OpenCL devices¶
OpenCL devices can be easily discovered and selected with the help of the
xopto.cl.clinfo module.
To list all available OpenCL devices use the
info() function. The human-readable output contains
information on the relevant properties of the installed OpenCL devices.
from xopto.cl import clinfo
clinfo.info()
An example output of the above command on a system with a single integrated Intel GPU:
===============================================================
Platform name: Intel(R) OpenCL HD Graphics
Platform profile: FULL_PROFILE
Platform vendor: Intel(R) Corporation
Platform version: OpenCL 2.1
---------------------------------------------------------------
Device name: Intel(R) Gen9 HD Graphics NEO
Device type: ALL | GPU
Device available: Yes
Device maximum clock frequency: 1150 MHz
Device default address space: 64 bit
Device maximum number of constant kernel arguments: 8
Device maximum kernel argument size: 1024 B
Device global memory: 12559 MB
Device global memory cache: 512 kB
Device maximum global memory allocation size: 4095 MB
Device local memory: 64 kB
Device constant memory: 4194296.0 kB
Device max clock speed: 1150 MHz
Device compute units: 24
Device maximum work group size: 256
Maximum number of work-item dimensions (global or local): 3
Maximum number of work-items in each dimension of the work group: [256, 256, 256]
Maximum number of image-objects that can be passed to a kernel: 128
Device profiling timer resolution: 83 ns
Device supported OpenCL extensions:
cl_khr_byte_addressable_store
cl_khr_fp16
cl_khr_global_int32_base_atomics
cl_khr_global_int32_extended_atomics
cl_khr_icd
cl_khr_local_int32_base_atomics
cl_khr_local_int32_extended_atomics
cl_intel_subgroups
cl_intel_required_subgroup_size
cl_intel_subgroups_short
cl_khr_spir
cl_intel_accelerator
cl_intel_driver_diagnostics
cl_khr_priority_hints
cl_khr_throttle_hints
cl_khr_create_command_queue
cl_intel_subgroups_char
cl_intel_subgroups_long
cl_khr_fp64
cl_khr_subgroups
cl_khr_il_program
cl_intel_spirv_device_side_avc_motion_estimation
cl_intel_spirv_media_block_io
cl_intel_spirv_subgroups
cl_khr_spirv_no_integer_wrap_decoration
cl_intel_unified_shared_memory_preview
cl_khr_mipmap_image
cl_khr_mipmap_image_writes
cl_intel_planar_yuv
cl_intel_packed_yuv
cl_intel_motion_estimation
cl_intel_device_side_avc_motion_estimation
cl_intel_advanced_motion_estimation
cl_khr_int64_base_atomics
cl_khr_int64_extended_atomics
cl_khr_image2d_from_buffer
cl_khr_depth_images
cl_intel_media_block_io
cl_khr_3d_image_writes
cl_intel_va_api_media_sharing
To get a list of all available OpenCL GPU devices use the
gpus() function. Two optional input arguments
allow filtering of the OpenCL devices by a platform name
(platform) and / or device name (device). The matching of the
given device and / or platform name is done in a case insensitive way. The
target OpenCL device and / or platform name must contain the given string.
In the following example we fetch a list of installed OpenCL GPU devices:
from xopto.cl import clinfo
gpus = clinfo.gpus()
In this example, we fetch only the Nvidia GPU devices:
from xopto.cl import clinfo
nvidia_gpus = clinfo.gpus(platform='nvidia')
Note that the gpus() returns None if no
devices are found that match the search criteria.
We can use the xopto.cl.clinfo.info() function to print out
human-readable information on the found OpenCL devices.
from xopto.cl import clinfo
nvidia_gpus = clinfo.gpus(platform='nvidia')
clinfo.info(devices=nvidia_gpus)
Alternatively, the gpu() function can be used to
retrieve the n-th OpenCL device that matches the search criteria. The following
example returns the second installed Nvidia GPU device. Note that the parameter
index that is used to select a particular device is zero-based
(0 - 1 st device, 1 - 2 nd device, …).
from xopto.cl import clinfo
nvidia_gpus = clinfo.gpu(platform='nvidia', index=1)
Use the cpus() and cpu()
functions to find or select CPU devices. The two functions take the same
parameters as the gpus() and
gpu() functions.
To retrieve OpenCL device information into a dict instead of
displaying the information in the console, use the
device_info() function.
from xopto.cl import clinfo
nvidia_gpu = clinfo.gpu(platform='nvidia')
gpu_info = clinfo.device_info(nvidia_gpu)
print(gpu_info['name'])
print(gpu_info['platform'])
To check if a device supports one or more OpenCL extensions, use the
device_extension() function that returns True if
the extension is supported and False if not (returns a list of boolean values
if multiple extensions are queried). See the
full list
of extensions supported by the OpenCL 1.2 programming language.
from xopto.cl import clinfo
nvidia_gpu = clinfo.gpu(platform='nvidia')
supports_double = clinfo.device_extension(nvidia_gpu, 'cl_khr_fp64')
supports_64bit_atomic_operations = clinfo.device_extension(
nvidia_gpu, 'cl_khr_int64_base_atomics')