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 (('pack_unorm_4x8', a
), ('pack_32_4x8', ('f2u8', ('fround_even', ('fmul', ('fsat', a
), 255.0))))),
36 # Allows us to schedule as a multiply by 2
37 (('~fadd', ('fadd', a
, b
), a
), ('fadd', ('fadd', a
, a
), b
)),
41 # ineg must be lowered late, but only for integers; floats will try to
42 # have modifiers attached... hence why this has to be here rather than
43 # a more standard lower_negate approach
45 (('ineg', a
), ('isub', 0, a
)),
47 # Likewise we want fsub lowered but not isub
48 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
))),
50 # These two special-cases save space/an op than the actual csel op +
51 # scheduler flexibility
53 (('b32csel', a
, 'b@32', 0), ('iand', a
, b
)),
54 (('b32csel', a
, 0, 'b@32'), ('iand', ('inot', a
), b
)),
56 # Fuse sat_signed. This should probably be shared with Bifrost
57 (('~fmin', ('fmax', a
, -1.0), 1.0), ('fsat_signed', a
)),
58 (('~fmax', ('fmin', a
, 1.0), -1.0), ('fsat_signed', a
)),
60 # Fuse clamp_positive. This should probably be shared with Utgard/bifrost
61 (('fmax', a
, 0.0), ('fclamp_pos', a
)),
63 (('ishl', 'a@16', b
), ('u2u16', ('ishl', ('u2u32', a
), b
))),
64 (('ishr', 'a@16', b
), ('i2i16', ('ishr', ('i2i32', a
), b
))),
65 (('ushr', 'a@16', b
), ('u2u16', ('ushr', ('u2u32', a
), b
))),
67 (('ishl', 'a@8', b
), ('u2u8', ('u2u16', ('ishl', ('u2u32', ('u2u16', a
)), b
)))),
68 (('ishr', 'a@8', b
), ('i2i8', ('i2i16', ('ishr', ('i2i32', ('i2i16', a
)), b
)))),
69 (('ushr', 'a@8', b
), ('u2u8', ('u2u16', ('ushr', ('u2u32', ('u2u16', a
)), b
)))),
71 # Canonical form. The scheduler will convert back if it makes sense.
72 (('fmul', a
, 2.0), ('fadd', a
, a
))
76 # Midgard is able to type convert down by only one "step" per instruction; if
77 # NIR wants more than one step, we need to break up into multiple instructions.
78 # Nevertheless, we can do both a size step and a floating/int step at once.
82 for op
in ('u2u', 'i2i', 'f2f', 'i2f', 'u2f', 'f2i', 'f2u'):
85 # 8 bit float doesn't exist
86 srcsz_min
= 8 if op
[0] != 'f' else 16
87 dstsz_min
= 8 if op
[2] != 'f' else 16
89 # Iterate over all possible destination and source sizes
90 while dstsz
<= dstsz_max
:
92 while srcsz
<= srcsz_max
:
93 # Size converter lowering is only needed if src and dst sizes are
94 # spaced by a factor > 2.
95 if srcsz
!= dstsz
and (srcsz
* 2 != dstsz
and srcsz
!= dstsz
* 2):
98 # When converting down we first do the type conversion followed
99 # by one or more size conversions. When converting up, we do
100 # the type conversion at the end. This way we don't have to
101 # deal with the fact that f2f8 doesn't exists.
102 sizeconvop
= op
[0] + '2' + op
[0] if srcsz
< dstsz
else op
[2] + '2' + op
[2]
103 if srcsz
> dstsz
and op
[0] != op
[2]:
104 rule
= (op
+ str(int(cursz
)), rule
)
105 while cursz
!= dstsz
:
106 cursz
= cursz
/ 2 if dstsz
< srcsz
else cursz
* 2
107 rule
= (sizeconvop
+ str(int(cursz
)), rule
)
108 if srcsz
< dstsz
and op
[0] != op
[2]:
109 rule
= (op
+ str(int(cursz
)), rule
)
110 converts
+= [((op
+ str(int(dstsz
)), 'a@' + str(int(srcsz
))), rule
)]
114 # Try to force constants to the right
116 # fge gets flipped to fle, so we invert to keep the order
117 (('fge', 'a', '#b'), (('inot', ('flt', a
, b
)))),
118 (('fge32', 'a', '#b'), (('inot', ('flt32', a
, b
)))),
119 (('ige32', 'a', '#b'), (('inot', ('ilt32', a
, b
)))),
120 (('uge32', 'a', '#b'), (('inot', ('ult32', a
, b
)))),
122 # fge gets mapped to fle with a flip
123 (('flt32', '#a', 'b'), ('inot', ('fge32', a
, b
))),
124 (('ilt32', '#a', 'b'), ('inot', ('ige32', a
, b
))),
125 (('ult32', '#a', 'b'), ('inot', ('uge32', a
, b
)))
128 # ..since the above switching happens after algebraic stuff is done
130 (('inot', ('inot', a
)), a
)
133 # Midgard scales fsin/fcos arguments by pi.
134 # Pass must be run only once, after the main loop
137 (('fsin', a
), ('fsin', ('fdiv', a
, math
.pi
))),
138 (('fcos', a
), ('fcos', ('fdiv', a
, math
.pi
))),
142 parser
= argparse
.ArgumentParser()
143 parser
.add_argument('-p', '--import-path', required
=True)
144 args
= parser
.parse_args()
145 sys
.path
.insert(0, args
.import_path
)
150 import nir_algebraic
# pylint: disable=import-error
152 print('#include "midgard_nir.h"')
154 print(nir_algebraic
.AlgebraicPass("midgard_nir_lower_algebraic_early",
157 print(nir_algebraic
.AlgebraicPass("midgard_nir_lower_algebraic_late",
158 algebraic_late
+ converts
+ constant_switch
).render())
160 print(nir_algebraic
.AlgebraicPass("midgard_nir_scale_trig",
161 scale_trig
).render())
163 print(nir_algebraic
.AlgebraicPass("midgard_nir_cancel_inot",
164 cancel_inot
).render())
167 if __name__
== '__main__':