Let us say, we have a thrust device vector of size 10^16 and another vector of size 10^8 containing some indices (not necessarily sorted). We want to sum all the elements of the first vector if it's index is in the second vector.
A naive approach to do so would be to use transform_reduce of thrust. However, I believe it'll involve iterating through all the elements of the first vector.
Is there an efficient way?
Following talonmies suggestion of permutation iterator, here is the main body of the code that implements reduction of a subset of a vector. I purposefully, chose small vector sizes to explain the idea. For reasonable sizes, it is faster than using inner_product
thrust::device_vector<double> vals(6); vals[0] = 2.0; vals[1] = 1.5; vals[2] = -1.2; vals[3] = 1.1; vals[4] = -4.3; vals[5] = 0.8; thrust::device_vector<int> indices(3); indices[0] = 1; indices[1] = 3; indices[2] = 5; thrust::device_vector<double> masks(6); for (auto elm:indices) masks[elm]=1.0; typedef thrust::device_vector<double>::iterator ValIterator; typedef thrust::device_vector<int>::iterator IndIterator; thrust::permutation_iterator<ValIterator, IndIterator> iter_begin(vals.begin(), indices.begin()); thrust::permutation_iterator<ValIterator, IndIterator> iter_end(vals.end(), indices.end()); double sum_reduce = thrust::reduce(iter_begin, iter_end); std::cout << "sum permutation iterator: " << sum_reduce << std::endl; double sum_inner_product = thrust::inner_product(vals.begin(), vals.end(), masks.begin(), 0.0); std::cout << "sum inner product: " << sum_inner_product << std::endl;