and the sky islands' meshes
update before most of their draws: about 200 updates inside passes in a field submission, each a pass end, a full
barrier and a pass begin.
An update inside a pass now goes to the scratch buffer only. Later bindings of a range it covers bind the scratch
bytes (ub_bind, from ubo_info and bind_range); a binding it covers only in part, or at an offset the alignment
refuses, gets a merged copy of the snapshot or guest memory with the submission's updates on top. The copies into the
real buffers are recorded behind one barrier of their own at the next full barrier (vk_exec_ub_flush from
barrier_to). A new update of a range replaces the older overlay and pending copy of it, so the lists stay as short as
the blocks. Copies, fills and storage writes of the same bytes retire the overlays they touch, and a storage binding
over a pending copy brings the copies forward with a pass restart. bind_sig hashes the overlay each of the draw's
uniform and storage bindings reads, so a draw of an updated block binds the newest bytes and the others still skip
their descriptor build. TOTK_UB_OVERLAY=0 copies at once as before; ab alternates; TOTK_TRACE=perf prints the
counts.
1.4.3, RTX 4070 Ti, 2560x1440 DLAA, standing still on Great Sky Island, TOTK_AB=10 (tools/ab_summary.py):
17.03 ms a frame with the copies at once, 14.39 ms with the overlays (-15.5%, 58.7 to 69.5 fps); replay time on the
CPU 8.81 to 8.30 ms.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BspNcezsiV8zT8wcnXrrME
A VMA allocation that needs a new memory block costs 8-11 ms on the replay thread, so finished submissions now return their chunks to a pool that chunk_new draws from. Chunk sizes are powers of two so that one fits the next request.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
nvn_pool_flushed kept one span from the lowest to the highest flush of a pool and mirror_sync copied all of it. In
Kakariko one 252 MiB pool copied 22,121 MiB for 847 MiB flushed, in spans of up to the whole pool. Over PCIe 3.0 x8
(RX 6600, Ryzen 5 5500, issue #698) a typical span takes about 30 ms of GPU time, which fits the frame spikes there.
A pool now keeps up to 32 sorted ranges and merges the two nearest above that. Writes between ranges still reach the
mirror through the write watch. Over a walk in Kakariko the upload fell from 472.8 to 125.5 MiB and no flush range
reached 16 MiB. TOTK_FLUSH_SPAN=1 restores the span. The mirror summary gains a line with the upload by source, and
--probe-vkexec checks the range list and that two far-apart flushes copy only their pages.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
SPIR-V leaves a register read before any write undefined and Mesa folds undefined values freely. The unconditional patch 0014 guarded against that and was taken out in #701 because the whole picture shook with DLSS on NVIDIA. Patch 0016 brings the zero start back behind a zero_locals job flag, which vk_caps_driver turns on for RADV, ANV and NVK. TOTK_ZERO_LOCALS overrides it, and gfx_shader_set_host mixes it into the translator identity so the variant has its own cache keys.
FXAA_FS gains the flag-on hash of the game's FXAA shader, so vk_dlaa_skip_draw knows it on Mesa too. A game run with TOTK_ZERO_LOCALS=1 on an RX 9070 produced that hash; a TOTK_FXAA=0 run logged no skip with either hash, so the switch itself is not confirmed.
All 29736 stored 1.4.3 inputs translate with the flag off and on, 0 spirv-val failures. FXAA hashes to 5c43ae18948d52cb with the flag off and aad42238801d31d8 with it on. Flag-on output matches the old translator byte for byte on a sample of 312 shaders. Not run on a Mesa driver.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
TotK's fire spread reads its 260x260 RG8 grass maps after a fence wait with no invalidate. The fence write-back took targets up to 64 KiB only, so the CPU saw stale maps and fire died where it hit (#132).
A target over wb_fence_limit now goes back unless the GPU drew it in each of the last 16 frames. Raising the limit instead pulled in four 256x256 targets drawn every frame, about 2.7 ms a frame.
Every level and layer of every texture in a write-back now shares one upload batch and one wait. A burst of 9 grass maps took 7.8 ms with one round trip per level and takes 2 to 2.5 ms. The fence write-back on the field replay fell from about 45 to 22 ms per second of play.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The icon capture draws into a colour target rendered for the first time, with the per-frame depth buffer. Under
pipe_skip any and fullscreen the recent depth made the pass look per-frame, so draws with a pipeline still building
were dropped and the icon kept missing parts (a fused weapon without its blade) until the item was captured again.
vk_exec_targets_recent now ignores a recent depth target while a colour target is rendered for the first time.
On fight2 with a warm driver cache no draw waited. With a cold cache the first world frame after loading waits for
the scene's new HDR target (2.9 s without prepared shaders), and one 64x64 capture waited 29 ms in the fight.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The inventory's preview light target is drawn without depth before the bloom chain; a pass like that with nothing armed ends the scene once the bloom follows, so DLSS no longer runs on the finished menu picture. Arming on that target stays allowed, because at some camera angles the field's effects composite comes before the first pass with the scene depth. The G-buffer width is the widest of the frame, a small pass ends the scene only into a scene colour target, and no pair narrower than the widest G-buffer arms: with the scope open the half-size effects pass armed and DLSS scaled the effects into the light target.
Every texture initialization got a 128-byte host copy for recordings to point at, kept for the whole run. The game
initializes its transient targets again every frame (about 29 new copies a frame on 1.4.3), so host memory grew by
about 44 MiB per minute of play, standing still too, until the commit limit ended long sessions (#670).
Each pin now counts the recording copies whose ops name it: counted in flatten_recording, uncounted when a copy or a
per-submission clone is freed. A pin with no copy left is freed NVN_PIN_AGE (1200) submits after its last use, so a
recording still in progress keeps its pins, and never while the replay holds it as a current render target, which
carries over to the next submission and is compared by pointer. TOTK_PIN_SWEEP=0 keeps every copy.
A 300 s standing-still run on the field save: host memory without VRAM grew 5 MiB, against about 220 MiB before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>