AMD Experimental Build 77-2 (a42b4d4)

Experimental AMD build — not the regular nx2 build. This is the second build from hotfix/amd-sparse-lazy-crash (build 94 was the first), at a42b4d4. It exists to test AMD driver crash fixes on Switch 2 titles. The branch is still a parallel line, not a superset: it has everything build 94 had plus the three changes below, and it is missing the regular build's newer fixes, which are listed further down. If you are not on an AMD GPU, use build 98 instead.

What's changed since build 94

Crashes and memory

• Running out of VRAM no longer leaves textures with nothing behind them. When device memory ran out, any image that failed to allocate was turned into a sparse image with nothing bound to it, and AMD's drivers crash when one of those gets sampled: an RX 6600 went down mid-flood, and an RX 7800 XT ended up with 211 of them, 8x8 ones included. Images that don't fit now go to host memory instead — slower, but real memory the GPU can read — and the sparse fallback is kept only for images too large to back at all. Each failed allocation is also counted, and the texture cache answers by keeping more room free: 512 MiB more per event up to 3 GiB, released 128 MiB at a time once 600 frames pass without a new one. In the log it shows as Out of device memory: keeping N MiB more free. (eff0fe3)

Rendering

• Splatoon Raiders: the missing fog was three pixel shaders being thrown out. Three of its pixel shaders used TXD and TXQ forms the Ampere→Maxwell lowering rejected, so their draws were skipped. Both checks were stricter than they needed to be: TXD refused a word-2 bit that its own emit already forwards as Maxwell's NODEP, and TXQ refused query mask 4 (depth), whose Maxwell mask bits are passed through exactly like the other masks. They lower now, so the fog draws. (a42b4d4)

Logging

• GameChat's camera polling no longer floods the log. GameChat asks nx2:camera command 29 for camera state every frame. It fell through the nx2 auto stub, which logged a Critical plus a Warning line on every call — 40,000 calls in one session. Commands 3, 21 and 29 are now answered directly with the same empty success and log once each. Nothing about the game changes; the log file just stops growing. (0b7b562)

What this build does not have

These are already in the regular build 98 and are not here, because this branch is a parallel line:

• AMD Boost (post_shaders/AMDBoost.fx — FSR 1 EASU upscaling, RCAS sharpening, optional denoise) — build 87.
• GPU buffer readback to the CPU, which fixes visible glitches such as missing collision on some objects — build 89.
• Pokopia's name entry (Switch 2 software keyboard) and its Linux Opus bundling (without it the game gets stuck on Ditto) — builds 96 and 98.
• Linux ARM64 builds, Linux .tar.gz packages and the Discord pre-release ping — CI only, builds 89 and later.

Downloads

x86_64 only, same as build 94: Linux AppImage and Windows x64 zip, each plain and PGO.

Windows: extract the zip and run eglantine.exe, or eden-cli.exe from a terminal.
Linux (AppImage): chmod +x it and run it; it bundles its libraries and needs glibc 2.39 or newer (Ubuntu 24.04, Fedora 40, Debian 13, Arch).
The -pgo builds are the same programs compiled with profile-guided optimization (clang plus Eden's profile); prefer them when your CPU supports it. The Windows PGO build is a MinGW build, so it carries its own DLLs and needs no Visual C++ redistributable.

Full diff against the previous build of this line: build-94...build-99

---

Full Changelog: build-96...build-99