Multimedia Extensions For General

1
Multimedia Extensions For General Purpose
Microprocessors
Theory and Comparative Evaluation
Abdullah Aldahlawi Dept. of Elect. Computer
Engineering The George Washington
University Dahlawi_at_ieee.org
2
Multimedia Extensions For General Purpose
Microprocessors
Three Microprocessor Families

• HP PA-RISC MAX (Multimedia Acceleration
  eXtension)
• Sun Ultra SPARC VIS (Visual Instruction Set)
• 3. Intel Pentium MMX (MultiMedia eXtension)

3
Contents

• Introduction
• Subword Parallelism
• Arithmetic Instructions
• 3.1 Add Subtract
• 3.2 Parallel Overflow and Saturation
  Arithmetic
• 3.3 Multiply
• Data Rearrangement Instructions
• 4.1 Permute
• 4.2 Mix
• Formatting Instructions
• Conditional Instructions
• Complex Instructions
• Memory Instructions
• Conclusion

4
Introduction

• Multimedia is an Integration Of Visual and Audio
  Information.
• Multimedia Consume Huge Storage and Computation
  Resources.
• - 24 fps ? MPEG Compression /
  Decompression
• 3. Special Type of Instructions are Needed to
  Speed up Compression and Decompression ? High
  Performance ? High Quality Video
• Most Microprocessors has extended their
  Instruction Set Architecture to with Multimedia
  Instruction
• HP PA-RISC MAX (Multimedia Acceleration
  eXtension)
• Sun Ultra SPARC VIS (Visual Instruction Set)
• Intel Pentium MMX (MultiMedia eXtension)

• First implemented 1994
• MPGE 24fps decoder
• on 80 Mhz WS

5
Subword Parallelism Fundamental Concept

• Pixel Oriented Media Processing Programs Show
  High Degree of Parallelism of Low-Precision
  Data.
• 2. Pixels Are Input in or Output from
  Computation as 16 bits Units.
• Parallelism Could be Achieved by Packing These
  units in 64-Bits Words and Process them in
  Parallel (Subwording).
• Multimedia Extended Instructions Take Advantage
  of Subwording and process These Subwords in
  Parallel (Similar to SIMD).
• The Performance gain is obvious, 64-bits Integer
  Adder Can be Divided into Four 16-Bits Subword
  Integer Adder ? 4 16-Bits Integer Operation can
  be Executed in Single Processor Cycle.
• Most Multimedia Instructions Extension are
  Designed to Execute Parallel Subwords in order to
  Enhanced the Multimedia Processing Performance.

6
(No Transcript)
7
Subword Arithmetic Instructions Add Multiply

• It is know that most frequent operations in
  Multimedia programs are Add and Subtract.
• 2. Multimedia Addition and Subtraction are
  usually 8 or 16 bits add Sub.
• Padd Psub are are the Extended instructions
  for addition and subtraction based on subword
  parallelism concept .
• MAX-2 Support 16-bits Padd Psub, VIS support
  16,32 and MMX support 8,16,32.

8
Parallel Overflows Saturation Arithmetic

• Key difference in multimedia computation is
  handling of overflow.
• Overflow is very likely because of small subword
  size ? inefficient handling of overflow ? effect
  quality of Multimedia data .
• Overflow can be handled by two ways ..
• - Modulo Arithmetic ? (2506) mod 256 ? 0
  , Dangerous !!.
• - Saturation Arithmetic ? (2506) ? 255
• MAX and MMX use Modulo arithmetic and saturation
  arithmetic.
• VIS uses only modulo arithmetic ? Disadvantage
  compared to MMX MAX

9
Subword Arithmetic Instructions Multiplication

• Multiply is a very important instruction, great
  difference performance difference can be found
  based on how different architecture implement
  Multiply instruction.
• Multiply can cause an overflow and takes much
  time to execute.
• VIS and MMX implement 16 bits H/W for
  multiplication (Pmul)
• VIS try to avoid overflow by allowing ONLY 168
  bits multiplications, it needs 3 instructions to
  multiply 1616 bits.
• MMX has parallel subword multiply and accumulate
  instruction (Pmadd)
• MAX support multiplication without adding H/W
  ..
• - two 32-bits FP multiply accumulate
  FMAC units ? Variable Operands ? 4
• instruction / Cycle
• - Integer Adder PshLadd PshRadd is
  used ? Constant Operands ?

? MMX Pmadd
10
Data Rearrangement Instructions

• It is sometimes required to change the position
  of subwords during multimedia computation.
• Permute is used in MAX-2 to change the position
  of subwords in the register.
• VIS and MMX do NOT have permute, they use other
  instruction to perform the same function.
• MixL MixR are used in MAX-2 to mix subwords
  from two registers and place the result into
  third register .
• MixL interleaves subwords starting from left most
  subwords in the source
• MixR interleaves subwords starting from right
  most subwords in the source .

MixL MixR ?
? Permute
11
Formatting Instructions

• Level of precision during Multimedia computation
  may change very often .
• Formatting instructions are used to convert one
  subword size to another.
• UnPack instruction is used to expand smaller
  subwords into large ones.
• Pack instruction is used to reduce large subwords
  into smaller ones.
• MMX could unpack 8 ? 16, 16?32, 32?64 subwords..
• MMX also pack 32 ? 16, 16 ? 8 (saturates).
• MAX uses PshR, Pshl and Mix Instructions to pack
  and unpack Subwords
• VIS Could Pack 32?16, 16?8, 32?8 subwords

Max VIS MMX
12
Conditional Instructions

• Parallel Subword Compare (Pcmp) instruction
  perform the test by comparing pair of subword in
  two source registers and
• - VIS generate one bit true of false
  indicator
• - MMX generate full mask 1s or 0s
  indicator
• The bit-mask instruction is used in VIS
  multimedia instruction set extension to control
  the Partial Store instruction (playing with
  colors)
• MAX do NOT need Partial Store , its architecture
  already has Store Byte instruction
• MMX has extension for Pcmp instructions

Max VIS MMX
13
Complex Instructions

• The sum of absolute differences, which is used in
  MPEG video encoding, is a complex operation that
  might effect the performance if not processed
  efficiently
• The Pdist instruction in VIS multimedia
  instruction set extension is used to calculate
  the sum of absolute difference (needed in MPEF
  encoding)
• MMX and MAX use generic Padd and Psub
  instructions to calculate the absolute
  differences which might requires more cycles,
  this gives an advantage to VIS over MAX and MMX
  in term of handling the same MPEG encoding
  computation.

14
Memory Instructions

• Most multimedia processing computations have very
  predictable memory access pattern (nature of
  Multimedia data)
• Multimedia programs benefit from the use of
  prefetch instruction in order to reduce the cache
  misses
• MAX has prefetch to cache instruction that would
  fetch the addressed word to cache
• VIS has a special purpose instruction for loading
  a block of eight registers (block load)

15
Conclusion

• Multimedia processing is based on the fundamental
  concept of subwording
• MAX, MMX and VIS all implement extended
  instruction for Multimedia processing.
• MMX and VIS have Parallel integer functional
  units , MAX does NOT.
• Some Multimedia instructions has multiple usage ,
  Mix ? Pack Mix
• The general purpose microprocessor with
  multimedia extension usually process difficult
  multimedia processing task.
• Future multimedia instructions set architecture
  are likely .
• The demand for high performance Multimedia
  computing will be required

End of Presentation