Integer-MPA-Coprocessor

Open-Source Coprocessor for Integer Multiple Precision Arithmetic (MPA)

Programming language: VHDL/C/C++/SystemVerilog

---

Introduction

The purpose of the project is to develop the VHDL code for FPGA allowing to accelerate MPA computations. The developed device should work as a coprocessor for standard CPU. As a starting point, the basic arithmetic operations are chosen for implementation in FPGA, i.e., addition, subtraction, multiplication. The device should support the integer operations based on the sign-magnitude representation of data (i.e., it should be similar to the GMP standard).

Scientific work

If the code is used in a scientific work, then reference should be made to the following publication:

Rudnicki, K.; Stefański, T.P.; Żebrowski, W. Open-Source Coprocessor for Integer Multiple Precision Arithmetic. Electronics 2020, 9, 1141. (https://www.mdpi.com/2079-9292/9/7/1141)

This publication also includes the description of the coprocessor architecture and results of benchmarks. It also includes the scheme of implementation on TySOM-1 board from Aldec.

---

Manual

The branch master includes three main parts:

./doc

./firmware

./software

The branch nf-tysom includes files specific for the implementation on TySOM-1 board from Aldec.

Doc

It includes a few documentation files.

Firmware

The Vivado Design Suite v2018.3 from Xilinx is employed as a software tool for this coprocessor.

Firmware is grouped within the following directories:

• bd - this directory contains the general board files (for zedboard)
• scripts - this directory contains scripts for firmaware generation
• sim - this directory contains simulation files
• src - this directory contains source files

Software

All software can be compiled with the use of the script: compile_all_programs.sh. Executable and object files can be removed with the use of the script: clean_all_programs.sh. For compilation, Linux with GCC and GMP library is required.

Software is grouped within the following directories:

• common - this directory contains useful functions used by all codes

• data_apps - this directory contains codes generating input files for the coprocessor

  • 1ddgf_data - this code generates files for computations of 1-D discrete Green's function
  • addition_data - this code tests adition in the coprocessor
  • factorial_data - this code generates files for computations of factorial
  • powernn_data - this code computes n^n
  • random_data - this code generates files for random computations on random data
  • progwriter - this code allows to generate the coprocessor code with the use of GUI

• DPI - this directory contains codes for DPI tests

  • dpi_module - this code allows us to test multiplication and addition in the coprocessor
  • dpi_core - this code allows to test the coprocessor core
  • dpi_test - this code allows us to test DPI interface

• emusrup - this code emulates the coprocessor with the use of GMP (i.e. EmuSRuP = "Emulate Stefanski Rudnicki MicroProcessor")

To use emusrup for verification, one needs to execute in SV testbech functions as follows:

* tbEmusrupStart - initializes emusrup for DPI debugging
* tbEmusrupProceed - emusrup proceeds a single step of the code
* tbEmusrupCheckLogic - returns values of logical registers in the coprocessor
* tbEmusrupCheckShadow - returns values of shadow registers in the coprocessor
* tbEmusrupStop - finalizes emusrup execution for DPI debugging

When tbEmusrupStart is called from DPI testbench, the arrays for storing handle and semaphore must be allocated at SV side. It stems from the fact that SV does not provide (**char) type allowing to return pointer to allocated memory. Only equivalent data types string<->(char *) are available. Hence, memory for string must be allocated in SV testbench.

Emusrup code is located in ./software/emusrup directory whereas its compilation for DPI verification is executed by script in ./software/DPI/emusrup_debug directory. Results of compilation go to ./firmware/sim/lib directory.

• helpers - some useful and helpful codes

  • bin2coe - this code converts *.bin files into *.coe type
  • bin2mem - this code converts *.bin files into *.mem type
  • mult_regs_address - this code presents multiplication operation for multiple-precision numbers

• runtime - this directory contains codes executable on CPU for the coprocessor benchmarking

  • 1ddgf_runtime - this code computes 1-D discrete Green's function on CPU
  • factorial_runtime - this code computes factorial on CPU
  • pownn_runtime - this code computes n^n on CPU
  • add - this code tests addition speed on CPU
  • div_vs_mult - this code allows one to compare division and multiplication

• vhdl_gens - this directory contains codes for vhdl generation

  • mux_gen - this code generates vhdl code of multiplexer
  • or_gen - this code generates or structure
  • wrapper_gen - this code generates vhdl wrapper

After compilation, one can check program arguments executing it without any argument.

Authors

The project is developed by

• Kamil Rudnicki - VHDL codes, firmware development
• Tomasz Stefanski - C/C++ codes, software development
• Wojciech Zebrowski - Implementation on TySOM-1 board, evaluation and benchmarking

License

This is an open-source code licensed under the Mozilla license.