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
|
STYLE REQUIREMENTS
==================
1. Most code in this sub-directory is expected to be upstreamed into glibc so
the GNU Coding Standard and glibc specific conventions should be followed
to ease upstreaming.
2. ABI and symbols: the code should be written so it is suitable for inclusion
into a libc with minimal changes. This e.g. means that internal symbols
should be hidden and in the implementation reserved namespace according to
ISO C and POSIX rules. If possible the built shared libraries and static
library archives should be usable to override libc symbols at link time (or
at runtime via LD_PRELOAD). This requires the symbols to follow the glibc ABI
(other than symbol versioning), this cannot be done reliably for static
linking so this is a best effort requirement.
3. API: include headers should be suitable for benchmarking and testing code
and should not conflict with libc headers.
CONTRIBUTION GUIDELINES FOR math SUB-DIRECTORY
==============================================
1. Math functions have quality and performance requirements.
2. Quality:
- Worst-case ULP error should be small in the entire input domain (for most
common double precision scalar functions the target is < 0.66 ULP error,
and < 1 ULP for single precision, even performance optimized function
variant should not have > 5 ULP error if the goal is to be a drop in
replacement for a standard math function), this should be tested
statistically (or on all inputs if possible in reasonable amount of time).
The ulp tool is for this and runulp.sh should be updated for new functions.
- All standard rounding modes need to be supported but in non-default rounding
modes the quality requirement can be relaxed. (Non-nearest rounded
computation can be slow and inaccurate but has to be correct for conformance
reasons.)
- Special cases and error handling need to follow ISO C Annex F requirements,
POSIX requirements, IEEE 754-2008 requirements and Glibc requiremnts:
https://www.gnu.org/software/libc/manual/html_mono/libc.html#Errors-in-Math-Functions
this should be tested by direct tests (glibc test system may be used for it).
- Error handling code should be decoupled from the approximation code as much
as possible. (There are helper functions, these take care of errno as well
as exception raising.)
- Vector math code does not need to work in non-nearest rounding mode and error
handling side effects need not happen (fenv exceptions and errno), but the
result should be correct (within quality requirements, which are lower for
vector code than for scalar code).
- Error bounds of the approximation should be clearly documented.
- The code should build and pass tests on arm, aarch64 and x86_64 GNU linux
systems. (Routines and features can be disabled on specific targets, but
the build must complete). On aarch64, both little- and big-endian targets
are supported as well as valid combinations of architecture extensions.
The configurations that should be tested depend on the contribution.
3. Performance:
- Common math code should be benchmarked on modern aarch64 microarchitectures
over typical inputs.
- Performance improvements should be documented (relative numbers can be
published; it is enough to use the mathbench microbenchmark tool which should
be updated for new functions).
- Attention should be paid to the compilation flags: for aarch64 fma
contraction should be on and math errno turned off so some builtins can be
inlined.
- The code should be reasonably performant on x86_64 too, e.g. some rounding
instructions and fma may not be available on x86_64, such builtins turn into
libc calls with slow code. Such slowdown is not acceptable, a faster fallback
should be present: glibc and bionic use the same code on all targets. (This
does not apply to vector math code).
|
