251 lines
6.8 KiB
Plaintext
251 lines
6.8 KiB
Plaintext
/*************************************************************************
|
|
*
|
|
* 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
|
|
* Copyright (c) 2022 University of Groningen - Madison Cotteret
|
|
*
|
|
*
|
|
* 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 "../../dataflow_neuro/primitives.act";
|
|
// import tmpl::dataflow_neuro;
|
|
// import tmpl::dataflow_neuro;
|
|
import std::channel;
|
|
open std::channel;
|
|
|
|
namespace tmpl {
|
|
namespace dataflow_neuro {
|
|
|
|
/**
|
|
* 2D decoder which uses a configurable delay from the VCtrees to buffer ack.
|
|
* Nx is the x size of the decoder array
|
|
* NxC is the number of wires in the x channel.
|
|
* Thus NxC should be something like NxC = ceil(log2(Nx))
|
|
* but my guess is that we can't do logs...
|
|
* N_dly_cfg is the number of config bits in the ACK delay line,
|
|
* with all bits high corresponding to 2**N_dly_cfg -1 DLY1_X4 cells.
|
|
*/
|
|
export template<pint NxC, NyC, Nx, Ny, N_dly_cfg>
|
|
defproc decoder_2d_dly (avMx1of2<NxC+NyC> in; bool? outx[Nx], outy[Ny],
|
|
dly_cfg[N_dly_cfg], reset_B; power supply) {
|
|
|
|
// Buffer to recieve concat(x,y) address packet
|
|
buffer<NxC+NyC> addr_buf(.in = in, .reset_B = reset_B, .supply = supply);
|
|
// NEED TO BUFFER OUTPUTS FROM BUFFER I RECKON
|
|
|
|
// Validity trees
|
|
vtree<NxC> vtree_x (.supply = supply);
|
|
vtree<NyC> vtree_y (.supply = supply);
|
|
(i:0..NxC-1:vtree_x.in.d[i].t = addr_buf.out.d.d[i].t;)
|
|
(i:0..NxC-1:vtree_x.in.d[i].f = addr_buf.out.d.d[i].f;)
|
|
(i:0..NyC-1:vtree_y.in.d[i].t = addr_buf.out.d.d[i+NxC].t;)
|
|
(i:0..NyC-1:vtree_y.in.d[i].f = addr_buf.out.d.d[i+NxC].f;)
|
|
|
|
|
|
// Delay ack line. Ack line is delayed (but not the val)
|
|
A_2C_B_X1 C2el(.c1 = vtree_x.out, .c2 = vtree_y.out, .vdd = supply.vdd, .vss = supply.vss);
|
|
addr_buf.out.v = C2el.y;
|
|
|
|
// delayprog<N_dly_cfg> dly(.in = tielow.y, .s = dly_cfg, .supply = supply);
|
|
delayprog<N_dly_cfg> dly(.in = C2el.y, .s = dly_cfg, .supply = supply);
|
|
|
|
// ACK MAY HAVE BEEN DISCONNECTED HERE
|
|
// FOR TESTING PURPOSES
|
|
// !!!!!!!!!!!!!!!!
|
|
dly.out = addr_buf.out.a;
|
|
// ACK MAY HAVE BEEN DISCONNECTED HERE
|
|
// FOR TESTING PURPOSES
|
|
// !!!!!!!!!!!!!!!!
|
|
|
|
// AND trees
|
|
pint bitval;
|
|
andtree<NxC> atree_x[Nx];
|
|
(k:0..Nx-1:atree_x[k].supply = supply;)
|
|
(i:0..Nx-1:
|
|
(j:0..NxC-1:
|
|
bitval = (i & ( 1 << j )) >> j; // Get binary digit of integer i, column j
|
|
[bitval = 1 ->
|
|
atree_x[i].in[j] = addr_buf.out.d.d[j].t;
|
|
[]bitval = 0 ->
|
|
atree_x[i].in[j] = addr_buf.out.d.d[j].f;
|
|
[]bitval >= 2 -> {false : "fuck"};
|
|
]
|
|
atree_x[i].out = outx[i];
|
|
)
|
|
)
|
|
|
|
andtree<NyC> atree_y[Ny];
|
|
(k:0..Ny-1:atree_y[k].supply = supply;)
|
|
(i:0..Ny-1:
|
|
(j:0..NyC-1:
|
|
bitval = (i & ( 1 << j )) >> j; // Get binary digit of integer i, column j
|
|
[bitval = 1 ->
|
|
atree_y[i].in[j] = addr_buf.out.d.d[j+NxC].t;
|
|
[]bitval = 0 ->
|
|
atree_y[i].in[j] = addr_buf.out.d.d[j+NxC].f;
|
|
]
|
|
atree_y[i].out = outy[i];
|
|
)
|
|
)
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
* Build an arbiter_handshake tree.
|
|
*/
|
|
export template<pint N>
|
|
defproc arbtree (a1of1 in[N]; a1of1 out; power supply)
|
|
{
|
|
bool tout;
|
|
|
|
{ N > 0 : "What?" };
|
|
|
|
pint i, end, j;
|
|
i = 0;
|
|
end = N-1;
|
|
|
|
pint arbCount;
|
|
arbCount = 0;
|
|
/* Pre"calculate" the number of C cells required, look below if confused */
|
|
*[ i != end ->
|
|
j = 0;
|
|
*[ i <= end ->
|
|
j = j + 1;
|
|
[i = end ->
|
|
i = end+1;
|
|
[] i+1 = end ->
|
|
i = end+1;
|
|
arbCount = arbCount +1;
|
|
[] else ->
|
|
i = i + 2;
|
|
arbCount = arbCount +1;
|
|
]
|
|
]
|
|
/*-- update range that has to be combined --*/
|
|
// i = end+1;
|
|
end = end+j;
|
|
]
|
|
|
|
/* array that holds ALL the nodes in the completion tree */
|
|
a1of1 tmp[end+1];
|
|
|
|
// Connecting the first nodes to the input
|
|
(l:N:
|
|
tmp[l] = in[l];
|
|
)
|
|
|
|
/* array to hold the actual C-elments, either A2C or A3C */
|
|
[arbCount > 0 ->
|
|
arbiter_handshake arbs[arbCount];
|
|
]
|
|
(h:arbCount:arbs[h].supply = supply;)
|
|
|
|
/* Reset the variables we just stole lol */
|
|
i = 0;
|
|
end = N-1;
|
|
j = 0;
|
|
pint arbIndex = 0;
|
|
|
|
/* Invariant: i <= end */
|
|
|
|
*[ i != end ->
|
|
/*
|
|
* Invariant: tmp[i..end] has the current signals that need to be
|
|
* combined together, and "isinv" specifies if they are the inverted
|
|
* sense or not
|
|
*/
|
|
j = 0;
|
|
*[ i <= end ->
|
|
/*-- there are still signals that need to be combined --*/
|
|
j = j + 1;
|
|
[ i = end ->
|
|
/*-- last piece: pipe input through to next layer --*/
|
|
tmp[end+j] = tmp[i];
|
|
i = end+1;
|
|
[] i+1 = end ->
|
|
/*-- last piece: use either a 2 input C-element --*/
|
|
arbs[arbIndex].in1 = tmp[i];
|
|
arbs[arbIndex].in2 = tmp[i+1];
|
|
arbs[arbIndex].out = tmp[end+j];
|
|
arbIndex = arbIndex +1;
|
|
i = end+1;
|
|
[] else ->
|
|
/*-- more to come; so use a two input C-element --*/
|
|
arbs[arbIndex].in1 = tmp[i];
|
|
arbs[arbIndex].in2 = tmp[i+1];
|
|
arbs[arbIndex].out = tmp[end+j];
|
|
arbIndex = arbIndex +1;
|
|
i = i + 2;
|
|
]
|
|
]
|
|
/*-- update range that has to be combined --*/
|
|
i = end+1;
|
|
end = end+j;
|
|
j = 0;
|
|
]
|
|
|
|
out = tmp[end];
|
|
|
|
}
|
|
|
|
|
|
template<pint N, pint M, pint ACK_STRENGTH>
|
|
defproc encoder2D(a1of1 x[N]; a1of1 y[M] ;avMx1of2<X> addr; bool! out_a; power supply)
|
|
{
|
|
// Arbiters
|
|
a1of1 _out_arb_x,_out_arb_y;
|
|
a1of1 _x_temp[N];
|
|
(i:N:
|
|
_x_temp[i].r = x[i].r;
|
|
)
|
|
(i:M:
|
|
_y_temp[i].r = y[i].r;
|
|
)
|
|
arbtree<N> Xarb(.in = _x_temp,.out = _out_arb_X,.supply = supply);
|
|
arbtree<M> Yarb(.in = _y_temp,.out = _out_arb_Y,.supply = supply);
|
|
|
|
sigbuf<ACK_STRENGTH> x_ack_arb[N];
|
|
sigbuf<ACK_STRENGTH> y_ack_arb[M];
|
|
(i:N:
|
|
x_ack_arb[i].in = _x_temp[i].a;
|
|
x_ack_arb[i].out[0] = x[i].a;
|
|
x_ack_arb[i].supply = supply;
|
|
)
|
|
(i:M:
|
|
y_ack_arb[i].in = _y_temp[i].a;
|
|
y_ack_arb[i].out[0] = y[i].a;
|
|
y_ack_arb[i].supply = supply;
|
|
)
|
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
}
|
|
|
|
} |