std::experimental::parallel::reduce

From cppreference.com
Defined in header <experimental/numeric>
template<class InputIt>

typename std::iterator_traits<InputIt>::value_type reduce(

    InputIt first, InputIt last);
(1) (parallelism TS)
template<class ExecutionPolicy, class InputIterator>

typename std::iterator_traits<InputIt>::value_type reduce(

    ExecutionPolicy&& policy, InputIt first, InputIt last);
(2) (parallelism TS)
template<class InputIt, class T>
T reduce(InputIt first, InputIt last, T init);
(3) (parallelism TS)
template<class ExecutionPolicy, class InputIt, class T>
T reduce(ExecutionPolicy&& policy, InputIt first, InputIt last, T init);
(4) (parallelism TS)
template<class InputIt, class T, class BinaryOp>
T reduce(InputIt first, InputIt last, T init, BinaryOp binary_op);
(5) (parallelism TS)
template<class ExecutionPolicy, class InputIt, class T, class BinaryOp>

T reduce(ExecutionPolicy&& policy,

         InputIt first, InputIt last, T init, BinaryOp binary_op);
(6) (parallelism TS)
1) same as reduce(first, last, typename std::iterator_traits<InputIt>::value_type{})
3) same as reduce(first, last, init, std::plus<>())
5) Reduces the range [first; last), possibly permuted and aggregated in unspecified manner, along with the initial value init over binary_op.
2,4,6) Same as (1,3,5), but executed according to policy

The behavior is non-deterministic if binary_op is not associative or not commutative.

The behavior is undefined if binary_op modifies any element or invalidates any iterator in [first; last).

Parameters

first, last - the range of elements to apply the algorithm to
init - the initial value of the generalized sum
policy - the execution policy
binary_op - binary FunctionObject that will be applied in unspecified order to the result of dereferencing the input iterators, the results of other binary_op and init.
Type requirements
-
InputIt must meet the requirements of InputIterator.

Return value

Generalized sum of init and *first, *(first+1), ... *(last-1) over binary_op,

where generalized sum GSUM(op, a
1
, ..., a
N
)
is defined as follows:

  • if N=1, a
    1
  • if N > 1, op(GSUM(op, b
    1
    , ..., b
    K
    ), GSUM(op, b
    M
    , ..., b
    N
    ))
    where
  • b
    1
    , ..., b
    N
    may be any permutation of a1, ..., aN and
  • 1 < K+1 = M ≤ N

in other words, the elements of the range may be grouped and rearranged in arbitrary order

Complexity

O(last - first) applications of binary_op.

Exceptions

  • If execution of a function invoked as part of the algorithm throws an exception,
  • if policy is parallel_vector_execution_policy, std::terminate is called
  • if policy is sequential_execution_policy or parallel_execution_policy, the algorithm exits with an exception_list containing all uncaught exceptions. If there was only one uncaught exception, the algorithm may rethrow it without wrapping in exception_list. It is unspecified how much work the algorithm will perform before returning after the first exception was encountered.
  • if policy is some other type, the behavior is implementation-defined
  • If the algorithm fails to allocate memory (either for itself or to construct an exception_list when handling a user exception), std::bad_alloc is thrown.

Notes

If the range is empty, init is returned, unmodified

  • If policy is an instance of sequential_execution_policy, all operations are performed in the calling thread.
  • If policy is an instance of parallel_execution_policy, operations may be performed in unspecified number of threads, indeterminately sequenced with each other
  • If policy is an instance of parallel_vector_execution_policy, execution may be both parallelized and vectorized: function body boundaries are not respected and user code may be overlapped and combined in arbitrary manner (in particular, this implies that a user-provided Callable must not acquire a mutex to access a shared resource)

Example

reduce is the out-of-order version of std::accumulate:

#include <iostream>
#include <chrono>
#include <vector>
#include <numeric>
#include <experimental/execution_policy>
#include <experimental/numeric>
 
int main()
{
    std::vector<double> v(10'000'007, 0.5);
 
    {
        auto t1 = std::chrono::high_resolution_clock::now();
        double result = std::accumulate(v.begin(), v.end(), 0.0);
        auto t2 = std::chrono::high_resolution_clock::now();
        std::chrono::duration<double, std::milli> ms = t2 - t1;
        std::cout << std::fixed << "std::accumulate result " << result
                  << " took " << ms.count() << " ms\n";
    }
 
    {
        auto t1 = std::chrono::high_resolution_clock::now();
        double result = std::experimental::parallel::reduce(
                            std::experimental::parallel::par,
                            v.begin(), v.end());
        auto t2 = std::chrono::high_resolution_clock::now();
        std::chrono::duration<double, std::milli> ms = t2 - t1;
        std::cout << "parallel::reduce result "
                  << result << " took " << ms.count() << " ms\n";
    }
}

Possible output:

std::accumulate result 5000003.50000 took 12.7365 ms
parallel::reduce result 5000003.50000 took 5.06423 ms

See also

sums up a range of elements
(function template)
applies a function to a range of elements
(function template)
(parallelism TS)
applies a functor, then reduces out of order
(function template)