xref: /aosp_15_r20/external/mesa3d/src/panfrost/compiler/bifrost_nir_algebraic.py (revision 6104692788411f58d303aa86923a9ff6ecaded22)
1# Copyright (C) 2021 Collabora, Ltd.
2# Copyright (C) 2016 Intel Corporation
3#
4# Permission is hereby granted, free of charge, to any person obtaining a
5# copy of this software and associated documentation files (the "Software"),
6# to deal in the Software without restriction, including without limitation
7# the rights to use, copy, modify, merge, publish, distribute, sublicense,
8# and/or sell copies of the Software, and to permit persons to whom the
9# Software is furnished to do so, subject to the following conditions:
10#
11# The above copyright notice and this permission notice (including the next
12# paragraph) shall be included in all copies or substantial portions of the
13# Software.
14#
15# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21# IN THE SOFTWARE.
22
23import argparse
24import sys
25import math
26
27a = 'a'
28b = 'b'
29c = 'c'
30d = 'd'
31
32# In general, bcsel is cheaper than bitwise arithmetic on Mali. On
33# Bifrost, we can implement bcsel as either CSEL or MUX to schedule to either
34# execution unit. On Valhall, bitwise arithmetic may be on the SFU whereas MUX
35# is on the higher throughput CVT unit. We get a zero argument for free relative
36# to the bitwise op, which would be LSHIFT_* internally taking a zero anyway.
37#
38# As such, it's beneficial to reexpress bitwise arithmetic of booleans as bcsel.
39opt_bool_bitwise = [
40    (('iand', 'a@1', 'b@1'), ('bcsel', a, b, False)),
41    (('ior', 'a@1', 'b@1'), ('bcsel', a, a, b)),
42    (('iand', 'a@1', ('inot', 'b@1')), ('bcsel', b, 0, a)),
43    (('ior', 'a@1', ('inot', 'b@1')), ('bcsel', b, a, True)),
44]
45
46algebraic_late = [
47    (('pack_32_4x8_split', a, b, c, d),
48     ('pack_32_2x16_split', ('ior', ('u2u16', a), ('ishl', ('u2u16', b), 8)),
49                            ('ior', ('u2u16', c), ('ishl', ('u2u16', d), 8)))),
50
51    # Canonical form. The scheduler will convert back if it makes sense.
52    (('fmul', a, 2.0), ('fadd', a, a)),
53
54    # Fuse Mali-specific clamps
55    (('fmin', ('fmax', a, -1.0), 1.0), ('fsat_signed_mali', a)),
56    (('fmax', ('fmin', a, 1.0), -1.0), ('fsat_signed_mali', a)),
57    (('fmax', a, 0.0), ('fclamp_pos_mali', a)),
58
59    (('b32csel', 'b@32', ('iadd', 'a@32', 1), a), ('iadd', a, ('b2i32', b))),
60
61    # We don't have an 8-bit CSEL, so this is the best we can do.
62    # Note that we use 8-bit booleans internally to preserve vectorization.
63    (('imin', 'a@8', 'b@8'), ('b8csel', ('ilt8', a, b), a, b)),
64    (('imax', 'a@8', 'b@8'), ('b8csel', ('ilt8', a, b), b, a)),
65    (('umin', 'a@8', 'b@8'), ('b8csel', ('ult8', a, b), a, b)),
66    (('umax', 'a@8', 'b@8'), ('b8csel', ('ult8', a, b), b, a)),
67
68    # Floats are at minimum 16-bit, which means when converting to an 8-bit
69    # integer, the vectorization changes. So there's no one-shot hardware
70    # instruction for f2i8. Instead, lower to two NIR instructions that map
71    # directly to the hardware.
72    (('f2i8', a), ('i2i8', ('f2i16', a))),
73    (('f2u8', a), ('u2u8', ('f2u16', a))),
74
75    # XXX: Duplicate of nir_lower_pack
76    (('unpack_64_2x32', a), ('vec2', ('unpack_64_2x32_split_x', a),
77                                     ('unpack_64_2x32_split_y', a))),
78]
79
80# Handling all combinations of boolean and float sizes for b2f is nontrivial.
81# bcsel has the same problem in more generality; lower b2f to bcsel in NIR to
82# reuse the efficient implementations of bcsel. This includes special handling
83# to allow vectorization in places the hardware does not directly.
84#
85# Because this lowering must happen late, NIR won't squash inot in
86# automatically. Do so explicitly. (The more specific pattern must be first.)
87for bsz in [8, 16, 32]:
88    for fsz in [16, 32]:
89        algebraic_late += [
90                ((f'b2f{fsz}', ('inot', f'a@{bsz}')), (f'b{bsz}csel', a, 0.0, 1.0)),
91                ((f'b2f{fsz}', f'a@{bsz}'), (f'b{bsz}csel', a, 1.0, 0.0)),
92        ]
93
94
95def main():
96    parser = argparse.ArgumentParser()
97    parser.add_argument('-p', '--import-path', required=True)
98    args = parser.parse_args()
99    sys.path.insert(0, args.import_path)
100    run()
101
102
103def run():
104    import nir_algebraic  # pylint: disable=import-error
105
106    print('#include "bifrost_nir.h"')
107
108    print(nir_algebraic.AlgebraicPass("bifrost_nir_opt_boolean_bitwise",
109                                      opt_bool_bitwise).render())
110    print(nir_algebraic.AlgebraicPass("bifrost_nir_lower_algebraic_late",
111                                      algebraic_late).render())
112
113
114if __name__ == '__main__':
115    main()
116