Citing

If you are using PULP in your academic work you can cite us:

@ARTICLE{8715500,
  author={Pullini, Antonio and Rossi, Davide and Loi, Igor and Tagliavini, Giuseppe and Benini, Luca},
  journal={IEEE Journal of Solid-State Circuits}, 
  title={Mr.Wolf: An Energy-Precision Scalable Parallel Ultra Low Power SoC for IoT Edge Processing}, 
  year={2019},
  volume={54},
  number={7},
  pages={1970-1981},
  doi={10.1109/JSSC.2019.2912307}}

PULP

PULP (Parallel Ultra-Low-Power) is an open-source multi-core computing platform part of the of the ongoing collaboration between ETH Zurich and the University of Bologna - started in 2013.

The PULP architecture targets IoT end-node applications requiring flexible processing of data streams generated by multiple sensors, such as accelerometers, low-resolution cameras, microphone arrays, vital signs monitors.

PULP consists of an advanced microcontroller architecture representing a significant step ahead in terms of completeness and complexity with respect to PULPino, taking care of autonomous I/O, advanced data pre-processing, external interrupts, and including a tightly-coupled cluster of processors to which compute-intensive kernels can be offloaded from a main processor. The PULP architecture includes:

• Either the RI5CY core or the zero-riscy one as main core
• Autonomous Input/Output subsystem (uDMA)
• New memory subsystem
• Support for Hardware Processing Engines (HWPEs)
• New simple interrupt controller
• New peripherals
• New parallel computing cluster
• New system DMA
• New event unit
• New SDK

RISCY is an in-order, single-issue core with 4 pipeline stages and it has an IPC close to 1, full support for the base integer instruction set (RV32I), compressed instructions (RV32C) and multiplication instruction set extension (RV32M). It can be configured to have single-precision floating-point instruction set extension (RV32F). It implements several ISA extensions such as: hardware loops, post-incrementing load and store instructions, bit-manipulation instructions, MAC operations, support fixed-point operations, packed-SIMD instructions and the dot product. It has been designed to increase the energy efficiency of in ultra-low-power signal processing applications. RISCY implementes a subset of the 1.9 privileged specification. Further information about the core can be found at http://ieeexplore.ieee.org/abstract/document/7864441/ and in the documentation of the IP.

zero-riscy is an in-order, single-issue core with 2 pipeline stages and it has full support for the base integer instruction set (RV32I) and compressed instructions (RV32C). It can be configured to have multiplication instruction set extension (RV32M) and the reduced number of registers extension (RV32E). It has been designed to target ultra-low-power and ultra-low-area constraints. zero-riscy implementes a subset of the 1.9 privileged specification. Further information about the core can be found at http://ieeexplore.ieee.org/document/8106976/ and in the documentation of the IP.

PULP includes a new efficient I/O subsystem via a uDMA (micro-DMA) which communicates with the peripherals autonomously. The core just needs to program the uDMA and wait for it to handle the transfer. Further information about the core can be found at http://ieeexplore.ieee.org/document/8106971/ and in the documentation of the IP.

PULP supports I/O on interfaces such as:

• SPI (as master)
• I2S
• Camera Interface (CPI)
• I2C
• UART
• JTAG

PULP also supports integration of hardware accelerators (Hardware Processing Engines) that share memory with the RI5CY core and are programmed on the memory map. An example accelerator, performing multiply-accumulate on a vector of fixed-point values, can be found in hwpe-mac-engine (after updating the IPs: see below in the Getting Started section). The hwpe-stream and hwpe-ctrl folders contain the IPs necessary to plug streaming accelerators into a PULP system on the data and control plane. For further information on how to design and integrate such accelerators, see hwpe-stream/doc and https://arxiv.org/abs/1612.05974.

Getting Started

Prerequisites

To be able to use the PULP platform, you need the PULP toolchain. The instructions to get it can be found here: https://github.com/pulp-platform/pulp-riscv-gnu-toolchain.

Building the RTL simulation platform

To build the RTL simulation platform, start by getting the latest version of the IPs composing the PULP system:

source setup/vsim.sh

make checkout

make scripts

make build

NOTE: An error might occur running the scripts (Failed to spawn child process.Too many open files (os error 24).) while a fix is WIP a workaround is to increase the number of processes avilable to your machine by setting for example ulimit to 4096 (ulimit -n 4096).

This command builds a version of the simulation platform with no dependencies on external models for peripherals. See below (Proprietary verification IPs) for details on how to plug in some models of real SPI, I2C, I2S peripherals.

Default dependency management is done using bender to gather IPs. If you would like to use the legacy IPApproX tool, set the IPAPPROX environment variable, e.g. by running export IPAPPROX=1, and continue at your own risk.

Working on IPs

The easiest way to work on an individual IP is to clone it using bender with the following command:

./bender clone $IP

This will checkout the IP to the working_dir directory, where it can be modified and the changes committed and pushed. The correct link will be set through an override in the Bender.local file, forcing the bender tool to use this version of the dependency. To build the platform, make sure to start at the make scripts step above after calling ./bender clone.

Once the changes are complete, please ensure the Bender.yml files in the packages calling the IP dependency are accordingly updated with the new version. The bender parents command can assist in determining which dependencies' Bender.yml files need updating. Please note that when modifying dependency versions, the ./bender update command needs to be called to re-resolve the correct versions. Once the update is complete, the corresponding line from Bender.local can be removed to revert to normal dependency resolution, no longer using the version in working_dir (be sure to call ./bender update). For more information check out the bender documentation.

Downloading and running simple C regression tests

Finally, you can download and run the tests; for that you can checkout the following repositories:

• Runtime tests: https://github.com/pulp-platform/regression_tests

• Pulp runtime: https://github.com/pulp-platform/pulp-runtime

Now you can change directory to your favourite test e.g.: for an hello world test, run

git clone https://github.com/pulp-platform/regression_tests.git

git clone https://github.com/pulp-platform/pulp-runtime.git

source pulp-runtime/configs/pulp.sh

export PATH=*path to riscv gcc toolchain*/bin:$PATH

export PULP_RISCV_GCC_TOOLCHAIN= *path to riscv gcc toolchain*

cd regression_tests/hello

mae clean all run gui=1

The open-source simulation platform relies on JTAG to emulate preloading of the PULP L2 memory. If you want to simulate a more realistic scenario (e.g. accessing an external SPI Flash), look at the sections below.

In case you want to see the Modelsim GUI, just type

make conf gui=1

before starting the simulation.

If you want to save a (compressed) VCD for further examination, type

make conf vsim/script=export_run.tcl

before starting the simulation. You will find the VCD in build/<SRC_FILE_NAME>/pulp/export.vcd.gz where <SRC_FILE_NAME> is the name of the C source of the test.

Proprietary verification IPs

The full simulation platform can take advantage of a few models of commercial SPI, I2C, I2S peripherals to attach to the open-source PULP simulation platform. In rtl/vip/spi_flash, rtl/vip/i2c_eeprom, rtl/vip/i2s you find the instructions to install SPI, I2C and I2S models.

When the SPI flash model is installed, it will be possible to switch to a more realistic boot simulation, where the internal ROM of PULP is used to perform an initial boot and to start to autonomously fetch the program from the SPI flash. To do this, the LOAD_L2 parameter of the testbench has to be switched from JTAG to STANDALONE.

PULP-SDK

If you are a software developer, you can find the PULP-SDK here: https://github.com/pulp-platform/pulp-sdk.

PULP platform structure

After being fully setup as explained in the Getting Started section, this root repository is structured as follows:

• rtl/tb contains the main platform testbench and the related files.
• rtl/vip contains the verification IPs used to emulate external peripherals, e.g. SPI flash and camera.
• rtl could also contain other material (e.g. global includes, top-level files)
• sim contains the ModelSim/QuestaSim simulation platform.
• pulp-sdk contains the PULP software development kit; pulp-sdk/tests contains all tests released with the SDK.
• Bender.yml contains all dependency and source file information for the bender tool.

Requirements

The RTL platform has the following requirements:

• Relatively recent Linux-based operating system; we tested Ubuntu 16.04 and CentOS 7.
• ModelSim in reasonably recent version (we tested it with version 10.6b).
• Python 3.4, with the pyyaml module installed (you can get that with pip3 install pyyaml).
• The SDK has its own dependencies, listed in https://github.com/pulp-platform/pulp-sdk/blob/master/README.md

Repository organization

The PULP platforms is highly hierarchical and the Git repositories for the various IPs follow the hierarchy structure to keep maximum flexibility. Most of the complexity of the IP updating system are hidden behind the bender tool; however, a few details are important to know:

• Do not assume that the master branch of an arbitrary IP is stable; many internal IPs could include unstable changes at a certain point of their history. Conversely, in top-level platforms (pulpissimo, pulp) we always use stable versions of the IPs. Therefore, you should be able to use the master branch of pulpissimo safely.
• By default, the IPs will be collected from GitHub using HTTPS. This makes it possible for everyone to clone them without first uploading an SSH key to GitHub. However, for development it is often easier to use SSH instead, particularly if you want to push changes back.

The tools used to collect IPs and create scripts for simulation have many features that are not necessarily intended for the end user, but can be useful for developers; if you want more information, e.g. to integrate your own repository into the flow, you can find documentation at https://github.com/pulp-platform/bender/blob/master/README.md

External contributions

The supported way to provide external contributions is by forking one of our repositories, applying your patch and submitting a pull request where you describe your changes in detail, along with motivations. The pull request will be evaluated and checked with our regression test suite for possible integration. If you want to replace our version of an IP with your GitHub fork, just add it to the corresponding Bender.yml file, or use an override in a Bender.local in the top repository. While we are quite relaxed in terms of coding style, please try to follow these recommendations: https://github.com/pulp-platform/ariane/blob/master/CONTRIBUTING.md

Known issues

The current version of the PULP platform does not include yet an FPGA port or example scripts for ASIC synthesis; both things may be deployed in the future. Simulation flows different from ModelSim/QuestaSim have only have limited testing.

Support & Questions

For support on any issue related to this platform or any of the IPs, please add an issue to our tracker on https://github.com/pulp-platform/pulpissimo/issues