Skip to content

Latest commit

 

History

History
168 lines (128 loc) · 8.95 KB

Building.md

File metadata and controls

168 lines (128 loc) · 8.95 KB
Raw

Building the code

Prerequisites

lld
cmake 3.20.6
ninja 1.8.2
Xilinx Vitis 2023.2
python 3.8.x and pip
virtualenv
pip3 install psutil rich pybind11 numpy
clang/llvm 14+ from source https://github.com/llvm/llvm-project

Xilinx Vitis can be downloaded and installed from the Xilinx Downloads site.

In order to successfully install Vitis on a fresh bare-bones Ubuntu install, some additional prerequisites are required, documented here. For Ubuntu 20.04, the installation should succeed if you additionally install the following packages: libncurses5 libtinfo5 libncurses5-dev libncursesw5-dev ncurses-compat-libs libstdc++6:i386 libgtk2.0-0:i386 dpkg-dev:i386 python3-pip Further note that the above mentioned cmake prerequisite is not satisfied by the package provided by Ubuntu; you will need to obtain a more current version.

NOTE: Using the Vitis recommended settings64.sh script to set up your environement can cause tool conflicts. Setup your environment in the following order for aietools and Vitis:

export PATH=$PATH:<Vitis_install_path>/Vitis/2023.2/aietools/bin:<Vitis_install_path>/Vitis/2023.2/bin

The cmake and python packages prerequisites can be satisfied by sourcing the utils/setup_python_packages.sh script. See step 2 of the build instructions. This script requires virtualenv.

clang/llvm 14+ are recommended to be built with the provided scripts. See step 3. of the build instructions.

When targetting the VCK5000 Versal device, you must build and install our experimental ROCm runtime which allows us to communicate with the AIEs. The ROCm-air-platforms repository contains documentation on how to install our experimental ROCm runtime. When targetting the VCK5000, it will be necessary to install a global version of ROCM 5.6. Details of all these steps can be found in the ROCm-air-platforms repo.

Building on X86 for mlir-aie development

  1. Clone the mlir-aie repository with its sub-modules:

    git clone --recurse-submodules https://github.com/Xilinx/mlir-aie.git
cd mlir-aie

    All subsequent steps should be run from inside the top-level directory of the mlir-aie repository cloned above.

  2. Source utils/setup_python_packages.sh to setup the prerequisite python packages. This script creates and installs the python packages listed in utils/requirements.txt in a virtual python environment called 'sandbox', then it enters the sandbox:

    source utils/setup_python_packages.sh

    If you need to exit the sandbox later, type deactivate. If you have a recent Linux distribution, you might not need this, as you are able to have all the required packages from the distribution.

  3. Clone and compile LLVM, with the ability to target AArch64 as a cross-compiler, and with MLIR enabled: in addition, we make some common build optimizations to use a linker (lld or gold) other than ld (which tends to be quite slow on large link jobs) and to link against libLLVM.so and libClang.so. You may find that other options are also useful. Note that due to changing MLIR APIs, only a particular revision is expected to work.

    To clone llvm, run utils/clone-llvm.sh (see utils/clone-llvm.sh for the correct llvm commit hash):

    ./utils/clone-llvm.sh

    If you have already an LLVM repository, you can instead of cloning just make a new worktree from it by using:

    ./utils/clone-llvm.sh --llvm-worktree <directory-of-existing-LLVM-repository>

    To build (compile and install) LLVM, run utils/build-llvm-local.sh in the directory where llvm has been cloned. See utils/build-llvm-local.sh for additional shell script arguments. (Note that build-llvm-local.sh and build-llvm.sh are a variation of the LLVM build script used for CI on GitHub and looking at the continuous integration recipe https://github.com/Xilinx/mlir-aie/blob/main/.github/workflows/buildAndTest.yml and output https://github.com/Xilinx/mlir-aie/actions/ might help in the case of compilation problem.)

    ./utils/build-llvm-local.sh

    This will build LLVM in llvm/build and install the LLVM binaries under llvm/install.

  4. Build the MLIR-AIE tools by calling utils/build-mlir-aie.sh or utils/build-mlir-aie-pcie.sh for Ryzen AI or Versal respectively with the path to the llvm/build directory. The Vitis environment will have to be set up for this to succeed.

    Building mlir-aie for a Ryzen AI system:

    source <Vitis Install Path>/settings64.sh
./utils/build-mlir-aie.sh <llvm dir>/<build dir>

    Building mlir-aie for a VCK5000 system:

    source <Vitis Install Path>/settings64.sh
./utils/build-mlir-aie-pcie.sh <llvm dir>/<build dir>

    This will create a build and install folder in the directory that you cloned MLIR AIE into.

    The MLIR AIE tools will be able to generate binaries targetting a combination of AIEngine and ARM/x86 processors.

  5. In order to run all the tools, it is necessary to add some paths into your environment. This can be done by sourcing the utils/env_setup.sh script with the paths to the install folders for mlir-aie and llvm.

    source utils/env_setup.sh <mlir-aie>/install <llvm dir>/install

  6. Only for the VCK5000: The PCIe AIR runtime requires the use of the AIR PCIe kernel driver. This directory contains documentation on how to compile and load the AIR PCIe kernel driver.

Note that when coming back to this install with a fresh environment, it is necessary to rerun the utils/env_setup.sh script to setup your environment as well as activate the Python virtual environment using the following command.

source sandbox/bin/activate

Building on X86 targetting the VCK5000 with seperate builds of ROCm runtime and aie-rt

The build-mlir-aie-pcie.sh script will automatically build a local install of the ROCm Runtime and aie-rt if the corresponding install directories are not provided. Another option is to build these externally and then point build-mlir-aie-pcie.sh to where they are installed. Below are the steps for how to build the tools using this second option.

We chose to install the aie-rt library in /opt/xaiengine but it is not required for the tools to be installed there. Just ensure that when building mlir-aie and mlir-air, that you point to the directory in which the aie-rt library was installed. Below are the steps to build the aie-rt library and move the installation to /opt/xaiengine.

git clone https://github.com/stephenneuendorffer/aie-rt
cd aie-rt
git checkout phoenix_v2023.2
cd driver/src
make -f Makefile.Linux CFLAGS="-D__AIEAMDAIR__"
sudo cp -r ../include /opt/xaiengine/
sudo cp libxaiengine.so* /opt/xaiengine/lib/
export LD_LIBRARY_PATH=/opt/xaiengine/lib:${LD_LIBRARY_PATH}

To install our experimental ROCm runtime which allows us to communicate with the AIEs on the VCK5000, it is necessary to install a global version of ROCM 5.6. Details of all these steps can be found in the ROCm-air-platforms. Afterwards, follow the steps in ROCm-air-platforms to build the experimental ROCm runtime that targets the AIEs in the VCK5000. You can run the following script to clone the ROCm-air-platforms repository:

./utils/clone-rocm-air-platforms.sh

Then, set ${ROCM_ROOT} to the ROCm install. Then, run the following command to build the mlir-aie toolchain targetting the VCK5000 pointing to the externally installed aie-rt and experimental ROCm runtime.

./utils/build-mlir-aie-pcie.sh llvm/build/ build install /opt/xaiengine ${ROCM_ROOT}/lib/cmake/hsa-runtime64/ ${ROCM_ROOT}/lib/cmake/hsakmt/

Sysroot

Since the AIE tools are cross-compiling, in order to actually compile code, we need a 'sysroot' directory, containing an ARM rootfs. This rootfs must match what will be available in the runtime environment. Note that copying the rootfs is often insufficient, since many root file systems include absolute links. Absolute symbolic links can be converted to relative symbolic links using symlinks.

cd /
sudo symlinks -rc .

Following the platform build steps will create such a sysroot based on PetaLinux. Note that those instructions require Vitis 2021.2 -- building a sysroot with Vitis 2023.2 will not currently succeed.

Copyright© 2019-2023 AMD/Xilinx