Livt packages

Livt.Math

Version 1.7.2

Reusable integer and fixed-point arithmetic, complex numbers, trigonometry, and statistics for FPGA applications written in Livt.

CHANGELOG.md

Changelog

All notable changes to this package are documented here.

1.7.2 - 2026-11-08

Metadata

  • Add a package description, repository and homepage links, discovery keywords, the MIT license identifier, and organization details to livt.toml.
  • Metadata-only release: no arithmetic API, implementation, timing, or dependency changes from 1.7.1.

1.7.1 - 2026-11-08

Packaging

  • Select the README, changelog, usage guide, and license in livt.toml so they can be distributed alongside the package's source and tests.
  • No arithmetic API or implementation changes from 1.7.0.

1.7.0 - 2026-11-06

Added

  • Arithmetic.CheckedBlockMac32<A_BITS, B_BITS, ACC_BITS, PRODUCT_STAGES> and its native CheckedBlockMacPrimitive: checked dot products and retained weighted sums for 32-element vectors processed in sixteen-int32 blocks.
  • Configurable signed input ranges (2..32 bits), accumulator range (2..64 bits), and one- or two-stage product pipelines. Pipeline choice changes latency, never arithmetic results or overflow checks.
  • Packed 65-bit result/status, ordered prefix-overflow detection, poisoned-bank recovery, reset cancellation and retained completion results.
  • A BRAM-backed weighted accumulator bank with validity bits and immediate-read forwarding; live input lane selection avoids redundant wide row registers.
  • Native integer-oracle and generated-wrapper tests covering numeric limits, both product pipelines, reset, retention, failure and recovery.

Compatibility

  • Existing arithmetic APIs and defaults are unchanged. The new worker is single-owner: serialize calls and hold native transaction inputs until completion. No fixed-point scaling, rounding or saturation is introduced.
  • Timing, DSP usage and BRAM inference depend on the target and integration; configured numeric ranges do not imply proportional hardware savings.

1.6.0 - 2026-11-03

Added

  • A globally stallable 64-stage unsigned division pipeline with exact quotient, remainder, divide-by-zero status, tag transport, and initiation interval one.
  • A globally stallable 16-stage unsigned floor square-root pipeline with tag transport and initiation interval one.
  • A two-stage exact population-variance numerator and denominator preparation pipeline for continuous moving-statistics datapaths.

Compatibility

  • Existing scheduled arithmetic APIs remain unchanged. The new direct valid/ready primitives are additive and intended for cycle-sensitive owners.

1.5.0 - 2026-11-03

Added

  • IntegerMath context-free signed minimum, maximum, clamp, magnitude, saturating absolute-value, addition, and subtraction helpers.
  • UnsignedMath context-free saturating uint addition and subtraction helpers.
  • Full-width boundary tests and usage documentation for the new integer helpers.
  • CordicSinCosQ15, a serialized 16-step rotation-mode worker for arbitrary signed full-turn phases with Q1.15 sine/cosine outputs and a 4-LSB error bound.
  • Numeric and lifecycle documentation for CordicSinCosQ15.
  • A narrow resource-shared CORDIC primitive with reset and retained-result behavior.
  • CordicAtan2MagnitudeQ15, a separate serialized vectoring worker for canonical phase and full-range unsigned magnitude from Q1.15 coordinates.
  • FPGA-oriented fixed-point special functions: serialized Q15 interpolation, normalized reciprocal, and paired base-2 logarithm/exponential components.
  • StreamingMomentsInt16, an exact bounded finite-stream accumulator for signed-16 mean, population variance, and standard deviation.

Changed

  • TanhQ15 reuses the serialized interpolation primitive used by the bounded special-function components.

Compatibility

  • Existing Livt.Math 1.4.0 APIs retain their public signatures and default behavior. The new components and helpers are additive.
  • Moving-window statistics and median filters remain outside this package and are reserved for Livt.Signal.

1.4.0

Added

  • Configurable three- or four-stage division rounding in WideArithmetic.
  • An optional exact fast path for positive power-of-two divisors using shifts and masks.
  • Optional registered unary/binary dispatch groups for reducing wide operand selector fan-in.
  • Scheduled wrapper tests for configured division, disabled operations, failure reporting and recovery.
  • Native RTL oracle coverage across every rounding/fast-path combination, including exact results, latency, reset cancellation and retained output.

Changed

  • WideArithmetic now has trailing ROUNDING_STAGES, POWER_OF_TWO_FAST_PATH and PIPELINED_DISPATCH parameters.
  • WideArithmeticPrimitive accepts constant synthesis controls for rounding stages and power-of-two classification. Its new native ports have defaults for compatibility with existing instantiations.
  • The power-of-two fast path accelerates both nearest-even and floor division. Other divisors require one classification clock when the fast path is enabled.

Compatibility and behavior

  • Existing WideArithmetic<OUTPUT_BITS, OPERATIONS> construction retains the previous result behavior and default division latency.
  • Configuration affects latency and generated hardware, never numeric results.
  • Three rounding stages remain the default. The fast path and registered dispatch are disabled by default.
  • Timing and resource effects depend on the target and synthesis tool; no routed clock-frequency or area claim is made by this package release.

Validation

  • The package test suite covers the public scheduled API and existing arithmetic components.
  • The native RTL oracle runs 5,232 exact comparisons across four configurations, including boundary values, random operands, reset cancellation and recovery.

Earlier releases

Earlier release history remains available in Git.