Entire CRT signature algorithm works by now.

Moved micro-operations handler into a separate module file, this way we don't
have any synthesized stuff in the top-level module, just instantiations. This
is more consistent from the design partitioning point of view. Btw, Xilinx
claims their tools work better that way too, but who knows...

Added optional simulation-only code to assist debugging. Un-comment the
ENABLE_DEBUG `define in 'rtl/modexpng_parameters.vh' to use, but don't ever
try to synthesize the core with debugging enabled.

2 files changed, 461 insertions, 3 deletions

diff --git a/bench/tb_core_full.v b/bench/tb_core_full_1024.v
index e592ac5..e6b1a66 100644
--- a/bench/tb_core_full.v
+++ b/bench/tb_core_full_1024.v
@@ -1,6 +1,6 @@
 `timescale 1ns / 1ps
 
-module tb_core_full;
+module tb_core_full_1024;
 
 
     //
@@ -274,7 +274,7 @@ module tb_core_full;
         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;
+        bit_index_last_pq = 9; //TB_MODULUS_LENGTH_N / 2 - 1;
         
         core_crt_mode      = 1'b1;
         
@@ -295,7 +295,7 @@ module tb_core_full;
         
         $display("Core output banks read.");
         
-        //verify;
+        verify;
     
     end
     
diff --git a/bench/tb_core_full_512.v b/bench/tb_core_full_512.v
new file mode 100644
index 0000000..221a2c6
--- /dev/null
+++ b/bench/tb_core_full_512.v
@@ -0,0 +1,458 @@
+`timescale 1ns / 1ps
+
+module tb_core_full_512;
+
+
+    //
+    // Headers
+    //
+    `include "../rtl/modexpng_parameters.vh"
+
+
+    //
+    // Test Vectors
+    //
+    localparam TB_MODULUS_LENGTH_N  = 512;
+    localparam TB_MODULUS_LENGTH_PQ = TB_MODULUS_LENGTH_N  / 2;
+    localparam TB_NUM_WORDS_PQ      = TB_MODULUS_LENGTH_PQ / BUS_DATA_W;
+    localparam TB_NUM_WORDS_N       = TB_MODULUS_LENGTH_N  / BUS_DATA_W;
+    localparam CORE_NUM_WORDS_PQ    = TB_MODULUS_LENGTH_PQ / WORD_W;
+    localparam CORE_NUM_WORDS_N     = TB_MODULUS_LENGTH_N  / WORD_W;
+
+    reg [31:0] M[0:TB_NUM_WORDS_N-1];
+    reg [31:0] N[0:TB_NUM_WORDS_N-1];
+    reg [31:0] N_FACTOR[0:TB_NUM_WORDS_N-1];
+    reg [31:0] N_COEFF[0:TB_NUM_WORDS_N];
+    reg [31:0] X[0:TB_NUM_WORDS_N-1];
+    reg [31:0] Y[0:TB_NUM_WORDS_N-1];
+    reg [31:0] P[0:TB_NUM_WORDS_PQ-1];
+    reg [31:0] Q[0:TB_NUM_WORDS_PQ-1];
+    reg [31:0] P_FACTOR[0:TB_NUM_WORDS_PQ-1];
+    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] 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];
+    reg [31:0] XM[0:TB_NUM_WORDS_N-1];
+    reg [31:0] YM[0:TB_NUM_WORDS_N-1];
+    reg [31:0] S[0:TB_NUM_WORDS_N-1];
+    reg [31:0] XM_READBACK[0:TB_NUM_WORDS_N-1];
+    reg [31:0] YM_READBACK[0:TB_NUM_WORDS_N-1];
+    reg [31:0] S_READBACK[0:TB_NUM_WORDS_N-1];
+    
+    initial begin
+        M[  0] = 32'h8d3b583b; M[  1] = 32'hc370f07e; M[  2] = 32'hb9078738; M[  3] = 32'haf37f86c;
+        M[  4] = 32'h02f0e161; M[  5] = 32'h0506a68a; M[  6] = 32'h1ae65107; M[  7] = 32'hcd3a97f1;
+        M[  8] = 32'hb27244b8; M[  9] = 32'h9bc3c400; M[ 10] = 32'he4d5636e; M[ 11] = 32'h35187c07;
+        M[ 12] = 32'h78a661c9; M[ 13] = 32'h1e7ec273; M[ 14] = 32'hcdc31041; M[ 15] = 32'h002291d8;
+        N[  0] = 32'hcb703101; N[  1] = 32'h82bc8290; N[  2] = 32'hdb2372c2; N[  3] = 32'hdeeb692e;
+        N[  4] = 32'ha3ee352a; N[  5] = 32'h81a711ba; N[  6] = 32'h14ee23bd; N[  7] = 32'h8ad351c0;
+        N[  8] = 32'h75ecd3d5; N[  9] = 32'h51c9b22f; N[ 10] = 32'hc1d3496e; N[ 11] = 32'h48176f3e;
+        N[ 12] = 32'hd2aca749; N[ 13] = 32'hf236cea9; N[ 14] = 32'h7f4525ed; N[ 15] = 32'hb4fc5067;        
+        N_FACTOR[  0] = 32'he253bfbf; N_FACTOR[  1] = 32'h8e0b26aa; N_FACTOR[  2] = 32'h0480b661; N_FACTOR[  3] = 32'h9a13f7a1;
+        N_FACTOR[  4] = 32'h464b7342; N_FACTOR[  5] = 32'hfb6f8e41; N_FACTOR[  6] = 32'h081208e4; N_FACTOR[  7] = 32'h63d8328a;
+        N_FACTOR[  8] = 32'h604d2b71; N_FACTOR[  9] = 32'hc987dabe; N_FACTOR[ 10] = 32'h8a474e35; N_FACTOR[ 11] = 32'hc053ba1c;
+        N_FACTOR[ 12] = 32'h15b82dd9; N_FACTOR[ 13] = 32'h42c2bbfa; N_FACTOR[ 14] = 32'h1681e95d; N_FACTOR[ 15] = 32'h07dee5fa;
+        N_COEFF[  0] = 32'h730f30ff; N_COEFF[  1] = 32'h50ed900a; N_COEFF[  2] = 32'h0b9038c5; N_COEFF[  3] = 32'h974ddd03;
+        N_COEFF[  4] = 32'he2c118c8; N_COEFF[  5] = 32'hbe1bc7e1; N_COEFF[  6] = 32'h224d548c; N_COEFF[  7] = 32'h48ea2ee4;
+        N_COEFF[  8] = 32'heb379247; N_COEFF[  9] = 32'had97b934; N_COEFF[ 10] = 32'hfc6dfd93; N_COEFF[ 11] = 32'h3a0246ef;
+        N_COEFF[ 12] = 32'h1baa167c; N_COEFF[ 13] = 32'h7d7ee254; N_COEFF[ 14] = 32'h657f0a53; N_COEFF[ 15] = 32'hea9e7245;
+        N_COEFF[ 16] = 32'h0000f88c;        
+        X[  0] = 32'h2d532c22; X[  1] = 32'h2d3c3b06; X[  2] = 32'he2862a8f; X[  3] = 32'he8616ce4;
+        X[  4] = 32'h5d77ee51; X[  5] = 32'he609de07; X[  6] = 32'hef718044; X[  7] = 32'h82f35f8b;
+        X[  8] = 32'hcdb9dcfe; X[  9] = 32'hff6ea364; X[ 10] = 32'h0994ae28; X[ 11] = 32'h409b369b;
+        X[ 12] = 32'hcfabda4e; X[ 13] = 32'h5cd52bbc; X[ 14] = 32'hd90e1715; X[ 15] = 32'h00f4dcf2;
+        Y[  0] = 32'h34fff653; Y[  1] = 32'h50f52544; Y[  2] = 32'h0ebf96a7; Y[  3] = 32'h98352265;
+        Y[  4] = 32'hbe372927; Y[  5] = 32'h5b2f6394; Y[  6] = 32'h9acfccb3; Y[  7] = 32'h7b5bd4b2;
+        Y[  8] = 32'h79b09448; Y[  9] = 32'h08f11fa6; Y[ 10] = 32'h8411d066; Y[ 11] = 32'h58ba5021;
+        Y[ 12] = 32'h03c1cb72; Y[ 13] = 32'hacf0689d; Y[ 14] = 32'h983c65bd; Y[ 15] = 32'h29a39dcc;
+        P[  0] = 32'hebfc2433; P[  1] = 32'ha2cfbc81; P[  2] = 32'hea08812b; P[  3] = 32'h0adf004f;
+        P[  4] = 32'hb987a8c6; P[  5] = 32'h2860f873; P[  6] = 32'haf2cfe12; P[  7] = 32'hddd53c3a;
+        Q[  0] = 32'hdc1981fb; Q[  1] = 32'h01184053; Q[  2] = 32'h7ab8d640; Q[  3] = 32'h62ba8a22;
+        Q[  4] = 32'h6cb226a1; Q[  5] = 32'he1f08e16; Q[  6] = 32'h13e990b5; Q[  7] = 32'hd0dc7ce3;        
+        P_FACTOR[  0] = 32'h043fb284; P_FACTOR[  1] = 32'hcaab7ce3; P_FACTOR[  2] = 32'h543c62ef; P_FACTOR[  3] = 32'h8aa74942;
+        P_FACTOR[  4] = 32'hefa2ea7b; P_FACTOR[  5] = 32'hdb8513b5; P_FACTOR[  6] = 32'h0ea607a4; P_FACTOR[  7] = 32'h6a59e5a7;
+        Q_FACTOR[  0] = 32'h089dcd43; Q_FACTOR[  1] = 32'h5b23611b; Q_FACTOR[  2] = 32'h02f0f47c; Q_FACTOR[  3] = 32'h952ababd;
+        Q_FACTOR[  4] = 32'hc4ee13fe; Q_FACTOR[  5] = 32'h3feb46fa; Q_FACTOR[  6] = 32'h96b679df; Q_FACTOR[  7] = 32'h831126dd;
+        P_COEFF[  0] = 32'h647c8905; P_COEFF[  1] = 32'hcb7c6b7d; P_COEFF[  2] = 32'h8053b8be; P_COEFF[  3] = 32'hb28f33a7;
+        P_COEFF[  4] = 32'hb3207e05; P_COEFF[  5] = 32'h4e3d416e; P_COEFF[  6] = 32'h1911d8d9; P_COEFF[  7] = 32'hd569156e;
+        P_COEFF[  8] = 32'h00003dd7;
+        Q_COEFF[  0] = 32'h5eee9ecd; Q_COEFF[  1] = 32'h085153b0; Q_COEFF[  2] = 32'h85326da6; Q_COEFF[  3] = 32'h7521931a;
+        Q_COEFF[  4] = 32'h99e0eef1; Q_COEFF[  5] = 32'ha219917b; Q_COEFF[  6] = 32'he8e9087a; Q_COEFF[  7] = 32'h5239d12b;
+        Q_COEFF[  8] = 32'h0000ed92;
+        DP[  0] = 32'h3891ed91; DP[  1] = 32'h775046c2; DP[  2] = 32'h60180c26; DP[  3] = 32'h5130700a;
+        DP[  4] = 32'hb13c8216; DP[  5] = 32'h833fcf78; DP[  6] = 32'h7ab89b12; DP[  7] = 32'hb976758c;
+        DQ[  0] = 32'h28cc59ad; DQ[  1] = 32'h3ce6ed45; DQ[  2] = 32'ha1f53aeb; DQ[  3] = 32'h06ca05e1;
+        DQ[  4] = 32'hc5195df6; DQ[  5] = 32'h42cf91f8; DQ[  6] = 32'h93d6f054; DQ[  7] = 32'h3d3bc769;
+        QINV[  0] = 32'h50201af6; QINV[  1] = 32'h85d97b7f; QINV[  2] = 32'h4247e697; QINV[  3] = 32'h9fd231fe;
+        QINV[  4] = 32'h21e98610; QINV[  5] = 32'ha0bc58dc; QINV[  6] = 32'ha86f266c; QINV[  7] = 32'h838688c8;
+        XM[  0] = 32'hf9980f33; XM[  1] = 32'hb444f483; XM[  2] = 32'h0a6f8294; XM[  3] = 32'h1c74da49;
+        XM[  4] = 32'h0aa4151f; XM[  5] = 32'ha1dfb66f; XM[  6] = 32'h1415da79; XM[  7] = 32'had7d3272;
+        XM[  8] = 32'h43d7b612; XM[  9] = 32'h56626cce; XM[ 10] = 32'ha65edef6; XM[ 11] = 32'h28c49eb8;
+        XM[ 12] = 32'h5364b7f8; XM[ 13] = 32'hd170915e; XM[ 14] = 32'h5a4c960d; XM[ 15] = 32'h27cb1911;
+        YM[  0] = 32'h7640c2ca; YM[  1] = 32'hf49d583a; YM[  2] = 32'he4ae0f22; YM[  3] = 32'ha3dad5ed;
+        YM[  4] = 32'hbe88ab4d; YM[  5] = 32'h9fb50b38; YM[  6] = 32'h223feceb; YM[  7] = 32'hfc4893ff;
+        YM[  8] = 32'hb40556c2; YM[  9] = 32'hb25b27fa; YM[ 10] = 32'h7e277535; YM[ 11] = 32'h42e1e9ab;
+        YM[ 12] = 32'hebd55ef2; YM[ 13] = 32'h8b6d8c0b; YM[ 14] = 32'h4d91ad9a; YM[ 15] = 32'h0e8bf565;
+        S[  0] = 32'h2f89f059; S[  1] = 32'hdbc41170; S[  2] = 32'h1d7ea6c0; S[  3] = 32'h1df9add6;
+        S[  4] = 32'ha619e2e1; S[  5] = 32'h253fcd88; S[  6] = 32'h6c03a351; S[  7] = 32'h795b1df0;
+        S[  8] = 32'h2854a51a; S[  9] = 32'h0245619b; S[ 10] = 32'hfb67ef8f; S[ 11] = 32'hcc5bdd4f;
+        S[ 12] = 32'ha70f58bd; S[ 13] = 32'h31f15702; S[ 14] = 32'hd6f36259; S[ 15] = 32'h280e67e0;
+    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_BUS_FREQUENCY_MHZ   (50.0)
+    `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;
+
+    always #`CLK_PERIOD_HALF_NS clk = ~clk;
+    
+    always #`CLK_BUS_PERIOD_HALF_NS clk_bus = ~clk_bus;
+    
+    
+    //
+    // Reset
+    //
+    reg rst = 1'b1;
+    
+    
+    //
+    // Control / Status
+    //
+    reg [ 7:0] word_index_last_n;
+    reg [ 7:0] word_index_last_pq;
+    reg [11:0] bit_index_last_n;
+    reg [11:0] bit_index_last_pq;
+    reg        core_next = 1'b0;
+    wire       core_valid;
+    reg        core_crt_mode;
+    
+    
+    //
+    // System Bus
+    //
+    reg         bus_ready;
+    reg         bus_cs = 1'b0;
+    reg         bus_we = 1'b0;
+    reg  [11:0] bus_addr;
+    reg  [31:0] bus_data_wr;
+    wire [31:0] bus_data_rd;
+
+    wire [ 1:0] bus_addr_sel  = bus_addr[11:10];
+    wire [ 2:0] bus_addr_bank = bus_addr[9:7];
+    wire [ 6:0] bus_addr_data = bus_addr[6:0];
+
+
+    //
+    // UUT
+    //
+    modexpng_core_top uut
+    (
+        .clk                (clk),
+        .clk_bus            (clk_bus),
+        
+        .rst                (rst),
+        
+        .next               (core_next),
+        .valid              (core_valid),
+        
+        .crt_mode           (core_crt_mode),
+        
+        .word_index_last_n  (word_index_last_n),
+        .word_index_last_pq (word_index_last_pq),
+        
+        .bit_index_last_n   (bit_index_last_n),
+        .bit_index_last_pq  (bit_index_last_pq),
+        
+        .bus_cs             (bus_cs),
+        .bus_we             (bus_we),
+        .bus_addr           (bus_addr),
+        .bus_data_wr        (bus_data_wr),
+        .bus_data_rd        (bus_data_rd)
+    );
+
+
+    //
+    // Routine (Bus)
+    //
+    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
+    
+    
+    //
+    // Routine (Control/Status, Bus)
+    //
+    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 = 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;
+
+        wait_clock_ticks(10);
+        get_output;
+        wait_clock_ticks(10);
+        
+        $display("Core output banks read.");
+        
+        verify;
+    
+    end
+    
+    
+    //
+    // Variables
+    //    
+    integer _w, _n;
+    
+
+    //
+    // set_input_1;
+    //
+    task 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]);
+        end
+    endtask
+    
+    
+    //
+    // set_input_2;
+    //
+    task 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]);
+        end
+    endtask
+    
+    
+    //
+    // get_output;
+    //
+    task 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]);
+        end
+    endtask
+
+
+    //
+    // verify;
+    //
+    task verify;
+        //
+        reg xm_ok;
+        reg ym_ok;
+        reg s_ok;
+            //
+        begin
+            //
+            xm_ok = 1;
+            ym_ok = 1;
+            s_ok = 1;
+            //
+            for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) begin
+               if (XM_READBACK[_w] !== XM[_w]) xm_ok = 0; 
+               if (YM_READBACK[_w] !== YM[_w]) ym_ok = 0;
+               if (S_READBACK[_w] !== S[_w]) s_ok = 0;
+            end
+            //
+            if (!xm_ok)
+                //
+                for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) begin
+                    $write("XM / XM_READBACK [%3d] = 0x%08x / 0x%08x", _w, XM[_w], XM_READBACK[_w]);
+                    if (XM[_w] !== XM_READBACK[_w]) $write(" <???: 0x%08x> ", XM[_w] ^ XM_READBACK[_w]);
+                    $write("\n");
+                end
+            //
+            if (!ym_ok)
+                //
+                for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) begin
+                    $write("YM / YM_READBACK [%3d] = 0x%08x / 0x%08x", _w, YM[_w], YM_READBACK[_w]);
+                    if (YM[_w] !== YM_READBACK[_w]) $write(" <???: 0x%08x> ", YM[_w] ^ YM_READBACK[_w]);
+                    $write("\n");
+                end
+            //
+            if (!s_ok)
+                //
+                for (_w=0; _w<TB_NUM_WORDS_N; _w=_w+1) begin
+                    $write("S / S_READBACK [%3d] = 0x%08x / 0x%08x", _w, S[_w], S_READBACK[_w]);
+                    if (S[_w] !== S_READBACK[_w]) $write(" <???: 0x%08x> ", S[_w] ^ S_READBACK[_w]);
+                    $write("\n");
+                end
+            //
+            $write("XM is ");
+            if (xm_ok) $write("OK.\n");
+            else       $write("WRONG!\n");
+            //
+            $write("YM is ");
+            if (ym_ok) $write("OK.\n");
+            else       $write("WRONG!\n");
+            //
+            $write("S is ");
+            if (s_ok) $write("OK.\n");
+            else       $write("WRONG!\n");                        
+            //
+        end
+        //
+    endtask
+    
+    
+
+    //
+    // _bus_drive()
+    //    
+    task _bus_drive;
+        input cs;
+        input we;
+        input [11:0] addr;
+        input [31:0] data;
+        {bus_cs, bus_we, bus_addr, bus_data_wr} <= {cs, we, addr, data};
+    endtask
+    
+    
+    //
+    // bus_write()
+    //
+    task bus_write;
+        input  [ 1:0] sel;
+        input  [ 2:0] bank;
+        input  [ 6:0] addr;
+        input  [31:0] data;
+        begin
+            _bus_drive(1'b1, 1'b1, {sel, bank, addr}, data);
+            wait_clock_bus_tick;
+            _bus_drive(1'b0, 1'b0, 12'hXXX, 32'hXXXXXXXX);
+        end
+    endtask
+    
+    
+    //
+    // bus_read()
+    //
+    task bus_read;
+        input  [ 1:0] sel;
+        input  [ 2:0] bank;
+        input  [ 6:0] addr;
+        output [31:0] data;
+        begin
+            _bus_drive(1'b1, 1'b0, {sel, bank, addr}, 32'hXXXXXXXX);
+            wait_clock_bus_tick;
+            data = bus_data_rd;
+            _bus_drive(1'b0, 1'b0, 12'hXXX, 32'hXXXXXXXX);
+        end
+    endtask
+    
+    
+    //
+    // _wait_half_clock_tick()
+    //
+    task _wait_half_clock_tick;
+        #`CLK_PERIOD_HALF_NS;
+    endtask
+    
+    //
+    // wait_clock_tick()
+    //
+    task wait_clock_tick;
+        begin
+            _wait_half_clock_tick;
+            _wait_half_clock_tick;
+        end
+    endtask
+    
+    
+    //
+    // wait_clock_bus_tick()
+    //
+    task wait_clock_bus_tick;
+        #`CLK_BUS_PERIOD_NS;
+    endtask
+
+
+    //
+    // wait_clock_ticks()
+    //    
+    task wait_clock_ticks;
+        input integer num_ticks;
+        for (_n=0; _n<num_ticks; _n=_n+1)
+            wait_clock_tick;
+    endtask
+    
+    
+    //
+    // wait_clock_bus_ticks()
+    //    
+    task wait_clock_bus_ticks;
+        input integer num_ticks;
+        for (_n=0; _n<num_ticks; _n=_n+1)
+            wait_clock_bus_tick;
+    endtask
+    
+endmodule