#include "mat_mul.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 slice_node, slice;

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_omp(){//int tid) {
	// TODO: parallelize & optimize matrix multiplication
	// Use num_threads per node

	#pragma omp parallel num_threads (num_threads) 
	{
		int tid = omp_get_thread_num();
		int is = slice_node / num_threads * tid + min(tid, slice_node % num_threads);
		int ie = slice_node / num_threads * (tid + 1) + min(tid + 1, slice_node % 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];
							int row_C = i * N;
							for (int j = jj; j < min(N, jj + JTILESIZE); ++j) {
								C[row_C + j] += ar * B[k * N + j];
							}
						}
					}
				}
			}
		}

		if (mpi_rank == 0 && tid == 0) {
			int edge = slice_node * mpi_world_size;
			if (edge < M) {
				int is = edge;
				int ie = M;

				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];
									int row_C = i * N;
									for (int j = jj; j < min(N, jj + JTILESIZE); ++j) {
										C[row_C + 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

	/*----------seperate---------------*/
	slice_node = M / mpi_world_size;	
	int numElements_A = slice_node * K;
	int numElements_C = slice_node * N;

	MPI_Status status;
	MPI_Request request = MPI_REQUEST_NULL;

	if(mpi_rank == 0){
		int offset = slice_node;

		for(int i=1; i<mpi_world_size; ++i){
			//MPI_Send(&A[offset * K], numElements_A, MPI_FLOAT, i, 0, MPI_COMM_WORLD );
			MPI_Isend(&A[offset * K], numElements_A, MPI_FLOAT, i, 0, MPI_COMM_WORLD, &request);
			offset += slice_node;
		}
	} else {
		A = (float*)malloc(numElements_A*sizeof(float));
		B = (float*)malloc(K*N*sizeof(float));
		C = (float*)malloc(numElements_C*sizeof(float));
		MPI_Recv (A, numElements_A, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE );
	}	
	MPI_Bcast(B, K*N, MPI_FLOAT, 0, MPI_COMM_WORLD);

	/*-------------calculate-----------------------*/
	// FIXME: for now, only root process runs the matrix multiplication.
	if (slice_node < num_threads) num_threads = slice_node;
	slice = slice_node / num_threads;

	if (mpi_rank != 0){
		for (int i=0; i<numElements_C; ++i) C[i] = 0;
	}
		
	mat_mul_omp();

	//MPI_Barrier(MPI_COMM_WORLD);
	/*--------Merge-----------------------*/
	if (mpi_rank == 0){
		int offset = slice_node;
		for (int i=1; i<mpi_world_size; ++i){
			MPI_Recv (&C[offset * N], numElements_C, MPI_FLOAT, i, 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE );
			offset += slice_node;
		}
	} else {
		//MPI_Send (C, numElements_C, MPI_FLOAT, 0, 0, MPI_COMM_WORLD );
		MPI_Isend (C, numElements_C, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &request );
	}
	//MPI_Barrier(MPI_COMM_WORLD);

}