CUDA optimization strategies for compute- and memory-bound neuroimaging algorithms.

Area of Science:

• Computer Science
• Neuroimaging
• High-Performance Computing

Background:

• Neuroimaging algorithms are increasing in complexity and resolution, outpacing traditional CPU performance.
• Graphical Processing Units (GPUs) offer a data-parallel architecture capable of significant computational speedups.
• Massively threaded GPU architectures present challenges when resources are exceeded.

Purpose of the Study:

• To present optimization strategies for compute- and memory-bound algorithms on the CUDA architecture.
• To address challenges posed by GPU resource limitations in neuroimaging computations.
• To enhance the performance of computationally intensive neuroimaging algorithms on GPUs.

Main Methods:

• For compute-bound algorithms: reduced registers via shared memory, increased data throughput via heavier thread workloads, and optimized thread configuration.
• For memory-bound algorithms: data reorganization into self-contained structures, multi-pass approach, and optimized memory resource selection for reduced latency.
• Strategies demonstrated on 3D unbiased nonlinear image registration and non-local means surface denoising.

Main Results:

• Optimized GPU implementations achieved up to 6x faster performance than unoptimized versions.
• Achieved peak GPU speedups of 129x for 3D unbiased nonlinear image registration compared to CPU.
• Achieved peak GPU speedups of 93x for non-local means surface denoising compared to CPU.

Conclusions:

• The presented optimization strategies effectively address GPU resource limitations for complex neuroimaging algorithms.
• Significant performance gains were realized for both compute- and memory-bound tasks, demonstrating the efficacy of the proposed methods.
• GPU acceleration, through these optimized strategies, is crucial for advancing neuroimaging research and applications.