# Splatoon Raiders (`0400380020056000`)

A Switch 2 game: update v65536 over the base game, SDK 21.4.4 built for the Switch 2 (`Ounce`), NVN2 API 231.39. Nintendo EPD's engine (sead / agl / gsys, Phive, Havok), modules `main`, `subsdk0` and `sdk`. **Plays on the Mac** (M4 Pro): the opening, character creation, the first level's beach, inking, grates and the first Salmonids, about 4.5 minutes in the last session with no crash. No IPA has been built.

## What it needs

| Setting | Where | Why |
|---|---|---|
| `mem_alias 0x400000000`, `mem_reserve 0x550000000` | [`game.tsv`](game.tsv) | with the 4.5 GB of an earlier profile the region ran out at the first level: [An SDK 21 Switch 2 game with a 4.5 GB alias region](../../../docs/reference/memory.md#an-sdk-21-switch-2-game-with-a-45-gb-alias-region) |
| `mem_stack 0x10000000`, `app_pool 0x240000000` | `game.tsv` | as Donkey Kong Bananza, the Switch 2's 9 GB |
| `sys_resource 0x2000000` | `game.tsv` | the NPDM declares 32 MB, so virtual address memory is on |
| `pad_layout full` | `game.tsv` | the game asks for most of the controller at once |

Render conditions, which the host evaluates for every game ([Conditional rendering evaluated at replay](../../../docs/reference/rendering.md#conditional-rendering-evaluated-at-replay); until 2026-11-12 the profile key `render_cond`): agl wraps its clears in render conditions; when those clears ran, the lit scene was wiped: the black dialogue scene ([Report tables that dropped entries](../../../docs/reference/rendering.md#report-tables-that-dropped-entries)).

## Files

- Built from a dump folder (`base/`, `updates/65536/`). The display version `v65536` is a placeholder: no control data gave a better one.
- Update RomFS: 46,354 files, 13 GB.
- Shaders: 27,192 in [`shader_map.tsv`](shader_map.tsv), from `Shader/*.Nin_Ounce_NVN2.bfsha.zs`, `Sample.Nin_Ounce_NVN2.release.sarc.zs` and `Lib/agl/*`. A full build takes about 34 minutes from scratch, 13 with the Metal cache. 22,147 pipeline pairs prebuild with none failing, in about 200 s.

## Shaders

27,192 of 27,195 built on chain `9bb5d8678204`. Of the 3 that did not, the one on instruction 0x481 was a false row, gone from the tables since they were split again (resolved, below); the other 2, compute shaders of `agl_technique_pfx` that Metal refused (`bracket nesting level exceeded`), have not been built since 04df08d4 raised Metal's bracket depth (unchecked). `fa072659664b2d38` in `agl_common`, which the game asks for at boot, builds now.

## Rendering

- The first level renders at 1707x960 (R11G11B10F), with what the pass inputs suggest is a spatial anti-aliasing step, and one compute program of agl's upscaler (`c1b8edc09e3313b1`, one texture in, no earlier frame) takes it to 2560x1440; that is copied to the 3840x2160 window. The dialogue scene renders at 2560x1440.
- agl's `agl_ndsr.pck` is the same file in the debug and the release library (header word `0x40100`). Its 20 compute variants read colour that is encoded already and come in twos: no decode at the end, or a half-float sRGB decode clamped to 0-1. All past logs name only the 10 that decode; the first level dispatches `c1b8edc09e3313b1` (1707x960 → 2560x1440) and `30d8ee33de9dee85`. With `DM_UPSCALER=metalfx` its GPU p50 was 8.05 ms against 9.1 ms with the network on the beach (M4 Pro, windows of 300 frames, a loaded Mac; the frame rate stays 60 either way), and the picture matches ([An upscaler's colour read from its code](../../../docs/reference/rendering.md#an-upscalers-colour-read-from-its-code)).
- Ink is a 6400x6400 RG8 map with a Depth32F-S8 (880x880 in the dialogue scene). A stamp writes depth and stencil in its rectangle, then paints with a barrier before each draw, sampling the map; then about 150-285 occlusion queries a frame measure the map, each its own `SetRenderTargets`, uniform updates and report. These drew the costs on this page.
- Frame and GPU time, M4 Pro, windows of 300 frames: the still beach about 15 ms of GPU at 60 fps (150-190 ms, about 6 fps, before the fixes below). The last play session: a median frame of 16.6-16.9 ms; the slowest 5 % at 17.1-17.5 ms in its last stretch and up to 22 ms in the heavier moments; GPU median 10-16 ms, slowest 5 % 13-22 ms. So 60 fps is not held everywhere (current), and that session shared the Mac with a regression run. The ink map's passes are a large part in inked scenes: 0.32 ms each ([A render area for passes on very large targets](../../../docs/reference/rendering.md#a-render-area-for-passes-on-very-large-targets)). Each stamp also took a whole-map snapshot (0.76 ms) and a pass of its own; it now reads its pixel of the map from the tile (resolved, [#1063](https://github.com/seedds/nxnative/issues/1063)): on the route to the first beach, 4.2 → 0.6 snapshots a frame and GPU p50 12.8 → 9.5 ms, M4 Pro. Inked stages are not measured here (the reporter's M2: 22–26 → 13–16 ms with their first patch). A few single frames take 130-535 ms; the worst show little submit time and no new pipelines, and one of 373 ms is nearly all submit time; not traced (open).
- The first boss fight's third stage (resolved, 2026-11-05, 16 GB M2 Air, docked): it measures its ink map with about 1,550 occlusion queries a frame, each after 7 inline updates of the query block and with colour writes masked off (`wm=0xfffffff0`), so those draws read the map directly rather than from the tile. Their 11,010 writes in one submission overflowed the host's list of queued writes, and from then on each query ended its render pass, at 1-2 fps ([the reference entry](../../../docs/reference/rendering.md#a-submission-past-8192-queued-writes-counted-as-writing-everywhere)). With a list that grows as needed (a larger fixed limit in the measured build), the stage holds a median of 17.7-19.9 ms with at most 118 render passes. A later stretch of the same run sat at 32-34 ms (frames 15,300-16,200); where in the game that was is not known (open).
- Focus: the game stops drawing its 3D scene when the window loses focus (10 passes, GPU 2 ms), so a frame rate read then says nothing.

## Dynamic resolution

The game writes 29 `ReportCounter(TIMESTAMP_TOP)` reports a frame and steers its dynamic resolution by the GPU time between them ([issue #1064](https://github.com/seedds/nxnative/issues/1064), M2 Air, 1920x1080 handheld). With one time for the whole submission it read about 0 and kept the 3D view at 1440x810 (render scale 0.75) at 26 ms of GPU a frame; with segment times it scaled between 0.39 and 1.0 of the render-scale size. With the GPU's idle gaps between command buffers counted it read about 16 ms where the GPU worked about 8 and held 0.50–0.57 of full size; without them its width median went to 0.94, and dropped frames from 4 to 35 per 300 ([issue #1290](https://github.com/seedds/nxnative/issues/1290)). Both are the host's default now ([GPU timestamps per submission segment](../../../docs/reference/frame-rate.md#gpu-timestamps-per-submission-segment)); not rerun on main.

## Driving it on the Mac

The work folder's `run/drive.sh <name> <secs> "<sec> <control line>"…` boots in a window with the HUD, raises it and appends control lines (`A~0.1`, `LUp~4`, `shot <path>`, `passdump 1`, `texdump <path with %d>`). With the save at the start of the dialogue, four presses of A get to the beach at about 75 s. `DM_NVN_SHOT=<frames>:<relative path>` takes consecutive frames: the path must have no space, as the driver's environment is split on them.

## Open

- A GPU address fault after 4-6 minutes of the tutorial on an M5, Metal 4, then a frozen picture while the game runs on ([issue #1289](https://github.com/seedds/nxnative/issues/1289)); not seen in the M4 Pro's sessions, the longest about 4.5 minutes. Cause unknown. It is not the cube in a cube array slot below: those 180 binds all come at frame 0. It did not reproduce on the M4 Pro to frame 31,200, played only as far as the ramp and moat wall near the first fort, so the stage the reporter names was not reached. `DM_RES_CHECK=1` to frame 13,500 reported no `released too early` line. The next log from a machine that faults has the `[RED] game shaders …` lines ([Naming the shaders of a GPU fault](../../../docs/reference/rendering.md#naming-the-shaders-of-a-gpu-fault)).
- An IPA: the 16 GB alias region is over iOS's limit until the SDK 21 cuts are ported.
- A crash dump once hung in `flockfile` under `shader_load`, with the log contended.
- `dm/games/profile_guess.py` does not recognise `Generic_21_4_4`. `pl::IsRunningOnOunce` answers false.

## Found on this game

- [GPU timestamps per submission segment](../../../docs/reference/frame-rate.md#gpu-timestamps-per-submission-segment): dynamic resolution that never moved, then held low by the GPU's idle gaps
- [Naming the shaders of a GPU fault](../../../docs/reference/rendering.md#naming-the-shaders-of-a-gpu-fault): added for the address fault above, which names no pass.
- [Compatible views for a mismatched declared texture type](../../../docs/reference/rendering.md#compatible-views-for-a-mismatched-declared-texture-type): resolved, a cube in a cube array slot. `d0ce73e71074df85_b` (NigoProc, layer index a literal 0) read texture ID 428 as `texture type mismatch … declares slot 0 as Metal type 6, … is type 5`, 180 binds at frame 0 and none after. With the view the count is 0. The opening looks the same, though the two runs are not frame-aligned, so no pixel comparison was possible.
- [The split shader tables differ from a game profile's](../../../docs/reference/packing.md#the-split-shader-tables-differ-from-a-game-profiles): NigoProc's `800215a0f8d0c44d` is a mesh shader, and every pack stopped until its row was split again ([issue #1263](https://github.com/seedds/nxnative/issues/1263)).
- [A control section's words read as a BAHS entry](../../../docs/reference/shaders.md#a-control-sections-words-read-as-a-bahs-entry): resolved, 2026-11-10. Three rows of `agl_resource.Nin_Ounce_NVN2.debug.sarc.zs` were no shaders (`cs_other` `68bbe98b3bfd9876` in `agl_common`, `ps_other` `18be578dfdaaffa0` in `agl_technique`, `vs_other` `7a4936fccf82094e` in `agl_technique_pfx`), and every NSP build stopped at the table check after 04df08d4 ([issue #1613](https://github.com/seedds/nxnative/issues/1613)). The tables were split again from the update: those rows and one fact went, the later rows of each package moved down one index, and nothing else changed.
- [Storage words read one by one](../../../docs/reference/shaders.md#storage-words-read-one-by-one): the lit pass on an M1, 70 → 16 ms.
- [A command memory grant over a list whose tail is lent](../../../docs/reference/nvn-objects.md#a-command-memory-grant-over-a-list-whose-tail-is-lent): the game grants more command memory than it writes; this host lends Raiders no command memory, so the bug never showed here.
- [A blocked audio device start freezes the game](../../../docs/reference/audio.md#a-blocked-audio-device-start-freezes-the-game) when the selected HDMI device refuses to play.

- [Goto elimination forcing loops around a whole program](../../../docs/reference/shaders.md#goto-elimination-forcing-loops-around-a-whole-program): the first shader build stalled for 5 h 40.
- [Ampere jump tables read past their end](../../../docs/reference/shaders.md#ampere-jump-tables-read-past-their-end).
- [MUFU.F16](../../../docs/reference/shaders.md#mufuf16) and [The NDSR pack's record start](../../../docs/reference/shaders.md#the-ndsr-packs-record-start): the black 3D scene in the opening.
- [The Switch 2's device values](../../../docs/reference/nvn-objects.md#the-switch-2s-device-values): the NVN versions read from the game's `sdk`.
- [Built-in fonts and the audio device notification of SDK 21](../../../docs/reference/services.md#built-in-fonts-and-the-audio-device-notification-of-sdk-21).
- [An SDK 21 Switch 2 game with a 4.5 GB alias region](../../../docs/reference/memory.md#an-sdk-21-switch-2-game-with-a-45-gb-alias-region).
- [Report tables that dropped entries](../../../docs/reference/rendering.md#report-tables-that-dropped-entries): the black dialogue scene.
- [Query slots running out inside a pass](../../../docs/reference/rendering.md#query-slots-running-out-inside-a-pass).
- [Setting the same targets again continues the pass](../../../docs/reference/rendering.md#setting-the-same-targets-again-continues-the-pass), [Inline updates of a running submission refused uniform shadows](../../../docs/reference/rendering.md#inline-updates-of-a-running-submission-refused-uniform-shadows) and [A render area for passes on very large targets](../../../docs/reference/rendering.md#a-render-area-for-passes-on-very-large-targets): about 6 fps in the first level, and the water's grey blocks on the way.
- [A store only a dispatch's first invocation makes](../../../docs/reference/shaders.md#a-store-only-a-dispatchs-first-invocation-makes): the flickering plants.
- [A submission past 8,192 queued writes counted as writing everywhere](../../../docs/reference/rendering.md#a-submission-past-8192-queued-writes-counted-as-writing-everywhere): the first boss fight's third stage at 1-2 fps.
- [Same-pass feedback read from the tile](../../../docs/reference/rendering.md#same-pass-feedback-read-from-the-tile): each ink stamp's one-draw pass and whole-map snapshot ([#1063](https://github.com/seedds/nxnative/issues/1063)).
- [An upscaler's colour read from its code](../../../docs/reference/rendering.md#an-upscalers-colour-read-from-its-code): the washed-out picture with `DM_UPSCALER=metalfx` ([#1156](https://github.com/seedds/nxnative/issues/1156)).
- [Unused border corrections retain sample registers](../../../docs/reference/rendering.md#unused-border-corrections-retain-sample-registers): after the first cutscene a 236-draw 1920x1080 pass took 20-21 ms of a 31.6 ms frame on an M1 in handheld, its pixel shaders spilling 592-720 bytes a thread for a correction none of its samplers needs; 9.3 ms and 19.8 ms with the corrections taken out at build time (PR #1203). Open: not measured again with the function constant that landed.