LIBRARY IEEE;
USE IEEE.STD_LOGIC_1164.ALL;
USE IEEE.NUMERIC_STD.ALL;
use std.textio.all; --  Imports the standard textio package.

use work.utils_pkg.all;

ENTITY emac_jb IS
	Generic
	(
		f_sysclk	: real		:= 100.0;
		TX_RAM_SIZE	: natural	:= 2048
	);
	Port
	(
		CLK_I		: in  STD_LOGIC;
		RST_I		: in  STD_LOGIC;
		INT_O		: out STD_LOGIC;
		CYC_I		: in  STD_LOGIC;
		STB_I		: in  STD_LOGIC;
		SEL_I		: in  unsigned(3 downto 0);
		WE_I		: in  STD_LOGIC;
		ACK_O		: out STD_LOGIC;
		SRDY_O		: out STD_LOGIC;
		MRDY_I		: in  STD_LOGIC;
		ADDR_I		: in  unsigned(31 downto 0);
		DAT_I		: in  unsigned(31 downto 0);
		DAT_O		: out unsigned(31 downto 0);
		mii_rx_clk	: in  STD_LOGIC;
		mii_rx_dv	: in  STD_LOGIC;
		mii_rx_er	: in  STD_LOGIC;
		mii_rx		: in unsigned(7 downto 0);
		mii_tx_clk	: in  STD_LOGIC;
		mii_tx_en	: out STD_LOGIC;
		mii_tx_er	: out STD_LOGIC;
		mii_tx		: out unsigned(7 downto 0);
		mii_gtx_clk	: out STD_LOGIC;
		mii_crs		: in  STD_LOGIC;
		mii_col		: in  STD_LOGIC
		
	);
END emac_jb;

ARCHITECTURE behavior OF emac_jb IS

	constant TX_RAM_ADDR_WIDTH	: natural := NextExpBaseTwo(TX_RAM_SIZE);
	
	subtype word_ptr_t is unsigned(TX_RAM_ADDR_WIDTH-1 downto 0);
	
	-- Signals for EMAC connections
	signal tx_fifo_din		: unsigned(35 downto 0);
	signal tx_fifo_dout		: unsigned(35 downto 0);
	signal tx_fifo_we		: std_logic;
	signal tx_fifo_re		: std_logic;
	signal tx_fifo_empty	: std_logic;
	signal tx_fifo_full		: std_logic;
	signal rx_int_en		: std_logic;
	signal tx_int_en		: std_logic;
	signal irq_rx			: std_logic;
	signal irq_tx			: std_logic;
	signal tx_err			: std_logic;

	signal tx_flush_en		: std_logic;
	signal tx_flush_rdy		: std_logic;
	signal tx_flush_ptr_set	: std_logic;

	signal tx_commit_en		: std_logic;
	signal tx_flush_ptr		: word_ptr_t;
	
	signal host_data_we		: std_logic;
	signal host_data_reg	: unsigned(31 downto 0);
	signal host_tx_size		: unsigned(15 downto 0);

	signal tx_fill_size		: word_ptr_t;
	signal tx_fill_cnt		: word_ptr_t;
	signal tx_fill_ptr		: word_ptr_t;
	signal tx_fill_rdy		: std_logic;
	signal tx_fill_cnt_set	: std_logic;
	signal tx_fill_cnt_adv	: std_logic;
	signal tx_fill_ptr_set	: std_logic;
	signal tx_fill_ptr_adv	: std_logic;

	signal tx_alloc_OK		: std_logic;
	signal tx_alloc_req_en	: std_logic;
	signal tx_alloc_req		: std_logic;
	signal tx_alloc_ack		: std_logic;
	
	signal tx_nwords_free	: word_ptr_t;
	signal tx_alloc_base	: word_ptr_t;
	signal tx_alloc_size	: word_ptr_t;
	signal tx_alloc_en		: std_logic;

	signal tx_free_en		: std_logic;
	signal tx_uncommit_en	: std_logic;
	signal tx_fill_remain	: unsigned(1 downto 0);

	signal tx_xfer_size		: word_ptr_t;
	signal tx_xfer_ptr		: word_ptr_t;
	signal tx_xfer_cnt		: word_ptr_t;
	signal tx_xfer_en		: std_logic;
	signal tx_xfer_ptr_set	: std_logic;
	signal tx_xfer_ptr_adv	: std_logic;
	signal tx_xfer_cnt_set	: std_logic;
	signal tx_xfer_cnt_adv	: std_logic;
	signal tx_xfer_rdy		: std_logic;
	signal xfer_tag			: unsigned(3 downto 0);
	signal xfer_remain		: unsigned(1 downto 0);
	signal xfer_odd			: std_logic;
	signal xfer_remain_en	: std_logic;

	signal tx_ram_en_a		: std_logic;
	signal tx_ram_we_a		: std_logic;
	signal tx_ram_addr_a	: unsigned(TX_RAM_ADDR_WIDTH-1 downto 0);
	signal tx_ram_din_a		: unsigned(35 downto 0);
	signal tx_ram_dout_a	: unsigned(35 downto 0);
	signal tx_ram_en_b		: std_logic;
	signal tx_ram_we_b		: std_logic;
	signal tx_ram_addr_b	: unsigned(TX_RAM_ADDR_WIDTH-1 downto 0);
	signal tx_ram_din_b		: unsigned(35 downto 0);
	signal tx_ram_dout_b	: unsigned(35 downto 0);

	signal piso_din_be		: unsigned(3 downto 0);
	signal piso_din_rdy		: std_logic;
	signal tx_data_vld		: std_logic;
	
	subtype inter_frame_gap_cnt_t is natural range 0 to natural(1.0*f_sysclk)-1;
	signal inter_frame_gap_cnt	: inter_frame_gap_cnt_t;
	signal inter_frame_gap_rdy	: std_logic;
	
	alias tx_ram_data_in_a is tx_ram_din_a(31 downto 0);
	alias tx_ram_tag_in_a is tx_ram_din_a(35 downto 32);
	alias tx_ram_data_out_a is tx_ram_dout_a(31 downto 0);
	alias tx_ram_tag_out_a is tx_ram_dout_a(35 downto 32);
	
	alias tx_ram_data_in_b is tx_ram_din_b(31 downto 0);
	alias tx_ram_tag_in_b is tx_ram_din_b(35 downto 32);
	alias tx_ram_data_out_b is tx_ram_dout_b(31 downto 0);
	alias tx_ram_tag_out_b is tx_ram_dout_b(35 downto 32);

	type host_tx_state_t is (host_tx_init, host_tx_flush, host_tx_idle, host_tx_alloc, host_tx_setup, host_tx_arm, host_tx_fill, host_tx_commit);
	signal host_tx_s, host_tx_sn : host_tx_state_t;
	
	type xfer_tx_state_t is (xfer_tx_init, xfer_tx_idle, xfer_tx_setup, xfer_wait, xfer_tx_active, xfer_tx_free);
	signal xfer_tx_s, xfer_tx_sn : xfer_tx_state_t;


	type piso_byte_mask_array_t is array (0 to 3) of unsigned (3 downto 0);
	constant piso_byte_mask_rom : piso_byte_mask_array_t :=
	(
		"1111",
		"0001",
		"0011",
		"0111"
	);
	
	
begin

	SRDY_O				<= CYC_I;
	INT_O				<= irq_rx or irq_tx;
	mii_gtx_clk			<= '0';
	mii_tx_er			<= tx_err;

	tx_ram_tag_in_b		<= "0000";
	tx_ram_en_b			<= '1';
	tx_ram_we_b			<= tx_uncommit_en or tx_flush_en;
	tx_ram_addr_b		<= tx_flush_ptr when (tx_flush_en = '1' or tx_uncommit_en = '1') else tx_xfer_ptr;
	xfer_tag			<= tx_ram_tag_out_b;
	
	tx_ram_en_a			<= '1';
	tx_ram_we_a			<= tx_commit_en or tx_fill_cnt_adv;
	tx_ram_tag_in_a		<= tx_fill_remain & "01" when tx_commit_en = '1' else "0000";
	tx_ram_data_in_a	<= resize(tx_fill_size, 16) & resize(tx_alloc_base, 16) when tx_commit_en = '1' else host_data_reg;
	tx_ram_addr_a		<= tx_alloc_base when tx_commit_en = '1' else tx_fill_ptr;

	piso_din_be			<= (others => '1');
	tx_fifo_re			<= piso_din_rdy;
	tx_data_vld			<= not tx_fifo_empty;
	
	tx_fifo_din(35 downto 32)	<= piso_byte_mask_rom(to_integer(xfer_remain)) when xfer_remain_en = '1' else "1111";
	tx_fifo_din(31 downto 0)	<= tx_ram_data_out_b;
	
------------------------------------------------------------------
host_tx_state_next:
	process(CLK_I)
	begin
		if rising_edge(CLK_I) then
			if RST_I = '1' then
				host_tx_s	<= host_tx_init;
			else
				host_tx_s	<= host_tx_sn;
			end if;
		end if;
	end process;	

host_tx_state:
	process(host_tx_s, tx_flush_rdy, tx_alloc_req, tx_alloc_OK, tx_fill_rdy, host_data_we)
	begin
	
		tx_flush_en			<= '0';
		tx_flush_ptr_set	<= '0';
		tx_alloc_ack		<= '0';
		tx_fill_cnt_set		<= '0';
		tx_fill_cnt_adv		<= '0';
		tx_fill_ptr_set		<= '0';
		tx_fill_ptr_adv		<= '0';
		tx_commit_en		<= '0';
		tx_alloc_en			<= '0';
		
		host_tx_sn			<= host_tx_s;
		
		case host_tx_s is
		
		when host_tx_init =>
			tx_flush_ptr_set	<= '1';
			host_tx_sn <= host_tx_flush;
		
		when host_tx_flush =>
			tx_flush_en	<= '1';
			if tx_flush_rdy = '1' then
				tx_flush_en	<= '0';
				host_tx_sn <= host_tx_idle;
			end if;
			
		when host_tx_idle =>
			if tx_alloc_req = '1' then
				host_tx_sn	<= host_tx_alloc;
			end if;

		when host_tx_alloc =>
			if tx_alloc_OK = '1' then
				host_tx_sn			<= host_tx_setup;
				tx_alloc_ack		<= '1';
			end if;

		when host_tx_setup =>
			tx_fill_cnt_set	<= '1';
			tx_fill_ptr_set	<= '1';
			host_tx_sn		<= host_tx_arm;

		when host_tx_arm =>
			tx_fill_ptr_adv	<= '1';
			tx_fill_cnt_adv	<= '1';
			host_tx_sn		<= host_tx_fill;

		when host_tx_fill =>
			tx_fill_cnt_adv	<= host_data_we;
			tx_fill_ptr_adv	<= not tx_fill_rdy and host_data_we;
			if tx_alloc_req = '1' then
				host_tx_sn	<= host_tx_alloc;
			elsif tx_fill_rdy = '1' then
				host_tx_sn		<= host_tx_commit;
			end if;
		
		when host_tx_commit =>
			tx_commit_en	<= '1';
			tx_alloc_en		<= '1';
			host_tx_sn		<= host_tx_idle;

		when others => 
			host_tx_sn	<= host_tx_idle;
		end case;

	end process;

------------------------------------------------------------------
-- Allocation stuff
------------------------------------------------------------------
alloc_request_logic:
	process(CLK_I)
	begin
		if rising_edge(CLK_I) then
			if RST_I = '1' then
				tx_alloc_req		<= '0';
			elsif tx_alloc_req_en = '1' and tx_alloc_req = '0' then
				tx_alloc_req		<= '1';
				tx_fill_remain		<= host_tx_size(1 downto 0);
				tx_fill_size		<= host_tx_size(word_ptr_t'left+2 downto 2);
				if host_tx_size(1 downto 0) /= "00" then
					tx_fill_size	<= host_tx_size(word_ptr_t'left+2 downto 2) + 1;
				end if;
			elsif tx_alloc_ack = '1' then
				tx_alloc_req		<= '0';
				tx_alloc_size		<= tx_fill_size;
			end if;
		end if;
	end process;	

------------------------------------------------------------------
alloc_base_logic:
	process(CLK_I)
	begin
		if rising_edge(CLK_I) then
			if tx_flush_en = '1' then
				tx_alloc_base		<= (others => '0');
			elsif tx_commit_en = '1' then
				tx_alloc_base		<= tx_fill_ptr;
			end if;
		end if;
	end process;	

------------------------------------------------------------------
alloc_eval_logic:
	process(CLK_I)
	begin
		if rising_edge(CLK_I) then
			tx_alloc_OK		<= '0';
			if tx_alloc_req = '1' then
				if tx_fill_size < tx_nwords_free then
					tx_alloc_OK		<= '1';
				end if;
			end if;
		end if;
	end process;	

------------------------------------------------------------------
tx_nwords_free_counter:
	process(CLK_I)
	begin
		if rising_edge(CLK_I) then
			if tx_flush_en = '1' then
				tx_nwords_free		<= (others => '1');
			elsif tx_alloc_en = '1' then
				tx_nwords_free		<= tx_nwords_free - tx_fill_size;
			elsif tx_free_en = '1' then
				tx_nwords_free		<= tx_nwords_free + tx_xfer_size;
			end if;
		end if;
	end process;	

------------------------------------------------------------------
tx_flush_counter:
	process(CLK_I)
	begin
		if rising_edge(CLK_I) then
			if tx_flush_en = '0' then
				tx_flush_rdy	<= '0';
				if tx_flush_ptr_set = '1' then
					tx_flush_ptr	<= (others => '1');
				elsif tx_xfer_ptr_set = '1' then
					tx_flush_ptr	<= tx_ram_data_out_b(tx_flush_ptr'left downto 0);
				end if;	
			elsif tx_flush_ptr /= 0 then
				tx_flush_ptr	<= tx_flush_ptr - 1;
			else
				tx_flush_rdy	<= '1';
			end if;
		end if;
	end process;	

------------------------------------------------------------------
-- Fill stuff
------------------------------------------------------------------
tx_fill_counter:
	process(CLK_I)
	begin
		if rising_edge(CLK_I) then
			if tx_fill_cnt_set = '1' then	
				tx_fill_rdy			<= '0';
				tx_fill_cnt			<= tx_alloc_size;
			elsif tx_fill_cnt_adv = '1' then
				if tx_fill_cnt /= 0 then
					tx_fill_cnt		<= tx_fill_cnt - 1;
				else
					tx_fill_rdy		<= '1';
				end if;
			end if;
		end if;
	end process;	

tx_fill_pointer:
	process(CLK_I)
	begin
		if rising_edge(CLK_I) then
			if tx_fill_ptr_set = '1' then
				tx_fill_ptr		<= tx_alloc_base;
			elsif tx_fill_ptr_adv = '1' then
				tx_fill_ptr		<= tx_fill_ptr + 1;
			end if;
		end if;
	end process;	

------------------------------------------------------------------
-- Transfer stuff
------------------------------------------------------------------
xfer_tx_state_next:
	process(CLK_I)
	begin
		if rising_edge(CLK_I) then
			if RST_I = '1' then
				xfer_tx_s	<= xfer_tx_init;
			else
				xfer_tx_s	<= xfer_tx_sn;
			end if;
		end if;
	end process;	

xfer_tx_state:
	process(xfer_tx_s, xfer_tag, xfer_odd, tx_xfer_rdy, inter_frame_gap_rdy, tx_alloc_en, tx_fifo_full)
	begin
	
		tx_xfer_cnt_set		<= '0';
		tx_xfer_cnt_adv		<= '0';
		tx_xfer_ptr_set		<= '0';
		tx_xfer_ptr_adv		<= '0';
		tx_xfer_en			<= '0';
		xfer_remain_en		<= '0';
		tx_free_en			<= '0';
		tx_uncommit_en		<= '0';
		
		xfer_tx_sn			<= xfer_tx_s;
		
		case xfer_tx_s is
		
		when xfer_tx_init =>
			xfer_tx_sn <= xfer_tx_idle;
		
		when xfer_tx_idle =>
			if xfer_tag(0) = '1' then
				tx_xfer_ptr_set	<= '1';
				tx_xfer_cnt_set	<= '1';
				xfer_tx_sn		<= xfer_wait;
			end if;
			
		when xfer_wait =>
			if inter_frame_gap_rdy = '1' then
				xfer_tx_sn			<= xfer_tx_setup;
			end if;			
		
		when xfer_tx_setup =>
			tx_uncommit_en		<= '1';
			tx_xfer_ptr_adv		<= '1';
			tx_xfer_cnt_adv		<= '1';
			xfer_tx_sn			<= xfer_tx_active;

		when xfer_tx_active =>
			tx_xfer_en			<= '1';
			tx_xfer_cnt_adv		<= not tx_fifo_full;
			tx_xfer_ptr_adv		<= not tx_fifo_full;
			if tx_xfer_rdy = '1' then
				xfer_remain_en	<= xfer_odd;
				tx_xfer_en		<= '0';
				tx_xfer_cnt_adv	<= '0';
				tx_xfer_ptr_adv	<= '0';
				xfer_tx_sn		<= xfer_tx_free;
			end if;	

		when xfer_tx_free =>
			tx_free_en			<= not tx_alloc_en;
			if tx_alloc_en = '0' then
				xfer_tx_sn			<= xfer_tx_idle;
			end if;

		when others => 
			xfer_tx_sn	<= xfer_tx_idle;
		end case;

	end process;

------------------------------------------------------------------
interframe_gap_counter:
	process(CLK_I)
	begin
		if rising_edge(CLK_I) then
			if tx_flush_en = '1' then
				inter_frame_gap_rdy	<= '1';
			elsif tx_fifo_empty = '0' then
				inter_frame_gap_rdy	<= '0';
				inter_frame_gap_cnt <= inter_frame_gap_cnt_t'high;
			elsif inter_frame_gap_cnt /= 0 then
				inter_frame_gap_cnt <= inter_frame_gap_cnt - 1;
			else
				inter_frame_gap_rdy	<= '1';
			end if;
		end if;
	end process;

------------------------------------------------------------------
tx_xfer_counter:
	process(CLK_I)
	begin
		if rising_edge(CLK_I) then
			if tx_xfer_cnt_set = '1' then
				tx_xfer_size	<= (others => '0');
				tx_xfer_rdy		<= '0';
				tx_xfer_cnt 	<= tx_ram_data_out_b(tx_xfer_ptr'left+16 downto 16);
			elsif tx_xfer_cnt_adv = '1' then
				if tx_xfer_cnt /= 0 then
					tx_xfer_cnt		<= tx_xfer_cnt - 1;
					tx_xfer_size	<= tx_xfer_size + 1;
				else
					tx_xfer_rdy		<= '1';
				end if;
			end if;
		end if;	
			
	end process;	
	
tx_xfer_pointer:
	process(CLK_I)
	begin
		if rising_edge(CLK_I) then
			tx_fifo_we			<= tx_xfer_en;
			if tx_xfer_ptr_set = '1' then
				tx_xfer_ptr		<= tx_ram_data_out_b(tx_xfer_ptr'left downto 0);
				xfer_remain		<= xfer_tag(3 downto 2);
				xfer_odd 		<= '0';
				if xfer_tag(3 downto 2) /= "00" then
					xfer_odd <= '1';
				end if;
			elsif tx_xfer_ptr_adv = '1' then
				tx_xfer_ptr		<= tx_xfer_ptr + 1;
			end if;
		end if;	
			
	end process;	
	
------------------------------------------------------------------
registers_write: 
	process(CLK_I)
	begin
		if rising_edge(CLK_I) then
			host_data_we	<= '0';
			tx_alloc_req_en	<= '0';
			if RST_I = '1' then
				rx_int_en		<= '0';
				tx_int_en		<= '0';
				tx_err			<= '0';
		elsif (STB_I and CYC_I and WE_I) = '1' then
				case ADDR_I(5 downto 2) is
				
				when "0000" =>
					tx_err				<= DAT_I(31);
					tx_alloc_req_en		<= DAT_I(16);
					tx_int_en			<= DAT_I(5);
					rx_int_en			<= DAT_I(4);
					
				when "0001" => 
					if tx_alloc_req = '0' then
						host_tx_size		<= DAT_I(31 downto 16);
					end if;	
					
				when "0010" =>
					host_data_we		<= '1';
					host_data_reg		<= DAT_I;
					
				when others => null;
				end case;
			end if;	
		end if;	
	end process;

registers_read: 
	process(CLK_I)
	begin
		if rising_edge(CLK_I) then
			ACK_O			<= '0';
			if (STB_I and CYC_I) = '1' then
				ACK_O		<= not WE_I;
				DAT_O		<= (others => '0');
				case ADDR_I(5 downto 2) is

				when "0000" =>
					DAT_O(31)	<= tx_err;
					DAT_O(30)	<= mii_rx_er;
					DAT_O(29)	<= mii_col;
					DAT_O(28)	<= mii_crs;
					DAT_O(16)	<= tx_alloc_req;
					DAT_O(5)	<= tx_int_en;
					DAT_O(4)	<= rx_int_en;

				when "0001" => 
					DAT_O(31 downto 16)	<= host_tx_size;

				when "0010" =>

				when others => null;
				end case;
			end if;
		end if;	
	end process;

inst_tx_ram : entity work.dpram_2w2r
	GENERIC MAP 
	(
		addr_width	=> TX_RAM_ADDR_WIDTH,
		data_width	=> 36
	)
	PORT MAP 
	(
		clk_a	=> CLK_I,
		clk_b	=> CLK_I,
		en_a	=> tx_ram_en_a,
		en_b	=> tx_ram_en_b,
		we_a	=> tx_ram_we_a,
		we_b	=> tx_ram_we_b,
		addr_a	=> tx_ram_addr_a,
		addr_b	=> tx_ram_addr_b,
		din_a	=> tx_ram_din_a,
		din_b	=> tx_ram_din_b,
		dout_a	=> tx_ram_dout_a,
		dout_b	=> tx_ram_dout_b
	);

-- Instantiate synchronous FIFO
inst_tx_fifo: entity work.fifo_async
	GENERIC MAP
	(
		addr_width			=> 5,
		data_width			=> 36,
		do_last_read_update	=> true
	)
	PORT MAP
	(
		rst			=>	RST_I,
		clk_w		=>	CLK_I,
		clk_r		=>	mii_tx_clk,
		we			=>	tx_fifo_we,
		re			=>	tx_fifo_re,
		fifo_full	=>	tx_fifo_full,
		fifo_empty	=>	tx_fifo_empty,
		fifo_afull	=>	open,
		fifo_aempty	=>	open,
		data_w		=>  tx_fifo_din,
		data_r		=>  tx_fifo_dout
	);

inst_piso : entity work.piso
	GENERIC MAP
	(
		data_width_in	=> 32,
		data_width_out	=> 8,
		msb_first		=> false
	)
	PORT MAP
	(
		rst			=> RST_I,
		clk			=> mii_tx_clk,
		dout_en		=> '1',
		din_vld		=> tx_data_vld,
		din_rdy		=> piso_din_rdy,
		din_be		=> tx_fifo_dout(35 downto 32),
		din			=> tx_fifo_dout(31 downto 0),
		dout_vld	=> mii_tx_en,
		dout		=> mii_tx
		
	);

irq_register:
	process(CLK_I)
	begin
		if rising_edge(CLK_I) then
			irq_tx		<= tx_int_en;
			irq_rx		<= rx_int_en;
		end if;
	end process;

end behavior;