Question

I want to implement a groupby reduction in OpenCL. For example, the input

a1 a2 a3 b1 b2 c3 c4

would produce

a6 b3 c7

The C pseudocode looks like this:

int data[n][2], result[n][2], result_count = -1, 
    added = 0, group = data[0][0];
for (int i = 0; i < n; i++) { 
  if (group == data[i][0]) {
    added += data[i][1];
  } else {
    result[++result_count][0] = group;
    result[result_count][1] = added;
    group = data[i][0];
    added = 0;
  }
} 
return result, result_count;

The only standard algorithm I know which goes in this direction is parallel reduction; however, it reduces to one number and not to a buffer of added values by group. I am not sure if parallel reduction could work with a dynamic result buffer (e.g. in local memory) and still be efficient in terms of performance.

Answer 1

A solution by Hashing

Phase 1) A hashing scheme can be used to hash the group value to a location, then an atomic add can sum the contents of the second value.

Phase 2) A prefix sum scan algorithm pass over the hash table to compact it.

Phase 3) Optionally sort the results

A solution by Sorting

Phase 1) Sort the data on the group value

Phase 2) Use a reduction to sum each group

Phase 3) A prefix sum scan to compact the sums