`timescale 1ns/1ps

module dct(clk, rst, in, out);

   parameter g_vec_size = 8;
   parameter g_data_width = 16;

   input     clk;
   input     rst;
   input [g_vec_size*g_data_width-1:0] in;
   output [g_vec_size*g_data_width-1:0] out;

   wire [g_data_width-1:0]              in_vec[0:g_vec_size-1];
   wire [g_data_width-1:0]              out_vec[0:g_vec_size-1];

   reg [g_data_width-1:0]               matrix[0:g_vec_size-1][0:g_vec_size-1];
   reg [g_data_width-1:0]               out_reg[0:g_vec_size-1];

   initial begin
      matrix[0][0] = 23170;
      matrix[0][1] = 23170;
      matrix[0][2] = 23170;
      matrix[0][3] = 23170;
      matrix[0][4] = 23170;
      matrix[0][5] = 23170;
      matrix[0][6] = 23170;
      matrix[0][7] = 23170;

      matrix[1][0] = 32138;
      matrix[1][1] = 27245;
      matrix[1][2] = 18204;
      matrix[1][3] = 6392;
      matrix[1][4] = -6392;
      matrix[1][5] = -18204;
      matrix[1][6] = -27245;
      matrix[1][7] = -32138;

      matrix[2][0] = 30273;
      matrix[2][1] = 12539;
      matrix[2][2] = -12539;
      matrix[2][3] = -30273;
      matrix[2][4] = -30273;
      matrix[2][5] = -12539;
      matrix[2][6] = -12539;
      matrix[2][7] = 30273;

      matrix[3][0] = 27245;
      matrix[3][1] = -6392;
      matrix[3][2] = -32138;
      matrix[3][3] = -18204;
      matrix[3][4] = 18204;
      matrix[3][5] = -32138;
      matrix[3][6] = 6392;
      matrix[3][7] = -27245;

      matrix[4][0] = 23170;
      matrix[4][1] = -23170;
      matrix[4][2] = -23170;
      matrix[4][3] = 23170;
      matrix[4][4] = 23170;
      matrix[4][5] = -23170;
      matrix[4][6] = -23170;
      matrix[4][7] = 23170;

      matrix[5][0] = 18204;
      matrix[5][1] = -32138;
      matrix[5][2] = 6392;
      matrix[5][3] = 27245;
      matrix[5][4] = -27245;
      matrix[5][5] = -6392;
      matrix[5][6] = 32138;
      matrix[5][7] = -18204;

      matrix[6][0] = 12539;
      matrix[6][1] = -30273;
      matrix[6][2] = 30273;
      matrix[6][3] = -12539;
      matrix[6][4] = -12539;
      matrix[6][5] = 30273;
      matrix[6][6] = -30273;
      matrix[6][7] = 12539;

      matrix[7][0] = 6392;
      matrix[7][1] = -18204;
      matrix[7][2] = 27245;
      matrix[7][3] = -32138;
      matrix[7][4] = 32138;
      matrix[7][5] = -27245;
      matrix[7][6] = 18204;
      matrix[7][7] = -6392;
   end

   genvar g;

   generate
      for (g = 0; g < g_vec_size; g = g + 1) begin
         assign in_vec[g] = in[g_data_width * (g + 1) - 1 : g_data_width * g];
         assign out[g_data_width * (g + 1) - 1 : g_data_width * g] = out_vec[g];
         assign out_vec[g] = out_reg[g];
      end
   endgenerate

   integer      i;
   integer      j;

   always @(posedge clk or posedge rst)
     begin
        if (rst == 1)
          for (i = 0; i < g_vec_size; i = i + 1) begin
             out_reg[i] = 0;
          end
        else
          for (i = 0; i < 8; i = i + 1) begin
             for (j = 0; j < 8; j = j + 1) begin
                out_reg[i] = out_reg[i] + matrix[i][j] * in_vec[j];
             end
          end
     end
endmodule