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

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

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

int min(int a, int b) { return a < b ? a : b; }

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

void mat_mul_omp(int is, int ie) {

#pragma omp parallel for num_threads(num_threads)
  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;

  if (mpi_rank != 0) {
    alloc_mat(&A, M, K);
    alloc_mat(&B, K, N);
    alloc_mat(&C, M, N);
  }

  // Calculate is and ie redundantly on every processes
  int is[mpi_world_size], ie[mpi_world_size];
  for (int i = 0; i < mpi_world_size; i++) {
    is[i] = M / mpi_world_size * i;
    ie[i] = M / mpi_world_size * (i + 1);
  }
  ie[mpi_world_size - 1] = M;

  // Scatter A
  if (mpi_rank == 0) {
    for (int i = 1; i < mpi_world_size; i++) {
      MPI_Send(A + is[i] * K, (ie[i] - is[i]) * K, MPI_FLOAT, i, 0,
               MPI_COMM_WORLD);
    }
  } else {
    MPI_Recv(A + is[mpi_rank] * K, (ie[mpi_rank] - is[mpi_rank]) * K, MPI_FLOAT,
             0, 0, MPI_COMM_WORLD, nullptr);
  }

  // Broadcast B
  MPI_Bcast(B, K * N, MPI_FLOAT, 0, MPI_COMM_WORLD);

  mat_mul_omp(is[mpi_rank], ie[mpi_rank]);

  // Gather C
  if (mpi_rank == 0) {
    for (int i = 1; i < mpi_world_size; i++) {
      MPI_Recv(C + is[i] * N, (ie[i] - is[i]) * N, MPI_FLOAT, i, 0,
               MPI_COMM_WORLD, nullptr);
    }
  } else {
    MPI_Send(C + is[mpi_rank] * N, (ie[mpi_rank] - is[mpi_rank]) * N, MPI_FLOAT, 0, 0,
             MPI_COMM_WORLD);
  }
}