path: root/rtl/modexpng_recombinator_block.v

module modexpng_recombinator_block
(
    clk, rst,
    ena, rdy,
    fsm_state_next,
    word_index_last,
    dsp_xy_ce_p,
    dsp_x_p, dsp_y_p,
    col_index, col_index_last,
    rd_narrow_xy_addr, rd_narrow_xy_bank,
    rcmb_wide_xy_bank,   rcmb_wide_xy_addr,   rcmb_wide_x_dout,   rcmb_wide_y_dout,   rcmb_wide_xy_valid,
    rcmb_narrow_xy_bank, rcmb_narrow_xy_addr, rcmb_narrow_x_dout, rcmb_narrow_y_dout, rcmb_narrow_xy_valid,
    rdct_narrow_xy_bank, rdct_narrow_xy_addr, rdct_narrow_x_dout, rdct_narrow_y_dout, rdct_narrow_xy_valid
);


    //
    // Headers
    //
    `include "modexpng_parameters.vh"
    `include "../rtl_1/modexpng_mmm_fsm_old.vh"
    //`include "../rtl_1/modexpng_parameters_x8_old.vh"


    input                        clk;
    input                        rst;
    input                        ena;
    output                       rdy;
    input  [FSM_STATE_WIDTH-1:0] fsm_state_next;
    input [7:0]                  word_index_last;
    input                        dsp_xy_ce_p;
    input  [9*47-1:0] dsp_x_p;
    input  [9*47-1:0] dsp_y_p;
    input  [     4:0] col_index;
    input  [     4:0] col_index_last;
    
    input  [     BANK_ADDR_W -1:0] rd_narrow_xy_bank;
    input  [     7:0] rd_narrow_xy_addr;

    output [     BANK_ADDR_W -1:0] rcmb_wide_xy_bank;
    output [     7:0] rcmb_wide_xy_addr;
    output [    17:0] rcmb_wide_x_dout;
    output [    17:0] rcmb_wide_y_dout;
    output            rcmb_wide_xy_valid;

    output [    BANK_ADDR_W -1:0] rcmb_narrow_xy_bank;
    output [     7:0] rcmb_narrow_xy_addr;
    output [    17:0] rcmb_narrow_x_dout;
    output [    17:0] rcmb_narrow_y_dout;
    output            rcmb_narrow_xy_valid;

    output [     BANK_ADDR_W -1:0] rdct_narrow_xy_bank;
    output [     7:0] rdct_narrow_xy_addr;
    output [    17:0] rdct_narrow_x_dout;
    output [    17:0] rdct_narrow_y_dout;
    output            rdct_narrow_xy_valid;


    //
    // Latches
    //
    reg  [1*47-1:0] dsp_x_p_latch[0:8];
    reg  [1*47-1:0] dsp_y_p_latch[0:8];


    //
    // Mapping
    //
    wire [46:0] dsp_x_p_split[0:8];
    wire [46:0] dsp_y_p_split[0:8];
    
    genvar z;
    generate for (z=0; z<(NUM_MULTS+1); z=z+1)
        begin : gen_dsp_xy_p_split
            assign dsp_x_p_split[z] = dsp_x_p[47*z+:47];
            assign dsp_y_p_split[z] = dsp_y_p[47*z+:47];
        end
    endgenerate


    //
    // Delays
    //
    reg dsp_xy_ce_p_dly1 = 1'b0;

    always @(posedge clk)
        //
        if (rst) dsp_xy_ce_p_dly1 <= 1'b0;
        else     dsp_xy_ce_p_dly1 <= dsp_xy_ce_p;


    //
    // Registers
    //
    
    // valid
    reg       xy_valid_lsb = 1'b0;
    reg       xy_aux_lsb   = 1'b0;
    reg       xy_valid_msb = 1'b0;
    
    // bitmap
    reg [7:0] xy_bitmap_lsb = {8{1'b0}};
    reg [7:0] xy_bitmap_msb = {8{1'b0}};
    
    // index
    reg [2:0] xy_index_lsb = 3'dX;
    
    // purge
    reg       xy_purge_lsb = 1'b0;
    reg       xy_purge_msb = 1'b0;
    
    // valid - latch
    reg       xy_valid_latch_lsb = 1'b0;
    
    // aux - latch
    reg       xy_aux_latch_lsb = 1'b0;
    
    // bitmap - latch
    reg [7:0] xy_bitmap_latch_lsb = {8{1'b0}};
    reg [7:0] xy_bitmap_latch_msb = {8{1'b0}};

    // index - latch
    reg [2:0] xy_index_latch_lsb = 3'dX;
    
    // purge - index
    reg       xy_purge_latch_lsb = 1'b0;
    reg       xy_purge_latch_msb = 1'b0;

    // 
    reg       xy_valid_lsb_adv[1:6];
    reg       xy_valid_msb_adv[1:6];
    reg       xy_aux_lsb_adv[1:6];
    reg [7:0] xy_bitmap_lsb_adv[1:6];
    reg [7:0] xy_bitmap_msb_adv[1:6];
    reg [2:0] xy_index_lsb_adv[1:6];
    reg [2:0] xy_index_msb_adv[1:6];
    reg       xy_purge_lsb_adv[1:6];
    reg       xy_purge_msb_adv[1:6];
    
    reg [1:0] rcmb_mode;
       
    always @(posedge clk)
       //
       if (ena)
           //
           case (fsm_state_next)
               FSM_STATE_MULT_SQUARE_COL_0_BUSY:        rcmb_mode <= 2'd1;
               FSM_STATE_MULT_TRIANGLE_COL_0_BUSY:      rcmb_mode <= 2'd2;
               FSM_STATE_MULT_RECTANGLE_COL_0_BUSY:     rcmb_mode <= 2'd3;
               default:                                 rcmb_mode <= 2'd0;
           endcase

               
    integer i;
    initial for (i=1; i<6; i=i+1) begin
        xy_valid_lsb_adv[i] = 1'b0;
        xy_valid_msb_adv[i] = 1'b0;
        xy_aux_lsb_adv[i] = 1'b0;
        xy_bitmap_lsb_adv[i] = {8{1'b0}};
        xy_bitmap_msb_adv[i] = {8{1'b0}};
        xy_index_lsb_adv[i] = 3'dX;
        xy_index_msb_adv[i] = 3'dX;
        xy_purge_lsb_adv[i] = 1'b0;
        xy_purge_msb_adv[i] = 1'b0;
    end
    
    function        calc_square_triangle_valid_lsb;
        input [4:0] col_index_value;
        input [4:0] col_index_last_value;
        input [BANK_ADDR_W -1:0] narrow_xy_bank_value;
        input [7:0] narrow_xy_addr_value;
        begin
            //
            if (narrow_xy_addr_value[7:3] == col_index_value)
                calc_square_triangle_valid_lsb = 1'b1;
            else
                calc_square_triangle_valid_lsb = 1'b0;
            //
        end
    endfunction

    function        calc_square_valid_lsb;
        input [4:0] col_index_value;
        input [4:0] col_index_last_value;
        input [BANK_ADDR_W -1:0] narrow_xy_bank_value;
        input [7:0] narrow_xy_addr_value;
        begin
            calc_square_valid_lsb = calc_square_triangle_valid_lsb(col_index_value, col_index_last_value, narrow_xy_bank_value, narrow_xy_addr_value);   
        end
    endfunction

    function        calc_triangle_valid_lsb;
        input [4:0] col_index_value;
        input [4:0] col_index_last_value;
        input [BANK_ADDR_W -1:0] narrow_xy_bank_value;
        input [7:0] narrow_xy_addr_value;
        begin
            calc_triangle_valid_lsb = calc_square_triangle_valid_lsb(col_index_value, col_index_last_value, narrow_xy_bank_value, narrow_xy_addr_value);   
        end
    endfunction
    
    function        calc_rectangle_valid_lsb;
        input [4:0] col_index_value;
        input [4:0] col_index_last_value;
        input [BANK_ADDR_W -1:0] narrow_xy_bank_value;
        input [7:0] narrow_xy_addr_value;
        begin
            //
            if (narrow_xy_addr_value[7:3] == col_index_value) 
                calc_rectangle_valid_lsb = narrow_xy_bank_value != BANK_NARROW_EXT;
            else
                calc_rectangle_valid_lsb = 1'b0;
            //
        end
    endfunction

    function        calc_triangle_aux_lsb;
        input [4:0] col_index_value;
        input [4:0] col_index_last_value;
        input [BANK_ADDR_W -1:0] narrow_xy_bank_value;
        input [7:0] narrow_xy_addr_value;
        begin
            //
            if (narrow_xy_bank_value == BANK_NARROW_EXT)
                calc_triangle_aux_lsb = 1'b1;
            else
                calc_triangle_aux_lsb = 1'b0;
            //
        end
    endfunction
    
    function  [7:0] calc_square_triangle_bitmap_lsb;
        input [4:0] col_index_value;
        input [4:0] col_index_last_value;
        input [BANK_ADDR_W -1:0] narrow_xy_bank_value;
        input [7:0] narrow_xy_addr_value;
        begin
            //
            if (narrow_xy_addr_value[7:3] == col_index_value)
                //
                case (narrow_xy_addr_value[2:0])
                    3'b000: calc_square_triangle_bitmap_lsb = 8'b00000001;
                    3'b001: calc_square_triangle_bitmap_lsb = 8'b00000010;
                    3'b010: calc_square_triangle_bitmap_lsb = 8'b00000100;
                    3'b011: calc_square_triangle_bitmap_lsb = 8'b00001000;
                    3'b100: calc_square_triangle_bitmap_lsb = 8'b00010000;
                    3'b101: calc_square_triangle_bitmap_lsb = 8'b00100000;
                    3'b110: calc_square_triangle_bitmap_lsb = 8'b01000000;
                    3'b111: calc_square_triangle_bitmap_lsb = 8'b10000000;
                endcase
                //
            else
                calc_square_triangle_bitmap_lsb = {8{1'b0}};
            //
        end
    endfunction

    function  [7:0] calc_square_bitmap_lsb;
        input [4:0] col_index_value;
        input [4:0] col_index_last_value;
        input [BANK_ADDR_W -1:0] narrow_xy_bank_value;
        input [7:0] narrow_xy_addr_value;
        begin
            calc_square_bitmap_lsb = calc_square_triangle_bitmap_lsb(col_index_value, col_index_last_value, narrow_xy_bank_value, narrow_xy_addr_value);
        end
    endfunction

    function  [7:0] calc_triangle_bitmap_lsb;
        input [4:0] col_index_value;
        input [4:0] col_index_last_value;
        input [BANK_ADDR_W -1:0] narrow_xy_bank_value;
        input [7:0] narrow_xy_addr_value;
        begin
            calc_triangle_bitmap_lsb = calc_square_triangle_bitmap_lsb(col_index_value, col_index_last_value, narrow_xy_bank_value, narrow_xy_addr_value);
        end
    endfunction

    function  [7:0] calc_rectangle_bitmap_lsb;
       input [4:0] col_index_value;
       input [4:0] col_index_last_value;
       input [BANK_ADDR_W -1:0] narrow_xy_bank_value;
       input [7:0] narrow_xy_addr_value;
       begin
           //
           if ((narrow_xy_addr_value[7:3] == col_index_value) && (narrow_xy_bank_value != BANK_NARROW_EXT))
               //
               case (narrow_xy_addr_value[2:0])
                   3'b000: calc_rectangle_bitmap_lsb = 8'b00000001;
                   3'b001: calc_rectangle_bitmap_lsb = 8'b00000010;
                   3'b010: calc_rectangle_bitmap_lsb = 8'b00000100;
                   3'b011: calc_rectangle_bitmap_lsb = 8'b00001000;
                   3'b100: calc_rectangle_bitmap_lsb = 8'b00010000;
                   3'b101: calc_rectangle_bitmap_lsb = 8'b00100000;
                   3'b110: calc_rectangle_bitmap_lsb = 8'b01000000;
                   3'b111: calc_rectangle_bitmap_lsb = 8'b10000000;
               endcase
               //
           else
               calc_rectangle_bitmap_lsb = {8{1'b0}};
           //
        end
    endfunction
       
       /*
        * These can be simplified (the difference between square/triangle and
        * rectangle is that the bank is checked or not). A universal function would
        * accept a parameter that tells it whether it should check the bank or not.
        * Let's do it later, too early to optimize now, it seems.
        *
        *
        */
       
    function  [2:0] calc_square_triangle_index_lsb;
        input [4:0] col_index_value;
        input [4:0] col_index_last_value;
        input [BANK_ADDR_W -1:0] narrow_xy_bank_value;
        input [7:0] narrow_xy_addr_value;
        begin
            //
            if (narrow_xy_addr_value[7:3] == col_index_value)
                //
                case (narrow_xy_addr_value[2:0])
                    3'b000: calc_square_triangle_index_lsb = 3'd0;
                    3'b001: calc_square_triangle_index_lsb = 3'd1;
                    3'b010: calc_square_triangle_index_lsb = 3'd2;
                    3'b011: calc_square_triangle_index_lsb = 3'd3;
                    3'b100: calc_square_triangle_index_lsb = 3'd4;
                    3'b101: calc_square_triangle_index_lsb = 3'd5;
                    3'b110: calc_square_triangle_index_lsb = 3'd6;
                    3'b111: calc_square_triangle_index_lsb = 3'd7;
                endcase
                //
            else
                calc_square_triangle_index_lsb = 3'dX;
            //
        end
    endfunction

    function  [2:0] calc_square_index_lsb;
        input [4:0] col_index_value;
        input [4:0] col_index_last_value;
        input [BANK_ADDR_W -1:0] narrow_xy_bank_value;
        input [7:0] narrow_xy_addr_value;
        begin
            calc_square_index_lsb = calc_square_triangle_index_lsb(col_index_value, col_index_last_value, narrow_xy_bank_value, narrow_xy_addr_value);
        end
    endfunction

    function  [2:0] calc_triangle_index_lsb;
        input [4:0] col_index_value;
        input [4:0] col_index_last_value;
        input [BANK_ADDR_W -1:0] narrow_xy_bank_value;
        input [7:0] narrow_xy_addr_value;
        begin
            calc_triangle_index_lsb = calc_square_triangle_index_lsb(col_index_value, col_index_last_value, narrow_xy_bank_value, narrow_xy_addr_value);
        end
    endfunction

    function  [2:0] calc_rectangle_index_lsb;
        input [4:0] col_index_value;
        input [4:0] col_index_last_value;
        input [BANK_ADDR_W -1:0] slim_bram_xy_bank_value;
        input [7:0] slim_bram_xy_addr_value;
        begin
            //
            if ((slim_bram_xy_addr_value[7:3] == col_index_value) && (slim_bram_xy_bank_value != BANK_NARROW_EXT))
                //
                case (slim_bram_xy_addr_value[2:0])
                    3'b000: calc_rectangle_index_lsb = 3'd0;
                    3'b001: calc_rectangle_index_lsb = 3'd1;
                    3'b010: calc_rectangle_index_lsb = 3'd2;
                    3'b011: calc_rectangle_index_lsb = 3'd3;
                    3'b100: calc_rectangle_index_lsb = 3'd4;
                    3'b101: calc_rectangle_index_lsb = 3'd5;
                    3'b110: calc_rectangle_index_lsb = 3'd6;
                    3'b111: calc_rectangle_index_lsb = 3'd7;
                endcase
                //
            else
                calc_rectangle_index_lsb = 3'dX;
            //
        end
    endfunction
    
    function        calc_square_rectangle_purge_lsb;
        input [4:0] col_index_value;
        input [4:0] col_index_last_value;
        input [BANK_ADDR_W -1:0] narrow_xy_bank_value;
        input [7:0] narrow_xy_addr_value;
        begin
            //
            if (narrow_xy_addr_value[7:3] == col_index_value)
                calc_square_rectangle_purge_lsb = narrow_xy_addr_value[7:3] == col_index_last_value;
            else
                calc_square_rectangle_purge_lsb = 1'b0;
            //
        end
    endfunction

    function        calc_square_purge_lsb;
        input [4:0] col_index_value;
        input [4:0] col_index_last_value;
        input [BANK_ADDR_W -1:0] narrow_xy_bank_value;
        input [7:0] narrow_xy_addr_value;
        begin
            calc_square_purge_lsb = calc_square_rectangle_purge_lsb(col_index_value, col_index_last_value, narrow_xy_bank_value, narrow_xy_addr_value);
        end
    endfunction

    function        calc_rectangle_purge_lsb;
        input [4:0] col_index_value;
        input [4:0] col_index_last_value;
        input [BANK_ADDR_W -1:0] narrow_xy_bank_value;
        input [7:0] narrow_xy_addr_value;
        begin
            calc_rectangle_purge_lsb = calc_square_rectangle_purge_lsb(col_index_value, col_index_last_value, narrow_xy_bank_value, narrow_xy_addr_value);
        end
    endfunction

    function        calc_square_valid_msb;
        input [4:0] col_index_value;
        input [4:0] col_index_last_value;
        input [BANK_ADDR_W -1:0] narrow_xy_bank_value;
        input [7:0] narrow_xy_addr_value;
        input [7:0] index_last_value;
        begin
            //
            if (narrow_xy_addr_value == index_last_value)
                calc_square_valid_msb = 1'b1;
            else
                calc_square_valid_msb = 1'b0;
            //
        end
    endfunction

    function        calc_rectangle_valid_msb;
        input [4:0] col_index_value;
        input [4:0] col_index_last_value;
        input [BANK_ADDR_W -1:0] narrow_xy_bank_value;
        input [7:0] narrow_xy_addr_value;
        input [7:0] index_last_value;
        begin
            //
            if ((narrow_xy_addr_value == 8'd1) && (narrow_xy_bank_value == BANK_NARROW_EXT))
                calc_rectangle_valid_msb = 1'b1;
            else
                calc_rectangle_valid_msb = 1'b0;
            //
        end
    endfunction
    
    function  [7:0] calc_square_bitmap_msb;
        input [4:0] col_index_value;
        input [4:0] col_index_last_value;
        input [BANK_ADDR_W -1:0] narrow_xy_bank_value;
        input [7:0] narrow_xy_addr_value;
        input [7:0] index_last_value;
        begin
            //
            if (narrow_xy_addr_value == index_last_value) begin
                calc_square_bitmap_msb[7] = col_index_value != col_index_last_value;
                calc_square_bitmap_msb[6:0] = 7'b1111111;
            end else
                calc_square_bitmap_msb[7:0] = 8'b00000000;
            //
        end
    endfunction

    function  [7:0] calc_rectangle_bitmap_msb;
        input [4:0] col_index_value;
        input [4:0] col_index_last_value;
        input [BANK_ADDR_W -1:0] narrow_xy_bank_value;
        input [7:0] narrow_xy_addr_value;
        input [7:0] index_last_value;
        begin
            //
            if ((narrow_xy_addr_value == 8'd1) && (narrow_xy_bank_value == BANK_NARROW_EXT)) begin
                calc_rectangle_bitmap_msb[7:0] = 8'b11111111;
            end else
                calc_rectangle_bitmap_msb[7:0] = 8'b00000000;
            //
        end
    endfunction

    function        calc_square_purge_msb;
        input [4:0] col_index_value;
        input [4:0] col_index_last_value;
        input [BANK_ADDR_W -1:0] narrow_xy_bank_value;
        input [7:0] narrow_xy_addr_value;
        input [7:0] index_last_value;
        begin
            //
            if (narrow_xy_addr_value == index_last_value)
                calc_square_purge_msb = col_index_value == col_index_last_value;
            else
                calc_square_purge_msb = 1'b0;
            //
        end
    endfunction

    function        calc_rectangle_purge_msb;
        input [4:0] col_index_value;
        input [4:0] col_index_last_value;
        input [BANK_ADDR_W -1:0] narrow_xy_bank_value;
        input [7:0] narrow_xy_addr_value;
        input [7:0] index_last_value;
        begin
            //
            if ((narrow_xy_addr_value == 8'd1) && (narrow_xy_bank_value == BANK_NARROW_EXT))
                calc_rectangle_purge_msb = col_index_value == col_index_last_value;
            else
                calc_rectangle_purge_msb = 1'b0;
            //
        end
    endfunction

    
    reg         rcmb_xy_lsb_ce = 1'b0;
    reg         rcmb_xy_lsb_ce_aux;
    reg  [ 2:0] rcmb_xy_lsb_ce_purge = 3'b000;
    wire        rcmb_xy_lsb_ce_combined = rcmb_xy_lsb_ce | rcmb_xy_lsb_ce_aux | rcmb_xy_lsb_ce_purge[0];
    reg         rcmb_xy_lsb_clr;

    reg  [46:0] rcmb_x_lsb_din;
    reg  [46:0] rcmb_y_lsb_din;
    wire [15:0] rcmb_x_lsb_dout;
    wire [15:0] rcmb_y_lsb_dout;

    reg         rcmb_xy_msb_ce = 1'b0;
    reg  [ 1:0] rcmb_xy_msb_ce_purge = 2'b00;
    wire        rcmb_xy_msb_ce_combined = rcmb_xy_msb_ce | rcmb_xy_msb_ce_purge[0];
    reg         rcmb_xy_msb_clr;
    
    reg  [46:0] rcmb_x_msb_din;
    reg  [46:0] rcmb_y_msb_din;
    wire [15:0] rcmb_x_msb_dout;
    wire [15:0] rcmb_y_msb_dout;
    
    modexpng_recombinator_cell recomb_x_lsb
    (
        .clk    (clk),
        .ce     (rcmb_xy_lsb_ce_combined),
        .clr    (rcmb_xy_lsb_clr),
        .din    (rcmb_x_lsb_din),
        .dout   (rcmb_x_lsb_dout)
    );
    modexpng_recombinator_cell recomb_y_lsb
    (
        .clk    (clk),
        .ce     (rcmb_xy_lsb_ce_combined),
        .clr    (rcmb_xy_lsb_clr),
        .din    (rcmb_y_lsb_din),
        .dout   (rcmb_y_lsb_dout)
    );

    modexpng_recombinator_cell recomb_x_msb
    (
        .clk    (clk),
        .ce     (rcmb_xy_msb_ce_combined),
        .clr    (rcmb_xy_msb_clr),
        .din    (rcmb_x_msb_din),
        .dout   (rcmb_x_msb_dout)
    );
    
    modexpng_recombinator_cell recomb_y_msb
    (
        .clk    (clk),
        .ce     (rcmb_xy_msb_ce_combined),
        .clr    (rcmb_xy_msb_clr),
        .din    (rcmb_y_msb_din),
        .dout   (rcmb_y_msb_dout)
    );

    always @(posedge clk) begin
        //
        rcmb_xy_lsb_ce <= xy_valid_latch_lsb;
        rcmb_xy_lsb_ce_aux <= xy_aux_latch_lsb;
        rcmb_xy_msb_ce <= xy_bitmap_latch_msb[0];
        //
        if (xy_purge_latch_lsb)
            rcmb_xy_lsb_ce_purge <= 3'b111;
        else
            rcmb_xy_lsb_ce_purge <= {1'b0, rcmb_xy_lsb_ce_purge[2:1]};
        //
        if (xy_purge_latch_msb && xy_bitmap_latch_msb[0] && !xy_bitmap_latch_msb[1])
            rcmb_xy_msb_ce_purge = 2'b11;
        else
            rcmb_xy_msb_ce_purge <= {1'b0, rcmb_xy_msb_ce_purge[1]};
        //
    end


    always @(posedge clk)
        //
        if (ena) begin
            rcmb_xy_lsb_clr <= 1'b1;
            rcmb_xy_msb_clr <= 1'b1;
        end else begin
            if (rcmb_xy_lsb_ce) rcmb_xy_lsb_clr <= 1'b0;
            if (rcmb_xy_msb_ce) rcmb_xy_msb_clr <= 1'b0;
        end

    always @(posedge clk)
        //
        if (xy_valid_latch_lsb) begin
            rcmb_x_lsb_din <= dsp_x_p_latch[xy_index_latch_lsb];
            rcmb_y_lsb_din <= dsp_y_p_latch[xy_index_latch_lsb];
        end else if (xy_aux_latch_lsb) begin
            rcmb_x_lsb_din <= dsp_x_p_latch[8];
            rcmb_y_lsb_din <= dsp_y_p_latch[8];
        end else begin
            rcmb_x_lsb_din <= {47{1'b0}};
            rcmb_y_lsb_din <= {47{1'b0}};
        end

    always @(posedge clk)
        //
        if (xy_bitmap_latch_msb[0]) begin
            rcmb_x_msb_din <= dsp_x_p_latch[0];
            rcmb_y_msb_din <= dsp_y_p_latch[0];
        end else begin
            rcmb_x_msb_din <= {47{1'b0}};
            rcmb_y_msb_din <= {47{1'b0}};
        end


    always @(posedge clk)
        //
        case (fsm_state_next)
            //
            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: begin
                //
                xy_valid_lsb_adv [6] <= calc_square_valid_lsb (col_index, col_index_last, rd_narrow_xy_bank, rd_narrow_xy_addr);
                xy_aux_lsb_adv   [6] <= 1'b0;
                xy_bitmap_lsb_adv[6] <= calc_square_bitmap_lsb(col_index, col_index_last, rd_narrow_xy_bank, rd_narrow_xy_addr);
                xy_index_lsb_adv [6] <= calc_square_index_lsb (col_index, col_index_last, rd_narrow_xy_bank, rd_narrow_xy_addr);
                xy_purge_lsb_adv [6] <= calc_square_purge_lsb (col_index, col_index_last, rd_narrow_xy_bank, rd_narrow_xy_addr);
                //
                xy_valid_msb_adv [6] <= calc_square_valid_msb (col_index, col_index_last, rd_narrow_xy_bank, rd_narrow_xy_addr, word_index_last);
                xy_bitmap_msb_adv[6] <= calc_square_bitmap_msb(col_index, col_index_last, rd_narrow_xy_bank, rd_narrow_xy_addr, word_index_last);
                xy_purge_msb_adv [6] <= calc_square_purge_msb (col_index, col_index_last, rd_narrow_xy_bank, rd_narrow_xy_addr, word_index_last);
                //
            end
            //
            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: begin
                //
                xy_valid_lsb_adv [6] <= calc_triangle_valid_lsb (col_index, col_index_last, rd_narrow_xy_bank, rd_narrow_xy_addr);
                xy_aux_lsb_adv   [6] <= calc_triangle_aux_lsb   (col_index, col_index_last, rd_narrow_xy_bank, rd_narrow_xy_addr);
                xy_bitmap_lsb_adv[6] <= calc_triangle_bitmap_lsb(col_index, col_index_last, rd_narrow_xy_bank, rd_narrow_xy_addr);
                xy_index_lsb_adv [6] <= calc_triangle_index_lsb (col_index, col_index_last, rd_narrow_xy_bank, rd_narrow_xy_addr);
                xy_purge_lsb_adv [6] <= 1'b0;
                //
                xy_valid_msb_adv [6] <= 1'b0;
                xy_bitmap_msb_adv[6] <= {8{1'b0}};
                xy_purge_msb_adv [6] <= 1'b0;
                //
            end
            //
            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: begin
                //
                xy_valid_lsb_adv [6] <= calc_rectangle_valid_lsb (col_index, col_index_last, rd_narrow_xy_bank, rd_narrow_xy_addr);
                xy_aux_lsb_adv   [6] <= 1'b0;
                xy_bitmap_lsb_adv[6] <= calc_rectangle_bitmap_lsb(col_index, col_index_last, rd_narrow_xy_bank, rd_narrow_xy_addr);
                xy_index_lsb_adv [6] <= calc_rectangle_index_lsb (col_index, col_index_last, rd_narrow_xy_bank, rd_narrow_xy_addr);
                xy_purge_lsb_adv [6] <= calc_rectangle_purge_lsb (col_index, col_index_last, rd_narrow_xy_bank, rd_narrow_xy_addr);
                //
                xy_valid_msb_adv [6] <= calc_rectangle_valid_msb (col_index, col_index_last, rd_narrow_xy_bank, rd_narrow_xy_addr, word_index_last);
                xy_bitmap_msb_adv[6] <= calc_rectangle_bitmap_msb(col_index, col_index_last, rd_narrow_xy_bank, rd_narrow_xy_addr, word_index_last);
                xy_purge_msb_adv [6] <= calc_rectangle_purge_msb (col_index, col_index_last, rd_narrow_xy_bank, rd_narrow_xy_addr, word_index_last);
                //
            end
            //
            default: begin
                //
                xy_valid_lsb_adv [6] <= 1'b0;
                xy_aux_lsb_adv   [6] <= 1'b0;
                xy_bitmap_lsb_adv[6] <= {8{1'b0}};
                xy_index_lsb_adv [6] <= 3'dX;
                xy_purge_lsb_adv [6] <= 1'b0;
                //
                xy_valid_msb_adv [6] <= 1'b0;
                xy_bitmap_msb_adv[6] <= {8{1'b0}};
                xy_purge_msb_adv [6] <= 1'b0;
                //
            end
            //
        endcase


    always @(posedge clk) begin
        //
        xy_valid_lsb  <= xy_valid_lsb_adv [1];
        xy_aux_lsb    <= xy_aux_lsb_adv   [1];
        xy_bitmap_lsb <= xy_bitmap_lsb_adv[1];
        xy_index_lsb  <= xy_index_lsb_adv [1];
        xy_purge_lsb  <= xy_purge_lsb_adv [1];
        //
        xy_valid_latch_lsb  <= xy_valid_lsb;
        xy_aux_latch_lsb    <= xy_aux_lsb;
        xy_bitmap_latch_lsb <= xy_bitmap_lsb;
        xy_index_latch_lsb  <= xy_index_lsb;
        xy_purge_latch_lsb  <= xy_purge_lsb;
        //
        xy_valid_msb  <= xy_valid_msb_adv[1];
        xy_bitmap_msb <= xy_bitmap_msb_adv[1];
        xy_purge_msb  <= xy_purge_msb_adv[1];
        //
        if (xy_valid_msb) begin
            xy_bitmap_latch_msb <= xy_bitmap_msb;
            xy_purge_latch_msb  <= xy_purge_msb;
        end else begin
            xy_bitmap_latch_msb <= {1'b0, xy_bitmap_latch_msb[7:1]};
        end
        //
        //
        for (i=1; i<6; i=i+1) begin
            xy_valid_lsb_adv [i] <= xy_valid_lsb_adv [i+1];
            xy_aux_lsb_adv   [i] <= xy_aux_lsb_adv   [i+1];
            xy_bitmap_lsb_adv[i] <= xy_bitmap_lsb_adv[i+1];
            xy_index_lsb_adv [i] <= xy_index_lsb_adv [i+1];
            xy_purge_lsb_adv [i] <= xy_purge_lsb_adv [i+1];
            //
            xy_valid_msb_adv [i] <= xy_valid_msb_adv [i+1];
            xy_bitmap_msb_adv[i] <= xy_bitmap_msb_adv[i+1];
            xy_purge_msb_adv [i] <= xy_purge_msb_adv [i+1];
        end
        //
    end

    always @(posedge clk)
        //
        if (xy_bitmap_latch_msb[1])   // only shift 7 times
            //
            for (i=0; i<8; i=i+1)
                //            
                if (i < 7) begin
                    dsp_x_p_latch[i] <= dsp_x_p_latch[i+1];
                    dsp_y_p_latch[i] <= dsp_y_p_latch[i+1];
                end else begin
                    dsp_x_p_latch[i] <= {47{1'bX}};
                    dsp_y_p_latch[i] <= {47{1'bX}};
                end
            //
        else if (dsp_xy_ce_p_dly1) begin
            //
            for (i=0; i<8; i=i+1)
                //
                if (xy_bitmap_lsb[i]) begin
                    dsp_x_p_latch[i] <= dsp_x_p_split[i];
                    dsp_y_p_latch[i] <= dsp_y_p_split[i];
                end else if (xy_valid_msb && xy_bitmap_msb[i]) begin
                    dsp_x_p_latch[i] <= dsp_x_p_split[i];
                    dsp_y_p_latch[i] <= dsp_y_p_split[i];
                end
            //
            if (xy_aux_lsb) begin
                dsp_x_p_latch[8] <= dsp_x_p_split[8];
                dsp_y_p_latch[8] <= dsp_y_p_split[8];
            end
            //
        end

    reg rcmb_xy_lsb_valid = 1'b0;
    reg rcmb_xy_msb_valid = 1'b0;

    always @(posedge clk)
        //
        if (rst) begin
            rcmb_xy_lsb_valid <= 1'b0;
            rcmb_xy_msb_valid <= 1'b0;        
        end else begin
            rcmb_xy_lsb_valid <= rcmb_xy_lsb_ce_combined;
            rcmb_xy_msb_valid <= rcmb_xy_msb_ce_combined;
        end        


    reg [ BANK_ADDR_W -1:0] wide_xy_bank;
    reg [ 7:0] wide_xy_addr;
    reg [17:0] wide_x_dout;
    reg [17:0] wide_y_dout;
    reg        wide_xy_valid = 1'b0;

    reg [ BANK_ADDR_W -1:0] narrow_xy_bank;
    reg [ 7:0] narrow_xy_addr;
    reg [17:0] narrow_x_dout;
    reg [17:0] narrow_y_dout;
    reg        narrow_xy_valid = 1'b0;

    reg [ BANK_ADDR_W -1:0] rdct_xy_bank;
    reg [ 7:0] rdct_xy_addr;
    reg [17:0] rdct_x_dout;
    reg [17:0] rdct_y_dout;
    reg        rdct_xy_valid = 1'b0;

    reg [ 7:0] cnt_lsb;
    reg [ 7:0] cnt_msb;
    
    reg        cnt_lsb_wrapped;
    reg        cnt_msb_wrapped;

    reg [31:0] rcmb_xy_msb_carry_0;
    reg [31:0] rcmb_xy_msb_carry_1;
    
    reg [31:0] rcmb_xy_msb_delay_0;
    reg [31:0] rcmb_xy_msb_delay_1;
    reg [31:0] rcmb_xy_msb_delay_2;
    
    reg [ 7:0] rcmb_msb_cnt_delay_0 = 8'd0;
    reg [ 7:0] rcmb_msb_cnt_delay_1 = 8'd0;
    reg [ 7:0] rcmb_msb_cnt_delay_2 = 8'd0;

    reg        rcmb_msb_flag_delay_0;
    reg        rcmb_msb_flag_delay_1;
    reg        rcmb_msb_flag_delay_2;
    
    assign rcmb_wide_xy_bank  = wide_xy_bank;
    assign rcmb_wide_xy_addr  = wide_xy_addr;
    assign rcmb_wide_x_dout   = wide_x_dout;
    assign rcmb_wide_y_dout   = wide_y_dout;
    assign rcmb_wide_xy_valid = wide_xy_valid;

    assign rcmb_narrow_xy_bank  = narrow_xy_bank;
    assign rcmb_narrow_xy_addr  = narrow_xy_addr;
    assign rcmb_narrow_x_dout   = narrow_x_dout;
    assign rcmb_narrow_y_dout   = narrow_y_dout;
    assign rcmb_narrow_xy_valid = narrow_xy_valid;

    assign rdct_narrow_xy_bank  = rdct_xy_bank;
    assign rdct_narrow_xy_addr  = rdct_xy_addr;
    assign rdct_narrow_x_dout   = rdct_x_dout;
    assign rdct_narrow_y_dout   = rdct_y_dout;
    assign rdct_narrow_xy_valid = rdct_xy_valid;
    
    reg rdy_reg = 1'b1;
    reg rdy_adv = 1'b1;
    
    assign rdy = rdy_reg;
    
    
    always @(posedge clk)
        //
        if (ena) rdy_reg <= 1'b0;
        else     rdy_reg <= rdy_adv;
            
    task advance_rcmb_msb_delay;
        input [15:0] dout_x;
        input [15:0] dout_y;
        input [ 7:0] cnt;
        input        flag;
        begin
            //
            rcmb_xy_msb_delay_0 <= {dout_y, dout_x};
            rcmb_xy_msb_delay_1 <= rcmb_xy_msb_delay_0;
            rcmb_xy_msb_delay_2 <= rcmb_xy_msb_delay_1;
            //
            rcmb_msb_cnt_delay_0 <= cnt;
            rcmb_msb_cnt_delay_1 <= rcmb_msb_cnt_delay_0;
            rcmb_msb_cnt_delay_2 <= rcmb_msb_cnt_delay_1;
            //
            rcmb_msb_flag_delay_0 <= flag;
            rcmb_msb_flag_delay_1 <= rcmb_msb_flag_delay_0;
            rcmb_msb_flag_delay_2 <= rcmb_msb_flag_delay_1;
            //
        end
    endtask
         
    task shift_rcmb_msb_carry;
        input [15:0] dout_x;
        input [15:0] dout_y;
        begin
            rcmb_xy_msb_carry_0 <= {dout_y, dout_x};
            rcmb_xy_msb_carry_1 <= rcmb_xy_msb_carry_0;
        end
    endtask
    
    task _update_wide;
        input [ BANK_ADDR_W -1:0] bank;
        input [ 7:0] addr;
        input [17:0] dout_x;
        input [17:0] dout_y;
        input        valid;
        begin
            wide_xy_bank       <= bank;
            wide_xy_addr       <= addr;
            wide_x_dout        <= dout_x;
            wide_y_dout        <= dout_y;
            wide_xy_valid <= valid;
        end
    endtask
    
    task _update_narrow;
        input [ BANK_ADDR_W -1:0] bank;
        input [ 7:0] addr;
        input [17:0] dout_x;
        input [17:0] dout_y;
        input        valid;
        begin
            narrow_xy_bank       <= bank;
            narrow_xy_addr       <= addr;
            narrow_x_dout        <= dout_x;
            narrow_y_dout        <= dout_y;
            narrow_xy_valid <= valid;
        end
    endtask

    task _update_rdct;
        input [ BANK_ADDR_W -1:0] bank;
        input [ 7:0] addr;
        input [17:0] dout_x;
        input [17:0] dout_y;
        input        valid;
        begin
            rdct_xy_bank       <= bank;
            rdct_xy_addr       <= addr;
            rdct_x_dout        <= dout_x;
            rdct_y_dout        <= dout_y;
            rdct_xy_valid <= valid;
        end
    endtask
            
    task set_wide;
        input [ BANK_ADDR_W -1:0] bank;
        input [ 7:0] addr;
        input [17:0] dout_x;
        input [17:0] dout_y;
        begin
            _update_wide(bank, addr, dout_x, dout_y, 1'b1);
        end
    endtask
    
    task set_narrow;
        input [ BANK_ADDR_W -1:0] bank;
        input [ 7:0] addr;
        input [17:0] dout_x;
        input [17:0] dout_y;
        begin
            _update_narrow(bank, addr, dout_x, dout_y, 1'b1);
        end
    endtask
    
    task set_rdct;
        input [ BANK_ADDR_W -1:0] bank;
        input [ 7:0] addr;
        input [17:0] dout_x;
        input [17:0] dout_y;
        begin
            _update_rdct(bank, addr, dout_x, dout_y, 1'b1);
        end
    endtask
    
    task clear_wide;
        begin
            _update_wide(BANK_DNC, OP_ADDR_DNC, WORD_EXT_DNC, WORD_EXT_DNC, 1'b0);
        end
    endtask

    task clear_narrow;
        begin
            _update_narrow(BANK_DNC, OP_ADDR_DNC, WORD_EXT_DNC, WORD_EXT_DNC, 1'b0);
        end
    endtask

    task clear_rdct;
        begin
            _update_rdct(BANK_DNC, OP_ADDR_DNC, WORD_EXT_DNC, WORD_EXT_DNC, 1'b0);
        end
    endtask
    
    task _set_cnt_lsb;
        input [7:0] cnt;
        input       wrapped;
        begin
            cnt_lsb <= cnt;
            cnt_lsb_wrapped <= wrapped;
        end
    endtask
    
    task _set_cnt_msb;
        input [7:0] cnt;
        input       wrapped;
        begin
            cnt_msb <= cnt;
            cnt_msb_wrapped <= wrapped;
        end
    endtask    

    task inc_cnt_lsb;
        begin
            if (cnt_lsb == word_index_last)
                _set_cnt_lsb(8'd0, 1'b1);
            else
                _set_cnt_lsb(cnt_lsb + 1'b1, cnt_lsb_wrapped);
        end
    endtask
    
    task inc_cnt_both;
        begin
            inc_cnt_lsb;
            inc_cnt_msb;
        end
    endtask
    
    task inc_cnt_msb;
        begin
            if (cnt_msb == word_index_last)
                _set_cnt_msb(8'd0, 1'b1);
            else
                _set_cnt_msb(cnt_msb + 1'b1, cnt_msb_wrapped);
        end
    endtask
    
    task clr_cnt_lsb;
        begin
            _set_cnt_lsb(8'd0, 1'b0);
        end
    endtask
    
    task clr_cnt_msb;
        begin
            _set_cnt_msb(8'd0, 1'b0);
        end
    endtask
    
   

    wire [1:0] rcmb_xy_valid = {rcmb_xy_msb_valid, rcmb_xy_lsb_valid}; 
    
    always @(posedge clk)
        //
        if (ena) begin
            clr_cnt_lsb();
            clr_cnt_msb();
        end else if (!rdy)
            //
            case (rcmb_mode)
                2'd1: recombine_square();
                2'd2: recombine_triangle();
                2'd3: recombine_rectangle();
            endcase
           
    wire [17:0] rcmb_x_lsb_dout_pad = {2'b00, rcmb_x_lsb_dout};
    wire [17:0] rcmb_y_lsb_dout_pad = {2'b00, rcmb_y_lsb_dout};

    wire [17:0] rcmb_x_msb_dout_pad = {2'b00, rcmb_x_msb_dout};
    wire [17:0] rcmb_y_msb_dout_pad = {2'b00, rcmb_y_msb_dout};
    
    wire [17:0] rcmb_x_msb_delay_2_pad = {2'b00, rcmb_xy_msb_delay_2[15: 0]};
    wire [17:0] rcmb_y_msb_delay_2_pad = {2'b00, rcmb_xy_msb_delay_2[31:16]};

    wire [17:0] rcmb_x_lsb_msb_carry_1_pad = {1'b0, {1'b0, rcmb_x_lsb_dout} + {1'b0, rcmb_xy_msb_carry_1[15: 0]}};
    wire [17:0] rcmb_y_lsb_msb_carry_1_pad = {1'b0, {1'b0, rcmb_y_lsb_dout} + {1'b0, rcmb_xy_msb_carry_1[31:16]}};
           
           
    task recombine_square;
        //
        begin
            //
            case (rcmb_xy_valid)
                //
                2'b01: inc_cnt_lsb; 
                2'b10: inc_cnt_msb;
                2'b11: inc_cnt_both;
                //
            endcase            
            //
            case (rcmb_xy_valid)
                //
                2'b00: if (rcmb_msb_flag_delay_2) set_wide(BANK_WIDE_H, rcmb_msb_cnt_delay_2, rcmb_x_msb_delay_2_pad, rcmb_y_msb_delay_2_pad);
                       else                       clear_wide;
                //
                2'b01:                            set_wide(BANK_WIDE_L, cnt_lsb, rcmb_x_lsb_dout_pad, rcmb_y_lsb_dout_pad);
                //
                2'b10: if (cnt_msb < 8'd2)        clear_wide;                        
                       else                       set_wide(BANK_WIDE_H, cnt_msb, rcmb_x_msb_dout_pad, rcmb_y_msb_dout_pad);
                //
                2'b11: if (cnt_lsb_wrapped)       set_wide(BANK_WIDE_H, cnt_lsb, rcmb_x_lsb_msb_carry_1_pad, rcmb_y_lsb_msb_carry_1_pad); 
                       else                       set_wide(BANK_WIDE_L, cnt_lsb, rcmb_x_lsb_dout_pad,        rcmb_y_lsb_dout_pad);
                //
            endcase            
            //
            case (rcmb_xy_valid)
                //
                2'b00: if (rcmb_msb_flag_delay_2) advance_rcmb_msb_delay(16'hXXXX, 16'hXXXX, 8'd0, 1'b0);
                2'b10: if (cnt_msb < 8'd2)        shift_rcmb_msb_carry(rcmb_x_msb_dout, rcmb_y_msb_dout);
                //
                2'b11: begin                      advance_rcmb_msb_delay(rcmb_x_msb_dout, rcmb_y_msb_dout, cnt_msb, 1'b1);
                       if (cnt_lsb_wrapped)       shift_rcmb_msb_carry({16{1'bX}}, {16{1'bX}});
                       end
                //
            endcase
            //        
        end
        //
    endtask
    
    
    task recombine_triangle;
        //
        begin
            //
            case (rcmb_xy_valid)
                //
                2'b01: inc_cnt_lsb(); 
               //
            endcase            
            //
            case (rcmb_xy_valid)
                //
                2'b00:                        clear_narrow;
                2'b01:  if (!cnt_lsb_wrapped) set_narrow(BANK_NARROW_Q,   cnt_lsb, rcmb_x_lsb_dout_pad, rcmb_y_lsb_dout_pad); 
                        else                  set_narrow(BANK_NARROW_EXT, 8'd1,    rcmb_x_lsb_dout_pad, rcmb_y_lsb_dout_pad);
                2'b10:                        clear_narrow;
                2'b11:                        clear_narrow;
                //
            endcase
            //        
        end
        //
    endtask


    task recombine_rectangle;
        //
        begin
            //
            case (rcmb_xy_valid)
                //
                2'b01:  inc_cnt_lsb; 
                2'b10:  inc_cnt_msb;
                2'b11:  inc_cnt_both;
                //
            endcase
//            //
            case (rcmb_xy_valid)
//                //
                2'b00:  if (rcmb_msb_flag_delay_2)  set_rdct(BANK_RCMB_MH, rcmb_msb_cnt_delay_2, rcmb_x_msb_delay_2_pad, rcmb_y_msb_delay_2_pad);
                        else                        clear_rdct;
                2'b01:                              set_rdct(BANK_RCMB_ML, cnt_lsb, rcmb_x_lsb_dout_pad, rcmb_y_lsb_dout_pad); 
                2'b10:  if (!cnt_msb_wrapped) begin 
                            if (cnt_msb < 8'd2)     clear_rdct;                        
                            else                    set_rdct(BANK_RCMB_MH, cnt_msb, rcmb_x_msb_dout_pad, rcmb_y_msb_dout_pad);
                        end else                    set_rdct(BANK_RCMB_EXT, 8'd0, rcmb_x_msb_dout_pad, rcmb_y_msb_dout_pad);
                            
                2'b11:  set_rdct(BANK_RCMB_MH, cnt_lsb, rcmb_x_lsb_msb_carry_1_pad, rcmb_y_lsb_msb_carry_1_pad); 
//                //
            endcase            
//            //
            case (rcmb_xy_valid)
//                //
                2'b00:  if (rcmb_msb_flag_delay_2)  advance_rcmb_msb_delay(16'hXXXX, 16'hXXXX, 8'd0, 1'b0);
                2'b10:  begin 
                            if ((cnt_msb < 8'd2) && !cnt_msb_wrapped) shift_rcmb_msb_carry(rcmb_x_msb_dout, rcmb_y_msb_dout);
                            if (cnt_msb_wrapped) advance_rcmb_msb_delay(16'hXXXX, 16'hXXXX, 8'd0, 1'b0);
                        end
//                //
                2'b11:  begin  advance_rcmb_msb_delay(rcmb_x_msb_dout, rcmb_y_msb_dout, cnt_msb, 1'b1);
                                   shift_rcmb_msb_carry({16{1'bX}}, {16{1'bX}});
                        end
//                //
            endcase
            //
        end
        //
    endtask
    
    
    always @(posedge clk)
        //
        if (ena) begin
            rdy_adv <= 1'b0;
        end else if (!rdy_reg) begin
            //
            case (rcmb_mode)
                //
                2'd1:   case (rcmb_xy_valid)
                            //
                            2'b00: begin
                                //
                                if (rcmb_msb_flag_delay_2) begin
                                    //
                                    rdy_adv <= ~rcmb_msb_flag_delay_1;
                                    //
                                end
                                //
                            end
                            //
                        endcase
                //
                2'd2:   case (rcmb_xy_valid)
                            //
                            2'b01: rdy_adv <= cnt_lsb_wrapped;                                //
                            //
                        endcase
                //
                2'd3: case (rcmb_xy_valid)
                                            //
                                            2'b00: begin
                                                //
                                                if (rcmb_msb_flag_delay_2) begin
                                                    //
                                                    rdy_adv <= ~rcmb_msb_flag_delay_1;
                                                    //
                                                end
                                                //
                                            end
                                            //
                                        endcase
                //
            endcase
            //        
        end
        
endmodule