iHaiDeeZ--SSBU/docs/PLAN.md

SSBU on PS4: plan

The base is WindowSmash (SNO2207): a static AArch64 → C recompiler for Super Smash Bros. Ultimate plus a native runtime (Maxwell GPU command streams and shaders, nn services, audio, VP9, netplay). Its Linux build already runs the same runtime on POSIX through a Win32 shim (runtime/linux) and draws its Direct3D 11 renderer through DXVK on Vulkan, compiling shaders with vkd3d-shader. This port does the same on the PS4, with SMW's PS4 stack underneath:

Layer	Windows (upstream)	Linux (upstream)	PS4 (this port)
Game code	clang-cl, AVX2/FMA	clang, AVX2/FMA	clang 21, -march=btver2 (Jaguar: AVX, F16C, BMI1; no AVX2, no FMA)
OS layer	Win32	runtime/linux/win32.c	that shim on the OpenOrbis libc + SMW's port/ (heap, threads/cores, lazy direct memory, crash log)
Graphics	D3D11	DXVK 3.1.1 native → Vulkan	DXVK built for PS4 → SMW's RADV (Crema's GNM build), WSI = SMW's ps4_vk_wsi (headless surface → VideoOut)
Shaders	d3dcompiler (DXBC)	vkd3d-shader (HLSL → DXBC)	vkd3d-shader built for PS4, plus a DXBC cache made on the PC
Audio	WASAPI	SDL2	sceAudioOut
Pads	XInput / SDL3	SDL2	scePad (up to 4 DS4)
Package	ssbu.exe	ssbu	eboot.bin in SSBU-vXX.YY.pkg, title SSBU00001, data in /data/SSBU

Milestones

1. Graphics proof (no game needed): done in v00.02. DXVK on RADV on the console: a D3D11 test app (tests/dxvk-test) clears the screen and draws a spinning triangle at 59.9 fps with GPU-side VideoOut flips. Lessons: the system modules must be loaded before any call into them (port/ps4_modules.cpp; v00.00/v00.01 died silently in sceVideoOutOpen), and only binary semaphores are dropped from submits (DXVK's timelines work).
2. Runtime on PS4: the Win32 shim on the OpenOrbis libc (futex → address waits, write watch by mprotect on 16 KiB pages as on macOS, fibers, files under /data/SSBU), the runtime linked with stub modules.
3. Game build (needs the user's SSBU dump: game + update, prod.keys/title.keys, firmware): upstream's recomp.py all --generate-only, then the generated C compiled for PS4. Fighter NROs: static link first; if the eboot is too big, prx modules loaded on demand.
4. First boot, then memory (≈5 GB on the PS4 vs upstream's 16 GB recommendation), speed, audio, pads.

Known risks

• CPU: Jaguar at 1.6 GHz; the game is locked to 60 fps logic. The recompiled code, the Maxwell command-stream translation and DXVK all run on it. NEON FMLA is compiled to fmaf(), which is software on Jaguar: needs a fast path (float FMA through double is exact for the product).
• Memory: upstream recommends 16 GB; the Linux build is a 1.2 GB executable before the game's own memory.
• Shaders: ~40,000 programs for 13.0.5; HLSL compiles on Jaguar are slow, so a PC-made cache is needed.
• Size: game data ~17 GB in /data/SSBU (RomFS read in place).