adriadam10/mkw-recomp v0.1.12

on GitHub

A new start screen and settings by @ViniciusStanula; 20 and 15 fps at the race's real speed, Low latency and frame
generation that splits the screen's refreshes evenly; a second fix for NVIDIA's "instruction pointer invalid" crashes,
with one part by @Elenedeath; Classic Grand Prix, faster audio and Steam Deck fixes by @andreknieriem; the Windows
setup sized by the PC's real memory by @fraustiz; and fixes for the setup's shaders step (cause found by @Talaric0),
the minimap on AMD (diagnosed by @elenotrindade) and game version 1.8.0 (photo mode, Thai, LAN play, button icons).
Thanks to @ViniciusStanula, @andreknieriem, @Elenedeath, @fraustiz, @SavageCore, @Talaric0 and @elenotrindade.

Changes

• The start screen's main page, with the update at the top right (#357 by @ViniciusStanula): the profile that
  plays is a chip beside Play (it opens Profiles) instead of a menu row, the first controller SDL sees sits under
  Play — with how many more, and a click through to Controls, refreshed when a pad is connected or removed — and the
  update is offered as a card in the upper right with the release, the size in use, Update and What's new. The
  desktop shortcut is offered once after setup instead of taking a menu entry, and the game files' state at the
  lower left names the version the last setup installed.
• Settings in five tabs plus Advanced and Hacks (#358 by @ViniciusStanula): the eight tabs become Screen,
  Graphics, Sound, Controls and General, with Game and Menu folded into General (both languages, LAN play, the update
  check) — every row kept its setting, only where it lives changed. Graphics opens with the Quality preset and the
  four rows it sets hanging under it, then a Resolution group and a Performance group, and the FSR sharpness row is
  greyed out unless the scaling is AMD FSR 1. The row described on the right is the one the player used last, the
  mouse's or the cursor's, never both; the start screen's rows show a changed value with a dot instead of the
  default and "after restart" notes (nothing there needs a restart), and Graphics gains Auto-set, the best preset,
  resolution and FSR for the card and screen in use, only when the player asks for it — the defaults stay the
  console's — with an Undo.
• The Game files page follows what is installed (#359 by @ViniciusStanula): with nothing set up it is three
  numbered steps — the player's copy, how its RomFS is kept as two cards saying what each costs, then Set up with the
  folder it installs to — and once installed a card of what is there (game version, the port version that set it up,
  how the RomFS is kept, the copy and the folder) with the steps folded under Set up again. A failed setup gets a red
  card above the steps with the reason, Try again (the steps already done are kept) and the log folder.
• A Tools page on the start screen (#360 by @ViniciusStanula): Back up saves zips the profile into the game
  folder's backups/, Add to Steam puts the game in every Steam user's shortcuts.vdf (the same executable
  replaced, never doubled), Undo update brings back the version the last update replaced (the files it replaced are
  kept for it), and Report a problem copies the system's details to the clipboard and opens the bug form. The
  scripts run off the UI thread, one job at a time. The start screen's 64 new texts come in all 20 menu languages.
• The Windows setup sizes its jobs by the PC's memory (#416 by @fraustiz): it read the memory size from
  /proc/meminfo, which only Linux has, so every Windows PC was taken for one with 8 GB (a 24-thread PC with 31 GB
  logged machine: 24 threads, 8.0 GB, 9 compile and 6 generate jobs, win32). The size now comes from Windows
  (GlobalMemoryStatusEx). With it alone, 24 generate workers held 26.6 GB (1.1 GB each, 0.42 on Linux) and left
  0.13 GB free, so a Windows generate worker now counts 1.2 GB, and on every system the jobs also fit in the memory
  free when the setup starts: 17 compiles and 7 generate workers on that PC, the setup in 222 s instead of 300 s.
• NVIDIA "instruction pointer invalid": the particles' integer abs (#395 by @Elenedeath, #22): NVIDIA's shader
  compiler mis-compiles an integer abs of a small select result followed by an add or xor, and the GPU jumps to an
  invalid address; in 1.8.0 it hits the title screen's and the Souk's particle programs on their first draw. The
  translator now writes that abs as x * ((x >> 31) | 1), the one form the driver does not turn back into it
  (bit-exact for every 32-bit input). It goes with the second fix below.
• LAN play on game version 1.8.0 (#394 by @Elenedeath, #309): 1.8.0's patch table had no lan set, so its title
  menu kept Wireless Play (the console's radio, which a PC does not have) and picking it ended in 2203-0022, the game
  never asking for a local network. The set is ported, and with LAN play on the menu opens with LAN Play chosen, as
  on 1.0.0.
• Classic Grand Prix (#402 by @andreknieriem): a new switch in Settings > Hacks (MKW_CLASSIC_GP, off). With it
  on, every race of a Grand Prix cup is laps on one course: the intermissions, one lap on the road to the next
  course, become that course for its laps (Mushroom Cup: Mario Bros. Circuit, then Crown City for 3 laps, and so on).
  Knockout Tour is left as it is. Turned on during play, it applies to races started after its scan (a few seconds).
• Audio decodes each sample once (#331 by @andreknieriem): a DSP-ADPCM voice decoded each sample about three
  times; now about once, 24 % fewer cycles in the mixer. Two errors in the history carried from one buffer to the next
  (at a buffer change, and at a pitch above 1) are fixed along the way, checked against an exact decode.
• Dynamic state recorded only when it changes (#349 by @andreknieriem): the 12 to 18 dynamic-state commands of a
  draw were recorded again after every pipeline bind; now only the ones whose value changed, 10.6 % fewer Vulkan calls
  on a Steam Deck's Grand Prix (no measurable change in frame time there, which the replay thread bounds).
• Out of video memory for a pool chunk, memory is freed first (#326 by @andreknieriem): on the Steam Deck (1 GB
  carve-out) a pool chunk that found no video memory failed outright; it now frees idle textures and waits for the
  GPU, as an image already did, and tries again (a wait of 1.9 ms at the median and 17.5 ms at most on the Deck).
• The GPU hang report comes before amdgpu resets the GPU (#327 by @andreknieriem): on the Steam Deck the kernel
  resets a GPU job after 5 s (SteamOS's amdgpu.lockup_timeout), as soon as the report of #297 was due, so three hangs
  ended with no line in mkw.log. On amdgpu the report now comes after two fifths of that timeout (2 s on the Deck);
  an unset or unavailable kernel timeout uses 1 s (the documented default has been as short as 2 s), elsewhere it stays
  at 5 s, and MKW_GPU_HANG_MS sets it. Each stuck episode reports once, not only the first, and one the GPU finishes
  after all says so ("it was slow, not hung").
• "Straight to" a Grand Prix (#348 by @andreknieriem): Settings > Hacks > Straight to gains Grand Prix (24 karts,
  through the menus by phase, skipping the course's flyover), beside time trial and free roam.
• Makefile shortcuts (#362 by @SavageCore): for building from source, make build, check, dist, appimage,
  windows and clean run the commands tools/build/README.md documents (make help lists them).
• The Steam Deck renders the race at 720 lines (#317 by @andreknieriem): the Deck's picture is 1280x720 on its
  screen, but the 3D scene still rendered at the console's docked 2560x1440. @andreknieriem measured it on a Deck: those
  passes took 16 of the GPU's 42 ms a frame. The Deck's profile now sets Resolution to 720, and races went from about
  36 to 24 ms a frame. The menu or mkw.ini still pick another one (a Deck on a TV, for instance).
• The start screen says how the last game ended: when it stopped with a known problem, it names the issue that
  follows it (no need to report it again) or the release that fixes it; an unknown error gets a button to the bug
  form, asking for logs/mkw.log.
• 20 and 15 fps, at the race's real speed, and Low latency: Settings > Graphics > Frame rate gains 20 and 15, and
  Automatic now drops to them (it stopped at 30) when frames take too long, so a slow graphics card no longer plays
  the race in slow motion: at 15 the time trial's clock runs at 0.998x real time and the karts drive as at 60. With
  Frame generation at 2x, 15 fps shows as 30. New Low latency (on): the game used to prepare up to two frames
  ahead of the screen; now it starts each one only as early as its work needs, measured frame by frame. On a Radeon
  660M at 15 fps a controller read reached its frame's display time in about 115 ms instead of 200. 20 and 15 apply
  only in races: menus and loading screens go at 30 at most, as on the console, and a cup's opening movie plays at
  60 from its first frame, so it can always be skipped (opened at 30 or 15 fps it was not skipped in 7 of 9 runs).
  Frame generation pauses while a movie plays: the movie keeps its own rate instead of repeated frames. At 15 fps
  with 2x the made frame and the game's now go out evenly, 33 ms apart (they went 25 and 42 ms apart), and the full
  performance overlay shows the game's frames and the screen's ("20/40 fps") instead of only the game's.
  Frame generation's Automatic now makes only as many frames as split the screen's refreshes evenly: on 60 Hz,
  2x at 30 fps, 3x at 20 and 4x at 15, all 60, where 2x at 20 (40 fps) held frames 1, 2, 1, 2 refreshes and
  juddered. A 75 Hz screen, where no game rate fits, shows the game's own frames. New Variable refresh screen
  (off): with FreeSync or G-Sync on, Automatic makes as many as fit in time instead. A fixed 2x, 3x or 4x that
  does not fit the screen shows a note under it.
• Scaling to the window is bilinear by default: AMD FSR 1 is on only when picked in Settings > Graphics or by
  Auto-set.
• Less copying of game memory on dedicated graphics cards: with the memory in mirror (what Automatic picks on a
  dedicated card), a page only the GPU wrote was uploaded again from the game's copy, and the shadow took back in what
  the GPU had just written out; neither does now. On a GTX 1050 Ti, 2.1 ms less per frame on average.
• Updating keeps the setup's choices: the setup that runs by itself after an update copied the whole RomFS
  (23 GB) even when the player had chosen to link it, and an update that found a leftover of the one before in
  update-previous/ stopped with an error instead of retrying.
• Issues: a report of a known error is closed at once as its duplicate (a reply from its author reopens it), and
  a report on an older release, quiet for 3 days, is asked to try the latest and closed after 7 days without a reply.
• Motion controls for every player (discussion 302): the gyroscope and accelerometer were read from player 1's
  controller only, so in local multiplayer players 2 to 8 had no motion; each player's controller now feeds its own.
• Thai no longer crashes on game version 1.8.0 (discussion 193): with the game in Thai, loading closed the game
  (no function at guest pc main+0x1bb0b34). Its text code leaves an error with the C library's longjmp, which
  goes back to where setjmp saved its place: in translated code that place is the middle of a function. A function
  that calls setjmp now keeps a host return point that the longjmp comes back to.
• Xbox, PlayStation and keyboard button icons on game version 1.8.0 (#19): 1.8.0's button sheet has two more
  icons, so the runtime did not recognise it and the prompts kept Nintendo's buttons. The other icons sit where
  1.0.0 has them, and both sheets are redrawn now.
• LAN play picks the home network's address (discussion 148): with several network adapters, the game could take
  one that leads nowhere, such as a VirtualBox host-only adapter (192.168.56.1) or one without a DHCP server
  (169.254.x.x), and two PCs on the same network never saw each other. The address on the network with the default
  route now comes first, then other local networks, then 169.254.x.x, and VPN /32 addresses last. The log lists the
  candidates (nifm: LAN candidate); lan_ip in mkw.ini still picks one by hand.
• "Choose game folder" in every menu language: its four texts came in English only in 0.1.11.
• Photo mode no longer crashes on game version 1.8.0 (#305): pressing Minus in free roam at 1440 lines closed the
  game (main+0xa66e0): 1.8.0's photo mode, with its stickers, needs render targets the game's buffer pool has no room
  for. The pool gets a second part of the runtime's own memory for them, as it did at 2160 lines (#255).
• No more silent exits in the C runtime on Windows (#310): a date past the year 3000 given to the game's time
  service ended the process with no message (code 0xC0000409 in ucrtbase.dll); it is now refused as on Linux, and any
  other call the Windows C runtime rejects writes a crash report with the game's frames to logs/mkw.log.
• The setup refuses exefs and romfs from different game versions (#315): 1.0.0's program with 1.8.0's game
  files passed the checks and the game crashed while warming up; the setup now says which version each one is.
• Unlock everything opens Mirror (discussion 292): Settings > Hacks > Unlock everything now also offers Mirror in
  the class select, as the game does after its "Mirror Mode has been unlocked!" notice.
• NVIDIA "instruction pointer invalid" crashes, second fix (#22, #231): 0.1.11's change did not cover Moo Moo
  Meadows. Every logged case was a pass stuck for 1-1.5 s whose draw has a fragment shader that discards pixels
  midway (8 of 8, against 30 % of the game's fragment shaders). A discarded pixel now stays as a helper to the end of
  the shader (Vulkan's demote, as Switch emulators and DXVK do) instead of ending in the middle of it. Verified on the
  captures (same images) and not on an NVIDIA card.
• A setup that failed once in the shaders step no longer fails every time after (#325, cause and fix by
  @Talaric0): when NVIDIA's disassembler gave a garbled capture, the rejected spec was already in sm86/, and every
  later setup reused it and stopped in translate.py (KeyError: 'BITS_…' or classes carry no Opcode constant). A
  capture is now parsed aside and kept only when accepted, and a spec left broken by an earlier release is built
  again. Before 0.1.12: delete the game folder's sm86 folder and run Set up again.
• The minimap without garbage on AMD (#365, diagnosed by @elenotrindade): AMD lays out some depth buffers larger
  than the Switch 2 does (the minimap's 676x676 one takes 0x2e0170 bytes for 0x204000), and bound at the game's address
  its extra bytes overwrote the minimap's colour target right after it. A depth or stencil buffer that does not fit its
  space now gets its own memory (MKW_POOL_SPILL=1 restores the old binding). Same picture on the race and title
  captures on a Radeon 660M, where 7 depth buffers did not fit.
• A processor without AVX2 gets a message instead of nothing (#262): the game is still built for AVX2 (a build
  without it is planned in discussion 82), and on an AMD FX or an Intel Core older than the 4th generation mkw.exe
  closed before writing any log. It now says so in a box before anything else runs.

Tested before publishing

From these packages, before publishing:

• The whole setup from the Linux AppImage (Radeon 660M, RADV, game 1.0.0) in 16 minutes (generate 57 s, compile 8 min,
  shaders 6 min), its warm-up to the title without a fault.
• The whole setup from the Windows zip (Windows 11, GTX 1050 Ti, 6 threads) in 23 minutes, its warm-up to the end
  without a fault; its log now names the PC's real memory (machine: 6 threads, 15.9 GB, 10.9 GB free, where 0.1.11
  said 8.0 GB).
• Before the tag, on main: the race and title captures replayed on the Radeon 660M with the minimap's depth change
  (the same images as without it: 0.053 and 0.0015 against the game), the shaders step's spec built and rejected on
  purpose (a rejected capture leaves no spec behind), and the setup's tests (40 passed).

Install

• Linux x86-64, glibc 2.31 or newer (Debian 11, Ubuntu 20.04, SteamOS and later): chmod +x the AppImage and run it.
  The game it builds lives in ~/.local/share/mkw-recomp. On 64-bit ARM (ARMv8.2: Snapdragon 8 Gen 1 and newer), the
  -aarch64.AppImage.
• Windows 10/11 x64: unpack the zip into a folder and run mkw.exe.
• Either: a Vulkan 1.3 GPU and about 30 GB free. Open Game files, point it at your own dump of Mario Kart World
  1.0.0 or 1.8.0 (its exefs/ and romfs/, or one .nup or .nx file) and press Set up. Nothing of the game ships here: the setup builds it on your
  machine and downloads NVIDIA's nvdisasm and ptxas (pinned by sha256; 84 MB on Linux, 126 MB on Windows) to
  describe the shader encoding.

Updating

From 0.1.5 on, the start screen offers a newer release (Update to vX) and installs it: it downloads it, checks it,
replaces the program and sets up again only what changed. From an older version, once by hand: Linux, run the new
AppImage; Windows, unpack the new zip over the same folder; then press Set up again (the steps already done are
skipped).

Known issues

• NVIDIA "instruction pointer invalid" (#22): the second fix is checked on the captures (same images), not on an
  NVIDIA card yet; skip_programs = 1c7a765b in mkw.ini still avoids the Shell Cup's crash if it comes back.
• Some AMD cards on Linux still hang later on, in a race or a menu, with the compute path (#50).
• Black ground on Intel Arc and AMD Polaris (#319, #314): the terrain's texture tiles past the 768 MiB budget stay
  empty there.
• A processor without AVX2 cannot run the game yet (discussion 82); it now says so instead of closing silently.

Reporting a problem

Open an issue with the bug form and attach the logs: logs/mkw.log (the game and the start screen) and
logs/setup.log (the setup's tools), in ~/.local/share/mkw-recomp on Linux and next to mkw.exe on Windows.