Merge pull request #771 from whitequark/techmap_cmp2lut

cmp2lut: new techmap pass

5 files changed, 139 insertions, 3 deletions

diff --git a/techlibs/common/Makefile.inc b/techlibs/common/Makefile.inc
index 70074f653..0e05620bc 100644
--- a/techlibs/common/Makefile.inc
+++ b/techlibs/common/Makefile.inc
@@ -26,5 +26,5 @@ $(eval $(call add_share_file,share,techlibs/common/pmux2mux.v))
 $(eval $(call add_share_file,share,techlibs/common/adff2dff.v))
 $(eval $(call add_share_file,share,techlibs/common/dff2ff.v))
 $(eval $(call add_share_file,share,techlibs/common/gate2lut.v))
+$(eval $(call add_share_file,share,techlibs/common/cmp2lut.v))
 $(eval $(call add_share_file,share,techlibs/common/cells.lib))
-
diff --git a/techlibs/common/cmp2lut.v b/techlibs/common/cmp2lut.v
new file mode 100644
index 000000000..8aa1eb957
--- /dev/null
+++ b/techlibs/common/cmp2lut.v
@@ -0,0 +1,105 @@
+// Certain arithmetic operations between a signal of width n and a constant can be directly mapped
+// to a single k-LUT (where n <= k). This is preferable to normal alumacc techmapping process
+// because for many targets, arithmetic techmapping creates hard logic (such as carry cells) which often
+// cannot be optimized further.
+//
+// TODO: Currently, only comparisons with 1-bit output are mapped. Potentially, all arithmetic cells
+// with n <= k inputs should be techmapped in this way, because this shortens the critical path
+// from n to 1 by avoiding carry chains.
+
+(* techmap_celltype = "$eq $ne $lt $le $gt $ge" *)
+module _90_lut_cmp_ (A, B, Y);
+
+parameter A_SIGNED = 0;
+parameter B_SIGNED = 0;
+parameter A_WIDTH = 0;
+parameter B_WIDTH = 0;
+parameter Y_WIDTH = 0;
+
+input [A_WIDTH-1:0] A;
+input [B_WIDTH-1:0] B;
+output [Y_WIDTH-1:0] Y;
+
+parameter _TECHMAP_CELLTYPE_ = "";
+
+parameter _TECHMAP_CONSTMSK_A_ = 0;
+parameter _TECHMAP_CONSTVAL_A_ = 0;
+parameter _TECHMAP_CONSTMSK_B_ = 0;
+parameter _TECHMAP_CONSTVAL_B_ = 0;
+
+function automatic integer gen_lut;
+	input integer width;
+	input integer operation;
+	input integer swap;
+	input integer sign;
+	input integer operand;
+	integer n, i_var, i_cst, lhs, rhs, o_bit;
+	begin
+		gen_lut = width'b0;
+		for (n = 0; n < (1 << width); n++) begin
+			if (sign)
+				i_var = n[width-1:0];
+			else
+				i_var = n;
+			i_cst = operand;
+			if (swap) begin
+				lhs = i_cst;
+				rhs = i_var;
+			end else begin
+				lhs = i_var;
+				rhs = i_cst;
+			end
+			if (operation == 0)
+				o_bit = (lhs <  rhs);
+			if (operation == 1)
+				o_bit = (lhs <= rhs);
+			if (operation == 2)
+				o_bit = (lhs >  rhs);
+			if (operation == 3)
+				o_bit = (lhs >= rhs);
+			if (operation == 4)
+				o_bit = (lhs == rhs);
+			if (operation == 5)
+				o_bit = (lhs != rhs);
+			gen_lut = gen_lut | (o_bit << n);
+		end
+	end
+endfunction
+
+generate
+	if (_TECHMAP_CELLTYPE_ == "$lt")
+		localparam operation = 0;
+	if (_TECHMAP_CELLTYPE_ == "$le")
+		localparam operation = 1;
+	if (_TECHMAP_CELLTYPE_ == "$gt")
+		localparam operation = 2;
+	if (_TECHMAP_CELLTYPE_ == "$ge")
+		localparam operation = 3;
+	if (_TECHMAP_CELLTYPE_ == "$eq")
+		localparam operation = 4;
+	if (_TECHMAP_CELLTYPE_ == "$ne")
+		localparam operation = 5;
+
+	if (A_WIDTH > `LUT_WIDTH || B_WIDTH > `LUT_WIDTH || Y_WIDTH != 1)
+		wire _TECHMAP_FAIL_ = 1;
+	else if (&_TECHMAP_CONSTMSK_B_)
+		\$lut #(
+			.WIDTH(A_WIDTH),
+			.LUT({ gen_lut(A_WIDTH, operation, 0, A_SIGNED && B_SIGNED, _TECHMAP_CONSTVAL_B_) })
+		) _TECHMAP_REPLACE_ (
+			.A(A),
+			.Y(Y)
+		);
+	else if (&_TECHMAP_CONSTMSK_A_)
+		\$lut #(
+			.WIDTH(B_WIDTH),
+			.LUT({ gen_lut(B_WIDTH, operation, 1, A_SIGNED && B_SIGNED, _TECHMAP_CONSTVAL_A_) })
+		) _TECHMAP_REPLACE_ (
+			.A(B),
+			.Y(Y)
+		);
+	else
+		wire _TECHMAP_FAIL_ = 1;
+endgenerate
+
+endmodule
diff --git a/tests/lut/check_map.ys b/tests/lut/check_map.ys
index dc0aaffc2..46854e82e 100644
--- a/tests/lut/check_map.ys
+++ b/tests/lut/check_map.ys
@@ -1,4 +1,6 @@
 simplemap
-equiv_opt -assert techmap -map +/gate2lut.v -D LUT_WIDTH=4
+equiv_opt -assert techmap -D LUT_WIDTH=4 -map +/cmp2lut.v
 design -load postopt
-select -assert-count 1 t:$lut
+equiv_opt -assert techmap -D LUT_WIDTH=4 -map +/gate2lut.v
+design -load postopt
+select -assert-count 0 t:* t:$lut %d
diff --git a/tests/lut/map_cmp.v b/tests/lut/map_cmp.v
new file mode 100644
index 000000000..5e413f894
--- /dev/null
+++ b/tests/lut/map_cmp.v
@@ -0,0 +1,29 @@
+module top(...);
+	input [3:0] a;
+
+	output o1_1 = 4'b1010 <= a;
+	output o1_2 = 4'b1010 <  a;
+	output o1_3 = 4'b1010 >= a;
+	output o1_4 = 4'b1010 >  a;
+	output o1_5 = 4'b1010 == a;
+	output o1_6 = 4'b1010 != a;
+
+	output o2_1 = a <= 4'b1010;
+	output o2_2 = a <  4'b1010;
+	output o2_3 = a >= 4'b1010;
+	output o2_4 = a >  4'b1010;
+	output o2_5 = a == 4'b1010;
+	output o2_6 = a != 4'b1010;
+
+	output o3_1 = 4'sb0101 <= $signed(a);
+	output o3_2 = 4'sb0101 <  $signed(a);
+	output o3_3 = 4'sb0101 >= $signed(a);
+	output o3_4 = 4'sb0101 >  $signed(a);
+	output o3_5 = 4'sb0101 == $signed(a);
+	output o3_6 = 4'sb0101 != $signed(a);
+
+	output o4_1 = $signed(a) <= 4'sb0000;
+	output o4_2 = $signed(a) <  4'sb0000;
+	output o4_3 = $signed(a) >= 4'sb0000;
+	output o4_4 = $signed(a) >  4'sb0000;
+endmodule
diff --git a/tests/lut/run-test.sh b/tests/lut/run-test.sh
index 207417fa6..207417fa6 100644..100755
--- a/tests/lut/run-test.sh
+++ b/tests/lut/run-test.sh