class DataFormat(Elaboratable):
def __init__(self):
# inputs
self.x_funct3 = Signal(3)
self.x_offset = Signal(2)
self.m_offset = Signal(2)
self.x_store_data = Signal(32) # raw data to store
self.m_data_r = Signal(32) # raw data from load
self.m_funct3 = Signal(3)
# outputs
self.x_byte_sel = Signal(4)
self.x_data_w = Signal(32) # formatted data to bus
self.x_misaligned = Signal()
self.m_load_data = Signal(32) # formatted data to pipeline
def elaborate(self, platform):
m = Module()
# create the byte selector
with m.Switch(self.x_funct3):
with m.Case(Funct3.B):
m.d.comb += self.x_byte_sel.eq(0b0001 << self.x_offset)
with m.Case(Funct3.H):
m.d.comb += self.x_byte_sel.eq(0b0011 << self.x_offset)
with m.Case(Funct3.W):
m.d.comb += self.x_byte_sel.eq(0b1111)
# format output data
with m.Switch(self.x_funct3):
with m.Case(Funct3.B):
m.d.comb += self.x_data_w.eq(Repl(self.x_store_data[:8], 4))
with m.Case(Funct3.H):
m.d.comb += self.x_data_w.eq(Repl(self.x_store_data[:16], 2))
with m.Case(Funct3.W):
m.d.comb += self.x_data_w.eq(self.x_store_data)
# format input data
_byte = Signal((8, True))
_half = Signal((16, True))
m.d.comb += [
_byte.eq(self.m_data_r.word_select(self.m_offset, 8)),
_half.eq(self.m_data_r.word_select(self.m_offset[1], 16)),
]
with m.Switch(self.m_funct3):
with m.Case(Funct3.B):
m.d.comb += self.m_load_data.eq(_byte)
with m.Case(Funct3.BU):
m.d.comb += self.m_load_data.eq(Cat(_byte,
0)) # make sign bit = 0
with m.Case(Funct3.H):
m.d.comb += self.m_load_data.eq(_half)
with m.Case(Funct3.HU):
m.d.comb += self.m_load_data.eq(Cat(_half,
0)) # make sign bit = 0
with m.Case(Funct3.W):
m.d.comb += self.m_load_data.eq(self.m_data_r)
# misalignment
with m.Switch(self.x_funct3):
with m.Case(Funct3.H, Funct3.HU):
m.d.comb += self.x_misaligned.eq(self.x_offset[0])
with m.Case(Funct3.W):
m.d.comb += self.x_misaligned.eq(self.x_offset != 0)
return m
class Cache(Elaboratable):
def __init__(
self,
nlines: int, # number of lines
nwords: int, # number of words x line x way
nways: int, # number of ways
start_addr: int = 0, # start of cacheable region
end_addr: int = 2**32, # end of cacheable region
enable_write: bool = True # enable writes to cache
) -> None:
# enable write -> data cache
if nlines == 0 or (nlines & (nlines - 1)):
raise ValueError(f'nlines must be a power of 2: {nlines}')
if nwords not in (4, 8, 16):
raise ValueError(f'nwords must be 4, 8 or 16: {nwords}')
if nways not in (1, 2):
raise ValueError(f'nways must be 1 or 2: {nways}')
self.enable_write = enable_write
self.nlines = nlines
self.nwords = nwords
self.nways = nways
self.start_addr = start_addr
self.end_addr = end_addr
offset_bits = log2_int(nwords)
line_bits = log2_int(nlines)
addr_bits = log2_int(end_addr - start_addr, need_pow2=False)
tag_bits = addr_bits - line_bits - offset_bits - 2 # -2 because word line.
extra_bits = 32 - tag_bits - line_bits - offset_bits - 2
self.pc_layout = [('byte', 2), ('offset', offset_bits),
('line', line_bits), ('tag', tag_bits)]
if extra_bits != 0:
self.pc_layout.append(('unused', extra_bits))
# -------------------------------------------------------------------------
# IO
self.s1_address = Record(self.pc_layout)
self.s1_flush = Signal()
self.s1_valid = Signal()
self.s1_stall = Signal()
self.s1_access = Signal()
self.s2_address = Record(self.pc_layout)
self.s2_evict = Signal()
self.s2_valid = Signal()
self.s2_stall = Signal()
self.s2_access = Signal()
self.s2_miss = Signal()
self.s2_rdata = Signal(32)
self.s2_re = Signal()
if enable_write:
self.s2_wdata = Signal(32)
self.s2_sel = Signal(4)
self.s2_we = Signal()
self.bus_addr = Record(self.pc_layout)
self.bus_valid = Signal()
self.bus_last = Signal()
self.bus_data = Signal(32)
self.bus_ack = Signal()
self.bus_err = Signal()
self.access_cnt = Signal(40)
self.miss_cnt = Signal(40)
# snoop bus
if not enable_write:
self.dcache_snoop = InternalSnoopPort(
name='cache_snoop'
) # RO cache. Implement the Internal snooping port
def elaborate(self, platform: Platform) -> Module:
m = Module()
snoop_addr = Record(self.pc_layout)
snoop_valid = Signal()
# -------------------------------------------------------------------------
# Performance counter
# TODO: connect to CSR's performance counter
with m.If(~self.s1_stall & self.s1_valid & self.s1_access):
m.d.sync += self.access_cnt.eq(self.access_cnt + 1)
with m.If(self.s2_valid & self.s2_miss & ~self.bus_valid
& self.s2_access):
m.d.sync += self.miss_cnt.eq(self.miss_cnt + 1)
# -------------------------------------------------------------------------
way_layout = [('data', 32 * self.nwords),
('tag', self.s1_address.tag.shape()), ('valid', 1),
('sel_lru', 1), ('snoop_hit', 1)]
if self.enable_write:
way_layout.append(('sel_we', 1))
ways = Array(
Record(way_layout, name='way_idx{}'.format(_way))
for _way in range(self.nways))
fill_cnt = Signal.like(self.s1_address.offset)
# Check hit/miss
way_hit = m.submodules.way_hit = Encoder(self.nways)
for idx, way in enumerate(ways):
m.d.comb += way_hit.i[idx].eq((way.tag == self.s2_address.tag)
& way.valid)
m.d.comb += self.s2_miss.eq(way_hit.n)
if self.enable_write:
# Asumiendo que hay un HIT, indicar que la vía que dió hit es en la cual se va a escribir
m.d.comb += ways[way_hit.o].sel_we.eq(self.s2_we & self.s2_valid)
# set the LRU
if self.nways == 1:
# One way: LRU is useless
lru = Const(0) # self.nlines
else:
# LRU es un vector de N bits, cada uno indicado el set a reemplazar
# como NWAY es máximo 2, cada LRU es de un bit
lru = Signal(self.nlines)
_lru = lru.bit_select(self.s2_address.line, 1)
write_ended = self.bus_valid & self.bus_ack & self.bus_last # err ^ ack = = 1
access_hit = ~self.s2_miss & self.s2_valid & (way_hit.o == _lru)
with m.If(write_ended | access_hit):
m.d.sync += _lru.eq(~_lru)
# read data from the cache
m.d.comb += self.s2_rdata.eq(ways[way_hit.o].data.word_select(
self.s2_address.offset, 32))
# Internal Snoop
snoop_use_cache = Signal()
snoop_tag_match = Signal()
snoop_line_match = Signal()
snoop_cancel_refill = Signal()
if not self.enable_write:
bits_range = log2_int(self.end_addr - self.start_addr,
need_pow2=False)
m.d.comb += [
snoop_addr.eq(self.dcache_snoop.addr), # aux
snoop_valid.eq(self.dcache_snoop.we & self.dcache_snoop.valid
& self.dcache_snoop.ack),
snoop_use_cache.eq(snoop_addr[bits_range:] == (
self.start_addr >> bits_range)),
snoop_tag_match.eq(snoop_addr.tag == self.s2_address.tag),
snoop_line_match.eq(snoop_addr.line == self.s2_address.line),
snoop_cancel_refill.eq(snoop_use_cache & snoop_valid
& snoop_line_match & snoop_tag_match),
]
else:
m.d.comb += snoop_cancel_refill.eq(0)
with m.FSM():
with m.State('READ'):
with m.If(self.s2_re & self.s2_miss & self.s2_valid):
m.d.sync += [
self.bus_addr.eq(self.s2_address),
self.bus_valid.eq(1),
fill_cnt.eq(self.s2_address.offset - 1)
]
m.next = 'REFILL'
with m.State('REFILL'):
m.d.comb += self.bus_last.eq(fill_cnt == self.bus_addr.offset)
with m.If(self.bus_ack):
m.d.sync += self.bus_addr.offset.eq(self.bus_addr.offset +
1)
with m.If(self.bus_ack & self.bus_last | self.bus_err):
m.d.sync += self.bus_valid.eq(0)
with m.If(~self.bus_valid | self.s1_flush
| snoop_cancel_refill):
m.next = 'READ'
m.d.sync += self.bus_valid.eq(0)
# mark the way to use (replace)
m.d.comb += ways[lru.bit_select(self.s2_address.line,
1)].sel_lru.eq(self.bus_valid)
# generate for N ways
for way in ways:
# create the memory structures for valid, tag and data.
valid = Signal(self.nlines) # Valid bits
tag_m = Memory(width=len(way.tag), depth=self.nlines) # tag memory
tag_rp = tag_m.read_port()
snoop_rp = tag_m.read_port()
tag_wp = tag_m.write_port()
m.submodules += tag_rp, tag_wp, snoop_rp
data_m = Memory(width=len(way.data),
depth=self.nlines) # data memory
data_rp = data_m.read_port()
data_wp = data_m.write_port(
granularity=32
) # implica que solo puedo escribir palabras de 32 bits.
m.submodules += data_rp, data_wp
# handle valid
with m.If(self.s1_flush & self.s1_valid): # flush
m.d.sync += valid.eq(0)
with m.Elif(way.sel_lru & self.bus_last
& self.bus_ack): # refill ok
m.d.sync += valid.bit_select(self.bus_addr.line, 1).eq(1)
with m.Elif(way.sel_lru & self.bus_err): # refill error
m.d.sync += valid.bit_select(self.bus_addr.line, 1).eq(0)
with m.Elif(self.s2_evict & self.s2_valid
& (way.tag == self.s2_address.tag)): # evict
m.d.sync += valid.bit_select(self.s2_address.line, 1).eq(0)
# assignments
m.d.comb += [
tag_rp.addr.eq(
Mux(self.s1_stall, self.s2_address.line,
self.s1_address.line)),
tag_wp.addr.eq(self.bus_addr.line),
tag_wp.data.eq(self.bus_addr.tag),
tag_wp.en.eq(way.sel_lru & self.bus_ack & self.bus_last),
data_rp.addr.eq(
Mux(self.s1_stall, self.s2_address.line,
self.s1_address.line)),
way.data.eq(data_rp.data),
way.tag.eq(tag_rp.data),
way.valid.eq(valid.bit_select(self.s2_address.line, 1))
]
# update cache: CPU or Refill
# El puerto de escritura se multiplexa debido a que la memoria solo puede tener un
# puerto de escritura.
if self.enable_write:
update_addr = Signal(len(data_wp.addr))
update_data = Signal(len(data_wp.data))
update_we = Signal(len(data_wp.en))
aux_wdata = Signal(32)
with m.If(self.bus_valid):
m.d.comb += [
update_addr.eq(self.bus_addr.line),
update_data.eq(Repl(self.bus_data, self.nwords)),
update_we.bit_select(self.bus_addr.offset,
1).eq(way.sel_lru & self.bus_ack),
]
with m.Else():
m.d.comb += [
update_addr.eq(self.s2_address.line),
update_data.eq(Repl(aux_wdata, self.nwords)),
update_we.bit_select(self.s2_address.offset,
1).eq(way.sel_we & ~self.s2_miss)
]
m.d.comb += [
# Aux data: no tengo granularidad de byte en el puerto de escritura. Así que para el
# caso en el cual el CPU tiene que escribir, hay que construir el dato (wrord) a reemplazar
aux_wdata.eq(
Cat(
Mux(self.s2_sel[0],
self.s2_wdata.word_select(0, 8),
self.s2_rdata.word_select(0, 8)),
Mux(self.s2_sel[1],
self.s2_wdata.word_select(1, 8),
self.s2_rdata.word_select(1, 8)),
Mux(self.s2_sel[2],
self.s2_wdata.word_select(2, 8),
self.s2_rdata.word_select(2, 8)),
Mux(self.s2_sel[3],
self.s2_wdata.word_select(3, 8),
self.s2_rdata.word_select(3, 8)))),
#
data_wp.addr.eq(update_addr),
data_wp.data.eq(update_data),
data_wp.en.eq(update_we),
]
else:
m.d.comb += [
data_wp.addr.eq(self.bus_addr.line),
data_wp.data.eq(Repl(self.bus_data, self.nwords)),
data_wp.en.bit_select(self.bus_addr.offset,
1).eq(way.sel_lru & self.bus_ack),
]
# --------------------------------------------------------------
# intenal snoop
# for FENCE.i instruction
_match_snoop = Signal()
m.d.comb += [
snoop_rp.addr.eq(snoop_addr.line), # read tag memory
_match_snoop.eq(snoop_rp.data == snoop_addr.tag),
way.snoop_hit.eq(snoop_use_cache & snoop_valid
& _match_snoop
& valid.bit_select(snoop_addr.line, 1)),
]
# check is the snoop match a write from this core
with m.If(way.snoop_hit):
m.d.sync += valid.bit_select(snoop_addr.line, 1).eq(0)
# --------------------------------------------------------------
return m
class Cache(Elaboratable):
def __init__(self, nlines, nwords, nways, start_addr=0, end_addr=2**32, enable_write=True):
if nlines == 0 or (nlines & (nlines - 1)):
raise ValueError(f'nlines must be a power of 2: {nlines}')
if nwords not in (4, 8, 16):
raise ValueError(f'nwords must be 4, 8 or 16: {nwords}')
if nways not in (1, 2):
raise ValueError(f'nways must be 1 or 2: {nways}')
self.enable_write = enable_write
self.nlines = nlines
self.nwords = nwords
self.nways = nways
offset_bits = log2_int(nwords)
line_bits = log2_int(nlines)
addr_bits = log2_int(end_addr - start_addr, need_pow2=False)
tag_bits = addr_bits - line_bits - offset_bits - 2 # -2 because word line.
extra_bits = 32 - tag_bits - line_bits - offset_bits - 2
pc_layout = [
('byte', 2),
('offset', offset_bits),
('line', line_bits),
('tag', tag_bits)
]
if (extra_bits != 0):
pc_layout.append(('unused', extra_bits))
self.s1_address = Record(pc_layout)
self.s1_flush = Signal()
self.s1_valid = Signal()
self.s1_stall = Signal()
self.s2_address = Record(pc_layout)
self.s2_evict = Signal()
self.s2_valid = Signal()
self.s2_miss = Signal()
self.s2_rdata = Signal(32)
self.s2_re = Signal()
if enable_write:
self.s2_wdata = Signal(32)
self.s2_sel = Signal(4)
self.s2_we = Signal()
self.bus_addr = Record(pc_layout)
self.bus_valid = Signal()
self.bus_last = Signal()
self.bus_data = Signal(32)
self.bus_ack = Signal()
self.bus_err = Signal()
def elaborate(self, platform):
m = Module()
way_layout = [
('data', 32 * self.nwords),
('tag', self.s1_address.tag.shape()),
('valid', 1),
('sel_lru', 1)
]
if self.enable_write:
way_layout.append(('sel_we', 1))
ways = Array(Record(way_layout) for _way in range(self.nways))
fill_cnt = Signal.like(self.s1_address.offset)
# set the LRU
if self.nways == 1:
lru = Const(0) # self.nlines
else:
lru = Signal(self.nlines)
with m.If(self.bus_valid & self.bus_ack & self.bus_last): # err ^ ack == 1
_lru = lru.bit_select(self.s2_address.line, 1)
m.d.sync += lru.bit_select(self.s2_address.line, 1).eq(~_lru)
# hit/miss
way_hit = m.submodules.way_hit = Encoder(self.nways)
for idx, way in enumerate(ways):
m.d.comb += way_hit.i[idx].eq((way.tag == self.s2_address.tag) & way.valid)
m.d.comb += self.s2_miss.eq(way_hit.n)
if self.enable_write:
m.d.comb += ways[way_hit.o].sel_we.eq(self.s2_we & self.s2_valid)
# read data
m.d.comb += self.s2_rdata.eq(ways[way_hit.o].data.word_select(self.s2_address.offset, 32))
with m.FSM():
with m.State('READ'):
with m.If(self.s2_re & self.s2_miss & self.s2_valid):
m.d.sync += [
self.bus_addr.eq(self.s2_address), # WARNING extra_bits
self.bus_valid.eq(1),
fill_cnt.eq(self.s2_address.offset - 1)
]
m.next = 'REFILL'
with m.State('REFILL'):
m.d.comb += self.bus_last.eq(fill_cnt == self.bus_addr.offset)
with m.If(self.bus_ack):
m.d.sync += self.bus_addr.offset.eq(self.bus_addr.offset + 1)
with m.If(self.bus_ack & self.bus_last | self.bus_err):
m.d.sync += self.bus_valid.eq(0)
with m.If(~self.bus_valid | self.s1_flush):
# in case of flush, abort ongoing refill.
m.next = 'READ'
m.d.sync += self.bus_valid.eq(0)
# mark the way to use (replace)
m.d.comb += ways[lru.bit_select(self.s2_address.line, 1)].sel_lru.eq(self.bus_valid)
# generate for N ways
for way in ways:
# create the memory structures for valid, tag and data.
valid = Signal(self.nlines)
tag_m = Memory(width=len(way.tag), depth=self.nlines)
tag_rp = tag_m.read_port()
tag_wp = tag_m.write_port()
m.submodules += tag_rp, tag_wp
data_m = Memory(width=len(way.data), depth=self.nlines)
data_rp = data_m.read_port()
data_wp = data_m.write_port(granularity=32)
m.submodules += data_rp, data_wp
# handle valid
with m.If(self.s1_flush & self.s1_valid): # flush
m.d.sync += valid.eq(0)
with m.Elif(way.sel_lru & self.bus_last & self.bus_ack): # refill ok
m.d.sync += valid.bit_select(self.bus_addr.line, 1).eq(1)
with m.Elif(way.sel_lru & self.bus_err): # refill error
m.d.sync += valid.bit_select(self.bus_addr.line, 1).eq(0)
with m.Elif(self.s2_evict & self.s2_valid & (way.tag == self.s2_address.tag)): # evict
m.d.sync += valid.bit_select(self.s2_address.line, 1).eq(0)
# assignments
m.d.comb += [
tag_rp.addr.eq(Mux(self.s1_stall, self.s2_address.line, self.s1_address.line)),
tag_wp.addr.eq(self.bus_addr.line),
tag_wp.data.eq(self.bus_addr.tag),
tag_wp.en.eq(way.sel_lru & self.bus_ack & self.bus_last),
data_rp.addr.eq(Mux(self.s1_stall, self.s2_address.line, self.s1_address.line)),
way.data.eq(data_rp.data),
way.tag.eq(tag_rp.data),
way.valid.eq(valid.bit_select(self.s2_address.line, 1))
]
# update cache: CPU or Refill
if self.enable_write:
update_addr = Signal(len(data_wp.addr))
update_data = Signal(len(data_wp.data))
update_we = Signal(len(data_wp.en))
aux_wdata = Signal(32)
with m.If(self.bus_valid):
m.d.comb += [
update_addr.eq(self.bus_addr.line),
update_data.eq(Repl(self.bus_data, self.nwords)),
update_we.bit_select(self.bus_addr.offset, 1).eq(way.sel_lru & self.bus_ack),
]
with m.Else():
m.d.comb += [
update_addr.eq(self.s2_address.line),
update_data.eq(Repl(aux_wdata, self.nwords)),
update_we.bit_select(self.s2_address.offset, 1).eq(way.sel_we & ~self.s2_miss)
]
m.d.comb += [
aux_wdata.eq(Cat(
Mux(self.s2_sel[0], self.s2_wdata.word_select(0, 8), self.s2_rdata.word_select(0, 8)),
Mux(self.s2_sel[1], self.s2_wdata.word_select(1, 8), self.s2_rdata.word_select(1, 8)),
Mux(self.s2_sel[2], self.s2_wdata.word_select(2, 8), self.s2_rdata.word_select(2, 8)),
Mux(self.s2_sel[3], self.s2_wdata.word_select(3, 8), self.s2_rdata.word_select(3, 8))
)),
#
data_wp.addr.eq(update_addr),
data_wp.data.eq(update_data),
data_wp.en.eq(update_we),
]
else:
m.d.comb += [
data_wp.addr.eq(self.bus_addr.line),
data_wp.data.eq(Repl(self.bus_data, self.nwords)),
data_wp.en.bit_select(self.bus_addr.offset, 1).eq(way.sel_lru & self.bus_ack),
]
return m