2 # 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
)),
40 # These two special-cases save space/an op than the actual csel op +
41 # scheduler flexibility
43 (('b32csel', a
, 'b@32', 0), ('iand', a
, b
)),
44 (('b32csel', a
, 0, 'b@32'), ('iand', ('inot', a
), b
)),
48 # Midgard is able to type convert down by only one "step" per instruction; if
49 # NIR wants more than one step, we need to break up into multiple instructions
53 for op
in ('u2u', 'i2i', 'f2f', 'i2f', 'u2f', 'f2i', 'f2u'):
56 # 8 bit float doesn't exist
57 srcsz_min
= 8 if op
[0] != 'f' else 16
58 dstsz_min
= 8 if op
[2] != 'f' else 16
60 # Iterate over all possible destination and source sizes
61 while dstsz
<= dstsz_max
:
63 while srcsz
<= srcsz_max
:
64 # Size converter lowering is only needed if src and dst sizes are
65 # spaced by a factor > 2.
66 # Type converter lowering is needed as soon as src_size != dst_size
67 if srcsz
!= dstsz
and ((srcsz
* 2 != dstsz
and srcsz
!= dstsz
* 2) or op
[0] != op
[2]):
70 # When converting down we first do the type conversion followed
71 # by one or more size conversions. When converting up, we do
72 # the type conversion at the end. This way we don't have to
73 # deal with the fact that f2f8 doesn't exists.
74 sizeconvop
= op
[0] + '2' + op
[0] if srcsz
< dstsz
else op
[2] + '2' + op
[2]
75 if srcsz
> dstsz
and op
[0] != op
[2]:
76 rule
= (op
+ str(int(cursz
)), rule
)
78 cursz
= cursz
/ 2 if dstsz
< srcsz
else cursz
* 2
79 rule
= (sizeconvop
+ str(int(cursz
)), rule
)
80 if srcsz
< dstsz
and op
[0] != op
[2]:
81 rule
= (op
+ str(int(cursz
)), rule
)
82 converts
+= [((op
+ str(int(dstsz
)), 'a@' + str(int(srcsz
))), rule
)]
86 # Midgard scales fsin/fcos arguments by pi.
87 # Pass must be run only once, after the main loop
90 (('fsin', a
), ('fsin', ('fdiv', a
, math
.pi
))),
91 (('fcos', a
), ('fcos', ('fdiv', a
, math
.pi
))),
95 parser
= argparse
.ArgumentParser()
96 parser
.add_argument('-p', '--import-path', required
=True)
97 args
= parser
.parse_args()
98 sys
.path
.insert(0, args
.import_path
)
103 import nir_algebraic
# pylint: disable=import-error
105 print('#include "midgard_nir.h"')
107 print(nir_algebraic
.AlgebraicPass("midgard_nir_lower_algebraic_late",
108 algebraic_late
+ converts
).render())
110 print(nir_algebraic
.AlgebraicPass("midgard_nir_scale_trig",
111 scale_trig
).render())
114 if __name__
== '__main__':