AMD/aocl-libm-ose: AMD libm | Hacker News Reader

AMD/aocl-libm-ose: AMD libm

github.com

27 pointslelf14 comments

Open articleView on HN

Here's a better description: https://github.com/amd/aocl-libm-ose/blob/master/docs/Releas...

Here's a page to download Linux binaries:

https://developer.amd.com/amd-aocl/amd-math-library-libm/

The license there looks much less friendly than the 3-clause BSD on Github.

I was unable to build this on macOS 10.13, because "aliases are not supported on darwin".

I'm curious to see the tests, which have not been checked in yet. There was a story a month ago about the Android calculator's library, and how it is used to test java.lang.Math:

https://news.ycombinator.com/item?id=24700705

I wonder what this library's accuracy tests look like.

I'm also curious, of course, to see what kind of performance difference there is. An old post by Agner Fog shows an 85% reduction in run time relative to glibc on AMD hardware:

https://www.agner.org/optimize/blog/read.php?i=209

What is this? The readme is sparse.

Presumably an implementation of the C math library based on the name (libm).

An implementation of the “core” math function of the C standard library, colloquially known as libm.

Think implementing cos(x) using only pure C (so essentially addition, subtraction, multiplication and division of floats).

It actually contains quite a lot of asm, for example this for avx2-capable processors

https://github.com/amd/aocl-libm-ose/tree/master/src/isa/avx...

i can't imagine myself writing this kidn of assembly code. is this generated code?

I'm going to guess no. There are lots of comments that provide context for the various statements, and there are lots of commented out lines that are probably from prior iterations. Additionally the formatting is not consistent (aka human like). I've definitely written vectorized ASM that looks like the before.

i dont work with asm so pls forgive me. how can one write sth like this?

``` DQ 03ff0842108421084h DQ 03ff07b9f29b8eae2h DQ 03ff073260a47f7c6h DQ 03ff06ab59c7912fbh DQ 03ff0624dd2f1a9fch DQ 03ff059eea0727586h DQ 03ff05197f7d73404h DQ 03ff04949cc1664c5h DQ 03ff0410410410410h DQ 03ff038c6b78247fch DQ 03ff03091b51f5e1ah ```

These are tables of values used for the calculations. DQ stands for define quadword. The tables were generated in another program and copied and pasted in here.

Thank you

Can anyone say what the meaning of the prefixes are:

vrd2, vrda, vrs4, vrsa

Are these different vector lengths?

Unfortunately no references or details are provided for the implementations that I can see.

I concur

Reply on news.ycombinator.com