AI_Slop/joeylib2
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
joeylib2/PERF.md

64 KiB
Raw Blame History

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.

2026-07-07, NATIVE-PERF PHASE 4 (IIgs arm) LANDED -- whole-map asm walker: iigsTileMapPasteMark (joeyDraw.s) replaces the IIgs stage-arm per-tile loop of jlTileMapPaste with ONE JSL that walks the index map: dst offset advances incrementally (+4/column, +1280/tile row) from a single gRowOffsetLut read, the register-starved inner lives in a 22-byte stacked D-frame (X survives the 16-store body and carries the dst offset; the map index sits in the frame because Y is clobbered), the idx << 5 tileset add carries into the bank byte, all three wrapping branches are brl, and ONE band union closes the call. Gated 50/50 + make iigs-verify. RESULT: IIgs batch row 60 -> 140 maps/s = 7,000 tiles/s (+133%), ~400 cyc/tile -- BELOW the 480-cyc single-tile hand-coded ceiling, because the walker amortizes the JSL/derive/mark that even a hand-coded per-tile call pays. Batch row across all ports: ST 8,150 / Amiga 7,650 / IIgs 7,000 / DOS 22,500 tiles/sec (the morning's per-call loop numbers were 1,600 / 2,300 / 1,750 / 10,000). The Amiga blitter cookie-cut sprite was recon'd and DEFERRED with a full design on record: the synchronous WaitBlit the hash contract requires erases the parallelism win at <= 16x16 (every current sprite row), and the chip-RAM pre-shifted source data is a net-new subsystem -- the honest experiment is a >= 32x32 golden'd row measured with and without the fence.

2026-07-07, NATIVE-PERF PHASE 4 CLOSED -- clean-buffer experiment PASSED, jlSurfaceCopyRect is public API: Scott signed off the jlTileMapPaste shape and approved the clean-buffer experiment; both landed same-day. jlSurfaceCopyRect(dst, src, x, y, w, h) is a SAME-POSITION rect copy (clip, then snap outward to the 16-px group grid on every port -- uniform snap keeps cross-port hashes byte-identical): core wrapper + dirty mark in surface.c, generic chunky row-memcpy engine (serves DOS and both IIgs surface flavors -- the IIgs stage pixels pointer reaches bank $01 through ordinary large-model stores, recon-verified against the sprite-restore precedent), ST override = strided interleaved row memcpys (the raw-span restore recipe), Amiga override = per-plane CPU row copies (CopyMemQuick is unusable at 16-px grain: word alignment only; the modulo-driven blitter variant is the recorded production upgrade). The STRATEGY row gameFrameClean composite (erase 3 sprite windows from a pristine clean surface + 3 draws + present; NO per-frame background repaint, NO save-under -- the static-background pattern, not an op-for-op gameFrame speedup) measures, vs gameFrame: ST 48 vs 22 = 2.18x (acceptance bar was 1.4x), DOS 195 vs 19 = 10.3x, IIgs 18 vs 10 = 1.8x, Amiga 34 vs 25 = 1.36x (Amiga gains least, exactly as the adversarial recon predicted: its fillRect was already blitter-fast, and clean->stage is a chip-to-chip CPU copy against display DMA -- the blitter rect variant is its lever). GOLDEN CONTRACT now 51 lines: regression gate exactly 50 match + 1 extra per port, 4-port cross-identity of the new row (hash C51FE171 x4), then freeze. Found along the way: make all does NOT regenerate joey.2mg, so bench-iigs.sh silently benched the previous UBER (clean 50/50, new row absent) until make iigs-disk -- recorded in the stale-binary memory.

2026-07-07, NATIVE-PERF PHASE 4 (ST arm) LANDED -- movep.l map engine: the ST jlTileMapPaste arm now dispatches to stTileMapPasteMovep (src/atarist/tileMap.s): per tile row, one move.l (a0)+ + movep.l d0,d16(a1) writes all 4 interleaved plane bytes in 36 cycles vs ~80 for the four compiled move.b -- movep IS the half-group byte-scatter pattern (the classic ST c2p trick), where movem is useless (byte-granular dst) and byte-pair word fusion loses (lsl.w #8 costs 22 cyc on a barrel-shifterless 68000, priced during recon). Cells walk in even/odd pairs sharing one group pointer (+8/pair, all row offsets compile-time d16); odd bx = head cell, odd width = tail cell. The C pair walker (stTileMapPaste in stTile.h, +17% alone -- adversarial recon correctly predicted the pure-C cap at +15-23% because gcc already hoists the mulu and the 632-cyc copy body dominates) stays as the odd-tileset-pointer fallback: move.l (a0)+ needs an even source, jlTileT is align-1, and the guard is one AND per call. RESULT: ST batch row 163 -> 346 ops/sec = 17,300 tiles/s (2.12x, 462 cyc/tile) -- ABOVE the Phase-0 12,700-15,500 "hand-coded ceiling", which under-counted by assuming byte-move bodies. Gate: ST 50/50 bit-identical (hash 52B9646B held through both engines) + a 12-case on-emulator proof (TILETEST, session scratchpad) covering odd-bx head, odd-width tail, w=1 columns, h=1, checkerboard/end/all skips, and both odd-pointer fallback paths, each with a border-ring containment check. Batch row across ports: DOS 22,500 / ST 17,300 / Amiga 7,650 / IIgs 7,000 tiles/s. IIgs PEI-slam map variant was EVALUATED and REJECTED: models ~+28% (~312 cyc/tile) but requires SEI for the whole map (~5.6 ms interrupts-off at 10x5, audio ISR jitter) AND the VBL-driven bench clock cannot honestly measure SEI ops (PERF artifact: SEI inflation) -- the walker already beats the 480-cyc per-tile ceiling without it.

2026-07-07, NATIVE-PERF PHASE 3 LANDED -- jlTileMapPaste batch API: one public entry pastes a wTiles x hTiles rectangle from a tileset + row-major uint8_t index map (TILE_MAP_SKIP = 0xFF leaves a cell untouched): one validation, one dirty union (68k/DOS; the IIgs stage-fused entries mark per pasted tile, and its loop hoists the stage test + does tileset indexing as base + (idx << 5) -- the natural tiles[idx] spelling cost a 32-bit indexed multiply per cell on the 65816 and erased the batch win until specialized). GOLDEN CONTRACT now 50 lines: refreeze followed the Phase-0 procedure (regression exactly 49 match + 1 extra per port, then 4-port cross-identity of the new jlTileMapPaste 10x5 row -- all four hash 52B9646B -- then freeze). RESULTS, tiles/sec batch vs per-call loop: ST 8,150 vs 4,000 (+106%, gap-to-ceiling 1.6-1.9x); Amiga 7,650 vs 3,750 (+104%, gap 1.40x -- INSIDE the ~1.5x program target); DOS 22,500 vs 12,900 (+74%, 2.5x); IIgs 3,000 vs 2,950 (+2% -- its Phase-2 fused entries already deleted the per-tile wrapper, so the remaining cost is ~93% inner JSL; a whole-map IIgs asm walker is the Phase-4 lever). Sprite batch APIs were DESCOPED from Phase 3: the phase gate measures only the tilemap row, and post-R5 sprite dispatch overhead is ~3-6% -- not worth new public surface without a measured need.

2026-07-07, NATIVE-PERF PHASE 2 COMPLETE -- Steps D3 (R6 + R10) close the phase: (R6) the IIgs masked-copy inner's 4 x 8 jsr/rts round-trips to tmaskByte were inlined via the TMASK_BYTE assembler macro (identical body; the row loop-back became brl -- the expanded body exceeds 8-bit branch range): jlTileCopyMasked 625 -> 681 (+9%, exactly the 384-cycle model). (R10) the stage arm of jlDrawPixel is one fused JSL (iigsDrawPixelMark: single-row widen + nibble RMW plot against the pinned stage base): jlDrawPixel 3,179 -> 5,576 (+75%, 502 cyc/call incl. wrapper vs the measured 880 before); the batching plotters (lines/circles, plotPixelNoMark) keep the unfused inner and were flat, as were tiles/sprites. Final gate 49/49 x4 + make iigs-verify. PHASE-2 EXIT vs its gate ("small-op rows move measurably toward Phase-0 ceilings on all ports; goldens hold"): tilemap row tiles/sec ST 1,600 -> 4,050 / Amiga 2,300 -> 3,750 / IIgs 1,750 -> 2,950 / DOS 10,000 -> 12,900; gameFrame ST 9 -> 21 / Amiga 18 -> 25 / IIgs 10 / DOS 18 -> 19; IIgs tilePaste ceiling gap 3.5x -> 2.15x, ST 2.6x. The remaining per-call floor is the public wrapper + one JSL/call -- Phase 3 batching territory. Optional leftovers R7 (saveAndDraw consolidation) and R12 (IIgs C-residue reshapes) are recorded in the plan, not phase-gating.

2026-07-07, NATIVE-PERF Phase 2 Step D2 LANDED -- sprite dispatch flattening (R5): the spriteCompiled* runtime dispatchers moved from spriteCompile.c into spriteDispatch.h as always-inline functions consumed only by sprite.c, deleting the cross-TU call per compiled sprite op and letting the compiler share the slot/field loads the caller already did. Gated 49/49 x4. Sprite rows: IIgs +6-13% (SaveAndDraw 376 -> 423, Draw 906 -> 966), ST SaveUnder +15% (2,611), Amiga +5-11% (SaveUnder 2,053), DOS SaveUnder +16% (10,743). Two measurement footnotes: gameFrame wobbles +/-1 fps between runs (single-digit iters per 16-frame window -- quantization, not signal; ST has bounced 21/22 across three consecutive gates), and DOS jlSpriteDraw unaligned is BIMODAL across runs of identical binaries (~3,800 vs ~5,100; seen 07-06 and again today; the all-phase sweep16 row is stable at ~362, so per-phase behavior is fine -- likely DOSBox dynamic-core state; do not chase single-run deltas on that row).

2026-07-07, NATIVE-PERF Phase 2 Step D1 LANDED -- IIgs fused tile entries (R2): jlTileFill/Copy/Paste on the stage now run ONE JSL each: new joeyDraw.s entries (iigsTile*Mark, section .text.iigsFusedTile) derive the destination in asm from the pinned $01:2000 stage base + gRowOffsetLut, write via long,X with the stage literal folded into each store, and widen the dirty band via a shared stageTileMark helper -- replacing C-side pointer math plus two JSLs (op + marker). Gated 49/49 (pixels) AND make iigs-verify (framebuffer -- the mark half is hash-blind, so the display check is load-bearing). RESULTS: IIgs tileFill 1,981 -> 3,251 (+64%), tilePaste 1,687 -> 2,710 (+61%), tileCopy 1,441 -> 1,988 (+38%), map row 35 -> 59 (+69% = 2,950 tiles/s), gameFrame 10 -> 11 fps; unfused controls (copyMasked, snap, sprites) flat. IIgs tilePaste is now ~1,030 cyc/call vs the 480-cyc hand-coded ceiling (gap 2.15x, was 3.5x). Toolchain landmine found en route: a 16-bit jsr between .section blocks can cross OMF segment banks (no link error, wild jump into $00CFxx) -- helpers must share their callers' section or be jsl/rtl; see the cross-section-jsr feedback memory.

2026-07-07, NATIVE-PERF Phase 2 Steps A-E LANDED (per-op tax): five hash-gated steps, each 49/49 x4. (A) plane-pointer access inlined into the 68k compiled-sprite dispatchers (amigaPlanar.h / stPlanar.h; the jlpSurfacePlanePtr cross-TU call -- ~190 cyc x4 per Amiga dispatch -- deleted): Amiga sprite ops +10-23%. (B) whole-tile dirty marks specialized (band = [bx2, bx2+1] with compile-time proof, unrolled 8-row widen): tiles +9-28% on 68k/DOS. (C) jlDrawRect raw-edge fills + ONE band-identical mark (+22-40%); the multi-row widen unroll is Amiga-only by measurement (gcc-mint compiled it WORSE, ST -5% -- see the per-port #if in surface.c). (D=R9) presents consume the core gStageScbDirty/gStagePaletteDirty flags instead of memcmp'ing 512-712-byte cached mirrors: gameFrame +1-2 fps on ST/Amiga/DOS. (E) ST/Amiga tile ops moved into always-inline per-port headers (stTile.h / amigaTile.h), deleting the second cdecl layer (ST tiles +9-14%, Amiga +5-11%), then the Amiga row loops unrolled to d16(An) form (tilePaste +36%, tileSnap +52%). NET SESSION MOVEMENT on the tile gate row (tiles/sec, map 10x5): ST 1,600 -> 4,000; Amiga 2,300 -> 3,750; DOS 10,000 -> 12,900. gameFrame: ST 9 -> 21, Amiga 23 -> 25, DOS 18 -> 19 fps (IIgs untouched at 10 -- its fused-asm glue is Phase 2 Step D, still open). The present-row jumps (ST 467, DOS 1,137) are the dirty-union idle early-out running flag-gated; gameFrame is the honest present signal.

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.

2026-07-09, Phase 7 P7-3 -- Amiga present sped up (src/amiga/hal.c only). The Amiga present was Phase 6's single largest out-of-bar item (2.8-5.3x). A probe forensic split it: a fixed per-present scan (P8 idle = 1.51 ms) plus, for large presents, a ~11 ms CPU copy (P1 present96 copied 12,288 B). Two bisectable changes, each P11-gated (presentPageSum = exhaustive displayed-page-vs-shadow byte parity, the gate for the hash-invisible present -- PASS on both): (A) the per-present snapshot pass was folded into the scan's COLD union-dirty path (a row clean in both frames already holds the clean sentinel in gPrev, so the hot path stays 2 loads -- a first attempt that touched every row REGRESSED and re-confirmed the recorded gcc-amigaos register-pressure trap), dropping the present from two 200-row passes to one (scanned 400 -> 200); (B) the large-run copy now runs on the blitter via the W3 amigaBlitterCopyPlanes (A->D, offset+modulo) above a 128-plane-word threshold, so present96 / gameFrame's 96-row bands move off the CPU while present3win's scattered windows stay on the CPU THIN path. RESULTS: present96 46 -> 71 (+54%), present3win 93 -> 103 (+11%), gameFrame 25 -> 31 (+28%), gameFrameClean 62 -> 71, jlStagePresent full 584 -> 609; Amiga 51/51 golden hashes hold (present changes are hash-invisible -- pixels unchanged). Remaining Amiga-present lever: a fused-scan present asm (ST presentSlam analogue over 4 plane buffers) for the scan-bound rows -- C fusion is now empirically register-pressure-capped, and the async-fence blitter contract stays parked behind a >=32x32 row.

2026-07-09 -- fused-scan present asm + a component probe that located the real cost. src/amiga/presentScan.s (hand-rolled 68000, the 4 band arrays walked in address registers) replaces jlpPresent's C union scan. The plain asm (no fast path) barely moved present3win (103->106), which first read as "the scan isn't the bottleneck" -- but a present-component skip-mask probe corrected that. It split present3win's 9.4 ms: SCAN ~4.9 ms, FLUSH ~3.3 ms (chip->chip CPU copy of the 3 tiny windows), FLIP ~0.4 ms, residual (copper-check/reset/WaitBlit/marks/harness) ~1.2 ms; the 3 marks alone are 0.64 ms. So the scan IS ~half the present -- the plain asm just wasn't faster than gcc because the scan is INSTRUCTION-COUNT-bound (~16 instr x 200 rows), not the register spill it was assumed to be. The ST presentSlam 4-row fast path (one move.l compares 4 clean-sentinel min bytes, cur AND prev, and skips the group; needs gPrev arrays aligned(2)) LANDED: present3win 106->128 (+21%), present96 ->78, gameFrame 32->34, gameFrameClean 62->82, P11 PASS, 51/51. A first cut passed all PIXEL gates but broke the gameFrameClean golden row -- a bisect proved the fast path's 'bra .rowLoop' skipped the loop-exit check, so a clean group ending at row 199 landed the row index on 200 and processed a PHANTOM out-of-bounds row 200 (writing gPrevStageMin/MaxWord[200] one byte past the arrays). That corrupts a neighbouring global invisibly to P11/hashes/counters yet breaks a later jlSurfaceCreate AllocMem -- gameFrameClean (the only op allocating a 2nd surface) was an accidental memory-integrity canary. FIX: the exit check now lives at the TOP of .rowLoop, guarding both re-entry paths. The FLUSH lever also LANDED: the ~3.3 ms chip->chip CPU copy of the small windows was contention-bound, so lowering AMIGA_PRESENT_BLIT_MIN_WORDS 128 -> 16 routes them to the blitter (parallel chip->chip DMA) -- present3win 128 -> 153 (+20%), gameFrameClean 82 -> 90, 51/51 + P11. An OWN-ONCE flush polish (the present OwnBlitters once instead of per window) also landed correct (51/51 + P11) but PERF-FLAT (present3win 153 -> 154, within noise): after the threshold drop the flush is blitter-DMA-bound, so the per-window Own/Disown it removes was already negligible -- kept for the right blitter discipline, not a speedup. Session present3win arc 106 -> 128 -> 153 -> 154 = +45%. No known remaining flush lever.

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 (2026-07-08): 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: all four ports 2026-07-08 from a full make clean rebuild (Phase 6 W1+W2+W3 landed: fused IIgs sprite/fillRect entries, sprite-mark specialization, per-row/fused presents on Amiga/ST/IIgs, ST movep tile singles, DOS map walker, Amiga blitter copyRect + map hoist); 51/51 golden hashes x4 on this exact build; Amiga solo TIMEOUT=600; IIgs from a fresh make iigs-disk image.

Op IIgs (ops/sec) Amiga (ops/sec, % of IIGS) Atari ST (ops/sec, % of IIGS) DOS (ops/sec, % of IIGS)
jlSurfaceClear 33 75 (227%) 88 (267%) 66 (200%)
jlPaletteSet 1561 11503 (737%) 10388 (665%) 29828 (1911%)
jlScbSetRange 163 4003 (2456%) 4833 (2965%) 8491 (5209%)
jlDrawPixel 5576 5903 (106%) 6522 (117%) 9009 (162%)
jlDrawLine H 1322 1753 (133%) 2670 (202%) 5682 (430%)
jlDrawLine V 104 303 (291%) 261 (251%) 456 (438%)
jlDrawLine diag 42 40 (95%) 88 (210%) 109 (260%)
jlDrawRect 100x100 110 285 (259%) 279 (254%) 470 (427%)
jlDrawCircle r=16 303 285 (94%) 359 (118%) 461 (152%)
jlDrawCircle r=80 70 63 (90%) 83 (119%) 98 (140%)
jlFillRect 16x16 785 1153 (147%) 1225 (156%) 3017 (384%)
jlFillRect 80x80 133 285 (214%) 240 (180%) 412 (310%)
jlFillRect 320x200 22 66 (300%) 77 (350%) 59 (268%)
jlFillCircle r=40 42 31 (74%) 155 (369%) 156 (371%)
jlSamplePixel 4797 7253 (151%) 9784 (204%) 46266 (964%)
jlTileFill 3239 3753 (116%) 3855 (119%) 11824 (365%)
jlTileCopy 1981 2903 (147%) 3626 (183%) 10910 (551%)
jlTileCopyMasked 681 1003 (147%) 1749 (257%) 3391 (498%)
jlTilePaste 2700 3753 (139%) 4685 (174%) 12575 (466%)
jlTileSnap 2520 6103 (242%) 8516 (338%) 19218 (763%)
jlSpriteSaveUnder 602 2053 (341%) 2611 (434%) 10718 (1780%)
jlSpriteDraw 1142 1153 (101%) 1341 (117%) 8146 (713%)
jlSpriteRestoreUnder 662 1603 (242%) 1865 (282%) 7826 (1182%)
jlSpriteSaveAndDraw 423 803 (190%) 934 (221%) 5478 (1295%)
jlStagePresent full 73 609 (834%) 407 (558%) 1132 (1551%)
jlInputPoll 229 4003 (1748%) 2788 (1217%) 3148 (1375%)
jlKeyDown 10191 39753 (390%) 44448 (436%) 77688 (762%)
jlKeyPressed 11432 38003 (332%) 44448 (389%) 74903 (655%)
jlMouseX 20138 97703 (485%) 117920 (586%) 155583 (773%)
joeyJoyConnected 10191 38903 (382%) 44260 (434%) 77697 (762%)
jlAudioFrameTick 27612 56503 (205%) 39885 (144%) 11490 (42%)
jlAudioIsPlayingMod 11029 56453 (512%) 60441 (548%) 127856 (1159%)
surfaceMarkDirtyRect (via jlFillRect 32x32) 397 853 (215%) 774 (195%) 1479 (373%)
jlSpriteDraw unaligned 1022 953 (93%) 1089 (107%) 6316 (618%)
jlSpriteDraw sweep16 67 58 (87%) 70 (104%) 463 (691%)
jlTilePaste map 10x5 59 75 (127%) 94 (159%) 257 (436%)
gameFrame composite 12 31 (258%) 32 (267%) 19 (158%)
jlTileMapPaste 10x5 140 191 (136%) 346 (247%) 1224 (874%)
gameFrameClean composite 20 71 (355%) 90 (450%) 192 (960%)

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.

IIgs full-screen writes: the paper ceiling is refuted in MAME (Phase 6 W2). The old note here derived ~34 ms / ~29 ops/sec for any full-screen op from 3 cyc/byte stores. The W2 fitted model -- which closes P1, P2, and the idle-present row simultaneously -- measured a shadowed push write at ~1.73 us (~4.8 fast cycles), so the honest MAME full-screen write floor is ~55 ms of writes alone; on real hardware (datasheet ~0.98 us/write) it is ~31 ms. The jlStagePresent full row is NOT a full-screen copy either way: it is the all-clean idle present (early-out scan only), which is why it reads far above any copy ceiling on every port. The GetTick clock itself is proven honest post-SEI-chunking (Lua cross-clock, tick-loss ~1.00); jlSurfaceClear remains the one mildly inflated row (1.21x).

Re-derived native-technique ceilings (NATIVE-PERF Phase 6 close-out, 2026-07-08)

The Phase-0 ceilings were paper models. Three measurement waves (W0 probes + ledgers, W1 IIgs sprite autopsy + fusion, W2 present fixes + fitted IIgs write model, W3 tile engines) invalidated or refined several of them; this section re-derives every headline ceiling from the MEASURED models and gates each row against the directive bar (measured within ~1.5x of hand-coded native technique for the row's exact work). "Ceiling" = the ops/sec a hand-coded native routine would achieve doing the row's exact work -- pixel/byte traffic at the measured machine rate, plus a minimal call and the dirty-band bookkeeping the present contract requires. Gap = ceiling / measured. Measured values: golden rows from the 2026-07-08 PERF.md capture; probe rows (P1/P2/P8) from the Phase-6 W0/W2 tracker entries. Gaps > 1.5x are bolded. Cells marked UNKNOWN have no recorded model and no honest derivation -- they are bounds, not ceilings.

Superseded by this section: (a) the "IIgs full-screen ceiling (~29 ops/sec)" callout above -- the 34 ms full-screen "PEI floor" is REFUTED in MAME: under the fitted write model (below), 32,000 shadowed bytes cost ~55.4 ms of writes alone, so the honest MAME full-present ceiling is ~14.5/s (~22/s real-HW variant), and the golden jlStagePresent full reading (73/s) is the idle-early-out artifact, not a copy; (b) the Phase-0 ST tile ceilings (12,700-15,500/s -- assumed byte-move bodies; the landed movep engine beats them); (c) the Phase-0 IIgs sprite-draw ceiling (4,300/s -- assumed no derive/mark/wrapper; W1 autopsy honest ceiling is 2,500-2,650/s); (d) the Phase-0 Amiga/ST/DOS "present full copy" ceilings (the timed row never re-dirtied; the P1/P2 probe models below are the real ones).

Row shapes (exact UBER geometry the ceilings are computed for): tile ops are one 8x8 tile; the map row is 10x5 = 50 tiles; sprite per-op rows are a 16x16 sprite at (40,30) -- x&15 = 8, so the 68k save/restore window is TWO 16-px groups = 256 B (IIgs/DOS chunky window is 128 B); gameFrame = 256x96 fill + 16 tile pastes + 3 saveAndDraw + 3 restores (sprites mod-16 aligned = 1-group windows) + a 96-row x 256-px dirty present; gameFrameClean = 3 16x16 copyRect erases + 3 draws + present; P1 = the 96-row x 256-px band alone (12,288 shadow bytes; 24,576 VGA bytes on DOS); P2 = three 16x16 windows (384 B; 768 VGA bytes). Clocks: ST 8.0 MHz, Amiga 7.09 MHz, IIgs 2.8 MHz; DOS has no cycle model -- only recorded rates are used.

Measured rate cards (replace the Phase-0 paper inputs)

  • ST (68000): bus floor 4.0 cyc/B copied; move.l inline 5.0 cyc/B; movem streams ~4.6-4.8 cyc/B (write-only fill ~2.2 cyc/B); landed movep tile bodies 36 cyc/tile-row (paste/snap) and 48 (copy); specialized whole-tile mark ~480 cyc; measured per-present idle base 0.87 ms (hunt-3 refresh: jlStagePresent full 1150/s post the group-skip landings; the old 2.2 ms / P8 452/s figure predates them -- every ceiling row below that adds "+ 2.2 ms base" is pessimistic by ~1.3 ms and ST P2's honest re-derived ceiling is ~775/s, which RE-OPENS that row). P1 machine ceiling ~100 ops/s (session model) carries the same staleness.
  • Amiga: blitter A->D ~2 chip cyc/word + ~230 cyc/plane setup + ~240 cyc Own/Disown; display-DMA contention ~2x on chip traffic (Phase-0 full-blit 110/s precedent); measured contended CPU tile body ~700 cyc/cell (backed out of the landed map walker, 191 maps/s = 742 cyc/tile incl. walk); P8 idle base 0.87 ms (hunt-3 refresh: jlStagePresent full 1143/s -- the old NEVER RECORDED note is resolved by the golden idle-dominated row post group-skips).
  • DOS: the W0 "~1.7 us/VGA byte -- pure hardware rate" model is REFUTED (hunt-3): the dword-fused expand measured probe present96 at 29 ops/s = ~1.40 us/B over 24,576 B, EXCEEDING the old ceiling (~24) -- under DOSBox's instruction-counting core the write loop was software-bound, not hardware-bound. DOS present ceilings should be re-fit from ~1.40 us/B; gameFrameClean is PASS-at-floor (245/s vs the probe-split-derived ~244 floor); CPU-side recorded models only: single-tile paste ~57,000/s (Phase-0, unrevised) and map walker ~76,000 tiles/s (W3: ~26 measured vs 18-20 hand instructions/cell).
  • IIgs: MAME-fitted write model (closed the P1/P2/P8 ledgers simultaneously): shadowed SHR push write wSh ~1.73 us/B, slow-I/O softswitch access ~1.0 us; measured per-present fixed base 2.75 ms (hunt-3 probe: presentIdle 363/s; the original 13.7 ms figure died with the band-memset fix -- any ceiling below still adding "+13.7" is pessimistic by ~11 ms). Real-HW datasheet variant: wSh ~0.977 us/B (fixed base assumed unchanged -- unverified on hardware). Tile hand-call floor 480 cyc; W1 sprite autopsy: honest full-op draw ~1,060-1,120 cyc (body ~670 + minimal per-row mark ~390 + 3-far-load derive + JSL; the fnAddr cache was built, measured a LOSS, and removed).

Atari ST (8.0 MHz 68000)

Row Ceiling (ops/s) Derivation Measured Gap Bar
jlTilePaste ~9,200 movep body 8x36=288 + mark ~480 + call ~100 = 868 cyc 4,685 2.0x FAIL
jlTileCopy ~8,300 movep body 8x48=384 + 480 + 100 = 964 cyc 3,626 2.3x FAIL
jlTileSnap ~20,600 body 288 + call ~100 = 388 cyc (no mark; measured overhead is ~651) 8,516 2.4x FAIL
jlTileMapPaste 10x5 ~440 50 x ~350 cyc/tile (movep-honest) + one mark 480 + one validation ~200 = 18.2k cyc 346 1.27x PASS
jlSpriteDraw 2,235-2,750 Phase-0 pre-shifted movem-RMW model, 2,909-3,580 cyc -- UNREVISED paper 1,341 1.7-2.1x FAIL
jlSpriteSaveUnder ~5,300 2-group window 256 B x 5.0 cyc/B + row stride ~230 = 1,510 cyc. P7-1 probe split (2026-07-09): full call 3,064 cyc = wrapper+validation 1,088 + dispatch/metadata ~660 + body 1,310; the window delta (g16 row) measured the body at 5.1 cyc/B = AT the move.l floor -- the whole gap is C glue 2,611 2.0x FAIL (lever: 68k wrapper trim/fusion)
jlSpriteRestoreUnder ~4,000 save copy 1,510 + sprite mark ~480 = 1,990 cyc. P7-1 probe: full 4,211 cyc, validation alone 422 (odd-bx row) 1,900 2.0x FAIL (same lever)
gameFrame ~58 fps fill 12,288 B x 2.2 = 27k + 16 pastes 13.9k + 3 saves 2.6k + 3 draws 9.3k (mid-range of the paper draw model) + 3 restores 4.1k + present96 80k = 136.9k cyc = 17.1 ms 32 1.8x FAIL
gameFrameClean ~245 3 copyRect x ~750 cyc + 3 draws 9.3k + P2-class present 20.8k = 32.4k cyc = 4.0 ms 90 2.7x FAIL
P1 present96 ~100 model: 12,288 B x 4.0 cyc/B = 6.1 ms + P8 base 2.2 ms + snapshot traffic in one fused 200-row scan/snapshot pass ~= 10 ms 73 1.37x PASS
P2 present3win ~385 measured idle base 2.2 ms (P8 452/s) + 384 B copy+snapshot x ~8 cyc/B ~= 2.6 ms 245 ~1.6x FAIL (marginal)

Amiga (7.09 MHz 68000 + blitter)

Row Ceiling (ops/s) Derivation Measured Gap Bar
jlTilePaste ~5,500 measured contended body ~700 + mark 480 + call ~100 = 1,280 cyc (uncontended instruction model gives ~6,100-6,400 -> gap 1.6-1.7; the Phase-0 18% pair discrepancy persists as this band) 3,753 ~1.48x PASS (at bar)
jlTileCopy ~4,800-5,300 chip-to-chip body ~750-900 + 480 + 100 cyc 2,903 1.7-1.8x FAIL (marginal)
jlTileSnap ~8,900 body ~700 + call ~100 = 800 cyc (no mark; writes go to fast RAM) 6,103 ~1.45x PASS (at bar)
jlTileMapPaste 10x5 ~200 50 x ~700 + mark 480 + validation ~200 = 35.7k cyc; the landed walker already beat its own design model (176-186 -> 191) 191 ~1.05x PASS
jlSpriteDraw ~1,350 Phase-0 compiled-CPU ceiling; the 3,600-4,300/s blitter cookie-cut is DEFERRED by recon -- sync WaitBlit forfeits the win at <= 16x16 (this row's size) 1,153 1.17x PASS
jlSpriteSaveUnder ~4,400-6,100 SUPERSEDES the blitter paper (6,300-7,400): P7-1 probe (2026-07-09) measured the CPU body at 3.8 cyc/B via the window delta (482 cyc/128 B), so body ~960-1,900 cyc (contention band) + call ~200. Full call 3,372 cyc = wrapper+validation 1,524 + dispatch/metadata ~880 + body -- the gap is C glue, not the copy 2,103 ~2.1-2.9x FAIL (lever: 68k wrapper trim/fusion)
jlSpriteRestoreUnder ~2,900-3,700 body + mark ~480 + call (same probe-derived band); full 4,423 cyc, validation alone 672 (odd-bx row) 1,603 ~1.8-2.3x FAIL (same lever)
gameFrame ~70 fps fill blit 13.4k x2 contention = 27k + 16 pastes 20.5k + 3 saves 3k + 3 draws 15.8k + 3 restores 4.4k (sprite components interpolated from the paper models) + present96 29k = 99.7k cyc = 14.1 ms 31 2.3x FAIL
gameFrameClean ~230 3 blit copyRects ~6.8k contended + 3 draws 15.8k + P2 present ~8k = 30.6k cyc = 4.3 ms 71 3.2x FAIL
P1 present96 ~245 4-plane modulo blit 4x(230 + 1,536 w x 2) + 240 = 13.4k cyc, x2 contention + union scan ~2k = 29k cyc = 4.1 ms (the landed P7-3 blitter flush now takes this path; CPU-copy variant was 12,288 B x 4 cyc/B + scan ~= 7.5 ms -> ~133/s) 78 ~3.1x FAIL
P2 present3win UNKNOWN (write-only bound ~890-1,300) 12 plane-blit setups 2.8k + 192 w x2 contention + Own/Disown + scan ~2k ~= 5.4-8k cyc = 890-1,310 ops/s; honest ceiling needs the Amiga P8 idle base = 661/s = 1.51 ms. Component split (2026-07-09 skip-mask probe): scan ~4.9 ms, flush ~3.3 ms (chip->chip CPU copy, below the 128-word blitter threshold), flip ~0.4 ms, residual ~1.2 ms. TWO levers landed: (1) presentScan.s 4-row fast path cut the scan (106 -> 128, +21%, after fixing a phantom-row-200 OOB); (2) the flush moved to the blitter by lowering AMIGA_PRESENT_BLIT_MIN_WORDS 128 -> 16 (chip->chip DMA vs contention-bound CPU copy): 128 -> 153 (+20%). Both 51/51 + P11 153 <= ~6-8x vs bound FAIL/UNKNOWN

DOS (AT-class VGA; no cycle model -- recorded rates only)

Row Ceiling (ops/s) Derivation Measured Gap Bar
jlTilePaste ~57,000 Phase-0 single-tile model -- UNREVISED (W3 walker data is consistent with it) 12,575 4.5x FAIL
jlTileCopy ~57,000 class same 64-B traffic as paste; no separate model recorded 10,910 ~5.2x FAIL
jlTileSnap >= ~57,000 class paste class minus the mark; no separate model recorded 19,218 ~3.0x FAIL
jlTileMapPaste 10x5 ~1,520 recorded W3 model: ~76k tiles/s (walker at 61,200 x 26/[18-20] hand instructions/cell) / 50 1,224 1.24x PASS (residue structural: map indirection + mark)
jlSpriteDraw ~11,000 Phase-0 model -- UNREVISED 8,146 1.35x PASS
jlSpriteSaveUnder 14,500-18,000 Phase-0 model -- UNREVISED 10,718 1.35-1.68x AT BAR (range straddles)
jlSpriteRestoreUnder 14,500-18,000 Phase-0 model -- UNREVISED (add mark: real ceiling somewhat lower) 7,826 1.9-2.3x FAIL
gameFrame ~22 fps present writes 24,576 B x 1.7 us = 41.8 ms + components ~2.5 ms ~= 44.3 ms (component figure not derivable from recorded rates -- no DOS fill rate; verdict insensitive: even at the measured P4 9.2 ms the ceiling is ~19.6 fps, gap ~1.03); W0 verdict: present-bound AT hardware 19 1.2x PASS
gameFrameClean UNKNOWN (upper bound ~500) erases+draws ~0.7 ms + present writes 768 B x 1.7 us = 1.3 ms; per-present fixed scan (~2.0 ms observed) never modeled 192 <= ~2.6x vs bound FAIL/UNKNOWN
P1 present96 ~24 24,576 VGA B x 1.7 us = 41.8 ms -- pure hardware rate 23 1.04x PASS (AT hardware)
P2 present3win UNKNOWN (upper bound ~766) 768 B x 1.7 us = 1.3 ms writes only; fixed base unmeasured (W0 closed DOS present work) 305 <= ~2.5x vs bound FAIL/UNKNOWN

Apple IIgs (2.8 MHz 65816; MAME-fitted write model)

Row Ceiling (ops/s) Derivation Measured Gap Bar
jlTilePaste ~5,830 recorded 480-cyc hand per-tile call 2,700 2.2x FAIL
jlTileCopy ~4,860 480 + ~96 cyc extra far-src reads = 576 cyc 1,981 2.5x FAIL
jlTileSnap ~5,800 480-cyc class (mark savings ~= far-read tax) 2,520 2.3x FAIL
jlTileMapPaste 10x5 ~140 50 x ~390 cyc/tile floor (no tracker source; within noise of the landed ~400-cyc/tile walker) + one mark ~= 19.9k cyc; the landed walker already runs BELOW the 480-cyc per-call ceiling 140 ~1.0x PASS
jlSpriteDraw 2,500-2,650 W1 autopsy honest full-op model: ~1,060-1,120 cyc = body ~670 + per-row mark ~390 + 3-far-load derive + JSL (fnAddr cache rejected by measurement) 1,142 ~2.2-2.3x FAIL
jlSpriteSaveUnder ~2,050 MVN window body ~1,111 cyc (128 B x ~7 cyc/B + row stepping; = Phase-0 2,520/s) + derive/JSL ~250 = 1,360 cyc 602 3.4x FAIL
jlSpriteRestoreUnder ~1,600 save body 1,111 + fused mark ~390 + derive/JSL ~250 = 1,750 cyc (matches the W1 "restore 1,600+" honest target) 662 2.4x FAIL
gameFrame ~16.5 fps (real HW ~20) P4 components ~20.4 ms (fill 6,144 w x 6 cyc = 13.2 + pastes 2.7 + saves 1.5 + draws 1.2 + restores 1.9) + present96 40.3 ms = 60.7 ms 12 1.37x PASS
gameFrameClean ~60 3 stage-copy erases ~1.2 ms + 3 draws ~1.2 ms + P2 present 14.4 ms ~= 16.7 ms 20 3.0x FAIL
P1 present96 ~25 MAME (real HW ~35) 15,360 B x 1.73 us = 26.6 ms writes + 13.7 ms measured base = 40.3 ms (landed Stage-B copies 13,824 B -> 37.6 ms -> ~27/s, gap 1.33x); minimal-copy variant (12,288 truly dirty B) 35.0 ms -> ~29/s, gap 1.43x -- PASS every way 20 1.24x PASS
P2 present3win ~70 (real HW ~71) 384 B x 1.73 us = 0.7 ms + 13.7 ms base = 14.4 ms 52 1.34x PASS

IIgs real-HW-datasheet variants (wSh ~0.977 us/B, MAME-measured 13.7 ms base assumed unchanged -- unverified on hardware): P1 15.0 + 13.7 = 28.7 ms -> ~35/s; P2 ~14.1 ms -> ~71/s; full 200-row present 31.3 + 13.7 = 45.0 ms -> ~22/s (vs ~14.5/s MAME-fitted); gameFrame ~49 ms -> ~20 fps. The MAME-fitted column is what the bench harness can verify; the real-HW column is the datasheet claim.

Rows still outside the 1.5x bar -- reasons and recorded levers

  • Tile singles, all four ports (1.7x-5.2x). Bodies are at machine floor (probe-validated splits, restated against the 2026-07-08 capture: ST paste overhead = 1,708 - 288 = 1,420 cyc, snap overhead 651 cyc with no mark; the W3-landing probe read 1,388/669); the gap is the per-call public wrapper + validation + whole-tile mark (Phase-2 exit finding: "remaining per-call floor = public wrapper + one JSL/call"). Recorded answer is the batch API, which PASSES on every port (1.0x-1.27x); no further singles lever is recorded -- the singles gap is accepted per-call API tax (absolute cost <= ~0.3 ms/call everywhere; IIgs tileCopy ~0.3 ms and tileSnap ~0.22 ms are the worst cases, all other rows <= ~0.2 ms).
  • IIgs sprite rows (2.2-2.3x/3.4x/2.4x). Autopsy buckets: validation ~22% + wrapper residue in the register-starved public wrappers; fusion already landed (+18-22%), the fnAddr cache measured a loss and was removed. VALIDATION LEVER LANDED 2026-07-12 via the trusted fast-path API (jlSprite{SaveUnder,RestoreUnder,SaveAndDraw} Trusted, extended from the 68k P7-1 work): the trusted entries skip the ~22% validation and route into the existing fused MVN entries. IIgs save 602->665 (+10.5%), restore 662->722 (+9%), saveAndDraw 423->455 (+7.6%); shrParity badRows=0, 51/51 golden (safe entries unchanged). Smaller than the 68k wins because 65816 validation is a smaller fraction. Do not resurrect the fnAddr cache without a compiler-level win (LLVM816-ASKS candidate).
  • ST sprite rows (~2.0x) / Amiga save-restore (~1.8-2.9x) / DOS restore (1.9x-2.3x). The P7-1 layer-split probe (2026-07-09) RESOLVED where the 68k gap lives: the copy bodies are AT the machine floor (window-delta-measured 5.1 cyc/B ST / 3.8 cyc/B Amiga), and ~55-70% of every call is the C wrapper+validation (1,088 cyc ST / 1,524 Amiga for save; 422/672 for restore's validation alone) plus dispatch/backup-metadata glue (~660-880 cyc). The old paper ceilings (Amiga blitter 6,300-7,400) are superseded by probe-derived CPU ceilings in the tables above; true gaps ~2x. LEVER ATTEMPT 1 (macro tightening, 2026-07-11): REVERTED, net loss. Both class macros were rewritten to their algebraically-minimal form, proven value-identical (102k save + 1.1M restore cases, 0 mismatches) and gated 51/51 -- benched save bit-identical on both ports, restore neutral on ST but -3% on Amiga (1603->1553; gcc-amigaos lowered the "simpler" C to +3 instructions). The wrapper is gated by register allocation (movem of 10-11 regs + interleaved validation), not arithmetic op count -- no C micro-opt moves these rows, same compiler-bound wall as the IIgs fnAddr-cache autopsy. LEVER ATTEMPT 2 (restore+draw mark fusion, 2026-07-11): LANDED, +5-6% on restore. Isolated the mark at ~26-29% of a 68k restore (probe: restore to a non-stage surface skips it). Three builds: a per-port body+mark asm shim REGRESSED both ports (broke the always-inline body's shared loads + added a call layer); a lean rolled-loop asm widen call REGRESSED (rolled dbra lost to gcc's unrolled indexed mark); INLINING the unrolled 16-row mark into jlSpriteRestoreUnder for the h==16 stage window (C only, no asm) WON -- ST restore 1900->1996, Amiga 1603->1703, both 51/51 + P11/P10/P5 parity. The recoverable win was the CALL BOUNDARY (only inlining removes it; asm always keeps a jsr). Draw bundled via the same inline (jlSpriteDraw + SaveAndDraw): Amiga draw +4% / saveAndDraw +6% / unaligned +5%; ST draw neutral (masked body dwarfs the mark); DOS untouched (guarded). gameFrame flat (present-bound). LEVER ATTEMPT 3 (TRUSTED fast-path entry, 2026-07-11): LANDED, the biggest 68k sprite win yet. New opt-in public API (jlSpriteSaveUnderTrusted / RestoreUnderTrusted / SaveAndDrawTrusted) that skips the per-call geometry-validation chain for a caller that pre-compiles its sprites and clamps positions on-screen (UB-if-violated contract; falls back to the safe sibling for anything unrecognised; Amiga/ST fast path, DOS just calls the safe entry). WINS ON ALL THREE COMPILED PORTS: save ST 705->817 (+16%) / Amiga 657->753 (+15%) / IIgs 602->665 (+10.5%); restore ST 529->609 (+15%) / Amiga 545->609 (+12%) / IIgs 662->722 (+9%); saveAndDraw ST 257->289 (+12.5%) / Amiga 273->289 (+6%) / IIgs 423->455 (+7.6%). On 68k the win EXCEEDS the raw ~422/672 cyc validation because dropping the checks also shrank gcc's frame/spills; IIgs routes into the fused MVN entries minus validation. Parity gates pass (P10 ST, P11 Amiga, shrParity IIgs); DOS 51/51 + IIgs 51/51 (falls back / safe entries unchanged). Batch APIs stay descoped (dispatch alone ~3-6%); the blitter cookie-cut stays deferred until a >= 32x32 golden row exists.
  • Amiga presents P1/P2 (+ both Amiga composites, 2.8x-5.3x; P2 bound-only). W2 fixed the bounding-box over-copy (P2 copied bytes 4,992 -> 384 exact; W0's conviction figure was 3.5x vs the ~1,536 truly-dirty estimate) but the ~11 ms per-present fixed scan survives (P2 = 11.1 ms to move 384 B; P1 even regressed 8% on added scan cost -- accepted deviation). Recorded levers: fused-scan present asm on the ST presentSlam pattern (W2 follow-up) and a modulo-blitter present flush (the landed W3 copyRect blitter arm is the template). This is the single largest remaining item in the program.
  • ST P2 (~1.6x, marginal) and ST composites (1.8x/2.7x). P2 residue is ~1.5 ms absolute (per-run move.l-chain entry + snapshot maintenance); W2 closed ST present work and no lever is recorded. gameFrame's gap is spread across sprite/fill wrapper taxes (no single lever); gameFrameClean's was dominated by jlSurfaceCopyRect's strided per-row memcpy dispatch -- the same overhead class the Amiga W3 blitter item deleted. LANDED 2026-07-12 (#3): an ST move.l-chain copyRect (src/atarist/copyRect.s, stCopyRectSpans68k) replaces the per-row memcpy CALL for short spans (the 48 8-byte-row window copies where mintlib memcpy's frame dwarfs the two move.l; long rows keep memcpy past the 64-byte crossover). gameFrameClean 90 -> 120 ops/sec (+33%), 51/51 + P10 displayMismatch=0; gameFrame unchanged (uses save/restore, not copyRect). Bigger than expected -- the small-window memcpy call overhead dominated the clean-buffer op.
  • IIgs gameFrameClean (was 3.0x). Recorded forensic: ~30 ms/frame unattributed, llvm-mos far-memcpy byte loops in the generic copyRect erase engine (finding #79 class). LEVER LANDED 2026-07-12: an IIgs-specific jlpSurfaceCopyRect override (src/iigs/hal.c, gated JL_HAS_SURFACE_COPY_RECT) does an inline word copy for short spans -- compiles to lda [dp],y / sta, NO far-memcpy call (verified in the -S output) -- for the 8-byte-row clean-buffer windows; long rows keep memcpy. gameFrameClean 18 -> 24 ops/sec (+33%), 51/51 + shrParity. NOTE: this was FIRST tried in the shared generic copyRect but that REGRESSED DOS -12% (x86 rep-movs memcpy already beats an inline word loop for 8-byte rows), so it is IIgs-only; DOS stays on generic memcpy. Remaining IIgs gameFrameClean gap (if any) would want the flagged llvm816 memcpy fix or an MVN block-copy for the long rows.
  • DOS gameFrameClean / P2 -- RESOLVED 2026-07-13 (perf-hunt-2 ranks 4/5/6). The missing display gate now exists: P12 vgaMismatch (an UBER_PROBE row + dosProbeVgaMismatch in src/dos/hal.c) re-derives every VGA byte independently of the expand LUT and reads 0/0/0 across the battery. The per-present scan was cut by the 4-row dword group skip (P8 idle 1137 -> 3881 ops/s, ~0.26 ms base) and the expand loop dword-fused (two LUT words per 32-bit store; DOSBox's instruction-counting core rewarded it directly): gameFrameClean 196 -> 243 ops/s (+24%), gameFrame 19 -> 23 (+21%), both 51/51. The probe component split answers the old unknown: gameFrameClean at 244/s = 4.10 ms/frame = present 3.06 ms + window erases 0.70 ms + sprite draws 0.34 ms -- the row is present-dominated and the remaining present cost is expand + per-row glue at the measured rate, not an unexplained residual. No further DOS present lever is recorded. (-fomit-frame-pointer also landed for flag parity with the 68k ports: input rows +28-29%, map paste +7%, zero regressions.)

Assumptions

  1. Clocks: ST 8.0 MHz, Amiga 7.09 MHz (PAL), IIgs 2.8 MHz; DOS has no cycle model, so DOS ceilings use only the recorded rates (1.7 us/VGA byte, 57k paste, 76k tiles/s walker).
  2. Irreducible per-call floors assumed: ~60-100 cyc call/shim on 68k, ~250 cyc JSL+derive on IIgs; the landed specialized dirty marks (~480 cyc 68k whole-tile, ~390 cyc IIgs fused) are treated as irreducible for row shapes that mark per call.
  3. Amiga display-DMA contention factor ~2x on chip-RAM traffic, taken from the Phase-0 full-blit 110/s precedent; uncontended variants are shown wherever they would change a verdict.
  4. The IIgs fixed present base was 13.7 ms MAME-measured (P8) when the ceilings below were derived. 2026-07-13: ~11 ms of that base was the dirty-band memset pair (see History) and is now ~2.75 ms -- ceilings that add "+ 13.7 ms base" are stale and pessimistic; re-derive with ~2.75 ms before gating new work on them. The real-HW column assumption (base carries over unchanged) is still unverified on hardware.
  5. Composite ceilings are sums of component ceilings at the rows' exact geometry (per-op sprite rows at x=40 use the 2-group 256-B 68k window; gameFrame's mod-16-aligned sprites use the 1-group 128-B window). Composite sprite components not separately recorded (ST draws, Amiga saves/restores/copyRects) are interpolations inside the quoted Phase-0 ranges.
  6. Ceilings labeled UNREVISED (ST/Amiga/DOS sprite models, DOS tile singles) are Phase-0 paper carried forward because no wave re-measured them; their gaps are quoted as ranges and should be re-derived before gating any new work on those rows. The ST P1 ~100 ops/s machine ceiling is a session model consistent with the P8 base + bus-floor copy, not a tracker-recorded number.
  7. Where no model or measurement exists (Amiga/DOS P8 idle bases, DOS copy/snap singles), the cell is an explicit upper bound or UNKNOWN -- no ceiling was invented.

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

  • 2026-07-20 (later still): music layer v2 = the complete feature set. jlAudioNoise percussion voice implemented natively on all four ports (ST YM noise generator on channel C with shared-volume guards both directions + a Supexec'd PSG all-off in shutdown; DOS OPL2 rhythm-mode snare on ch 7 with pitch via the hi-hat modulator and a shared adlibFreqToFnum factored out of jlpAudioVoice; Amiga 1024-byte LFSR chip sample looped on Paula ch 3, mt_musicmask re-applied after every mt_end; IIgs resident LFSR sample on SFX slot 3 with retrigger phase rotation -- and a real guard fix: jlpAudioVoice bounded by SFX_SLOTS could stamp tone squares over slots 3-4). Sequencer v2: noise as uniform voice 3, tempo-change opcode, jlMusicSetAtten fades, jlMusicPause/Resume. songbake v2: noise column, repeat blocks, mid-song tickms, ST-overlap warning. AGI's always-parsed-never-sounded SN76489 noise channel is now audible via psgNoiseApply. Design recon = 10-agent workflow (5 design + 5 adversarial verify) that caught 3 hardware-level bugs pre-implementation. Gates: builds x4, AGI host suite 1/1, sequencer host harness all-checks, golden 51/51 x4 cross-diff.

  • 2026-07-20 (later): portable chip-tracker music layer landed (the cross-platform answer to the ST MOD tax -- one authored song, four native chip backends, no CPU mixing anywhere). New: src/core/music.c jlMusicPlay/Stop/IsPlaying sequencer over the JYM1 event stream (tools/songbake bakes text notation; pacing = jlMillisElapsed so tempo holds at 50/60/70 Hz; pumped by jlAudioFrameTick above the gAudioReady gate). Hunt-3 rank 4 landed as the IIgs backend: jlpAudioVoice rewritten from synthesize-per-call (~120-150 ms: C toneFill + forced 2 KB XOR recopy, plus it burned that cost even with audio uninitialized) to a resident DOC-unsigned square with freqWord/length/volume struct pokes + one NTPstreamsound JSL per note (~10 far stores). Amiga jlpAudioVoice implemented for the first time (was a stub): Paula ch 0-2 looping a 2-byte chip-RAM square, DMA auto-reload sustains for free. Voices vs MOD stay mutually exclusive on ST/Amiga (same chip). jlAudioFrameTick gains a no-song boolean check (musicPump) -- DOS row 57969 ops/s, unchanged within noise. Gates: 4-port builds clean, DOS/ST/Amiga/ IIgs golden 51/51 cross-diff OK.

  • 2026-07-20: hunt-3 top ranks landed. Rank 1 probe re-baseline (IIgs+ST -DUBER_PROBE battery): the IIgs gameFrameClean mystery SOLVED -- the split rows (newly wired into the IIgs timeOp arm) attributed the 28.6 ms as erases 15.9 + draws 3.3 + present 9.4, exposing jlSurfaceCopyRect at ~5.3 ms PER 16x16 call: the Phase-7 C word-copy override had regressed into the dead-lever-#12 shape (llvm-mos re-deriving both far pointers per word, ~30+ instr/word, visible in the current -S). Fix = iigsCopyRectRows asm ([dp],y pairs, ~21 cyc/word): gameFrameClean 35 -> 56 ops/s (+60%), hash C51FE171 unchanged, 51/51. Rank 5: jlpInputPoll idle clear memset -> iigsByteFill: jlInputPoll 228 -> 662 ops/s (+190%). Rank 3: AGI restoreRectFromBackdrop -> jlSurfaceCopyRect (the stale "planar = rare (Amiga only)" fallback full-copied 32KB per object erase on BOTH 68k ports and forced full-dirty presents); builds clean x4, host AGI suite 1/1 (harness stub rot also repaired). Rank 2: rate cards refreshed in place (IIgs base 2.75 ms, ST idle 0.87 ms, Amiga idle 0.87 ms recorded for the first time, DOS 1.7 us/B model REFUTED -- measured 29/s EXCEEDS the old ceiling, refit ~1.40 us/B; ST P2 RE-OPENED at ~775/s re-derived ceiling). IIgs probe battery also refreshed: present96 26/s, present3win 133/s, trusted save 795/s, drawTrusted 1262/s, shrParity OK, band fingerprints cross-port identical.

  • 2026-07-19: perf-hunt-2 close-out. Ranks 8/9/10/12/13/14/17 landed (ST draw-JIT group rewrite: jlSpriteDraw 1366 -> 1807 +32%, saveAndDraw +21%, gameFrameClean +14%; jlSpriteDrawTrusted on all ports; IIgs fused save: saveUnder 602 -> 722 +20%; ST PS_ROW inline arm; DOS inline-mark + full trusted family: safe restore +5%, trusted restore +21%; Amiga fillCircle exact-disk walk 31 -> 40 +29%; cleanup batch incl. 115 lines of unreachable ST walkers). Rank 11 v2 landed after a re-probe (presentScan.s 50-fixed-group restructure; Amiga gameFrameClean 91 -> 97 +6.6%; the re-measured split: scan 3.24 / flush 1.61 / flip 0.36 / residual 1.33 ms of a 6.54 ms present3win). Ranks 15/16 killed by their own gates. P7-2 stage 2 (metadata fills into a 10-arg asm entry) was built, measured 722 -> 602, and REVERTED -- llvm-mos argument marshalling costs more than the far stores it replaces (the fnAddr-cache wall; P7-2 is closed, both directions measured). All landings 51/51 cross-port + P5/P11/P12 parity.

  • 2026-07-13 (later): perf-hunt-2 ranks 2/4/5/6/7 landed. ST+Amiga idle-present group scan (4 rows per longword vs the 0xFF clean sentinel): jlStagePresent full ST 407 -> 1150/s, Amiga 609 -> 1131/s; composites unchanged. DOS: -fomit-frame-pointer (flag parity; input rows +28-29%), 4-row group skip in jlpPresent (P8 idle 1137 -> 3881/s), dword-fused VGA expand (gameFrameClean 196 -> 243, gameFrame 19 -> 23), NEW P12 vgaMismatch display gate (dosProbeVgaMismatch + UBER_PROBE battery, 0/0/0), and the gameFrameClean component split (present 3.06 / erases 0.70 / draws 0.34 ms). All four ports 51/51 cross-diffed after each landing.

  • 2026-07-13: IIgs dirty-band memset fix (perf-hunt-2 rank 1+3). The "13.7 ms irreducible present base" was mostly two libc memset far calls: stageDirtyClearAll (runs after EVERY present) memset both 200-byte band arrays, and llvm-mos lowers memset to jsl memset -- libc's while (n--) *d++ = c byte loop (the old "MVN-seeded" comments were wrong; finding-#79 class). Replaced with asm word fills in joeyDraw.s (iigsStageBandsClearAll / MarkAll, abs,x descending; a C word-fill emitted ~35 instr/store and was rejected) plus iigsByteFill ([dp],y) for jlScbSetRange. Measured (MAME): jlStagePresent full 67 -> 363/s (fixed base ~13.7 -> ~2.75 ms), jlScbSetRange 163 -> 1322/s, jlSurfaceClear 33 -> 56/s, gameFrameClean 24 -> 35/s, gameFrame 12 -> 14/s. 51/51 golden vs Amiga reference; verify-iigs UBER framebuffer checksum byte- identical to the memset build. NOTE: every IIgs ceiling that bakes in the 13.7 ms base (P1/P2 rows, assumption 4, the rate-card fixed base) is now stale on the optimistic side for the library and needs re-derivation with the ~2.75 ms base; the 2026-07-08 baseline table rows for the five ops above are superseded by the numbers here.

  • 2026-07-08: Phase 6 W1+W2+W3 landed (fused IIgs sprite/fillRect entries + sprite-mark specialization; per-row/fused presents on Amiga/ST/IIgs with display-parity probe gates P8-P11; ST movep tile singles, DOS map walker, Amiga blitter copyRect + map hoist). Table regenerated from a clean rebuild, 51/51 x4. Ceilings re-derived from the measured models (fitted IIgs write cost, ST bus floor, movep/blitter rates) -- the old ~29 ops/s IIgs full-screen note is superseded in place.

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).