v0.0.3-dev
gba emulator witten in c++23.
the arm7tdmi interptreter supports different "backends" for executing instructions.
to select one of the backends, simple build with -DINTERPRETER=INTERPRETER_X where X is one of the following
- INTERPRETER_TABLE
- INTERPRETER_SWITCH
(default) - INTERPRETER_GOTO
for conveince, you can build with a cmake preset like so:
cmake --preset sdl2-table
cmake --build --preset sdl2-tableBelow are a few tests i tried using the various benchmark presets in CMakePresets.json. The fps is an average after running for 20s, accumulated after 3 runs.
| backend | speed | size | preset |
|---|---|---|---|
table |
1080 fps | 442.6 KiB (453,184) | benchmark-table |
switch |
1120 fps | 414.5 KiB (424,432) | benchmark-switch |
goto |
1020 fps | 1.3 MiB (1,380,776) | benchmark-goto |
NOTE: Ofast was slower than O3 in both table and switch builds, but faster in goto build (Ofast=1146 fps, O3=1115 fps).
| backend | speed | size | preset |
|---|---|---|---|
table |
1060 fps | 442.0 KiB (452,600) | benchmark-table-lto |
switch |
1095 fps | 409.0 KiB (418,840) | benchmark-switch-lto |
goto |
985 fps | 1.3 MiB (1,368,784) | benchmark-goto-lto |
NOTE: unlike the non-lto builds, O3 was always faster than Ofast.
| backend | speed | size | preset |
|---|---|---|---|
table |
1022 fps | 503.7 KiB (515,816) | benchmark-clang-table |
switch |
1065 fps | 489.6 KiB (501,320) | benchmark-clang-switch |
goto |
1010 fps | 860.9 KiB (881,568) | benchmark-clang-goto |
NOTE: lto wouldn't work for clang build for some reason, this is in an issue on my platform, not with my emulator / cmake.
web builds are the easiest way to quickly test a game. builds are automatically built from master. please report any bugs you find, giving as much info as possible such as browser, os, game etc.
gh-pages version doesn't support threads / mutexs. may crash.
netlify version supports threads / mutexes, won't crash.
- armwrestler
- thumb.gba
- arm.gba
- memory.gba
- nes.gba
- dma3 special
- unused bits
- unused region read
- bios read (when pc is not in bios)
- proper r/w timing access
- proper fifo mgba-emu/mgba#1847
- obj in obj window
- obj not in obj window
- bg affine
- obj affine
- obj mosaic
- window wrapping
- obj 4bpp
- better optimise vram r/w (somehow adjust the masking so it always works).
- refactor code. make functions into methods where possible (and if it makes sense).
- further reduce template bloat for arm/thumb instructions. this will help for platforms that have a small-ish icache.
- faster path for apu sampling, ie, if channel is disabled, don't sample it. if all dmg channels will output zero, stop sampling further.
- measure dmg apu channels on scheduler VS ticking them on sample(), likely it's faster to tick on sample().
- batch samples. basically, on sample callback, save everything needed (about 8 bytes) to an array. after X amount of entries added, process all samples at once. this can be processed on the audio thread (frontend) or still in emu core.
- impl computed goto. along side this (or first), impl switch() version as well. measure all 3 versions!
- prioritise ewram access first as thats the most common path.
- check for fast paths in dma's. check if they can be memcpy (doesnt cross boundries) / memset (if the src addr doesnt change). cache this check if R (repeat) is set, so on next dma fire, we know its fast path already.
- measure different impl of mem r/w. current impl uses fastmem, with a fallback to a function table. should be very fast, but maybe the indirection of all r/w access is too slow.
- fastmem for reads. ie, array of pointers to arrays.
- adding [[likely]] to fastmen reads, this gave ~200fps in openlara.
- scheduler. HUGE fps increase.
- halt skipping. when in halt, fast forward to the next event in a loop.
- tick arm/thumb in a loop. break out of this loop only on new_frame_event or state changed. this avoids the switch(state) on each ittr.
- switch instead of table, switch was slower (see table_switch_goto branch).
- computed goto instead of table, computed goto was slower (see table_switch_goto branch).
- array of functions pointers for reads. this was much slower than having 2 arrays, 1 of pointers, one of functions.
- cowbite spec https://www.cs.rit.edu/~tjh8300/CowBite/CowBiteSpec.htm
- gbatek https://problemkaputt.de/gbatek.htm
- tonc https://www.coranac.com/tonc/text/toc.htm
- belogic for apu http://belogic.com/gba/
- dillion for flashmem https://dillonbeliveau.com/2020/06/05/GBA-FLASH.html
- dennis for eeprom https://densinh.github.io/DenSinH/emulation/2021/02/01/gba-eeprom.html
- emudev discord
- jsmolka for their very helpful tests https://github.com/jsmolka/gba-tests
- normmatt (and vba) for the builtin bios https://github.com/Nebuleon/ReGBA/tree/master/bios
- endrift for mgba https://github.com/mgba-emu/mgba
- ocornut for imgui https://github.com/ocornut/imgui
- ocornut for imgui_club https://github.com/ocornut/imgui_club
- everyone that has contributed to the bios decomp https://github.com/Gericom/gba_bios
- xproger for openlara (fixed several bugs in my emu) https://github.com/XProger/OpenLara
- zayd for info on rtc https://beanmachine.alt.icu/post/rtc/
- pokeemerald for a being a good reference https://github.com/pret/pokeemerald
- kenney for the onscreen control buttons https://www.kenney.nl/assets/onscreen-controls