the class turns into a SIMD variant of Signal. *this is dynamic*.
the basic fundamental idea is: write code once, and if you want a SIMD
-version of it, use PartitionedSignal in place of Signal. job done.
+version of it, use SimdSignal in place of Signal. job done.
this however requires the code to *not* be designed to use nmigen.If,
nmigen.Case, or other constructs: only Mux and other logic.
from ieee754.part_mul_add.adder import PartitionedAdder
from ieee754.part_cmp.eq_gt_ge import PartitionedEqGtGe
+from ieee754.part_bits.xor import PartitionedXOR
+from ieee754.part_bits.bool import PartitionedBool
+from ieee754.part_bits.all import PartitionedAll
from ieee754.part_shift.part_shift_dynamic import PartitionedDynamicShift
from ieee754.part_shift.part_shift_scalar import PartitionedScalarShift
-from ieee754.part_mul_add.partpoints import make_partition
+from ieee754.part_mul_add.partpoints import make_partition2, PartitionPoints
+from ieee754.part_mux.part_mux import PMux
+from ieee754.part_ass.passign import PAssign
+from ieee754.part_cat.pcat import PCat
+from ieee754.part_repl.prepl import PRepl
+from ieee754.part.simd_scope import SimdScope
+from ieee754.part.layout_experiment import layout
from operator import or_, xor, and_, not_
-from nmigen import (Signal, Const)
+from nmigen import (Signal, Const, Cat)
+from nmigen.hdl.ast import UserValue, Shape
def getsig(op1):
- if isinstance(op1, PartitionedSignal):
+ if isinstance(op1, SimdSignal):
op1 = op1.sig
return op1
def applyop(op1, op2, op):
- return op(getsig(op1), getsig(op2))
-
-
-class PartitionedSignal:
- def __init__(self, mask, *args, **kwargs):
- self.sig = Signal(*args, **kwargs)
- width = self.sig.shape()[0] # get signal width
+ if isinstance(op1, SimdSignal):
+ result = SimdSignal.like(op1)
+ else:
+ result = SimdSignal.like(op2)
+ result.m.d.comb += result.sig.eq(op(getsig(op1), getsig(op2)))
+ return result
+
+
+global modnames
+modnames = {}
+# for sub-modules to be created on-demand. Mux is done slightly
+# differently (has its own global)
+for name in ['add', 'eq', 'gt', 'ge', 'ls', 'xor', 'bool', 'all']:
+ modnames[name] = 0
+
+
+# Prototype https://bugs.libre-soc.org/show_bug.cgi?id=713#c53
+# this provides a "compatibility" layer with existing SimdSignal
+# behaviour. the idea is that this interface defines which "combinations"
+# of partition selections are relevant, and as an added bonus it says
+# which partition lanes are completely irrelevant (padding, blank).
+class PartType: # TODO decide name
+ def __init__(self, psig):
+ self.psig = psig
+
+ def get_mask(self):
+ return list(self.psig.partpoints.values())
+
+ def get_switch(self):
+ return Cat(self.get_mask())
+
+ def get_cases(self):
+ return range(1 << len(self.get_mask()))
+
+ @property
+ def blanklanes(self):
+ return 0
+
+
+# this one would be an elwidth version
+# see https://bugs.libre-soc.org/show_bug.cgi?id=713#c34
+# it requires an "adapter" which is the layout() function
+# where the PartitionPoints was *created* by the layout()
+# function and this class then "understands" the relationship
+# between elwidth and the PartitionPoints that were created
+# by layout()
+class ElwidPartType: # TODO decide name
+ def __init__(self, psig):
+ self.psig = psig
+
+ def get_mask(self):
+ return list(self.psig._shape.partpoints.values()) # i think
+
+ def get_switch(self):
+ return self.psig.scope.elwid # switch on elwid: match get_cases()
+
+ def get_cases(self):
+ return self.psig._shape.bitp.keys() # all possible values of elwid
+
+ @property
+ def blanklanes(self):
+ return self.psig.shape.blankmask
+
+
+class SimdShape(Shape):
+ """a SIMD variant of Shape. supports:
+ * fixed overall width with variable (maxed-out) element lengths
+ * fixed element widths with overall size auto-determined
+ * both fixed overall width and fixed element widths
+
+ naming is preserved to be compatible with Shape().
+ """
+ def __init__(self, scope, width=None, # this is actually widths_at_elwid
+ signed=False,
+ fixed_width=None): # fixed overall width
+ widths_at_elwid = width
+ print ("SimdShape width", width, "fixed_width", fixed_width)
+ # this check is done inside layout but do it again here anyway
+ assert fixed_width != None or widths_at_elwid != None, \
+ "both width (widths_at_elwid) and fixed_width cannot be None"
+ (pp, bitp, lpoints, bmask, fixed_width, lane_shapes, part_wid) = \
+ layout(scope.elwid,
+ scope.vec_el_counts,
+ widths_at_elwid,
+ fixed_width)
+ self.partpoints = pp
+ self.bitp = bitp # binary values for partpoints at each elwidth
+ self.lpoints = lpoints # layout ranges
+ self.blankmask = bmask # blanking mask (partitions always padding)
+ self.partwid = part_wid # smallest alignment start point for elements
+
+ # pass through the calculated width to Shape() so that when/if
+ # objects using this Shape are downcast, they know exactly how to
+ # get *all* bits and need know absolutely nothing about SIMD at all
+ Shape.__init__(self, fixed_width, signed)
+
+
+class SimdSignal(UserValue):
+ # XXX ################################################### XXX
+ # XXX Keep these functions in the same order as ast.Value XXX
+ # XXX ################################################### XXX
+ def __init__(self, mask, shape=None, *args,
+ src_loc_at=0, fixed_width=None, **kwargs):
+ super().__init__(src_loc_at=src_loc_at)
+ print ("SimdSignal shape", shape)
# create partition points
- self.partpoints = make_partition(mask, width)
- self.modnames = {}
- for name in ['add', 'eq', 'gt', 'ge', 'ls']:
- self.modnames[name] = 0
+ if isinstance(mask, SimdScope): # mask parameter is a SimdScope
+ self.scope = mask
+ self.ptype = ElwidPartType(self)
+ # adapt shape to a SimdShape
+ if not isinstance(shape, SimdShape):
+ shape = SimdShape(self.scope, shape, fixed_width=fixed_width)
+ self._shape = shape
+ self.sig = Signal(shape, *args, **kwargs)
+ # get partpoints from SimdShape
+ self.partpoints = shape.partpoints
+ else:
+ self.sig = Signal(shape, *args, **kwargs)
+ width = len(self.sig) # get signal width
+ if isinstance(mask, PartitionPoints):
+ self.partpoints = mask
+ else:
+ self.partpoints = make_partition2(mask, width)
+ self.ptype = PartType(self)
def set_module(self, m):
self.m = m
def get_modname(self, category):
- self.modnames[category] += 1
- return "%s%d" % (category, self.modnames[category])
+ modnames[category] += 1
+ return "%s_%d" % (category, modnames[category])
+
+ @staticmethod
+ def like(other, *args, **kwargs):
+ """Builds a new SimdSignal with the same PartitionPoints and
+ Signal properties as the other"""
+ result = SimdSignal(PartitionPoints(other.partpoints))
+ result.sig = Signal.like(other.sig, *args, **kwargs)
+ result.m = other.m
+ return result
- def eq(self, val):
- return self.sig.eq(getsig(val))
+ def lower(self):
+ return self.sig
+
+ # nmigen-redirected constructs (Mux, Cat, Switch, Assign)
+
+ # TODO, http://bugs.libre-riscv.org/show_bug.cgi?id=716
+ #def __Part__(self, offset, width, stride=1, *, src_loc_at=0):
+ raise NotImplementedError("TODO: implement as "
+ "(self>>(offset*stride)[:width]")
+ # TODO, http://bugs.libre-riscv.org/show_bug.cgi?id=716
+ def __Slice__(self, start, stop, *, src_loc_at=0):
+ # NO. Swizzled shall NOT be deployed, it violates
+ # Project Development Practices
+ raise NotImplementedError("TODO: need PartitionedSlice")
+
+ def __Repl__(self, count, *, src_loc_at=0):
+ return PRepl(self.m, self, count, self.ptype)
+
+ def __Cat__(self, *args, src_loc_at=0):
+ print ("partsig cat", self, args)
+ # TODO: need SwizzledSimdValue-aware Cat
+ args = [self] + list(args)
+ for sig in args:
+ assert isinstance(sig, SimdSignal), \
+ "All SimdSignal.__Cat__ arguments must be " \
+ "a SimdSignal. %s is not." % repr(sig)
+ return PCat(self.m, args, self.ptype)
+
+ def __Mux__(self, val1, val2):
+ # print ("partsig mux", self, val1, val2)
+ assert len(val1) == len(val2), \
+ "SimdSignal width sources must be the same " \
+ "val1 == %d, val2 == %d" % (len(val1), len(val2))
+ return PMux(self.m, self.partpoints, self, val1, val2, self.ptype)
+
+ def __Assign__(self, val, *, src_loc_at=0):
+ print ("partsig assign", self, val)
+ # this is a truly awful hack, outlined here:
+ # https://bugs.libre-soc.org/show_bug.cgi?id=731#c13
+ # during the period between constructing Simd-aware sub-modules
+ # and the elaborate() being called on them there is a window of
+ # opportunity to indicate which of those submodules is LHS and
+ # which is RHS. manic laughter is permitted. *gibber*.
+ if hasattr(self, "_hack_submodule"):
+ self._hack_submodule.set_lhs_mode(True)
+ if hasattr(val, "_hack_submodule"):
+ val._hack_submodule.set_lhs_mode(False)
+ return PAssign(self.m, self, val, self.ptype)
+
+ # TODO, http://bugs.libre-riscv.org/show_bug.cgi?id=458
+ # def __Switch__(self, cases, *, src_loc=None, src_loc_at=0,
+ # case_src_locs={}):
+
+ # no override needed, Value.__bool__ sufficient
+ # def __bool__(self):
# unary ops that do not require partitioning
def __invert__(self):
- return ~self.sig
+ result = SimdSignal.like(self)
+ self.m.d.comb += result.sig.eq(~self.sig)
+ return result
# unary ops that require partitioning
def __neg__(self):
- result, _ = self.add_op(self, ~0, carry=0) # TODO, subop
+ z = Const(0, len(self.sig))
+ result, _ = self.sub_op(z, self)
return result
- # binary ops that don't require partitioning
-
- def __and__(self, other):
- return applyop(self, other, and_)
-
- def __rand__(self, other):
- return applyop(other, self, and_)
-
- def __or__(self, other):
- return applyop(self, other, or_)
-
- def __ror__(self, other):
- return applyop(other, self, or_)
-
- def __xor__(self, other):
- return applyop(self, other, xor)
-
- def __rxor__(self, other):
- return applyop(other, self, xor)
-
# binary ops that need partitioning
- # TODO: detect if the 2nd operand is a Const, a Signal or a
- # PartitionedSignal. if it's a Const or a Signal, a global shift
- # can occur. if it's a PartitionedSignal, that's much more interesting.
- def ls_op(self, op1, op2, carry):
- op1 = getsig(op1)
- if isinstance(op2, Const) or isinstance(op2, Signal):
- scalar = True
- shape = op1.shape()
- pa = PartitionedScalarShift(shape[0], self.partpoints)
- else:
- scalar = False
- op2 = getsig(op2)
- shape = op1.shape()
- pa = PartitionedDynamicShift(shape[0], self.partpoints)
- setattr(self.m.submodules, self.get_modname('ls'), pa)
- comb = self.m.d.comb
- if scalar:
- comb += pa.data.eq(op1)
- comb += pa.shifter.eq(op2)
- else:
- comb += pa.a.eq(op1)
- comb += pa.b.eq(op2)
- # XXX TODO: carry-in, carry-out
- #comb += pa.carry_in.eq(carry)
- return (pa.output, 0)
-
- def __lshift__(self, other):
- result, _ = self.ls_op(self, other, carry=0)
- return result
-
- def __rlshift__(self, other):
- raise NotImplementedError
- return Operator("<<", [other, self])
-
- def __rshift__(self, other):
- raise NotImplementedError
- return Operator(">>", [self, other])
-
- def __rrshift__(self, other):
- raise NotImplementedError
- return Operator(">>", [other, self])
-
def add_op(self, op1, op2, carry):
op1 = getsig(op1)
op2 = getsig(op2)
- shape = op1.shape()
- pa = PartitionedAdder(shape[0], self.partpoints)
+ pa = PartitionedAdder(len(op1), self.partpoints)
setattr(self.m.submodules, self.get_modname('add'), pa)
comb = self.m.d.comb
comb += pa.a.eq(op1)
comb += pa.b.eq(op2)
comb += pa.carry_in.eq(carry)
- return (pa.output, pa.carry_out)
+ result = SimdSignal.like(self)
+ comb += result.sig.eq(pa.output)
+ return result, pa.carry_out
def sub_op(self, op1, op2, carry=~0):
op1 = getsig(op1)
op2 = getsig(op2)
- shape = op1.shape()
- pa = PartitionedAdder(shape[0], self.partpoints)
+ pa = PartitionedAdder(len(op1), self.partpoints)
setattr(self.m.submodules, self.get_modname('add'), pa)
comb = self.m.d.comb
comb += pa.a.eq(op1)
comb += pa.b.eq(~op2)
comb += pa.carry_in.eq(carry)
- return (pa.output, pa.carry_out)
+ result = SimdSignal.like(self)
+ comb += result.sig.eq(pa.output)
+ return result, pa.carry_out
def __add__(self, other):
result, _ = self.add_op(self, other, carry=0)
return result
def __radd__(self, other):
+ # https://bugs.libre-soc.org/show_bug.cgi?id=718
result, _ = self.add_op(other, self)
return result
return result
def __rsub__(self, other):
+ # https://bugs.libre-soc.org/show_bug.cgi?id=718
result, _ = self.sub_op(other, self)
return result
def __mul__(self, other):
+ raise NotImplementedError # too complicated at the moment
return Operator("*", [self, other])
def __rmul__(self, other):
+ raise NotImplementedError # too complicated at the moment
return Operator("*", [other, self])
- def __check_divisor(self):
- width, signed = self.shape()
- if signed:
- # Python's division semantics and Verilog's division semantics
- # differ for negative divisors (Python uses div/mod, Verilog
- # uses quo/rem); for now, avoid the issue
- # completely by prohibiting such division operations.
- raise NotImplementedError(
- "Division by a signed value is not supported")
+ # not needed: same as Value.__check_divisor
+ # def __check_divisor(self):
def __mod__(self, other):
raise NotImplementedError
self.__check_divisor()
return Operator("//", [other, self])
+ # not needed: same as Value.__check_shamt
+ # def __check_shamt(self):
+
+ # TODO: detect if the 2nd operand is a Const, a Signal or a
+ # SimdSignal. if it's a Const or a Signal, a global shift
+ # can occur. if it's a SimdSignal, that's much more interesting.
+ def ls_op(self, op1, op2, carry, shr_flag=0):
+ op1 = getsig(op1)
+ if isinstance(op2, Const) or isinstance(op2, Signal):
+ scalar = True
+ pa = PartitionedScalarShift(len(op1), self.partpoints)
+ else:
+ scalar = False
+ op2 = getsig(op2)
+ pa = PartitionedDynamicShift(len(op1), self.partpoints)
+ # else:
+ # TODO: case where the *shifter* is a SimdSignal but
+ # the thing *being* Shifted is a scalar (Signal, expression)
+ # https://bugs.libre-soc.org/show_bug.cgi?id=718
+ setattr(self.m.submodules, self.get_modname('ls'), pa)
+ comb = self.m.d.comb
+ if scalar:
+ comb += pa.data.eq(op1)
+ comb += pa.shifter.eq(op2)
+ comb += pa.shift_right.eq(shr_flag)
+ else:
+ comb += pa.a.eq(op1)
+ comb += pa.b.eq(op2)
+ comb += pa.shift_right.eq(shr_flag)
+ # XXX TODO: carry-in, carry-out (for arithmetic shift)
+ #comb += pa.carry_in.eq(carry)
+ return (pa.output, 0)
+
+ def __lshift__(self, other):
+ z = Const(0, len(self.partpoints)+1)
+ result, _ = self.ls_op(self, other, carry=z) # TODO, carry
+ return result
+
+ def __rlshift__(self, other):
+ # https://bugs.libre-soc.org/show_bug.cgi?id=718
+ raise NotImplementedError
+ return Operator("<<", [other, self])
+
+ def __rshift__(self, other):
+ z = Const(0, len(self.partpoints)+1)
+ result, _ = self.ls_op(self, other, carry=z, shr_flag=1) # TODO, carry
+ return result
+
+ def __rrshift__(self, other):
+ # https://bugs.libre-soc.org/show_bug.cgi?id=718
+ raise NotImplementedError
+ return Operator(">>", [other, self])
+
+ # binary ops that don't require partitioning
+
+ def __and__(self, other):
+ return applyop(self, other, and_)
+
+ def __rand__(self, other):
+ return applyop(other, self, and_)
+
+ def __or__(self, other):
+ return applyop(self, other, or_)
+
+ def __ror__(self, other):
+ return applyop(other, self, or_)
+
+ def __xor__(self, other):
+ return applyop(self, other, xor)
+
+ def __rxor__(self, other):
+ return applyop(other, self, xor)
+
# binary comparison ops that need partitioning
def _compare(self, width, op1, op2, opname, optype):
setattr(self.m.submodules, self.get_modname(opname), pa)
comb = self.m.d.comb
comb += pa.opcode.eq(optype) # set opcode
- if isinstance(op1, PartitionedSignal):
+ if isinstance(op1, SimdSignal):
comb += pa.a.eq(op1.sig)
else:
comb += pa.a.eq(op1)
- if isinstance(op2, PartitionedSignal):
+ if isinstance(op2, SimdSignal):
comb += pa.b.eq(op2.sig)
else:
comb += pa.b.eq(op2)
return pa.output
def __eq__(self, other):
- width = self.sig.shape()[0]
+ width = len(self.sig)
return self._compare(width, self, other, "eq", PartitionedEqGtGe.EQ)
def __ne__(self, other):
- width = self.sig.shape()[0]
+ width = len(self.sig)
eq = self._compare(width, self, other, "eq", PartitionedEqGtGe.EQ)
ne = Signal(eq.width)
self.m.d.comb += ne.eq(~eq)
return ne
- def __gt__(self, other):
- width = self.sig.shape()[0]
- return self._compare(width, self, other, "gt", PartitionedEqGtGe.GT)
-
def __lt__(self, other):
- width = self.sig.shape()[0]
+ width = len(self.sig)
+ # swap operands, use gt to do lt
return self._compare(width, other, self, "gt", PartitionedEqGtGe.GT)
+ def __le__(self, other):
+ width = len(self.sig)
+ # swap operands, use ge to do le
+ return self._compare(width, other, self, "ge", PartitionedEqGtGe.GE)
+
+ def __gt__(self, other):
+ width = len(self.sig)
+ return self._compare(width, self, other, "gt", PartitionedEqGtGe.GT)
+
def __ge__(self, other):
- width = self.sig.shape()[0]
+ width = len(self.sig)
return self._compare(width, self, other, "ge", PartitionedEqGtGe.GE)
- def __le__(self, other):
- width = self.sig.shape()[0]
- return self._compare(width, other, self, "ge", PartitionedEqGtGe.GE)
+ # no override needed: Value.__abs__ is general enough it does the job
+ # def __abs__(self):
+
+ def __len__(self):
+ return len(self.sig)
+
+ # TODO, http://bugs.libre-riscv.org/show_bug.cgi?id=716
+ # def __getitem__(self, key):
+
+ def __new_sign(self, signed):
+ # XXX NO - SimdShape not Shape
+ print ("XXX requires SimdShape not Shape")
+ shape = Shape(len(self), signed=signed)
+ result = SimdSignal.like(self, shape=shape)
+ self.m.d.comb += result.sig.eq(self.sig)
+ return result
+
+ # http://bugs.libre-riscv.org/show_bug.cgi?id=719
+ def as_unsigned(self):
+ return self.__new_sign(False)
+
+ def as_signed(self):
+ return self.__new_sign(True)
# useful operators
Value, out
``1`` if any bits are set, ``0`` otherwise.
"""
- raise NotImplementedError
- return Operator("b", [self])
+ width = len(self.sig)
+ pa = PartitionedBool(width, self.partpoints)
+ setattr(self.m.submodules, self.get_modname("bool"), pa)
+ self.m.d.comb += pa.a.eq(self.sig)
+ return pa.output
def any(self):
"""Check if any bits are ``1``.
Value, out
``1`` if any bits are set, ``0`` otherwise.
"""
- raise NotImplementedError
+ return self != Const(0) # leverage the __ne__ operator here
return Operator("r|", [self])
def all(self):
Value, out
``1`` if all bits are set, ``0`` otherwise.
"""
- raise NotImplementedError
- return Operator("r&", [self])
+ # something wrong with PartitionedAll, but self == Const(-1)"
+ # XXX https://bugs.libre-soc.org/show_bug.cgi?id=176#c17
+ #width = len(self.sig)
+ #pa = PartitionedAll(width, self.partpoints)
+ #setattr(self.m.submodules, self.get_modname("all"), pa)
+ #self.m.d.comb += pa.a.eq(self.sig)
+ # return pa.output
+ return self == Const(-1) # leverage the __eq__ operator here
def xor(self):
"""Compute pairwise exclusive-or of every bit.
``1`` if an odd number of bits are set, ``0`` if an
even number of bits are set.
"""
- # XXXX TODO: return partition-mask-sized set of bits
- raise NotImplementedError
- return Operator("r^", [self])
+ width = len(self.sig)
+ pa = PartitionedXOR(width, self.partpoints)
+ setattr(self.m.submodules, self.get_modname("xor"), pa)
+ self.m.d.comb += pa.a.eq(self.sig)
+ return pa.output
- def implies(premise, conclusion):
- """Implication.
+ # not needed: Value.implies does the job
+ # def implies(premise, conclusion):
- Returns
- -------
- Value, out
- ``0`` if ``premise`` is true and ``conclusion`` is not,
- ``1`` otherwise.
- """
- # amazingly, this should actually work.
- return ~premise | conclusion
+ # TODO. contains a Value.cast which means an override is needed (on both)
+ # def bit_select(self, offset, width):
+ # def word_select(self, offset, width):
+
+ # not needed: Value.matches, amazingly, should do the job
+ # def matches(self, *patterns):
+
+ # TODO, http://bugs.libre-riscv.org/show_bug.cgi?id=713
+ def shape(self):
+ return self.sig.shape()
projects / ieee754fpu.git / blobdiff