install.rst

Installation

Install Requirements

In order to build pocl, you need the following support libraries and tools:

• A supported version of LLVM & Clang (check release notes)
• development files for LLVM & Clang + their transitive dependencies (e.g. libclang-dev, libllvm-dev, zlib1g-dev, libtinfo-dev...)
• CMake
• GNU make or ninja
• pkg-config
• pthread (should be installed by default)
• hwloc v1.0 or newer (e.g. libhwloc-dev) - optional
• python3 (for support of LLVM bitcode with SPIR target; optional but enabled by default)
• llvm-spirv (version-compatible with LLVM) and spirv-tools (optional; required for SPIR-V support in CPU / CUDA; Vulkan driver supports SPIR-V through clspv)

Install requirements on Linux

Note: For Ubuntu/Debian, PoCL supports the binary packages from https://apt.llvm.org/ - these usually support more (and newer) LLVM versions than Ubuntu/Debian.

Install requirements (e.g with LLVM 19) for Ubuntu/Debian:

export LLVM_VERSION=19
apt install -y python3-dev libpython3-dev build-essential ocl-icd-libopencl1 \
cmake git pkg-config libclang-${LLVM_VERSION}-dev clang-${LLVM_VERSION} \
llvm-${LLVM_VERSION} make ninja-build ocl-icd-libopencl1 ocl-icd-dev \
ocl-icd-opencl-dev libhwloc-dev zlib1g zlib1g-dev clinfo dialog apt-utils \
libxml2-dev libclang-cpp${LLVM_VERSION}-dev libclang-cpp${LLVM_VERSION} \
llvm-${LLVM_VERSION}-dev

If LLVM is linked to PoCL statically (-DSTATIC_LLVM=ON cmake option, default OFF), then the libpolly-${LLVM_VERSION}-dev libzstd-dev packages might be also needed (at least on Ubuntu 22.04 with packages from apt.llvm.org).

Install requirements for Arch Linux:

pacman -S gcc patch hwloc cmake git pkg-config make ninja ocl-icd clang llvm llvm-libs clinfo opencl-headers

Install requirements for Fedora:

dnf install gcc gcc-c++ clinfo hwloc-devel hwloc-libs cmake git-core pkgconfig make ninja-build ocl-icd ocl-icd-devel clang clang-devel clang-libs llvm llvm-devel llvm-libs patch redhat-rpm-config findutils

There are also Dockerfiles available for a few most common linux distributions in tools/docker, looking into them might be helpful.

ICD / OpenCL 3.0 support

PoCL's build by default builds the runtime library, tests and examples. The tests and examples cannot be built against ICD loader which doesn't support OpenCL 3.0. In other words, they can only be built with disabled ICD (linking directly to PoCL) or with ICD supporting OpenCL 3.0.

The loader in Ubuntu (ocl-icd) supports OpenCL 3.0 since version 2.3.0; for Ubuntu earlier than 24.04, it can be installed from this PPA: https://launchpad.net/~ocl-icd/+archive/ubuntu/ppa

If you don't have a sufficiently new ICD, and you want the tests/examples, then either:

• disable building with ICD (-DENABLE_ICD=0)
• disable the tests & examples (-DENABLE_TESTS=0 -DENABLE_EXAMPLES=0)

Configure & Build

CMake version 3.15 or higher is required.

The build+install is the usual CMake way:

cd <directory-with-pocl-sources>
mkdir build
cd build
cmake [-D<option>=<value> ...] ..
make && make install

To see the default detected values, run cmake .. without any options, it will produce a summary.

Compiler support

Since pocl includes a compiler for the kernels, it both compiles (producing code) and is compiled (it consists of code). This distinction typically called "host" and "target": The host is where pocl is running, the target is where the OpenCL code will be running. These two systems can be wildly different.

Host compiler used to compile pocl can be GCC or Clang; the target compiler is always Clang+LLVM since pocl uses Clang/LLVM internally. For host compiler, you should use the one which your LLVM was compiled with (because the LLVM-related parts of pocl take LLVM's CXXFLAGS from llvm-config and pass them to the host compiler).

Clang / LLVM: target compiler supported versions

Note that pocl aims to support the latest LLVM version at the time of pocl release, plus the previous LLVM version. All older LLVM versions are supported with "best effort" basis; there might not be CI continuously testing the code base nor anyone fixing their possible breakage.

CMake options

For multiple-item options like KERNELLIB_HOST_CPU_VARIANTS, use ";" as separator (you'll have to escape it for bash).

• -DWITH_LLVM_CONFIG=<path-to-llvm-config> IMPORTANT Path to a llvm-config binary. This determines the LLVM installation used by pocl. If not specified, pocl will try to find and link against llvm-config in PATH env var (usually means your system LLVM).

• -DSTATIC_LLVM pocl uses llvm-config --libs to get list of LLVM libraries it should link to. With this flag enabled, it additionally passes --link-static to llvm-config; otherwise it passes --link-shared. Default is OFF (=shared).

• -DLLVM_SPIRV Path to a llvm-spirv binary of SPIRV-LLVM-Translator. Note that only the builds of open-source Khronos Translator are supported, and the binary must be built for the same LLVM version as PoCL is being built against. If not given, PoCL will try to autodetect a llvm-spirv binary with suitable version. Note that PoCL will also try to detect LLVMSPIRVLib library, and if found, it will take precedence over llvm-spirv binary.

• -DENABLE_ICD By default pocl's buildsystem will try to find an ICD and build pocl as a dynamic library named "libpocl". This option is useful if you want to avoid ICD and build pocl directly as libOpenCL library. See also :ref:`linking-with-icd`

• -DPOCL_INSTALL_<something>_DIR The equivalent of --bindir, --sbindir etc fine-tuning of paths for autotools. See the beginning of toplevel CMakeLists.txt for all the variables.

  Note that if CMAKE_INSTALL_PREFIX equals /usr then pocl.icd is installed to /etc/OpenCL/vendors, otherwise it's installed to ${CMAKE_INSTALL_PREFIX}/etc/OpenCL/vendors.

• -DLLC_HOST_CPU=<something> Defaults to auto-detection via llc. Run llc -mcpu=help for valid values. The CPU type is required to compile the "target" (kernel library) part of CPU backend.

  This variable overrides LLVM's autodetected host CPU at configure time. Useful when llc fails to detect the CPU (often happens on non-x86 platforms, or x86 with CPU newer than LLVM).

  Note that when this is set (set by default) and the KERNELLIB_HOST_CPU_VARIANTS variable is not distro, pocl will first try to find compiled kernel library for runtime-detected CPU then fallback to LLC_HOST_CPU. This works well if pocl is run where it was built, or the actual CPU is in the KERNELLIB_HOST_CPU_VARIANTS list, or the actual CPU is >= LLC_HOST_CPU feature-wise; otherwise it will likely fail with illegal instruction at runtime.

• -DKERNELLIB_HOST_CPU_VARIANTS You can control which CPUs the "target" part of CPU backend will be built for. Unlike LLC_HOST_CPU, this variable is useful if you plan to build for multiple CPUs. Defaults to "native" which is automagically replaced by LLC_HOST_CPU. Available CPUs are listed by llc -mcpu=help. See above for runtime CPU detection rules.

  Note that there's another available value on certain (x86, ppc) platforms. If set to distro, on x86, the KERNELLIB_HOST_CPU_VARIANTS variable will be set up with a few preselected sse/avx variants covering 99.99% of x86 processors, and the runtime CPU detection is slightly altered: pocl will find the suitable compiled library based on detected CPU features, so it cannot fail (at worst it'll degrade to SSE2 library).

• -DLLC_TRIPLE=<something> Controls what target triple pocl is built for. You can set this manually in case the autodetection fails. Example value: x86_64-pc-linux-gnu

• -DENABLE_TESTSUITES Which external (source outside pocl) testsuites to enable. For the list of testsuites, see examples/CMakeLists.txt or the examples directory. Set to all and pocl will try to autodetect & enable everything it can.

  Note that you may build testsuites outside pocl's build tree, and test multiple pocl builds with a single testsuite directory. To use this, run cmake with -DTESTSUITE_BASEDIR=<tests-builddir> and -DTESTSUITE_SOURCE_BASEDIR=<tests-sourcedir>. The directory structure mirrors that of pocl/examples. So to build e.g. AMD SDK 2.9 with -DTESTSUITE_BASEDIR=/home/pocltest-build -DTESTSUITE_SOURCE_BASEDIR=/home/pocltest-src, place the AMD-APP-SDK-v2.9-RC-lnx64.tgz file into /home/pocltest-src/AMDSDK2.9 directory.

• -DENABLE_TESTS=ON/OFF enable/disable compilation of internal tests.

• -DENABLE_EXAMPLES=ON/OFF enable/disable compilation of all examples. Disabling this makes ENABLE_TESTSUITES option unavailable.

• -DENABLE_POCLCC=ON/OFF enable/disable compilation of poclcc.

• -DENABLE_POCL_BUILDING=OFF When OFF, POCL_BUILDING option (which causes pocl to look for required files in build / source directories) will be ignored and pocl will always look in installed paths only.

• -DENABLE_CONFORMANCE=ON/OFF Ensures that certain build options which would result in non-conformant pocl build stay disabled. Defaults to OFF. Note that this does not quarantee a fully conformant build of pocl by itself. See :ref:`pocl-conformance` for details.

• -DENABLE_{A,L,T,UB}SAN - compiles pocl's host code (and tests + examples) with various sanitizers. Using more than one sanitizer at a time is untested. Using together with -DENABLE_ICD=OFF -DENABLE_LOADABLE_DRIVERS=OFF is highly recommended to avoid issues with loading order of sanitizer libraries.

• -DENABLE_LTTNG=ON/OFF - compile pocl with LTTng support for tracing. Requires LTTng to be installed on the host machine.

• -DENABLE_{CUDA,TCE,HSA,VULKAN,LEVEL0}=ON/OFF - enable various (non-CPU) backends. Usually requires some additional build dependencies; see their documentation.

• -DPOCL_DEBUG_MESSAGES=ON - when disabled, pocl is compiled without debug messages (POCL_DEBUG env var) support.

• -DEXAMPLES_USE_GIT_MASTER=ON - when enabled, examples (external programs in examples/ directory) are built from their git branches (if available), as opposed to default: building from release tars.

• -DINTEL_SDE_AVX512=<PATH> Path to Intel® Software Development Emulator. When this option is given, the LLVM host CPU is forcibly set to 'skylake-avx512', and the internal tests are run through the Emulator. Mostly useful to test AVX512.

• -DPOCL_ICD_ABSOLUTE_PATH=OFF The pocl.icd file (which the ICD loader uses to load the pocl lib) by default has a full path to the installed libpocl.so file. Set this option to OFF and pocl will only put the dynamic library name into pocl.icd.

Advanced CMake options: using extra flags for the builtin library

All of these empty by default. There are hardcoded defaults which may be overridden by setting these variables (rarely needed).

Extra parameters to llc

EXTRA_HOST_LLC_FLAGS

Extra parameters to clang

EXTRA_HOST_CLANG_FLAGS

Extra parameters to linker (links kernel to shared library which is then dlopened):

EXTRA_HOST_LD_FLAGS

EXTRA_KERNEL_FLAGS

is applied to all kernel library compilation commands, IOW it's for language-independent options

EXTRA_KERNEL_{C,CL,CXX}_FLAGS

cmake variables for per-language options for kernel library compilation

CMake: LLVM-less build

You can build a runtime-only pocl to run prebuilt pocl binaries on a CPU driver, or if you intend to only use PoCL with a driver that does not depend on LLVM.

When building a LLVM-less build with a CPU driver:

• First, build PoCL with LLVM somewhere.

• on that machine, set up env vars required for your device (if any), then run bin/poclcc -l. That should print something like:

  LIST OF DEVICES:
  0:
   Vendor:   AuthenticAMD
     Name:   pthread-AMD A10-7800 Radeon R7, 12 Compute Cores 4C+8G
  Version:   OpenCL 2.0 pocl HSTR: pthread-x86_64-unknown-linux-gnu-bdver3

The string after "HSTR:" is the device build hash.

• now build the LLVM-less pocl. You will need the device build hash from previous step:

  cmake -DENABLE_LLVM=0 -DHOST_DEVICE_BUILD_HASH=<something> ...

  This is required because pocl binaries contain a device hash, and the LLVM-less pocl needs to know which binaries it can load.

NOTE: If you've enabled the :ref:`almaif device <almaif_usage>` , HOST_DEVICE_BUILD_HASH can be set to anything you want. Reason being, fixed function accelerators don't require compiling OpenCL kernels, therefore, no hash will ever be matched.

Packaging PoCL

PoCL supports CPACK. Additionally, these CMake options are of interest:

• -DKERNELLIB_HOST_CPU_VARIANTS=distro to enable support for most multiple CPU families with runtime detection (cpu driver)
• -DPOCL_ICD_ABSOLUTE_PATH=OFF to not put absolute path into pocl.icd
• -DENABLE_POCL_BUILDING=OFF to disable embedding the build path into libpocl
• -DSTATIC_LLVM to link against Clang & LLVM static component libraries. This may help avoid symbol clashes with other LLVM libraries linked in the same executable.

Cross-compile PoCL

It's now possible to cross-compile pocl on x86-64 to run on ARM/MIPS/etc, There is a ToolchainExample.cmake file; copy it under different name, then follow the instructions in the file.

Building & running in Docker

Required storage for standard pocl build is about 1.5 GB per container)

Build & start Pocl container

• cd tools/docker
• pick a Dockerfile from tools/docker, e.g. Fedora/default
• to build PoCL: sudo docker build -t TAG -f Fedora/default ., where TAG is a name you choose for the build (must be lowercase)
• to run the tests on the built PoCL: sudo docker run -t TAG
• this will by default use master branch of pocl git; to use a different branch/commit, run docker build with --build-arg GIT_COMMIT=<branch/commit>

Dockerfiles

Dockerfiles are named according to what they build, or the release they're based on:

• default: builds pocl, then runs the internal tests from build dir.

  Uses latest release of a distribution, with whatever is the default version of LLVM.

• distro: does a distribution-friendly build: enables runtime detection of CPU,

  installs pocl into system path, then runs the internal tests

• <release>: same as above, except uses specific release and specific LLVM version (the latest available in that release).
• conformance: builds & installs Pocl, then runs conformance test suite (the shortest version of it)

ARM 32/64bit support

Status:

PoCL builds (as of Dec 2023) on ODROID XU3 and ODROID C2 and almost all tests pass.

ARM specific build notes:

• DO NOT use Clang/LLVM downloaded directly from llvm.org, they only work on the distro where they were compiled. Ubuntu LTS these days ships multiple llvm versions even quite recent ones; get Clang+LLVM from your distro's package manager or build it yourself.
• LLVM might not recognize your cpu, in which case CMake will complain. Run cmake with -DLLC_HOST_CPU=<yourcpu>. "yourcpu" must be something LLVM recognizes, usually it's simply "cortex-aXX" like cortex-a15 etc. You can get the full list by running llc -mcpu=help.

RISC-V support

The RISC-V support has been tested (as of Dec 2023) on Starfive VisionFive 2 using Ubuntu 23.10 preinstalled image, with LLVM 17 and GCC 13.2; of the internal tests, 98% tests pass, 4 tests fail out of 253. In particular, tests using printf with vector arguments are broken ATM. Other boards / CPUs have not been tested. RISC Vector extension is not supported.

RISC-V specific build notes:

• Avoid older LLVM and GCC versions (like GCC 11 / Clang 14 on the official Starfive Debian images) as much as possible. Code generation is much better with recent versions, and your experience will generally better

• LLVM might not recognize your CPU, in which case CMake will complain. Run cmake with -DLLC_HOST_CPU=<yourcpu>. "yourcpu" must be something LLVM recognizes; you can get the full list by running llc -mcpu=help.

• on RISC-V, PoCL additionally needs to pass a target ABI flag to the compiler. There is some autodetection in PoCL but right now it's limited, and Clang unfortunately does not always get the defaults correctly. If you get errors similar to:

  "can't link double-float modules with soft-float modules"

  from linker, then most likely PoCL used the incorrect ABI. You can explicitly specify the ABI to use with the HOST_CPU_TARGET_ABI CMake option.

PowerPC support

PoCL is used to provide OpenCL on IBM AC922 computers featuring IBM Power9 processors and Nvidia Tesla V100 GPU interconnected with NVlink v2 (up to 72 GByte/s). This has been tested under debian_11 and Ubuntu_20.04.

Officially, Nvidia does not support OpenCL on this platform and the driver they are shipping is lacking the compiler part.

## Building tricks (as of 04/2023):

The PPC64le features 128-bit vector unit (Altivec/VSX) which are easily confused by the C++ compiler with the C++ vector instruction when using the compile option -std=c++XX. The corresponding code usually fails compiling. The trick is to pass the option -std=gnu++XX.

For example, when configuring pocl: ` cmake .. -DLLVM_CXXFLAGS="-std=gnu++14 ..." -DENABLE_CUDA=ON `

The full list of options for the CXXFLAGS is obtained with: ` llvm-config --cxxflags|sed -e "s/std=c/std=gnu/" ` Later on, the build continues with make ...

Note the CUDA option to enable the GPU support on those computers.