CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

The format is inspired by Keep a Changelog, adapted to fit fxpmath release history and compatibility tracking.

[0.5.0]

Compatibility

• Removed default import-time mutation of NUMPY_EXPERIMENTAL_ARRAY_FUNCTION and fixed its environment parsing bug.
• Defined support policy for Python 3.9 to 3.14 and NumPy 1.26.4 to <3 in packaging metadata and dependency files.
• Expanded CI and release-gate workflows with explicit NumPy 1.26.4 and latest 2.x matrix lanes across valid Python versions.
• Normalized out=... handling across operator helpers so NumPy ellipsis is treated as an omitted output argument.
• Kept strict Fxp-only out contract for fxpmath-managed dispatch and explicit errors for invalid out and out_like types.
• Updated _wrapped_numpy_func output routing to make __array_ufunc__ and __array_function__ paths consistent for out and out_like handling.
• Modernized NumPy protocol hooks: Fxp.__array__ now supports NumPy 2.x dtype and copy behavior while preserving legacy positional compatibility.
• Removed legacy __array_prepare__ hook and extended Fxp.__array_wrap__ with return_scalar support for NumPy 2.x compatibility.

Added

• Added isolated import regression coverage to verify import fxpmath preserves absent, 0, and 1 states for NUMPY_EXPERIMENTAL_ARRAY_FUNCTION.
• Added mixed-input NumPy divide regression coverage.
• Added focused feature tests for out=..., scalar and 0-D output paths, strict output-type validation, and array protocol compatibility in NumPy-facing and function-level test files.
• Added user-facing compatibility notes and NumPy migration guidance in docs/compatibility_notes.md.
• Started internal module reorganization for 0.5.0 with fxpmath/operators/ split into core, arithmetic, reduction, and array.
• Extracted Config from fxpmath/objects.py into fxpmath/config.py.

Changed

• Narrowed fallback conversion retries in _wrapped_numpy_func so unsupported kwargs raise cleanly instead of being retried through broad implicit coercion.
• Converted fxpmath/functions.py into a backward-compatible public facade re-exporting operator functions from fxpmath.operators.
• Converted monolithic fxpmath/utils.py into a package-based layout under fxpmath/utils/ with common, parse, repr, bitwise, and numeric modules.
• Preserved compatibility for from fxpmath import utils, from fxpmath import Config, and from fxpmath.objects import Config during the reorganization.
• Kept NumPy dispatch registrations and public operation signatures stable during the refactor.

Fixed

• Fixed unsigned subtraction underflow in raw arithmetic dispatch by forcing object-backed intermediates when both operands are unsigned, preventing pre-saturation wraparound on 32-bit paths and 64-bit boundary cases (issue #73).

Tests

• Expanded issue #73 regression coverage with explicit cross-platform paths in tests/test_issues.py: default behavior, emulated n_word_max=32, and emulated n_word_max=64 boundary conditions.

Documentation

• Documented supported Python and NumPy ranges plus finalized NumPy dispatch output behavior.
• Moved compatibility guidance out of the README into dedicated documentation.

Validation

• Validated targeted compatibility test groups on NumPy 1.26.4 and latest 2.x.
• Added stable CI gate jobs for required NumPy compatibility and installation smoke checks, plus artifact uploads for lint/test traceability.
• Added a scheduled nightly workflow against NumPy pre-release wheels.
• Ran full-suite validation successfully on NumPy 1.26.4 and latest 2.x environments (143 passed).

[0.4.10]

Added

• Added installation smoke tests in tests/test_installation.py to create an isolated virtual environment, install from source, and verify import path and version.
• Added regression test coverage for complex uraw conversion parity (issue #102).
• Added regression coverage for repeated int() casting across Linux sizing paths (issue #98).
• Added Fxp.reshape_inplace(shape, order='C') as the explicit in-place reshape API.

Changed

• Modernized packaging to PEP 517 and PEP 621 using pyproject.toml as canonical metadata source.
• Added dynamic versioning from fxpmath.__version__ and setuptools package discovery via find.
• Cleaned up development-plan tracking by removing a duplicated issue #86 status entry.
• Deprecated Fxp.shape = ... assignment with DeprecationWarning; prefer reshape() or reshape_inplace().

Fixed

• Fixed complex uraw conversion by applying unsigned two's-complement independently to real and imaginary components (issue #102).
• Fixed scalar integer casting after repeated in-place increments by using scalar-safe integer conversion in astype and object-cast set_val paths (issue #98).
• Defined complex bitwise semantics for issue #55: mixed complex and non-complex operands now warn once and apply the real operand to both components, while complex-complex applies operations part-wise.
• Refactored bitwise internals to support array-vs-array operands with NumPy broadcasting semantics for &, |, ^, and ~.
• Aligned Fxp.reshape with NumPy by returning a reshaped copy instead of mutating the source object.
• Improved reshape compatibility by using positional ndarray reshape arguments across NumPy versions.

Tests

• Kept issue #55 regression minimal in tests/test_issues.py and moved broader behavior coverage to tests/test_complex_numbers.py.
• Expanded bitwise operator coverage in tests/test_operators.py for multidimensional arrays, broadcast paths, scalar masks, and >64-bit cases.

[0.4.9]

Added

• Added from_bin method and function for creating or setting Fxp values from binary strings (issue #49).
• Added support for complex binary strings as input format.
• Added prefix selection for binary and hexadecimal representations.

Changed

• Forced config.op_input_size = 'best' when power uses a constant non-Fxp operand and added a warning (issue #89).

Fixed

• Fixed wrap behavior beyond n_word_max (issue #41).
• Propagated the inaccuracy flag to new or resulting Fxp objects when any input is inaccurate (issue #48).
• Fixed cumsum for sizes larger than 32 bits on Windows and 64 bits on Linux (issue #76).
• Fixed numpy.reshape handling so the default sizing remains same instead of optimal (issue #77).
• Fixed negative number parsing in dtype strings (issue #80).

[0.4.8]

Fixed

• Fixed value dtype handling for Windows and uint values treated as 32 bits.
• Updated __getitem__ so it returns an Fxp with raw value as a view, solving issue #62.

Tests

• Added tests for issues #60 and #62.

[0.4.7]

Added

• Added abs method.

Changed

• Kept val as the original vdtype instead of forcing object when it is safe to do so.
• Converted vdtype to float when scaling transformation is applied.

[0.4.6]

Fixed

• Fixed complex truediv and floordiv methods (issue #53).
• Fixed complex binary, hexadecimal, and base representations for arrays (issue #56).
• Fixed dtype-based initialization when the dtype is complex (issue #58).

Tests

• Updated tests for complex representations.

[0.4.5]

Added

• Added dtype argument to resize.
• Added get_dtype method with notation argument.
• Added docstring updates for get_dtype, resize, reshape, flatten, set_val, astype, get_val, raw, uraw, and equal.
• Added order parameter support for reshape and flatten.

Fixed

• Fixed FutureWarning in subdtype 'str' comparison (issue #45).

[0.4.4]

Fixed

• Fixed wrapping and scaling behavior (issue #44).

[0.4.3]

Added

• Added optional config init parameter to Fxp.
• Added copy and deepcopy methods to Config.
• Linked functions.truediv with numpy.divide dispatch.
• Added axes parameter to transpose.

Changed

• Functions operating on one or two values now copy the first operand configuration by default.

Fixed

• Fixed wrap overflow after operations (issue #42).
• Fixed utils.wrap when data exceeds n_word_max (issue #41).
• Fixed warning about internal complex checks (issue #39).

[0.4.2]

Added

• Added template support to Config.

Fixed

• Fixed config update behavior from an Fxp template.

[0.4.1]

Added

• Added NumPy precision testing at init.
• Defined n_word_max and config error handling independently.
• Added support for NumPy arrays of strings.

Fixed

• Fixed astype casting when dtype and vdtype are None.
• Fixed casting to object in extended cumprod, extended precision multiply, and constructor paths.
• Fixed integer representation behavior.
• Fixed issues #26 and #31.

[0.4.0]

Added

• Added extended precision support beyond 64 bits for add, sub, mult, truediv, floordiv, mod, and pow.
• Added support for tuple and Decimal input formats.
• Added dtype-based initialization for fxp and Q formats.
• Added template functionality for defining default formats for new Fxp objects.
• Added Config class to manage low-level object behavior.
• Added extended precision status flag (extended_prec).
• Added new Fxp properties: size, shape, ndim, and T.
• Added new Fxp methods: reshape, flatten, raw, and uraw.
• Added NumPy-array-style methods on Fxp: all, any, argmax, argmin, argpartition, argsort, max, min, mean, std, var, sum, cumsum, cumprod, ravel, tolist, sort, conjugate, conj, transpose, item, clip, diagonal, trace, prod, dot, and nonzero.
• Added support for bool, int, float, and complex casting.
• Added dtype information to Fxp.__repr__.
• Added basic callbacks: on_value_change, on_status_overflow, on_status_underflow, and on_status_inaccuracy.
• Added NumPy function dispatch so Fxp can be used as both NumPy inputs and outputs.

Changed

• Migrated several core calculations and methods from float to int to support extended precision.
• Results of ALU operations with constants now keep the same size as the Fxp object by default.
• like initialization now supports overwriting size parameters.
• Improved performance in several core paths.

Notes

• NumPy >= 1.17.x was recommended to enable function dispatch by default.
• With older NumPy versions, users were advised to import fxpmath before NumPy or set NUMPY_EXPERIMENTAL_ARRAY_FUNCTION=1 before importing NumPy.

[0.3.9]

Fixed

• Fixed wrapping behavior (issue #20).

[0.3.8]

Added

• Added padding support for hexadecimal representations.

Fixed

• Fixed bug with 'b' in hexadecimal strings (issue #15).
• Fixed spelling/wording of inaccuracy in status (issue #16).
• Fixed __pow__ behavior (issue #17).
• Fixed bug where indexed complex assignment stored only the imaginary part (issue #19).

[0.3.7]

Added

• Added right-shift support for shifting='expand' mode.
• Added sum function in functions.
• Added support for Fxp words and values larger than 64 bits.

Changed

• Improved performance in indexed assignment.
• Fxp slicing now returns an Fxp object when values are arrays.
• Fxp now accepts another Fxp as input value.
• Updated docstrings.

[0.3.6]

Fixed

• Fixed setup.py dependency handling so package dependencies install correctly.

[0.3.5]

Fixed

• Fixed wrap error for ndarrays.

[0.3.4]

Fixed

• Fixed wrap error for unsigned Fxp.

[0.3.3]

Fixed

• Fixed bitwise shifting of unsigned values.
• Fixed wrap error when the input is a float.

[0.3.2]

Fixed

• Fixed support for N-dimensional array inputs (N > 1).

[0.3.1]

Fixed

• Fixed initialization and best_size_calc behavior for Win32 platforms.

[0.3.0]

Added

• Added logical (bitwise) operators.
• Added binary shifting.
• Added value scaling (scale and offset).
• Added comparison support.
• Added inaccuracy status flag to indicate stored value differs from input.
• Added verbosity support to info().
• Added reset() to clear status flags.
• Added raw-value support to set_val().
• Added power operator support.
• Added inplace operator support.

Fixed

• Fixed several bugs.

[0.2.0]

Added

• Added binary, hexadecimal, and base representations.
• Added upper, lower, and precision properties.
• Added arithmetic operator support.
• Added support for string inputs including binary, hexadecimal, and fractional formats.
• Added Fxp copying.
• Added initialization from another Fxp.
• Added indexing support.

[0.1.0]

Added

• First public release.