Transcript Introdution to SSE or How to put your algorithms on steroids!

Introdution to SSE
or
How to put your algorithms on steroids!
Christian Kerl
11.05.2012
1
Outline
•
•
•
•
•
What is SSE?
Basic Operations
Example: Image Pyramid
Summary
Further Resources
11.05.2012
2
What is SSE?
• SSE
• SIMD
= Streaming SIMD Extensions
= Single Instruction, Multiple Data
• Developed by Intel in 1999
• Further extensions SSE2, SSE3 (SSSE3, SSE4)
• Allows parallel processing of multiple integer
or floating point values
11.05.2012
3
What is SSE?
• 8 XMM registers (special CPU registers, 16 on
64 bit)
• Each XMM register is 128 bits wide
– 2 int64 / doubles
– 4 int32 / floats
– 8 int16
– 16 int8
11.05.2012
4
What is SSE?
• Special instructions working on XMM registers
– SSE
70,
SSE2 144, SSE3 13
instructions
• Different instructions for each data type
• Usable in
– Assembly
– C/C++ through SSE “intrinsics”
11.05.2012
5
Basic Operations
•
•
•
•
•
Load / Store
Arithmetic
Comparison / Logical
Type conversion
…
11.05.2012
6
Basic Operations
• Requirements on memory layout for loading
and storing data
• Memory addresses (pointers) need to be 16
byte aligned!
11.05.2012
7
Example: Image Pyramid
• Performance on 2560x1920 image
– Standard C++ version: 7.4 ms
– SSE optimized version: 1.62 ms
=> ≈ 4.5x speedup
11.05.2012
8
Summary
• SSE available on all modern x86 CPUs
• Good for sequential data processing
• Provides considerable speedups (2-4x)
• SSE intrinsic code harder to program and read
=> Use wrapper library, e.g. EasySSE, ut-sse
– Need to evaluate / extend / write one
11.05.2012
9
Further Resources
• Tutorials:
– http://supercomputingblog.com/optimization/getting-started-with-sseprogramming/
– http://www.codeproject.com/Articles/4522/Introduction-to-SSEProgramming
– http://sci.tuomastonteri.fi/programming/sse
• MSDN: good reference manual for intrinsics
– http://msdn.microsoft.com/de-de/library/y0dh78ez
• Wrapper Libraries:
– http://sourceforge.net/projects/easysse/
– http://code.google.com/p/ut-sse/
11.05.2012
10