Device Selection
Offloading code to a
device (such as a CPU, GPU, or FPGA) is available for both DPC++ and OpenMP* applications.
DPC++ Device Selection in the Host Code
Host code can explicitly select a device type. To do this, select a queue and initialize its device with one of the following:
- default_selector
- cpu_selector
- gpu_selector
- accelerator_selector
If
default_selector
is used, the kernel runs based on a heuristic that chooses from available compute devices (all, or a subset based on the value of the
SYCL_DEVICE_FILTER
environment variable).
If a specific device type (such as
cpu_selector
or
gpu_selector
) is used, then it is expected that the specified device type is available in the platform or included in the filter specified by
SYCL_DEVICE_FILTER
. If such a device is not available, then the runtime system throws an exception indicating that the requested device is not available. This error can be thrown in the situation where an ahead-of-time (AOT) compiled binary is run in a platform that does not contain the specified device type.
While
applications can run on any supported target hardware, tuning is required to derive the best performance advantage on a given target architecture. For example, code tuned for a CPU likely will not run as fast on a GPU accelerator without modification.
DPC++
SYCL_DEVICE_FILTER
is a complex environment variable that allows you to limit the runtimes, compute device types, and compute device IDs that may be used by the DPC++ runtime to a subset of all available combinations. The compute device IDs correspond to those returned by the SYCL API,
clinfo
, or
sycl-ls
(with the numbering starting at 0). They have no relation to whether the device with that ID is of a certain type or supports a specific runtime. Using a programmatic special selector (like
gpu_selector
) to request a filtered out device will cause an exception to be thrown. Refer to the environment variable description in GitHub for details on use and example values:
https://github.com/intel/llvm/blob/sycl/sycl/doc/EnvironmentVariables.md.
Additional information about device selection is available from the
DPC++ Language Guide and API Reference.
OpenMP* Device Query and Selection in the Host Code
OpenMP provided a set of APIs for programmers to query and set device for running code on the device. Host code can explicitly select and set a device num. For each offloading region, a programmer can also use a
device
clause to specify the target device that is to be used for executing the offloading region.
- int omp_get_num_procs (void)routine returns the number of processors available to the device
- void omp_set_default_device(int device_num)routine controls the default target device
- int omp_get_default_device(void)routine returns the default target device
- int omp_get_num_devices(void)routine returns the number of non-host devices available for offloading code or data.
- int omp_get_device_num(void)routine returns the device number of the device on which the calling thread is executing.
- int omp_is_initial_device(int device_num)routine returnstrueif the current task is executing on the host device; otherwise, it returnsfalse.
- int omp_get_initial_device(void)routine returns a device number that represents the host device.
A programmer can use the environment variable
LIBOMPTARGET_DEVICETYPE = [ CPU | GPU ]
to perform a device type selection. If a specific device type such as CPU or GPU is specified, then it is expected that the specified device type is available in the platform. If such a device is not available, then the runtime system throws an error that the requested device type is not available if the environment variable
OMP_TARGET_OFFLOAD=mandatory
, otherwise, the execution will have a fallback execution on its initial device. Additional information about device selection is available from the OpenMP 5.1 specification. Details about environment variables are available from GitHub:
https://github.com/intel/llvm/blob/sycl/sycl/doc/EnvironmentVariables.md.