Improve iCE40 SB_MAC16 model

Signed-off-by: Clifford Wolf <clifford@clifford.at>

5 files changed, 179 insertions, 121 deletions

diff --git a/techlibs/ice40/cells_sim.v b/techlibs/ice40/cells_sim.v
index 4a92fccdf..2041693cc 100644
--- a/techlibs/ice40/cells_sim.v
+++ b/techlibs/ice40/cells_sim.v
@@ -1283,7 +1283,7 @@ module SB_MAC16 (
 
 	// Regs B and D
 	reg [15:0] rB, rD;
-	always @(posedge clock, posedge IRSTTOP) begin
+	always @(posedge clock, posedge IRSTBOT) begin
 		if (IRSTBOT) begin
 			rB <= 0;
 			rD <= 0;
@@ -1298,10 +1298,10 @@ module SB_MAC16 (
 	// Multiplier Stage
 	wire [15:0] p_Ah_Bh, p_Al_Bh, p_Ah_Bl, p_Al_Bl;
 	wire [15:0] Ah, Al, Bh, Bl;
-	assign Ah = A_SIGNED ? {{8{iA[15]}}, iA[15: 8]} : iA[15: 8];
-	assign Al = A_SIGNED ? {{8{iA[ 7]}}, iA[ 7: 0]} : iA[15: 8];
-	assign Bh = B_SIGNED ? {{8{iB[15]}}, iB[15: 8]} : iB[15: 8];
-	assign Bl = B_SIGNED ? {{8{iB[ 7]}}, iB[ 7: 0]} : iB[15: 8];
+	assign Ah = {A_SIGNED ? {8{iA[15]}} : 8'b0, iA[15: 8]};
+	assign Al = {A_SIGNED ? {8{iA[ 7]}} : 8'b0, iA[ 7: 0]};
+	assign Bh = {B_SIGNED ? {8{iB[15]}} : 8'b0, iB[15: 8]};
+	assign Bl = {B_SIGNED ? {8{iB[ 7]}} : 8'b0, iB[ 7: 0]};
 	assign p_Ah_Bh = Ah * Bh;
 	assign p_Al_Bh = Al * Bh;
 	assign p_Ah_Bl = Ah * Bl;
@@ -1336,17 +1336,10 @@ module SB_MAC16 (
 	assign iG = BOT_8x8_MULT_REG ? rG : p_Al_Bl;
 
 	// Adder Stage
-	reg [31:0] P;
-	always @* begin
-		P = iG[7:0];
-		P = P + (iG[15:8] + iK[7:0]) << 8;
-		P = P + (iK[15:8] + iJ[7:0]) << 16;
-		P = P + (iJ[15:8] + iF[7:0]) << 24;
-	end
-	assign iL = P;
+	assign iL = iG + (iK << 8) + (iJ << 8) + (iF << 16);
 
 	// Reg H
-	reg [15:0] rH;
+	reg [31:0] rH;
 	always @(posedge clock, posedge IRSTBOT) begin
 		if (IRSTBOT) begin
 			rH <= 0;
@@ -1359,7 +1352,7 @@ module SB_MAC16 (
 	// Hi Output Stage
 	wire [15:0] XW, Oh;
 	reg [15:0] rQ;
-	assign iW = TOPADDSUB_UPPERINPUT ? iC : iQ[31:16];
+	assign iW = TOPADDSUB_UPPERINPUT ? iC : iQ;
 	assign iX = (TOPADDSUB_LOWERINPUT == 0) ? iA : (TOPADDSUB_LOWERINPUT == 1) ? iF : (TOPADDSUB_LOWERINPUT == 2) ? iH[31:16] : {16{iZ[15]}};
 	assign {ACCUMCO, XW} = iX + (iW ^ {16{ADDSUBTOP}}) + HCI;
 	assign CO = ACCUMCO ^ ADDSUBTOP;
@@ -1379,7 +1372,7 @@ module SB_MAC16 (
 	// Lo Output Stage
 	wire [15:0] YZ, Ol;
 	reg [15:0] rS;
-	assign iY = BOTADDSUB_UPPERINPUT ? iD : iQ[15:0];
+	assign iY = BOTADDSUB_UPPERINPUT ? iD : iS;
 	assign iZ = (BOTADDSUB_LOWERINPUT == 0) ? iB : (BOTADDSUB_LOWERINPUT == 1) ? iG : (BOTADDSUB_LOWERINPUT == 2) ? iH[15:0] : {16{SIGNEXTIN}};
 	assign {LCO, YZ} = iZ + (iY ^ {16{ADDSUBBOT}}) + LCI;
 	assign iR = OLOADBOT ? iD : YZ ^ {16{ADDSUBBOT}};
@@ -1387,7 +1380,7 @@ module SB_MAC16 (
 		if (ORSTBOT) begin
 			rS <= 0;
 		end else if (CE) begin
-			if (!OHOLDTOP) rS <= iR;
+			if (!OHOLDBOT) rS <= iR;
 		end
 	end
 	assign iS = rS;
diff --git a/techlibs/ice40/tests/.gitignore b/techlibs/ice40/tests/.gitignore
index b58f9ad4a..9795152c2 100644
--- a/techlibs/ice40/tests/.gitignore
+++ b/techlibs/ice40/tests/.gitignore
@@ -1,2 +1,6 @@
-test_ffs_[01][01][01][01][01]_*
-test_bram_[0-9]*
+/test_ffs_[01][01][01][01][01]_*
+/test_bram_[0-9]*
+/test_dsp_model
+/test_dsp_model.vcd
+/test_dsp_model_ref.v
+/test_dsp_model_uut.v
diff --git a/techlibs/ice40/tests/test_dsp_model.gtkw b/techlibs/ice40/tests/test_dsp_model.gtkw
deleted file mode 100644
index dafe8916f..000000000
--- a/techlibs/ice40/tests/test_dsp_model.gtkw
+++ /dev/null
@@ -1,87 +0,0 @@
-[*]
-[*] GTKWave Analyzer v3.3.86 (w)1999-2017 BSI
-[*] Tue Feb 19 13:33:31 2019
-[*]
-[dumpfile] "/home/clifford/Work/yosys/techlibs/ice40/tests/test_dsp_model.vcd"
-[dumpfile_mtime] "Tue Feb 19 13:29:34 2019"
-[dumpfile_size] 119605
-[savefile] "/home/clifford/Work/yosys/techlibs/ice40/tests/test_dsp_model.gtkw"
-[timestart] 0
-[size] 1850 1362
-[pos] 1816 32
-*-16.399944 42300 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1
-[sst_width] 223
-[signals_width] 142
-[sst_expanded] 1
-[sst_vpaned_height] 420
-@28
-testbench.CLK
-testbench.CE
-@200
--
-@28
-testbench.REF_ACCUMCO
-testbench.UUT_ACCUMCO
-@200
--
-@28
-testbench.REF_CO
-testbench.UUT_CO
-@200
--
-@22
-testbench.REF_O[31:0]
-testbench.UUT_O[31:0]
-@200
--
-@28
-testbench.REF_SIGNEXTOUT
-testbench.UUT_SIGNEXTOUT
-@200
--
-@22
-testbench.A[15:0]
-testbench.B[15:0]
-testbench.C[15:0]
-testbench.D[15:0]
-@200
--
-@28
-testbench.AHOLD
-testbench.BHOLD
-testbench.CHOLD
-testbench.DHOLD
-@200
--
-@28
-testbench.SIGNEXTIN
-testbench.ACCUMCI
-testbench.CI
-@200
--
-@28
-testbench.ADDSUBTOP
-testbench.ADDSUBBOT
-@200
--
-@28
-testbench.IRSTTOP
-testbench.IRSTBOT
-@200
--
-@29
-testbench.OHOLDTOP
-@28
-testbench.OHOLDBOT
-@200
--
-@28
-testbench.OLOADTOP
-testbench.OLOADBOT
-@200
--
-@28
-testbench.ORSTTOP
-testbench.ORSTBOT
-[pattern_trace] 1
-[pattern_trace] 0
diff --git a/techlibs/ice40/tests/test_dsp_model.sh b/techlibs/ice40/tests/test_dsp_model.sh
index ad079b2b6..1bc0cc688 100644
--- a/techlibs/ice40/tests/test_dsp_model.sh
+++ b/techlibs/ice40/tests/test_dsp_model.sh
@@ -2,5 +2,10 @@
 set -ex
 sed 's/SB_MAC16/SB_MAC16_UUT/; /SB_MAC16_UUT/,/endmodule/ p; d;' < ../cells_sim.v > test_dsp_model_uut.v
 cat /opt/lscc/iCEcube2.2017.01/verilog/sb_ice_syn.v > test_dsp_model_ref.v
-iverilog -s testbench -o test_dsp_model test_dsp_model.v test_dsp_model_uut.v test_dsp_model_ref.v
-./test_dsp_model
+for tb in testbench \
+		testbench_comb_8x8_A testbench_comb_8x8_B testbench_comb_16x16 \
+		testbench_seq_16x16_A testbench_seq_16x16_B
+do
+	iverilog -s $tb -o test_dsp_model test_dsp_model.v test_dsp_model_uut.v test_dsp_model_ref.v
+	vvp -N ./test_dsp_model
+done
diff --git a/techlibs/ice40/tests/test_dsp_model.v b/techlibs/ice40/tests/test_dsp_model.v
index 919bb7917..594bd4ad3 100644
--- a/techlibs/ice40/tests/test_dsp_model.v
+++ b/techlibs/ice40/tests/test_dsp_model.v
@@ -11,13 +11,13 @@ module testbench;
 	parameter [0:0] PIPELINE_16x16_MULT_REG1 = 0;
 	parameter [0:0] PIPELINE_16x16_MULT_REG2 = 0;
 	parameter [1:0] TOPOUTPUT_SELECT = 0;
-	parameter [1:0] TOPADDSUB_LOWERINPUT = 2;
+	parameter [1:0] TOPADDSUB_LOWERINPUT = 0;
 	parameter [0:0] TOPADDSUB_UPPERINPUT = 1;
-	parameter [1:0] TOPADDSUB_CARRYSELECT = 2;
+	parameter [1:0] TOPADDSUB_CARRYSELECT = 0;
 	parameter [1:0] BOTOUTPUT_SELECT = 0;
-	parameter [1:0] BOTADDSUB_LOWERINPUT = 2;
+	parameter [1:0] BOTADDSUB_LOWERINPUT = 0;
 	parameter [0:0] BOTADDSUB_UPPERINPUT = 1;
-	parameter [1:0] BOTADDSUB_CARRYSELECT = 2;
+	parameter [1:0] BOTADDSUB_CARRYSELECT = 0;
 	parameter [0:0] MODE_8x8 = 0;
 	parameter [0:0] A_SIGNED = 0;
 	parameter [0:0] B_SIGNED = 0;
@@ -46,7 +46,7 @@ module testbench;
 			CLK = ~CLK;
 			#2;
 			if (REF_O !== UUT_O) begin
-				$display("ERROR at %1t: REF_O=%b UUT_O=%b", $time, REF_O, UUT_O);
+				$display("ERROR at %1t: REF_O=%b UUT_O=%b DIFF=%b", $time, REF_O, UUT_O, REF_O ^ UUT_O);
 				errcount = errcount + 1;
 			end
 			if (REF_CO !== UUT_CO) begin
@@ -69,29 +69,47 @@ module testbench;
 		$dumpfile("test_dsp_model.vcd");
 		$dumpvars(0, testbench);
 
-		#5;
+		#2;
 		CLK = NEG_TRIGGER;
 		CE = 1;
 		{C, A, B, D} = 0;
 		{AHOLD, BHOLD, CHOLD, DHOLD} = 0;
-		{IRSTTOP, IRSTBOT} = 0;
-		{ORSTTOP, ORSTBOT} = 0;
 		{OLOADTOP, OLOADBOT} = 0;
 		{ADDSUBTOP, ADDSUBBOT} = 0;
 		{OHOLDTOP, OHOLDBOT} = 0;
 		{CI, ACCUMCI, SIGNEXTIN} = 0;
 
-		// C = 10;
-		// A = 15;
-		// B = 22;
-		// D = 27;
+		{IRSTTOP, IRSTBOT} = ~0;
+		{ORSTTOP, ORSTBOT} = ~0;
+
+		#3;
+		{IRSTTOP, IRSTBOT} = 0;
+		{ORSTTOP, ORSTBOT} = 0;
+
+		repeat (300) begin
+			clkcycle;
+
+			A = $urandom;
+			B = $urandom;
+			C = $urandom;
+			D = $urandom;
 
-		repeat (10) clkcycle;
+			{AHOLD, BHOLD, CHOLD, DHOLD} = $urandom & $urandom & $urandom;
+			{OLOADTOP, OLOADBOT} = $urandom & $urandom & $urandom;
+			{ADDSUBTOP, ADDSUBBOT} = $urandom & $urandom & $urandom;
+			{OHOLDTOP, OHOLDBOT} = $urandom & $urandom & $urandom;
+			{CI, ACCUMCI, SIGNEXTIN} = $urandom & $urandom & $urandom;
+
+			{IRSTTOP, IRSTBOT} = $urandom & $urandom & $urandom;
+			{ORSTTOP, ORSTBOT} = $urandom & $urandom & $urandom;
+		end
 
 		if (errcount == 0) begin
 			$display("All tests passed.");
+			$finish;
 		end else begin
 			$display("Caught %1d errors.", errcount);
+			$stop;
 		end
 	end
 
@@ -197,3 +215,128 @@ module testbench;
 		.SIGNEXTOUT (UUT_SIGNEXTOUT)
 	);
 endmodule
+
+module testbench_comb_8x8_A;
+	testbench #(
+		.NEG_TRIGGER               (0),
+		.C_REG                     (0),
+		.A_REG                     (0),
+		.B_REG                     (0),
+		.D_REG                     (0),
+		.TOP_8x8_MULT_REG          (0),
+		.BOT_8x8_MULT_REG          (0),
+		.PIPELINE_16x16_MULT_REG1  (0),
+		.PIPELINE_16x16_MULT_REG2  (0),
+		.TOPOUTPUT_SELECT          (2),   // 0=P, 1=Q, 2=8x8, 3=16x16
+		.TOPADDSUB_LOWERINPUT      (0),   // 0=A, 1=8x8, 2=16x16, 3=S-EXT
+		.TOPADDSUB_UPPERINPUT      (0),   // 0=Q, 1=C
+		.TOPADDSUB_CARRYSELECT     (0),   // 0=0, 1=1, 2=ACI, 3=CI
+		.BOTOUTPUT_SELECT          (2),   // 0=R, 1=S, 2=8x8, 3=16x16
+		.BOTADDSUB_LOWERINPUT      (0),   // 0=B, 1=8x8, 2=16x16, 3=S-EXT
+		.BOTADDSUB_UPPERINPUT      (0),   // 0=S, 1=D
+		.BOTADDSUB_CARRYSELECT     (0),   // 0=0, 1=1, 2=ACI, 3=CI
+		.MODE_8x8                  (0),
+		.A_SIGNED                  (0),
+		.B_SIGNED                  (0)
+	) testbench ();
+endmodule
+
+module testbench_comb_8x8_B;
+	testbench #(
+		.NEG_TRIGGER               (0),
+		.C_REG                     (0),
+		.A_REG                     (0),
+		.B_REG                     (0),
+		.D_REG                     (0),
+		.TOP_8x8_MULT_REG          (0),
+		.BOT_8x8_MULT_REG          (0),
+		.PIPELINE_16x16_MULT_REG1  (0),
+		.PIPELINE_16x16_MULT_REG2  (0),
+		.TOPOUTPUT_SELECT          (0),   // 0=P, 1=Q, 2=8x8, 3=16x16
+		.TOPADDSUB_LOWERINPUT      (1),   // 0=A, 1=8x8, 2=16x16, 3=S-EXT
+		.TOPADDSUB_UPPERINPUT      (1),   // 0=Q, 1=C
+		.TOPADDSUB_CARRYSELECT     (0),   // 0=0, 1=1, 2=ACI, 3=CI
+		.BOTOUTPUT_SELECT          (0),   // 0=R, 1=S, 2=8x8, 3=16x16
+		.BOTADDSUB_LOWERINPUT      (1),   // 0=B, 1=8x8, 2=16x16, 3=S-EXT
+		.BOTADDSUB_UPPERINPUT      (1),   // 0=S, 1=D
+		.BOTADDSUB_CARRYSELECT     (0),   // 0=0, 1=1, 2=ACI, 3=CI
+		.MODE_8x8                  (0),
+		.A_SIGNED                  (0),
+		.B_SIGNED                  (0)
+	) testbench ();
+endmodule
+
+module testbench_comb_16x16;
+	testbench #(
+		.NEG_TRIGGER               (0),
+		.C_REG                     (0),
+		.A_REG                     (0),
+		.B_REG                     (0),
+		.D_REG                     (0),
+		.TOP_8x8_MULT_REG          (0),
+		.BOT_8x8_MULT_REG          (0),
+		.PIPELINE_16x16_MULT_REG1  (0),
+		.PIPELINE_16x16_MULT_REG2  (0),
+		.TOPOUTPUT_SELECT          (0),   // 0=P, 1=Q, 2=8x8, 3=16x16
+		.TOPADDSUB_LOWERINPUT      (2),   // 0=A, 1=8x8, 2=16x16, 3=S-EXT
+		.TOPADDSUB_UPPERINPUT      (1),   // 0=Q, 1=C
+		.TOPADDSUB_CARRYSELECT     (2),   // 0=0, 1=1, 2=ACI, 3=CI
+		.BOTOUTPUT_SELECT          (0),   // 0=R, 1=S, 2=8x8, 3=16x16
+		.BOTADDSUB_LOWERINPUT      (2),   // 0=B, 1=8x8, 2=16x16, 3=S-EXT
+		.BOTADDSUB_UPPERINPUT      (1),   // 0=S, 1=D
+		.BOTADDSUB_CARRYSELECT     (2),   // 0=0, 1=1, 2=ACI, 3=CI
+		.MODE_8x8                  (0),
+		.A_SIGNED                  (0),
+		.B_SIGNED                  (0)
+	) testbench ();
+endmodule
+
+module testbench_seq_16x16_A;
+	testbench #(
+		.NEG_TRIGGER               (0),
+		.C_REG                     (1),
+		.A_REG                     (1),
+		.B_REG                     (1),
+		.D_REG                     (1),
+		.TOP_8x8_MULT_REG          (1),
+		.BOT_8x8_MULT_REG          (1),
+		.PIPELINE_16x16_MULT_REG1  (1),
+		.PIPELINE_16x16_MULT_REG2  (1),
+		.TOPOUTPUT_SELECT          (0),   // 0=P, 1=Q, 2=8x8, 3=16x16
+		.TOPADDSUB_LOWERINPUT      (2),   // 0=A, 1=8x8, 2=16x16, 3=S-EXT
+		.TOPADDSUB_UPPERINPUT      (1),   // 0=Q, 1=C
+		.TOPADDSUB_CARRYSELECT     (2),   // 0=0, 1=1, 2=ACI, 3=CI
+		.BOTOUTPUT_SELECT          (0),   // 0=R, 1=S, 2=8x8, 3=16x16
+		.BOTADDSUB_LOWERINPUT      (2),   // 0=B, 1=8x8, 2=16x16, 3=S-EXT
+		.BOTADDSUB_UPPERINPUT      (1),   // 0=S, 1=D
+		.BOTADDSUB_CARRYSELECT     (2),   // 0=0, 1=1, 2=ACI, 3=CI
+		.MODE_8x8                  (0),
+		.A_SIGNED                  (0),
+		.B_SIGNED                  (0)
+	) testbench ();
+endmodule
+
+module testbench_seq_16x16_B;
+	testbench #(
+		.NEG_TRIGGER               (0),
+		.C_REG                     (1),
+		.A_REG                     (1),
+		.B_REG                     (1),
+		.D_REG                     (1),
+		.TOP_8x8_MULT_REG          (1),
+		.BOT_8x8_MULT_REG          (1),
+		.PIPELINE_16x16_MULT_REG1  (1),
+		.PIPELINE_16x16_MULT_REG2  (0),
+		.TOPOUTPUT_SELECT          (1),   // 0=P, 1=Q, 2=8x8, 3=16x16
+		.TOPADDSUB_LOWERINPUT      (2),   // 0=A, 1=8x8, 2=16x16, 3=S-EXT
+		.TOPADDSUB_UPPERINPUT      (0),   // 0=Q, 1=C
+		.TOPADDSUB_CARRYSELECT     (2),   // 0=0, 1=1, 2=ACI, 3=CI
+		.BOTOUTPUT_SELECT          (1),   // 0=R, 1=S, 2=8x8, 3=16x16
+		.BOTADDSUB_LOWERINPUT      (2),   // 0=B, 1=8x8, 2=16x16, 3=S-EXT
+		.BOTADDSUB_UPPERINPUT      (0),   // 0=S, 1=D
+		.BOTADDSUB_CARRYSELECT     (2),   // 0=0, 1=1, 2=ACI, 3=CI
+		.MODE_8x8                  (0),
+		.A_SIGNED                  (0),
+		.B_SIGNED                  (0)
+	) testbench ();
+endmodule