OpenMP executor#

gko::OmpExecutor is a multi-threaded host executor that uses OpenMP for parallelism.

Construction#

auto host = gko::OmpExecutor::create();

No arguments. Requires the build to be configured with an OpenMP-capable compiler.

What it parallelizes#

Most performance-critical host kernels (SpMV, vector operations, dense factorisations) include OpenMP-parallel regions. Element-wise reductions and simple array operations are also parallelized via the unified kernel infrastructure that backs both OMP and the GPU backends.

Memory model#

Plain host memory via new[]/delete[]. Memory is shared with ReferenceExecutor — the two are memory-accessible to each other.

Using as a master executor#

OmpExecutor is the typical master (host-side) executor when constructing a GPU executor:

auto host = gko::OmpExecutor::create();
auto gpu  = gko::CudaExecutor::create(0, host);

Choosing OMP over Reference for the master means any host-side fallback or post-processing work runs in parallel.