Reworked testbench, clk_sys and clk_core can now have any ratio, not

necessarily 1:2.

Fixed compile-time issue where ISE fails to place two DSP slices next to each
other, if A and/or B cascade path(s) between then are partially connected.
Basically, if cascade is used, entire bus must be connected.

7 files changed, 381 insertions, 222 deletions

diff --git a/bench/tb_core_full_1024.v b/bench/tb_core_full_1024.v
index e6b1a66..87eac79 100644
--- a/bench/tb_core_full_1024.v
+++ b/bench/tb_core_full_1024.v
@@ -31,6 +31,7 @@ module tb_core_full_1024;
     reg [31:0] Q_FACTOR[0:TB_NUM_WORDS_PQ-1];
     reg [31:0] P_COEFF[0:TB_NUM_WORDS_PQ];
     reg [31:0] Q_COEFF[0:TB_NUM_WORDS_PQ];
+    reg [31:0] D[0:TB_NUM_WORDS_N-1];
     reg [31:0] DP[0:TB_NUM_WORDS_PQ-1];
     reg [31:0] DQ[0:TB_NUM_WORDS_PQ-1];
     reg [31:0] QINV[0:TB_NUM_WORDS_PQ-1];
@@ -117,6 +118,14 @@ module tb_core_full_1024;
         Q_COEFF[  8] = 32'h0cf1175c; Q_COEFF[  9] = 32'h4911b74e; Q_COEFF[ 10] = 32'h331e61cb; Q_COEFF[ 11] = 32'he9527ead;
         Q_COEFF[ 12] = 32'h8d6a5911; Q_COEFF[ 13] = 32'hae42d654; Q_COEFF[ 14] = 32'he10d29a8; Q_COEFF[ 15] = 32'h50a5dd76;
         Q_COEFF[ 16] = 32'h0000ed75;
+        D[  0] = 32'h6d4e7831; D[  1] = 32'ha96f72d6; D[  2] = 32'h3b47788f; D[  3] = 32'h07207984;
+        D[  4] = 32'h3247b99f; D[  5] = 32'hfe6ade84; D[  6] = 32'h0e0fdec4; D[  7] = 32'hc31450bd;
+        D[  8] = 32'hb8624e9a; D[  9] = 32'h362966c6; D[ 10] = 32'h4d8a15bf; D[ 11] = 32'h845b6e48;
+        D[ 12] = 32'h0b2259bf; D[ 13] = 32'h793bfff7; D[ 14] = 32'h28eeceb0; D[ 15] = 32'h8d6fde7f;
+        D[ 16] = 32'h0d618632; D[ 17] = 32'hfb371e4b; D[ 18] = 32'h0b80a9a0; D[ 19] = 32'h5c1a1381;
+        D[ 20] = 32'h8dfe755c; D[ 21] = 32'h82564b31; D[ 22] = 32'heee93cbd; D[ 23] = 32'h22bb686e;
+        D[ 24] = 32'hf2ccae97; D[ 25] = 32'he3400736; D[ 26] = 32'h11c4a85e; D[ 27] = 32'ha0c904d0;
+        D[ 28] = 32'hb495cecf; D[ 29] = 32'h32ceb818; D[ 30] = 32'ha2e46c28; D[ 31] = 32'h6915229e;
         DP[  0] = 32'h63d165e5; DP[  1] = 32'h856ac81e; DP[  2] = 32'hc4b8779d; DP[  3] = 32'h8b119544;
         DP[  4] = 32'had780837; DP[  5] = 32'h3e920266; DP[  6] = 32'he9d10f2e; DP[  7] = 32'h7c1b42b2;
         DP[  8] = 32'hc7daca3b; DP[  9] = 32'h7883be11; DP[ 10] = 32'ha384548d; DP[ 11] = 32'he0848b23;
@@ -124,7 +133,8 @@ module tb_core_full_1024;
         DQ[  0] = 32'hd7ffdc71; DQ[  1] = 32'hed01b8aa; DQ[  2] = 32'h2f99d3a6; DQ[  3] = 32'h8ccb4428;
         DQ[  4] = 32'hb1574616; DQ[  5] = 32'hfc218e36; DQ[  6] = 32'h4fe24f91; DQ[  7] = 32'h9c367c42;
         DQ[  8] = 32'h69dfa208; DQ[  9] = 32'h3ee3de79; DQ[ 10] = 32'h54ded59b; DQ[ 11] = 32'hcb3b487d;
-        DQ[ 12] = 32'hbcc0db4e; DQ[ 13] = 32'hb3e6678c; DQ[ 14] = 32'h3d13ec03; DQ[ 15] = 32'h99e0f684;        QINV[  0] = 32'h9a2f0db2; QINV[  1] = 32'h4a8075a5; QINV[  2] = 32'hb61201fa; QINV[  3] = 32'h0e876a42;
+        DQ[ 12] = 32'hbcc0db4e; DQ[ 13] = 32'hb3e6678c; DQ[ 14] = 32'h3d13ec03; DQ[ 15] = 32'h99e0f684;
+        QINV[  0] = 32'h9a2f0db2; QINV[  1] = 32'h4a8075a5; QINV[  2] = 32'hb61201fa; QINV[  3] = 32'h0e876a42;
         QINV[  4] = 32'h94667476; QINV[  5] = 32'h7538b796; QINV[  6] = 32'h8d8dfa35; QINV[  7] = 32'h689ee4a7;
         QINV[  8] = 32'h6779dd63; QINV[  9] = 32'he15b6b5e; QINV[ 10] = 32'h8275500c; QINV[ 11] = 32'he4dcd058;
         QINV[ 12] = 32'haf54b86c; QINV[ 13] = 32'hba76dc50; QINV[ 14] = 32'h473d0d6d; QINV[ 15] = 32'ha023ba44;
@@ -155,30 +165,50 @@ module tb_core_full_1024;
     end
     
 
-
     //
     // Clocks
     //
     `define CLK_FREQUENCY_MHZ   (100.0)
     `define CLK_PERIOD_NS       (1000.0 / `CLK_FREQUENCY_MHZ)
     `define CLK_PERIOD_HALF_NS  (0.5 * `CLK_PERIOD_NS)
+    `define CLK_PERIOD_QUARTER_NS (0.5    * `CLK_PERIOD_HALF_NS)
 
-    `define CLK_BUS_FREQUENCY_MHZ   (50.0)
+    `define CLK_BUS_FREQUENCY_MHZ   (25.0)
     `define CLK_BUS_PERIOD_NS       (1000.0 / `CLK_BUS_FREQUENCY_MHZ)
     `define CLK_BUS_PERIOD_HALF_NS  (0.5 * `CLK_BUS_PERIOD_NS)
     
-	reg clk = 1'b1;
-	reg clk_bus = 1'b0;
+	reg  clk      = 1'b1;
+	reg  clk_dly  = 1'b0;
+	wire clk_idle = clk & clk_dly; 
+	
+	reg  clk_bus      = 1'b1;
+	reg  clk_bus_dly  = 1'b0;
+	wire clk_bus_idle = clk_bus & clk_bus_dly;  
+
+    always #`CLK_PERIOD_HALF_NS     clk     <= ~clk;
+    always #`CLK_BUS_PERIOD_HALF_NS clk_bus <= ~clk_bus;
+    
+    always @(clk    ) clk_dly     <= #(`CLK_PERIOD_HALF_NS     - `CLK_PERIOD_QUARTER_NS) clk;
+    always @(clk_bus) clk_bus_dly <= #(`CLK_BUS_PERIOD_HALF_NS - `CLK_PERIOD_QUARTER_NS) clk_bus;  
 
-    always #`CLK_PERIOD_HALF_NS clk = ~clk;
     
-    always #`CLK_BUS_PERIOD_HALF_NS clk_bus = ~clk_bus;
+    //
+    // Clock Sync
+    //
+    task sync_clk;
+        while (clk_idle !== 1) _wait_quarter_clk_tick;
+    endtask
+    
+    task sync_clk_bus;
+        while (clk_bus_idle !== 1) _wait_quarter_clk_tick;
+    endtask
     
     
     //
     // Reset
     //
     reg rst = 1'b1;
+    wire rst_n = ~rst;
     
     
     //
@@ -196,7 +226,6 @@ module tb_core_full_1024;
     //
     // System Bus
     //
-    reg         bus_ready;
     reg         bus_cs = 1'b0;
     reg         bus_we = 1'b0;
     reg  [11:0] bus_addr;
@@ -216,7 +245,7 @@ module tb_core_full_1024;
         .clk                (clk),
         .clk_bus            (clk_bus),
         
-        .rst                (rst),
+        .rst_n              (rst_n),
         
         .next               (core_next),
         .valid              (core_valid),
@@ -238,66 +267,106 @@ module tb_core_full_1024;
 
 
     //
-    // Routine (Bus)
+    // Bus Init Routine
     //
-    initial begin
-    
-        bus_ready = 1'b0;
-    
-        while (rst) wait_clock_bus_tick;
-        wait_clock_bus_ticks(10);
-        $display("Core came out of reset.");
-         
-        set_input_1;
-        set_input_2;
-        
-        wait_clock_bus_ticks(10);
-        bus_ready = 1'b1;
-    
-    end
-    
+    task core_set_input;
+        begin         
+            core_set_input_1;
+            core_set_input_2;
+            wait_clk_bus_ticks(10);
+            $display("Core input banks written.");
+        end
+    endtask
+
     
     //
-    // Routine (Control/Status, Bus)
+    // Script
     //
-    initial begin
-    
-        _wait_half_clock_tick;
-        wait_clock_ticks(100);
-        rst = 1'b0;
-        
-        while (!bus_ready) wait_clock_tick;
-        wait_clock_ticks(10);
-        $display("Core input banks written.");
-        
-        word_index_last_n  = CORE_NUM_WORDS_N - 1;
-        word_index_last_pq = CORE_NUM_WORDS_PQ - 1;
-        
-        bit_index_last_n  = TB_MODULUS_LENGTH_N - 1;
-        bit_index_last_pq = 9; //TB_MODULUS_LENGTH_N / 2 - 1;
-        
-        core_crt_mode      = 1'b1;
-        
-        core_next = 1'b1;
-        wait_clock_tick;
-        core_next = 1'b0;
-        $display("Pulsed 'next' control signal.");
-    
-        while (!core_valid) wait_clock_tick;
-        wait_clock_ticks(10);
-        
-        $display("Detected high 'valid' status signal.");
-        core_crt_mode      = 1'bX;
+    initial main;
+
+
+    //
+    // Main Routine (Control/Status, Bus)
+    //
+    integer i, j, k;
+    task main;
+        begin
 
-        wait_clock_ticks(10);
-        get_output;
-        wait_clock_ticks(10);
         
-        $display("Core output banks read.");
+            sync_clk;               // switch to fast core clock
+            core_reset;             // reset core
+            core_set_params;        // set parameters (modulus width, exponent length)
         
-        verify;
+            sync_clk_bus;           // switch to slow bus clock
+            core_set_input;         // write to core input banks
+            
+            //sync_clk;               // switch to fast core clock
+            //core_set_crt_mode(1);   // enable CRT signing
+            //core_pulse_next;        // assert 'next' bit for one cycle
+            //core_wait_valid;        // wait till 'valid' bit gets asserted
+    
+            //sync_clk_bus;           // switch to slow bus clock
+            //core_get_output;        // read from core output banks
+            //core_verify_output;     // check, whether core output matches precomputed known good refrence values
+            
+            bit_index_last_n   = 16;
+            bus_write(2'd2, BANK_IN_2_D, 7'd0, 32'h00010001);
+            
+            sync_clk;               // switch to fast core clock
+            core_set_crt_mode(0);   // disable CRT signing
+            core_pulse_next;        // assert 'next' bit for one cycle
+            core_wait_valid;        // wait till 'valid' bit gets asserted
+    
+            sync_clk_bus;           // switch to slow bus clock
+            core_get_output;        // read from core output banks
+            core_verify_output;     // check, whether core output matches precomputed known good refrence values
+        end
+    endtask
     
-    end
+    task core_reset;
+        begin
+            wait_clk_ticks(100);
+            rst = 1'b0;
+            wait_clk_ticks(10);
+            $display("Core reset finished.");
+        end
+    endtask
+    
+    task core_set_params;
+        begin
+            word_index_last_n  = CORE_NUM_WORDS_N - 1;
+            word_index_last_pq = CORE_NUM_WORDS_PQ - 1;
+            bit_index_last_n   = TB_MODULUS_LENGTH_N - 1;
+            bit_index_last_pq  = TB_MODULUS_LENGTH_N / 2 - 1;
+            $display("Core parameters set.");
+        end
+    endtask
+    
+    task core_set_crt_mode;
+        input _crt;
+        begin
+            core_crt_mode = _crt;
+            if (_crt) $display("Enabled CRT mode.");
+            else      $display("Disabled CRT mode.");
+        end
+    endtask
+    
+    task core_pulse_next;
+        begin
+            core_next = 1'b1;
+            wait_clk_tick;
+            core_next = 1'b0;
+            $display("Pulsed core 'next' control signal.");
+        end
+    endtask
+
+    task core_wait_valid;
+        begin
+            while (!core_valid) wait_clk_tick;
+            wait_clk_ticks(10);
+            $display("Detected high core 'valid' status signal.");
+        end
+    endtask
     
     
     //
@@ -307,58 +376,60 @@ module tb_core_full_1024;
     
 
     //
-    // set_input_1;
+    // core_set_input_1
     //
-    task set_input_1;
+    task core_set_input_1;
         reg [9:0] _tn; 
         begin
             _tn = BANK_IN_1_N_COEFF * 2 ** BUS_OP_ADDR_W + TB_NUM_WORDS_N; // trick to write extra trailer word
-            for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) bus_write(2'd0, BANK_IN_1_M,        _w[6:0],  M[_w]);
-            for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) bus_write(2'd0, BANK_IN_1_N,        _w[6:0],  N[_w]);
-            for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) bus_write(2'd0, BANK_IN_1_N_FACTOR, _w[6:0],  N_FACTOR[_w]);
-            for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) bus_write(2'd0, BANK_IN_1_N_COEFF,  _w[6:0],  N_COEFF[_w]);
-                                                   bus_write(2'd0, _tn[9:7],           _tn[6:0], N_COEFF[TB_NUM_WORDS_N]);
-            for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) bus_write(2'd0, BANK_IN_1_X,        _w[6:0],  X[_w]);
-            for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) bus_write(2'd0, BANK_IN_1_Y,        _w[6:0],  Y[_w]);
+            for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) bus_write(2'd1, BANK_IN_1_M,        _w[6:0],  M[_w]);
+            for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) bus_write(2'd1, BANK_IN_1_N,        _w[6:0],  N[_w]);
+            for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) bus_write(2'd1, BANK_IN_1_N_FACTOR, _w[6:0],  N_FACTOR[_w]);
+            for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) bus_write(2'd1, BANK_IN_1_N_COEFF,  _w[6:0],  N_COEFF[_w]);
+                                                   bus_write(2'd1, _tn[9:7],           _tn[6:0], N_COEFF[TB_NUM_WORDS_N]);
+            for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) bus_write(2'd1, BANK_IN_1_X,        _w[6:0],  X[_w]);
+            for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) bus_write(2'd1, BANK_IN_1_Y,        _w[6:0],  Y[_w]);
         end
     endtask
     
     
     //
-    // set_input_2;
+    // core_set_input_2
     //
-    task set_input_2;
+    task core_set_input_2;
         begin
-//            for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) bus_write(2'd0, BANK_IN_1_M,        _w[6:0], M[_w]);
-            for (_w=0; _w< TB_NUM_WORDS_PQ; _w=_w+1) bus_write(2'd1, BANK_IN_2_P,        {1'b0, _w[5:0]}, P       [_w]);
-            for (_w=0; _w< TB_NUM_WORDS_PQ; _w=_w+1) bus_write(2'd1, BANK_IN_2_P,        {1'b1, _w[5:0]}, DP      [_w]);
-            for (_w=0; _w< TB_NUM_WORDS_PQ; _w=_w+1) bus_write(2'd1, BANK_IN_2_P_FACTOR, {      _w[6:0]}, P_FACTOR[_w]);
-            for (_w=0; _w<=TB_NUM_WORDS_PQ; _w=_w+1) bus_write(2'd1, BANK_IN_2_P_COEFF,  {      _w[6:0]}, P_COEFF [_w]);
-            for (_w=0; _w< TB_NUM_WORDS_PQ; _w=_w+1) bus_write(2'd1, BANK_IN_2_Q,        {1'b0, _w[5:0]}, Q       [_w]);
-            for (_w=0; _w< TB_NUM_WORDS_PQ; _w=_w+1) bus_write(2'd1, BANK_IN_2_Q,        {1'b1, _w[5:0]}, DQ      [_w]);
-            for (_w=0; _w< TB_NUM_WORDS_PQ; _w=_w+1) bus_write(2'd1, BANK_IN_2_Q_FACTOR, {      _w[6:0]}, Q_FACTOR[_w]);
-            for (_w=0; _w<=TB_NUM_WORDS_PQ; _w=_w+1) bus_write(2'd1, BANK_IN_2_Q_COEFF,  {      _w[6:0]}, Q_COEFF [_w]);            
-            for (_w=0; _w< TB_NUM_WORDS_PQ; _w=_w+1) bus_write(2'd1, BANK_IN_2_QINV,     {      _w[6:0]}, QINV    [_w]);
+            for (_w=0; _w< TB_NUM_WORDS_N;  _w=_w+1) bus_write(2'd2, BANK_IN_2_D,        {      _w[6:0]}, D       [_w]);
+            for (_w=0; _w< TB_NUM_WORDS_PQ; _w=_w+1) bus_write(2'd2, BANK_IN_2_P,        {1'b0, _w[5:0]}, P       [_w]);
+            for (_w=0; _w< TB_NUM_WORDS_PQ; _w=_w+1) bus_write(2'd2, BANK_IN_2_P,        {1'b1, _w[5:0]}, DP      [_w]);
+            for (_w=0; _w< TB_NUM_WORDS_PQ; _w=_w+1) bus_write(2'd2, BANK_IN_2_P_FACTOR, {      _w[6:0]}, P_FACTOR[_w]);
+            for (_w=0; _w<=TB_NUM_WORDS_PQ; _w=_w+1) bus_write(2'd2, BANK_IN_2_P_COEFF,  {      _w[6:0]}, P_COEFF [_w]);
+            for (_w=0; _w< TB_NUM_WORDS_PQ; _w=_w+1) bus_write(2'd2, BANK_IN_2_Q,        {1'b0, _w[5:0]}, Q       [_w]);
+            for (_w=0; _w< TB_NUM_WORDS_PQ; _w=_w+1) bus_write(2'd2, BANK_IN_2_Q,        {1'b1, _w[5:0]}, DQ      [_w]);
+            for (_w=0; _w< TB_NUM_WORDS_PQ; _w=_w+1) bus_write(2'd2, BANK_IN_2_Q_FACTOR, {      _w[6:0]}, Q_FACTOR[_w]);
+            for (_w=0; _w<=TB_NUM_WORDS_PQ; _w=_w+1) bus_write(2'd2, BANK_IN_2_Q_COEFF,  {      _w[6:0]}, Q_COEFF [_w]);            
+            for (_w=0; _w< TB_NUM_WORDS_PQ; _w=_w+1) bus_write(2'd2, BANK_IN_2_QINV,     {      _w[6:0]}, QINV    [_w]);
         end
     endtask
     
     
     //
-    // get_output;
+    // core_get_output
     //
-    task get_output;
+    task core_get_output;
         begin
-            for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) bus_read(2'd2, BANK_OUT_XM, _w[6:0], XM_READBACK[_w]);
-            for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) bus_read(2'd2, BANK_OUT_YM, _w[6:0], YM_READBACK[_w]);
-            for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) bus_read(2'd2, BANK_OUT_S,  _w[6:0], S_READBACK[_w]);
+            for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) bus_read(2'd3, BANK_OUT_XM, _w[6:0], XM_READBACK[_w]);
+            for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) bus_read(2'd3, BANK_OUT_YM, _w[6:0], YM_READBACK[_w]);
+            for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) bus_read(2'd3, BANK_OUT_S,  _w[6:0], S_READBACK[_w]);
+            wait_clk_bus_ticks(10);
+            $display("Core output banks read.");
         end
     endtask
 
 
     //
-    // verify;
+    // core_verify_output
     //
-    task verify;
+    task core_verify_output;
         //
         reg xm_ok;
         reg ym_ok;
@@ -440,7 +511,7 @@ module tb_core_full_1024;
         input  [31:0] data;
         begin
             _bus_drive(1'b1, 1'b1, {sel, bank, addr}, data);
-            wait_clock_bus_tick;
+            wait_clk_bus_tick;
             _bus_drive(1'b0, 1'b0, 12'hXXX, 32'hXXXXXXXX);
         end
     endtask
@@ -456,7 +527,7 @@ module tb_core_full_1024;
         output [31:0] data;
         begin
             _bus_drive(1'b1, 1'b0, {sel, bank, addr}, 32'hXXXXXXXX);
-            wait_clock_bus_tick;
+            wait_clk_bus_tick;
             data = bus_data_rd;
             _bus_drive(1'b0, 1'b0, 12'hXXX, 32'hXXXXXXXX);
         end
@@ -464,48 +535,61 @@ module tb_core_full_1024;
     
     
     //
-    // _wait_half_clock_tick()
+    // _wait_quarter_clk_tick()
     //
-    task _wait_half_clock_tick;
-        #`CLK_PERIOD_HALF_NS;
+    task _wait_quarter_clk_tick;
+        #`CLK_PERIOD_QUARTER_NS;
+    endtask
+
+
+    //
+    // _wait_half_clk_tick()
+    //
+    task _wait_half_clk_tick;
+        begin
+            _wait_quarter_clk_tick;
+            _wait_quarter_clk_tick;
+        end
     endtask
     
+
     //
-    // wait_clock_tick()
+    // wait_clk_tick()
     //
-    task wait_clock_tick;
+    task wait_clk_tick;
         begin
-            _wait_half_clock_tick;
-            _wait_half_clock_tick;
+            _wait_half_clk_tick;
+            _wait_half_clk_tick;
         end
     endtask
     
     
     //
-    // wait_clock_bus_tick()
+    // wait_clk_bus_tick()
     //
-    task wait_clock_bus_tick;
+    task wait_clk_bus_tick;
         #`CLK_BUS_PERIOD_NS;
     endtask
 
 
     //
-    // wait_clock_ticks()
+    // wait_clk_ticks()
     //    
-    task wait_clock_ticks;
+    task wait_clk_ticks;
         input integer num_ticks;
         for (_n=0; _n<num_ticks; _n=_n+1)
-            wait_clock_tick;
+            wait_clk_tick;
     endtask
     
     
     //
-    // wait_clock_bus_ticks()
+    // wait_clk_bus_ticks()
     //    
-    task wait_clock_bus_ticks;
+    task wait_clk_bus_ticks;
         input integer num_ticks;
         for (_n=0; _n<num_ticks; _n=_n+1)
-            wait_clock_bus_tick;
+            wait_clk_bus_tick;
     endtask
     
+
 endmodule
diff --git a/bench/tb_core_full_512.v b/bench/tb_core_full_512.v
index cbcdd1d..6357d4e 100644
--- a/bench/tb_core_full_512.v
+++ b/bench/tb_core_full_512.v
@@ -150,6 +150,7 @@ module tb_core_full_512;
     // Reset
     //
     reg rst = 1'b1;
+    wire rst_n = ~rst;
     
     
     //
@@ -186,7 +187,7 @@ module tb_core_full_512;
         .clk                (clk),
         .clk_bus            (clk_bus),
         
-        .rst                (rst),
+        .rst_n              (rst_n),
         
         .next               (core_next),
         .valid              (core_valid),
@@ -229,8 +230,10 @@ module tb_core_full_512;
     //
     // Main Routine (Control/Status, Bus)
     //
+    integer i, j, k;
     task main;
         begin
+
         
             sync_clk;               // switch to fast core clock
             core_reset;             // reset core
@@ -316,7 +319,6 @@ module tb_core_full_512;
     //
     task core_set_input_1;
         reg [9:0] _tn;
-        reg [31:0] zzz; 
         begin
             _tn = BANK_IN_1_N_COEFF * 2 ** BUS_OP_ADDR_W + TB_NUM_WORDS_N; // trick to write extra trailer word
             for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) bus_write(2'd1, BANK_IN_1_M,        _w[6:0],  M[_w]);
diff --git a/bench/tb_wrapper.v b/bench/tb_wrapper.v
index 2000ec5..0c51228 100644
--- a/bench/tb_wrapper.v
+++ b/bench/tb_wrapper.v
@@ -58,7 +58,7 @@ module tb_wrapper;
         .clk        (clk_bus),
         .rst_n      (rst_n),
     
-        .clk_core   (clk),
+        .clk_core   (clk_bus),
 
         .cs         (bus_cs),
         .we         (bus_we),
@@ -83,10 +83,11 @@ module tb_wrapper;
         begin
         
             sync_clk_bus;
-            wait_clk_bus_ticks(10);
+            wait_clk_bus_ticks(100);
             rst_n = 1'b1;
-            wait_clk_bus_ticks(10);
+            wait_clk_bus_ticks(100);
             
+            /*
                                             bus_read('h11, to); $display("modulus_bits = %0d",            to);
             ti =  100; bus_write('h11, ti); bus_read('h11, to); $display("modulus_bits = %0d -> %0d", ti, to);
             ti =  510; bus_write('h11, ti); bus_read('h11, to); $display("modulus_bits = %0d -> %0d", ti, to);
@@ -103,8 +104,8 @@ module tb_wrapper;
             ti = 1026; bus_write('h11, ti); bus_read('h11, to); $display("modulus_bits = %0d -> %0d", ti, to);
             
             ti = 4094; bus_write('h11, ti); bus_read('h11, to); $display("modulus_bits = %0d -> %0d", ti, to);
-            ti = 4095; bus_write('h11, ti); bus_read('h11, to); $display("modulus_bits = %0d -> %0d", ti, to);
-            ti = 4096; bus_write('h11, ti); bus_read('h11, to); $display("modulus_bits = %0d -> %0d", ti, to);
+            ti = 4095; bus_write('h11, ti); bus_read('h11, to); $display("modulus_bits = %0d -> %0d", ti, to);*/
+            ti = 4096; bus_write('h11, ti); bus_read('h11, to); $display("modulus_bits = %0d -> %0d", ti, to);/*
             ti = 4097; bus_write('h11, ti); bus_read('h11, to); $display("modulus_bits = %0d -> %0d", ti, to);
             ti = 4098; bus_write('h11, ti); bus_read('h11, to); $display("modulus_bits = %0d -> %0d", ti, to);
 
@@ -135,11 +136,17 @@ module tb_wrapper;
             ti = 1026; bus_write('h12, ti); bus_read('h12, to); $display("exponent_bits = %0d -> %0d", ti, to);
             
             ti = 4094; bus_write('h12, ti); bus_read('h12, to); $display("exponent_bits = %0d -> %0d", ti, to);
-            ti = 4095; bus_write('h12, ti); bus_read('h12, to); $display("exponent_bits = %0d -> %0d", ti, to);
-            ti = 4096; bus_write('h12, ti); bus_read('h12, to); $display("exponent_bits = %0d -> %0d", ti, to);
+            ti = 4095; bus_write('h12, ti); bus_read('h12, to); $display("exponent_bits = %0d -> %0d", ti, to);*/
+            ti = 4096; bus_write('h12, ti); bus_read('h12, to); $display("exponent_bits = %0d -> %0d", ti, to);/*
             ti = 4097; bus_write('h12, ti); bus_read('h12, to); $display("exponent_bits = %0d -> %0d", ti, to);
             ti = 4098; bus_write('h12, ti); bus_read('h12, to); $display("exponent_bits = %0d -> %0d", ti, to);
-
+            */
+            //ti = 0; bus_write('h08, ti);
+            //ti = 2; bus_write('h08, ti);
+            //ti = 0; bus_write('h08, ti);
+            //ti = 2; bus_write('h08, ti);
+            //ti = 0; bus_write('h08, ti);
+            //ti = 2; bus_write('h08, ti);
 
         end
     endtask
diff --git a/rtl/modexpng_dsp_array_block.v b/rtl/modexpng_dsp_array_block.v
index 1444aa7..2724a68 100644
--- a/rtl/modexpng_dsp_array_block.v
+++ b/rtl/modexpng_dsp_array_block.v
@@ -24,8 +24,8 @@ module modexpng_dsp_array_block
     input  [                     WORD_W        -1:0] b;
     output [NUM_MULTS_AUX      * MAC_W         -1:0] p;
     
-    wire [WORD_EXT_W -1:0] casc_a[0:NUM_MULTS_HALF-1];
-    wire [    WORD_W -1:0] casc_b[0:NUM_MULTS_HALF-1];
+    wire [DSP48E1_A_W -1:0] casc_a[0:NUM_MULTS_HALF-1];
+    wire [DSP48E1_B_W -1:0] casc_b[0:NUM_MULTS_HALF-1];
     
     wire ce_a0 = ce_a;
     reg  ce_a1 = 1'b0;
@@ -70,8 +70,8 @@ module modexpng_dsp_array_block
                 .opmode         ({1'b0, mode_z[2*z], 1'b0, 2'b01, 2'b01}),
                 .alumode        ({DSP48E1_ALUMODE_W{1'b0}}),
                 
-                .casc_a_in      (WORD_EXT_ZERO),
-                .casc_b_in      (WORD_ZERO),
+                .casc_a_in      (),
+                .casc_b_in      (),
                 
                 .casc_a_out     (casc_a[z]),
                 .casc_b_out     (casc_b[z])
@@ -138,8 +138,8 @@ module modexpng_dsp_array_block
         .opmode         ({1'b0, mode_z[2*NUM_MULTS_HALF], 1'b0, 2'b01, 2'b01}),
         .alumode        ({DSP48E1_ALUMODE_W{1'b0}}),
         
-        .casc_a_in      (WORD_EXT_ZERO),
-        .casc_b_in      (WORD_ZERO),
+        .casc_a_in      (),
+        .casc_b_in      (),
         
         .casc_a_out     (),
         .casc_b_out     ()
diff --git a/rtl/modexpng_dsp_slice_wrapper_xilinx.v b/rtl/modexpng_dsp_slice_wrapper_xilinx.v
index 9c1a60d..8c0b969 100644
--- a/rtl/modexpng_dsp_slice_wrapper_xilinx.v
+++ b/rtl/modexpng_dsp_slice_wrapper_xilinx.v
@@ -30,13 +30,11 @@ module modexpng_dsp_slice_wrapper_xilinx #
     input  [ DSP48E1_INMODE_W -1:0] inmode;
     input  [ DSP48E1_OPMODE_W -1:0] opmode;
     input  [DSP48E1_ALUMODE_W -1:0] alumode;
-    input  [       WORD_EXT_W -1:0] casc_a_in;
-    input  [           WORD_W -1:0] casc_b_in;
-    output [       WORD_EXT_W -1:0] casc_a_out;
-    output [           WORD_W -1:0] casc_b_out;
+    input  [      DSP48E1_A_W -1:0] casc_a_in;
+    input  [      DSP48E1_B_W -1:0] casc_b_in;
+    output [      DSP48E1_A_W -1:0] casc_a_out;
+    output [      DSP48E1_B_W -1:0] casc_b_out;
     
-    wire [DSP48E1_A_W - WORD_EXT_W -1:0] casc_a_dummy;
-    wire [DSP48E1_B_W - WORD_W     -1:0] casc_b_dummy;
     wire [DSP48E1_P_W - MAC_W      -1:0] p_dummy;
 
     DSP48E1 #
@@ -99,10 +97,10 @@ module modexpng_dsp_slice_wrapper_xilinx #
         .OPMODE             (opmode),
         .ALUMODE            (alumode),
 
-        .ACIN               ({{(DSP48E1_A_W-WORD_EXT_W){1'b0}}, casc_a_in}),
-        .BCIN               ({{(DSP48E1_B_W-WORD_W){1'b0}}, casc_b_in}),
-        .ACOUT              ({casc_a_dummy, casc_a_out}),
-        .BCOUT              ({casc_b_dummy, casc_b_out}),
+        .ACIN               (casc_a_in),
+        .BCIN               (casc_b_in),
+        .ACOUT              (casc_a_out),
+        .BCOUT              (casc_b_out),
         .PCIN               ({DSP48E1_P_W{1'b0}}),
         .PCOUT              (),
         .CARRYCASCIN        (1'b0),
diff --git a/rtl/modexpng_uop_rom.v b/rtl/modexpng_uop_rom.v
index c15f608..74f7ea3 100644
--- a/rtl/modexpng_uop_rom.v
+++ b/rtl/modexpng_uop_rom.v
@@ -134,7 +134,7 @@ module modexpng_uop_rom
 
             7'd080:  data <= {UOP_OPCODE_OUTPUT_FROM_NARROW,  UOP_CRT_X,   UOP_NPQ_N,   UOP_AUX_1,   UOP_LADDER_DNC, BANK_DNC,      BANK_NARROW_D,      BANK_DNC,    BANK_OUT_XM      }; //
             7'd081:  data <= {UOP_OPCODE_OUTPUT_FROM_NARROW,  UOP_CRT_Y,   UOP_NPQ_N,   UOP_AUX_1,   UOP_LADDER_DNC, BANK_DNC,      BANK_NARROW_D,      BANK_DNC,    BANK_OUT_YM      }; //
-                                                                                                                                                                                         //            
+            
             7'd082:  data <= {UOP_OPCODE_MODULAR_MULTIPLY,    UOP_CRT_DNC, UOP_NPQ_N,   UOP_AUX_1,   UOP_LADDER_11,  BANK_WIDE_E,   BANK_NARROW_B,      BANK_WIDE_C, BANK_NARROW_C    }; //
 
             7'd083:  data <= {UOP_OPCODE_INPUT_TO_WIDE,       UOP_CRT_X,   UOP_NPQ_N,   UOP_AUX_1,   UOP_LADDER_DNC, BANK_DNC,      BANK_IN_1_N_FACTOR, BANK_WIDE_A, BANK_DNC         }; //
diff --git a/rtl/modexpng_wrapper.v b/rtl/modexpng_wrapper.v
index 687a963..0af6c32 100644
--- a/rtl/modexpng_wrapper.v
+++ b/rtl/modexpng_wrapper.v
@@ -111,7 +111,8 @@ module modexpng_wrapper
     //
     localparam MIN_OP_W = 2 * NUM_MULTS * WORD_W * 2;
     localparam MIN_EXP_W = 2 * 2;
-    localparam ZEROES_BIT_INDEX_W = 1 + cryptech_clog2(NUM_MULTS) + cryptech_clog2(WORD_W);
+    localparam LSB_BIT_INDEX_W = 1 + cryptech_clog2(NUM_MULTS) + cryptech_clog2(WORD_W);
+    localparam MSB_BIT_INDEX_W = BIT_INDEX_W - LSB_BIT_INDEX_W;
     
     
     //
@@ -125,35 +126,29 @@ module modexpng_wrapper
     //
     // Registers
     //
-    reg wrap_reg_control     = `MODEXPNG_DEFAULT_CONTROL;
-    reg sync_reg_control     = `MODEXPNG_DEFAULT_CONTROL;
-    reg sync_reg_control_dly = `MODEXPNG_DEFAULT_CONTROL;
-    reg core_reg_control     = `MODEXPNG_DEFAULT_CONTROL;
-    reg wrap_reg_mode        = `MODEXPNG_DEFAULT_MODE;
+    reg wrap_reg_control = `MODEXPNG_DEFAULT_CONTROL;
+    reg core_reg_control = `MODEXPNG_DEFAULT_CONTROL;
+    reg wrap_reg_mode    = `MODEXPNG_DEFAULT_MODE;
     reg sync_reg_mode;
     reg core_reg_mode;
         
-    reg [BIT_INDEX_W:ZEROES_BIT_INDEX_W] wrap_modulus_bits_msb;
-    reg [BIT_INDEX_W:                 0] wrap_exponent_bits;
+    reg [BIT_INDEX_W:LSB_BIT_INDEX_W] wrap_modulus_bits_msb;
+    reg [BIT_INDEX_W:              0] wrap_exponent_bits;
 
-    initial write_modulus_bits(`MODEXPNG_DEFAULT_MODULUS_BITS);
-    initial write_exponent_bits(`MODEXPNG_DEFAULT_EXPONENT_BITS);
-
-    wire sync_reg_control_rising = sync_reg_control & ~sync_reg_control_dly;
+    initial update_modulus_bits_user;
+    initial update_exponent_bits_user;
+    
+    //wire sync_reg_control_rising = sync_reg_control & ~sync_reg_control_dly;
 
 
     //
     // Wires
     //
-    reg  wrap_reg_status = 1'b1;
-    reg  sync_reg_status = 1'b1;
+    reg  wrap_reg_status     = 1'b1;
+    reg  sync_reg_status     = 1'b1;
+    reg  sync_reg_status_dly = 1'b1;
     wire core_reg_status;
     
-    always @(posedge clk or negedge rst_n)
-        //
-        if (!rst_n) {wrap_reg_status, sync_reg_status} <= {           1'b1,            1'b1};
-        else        {wrap_reg_status, sync_reg_status} <= {sync_reg_status, core_reg_status};
-
 
     //
     // Output Mux
@@ -175,8 +170,76 @@ module modexpng_wrapper
         else        core_rst_shreg <= {core_rst_shreg[14:0], 1'b1};
 
     assign core_rst_n = core_rst_shreg[15];
+     
+    
+    //
+    // Trigger Logic
+    //
+    reg wrap_reg_control_dly     = `MODEXPNG_DEFAULT_CONTROL;
+    reg wrap_reg_control_posedge = 1'b0;
+    reg sync_reg_control_posedge = 1'b0;
+    reg core_reg_control_posedge = 1'b0;
+    reg core_reg_control_posedge_dly = 1'b0;
+    reg sync_reg_control_posedge_ack = 1'b0;
+    reg wrap_reg_control_posedge_ack = 1'b0;
+    
+    always @(posedge clk or negedge rst_n)
+        if (!rst_n) wrap_reg_control_dly <= `MODEXPNG_DEFAULT_CONTROL;
+        else        wrap_reg_control_dly <= wrap_reg_control;
+        
+    always @(posedge clk or negedge rst_n)
+        if (!rst_n) wrap_reg_control_posedge <= 1'b0;
+        else begin
+            if (!wrap_reg_control_posedge) begin
+                if (wrap_reg_control && !wrap_reg_control_dly) wrap_reg_control_posedge <= 1'b1;
+            end else begin
+                if (wrap_reg_control_posedge_ack) wrap_reg_control_posedge <= 1'b0;
+            end
+        end
+    
+    always @(posedge clk_core or negedge core_rst_n)
+        if (!core_rst_n) sync_reg_control_posedge <= 1'b0;
+        else             sync_reg_control_posedge <= wrap_reg_control_posedge;
+    
+    always @(posedge clk_core or negedge core_rst_n)
+        if (!core_rst_n) core_reg_control_posedge <= 1'b0;
+        else             core_reg_control_posedge <= sync_reg_control_posedge;
+        
+    always @(posedge clk_core or negedge core_rst_n)
+        if (!core_rst_n) core_reg_control_posedge_dly <= 1'b0;
+        else             core_reg_control_posedge_dly <= core_reg_control_posedge;
+    
+    always @(posedge clk or negedge rst_n)
+        if (!rst_n) sync_reg_control_posedge_ack <= 1'b0;
+        else        sync_reg_control_posedge_ack <= core_reg_control_posedge;
+        
+    always @(posedge clk or negedge rst_n)
+        if (!rst_n) wrap_reg_control_posedge_ack <= 1'b0;
+        else        wrap_reg_control_posedge_ack <= sync_reg_control_posedge_ack;
+
+    always @(posedge clk_core or negedge core_rst_n)
+        if (!core_rst_n) core_reg_control <= `MODEXPNG_DEFAULT_CONTROL;
+        else             core_reg_control <= core_reg_control_posedge && !core_reg_control_posedge_dly;
     
+    always @(posedge clk or negedge rst_n)
+        if (!rst_n) sync_reg_status <= 1'b1;
+        else        sync_reg_status <= core_reg_status;
+        
+    always @(posedge clk or negedge rst_n)
+        if (!rst_n) sync_reg_status_dly <= 1'b1;
+        else        sync_reg_status_dly <= sync_reg_status;
+        
+    always @(posedge clk or negedge rst_n)
+        if (!rst_n) wrap_reg_status <= 1'b1;
+        else begin
+            if (wrap_reg_status) begin
+                if (wrap_reg_control && !wrap_reg_control_dly) wrap_reg_status <= 1'b0;
+            end else begin
+                if (!wrap_reg_control_posedge && !wrap_reg_control_posedge_ack && sync_reg_status_dly) wrap_reg_status <= 1'b1;
+            end
+        end
 
+    
     //
     // Parameters Resync
     //
@@ -212,7 +275,7 @@ module modexpng_wrapper
     
     always @(posedge clk_core)
         //
-        if (sync_reg_control_rising) begin
+        if (core_reg_control_posedge && !core_reg_control_posedge_dly) begin
             //
             core_reg_mode <= sync_reg_mode;
             //
@@ -224,21 +287,6 @@ module modexpng_wrapper
             //
         end
     
-    
-    //
-    // Trigger Logic
-    //    
-    always @(posedge clk_core or negedge rst_n)
-        //
-        if (!rst_n) {sync_reg_control_dly, sync_reg_control} <= {`MODEXPNG_DEFAULT_CONTROL, `MODEXPNG_DEFAULT_CONTROL};
-        else        {sync_reg_control_dly, sync_reg_control} <= {         sync_reg_control,          wrap_reg_control};
-
-    always @(posedge clk_core or negedge rst_n)
-        //
-        if (!rst_n) core_reg_control <= `MODEXPNG_DEFAULT_CONTROL;
-        else        core_reg_control <= sync_reg_control_rising;     
-    
-    
     //
     // ModExpNG
     //
@@ -271,6 +319,63 @@ module modexpng_wrapper
     //
     // Write Interface
     //
+    
+    wire [    BIT_INDEX_W  :0] corrected_modulus_bits_user     = correct_modulus_bits(write_data[BIT_INDEX_W:0]);
+    wire [MSB_BIT_INDEX_W  :0] corrected_modulus_bits_msb_user = corrected_modulus_bits_user[BIT_INDEX_W:LSB_BIT_INDEX_W];
+    wire [      OP_ADDR_W  :0] modulus_num_words_n_user        = {corrected_modulus_bits_msb_user, {(LSB_BIT_INDEX_W-WORD_MUX_W){1'b0}}} - 1'b1;
+    wire [      OP_ADDR_W-1:0] modulus_num_words_n_lsb_user    = modulus_num_words_n_user[OP_ADDR_W-1:0];
+    wire [      OP_ADDR_W-1:0] modulus_num_words_pq_user       = {corrected_modulus_bits_msb_user, {(LSB_BIT_INDEX_W-WORD_MUX_W-1){1'b0}}} - 1'b1;
+
+    wire [    BIT_INDEX_W  :0] corrected_modulus_bits_default     = `MODEXPNG_DEFAULT_MODULUS_BITS;
+    wire [MSB_BIT_INDEX_W  :0] corrected_modulus_bits_msb_default = corrected_modulus_bits_default[BIT_INDEX_W:LSB_BIT_INDEX_W];
+    wire [      OP_ADDR_W  :0] modulus_num_words_n_default        = {corrected_modulus_bits_msb_default, {(LSB_BIT_INDEX_W-WORD_MUX_W){1'b0}}} - 1'b1;
+    wire [      OP_ADDR_W-1:0] modulus_num_words_n_lsb_default    = modulus_num_words_n_default[OP_ADDR_W-1:0];
+    wire [      OP_ADDR_W-1:0] modulus_num_words_pq_default       = {corrected_modulus_bits_msb_default, {(LSB_BIT_INDEX_W-WORD_MUX_W-1){1'b0}}} - 1'b1;
+    
+    wire [    BIT_INDEX_W  :0] corrected_exponent_bits_user     = correct_exponent_bits(write_data[BIT_INDEX_W:0]);
+    wire [    BIT_INDEX_W-1:0] corrected_exponent_bits_msb_user = corrected_exponent_bits_user[BIT_INDEX_W:1];
+    wire [    BIT_INDEX_W  :0] exponent_num_bits_n_user         = corrected_exponent_bits_user - 1'b1;
+    wire [    BIT_INDEX_W-1:0] exponent_num_bits_n_lsb_user     = exponent_num_bits_n_user[BIT_INDEX_W-1:0];
+    wire [    BIT_INDEX_W-1:0] exponent_num_bits_pq_user        = corrected_exponent_bits_msb_user - 1'b1;
+    
+    wire [    BIT_INDEX_W  :0] corrected_exponent_bits_default     = `MODEXPNG_DEFAULT_EXPONENT_BITS;
+    wire [    BIT_INDEX_W-1:0] corrected_exponent_bits_msb_default = corrected_exponent_bits_default[BIT_INDEX_W:1];
+    wire [    BIT_INDEX_W  :0] exponent_num_bits_n_default         = corrected_exponent_bits_default - 1'b1;
+    wire [    BIT_INDEX_W-1:0] exponent_num_bits_n_lsb_default     = exponent_num_bits_n_default[BIT_INDEX_W-1:0];
+    wire [    BIT_INDEX_W-1:0] exponent_num_bits_pq_default        = corrected_exponent_bits_msb_default - 1'b1;
+    
+    task update_modulus_bits_user;
+        begin
+            wrap_modulus_bits_msb   <= corrected_modulus_bits_msb_user;
+            wrap_word_index_last_n  <= modulus_num_words_n_lsb_user;
+            wrap_word_index_last_pq <= modulus_num_words_pq_user;
+        end
+    endtask
+    
+    task update_modulus_bits_default;
+        begin
+            wrap_modulus_bits_msb   <= corrected_modulus_bits_msb_default;
+            wrap_word_index_last_n  <= modulus_num_words_n_lsb_default;
+            wrap_word_index_last_pq <= modulus_num_words_pq_default;
+        end
+    endtask
+    
+    task update_exponent_bits_user;
+        begin
+            wrap_exponent_bits     <= corrected_exponent_bits_user;
+            wrap_bit_index_last_n  <= exponent_num_bits_n_lsb_user;
+            wrap_bit_index_last_pq <= exponent_num_bits_pq_user;
+        end
+    endtask
+    
+    task update_exponent_bits_default;
+        begin
+            wrap_exponent_bits     <= corrected_exponent_bits_default;
+            wrap_bit_index_last_n  <= exponent_num_bits_n_lsb_default;
+            wrap_bit_index_last_pq <= exponent_num_bits_pq_default;
+        end
+    endtask
+    
     always @(posedge clk or negedge rst_n)
         //
         if (!rst_n) begin
@@ -278,77 +383,40 @@ module modexpng_wrapper
             wrap_reg_control   <= `MODEXPNG_DEFAULT_CONTROL;
             wrap_reg_mode      <= `MODEXPNG_DEFAULT_MODE;
             //
-            write_modulus_bits(`MODEXPNG_DEFAULT_MODULUS_BITS);
-            write_exponent_bits(`MODEXPNG_DEFAULT_EXPONENT_BITS);
+            update_modulus_bits_default;
+            update_exponent_bits_default;
             //
         end else if (cs && we && addr_msb_is_wrap)
             //
             case (addr_lsb)
                 ADDR_CONTROL:       wrap_reg_control <= write_data[CONTROL_NEXT_BIT];
                 ADDR_MODE:          wrap_reg_mode    <= write_data[MODE_FULLCRT_BIT];
-                ADDR_MODULUS_BITS:  write_modulus_bits(write_data[BIT_INDEX_W:0]);
-                ADDR_EXPONENT_BITS: write_exponent_bits(write_data[BIT_INDEX_W:0]);
+                ADDR_MODULUS_BITS:  update_modulus_bits_user;
+                ADDR_EXPONENT_BITS: update_exponent_bits_user;
             endcase
 
 
     //
-    // Update modulus width
+    // Only accept correct modulus width
     // 
-    function [BIT_INDEX_W:ZEROES_BIT_INDEX_W] fix_modulus_bits;
-       input [BIT_INDEX_W:                 0] width;
-            if      (width < MIN_OP_W) fix_modulus_bits = MIN_OP_W[BIT_INDEX_W:ZEROES_BIT_INDEX_W];
-            else if (width > MAX_OP_W) fix_modulus_bits = MAX_OP_W[BIT_INDEX_W:ZEROES_BIT_INDEX_W];
-            else                       fix_modulus_bits = width   [BIT_INDEX_W:ZEROES_BIT_INDEX_W];
-    endfunction
-    
-    function [OP_ADDR_W-1:                 0] calc_modulus_num_words_n;
-       input [BIT_INDEX_W:ZEROES_BIT_INDEX_W] width;
-       calc_modulus_num_words_n = {width, {(ZEROES_BIT_INDEX_W-WORD_MUX_W){1'b0}}} - 1'b1; // truncates msb
-    endfunction
-    
-    function [OP_ADDR_W-1:                 0] calc_modulus_num_words_pq;
-       input [BIT_INDEX_W:ZEROES_BIT_INDEX_W] width;
-       calc_modulus_num_words_pq = {width, {(ZEROES_BIT_INDEX_W-WORD_MUX_W-1){1'b0}}} - 1'b1; // fits exactly
-    endfunction
-   
-    task write_modulus_bits;
+    function  [BIT_INDEX_W:0] correct_modulus_bits;
         input [BIT_INDEX_W:0] width;
-        begin
-            wrap_modulus_bits_msb   <= fix_modulus_bits(width);
-            wrap_word_index_last_n  <= calc_modulus_num_words_n(fix_modulus_bits(width));
-            wrap_word_index_last_pq <= calc_modulus_num_words_pq(fix_modulus_bits(width));
-        end
-    endtask
+            if      (width < MIN_OP_W) correct_modulus_bits = MIN_OP_W;
+            else if (width > MAX_OP_W) correct_modulus_bits = MAX_OP_W;
+            else                       correct_modulus_bits = width;
+    endfunction
     
     
     //
-    // Update exponent width 
+    // Only accept correct exponent width 
     //
-    function [BIT_INDEX_W:0] fix_exponent_bits;
+    function [BIT_INDEX_W:0] correct_exponent_bits;
        input [BIT_INDEX_W:0] width;
-            if      (width < MIN_EXP_W) fix_exponent_bits = MIN_EXP_W;
-            else if (width > MAX_OP_W ) fix_exponent_bits = MAX_OP_W;
-            else                        fix_exponent_bits = width;
-    endfunction
-    
-    function [BIT_INDEX_W-1:0] calc_exponent_num_bits_n;
-       input [BIT_INDEX_W  :0] width;
-       calc_exponent_num_bits_n = width - 1'b1;  // truncates msb
-    endfunction
-    
-    function [BIT_INDEX_W-1:0] calc_exponent_num_bits_pq;
-       input [BIT_INDEX_W:  0] width;
-       calc_exponent_num_bits_pq = width[BIT_INDEX_W:1] - 1'b1; // fits exactly
+            if      (width < MIN_EXP_W) correct_exponent_bits = MIN_EXP_W;
+            else if (width > MAX_OP_W ) correct_exponent_bits = MAX_OP_W;
+            else                        correct_exponent_bits = width;
     endfunction
-   
-    task write_exponent_bits;
-        input [BIT_INDEX_W:0] width;
-        begin
-            wrap_exponent_bits     <= fix_exponent_bits(width);
-            wrap_bit_index_last_n  <= calc_exponent_num_bits_n(fix_exponent_bits(width));
-            wrap_bit_index_last_pq <= calc_exponent_num_bits_pq(fix_exponent_bits(width));
-        end
-    endtask
+  
     
     
     //
@@ -367,7 +435,7 @@ module modexpng_wrapper
                 ADDR_STATUS:        wrap_read_data <= {{30{1'b0}}, wrap_reg_status,  1'b1};
                 //
                 ADDR_MODE:          wrap_read_data <= {{30{1'b0}}, wrap_reg_mode, 1'b0};
-                ADDR_MODULUS_BITS:  wrap_read_data <= {{(31-BIT_INDEX_W){1'b0}}, wrap_modulus_bits_msb, {ZEROES_BIT_INDEX_W{1'b0}}};
+                ADDR_MODULUS_BITS:  wrap_read_data <= {{(31-BIT_INDEX_W){1'b0}}, wrap_modulus_bits_msb, {LSB_BIT_INDEX_W{1'b0}}};
                 ADDR_EXPONENT_BITS: wrap_read_data <= {{(31-BIT_INDEX_W){1'b0}}, wrap_exponent_bits};
                 ADDR_BANK_BITS:     wrap_read_data <= MAX_OP_W;
                 ADDR_NUM_MULTS:     wrap_read_data <= NUM_MULTS;