CUDA Thrust: reduce_by_key on only some values in an array, based off values in a “key” array

Question

Let\'s say I have two device_vector arrays, d_keys and d_data.

If d_data is, for example, a flattened 2D 3x5 array

Answer 1

Based on the additional comment that instead of 3 rows there are thousands of rows, we can write a transform functor that sums an entire row. Based on the fact that there are thousands of rows, this should keep the machine pretty busy:

#include 
#include 
#include 
#include 
#include 
#include 

#define ROW   20
#define COL   10

__device__ int *vals;
__device__ int *keys;

struct test_functor
{
  const int a;

  test_functor(int _a) : a(_a) {}

  __device__
  int operator()(int& x, int& y ) {
    int temp = 0;
    for (int i = 0; i h_vals(ROW*COL);
  thrust::host_vector h_keys(COL);
  thrust::sequence(h_vals.begin(), h_vals.end());
  thrust::fill(h_keys.begin(), h_keys.end(), 1);
  h_keys[0] = 0;
  thrust::device_vector d_vals = h_vals;
  thrust::device_vector d_keys = h_keys;
  thrust::device_vector d_sums(ROW);
  thrust::fill(d_sums.begin(), d_sums.end(), 0);
  s_vals = thrust::raw_pointer_cast(&d_vals[0]);
  s_keys = thrust::raw_pointer_cast(&d_keys[0]);
  cudaMemcpyToSymbol(vals, &s_vals, sizeof(int *));
  cudaMemcpyToSymbol(keys, &s_keys, sizeof(int *));
  thrust::device_vector d_idx(ROW);
  thrust::sequence(d_idx.begin(), d_idx.end());
  thrust::transform(d_sums.begin(), d_sums.end(), d_idx.begin(),  d_sums.begin(), test_functor(COL));
  thrust::host_vector h_sums = d_sums;
  std::cout << "Results :" << std::endl;
  for (unsigned i = 0; i