AMD's RDNA2 driver places the depth compression metadata after the depth data. Above 1080p that tail of a depth
target bound into the game's texture heap lay over the next texture, and every depth write scrambled it: a garbage
band at the top of the screen, a pink glow, blocks in the grass and a blue checkered scene in the menu (RX 6600 to
6800, issue #185). Reproduced at 2560x1440 on an RDNA2 iGPU with driver 26.10.1; all 63 oversize images there were
D16 or D32 depth targets. Colour images keep their place in the pool.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Eight places sized a pool (pipeline workers, library prebuild, translator prewarm and seeding, the multithreaded ASTC decoder) by GetActiveProcessorCount, every logical processor of the machine. A process limited with start /affinity, Task Manager or a launcher then started a pool for 16 CPUs on 4. plat_cpu_usable counts the affinity mask (all active processors for a process that spans processor groups); plat_cpu_count, the machine's count, stays for the system info. Idea from MMadmer's UsableCpuCount in a fork of the Shadow of the Colossus PC port.
The game pauses whenever its window loses focus, as it does on the Switch when Home is pressed, and ASSUME_FOCUS was the
only way to change that: a known key with no row in the settings table. The runtime reads it at each focus change rather
than once at start, so the row takes effect the next time the window is left, and the build guide says what an unfocused
window now does.
A Chinese IME switched on over the game attached to the window, then showed its candidate window and took the keys it
composes although there is nothing to type into here; Shift, which a player may have bound as the walk key, is one of
the keys such an IME takes. The software keyboard's edit box is a window of its own and keeps the IME it needs, so only
the game's window is detached.
AMD's Windows driver branch for RDNA 2 and older (vk_caps.amd_sc, driverInfo
"(AMD proprietary shader compiler)") applies alpha to coverage even while
vkCmdSetAlphaToCoverageEnableEXT sets it to false. TotK never turns alpha to
coverage on (a TOTK_DUMP frame showed it off in all 1170 pipelines), yet on an
RX 590 with driver 24.9.1 the character draws kept their depth and lost their
colour in the G-buffer, so Link and the NPCs were black silhouettes, and other
surfaces were dithered into checkerboards (issue #67, RX 470/570/580).
On that driver branch vk_pipeline.c ensure_init now leaves alpha to coverage out
of the dynamic state, so it goes into the pipeline key as on GPUs without EDS3.
TOTK_VK_DYN_A2C=0/1 (ini vk_dyn_a2c) forces either behaviour. NVIDIA and RDNA 3/4
are unchanged. Tested on the RX 590 with 1.4.3: characters, the checkerboards and
the black blocks all render correctly.
Fixes#67.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DcYzeu7eb2txzmu4brvcgh
tools/ultrawide_layout.py computes every edit from the canvas aspect, so 32:9 needs no new layout data: the installer now builds romfs-patches-3840x1080 and -5120x1440 beside the four 21:9 sizes, and the menu lists both. Before, a 32:9 output_res got the wide camera with the 16:9 HUD stretched over it (issue #181).
The pane copied each description into 512 bytes. The Russian button layout description is 840 bytes of UTF-8, so it stopped mid-sentence, and the English one (507 bytes) lost the note appended after it. The pane now holds UI_DESC_BYTES (2048), and probe_menu_text checks every English and Russian menu text leaves room for a note.
TOTK_DLSS (dlaa|quality|balanced|performance, read at start) sets the scene to
67, 58 or 50 % of the output. DLSS then runs at the begin of the game's scaling
pass (the FXAA shader drawing the light target into the output-sized target,
before the HUD) and its result replaces that draw; the HUD is drawn afterwards
and never jittered or passed through DLSS. The jitter now ends where the scene
ends (the bloom), separately from where DLSS runs. Scaled modes add a mip bias
of log2(scale) to repeating, mipmapped material samplers.
With a scene smaller than the output, fxaa = 0 no longer skips the game's
scaling draw (it left the 3D picture out).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The menu had English and Russian. pt-BR gets its own row with every menu id, a Português (Brasil) entry in the menu language list (TOTK_MENU_LANG = pt-BR), and is picked automatically when the Windows display language is pt-BR. The game language pt-BR now maps to this row instead of Portugal's pt, so the software keyboard shows Brazilian text too; pt keeps its row and the English menu. probe_menu_text checks the row is complete, keeps its overlay formats and follows the setting.
The aniso setting raised every filtered, mipmapped colour sampler to the chosen level. Some of those read screen-sized
buffers, among them the ones the inventory's character preview is drawn from. Their coordinates stretch steeply at the
model's edges, so the anisotropic taps reach across the edge and leave a speckled fringe around the character. The
level is now raised only for samplers that repeat on U and V, which the surface textures it is meant for do; clamped
samplers keep the game's level. Distant ground stays as sharp as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
stick_deadzone sets one dead zone for both sticks, and finding the value that stops a drift means guessing. A worn
stick usually drifts on its own while the other one is fine. lstick_deadzone and rstick_deadzone (0 to 50 percent,
live) set one stick each and fall back to stick_deadzone while unset, so existing totk.ini files behave as before.
input_pad_deadzone in input_map.c picks the value; input_pad_stick and its scaling stay as they are.
The Controls tab shows Left and Right stick dead zone rows in place of the single one. With a gamepad connected each
row shows its stick's deflection now, before the dead zone ("stick now"), so a drifting stick's resting value can be
read off and the zone set just above it. opt_slider takes the note line and a fallback key for that: a row whose key is
unset shows the fallback's value, and a change sets the key. probe-cfg checks the fallback, the clamp and
input_stick_percent. The menu text is in English and Russian.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The swapped letters are baked into the game's textures, fonts and messages, so a button remapped with a bind_ line or on the Controls tab keeps the old letter on screen. The menu description (English and Russian) and docs/controls.md now say so.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Xbox prompts folder is optional, so a manifest that no longer matches the RomFS logs the reason and drops that folder instead of stopping the game. The main romfs-patches and the mods folder keep their fatal checks.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With pad_layout = nintendo on an Xbox pad the right button does Switch A but is labelled B, the bottom one does B but is
labelled A, and the game's prompts still name the Switch buttons. A fourth Button layout choice, Xbox prompts (A right)
(pad_layout = xbox_prompts), maps the buttons as nintendo and swaps the prompts to the labels, so each prompt names the
button to press. Automatic, Nintendo and Xbox are unchanged and keep the game's own prompts. The buttons change at once;
the prompts at the next start, as the game reads its UI archive and fonts at boot (the menu says so).
At start, xbox_prompts loads romfs-patches-xbox in place of romfs-patches (runtime/romfs.c overlay_init).
tools/make_romfs_patches.py --xbox-prompts writes it from the player's own RomFS, with the 60 fps patches; nothing from
the game or a mod is added to the repository. It swaps A with B and X with Y in:
- the letter images Nt_KeyTexA/B/X/Y_00 of the action prompt (DoCommand_00), in the Common archive's __Combined.bntx
(same format, size and tiling, so the image data trades places);
- the action prompt's diamond of the other face buttons: one-letter text panes the game fills from
LayoutMsg/DoCommand_00.msbt in each language's Mals archive (both the messages and the layout's own text);
- the button glyphs U+E040 to U+E043 of scft/External.bfotf in each Font/Font*.bfarc, which every other prompt uses
(CFF outlines and metrics in the deobfuscated OpenType font, then its checksums).
The diamond's shape shows the button's position and stays. The setup writes both folders; the menu notes when the Xbox
prompt files are missing and when the prompts in use differ from the choice until a restart.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
1.2.1 has no camera aspect patch, so the installer no longer builds romfs-patches-<w>x<h> for it (the layout tool is checked against 1.4.3 and could fail the whole install), and romfs.c overlay_init skips that overlay on 1.2.1 through res_wants_wide_overlay. overlay_init and res_init now both call res_is_16x9 (1% tolerance), so 1366x768 counts as 16:9 and logs no missing overlay.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An output_res such as 3440x1440 now widens the camera's view to that shape on 1.4.3: patches-v143.txt
sets the aspect ratio at the camera and viewport sites that hand it to the perspective projection.
The HUD and menus come from a romfs-patches-<w>x<h> overlay that make_romfs_patches.py --aspect
builds through tools/ultrawide_layout.py, and romfs.c overlay_init picks that overlay when the game
resolution is not 16:9. The installer prepares it for 2560x1080, 3440x1440, 3840x1600 and 5120x2160.
When the window's shape differs from the game resolution's, the new aspect setting keeps the
picture's shape with black bars (fit, the default) or stretches it over the window.
A BC3 image takes whole 4x4 blocks or a part ending on the mip's edge; a WriteTexels region of a 5x5 or 6x6 footprint may end inside one. --probe-astc encodes every footprint to BC3 and holds the blocks to alpha, palette-place, flat-colour and PSNR bounds.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A GPU that cannot sample ASTC gets the textures decoded to RGBA8, four times the video memory of the ASTC blocks. With astc_bc = 1 (default 0, GPU decoder only) a second compute pass, bc3_encode.comp, encodes the decoded texels to BC3 and the image is BC3_UNORM or BC3_SRGB: 8 bits per texel instead of 32. The encoder is the usual bounding-box method with the box diagonal turned to the sign of the colour covariance, and the eight-value alpha mode.
A region of a WriteTexels that does not start on a 4x4 block corner is left as it was and logged. The setting is in totk.ini and the Advanced tab. The SPIR-V header was built with glslang 16.6 (build_shaders.py uses glslc) with the same source hash.
On a GTX 1650 (4 GiB) in the Fire Temple the GPU time per frame went from 10-23 ms to 3-12 ms and the frame rate held at 59-60 fps.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
SDL reports the back buttons of an Xbox Elite, DualSense Edge or Steam
Controller as four paddle buttons, but the pad table stopped at the
triggers, so they could not be mapped. They are now pad inputs lpaddle1,
rpaddle1, lpaddle2 and rpaddle2, with no default action, bindable from
totk.ini or the remapping rows.
#94 ended on a realloc of 1 MiB that failed with 64 GB of commit allowed, and the log could not tell whether the
process had grown towards the limit or the system ran short. rt_fatal now adds one line with the process's private
bytes and the system's commit in use and limit (plat_commit_info).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The render thread ran on its own 5 ms timer while the device consumed on its own clock, so the queue crept: on
DirectSound it grew 2.7 ms every 5 s, which fills the 340 ms ring in about ten minutes of play. The render thread now
renders the next frame when the device has taken the ring below max(prefill, two device periods), as the audio
hardware paces the Switch's DSP and as LASRecomp and SavageWindRecomp pace their renderers by the SDL stream. Without a
device, while the renderer is stopped and while the device takes nothing for 50 ms (at least four periods) it keeps the
timer, so a stalled device does not stall the game; the ring trim then drops what piled up, measured from the target.
60 s on the monsters save: WASAPI 15 ms queued, DirectSound 27-40 ms without growth, no underruns. With the SDL audio
thread suspended for 1.5 s the game held 60 fps and the queue was back at 15 ms after one trim.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The value is now a count (1 still means 16, capped at 4096), so blocks such as the deferred lighting shaders' camera
block, whose shadow cascade matrices sit at vec4 42 to 58, can be read from a frame dump.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The deferred lighting shaders soften the sun shadow with four compares around the receiver, offset by element 56 .xy of
their lighting block. The offset is in UV and the game sized it for its own 960x960 cascades, so with shadow_scale above
1 it spans that many more texels and the near cascade shows four separate copies of each edge. Translator patch 0011
multiplies those reads by the support buffer's new shadow_spread field in fragment shaders that sample a 2D array
shadow texture, leaving fp_c4 (the cascade matrices) alone. support_fill writes 1 / shadow_scale there, so the spacing
stays the game's in texels; at shadow_scale 1 the shaders behave as before. The support buffer grows to 1440 bytes and
build_xlate.py expects eleven patches.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The game's five camera sensitivities (Options, Very Low to Very High) turn the camera at the rates
in the StickSensitivity array of PlayerCamera.game__component__PlayerCameraParam.bgyml, 0.8 to 2.4.
stick_sens (0.1 to 3, read at start) multiplies that array: camera_speed.c decompresses
Pack/Actor/PlayerCamera.pack.zs with the RomFS's pack.zsdic, scales the floats in place and
compresses it again with the same dictionary, and romfs.c serves the result from
<profile>/camera/. Scaling the stick's deflection instead could not raise the top speed, since a
full tilt cannot go further.
The mouse camera reaches the game as a right-stick deflection, so input_win.c mouse_camera divides
it by the same factor and mouse_sens alone keeps setting the mouse's speed.
romfs.c heapdef_overlay is split into read_served and serve_generated so both generated files
share the reading, writing and overlay bookkeeping.
TOTK_TEST_PAD=<lx>,<ly>,<rx>,<ry> attaches an SDL virtual gamepad at start with its sticks at those raw values, so the
dead zone and the gamepad path can be tested in a run. On the field save: an 11% drift reached the game as 3.3% at the
default 8% zone (the old per-axis cutoff passed 9% on x and dropped y) and not at all at 15% (Link stood still); full
deflection reached it as full at 15%.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The game resolution row started at 1920x1080 although resolution.c accepts sizes from 1280x720, so weak GPUs and
integrated graphics had to edit totk.ini to draw the game smaller. A 1280x720 run on 1.4.3 renders the field with a
1067x600 scene and an intact interface.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
motion_restart only flagged the next sample to start a fresh frame, so a read in the frame of the Start returned the
previous session's 16 states as live, and for the next 60 ms 15 of the 16 were from before it, with the angle jumping
back to zero between two reads. Motion aim was dead when the bow was first drawn and lagged after that. A restart now
hides the older samples and zeroes the angle; sampling numbers keep counting.
--probe-services checks a restart in the middle of a turn: 16 old states in the Start frame before, none after.
Reported with the cause and a harness in #104.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>