DSP Architectures

Professor-- Sanjay Khadagade

EC----Department
[email protected]
 A Digital Signal Processing System

Analog D Analog
Antialiasing Sample Reconst.
Signal Filter and Hold A/D S D/A Filter
Signal
in out
P
A perspective of the Digital Signal Processing problem
Application areas
Medical Radar Speech Seismic Image
•••

Digital signal processing theory

Theoretical
problem Basic functions
modelling

Algorithms

Architechtures
Processor
Implementation instruction sets
and/or hardware
functions

Component technology
 DSP APPLICATION CHARACTERISTICS

APPLICATION REQUIREMENT PROCESSOR ATTRIBUTES

REAL-TIME PROCESSING HIGH SPEED, HIGH THROUGHPUT

LARGE ARRAY OF DATA INSTRUCTIONS TO MOVE AND

PROCESS LARGE DATA ARRAYS

ALGORITHM INTENSIVE FAST MATHEMATICAL

COMPUTATIONS, SINGLE CYCLE
DSP OPERATIONS (MACD)

SYSTEM FLEXIBILITY GENERAL PURPOSE

PROGRAMMABILITY, EPROM,
MC/MP MODE
 Different approaches to hardware
implementation
1. HIGH SPEED GENERAL PURPOSE COMPUTERS

Programmable Expensive
Can be configured for Complex control
different applications I/O overheads

2. CUSTOM-DESIGNED VLSI COMPONENTS

Efficient design Application specific

Large throughputs High development cost
3. GENERAL PURPOSE DIGITAL SIGNAL PROCESSORS

Combine the Programmability & Control features of general

purpose computers and the Architectural innovations of
special purpose chips.

GOALS: HIGH SPEED, LOW POWER AND LOW COST

 General purpose computers

1. Flexible
2. Suitable for Internet and
Multimedia application
3. Software Intensive
4. Slow for high speed application
5. Too bulky
6. Power hungry
 Why are conventional Processors not
suitable for DSP?

 Caches are a waste of chip area

 Small register files force lots of memory

accesses

- these are different from cache since

these are program managed

 Complex instruction issue logic, branch prediction,

speculation etc. are not needed for DSP

 Not enough ALU function

 Not adequate dynamic range and precision

 Data Processing vs
Signal Processing
• General-purpose microprocessors are designed
primarily for Data Processing.
– The primary burden is Data Read/Write

• Digital Signal Processors are Microprocessors

specifically designed for Signal Processing.
– The primary burden is Mathematical operation

• DSP architecture therefore incorporates certain

features not found in general-purpose P’s.
 DSP Requirements
• Emphasis is on mathematical operations rather
than data manipulation operations like word
processing, database management etc.
– Design is optimized for DSP algorithms which
implement FIR filter, FFT generator etc.

• Processing is real-time, i.e. the input signal comes

continuously, and the output signal is also
produced continuously as the input is acquired.

• Dominant mathematical operation is Multiply and

Accumulate (MAC), on separate inputs in parallel.
 Digital Signal Processor features

 Caters to high arithmetic demands

 Real time operation

 Analog input / output

 Large number of functional units for a

given size

 Small control Logic

Typical MAC Execution Cycle
• Obtain a sample of the Input signal
• Move the input sample into the input buffer
• Fetch the co-efficient from internal buffer
• Multiply the input sample by the co-efficient
• Add the product to the Accumulator
• Move the output to the output buffer
• Send it out as a sample of the output signal
 MAC Execution Hardware

Data Read
Program Address
Counter
Data Write
Address

*
Program/ Data
Coefficient Memory
Memory

CPU ACC
 Architecture of Digital Signal
Processors
• General-purpose processors are based on the Von
Neumann architecture (single memory bank and
processor accesses this memory bank thro’ single
set of address and data lines)

• Harvard architecture commonly used in DSP

processors
– Separate Data and Program memories (two memory
banks)
– Separate Address Buses for Data and Program
memories
 Additional Features in a DSP
Processor
• Instruction Cache and Pipelined processor
as in any modern microprocessor, but no
Data Cache

• Separate ALU, Multiplier and Shifter,

connected through multiple internal data
buses, enabling fast MAC operations
 DSP, CISC and RISC
• DSP Processors can’t be called truly as
CISC or RISC-type of processors

• Some features present in a RISC processor

may exist. However, DSP processors are
“tuned” towards operations encountered in
signal processing applications
 DSP IMPLEMENTATION APPROACHES

Important desirable characteristics

 Adequate word length

 Fast multiply & accumulate
 High speed RAM
 Fast Coefficient table addressing
 Fast new sample fetch mechanism
DSP functions implemented with
IC chips
Issues:

Speed Architectural features

Accuracy Register lengths and

floating point capability

Cost Advances in VLSI

techniques
 GENERAL PURPOSE DSP FEATURES
1. PARALLELISM: Multiple Functional Units
Multiple Buses
Multiple Memories

2. PIPELINING
3. HARDWARE MULTIPLIERS AND OTHER
ARITHMETIC FUNCTIONS
4. ON-CHIP AND CACHE MEMORIES
5. A VARIETY OF ADDRESSING MODES
7. INSTRUCTIONS THAT PACK SEVERAL
OPERATIONS
8. ZERO-OVERHEAD LOOPING
9. I/O FEATURES SUCH AS INTERRUPT, SERIAL
I/O, DMA
10. OTHER CONTROL FUNCTIONS SUCH AS WAIT
STATES
 A second order FIR filter

x(n) x(n-1) x(n-2)

Delay Delay

h(0) h(1) h(2)

y(n)
+ +

y(n) = h(0)x(n) + h(1)x(n -1) + h(2)x(n-2)

 x(n+1)

x(n) Delay h(0)

x(n-1) h(1)
Delay
ar1 x(n-2) ar2 h(2)

MAC

y(n)
Organization of signal samples and filter coefficients
for a second order FIR filter implementation
 An Nth order FIR filter implementation

A[0] X[n]
A[1] X[n-1]
A[2] X[n-2]
*
•• ••
•• P ••
•• ••

A[N-1] +
X[n-N+1]
ACC y[n]

Coefficient Data
Memory Memory
 FIR Filter pseudo-code
Load loop count
Initialize coefficient and data addr
regs
Zero Acc and P registers
LOOP: Pnew = A[i] . X[n-i]
Accnew = ACCold + Pold
Decrement coefficient and data addr
regs
X[n-i] X[n-i-1] {for next iteration}
Decr loop count
BNZ LOOP
Acc Y[n]
 A Typical DSP Architecture
PM Data DM Data
PM Address Address Address DM Address
Program
Memory Generator Generator Data
(PM) Memory
(DM)
Instruc- Program Sequencer
tions & Instruction Cache Data
secondary PM Data DM Data
only
data
Registers DMA Bus
I/O
Multiplier Controller
(DMA)
ALU

Shifter
Input/Output
 Salient Features
• REPEAT-MAC instruction
- Performs auto-increment of both coefficient and
data pointers
- Frees up program memory bus for fetching
coefficients
• Circular buffer
- to manage data movement at the end of every
output computation
• Handling precision
- Accumulator guard bits
- Saturation mode
- Shifters (both right and left shift)
 Types of multipliers used

• Array multipliers
• Multipliers based on modified Booth’s
algorithm
Product Computation Unit of a
simple multiplier for 4-bit
unsigned numbers X and Y
Multiplying X and Y: summation
unit of the simple multiplier
Combinational array for Booth’s
algorithm – Basic cell B
Array Multiplier for 4x4-bit
numbers using basic cell B
 Arithmetic
Fixed point Vs Floating point
Array indices, Loop Wider dynamic range
counters etc. frees user from scaling
concerns
Less sensitive to error
accumulation

Overflow/underflow 50% slower for same

management technology
Error budget for Higher Cost
word length growth Normalize after each
operation
Mantissa round off (some
accuracy is traded)
Fixed point does not always limit performance:

e.g., for dynamic range of 50 to 60 dB, 12 -bit

quantization (step size of -72 dB) is more than
adequate. For Hi-fi audio with 80 dB dynamic
range, 16 bits (-96 dB) are more than adequate
 Overflow Management
SHIFT

Left shift removes redundant sign bit after 2’s

complement multiplication
Right shift down scales numbers as word growth is
detected

Unbiased rounding

Prevents accumulation of a small dc bias from

outputs which fall just half way between adjacent
rounded values
Saturation Logic

Sets the contents of register to maximize the

value if overflow occurs

Block Floating Point

Scaling logic + exponent register: If overflow

condition of any point is detected, the entire
array is rescaled downwards and the scaling is
stored in the block exponent register.
 SHIFTERS
- Scales numbers to prevent overflow/underflow
- Conversion between fixed point and floating
point
- Many bits must be shifted in a single cycle to
preserve single cycle computational speed
(Barrel Shifter)
- Logical shift assumes unsigned data and fills
with zeroes left or right
- Arithmetic shift scales numbers upwards
(left) or downward (right)

zero fills sign extend

- Normalization/de-normalization for block
floating point
 Memory
Traditional µPs : register to register
(limited memory bandwidth)
DSPs : memory to memory
(higher memory bandwidth)
upto six memory fetches in an inst. cycle
Parallel memory banks: small, fast and simple
memories.

Internal Vs External

Pincount limitation
Speed penalty  Off-chip bussing

Internal busses are multiplexed to the outside.

Expand only one memory off-chip.
 MEMORY ORGANISATION - I

BASIC HARVARD ARCHITECTURE

PROGRAM DATA
MEMORY MEMORY

MODIFICATION #1 MODIFICATION#2
PROGRAM MULTI-PORT
DATA PROGRAM
/DATA DATA
MEMORY MEMORY
MEMORY MEMORY
 MEMORY ORGANISATION - II
MODIFICATION #3
PROGRAM/
PROGRA DATA
DATA
M CACHE MEMORY
MEMORY

MODIFICATION #4

PROGRA DATA DATA

M MEMORY MEMORY
MEMORY

MODIFICATION #5

I/O PROGRAM DATA DATA DATA

DATA
PROGRAM MEMORY MEMORY MEMORY MEMORY
MEMORY
MEMORY
 Addressing modes
Parallel memory inst. must specify upto 3 memory
accesses
Number of bits required is very large
More memory, wider busses, more memory cycle

Solution: register - indirect addressing modes

Many addresses in one word inst.
Only a few bits are required since register
bank is small ; parallel hardware to update
registers containing memory addresses.
 Addressing Modes (contd.)
• Short immediate addressing mode
• Short direct addressing mode
• Memory mapped addressing mode
• Circular buffer addressing mode
• Bit-reversed addressing mode
 Instruction Level Parallelism

VLIW architecture
• Each instruction specifies several operations
to be done in parallel
• Advantages : Simple hardware
compilers can spot ILP easily
• Disadvantages : Little compatibilty between
generations
Explicit NOPs bloat code size
 Super scalar architecture

• Hardware responsible for finding ILP in a

sequential program

• Advantage : Compatibility between

generations

• Disadvantage : Very complex hardware

 Explicitly Parallel Instruction Computing
(EPIC)
• Combines VLIW and super scalar
architectures
• Instructions are grouped into 3 operating
blocks and a template block
• Template block tells hardware if
instructions can be executed in parallel
• Also gives information whether the block
can be executed in parallel
 ILP versus Power

Increasing instructions / cycle

 Requires fewer cycles to execute a task
 Uses longer clock for same performance
 Uses lower supply voltage
 And hence uses less power
However, too many functional units and too
many transitions per clock cycle increase
power consumption.
 Low Power architecture

 Power consumed by additional circuits vs. ability to

lower clock rate while maintaining performance

 Circuits must be highly used

 Move complexity into software

 Voltage scaling : Reduce Vdd

 Clock gating : Turn off clock when chip

is not in use ( applies to
sub-modules of chip also)
 VLIW is more suitable than super scalar for
low power
- VLIW is smaller for same number of
functional units
- Compiler is better at finding parallelism
than hardware
 Put multiple processors on chip rather than
lots of functional units in one processor
 Helps in running independent tasks
 Improvement of Speed by
Pipelining
• Processor speed can be enhanced by having separate
hardware units for the different functional blocks,
with buffers between the successive units.
– The number of unit operations into which the instruction
cycle of a processor can be divided for this purpose
defines the number of stages in the pipeline.
– A processor having an n-stage pipeline would have up to
n instructions simultaneously being processed by the
different functional units of the processor.
• Effective processor speed increases ideally by a
factor equal to the number of pipelining stages.
A Four-stage Pipeline
 Data Dependency in Pipelining
If the input data for an instruction depends on the
outcome of the previous instruction, the Write cycle of
the previous instruction has to be over before the
Operate cycle of the next instruction can start. The
pipeline effectively idles through one instruction,
creating a bubble in the pipeline which persists for
several instructions.
F1 D1 O1 W1
F2 D2 idle O2 W2
F3 idle D3 O3 W3
Bubble
ends
here
F4 D4 O4 W4
 Example of dependency
• A  3 + A; B  4 x A
Can’t perform these two in parallel
• Another case: A = B + A; B = A – B; A =
A – B (swapping without temp) ; examine
how you can handle this.
Branch Dependency in Pipelining
A Branch instruction can cause a pipeline stall if the
branch is taken, as the next instruction has to be
aborted in that case. If I1 is an unconditional branch
instruction, the next Fetch cycle (F2) can start after D1.
But if I1 is a conditional branch instruction, F2 has to
wait until O1 for the decision as to whether the branch
will be taken or not.
F1 D1 O1 W1 branch instruction

F2 D2 O2 W2 executed if branch is not taken

executed for
F2 D2 O2 W2 unconditional branch

F2 D2 O2 W2 for conditional
branch, if taken
 Pipeline in ADSP 219x
Processors
6-Stage Instruction Pipeline with Single-cycle
Computational Units:
• Look-Ahead: places the address of an instruction
that is going to be executed several cycles down the
road, on the program-memory address (PMA) bus
• Pre-fetch: Pre-fetches an instruction if the
instruction is already in the instruction cache
• Fetch: Fetches the instruction that was “looked
ahead” 2 cycles ago
• Address-decode: Decoding of the DAG operand
fields in the opcode in this cycle
• Decode: The second stage of the instruction
decoding process, where the rest of the opcode is
decoded
• Execute: Instruction is executed, status updated,
results written to destinations
 Causes for Pipeline Stalls
 Memory block conflicts: If both instruction and data are to be
fetched from the same block of memory, a stall is
automatically inserted
 DAG usage immediately (or within 2 cycles) after
initialization. e.g.
 I2 = 0x1234;
 AX0 = DM(I2,M2);
 Bus conflicts: Instructions which use the PMA/PMD buses for
data transfer may cause bus conflict. e.g.
 PM(I5,M7)=M3;
 Avoiding DAG-related Pipeline
Stalls
• Note that
– I2 = 0x1234;
– I3 = 0x0001;
– I1 = 0x002;
– AX0 = DM(I2,M2);
will NOT cause a stall.
• Also, note that switching DAG register
banks (primary  secondary) immediately
before using them will NOT cause a stall.
 TMS320C25 KEY FEATURES
+5 v GND

100 ns INSTRUCTION CYCLE TIME

INTERRUPTS DATA
288 x 16 256 x 16
DATA RAM DATA/
16 128K-WORDS TOTAL MEMORY
PROGRAM
MULTI- SPACE
PROCESSOR
INTERFACE
THREE PARALLEL SHIFTERS
4K x 16 PROGRAM ROM
SERIAL 133 GENERAL PURPOSE AND DSP
INTERFACE INSTRUCTIONS
32-BIT ALU/ACC
S/W UPWARD COMPATIBLE WITH
ADDRESS PREVIOUS FAMILY MEMBERS
16 x 16
MULTIPLIER 16 1.8u CMOS: 68-PIN PLCC / PGA
 TMS320C25 GENERAL PURPOSE FEATURES

COMPREHENSIVE INSTRUCTION SET-

133 INSTRUCTIONS INCLUDING
- NUMERICAL (34)
X=X-Y - LOGICAL (15)
- MEMORY MANAGEMENT (33)
- BRANCHES (20)
- PROGRAM/MODE CONTROL (31)
EXTENDED-PRECISION ARITHMETIC
SERIAL PORT (DOUBLE BUFFERED,
BIT 16=0 STATIC)
NO
MULTIPROCESSOR INTERFACES
(CONCURRENT DMA, GLOBAL DATA
MEMORY)
YES
BLOCK MOVES (UP TO 10 M WORDS/SEC)
ON-CHIP TIMER
OUTPUT X
THREE EXTERNAL MASKABLE
INTERRUPTS
POWERDOWN MODE
TMS320C25 ALU

DESIGN & OPERATION

32-BIT ALU & ACCUMULATOR
CARRY BIT FOR EXTENDED
PRECISION
OVERFLOW DETECTION &
SATURATION
SIGN EXTENSION OPTION
0-16 BIT PARALLEL SHIFTER
FOR LOADS AND ARITHMETIC
OPS
SHIFTERS ON PRODUCT
REGISTER OUTPUT DATA
0-7 BIT PARALLEL SHIFTER FOR
ACCUMULATOR STORES
TMS320C25 - MULTIPLY INSTRUCTIONS II

MAC MPY data memory * program

memory & add past P-Reg to
ACC

MACD MPY data memory * program

memory, add past P-Reg to
ACC, & move data memory

SQRA Square data memory value &

add past P-Reg to ACC

SQRS Square data memory value &

sub past P-Reg from ACC
 TMS320C25 SPECIAL PURPOSE FEATURES

Xi SINGLE-CYCLE
Z-1 Z-1 Z-1 Z-1 MULTIPLY/ACCUMULATE
MULTIPLY/ACCUMULAT
E USING EXTERNAL
PROGRAM MEMORY
REPEAT INSTRUCTION
ADAPTIVE FILTERING
Yi INSTRUCTIONS

BIT-REVERSED
0-16 BIT SCALING SHIFTER (SIGNED ADDRESSING
OR UNSIGNED) AUTOMATIC DATA-
OVERFLOW MANAGEMENT MOVE IN MEMORY (Z-1)
-SATURATION MODE
-BRANCH ON OVERFLOW
-PRODUCT RIGHT SHIFT
TMS320C25 - HIGHER PERFORMANCE AT LESS CODE SPACE

xn
Z-1 Z-1 Z-1 Z-1

x x x x

Yn

N

Yn = b K X(n-K) TMS320C25
K=0
RPTK 49
MACD

3 WORDS PROG MEMORY

53 CYCLES
 TMS320C25 ADDRESSING MODES
 IMMEDIATE ADDRESSING Program memory
- BOTH LONG AND SHORT CONSTANTS
ADDK 5
- EXAMPLES:
ADDLK
ADDK 5
1325
ADLK > 1325
From
 DIRECT ADDRESSING DP instruction

- SAME AS TMS320C1X BUT DP IS 9 BITS 9 BITS 7 BITS

- 512 “BANKS” OF 128 WORDS
- USED OFTEN FOR LONG SEQUENCES OPERAND ADDRESS
OF IN-LINE CODE

 INDIRECT ADDRESSING
- B AUXILIARY REGISTERS
- USED OFTEN IN PROGRAM LOOPS WITH AUTO
INC/DEC OPTIONS
 Addressing Mode (contd.)

• Circular buffer addressing mode

• Bit-reversed addressing mode

TMS320C25 AUXILIARY REGISTER INSTRUCTIONS

LAR Load aux-reg w/data

LARK Load AR w/8-bit constant

LRLK Load AR w/16-bit constant

MAR Modify auxiliary register

SAR Store auxiliary register

ADRK Add 8-bit constant to AR

SBRK Sub 8-bit constant from

AR

LARP Load auxiliary register

pointer
TMS320C25 ON-CHIP MEMORY

MEMORY ORGANIZATION
4K WORDS ON-CHIP
MASKED ROM
544 WORDS ON-CHIP
DATA RAM
256 WORDS ON-CHIP
RAM RECONFIGURABLE
AS DATA/PROGRAM
MEMORY
BLOCK TRANSFERS IN
MEMORY
DIRECT, INDIRECT, AND
IMMEDIATE ADDRESSING
MODES
BLOCK DIAGRAM OF A TMS320C5X DSP
 General-Purpose Microprocessor
circa 1984 : Intel 8088

 ~100,000 transistors
 Clock speed : ~ 5 MHz
 Address space : 20 bits
 Bus width : 8 bits
 100+ instructions
 2-35 cycles per instruction
 Micro-coded architecture
 DSP TMS 32010 1984
 Clock 20 MHz
 16 bits
 8, 12 bits addressing space
 ~ 50 k transistors
 ~ 35 instructions
 Harvard architecture
 Hardware multiplier
 Double length accumulator with saturation
 A few special DSP instructions
 Relatively inexpensive
General Purpose Microprocessor 2000
 GHz clock speed
 32-bit address or more
 32-bit bus, 128-bit instructions
 Complex MMU
 Super scalar CPU
 MMX instructions
 On chip cache
 Single cycle execution
 32-bit floating point ALU on board
 Very expensive
 10s of watts of power
 DSP in 2000

 Clock 100 ~ 200 MHz

 16-bit floating point or 32-bit floating point
 16-24 bits address space
 Large on-chip and off-chip memories
 Single cycle execution of most instructions
 Harvard architecture
 Lots of special DSP instructions
 50 mw to 2w power
 Cheap
 Future of DSP Microprocessor

 Sufficiently unique for an independent

class of applications (HDD, cell phone)
 Low power consumption, low cost
 High performance within power, cost
constraints (MIPS/mw, MIPS/$)
 Fixed point & floating point
 Better compilers - but users must be
informed
 Hybrid DSP/ GP systems