Skip to content

Releases

Steam Deck build 2026-11-13b — background thread held, chunks of 500, sample-window settings

Choose a tag to compare

@RufflezAU RufflezAU released this 13 Nov 12:38

⬇ Download for Linux / Steam Deck


Measured on a Steam Deck OLED (Van Gogh, RADV), in Game Mode and through a headless harness, at a
matched scene of 707 draws a frame (87 samples a side, same config both sides, measured twice where
a config line was the variable).

The frame, at that scene

replay whole frame fps
previous build 19.19 ms 45.25 ms 22.1
this release 12.87 ms 35.46 ms 28.2

That is −33 % of the replay, −22 % of the frame, +28 % of frame rate, with the GPU idling at
5.8–12.9 ms of it.

What produced it

The renderer (hookshot_vk.so, attached, plus the settings below)

  • the recording is cut every 500 draws instead of 120: about 21 cuts a frame, each re-sending all
    of the state. On its own: replay 14.60 → 12.87 ms.
  • push constants already set on the command buffer are not pushed again (a draw's push was ~5 % of
    the replay's commands).
  • texture copies barrier the transfer stages only, not every command.
  • uniform-update payloads live in one flat arena instead of a std::vector per update.

The runtime (hookshot, built from the branch — not in these files)

  • the game's background model thread is held to 25 % of a processor: it used 66 % of one, ahead of
    the renderer, while the game had 4.0–4.4 processors of work for 4 cores.
  • the game's eleven idle job workers sleep on a timer instead of spinning (13–26 % of a core each).
  • the SVC release spin stays at 20 µs: at 1 µs the frame was 45.25 → 44.74 ms; at 20 µs 45.25 → 36.23.
  • waiter tables hashed, the audio parameter queue lock-free, the NVN block pool per thread.

Settings, and why — deck-settings.env is in this release:
RX_VK_CHUNK=500, RX_VK_FRAMEGEN=0 (at 2× it cost 8.5 ms of GPU a frame and doubled shadows),
RX_VK_CHECKPOINTS=0 (a marker write before every draw and dispatch, ~2 300 a frame),
HOS_SHORT_SLEEP_TIMER=1, HOS_SPIN_US=20.

The AppImage is here. Hookshot-deck-2026-11-13b-x86_64.AppImage builds the game from your own copy, with a
hookshot runtime that carries the runtime-side changes above. Its sysroot is the previous release's,
glibc 2.35, so it runs where the last one did. Verified before upload: it extracts under its own
runtime, carries AppRun, the emsdk clang, the bundled sysroot, Python 3.12 and the renderer above,
and the bundled compiler builds and runs a C program against that sysroot. One limitation inherited
from the image it was repacked from: the sysroot has no C++ headers, so a build with DLSS ticked (the
only path that compiles the renderer) needs the sources from the branch instead.

No game data is in this release, and none is in any of them.

Steam Deck build - verified stable (rebased, 75% default)

Choose a tag to compare

@RufflezAU RufflezAU released this 12 Nov 10:37

⬇ Download for Linux / Steam Deck


What this is

The Steam Deck build, rebased onto upstream main — 1,771 upstream commits replayed under our Deck patches. Only 7 files / 89 lines of local patch survived the rebase, because most of what we had patched by hand had been fixed properly upstream in the meantime.

This release ships the module plus the exact settings bundle it was measured with, so the numbers below are reproducible.

Measured against the previous stable build

Same session, same save, same area, same settings (50 % internal resolution), run back to back:

build fps p50 p99
previous stable (pre-merge, 1a60c0ff) 23.2 43.10 ms 43.86 ms
this build (8a215c7f / 7db6d5f6) 24.0 41.67 ms 42.92 ms

+3.4 % frame rate and a tighter p99. The rebase did not cost performance — it gained it, along with the upstream fixes.

Settings

75 % internal resolution is the default in the bundled config: it is the sharpness/frame-rate compromise the player picked, and it costs about 2.5 fps over 50 %.

  • Frame generation is off. Measured: it cost 5 ms of base frame time and produced ghosting and doubled shadows.
  • Presenter/replay thread left at the upstream default (off) — upstream measured 28 black frames per 1,000 with it on.
  • Six emulated guest cores (fewer breaks boot).

Assets

  • hookshot_vk.so — the Vulkan module, drop it beside hookshot.
  • stable.env / stable.ini — the measured settings, beside the executable.
  • SHA256SUMS — checksums.

The full AppImage bundle containing the identical binaries is in the deck-rebase-main release.

Known issues

  • Distant water/river can flicker — under investigation.
  • A brief stutter the first time a new area's shaders are seen (one-time pipeline translation, not a steady-state cost).
  • Frame rate is thermally sensitive on this device: the same binary measures 26.6 fps cool and 21–23 fps hot, so only back-to-back comparisons are meaningful.

Which file do I download?

On Download
Windows HookshotBuilder.exe, the setup
Linux, Steam Deck the .AppImage

You need your own copy of the game. hookshot-setup.txt, hookshot-linux.txt and the .zsync are for the updaters, and the -Windows-x64.zip (the setup and its recomplauncher.json) for RecompLauncher: on Windows, download HookshotBuilder.exe instead. To check a download: sha256sum -c --ignore-missing SHA256SUMS, or gh attestation verify <file> -R RufflezAU/Hookshot.

Steam Deck build 2026-11-12 - rebased onto upstream main (+1771 commits)

Choose a tag to compare

@RufflezAU RufflezAU released this 12 Nov 05:19

⬇ Download for Linux / Steam Deck


Rebased onto upstream 55b976c0. Our patch set is 7 files / 89 lines; everything else is upstream's.

old base (fallback) this build
fps 27.7 26.8
frame p50 36.1 ms 37.04 ms
p99 36.63 ms 39.53 ms
GPU time 14.6 ms 12.6 ms
replay 19.1 ms 16.4 ms

3% slower in total, but the renderer is 2 ms and the replay 2.7 ms faster - upstream does more per frame elsewhere. In exchange: 1771 upstream fixes, a working pipeline cache, and the lighting fix that removes our fog workaround.

Our patches, all in and verified live in the log

  • HOS_REALTIME_FLOOR - real time below 15 fps (upstream still hardcodes it). Log: vi: real-time step floor 10 fps
  • RX_VK_BATCH_IDLE_MIN - the throttle on NotRexed#1993's idle flush. Raw, it stalls this Deck for 35.9 s. Worth +20%.
  • NOP_SET_RT compares before bumping rt_serial
  • push-constant dedupe, RX_VK_GRID_CAP
  • alpha test from the NVN colour state (upstream has the emitter; nothing set it)
  • texture refault pin (RX_VK_STICKY_MB)
  • frame-gen pacer's 500/250 ms bounds

Build fixes needed for this toolchain (not upstream, not ours)

  • nx2/build.py: HOOKSHOT_CXX_INCLUDE / HOOKSHOT_CXX_LIB for the C++ headers and the static libstdc++ - without them every C++ file fails with 'algorithm' file not found
  • host/posix/glibc_compat.c: __isoc23_strtoul and arc4random, which the toolchain's libstdc++.a references and the 2.35 sysroot does not export

Dropped because upstream has them (or because they were proven broken)

tree-shadow shader (NotRexed#1581), the handle guard (758e563), NotRexed#1338 noise volumes, NotRexed#1507 64-lane warps, NotRexed#1993, the count cap, the index dedupe, the PGO plumbing - all upstream, several as supersets. Dropped after measurement: the safe-mode interpolation, a sampler cache by id (broke the game), the fused FMA (broke the game). Dropped because upstream fixed the root cause: the OoT fog/exposure/blur workaround (NotRexed#2042) - RX_VK_OOT_VISUAL/OOT_FOG are now no-ops.

Known issues

  • Frame time is CPU-bound and serialised: replay 16.4 ms + GPU 12.6 ms + game code. Overlapping the CPU and GPU is the remaining large win.
  • Occasional fog/cloud/background flicker; this build runs the game's own lighting, unmodified, which is upstream's fix rather than our band-aid.

Files

hookshot_vk.so is the module from this source (md5 7db6d5f6). The game binary is not distributed - it is the user's own dump.

Which file do I download?

On Download
Windows HookshotBuilder.exe, the setup
Linux, Steam Deck the .AppImage

You need your own copy of the game. hookshot-setup.txt, hookshot-linux.txt and the .zsync are for the updaters, and the -Windows-x64.zip (the setup and its recomplauncher.json) for RecompLauncher: on Windows, download HookshotBuilder.exe instead. To check a download: sha256sum -c --ignore-missing SHA256SUMS, or gh attestation verify <file> -R RufflezAU/Hookshot.

Steam Deck build 2026-11-13 - camera stall gone, exit freeze gone, shaders prepared

Choose a tag to compare

@RufflezAU RufflezAU released this 12 Nov 16:32

⬇ Download for Linux / Steam Deck


What this build fixes

Three things found and fixed by measurement, all on top of the rebased tree:

Turning the camera no longer stalls. New textures streaming in used to cost a staging arena each time: making one with ~2.5 GB of pools and textures held took 12–25 ms, and the arena the last burst used had been given back, so every burst made another — 28 times in one session, each a stall of the replay thread. Four arenas are now made at replay start on an empty device and kept. Measured over 240 s: 28 slow creations before, 1 after.

Closing the game no longer freezes the desktop. The emulator aborts on the heap during shutdown (malloc(): unsorted double linked list corrupted). That alone is harmless — but ulimit -c is unlimited on SteamOS and the kernel pipes cores to a Python submitter, so it was writing a core of a 495 MB process with ~2.5 GB mapped. The launcher now sets ulimit -c 0. The corruption itself is still there and still being chased; it is exit-only.

Sharper frames at the same rate. A direct-mapped front cache now answers repeated texture-view lookups (10,600 a frame, each was a hash-map find). Average fps is unchanged; p99 41.32 ms vs 42.37 ms and 1 % low 24.2 vs 23.6 in a back-to-back A/B.

Shader preparation

hookshot --prepare-shaders was run with these settings: 17,645 pipelines predicted and built (17,741 asked, 96 duplicates), 786 programs, 1,783 graphics + 109 compute pipelines. The effect, per frame:

before after
waits for shaders or pipelines 9.4 ms 0.0 ms
draws left out (not ready) 1.1 0.0
translation, whole session 4,608 ms 202 ms

It costs nothing in steady state (21.8 vs 21.6 fps back to back). The caches are not shipped here — run hookshot --prepare-shaders once and they are written beside the executable.

Measured

75 % internal resolution, the settings in this release: the emulator's replay is 21 ms of a 42–45 ms frame. Frame rate on this device swings with load and heat — the same binary measured 24.6 fps cool and 16.3 fps in a heat-soaked session — so treat any single number with suspicion. The replay time is the stable metric (±8 %).

Known issues

  • Distant water/river can flicker.
  • The exit-time heap corruption above (no longer costs anything).
  • The prepared program cache is 4.3 GB on disk.

Install

hookshot_vk.so beside hookshot; deck-settings.env and deck-settings.ini beside it as hookshot_env.txt and hookshot_options.ini. The main hookshot binary is unchanged from deck-stable-75.

Which file do I download?

On Download
Windows HookshotBuilder.exe, the setup
Linux, Steam Deck the .AppImage

You need your own copy of the game. hookshot-setup.txt, hookshot-linux.txt and the .zsync are for the updaters, and the -Windows-x64.zip (the setup and its recomplauncher.json) for RecompLauncher: on Windows, download HookshotBuilder.exe instead. To check a download: sha256sum -c --ignore-missing SHA256SUMS, or gh attestation verify <file> -R RufflezAU/Hookshot.

Note on frame generation

The settings in this release leave frame generation off (RX_VK_FRAMEGEN=0) — that is the known-good baseline.

It is being tested on the Deck separately: it needs a real window, so it cannot be A/B'd from a script (the test
harness runs the game offscreen, where the emulator disables it). To try it, set RX_VK_FRAMEGEN=1 in
hookshot_env.txt; the in-game Frame generation row then decides. Worth watching: the base frame rate before
interpolation, and ghosting or doubled shadows.

The launcher and settings this build was measured with are versioned in the repository under deck/, with a README
recording the two things a rebuilt game folder must not lose — the ulimit -c 0 in the launcher, and
./hookshot --prepare-shaders run once.

Linux builder parts

Linux builder parts Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 12 Nov 05:29

Parts Hookshot's Linux builder downloads. A file's name is its checksum; a part that did not change is not uploaded again. This release is not a version of the game.

Steam Deck build 2026-11-12 - 2x frame rate, boxes gone, stable

Choose a tag to compare

@RufflezAU RufflezAU released this 12 Nov 00:55

Measured on the device: Link's house and a heavy 1437-draw area, same save, 240 s runs.

start of work this build
base fps 11.4-15.7 27.5
frame 82-87 ms 36.0 ms
p99 95 ms 41.5 ms
1% low 10.5 24.1
GPU time/frame 46-52 ms 14.6 ms

With 2x frame generation that is ~55 shown per second.

Fixed in this build

  • The fairy and NPC boxes, the water flicker, objects vanishing - the emulator did not have upstream 758e563's storage-pointer handle guard at all: a texture whose handle is really a storage-buffer address was bound as a texture and sampled garbage. Fragment stages only, as upstream refined it.
  • The GPU idling 30 ms a frame - NotRexed#1993's batch-out-at-an-idle-GPU, with a throttle (RX_VK_BATCH_IDLE_MIN, default 2). Raw it stalled this Deck (35.9 s of no input); throttled it is +20% fps.
  • Textures flickering in and out - the texture cache ran 2000 MB against a ~2760 MB working set, so eviction ran every 10 frames forever. A refault pin plus a 2800 MB budget: continuous evictions -> 10 per 240 s.
  • Foliage bare at mid distance - upstream NotRexed#1338: BC raw colour attachments and OoT's two 64^3 noise volumes, whose dither the foliage LOD crossfade reads.
  • Wrong lighting - upstream NotRexed#1507: AMD's graphics stages run at 64 lanes, and the translator was told 32 everywhere, so the game's fragment ballots and shuffles spanned both halves of a wave.
  • Slow motion below 15 fps - HOS_REALTIME_FLOOR (5..60, default 15).
  • Frame generation - the pacer's frame-time estimate was clamped to 50 ms and ignored frames 100 ms apart, so made frames went out in 9-12 ms bursts. Now 500/250 ms, and the made frame is the residual-ratio interpolation (SAFE=3): per pixel the warp's error against a held frame's error, never a mix of two frames.

Rejected after measurement (do not re-apply)

  • NotRexed#2383 fused FMA - the game never leaves the title. Reverted.
  • A sampler cache by id - the same failure. Reverted.
  • NotRexed#1993 raw (no throttle) - stalls the Deck.
  • SMT off - 35% slower.

Known issues

  • Every couple of seconds, with fog or many objects on screen, the picture glitches for a moment (objects shifting then settling). Being investigated.
  • Motion ghosting from the 2x interpolation (RX_VK_FRAMEGEN_SAFE=1 gives a plain blend, =2 holds each frame, =0 the old flow warp).
  • Frame time is CPU-bound: 19.1 ms emulator replay + 14.6 ms GPU + ~9.7 ms game code, run in series. The next win is overlapping the CPU and GPU.

Files

hookshot_vk.so is the emulator's Vulkan backend from this source (md5 461257b1). The game binary is not distributed - it is the user's own dump.

Full Changelog: deck-2026-11-11...deck-2026-11-12

Steam Deck build 2026-11-11 - 2x frame rate, HUD flicker gone, texture thrash gone

Choose a tag to compare

@RufflezAU RufflezAU released this 11 Nov 11:51

Steam Deck OLED tuning, measured on the device (Link's house and a 1437-draw area, same save, 240-280 s runs).

Results

before after
frame 82-87 ms 48-50 ms
base fps 11.4-15.7 20.2
p99 95 ms 49.5-52.9 ms
1% low 10.5 18.9-20.2
GPU time/frame 46-52 ms 14.8 ms

With 2x frame generation that is ~40 shown per second, evenly paced.

What was wrong, and the fix

The flicker was never a texture bug. The frame-generation pacer laid the made frames out on a frame-time estimate that ignored frames 100 ms or more apart and clamped at 50 ms, so below 10 fps it never learned the rate and at 15 fps it read 20: a set of two went out 9-12 ms apart in bursts. And the interpolator warped HUD icons by a 16-px-block flow field median-smoothed over 48 px, so a 32-px icon inherited the background's motion and the other frame's icon was shown for one made frame. Fixed: the pacer's bounds (500 ms / 250 ms), and high-contrast content now takes the nearer frame and is never warped.

Textures flickering in and out was the texture cache: a 2000 MB budget against a ~2760 MB working set, so eviction ran every 10 frames for ever and any texture reused just past the age floor was evicted before every reuse. Fixed: a refault pin (a texture made again after an eviction is pinned; pins give way beyond RX_VK_STICKY_MB past the budget) and the budget raised to 2800 MB. Evictions went from continuous to 10 in 240 s.

Slow motion below 15 fps was a hard clamp in the real-time step (ratio > 4.0). Now HOS_REALTIME_FLOOR= (5..60, default 15).

Build quality: this build was compiled with no profile at all. It now uses hookshot_pieces.prof + hot_funcs.txt + hot_order.txt and the hot chunk, like the official release.

Ported from upstream

  • NotRexed#1993 (batch out at an idle GPU) is included but off by default (RX_VK_BATCH_IDLE=1 to try): it stalled this Deck - the game read no input for 35.9 s - and upstream merged it without testing on RADV.

Safety

  • RX_VK_COUNT_CAP (16M) and RX_VK_GRID_CAP (1M groups): a misparsed draw count (2432696322 vertices seen) or dispatch grid is logged and skipped instead of walked.

Not fixed yet

Lighting/fog instability: the game's per-frame lighting/fog accumulation is stochastic and the temporal average that hides it on console (the TAAU network) cannot run here, so the low-contrast field is re-randomised every frame and the interpolator then slides it. Diagnosis and the one-run tests are in deck/STATE.md.

Files

hookshot_vk.so is the emulator's Vulkan backend from this source. The game binary is not distributed (it is the user's own dump).

Full Changelog: https://github.com/RufflezAU/Hookshot/commits/deck-2026-11-11