Virtual Jaguar libretro v3.5.1 — Settings clarity

A small fast-follow to v3.5.0. No emulation behaviour changes: this release makes the core options honest about what they do to each other, shortens the descriptions that had grown into walls of text, and fixes a test fixture that kept make test red on a clean tree.

The compounding-settings warning (#595)

DSP Idle-Loop Fast-Forward (virtualjaguar_risc_idle_skip) is the largest speed lever the core has — 66-87% less DSP interpretation on the titles measured in #569 — and DSPExec() switches it off entirely while certain other options are active. Nothing said so. A user who raised the RISC overclock on a borderline-slow title paid the overclock's own cost and silently forfeited the bigger win, which reads as "overclocking made it much slower" with no explanation available anywhere.

The core now logs one [perf] line naming the actual suppressor:

[perf] virtualjaguar_risc_idle_skip is enabled but suppressed by:
virtualjaguar_risc_clock_scale -- the DSP idle-loop fast-forward is doing
nothing, so this combination can run slower than idle-skip alone. Your
settings are honored, not overridden.

It warns; it never overrides. User-set values always win — the same rule the per-title database already honours for known-bad entries (#464). Latched once per content load, so it does not repeat every time the options menu is opened.

The four suppressors are virtualjaguar_risc_clock_scale (any non-1x value), virtualjaguar_dram_timing, virtualjaguar_gpu_pipeline_timing, and virtualjaguar_blit_memo. Each of those options now says so in its own description, and the idle-skip option lists what turns it off.

Two options that look like suppressors and are not

Worth stating plainly, because both this feature's original issue text and some earlier notes claimed otherwise before the gate was read and measured:

virtualjaguar_m68k_clock_scale does NOT disable idle-skip. m68kClockScalePct appears in the DSP's budget arithmetic but nowhere in the gate. The 68K overclock and idle-skip compose fine.
virtualjaguar_blitter_timing does not either. vjs.blitterTiming never reaches busArbiter.enabled; only virtualjaguar_dram_timing sets that.

Measured on Alien vs Predator, 120 frames:

Configuration	Idle-skip fires	DSP opcodes/frame
idle-skip alone	112,398	419,949
+ `m68k_clock_scale` 3x	127,525	476,343
+ `blitter_timing`	112,316	419,776
+ `risc_clock_scale` 2x	0	917,593
+ `dram_timing`	0	363,628

Warning on either of the first two would have false-positived on the most likely overclock combination there is.

Note also that despite the risc_ prefix, idle-skip is DSP-only — the GPU idles too and gets nothing from it — and it still ships off by default pending a listening pass, so it has to be enabled by hand.

Core-option text pass (#596)

23 info strings rewritten; the rest already read as a single idea and were left alone. Net −700 characters even after adding the new interaction text above.

Cut: internal issue numbers, source and doc paths, wire-protocol names, and boilerplate restated on option after option. Kept every load-bearing caveat — "takes effect on restart", "off by default", "bit-identical either way", cartridge-vs-CD applicability, and "report bugs only at 1x".

Largest reductions: Network Link Speed (1090→893), Cart BIOS (682→513), CD Boot Mode (734→651), Voice Chat (599→481), Widescreen (297→239).

Category descriptions were considered as a home for the repeated presentation-only boilerplate and deliberately not used: libretro.h documents retro_core_option_v2_category.info only as "secondary help text", which is spec intent rather than evidence that any frontend renders it somewhere a user actually reads. Losing a caveat costs more than duplicating one.

`make test` exits 0 again (#590)

test_pertitle_db cases 4 and 7 asserted against Alien vs Predator's shipped titledb row expecting a virtualjaguar_true_color pair. #551 removed that pair after a pixel-diffed A/B found it changed 0.0000% of pixels — so the test had been failing on unmodified develop ever since, and every contributor since has had to know which two failures were "expected".

Fixed at the fixture, not the table: a new TitleDBSetPairsForTest() override (mirroring the existing hooks and negative-pair overrides) lets those two cases drive off a synthetic row instead of real, evolving shipped data. A test that breaks every time the database is legitimately edited was the actual defect.

Known limitations

The warning's mid-session path — raising the RISC overclock from the options menu during play, rather than having it set at load — is reasoned from content_loaded being true at that point, not measured. The load-time path is verified headlessly.
Nothing here changes emulation. Everything still open against v3.5.0 stays open, including World Tour Racing's second freeze (#589) and the on-device A10X capture (#601).

Stats

6 files changed, 232 insertions(+), 29 deletions(-)
3 commits since v3.5.0

Maintainers

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

5.3 KiB Raw Permalink Blame History Unescape Escape