Abstract
The computation of Lagrangian coherent structures (LCS) has become a standard tool for the analysis of advective transport in unsteady flow applications. LCS identification is primarily accomplished by evaluating measures based on the finite-time Cauchy Green (CG) strain tensor over the fluid domain. Sampling the CG tensor requires the advection of large numbers of fluid tracers, which can be computationally intensive, but presents a large degree of data parallelism. We discuss the use of Visualization Toolkit (VTK) libraries as a foundation for object-oriented LCS computation, and how this framework facilitates integration into flow visualization software such as ParaView. We also discuss the development of CUDA GPU kernels for efficient parallel spatial sampling of the flow map, including optimizing these kernels for better utilization.
Introduction
In fluid dynamics, understanding how substances transport and mix in unsteady (time-varying) flows is a major challenge. Lagrangian Coherent Structures (LCS) serve as the "skeletons" of these flows—stable and unstable manifolds that govern fluid motion. However, identifying LCS is computationally demanding, requiring the seeding of millions of virtual tracers and integrating their trajectories to calculate the Finite-Time Lyapunov Exponent (FTLE) or the Cauchy-Green (CG) strain tensor. Historically, researchers faced hurdles regarding both computational cost and the usability of specialized codes that lacked integration with standard visualization tools.
Modular Pipeline Architecture
To address these challenges, a modular, object-oriented pipeline was developed using the Visualization Toolkit (VTK). The architecture is organized into a series of filters that handle data ingestion, computation, and visualization. This modularity allows for easy extension, such as adding new interpolators or specialized LCS measures.
By building on VTK, the tool can be compiled as a ParaView plugin. This enables researchers to compute LCS and visualize them alongside other flow data—such as pressure or velocity—in a single, unified environment, significantly streamlining the research workflow.
GPU Acceleration and Optimization
The most computationally intensive components—particle advection and flow map computation—are accelerated using NVIDIA’s CUDA. A key technical innovation in this pipeline is the use of "kernel splitting." To maximize GPU performance, the 4th-order Runge-Kutta (RK4) integration was broken into smaller, separate kernels. This approach reduces "register pressure" (the amount of fast memory each thread uses), allowing for higher occupancy—meaning more threads can run simultaneously on the GPU hardware.
As shown in the performance data, splitting a single complex integration kernel into smaller ones resulted in up to a 30% decrease in runtime for complex unstructured grids. This optimization is particularly effective as the number of trajectories scales into the millions.
Precision and Performance Results
The pipeline was evaluated using benchmarks ranging from 2D Cartesian grids to complex 3D unstructured grids, such as a patient-specific Abdominal Aortic Aneurysm (AAA) model.
- Speedups: Compared to serial CPU implementations, the GPU pipeline achieved approximately a 70x speedup for 2D Cartesian data and a 12x speedup for the most complex 3D unstructured data.
- Numerical Accuracy: The study highlighted the necessity of double-precision (64-bit) math for complex 3D flows. While single-precision is faster, it proved unreliable for the AAA model, where errors were large enough to advect tracers into incorrect arterial branches. Using double precision, the GPU results were identical to CPU results (error < $1 \times 10^{-10}$).
Conclusion
This framework provides a high-performance, accessible solution for LCS analysis on standard workstations. While currently limited by GPU VRAM for extremely large datasets, the modular VTK-based approach ensures that the pipeline can be easily updated with new features, such as strainline computation or spatial streaming, to further enhance its utility in fluid mechanics research.
Summary updated