litedram: l2: Latency improvements

This implements in the L2 cache the feature already in the L1s
allowing a request to be completed before the end of a refill
using partial line valid bits, and starting a refill from the
row of the first miss on that line instead of the beginning of
the line.

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
pull/229/head
Benjamin Herrenschmidt 4 years ago
parent b0241d9f2d
commit 1441b2a859

@ -293,6 +293,43 @@ begin
check_data(make_pattern(i));
end loop;

report "Pre-fill a line";
a(11) := '1';
clr_acks;
wb_write(add_off(a, 0), x"1111111100000000", x"ff");
wb_write(add_off(a, 8), x"3333333322222222", x"ff");
wb_write(add_off(a, 16), x"5555555544444444", x"ff");
wb_write(add_off(a, 24), x"7777777766666666", x"ff");
wb_write(add_off(a, 32), x"9999999988888888", x"ff");
wb_write(add_off(a, 40), x"bbbbbbbbaaaaaaaa", x"ff");
wb_write(add_off(a, 48), x"ddddddddcccccccc", x"ff");
wb_write(add_off(a, 56), x"ffffffffeeeeeeee", x"ff");
wb_write(add_off(a, 64), x"1111111100000000", x"ff");
wb_write(add_off(a, 72), x"3333333322222222", x"ff");
wb_write(add_off(a, 80), x"5555555544444444", x"ff");
wb_write(add_off(a, 88), x"7777777766666666", x"ff");
wb_write(add_off(a, 96), x"9999999988888888", x"ff");
wb_write(add_off(a,104), x"bbbbbbbbaaaaaaaa", x"ff");
wb_write(add_off(a,112), x"ddddddddcccccccc", x"ff");
wb_write(add_off(a,120), x"ffffffffeeeeeeee", x"ff");
wait_acks(16);

report "Scattered from middle of line...";
clr_acks;
wb_read(add_off(a,24));
wb_read(add_off(a,32));
wb_read(add_off(a, 0));
wb_read(add_off(a,16));
wait_acks(4);
read_data(d);
assert d = x"7777777766666666" report "bad data (24), got " & to_hstring(d) severity failure;
read_data(d);
assert d = x"9999999988888888" report "bad data (32), got " & to_hstring(d) severity failure;
read_data(d);
assert d = x"1111111100000000" report "bad data (0), got " & to_hstring(d) severity failure;
read_data(d);
assert d = x"5555555544444444" report "bad data (16), got " & to_hstring(d) severity failure;

std.env.finish;
end process;
end architecture;

@ -189,6 +189,7 @@ architecture behaviour of litedram_wrapper is
subtype row_t is integer range 0 to BRAM_ROWS-1;
subtype index_t is integer range 0 to NUM_LINES-1;
subtype way_t is integer range 0 to NUM_WAYS-1;
subtype row_in_line_t is unsigned(ROW_LINEBITS-1 downto 0);

-- The cache data BRAM organized as described above for each way
subtype cache_row_t is std_ulogic_vector(DRAM_DBITS-1 downto 0);
@ -207,6 +208,9 @@ architecture behaviour of litedram_wrapper is
subtype cache_way_valids_t is std_ulogic_vector(NUM_WAYS-1 downto 0);
type cache_valids_t is array(index_t) of cache_way_valids_t;

-- "Temporary" valid bits for the rows of the currently refilled line
type row_per_line_valid_t is array(0 to ROW_PER_LINE - 1) of std_ulogic;

-- Storage. Hopefully "cache_rows" is a BRAM, the rest is LUTs
signal cache_tags : cache_tags_array_t;
signal cache_valids : cache_valids_t;
@ -233,7 +237,7 @@ architecture behaviour of litedram_wrapper is
-- Cache management signals
--

-- Cache state machine
-- Cache state machine
type state_t is (IDLE, -- Normal load hit processing
REFILL_CLR_TAG, -- Cache refill clear tag
REFILL_WAIT_ACK); -- Cache refill wait ack
@ -261,7 +265,8 @@ architecture behaviour of litedram_wrapper is
OP_LOAD_HIT,
OP_LOAD_MISS,
OP_STORE_HIT,
OP_STORE_MISS);
OP_STORE_MISS,
OP_STORE_DELAYED);

signal req_index : index_t;
signal req_row : row_t;
@ -272,7 +277,6 @@ architecture behaviour of litedram_wrapper is
signal req_ad3 : std_ulogic;
signal req_we : std_ulogic_vector(DRAM_SBITS-1 downto 0);
signal req_wdata : std_ulogic_vector(DRAM_DBITS-1 downto 0);
signal accept_store : std_ulogic;
signal stall : std_ulogic;

-- Line refill command signals and latches
@ -281,6 +285,8 @@ architecture behaviour of litedram_wrapper is
signal refill_way : way_t;
signal refill_index : index_t;
signal refill_row : row_t;
signal refill_end_row : row_in_line_t;
signal refill_rows_vlid : row_per_line_valid_t;

-- Cache RAM interface
type cache_ram_out_t is array(way_t) of cache_row_t;
@ -306,21 +312,25 @@ architecture behaviour of litedram_wrapper is
return to_integer(unsigned(addr(SET_SIZE_BITS - 1 downto ROW_OFF_BITS)));
end;

-- Return the index of a row within a line
function get_row_of_line(row: row_t) return row_in_line_t is
variable row_v : unsigned(ROW_BITS-1 downto 0);
begin
row_v := to_unsigned(row, ROW_BITS);
return row_v(ROW_LINEBITS-1 downto 0);
end;
-- Returns whether this is the last row of a line. It takes a DRAM address
function is_last_row_addr(addr: std_ulogic_vector(REAL_ADDR_BITS-1 downto ROW_OFF_BITS))
function is_last_row_addr(addr: std_ulogic_vector(REAL_ADDR_BITS-1 downto ROW_OFF_BITS);
last: row_in_line_t)
return boolean is
constant ones : std_ulogic_vector(ROW_LINEBITS-1 downto 0) := (others => '1');
begin
return addr(LINE_OFF_BITS-1 downto ROW_OFF_BITS) = ones;
return unsigned(addr(LINE_OFF_BITS-1 downto ROW_OFF_BITS)) = last;
end;

-- Returns whether this is the last row of a line
function is_last_row(row: row_t) return boolean is
variable row_v : std_ulogic_vector(ROW_BITS-1 downto 0);
constant ones : std_ulogic_vector(ROW_LINEBITS-1 downto 0) := (others => '1');
function is_last_row(row: row_t; last: row_in_line_t) return boolean is
begin
row_v := std_ulogic_vector(to_unsigned(row, ROW_BITS));
return row_v(ROW_LINEBITS-1 downto 0) = ones;
return get_row_of_line(row) = last;
end;

-- Return the address of the next row in the current cache line. It takes a
@ -493,7 +503,7 @@ begin
--
-- Write mux: cache refills from DRAM or writes from Wishbone
--
if state = IDLE then
if req_op = OP_STORE_HIT and req_hit_way = i then
-- Write from wishbone
wr_addr <= std_ulogic_vector(to_unsigned(req_row, ROW_BITS));
wr_data <= req_wdata;
@ -570,7 +580,7 @@ begin
end generate;

--
-- Wishbone interface:
-- Wishbone request interface:
--
-- - Incoming wishbone request latch (to help with timing)
-- - Read response pipeline (to match BRAM output buffer delay)
@ -632,7 +642,6 @@ begin
--
-- Read response pipeline
--
-- XXX Might have to put store acks in there too (see comment in wb_response)
read_pipe: process(system_clk)
begin
if rising_edge(system_clk) then
@ -670,54 +679,46 @@ begin
end if;
end process;

--
-- Wishbone response generation
--

wb_reponse: process(all)
variable rdata : std_ulogic_vector(DRAM_DBITS-1 downto 0);
variable store_done : std_ulogic;
variable rdata : std_ulogic_vector(DRAM_DBITS-1 downto 0);
variable store_done : std_ulogic;
variable accept_store : std_ulogic;
begin
-- Can we accept a store ? This is set when IDLE and the store
-- queue & command queue are not full.
-- Can we accept a store ? This is set when the store queue & command
-- queue are not full.
--
-- Note: This is only used to control the WB request latch, stall
-- and store "early complete". We don't want to use this to control
-- cmd_valid to DRAM as this would create a circular dependency inside
-- LiteDRAM as cmd_ready I think is driven from cmd_valid.
-- This does *not* mean that we will accept the store, there are other
-- reasons to delay them (see OP_STORE_DELAYED).
--
-- The state machine that controls the command queue must thus
-- reproduce this logic at least partially.
-- A store is fully accepted when *both* req_op is not OP_STORE_DELAYED
-- and accept_store is '1'.
--
-- Note also that user_port0_cmd_ready from LiteDRAM is combinational
-- from user_port0_cmd_valid. IE. we won't know that LiteDRAM cannot
-- accept a command until we try to send one.
-- The reason for this split is to avoid a circular dependency inside
-- LiteDRAM, since cmd_ready from litedram is driven from cmd_valid (*)
-- we don't want to generate cmd_valid from cmd_ready. So we generate
-- it instead from all the *other* conditions that make a store valid.
--
if state = IDLE then
accept_store <= user_port0_cmd_ready and storeq_wr_ready;

-- Corner case !!! The read acks pipeline takes two extra cycles
-- which means a store ack can collide with a previous load hit
-- ack. Thus we stall stores if we have a load ack pending.
if read_ack_0 = '1' or read_ack_1 = '1' then
accept_store <= '0';
end if;
-- (*) It's my understanding that user_port0_cmd_ready from LiteDRAM is
-- ombinational from user_port0_cmd_valid along with a bunch of other
-- internal signals. IE. we won't know that LiteDRAM cannot accept a
-- command until we try to send one.
--
accept_store := user_port0_cmd_ready and storeq_wr_ready;

-- Generate stalls. For stores we stall if we can't accept it.
-- For loads, we stall if we are going to take a load miss or
-- are in the middle of a refill and it isn't a partial hit.
if req_op = OP_STORE_MISS or req_op = OP_STORE_HIT then
stall <= not accept_store;
elsif req_op = OP_LOAD_MISS or req_op = OP_STORE_DELAYED then
stall <= '1';
else
accept_store <= '0';
stall <= '0';
end if;

-- Generate stalls. For loads, we stall if we are going to take a load
-- miss or are in the middle of a refill. For stores, if we can't
-- accept it.
case state is
when IDLE =>
case req_op is
when OP_LOAD_MISS =>
stall <= '1';
when OP_STORE_MISS | OP_STORE_HIT =>
stall <= not accept_store;
when others =>
stall <= '0';
end case;
when others =>
stall <= '1';
end case;
-- Data out mux
rdata := cache_out(read_way_1);
@ -736,7 +737,8 @@ begin
-- Generate Wishbone ACKs on read hits and store complete
--
-- This can happen on store right behind loads ! This is why
-- we don't accept a new store right behind a load ack above.
-- we delay a store when a load ack is in the pipeline in the
-- request decoder below.
--
wb_out.ack <= read_ack_1 or store_ack_1;
assert read_ack_1 = '0' or store_ack_1 = '0' report
@ -748,27 +750,24 @@ begin
-- Cache request decode
--
request_decode: process(all)
variable valid : std_ulogic;
variable is_hit : std_ulogic;
variable hit_way : way_t;
variable valid : boolean;
variable is_hit : boolean;
variable store_delay : boolean;
variable hit_way : way_t;
begin
-- Extract line, row and tag from request
req_index <= get_index(wb_req.adr);
req_row <= get_row(wb_req.adr(REAL_ADDR_BITS-1 downto 0));
req_tag <= get_tag(wb_req.adr);

-- Calculate address of beginning of cache line, will be
-- Calculate address of beginning of cache row, will be
-- used for cache miss processing if needed
req_laddr <= wb_req.adr(REAL_ADDR_BITS - 1 downto LINE_OFF_BITS) &
(LINE_OFF_BITS-1 downto 0 => '0');
req_laddr <= wb_req.adr(REAL_ADDR_BITS - 1 downto ROW_OFF_BITS) &
(ROW_OFF_BITS-1 downto 0 => '0');


-- Do we have a valid request in the WB latch ?
if state = IDLE then
valid := wb_req.cyc and wb_req.stb;
else
valid := '0';
end if;
valid := wb_req.cyc = '1' and wb_req.stb = '1';

-- Store signals
req_ad3 <= wb_req.adr(3);
@ -778,30 +777,67 @@ begin

-- Test if pending request is a hit on any way
hit_way := 0;
is_hit := '0';
is_hit := false;
for i in way_t loop
if valid = '1' and cache_valids(req_index)(i) = '1' then
if valid and
(cache_valids(req_index)(i) = '1' or
(state = REFILL_WAIT_ACK and
req_index = refill_index and i = refill_way and
refill_rows_vlid(req_row mod ROW_PER_LINE) = '1')) then
if read_tag(i, cache_tags(req_index)) = req_tag then
hit_way := i;
is_hit := '1';
is_hit := true;
end if;
end if;
end loop;

-- We need to delay stores under some circumstances to avoid
-- collisions with the refill machine.
--
-- Corner case !!! The read acks pipeline takes two extra cycles
-- which means a store ack can collide with a previous load hit
-- ack. Thus we stall stores if we have a load ack pending.
--
if read_ack_0 = '1' or read_ack_1 = '1' then
-- Clash with pending read acks, delay..
store_delay := true;
elsif state /= IDLE then
-- If the reload machine is active, we cannot accept a store
-- for now.
--
-- We could improve this a bit by allowing stores if we have sent
-- all the requests down to litedram (we are only waiting for the
-- responses) *and* either of those conditions is true:
--
-- * It's a miss (doesn't require a write to BRAM) and isn't
-- for the line being reloaded (otherwise we might reload
-- stale data into the cache).
-- * It's a hit on a different way than the one being reloaded
-- in which case there is no conflict for BRAM access.
--
-- Otherwise we delay it...
--
store_delay := true;
else
store_delay := false;
end if;

-- Generate the req op. We only allow OP_LOAD_* when in the
-- IDLE state as our PLRU and ACK generation rely on this,
-- stores are allowed in IDLE state.
--
req_op <= OP_NONE;
if valid = '1' then
if valid then
if wb_req.we = '1' then
if is_hit = '1' then
if store_delay then
req_op <= OP_STORE_DELAYED;
elsif is_hit then
req_op <= OP_STORE_HIT;
else
req_op <= OP_STORE_MISS;
end if;
else
if is_hit = '1' then
if is_hit then
req_op <= OP_LOAD_HIT;
else
req_op <= OP_LOAD_MISS;
@ -837,7 +873,7 @@ begin
begin
storeq_wr_data <= wb_req.dat & req_we;

-- Only accept store if we can send a command
-- Only queue stores if we can also send a command
if req_op = OP_STORE_HIT or req_op = OP_STORE_MISS then
storeq_wr_valid <= user_port0_cmd_ready;
else
@ -858,13 +894,13 @@ begin
to_hstring(wb_req.adr(DRAM_ABITS+3 downto 0)) &
" data:" & to_hstring(req_wdata) &
" we:" & to_hstring(req_we) &
" V:" & std_ulogic'image(accept_store);
" V:" & std_ulogic'image(user_port0_cmd_ready);
else
report "Store miss to:" &
to_hstring(wb_req.adr(DRAM_ABITS+3 downto 0)) &
" data:" & to_hstring(req_wdata) &
" we:" & to_hstring(req_we) &
" V:" & std_ulogic'image(accept_store);
" V:" & std_ulogic'image(user_port0_cmd_ready);
end if;
if storeq_wr_valid = '1' and storeq_wr_ready = '1' then
report "storeq push " & to_hstring(storeq_wr_data);
@ -879,9 +915,9 @@ begin
-- LiteDRAM command mux
dram_commands: process(all)
begin
if state = IDLE and (req_op = OP_STORE_HIT or req_op = OP_STORE_MISS) then
if req_op = OP_STORE_HIT or req_op = OP_STORE_MISS then
-- For stores, forward signals directly. Only send command if
-- the FIFO can accept a store
-- the FIFO can accept a store.
user_port0_cmd_addr <= wb_req.adr(DRAM_ABITS+3 downto 4);
user_port0_cmd_we <= '1';
user_port0_cmd_valid <= storeq_wr_ready;
@ -918,6 +954,11 @@ begin
assert refill_cmd_valid = '0' report "refill cmd valid in IDLE state !"
severity failure;

-- Reset per-row valid flags, only used in WAIT_ACK
for i in 0 to ROW_PER_LINE - 1 loop
refill_rows_vlid(i) <= '0';
end loop;

-- If NO_LS_OVERLAP is set, disallow a load miss if the store
-- queue still has data in it.
wait_qdrain := false;
@ -931,8 +972,9 @@ begin
refill_way <= to_integer(unsigned(plru_victim(req_index)));

-- Keep track of our index and way for subsequent stores
refill_index <= req_index;
refill_row <= get_row(req_laddr);
refill_index <= req_index;
refill_row <= get_row(req_laddr);
refill_end_row <= get_row_of_line(get_row(req_laddr)) - 1;

-- Prep for first DRAM read
--
@ -982,7 +1024,7 @@ begin
if TRACE then
report "got refill cmd ack !";
end if;
if is_last_row_addr(refill_cmd_addr) then
if is_last_row_addr(refill_cmd_addr, refill_end_row) then
refill_cmd_valid <= '0';
cmds_done := true;
if TRACE then
@ -1003,8 +1045,12 @@ begin
if TRACE then
report "got refill data ack !";
end if;

-- Mark partial line valid
refill_rows_vlid(refill_row mod ROW_PER_LINE) <= '1';

-- Check for completion
if cmds_done and is_last_row(refill_row) then
if cmds_done and is_last_row(refill_row, refill_end_row) then
if TRACE then
report "all refill data done !";
end if;

@ -1,17 +1,18 @@
[*]
[*] GTKWave Analyzer v3.3.86 (w)1999-2017 BSI
[*] Sun May 31 12:53:52 2020
[*] Mon Jun 22 06:32:16 2020
[*]
[dumpfile] "/home/ANT.AMAZON.COM/benh/hackplace/microwatt/foo.ghw"
[dumpfile_mtime] "Sun May 31 12:50:15 2020"
[dumpfile_size] 1134118
[dumpfile_mtime] "Mon Jun 22 06:28:35 2020"
[dumpfile_size] 1680014
[savefile] "/home/ANT.AMAZON.COM/benh/hackplace/microwatt/litedram/extras/wave_tb.gtkw"
[timestart] 1312950000
[timestart] 1920580000
[size] 2509 1371
[pos] -1 -1
*-24.248457 1386890000 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1
*-24.248457 1935000000 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1
[treeopen] top.
[treeopen] top.dram_tb.
[treeopen] top.dram_tb.dram.
[sst_width] 301
[signals_width] 433
[sst_expanded] 1
@ -26,7 +27,6 @@ top.dram_tb.clk
@22
#{top.dram_tb.wb_in.dat[63:0]} top.dram_tb.wb_in.dat[63] top.dram_tb.wb_in.dat[62] top.dram_tb.wb_in.dat[61] top.dram_tb.wb_in.dat[60] top.dram_tb.wb_in.dat[59] top.dram_tb.wb_in.dat[58] top.dram_tb.wb_in.dat[57] top.dram_tb.wb_in.dat[56] top.dram_tb.wb_in.dat[55] top.dram_tb.wb_in.dat[54] top.dram_tb.wb_in.dat[53] top.dram_tb.wb_in.dat[52] top.dram_tb.wb_in.dat[51] top.dram_tb.wb_in.dat[50] top.dram_tb.wb_in.dat[49] top.dram_tb.wb_in.dat[48] top.dram_tb.wb_in.dat[47] top.dram_tb.wb_in.dat[46] top.dram_tb.wb_in.dat[45] top.dram_tb.wb_in.dat[44] top.dram_tb.wb_in.dat[43] top.dram_tb.wb_in.dat[42] top.dram_tb.wb_in.dat[41] top.dram_tb.wb_in.dat[40] top.dram_tb.wb_in.dat[39] top.dram_tb.wb_in.dat[38] top.dram_tb.wb_in.dat[37] top.dram_tb.wb_in.dat[36] top.dram_tb.wb_in.dat[35] top.dram_tb.wb_in.dat[34] top.dram_tb.wb_in.dat[33] top.dram_tb.wb_in.dat[32] top.dram_tb.wb_in.dat[31] top.dram_tb.wb_in.dat[30] top.dram_tb.wb_in.dat[29] top.dram_tb.wb_in.dat[28] top.dram_tb.wb_in.dat[27] top.dram_tb.wb_in.dat[26] top.dram_tb.wb_in.dat[25] top.dram_tb.wb_in.dat[24] top.dram_tb.wb_in.dat[23] top.dram_tb.wb_in.dat[22] top.dram_tb.wb_in.dat[21] top.dram_tb.wb_in.dat[20] top.dram_tb.wb_in.dat[19] top.dram_tb.wb_in.dat[18] top.dram_tb.wb_in.dat[17] top.dram_tb.wb_in.dat[16] top.dram_tb.wb_in.dat[15] top.dram_tb.wb_in.dat[14] top.dram_tb.wb_in.dat[13] top.dram_tb.wb_in.dat[12] top.dram_tb.wb_in.dat[11] top.dram_tb.wb_in.dat[10] top.dram_tb.wb_in.dat[9] top.dram_tb.wb_in.dat[8] top.dram_tb.wb_in.dat[7] top.dram_tb.wb_in.dat[6] top.dram_tb.wb_in.dat[5] top.dram_tb.wb_in.dat[4] top.dram_tb.wb_in.dat[3] top.dram_tb.wb_in.dat[2] top.dram_tb.wb_in.dat[1] top.dram_tb.wb_in.dat[0]
#{top.dram_tb.wb_in.adr[31:0]} top.dram_tb.wb_in.adr[31] top.dram_tb.wb_in.adr[30] top.dram_tb.wb_in.adr[29] top.dram_tb.wb_in.adr[28] top.dram_tb.wb_in.adr[27] top.dram_tb.wb_in.adr[26] top.dram_tb.wb_in.adr[25] top.dram_tb.wb_in.adr[24] top.dram_tb.wb_in.adr[23] top.dram_tb.wb_in.adr[22] top.dram_tb.wb_in.adr[21] top.dram_tb.wb_in.adr[20] top.dram_tb.wb_in.adr[19] top.dram_tb.wb_in.adr[18] top.dram_tb.wb_in.adr[17] top.dram_tb.wb_in.adr[16] top.dram_tb.wb_in.adr[15] top.dram_tb.wb_in.adr[14] top.dram_tb.wb_in.adr[13] top.dram_tb.wb_in.adr[12] top.dram_tb.wb_in.adr[11] top.dram_tb.wb_in.adr[10] top.dram_tb.wb_in.adr[9] top.dram_tb.wb_in.adr[8] top.dram_tb.wb_in.adr[7] top.dram_tb.wb_in.adr[6] top.dram_tb.wb_in.adr[5] top.dram_tb.wb_in.adr[4] top.dram_tb.wb_in.adr[3] top.dram_tb.wb_in.adr[2] top.dram_tb.wb_in.adr[1] top.dram_tb.wb_in.adr[0]
@23
#{top.dram_tb.wb_in.sel[7:0]} top.dram_tb.wb_in.sel[7] top.dram_tb.wb_in.sel[6] top.dram_tb.wb_in.sel[5] top.dram_tb.wb_in.sel[4] top.dram_tb.wb_in.sel[3] top.dram_tb.wb_in.sel[2] top.dram_tb.wb_in.sel[1] top.dram_tb.wb_in.sel[0]
@28
top.dram_tb.wb_in.cyc
@ -67,6 +67,9 @@ top.dram_tb.dram.user_port0_cmd_valid
top.dram_tb.dram.refill_cmd_valid
@420
top.dram_tb.dram.req_index
@421
top.dram_tb.dram.req_row
@420
top.dram_tb.dram.req_hit_way
@28
top.dram_tb.dram.req_ad3

Loading…
Cancel
Save