Portable performance and benchmark policy

Version 1.8 keeps native Pascal scalar code as the correctness baseline while adding a reviewable cache-blocked dense multiplication path.

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 1.8, so every supported target retains the complete portable implementation.

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.

Compile with:

fpc -B -O3 -FcUTF8 -Fusrc -FEbuild-temp/benchmarks \

  -FUbuild-temp/benchmarks benchmarks/BenchmarkRunner.lpr

./build-temp/benchmarks/BenchmarkRunner

Compare timing only on the same CPU, compiler, flags, power policy, and background-load conditions. Release qualification publishes those settings, checksums, working-storage rules, and deltas from the previous stable release. A material regression needs an explanation; CI compiles benchmarks but does not use noisy wall-clock thresholds as a correctness test.

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.