Generated from the // comments on the declarations under src/. The prose is the same either way — reading it in the header is often quicker — but the generated site adds two things a header cannot show: which layer includes which, and a link from every declaration to the places that use it.

The narrative documentation — architecture, benchmarks, sanitizers, testing — lives in the repository's docs/ directory. This site is the reference half.

Layers

Layer	Contents
`core/`	`Span`, `GemmShape`, `MatrixUtils`, `Stopwatch`, `CliOptions` — everything the three backends share
`cpu/`	Serial kernels: five GEMM variants, two sum reductions
`omp/`	OpenMP kernels and the thread-count control used to pin and report them
`cuda/`	Device kernels plus the RAII owners for device memory and events

The dependency edges run one way: cpu/, omp/ and cuda/ all include core/, and none of them includes another. omp/gemm_omp.h includes cpu/cpu_gemm.h for the shared tile-size constant and nothing else.

Where to start

cpu/cpu_gemm.h — the five serial GEMM variants and what distinguishes them
omp/gemm_omp.h — why only the fastest serial kernels have parallel counterparts
cpu/reduction.h — why a serial optimised reduction exists beside the baseline
cuda/device_buffer.cuh — the ownership discipline the error paths depend on