(mirrors: philippos.info, GitHub)
This repository contains a softcore built from scratch for supporting custom SIMD instruction development by Philippos Papaphilippou @ Imperial College London. See below for additional material from the publication, including a video tutorial.
The source code is divided into 4 directories. A separate README file accompanies each directory, to explain how the source code can be used to reproduce the experiments in the paper.
RTL_and_simulation This folder contains the main code of the softcore for simulation, as well as the template and examples for custom SIMD instructions. Benchmarks This directory contains the benchmarks and micro-benchmarks (sort & prefix sum) used in the evaluation section.FPGA_Implementation_AXI_peripheral The FPGA directory provides additional Verilog code for interfacing with the an AXI interconnect (such as for Avnet's Ultra96 with the Xilinx device ZU3EG) and the relevant information for creating a Vivado projectARM_Linux_Client_for_Ultra96 The optional Linux (for ARM) client mainly provides C code for interfacing with the softcore in our heterogeneous platform. Having Linux and the specific board is not necessary, as AXI is a versatile protocol, at least for Xilinx boards. Though bridges to other protocols exist as well. In order to run the simulations, Icarus Verilog is the main requirement for running the provided examples (tested with version 11.0 (stable) (v11_0)). For the debugging functionality through waveforms, a waveform viewer is recommended, such as GTKWave Analyzer (tested on version 3.3.108).
For experimentation on a Xilinx FPGA, the freely-available Webpack version of Vivado is enough, given that your FPGA board is mentioned in the release notes (see here).
For software development for the softcore, we need to install the RISC-V GNU Compiler Toolchain, and modify it for supporting the custom SIMD instructions for inline assembly. Note that for simply testing that simulator works there already is a sample binary at RTL_and_simulation by default.
First we fetch the code using the bash command (6.65GB of data, about 10GB after compilation):
$ git clone https://github.com/riscv/riscv-gnu-toolchainand change directory:
$ cd riscv-gnu-toolchainUpdate - the file structure and behaviour of the RISC-V GNU toolchain has changed since the softcore has been developped. Revert to a tested version of that repository using the following commands:
$ git switch b715e4f --detachThen we provide the target directory where GCC and related tools are installed. This is for bare-metal applications, based on the 32-bit "RV32I" base specification and "M" multiplication/division extension. The size of the prefix directory will become about 1.1 GB after compilation.
$ ./configure --prefix=/opt/riscv32im_custom --with-arch=rv32im --with-abi=ilp32 Before compiling, we need to add the support of custom instructions to binutils (see next section).
Finally, we compile the tools by running make, preferably with multiple-threads (15 in the example), as it can take a while.
$ make newlib -j15In order to do that, we add the following lines to the C file riscv-gnu-toolchain/riscv-binutils/opcodes/riscv-opc.c inside the lengthy const struct riscv_opcode riscv_opcodes[] array:
{"c0_lv", 0, INSN_CLASS_I, "d,s,t,j", MATCH_CUSTOM0, MASK_CUSTOM0, match_opcode, 0 },
{"c0_sv", 0, INSN_CLASS_I, "d,s,t,j", MATCH_CUSTOM0_RS1, MASK_CUSTOM0_RS1, match_opcode, 0 },
{"c1", 0, INSN_CLASS_I, "d,s,j", MATCH_CUSTOM1, MASK_CUSTOM1, match_opcode, 0 },
{"c2", 0, INSN_CLASS_I, "d,s,j", MATCH_CUSTOM2, MASK_CUSTOM2, match_opcode, 0 },
{"c3", 0, INSN_CLASS_I, "d,s,j", MATCH_CUSTOM3, MASK_CUSTOM3, match_opcode, 0 },where c0_lv, c0_sv are two custom instructions for vector load and store respectively, using the same opcode and the S' instruction type. c1, c2 and c3 are for the custom instructions, and are using the I' instruction type. Currently, c1, c2 and c3 are used for the merge, sort and prefix SIMD instruction examples, but can be replaced as well as increased and aliased. For more information on the pre-defined opcodes for custom instructions, and for adding more, see file riscv-gnu-toolchain/riscv-binutils/include/opcode/riscv-opc.h. The equivalent files are also included in the riscv-gnu-toolchain/riscv-gdb, should this be of use.
In the later versions of the riscv-gnu-toolchain, the custom opcodes need to be added manually inside the file riscv-gnu-toolchain/riscv-binutils/include/opcode/riscv-opc.h. For example, the following snippet can be copied from an earlier version of riscv-opc.h from here, as shown below.
#define MATCH_CUSTOM0 0xb
#define MASK_CUSTOM0 0x707f
#define MATCH_CUSTOM0_RS1 0x200b
#define MASK_CUSTOM0_RS1 0x707f // ...
#define MATCH_CUSTOM1 0x2b
#define MASK_CUSTOM1 0x707f // ...
#define MATCH_CUSTOM2 0x5b
#define MASK_CUSTOM2 0x707f // ...
#define MATCH_CUSTOM3 0x7b
#define MASK_CUSTOM3 0x707fAs before, the toolchain now needs to be recompiled to include the changes:
$ make newlib -j15The following license applies to this source code, except on code of other projects like some benchmarks in the Benchmarks folder.
Copyright 2021-2024 Philippos Papaphilippou
Simodense Software License v1.0
You are free to use, learn from, modify and distribute this creative work under
the following conditions:
1. No code, text, or other elements of this work can be used as an input to models
for the purposes of training. Generative A.I. and language models are prohibited
from using this work. These do not include search engines and related tools whose
main purpose is to index and appropriately point to the original source and author.
2. It comes under no warranties.
3. Any modification to the code shall be made available in the style of GNU General
Public License v2.0 (GPL-2.0), but published under the same license (not GPL-2.0,
and clause 7 still refers to the original copyright holder).
4. A proper attribution of the author (copyright holder) is required when the work
is used. (If applicable, it would be appreciated to cite the related academic
publication that introduced the work.)
5. It cannot be used for commercial purposes unless specific permission is given by
the author.
6. A redistribution shall include this license in its entirety.
7. The author has the right to change this license for future versions of this work,
as well as to update and clarify the author's original intentions of the current
version, such as with regard to what is considered "fair use" by the author for
future entities and technologies. Philippos Papaphilippou, Myrtle Shah "FPGA-extended General Purpose Computer Architecture" The 18th International Symposium on Applied Reconfigurable Computing (ARC) 2022 pdf link source (framework is based on Simodense) program slides video bib
(cite this) Philippos Papaphilippou, Paul H. J. Kelly and Wayne Luk "Simodense: a RISC-V softcore optimised for exploring custom SIMD instructions" The International Conference on Field-Programmable Logic and Applications (FPL) 2021 pdf link source video program bib
Philippos Papaphilippou, Paul H. J. Kelly and Wayne Luk "Demonstrating custom SIMD instruction development for a RISC-V softcore (demo abstract)" The International Conference on Field-Programmable Logic and Applications (FPL) 2021 pdf link source video bib
Philippos Papaphilippou, Paul H. J. Kelly and Wayne Luk "Extending the RISC-V ISA for exploring advanced reconfigurable SIMD instructions" The Fifth Workshop on Computer Architecture Research with RISC-V (CARRV 2021) (co-located with ISCA 2021) pdf slides source video program bib
General comments or feedback on this repository has been provided by: Tim Todman, Hakam Atassi, Yutaka Nagashima