* shader_recompiler: Lower user clip planes to clip distances
* shader_recompiler: Skip user clip plane lowering when the guest exports clip distances
* vk_pipeline_serialization: Skip preloading permutations with conflicting indices
* shader_recompiler: Emulate lowered user clip planes on NVIDIA
* video_core: bind null clip planes when clipping is disabled
After some changes that happened at some point, logged in users are only properly registered as logged in after libScePad initializes. shadNet initialization relies on logged in users, so this needed fixing.
- Analyze EH function boundaries and relocate faultable guest instructions through red-zone-safe trampolines.
- Integrate Windows exception handling and existing CPU patches without an external debugger.
- Add a Windows-only per-game setting, coverage diagnostics, and configuration tests.
Co-authored-by: w1naenator <valdis.bogdans@hotmail.com>
* add a bunch of signal stuff qemu needs
* fix build
* everyone is doing some bullshit except linux and freebsd
* openorbis strikes again
* better logging
* eintr
* fix sigaltstack nullptr param handling and fix errno-s
* eintr is eintr
* Big Picture: launch games in a clean process
- prevent frontend renderer state from affecting game texture loading
- wait for Big Picture shutdown before initializing the game process
- reuse relaunch handling for IPC and native process creation
- preserve Unicode game paths on Windows
- correct SDL renderer and window destruction order
* Clang format fix
---------
Co-authored-by: w1naenator <valdis.bogdans@hotmail.com>
* UpdatePlayTime now also appends unix epoch after playtime in play_time.txt
* Added explicit include of ctime, cleaned variable initialization
* Fix clang formatting
* input: reduce pad state latency for single-sample reads
* Restore original disconnected pad handling
* input: restore axis smoothing default in FinalizeUpdate
Explicitly passing false to GameController::Axis() unconditionally
disabled smoothing for all input sources. Controller hardware is
unaffected (the <16-unit guard in UpdateAxisSmoothing already skips
interpolation for fine-grained inputs), but KBM users rely on it for
smooth analog motion (e.g. Bloodborne). Revert to the default smooth=true.
* input: correct controller state copyright year
* abstract fs part1
* ops ifile :P
* added backedn to fs/file_system
* memory handling for non hosted backends
* remove ZArchive
* total remove
* make linux happy
* refactor loader to support backends
* AVIO streamer for ffmpeg
* Improved IterateDirectory so it can getdents in zarchieve works
* support for fstat in zachieve
* fixed module loading
* fs : added ReadFile
* improve file reading
* fixed trophy extraction , npbind,playgo
* argg
* fixed headers
* zarchive improvements for concurrent access
* playgo fix
* system service fix
* fixed restart from zar
* fixed save data icon read from zarchive
* fonts fix
* some stephen's comments fixups
* improved
* fixed rest of stephen's comments
* even more stephen review fixes
* possible linux fix
* fixed stephen's issue with multi games
* added thread worker to zarchive to solve some race conditions videos should be working fine now
* fixed paths for updates,mods
* fixes for updates , splash screen , icon
* fixed case sentitive issues in linux?
* Use pthread* instead of std::mutex in PthreacOnce
mutex goes unused, so this is all that needs fixing for this to be accurate. pthread_once struct is something guest applications can use, so it's relatively valuable to have it done correctly.
* Some days I wonder how absent-minded I can be
oops
* Implement min_lod support in image view creation
* Add min_lod member to ImageViewInfo struct
* Add ImageViewMinLod feature support
* Add support check for minLod feature
* log
* Update FindView to use modified lookup_info
* cleanup
* Adress comments
more cleanup
* Change buffer sharp struct
Hardware testing confirms that the last 4 bits of base_address (which would be outside the GPU address space) are ignored.
* Do not bind buffers with an address of 0
We check for this in most places, but apparently not here.
* Log fault addresses in SrtWalkerSignalHandler
* Bump SRT walker patch to warning
Whenever this hack triggers, there's usually more trouble. Not sure why it's a debug, but it should at least be a warning to alert developers.
* Revert BufferCache::ObtainBuffer "optimization"
Was a change done for performance in BB, but it broke UE games.
* kernel: Enhance condition variable wait and signal handling with improved timeout logic and error checks
* kernel: Enhance logging in condition variable and mutex handling for better debugging
* Add vectored exception handler stack support for Windows
* I love you, CLang!
* Fix the resource release order in Pools::ReleaseContents.
This file was accidentally omitted from the previous commit.
* Implement exception stack management for Windows threads
* Fix stale pthread condvar wakes across platforms
* kernel: align pthread synchronization with libkernel
Updated:
- Replace generation-based condvar wakes with PS4-style FIFO and deferred wake behavior.
- Rework semaphore waits, posts, cancellation, and timeout races.
- Harden mutex ownership and Windows VEH stack lifecycle.
Added:
- pthread cancellation, cancel-type, testcancel, and timed-join exports.
- Native cross-thread cancellation delivery on Windows, Linux, and macOS.
- Relative timed semaphore waits.
* kernel: isolate synchronization race fixes
Remove unrelated VEH stack and shader changes, keeping this branch focused on pthread synchronization and cancellation behavior.
* some updates.
---------
Co-authored-by: w1naenator <valdis.bogdans@hotmail.com>
Thanks to recent changes to readbackLinearImages, we get a lot more of these pending ops, and a lot more of these PopPendingOperations calls. This exposed a race condition where calls to PopPendingOperations would run simultaneously with our PriorityPendingOpsRunner threads, and potentially call a callback while a runner pops it off the pending ops queue.
* Read files into a host buffer, then copy to guest.
For some reason this might fix the fread errors we see?
* Reduce unnecessary logging
* Assert on failed malloc
No reason not to do this.
* Use a static vector instead of allocating each read
* Make file_buf thread local
Shouldn't matter, but prevents any possible concurrency issues (though ReadFile was always called under a lock anyway)
* Assert on failure instead of logging an error
Also adds checks to other unchecked reads and writes
* Flush on all writes
Might as well do this while I'm here.
* avplayer: align frame sizes to 16
* align pitch to 64, calc num audio buffers
* green line rebranding into black line
* extend data to the bottom of the video, sync vdec impl
* fixed build
Was hoping to get away with lower overhead, but we need to have a lower-bounded check too, or Windows with it's somewhat sparse address space will have issues with host mappings going through CopySparseMemory and asserting.
Intentionally not checking size, as the more robust mapping check would be a huge performance loss for a hot code path like this.
Create the window directly in fullscreen when configured, so there is
no visible windowed -> fullscreen transition for the compositor to
animate. Refresh the cached window size after SDL_SyncWindow so the
first swapchain and the splashscreen use the real drawable size, handle
enter/leave fullscreen events like resizes, and fix a dead error check
on the unsigned display id.