gpufft (Python)

Cross-vendor GPU FFT for Python. One API over two battle-tested libraries:

• cuFFT on NVIDIA (backend="cuda")
• VkFFT on any Vulkan device (backend="vulkan"): AMD, Intel, Apple Silicon (MoltenVK), NVIDIA

pip install gpufft

import numpy as np
import gpufft

x = (np.random.randn(4096) + 1j * np.random.randn(4096)).astype(np.complex64)

# NVIDIA via cuFFT
X = gpufft.fft.fft_1d_c2c_pow2(x, log_n=12, backend="cuda")

# Same call on any Vulkan GPU via VkFFT
X = gpufft.fft.fft_1d_c2c_pow2(x, log_n=12, backend="vulkan")

# Arbitrary (non-power-of-two) sizes
y = (np.random.randn(1000) + 1j * np.random.randn(1000)).astype(np.complex64)
Y = gpufft.fft.fft_1d_c2c(y, n=1000, backend="cuda")

The fft_1d_c2c_pow2 signature matches ferrum-gpu's, so code written against one package runs on the other. ferrum-gpu is the pure-Rust kernel path (transparent, hackable); gpufft is the vendor-library path (near-parity performance, cross-vendor today).

Persistent device handles

dev = gpufft.cuda.Device(0)        # or gpufft.vulkan.Device(0)
X = gpufft.fft.fft_1d_c2c_pow2(x, log_n=12, device=dev)

Runtime requirements

• NVIDIA path: an NVIDIA driver providing libcufft.so (CUDA 13.x).
• Vulkan path: a Vulkan 1.x loader (libvulkan.so.1) and an installed ICD.

Neither library is bundled; both are loaded from the system at runtime.

Status

This release ships 1D complex-to-complex transforms on both backends. 2D/3D, real transforms (R2C/C2R), and f64 are on the roadmap.

Apache-2.0.