ESIMD-Based Kernel Templates

The ESIMD kernel templates are based on Explicit SIMD SYCL extension of oneAPI DPC++ Compiler. This technology only supports Intel GPU devices.

Kernel Parameter Interpretation

The kernel_param parameters map to Intel GPU hardware as follows:

• data_per_workitem: The number of data elements processed by a single hardware thread (work-item in ESIMD terminology), which issues explicit vector operations. The hardware thread processes these elements using SIMD instructions with an implementation-defined vector length.

• workgroup_size: The number of hardware threads in a work-group. Each hardware thread executes scalar and SIMD operations independently.

Available Templates

These templates are available in the oneapi::dpl::experimental::kt::gpu::esimd namespace. The following are implemented:

• radix_sort

• radix_sort_by_key