#include "mat_mul.h"

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

#define MASTER 		(0)
#define FROM_MASTER	(1)
#define FROM_WORKER	(2)

#define ITILESIZE (32)
#define JTILESIZE (1024)
#define KTILESIZE (1024)

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

void _alloc_mat(float **m, int R, int C) {
  *m = (float *)aligned_alloc(32, sizeof(float) * R * C);
  if (*m == NULL) {
    printf("Failed to allocate memory for matrix.\n");
    exit(0);
  }
}
void _zero_mat(float *m, int R, int C) { memset(m, 0, sizeof(float) * R * C); }

static int min(int x, int y) {
  return x < y ? x : y;
}

static void mat_mul_omp() {
  // TODO: parallelize & optimize matrix multiplication
  // Use num_threads per node
 int is = 0;
 int ie = rows;

#pragma omp parallel for num_threads(num_threads) schedule(dynamic)
 for (int ii = is; ii < ie; ii += ITILESIZE) {
    for (int jj = 0; jj < N; jj += JTILESIZE) {
      for (int kk = 0; kk < K; kk += KTILESIZE) {

        for (int k = kk; k < min(K, kk + KTILESIZE); k++) {
          for (int i = ii; i < min(ie, ii + ITILESIZE); i++) {
            float ar = A[i * K + k];
            for (int 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;

  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

  // FIXME: for now, only root process runs the matrix multiplication.
  if (mpi_rank == 0)
  {
    int row = M / mpi_world_size;
    int start, end;

    for(int dest = 1; dest < mpi_world_size; dest++)
    {
      start = offset = dest * row;
      end = dest == mpi_world_size -1 ? M : (dest+1)*row;
      rows = end - start;
      MPI_Isend(&offset, 1, MPI_INT, dest, FROM_MASTER, MPI_COMM_WORLD, &request);
      MPI_Isend(&rows, 1, MPI_INT, dest, FROM_MASTER, MPI_COMM_WORLD, &request);
      MPI_Isend(&A[offset*K], rows*K, MPI_FLOAT, dest, FROM_MASTER, MPI_COMM_WORLD, &request);
      MPI_Isend(B, K*N, MPI_FLOAT, dest, FROM_MASTER, MPI_COMM_WORLD, &request);
    }

    rows = row;
    mat_mul_omp();

    for(int dest = 1; dest < mpi_world_size; dest++)
    {
      MPI_Recv(&offset, 1, MPI_INT, dest, FROM_WORKER, MPI_COMM_WORLD, &status);
      MPI_Recv(&rows, 1, MPI_INT, dest, FROM_WORKER, MPI_COMM_WORLD, &status);
      MPI_Recv(&C[offset*N], rows*N, MPI_FLOAT, dest, FROM_WORKER, 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, MASTER, FROM_MASTER, MPI_COMM_WORLD, &status);
    MPI_Recv(&rows, 1, MPI_INT, MASTER, FROM_MASTER, MPI_COMM_WORLD, &status);
    MPI_Recv(A, rows*K, MPI_FLOAT, MASTER, FROM_MASTER, MPI_COMM_WORLD, &status);
    MPI_Recv(B, K*N, MPI_FLOAT, MASTER, FROM_MASTER, MPI_COMM_WORLD, &status);

    mat_mul_omp();

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