#ifndef _LOOP_PIPELINE_H_
#define _LOOP_PIPELINE_H_

#include <stdio.h>
#include "ap_cint.h"
#define N 20
#define NUM_TRANS 20
#define L 4
#define N_L 2

typedef int20 din_t;
typedef int20 dout_t;

#endif

void loop_pipeline(din_t input[N_L], din_t output[N_L], din_t coeff[L*N_L]) {

  int i,j,k;
  static dout_t acc;
  din_t first_out[N_L];
  din_t second_out[N_L];
  din_t third_out[N_L];

    //First Layer
    LOOP_I1: for(i=0; i < N_L; i++){
	LOOP_J1: for(j=0; j < N_L; j++){
			    first_out[i] += input[j]*coeff[i];
	}
	}

    //Second Layer
    LOOP_I2: for(i=0; i < N_L; i++){
    LOOP_J2: for(j=0; j < N_L; j++){
    	        second_out[i] += first_out[j]*coeff[N_L+i];
    }
    }

    //Third Layer
    LOOP_I3: for(i=0; i < N_L; i++){
    LOOP_J3: for(j=0; j < N_L; j++){
        	    third_out[i] += second_out[j]*coeff[N_L+N_L+i];
    }
    }

    //Fourth Layer
    LOOP_I4: for(i=0; i < N_L; i++){
    LOOP_J4: for(j=0; j < N_L; j++){
            	output[i] += third_out[j]*coeff[N_L+N_L+N_L+i];
    }
    }

	//return acc;
}