vecasm

Hand-written NASM dense float kernels with runtime ISA dispatch (optional calibration). C API + thin C++ wrapper.

Cross-platform x86_64: Windows (Win64 ABI), Linux/macOS (System V).

Advantage

1. Detect CPU/OS features (CPUID + XGETBV). avx512icl only on GenuineIntel + FMA + VNNI + VBMI2.
2. Default AUTO = highest supported ISA (icl > avx512 > avx2 > scalar). Instant, no heap spike.
3. Optional vecasm_calibrate() builds a {op,tier} timing table and switches to VECASM_DISPATCH_CALIBRATED (~0.5s / up to ~192 MiB). Never runs by itself.
4. Hot path: vecasm_resolve_* / vecasm_resolve_dense return direct function pointers (no per-call atomics).

vecasm_set_dispatch(VECASM_DISPATCH_ISA); /* default, good for games */

vecasm_dense_fns fn;
vecasm_resolve_dense(&fn, n);
for (;;)
    acc += fn.dot(a, b, n);

/* optional offline tune */
vecasm_calibrate();

API

Kernel	Meaning
vecasm_dot_f32 / sum / axpy / norm2	dense (dispatched each call)
vecasm_resolve_dot/sum/axpy / resolve_dense	bind kernel once
vecasm_dot3_f32 / cross3 / normalize3	vec3
vecasm_set_dispatch / calibrate	ISA vs calibrated

Build

cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release
cmake --build build
./build/vecasm_test_dot
./build/vecasm_bench_dot

License

vecasm Attribution License (LICENSE + NOTICE).

Credit: LucPrusPPi (https://github.com/LucPrusPPi/vecasm). Do not hide origin.