module hub75e (
    input clk,
    input write_trig,
    output reg [2:0] panel_rgb0,
    output reg [2:0] panel_rgb1,
    output reg display_clk = 0,
    output reg out_enable = 1,
    output reg latch = 0,
    output reg done = 0,

    // bram interface (using clk)
    output reg [8:0] pixbuf_addr,
    input [35:0] pixbuf_data
);

  parameter ROW_DEPTH = 128, BIT_DEPTH = 8, BCM_LEN = 32;

  reg [31:0] counter = 0;
  // which bit of the colors are we currently exposing.
  reg [$clog2(BIT_DEPTH) - 1:0] bcm_shift = 7;

  localparam StateInit = 0;
  localparam StateWriteRow = 1;
  localparam StateLatchout = 2;
  // the last data that we clock out for the row won't be exposed
  // in the next writerow state because we'll change addresses.
  localparam StateFinishExpose = 3;

  // this state is used to prefetch the first pixel so the cycle can work.
  localparam StatePreload = 4;

  reg [7:0] state = StateInit;  // our state

  // The FSM is a bit confusing since it's optimized for *speed*
  // We can basically display the previous line of data while we write the
  // next one. So instead of having WRITEROW -> LATCH -> EXPOSE
  // like most modules do, we can instead do (WRITEROW + EXPOSE_PREV) -> LATCH
  // There is an edge case for the first/last bits of the color depth.
  // When we start writing a line, we can't flash anything since there's no
  // previous. Likewise, when we end a line, we have to have an extra expose
  // period since there is no next writerow for this address. As a result,
  // we want to go MSB to LSB for our BCM so that the trailing expose time is
  // short!

  wire should_clock, should_expose;
  // the plus 1 is for the falling edge!
  assign should_clock  = (counter < (ROW_DEPTH << 2) + 1);
  assign should_expose = (counter < (BCM_LEN << bcm_shift + 1)) && (bcm_shift != 7);

  wire [7:0] ram_r, ram_g, ram_b;
  assign ram_r = pixbuf_data[23:16];
  assign ram_g = pixbuf_data[15:8];
  assign ram_b = pixbuf_data[7:0];
  wire [2:0] ram_rgb_slice;
  assign ram_rgb_slice = {
    (ram_r[bcm_shift]), (ram_g[bcm_shift]), (ram_b[bcm_shift])
  };


  reg [7:0] pixnum;
  reg pixrow = 0;

  assign pixbuf_addr = {pixrow, pixnum};

  always_ff @(posedge clk) begin
    counter <= counter + 1;
    case (state)
      StateInit: begin
        bcm_shift <= 7;
        counter <= 0;
        done <= 0;
        pixnum <= ROW_DEPTH - 1;
        pixrow <= 0;
        // wait for the signal to write out our lines.
        if (write_trig) begin
          state <= StatePreload;
        end
      end

      StatePreload: begin
        case (counter)
          0: begin
            pixrow <= 0;
          end
          1: begin
            // load pix 1
            panel_rgb0 <= ram_rgb_slice;
            // wait for pix 1
            pixrow <= 1;
          end
          2: begin
            panel_rgb1 <= ram_rgb_slice;
            pixrow <= 0;
            counter <= 0;
            state <= StateWriteRow;
          end
          default: begin
            counter <= 0;
          end
        endcase
      end

      StateWriteRow: begin
        if (should_clock) begin
          // we have data to  clock
          display_clk <= counter[1];
          case (counter[1:0])
            2'b10: begin  // rising edge
              // fetch pixel 1
              pixrow <= 0;
            end
            2'b11: begin  // midpoint of high clk.
              // fetch pixel 2, load pixel 1
              pixrow <= 1;
              panel_rgb0 <= ram_rgb_slice;
            end
            2'b00: begin  // falling edge
              // load pixel 2
            end
            2'b01: begin  // midpoint of low clk
              panel_rgb1 <= ram_rgb_slice;
              // decrement pixnum
              pixnum <= pixnum - 1;
              pixrow <= 0;
            end
            default: begin
            end
          endcase
        end
        if (should_expose) begin
          out_enable <= 0;
        end else begin
          out_enable <= 1;
        end
        // if we're done with our data clock out and also done with exposing
        // the previous line, go to the latchout stage.
        if (~should_clock && ~should_expose) begin
          counter <= 0;
          pixnum <= 0;
          state <= StateLatchout;
          display_clk <= 0;
        end
      end

      StateLatchout: begin
        // raise latch high; compute next bcm.
        latch <= 1;
        out_enable <= 1;
        pixnum <= ROW_DEPTH - 1;
        if (counter > 3) begin
          counter <= 0;
          if (bcm_shift == 0) begin
            // we've reached the lsb of this data, go to the next one!
            state <= StateFinishExpose;
            latch <= 0;
          end else begin
            bcm_shift <= bcm_shift - 1;
            state <= StatePreload;
            latch <= 0;
          end
        end
      end

      StateFinishExpose: begin
        assert (bcm_shift == 0);
        if (counter < (BCM_LEN << bcm_shift)) begin
          out_enable <= 0;
        end else begin
          out_enable <= 1;
          state <= StateInit;
          done <= 1;
          // we are done!
        end
      end
      default: begin
        state <= StateInit;
      end
    endcase
  end

endmodule