1 # SPDX-License-Identifier: LGPL-2.1-or-later
2 # See Notices.txt for copyright information
3 """Integer Multiplication."""
5 from nmigen
import Signal
, Module
, Value
, Elaboratable
, Cat
, C
, Mux
, Repl
6 from nmigen
.hdl
.ast
import Assign
7 from abc
import ABCMeta
, abstractmethod
8 from nmigen
.cli
import main
9 from functools
import reduce
10 from operator
import or_
13 class PartitionPoints(dict):
14 """Partition points and corresponding ``Value``s.
16 The points at where an ALU is partitioned along with ``Value``s that
17 specify if the corresponding partition points are enabled.
19 For example: ``{1: True, 5: True, 10: True}`` with
20 ``width == 16`` specifies that the ALU is split into 4 sections:
23 * bits 5 <= ``i`` < 10
24 * bits 10 <= ``i`` < 16
26 If the partition_points were instead ``{1: True, 5: a, 10: True}``
27 where ``a`` is a 1-bit ``Signal``:
28 * If ``a`` is asserted:
31 * bits 5 <= ``i`` < 10
32 * bits 10 <= ``i`` < 16
35 * bits 1 <= ``i`` < 10
36 * bits 10 <= ``i`` < 16
39 def __init__(self
, partition_points
=None):
40 """Create a new ``PartitionPoints``.
42 :param partition_points: the input partition points to values mapping.
45 if partition_points
is not None:
46 for point
, enabled
in partition_points
.items():
47 if not isinstance(point
, int):
48 raise TypeError("point must be a non-negative integer")
50 raise ValueError("point must be a non-negative integer")
51 self
[point
] = Value
.wrap(enabled
)
53 def like(self
, name
=None, src_loc_at
=0, mul
=1):
54 """Create a new ``PartitionPoints`` with ``Signal``s for all values.
56 :param name: the base name for the new ``Signal``s.
57 :param mul: a multiplication factor on the indices
60 name
= Signal(src_loc_at
=1+src_loc_at
).name
# get variable name
61 retval
= PartitionPoints()
62 for point
, enabled
in self
.items():
64 retval
[point
] = Signal(enabled
.shape(), name
=f
"{name}_{point}")
68 """Assign ``PartitionPoints`` using ``Signal.eq``."""
69 if set(self
.keys()) != set(rhs
.keys()):
70 raise ValueError("incompatible point set")
71 for point
, enabled
in self
.items():
72 yield enabled
.eq(rhs
[point
])
74 def as_mask(self
, width
, mul
=1):
75 """Create a bit-mask from `self`.
77 Each bit in the returned mask is clear only if the partition point at
78 the same bit-index is enabled.
80 :param width: the bit width of the resulting mask
81 :param mul: a "multiplier" which in-place expands the partition points
82 typically set to "2" when used for multipliers
85 for i
in range(width
):
87 if i
.is_integer() and int(i
) in self
:
93 def get_max_partition_count(self
, width
):
94 """Get the maximum number of partitions.
96 Gets the number of partitions when all partition points are enabled.
99 for point
in self
.keys():
104 def fits_in_width(self
, width
):
105 """Check if all partition points are smaller than `width`."""
106 for point
in self
.keys():
111 def part_byte(self
, index
, mfactor
=1): # mfactor used for "expanding"
112 if index
== -1 or index
== 7:
114 assert index
>= 0 and index
< 8
115 return self
[(index
* 8 + 8)*mfactor
]
118 class FullAdder(Elaboratable
):
121 :attribute in0: the first input
122 :attribute in1: the second input
123 :attribute in2: the third input
124 :attribute sum: the sum output
125 :attribute carry: the carry output
127 Rather than do individual full adders (and have an array of them,
128 which would be very slow to simulate), this module can specify the
129 bit width of the inputs and outputs: in effect it performs multiple
130 Full 3-2 Add operations "in parallel".
133 def __init__(self
, width
):
134 """Create a ``FullAdder``.
136 :param width: the bit width of the input and output
138 self
.in0
= Signal(width
, reset_less
=True)
139 self
.in1
= Signal(width
, reset_less
=True)
140 self
.in2
= Signal(width
, reset_less
=True)
141 self
.sum = Signal(width
, reset_less
=True)
142 self
.carry
= Signal(width
, reset_less
=True)
144 def elaborate(self
, platform
):
145 """Elaborate this module."""
147 m
.d
.comb
+= self
.sum.eq(self
.in0 ^ self
.in1 ^ self
.in2
)
148 m
.d
.comb
+= self
.carry
.eq((self
.in0
& self
.in1
)
149 |
(self
.in1
& self
.in2
)
150 |
(self
.in2
& self
.in0
))
154 class MaskedFullAdder(Elaboratable
):
155 """Masked Full Adder.
157 :attribute mask: the carry partition mask
158 :attribute in0: the first input
159 :attribute in1: the second input
160 :attribute in2: the third input
161 :attribute sum: the sum output
162 :attribute mcarry: the masked carry output
164 FullAdders are always used with a "mask" on the output. To keep
165 the graphviz "clean", this class performs the masking here rather
166 than inside a large for-loop.
168 See the following discussion as to why this is no longer derived
169 from FullAdder. Each carry is shifted here *before* being ANDed
170 with the mask, so that an AOI cell may be used (which is more
172 https://en.wikipedia.org/wiki/AND-OR-Invert
173 https://groups.google.com/d/msg/comp.arch/fcq-GLQqvas/vTxmcA0QAgAJ
176 def __init__(self
, width
):
177 """Create a ``MaskedFullAdder``.
179 :param width: the bit width of the input and output
182 self
.mask
= Signal(width
, reset_less
=True)
183 self
.mcarry
= Signal(width
, reset_less
=True)
184 self
.in0
= Signal(width
, reset_less
=True)
185 self
.in1
= Signal(width
, reset_less
=True)
186 self
.in2
= Signal(width
, reset_less
=True)
187 self
.sum = Signal(width
, reset_less
=True)
189 def elaborate(self
, platform
):
190 """Elaborate this module."""
192 s1
= Signal(self
.width
, reset_less
=True)
193 s2
= Signal(self
.width
, reset_less
=True)
194 s3
= Signal(self
.width
, reset_less
=True)
195 c1
= Signal(self
.width
, reset_less
=True)
196 c2
= Signal(self
.width
, reset_less
=True)
197 c3
= Signal(self
.width
, reset_less
=True)
198 m
.d
.comb
+= self
.sum.eq(self
.in0 ^ self
.in1 ^ self
.in2
)
199 m
.d
.comb
+= s1
.eq(Cat(0, self
.in0
))
200 m
.d
.comb
+= s2
.eq(Cat(0, self
.in1
))
201 m
.d
.comb
+= s3
.eq(Cat(0, self
.in2
))
202 m
.d
.comb
+= c1
.eq(s1
& s2
& self
.mask
)
203 m
.d
.comb
+= c2
.eq(s2
& s3
& self
.mask
)
204 m
.d
.comb
+= c3
.eq(s3
& s1
& self
.mask
)
205 m
.d
.comb
+= self
.mcarry
.eq(c1 | c2 | c3
)
209 class PartitionedAdder(Elaboratable
):
210 """Partitioned Adder.
212 Performs the final add. The partition points are included in the
213 actual add (in one of the operands only), which causes a carry over
214 to the next bit. Then the final output *removes* the extra bits from
217 partition: .... P... P... P... P... (32 bits)
218 a : .... .... .... .... .... (32 bits)
219 b : .... .... .... .... .... (32 bits)
220 exp-a : ....P....P....P....P.... (32+4 bits, P=1 if no partition)
221 exp-b : ....0....0....0....0.... (32 bits plus 4 zeros)
222 exp-o : ....xN...xN...xN...xN... (32+4 bits - x to be discarded)
223 o : .... N... N... N... N... (32 bits - x ignored, N is carry-over)
225 :attribute width: the bit width of the input and output. Read-only.
226 :attribute a: the first input to the adder
227 :attribute b: the second input to the adder
228 :attribute output: the sum output
229 :attribute partition_points: the input partition points. Modification not
230 supported, except for by ``Signal.eq``.
233 def __init__(self
, width
, partition_points
, partition_step
=1):
234 """Create a ``PartitionedAdder``.
236 :param width: the bit width of the input and output
237 :param partition_points: the input partition points
238 :param partition_step: a multiplier (typically double) step
239 which in-place "expands" the partition points
242 self
.pmul
= partition_step
243 self
.a
= Signal(width
, reset_less
=True)
244 self
.b
= Signal(width
, reset_less
=True)
245 self
.output
= Signal(width
, reset_less
=True)
246 self
.partition_points
= PartitionPoints(partition_points
)
247 if not self
.partition_points
.fits_in_width(width
):
248 raise ValueError("partition_points doesn't fit in width")
250 for i
in range(self
.width
):
251 if i
in self
.partition_points
:
254 self
._expanded
_width
= expanded_width
256 def elaborate(self
, platform
):
257 """Elaborate this module."""
259 expanded_a
= Signal(self
._expanded
_width
, reset_less
=True)
260 expanded_b
= Signal(self
._expanded
_width
, reset_less
=True)
261 expanded_o
= Signal(self
._expanded
_width
, reset_less
=True)
264 # store bits in a list, use Cat later. graphviz is much cleaner
265 al
, bl
, ol
, ea
, eb
, eo
= [],[],[],[],[],[]
267 # partition points are "breaks" (extra zeros or 1s) in what would
268 # otherwise be a massive long add. when the "break" points are 0,
269 # whatever is in it (in the output) is discarded. however when
270 # there is a "1", it causes a roll-over carry to the *next* bit.
271 # we still ignore the "break" bit in the [intermediate] output,
272 # however by that time we've got the effect that we wanted: the
273 # carry has been carried *over* the break point.
275 for i
in range(self
.width
):
276 pi
= i
/self
.pmul
# double the range of the partition point test
277 if pi
.is_integer() and pi
in self
.partition_points
:
278 # add extra bit set to 0 + 0 for enabled partition points
279 # and 1 + 0 for disabled partition points
280 ea
.append(expanded_a
[expanded_index
])
281 al
.append(~self
.partition_points
[pi
]) # add extra bit in a
282 eb
.append(expanded_b
[expanded_index
])
283 bl
.append(C(0)) # yes, add a zero
284 expanded_index
+= 1 # skip the extra point. NOT in the output
285 ea
.append(expanded_a
[expanded_index
])
286 eb
.append(expanded_b
[expanded_index
])
287 eo
.append(expanded_o
[expanded_index
])
290 ol
.append(self
.output
[i
])
293 # combine above using Cat
294 m
.d
.comb
+= Cat(*ea
).eq(Cat(*al
))
295 m
.d
.comb
+= Cat(*eb
).eq(Cat(*bl
))
296 m
.d
.comb
+= Cat(*ol
).eq(Cat(*eo
))
298 # use only one addition to take advantage of look-ahead carry and
299 # special hardware on FPGAs
300 m
.d
.comb
+= expanded_o
.eq(expanded_a
+ expanded_b
)
304 FULL_ADDER_INPUT_COUNT
= 3
308 def __init__(self
, ppoints
, n_inputs
, output_width
, n_parts
):
309 self
.part_ops
= [Signal(2, name
=f
"part_ops_{i}", reset_less
=True)
310 for i
in range(n_parts
)]
311 self
.inputs
= [Signal(output_width
, name
=f
"inputs_{i}",
313 for i
in range(n_inputs
)]
314 self
.reg_partition_points
= ppoints
.like()
316 def eq_from(self
, reg_partition_points
, inputs
, part_ops
):
317 return [self
.reg_partition_points
.eq(reg_partition_points
)] + \
318 [self
.inputs
[i
].eq(inputs
[i
])
319 for i
in range(len(self
.inputs
))] + \
320 [self
.part_ops
[i
].eq(part_ops
[i
])
321 for i
in range(len(self
.part_ops
))]
324 return self
.eq_from(rhs
.reg_partition_points
, rhs
.inputs
, rhs
.part_ops
)
327 class FinalReduceData
:
329 def __init__(self
, ppoints
, output_width
, n_parts
):
330 self
.part_ops
= [Signal(2, name
=f
"part_ops_{i}", reset_less
=True)
331 for i
in range(n_parts
)]
332 self
.output
= Signal(output_width
, reset_less
=True)
333 self
.reg_partition_points
= ppoints
.like()
335 def eq_from(self
, reg_partition_points
, output
, part_ops
):
336 return [self
.reg_partition_points
.eq(reg_partition_points
)] + \
337 [self
.output
.eq(output
)] + \
338 [self
.part_ops
[i
].eq(part_ops
[i
])
339 for i
in range(len(self
.part_ops
))]
342 return self
.eq_from(rhs
.reg_partition_points
, rhs
.output
, rhs
.part_ops
)
345 class FinalAdd(Elaboratable
):
346 """ Final stage of add reduce
349 def __init__(self
, n_inputs
, output_width
, n_parts
, register_levels
,
351 self
.i
= AddReduceData(partition_points
, n_inputs
,
352 output_width
, n_parts
)
353 self
.o
= FinalReduceData(partition_points
, output_width
, n_parts
)
354 self
.output_width
= output_width
355 self
.n_inputs
= n_inputs
356 self
.n_parts
= n_parts
357 self
.register_levels
= list(register_levels
)
358 self
.partition_points
= PartitionPoints(partition_points
)
359 if not self
.partition_points
.fits_in_width(output_width
):
360 raise ValueError("partition_points doesn't fit in output_width")
362 def elaborate(self
, platform
):
363 """Elaborate this module."""
366 output_width
= self
.output_width
367 output
= Signal(output_width
, reset_less
=True)
368 if self
.n_inputs
== 0:
369 # use 0 as the default output value
370 m
.d
.comb
+= output
.eq(0)
371 elif self
.n_inputs
== 1:
372 # handle single input
373 m
.d
.comb
+= output
.eq(self
.i
.inputs
[0])
375 # base case for adding 2 inputs
376 assert self
.n_inputs
== 2
377 adder
= PartitionedAdder(output_width
, self
.i
.reg_partition_points
)
378 m
.submodules
.final_adder
= adder
379 m
.d
.comb
+= adder
.a
.eq(self
.i
.inputs
[0])
380 m
.d
.comb
+= adder
.b
.eq(self
.i
.inputs
[1])
381 m
.d
.comb
+= output
.eq(adder
.output
)
384 m
.d
.comb
+= self
.o
.eq_from(self
.i
.reg_partition_points
, output
,
390 class AddReduceSingle(Elaboratable
):
391 """Add list of numbers together.
393 :attribute inputs: input ``Signal``s to be summed. Modification not
394 supported, except for by ``Signal.eq``.
395 :attribute register_levels: List of nesting levels that should have
397 :attribute output: output sum.
398 :attribute partition_points: the input partition points. Modification not
399 supported, except for by ``Signal.eq``.
402 def __init__(self
, n_inputs
, output_width
, n_parts
, register_levels
,
404 """Create an ``AddReduce``.
406 :param inputs: input ``Signal``s to be summed.
407 :param output_width: bit-width of ``output``.
408 :param register_levels: List of nesting levels that should have
410 :param partition_points: the input partition points.
412 self
.n_inputs
= n_inputs
413 self
.n_parts
= n_parts
414 self
.output_width
= output_width
415 self
.i
= AddReduceData(partition_points
, n_inputs
,
416 output_width
, n_parts
)
417 self
.register_levels
= list(register_levels
)
418 self
.partition_points
= PartitionPoints(partition_points
)
419 if not self
.partition_points
.fits_in_width(output_width
):
420 raise ValueError("partition_points doesn't fit in output_width")
422 max_level
= AddReduceSingle
.get_max_level(n_inputs
)
423 for level
in self
.register_levels
:
424 if level
> max_level
:
426 "not enough adder levels for specified register levels")
428 self
.groups
= AddReduceSingle
.full_adder_groups(n_inputs
)
429 n_terms
= AddReduceSingle
.calc_n_inputs(n_inputs
, self
.groups
)
430 self
.o
= AddReduceData(partition_points
, n_terms
, output_width
, n_parts
)
433 def calc_n_inputs(n_inputs
, groups
):
434 retval
= len(groups
)*2
435 if n_inputs
% FULL_ADDER_INPUT_COUNT
== 1:
437 elif n_inputs
% FULL_ADDER_INPUT_COUNT
== 2:
440 assert n_inputs
% FULL_ADDER_INPUT_COUNT
== 0
444 def get_max_level(input_count
):
445 """Get the maximum level.
447 All ``register_levels`` must be less than or equal to the maximum
452 groups
= AddReduceSingle
.full_adder_groups(input_count
)
455 input_count
%= FULL_ADDER_INPUT_COUNT
456 input_count
+= 2 * len(groups
)
460 def full_adder_groups(input_count
):
461 """Get ``inputs`` indices for which a full adder should be built."""
463 input_count
- FULL_ADDER_INPUT_COUNT
+ 1,
464 FULL_ADDER_INPUT_COUNT
)
466 def create_next_terms(self
):
467 """ create next intermediate terms, for linking up in elaborate, below
472 # create full adders for this recursive level.
473 # this shrinks N terms to 2 * (N // 3) plus the remainder
474 for i
in self
.groups
:
475 adder_i
= MaskedFullAdder(self
.output_width
)
476 adders
.append((i
, adder_i
))
477 # add both the sum and the masked-carry to the next level.
478 # 3 inputs have now been reduced to 2...
479 terms
.append(adder_i
.sum)
480 terms
.append(adder_i
.mcarry
)
481 # handle the remaining inputs.
482 if self
.n_inputs
% FULL_ADDER_INPUT_COUNT
== 1:
483 terms
.append(self
.i
.inputs
[-1])
484 elif self
.n_inputs
% FULL_ADDER_INPUT_COUNT
== 2:
485 # Just pass the terms to the next layer, since we wouldn't gain
486 # anything by using a half adder since there would still be 2 terms
487 # and just passing the terms to the next layer saves gates.
488 terms
.append(self
.i
.inputs
[-2])
489 terms
.append(self
.i
.inputs
[-1])
491 assert self
.n_inputs
% FULL_ADDER_INPUT_COUNT
== 0
495 def elaborate(self
, platform
):
496 """Elaborate this module."""
499 terms
, adders
= self
.create_next_terms()
501 # copy the intermediate terms to the output
502 for i
, value
in enumerate(terms
):
503 m
.d
.comb
+= self
.o
.inputs
[i
].eq(value
)
505 # copy reg part points and part ops to output
506 m
.d
.comb
+= self
.o
.reg_partition_points
.eq(self
.i
.reg_partition_points
)
507 m
.d
.comb
+= [self
.o
.part_ops
[i
].eq(self
.i
.part_ops
[i
])
508 for i
in range(len(self
.i
.part_ops
))]
510 # set up the partition mask (for the adders)
511 part_mask
= Signal(self
.output_width
, reset_less
=True)
513 # get partition points as a mask
514 mask
= self
.i
.reg_partition_points
.as_mask(self
.output_width
, mul
=2)
515 m
.d
.comb
+= part_mask
.eq(mask
)
517 # add and link the intermediate term modules
518 for i
, (iidx
, adder_i
) in enumerate(adders
):
519 setattr(m
.submodules
, f
"adder_{i}", adder_i
)
521 m
.d
.comb
+= adder_i
.in0
.eq(self
.i
.inputs
[iidx
])
522 m
.d
.comb
+= adder_i
.in1
.eq(self
.i
.inputs
[iidx
+ 1])
523 m
.d
.comb
+= adder_i
.in2
.eq(self
.i
.inputs
[iidx
+ 2])
524 m
.d
.comb
+= adder_i
.mask
.eq(part_mask
)
529 class AddReduce(Elaboratable
):
530 """Recursively Add list of numbers together.
532 :attribute inputs: input ``Signal``s to be summed. Modification not
533 supported, except for by ``Signal.eq``.
534 :attribute register_levels: List of nesting levels that should have
536 :attribute output: output sum.
537 :attribute partition_points: the input partition points. Modification not
538 supported, except for by ``Signal.eq``.
541 def __init__(self
, inputs
, output_width
, register_levels
, partition_points
,
543 """Create an ``AddReduce``.
545 :param inputs: input ``Signal``s to be summed.
546 :param output_width: bit-width of ``output``.
547 :param register_levels: List of nesting levels that should have
549 :param partition_points: the input partition points.
552 self
.part_ops
= part_ops
553 n_parts
= len(part_ops
)
554 self
.o
= FinalReduceData(partition_points
, output_width
, n_parts
)
555 self
.output_width
= output_width
556 self
.register_levels
= register_levels
557 self
.partition_points
= partition_points
562 def get_max_level(input_count
):
563 return AddReduceSingle
.get_max_level(input_count
)
566 def next_register_levels(register_levels
):
567 """``Iterable`` of ``register_levels`` for next recursive level."""
568 for level
in register_levels
:
572 def create_levels(self
):
573 """creates reduction levels"""
576 next_levels
= self
.register_levels
577 partition_points
= self
.partition_points
578 part_ops
= self
.part_ops
579 n_parts
= len(part_ops
)
583 groups
= AddReduceSingle
.full_adder_groups(len(inputs
))
586 next_level
= AddReduceSingle(ilen
, self
.output_width
, n_parts
,
587 next_levels
, partition_points
)
588 mods
.append(next_level
)
589 next_levels
= list(AddReduce
.next_register_levels(next_levels
))
590 partition_points
= next_level
.i
.reg_partition_points
591 inputs
= next_level
.o
.inputs
593 part_ops
= next_level
.i
.part_ops
595 next_level
= FinalAdd(ilen
, self
.output_width
, n_parts
,
596 next_levels
, partition_points
)
597 mods
.append(next_level
)
601 def elaborate(self
, platform
):
602 """Elaborate this module."""
605 for i
, next_level
in enumerate(self
.levels
):
606 setattr(m
.submodules
, "next_level%d" % i
, next_level
)
608 partition_points
= self
.partition_points
610 part_ops
= self
.part_ops
611 n_parts
= len(part_ops
)
612 n_inputs
= len(inputs
)
613 output_width
= self
.output_width
614 i
= AddReduceData(partition_points
, n_inputs
, output_width
, n_parts
)
615 m
.d
.comb
+= i
.eq_from(partition_points
, inputs
, part_ops
)
616 for idx
in range(len(self
.levels
)):
617 mcur
= self
.levels
[idx
]
618 if 0 in mcur
.register_levels
:
619 m
.d
.sync
+= mcur
.i
.eq(i
)
621 m
.d
.comb
+= mcur
.i
.eq(i
)
622 i
= mcur
.o
# for next loop
624 # output comes from last module
625 m
.d
.comb
+= self
.o
.eq(i
)
631 OP_MUL_SIGNED_HIGH
= 1
632 OP_MUL_SIGNED_UNSIGNED_HIGH
= 2 # a is signed, b is unsigned
633 OP_MUL_UNSIGNED_HIGH
= 3
636 def get_term(value
, shift
=0, enabled
=None):
637 if enabled
is not None:
638 value
= Mux(enabled
, value
, 0)
640 value
= Cat(Repl(C(0, 1), shift
), value
)
646 class ProductTerm(Elaboratable
):
647 """ this class creates a single product term (a[..]*b[..]).
648 it has a design flaw in that is the *output* that is selected,
649 where the multiplication(s) are combinatorially generated
653 def __init__(self
, width
, twidth
, pbwid
, a_index
, b_index
):
654 self
.a_index
= a_index
655 self
.b_index
= b_index
656 shift
= 8 * (self
.a_index
+ self
.b_index
)
662 self
.ti
= Signal(self
.width
, reset_less
=True)
663 self
.term
= Signal(twidth
, reset_less
=True)
664 self
.a
= Signal(twidth
//2, reset_less
=True)
665 self
.b
= Signal(twidth
//2, reset_less
=True)
666 self
.pb_en
= Signal(pbwid
, reset_less
=True)
669 min_index
= min(self
.a_index
, self
.b_index
)
670 max_index
= max(self
.a_index
, self
.b_index
)
671 for i
in range(min_index
, max_index
):
672 tl
.append(self
.pb_en
[i
])
673 name
= "te_%d_%d" % (self
.a_index
, self
.b_index
)
675 term_enabled
= Signal(name
=name
, reset_less
=True)
678 self
.enabled
= term_enabled
679 self
.term
.name
= "term_%d_%d" % (a_index
, b_index
) # rename
681 def elaborate(self
, platform
):
684 if self
.enabled
is not None:
685 m
.d
.comb
+= self
.enabled
.eq(~
(Cat(*self
.tl
).bool()))
687 bsa
= Signal(self
.width
, reset_less
=True)
688 bsb
= Signal(self
.width
, reset_less
=True)
689 a_index
, b_index
= self
.a_index
, self
.b_index
691 m
.d
.comb
+= bsa
.eq(self
.a
.part(a_index
* pwidth
, pwidth
))
692 m
.d
.comb
+= bsb
.eq(self
.b
.part(b_index
* pwidth
, pwidth
))
693 m
.d
.comb
+= self
.ti
.eq(bsa
* bsb
)
694 m
.d
.comb
+= self
.term
.eq(get_term(self
.ti
, self
.shift
, self
.enabled
))
696 #TODO: sort out width issues, get inputs a/b switched on/off.
697 #data going into Muxes is 1/2 the required width
701 bsa = Signal(self.twidth//2, reset_less=True)
702 bsb = Signal(self.twidth//2, reset_less=True)
703 asel = Signal(width, reset_less=True)
704 bsel = Signal(width, reset_less=True)
705 a_index, b_index = self.a_index, self.b_index
706 m.d.comb += asel.eq(self.a.part(a_index * pwidth, pwidth))
707 m.d.comb += bsel.eq(self.b.part(b_index * pwidth, pwidth))
708 m.d.comb += bsa.eq(get_term(asel, self.shift, self.enabled))
709 m.d.comb += bsb.eq(get_term(bsel, self.shift, self.enabled))
710 m.d.comb += self.ti.eq(bsa * bsb)
711 m.d.comb += self.term.eq(self.ti)
717 class ProductTerms(Elaboratable
):
718 """ creates a bank of product terms. also performs the actual bit-selection
719 this class is to be wrapped with a for-loop on the "a" operand.
720 it creates a second-level for-loop on the "b" operand.
722 def __init__(self
, width
, twidth
, pbwid
, a_index
, blen
):
723 self
.a_index
= a_index
728 self
.a
= Signal(twidth
//2, reset_less
=True)
729 self
.b
= Signal(twidth
//2, reset_less
=True)
730 self
.pb_en
= Signal(pbwid
, reset_less
=True)
731 self
.terms
= [Signal(twidth
, name
="term%d"%i, reset_less
=True) \
732 for i
in range(blen
)]
734 def elaborate(self
, platform
):
738 for b_index
in range(self
.blen
):
739 t
= ProductTerm(self
.pwidth
, self
.twidth
, self
.pbwid
,
740 self
.a_index
, b_index
)
741 setattr(m
.submodules
, "term_%d" % b_index
, t
)
743 m
.d
.comb
+= t
.a
.eq(self
.a
)
744 m
.d
.comb
+= t
.b
.eq(self
.b
)
745 m
.d
.comb
+= t
.pb_en
.eq(self
.pb_en
)
747 m
.d
.comb
+= self
.terms
[b_index
].eq(t
.term
)
752 class LSBNegTerm(Elaboratable
):
754 def __init__(self
, bit_width
):
755 self
.bit_width
= bit_width
756 self
.part
= Signal(reset_less
=True)
757 self
.signed
= Signal(reset_less
=True)
758 self
.op
= Signal(bit_width
, reset_less
=True)
759 self
.msb
= Signal(reset_less
=True)
760 self
.nt
= Signal(bit_width
*2, reset_less
=True)
761 self
.nl
= Signal(bit_width
*2, reset_less
=True)
763 def elaborate(self
, platform
):
766 bit_wid
= self
.bit_width
767 ext
= Repl(0, bit_wid
) # extend output to HI part
769 # determine sign of each incoming number *in this partition*
770 enabled
= Signal(reset_less
=True)
771 m
.d
.comb
+= enabled
.eq(self
.part
& self
.msb
& self
.signed
)
773 # for 8-bit values: form a * 0xFF00 by using -a * 0x100, the
774 # negation operation is split into a bitwise not and a +1.
775 # likewise for 16, 32, and 64-bit values.
777 # width-extended 1s complement if a is signed, otherwise zero
778 comb
+= self
.nt
.eq(Mux(enabled
, Cat(ext
, ~self
.op
), 0))
780 # add 1 if signed, otherwise add zero
781 comb
+= self
.nl
.eq(Cat(ext
, enabled
, Repl(0, bit_wid
-1)))
786 class Parts(Elaboratable
):
788 def __init__(self
, pbwid
, epps
, n_parts
):
791 self
.epps
= PartitionPoints
.like(epps
, name
="epps") # expanded points
793 self
.parts
= [Signal(name
=f
"part_{i}", reset_less
=True)
794 for i
in range(n_parts
)]
796 def elaborate(self
, platform
):
799 epps
, parts
= self
.epps
, self
.parts
800 # collect part-bytes (double factor because the input is extended)
801 pbs
= Signal(self
.pbwid
, reset_less
=True)
803 for i
in range(self
.pbwid
):
804 pb
= Signal(name
="pb%d" % i
, reset_less
=True)
805 m
.d
.comb
+= pb
.eq(epps
.part_byte(i
))
807 m
.d
.comb
+= pbs
.eq(Cat(*tl
))
809 # negated-temporary copy of partition bits
810 npbs
= Signal
.like(pbs
, reset_less
=True)
811 m
.d
.comb
+= npbs
.eq(~pbs
)
812 byte_count
= 8 // len(parts
)
813 for i
in range(len(parts
)):
815 pbl
.append(npbs
[i
* byte_count
- 1])
816 for j
in range(i
* byte_count
, (i
+ 1) * byte_count
- 1):
818 pbl
.append(npbs
[(i
+ 1) * byte_count
- 1])
819 value
= Signal(len(pbl
), name
="value_%d" % i
, reset_less
=True)
820 m
.d
.comb
+= value
.eq(Cat(*pbl
))
821 m
.d
.comb
+= parts
[i
].eq(~
(value
).bool())
826 class Part(Elaboratable
):
827 """ a key class which, depending on the partitioning, will determine
828 what action to take when parts of the output are signed or unsigned.
830 this requires 2 pieces of data *per operand, per partition*:
831 whether the MSB is HI/LO (per partition!), and whether a signed
832 or unsigned operation has been *requested*.
834 once that is determined, signed is basically carried out
835 by splitting 2's complement into 1's complement plus one.
836 1's complement is just a bit-inversion.
838 the extra terms - as separate terms - are then thrown at the
839 AddReduce alongside the multiplication part-results.
841 def __init__(self
, epps
, width
, n_parts
, n_levels
, pbwid
):
847 self
.a
= Signal(64, reset_less
=True)
848 self
.b
= Signal(64, reset_less
=True)
849 self
.a_signed
= [Signal(name
=f
"a_signed_{i}", reset_less
=True)
851 self
.b_signed
= [Signal(name
=f
"_b_signed_{i}", reset_less
=True)
853 self
.pbs
= Signal(pbwid
, reset_less
=True)
856 self
.parts
= [Signal(name
=f
"part_{i}", reset_less
=True)
857 for i
in range(n_parts
)]
859 self
.not_a_term
= Signal(width
, reset_less
=True)
860 self
.neg_lsb_a_term
= Signal(width
, reset_less
=True)
861 self
.not_b_term
= Signal(width
, reset_less
=True)
862 self
.neg_lsb_b_term
= Signal(width
, reset_less
=True)
864 def elaborate(self
, platform
):
867 pbs
, parts
= self
.pbs
, self
.parts
869 m
.submodules
.p
= p
= Parts(self
.pbwid
, epps
, len(parts
))
870 m
.d
.comb
+= p
.epps
.eq(epps
)
873 byte_count
= 8 // len(parts
)
875 not_a_term
, neg_lsb_a_term
, not_b_term
, neg_lsb_b_term
= (
876 self
.not_a_term
, self
.neg_lsb_a_term
,
877 self
.not_b_term
, self
.neg_lsb_b_term
)
879 byte_width
= 8 // len(parts
) # byte width
880 bit_wid
= 8 * byte_width
# bit width
881 nat
, nbt
, nla
, nlb
= [], [], [], []
882 for i
in range(len(parts
)):
883 # work out bit-inverted and +1 term for a.
884 pa
= LSBNegTerm(bit_wid
)
885 setattr(m
.submodules
, "lnt_%d_a_%d" % (bit_wid
, i
), pa
)
886 m
.d
.comb
+= pa
.part
.eq(parts
[i
])
887 m
.d
.comb
+= pa
.op
.eq(self
.a
.part(bit_wid
* i
, bit_wid
))
888 m
.d
.comb
+= pa
.signed
.eq(self
.b_signed
[i
* byte_width
]) # yes b
889 m
.d
.comb
+= pa
.msb
.eq(self
.b
[(i
+ 1) * bit_wid
- 1]) # really, b
893 # work out bit-inverted and +1 term for b
894 pb
= LSBNegTerm(bit_wid
)
895 setattr(m
.submodules
, "lnt_%d_b_%d" % (bit_wid
, i
), pb
)
896 m
.d
.comb
+= pb
.part
.eq(parts
[i
])
897 m
.d
.comb
+= pb
.op
.eq(self
.b
.part(bit_wid
* i
, bit_wid
))
898 m
.d
.comb
+= pb
.signed
.eq(self
.a_signed
[i
* byte_width
]) # yes a
899 m
.d
.comb
+= pb
.msb
.eq(self
.a
[(i
+ 1) * bit_wid
- 1]) # really, a
903 # concatenate together and return all 4 results.
904 m
.d
.comb
+= [not_a_term
.eq(Cat(*nat
)),
905 not_b_term
.eq(Cat(*nbt
)),
906 neg_lsb_a_term
.eq(Cat(*nla
)),
907 neg_lsb_b_term
.eq(Cat(*nlb
)),
913 class IntermediateOut(Elaboratable
):
914 """ selects the HI/LO part of the multiplication, for a given bit-width
915 the output is also reconstructed in its SIMD (partition) lanes.
917 def __init__(self
, width
, out_wid
, n_parts
):
919 self
.n_parts
= n_parts
920 self
.part_ops
= [Signal(2, name
="dpop%d" % i
, reset_less
=True)
922 self
.intermed
= Signal(out_wid
, reset_less
=True)
923 self
.output
= Signal(out_wid
//2, reset_less
=True)
925 def elaborate(self
, platform
):
931 for i
in range(self
.n_parts
):
932 op
= Signal(w
, reset_less
=True, name
="op%d_%d" % (w
, i
))
934 Mux(self
.part_ops
[sel
* i
] == OP_MUL_LOW
,
935 self
.intermed
.part(i
* w
*2, w
),
936 self
.intermed
.part(i
* w
*2 + w
, w
)))
938 m
.d
.comb
+= self
.output
.eq(Cat(*ol
))
943 class FinalOut(Elaboratable
):
944 """ selects the final output based on the partitioning.
946 each byte is selectable independently, i.e. it is possible
947 that some partitions requested 8-bit computation whilst others
948 requested 16 or 32 bit.
950 def __init__(self
, output_width
, n_parts
, partition_points
):
951 self
.expanded_part_points
= partition_points
952 self
.i
= IntermediateData(partition_points
, output_width
, n_parts
)
953 self
.out_wid
= output_width
//2
955 self
.out
= Signal(self
.out_wid
, reset_less
=True)
956 self
.intermediate_output
= Signal(output_width
, reset_less
=True)
958 def elaborate(self
, platform
):
961 eps
= self
.expanded_part_points
962 m
.submodules
.p_8
= p_8
= Parts(8, eps
, 8)
963 m
.submodules
.p_16
= p_16
= Parts(8, eps
, 4)
964 m
.submodules
.p_32
= p_32
= Parts(8, eps
, 2)
965 m
.submodules
.p_64
= p_64
= Parts(8, eps
, 1)
967 out_part_pts
= self
.i
.reg_partition_points
970 d8
= [Signal(name
=f
"d8_{i}", reset_less
=True) for i
in range(8)]
971 d16
= [Signal(name
=f
"d16_{i}", reset_less
=True) for i
in range(4)]
972 d32
= [Signal(name
=f
"d32_{i}", reset_less
=True) for i
in range(2)]
974 i8
= Signal(self
.out_wid
, reset_less
=True)
975 i16
= Signal(self
.out_wid
, reset_less
=True)
976 i32
= Signal(self
.out_wid
, reset_less
=True)
977 i64
= Signal(self
.out_wid
, reset_less
=True)
979 m
.d
.comb
+= p_8
.epps
.eq(out_part_pts
)
980 m
.d
.comb
+= p_16
.epps
.eq(out_part_pts
)
981 m
.d
.comb
+= p_32
.epps
.eq(out_part_pts
)
982 m
.d
.comb
+= p_64
.epps
.eq(out_part_pts
)
984 for i
in range(len(p_8
.parts
)):
985 m
.d
.comb
+= d8
[i
].eq(p_8
.parts
[i
])
986 for i
in range(len(p_16
.parts
)):
987 m
.d
.comb
+= d16
[i
].eq(p_16
.parts
[i
])
988 for i
in range(len(p_32
.parts
)):
989 m
.d
.comb
+= d32
[i
].eq(p_32
.parts
[i
])
990 m
.d
.comb
+= i8
.eq(self
.i
.outputs
[0])
991 m
.d
.comb
+= i16
.eq(self
.i
.outputs
[1])
992 m
.d
.comb
+= i32
.eq(self
.i
.outputs
[2])
993 m
.d
.comb
+= i64
.eq(self
.i
.outputs
[3])
997 # select one of the outputs: d8 selects i8, d16 selects i16
998 # d32 selects i32, and the default is i64.
999 # d8 and d16 are ORed together in the first Mux
1000 # then the 2nd selects either i8 or i16.
1001 # if neither d8 nor d16 are set, d32 selects either i32 or i64.
1002 op
= Signal(8, reset_less
=True, name
="op_%d" % i
)
1004 Mux(d8
[i
] | d16
[i
// 2],
1005 Mux(d8
[i
], i8
.part(i
* 8, 8), i16
.part(i
* 8, 8)),
1006 Mux(d32
[i
// 4], i32
.part(i
* 8, 8), i64
.part(i
* 8, 8))))
1008 m
.d
.comb
+= self
.out
.eq(Cat(*ol
))
1009 m
.d
.comb
+= self
.intermediate_output
.eq(self
.i
.intermediate_output
)
1013 class OrMod(Elaboratable
):
1014 """ ORs four values together in a hierarchical tree
1016 def __init__(self
, wid
):
1018 self
.orin
= [Signal(wid
, name
="orin%d" % i
, reset_less
=True)
1020 self
.orout
= Signal(wid
, reset_less
=True)
1022 def elaborate(self
, platform
):
1024 or1
= Signal(self
.wid
, reset_less
=True)
1025 or2
= Signal(self
.wid
, reset_less
=True)
1026 m
.d
.comb
+= or1
.eq(self
.orin
[0] | self
.orin
[1])
1027 m
.d
.comb
+= or2
.eq(self
.orin
[2] | self
.orin
[3])
1028 m
.d
.comb
+= self
.orout
.eq(or1 | or2
)
1033 class Signs(Elaboratable
):
1034 """ determines whether a or b are signed numbers
1035 based on the required operation type (OP_MUL_*)
1039 self
.part_ops
= Signal(2, reset_less
=True)
1040 self
.a_signed
= Signal(reset_less
=True)
1041 self
.b_signed
= Signal(reset_less
=True)
1043 def elaborate(self
, platform
):
1047 asig
= self
.part_ops
!= OP_MUL_UNSIGNED_HIGH
1048 bsig
= (self
.part_ops
== OP_MUL_LOW
) \
1049 |
(self
.part_ops
== OP_MUL_SIGNED_HIGH
)
1050 m
.d
.comb
+= self
.a_signed
.eq(asig
)
1051 m
.d
.comb
+= self
.b_signed
.eq(bsig
)
1056 class IntermediateData
:
1058 def __init__(self
, ppoints
, output_width
, n_parts
):
1059 self
.part_ops
= [Signal(2, name
=f
"part_ops_{i}", reset_less
=True)
1060 for i
in range(n_parts
)]
1061 self
.reg_partition_points
= ppoints
.like()
1062 self
.outputs
= [Signal(output_width
, name
="io%d" % i
, reset_less
=True)
1064 # intermediates (needed for unit tests)
1065 self
.intermediate_output
= Signal(output_width
)
1067 def eq_from(self
, reg_partition_points
, outputs
, intermediate_output
,
1069 return [self
.reg_partition_points
.eq(reg_partition_points
)] + \
1070 [self
.intermediate_output
.eq(intermediate_output
)] + \
1071 [self
.outputs
[i
].eq(outputs
[i
])
1072 for i
in range(4)] + \
1073 [self
.part_ops
[i
].eq(part_ops
[i
])
1074 for i
in range(len(self
.part_ops
))]
1077 return self
.eq_from(rhs
.reg_partition_points
, rhs
.outputs
,
1078 rhs
.intermediate_output
, rhs
.part_ops
)
1083 def __init__(self
, partition_points
):
1086 self
.epps
= partition_points
.like()
1087 self
.part_ops
= [Signal(2, name
=f
"part_ops_{i}") for i
in range(8)]
1089 def eq_from(self
, epps
, inputs
, part_ops
):
1090 return [self
.epps
.eq(epps
)] + \
1091 [self
.a
.eq(a
), self
.b
.eq(b
)] + \
1092 [self
.part_ops
[i
].eq(part_ops
[i
])
1093 for i
in range(len(self
.part_ops
))]
1096 return self
.eq_from(rhs
.epps
, rhs
.a
, rhs
.b
, rhs
.part_ops
)
1099 class AllTerms(Elaboratable
):
1100 """Set of terms to be added together
1103 def __init__(self
, n_inputs
, output_width
, n_parts
, register_levels
,
1105 """Create an ``AddReduce``.
1107 :param inputs: input ``Signal``s to be summed.
1108 :param output_width: bit-width of ``output``.
1109 :param register_levels: List of nesting levels that should have
1111 :param partition_points: the input partition points.
1113 self
.i
= AllTermsData(partition_points
)
1114 self
.register_levels
= register_levels
1115 self
.n_inputs
= n_inputs
1116 self
.n_parts
= n_parts
1117 self
.output_width
= output_width
1118 self
.o
= AddReduceData(self
.i
.epps
, n_inputs
,
1119 output_width
, n_parts
)
1121 def elaborate(self
, platform
):
1126 # collect part-bytes
1127 pbs
= Signal(8, reset_less
=True)
1130 pb
= Signal(name
="pb%d" % i
, reset_less
=True)
1131 m
.d
.comb
+= pb
.eq(eps
.part_byte(i
))
1133 m
.d
.comb
+= pbs
.eq(Cat(*tl
))
1140 setattr(m
.submodules
, "signs%d" % i
, s
)
1141 m
.d
.comb
+= s
.part_ops
.eq(self
.i
.part_ops
[i
])
1143 n_levels
= len(self
.register_levels
)+1
1144 m
.submodules
.part_8
= part_8
= Part(eps
, 128, 8, n_levels
, 8)
1145 m
.submodules
.part_16
= part_16
= Part(eps
, 128, 4, n_levels
, 8)
1146 m
.submodules
.part_32
= part_32
= Part(eps
, 128, 2, n_levels
, 8)
1147 m
.submodules
.part_64
= part_64
= Part(eps
, 128, 1, n_levels
, 8)
1148 nat_l
, nbt_l
, nla_l
, nlb_l
= [], [], [], []
1149 for mod
in [part_8
, part_16
, part_32
, part_64
]:
1150 m
.d
.comb
+= mod
.a
.eq(self
.i
.a
)
1151 m
.d
.comb
+= mod
.b
.eq(self
.i
.b
)
1152 for i
in range(len(signs
)):
1153 m
.d
.comb
+= mod
.a_signed
[i
].eq(signs
[i
].a_signed
)
1154 m
.d
.comb
+= mod
.b_signed
[i
].eq(signs
[i
].b_signed
)
1155 m
.d
.comb
+= mod
.pbs
.eq(pbs
)
1156 nat_l
.append(mod
.not_a_term
)
1157 nbt_l
.append(mod
.not_b_term
)
1158 nla_l
.append(mod
.neg_lsb_a_term
)
1159 nlb_l
.append(mod
.neg_lsb_b_term
)
1163 for a_index
in range(8):
1164 t
= ProductTerms(8, 128, 8, a_index
, 8)
1165 setattr(m
.submodules
, "terms_%d" % a_index
, t
)
1167 m
.d
.comb
+= t
.a
.eq(self
.i
.a
)
1168 m
.d
.comb
+= t
.b
.eq(self
.i
.b
)
1169 m
.d
.comb
+= t
.pb_en
.eq(pbs
)
1171 for term
in t
.terms
:
1174 # it's fine to bitwise-or data together since they are never enabled
1176 m
.submodules
.nat_or
= nat_or
= OrMod(128)
1177 m
.submodules
.nbt_or
= nbt_or
= OrMod(128)
1178 m
.submodules
.nla_or
= nla_or
= OrMod(128)
1179 m
.submodules
.nlb_or
= nlb_or
= OrMod(128)
1180 for l
, mod
in [(nat_l
, nat_or
),
1184 for i
in range(len(l
)):
1185 m
.d
.comb
+= mod
.orin
[i
].eq(l
[i
])
1186 terms
.append(mod
.orout
)
1188 # copy the intermediate terms to the output
1189 for i
, value
in enumerate(terms
):
1190 m
.d
.comb
+= self
.o
.inputs
[i
].eq(value
)
1192 # copy reg part points and part ops to output
1193 m
.d
.comb
+= self
.o
.reg_partition_points
.eq(eps
)
1194 m
.d
.comb
+= [self
.o
.part_ops
[i
].eq(self
.i
.part_ops
[i
])
1195 for i
in range(len(self
.i
.part_ops
))]
1200 class Intermediates(Elaboratable
):
1201 """ Intermediate output modules
1204 def __init__(self
, output_width
, n_parts
, partition_points
):
1205 self
.i
= FinalReduceData(partition_points
, output_width
, n_parts
)
1206 self
.o
= IntermediateData(partition_points
, output_width
, n_parts
)
1208 def elaborate(self
, platform
):
1211 out_part_ops
= self
.i
.part_ops
1212 out_part_pts
= self
.i
.reg_partition_points
1215 m
.submodules
.io64
= io64
= IntermediateOut(64, 128, 1)
1216 m
.d
.comb
+= io64
.intermed
.eq(self
.i
.output
)
1218 m
.d
.comb
+= io64
.part_ops
[i
].eq(out_part_ops
[i
])
1219 m
.d
.comb
+= self
.o
.outputs
[3].eq(io64
.output
)
1222 m
.submodules
.io32
= io32
= IntermediateOut(32, 128, 2)
1223 m
.d
.comb
+= io32
.intermed
.eq(self
.i
.output
)
1225 m
.d
.comb
+= io32
.part_ops
[i
].eq(out_part_ops
[i
])
1226 m
.d
.comb
+= self
.o
.outputs
[2].eq(io32
.output
)
1229 m
.submodules
.io16
= io16
= IntermediateOut(16, 128, 4)
1230 m
.d
.comb
+= io16
.intermed
.eq(self
.i
.output
)
1232 m
.d
.comb
+= io16
.part_ops
[i
].eq(out_part_ops
[i
])
1233 m
.d
.comb
+= self
.o
.outputs
[1].eq(io16
.output
)
1236 m
.submodules
.io8
= io8
= IntermediateOut(8, 128, 8)
1237 m
.d
.comb
+= io8
.intermed
.eq(self
.i
.output
)
1239 m
.d
.comb
+= io8
.part_ops
[i
].eq(out_part_ops
[i
])
1240 m
.d
.comb
+= self
.o
.outputs
[0].eq(io8
.output
)
1243 m
.d
.comb
+= self
.o
.part_ops
[i
].eq(out_part_ops
[i
])
1244 m
.d
.comb
+= self
.o
.reg_partition_points
.eq(out_part_pts
)
1245 m
.d
.comb
+= self
.o
.intermediate_output
.eq(self
.i
.output
)
1250 class Mul8_16_32_64(Elaboratable
):
1251 """Signed/Unsigned 8/16/32/64-bit partitioned integer multiplier.
1253 Supports partitioning into any combination of 8, 16, 32, and 64-bit
1254 partitions on naturally-aligned boundaries. Supports the operation being
1255 set for each partition independently.
1257 :attribute part_pts: the input partition points. Has a partition point at
1258 multiples of 8 in 0 < i < 64. Each partition point's associated
1259 ``Value`` is a ``Signal``. Modification not supported, except for by
1261 :attribute part_ops: the operation for each byte. The operation for a
1262 particular partition is selected by assigning the selected operation
1263 code to each byte in the partition. The allowed operation codes are:
1265 :attribute OP_MUL_LOW: the LSB half of the product. Equivalent to
1266 RISC-V's `mul` instruction.
1267 :attribute OP_MUL_SIGNED_HIGH: the MSB half of the product where both
1268 ``a`` and ``b`` are signed. Equivalent to RISC-V's `mulh`
1270 :attribute OP_MUL_SIGNED_UNSIGNED_HIGH: the MSB half of the product
1271 where ``a`` is signed and ``b`` is unsigned. Equivalent to RISC-V's
1272 `mulhsu` instruction.
1273 :attribute OP_MUL_UNSIGNED_HIGH: the MSB half of the product where both
1274 ``a`` and ``b`` are unsigned. Equivalent to RISC-V's `mulhu`
1278 def __init__(self
, register_levels
=()):
1279 """ register_levels: specifies the points in the cascade at which
1280 flip-flops are to be inserted.
1284 self
.register_levels
= list(register_levels
)
1287 self
.part_pts
= PartitionPoints()
1288 for i
in range(8, 64, 8):
1289 self
.part_pts
[i
] = Signal(name
=f
"part_pts_{i}")
1290 self
.part_ops
= [Signal(2, name
=f
"part_ops_{i}") for i
in range(8)]
1294 # intermediates (needed for unit tests)
1295 self
.intermediate_output
= Signal(128)
1298 self
.output
= Signal(64)
1300 def elaborate(self
, platform
):
1303 # create (doubled) PartitionPoints (output is double input width)
1304 expanded_part_pts
= eps
= PartitionPoints()
1305 for i
, v
in self
.part_pts
.items():
1306 ep
= Signal(name
=f
"expanded_part_pts_{i*2}", reset_less
=True)
1307 expanded_part_pts
[i
] = ep
1308 m
.d
.comb
+= ep
.eq(v
)
1311 n_parts
= 8 #len(self.part_pts)
1312 t
= AllTerms(n_inputs
, 128, n_parts
, self
.register_levels
,
1314 m
.submodules
.allterms
= t
1315 m
.d
.comb
+= t
.i
.a
.eq(self
.a
)
1316 m
.d
.comb
+= t
.i
.b
.eq(self
.b
)
1317 m
.d
.comb
+= t
.i
.epps
.eq(eps
)
1319 m
.d
.comb
+= t
.i
.part_ops
[i
].eq(self
.part_ops
[i
])
1323 add_reduce
= AddReduce(terms
,
1325 self
.register_levels
,
1326 t
.o
.reg_partition_points
,
1329 out_part_ops
= add_reduce
.o
.part_ops
1330 out_part_pts
= add_reduce
.o
.reg_partition_points
1332 m
.submodules
.add_reduce
= add_reduce
1334 interm
= Intermediates(128, 8, expanded_part_pts
)
1335 m
.submodules
.intermediates
= interm
1336 m
.d
.comb
+= interm
.i
.eq(add_reduce
.o
)
1339 m
.submodules
.finalout
= finalout
= FinalOut(128, 8, expanded_part_pts
)
1340 m
.d
.comb
+= finalout
.i
.eq(interm
.o
)
1341 m
.d
.comb
+= self
.output
.eq(finalout
.out
)
1342 m
.d
.comb
+= self
.intermediate_output
.eq(finalout
.intermediate_output
)
1347 if __name__
== "__main__":
1351 m
.intermediate_output
,
1354 *m
.part_pts
.values()])