AltiVec Optimizations

1
AltiVec Optimizations
for Linux/PowerPC
2

• What is AltiVec?
• Where can it be used?
• How does it work?
• How is it used?
• Is it worth it?
• Why hasn't it be used so far?
• Can it be integrated into Linux?
• What's Debian's role?
• Altivec coding examples

3

• AltiVec is a SIMD unit for PowerPC CPUs (G4, G5)
• It's similar to other SIMDs (MMX/SSEx/3dNow!/etc).
  
• It's the most complete SIMD implementation
  available
• It doesn't do things faster, but it does more
  things in the same time!
• It's a consistent architecture that just works!
• It has plenty of registers and it uses them
  wisely -)(every instruction can use up to 3
  operands and the result is placed in a separate
  register)

4

• When power consumption is an important factor.
• Where Raw Power is just not enough and other
  methods must be used to achieve good performance.
• In security, using vectorized encryption
  algorithms
• In databases using vectorized hashing
• In video codecs, using vectorized
  encoding/decoding
• In scientific computing (eg. linear algebra,
  bioinformatics)
• Also, generic computing where lots of data
  processing gets done

5
Scalar vs. SIMD
6
How is it used?

• Using C? very easily.
• No need to spend lots of time writing hand
  written asm functions.
• Code vectorization can be tricky but it's not
  that difficult.
• Consider the following simple example in C code

7
Why hasn't it been used more?

• It is used extensively in MacOS X throughout the
  whole OS, and lots of ready made libraries are
  available.
• But It has had very limited adoption in Linux.
• Very few applications in Linux are
  AltiVec-optimized.
• No system-wide AltiVec optimizations.
• On Linux, AltiVec is one of the most underused
  SIMD engines available.

8
Can we benefit from it?

• YES! Thanks to funding from Genesi and support
  from Freescale, a process has started of
  AltiVec-optimizing commonly used Linux
  components.
• The TODO list is long, but we have started with
  the following
• libfreevec GNU libC (functions mem() , str(),
  swab()).
• Hashing Algorithms (Adler32, BDB hashing
  functions)
• Sorting Algorithms (Insertion Sort, Merge Sort,
  Quick Sort)
• TODO C STL, libmcrypt (common encryption
  functions AES, RC5, DES, 3DES, etc), UMAC
  parallelizable hashing algorithm (a replacement
  for MD5), CRC8/16/24/32, etc.

9

• Debian has very good communication channels with
  upstream and an excellent BTS, so the AltiVec
  patches can be pushed upstream fairly quickly.
• Debian is full of knowledgeable (and helping -)
  people that will provide the necessary feedback
  for this effort.
• Debian has a very strong PowerPC community that
  will directly benefit from these optimizations.
  It will also make Debian the de facto
  distribution of choice for PowerPC users.

10
What about Performance?

• memcpy() is 4x faster
• memchr()/strnlen() is up to 10x faster
• memset() is 10x faster
• memfrob() is 24x faster!
• swab() is 11x faster
• hashing algorithms are from 2-7x faster
• Adler32 is 2.5x faster
• Zlib is 25 faster (compression) with only 2/6
  functions vectorized!
• Insertion Sort is from 2-54x(!) faster!

11
Example memcpy()
12
Example Insertion Sort
13
Example Merge Sort
14
More about Sorting

• Integer sorting is used in the kernel
  (filesystems) and in other applications.
• AltiVec Insertion Sort in typical integer sorting
  is 4x faster than the scalar version.
• Insertion Sort is up to 54x faster for char
  sorting than the scalar version!!
• Next to vectorize Quicksort. Estimated
  performance gain 400!

15

• AltiVec is a very useful tool, which until
  recently has been greatly underused
• The future libfreevec 0.8 has been released,
  with 0.9 imminent. The goal is that with 1.0,
  many routines in GNU libC will have been replaced
  by high performance Altivec replacements.
  Further, other libC projects (like dietlibc,
  uClibc, etc) might benefit.
• Other libraries will follow, like zlib, mcrypt,
  C STL, libdv, etc. The list is very long!

16

• For more Altivec information check
• http//arstechnica.com/articles/paedia/cpu/p4andg4
  e2.ars/1
• http//developer.apple.com/hardware/ve/quickstart.
  html
• http//www.penguinppc.org/dev/altivec
• http//arstechnica.com/01q2/p4andg4e/p4andg4e-1.ht
  ml
• http//www.freescale.com/files/32bit/doc/reports_p
  resentations/LNXDEVSYS_PPT1.pdf
• http//www.cs.ucdavis.edu/rogaway/umac/umac_full.
  pdf
• http//www.ppczone.org/
• http//www.freevec.org/ (libfreevec official site)