Sorting objects with Thrust CUDA

Question

Is it possible to sort objects using the Thrust library? I have the following struct:

struct OB{
  int N;
  Cls *C; //CLS is another struct.
}

Is it possible to use thrust in order to sort an array of OB according to N? Can you provide a simple example on using thrust to sort objects? If thrust is not able to do so, is there any other CUDA libraries that allows me to do so?

Answer 1

I have not tried Thrust yet, but there is a similar sort function in CUDPP called cudppSort. You cannot directly sort structures using cudppSort, it can only handle integers or floats.

So, one way to sort array of structures is to sort the keys (of your structure) and an index array of values along with it. Later, use the sorted index array to move the structures to their final sorted locations. I have described how to do this for the cudppCompact compaction algorithm in a blog post here. The technique should be similar for cudppSort too.

Answer 2

Even though you may sort the objects by using special struct definitions, using a struct as functor, it will make thrust to change the sort algorithm from radix-sort to merge-sort. Speed of radix-sort is noticeably faster than merge-sort. So when using thrust, try to use integer types as key values as possible.

I may suggest you using "thrust::sory_by_key(..)" function.

You should change your struct from AOS to SOA structure.

struct OB{
  int N;
  Cls *C; //CLS is another struct.
}

to

struct OBs{
   int []Ns; -> thrust::device_vector<int> indices;
   Cls *C[]; -> thrust::device_vector<Cls> values;
}

When you sort the indices with sort_by_key, the values will already be sorted.

thrust::sort_by_key(indices.begin(), indices.end(), values.begin());

Answer 3

The docs for thrust::sort show it accepts a comparison operator. See in their example how those are defined and used. I haven't tested this, but based on the example, all you would need is a struct that looks something like this:

struct OBCmp {
  __host__ __device__
  bool operator()(const OB& o1, const OB& o2) {
      return o1.N < o2.N;
  }
};

and then just invoke thrust::sort(obs.begin(), obs.end(), OBCmp()).

Answer 4

Up until now you cannot sort custom objects. You can do key based sorting but not of custom objects like the struct you mentioned. There are a couple of other open CUDA based algorithms available for doing this but that too requires some modification etc to make them work for you.

Answer 5

you can sort objects by overloading operator< . For example:

__host__ __device__ struct Color{
  double blue, green, red;
  double distance;
  void dist()
  {
    distance = sqrt(blue*blue + green*green + red*red);
  }
};

__host__ __device__ bool operator<(const Color &lhs, const Color &rhs) 
{
   return lhs.distance < rhs.distance;
}

int main(void)
{
   thrust::device_vector<Color> cd;
   thrust::host_vector<Color> ch;
   for (int i = 0; i<6; i++)
   {
      Color c;
      c.blue = rand()*255;
      c.green = rand()*255;
      c.red = rand()*255;
      c.dist();
      ch.push_back(c);
   }
   cd = ch;
   thrust::sort(cd.begin(), cd.end());
   ch = cd;
   return 0;
}

the objects will be sorted after the distance.