offload.h

Go to the documentation of this file.

/*Copyright (c) 2014, Edgar Solomonik, all rights reserved.*/

#ifndef __OFFLOAD_H__
#define __OFFLOAD_H__

//#include "../interface/common.h"

namespace CTF_int{
  class algstrct;
  
  void offload_init();
  void offload_exit();

  double estimate_download_time(int64_t size);

  double estimate_upload_time(int64_t size);
 
  class offload_arr {
    public:
      char * dev_spr;
      int64_t nbytes;
      
      offload_arr(int64_t nbytes);
  
      ~offload_arr();
  
      void download(char * host_spr);
  
      void upload(char const * host_spr);
  };

  class offload_tsr : public offload_arr {
    public:
      algstrct const * sr;
      int64_t size;
      
      offload_tsr(algstrct const * sr, int64_t size);

      void set_zero();
  };

  
  void host_pinned_alloc(void ** ptr, int64_t size);
  
  void host_pinned_free(void * ptr);
  
  template <typename dtype>
  void offload_gemm(char          tA,
                    char          tB,
                    int           m,
                    int           n,
                    int           k,
                    dtype         alpha,
                    offload_tsr & A,
                    int           lda_A,
                    offload_tsr & B,
                    int           lda_B,
                    dtype         beta,
                    offload_tsr & C,
                    int           lda_C);
  
  template <typename dtype>
  void offload_gemm(char                 tA,
                    char                 tB,
                    int                  m,
                    int                  n,
                    int                  k,
                    dtype                alpha,
                    dtype const        * dev_A,
                    int                  lda_A,
                    dtype const        * dev_B,
                    int                  lda_B,
                    dtype                beta,
                    dtype              * dev_C,
                    int                  lda_C);
}  
#endif