Skip to content

Witch Time v0.8.0

Choose a tag to compare

@github-actions github-actions released this 05 Nov 15:16
· 70 commits to main since this release
5be114c

Witch Time v0.8.0

This update adds FSR 3.1 (#118), as anti-aliasing or to upscale. It's experimental and off by default. It also raises the frame rate in Shibuya when the GPU has a lot to draw (#132, #133). At 90 and 120 fps, stick moves and the windows around Witch Time now last as long as at 60 (#134). And on Windows, REBUILD no longer stops at "Access is denied" when an old copy of the game is still running (#127).

Coming from v0.7.0, you'll rebuild once more, but there's no PREPARING SHADERS this time. The first start after updating says the game was built by an older version, so press REBUILD and let it finish. Only 4 of the game's 409 code files changed, and only those compile again. In test builds on a Ryzen 7 7800X3D, 3 such files compiled in 9 seconds, against about 3 minutes for all of them. Your dump folder doesn't need to be there. The game's shaders haven't changed, and your converted textures are kept, so there's no new conversion. If you built the picture book, REBUILD builds it again from the files it already has. If you didn't, you can put its program1/ folder in your dump before you press REBUILD (see the README's The picture book). The first time Bayonetta 3 starts, it adds the new upscaler key to your config/graphics.ini, set to off. Your values stay, and the file as it was goes to graphics.ini.bak. Coming from v0.6.1, v0.6.0 or v0.5.2, PREPARING SHADERS shows once more, on every GPU, because v0.7.0's golden light fix changed the game's shaders. Coming from v0.5.2 or older, rename any SSBU_* environment variable you set to WT_* (see the v0.6.0 notes). Coming from an older version, you'll also get the one-time steps listed in the v0.5.0 notes.

Linux (x86_64) and Windows 10/11 (64-bit). macOS and Android are planned.

Download and run

Linux

  1. Download WitchTime-v0.8.0-linux-x86_64.tar.gz and, if you like, check it against the .sha256 file.
  2. Unpack it and run ./witch-time.

Windows

  1. Download WitchTime-v0.8.0-windows-x86_64.zip and, if you like, check it against the .sha256 file.
  2. Unzip it to a folder you can write to. A short path with no special characters works best.
  3. Run witch-time.exe.

Windows Defender might flag Witch Time as a Trojan (#64). The release's programs aren't code-signed, and each release has new ones Defender hasn't seen. Download it only from this page and check the zip's hash: certutil -hashfile <zip> SHA256.

Already on v0.4.0 or later? The launcher's title menu offers Update, tagged v0.8.0 (UPDATE TO v0.8.0 on launchers before v0.6.0), and keeps your config/ files. If you unpack v0.8.0 over your old folder by hand instead, copy your config/ folder somewhere first, because the archive has the default files.

What you need is the same as before: a Vulkan-capable GPU and driver, plus OpenGL and FreeType on Linux, or the Media Feature Pack on Windows "N" editions for the album's video clips. For the picture book you also need program 1 from your copy of the game's update 1.2.0 (see the README).

Changes

  • FSR 3.1, as anti-aliasing or to upscale (#118, #129, #131). AMD FSR 3.1.4 is now built in. It's experimental and off by default. Over several frames it smooths jagged edges, shimmer and the screen-door pattern of fading objects. To turn it on, open the PC settings panel (F1 or Start + Select) and use the FSR 3.1 row, just below Film grain: Off, Native AA, Quality, Balanced, Performance or Ultra Performance. A change applies when the panel closes, like Internal resolution. The launcher has the same row under Settings, Graphics. In config/graphics.ini the key is upscaler, with the values off, fsr3 (Native AA), fsr3_quality, fsr3_balanced, fsr3_performance and fsr3_ultra_performance. A value it doesn't know counts as off. It works with the Vulkan renderer only, and it isn't a row of the game's own Options screen. Witch Time ships no AMD program: AMD's MIT-licensed shader code from the FidelityFX SDK v1.1.4 is compiled on your PC.
    • Native AA runs FSR at the internal resolution, with no upscaling. On the test route, rails, platform edges and diagonal leaves lost their steps and stopped crawling. Gravel looks a little softer. At 2160p (resolution_scale = 2.4) on an RX 9070 XT (Linux, RADV), GPU time per fight frame went from 4.79 ms with FSR off to 6.70 ms. Its cost at 900p, or on a weaker GPU, hasn't been measured.
    • The screen door of fading objects (the dither in #118's screenshots). Objects that fade, such as the game's cross-fades between detail levels and objects close to the camera, are drawn through a fixed 64 x 64 pattern of holes. FSR alone would keep that pattern, because it's the same every frame. With FSR on, the pattern changes every frame (13 versions), and each version covers exactly the share of pixels the game asked for, so FSR blends them into a smooth fade. We tested a billboard in a cutscene cross-fade, with the fade held at half by a test switch. The holes kept 0.76 to 0.93 of their strength without FSR, 0.37 with FSR and the game's fixed pattern, and 0.13 with the changing one. With FSR off, the game's pattern is left as it is.
    • Quality, Balanced, Performance and Ultra Performance draw the 3D scene smaller: 1/1.5, 1/1.7, 1/2 and 1/3 of the internal resolution across. FSR then upscales it to the internal resolution, not to your window. The HUD, film grain and FXAA are still drawn at the full internal resolution. At 2160p on the RX 9070 XT, GPU ms per fight frame: Off 4.79, Native AA 6.70, Quality 4.43 (scene at 2560x1434), Performance 3.66 (1920x1075), Ultra Performance 3.31 (1600x896). Balanced wasn't measured. Quality looked close to Native AA on the route. Performance and Ultra Performance soften the small text on signs.
    • The upscaling modes only do something above 900p. The scene is never drawn smaller than the game itself draws it: 900p, or less when the game lowers its own resolution in a busy moment, because two of the game's shaders keep images at its own pixel size. At 900p or lower, every mode works like Native AA. At 1080p all four upscaling modes draw at 900p. At 1440p, Quality draws about 1707x956 and the other three draw at 900p. At 2160p, Ultra Performance stops at 900p too. Only 2160p was measured; the other sizes come from the code. So FSR can't make the game faster at the default 900p or below, profile = low included.
    • On the test PC the frame rate stayed the same in every mode. The route was limited by the CPU, so the smaller modes gave free GPU time. A PC that waits on its GPU at a high internal resolution is where Quality and below should help.
    • Turning it on or off needs no REBUILD and no PREPARING SHADERS. FSR's 8 passes are compiled the first time FSR runs, which takes about half a second once, so the game may pause for it. After that they come from the shader cache. If your GPU lacks a feature FSR needs, FSR stays off and the scene is drawn at the full internal resolution, in every mode. build/run/witch-time-game.log then has a line starting nvn_vk_fsr: FSR is off: that says why. When FSR works, it says nvn_vk_fsr: FSR 3.1.4 ready. Please include that log if you report a problem with FSR.
  • Higher frame rate in Shibuya when the GPU has a lot to draw (#133, #132, #137). In the city, Witch Time's render thread kept stopping to wait for results the GPU had to produce, and each time it waited for everything the GPU had queued, the previous frame included. So the CPU and the GPU took turns, a frame cost both of their times added up, and neither one looked busy. That fits what #133 shows: 35 fps in Shibuya on an RTX 5080 with an i9-10850K at 3440x1440, with the CPU at 19 % and the GPU at 51 %. Now the GPU reads those results itself (which objects are visible, the particle draws, and the brightness the last frame measured), and the CPU goes on with the frame meanwhile.
    • On an RX 9070 XT with a Ryzen 7 7800X3D on Linux, at 21:9 and Internal resolution 2160p (resolution_scale = 2.4), over 200 s of the Shibuya street fight, about 57.5 fps became 60.0. Frames on time for 60 went from 82 % to 98.4-98.8 %, and the 1% low from about 44.5 to 51.2-52.8 fps. The render thread's wait for the GPU went from 6.9 ms a frame to 0.04 ms.
    • At the default 900p and 16:9, that PC already held 60, and it still does. Slightly more frames take 18.5-20 ms (98.7 % on time, against 99.2 %), but fewer take longer than that.
    • It's on by default, with nothing to set. It was measured only in Shibuya, at frame_rate = 60 with FSR off, on AMD with RADV on Linux. For #133 on NVIDIA and Windows, and for #132 on the Steam Deck, it should help, but nobody has confirmed it yet (see Not tested yet).
  • Stick moves at 90 and 120 fps (#134, #135). At Frame rate 120, moves that need a quick stick motion, such as back, forward + attack and forward, forward + attack, rarely came out. At 60 they came out reliably. The game counts their timing windows in its own logic steps, and above 60 it runs one step a frame, so at 120 the window for the second flick lasted 83 ms instead of 167 ms, and at 90 it was a third shorter. These windows now count the 60 Hz frames of game time, so they last as long above 60 as at 60.
    • We tested with scripted flicks 0.05 to 0.25 s apart on an RX 9070 XT on Linux, and checked whether the game registered the second flick. At 60 it did with the flicks up to 0.14 s apart, and not at 0.18 s. Before this fix, 120 only managed up to 0.08 s. Now 120 matches 60 in all 24 trials.
    • The stick-rotation windows get the same fix. So does the press buffer: an attack you press during a recovery now waits as long at 120 as at 60.
  • The windows around Witch Time at 90 and 120 (#134, #135). Some of them ran half as long at 120: the short slow-down just before Witch Time starts, the short window after a hit lands in which Viola's parry, Bat Within and Mahaa-Kalaa start Witch Time, the slow-downs on finishing hits, the near-freeze when Bayonetta is hit and the long slow-down of a fatal hit. They now count 60 Hz frames too. We measured only the slow-down before Witch Time: 30 frames at 60, 30 at 120 before the fix (so half the time), and 60 at 120 now. The dodge window that sets off Witch Time already kept real time above 60, and so did Witch Time's length.
  • Timing windows cut short below 60 fps with Frame rate above 60 (#135). With Frame rate at 90 or 120, when the game ran below 60 fps, a countdown sitting at 0 kept collecting frames it owed, up to 8, and the next window it counted paid them all as soon as it opened. In our test, slowed to 44 fps on purpose, a 15-frame window, a quarter of a second, dropped to 7 frames at once. This could also hit the demon form's double-tap dodge, converted for 90 and 120 in v0.2.5. Now the window lasts its 0.25 s at 44 fps.
    • Frame rate is 60 by default, and at 60 none of these three changes does anything. They work only above 60: Frame rate (experimental) in the launcher's Settings or the F1 panel, Experimental Frame Rate on the game's Options screen, or frame_rate = 90 or 120 in config/graphics.ini. It applies at the next start. To check whether a timing problem comes from these conversions, start the game with WT_FPS_COUNTERS=0 set, for example WT_FPS_COUNTERS=0 ./witch-time on Linux. Above 60 that puts every converted counter back to the game's own count of 1 a frame, the ones from v0.2.5 included.
  • REBUILD no longer stops at "Access is denied" when an old copy of the game is still running on Windows (#127, #128). The link step stopped at its first file with avcodec-61.dll: Access is denied. (os error 5), and again at every REBUILD. A witch-time-game.exe was still running in the background, perhaps left over from the last time you played, or one that froze while closing. Windows won't replace a program or DLL that a running program has loaded, but it does let one be moved. So the link step now moves such a file into build\old\ and puts the new one in its place. build\logs\link.log says so for each file it moved, and the next link step deletes what nothing holds any more. Linux could always replace these files. The reporter got past it by restarting the PC before this change existed, so they haven't run the fix.

Known issues

  • If the book step fails, Play is blocked for Bayonetta 3 too. This update's REBUILD runs that step again for everyone who built the book. The launcher keeps saying the game was built by an older version. This happens when the program1 files don't match, a recompressed one included, or when the prologue fails to build. If the files don't match, the launcher's message names the file and the size and SHA-256 it expected. If the prologue fails to build, its logs are build/logs/book-*.log. Take program1/ out of your dump and press REBUILD, and the rest builds without the book. Please tell us on #115 and include the message or those logs.
  • FSR 3.1 is experimental (#118). With it on, Bayonetta's hair, fast limbs and effects that move on their own can leave faint trails, because the game draws no motion vectors, so FSR only knows how the camera moves. FSR starts its smoothing over each time the game changes its own dynamic resolution, and at what looks like a camera cut. There's no sharpening, no special handling of see-through effects and no texture bias for the smaller modes yet. Only objects' fade pattern changes: fading shadows keep the game's fixed one. Frame generation (#119), FSR 4, DLSS and XeSS aren't included.
  • FSR's upscaling modes probably don't work in the picture book's prologue. FSR finds its place in a frame by Bayonetta 3's own shaders, and the camera jitter it needs is only added to Bayonetta 3. Above 900p, Quality and below may still draw the prologue's scene smaller with nothing to upscale it. Nobody has tried it. Set FSR 3.1 to Off before you open the book.
  • Above 60 fps, hit-stops are shorter (#134). The short freeze when a hit lands still counts frames, so at frame_rate = 120 it lasts half as long as at 60, and hits may feel lighter. This is a different thing from the near-freeze when Bayonetta is hit, which is fixed. We left hit-stops out on purpose: they need a combat test before they change. Other frame-locked timers haven't been read or converted yet either, so above 60 some effects and countdowns may still run fast, which is why Frame rate above 60 is still experimental. The default, 60, isn't affected.
  • Cutscenes can stop for a few seconds the first time (#136). On an RX 6600 on Windows 11 with v0.7.0, the first time a cutscene played, the picture could stop for a second or more and the sound fell out of step. The reporter says the same cutscene stutters less the second time and sometimes not at all the third, and that gameplay is smooth. In the log of that run the picture stopped three times, for up to 4.6 s. No shaders were being made, so it doesn't look like shader compiling, and we don't know the cause yet. If it happens to you, please attach build/run/witch-time-game.log to #136.
  • Blocky effects on an RX 6700 XT on Windows 10 (#126). Several effects look blocky, like the smoke in the report's screenshot, and some textures don't show right. Raising texture filtering doesn't change it. We haven't looked into it yet. If you see it, please add a screenshot, where in the game it was, and build/run/witch-time-game.log to #126.
  • Combo videos play only once (#138). In the menus that show a combo's demonstration video, the video stops when it ends instead of starting again. We haven't checked yet whether the Switch version loops them.
  • Small dark marks on character outlines (#93). The large black triangles were fixed in v0.5.2, but a few much smaller ones remain in some shots. Blocky marks in Bayonetta's hair remain too.
  • Frame rate drops in some areas. The Shibuya change above was measured only on an RX 9070 XT on Linux, so we don't know yet how much it helps on an RTX 5080 on Windows (#133) or a Steam Deck (#132). One more kind of GPU result is still read on the CPU with a wait. None happen in Shibuya, but other chapters weren't checked. Drops have also been reported on an RX 580 (#50), on an RX 6600 whose GPU wasn't fully used (#89), and in two cutscenes on an RTX 3070, from 60 to about 46 fps (#93).
  • Frame rate in the ride of Chapter 1, Verse 5 (#120). It drops to about 47 fps and feels much slower, at frame_rate 60, 90 and 120 alike in the report. We haven't reproduced it yet, and the cause isn't known. If you see it, please tell us on #120 your frame_rate and cutscene_fps, and attach build/run/witch-time-game.log from a run that reaches the ride.
  • HDR looked dim on one PC (#130). On Windows 11 the picture was dim and dull with hdr = auto, the default. hdr = on in config\graphics.ini fixed it, and after that auto worked too, so we don't know what went wrong. If HDR looks dim for you, try hdr = on and send us build\run\witch-time-game.log from that run.
  • On Windows, a copy of the game can keep running after you've stopped playing (#127). We don't know why yet. REBUILD gets past it now, but the old copy keeps running. To end it, open Task Manager, go to Details and end every witch-time-game.exe, or restart the PC.
  • The conversion runs even if your GPU doesn't need it. One that reads ASTC itself (most Intel GPUs) gets a pack it never uses. REMOVE in Settings gives the space back.
  • A stopped conversion redoes the group of archives it was in, and a mod's textures aren't converted.
  • Below about 41 fps the spider ride falls behind. It needs 60 logic steps a second, and a frame runs at most two: at 25 fps it gets about 50.
  • When you rebind buttons, the ride's four-circle prompt shows the right button in the old button's place. For example, Jump moved to B shows a B at the bottom, not on the right where an Xbox pad has it.
  • Other long stalls. Only one extra shader variant is prepared, so another kind in a later chapter would still compile at its first draw. Any stall over 3 s in a cutscene still loses the part past 3 s.
  • The prologue stutters on its first run. PREPARING SHADERS doesn't run for it, because none of its shader containers are found, so each shader compiles at its first draw.
  • Intel UHD 620 (#55). The game hangs and closes during the loading clock.
  • RTX 5070 on Windows 11 (#20) is still open. Other players with an RTX 5070 on Windows 11 play normally, and the reporter hasn't posted since v0.4.0.
  • Rendering artifacts with large mods (#46) is still open.

Not tested yet

Please report anything you see with these:

  • FSR 3.1 in this exact build, and anywhere but Linux on AMD. It was run only on an RX 9070 XT with RADV on Linux, and before this release's Shibuya change (#137), which moved where in a frame FSR starts. Nobody has run FSR on this build since. It hasn't run on Windows, NVIDIA, Intel or a Steam Deck, the #118 reporter's RDNA 2 card on Windows included. Balanced, the modes below 2160p, the long fade close to the camera from #118's screenshots, FSR at frame_rate 90 or 120, at 21:9 or with HDR, and the launcher's new row on screen haven't been checked either. If something looks wrong, set FSR 3.1 back to Off and tell us on #118 your GPU, the mode and your internal resolution, with a screenshot and build/run/witch-time-game.log.
  • The Shibuya change on Windows, NVIDIA, Intel and the Steam Deck (#133, #132). It has only been run on an RX 9070 XT with RADV on Linux. For Windows, CI only built it and ran the build tool's tests. If you see something v0.7.0 didn't do, for example objects or particles missing or flickering, or the brightness jumping, turn the change off: start Witch Time with all four of WT_VK_GPUCOUNT, WT_VK_CBLAZY, WT_VK_CBDISCARD and WT_VK_GPUCOPY set to 0. They're read once at start, and graphics.ini has no key for them.
    • On Linux: WT_VK_GPUCOUNT=0 WT_VK_CBLAZY=0 WT_VK_CBDISCARD=0 WT_VK_GPUCOPY=0 ./witch-time
    • On Windows, open a Command Prompt (cmd, not PowerShell, where set does nothing) and type cd /d "<your Witch Time folder>", then set WT_VK_GPUCOUNT=0, set WT_VK_CBLAZY=0, set WT_VK_CBDISCARD=0, set WT_VK_GPUCOPY=0, then witch-time.exe, all in that same window.
    • Please tell us on #133, or on #132 for a Deck, and say whether the frame rate changed with the switches. Attach build/run/witch-time-game.log, and copy it before you start the game again, because each start overwrites it.
  • The 90 and 120 fps timing fixes, played by hand (#134, #135). Back, forward + attack and forward, forward + attack were timed with scripted inputs on Linux, and the slow-down before Witch Time was counted there too. The window after a hit, the finishing-hit and fatal-hit slow-downs and the near-freeze never came up in a test run, and the demon form's double-tap dodge wasn't tried again. The 44 fps case was made by slowing a fast PC on purpose. Nothing has been played by hand, run at 90 or run on Windows. If moves or Witch Time still feel off above 60, please tell us on #134.
  • REBUILD with a real game running on Windows (#127, #128). CI's Windows test holds the file with a stand-in program, not a running game. If REBUILD still stops with "Access is denied", end every witch-time-game.exe in Task Manager, press REBUILD again, and open an issue with build\logs\link.log.
  • The update from v0.7.0 itself. Nobody has run it through the launcher yet, with or without the picture book, so the rebuild time above comes from test builds.
  • The picture book's round trip. Nobody has opened the book through the launcher and played the prologue to its end, so what the prologue writes to the shared save hasn't been checked either. Copy build/userdata (your saves) somewhere safe before you open the book. Opening the book without the picture book built (the NOT READY message) has never been run. The picture book has never been built or run on Windows. If you try any of these, please tell us on #115.
  • The golden light fix on NVIDIA and Intel (#100). It was measured on AMD with RADV on Linux. The reporter confirmed an earlier form of it on an RX 6700 XT on Windows. If the gold cracks look wrong, please tell us on #100.
  • The cloud fix on Windows (#107, #112). The issue is closed, but nobody has run the fix on AMD on Windows, the RX 580 it was reported on included, or on NVIDIA or Intel.
  • The v0.6.1 update fix on Windows (#114, #116). If REBUILD fails after updating, unzip v0.8.0 into a new, empty folder. Copy your config folder and build\userdata (your saves) into the same places there, pick your dump, let it build, and tell us on #114.
  • A real ultrawide monitor (#77), including Unlocked in borderless and fullscreen.
  • Carried over from v0.6.0: v0.6.0's new features on Windows (aspect ratios, Discord, Model Detail, Film Grain and the new launcher), Model Detail on a weaker GPU, Discord past chapter 2, the new launcher with a real controller and the digest check in the updater.
  • Carried over from v0.5.2: Viola's rebound layout and resolution_scale below 1.0 with the depth change.
  • Carried over from v0.5.1: changing resolution from the game's Options rows and live resolution on Windows.
  • Carried over from v0.5.0: converted textures in play on a Steam Deck and the graphics.ini update on your own edited file, which this update runs for every player.

If something fails, include build/run/witch-time-game.log, the files in build/logs/, your OS and your GPU.

This release contains no game code or assets. Everything derived from the game is made on your PC from your own dump. Please never share it.

What's Changed

  • Switch the README logo to the Cinzel Decorative wordmark (#123)
  • Link step: a file a running game has loaded is moved aside, not renamed over (#128)
  • FSR 3.1 as anti-aliasing at the internal resolution, and a dither FSR can fade (#129)
  • FSR 3.1 quality modes: the scene drawn smaller and upscaled (#131)
  • 120 fps: stick-motion moves and Witch Time's frame windows count 60 Hz frames (#135)
  • The render thread and the GPU overlap: indirect counts, records and buffer copies on the GPU (#137)

Full Changelog: v0.7.0...v0.8.0