AIPS BOOM

cuFFTW

2 min read

cuFFTW

Type: Library Tags: NVIDIA, CUDA, cuFFT, cuFFTW, FFTW, FFT, porting, signal processing, HPC Related: cuFFT, cuFFTDx, cuFFTMp, NVPL-FFT, nvmath-python, NVIDIA-CUDA Sources: https://docs.nvidia.com/cuda/cufft/index.html#fftw-interface-to-cufft Last Updated: 2026-04-29

Summary

cuFFTW is NVIDIA’s FFTW3-compatible interface layer for cuFFT. It lets applications that already use FFTW begin using NVIDIA GPUs with minimal source changes by including cufftw.h and linking against cuFFT and cuFFTW libraries.

Detail

Purpose

Many scientific and engineering applications were written against FFTW before GPU acceleration was introduced. cuFFTW gives those applications a migration bridge: keep an FFTW-style API surface, link to NVIDIA CUDA libraries, and then progressively move data placement and performance-sensitive paths toward native cuFFT usage.

Key capabilities

  • FFTW3-style interface to the cuFFT library.
  • Header and library path through cufftw.h and libcufftw.
  • Double-precision and single-precision FFTW-style functions for supported transform families.
  • Complex and real DFT plan coverage for common 1D, 2D, 3D, and many-transform cases.
  • Compatibility path for FFTW applications before deeper CUDA/cuFFT-specific optimization.
  • NVIDIA docs recommend moving data to and from the GPU and using native cuFFT conversion-guide routines for best performance after the application is working.

Important limits

  • cuFFTW does not implement every FFTW feature.
  • Long double and quad precision are not supported.
  • Multi-threaded FFTW and distributed-memory FFTW MPI interfaces are not supported.
  • Some FFTW concepts exist but are not functional in the cuFFTW interface, such as selected wisdom and cost/plan-inspection functions.

NVIDIA context

cuFFTW sits between legacy CPU FFTW code and the NVIDIA CUDA FFT stack. Use it when source compatibility and incremental porting matter. Use cuFFT directly for full CUDA control, cuFFTDx for in-kernel FFT fusion, cuFFTMp for distributed multi-process FFTs, and NVPL-FFT when the FFTW-compatible target is the NVIDIA Grace CPU rather than a GPU.

Connections

  • cuFFT - cuFFTW is the FFTW3-compatible interface layer for cuFFT.
  • cuFFTDx - device-side FFT sibling for fusing FFTs into CUDA kernels.
  • cuFFTMp - distributed FFT sibling for multi-GPU and multi-node HPC workloads.
  • NVPL-FFT - CPU-side Grace FFT library with FFTW compatibility, complementary to cuFFTW’s GPU migration role.
  • nvmath-python - Python-facing math layer for NVIDIA FFT workflows.
  • NVIDIA-CUDA - platform and toolkit context for cuFFT/cuFFTW libraries.

Source Excerpts

  • NVIDIA’s current cuFFT documentation describes cuFFT as consisting of cuFFT and cuFFTW libraries.
  • The cuFFT docs present cuFFTW as an FFTW3 interface intended to let FFTW applications use NVIDIA GPUs with minimal source changes.

Graph View

Backlinks