This project ships with a small benchmark executable so that algorithm changes can be evaluated with reproducible numbers instead of assumptions.
Generate the markdown report from CMake:
cmake --build build --config Release --target benchmark_reportRun the benchmark executable directly:
./build/dsp_core/dsp_core_benchmark --iterations 20 --report build/reports/dsp_performance.mdThe report is written as a markdown table to make it easy to review in CI artifacts or commit into documentation when needed.
The current benchmark compares representative kernels:
- direct FIR convolution
- FFT overlap-save convolution
- scalar Goertzel evaluation
- batched Goertzel evaluation
These measurements are intended as engineering guidance, not as a vendor-grade microbenchmark suite.
The generated table includes:
- average execution time in milliseconds
- relative speedup against the direct FIR baseline
Representative example:
| Method | Time (ms) | Speedup |
|---|---|---|
| FIR direct | 2.693 | 1.00x |
| FIR overlap-save FFT | 1.816 | 1.48x |
| Goertzel scalar | 0.130 | 20.72x |
| Goertzel batched | 0.485 | 5.55x |
The interpretation matters more than the raw numbers:
short filters -> direct FIR can stay competitive
longer filters -> FFT methods become more attractive
single tone -> scalar Goertzel is simple and efficient
multiple tones -> batching can simplify multi-target analysis
Benchmark numbers depend on:
- compiler and flags
- CPU microarchitecture
- memory subsystem behavior
- selected iteration count
- signal size and filter length
Always compare runs on the same hardware and toolchain before drawing conclusions.
