On this page

mathlib-fp 1.9.5 performance evidence

This report maps every v1.9.5 performance and memory statement to a canonical benchmark row. The machine-readable contract is performance-evidence-1.9.5.json; the validator compiles the runner with -B -O3 -FcUTF8, captures host and compiler metadata, checks exact gates, and writes a versioned result artifact.

Evidence contract

Each PERF|... row records its domain, scale, scalar kind, shape, cold time, warmed aggregate time and repetition count, allocation metric, retained bytes, logical working elements, dense-shape elements, checksum, tolerance, and setup. Cold time is the first operation after deterministic fixture construction. Warmed time follows an unmeasured warm-up and uses prepared inputs or workspace as named by the row.

The following are hard failures: missing or duplicate rows, incorrect metadata, checksum drift beyond tolerance, allocation or retained-memory ceilings, logical working-storage ceilings, and any sparse/matrix-free dense-shape allocation. Timing ratios are advisory. A ratio outside its review band makes the result review; it does not turn a variable hosted runner into a false precision gate.

Checked rows and claims

Row	Scale and scalar	Checked claim
`dense-gemm-small-portable`	16×16×16, `Double`	Prepared portable `Into` overhead; zero result allocations and retained bytes.
`dense-gemm-large-portable`	127×129×65, `Double`	Portable dense reference with bounded logical working storage.
`dense-gemm-large-auto`	127×129×65, `Double`	Automatic dispatch, checked against the portable row in the same run.
`sparse-cg-large`	100,000×100,000 CSR, 100,000 nonzeros, `Double`	Prepared repeated solve retains linear workspace and exactly zero dense-shape elements.
`iterative-cg-large`	dimension 200,000, matrix-free, `Double`	Prepared repeated solve retains linear workspace and exactly zero dense-shape elements.
`dsp-convolution-small-direct`	lengths 8 and 3, `Double`	Small allocating direct-convolution overhead and checksum.
`dsp-fft-large-baseline`	65,536 complex values, `Double` components	Existing native complex radix-2 baseline.
`dsp-fft-large-candidate`	65,536 complex values, `Double` components	Public applied transform, checked against the baseline in the same run.
`modelling-integral-small`	scalar interval, `Double`	Adaptive integration setup-sensitive path at documented default tolerances.
`modelling-qmc-large`	2-D, 100,000 samples, `Double`	Seeded Halton quasi-Monte Carlo throughput with constant-dimensional workspace.
`statistics-online-small`	64 values, `Double`	Fresh streaming-statistics record overhead and six-scalar state.
`statistics-online-large`	2,000,000 values, `Double`	Streaming throughput with the same six-scalar state.
`data-analysis-pca-small`	32×4, 2 components, `Double`	Setup-sensitive allocating PCA and bounded retained output.
`data-analysis-pca-large`	1024×8, 4 components, `Double`	Larger allocating PCA throughput and bounded storage.

The exact fixture, checksum, tolerance, ceiling, timing semantics, and claim identifier for each row live beside it in the machine-readable contract.

Profile-led DSP change

The host-matched 1.9.4 observation records dsp-fft-large-candidate at 78.2 ms per prepared call. Inspection located per-butterfly trigonometric evaluation in the applied radix-2 kernel. Reusing the already tested native Pascal radix-2 implementation measured 9.4 ms per call in final qualification; the same run measured dsp-fft-large-baseline at 9.2 ms. The checked candidate/prior ratio was approximately 0.120 and the same-run ratio was 1.022, both within their advisory bands. These are observations on one Windows x86-64 machine with FPC 3.2.2, not portable throughput promises.

The change does not alter a public declaration. A new reference and round-trip test covers the public power-of-two path, the complete FPCUnit suite remains green, and the SizeInt fallback retains an explicit portable twiddle recurrence for array sizes outside the reused kernel's Integer range.

Reproduction

From the repository root:


python tools/test_performance_evidence.py

python tools/check_performance_evidence.py --compiler fpc \

  --work-dir build-temp/performance

The result is build-temp/performance/performance-results.json. Compare timings only when CPU, OS, compiler version, flags, scalar kind, dimensions, tolerance, setup, and output validation match. Allocation ceilings are logical counts or explicitly labelled sampled live-heap deltas; they are not claims about every short-lived RTL allocation event.

Limits

The suite is representative, not an exhaustive performance model. Timer resolution can report zero for individual small cold calls; warmed repetitions remain available. No external library is needed or used. Linux and Windows clean-archive qualification for the exact candidate commit remains mandatory before tagging.