#include "mat_mul.h"

#include <cstdio>
#include <cstdlib>
#include <mpi.h>
#include <omp.h>

static float *A, *B, *C;
static int M, N, K;
static int num_threads;
static int mpi_rank, mpi_world_size;
static int min(int x, int y) {
  return x < y ? x : y;
}
MPI_Request request;
MPI_Status status;
#define ITILESIZE (32)
#define JTILESIZE (1024)
#define KTILESIZE (1024)
static void mat_mul_omp(int tid,int bounds) {
  // TODO: parallelize & optimize matrix multiplication
  // Use num_threads per node
int i,j,k,ii,jj,kk;
float ar=0.0;
int is, ie;
//int per_process = M/mpi_world_size;
//MPI_Status status;

//omp_set_num_threads(num_threads);

//int tid;
//   #pragma omp parallel shared(A,B,C) private(i,j,k,ii,jj,kk,tid,is,ie,ar)
//{
//  tid = omp_get_thread_num();
  
  is =  bounds / num_threads * tid + min(tid, bounds % num_threads);
  ie = bounds / num_threads * (tid + 1) + min(tid + 1, bounds % num_threads);
  //int interval = M/mpi_world_size;
  //if(mpi_rank == mpi_world_size-1) interval = interval + M%mpi_world_size;
  for (ii = is; ii < ie; ii += ITILESIZE) {
    for (jj = 0; jj < N; jj += JTILESIZE) {
      for (kk = 0; kk < K; kk += KTILESIZE) {

        for (k = kk; k < min(K, kk + KTILESIZE); k++) {
          for (i = ii; i < min(ie, ii + ITILESIZE); i++) {
            ar = A[i * K + k];
            for (j = jj; j < min(N, jj + JTILESIZE); j+=1) {
              C[i * N + j] += ar * B[k * N + j];
            }
          }
        }

      }
    }
  }
//}

}


void mat_mul(float *_A, float *_B, float *_C, int _M, int _N, int _K,
             int _num_threads, int _mpi_rank, int _mpi_world_size) {//(
  A = _A, B = _B, C = _C;
  M = _M, N = _N, K = _K;
  num_threads = _num_threads, mpi_rank = _mpi_rank,
  mpi_world_size = _mpi_world_size;

  int i;
  int cal_node = mpi_world_size;
  int offset = M/cal_node;
  int bound  = (mpi_rank == mpi_world_size-1)? M/cal_node + M%cal_node : M/cal_node; 
  // TODO: parallelize & optimize matrix multiplication on multi-node
  // You must allocate & initialize A, B, C for non-root processes
  // FIXME: for now, only root process runs the matrix multiplication.
  if (mpi_rank == 0) {
    offset = M/cal_node;    
    for(i=1; i<mpi_world_size; ++i)
    {
       bound = (i == mpi_world_size-1)? M/cal_node +(M%cal_node) : M/cal_node; 
       MPI_Isend(&A[offset*K],bound*K,MPI_FLOAT,i,0,MPI_COMM_WORLD,&request);
       MPI_Isend(B,K*N,MPI_FLOAT,i,0,MPI_COMM_WORLD,&request); 
       //MPI_Send(&A[offset*K],bound*K,MPI_FLOAT,i,0,MPI_COMM_WORLD);
       //MPI_Send(B,K*N,MPI_FLOAT,i,0,MPI_COMM_WORLD); 
       //MPI_Irecv(&C[offset*N],bound*N,MPI_FLOAT,i,0,MPI_COMM_WORLD,&request);
       //MPI_Recv(&C[offset*N],bound*N,MPI_FLOAT,i,0,MPI_COMM_WORLD,&status);
       offset += bound;
    }
    //MPI_Wait(&request, &status);
    #pragma omp parallel num_threads(num_threads)
    {
       #pragma omp for nowait
       for(i=0; i<num_threads; ++i)
       mat_mul_omp(i,bound);
       #pragma omp barrier

    } 
    

    
  /*  for(i=1; i<mpi_world_size; ++i) { 
       //MPI_Irecv(&C[offset*N],bound*N,MPI_FLOAT,i,0,MPI_COMM_WORLD,&request);
       MPI_Recv(&C[offset*N],bound*N,MPI_FLOAT,i,0,MPI_COMM_WORLD,&status);
       
       //MPI_Wait(&request, &status); 
    } */ 

      // MPI_Wait(&request, &status); 
  } 
  else { 
    A = (float*)malloc(sizeof(float)*bound*K);
    B = (float*)malloc(sizeof(float)*K*N);
    C = (float*)malloc(sizeof(float)*bound*N);

    //MPI_Irecv(A,bound*K,MPI_FLOAT,0,0,MPI_COMM_WORLD,&request);
    //MPI_Irecv(B,K*N,MPI_FLOAT,0,0,MPI_COMM_WORLD,&request);

    MPI_Recv(A,bound*K,MPI_FLOAT,0,0,MPI_COMM_WORLD,&status);
    MPI_Recv(B,K*N,MPI_FLOAT,0,0,MPI_COMM_WORLD,&status);
    #pragma omp parallel num_threads(num_threads)
    {  
      
       #pragma omp for nowait
       for(i=0; i<num_threads; ++i)
        mat_mul_omp(i,bound);
       #pragma omp barrier 
    } 

    //MPI_Isend(C, bound*N, MPI_FLOAT, 0, 0, MPI_COMM_WORLD,&request);
    //MPI_Send(C, bound*N, MPI_FLOAT, 0, 0, MPI_COMM_WORLD);
    //MPI_Wait(&request, &status); 
  }   


  

  //#pragma omp parallel num_threads(num_threads)
    //{
    //   #pragma omp for nowait
    //   for(i=0; i<num_threads; ++i)
    //   mat_mul_omp(i,bound);
    //   #pragma omp barrier

    //}   



  if (mpi_rank == 0) {
    offset = M/cal_node;
    for(i=1; i<mpi_world_size; ++i){  
       bound = (i == mpi_world_size-1)? M/cal_node + (M%cal_node) : M/cal_node; 
       //MPI_Recv(&C[offset*N],bound*N,MPI_FLOAT,i,0,MPI_COMM_WORLD,MPI_STATUS_IGNORE);
       MPI_Irecv(&C[offset*N],bound*N,MPI_FLOAT,i,0,MPI_COMM_WORLD,&request);
       offset += bound;
    }
       //MPI_Wait(&request, &status); 
    
  }
  else {	  
    MPI_Send(C, bound*N, MPI_FLOAT, 0, 0, MPI_COMM_WORLD);
    //MPI_Isend(C, bound*N, MPI_FLOAT, 0, 0, MPI_COMM_WORLD,&request);
  
    //MPI_Wait(&request, &status); 
  } 
  //MPI_Finalize();
} //)