6f3f6eee1a901fe5a214322ac2b9549f973267ed
2 # Copyright (C) 2020 Collabora, Ltd.
3 # Copyright (C) 2018 Alyssa Rosenzweig
4 # Copyright (C) 2016 Intel Corporation
6 # Permission is hereby granted, free of charge, to any person obtaining a
7 # copy of this software and associated documentation files (the "Software"),
8 # to deal in the Software without restriction, including without limitation
9 # the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 # and/or sell copies of the Software, and to permit persons to whom the
11 # Software is furnished to do so, subject to the following conditions:
13 # The above copyright notice and this permission notice (including the next
14 # paragraph) shall be included in all copies or substantial portions of the
17 # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 # THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
22 # FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
34 # ineg must be lowered late, but only for integers; floats will try to
35 # have modifiers attached... hence why this has to be here rather than
36 # a more standard lower_negate approach
38 (('ineg', a
), ('isub', 0, a
)),
41 for isz
in ('8', '16', '32'):
42 for osz
in ('16', '32', '64'):
43 algebraic_late
+= [(('b2f' + osz
, 'a@' + isz
), ('b' + isz
+ 'csel', a
, 1.0, 0.0))]
45 # Midgard is able to type convert down by only one "step" per instruction; if
46 # NIR wants more than one step, we need to break up into multiple instructions
50 for op
in ('u2u', 'i2i', 'f2f', 'i2f', 'u2f', 'f2i', 'f2u'):
53 # 8 bit float doesn't exist
54 srcsz_min
= 8 if op
[0] != 'f' else 16
55 dstsz_min
= 8 if op
[2] != 'f' else 16
57 # Iterate over all possible destination and source sizes
58 while dstsz
<= dstsz_max
:
60 while srcsz
<= srcsz_max
:
61 # Size converter lowering is only needed if src and dst sizes are
62 # spaced by a factor > 2.
63 # Type converter lowering is needed as soon as src_size != dst_size
64 if srcsz
!= dstsz
and ((srcsz
* 2 != dstsz
and srcsz
!= dstsz
* 2) or op
[0] != op
[2]):
67 # When converting down we first do the type conversion followed
68 # by one or more size conversions. When converting up, we do
69 # the type conversion at the end. This way we don't have to
70 # deal with the fact that f2f8 doesn't exists.
71 sizeconvop
= op
[0] + '2' + op
[0] if srcsz
< dstsz
else op
[2] + '2' + op
[2]
72 if srcsz
> dstsz
and op
[0] != op
[2]:
73 rule
= (op
+ str(int(cursz
)), rule
)
75 cursz
= cursz
/ 2 if dstsz
< srcsz
else cursz
* 2
76 rule
= (sizeconvop
+ str(int(cursz
)), rule
)
77 if srcsz
< dstsz
and op
[0] != op
[2]:
78 rule
= (op
+ str(int(cursz
)), rule
)
79 converts
+= [((op
+ str(int(dstsz
)), 'a@' + str(int(srcsz
))), rule
)]
83 # Bifrost doesn't have fp16 for a lot of special ops
84 SPECIAL
= ['fexp2', 'flog2', 'fsin', 'fcos']
87 converts
+= [((op
+ '@16', a
), ('f2f16', (op
, ('f2f32', a
))))]
89 converts
+= [(('f2b32', a
), ('fne32', a
, 0.0)),
90 (('i2b32', a
), ('ine32', a
, 0)),
91 (('b2i32', a
), ('iand', 'a@32', 1))]
94 parser
= argparse
.ArgumentParser()
95 parser
.add_argument('-p', '--import-path', required
=True)
96 args
= parser
.parse_args()
97 sys
.path
.insert(0, args
.import_path
)
102 import nir_algebraic
# pylint: disable=import-error
104 print('#include "bifrost_nir.h"')
106 print(nir_algebraic
.AlgebraicPass("bifrost_nir_lower_algebraic_late",
107 algebraic_late
+ converts
).render())
109 if __name__
== '__main__':