使用CUDA Thrust确定每个矩阵列中的最小元素及其位置

Question

I have a fairly simple problem but I cannot figure out an elegant solution to it.

I have a Thrust code which produces c vectors of same size containing values. Let say each of these c vectors have an index. I would like for each vector position to get the index of the c vector for which the value is the lowest:

Example:

C0 =     (0,10,20,3,40)
C1 =     (1,2 ,3 ,5,10)

I would get as result a vector containing the index of the C vector which has the lowest value:

result = (0,1 ,1 ,0,1)

I have thought about doing it using thrust zip iterators, but have come accross issues: I could zip all the c vectors and implement an arbitrary transformation which takes a tuple and returns the index of its lowest value, but:

1. How to iterate over the contents of a tuple?
2. As I understand tuples can only store up to 10 elements and there can be much more than 10 c vectors.

I have then thought about doing it this way: Instead of having c separate vectors, append them all in a single vector C , then generate keys referencing the positions and perform a stable sort by key which will regroup the vector entries from a same position together. In the example that would give:

C =      (0,10,20,3,40,1,2,3,5,10)
keys =   (0,1 ,2 ,3,4 ,0,1,2,3,4 )
after stable sort by key:
output = (0,1,10,2,20,3,3,5,40,10)
keys =   (0,0,1 ,1,2 ,2,3,3,4 ,4 )

Then generate keys with the positions in the vector, zip the output with the index of the c vectors and then perform a reduce by key with a custom functor which for each reduction outputs the index with the lowest value. In the example:

input =  (0,1,10,2,20,3,3,5,40,10)
indexes= (0,1,0 ,1,0 ,1,0,1,0 ,1)
keys =   (0,0,1 ,1,2 ,2,3,3,4 ,4)
after reduce by keys on zipped input and indexes:
output = (0,1,1,0,1)

However, how to write such functor for the reduce by key operation?

Answer 1

One possible idea, building on the vectorized sort idea here

1. Suppose I have vectors like this:

    values: C = ( 0,10,20, 3,40, 1, 2, 3, 5,10) keys: K = ( 0, 1, 2, 3, 4, 0, 1, 2, 3, 4) segments: S = ( 0, 0, 0, 0, 0, 1, 1, 1, 1, 1) 

2. zip together K and S to create KS

3. stable_sort_by_key using C as the keys, and KS as the values:

    stable_sort_by_key(C.begin(), C.end(), KS_begin); 

4. zip together the reordered C and K vectors, to create CK

5. stable_sort_by_key using the reordered S as the keys, and CK as the values:

    stable_sort_by_key(S.begin(), S.end(), CK_begin); 

6. use a permutation iterator or a strided range iterator to access every Nth element (0, N, 2N, ...) of the newly re-ordered K vector, to retrieve a vector of the indices of the min element in each segment, where N is the length of the segments.

I haven't actually implemented this, right now it's just an idea. Maybe it won't work for some reason I haven't observed yet.

segments ( S ) and keys ( K ) are effectively row and column indices.

And your question seems wierd to me, because your title mentions "find index of max value" but most of your question seems to be referring to "lowest value". Regardless, with a change to step 6 of my algorithm, you can find either value.

Answer 2

Since the length of your vectors has to be the same. It's better to concatenate them together and treat them as a matrix C.

Then your problem becomes finding the indices of the min element of each column in a row-major matrix. It can be solved as follows.

1. change the row-major to col-major;
2. find indices for each column.

In step 1, you proposed to use stable_sort_by_key to rearrange the element order, which is not a effective method. Since the rearrangement can be directly calculated given the #row and #col of the matrix. In thrust, it can be done with permutation iterators as:

thrust::make_permutation_iterator(
    c.begin(),
    thrust::make_transform_iterator(
        thrust::make_counting_iterator((int) 0),
        (_1 % row) * col + _1 / row)
)

In step 2, reduce_by_key can do exactly what you want. In your case the reduction binary-op functor is easy, since comparison on tuple (element of your zipped vector) has already been defined to compare the 1st element of the tuple, and it's supported by thrust as

thrust::minimum< thrust::tuple<float, int> >()

The whole program is shown as follows. Thrust 1.6.0+ is required since I use placeholders in fancy iterators.

#include <iterator>
#include <algorithm>

#include <thrust/device_vector.h>
#include <thrust/iterator/counting_iterator.h>
#include <thrust/iterator/transform_iterator.h>
#include <thrust/iterator/permutation_iterator.h>
#include <thrust/iterator/zip_iterator.h>
#include <thrust/iterator/discard_iterator.h>
#include <thrust/reduce.h>
#include <thrust/functional.h>

using namespace thrust::placeholders;

int main()
{

    const int row = 2;
    const int col = 5;
    float initc[] =
            { 0, 10, 20, 3, 40, 1, 2, 3, 5, 10 };
    thrust::device_vector<float> c(initc, initc + row * col);

    thrust::device_vector<float> minval(col);
    thrust::device_vector<int> minidx(col);

    thrust::reduce_by_key(
            thrust::make_transform_iterator(
                    thrust::make_counting_iterator((int) 0),
                    _1 / row),
            thrust::make_transform_iterator(
                    thrust::make_counting_iterator((int) 0),
                    _1 / row) + row * col,
            thrust::make_zip_iterator(
                    thrust::make_tuple(
                            thrust::make_permutation_iterator(
                                    c.begin(),
                                    thrust::make_transform_iterator(
                                            thrust::make_counting_iterator((int) 0), (_1 % row) * col + _1 / row)),
                            thrust::make_transform_iterator(
                                    thrust::make_counting_iterator((int) 0), _1 % row))),
            thrust::make_discard_iterator(),
            thrust::make_zip_iterator(
                    thrust::make_tuple(
                            minval.begin(),
                            minidx.begin())),
            thrust::equal_to<int>(),
            thrust::minimum<thrust::tuple<float, int> >()
    );

    std::copy(minidx.begin(), minidx.end(), std::ostream_iterator<int>(std::cout, " "));
    std::cout << std::endl;
    return 0;
}

Two remaining issues may affect the performance.

1. min values have to be outputted, which is not required;
2. reduce_by_key is designed for segments with variant lengths, it may not be the fastest algorithm for reduction on segments with same length.

Writing your own kernel could be the best solution for highest performance.

Answer 3

I had the curiosity to test which one of the previous approaches was faster. So, I implemented Robert Crovella's idea in the code below which reports, for the sake of completeness, also Eric's approach.

#include <iterator>
#include <algorithm>

#include <thrust/random.h>
#include <thrust/device_vector.h>
#include <thrust/iterator/counting_iterator.h>
#include <thrust/iterator/transform_iterator.h>
#include <thrust/iterator/permutation_iterator.h>
#include <thrust/iterator/zip_iterator.h>
#include <thrust/iterator/discard_iterator.h>
#include <thrust/reduce.h>
#include <thrust/functional.h>
#include <thrust/sort.h>

#include "TimingGPU.cuh"

using namespace thrust::placeholders;

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;
};


/**************************************************************/
/* CONVERT LINEAR INDEX TO ROW INDEX - NEEDED FOR APPROACH #1 */
/**************************************************************/
template< typename T >
struct mod_functor {
    __host__ __device__ T operator()(T a, T b) { return a % b; }
};

/********/
/* MAIN */
/********/
int main()
{
    /***********************/
    /* SETTING THE PROBLEM */
    /***********************/
    const int Nrows = 200;
    const int Ncols = 200;

    // --- Random uniform integer distribution between 10 and 99
    thrust::default_random_engine rng;
    thrust::uniform_int_distribution<int> dist(10, 99);

    // --- Matrix allocation and initialization
    thrust::device_vector<float> d_matrix(Nrows * Ncols);
    for (size_t i = 0; i < d_matrix.size(); i++) d_matrix[i] = (float)dist(rng);

    TimingGPU timerGPU;

    /******************/
    /* APPROACH NR. 1 */
    /******************/
    timerGPU.StartCounter();

    thrust::device_vector<float>    d_min_values(Ncols);
    thrust::device_vector<int>      d_min_indices_1(Ncols);

    thrust::reduce_by_key(
            thrust::make_transform_iterator(
                    thrust::make_counting_iterator((int) 0),
                    _1 / Nrows),
            thrust::make_transform_iterator(
                    thrust::make_counting_iterator((int) 0),
                    _1 / Nrows) + Nrows * Ncols,
            thrust::make_zip_iterator(
                    thrust::make_tuple(
                            thrust::make_permutation_iterator(
                                    d_matrix.begin(),
                                    thrust::make_transform_iterator(
                                            thrust::make_counting_iterator((int) 0), (_1 % Nrows) * Ncols + _1 / Nrows)),
                            thrust::make_transform_iterator(
                                    thrust::make_counting_iterator((int) 0), _1 % Nrows))),
            thrust::make_discard_iterator(),
            thrust::make_zip_iterator(
                    thrust::make_tuple(
                            d_min_values.begin(),
                            d_min_indices_1.begin())),
            thrust::equal_to<int>(),
            thrust::minimum<thrust::tuple<float, int> >()
    );

    printf("Timing for approach #1 = %f\n", timerGPU.GetCounter());

    /******************/
    /* APPROACH NR. 2 */
    /******************/
    timerGPU.StartCounter();

    // --- Computing row indices vector
    thrust::device_vector<int> d_row_indices(Nrows * Ncols);
    thrust::transform(thrust::make_counting_iterator(0), thrust::make_counting_iterator(Nrows * Ncols), thrust::make_constant_iterator(Ncols), d_row_indices.begin(), thrust::divides<int>() );

    // --- Computing column indices vector
    thrust::device_vector<int> d_column_indices(Nrows * Ncols);
    thrust::transform(thrust::make_counting_iterator(0), thrust::make_counting_iterator(Nrows * Ncols), thrust::make_constant_iterator(Ncols), d_column_indices.begin(), mod_functor<int>());

    // --- int and float iterators
    typedef thrust::device_vector<int>::iterator        IntIterator;
    typedef thrust::device_vector<float>::iterator      FloatIterator;

    // --- Relevant tuples of int and float iterators
    typedef thrust::tuple<IntIterator, IntIterator>     IteratorTuple1;
    typedef thrust::tuple<FloatIterator, IntIterator>   IteratorTuple2;

    // --- zip_iterator of the relevant tuples
    typedef thrust::zip_iterator<IteratorTuple1>        ZipIterator1;
    typedef thrust::zip_iterator<IteratorTuple2>        ZipIterator2;

    // --- zip_iterator creation
    ZipIterator1 iter1(thrust::make_tuple(d_column_indices.begin(), d_row_indices.begin()));

    thrust::stable_sort_by_key(d_matrix.begin(), d_matrix.end(), iter1);

    ZipIterator2 iter2(thrust::make_tuple(d_matrix.begin(), d_row_indices.begin()));

    thrust::stable_sort_by_key(d_column_indices.begin(), d_column_indices.end(), iter2);

    typedef thrust::device_vector<int>::iterator Iterator;

    // --- Strided access to the sorted array
    strided_range<Iterator> d_min_indices_2(d_row_indices.begin(), d_row_indices.end(), Nrows);

    printf("Timing for approach #2 = %f\n", timerGPU.GetCounter());

    printf("\n\n");
    std::copy(d_min_indices_2.begin(), d_min_indices_2.end(), std::ostream_iterator<int>(std::cout, " "));
    std::cout << std::endl;

    return 0;
}

Testing the two approaches for the case of 2000x2000 sized matrices, this has been the result on a Kepler K20c card:

Eric's             :  8.4s
Robert Crovella's  : 33.4s