joeylib2/PERF.md

JoeyLib cross-port performance (UBER benchmark)

Regenerated 2026-07-05 (PERF-AUDIT-PLAN.md Phase 8, #25b) from same-day captures of all four ports. This supersedes the 2026-07-04 Phase 0 baseline; every phase of the audit fix plan (0-8) has landed, so these are the numbers the library actually ships with.

2026-07-06 update: the IIgs sprite column is now COMPILED. Finding #77 (the loader/segment-BSS + malloc-heap reservation that forced the codegen arena into the app's own BSS) was fixed toolchain-side and JOEY_IIGS_SPRITE_CODEGEN was re-enabled; the four IIgs sprite rows were re-captured (bench-iigs.sh) and jumped 5-25x (SaveUnder 104->543, Draw 48->907, RestoreUnder 110->568, SaveAndDraw 15->376 ops/sec), bit-identical to the interpreter goldens (diff-uber-hashes 45/45). The ST and DOS columns are unchanged from the 2026-07-05 capture.

Also 2026-07-06: the Amiga jlpFillRect partial-width path was moved onto the blitter (BLTAFWM/BLTALWM hardware edge masks + masked minterm), which lifted jlFillRect 16x16 567->1153, 80x80 71->269 (both from below-floor to well above it), and -- because they build on fillRect -- jlDrawLine V 144->269 and jlDrawRect 100x100 122->203, all hash-identical (Amiga 45/45). The one remaining Amiga below-floor cell is jlFillCircle r=40 (74%); a whole-circle asm routine was tried and did NOT help (per-span dispatch is not the cost -- the 8-way-symmetry overlap fill dominates), so it stays on the C span walker for now.

2026-07-06, NATIVE-PERF Phase 0: four WORKLOAD rows were added (see NATIVE-PERF-PLAN.md) and the goldens refrozen at 49 lines after a clean regression gate (45 match + exactly 4 extras per port) and full 4-port cross-identity. The new rows measure what games actually do, and they expose the real problem the per-op rows hid: jlSpriteDraw unaligned runs at 5% (ST) / 9% (Amiga) of the IIgs floor -- a 15x/12x cliff vs the aligned row, because only x%8==0 draws are compiled on the 68k ports (see the plan's code-truth table). The gameFrame composite row is the bottom line: ST 9 / Amiga 18 / IIgs 10 / DOS 18 frames/sec for a modest game frame. Bonus finding: cross-identity of the sweep16 row validated the ST/Amiga shifted c2p walkers as pixel-perfect at all 14 previously-untested x phases. (Re-capture note: IIgs input/audio rows wobbled vs the prior MAME run -- e.g. jlAudioFrameTick 50223 -> 27612 -- run-to-run MAME variance on cheap ops; hashes are unaffected and those rows gate nothing.)

2026-07-06, NATIVE-PERF Phase 1 (draw side) LANDED: all-shift compiled sprite DRAW on both 68k ports. Amiga emits pre-shifted JIT variants for all 8 bit phases; the ST emits per-phase thunk + data tables driven by a shared masked-movem walker (spriteWalk.s) for all 16 intra-group phases, with the dispatcher's ST shift contract changed to x & 15. jlSpriteCompile is now plan-based: it measures every (shift, op) variant, greedily fits them under the uint16 offset cap (draws first, save/restore as atomic pairs), and degrades to the byte-aligned phases when the arena is too small instead of losing the sprite. RESULT -- the alignment cliff is CLOSED: jlSpriteDraw unaligned ST 40 -> 500 ops/sec (12.5x), Amiga 75 -> 803 (10.7x); sweep16 ST 2 -> 32 sweeps/sec (512 draws/sec), Amiga 5 -> 50 (800 draws/sec). Unaligned now performs like aligned on both ports (ST 82% / Amiga 84% of their aligned rows), and every one of the 49 golden hashes still matches -- the compiled paths are bit-identical to the interpreted walkers they replace, on all four ports. (Found along the way: the Amiga bench task stack was ~4 KB -- one deep call chain from overflow -- now 32 KB in bench-amiga.sh; the emitters' accumulators are file statics.)

Phase 1 step 5 (save/restore) ALSO LANDED, completing Phase 1: compiled save/restore on both 68k ports operate on the 16-px-group window (two width classes per sprite, like the IIgs MVN pair), the backup format is a pure function of the recorded geometry (identical for compiled and interpreted producers/consumers -- the ST HAL gained a raw-span interpreted branch, switched atomically with restore), and the gates route by SPRITE_SAVE_CLASS / SPRITE_RESTORE_CLASS with an odd- buffer-pointer demotion (68000 word ops need even addresses). The combined jlSpriteSaveAndDraw path now fires on 68k at EVERY x for the first time. RESULTS (floor-relative): Amiga SaveUnder 104% -> 276%, RestoreUnder 98% -> 236%, SaveAndDraw 94% -> 160%; ST SaveUnder 158% -> 203%, RestoreUnder 122% -> 163%, SaveAndDraw 87% -> 107%; Amiga gameFrame 18 -> 23 fps. Backup capacity contract: JOEY_SPRITE_BACKUP_BYTES in sprite.h (window worst case; all in-repo examples migrated + 2-aligned). DEFAULT_CODEGEN_BYTES is 64 KB on Amiga/ST (full variant sets). All 49 golden hashes held throughout.

2026-07-07, THE ST COLUMN DOUBLED -- idle-audio ISR tax found and fixed: every previous ST capture ran with ~49.5% of the CPU silently consumed by the 12.3 kHz Timer-A PWM ISR that jlpAudioInit started unconditionally -- it streamed buffer silence to the YM even with nothing playing, and UBER inits audio before its timed run. Found by a calibrated probe (known-cycle dbra spin + per-layer direct-call rows under the bench Hatari config): every probe row slowed by exactly 1.98x once audio was initialized, and the probe's un-taxed tilePaste cycles/call matched the instruction-count model that UBER's number mysteriously doubled. Fix: src/atarist/audio.c Timer A is now demand-driven -- installed on the first PlayMod/PlaySfx, parked again once the mix buffer drains back to silence with nothing live (drain-to-silence handshake in jlpAudioFrameTick; StopMod parks immediately when no SFX is riding the pump). Probe-verified: idle rows match the never-initialized rows exactly, the tax correctly returns while an SFX drains, and it vanishes after the drain. Full ST re-bench: 49/49 hashes bit-identical (pixels untouched) and EVERY row ~2x: gameFrame 9 -> 19 fps, tilePaste 1,574 -> 3,203/s, sprite draw 607 -> 1,188/s, unaligned draw 508 -> 992/s. The other three ports were audited for the same bug class and are clean: Amiga ~1.8% idle tax (PTPlayer rides Paula DMA), DOS ~0.5% (SB auto-init DMA, ~11 Hz half-buffer IRQ), IIgs 0% (no handler installed until PlayMod). ALL pre-2026-07-07 ST comparisons -- including the Phase-0 ceiling gaps and the Phase-1 relative results -- understated the ST by ~2x.

Correctness statement backing this table: all four ports produce BYTE-IDENTICAL pixels on every measured op and every correctness check -- 49/49 surface hashes match the frozen goldens in tests/goldens/uber/ (37 timed ops + 12 hashed checks; the PASS/FAIL checks -- palette round-trip, PRNG golden sequence, sprite clip round-trip, arena churn, sprite-from-surface -- pass everywhere). The Phase 0-era divergences (#75 tilePasteMono three-way, #78 ST fillCircle/tileFill, #87 IIgs circle outline) are all fixed.

Measurement honesty (read before trusting any number)

Four measurement artifacts have shaped this table's history. Three are fixed, one is inherent -- know them before comparing rows:

1. Per-iteration poll tax -- FIXED (Phase 0). The pre-audit bench loop called jlFrameCount() between every op; on the IIgs that was a GetTick toolbox chain (~600-800 cycles), so every small-op number carried the harness cost, not the op cost. Ops now run in adaptive batches per poll. Small-op numbers jumped 2-10x on every port when this landed; never compare against pre-Phase-0 captures.
2. SEI tick loss -- INHERENT on IIgs, rows tagged below. The benchmark clock is jlFrameCount = the VBL-interrupt-driven GetTick. An op that holds SEI while it runs suppresses those interrupts, the clock loses ticks, and the op over-reports. The three SEI-holding rows -- jlSurfaceClear, jlFillRect 320x200, and jlStagePresent full -- are CLOCK-UNRELIABLE in the IIgs column. The physics check: a full-screen op moves 32000 bytes at ~6 cyc/word, so ~29 ops/sec is the hard ceiling at 2.8 MHz. The present row's 68 and the clear row's 35 sit above it -- inflated; the honest values are at or below the ceiling.
3. ST frame-rate miscalibration -- FOUND and FIXED 2026-07-06; the ST column below is the corrected (millis-clock) capture. Before the fix, ops/sec was iters * jlFrameHz() / actualFrames and the ST's jlpFrameHz() hardcoded 50 -- but Hatari runs the emulated ST at ~60 Hz VBL, so every ST number was scaled by 50/60 ~= 0.83 (~17% pessimistic). Caught by cross-checking against UBER's independent Timer-C 200 Hz millis clock (refresh-independent), which ran ~16-22% faster; the Amiga's two clocks agreed exactly (FS-UAE = PAL 50 Hz), confirming it was ST-only. Two fixes landed: (a) UBER now computes ops/sec from jlMillisElapsed (refresh-independent on all four ports; identical to the old formula on IIgs where millis is GetTick-derived), and (b) ST jlpFrameHz() now reads the shifter resolution + sync-mode registers to return the true 50/60/71 Hz. The per-op UBER-CLK: log line reports frame-vs-millis as a standing cross-check (they now agree). ST cells rose ~17% (e.g. jlSpriteDraw 503->585, jlFillRect 80x80 97->115); DOS/IIgs/Amiga unchanged.
4. ST idle-audio ISR tax -- FOUND and FIXED 2026-07-07; the ST column below is the post-fix capture. jlpAudioInit started the 12.3 kHz Timer-A PWM ISR at init and left it running forever, even silent (~320 cycles/tick = ~49.5% of the CPU), and UBER inits audio before timing anything. Every ST capture before 2026-07-07 is ~2x pessimistic. Timer A is demand-driven now (runs only while a mod or SFX is audible), so the benches measure the op and an idle game gets the whole CPU. Amiga/DOS/IIgs audited clean for the same bug class (~1.8% / ~0.5% / 0% idle cost).

Baseline (2026-07-07): absolute ops/sec per port

The IIgs is the reference and the perf floor (every other port must meet or beat it -- project directive). Non-IIgs cells show absolute ops/sec plus percentage of the IIgs number; sub-100% cells are bolded as below-floor. tools/uber-perf-table emits this table directly; regenerate after any re-capture (Amiga needs TIMEOUT=600 -- the default cuts off the final checks). Capture provenance: ST 2026-07-07 (post-audio-fix); IIgs/Amiga/DOS 2026-07-06 (recovery-gate rerun, unaffected by the ST fix).

Op	IIgs (ops/sec)	Amiga (ops/sec, % of IIGS)	Atari ST (ops/sec, % of IIGS)	DOS (ops/sec, % of IIGS)
jlSurfaceClear	33	75 (227%)	89 (270%)	67 (203%)
jlPaletteSet	1561	11453 (734%)	10255 (657%)	29828 (1911%)
jlScbSetRange	163	4003 (2456%)	4744 (2910%)	8491 (5209%)
jlDrawPixel	3179	5853 (184%)	6645 (209%)	23689 (745%)
jlDrawLine H	1142	1703 (149%)	2660 (233%)	5708 (500%)
jlDrawLine V	99	269 (272%)	265 (268%)	456 (461%)
jlDrawLine diag	42	40 (95%)	86 (205%)	108 (257%)
jlDrawRect 100x100	89	203 (228%)	216 (243%)	378 (425%)
jlDrawCircle r=16	303	285 (94%)	359 (118%)	461 (152%)
jlDrawCircle r=80	70	61 (87%)	83 (119%)	98 (140%)
jlFillRect 16x16	605	1153 (191%)	1211 (200%)	3017 (499%)
jlFillRect 80x80	110	269 (245%)	236 (215%)	412 (375%)
jlFillRect 320x200	21	66 (314%)	78 (371%)	59 (281%)
jlFillCircle r=40	42	31 (74%)	155 (369%)	156 (371%)
jlSamplePixel	4815	7253 (151%)	9603 (199%)	46266 (961%)
jlTileFill	1981	2253 (114%)	3144 (159%)	9301 (470%)
jlTileCopy	1441	1853 (129%)	2720 (189%)	8764 (608%)
jlTileCopyMasked	623	953 (153%)	1366 (219%)	3147 (505%)
jlTilePaste	1741	2253 (129%)	3203 (184%)	9834 (565%)
jlTileSnap	2520	3503 (139%)	5751 (228%)	19290 (765%)
jlSpriteSaveUnder	545	1503 (276%)	2196 (403%)	9248 (1697%)
jlSpriteDraw	902	953 (106%)	1188 (132%)	6234 (691%)
jlSpriteRestoreUnder	510	1203 (236%)	1662 (326%)	5925 (1162%)
jlSpriteSaveAndDraw	376	603 (160%)	803 (214%)	4271 (1136%)
jlStagePresent full	67	359 (536%)	125 (187%)	360 (537%)
jlInputPoll	228	4003 (1756%)	2788 (1223%)	3148 (1381%)
jlKeyDown	10191	39303 (386%)	45286 (444%)	77548 (761%)
jlKeyPressed	11389	38053 (334%)	45286 (398%)	75092 (659%)
jlMouseX	20214	96703 (478%)	117981 (584%)	155372 (769%)
joeyJoyConnected	10191	39303 (386%)	43440 (426%)	77548 (761%)
jlAudioFrameTick	27509	55853 (203%)	40637 (148%)	11527 (42%)
jlAudioIsPlayingMod	17627	56453 (320%)	60441 (343%)	127708 (725%)
surfaceMarkDirtyRect (via jlFillRect 32x32)	323	853 (264%)	774 (240%)	1472 (456%)
jlSpriteDraw unaligned	845	803 (95%)	992 (117%)	5118 (606%)
jlSpriteDraw sweep16	56	50 (89%)	64 (114%)	363 (648%)
jlTilePaste map 10x5	35	46 (131%)	62 (177%)	200 (571%)
gameFrame composite	10	23 (230%)	19 (190%)	18 (180%)

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.

Reading notes

• Sprite rows measure COMPILED code on all four ports since NATIVE-PERF Phase 1 (2026-07-06): the 68k ports JIT all-shift draw variants plus group-window save/restore (previously stubs that fell to interpreters), IIgs codegen re-enabled via finding #77, DOS via finding #74. The interpreted walkers remain only as arena-full degrade paths -- and are proven bit-identical to the compiled code by the 49-hash gate.
• The below-floor cells are the standing punch list, and after the 2026-07-07 audio-tax fix the ST has NONE -- every remaining below-floor graphics cell is the Amiga's: the circle family (diag line 95%, circle r=16 94%, r=80 87%, fillCircle 74% -- planar 4-plane RMW tax) and the two sprite workload rows a hair under floor (unaligned 95%, sweep16 89%; chip-RAM display-DMA contention that the fast-RAM-code ST does not pay). DOS jlAudioFrameTick (42%) is the SB half-buffer silence refill on an otherwise-trivial call -- it gates nothing. Floor parity is effectively DONE; the open work is ceiling-relative (NATIVE-PERF-PLAN.md Phase 2+), where every port still runs 2-5x below hand-coded technique on tile/sprite glue. Amiga jlFillCircle r=40 (74%) was probed 2026-07-06 with a full whole-circle asm routine (Bresenham + inline 4-plane long-fill span): it held 45/45 hashes but did NOT move the number (34 vs 31, within the coarse 11-iter sample), so the per-span C dispatch is NOT the bottleneck -- the asm was reverted. The real cost is the midpoint fill's 8-way symmetry filling each row 2-4x (idempotent overlap): the true lever is a scanline rewrite (one span per row, no overlap), which would roughly halve the byte writes and stay hash-identical. NB: this op is timed at batch=1 (~29 ms each) by the 50 Hz frame counter, a coarse measurement. (The old note here blaming a "4-5x model-vs-measured gap" partly on measurement was vindicated 2026-07-07: on the ST that gap was the idle-audio tax, artifact #4.)
• jlRandom has no timed row. Phase 8 (#43) rewrote its xorshift32 update in 16-bit halves on the IIgs only (~2-3x fewer cycles per call on the 65816; 32-bit shifts there are per-bit helper loops). Output is bit-identical -- host-proven over the golden sequence and a 50M-state sweep, and gated by UBER's random-golden check on all four ports.
• Input predicates stay functions (Phase 8 #40 decision): the post-clock-fix numbers (12K-77K ops/sec) put call overhead nowhere near a real frame budget; macro forms in the public header were rejected as not worth the API exposure.
• IIgs libc memset/memcpy remain ~4x slow (finding #79, toolchain-owned): llvm-mos runtime byte loops vs ORCA's MVN. The jlScbSetRange and jlPaletteSet IIgs rows carry that tax; library-side mitigations landed (#16 unrolled word copy, #21), the libc fix is flagged to the llvm816 session.

History

Pre-2026-07-04 tables (ORCA-era IIgs column, artifact-inflated DOS numbers, un-batched loops) live in git history only -- every conclusion drawn from them was re-derived from Phase 0 onward. The Phase 0 baseline (2026-07-04) and per-phase deltas are recorded in PERF-AUDIT-PLAN.md's progress tracker; headline gains since Phase 0: IIgs interpreted spriteDraw 18 -> 48 ops/sec (Phase 4), IIgs tilePasteMono asm 3.35x (Phase 7 #3), ST flood plane hooks 5.6x (Phase 7 #4), DOS paletteSet +92% / tileSnap +42% (Phases 2/5), DOS compiled sprites 284 -> ~6100 ops/sec (#74).