Out-of-core GPU 2D-shift-FFT algorithm for ultra-high-resolution hologram generation.

Area of Science:

• Computational Imaging
• Computer Graphics
• High-Performance Computing

Background:

• Generating ultra-high-resolution holograms necessitates processing large complex matrices that often exceed standard GPU memory capacities.
• Existing methods struggle with memory constraints when dealing with massive datasets required for advanced holographic applications.

Purpose of the Study:

• To develop a novel out-of-core GPU algorithm for 2D-Shift-FFT (Fast Fourier Transform) to efficiently generate ultra-high-resolution holograms.
• To address the challenge of limited GPU memory when processing large complex matrices (e.g., 100K^2) for holographic applications.

Main Methods:

• Employs a 1D-FFT based 2D-FFT computation method for out-of-core processing.
• Optimizes performance through data transposition for continuous memory layout, combined FFT-shift and transposition, concurrent execution on heterogeneous systems, cache-friendly chunk generation, and pinned-memory buffers.

Main Results:

• Achieved up to 3.24x (single-precision) and 3.06x (double-precision) performance improvement over conventional cuFFT implementations for large-scale complex matrices.
• Reduced hologram generation time by up to 28% in layer-based computer-generated holography (CGH) for 100K^2 holograms.

Conclusions:

• The proposed out-of-core GPU algorithm offers significant efficiency and usefulness for generating ultra-high-resolution holograms.
• Demonstrates a practical solution for memory-bound computational challenges in advanced holographic display technologies.