Changelog

All notable changes to HPCSeries Core will be documented here.

Version 0.8.0 (2025-01-06)

Major Features

Execution Mode System

Safety vs performance trade-offs with three execution modes:

  • MODE_SAFE (default): IEEE 754 compliant, full error checking

  • MODE_FAST: Relaxed math optimizations, minimal validation

  • MODE_DETERMINISTIC: Bit-exact reproducibility, no OpenMP parallelization

Control via set_execution_mode() and get_execution_mode() API, or per-call mode parameter.

New Functions (v0.8)

  • ewma() - Exponentially weighted moving average (15-60x faster than pandas)

  • ewvar() - Exponentially weighted variance

  • ewstd() - Exponentially weighted standard deviation

  • diff() - Finite differencing (arbitrary order)

  • cumulative_min() - Running minimum

  • cumulative_max() - Running maximum

  • convolve_valid() - 1D FIR convolution with template specialization (kernel sizes 3-15)

  • trimmed_mean() - Robust mean (10-15x faster than SciPy)

  • winsorized_mean() - Robust mean with winsorization

Implementation

  • Thread-safe execution mode using OpenMP threadprivate storage

  • Zero overhead via compile-time dispatcher pattern

  • Hybrid Fortran/C++ architecture

  • Template-specialized convolution for common kernel sizes

Documentation

  • Fixed GitHub URLs (your-org → hpcseries)

  • Fixed Sphinx autodoc to show proper function signatures (not MagicMock)

  • Added execution mode API complete documentation

  • New notebook: 10_exponential_weighted_statistics.ipynb

  • Updated API reference with all v0.8.0 functions

Testing

  • 400+ line test suite with 40+ tests (24 passed, 3 skipped)

  • Reference comparisons vs NumPy/pandas/SciPy

  • All C/Fortran tests passing

License

  • Changed from MIT to Apache 2.0 for commercialization

Version 0.7.0 (2025-12-17)

Major Features

  • Python API with comprehensive Cython bindings

  • Complete set of Jupyter notebook examples (00-09)

  • SIMD-accelerated reductions (sum, mean, var, std, min, max)

  • Fast rolling operations (mean, median, std, var, MAD, z-score)

  • Robust statistics (median, MAD, robust z-score)

  • Anomaly detection (statistical and robust methods)

  • Axis operations for 2D arrays

  • Masked operations for handling missing data

  • Performance calibration and auto-tuning

  • CPU topology detection

  • CLI tool (hpcs cpuinfo)

Performance

  • 2-5x faster than NumPy for basic reductions

  • 50-100x faster than Pandas for rolling operations

  • Sub-microsecond latency for small arrays

  • Automatic SIMD vectorization (AVX2/AVX-512/SSE2)

  • OpenMP parallelization for large datasets

Documentation

  • Complete API reference

  • 12 Jupyter notebook tutorials

  • Installation and quick start guides

  • Migration guide from NumPy/Pandas

Version 0.6.0 (2024-12-01)

Features

  • SIMD dispatch system for runtime ISA detection

  • Fortran-C bridge for SIMD operations

  • Axis reductions (per-row/per-column statistics)

  • Masked reductions for missing data

  • Rolling operations with batching support

Version 0.5.0 (2024-11-01)

Features

  • Performance calibration system

  • Configuration persistence (~/.hpcs/config.json)

  • CPU detection (cores, cache, NUMA)

  • Adaptive auto-tuning

Version 0.4.0 (2024-10-01)

Features

  • 2D array operations (axis reductions)

  • Batched rolling operations

  • Masked array support

  • Quality indicators

Version 0.3.0 (2024-09-01)

Features

  • Rolling operations (mean, median, std, MAD)

  • Robust statistics (median, MAD, quantile)

  • Z-score normalization

  • C++ optimized rolling implementations

Version 0.2.0 (2024-08-01)

Features

  • Fortran HPC kernels for core reductions

  • OpenMP parallelization

  • Anomaly detection

Version 0.1.0 (2024-07-01)

Initial Release

  • Basic reduction operations (sum, mean, var, std, min, max)

  • Fortran implementation

  • C/Python bindings