github adriadam10/mkw-recomp v0.1.14

one hour ago

Races at 60 fps on AMD cards under Linux (42.6 to 16.9 ms a frame on an RX 6700 XT), trees and bushes that no longer
drop out at random, black textures on Intel Iris Xe and a crash on Vega, the setup that now says when the game files
are gone, finds your copies by itself and checks the free space first (also from Steam's Gaming Mode), the game in
Turkish, every controller's battery, an update that sets itself up while you play, and a first start that builds every
shader library in the setup. New: Steam Big Picture and Gaming Mode start by @AngeloAscheri, Cemuhook motion by
@jony25, the HUD at the screen edges by @simonwjackson, a menu shortcut setting by @SavageCore, a 120 fps frame
generation target by @Hampter588, and the README in German by @andreknieriem. Thanks to @AngeloAscheri,
@simonwjackson, @SavageCore, @louisdibiasi, @andreknieriem, @jony25, @Gael50, @Gag21, @Hampter588, @VictorMerk and
@MarcTheFennec.

Changes

  • The start screen says when the game files are gone, also in Gaming Mode (#493 by @AngeloAscheri, #486): a Steam
    Deck player linked the game's .nup in desktop mode, the package was no longer there, and Play just closed the game
    with nothing said. The start screen now checks once a second that the RomFS and exefs/ are there; when they are
    not, Play stays off, the state line reads "Game files not found" and the Game files page opens, saying where the
    setup left them and what brings them back (connect the card or drive, or set up again). A card connected while it
    is open is seen within a second. A missing RomFS is also named in the log. In Gaming Mode, where the start screen is
    skipped, it now opens when the files are gone (by us).
  • Each controller's battery (discussion #516): the Controls tab lists each player's controller and its battery,
    the start screen adds the first one's to its controller line (every one's when hovered), and a notice says once
    when a battery drops to 15 %. A Switch 2 controller over Bluetooth gives its voltage in three bands, as no curve to
    a percentage is known.
  • The game in Turkish (discussion #593): 1 162 of 1.0.0's 1 328 game texts and 1 309 of 1.8.0's 1 796 are
    translated (song titles and names stay in English), chosen in Settings > Game > Language. The game's fonts lack
    ğ Ğ ş Ş İ, so the setup composes them into each font's own overlay from its breve, dot and cedilla; the composed
    fonts are kept between setups (24 s to 1 s).
  • Trees and bushes stopped dropping out (#524): four compute programs reset an append counter from the dispatch's
    first invocation while every workgroup adds to it, and nothing ordered workgroup 0 first, so the scenery lists
    lost entries at random from frame to frame. The first invocation's reset now runs on its own before the rest
    (MKW_FIRST_STORE, on by default, 0 turns it off); 11 561 of 11 561 programs stay valid and only those four
    change. Found by nxnative, which measured 50 % of a rebuild lost on Apple GPUs.
  • The replay thread no longer takes a core from a hybrid CPU with few P cores (#497): on 4 P and 4 E cores
    (12 processors, RTX 3050 laptop) 0.1.13's fast-core pin left the game and its three recording chunks on 3 P cores:
    30 fps in 0.1.12 became 18. A fast core is now taken only when 8 fast processors are left besides it; otherwise
    the last physical core, as in 0.1.12. Not measured on a hybrid CPU here; on the i7-14700F nothing changes.
  • The GPU check's texts in every language, and Windows' own advice (#621): the start screen says when no GPU here
    can run the game (see #621 by @AngeloAscheri below); its eight texts are now in all 20 languages, and Windows
    players are told to install the latest driver from the card's maker (NVIDIA 510 or newer) instead of being sent to
    Mesa, Debian and Ubuntu.
  • Intel Iris Xe: black textures on 0.1.13 (#540): Intel's Windows driver gives ASTC 8x8 sparse granularity as
    64x64 blocks, as it does for BC, and only BC was converted: an Iris Xe left 524 288 tiles without memory. Any ASTC
    block size is converted now.
  • Setup: a copy made in copy mode whose source was deleted sets up again (#501): the setup repeats the last one
    from its recorded source, and a deleted source stopped it with "moved". The game folder, which already holds
    exefs/main and romfs/, now stands in.
  • Setup on Windows with a non-ASCII path (#527): zig prints the player's path in UTF-8 ("Área de Trabalho") and
    it was read as cp1252, which failed the compile step. zig's and mkw --unpack's output is read as UTF-8 and
    every tool's output replaces a byte it cannot decode instead of failing.
  • Setup: a dead worker fails instead of hanging (#506): gen.py waited forever for a chunk lost with a worker
    (0 % CPU, nothing but stack dumps in setup.log). It now exits with an error, and a compiler crash (ICE) is
    retried once for that chunk, with the head of its error in setup.log.
  • The pipeline cache autosave waits for the race to end (#533, #596): each save copied the whole driver cache
    under the driver's lock, which the replay's fast links wait on: 0.3-1.1 s hitches in a race on a Steam Deck. New
    pipeline keys still go every 2 s and the cache follows at the next menu or load; each save logs its size and times.
    On a GTX 1050 Ti (45 s of race, a 498 MB cache) the 4 saves all came in menus and the load, none in the race.
  • HUD position, translated (#604, #605): the setting's texts are in every language (by us; see the entries by
    @simonwjackson below).
  • The build is attested (#541): the release now attests the three packages and mkw.exe inside the zip, so a
    player can check with gh attestation verify that a file came from this repository's workflow (one antivirus
    engine of 70 flags mkw.exe by its behaviour during the setup).
  • 60 fps races on AMD GPUs under Linux (#521 by @AngeloAscheri, #496): amdgpu chose a clock by how busy the GPU
    was, and the race frame keeps it half idle: 500 MHz of 2855 on an RX 6700 XT. The kernel's stable pstate is now set
    on a context of our own, with no root: the race median went from 42.6 to 16.9 ms, the GPU's busy time per frame
    from 19 to 5.3 ms. MKW_GPU_CLOCKS: auto (default: the peak on a discrete AMD GPU; off on an integrated one,
    such as the Steam Deck or the 660M, until measured there), peak, standard, off. Nothing changes on Windows
    or with other vendors. We set it only while the card is at amdgpu's automatic level, so LACT's or CoreCtrl's own
    setting is never lost (by us); --probe-gpu-clocks shows on any GPU whether the driver lowers the clock in idle
    gaps.
  • The frame's threads on a hybrid CPU's fast cores (#490 by @AngeloAscheri, #480): the frame's threads stay on
    the fastest cores while background work keeps the efficiency cores. On an i7-14700F, in-run A/B: -6.3 ms a frame
    (means), B faster in all four runs; the race's heavy first seconds go from 27-30 to about 20 ms. MKW_FAST_CORES
    is on by default on Linux with a hybrid CPU that has 8 fast processors to spare, off elsewhere (Windows stays off
    until measured there); 0 turns it off (follow-up #497 by us above).
  • vk submit and vk present run with the frame's threads (#584 by @AngeloAscheri): they started on the replay
    thread's own processor. A part's wait for its vkQueueSubmit2 falls from 34-39 to 11-13 us. MKW_SUBMIT_APART
    is on by default (0 keeps the old placement). The frame did not move on the RX 6700 XT at its 60 fps cap; the
    gain is expected on a CPU-bound machine, not measured.
  • Course loads: the ASTC alpha scan and the texture deswizzle on worker threads (#581 by @AngeloAscheri): in a
    Grand Prix load the alpha scan on the replay thread goes from 215 to 138 ms and the deswizzle of 1.1 GiB of larger
    textures from 139 to 77 ms, with the same images; the race itself is unchanged. MKW_TEX_PAR=0 starts no workers.
  • The shadow write-back compares 64 bytes at a time (#612 by @AngeloAscheri): the byte merge of what the GPU
    changed took 2.6 ms of every frame in a 2-player race on an i7-14700F. With AVX2 masks and whole-run copies it is
    1.5-1.7 ms and the frame gains 0.52 ms (the 60 fps cap hides the rest). MKW_MERGE_SIMD=0 keeps the old loop.
  • Setup finds your copies in the usual places (#514 by @AngeloAscheri): Steam's Gaming Mode has no file dialogs,
    so a player with only a controller could not choose the copy. The start screen now lists the .nup, .nx and
    loose copies it finds in Downloads, Desktop, Documents, Games, EmuDeck's Emulation, Steam's Proton prefixes and
    /run/media, /media and /mnt (on Windows, the profile's folders and each drive's root), with version and path,
    one button each; the first is chosen, so Set up is one press away (the first of a known version, by us).
  • Add to Steam for Big Picture and Gaming Mode (#518 by @AngeloAscheri): the shortcut gets library art drawn in
    the start screen's look (600x900 capsule, 920x430 wide capsule, 1920x620 hero, logo, icon) instead of a grey tile,
    the button is also on the card offered after the setup, and adding again from another program path replaces the
    shortcut instead of doubling it.
  • Big Picture and Gaming Mode: the game starts straight in (#525 by @AngeloAscheri), fullscreen and with the
    controller working: the start screen was asking for one more Play, the game opened in a window and the controller
    did nothing. MKW_START_SCREEN (Settings > General: Automatic, Always, Skip; default Automatic) skips the start
    screen when Steam started the game, unless it has something to do or say (files not set up, a setup to repeat, a
    code mod to build, a fatal error, or View/Select held in the first 400 ms). On a Wayland desktop the game uses
    X11 so Steam can hand it the controller; we keep native Wayland when a display is in HDR (MKW_STEAM_X11, by us).
  • Free space is checked before the RomFS copy (#519 by @AngeloAscheri): 21.9 GiB in 206 202 files for 1.8.0 plus
    about 3 GiB for the first build. A drive that filled half-way (a 64 GB Deck, a full SD card) failed in English
    after up to 20 GiB. The setup now says in the player's language what is needed and what is free, and suggests
    "Use it where it is".
  • An Xbox controller's sticks get Microsoft's own dead zones (#523 by @AngeloAscheri): an Xbox Elite 2 through
    xone came to rest at 11-21 % and the course select's cursor kept moving. Xbox 360 and One/Series pads now get
    24 % and 26.5 % by default (MKW_STICK_DEADZONE still wins; every other pad and Steam's virtual gamepad keep
    8 %). The slider shows the open pad's own default while unset (by us).
  • Audio resamples to the rate the device really drains (#553 by @AngeloAscheri, #544): a device pulling 44.1 kHz
    while the game wrote 48 000 a second dropped about 8 % of the blocks (102 overruns in 10 s). The rate is measured
    every second and the ratio corrected up to 1 %: with a skewed device, 19-20 overruns, all before the first
    estimate at 5 s, then none. The prefill now covers the device period and SDL's whole request, and underruns are
    logged (by us).
  • The first recording chunk no longer gets a depth target in the wrong layout (#556 by @AngeloAscheri): one
    validation error in a Grand Prix run, three times, on the race's depth target.
  • Setup shaders and prewarm (by us): the setup now builds the shader libraries of every vertex and fragment pair
    in the copy and every compute pipeline (753 prewarmed on a Windows PC, 0 left to a dispatch, 18 before; #500), so
    a new place no longer skips draws while its libraries compile. The alpha test reads its reference from the support
    buffer, so its fade steps are no longer pipelines of their own. The prewarm reads one shaders.pack: 24.0 to
    2.0 s on the Windows PC with a cold file cache, and the guest starts 27.2 to 11.4 s after the process. Programs
    only the prewarm read are released after it (745 to 417 MiB resident on a 660M). A compute program loaded by key
    keeps its resource table (on a GTX 1050 Ti one dispatch looped 65 536 times: 403-466 ms a frame in the character
    select, 15-17 ms now).
  • An update is set up in the background (by us, discussion #429): the download no longer replaces the game
    folder at once. prepare.py --next sets it up in setup-next/, at low priority and with half the memory, Play
    stays the first button and works all along, and the card offers Restart with it. Its warm step runs there too
    (it was 106 s at the first start on the Windows PC). The update's shaders step runs beside generate and compile,
    and the swap waits 10 s instead of 18; 0.1.13 to this version on the 6-thread Windows PC: 390 to 329 s, Linux
    284 to 255 s. The compiler's key leaves out its folder and the host, so an update on Linux or a kernel update no
    longer compiles all ~360 chunks again (a 0.1.13 object is reused). MKW_UPDATE_API asks another server for the
    latest release, to try an update before it is released.
  • A faster first setup (by us): the game's chunks compile on LLVM's own clang with zig's -cc1 line (a 3.5 MB
    chunk 6.9 to 4.75 s on the 6-thread Windows PC), the translation of the 11 561 shader programs goes from 366 to
    140 s on 12 threads (glslc -O0; the frame's GPU time stayed 19.5 to 19.2 ms on a GTX 1050 Ti), pack.py reads
    on 16 threads (29.3 to 17.1 s), and setup.log gives each tool's time and each step's CPU, also on Windows.
    MKW_SETUP_CPU=v2 builds the game code for x86-64-v2 for CPUs without AVX2 (#262): FMA3-less CPUs get a correctly
    rounded fma; not yet measured against v3.
  • Racing wheels without a gamepad mapping work as a controller (#488, by us): SDL knows the MOZA R3/R5 as wheels
    but guesses no mapping, so the game never saw them. The wheel axis goes to the left stick, the first buttons to
    A/B/X/Y, shoulders, Minus, Plus and the first hat to the D-pad; pedals are unmapped, and a gamecontrollerdb.txt
    in the game folder is read for them.
  • Knockout Tour's checkpoint rainbow covers the whole width at 21:9 (#425, by us), on 1.0.0 and 1.8.0.
  • Mods can ship native plugins (by us): a mod's plugin/*.so|dll is loaded before the game code with a
    versioned API (overrides by symbol or address, on_game_loaded, notices and an overlay); the Mods page warns that
    the mod runs native code. A code request for a number gets an on-screen number pad (D-pad or stick and A, B deletes,
    Plus confirms; also the mouse and keys).
  • The start screen's GitHub and Discord marks (by us): two icons, lit on hover, replace the GitHub text link.
  • GetThreadCoreMask never reports a core outside the thread's mask (#343 by @Gael50): the game's network
    thread copies its (core, mask) pair and SetThreadCoreMask refuses a core outside the mask; a thread given
    SetThreadCoreMask(-1, 0x4) read back (0, 0x4). It now reads back the ideal core the game set, -1 there, as
    the console does (the guest's own ideal core rather than our scheduler's hint, by us).
  • LAN play over a VPN (#345 by @Gael50): MKW_LAN_PEERS in mkw.ini lists the other PCs for a network with no
    broadcast, such as Tailscale; tested by NoxiousGit with 3 races over Tailscale on two PCs. The 0.1.13 notes already
    listed it, but its code merged after that tag.
  • Character select on a Radeon RX 570 on Windows (#560 by @VictorMerk): stippled transparent panels
    and broken rounded edges around the character-selection cards, caused by dynamic alpha-to-coverage on Ellesmere
    (0x67df) with AMD's Windows driver. Alpha-to-coverage alone is now baked into the pipeline on that GPU;
    MKW_VK_DYN_A2C=0 picks static and =1 restores dynamic for comparing.
  • Radeon Vega 7 and older GFX9 on RADV: crash at frame 150 (#385 by @MarcTheFennec): a texture resolving as slot
    337 killed the process inside vkCreateImage (an assert in Mesa's addrlib for a block-compressed 3D texture with
    mips). Those images are now made without COLOR_ATTACHMENT, and the usage query's extent, mip, layer and sample
    limits are checked. Confirmed by the reporter on a Vega 7.
  • Window size in pixels on Windows with display scaling (#575 by @Gag21): at
    125 % scaling a window_size = 2560x1440 opened at 2048x1152; it now divides by the pixel density, not the display
    scale.
  • Straight to works on a player's install (#572 by @Gag21): Settings > Hacks > Straight to (MKW_QUICK_RACE)
    waits for the game's phases, named by a table only a source checkout made. The setup now writes it from
    exefs/main (285 phases on 1.8.0, about 1 s) (an optional table, by us).
  • Setup on Windows defaults to llvm-mingw's clang when there is no cc (#529 by @Gag21): the shader spec step
    failed with a source checkout and no compiler on PATH.
  • Source builds detect the game's version (#585 by @SavageCore, #619 by @SavageCore):
    make build and make generated pick build/generated and the patch list from the exefs version (1.0.0 or
    1.8.0), and build.py fails early when the layout does not match; --out and --patches given by hand no longer
    trigger the detection (duplicate imports fixed by us).
  • Menu shortcut setting with a Guide button option (#579 by @SavageCore, #563, #613 by @SavageCore): a setting
    for the button that opens the menu, with a dim note under the description pane; the Xbox Game Bar note is in its
    help text.
  • The overlay opens on Continue (#587 by @SavageCore, #577, #611 by @SavageCore): opening the menu focused the
    first setting, so an accidental D-pad press retuned it. Continue is focused instead, and its description pane is
    cleared while Continue or Quit has the cursor.
  • Xbox battery on Linux with the xone driver (#583 by @SavageCore, #623 by @SavageCore): SDL reports it as
    unknown, so the Controls tab and the start screen showed nothing. The sysfs level band is read instead, shown as
    High, Medium or Low (never a percent), matched by Xbox order (with a non-Xbox pad as player 1 it looked up another
    pad's); the help text names both sources in all 20 languages.
  • Cemuhook / DSU motion for gamepads without a gyro (#535 by @jony25, discussion #435): a pad in XInput mode can
    have its gyro and accelerometer read from a local DSU server (DS4Windows, BetterJoy, PadForge) so motion steering
    and tricks work. Settings > Controls (MKW_DSU, off by default); MKW_DSU_SERVER (127.0.0.1:26760) and
    MKW_DSU_SLOT are in mkw.ini. Its own thread, with CRC32-checked packets, never the game's.
  • A 120 fps output target for frame generation (#491 by @Hampter588): validated without game files or a GPU run;
    not measured in a race.
  • Audio voice updates stay tied to their input slots (#495 by @Hampter588): an unused voice record can be all
    zero, including its ID, so every one of them reset playing voice 0 instead of its own slot; filter updates also
    hit the ID, not the slot. May contribute to audio dropouts; not claimed to fix the reported movie-audio cuts.
  • 1.8.0's picture takes the window's shape (#601 by @simonwjackson): 1.8.0 sized its
    render targets from the display object, so a window that is not 16:9 got a 2560x1440 picture stretched on one axis:
    2x horizontally at 32:9, 1.28x vertically at 2448x1848. The game now renders the rows or columns the window shows
    (2560x1934 at 2448x1848, 3414x1440 at 21:9, 5120x1440 at 32:9), which costs +34 %, +33 % and +100 % pixels at
    original resolution; MKW_DISPLAY_PATCH=0 keeps the stretched picture.
  • Character and kart lists on wider outputs (#602 by @simonwjackson): on a window wider than 16:9 1.8.0's
    character and kart select showed a hard vertical seam at the list's left edge. The two capture masks now sample the
    backdrop under themselves.
  • Race and pause prompts keep their shape (#603 by @simonwjackson): on 1.8.0 with a window that is not 16:9, the
    "Photo Mode" prompt box was 15.59 wide against its height at 5120x1440 and 5.80 (squashed) at 2448x1848; they now
    sit in the same corners with the 16:9 shape.
  • HUD position (#604 by @simonwjackson): Settings > Screen > HUD position (MKW_HUD_POSITION), off by default
    (16:9 area): with Screen edges, the 1-player race HUD of game 1.8.0 goes to the screen's edges and corners (item
    slot top-left, coins and lap bottom-left, place bottom-right, the Time Trial timer top-right, the minimap at the
    right edge).
  • HUD position on every 1.8.0 screen (#605 by @simonwjackson): with the setting still off by default, a reviewed
    table of 1.8.0's layouts puts the HUD pieces and the menus' chrome of every 1-player screen at their edge (the
    loading bands, the cup and cc band, the face, character and kart select footers, the Knockout Tour, battle, free
    roam, mission, spectate, replay and photo mode pieces) while lists, grids, 3D previews and dialogs stay in the
    middle.
  • The start screen's GPU check (#621 by @AngeloAscheri): with no GPU able to run the game (Vulkan 1.3, three
    extensions, graphics and compute queues), the window closed at its first frame after a setup of many minutes: the
    case of Debian 11 and Ubuntu 20.04 with their own Mesa (Vulkan 1.2). The start screen now reports each GPU's
    Vulkan version, driver and what it lacks, with a card on the Game files page, and warns about a processor-only
    renderer (llvmpipe) before the setup and Play.
  • Linux x86_64 AppImage draws through Wayland again (#620 by @AngeloAscheri): the AppImage's SDL had no Wayland
    backend (the Wayland scanner aborted in the build), so every Linux release up to 0.1.13 drew through Xwayland, and
    a session without it could not open the start screen at all ("the game files are not set up").
  • Update on AArch64 Linux (#492 by @louisdibiasi): update.py fetch tried to open the AArch64 AppImage as a
    Windows zip and exited with "File is not a zip file"; both Linux targets now take the AppImage path.
  • Linked RomFS refreshed after setup (#499 by @louisdibiasi): after setting up with a .nup kept outside the
    installation, installing a mod in the same session failed with "is not the game's RomFS".
  • Backups made in the same second (#565 by @louisdibiasi): two backups of one player within a second overwrote
    each other.
  • Spaces in imported mod names (#566 by @louisdibiasi): the Mods page stopped the name at the first space.
  • Mod import in UTF-8 (#567 by @louisdibiasi): on Windows with cp1252 output the importer could emit bytes the
    launcher cannot decode.
  • Truncated shortcuts.vdf (#568 by @louisdibiasi) and missing Steam profiles (#616 by @louisdibiasi): Add to
    Steam reports an error instead of exiting silently or finding no userdata.
  • Switching versions from a loose RomFS (#569 by @louisdibiasi): a shared file of equal size but other contents
    was kept from the old version; equal-sized files are now compared.
  • A failed staged promotion is rolled back (#570 by @louisdibiasi): the live installation is restored (our
    first-version backup recovery was dropped, by us).
  • Bounded copy discovery (#571 by @louisdibiasi): one directory could give more than the 16-result cap or keep
    scanning past the deadline.
  • Lazy pool backing failures are counted (#574 by @louisdibiasi, #438, #562): the memory census now counts them
    (923 on the AMD log of #438); the rendering cause of #562 is still open.
  • A failed source scan no longer deletes files (#614 by @louisdibiasi): files missing from a skipped directory
    were treated as stale and removed from the installation.
  • Interrupted package downloads (#615 by @louisdibiasi): a connection reset reports the failed fetch and removes
    the partial file.
  • Discovered dump paths in UTF-8 (#617 by @louisdibiasi): an accented path reached the launcher with invalid
    bytes and a Chinese name stopped discovery on Windows.
  • Damaged GameBanana cache entries (#618 by @louisdibiasi): a truncated entry made the Mods page fail every time
    until the cache expired.
  • README in German (#484 by @andreknieriem, "Deutsch" in every language bar): the button and menu names are the
    German start screen's own.
  • Pipeline cache autosave interval (#596 by @andreknieriem): MKW_PIPE_AUTOSAVE_S (default 20; 0 never while
    running). On a Steam Deck LCD, 9 of 17 default runs had a hitch over 300 ms after frame 1500 (15 hitches) against 0
    in 3 runs with 0; three runs are not a proof (our fix #533 above makes the wait automatic, by us).
  • 1.8.0's run checks (#487 by @AngeloAscheri): the NVN2 procs new in 1.8.0's SDK are never called, its 11 604
    shader stages translate, and the c24 push constants are passed. Documentation only.
  • Race measurements on a hybrid CPU (#489 by @AngeloAscheri): the i7-14700F and RX 6700 XT numbers, and two test
    defaults that mislead on a discrete card. Documentation only.
  • Dependabot grouped updates (#629 by @app/dependabot): numpy 2.5.5 and cogapp.
  • Smaller: MKW_STEAM_X11 is a known key of mkw.ini, and the X11 switch from Steam is skipped with HDR on;
    the audio underruns of #544 are logged without MKW_TRACE; the Xbox pad's own stick default shows on the slider
    (#523); MKW_REPLAY_PREWARM makes MKW_PIPE_SYNC default to 0, so the replay's prewarm switch is
    no longer silently skipped; the 1.8.0 setup follow-ups of the merged PRs (#572 optional table, #579 note length,
    #596 strict value, #566 names of 64, #560 MKW_VK_DYN_A2C known in mkw.ini).

Tested before publishing

From the packages Actions built for this tag, before publishing:

  • The whole setup from the Linux AppImage (12 threads, game 1.0.0) in about 12 minutes; its warm-up built all 13 630
    shader libraries and compute pipelines (761 compute) with none failed.
  • The whole setup from the Windows zip (the 6-thread Windows PC, game 1.0.0) in about 15 minutes; its warm-up built
    the same 13 630 with none failed.
  • #596's autosave on a GTX 1050 Ti (45 s of race): its 4 saves came in the menus and the load, none in the race.

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.
  • Every package and the mkw.exe in the zip carry a build attestation: gh attestation verify mkw.exe --repo adriadam10/mkw-recomp shows it was built by this repository's GitHub Actions. An antivirus that flags it by its
    behaviour (it compiles and loads the game's code) is a false positive: report it to that antivirus.

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

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.

Don't miss a new mkw-recomp release

NewReleases is sending notifications on new releases.