wangmaolin/rsqp

The Reconfigurable Solver for QP

The Reconfigurable Solver for QP

RSQP is a framework for generating FPGA solvers for QPs. Some off-the-shelf bitstreams are provided to solve generic QPs. If the sparsity strucutre of QPs to be solved are known, RSQP can analyze the sparsity structure and build a customized FPGA architecture for extra speed boost while solving a series of QPs with the same sparsity pattern.

System Requirements

This framework requries the following system specifications to run:

• Operating System: Ubuntu 18.04
• FPGA Card: Alevo U50 or U280

Please check this guide for detailed steps of installing the FPGA card. To verify the installation, first run the following command to make sure the card is visable:

xbutil query

Part 1. Hello Solver

A pre-synthesised FPGA architecture is in the bitstream folder. The pre-compiled binary of QP problem to run on this architecture is in the elf folder. Run the following script to see the solver in action 🚀

./fpga-solve.sh -sw=svm-s10-2-bf-949-194.fpga -hw=u50-2-bf-3259-275.xclbin

Part 2. Compile a QP to Run on an Off-the-shelf Architecture

Generating the elf file:

./sw-elf-gen.sh -c=2 -a=bf -start=10 -end=10

To check solver states (values of each vector) in each step, run the functional simulation of the elf:

./hw-hls-gen.sh -v=1 -c=2 -a=bf -cvb=3259 -vec=275

./func-sim.sh -sw=svm-s10-2-bf-949-194.fpga -hw=hbm2-bf-3259-275

Part 3. Synthesis a Customzied Architecture for a Specific Sparsity Pattern

To synthesis a new FPGA architecture, Vitis 2021.1 is needed. Install the tool using this guide.

./hw-hls-gen.sh -c=2 -a=bf -cvb=3259 -vec=275 -b=u50 -g=1

Results

Citation

If you found our work useful in your research, please consider citing our paper at

@inproceedings{rsqp,
Author = {Wang, Maolin and McInerney, Ian and Stellato, Bartolomeo and Boyd, Stephen and So, Hayden},
Title = {RSQP: Problem-specific Architectural Customization for Accelerated Convex Quadratic Optimization},
Booktitle  = {2023 ACM/IEEE 50th Annual International Symposium on Computer Architecture (ISCA)},
Year = {2023}
}