#include "mat_mul.h"

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

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

#define FROM_MASTER 1
#define FROM_WORKER 2

#ifndef max
#define max(a,b)  (((a) > (b)) ? (a) : (b))
#endif

#ifndef min
#define min(a,b)  (((a) < (b)) ? (a) : (b))
#endif

void alloc_mat_local(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_local(float *m, int R, int C) { memset(m, 0, sizeof(float) * R * C); }

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

 #pragma omp parallel
 {
    int idx = omp_get_thread_num();
    int slice = rows / num_threads;
    int start = offset + (idx * slice);
    int end = idx == num_threads - 1 ? offset+rows : offset + (idx + 1) * slice;

    float Aik;
    int iBS = 32;
    int jBS = 1024;
    int kBS = 1024;

    for (int kk = 0; kk < K; kk += iBS) {
      for (int jj = 0; jj < N; jj += jBS) {
        for(int ii=start;ii<end;ii+=kBS){
          for (int i = ii; i < min(ii+kBS,end); ++i) {
            for (int k = kk; k < min(kk + iBS, K); ++k) {
              Aik = A[i * K + k];
              for (int j = jj; j < min(jj + jBS, N); ++j) {
                C[i * N + j] += Aik * 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;
  int i, rows, mpi_start, mpi_end, mpi_rows;

  // 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.

  MPI_Status stat;
  MPI_Request req;
  rows = M / mpi_world_size;
  omp_set_num_threads(num_threads);

  if (mpi_rank == 0)
  {
    for(i=1;i<mpi_world_size;i++)
    {
      mpi_start = i * rows;
      mpi_end = i == mpi_world_size - 1 ? M : (i + 1) * rows; 
      mpi_rows = mpi_end - mpi_start;
      //int MPI_Send(const void *buf, int count, MPI_Datatype datatype, int dest, int tag, MPI_Comm comm)
      // MPI_Send(&mpi_start, 1, MPI_INT, i, FROM_MASTER, MPI_COMM_WORLD);
      // MPI_Send(&mpi_rows, 1, MPI_INT, i,FROM_MASTER,MPI_COMM_WORLD);
      // MPI_Send(A+(mpi_start*K), mpi_rows*K, MPI_FLOAT, i, FROM_MASTER, MPI_COMM_WORLD);
      // MPI_Send(B,K*N, MPI_FLOAT, i, FROM_MASTER, MPI_COMM_WORLD);

      MPI_Isend(&mpi_start, 1, MPI_INT, i, FROM_MASTER, MPI_COMM_WORLD,&req);
      MPI_Isend(&mpi_rows, 1, MPI_INT, i,FROM_MASTER,MPI_COMM_WORLD,&req);
      MPI_Isend(A+(mpi_start*K), mpi_rows*K, MPI_FLOAT, i, FROM_MASTER, MPI_COMM_WORLD,&req);
      MPI_Isend(B,K*N, MPI_FLOAT, i, FROM_MASTER, MPI_COMM_WORLD,&req);
    }

    mat_mul_omp(0, rows);

    for(i=1;i<mpi_world_size;i++)
    {
      //int MPI_Recv(void *buf, int count, MPI_Datatype datatype, int source, int tag, MPI_Comm comm, MPI_Status *status)
      MPI_Recv(&mpi_start, 1, MPI_INT, i, FROM_WORKER, MPI_COMM_WORLD,&stat);
      MPI_Recv(&mpi_rows, 1, MPI_INT, i,FROM_WORKER,MPI_COMM_WORLD,&stat);
      MPI_Recv(C+(mpi_start*N), mpi_rows*N, MPI_FLOAT, i, FROM_WORKER, MPI_COMM_WORLD,&stat);
    }
  }
  else
  {
    alloc_mat_local(&A, M, K);
    alloc_mat_local(&B, K, N);
    alloc_mat_local(&C, M, N);
    //zero_mat_local(A, M, K);
    //zero_mat_local(B, K, N);
    zero_mat_local(C, M, N);

    MPI_Recv(&mpi_start, 1, MPI_INT, 0, FROM_MASTER, MPI_COMM_WORLD,&stat);
    MPI_Recv(&mpi_rows, 1, MPI_INT, 0,FROM_MASTER,MPI_COMM_WORLD,&stat);
    MPI_Recv(A+(mpi_start*K), mpi_rows*K, MPI_FLOAT, 0, FROM_MASTER, MPI_COMM_WORLD,&stat);
    MPI_Recv(B, K*N, MPI_FLOAT, 0, FROM_MASTER, MPI_COMM_WORLD,&stat);
    
    mat_mul_omp(mpi_start, mpi_rows);
    
    // MPI_Send(&mpi_start, 1, MPI_INT, 0, FROM_WORKER, MPI_COMM_WORLD);
    // MPI_Send(&mpi_rows, 1, MPI_INT, 0,FROM_WORKER,MPI_COMM_WORLD);
    // MPI_Send(C+(mpi_start*N), mpi_rows*N, MPI_FLOAT, 0, FROM_WORKER, MPI_COMM_WORLD);
    
    MPI_Isend(&mpi_start, 1, MPI_INT, 0, FROM_WORKER, MPI_COMM_WORLD,&req);
    MPI_Isend(&mpi_rows, 1, MPI_INT, 0,FROM_WORKER,MPI_COMM_WORLD,&req);
    MPI_Isend(C+(mpi_start*N), mpi_rows*N, MPI_FLOAT, 0, FROM_WORKER, MPI_COMM_WORLD,&req);
  }
}