Portable performance and benchmark policy

Version 1.9.5 keeps native Pascal code as the correctness baseline and makes representative timing, allocation, retained state, and logical complexity reviewable through a checked benchmark contract.

Checked v1.9.5 evidence

The canonical runner emits 14 PERF|key=value rows across dense, sparse, matrix-free iterative, DSP, numerical-modelling, statistics, and data-analysis workflows. Small setup-sensitive rows and larger throughput/scale rows state their scalar kind, dimensions, cold and warmed timing semantics, setup, tolerance, allocation metric, retained bytes, working elements, dense-shape elements, and deterministic checksum.

docs/performance-evidence-1.9.5.json is the release contract. Missing rows, checksum drift, and exact allocation/storage/complexity ceilings fail. Same-run and matching-host prior timing ratios are advisory: a movement beyond its band requests review rather than pretending shared-runner wall time is mathematically exact. The human claim map and limitations are in PERFORMANCE_EVIDENCE_1.9.5.md.

Matrix multiplication paths

MultiplyInto remains the portable compensated oracle. MultiplyBlockedInto traverses output tiles with a caller-selected positive block size (default DENSE_MULTIPLY_BLOCK_SIZE = 32) while preserving each dot product's increasing-K order and compensated accumulation.

MultiplyAutoInto calls SelectedMultiplyPath. Dispatch is deterministic: workloads at or above DENSE_MULTIPLY_AUTO_THRESHOLD = 131072 scalar multiply-add positions select dmpBlocked; smaller workloads select dmpPortable. CPU model, timing, thread count, and global settings do not affect the choice.

All four typed scalar paths—single/double real and single/double complex—have the same validation, exact destination shape, overlap-safe temporary, and finite-input contracts. Cross-path tests use the portable result as oracle and also check deterministic repeat results and unchanged destinations after validation failure.

The blocked implementation is serial. No parallel or SIMD path is stable in 1.9.5, so every supported target retains the complete portable implementation.

Applied power-of-two DSP transforms reuse the tested native radix-2 FFT kernel for supported array sizes. The explicit portable twiddle-recurrence fallback remains available for larger SizeInt ranges, and public reference/round-trip tests cover the dispatch without changing the API.

Reproducing benchmarks

The benchmark runner reports:

• small-call FFT and streaming overhead;
• arbitrary-length and power-of-two transform throughput;
• direct and FFT convolution;
• medium applied-analysis workloads;
• portable, blocked, and automatic typed matrix multiplication;
• allocation/state sizes and numerical checksums; and
• the retained 1.7 scalar, geometry, vector, dense solve, SVD, and eigen baselines.

Run the checked contract with:

python tools/test_performance_evidence.py

python tools/check_performance_evidence.py --compiler fpc \

  --work-dir build-temp/performance

Compare timing only on the same CPU, OS, compiler, flags, power policy, background load, scalar kind, shape, tolerance, setup, and result-validation conditions. Release qualification retains those settings, exact-gate outcomes, same-run comparisons, and matching-host deltas from the previous stable release in performance-results.json.

Allocation and scale guidance

• Prefer ...Into matrix kernels when destinations can be reused.
• Reuse TStreamingFIR, TOnlineStatistics, TLocalRandom, and TScalarKalmanFilter state instead of concatenating an unbounded history.
• Bluestein FFT workspace is the next power of two at least 2N-1; check this cost before very large arbitrary-length transforms.
• Welch and STFT retain their output plus frame-sized work. PCA and k-means++ are in-memory dense algorithms.
• Binary interchange validates checked QWord sizes, caller limits, and platform SizeInt limits before result construction.

Win32 remains part of release qualification because address-space and dimension overflow can matter before numerical throughput does.