GPU-accelerated motion magnification — implementation tracker

[joeljose]

Feature: GPU-Accelerated Motion Magnification

PRD: #10
Design Doc: docs/design/gpu-acceleration.md

Adds --gpu flag to motion_mag.py for ~17x speedup via pytorch_wavelets + cuFFT. Includes CPU improvements (FFT filter optimization already applied, wavelet filter upgrade).

Slices

• Slice 1: Switch default wavelet filters + add --biort/--qshift flags #12 — Slice 1: Switch default wavelet filters + add --biort/--qshift flags
• Slice 2: GPU infrastructure — Dockerfile, deps, --gpu/--device flags, memory check #13 — Slice 2: GPU infrastructure — Dockerfile, deps, --gpu/--device flags, memory check
• Slice 3: GPU Pass 1 — batched forward DTCWT + phase extraction #14 — Slice 3: GPU Pass 1 — batched forward DTCWT + phase extraction
• Slice 4: GPU temporal filter — chunked cuFFT #15 — Slice 4: GPU temporal filter — chunked cuFFT
• Slice 5: GPU Pass 2 — coefficient reconstruction + inverse DTCWT #16 — Slice 5: GPU Pass 2 — coefficient reconstruction + inverse DTCWT
• Slice 6: GPU integration — magnify_motions_gpu(), tests, docs #17 — Slice 6: GPU integration — magnify_motions_gpu(), tests, docs

Dependency Graph

Slice 1 (filters)  ──────────────────────────────────┐
Slice 2 (infra) → Slice 3 (Pass 1) → Slice 4 (filter) → Slice 5 (Pass 2) → Slice 6 (integration)

Slices 1 and 2 can be done in parallel.

Key Decisions (from design doc)

• Single file (motion_mag.py) with --gpu flag, not separate file
• Two-pass batched GPU pipeline (phase on CPU between passes)
• 3× C=1 sequential channels (not C=3 batching)
• Chunked cuFFT for temporal filtering (auto-tuned to VRAM)
• Re-run forward DTCWT in Pass 2 (no amplitude/Yl storage)
• near_sym_b/qshift_b as new default filters
• Float32 everywhere on GPU (verified negligible error)
• Pre-flight CPU RAM + VRAM check

[joeljose]

Completed in PR #18, released in v2.0.0