Further work:

 - added core wrapper
 - fixed module resets across entire core (all the resets are now consistently
   active-low)
 - continued refactoring

1 files changed, 247 insertions, 249 deletions

diff --git a/rtl/modexpng_mmm_dual.v b/rtl/modexpng_mmm_dual.v
index 13a8773..a868362 100644
--- a/rtl/modexpng_mmm_dual.v
+++ b/rtl/modexpng_mmm_dual.v
@@ -1,6 +1,6 @@
 module modexpng_mmm_dual
 (
-    clk, rst,
+    clk, rst_n,
     
     ena, rdy,
         
@@ -56,16 +56,14 @@ module modexpng_mmm_dual
     // Headers
     //
     `include "modexpng_parameters.vh"
-    `include "../rtl_1/modexpng_mmm_fsm_old.vh"
-    //`include "../rtl_1/modexpng_parameters_old.vh"
-    //`include "../rtl_1/modexpng_parameters_x8_old.vh"
+    `include "modexpng_mmm_dual_fsm.vh"
 
 
     //
     // Ports
     //
     input                        clk;
-    input                        rst;
+    input                        rst_n;
     
     input                        ena;
     output                       rdy;
@@ -122,23 +120,23 @@ module modexpng_mmm_dual
     //
     // FSM Declaration
     //
-    reg [FSM_STATE_WIDTH-1:0] fsm_state = FSM_STATE_IDLE;
-    reg [FSM_STATE_WIDTH-1:0] fsm_state_next;
+    reg [MMM_FSM_STATE_W-1:0] fsm_state = MMM_FSM_STATE_IDLE;
+    reg [MMM_FSM_STATE_W-1:0] fsm_state_next;
 
-    wire [FSM_STATE_WIDTH-1:0] fsm_state_after_idle;    
-    wire [FSM_STATE_WIDTH-1:0] fsm_state_after_mult_square;
-    wire [FSM_STATE_WIDTH-1:0] fsm_state_after_mult_triangle;
-    wire [FSM_STATE_WIDTH-1:0] fsm_state_after_mult_rectangle;
-    wire [FSM_STATE_WIDTH-1:0] fsm_state_after_square_holdoff;
+    wire [MMM_FSM_STATE_W-1:0] fsm_state_after_idle;    
+    wire [MMM_FSM_STATE_W-1:0] fsm_state_after_mult_square;
+    wire [MMM_FSM_STATE_W-1:0] fsm_state_after_mult_triangle;
+    wire [MMM_FSM_STATE_W-1:0] fsm_state_after_mult_rectangle;
+    wire [MMM_FSM_STATE_W-1:0] fsm_state_after_square_holdoff;
 
     
     //
     // FSM Process
     //
-    always @(posedge clk)
+    always @(posedge clk or negedge rst_n)
         //
-        if (rst) fsm_state <= FSM_STATE_IDLE;
-        else     fsm_state <= fsm_state_next;
+        if (!rst_n) fsm_state <= MMM_FSM_STATE_IDLE;
+        else        fsm_state <= fsm_state_next;
 
         
     //
@@ -193,9 +191,9 @@ module modexpng_mmm_dual
         //
         case (fsm_state_next)
             //
-            FSM_STATE_MULT_SQUARE_COL_0_INIT,
-            FSM_STATE_MULT_TRIANGLE_COL_0_INIT,
-            FSM_STATE_MULT_RECTANGLE_COL_0_INIT: begin
+            MMM_FSM_STATE_MULT_SQUARE_COL_0_INIT,
+            MMM_FSM_STATE_MULT_TRIANGLE_COL_0_INIT,
+            MMM_FSM_STATE_MULT_RECTANGLE_COL_0_INIT: begin
                 col_index       <= 5'd0;
                 col_index_last  <= word_index_last[7:3];
                 col_index_next  <= 5'd1;
@@ -203,9 +201,9 @@ module modexpng_mmm_dual
                 
             end
             //
-            FSM_STATE_MULT_SQUARE_COL_N_INIT,
-            FSM_STATE_MULT_TRIANGLE_COL_N_INIT,
-            FSM_STATE_MULT_RECTANGLE_COL_N_INIT: begin
+            MMM_FSM_STATE_MULT_SQUARE_COL_N_INIT,
+            MMM_FSM_STATE_MULT_TRIANGLE_COL_N_INIT,
+            MMM_FSM_STATE_MULT_RECTANGLE_COL_N_INIT: begin
                 col_index <= col_index_next;
                 col_is_last <= col_index_next == col_index_last;
                 col_index_next <= col_index_next == col_index_last ? 5'd0 : col_index_next + 5'd1;   
@@ -242,8 +240,8 @@ module modexpng_mmm_dual
         //
         case (fsm_state)
             //
-            FSM_STATE_MULT_SQUARE_COL_0_BUSY,
-            FSM_STATE_MULT_SQUARE_COL_N_BUSY:
+            MMM_FSM_STATE_MULT_SQUARE_COL_0_BUSY,
+            MMM_FSM_STATE_MULT_SQUARE_COL_N_BUSY:
                 square_almost_done_flop <= square_almost_done_comb;
             //
             default:
@@ -262,8 +260,8 @@ module modexpng_mmm_dual
         //
         case (fsm_state)
             //
-            FSM_STATE_MULT_TRIANGLE_COL_0_BUSY,
-            FSM_STATE_MULT_TRIANGLE_COL_N_BUSY:
+            MMM_FSM_STATE_MULT_TRIANGLE_COL_0_BUSY,
+            MMM_FSM_STATE_MULT_TRIANGLE_COL_N_BUSY:
                 triangle_almost_done_flop <= triangle_almost_done_comb;
             //
             default:
@@ -283,8 +281,8 @@ module modexpng_mmm_dual
         //
         case (fsm_state)
             //
-            FSM_STATE_MULT_RECTANGLE_COL_0_BUSY,
-            FSM_STATE_MULT_RECTANGLE_COL_N_BUSY:
+            MMM_FSM_STATE_MULT_RECTANGLE_COL_0_BUSY,
+            MMM_FSM_STATE_MULT_RECTANGLE_COL_N_BUSY:
                 rectangle_almost_done_flop <= rectangle_almost_done_comb;
             //
             default:
@@ -301,36 +299,36 @@ module modexpng_mmm_dual
     //
     // Narrow Storage Control Logic
     //
-    always @(posedge clk)
+    always @(posedge clk or negedge rst_n)
         //
-        if (rst) narrow_xy_ena <= 1'b0;
+        if (!rst_n) narrow_xy_ena <= 1'b0;
         else begin
             //
             // Narrow Address
             //
             case (fsm_state_next)
                 //
-                FSM_STATE_MULT_SQUARE_COL_0_INIT,
-                FSM_STATE_MULT_SQUARE_COL_N_INIT:   narrow_xy_addr <= 8'd0;
-                FSM_STATE_MULT_SQUARE_COL_0_TRIG,
-                FSM_STATE_MULT_SQUARE_COL_N_TRIG,
-                FSM_STATE_MULT_SQUARE_COL_0_BUSY,
-                FSM_STATE_MULT_SQUARE_COL_N_BUSY:   narrow_xy_addr <= !square_almost_done_flop ? narrow_xy_addr + 1'b1 : 8'd0;
+                MMM_FSM_STATE_MULT_SQUARE_COL_0_INIT,
+                MMM_FSM_STATE_MULT_SQUARE_COL_N_INIT:   narrow_xy_addr <= 8'd0;
+                MMM_FSM_STATE_MULT_SQUARE_COL_0_TRIG,
+                MMM_FSM_STATE_MULT_SQUARE_COL_N_TRIG,
+                MMM_FSM_STATE_MULT_SQUARE_COL_0_BUSY,
+                MMM_FSM_STATE_MULT_SQUARE_COL_N_BUSY:   narrow_xy_addr <= !square_almost_done_flop ? narrow_xy_addr + 1'b1 : 8'd0;
                 //
-                FSM_STATE_MULT_TRIANGLE_COL_0_INIT,
-                FSM_STATE_MULT_TRIANGLE_COL_N_INIT: narrow_xy_addr <= 8'd0;
-                FSM_STATE_MULT_TRIANGLE_COL_0_TRIG,
-                FSM_STATE_MULT_TRIANGLE_COL_N_TRIG,
-                FSM_STATE_MULT_TRIANGLE_COL_0_BUSY,
-                FSM_STATE_MULT_TRIANGLE_COL_N_BUSY: narrow_xy_addr <= triangle_almost_done_flop || (col_is_last && triangle_surely_done_flop) ?
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_INIT,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_INIT: narrow_xy_addr <= 8'd0;
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_TRIG,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_TRIG,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_BUSY,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_BUSY: narrow_xy_addr <= triangle_almost_done_flop || (col_is_last && triangle_surely_done_flop) ?
                     8'd0 :  narrow_xy_addr + 1'b1;
                 //
-                FSM_STATE_MULT_RECTANGLE_COL_0_INIT,
-                FSM_STATE_MULT_RECTANGLE_COL_N_INIT: narrow_xy_addr <= 8'd0;
-                FSM_STATE_MULT_RECTANGLE_COL_0_TRIG,
-                FSM_STATE_MULT_RECTANGLE_COL_N_TRIG,
-                FSM_STATE_MULT_RECTANGLE_COL_0_BUSY,
-                FSM_STATE_MULT_RECTANGLE_COL_N_BUSY: narrow_xy_addr <= rectangle_almost_done_flop || rectangle_surely_done_flop ?
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_INIT,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_INIT: narrow_xy_addr <= 8'd0;
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_TRIG,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_TRIG,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_BUSY,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_BUSY: narrow_xy_addr <= rectangle_almost_done_flop || rectangle_surely_done_flop ?
                     8'd1 :  narrow_xy_addr + 1'b1;            
                 //
                 default:                            narrow_xy_addr <= 8'dX;
@@ -341,27 +339,27 @@ module modexpng_mmm_dual
             //
             case (fsm_state_next)
                 //
-                FSM_STATE_MULT_SQUARE_COL_0_INIT,
-                FSM_STATE_MULT_SQUARE_COL_N_INIT,
-                FSM_STATE_MULT_SQUARE_COL_0_TRIG,
-                FSM_STATE_MULT_SQUARE_COL_N_TRIG,
-                FSM_STATE_MULT_SQUARE_COL_0_BUSY,
-                FSM_STATE_MULT_SQUARE_COL_N_BUSY:   narrow_xy_bank <= sel_narrow_in;
+                MMM_FSM_STATE_MULT_SQUARE_COL_0_INIT,
+                MMM_FSM_STATE_MULT_SQUARE_COL_N_INIT,
+                MMM_FSM_STATE_MULT_SQUARE_COL_0_TRIG,
+                MMM_FSM_STATE_MULT_SQUARE_COL_N_TRIG,
+                MMM_FSM_STATE_MULT_SQUARE_COL_0_BUSY,
+                MMM_FSM_STATE_MULT_SQUARE_COL_N_BUSY:   narrow_xy_bank <= sel_narrow_in;
                 //
-                FSM_STATE_MULT_TRIANGLE_COL_0_INIT,
-                FSM_STATE_MULT_TRIANGLE_COL_N_INIT,
-                FSM_STATE_MULT_TRIANGLE_COL_0_TRIG,
-                FSM_STATE_MULT_TRIANGLE_COL_N_TRIG,
-                FSM_STATE_MULT_TRIANGLE_COL_0_BUSY,
-                FSM_STATE_MULT_TRIANGLE_COL_N_BUSY: narrow_xy_bank <= col_is_last && (triangle_almost_done_flop || triangle_surely_done_flop) ?
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_INIT,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_INIT,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_TRIG,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_TRIG,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_BUSY,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_BUSY: narrow_xy_bank <= col_is_last && (triangle_almost_done_flop || triangle_surely_done_flop) ?
                     BANK_NARROW_EXT : BANK_NARROW_COEFF;
                 //
-                FSM_STATE_MULT_RECTANGLE_COL_0_INIT,
-                FSM_STATE_MULT_RECTANGLE_COL_N_INIT,
-                FSM_STATE_MULT_RECTANGLE_COL_0_TRIG,
-                FSM_STATE_MULT_RECTANGLE_COL_N_TRIG,
-                FSM_STATE_MULT_RECTANGLE_COL_0_BUSY,
-                FSM_STATE_MULT_RECTANGLE_COL_N_BUSY: narrow_xy_bank <= rectangle_almost_done_flop || rectangle_surely_done_flop ?
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_INIT,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_INIT,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_TRIG,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_TRIG,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_BUSY,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_BUSY: narrow_xy_bank <= rectangle_almost_done_flop || rectangle_surely_done_flop ?
                     BANK_NARROW_EXT : BANK_NARROW_Q;            
                 //
                 default:                            narrow_xy_bank <= 2'bXX;
@@ -370,24 +368,24 @@ module modexpng_mmm_dual
             //
             case (fsm_state_next)
                 //
-                FSM_STATE_MULT_SQUARE_COL_0_INIT,
-                FSM_STATE_MULT_SQUARE_COL_N_INIT,
-                FSM_STATE_MULT_SQUARE_COL_0_TRIG,
-                FSM_STATE_MULT_SQUARE_COL_N_TRIG:   narrow_xy_ena <= 1'b1;
-                FSM_STATE_MULT_SQUARE_COL_0_BUSY,
-                FSM_STATE_MULT_SQUARE_COL_N_BUSY:   narrow_xy_ena <= ~square_almost_done_flop;
-                FSM_STATE_MULT_TRIANGLE_COL_0_INIT,
-                FSM_STATE_MULT_TRIANGLE_COL_N_INIT,
-                FSM_STATE_MULT_TRIANGLE_COL_0_TRIG,
-                FSM_STATE_MULT_TRIANGLE_COL_N_TRIG:   narrow_xy_ena <= 1'b1;
-                FSM_STATE_MULT_TRIANGLE_COL_0_BUSY,
-                FSM_STATE_MULT_TRIANGLE_COL_N_BUSY:   narrow_xy_ena <= !col_is_last ? ~triangle_almost_done_flop : ~triangle_surely_done_flop; 
-                FSM_STATE_MULT_RECTANGLE_COL_0_INIT,
-                FSM_STATE_MULT_RECTANGLE_COL_N_INIT,
-                FSM_STATE_MULT_RECTANGLE_COL_0_TRIG,
-                FSM_STATE_MULT_RECTANGLE_COL_N_TRIG:   narrow_xy_ena <= 1'b1;
-                FSM_STATE_MULT_RECTANGLE_COL_0_BUSY,
-                FSM_STATE_MULT_RECTANGLE_COL_N_BUSY:   narrow_xy_ena <= ~rectangle_surely_done_flop;
+                MMM_FSM_STATE_MULT_SQUARE_COL_0_INIT,
+                MMM_FSM_STATE_MULT_SQUARE_COL_N_INIT,
+                MMM_FSM_STATE_MULT_SQUARE_COL_0_TRIG,
+                MMM_FSM_STATE_MULT_SQUARE_COL_N_TRIG:   narrow_xy_ena <= 1'b1;
+                MMM_FSM_STATE_MULT_SQUARE_COL_0_BUSY,
+                MMM_FSM_STATE_MULT_SQUARE_COL_N_BUSY:   narrow_xy_ena <= ~square_almost_done_flop;
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_INIT,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_INIT,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_TRIG,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_TRIG:   narrow_xy_ena <= 1'b1;
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_BUSY,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_BUSY:   narrow_xy_ena <= !col_is_last ? ~triangle_almost_done_flop : ~triangle_surely_done_flop; 
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_INIT,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_INIT,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_TRIG,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_TRIG:   narrow_xy_ena <= 1'b1;
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_BUSY,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_BUSY:   narrow_xy_ena <= ~rectangle_surely_done_flop;
                 //
                 default:                              narrow_xy_ena <= 1'b0;
                 //
@@ -420,9 +418,9 @@ module modexpng_mmm_dual
     endfunction
     
     integer j;
-    always @(posedge clk)
+    always @(posedge clk or negedge rst_n)
         //
-        if (rst) begin
+        if (!rst_n) begin
             wide_xy_ena <= 1'b0;
             wide_xy_ena_aux <= 1'b0;
         end else begin
@@ -435,26 +433,26 @@ module modexpng_mmm_dual
                     //
                     // this can be reworked by having 8 address regs instead of 4 and using shifts instead of subtractions!
                     //
-                    FSM_STATE_MULT_SQUARE_COL_0_INIT:   wide_xy_addr[j] <= {5'd0, wide_offset_rom[j]};
-                    FSM_STATE_MULT_SQUARE_COL_N_INIT:   wide_xy_addr[j] <= {col_index_next, wide_offset_rom[j]};
-                    FSM_STATE_MULT_SQUARE_COL_0_TRIG,
-                    FSM_STATE_MULT_SQUARE_COL_N_TRIG,
-                    FSM_STATE_MULT_SQUARE_COL_0_BUSY,
-                    FSM_STATE_MULT_SQUARE_COL_N_BUSY:   wide_xy_addr[j] <= wide_xy_addr_next(wide_xy_addr[j], word_index_last);
+                    MMM_FSM_STATE_MULT_SQUARE_COL_0_INIT:   wide_xy_addr[j] <= {5'd0, wide_offset_rom[j]};
+                    MMM_FSM_STATE_MULT_SQUARE_COL_N_INIT:   wide_xy_addr[j] <= {col_index_next, wide_offset_rom[j]};
+                    MMM_FSM_STATE_MULT_SQUARE_COL_0_TRIG,
+                    MMM_FSM_STATE_MULT_SQUARE_COL_N_TRIG,
+                    MMM_FSM_STATE_MULT_SQUARE_COL_0_BUSY,
+                    MMM_FSM_STATE_MULT_SQUARE_COL_N_BUSY:   wide_xy_addr[j] <= wide_xy_addr_next(wide_xy_addr[j], word_index_last);
                     //
-                    FSM_STATE_MULT_TRIANGLE_COL_0_INIT:   wide_xy_addr[j] <= {5'd0, wide_offset_rom[j]};
-                    FSM_STATE_MULT_TRIANGLE_COL_N_INIT:   wide_xy_addr[j] <= {col_index_next, wide_offset_rom[j]};
-                    FSM_STATE_MULT_TRIANGLE_COL_0_TRIG,
-                    FSM_STATE_MULT_TRIANGLE_COL_N_TRIG,
-                    FSM_STATE_MULT_TRIANGLE_COL_0_BUSY,
-                    FSM_STATE_MULT_TRIANGLE_COL_N_BUSY:   wide_xy_addr[j] <= wide_xy_addr_next(wide_xy_addr[j], word_index_last);
+                    MMM_FSM_STATE_MULT_TRIANGLE_COL_0_INIT:   wide_xy_addr[j] <= {5'd0, wide_offset_rom[j]};
+                    MMM_FSM_STATE_MULT_TRIANGLE_COL_N_INIT:   wide_xy_addr[j] <= {col_index_next, wide_offset_rom[j]};
+                    MMM_FSM_STATE_MULT_TRIANGLE_COL_0_TRIG,
+                    MMM_FSM_STATE_MULT_TRIANGLE_COL_N_TRIG,
+                    MMM_FSM_STATE_MULT_TRIANGLE_COL_0_BUSY,
+                    MMM_FSM_STATE_MULT_TRIANGLE_COL_N_BUSY:   wide_xy_addr[j] <= wide_xy_addr_next(wide_xy_addr[j], word_index_last);
                     //
-                    FSM_STATE_MULT_RECTANGLE_COL_0_INIT:   wide_xy_addr[j] <= {5'd0, wide_offset_rom[j]};
-                    FSM_STATE_MULT_RECTANGLE_COL_N_INIT:   wide_xy_addr[j] <= {col_index_next, wide_offset_rom[j]};
-                    FSM_STATE_MULT_RECTANGLE_COL_0_TRIG,
-                    FSM_STATE_MULT_RECTANGLE_COL_N_TRIG,
-                    FSM_STATE_MULT_RECTANGLE_COL_0_BUSY,
-                    FSM_STATE_MULT_RECTANGLE_COL_N_BUSY:   wide_xy_addr[j] <= wide_xy_addr_next(wide_xy_addr[j], word_index_last);
+                    MMM_FSM_STATE_MULT_RECTANGLE_COL_0_INIT:   wide_xy_addr[j] <= {5'd0, wide_offset_rom[j]};
+                    MMM_FSM_STATE_MULT_RECTANGLE_COL_N_INIT:   wide_xy_addr[j] <= {col_index_next, wide_offset_rom[j]};
+                    MMM_FSM_STATE_MULT_RECTANGLE_COL_0_TRIG,
+                    MMM_FSM_STATE_MULT_RECTANGLE_COL_N_TRIG,
+                    MMM_FSM_STATE_MULT_RECTANGLE_COL_0_BUSY,
+                    MMM_FSM_STATE_MULT_RECTANGLE_COL_N_BUSY:   wide_xy_addr[j] <= wide_xy_addr_next(wide_xy_addr[j], word_index_last);
                     //
                     default:                            wide_xy_addr[j] <= 8'dX;
                 endcase
@@ -465,27 +463,27 @@ module modexpng_mmm_dual
                 //
                 // this can be reworked by having 8 address regs instead of 4 and using shifts instead of subtractions!
                 //
-                FSM_STATE_MULT_SQUARE_COL_0_INIT:   wide_xy_addr_aux <= {5'd0, 3'd1};
-                FSM_STATE_MULT_SQUARE_COL_N_INIT:   wide_xy_addr_aux <= {5'd0, 3'd1};
-                FSM_STATE_MULT_SQUARE_COL_0_TRIG,
-                FSM_STATE_MULT_SQUARE_COL_N_TRIG,
-                FSM_STATE_MULT_SQUARE_COL_0_BUSY,
-                FSM_STATE_MULT_SQUARE_COL_N_BUSY:   wide_xy_addr_aux <= wide_xy_addr_next(wide_xy_addr_aux, word_index_last);
+                MMM_FSM_STATE_MULT_SQUARE_COL_0_INIT:   wide_xy_addr_aux <= {5'd0, 3'd1};
+                MMM_FSM_STATE_MULT_SQUARE_COL_N_INIT:   wide_xy_addr_aux <= {5'd0, 3'd1};
+                MMM_FSM_STATE_MULT_SQUARE_COL_0_TRIG,
+                MMM_FSM_STATE_MULT_SQUARE_COL_N_TRIG,
+                MMM_FSM_STATE_MULT_SQUARE_COL_0_BUSY,
+                MMM_FSM_STATE_MULT_SQUARE_COL_N_BUSY:   wide_xy_addr_aux <= wide_xy_addr_next(wide_xy_addr_aux, word_index_last);
                 //
-                FSM_STATE_MULT_TRIANGLE_COL_0_INIT:   wide_xy_addr_aux <= {5'd0, 3'd1};
-                FSM_STATE_MULT_TRIANGLE_COL_N_INIT:   wide_xy_addr_aux <= {5'd0, 3'd1};
-                FSM_STATE_MULT_TRIANGLE_COL_0_TRIG,
-                FSM_STATE_MULT_TRIANGLE_COL_N_TRIG,
-                FSM_STATE_MULT_TRIANGLE_COL_0_BUSY,
-                FSM_STATE_MULT_TRIANGLE_COL_N_BUSY:   wide_xy_addr_aux <= wide_xy_addr_next(wide_xy_addr_aux, word_index_last);
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_INIT:   wide_xy_addr_aux <= {5'd0, 3'd1};
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_INIT:   wide_xy_addr_aux <= {5'd0, 3'd1};
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_TRIG,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_TRIG,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_BUSY,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_BUSY:   wide_xy_addr_aux <= wide_xy_addr_next(wide_xy_addr_aux, word_index_last);
                 //
-                FSM_STATE_MULT_RECTANGLE_COL_0_INIT:   wide_xy_addr_aux <= 8'dX;//{5'd0, 3'd0};
-                FSM_STATE_MULT_RECTANGLE_COL_N_INIT,
-                FSM_STATE_MULT_RECTANGLE_COL_0_TRIG,
-                FSM_STATE_MULT_RECTANGLE_COL_N_TRIG,
-                FSM_STATE_MULT_RECTANGLE_COL_0_BUSY,
-                FSM_STATE_MULT_RECTANGLE_COL_N_BUSY,
-                FSM_STATE_MULT_RECTANGLE_HOLDOFF:    wide_xy_addr_aux <= rcmb_xy_valid ? rcmb_xy_addr : 8'dX;
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_INIT:   wide_xy_addr_aux <= 8'dX;//{5'd0, 3'd0};
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_INIT,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_TRIG,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_TRIG,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_BUSY,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_BUSY,
+                MMM_FSM_STATE_MULT_RECTANGLE_HOLDOFF:    wide_xy_addr_aux <= rcmb_xy_valid ? rcmb_xy_addr : 8'dX;
                 //recomb_fat_bram_xy_dout_valid && (recomb_fat_bram_xy_bank == BANK_FAT_ML) ?
                     //mac_fat_bram_xy_addr[4] + 1'b1 : mac_fat_bram_xy_addr[4];
                 //
@@ -495,49 +493,49 @@ module modexpng_mmm_dual
             // Wide Bank
             //
             case (fsm_state_next)
-                FSM_STATE_MULT_SQUARE_COL_0_INIT,
-                FSM_STATE_MULT_SQUARE_COL_N_INIT,
-                FSM_STATE_MULT_SQUARE_COL_0_TRIG,
-                FSM_STATE_MULT_SQUARE_COL_N_TRIG,
-                FSM_STATE_MULT_SQUARE_COL_0_BUSY,
-                FSM_STATE_MULT_SQUARE_COL_N_BUSY:    wide_xy_bank <= sel_wide_in;
-                FSM_STATE_MULT_TRIANGLE_COL_0_INIT,
-                FSM_STATE_MULT_TRIANGLE_COL_N_INIT,
-                FSM_STATE_MULT_TRIANGLE_COL_0_TRIG,
-                FSM_STATE_MULT_TRIANGLE_COL_N_TRIG:  wide_xy_bank <= BANK_WIDE_L;
-                FSM_STATE_MULT_TRIANGLE_COL_0_BUSY,
-                FSM_STATE_MULT_TRIANGLE_COL_N_BUSY:   wide_xy_bank <= BANK_WIDE_L;
-                FSM_STATE_MULT_RECTANGLE_COL_0_INIT,
-                FSM_STATE_MULT_RECTANGLE_COL_N_INIT,
-                FSM_STATE_MULT_RECTANGLE_COL_0_TRIG,
-                FSM_STATE_MULT_RECTANGLE_COL_N_TRIG,    
-                FSM_STATE_MULT_RECTANGLE_COL_0_BUSY,
-                FSM_STATE_MULT_RECTANGLE_COL_N_BUSY:   wide_xy_bank <= BANK_WIDE_N;            
+                MMM_FSM_STATE_MULT_SQUARE_COL_0_INIT,
+                MMM_FSM_STATE_MULT_SQUARE_COL_N_INIT,
+                MMM_FSM_STATE_MULT_SQUARE_COL_0_TRIG,
+                MMM_FSM_STATE_MULT_SQUARE_COL_N_TRIG,
+                MMM_FSM_STATE_MULT_SQUARE_COL_0_BUSY,
+                MMM_FSM_STATE_MULT_SQUARE_COL_N_BUSY:    wide_xy_bank <= sel_wide_in;
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_INIT,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_INIT,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_TRIG,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_TRIG:  wide_xy_bank <= BANK_WIDE_L;
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_BUSY,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_BUSY:   wide_xy_bank <= BANK_WIDE_L;
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_INIT,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_INIT,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_TRIG,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_TRIG,    
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_BUSY,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_BUSY:   wide_xy_bank <= BANK_WIDE_N;            
                 default:                             wide_xy_bank <= 3'bXXX;
             endcase
             //
             // Wide Aux Bank
             //
             case (fsm_state_next)
-                FSM_STATE_MULT_SQUARE_COL_0_INIT,
-                FSM_STATE_MULT_SQUARE_COL_N_INIT,
-                FSM_STATE_MULT_SQUARE_COL_0_TRIG,
-                FSM_STATE_MULT_SQUARE_COL_N_TRIG,
-                FSM_STATE_MULT_SQUARE_COL_0_BUSY,
-                FSM_STATE_MULT_SQUARE_COL_N_BUSY:   wide_xy_bank_aux <= sel_wide_in;
-                FSM_STATE_MULT_TRIANGLE_COL_0_INIT,
-                FSM_STATE_MULT_TRIANGLE_COL_N_INIT,
-                FSM_STATE_MULT_TRIANGLE_COL_0_TRIG,
-                FSM_STATE_MULT_TRIANGLE_COL_N_TRIG: wide_xy_bank_aux <= BANK_WIDE_H;
-                FSM_STATE_MULT_TRIANGLE_COL_0_BUSY,
-                FSM_STATE_MULT_TRIANGLE_COL_N_BUSY:   wide_xy_bank_aux <= BANK_WIDE_L;
-                FSM_STATE_MULT_RECTANGLE_COL_0_INIT,
-                FSM_STATE_MULT_RECTANGLE_COL_N_INIT,
-                FSM_STATE_MULT_RECTANGLE_COL_0_TRIG,
-                FSM_STATE_MULT_RECTANGLE_COL_N_TRIG,    
-                FSM_STATE_MULT_RECTANGLE_COL_0_BUSY,
-                FSM_STATE_MULT_RECTANGLE_COL_N_BUSY,
-                FSM_STATE_MULT_RECTANGLE_HOLDOFF:   if (rcmb_xy_valid) // rewrite using "Kolya-style" here (get rid of too many xxx's)
+                MMM_FSM_STATE_MULT_SQUARE_COL_0_INIT,
+                MMM_FSM_STATE_MULT_SQUARE_COL_N_INIT,
+                MMM_FSM_STATE_MULT_SQUARE_COL_0_TRIG,
+                MMM_FSM_STATE_MULT_SQUARE_COL_N_TRIG,
+                MMM_FSM_STATE_MULT_SQUARE_COL_0_BUSY,
+                MMM_FSM_STATE_MULT_SQUARE_COL_N_BUSY:   wide_xy_bank_aux <= sel_wide_in;
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_INIT,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_INIT,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_TRIG,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_TRIG: wide_xy_bank_aux <= BANK_WIDE_H;
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_BUSY,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_BUSY:   wide_xy_bank_aux <= BANK_WIDE_L;
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_INIT,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_INIT,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_TRIG,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_TRIG,    
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_BUSY,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_BUSY,
+                MMM_FSM_STATE_MULT_RECTANGLE_HOLDOFF:   if (rcmb_xy_valid) // rewrite using "Kolya-style" here (get rid of too many xxx's)
                     case (rcmb_xy_bank)
                         BANK_RCMB_ML: wide_xy_bank_aux <= BANK_WIDE_L;
                         BANK_RCMB_MH: wide_xy_bank_aux <= BANK_WIDE_H;
@@ -551,43 +549,43 @@ module modexpng_mmm_dual
             // Wide Enable
             //
             case (fsm_state_next)
-                FSM_STATE_MULT_SQUARE_COL_0_INIT,
-                FSM_STATE_MULT_SQUARE_COL_N_INIT,
-                FSM_STATE_MULT_SQUARE_COL_0_TRIG,
-                FSM_STATE_MULT_SQUARE_COL_N_TRIG,
-                FSM_STATE_MULT_SQUARE_COL_0_BUSY,
-                FSM_STATE_MULT_SQUARE_COL_N_BUSY,
-                FSM_STATE_MULT_TRIANGLE_COL_0_INIT,
-                FSM_STATE_MULT_TRIANGLE_COL_N_INIT,
-                FSM_STATE_MULT_TRIANGLE_COL_0_TRIG,
-                FSM_STATE_MULT_TRIANGLE_COL_N_TRIG,
-                FSM_STATE_MULT_TRIANGLE_COL_0_BUSY,
-                FSM_STATE_MULT_TRIANGLE_COL_N_BUSY,
-                FSM_STATE_MULT_RECTANGLE_COL_0_INIT,
-                FSM_STATE_MULT_RECTANGLE_COL_N_INIT,
-                FSM_STATE_MULT_RECTANGLE_COL_0_TRIG,
-                FSM_STATE_MULT_RECTANGLE_COL_N_TRIG,
-                FSM_STATE_MULT_RECTANGLE_COL_0_BUSY,
-                FSM_STATE_MULT_RECTANGLE_COL_N_BUSY:   wide_xy_ena <= 1'b1;
+                MMM_FSM_STATE_MULT_SQUARE_COL_0_INIT,
+                MMM_FSM_STATE_MULT_SQUARE_COL_N_INIT,
+                MMM_FSM_STATE_MULT_SQUARE_COL_0_TRIG,
+                MMM_FSM_STATE_MULT_SQUARE_COL_N_TRIG,
+                MMM_FSM_STATE_MULT_SQUARE_COL_0_BUSY,
+                MMM_FSM_STATE_MULT_SQUARE_COL_N_BUSY,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_INIT,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_INIT,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_TRIG,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_TRIG,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_BUSY,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_BUSY,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_INIT,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_INIT,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_TRIG,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_TRIG,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_BUSY,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_BUSY:   wide_xy_ena <= 1'b1;
                 default:                               wide_xy_ena <= 1'b0;
             endcase
             //
             // Wide Aux Enable
             //
             case (fsm_state_next)
-                FSM_STATE_MULT_TRIANGLE_COL_0_INIT,
-                FSM_STATE_MULT_TRIANGLE_COL_N_INIT,
-                FSM_STATE_MULT_TRIANGLE_COL_0_TRIG,
-                FSM_STATE_MULT_TRIANGLE_COL_N_TRIG,
-                FSM_STATE_MULT_TRIANGLE_COL_0_BUSY,
-                FSM_STATE_MULT_TRIANGLE_COL_N_BUSY:  wide_xy_ena_aux <= 1'b1;
-                FSM_STATE_MULT_RECTANGLE_COL_0_INIT: wide_xy_ena_aux <= 1'b0;//1'b1;
-                FSM_STATE_MULT_RECTANGLE_COL_N_INIT,
-                FSM_STATE_MULT_RECTANGLE_COL_0_TRIG,
-                FSM_STATE_MULT_RECTANGLE_COL_N_TRIG,
-                FSM_STATE_MULT_RECTANGLE_COL_0_BUSY,
-                FSM_STATE_MULT_RECTANGLE_COL_N_BUSY,
-                FSM_STATE_MULT_RECTANGLE_HOLDOFF:   wide_xy_ena_aux <= rcmb_xy_valid;// && (recomb_fat_bram_xy_bank == BANK_FAT_ML);
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_INIT,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_INIT,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_TRIG,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_TRIG,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_0_BUSY,
+                MMM_FSM_STATE_MULT_TRIANGLE_COL_N_BUSY:  wide_xy_ena_aux <= 1'b1;
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_INIT: wide_xy_ena_aux <= 1'b0;//1'b1;
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_INIT,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_TRIG,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_TRIG,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_0_BUSY,
+                MMM_FSM_STATE_MULT_RECTANGLE_COL_N_BUSY,
+                MMM_FSM_STATE_MULT_RECTANGLE_HOLDOFF:   wide_xy_ena_aux <= rcmb_xy_valid;// && (recomb_fat_bram_xy_bank == BANK_FAT_ML);
                 default:                            wide_xy_ena_aux <= 1'b0;
             endcase
             //
@@ -686,9 +684,9 @@ module modexpng_mmm_dual
     reg narrow_xy_ena_dly1 = 1'b0;
     reg narrow_xy_ena_dly2 = 1'b0;
     
-    always @(posedge clk)
+    always @(posedge clk or negedge rst_n)
         //
-        if (rst) begin
+        if (!rst_n) begin
             //
             narrow_xy_ena_dly1 <= 1'b0;
             narrow_xy_ena_dly2 <= 1'b0;
@@ -721,13 +719,13 @@ module modexpng_mmm_dual
     always @(posedge clk) begin
         //
         case (fsm_state)
-            FSM_STATE_MULT_SQUARE_COL_0_TRIG:   dsp_merge_xy_b <= 1'b1;
-            FSM_STATE_MULT_TRIANGLE_COL_0_TRIG: dsp_merge_xy_b <= 1'b0;
+            MMM_FSM_STATE_MULT_SQUARE_COL_0_TRIG:   dsp_merge_xy_b <= 1'b1;
+            MMM_FSM_STATE_MULT_TRIANGLE_COL_0_TRIG: dsp_merge_xy_b <= 1'b0;
         endcase
         //
         case (fsm_state)
-            FSM_STATE_MULT_SQUARE_COL_0_TRIG,
-            FSM_STATE_MULT_SQUARE_COL_N_TRIG: dsp_merge_xy_b_first <= 1'b1;
+            MMM_FSM_STATE_MULT_SQUARE_COL_0_TRIG,
+            MMM_FSM_STATE_MULT_SQUARE_COL_N_TRIG: dsp_merge_xy_b_first <= 1'b1;
             default:                          dsp_merge_xy_b_first <= 1'b0;
         endcase
         //
@@ -815,18 +813,18 @@ module modexpng_mmm_dual
     always @(posedge clk)
         //
         case (fsm_state_next)
-            FSM_STATE_MULT_SQUARE_COL_0_TRIG,
-            FSM_STATE_MULT_SQUARE_COL_N_TRIG:       dsp_xy_mode_z_adv4 <= {9{1'b0}};
-            FSM_STATE_MULT_SQUARE_COL_0_BUSY,
-            FSM_STATE_MULT_SQUARE_COL_N_BUSY:       dsp_xy_mode_z_adv4 <= calc_mac_mode_z_square(col_index_prev, narrow_xy_addr_dly);
-            FSM_STATE_MULT_TRIANGLE_COL_0_TRIG,
-            FSM_STATE_MULT_TRIANGLE_COL_N_TRIG:     dsp_xy_mode_z_adv4 <= {9{1'b0}};    // so easy
-            FSM_STATE_MULT_TRIANGLE_COL_0_BUSY,
-            FSM_STATE_MULT_TRIANGLE_COL_N_BUSY:     dsp_xy_mode_z_adv4 <= {9{1'b1}};
-            FSM_STATE_MULT_RECTANGLE_COL_0_TRIG,
-            FSM_STATE_MULT_RECTANGLE_COL_N_TRIG:     dsp_xy_mode_z_adv4 <= {9{1'b0}};    // so easy
-            FSM_STATE_MULT_RECTANGLE_COL_0_BUSY,
-            FSM_STATE_MULT_RECTANGLE_COL_N_BUSY:     dsp_xy_mode_z_adv4 <= calc_mac_mode_z_rectangle(col_index_prev, narrow_xy_addr_dly);
+            MMM_FSM_STATE_MULT_SQUARE_COL_0_TRIG,
+            MMM_FSM_STATE_MULT_SQUARE_COL_N_TRIG:       dsp_xy_mode_z_adv4 <= {9{1'b0}};
+            MMM_FSM_STATE_MULT_SQUARE_COL_0_BUSY,
+            MMM_FSM_STATE_MULT_SQUARE_COL_N_BUSY:       dsp_xy_mode_z_adv4 <= calc_mac_mode_z_square(col_index_prev, narrow_xy_addr_dly);
+            MMM_FSM_STATE_MULT_TRIANGLE_COL_0_TRIG,
+            MMM_FSM_STATE_MULT_TRIANGLE_COL_N_TRIG:     dsp_xy_mode_z_adv4 <= {9{1'b0}};    // so easy
+            MMM_FSM_STATE_MULT_TRIANGLE_COL_0_BUSY,
+            MMM_FSM_STATE_MULT_TRIANGLE_COL_N_BUSY:     dsp_xy_mode_z_adv4 <= {9{1'b1}};
+            MMM_FSM_STATE_MULT_RECTANGLE_COL_0_TRIG,
+            MMM_FSM_STATE_MULT_RECTANGLE_COL_N_TRIG:     dsp_xy_mode_z_adv4 <= {9{1'b0}};    // so easy
+            MMM_FSM_STATE_MULT_RECTANGLE_COL_0_BUSY,
+            MMM_FSM_STATE_MULT_RECTANGLE_COL_N_BUSY:     dsp_xy_mode_z_adv4 <= calc_mac_mode_z_rectangle(col_index_prev, narrow_xy_addr_dly);
             default:                                dsp_xy_mode_z_adv4 <= {9{1'b1}};
         endcase
 
@@ -851,7 +849,7 @@ module modexpng_mmm_dual
     modexpng_recombinator_block recombinator_block
     (
         .clk                            (clk),
-        .rst                            (rst),
+        .rst_n                          (rst_n),
         
         .ena                            (rcmb_ena),
         .rdy                            (rcmb_rdy),
@@ -894,10 +892,10 @@ module modexpng_mmm_dual
     //
     // Recombinator Enable Logic
     //    
-    always @(posedge clk)
+    always @(posedge clk or negedge rst_n)
         //
-        if (rst) rcmb_ena <= 1'b0;
-        else     rcmb_ena <= dsp_xy_ce_a && !dsp_xy_ce_b && !dsp_xy_ce_m && !dsp_xy_ce_p;
+        if (!rst_n) rcmb_ena <= 1'b0;
+        else        rcmb_ena <= dsp_xy_ce_a && !dsp_xy_ce_b && !dsp_xy_ce_m && !dsp_xy_ce_p;
 
         
     //
@@ -911,55 +909,55 @@ module modexpng_mmm_dual
     //
     // FSM Transition Logic
     //
-    assign fsm_state_after_idle = !only_reduce ? FSM_STATE_MULT_SQUARE_COL_0_INIT : FSM_STATE_MULT_TRIANGLE_COL_0_INIT;
-    assign fsm_state_after_mult_square    = col_is_last ? FSM_STATE_MULT_SQUARE_HOLDOFF   : FSM_STATE_MULT_SQUARE_COL_N_INIT;
-    assign fsm_state_after_mult_triangle  = col_is_last ? FSM_STATE_MULT_TRIANGLE_HOLDOFF : FSM_STATE_MULT_TRIANGLE_COL_N_INIT;
-    assign fsm_state_after_mult_rectangle = col_is_last ? FSM_STATE_MULT_RECTANGLE_HOLDOFF : FSM_STATE_MULT_RECTANGLE_COL_N_INIT;
-    assign fsm_state_after_square_holdoff = just_multiply ? FSM_STATE_STOP : FSM_STATE_MULT_TRIANGLE_COL_0_INIT;
+    assign fsm_state_after_idle = !only_reduce ? MMM_FSM_STATE_MULT_SQUARE_COL_0_INIT : MMM_FSM_STATE_MULT_TRIANGLE_COL_0_INIT;
+    assign fsm_state_after_mult_square    = col_is_last ? MMM_FSM_STATE_MULT_SQUARE_HOLDOFF   : MMM_FSM_STATE_MULT_SQUARE_COL_N_INIT;
+    assign fsm_state_after_mult_triangle  = col_is_last ? MMM_FSM_STATE_MULT_TRIANGLE_HOLDOFF : MMM_FSM_STATE_MULT_TRIANGLE_COL_N_INIT;
+    assign fsm_state_after_mult_rectangle = col_is_last ? MMM_FSM_STATE_MULT_RECTANGLE_HOLDOFF : MMM_FSM_STATE_MULT_RECTANGLE_COL_N_INIT;
+    assign fsm_state_after_square_holdoff = just_multiply ? MMM_FSM_STATE_STOP : MMM_FSM_STATE_MULT_TRIANGLE_COL_0_INIT;
      
 
     always @* begin
         //
-        fsm_state_next = FSM_STATE_IDLE;
+        fsm_state_next = MMM_FSM_STATE_IDLE;
         //
         case (fsm_state)
-            FSM_STATE_IDLE:                   fsm_state_next = ena                   ? fsm_state_after_idle /*FSM_STATE_MULT_SQUARE_COL_0_INIT*/ : FSM_STATE_IDLE;
+            MMM_FSM_STATE_IDLE:                   fsm_state_next = ena                   ? fsm_state_after_idle /*MMM_FSM_STATE_MULT_SQUARE_COL_0_INIT*/ : MMM_FSM_STATE_IDLE;
                         
-            FSM_STATE_MULT_SQUARE_COL_0_INIT: fsm_state_next =                         FSM_STATE_MULT_SQUARE_COL_0_TRIG ;
-            FSM_STATE_MULT_SQUARE_COL_0_TRIG: fsm_state_next =                         FSM_STATE_MULT_SQUARE_COL_0_BUSY ;
-            FSM_STATE_MULT_SQUARE_COL_0_BUSY: fsm_state_next = square_done ? FSM_STATE_MULT_SQUARE_COL_N_INIT : FSM_STATE_MULT_SQUARE_COL_0_BUSY;
+            MMM_FSM_STATE_MULT_SQUARE_COL_0_INIT: fsm_state_next =                         MMM_FSM_STATE_MULT_SQUARE_COL_0_TRIG ;
+            MMM_FSM_STATE_MULT_SQUARE_COL_0_TRIG: fsm_state_next =                         MMM_FSM_STATE_MULT_SQUARE_COL_0_BUSY ;
+            MMM_FSM_STATE_MULT_SQUARE_COL_0_BUSY: fsm_state_next = square_done ? MMM_FSM_STATE_MULT_SQUARE_COL_N_INIT : MMM_FSM_STATE_MULT_SQUARE_COL_0_BUSY;
             
-            FSM_STATE_MULT_SQUARE_COL_N_INIT: fsm_state_next =                         FSM_STATE_MULT_SQUARE_COL_N_TRIG ;
-            FSM_STATE_MULT_SQUARE_COL_N_TRIG: fsm_state_next =                         FSM_STATE_MULT_SQUARE_COL_N_BUSY ;
-            FSM_STATE_MULT_SQUARE_COL_N_BUSY: fsm_state_next = square_done ? fsm_state_after_mult_square    : FSM_STATE_MULT_SQUARE_COL_N_BUSY;
+            MMM_FSM_STATE_MULT_SQUARE_COL_N_INIT: fsm_state_next =                         MMM_FSM_STATE_MULT_SQUARE_COL_N_TRIG ;
+            MMM_FSM_STATE_MULT_SQUARE_COL_N_TRIG: fsm_state_next =                         MMM_FSM_STATE_MULT_SQUARE_COL_N_BUSY ;
+            MMM_FSM_STATE_MULT_SQUARE_COL_N_BUSY: fsm_state_next = square_done ? fsm_state_after_mult_square    : MMM_FSM_STATE_MULT_SQUARE_COL_N_BUSY;
             
-            FSM_STATE_MULT_SQUARE_HOLDOFF:    fsm_state_next =                         rcmb_rdy ? fsm_state_after_square_holdoff : FSM_STATE_MULT_SQUARE_HOLDOFF;
+            MMM_FSM_STATE_MULT_SQUARE_HOLDOFF:    fsm_state_next =                         rcmb_rdy ? fsm_state_after_square_holdoff : MMM_FSM_STATE_MULT_SQUARE_HOLDOFF;
 
-            FSM_STATE_MULT_TRIANGLE_COL_0_INIT: fsm_state_next =                         FSM_STATE_MULT_TRIANGLE_COL_0_TRIG ;
-            FSM_STATE_MULT_TRIANGLE_COL_0_TRIG: fsm_state_next =                         FSM_STATE_MULT_TRIANGLE_COL_0_BUSY ;
-            FSM_STATE_MULT_TRIANGLE_COL_0_BUSY: fsm_state_next = triangle_done ? FSM_STATE_MULT_TRIANGLE_COL_N_INIT : FSM_STATE_MULT_TRIANGLE_COL_0_BUSY;     
+            MMM_FSM_STATE_MULT_TRIANGLE_COL_0_INIT: fsm_state_next =                         MMM_FSM_STATE_MULT_TRIANGLE_COL_0_TRIG ;
+            MMM_FSM_STATE_MULT_TRIANGLE_COL_0_TRIG: fsm_state_next =                         MMM_FSM_STATE_MULT_TRIANGLE_COL_0_BUSY ;
+            MMM_FSM_STATE_MULT_TRIANGLE_COL_0_BUSY: fsm_state_next = triangle_done ? MMM_FSM_STATE_MULT_TRIANGLE_COL_N_INIT : MMM_FSM_STATE_MULT_TRIANGLE_COL_0_BUSY;     
             
-            FSM_STATE_MULT_TRIANGLE_COL_N_INIT: fsm_state_next =                         FSM_STATE_MULT_TRIANGLE_COL_N_TRIG ;
-            FSM_STATE_MULT_TRIANGLE_COL_N_TRIG: fsm_state_next =                         FSM_STATE_MULT_TRIANGLE_COL_N_BUSY ;
-            FSM_STATE_MULT_TRIANGLE_COL_N_BUSY: fsm_state_next = triangle_done ? fsm_state_after_mult_triangle : FSM_STATE_MULT_TRIANGLE_COL_N_BUSY;
+            MMM_FSM_STATE_MULT_TRIANGLE_COL_N_INIT: fsm_state_next =                         MMM_FSM_STATE_MULT_TRIANGLE_COL_N_TRIG ;
+            MMM_FSM_STATE_MULT_TRIANGLE_COL_N_TRIG: fsm_state_next =                         MMM_FSM_STATE_MULT_TRIANGLE_COL_N_BUSY ;
+            MMM_FSM_STATE_MULT_TRIANGLE_COL_N_BUSY: fsm_state_next = triangle_done ? fsm_state_after_mult_triangle : MMM_FSM_STATE_MULT_TRIANGLE_COL_N_BUSY;
             
-            FSM_STATE_MULT_TRIANGLE_HOLDOFF:    fsm_state_next =                         rcmb_rdy ? FSM_STATE_MULT_RECTANGLE_COL_0_INIT : FSM_STATE_MULT_TRIANGLE_HOLDOFF;
+            MMM_FSM_STATE_MULT_TRIANGLE_HOLDOFF:    fsm_state_next =                         rcmb_rdy ? MMM_FSM_STATE_MULT_RECTANGLE_COL_0_INIT : MMM_FSM_STATE_MULT_TRIANGLE_HOLDOFF;
 
-            FSM_STATE_MULT_RECTANGLE_COL_0_INIT: fsm_state_next =                         FSM_STATE_MULT_RECTANGLE_COL_0_TRIG ;
-            FSM_STATE_MULT_RECTANGLE_COL_0_TRIG: fsm_state_next =                         FSM_STATE_MULT_RECTANGLE_COL_0_BUSY ;
-            FSM_STATE_MULT_RECTANGLE_COL_0_BUSY: fsm_state_next = rectangle_done ? FSM_STATE_MULT_RECTANGLE_COL_N_INIT : FSM_STATE_MULT_RECTANGLE_COL_0_BUSY;     
+            MMM_FSM_STATE_MULT_RECTANGLE_COL_0_INIT: fsm_state_next =                         MMM_FSM_STATE_MULT_RECTANGLE_COL_0_TRIG ;
+            MMM_FSM_STATE_MULT_RECTANGLE_COL_0_TRIG: fsm_state_next =                         MMM_FSM_STATE_MULT_RECTANGLE_COL_0_BUSY ;
+            MMM_FSM_STATE_MULT_RECTANGLE_COL_0_BUSY: fsm_state_next = rectangle_done ? MMM_FSM_STATE_MULT_RECTANGLE_COL_N_INIT : MMM_FSM_STATE_MULT_RECTANGLE_COL_0_BUSY;     
             
-            FSM_STATE_MULT_RECTANGLE_COL_N_INIT: fsm_state_next =                         FSM_STATE_MULT_RECTANGLE_COL_N_TRIG ;
-            FSM_STATE_MULT_RECTANGLE_COL_N_TRIG: fsm_state_next =                         FSM_STATE_MULT_RECTANGLE_COL_N_BUSY ;
-            FSM_STATE_MULT_RECTANGLE_COL_N_BUSY: fsm_state_next = rectangle_done ? fsm_state_after_mult_rectangle : FSM_STATE_MULT_RECTANGLE_COL_N_BUSY;
+            MMM_FSM_STATE_MULT_RECTANGLE_COL_N_INIT: fsm_state_next =                         MMM_FSM_STATE_MULT_RECTANGLE_COL_N_TRIG ;
+            MMM_FSM_STATE_MULT_RECTANGLE_COL_N_TRIG: fsm_state_next =                         MMM_FSM_STATE_MULT_RECTANGLE_COL_N_BUSY ;
+            MMM_FSM_STATE_MULT_RECTANGLE_COL_N_BUSY: fsm_state_next = rectangle_done ? fsm_state_after_mult_rectangle : MMM_FSM_STATE_MULT_RECTANGLE_COL_N_BUSY;
             
-            FSM_STATE_MULT_RECTANGLE_HOLDOFF:    fsm_state_next =                         rcmb_rdy ? FSM_STATE_WAIT_REDUCTOR : FSM_STATE_MULT_RECTANGLE_HOLDOFF;
+            MMM_FSM_STATE_MULT_RECTANGLE_HOLDOFF:    fsm_state_next =                         rcmb_rdy ? MMM_FSM_STATE_WAIT_REDUCTOR : MMM_FSM_STATE_MULT_RECTANGLE_HOLDOFF;
             
-            FSM_STATE_WAIT_REDUCTOR:             fsm_state_next =                         rdct_rdy ? FSM_STATE_STOP : FSM_STATE_WAIT_REDUCTOR;
+            MMM_FSM_STATE_WAIT_REDUCTOR:             fsm_state_next =                         rdct_rdy ? MMM_FSM_STATE_STOP : MMM_FSM_STATE_WAIT_REDUCTOR;
             
-            FSM_STATE_STOP:                     fsm_state_next =                         FSM_STATE_IDLE                   ;
+            MMM_FSM_STATE_STOP:                     fsm_state_next =                         MMM_FSM_STATE_IDLE                   ;
             
-            default:                             fsm_state_next =                         FSM_STATE_IDLE                   ;
+            default:                             fsm_state_next =                         MMM_FSM_STATE_IDLE                   ;
 
         endcase
         //
@@ -973,11 +971,11 @@ module modexpng_mmm_dual
 
     assign rdct_ena = rdct_ena_reg; 
     
-    always @(posedge clk)
+    always @(posedge clk or negedge rst_n)
         //
-        if (rst)                                rdct_ena_reg <= 1'b0;
+        if (!rst_n)                                rdct_ena_reg <= 1'b0;
         else case (fsm_state)
-           FSM_STATE_MULT_RECTANGLE_COL_0_INIT: rdct_ena_reg <= 1'b1;
+           MMM_FSM_STATE_MULT_RECTANGLE_COL_0_INIT: rdct_ena_reg <= 1'b1;
            default:                             rdct_ena_reg <= 1'b0;
         endcase
     
@@ -989,12 +987,12 @@ module modexpng_mmm_dual
     
     assign rdy = rdy_reg;
     
-    always @(posedge clk)
+    always @(posedge clk or negedge rst_n)
         //
-        if (rst) rdy_reg <= 1'b1;
+        if (!rst_n) rdy_reg <= 1'b1;
         else begin
             if (rdy && ena) rdy_reg <= 1'b0;
-            if (!rdy && (fsm_state == FSM_STATE_STOP)) rdy_reg <= 1'b1; 
+            if (!rdy && (fsm_state == MMM_FSM_STATE_STOP)) rdy_reg <= 1'b1; 
         end
     
 endmodule