Fast math functions - BelaPlatform/Bela GitHub Wiki

This material has been superseded. Visit learn.bela.io for the maintained version.

When programming in C++ on Bela, you have access to a library of fast, approximated math function: math_neon. This is a library that aims at replacing the accurate math functions implemented by the libmath library with faster, optimized and less accurate functions.

In order to use it, you will have to add

# include <libraries/math_neon/math_neon.h>

at the top of your file and then you can invoke the functions by adding _neon at the end of the name of a regular math.h function, e.g.: sinf(float) becomes sinf_neon(float)

Being these functions approximate, they may not suit all purposes and it is up to you to verify that they are accurate for your specific case. A list of the supported functions and a summary benchmark is in the table below. "rate" is the relative improvement of the "fast" version compared to the one in libmath.

Function Range Number ABS Max Err REL Max Err(%) RMS Err Time Rate
sinf -3.14, 3.14 500000 0.00000000 0.00000000 0.00000000 262041 1.00
sinf_neon -3.14, 3.14 500000 0.00000083 0.85972977 0.00000041 66440 3.94
cosf -3.14, 3.14 500000 0.00000000 0.00000000 0.00000000 215722 1.00
cosf_neon -3.14, 3.14 500000 0.00000083 0.67412275 0.00000042 77271 2.79
tanf -0.79, 0.79 500000 0.00000000 0.00000000 0.00000000 250169 1.00
tanf_neon -0.79, 0.79 500000 0.00000191 0.00036218 0.00000067 97411 2.57
asinf -1.00, 1.00 500000 0.00000000 0.00000000 0.00000000 229272 1.00
asinf_neon -1.00, 1.00 500000 0.00004661 0.00889725 nan 121673 1.88
acosf -1.00, 1.00 500000 0.00000000 0.00000000 0.00000000 277243 1.00
acosf_neon -1.00, 1.00 500000 0.00004673 0.00634698 nan 123174 2.25
atanf -1.00, 1.00 500000 0.00000000 0.00000000 0.00000000 206506 1.00
atanf_neon -1.00, 1.00 500000 0.00016665 0.02123393 0.00007395 89627 2.30
sinhf -3.14, 3.14 500000 0.00000000 0.00000000 0.00000000 313378 1.00
sinhf_neon -3.14, 3.14 500000 0.00000191 0.15247960 0.00000019 75931 4.13
coshf -3.14, 3.14 500000 0.00000000 0.00000000 0.00000000 327628 1.00
coshf_neon -3.14, 3.14 500000 0.00000191 0.00002146 0.00000017 75932 4.31
tanhf -3.14, 3.14 500000 0.00000000 0.00000000 0.00000000 322052 1.00
tanhf_neon -3.14, 3.14 500000 0.00000024 0.24726489 0.00000005 74937 4.30
expf 0.00, 10.00 500000 0.00000000 0.00000000 0.00000000 288120 1.00
expf_neon 0.00, 10.00 500000 0.00976562 0.00006549 0.00164461 51935 5.55
logf 1.00, 1000.00 500000 0.00000000 0.00000000 0.00000000 220097 1.00
logf_neon 1.00, 1000.00 500000 0.00000763 0.01027698 0.00000107 52387 4.20
log10f 1.00, 1000.00 500000 0.00000000 0.00000000 0.00000000 273479 1.00
log10f_neon 1.00, 1000.00 500000 0.00000334 0.00668141 0.00000048 52388 5.22
floorf 1.00, 1000.00 5000000 0.00000000 0.00000000 0.00000000 566063 1.00
floorf_neon 1.00, 1000.00 5000000 0.00000000 0.00000000 0.00000000 339748 1.67
ceilf 1.00, 1000.00 5000000 0.00000000 0.00000000 0.00000000 617084 1.00
ceilf_neon 1.00, 1000.00 5000000 0.00000000 0.00000000 0.00000000 339745 1.82
fabsf 1.00, 1000.00 5000000 0.00000000 0.00000000 0.00000000 332227 1.00
fabsf_neon 1.00, 1000.00 5000000 0.00000000 0.00000000 0.00000000 226497 1.47
sqrtf 1.00, 1000.00 500000 0.00000000 0.00000000 0.00000000 152492 1.00
sqrtf_neon 1.00, 1000.00 500000 0.00000763 0.00002913 0.00000160 53386 2.86
invsqrtf 1.00, 1000.00 500000 0.00000000 0.00000000 0.00000000 54986 1.00
invsqrtf_neon 1.00, 1000.00 500000 0.00000012 0.00002124 0.00000000 38418 1.43
atan2f 0.10, 10.00 10000 0.00000000 0.00000000 0.00000000 511718 1.00
atan2f_neon 0.10, 10.00 10000 0.00016665 0.02124334 0.00000000 203125 2.52
powf 1.00, 10.00 10000 0.00000000 0.00000000 0.00000000 1783204 1.00
powf_neon 1.00, 10.00 10000 136192.00000000 0.00587880 0.00000000 201172 8.86
fmodf 1.00, 10.00 10000 0.00000000 0.00000000 0.00000000 140625 1.00
fmodf_neon 1.00, 10.00 10000 9.97472000 0.08058017 0.00000000 148437 0.95