Multimedia Processor: Architectures and Applications

1
Multimedia Processor Architectures and
Applications

• By
• Surin Kittitornkun
• 14 March, 1998

2
Contents

• Programmable Multimedia Processor Why ?
• TMS320C8x C80 C82
• MPACT-R-2
• Targeted Applications
• Application H.324 on TMS320C82
• Current Multimedia Processors
• References

3
Programmable Multimedia Processor Why ?

• More flexibility changes can be made in software
• Less complexity shorter time to market
• More cost efficient than ASIC design
• Requires software development
• May consume more power
• PC and consumer product

4
TMS320C8x Overview

• RISC Master processor _at_ 50 and 60 MHz
• Parallel processors x2 (4 for c80)
• Transfer controller DMA and memory controller
• Video controller (C80 only)

5
TMS320C8x Master Processor

• 32-bit RISC instruction/64-bit data
• Scoreboarded 31 GP registers and a zero register
• IEEE 754 floating point
• Supports vector FP operations
• Performs single precision FP MAC in 1 cycle 100
  MFLOPS (_at_50 MHz)
• Suitable for control protocols and FP intensive
  algorithms

6
TMS320C8x Processor communication

• Shared memory multiprocessor
• MP sends commands through command buffers located
  in shared memory

7
TMS320C8x Parallel Processor

• Data unit 32-bit datapath, ALU, multiplier,
  etc.
• 2 Independent Address units global and local
• Single cycle on-chip memory access (no conflict)
• Single cycle load/store of byte, halfword, and
  word
• Internal adder can offload data unit computation
• Program flow control unit
• 3-stage pipelining
• Instruction fetch
• Address generation, and
• Operation execution
• Supports conditional operation of data unit
  operations, moves, load from memory and branches
• PC is mapped into register file
• To minimize overhead Loop controller supports 3
  levels of nested loops

8
TMS320C8x Parallel Procesor Data Unit

• Split 32-bit 3-input ALU Boolean and arithmetic
  operations
• Split and rounded multiplier dual 8x816,
  16x1632
• Flexible datapath barrel rotator, mask
  generator
• Supports signed, unsigned and saturate arithmetic

9
TMS320C8x Parallel Procesor Data Unit
3-input ALU

• Supports totally 512 operations Boolean 256
  Arith. 256
• Boolean F0 (ABC) F1 (ABC) F2
  (ABC) F3 (ABC) F4 (ABC)
  F5 (ABC) F6 (ABC) F7 (ABC)
• Arithmetic A f1(B,C) f2(B,C) 1
• Example
• AB1
• (AB)(BC) Mask A and B by C and then add
• A((BC) (-BC)) Multiple-byte AB
• A-((BC) (-BC)) Multiple-byte A-B

10
TMS320C8x Parallel Processor Instruction Set

• 64-bit opcode contains multiple subinstructions
  for
• Data unit
• Global address unit and
• Local address unit
• Ex d4d5d6gtgtd0 a8d7 d0(a0x1)

11
TMS320C8x Transfer Controller

• Prioritizes, schedules, and transfers data cache
  between on- and off-chip memories
• Handles data cache (on chip RAM) miss and
  instruction cache
• Supports multidimensional data transfers
• simple contiguous linear sequence up to 3D region
• Memory interface supports a wide range of memory
  system
• DRAM, SDRAM, Video RAM and SRAM

12
TMS320C8x Video Controller (c80 only)

• Provides simultaneous control over two
  independent capture or display systems and frame
  grabber or frame buffer image storage
• Dual-frame timers
• Programmable timing and control registers
• Programmable line interrupt to MP

13
TMS320C8x Development Tools

• C-like compilers and assemblers for both master
  and parallel processor
• Register allocator
• identifies live and free registers
• allows using variable names in assembly code and
• assigns specific register to variable
• Code compactor converts straight-line assembly
  codes into parallel codes
• Optimization can be done by hand for
  time-critical parallel code

14
TMS320C8x Execution Time for 256-Point FFT
-C" indicates performance with the cache
pre-loaded - Benchmark results for the TMS320C80
are for one of the on-chip DSP processors
15
MPACT-R-2 Overview

• VLIW CPU
• Multimedia ISA

• Hardware/Software relationship
• Variety of high speed I/O interface

16
MPACT-R-2 CPU Datapath

• Data size multiple of 9 bits
• 512 72-bit register file with 4 read and 4 write
  ports
• ALU1 - shift and align
• ALU2 - add and logic
• ALU3 - arithmetic and logic
• ALU4 - stage 1 of multiplication
• ALU5 - motion estimation
• Full crossbar between ALU outputs, inputs,
  register read and write ports

17
MPACT-R-2 CPU Datapath
18
MPACT-R-2 Multimedia ISA

• Issues two instruction pack of 72 bits every
  cycle
• Data forwarding from one ALU to one another
• Vector instruction (length upto 255)
• Multimedia data byte of 9, 18, 27, and 36 bits
• Supports signed , unsigned and saturating
  arithmetic
• MPACT 2 includes single-precision FP for 3D
  graphics
• Flow control branch, jump and calls
• Special purpose instruction
• Motion Estimation
• IDCT
• Butterfly FFT, etc.

19
MPACT-R-2 Hardware/Software Relationship

• Requires a host x86 CPU
• Mediaware- uses standard APIs
• RM Resource Manager running under Windows
• MRK MPACT real-time kernel
• Nearest deadline scheduling algorithm
• Interrupt-driven kernel with 4-us context switch
  time in the worst case

20
MPACT-R-2 Hardware/Software Relationship
21
MPACT-R-2 High speed I/O interface

• PCI bus or AGP (Accelerated Graphics Port)
• x86 Host CPU bus
• 66 MHz gt 264 Mbytes/s
• Rambus Memory Interface
• 300 MHz bus (9-bit wide) on both edge600Mbytes/s
• Requires 2-4 Mbytes
• Display Controller
• 24-bit RAMDAC
• High resolution up to 1280x1024 24-bit or
  1600x1200 16-bit
• Video Interface
• Accepts NTSC and PAL format video or
• DVD input through PCI or AGP
• Programmable Peripheral I/O Interface
• Supports connection to several devices

22
MPACT-R-2 Architecture trade-offs

• High speed I/O to move data inout
• No Data cache but large register file
• multimedia data has poor locality
• Based on standard APIs (Application Program
  Interface) of Microsoft Windows no proprietary
  API
• Pin counts vs. high memory bandwidth/low latency
• RDRAM is chosen
• PC and Consumer market

23
Targeted Applications

• Video - DVD, MPEG 1 2 decoding, H.261/263
• Audio - Dolby AC-3, 3D Audio, MPEG Decode,
  Wavetable Synthesis
• Graphics - 2D 3D acceleration
• Communication
• Fax/MODEM V.34, 56k
• Desktop Videoconferencing
• H.320 ISDN
• H.324 on POTS (Plain Old Telephone System)

24
H.324 on TMS320C82 Overview

• ITU-T H.324 Low-bit-rate multimedia
  teleconferencing on circuit-switched network
  includes
• G.723 Audio coding at 5.3-6.4 kbps requires 18-20
  fixed-point MIPS
• H.263 Video coding based on H.261 includes some
  enhancements
• H.223 MUX/DEMUX control
• H.245 Control protocol
• V.34 Modem up to 33.6 kbps
• Other related standards H.320 (ISDN), H.323
  (LAN), and H.310 (ATM/B-ISDN)

25
H.324 on TMS320C82 Overview
26
H.324 on TMS320C82 Task Partitioning

• Video Processing (H.263)
• Encoding
• Pre-processing MP
• Motion estimation PP0
• DCT PP0
• Decoding
• Huffman or arithmetic decode, IDCT, etc. PP0
• Post processing PP0
• Audio Processing and AEC (Acoustic Echo
  Cancellation) - PP1
• G.723
• Encoding 22 MIPS
• Decoding 3 MIPS
• AEC LMS algorithm up to 64-ms echo 10MIPS
• MODEM V.34 20 MIPS - PP1

27
H.324 on TMS320C82 Task Partitioning
28
Current Multimedia Processors
29
References

• J. Golston, Single-chip H.324 video
  conferencing, IEEE Micro, August 1996, pp. 42-50
• Texas Instrument, TMSC320C80 Data Sheet, 1997
• http//www.ti.com/../sprs023b.pdf
• P. Lapseley and G. Blalock, How to estimate DPS
  processor performance, IEEE Spectrum, July 1996,
  pp. 74-78
• HTML file http//www.bdti.com/../wpeval.html
• P. Kalapathy, Hardware-software interfacing on
  Mpact, IEEE Micro, March 1997, pp. 20-26
• Presentation file http//infopad.eecs.berkeley.ed
  u/HotChips8/
• Chromatic Research, MPACT2 Preliminary Data
  Sheet, Feb. 1998
• http//www.mpact.com/../mpact2.pdf
• Toshiba, TOSHIBA ANNOUNCES ITS NEXT-GENERATION
  MPACT MEDIA PROCESSOR, September 22, 1997
• http//www.toshiba.com/taec/../to-628.htm

30
References

• G. A. Slavenburg, The Trimedia TM-1 PCI VLIW
  Mediaprocessor, IEEE Hot Chips 8 Symposium on
  High-Performance Chips, Aug. 1996
• http//infopad.eecs.berkeley.edu/HotChips8/
• L. T.Nguyen, M. Mohamed, H. Park, Y. Pal, R.
  Wong, A. Qureshi, P. Psong, F. Valesco, H. D.
  Truong, C. Reader, Multi-media Signal Processor
  (MSP) Summary , IEEE Hot Chips 8 Symposium on
  High-Performance Chips, Aug. 1996
• http//infopad.eecs.berkeley.edu/HotChips8/
• D. Lindbergh, The H.324 multimedia communication
  standard, IEEE Communication Magazine, December
  1996, pp. 46-51
• K. Rijkse, H.263 Video coding for low-bit-rate
  communication, IEEE Communication Magazine,
  December 1996, pp. 42-45

31
Useful links

• CPU Information Center
• http//infopad.eecs.berkeley.edu/CIC/
• Microprocessor Report
• http//www.chipanalyst.com/q/
• Berkeley Design Technology Inc.
• http//www.bdti.com/
• Peter Pirschs research group
• http//www.mst.uni-hannover.de/