tests/bench/ is a host microbenchmark suite. just quality::local::bench builds every bench binary and runs it; each binary prints CSV to stdout.

It has no third-party benchmark dependency on purpose – no Google Benchmark, no JSON writer, no plotting. A bench that needs a package installed before it will build is a bench nobody runs on a fresh checkout, and the whole value of this suite is that it runs anywhere the host unit tests run, including CI, with nothing extra present.

The numbers are host x86_64 numbers, not RA8D2 numbers. They bound the algorithmic cost of a routine and catch a regression in the off-target path. They say nothing about the same code on a Cortex-M85, where the CRC, JPEG and graphics paths all have on-die accelerators underneath them.

Reading a result

Each binary writes a one-line CSV header followed by one row per measurement:

name,iterations,ns_per_op,MB_per_s

Column	Definition
name	Bench label as printed by the binary; never contains a comma.
iterations	Inner-loop iterations the harness chose to reach its minimum wall time.
ns_per_op	Mean wall-clock nanoseconds per iteration (total elapsed / iterations).
MB_per_s	Throughput where 1 MB = 10^6 bytes (decimal).

The harness pre-warms its working buffers outside the timed region, auto-scales iterations until the measurement runs long enough that the work dominates the clock, and reads clock_gettime(CLOCK_MONOTONIC).

There is no variance, median or percentile column, because at this granularity the spread is OS scheduler noise rather than microarchitectural jitter. Read a row as an order of magnitude: a result that moves by a factor of two on a different machine is the machine, and a result that moves by an order of magnitude on the same machine is a regression worth chasing. If a bench ever becomes precise enough that percentiles would mean something, that is the point to vendor a real benchmark library for it – not before.

How the suite is wired

The harness is tests/bench/inc/ra8_bench.h: single-header, no allocation, no init. Include it and call RA8_BENCH_TIME(...) from main().

Bench binaries link the same object library as the host unit tests, so they pick up the same fake MMIO. What a bench measures is therefore exactly the code the tests exercise; there is deliberately no parallel "perf-only" build variant that could drift away from the tested one.

Bench translation units are never instrumented: the bench configure path forces coverage and MC/DC off regardless of what the caller asked for, because an instrumented timing run measures the instrumentation.