CLINFO(1) General Commands Manual CLINFO(1)

clinfo - show OpenCL platforms and devices

clinfo [options ...]

clinfo prints all available information about all OpenCL platforms available on the system and the devices they expose.

clinfo accepts the following options:

try to retrieve all properties, even those not officially supported (e.g. because they require specific extensions), but only show them if the property could be retrieved successfully; see also the LIMITATIONS section below; note that even though this may reveal hidden properties, there is no guarantee that the returned values are meaningful, nor that the corresponding feature is actually available at all;
like -a, but also show errors;
produce human-friendly output; this is the default (except as noted below);
produce machine-friendly output; this is the default if clinfo is invoked with a name that contains the string “raw”;
outputs the raw data (cf. the --raw option) in JSON format; support for this option is experimental, as the representation of some of the values is not finalized;
shows also offline devices for platforms that expose this feature;
list platforms and devices by name, with no (other) properties;
only show properties for the specified device in the specified platform;
only show properties whose symbolic name matches (contains as a substring) the given property-name; the name is normalized as upper-case and with minus sign (-) replaced by underscore signs (_); when this flag is specified, raw mode is forced;
show usage;
show program version.

OpenCL 1.1, OpenCL 1.2, OpenCL 2.0, OpenCL 2.1, OpenCL 2.2, OpenCL 3.0.

Supported OpenCL extensions:

for the UUID, LUID and node mask of the device;
for the extended platform, device, extension and IL versioned properties backported from OpenCL 3.0 to previous OpenCL versions;
for information about support for half-precision and double-precision floating-point data types;
for information about the base address and pitch alignment requirements of buffers to be used as base for 2D images;
for information about the supported IL (Intermediate Language) representations;
for information about the supported SPIR (Standard Portable Intermediate Representation) versions;
for the suffix of vendor extensions functions;
for the maximum number of named sub-group barriers;
for the terminate capabilities for the device;
for device fission support in OpenCL 1.1 devices;
for the atomic counter extension;
for the version of the C++ for OpenCL language supported by the device compiler;
for AMD-specific device attributes;
to show the maximum number of keys supported by the platform;
to show offline devices exposed by the platform, if requested (see --offline option);
to show the number and IDs of available P2P devices;
for Shared Virtual Memory (SVM) capabilities in OpenCL 1.2 devices;
to show the (potentially sparse) list of the core IDs that the device may return;
to show the (potentially sparse) list of available job slots for command submission;
for NVIDIA-specific device attributes;
for the Intel extension allowing CPU devices to run kernels as part of the current host thread;
for the version of the Intel Motion Estimation accelerator version;
for the version and supported features of Intel's device-side AVC Motion;
for the maximum dimensions of planar YUV images;
for simultaneous CL/GL/DirectX context sharing (only partial support);
to enumerate allowed sub-group sizes;
for the Altera extension to query the core temperature of the device;
for the QUALCOMM extension to query page size and required padding in external memory allocation.

Some information is duplicated when available from multiple sources. Examples:

  • supported device partition types and domains as obtained using the cl_ext_device_fission extension typically match the ones obtained using the core OpenCL 1.2 device partition feature;
  • the preferred work-group size multiple matches the NVIDIA warp size (on NVIDIA devices) or the AMD wavefront width (on AMD devices).

Some floating-point configuration flags may only be meaningful for specific precisions and/or specific OpenCL versions. For example, CL_FP_CORRECTLY_ROUNDED_DIVIDE_SQRT is only relevant for single precision in OpenCL 1.2 devices.

The implementation-defined behavior for NULL platform or context properties is tested for the following API calls:

by trying to show the platform name;
by trying to enumerate devices; the corresponding platform (if any) is then detected by querying the device platform of the first device;
by trying to create a context from a device from the previous list (if any), and a context from a device from a different platform;
by trying to create contexts for each device type (except DEFAULT).

Support for OpenCL 2.x properties is not fully tested.

Support for cl_khr_subgroup_named_barrier is experimental due to missing definitions in the official OpenCL headers.

Raw (machine-parsable) output is considered experimental, the output format might still undergo changes.

The properties of the ICD loader will also be queried if the clGetICDLoaderInfoOCLICD extension function is found.

Support for the properties exposed by cl_amd_copy_buffer_p2p is experimental.

Support for some (documented and undocumented) properties exposed by cl_amd_device_attribute_query is experimental (see also LIMITATIONS).

Support for the interop lists exposed by cl_intel_simultaneous_sharing is experimental.

The highest OpenCL version supported by the ICD loader is detected with some trivial heuristics (symbols found); a notice is output if this is lower than the highest platform OpenCL version, or if the detected version doesn't match the one declared by the ICD loader itself.

OpenCL provides no explicit mean to detect the supported version of any extension exposed by a device, which makes it impossible to determine a priori if it will be possible to successfully query a device about a specific property. Additionally, the actual size and meaning of some properties are not officially declared anywhere.

Most notably, this affects extensions such as cl_amd_device_attribute_query, cl_nv_device_attribute_query and cl_arm_core_id. Heuristics based on standard version support are partially used in the code to determine which version may be supported.

Properties which are known to be affected by these limitations include:

documented in v3 of the cl_amd_device_attribute_query extension specification as being the global free memory in KBytes, without any explanation given on why there are two values, although in the source code of the ROCm stack the second value is documented as being the largest free block;
documented in v3 of the cl_amd_device_attribute_query extension specification, but not reported by drivers supporting other v3 properties. This has now been enabled for drivers assumed to support v4 of the same extension;
exposed by the cl_khr_terminate_context has changed value between OpenCL 1.x and 2.x, and it's allegedly a bitfield, whose values are however not defined anywhere.

Please report any issues on the project tracker on GitHub.

If multiple OpenCL platforms using shared LLVM libraries are present in the system, clinfo (and other OpenCL application) may crash with errors to the tune of

: CommandLine Error: Option '(some option name)' registered more than once!
LLVM ERROR: inconsistency in registered CommandLine options

or similar. This is not an issue in clinfo, or in any OpenCL platform or application, but it is due to the way LLVM handles its own command-line options parsing. The issue has been reported upstream as issue #30587. See the next point for possible workarounds and assistance in identifying the conflicting platforms.

Faulty OpenCL platforms may cause segmentation faults in clinfo during the information gathering phase, sometimes even before any output is shown. There is very little clinfo can do to avoid this. If you see this happening, try disabling all platforms and then re-enabling them one by one until you experience the crash again. Chances are the last platform you enabled is defective in some way (either by being incompatible with other platforms or by missing necessary components and not handling their absence gracefully).

To selectively enable/disable platforms, one way is to move or rename the *.icd files present in /etc/OpenCL/vendors/ and then restoring them one by one. When using the free-software ocl-icd OpenCL library, a similar effect can be achieved by setting the OPENCL_VENDOR_PATH or OCL_ICD_VENDORS environment variables, as documented in libOpenCL(7). Other implementations of libOpenCL are known to support OPENCL_VENDOR_PATH too.

find /etc/OpenCL/vendors/ -name '*.icd' | while read OPENCL_VENDOR_PATH ; do clinfo -l > /dev/null ; echo "$? ${OPENCL_VENDOR_PATH}" ; done

This one liner will run clinfo -l for each platform individually (hiding the normal output), and report the .icd path prefixed by 0 for successful runs, and a non-zero value for faulty platforms.

2021-02-21 clinfo