• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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'
30
31# In general, bcsel is cheaper than bitwise arithmetic on Mali. On
32# Bifrost, we can implement bcsel as either CSEL or MUX to schedule to either
33# execution unit. On Valhall, bitwise arithmetic may be on the SFU whereas MUX
34# is on the higher throughput CVT unit. We get a zero argument for free relative
35# to the bitwise op, which would be LSHIFT_* internally taking a zero anyway.
36#
37# As such, it's beneficial to reexpress bitwise arithmetic of booleans as bcsel.
38opt_bool_bitwise = [
39    (('iand', 'a@1', 'b@1'), ('bcsel', a, b, False)),
40    (('ior', 'a@1', 'b@1'), ('bcsel', a, a, b)),
41    (('iand', 'a@1', ('inot', 'b@1')), ('bcsel', b, 0, a)),
42    (('ior', 'a@1', ('inot', 'b@1')), ('bcsel', b, a, True)),
43]
44
45algebraic_late = [
46    # Canonical form. The scheduler will convert back if it makes sense.
47    (('fmul', a, 2.0), ('fadd', a, a)),
48
49    # Fuse Mali-specific clamps
50    (('fmin', ('fmax', a, -1.0), 1.0), ('fsat_signed_mali', a)),
51    (('fmax', ('fmin', a, 1.0), -1.0), ('fsat_signed_mali', a)),
52    (('fmax', a, 0.0), ('fclamp_pos_mali', a)),
53
54    (('fabs', ('fddx', a)), ('fabs', ('fddx_must_abs_mali', a))),
55    (('fabs', ('fddy', b)), ('fabs', ('fddy_must_abs_mali', b))),
56
57    (('b32csel', 'b@32', ('iadd', 'a@32', 1), a), ('iadd', a, ('b2i32', b))),
58
59    # We don't have an 8-bit CSEL, so this is the best we can do.
60    # Note that we use 8-bit booleans internally to preserve vectorization.
61    (('imin', 'a@8', 'b@8'), ('b8csel', ('ilt8', a, b), a, b)),
62    (('imax', 'a@8', 'b@8'), ('b8csel', ('ilt8', a, b), b, a)),
63    (('umin', 'a@8', 'b@8'), ('b8csel', ('ult8', a, b), a, b)),
64    (('umax', 'a@8', 'b@8'), ('b8csel', ('ult8', a, b), b, a)),
65
66    # Floats are at minimum 16-bit, which means when converting to an 8-bit
67    # integer, the vectorization changes. So there's no one-shot hardware
68    # instruction for f2i8. Instead, lower to two NIR instructions that map
69    # directly to the hardware.
70    (('f2i8', a), ('i2i8', ('f2i16', a))),
71    (('f2u8', a), ('u2u8', ('f2u16', a))),
72
73    # XXX: Duplicate of nir_lower_pack
74    (('unpack_64_2x32', a), ('vec2', ('unpack_64_2x32_split_x', a),
75                                     ('unpack_64_2x32_split_y', a))),
76]
77
78# Handling all combinations of boolean and float sizes for b2f is nontrivial.
79# bcsel has the same problem in more generality; lower b2f to bcsel in NIR to
80# reuse the efficient implementations of bcsel. This includes special handling
81# to allow vectorization in places the hardware does not directly.
82#
83# Because this lowering must happen late, NIR won't squash inot in
84# automatically. Do so explicitly. (The more specific pattern must be first.)
85for bsz in [8, 16, 32]:
86    for fsz in [16, 32]:
87        algebraic_late += [
88                ((f'b2f{fsz}', ('inot', f'a@{bsz}')), (f'b{bsz}csel', a, 0.0, 1.0)),
89                ((f'b2f{fsz}', f'a@{bsz}'), (f'b{bsz}csel', a, 1.0, 0.0)),
90        ]
91
92
93def main():
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)
98    run()
99
100
101def run():
102    import nir_algebraic  # pylint: disable=import-error
103
104    print('#include "bifrost_nir.h"')
105
106    print(nir_algebraic.AlgebraicPass("bifrost_nir_opt_boolean_bitwise",
107                                      opt_bool_bitwise).render())
108    print(nir_algebraic.AlgebraicPass("bifrost_nir_lower_algebraic_late",
109                                      algebraic_late).render())
110
111
112if __name__ == '__main__':
113    main()
114