[tlbuild] ARM versions

[Johannes Hielscher]

Am Sat, 29 Apr 2017 07:57:24 +0100
schrieb Simon Dales <simon at getThingsFixed.co.uk>:

> Anybody know about the relative performances between different ARM
> binaries? WRT TeX.
You can find some chatter about different performance due to Thumb 1/2
and NEON out in the net, mainly regarding floating-point calculations.
Hence mostly inconclusive for TeX.

> 
> I wish to retain backward compatibility so my binaries are for
> Raspberry Pi version 1; which is armv6.
> 
> The RPi 2 and 3 have v7 chips. Eventually a RPi 4 with some other arm
> chip.
> 
> How inefficiently does an armv6 binary run on a v7 machine relative to
> native v7?

Out of curiosity, I quickly tested this on my Archlinux armv7h, as
in their repos, they have both armv7h and armv6h binaries for TL2016.

Indeed I have seen a (somewhat minor) lead for armv7h, being 5% faster
(4.48(6) s instead of armv6h's 4.72(3) s, average over five runs) for
some random TeX file. This can be due to *anything* including alarm's
own compile toolchain magic, my CPU or whatever.
> 
> So is there a case for building multiple binaries, like we do for the
> Intels?

Decide yourself if this effect is worth further studies (and maybe
test it on a spare SD card with Archlinux ARM, where the old
architecture is just a deliberately wrong Architecture field
in /etc/pacman.conf away [you don't have heard this trick from me...]).

At the end, serious TeXing on RPis will likely notice their limited
RAM, so (depending on what you intend to do) the speed loss is mostly
irrelevant for small documents, and for large ones merely about running
out of memory some seconds later...
> 
> /////////////////
> 
> Simon
> 

Best,
Johannes