#include "mat_mul.h"
#include <omp.h>
#include <cstdio>
#include <cstdlib>
#include <mpi.h>
#include "util.h"
#define min(a,b) (a)>(b)?(b):(a)

#define ITILESIZE (32)
#define JTILESIZE (2048)
#define KTILESIZE (16)

static float *A, *B, *C;
static int M, N, K;
static int num_threads;
static int mpi_rank, mpi_world_size;
static int source, dest;
static int rows;
static int averow, extra, offset;

MPI_Status status;

static void mat_mul_omp() {
  // TODO: parallelize & optimize matrix multiplication
  // Use num_threads per node
int tid;
int ii,jj,kk,i,j,k;

omp_set_num_threads(num_threads);

#pragma omp parallel private(tid,ii,jj,kk,i,j,k)
{
  tid = omp_get_thread_num();
  int slice = rows / num_threads;
  int is = tid*slice;
  int ie = (tid==num_threads-1)? rows : (tid+1)*slice;

  for ( ii = is; ii < ie; ii += ITILESIZE) {
    for ( kk = 0; kk < K; kk += KTILESIZE) {
      for ( jj = 0; jj < N; jj += JTILESIZE) {
        
        for(i=ii; i< (min(ie,ii+ITILESIZE));i++){
	  for(k=kk; k< (min(K,kk+KTILESIZE)); k++){
	    float 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;

  // 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){
    averow = M / mpi_world_size;
    extra = M % mpi_world_size;
    offset = averow;

    for(dest=1; dest<mpi_world_size; dest++){
	    rows = (dest <= extra) ? averow+1 : averow;
	    MPI_Send(&offset, 1, MPI_INT, dest, 0, MPI_COMM_WORLD);
	    MPI_Send(&rows, 1, MPI_INT, dest, 0, MPI_COMM_WORLD);
	    MPI_Send(&A[offset*K], rows*K, MPI_FLOAT, dest, 0, MPI_COMM_WORLD);
	    MPI_Send(B, N*K, MPI_FLOAT, dest, 0, MPI_COMM_WORLD);
	    MPI_Send(&C[offset*N], rows*N, MPI_FLOAT, dest, 0, MPI_COMM_WORLD);
	    offset = offset + rows;
    }
    rows = averow;
  }
  else if(mpi_rank>0){
	 alloc_mat(&A,M,K);
         alloc_mat(&B,K,N);
	 alloc_mat(&C,M,N);
	 MPI_Recv(&offset, 1, MPI_INT, 0, 0, MPI_COMM_WORLD, &status);
	 MPI_Recv(&rows, 1, MPI_INT,0, 0, MPI_COMM_WORLD, &status);
	 MPI_Recv(A, rows*K, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &status);
	 MPI_Recv(B, N*K, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &status);
	 MPI_Recv(C, rows*N, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &status);
  }

 // printf("MPIrank :%d, rows:%d\n",mpi_rank,rows);

  mat_mul_omp();

  if(mpi_rank>0){
	 MPI_Send(&offset, 1, MPI_INT, 0, 0, MPI_COMM_WORLD);
	 MPI_Send(&rows, 1, MPI_INT, 0, 0, MPI_COMM_WORLD);
	 MPI_Send(C, rows*N, MPI_FLOAT, 0, 0, MPI_COMM_WORLD);  
 }
 else if(mpi_rank==0){
        for(source=1; source<mpi_world_size; source++){
	    MPI_Recv(&offset, 1, MPI_INT, source, 0, MPI_COMM_WORLD, &status);
	    MPI_Recv(&rows, 1, MPI_INT, source, 0, MPI_COMM_WORLD, &status);
	    MPI_Recv(&C[offset*N], rows*N, MPI_FLOAT, source, 0, MPI_COMM_WORLD, &status);
          }
 }
}