Usage

The basic usage of pocl should be as easy as any other OpenCL implementation.

While it is possible to link against pocl directly, the recommended way is to use the ICD interface.

Linking your program with pocl through an icd loader

You can link your OpenCL program against an ICD loader. If your ICD loader is correctly configured to load pocl, your program will be able to use pocl. See the section below for more information about ICD and ICD loaders.

Example of compiling an OpenCL host program using the free ocl-icd loader:

gcc example1.c -o example `pkg-config --libs --cflags OpenCL`

Example of compiling an OpenCL host program using the AMD ICD loader (no pkg-config support):

gcc example1.c -o example -lOpenCL

Installable client driver (ICD)

pocl is built with the ICD extensions of OpenCL by default. This allows you to have several OpenCL implementations concurrently on your computer, and select the one to use at runtime by selecting the corresponding cl_platform. ICD support can be disabled by adding the flag:

-DENABLE_ICD=OFF

to the CMake invocation.

The ocl-icd ICD loader allows to use the OCL_ICD_VENDORS environment variable to specify a (non-standard) replacement for the /etc/OpenCL/vendors directory.

An ICD loader is an OpenCL library acting as a “proxy” to one of the various OpenCL implementations installed in the system. pocl does not provide an ICD loader itself, but NVidia, AMD, Intel, Khronos, and the free ocl-icd project each provides one.

Linking your program directly with pocl

Passing the appropriate linker flags is enough to use pocl in your program. However, please bear in mind that:

The pkg-config tool is used to locate the libraries and headers in the installation directory.

Example of compiling an OpenCL host program against pocl using the pkg-config:

gcc example1.c -o example `pkg-config --libs --cflags pocl`

In this link mode, your program will always require the pocl OpenCL library. It wont be able to run with another OpenCL implementation without recompilation.

Using pocl on MacOSX

On MacOSX, you can either link your program directly with pocl or link through the ICD loader by KhronosGroup.

Even if you use an ICD loader, the Apple OpenCL implementation will still be invisible, unless you use a wrapper library to expose the Apple OpenCL implementation as an ICD.

When ENABLE_ICD is turned off and an application links directly with PoCL, the only platform that is visible to the application will be PoCL.

Tuning pocl behavior with ENV variables

The behavior of pocl can be controlled with multiple environment variables listed below. The variables are helpful both when using and when developing pocl.

  • POCL_AFFINITY

    Linux-only, specific to ‘cpu’ driver. If set to 1, each thread of the driver sets its affinity to its index. This may be useful with very long running kernels, or when using subdevices. Defaults to 0 (most people don’t need this).

  • POCL_BINARY_SPECIALIZE_WG

    By default the PoCL program binaries store generic kernel binaries which can be executed across any grid dimensions. This configuration variable can be used to also include specialized work-group functions in the binaries, by defining a comma separated list of strings that describe the specialized versions. The strings adhere to the directory names in the PoCL cache from which the binaries are captured.

    Example:

    POCL_BINARY_SPECIALIZE_WG='2-1-1,0-0-0-goffs0,13-1-1-smallgrid,128-2-1-goffs0-smallgrid' poclcc [...]

    This makes poclcc generate a binary which contains the generic work-group function binary, a work-group function that is specialized for local size of 2x1x1, another with generic local size but specialized for the global offset at origo, one with local size of 13x1x1, but which is specialized for a “small grid” (size defined by the device driver), and finally one that is specialized for local size 128x2x1, an origo global offset and a small grid.

  • POCL_BITCODE_FINALIZER

    Defines a custom command that can manipulate the final kernel work-group function bitcode produced after all LLVM optimizations and before entering code generation. This can be useful, for example, to add instrumentation to the LLVM bitcode before proceeding to the backend.

    Example:

    POCL_BITCODE_FINALIZER='verificarlo %(bc) --emit-llvm -o %(bc)' examples/example1/example1

    This results in running the above command with ‘%(bc)’ strings replaced with the path of the final bitcode’s temporary file. Note that the modified bitcode should be written over the same file for it to get picked to the code generation.

    Please note that setting the env doesn’t force regeneration of the kernel binaries if they are found in the kernel compiler cache. You can either use POCL_KERNEL_CACHE=0 to disable the kernel cache, or wipe the kernel cache directory manually to force kernel binary rebuild.

  • POCL_BUILDING

If set, the pocl helper scripts, kernel library and headers are searched first from the pocl build directory. Only has effect if ENABLE_POCL_BUILDING was enabled at build (by default it is).

  • POCL_CACHE_DIR

If this is set to an existing directory, pocl uses it as the cache directory for all compilation results. This allows reusing compilation results between pocl invocations. If this env is not set, then the default cache directory will be used, which is $XDG_CACHE_HOME/pocl/kcache (if set) or $HOME/.cache/pocl/kcache/ on Unix-like systems.

  • POCL_CPU_LOCAL_MEM_SIZE

Set the local memory size of the CPU devices (cpu, cpu-minimal) to the given amount in bytes instead of the default one.

  • POCL_CPU_MAX_CU_COUNT

The maximum number of threads created for work group execution in the ‘cpu’ device driver. The default is to determine this from the number of hardware threads available in the CPU.

  • POCL_DEBUG

Enables debug messages to stderr. This will be mostly messages from error condition checks in OpenCL API calls and Event/API timing information. Useful to e.g. distinguish between various reasons a call could return CL_INVALID_VALUE. If clock_gettime is available, messages will include a timestamp.

The old way (setting POCL_DEBUG to 1) has been updated to support categories. Using this limits the amount of debug messages produced. Current options are: error,warning,general,memory,llvm,events,cache,locking,refcounts,timing,hsa,tce,cuda,vulkan,proxy,all. Note: setting POCL_DEBUG to 1 still works and equals error+warning+general.

  • POCL_DEBUG_LLVM_PASSES

When set to 1, enables debug output from LLVM passes during optimization.

  • POCL_DEVICES and POCL_x_PARAMETERS

POCL_DEVICES is a space separated list of the device instances to be enabled. This environment variable is used for the following devices:

  • cpu-minimal A minimalistic example device driver for executing

    kernels on the host CPU. No multithreading.

  • cpu Execution of OpenCL kernels on the host CPU using

    (by default) all available CPU threads.

  • cuda An experimental driver that uses libcuda to execute on NVIDIA GPUs.

  • hsa Uses HSA Runtime API to control HSA-compliant

    kernel agents that support HSAIL finalization (deprecated).

  • vulkan An experimental driver that uses Vulkan and SPIR-V for executing on

    Vulkan supported devices.

  • ttasim Device that simulates a TTA device using the

    TCE’s ttasim library. Enabled only if TCE libraries installed.

  • level0 An experimental driver that uses libze to execute on Intel GPUs.

If POCL_DEVICES is not set, one cpu device will be used. To specify parameters for drivers, the POCL_<drivername><instance>_PARAMETERS environment variable can be specified (where drivername is in uppercase). Example:

export POCL_DEVICES="cpu ttasim ttasim"
export POCL_TTASIM0_PARAMETERS="/path/to/my/machine0.adf"
export POCL_TTASIM1_PARAMETERS="/path/to/my/machine1.adf"

Creates three devices, one ‘cpu’ device with multithreading and two TTA device simulated with the ttasim. The ttasim devices gets a path to the architecture description file of the tta to simulate as a parameter. POCL_TTASIM0_PARAMETERS will be passed to the first ttasim driver instantiated and POCL_TTASIM1_PARAMETERS to the second one.

  • POCL_EXTRA_BUILD_FLAGS

Adds the contents of the environment variable to all clBuildProgram() calls. E.g. POCL_EXTRA_BUILD_FLAGS="-g -cl-opt-disable" can be useful for force adding debug data all the built kernels to help debugging kernel issues with tools such as gdb or valgrind.

  • POCL_KERNEL_CACHE

If this is set to 0 at runtime, kernel compilation files will be deleted at clReleaseProgram(). Note that it’s currently not possible for pocl to avoid interacting with LLVM via on-disk files, so pocl requires some disk space at least temporarily (at runtime).

  • POCL_LEAVE_KERNEL_COMPILER_TEMP_FILES

If this is set to 1, the kernel compiler cache/temporary directory that contains all the intermediate compiler files are left as it is. This will be handy for debugging

  • POCL_LEVEL0_JIT

Sets up Just-In-Time compilation in the Level0 driver. (see Level Zero driver for details) Accepted values: {0,1,auto}

  • 0 = always disable JIT

  • 1 = always use JIT,

  • auto (default) = guess based on program’s kernel count & SPIR-V size.

  • POCL_LLVM_VERIFY

    if enabled, some drivers (CUDA, CPU, Level0) use an extra step of verification of LLVM modules at certain stages (program.bc always, kernel bitcode (parallel.bc) only with some drivers). Defaults to 0 if CMAKE_BUILD_TYPE=Debug and 1 otherwise.

  • POCL_MAX_WORK_GROUP_SIZE

Forces the maximum WG size returned by the device or kernel work group queries to be at most this number. For certain devices, this is can only be lower than their hardware limits.

  • POCL_MEMORY_LIMIT

Integer option, unit: gigabytes. Limits the total global memory size reported by pocl for the CPU devices (this will also affect local/constant/max-alloc-size numbers, since these are derived from global mem size).

  • POCL_OFFLINE_COMPILE

Bool. When enabled(==1), some drivers will create virtual devices which are only good for creating pocl binaries. Requires those drivers to be compiled with support for compilation for those devices.

  • POCL_SIGFPE_HANDLER

Defaults to 1. If set to 0, pocl will not install the SIGFPE handler. See Known issues related to CTS

  • POCL_SIGUSR2_HANDLER

When set to 1 (default 0), pocl installs a SIGUSR2 handler that will print some debugging information. Currently it prints the count of live cl_* objects by type (buffers, events, etc).

  • POCL_STARTUP_DELAY

    Default 0. If set to an integer N > 0, libpocl will make a pause of N seconds once, when it’s loading. Useful e.g. to set up a LTTNG tracing session.

  • POCL_TRACING, POCL_TRACING_OPT and POCL_TRACING_FILTER

If POCL_TRACING is set to some tracer name, then all events will be traced automatically. Depending on the backend, traces may be output in different formats and collected in a different way. POCL_TRACING_FILTER is a comma separated list of string to indicate which event status should be filtered. For instance to trace complete and running events POCL_TRACING_FILTER should be set to “complete,running”. Default behavior is to trace all events.

cq – Dumps a simple per-kernel execution time statistics at the

program exit time which is collected from command queue start and finish time stamps. Useful for quick and easy profiling purposes with accurate kernel execution time stamps produced in a per device way. Currently only tracks kernel timings, and POCL_TRACING_FILTER has no effect.

text – Basic text logger for each events state

Use POCL_TRACING_OPT=<file> to set the output file. If not specified, it defaults to pocl_trace_event.log

lttng – LTTNG tracepoint support. When activated, a lttng session
must be started. The following tracepoints are available:

  • pocl_trace:ndrange_kernel -> Kernel execution

  • pocl_trace:read_buffer -> Read buffer

  • pocl_trace:write_buffer -> Write buffer

  • pocl_trace:copy_buffer -> Copy buffer

  • pocl_trace:map -> Map image/buffer

  • pocl_trace:command -> other commands

For more information, please see lttng documentation: http://lttng.org/docs/#doc-tracing-your-own-user-application

  • POCL_VECTORIZER_REMARKS

When set to 1, prints out remarks produced by the loop vectorizer of LLVM during kernel compilation.

  • POCL_VULKAN_VALIDATE

When set to 1, and the Vulkan implementation has the validation layers, enables the validation layers in the driver. You will also need POCL_DEBUG=vulkan or POCL_DEBUG=all to see the output printed.

  • POCL_WORK_GROUP_METHOD

The kernel compiler method to produce the work group functions from multiple work items. Legal values:

auto – Choose the best available method depending on the

kernel and the work group size. Use POCL_FULL_REPLICATION_THRESHOLD=N to set the maximum local size for a work group to be replicated fully with ‘repl’. Otherwise, ‘loops’ is used.

loops – Create for-loops that execute the work items

(under stabilization). The drawback is the need to save the thread contexts in arrays.

The loops will be unrolled a certain number of times of which maximum can be controlled with POCL_WILOOPS_MAX_UNROLL_COUNT=N environment variable (default is to not perform unrolling).

loopvec – Create work-item for-loops (see ‘loops’) and execute

the LLVM LoopVectorizer. The loops are not unrolled but the unrolling decision is left to the generic LLVM passes (the default).

repl – Replicate and chain all work items. This results

in more easily scalarizable private variables, thus might avoid storing work-item context to memory. However, the code bloat is increased with larger WG sizes.

cbs – Use continuation-based synchronization to execute work-items

on non-SPMD devices. CBS is expected to work for kernels that ‘loops’ does not support. For most other kernels it is expected to perform slightly worse. Also enables the LLVM LoopVectorizer.

An in-depth explanation of the implementation of CBS and how it compares to the other approaches can be found in [this thesis](https://joameyer.de/hipsycl/Thesis_JoachimMeyer.pdf).

  • POCL_WORK_GROUP_SPECIALIZATION

    PoCL specializes work-groups at kernel command launch time by default to optimize the execution performance with the cost of cached variations of the kernels with the different specialization values.

    The kernel command parameters PoCL currently specializes with include the local size, global offset zero or non-zero and maximum grid size. The specialization can be disabled by setting this environment variable to 0.