gowin: dsp: Add basic DSP block inferencing for various MULT cells

techlibs/gowin/Makefile.inc

@@ -11,3 +11,4 @@ $(eval $(call add_share_file,share/gowin,techlibs/gowin/brams_map.v))
 $(eval $(call add_share_file,share/gowin,techlibs/gowin/brams.txt))
 $(eval $(call add_share_file,share/gowin,techlibs/gowin/lutrams_map.v))
 $(eval $(call add_share_file,share/gowin,techlibs/gowin/lutrams.txt))
+$(eval $(call add_share_file,share/gowin,techlibs/gowin/dsp_map.v))

techlibs/gowin/dsp_map.v

@@ -0,0 +1,65 @@
+module \$__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;
+
+	wire [8:0] soa;
+	wire [8:0] sob;
+
+	MULT9X9 _TECHMAP_REPLACE_ (
+		.A(A),
+		.B(B),
+		.SIA(8'b0),
+		.SIB(8'b0),
+		.ASIGN(A_SIGNED),
+		.BSIGN(B_SIGNED),
+		.ASEL(1'b0),
+		.BSEL(1'b0),
+		.SOA(soa),
+		.SOB(sob),
+		.DOUT(Y)
+	);
+endmodule
+
+module \$__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;
+
+	wire [17:0] soa;
+	wire [17:0] sob;
+
+	MULT18X18 _TECHMAP_REPLACE_ (
+		.A(A),
+		.B(B),
+		.SIA(18'b0),
+		.SIB(18'b0),
+		.ASIGN(A_SIGNED),
+		.BSIGN(B_SIGNED),
+		.ASEL(1'b0),
+		.BSEL(1'b0),
+		.SOA(soa),
+		.SOB(sob),
+		.DOUT(Y)
+	);
+endmodule
+
+module \$__MUL36X36 (input [35:0] A, input [35:0] B, output [72:0] Y);
+	parameter A_WIDTH = 36;
+	parameter B_WIDTH = 36;
+	parameter Y_WIDTH = 72;
+	parameter A_SIGNED = 0;
+	parameter B_SIGNED = 0;
+
+	MULT36X36 _TECHMAP_REPLACE_ (
+		.A(A),
+		.B(B),
+		.ASIGN(A_SIGNED),
+		.BSIGN(B_SIGNED),
+		.DOUT(Y)
+	);
+endmodule

techlibs/gowin/synth_gowin.cc

@@ -91,13 +91,16 @@ struct SynthGowinPass : public ScriptPass
 		log("		  The following families are supported:\n");
 		log("        'gw1n', 'gw2a', 'gw5a'.\n");
 		log("\n");
+		log("    -nodsp\n");
+		log("        do not infer DSP multipliers.\n");
+		log("\n");
 		log("The following commands are executed by this synthesis command:\n");
 		help_script();
 		log("\n");
 	}
 
 	string top_opt, vout_file, json_file, family;
-	bool retime, nobram, nolutram, flatten, nodffe, nowidelut, abc9, noiopads, noalu, no_rw_check;
+	bool retime, nobram, nolutram, flatten, nodffe, nowidelut, abc9, noiopads, noalu, no_rw_check, nodsp;
 
 	void clear_flags() override
 	{
@@ -115,6 +118,7 @@ struct SynthGowinPass : public ScriptPass
 		noiopads = false;
 		noalu = false;
 		no_rw_check = false;
+		nodsp = false;
 	}
 
 	void execute(std::vector<std::string> args, RTLIL::Design *design) override
@@ -193,6 +197,10 @@ struct SynthGowinPass : public ScriptPass
 				no_rw_check = true;
 				continue;
 			}
+			if (args[argidx] == "-nodsp") {
+				nodsp = true;
+				continue;
+			}
 			break;
 		}
 		extra_args(args, argidx, design);
@@ -208,6 +216,24 @@ struct SynthGowinPass : public ScriptPass
 		log_pop();
 	}
 
+	// DSP mapping rules for mul2dsp
+	struct DSPRule {
+		int a_maxwidth;
+		int b_maxwidth;
+		int a_minwidth;
+		int b_minwidth;
+		std::string prim;
+	};
+
+	// gw1n and gw2a
+	const std::vector<DSPRule> dsp_rules = {
+		{36, 36, 18, 18, "$__MUL36X36"},
+		{18, 18, 10,  4, "$__MUL18X18"},
+		{18, 18,  4, 10, "$__MUL18X18"},
+		{ 9,  9,  4,  4, "$__MUL9X9"},
+	};
+	// TODO: gw5a (MULT12X12, MULT27x36)
+
 	void script() override
 	{
 		std::string no_rw_check_opt = "";
@@ -233,6 +259,23 @@ struct SynthGowinPass : public ScriptPass
 
 		if (check_label("coarse"))
 		{
+			if (help_mode)
+			{
+				run("techmap -map +/mul2dsp.v [...]", "(unless -nodsp)");
+				run("techmap -map +/gowin/dsp_map.v [...]", "(unless -nodsp)");
+			} else if (!nodsp) {
+				if (family == "gw1n" || family == "gw2a")
+				{
+					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 +/gowin/dsp_map.v");
+				}
+			}
+
 			run("synth -run coarse" + no_rw_check_opt);
 		}
 

tests/arch/gowin/mul_gw1n.ys

@@ -0,0 +1,47 @@
+read_verilog ../common/mul.v
+chparam -set X_WIDTH 8 -set Y_WIDTH 8 -set A_WIDTH 16
+hierarchy -top top
+proc
+# equivalence checking is somewhat slow (and missing simulation models)
+synth_gowin -family gw1n
+
+cd top # Constrain all select calls below inside the top module
+select -assert-count 1 t:MULT9X9
+# XXX: Whats's `top/x_IBUF_I_O_MULT9X9_A_A_GND_G` ??
+# select -assert-none t:IBUF t:OBUF t:MULT9X9 %% t:* %D
+
+design -reset
+read_verilog ../common/mul.v
+chparam -set X_WIDTH 16 -set Y_WIDTH 16 -set A_WIDTH 32
+hierarchy -top top
+proc
+synth_gowin -family gw1n
+cd top # Constrain all select calls below inside the top module
+select -assert-count 1 t:MULT18X18
+# XXX: top/x_IBUF_I_O_MULT18X18_A_A_GND_G ???
+# select -assert-none t:IBUF t:OBUF t:MULT18X18 %% t:* %D
+
+
+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_gowin
+cd top # Constrain all select calls below inside the top module
+select -assert-count 1 t:MULT36X36
+# XXX: top/x_IBUF_I_O_MULT36X36_A_A_GND_G
+# select -assert-none t:IBUF t:OBUF t:MULT36X36 %% t:* %D
+
+# TODO: We end up with two 18x18 multipliers
+# 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_gowin
+# cd top # Constrain all select calls below inside the top module
+# select -assert-count 2 t:MULT18X18
+# select -assert-none t:IBUF t:OBUF t:MULT36X36 %% t:* %D

tests/arch/gowin/mul_gw2a.ys

@@ -0,0 +1,47 @@
+read_verilog ../common/mul.v
+chparam -set X_WIDTH 8 -set Y_WIDTH 8 -set A_WIDTH 16
+hierarchy -top top
+proc
+# equivalence checking is somewhat slow (and missing simulation models)
+synth_gowin -family gw2a
+
+cd top # Constrain all select calls below inside the top module
+select -assert-count 1 t:MULT9X9
+# XXX: Whats's `top/x_IBUF_I_O_MULT9X9_A_A_GND_G` ??
+# select -assert-none t:IBUF t:OBUF t:MULT9X9 %% t:* %D
+
+design -reset
+read_verilog ../common/mul.v
+chparam -set X_WIDTH 16 -set Y_WIDTH 16 -set A_WIDTH 32
+hierarchy -top top
+proc
+synth_gowin -family gw2a
+cd top # Constrain all select calls below inside the top module
+select -assert-count 1 t:MULT18X18
+# XXX: top/x_IBUF_I_O_MULT18X18_A_A_GND_G ???
+# select -assert-none t:IBUF t:OBUF t:MULT18X18 %% t:* %D
+
+
+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_gowin -family gw2a
+cd top # Constrain all select calls below inside the top module
+select -assert-count 1 t:MULT36X36
+# XXX: top/x_IBUF_I_O_MULT36X36_A_A_GND_G
+# select -assert-none t:IBUF t:OBUF t:MULT36X36 %% t:* %D
+
+# TODO: We end up with two 18x18 multipliers
+# 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_gowin -family gw2a
+# cd top # Constrain all select calls below inside the top module
+# select -assert-count 2 t:MULT18X18
+# select -assert-none t:IBUF t:OBUF t:MULT36X36 %% t:* %D