#include "mat_mul.h"

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

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_omp2() {
  // TODO: parallelize & optimize matrix multiplication
  // Use num_threads per node
#pragma omp parallel for
  for (int i = 0; i < M; ++i) {
    for (int j = 0; j < N; ++j) {
      for (int k = 0; k < K; ++k) {
        C[i * N + j] += A[i * K + k] * B[k * N + j];
      }
    }
  }
}*/

#define min(a,b) (a>b?b:a)
static void mat_mul_omp() {
  float Aik;
  int bs_k = 32;
  int bs_j = 2048;
  //omp_set_num_threads(num_threads);
  for (int kk = 0; kk < K; kk += bs_k) 
   for (int jj = 0; jj < N; jj += bs_j) 
    #pragma omp parallel for num_threads(num_threads)
    for (int i = 0; i < M; i++)
      for (int k = kk; k < min(kk + bs_k, K); k++) {
        Aik = A[i*K + k];
        //for (int j =0; j < N; j++)
        for (int j = jj; j < min(jj + bs_j, N); j++)
          C[i*N + j + 0] += Aik * B[k*N + j + 0];
        //for (int j =0; j < (N/8)*8; j+=8) {
        //  C[i*N + j + 0] += Aik * B[k*N + j + 0];
        //  C[i*N + j + 1] += Aik * B[k*N + j + 1];
        //  C[i*N + j + 2] += Aik * B[k*N + j + 2];
        //  C[i*N + j + 3] += Aik * B[k*N + j + 3];
        //  C[i*N + j + 4] += Aik * B[k*N + j + 4];
        //  C[i*N + j + 5] += Aik * B[k*N + j + 5];
        //  C[i*N + j + 6] += Aik * B[k*N + j + 6];
        //  C[i*N + j + 7] += Aik * B[k*N + j + 7];
        //}
        //for (; j < N; j++)
        //  C[i*N + j] += Aik * B[k*N + j];
      }

  return;
}

/*static void mat_mul_omp() {
  // TODO: parallelize & optimize matrix multiplication
  int bs = 64;
  for (int jj = 0; jj < K; jj += bs) 
   for (int kk = 0; kk < K; kk += bs) 
    for (int i = 0; i < M; ++i) 
    #pragma omp parallel for
     for (int j =jj; j < min(jj+bs, N); j++)
      for (int k = kk; k < min(kk+bs, K); k++) 
          //C[i*N + j + 0] += Aik * B[k*N + j + 0];
	  C[i*N + j] += A[i*K + k] * B[k*N + j];

  return;
}*/


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 default_div_size = M/mpi_world_size;
  MPI_Status status;
  MPI_Request request;
  if(mpi_rank == 0){
	  //1. Send part of A and B to each node

	  for(int target_rank = 1; target_rank < mpi_world_size; target_rank++){
		  int div_start, div_size;
		  div_start = target_rank * default_div_size;
		  div_size = default_div_size;
		  if(target_rank == (mpi_world_size - 1))
			  div_size += M - (default_div_size * mpi_world_size);

		  //printf("send. rank=%d, div_start=%d, div_size=%d\n", target_rank, div_start, div_size);

		  //Send part of array A
		  //MPI_Send(A + (div_start * K), div_size * K, MPI_FLOAT, target_rank, 0, MPI_COMM_WORLD);
		  MPI_Isend(A + (div_start * K), div_size * K, MPI_FLOAT, target_rank, 0, MPI_COMM_WORLD, &request);

		  //MPI_Isend(B, K * N, MPI_FLOAT, target_rank, 0, MPI_COMM_WORLD, &request);
	  }
	  MPI_Bcast(B, K * N, MPI_FLOAT, 0, MPI_COMM_WORLD);
	  //MPI_Wait(&request, &status);

	  //2. Calc rank0 mat mul
	  int original_M = M;
	  M = default_div_size; // make mat_mal compute the first division of M.

	  //printf("main. my rank=%d, div_size=%d\n", mpi_rank, default_div_size);
	  //timer_start(1);
	  mat_mul_omp();
	  //double elapsed_time = timer_stop(1);
	  //printf("[rank %d] time: %f sec\n", mpi_rank, elapsed_time);
	  M = original_M; // Restore M

	  //3. Recv C from other nodes
	  MPI_Request arrC_req[4];
	  MPI_Status arrC_status[4];
	  for(int target_rank = 1; target_rank < mpi_world_size; target_rank++){
		  int div_start, div_size;
		  div_start = target_rank * default_div_size;
		  div_size = default_div_size;
		  if(target_rank == (mpi_world_size - 1))
			  div_size += M - (default_div_size * mpi_world_size);

		  //MPI_Recv(C + div_start * N, div_size * N, MPI_FLOAT, target_rank, 0, MPI_COMM_WORLD, &status);
		  MPI_Irecv(C + div_start * N, div_size * N, MPI_FLOAT, target_rank, 0, MPI_COMM_WORLD, &arrC_req[target_rank-1]);
	  }
	  MPI_Waitall(mpi_world_size-1, arrC_req, arrC_status);
	  //printf("main. wait. end\n");
  }else{
	  //0. alloc local memory
	  int div_size;

	  div_size = default_div_size;
	  if(mpi_rank == (mpi_world_size - 1))
		  div_size += M - (default_div_size * mpi_world_size);
	  M = div_size; // Adjust M size

	  alloc_mat(&A, M, K);
	  alloc_mat(&B, K, N);
	  alloc_mat(&C, M, N);



	  //1. Recv part of A
	  MPI_Recv(A, M * K, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &status);

	  //2. Recv full B
	  //MPI_Recv(B, K * N, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &status);
	  MPI_Bcast(B, K * N, MPI_FLOAT, 0, MPI_COMM_WORLD);

	  //3. Calc sub matrix
	  //printf("sub. my rank=%d, div_size=%d\n", mpi_rank, div_size);
	  //timer_start(1);
	  mat_mul_omp();
	  //double elapsed_time = timer_stop(1);
	  //printf("[rank %d] time: %f sec\n", mpi_rank, elapsed_time);

	  //4. Send C to rank 0 node.
	  //printf("sub. end. my rank=%d, div_size=%d\n", mpi_rank, div_size);
	  MPI_Send(C, M * N, MPI_FLOAT, 0, 0, MPI_COMM_WORLD);
	  //MPI_Isend(C, M * N, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &request);

  }
}