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

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

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

#define min(x, y) (((x) > (y)) ? (y) : (x))

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

//#define ITILESIZE (16)
//#define JTILESIZE (512)
//#define KTILESIZE (512)

static void mat_mul_omp(int limit) {


	//int tid = (long)data;
	//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 = M / mpi_world_size * mpi_rank + min(mpi_rank, M % mpi_world_size);
	//int ie = M / mpi_world_size * (mpi_rank + 1) + min(mpi_rank + 1, M % mpi_world_size);

//	printf("[mat_mul_omp] _mpi_rank : %d / is : %d / ie : %d\n", mpi_rank, is, ie);

	//printf("[mat_mul_omp] %d %d\n", mpi_rank, limit);

#pragma omp parallel for num_threads(num_threads)
	for (int ii = 0; ii < limit; 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(limit, ii + ITILESIZE); i++) {

						float ar = A[i * K + k];
						for (int j = jj; j < min(N, jj + JTILESIZE); j++) {
							C[i * N + j] += ar * B[k * N + j];
						}

					}
				}

			}
		}
	}
	//for(int i = 0; i < limit*K; i++){
	//	printf("A %d %d: %lf\n", mpi_rank, i, A[i]);
	//}

	//for(int i = 0; i < limit*N; i++){
	//	printf("C %d %d : %lf\n", mpi_rank, i, C[i]);
	//}

}

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) {

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

	MPI_Status status;
	int is[mpi_world_size];
	int ie[mpi_world_size];

	MPI_Request a_req;
	MPI_Request b_req;

	if(mpi_rank == 0){
		A = _A, B = _B, C = _C;
		for(int i = 1; i < mpi_world_size; i++){
			is[i] = M / mpi_world_size * i + min(i, M % mpi_world_size);
			ie[i] = M / mpi_world_size * (i + 1) + min(i + 1, M % mpi_world_size);
			//MPI_Send(A + is[i]*K, (ie[i] - is[i])*K   , MPI_FLOAT, i, 0, MPI_COMM_WORLD);
			//MPI_Send(B       , K*N           , MPI_FLOAT, i, 1, MPI_COMM_WORLD);
			MPI_Isend(A + is[i]*K, (ie[i] - is[i])*K   , MPI_FLOAT, i, 0, MPI_COMM_WORLD, &a_req);
			MPI_Isend(B       , K*N           , MPI_FLOAT, i, 1, MPI_COMM_WORLD, &b_req);
			//printf("mpi send AB %d\n", i);
		}
		is[0] = 0;
		ie[0] = M / mpi_world_size * 1 + min(1, M % mpi_world_size);

	} else{
		is[0] = M / mpi_world_size * mpi_rank + min(mpi_rank, M % mpi_world_size);
		ie[0] = M / mpi_world_size * (mpi_rank + 1) + min(mpi_rank + 1, M % mpi_world_size);
		//printf("malloc size %d : %d\n", mpi_rank, ie - is);
		

		A = (float*) aligned_alloc(32, sizeof(float)*(ie[0] - is[0])*K);
		B = (float*) aligned_alloc(32, sizeof(float)*K*N);
		C = (float*) aligned_alloc(32, sizeof(float)*(ie[0] - is[0])*N);
		//zero_mat(C, (ie[0] - is[0]), N);

		MPI_Recv(A, (ie[0] - is[0])*K, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &status);
		MPI_Recv(B, K*N, MPI_FLOAT, 0, 1, MPI_COMM_WORLD, &status);

		//printf("mpi receive AB %d\n", mpi_rank);

	}

	mat_mul_omp(ie[0] - is[0]);

	// A : M*K
	// B : K*N
	// C : M*N
	

	//for(int i = 0; i < M*N; i++){
	//	printf("A %d %d: %lf\n", mpi_rank, i, A[i]);
	//}

	//for(int i = 0; i < K*N; i++){
	//	printf("B %d %d : %lf\n", mpi_rank, i, B[i]);
	//}

	//for(int i = 0; i < M*N; i++){
	//	printf("C %d %d : %lf\n", mpi_rank, i, C[i]);
	//}

	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, 2, MPI_COMM_WORLD, &status);
			//printf("mpi receive C %d\n", i);
		}
	}
	else{
		//MPI_Send(C, (ie[0] - is[0])*N, MPI_FLOAT, 0, 2, MPI_COMM_WORLD);
		MPI_Isend(C, (ie[0] - is[0])*N, MPI_FLOAT, 0, 2, MPI_COMM_WORLD, &a_req);
		//printf("mpi send C %d\n", mpi_rank);
		free(A);
		free(B);
		free(C);
	}

}