joeylib2/PERF.md

JoeyLib cross-port performance (UBER benchmark)

Regenerated 2026-06-30. ST/Amiga/DOS columns are fresh measurements from the headless harness (scripts/bench-atarist.sh Hatari, bench-amiga.sh FS-UAE, bench-dos.sh DOSBox). The IIgs column is the perf reference: the generic+override redesign was bit-identical on IIgs (MAME iigs-verify checksum 0817FE), so its codegen — and thus its ops/sec — is unchanged from the last IIgs capture. Every non-IIgs cell is a ratio (port / IIgs); the directive is that every target meet or beat the IIgs floor (>= 1.0x).

Sprite ops now meet or beat the IIgs on BOTH 68k targets after this session's work (compiled DRAW emitters + raw-plane-byte save/restore); they were the dominant violations (down to 0.13x). The ST column is honest for the first time — the ST port was crashing on first present (user-mode I/O poke) until the Super(0L) fix in jlpInit.

Genuine remaining sub-1.0x cells and why:

• ST jlDrawCircle (0.61-0.65x) — inherent planar-outline cost: a 4-plane word RMW per sparse outline pixel (8 memory accesses) vs the IIgs single nibble, amplified by ST video-DMA contention. Cycle-bound; ~0.80x ceiling even fully optimized. Not winnable in this format.
• Amiga jlFillCircle (0.82x) — already asm span-fill (long-store middles); closing the gap needs a blitter area-fill rewrite (risk: the blitter has wedged the system before).
• ST jlSpriteSaveAndDraw (0.97x) — ST runs save+draw as two dispatched ops; the IIgs fuses them. Marginal.
• ST jlTileFill (0.98x), Amiga jlFillRect 80x80 (0.95x) — within noise.
• ST jlFillRect 320x200 (0.53x) — ARTIFACT: the IIgs reference (60) is the SEI-inflated full-screen reading (see the ceiling note below; honest ~28). Against the honest reference ST is ~1.1x. Same caveat for surfaceClear / stagePresent full.

Op	IIgs (ops/sec)	Amiga (vs IIGS)	Atari ST (vs IIGS)	DOS (vs IIGS)
jlSurfaceClear	28	2.68x	1.32x	449.21x
jlPaletteSet	678	9.09x	4.27x	22.01x
jlScbSetRange	1005	3.62x	1.86x	14.20x
jlDrawPixel	1755	2.04x	1.18x	9.40x
jlDrawLine H	682	2.26x	1.61x	26.93x
jlDrawLine V	90	1.70x	1.14x	145.83x
jlDrawLine diag	35	1.14x	1.06x	475.37x
jlDrawRect 100x100	75	1.67x	1.16x	206.49x
jlDrawCircle r=16	232	1.20x	0.65x	64.79x
jlDrawCircle r=80	56	1.16x	0.61x	301.25x
jlFillRect 16x16	450	1.24x	1.11x	36.04x
jlFillRect 80x80	75	0.95x	1.33x	205.09x
jlFillRect 320x200	60	1.13x	0.53x	188.92x
jlFillCircle r=40	38	0.82x	1.39x	398.24x
jlSamplePixel	1916	3.42x	1.92x	18.04x
jlTileFill	1252	1.74x	0.98x	10.60x
jlTileCopy	997	1.80x	1.04x	7.88x
jlTileCopyMasked	498	1.81x	1.10x	32.48x
jlTilePaste	1106	1.99x	1.12x	14.64x
jlTileSnap	1473	2.29x	1.53x	9.96x
jlSpriteSaveUnder	528	1.05x	1.29x	23.69x
jlSpriteDraw	438	2.09x	1.09x	23.85x
jlSpriteRestoreUnder	487	1.08x	1.11x	34.83x
jlSpriteSaveAndDraw	277	1.23x	0.97x	53.79x
jlStagePresent full	42	7.81x	1.10x	412.17x
jlInputPoll	273	13.94x	4.19x	37.87x
jlKeyDown	3382	7.75x	3.76x	18.94x
jlKeyPressed	3345	7.66x	3.81x	18.73x
jlMouseX	4170	10.40x	5.22x	26.30x
joeyJoyConnected	3378	7.71x	3.71x	18.96x
jlAudioFrameTick	4106	8.00x	2.50x	12.42x
jlAudioIsPlayingMod	3536	9.30x	4.33x	26.78x
surfaceMarkDirtyRect (via jlFillRect 32x32)	240	1.42x	1.31x	47.12x

IIgs full-screen ceiling (~29 ops/sec). A full-screen op moves 32000 bytes = 16000 word-stores x 6 cyc (STA long,X / PEI, both 3 cyc/byte) / 2.8 MHz ~= 34 ms ~= 29 ops/sec maximum. So jlSurfaceClear, jlFillRect 320x200, and jlStagePresent full cannot legitimately exceed ~29; a higher reading means the benchmark clock was perturbed. The clock is jlFrameCount = the VBL-interrupt- driven GetTick, so an op that holds SEI while it runs can over- report its ops/sec. Treat full-screen IIgs numbers above ~29 as inflated and cross-check against the cycle model.