NOTE: VC4CL will NOT work with Raspberry Pi 4, since it has an incompatible GPU!
VC4CL is an implementation of the OpenCL 1.2 standard for the VideoCore IV GPU (found in Raspberry Pi 1 - 3 models).
The implementation consists of:
- The VC4CL OpenCL runtime library, running on the host CPU to compile, run and interact with OpenCL kernels.
- The VC4C compiler, converting OpenCL kernels into machine code. This compiler also provides an implementation of the OpenCL built-in functions.
- The VC4CLStdLib, the platform-specific implementation of the OpenCL C standard library, is linked in with the kernel by VC4C
The VC4CL implementation supports the EMBEDDED PROFILE of the OpenCL standard version 1.2.
Additionally the cl_khr_icd
extension is supported, to allow VC4CL to be found by an installable client driver loader (ICD). This enables VC4CL to be used in parallel with another OpenCL implementation, e.g. pocl, which executes OpenCL code on the host CPU.
The OpenCL version 1.2 was selected as target standard version, since it is the last version of the OpenCL standard where all mandatory features can be supported.
VC4CL supports the EMBEDDED PROFILE of the OpenCL-standard, which is a trimmed version of the default FULL PROFILE. The most notable features, which are not supported by the VC4CL implementation are images, the long
and double
data-types, device-side printf
and partitioning devices. See RuntimeLibrary for more details of (not) supported features.
The VideoCore IV GPU, in the configuration as found in the Raspberry Pi models, has a theoretical maximum performance of 24 GPFLOS and is therefore very powerful in comparison to the host CPU. The GPU (which is located on the same chip as the CPU) has 12 cores, able of running independent instructions each, supports a SIMD vector-width of 16 elements natively and can access the RAM directly via DMA.
- A C++14-capable compiler (e.g. GCC 6.3 or clang from the Raspbian repositories)
- The VC4C compiler to compile OpenCL C-code
- The Raspbian development files (available in the official Raspbian repository as
sudo apt-get install libraspberrypi-dev
) - The Khronos ICD Loader (available in the official Raspbian repository as
sudo apt-get install ocl-icd-opencl-dev ocl-icd-dev
) for building with ICD-support (e.g. allows to run several OpenCL implementations on one machine) - The OpenCL headers in version >= 1.2 (available in the Raspbian repositories as
sudo apt-get install opencl-headers
) - The Raspberry Pi firmware GPU-side and host-side binaries as well as the mailbox kernel module. These are already provided by default on Raspbian OS. For other Linux distributions, see here.
The following configuration options are available in CMake:
BUILD_TESTING
toggles building of test program (when configured, can be built withmake TestVC4CL
)CMAKE_BUILD_TYPE
set to Debug, Release, RelWithDebInfo, asan, tsan or usbsan to build for the given debug/release mode or with sanitizers enabledCROSS_COMPILE
toggles whether to cross-compile for the Raspberry Pi. NOTE: The Raspberry Pi cross-compiler is no longer supported!CROSS_COMPILER_PATH
sets the root path to the Raspberry Pi cross compiler (.g.arm-linux-gnueabihf-g++-6
)INCLUDE_COMPILER
whether to include the VC4C compiler. For the compiler to be actually included, the VC4C header and library needs to be found tooVC4C_HEADER_PATH
sets the path to the VC4C include headers, defaults to../VC4C/include/VC4C.h
orlib/vc4c/include/VC4C.h
VC4CC_LIBRARY
sets the path to the VC4C compiler library, defaults to../VC4C/build/libVC4CC.xxx
orlib/vc4c/build/libVC4CC.xxx
BUILD_ICD
toggles whether to build with support for the Khronos ICD loader, requires the ICD loader to be installed system-wideIMAGE_SUPPORT
toggles whether to enable the very experimental image-support
The Khronos ICD Loaders allows multiple OpenCL implementation to be used in parallel (e.g. VC4CL and pocl), but requires a bit of manual configuration:
Create a file /etc/OpenCL/vendors/VC4CL.icd
with a single line containing the absolute path to the VC4CL library.
The program clinfo can be used to test, whether the ICD loader finds the VC4CL implementation. Note: the program version in the official Raspbian repository is too old and has a bug (see fix), so it must be compiled from the github repository.
Because of the DMA-interface which has no MMU between the GPU and the RAM, code executed on the GPU can access any part of the main memory! This means, an OpenCL kernel could be used to read sensitive data or write into kernel memory!
Depending on the configuration used for the VC4CL (see Experimental Features below), the process using the VC4CL library needs to be either root
(e.g. via sudo <program>
) or be in the video
group).
The v3d_info
and v3d_profiling
tools in this project need to be run as root to give the maximum amount of information.
Since this software is still in development, some functionality might not work. For curious users or to be able to provide more information for bugs, additional debug information can be generated if desired.
To generate debug information, set the VC4CL_DEBUG
environment variable to one (or multiple, separated by commas) of the following strings:
api
enables logging OpenCL API calls, parameters and non-success errors to the standard outputcode
enables dumping OpenCL C (or IR) source code as well as generated VC4C binary code into temporary files for the compilation functionssyscall
enables logging detailed information for syscalls (e.g. mailbox) to the standard outputexecution
enables logging of detailed kernel execution information (parameters, execution cycle, return codes) to the standard outputevents
enables logging of asynchronous event processing information to the standard outputobjects
enables logging of lifetime begin/end and leaks of OpenCL API objects to the standard outputperf
enables collection and logging of V3D hardware performance counters to the standard outputsystem
enables logging high-level information about the selected system interfaces (e.g. mailbox, V3D, VCSM, etc.) to the standard outputmemory
enables logging of memory (buffer) management related information to the standard outputall
enables all above logging modes
Mostly for development, performance comparison and debugging purposes, the system interfaces used for specific system accesses can be selected via following environment variables:
VC4CL_EMULATOR
forces to use the emulator, does not actually access any specific VideoCore hardwareVC4CL_EXECUTE_MAILBOX
explicitly uses the mailbox interface to execute kernels, which has a system-wide lock on the GPU accessVC4CL_EXECUTE_REGISTER_POKING
explicitly directly writes the V3D registers, which is faster, but less compatible with other applications using the VideoCore IV hardwareVC4CL_EXECUTE_VCHI
explicitly uses the VCHI "GPUS" service to execute kernelsVC4CL_MEMORY_CMA
explicitly uses the newer VCSM CMA interface (with fall-back to the VCSM interface) to manage GPU-accessible memoryVC4CL_MEMORY_VCSM
explicitly uses the older VCSM interface (with fall-back to the VCSM CMA interface) to manage GPU-accessible memoryVC4CL_MEMORY_MAILBOX
explicitly uses the mailbox interface to manage GPU-accessible memoryVC4CL_NO_<COMPONENT>
with<COMPONENT>
eitherMAILBOX
,V3D
,VCSM
orVCHI
disables the given component completelyVC4CL_CACHE_FORCE=<VAL>
forces the buffer caching behavior to uncached (<VAL> = 0
), host-cached (<VAL> = 1
), GPU-cached (<VAL> = 2
) or host- and GPU-cached (<VAL> = 3
)