Efficient FPGA Implementation of Conjugate Gradient Methods for Laplacian System using HLS

TL;DR

This paper presents an FPGA-based pipelined and superscalar implementation of conjugate gradient methods for solving Laplacian systems, achieving significant latency reductions and speedups over software solutions through novel domain decomposition and parallel stencil operations.

Contribution

The paper introduces a high-level synthesis FPGA design for conjugate gradient methods with optimized 2D domain decomposition, reducing latency and improving speed compared to traditional software implementations.

Findings

FPGA implementation reduces latency significantly.

New 2D domain traversal halves padding latency.

Implementation is 10 times faster than software for large systems.

Abstract

In this paper, we study FPGA based pipelined and superscalar design of two variants of conjugate gradient methods for solving Laplacian equation on a discrete grid; the first version corresponds to the original conjugate gradient algorithm, and the second version corresponds to a slightly modified version of the same. In conjugate gradient method to solve partial differential equations, matrix vector operations are required in each iteration; these operations can be implemented as 5 point stencil operations on the grid without explicitely constructing the matrix. We show that a pipelined and superscalar design using high level synthesis written in C language leads to a significant reduction in latencies for both methods. When comparing these two, we show that the later has roughly two times lower latency than the former given the same degree of superscalarity. These reductions in latencies for the newer variant of CG is due to parallel implementations of stencil operation on subdomains of the grid, and dut to overlap of these stencil operations with dot product operations. In a superscalar design, domain needs to be partitioned, and boundary data needs to be copied, which requires padding. In 1D partition, the padding latency increases as the number of partitions increase. For a streaming data flow model, we propose a novel traversal of the grid for 2D domain decomposition that leads to 2 times reduction in latency cost involved with padding compared to 1D partitions. Our implementation is roughly 10 times faster than software implementation for linear system of dimension $10000 \times 10000.$