/*************************************************************************
 *
 *  This file is part of ACT dataflow neuro library
 *
 *  Copyright (c) 2022 University of Groningen - Ole Richter
 *  Copyright (c) 2022 University of Groningen - Michele Mastella
 *  Copyright (c) 2022 University of Groningen - Hugh Greatorex 
 *
 *  This source describes Open Hardware and is licensed under the CERN-OHL-W v2 or later
 *
 *  You may redistribute and modify this documentation and make products
 *  using it under the terms of the CERN-OHL-W v2 (https:/cern.ch/cern-ohl).
 *  This documentation is distributed WITHOUT ANY EXPRESS OR IMPLIED
 *  WARRANTY, INCLUDING OF MERCHANTABILITY, SATISFACTORY QUALITY
 *  AND FITNESS FOR A PARTICULAR PURPOSE. Please see the CERN-OHL-W v2
 *  for applicable conditions.
 *
 *  Source location: https://git.web.rug.nl/bics/actlib_dataflow_neuro
 *
 *  As per CERN-OHL-W v2 section 4.1, should You produce hardware based on
 *  these sources, You must maintain the Source Location visible in its
 *  documentation.
 *
 **************************************************************************
 */
import "../../dataflow_neuro/cell_lib_async.act";
import "../../dataflow_neuro/cell_lib_std.act";
import "../../dataflow_neuro/treegates.act";
// import tmpl::dataflow_neuro;
// import tmpl::dataflow_neuro;
import std::channel;
open std::channel;

namespace tmpl {
    namespace dataflow_neuro {

        // @ole talk to rajit, we use valid the wrong way arround according to stdlib
        template<pbool reset; pint V; pint M>
        defchan gen_avMx1of2 <: chan(int<M>) (std::data::Mx1of2?!<M> d; bool!? a; bool!? v)
        {
          { 0 <= V & std::ceil_log2(V) < M : "Initial token value out of range" };
          
          methods {
            /*-- initialize channel, sender end --*/
            send_init {
              [ reset -> (,i:M: [ ((V >> i) & 1) = 0 -> d.d[i].f+ [] else -> d.d[i].t+ ]);[v]
              [] else -> (,i:M: d.d[i].t-,d.d[i].f-);[~v]
              ]
            }
        
            /*-- set output data --*/
            set {
              (,i:M: [((self >> i) & 1) = 0 -> d.d[i].f+ [] else -> d.d[i].t+ ]);[v]
            }
        
            /*-- finish synchronization --*/
            send_up {
              [a]
            }
        
            /*-- reset part of the protocol --*/
            send_rest {
              (,i:M: d.d[i].t-,d.d[i].f-);[~v],[~a]
            }
        
            /*-- initialize channel, receiver end --*/
            recv_init {
              v-;a-
            }
        
            /*-- get value --*/
            get {
              [(&i:M: d.d[i].t | d.d[i].f)];
              self := 0;
              (;i:M: [ d.d[i].t -> self := self | (1 << i)
        	     [] else -> skip
        	     ]
              )
            }
        
            /*-- finish synchronization action --*/
            recv_up {
              v+,a+
            }
        
            /*-- reset part of the protocol --*/
            recv_rest {
              [(&i:M:~d.d[i].t & ~d.d[i].f)];v-,a-
            }
        
            /*-- probe expression on receiver --*/
            // i think this deadlocks with recv_up
            recv_probe = v;
        
            // no sender probe
          }
        }
        export defchan avMx1of2 <: gen_avMx1of2<false,0> () { }
        export defchan avrMx1of2 <: gen_avMx1of2<true,0> () { }
        

        /**
         * the buffer template gives you a standart buffer of bitwidth N
         *
         */
        export template<pint N>
        defproc buffer (avMx1of2<N> in; avMx1of2<N> out; bool? reset_B; power supply) {
            //control
            bool _en, _reset_BX,_reset_BXX[N];
            A_3C_RB_X4 inack_ctl(.c1=_en,.c2=in.v,.c3=out.v,.y=in.a,.pr_B=reset_B,.sr_B=reset_B,.vdd=supply.vdd,.vss=supply.vss);
            A_1C1P_X1 en_ctl(.c1=in.a,.p1=out.v,.y=_en,.vdd=supply.vdd,.vss=supply.vss);
            
            //reset_buffers
	          sigbuf<N> reset_bufarray(.in=_reset_BX, .out=_reset_BXX);
	          BUF_X1 reset_buf(.a=reset_B, .y=_reset_BX,.vdd=supply.vdd,.vss=supply.vss);
            
            //validity
            bool _in_v;
            ctree<N> vc(.in=in.d,.out=_in_v,.supply=supply);
            BUF_X4 in_v_buf(.a=_in_v, .y=in.v,.vdd=supply.vdd,.vss=supply.vss);
            
            //function
            bool _out_a_BX_t[N],_out_a_BX_f[N],_out_a_B,_en_X_t[N],_en_X_f[N];
            A_2C1N_RB_X4 f_buf_func[N];
            A_2C1N_RB_X4 t_buf_func[N];
            sigbuf<N> en_buf_t(.in=_en, .out=_en_X_t, .supply=supply);
            sigbuf<N> en_buf_f(.in=_en, .out=_en_X_f, .supply=supply);
            INV_X1 out_a_inv(.a=out.a,.y=_out_a_B);
            sigbuf<N> out_a_B_buf_f(.in=_out_a_B,.out=_out_a_BX_t);
            sigbuf<N> out_a_B_buf_t(.in=_out_a_B,.out=_out_a_BX_f);
            // check if you can also do single var to array connect a=b[N]
            // and remove them from the loop
            (i:N: 
                f_buf_func[i].y=out.d.d[i].f;
                t_buf_func[i].y=out.d.d[i].t;
                f_buf_func[i].c1=_en_X_f[i];
                t_buf_func[i].c1=_en_X_t[i];
                f_buf_func[i].c2=_out_a_BX_f[i];
                t_buf_func[i].c2=_out_a_BX_t[i];
                f_buf_func[i].n1=in.d.d[i].f;
                t_buf_func[i].n1=in.d.d[i].t;
                f_buf_func[i].vdd=supply.vdd;
                t_buf_func[i].vdd=supply.vdd;
                f_buf_func[i].vss=supply.vss;
                t_buf_func[i].vss=supply.vss;
                t_buf_func[i].pr_B = _reset_BXX[i];
                t_buf_func[i].sr_B = _reset_BXX[i];
                f_buf_func[i].pr_B = _reset_BXX[i];
                f_buf_func[i].sr_B = _reset_BXX[i];
            )
        }

        export template<pint N>
        defproc fork (avMx1of2<N> in; avMx1of2<N> out1; avMx1of2<N> out2 ; bool? reset_B; power supply) {
          // control 
          bool _en, _reset_BX,_reset_BXX[N*2];
          A_4C_RB_X4 inack_ctl(.c1=_en,.c2=in.v,.c3=out1.v,.c4=out2.v,.y=in.a,.pr_B=reset_B,.sr_B=reset_B,.vdd=supply.vdd,.vss=supply.vss);
          A_1C2P_X1 en_ctl(.c1=in.a,.p1=out1.v,.p2=out2.v,.y=_en,.vdd=supply.vdd,.vss=supply.vss);

          //reset_buffers
          sigbuf<N*2> reset_bufarray(.in=_reset_BX, .out=_reset_BXX);
          BUF_X1 reset_buf(.a=reset_B, .y=_reset_BX,.vdd=supply.vdd,.vss=supply.vss);

          //validity
          bool _in_v;
          ctree<N> vc(.in=in.d,.out=_in_v,.supply=supply);
          BUF_X4 in_v_buf(.a=_in_v, .y=in.v,.vdd=supply.vdd,.vss=supply.vss);

          //function
            //func buffer out1
            bool _out1_a_BX_t[N],_out1_a_BX_f[N],_out1_a_B,_en1_X_t[N],_en1_X_f[N];
            A_2C1N_RB_X4 out1_f_buf_func[N];
            A_2C1N_RB_X4 out1_t_buf_func[N];
            sigbuf<N> out1_en_buf_t(.in=_en, .out=_en1_X_t, .supply=supply);
            sigbuf<N> out1_en_buf_f(.in=_en, .out=_en1_X_f, .supply=supply);
            INV_X1 out1_a_inv(.a=out1.a,.y=_out1_a_B);
            sigbuf<N> out1_a_B_buf_f(.in=_out1_a_B,.out=_out1_a_BX_t);
            sigbuf<N> out1_a_B_buf_t(.in=_out1_a_B,.out=_out1_a_BX_f);
            (i:N: 
                out1_f_buf_func[i].y=out1.d.d[i].f;
                out1_t_buf_func[i].y=out1.d.d[i].t;
                out1_f_buf_func[i].c1=_en1_X_f[i];
                out1_t_buf_func[i].c1=_en1_X_t[i];
                out1_f_buf_func[i].c2=_out1_a_BX_f[i];
                out1_t_buf_func[i].c2=_out1_a_BX_t[i];
                out1_f_buf_func[i].n1=in.d.d[i].f;
                out1_t_buf_func[i].n1=in.d.d[i].t;
                out1_f_buf_func[i].vdd=supply.vdd;
                out1_t_buf_func[i].vdd=supply.vdd;
                out1_f_buf_func[i].vss=supply.vss;
                out1_t_buf_func[i].vss=supply.vss;
                out1_t_buf_func[i].pr_B = _reset_BXX[i];
                out1_t_buf_func[i].sr_B = _reset_BXX[i];
                out1_f_buf_func[i].pr_B = _reset_BXX[i];
                out1_f_buf_func[i].sr_B = _reset_BXX[i];
            )
            //func buffer out2
            bool _out2_a_BX_t[N],_out2_a_BX_f[N],_out2_a_B,_en2_X_t[N],_en2_X_f[N];
            A_2C1N_RB_X4 out2_f_buf_func[N];
            A_2C1N_RB_X4 out2_t_buf_func[N];
            sigbuf<N> out2_en_buf_t(.in=_en, .out=_en2_X_t, .supply=supply);
            sigbuf<N> out2_en_buf_f(.in=_en, .out=_en2_X_f, .supply=supply);
            INV_X1 out2_a_inv(.a=out2.a,.y=_out2_a_B);
            sigbuf<N> out2_a_B_buf_f(.in=_out2_a_B,.out=_out2_a_BX_t);
            sigbuf<N> out2_a_B_buf_t(.in=_out2_a_B,.out=_out2_a_BX_f);
            (i:N: 
                out2_f_buf_func[i].y=out2.d.d[i].f;
                out2_t_buf_func[i].y=out2.d.d[i].t;
                out2_f_buf_func[i].c1=_en2_X_f[i];
                out2_t_buf_func[i].c1=_en2_X_t[i];
                out2_f_buf_func[i].c2=_out2_a_BX_f[i];
                out2_t_buf_func[i].c2=_out2_a_BX_t[i];
                out2_f_buf_func[i].n1=in.d.d[i].f;
                out2_t_buf_func[i].n1=in.d.d[i].t;
                out2_f_buf_func[i].vdd=supply.vdd;
                out2_t_buf_func[i].vdd=supply.vdd;
                out2_f_buf_func[i].vss=supply.vss;
                out2_t_buf_func[i].vss=supply.vss;
                out2_t_buf_func[i].pr_B = _reset_BXX[i+N-1];
                out2_t_buf_func[i].sr_B = _reset_BXX[i+N-1];
                out2_f_buf_func[i].pr_B = _reset_BXX[i+N-1];
                out2_f_buf_func[i].sr_B = _reset_BXX[i+N-1];
            )
        }
    }
}