def verify_btype(self, m):
sig = self.build_signal(m, "B", [(0, 6, "1001011"), (7, 7, "1"),
(8, 11, "0000"), (12, 14, "001"),
(15, 19, "00010"), (20, 24, "00011"),
(25, 30, "011111"), (31, 31, "1")])
b = BType("btype")
b.elaborate(m.d.comb, sig)
m.d.comb += Assert(b.opcode == Const(0b1001011, 7))
m.d.comb += Assert(b.funct3 == Const(1, 3))
m.d.comb += Assert(b.rs1 == Const(2, 5))
m.d.comb += Assert(b.rs2 == Const(3, 5))
m.d.comb += Assert(
b.imm == Cat(Const(0b1101111100000, 13), Repl(1, 19)))
m.d.comb += Assert(b.match(opcode=0b1001011))
m.d.comb += Assert(b.match(rs1=2))
m.d.comb += Assert(b.match(rs2=3))
m.d.comb += Assert(b.match(funct3=1))
m.d.comb += Assert(b.match(imm=0b11111111111111111111101111100000))
m.d.comb += Assert(
b.match(opcode=0b1001011,
funct3=1,
rs1=2,
rs2=3,
imm=0b11111111111111111111101111100000))
m.d.comb += Assert(~b.match(opcode=0b1001011,
funct3=3,
rs1=1,
rs2=5,
imm=0b11111111111111111111101111100000))
b_builder_check = Signal(32)
b_builder_opcode = Signal(7)
b_builder_f3 = Signal(3)
b_builder_rs1 = Signal(5)
b_builder_rs2 = Signal(5)
b_builder_imm = Signal(13)
m.d.comb += Assume(b_builder_imm[0] == 0)
built_btype = BType.build_i32(opcode=b_builder_opcode,
funct3=b_builder_f3,
rs1=b_builder_rs1,
rs2=b_builder_rs2,
imm=b_builder_imm)
m.d.comb += b_builder_check.eq(built_btype)
b = BType("btype.build")
b.elaborate(m.d.comb, built_btype)
m.d.comb += Assert(b_builder_opcode == b.opcode)
m.d.comb += Assert(b_builder_imm == Cat(Const(0, 1), b.imm[1:13]))
m.d.comb += Assert(b.imm[13:32] == Repl(b_builder_imm[12], 32 - 13))
return [
b_builder_check, b_builder_opcode, b_builder_f3, b_builder_rs1,
b_builder_rs2, b_builder_imm
]
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
def elaborate(self, platform: Platform):
led1 = platform.request("led", 0)
led2 = platform.request("led", 1)
led3 = platform.request("led", 2)
led4 = platform.request("led", 3)
ft600_resource = platform.request("ft600")
m = Module()
# Connect pseudo power pins for the FT600 and DDR3 banks
pseudo_power = platform.request("pseudo_power")
m.d.comb += pseudo_power.ddr.o.eq(Repl(1, len(pseudo_power.ddr)))
m.d.comb += pseudo_power.ft.o.eq(Repl(1, len(pseudo_power.ft)))
m.submodules.pll = ECP5PLL(clock_signal_name="pll_clk25",
clock_config=[
ECP5PLLConfig("sync", 25),
])
m.domains += ClockDomain("ft600")
m.d.comb += ClockSignal("ft600").eq(ft600_resource.clk)
m.submodules.ft600 = ft600 = DomainRenamer("ft600")(FT600(
ft600_resource, ))
m.submodules.fifo = fifo = AsyncFIFOBuffered(width=16,
depth=2048,
r_domain="ft600",
w_domain="sync")
# FT to Write FIFO
m.d.comb += ft600.input_payload.eq(fifo.r_data)
m.d.comb += fifo.r_en.eq(ft600.input_ready)
m.d.comb += ft600.input_valid.eq(fifo.r_rdy)
# Write data into FIFO
m.d.comb += fifo.w_data.eq(0xABCD)
m.d.comb += fifo.w_en.eq(1)
led_counter = Signal(10)
with m.If(fifo.w_rdy):
m.d.sync += led_counter.eq(led_counter + 1)
# Connect LEDs
m.d.comb += led1.o.eq(ft600_resource.write)
m.d.comb += led2.o.eq(ft600_resource.txe)
m.d.comb += led3.o.eq(fifo.w_level > 2000)
m.d.comb += led4.o.eq(led_counter[-1])
return m
def elaborate(self, platform: Platform) -> Module:
m = Module()
sign_fill = Signal()
operand = Signal(32)
r_direction = Signal()
r_result = Signal(32)
shdata = Signal(64) # temp data
# pre-invert the value, if needed
# Direction: 1 = right. 0 = left.
m.d.comb += [
operand.eq(Mux(self.direction, self.dat, self.dat[::-1])),
sign_fill.eq(Mux(self.direction & self.sign_ext, self.dat[-1], 0)),
shdata.eq(Cat(operand, Repl(sign_fill, 32)))
]
with m.If(~self.stall):
m.d.sync += [
r_direction.eq(self.direction),
r_result.eq(shdata >> self.shamt)
]
m.d.comb += self.result.eq(Mux(r_direction, r_result, r_result[::-1]))
return m
def process_load(self, input_value):
lh_value = Signal(32)
comb = self.core.current_module.d.comb
bit_to_replicate=Mux(self.core.itype.funct3[2], Const(0,1), input_value[15])
comb += lh_value.eq(Cat(input_value[0:16], Repl(bit_to_replicate, 16)))
return lh_value
def elaborate(self, _: Platform) -> Module:
"""Implements the logic for the conditional right-shifter."""
m = Module()
N = self._N
with m.If(self.en):
# The high N bits get either 0 or the most significant
# bit in data_in, depending on arithmetic.
msb = self.data_in[-1]
w = len(self.data_out)
# m.d.comb += self.data_out.bit_select(w-N, N).eq(
# Mux(self.arithmetic, Repl(msb, N), 0))
with m.If(self.arithmetic):
m.d.comb += self.data_out[w - N:].eq(Repl(msb, N))
with m.Else():
m.d.comb += self.data_out[w - N:].eq(0)
# The rest are moved over.
# m.d.comb += self.data_out.bit_select(0, w-N).eq(
# self.data_in.bit_select(N, w-N))
m.d.comb += self.data_out[:w - N].eq(self.data_in[N:])
with m.Else():
m.d.comb += self.data_out.eq(self.data_in)
return m
def elaborate(self, comb: List[Statement], input: Signal):
comb += self.opcode.eq(input[0:7])
comb += self.funct3.eq(input[12:15])
comb += self.rs1.eq(input[15:20])
comb += self.rs2.eq(input[20:25])
comb += self.imm.eq(
Cat(Const(0, 1), input[8:12], input[25:31], input[7],
Repl(input[31], 20)))
def elaborate(self, platform: Platform) -> Module:
m = Module()
a = Signal(signed(33))
b = Signal(signed(33))
result_ll = Signal(32)
result_lh = Signal(33)
result_hl = Signal(33)
result_hh = Signal(33)
result_3 = Signal(64)
result_4 = Signal(64)
active = Signal(5)
is_signed = Signal()
a_is_signed = Signal()
b_is_signed = Signal()
low = Signal()
m.d.sync += [
is_signed.eq(a_is_signed ^ b_is_signed),
active.eq(Cat(self.valid & (active == 0), active)),
low.eq(self.op == Funct3.MUL)
]
# ----------------------------------------------------------------------
# fist state
m.d.comb += [
a_is_signed.eq((
(self.op == Funct3.MULH) | (self.op == Funct3.MULHSU))
& self.dat1[-1]),
b_is_signed.eq((self.op == Funct3.MULH) & self.dat2[-1])
]
m.d.sync += [
a.eq(Mux(a_is_signed, -Cat(self.dat1, 1), self.dat1)),
b.eq(Mux(b_is_signed, -Cat(self.dat2, 1), self.dat2)),
]
# ----------------------------------------------------------------------
# second state
m.d.sync += [
result_ll.eq(a[0:16] * b[0:16]),
result_lh.eq(a[0:16] * b[16:33]),
result_hl.eq(a[16:33] * b[0:16]),
result_hh.eq(a[16:33] * b[16:33])
]
# ----------------------------------------------------------------------
# third state
m.d.sync += [
result_3.eq(
Cat(result_ll, result_hh) +
Cat(Repl(0, 16), (result_lh + result_hl)))
]
# ----------------------------------------------------------------------
# fourth state
m.d.sync += [
result_4.eq(Mux(is_signed, -result_3, result_3)),
self.result.eq(Mux(low, result_4[:32], result_4[32:64]))
]
m.d.comb += self.ready.eq(active[-1])
return m
def elaborate(self, platform: Platform):
led1 = platform.request("led", 0)
led4 = platform.request("led", 3)
timer1 = Signal(25)
fifo_buf = Signal(16)
m = Module()
# Connect pseudo power pins for the FT600 and DDR3 banks
pseudo_power = platform.request("pseudo_power")
m.d.comb += pseudo_power.ddr.o.eq(Repl(1, len(pseudo_power.ddr)))
m.d.comb += pseudo_power.ft.o.eq(Repl(1, len(pseudo_power.ft)))
m.submodules.pll = ECP5PLL(clock_signal_name="pll_clk25",
clock_config=[
ECP5PLLConfig("sync", 25),
ECP5PLLConfig("fast", 100, error=0),
ECP5PLLConfig("fast2", 100, error=0),
ECP5PLLConfig("fast3", 150, error=0),
])
m.submodules.fifo = fifo = AsyncFIFOBuffered(width=16,
depth=1024,
r_domain="fast",
w_domain="sync")
# Write the FIFO using the timer data
m.d.sync += timer1.eq(timer1 + 1)
with m.If(fifo.w_rdy):
m.d.comb += fifo.w_data.eq(timer1[9:25])
m.d.comb += fifo.w_en.eq(1)
# Read the FIFO in the `fast` domain, the LEDs should blink at the same time
with m.If(fifo.r_rdy):
m.d.fast += fifo_buf.eq(fifo.r_data)
m.d.comb += fifo.r_en.eq(1)
# Combinatorial logic
m.d.comb += led1.o.eq(timer1[-1])
m.d.comb += led4.o.eq(fifo_buf[-1])
return m
def verify_utype(self, m):
sig = self.build_signal(m, "U", [(0, 6, "1001011"), (7, 11, "10000"),
(12, 31, "00100010000110111111")])
u = UType("utype")
u.elaborate(m.d.comb, sig)
m.d.comb += Assert(u.opcode == Const(0b1001011, 7))
m.d.comb += Assert(u.rd == Const(0b10000, 5))
m.d.comb += Assert(
u.imm == Const(0b00100010000110111111000000000000, 32))
m.d.comb += Assert(u.match(opcode=0b1001011))
m.d.comb += Assert(u.match(rd=0b10000))
m.d.comb += Assert(u.match(imm=0b00100010000110111111000000000000))
m.d.comb += Assert(u.match(opcode=0b1011011) == 0)
m.d.comb += Assert(u.match(rd=0b11000) == 0)
m.d.comb += Assert(
u.match(imm=0b10100010000110111111000000000000) == 0)
m.d.comb += Assert(
u.match(opcode=0b1001011,
rd=0b10000,
imm=0b00100010000110111111000000000000))
u = UType("utype.sign")
u.elaborate(m.d.comb, Const(0x8000_0000, 32))
m.d.comb += Assert(u.imm[31] == 1)
u = UType("utype.trailzero")
u.elaborate(m.d.comb, Const(0xFFFF_FFFF, 32))
m.d.comb += Assert(u.imm.bit_select(0, 12) == 0)
u_builder_check = Signal(32)
u_builder_opcode = Signal(7)
u_builder_rd = Signal(5)
u_builder_imm = Signal(20)
m.d.comb += Assume(u_builder_imm[0:12] == 0)
built_utype = UType.build_i32(opcode=u_builder_opcode,
rd=u_builder_rd,
imm=u_builder_imm)
m.d.comb += u_builder_check.eq(built_utype)
u = UType("utype.build")
u.elaborate(m.d.comb, built_utype)
m.d.comb += Assert(u_builder_opcode == u.opcode)
m.d.comb += Assert(u_builder_rd == u.rd)
m.d.comb += Assert(Cat(Repl(0, 12), u_builder_imm[12:32]) == u.imm)
return [u_builder_check, u_builder_opcode, u_builder_rd, u_builder_imm]
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
def requant(mod: Module, sig_i: Signal, QI: dict, QO: dict) -> Signal:
"""
Change word length of input signal `sig_i` to `WO` bits, using the
quantization and saturation methods specified by ``QO['quant']`` and
``QO['ovfl']``.
Parameters
----------
mod: Module (nmigen)
instance of migen module
sig_i: Signal (nmigen)
Signal to be requantized
QI: dict
Quantization dict for input word, only the keys 'WI' and 'WF' for integer
and fractional wordlength are evaluated. QI['WI'] = 2 and QI['WF'] = 13
e.g. define Q-Format '2.13' with 2 integer, 13 fractional bits and 1 implied
sign bit = 16 bits total.
QO: dict
Quantization dict for output word format; the keys 'WI' and 'WF' for
integer and fractional wordlength are evaluated as well as the keys 'quant'
and 'ovfl' describing requantization and overflow behaviour.
Returns
-------
sig_o: Signal (nmigen)
Requantized signal
Documentation
-------------
**Input and output word are aligned at their binary points.**
The following shows an example of rescaling an input word from Q2.4 to Q0.3
using wrap-around and truncation. It's easy to see that for simple wrap-around
logic, the sign of the result may change.
::
S | WI1 | WI0 * WF0 | WF1 | WF2 | WF3 : WI = 2, WF = 4, W = 7
0 | 1 | 0 * 1 | 0 | 1 | 1 = 43 (dec) or 43/16 = 2 + 11/16 (float)
*
| S * WF0 | WF1 | WF2 : WI = 0, WF = 3, W = 4
0 * 1 | 0 | 1 = 7 (dec) or 7/8 (float)
The float or "real (world) value" is calculated by multiplying the integer
value by 2 ** (-WF).
For requantizing two numbers to the same WI and WF, imagine both binary numbers
to be right-aligned. Changes in the number of integer bits `dWI` and fractional
bits `dWF` are handled separately.
Fractional Bits
---------------
- For reducing the number of fractional bits by `dWF`, simply right-shift the
integer number by `dWF`. For rounding, add '1' to the bit below the truncation
point before right-shifting.
- Extend the number of fractional bits by left-shifting the integer by `dWF`,
LSB's are filled with zeros.
Integer Bits
------------
- For reducing the number of integer bits by `dWI`, simply right-shift the
integer by `dWI`.
- The number of fractional bits is SIGN-EXTENDED by filling up the left-most
bits with the sign bit.
"""
WI_I = QI['WI'] # number of integer bits (input signal)
WI_F = QI['WF'] # number of integer bits (output signal)
WI = WI_I + WI_F + 1 # total word length (input signal)
WO_I = QO['WI'] # number of integer bits (input signal)
WO_F = QO['WF'] # number of integer bits (output signal)
WO = WO_I + WO_F + 1 # total word length (input signal)
dWF = WI_F - WO_F # difference of fractional lengths
dWI = WI_I - WO_I # difference of integer lengths
# max. resp. min, output values
MIN_o = -1 << (WO - 1)
MAX_o = -MIN_o - 1
# intermediate signal with requantized fractional part
sig_i_q = Signal(signed(max(WI, WO)))
sig_o = Signal(signed(WO))
# logger.debug(f"rescale: dWI={dWI}, dWF={dWF}, Qu:{QO['quant']}, Ov:{QO['ovfl']}")
# -----------------------------------------------------------------------
# Requantize fractional part
# -----------------------------------------------------------------------
if dWF <= 0: # Extend fractional word length of output word by
# multiplying with 2^dWF
mod.d.comb += sig_i_q.eq(sig_i << -dWF) # shift input left by -dWF
else: # dWF > 0, reduce fract. word length by dividing by 2^dWF
# (shift right by dWF)
if QO['quant'] == 'round':
# add half an LSB (1 << (dWF - 1)) before right shift
mod.d.comb += sig_i_q.eq((sig_i + (1 << (dWF - 1))) >> dWF)
elif QO['quant'] == 'floor': # just shift right
mod.d.comb += sig_i_q.eq(sig_i >> dWF)
elif QO['quant'] == 'fix':
# add sign bit (sig_i[-1]) as LSB (1 << dWF) before right shift
mod.d.comb += sig_i_q.eq((sig_i + (sig_i[-1] << dWF)) >> dWF)
else:
raise Exception(u'Unknown quantization method "%s"!' %
(QI['quant']))
# -----------------------------------------------------------------------
# Requantize integer part
# -----------------------------------------------------------------------
if dWI < 0: # WI_I < WO_I, sign extend integer part (prepend copies of sign bit)
mod.d.comb += sig_o.eq(Cat(sig_i_q, Repl(sig_i_q[-1], -dWI)))
elif dWI == 0: # WI = WO, don't change integer part
mod.d.comb += sig_o.eq(sig_i_q)
elif QO['ovfl'] == 'sat':
with mod.If(sig_i_q[-1] == 1):
with mod.If(sig_i_q < MIN_o):
mod.d.comb += sig_o.eq(MIN_o)
with mod.Else():
mod.d.comb += sig_o.eq(sig_i_q)
with mod.Elif(sig_i_q > MAX_o): # sig_i_q[-1] == 0
mod.d.comb += sig_o.eq(MAX_o)
with mod.Else():
mod.d.comb += sig_o.eq(sig_i_q)
elif QO['ovfl'] == 'wrap': # wrap around (shift left)
mod.d.comb += sig_o.eq(sig_i_q)
else:
raise Exception(u'Unknown overflow method "%s"!' % (QI['ovfl']))
return sig_o
def elaborate(self, comb: List[Statement], input: Signal):
comb += self.opcode.eq(input[0:7])
comb += self.rd.eq(input[7:12])
comb += self.funct3.eq(input[12:15])
comb += self.rs1.eq(input[15:20])
comb += self.imm.eq(Cat(input[20:32], Repl(input[31], 20)))
def elaborate(self, comb: List[Statement], input: Signal):
comb += self.opcode.eq(input[0:7])
comb += self.rd.eq(input[7:12])
comb += self.imm.eq(
Cat(Const(0, 1), input[21:31], input[20], input[12:20],
Repl(input[31], 12)))
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
def decode_imm_chips(self, m: Module):
mux = IC_mux32(N=6, faster=True)
gal = IC_GAL_imm_format_decoder()
m.submodules += mux
m.submodules += gal
m.d.comb += gal.opcode.eq(self._opcode)
m.d.comb += mux.n_sel[0].eq(gal.i_n_oe)
m.d.comb += mux.n_sel[1].eq(gal.s_n_oe)
m.d.comb += mux.n_sel[2].eq(gal.u_n_oe)
m.d.comb += mux.n_sel[3].eq(gal.b_n_oe)
m.d.comb += mux.n_sel[4].eq(gal.j_n_oe)
m.d.comb += mux.n_sel[5].eq(gal.sys_n_oe)
instr = self.state._instr
# Format I
m.d.comb += [
mux.a[0][0:12].eq(instr[20:]),
mux.a[0][12:].eq(Repl(instr[31], 32)), # sext
]
# Format S
m.d.comb += [
mux.a[1][0:5].eq(instr[7:]),
mux.a[1][5:11].eq(instr[25:]),
mux.a[1][11:].eq(Repl(instr[31], 32)), # sext
]
# Format U
m.d.comb += [
mux.a[2][0:12].eq(0),
mux.a[2][12:].eq(instr[12:]),
]
# Format B
m.d.comb += [
mux.a[3][0].eq(0),
mux.a[3][1:5].eq(instr[8:]),
mux.a[3][5:11].eq(instr[25:]),
mux.a[3][11].eq(instr[7]),
mux.a[3][12:].eq(Repl(instr[31], 32)), # sext
]
# Format J
m.d.comb += [
mux.a[4][0].eq(0),
mux.a[4][1:11].eq(instr[21:]),
mux.a[4][11].eq(instr[20]),
mux.a[4][12:20].eq(instr[12:]),
mux.a[4][20:].eq(Repl(instr[31], 32)), # sext
]
# Format SYS
m.d.comb += [
mux.a[5][0:5].eq(instr[15:]),
mux.a[5][5:].eq(0),
]
m.d.comb += self._imm.eq(mux.y)
