FFT-op: An FFT library with hardware-specific optimizations

FFT-op provides a common FFT interface with implementations optimized for different hardware. The goal is to explore how algorithm choice, parallelism, vectorization, and memory access affect FFT performance across CPUs and GPUs.

Usage

cmake -B build && cmake --build build

CUDA is enabled automatically when a CUDA compiler is available. Pass -DFFTOP_ENABLE_CUDA=OFF to force a CPU-only build.

The main API is FFTop::fft():

FFTop::fft(x);                                          
// Use Backend::Auto to automatically resolves
FFTop::fft(x, {Backend::CPU, Direction::Forward});
FFTop::fft(x, {Backend::GPU, Direction::Inverse});

Inverse FFTs are unnormalized (no /N).

Run tests and benchmarks with:

ctest --test-dir build
./build/bench/fftop_bench

Implementations

FFT-op detects available CPU features, thread count, SIMD support (SSE2 / AVX2 / AVX-512 / NEON), OpenMP, and CUDA devices at runtime. It uses this information to select kernels and resolve Backend::Auto.

CPU

• Radix-2 and radix-4
• Scalar, AVX2, and NEON kernels
• Optional OpenMP parallelism
• Iterative and recursive implementations

GPU

• CUDA radix-2 and radix-4
• Stockham radix-2
• Naive DFT for reference

FFT plans are cached and reused across repeated transforms with the same configuration.