Work in progress.

4 files changed, 246 insertions, 132 deletions

diff --git a/src/rtl/modexpa7_exponentiator.v b/src/rtl/modexpa7_exponentiator.v
index b33360a..93c8047 100644
--- a/src/rtl/modexpa7_exponentiator.v
+++ b/src/rtl/modexpa7_exponentiator.v
@@ -58,6 +58,8 @@ module modexpa7_exponentiator #
 		input											ena,

 		output										rdy,

 		

+		input											crt,

+		

 		output	[OPERAND_ADDR_WIDTH-1:0]	m_bram_addr,

 		output	[OPERAND_ADDR_WIDTH-1:0]	d_bram_addr,

 		output	[OPERAND_ADDR_WIDTH-1:0]	f_bram_addr,

@@ -86,76 +88,120 @@ module modexpa7_exponentiator #
 		//

 		// FSM Declaration

 		//

-	localparam	[ 7: 0]	FSM_STATE_EXP_IDLE	= 8'h00;

+	localparam	[ 7: 0]	FSM_STATE_EXP_IDLE		= 8'h00;

+	//

+	localparam	[ 7: 0]	FSM_STATE_EXP_INIT_1		= 8'hA1;

+	localparam	[ 7: 0]	FSM_STATE_EXP_INIT_2		= 8'hA2;

+	localparam	[ 7: 0]	FSM_STATE_EXP_INIT_3		= 8'hA3;

+	localparam	[ 7: 0]	FSM_STATE_EXP_INIT_4		= 8'hA4;

+

+	localparam	[ 7: 0]	FSM_STATE_EXP_LOAD_1		= 8'hB1;

+	localparam	[ 7: 0]	FSM_STATE_EXP_LOAD_2		= 8'hB2;

+	localparam	[ 7: 0]	FSM_STATE_EXP_LOAD_3		= 8'hB3;

+	localparam	[ 7: 0]	FSM_STATE_EXP_LOAD_4		= 8'hB4;

+

+	localparam	[ 7: 0]	FSM_STATE_EXP_CALC_1		= 8'hC1;

+	localparam	[ 7: 0]	FSM_STATE_EXP_CALC_2		= 8'hC2;

+	localparam	[ 7: 0]	FSM_STATE_EXP_CALC_3		= 8'hC3;

+

+	localparam	[ 7: 0]	FSM_STATE_EXP_FILL_1		= 8'hD1;

+	localparam	[ 7: 0]	FSM_STATE_EXP_FILL_2		= 8'hD2;

+	localparam	[ 7: 0]	FSM_STATE_EXP_FILL_3		= 8'hD3;

+	localparam	[ 7: 0]	FSM_STATE_EXP_FILL_4		= 8'hD4;

+

+	localparam	[ 7: 0]	FSM_STATE_EXP_NEXT		= 8'hE0;

+

+	localparam	[ 7: 0]	FSM_STATE_EXP_SAVE_1		= 8'hF1;

+	localparam	[ 7: 0]	FSM_STATE_EXP_SAVE_2		= 8'hF2;

+	localparam	[ 7: 0]	FSM_STATE_EXP_SAVE_3		= 8'hF3;

+	localparam	[ 7: 0]	FSM_STATE_EXP_SAVE_4		= 8'hF4;

+	//

+	localparam	[ 7: 0]	FSM_STATE_MUL_INIT_1		= 8'h11;

+	localparam	[ 7: 0]	FSM_STATE_MUL_INIT_2		= 8'h12;

+	localparam	[ 7: 0]	FSM_STATE_MUL_INIT_3		= 8'h13;

+	localparam	[ 7: 0]	FSM_STATE_MUL_INIT_4		= 8'h14;

+

+	localparam	[ 7: 0]	FSM_STATE_MUL_CALC_1		= 8'h21;

+	localparam	[ 7: 0]	FSM_STATE_MUL_CALC_2		= 8'h22;

+	localparam	[ 7: 0]	FSM_STATE_MUL_CALC_3		= 8'h23;

 	//

-	localparam	[ 7: 0]	FSM_STATE_EXP_INIT_1	= 8'hA1;

-	localparam	[ 7: 0]	FSM_STATE_EXP_INIT_2	= 8'hA2;

-	localparam	[ 7: 0]	FSM_STATE_EXP_INIT_3	= 8'hA3;

-	localparam	[ 7: 0]	FSM_STATE_EXP_INIT_4	= 8'hA4;

-

-	localparam	[ 7: 0]	FSM_STATE_EXP_LOAD_1	= 8'hB1;

-	localparam	[ 7: 0]	FSM_STATE_EXP_LOAD_2	= 8'hB2;

-	localparam	[ 7: 0]	FSM_STATE_EXP_LOAD_3	= 8'hB3;

-	localparam	[ 7: 0]	FSM_STATE_EXP_LOAD_4	= 8'hB4;

-

-	localparam	[ 7: 0]	FSM_STATE_EXP_CALC_1	= 8'hC1;

-	localparam	[ 7: 0]	FSM_STATE_EXP_CALC_2	= 8'hC2;

-	localparam	[ 7: 0]	FSM_STATE_EXP_CALC_3	= 8'hC3;

-

-	localparam	[ 7: 0]	FSM_STATE_EXP_FILL_1	= 8'hD1;

-	localparam	[ 7: 0]	FSM_STATE_EXP_FILL_2	= 8'hD2;

-	localparam	[ 7: 0]	FSM_STATE_EXP_FILL_3	= 8'hD3;

-	localparam	[ 7: 0]	FSM_STATE_EXP_FILL_4	= 8'hD4;

-

-	localparam	[ 7: 0]	FSM_STATE_EXP_NEXT	= 8'hE0;

-

-	localparam	[ 7: 0]	FSM_STATE_EXP_SAVE_1	= 8'hF1;

-	localparam	[ 7: 0]	FSM_STATE_EXP_SAVE_2	= 8'hF2;

-	localparam	[ 7: 0]	FSM_STATE_EXP_SAVE_3	= 8'hF3;

-	localparam	[ 7: 0]	FSM_STATE_EXP_SAVE_4	= 8'hF4;

+	localparam	[ 7: 0]	FSM_STATE_CRT_INIT_A_1	= 8'h31;

+	localparam	[ 7: 0]	FSM_STATE_CRT_INIT_A_2	= 8'h32;

+	localparam	[ 7: 0]	FSM_STATE_CRT_INIT_A_3	= 8'h33;

+	localparam	[ 7: 0]	FSM_STATE_CRT_INIT_A_4	= 8'h34;

+

+	localparam	[ 7: 0]	FSM_STATE_CRT_CALC_A_1	= 8'h41;

+	localparam	[ 7: 0]	FSM_STATE_CRT_CALC_A_2	= 8'h42;

+	localparam	[ 7: 0]	FSM_STATE_CRT_CALC_A_3	= 8'h43;

 	//

-	localparam	[ 7: 0]	FSM_STATE_MUL_INIT_1	= 8'h11;

-	localparam	[ 7: 0]	FSM_STATE_MUL_INIT_2	= 8'h12;

-	localparam	[ 7: 0]	FSM_STATE_MUL_INIT_3	= 8'h13;

-	localparam	[ 7: 0]	FSM_STATE_MUL_INIT_4	= 8'h14;

-

-	localparam	[ 7: 0]	FSM_STATE_MUL_CALC_1	= 8'h21;

-	localparam	[ 7: 0]	FSM_STATE_MUL_CALC_2	= 8'h22;

-	localparam	[ 7: 0]	FSM_STATE_MUL_CALC_3	= 8'h23;

+	localparam	[ 7: 0]	FSM_STATE_CRT_INIT_B_1	= 8'h51;

+	localparam	[ 7: 0]	FSM_STATE_CRT_INIT_B_2	= 8'h52;

+	localparam	[ 7: 0]	FSM_STATE_CRT_INIT_B_3	= 8'h53;

+	localparam	[ 7: 0]	FSM_STATE_CRT_INIT_B_4	= 8'h54;

+

+	localparam	[ 7: 0]	FSM_STATE_CRT_CALC_B_1	= 8'h61;

+	localparam	[ 7: 0]	FSM_STATE_CRT_CALC_B_2	= 8'h62;

+	localparam	[ 7: 0]	FSM_STATE_CRT_CALC_B_3	= 8'h63;

 	//

-	localparam	[ 7: 0]	FSM_STATE_EXP_STOP	= 8'hFF;

+	localparam	[ 7: 0]	FSM_STATE_CRT_INIT_C_1	= 8'h71;

+	localparam	[ 7: 0]	FSM_STATE_CRT_INIT_C_2	= 8'h72;

+	localparam	[ 7: 0]	FSM_STATE_CRT_INIT_C_3	= 8'h73;

+	localparam	[ 7: 0]	FSM_STATE_CRT_INIT_C_4	= 8'h74;

+

+	localparam	[ 7: 0]	FSM_STATE_CRT_CALC_C_1	= 8'h81;

+	localparam	[ 7: 0]	FSM_STATE_CRT_CALC_C_2	= 8'h82;

+	localparam	[ 7: 0]	FSM_STATE_CRT_CALC_C_3	= 8'h83;

+	//

+	localparam	[ 7: 0]	FSM_STATE_EXP_STOP		= 8'hFF;

 

 

 	/*

 	 *  //

 	 *

-	 *  MUL_INIT:	P1 = F

-	 *             P2 = F

-	 *             P3 = F

-	 *             T2 = M

+	 *  MUL_INIT:		P1 <= F

+	 *             	P2 <= F

+	 *             	P3 <= F

+	 *            		T2 <= M

 	 *

-	 *  MUL_CALC:	TP = T2 * P3

+	 *  MUL_CALC:		TP = T2 * P3

 	 *

 	 *  //

 	 *

-	 *  EXP_INIT:	P1 <= TP

-	 *					P2 <= TP

-	 *					P3 <= TP

-	 *					T1 <= 1

-	 *					T2 <= 1

+	 *  CRT_INIT_A:	T2 <= M

+	 *

+	 *  CRT_CALC_A:	TP = T2 * P3 ("reduce only")

 	 *

-	 *  EXP_LOAD:	T0 <= T1

+	 *  CRT_INIT_B:  	P1 <= F

+	 *						P2 <= F

+	 *						P3 <= F

+	 *						T2 <= TP

 	 *

-	 *  EXP_CALC:	PP = P1 * P2

-	 *					TP = T2 * P3

+	 *  CRT_CALC_B:	TP = T2 * P3

 	 *

-	 *  EXP_FILL:	P1 <= PP

-	 *					P2 <= PP

-	 *					P3 <= PP

-	 *					T1 <= D[i] ? TP : T0

-	 *					T2 <= D[i] ? TP : T0

+	 *  CRT_INIT_C:  	T2 <= TP

 	 *

-	 *  EXP_SAVE:	R  <=  T1

+	 *  CRT_CALC_C:	TP = T2 * P3

+	 *

+	 *  //

+	 *

+	 *  EXP_INIT:		P1 <= TP

+	 *						P2 <= TP

+	 *						P3 <= TP

+	 *						T1 <= 1

+	 *						T2 <= 1

+	 *

+	 *  EXP_LOAD:		T0 <= T1

+	 *

+	 *  EXP_CALC:		PP = P1 * P2

+	 *						TP = T2 * P3

+	 *

+	 *  EXP_FILL:		P1 <= PP

+	 *						P2 <= PP

+	 *						P3 <= PP

+	 *						T1 <= D[i] ? TP : T0

+	 *						T2 <= D[i] ? TP : T0

+	 *

+	 *  EXP_SAVE:		R  <=  T1

 	 *

 	 *  //

 	 *

@@ -225,10 +271,12 @@ module modexpa7_exponentiator #
 		 */

 		

 		/* the very first addresses */

-	wire	[OPERAND_ADDR_WIDTH-1:0]	bram_addr_zero		= {{OPERAND_ADDR_WIDTH{1'b0}}};

+	wire	[OPERAND_ADDR_WIDTH-1:0]	bram_addr_zero			= {{OPERAND_ADDR_WIDTH{1'b0}}};

 	

 		/* the very last addresses */

-	wire	[OPERAND_ADDR_WIDTH-1:0]	bram_addr_last		= {m_num_words_latch};

+	wire	[OPERAND_ADDR_WIDTH-1:0]	bram_addr_last			= {m_num_words_latch};

+	wire	[OPERAND_ADDR_WIDTH-1:0]	bram_addr_last_crt	=

+		{m_num_words_latch[OPERAND_ADDR_WIDTH-2:0], 1'b1};

 

 		/* address registers */

 	reg	[OPERAND_ADDR_WIDTH-1:0]	m_addr;

@@ -261,16 +309,18 @@ module modexpa7_exponentiator #
 	wire	[OPERAND_ADDR_WIDTH-1:0]	tp_addr_rd_next	= tp_addr_rd + 1'b1;

 	

 		/* handy stop flags */

-	wire	m_addr_done			= (m_addr     == bram_addr_last) ? 1'b1 : 1'b0;

-	wire	d_addr_done			= (d_addr     == bram_addr_last) ? 1'b1 : 1'b0;

-	wire	f_addr_done			= (f_addr     == bram_addr_last) ? 1'b1 : 1'b0;

-	wire	r_addr_done			= (r_addr     == bram_addr_last) ? 1'b1 : 1'b0;

-	wire	t0_addr_done		= (t0_addr    == bram_addr_last) ? 1'b1 : 1'b0;

-	wire	t1_addr_done		= (t1_addr    == bram_addr_last) ? 1'b1 : 1'b0;

-	wire	t2_addr_wr_done	= (t2_addr_wr == bram_addr_last) ? 1'b1 : 1'b0;

-	wire	p_addr_wr_done		= (p_addr_wr  == bram_addr_last) ? 1'b1 : 1'b0;

-	wire	pp_addr_rd_done	= (pp_addr_rd == bram_addr_last) ? 1'b1 : 1'b0;

-	wire	tp_addr_rd_done	= (tp_addr_rd == bram_addr_last) ? 1'b1 : 1'b0;

+	wire	m_addr_done				= (m_addr     == bram_addr_last)     ? 1'b1 : 1'b0;

+	wire	m_addr_done_crt		= (m_addr     == bram_addr_last_crt) ? 1'b1 : 1'b0;

+	wire	d_addr_done				= (d_addr     == bram_addr_last)     ? 1'b1 : 1'b0;

+	wire	f_addr_done				= (f_addr     == bram_addr_last)     ? 1'b1 : 1'b0;

+	wire	r_addr_done				= (r_addr     == bram_addr_last)     ? 1'b1 : 1'b0;

+	wire	t0_addr_done			= (t0_addr    == bram_addr_last)     ? 1'b1 : 1'b0;

+	wire	t1_addr_done			= (t1_addr    == bram_addr_last)     ? 1'b1 : 1'b0;

+	wire	t2_addr_wr_done		= (t2_addr_wr == bram_addr_last)     ? 1'b1 : 1'b0;

+	wire	t2_addr_wr_done_crt	= (t2_addr_wr == bram_addr_last_crt) ? 1'b1 : 1'b0;

+	wire	p_addr_wr_done			= (p_addr_wr  == bram_addr_last)     ? 1'b1 : 1'b0;

+	wire	pp_addr_rd_done		= (pp_addr_rd == bram_addr_last)     ? 1'b1 : 1'b0;

+	wire	tp_addr_rd_done		= (tp_addr_rd == bram_addr_last)     ? 1'b1 : 1'b0;

 				

 		/* map registers to top-level ports */

 	assign m_bram_addr = m_addr;

@@ -392,10 +442,15 @@ module modexpa7_exponentiator #
 		// m_addr

 		//

 		case (fsm_next_state)

-			FSM_STATE_MUL_INIT_1: 	m_addr <= bram_addr_zero;

+			FSM_STATE_MUL_INIT_1: 		m_addr <= bram_addr_zero;

 			FSM_STATE_MUL_INIT_2,

 			FSM_STATE_MUL_INIT_3,

-			FSM_STATE_MUL_INIT_4:	m_addr <= !m_addr_done ? m_addr_next : m_addr;

+			FSM_STATE_MUL_INIT_4:		m_addr <= !m_addr_done ? m_addr_next : m_addr;

+			//

+			FSM_STATE_CRT_INIT_A_1: 	m_addr <= bram_addr_zero;

+			FSM_STATE_CRT_INIT_A_2,

+			FSM_STATE_CRT_INIT_A_3,

+			FSM_STATE_CRT_INIT_A_4:		m_addr <= !m_addr_done_crt ? m_addr_next : m_addr;

 		endcase

 		//

 		// d_addr

@@ -472,7 +527,10 @@ module modexpa7_exponentiator #
 			//

 			FSM_STATE_MUL_INIT_3:	t2_addr_wr <= bram_addr_zero;

 			FSM_STATE_MUL_INIT_4:	t2_addr_wr <= t2_addr_wr_next;

-

+			//

+			FSM_STATE_CRT_INIT_A_3:	t2_addr_wr <= bram_addr_zero;

+			FSM_STATE_CRT_INIT_A_4:	t2_addr_wr <= t2_addr_wr_next;

+			//

 			FSM_STATE_EXP_INIT_3:	t2_addr_wr <= bram_addr_zero;

 			FSM_STATE_EXP_INIT_4:	t2_addr_wr <= t2_addr_wr_next;

 			//

@@ -554,6 +612,8 @@ module modexpa7_exponentiator #
 		case (fsm_next_state)

 			FSM_STATE_MUL_INIT_3,

 			FSM_STATE_MUL_INIT_4,

+			FSM_STATE_CRT_INIT_A_3,

+			FSM_STATE_CRT_INIT_A_4,

 			FSM_STATE_EXP_INIT_3,		

 			FSM_STATE_EXP_INIT_4,

 			FSM_STATE_EXP_FILL_3,

@@ -616,15 +676,19 @@ module modexpa7_exponentiator #
 		//

 		case (fsm_next_state)

 			//

-			FSM_STATE_MUL_INIT_3,	

-			FSM_STATE_MUL_INIT_4:	t2_data_in <= m_bram_out;

+			FSM_STATE_MUL_INIT_3,

+			FSM_STATE_MUL_INIT_4:		t2_data_in <= m_bram_out;

+			//

+			FSM_STATE_CRT_INIT_A_3,

+			FSM_STATE_CRT_INIT_A_4:		t2_data_in <= m_bram_out;

+

 			//

-			FSM_STATE_EXP_INIT_3:	t2_data_in <= 32'd1;

-			FSM_STATE_EXP_INIT_4:	t2_data_in <= 32'd0;

+			FSM_STATE_EXP_INIT_3:		t2_data_in <= 32'd1;

+			FSM_STATE_EXP_INIT_4:		t2_data_in <= 32'd0;

 			//

 			FSM_STATE_EXP_FILL_3,

-			FSM_STATE_EXP_FILL_4:	t2_data_in <= flag_update_r ? tp_data_out : t0_data_out;

-			default:						t2_data_in <= 32'dX;

+			FSM_STATE_EXP_FILL_4:		t2_data_in <= flag_update_r ? tp_data_out : t0_data_out;

+			default:							t2_data_in <= 32'dX;

 		endcase		

 		//

 	end

@@ -634,6 +698,7 @@ module modexpa7_exponentiator #
 		// Double Multiplier

 		//

 	reg	mul_ena;

+	reg	mul_crt;

 	wire	mul_rdy_pp;

 	wire	mul_rdy_tp;

 	wire	mul_rdy_all = mul_rdy_pp & mul_rdy_tp;

@@ -651,6 +716,8 @@ module modexpa7_exponentiator #
 		.ena						(mul_ena),

 		.rdy						(mul_rdy_pp),

 

+		.reduce_only			(1'b0),

+

 		.a_bram_addr			(p1_addr_rd),

 		.b_bram_addr			(p2_addr_rd),

 		.n_bram_addr			(n1_bram_addr),

@@ -681,6 +748,8 @@ module modexpa7_exponentiator #
 		.ena						(mul_ena),

 		.rdy						(mul_rdy_tp),

 

+		.reduce_only			(mul_crt),

+

 		.a_bram_addr			(t2_addr_rd),

 		.b_bram_addr			(p3_addr_rd),

 		.n_bram_addr			(n2_bram_addr),

@@ -703,8 +772,18 @@ module modexpa7_exponentiator #
 		//

 		case (fsm_next_state)

 			FSM_STATE_MUL_CALC_1,

-			FSM_STATE_EXP_CALC_1:	mul_ena <= 1'b1;

-			default:						mul_ena <= 1'b0;

+			FSM_STATE_CRT_CALC_A_1,

+			FSM_STATE_CRT_CALC_B_1,

+			FSM_STATE_CRT_CALC_C_1,

+			FSM_STATE_EXP_CALC_1:		mul_ena <= 1'b1;

+			default:							mul_ena <= 1'b0;

+		endcase

+

+	always @(posedge clk)

+		//

+		case (fsm_next_state)

+			FSM_STATE_CRT_CALC_A_1:		mul_crt <= 1'b1;

+			default:							mul_crt <= 1'b0;

 		endcase

 			

 

@@ -726,53 +805,70 @@ module modexpa7_exponentiator #
 		//

 		case (fsm_state)

 			//

-			FSM_STATE_MUL_INIT_1:									fsm_next_state = FSM_STATE_MUL_INIT_2;

-			FSM_STATE_MUL_INIT_2:									fsm_next_state = FSM_STATE_MUL_INIT_3;

-			FSM_STATE_MUL_INIT_3:									fsm_next_state = FSM_STATE_MUL_INIT_4;

-			FSM_STATE_MUL_INIT_4:	if (t2_addr_wr_done)		fsm_next_state = FSM_STATE_MUL_CALC_1;

-											else							fsm_next_state = FSM_STATE_MUL_INIT_4;

-			//

-			FSM_STATE_MUL_CALC_1:									fsm_next_state = FSM_STATE_MUL_CALC_2;

-			FSM_STATE_MUL_CALC_2:	if (mul_rdy_tp)			fsm_next_state = FSM_STATE_MUL_CALC_3;

-											else							fsm_next_state = FSM_STATE_MUL_CALC_2;

-			FSM_STATE_MUL_CALC_3:									fsm_next_state = FSM_STATE_EXP_INIT_1;

-			//

-			FSM_STATE_EXP_IDLE:		if (ena_trig)				fsm_next_state = FSM_STATE_MUL_INIT_1;

-											else							fsm_next_state = FSM_STATE_EXP_IDLE;

-			//

-			FSM_STATE_EXP_INIT_1:									fsm_next_state = FSM_STATE_EXP_INIT_2;

-			FSM_STATE_EXP_INIT_2:									fsm_next_state = FSM_STATE_EXP_INIT_3;

-			FSM_STATE_EXP_INIT_3:									fsm_next_state = FSM_STATE_EXP_INIT_4;

-			FSM_STATE_EXP_INIT_4:	if (t1_addr_done)			fsm_next_state = FSM_STATE_EXP_LOAD_1;

-											else							fsm_next_state = FSM_STATE_EXP_INIT_4;

-			//

-			FSM_STATE_EXP_LOAD_1:									fsm_next_state = FSM_STATE_EXP_LOAD_2;

-			FSM_STATE_EXP_LOAD_2:									fsm_next_state = FSM_STATE_EXP_LOAD_3;

-			FSM_STATE_EXP_LOAD_3:									fsm_next_state = FSM_STATE_EXP_LOAD_4;

-			FSM_STATE_EXP_LOAD_4:	if (t0_addr_done)			fsm_next_state = FSM_STATE_EXP_CALC_1;

-											else							fsm_next_state = FSM_STATE_EXP_LOAD_4;

-			//

-			FSM_STATE_EXP_CALC_1:									fsm_next_state = FSM_STATE_EXP_CALC_2;

-			FSM_STATE_EXP_CALC_2:	if (mul_rdy_all)			fsm_next_state = FSM_STATE_EXP_CALC_3;

-											else							fsm_next_state = FSM_STATE_EXP_CALC_2;

-			FSM_STATE_EXP_CALC_3:									fsm_next_state = FSM_STATE_EXP_FILL_1;

-			//

-			FSM_STATE_EXP_FILL_1:									fsm_next_state = FSM_STATE_EXP_FILL_2;

-			FSM_STATE_EXP_FILL_2:									fsm_next_state = FSM_STATE_EXP_FILL_3;

-			FSM_STATE_EXP_FILL_3:									fsm_next_state = FSM_STATE_EXP_FILL_4;

-			FSM_STATE_EXP_FILL_4:	if (p_addr_wr_done)		fsm_next_state = FSM_STATE_EXP_NEXT;

-											else							fsm_next_state = FSM_STATE_EXP_FILL_4;			

-			//

-			FSM_STATE_EXP_NEXT:		if (bit_cnt_done)			fsm_next_state = FSM_STATE_EXP_SAVE_1;

-											else							fsm_next_state = FSM_STATE_EXP_LOAD_1;

-			//

-			FSM_STATE_EXP_SAVE_1:									fsm_next_state = FSM_STATE_EXP_SAVE_2;

-			FSM_STATE_EXP_SAVE_2:									fsm_next_state = FSM_STATE_EXP_SAVE_3;

-			FSM_STATE_EXP_SAVE_3:									fsm_next_state = FSM_STATE_EXP_SAVE_4;

-			FSM_STATE_EXP_SAVE_4:	if (r_addr_done)			fsm_next_state = FSM_STATE_EXP_STOP;

-											else							fsm_next_state = FSM_STATE_EXP_SAVE_4;

-			//

-			FSM_STATE_EXP_STOP:										fsm_next_state = FSM_STATE_EXP_IDLE;

+			//

+			FSM_STATE_MUL_INIT_1:										fsm_next_state = FSM_STATE_MUL_INIT_2;

+			FSM_STATE_MUL_INIT_2:										fsm_next_state = FSM_STATE_MUL_INIT_3;

+			FSM_STATE_MUL_INIT_3:										fsm_next_state = FSM_STATE_MUL_INIT_4;

+			FSM_STATE_MUL_INIT_4:	if (t2_addr_wr_done)			fsm_next_state = FSM_STATE_MUL_CALC_1;

+											else								fsm_next_state = FSM_STATE_MUL_INIT_4;

+			//

+			FSM_STATE_MUL_CALC_1:										fsm_next_state = FSM_STATE_MUL_CALC_2;

+			FSM_STATE_MUL_CALC_2:	if (mul_rdy_tp)				fsm_next_state = FSM_STATE_MUL_CALC_3;

+											else								fsm_next_state = FSM_STATE_MUL_CALC_2;

+			FSM_STATE_MUL_CALC_3:										fsm_next_state = FSM_STATE_EXP_INIT_1;

+			//

+			//

+			FSM_STATE_CRT_INIT_A_1:										fsm_next_state = FSM_STATE_CRT_INIT_A_2;

+			FSM_STATE_CRT_INIT_A_2:										fsm_next_state = FSM_STATE_CRT_INIT_A_3;

+			FSM_STATE_CRT_INIT_A_3:										fsm_next_state = FSM_STATE_CRT_INIT_A_4;

+			FSM_STATE_CRT_INIT_A_4:	if (t2_addr_wr_done_crt)	fsm_next_state = FSM_STATE_CRT_CALC_A_1;

+											else								fsm_next_state = FSM_STATE_CRT_INIT_A_4;

+

+			//

+			FSM_STATE_CRT_CALC_A_1:										fsm_next_state = FSM_STATE_CRT_CALC_A_2;

+			FSM_STATE_CRT_CALC_A_2:	if (mul_rdy_tp)				fsm_next_state = FSM_STATE_CRT_CALC_A_3;

+											else								fsm_next_state = FSM_STATE_CRT_CALC_A_2;

+			FSM_STATE_CRT_CALC_A_3:										fsm_next_state = FSM_STATE_EXP_INIT_1;

+			//

+			//

+			FSM_STATE_EXP_IDLE:		if (ena_trig)					fsm_next_state = crt ?

+																					FSM_STATE_CRT_INIT_A_1 : FSM_STATE_MUL_INIT_1;

+											else								fsm_next_state = FSM_STATE_EXP_IDLE;

+			//

+			//

+			FSM_STATE_EXP_INIT_1:										fsm_next_state = FSM_STATE_EXP_INIT_2;

+			FSM_STATE_EXP_INIT_2:										fsm_next_state = FSM_STATE_EXP_INIT_3;

+			FSM_STATE_EXP_INIT_3:										fsm_next_state = FSM_STATE_EXP_INIT_4;

+			FSM_STATE_EXP_INIT_4:	if (t1_addr_done)				fsm_next_state = FSM_STATE_EXP_LOAD_1;

+											else								fsm_next_state = FSM_STATE_EXP_INIT_4;

+			//

+			FSM_STATE_EXP_LOAD_1:										fsm_next_state = FSM_STATE_EXP_LOAD_2;

+			FSM_STATE_EXP_LOAD_2:										fsm_next_state = FSM_STATE_EXP_LOAD_3;

+			FSM_STATE_EXP_LOAD_3:										fsm_next_state = FSM_STATE_EXP_LOAD_4;

+			FSM_STATE_EXP_LOAD_4:	if (t0_addr_done)				fsm_next_state = FSM_STATE_EXP_CALC_1;

+											else								fsm_next_state = FSM_STATE_EXP_LOAD_4;

+			//

+			FSM_STATE_EXP_CALC_1:										fsm_next_state = FSM_STATE_EXP_CALC_2;

+			FSM_STATE_EXP_CALC_2:	if (mul_rdy_all)				fsm_next_state = FSM_STATE_EXP_CALC_3;

+											else								fsm_next_state = FSM_STATE_EXP_CALC_2;

+			FSM_STATE_EXP_CALC_3:										fsm_next_state = FSM_STATE_EXP_FILL_1;

+			//

+			FSM_STATE_EXP_FILL_1:										fsm_next_state = FSM_STATE_EXP_FILL_2;

+			FSM_STATE_EXP_FILL_2:										fsm_next_state = FSM_STATE_EXP_FILL_3;

+			FSM_STATE_EXP_FILL_3:										fsm_next_state = FSM_STATE_EXP_FILL_4;

+			FSM_STATE_EXP_FILL_4:	if (p_addr_wr_done)			fsm_next_state = FSM_STATE_EXP_NEXT;

+											else								fsm_next_state = FSM_STATE_EXP_FILL_4;			

+			//

+			FSM_STATE_EXP_NEXT:		if (bit_cnt_done)				fsm_next_state = FSM_STATE_EXP_SAVE_1;

+											else								fsm_next_state = FSM_STATE_EXP_LOAD_1;

+			//

+			FSM_STATE_EXP_SAVE_1:										fsm_next_state = FSM_STATE_EXP_SAVE_2;

+			FSM_STATE_EXP_SAVE_2:										fsm_next_state = FSM_STATE_EXP_SAVE_3;

+			FSM_STATE_EXP_SAVE_3:										fsm_next_state = FSM_STATE_EXP_SAVE_4;

+			FSM_STATE_EXP_SAVE_4:	if (r_addr_done)				fsm_next_state = FSM_STATE_EXP_STOP;

+											else								fsm_next_state = FSM_STATE_EXP_SAVE_4;

+			//

+			FSM_STATE_EXP_STOP:											fsm_next_state = FSM_STATE_EXP_IDLE;

 			//

 		endcase

 		//

diff --git a/src/rtl/modexpa7_systolic_multiplier.v b/src/rtl/modexpa7_systolic_multiplier.v
index 7293998..444693d 100644
--- a/src/rtl/modexpa7_systolic_multiplier.v
+++ b/src/rtl/modexpa7_systolic_multiplier.v
@@ -57,6 +57,8 @@ module modexpa7_systolic_multiplier #
 
 		input											ena,
 		output										rdy,
+

+		input											reduce_only,

 
 		output	[OPERAND_ADDR_WIDTH-1:0]	a_bram_addr,
 		output	[OPERAND_ADDR_WIDTH-1:0]	b_bram_addr,
@@ -155,7 +157,8 @@ module modexpa7_systolic_multiplier #
 		 * Parameters Latch
 		 */
 	reg	[OPERAND_ADDR_WIDTH-1:0]	n_num_words_latch;
-	reg	[OPERAND_ADDR_WIDTH  :0]	p_num_words_latch;
+	reg	[OPERAND_ADDR_WIDTH  :0]	p_num_words_latch;

+	reg										reduce_only_latch;
 
 		// save number of words in n when new operation starts
 	always @(posedge clk)
@@ -163,7 +166,12 @@ module modexpa7_systolic_multiplier #
 		if ((fsm_state == FSM_STATE_IDLE) && ena_trig)
 			n_num_words_latch <= n_num_words;
 			

+	always @(posedge clk)

+		//

+		if ((fsm_state == FSM_STATE_IDLE) && ena_trig)

+			reduce_only_latch <= reduce_only;

 			

+		

 		/*

 		 * Multiplication Phase

 		 */

@@ -174,6 +182,7 @@ module modexpa7_systolic_multiplier #
 	

 	reg	[ 1: 0]	mult_phase;

 	

+	wire	mult_phase_ab   = (mult_phase == MULT_PHASE_A_B)   ? 1'b1 : 1'b0;

 	wire	mult_phase_done = (mult_phase == MULT_PHASE_STALL) ? 1'b1 : 1'b0;

 	
    always @(posedge clk)
@@ -296,6 +305,7 @@ module modexpa7_systolic_multiplier #
 	wire	[OPERAND_ADDR_WIDTH  :0]	bram_addr_ext_last = {n_num_words_latch, 1'b1};
 
 		// address registers

+	wire	[OPERAND_ADDR_WIDTH-1:0]	a_addr;

 	reg	[OPERAND_ADDR_WIDTH-1:0]	b_addr;
 	reg	[OPERAND_ADDR_WIDTH-1:0]	n_addr;
 	wire	[OPERAND_ADDR_WIDTH  :0]	p_addr_ext_wr;
@@ -570,8 +580,9 @@ module modexpa7_systolic_multiplier #
 				MULT_PHASE_Q_N:			p_num_words_latch <= {n_num_words_latch, 1'b1};
 			endcase

 			

-	assign n_coeff_bram_addr = a_bram_addr;

-	assign q_addr_rd = a_bram_addr;

+	assign a_bram_addr = a_addr;

+	assign n_coeff_bram_addr = a_addr;

+	assign q_addr_rd = a_addr;

 	

 	reg	[31: 0]	a_data_out;

 	

@@ -597,12 +608,14 @@ module modexpa7_systolic_multiplier #
 		.ena					(pe_array_ena),
 		.rdy					(pe_array_rdy),
 

+		.crt					(reduce_only_latch && mult_phase_ab),

+

 		.loader_addr_rd	(loader_addr_rd),

 		

 		.pe_a_wide			({SYSTOLIC_ARRAY_LENGTH{a_data_out}}),

 		.pe_b_wide			(pe_b_wide),

 		

-		.a_bram_addr		(a_bram_addr),

+		.a_bram_addr		(a_addr),

 		

 		.p_bram_addr		(p_addr_ext_wr),
 		.p_bram_in			(p_data_in),
diff --git a/src/rtl/modexpa7_systolic_multiplier_array.v b/src/rtl/modexpa7_systolic_multiplier_array.v
index 754203d..3280010 100644
--- a/src/rtl/modexpa7_systolic_multiplier_array.v
+++ b/src/rtl/modexpa7_systolic_multiplier_array.v
@@ -48,6 +48,8 @@ module modexpa7_systolic_multiplier_array #
 		input																				ena,
 		output																			rdy,
 

+		input																				crt,

+

 		output	[OPERAND_ADDR_WIDTH - SYSTOLIC_ARRAY_POWER - 1 : 0]	loader_addr_rd,
 

 		input		[         32 * (2 ** SYSTOLIC_ARRAY_POWER) - 1 : 0]	pe_a_wide,

@@ -385,6 +387,8 @@ module modexpa7_systolic_multiplier_array #
 	

 		// the very last address

 	wire	[OPERAND_ADDR_WIDTH - 1 : 0]	bram_addr_last     = n_num_words_latch;

+	wire	[OPERAND_ADDR_WIDTH - 1 : 0]	bram_addr_last_crt =

+		{n_num_words_latch[OPERAND_ADDR_WIDTH-2:0], 1'b1};

 	wire	[OPERAND_ADDR_WIDTH     : 0]	bram_addr_ext_last = p_num_words_latch;

 		

 		// registers

@@ -398,8 +402,9 @@ module modexpa7_systolic_multiplier_array #
 	wire	[OPERAND_ADDR_WIDTH     : 0]	p_addr_next = p_addr + 1'b1;

 	

 		// handy flags

-	wire	a_addr_done = (a_addr == bram_addr_last)     ? 1'b1 : 1'b0;

-	wire	p_addr_done = (p_addr == bram_addr_ext_last) ? 1'b1 : 1'b0;

+	wire	a_addr_done     = (a_addr == bram_addr_last)     ? 1'b1 : 1'b0;

+	wire	a_addr_done_crt = (a_addr == bram_addr_last_crt) ? 1'b1 : 1'b0;

+	wire	p_addr_done     = (p_addr == bram_addr_ext_last) ? 1'b1 : 1'b0;

 	

 		// map top-level ports to internal registers

 	assign a_bram_addr	= a_addr;

@@ -452,7 +457,7 @@ module modexpa7_systolic_multiplier_array #
 		//

 		case (fsm_next_state)
 			FSM_STATE_MULT_START:	a_addr <= bram_addr_zero;
-			FSM_STATE_MULT_RELOAD:	a_addr <= !a_addr_done ? a_addr_next : a_addr;
+			FSM_STATE_MULT_RELOAD:	crt ? //a_addr <= !a_addr_done ? a_addr_next : a_addr;
 		endcase
 		//
 	end
diff --git a/src/rtl/pe/modexpa7_primitive_switch.v b/src/rtl/pe/modexpa7_primitive_switch.v
index fa958ec..17e8264 100644
--- a/src/rtl/pe/modexpa7_primitive_switch.v
+++ b/src/rtl/pe/modexpa7_primitive_switch.v
@@ -1,4 +1,4 @@
-`define USE_VENDOR_PRIMITIVES

+//`define USE_VENDOR_PRIMITIVES

 

 `ifdef USE_VENDOR_PRIMITIVES