Prototype for mapping abstract compute kernels to CPUs. We introduce a
new data type WorkGroup that holds data for a whole workgroup. This
is similar to a SIMD vector, but is more flexible in case we need to
represent types that cannot be handle by SIMD.jl.
When translating a kernel, we would:
- Choose a fixed workgroup size
N(maybe 16?) - Wrap each literal with a call to
WorkGroup{N} - Translate each function call to calling
map
This package is a demonstration. The WorkGroups module implements
this idea (roughly), the test cases show a simple scalar kernel
(axpy!) and how it would look like after translation (axpyN!).