Skip to content

Add the box-emu cores and Star Trek: 25th Anniversary (ScummVM) - #23

Merged
finger563 merged 16 commits into
mainfrom
feat/box-cores
Nov 16, 2026
Merged

finger563 merged 16 commits into
mainfrom
feat/box-cores

Conversation

@finger563

Copy link
Copy Markdown
Contributor

Summary

Adds the box-emu cores and Star Trek: 25th Anniversary (through ScummVM) to the Tab5, and fixes the memory and teardown problems that showed up when switching between games.

New cores

  • NES (nofrendo), Game Boy / Color (gnuboy), Master System / Game Gear (smsplus), MSX (fmsx), Doom (prboom) and Dark Forces (TheForceEngine), from the shared esp-cpp core repos.
    • A shim makes BoxEmu an alias for Tab5Emu.
    • Clashing symbols are renamed with objcopy --redefine-syms.
    • Each core's .bss/.data is moved to PSRAM.
    • Tab5Emu::reserve_romdata provides the 4 MB pool that MSX, Doom and Dark Forces expect.
  • NES runs at full speed: nofrendo's audio frame size is fixed.
  • Star Trek: 25th Anniversary runs on the esp-cpp/scummvm fork, branch esp32p4.
    • Only the startrek engine is built, without the ScummVM GUI and without RTTI.
    • It has a minimal ESP32 OSystem backend.
    • The game is a .scummvm marker file in its folder.
    • Saves are taken while the game waits for input; loading restarts the engine with save_slot.
    • The pause menu parks the engine task.

Memory and teardown fixes

  • ScummVM allocations go to PSRAM.
  • Each game run gets its own arena, which is freed whole when the game ends.
  • The singletons are created once at boot.
  • Fixed Star Trek leaks: LZSS input streams, the bridge action queue and the load thumbnail.
  • Dark Forces' free wrapper no longer swallows other games' frees (darkforces_forget_block()).
  • JK and TPM now really free their per-session buffers. vector::shrink_to_fit() does nothing under -fno-exceptions, so they now swap with an empty vector.
  • libsmacker double free fixed (OpenJKDF2 fd47e34).
  • Doom's 2 KB audio task overflowed its stack because of the HAL's 30 s audio log line, which is now debug-level.

Star Trek fixes from testing

  • Speech, and sound effects after the intro, played at volume 0. The backend now registers ScummVM's config defaults.
  • The intro stalled because an unimplemented drawing branch printed a warning per row to the slow console.
    • Repeated log lines are now counted instead of printed.
    • Close shapes are now drawn.
    • Shapes behind the camera are culled.

Notes

  • The nofrendo, gnuboy and TheForceEngine pointers are on their fix/riscv-build branches, which hold the RISC-V build fixes and the NES audio fix. Those branches should be merged to the core repos' main, which the box can then pick up too.
  • ScummVM's AI guidelines forbid AI-authored upstream contributions, so the scummvm fork stays a fork.
  • Known gaps in upstream ScummVM's Star Trek engine, the same on a desktop build:
    • the Enterprise is never drawn in the intro;
    • the bridge viewscreen and space flight aren't implemented;
    • the captions have grey bars behind them.

Testing (Tab5 hardware)

  • Every core was launched and checked for frame rate and for PSRAM returned after quitting:
    • NES, GB/GBC, SMS/GG, MSX, Doom, Dark Forces, Star Trek, JK, MGS, TPM, GBA, Genesis, SNES and PICO-8.
  • Switching sequences, all working:
    • JK → SNES → TPM → SNES → JK (JK hands back all of its PSRAM);
    • Doom → DF → ST → JK.
  • Star Trek:
    • the intro plays through to the bridge;
    • speech plays (checked by ear);
    • two runs in a row leave PSRAM at the same level.

🤖 Generated with Claude Code

https://claude.ai/code/session_01SFCMzomCZ5wRHJN4sxwkvY

finger563 and others added 10 commits November 14, 2026 10:27
…nd Dark Forces cores

The cores esp-box-emu moved to submodules (each repo holds the core plus
its port/ glue, written against BoxEmu, which the Tab5 aliases to
Tab5Emu), added the same way as the Genesis and PICO-8 ones:

- components/{nes,gbc,sms,msx,doom,darkforces}: the submodules plus their
  component CMake from the box, with the static state (.bss and .data)
  moved to PSRAM by tools/bss_to_psram.py (the P4 has no extern_ram
  placeholder for ldgen). fMSX and gnuboy define globals that snes9x, MGS,
  the Genesis core and the TPM runtime also define; those are renamed in
  their libraries with objcopy --redefine-syms.
- main/: the carts, ported from the box, with the Tab5's video settings
  (original = integer scale, fit, fill).
- rom_info: the new platforms (.nes, .gb, .gbc, .sms, .gg, .rom, .wad,
  DARK.GOB) and consoles (NES, Game Boy, Game Boy Color, Master System,
  Game Gear, MSX), in the box's order and colors.
- Tab5Emu: make_color(), osd_getromdata(), no-op haptics, and
  reserve_romdata(): MSX, Doom and Dark Forces run their allocator in the
  box's fixed 4 MB ROM block, which the Tab5 allocates per cart.
- pool_allocator from the box (JK's private pool functions are renamed
  jk_pool_* so they don't collide with it).

The core repos needed three RISC-V build fixes (fix/riscv-build in
esp-cpp/gnuboy, nofrendo and TheForceEngine): Xtensa-only inline asm and
an Xtensa-only ESP-IDF backtrace call.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SFCMzomCZ5wRHJN4sxwkvY
The NES ran at ~30 fps: the glue made 1600 audio samples a frame instead
of 367, and the Tab5 paces to the DAC. Measured after the fix: ~3 ms of
emulation per frame, 60 emulated / ~30 presented fps (nofrendo's frame
skip).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SFCMzomCZ5wRHJN4sxwkvY
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SFCMzomCZ5wRHJN4sxwkvY
- components/scummvm: esp-cpp/scummvm (branch esp32p4) as a submodule, with
  the ESP32 backend in backends/platform/esp32; built with the startrek
  engine only, no RTTI and no ScummVM GUI (~1.1 MB of flash).
- A game is a marker file *.scummvm in the game's folder (Emulator::SCUMMVM,
  shown under PC); ScummVM detects which game the folder holds.
- ScummvmCart: the engine runs in its own task; the pause menu parks it and
  saves / loads through the engine's save states into the slot files; the
  4:3 games' 320x200 is shown at 4:3. Touch is the mouse (two fingers = right
  click), as are a USB mouse and the gamepad (A / B = left / right click).
- Tab5Emu::display_size() getter (touch -> game coordinates).
- CI: checks out the new core submodules (ScummVM shallow).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SFCMzomCZ5wRHJN4sxwkvY
- esp-cpp/scummvm esp32p4 66a0f24: the backend's quit / load / task fixes
  and the startrek engine's save states and teardown fixes.
- The component renames ScummVM's allocation functions to the backend's
  PSRAM ones (backends/platform/esp32/psram_alloc.syms).

Tested on the Tab5 with the 25th Anniversary Enhanced CD (English + speech):
intro, bridge and briefings; pause-menu save (while the game waits for
input) and load (restarts the game with the state, from anywhere);
repeated launches and quits, including during the intro. Known: each engine
instance leaves ~90 KB of PSRAM behind (the engine's own teardown), until
the next boot.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SFCMzomCZ5wRHJN4sxwkvY
…tion

- esp-cpp/scummvm esp32p4 7c81792: three startrek engine leaks fixed (LZSS
  input streams, the bridge's ever-growing action queue, the save
  thumbnail) and the backend's detection data left on a jumped-over stack.
- idf.py -DSCUMMVM_TRACK_ALLOCS=1 build: tracks ScummVM's allocations and
  writes what each run leaves to /sdcard/saves/scummvm_leaks.txt.

Measured on the Tab5: free PSRAM flat during bridge play (was -600 B/s), and
a few hundred bytes a game run (was 50-130 KB).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SFCMzomCZ5wRHJN4sxwkvY
…oom's crash

- Dark Forces: its free() wrapper kept claiming its ended session's 4 MB
  block, which the Tab5 gives back to the heap with the cart: every free()
  the other games later made into that range was swallowed, and Jedi Knight
  hung for want of memory. The cart now calls darkforces_forget_block()
  (TheForceEngine 03417ff) before the block goes.
- ScummVM (esp-cpp/scummvm a0a35c3): a game runs in its own PSRAM arena,
  given back whole at its end, and the long-lived set-up is done at boot
  (scummvm_prepare() in main): Star Trek no longer leaves blocks mid-heap
  that split the big ones Jedi Knight wants.
- Tab5Emu::play_audio()'s 30 s report is debug-level: it is formatted on the
  game's audio task, and Doom's 2 KB one overflowed (a stack protection
  fault after 30 s of play).

Checked on the Tab5: Doom (60+ s), Dark Forces, Star Trek, then Jedi Knight
(20 fps), MGS, TPM, GBA, Genesis, PICO-8, NES, GB/GBC, SMS/GG, MSX.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SFCMzomCZ5wRHJN4sxwkvY
…d fails late)

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SFCMzomCZ5wRHJN4sxwkvY
…-op here)

Built without exceptions (ESP-IDF), libstdc++'s vector::shrink_to_fit()
does nothing: Jedi Knight's presentation buffers (~1.8 MB: the RGB565 frame,
two 8-bit frames, two overlays) and TPM's copy of the game executable stayed
allocated after the game. With ~1.35 MB less PSRAM since the new cores and
ScummVM, the SNES core then had too little left after Jedi Knight. They are
swapped with empty vectors now.

Checked on the Tab5: free PSRAM after quitting Jedi Knight is back to what
it was before (7.28 MB, largest block 7.17 MB); JK -> SNES -> TPM -> SNES ->
JK all run.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SFCMzomCZ5wRHJN4sxwkvY
Bumps the esp-cpp/scummvm fork: config defaults registered (speech and
effects played at volume 0 once the engine synced its settings), repeated
log lines counted (the per-row drawR3Shape warning stalled the intro), close
shapes drawn and shapes behind the camera culled (the intro's planet).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SFCMzomCZ5wRHJN4sxwkvY
@github-actions

Copy link
Copy Markdown

ESP-IDF Size Report for 'Tab5 Emu'

Metric Base PR Delta
FLASH 0 bytes (0.00%) 0 bytes (0.00%) 0 bytes (0.00%)
DRAM 267,732 bytes (60.67%) 269,520 bytes (61.07%) 🔺 +1,788 bytes (+0.40%)
IRAM 0 bytes 0 bytes 0 bytes
RAM (DRAM+IRAM) 267,732 bytes 269,520 bytes 🔺 +1,788 bytes (+0.40%)

FLASH uses app .bin size or json2 flash sum. RAM sums DRAM+IRAM via idf_size. Percentages shown when totals are available.
DRAM/IRAM usage does not include memory used by the heap allocator at runtime.
This report was generated by esp-idf-size-delta.

finger563 and others added 6 commits November 15, 2026 00:08
Bumps the esp-cpp/scummvm fork to the reverse-engineered ship rendering
(views picked by angle, the original's 2D transform, dithered close-ups),
the view check and draw order, and the intro's movement, credits and
captions as the original plays them.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SFCMzomCZ5wRHJN4sxwkvY
Bumps the esp-cpp/scummvm fork: the bridge's viewscreen shows space, the
star systems' planets (generated as the original does), entering and
leaving orbit and the mission start fly-bys as the original plays them.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SFCMzomCZ5wRHJN4sxwkvY
…is known

The ScummVM cart applied fit/original before the engine had set up its
screen (0x0 then, ScummVM being prepared at boot), so a game started
filling the screen until the pause menu re-applied it. It now follows the
screen's size whenever that changes.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SFCMzomCZ5wRHJN4sxwkvY
Bumps the fork: ships fly and fight as the original's, the captain flies
and fires from the viewscreen (tap: phasers, two fingers: torpedoes), and
the missions' bridge scripts with their battles.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SFCMzomCZ5wRHJN4sxwkvY
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SFCMzomCZ5wRHJN4sxwkvY
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SFCMzomCZ5wRHJN4sxwkvY
@finger563
finger563 merged commit 5bab8f8 into main Nov 16, 2026
1 check passed
@finger563
finger563 deleted the feat/box-cores branch November 16, 2026 02:33
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant