matmul-cpu: a 512 x 512 float matmul in C++, timed on a laptop CPU

Level: L1. Kind: timing. Objective: medianMs, lower is better. Hardware class: cpu-generic. Agreement band: 10 percent. Time budget: 60 s. On when DMI_ENABLE_CONTEST=1.

What the challenge is

The smallest possible timing challenge. It exists so the whole timing path, from the static rules through the compute-node agreement band to the ledger, can be run on a laptop with no GPU. It has the shape every kernel challenge has: a fixed kernel signature, a driver the challenge provides, a reference implementation, hidden inputs, a correctness gate, and a median wall time.

The submission is one C++ file that defines

extern "C" void dmi_matmul(const float* A, const float* B, float* C, int n);

Row major, C is zeroed before every call, C = A * B, single thread.

Objective

The driver (run.cpp) generates two pairs of 512 x 512 matrices from the seed (mulberry32, values in [-1, 1)), calls the kernel once to warm up, then five times, timing each pass over both cases with steady_clock. medianMs is the median of the five passes. Before that, one call per case is checked against the reference kernel (baseline.cpp, compiled with the same driver in a separate process): every element must be within 1e-3 + 1e-3 * |ref|. A miss is rejected before timing. The outputs of the timed pass are checked too.

Timing is not deterministic, so two scorers agree when their medians are within 10 percent on cpu-generic, the settled value is the median of the agreeing medians, and a submission promotes only when it beats the frontier by more than 10 percent. See coordinator/compute.js (resultsAgree, settledResult) and beatsBar in coordinator/server.js.

Rules

Checked by loadPolicySource() before compiling, each with a message that names the rule:

• under 64 KB, defines the entry point above
• includes only from: cstddef, cstdint, cmath, cstring, algorithm, immintrin.h, arm_neon.h, x86intrin.h and their C spellings
• no asm, no threads or OpenMP, no system, fopen, exec, fork, dlopen, mmap, signal, exit, getenv, no main, no __attribute__, no #pragma

Compiled with c++ -O2 -std=c++17 run.cpp kernel.cpp. Compile plus the reference run, the correctness run and the five timed passes must finish inside 60,000 ms.

Measured numbers (Apple silicon laptop, Apple clang 17, 2026-09-08)

The example is 14x faster. Run to run, medians on this laptop move by 1 to 5 percent, which is why the band is 10.

Running it locally

node coordinator/challenges/matmul-cpu/harness.js examples/matmul-blocked.cpp
node coordinator/challenges/matmul-cpu/harness.js coordinator/challenges/matmul-cpu/baseline.cpp
DMI_ENABLE_CONTEST=1 node --test test/contest.test.mjs

Provenance

• Driver run.cpp, reference baseline.cpp, inputs public-inputs.json ({"seed":1,"n":512,"cases":2}), all in this directory. The hidden inputs are the same shape with the coordinator's hidden seed, written by ensureTraces().
• Generic harness: coordinator/challenges/_contest/harness.js.