Multimedia Processor Design

1
Multimedia Processor Design

• -An Overview of Design Issues

2
Discussion ..

• Multimedia in general
• Approaches in designing SIMD and
• VLIW
• Evaluation methods and Parameters
• A focus on SIMD
• Conclusion

3
Why Multimedia Processors?

• Video Audio Compression
• Image Processing 3D graphics
• Signal Processing
• Communication Networking and ,
• World Wide Web

4
Features of Multimedia Processing..

• Heavy Computation
• High Memory Bandwidth
• Real Time Processing Video Conferencing
• Inherent Data and Instruction Parallelism
• A lot of execution time spent in small loops

5
Approaches..

• SIMD Single Instruction Multiple Data
• VLIW Very Long Instruction Word
• ASICS Application Specific Ics.

6
SIMD?

• Single Instruction Multiple Data
• Exploitation of Data Parallelism
• Multimedia Extensions to GPPs
• MMX for Pentium II Processor
• Similar operation on different data
  Simultaneously

7
SIMD .What is the idea?

• Multimedia processing involves smaller data types
  (typically 8 or 16 bit)
• Packing SMALLER data elements in
• WIDER data paths (typically 32 or 64
• bit)

8
ADD Instruction in SIMD4 way Parallelism in
64 bit SIMD
9
Another Example..
10
VLIW .Very Long Instruction Word

• Exploits Instruction Parallelism
• Multiple Functional Units in Data Path
• A Single instruction triggers Multiple
  Operations
• Dedicated Multimedia Processors
• SONY Play station 2 3D gaming

11
What a Lengthy Instruction does?
12
ASICs

• Application Specific Integrated Circuits
• Highly Optimized
• Reduced Flexibility
• Ex ADV- JP2000 Wavelet Processor Entropy
  Codec
• Not in much use compared to SIMD and
• VLIW

13
Parameters for Evaluation

• Execution time Speed Up
• Performance improvement of technique x over y
  is
• Execution time of y
• Execution time of x

14
Parameters (contd..)

• IPC Instructions retired per cycle
• Frequency of Operations
• Effective Branch prediction
• Cache hit rates
• Data size

15
Parameters (contd)

• Resource and Instruction stream stalls
• Average Basic Block size local parallelism
• Floating point operations

16
Testing Methodologies

• How to evaluate performance?
• Benchmarks
• Standard Multimedia Applications
• Kernels
• Kernels are operations that dominate
• overall processing time

17

• Kernels
• Dot-product, matrix-vector products
• FIR and IIR filters
• FFT and DCT
• Applications
• Speech compression algorithms
• Image Processing applications (JPEG)
• 3D graphics - Quake II
• Video processing (MPEG) - Real Video

18
Bottlenecks in SIMD.

• Most of the DLP in SIMD lies in data present
  within nested loops.
• As an example of data in multimedia processing
  consider this..

19
A 2D Data structure in multimedia..
20
The problem..

• DLP in the outer loops is not exploited
• Hardware to generate multiple address sequences
  is not complicated, but supporting them using
  general-purpose instruction sets is difficult due
  to limitation of addressing modes.
• Not enough support for keeping track of multiple
  indices/strides efficiently in GPPs.

21
Instructions of two types..

• Computationally Useful Instructions
• EX MAC used in the algorithm
• Overhead instructions
• Address generation
• Branch
• Address transformation
• Load store

22
What dominates??

• Overhead instructions consume 75 of
• execution time!!!!
• Peak throughput of SIMD is not achieved.
• Hardware to process Overhead Instructions has to
  be scaled.

23
A Solution..

• Scaling SIMD units will not help
• More Parallelism is extracted by scaling
• NON-SIMD part of the architecture..
• Dedicated architecture like Media Breeze along
  with the SIMD will yield good results.

24
An important issue

• Architecture and compiler should complement each
  other
• Compiler should not saturate certain
  architectural resources while rarely using other.

25
A sample analysisRatio of Execution Times
26
VLIW or SIMD ??

• Either processor is not a replacement for the
  other
• SIMD and VLIW and
• not SIMD vs VLIW!
• Infact both can be implemented on the same
  processor. Ex ADIs TigerSharc

27

• Thank You