What is the fastest way to compute sin and cos together?

Question

I would like to compute both the sine and co-sine of a value together (for example to create a rotation matrix). Of course I could compute them separately one after another like a = cos(x); b = sin(x);, but I wonder if there is a faster way when needing both values.

Edit: To summarize the answers so far:

• Vlad said, that there is the asm command FSINCOS computing both of them (in almost the same time as a call to FSIN alone)

• Like Chi noticed, this optimization is sometimes already done by the compiler (when using optimization flags).

• caf pointed out, that functions sincos and sincosf are probably available and can be called directly by just including math.h

• tanascius approach of using a look-up table is discussed controversial. (However on my computer and in a benchmark scenario it runs 3x faster than sincos with almost the same accuracy for 32-bit floating points.)

• Joel Goodwin linked to an interesting approach of an extremly fast approximation technique with quite good accuray (for me, this is even faster then the table look-up)

Answer 1

You could compute either and then use the identity:

cos(x)2 = 1 - sin(x)2

but as @tanascius says, a precomputed table is the way to go.

Answer 2

When you need performance, you could use a precalculated sin/cos table (one table will do, stored as a Dictionary). Well, it depends on the accuracy you need (maybe the table would be to big), but it should be really fast.

Answer 3

When performance is critical for this kind of thing it is not unusual to introduce a lookup table.

Answer 4

Have you thought of declaring lookup tables for the two functions? You'd still have to "calculate" sin(x) and cos(x), but it'd be decidedly faster, if you don't need a high degree of accuracy.

Answer 5

Modern Intel/AMD processors have instruction FSINCOS for calculating sine and cosine functions simultaneously. If you need strong optimization, perhaps you should use it.

Here is a small example: http://home.broadpark.no/~alein/fsincos.html

Here is another example (for MSVC): http://www.codeguru.com/forum/showthread.php?t=328669

Here is yet another example (with gcc): http://www.allegro.cc/forums/thread/588470

Hope one of them helps. (I didn't use this instruction myself, sorry.)

As they are supported on processor level, I expect them to be way much faster than table lookups.

Edit:
Wikipedia suggests that FSINCOS was added at 387 processors, so you can hardly find a processor which doesn't support it.

Edit:
Intel's documentation states that FSINCOS is just about 5 times slower than FDIV (i.e., floating point division).

Edit:
Please note that not all modern compilers optimize calculation of sine and cosine into a call to FSINCOS. In particular, my VS 2008 didn't do it that way.

Answer 6

Technically, you’d achieve this by using complex numbers and Euler’s Formula. Thus, something like (C++)

complex<double> res = exp(complex<double>(0, x));
// or equivalent
complex<double> res = polar<double>(1, x);
double sin_x = res.imag();
double cos_x = res.real();

should give you sine and cosine in one step. How this is done internally is a question of the compiler and library being used. It could (and might) well take longer to do it this way (just because Euler’s Formula is mostly used to compute the complex exp using sin and cos – and not the other way round) but there might be some theoretical optimisation possible.

---

Edit

The headers in <complex> for GNU C++ 4.2 are using explicit calculations of sin and cos inside polar, so it doesn’t look too good for optimisations there unless the compiler does some magic (see the -ffast-math and -mfpmath switches as written in Chi’s answer).

Answer 7

I don't believe that lookup tables are necessarily a good idea for this problem. Unless your accuracy requirements are very low the table needs to be very large. And modern CPUs can do a lot of computation while a value is fetched from main memory. This is not one of those questions which can be properly answered by argument (not even mine), test and measure and consider the data.

But I'd look to the fast implementations of SinCos that you find in libraries such as AMD's ACML and Intel's MKL.

Answer 8

There is very interesting stuff on this forum page, which is focused on finding good approximations that are fast: http://www.devmaster.net/forums/showthread.php?t=5784

Disclaimer: Not used any of this stuff myself.

Answer 9

For a creative approach, how about expanding the Taylor series? Since they have similar terms, you could do something like the following pseudo:

numerator = x
denominator = 1
sine = x
cosine = 1
op = -1
fact = 1

while (not enough precision) {
    fact++
    denominator *= fact
    numerator *= x

    cosine += op * numerator / denominator

    fact++
    denominator *= fact
    numerator *= x

    sine += op * numerator / denominator

    op *= -1
}

This means you do something like this: starting at x and 1 for sin and cosine, follow the pattern - subtract x^2 / 2! from cosine, subtract x^3 / 3! from sine, add x^4 / 4! to cosine, add x^5 / 5! to sine...

I have no idea whether this would be performant. If you need less precision than the built in sin() and cos() give you, it may be an option.

Answer 10

Modern x86 processors have a fsincos instruction which will do exactly what you're asking - calculate sin and cos at the same time. A good optimizing compiler should detect code which calculates sin and cos for the same value and use the fsincos command to execute this.

It took some twiddling of compiler flags for this to work, but:

$ gcc --version
i686-apple-darwin9-gcc-4.0.1 (GCC) 4.0.1 (Apple Inc. build 5488)
Copyright (C) 2005 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

$ cat main.c
#include <math.h> 

struct Sin_cos {double sin; double cos;};

struct Sin_cos fsincos(double val) {
  struct Sin_cos r;
  r.sin = sin(val);
  r.cos = cos(val);
  return r;
}

$ gcc -c -S -O3 -ffast-math -mfpmath=387 main.c -o main.s

$ cat main.s
    .text
    .align 4,0x90
.globl _fsincos
_fsincos:
    pushl   %ebp
    movl    %esp, %ebp
    fldl    12(%ebp)
    fsincos
    movl    8(%ebp), %eax
    fstpl   8(%eax)
    fstpl   (%eax)
    leave
    ret $4
    .subsections_via_symbols

Tada, it uses the fsincos instruction!

Answer 11

If you are willing to use a commercial product, and are calculating a number of sin/cos calculations at the same time (so you can use vectored functions), you should check out Intel's Math Kernel Library.

It has a sincos function

According to that documentation, it averages 13.08 clocks/element on core 2 duo in high accuracy mode, which i think will be even faster than fsincos.

Answer 12

If you use the GNU C library, then you can do:

#define _GNU_SOURCE
#include <math.h>

and you will get declarations of the sincos(), sincosf() and sincosl() functions that calculate both values together - presumably in the fastest way for your target architecture.

Answer 13

In old assembly days, I simply store the sin/cos in two 256 arrays.

This is the fastest way if you don't care about the precision too much.

Answer 14

Many C math libraries, as caf indicates, already have sincos(). The notable exception is MSVC.

• Sun has had sincos() since at least 1987 (twenty-three years; I have a hard-copy man page)
• HPUX 11 had it in 1997 (but isn't in HPUX 10.20)
• Added to glibc in version 2.1 (Feb 1999)
• Became a built-in in gcc 3.4 (2004), __builtin_sincos().

And regarding look-up, Eric S. Raymond in the Art of Unix Programming (2004) (Chapter 12) says explicitly this a Bad Idea (at the present moment in time):

"Another example is precomputing small tables--for example, a table of sin(x) by degree for optimizing rotations in a 3D graphics engine will take 365 × 4 bytes on a modern machine. Before processors got enough faster than memory to demand caching, this was an obvious speed optimization. Nowadays it may be faster to recompute each time rather than pay for the percentage of additional cache misses caused by the table.

"But in the future, this might turn around again as caches grow larger. More generally, many optimizations are temporary and can easily turn into pessimizations as cost ratios change. The only way to know is to measure and see." (from the Art of Unix Programming)

But, judging from the discussion above, not everyone agrees.

Answer 15

I have posted a solution involving inline ARM assembly capable of computing both the sine and cosine of two angles at a time here: Fast sine/cosine for ARMv7+NEON

Answer 16

This article shows how to construct a parabolic algorithm that generates both the sine and the cosine:

DSP Trick: Simultaneous Parabolic Approximation of Sin and Cos

http://www.dspguru.com/dsp/tricks/parabolic-approximation-of-sin-and-cos