// FFT controll

// point=512
//`define THE_LAST_STAGE 4'b1000   // 9-1=8
//`define THE_LAST_BFLY  8'hFF     // 512/2-1=255
//point = 2048
//`define THE_LAST_STAGE 4'b1010;   // point=2048
//`define THE_LAST_BFLY  10'h3FF;   // 2048/2-1=1023


module fft_ctrl (
  rst,
  clk,
  frame_in_dav,
  frame_in_enb,
  frame_in_sop,
  frame_out_dav,
  frame_out_enb,
  frame_out_sop,
  frame_out_eop,

  ram_up_wen,
  ram_up_ren,
  ram_up_waddr,
  ram_up_raddr,

  ram_dn_wen,
  ram_dn_ren,
  ram_dn_waddr,
  ram_dn_raddr,
   
  rom_ren,
  rom_raddr,
  
  ram_up_wsel,
  ram_up_rsel,
  frame_input_on,   // to input multiplexer
  ram_rdata_valid,
  wr_stage_cmplt,
  bfly_finish
);

input rst, clk;
input frame_in_dav, frame_out_dav;
output frame_in_enb,frame_out_enb;
input frame_in_sop;
output frame_out_sop, frame_out_eop;
output ram_up_wen, ram_up_ren;
output [9:0] ram_up_waddr, ram_up_raddr;
output ram_dn_wen, ram_dn_ren;
output [9:0] ram_dn_waddr, ram_dn_raddr;
output rom_ren;
output [9:0] rom_raddr;
output ram_up_wsel;
output ram_up_rsel;
output frame_input_on;
output ram_rdata_valid;
output wr_stage_cmplt, bfly_finish;

reg frame_in_enb, frame_out_enb;
reg frame_out_sop, frame_out_eop;
wire ram_up_wsel;
reg ram_up_rsel;
wire frame_input_on;
reg ram_rdata_valid;
wire wr_stage_cmplt;
wire bfly_finish;

//=================================================
// Input control
//=================================================
reg ram_full;
reg [10:0] frame_in_cnt;  //2^11= 2048
reg frame_in_cnt_enb;
wire frame_in_eop;

always @(posedge rst or posedge clk) begin
  if(rst)
	  ram_full <= 1'b0;
  else if(frame_in_eop)
	  ram_full <= 1'b1;
  else if(frame_out_eop)
	  ram_full <= 1'b0;
end

always @(posedge rst or posedge clk) begin
  if(rst)
	  frame_in_enb <= 1'b0;
  else if(frame_in_eop)
	  frame_in_enb <= 1'b0;
  else if(frame_in_dav && !ram_full)
	  frame_in_enb <= 1'b1;
end

always @(posedge rst or posedge clk) begin
  if(rst)
	  frame_in_cnt_enb <= 1'b0;
  else if(frame_in_eop)
	  frame_in_cnt_enb <= 1'b0;
  else if(frame_in_enb && frame_in_sop)
	  frame_in_cnt_enb <= 1'b1;
end

// input switch control
assign frame_input_on = frame_in_sop | frame_in_cnt_enb;

// fram_in_cnt
always @(posedge rst or posedge clk) begin
  if(rst)
	  frame_in_cnt <= 11'h0;  // point=2048
  else if(frame_in_enb && frame_input_on)
	  frame_in_cnt <= frame_in_cnt + 1'b1;
end

assign frame_in_eop = (frame_in_cnt==11'h7FF);  // 2^11=2048

// input writing address
wire [10:0] ram_input_addr;

assign ram_input_addr = { frame_in_cnt[0],
	                  frame_in_cnt[1],
	                  frame_in_cnt[2],
	                  frame_in_cnt[3],
	                  frame_in_cnt[4],
	                  frame_in_cnt[5],
	                  frame_in_cnt[6],
	                  frame_in_cnt[7],
	                  frame_in_cnt[8],
	                  frame_in_cnt[9],
	                  frame_in_cnt[10]
			 };

//=================================================
// RAM read control in butterfly
//=================================================
reg [13:0] rd_cnt;
reg rd_cnt_enb;
wire [3:0] rd_stage;
wire [9:0] rd_index;
wire rd_index_eop;
wire rd_stage_eop;
reg [2:0] stage_end_delay;
reg rd_cnt_wait;

always @(posedge rst or posedge clk) begin
  if(rst)
	rd_cnt <= 14'h0;
  else if(frame_in_eop)
	rd_cnt <= 14'h0;
  else if(rd_cnt_enb && !rd_cnt_wait )
	rd_cnt <= rd_cnt + 1'b1;
end

assign rd_stage = rd_cnt[13:10];
assign rd_index = rd_cnt[9:0];

assign rd_index_eop = (rd_index==10'b11_1111_1111) && rd_cnt_enb;    // the last butterfly in the stage  
assign rd_stage_eop = (rd_stage==4'b1010) && rd_index_eop;    // the last stage 11-1=10

always @(posedge rst or posedge clk) begin
  if(rst)
	  stage_end_delay <= 3'b111;
  else if(rd_index_eop)            
	  stage_end_delay <= 3'b000;
  else if(stage_end_delay!=3'b111)
	  stage_end_delay <= stage_end_delay + 1'b1;
end 

always @(posedge rst or posedge clk) begin
  if(rst)
	  rd_cnt_wait <= 1'b0;
  else if(rd_index_eop)            
	  rd_cnt_wait <= 1'b1;
  else if(stage_end_delay==3'b101)  // end of writing delay 6clk(0~5)
	  rd_cnt_wait <= 1'b0;
end 

always @(posedge rst or posedge clk) begin
  if(rst)
	rd_cnt_enb <= 1'b0;
  else if(frame_in_eop)
	rd_cnt_enb <= 1'b1;
  else if(rd_stage_eop )
	rd_cnt_enb <= 1'b0;
end

//-------------------------------------------------
// map the RAM read address
reg [9:0] raddr_up;  //1024
wire [9:0] raddr_dn;
wire ren_up;
wire ren_dn;

assign ren_up = rd_cnt_enb & ~rd_cnt_wait;

// circle left shift
always @(rd_stage or rd_index) begin
	case(rd_stage)
        	4'b0000,
		4'b1010: raddr_up <= rd_index[9:0];
        	4'b0001: raddr_up <= {rd_index[8:0],rd_index[9]};
        	4'b0010: raddr_up <= {rd_index[7:0],rd_index[9:8]};
        	4'b0011: raddr_up <= {rd_index[6:0],rd_index[9:7]};
        	4'b0100: raddr_up <= {rd_index[5:0],rd_index[9:6]};
        	4'b0101: raddr_up <= {rd_index[4:0],rd_index[9:5]};
        	4'b0110: raddr_up <= {rd_index[3:0],rd_index[9:4]};
        	4'b0111: raddr_up <= {rd_index[2:0],rd_index[9:3]};
        	4'b1000: raddr_up <= {rd_index[1:0],rd_index[9:2]};
        	4'b1001: raddr_up <= {rd_index[0],rd_index[9:1]};
        	default: raddr_up <= 10'b00_0000_0000;
	endcase
end

assign ren_dn = ren_up;
assign raddr_dn = raddr_up;

//=================================================
// RAM write control in butterfly
//=================================================
reg [13:0] wr_cnt;
reg wr_cnt_enb;
wire [3:0] wr_stage;
wire [9:0] wr_index;
wire wr_index_sop;
wire wr_index_eop;
wire wr_stage_eop;
reg [2:0] stage_start_delay;
reg wr_cnt_wait;

wire wen_up;
reg wen_dn;

always @(posedge rst or posedge clk) begin
  if(rst)
	wr_cnt <= 14'h0;
  else if(frame_in_eop)   // cleared by input finish
	wr_cnt <= 14'h0;
  else if(wr_cnt_enb && !wr_cnt_wait )
	wr_cnt <= wr_cnt + 1'b1;
end

assign wr_stage = wr_cnt[13:10];
assign wr_index = wr_cnt[9:0];

assign wr_index_sop = ((stage_end_delay==3'b101) && wr_cnt_enb) || frame_in_eop;   // the last butterfly in the stage 

assign wr_index_eop = (wr_index==10'b11_1111_1111) && wr_cnt_enb;
assign wr_stage_eop = (wr_stage==4'b1010) && wr_index_eop; // 11-1=10  

always @(posedge rst or posedge clk) begin
  if(rst)
	  stage_start_delay <= 3'b111;
  //else if(wr_index_sop || frame_in_eop)  // the start of each write stage or the end of frame_in   
  else if(wr_index_sop)  // the start of each write stage or the end of frame_in   
	  stage_start_delay <= 3'b000;
  else if(stage_start_delay!=3'b111)
	  stage_start_delay <= stage_start_delay + 1'b1;
end 

always @(posedge rst or posedge clk) begin
  if(rst)
	  wr_cnt_wait <= 1'b0;
  else if(wr_index_eop || frame_in_eop)            
	  wr_cnt_wait <= 1'b1;
  else if(stage_start_delay==3'b100)  // end of writing delay 5clk(0~4)
	  wr_cnt_wait <= 1'b0;
end 

always @(posedge rst or posedge clk) begin
  if(rst)
	wr_cnt_enb <= 1'b0;
else if(frame_in_eop)
	wr_cnt_enb <= 1'b1;
  else if(wr_stage_eop )  // the last writing stage
	wr_cnt_enb <= 1'b0;
end

assign wen_up = wr_cnt_enb & ~wr_cnt_wait;

always @(posedge rst or posedge clk) begin
  if(rst)
	  wen_dn <= 1'b0;
  else 
	  wen_dn <= wen_up;  // 1clk dealy
end 

//assign ram_up_wsel = ~wr_cnt[0] & wen_up;
//assign ram_up_wsel = ~wr_cnt[0] & wen_up & ~wr_index_eop;  // avoid glach at the end of every stage
assign ram_up_wsel = ~wr_cnt[0] & (wen_up | wen_dn); 

//-------------------------------------------------
// map the RAM write address
reg [9:0] waddr1, waddr2;

always @(wr_stage or wr_index) begin
	case(wr_stage)
        	4'b0000,4'b1010: begin
			waddr1 <= {wr_index[9:1],1'b0};
			waddr2 <= {wr_index[9:1],1'b1};
		end
        	4'b0001: begin
			waddr1 <= {wr_index[8:1],1'b0,wr_index[9]};
			waddr2 <= {wr_index[8:1],1'b1,wr_index[9]};
		end
        	4'b0010: begin
			waddr1 <= {wr_index[7:1],1'b0,wr_index[9:8]};
			waddr2 <= {wr_index[7:1],1'b1,wr_index[9:8]};
		end
        	4'b0011: begin
			waddr1 <= {wr_index[6:1],1'b0,wr_index[9:7]};
			waddr2 <= {wr_index[6:1],1'b1,wr_index[9:7]};
		end
        	4'b0100: begin
			waddr1 <= {wr_index[5:1],1'b0,wr_index[9:6]};
			waddr2 <= {wr_index[5:1],1'b1,wr_index[9:6]};
		end
        	4'b0101: begin
			waddr1 <= {wr_index[4:1],1'b0,wr_index[9:5]};
			waddr2 <= {wr_index[4:1],1'b1,wr_index[9:5]};
		end
        	4'b0110: begin
			waddr1 <= {wr_index[3:1],1'b0,wr_index[9:4]};
			waddr2 <= {wr_index[3:1],1'b1,wr_index[9:4]};
		end
        	4'b0111: begin
			waddr1 <= {wr_index[2:1],1'b0,wr_index[9:3]};
			waddr2 <= {wr_index[2:1],1'b1,wr_index[9:3]};
		end
        	4'b1000: begin
			waddr1 <= {wr_index[1],1'b0,wr_index[9:2]};
			waddr2 <= {wr_index[1],1'b1,wr_index[9:2]};
		end
        	4'b1001: begin
			waddr1 <= {1'b0,wr_index[9:1]};
			waddr2 <= {1'b1,wr_index[9:1]};
		end
        	default: begin
			waddr1 <= 10'b00_0000_0000;
			waddr2 <= 10'b00_0000_0000;
		end
	endcase
end

// waddr2 buffer
reg [9:0] waddr2_d1t_up, waddr2_d1t_dn;

always @(posedge rst or posedge clk) begin
  if(rst)
	waddr2_d1t_up = 10'h0;
  else if(!wr_cnt[0] && wr_cnt_enb)
	waddr2_d1t_up = waddr2;  // write after 1 clk delay
end

always @(posedge rst or posedge clk) begin
  if(rst)
	waddr2_d1t_dn = 10'h0;
  else if(wr_cnt[0] && wr_cnt_enb)
	waddr2_d1t_dn = waddr2;  // write after 1 clk delay
end

//=================================================
// output control
//=================================================
reg bfly_finish_flg;
reg [10:0] frame_out_cnt; //2^11=2048
reg frame_out_cnt_enb;
wire frame_out_cnt_sop;
wire frame_out_cnt_eop;

//assign bfly_finish = (wr_stage==4'b1000) && (waddr2_d1t_dn==8'b1111_1111);
assign bfly_finish = (wr_stage==4'b1010) && (waddr2==10'b11_1111_1111);

always @(posedge rst or posedge clk) begin
  if(rst)
	  bfly_finish_flg <= 1'b0;
  else if(bfly_finish)
	  bfly_finish_flg <= 1'b1;
  else if(frame_out_dav)
	  bfly_finish_flg <= 1'b0;
end
  
always @(posedge rst or posedge clk) begin
  if(rst)
	  frame_out_cnt_enb <= 1'b0;
  else if(frame_out_dav && bfly_finish_flg)
	  frame_out_cnt_enb <= 1'b1;
  else if(frame_out_cnt_eop)
	  frame_out_cnt_enb <= 1'b0;
end

assign frame_out_cnt_sop = (frame_out_cnt==11'h000) && frame_out_cnt_enb;
assign frame_out_cnt_eop = (frame_out_cnt==11'h7FF) && frame_out_cnt_enb;

always @(posedge rst or posedge clk) begin
  if(rst)
	  frame_out_cnt <= 11'h0;
  else if(frame_out_cnt_eop)
	  frame_out_cnt <= 11'h0;
  else if(frame_out_cnt_enb)
	  frame_out_cnt <= frame_out_cnt + 1'b1;
end

// delay 2clk output
reg frame_out_cnt_enb_d1t;
reg frame_out_cnt_sop_d1t;
reg frame_out_cnt_eop_d1t;

always @(posedge rst or posedge clk) begin
  if(rst) begin
	  frame_out_cnt_enb_d1t <= 1'b0;
	  frame_out_cnt_sop_d1t <= 1'b0;
	  frame_out_cnt_eop_d1t <= 1'b0;
	  frame_out_enb <= 1'b0;
	  frame_out_sop <= 1'b0;
	  frame_out_eop <= 1'b0;
  end
  else begin
	  frame_out_cnt_enb_d1t <= frame_out_cnt_enb;
	  frame_out_cnt_sop_d1t <= frame_out_cnt_sop;
	  frame_out_cnt_eop_d1t <= frame_out_cnt_eop;
	  frame_out_enb <= frame_out_cnt_enb_d1t;
	  frame_out_sop <= frame_out_cnt_sop_d1t;
	  frame_out_eop <= frame_out_cnt_eop_d1t;
  end
end


//=================================================
// RAM control output 
//=================================================
assign ram_up_ren = ren_up | (frame_out_cnt_enb & ~frame_out_cnt[0]);
assign ram_dn_ren = ren_dn | (frame_out_cnt_enb &  frame_out_cnt[0]);
assign ram_up_raddr = (frame_out_cnt_enb) ? {frame_out_cnt[9:1],frame_out_cnt[10]} : 
	                                    raddr_up;
assign ram_dn_raddr = (frame_out_cnt_enb) ? {frame_out_cnt[9:1],frame_out_cnt[10]} : 
	                                    raddr_dn;

assign ram_up_wen = (~ram_input_addr[0] & frame_input_on) | wen_up ; 
assign ram_dn_wen = ( ram_input_addr[0] & frame_input_on) | wen_dn ; 
assign ram_up_waddr = (frame_input_on) ? ram_input_addr[10:1] : 
	                                (wr_cnt[0]) ? waddr2_d1t_up : waddr1;
assign ram_dn_waddr = (frame_input_on) ? ram_input_addr[10:1] : 
	                                (wr_cnt[0]) ? waddr1 : waddr2_d1t_dn;

				
//=================================================
// map the ROM read address
//=================================================
reg rom_ren;
reg [10:0] right_shift;
reg [9:0] rom_raddr;
wire [9:0] rom_offset;
reg [9:0] tw_cnt;
wire [9:0] field_num;

always @(posedge rst or posedge clk) begin
  if(rst)
        rom_ren <= 1'b0;
  else if(stage_start_delay==3'b000)  // 1clk delay with ram_ren
        rom_ren <= 1'b1;
  else if(stage_end_delay==3'b000)
        rom_ren <= 1'b0;
end

always @(posedge rst or posedge clk) begin
  if(rst)
	right_shift <= 11'b000_0000_0000;
  else if(frame_in_eop)
	right_shift <= 11'b100_0000_0000;
  //else if(rd_index_eop)
  else if(stage_end_delay==3'b001)
	right_shift <= {1'b0,right_shift[10:1]};
end

assign rom_offset = right_shift[9:0];

assign field_num = rom_offset - 1'b1;

always @(posedge rst or posedge clk) begin
  if(rst) begin
	  tw_cnt <= 9'h0;
  end
  else begin 
	  if( (tw_cnt == field_num) && rom_ren )
	  	tw_cnt <= 9'h0;
	  else if (rom_ren)
	  	tw_cnt <= tw_cnt + 1'b1;
  end
end
	 
always @(posedge rst or posedge clk) begin
  if(rst)
	rom_raddr <= 10'b00_0000_0000;
  else if(frame_in_eop)
	rom_raddr <= 10'b00_0000_0000;
  else if(rom_ren && (tw_cnt == field_num) )
	rom_raddr <= rom_raddr + rom_offset;
end

//=================================================
// internal control out
//=================================================
reg frame_out_ram_up;

always @(posedge rst or posedge clk) begin
  if(rst)
	  ram_rdata_valid <= 1'b0;
  else if(stage_start_delay==3'b001) // 2clk delay after ram_ren
	  ram_rdata_valid <= 1'b1;
  else if(stage_end_delay==3'b001)
	  ram_rdata_valid <= 1'b0;
end 

assign wr_stage_cmplt = wr_index_eop;

always @(posedge rst or posedge clk) begin
  if(rst) begin
        frame_out_ram_up <= 1'b0;
	ram_up_rsel      <= 1'b0;
  end
  else begin
        frame_out_ram_up <= frame_out_cnt_enb & ~frame_out_cnt[0];
	ram_up_rsel      <= frame_out_ram_up;
  end
end

endmodule