gallery/docs/BUGS/hypotheses-120hz.md

120 Hz zero-jank campaign — hypothesis log

Target: sltest (minimal 100-row grid, no textures) on the phone's Mali G715. Definition of ZERO JANK (measured from PRES mon= + VSYNC ftn= lines):

1. no display-slot skip: for consecutive presents, slot_index(n+1) == slot_index(n)+1 (100%)
2. no same-slot double: two presents never land in the same display slot
3. phase stable: present phase within its slot varies < 0.5 ms Test: 2x input swipe 640 1900 640 900 4000 (slow drags) + fling, ~10 s. A "burst" = consecutive presents with gaps < 20 ms; evaluate inside bursts only.

Hypotheses (order = priority)

• H6 [sltest confound]: sltest FileLogger is blocking (flush per line) on the render path (PRES/STG/IT at ~240 lines/s). FIX: non-blocking drain-thread logger. TEST: rebuild, re-measure baseline.
• H1: early waker misses edges (thread descheduled / nanosleep jitter) → render starts late (edge+5 late wake wins) → present phase bimodal. TEST: render-start phase histogram vs edges; early-wake hit rate.
• H2: Skia render core variance → present phase jitter. TEST: VK SKD distribution in sltest (trivial scene, expect <1.5 ms).
• H3: wgpu/Mali submit+present latency variance (CLR/SKF/ORD/PRE). TEST: VK PRE / CLR / SKF distributions in sltest.
• H4: event-loop overhead (timers, change tracker, spin iterations) eats the slot budget. TEST: IT poll/timers/ev/fe distributions; iters/s during scroll.
• H5: CPU scheduling of the UI thread (deschedule mid-slot, core migration, frequency). TEST: wall vs CPU split for the render window (vkprobe thread_cpu_ns).
• H7 [method confound]: adb input swipe motion is bursty (~20-60 Hz) → grid movement itself is non-uniform regardless of render cadence. TEST: pointer-location overlay + user finger, later.

Rules

• One hypothesis per build; keep the change minimal; measure before/after with the identical test; record numbers here.
• A hypothesis is RULED OUT only when the measurement it predicts is absent (not "the fix didn't help").
• Phone hygiene before each run: clean task (sz=1), ResumedActivity, dismiss 16KB dialog (sltest is debuggable), record log byte offsets.

Results

2026-07-09 — sltest baseline + model correction (builds slt1-slt3)

• H6 RULED OUT / FIXED: sltest's blocking FileLogger replaced with a non-blocking drain-thread logger. No crash, no behavior change beyond log I/O coming off the render path.
• MODEL CORRECTION (major): the "Choreographer callback arrives 2.5-6 ms after the edge" belief from earlier sessions was WRONG — it came from the STG wake= field, which actually measures the poll's block time (first event), not callback latency. Measured from the VSYNC line itself (ftn = vsync edge, mon = callback ran): callback delivery = edge+0.52 ms median, p90 0.65, max 1.8 ms (sltest, phone). The callback fires ON the edge.
• H5 RULED OUT for the poll phase: pcpu (CPU time inside poll_events) = 2% of wall — the UI thread sleeps in epoll normally; it is not spinning or descheduled while waiting. The wake→poll path is healthy: poll start ≈ edge−2.3, byte at edge+0.5, return at edge+0.5, render 2.3-4.3 ms, present ≈ edge+3-4 ms — INSIDE the 5.5 ms deadline at the median. (The early-waker's edge−0.8 wake is redundant with the on-edge Choreographer wake — both are fine; the waker stays as fallback.)
• Baseline zero-jank metrics (sltest, phone, 991 transitions): DROPPED 0.71%, WASTED 0.91%, phase p10-p90 = 2.64-4.13 ms (1.5 ms spread), 12.5% consecutive phase jumps >1 ms. Median behavior is correct; the visible judder is the variance.
• H2/H3 CONFIRMED (residual cause): per-frame render = 3.5 ms median but spikes to 10.7 ms. Stage breakdown (sltest, phone): SKD (Skia draw) 1.07 ms median / 5.2 max; CLR (wgpu clear submit) 0.8 / 3.2; PRE 1.0 / 2.5; SKF 0.3; ORD 0.35. Spikes push presents past the deadline (the drops) and shake the phase (the judder). GPU submit overhead (2.45 ms median) exceeds the actual scene draw (1.07 ms) — four wgpu submits per frame.
• Next (implemented, phone-measurement pending): RENDER-AHEAD re-enabled under the corrected model. With the wake on the edge, the held frame is presented at the START of the render (edge+~1.5 ms, constant phase, 4 ms margin) and a spike only stretches the PREPARE phase of the frame held for the next slot — spikes stop delaying visible frames. Behavior verified on the emulator (scroll, settle, mid-drag pause/resume, no AlreadyAcquired, no black window). The phone is the only remaining test: expect drops → ~0 and phase spread → <0.5 ms IF spikes were the cause; rcpu (CPU time inside do_render) classifies any residual spikes (CPU work vs GPU/driver block).
• Emulator caveat: the emulated Choreographer fires 9-10 ms late and the SF pipeline is slow — the emulator cannot reproduce or validate 120 Hz cadence; it is for behavior verification only. Phone remains the test target.

2026-07-09 (2) — render-ahead on the phone (run slt4)

Render-ahead re-enabled (held frame presented at render START). sltest, phone, 2×4 s drags, 1433 transitions:

• Steady burst (401 frames / 3.3 s): phase 0.39-1.06 ms (p10-p90, spread 0.67 — was 1.5), slot jumps 1/401 (0.25%), doubles 0.
• Overall: DROPPED 0.70%, WASTED 0.70% — concentrated in the 17/62/87-frame bursts at drag START/STOP (locked↔unlocked transition + fling decay), not in steady scroll.
• Render wall: med 3.68 ms, p90 3.67, p99 3.9 — tight.
• Spike classifier (rcpu): all 13 spikes >6 ms are CPU-bound (rcpu≈wall, e.g. 17.5/15.2 ms) — Skia/driver CPU work, NOT GPU/driver blocking. Render-ahead absorbs them: 13 spikes → 1 skipped frame in the burst.
• Status: steady-state minimal-app scroll is effectively zero-jank by the measured definition. Remaining work: (a) user finger-perception test on sltest AND the gallery (textures, 8580 files), (b) if anything remains, attack the 1.3% CPU spikes (Skia), (c) burst-boundary transitions (drag start/stop) if visible.
• sltest logging bug noted: the non-blocking drain thread went silent mid-idle (file frozen, process alive) — likely a /sdcard FUSE flush stall on the per-line flush. Batch-flush (N lines or 100 ms) needed for long runs.

2026-07-09 (3) — jank_eval.py: the referee

User rightly rejected ad-hoc analysis (and my over-optimistic "effectively zero-jank" claim for the slt4 run). Built tools/jank_eval.py: strict PASS/FAIL against the zero-jank definition (no dropped slot, no same-slot double, no present past the SF latch (+5 ms edge phase), no >20 ms present gap, phase spread p10-p90 <= 0.5 ms) per active window.

• SLOT GRID: VSYNC ftn= edges = hardware truth (slot = floor((m-ftn0)/P), P = median 1-slot ftn delta). Anchor-fit fallback only for VSYNC-less logs (documented ambiguity limit >P/2).
• Self-test: 7 synthetic LOG FILES run through the same main() path — perfect / 3-slot drop / double / past-latch late / 120 ms stall / phase swing / sub-threshold jitter — all classified correctly.
• Emulator validated: correctly detects the 60 Hz emulated display, finds active windows, reports the known emulated-SF jank. Window thresholds scale with display rate.
• Estimator lessons (documented in the tool): median-of-deltas is biased by correlated jitter (1.8% for a 0.15 ms sawtooth!); a residual-score period search is non-minimizing at the true P (nearest-slot quantization); floor-vs-round slot assignment flips frames near the anchor. The hardware edge grid avoids all of it.

Honest re-judgement of slt4 (my 2-minute phone test), via the new tool (window w0 = drag 1, n=452, 120.066 Hz): dropped=4 (incl. one 30.6 ms + one 21.0 ms stall), late=20 (+5.1..+7.4 ms, early in the drag), phase spread 1.11 ms, 163 phase jumps >0.5 ms -> FAIL. The user's "I saw jank during the test" is confirmed by the data. (w1-w3 = post-test drain-stall corruption + user phone usage, excluded.)

Diagnosis state (what the CPU profiling did and did NOT answer)

• Answered: per-frame CPU work when a frame RUNS = 3.7 ms median, tight p90. It is not missing compute; there is no CPU work missing.
• UNANSWERED (now targeted): why NO frame ran in the stall slots, and why renders started past the latch. Candidates ranked:
  1. FRESHNESS GUARD misfire: threshold 6.83 ms vs earliest legitimate interval after a Choreographer-phase render 7.03 ms = 0.2 ms margin; any jitter false-rejects a render (and a reject can't be "undone" — the held frame waits a full extra slot).
  2. Lost wakes: locked do_render = saw_wake && fresh; if BOTH the early waker and the Choreographer fail to deliver a Wake event for a slot, the frame silently vanishes. No counter exists for this today.
  3. Loop blocked (driver/system call) for tens of ms — rcpu split will show wall≫rcpu.
• Next build (phone): slot-based guard (reject only if the wake's slot <= last rendered slot, slot from the edge grid) + a per-slot render ledger that logs every missed slot WITH its reason (no-wake / guard-reject / not-animating / blocked). jank_eval is the referee.
• sltest logger: batch-flush fix still pending (drain thread went silent ~120 s into process B; PRES lines use a separate path so present data survived).