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 # Likewise we want fsub lowered but not isub
41 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
))),
43 # These two special-cases save space/an op than the actual csel op +
44 # scheduler flexibility
46 (('b32csel', a
, 'b@32', 0), ('iand', a
, b
)),
47 (('b32csel', a
, 0, 'b@32'), ('iand', ('inot', a
), b
)),
49 # Fuse sat_signed. This should probably be shared with Bifrost
50 (('~fmin', ('fmax', a
, -1.0), 1.0), ('fsat_signed', a
)),
51 (('~fmax', ('fmin', a
, 1.0), -1.0), ('fsat_signed', a
)),
53 # Fuse clamp_positive. This should probably be shared with Utgard/bifrost
54 (('fmax', a
, 0.0), ('fclamp_pos', a
)),
58 # Midgard is able to type convert down by only one "step" per instruction; if
59 # NIR wants more than one step, we need to break up into multiple instructions
63 for op
in ('u2u', 'i2i', 'f2f', 'i2f', 'u2f', 'f2i', 'f2u'):
66 # 8 bit float doesn't exist
67 srcsz_min
= 8 if op
[0] != 'f' else 16
68 dstsz_min
= 8 if op
[2] != 'f' else 16
70 # Iterate over all possible destination and source sizes
71 while dstsz
<= dstsz_max
:
73 while srcsz
<= srcsz_max
:
74 # Size converter lowering is only needed if src and dst sizes are
75 # spaced by a factor > 2.
76 # Type converter lowering is needed as soon as src_size != dst_size
77 if srcsz
!= dstsz
and ((srcsz
* 2 != dstsz
and srcsz
!= dstsz
* 2) or op
[0] != op
[2]):
80 # When converting down we first do the type conversion followed
81 # by one or more size conversions. When converting up, we do
82 # the type conversion at the end. This way we don't have to
83 # deal with the fact that f2f8 doesn't exists.
84 sizeconvop
= op
[0] + '2' + op
[0] if srcsz
< dstsz
else op
[2] + '2' + op
[2]
85 if srcsz
> dstsz
and op
[0] != op
[2]:
86 rule
= (op
+ str(int(cursz
)), rule
)
88 cursz
= cursz
/ 2 if dstsz
< srcsz
else cursz
* 2
89 rule
= (sizeconvop
+ str(int(cursz
)), rule
)
90 if srcsz
< dstsz
and op
[0] != op
[2]:
91 rule
= (op
+ str(int(cursz
)), rule
)
92 converts
+= [((op
+ str(int(dstsz
)), 'a@' + str(int(srcsz
))), rule
)]
96 # Midgard outputs fp32 for specials. The f2f32 will be folded in later.
97 SPECIAL
= ['fexp2', 'flog2', 'fsin', 'fcos', 'frcp', 'frsq']
100 converts
+= [((op
+ '@16', a
), ('f2f16', (op
, ('f2f32', a
))))]
102 # Try to force constants to the right
104 # fge gets flipped to fle, so we invert to keep the order
105 (('fge', 'a', '#b'), (('inot', ('flt', a
, b
)))),
106 (('fge32', 'a', '#b'), (('inot', ('flt32', a
, b
)))),
107 (('ige32', 'a', '#b'), (('inot', ('ilt32', a
, b
)))),
108 (('uge32', 'a', '#b'), (('inot', ('ult32', a
, b
)))),
110 # fge gets mapped to fle with a flip
111 (('flt32', '#a', 'b'), ('inot', ('fge32', a
, b
))),
112 (('ilt32', '#a', 'b'), ('inot', ('ige32', a
, b
))),
113 (('ult32', '#a', 'b'), ('inot', ('uge32', a
, b
)))
116 # ..since the above switching happens after algebraic stuff is done
118 (('inot', ('inot', a
)), a
)
121 # Midgard scales fsin/fcos arguments by pi.
122 # Pass must be run only once, after the main loop
125 (('fsin', a
), ('fsin', ('fdiv', a
, math
.pi
))),
126 (('fcos', a
), ('fcos', ('fdiv', a
, math
.pi
))),
130 parser
= argparse
.ArgumentParser()
131 parser
.add_argument('-p', '--import-path', required
=True)
132 args
= parser
.parse_args()
133 sys
.path
.insert(0, args
.import_path
)
138 import nir_algebraic
# pylint: disable=import-error
140 print('#include "midgard_nir.h"')
142 print(nir_algebraic
.AlgebraicPass("midgard_nir_lower_algebraic_late",
143 algebraic_late
+ converts
+ constant_switch
).render())
145 print(nir_algebraic
.AlgebraicPass("midgard_nir_scale_trig",
146 scale_trig
).render())
148 print(nir_algebraic
.AlgebraicPass("midgard_nir_cancel_inot",
149 cancel_inot
).render())
152 if __name__
== '__main__':