ternary)¶

Date: 2026-07-24
Machine: Windows x64, MSVC /O2 /openmp, hardware __popcnt64
Protocol: scripts/benchmark.py — weights pre-packed once; warmup=5; reps=10; BNN_NUM_THREADS=4 for headline compute; thread sweep 1/2/4/8.

Thesis lock unchanged: 32× is packing density, not a guaranteed e2e wall-clock claim. Numbers below are kernel measurements.

Before → after (this pass)¶

Item	Before	After
Native GEMM	Single-thread (or OpenMP stub without thread API)	OpenMP over output rows + `BNN_NUM_THREADS` / `--threads`
Pack	Multiply-sum over 64 bits	`numpy.packbits` little-endian uint64
Ternary compute	Dequant → FP GEMM only	Bitplane `ternary_gemm_u64` (native) + NumPy fallback
Compile	MSVC vcvars list	+ vswhere, `/DEF` exports, Linux `-fopenmp` path
Bench	Fixed sizes, no scaling curve	Warmups, pack-vs-compute, thread curve, ternary

Headline speedups vs NumPy FP32 (compute-only, committed `results/benchmark.json`)¶

Shape (B×N×M)	Compute ms	NumPy FP32 ms	S_compute
128×2048×2048	0.61	7.34	11.99×
64×4096×4096	0.60	14.37	23.86×
32×8192×8192	0.94	27.35	29.25×

Prior single-thread-era golden at 64×4096×4096 was ~3.7× vs NumPy FP32.

Thread scaling (compute-only ms → speedup vs 1 thread)¶

Committed artifact: results/benchmark.json / results/benchmark.md (W13.T04). Soft CI check: bnn.profile.check_committed_bench_soft_floors requires ≥2 thread points per published shape. Re-run does not invent new golden shapes.

Shape	1	2	4	8
128×2048×2048	1.93 (1.00×)	1.19 (1.62×)	0.85 (2.28×)	1.31 (1.48×)
64×4096×4096	2.22 (1.00×)	1.66 (1.34×)	1.07 (2.08×)	0.74 (2.99×)
32×8192×8192	3.69 (1.00×)	2.52 (1.46×)	1.65 (2.23×)	1.00 (3.68×)

Plateau / non-monotonic 8-thread on smaller shapes is expected (memory bandwidth / OpenMP overhead). Prefer BNN_NUM_THREADS=4–8; avoid blindly matching logical CPU count. Numbers above are wall-clock on the commit machine — floats need not match bit-identically elsewhere; conclusions must.

Ternary bitplane vs dequant FP (same shapes, err=0)¶

Shape	Bitplane ms	Dequant FP ms	S
128×2048×2048	3.31	21.54	6.50×
64×4096×4096	4.56	54.64	11.99×
32×8192×8192	11.20	188.70	16.85×

Bitplane path requires ±1 activations. Full-precision X still uses dequant FP (honest pedagogy).

Pack once vs compute¶

Shape	Pack W ms	Pack X ms	Compute ms
128×2048×2048	5.31	0.22	1.79
64×4096×4096	20.13	0.33	2.92
32×8192×8192	90.67	0.39	6.04

Weights are deploy-time; fair inference benches must not re-pack W every call.

How to reproduce¶

python -m bnn.kernels.compile_native --force
set BNN_NUM_THREADS=4
python scripts/benchmark.py --reps 10 --warmup 5 --threads 1,2,4,8
bnn validate-native   # err_nat=0

Artifacts: results/benchmark.json, results/benchmark.md.

Other gaps closed here¶

binary_gemm_set_num_threads / get / openmp_enabled + .def exports (MSVC)
Linux/GCC -fopenmp in compile_native.py
Eval-only fuse_bireal_bn_ (BN → binary α/bias)
DataLoader num_workers=0 on Windows (BNN_NUM_WORKERS override)
G11 (OpenMP polish) no longer “deferred non-goal” for this lab — shipped with measured curve