ICME 2004© 2004 Marilyn Wolf Architectures for Multimedia Computing 3: DSPs and Systems-on- Chips Marilyn Wolf Dept. of Electrical Engineering Princeton.

ICME 2004 © 2004 Marilyn Wolf

Architectures for Multimedia Computing3: DSPs and Systems-on-Chips

Marilyn Wolf

Dept. of Electrical Engineering

Princeton University

What is a DSP?

Traditional definition: Harvard architecture. Multiplier-accumulator.

Today: Marketing term (every CPU is a DSP).

DSP goals

Real-time computation on streaming data. Low latency.

Small code size. Low energy/power.

TI C54x architectural features

40-bit ALU + barrel shifter. Multiple internal busses: 1

instruction, 3 data, 4 address. 17 x 17 multiplier. Single-cycle exponent encoder. Two address generators with

dedicated registers.

TI C54x instruction set features

Specialized instructions for Viterbi. Repeat and block repeat

instructions. Instructions that read 2, 3 operands

simultaneously. Conditional store. Fast return from interrupt.

Sample configurations

C5409: 100 MIPS @ 100 MHz. Address space sizes:

Data 64K words. Program 8M words.

On-board memory: RAM 16K words. ROM 32K words.

3 buffered serial ports. 1 timer.

C5510: 400 MIPS @ 200 MHz. Total address space:

8 M Words.

On-board memory: RAM 160K words. ROM 16K words.

3 McBSP serial ports. 2 timers.

Single-instruction repeat loop example

STM #4000h,AR2

; load pointer to source

STM #100h,AR3

; load pointer to destination

RPT #(1024-1)

MVDD *AR2+,*AR3+

; move

C55x organization

Instructionunit

Programflowunit

Addressunit

Dataunit

3 data read busses

3 data read address busses

program address bus

programread bus

2 data write busses

2 data write address busses

16

24

24

16

24

32

InstructionfetchData readfrom memory

D bus

Single operandread

C, D busses

Dual operandread

B bus

Dual-multiplycoefficientWrites

C55x low power features

Parallel execution units---longer idle shutdown times.

Multiple data widths: 16-bit ALU vs. 40-bit ALU.

Instruction caches minimizes main memory accesses.

Power management: Function unit idle detection. Memory idle detection. User-configurable IDLE domains allow

programmer control of what hardware is shut down.

C5x McBSP serial ports

Designed to interface to devices. Full, duplex; double-buffered transmission;

triple-buffered reception. Works with interrupts or DMA. Separate xmit, rcv clock, highly

programmable. -law, A-law companding. Direct interface to codecs, T1/E1 framers,

MVIP framers, etc.

McBSP block diagram16-bit peripheral bus

Registersfor sync

control andmonitoring

Registers formultichannelcontrol andmonitoringco

mpa

nd

Data rcv, xmit clock

interrupt DMA

Clock/frame sync

Image/video hardware extensions

Available in 5509 and 5510. Equivalent C-callable functions for

other devices.

Available extensions: DCT/IDCT. Pixel interpolation Motion estimation.

DCT/IDCT

2-D DCT/IDCT is computed from two 1-D DCT/IDCT. Put data in

different banks to maximize throughput.

block

Column DCT

interim DCTRowDCT

C55 DCT/IDCT coprocessor extensions

Load, compute, transfer to accumulators: ACy=copr(k8,ACx,Xmem,Ymem)

Compute, transfer, mem write: ACy=copr(k8,ACx,ACy), Lmem=ACz

Special: ACy=copr(k8,ACx,ACy)

Software pipelined load/compute/store for DCT

Iteration i-1

Dual_load

8compute

4Long_store

4 empty

3Dual_load

empty

Iteration i

Dual_load

8compute

4Long_store

4 empty

3Dual_load

empty

Iteration i+1

Dual_load

8compute

4Long_store

4 empty

3Dual_load

empty

op_i(0), load_i+1(0,1)op_i(1), store_i-1(0,1)op_i(2), store_i-1(2,3)op_i(2), store_i-1(4,5)op_i(2), store_i-1(6,7)op_i(2), load_i+1(2,3)…

C55 motion estimation

Search strategy: Full vs. non-full.

Accuracy: Full-pixel vs. half-pixel.

Number of returned motion vectors: 1 (one 16x16) vs. 4 (four 8x8).

Algorithms: 3-step algorithm (distance 4,2,1). 4-step algorithm (distance 8,4,2,1). 4-step with half-pixel refinement.

Four-step motion estimation breakdown

d = {8,4,2,1};

for (i=0; i<4; i++) {

compute 3 upper differences for d[i];

compute 3 middle differences for d[i];

compute 3 lower differences for d[i];

compute minimum value;

move to next d;

}

X

C55 motion estimation accelerator

Includes 3 16-bit pixel data paths, 3 16-bit absolute differences (ADs).

Basic operation: [ACx,ACy] =

copr(k8,ACx,ACy,Xmem,Ymem,Coeff) K8 = control bits (enable AD units, etc.) ACx, ACy = accumulated absolute differences Xmem, Ymem = pointers to odd, even lines of

the search window Pointer to two adjacent pixels from reference

window

C55 pixel interpolation

Given four pixels A, B, C, D, interpolate three half-pixels:

A B

C D

U

M R

Pixel interpolation coprocessor operations

Load pixels and compute: ACy=copr(k8,AC,Lmem)

Load pixels, compute, and store: ACy=copr(k8,AACx,Lmem) ||

Lmem=ACz

Motorola 56301

24-bit word. 80 MIPS @ 80 MHz, 100 MIPS @ 100 MHz. 24x24 multiplier, 56-bit barrel shifter. On-chip instruction cache controller. Nested hardware DO loops. On-chip memory:

RAM 1024 -> 4096 words. X data RAM 2048 -> 3072 words. Y data RAM 2048 -> 3072 words.

TI C62/C67

Up to 8 instructions/cycle.32 32-bit registers.Function units:

Two multipliers.Six ALUs.

All instructions execute conditionally.

TI C6x data operations

8/16/32-bit arithmetic.40-bit operations.Bit manipulation operations.

C6x system

On-chip RAM.32-bit external memory: SDRAM,

SRAM, etc.Host port.Multiple serial ports.Multichannel DMA.32-bit timer.

C6x block diagram

Data path 1/Reg file 1

Data path 2/Reg file 2

Execute DMA

timers

Serial

Program RAM/cache512K bits

Data RAM512K bits

JTAG

PLL

bus

C6x data paths

General-purpose register files (A and B, 16 words each).

Eight function units:.L1, .L2, .S1, .S2, .M1, .M2, .D1, .D2

Two load units (LD1, LD2).Two store units (ST1, ST2).Two register file cross paths (1X and 2X).Two data address paths (DA1 and DA2).

C6x function units

.L 32/40-bit arithmetic. Leftmost 1 counting. Logical ops.

.S 32-bit arithmetic. 32/40-bit shift and 32-bit field. Branches. Constants.

.M 16 x 16 multiply.

.D 32-bit add, subtract, circular address. Load, store with 5/15-bit constant offset.

Motorola Starcore SC140

DALU includes 4 ALUs, 1 register file. AGU includes 2 address arithmetic

units (AAU), 1 address register file. Program sequencer and control unit

(PSEQ). Performance:

4 million MACs per second per MHz clock. 10 RISC MIPS per MHz clock.

Typical SC140 configuration

Level 1 memory expansionRAM, ROM

peripherals

DMA,Cache,

Interrupts,Level 2 mem,

Etc.

Program sequencer

ALUALUALUALU

AAUAAU

SC140 Core

Programsequencer

Powermgt

Clock/PLL

AGU DALU

AddressRegister file

2 AAUs BMU

DataRegister file

4 ALUs

Instruction format

16-bit instructions. Up to 6 instructions per cycle. Instructions are grouped to define

allowable simultaneous operations.

MACR –D0,D1,D7 AND D4,D5 MOVE.L (R0),+N0,R6 ADDA R2,R3

DALU AGU

AGU addressing

Allowable addressing modes: Linear. Modulo. Reverse-carry. Automatic updating during register

indirect. Stack.

Array addressing: base, offsett, modifier registers.

Motorola MC8126 multiprocessor

Four SC140 cores. 224 KB M1 memory with zero wait states.

Shared M2 memory. 476 KB running at core frequency. M2-accessible multi-core bus.

64/32-bit data, 32-bit address 60x bus. Eight-bank memory controller. Hardware semaphores.

TM-1 characteristics

Characteristics Floating point

support Sub-word

parallelism support If Conversion Additional custom

operations

Trimedia TM-1

memory interface

video in

audio in

I2C

timers

image co-p

PCI

video out

audio out

serial

VLD co-p

VLIW CPU

TM-1 VLIW CPU

register file

read/write crossbar

FU1 FU27

slot 1 slot 2 slot 3 slot 4 slot 5

...

Embedded vs. scientific applications

Embedded applications provide task-level parallelism.

Embedded applications run many different types of algorithms at once.

Scientific numerical calculation

Partition a large calculation across multiple processors:

CPU CPU CPU…

Task level parallelism

Map tasks on to multiprocessor:

PE 1 PE 2 PE 3+

application architecture

What is an MPSoC?

Multiple processors. CPUs, DSPs, etc. Hardwired blocks. Mixed-signal.

Custom memory system. Lots of software.

MPSoC applications

Sophisticated markets: High volume. Demanding performance, power

requirements. Strict price restrictions.

Often standards-driven. Multimedia is an ideal application.

Why multiple platforms?

People still care about cost.People care about power

consumption.Sufficiently general solutions don’t fit

on one chip.

CD/MP3 player

AudioCPU

amp

Jogmemory

Errorcorrector

ServoCPU

Analogin

Analogout

FE, TE, amp

focus,tracking,sled,motor

head

drive

memory

memory

display

DAC

I2S

CD mechanism

Laser, lens, sled:

lase

r

CD

detectorsdiffraction

gratingsled

track

track

focus

Servo control

Four main signals:focus (laser) @ 245 kHz;tracking (laser) @ 245 kHz;sled (motor): @ 800 Hz;Disc motor.

Optical pickup

Control and error correction

Skips caused by physical disturbance. Wait for disturbance to subside. Retry.

Read errors caused by disc/servo problems. Detect error. Choose location for retry. Retry. Fail and interpolate.

DVD player

Yasuda, et al (Matsushita): two-chip solution for DVD back-end.Video MPEG 2 at Main Level (MP2@ML).Audio AC-3.External 16 Mb SDRAM.

DVD back-end

MPEG2stream

demux

AC-3Audio decoder

MPEG2Video decoder

Subpicturedecoder

Colorencoder

+

SDRAM

Video decoder

Hostinterface

RISCcontroller

Micro-program

Streamparser

Macroblockpipeline

VideoOutput

core

PLL SDRAM interface

system

audio

Macroblock pipeline

Allowed processing time:[30 rows*45 macroblocks/row]/29.97 Hz

=24.71 s.

Three stages in pipeline:Variable-length decode.Inverse quantize/inverse DCT, motion

compenstation.Reconstruction combines IQ/IDCT + MC.

Viper set-top box SoC

MIPS Trimedia

Off-chip SDRAM

MC bridge TC bridge

Bus ctrlBus ctrl

Clocks,DMA,Reset,debug

I2C, Smcard

PCI

USB, 1394

MMIbus

MBS

2D AICP

MPEG

SPDIF

GPIO

C bridge

Bus requirements

Separate cache and register-access traffic.

Separate register-access traffic of the two CPUs.

Each CPU has a high-performance, low-latency path to memory.

Each CPU has low-latency access to its peripherals.

Bus structure

MMI bus goes to bulk memory. Point-to-point design. 64-bit, 143 MHz. Round-robin arbitration. Controller separates on-chip accesses.

PI bus for peripherals. Tri-state design. Compatible with existing IP for peripherals. Fast-PI for high-speed MIPS peripherals. MIPS-PI for generic MIPS peripherals. T-PI for Trimedia peripherals.

TI OMAP

Targets communications, multimedia.

Multiprocessor with DSP, RISC.

C55x DSP

OMAP 5910:

ARM9

MMU

Memory ctrl

MPUinterface

SystemDMA

control

bridge

I/O

ST Nomadik

Targets mobile multimedia.

A multiprocessor-of-multiprocessors.

ARM9

Mem

ory

syst

em

I/O

bri

dges

Audioaccelerator

Videoaccelerator

heterogeneousmultiprocessors

ST MMDSP+

Embedded processor core used in multiple chips: Runs at 175 MHz. 1 cycle per instruction. 2-level instruction cache. 16/24-bit fixed point. 32-bit floating point. C programmed Fully synthesizable.

Nomadik video accelerator

MMDSP+data

RAMinstrRAM

Xbus

Interruptcontroller

Picturepost

processing

Videocodec

Pictureinput

processing

Localdatabus

MasterAHBDMA

Nomadik audio processor

SlaveAHB

Timers, GPIO, etc.

MMDSP+Y RAMX RAMInstrcache

ARMDMA

DMA1 DMA2

MasterAHB

X Bus

Y Bus

Nomadik support

Supports OMAPI standard. Standard includes mid-level driver

APIs.

Nomadik defines: Base operating system. Base drivers. Multimedia accelerators and their

drivers.

FPGAs

Can be used to provide hardware acceleration. First-generation Replay.

Platform FPGAs provide CPU, FPGA fabric.

FPGA trade-offs: Low design cost. Higher incremental cost in large volumes. Higher power consumption than custom.

Xilinx Virtex II Pro

Up to 4 PowerPC 405s per chip:

5 stage pipe, static branch prediction, etc.

Separate instruction, data caches.

MMU. Timers.

PowerPC

FPGA fabric