Google Summer of Code 2022 - Ideas for Projects

FOSSi Foundation is applying as an umbrella organization for Google Summer of Code 2022. That means that we give small projects the chance to participate in the program. Below you can find a list of ideas that community mentors had, but mentees are encouraged to propose their own ideas. These projects are mostly open-ended and can be tailored to your level of experience, assuming that you have the appropriate set of required skills for the particular project idea.

Whether you’re an aspiring mentee or mentor, feel free to contact us, either through the private GSoC-specific mailing list gsoc@fossi-foundation.org, through the public discussion mailing list, or through the mentors listed for each project below. We are also available on Gitter in librecores/Lobby.

Looking forward to meet you all!

TOC {:toc}

Jupyter Notebook add-ons for Open Silicon development

Details: Jupyter Notebook allows developers to share interactive documents that allow to reproduce a given computation, by capturing in a single file: the environment with the tools installation the code with a given set of parameters the results and their analysis

This is a natural fit for Open Silicon design that often tuning input parameters, inspecting visual results and comparing metrics and simulation measurements.

Possible contributions to this project:

Create Jupyter extensions that interface the common Open Silicon data formats (GDS, LEF, DEF, Waveform, Timing and Power analysis report) with its display primitives.
Create Jupyter kernels to support writing cells in HDL and HSL languages and evaluating their testbenches.
Create a corpus of notebook examples that capture popular Open Silicon flows (analog, digital, hls-based). Bootstrap the developer infrastructure and community to maintain those projects.

Skill level: Beginner to Intermediate

Duration: ~350 hrs.

Language/Tools: Python, Typescript

Mentor: Johan Euphrosine

TinyParrot: A minimal BlackParrot RISC-V Multicore variant

Details: BlackParrot aims to be the default open-source, Linux-capable, cache-coherent, RV64GC multicore used by the world. It has been FPGA and Silicon-validated as an industry-strength design with leading efficiency. We wish to support open-source toolchains for FPGA and ASIC flows, but current open-source SRAM generators have not quite matured to make this seamless. This project will shrink BlackParrot to have extremely tiny caches, provide minimal ISA support and have other logic-saving optimizations with the goal of being able to easily synthesize to gates.

Skill level: beginner-intermediate

Project length: medium (175 hours)

Mentors: Dan Petrisko, Michael Taylor

Language/Tools: SystemVerilog, some knowledge of computer architecture, RISC-V knowledge preferred but not required

Logical Equivalence Checks with LLHD

The Low Level Hardware Description language, or LLHD in short, is simple intermediate representation that is able to capture the semantics of today's hardware description languages including their behavioral, verification, and testing features. The IR has been developed alongside the Moore compiler and published as a paper at PLDI 2020. Since then, we have been busy merging LLHD into the CIRCT project, a larger joint effort to develop Circuit IR Compilers and Tools.

When working with LLHD, you can assume that all the nasty parts of languages like SystemVerilog or VHDL have already been taken care of by a language frontend, and you are dealing with a much simpler but still complete representation of a circuit. As of today there are already a basic reference simulator as well as a faster LLVM-JIT-compilation-based simulator available for LLHD.

As a proof of concept for LLHD's verification prowess we would like you to implement a basic Logical Equivalence Check (LEC) for LLHD designs. This is essentially the process of taking two circuits and formally proving (or disproving) that they logically do the same thing, even if they use a different combination of logic gates to do so. This is a crucial step in chip design flows, to ensure that a synthesizer has properly translated RTL into a gate-level implementation, and that further place-and-route work did not violate the operation of a design.

In this project you will take two LLHD designs, translate them into the fundamental boolean equations, and formulate a satisfiability problem for an existing optimized SAT/SMT solver with them. You can scale this project's complexity to your liking, for example simplifying by looking only at combinational circuits, or complicating by finding ways to use registers as invariant anchor points in a circuit.

SAT/SMT solvers are tools that are extremely fun and satisfying to work with, but generally don't get the love they deserve in the circuit design community. With this project you'll be able to contribute to a key process in the ASIC design flow and generally show that, with good existing abstractions such as LLHD and the standardized solver file formats, it is very easy to create seemingly daunting formal tools for an open ASIC world.

Skill Level: Intermediate to Advanced

Duration: 175 hours or 350 hours

Language/Tools: C++, SAT/SMT Solver (e.g. Z3, Boolector, Bitwuzla, or others)

Mentor: Fabian Schuiki

Building Manycore SoCs with OpenPiton + LiteX

LiteX makes building FPGA-based SoCs easy. Using the Python hardware design library Migen, LiteX provides a variety of peripherals to enable users to build a complex SoC around a core of their choice. For this project, we would like to connect a manycore OpenPiton processor design in order to build a new manycore LiteX SoC.

Skill level: Intermediate

Duration: 175 hours or 350 hours

Language/Tools: Python (Migen), Verilog

Mentor: Jonathan Balkind

Minimal RISC-V core with AI Acceleration synthesizable with open source tools

A series of projects between Embecosm and Southampton University over the past and a FOSSi GSoC project from 2021 have developed a simple (8 instruction) open source ISA extension for the CV32E40P and CV32E40X RISC-V cores to accelerate neural network inference acceleration. The project GitHub has demonstrated a 5 fold increase in inference performance. YouTube has the most recent presentation of the project, given at the London Open Source RISC-V meetup in January 2022.

This demonstrated the potential benefit of this approach, but could only be synthesized using proprietary FPGA synthesis tools. This project proposes trying to create a stripped down version which can be synthesized using the Yosys open source tool suite.

The core is delivered within the CORE-V MCU, a SoC wrapper which provides a debug interface, memory and a range of peripherals. In order to make this project tractable, this will be stripped down to a minimal MCU providing just the debug interface and memory alongside the core. CORE-V MCU uses FuseSoc for ease of development, but the AVA work has not picked this up, so cleaning up the code to use FuseSoc will be an important starting point.

There are a series of steps in this project, so it can be provided as either a medium or long project. It is also quite open ended, so applicants should consider how they would wish to craft the project to directions that most interest them.

The deliverables of this project would be for the medium project:

the stripped down, synthesizable with Verilator as a model;
the stripped down CORE-V MCU with the CV32E40P core, synthasizable for FPGA with Yosys; and
Embench benchmarking of the stripped down CORE-V MCU on FPGA.

The long project would then extend these deliverables with the AI interface using the CV32E40X core:

a modified stripped down CORE-V MCU using the CV32E40X core with AI instruction set extensions cleaned up to build completely under FuseSoc.
a modified stripped down CORE-V MCU using the CV32E40X core with AI instruction set extensions as a Verilator model;
a modified stripped down CORE-V MCU using the CV32E40X core with AI instruction set extensions, synthasizable for FPGA with Yosys; and
measurement of the impact on AI inference using TinyMLPerf.

Skill level: Advanced (particularly FPGA synthesis using Yosys)

Project length: medium or long

Mentors: Jeremy Bennett (email) and William Jones (email)

Language/Tools: Verilog, SystemVerilog, RISC-V, Yosys open source tool suite, TinyML benchmark suite

OpenRISC Benchmarking and Performance improvements

The OpenRISC CPU architecture has multiple CPU implementations including the mor1kx and marocchino. Recent testing has shown that memory access on the marocchino is slightly slower compared to the mor1kx.

This project will have the student:

Use tools like the Embench modern benchmark suite to measure OpenRISC processor and compiler toolchain performance.
Document the OpenRISC performance at Embench IoT results to be able to compare OpenRISC vs other popular CPUs.
Track down and improve OpenRISC CPU performance by finding and fixing deficiencies in the verilog designed cores.

Skill level: Advanced

Project Length: large

Language/Tools: Verilog, Shell scripting, C, Assembly, Python

Mentor: Stafford Horne

Circuit Visualization

Details: The Makerchip platform for open-source circuit design has introduced support for custom visualization of circuit simulations. This has been used to great success in RISC-V training and is useful for circuit debug and conveying microarchitectural behavior as well. Any circuit can be visualized in useful ways. Find a few examples here.

Some circuits we would like to provide visibility into include:

SweRV
SERV
TL-Verilog flow library
BaseJump STL
basic circuits for instructional purposes
virtualized FPGA boards (completed in GSoC 2021)
this neural network
gamified circuit design challenges

Visualization is written in (very straight-forward) JavaScript that has access to trace data from simulation and a canvas to draw on, described here.

Skill level: Beginner/Intermediate

Duration: ~175 or ~350 hrs.

Language/Tools: JavaScript, Verilog/SystemVerilog/TL-Verilog

Mentor: Steve Hoover (email), Alternate Mentor (for gamified circuit education) Gayatri Mehta (email) (or other alternates depending on the specific project)

LiteX SMP SoC for OpenRISC

The LiteX project makes creating FPGA-based SoCs easy. LiteX supports creating SoCs containing OpenRISC CPU cores. Up until now however, there have been no LiteX SoCs that support running OpenRISC multicore/SMP Linux. The linux-on-litex-vexrisc project provides a good example of how to develop and document getting Linux up and running on a LiteX SoC; including multicore.

Using linux-on-litex-vexrisc as an example, this project will have the student creating a project to help people get up and running with OpenRISC. The final goal shall be to have a documented multicore OpenRISC LiteX SoC running Linux SMP.

Skill level: Advanced

Project Length: large

Language/Tools: Verilog, LiteX, Linux, Python, OpenRISC architecture

Mentor: Stafford Horne

The World's smallest AI processor

The prize winning bit-serial SERV core, created by Olof Kindgren, is the world's smallest RISC-V processor. In this project, we propose combining SERV with the miinimal AI instruction set extension developed under the AVA project (https://github.com/AI-Vector-Accelerator). This could for example use the CFU-LI interface.

The goal would be to demonstrate the world's smallest processor optimized for AI inference.

Skill level: Intermediate (particularly FPGA synthesis using Yosys)

Project length: medium

Mentors: Jeremy Bennett (email) and Olof Kindgren (email)

Language/Tools: Verilog/SystemVerilog, RISC-V, Yosys open source tool suite

1st CLaaS for Local FPGAs

1st ClaaS supports web application communication with FPGA logic in the cloud. It would be nice to support local FPGA use cases as well. This is useful for local applications as well as to support development that will ultimately be deployed to the cloud.

FuseSoC and Apio are worth considering for vendor-agnostic FPGA support.

Discussion: TL-Verilog User's Slack workspace #1st-claas channel.

Skill level: Advanced

Languages/Tools: FPGAs, Linux, Scripting, Makefiles

Project length:: ~350 hours

Mentors: Theo Omtzigt (email), Steve Hoover (email)

Giving AnyCore an Open-Source FPU

AnyCore is an advanced superscalar processor developed at NC State University, designed to be highly configurable across parameters like issue width and pipeline depth. This project would entail connecting an open-source Floating Point Unit (FPU) to the high performance AnyCore processor, which runs the RISC-V ISA.

Skill level: Beginner/Intermediate

Duration: 175 hours

Language/Tools: Verilog/SystemVerilog, RISC-V

Mentor: Jonathan Balkind

If you are interested in the TL-Verilog ecosystem, you might also consider any of these TL-Verilog project ideas. All would make excellent GSoC projects and mentors can be identified.

Architectural Improvements to OpenPiton+Ariane

OpenPiton+Ariane is a permissively-licensed RISC-V manycore processor, built as a collaboration between the PULP Platform from ETH Zürich and the OpenPiton Platform from Princeton University. We would like to co-optimise OpenPiton and Ariane in their combined platform, to improve performance of the processor both in FPGA emulation systems and for eventual silicon chips. One possible idea this year is to implement the RISC-V cache management operation (CMO) extensions. We are open to many potential projects aimed at improving the performance of aspects of either Ariane or OpenPiton and encourage you to get in touch to discuss ideas.

Skill level: Intermediate

Duration: 175 or 350 hours

Language/Tools: Verilog, SystemVerilog, RISC-V

Mentor: Jonathan Balkind, Nils Wistoff

OpenRISC Linux Feature Development

The OpenRISC Linux kernel support is under constant development but there are certain Linux facilities that are not yet used or available on the OpenRISC platform.

This project will have the student developing, testing and sending patches up to the Linux kernel. This includes:

Develop rseq support to OpenRISC Linux and glibc.
Add floating point support to OpenRISC Linux.
Use the cacheinfo API for reporting CPU details in OpenRISC Linux.
Add virtio support to OpenRISC Linux which would allow running OpenRISC containers.

Skill level: Advanced

Project Length: large

Language/Tools: Linux, C, Assembly, OpenRISC architecture

Mentor: Stafford Horne

Push a Design through the Moore HDL Compiler

Moore is a compiler for SystemVerilog and VHDL hardware designs. Its goal is world domination by finally moving the burden of implementing SV and VHDL out of tools like synthesizers and simulators and into a separate frontend, very much like what Clang and LLVM did for the software world. In contrast to other projects that focus on specific use cases such as synthesis or netlists processing, Moore strives to support the entirety of the SV and VHDL languages. As output the compiler produces LLHD IR, a simple intermediate representation that is able to capture the semantics of today's hardware description languages including their behavioral, verification, and testing features. (See also the LLHD paper from PLDI 2020 for details.)

We would love to have you help make Moore even better! As a proof of concept that goes beyond the single RISC-V core that Moore tackles successfully, we would like to push a larger, more complex compute cluster through the compiler. This could be for example a Snitch compute cluster of ETH Zurich which contains complex caches, interconnects, memory systems, multiple processor cores and large floating-point data paths. The goal of this project is to take this cluster's SV source code and implement the pieces Moore currently lacks in order to be able to compile the cluster. You'll be able to look for frequent sources of errors and juicy implementation targets, implement them in the compiler, and see the errors disappear one by one. If it turns out that not a lot of pieces are missing, you'll be able to fully simulate the cluster with an LLHD-based simulator and push the project forward significantly.

The Moore compiler is written in Rust. Don't be scared if you haven't touched Rust before -- if you know C or C++ you'll feel right at home. The main language we're currently tackling with Moore right now is SystemVerilog, so either knowing the language a bit or not being scared of looking into language reference manual will be useful.

Skill Level: Advanced

Duration: 175 hours or 350 hours

Language/Tools: Rust, SystemVerilog

Mentor: Fabian Schuiki

Embench DSP Extensions

We will be developing a set of new digital signal processing benchmarks to add to the current suite of Embench DSP benchmarks. The starting point will be the current suite of DSP benchmarks (FIR, IIR, and FFT). New benchmarks will be run on a microprocessor and its performance evaluated.

These new benchmarks will be added to the Embench DSP suite and be made available to the world.

There will the opportunity to engage with students at Rice University who are also working on the project.

Skill level: Beginner (accessible to students who have some programming experience and are willing to learn)

Project length: medium

Mentors: Ray Simar (email).

Language/Tools: C, embedded platforms on which to evaluate the benchmarks.

Extend LibreCores.org

LibreCores is a community web site with the goal of providing an overview of IP cores and the corresponding ecosystem. We strive for LibreCores to be the resource for free and open source IP: it should be easy to find, integrate, and contribute to the projects found there -- to make digital design projects as easy as writing software. For further information on our goals, see the FOSDEM Presentation slides announcing LibreCores. The full site source code is available on GitHub.

You can find a non-exhaustive list of available tasks in our documentation. Please talk to Philipp if you have other ideas, or didn't find an interesting project. We welcome your own ideas!

Skill Level: Intermediate

Duration: 175 or 350 hours

Language/Tools: PHP with the Symfony Framework, MySQL, HTML/JS

Mentor: Philipp Wagner

Adding Language Servers (VHDL/Verilog/SV) to TerosHDL

Details: TerosHDL (https://terostechnology.github.io/) is an open source IDE for FPGA/ASIC development based in VSCode. It includes support for VHDL, Verilog and SystemVerilog. One of the most important features in a code editor is related to “go to definition”, linting… Currently TerosHDL supports those features partially. Tha main idea of this project is to integrate open source HDL language servers (Verible LSP, GHDL LSP, rusthdl…) in TerosHDL and provide all their capabilities. The language servers should work on Windows, Linux and Mac operating systems. Also create an interface for the plugin that allows users to enable/disable each feature independently.

Skill level: Beginner

Duration: 175h

Language/Tool: Javascript, NodeJS

Mentor: Carlos Alberto Ruiz, Ismael Pérez

Improve Test Coverage of the Moore HDL Compiler

We would love to have you help make Moore even better! Since the input languages are very complex, a key aspect of Moore is to perform well on existing test suites and benchmarks. The SymbiFlow project maintains a large suite of SystemVerilog tests where Moore is also represented. The goal of this project is to go into these tests and extend and improve the Moore compiler to support more of the use cases that are currently failing. You'll be able to look for juicy optimization targets and common reasons of failure, fix them, and reap the benefits of seeing a lot more green on this dashboard!

Skill Level: Advanced

Duration: 175 hours or 350 hours

Language/Tools: Rust, SystemVerilog

Mentor: Fabian Schuiki

Embench Interrupt Latency benchmark

Embench has always envisaged measuring the dynamic behavior of systems as well as computation throughput. This is the Embench-RT benchmark suite.

One of the benchmarks on Embench-RT is IRQ latency. The benchmark goal is to measure an MCU interrupt latency. This factor is crucial for RT FW/SW that is driven by interrupts.

Currently, this benchmark can run on RISCV cores; the project will be to integrate the benchmark for the reference Arm Cortex M4 board used for the main Embench IoT benchmark. This will act as the reference against which all others will be compared.

Skill level: Intermediate (some experience with embedded devices and board support packages).

Project length: long

Mentors: Ofer Shinaar (email) assisted by one of his colleagues (name TBC).

Language/Tools: C, embedded platforms on which to evaluate the benchmarks.

Python bindings for OpenROAD

Details: OpenROAD is a standalone application capable of performing RTL-to-GDSII SoC design, from floorplanning through detailed routing with metal fill insertion, signoff parasitic extraction and timing analysis. It currently provides a comprehensive high-level scriptable layer in TCL, but also embeds a Python interpreter. The embedded Python interpreter provides support only for the database (odb) and is a subset of what is available in TCL.

Possible contributions to this project:

Compile odb as a standalone python module
Extend the bindings to other components of OpenROAD
Write Python and/or Conda packaging recipes for the package and its dependencies
Bootstrap infrastructure to continuously build and publish the package to PyPI and/or Conda.

Skill level: Intermediate to Advanced

Duration: ~350 hrs.

Language/Tools: Python, C++, Conda

Mentor: Johan Euphrosine, Matt Liberty

Support for multiple RISC-V cores on a mini MCU

OpenHW Group is a not-for-profit organization that develops industrial-quality open-source HW. The CORE-V RISC-V cores family includes embedded and application class processors. The embedded cores are customized for performance CV32E40P, security CV32E40S, extendability, CV32E40X, and for code size CV32E41P. core-v-mini-mcu is a simple MCU meant to quickly try out new core HW and SW features. It targets simulation with Verilator, Questasim, etc, as well as real implementation on FPGAs and ASICs. FuseSoc is used to handle tools and targets.

In this project, we propose extending the core-v-mini-mcu to support all the OpenHW Group embedded class cores (as today we only use CV32E40P) and to demonstrate performance and code size differences across them.

Skill level: Medium (SystemVerilog)

Project length: medium

Mentors: Davide Schiavone (email) and Veronia Bahaa (email)

Language/Tools: SystemVerilog, RISC-V, C, Verilator

Development Framework for Open MPW Shuttles

Efabless, Google, and SkyWater Technologies, through the Open MPW and ChipIgnite programs, provide free ASIC fabrication. To help streamline the development flow, this project will provide a starting template for development within the free online Makerchip IDE. This will include:

visualization of the Caravel harness
support for FPGA prototyping, leveraging this environment
default stimulus
build flow enhancements to the Caravel sample user project, adding support for TL-Verilog and Makerchip
clear development flow documentation

Skill level: Intermediate/Advanced

Languages/Tools: Caravel (https://github.com/efabless/caravel), FPGAs, TL-Verilog, Makerchip

Project length: ~175 or ~350 hours

Mentors: Mohamed Kassem, Steve Hoover (email)

Improving state of code generation for RISC-V based manycore architectures in LLVM

The RISC-V ISA will transform the world. Recently, UW taped out an open source RISC-V manycore processor with thousands of cores which executes trillions of RISC-V instructions per second in one chip. We are working on finding answers to questions like how to efficiently program and run code on thousands of cores?

We want to improve the state of the code generation for manycore architectures in LLVM by developing techniques to better specialize different parts of the compiler for manycore architectures, for example, congestion awareness, instruction scheduling and register allocation. We plan to do this by analyzing different workloads or comparing with other compilers like gcc for finding generalized improvement opportunities.

Skill level: intermediate

Project length: long (350 hours)

Mentors: Reshabh K Sharma, Michael Taylor

Language/Tools: C++, LLVM

Enhancing the Sootty Terminal-based Graphical Waveform Viewer

Sootty is a terminal-based waveform viewer which takes advantage of modern terminals' capabilities to give a full graphical display of the users' waveforms. Combined with this, the tool introduces a textual query language for searching waveforms for interesting events to assist in debugging. Sootty is in a fairly early stage of development and so there are a variety of new features that could be added. Get in touch and we can figure out some fun and useful new features for our users!

Skill level: Beginner

Duration: 175 or 350 hours

Language/Tools: Python, Verilog/VHDL/...

Mentors: Ben Darnell, Jonathan Balkind

Javascript library for FuseSoC core files

Details: FuseSoC core description files are a language-agnostic description of an IP core. So far, the only implementations for working with core description files are written in Python (e.g. FuseSoC). A javascript library would make it easier to parse interesting information from the core description files and present on websites.

Skill level: beginner-intermediate

Project length: medium

Language/Tools: JavaScript. Python knowledge is useful for looking at the current implementation

Mentors: Olof Kindgren

Conda packages for Open Silicon tooling

Details: Bootstrapping a working environment for Open Silicon development is often a tedious task that involves installing multiple tools written in many different languages from various distribution channels (PyPI, debian, homebrew, containers image) and more often than not building them directly from sources!

The https://github.com/hdl/conda-eda/ project aims to fix this by providing conda.io packages for popular FPGA and ASIC tools and distributing them on a dedicated channel: https://anaconda.org/LiteX-Hub/.

Possible contributions to this project:

Make an inventory of the tools commonly used in popular Open Silicon flows.
Identify the tools (and their dependencies) that are not yet packaged.
Add new recipes to the project to cover those gaps.
Moving the package to conda-forge.org to provide automated build and distribution for Linux, MacOSX and Windows.
Creating meta-packages that capture the environment for popular Open Silicon flows (analog, digital, hls-based).

Skill level: Beginner

Duration: ~175 hrs.

Language/Tools: Conda, Shell scripting, Git

Mentor: Johan Euphrosine

Integration of WARP-V with OpenPiton

Details: In GSoC 2020, one of Shivam Potdar's accomplishments was to prepare the WARP-V CPU core for integration with the OpenPiton heterogeneous many-core framework. This project would aim to complete that integration.

This will enable academic exploration of the combined flexibility benefits of OpenPiton and WARP-V. OpenPiton provides flexibility through the integration of different CPU cores. WARP-V provides flexibility of the CPU core itself. WARP-V can be optimized for the cycle-time of OpenPiton, and differently-configured WARP-V cores could be run within the same system along with other CPU cores.

Duration: ~350 hrs.

Skill level: Advanced

Languages/Tools: Verilog, TL-Verilog, Makerchip, GNU Make

Co-mentors: Steve Hoover (email), Jonathan Balkind (email)

Discussion: WARP-V LibreCores Gitter Room, TL-Verilog Users Slack Workspace, OpenPiton LibreCores Gitter Room

Porting Compiler test-suite (micro-benchmarks) to Embench

Embench contains several benchmarks for IoT, benchmarks targeting performance, and code-size comparison. One missing piece is a compiler test-suite.

The project goal is to port the riscv32-Code-density-test-bench to Embench IoT new repo. The effort will include porting the Embench build system, bringing up the repo for users (instructions, build, and full run), adding "correctness testing" to the test suite.

Skill level: Beginner (some experience with C Python and Git).

Project length: medium

Mentors: Ofer Shinaar (email) assisted by one of his colleagues (name TBC).

Language/Tools: C, embedded platforms on which to evaluate the benchmarks.

Create your own LibreCores, or contribute to an existing one

Details: Our main goal is to grow the community around open source silicon designs. LibreCores are IP cores, but they are free and open. While there are many projects you can contribute too, you may have your own great idea for a LibreCore. All projects start small, and we see this is a great chance to bring forward new ideas and start building new tiny bits and pieces that enable free and open source chips.

We are happy to mentor you with your own idea, but it is important that it is re-usable and contains everything needed for simple and flexible integration, like testbenches, the required software drivers etc. So, it is important that you discuss a proposal intensively.

Skill level: All

Duration: ~175 or ~350 hrs.

Language/Tool: Verilog, VHDL, Chisel, TL-Verilog, ...

Mentor: We will find the mentor with you, LibreCores GSoC team

Klayout ASAP7 DRC and LVS decks

ASAP7 PDK is a 7-nm predictive process design kit (PDK) c, developed by Arizona State University in collaboration with ARM Ltd. for academic use. The PDK is realistic, based on current assumptions for the 7-nm technology node. Currently, there is an effort to port ASAP7 to OpenLane. To complete the port, ASA7 technology files must be ported to physical verification tools used by OpenLane.

The objective of this project is to develop ASAP7 Design Rules Checker (DRC) and Layout vs. Schematic (LVS) technology files for Klayout. Also, the scope of the project includes the development of a CI regression test suite.

Skill level: intermediate

Project length: medium

Language/Tools: Python, Ruby and Klayout

Other requirements: Good EE background

Mentors: Mohamed Shalan, Mohamed Kassem

Fault Silicon Validation and Open MPW integration

Design for Testability (DFT) is a technique which facilitates a design to become testable after fabrication. “Extra” logic is added to the design during the implementation process, which helps post-fabrication testing. Post-fabrication testing is necessary because IC fabrication is not 100% error free; thus, identifying faulty silicon as early as possible and before reaching the end-user is essential to reduce its impact.

The objective of the project is to Fully integrate Fault into OpenLane to enable DFT. Also, the project aims at adding DFT support to Caravel chip to enable post fabrication testing using Automatic Testing Equipment (ATE) before packaging to exclude faulty chips.

Skill level: intermediate

Project length: medium

Language/Tools: Python, Verilog and OpenLane

Other requirements: Good knowledge of RTL design as well as DFT.

Mentors: Mohamed Shalan, Kareem Farid

IRSIM dynamic power analysis and other improvements:

Summary: Dynamic power analysis is a requirement for digital designs that need to be power efficient. Static timing analysis tools do not understand actual circuit operation, so they can only estimate dynamic power usage, and will not be able to analyze specific power mitigation techniques, like having unused portions of a circuit shut down into a low-power or sleep mode. Proper dynamic power analysis requires simulating a circuit while computing switching power, with simulation setups designed to put the circuit into specific operating modes. SPICE analysis can do this but is prohibitively compute-expensive for large digital designs. Verilog simulators do not have the information to compute power consumption. Currently there are no open-source tools for dynamic power analysis.

IRSIM is an open source switch-level simulator that analyzes digital circuits on the transistor level, but with a simplified linear model of transistors as resistive switches. Because digital circuits operate by treating FET transistors as switches in one of two states, "on" (saturated) or "off", the simplified linear model works quite well, and changes in the state of a switch can be applied as "events" in a queue, similar to the way verilog simulators operate. Consequently, IRSIM runs only about an order of magnitude slower than a verilog gate-level simulator, but many orders of magnitude faster than a SPICE simulator. The slow speed of simulation relative to verilog simulators is one reason that IRSIM has largely fallen out of use.

However, IRSIM has had built-in power analysis for decades. But it has not been exercised for a long time, with considerable development in IRSIM happening without checking against the power analysis code.

The purpose of the project is to test the operation of power analysis in IRSIM, and to upgrade it where needed to make use of features in IRSIM added in recent years. That includes defining a Tcl command-line interface for power analysis.

Time permitting (e.g., 330 hour internship instead of 165 hour), an additional needed feature for IRSIM is to be able to handle multiple transistor types and, by extension, multiple power domains. The latter is particularly important for power analysis because many circuits operate in "sleep" modes by lowering the power supply to a minimum that can trigger a "wake-up" mechanism. Proper power analysis depends on understanding the different voltage domains. The implementation should be straightforward, as it adds nothing new to the analyses done by IRSIM, but is mostly just record-keeping.

Skill level: intermediate

Project length: 175 or 350 hrs

Language/Tools: C, Tcl/Tk, iverilog, ngspice

Mentors: Tim Edwards

Parameterized device layouts for Sky130

Summary: The purpose of this internship is to expand the set of available parameterized devices available in magic for use with the SkyWater Sky130 foundry process. These devices are created using an existing framework written in Tcl/Tk, although most of the code involves magic command-line commands (which are implemented in Tcl). Only a basic understanding of Tcl variables, conditionals, loops, and subroutine calls is needed.

New devices of interest that have not yet been done in parameterized cells include (but are not limited to): The photodiode, extended-drain MOSFETs, bipolar transistors, ESD transistors, inductors, metal fuses, UHV (ultra-high-voltage) devices.

Each new device should follow the design specifications for device layout from the SkyWater DRC manual, and needs to implement five routines that (1) define the device parameters and limits, (2) convert parameters from a SPICE netlist, (3) define the user interface dialog (UI) for setting those parameters, (4) check and enforce parameter limits, and (5) draw the device. After implementing, each device needs to be checked for DRC correctness by generating a "torture test" of a large array of devices with different sets of parameters which can be passed to the DRC checkers to make sure that the drawing routine produces DRC clean layouts.

Time permitting (e.g., 330 hour internship instead of 165 hour), the internship can include generating an analog layout implementing many different devices that can be tested from a probe station. This layout would be placed on an Efabless ChipIgnite run and tested at NIST.

Skill level: intermediate

Project length: 175 or 350 hrs

Language/Tools: Tcl/Tk, Magic

Mentors: Tim Edwards

RISC-V TensorCore for Edge AI

The goals of the RISC-V TensorCore project are to create RISC-V V-extension-based hardware accelerators that leverage custom numerics to gain energy-efficiency, performance, reproducibility, or reliable computations for robotics, model-predictive control, AI/ML, Reinforcement Learning, data acquisition and signal processing applications.

As many embedded intelligence applications will reside on the edge, and hardware experimentation requires a cost-effective and flexible design environment, most of the vector engines presented here are targeted as softcores on different FPGA platforms.

As the reconfigurability of an FPGA adds overhead to the realization of the logic for a computational transformation, it is paramount for energy-efficiency that the computational engine takes full advantage of the hardmacros available in the FPGA. The micro-architecture of a vector engine maps well to the DSP-slice architecture of most FPGAs, they are a productive technology target to deliver on custom compute engines with custom numerics.

Furthermore, when introducing custom numerics, the language support for these new types will always lag by many years, and sometimes the language standard committee will never be motivated to adapt. This means that only languages that offer user-defined types, such as Julia and C++, are ready to take advantage of custom hardware accelerators that differentiate through custom numerics. The user-defined type can be emulated by the scalar core, and computational kernels can be executed faithfully in the custom arithmetic by the vector engine. A user-defined type library, such as Universal will function as the custom type emulation layer, and higher level libraries, such as hprBLAS will function as the kernel dispatch layer.

Skill level: intermediate

Project length: 175 or 350 hrs

Language/Tools: TL-V, SV, Verilog

Mentors: Theodore Omtzigt

RISC-V core benchmarks using SiliconCompiler

SiliconCompiler is an open-source build framework for chip design that automates translation from source code to silicon. For this project, we would like to use SiliconCompiler to compile a handful of open-source RISC-V cores to open technology targets like FreePDK45 or Skywater130, and document optimal performance and area for each.

In order to demonstrate SiliconCompiler's flexibility, we would like to pick an interesting range of cores implemented in different HDLs. Some cores we are considering include:

PicoRV32 (Verilog)
Ibex (SystemVerilog)
VexRiscv (SpinalHDL)
RISC-V Mini (Chisel)
Flute (Bluespec)

However, we are open to suggestions from our mentee!

The expected outcome of this project is to pick 3-6 cores and write a SiliconCompiler build script for each one (with tuned build parameters), and then publish the build scripts along with layout screenshots and benchmarking results. In addition, this project may entail contributing back any changes to SiliconCompiler needed to complete the task. In particular, this project will likely involve work on tool drivers for any non-Verilog frontends used.

Skill Level: Intermediate

Duration: 175 hours (3 cores) or 350 hours (6 cores)

Language/tools: Python, Yosys, OpenROAD, Verilog, other HDLs depending on cores chosen

Mentors: Noah Moroze, Will Ransohoff

Interactive SiliconCompiler design summaries

SiliconCompiler produces automated reports for its designs, and one format that we are testing is a browser-based visualization of a job's results. This lets users view a small preview of the GDS result, alongside various metrics reported by each task in the physical design flow.

We would like to improve the way that this data is presented, and add more information from the SiliconCompiler job manifests. This would be a great project for someone interested in frontend development and/or data visualization!

This is what our reports currently look like. We have a few improvements in mind, but we would also welcome suggestions:

Past Results: SiliconCompiler retains metrics from past builds, which we would like to display.
Data Visualization: We would like to explore options beyond simple tables for visualizing metrics within and across job runs.
Input Parameters: In addition to a job's results, we would like to include the values of configurable inputs in our reports.
Modal Dialogs: There is some content which we would like to display in 'hover-over' modals, rather than linking to it.
"Flowgraph" Visualization: SiliconCompiler can generate diagrams describing which tasks ran over the course of a job, and we would like to include those in our reports.

Skill Level: Beginner

Duration: 175 hours

Languate/tools: Python (+Jinja), JS, HTML/CSS

Mentors: Will Ransohoff