#include "mat_mul.h"

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

#include "util.h"

#define MASTER_TO_SLAVE_TAG 1 //tag for messages sent from master to slaves
#define SLAVE_TO_MASTER_TAG 4 //tag for messages sent from slaves to master

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;
}

/*
Node Num 1
[rank 0] Avg. time: 3.782508 sec
[rank 0] Avg. throughput: 290.683216 GFLOPS
Node Num 2
[rank 0] Avg. time: 2.597820 sec
[rank 0] Avg. throughput: 423.243954 GFLOPS
Node Num 3
[rank 0] Avg. time: 1.980322 sec
[rank 0] Avg. throughput: 555.218642 GFLOPS
Node Num 4
[rank 0] Avg. time: 1.788115 sec
[rank 0] Avg. throughput: 614.899922 GFLOPS
*/

static void mat_mul_omp(int start_m, int end_m) {
  // TODO: parallelize & optimize matrix multiplication
  // Use num_threads per node
    /*
    #pragma omp parallel for num_threads(num_threads) schedule(guided, 1)
    for (int i = start_m; i < end_m; ++i) {
        for (int k = 0; k < K; ++k) {          
          float ar = A[i * K + k];
          for (int j = 0; j < N; ++j){
            //printf("thread %d, i (%d) , k (%d), j(%d) \n", omp_get_thread_num(), i, k, j );
            C[i * N + j] += ar * B[k * N + j];
          }
      }
    }
    */
    int ITILESIZE = 16;
    int JTILESIZE = 128;
    int KTILESIZE = 128;
    if(K == 8192 && N ==8192 && M == 8192){
      ITILESIZE = 32;
      JTILESIZE = 1024;
      KTILESIZE = 1024;   
    }

    
    int size = (end_m - start_m);
    
   #pragma omp parallel num_threads(num_threads) firstprivate(ITILESIZE) firstprivate(JTILESIZE) firstprivate(KTILESIZE)
    {   
    
    int tid = omp_get_thread_num();
//    int is = M / num_threads * tid + min(tid, M % num_threads);
//    int ie = M / num_threads * (tid + 1) + min(tid + 1, M % num_threads);
  
    int is = start_m + ((size / num_threads) * tid);
    int ie;
    if(tid == (num_threads-1)){
      ie = start_m + ((size / num_threads) * (tid + 1) + (size % num_threads));
    }
    else{
      ie = start_m + ((size / num_threads) * (tid + 1));
    }
    //printf("tid (%d) size(%d) is(%d), ie(%d)\n", tid, size, is, ie);
    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;

  // 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)
//    mat_mul_omp();

    int size_m;
    int start_m;
    int end_m;
    
    int start_m_no0;
    int end_m_no0; 
    
    if(mpi_rank == 0){    
      
      for(int node=0; node< mpi_world_size; node++){
        
        size_m = M / (mpi_world_size);
        start_m = (node)*size_m;
        
        if(node == 0) start_m_no0 = start_m;
        
        if(((node+1) == mpi_world_size) && ((M%(mpi_world_size)) != 0)){
          end_m = M;
          if(node == 0) end_m_no0 = end_m;
        }
        else{
          end_m = start_m + size_m;
          if(node == 0) end_m_no0 = end_m;
        }
        
        if(node >= 1){
          //send the low bound first blocking, to the intended slave
          MPI_Send(&start_m, 1, MPI_INT, node, MASTER_TO_SLAVE_TAG, MPI_COMM_WORLD);
          //printf("SEND start_m (%d) to (%d)\n", start_m, node);
          
          //next send the upper bound blocking, to the intended slave        
          MPI_Send(&end_m, 1, MPI_INT, node, MASTER_TO_SLAVE_TAG + 1, MPI_COMM_WORLD);
          //printf("SEND end_m (%d) to (%d)\n", end_m, node);
          
          //finally send the allocated row portion of [A] blocking, to the intended slave    
          int sizeBuf = (end_m - start_m) * K;
          //printf("SEND A Buff Len : (%d) to (%d)\n", sizeBuf, node);
          MPI_Send(&A[start_m*K], sizeBuf, MPI_FLOAT, node, MASTER_TO_SLAVE_TAG + 2, MPI_COMM_WORLD);
          //printf("SEND A size (%d) to (%d)\n", sizeBuf, node);
          //print_mat(&A[start_m], (end_m - start_m), K);
        }
      }
    }
    else{
        alloc_mat(&A, M, K);        
        
        //receive low bound from the master
        MPI_Recv(&start_m, 1, MPI_INT, 0, MASTER_TO_SLAVE_TAG, MPI_COMM_WORLD, &status);
        //printf("RECV start_m (%d) at (%d)\n", start_m, mpi_rank);
        
        //next receive upper bound from the master
        MPI_Recv(&end_m, 1, MPI_INT, 0, MASTER_TO_SLAVE_TAG + 1, MPI_COMM_WORLD, &status);
        //printf("RECV end_m (%d) at (%d)\n", end_m, mpi_rank);
        
        //finally receive row portion of [A] to be processed from the master
        int sizeBuf = (end_m - start_m) * K;
        //printf("RECV A Buff Len : %d \n", sizeBuf);
        MPI_Recv(&A[start_m*K], sizeBuf, MPI_FLOAT, 0, MASTER_TO_SLAVE_TAG + 2, MPI_COMM_WORLD, &status);       
        //printf("RECV A size (%d) at (%d)\n", sizeBuf, mpi_rank);
        //print_mat(&A[start_m], (end_m - start_m), K);
        
    }
    
    //MPI_Barrier(MPI_COMM_WORLD);
    
    if(mpi_rank > 0){
      alloc_mat(&B, K, N);
    }
    MPI_Bcast(&B[0], K*N, MPI_FLOAT, 0, MPI_COMM_WORLD);
    //printf("Bcast B (%d) to (%d) \n", K*N, mpi_rank);
    //print_mat(&B[0], K, N);
    
    if(mpi_rank > 0){
      alloc_mat(&C, M, N);
    }
    MPI_Bcast(&C[0], M*N, MPI_FLOAT, 0, MPI_COMM_WORLD);
    //printf("Bcast C (%d) to (%d) \n", M*N, mpi_rank);
    //print_mat(&C[0], M, N);
    
    MPI_Barrier(MPI_COMM_WORLD);
    
    if(mpi_rank == 0){
        mat_mul_omp(start_m_no0, end_m_no0);
    }
     else{
        mat_mul_omp(start_m, end_m);
        if(mpi_rank >= 1){
          //send the low bound first blocking, to the intended slave
          MPI_Send(&start_m, 1, MPI_INT, 0, SLAVE_TO_MASTER_TAG, MPI_COMM_WORLD);
          //printf("SEND C start_m (%d) to (%d)\n", start_m, 0);
          
          //next send the upper bound blocking, to the intended slave        
          MPI_Send(&end_m, 1, MPI_INT, 0, SLAVE_TO_MASTER_TAG + 1, MPI_COMM_WORLD);
          //printf("SEND C end_m (%d) to (%d)\n", end_m, 0);
          
          //finally send the allocated row portion of [A] blocking, to the intended slave    
          int sizeBuf = (end_m - start_m) * N;
          //printf("SEND C Buff Len : (%d) to (%d)\n", sizeBuf, 0);
          MPI_Send(&C[start_m*N], sizeBuf, MPI_FLOAT, 0, SLAVE_TO_MASTER_TAG + 2, MPI_COMM_WORLD);
          //printf("SEND C size (%d) to (%d)\n", sizeBuf, 0);
          //print_mat(&C[start_m], (end_m - start_m), N);
        }
    }
    
    if(mpi_rank == 0){
      for (int node = 1; node < mpi_world_size; node++) {// untill all slaves have handed back the processed data
            //receive low bound from a slave
            MPI_Recv(&start_m, 1, MPI_INT, node, SLAVE_TO_MASTER_TAG, MPI_COMM_WORLD, &status);
            //receive upper bound from a slave
            MPI_Recv(&end_m, 1, MPI_INT, node, SLAVE_TO_MASTER_TAG + 1, MPI_COMM_WORLD, &status);
            //receive processed data from a slave
            MPI_Recv(&C[start_m*N], (end_m - start_m) * N, MPI_FLOAT, node, SLAVE_TO_MASTER_TAG + 2, MPI_COMM_WORLD, &status);
        }
    }
    
    
}