how to create an appropriate temp vector?

[bricerebsamen]

hello

I have a function that uses thrust and does something like this:

void foo(Point* points, int num_pts) {
  thrust::device_vector<int> tags(num_pts); // temp data
  thrust::fill(tags.begin(), tags.end(), 0);
  // .. work on tags and points together using a zip iterator and thrust::sort
}

As it is, it's designed to work on the GPU only, because it uses a device_vector. So it assumes points is a device pointer. How could I modify this function so that it could accept either a device or host vector for points and work on cpu or gpu accordingly.

perhaps the new signature could take iterators instead, which would encode "where" the input points are located

void foo(PointIterator points_bgn, PointIterator points_end);

related to https://stackoverflow.com/questions/76597522/thrust-execution-policy-with-device-vector

[gevtushenko]

Hello @bricerebsamen!

You could try template template parameters to achieve that:

template <template <class, class> class VectorT, 
          template <class> class AllocatorT,
          class T> 
void foo(VectorT<T, AllocatorT<T>> &vec) {
  using vec_t = VectorT<T, AllocatorT<T>>;

  vec_t tmp(vec.size() /* size */, 0 /* fill value */);
}

int main() {
  {
    thrust::host_vector<double> points(100);
    foo(points);
  }

  {
    thrust::device_vector<double> points(100);
    foo(points);
  }
}

[miscco]

So in most cases it would be a bit of an overkill to go the full template template way. If you do not need the access to all the template arguments then just making it a simple template suffices:

template <class Container> 
void foo(Container& vec) {
  Container tmp(vec.size() /* size */, 0 /* fill value */);
}

[pauleonix]

Just FYI, the vector (both Thrust and STL) constructor taking only the number of elements always value-initializes these elements, i.e. tags doesn't need to be filled with 0 after construction. If you want to be more explicit you can use the constructor with an explicit value, i.e. thrust::device_vector<int> tags(num_pts, 0);. Filling it afterwards is a waste (and at least for device_vector in wont be optimized away either).