[Paper Review] Optimized M2L Kernels for the Chebyshev Interpolation based Fast Multipole Method

Matthias Messner, Bérenger Bramas+2 more|arXiv (Cornell University)|Oct 27, 2012

Electromagnetic Scattering and Analysis13 references17 citations

TL;DR

This paper optimizes the multiple-to-local (M2L) kernel in the Chebyshev interpolation-based fast multipole method (bbFMM/dFMM) by introducing individual low-rank approximations per cluster pair and symmetry-based permutations, reducing precomputation time by up to 340× and accelerating matrix-vector products by 4–6× via optimized BLAS-level matrix-matrix operations, with IAblk variant outperforming SArcmp in precomputation and memory, especially for single matrix-vector products.

ABSTRACT

A fast multipole method (FMM) for asymptotically smooth kernel functions (1/r, 1/r^4, Gauss and Stokes kernels, radial basis functions, etc.) based on a Chebyshev interpolation scheme has been introduced in [Fong et al., 2009]. The method has been extended to oscillatory kernels (e.g., Helmholtz kernel) in [Messner et al., 2012]. Beside its generality this FMM turns out to be favorable due to its easy implementation and its high performance based on intensive use of highly optimized BLAS libraries. However, one of its bottlenecks is the precomputation of the multiple-to-local (M2L) operator, and its higher number of floating point operations (flops) compared to other FMM formulations. Here, we present several optimizations for that operator, which is known to be the costliest FMM operator. The most efficient ones do not only reduce the precomputation time by a factor up to 340 but they also speed up the matrix-vector product. We conclude with comparisons and numerical validations of all presented optimizations.

Motivation & Objective

• To reduce the high precomputation cost of the M2L operator in the black-box FMM (bbFMM) and directional FMM (dFMM), which is the most expensive FMM component.
• To improve performance by exploiting geometric symmetries in cluster interactions to minimize redundant computation and storage of M2L operators.
• To accelerate matrix-vector products by transforming repeated matrix-vector operations into optimized matrix-matrix products using symmetry-induced permutations.
• To evaluate the trade-offs between precomputation cost and runtime efficiency for different M2L variants in both smooth (bbFMM) and oscillatory (dFMM) kernel scenarios.
• To identify optimal algorithmic variants for two distinct use cases: single matrix-vector product (fast precomputation) vs. iterative linear system solves (fast M2L application).

Proposed method

• Introduces individual low-rank approximation of M2L operators per cluster pair, replacing the global SVD used in prior work, to achieve optimal compression per interaction.
• Identifies and exploits spatial symmetries (e.g., reflections across x=0, y=0, z=0 planes) to represent all 316 M2L operators in bbFMM as permutations of just 16 unique operators.
• Uses symmetry permutations to convert 189 matrix-vector products into 16 matrix-matrix products, enabling use of highly optimized BLAS libraries like Intel MKL.
• Proposes two new variants: IAblk (individual approximation + symmetry-based blocking) and SArcmp (recompression of suboptimal SVD-based M2L), both improving on the original SA method.
• Employs transfer vectors to uniquely identify M2L operators and maps kernel type (smooth or oscillatory) to interaction list structure and operator behavior.
• Validates performance using three geometries (sphere, oblate spheroid, prolate spheroid) and measures precomputation time, M2L application time, and accuracy across varying accuracy levels.

Experimental results

Research questions

• RQ1 Can individual low-rank compression of M2L operators per cluster pair reduce computational cost below the global SVD-based approach?
• RQ2 To what extent can geometric symmetries in the FMM tree structure reduce the number of unique M2L operators that need to be precomputed?
• RQ3 How do symmetry-based permutations enable the use of optimized matrix-matrix operations, and what performance gains do they yield?
• RQ4 What is the trade-off between precomputation time and M2L application speed for different M2L variants in bbFMM and dFMM?
• RQ5 Which variant—IAblk or SArcmp—is optimal for single matrix-vector products versus iterative linear system solves?

Key findings

• The IAblk variant reduces precomputation time by up to 340× compared to the original SA method, primarily due to symmetry-based reduction from 316 to 16 unique M2L operators.
• IAblk accelerates M2L application by 4–6× in bbFMM and 1.2–2.7× in dFMM by enabling optimized matrix-matrix operations via BLAS libraries.
• For single matrix-vector products, IAblk is optimal: precomputation time is 0.4 s vs. 69.1 s for SArcmp, and total time is 10.4 s vs. 75.4 s.
• For iterative linear system solves, SArcmp becomes faster than IAblk after approximately 19 matrix-vector products at accuracy Acc=5, and after 26 products at Acc=6.
• The IAblk variant achieves the lowest computational cost and memory footprint, especially beneficial for non-directional expansions (bbFMM and low-frequency dFMM).
• The individual low-rank approximation (IA variants) outperforms the global SVD-based approach (SA) in both precomputation and runtime, with IAblk being the most efficient overall for single-use scenarios.