1 # SPDX-License-Identifier: LGPL-2.1-or-later
2 # See Notices.txt for copyright information
5 Copyright (C) 2021 Luke Kenneth Casson Leighton <lkcl@lkcl.net>
7 dynamically-partitionable "repl" class, directly equivalent
12 * http://libre-riscv.org/3d_gpu/architecture/dynamic_simd/repl
13 * http://bugs.libre-riscv.org/show_bug.cgi?id=709
17 from nmigen
import Signal
, Module
, Elaboratable
, Cat
, Repl
18 from nmigen
.back
.pysim
import Simulator
, Settle
19 from nmigen
.cli
import rtlil
21 from ieee754
.part_mul_add
.partpoints
import PartitionPoints
22 from ieee754
.part
.partsig
import PartitionedSignal
25 def get_runlengths(pbit
, size
):
28 # identify where the 1s are, which indicates "start of a new partition"
29 # we want a list of the lengths of all partitions
31 if pbit
& (1<<i
): # it's a 1: ends old partition, starts new
32 res
.append(count
) # add partition
33 count
= 1 # start again
36 # end reached, add whatever is left. could have done this by creating
37 # "fake" extra bit on the partitions, but hey
40 print ("get_runlengths", bin(pbit
), size
, res
)
45 class PartitionedRepl(Elaboratable
):
46 def __init__(self
, repl
, qty
, mask
):
47 """Create a ``PartitionedRepl`` operator
49 # work out the length (total of all PartitionedSignals)
52 width
, signed
= repl
.shape()
53 if isinstance(mask
, dict):
54 mask
= list(mask
.values())
56 self
.shape
= (width
* qty
), signed
57 self
.output
= PartitionedSignal(mask
, self
.shape
, reset_less
=True)
58 self
.partition_points
= self
.output
.partpoints
59 self
.mwidth
= len(self
.partition_points
)+1
61 def get_chunk(self
, y
, numparts
):
63 if not isinstance(x
, PartitionedSignal
):
64 # assume Scalar. totally different rules
65 end
= numparts
* (len(x
) // self
.mwidth
)
67 # PartitionedSignal: start at partition point
68 keys
= [0] + list(x
.partpoints
.keys()) + [len(x
)]
69 # get current index and increment it (for next Repl chunk)
72 print ("getting", upto
, numparts
, keys
, len(x
))
73 # get the partition point as far as we are up to
75 end
= keys
[upto
+numparts
]
76 print ("start end", start
, end
, len(x
))
79 def elaborate(self
, platform
):
83 keys
= list(self
.partition_points
.keys())
84 print ("keys", keys
, "values", self
.partition_points
.values())
85 print ("mask", self
.mask
)
86 outpartsize
= len(self
.output
) // self
.mwidth
87 width
, signed
= self
.output
.shape()
88 print ("width, signed", width
, signed
)
90 with m
.Switch(Cat(self
.mask
)):
91 # for each partition possibility, create a Repl sequence
92 for pbit
in range(1<<len(keys
)):
93 # set up some indices pointing to where things have got
94 # then when called below in the inner nested loop they give
95 # the relevant sequential chunk
98 # get a list of the length of each partition run
99 runlengths
= get_runlengths(pbit
, len(keys
))
100 print ("pbit", bin(pbit
), "runs", runlengths
)
101 for i
in runlengths
: # for each partition
102 thing
= self
.get_chunk(y
, i
) # get sequential chunk
103 output
.append(Repl(thing
, self
.qty
)) # and replicate it
105 # direct access to the underlying Signal
106 comb
+= self
.output
.sig
.eq(Cat(*output
)) # cat all chunks
111 if isinstance(self
.repl
, PartitionedSignal
):
112 return [self
.repl
.lower(), self
.output
.lower()]
113 return [self
.repl
, self
.output
.lower()]
116 if __name__
== "__main__":
117 from ieee754
.part
.test
.test_partsig
import create_simulator
120 a
= PartitionedSignal(mask
, 32)
121 m
.submodules
.repl
= repl
= PartitionedRepl(a
, 2, mask
)
122 omask
= (1<<len(repl
.output
))-1
124 traces
= repl
.ports()
125 vl
= rtlil
.convert(repl
, ports
=traces
)
126 with
open("part_repl.il", "w") as f
:
129 sim
= create_simulator(m
, traces
, "partrepl")
133 yield a
.sig
.eq(0xa12345c7)
135 out
= yield repl
.output
.sig
136 print("out 000", bin(out
&omask
), hex(out
&omask
))
139 out
= yield repl
.output
.sig
140 print("out 010", bin(out
&omask
), hex(out
&omask
))
143 out
= yield repl
.output
.sig
144 print("out 110", bin(out
&omask
), hex(out
&omask
))
147 out
= yield repl
.output
.sig
148 print("out 111", bin(out
&omask
), hex(out
&omask
))
150 sim
.add_process(process
)
151 with sim
.write_vcd("partition_repl.vcd", "partition_repl.gtkw",
159 m
.submodules
.ass
= ass
= PartitionedRepl(a
, 2, mask
)
160 omask
= (1<<len(ass
.output
))-1
163 sim
= create_simulator(m
, traces
, "partass")
167 yield a
.eq(0xa12345c7)
169 out
= yield ass
.output
.sig
170 print("out 000", bin(out
&omask
), hex(out
&omask
))
173 out
= yield ass
.output
.sig
174 print("out 010", bin(out
&omask
), hex(out
&omask
))
177 out
= yield ass
.output
.sig
178 print("out 110", bin(out
&omask
), hex(out
&omask
))
181 out
= yield ass
.output
.sig
182 print("out 111", bin(out
&omask
), hex(out
&omask
))
184 sim
.add_process(process
)
185 with sim
.write_vcd("partition_repl_scalar.vcd",
186 "partition_repl_scalar.gtkw",