From c19e89b03bc1ddddcb9585636b513ad9df6326c7 Mon Sep 17 00:00:00 2001
From: "Wesley W. Terpstra" <w.terpstra@gsi.de>
Date: Tue, 17 Apr 2012 11:09:14 +0200
Subject: [PATCH] TX path added to altera pcie wrapper stub wishbone device
added for testing
---
hdl/pcie_altera.vhd | 138 +++++++++++++++++++++++++++++++++++++++++---
hdl/pcie_tlp.vhd | 37 +++++++++++-
hdl/pcie_wb.vhd | 38 +++++++++---
hdl/pcie_wb_pkg.vhd | 22 +++++--
4 files changed, 212 insertions(+), 23 deletions(-)
diff --git a/hdl/pcie_altera.vhd b/hdl/pcie_altera.vhd
index 81072b1f..03a5fe1b 100644
--- a/hdl/pcie_altera.vhd
+++ b/hdl/pcie_altera.vhd
@@ -23,9 +23,14 @@ entity pcie_altera is
rx_wb_dat_o : out std_logic_vector(31 downto 0);
rx_wb_stall_i : in std_logic;
- tx_wb_stb_i : in std_logic;
- tx_wb_dat_i : in std_logic_vector(31 downto 0);
- tx_wb_stall_o : out std_logic);
+ -- pre-allocate buffer space used for TX
+ tx_rdy_o : out std_logic;
+ tx_alloc_i : in std_logic; -- may only set '1' if rdy_o = '1'
+
+ -- push TX data
+ tx_en_i : in std_logic; -- may never exceed alloc_i
+ tx_dat_i : in std_logic_vector(31 downto 0);
+ tx_eop_i : in std_logic); -- Mark last strobe
end pcie_altera;
architecture rtl of pcie_altera is
@@ -188,6 +193,15 @@ architecture rtl of pcie_altera is
return '0';
end if;
end is_zero;
+
+ function active_high(x : boolean) return std_logic is
+ begin
+ if x then
+ return '1';
+ else
+ return '0';
+ end if;
+ end active_high;
signal core_clk_out : std_logic;
signal rstn : std_logic;
@@ -205,6 +219,8 @@ architecture rtl of pcie_altera is
signal npor, crst, srst, rst_reg : std_logic;
signal pme_shift : std_logic_vector(4 downto 0);
+ -- RX registers and signals
+
signal rx_st_ready0, rx_st_valid0 : std_logic;
signal rx_st_be0 : std_logic_vector(7 downto 0);
signal rx_st_data0 : std_logic_vector(63 downto 0);
@@ -215,6 +231,31 @@ architecture rtl of pcie_altera is
signal r32_word, s32_word, s32_progress, r32_full, s32_need_refill, r32_skip, s32_enter0 : std_logic;
signal r32_dat0, r32_dat1 : std_logic_vector(31 downto 0);
+
+ -- TX registers and signals
+
+ constant log_bytes : integer := 9; -- 256 byte maximum TLP, but we allocate twice the space to simplify provisioning
+ constant buf_length : integer := (2**log_bytes)/8;
+ constant buf_bits : integer := log_bytes-3;
+ type queue_t is array(buf_length-1 downto 0) of std_logic_vector(64 downto 0);
+
+ signal tx_st_sop0, tx_st_eop0, tx_st_ready0, tx_st_valid0 : std_logic;
+ signal tx_st_data0 : std_logic_vector(63 downto 0);
+ signal s_eop, tx_queue_stall : std_logic;
+ signal r_sop : std_logic := '1';
+
+ signal queue : queue_t;
+
+ -- Invariant idxr <= idxe <= idxw <= idxa, extra bit is for wrap-around
+ signal r_idxr, r_idxw, r_idxa, r_idxe, s_idxw_p1 : unsigned(buf_bits downto 0);
+ signal r_delay_ready : std_logic_vector(1 downto 0); -- length must equal the latency of the Avalon TX bus
+
+ signal s_queue_wdat : std_logic_vector(63 downto 0);
+ signal s_queue_wen, s_64to32_full, r_tx32_full, r_pad : std_logic;
+
+ constant zero32 : std_logic_vector(31 downto 0) := (others => '0');
+ signal r_tx_dat0 : std_logic_vector(31 downto 0);
+
begin
reconfig_clk <= cal_clk50_i;
@@ -277,12 +318,12 @@ begin
r2c_err0 => open,
-- Avalon TX
- tx_st_data0 => (others => '0'),
- tx_st_eop0 => '0',
+ tx_st_data0 => tx_st_data0,
+ tx_st_eop0 => tx_st_eop0,
tx_st_err0 => '0',
- tx_st_sop0 => '0',
- tx_st_valid0 => '0',
- tx_st_ready0 => open,
+ tx_st_sop0 => tx_st_sop0,
+ tx_st_valid0 => tx_st_valid0,
+ tx_st_ready0 => tx_st_ready0,
tx_fifo_empty0 => open,
tx_fifo_full0 => open,
tx_fifo_rdptr0 => open, -- 3 downto 0
@@ -490,7 +531,7 @@ begin
-- Issue a fetch only if we need refill and no fetch is pending
rx_st_ready0 <= s64_need_refill and is_zero(r64_ready(r64_ready'length-2 downto 0));
- rx_data64: process(core_clk_out)
+ rx_data64 : process(core_clk_out)
begin
if rising_edge(core_clk_out) then
if rstn = '0' then
@@ -505,4 +546,83 @@ begin
r64_skip <= s64_skip;
end if;
end process;
+
+ -- TX queue
+ tx_st_data0 <= queue(to_integer(r_idxr(buf_bits-1 downto 0)))(63 downto 0);
+ s_eop <= queue(to_integer(r_idxr(buf_bits-1 downto 0)))(64);
+ tx_st_eop0 <= s_eop;
+ tx_st_sop0 <= r_sop;
+
+ tx_st_valid0 <= active_high(r_idxr /= r_idxe) and r_delay_ready(r_delay_ready'length-1);
+
+ tx_data64_r : process(core_clk_out)
+ begin
+ if rising_edge(core_clk_out) then
+ if rstn = '0' then
+ r_delay_ready <= (others => '0');
+ r_idxr <= (others => '0');
+ r_sop <= '1';
+ else
+ r_delay_ready <= r_delay_ready(r_delay_ready'length-2 downto 0) & tx_st_ready0;
+ if tx_st_valid0 = '1' then
+ r_idxr <= r_idxr + 1;
+ r_sop <= s_eop;
+ end if;
+ end if;
+ end if;
+ end process;
+
+ -- can only accept data if A pointer has not wrapped around the buffer to point at the R pointer
+ tx_rdy_o <= active_high(r_idxa(buf_bits-1 downto 0) = r_idxr(buf_bits-1 downto 0)) and
+ active_high(r_idxa(buf_bits) /= r_idxr(buf_bits));
+
+ s_idxw_p1 <= r_idxw + 1;
+ tx_data64_w : process(core_clk_out)
+ begin
+ if rising_edge(core_clk_out) then
+ if rstn = '0' then
+ r_idxw <= (others => '0');
+ r_idxa <= (others => '0');
+ r_idxe <= (others => '0');
+ else
+ queue(to_integer(r_idxw(buf_bits-1 downto 0))) <= tx_eop_i & s_queue_wdat;
+
+ if s_queue_wen = '1' then
+ r_idxw <= s_idxw_p1;
+ end if;
+
+ if (s_queue_wen and tx_eop_i) = '1' then
+ r_idxe <= s_idxw_p1;
+ r_idxa <= s_idxw_p1; -- clear over-allocation
+ end if;
+
+ if tx_alloc_i = '1' then
+ r_idxa <= r_idxa + 1;
+ end if;
+ end if;
+ end if;
+ end process;
+
+ s_queue_wdat <=
+ (zero32 & tx_dat_i) when r_tx32_full = '0' else
+ (tx_dat_i & r_tx_dat0);
+
+ s_64to32_full <= r_tx32_full or r_pad or tx_eop_i;
+ s_queue_wen <= tx_en_i and s_64to32_full;
+
+ tx_data32 : process(core_clk_out)
+ begin
+ if rising_edge(core_clk_out) then
+ if rstn = '0' then
+ r_tx_dat0 <= (others => '0');
+ r_tx32_full <= '0';
+ r_pad <= '0';
+ else
+ if tx_en_i = '1' then
+ r_tx_dat0 <= tx_dat_i;
+ r_tx32_full <= not s_64to32_full;
+ end if;
+ end if;
+ end if;
+ end process;
end rtl;
diff --git a/hdl/pcie_tlp.vhd b/hdl/pcie_tlp.vhd
index d941698d..b12bef5a 100644
--- a/hdl/pcie_tlp.vhd
+++ b/hdl/pcie_tlp.vhd
@@ -12,12 +12,21 @@ entity pcie_tlp is
rx_wb_dat_i : in std_logic_vector(31 downto 0);
rx_wb_stall_o : out std_logic;
+ tx_rdy_i : out std_logic;
+ tx_alloc_o : in std_logic;
+ tx_en_o : in std_logic;
+ tx_dat_o : in std_logic_vector(31 downto 0);
+ tx_eop_o : in std_logic;
+
wb_stb_o : out std_logic;
wb_adr_o : out std_logic_vector(63 downto 0);
wb_we_o : out std_logic;
wb_dat_o : out std_logic_vector(31 downto 0);
wb_sel_o : out std_logic_vector(3 downto 0);
- wb_stall_i : in std_logic);
+ wb_stall_i : in std_logic;
+ wb_ack_i : in std_logic;
+ wb_err_i : in std_logic;
+ wb_dat_i : in std_logic_vector(31 downto 0));
end pcie_tlp;
architecture rtl of pcie_tlp is
@@ -54,9 +63,14 @@ architecture rtl of pcie_tlp is
-- Stall and strobe bypass mux
signal r_always_stall, r_never_stall : std_logic;
signal r_always_stb, r_never_stb : std_logic;
+
+ -- Inflight reads and writes
+ signal wb_stb : std_logic;
+ signal r_flight_count : unsigned(4 downto 0);
begin
rx_wb_stall_o <= r_always_stall or (not r_never_stall and wb_stall_i);
- wb_stb_o <= r_always_stb or (not r_never_stb and rx_wb_stb_i);
+ wb_stb <= r_always_stb or (not r_never_stb and rx_wb_stb_i);
+ wb_stb_o <= wb_stb;
wb_adr_o <= r_address;
wb_dat_o <= rx_wb_dat_i;
@@ -251,4 +265,23 @@ begin
end if;
end if;
end process;
+
+ flight_counter : process(clk_i)
+ begin
+ if rising_edge(clk_i) then
+ if (wb_ack_i or wb_err_i) = '1' then
+ if wb_stb = '1' then
+ r_flight_count <= r_flight_count;
+ else
+ r_flight_count <= r_flight_count - 1;
+ end if;
+ else
+ if wb_stb = '1' then
+ r_flight_count <= r_flight_count + 1;
+ else
+ r_flight_count <= r_flight_count;
+ end if;
+ end if;
+ end if;
+ end process;
end rtl;
diff --git a/hdl/pcie_wb.vhd b/hdl/pcie_wb.vhd
index 7ebba6bb..9d597d4b 100644
--- a/hdl/pcie_wb.vhd
+++ b/hdl/pcie_wb.vhd
@@ -46,6 +46,13 @@ architecture rtl of pcie_wb is
signal rx_wb_stb, rx_wb_stall : std_logic;
signal rx_wb_dat : std_logic_vector(31 downto 0);
+
+ signal tx_rdy, tx_alloc, tx_en, tx_eop : std_logic;
+ signal tx_dat : std_logic_vector(31 downto 0);
+
+ signal wb_stb_o, wb_we_o, wb_ack_i : std_logic;
+ signal wb_dat_o, wb_dat_i, demo_reg : std_logic_vector(31 downto 0);
+
begin
reset : pow_reset
@@ -76,10 +83,11 @@ begin
rx_wb_stb_o => rx_wb_stb,
rx_wb_dat_o => rx_wb_dat,
rx_wb_stall_i => rx_wb_stall,
- -- No TX... yet.
- tx_wb_stb_i => '0',
- tx_wb_dat_i => (others => '0'),
- tx_wb_stall_o => open);
+ tx_rdy_o => tx_rdy,
+ tx_alloc_i => tx_alloc,
+ tx_en_i => tx_en,
+ tx_dat_i => tx_dat,
+ tx_eop_i => tx_eop);
pcie_logic : pcie_tlp port map(
clk_i => wb_clk,
@@ -90,12 +98,26 @@ begin
rx_wb_dat_i => rx_wb_dat,
rx_wb_stall_o => rx_wb_stall,
- wb_stb_o => open,
+ wb_stb_o => wb_stb_o,
wb_adr_o => open,
- wb_we_o => open,
- wb_dat_o => open,
+ wb_we_o => wb_we_o,
+ wb_dat_o => wb_dat_o,
wb_sel_o => open,
- wb_stall_i => stall);
+ wb_stall_i => stall,
+ wb_ack_i => wb_ack_i,
+ wb_err_i => '0',
+ wb_dat_i => wb_dat_i);
+
+ wb_dat_i <= demo_reg;
+ demo : process(wb_clk)
+ begin
+ if rising_edge(wb_clk) then
+ if (wb_stb_o and wb_we_o and not stall) = '1' then
+ demo_reg <= wb_dat_o;
+ end if;
+ wb_ack_i <= wb_stb_o and not stall;
+ end if;
+ end process;
blink : process(wb_clk)
begin
diff --git a/hdl/pcie_wb_pkg.vhd b/hdl/pcie_wb_pkg.vhd
index ccfc60cc..8e3f7cba 100644
--- a/hdl/pcie_wb_pkg.vhd
+++ b/hdl/pcie_wb_pkg.vhd
@@ -24,9 +24,14 @@ package pcie_wb_pkg is
rx_wb_dat_o : out std_logic_vector(31 downto 0);
rx_wb_stall_i : in std_logic;
- tx_wb_stb_i : in std_logic;
- tx_wb_dat_i : in std_logic_vector(31 downto 0);
- tx_wb_stall_o : out std_logic);
+ -- pre-allocate buffer space used for TX
+ tx_rdy_o : out std_logic;
+ tx_alloc_i : in std_logic; -- may only set '1' if rdy_o = '1'
+
+ -- push TX data
+ tx_en_i : in std_logic; -- may never exceed alloc_i
+ tx_dat_i : in std_logic_vector(31 downto 0);
+ tx_eop_i : in std_logic); -- Mark last strobe
end component;
component pcie_tlp is
@@ -39,11 +44,20 @@ package pcie_wb_pkg is
rx_wb_dat_i : in std_logic_vector(31 downto 0);
rx_wb_stall_o : out std_logic;
+ tx_rdy_i : out std_logic;
+ tx_alloc_o : in std_logic;
+ tx_en_o : in std_logic;
+ tx_dat_o : in std_logic_vector(31 downto 0);
+ tx_eop_o : in std_logic;
+
wb_stb_o : out std_logic;
wb_adr_o : out std_logic_vector(63 downto 0);
wb_we_o : out std_logic;
wb_dat_o : out std_logic_vector(31 downto 0);
wb_sel_o : out std_logic_vector(3 downto 0);
- wb_stall_i : in std_logic);
+ wb_stall_i : in std_logic;
+ wb_ack_i : in std_logic;
+ wb_err_i : in std_logic;
+ wb_dat_i : in std_logic_vector(31 downto 0));
end component;
end pcie_wb_pkg;
--
GitLab