Debian on ARM

Wookey & Hector Oron

Fosdem, February 2011

What we’ll tell you about

• Nomenclature
• ABIs
• Debian ports
• New armhf port
• Future developments

Nomenclature

• ARM CPU numbers

  • ARM2-ARM11,
  • Cortex-A/R/M series A8, A9, A15

• ARM ISA names: armv1-armv7

  • ARM + extensions
  • FPA, jazelle, thumb, thumb2, VFPv1–3, Neon
    
  • Thumb (introduced in v4t) is partial 16-bit ISA
  • Thumb2 (introduced in v7) is complete ISA

What’s an ABI?

• Endianness, Architecture, C-library calling convention
• Not instruction set within arch
• Not CPU optimisations

Floating point units

• FPA (1996) - mixed endianness values
• iwMMXt (Intel, 2004)
• Maverick crunch (Cirrus, 2004)
• VFP option since v5, now standard
• Neon

Floating point calling convention

• Using FPU normally is best ‘hard-float’
• But emulation is slow, instruction trapping extremely slow

• Soft-float has small overhead due to register transfer

  • Hard-float: FP used, FP arguments are passed directly in FPU registers
  • Softfp: FP used, FP arguments passed in integer registers
  • Soft: FP instructions not used, emulation compiled in.

FP Calling convention options

• VFPv3_D16 available on all v7 CPUs
• VFPv3_D32 on some A8/A9s (v7)
• half-precision VFPv3-f16 on some v7 CPUs
• Neon support uses the FP registers, thus compatible

Endianness

• All arms since ARM7 (armv3) switchable
• Only really used on IXP network processor series
• Not much use anymore as there is a swap instruction
• Big-endian discouraged

Debian ‘arm’ port

• ‘arm’ architecture added to Debian GNU/Linux 2.2 (Potato) in 2000

• End of life after Debian GNU 6.0 (Squeeze) gets released.

  • TODAY!
• Supported armv3 instruction set, later changed to v4

• Old ABI (APCS), FPA/FPE FPU

Debian ‘armel’ port

• ‘armel’ arch added to Debian GNU/Linux 5.0 (Lenny) in 2008
• Change the ABI: to ‘Embedded’ ABI ‘EABI’
• Built for armv4t instruction set
• ‘Soft-float’ for FPU ops - works with/without FPU
  • http://wiki.debian.org/ArmEabiPort
  • There are no plans to deprecate it.
• Ubuntu armel is different: v7, thumb2, softfp

Debian ‘armeb’ port

• ‘armeb’ has never been a Debian official architecture
• Was done for the NSLU2 due to proprietary big-endian ethernet driver
• Nobody currently working on it
• Old ABI, not EABI

Debian ‘armhf’ port

• Why yet another ARM architecture?

  • Thumb2 is faster
  • Hard-float is faster
  • v7 optimisations - Current netbooks, nettops, mobile devices
  • http://wiki.debian.org/ArmHardFloatPort

Debian ‘armhf’ performance

• Vorbisgain test by Clint Adams

  vgsoft() {
    su oem -c 'time vorbisgain -q -a -r ~/armhf.chroot/tmp/gaintest/*'
  }
  
  vghard() {
    chroot ~/armhf.chroot su oem -c 'time vorbisgain -q -a -r /tmp/gaintest/*'
  }
  

Debian ‘armhf’ performance

• Debian armel (‘soft-float’) vs Debian/Ubuntu (‘softfp’)

Sat Aug 21 22:46:06 CDT 2010 #1S
real    15m54.630s
user    15m52.050s
sys 0m2.370s

Sat Aug 21 23:02:01 CDT 2010 #1SFP
real    9m55.798s
user    9m53.000s
sys 0m2.490s

Debian ‘armhf’ performance

• Debian armel (‘soft-float’) vs Debian armhf (‘hard-float’)

Tue Aug 10 17:20:31 CDT 2010 #1S
real    16m41.598s
user    15m57.740s
sys     0m2.640s

Tue Aug 10 17:37:12 CDT 2010 #1H
real    7m47.515s
user    7m24.980s
sys     0m2.150s

Debian ‘armhf’ performance

• Debian armel (‘soft-float’) vs Debian/Ubuntu (‘softfp’)

softfp:
$ glxgears
80 frames in 5.0 seconds
119 frames in 5.0 seconds
118 frames in 5.0 seconds
120 frames in 5.0 seconds 

hardfp:
136 frames in 5.0 seconds
144 frames in 5.0 seconds
143 frames in 5.0 seconds
139 frames in 5.0 seconds 

Debian ‘armhf’ performance

• armv7, hard-float (VFP_D16), thumb2
• Up to 40% faster (povray)
• Mesa (20%)
• cfft (30%)

Debian ‘armhf’ performance

• Vorbis tests

  • http://people.debian.org/~schizo/tmp/armhf

• Konstantinos Magaritis’ tests:

  • http://www.powerdeveloper.org/forums/viewtopic.php?p=13609
  • http://svn.arhuaco.org/svn/src/emqbit/tools/emqbit-bench/
  • 30%–40% better performance on hard-float if uses floating point
  • 1% better performance on hard-float if does not use floating point

Debian Ports stats

Debian Ports status

• ‘armhf’ is the green line

Debian armhf TODO

• TODO:

  • http://wiki.debian.org/ArmHardFloatTodo
• Mass bug filing
• Sprints:
  • February: Cambridge, UK
  • April: San Antonio, TX

Optimizations and subarches

• Can optimise for CPU within ABI (v5,v6,v7, cortex-A8, A9)

• And add NEON support for multimedia

  • Only needed for a few packages: e.g. codecs
  • Currently done in package namespace: e.g mplayer-i686, libc6-amd64

• Better to be able to rebuild subset of packages for new ‘flavour’

  • GCC has support for this in linker ‘multilib’
  • see http://wiki.debian.org/Multiarch/Spec

Sponsors

Thanks and comments

• Thanks to all people involved:

  • Konstantinos Margaritis (Genesi-USA)
  • Debian-ARM and lots of individuals
  • Linaro GCC Toolchain people

• Comments:

Remember to have fun!

Remember to have fun