Yaba Sanshiro — libretro core for low-end ARM handhelds

This is a fork. It forks sydarn/yabause (a fork of Yabause / Yaba Sanshiro by devmiyax), pinned at a40dace1 (devmiyax B2_1_11, standalone 1.11.0 lineage). Branch h700 is the only maintained branch here; upstream development on this line has stopped. The Sega Titan Video (ST-V) support was developed on a stv branch and merged into h700 in aa20951b (2026-10-02); that branch was then deleted from the remote and is kept locally only.

It targets low-end ARM handhelds — verified on the Anbernic RGSP (Allwinner H700, 4× Cortex-A53, Mali-G31 / libmali, glibc 2.35) with the NextUI/minarch-gl frontend — and is loaded as yabasanshiro_libretro.so. It plays Sega Saturn discs and Sega Titan Video (ST-V) arcade ROMs (the latter ported from the kronos branch of libretro/yabause, see Sega Titan Video (ST-V) below).

What this branch adds

Make the abandoned port work, and make it fast (a40dace1..h700):

• 0be6fb9e — make the libretro port in this pin build again: libchdr wiring, vdp1/vdp2 compiled as C++, a GLES compatibility header, YabNanosleep, …
• 258a192a — run the async VDP thread through a core-owned shared EGL context: eglCreateContext(share = frontend context), and the engine draws through a mirror FBO whose colour attachment is the frontend's texture. VDP rendering moves to a worker thread while video_cb() stays on the frontend's thread. Without it the port is synchronous and pinned to one CPU core; with it the emulation thread drops from ~1.0 to ~0.7 cores and the frame rate gains 3–9 % depending on the game (see Measured).
• b2bb246e — hand the frontend's GL context back after borrowing the shared one (SDL tracks "which context is current" per thread in TLS; bypassing it left the frontend without a context).

GL / correctness on this driver (textures are shared between the two contexts, FBOs are not):

• 074c4034 — give the core-owned mirror FBO the depth and stencil the engine's VDP2 compositor needs: it clears depth to 0 and draws every layer with GL_GEQUAL + per-layer z, and implements VDP2 windows through the stencil buffer. Without them a windowed background layer covers the whole screen (Demon Castle Dracula X: walls and stairs vanish).
• 9e9593f6 — re-assert the mirror's attachments on every hand-back, and re-read the frontend's colour attachment every frame (a frontend may recreate its FBO under the same name).
• caf6c0f2 — size that attachment to the render area, not to the frontend's power-of-two texture. A framebuffer is as large as the intersection of its attachments and the engine's per-frame clear is not scissored, so this removes a 4 MB depth+stencil clear per frame: the depth/stencil work is now free (back to the pre-fix frame rate).
• 26364f71 — resolve the ES 3.1 entry points (glBindImageTexture/glDispatchCompute/ glMemoryBarrier) that the RBG compute path calls; without them Virtua Fighter 2's hi-res title screen crashed the core about 9 s in (PC=0).

libretro feature parity with the reference core (the yabasanshiro branch of libretro/yabause, tip 09ed8e5b):

• 8d8cab15, 247ee0f4, ea028ee4 — savestates, plus SAVE_RAM/SYSTEM_RAM and the Saturn memory map. ea028ee4 also fixes silent audio when loading a state on a cold start: the SCSP chunk never carried the sound CPU's registers (the Musashi hooks were empty) and restoring SP before SR left the CPU in the wrong mode. Savestate SCSP chunk version 4 → 5; a core older than this cannot read new states (re-save if you switch back).
• 5e408a8b — resolution_mode = original maps to RES_NATIVE (draw straight into the framebuffer) and Resize(..., ORIGINAL), like the reference core, whose comment notes the upscale path is broken under a libretro HW context.
• c42b9781 — core options v2 (categories, submenus, visibility, translations), libretro.h up to the v2 option commands, yabasanshiro_libretro.info.
• 50e91159 — the disk control interface (v1 + extended) and .m3u playlists, so multi-disc games swap discs from the frontend menu.
• 610bb618 — the reference core's low-risk robustness fixes (error.c allocation, the thread CPU-index mutex, BackupManager flush, HLE BIOS default Slave-CPU proc, debug.c log mutex).
• 6cf146a3 (restored by 843509a2) — align the two rendering differences with the reference core: the VDP1 shadow-discard threshold and the RGB shader bit ops. An earlier revert (cfcba53f) had blamed the threshold for SotN's missing walls; that premise was disproven — the walls were lost to the missing depth/stencil on the async mirror FBO (074c4034).

SCSP threading (68821ac4, 8fd15b47, b000abc0) — the async sound CPU thread used ctime() as a struct timespec *, so its timed wait returned EINVAL immediately and the thread span at ~1.09 cores; fixing the clock source restored the on-demand synchronisation path (Golden Axe: 2.57 → 1.87 cores, +2 % frames). The pending main-CPU interrupt word is atomic now. An optional realtime priority switch exists (YAB_SCSP_RT_PRIO/YAB_SCSP_RT_NICE, off by default: it measured neutral on this device).

Two build switches, both off by default, both required for the async build:

Define	Meaning
YAB_ASYNC_RENDERING	upstream's VDP render thread (its #define in vdp1.h is commented out); on by default in this branch
YAB_CORE_SHARED_CONTEXT	the core-owned shared context added here (no-op when undefined); on by default in this branch

Both are added by yabause/src/libretro/Makefile.common now, so a plain make platform=arm64 gives the shipping asynchronous core. Build the old synchronous one with make platform=arm64 YAB_SYNC=1 (kept for A/B comparisons -- it is how the ~+20 % below was measured).

Sega Titan Video (ST-V)

Ported from the kronos branch of libretro/yabause (tip 3791ffb, "Merge pull request #328 from WizzardSK/aarch64-kronos"; that branch is in turn the libretro port of FCare/Kronos). Both are GPL-2.0, like this tree. What came over:

• stv.c/stv.h — the ST-V BIOS table (BiosList: file name + CRC32 + region) and the game table (GameList, 103 titles), zip ROM/BIOS loading, the 48 MiB ROM-board image, the NV-RAM template.
• eeprom.c/eeprom.h and the SMPC bit-bang that drives it (PDR1 0x3f write, DO bit read back through PDR2) — without it the ST-V BIOS spins forever waiting for the EEPROM.
• decrypt.c/decrypt.h, the cs1 decryption channel in cs0.c, and junzip.c/junzip.h for reading game ROMs and stvbios.zip straight out of the zip.
• cs0.c's CART_ROMSTV cartridge type (48 MiB) and the memory.c DMA read path the decryption needs.
• The IOGA / JAMMA cabinet input: peripheral.c keeps IOPORT/PerCabAdd and the games read it at 0x04000000 (0x01 P1, 0x03 P2, 0x05 system = coin/test/service/START); libretro.c binds the frontend's joypad to the PERJAMMA_* keys.
• The memory.c/smpc.c/scu.c/yabause.c alignments the port needs (yabsys.isSTV gating of the Saturn-only init paths, the OREG[31] handshake, the PDR2 sound-CPU wire) and the libretro hooks: stv_mode detected from the zip entry names + CRC32, stv_favorite_region, the ST-V key descriptors.

Two deliberate differences from Kronos, both measured:

• The backup-RAM card layout stays the classic yabause one (header[32] = FF 'B' FF 'a' …, card byte N at offset 2N+1). Kronos' dense layout made every <rom>.sav written by an earlier build unreadable, and a real BIOS then answers "The System Memory is not ready for use. Please clear all files using Sega Saturn's Memory Manager." (Guardian Heroes, Castlevania SotN). 6407c37d reverts just that group; the extend_backup value and the real word/long handlers from the same commit are kept.
• SmpcSetTiming() gives the INTBACK case a timing for every IREG[0] (56753e92). The old code handled only IREG[0] == 0x01/0x00; cotton2 asks with 0xFF/0x80, so SmpcExec() never ran the command, SR stayed 0 and the game never read the pad. That is why the picture-good line never saw coin or START (the only build that did was the full Kronos SMPC port, which came with a black picture and a deadlock at the resolution switch). Coin, START and the in-game 1P start were verified on this tree afterwards.

To run an ST-V game, put stvbios.zip in the frontend's system directory (next to saturn_bios.bin) and the game zips in the ROM directory. The board is recognised from the zip entry names + CRC32, so a re-zipped or renamed stvbios.zip silently falls back to Saturn mode. yabasanshiro_libretro.info declares that file the way Kronos' .info does (firmware1, optional, because Saturn titles do not need it), so RetroArch's Information → Firmware list names it too.

The core advertises block_extract = true, so the frontend has to hand the romset zip over unextracted — the core opens it itself (RetroArch would otherwise extract a member to /tmp/cache/… and the board detection would never run). A side effect is that Saturn CD images must be uncompressed (.cue/.bin/.iso/.chd); a zipped disc is not something this core can mount.

The ST-V SRAM is per game: <save>/stv/<game>.ram, next to the <game>.nv EEPROM, matching Kronos' naming (the layout is still our classic 64 KB card image, not Kronos' dense 32 KB one). A Saturn keeps one shared internal memory card, as on hardware. Making that file actually appear re-exposed a latent bug: bupfilename was assigned only on the extend_backup path, so with the extend_backup = 0 the ST-V credit fix needs, YabauseDeInit() called T123Save(..., NULL) and the core never persisted its card at all — only the frontend's per-ROM .sav carried it.

The probes and recorders that the porting effort needed (SMPC/IOGA loggers, black-box recorders, per-frame stat dumps, the autopress.txt hook) were removed in 09b9f8b8: the shipping core runs no test instrumentation.

Building

cd yabause/src/libretro

# asynchronous build (the default now: Makefile.common adds both defines)
make platform=arm64

# synchronous build (A/B reference)
make platform=arm64 YAB_SYNC=1

(Before this branch made them the default you had to pass them yourself, and only in the environment -- a CFLAGS= on the command line overrides the makefile's own CFLAGS += $(FLAGS) and would drop the platform flags (-DAARCH64, -D_OGLES3_, -I…, -D__LIBRETRO__, -std=gnu99, …).)

platform=arm64 selects -DAARCH64 -march=armv8-a+crc+fp+simd -mcpu=cortex-a53, FORCE_GLES=1, -D_OGLES3_ -DHAVE_OPENGLES -DHAVE_OPENGLES3 and HAVE_GLSYM_PRIVATE. Output: yabause/src/libretro/yabasanshiro_libretro.so.

Cross-compiling for the device, two requirements are easy to miss:

• Link libmali explicitly (-l:libmali.so.0 in LDFLAGS): libGLESv2.so.2 on the device is a stub that exports no GL entry points; the real symbols live in libmali.so.0.
• Add -ldl when the sysroot is glibc < 2.34 (for example the gcc 8.3 / glibc 2.28 toolchain used to build the NextUI cores): the shared-context path uses dlopen/dlsym and libdl is still a separate library there. Newer sysroots merge it into libc, so the same source can build locally without it.

The built file must keep an _ in its name when you drop it into a frontend: minarch derives the per-core configuration directory from the part before the first _ (and segfaults without one).

Notes for frontends

• The core renders into a core-owned mirror FBO over the colour texture the frontend hands it (FBOs are not shared between the two EGL contexts on this driver, textures are). It attaches its own depth+stencil, sized to the render area.
• It asks for depth and stencil in retro_hw_render_callback; a frontend that builds the per-slot FBOs itself should attach GL_DEPTH24_STENCIL8 (RetroArch's gl3 and NextUI's minarch-gl do).
• If the frontend rotates hw-render FBOs between slots (RetroArch's gfx/video_thread_hw.c ring, minarch-gl's three-slot ring), the core follows it: it re-reads hw_render.get_current_framebuffer() on every retro_run and re-points its mirror at the new texture. A frontend must not rotate slots for a core that caches the framebuffer once at GLSM_CTL_STATE_SETUP — glsm-based cores (flycast, mupen64plus_next) do exactly that and would render into a slot nobody presents.
• video_cb(RETRO_HW_FRAME_BUFFER_VALID) is issued on the frontend's thread; the render worker only glFinish()es and marks the frame, so the frontend keeps its own context for its shader chain and swap.

Measured

Anbernic RGSP (Allwinner H700, Mali-G31), minarch-gl hardware render, CPU governor pinned to performance (1512 MHz on all four cores), 4000-frame windows, frameskip disabled, debug HUD off. Values are fps at frame 3600 / average over the window; the control is the same tree built synchronously (no YAB_ASYNC_RENDERING/YAB_CORE_SHARED_CONTEXT):

Build	Golden Axe	Daytona USA
synchronous	50.3 / 54.3	33.6 / 39.1
async + shared context	60.2 / 60.4	40.4 / 43.6

⇒ +19.7 % (Golden Axe) and +20.2 % (Daytona USA) at frame 3600, or +11.2 % / +11.5 % on the window average. Whole-process CPU (all threads) measures 1.41 → 1.71 cores on Golden Axe and ~1.9 on Daytona, and the async build runs one thread more (9 vs 8): the VDP worker now does the rendering the emulation thread used to do inline, so the emulation thread is no longer pinned to a single core. (Re-measured 2026-10-02 on the merged h700; these numbers supersede the earlier 58.4 / 40.3, which came from a different sampling point.) Both windows are no-input runs, i.e. the titles' attract/title screens — treat them as a rendering/threading benchmark, not a gameplay one.

Known costs: the worker performs one glFinish() per frame (a whole-pipeline drain on the render thread, so the frame is complete in the shared texture before the frontend samples it from its own context — this is not the frontend's ring fence, which is a glFenceSync/glClientWaitSync pair guarding a slot against being overwritten while the GPU still reads it, and which measured as never blocking), and the presentation path adds a normalise pass in the frontend.

Upstream, credits and license

• Upstream: sydarn/yabause → devmiyax's Yaba Sanshiro → Yabause. The upstream website (uoyabause.org) is offline.
• ST-V support is ported from the kronos branch of libretro/yabause (tip 3791ffb), which is the libretro port of FCare/Kronos. The stv.c/eeprom.c/ decrypt.c/junzip.c files, the cs0/cs1 cartridge and decryption channels, the IOGA/JAMMA input model and the SMPC/peripheral alignments come from there (GPL-2.0, same as this tree).
• The upstream tree's own documentation is kept as-is: yabause/README, README.LIN, README.QT, README.WIN, README.DC, README.MAC and yabause/doc/.
• Upstream community, not maintained by this fork: Discord · Liberapay. The badges that used to be at the top of this file were removed: they reported the upstream author's Travis / AppVeyor / Snapcraft accounts, and none of those report anything for this branch.
• License: GPL-2.0 (LICENSE, unchanged). Everything added on this branch is GPL-2.0 as well.