# Copyright (C) Jonathan P Dawson 2013
# 2013-12-12
-from nmigen import Module, Signal, Cat
-from nmigen.cli import main
+from nmigen import Module, Signal, Cat, Const
+from nmigen.cli import main, verilog
class FPNum:
+ """ Floating-point Number Class, variable-width TODO (currently 32-bit)
+
+ Contains signals for an incoming copy of the value, decoded into
+ sign / exponent / mantissa.
+ Also contains encoding functions, creation and recognition of
+ zero, NaN and inf (all signed)
+
+ Four extra bits are included in the mantissa: the top bit
+ (m[-1]) is effectively a carry-overflow. The other three are
+ guard (m[2]), round (m[1]), and sticky (m[0])
+ """
def __init__(self, width, m_width=None):
self.width = width
if m_width is None:
self.e = Signal((10, True)) # Exponent: 10 bits, signed
self.s = Signal() # Sign bit
+ self.mzero = Const(0, (m_width, False))
+ self.m1s = Const(-1, (m_width, False))
+ self.P128 = Const(128, (10, True))
+ self.P127 = Const(127, (10, True))
+ self.N127 = Const(-127, (10, True))
+ self.N126 = Const(-126, (10, True))
+
def decode(self):
+ """ decodes a latched value into sign / exponent / mantissa
+
+ bias is subtracted here, from the exponent. exponent
+ is extended to 10 bits so that subtract 127 is done on
+ a 10-bit number
+ """
v = self.v
return [self.m.eq(Cat(0, 0, 0, v[0:23])), # mantissa
- self.e.eq(Cat(v[23:31]) - 127), # exponent (take off bias)
- self.s.eq(Cat(v[31])), # sign
+ self.e.eq(v[23:31] - self.P127), # exp (minus bias)
+ self.s.eq(v[31]), # sign
]
def create(self, s, e, m):
+ """ creates a value from sign / exponent / mantissa
+
+ bias is added here, to the exponent
+ """
return [
self.v[31].eq(s), # sign
- self.v[23:31].eq(e + 127), # exp
+ self.v[23:31].eq(e + self.P127), # exp (add on bias)
self.v[0:23].eq(m) # mantissa
]
def shift_down(self):
- return self.create(self.s,
- self.e + 1,
- Cat(self.m[0] | self.m[1], self.m[1:-5], 0))
+ """ shifts a mantissa down by one. exponent is increased to compensate
+
+ accuracy is lost as a result in the mantissa however there are 3
+ guard bits (the latter of which is the "sticky" bit)
+ """
+ return [self.e.eq(self.e + 1),
+ self.m.eq(Cat(self.m[0] | self.m[1], self.m[2:], 0))
+ ]
def nan(self, s):
- return self.create(s, 0xff, 1<<22)
+ return self.create(s, self.P128, 1<<22)
def inf(self, s):
- return self.create(s, 0xff, 0)
+ return self.create(s, self.P128, 0)
+
+ def zero(self, s):
+ return self.create(s, self.N127, 0)
def is_nan(self):
- return (self.e == 128) & (self.m != 0)
+ return (self.e == self.P128) & (self.m != 0)
def is_inf(self):
- return (self.e == 128) & (self.m == 0)
+ return (self.e == self.P128) & (self.m == 0)
def is_zero(self):
- return (self.e == -127) & (self.m == 0)
+ return (self.e == self.N127) & (self.m == self.mzero)
+ def is_overflowed(self):
+ return (self.e > self.P127)
-class FPADD:
+ def is_denormalised(self):
+ return (self.e == self.N126) & (self.m[23] == 0)
+
+class FPOp:
def __init__(self, width):
self.width = width
- self.in_a = Signal(width)
- self.in_a_stb = Signal()
- self.in_a_ack = Signal()
+ self.v = Signal(width)
+ self.stb = Signal()
+ self.ack = Signal()
+
+ def ports(self):
+ return [self.v, self.stb, self.ack]
+
+
+class Overflow:
+ def __init__(self):
+ self.guard = Signal() # tot[2]
+ self.round_bit = Signal() # tot[1]
+ self.sticky = Signal() # tot[0]
- self.in_b = Signal(width)
- self.in_b_stb = Signal()
- self.in_b_ack = Signal()
- self.out_z = Signal(width)
- self.out_z_stb = Signal()
- self.out_z_ack = Signal()
+class FPADD:
+ def __init__(self, width):
+ self.width = width
- def get_fragment(self, platform):
+ self.in_a = FPOp(width)
+ self.in_b = FPOp(width)
+ self.out_z = FPOp(width)
+
+ def get_op(self, m, op, v, next_state):
+ with m.If((op.ack) & (op.stb)):
+ m.next = next_state
+ m.d.sync += [
+ v.eq(op.v),
+ op.ack.eq(0)
+ ]
+ with m.Else():
+ m.d.sync += op.ack.eq(1)
+
+ def normalise_1(self, m, z, of, next_state):
+ with m.If((z.m[-1] == 0) & (z.e > z.N126)):
+ m.d.sync +=[
+ z.e.eq(z.e - 1), # DECREASE exponent
+ z.m.eq(z.m << 1), # shift mantissa UP
+ z.m[0].eq(of.guard), # steal guard bit (was tot[2])
+ of.guard.eq(of.round_bit), # steal round_bit (was tot[1])
+ of.round_bit.eq(0), # reset round bit
+ ]
+ with m.Else():
+ m.next = next_state
+
+ def get_fragment(self, platform=None):
m = Module()
# Latches
b = FPNum(self.width)
z = FPNum(self.width, 24)
- guard = Signal()
- round_bit = Signal()
- sticky = Signal()
+ tot = Signal(28) # sticky/round/guard bits, 23 result, 1 overflow
- tot = Signal(28)
+ of = Overflow()
with m.FSM() as fsm:
# gets operand a
with m.State("get_a"):
- with m.If((self.in_a_ack) & (self.in_a_stb)):
- m.next = "get_b"
- m.d.sync += [
- a.v.eq(self.in_a),
- self.in_a_ack.eq(0)
- ]
- with m.Else():
- m.d.sync += self.in_a_ack.eq(1)
+ self.get_op(m, self.in_a, a.v, "get_b")
# ******
# gets operand b
with m.State("get_b"):
- with m.If((self.in_b_ack) & (self.in_b_stb)):
- m.next = "get_a"
- m.d.sync += [
- b.v.eq(self.in_b),
- self.in_b_ack.eq(0)
- ]
- with m.Else():
- m.d.sync += self.in_b_ack.eq(1)
+ self.get_op(m, self.in_b, b.v, "unpack")
# ******
# unpacks operands into sign, mantissa and exponent
m.next = "put_z"
m.d.sync += z.inf(a.s)
# if a is inf and signs don't match return NaN
- with m.If((b.e == 128) & (a.s != b.s)):
+ with m.If((b.e == b.P128) & (a.s != b.s)):
m.d.sync += z.nan(b.s)
# if b is inf return inf
# if a is zero and b zero return signed-a/b
with m.Elif(a.is_zero() & b.is_zero()):
m.next = "put_z"
- m.d.sync += z.create(a.s & b.s, b.e[0:8], b.m[3:26])
+ m.d.sync += z.create(a.s & b.s, b.e[0:8], b.m[3:-1])
# if a is zero return b
with m.Elif(a.is_zero()):
m.next = "put_z"
- m.d.sync += z.create(b.s, b.e[0:8], b.m[3:26])
+ m.d.sync += z.create(b.s, b.e[0:8], b.m[3:-1])
# if b is zero return a
with m.Elif(b.is_zero()):
m.next = "put_z"
- m.d.sync += z.create(a.s, a.e[0:8], a.m[3:26])
+ m.d.sync += z.create(a.s, a.e[0:8], a.m[3:-1])
# Denormalised Number checks
with m.Else():
m.next = "align"
# denormalise a check
- with m.If(a.e == -127):
+ with m.If(a.e == a.N127):
m.d.sync += a.e.eq(-126) # limit a exponent
with m.Else():
- m.d.sync += a.m[26].eq(1) # set highest mantissa bit
+ m.d.sync += a.m[-1].eq(1) # set top mantissa bit
# denormalise b check
- with m.If(b.e == -127):
+ with m.If(b.e == a.N127):
m.d.sync += b.e.eq(-126) # limit b exponent
with m.Else():
- m.d.sync += b.m[26].eq(1) # set highest mantissa bit
+ m.d.sync += b.m[-1].eq(1) # set top mantissa bit
# ******
# align. NOTE: this does *not* do single-cycle multi-shifting,
with m.If(tot[27]):
m.d.sync += [
z.m.eq(tot[4:28]),
- guard.eq(tot[3]),
- round_bit.eq(tot[2]),
- sticky.eq(tot[1] | tot[0]),
+ of.guard.eq(tot[3]),
+ of.round_bit.eq(tot[2]),
+ of.sticky.eq(tot[1] | tot[0]),
z.e.eq(z.e + 1)
]
# tot[27] zero case
with m.Else():
m.d.sync += [
z.m.eq(tot[3:27]),
- guard.eq(tot[2]),
- round_bit.eq(tot[1]),
- sticky.eq(tot[0])
+ of.guard.eq(tot[2]),
+ of.round_bit.eq(tot[1]),
+ of.sticky.eq(tot[0])
]
# ******
# the extra mantissa bits coming from tot[0..2]
with m.State("normalise_1"):
- with m.If((z.m[23] == 0) & (z.e > -126)):
+ with m.If((z.m[-1] == 0) & (z.e > z.N126)):
m.d.sync +=[
z.e.eq(z.e - 1), # DECREASE exponent
z.m.eq(z.m << 1), # shift mantissa UP
- z.m[0].eq(guard), # steal guard bit (was tot[2])
- guard.eq(round_bit), # steal round_bit (was tot[1])
+ z.m[0].eq(of.guard), # steal guard bit (was tot[2])
+ of.guard.eq(of.round_bit), # steal round_bit was tot[1])
]
with m.Else():
- m.next = "normalize_2"
+ m.next = "normalise_2"
# ******
# Second stage of normalisation.
# the extra mantissa bits coming from tot[0..2]
with m.State("normalise_2"):
- with m.If(z.e < -126):
+ with m.If(z.e < z.N126):
m.d.sync +=[
z.e.eq(z.e + 1), # INCREASE exponent
z.m.eq(z.m >> 1), # shift mantissa DOWN
- guard.eq(z.m[0]),
- round_bit.eq(guard),
- sticky.eq(sticky | round_bit)
+ of.guard.eq(z.m[0]),
+ of.round_bit.eq(of.guard),
+ of.sticky.eq(of.sticky | of.round_bit)
]
with m.Else():
m.next = "round"
# rounding stage
with m.State("round"):
- m.next = "pack"
- with m.If(guard & (round_bit | sticky | z.m[0])):
+ m.next = "corrections"
+ with m.If(of.guard & (of.round_bit | of.sticky | z.m[0])):
m.d.sync += z.m.eq(z.m + 1) # mantissa rounds up
- with m.If(z.m == 0xffffff): # all 1s
+ with m.If(z.m == z.m1s): # all 1s
m.d.sync += z.e.eq(z.e + 1) # exponent rounds up
+ # ******
+ # correction stage
+
+ with m.State("corrections"):
+ m.next = "pack"
+ # denormalised, correct exponent to zero
+ with m.If(z.is_denormalised()):
+ m.d.sync += z.m.eq(-127)
+ # FIX SIGN BUG: -a + a = +0.
+ with m.If((z.e == z.N126) & (z.m[0:] == 0)):
+ m.d.sync += z.s.eq(0)
+
# ******
# pack stage
- """ TODO: see if z.create can be used *later*. convert
- verilog first (and commit), *second* phase, convert nmigen
- code to use FPNum.create() (as a separate commit)
-
- pack:
- begin
- z[22 : 0] <= z_m[22:0];
- z[30 : 23] <= z_e[7:0] + 127;
- z[31] <= z_s;
- if ($signed(z_e) == -126 && z_m[23] == 0) begin
- z[30 : 23] <= 0;
- end
- if ($signed(z_e) == -126 && z_m[23:0] == 24'h0) begin
- z[31] <= 1'b0; // FIX SIGN BUG: -a + a = +0.
- end
- //if overflow occurs, return inf
- if ($signed(z_e) > 127) begin
- z[22 : 0] <= 0;
- z[30 : 23] <= 255;
- z[31] <= z_s;
- end
- state <= put_z;
- end
- """
+ with m.State("pack"):
+ m.next = "put_z"
+ # if overflow occurs, return inf
+ with m.If(z.is_overflowed()):
+ m.d.sync += z.inf(0)
+ with m.Else():
+ m.d.sync += z.create(z.s, z.e, z.m)
# ******
# put_z stage
- """
- put_z:
- begin
- s_out_z_stb <= 1;
- s_out_z <= z;
- if (s_out_z_stb && out_z_ack) begin
- s_out_z_stb <= 0;
- state <= get_a;
- end
- end
- """
+ with m.State("put_z"):
+ m.d.sync += [
+ self.out_z.stb.eq(1),
+ self.out_z.v.eq(z.v)
+ ]
+ with m.If(self.out_z.stb & self.out_z.ack):
+ m.d.sync += self.out_z.stb.eq(0)
+ m.next = "get_a"
return m
-"""
- always @(posedge clk)
- begin
-
- case(state)
-
- get_a:
- begin
- s_in_a_ack <= 1;
- if (s_in_a_ack && in_a_stb) begin
- a <= in_a;
- s_in_a_ack <= 0;
- state <= get_b;
- end
- end
-
- get_b:
- begin
- s_in_b_ack <= 1;
- if (s_in_b_ack && in_b_stb) begin
- b <= in_b;
- s_in_b_ack <= 0;
- state <= unpack;
- end
- end
-
- unpack:
- begin
- a_m <= {a[22 : 0], 3'd0};
- b_m <= {b[22 : 0], 3'd0};
- a_e <= a[30 : 23] - 127;
- b_e <= b[30 : 23] - 127;
- a_s <= a[31];
- b_s <= b[31];
- state <= special_cases;
- end
-
- special_cases:
- begin
- //if a is NaN or b is NaN return NaN
- if ((a_e == 128 && a_m != 0) || (b_e == 128 && b_m != 0)) begin
- z[31] <= 1;
- z[30:23] <= 255;
- z[22] <= 1;
- z[21:0] <= 0;
- state <= put_z;
- //if a is inf return inf
- end else if (a_e == 128) begin
- z[31] <= a_s;
- z[30:23] <= 255;
- z[22:0] <= 0;
- //if a is inf and signs don't match return nan
- if ((b_e == 128) && (a_s != b_s)) begin
- z[31] <= b_s;
- z[30:23] <= 255;
- z[22] <= 1;
- z[21:0] <= 0;
- end
- state <= put_z;
- //if b is inf return inf
- end else if (b_e == 128) begin
- z[31] <= b_s;
- z[30:23] <= 255;
- z[22:0] <= 0;
- state <= put_z;
- //if a is zero return b
- end else if ((($signed(a_e) == -127) && (a_m == 0)) && (($signed(b_e) == -127) && (b_m == 0))) begin
- z[31] <= a_s & b_s;
- z[30:23] <= b_e[7:0] + 127;
- z[22:0] <= b_m[26:3];
- state <= put_z;
- //if a is zero return b
- end else if (($signed(a_e) == -127) && (a_m == 0)) begin
- z[31] <= b_s;
- z[30:23] <= b_e[7:0] + 127;
- z[22:0] <= b_m[26:3];
- state <= put_z;
- //if b is zero return a
- end else if (($signed(b_e) == -127) && (b_m == 0)) begin
- z[31] <= a_s;
- z[30:23] <= a_e[7:0] + 127;
- z[22:0] <= a_m[26:3];
- state <= put_z;
- end else begin
- //Denormalised Number
- if ($signed(a_e) == -127) begin
- a_e <= -126;
- end else begin
- a_m[26] <= 1;
- end
- //Denormalised Number
- if ($signed(b_e) == -127) begin
- b_e <= -126;
- end else begin
- b_m[26] <= 1;
- end
- state <= align;
- end
- end
-
- align:
- begin
- if ($signed(a_e) > $signed(b_e)) begin
- b_e <= b_e + 1;
- b_m <= b_m >> 1;
- b_m[0] <= b_m[0] | b_m[1];
- end else if ($signed(a_e) < $signed(b_e)) begin
- a_e <= a_e + 1;
- a_m <= a_m >> 1;
- a_m[0] <= a_m[0] | a_m[1];
- end else begin
- state <= add_0;
- end
- end
-
- add_0:
- begin
- z_e <= a_e;
- if (a_s == b_s) begin
- tot <= a_m + b_m;
- z_s <= a_s;
- end else begin
- if (a_m >= b_m) begin
- tot <= a_m - b_m;
- z_s <= a_s;
- end else begin
- tot <= b_m - a_m;
- z_s <= b_s;
- end
- end
- state <= add_1;
- end
-
- add_1:
- begin
- if (tot[27]) begin
- z_m <= tot[27:4];
- guard <= tot[3];
- round_bit <= tot[2];
- sticky <= tot[1] | tot[0];
- z_e <= z_e + 1;
- end else begin
- z_m <= tot[26:3];
- guard <= tot[2];
- round_bit <= tot[1];
- sticky <= tot[0];
- end
- state <= normalise_1;
- end
-
- normalise_1:
- begin
- if (z_m[23] == 0 && $signed(z_e) > -126) begin
- z_e <= z_e - 1;
- z_m <= z_m << 1;
- z_m[0] <= guard;
- guard <= round_bit;
- round_bit <= 0;
- end else begin
- state <= normalise_2;
- end
- end
-
- normalise_2:
- begin
- if ($signed(z_e) < -126) begin
- z_e <= z_e + 1;
- z_m <= z_m >> 1;
- guard <= z_m[0];
- round_bit <= guard;
- sticky <= sticky | round_bit;
- end else begin
- state <= round;
- end
- end
-
- round:
- begin
- if (guard && (round_bit | sticky | z_m[0])) begin
- z_m <= z_m + 1;
- if (z_m == 24'hffffff) begin
- z_e <=z_e + 1;
- end
- end
- state <= pack;
- end
-
- pack:
- begin
- z[22 : 0] <= z_m[22:0];
- z[30 : 23] <= z_e[7:0] + 127;
- z[31] <= z_s;
- if ($signed(z_e) == -126 && z_m[23] == 0) begin
- z[30 : 23] <= 0;
- end
- if ($signed(z_e) == -126 && z_m[23:0] == 24'h0) begin
- z[31] <= 1'b0; // FIX SIGN BUG: -a + a = +0.
- end
- //if overflow occurs, return inf
- if ($signed(z_e) > 127) begin
- z[22 : 0] <= 0;
- z[30 : 23] <= 255;
- z[31] <= z_s;
- end
- state <= put_z;
- end
-
- put_z:
- begin
- s_out_z_stb <= 1;
- s_out_z <= z;
- if (s_out_z_stb && out_z_ack) begin
- s_out_z_stb <= 0;
- state <= get_a;
- end
- end
-
- endcase
-
- if (rst == 1) begin
- state <= get_a;
- s_in_a_ack <= 0;
- s_in_b_ack <= 0;
- s_out_z_stb <= 0;
- end
-
- end
- assign in_a_ack = s_in_a_ack;
- assign in_b_ack = s_in_b_ack;
- assign out_z_stb = s_out_z_stb;
- assign out_z = s_out_z;
-
-endmodule
-"""
if __name__ == "__main__":
alu = FPADD(width=32)
- main(alu, ports=[
- alu.in_a, alu.in_a_stb, alu.in_a_ack,
- alu.in_b, alu.in_b_stb, alu.in_b_ack,
- alu.out_z, alu.out_z_stb, alu.out_z_ack,
- ])
+ main(alu, ports=alu.in_a.ports() + alu.in_b.ports() + alu.out_z.ports())
+
+
+ # works... but don't use, just do "python fname.py convert -t v"
+ #print (verilog.convert(alu, ports=[
+ # ports=alu.in_a.ports() + \
+ # alu.in_b.ports() + \
+ # alu.out_z.ports())