Improving the fast inverse square root (2010)

#1

Post by cpdt » Mon, Jul 09, 2018, 2:06 AM UTC

Improving the fast inverse square root (2010)
rrrola.wz.cz

Re: Improving the fast inverse square root (2010)

#2

Post by oranlooney » Mon, Jul 09, 2018, 3:29 AM UTC

That's great and all, but nobody needs a 32-bit anything in 2018. This undergraduate paper provides a magic number and associated error bound for 64-bit doubles:

https://cs.uwaterloo.ca/~m32rober/rsqrt.pdf

Re: Improving the fast inverse square root (2010)

#3

Post by minton » Mon, Jul 09, 2018, 3:40 AM UTC

oranlooney wrote: post #2

That's great and all, but nobody needs a 32-bit anything in 2018. This undergraduate paper provides a magic number and associated error bound for 64-bit doubles: https://cs.uwaterloo.ca/~m32rober/rsqrt.pdf

I think this article is from 2010.

Re: Improving the fast inverse square root (2010)

#4

Post by dleslie » Mon, Jul 09, 2018, 3:46 AM UTC

oranlooney wrote: post #2

That's great and all, but nobody needs a 32-bit anything in 2018. This undergraduate paper provides a magic number and associated error bound for 64-bit doubles: https://cs.uwaterloo.ca/~m32rober/rsqrt.pdf

TIL, no one in the gaming industry uses 32 bit floats any longer. /s

Re: Improving the fast inverse square root (2010)

#5

Post by bananaboy » Mon, Jul 09, 2018, 3:50 AM UTC

oranlooney wrote: post #2

That's great and all, but nobody needs a 32-bit anything in 2018. This undergraduate paper provides a magic number and associated error bound for 64-bit doubles: https://cs.uwaterloo.ca/~m32rober/rsqrt.pdf

This is not true. In games 32-bit floats are extremely common.

Re: Improving the fast inverse square root (2010)

#6

Post by oppositelock » Mon, Jul 09, 2018, 4:02 AM UTC

oranlooney wrote: post #2

That's great and all, but nobody needs a 32-bit anything in 2018. This undergraduate paper provides a magic number and associated error bound for 64-bit doubles: https://cs.uwaterloo.ca/~m32rober/rsqrt.pdf

Realtime 3D still uses floats, but only when we can afford something so big, s10e5 is better where available.

Re: Improving the fast inverse square root (2010)

#7

Post by perfmode » Mon, Jul 09, 2018, 4:07 AM UTC

oranlooney wrote: post #2

That's great and all, but nobody needs a 32-bit anything in 2018. This undergraduate paper provides a magic number and associated error bound for 64-bit doubles: https://cs.uwaterloo.ca/~m32rober/rsqrt.pdf

deep learning uses low precision floats

sometimes as few as 8 bits are needed

Re: Improving the fast inverse square root (2010)

#8

Post by dagenix » Mon, Jul 09, 2018, 4:16 AM UTC

oranlooney wrote: post #2

That's great and all, but nobody needs a 32-bit anything in 2018. This undergraduate paper provides a magic number and associated error bound for 64-bit doubles: https://cs.uwaterloo.ca/~m32rober/rsqrt.pdf

That's not really accurate. Even in cases were 32 bit and 64 bit operations are equally fast on the CPU, 32 bit values still take up half the memory. For many workloads, the limiting factor is cache space. So, if you can use 32 but values, you can get much better performance for those workloads.

Re: Improving the fast inverse square root (2010)

#9

Post by simonbyrne » Mon, Jul 09, 2018, 4:19 AM UTC

It is worth noting that with AVX-512, Intel has introduced a native inverse sqrt approximation (VRSQRT14).

Re: Improving the fast inverse square root (2010)

#10

Post by dnautics » Mon, Jul 09, 2018, 4:22 AM UTC

oranlooney wrote: post #2

That's great and all, but nobody needs a 32-bit anything in 2018. This undergraduate paper provides a magic number and associated error bound for 64-bit doubles: https://cs.uwaterloo.ca/~m32rober/rsqrt.pdf

Even scientific calculation would be fine with 32 bit floats, but average floating point error due to representation creeps with ON (iirc) over N multiplications, so you have to use 64 bit for many scientific applications to get satisfactory results after a million or a trillion multiplications.