Thrust: Reduce, Scan und Sort

Reductions, scans, and sorts are tricky to write fast by hand. Thrust gives you tuned versions through a single function call. 🎁

thrust::reduce combines every element into a single result, like summing an array, all in parallel under the hood.

Reduce defaults to addition, but you can pass an init value and a binary op to compute a product, max, or anything associative.

A scan, or prefix sum, outputs the running total at each position. It is the backbone of compaction, sorting, and stream allocation.

inclusive_scan includes the current element in its sum; exclusive_scan does not. Picking the right one avoids an off-by-one bug.

A prefix sum looks sequential, yet Thrust runs it in parallel with a clever tree algorithm you never have to write yourself.

thrust::sort orders a device_vector in place using a fast GPU radix or merge sort, far quicker than a CPU sort on big data.

sort_by_key sorts one array and reorders a second values array to match, perfect for keeping records aligned with their keys.

Real pipelines chain these: transform then reduce, or sort then scan. Each step is one tuned call, so you focus on the logic.

transform_reduce maps and sums in one pass, computing things like a dot product or sum of squares without a temporary array.

These primitives are heavily optimized by NVIDIA. Reaching for them first usually beats a custom kernel and saves hours of work.

Recall what a prefix sum produces.

You collapsed data with reduce, built running totals with scan, ordered arrays with sort, and fused steps with transform_reduce. 🏁