• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1#
2# Copyright (C) 2018 Alyssa Rosenzweig
3# Copyright (C) 2019-2020 Collabora, Ltd.
4#
5# Copyright (C) 2016 Intel Corporation
6#
7# Permission is hereby granted, free of charge, to any person obtaining a
8# copy of this software and associated documentation files (the "Software"),
9# to deal in the Software without restriction, including without limitation
10# the rights to use, copy, modify, merge, publish, distribute, sublicense,
11# and/or sell copies of the Software, and to permit persons to whom the
12# Software is furnished to do so, subject to the following conditions:
13#
14# The above copyright notice and this permission notice (including the next
15# paragraph) shall be included in all copies or substantial portions of the
16# Software.
17#
18# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
21# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
22# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
23# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24# IN THE SOFTWARE.
25
26import argparse
27import sys
28import math
29
30a = 'a'
31b = 'b'
32c = 'c'
33
34algebraic = [
35   (('pack_unorm_4x8', a), ('pack_32_4x8', ('f2u8', ('fround_even', ('fmul', ('fsat', a), 255.0))))),
36
37   # Allows us to schedule as a multiply by 2
38   (('~fadd', ('fadd', a, b), a), ('fadd', ('fadd', a, a), b)),
39]
40
41algebraic_late = [
42    # ineg must be lowered late, but only for integers; floats will try to
43    # have modifiers attached... hence why this has to be here rather than
44    # a more standard lower_negate approach
45
46    (('ineg', a), ('isub', 0, a)),
47
48    # Likewise we want fsub lowered but not isub
49    (('fsub', a, b), ('fadd', a, ('fneg', b))),
50
51    # These two special-cases save space/an op than the actual csel op +
52    # scheduler flexibility
53
54    (('b32csel', a, 'b@32', 0), ('iand', a, b)),
55    (('b32csel', a, 0, 'b@32'), ('iand', ('inot', a), b)),
56
57    # Fuse sat_signed. This should probably be shared with Bifrost
58    (('~fmin', ('fmax', a, -1.0), 1.0), ('fsat_signed', a)),
59    (('~fmax', ('fmin', a, 1.0), -1.0), ('fsat_signed', a)),
60
61    # Fuse clamp_positive. This should probably be shared with Utgard/bifrost
62    (('fmax', a, 0.0), ('fclamp_pos', a)),
63
64    (('ishl', 'a@16', b), ('u2u16', ('ishl', ('u2u32', a), b))),
65    (('ishr', 'a@16', b), ('i2i16', ('ishr', ('i2i32', a), b))),
66    (('ushr', 'a@16', b), ('u2u16', ('ushr', ('u2u32', a), b))),
67
68    (('ishl', 'a@8', b), ('u2u8', ('u2u16', ('ishl', ('u2u32', ('u2u16', a)), b)))),
69    (('ishr', 'a@8', b), ('i2i8', ('i2i16', ('ishr', ('i2i32', ('i2i16', a)), b)))),
70    (('ushr', 'a@8', b), ('u2u8', ('u2u16', ('ushr', ('u2u32', ('u2u16', a)), b)))),
71
72    # Canonical form. The scheduler will convert back if it makes sense.
73    (('fmul', a, 2.0), ('fadd', a, a))
74]
75
76
77# Midgard is able to type convert down by only one "step" per instruction; if
78# NIR wants more than one step, we need to break up into multiple instructions.
79# Nevertheless, we can do both a size step and a floating/int step at once.
80
81converts = []
82
83for op in ('u2u', 'i2i', 'f2f', 'i2f', 'u2f', 'f2i', 'f2u'):
84    srcsz_max = 64
85    dstsz_max = 64
86    # 8 bit float doesn't exist
87    srcsz_min = 8 if op[0] != 'f' else 16
88    dstsz_min = 8 if op[2] != 'f' else 16
89    dstsz = dstsz_min
90    # Iterate over all possible destination and source sizes
91    while dstsz <= dstsz_max:
92        srcsz = srcsz_min
93        while srcsz <= srcsz_max:
94            # Size converter lowering is only needed if src and dst sizes are
95            # spaced by a factor > 2.
96            if srcsz != dstsz and (srcsz * 2 != dstsz and srcsz != dstsz * 2):
97                cursz = srcsz
98                rule = a
99                # When converting down we first do the type conversion followed
100                # by one or more size conversions. When converting up, we do
101                # the type conversion at the end. This way we don't have to
102                # deal with the fact that f2f8 doesn't exists.
103                sizeconvop = op[0] + '2' + op[0] if srcsz < dstsz else op[2] + '2' + op[2]
104                if srcsz > dstsz and op[0] != op[2]:
105                    rule = (op + str(int(cursz)), rule)
106                while cursz != dstsz:
107                    cursz = cursz / 2 if dstsz < srcsz else cursz * 2
108                    rule = (sizeconvop + str(int(cursz)), rule)
109                if srcsz < dstsz and op[0] != op[2]:
110                    rule = (op + str(int(cursz)), rule)
111                converts += [((op + str(int(dstsz)), 'a@' + str(int(srcsz))), rule)]
112            srcsz *= 2
113        dstsz *= 2
114
115# Try to force constants to the right
116constant_switch = [
117        # fge gets flipped to fle, so we invert to keep the order
118        (('fge', 'a', '#b'), (('inot', ('flt', a, b)))),
119        (('fge32', 'a', '#b'), (('inot', ('flt32', a, b)))),
120        (('ige32', 'a', '#b'), (('inot', ('ilt32', a, b)))),
121        (('uge32', 'a', '#b'), (('inot', ('ult32', a, b)))),
122
123        # fge gets mapped to fle with a flip
124        (('flt32', '#a', 'b'), ('inot', ('fge32', a, b))),
125        (('ilt32', '#a', 'b'), ('inot', ('ige32', a, b))),
126        (('ult32', '#a', 'b'), ('inot', ('uge32', a, b)))
127]
128
129# ..since the above switching happens after algebraic stuff is done
130cancel_inot = [
131        (('inot', ('inot', a)), a)
132]
133
134# Midgard scales fsin/fcos arguments by pi.
135# Pass must be run only once, after the main loop
136
137scale_trig = [
138        (('fsin', a), ('fsin', ('fdiv', a, math.pi))),
139        (('fcos', a), ('fcos', ('fdiv', a, math.pi))),
140]
141
142def main():
143    parser = argparse.ArgumentParser()
144    parser.add_argument('-p', '--import-path', required=True)
145    args = parser.parse_args()
146    sys.path.insert(0, args.import_path)
147    run()
148
149
150def run():
151    import nir_algebraic  # pylint: disable=import-error
152
153    print('#include "midgard_nir.h"')
154
155    print(nir_algebraic.AlgebraicPass("midgard_nir_lower_algebraic_early",
156                                      algebraic).render())
157
158    print(nir_algebraic.AlgebraicPass("midgard_nir_lower_algebraic_late",
159                                      algebraic_late + converts + constant_switch).render())
160
161    print(nir_algebraic.AlgebraicPass("midgard_nir_scale_trig",
162                                      scale_trig).render())
163
164    print(nir_algebraic.AlgebraicPass("midgard_nir_cancel_inot",
165                                      cancel_inot).render())
166
167
168if __name__ == '__main__':
169    main()
170