Repository navigation
Releases
HeyListen v0.4.3-test1 (test build: AMD on Windows)
Test build for Windows players with an AMD Radeon GPU. Not a release: don't install it over your normal game.
This is v0.4.2's main plus one change (pull request #85, a port of a Hookshot fix): on Windows, AMD GPUs now give back the memory imports the game no longer uses, as NVIDIA GPUs do. Without it, after a few area changes the driver can refuse new imports and parts of the world, like Hyrule Field's ground, stop being drawn until a restart.
How to test
- Download
HeyListen-v0.4.3-test1-Windows-x64.zip, unzip it in a new folder (not your normal HeyListen folder), and build the game with your own copy as usual. - In the new game folder (the one with
image.bin), createhookshot_env.txtwith this line:RX_VK_IMPORT_DIAG=1 - Play 20-30 minutes with many area changes: Kokiri Forest, Hyrule Field, the Market, Kakariko, back to Hyrule Field...
- Check: is Hyrule Field's ground still there after several area changes? Any crash, freeze or black screen? Frame rate normal?
- If it crashes or freezes: add a second line
RX_VK_KEEP_RETIRED=1(the old behaviour) and tell us if that run is fine. - Send
hookshot_log.txt, your GPU model and driver version, on Discord #test-builds or on GitHub pull request #85.
HeyListen v0.4.2
HeyListen v0.4.2
HeyListen turns your own Switch 2 copy of The Legend of Zelda: Ocarina of Time into a native Windows and Linux
game. The release holds only the HeyListen Builder: no game code, data, keys or shaders. You dump your own game as a
.nup, the builder reads it, translates and compiles the game on your PC, prepares its shaders, and starts it. See the
user guide.
🎮 To play, download the HeyListen Builder
🪟 Download the Builder for Windows
🐧 Download the Builder for Linux (AppImage)
Run it, pick your own .nup (Switch 2, version 131072): it builds the game on your PC and starts it. Nothing of the game is in
these files.
Also on RecompLauncher, if you use it: HeyListen is in its library, with the same build done for you
(HeyListen-v0.4.2-Windows-x64.zip below is what the launcher installs; you don't need to download it).
Shaders are prepared again once with this release: the first start takes a few minutes longer, and the first session
builds the driver's pipelines as you play, so it stutters more than usual. The second start is smooth again.
v0.4.2, new since v0.4.1
Features
- perf: the capture copies only the memory the GPU can read (+13% fps on an RTX 3050)
- perf: the hybrid engine's native main loop, physics units and program-pick memo are on by default
- builder: the AMD FSR 4 tick downloads AMD's runtime, checks it and puts it in place, as the launcher does
- crash reports: the crash window's "report" button opens an issue that already holds the report (no paste needed)
Fixes
- crash: the first start no longer uses up to 39 GB of memory while it builds pipelines: the game's two largest shader programs are left to the draws that need them (crashes at the title on 16 GB PCs)
- crash: at Windows' memory limit (RAM plus the page file) the game waits for memory and, if none comes back, stops with an out-of-memory message instead of crashing; the textures give memory back before the limit (#62)
- crash: changing the frame generation multiplier while playing no longer crashes NVIDIA and AMD drivers
- crash: a model's full memory heap is tried in its parent heaps instead of handing the game an empty pointer
- vk: a pipeline is drawn only with the program it was built for: the device lost on AMD RDNA 2/3 and the Steam Deck at the same draw (#14, #51, #63, tested in #65)
- vk: GPU fault on RDNA 4 under Linux in Hyrule Field: buffer views go with the memory they were made on (#61)
- vk: no more dark blocks on the water and floor in Jabu-Jabu and around characters (#16, #59)
- vk: three Vulkan usage errors the validation layer reported: the overlay's functions, 8/16-bit integers on cards without them, overlapping uniform updates (#64)
- vk: fixes from Longshot rc6: AMD and Intel ports, RX 6800 (Navi 21) bounds, the ground texture garbling after minutes in one area
Credits for every port are in the README.
HeyListen v0.4.1
HeyListen v0.4.1
HeyListen turns your own Switch 2 copy of The Legend of Zelda: Ocarina of Time into a native Windows and Linux
game. The release holds only the HeyListen Builder: no game code, data, keys or shaders. You dump your own game as a
.nup, the builder reads it, translates and compiles the game on your PC, prepares its shaders, and starts it. See the
user guide.
🎮 To play, download the HeyListen Builder
🪟 Download the Builder for Windows
🐧 Download the Builder for Linux (AppImage)
Run it, pick your own .nup (Switch 2, version 131072): it builds the game on your PC and starts it. Nothing of the game is in
these files.
Also on RecompLauncher, if you use it: HeyListen is in its library, with the same build done for you
(HeyListen-v0.4.1-Windows-x64.zip below is what the launcher installs; you don't need to download it).
Shaders are prepared again once with this release (the mesh programs are translated differently): the first start
takes a few minutes longer, later starts are as before.
v0.4.1, new since v0.4.0
Features
- builder: the progress bar moves while your
.nupis read (it used to sit at the start for minutes); a RecompLauncher setup also prepares the shaders and pipelines, with their progress, instead of the first start - telemetry: opt-in crash reports and performance statistics, both off by default (Privacy settings in the game and in the launcher); no save, no game data, no account, user name and paths removed
- vk:
RX_VK_MESH_SHADERS=nativeorcomputepicks the mesh path by hand - perf: smaller mesh shaders: big mesh programs keep their memory helpers out of line, constant flag operations fold, each mesh invocation works its buffer ranges out once (-1.1 ms of GPU in Kokiri Forest on an RTX 3050)
- perf: fewer textures held back at area loads, the upload budget counts per thread (#58)
- perf: upstream Hookshot's batch: background workers on 16-thread CPUs, texture spans aged once a frame, fewer copies and write-backs in the replay, the pipeline cache written only when it changed, passes ordered by the game's barriers on NVIDIA
- perf: fewer stalls: flat per-draw caches, only changed vertex streams bound, recording threads wait less on the copy lock (#38 to #42)
- perf: the program cache is packed (6.4 GB became 1.7 GB)
- build: the Vulkan back end can be built with a play profile (#54)
- launcher: HeyListen can also be installed, started and set up from RecompLauncher: the game is built from your own copy with no window, settings are on the game's page, saves are found
- launcher: NVIDIA DLSS and AMD FSR 4 are ticks on the setup screen; NVIDIA's and AMD's files are downloaded from them, checked and put in place for you
- launcher: a
logsfolder keeps the last five runs' logs for the launcher's crash reports
Fixes
- vk: AMD RX 6000 and the Steam Deck on Linux run mesh work on the compute path, the fix for device-lost crashes (#51)
- vk: RX 6700 / 6800 on Windows draw task and mesh work on the compute path (#57)
- vk: AMD lighting: storage buffers at the driver's 16-byte alignment (#45)
- fps: above a 60 fps cap, sounds, moving platforms, the pick-up counter, the time of day, water ripples, the Deku Tree's web and the opening movie's cues keep their speed (#46)
- host: no more quarter-second frames at a cap: no virtual core held while a thread sleeps on a lock (#56)
- vk: a game folder built by an older version no longer aborts in the shader preparation (#48)
- vk: fixes from Longshot: Intel compile faults and Arc B-series device loss, ASTC decoded on the CPU for AMD's integrated graphics (no more GPU resets in Kokiri Forest), the Steam Deck's Game Mode recompiling every pipeline, shader programs over 1 MiB, samplers past the driver's limit, read-backs after a timeout
- linux: the sound output is closed before it opens again (#44); short sleeps wait on the timer (#43)
- vk: a stuck GPU's report says what it was running;
hookshot_env.txttakes the hybrid engine's switches
Credits for every port are in the README.
HeyListen v0.4.0
HeyListen v0.4.0
HeyListen turns your own Switch 2 copy of The Legend of Zelda: Ocarina of Time into a native Windows and Linux
game. The release holds only the HeyListen Builder: no game code, data, keys or shaders. You dump your own game as a
.nup, the builder reads it, translates and compiles the game on your PC, prepares its shaders, and starts it. See the
user guide.
Download
- Windows: HeyListenBuilder.exe, the setup. Run it; it downloads the builder and its tools (checked by SHA-256) into
%LOCALAPPDATA%\HeyListenBuilderand opens it. Already installed: the builder offers the update itself. - Linux:
HeyListen-v0.4.0-x86_64.AppImage(glibc 2.35 and newer; see the Linux section of the user guide). heylisten-setup.txt,heylisten-linux.txt: the lists of parts the setup and the AppImage's update read (you do not
need to download them).SHA256SUMS.txt,SHA256SUMS-linux.txt: checksums.
Shaders are prepared again once with this release. The way prepared shaders are recorded changed (below): the first
start after the update prepares them as a new install does. Later updates start at once.
New in this release
No more outlines and light rectangles around the HUD (issue 16). The game's ASTC textures are re-encoded to BC7 on
your PC to save video memory, and the encoder stored fully transparent texels as white at alpha 1/255: on dark scenes
the whole quad of each HUD sprite (the map, item icons, the title logo) was lifted by a few levels. Alpha 0 and 255 are
now exact; the rectangles are gone, as they were with RX_VK_ASTC_RGBA8=1, at no cost in video memory.
No more one-frame black world in Zora's Domain (issues 3 and 14). When a frame's drawing was recorded on several
threads, a pass that only cleared two of the scene's images was not counted as having drawn them, and another thread
copied stale memory over them: the world went black for one frame, mostly while moving the camera. Measured with the
reporter's save: 24 to 36 black frames in every 10,000 before, none after (also confirmed by eye).
A drawn image is no longer overwritten from memory (issue 3). For textures in the game's virtual memory pools, the
engine did not note which pages back an image the GPU drew, and later uploaded the whole image from memory that never
held it (a black or stale surface for a moment). In Zora's Domain: 103 such re-uploads in a few minutes before, 1 after.
Faster frames on mid-range PCs. A copy step was hashing the game's memory twice for every submission (a regression
since sprint 5): on an RTX 3050 / i5-10400F, +13% in Castle Town and +9.5% in Kokiri Forest. On Linux with AMD
(RADV), the readback wait now covers only the light meter: +26% on an RX 9070 XT (a contributor's pull request, #25).
Updates no longer wait for the shaders. Each prepared program now carries a key of what the driver compiles; after
an update the game starts at once and the programs are checked in the background, and only those that really changed
go to the driver again. (This release itself still prepares once: the record starts with it.)
Skip title screen (Insert menu, Gameplay, off by default): a start goes straight into the most recent save.
exeFS (code) mods: the builder can apply IPS, IPS32 and pchtxt code mods before the translation (its "Code mods"
option); functions a mod patches are left to the translated code, and the log says which.
Intel GPUs (issue 23): Longshot's fixes for Intel's Windows driver: compute programs built without the optimizer
first (the freeze after loading a save on Intel laptops), a crash while saving the pipeline cache, white blocks on
integrated GPUs, a device loss on Arc B-series (Battlemage), and one shader compile at a time in Intel's compiler.
Less memory on 16 GB PCs and the Steam Deck: under 32 GB of RAM the game is told the memory it really maps, as on the
console, and stops growing its caches (1.2 to 2.6 GB less commit, up to 1.1 GB less RAM in Hyrule Field).
Linux: the AppImage now builds the game with the native engine's C++ (it failed with "'cstddef' file not found"
in the test builds); the title movie no longer judders (#22); the Full overlay shows AMD GPUs' load, temperature, power,
clock and VRAM (#24).
Also: the window's downscale averages a larger picture instead of sampling it (from Longshot); the Frame rate limit moved to
an option id of its own; the stall markers name the command a stuck GPU never finished (RX_VK_CHECKPOINTS=1).
AMD FSR 4 as an Upscaler choice (and FSR 4's native anti-aliasing), next to DLSS, DLAA and FSR 3.1. It needs an
RDNA 4 card (RX 9000 series); we had none to try it on, so it ships untested: please report how it looks and runs.
It is not in a game folder unless asked for, and AMD's files are not included. To get it:
- Windows: tick AMD FSR 4 in the Builder before you build (it is off by default), then build again (updating the folder is enough).
- Put AMD's
amd_fidelityfx_loader_vk.dllandamd_fidelityfx_upscaler_vk.dllfrom the
FidelityFX SDK 2.4 release (Kits/FidelityFX/signedbin) in
the game folder or next tobin\HeyListen.exe. On Linux, a Vulkan FSR 4 library (amd_fidelityfx_upscaler_vk.so) beside
the game takes their place; the AppImage is built with FSR 4.
With the tick and without the files, the Upscaler row says the files are missing; without the tick it offers FSR 3.1 only.
The overlay's Scaling line now names the upscaler that draws (it said "bilinear" alone while DLSS drew).
Linux diagnostics (#26): job traces, readback call sites, the overlay's shader and pipeline waits counted apart,
the CPU's temperature in the Full overlay, and the early segments' GPU time in the log.
Performance
In-game frame rates only, from the moment play begins (no title, loading or shader preparation): 2560x1440, 3D scene
at 100 %, no upscaling, no frame generation, no HDR, uncapped, Link idle in each zone loaded from the same saves.
PresentMon over the 40 s after each build's "in play" marker, without its first 3 s; the mean of two runs per zone.
Compared with NX2navi 0.11.0-rc4 and Longshot v3.0.0-rc.4. Full method, the whole-window tables and every caveat:
docs/benchmarks/2026-11-16-v0.4.0-zones.md.
| Machine | Zones | NX2navi fps | NX2navi 1% low | Longshot fps | Longshot 1% low | HeyListen fps | HeyListen 1% low |
|---|---|---|---|---|---|---|---|
| i5-10400F, RTX 3050 8 GB | 17 | 29.7 | 22.6 | 35.6 | 28.1 | 40.8 | 24.4 |
| Ryzen 7 7800X3D, RTX 5070 12 GB | 13 | 68.0 | 35.5 | 86.0 | 51.4 | 96.2 | 43.6 |
HeyListen has the highest average frame rate in 14 of 17 zones on the RTX 3050 and in 10 of 13 on the RTX 5070 (+15 %
and +12 % over Longshot, +37 % and +41 % over NX2navi). Its 1% lows are still below Longshot's (some 100-350 ms frames in
steady play); that is the next thing we work on.
Per zone, RTX 3050 + i5-10400F
| Zone | NX2navi fps | NX2navi 1% low | Longshot fps | Longshot 1% low | HeyListen fps | HeyListen 1% low | HeyListen vs the faster |
|---|---|---|---|---|---|---|---|
| Hyrule Field | 38.2 | 29.1 | 31.6 | 26.0 | 44.8 | 20.5 | +17% |
| Lost Woods | 30.0 | 17.0 | 35.7 | 23.4 | 38.1 | 22.9 | +7% |
| Sacred Forest Meadow | 27.3 | 21.7 | 32.9 | 28.0 | 37.2 | 19.8 | +13% |
| Great Fairy Fountain | 66.2 | 44.0 | 55.3 | 44.2 | 56.4 | 37.7 | -15% |
| Kokiri Forest | 18.8 | 15.5 | 31.7 | 26.4 | 36.7 | 21.6 | +16% |
| Inside the Deku Tree | 39.3 | 29.6 | 38.3 | 30.6 | 43.5 * | 36.4 | +11% |
| Deku Tree meadow | 20.3 | 16.8 | 31.3 | 26.7 | 33.8 | 25.0 | +8% |
| Lake Hylia | 20.6 | 17.6 | 27.9 | 21.0 | 38.2 | 22.0 | +37% |
| Gerudo Valley | 23.3 | 19.1 | 30.3 | 26.2 | 35.2 | 21.2 | +16% |
| Lon Lon Ranch | 27.0 | 21.8 | 30.3 | 18.1 | 40.5 | 15.0 | +33% |
| Market | 25.5 | 20.2 | 33.7 | 25.9 | 41.7 | 25.5 | +24% |
| Temple of Time | 43.3 | 30.3 | 50.7 | 35.2 | 47.8 | 21.9 | -6% |
| Hyrule Castle | 25.9 | 21.0 | 33.8 | 29.6 | 37.8 | 24.7 | +12% |
| Kakariko Village | 23.1 | 18.8 | 35.2 | 28.4 | 44.7 | 22.5 | +27% |
| Graveyard | 22.0 | 18.5 | 29.8 | 24.5 | 31.0 | 14.6 | +4% |
| Death Mountain Trail | 33.8 | 25.8 | 43.8 | 37.8 | 49.7 | 40.3 | +14% |
| Zora's River | 19.9 | 16.6 | 32.4 | 25.9 | 36.4 | 23.7 | +13% |
| Mean of 17 zones | 29.7 | 22.6 | 35.6 | 28.1 | 40.8 | 24.4 |
Per zone, RTX 5070 + Ryzen 7 7800X3D (13 zones; read the note below the table)
| Zone | NX2navi fps | NX2navi 1% low | Longshot fps | Longshot 1% low | HeyListen fps | HeyListen 1% low | HeyListen vs the faster |
|---|---|---|---|---|---|---|---|
| Hyrule Field | 63.1 | 29.4 | 93.7 | 61.9 | 111.9 | 41.1 | +19% |
| Lost Woods | 72.1 | 36.4 | 72.8 | 20.5 | 92.0 | 51.4 | +26% |
| Sacred Forest Meadow | 65.6 | 44.9 | 91.5 | 68.5 | 104.5 | 71.5 | +14% |
| Great Fairy Fountain | 106.9 | 61.7 | 111.3 | 79.9 | 117.7 | 51.5 | +6% |
| Kokiri Forest | 61.2 | 23.7 | 77.8 | 34.0 | 90.5 | 33.8 | +16% |
| Inside the Deku Tree | 83.9 | 43.2 | 108.5 | 86.0 | 122.2 | 83.7 | +13% |
| Deku Tree meadow | 64.6 | 28.7 | 85.8 | 63.0 | 91.5 | 54.8 | +7% |
| Lake Hylia | 43.8 | 22.8 | 38.9 | 0.8 | 73.0 | 31.7 | +67% |
| Gerudo Valley | 72.6 | 44.7 | 85.6 | 64.1 | 93.5 | 35.8 | +9% |
| Lon Lon Ranch | 57.6 * | 21.1 | 84.6 | 51.3 | 88.7 | 23.5 | +5% |
| Market | 51.6 | 23.1 | 77.6 | 46.1 | 76.9 | 30.9 | -1% |
| Temple of Time | 84.9 | 50.0 | 104.8 | 32.3 | 102.5 | 17.5 | -2% |
| Hyrule Castle | 55.9 | 31.8 | 85.0 | 60.1 | 85.0 | 39.6 | 0% |
| Mean of 13 zones | 68.0 | 35.5 | 86.0 | 51.4 | 96.2 | 43.6 |
On this PC the...
HeyListen v0.3.2
HeyListen v0.3.2
HeyListen turns your own Switch 2 copy of The Legend of Zelda: Ocarina of Time into a native Windows game. The release holds
only the HeyListen Builder: no game code, data, keys or shaders. You dump your own game as a .nup, the builder reads it,
translates and compiles the game on your PC, prepares its shaders, and starts it. See the
user guide.
HeyListen is a fork of Hookshot by NotRexed and its contributors (GPL-3.0); all credit
for the recompiler, the builder and the Vulkan back end goes to them. HeyListen is GPL-3.0-or-later; the source is this
repository (the engine under engine/).
Download
- HeyListenBuilder.exe: the setup. Run it; it downloads the builder and its tools (checked by SHA-256) into
%LOCALAPPDATA%\HeyListenBuilderand opens it. heylisten-setup.txt: the list of parts the setup reads (you do not need to download it).SHA256SUMS.txt: checksums of the files above.
New in this release
v0.3.2 fixes crashes when moving between areas, which have been there since v0.3.0. When graphics memory got tight
(more likely at high resolutions and after several areas), a memory-saving step could copy stale GPU data over memory the
game had just reused for its own objects, and the game crashed shortly after, usually while an area loaded or when an
area was entered again. The data is now written back only into memory the game still uses for graphics, and only if the
game has not written it since. Played by hand for 35, 20, 45 and 28 minutes through every area reachable from our save
without a crash, where every earlier session crashed within 3 to 21 minutes.
A first play no longer starts with missing buildings. Until now each area's graphics pipelines were compiled the first
time they were seen, which showed as buildings and ground missing over a white background, low frame rates and stalls of
up to a minute in a new area. The builder now ships a list of 9,854 pipeline states (hashes, formats and draw state only:
no game code or data) recorded over 17 areas, and compiles them once: at the end of the build, or at the first start after
this update, with a progress count (about one to two minutes on a recent PC, longer on a slow one). Measured on a cold
start in Castle Town: draws left out 75,574 -> 332, the longest stall 21 s -> under 8 s. That last stall is not gone: about
70 pipelines of a first visit cannot be known ahead yet and are still compiled when first seen.
GPUs without mesh shaders (GeForce GTX 10 series and older, older Radeon and integrated graphics): three errors in how
the engine used Vulkan on such cards are fixed (shader execution modes declared without checking that the card has them,
start-up programs prepared as if the card had mesh shaders, a missing vertex stream). Current drivers of newer cards
tolerated them; an older one may lose the GPU device instead, as reported after the first cutscene on a GTX 1060. We
could not reproduce that loss on our hardware, so this is a likely fix, not a confirmed one: if the game still closes
there, docs/support/no-mesh-diagnostic.md lists three settings to try and what to send us. On these cards the shader
cache is rebuilt once after the update.
Also fixed: a graphics memory pool destroyed twice by the game could make the engine forget a newer pool made at the same
address.
An interrupted shader preparation no longer leaves the game stuttering (a contributor's pull request, #5). The record
that said "the shaders are prepared" did not know whether the shader cache still held them: a start under other shader
settings, or one closed before its preparation ended, could leave an empty cache that the next start trusted, and every
shader was then translated during play. The next start now sees that the cache lost what was prepared and prepares again.
Linux (new, little tested). This release also carries a Linux AppImage, HeyListen-<version>-x86_64.AppImage
(built by this repository's CI in an Ubuntu 22.04 container, glibc 2.35 and newer), with heylisten-linux.txt, the list its
update reads. It builds the game from your own copy on your PC, like the Windows builder; see the Linux section of the
user guide. It is Hookshot's Linux port, made
buildable again on this release by a contributor's pull request (#4) and given HeyListen's names, folders
(~/.local/share/heylisten, ~/.config/heylisten) and update links. What was and was not tested:
- Tested: the game-free test suite on Ubuntu 26.04 (WSL2); the AppImage built there with the repository's scripts; a
complete player's build from it with a real game copy (32 minutes on 8 cores); the builder window starting and showing
HeyListen's name; an update installed from a test list (parts unchanged, so nothing was downloaded). - WSL has no hardware Vulkan on the test PC (only Mesa's CPU renderer), so the game was not played there. The built game
starts, opens a window named HeyListen, creates the Vulkan device and begins the first start's shader preparation (2,300 of
38,809 shaders in the 100 seconds it was left running, then stopped); the CPU renderer is far too slow to go further, and
it reports some programs as not translated for that renderer, which says nothing about a real GPU. - A contributor reported the Linux build of the
mainbranch working on CachyOS with a Radeon RX 9070 XT. That is not this
release's code, and no one has run this AppImage on a PC with a Linux GPU: we have no native Linux machine. - The Linux update (a new AppImage's parts downloaded from the
linux-partsrelease) has never run against a real release.
The update downloads only the builder's own few MB; the game is built again once.
Known issues
- Above 60 fps, cutscenes and some enemies run too fast: the voice lines of the opening movie start almost together
and the picture goes black for the rest of its time, and a boss's behaviour speeds up. Limiting the frame rate to 60
(the builder's and the in-game settings) avoids it. This is not new in this release. - Memory grows with each area visited (several GB over a long session at 2560x1440) and is only given back when
graphics memory runs short. No crash is known from it since this release's fix; it is being worked on. - The first entry into an area can still stall for a few seconds (see above), and entering the Deku Tree was seen to run
at about 30 fps for a minute on a first visit. - On an older integrated Radeon (Ryzen 5700U) the picture is covered by a regular grid of dark lines. Not fixed.
- A builder that is newer than the latest published release offers to "update" to the older one. Ignore it.
v0.3.1 repaired the installation: the setup downloaded four of the builder's tools from the Hookshot repository, which is
no longer online; they are now served from this repository, byte for byte the same files.
The rest of these notes is v0.3.0's. Everything below is new since v0.2.0. It is a speed release: nothing changes in how the game looks or plays.
Measured on the fast test PC (Ryzen 7 7800X3D + RTX 5070), frame generation off, the frame limit lifted, v0.2.0 and this
release run in turn on the same day:
| Scene | v0.2.0 | this release |
|---|---|---|
| Castle Town, moving (3 runs each) | 77.7 fps | 83.7 fps |
| Kokiri Forest, standing (1 run each) | 80.6 fps | 84.5 fps |
| Hyrule Field, walking (1 run each) | 79.5 fps | 83.1 fps |
One run differs from the next by 1 to 2 fps, and the slowest frames (1% low, about 46 fps in Castle Town) are where they
were. The picture (screenshots against v0.2.0's) and the game's speed were checked in the three scenes.
What changed:
- Faster floating-point code. Straight runs of the game's floating-point instructions now execute as one block on the
processor's SIMD registers, with a single check that falls back to the exact instruction-by-instruction code. The builder
ships an optimization profile recorded on that code. Results are bit-identical to before (2 million random test runs,
0 differences). - The frame's memory copy starts earlier. The pages the game changed are copied ahead, in parallel, while its job
threads wait at the end of the frame; when the frame is submitted only what changed since is copied. - Less waiting between the game and the GPU. A page written late into a texture that is still in use no longer makes
the copy wait for the GPU; finished GPU work is recognized without asking the driver again; the engine no longer repeats
barriers the game already asked for. - Lighter bookkeeping every frame. The mirror syncs, the copy's page stamps, the hazard scan and the set-up of the
parallel recording each do less work, and so do the waits of the game's threads (the switches measured in tickets 056
to 062, now on by default). - A background thread of the game no longer takes a processor from the frame's threads (it runs in short slices).
- Each engine change above has a switch that turns it off (
=0in the game folder'shookshot_env.txt;
docs/tickets/SPRINT-5.mdlists them). The floating-point blocks are chosen when the game is built.
Not measured again for this release: the slow-link test PC (RTX 3050 on PCIe 3.0 x4). Its copy-ahead stays off (it keeps
v0.2.0's chunk-diff syncs); the floating-point blocks and the tickets' switches were checked there when they were built, the
late texture pages and the background thread's slices were not.
For developers: docs/tickets/SPRINT-5.md is the log of this work, with what holds the frame near 85 fps on the test PC
(three chains of 11-12 ms each: the GPU round trip, the game's own job chain, the processors) and the experiments that did
not pay: a copy by written 512-byte granules (in the code, off), a light-meter readback one frame late and twelve job
executors instead of si...
Linux builder parts
Parts HeyListen'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.
HeyListen v0.3.1
HeyListen v0.3.1
HeyListen turns your own Switch 2 copy of The Legend of Zelda: Ocarina of Time into a native Windows game. The release holds
only the HeyListen Builder: no game code, data, keys or shaders. You dump your own game as a .nup, the builder reads it,
translates and compiles the game on your PC, prepares its shaders, and starts it. See the
user guide.
HeyListen is a fork of Hookshot by NotRexed and its contributors (GPL-3.0); all credit
for the recompiler, the builder and the Vulkan back end goes to them. HeyListen is GPL-3.0-or-later; the source is this
repository (the engine under engine/).
Download
- HeyListenBuilder.exe: the setup. Run it; it downloads the builder and its tools (checked by SHA-256) into
%LOCALAPPDATA%\HeyListenBuilderand opens it. heylisten-setup.txt: the list of parts the setup reads (you do not need to download it).SHA256SUMS.txt: checksums of the files above.
New in this release
v0.3.1 repairs the installation. The setup downloaded four of the builder's tools (Python, emsdk, the CPM packages and
wgpu-native) from the Hookshot repository, which is no longer online, so a new install of v0.3.0 stopped with a download
error. Those parts are now served from this repository, byte for byte the same files (same SHA-256), and the release tool
refuses a list that names any other host than this repository and llvm-mingw's official release. The game and the builder
are those of v0.3.0: if v0.3.0 is installed and working, the update downloads only the builder's own few MB and you need
not build again for anything new.
The rest of these notes is v0.3.0's. Everything below is new since v0.2.0. It is a speed release: nothing changes in how the game looks or plays.
Measured on the fast test PC (Ryzen 7 7800X3D + RTX 5070), frame generation off, the frame limit lifted, v0.2.0 and this
release run in turn on the same day:
| Scene | v0.2.0 | this release |
|---|---|---|
| Castle Town, moving (3 runs each) | 77.7 fps | 83.7 fps |
| Kokiri Forest, standing (1 run each) | 80.6 fps | 84.5 fps |
| Hyrule Field, walking (1 run each) | 79.5 fps | 83.1 fps |
One run differs from the next by 1 to 2 fps, and the slowest frames (1% low, about 46 fps in Castle Town) are where they
were. The picture (screenshots against v0.2.0's) and the game's speed were checked in the three scenes.
What changed:
- Faster floating-point code. Straight runs of the game's floating-point instructions now execute as one block on the
processor's SIMD registers, with a single check that falls back to the exact instruction-by-instruction code. The builder
ships an optimization profile recorded on that code. Results are bit-identical to before (2 million random test runs,
0 differences). - The frame's memory copy starts earlier. The pages the game changed are copied ahead, in parallel, while its job
threads wait at the end of the frame; when the frame is submitted only what changed since is copied. - Less waiting between the game and the GPU. A page written late into a texture that is still in use no longer makes
the copy wait for the GPU; finished GPU work is recognized without asking the driver again; the engine no longer repeats
barriers the game already asked for. - Lighter bookkeeping every frame. The mirror syncs, the copy's page stamps, the hazard scan and the set-up of the
parallel recording each do less work, and so do the waits of the game's threads (the switches measured in tickets 056
to 062, now on by default). - A background thread of the game no longer takes a processor from the frame's threads (it runs in short slices).
- Each engine change above has a switch that turns it off (
=0in the game folder'shookshot_env.txt;
docs/tickets/SPRINT-5.mdlists them). The floating-point blocks are chosen when the game is built.
Not measured again for this release: the slow-link test PC (RTX 3050 on PCIe 3.0 x4). Its copy-ahead stays off (it keeps
v0.2.0's chunk-diff syncs); the floating-point blocks and the tickets' switches were checked there when they were built, the
late texture pages and the background thread's slices were not.
For developers: docs/tickets/SPRINT-5.md is the log of this work, with what holds the frame near 85 fps on the test PC
(three chains of 11-12 ms each: the GPU round trip, the game's own job chain, the processors) and the experiments that did
not pay: a copy by written 512-byte granules (in the code, off), a light-meter readback one frame late and twelve job
executors instead of six (patches in docs/research/).
What HeyListen adds to Hookshot
- Frame rate: game speed correct at 60-360 fps (was +12% too fast at 120); engine copies cut (write-backs 67.5 -> 13.9 MB
a frame, mirror syncs as one compute pass); the replay of the game's draws recorded in parallel and overlapping the game's
next frame; floating-point code run as real SIMD, a block at a time, with its own optimization profile; Castle Town
turn/move drops fixed (31-43 -> 51-57 fps at the time). - Slow graphics links (PCIe x4 cards): the CPU records the next frame while the GPU finishes its write-backs, the copy
hashes pages for the syncs, and the syncs send only the 256-byte chunks that changed (+18-21% on the RTX 3050 test PC). - Fewer stutters: pipelines used once are noted and precompiled at the next start; texture descriptor cache; ASTC
transcodes batched, their BC copies cached in system RAM. - Memory: a VRAM breakdown in the log, texture budget from the driver, big images in their own memory, idle mirrors
freed, ASTC textures kept as BC1/BC7 (decoded textures 3.0 -> 1.1 GB); about 2 GB less RAM. - Picture fixes: the Castle Town Market ground while walking (no more flat mint, white or teal ground).
- Builder: the HeyListen Builder builds this fork (its own releases and updates, its own work folder), and copies Ocarina
of Time saves from and to emulators (Lichii, Eden and its forks, NX2navi, or any folder) with a backup first
(upstream Hookshot'soot_saves.py). - An original icon (a blue fairy orb), on the builder, the setup and the game; the game is
HeyListen.exe; the Discord
status shows "Playing HeyListen"; a frame limit row in the menu and an FPS counter (F10).
Known issues
- 120 fps is not reached. The test PC runs the three scenes at about 84 fps on average with drops to about 46; the work
goes on (docs/tickets/SPRINT-5.md). - Hookshot's newest engine work (after v2.0.13) is not in this release. Merged with HeyListen, it is about 29% slower
here for now (a readback wait and thread scheduling changes); it waits on thesync/hookshot-12branch. - Occlusion is off (a fixed value, as in Hookshot): real occlusion queries break the picture; some scenes draw more than
they need. - 12 GB of VRAM recommended. 8 GB cards are untested.
- Tested on NVIDIA RTX only. AMD and Intel cards are untested in HeyListen (see Hookshot's notes for them).
- Frame generation is off by default (turn it on in the Insert menu if you want to try it).
- The first build takes a long time (tens of minutes, then shader preparation); later updates rebuild only what changed.
- The in-game update offer is opt-in (set
HOOKSHOT_UPDATE_CHECK=1in the game folder'shookshot_env.txt); the builder
itself says when a new release is out.
HeyListen v0.3.0
HeyListen v0.3.0
HeyListen turns your own Switch 2 copy of The Legend of Zelda: Ocarina of Time into a native Windows game. The release holds
only the HeyListen Builder: no game code, data, keys or shaders. You dump your own game as a .nup, the builder reads it,
translates and compiles the game on your PC, prepares its shaders, and starts it. See the
user guide.
HeyListen is a fork of Hookshot by NotRexed and its contributors (GPL-3.0); all credit
for the recompiler, the builder and the Vulkan back end goes to them. HeyListen is GPL-3.0-or-later; the source is this
repository (the engine under engine/).
Download
- HeyListenBuilder.exe: the setup. Run it; it downloads the builder and its tools (checked by SHA-256) into
%LOCALAPPDATA%\HeyListenBuilderand opens it. heylisten-setup.txt: the list of parts the setup reads (you do not need to download it).SHA256SUMS.txt: checksums of the files above.
New in this release
Everything below is new since v0.2.0. It is a speed release: nothing changes in how the game looks or plays.
Measured on the fast test PC (Ryzen 7 7800X3D + RTX 5070), frame generation off, the frame limit lifted, v0.2.0 and this
release run in turn on the same day:
| Scene | v0.2.0 | this release |
|---|---|---|
| Castle Town, moving (3 runs each) | 77.7 fps | 83.7 fps |
| Kokiri Forest, standing (1 run each) | 80.6 fps | 84.5 fps |
| Hyrule Field, walking (1 run each) | 79.5 fps | 83.1 fps |
One run differs from the next by 1 to 2 fps, and the slowest frames (1% low, about 46 fps in Castle Town) are where they
were. The picture (screenshots against v0.2.0's) and the game's speed were checked in the three scenes.
What changed:
- Faster floating-point code. Straight runs of the game's floating-point instructions now execute as one block on the
processor's SIMD registers, with a single check that falls back to the exact instruction-by-instruction code. The builder
ships an optimization profile recorded on that code. Results are bit-identical to before (2 million random test runs,
0 differences). - The frame's memory copy starts earlier. The pages the game changed are copied ahead, in parallel, while its job
threads wait at the end of the frame; when the frame is submitted only what changed since is copied. - Less waiting between the game and the GPU. A page written late into a texture that is still in use no longer makes
the copy wait for the GPU; finished GPU work is recognized without asking the driver again; the engine no longer repeats
barriers the game already asked for. - Lighter bookkeeping every frame. The mirror syncs, the copy's page stamps, the hazard scan and the set-up of the
parallel recording each do less work, and so do the waits of the game's threads (the switches measured in tickets 056
to 062, now on by default). - A background thread of the game no longer takes a processor from the frame's threads (it runs in short slices).
- Each engine change above has a switch that turns it off (
=0in the game folder'shookshot_env.txt;
docs/tickets/SPRINT-5.mdlists them). The floating-point blocks are chosen when the game is built.
Not measured again for this release: the slow-link test PC (RTX 3050 on PCIe 3.0 x4). Its copy-ahead stays off (it keeps
v0.2.0's chunk-diff syncs); the floating-point blocks and the tickets' switches were checked there when they were built, the
late texture pages and the background thread's slices were not.
For developers: docs/tickets/SPRINT-5.md is the log of this work, with what holds the frame near 85 fps on the test PC
(three chains of 11-12 ms each: the GPU round trip, the game's own job chain, the processors) and the experiments that did
not pay: a copy by written 512-byte granules (in the code, off), a light-meter readback one frame late and twelve job
executors instead of six (patches in docs/research/).
What HeyListen adds to Hookshot
- Frame rate: game speed correct at 60-360 fps (was +12% too fast at 120); engine copies cut (write-backs 67.5 -> 13.9 MB
a frame, mirror syncs as one compute pass); the replay of the game's draws recorded in parallel and overlapping the game's
next frame; floating-point code run as real SIMD, a block at a time, with its own optimization profile; Castle Town
turn/move drops fixed (31-43 -> 51-57 fps at the time). - Slow graphics links (PCIe x4 cards): the CPU records the next frame while the GPU finishes its write-backs, the copy
hashes pages for the syncs, and the syncs send only the 256-byte chunks that changed (+18-21% on the RTX 3050 test PC). - Fewer stutters: pipelines used once are noted and precompiled at the next start; texture descriptor cache; ASTC
transcodes batched, their BC copies cached in system RAM. - Memory: a VRAM breakdown in the log, texture budget from the driver, big images in their own memory, idle mirrors
freed, ASTC textures kept as BC1/BC7 (decoded textures 3.0 -> 1.1 GB); about 2 GB less RAM. - Picture fixes: the Castle Town Market ground while walking (no more flat mint, white or teal ground).
- Builder: the HeyListen Builder builds this fork (its own releases and updates, its own work folder), and copies Ocarina
of Time saves from and to emulators (Lichii, Eden and its forks, NX2navi, or any folder) with a backup first
(upstream Hookshot'soot_saves.py). - An original icon (a blue fairy orb), on the builder, the setup and the game; the game is
HeyListen.exe; the Discord
status shows "Playing HeyListen"; a frame limit row in the menu and an FPS counter (F10).
Known issues
- 120 fps is not reached. The test PC runs the three scenes at about 84 fps on average with drops to about 46; the work
goes on (docs/tickets/SPRINT-5.md). - Hookshot's newest engine work (after v2.0.13) is not in this release. Merged with HeyListen, it is about 29% slower
here for now (a readback wait and thread scheduling changes); it waits on thesync/hookshot-12branch. - Occlusion is off (a fixed value, as in Hookshot): real occlusion queries break the picture; some scenes draw more than
they need. - 12 GB of VRAM recommended. 8 GB cards are untested.
- Tested on NVIDIA RTX only. AMD and Intel cards are untested in HeyListen (see Hookshot's notes for them).
- Frame generation is off by default (turn it on in the Insert menu if you want to try it).
- The first build takes a long time (tens of minutes, then shader preparation); later updates rebuild only what changed.
- The in-game update offer is opt-in (set
HOOKSHOT_UPDATE_CHECK=1in the game folder'shookshot_env.txt); the builder
itself says when a new release is out.
HeyListen v0.2.0
HeyListen v0.2.0
HeyListen turns your own Switch 2 copy of The Legend of Zelda: Ocarina of Time into a native Windows game. The release holds
only the HeyListen Builder: no game code, data, keys or shaders. You dump your own game as a .nup, the builder reads it,
translates and compiles the game on your PC, prepares its shaders, and starts it. See the
user guide.
HeyListen is a fork of Hookshot by NotRexed and its contributors (GPL-3.0); all credit
for the recompiler, the builder and the Vulkan back end goes to them. HeyListen is GPL-3.0-or-later; the source is this
repository (the engine under engine/).
Download
- HeyListenBuilder.exe: the setup. Run it; it downloads the builder and its tools (checked by SHA-256) into
%LOCALAPPDATA%\HeyListenBuilderand opens it. heylisten-setup.txt: the list of parts the setup reads (you do not need to download it).SHA256SUMS.txt: checksums of the files above.
New in this release
Everything below is new since v0.1.0 (v0.2.0 and v0.2.1 were built but never published). Figures are from the two test PCs:
a Ryzen 7 7800X3D + RTX 5070 on PCIe 5.0 x16, and an i5-10400F + RTX 3050 on PCIe 3.0 x4.
- Much faster on PCs with a slow graphics link (the RTX 3050 test PC, PCIe 3.0 x4):
- The CPU now records the next frame while the GPU finishes its write-backs, instead of waiting for them (
RX_VK_LATE_DEFER). - The capture copy also hashes the pages for the mirror syncs, so the syncs no longer hash them again (
RX_VK_CAPTURE_HASH). - Mirror syncs copy only the 256-byte chunks that changed (32 -> 4.4 MB a frame).
- These are on by default where the link is 8 GB/s or less (read from the driver). Together they give +18-21% fps there:
Kokiri Forest 30.9 -> 37.2, Castle Town 29.1 -> 34.4, Hyrule Field 31.4 -> 37.2 fps at 1440p. Fast links keep the old
copies and the same speed.
- The CPU now records the next frame while the GPU finishes its write-backs, instead of waiting for them (
- Faster frames on every PC:
- The game's frame and the replay of its draws overlap better: the replay is held and recorded in parallel. That gives
+1.9 to +3.8 fps in Kokiri Forest, Castle Town and Hyrule Field on the 7800X3D. - The capture copies use streaming stores (+2.4 to +3.3 fps).
- The game's frame and the replay of its draws overlap better: the replay is held and recorded in parallel. That gives
- Faster floating-point code:
- The game's vector math runs as real SIMD instead of lane by lane, and its slow paths (quiet NaN lanes, estimates) are cut.
- The builder ships an optimization profile recorded on that code.
- Market about 75 -> 78 fps standing, 68 -> 73 walking. Results are bit-identical to before.
- Fewer stutters on mid-range PCs: ASTC texture transcodes are batched and their BC copies are cached in system RAM.
- Castle Town Market ground fixed while walking (no more flat mint, white or teal ground):
- ASTC textures are kept as BC1/BC7 (decoded textures 3.0 -> 1.1 GB, VRAM peak -14%).
- A pool-mirror ordering bug is fixed.
- Less RAM: about 2 GB less (compute pipelines are built on first use and their modules dropped); shaders start faster.
- Hookshot v2.0.12/v2.0.13:
- The new Dear ImGui menu (Insert), keyboard shortcut choice, capture and screenshot fixes, and builder hardening.
- HeyListen keeps its frame limit row and FPS counter (F10).
- Discord: the status now shows "Playing HeyListen" with its own icon (the Insert menu's "Discord status" row; it does
nothing without a Discord client running). - For developers:
- Many measured performance switches, off by default, each with its ticket and benchmark in
docs/. - Frame tracing tools (job trace, critical path, frame breakdown) and an analysis of where frame time can still be saved
(docs/research/suspects/).
- Many measured performance switches, off by default, each with its ticket and benchmark in
What HeyListen adds to Hookshot
- Frame rate: game speed correct at 60-360 fps (was +12% too fast at 120); engine copies cut (write-backs 67.5 -> 13.9 MB
a frame, mirror syncs as one compute pass); Hyrule Field about 68 -> 74 fps on the test PC; Castle Town turn/move drops
fixed (31-43 -> 51-57 fps). - Fewer stutters: pipelines used once are noted and precompiled at the next start; texture descriptor cache.
- VRAM: a VRAM breakdown in the log, texture budget from the driver, big images in their own memory, idle mirrors freed
(peak use about 88% -> 81-87%); the teal ground regression fixed. - Builder: the HeyListen Builder builds this fork (its own releases and updates, its own work folder), and copies Ocarina
of Time saves from and to emulators (Lichii, Eden and its forks, NX2navi, or any folder) with a backup first
(upstream Hookshot'soot_saves.py). - An original icon (a blue fairy orb), on the builder, the setup and the game; the game is
HeyListen.exe(a folder built
with an earlier build hadhookshot.exe: rebuilding it replaces that).
Known issues
- Hookshot's newest engine work (after v2.0.13) is not in this release. Merged with HeyListen, it is about 29% slower
here for now (a readback wait and thread scheduling changes); it waits on thesync/hookshot-12branch. - Occlusion is off (a fixed value, as in Hookshot): real occlusion queries break the picture; some scenes draw more than
they need. - 12 GB of VRAM recommended. 8 GB cards are untested.
- Tested on NVIDIA RTX only. AMD and Intel cards are untested in HeyListen (see Hookshot's notes for them).
- Frame generation is off by default (turn it on in the Insert menu if you want to try it).
- The first build takes a long time (tens of minutes, then shader preparation); later updates rebuild only what changed.
- The in-game update offer is opt-in (set
HOOKSHOT_UPDATE_CHECK=1in the game folder'shookshot_env.txt); the builder
itself says when a new release is out.
HeyListen v0.1.0
HeyListen v0.1.0
HeyListen turns your own Switch 2 copy of The Legend of Zelda: Ocarina of Time into a native Windows game. The release holds
only the HeyListen Builder: no game code, data, keys or shaders. You dump your own game as a .nup, the builder reads it,
translates and compiles the game on your PC, prepares its shaders, and starts it. See the
user guide.
HeyListen is a fork of Hookshot by NotRexed and its contributors (GPL-3.0); all credit
for the recompiler, the builder and the Vulkan back end goes to them. HeyListen is GPL-3.0-or-later; the source is this
repository (the engine under engine/).
Download
- HeyListenBuilder.exe: the setup. Run it; it downloads the builder and its tools (checked by SHA-256) into
%LOCALAPPDATA%\HeyListenBuilderand opens it. heylisten-setup.txt: the list of parts the setup reads (you do not need to download it).SHA256SUMS.txt: checksums of the files above.
What HeyListen adds to Hookshot
- Frame rate: game speed correct at 60-360 fps (was +12% too fast at 120); engine copies cut (write-backs 67.5 -> 13.9 MB
a frame, mirror syncs as one compute pass); Hyrule Field about 68 -> 74 fps on the test PC; Castle Town turn/move drops
fixed (31-43 -> 51-57 fps). - Fewer stutters: pipelines used once are noted and precompiled at the next start; texture descriptor cache.
- VRAM: a VRAM breakdown in the log, texture budget from the driver, big images in their own memory, idle mirrors freed
(peak use about 88% -> 81-87%); the teal ground regression fixed. - Builder: the HeyListen Builder builds this fork (its own releases and updates, its own work folder), and copies Ocarina
of Time saves from and to emulators (Lichii, Eden and its forks, NX2navi, or any folder) with a backup first
(upstream Hookshot'soot_saves.py). - An original icon (a blue fairy orb), on the builder, the setup and the game; the game is
HeyListen.exe(a folder built
with an earlier build hadhookshot.exe: rebuilding it replaces that).
Known issues
- Occlusion is off (a fixed value, as in Hookshot): real occlusion queries break the picture; some scenes draw more than
they need. - Missing ground textures while walking in Castle Town Market (flat mint/white or teal ground; buildings and people are fine): VRAM runs short there. A fix (smaller texture formats for the game's ASTC textures) is planned. Standing still renders correctly.
- 12 GB of VRAM recommended. 8 GB cards may show the issue above more often.
- Tested on NVIDIA RTX only. AMD and Intel cards are untested in HeyListen (see Hookshot's notes for them).
- Frame generation is off by default (turn it on in the Insert menu if you want to try it).
- The first build takes a long time (tens of minutes, then shader preparation); later updates rebuild only what changed.
- The in-game update offer is opt-in (set
HOOKSHOT_UPDATE_CHECK=1in the game folder'shookshot_env.txt); the builder
itself says when a new release is out.