1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
|
#
# Copyright (C) 2020 Microsoft Corporation
#
# Copyright (C) 2018 Alyssa Rosenzweig
#
# Copyright (C) 2016 Intel Corporation
#
# Permission is hereby granted, free of charge, to any person obtaining a
# copy of this software and associated documentation files (the "Software"),
# to deal in the Software without restriction, including without limitation
# the rights to use, copy, modify, merge, publish, distribute, sublicense,
# and/or sell copies of the Software, and to permit persons to whom the
# Software is furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice (including the next
# paragraph) shall be included in all copies or substantial portions of the
# Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
# IN THE SOFTWARE.
import argparse
import sys
import math
a = 'a'
b = 'b'
c = 'c'
# The nir_lower_bit_size() pass gets rid of all 8bit ALUs but insert new u2u8
# and i2i8 operations to convert the result back to the original type after the
# arithmetic operation is done. Those u2u8 and i2i8 operations, as any other
# 8bit operations, are not supported by DXIL and needs to be discarded. The
# dxil_nir_lower_8bit_conv() pass is here for that.
# Similarly, some hardware doesn't support 16bit values
no_8bit_conv = []
no_16bit_conv = []
def remove_unsupported_casts(arr, bit_size, mask, max_unsigned_float, min_signed_float, max_signed_float):
for outer_op_type in ('u2u', 'i2i', 'u2f', 'i2f'):
for outer_op_sz in (16, 32, 64):
if outer_op_sz == bit_size:
continue
outer_op = outer_op_type + str(int(outer_op_sz))
for inner_op_type in ('u2u', 'i2i'):
inner_op = inner_op_type + str(int(bit_size))
for src_sz in (16, 32, 64):
if (src_sz == bit_size):
continue
# Coming from integral, truncate appropriately
orig_seq = (outer_op, (inner_op, 'a@' + str(int(src_sz))))
if (outer_op[0] == 'u'):
new_seq = ('iand', a, mask)
else:
shift = src_sz - bit_size
new_seq = ('ishr', ('ishl', a, shift), shift)
# Make sure the destination is the right type/size
if outer_op_sz != src_sz or outer_op[2] != inner_op[0]:
new_seq = (outer_op, new_seq)
arr += [(orig_seq, new_seq)]
for inner_op_type in ('f2u', 'f2i'):
inner_op = inner_op_type + str(int(bit_size))
if (outer_op[2] == 'f'):
# From float and to float, just truncate via min/max, and ensure the right float size
for src_sz in (16, 32, 64):
if (src_sz == bit_size):
continue
orig_seq = (outer_op, (inner_op, 'a@' + str(int(src_sz))))
if (outer_op[0] == 'u'):
new_seq = ('fmin', ('fmax', a, 0.0), max_unsigned_float)
else:
new_seq = ('fmin', ('fmax', a, min_signed_float), max_signed_float)
if outer_op_sz != src_sz:
new_seq = ('f2f' + str(int(outer_op_sz)), new_seq)
arr += [(orig_seq, new_seq)]
else:
# From float to integral, convert to integral type first, then truncate
orig_seq = (outer_op, (inner_op, a))
float_conv = ('f2' + inner_op[2] + str(int(outer_op_sz)), a)
if (outer_op[0] == 'u'):
new_seq = ('iand', float_conv, mask)
else:
shift = outer_op_sz - bit_size
new_seq = ('ishr', ('ishl', float_conv, shift), shift)
arr += [(orig_seq, new_seq)]
remove_unsupported_casts(no_8bit_conv, 8, 0xff, 255.0, -128.0, 127.0)
remove_unsupported_casts(no_16bit_conv, 16, 0xffff, 65535.0, -32768.0, 32767.0)
algebraic_ops = [
(('b2b32', 'a'), ('b2i32', 'a')),
(('b2b1', 'a'), ('ine', ('b2i32', a), 0)),
# We don't support the saturating versions of these
(('sdot_4x8_iadd_sat', a, b, c), ('iadd_sat', ('sdot_4x8_iadd', a, b, 0), c)),
(('udot_4x8_uadd_sat', a, b, c), ('uadd_sat', ('udot_4x8_uadd', a, b, 0), c)),
]
no_16bit_conv += [
(('f2f32', ('u2u16', 'a@32')), ('unpack_half_2x16_split_x', 'a')),
(('u2u32', ('f2f16_rtz', 'a@32')), ('pack_half_2x16_split', 'a', 0)),
]
def main():
parser = argparse.ArgumentParser()
parser.add_argument('-p', '--import-path', required=True)
args = parser.parse_args()
sys.path.insert(0, args.import_path)
run()
def run():
import nir_algebraic # pylint: disable=import-error
print('#include "dxil_nir.h"')
print(nir_algebraic.AlgebraicPass("dxil_nir_lower_8bit_conv",
no_8bit_conv).render())
print(nir_algebraic.AlgebraicPass("dxil_nir_lower_16bit_conv",
no_16bit_conv).render())
print(nir_algebraic.AlgebraicPass("dxil_nir_algebraic",
algebraic_ops).render())
if __name__ == '__main__':
main()
|
