DPC++ executor#

gko::DpcppExecutor runs Ginkgo kernels on SYCL-capable accelerators using Intel’s DPC++ implementation. The most common target is Intel GPUs, but SYCL also covers CPUs and (vendor-permitting) NVIDIA / AMD devices through alternative SYCL implementations.

Construction#

auto host = gko::OmpExecutor::create();
auto gpu  = gko::DpcppExecutor::create(0, host);   // device id 0

A device-type filter can target a specific kind of SYCL device (e.g., "gpu", "cpu", "all"):

auto gpu  = gko::DpcppExecutor::create(0, host, "gpu");

Build requirements#

Intel oneAPI DPC++ or any SYCL-2020-conformant compiler.
For Intel GPUs: appropriate Level Zero / OpenCL runtime.

Memory model#

Device memory via SYCL Unified Shared Memory (USM). Cross-executor copies use SYCL queue operations and the executor’s master for orchestration.

Discovering devices#

auto count = gko::DpcppExecutor::get_num_devices("gpu");