#include "headers.h"
int send_row(int dst, int nproc, int* row){
  return MPI_Send(row, nproc, MPI_INT, dst, ROWTAG, MPI_COMM_WORLD);
}
int receive_row(int src, int nproc, int* row){
  MPI_Status status;
  return MPI_Recv(row, nproc, MPI_INT, src, ROWTAG, MPI_COMM_WORLD, &status);
}
int main(int argc,  char *argv[]){
  int me, row, fd[3], i, j, nproc, square;
  int *Arow, *Brow, *Crow, *A, *B, *C;
  MPI_Init(&argc, &argv);
  MPI_Comm_rank(MPI_COMM_WORLD, &me);
  MPI_Comm_size(MPI_COMM_WORLD, &nproc);
  if(allocate_rows(&Arow, &Brow, &Crow, nproc + 1) < 0) goto fail;
  row = me + 1;
  square = nproc * nproc;
  if(me  ==  BossMan){
    /* parent code */
    if(parse_args(argc, argv, fd, &nproc) < 0) goto fail;
    /* make the matrices */
    if(allocate_rows(&A, &B, &C, square) < 0) goto fail;
    /* initialize A */
    if(get_row(fd[0], square, row, A) == -1){
      fprintf(stderr,"Initialization of A failed at row %d\n", row);
      goto fail;
    }
    /* initialize B */
    if(get_row(fd[1], square, row, B) == -1){
      fprintf(stderr,"Initialization of B failed at row %d\n", row);
      goto fail;
    }
  }
  if(DEBUG || (me == BossMan))fprintf(stderr, "Scattering A\n");
  /* Scatter A */
  if(MPI_Scatter(A,
		 nproc,
		 MPI_INT,
		 &Arow[1],
		 nproc,
		 MPI_INT,
		 BossMan,
		 MPI_COMM_WORLD) != MPI_SUCCESS){
    fprintf(stderr,"Scattering of A failed\n");
    goto fail;
  }
  if(DEBUG || (me == BossMan))fprintf(stderr, "Scattering B\n");
  /* Scatter B */
  if(MPI_Scatter(B,
                 nproc,
                 MPI_INT,
                 &Brow[1],
                 nproc,
                 MPI_INT,
                 BossMan,
                 MPI_COMM_WORLD) != MPI_SUCCESS){
    fprintf(stderr,"Scattering of B failed\n");
    goto fail;
  }
  /* ring code */
  {
    int m;
    int dst = ( (me == (nproc - 1)) ? 0 : me + 1 );
    int src = ( (me == 0) ? (nproc - 1) : me - 1 );
    /* put a barrier up for fun */
    if(SYNCHRO)MPI_Barrier(MPI_COMM_WORLD);
    /* initialize C */
    if(DEBUG || (me == BossMan))fprintf(stderr, "Clone %i initializing C\n",  me);
    for(j = 1; j <= nproc; j++) Crow[j] = Arow[row] * Brow[j];
    if(DEBUG || (me == BossMan))
      fprintf(stderr, "Clone %i finished initializing C\n", me);
    /* put a barrier up for fun */
    if(SYNCHRO)MPI_Barrier(MPI_COMM_WORLD);
    /* pass the rows of B around the ring */
    m = row;
    for(i = 1; i < nproc; i++){
      if(DEBUG || (me == BossMan))fprintf(stderr, "Clone %i sending row %i\n",  me, i);
      if(send_row(dst, nproc, &Brow[1]) != MPI_SUCCESS){
        fprintf(stderr,"Write to ring by clone %d failed\n", me);
        goto fail;
      }
      if(DEBUG || (me == BossMan))fprintf(stderr, "Clone %i sent row %i\n",  me, i);
      if(DEBUG || (me == BossMan))fprintf(stderr, "Clone %i receiving row %i\n",  me, i);
      if(receive_row(src, nproc, &Brow[1]) != MPI_SUCCESS){
        fprintf(stderr,"Read from ring by clone %d failed\n", me);
        goto fail;
      }
      if(DEBUG || (me == BossMan))fprintf(stderr, "Clone %i got row %i\n",  me, i);
      if((m -= 1) == 0) m = nproc;
      for(j = 1; j <= nproc; j++) {
        Crow[j] += Arow[m] * Brow[j]; }
    }
    /* put a barrier up for fun */
    if(SYNCHRO)MPI_Barrier(MPI_COMM_WORLD);
    /* Gathering C */
    if(DEBUG || (me == BossMan))fprintf(stderr, "Gathering C\n");
    if(MPI_Gather(&Crow[1],
                  nproc,
                  MPI_INT,
                  C,
		  nproc,
                  MPI_INT,
                  BossMan,
                  MPI_COMM_WORLD) != MPI_SUCCESS){
      fprintf(stderr, "Gathering of C failed\n");
      goto fail;
    }
    if(DEBUG || (me == BossMan))
      if(me == BossMan )
        for(i = 0; i < square; i++)
          fprintf(stdout, "C[%d] = %d\n", i, C[i]);                     
    if(DEBUG || (me == BossMan))fprintf(stderr, "C Gathered\n");
    if(me == BossMan ){
      if(DEBUG || (me == BossMan))fprintf(stderr, "Writing C out\n");
      if(set_row(fd[2], square, row, C) == -1)
        fprintf(stderr,"Writing of matrix C failed\n");
    }
    /* Put a barrier up for fun */
    if(SYNCHRO)MPI_Barrier(MPI_COMM_WORLD);
    MPI_Finalize();
    exit(EXIT_SUCCESS);
  } /* ring code */

 fail:
  MPI_Finalize();
  exit(EXIT_FAILURE);
  
}