Skip to content

Latest commit

 

History

History
234 lines (154 loc) · 11.8 KB

Changelog.md

File metadata and controls

234 lines (154 loc) · 11.8 KB

Version 0.10.2

  • Switched from packed_simd to std::simd. Warning: I no longer have a CPU which supports AVX512, so this option is untested and likely broken.

Version 0.10.1

  • Updated dependencies, thanks to Alex Touchet for his help!
  • Increased minimum Rust version to 1.67 as this is required by the latest versions of our dependencies
  • Fixed panic in complex to real vector operations, #51

Version 0.10.0

  • Fixed compilation error on ARM Neon which includes Apples M1 (#49), thanks to Andy Bailey for reporting this issue and for doing tests!
  • Update to Rust edition 2021
  • Increased minimum Rust version to 1.65 as this is required by the latest versions of our dependencies

Version 0.9.5

  • Updated time dependency

Version 0.9.4

  • Fixed compilation error with no_std option

Version 0.9.3

  • Updated rustfft dependency

Version 0.9.2

  • Updated rustfft dependency
  • Memory usage: With rustfft version 5 FFTs might require more memory. FixedLenBuffers will need to be resized. The default SingleBuffer will adjust itself dynamically. Refer to ejmahler/RustFFT#45 for a more detailed disucssion.
  • Buffers now resize correctly if the underlying data type allows for it
  • Fixed "panic safety issue in impl TransformContent<S, D> for [S; (2|3|4)]" (#47). Thanks to Youngsuk Kim.

Version 0.9.1

  • Updated rustfft and num-complex dependencies
  • Switched from packed_simd to packed_simd_2 to avoid build failures for now
  • Fixed warnings in interop crate
  • Implemented From and Intofor the conversion of a DspVec into a std::Vec to address #46
  • Increased minimum Rust version to 1.43 as the deeper integration with standard types requires https://rust-lang.github.io/rfcs/2451-re-rebalancing-coherence.html
  • Fix: in zero_pad and zero_pad_b used a different definiton for PaddingOption::Center if a complex vector with even points is padded to a resulting vector with odd points

Version 0.9.0

  • Updated rustfft and num-complex dependencies
  • Switched from packed_simd to packed_simd_2 to avoid build failures for now
  • Fixed warnings in interop crate

Version 0.8.2

  • Disabled AVX512 as it right now requires an unstable Rust feature
  • Updated time and linreg dependencies
  • Fixed "UB due to misuse of mem::uninitialized (will soon lead to panic) #44"

Version 0.8.1

  • Updated lazy_static and arrayvec dependencies.

Version 0.8.0

This release aims to improve the conversions from Rust standard library types to/from DSP vectors.

  • Renamed complex(..) function on vectors to cdata(..)
  • Redefined index functions for better usability. If the old functionality is required then replace calls to indexers with call to .data(..) and .data_mut(..), e.g. (&vec[5..10] becomes vec.data(5..10) and &mut vec[5..10] becomes vec.data_mut(5..10))
  • Changed FromVector trait. Call getf instead of get if the old functionality is required.

Version 0.7.2

  • Fixed #40 . Thanks to Carlos Tuñón.
  • Replaced simd with packed_simd

Version 0.7.1

  • Removed simd from default compilation options due to rust-lang/rust#57940 . Compile with --features use_sse2,use_avx2 to get the previous behavior.

Version 0.7.0

  • Update to Rust 2018
  • Updated dependencies, e.g. rustfft

Version 0.6.0

  • Removed combined_ops module as it at best showed performance improvements of 10% which didn't justify the complexity within it.
  • to_real function of ComplexTimeLinearTableLookup and ComplexFrequencyLinearTableLookup now takes &self instead of self
  • Renamed feature use_sse to use_sse2 and feature use_avx to use_avx2 to avoid confusion with SIMD naming. use_sse2 and use_avx2 are now active on default. Compile with --no-default-features --features std to revert to the compile options before this update.
  • ComplexData is now in meta::Complex. RealData, RealOrComplexData, TimeData, FrequencyData and TimeOrFrequencyData have been moved similarly. If you used those types then consider to implement against the vector type definitions or traits instead.
  • MultiCoreSettings now allow more fine grained control over the threading behavior.

Version 0.5.6

Behaviour changes:

The default MultiCoreSettings now have a core_limit of 1. That's because enabling multi threading should be a conscious decision, e.g. due to the overhead of multi threading. The conversion functions with the _par suffix (e.g. to_real_time_vec_par) can be used to get the old behaviour.

Changes:

  • Fixed performance issue for small vectors
  • If core_limit > 1 then at first call to a vector function which utilizes multi threading a small internal performance test will be executed to determine reasonable values for how work will be spread among the available threads. Running this benchmark will of course slow down this call (around 200 to 500ms). The assumption is that this is okay as it only happens once.
  • Only features use_sse or use_avx: Fixed index out of bounds exceptions in if the input vector is smaller than the SIMD register length.
  • Fixed calculation error in dot_product and dot_product_prec for vector in real number space.
  • AVX fixes (only feature use_avx): Fixed various calculations for the AVX optimizations. The AVX feature now passes the same tests as SSE and without optimizations. There is only one error in the doc tests left which seems to be related to rust-lang/stdarch#253 although that should be closed.

Version 0.5.5

It's now decided at runtime which SIMD optimization should be used. Also updated clFFT dependency.

Version 0.5.4

  • Went back from stdsimd to simd crate.

Version 0.5.3

  • Switched from simd to stdsimd crate.
  • Updated crossbeam, ocl and clFFT dependencies.

Version 0.5.2

Updated to crossbeam, arrayvec, ocl and clFFT dependencies.

Version 0.5.1

Updated to rustfft version 2.0.0.

Version 0.5.0

API cleanup and fixes. Most API changes should be transparent to users. A few tips for the version update:

  • Deprecated functions have been removed. The traits have replacements available, so the documentation should provide an idea what's the intended replacement.
  • In some cases in was possible to pass references of references (e.g. &&vec) as argument for binary ops (e.g. div) and that's no longer legal.
  • Some traits have been redefined. It's assumed that most users won't use the traits directly except for calling functions on vectors. And so most users shouldn't be affected by this change. If this doesn't hold true for your project then feedback in form of a bug or enhancement request is welcomed.

Changes:

  • zero_pad_b now returns a result, which may contain an error if the passed argument is smaller than the actual vector length
  • Renamed interpolate_vector to interpolate_signal
  • Removed deprecated functions swap_halves_b, statistics_splitted
  • Removed MapInplaceNoArgsOps for combined ops since they seem to not fit in their independent on how this module progresses.
  • Restructured statistics traits.
  • Mapping operations now only borrow a function. That also allows them to work internally without any reference counting.
  • Binary operations are now more flexible regarding what kind of argument they accept.
  • Restructured buffers to avoid two classes of errors.

Version 0.4.3

Minor additions.

  • Updated simd dependency to version 0.2.
  • New feature: Added ApproximatedOps trait.
  • New feature: New interpolate method which offers the same features as interpolatef, but the performance should be closer to interpolatei.
  • Fix: multiply_complex_exponential didn't consistently took delta into account.

Version 0.4.2

Bugfix release.

  • Fix: Not all implementations of to_complex_time_vec and to_complex_freq_vec set the vector length to 0 if a storage of odd length is passed.
  • Fix: Offset in zero_pad if Surround or Center was chosen and the original vector length is odd. As a result the convolution theorem didn't hold true, that means a multiplication in frequency domain didn't give the same results as a convolution.
  • Fix: zero_pad_b failed to copy the vector completely if Center was the selected option.
  • Fix: Convolution for vectors produced random results if the allocated size of a vector was different than it's actual size.
  • Fix: Buffer convolve_vector for matrices now returns all buffers back to the pool.
  • Performance: Convolution and interpolation now rely on the overlap and add algorithm or spawn worker threads for larger data sizes.
  • Fix: use_sse and use_avx failed to pick the faster implementations. However with the current status of the simd crate it's not recommended to use those feature flags. The Rust lib team is right now discussing about the future of the simdcrate.
  • Fix: Out or range panic in add, sub, mul and div for large vectors.
  • New feature: Added a TypeMetaData struct which allows to create a new vector with the same meta data as an existing vector.

Version 0.4.1

Minor additions and improvements.

  • Added a method to convolve a matrix with a matrix of impulse responses.
  • Added traits with more precise versions of sum, statistics and dot products.
  • Performance improvement for swap_halves and swap_halves_b, swap_halves_b at the same time has been marked as deprecated since it offers no advantage to swap_halves

Version 0.4.0 (Breaking changes)

Added support for matrix operations. In order to allow matrices and vectors to implement the same traits the existing traits had to be renamed and restructured.

  • Breaking change: Reorganized existing interfaces so that they can be reused for the matrix types. For that traits have been renamed and sometimes traits have been split in several smaller traits.
  • Breaking change: Interop facade is now only compiled with --features interop
  • Breaking change: Removed complex_data, data and override_data (which also had a spelling mistake) in favor for implementations of several Indexer traits.
  • Crate now compiles with Rust stable

Version 0.3.2

Bugfix release:

  • Bugfix: Potential SEGVAULT in add, sub, mul, div, add_smaller, sub_smaller, mul_smaller and div_smaller methods
  • Bugfix: SEGVAULT in get_magnitude and complex_dot_product

Version 0.3.1

Bugfix release with a minor enhancement:

  • New feature: Added sum and sum_sq operations.
  • Bugfix: SEGVAULT in complex magnitude op with recent versions of Rust nightly

Version 0.3.0 (Breaking changes)

Added prepared operations/multi operations, see combined_ops

  • Breaking change: Renamed a lot of operations so that their names match (more closely) with the num crate. Also renamed traits so that all traits which mainly implement operations have similiar names.
  • Breaking change: RededicateVector trait now defines exactly one conversion. Vectors implement the trait several times so the same functionality is still available.
  • Breaking change: perform_operations and related types appear on the API differently now.

Version 0.2.2

Added Apache-2.0 as license option.

Version 0.2.1

Bugfix release: Fixed implementation of AVX operations.

Version 0.2.0

First release: Vectors with a couple of operations are available.