nexus: DSP inference support

Signed-off-by: David Shah <dave@ds0.me>

4 files changed, 151 insertions, 13 deletions

diff --git a/techlibs/nexus/Makefile.inc b/techlibs/nexus/Makefile.inc
index e0ff40e15..c9a9ad4ff 100644
--- a/techlibs/nexus/Makefile.inc
+++ b/techlibs/nexus/Makefile.inc
@@ -11,4 +11,5 @@ $(eval $(call add_share_file,share/nexus,techlibs/nexus/brams_map.v))
 $(eval $(call add_share_file,share/nexus,techlibs/nexus/brams.txt))
 $(eval $(call add_share_file,share/nexus,techlibs/nexus/arith_map.v))
 $(eval $(call add_share_file,share/nexus,techlibs/nexus/latches_map.v))
+$(eval $(call add_share_file,share/nexus,techlibs/nexus/dsp_map.v))
 
diff --git a/techlibs/nexus/dsp_map.v b/techlibs/nexus/dsp_map.v
new file mode 100644
index 000000000..b12528309
--- /dev/null
+++ b/techlibs/nexus/dsp_map.v
@@ -0,0 +1,79 @@
+module \$__NX_MUL36X36 (input [35:0] A, input [35:0] B, output [71:0] Y);
+
+	parameter A_WIDTH = 36;
+	parameter B_WIDTH = 36;
+	parameter Y_WIDTH = 72;
+	parameter A_SIGNED = 0;
+	parameter B_SIGNED = 0;
+
+	MULT36X36 #(
+		.REGINPUTA("BYPASS"),
+		.REGINPUTB("BYPASS"),
+		.REGOUTPUT("BYPASS")
+	) _TECHMAP_REPLACE_ (
+		.A(A), .B(B),
+		.SIGNEDA(A_SIGNED ? 1'b1 : 1'b0),
+		.SIGNEDB(B_SIGNED ? 1'b1 : 1'b0),
+		.Z(Y)
+	);
+endmodule
+
+module \$__NX_MUL36X18 (input [35:0] A, input [17:0] B, output [53:0] Y);
+
+	parameter A_WIDTH = 36;
+	parameter B_WIDTH = 18;
+	parameter Y_WIDTH = 54;
+	parameter A_SIGNED = 0;
+	parameter B_SIGNED = 0;
+
+	MULT18X36 #(
+		.REGINPUTA("BYPASS"),
+		.REGINPUTB("BYPASS"),
+		.REGOUTPUT("BYPASS")
+	) _TECHMAP_REPLACE_ (
+		.A(B), .B(A),
+		.SIGNEDA(B_SIGNED ? 1'b1 : 1'b0),
+		.SIGNEDB(A_SIGNED ? 1'b1 : 1'b0),
+		.Z(Y)
+	);
+endmodule
+
+module \$__NX_MUL18X18 (input [17:0] A, input [17:0] B, output [35:0] Y);
+
+	parameter A_WIDTH = 18;
+	parameter B_WIDTH = 18;
+	parameter Y_WIDTH = 36;
+	parameter A_SIGNED = 0;
+	parameter B_SIGNED = 0;
+
+	MULT18X18 #(
+		.REGINPUTA("BYPASS"),
+		.REGINPUTB("BYPASS"),
+		.REGOUTPUT("BYPASS")
+	) _TECHMAP_REPLACE_ (
+		.A(A), .B(B),
+		.SIGNEDA(A_SIGNED ? 1'b1 : 1'b0),
+		.SIGNEDB(B_SIGNED ? 1'b1 : 1'b0),
+		.Z(Y)
+	);
+endmodule
+
+module \$__NX_MUL9X9 (input [8:0] A, input [8:0] B, output [17:0] Y);
+
+	parameter A_WIDTH = 9;
+	parameter B_WIDTH = 9;
+	parameter Y_WIDTH = 18;
+	parameter A_SIGNED = 0;
+	parameter B_SIGNED = 0;
+
+	MULT9X9 #(
+		.REGINPUTA("BYPASS"),
+		.REGINPUTB("BYPASS"),
+		.REGOUTPUT("BYPASS")
+	) _TECHMAP_REPLACE_ (
+		.A(A), .B(B),
+		.SIGNEDA(A_SIGNED ? 1'b1 : 1'b0),
+		.SIGNEDB(B_SIGNED ? 1'b1 : 1'b0),
+		.Z(Y)
+	);
+endmodule
diff --git a/techlibs/nexus/synth_nexus.cc b/techlibs/nexus/synth_nexus.cc
index 7e2185ab6..9eabbace7 100644
--- a/techlibs/nexus/synth_nexus.cc
+++ b/techlibs/nexus/synth_nexus.cc
@@ -89,6 +89,9 @@ struct SynthNexusPass : public ScriptPass
 		log("    -noiopad\n");
 		log("        do not insert IO buffers\n");
 		log("\n");
+		log("    -nodsp\n");
+		log("        do not infer DSP multipliers\n");
+		log("\n");
 		log("    -abc9\n");
 		log("        use new ABC9 flow (EXPERIMENTAL)\n");
 		log("\n");
@@ -98,7 +101,7 @@ struct SynthNexusPass : public ScriptPass
 	}
 
 	string top_opt, json_file, vm_file, family;
-	bool noccu2, nodffe, nobram, nolutram, nowidelut, noiopad, flatten, dff, retime, abc9;
+	bool noccu2, nodffe, nobram, nolutram, nowidelut, noiopad, nodsp, flatten, dff, retime, abc9;
 
 	void clear_flags() override
 	{
@@ -112,6 +115,7 @@ struct SynthNexusPass : public ScriptPass
 		nolutram = false;
 		nowidelut = false;
 		noiopad = false;
+		nodsp = false;
 		flatten = true;
 		dff = false;
 		retime = false;
@@ -161,6 +165,10 @@ struct SynthNexusPass : public ScriptPass
 				dff = true;
 				continue;
 			}
+			if (args[argidx] == "-nodsp") {
+				nodsp = true;
+				continue;
+			}
 			if (args[argidx] == "-retime") {
 				retime = true;
 				continue;
@@ -211,6 +219,22 @@ struct SynthNexusPass : public ScriptPass
 		log_pop();
 	}
 
+	struct DSPRule {
+		int a_maxwidth;
+		int b_maxwidth;
+		int a_minwidth;
+		int b_minwidth;
+		std::string prim;
+	};
+
+	const std::vector<DSPRule> dsp_rules = {
+		{36, 36, 22, 22, "$__NX_MUL36X36"},
+		{36, 18, 22, 10, "$__NX_MUL36X18"},
+		{18, 18, 10,  4, "$__NX_MUL18X18"},
+		{18, 18,  4, 10, "$__NX_MUL18X18"},
+		{ 9,  9,  4,  4, "$__NX_MUL9X9"},
+	};
+
 	void script() override
 	{
 
@@ -244,6 +268,18 @@ struct SynthNexusPass : public ScriptPass
 			run("opt_expr");
 			run("opt_clean");
 
+			if (help_mode) {
+				run("techmap -map +/mul2dsp.v [...]", "(unless -nodsp)");
+				run("techmap -map +/nexus/dsp_map.v", "(unless -nodsp)");
+			} else if (!nodsp) {
+				for (const auto &rule : dsp_rules) {
+					run(stringf("techmap -map +/mul2dsp.v -D DSP_A_MAXWIDTH=%d -D DSP_B_MAXWIDTH=%d -D DSP_A_MINWIDTH=%d -D DSP_B_MINWIDTH=%d -D DSP_NAME=%s",
+						rule.a_maxwidth, rule.b_maxwidth, rule.a_minwidth, rule.b_minwidth, rule.prim.c_str()));
+					run("chtype -set $mul t:$__soft_mul");
+				}
+				run("techmap -map +/nexus/dsp_map.v");
+			}
+
 			run("alumacc");
 			run("opt");
 			run("memory -nomap");
diff --git a/tests/arch/nexus/mul.ys b/tests/arch/nexus/mul.ys
index 27ea3e04e..65a2fd8c3 100644
--- a/tests/arch/nexus/mul.ys
+++ b/tests/arch/nexus/mul.ys
@@ -1,4 +1,5 @@
 read_verilog ../common/mul.v
+chparam -set X_WIDTH 8 -set Y_WIDTH 8 -set A_WIDTH 16
 hierarchy -top top
 proc
 
@@ -7,22 +8,43 @@ design -save read
 equiv_opt -assert -map +/nexus/cells_sim.v synth_nexus
 design -load postopt # load the post-opt design (otherwise equiv_opt loads the pre-opt design)
 cd top # Constrain all select calls below inside the top module
-select -assert-count 7 t:CCU2
-select -assert-max 5 t:WIDEFN9
-select -assert-max 62 t:LUT4
+select -assert-count 1 t:MULT9X9
 
-select -assert-none t:IB t:OB t:VLO t:VHI t:LUT4 t:CCU2 t:WIDEFN9 %% t:* %D
+select -assert-none t:IB t:OB t:VLO t:VHI t:MULT9X9 %% t:* %D
 
-design -load read
 
-equiv_opt -assert -map +/nexus/cells_sim.v synth_nexus -abc9
-design -load postopt # load the post-opt design (otherwise equiv_opt loads the pre-opt design)
+design -reset
+read_verilog ../common/mul.v
+chparam -set X_WIDTH 16 -set Y_WIDTH 16 -set A_WIDTH 32
+hierarchy -top top
+proc
+# equivalence checking is too slow here
+synth_nexus
+cd top # Constrain all select calls below inside the top module
+select -assert-count 1 t:MULT18X18
+select -assert-none t:IB t:OB t:VLO t:VHI t:MULT18X18 %% t:* %D
+
+
+design -reset
+read_verilog ../common/mul.v
+chparam -set X_WIDTH 32 -set Y_WIDTH 16 -set A_WIDTH 48
+hierarchy -top top
+proc
+# equivalence checking is too slow here
+synth_nexus
 cd top # Constrain all select calls below inside the top module
+select -assert-count 1 t:MULT18X36
+select -assert-none t:IB t:OB t:VLO t:VHI t:MULT18X36 %% t:* %D
 
-stat
 
-select -assert-count 7 t:CCU2
-select -assert-max 12 t:WIDEFN9
-select -assert-max 58 t:LUT4
+design -reset
+read_verilog ../common/mul.v
+chparam -set X_WIDTH 32 -set Y_WIDTH 32 -set A_WIDTH 64
+hierarchy -top top
+proc
+# equivalence checking is too slow here
+synth_nexus
+cd top # Constrain all select calls below inside the top module
+select -assert-count 1 t:MULT36X36
 
-select -assert-none t:IB t:OB t:VLO t:VHI t:LUT4 t:CCU2 t:WIDEFN9 %% t:* %D
+select -assert-none t:IB t:OB t:VLO t:VHI t:MULT36X36 %% t:* %D