71 commits from 474ad59818 to 2cf8dabe6b, triaged rather than taken wholesale.
Declined, resolved to ours:
- AGENTS.md: upstream's AI-agent instructions; we carry our own and do not
want a second, conflicting policy file.
- CI deps bump (setup-node, labeler): both target workflows are absent here,
and the labeler job is gated on the repository being PCSX2/pcsx2.
- KDDockWidgets 2.4.1: two of the six files do not exist here; we already
build 2.4.0 against a 2.3.0 floor, so there is nothing to gain.
- The FullscreenUI Achievements-layout realignment: our section already
carries the same settings, and ours is the branded copy.
- The GS draw/vertex-buffer cluster (7887919e74, b2fa00844e, 99cfbb49c1,
5c611f85e1, 9945046a49, af48193ebb, d88510e3a6, 8c1bb5742e). Our vertex
kick is an ARM64 rewrite of the same hot path -- register-resident cursor,
fused min/max with a rewind watermark, and a scalar cull mirror that
dual-issues against the NEON parse -- so upstream's generic pointer-logic
optimisation is a variant of work already banked here, and their growth
restructure replaces per-buffer capacity with a single global value, which
the pooled draw-node model cannot express. Two of the four August commits
in that cluster repair regressions the July rewrite introduced, and the
third's genuine fix (staging arrays sized from an unrelated buffer) we had
already made independently.
Taken with adjustment:
- EATAN coefficients (aae9438f98). Upstream relabelled mVU_Globals so the
names match the powers; we had fixed the same defect by ordering the arm64
call sites by power instead. Both fixes are correct alone and CANCEL when
combined, so the arm64 call sites move to plain ascending order in the same
commit. The values never moved, so this emits an identical instruction
sequence. Their fix also repairs the x86 mVU we still carry.
- Shader cache version: upstream numbered their tfx.glsl change 109, which is
below our 110. Taking their value would hand every user a stale blob, so
this lands as 111.
- FullscreenUI: took the two readback-spin toggles, placed outside our
non-Apple guard rather than inside upstream's unguarded run.
- Restored tools/generate_fullscreen_ui_translation_strings.py, dropped by
431ca0c063, and regenerated both string areas. That also registers the Big
Picture setup-wizard strings, which had never been extractable.
GameDB: the three serials upstream gave gsHWFixes (SLES-53869, PAPX-90020,
SCPS-15064) are absent from the mobile overlay, so no fix is silently erased
on handhelds.
OboeAudioStream had no synchronisation at all. onError() runs on Oboe's own
callback thread and tears the stream down and back up, while the CPU thread
can be inside SetPaused() on the same object - its null check and dereference
of m_stream are not atomic against onError's reset(), so the stream could be
destroyed between them. Android reclaims an idle low-latency stream a few
seconds into the pause menu, so touching any setting at that moment re-entered
SPU2 and crashed. Fixed with a recursive lock over the lifecycle.
setOutputPauseSuppressed and onNativeSurfaceDestroyed were both fully
implemented, both documented as having callers, and neither was ever called.
Without the first, every overlay pause let Android reclaim the audio device
and the RESUME had to rebuild it inline on the CPU thread ahead of
Host::OnVMResumed - seconds of apparent hang. Without the second, the GS
thread was never told the surface died and could block in
vkAcquireNextImageKHR forever, which is the permanent freeze on backgrounding.
pauseForOverlay now goes through the same executor as resume, so the FIFO
ordering both comments claimed is actually true.
Save states: every refusal was silent. MemcardBusy is decremented only by
VSyncStart, so it is frozen while the pause menu is up and waiting in the menu
can never clear it - only resuming can, which is why it appeared to need two
or three tries. The refusal is correct; the silence was not. Also adds
Achievements hardcore reporting and a per-exit diagnostic line.
Includes Brian Degenhardt's PR #426 (JNI thread ownership).
Audio
- Optional OpenSL ES output backend for devices where the default AAudio path
crackles, glitches or won't initialise (Settings -> Audio), plus a lightweight
SPU2 mode that skips the reverb pipeline to free CPU on low-end devices.
- Keep the audio device alive across the in-game menu pause so Android no longer
reclaims the idle stream and drops sound after the menu sits open (#333).
Settings
- Restored the per-setting descriptions under every GameDB Fix and Advanced
Speedhack toggle (lost in the settings redesign).
- Per-game Reset now clears the native per-game INI, so it truly reverts to the
global values instead of the game keeping stale overrides.
- On-screen display now defaults off; Custom stats appear on boot without a
reset (#385).
Controls / RetroAchievements
- Vibration Strength slider scaling all rumble and touch haptics 0-200%.
- Achievement Sound Volume slider; points now show in the menu before a game
loads; unlock sounds play with Do Not Disturb enabled.
Misc
- Drop the compiled GS shader/pipeline cache automatically on app update to
avoid post-update graphical corruption.
- Animated XMB library-background fallback for GPUs without float-texture
filtering.
GS correctness (ported from sashkinbro/EmuCoreX)
- Reset per-game hardware-hack HLE state on game change (Burnout bloom,
IRem/GT channel-shuffle) so it no longer leaks across in-app game switches.
- Fix a non-strict-weak-ordering comparator in SortMultiStretchRects.
- Free the leaked m_expand_vao on the OpenGL device teardown path.
Merges yaps2/main (github.com/yaps2/yaps2, c16b88cb7) into ARMSX2,
replacing the arm64 recompiler family with the yaps2 JITs and importing
the yaps2 testing, perf, and libretro infrastructure. Common ancestor is
upstream PCSX2 342db5152 (2026-06-19); git auto-merged all but 38 files.
Replaced (deleted in this merge, recoverable from history):
- arm64/aR5900*, aR3000A*, aVU* -> arm64/iR5900*/iR3000A*/microVU*-arm64:
EE static-pin register file with lazy dirty tracking, dual-residence
allocator, IOP block linking, native COP2 macro ops, inline unaligned
fastmem, persisted VU program cache, call-ret shadow ring, VU0 spin
fast-forward.
- MVU_DIFF shadow-run hooks in shared VU interpreter TUs (superseded by
the offline vurunner JIT-vs-interp oracle).
Imported from yaps2:
- tests/ctest/core/recompilers: ~80 gtest suites (EE/IOP/VU differential
harnesses, fuzzers, ABI digest tripwire, capture format pins) plus the
gs_vertex_tests kernel oracle.
- pcsx2-vurunner / pcsx2-eerunner headless capture-replay runners.
- tools/perf counter-based A/B rigs, perf jitdump productionization,
PmuCounters, clang-perf/clang-handheld presets.
- pcsx2-libretro core (ENABLE_LIBRETRO, default OFF; rename pending).
- GS vertex-kick fast path (GV series): TBL-based packed parse,
register-resident kick, scalar-outcode cull, fused draw-rect/FindMinMax.
- Null renderer, VK_KHR_display direct WSI, swapchain PresentStats.
- SPU2 NEON mixer vectorization, EE timer read clamp (NFL 2K5 hang),
IOP ioman signed-compare fix, assorted UB fixes.
Kept from ARMSX2 in the both-touched files:
- iOS dual-map W^X and fastmem-unavailable resilience (Memory, HostSys,
vtlb). The split data/code area model is retained; both areas now take
fixed VA hints so cached VU JIT code stays deterministic on Linux.
- Android thread-affinity model, VMState shutdown early-outs, all
platform frontends, branding, CI, RetroAchievements identity/policy.
- GSDeviceVK: ARMSX2's push-descriptor decision logic (Mali crash gate,
proprietary-vs-turnip Adreno split) merged with yaps2's descriptor-pool
exhaustion recovery (flush + render-pass restart instead of dropped
binds). Vendor feature policy is the union: Mali fbfetch policy with
MediaTek/G57/Xclipse gates from ARMSX2; Adreno stencil/ROV/
test-and-sample-depth hang avoidance and no_ps2_z_quantization from
yaps2.
Build-system notes:
- The Qt debugger is now gated behind ENABLE_QT_DEBUGGER (default off on
arm64) so handheld builds drop the KDDockWidgets dependency.
- GSDeviceNone and remaining yaps2 GS code were ported to the newer
upstream GSTexture Usage-flags API.
The replaced backend's interpreter-fallback glue (intExecuteOneInst,
AndroidEEOpHist) and the EEDiffVerify runtime differ are retained for
now; dead pieces will be removed in a follow-up commit.
Renderer:
- RetroArch (.slangp) shader chains on both Vulkan and OpenGL/GLES via
librashader, hooked into the GS post-process chain after FXAA.
- Fix the Vulkan chain crashing the GS thread. Merge() runs the chain before
VSync decides whether to present, so a skipped present never submits the
command buffer librashader recorded into. Its resource ring keeps advancing
and frees image views the pending buffer still references
(VUID-vkDestroyImageView-imageView-01026). SkipDuplicateFrames defaults on
with a limit of 3, exactly the ring depth, so three skips in a row destroy
views mid-recording. Submit once per chain frame so both rings advance
together.
- librashader is built from source at configure time and is optional: with no
Rust toolchain the feature compiles out. Ship libc++_shared.so alongside it,
which nothing else in the app pulls in.
Shader picker:
- Download shader packs in-app, mirroring the GPU driver manager.
- Group presets by family (Mega Bezel, koko-aio, CRT...), collapsed by default,
each family split by its own subfolders so no group composes thousands of
rows. Show each preset's pass count as information only; it does not predict
cost and is not used to rank.
- Available from Settings and the in-game menu.
Fixes:
- Add to home screen never pinned anything: the launcher rejected the
full-size cover as a shortcut icon, the exception was swallowed, and the
legacy INSTALL_SHORTCUT fallback (ignored since Android 8) reported success
so the failure toast never fired. Scale the icon and report real failures.
- Audio stayed muted after the in-game menu was left open, when Android
reclaimed the stream. Reopen it on resume.
Other:
- About section with Discord, GitHub and website links.
- ForceMaliFramebufferFetch (default off) so MediaTek Mali users can test
whether their driver's ROAA is actually broken, instead of always taking the
per-primitive barrier path.
* Libretro: Scaffold yaps2_libretro core (M1: builds + dlopens)
New pcsx2-libretro/ target producing yaps2_libretro.so (ENABLE_LIBRETRO=ON,
or built on demand). Milestone 1 of the libretro port:
- Full libretro v1 entry-point surface, version-script-restricted to
retro_* exports (PCSX2 internals must not collide with the frontend).
- Host:: implementations adapted from pcsx2-sdl: same CPU-thread state
machine (VMManager::Execute on a dedicated thread, RunOnCPUThread queue),
no windowing/clipboard/file-picker.
- retro_load_game boots the VM headlessly: GS renderer forced to Null,
Surfaceless WindowInfo, null audio; config+data self-contained under
<system dir>/pcsx2 (yaps2-libretro.ini).
- retro_run presents a placeholder XRGB8888 frame; no pacing yet.
- POSITION_INDEPENDENT_CODE forced ON (the pcsx2-sdl ET_EXEC persisted-JIT
trick cannot apply to a shared core).
Verified: builds in org.kde.Sdk 6.10 (gcc), dlopen + retro_api_version +
retro_get_system_info OK.
Next: M2 Vulkan negotiation interface (wrap vkCreateInstance/Device so
GSDeviceVK inits against the frontend-shared device, set_image handoff),
M3 frame pacing + libretro input/audio, M4 savestates/disk control.
* Libretro: M2 Vulkan context sharing + frame handoff (first light)
The lrps2-libretro pattern ported to the modern GSDeviceVK:
- VKLibretro.{h,cpp}: the loader's global vkGetInstanceProcAddr is swapped
for a wrapper that intercepts vkCreateDevice (merges the frontend's
required extensions/layers/features into GS's create info and captures
the shared VkDevice) and vkQueueSubmit (serialises against the frontend
through the HW-render interface queue lock).
- GSDeviceVK: adopts the negotiation-provided VkInstance/VkPhysicalDevice
instead of creating its own (and never destroys the frontend's instance);
the surfaceless BeginPresent branch publishes the merged display texture
(ShaderReadOnly + ExecuteCommandBuffer) into a mutex-guarded slot.
- Main.cpp: RETRO_HW_CONTEXT_VULKAN + context negotiation interface
(create_device opens MTGS from the frontend thread, so GSDeviceVK is
fully constructed before the context reply); the VM boot parks until
context_reset delivers the retro_hw_render_interface_vulkan; retro_run
consumes at most one published frame per call and forwards it via
set_image + video_cb(RETRO_HW_FRAME_BUFFER_VALID), dupe otherwise.
First light verified on Turnip Adreno 618 (and llvmpipe): GT3 boots
(VMManager::Initialize StartupSuccess), frames arrive hw_valid with a
non-black 640x448 readback.
Known gaps for M3: no frame pacing (VM free-runs; YAPS2_RUN_SLEEP=1 paces
headless runs), no libretro input/audio, m_current is sampled by the
frontend while GS may already be rendering the next frame (single
buffered), no OSD.
* Libretro: M3 frame pacing, joypad input, audio
- Pacing: PublishFrame now blocks the GS thread until retro_run consumes
the frame (one presented frame per retro_run; the frontend's cadence is
the emulation's vsync). Enabled at context_reset, aborted before any
path that could otherwise leave the GS thread parked (context_destroy,
retro_unload_game, VKLibretro::Shutdown). Static screens still dupe
thanks to yaps2's SkipDuplicateFrames -- the VM keeps 100% speed and the
frontend gets video_cb(NULL) for unchanged frames.
- Input: retro_run forwards the libretro joypad + both analogs straight
into the DualShock2 bind slots (Pad::GetPad(0)->Set), bypassing
InputManager; SDL input source stays disabled.
- Audio: AudioStream grows a public PullFrames() (frontend-driven pull
from the ring), SPU2 exposes GetOutputStream(), the config pins the
Null backend (mix into the ring, no device thread), and retro_run
drains the ring into audio_batch_cb with float->s16 conversion.
- Config: EmuFolders::Settings is now set explicitly (the libretro path
bypasses SetDataDirectory, so the INI used to land in the cwd).
Verified on Adreno 618 with GT3: internal fps 59.9, speed 100%, frames
publish on change, INI persists under <system>/pcsx2/inis.
* Libretro: Stable backbuffer ring + dynamic PAL av_info
Fixes the RetroArch heap-corruption crash ~2 minutes in (GT3 FMV -> demo
race transition): the published frame was the pooled m_current texture,
which GSDeviceVK recycles while the frontend still samples the view for
cached-frame replays. Frames are now copied into a dedicated ring of
three backbuffer textures owned outside the pool; on a resolution change
the displaced buffer is retired (kept alive until device teardown)
instead of destroyed, because the frontend can replay the old image
indefinitely (e.g. while its menu is open). Verified: headless harness
clean over 4+ minutes at ~100% speed, RetroArch session stable past the
previous crash point.
Also: when the booted VM reports a vertical frequency different from the
NTSC default (PAL 50Hz, progressive), retro_run pushes an updated
retro_system_av_info to the frontend.
* Libretro: M4 save states (retro_serialize/retro_unserialize)
SaveState grows in-memory zip variants sharing the existing disk code:
SaveState_ZipToBuffer writes the ArchiveEntryList into a libzip
buffer source (zip_source_keep + read-back after close), and
SaveState_UnzipFromBuffer opens one over the incoming blob; the whole
entry/version/screenshot pipeline is reused via a shared
SaveState_UnzipFromZip body, so the on-disk and in-memory formats are
identical (a retro state is a valid .p2s payload).
retro_serialize runs SaveState_DownloadState + ZipToBuffer as a blocking
RunOnCPUThread job with frame pacing temporarily disabled -- while the
frontend is inside retro_serialize it is not calling retro_run, so a
parked PublishFrame would deadlock the GS freeze. The fixed
retro_serialize_size bound is 68 MiB (DownloadState's 64 MiB working
buffer + slack); the actual zip length travels as a leading u64 inside
the block.
Verified on GT3 (Adreno 618): serialize 587 ms, unserialize 131 ms,
emulation continues cleanly after the in-place load.
* Libretro: M4 core options + disk control (m3u multi-disc)
Core options (RETRO_ENVIRONMENT_SET_VARIABLES): GS renderer
(Vulkan/Software; the software renderer still presents through the
shared Vulkan context, so the negotiation path is unchanged), internal
resolution 1x-4x (EmuCore/GS upscale_multiplier, live), fast boot and
widescreen patches. Startup values apply before LoadStartupSettings;
later changes re-apply via VMManager::ApplySettings on the CPU thread.
Disk control (SET_DISK_CONTROL_EXT_INTERFACE): .m3u playlists parse
into a disc list (relative entries resolved against the playlist dir),
single-disc content registers as a one-entry list, and closing the tray
swaps via VMManager::ChangeDisc on the CPU thread. m3u added to
valid_extensions.
Regression-tested on GT3: boot, savestate roundtrip and rendering
unchanged.
* Libretro: OSD via real present path into the backbuffer
The surfaceless BeginPresent no longer copies m_current and skips the
frame -- with the libretro context active it begins an actual present
render pass targeting the dedicated backbuffer (clear + viewport/scissor,
same sequence as the swapchain path) and returns PresentResult::OK. The
whole standard presentation pipeline now runs unchanged: PresentRect
draws the display aspect-corrected with TV shaders/linear filtering,
FullscreenUI::Render and ImGuiManager::RenderOSD draw the overlay, and
EndPresent (libretro branch) finishes the backbuffer, submits without
swapchain semantics and publishes the image to the frontend. The old
copy-based publish is gone.
New core option: yaps2_show_fps (EmuCore/GS OsdShowFPS).
Verified on GT3/Adreno 618: readback shows the ImGui FPS counter drawn
over the aspect-corrected frame; boot/savestate regression clean.
* Libretro: Fix retro_serialize bounds check to include the u64 length header
The check compared buffer.size() against the caller's buffer size, but the
write is sizeof(u64) + buffer.size() — a state within 8 bytes of the
reported serialize size would overflow the frontend's buffer.
* Libretro: Size the output canvas to the internal resolution
The present backbuffer was sized from the fixed 640x448 window info, so
the upscale option rendered internally at 2x-4x and then got scaled back
down before the frontend ever saw the frame.
The present path now tracks the merged frame: expand it to the target
aspect ratio (the internal-resolution screenshot rule), clamp it to the
advertised max geometry (2732x2048, 4x PAL at 4:3), and resize the
surfaceless "window" before the draw rect is computed so the whole
frame stays consistent. ResizeWindow learns to adopt a new size with no
swap chain, and retro_run reports geometry changes with SET_GEOMETRY so
the frontend keeps scaling correctly. Resizes are rare in practice (boot,
FMV/interlace switches, option changes) and reuse the existing
retire-don't-destroy backbuffer ring.
* Libretro: Core options v2 + video/performance options
Registers options through SET_CORE_OPTIONS_V2 with Video/Performance/
System categories and per-option help text (legacy SET_VARIABLES kept as
the fallback), and adds:
- Aspect ratio (Auto 4:3/3:2, 4:3, 16:9, Stretch) — the canvas sizing
follows it, and 16:9 pairs with the widescreen patches option
- Deinterlacing mode (Automatic/Off/Weave/Bob/Blend/Adaptive)
- No-interlacing patches (progressive output for supported games)
- Blending accuracy (Minimum-Maximum)
- EE cycle rate (50%-300%) and EE cycle skip speed hacks
* Libretro: Don't double-load the Vulkan library in EnumerateGPUs
The frontend preloads libvulkan for the context negotiation, and
EnumerateGPUs asserted (and would have unloaded the host's library) when
called with no device open — which is exactly the Software renderer
path, via D3D::GetPreferredRenderer. Use the already-loaded library and
leave it loaded.
* Libretro: Second wave of core options
Video: texture filtering, anisotropic filtering, software renderer
threads. Performance: hardware download mode (readbacks are expensive
on tile-based mobile GPUs), MTVU and Instant VU1 toggles. System: BIOS
selection (scanned from <system>/pcsx2/bios at option registration,
auto = first valid image) and cheats (.pnach loading).
* Libretro: Fix crash on content close
Two shutdown bugs, both hit on every quit-from-menu:
- GSDeviceVK::Destroy destroyed the negotiated VkDevice, but that device
belongs to the frontend (it made the vkCreateDevice call) and the
frontend tears it down after context_destroy — RetroArch was left
waiting on and destroying a dead device (freeze, then segfault).
Guard it like the adopted VkInstance already was.
- The frontend replays the last set_image indefinitely (menu background,
duped frames), so retro_unload_game now retracts the image and waits
for the GPU before VM teardown destroys the textures it points at.
Verified: 6/6 clean RetroArch exits after a full Vulkan content run
(was a reliable SIGSEGV on close before).
* Libretro: Reclaim retired presentation backbuffers
A resolution change retires the displaced backbuffer instead of freeing
it, because the frontend may still be replaying its image for a few
cached/duped frames. But the retired list was only cleared at device
teardown, so every interlace<->progressive switch (frequent in FMV-heavy
games) leaked a full-resolution render target for the rest of the
session.
Each retired backbuffer is now tagged with a monotonic present count and
reclaimed once kLibretroRetireFrames (6) presents have gone by --
comfortably beyond any libretro frontend's swapchain depth (2-3), and
GSTextureVK destruction is itself fence-deferred, so the underlying
Vulkan objects aren't freed until the GPU is done with them either.
Verified: GT3 boots/renders through its 640x448<->640x480 interlace
switches with no crash (harness + real RetroArch, clean exit).
* Libretro: Move libretro headers to 3rdparty/libretro
The libretro Vulkan HW-render interface header was pulled in by adding
pcsx2-libretro/ to PCSX2_FLAGS's INTERFACE include path, which leaked
onto every target that consumes PCSX2_FLAGS (and had the GS backend
reaching into the frontend's source dir).
Move libretro.h + libretro_vulkan.h into 3rdparty/libretro/ behind a
header-only INTERFACE library (libretro-headers), and link it PRIVATE
to the two targets that actually need it: PCSX2 (for VKLibretro.cpp) and
the pcsx2-libretro frontend. No other target sees the headers now.
Vulkan render fixes (GS device backends), Oboe audio backend, GameDB armsx2_overrides.yaml override loader, EE+VU mac-port recompiler graft, PGO=optimize, and the VU-slam fix: an Android-only force-float in VMManager::SetEmuThreadAffinities (the slam was thread pinning locking the VU1 worker off the prime core, not the VU codegen). Also: CPU-name SoC-property fallback, Mali VK attachment-feedback-loop crash gate, MediaTek fbfetch disable, and on-device thread-placement diagnostic helpers.
The recompiler grafts and GS deltas are unguarded and land in the shared core (they reach the mac/Linux arm64 builds) - see REFACTOR_STATUS.md items 2 and 3 for the reconciliation this still needs; the VU graft is reducible to just the force-float fix.
Snapshot the refresh-experimental Android app (Gradle + JNI + Android-only
3rdparty) into platforms/android/. Delete its vendored PCSX2 core copy and
relocate the ~24 genuinely Android-specific core additions (Oboe audio, Android
stubs, EGL-Android GL context, NEON SPU2, GSGPUProfile, VU1Fingerprint, Android
HTTP downloader) into the root core, guarded by if(ANDROID) in
pcsx2/CMakeLists.txt and common/CMakeLists.txt.
The superseded arm64 JIT experiment (arm64/mac/* IR-VU backend, split
aVU0/aDMAC/aVTLB/aR5900COP*) is dropped: the root macOS arm64 JIT (127 unique
commits, newer) is the canonical recompiler.
Rewire the Android native build to a thin CMakeLists that sources the root
{common,pcsx2,3rdparty} instead of the deleted vendored copy.
NOT yet compiled against a real NDK -- build validation is CI's job.
See REFACTOR_STATUS.md.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Brian Degenhardt
jpolo1224
jpolo1224
chaoticgd
WizzardSK
David IsztlandClaude Opus 4.8
Ziemas
SternXD
KamFretoZ
Ty Lamontagne
Silent
JordanTheToaster
TheLastRarandKamFretoZ
GovanifY