github Lucas-blip-png/dkb-pc-recomp v0.4.1
DKB Recomp v0.4.1

one hour ago

DKB Recomp 0.4.1: a static recompilation of Donkey Kong Bananza (Switch 2) to a native Windows exe with a Vulkan renderer. It is still experimental (0.x). This is a smaller release: what landed in the two days after 0.4.0, most of it from contributors.

The packages hold no game data, and no release ever will. You build the game on your own PC from your own dump of 1.0.0 or the 3.2.0 update, given as a .nup or .nx file or as ExeFS and RomFS folders. Other updates are not supported.

Not run in play by us: the Lanczos window filter at 1440p, AMD FSR 3.1 frame interpolation, AMD Anti-Lag 2 on AMD hardware, FSR 4 on AMD hardware and HD rumble on real controllers. The CRT and window filters and the frame interpolation come from community pull requests and have had little testing in play.

What is new since 0.4.0

  • Crashes. A null edge target in the job graph builder is guarded on 1.0.0 and 3.2.0 (#317). A call that crashed after the Forest Layer snow tunnel gets a guard that skips it when its slot is null (#136, still open until the reporter retests). A GPU that finishes nothing for 30 seconds is reported as a device loss instead of a silent freeze (#329). A failed swapchain rebuild destroys the retired swapchain (#332). The frame generation display thread no longer takes its start lock under the present lock (#343).
  • Picture and upscalers. upscaler = fsr3 is AMD FSR 3.1 temporal upscaling and upscaler = fsr4 is AMD FSR 4 through your own AMD DLLs (#270, #333); both are off by default and fsr4 has not been run on AMD hardware by us. DLSS, FSR 3.1 and FSR 4 also run as native anti-aliasing at scene_res = full on 3.2.0 (#345). Only draws that test or write depth are jittered (#321, #325), and the upscaler dispatch is found by its shape when its hash is not a listed one (#323). The upscaler row stays open on Bananza, and DLSS says in the log why it does not run (#152, #306). Window scaling filter: window_filter = fsr1, linear, catmull_rom or lanczos, with Sharpness at any window size (#340); fsr1 stays the default, and Lanczos at 1440p has not been run by us. CRT filters: crt_filter = off (default), lottes_grille, lottes_tv, hyllian or zfast, with crt_lines (#342). ASTC textures stored as BC3 are copied instead of skipped, which is meant to bring back the Ingot Isle textures (#310, awaiting the reporter). Text above 1080p is sharper behind DKB_FONT_SCALE=1 (#214, off by default). Dug terrain builds on RADV (#322).
  • Frame generation. frame_gen has an Auto mode and 3x to 8x besides 2x, with the pacing they need (#334); it is off by default. At 2x, frame_gen_by = fsr3 makes the frame with AMD FSR 3.1 frame interpolation instead of the built-in one (#346, off by default, not run by us), and it gets the frames without their HUD (#348). Frames are also made from the game's own HDR picture (#337). The Display tab has a Frame generation row and a Made by row (#349), and a set that waited is counted in the log (#352).
  • Performance. The uniform snapshot takes shortcuts (TOTK_USNAP_FAST) and deferred guest reads are written back without waiting, on by default, about 1.5 ms a frame in Canyon Layer (#312). TOTK_MIRROR_OWN leaves the VRAM mirror's own GPU writes out of the sync (#295), TOTK_CALL_REF records some CallCommands as one call (#307), and a submission's GPU writes are copied back in one pass (#328). TOTK_DRAW_MERGE draws runs of c4-only draws as instanced draws and is off by default (#300). The end of a sliced pass is held back without the recorder thread too (#292), and sliced passes are recorded by default on AMD's proprietary driver on the RX 9070 and 9070 XT, the chip measured (#298; RDNA 1 and 2 are left out, #84, #155). An upscaler that is no longer chosen releases its images after 120 frames (#351). On Linux the shader disk caches get 10 GiB of room unless you set a size (#330), and on Intel under Windows the pipeline cache data is fetched only while no pipeline is being built (#331).
  • Sound. The audio renderer is paced evenly instead of in bursts after each device period, which is meant to remove the crackling on a 1024-frame device (#133, #272, #347); #272 awaits the reporter.
  • Languages and text. Polish (pl, EUpl, pl-PL) is a language choice (#335).
  • Input. pad:touchpad can be bound, so the DualSense or DualShock 4 touchpad click can open the map (#291). HD rumble goes to a Switch Pro Controller or Joy-Con as amplitude and frequency (#285, not run on real controllers by us). The Controls tab builds against SDL older than 3.4.18 (#324).
  • Installer and Linux. The setup's host check builds against the release sysroot on Linux, so the AppImage no longer fails at step 1 (#293), and its last check accepts a CPU that has no AVX2 when you build for x86-64-v2 (#318). vulkan_driver runs a Mesa build of your own on Linux without a wrapper script (#339). The --warm-shaders and --prepare-shaders library builds run at below-normal priority on one core fewer than the machine has, from 4 cores up (#3).
  • Diagnostics. The performance overlay has a Breakdown mode that stacks each frame's parts in the graph (#341) and a Counter mode, a large FPS number with a small label (#350); both are values of perf_overlay. AMD Anti-Lag 2 is antilag = 1, off by default, on an AMD GPU whose driver has VK_AMD_anti_lag (#338); not run on AMD hardware by us. TOTK_MEM_REPORT=N logs where the process's commit goes, and tools/lowend_bench.py runs a low-end benchmark. The perf trace counts the waits on the generation-table lock.
  • Long sessions. The VRAM mirror sync asks Windows for written pages before it takes the mirror lock, so the replay and its slices wait less on it (#317, partial: the query itself gets slower as the process ages).

Pull requests in this release

Thank you to everyone who sent code. Each one below is new since 0.4.0:

  • #292 Hold back the end of a sliced pass without the recorder thread too (@GlebShekovtsov)
  • #294 Read page stamps under gen_lock in TOTK_MEMORY=verify (@Zorkats)
  • #295 Leave the VRAM mirror's own GPU writes out of mirror_sync (TOTK_MIRROR_OWN) (@Zorkats)
  • #296 Commit the probe's heap before the freeze checks (exe-probe-gfx-abi with the lazy heap) (@thatbananaman08)
  • #298 Record sliced passes by default on AMD's proprietary driver too (@GlebShekovtsov)
  • #300 Draw runs of c4-only draws as instanced draws (TOTK_DRAW_MERGE, off by default) (@Zorkats)
  • #307 Record CallCommands of a recording without deferred records as one call instead of a copy (TOTK_CALL_REF) (@thatbananaman08)
  • #312 Take shortcuts in the uniform snapshot and write deferred reads back without waiting, about 1.5 ms a frame in Canyon Layer (@Zorkats)
  • #321 Jitter only the draws that test or write depth, so the temporal upscalers are not handed a picture shifted twice (@roverdrubber)
  • #322 Build the geometry stage of dug terrain on RADV, whose input locations past 15 the location check refused (@roverdrubber)
  • #323 Take the game's upscaler dispatch for DLSS, FSR 3.1 and FSR 4 by its shape when its hash is not one of the listed ones (@roverdrubber, landed by hand)
  • #324 Build the Controls tab against SDL older than 3.4.18, whose headers have no SDL_GAMEPAD_TYPE_STEAM (@roverdrubber)
  • #325 Jitter the scene draws by their depth test alone, as the review of #321 noted, and say which draws that leaves out (@roverdrubber)
  • #326 Say in geometry_validate that taking the vertex stage's location limit on RADV is deliberate, as the review of #322 asked (@roverdrubber)
  • #328 Copy a submission's GPU writes back to guest memory in one pass (@roverdrubber, landed by hand)
  • #329 Report a GPU that finishes nothing for 30 seconds as a device loss (@roverdrubber)
  • #330 Give Mesa's and NVIDIA's shader disk caches 10 GiB of room on Linux unless the player set a size (@roverdrubber, landed by hand)
  • #331 Fetch the pipeline cache's data only while no pipeline is being built on Intel under Windows (@roverdrubber, landed by hand)
  • #332 Destroy the retired swapchain when creating its replacement fails (@roverdrubber, landed by hand)
  • #333 Give FSR 4 an input pass, FSR 3.1's output pass, a Linux library and depth flags (@roverdrubber)
  • #334 Add frame generation's Auto mode and the 3x to 8x multipliers, with the pacing they need (@roverdrubber, landed by hand)
  • #337 Make frames from the game's own HDR picture too (@roverdrubber)
  • #338 Add AMD Anti-Lag 2 as an option (antilag = 1) (@roverdrubber)
  • #339 Add an optional vulkan_driver setting so a Mesa build of the player's own can run the game on Linux without a wrapper script (@roverdrubber)
  • #340 Give the window a choice of scaling filter, with Sharpness at any window size and the game's own HDR picture for the filters chosen (@roverdrubber)
  • #341 Add a Breakdown mode to the performance overlay (@roverdrubber)
  • #342 Add Hookshot's CRT filters to the present (crt_filter) (@roverdrubber)
  • #343 Keep the frame generation display thread from taking the start lock under the present lock (@roverdrubber)
  • #345 Run DLSS, FSR 3.1 and FSR 4 as native anti-aliasing at scene_res = full (@roverdrubber)
  • #346 Add AMD FSR 3.1 frame interpolation as a way to make frame generation's 2x frames (frame_gen_by = fsr3) (@roverdrubber)
  • #347 Pace the audio renderer evenly instead of in bursts after each device period (@Gabinou22s)
  • #356 Refill a behind audio queue at once, drop the dead device pacing path (@Gabinou22s)
  • #348 Give AMD FSR 3.1 frame interpolation the frames without their HUD, so the HUD is not moved with the scene (@roverdrubber)
  • #349 Add Frame generation and its Made by row to the Display tab (@roverdrubber)
  • #350 Add a Counter value to the performance overlay: a large FPS number with a small label, after Hookshot's (@roverdrubber)
  • #351 Let an upscaler no longer chosen release its images, feature or context after 120 frames (@roverdrubber)
  • #353 Slice indirect draws and cut the slice threads' lock waits, Canyon Layer replay 11.8 to 10.0 ms a frame (TOTK_PAR_INDIRECT) (@Zorkats)
  • #358 Say in inline_recording's comment that TOTK_CALL_REF records a call instead of a copy (@thatbananaman08)
  • #352 Count frame generation's set waits, name Auto's flight thumbnails, and correct its pacing comment and docs (@roverdrubber)

Issues closed or answered by this release

Thank you to everyone who reported a problem. Open ones wait for the reporter to retest.

Thanks

Pull requests since 0.4.0: roverdrubber, Zorkats, GlebShekovtsov, thatbananaman08, Gabinou22s. Thank you.

Download

Package For Status
DKB-Setup-<version>-windows.zip Windows 10/11 x64 Supported. Carries its own tools and the installer window: unpack, run DKB-Setup.exe, pick your dump.
DKB-Setup-<version>-linux-x86_64.tar.gz Linux x86-64, kernel 6.7 or later Preview. Source package with setup.sh: you install the system packages it lists (./setup.sh --check), then it runs the command-line installer.
DKB-Setup-<version>-macos-arm64.zip macOS 14.4 or later, Apple silicon Preview. Same idea with setup.command. Play is not verified and the destructible terrain is missing (Metal has no geometry stage).

Requirements (Windows)

  • NVIDIA GPU with a recent driver (tested on an RTX 5060 Ti and an RTX 3070 Ti Laptop). AMD and Intel are experimental: on AMD's Windows driver the device loss most players hit is fixed (recorder thread off), and RADV on Linux can still hang (#99, #122, #198, #204). Intel Arc: #3.
  • 16 GB of RAM plus a large page file (the process commits about 9 GB), 32 GB is comfortable.
  • A controller is recommended.

Known issues

  • GPU hangs under RADV on Linux (#99, #122, #198, #204), and a driver fault on Intel Arc (#3).
  • Crashes reported without a reproduction here: #126 (Lagoon to Hilltop tunnel, retest asked), #136 (Forest snow tunnel landing: needs a guard on a call, after this release), #148 (Linux, after the first boss: the next log will name the function).
  • Black shards where salt or acid meets slime (#139, #266).
  • Text is soft at 1440p and above: the game's fonts are drawn at a fixed size (#214; DKB_FONT_SCALE=1 helps small text).
  • A device loss on an RTX 5070 in the Landfill Layer that we cannot reproduce (#304), and frame rate falling after a long session (#317).
  • Heavy scenes can still fall under 60 fps, and the first visit to an area can stutter.
  • Checkerboard textures (#5), #112 and unrecorded NVN 2 calls (#6).

Reporting a bug

Open an issue with the log (dkb.log), the build line at its top (dkb 0.4.1+<commit>), your GPU and driver, and what you did before it failed. Never share generated files, keys or screenshots that show game content.

Roadmap: #10.

What's Changed

  • Commit the probe's heap before the freeze checks (exe-probe-gfx-abi with the lazy heap) by @thatbananaman08 in #296
  • Draw runs of c4-only draws as instanced draws (TOTK_DRAW_MERGE, off by default) by @Zorkats in #300
  • Read page stamps under gen_lock in TOTK_MEMORY=verify by @Zorkats in #294
  • Leave the VRAM mirror's own GPU writes out of mirror_sync (TOTK_MIRROR_OWN) by @Zorkats in #295
  • Take shortcuts in the uniform snapshot and write deferred reads back without waiting, about 1.5 ms a frame in Canyon Layer by @Zorkats in #312
  • Jitter only the draws that test or write depth, so the temporal upscalers are not handed a picture shifted twice by @roverdrubber in #321
  • Build the geometry stage of dug terrain on RADV, whose input locations past 15 the location check refused by @roverdrubber in #322
  • Build the Controls tab against SDL older than 3.4.18, whose headers have no SDL_GAMEPAD_TYPE_STEAM by @roverdrubber in #324
  • Jitter the scene draws by their depth test alone, as the review of #321 noted, and say which draws that leaves out by @roverdrubber in #325
  • Say in geometry_validate that taking the vertex stage's location limit on RADV is deliberate, as the review of #322 asked by @roverdrubber in #326
  • Report a GPU that finishes nothing for 30 seconds as a device loss by @roverdrubber in #329
  • Add an optional vulkan_driver setting so a Mesa build of the player's own can run the game on Linux without a wrapper script by @roverdrubber in #339
  • Hold back the end of a sliced pass without the recorder thread too by @GlebShekovtsov in #292
  • Pace the audio renderer evenly instead of in bursts after each device period by @Gabinou22s in #347
  • Keep the frame generation display thread from taking the start lock under the present lock by @roverdrubber in #343
  • Add AMD FSR 3.1 frame interpolation as a way to make frame generation's 2x frames (frame_gen_by = fsr3) by @roverdrubber in #346
  • Run DLSS, FSR 3.1 and FSR 4 as native anti-aliasing at scene_res = full by @roverdrubber in #345
  • Give FSR 4 an input pass, FSR 3.1's output pass, a Linux library and depth flags by @roverdrubber in #333
  • Give the window a choice of scaling filter, with Sharpness at any window size and the game's own HDR picture for the filters chosen by @roverdrubber in #340
  • Add a Breakdown mode to the performance overlay by @roverdrubber in #341
  • Add Hookshot's CRT filters to the present (crt_filter) by @roverdrubber in #342
  • Record sliced passes by default on AMD's proprietary driver too by @GlebShekovtsov in #298
  • Add Frame generation and its Made by row to the Display tab by @roverdrubber in #349
  • Let an upscaler no longer chosen release its images, feature or context after 120 frames by @roverdrubber in #351
  • Add a Counter value to the performance overlay: a large FPS number with a small label, after Hookshot's by @roverdrubber in #350
  • Make frames from the game's own HDR picture too by @roverdrubber in #337
  • Add AMD Anti-Lag 2 as an option (antilag = 1) by @roverdrubber in #338
  • Give AMD FSR 3.1 frame interpolation the frames without their HUD, so the HUD is not moved with the scene by @roverdrubber in #348
  • Count frame generation's set waits, name Auto's flight thumbnails, and correct its pacing comment and docs by @roverdrubber in #352
  • Record CallCommands of a recording without deferred records as one call instead of a copy (TOTK_CALL_REF) by @thatbananaman08 in #307
  • Refill a behind audio queue at once, drop the dead device pacing path by @Gabinou22s in #356
  • Slice indirect draws and cut the slice threads' lock waits, Canyon Layer replay 11.8 to 10.0 ms a frame by @Zorkats in #353
  • Say in inline_recording's comment that TOTK_CALL_REF records a call instead of a copy by @thatbananaman08 in #358

New Contributors

Full Changelog: v0.4.0...v0.4.1

Don't miss a new dkb-pc-recomp release

NewReleases is sending notifications on new releases.