#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 count;

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

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


static void* mat_mul_op(){

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

			}
		}
	}
}
return NULL;
}

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;

	MPI_Status status;
	// 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.
	

	int offset = (M / mpi_world_size) + 1;
	int is = mpi_rank * offset;
	int ie = min(is + offset, M);
	count = ie - is;

	if (mpi_rank == 0) {
		int s, e, c;

		for (int i = 1; i < mpi_world_size; i++) {
			s = i * offset;
			e = min(s + offset, M);
			c = e - s;
			if (c > 0) {
				MPI_Send(&A[s * K + 0], c * K, MPI_FLOAT, i, i, MPI_COMM_WORLD);
				MPI_Send(B, K * N, MPI_FLOAT, i, i, MPI_COMM_WORLD);
				MPI_Send(C, c * N, MPI_FLOAT, i, i, MPI_COMM_WORLD);
			}
		}
		mat_mul_op();
		for (int i = 1; i < mpi_world_size; i++) {
			s = i * offset;
			e = min(s + offset, M);
			c = e - s;

			if (c > 0) {
				MPI_Recv(&C[s * N + 0], c * N, MPI_FLOAT, i, 0, MPI_COMM_WORLD, &status);
			}
		}
	} else {
		if (count > 0) {
			alloc_mat(&A, count, K);
			alloc_mat(&B, K, N);
			alloc_mat(&C, count, N);
			MPI_Recv(A, count * K, MPI_FLOAT, 0, mpi_rank, MPI_COMM_WORLD, &status);
			MPI_Recv(B, K * N, MPI_FLOAT, 0, mpi_rank, MPI_COMM_WORLD, &status);
			MPI_Recv(C, count * N, MPI_FLOAT, 0, mpi_rank, MPI_COMM_WORLD, &status);
			mat_mul_op();
			MPI_Send(C, count * N, MPI_FLOAT, 0, 0, MPI_COMM_WORLD);
		}
	}
}