bics/actlib_dataflow_neuro
20
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

vastly improved lazy synapse handshakes

Browse Source
This commit is contained in:
alexmadison 2022-03-31 12:44:00 +02:00
parent cd5d41d7f8
commit 2e4cdd5029
1 changed files with 156 additions and 193 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

349
dataflow_neuro/coders.act
View File

@ -53,223 +53,186 @@ namespace tmpl {
* Nc is the number of dualrail input channels. * Nc is the number of dualrail input channels.
* Then builds N output AND gates, connecting to the right input wires. * Then builds N output AND gates, connecting to the right input wires.
*/ */
export template<pint Nc, N> export template<pint Nc, N>
defproc decoder_dualrail (Mx1of2<Nc> in; bool? out[N]; power supply) { defproc decoder_dualrail (Mx1of2<Nc> in; bool? out[N]; power supply) {
// signal buffers // signal buffers
sigbuf<N> in_tX[Nc]; sigbuf<N> in_tX[Nc];
sigbuf<N> in_fX[Nc]; sigbuf<N> in_fX[Nc];
(i:Nc: (i:Nc:
in_tX[i].supply = supply; in_tX[i].supply = supply;
in_tX[i].in = in.d[i].t; in_tX[i].in = in.d[i].t;
in_fX[i].supply = supply; in_fX[i].supply = supply;
in_fX[i].in = in.d[i].f; in_fX[i].in = in.d[i].f;
)
// AND trees
pint bitval;
andtree<Nc> atree[N];
(k:0..N-1:atree[k].supply = supply;)
(i:0..N-1:
(j:0..Nc-1:
bitval = (i & ( 1 << j )) >> j; // Get binary digit of integer i, column j
[bitval = 1 ->
atree[i].in[j] = in_tX[j].out[i];
// atree[i].in[j] = addr_buf.out.d.d[j].t;
[]bitval = 0 ->
atree[i].in[j] = in_fX[j].out[i];
// atree[i].in[j] = addr_buf.out.d.d[j].f;
[]bitval >= 2 -> {false : "fuck"};
]
atree[i].out = out[i];
) )
)
// AND trees }
pint bitval;
andtree<Nc> atree[N];
(k:0..N-1:atree[k].supply = supply;)
(i:0..N-1:
(j:0..Nc-1:
bitval = (i & ( 1 << j )) >> j; // Get binary digit of integer i, column j
[bitval = 1 ->
atree[i].in[j] = in_tX[j].out[i];
// atree[i].in[j] = addr_buf.out.d.d[j].t;
[]bitval = 0 ->
atree[i].in[j] = in_fX[j].out[i];
// atree[i].in[j] = addr_buf.out.d.d[j].f;
[]bitval >= 2 -> {false : "fuck"};
]
atree[i].out = out[i];
)
)
}
/** /**
* 2D decoder which uses a configurable delay from the VCtrees to buffer ack. * 2D decoder which uses a configurable delay from the VCtrees to buffer ack.
* Nx is the x size of the decoder array * Nx is the x size of the decoder array
* NxC is the number of wires in the x channel. * NxC is the number of wires in the x channel.
* Thus NxC should be something like NxC = ceil(log2(Nx)) * Thus NxC should be something like NxC = ceil(log2(Nx))
* but my guess is that we can't do logs... * but my guess is that we can't do logs...
* N_dly_cfg is the number of config bits in the ACK delay line, * 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 DLY4_X1 cells. * with all bits high corresponding to 2**N_dly_cfg -1 DLY4_X1 cells.
*/ */
export template<pint NxC, NyC, Nx, Ny, N_dly_cfg> export template<pint NxC, NyC, Nx, Ny, N_dly_cfg>
defproc decoder_2d_dly (avMx1of2<NxC+NyC> in; bool? outx[Nx], outy[Ny], defproc decoder_2d_dly (avMx1of2<NxC+NyC> in; bool? outx[Nx], outy[Ny],
dly_cfg[N_dly_cfg], reset_B; power supply) { dly_cfg[N_dly_cfg], reset_B; power supply) {
// Buffer to recieve concat(x,y) address packet // Buffer to recieve concat(x,y) address packet
buffer<NxC+NyC> addr_buf(.in = in, .reset_B = reset_B, .supply = supply); buffer<NxC+NyC> addr_buf(.in = in, .reset_B = reset_B, .supply = supply);
// Validity trees // Validity trees
vtree<NxC> vtree_x (.supply = supply); vtree<NxC> vtree_x (.supply = supply);
vtree<NyC> vtree_y (.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].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..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].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;) (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) // 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); 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; 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 = tielow.y, .s = dly_cfg, .supply = supply);
delayprog<N_dly_cfg> dly(.in = C2el.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 // ACK MAY HAVE BEEN DISCONNECTED HERE
// FOR TESTING PURPOSES // FOR TESTING PURPOSES
// !!!!!!!!!!!!!!!! // !!!!!!!!!!!!!!!!
dly.out = addr_buf.out.a; dly.out = addr_buf.out.a;
// ACK MAY HAVE BEEN DISCONNECTED HERE // ACK MAY HAVE BEEN DISCONNECTED HERE
// FOR TESTING PURPOSES // FOR TESTING PURPOSES
// !!!!!!!!!!!!!!!! // !!!!!!!!!!!!!!!!
// Decoder X/Y And trees // Decoder X/Y And trees
decoder_dualrail<NxC,Nx> d_dr_x(.out = outx, .supply = supply); decoder_dualrail<NxC,Nx> d_dr_x(.out = outx, .supply = supply);
(i:0..NxC-1:d_dr_x.in.d[i] = addr_buf.out.d.d[i];) (i:0..NxC-1:d_dr_x.in.d[i] = addr_buf.out.d.d[i];)
decoder_dualrail<NyC,Ny> d_dr_y(.out = outy, .supply = supply); decoder_dualrail<NyC,Ny> d_dr_y(.out = outy, .supply = supply);
(i:0..NyC-1:d_dr_y.in.d[i] = addr_buf.out.d.d[i+NxC];) (i:0..NyC-1:d_dr_y.in.d[i] = addr_buf.out.d.d[i+NxC];)
} }
export template<pint Nx, Ny> export template<pint Nx, Ny>
defproc and_grid(bool! out[Nx*Ny]; bool? inx[Nx], iny[Ny]; power supply) { defproc and_grid(bool! out[Nx*Ny]; bool? inx[Nx], iny[Ny]; power supply) {
AND2_X1 ands[Nx*Ny]; AND2_X1 ands[Nx*Ny];
(i:0..Nx*Ny-1:ands[i].vss = supply.vss; ands[i].vdd = supply.vdd;) (i:0..Nx*Ny-1:ands[i].vss = supply.vss; ands[i].vdd = supply.vdd;)
(x:0..Nx-1: (x:0..Nx-1:
(y:0..Ny-1: (y:0..Ny-1:
ands[x + y*Nx].a = inx[x]; ands[x + y*Nx].a = inx[x];
ands[x + y*Nx].b = iny[y]; ands[x + y*Nx].b = iny[y];
ands[x + y*Nx].y = out[x + y*Nx]; ands[x + y*Nx].y = out[x + y*Nx];
)
)
}
/**
* 2D decoder which uses synapse handshaking using line pulldowns.
* Nx is the x size of the decoder array
* NxC is the number of wires in the x channel.
* but my guess is that we can't do logs...
* the req on a1of1 out is the req to each synapse.
* The ack back from each line should go high when the synapse is charged.
* N_dly is a hard coded delay of the pull down circuit.
* It can be set to 0.
*/
export template<pint NxC, NyC, Nx, Ny, N_dly>
defproc decoder_2d_hs (avMx1of2<NxC+NyC> in; a1of1 out[Nx*Ny]; bool? 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);
// Decoder X/Y And trees
decoder_dualrail<NxC,Nx> d_dr_x(.supply = supply);
(i:0..NxC-1:d_dr_x.in.d[i] = addr_buf.out.d.d[i];)
decoder_dualrail<NyC,Ny> d_dr_y(.supply = supply);
(i:0..NyC-1:d_dr_y.in.d[i] = addr_buf.out.d.d[i+NxC];)
// Validity
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;)
A_2C_B_X1 C2el(.c1 = vtree_x.out, .c2 = vtree_y.out, .y = addr_buf.out.v,
.vdd = supply.vdd, .vss = supply.vss);
// and grid for reqs into synapses
and_grid<Nx, Ny> _and_grid(.inx = d_dr_x.out, .iny = d_dr_y.out, .supply = supply);
(i:Nx*Ny: out[i].r = _and_grid.out[i];)
// Acknowledge pull down time
// Pull DOWNs on the reqB lines by synapses (easier to invert).
bool _out_reqsB[Nx], _out_acksB[Nx]; // The vertical output ack lines from each syn.
PULLDOWN2_X4 req_pulldowns[Nx*Ny];
pint index;
(i:Nx:
(j:Ny:
index = i + Nx*j;
req_pulldowns[index].a = out[index].a;
req_pulldowns[index].b = _out_acksB[i];
req_pulldowns[index].y = _out_reqsB[i];
req_pulldowns[index].vss = supply.vss;
req_pulldowns[index].vdd = supply.vdd;
)
) )
)
// ReqB keep cells }
KEEP_X1 req_keeps[Nx];
(i:Nx:
req_keeps[i].y = _out_reqsB[i];
req_keeps[i].vdd = supply.vdd;
req_keeps[i].vss = supply.vss;
)
// req-ack buffers
// Delay needed here, since otherwise the pull up of reqB happens too quickly.
// Means that the pull up may start fighting the synapse,
// since the synapse has not yet retracted its ack.
// Also there is the possibility, if really fast, that the line pull up block
// doesn't yet see that the input is valid, and starts pulling up.
// In any case, this delay is important.
sigbuf<Ny> req_bufs[Nx];
delay_chain<N_dly> ack_delays[Nx];
(i:Nx:
ack_delays[i].in = _out_reqsB[i];
ack_delays[i].supply = supply;
// req_bufs[i].in = _out_reqsB[i];
req_bufs[i].in = ack_delays[i].out;
req_bufs[i].out[0] = _out_acksB[i]; // DANGER DANGER
req_bufs[i].supply = supply;
/**
* 2D decoder which uses synapse handshaking using line pulldowns.
* Nx is the x size of the decoder array
* NxC is the number of wires in the x channel.
* but my guess is that we can't do logs...
* the req on a1of1 out is the req to each synapse.
* The ack back from each line should go high when the synapse is charged.
* N_dly is a hard coded delay of the pull down circuit.
* It can be set to 0.
*/
export template<pint NxC, NyC, Nx, Ny, N_dly>
defproc decoder_2d_hs (avMx1of2<NxC+NyC> in; a1of1 out[Nx*Ny]; bool? 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);
// Line end pull UPs (triggered once synapse reqs removed) // Decoder X/Y And trees
OR2_X1 pu_ORs[Nx]; decoder_dualrail<NxC,Nx> d_dr_x(.supply = supply);
PULLUP_X4 pu[Nx]; // TODO probably replace this with variable strength PU (i:0..NxC-1:d_dr_x.in.d[i] = addr_buf.out.d.d[i];)
AND2_X1 pu_ANDs[Nx]; decoder_dualrail<NyC,Ny> d_dr_y(.supply = supply);
(i:Nx: (i:0..NyC-1:d_dr_y.in.d[i] = addr_buf.out.d.d[i+NxC];)
pu_ORs[i].a = _out_acksB[i];
pu_ORs[i].b = d_dr_x.out[i];
pu_ORs[i].vdd = supply.vdd;
pu_ORs[i].vss = supply.vss;
pu_ANDs[i].a = pu_ORs[i].y; // Validity
pu_ANDs[i].b = reset_B; // TODO buffer vtree<NxC> vtree_x (.supply = supply);
pu_ANDs[i].vdd = supply.vdd; vtree<NyC> vtree_y (.supply = supply);
pu_ANDs[i].vss = supply.vss; (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;)
A_2C_B_X1 C2el(.c1 = vtree_x.out, .c2 = vtree_y.out, .y = addr_buf.out.v,
.vdd = supply.vdd, .vss = supply.vss);
pu[i].a = pu_ANDs[i].y;
pu[i].y = _out_reqsB[i];
pu[i].vdd = supply.vdd;
pu[i].vss = supply.vss;
)
// ORtree from all output reqs, back to the buffer ack. // and grid for reqs into synapses
// This is instead of the ack that came from the delayed validity trees, and_grid<Nx, Ny> _and_grid(.inx = d_dr_x.out, .iny = d_dr_y.out, .supply = supply);
// in decoder_2d_dly. (i:Nx*Ny: out[i].r = _and_grid.out[i];)
ortree<Nx> _ortree(.out = addr_buf.out.a, .supply = supply);
INV_X1 out_req_invs[Nx];
(i:Nx:
out_req_invs[i].a = _out_reqsB[i];
out_req_invs[i].vdd = supply.vdd;
out_req_invs[i].vss = supply.vss;
_ortree.in[i] = out_req_invs[i].y; // Acknowledge pull down time
)
}
// Pull DOWNs on the ackB lines by synapses (easier to invert).
bool _out_acksB[Nx]; // The vertical output ack lines from each syn.
PULLDOWN2_X4 ack_pulldowns[Nx*Ny];
pint index;
(i:Nx:
(j:Ny:
index = i + Nx*j;
ack_pulldowns[index].a = out[index].a;
ack_pulldowns[index].b = d_dr_x.out[i];
ack_pulldowns[index].y = _out_acksB[i];
ack_pulldowns[index].vss = supply.vss;
ack_pulldowns[index].vdd = supply.vdd;
)
)
// Line end pull UPs (triggered once reqs removed)
PULLUP_X4 pu[Nx]; // TODO probably replace this with variable strength PU
AND2_X1 pu_ANDs[Nx];
(i:Nx:
pu_ANDs[i].a = d_dr_x.out[i];
pu_ANDs[i].b = reset_B; // TODO buffer
pu_ANDs[i].vdd = supply.vdd;
pu_ANDs[i].vss = supply.vss;
pu[i].a = pu_ANDs[i].y;
pu[i].y = _out_acksB[i];
pu[i].vdd = supply.vdd;
pu[i].vss = supply.vss;
)
// ORtree from all output acks, back to the buffer ack.
// This is instead of the ack that came from the delayed validity trees,
// in decoder_2d_dly.
ortree<Nx> _ortree(.out = addr_buf.out.a, .supply = supply);
INV_X1 out_ack_invs[Nx];
(i:Nx:
out_ack_invs[i].a = _out_acksB[i];
out_ack_invs[i].vdd = supply.vdd;
out_ack_invs[i].vss = supply.vss;
_ortree.in[i] = out_ack_invs[i].y;
)
}