cuda_execution_policy.hpp

Go to the documentation of this file.

#pragma once
 
#include "cuda_error.hpp"
 
namespace tf {
 
template<unsigned NT, unsigned VT>
class cudaExecutionPolicy {
 
  static_assert(is_pow2(NT), "max # threads per block must be a power of two");
 
  public:
 
  const static unsigned nt = NT;
 
  const static unsigned vt = VT;
 
  const static unsigned nv = NT*VT;
 
  cudaExecutionPolicy() = default;
 
  explicit cudaExecutionPolicy(cudaStream_t s) : _stream{s} {}
  
  cudaStream_t stream() noexcept { return _stream; };
 
  void stream(cudaStream_t stream) noexcept { _stream = stream; }
 
  private:
 
  cudaStream_t _stream {0};
};
 
using cudaDefaultExecutionPolicy = cudaExecutionPolicy<512, 9>;
 
}  // end of namespace tf -----------------------------------------------------