Multimedia Support - PowerPoint PPT Presentation

About This Presentation
Title:

Multimedia Support

Description:

... x 242 x 24 bit MPEG video requires a ... Product of two 8-bit numbers is a 16-bit number ... provide hardware to operate on both 32-bit components in parallel ... – PowerPoint PPT presentation

Number of Views:18
Avg rating:3.0/5.0
Slides: 12
Provided by: bar72
Category:

less

Transcript and Presenter's Notes

Title: Multimedia Support


1
Multimedia Support
Carnegie Mellon University Computer Science
Department Computer Architecture F01
v
  • Jernej Barbic
  • Overview of multimedia and processors
  • Luis von Ahn
  • Altivec

2
Multimedia means
Creating, displaying, playing, manipulating,
compressing, decompressing
  • Still images,
  • Audio,
  • Video.

Each of these require different type of CPU
operations.
3
Typical multimedia formats
  • JPEG
  • MP3
  • MPEG
  • MPEG-4 just released
  • Decompression significantly slower than
    compression
  • DVD

4
Multimedia operations are a bottleneck
In this presentation, we focus on this
  • CPUs cannot compute fast enough
  • Memory/disk latencies too large
  • Network bandwidth too low
  • Playing 352 x 242 x 24 bit MPEG video requires a
    T1 line (1.5 Mbps)

5
The majority of CPU multimedia operations are
  • Matrix multiplications, inner products of vectors
  • FFT, DCT (discrete cosine transformation)
  • JPEG, MPEG, MP3
  • Multiply-accumulate instructions
  • Calculating the sum of several products (e.g. in
    an inner product)
  • Overflow problems
  • Product of two 8-bit numbers is a 16-bit number
  • Some multimedia architectures incorporate special
    hardware to deal with overflows (saturated
    arithmetics)
  • Single precision floating point operations
  • Single precision 32-bit
  • geometry 3D processing

6
The majority of CPU multimedia operations are
(contd.)
  • Calculating the sum of absolute differences
  • MPEG compression detecting motion
  • Implemented in hardware only in Alphas MVI

detecting motion
7
Multimedia features of modern processors
  • PS (paired single) floating point (fp) format
  • Problem typical fp registers are 64-bit, but a
    lot of data is 32-bit
  • Solution
  • pack two 32-bit single precision fp numbers into
    a 64-bit register
  • provide hardware to operate on both 32-bit
    components in parallel
  • Many geometric operations require only single fp
    precision
  • Useful also for non-multimedia tasks (scientific
    computing)
  • Using PS increases performance by a factor of two

8
Multimedia features of modern processors (contd.)
  • SIMD (single instruction, multiple data)
  • SIMD is the integer version of PS
  • Divide 64-bit integer registers into eight bytes
    (8 bit) or four half-words (16 bit)
  • Operate on this data in parallel
  • Parallel arithmetic instructions
  • Instructions to store, load, rearrange individual
    bytes

9
Multimedia architectures
  • Alpha MVI (motion-video instructions)
    multimedia extension
  • Alphas are fast and dont need extensive
    multimedia additions
  • Fast video (MPEG) compression enabled in MVI
  • Limited SIMD support, little parallel arithmetic
    instructions,
  • but already fast without them.
  • Special motion detecting instruction
  • PERR Ra, Rb, Rc
  • Calculates

10
Multimedia architectures (contd.)
  • MIPS V
  • Fully support SIMD and PS
  • Instructions add little to the complexity of
    circuitry
  • MIPS MDMX
  • More extensive than MIPS V
  • Supports 32 media registers
  • Media registers mapped to 64-bit floating point
    registers
  • SIMD formats oct-byte (8x 8-bit), quad-half (4x
    16-bit)
  • Saturated arithmetics avoids underflows and
    overflows
  • Wide 192-bit accumulator to support extensive
    multiply-accumulate instructions

11
Multimedia architectures (contd.)
  • HP
  • Intel
  • MMX
  • Sun
  • VIS (visual instruction set)
  • AltiVec
Write a Comment
User Comments (0)
About PowerShow.com