#include "mat_mul.h"
#include <omp.h>
#include "util.h"
#include <cstdlib>
#include <cstdio>
#include <algorithm>
#include <mpi.h>
using namespace std;

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


static void mat_mul_omp() {
  // TODO: parallelize & optimize matrix multiplication
  int iblock = 32;
  int jblock = 1024;//256;
  int kblock = 1024;//30;//32;//28;
 
  #pragma omp parallel num_threads(num_threads) 
  #pragma omp for
  for(int ii = 0; ii < rows; ii+=iblock){ 
     for(int jj = 0; jj < N; jj+=jblock){
	  for(int kk = 0; kk < K; kk+=kblock){
	      for(int k = kk; k < min(kk+kblock, K); k++){	  
              for(int i = ii; i < min(ii+iblock, rows); i++){
	          float temp = A[i * K + k];
                  for(int j = jj; j < min(jj+jblock,N); j+=1) {
                     C[i * N + j] += temp * B[k * N + j];
                  }
		}
	      }
          }
     }
  }
  return; 
}

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;
  MPI_Status status;
  MPI_Request request;

  if(mpi_rank == 0){
     int row_size = M/mpi_world_size;
     int start, end;

     for(int i = 1; i < mpi_world_size; i++){
	  start = offset = i * row_size;    
	  end = i == mpi_world_size - 1 ? M : (i+1)*row_size;
	  rows = end - start;

	  MPI_Isend(&offset, 1, MPI_INT, i, 1, MPI_COMM_WORLD, &request);
	  MPI_Isend(&rows, 1, MPI_INT, i, 1, MPI_COMM_WORLD, &request);
	  MPI_Isend(&A[offset*K], rows*K, MPI_FLOAT, i, 1, MPI_COMM_WORLD,&request);
          MPI_Isend(B, K*N, MPI_FLOAT, i, 1, MPI_COMM_WORLD, &request);	  
      }
      rows = row_size;
      mat_mul_omp();

      for(int i = 1; i < mpi_world_size; i++){
	  MPI_Recv(&offset, 1, MPI_INT, i, 2, MPI_COMM_WORLD, &status);
	  MPI_Recv(&rows, 1, MPI_INT, i, 2, MPI_COMM_WORLD, &status);
	  MPI_Recv(&C[offset*N], rows*N, MPI_FLOAT, i, 2, MPI_COMM_WORLD, &status);
      }
  }
  else{
      alloc_mat(&A,M,K);
      alloc_mat(&B,K,N);
      alloc_mat(&C,M,N);
      zero_mat(C,M,N);

      MPI_Recv(&offset, 1, MPI_INT, 0, 1, MPI_COMM_WORLD, &status);
      MPI_Recv(&rows, 1, MPI_INT, 0, 1, MPI_COMM_WORLD, &status);
      MPI_Recv(A, rows*K, MPI_FLOAT, 0, 1, MPI_COMM_WORLD, &status);
      MPI_Recv(B, K*N, MPI_FLOAT, 0 , 1, MPI_COMM_WORLD, &status);

      mat_mul_omp();

      MPI_Isend(&offset, 1, MPI_INT, 0, 2, MPI_COMM_WORLD,&request);
      MPI_Isend(&rows, 1, MPI_INT, 0, 2, MPI_COMM_WORLD, &request);
      MPI_Isend(C, rows*N, MPI_FLOAT, 0, 2, MPI_COMM_WORLD, &request);
  }
}