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

static int min(int x, int y) {
  return x < y ? x : y;
}

static int rows;
#define ITILESIZE (25)
#define JTILESIZE (1024)
#define KTILESIZE (1024)
static void mat_mul_omp() {
#pragma omp parallel num_threads(num_threads)
    {
        int tid = omp_get_thread_num();
        int is = rows / num_threads * tid + min(tid, rows % num_threads);
        int ie = rows / num_threads * (tid + 1) + min(tid + 1, rows % 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];
                            }
                        }
                    }

                }
            }
        }
    }
}

#define MASTER 0
#define FROM_MASTER 1
#define FROM_WORKER 2
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) {
  // TODO: parallelize & optimize matrix multiplication on multi-node
  // You must allocate & initialize A, B, C for non-root processes

    M = _M, N = _N, K = _K;
    num_threads = _num_threads, mpi_rank = _mpi_rank,
                mpi_world_size = _mpi_world_size;

    int numworkers = mpi_world_size -1;
    int averrow, extra, offset;
    int NCA = K, NCB = N;
    int mtype;
    MPI_Status status;
    MPI_Request req;
    int dest, source;

    averrow = M/mpi_world_size;
    extra = M%mpi_world_size;
    rows = (mpi_rank == numworkers) ? averrow+extra : averrow;
    offset = averrow;
    int off[4], rrr[4];
    if (mpi_rank == MASTER) {
        A = _A, B = _B, C = _C;
        mtype = FROM_MASTER;
        for(dest = 1; dest<=numworkers; dest++) {
            rows = (dest == numworkers) ? averrow + extra : averrow;
            rrr[dest] = rows;
            off[dest] = offset;
            MPI_Isend(&offset, 1, MPI_INT, dest, mtype, MPI_COMM_WORLD, &req);
            MPI_Isend(&rows, 1, MPI_INT, dest, mtype, MPI_COMM_WORLD, &req);
            MPI_Isend(&A[offset*NCA], rows*NCA, MPI_FLOAT, dest, mtype, MPI_COMM_WORLD, &req);
            MPI_Isend(&B[0], NCA*NCB, MPI_FLOAT, dest, mtype, MPI_COMM_WORLD, &req);
            offset += rows;
        }
        mtype = FROM_WORKER;
        for(int i = 1; i<=numworkers; i++) {
            source = i;
            MPI_Irecv(&C[off[i]*NCB], rrr[i]*NCB, MPI_FLOAT, source, mtype, MPI_COMM_WORLD, &req);
        }

        rows = averrow;
        mat_mul_omp();
        for(int i = 1; i<=numworkers; i++) {
            MPI_Wait(&req, &status);
        }
    }

    if (mpi_rank > MASTER) {
        alloc_mat(&A, rows, K);
        alloc_mat(&B, K, N);
        alloc_mat(&C, rows, N);
        zero_mat(C, rows, N);

        mtype = FROM_MASTER;
        MPI_Recv(&offset, 1, MPI_INT, MASTER, mtype, MPI_COMM_WORLD, &status);
        MPI_Recv(&rows, 1, MPI_INT, MASTER, mtype, MPI_COMM_WORLD, &status);

        MPI_Recv(&A[0], rows*NCA, MPI_FLOAT, MASTER, mtype, MPI_COMM_WORLD, &status);
        MPI_Recv(&B[0], NCA*NCB, MPI_FLOAT, MASTER, mtype, MPI_COMM_WORLD, &status);

#pragma omp parallel num_threads(num_threads)
        {
            int tid = omp_get_thread_num();
            int is = rows / num_threads * tid + min(tid, rows % num_threads);
            int ie = rows / num_threads * (tid + 1) + min(tid + 1, rows % 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];
                                }
                            }
                        }

                    }
                }
            }
        }

        mtype = FROM_WORKER;
        MPI_Send(&C[0], rows*NCB, MPI_FLOAT, MASTER, mtype, MPI_COMM_WORLD);
    }
}