Power analysis of H.264/AVCfor mobile platforms

Student: Alexandru IOVANOVICI, B.Eng.Supervisor: Lucian PRODAN, Ph.D, B.Eng.

Master’s ThesisJune 2011

What ... is the problem ?

1. Customers are mobile

2. Customers expect PC-like featuresfrom their „cell-phones”

What ... solutions do we have ?

CODECs

Video CODECs basics

RAW video -> minimum 216Mbps [Richardson2010] Lossless compression = removing data redundancy

(3-4 times less space); Lossy compression = removing subjective

redundancy; much higher compression rates;

Video ENCODER basics

Prediction model:exploiting the similarities between neighbouring video frames;

Spatial model: compressing the residual; transform coeficients;

Entropy encoder:removing statistical redundancy;

Typical ENCODER architecture

DPCM/DCT architecture: basis for all modern encoders, including H.264;

Typical DECODER architecture

The H.264/AVC

H.264/AVC gives specifications only for the decoder; Published by ITU in 2003 with several revisions; Based on MPEG-4 Visual; 10 to 50 times better compression ratio at same visual quality

[Richardson2010, Xe2007] 2 to 10 times more power dissipation than MPEG-4 [Xe2007]

The H.264 ENCODER

Transform and quantization

H.264 Profiles

There is a great flexibility in choosing the apropriate combination of tools that best suit the specific needs of a particular tasks [eg. CAVLC+DCT];

H.264 profiles define a specific set of tools; A profile compliant decoder must be able to

decode with all the tools in that profile → constraints on the capabilities required by a decoder;

The Zexia encoderNon-interlaced Base

Profile;

Only OpenSource available encoder for H.264;

Modular, interface-based;

VHDL configuration-based;

Better suited for Spartan 3 than for Cyclone III: Spartan has more on-chip memory;

Two Clock lines;

RAM: an entire image must be loaded by the external controller; and the reference image too; Smaller reqirements if we use intraprediction only

[Richardson2007]; Predictions components:

SAD comparaison; Only p-frames prediction (licensing issues);

The Zexia encoder

Simulation based – lack of hardware resources on Altera DE2 board;

Quartus II with Advanced PowerPlay Early Estimator;

VCD files based on statistical distribution of line transitions (12.5% [Xe2007]).

Experimental results

Experimental results

Chip flooplan area distribution In number of LEs

Buffer

CAVLC

Core transform

DC transform

Dequantise

Header

Intra 4x4

Invtransform

Quantise

Reconstruction

ToBytes

Total power distribution by functional unit

Buffer

CAVLC

Core transform

DC transform

Dequantise

Header

Intra 4x4

Intra 8x8

Invtransform

Quantise

Reconstruction

ToBytes

Experimental results

FPGAs are bad at low-power optimizations

Augmented Cell PhoneA new architecture

Coprocessor based cell-phone;

“Program repository” on Flash;

Marketplace for “programs”;

High power requirements but even higher customer satisfaction

better user experience;

→ more devices sold;

Conclusions• H.264 is a power intensive algorithm;

• Less than 30% is parallel [52];

• ASIC is the best option but is „expensive”;

• FPGAs are not good for low-power techniques• Large routing grids;

• A lot of cells are powered on even in neutral functions;

• Good for comparing two similar designs in HDL in terms of simulated performance;

• ACP: balance between power requirements and user satisfaction

• Need considerable rethinking on the HW level and the OS;

• Need skilled developers for “soft-coprocessors”