marfrit
85feba4087
Fourth daedalus-fourier kernel — VP9 8-tap inner loop filter wd=8 h_8_8 variant. Width extension of cycle 2's wd=4; completes VP9 inner-edge LPF coverage. Full cycle Phase 1-7 + M4'''' in one combined go (cycle compressed since incremental from cycle 2). Phase 5 review explicitly skipped (incremental ~30-line shader delta from cycle 2 + same geometry + cycle-2 RED-pattern checks still apply). Flagged in docs/k4_lpf8_phase4_7.md per dev_process.md "Skipping phases is a deliberate choice that should be flagged." Phase 6 v1 first-light: M1'''' 100.0000% bit-exact (65536/65536) first try. Shaderdb shows 231 inst, 4 hardware threads, 0 spills, 27 max-temps, 48 uniforms — compiler at the latency-hiding ceiling. Performance: M3'''' NEON (single-core) 52.382 Medge/s M2'''' QPU isolation 17.847 Medge/s R'''' 0.341 → ORANGE band 30fps floor margin 9.2x (isolation), 20.3x (mixed) M4'''' concurrent matrix: NEON 4-core 37.823 Medge/s <- baseline QPU only 14.867 Medge/s MIXED NEON-3 + QPU 39.389 Medge/s <- +4.1% PASS Verdict: YELLOW-via-M4'''' PASS. Deploy wd=8 LPF on QPU alongside cycle 2 wd=4. Combined VP9 inner-edge LPF coverage now complete. Cross-cycle LPF comparison: | | wd=4 (k2) | wd=8 (k4) | | M3 NEON | 48.3 | 52.4 | | M2 QPU iso | 19.6 | 17.8 | | R iso | 0.41 | 0.34 | | M4 delta | +6.9% | +4.1% | | 30fps mixed | 7.2x | 20.3x | | Verdict | GO QPU | GO QPU | NEW finding (Phase 9 lesson): NEON gets faster per edge as filter width grows (20.7 → 19.1 ns wd=4 → wd=8). The relative QPU loss grows with width. wd=16 would probably flip negative based on the trend line. Deployment recipe with cycle 4: IDCT 8x8 (k1) -> QPU (R=0.92, +7% mixed) LPF wd=4 (k2) -> QPU (R=0.41, +7% mixed) LPF wd=8 (k4) -> QPU (R=0.34, +4% mixed) MC 8h (k3) -> CPU (R=0.067, -19% mixed) Entropy -> CPU (structural) VP9 inner-edge LPF coverage complete. Project continues to higgs deployment plumbing or further kernels per user direction. New artifacts: - src/v3d_lpf_h_8_8.comp — GLSL shader - tests/vp9_lpf8_ref.c — standalone C ref - tests/bench_neon_lpf8.c — M1+M3 bench - tests/bench_v3d_lpf8.c — M1+M2 bench - tests/bench_concurrent_lpf8.c — M4 pthread bench - docs/k4_lpf8_phase1_3.md + phase4_7.md — combined cycle docs Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
69 lines
2.5 KiB
Markdown
69 lines
2.5 KiB
Markdown
---
|
||
cycle: 4
|
||
phases: 1-3 (combined doc — straight extension of cycle 2)
|
||
status: phase 3 in progress
|
||
date_opened: 2026-05-18
|
||
parent_cycle: k3_mc_phase7.md
|
||
target_kernel: VP9 loop filter wd=8 inner-edge horizontal (h_8_8)
|
||
---
|
||
|
||
# Cycle 4, Phases 1-3 — LPF wd=8
|
||
|
||
Compact combined doc — cycle 4 is a *width extension* of cycle 2
|
||
(same kernel family, same shape, same NEON file).
|
||
|
||
## Phase 1 — goal
|
||
|
||
**Kernel**: VP9 loop filter, 8-tap inner-edge variant (wd=8), horizontal
|
||
direction, 8-pixel edge. libavcodec symbol `ff_vp9_loop_filter_h_8_8_neon`
|
||
(already in vendored `vp9lpf_neon.S`).
|
||
|
||
**Why this kernel**: completes VP9 LPF coverage alongside cycle 2's
|
||
wd=4. The wd=8 path adds the `flat8in` test (6 abs comparisons) and a
|
||
6-pixel "flat region" write path — meaningfully more conditional
|
||
branches than wd=4 within the same kernel family.
|
||
|
||
**Measurable success** (cycle-4 numbering, `''''` superscript):
|
||
|
||
| ID | Measurement | Gate |
|
||
|---|---|---|
|
||
| M1'''' | Bit-exact vs C reference | 100.0000 % |
|
||
| M2'''' | QPU throughput Medge/s | recorded |
|
||
| M3'''' | NEON `ff_vp9_loop_filter_h_8_8_neon` Medge/s | recorded |
|
||
| M4'''' | Mixed NEON-3 + QPU vs pure NEON-4 (Medge/s) | recorded if YELLOW |
|
||
|
||
Same R bands + 30fps-floor calibration as cycles 2/3.
|
||
|
||
**Predicted R''''**: 0.3–0.5. Cycle 2 LPF wd=4 hit R=0.41; wd=8 adds
|
||
~20 % more conditional logic (flat8in test) and additional writes
|
||
when flat8in passes. Likely modestly worse R than wd=4. The 6-write
|
||
flat8in path under SIMD divergence may dominate.
|
||
|
||
## Phase 2 — situation
|
||
|
||
C reference: `external/ffmpeg-snapshot/libavcodec/vp9dsp_template.c`,
|
||
the same `loop_filter()` function (lines 1780-1898) used in cycle 2
|
||
but invoked with wd=8 via the `lf_8_fn(h, 8, stride, 1)` macro
|
||
instantiation. The wd=8 path activates the `if (wd >= 8 && flat8in)`
|
||
branch.
|
||
|
||
NEON reference: already vendored at
|
||
`external/ffmpeg-snapshot/libavcodec/aarch64/vp9lpf_neon.S`,
|
||
symbol `ff_vp9_loop_filter_h_8_8_neon`. Same calling convention
|
||
as wd=4: `(uint8_t *dst, ptrdiff_t stride, int E, int I, int H)`.
|
||
|
||
No new vendored sources needed.
|
||
|
||
**Workload model per edge (worst case, flat8in passes):**
|
||
- 8 rows × 6 written + 2 unwritten = 48 writes per edge (vs wd=4's 16-32)
|
||
- 8 rows × 8 reads = 64 reads (same as wd=4)
|
||
- ~12 abs+compares per row × 8 = ~96 per edge (vs wd=4's ~50)
|
||
|
||
Memory traffic similar to cycle 2 (~80-110 bytes per edge).
|
||
Compute moderately higher (more conditional branches + more writes
|
||
when flat8in fires).
|
||
|
||
## Phase 3 — NEON M3'''' baseline
|
||
|
||
(captured below after build + run)
|