The current state of increasing PC calculation performance

Ways to improve:

- raise the CPU frequency (on limit)

- add more CPU's (no single thread speed-up)

- add more processor cache (no reason)  

- speed-up I/O, RAM, etc... (on the way)

- code optimization (need human resources)

- something else?

The Free Lunch Is Over

Herb Sutter (December 2004)

What is SIMD?

SIMD is Single Instruction Multiple Data

It's concept of data level parallelism

In other words SIMD is way to manipulate with a big chunk of data using identical instruction from some instruction pull.

SIMD realization on PC

1997y Intel: MMX (Multimedia Extensions) provide in i860

1998y AMD: 3DNow!​ provide in AMD K6-2 [dead from 2010]

1999y Intel: SSE (Streaming SIMD Extensions) in Pentium III

2001y Intel: SSE2 in Pentium IV

2004y Intel: SSE3 in Pentium IV (Prescott revision) 

2005y Intel: SSSE3 (Supplemental SSE3) in 5100 Xeon

2006y Intel: SSE4 in Nehalem-based Core i7

2006-2008y: SSE4a, SSE4.1, SSE4.2, POPCNT, LZCNT

2008y Intel: AVX (Advanced Vector Extensions) but first processor at Q1 2011 (Sandy Bridge)

2013y Intel: AVX2 in Haswell

2013y Intel: FMA1 - FMA4 (Fused Multiply-Add) ​in Haswell

2013y Intel: AVX-512 in Xeon Phi (Knights Landing)

NEON on ARM architecture

Language with SIMD API

• ​C/C++/D
• (CIL) C#  (2015) Mono from 2008
• Java script (2013)
• Dart

Lets make

"hello world"

sum of 4 integers program

Vectors

+

=

64bit

128bit

256bit

512bit

Containt:

• Integer
• Char
• Float
• Double
• Bitmask
• Complex
• Undefine
• ​NaN

Is it really faster?

Time to ask questions

Operation pull

• Arithmetic
• Bit Manipulation
• Cast
• Double
• Compare
• Convert
• Cryptography
• Elementary Math Functions
• General Support
• Load
• Logical
• Mask

• Move
• Miscellaneous
• Random
• Probability/Statistics
• OS-Targeted
• Set
• Shift
• Special Math Functions
• Store
• String Compare
• Swizzle
• Trigonometry

Ways to use it

• Let compiler/interpreter do it
• Use assembler directly
• Use intrinsic

Let's see something interesting

Time to ask questions

Logic operations concept

<

=

OR

Data preparing cost

RGBA to grey strategy

The most important things

• Parallel code != multithreaded code
• SIMD instructions are hardware dependent. But it's not problem
• If Math calculation is bottleneck SIMD can be a solution​

Useful links

• Intel Intrinsics Guide  - list of all possible intrinsics 
• kernel.org SIMD basic tutor - good guide for SIMD beginners
• ​IDF14 - JS demo's repo
• (Auto)Vectorization tutorial with Intel compiler
• google.com - nice service to find SIMD solution and tutorials
• Coders guild repo​ - all examples from this talk 

Time to ask questions