* dos: Some initial work.
* dos: Turn off buffer on stdio SDL_IOStreams.
Seeking breaks otherwise. We might be able to just fflush() before or seeking
instead?
* dos: Audio implementation using the Sound Blaster 16.
* dos: remove audio Pump interface.
Turns out DosBox-X was having trouble with the Sound Blaster or something;
standard DosBox works correctly directly from the interrupt handler, and
without doubling the buffer size.
* dos: just dump and restore the stdio buffer when seeking.
This is MUCH faster than just leaving buffering disabled, and also works
around getting bogus reads after an fseek. SDL_LoadWAV on test/sample.wav
no longer takes several seconds to finish, and comes up with the correct
data.
I wonder if we're triggering this in LoadWAV because we're malloc'ing data
between seeks/reads, and it's causing the djgpp transfer buffer to change. Or
maybe the Fat DS trick is confusing it? I don't know, I haven't had time to
debug it, it might just be a legit libc bug in djgpp too, for all I know.
* dos: Protect audio device "thread" iterations when streams are locked.
This uses an old trick we used in SDL 1.2 for MacOS Classic, which did its
audio callback in a hardware interrupt. If the audio is locked when the
interrupt fires, make a note of it and return immediately. When the lock is
released, if the interrupt has been fired, run the audio device iteration
right then.
Since there isn't a big device lock in SDL3 (available to the app, at least),
this keeps a counter of when any SDL_AudioStream is locked, which is probably
good enough.
* dos: Implemented initial video subsystem.
This uses VESA interfaces to manage the display and works with the software
renderer.
Events aren't hooked up yet, so prepare to close DosBox on each run. :)
* dos: Whoops, forgot to add these to revision control. Core and Main support.
* dos: Wired up basic filesystem support.
This gets most of the rendering examples, which use SDL_GetBasePath() to
find textures to load, working.
* dos: Fixed compiler warning.
* dos: Initial mouse support!
* dos: Move interrupt hooking code into core/dos.
* dos: Initial keyboard support!
* dos: Use a simple ring buffer for keyboard events.
Of course Quake 1 solved this better, haha. It's smart: less memory, dirt
simple, and you don't even have to worry about synchronizing with the
interrupt handler, because it's safe for both sides no matter when an
interrupt fires.
* ci: add djgpp job
[sdl-ci-filter djgpp]
[sdl-ci-artifacts]
* dos: Fix build issues after rebase onto current main
- SDL_runapp.c: Add SDL_PLATFORM_DOS to the exclusion list so the
generic
SDL_RunApp() is disabled when the DOS-specific one is compiled.
- SDL.c: Exclude SDL_Gtk_Quit() on DOS. DJGPP defines __unix__ which
sets
SDL_PLATFORM_UNIX, but DOS has no GTK/display server. The GTK source
is not compiled (CMake UNIX is false for DOS) so this was a link
error.
- sdlplatform.cmake: Add DOS case to SDL_DetectCMakePlatform so the
platform is properly detected from CMAKE_SYSTEM_NAME=DOS.
- i586-pc-msdosdjgpp.cmake: Add i386-pc-msdosdjgpp-gcc as a fallback
compiler name, since some DJGPP toolchain builds use the i386 prefix.
* Add 8-bit palette support to DOS VESA driver
* Add VBE page-flipping, state restore, and robust keyboard handling
- Implement double-buffered page-flipping for VBE modes with >1 image
page
- Save and restore full VBE state on video init/quit for clean mode
switching
- Improve DOS keyboard handling: support extended scancodes and Pause
key
- Lock ISR code/data to prevent page faults during interrupts
- Always vsync when blitting in single-buffered modes to reduce tearing
* Refactor Sound Blaster audio mixing to main loop
Move audio mixing out of IRQ handler to main loop for improved
stability and to avoid reentrancy issues. Add SDL_DOS_PumpAudio
function, update DMA buffer handling, and adjust sample rate to 22050
Hz.
Silence stale DMA buffer halves to prevent stutter during load.
* Add DOS timer support and update build config
* Add support for pre-SB16 8-bit mono Sound Blaster audio
Detect SB version and select 8-bit mono or 16-bit stereo mode.
Handle DMA and DSP setup for both SB16 and pre-SB16 hardware.
Add FORCE_SB_8BIT option for testing in DOSBox.
* Add SB Pro stereo support and simplify IRQ handler
* Add DOS joystick driver support
* Improve DOS hardware handling and clarify memory allocation
- Poll Sound Blaster DSP status instead of fixed delay after speaker-on
- Clarify DPMI conventional memory is always locked; update comments
- Document and justify DMA memory allocation strategy
- Free IRET wrapper after restoring interrupt vector to avoid leaks
- Throttle joystick axis polling to ~60 Hz to reduce BIOS timing loop
cost
- Always poll joystick buttons directly for responsiveness
* Query and use mouse sensitivity from INT 33h function 0x1B
* Add support for VESA banked framebuffer modes
Implement banked framebuffer access for VBE 1.2+ modes without LFB.
Detect and initialize banked modes, copy framebuffer data using bank
switching, and blank the framebuffer on mode set. Page-flipping is
disabled in banked mode.
* Add optional vsync to page flipping in DOS VESA driver
* Add cooperative threading support for DOS platform
* Move SoundBlaster audio mixing to SDL audio thread
* Fix DOS platform comments and workarounds for DJGPP support
* Fix SoundBlaster IRQ handling and DMA setup for DOS
- Pass IRQ number to DOS_EndOfInterrupt and handle slave PIC EOI
- Validate DMA channel from BLASTER variable
- Correct DMA page register selection for SB16
- Improve BLASTER variable parsing and error messages
- Unmask/mask IRQs on correct PIC in DOS_HookInterrupt
- Rename SDL_dosjoystick.c to SDL_sysjoystick.c
- Include SDL_main_callbacks.h in SDL_sysmain_runapp.c
- Add include guard to SDL_systhread_c.h
* Add DOS platform options and preseed cache for DJGPP
Disable unsupported SDL features when building for DOS. Add
PreseedDOSCache.cmake to pre-populate CMake cache variables for DJGPP.
* cmake: use a 8.3 naming scheme for tests on DOS
* Apply code style
* Update include/SDL3/SDL_platform_defines.h
Co-authored-by: Anonymous Maarten <madebr@users.noreply.github.com>
* Code review clean up
- Split DOS VESA mode-setting into its own file
- Replace magic numbers with named constants
- Update copyright dates to 2026
- Substract time taken by other threads form delays
* Fix DOS bugs and improve compatibility
- Disable fseeko64 for DJGPP due to broken implementation
- Refactor DOS timer delay to always yield and avoid busy-waiting
- Fix animated cursor rendering in DOS VESA backend
- Always set display mode when creating DOS VESA window
- Work around DJGPP allowing invalid file access in testfile.c
- Bump max threads to 16
- Apply workarounds for threading tests
* Add DOS platform documentation and fix a few issues
- Fix fullscreen default resolution
- Improve best mode matching
- Fix builds on GCC older than 7.0
- Fix text input events
* Fix keyboard mapping of "*"
* Fix running, and existing, under PCem
* Apply suggestions from code review
Co-authored-by: Cameron Cawley <ccawley2011@gmail.com>
* Pre-mix audio in ring buffer and copy to DMA via IRQ thread
* Video fixes and optimizations
* DOS: Fix Intel 740 and VGA compatability
* DOS: Update readme
* DOS: Fix thread ID, get GPU name
* DOS: Cap mouse range
* DOS: Map test resources to 8.3 names
* DOS: Skip unsupported WM color modes
* Fix "windowed" resolution selection
* DOS: Hide INDEX8 modes behind SDL_DOS_ALLOW_INDEX8_MODES
* Remove SDL_HINT_DOS_ALLOW_INDEX8_MODES and order modes logically
* Don't convert cursor if dest is not INDEX8
---------
Co-authored-by: Ryan C. Gordon <icculus@icculus.org>
Co-authored-by: Anonymous Maarten <anonymous.maarten@gmail.com>
Co-authored-by: Cameron Cawley <ccawley2011@gmail.com>
Co-authored-by: Gleb Mazovetskiy <glex.spb@gmail.com>
Co-authored-by: Jay Petacat <jay@jayschwa.net>
Tested-by: Cameron Cawley <ccawley2011@gmail.com>
This avoids requiring inline assembly for each architecture.
It also fixes some weakly ordered architectures which lacked
said inline assembly (RISC-V, MIPS, LoongArch, etc) and were
thus disasterously broken.
This reverts commit 0231ff03de.
This causes SDL to return 1024x768 when asking for a best fit to 640x480 (thanks @AJenbo), so I'm reverting this until we can investigate more.
This PR modifies the Emscripten joystick backend to detect the user's OS and store its ID in the GUID, because different OSes might need different mappings for the same controllers.
I'm not sure if different browsers on the same OS can also have different mappings, but if they can, browser detection can be added to the GUID too if needed.
This PR also makes the GUID use `SDL_HARDWARE_BUS_USB` instead of `SDL_HARDWARE_BUS_UNKNOWN`, similarly to how Android and MFI backends always use `SDL_HARDWARE_BUS_BLUETOOTH` and GameInput, XInput, and RawInput backends always use `SDL_HARDWARE_BUS_USB`.
On most implementations, filter pattern matching is case-sensitive. For case-insensitive matching of a pattern such as '*.png', the pattern *.[pP][nN][gG]' must be used.
* [tray/dbus] set tray session name to app subname in flatpak env
* [tray/dbus] have dbus service name come from the app id. app id falls
back to flatpak if in a flatpak environment
* [tray/dbus] change dbus menu path to work with apparmor
Returning true with an unavailable interface in no-op cases can prevent fallback to other inhibition methods. If the inhibitor interface was previously tried and marked as unavailable, just return false.
Currently, the SDL_Tray* structs in the unix backend are subclassed for each
implementation (only dbus so far). This means that the DBus-specific members
are in their own structs (SDL_Tray*DBus), which all are required to have the
corresponding 'parent' struct as their first member, so that they can be cast
easily and used in the more generic code.
However, other SDL systems which have pluggable backends work the other way
around: the 'generic' struct has an 'internal' member, which individual
drivers can use to store a pointer to any internal state. This is a bit
simpler to wrap one's head around -- particularly because it's consistent --
but does typically involve more memory allocations.
Change the unix/DBus tray implementations to use an 'internal' pointer to
match other subsystems.
When the pointer isn't being scaled, make sure the cursor scale factor is set to that of the window to avoid blurry cursors on high-DPI desktops, and use the inverse of the pointer scale value when selecting buffers for size-adjusted cursors. Fixes a regression from adjusting custom cursor sizes when using scale to display mode, and ensures that the best buffer size for the scaled cursor is always selected.
__sync_lock_test_and_set() is designed for creating locks, not as
a general atomic exchange function. As a result, it only provides
an acquire memory barrier and isn't guaranteed to actually store
the provided value (though it does on architectures we care about).
__atomic_exchange_n() is supported on GCC/Clang for the last ~10
years, so let's use that instead if available. We will keep the
__sync_lock_test_and_set() fallback around for ancient platforms,
but add a full memory barrier to match the documented behavior.
[N-Gage] Set proper brush style to draw filled rects properly.
[N-Gage] Add persistent buffers to avoid per-frame memory allocations (which are expensive)
[N-Gage] Add support for SDL_TEXTURE_ACCESS_TARGET, fixes#13165
[N-Gage] Update README, add hint that the compiler does not support aggregate initializations for structs (knowing this, avoids a lot of headache during debugging)
[N-Gage] Add basic fast-path optimisations for render operations.
[N-Gage] Fix line drawing.
XInput2 keyboard handling has limitations: system keys that shouldn't be passed through when the keyboard isn't grabbed can be seen, and the text input system needs key events to flow through the X server to function properly (passing synthesized events through the filter function is not sufficient and doesn't work with non-Latin character sets).
The primary bit of information missing from the core X key events that XInput2 provides is the source device, so use the XInput2 slave keyboard device events to store that value, and apply it to core X key events with the same serial. XInput2 events always arrive before core events so this works universally.
- Change active object priority from EPriorityLow to EPriorityStandard
- Process all events in batch before SDL_AppIterate() to reduce input lag
- Remove redundant SDL_PumpEvents() call (already done by SDL_PollEvent)
- Move clean-up logic into ShutdownApp() helper function
- Skip SDL_GetRenderScale call in Copy() fast path
- Cache last clear color to avoid redundant SetBrushColor calls
- Add whole-image bounds pre-check to skip per-pixel checks in rotation
- Simplify color packing in DrawPoints/FillRects to reduce overhead
- Remove redundant null checks
- Use cached texture properties instead of API calls (GetBitmapWidth/Height/Pitch)
- Eliminate duplicate SDL_GetRenderScale() call in Copy()
- Reorder CopyEx() fast paths to check no-transform case first
- Combine operations in NGAGE_ConvertColor() to reduce intermediate steps
- Add reusable line points buffer to eliminate per-call heap allocations in DrawLines.
- Cache last draw color to skip redundant SetPenColor/SetBrushColor calls.
- Pre-compute cardinal angle constants (0°, 90°, 180°, 270°) for CopyEx fast-path.
- Cache color modulation state to avoid redundant LUT rebuilds.
- Add missing break statement in HandleEvent.
- Initialize previously uninitialized lastTime variable in UpdateFPS.
Someone mentioned that Claude sees this file's previous text and adds "look at
AGENTS.md" to it's TODO list instead of treating it as the thing it should do
first.
If everything decides to prefer AGENTS.md at some point, we'll delete this
file.
A symlink would apparently also work, but it would be the only symlink in the
SDL git repo, so it seemed less disruptive to just copy the file over for now.
Apparently Claude Code won't read AGENTS.md by default at this moment, but
since that's the direction we're moving, let's just add the file Claude
currently looks for and tell it that the actual instructions are in AGENTS.md.
I thought these things were supposed to be less complicated that interacting
with humans, lol.
- Replace FixDiv with inverse scale factors in ApplyScale
- Improve incremental DDA in ApplyRotation
- Optimize ApplyColorMod bit manipulation and LUT addressing
- Batch color changes in DrawPoints and FillRects to reduce API overhead
- Add early-exit optimizations to Copy/CopyEx for common cases
- Streamline Flip function by removing unnecessary API calls
- Fix vertex indexing bug in FillRects
- Implement lookup tables for faster color modulation
- Cache 0°/90°/180°/270° rotations for speedup on common angles
- Add dirty rectangle tracking infrastructure
- Process 4 pixels at a time in all transform operations
- Remove SDL_Surface member from NGAGE_TextureData structure and update all functions that currently use
surface->pixels to instead access bitmap->DataAddress() directly. This eliminates the intermediate copy
step (Mem::Copy from surface to bitmap) in rendering operations.
- Eliminate per-frame allocations in Copy/CopyEx methods. These buffers are now allocated once and resized
only when needed.
cursor-shape-v1 version 2 adds dnd_ask and all_resize, but SDL_SystemCursor does not expose matching cursor types yet. Bind the protocol at version 2 now so SDL negotiates the updated interface correctly while keeping the current cursor mapping unchanged.
