QuadriSparse

QuadriSparse is a sparse dense matrix mulitplication (SpMM) accelerator and RISC-V ISA extention based on the matrix multiplication co-processor Quadrilatero. It uses the CORE-V-X-IF interface to interface with OpenHW Group CPUs and the OBI protocol to interface with memories.

This project was developed as part of a master's thesis at Chalmers Univeristy of Technology.

Dependencies

• Verilator: SV simulator
• Bender: dependency management tool available here
• Make
• Python3

Usage

Setup

Ensure the dependencies above are installed:

verilator -V
bender -V

Setup Python virtual envirmoment:

python -m venv venv
source venv/bin/activate
pip install numpy

Running simulation

Generate test data. SIZE: number of rows, SPARSITY: Ammount of sparsity 0-1, MAXVAL: Maximum value of the elements (optional).

make matgen SIZE=16 SPARSITY=0.8 MAXVAL=15

Compile and run. MODE (sparse, dense) switches between the new sparse optimized multiplication and the traditional dense.

make run DATA_PREFIX=mat_16_0.8 SIZE=16 MODE=sparse

Notes

If you want to bring your own test data it has to be formatted as follows:

• All files are flat text files containing one hex formatted 32 bit number per row
• The sparse matrix in CSR format consisting of 3 files, xx_a_row.hex, xx_a_col.hex, xx_a_val.hex
• The dense matrix file: xx_b.hex
• The result reference matrix: xx_ref.hex
• xx is the DATA_PREFIX argument in the run command

ISA Extention

QuadriSparse is based on a RISC-V matrix extension available here. Below are listed the instructions added by this project and their encodings.

All instructions share 7'b0101011 (CUSTOM 1) as the major opcode, and func3 is 3'b000.

mnemonic	31-27	26-25	24	23-21	20-18	17-15	14-12	11-10	9-7	6-0
SPLD_W	00100	00	0	000	000	nnz $^3$	func3	10	md	0101011
DLD_W	00010	00	0	000	000	ms1 $^1$	func3	10	md	0101011
SPMAC_W	11110	00	0	ms1 $^1$	ms2 $^2$	md	func3	10	000	0101011

1. Sparse register
2. Dense register
3. The numeber of non zero elements to load form the CSR values array

Directory Structure

• /rtl contains the SystemVerilog files describing the co-processor
• /sw contains example programs that can be used with the x-heep platform as well as helper functions to generate test data
• /tb contains a standalone testbench which can be used to verify the functionality of the accelerator

Limitations

• DLD_W instruction requires the dense matrix's number of columns to be divisible by 4 or zero-padded
• The instructions are completely un-pipelined

Licence

Unless otherwise specified in their respective file headers all files in this repository are made available under Apache License v2.0 (Apache-2.0). Most RTL files are licenced under the Solderpad Hardware License v2.1 (SHL-2.1), see LICENCE.md.