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

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

//#define MASTER 0
//#define FROM_MASTER 1
//#define FROM_WORKER 2
#define TILE_M	64
#define TILE_K	16
#define TILE_N	2048

static float *A, *B, *C;
static int M, N, K;
static int num_threads;
static int mpi_rank, mpi_world_size;
static int rows[4] = {0,};
static int offset[4] = {0,};

static void mat_mul_omp() {
	int begin = 0;
	int end = rows[mpi_rank];
   
  // TODO: parallelize & optimize matrix multiplication
  // Use num_threads per node
#pragma omp parallel for num_threads(num_threads) schedule(dynamic)
for (int ii = begin; ii < end; ii+= TILE_M) {
	for (int kk = 0; kk < K; kk+= TILE_K) {
		for (int jj = 0; jj<N; jj += TILE_N) {
			int ek = kk + TILE_K < K ? (kk + TILE_K) : K;
			int em = ii + TILE_M < M ? (ii + TILE_M) : M;
			int en = jj + TILE_N < N ? (jj + TILE_N) : N;
		    for (int i = ii; i< em; ++i) {
				for (int k = kk; k< ek; ++k) {
					for (int j = jj; j<en; ++j) {
						C[i*N+j] += A[i*K + k] * 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;
  // TODO: parallelize & optimize matrix multiplication on multi-node
  // You must allocate & initialize A, B, C for non-root processes

	int row_num = M / mpi_world_size;
	
	for (int i =0; i<mpi_world_size; i++){
		rows[i] = (i == mpi_world_size - 1) ? (M -(row_num *(mpi_world_size - 1))) : row_num;
	}
	
	for (int i = 0; i<mpi_world_size -1; i++){
	    offset[i+1] = offset[i] + rows[i];
	}

	if (mpi_rank!=0){
		M = rows[mpi_rank];
		alloc_mat(&A, rows[mpi_rank], K);
		alloc_mat(&B, K, N);
		alloc_mat(&C, rows[mpi_rank], 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++)
			MPI_Isend(&A[offset[i]*K], rows[i]*K, MPI_FLOAT, i, 0, MPI_COMM_WORLD, &request);
	}	else {
			MPI_Recv(A, rows[mpi_rank]*K, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &status);
	}

	mat_mul_omp();

	if (mpi_rank!=0) {
		MPI_Isend(C, rows[mpi_rank]*N, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &request);
	}	else{
		for (int i =1; i< mpi_world_size; i++){
		MPI_Recv(&C[offset[i] *N], rows[i]* N, MPI_FLOAT, i, 0, MPI_COMM_WORLD, &status);
		}
	}
//	free(A);
//	free(B);
//	free(C);
//   MPI_Finalize();
}