1. Install OmpSs-2@FPGA toolchain

This page should help you install the OmpSs-2@FPGA toolchain. However, it is preferable using the pre-build Docker image with the latest stable toolchain. They are available at DockerHUB (https://hub.docker.com/r/bscpm/ompss_2_at_fpga). Moreover, we distribute pre-built SD images for some SoC. Do not hesitate to contact us at <ompss-fpga-support @ bsc.es> if you need help.

First, it describes the prerequisites to do the toolcahin installation. After that, the following sections explain different approaches to do the installation.

1.1. Prerequisites

1.1.1. Git Large File Storage

AIT repository uses Git Large File Storage to handle relatively-large files that are frequently updated (i.e. hardware runtime IP files) to avoid increasing the history size unnecessarily. You must install it so Git is able to download these files.

Follow instructions on their website to install it.

1.1.2. Vendor backends - Xilinx Vivado

Follow the installation instructions from Xilinx Vitis HLS and Vivado. You will need to enable support for the devices you’re working on, as well as install the board files for the given devices.

1.2. Stable release

There is a meta-repository that points to latest stable version of all tools: https://github.com/bsc-pm-ompss-at-fpga/ompss-2-at-fpga-releases. It contains a Makefile which, based on some environment variables, will compile and install the toolchain. The environment variables are:

  • TARGET [Def: aarch64-linux-gnu] Linux architecture that toolchain will target
  • PREFIX_HOST [Def: /] Installation prefix for the host tools (e.g. llvm, ait)
  • PREFIX_TARGET [Def: /] Installation prefix for the target tools (e.g. nanos6, libxdma, libxtasks)
  • BUILDCPUS [Def: nproc] Number of processes used for building

The following example will cross-build the toolchain for the aarch64-linux-gnu architecture and install it in /opt/bsc/host-arm64/ompss-2 and /opt/bsc/arm64/ompss-2:

git clone --recursive https://github.com/bsc-pm-ompss-at-fpga/ompss-2-at-fpga-releases.git
cd ompss-2-at-fpga-releases
export TARGET=aarch64-linux-gnu
export PREFIX_HOST=/opt/bsc/host-arm64/ompss-2
export PREFIX_TARGET=/opt/bsc/arm64/ompss-2
make

1.3. Individual git repositories

The master branches of all tools should generate a compatible toolchain. Each package should contain information about how to compile/install itself, look for the README files. The following points briefly describe each tool and provide a possible build configuration/setup for each one. We assume that all packages will be installed in a Linux OS in the /opt/bsc/arm64/ompss-2 folder. Moreover, we assume that the packages will be cross-compiled from an Intel machine to be run on an ARM64 embedded board.

List of tools to install:

1.3.1. Accelerator Integration Tool (AIT)

You can install the AIT package through the pip repository python3 -m pip install ait-bsc or cloning the git repository:

git clone https://github.com/bsc-pm-ompss-at-fpga/ait
cd ait
git lfs install
git lfs pull
export AIT_HOME="/path/to/install/ait"
export DEB_PYTHON_INSTALL_LAYOUT=deb_system
python3 -m pip install . -t $AIT_HOME

export PATH=PREFIX/ait/:$PATH
export PYTHONPATH=$AIT_HOME:$PYTHONPATH

1.3.2. Kernel module

The driver is only needed to execute the applications. To compile them, the library must be installed on the host but the kernel module may not be loaded. Example to cross-compile the driver:

git clone https://github.com/bsc-pm-ompss-at-fpga/ompss-at-fpga-kernel-module.git
cd ompss-at-fpga-kernel-module
export CROSS_COMPILE=aarch64-linux-gnu-
export KDIR=/home/my_user/kernel-headers
export ARCH=arm64
make

1.3.3. XDMA

Example to cross-compile the library and install it in the /opt/bsc/arm64/ompss-2/libxdma folder:

git clone https://github.com/bsc-pm-ompss-at-fpga/xdma.git
cd xdma/src/zynq
export CROSS_COMPILE=aarch64-linux-gnu-
export KERNEL_MODULE_DIR=/path/to/ompss-at-fpga/kernel/module/src
make
make PREFIX=/opt/bsc/arm64/ompss-2/libxdma install

1.3.4. xTasks

Example to cross-compile the library and install it in the /opt/bsc/arm64/ompss-2/libxtasks folder:

git clone https://github.com/bsc-pm-ompss-at-fpga/xtasks.git
cd xtasks/src/zynq
export CROSS_COMPILE=aarch64-linux-gnu-
export LIBXDMA_DIR=/opt/bsc/arm64/ompss-2/libxdma
make
make PREFIX=/opt/bsc/arm64/ompss-2/libxtasks install

1.3.5. Nanos6-fpga

Example to cross-compile the runtime library and install it in the /opt/bsc/arm64/ompss-2/nanos6-fpga folder:

git clone https://github.com/bsc-pm-ompss-at-fpga/nanos6-fpga.git
cd nanos6-fpga
./autogen.sh
mkdir build-fpga-arm64
cd build-fpga-arm64
../configure --prefix=/opt/bsc/arm64/ompss-2/nanos6-fpga --host=aarch64-linux-gnu --enable-fpga --with-xtasks=/opt/bsc/arm64/ompss-2/libxtasks --disable-discrete-deps --disable-all-instrumentations --enable-stats-instrumentation --enable-verbose-instrumentation
make
make install

1.3.6. LLVM/Clang

Example to build a LLVM/Clang cross-compiler that runs on the host and creates binaries for another platform (ARM64 in the example):

git clone https://github.com/bsc-pm-ompss-at-fpga/llvm.git
cd llvm
mkdir build-fpga
cd build-fpga
cmake -G Ninja -DCMAKE_INSTALL_PREFIX=/opt/bsc/host-arm64/ompss-2/llvm -DLLVM_TARGETS_TO_BUILD="AArch64" -DCMAKE_BUILD_TYPE=Release -DCLANG_DEFAULT_NANOS6_HOME=/opt/bsc/arm64/ompss-2/nanos6-fpga -DLLVM_USE_SPLIT_DWARF=ON -DLLVM_ENABLE_PROJECTS="clang" -DLLVM_INSTALL_TOOLCHAIN_ONLY=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DLLVM_USE_LINKER=lld ../llvm/llvm
nina
ninja install