#include "mat_mul.h"
#include "util.h"

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

#define BLOCKSIZE_I 20
#define BLOCKSIZE_J 1024
#define BLOCKSIZE_K 512
#define FROM_MASTER 1
#define FROM_SLAVE 2

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


static void mat_mul_omp(int size_of_M) {
  // TODO: parallelize & optimize matrix multiplication
  // Use num_threads per node

  int temp_k;
  int temp_i;
  int temp_j;

  #pragma omp parallel for num_threads(num_threads) private(temp_i, temp_j, temp_k) shared(A, B, C, K, N, size_of_M)
  for (int ii = 0; ii < size_of_M; ii += BLOCKSIZE_I) {
	if (ii+BLOCKSIZE_I <= size_of_M)	temp_i = ii+BLOCKSIZE_I;
	else					temp_i = size_of_M;
	for (int jj = 0; jj < N; jj += BLOCKSIZE_J) {
		if (jj+BLOCKSIZE_J <= N)	temp_j = jj+BLOCKSIZE_J;
		else				temp_j = N;
		for (int kk = 0; kk < K; kk += BLOCKSIZE_K) {
			if (kk+BLOCKSIZE_K <= K)	temp_k = kk+BLOCKSIZE_K;
			else				temp_k = K;
			for (int k = kk; k < temp_k; k++) {
				for (int i = ii; i < temp_i; i++) {
				float ar = A[i * K + k];
					for (int j = jj; j < temp_j; j++) {
						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;

  MPI_Status status;
  MPI_Request request;

  int divided_M = M / mpi_world_size;
  int modular_M = M % mpi_world_size;
  int stride_M[4];
  int offset[4];

  int stride;

  // TODO: parallelize & optimize matrix multiplication on multi-node
  // You must allocate & initialize A, B, C for non-root processes
  if (mpi_rank != 0) {
	  alloc_mat(&A, divided_M+1, K);
	  alloc_mat(&B, K, N);
	  alloc_mat(&C, divided_M+1, N);
	  zero_mat(C, divided_M+1, N);
  }

  MPI_Bcast (B, K*N, MPI_FLOAT, 0, MPI_COMM_WORLD);

  if (mpi_rank == 0) {
  	for (int i = 1; i < mpi_world_size; i++) {
		if (i <= modular_M) stride_M[i] = divided_M + 1;
		else stride_M[i] = divided_M;
		offset[1] = divided_M;
		if (i > 1) offset[i] = offset[i-1] + stride_M[i-1]; 
	}
  	for (int i = 1; i < mpi_world_size; i++) {
		MPI_Isend (&stride_M[i], 1, MPI_INT, i, FROM_MASTER, MPI_COMM_WORLD, &request); 
		MPI_Isend (&A[offset[i]*K], stride_M[i]*K, MPI_FLOAT, i, FROM_MASTER, MPI_COMM_WORLD, &request);
  	}
	mat_mul_omp(divided_M);
  	for (int i = 1; i < mpi_world_size; i++) {
		MPI_Recv (&stride_M[i], 1, MPI_INT, i, FROM_SLAVE, MPI_COMM_WORLD, &status);
		MPI_Recv (&C[offset[i]*N], stride_M[i]*N, MPI_FLOAT, i, FROM_SLAVE, MPI_COMM_WORLD, &status);
	}
  }
		
  // FIXME: for now, only root process runs the matrix multiplication.
  if (mpi_rank != 0) {
	MPI_Recv (&stride, 1, MPI_INT, 0, FROM_MASTER, MPI_COMM_WORLD, &status);
	MPI_Recv (A, stride*K, MPI_FLOAT, 0, FROM_MASTER, MPI_COMM_WORLD, &status);
	mat_mul_omp(stride);
	MPI_Send (&stride, 1, MPI_INT, 0, FROM_SLAVE, MPI_COMM_WORLD);
	MPI_Send (C, stride*N, MPI_FLOAT, 0, FROM_SLAVE, MPI_COMM_WORLD);
  }
}