AOCL-FFTZ

AOCL-FFTZ is a high-performance Fast Fourier Transform (FFT) library optimized for AMD “Zen”-based CPUs. It efficiently computes FFTs for both complex and real data, supporting a wide range of problem sizes and dimensions. The library is tuned for both single-threaded and multi-threaded execution. AOCL-FFTZ can serve as a drop-in replacement for FFTW in applications like VASP (Vienna Ab initio Simulation Package), offering a similar API with enhanced functionality.

Key Features

• Unified Library Design
  Single binary supporting single/double precision and LP64/ILP64 models
• Versatile FFT Capabilities
  Handles real/complex inputs, in-place/out-of-place transforms, multi-dimensional and batched FFTs
• Performance Optimizations
  Vectorized compute kernels (AVX128/256/512) with OpenMP multi-threading and dynamic CPU feature dispatch
• Robust Testing & Validation
  Integrated test bench, unit tests (GTest), fuzz testing, memory analysis (Valgrind/ASAN), and code coverage tracking
• Developer-Friendly Features
  FFTW wrapper for seamless migration, comprehensive logging, auto-generated Sphinx docs
  

Special Highlights

• Optimized to deliver best-in-class performance for VASP workloads
• Performance and portability ensured through automatic runtime selection of the highest-performing kernel available on the host CPU
• Portable C fallback guarantees correctness across all x86-compatible processors
• Flexible descriptor API supports LP64 and ILP64 systems with arbitrarily batched and multi-dimensional transforms
• Memory-efficient design using half-complex output layout for real transforms to minimize footprint

Documentation

• AOCL-FFTZ API Guide
• Source Code: GitHub

Downloads