Re: [gentoo-user] Binary packages for a different amd64 flavor

[David Haller]

Hello,

On Tue, 12 Jun 2018, Mick wrote:
>On Tuesday, 12 June 2018 19:31:32 BST Ian Zimmerman wrote:
>> I have had it with compiling stuff from source on my laptop.  It is just
>> too slow.  So I would like to create binary packages on my desktop and
>> then just tell the laptop to use them.
>>
>> Simple enough, except that the desktop is AMD Phenom, and the laptop is
>> Intel 64 bit Atom.  Up to now, each system had unique CFLAGS to squeeze
>> as much performance as possible.
>>
>> On the desktop:
>> CFLAGS="-march=barcelona --param l1-cache-size=64 --param
>>  l1-cache-line-size=64 --param l2-cache-size=512 -O2 -pipe"
>>
>> On the laptop:
>> CFLAGS="-march=ivybridge --param l1-cache-size=32 --param
>>  l1-cache-line-size=64 --param l2-cache-size=4096 -O2 -pipe"

First of all, I've seen quite a big boost just by using '-O3'!

My speeding-up-relevant flags (for a Athlon II X2 250 "Regor") are:

-O3 -march=native -mtune=native -mfpmath=sse -msse4a \
    -funwind-tables -fasynchronous-unwind-tables

You'll have to adjust -march=/-mtune= for the Atom and -m* for both,
crosscheck with e.g. 

$ grep ^flags /proc/cpuinfo |tr ' ' '\n' |grep sse | sort -u
misalignsse
sse
sse2
sse4a

or app-portage/cpuid2cpuflags.

>> I don't want to give up these tunings, but from the wiki page [1] I can
>> see no straightforward way to have different CFLAGS when compiling binary
>> packages, from the normal CFLAGS when installing directly from source on
>> the host system.  Is the only way of doing this to set up a full-blown
>> cross-development environment?
>>
>> [1]
>> https://wiki.gentoo.org/wiki/Binary_package_guide
>
>On the desktop you could just use -march=native for its own compiles, not sure
>if there is a benefit or good reason to use '-march=barcelona', but I'm
>digressing.
>
>If you are compiling binary packages for the laptop with a single stanza on
>the CLI, then you can run:
>
>CFLAGS="-march=ivybridge ..." CXXFLAGS="-march=ivybridge ..."
>FEATURES="buildpkg" PKGDIR="/tmp/binpkg_dir emerge -uaNDv --buildpkg world
>
>I understand you will need the complete CFLAGS & CXXFLAGS for the guest's
>hardware - others should confirm if this is so.

Depends on how far you want to go on optimization ;)

>I find it neater/easier to copy the guest's fs over to the faster host, then
>chroot into it, sync portage and emerge with --buildpkg world.  There are
>other solutions, NFS mounts of the guest over the network, using a VM
>mirroring the laptop build on the host, but they are more complicated for my
>use case of a single guest.

And/or setup some wrappers/shell-functions, that setup flags like:

==== emerge-atom ===
#!/bin/sh
FEATURES="buildpkg"
PKGDIR="${PORTDIR}/laptop_binpkg_dir"
export CFLAGS="..."
export CXXFLAGS="..."
...
emerge "$@"
====

and then use 'emerge ....' for the desktop and 'emerge-atom' for the
laptop. Or something along those lines.

On the other paw, one could run some benchmarks, comparing both native
flags vs. the "common demnominator" (and -O2 vs -O3 ;)...

HTH,
-dnh

-- 
I like offending people, because I think people that get offended
should be offended.     -- Linus Torvalds, June 2012 at Aalto Uni