Merge pull request #1648 from YosysHQ/eddie/cmp2lcu

"techmap -map +/cmp2lcu.v" for decomposing arithmetic compares to $lcu

6 files changed, 173 insertions, 13 deletions

diff --git a/techlibs/common/Makefile.inc b/techlibs/common/Makefile.inc
index d5e69a241..7b1e4b430 100644
--- a/techlibs/common/Makefile.inc
+++ b/techlibs/common/Makefile.inc
@@ -30,3 +30,4 @@ $(eval $(call add_share_file,share,techlibs/common/cmp2lut.v))
 $(eval $(call add_share_file,share,techlibs/common/cells.lib))
 $(eval $(call add_share_file,share,techlibs/common/mul2dsp.v))
 $(eval $(call add_share_file,share,techlibs/common/abc9_model.v))
+$(eval $(call add_share_file,share,techlibs/common/cmp2lcu.v))
diff --git a/techlibs/common/cmp2lcu.v b/techlibs/common/cmp2lcu.v
new file mode 100644
index 000000000..b6f4aeed6
--- /dev/null
+++ b/techlibs/common/cmp2lcu.v
@@ -0,0 +1,116 @@
+// This pass performs an optimisation that decomposes wide arithmetic
+//   comparisons into LUT-size chunks (as guided by the `LUT_WIDTH
+//   macro) connected to a single lookahead-carry-unit $lcu cell,
+//   which is typically mapped to dedicated (and fast) FPGA
+//   carry-chains.
+(* techmap_celltype = "$lt $le $gt $ge" *)
+module _80_lcu_cmp_ (A, B, Y);
+
+parameter A_SIGNED = 0;
+parameter B_SIGNED = 0;
+parameter A_WIDTH = 0;
+parameter B_WIDTH = 0;
+parameter Y_WIDTH = 0;
+
+input [A_WIDTH-1:0] A;
+input [B_WIDTH-1:0] B;
+output [Y_WIDTH-1:0] Y;
+
+parameter _TECHMAP_CELLTYPE_ = "";
+
+generate
+    if (_TECHMAP_CELLTYPE_ == "" || `LUT_WIDTH < 2)
+        wire _TECHMAP_FAIL_ = 1;
+    else if (_TECHMAP_CELLTYPE_ == "$lt") begin
+        // Transform $lt into $gt by swapping A and B
+        $gt #(.A_SIGNED(B_SIGNED), .B_SIGNED(A_SIGNED), .A_WIDTH(B_WIDTH), .B_WIDTH(A_WIDTH), .Y_WIDTH(Y_WIDTH)) _TECHMAP_REPLACE_ (.A(B), .B(A), .Y(Y));
+    end
+    else if (_TECHMAP_CELLTYPE_ == "$le") begin
+        // Transform $le into $ge by swapping A and B
+        $ge #(.A_SIGNED(B_SIGNED), .B_SIGNED(A_SIGNED), .A_WIDTH(B_WIDTH), .B_WIDTH(A_WIDTH), .Y_WIDTH(Y_WIDTH)) _TECHMAP_REPLACE_ (.A(B), .B(A), .Y(Y));
+    end
+    else begin
+        // Perform sign extension on A and B
+        localparam WIDTH = A_WIDTH > B_WIDTH ? A_WIDTH : B_WIDTH;
+        wire [WIDTH-1:0] AA = {{(WIDTH-A_WIDTH){A_SIGNED ? A[A_WIDTH-1] : 1'b0}}, A};
+        wire [WIDTH-1:0] BB = {{(WIDTH-B_WIDTH){B_SIGNED ? B[B_WIDTH-1] : 1'b0}}, B};
+        // For $ge operation, start with the assumption that A and B are
+        //   equal (propagating this equality if A and B turn out to be so)
+        if (_TECHMAP_CELLTYPE_ == "$ge")
+            localparam CI = 1'b1;
+        else
+            localparam CI = 1'b0;
+        $__CMP2LCU #(.AB_WIDTH(WIDTH), .AB_SIGNED(A_SIGNED && B_SIGNED), .LCU_WIDTH(1), .BUDGET(`LUT_WIDTH), .CI(CI))
+            _TECHMAP_REPLACE_ (.A(AA), .B(BB), .P(1'b1), .G(1'b0), .Y(Y));
+    end
+endgenerate
+endmodule
+
+module $__CMP2LCU (A, B, P, G, Y);
+
+parameter AB_WIDTH = 0;
+parameter AB_SIGNED = 0;
+parameter LCU_WIDTH = 1;
+parameter BUDGET = 0;
+parameter CI = 0;
+
+input [AB_WIDTH-1:0] A; // A from original $gt/$ge
+input [AB_WIDTH-1:0] B; // B from original $gt/$ge
+input [LCU_WIDTH-1:0] P; // P of $lcu
+input [LCU_WIDTH-1:0] G; // G of $lcu
+output Y;
+
+parameter [AB_WIDTH-1:0] _TECHMAP_CONSTMSK_A_ = 0;
+parameter [AB_WIDTH-1:0] _TECHMAP_CONSTMSK_B_ = 0;
+parameter [LCU_WIDTH-1:0] _TECHMAP_CONSTMSK_P_ = 0;
+
+generate
+    if (AB_WIDTH == 0) begin
+        wire [LCU_WIDTH-1:0] CO;
+        $lcu #(.WIDTH(LCU_WIDTH)) _TECHMAP_REPLACE_ (.P(P), .G(G), .CI(CI), .CO(CO));
+        assign Y = CO[LCU_WIDTH-1];
+    end
+    else begin
+        if (_TECHMAP_CONSTMSK_A_[AB_WIDTH-1:0] && _TECHMAP_CONSTMSK_B_[AB_WIDTH-1:0])
+            localparam COST = 0;
+        else if (_TECHMAP_CONSTMSK_A_[AB_WIDTH-1:0] || _TECHMAP_CONSTMSK_B_[AB_WIDTH-1:0])
+            localparam COST = 1;
+        else
+            localparam COST = 2;
+
+        if (BUDGET < COST)
+             $__CMP2LCU #(.AB_WIDTH(AB_WIDTH), .AB_SIGNED(AB_SIGNED), .LCU_WIDTH(LCU_WIDTH+1), .BUDGET(`LUT_WIDTH), .CI(CI))
+                _TECHMAP_REPLACE_ (.A(A), .B(B), .P({P, 1'b1}), .G({G, 1'b0}), .Y(Y));
+        else begin
+            wire PP, GG;
+            // Bit-wise equality (xnor) of A and B
+            assign PP = A[AB_WIDTH-1] ^~ B[AB_WIDTH-1];
+            if (AB_SIGNED)
+                assign GG = ~A[AB_WIDTH-1] & B[AB_WIDTH-1];
+            else if (_TECHMAP_CONSTMSK_P_[LCU_WIDTH-1]) // First compare for LUT if P (and G) is constant
+                assign GG = A[AB_WIDTH-1] & ~B[AB_WIDTH-1];
+            else
+                // Priority "encoder" that checks A[i] == 1'b1 && B[i] == 1'b0
+                //   from MSB down, deferring to less significant bits if the
+                //   MSBs are equal
+                assign GG = P[0] & (A[AB_WIDTH-1] & ~B[AB_WIDTH-1]);
+            if (LCU_WIDTH == 1) begin
+                // Propagate only if all pairs are equal
+                //   (inconclusive evidence to say A >= B)
+                wire P_ = P[0] & PP;
+                // Generate if any comparisons call for it
+                wire G_ = G[0] | GG;
+            end
+            else begin
+                // Propagate only if all pairs are equal
+                //   (inconclusive evidence to say A >= B)
+                wire [LCU_WIDTH-1:0] P_ = {P[LCU_WIDTH-1:1], P[0] & PP};
+                // Generate if any comparisons call for it
+                wire [LCU_WIDTH-1:0] G_ = {G[LCU_WIDTH-1:1], G[0] | GG};
+            end
+            $__CMP2LCU #(.AB_WIDTH(AB_WIDTH-1), .AB_SIGNED(1'b0), .LCU_WIDTH(LCU_WIDTH), .BUDGET(BUDGET-COST), .CI(CI))
+                _TECHMAP_REPLACE_ (.A(A[AB_WIDTH-2:0]), .B(B[AB_WIDTH-2:0]), .P(P_), .G(G_), .Y(Y));
+        end
+    end
+endgenerate
+endmodule
diff --git a/techlibs/common/cmp2lut.v b/techlibs/common/cmp2lut.v
index 1c8192b85..8ecd356cc 100644
--- a/techlibs/common/cmp2lut.v
+++ b/techlibs/common/cmp2lut.v
@@ -57,10 +57,6 @@ function automatic [(1 << `LUT_WIDTH)-1:0] gen_lut;
 				o_bit = (lhs >  rhs);
 			if (operation == 3)
 				o_bit = (lhs >= rhs);
-			if (operation == 4)
-				o_bit = (lhs == rhs);
-			if (operation == 5)
-				o_bit = (lhs != rhs);
 			gen_lut = gen_lut | (o_bit << n);
 		end
 	end
@@ -75,10 +71,6 @@ generate
 		localparam operation = 2;
 	if (_TECHMAP_CELLTYPE_ == "$ge")
 		localparam operation = 3;
-	if (_TECHMAP_CELLTYPE_ == "$eq")
-		localparam operation = 4;
-	if (_TECHMAP_CELLTYPE_ == "$ne")
-		localparam operation = 5;
 
 	if (A_WIDTH > `LUT_WIDTH || B_WIDTH > `LUT_WIDTH || Y_WIDTH != 1)
 		wire _TECHMAP_FAIL_ = 1;
diff --git a/techlibs/common/synth.cc b/techlibs/common/synth.cc
index e7a192c07..d6dffdd7f 100644
--- a/techlibs/common/synth.cc
+++ b/techlibs/common/synth.cc
@@ -225,9 +225,9 @@ struct SynthPass : public ScriptPass
 			run("peepopt");
 			run("opt_clean");
 			if (help_mode)
-				run("techmap -map +/cmp2lut.v", " (if -lut)");
-			else
-				run(stringf("techmap -map +/cmp2lut.v -D LUT_WIDTH=%d", lut));
+				run("techmap -map +/cmp2lut.v -map +/cmp2lcu.v", " (if -lut)");
+			else if (lut)
+				run(stringf("techmap -map +/cmp2lut.v -map +/cmp2lcu.v -D LUT_WIDTH=%d", lut));
 			if (!noalumacc)
 				run("alumacc", "  (unless -noalumacc)");
 			if (!noshare)
diff --git a/techlibs/xilinx/synth_xilinx.cc b/techlibs/xilinx/synth_xilinx.cc
index 8553efd6b..ac4f5bcf4 100644
--- a/techlibs/xilinx/synth_xilinx.cc
+++ b/techlibs/xilinx/synth_xilinx.cc
@@ -393,8 +393,6 @@ struct SynthXilinxPass : public ScriptPass
 				run("pmux2shiftx", "(skip if '-nosrl' and '-widemux=0')");
 				run("clean", "      (skip if '-nosrl' and '-widemux=0')");
 			}
-
-			run("techmap -map +/cmp2lut.v -D LUT_WIDTH=" + lut_size_s);
 		}
 
 		if (check_label("map_dsp", "(skip if '-nodsp')")) {
@@ -460,6 +458,7 @@ struct SynthXilinxPass : public ScriptPass
 		}
 
 		if (check_label("coarse")) {
+			run("techmap -map +/cmp2lut.v -map +/cmp2lcu.v -D LUT_WIDTH=" + lut_size_s);
 			run("alumacc");
 			run("share");
 			run("opt");
diff --git a/tests/techmap/cmp2lcu.ys b/tests/techmap/cmp2lcu.ys
new file mode 100644
index 000000000..7c8a63692
--- /dev/null
+++ b/tests/techmap/cmp2lcu.ys
@@ -0,0 +1,52 @@
+read_verilog <<EOT
+module top(input [12:0] a, b, output gtu, gts, ltu, lts, geu, ges, leu, les);
+assign gtu = a > b;
+assign gts = $signed(a) > $signed(b);
+assign ltu = a < b;
+assign lts = $signed(a) < $signed(b);
+assign geu = a >= b;
+assign ges = $signed(a) >= $signed(b);
+assign leu = a <= b;
+assign les = $signed(a) <= $signed(b);
+endmodule
+EOT
+
+equiv_opt -assert techmap -map +/cmp2lcu.v -D LUT_WIDTH=6
+design -load postopt
+select -assert-count 8 t:$lcu r:WIDTH=5 %i
+select -assert-none t:$gt t:$ge t:$lt t:$le
+
+design -load preopt
+equiv_opt -assert techmap -map +/cmp2lcu.v -D LUT_WIDTH=4
+design -load postopt
+select -assert-count 8 t:$lcu r:WIDTH=7 %i
+select -assert-none t:$gt t:$ge t:$lt t:$le
+
+
+design -reset
+read_verilog <<EOT
+module top(input [8:0] a, b, output gtu, gts, ltu, lts, geu, ges, leu, les);
+wire [13:0] c = {a[8:6], 3'b101, a[5:4], 2'b11, a[3:0]};
+wire [13:0] d = {b[8], 3'b101, b[7:4], 2'b01, b[3:0]};
+assign gtu = c > d;
+assign gts = $signed(c) > $signed(d);
+assign ltu = c < d;
+assign lts = $signed(c) < $signed(d);
+assign geu = c >= d;
+assign ges = $signed(c) >= $signed(d);
+assign leu = c <= d;
+assign les = $signed(c) <= $signed(d);
+endmodule
+EOT
+design -save gold
+
+equiv_opt -assert techmap -map +/cmp2lcu.v -D LUT_WIDTH=5
+design -load postopt
+select -assert-count 8 t:$lcu r:WIDTH=2 %i
+select -assert-none t:$gt t:$ge t:$lt t:$le
+
+design -load preopt
+equiv_opt -assert techmap -map +/cmp2lcu.v -D LUT_WIDTH=3
+design -load postopt
+select -assert-count 8 t:$lcu r:WIDTH=4 %i
+select -assert-none t:$gt t:$ge t:$lt t:$le