From 6dfc941993a92cea2d2dd94912ec40fc979b4b75 Mon Sep 17 00:00:00 2001
From: alexmadison <git@mdsn.co.uk>
Date: Thu, 31 Mar 2022 15:10:47 +0200
Subject: [PATCH] added Cel to buf out ack, so that ack can't be removed before
 the bufs output data is fully invalidated

---
 dataflow_neuro/coders.act | 40 +++++++++++++++++++++++++++++++++++----
 1 file changed, 36 insertions(+), 4 deletions(-)

diff --git a/dataflow_neuro/coders.act b/dataflow_neuro/coders.act
index 7e3730f..14a6264 100644
--- a/dataflow_neuro/coders.act
+++ b/dataflow_neuro/coders.act
@@ -130,12 +130,24 @@ defproc decoder_2d_dly (avMx1of2<NxC+NyC> in; bool? outx[Nx], outy[Ny],
 
 export template<pint Nx, Ny>
 defproc and_grid(bool! out[Nx*Ny]; bool? inx[Nx], iny[Ny]; power supply) {
+	// Buffer inputs
+	sigbuf<Ny> xbuf[Nx];
+	sigbuf<Nx> ybuf[Ny];
+	(i:Nx:
+		xbuf[i].in = inx[i];
+		xbuf[i].supply = supply;
+	)
+	(i:Ny:
+		ybuf[i].in = iny[i];
+		ybuf[i].supply = supply;
+	)
+
 	AND2_X1 ands[Nx*Ny];
 	(i:0..Nx*Ny-1:ands[i].vss = supply.vss; ands[i].vdd = supply.vdd;)
 	(x:0..Nx-1:
 		(y:0..Ny-1:
-			ands[x + y*Nx].a = inx[x];
-			ands[x + y*Nx].b = iny[y];
+			ands[x + y*Nx].a = xbuf[x].out[y];
+			ands[x + y*Nx].b = ybuf[y].out[x];
 			ands[x + y*Nx].y = out[x + y*Nx];
 			)
 		)
@@ -164,6 +176,18 @@ defproc decoder_2d_hs (avMx1of2<NxC+NyC> in; a1of1 out[Nx*Ny]; bool? reset_B; po
 	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];)
 
+	// Buffer dualrail decoder outputs
+	// sigbuf<Ny*2> d_dr_xX[Nx]; // x2 because X req lines go to twice the num of gates
+	// sigbuf<Nx> d_dr_yX[Ny];
+	// (i:Nx:
+	// 	d_dr_xX[i].in = d_dr_x.out[i];
+	// 	d_dr_xX[i].supply = supply;
+	// )
+	// (i:Ny:
+	// 	d_dr_yX[i].in = d_dr_y.out[i];
+	// 	d_dr_yX[i].supply = supply;
+	// )
+
 	// Validity 
 	vtree<NxC> vtree_x (.supply = supply);
 	vtree<NyC> vtree_y (.supply = supply);
@@ -171,7 +195,7 @@ defproc decoder_2d_hs (avMx1of2<NxC+NyC> in; a1of1 out[Nx*Ny]; bool? reset_B; po
 	(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,
+	A_2C_B_X1 valid_Cel(.c1 = vtree_x.out, .c2 = vtree_y.out, .y = addr_buf.out.v,
 		.vdd = supply.vdd, .vss = supply.vss);
 
 
@@ -214,7 +238,7 @@ defproc decoder_2d_hs (avMx1of2<NxC+NyC> in; a1of1 out[Nx*Ny]; bool? reset_B; po
 	// 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);
+	ortree<Nx>  _ortree(.supply = supply);
 	INV_X1 out_ack_invs[Nx];
 	(i:Nx:
 		out_ack_invs[i].a = _out_acksB[i];
@@ -223,6 +247,14 @@ defproc decoder_2d_hs (avMx1of2<NxC+NyC> in; a1of1 out[Nx*Ny]; bool? reset_B; po
 
 		_ortree.in[i] = out_ack_invs[i].y;
 	)
+
+	// NEED TO ADD C EL
+	// SUCH THAT ACK IS LOWERED ONLY WHEN
+	// VAL IS LOWERED TOO
+	// i.e. buffer.out.a = Cel(ortree_ack, valid)
+	A_2C_B_X1 buf_ack_Cel(.c1 = _ortree.out, .c2 = valid_Cel.y, .y = addr_buf.out.a,
+		.vdd = supply.vdd, .vss = supply.vss);
+
 }