2 # SPDX-License-Identifier: LGPL-3-or-later
3 # See Notices.txt for copyright information
6 * https://libre-soc.org/3d_gpu/architecture/dynamic_simd/shape/
7 * https://bugs.libre-soc.org/show_bug.cgi?id=713#c20
8 * https://bugs.libre-soc.org/show_bug.cgi?id=713#c30
9 * https://bugs.libre-soc.org/show_bug.cgi?id=713#c34
10 * https://bugs.libre-soc.org/show_bug.cgi?id=713#c47
11 * https://bugs.libre-soc.org/show_bug.cgi?id=713#c22
12 * https://bugs.libre-soc.org/show_bug.cgi?id=713#c67
15 from nmigen
import Signal
, Module
, Elaboratable
, Mux
, Cat
, Shape
, Repl
16 from nmigen
.back
.pysim
import Simulator
, Delay
, Settle
17 from nmigen
.cli
import rtlil
19 from collections
.abc
import Mapping
20 from functools
import reduce
22 from collections
import defaultdict
23 from pprint
import pprint
25 from ieee754
.part_mul_add
.partpoints
import PartitionPoints
28 # main fn, which started out here in the bugtracker:
29 # https://bugs.libre-soc.org/show_bug.cgi?id=713#c20
30 def layout(elwid
, signed
, vec_el_counts
, lane_shapes
=None, fixed_width
=None):
31 """calculate a SIMD layout.
34 * element: a single scalar value that is an element of a SIMD vector.
35 it has a width in bits, and a signedness. Every element is made of 1 or
37 * ElWid: the element-width (really the element type) of an instruction.
38 Either an integer or a FP type. Integer `ElWid`s are sign-agnostic.
39 In Python, `ElWid` is either an enum type or is `int`.
40 Example `ElWid` definition for integers:
43 I64 = ... # SVP64 value 0b00
44 I32 = ... # SVP64 value 0b01
45 I16 = ... # SVP64 value 0b10
46 I8 = ... # SVP64 value 0b11
48 Example `ElWid` definition for floats:
51 F64 = ... # SVP64 value 0b00
52 F32 = ... # SVP64 value 0b01
53 F16 = ... # SVP64 value 0b10
54 BF16 = ... # SVP64 value 0b11
56 # XXX this is redundant and out-of-date with respect to the
57 # clarification that the input is in counts of *elements*
58 # *NOT* "fixed width parts".
59 # fixed-width parts results in 14 such parts being created
60 # when 5 will do, for a simple example 5-6-6-6
61 * part: A piece of a SIMD vector, every SIMD vector is made of a
62 non-negative integer of parts. Elements are made of a power-of-two
63 number of parts. A part is a fixed number of bits wide for each
64 different SIMD layout, it doesn't vary when `elwid` changes. A part
65 can have a bit width of any non-negative integer, it is not restricted
66 to power-of-two. SIMD vectors should have as few parts as necessary,
67 since some circuits have size proportional to the number of parts.
69 * elwid: ElWid or nmigen Value with ElWid as the shape
70 the current element-width
72 the signedness of all elements in a SIMD layout
73 * vec_el_counts: dict[ElWid, int]
74 a map from `ElWid` values `k` to the number of vector elements
75 required within a partition when `elwid == k`.
78 vec_el_counts = {ElWid.I8(==0b11): 8, # 8 vector elements
79 ElWid.I16(==0b10): 4, # 4 vector elements
80 ElWid.I32(==0b01): 2, # 2 vector elements
81 ElWid.I64(==0b00): 1} # 1 vector (aka scalar) element
85 vec_el_counts = {ElWid.BF16(==0b11): 4,
90 * lane_shapes: int or Mapping[ElWid, int] (optional)
91 the bit-width of all elements in a SIMD layout.
93 * fixed_width: int (optional)
94 the total width of a SIMD vector. One or both of lane_shapes or
95 fixed_width may be provided. Both may not be left out.
97 # when there are no lane_shapes specified, this indicates a
98 # desire to use the maximum available space based on the fixed width
99 # https://bugs.libre-soc.org/show_bug.cgi?id=713#c67
100 if lane_shapes
is None:
101 assert fixed_width
is not None, \
102 "both fixed_width and lane_shapes cannot be None"
103 lane_shapes
= {i
: fixed_width
// vec_el_counts
[i
]
104 for i
in vec_el_counts
}
105 print("lane_shapes", fixed_width
, lane_shapes
)
106 # identify if the lane_shapes is a mapping (dict, etc.)
107 # if not, then assume that it is an integer (width) that
108 # needs to be requested across all partitions
109 if not isinstance(lane_shapes
, Mapping
):
110 lane_shapes
= {i
: lane_shapes
for i
in vec_el_counts
}
111 # compute a set of partition widths
112 print("lane_shapes", lane_shapes
, "vec_el_counts", vec_el_counts
)
113 cpart_wid
= max(lane_shapes
.values())
114 part_count
= max(vec_el_counts
.values())
115 # calculate the minumum width required
116 width
= cpart_wid
* part_count
117 print("width", width
, cpart_wid
, part_count
)
118 if fixed_width
is not None: # override the width and part_wid
119 assert width
< fixed_width
, "not enough space to fit partitions"
120 part_wid
= fixed_width
// part_count
121 assert part_wid
* part_count
== fixed_width
, \
122 "calculated width not aligned multiples"
124 print("part_wid", part_wid
, "count", part_count
)
126 # go with computed width
128 # create the breakpoints dictionary.
129 # do multi-stage version https://bugs.libre-soc.org/show_bug.cgi?id=713#c34
130 # https://stackoverflow.com/questions/26367812/
131 dpoints
= defaultdict(list) # if empty key, create a (empty) list
132 for i
, c
in vec_el_counts
.items():
134 dpoints
[p
].append(i
) # auto-creates list if key non-existent
135 for start
in range(0, part_count
, c
):
136 add_p(start
* part_wid
) # start of lane
137 add_p(start
* part_wid
+ lane_shapes
[i
]) # start of padding
138 # do not need the breakpoints at the very start or the very end
140 dpoints
.pop(width
, None)
141 plist
= list(dpoints
.keys())
144 pprint(dict(dpoints
))
145 # second stage, add (map to) the elwidth==i expressions.
146 # TODO: use nmutil.treereduce?
149 points
[p
] = map(lambda i
: elwid
== i
, dpoints
[p
])
150 points
[p
] = reduce(operator
.or_
, points
[p
])
151 # third stage, create the binary values which *if* elwidth is set to i
152 # *would* result in the mask at that elwidth being set to this value
153 # these can easily be double-checked through Assertion
155 for i
in vec_el_counts
.keys():
157 for p
, elwidths
in dpoints
.items():
159 bitpos
= plist
.index(p
)
160 bitp
[i
] |
= 1 << bitpos
161 # fourth stage: determine which partitions are 100% unused.
162 # these can then be "blanked out"
163 bmask
= (1 << len(plist
))-1
164 for p
in bitp
.values():
166 return (PartitionPoints(points
), bitp
, bmask
, width
, lane_shapes
,
167 part_wid
, part_count
)
170 if __name__
== '__main__':
172 # for each element-width (elwidth 0-3) the number of Vector Elements is:
173 # elwidth=0b00 QTY 1 partitions: | ? |
174 # elwidth=0b01 QTY 1 partitions: | ? |
175 # elwidth=0b10 QTY 2 partitions: | ? | ? |
176 # elwidth=0b11 QTY 4 partitions: | ? | ? | ? | ? |
177 # actual widths of Signals *within* those partitions is given separately
185 # width=3 indicates "same width Vector Elements (3) at all elwidths"
186 # elwidth=0b00 1x 5-bit | unused xx ..3 |
187 # elwidth=0b01 1x 6-bit | unused xx ..3 |
188 # elwidth=0b10 2x 12-bit | xxx ..3 | xxx ..3 |
189 # elwidth=0b11 3x 24-bit | ..3| ..3 | ..3 |..3 |
190 # expected partitions (^) | | | (^)
191 # to be at these points: (|) | | | |
192 width_in_all_parts
= 3
195 pprint((i
, layout(i
, True, vec_el_counts
, width_in_all_parts
)))
197 # fixed_width=32 and no lane_widths says "allocate maximum"
198 # i.e. Vector Element Widths are auto-allocated
199 # elwidth=0b00 1x 32-bit | .................32 |
200 # elwidth=0b01 1x 32-bit | .................32 |
201 # elwidth=0b10 2x 12-bit | ......16 | ......16 |
202 # elwidth=0b11 3x 24-bit | ..8| ..8 | ..8 |..8 |
203 # expected partitions (^) | | | (^)
204 # to be at these points: (|) | | | |
206 # TODO, fix this so that it is correct
207 #print ("maximum allocation from fixed_width=32")
209 # pprint((i, layout(i, True, vec_el_counts, fixed_width=32)))
211 # specify that the Vector Element lengths are to be *different* at
212 # each of the elwidths.
213 # combined with vec_el_counts we have:
214 # elwidth=0b00 1x 5-bit | <-- unused -->....5 |
215 # elwidth=0b01 1x 6-bit | <-- unused -->.....6 |
216 # elwidth=0b10 2x 12-bit | unused .....6 | unused .....6 |
217 # elwidth=0b11 3x 24-bit | .....6 | .....6 | .....6 | .....6 |
218 # expected partitions (^) ^ ^ ^^ (^)
219 # to be at these points: (|) | | || (|)
220 widths_at_elwidth
= {
227 print ("5,6,6,6 elements", widths_at_elwidth
)
229 pprint((i
, layout(i
, False, vec_el_counts
, widths_at_elwidth
)))
231 # this tests elwidth as an actual Signal. layout is allowed to
232 # determine arbitrarily the overall length
233 # https://bugs.libre-soc.org/show_bug.cgi?id=713#c30
236 pp
, bitp
, bm
, b
, c
, d
, e
= layout(
237 elwid
, False, vec_el_counts
, widths_at_elwidth
)
238 pprint((pp
, b
, c
, d
, e
))
239 for k
, v
in bitp
.items():
240 print("bitp elwidth=%d" % k
, bin(v
))
241 print("bmask", bin(bm
))
250 for pval
in list(pp
.values()):
251 val
= yield pval
# get nmigen to evaluate pp
253 pprint((i
, (ppt
, b
, c
, d
, e
)))
254 # check the results against bitp static-expected partition points
255 # https://bugs.libre-soc.org/show_bug.cgi?id=713#c47
256 # https://stackoverflow.com/a/27165694
257 ival
= int(''.join(map(str, ppt
[::-1])), 2)
258 assert ival
== bitp
[i
]
261 sim
.add_process(process
)
264 # this tests elwidth as an actual Signal. layout is *not* allowed to
265 # determine arbitrarily the overall length, it is fixed to 64
266 # https://bugs.libre-soc.org/show_bug.cgi?id=713#c22
269 pp
, bitp
, bm
, b
, c
, d
, e
= layout(elwid
, False, vec_el_counts
,
272 pprint((pp
, b
, c
, d
, e
))
273 for k
, v
in bitp
.items():
274 print("bitp elwidth=%d" % k
, bin(v
))
275 print("bmask", bin(bm
))
284 for pval
in list(pp
.values()):
285 val
= yield pval
# get nmigen to evaluate pp
287 print("test elwidth=%d" % i
)
288 pprint((i
, (ppt
, b
, c
, d
, e
)))
289 # check the results against bitp static-expected partition points
290 # https://bugs.libre-soc.org/show_bug.cgi?id=713#c47
291 # https://stackoverflow.com/a/27165694
292 ival
= int(''.join(map(str, ppt
[::-1])), 2)
293 assert ival
== bitp
[i
], "ival %s actual %s" % (bin(ival
),
297 sim
.add_process(process
)