pcx

pcx is a library for performing fft, written in C++23.
Currently only one-dimensional complex to complex fft is supported.
Transforms over contiguous or parallel data are supportted.
Higher dimensional transforms can be perfromed by combining multiple one-dimensional transforms.
Currently only 32-bit and 64-bit floating point numbers are supported.
Currently only modern x86 platforms are supported (SSE4.1, AVX2, AVX512).

Usage

The simplest way to use is with FetchContent_MakeAvailable or add_subdirectory. Since the implementation is very abstracted the performance in unoptimized e.g. Debug modes is poor. Installing locally using CMake ExternalProject_Add or otherwise can be beneficial.

The target architecture is selected using PCX_ARCH option. Currently see41, avx2, avx512 values are supported.

FFT

There are two classes that perform fft: pcx::fft_plan and pcx::par_fft_plan. pcx::fft_plan is for perfroming fft over contiguous complex data.
pcx::par_fft_plan is for perfroming fft over parallel complex data.

An example of fast convolution:

    auto plan = pcx::fft_plan<float>(1024);
    auto data = std::span<const std::complex<float>>{};
    auto result = std::vector<std::complex<flaot>>(1024);
    // acquire data
    plan.fft(result, data);
    // multiply `result` by convulution kernel
    plan.ifft(result);

Design

pcx intially stands for packed complex, a represntation of complex data where real and imaginary values are stored in interleaved packs. std::complex can be viewed as packed complex data with pack size of 1. This representation is better suited for simd processing. Currently pcx APIs use std::complex, but the goal in the future is to provide a collection of containers/wrappers/views with user-provided pack size, as well as instruments to perfrom common mathematical operations over the data.

pcx fft implementation is designed to be very flexible through heavy use of templates. The goal is to adapt the library to all common platforms with simd. std::experimental::simd from libstdc++ and C++26 std::simd are planned to be added as backends for simd implementation.

The end goal is to provide a number of utilities for processing of complex data with main application being processing of radio data.

Future ideas

• implement radix-4 based nodes
• implement better size-alignment policy
• implement permuted pack policy for faster interleaved-interleaved operations