GeezerVison is showing libc.llvm.org — Math Functions - The LLVM C Library
Math Functions#
Implementation Requirements / Goals#
The highest priority is to be as accurate as possible, according to the C and
IEEE 754 standards. By default, we will aim to be correctly rounded for all rounding modes.
The current rounding mode of the floating point environment is used to perform
computations and produce the final results.
Our next requirement is that the outputs are consistent across all platforms.
Notice that the consistency requirement will be satisfied automatically if the
implementation is correctly rounded. Our last requirement for the implementations is to have good and predicable
performance:
The average performance should be comparable to other libc
implementations. The worst case performance should be within 10X-20X of the average. Platform-specific implementations or instructions could be added whenever it
makes sense and provides significant performance boost.
For other use cases that have strict requirements on the code size, memory
footprint, or latency, such as embedded systems, we will aim to be as accurate
as possible within the memory or latency budgets, and consistent across all
platforms.
Add a new math function to LLVM libc#
Implementation Status#
C23 and POSIX.1-2024 header status tables: math.h To check math functions enabled for Linux:
To check math functions enabled for Windows:
To check math functions enabled for macOS:
To check math functions enabled for GPU:
To check math functions enabled for embedded system:
GPU Conformance#
Conformance tests are located at: offload/unittests/Conformance. The math functions for GPUs are compiled with the following optimization options: LIBC_MATH_SKIP_ACCURATE_PASS, LIBC_MATH_INTERMEDIATE_COMP_IN_FLOAT, LIBC_MATH_SMALL_TABLES, LIBC_MATH_NO_ERRNO, and LIBC_MATH_NO_EXCEPT. The conformance test results for higher math functions on GPUs are reported in the table below. The results show the maximum observed ULP distance when comparing a given GPU implementation against the corresponding correctly rounded implementation from LLVM libc, which is computed on the host CPU and serves as the reference. For comparison purposes, results for CUDA Math and HIP Math against the same reference are also included.
Function |
Test Method |
ULP Tolerance |
Max ULP Distance |
|---|
LLVM libc
(AMDGPU) |
LLVM libc
(CUDA) |
CUDA Math
(CUDA) |
HIP Math
(AMDGPU) |
|---|
acos |
Randomized |
4 |
6 (FAILED) |
6 (FAILED) |
1 |
1 |
acosf |
Exhaustive |
4 |
1 |
1 |
1 |
1 |
acosf16 |
Exhaustive |
2 |
1 |
1 |
|
1 |
acoshf |
Exhaustive |
4 |
1 |
1 |
2 |
1 |
acoshf16 |
Exhaustive |
2 |
0 |
0 |
|
0 |
acospif16 |
Exhaustive |
2 |
0 |
0 |
|
|
asin |
Randomized |
4 |
6 (FAILED) |
6 (FAILED) |
2 |
1 |
asinf |
Exhaustive |
4 |
1 |
1 |
1 |
3 |
asinf16 |
Exhaustive |
2 |
0 |
0 |
|
2 |
asinhf |
Exhaustive |
4 |
1 |
1 |
2 |
1 |
asinhf16 |
Exhaustive |
2 |
1 |
1 |
|
1 |
atanf |
Exhaustive |
5 |
0 |
0 |
1 |
2 |
atanf16 |
Exhaustive |
2 |
1 |
1 |
|
1 |
atan2f |
Randomized |
6 |
1 |
1 |
2 |
3 |
atanhf |
Exhaustive |
5 |
0 |
0 |
3 |
1 |
atanhf16 |
Exhaustive |
2 |
0 |
0 |
|
1 |
cbrt |
Randomized |
2 |
1 |
1 |
1 |
1 |
cbrtf |
Exhaustive |
2 |
0 |
0 |
1 |
1 |
cos |
Randomized |
4 |
1 |
1 |
2 |
1 |
cosf |
Exhaustive |
4 |
1 |
1 |
2 |
2 |
cosf16 |
Exhaustive |
2 |
1 |
1 |
|
1 |
coshf |
Exhaustive |
4 |
0 |
0 |
2 |
1 |
coshf16 |
Exhaustive |
2 |
1 |
0 |
|
1 |
cospif |
Exhaustive |
4 |
0 |
0 |
1 |
1 |
cospif16 |
Exhaustive |
2 |
0 |
0 |
|
|
erff |
Exhaustive |
16 |
0 |
0 |
1 |
2 |
exp |
Randomized |
3 |
1 |
1 |
1 |
1 |
expf |
Exhaustive |
3 |
0 |
0 |
2 |
1 |
expf16 |
Exhaustive |
2 |
1 |
1 |
|
1 |
exp10 |
Randomized |
3 |
1 |
1 |
1 |
1 |
exp10f |
Exhaustive |
3 |
0 |
0 |
2 |
1 |
exp10f16 |
Exhaustive |
2 |
1 |
1 |
|
1 |
exp2 |
Randomized |
3 |
1 |
1 |
1 |
1 |
exp2f |
Exhaustive |
3 |
1 |
1 |
2 |
1 |
exp2f16 |
Exhaustive |
2 |
1 |
1 |
|
0 |
expm1 |
Randomized |
3 |
0 |
0 |
1 |
2 |
expm1f |
Exhaustive |
3 |
1 |
1 |
1 |
1 |
expm1f16 |
Exhaustive |
2 |
1 |
1 |
|
1 |
hypot |
Randomized |
4 |
0 |
0 |
2 |
1 |
hypotf |
Randomized |
4 |
0 |
0 |
1 |
2 |
hypotf16 |
Exhaustive |
2 |
0 |
0 |
|
|
log |
Randomized |
3 |
1 |
1 |
1 |
1 |
logf |
Exhaustive |
3 |
1 |
1 |
1 |
2 |
logf16 |
Exhaustive |
2 |
1 |
1 |
|
1 |
log10 |
Randomized |
3 |
1 |
1 |
1 |
1 |
log10f |
Exhaustive |
3 |
1 |
1 |
2 |
2 |
log10f16 |
Exhaustive |
2 |
1 |
1 |
|
1 |
log1p |
Randomized |
2 |
1 |
1 |
1 |
1 |
log1pf |
Exhaustive |
2 |
1 |
1 |
1 |
1 |
log2 |
Randomized |
3 |
1 |
1 |
1 |
1 |
log2f |
Exhaustive |
3 |
0 |
0 |
1 |
1 |
log2f16 |
Exhaustive |
2 |
1 |
1 |
|
0 |
powf (integer exp.) |
Randomized |
16 |
0 |
0 |
2 |
1 |
powf (real exp.) |
Randomized |
16 |
0 |
0 |
2 |
1 |
sin |
Randomized |
4 |
1 |
1 |
1 |
1 |
sinf |
Exhaustive |
4 |
1 |
1 |
1 |
2 |
sinf16 |
Exhaustive |
2 |
1 |
1 |
|
1 |
sincos (cos part) |
Randomized |
4 |
1 |
1 |
2 |
1 |
sincos (sin part) |
Randomized |
4 |
1 |
1 |
1 |
1 |
sincosf (cos part) |
Exhaustive |
4 |
1 |
1 |
2 |
2 |
sincosf (sin part) |
Exhaustive |
4 |
1 |
1 |
1 |
2 |
sinhf |
Exhaustive |
4 |
1 |
1 |
3 |
1 |
sinhf16 |
Exhaustive |
2 |
1 |
1 |
|
1 |
sinpif |
Exhaustive |
4 |
0 |
0 |
1 |
1 |
sinpif16 |
Exhaustive |
2 |
0 |
0 |
|
|
tan |
Randomized |
5 |
2 |
2 |
2 |
1 |
tanf |
Exhaustive |
5 |
0 |
0 |
3 |
2 |
tanf16 |
Exhaustive |
2 |
1 |
1 |
|
2 |
tanhf |
Exhaustive |
5 |
0 |
0 |
2 |
1 |
tanhf16 |
Exhaustive |
2 |
0 |
0 |
|
1 |
tanpif |
Exhaustive |
6 |
0 |
0 |
|
|
tanpif16 |
Exhaustive |
2 |
1 |
1 |
|
|
Notes:
Exhaustive tests check every representable point in the input space. This method is used for half-precision functions and single-precision univariate functions. Randomized tests check a large, deterministic subset of the input space, typically using 232 samples. This method is used for functions with larger input spaces, such as single-precision bivariate and double-precision functions. ULP tolerances are based on The Khronos Group, The OpenCL C Specification v3.0.19, Sec. 7.4, Khronos Registry [July 10, 2025]. The AMD GPU used for testing is AMD Radeon RX 6950 XT. The NVIDIA GPU used for testing is NVIDIA RTX 4000 SFF Ada Generation.
Performance#
Simple performance testings are located at: libc/test/src/math/performance_testing. We also use the perf tool from the CORE-MATH
project: link.
The performance results from the CORE-MATH’s perf tool are reported in the
table below, using the system library as reference (such as the GNU C library
on Linux). Fmod performance results obtained with “performance_testing”.
<Func> |
Reciprocal throughput (clk) |
Latency (clk) |
Testing ranges |
Testing configuration |
|---|
LLVM libc |
Reference (glibc) |
LLVM libc |
Reference (glibc) |
CPU |
OS |
Compiler |
Special flags |
|---|
acosf |
24 |
29 |
62 |
77 |
\([-1, 1]\) |
Ryzen 1700 |
Ubuntu 22.04 LTS x86_64 |
Clang 14.0.0 |
FMA |
acoshf |
18 |
26 |
73 |
74 |
\([1, 21]\) |
Ryzen 1700 |
Ubuntu 22.04 LTS x86_64 |
Clang 14.0.0 |
FMA |
asinf |
23 |
27 |
62 |
62 |
\([-1, 1]\) |
Ryzen 1700 |
Ubuntu 22.04 LTS x86_64 |
Clang 14.0.0 |
FMA |
asinhf |
21 |
39 |
77 |
91 |
\([-10, 10]\) |
Ryzen 1700 |
Ubuntu 22.04 LTS x86_64 |
Clang 14.0.0 |
FMA |
atanf |
27 |
29 |
79 |
68 |
\([-10, 10]\) |
Ryzen 1700 |
Ubuntu 22.04 LTS x86_64 |
Clang 14.0.0 |
FMA |
atanhf |
18 |
66 |
68 |
133 |
\([-1, 1]\) |
Ryzen 1700 |
Ubuntu 22.04 LTS x86_64 |
Clang 14.0.0 |
FMA |
cosf |
13 |
32 |
53 |
59 |
\([0, 2\pi]\) |
Ryzen 1700 |
Ubuntu 20.04 LTS x86_64 |
Clang 12.0.0 |
FMA |
coshf |
14 |
20 |
50 |
48 |
\([-10, 10]\) |
Ryzen 1700 |
Ubuntu 22.04 LTS x86_64 |
Clang 14.0.0 |
FMA |
expf |
9 |
7 |
44 |
38 |
\([-10, 10]\) |
Ryzen 1700 |
Ubuntu 20.04 LTS x86_64 |
Clang 12.0.0 |
FMA |
exp10f |
10 |
8 |
40 |
38 |
\([-10, 10]\) |
Ryzen 1700 |
Ubuntu 22.04 LTS x86_64 |
Clang 14.0.0 |
FMA |
exp2f |
9 |
6 |
35 |
31 |
\([-10, 10]\) |
Ryzen 1700 |
Ubuntu 22.04 LTS x86_64 |
Clang 14.0.0 |
FMA |
expm1f |
9 |
44 |
42 |
121 |
\([-10, 10]\) |
Ryzen 1700 |
Ubuntu 20.04 LTS x86_64 |
Clang 12.0.0 |
FMA |
fmodf |
73 |
263 |
|
|
[MIN_NORMAL, MAX_NORMAL] |
i5 mobile |
Ubuntu 20.04 LTS x86_64 |
Clang 12.0.0 |
|
9 |
11 |
|
|
[0, MAX_SUBNORMAL] |
i5 mobile |
Ubuntu 20.04 LTS x86_64 |
Clang 12.0.0 |
|
fmod |
595 |
3297 |
|
|
[MIN_NORMAL, MAX_NORMAL] |
i5 mobile |
Ubuntu 20.04 LTS x86_64 |
Clang 12.0.0 |
|
14 |
13 |
|
|
[0, MAX_SUBNORMAL] |
i5 mobile |
Ubuntu 20.04 LTS x86_64 |
Clang 12.0.0 |
|
hypotf |
25 |
15 |
64 |
49 |
\([-10, 10] \times [-10, 10]\) |
Ryzen 1700 |
Ubuntu 20.04 LTS x86_64 |
Clang 12.0.0 |
|
logf |
12 |
10 |
56 |
46 |
\([e^{-1}, e]\) |
Ryzen 1700 |
Ubuntu 20.04 LTS x86_64 |
Clang 12.0.0 |
FMA |
log10f |
9 |
17 |
35 |
48 |
\([e^{-1}, e]\) |
Ryzen 5900X |
Ubuntu 22.04 LTS x86_64 |
Clang 15.0.6 |
FMA |
log1pf |
16 |
33 |
61 |
97 |
\([e^{-0.5} - 1, e^{0.5} - 1]\) |
Ryzen 1700 |
Ubuntu 20.04 LTS x86_64 |
Clang 12.0.0 |
FMA |
log2f |
13 |
10 |
57 |
46 |
\([e^{-1}, e]\) |
Ryzen 1700 |
Ubuntu 20.04 LTS x86_64 |
Clang 12.0.0 |
FMA |
sinf |
12 |
25 |
51 |
57 |
\([-\pi, \pi]\) |
Ryzen 1700 |
Ubuntu 20.04 LTS x86_64 |
Clang 12.0.0 |
FMA |
sincosf |
19 |
30 |
57 |
68 |
\([-\pi, \pi]\) |
Ryzen 1700 |
Ubuntu 20.04 LTS x86_64 |
Clang 12.0.0 |
FMA |
sinhf |
13 |
63 |
48 |
137 |
\([-10, 10]\) |
Ryzen 1700 |
Ubuntu 22.04 LTS x86_64 |
Clang 14.0.0 |
FMA |
tanf |
16 |
50 |
61 |
107 |
\([-\pi, \pi]\) |
Ryzen 1700 |
Ubuntu 22.04 LTS x86_64 |
Clang 14.0.0 |
FMA |
tanhf |
13 |
55 |
57 |
123 |
\([-10, 10]\) |
Ryzen 1700 |
Ubuntu 22.04 LTS x86_64 |
Clang 14.0.0 |
FMA |
Algorithms + Implementation Details#
|