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

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

#define MIN(a, b) (((a)<(b)) ? (a):(b))
static float *A, *B, *C;
static int M, N, K;
static int num_threads;
static int mpi_rank, mpi_world_size;
static int rows;
static int kBS=32;
static int nBS=2048;
static int BS=32;

static void mat_mul_omp() {
	// TODO: parallelize & optimize matrix multiplication
	// Use num_threads per node
#pragma omp parallel num_threads(num_threads)
#pragma omp for
	for(int ii = 0; ii < rows; ii += BS){
		for (int bk = 0; bk < K; bk += kBS) { 
			for (int bn = 0; bn < N; bn += nBS) { 
				for(int i = ii; i < MIN(rows, ii+BS); i++) {
					for (int k = bk; k < MIN(bk+kBS, K); ++k) {
						float a = A[i * K + k];
						for (int j = bn; j < MIN(bn+nBS, N); j+=1) {
							C[i * N + j] += a * 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.
	// NRA = M, NCA = K, NCB = N
	int offset;
	offset = 0;
	MPI_Request request;

	if (mpi_rank == 0) {
		int averow, si, ei;
		averow = M / mpi_world_size;
		for (int i = 1; i < mpi_world_size; i++) { 
			si = offset = i * averow;
			ei = i == mpi_world_size -1 ? M : (i+1)*averow;
			rows = ei - si;
			MPI_Isend(&offset, 1, MPI_INT, i, 1, MPI_COMM_WORLD, &request);
			MPI_Isend(&rows, 1, MPI_INT, i, 1, MPI_COMM_WORLD, &request);
			MPI_Isend(&A[offset*K], rows*K, MPI_FLOAT, i, 1, MPI_COMM_WORLD, &request);
			MPI_Isend(B, K*N, MPI_FLOAT, i, 1, MPI_COMM_WORLD, &request);
		}
		rows = averow;
		mat_mul_omp();
		//offset = offset + rows;

		MPI_Status status;
		for (int i = 1; i < mpi_world_size; i++) { 
			MPI_Recv(&offset, 1, MPI_INT, i, 2, MPI_COMM_WORLD, &status);
			MPI_Recv(&rows, 1, MPI_INT, i, 2, MPI_COMM_WORLD, &status);
			MPI_Recv(&C[offset*N], rows*N, MPI_FLOAT, i, 2, MPI_COMM_WORLD, &status);
		}
	} else {
		alloc_mat(&A, M, K);
		alloc_mat(&B, K, N);
		alloc_mat(&C, M, N);
		zero_mat(C, M, N);

		MPI_Status status;
		MPI_Recv(&offset, 1, MPI_INT, 0, 1, MPI_COMM_WORLD, &status);
		MPI_Recv(&rows, 1, MPI_INT, 0, 1, MPI_COMM_WORLD, &status);
		MPI_Recv(A, rows*K, MPI_FLOAT, 0, 1, MPI_COMM_WORLD, &status);
		MPI_Recv(B, K*N, MPI_FLOAT, 0, 1, MPI_COMM_WORLD, &status);

		mat_mul_omp();

		MPI_Isend(&offset, 1, MPI_INT, 0, 2, MPI_COMM_WORLD, &request);
		MPI_Isend(&rows, 1, MPI_INT, 0, 2, MPI_COMM_WORLD, &request);
		MPI_Isend(C, rows*N, MPI_FLOAT, 0, 2, MPI_COMM_WORLD, &request);
	}
}