Replicate a vector multiple times using CUDA Thrust

Question

I am trying to solve a problem using CUDA Thrust.

I have a host array with 3 elements. Is it possible, using Thrust, to create a device array of 384 elements in which the 3 elements in my host array is repeated 128 times (128 x 3 = 384)?

Generally speaking, starting from an array of 3 elements, how can I use Thrust to generate a device array of size X, where X = Y x 3, i.e. Y is the number of repetitions?

Answer 1

One possible approach:

1. create a device vector of appropriate size
2. create 3 strided ranges, one for each of the element positions {1, 2, 3} in the final output (device) vector
3. use thrust::fill to fill each of the 3 strided ranges with the appropriate (host vector) element {1, 2, 3}

This code is a trivial modification of the strided range example to demonstrate. You can change the REPS define to 128 to see the full expansion to 384 output elements:

#include <thrust/iterator/counting_iterator.h>
#include <thrust/iterator/transform_iterator.h>
#include <thrust/iterator/permutation_iterator.h>
#include <thrust/functional.h>

#include <thrust/fill.h>
#include <thrust/device_vector.h>
#include <thrust/host_vector.h>

// for printing
#include <thrust/copy.h>
#include <ostream>


#define STRIDE 3
#define REPS  15  // change to 128 if you like
#define DSIZE (STRIDE*REPS)

// this example illustrates how to make strided access to a range of values
// examples:
//   strided_range([0, 1, 2, 3, 4, 5, 6], 1) -> [0, 1, 2, 3, 4, 5, 6]
//   strided_range([0, 1, 2, 3, 4, 5, 6], 2) -> [0, 2, 4, 6]
//   strided_range([0, 1, 2, 3, 4, 5, 6], 3) -> [0, 3, 6]
//   ...

template <typename Iterator>
class strided_range
{
    public:

    typedef typename thrust::iterator_difference<Iterator>::type difference_type;

    struct stride_functor : public thrust::unary_function<difference_type,difference_type>
    {
        difference_type stride;

        stride_functor(difference_type stride)
            : stride(stride) {}

        __host__ __device__
        difference_type operator()(const difference_type& i) const
        {
            return stride * i;
        }
    };

    typedef typename thrust::counting_iterator<difference_type>                   CountingIterator;
    typedef typename thrust::transform_iterator<stride_functor, CountingIterator> TransformIterator;
    typedef typename thrust::permutation_iterator<Iterator,TransformIterator>     PermutationIterator;

    // type of the strided_range iterator
    typedef PermutationIterator iterator;

    // construct strided_range for the range [first,last)
    strided_range(Iterator first, Iterator last, difference_type stride)
        : first(first), last(last), stride(stride) {}

    iterator begin(void) const
    {
        return PermutationIterator(first, TransformIterator(CountingIterator(0), stride_functor(stride)));
    }

    iterator end(void) const
    {
        return begin() + ((last - first) + (stride - 1)) / stride;
    }

    protected:
    Iterator first;
    Iterator last;
    difference_type stride;
};

int main(void)
{
    thrust::host_vector<int> h_data(STRIDE);
    h_data[0] = 1;
    h_data[1] = 2;
    h_data[2] = 3;

    thrust::device_vector<int> data(DSIZE);

    typedef thrust::device_vector<int>::iterator Iterator;
    strided_range<Iterator> pos1(data.begin(), data.end(), STRIDE);
    strided_range<Iterator> pos2(data.begin()+1, data.end(), STRIDE);
    strided_range<Iterator> pos3(data.begin()+2, data.end(), STRIDE);

    thrust::fill(pos1.begin(), pos1.end(), h_data[0]);
    thrust::fill(pos2.begin(), pos2.end(), h_data[1]);
    thrust::fill(pos3.begin(), pos3.end(), h_data[2]);


    // print the generated data
    std::cout << "data: ";
    thrust::copy(data.begin(), data.end(), std::ostream_iterator<int>(std::cout, " "));  std::cout << std::endl;

    return 0;
}

Answer 2

Robert Crovella has already answered this question using strided ranges. He has also pointed out the possibility of using the expand operator.

Below, I'm providing a worked example using the expand operator. Opposite to the use of strided ranges, it avoids the need of for loops.

#include <thrust/device_vector.h>
#include <thrust/gather.h>
#include <thrust/sequence.h>
#include <stdio.h>

using namespace thrust::placeholders;

/*************************************/
/* CONVERT LINEAR INDEX TO ROW INDEX */
/*************************************/
template <typename T>
struct linear_index_to_row_index : public thrust::unary_function<T,T> {

    T Ncols; // --- Number of columns

    __host__ __device__ linear_index_to_row_index(T Ncols) : Ncols(Ncols) {}

    __host__ __device__ T operator()(T i) { return i / Ncols; }
};

/*******************/
/* EXPAND OPERATOR */
/*******************/
template <typename InputIterator1, typename InputIterator2, typename OutputIterator>
OutputIterator expand(InputIterator1 first1,
                      InputIterator1 last1,
                      InputIterator2 first2,
                      OutputIterator output)
{
    typedef typename thrust::iterator_difference<InputIterator1>::type difference_type;

    difference_type input_size  = thrust::distance(first1, last1);
    difference_type output_size = thrust::reduce(first1, last1);

    // scan the counts to obtain output offsets for each input element
    thrust::device_vector<difference_type> output_offsets(input_size, 0);
    thrust::exclusive_scan(first1, last1, output_offsets.begin()); 

    // scatter the nonzero counts into their corresponding output positions
    thrust::device_vector<difference_type> output_indices(output_size, 0);
    thrust::scatter_if(thrust::counting_iterator<difference_type>(0), thrust::counting_iterator<difference_type>(input_size),
                       output_offsets.begin(), first1, output_indices.begin());

    // compute max-scan over the output indices, filling in the holes
    thrust::inclusive_scan(output_indices.begin(), output_indices.end(), output_indices.begin(), thrust::maximum<difference_type>());

    // gather input values according to index array (output = first2[output_indices])
    OutputIterator output_end = output; thrust::advance(output_end, output_size);
    thrust::gather(output_indices.begin(), output_indices.end(), first2, output);

    // return output + output_size
    thrust::advance(output, output_size);

    return output;
}

/**************************/
/* STRIDED RANGE OPERATOR */
/**************************/
template <typename Iterator>
class strided_range
{
    public:

    typedef typename thrust::iterator_difference<Iterator>::type difference_type;

    struct stride_functor : public thrust::unary_function<difference_type,difference_type>
    {
        difference_type stride;

        stride_functor(difference_type stride)
            : stride(stride) {}

        __host__ __device__
        difference_type operator()(const difference_type& i) const
        {
            return stride * i;
        }
    };

    typedef typename thrust::counting_iterator<difference_type>                   CountingIterator;
    typedef typename thrust::transform_iterator<stride_functor, CountingIterator> TransformIterator;
    typedef typename thrust::permutation_iterator<Iterator,TransformIterator>     PermutationIterator;

    // type of the strided_range iterator
    typedef PermutationIterator iterator;

    // construct strided_range for the range [first,last)
    strided_range(Iterator first, Iterator last, difference_type stride)
        : first(first), last(last), stride(stride) {}

    iterator begin(void) const
    {
        return PermutationIterator(first, TransformIterator(CountingIterator(0), stride_functor(stride)));
    }

    iterator end(void) const
    {
        return begin() + ((last - first) + (stride - 1)) / stride;
    }

    protected:
    Iterator first;
    Iterator last;
    difference_type stride;
};

/********/
/* MAIN */
/********/
int main(){

    /**************************/
    /* SETTING UP THE PROBLEM */
    /**************************/

    const int Nrows = 10;           // --- Number of objects
    const int Ncols =  3;           // --- Number of centroids  

    thrust::device_vector<int> d_sequence(Nrows * Ncols);
    thrust::device_vector<int> d_counts(Ncols, Nrows);
    thrust::sequence(d_sequence.begin(), d_sequence.begin() + Ncols);
    expand(d_counts.begin(), d_counts.end(), d_sequence.begin(), 
        thrust::make_permutation_iterator(
                                d_sequence.begin(),
                                thrust::make_transform_iterator(thrust::make_counting_iterator(0),(_1 % Nrows) * Ncols + _1 / Nrows)));

    printf("\n\nCentroid indices\n");
    for(int i = 0; i < Nrows; i++) {
        std::cout << " [ ";
        for(int j = 0; j < Ncols; j++)
            std::cout << d_sequence[i * Ncols + j] << " ";
        std::cout << "]\n";
    }

    return 0;
}