clvk

clvk is a conformant implementation of OpenCL 3.0 on top of Vulkan using clspv as the compiler.

Supported applications

Full list

Getting dependencies

clvk depends on the following external projects:

• clspv and its dependencies
• OpenCL-Headers
• SPIRV-Headers
• SPIRV-Tools
• SPIRV-LLVM-Translator

clvk also (obviously) depends on a Vulkan implementation. The build system supports a number of options there (see Building section).

To fetch all the dependencies needed to build and run clvk, please run:

git submodule update --init --recursive
./external/clspv/utils/fetch_sources.py --deps llvm

Building

clvk uses CMake for its build system.

Getting started

To build with the default configuration options, just use following:

mkdir -p build
cd build
cmake ../
make -j$(nproc)

Build options

The build system allows a number of things to be configured.

Vulkan implementation

You can select the Vulkan implementation that clvk will target with the CLVK_VULKAN_IMPLEMENTATION build system option. Two options are currently supported:

• -DCLVK_VULKAN_IMPLEMENTATION=system instructs the build system to use the Vulkan implementation provided by your system as detected by CMake's find_package(Vulkan). This is the default. You can use the Vulkan SDK to provide headers and libraries to build against.

• -DCLVK_VULKAN_IMPLEMENTATION=custom instructs the build system to use the values provided by the user manually using -DVulkan_INCLUDE_DIRS and -DVulkan_LIBRARIES.

Tests

It is possible to disable the build of the tests by passing -DCLVK_BUILD_TESTS=OFF.

It is also possible to disable only the build of the tests linking with the static OpenCL library by passing -DCLVK_BUILD_STATIC_TESTS=OFF.

By default, tests needing gtest are linked with the libraries coming from llvm (through clspv). It is possible to use other libraries by passing -DCLVK_GTEST_LIBRARIES=<lib1>;<lib2> (semicolumn separated list).

Assertions

Assertions can be controlled with the CLVK_ENABLE_ASSERTIONS build option. They are enabled by default in Debug builds and disabled in other build types.

OpenCL conformance tests

Passing -DCLVK_BUILD_CONFORMANCE_TESTS=ON will instruct CMake to build the OpenCL conformance tests. This is not expected to work out-of-the box at the moment.

It is also possible to build GL and GLES interroperability tests by passing -DCLVK_BUILD_CONFORMANCE_TESTS_GL_GLES_SUPPORTED=ON.

Clspv compilation

You can select the compilation style that clvk will use with Clspv via the CLVK_CLSPV_ONLINE_COMPILER option. By default, Clspv is run in a separate process.

• -DCLVK_CLSPV_ONLINE_COMPILER=1 will cause clvk to compile kernels in the same process via the Clspv C++ API.

You can build clvk using an external Clspv source tree by setting -DCLSPV_SOURCE_DIR=/path/to/clspv/source/.

SPIRV components

All needed SPIRV components are added to clvk using git submodules. It is possible to disable the build of those component or to reuse already existing sources:

SPIRV-Headers

SPIRV_HEADERS_SOURCE_DIR can be overriden to use another SPIRV-Headers repository.

SPIRV-Tools

SPIRV_TOOLS_SOURCE_DIR can be overriden to use another SPIRV-Tools repository. You can also disable the build of SPIRV-Tools by setting -DCLVK_BUILD_SPIRV_TOOLS=OFF.

SPIRV-LLVM-Translator

LLVM_SPIRV_SOURCE can be overriden to use another SPIRV-LLVM-Translator repository. Note that it is not used if the compiler support is disabled (enabled by default).

Sanitizers

Support for sanitizers is integrated into the build system:

• CLVK_ENABLE_ASAN can be used to enable AddressSanitizer.
• CLVK_ENABLE_TSAN can be used to enable ThreadSanitizer.
• CLVK_ENABLE_UBSAN can be used to enable UndefinedBehaviorSanitizer.

Building for Android

clvk can be built for Android using the Android NDK toolchain.

1. Download and extract the NDK toolchain to a directory (/path/to/ndk)
2. Pass the following options to CMake:
   • -DCMAKE_TOOLCHAIN_FILE=/path/to/ndk/build/cmake/android.toolchain.cmake
   • -DANDROID_ABI=<ABI_FOR_THE_TARGET_DEVICE>, most likely arm64-v8a
   • -DVulkan_LIBRARY=/path/to/ndk/**/<api-level>/libvulkan.so
3. That should be it!

Using

Via the OpenCL ICD Loader

clvk supports the cl_khr_icd OpenCL extension that makes it possible to use the OpenCL ICD Loader.

Directly

To use clvk to run an OpenCL application, you just need to make sure that the clvk shared library is picked up by the dynamic linker.

When clspv is not built into the shared library (which is currently the default), you also need to make sure that clvk has access to the clspv binary. If you wish to move the built library and clspv binary out of the build tree, you will need to make sure that you provide clvk with a path to the clspv binary via the CLVK_CLSPV_PATH environment variable (see Environment variables).

Unix-like systems (Linux, macOS)

The following ought to work on Unix-like systems:

$ LD_LIBRARY_PATH=/path/to/build /path/to/application

# Running the included simple test
$ LD_LIBRARY_PATH=./build ./build/simple_test

With perfetto traces

Perfetto is a production-grade open-source stack for performance instrumentation and trace analysis. It offers services and libraries and for recording system-level and app-level traces, native + java heap profiling, a library for analyzing traces using SQL and a web-based UI to visualize and explore multi-GB traces.

-- https://github.com/google/perfetto/tree/v39.0#perfetto---system-profiling-app-tracing-and-trace-analysis

Perfetto can be enabled by passing the following options to CMake:

• -DCLVK_PERFETTO_ENABLE=ON
• -DCLVK_PERFETTO_SDK_DIR=/path/to/perfetto/sdk

The perfetto SDK can be found in the Perfetto Github repository

If you already have a perfetto library in your system, you still need to provide the path to the SDK directory so the build system can find perfetto.h. But you should also provide the following option to CMake:

• -DCLVK_PERFETTO_LIBRARY=<your_perfetto_library_name>

By default, clvk will use Perfetto's InProcess backend, which means that you just have to run your application to generate traces. Environment variables can be used to control the maximum size of traces and what file they are saved to.

If you'd rather use Perfetto's System backend, pass the following option to CMake:

• -DCLVK_PERFETTO_BACKEND=System

Once traces have been generated, you can view them using the perfetto trace viewer.

Windows

Copy OpenCL.dll into a system location or alongside the application executable you want to run with clvk.

Raspberry Pi

Make sure you have an up-to-date Mesa installed on your system. At the time of writing (May 2023) RaspberryPi OS (Debian 11) did not, but Ubuntu 23.04 does have a compatible vulkan driver.

Install the prerequisites:

$ sudo apt install mesa-vulkan-drivers vulkan-tools libvulkan-dev git cmake clang clinfo

Check if your vulkan implementation has the VK_KHR_storage_buffer_storage_class extension or supports Vulkan 1.1. Note that it's not enough if this is the case for llvmpipe. You need v3dv to support this too. If it does not, your Mesa is too old.

To fetch the dependencies, do python3 ./external/clspv/utils/fetch_sources.py because the python interpreter may not be found if not explicitly called as python3.

Building will take many hours on a rPi4. Maybe you can skip some tool building, but building with default settings works, at least.

Once the libOpenCL.so library has been built, verify with:

LD_LIBRARY_PATH=/path/to/build clinfo

With global timing information

Global timing information about API functions as well as some internal functions can be logged at the end of the execution.

To enable it, pass the following option to CMake:

• -DCLVK_ENABLE_TIMING=ON

Here is an example of what to expect running simple_test:

[CLVK] 0.00 ms -> clReleaseContext (1 blocks, avg 0.001 ms)
[CLVK] 0.02 ms -> clReleaseProgram (1 blocks, avg 0.024 ms)
[CLVK] 0.01 ms -> clReleaseKernel (1 blocks, avg 0.008 ms)
[CLVK] 0.00 ms -> clReleaseCommandQueue (1 blocks, avg 0.005 ms)
[CLVK] 0.01 ms -> clReleaseMemObject (1 blocks, avg 0.007 ms)
[CLVK] 0.00 ms -> clEnqueueUnmapMemObject (1 blocks, avg 0.002 ms)
[CLVK] 0.03 ms -> clEnqueueMapBuffer (1 blocks, avg 0.034 ms)
[CLVK] 5.04 ms -> vkQueueWaitIdle (1 blocks, avg 5.043 ms)
[CLVK] 5.13 ms -> executor_wait (1 blocks, avg 5.126 ms)
[CLVK] 0.02 ms -> vkQueueSubmit (1 blocks, avg 0.022 ms)
[CLVK] 5.07 ms -> execute_cmd: CLVK_COMMAND_BATCH (3 blocks, avg 1.690 ms)
[CLVK] 5.09 ms -> execute_cmds (3 blocks, avg 1.698 ms)
[CLVK] 0.00 ms -> extract_cmds_required_by (3 blocks, avg 0.001 ms)
[CLVK] 0.00 ms -> enqueue_command (3 blocks, avg 0.001 ms)
[CLVK] 0.00 ms -> end_current_command_batch (1 blocks, avg 0.002 ms)
[CLVK] 0.12 ms -> flush_no_lock (4 blocks, avg 0.030 ms)
[CLVK] 5.20 ms -> clFinish (2 blocks, avg 2.602 ms)
[CLVK] 97.06 ms -> clEnqueueNDRangeKernel (1 blocks, avg 97.063 ms)
[CLVK] 0.00 ms -> clSetKernelArg (1 blocks, avg 0.002 ms)
[CLVK] 0.01 ms -> clCreateBuffer (1 blocks, avg 0.012 ms)
[CLVK] 0.01 ms -> clCreateCommandQueue (1 blocks, avg 0.009 ms)
[CLVK] 0.19 ms -> clCreateKernel (1 blocks, avg 0.187 ms)
[CLVK] 237.71 ms -> clBuildProgram (1 blocks, avg 237.713 ms)
[CLVK] 0.02 ms -> clCreateProgramWithSource (1 blocks, avg 0.016 ms)
[CLVK] 0.00 ms -> clCreateContext (1 blocks, avg 0.000 ms)
[CLVK] 0.00 ms -> clGetDeviceInfo (1 blocks, avg 0.001 ms)
[CLVK] 0.00 ms -> clGetDeviceIDs (1 blocks, avg 0.001 ms)
[CLVK] 0.00 ms -> clGetPlatformInfo (1 blocks, avg 0.002 ms)
[CLVK] 0.00 ms -> clGetPlatformIDs (1 blocks, avg 0.000 ms)

Tuning clvk

clvk can be tuned to improve the performance of specific workloads or on specific platforms. While we try to have the default parameters set at their best values for each platform, they can be changed for specific applications. One of the best way to know whether something can be improved is to use traces to understand what should be changed.

Group size

clvk is grouping commands and waiting for a call to clFlush or any blocking calls (clFinish, clWaitForEvents, etc.) to submit those groups for execution.

clvk's default group flushing behaviour can be controlled using the following two variables to flush groups as soon as a given number of commands have been grouped:

• CLVK_MAX_CMD_GROUP_SIZE
• CLVK_MAX_FIRST_CMD_GROUP_SIZE

Batch size

clvk relies on vulkan to offload workoad to the GPU. As such, it is better to batch OpenCL commands (translated into vulkan commands) into a vulkan command buffer. But doing that may increase the latency to start running commands.

The size of those batches can be controlled using the following two variables:

• CLVK_MAX_CMD_BATCH_SIZE
• CLVK_MAX_FIRST_CMD_BATCH_SIZE

Configuration

Many aspects of clvk's behaviour can be configured using configuration files and/or environment variables. clvk attempts to get its configuration from the following sources (in the order documented here). Values obtained from each source take precedence over previously obtained values.

1. Configuration file in /usr/local/etc/clvk.conf
2. System-wide configuration in /etc/clvk.conf
3. Per-user configuration in ~/.config/clvk.conf
4. clvk.conf in the current directory
5. An additional configuration file specified using the CLVK_CONFIG_FILE environment variable, if provided
6. Environment variables for individual configuration options

Configuration files use a key-value format and allow comments beginning with #:

# Here's a comment
option = value

other_option = 42

Options names are lowercase (e.g myoption) in configuration files but uppercase and prefixed with CLVK_ in environment variables (e.g. CLVK_MYOPTION).

Here is a list of all the configuration options that clvk supports documented using the name of the corresponding environment variable.

• CLVK_CONFIG_FILE specifies the path to an additional configuration file.

• CLVK_LOG controls the level of logging

  • 0: only print fatal messages (default)
  • 1: print errors as well
  • 2: print warnings as well
  • 3: print information messages as well
  • 4: print all debug messages

• CLVK_LOG_COLOUR controls colour logging

  • 0: disabled
  • 1: enabled (default when the output is a terminal)

• CLVK_LOG_DEST controls where the logging output goes

  • stderr: logging goes to the standard error (default)
  • stdout: logging goes to the standard output
  • file:<fname>: logging goes to <fname>. The file will be created if it does not exist and will be truncated.

• CLVK_LOG_GROUPS controls what logging groups are enabled. A comma-separated list of group enable/disable requests is accepted. A group is enabled by giving its name and disabled by giving its name with a - prefix. A few examples:

  • api enables logging of the OpenCL API calls encountered and only that.
  • -refcounting disables logging of object reference counting but keeps all other groups enabled by default.

  All groups are enabled by default. The first group enabled replaces the default.

• CLVK_CLSPV_PATH to provide a path to the clspv binary to use

• CLVK_LLVMSPIRV_BIN to provide a path to the llvm-spirv binary to use

• CLVK_ENABLE_SPIRV_IL to enable support for SPIR-V as an intermediate language

  • 0: disabled
  • 1: enabled (default)

• CLVK_VALIDATION_LAYERS allows to enable Vulkan validation layers

  • 0: disabled (default)
  • 1: enabled

• CLVK_CLSPV_OPTIONS to provide additional options to pass to clspv

• CLVK_CLSPV_NATIVE_BUILTINS comma separated list of builtins that will use the native implementation by default

• CLVK_CLSPV_LIBRARY_BUILTINS comma separated list of builtins that will be forced to use the libclc implementation

• CLVK_QUEUE_PROFILING_USE_TIMESTAMP_QUERIES to use timestamp queries to measure the CL_PROFILING_COMMAND_{START,END} profiling infos on devices that do not support VK_EXT_calibrated_timestamps.

  • 0: disabled (default)
  • 1: enabled

  WARNING: the values will not use the same time base as that used for CL_PROFILING_COMMAND_{QUEUED,SUBMIT} but this allows to get closer-to-the-execution timestamps.

• CLVK_SPIRV_VALIDATION controls SPIR-V validation behaviour.

  • 0: skip validation
  • 1: warn when validation fails
  • 2: fail compilation and report an error when validation fails (default)

• CLVK_SKIP_SPIRV_CAPABILITY_CHECK to avoid checking whether the Vulkan device supports all of the SPIR-V capabilities declared by each SPIR-V module.

• CLVK_MAX_BATCH_SIZE to control the maximum number of commands that can be recorded in a single command buffer.

• CLVK_KEEP_TEMPORARIES to keep temporary files created during program build, compilation and link operations.

  • 0: disabled (default)
  • 1: enabled

• CLVK_CACHE_DIR specifies a directory used for caching compiled program data between applications runs. The user is responsible for ensuring that this directory is not used concurrently by more than one application.

• CLVK_COMPLIER_TEMP_DIR specifies a directory used to create a temporary folder to store compiled program data used in a single run. This folder shall have write permission (default: current directory).

• CLVK_MAX_CMD_GROUP_SIZE specifies the maximum number of commands in a group. When a group reaches this number, it is automatically flushed.

• CLVK_MAX_FIRST_CMD_GROUP_SIZE specifies the maximum number of commands in a group when there is no group to be processed or being processed in the queue.

• CLVK_MAX_CMD_BATCH_SIZE specifies the maximum number of commands per batch. When this number is reached, the batch will be added to the current group of commands, and a new batch will be created.

• CLVK_MAX_FIRST_CMD_BATCH_SIZE specifies the maximum number of commands per batch when there is no batch to be processed or being processed in the queue.

• CLVK_PERFETTO_TRACE_MAX_SIZE specifies the maximum size (in kB) of traces generated by Perfetto. It only applies when using Perfetto with the InProcess backend.

• CLVK_PERFETTO_TRACE_DEST specifies the filename to use for the traces generated by Perfetto (default: clvk.perfetto-trace).

• CLVK_OPENCL_VERSION specifies the opencl version reported by clvk. The version needs to follow the layout used by CL_MAKE_VERSION from the OpenCL Headers:

  • 0x00c00000: meaning CL3.0 (default)
  • 0x00402000: meaning CL1.2

• CLVK_SUPPORTS_FILTER_LINEAR specifies whether using samplers with CL_FILTER_LINEAR is supported (default: true). Note that this is not required for OpenCL conformance on GPU devices and thus can be disabled to pass conformance on devices that do not support linear filtering with all the image formats required for conformance.

• CLVK_PREFERRED_SUBGROUP_SIZE specifies the subgroup size to use if nothing is specified in the kernel. When not set use the default value reported by the Vulkan driver.

• CLVK_FORCE_SUBGROUP_SIZE specifies the subgroup size to use, overriding everything.

• CLVK_QUEUE_GLOBAL_PRIORITY specifies the queue global priority to use if it is supported by the driver:

  • 0: VK_QUEUE_GLOBAL_PRIORITY_LOW_KHR
  • 1: VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_KHR (default)
  • 2: VK_QUEUE_GLOBAL_PRIORITY_HIGH_KHR
  • 3: VK_QUEUE_GLOBAL_PRIORITY_REALTIME_KHR

• CLVK_MAX_ENTRY_POINTS_INSTANCES specifies the number of instances of a kernel that can be in flight at the same time. Increasing this value has an impact on the memory usage as it will allocate more descriptor sets per kernel (default: 2048).

• CLVK_ENQUEUE_COMMAND_RETRY_SLEEP_US specifies the time to wait between two attempts to enqueue a command. It is disabled by default, meaning that if an enqueue fails, it returns an error. When specified, it will retry as long as there are groups in flight (commands being processed).

Limitations

• Only one device per CL context
• No support for out-of-order queues
• No support for device partitioning
• No support for native kernels
• All the limitations implied by the use of clspv
• ... and problably others