libraries-openly-fused/fusedkernellibrary/fkl-using-operations

Using FKL Operations and IOps (consumer side)

fkl-implementing-operations covers how to AUTHOR an Operation struct. This skill covers the other half: how to create Instantiable Operations (IOps) on the host, compose them into a pipeline, and pass them to an Executor.

(Note: For the device-side implementation of how a DPP actually invokes these operations internally, see the `fkl-implementing-data-parallel-patterns` skill).

Creating the operations and executing them in order (Host Side)

Before launching the DPP, the host code must build the IOps. The executor handles fusing the sequence automatically:

// 1. Build the input Read IOp
auto input = fk::PerThreadRead<fk::ND::_2D, float>::build(in_ptr);
 
// 2. Build the compute IOp (the epilogue)
auto compute_iop = fk::Add<float>::build(5.0f);
 
// 3. Build the destination write IOp
auto write_iop = fk::PerThreadWrite<fk::ND::_2D, float>::build(out_ptr);
 
// 4. Pass them sequentially to the Executor, providing the DPP as a template parameter
fk::executeOperations<MyDPP>(stream, input, compute_iop, write_iop);

See also

• fkl-implementing-operations — authoring the op structs.
• fkl-implementing-data-parallel-patterns — the device-side DPP exec(...) contract where these IOps are actually invoked.
• fkl-fusion-techniques — understanding fusion techniques for discussion.