// Fixed geometry teaching core: one 4x4 signed-int8 frame after each reset.
// Valid 3x3, stride=1, dilation=1, one channel, no padding.
module cnn_window3x3 (
  input  logic clk, reset,
  input  logic in_valid,
  output logic in_ready,
  input  logic signed [7:0] in_pixel,
  output logic out_valid,
  input  logic out_ready,
  output logic [1:0] out_row, out_col,
  output logic signed [7:0] w00, w01, w02,
  output logic signed [7:0] w10, w11, w12,
  output logic signed [7:0] w20, w21, w22,
  output logic input_done
);
  logic signed [7:0] previous [0:3];
  logic signed [7:0] older [0:3];
  logic signed [7:0] h00, h01, h02;
  logic signed [7:0] h10, h11, h12;
  logic signed [7:0] h20, h21, h22;
  logic [1:0] row, col;
  integer i;

  // A complete unconsumed window blocks all pixel-state advancement.
  assign in_ready = !input_done && (!out_valid || out_ready);

  always_ff @(posedge clk) begin
    if (reset) begin
      row <= 0; col <= 0; input_done <= 0; out_valid <= 0;
      out_row <= 0; out_col <= 0;
      h00 <= 0; h01 <= 0; h02 <= 0;
      h10 <= 0; h11 <= 0; h12 <= 0;
      h20 <= 0; h21 <= 0; h22 <= 0;
      w00 <= 0; w01 <= 0; w02 <= 0;
      w10 <= 0; w11 <= 0; w12 <= 0;
      w20 <= 0; w21 <= 0; w22 <= 0;
      for (i = 0; i < 4; i = i + 1) begin
        previous[i] <= 0;
        older[i] <= 0;
      end
    end else begin
      if (out_valid && out_ready) out_valid <= 0;
      if (in_valid && in_ready) begin
        older[col] <= previous[col];
        previous[col] <= in_pixel;
        if (col == 0) begin
          // New row: old horizontal columns must not enter its first windows.
          h00 <= 0; h01 <= 0; h02 <= older[col];
          h10 <= 0; h11 <= 0; h12 <= previous[col];
          h20 <= 0; h21 <= 0; h22 <= in_pixel;
        end else begin
          h00 <= h01; h01 <= h02; h02 <= older[col];
          h10 <= h11; h11 <= h12; h12 <= previous[col];
          h20 <= h21; h21 <= h22; h22 <= in_pixel;
        end
        if (row >= 2 && col >= 2) begin
          // Nonblocking RHS uses the preceding two horizontal samples.
          out_valid <= 1;
          out_row <= row - 2; out_col <= col - 2;
          w00 <= h01; w01 <= h02; w02 <= older[col];
          w10 <= h11; w11 <= h12; w12 <= previous[col];
          w20 <= h21; w21 <= h22; w22 <= in_pixel;
        end
        if (col == 3) begin
          col <= 0;
          if (row == 3) input_done <= 1;
          else row <= row + 1;
        end else col <= col + 1;
      end
    end
  end
endmodule