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

Here is some sample code that does something like what you are after, using the approach I outlined in my comment below your question. In fact we want to use 4-tuples, to pick up your key value. Reproducing the suitably modified comment here:

You could make a zip iterator that zips your 3 rows together plus the key "row" and passes a 4-tuple to a special functor. Your special functor would then do a reduction on the array of 3-tuples (using the key also) and return a result that is a 4-tuple. The thrust dot product example may give you some ideas.

This is one possible approach:

#include <thrust/host_vector.h>
#include <thrust/iterator/zip_iterator.h>
#include <thrust/sequence.h>
#include <thrust/fill.h>
#include <thrust/tuple.h>

#define N 30  // make this evenly divisible by 3 for this example

typedef thrust::tuple<int, int, int, int>  tpl4int;
typedef thrust::host_vector<int>::iterator intiter;
typedef thrust::tuple<intiter, intiter, intiter, intiter>  tpl4intiter;
typedef thrust::zip_iterator<tpl4intiter>  int4zip;



struct r3key_unary_op : public thrust::unary_function<tpl4int, tpl4int>
{
  __host__ __device__
  tpl4int operator()(const tpl4int& x) const
  {
    tpl4int result;
    thrust::get<0>(result) = x.get<0>()*x.get<3>();
    thrust::get<1>(result) = x.get<1>()*x.get<3>();
    thrust::get<2>(result) = x.get<2>()*x.get<3>();
    thrust::get<3>(result) = 1;
    return result;
   }
};

struct r3key_binary_op : public thrust::binary_function<tpl4int, tpl4int, tpl4int>
{
  __host__ __device__
  tpl4int operator()(const tpl4int& x, const tpl4int& y) const
  {
    tpl4int result;
    thrust::get<0>(result) = x.get<0>()*x.get<3>() + y.get<0>()*y.get<3>();
    thrust::get<1>(result) = x.get<1>()*x.get<3>() + y.get<1>()*y.get<3>();
    thrust::get<2>(result) = x.get<2>()*x.get<3>() + y.get<2>()*y.get<3>();
    thrust::get<3>(result) = 1;
    return result;
  }
};


int main() {

  thrust::host_vector<int> A(N);  // values, in 3 "rows" flattened
  thrust::sequence(A.begin(), A.end());
  thrust::host_vector<int> K(N/3);   // keys in one row
  thrust::fill(K.begin(), K.end(), 1);  // set some keys to 1
  K[9] = 0;  // set some keys to zero

  int4zip first = thrust::make_zip_iterator(thrust::make_tuple(A.begin(), A.begin() + N/3, A.begin() + 2*N/3, K.begin()));
  int4zip  last = thrust::make_zip_iterator(thrust::make_tuple(A.begin() + N/3, A.begin() + 2*N/3, A.end(), K.end()));
  r3key_unary_op my_unary_op;
  r3key_binary_op my_binary_op;
  tpl4int init = my_unary_op(*first);
  // init = thrust::make_tuple((int) 0, (int) 0, (int) 0, (int) 0);
  tpl4int result = thrust::transform_reduce(first, last, my_unary_op, init, my_binary_op);
  std::cout << "row 0 = " << result.get<0>() << std::endl;
  std::cout << "row 1 = " << result.get<1>() << std::endl;
  std::cout << "row 2 = " << result.get<2>() << std::endl;
  return 0;

}

Notes:

1. This is just using host_vector. Extending it to work with device_vector, or templatizing it to work with something other than int should be straightforward.
2. For completeness, I am using the unary functor to provide an init value other than zero for the sum reduction of each row. You might want to change the init value to zero (a 4-tuple of zeros).

Answer 2

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 <iostream>
#include <thrust/host_vector.h>
#include <thrust/device_vector.h>
#include <thrust/transform.h>
#include <thrust/sequence.h>
#include <thrust/fill.h>

#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<a; i++)
      temp += vals[i + (y*a)] * keys[i];
    return temp;
    }
};

int main(){
  int *s_vals, *s_keys;
  thrust::host_vector<int> h_vals(ROW*COL);
  thrust::host_vector<int> 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<int> d_vals = h_vals;
  thrust::device_vector<int> d_keys = h_keys;
  thrust::device_vector<int> 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<int> 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<int> h_sums = d_sums;
  std::cout << "Results :" << std::endl;
  for (unsigned i = 0; i<ROW; i++)
    std::cout<<"h_sums["<<i<<"] = " << h_sums[i] << std::endl;
  return 0;
}

This approach has the drawback that in general accesses to the vals array will not be coalesced. However for a few thousand rows the cache may offer significant relief. We can fix this problem by re-ordering the data to be stored in column-major form in the flattened array, and change our indexing method in the loop in the functor to be like this:

for (int i=0; i<a; i++)
  temp += vals[(i*ROW)+y]*keys[i];

If preferred, you can pass ROW as an additional parameter to the functor.