bics
/
actlib_dataflow_neuro
18
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

commented nrn hs timing assumption, added encoder1d not tested

This commit is contained in:
alexmadison 2022-04-11 18:36:30 +02:00
parent c1ea7be83d
commit b21e64a22a
2 changed files with 52 additions and 19 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
dataflow_neuro/chips.act
View File

@ -106,7 +106,7 @@ defproc chip_texel (bd<N_IN> in, out;
NC_NRN = NC_NRN_X + NC_NRN_Y;
nrn_hs_2d_array<N_NRN_X,N_NRN_Y,N_LINE_PD_DLY> nrn_grid(.in = neurons,
.supply = supply, .reset_B = reset_B);
encoder2d_simple<NC_NRN_X, NC_NRN_Y, N_NRN_X, N_NRN_Y, 16> encoder(
encoder2d_simple<NC_NRN_X, NC_NRN_Y, N_NRN_X, N_NRN_Y> encoder(
.inx = nrn_grid.outx,
.iny = nrn_grid.outy,
.reset_B = reset_B, .supply = supply

69
dataflow_neuro/coders.act
View File

@ -774,7 +774,7 @@ defproc decoder_2d_hybrid (avMx1of2<NxC+NyC> in; a1of1 out[Nx*Ny]; bool? dly_cfg
}
export template<pint NxC, NyC, Nx, Ny, ACK_STRENGTH>
export template<pint NxC, NyC, Nx, Ny>
defproc encoder2d_simple(a1of1 inx[Nx]; a1of1 iny[Ny]; avMx1of2<(NxC + NyC)> out;
power supply; bool reset_B) {
@ -827,6 +827,42 @@ defproc decoder_2d_hybrid (avMx1of2<NxC+NyC> in; a1of1 out[Nx*Ny]; bool? dly_cfg
}
export template<pint Nc, N>
defproc encoder1d_simple(a1of1 in[N]; avMx1of2<Nc> out;
power supply; bool reset_B) {
bool _a_x, _r_x;
bool _r_x_B;
buffer<Nc> buf(.out = out, .supply = supply, .reset_B = reset_B);
// Arbiters
arbtree<N> Xarb(.supply = supply);
Xarb.out.a = _a_x;
Xarb.out.r = _r_x;
// Encoders
dualrail_encoder<Nc, N> Xenc(.supply = supply);
// Wire up inputs to encoders and arb
(i:N:
Xarb.in[i].r = in[i].r;
Xarb.in[i].a = in[i].a;
Xenc.in[i] = in[i].a;
)
INV_X2 inv_buf(.a = buf.in.a, .vdd = supply.vdd, .vss = supply.vss);
A_2C_RB_X1 a_x_Cel(.c1 = inv_buf.y, .c2 = _r_x, .y = _a_x,
.sr_B = reset_B, .pr_B = reset_B, .vdd = supply.vdd, .vss = supply.vss);
// Wire up encoder to buffer
(i:Nc:
Xenc.out.d[i] = buf.in.d.d[i];
)
}
/**
* Neuron handshaking.
@ -839,14 +875,7 @@ defproc decoder_2d_hybrid (avMx1of2<NxC+NyC> in; a1of1 out[Nx*Ny]; bool? dly_cfg
BUF_X2 reset_buf(.a = reset_B, .y = _reset_BX, .vdd = supply.vdd, .vss = supply.vss);
bool _en, _req;
// A_1C2N_RB_X1 A_ack(.c1 = _en, .n1 = _req, .n2 = in.r, .y = in.a,
// .pr_B = _reset_BX, .sr_B = _reset_BX, .vss = supply.vss, .vdd = supply.vdd);
// Switched it back
// Because had the problem that if the req was not removed in time,
// it would be recounted as a double spike,
// since in.req is still high after the out has been dealt with.
A_2C1N_RB_X1 A_ack(.c1 = _en, .c2 = in.r, .n1 = _req, .y = in.a,
.pr_B = _reset_BX, .sr_B = _reset_BX, .vss = supply.vss, .vdd = supply.vdd);
@ -858,19 +887,23 @@ defproc decoder_2d_hybrid (avMx1of2<NxC+NyC> in; a1of1 out[Nx*Ny]; bool? dly_cfg
INV_X2 inv_x(.a = outx.a, .y = _x_a_B, .vss = supply.vss, .vdd = supply.vdd);
INV_X2 inv_y(.a = outy.a, .y = _y_a_B, .vss = supply.vss, .vdd = supply.vdd);
// WARNUNG WARNUNG WARNUNG //
// This neuron hs design has a fat timing assumption.
// Say that the neuron has sent out both reqs, and is now receiving the acks.
// _x_a_B and _y_a_B are then low, and _req starts to be pulled down to reset the hs.
// However, if the line pull downs at the end of the neuron row/column are fast enough,
// then seeing the high acks, they will pull the ack lines down. If the arbiter tree
// is sufficiently fast enough, then it will remove the ack lines.
// If this cell were rather tardy, then _req's pd would be cancelled midway,
// it missed its window of opportunity to switch, and would probably make the system hang.
// Or starts oscillating with the line pull down and goes brrrrapppppppp.
// This issue may be somewhat unavoidable, as from a black box perspective,
// we are giving the neuron acks, but then not listening to it at all to check
// that it has had time to act upon these acks.
A_2C1P1N_RB_X1 A_req(.p1 = _x_a_B, .c1 = _en, .c2 = _y_a_B, .n1 = in.r, .y = _req,
.pr_B = _reset_BX, .sr_B = _reset_BX, .vdd = supply.vdd, .vss = supply.vss);
// // y_req pull up
// NAND2_X1 nand_y(.a = _y_a_B, .b = _req, .vdd = supply.vdd, .vss = supply.vss);
// A_1P_U_X4 pu_y(.a = nand_y.y, .y = outy.r, .vdd = supply.vdd, .vss = supply.vss);
// // x_req pull up
// NAND3_X1 nand_x(.a = _x_a_B, .b = _req, .c = outy.a, .vdd = supply.vdd, .vss = supply.vss);
// A_1P_U_X4 pu_x(.a = nand_x.y, .y = outx.r, .vdd = supply.vdd, .vss = supply.vss);
// Better version with fewer timing assumptions
// Core change is that the out acks stop the pullups without any delay.
// y_req pull up
bool _reqB;
INV_X1 req_inv(.a = _req, .y = _reqB, .vdd= supply.vdd, .vss = supply.vss);