Documentation

SplitSpace o w i elems_per_thread.

Computes how to divide array elements to threads in a kernel. Returns the number of elements in the chunk that the current thread should take.

w is the length of the outer dimension in the array. i is the current thread index. Each thread takes at most elems_per_thread elements.

If the order o is SplitContiguous, thread with index i should receive elements i*elems_per_tread, i*elems_per_thread + 1, ..., i*elems_per_thread + (elems_per_thread-1).

If the order o is SplitStrided stride, the thread will receive elements i, i+stride, i+2*stride, ..., i+(elems_per_thread-1)*stride.

Combine cspace ts aspace body will combine values from threads to a single (multidimensional) array. If we define (is, ws) = unzip cspace, then ws is defined the same accross all threads. The cspace defines the shape of the resulting array, and the identifiers used to identify each individual element. Only threads for which all ((i,w) -> i < w) aspace is true will provide a value (of type ts), which is generated by body.

The result of a combine is always stored in local memory (OpenCL terminology)

The same thread may be assigned to multiple elements of Combine, if the size of the CombineSpace exceeds the group size.

GroupReduce w lam input (with (nes, arrs) = unzip input), will perform a reduction of the arrays arrs using the associative reduction operator lam and the neutral elements nes.

The arrays arrs must all have outer dimension w, which must not be larger than the group size.

Currently a GroupReduce consumes the input arrays, as it uses them for scratch space to store temporary results

All threads in a group must participate in a GroupReduce (due to barriers)

The length of the arrays w can be smaller than the number of elements in a group (neutral element will be filled in), but w can never be larger than the group size.

Same restrictions as with GroupReduce.

GroupGenReduce length destarrays op bucket values arrays

HACK: Semantically identity, but inserts a barrier afterwards. This reflects a weakness in our kernel representation.