DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
An Overview
ace DSPs?
ts reserved.
Page 1 of 3211/18/97slides_pdf.fm5
Page 1 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
DSP on General-Purpose Processors — or
Can General-Purpose Processors Repl
MicroDesign Resources DInner MeetingSanta Clara, California
January 9, 1997
Jeff Bier
Berkeley Design Technology, Inc.Berkeley, California
E-mail: [email protected]: www.bdti.comTel: (510) 665-1600
Copyright © 1997 Berkeley Design Technology, Inc. All righ
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
deo, motor control,
mmable) in 1996
rward Concepts]
rocessors (GPPs).
bsume the other
Page 2 of 3211/18/97slides_pdf.fm5
Page 2 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
Why is this Important?
DSP applications are proliferating dramatically:
• Wired and wireless comms., music, speech, vinoise cancellation, navigation, ...
• Many products are becoming DSP-intensive
The DSP chip market is booming:
• $2.3B (programmable) and $3.6B (non-progra
• Growth rate > 30% per year [Source: Fo
Many products contain DSPs and general-purpose p
• With increasing integration, one will tend to su
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
n
rs
Page 3 of 3211/18/97slides_pdf.fm5
Page 3 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
Outline of Today’s Presentatio
Focus: DSP capabilities of general-purpose processo
• Applications and system architectures
• Processor architectural approaches
• Evaluation criteria
• Evaluating GPPs for DSP applications
• Conclusions and Trends
Not included:
• “Media” processors
• Non-programmable devices
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
t
y ?
ony,
gration, cost- many applications
be problematic
ifficult
Page 4 of 3211/18/97slides_pdf.fm5
Page 4 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
Short Answer for the Impatien
Can general-purpose processors replace DSPs toda
Yes:
• Especially in personal computers
• Especially for certain applications; e.g., telephvideoconferencing
But, you may not want to, because:
• DSPs have an unbeatable combination of inteperformance, low power, and infrastructure for
• Ensuring strict real-time behavior on GPPs can
• DSP software development on GPPs can be d
Requires application-by-application analysis
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
tures
Page 5 of 3211/18/97slides_pdf.fm5
Page 5 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
Applications and System Architec
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
tions?
int
Page 6 of 3211/18/97slides_pdf.fm5
Page 6 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
What’s Special About DSP Applica
Demands:
• Lots of number crunching
• High data bandwidth; limited data locality
• Real-time constraints
• Attention to subtle numeric effects (in fixed-poimplementations)
• Specialized peripherals/interfaces
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
categories:
gnition/synthesis, on
nes, pagers, motor ering machines
eneral-purpose
ow, but embedded
Page 7 of 3211/18/97slides_pdf.fm5
Page 7 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
Applications
For this analysis, we divide DSP applications into two
• Personal-computer-based
• E.g., modems, speech compression/recomusic/sound synthesis, video compressi
• Embedded
• E.g., disk drive servo control, cellular phocontrol, navigation, modem banks, answ
Both classes are candidates for implementation on gprocessors.
PC-based applications are receiving more attention napplications are and will be far more numerous.
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
rocessor can be
onsumption
Page 8 of 3211/18/97slides_pdf.fm5
Page 8 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
System Architectures
Many existing or emerging products:
• Already contain a µP or µC
• Already contain a µP or µC plus a DSP
• Require µP/µC and DSP functionality
Merging all programmable functionality into a single pattractive:
• High integration can reduce size, cost, power c
• Leverages existing software, tools, know-how
• Few modifications to existing system hardware
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
hes to DSP
Page 9 of 3211/18/97slides_pdf.fm5
Page 9 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
Overview of GPP Architectural Approac
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
DSP
iety of approaches
P features)
nted features
Page 10 of 3211/18/97slides_pdf.fm5
Page 10 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
GPP Architectural Approaches to
General-purpose processor vendors have taken a varto addressing DSP performance:
• Baseline GPPs (moderate performance, no DS
• High-performance GPPs with few/no DSP-orie
• GPPs with major DSP-oriented features
• SIMD
• DSP-processor-like
• DSP co-processors
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
s
I
atures perform
tection,
Page 11 of 3211/18/97slides_pdf.fm5
Page 11 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
I: Baseline GPP Architecture
Example: Advanced RISC Machines’ ARM7TDM
Typical moderate-performance GPPs with no DSP fepoorly on DSP tasks.
The main reasons for this are:
• Poor multiplication throughput
• Limited memory bandwidth
• Loop overhead
• Address generation overhead
Also, on fixed-point processors:
• Lack of hardware support for fast overflow proconvergent rounding, etc.
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
DSP Features
R4650
s.
ical DSPs)
ations
and multi-issue
ia multi-issue
f DSP code and
Page 12 of 3211/18/97slides_pdf.fm5
Page 12 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
II: High-Performance GPPs with No/Few
Example: Pentium (P54C), PowerPC 604e, IDT
These processors can perform very well on DSP task
The main reasons for this are:
• High clock rates (200+ MHz; 2-5 X those of typ
• Single-cycle multiplication and arithmetic oper
• Good memory bandwidth
• Loop overhead reduced via branch prediction
• Address generation, other overhead reduced v
However, dynamic features complicate optimization oreal-time development.
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
(1 of 2)
ns
nce by combining s (group II) with
itioned one)
e.g., 4 multiplies)
data types
ecture can be
Page 13 of 3211/18/97slides_pdf.fm5
Page 13 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
III: GPPs with Major DSP Features
Approach A: Single-instruction, multiple-data operatio
Example: Intel MMX Pentium (P55C)
These processors achieve outstanding DSP performathe features of “conventional” high-performance GPPnew SIMD capabilities:
• Partition existing data path (or add a new, part
• Multiple operations/cycle on small data types (
• Single-cycle operations on various fixed-point
• Specialized instructions
Integration of these features into a pre-existing architawkward.
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
(2 of 2)
es
y mimicking DSP
ated registers
addressing,
ecture can be
Page 14 of 3211/18/97slides_pdf.fm5
Page 14 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
III: GPPs with Major DSP Features
Approach B: Integration of DSP-processor-like featur
Example: Hitachi SH-DSP
These processors achieve good DSP performance bprocessors.
To a conventional GPP architecture, they add:
• A DSP-oriented data path, complete with dedic
• Address generators, hardware looping, modulosaturation, etc.
Integration of these features into a pre-existing architawkward.
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
on DSP tasks.
Page 15 of 3211/18/97slides_pdf.fm5
Page 15 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
IV: DSP Co-processors
Example: ARM Piccolo
These processors should achieve good performanceNone widely deployed yet.
Approach is similar to IV(B), but:
• Programming can be more complicated
• More parallelism may be possible
Contrast with DSP + GPP on one chip:
• Motorola MC68356
• Texas Instruments TMS320C54x + ARM7
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
Page 16 of 3211/18/97slides_pdf.fm5
Page 16 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
Evaluating GPPs for DSP
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
de off vs. speed)
Page 17 of 3211/18/97slides_pdf.fm5
Page 17 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
What’s Most Important?
• DSP speed
• DSP numeric performance (can sometimes tra
• Cost
• Cost/performance
• Power consumption
• Real-time suitability
• Product development time and cost
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
n Time ( µs)
PP
C 6
04e
200
MH
z
PP
C 6
04e-
C20
0 M
Hz
AD
SP
-210
6240
MH
z
AD
SP
-210
62-C
40 M
Hz
110
87
159 158
Floating-Point
on-chip caches preloaded
Page 18 of 3211/18/97slides_pdf.fm5
Page 18 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
Speed: BDTI FFT Benchmark Executio
AR
M7T
DM
I40
MH
z
SH
-DS
P60
MH
z
P55
C20
0 M
Hz
P55
C-C
200
MH
z
IDT
R46
5013
3 M
Hz
TM
S32
0C54
x66
MH
z
P54
C20
0 M
Hz
P54
C-C
200
MH
z
1098
211
89 86
217
199
159151
1100
300
200
100
0
Measured
Fixed-Point
Estimated “-C” indicates
(Lower is Better)
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
effects.essors.
t support
ith software, but
Page 19 of 3211/18/97slides_pdf.fm5
Page 19 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
DSP Numeric Performance
DSP applications are often very sensitive to numeric This is typically not a concern with floating-point proc
On fixed-point processors, key issues include:
• Selection of appropriate word widths
• Overflow protection
• Convergent rounding
• Multi-precision/block floating-point/floating-poin
Lack of needed hardware support can be overcome wthe cost may be high.
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
.
rom processor
processor plus its hanced GPP.
PPs vs. DSPs:
mance penalty for rice/perf. at high end.
Page 20 of 3211/18/97slides_pdf.fm5
Page 20 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
Cost
Cost is surprisingly tricky to analyze.
• Processor cost alone is often not very relevant
Need to compare the overall system costs resulting fchoices.
• E.g., may need to compare the cost of a DSP own memory vs. the cost of upgrading to an en
• Memory usage plays an important role.
Pricing strategies are very different for PC-oriented G
• PC-oriented GPPs command > 2X price/perforthe fastest versions. DSPs have their best p
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
Speed(MHz)
Unit Price(Qty. 1,000)
60 $45
133 $63
133 $34
200 $550
66 $35
200 $509
100 $106
225 $620
100 $173
40 $170
Page 21 of 3211/18/97slides_pdf.fm5
Page 21 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
Example Cost
Type Vendor Processor
Fixed-Point
Hitachi SH-DSP
IDTR4650
R4640
IntelMMX Pentium
(projected)
Texas Instr. TMS320C548
Floating-Point
Intel Pentium
Motorola PowerPC 604e
Analog Devices ADSP-21062
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
Benchmark
100
MH
z
PP
C 6
04e-
C13
3 M
Hz
AD
SP
-210
62-C
33 M
Hz
1.25
1.77
ating-Point
on-chip caches preloaded
Page 22 of 3211/18/97slides_pdf.fm5
Page 22 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
Cost-Execution Time Product, Block FIR
SH
-DS
P60
MH
z
P55
C-C
200
MH
z
IDT
R46
5013
3 M
Hz
TM
S32
0C54
x50
MH
z
P54
C-C
0.34
1.25
0.18
0.74
1.48
2.00
1.50
1.00
0.50
0
($-µs, Lower is Better)
Measured Estimated
FloFixed-Point
“-C” indicates
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
ria in many
ance with very low e power
geTypical Power Consumption
(W)
.0 0.20 (est.)
3 2.1
.0 0.11
Page 23 of 3211/18/97slides_pdf.fm5
Page 23 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
Power Consumption
Power consumption is a key processor selection criteimportant DSP applications.
• Today, only DSPs combine good DSP performpower consumption and application-appropriatmanagement.
Example data:
Vendor ProcessorSpeed(MHz)
Volta(V)
Hitachi SH-DSP 40 3
IDT R4650 133 3.
Texas Instr. TMS320C54x 50 3
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
general-duction?
Page 24 of 3211/18/97slides_pdf.fm5
Page 24 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
Real-Time Suitability
Q: Why is running DSP code onpurpose processors like alien ab
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
me.
Page 25 of 3211/18/97slides_pdf.fm5
Page 25 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
Real-Time Suitability
A: Both result in inexplicable gaps in ti
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
pplications.
: failure to meet a n.
features:
ar execution, data-
to be stochastic.
P applications.
real-time support
Page 26 of 3211/18/97slides_pdf.fm5
Page 26 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
Real-Time Suitability
The most important DSP applications are real-time a
• Many of these are “hard real-time” applicationsreal-time deadline creates a serious malfunctio
High-performance GPPs make heavy use of dynamic
• Caches, branch prediction, dynamic superscaldependent instruction execution times, etc.
These features result in timing behavior that appears
• This seriously complicates development of DS
PC applications are further complicated by the lack ofin PC operating systems.
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
Cost
ent effort:
n
s requires using
ogram for DSP
l support, but DSP-
ar between
Page 27 of 3211/18/97slides_pdf.fm5
Page 27 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
Product Development Time and
Among the most important factors affecting developm
• Breadth and quality of tools and documentatio
• Processor ease of use
• Availability of off-the-shelf software libraries
Developing DSP code for general-purpose processorassembly language when efficiency is important.
• High-performance GPPs are very difficult to pr
• The most popular GPPs enjoy unparalleled toooriented tools are rare
• DSP software libraries for GPPs are few and f
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
nge
update
update
eration
pdate
Page 28 of 3211/18/97slides_pdf.fm5
Page 28 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
Example of Optimization Challe
Vector addition on PowerPC 604e:
@vec_add_loop:
lfsu fpTemp1,4(rAAddr) # Load A data, ptr.
lfsu fpTemp2,4(rBAddr) # Load B data, ptr.
fadds fpSum,fpTemp1,fpTemp2 # Perform add op
stfsu fpSum,4(rCAddr) # Store sum, ptr. u
bdnz @vec_add_loop # loop
Q: How many instruction cycles per iteration?
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
Page 29 of 3211/18/97slides_pdf.fm5
Page 29 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
Conclusions and Trends
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
DSPs?
s strong. ls are weaknesses.
lly where a µP or s are
eeds are modest.
ing the best quality DSPs and other
plications, today’s lled together all of e.
Page 30 of 3211/18/97slides_pdf.fm5
Page 30 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
Conclusions: Can GPPs Replace
Today:
Yes:
• In some PC DSP applications, the case iReal-time behavior, OS support, and too
• In some embedded applications, especiaµC is already established, DSP algorithmstraightforward, and DSP performance n
And, no:
• In many PC DSP applications, users needand highest performance will benefit fromspecialized processors.
• In the most important embedded DSP apGPPs cannot compete: they have not puthe necessary attributes and infrastructur
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
asingly important
iented
ructure for GPPs -start
for architectural
ors; DSP and GPP
ed for the wide
f DSP applications
Page 31 of 3211/18/97slides_pdf.fm5
Page 31 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
Trends
• DSP applications will continue to become incre
• GPPs will continue to add and expand DSP-orenhancements
• DSP-oriented tools, software, and other infrastwill develop, but DSPs have a significant head
• DSPs will not stand still; there is fertile groundinnovation, clock speed increases, etc.
• There will be an expanding diversity of processfamily trees will mix
• Capabilities will become increasingly specializrange of important DSP applications
• GPPs will be suitable for an expanding range o
DSP on General-Purpose Processors
Berkeley Design Technology, Inc.
eleased)
96, pp. 12-15)
P Strate gies 2000
T, DSP World, etc.
Page 32 of 3211/18/97slides_pdf.fm5
Page 32 of 3211/18/97Copyright © 1997 Berkeley Design Technology, Inc.
Further Resources
• BDTI technical reports:
• DSP on General-Purpose Processors (just r
• Buyer’s Guide to DSP Processors
• Microprocessor Report articles (especially 12/30/
• BDTI’s web site: www.bdti.com
• Forward Concepts market research reports: DS
• DSP-oriented trade shows and conferences: ICSPA
• Join BDTI ... we’re hiring
Top Related