libretro/virtualjaguar-libretro
1
0 Fork 0
mirror of https://github.com/libretro/virtualjaguar-libretro.git synced 2026-11-07 04:36:00 -10:00
Code Issues Packages Projects Releases 1 Wiki Activity
Files
master
virtualjaguar-libretro/docs/RELEASE_NOTES_v3.7.1.md
T

18 KiB
Raw Permalink Blame History

Virtual Jaguar libretro v3.7.1

Broken-games release, continued. Xenowings is now playable, Club Drive stops crashing when you drive, and the homebrew carts that froze the Accurate blitter now run.

v3.7.1 is the second broken-games release (milestone "v3.7.1, Broken games, continued"). It fixes titles that froze, crashed, showed a black screen or ignored input. There is no performance work and no new core option. The savestate format moves from v15 to v16; v3.7.0 and older states still load (see Savestates).

The Xenowings work makes the core emulate three more entries from Atari's own hardware-bug list (TOM bug 13, JERRY bug 13, JERRY bug 2), and a new document lists where the manual is wrong or silent (see Documentation).

Highlights

Game fixes

Cartridge titles

  • Xenowings (https://github.com/libretro/virtualjaguar-libretro/issues/811). Four separate faults stood between the title and a playable game. All are headless-verified (scripted button presses and framebuffer checks); none is yet confirmed on a RetroArch run.
    • Black gameplay, TOM bug 13 (https://github.com/libretro/virtualjaguar-libretro/pull/819). At the start of gameplay the GPU swaps its object-list buffer with load (r0),r2 followed by moveq #0,r2. Atari's bug list (JTRM Rev 8 p.136, "Scoreboard failure on successive writes") says the late load data lands after the moveq, so the object list pointer gets the real address. The core wrote the load data first, so the pointer became 0 and the screen stayed black. After the fix the stage 1 playfield renders (94% of the frame lit in the PR's headless check).
    • Back to the loader, 68000 prefetch (https://github.com/libretro/virtualjaguar-libretro/pull/823). A few frames into gameplay the game runs a self-modifying anti-tamper check. On a real 68000 the two words it patches are already in the prefetch queue, so the old instructions still execute once. The core now models that, only where it is observable. The PR measured instructions retired within 0.07% of the base on Doom, Alien vs Predator and Tempest 2000. Gameplay then stays up (990 consecutive lit gameplay frames, in both HLE and BIOS boot).
    • Input did nothing, JERRY bug 13 and bug 2 (https://github.com/libretro/virtualjaguar-libretro/pull/822, https://github.com/libretro/virtualjaguar-libretro/pull/825). The DSP joypad reader hit the same late-load rule on the DSP (https://github.com/libretro/virtualjaguar-libretro/pull/822) and then bug 2, "Scoreboard failure on indexed addressing mode stores" (JTRM Rev 8 p.134): a div followed by an indexed store writes the register's old value, not the quotient. After both, 835 frames differ from a no-input run in the PR's check, and screenshots show the ship moving, firing and scoring.
    • Guard (https://github.com/libretro/virtualjaguar-libretro/pull/824). The GPU idle-loop fast-forward now refuses a loop in which the late write-back fired, so the bug-13 rule and the fast-forward cannot disagree. In Skyhammer the rule fires 94 times in 3000 frames and nothing observable changed: identical frame hashes, audio and savestates.
  • Club Drive (https://github.com/libretro/virtualjaguar-libretro/issues/611, https://github.com/libretro/virtualjaguar-libretro/pull/828, https://github.com/libretro/virtualjaguar-libretro/pull/829). The game froze or crashed a few seconds into driving. GPU NORMI returned one too high for every input ($80000000 gave 9 where the chip gives 8, zero gave 0 where the chip gives -32). The jag_sim chip netlists decode to the lower values. Club Drive uses NORMI to pick a corner for each polygon it clips against the screen, so a polygon crossing two screen edges came out twisted, the edge walker bailed, and the GPU returned into garbage. With the fix, six driving input patterns from a saved state each ran 2400 frames clean with the picture moving; on the previous build four of the six crashed. The DSP has the same NORMI block and gets the same fix (https://github.com/libretro/virtualjaguar-libretro/pull/829; only Music Demo executes DSP NORMI in the cart corpus, with no visible or audible change). Club Drive's indoor world (Jerome's Pad) now draws its ceiling and walls instead of the flat clear colour, and the outdoor worlds' sky boxes render as before (https://github.com/libretro/virtualjaguar-libretro/issues/831 was this, not a missing sky). See Testing for the corpus effects.
  • Seven homebrew carts hung the Accurate blitter (https://github.com/libretro/virtualjaguar-libretro/issues/800, https://github.com/libretro/virtualjaguar-libretro/pull/817). Chroma-Luma (bin), DEMO1 (bin), DEMO1B (two dumps), JagMania (Jul 8), Ladybug Demo and Ladybug Demo (rom) each froze inside one blit under the default Accurate blitter. The blitter was not slow: these are uploaded executables with no vector table, and several end their 68K main with ILLEGAL. The 68K then ran the vector table as code and walked its stack down into the blitter registers. The core now parks the ILLEGAL vector at a loop at $1000 for any image that is neither a cartridge nor a CD. All seven now run to the end of a 600 frame test under the Accurate blitter, in both boot modes. Some of these demos used to produce noise from the runaway 68K; they now sit parked, so a few show a black or silent screen where they used to show garbage.
  • A blit that never ends no longer freezes the host (https://github.com/libretro/virtualjaguar-libretro/issues/794, https://github.com/libretro/virtualjaguar-libretro/pull/817). Some blits never finish on real hardware (phrase mode below 8bpp whose X step never reaches the count). The core now hangs the emulated blitter the way the chip does: status reads busy, further starts are ignored, and it clears on reset. RetroArch keeps running. Where repeating the step is not idempotent the core runs one defined pass, and the crash watchdog logs approx=1. Across the 155-cart corpus, both blitters, both boot modes, there are now no timeouts.
  • Six homebrew binaries loaded $1000 low (https://github.com/libretro/virtualjaguar-libretro/issues/818, https://github.com/libretro/virtualjaguar-libretro/pull/826). Headerless 68K binaries linked at $5000 were being loaded at $4000, so their first JSR ran font data. The loader gained $5000 as a candidate and breaks ties by which base makes JSR/JMP targets land on a function start. Exactly six of the 155 corpus titles change: Chroma-Luma (bin), JagMania (Jul 8), JagMarble (2000), FORCE Design Legion Force Jidai Intro Demo 0!, Joypad-TeamTap Tester [a1] and Music Demo (ScatoLOGIC). Each goes from a black screen to video (Music Demo also plays audio). The Music Demo fix is also the actual cause of https://github.com/libretro/virtualjaguar-libretro/issues/794: loaded low in BIOS mode its 68K sprayed the blitter registers. One thing remains for Music Demo: in BIOS mode it logs a single dsp_pc_escape at frame 5, while HLE does not.
  • White Men Can't Jump intro logo (https://github.com/libretro/virtualjaguar-libretro/issues/736, https://github.com/libretro/virtualjaguar-libretro/pull/827). Only the apex of the first logo (the Trimark pyramid) showed, and the High Voltage logo never appeared. The game starts the GPU decoding the logo and then fades it in with 128 blits without waiting. On real hardware each blit stalls the 68000 so the decode wins; with zero-time blits the fade finished while the logo was still decoding. The title now turns on Blitter Bus Timing by default through a per-title row. It is limited to this one title; it is not on for everything. Fast vs Accurate blitter makes no difference. The new wmcj_intro_logo check, which needs the first held picture to be at least 25% lit, went from 6.4% to 43.8%. A 6000-frame driven run reached the menus, team select, the city map and a live match with no new freeze, but https://github.com/libretro/virtualjaguar-libretro/issues/736 covers more than the logo and this release does not claim the whole game is clean.

Jaguar CD titles

  • Memory Track saves after a no-content boot plus a disc insert (https://github.com/libretro/virtualjaguar-libretro/issues/810, https://github.com/libretro/virtualjaguar-libretro/pull/815). Two faults. The core reported save RAM only when content was loaded, so the frontend saw nothing to save on a no-content launch. And the insert path never reinstalled the NVM BIOS module, which the insert's reboot wipes, so a game started by an insert saw the Memory Track with no NVM BIOS behind it. Save RAM is now exposed whenever the Memory Track is enabled, and the insert reinstalls the module. Covered by new test_disk_control cases (marker written before the insert survives it). Not yet verified in RetroArch: the actual .srm file name for a no-content launch, and a real game saving then reloading after an insert. With the Memory Track disabled, nothing is exposed.
  • Real-BIOS CD run-ahead and rewind (https://github.com/libretro/virtualjaguar-libretro/issues/804, https://github.com/libretro/virtualjaguar-libretro/pull/814). With cd_boot_mode=bios, Hover Strike diverged under run-ahead because two pieces of state lived outside the savestate: the flag that gates the 68K boot-stub injection, and the remainder that decides whether a CD FIFO refill takes 2 or 3 ticks. Both are now saved. Run-ahead at warmup 400, 150, 600 and 800 now agrees with a plain run. This is the savestate change below.

Savestates

  • States are now v16. v3.7.0 (v15) and older states still load, with the new fields left at their live values (exactly v3.7.0 behaviour). v16 states do not load in older builds. A v15 file written by a v3.7.0 build loads in v3.7.1 with video identical to v3.7.0.
  • Three chunks are appended after the HLE chunk: CDB1 and CDX1 (real-BIOS CD boot flag and refill phase, https://github.com/libretro/virtualjaguar-libretro/pull/814) and BLH1 (the "blitter hung" flag and its stuck count, https://github.com/libretro/virtualjaguar-libretro/pull/817). A missing or unrecognised chunk loads as the default, so a v16 state that lacks BLH1 loads with a normal blitter.
  • Format details and the chunk layout: docs/savestate-compat.md.

Documentation

Testing

Release gate, v3.7.1 against the v3.7.0 tag, each built from a clean tree.

  • Cart corpus, 155 titles. Fast blitter: 155/155 reach game code in HLE and 154/155 in BIOS (was 155 and 153; the v3.7.0 LOAD_FAIL, JagMarble BIOS, now completes). Accurate blitter: 155/155 in both modes, up from 148 HLE and 136 BIOS in v3.7.0.
  • CD corpus: 28 of 28 discs reach game code in both modes, unchanged from v3.7.0.
  • NORMI corpus sweep (https://github.com/libretro/virtualjaguar-libretro/pull/828). Across 155 ROMs, HLE and BIOS, both blitters, the only titles that change are Club Drive, Checkered Flag and Native Demo. Club Drive's rendering was confirmed correct by the maintainer. Native Demo (bin), HLE, was already a runaway GPU from frame 1; its garbage path reshuffles to a black, silent screen.
  • Several Fast-blitter rows for homebrew demos (DEMO1, DEMO1B, Ladybug and others) moved from audio or static video to silent or black. The matrix bisect (https://github.com/libretro/virtualjaguar-libretro/pull/830) puts the flip at https://github.com/libretro/virtualjaguar-libretro/pull/817: in each case the old signal was a runaway 68K, not real output.
  • The cart boot matrix runs only the Fast blitter by default; the Accurate sweep is not committed (it needs CART_MATRIX_PROBE_ARGS). The matrices record a stage plus coarse video and audio flags, so they cannot see most of what changed here. Xenowings is not in the corpus.

Developer and CI

Known issues

Upgrading

  • Savestates: v15 and older load; v16 states do not load in older builds.
  • No core option changes.

Downloads

Pre-built libretro cores are attached to the GitHub release, built by release.yml (24 targets), Each binary has a matching debug archive and SHA256 checksums are in SHA256SUMS.txt. Every push to develop publishes a nightly pre-release: https://github.com/libretro/virtualjaguar-libretro/releases/tag/nightly

Maintainers

Joseph Mattiello, with the Virtual Jaguar libretro contributors. Original Virtual Jaguar by David Raingeard (Potato Emulation) and James Hammons.

Reference in new issue View Git Blame Copy Permalink