#include "mat_mul.h"

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

#include "util.h"
#include <immintrin.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;
MPI_Request request;
MPI_Status status;

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
#pragma omp parallel for num_threads(num_threads)
  for (int ii = 0; ii < M; 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(M, 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;

  // TODO: parallelize & optimize matrix multiplication on multi-node
  // You must allocate & initialize A, B, C for non-root processes
  int base = M / mpi_world_size;
  int extra = M % mpi_world_size;
  int offset; 

 if (mpi_rank== 0) { //Master
  offset = base + extra;
  M = offset;
   for(int node=1; node < mpi_world_size; node++) {
     MPI_Isend(&offset, 1, MPI_INT, node, FROM_MASTER, MPI_COMM_WORLD, &request);
     MPI_Isend(&A[offset*K], base*K, MPI_FLOAT, node, FROM_MASTER, MPI_COMM_WORLD, &request);
     MPI_Isend(B, K*N, MPI_FLOAT, node, FROM_MASTER, MPI_COMM_WORLD, &request);

     offset = offset + base;
   }
  
   mat_mul_omp();

   for(int node=1; node < mpi_world_size; node++) {
     MPI_Recv(&offset, 1, MPI_INT, node, FROM_WORKER, MPI_COMM_WORLD, &status);
     MPI_Recv(&C[offset*N], base*N, MPI_FLOAT, node, FROM_WORKER, MPI_COMM_WORLD, &status);
   }
 } //if(mpi_rank==0)

 else { //Worker
   alloc_mat(&A, base, K);
   alloc_mat(&B, K, N);
   alloc_mat(&C, base, N);
   zero_mat(C ,base, N);

   MPI_Recv(&offset, 1, MPI_INT, MASTER, FROM_MASTER, MPI_COMM_WORLD ,&status);
   MPI_Recv(A, base*K, MPI_FLOAT, MASTER, FROM_MASTER, MPI_COMM_WORLD, &status);
   MPI_Recv(B, K*N, MPI_FLOAT, MASTER, FROM_MASTER, MPI_COMM_WORLD, &status);

   M = base;
   mat_mul_omp();
   
   MPI_Isend(&offset, 1, MPI_INT, MASTER, FROM_WORKER, MPI_COMM_WORLD ,&request);
   MPI_Isend(C, base*N, MPI_FLOAT, MASTER ,FROM_WORKER, MPI_COMM_WORLD, &request);
  } //else

}