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
|
source_root = meson.project_source_root()
mod_features = import('features')
NEON = mod_features.new(
'NEON', 1,
test_code: files(source_root + '/numpy/distutils/checks/cpu_neon.c')[0]
)
NEON_FP16 = mod_features.new(
'NEON_FP16', 2, implies: NEON,
test_code: files(source_root + '/numpy/distutils/checks/cpu_neon_fp16.c')[0]
)
# FMA
NEON_VFPV4 = mod_features.new(
'NEON_VFPV4', 3, implies: NEON_FP16,
test_code: files(source_root + '/numpy/distutils/checks/cpu_neon_vfpv4.c')[0]
)
# Advanced SIMD
ASIMD = mod_features.new(
'ASIMD', 4, implies: NEON_VFPV4, detect: {'val': 'ASIMD', 'match': 'NEON.*'},
test_code: files(source_root + '/numpy/distutils/checks/cpu_asimd.c')[0]
)
cpu_family = host_machine.cpu_family()
if cpu_family == 'aarch64'
# hardware baseline
NEON.update(implies: [NEON_FP16, NEON_VFPV4, ASIMD])
NEON_FP16.update(implies: [NEON, NEON_VFPV4, ASIMD])
NEON_VFPV4.update(implies: [NEON, NEON_FP16, ASIMD])
elif cpu_family == 'arm'
NEON.update(args: '-mfpu=neon')
NEON_FP16.update(args: ['-mfp16-format=ieee', {'val': '-mfpu=neon-fp16', 'match': '-mfpu=.*'}])
NEON_VFPV4.update(args: [{'val': '-mfpu=neon-vfpv4', 'match': '-mfpu=.*'}])
ASIMD.update(args: [
{'val': '-mfpu=neon-fp-armv8', 'match': '-mfpu=.*'},
'-march=armv8-a+simd'
])
endif
# ARMv8.2 half-precision & vector arithm
ASIMDHP = mod_features.new(
'ASIMDHP', 5, implies: ASIMD,
args: {'val': '-march=armv8.2-a+fp16', 'match': '-march=.*', 'mfilter': '\+.*'},
test_code: files(source_root + '/numpy/distutils/checks/cpu_asimdhp.c')[0]
)
## ARMv8.2 dot product
ASIMDDP = mod_features.new(
'ASIMDDP', 6, implies: ASIMD,
args: {'val': '-march=armv8.2-a+dotprod', 'match': '-march=.*', 'mfilter': '\+.*'},
test_code: files(source_root + '/numpy/distutils/checks/cpu_asimddp.c')[0]
)
## ARMv8.2 Single & half-precision Multiply
ASIMDFHM = mod_features.new(
'ASIMDFHM', 7, implies: ASIMDHP,
args: {'val': '-march=armv8.2-a+fp16fml', 'match': '-march=.*', 'mfilter': '\+.*'},
test_code: files(source_root + '/numpy/distutils/checks/cpu_asimdfhm.c')[0]
)
## Scalable Vector Extensions (SVE)
SVE = mod_features.new(
'SVE', 8, implies: ASIMDHP,
args: {'val': '-march=armv8.2-a+sve', 'match': '-march=.*', 'mfilter': '\+.*'},
test_code: files(source_root + '/numpy/distutils/checks/cpu_sve.c')[0]
)
# TODO: Add support for MSVC
ARM_FEATURES = {
'NEON': NEON, 'NEON_FP16': NEON_FP16, 'NEON_VFPV4': NEON_VFPV4,
'ASIMD': ASIMD, 'ASIMDHP': ASIMDHP, 'ASIMDFHM': ASIMDFHM,
'SVE': SVE
}
|
