yvantor / hero

Heterogeneous Research Platform (HERO) for exploration of heterogeneous computers consisting of programmable many-core accelerators and an application-class host CPU, including full-stack software and hardware.

Heterogeneous Research Platform (HERO)

HERO is an FPGA-based research platform that enables accurate and fast exploration of heterogeneous computers consisting of programmable many-core accelerators and an application-class host CPU. Currently, 32-bit RISC-V cores are supported in the accelerator and 64-bit ARMv8 or RISC-V cores as host CPU. HERO allows to seamlessly share data between host and accelerator through a unified heterogeneous programming interface based on OpenMP 4.5 and a mixed-data-model, mixed-ISA heterogeneous compiler based on LLVM.

HERO's hardware architecture, shown below, combines a general-purpose host CPU (in the upper left corner) with a domain-specific programmable many-core accelerator (on the right side) so that data in the main memory (in the lower left corner) can be shared effectively.

HERO's runtime software stack, shown below, seamlessly integrates accelerators (with their runtime stack on the right) into the OS running on the host (runtime stack on the left) to enable heterogeneous applications with transparent offloading of accelerated kernels (at the top).

All of HERO's software and most of its hardware (everything except the components of the Xilinx Zynq MPSoC) are available under open-source licenses. This includes:

• cross compiler toolchains
• root file system configurations
• boot images
• Linux kernel configurations and drivers
• runtime libraries
• benchmarks
• OpenMP applications
• RISC-V cores
• on-chip memory controllers
• caches
• interconnects
• DMA engines
• IOMMU

This repository contains the source code for the top-level logic of HERO and many of its components. The source code of other components, such as the Linux kernel and the cross compilers, is not directly stored in this repository. Instead, that code is stored in other open-source repositories to which this repository links.

The most mature implementation target currently is Xilinx' ZCU102 featuring a Xilinx Zynq UltraScale+ MPSoC. The 64-bit ARMv8 Cortex-A53 of the processing system (PS) serves as host CPU, and a multi-core 32-bit RISC-V accelerator is instantiated in the programmable logic (PL).

A second implementation target currently under development is Xilinx' VCU128, where a 64-bit RISC-V CVA6 host CPU and a multi-core 32-bit RISC-V accelerator are combined on an FPGA.

This repository is published on GitHub, including all releases and the main branch. As continuous integration and verification (partially) relies on commercial software and FPGA boards, it is hosted on servers of the maintainers. For the same reason, issues are tracked there. If you have questions or would like to provide feedback, please e-mail hero (at) iis ee ethz ch (replace the spaces in the domain name with dots).

The design and architecture of the components of HEROv2 are described in this paper (preprint). If you use our work in your research, please cite it.

Getting Started

Prerequisites

All build steps should work on any recent Linux distribution (we have automated tests on CentOS 7). Most of the core dependencies are installed automatically as part of the toolchain. Required is at least GCC 4.8 / Clang 3.3 to bootstrap the toolchain installation. More details can be found here.

External sources

HERO uses Git submodules that have to be initialized. Either clone the repository recursively using

git clone --recursive <url>

or fetch the submodules afterwards in the repository

git submodule update --init --recursive

Setup

All toolchains and SDKs are installed automatically to the location pointed to by the HERO_INSTALL environment variable. Please set it to your preferred installation location before continuing with the next step

export HERO_INSTALL=<your_path>

We recommend you create an install subdirectory in this repository and set HERO_INSTALL to that.

If you want to build the HERO SDK and the board images with PREM support, please run

make prem-set

before continuing. This writes the necessary commands to your local configuration file. The command must be run before building the sdk-h*, br-h* targets. To disable PREM, run

make prem-unset

and then rebuild your sdk-h* and br-h* targets.

PREM support is disabled by default, and no action is necessary to setup HERO without PREM support. PREM has only been tested for exilzcu102, and is known to not work correctly with the teams distribute OpenMP construct.

Toolchains

HERO has configurations for multiple base toolchains. The project relies on Crosstool-NG for easily building the different toolchains. The toolchains are installed to the HERO_INSTALL directory.

All heterogeneous installations require a toolchain for the PULP accelerator, which can be built with:

make tc-pulp

Furthermore, both Linux and bare-metal toolchains are available for both the RISC-V Ariane host and for the ARMv8 Cortex-A53 host. These can be separately built as follows:

make tc-hrv-obare   # 64-bit RISC-V bare-metal toolchain
make tc-hrv-olinux  # 64-bit RISC-V Linux toolchain
make tc-har-obare   # 64-bit AArch64 bare-metal toolchain
make tc-har-olinux  # 64-bit AArch64 Linux toolchain

Current setups require only the Linux toolchains (-olinux suffix).

PULP SDK

The PULP SDK is required to build applications for PULP. This has to be set up before installing the HERO host SDKs because heterogeneous libraries rely on it. The required configuration can be built with

make sdk-pulp

Host SDKs

Host Software Development Kits (SDKs) for both RISC-V and ARMv8 are necessary for the infrastructure of further build steps, for the development of heterogeneous applications, and for development on the HERO software stack itself. The SDKs are generated using Buildroot and installed to the HERO_INSTALL directory. These SDKs can be built as follows:

make sdk-hrv  # SDK for RISC-V host
make sdk-har  # SDK for ARMv8 host

Heterogeneous Toolchain

After the base toolchains and accelerator and host SDKs are installed, the HERO heterogeneous toolchain can be installed. The heterogeneous toolchain is based on LLVM 12 and includes support in Clang and OpenMP to create heterogeneous applications for the HERO target platform. The toolchain will be installed in the HERO_INSTALL toolchain folder together with the core toolchains and SDKs. To build and install the LLVM toolchain, run:

make tc-llvm

Note that the LLVM OpenMP target offloading plugin for PULP (i.e., libomptarget.rtl.pulp.so) is not compiled as part of LLVM. Instead, it is compiled as package in the host SDK, because it must be compiled for a specific host architecture. You can thus safely ignore the message LIBOMPTARGET: Not building PULP offloading plugin: build disabled..

Environments

With the toolchains and SDKs installed, the development environment for an implementation target can be created (see introduction for an overview of the implementation targets). This section provides the instructions for each supported implementation target.

Xilinx ZCU102

A complete Linux environment with kernel and the base root filesystem for the ARMv8 host on the Xilinx Zynq UltraScale+ MPSoC ZCU102 can be built as follows.

First, add a line to the local.cfg file at the root of your repository (create that file if it does not exist), which defines the path to the bitstream:

BR2_HERO_BITSTREAM=/absolute/path/to/hero_exilzcu102_wrapper.bit

Then, generate the SD card image files with

make br-har-exilzcu102

This creates an SD card image from the Buildroot filesystem and additionally invokes PetaLinux to create bootable image files. Due to security limitations in the latest XCZU9EG revisions, we currently have to use images generated by Xilinx PetaLinux to boot Linux. When the images are generated, insert the SD card of the ZCU102 into your development workstation and

1. Initialize the SD card with a master boot record (MBR, also known as "DOS partition table") and create two partitions:
   1. FAT32 (0xc), 256 MiB, bootable
   2. Linux (0x83), remaining space
2. Format the first partition with vfat and the second partition with ext4.
3. Mount the first partition, copy BOOT.BIN and image.ub from output/br-har-exilzcu102/images/ to the mounted partition, and unmount it.

The system can then be booted up after configuring the ZCU102 board jumpers to boot from the SD card. On the first boot, change the U-Boot variables as described in issue #107.

The root filesystem mounted after boot is provided by PetaLinux. It is loaded into memory during boot, and changes are not persisted to the SD card. Furthermore, that root filesystem does not include the HERO libraries. A persistent filesystem, which includes the HERO libraries, resides on the second partition of the SD card. We recommend mounting it at /mnt

mount /dev/mmcblk0p2 /mnt

and using its dynamically-linked shared libraries through

export LD_LIBRARY_PATH=/mnt/usr/lib:/mnt/lib

in every session on the board. You may also want to place your binaries and data under /mnt to not lose them on a reboot (or kernel panic).

Once the board boots from the SD card and you have SSH access to it, you can remotely update the image files and reboot the board:

scp output/br-har-exilzcu102/images/{BOOT.BIN,image.ub} root@zcu102-hostname:/run/media/mmcblk0p1/
ssh root@zcu102-hostname /sbin/reboot

To develop applications for this setup, the dynamic environment on the development machine can be loaded using source env/exilzcu102.sh. Afterwards applications can be built and transferred directly to the board.

QEMU RISC-V

For host debugging it can be useful to test the environment first with the QEMU machine emulator. A clone of the RISC-V Ariane environment without specific hardware patches and including virtual drivers can be built with:

make br-hrv-eqemu

A recent version of QEMU is required to be able to emulate RISC-V, it is currently recommended to build it from source. After installation the following configuration can be used to boot:

qemu-system-riscv64 \
   -machine virt -nographic \
   -bios output/hrv-eqemu-base-fw_jump.bin \
   -kernel output/hrv-eqemu-base-Image \
   -append "root=/dev/vda ro" \
   -drive file=output/hrv-eqemu-base-rootfs.ext2,format=raw,id=hd0 \
   -device virtio-blk-device,drive=hd0 \
   -netdev user,id=net0,hostfwd=tcp::5555-:22 \
   -device virtio-net-device,netdev=net0 \
   -device virtio-rng-device \
   -gdb tcp::3332

It it possible to SSH to the machine on port 5555 and debug with GDB on port 3332.

A root filesystem with the entire HERO infrastructure can be loaded in QEMU by adding

   -drive file=output/hrv-rootfs.ext4,format=raw,id=hd1 \
   -device virtio-blk-device,drive=hd1 \

to the command to run above. Afterwards chroot /mnt can be executed to change to the HERO environment.

RTL Simulation

An environment is provided to simulate the accelerator in RTL. If QuestaSim is installed, the simulation infrastructure can be initialized as follows:

cd hardware
make
cd vsim
./compile.sh

For building applications for the simulation, the dynamic environment can be loaded with source env/esim-exilzcu102.sh. When building OpenMP applications, it has to be specified to build them only for the PULP accelerator using make only=pulp, which will also generate SLM files to initialize memory in simulation. Afterwards, the binary can be loaded in the simulator with:

cd hardware/vsim
../test/gen_slm_files.sh <app_name>
./start_sim.sh

where <app_name> is the path to the directory from the openmp-examples directory (for example polybench-acc/2mm).

Utilities for the Development Machine

util contains utilities for the development machine to facilitate working with HERO. One of them, the OpenOCD debugger for RISC-V, can be installed to HERO_INSTALL with

make util-hrv-openocd