Use parallel implementations for CPU execution policies #831
Description
Execution policies like TBB and OMP are currently using serial implementation for some algorithms. For instance:
#define THRUST_HOST_SYSTEM THRUST_HOST_SYSTEM_OMP
#include <thrust/execution_policy.h>
#include <thrust/host_vector.h>
#include <thrust/scan.h>
int main() {
thrust::host_vector<int> keys(1 << 20, 1);
thrust::host_vector<int> vals(1 << 20, 1);
thrust::exclusive_scan_by_key(keys.begin(), keys.end(), vals.begin(),
vals.begin(), 0, thrust::plus<int>(),
thrust::equal_to<int>());
}We have generic parallel implementation of this algorithm, that's unused.
Tasks
- Use generic implementation of
*scan_by_keyfor TBB and OMP execution policies - Implement parallel version of
*scanfor OMP execution policy