Pipelining and Vector Processing 1

PIPELINING AND VECTOR PROCESSING

• Parallel Processing

• Pipelining

• Arithmetic Pipeline

• Instruction Pipeline

• RISC Pipeline

• Vector Processing

• Array Processors
Pipelining and Vector Processing 2 Parallel Processing

PARALLEL PROCESSING

Execution of Concurrent Events in the computing

process to achieve faster Computational Speed

Levels of Parallel Processing

- Job or Program level

- Task or Procedure level

- Inter-Instruction level

- Intra-Instruction level
Pipelining and Vector Processing 3

•Simultaneous data processing tasks for the purpose of increasing the computational speed
•Perform concurrent data processing to achieve faster execution time
•Multiple Functional Unit :
Separate the execution unit into eight functional units operating in parallel

Adder-subtractor

Integer m ultiply

Logic unit

Shift unit

To Mem ory

Increm enter
Processor
registers
Floatint-point
add-subtract

Floatint-point
m ultiply

Floatint-point
divide
Pipelining and Vector Processing 4 Parallel Processing

PARALLEL COMPUTERS
Architectural Classification

– Flynn's classification
» Based on the multiplicity of Instruction Streams and
Data Streams
» Instruction Stream
• Sequence of Instructions read from memory
» Data Stream
• Operations performed on the data in the processor

Number of Data Streams

Single Multiple

Number of Single SISD SIMD

Instruction
Streams Multiple MISD MIMD
Pipelining and Vector Processing 5 Parallel Processing

SISD COMPUTER SYSTEMS

Control Processor Data stream

Memory
Unit Unit

Instruction stream

Characteristics
- Standard von Neumann machine
- Instructions and data are stored in memory
- One operation at a time

Limitations
Von Neumann bottleneck

Maximum speed of the system is limited by the

Memory Bandwidth (bits/sec or bytes/sec)

- Limitation on Memory Bandwidth

- Memory is shared by CPU and I/O
Pipelining and Vector Processing 6 Parallel Processing

SISD PERFORMANCE IMPROVEMENTS

• Multiprogramming

• Multifunction processor
• Pipelining
• Exploiting instruction-level parallelism
Pipelining and Vector Processing 7 Parallel Processing

MISD COMPUTER SYSTEMS

M CU P

M CU P Memory
• •
• •
• •

M CU P Data stream

Instruction stream

Characteristics
- There is no computer at present that can be
classified as MISD
Pipelining and Vector Processing 8 Parallel Processing

SIMD COMPUTER SYSTEMS

Memory
Data bus

Control Unit
Instruction stream

P P ••• P Processor units

Data stream

Alignment network

M M ••• M Memory modules

Characteristics
- Only one copy of the program exists
- A single controller executes one instruction at a time
Pipelining and Vector Processing 9 Parallel Processing

MIMD COMPUTER SYSTEMS

P M P M ••• P M

Interconnection Network

Shared Memory

Characteristics
- Multiple processing units

- Execution of multiple instructions on multiple data

Types of MIMD computer systems

- Shared memory multiprocessors

- Message-passing multicomputers
Pipelining and Vector Processing 10 Pipelining

PIPELINING
A technique of decomposing a sequential process
into suboperations, with each subprocess being
executed in a partial dedicated segment that
operates concurrently with all other segments.
Ai * Bi + Ci for i = 1, 2, 3, ... , 7
Ai Bi Memory Ci
Segment 1
R1 R2

Multiplier
Segment 2

R3 R4

Adder
Segment 3

R5

R1  Ai, R2  Bi Load Ai and Bi

R3  R1 * R2, R4  Ci Multiply and load Ci
R5  R3 + R4 Add
Pipelining and Vector Processing 11 Pipelining

OPERATIONS IN EACH PIPELINE STAGE

Clock Segment 1 Segment 2 Segment 3

Pulse
Number R1 R2 R3 R4 R5
1 A1 B1
2 A2 B2 A1 * B1 C1
3 A3 B3 A2 * B2 C2 A1 * B1 + C1
4 A4 B4 A3 * B3 C3 A2 * B2 + C2
5 A5 B5 A4 * B4 C4 A3 * B3 + C3
6 A6 B6 A5 * B5 C5 A4 * B4 + C4
7 A7 B7 A6 * B6 C6 A5 * B5 + C5
8 A7 * B7 C7 A6 * B6 + C6
9 A7 * B7 + C7
Pipelining and Vector Processing 12 Pipelining

GENERAL PIPELINE
General Structure of a 4-Segment Pipeline
Clock

Input S1 R1 S2 R2 S3 R3 S4 R4

Space-Time Diagram
1 2 3 4 5 6 7 8 9 Clock cycles
Segment 1 T1 T2 T3 T4 T5 T6
2 T1 T2 T3 T4 T5 T6
3 T1 T2 T3 T4 T5 T6
4 T1 T2 T3 T4 T5 T6
Pipelining and Vector Processing 13 Pipelining

PIPELINE SPEEDUP
n: Number of tasks to be performed

Conventional Machine (Non-Pipelined)

tn: Clock cycle
: Time required to complete the n tasks
 = n * t n

Pipelined Machine (k stages)

tp: Clock cycle (time to complete each suboperation)
: Time required to complete the n tasks
 = (k + n - 1) * tp

Speedup
Sk: Speedup

Sk = n*tn / (k + n - 1)*tp
tn
lim Sk = ( = k, if tn = k * tp )
n tp
Pipelining and Vector Processing 14 Pipelining

PIPELINE AND MULTIPLE FUNCTION UNITS

Example
- 4-stage pipeline
- subopertion in each stage; tp = 20nS
- 100 tasks to be executed
- 1 task in non-pipelined system; 20*4 = 80nS

Pipelined System
(k + n - 1)*tp = (4 + 99) * 20 = 2060nS

Non-Pipelined System
n*k*tp = 100 * 80 = 8000nS

Speedup
Sk = 8000 / 2060 = 3.88

4-Stage Pipeline is basically identical to the system

Ii Ii+1 I i+2 I i+3
with 4 identical function units

Multiple Functional Units P1 P2 P3 P4

Pipelining and Vector Processing 15 Arithmetic Pipeline

ARITHMETIC PIPELINE
Floating-point adder Exponents
a b
Mantissas
A B
X = A x 2a
Y = B x 2b R R

[1] Compare the exponents Compare Difference

Segment 1: exponents
[2] Align the mantissa by subtraction
[3] Add/sub the mantissa
[4] Normalize the result
R

Segment 2: Choose exponent Align mantissa

Segment 3: Add or subtract

mantissas

R R

Segment 4: Adjust Normalize

exponent result

R R
Pipelining and Vector Processing 16 Instruction Pipeline

INSTRUCTION CYCLE
Six Phases* in an Instruction Cycle
[1] Fetch an instruction from memory
[2] Decode the instruction
[3] Calculate the effective address of the operand
[4] Fetch the operands from memory
[5] Execute the operation
[6] Store the result in the proper place

* Some instructions skip some phases

* Effective address calculation can be done in
the part of the decoding phase
* Storage of the operation result into a register
is done automatically in the execution phase

==> 4-Stage Pipeline

[1] FI: Fetch an instruction from memory

[2] DA: Decode the instruction and calculate
the effective address of the operand
[3] FO: Fetch the operand
[4] EX: Execute the operation
Pipelining and Vector Processing 17 Instruction Pipeline

INSTRUCTION PIPELINE

Execution of Three Instructions in a 4-Stage Pipeline

Conventional

i FI DA FO EX

i+1 FI DA FO EX

i+2 FI DA FO EX

Pipelined

i FI DA FO EX
i+1 FI DA FO EX
i+2 FI DA FO EX
Pipelining and Vector Processing 18 Instruction Pipeline

INSTRUCTION EXECUTION IN A 4-STAGE PIPELINE

Segment1: Fetch instruction

from memory

Decode instruction
Segment2: and calculate
effective address

yes Branch?
no
Segment3: Fetch operand
from memory

Segment4: Execute instruction

Interrupt yes
Interrupt?
handling
no
Update PC

Empty pipe
Step: 1 2 3 4 5 6 7 8 9 10 11 12 13
Instruction 1 FI DA FO EX
2 FI DA FO EX
(Branch) 3 FI DA FO EX
4 FI FI DA FO EX
5 FI DA FO EX
6 FI DA FO EX
7 FI DA FO EX
Pipelining and Vector Processing 19 Instruction Pipeline

MAJOR HAZARDS IN PIPELINED EXECUTION

Structural hazards(Resource Conflicts)
Hardware Resources required by the instructions in
simultaneous overlapped execution cannot be met
Data hazards (Data Dependency Conflicts)
An instruction scheduled to be executed in the pipeline requires the
result of a previous instruction, which is not yet available
R1 <- B + C ADD DA B,C + Data dependency

R1 <- R1 + 1
INC DA bubble R1 +1

Control hazards
Branches and other instructions that change the PC
make the fetch of the next instruction to be delayed
JMP ID PC + PC Branch address dependency

bubble IF ID OF OE OS

Hazards in pipelines may make it Pipeline Interlock:

necessary to stall the pipeline Detect Hazards Stall until it is cleared
Pipelining and Vector Processing 20 Instruction Pipeline

STRUCTURAL HAZARDS
Structural Hazards
Occur when some resource has not been
duplicated enough to allow all combinations
of instructions in the pipeline to execute

Example: With one memory-port, a data and an instruction fetch

cannot be initiated in the same clock
i FI DA FO EX

i+1 FI DA FO EX

i+2 stall stall FI DA FO EX

The Pipeline is stalled for a structural hazard

<- Two Loads with one port memory
-> Two-port memory will serve without stall
Pipelining and Vector Processing 21 Instruction Pipeline

DATA HAZARDS
Data Hazards

Occurs when the execution of an instruction

depends on the results of a previous instruction
ADD R1, R2, R3
SUB R4, R1, R5
Data hazard can be dealt with either hardware
techniques or software technique
Hardware Technique

Interlock
- hardware detects the data dependencies and delays the scheduling
of the dependent instruction by stalling enough clock cycles
Forwarding (bypassing, short-circuiting)
- Accomplished by a data path that routes a value from a source
(usually an ALU) to a user, bypassing a designated register. This
allows the value to be produced to be used at an earlier stage in the
pipeline than would otherwise be possible

Software Technique
Instruction Scheduling(compiler) for delayed load
Pipelining and Vector Processing 22 Instruction Pipeline

FORWARDING HARDWARE

Example: Register
file
ADD R1, R2, R3
SUB R4, R1, R5

3-stage Pipeline MUX MUX Bypass

path
Result
I: Instruction Fetch write bus
A: Decode, Read Registers, ALU
ALU Operations
E: Write the result to the
destination register R4

ALU result buffer

ADD I A E

SUB I A E Without Bypassing

SUB I A E With Bypassing

Pipelining and Vector Processing 23 Instruction Pipeline

INSTRUCTION SCHEDULING
a = b + c;
d = e - f;

Unscheduled code: Scheduled Code:

LW Rb, b LW Rb, b
LW Rc, c LW Rc, c
ADD Ra, Rb, Rc LW Re, e
SW a, Ra ADD Ra, Rb, Rc
LW Re, e LW Rf, f
LW Rf, f SW a, Ra
SUB Rd, Re, Rf SUB Rd, Re, Rf
SW d, Rd SW d, Rd

Delayed Load
A load requiring that the following instruction not use its result
Pipelining and Vector Processing 24 Instruction Pipeline

CONTROL HAZARDS
Branch Instructions

- Branch target address is not known until

the branch instruction is completed
Branch
FI DA FO EX
Instruction
Next FI DA FO EX
Instruction

Target address available

- Stall -> waste of cycle times

Dealing with Control Hazards

* Prefetch Target Instruction

* Branch Target Buffer
* Loop Buffer
* Branch Prediction
* Delayed Branch
Pipelining and Vector Processing 25 Instruction Pipeline

CONTROL HAZARDS
Prefetch Target Instruction
– Fetch instructions in both streams, branch not taken and branch taken
– Both are saved until branch branch is executed. Then, select the right
instruction stream and discard the wrong stream
Branch Target Buffer(BTB; Associative Memory)
– Entry: Addr of previously executed branches; Target instruction
and the next few instructions
– When fetching an instruction, search BTB.
– If found, fetch the instruction stream in BTB;
– If not, new stream is fetched and update BTB
Loop Buffer(High Speed Register file)
– Storage of entire loop that allows to execute a loop without accessing memory
Branch Prediction
– Guessing the branch condition, and fetch an instruction stream based on
the guess. Correct guess eliminates the branch penalty
Delayed Branch
– Compiler detects the branch and rearranges the instruction sequence
by inserting useful instructions that keep the pipeline busy
in the presence of a branch instruction
Pipelining and Vector Processing 26 RISC Pipeline

RISC PIPELINE
RISC
- Machine with a very fast clock cycle that
executes at the rate of one instruction per cycle
<- Simple Instruction Set
Fixed Length Instruction Format
Register-to-Register Operations

Instruction Cycles of Three-Stage Instruction Pipeline

Data Manipulation Instructions
I: Instruction Fetch
A: Decode, Read Registers, ALU Operations
E: Write a Register

Load and Store Instructions

I: Instruction Fetch
A: Decode, Evaluate Effective Address
E: Register-to-Memory or Memory-to-Register

Program Control Instructions

I: Instruction Fetch
A: Decode, Evaluate Branch Address
E: Write Register(PC)
Pipelining and Vector Processing 27 RISC Pipeline

DELAYED LOAD
LOAD: R1  M[address 1]
LOAD: R2  M[address 2]
ADD: R3  R1 + R2
STORE: M[address 3]  R3
Three-segment pipeline timing
Pipeline timing with data conflict

clock cycle 1 2 3 4 5 6
Load R1 I A E
Load R2 I A E
Add R1+R2 I A E
Store R3 I A E

Pipeline timing with delayed load

clock cycle 1 2 3 4 5 6 7
Load R1 I A E
The data dependency is taken
Load R2 I A E care by the compiler rather
NOP I A E than the hardware
Add R1+R2 I A E
Store R3 I A E
Pipelining and Vector Processing 28 RISC Pipeline

DELAYED BRANCH
Compiler analyzes the instructions before and after
the branch and rearranges the program sequence by
inserting useful instructions in the delay steps

Using no-operation instructions

Clock cycles: 1 2 3 4 5 6 7 8 9 10
1. Load I A E
2. Increment I A E
3. Add I A E
4. Subtract I A E
5. Branch to X I A E
6. NOP I A E
7. NOP I A E
8. Instr. in X I A E

Rearranging the instructions

Clock cycles: 1 2 3 4 5 6 7 8
1. Load I A E
2. Increment I A E
3. Branch to X I A E
4. Add I A E
5. Subtract I A E
6. Instr. in X I A E
Pipelining and Vector Processing 29 Vector Processing

VECTOR PROCESSING
Vector Processing Applications
• Problems that can be efficiently formulated in terms of vectors
– Long-range weather forecasting
– Petroleum explorations
– Seismic data analysis
– Medical diagnosis
– Aerodynamics and space flight simulations
– Artificial intelligence and expert systems
– Mapping the human genome
– Image processing

Vector Processor (computer)

Ability to process vectors, and related data structures such as matrices
and multi-dimensional arrays, much faster than conventional computers

Vector Processors may also be pipelined

Pipelining and Vector Processing 30 Vector Processing

VECTOR PROGRAMMING

DO 20 I = 1, 100
20 C(I) = B(I) + A(I)

Conventional computer

Initialize I = 0
20 Read A(I)
Read B(I)
Store C(I) = A(I) + B(I)
Increment I = i + 1
If I  100 goto 20

Vector computer

C(1:100) = A(1:100) + B(1:100)

Pipelining and Vector Processing 31 Vector Processing

VECTOR INSTRUCTION FORMAT

Vector Instruction Format

Operation Base address Base address Base address Vector
code source 1 source 2 destination length

Pipeline for Inner Product

Source
A

Source Multiplier Adder

B pipeline pipeline
Pipelining and Vector Processing 32

– Vector Instruction Format

Operation Base address Base address Base address Vector

code source 1 source 2 destination length

ADD A B C 100

– Matrix Multiplication
» 3 x 3 matrices multiplication : n2 = 9 inner product

 a11 a12 a13   b11 b12 b13   c11 c12 c13 

a a a23   b21 b22 b23   c21 c22 c23 
 21 22     
a31 a32 a33  b31 b32 b33  c31 c32 c33 
•
c•
11  a11 b11  a12 b21  a13 b31: inner product 9
» Cumulative multiply-add operation : n3 = 27 multiply-add
•
c  c  a b

    
c11  c11  a11 b11  a12 b21  a13 b31
– multiply-add 3
C11 = 0 – 9 X 3 multiply-add = 27
Pipelining and Vector Processing 33

– Pipeline for calculating an inner product :

» Floating point multiplier pipeline : 4 segment

C  A1B1  A2 B2  A3 B3    Ak Bk
after 1st clock input

Source
A
» after 4th clock input
Source
A1B1 A

A4B4 A3B3 A2B2 A1B1

Source Multiplier Adder
B pipeline pipeline
Source Multiplier Adder
• after 8th clock input B pipeline pipeline

» after 9th, 10th, 11th ,...

Source
Source A
A
A8B8 A7B7 A6B6 A5B5 A4B4 A3B3 A2B2 A1B1
A8B8 A7B7 A6B6 A5B5 A4B4 A3B3 A2B2 A1B1
Source Multiplier Adder
Source Multiplier Adder B pipeline pipeline
B pipeline pipeline

• C  section
Four A1B1  summation
A5 B5  A9 B9  A13B13   A2 B2  A6 B6 A1B1  A5 B5
 A2 B2  A6 B6  A10 B10  A14 B14   ,,,
 A3 B3  A7 B7  A11B11  A15B15  
 A4 B4  A8 B8  A12 B12  A16 B16  
Pipelining and Vector Processing 34

Address bus

– Memory Interleaving : AR AR AR AR

» Simultaneous access to memory from two or

more source using one memory bus system Memory
array
Memory
array
Memory
array
Memory
array

» AR 2 bit memory module

» Even / Odd Address Memory Access DR DR DR DR

 Supercomputer Data bus

 Supercomputer = Vector Instruction + Pipelined floating-point arithmetic

 Performance Evaluation Index
» MIPS : Million Instruction Per Second
» FLOPS : Floating-point Operation Per Second
 megaflops : 106, gigaflops : 109

 Cray supercomputer : Cray Research

» Clay-1 : 80 megaflops, 4 million 64 bit words memory
» Clay-2 : 12 times more powerful than the clay-1
 VP supercomputer : Fujitsu
» VP-200 : 300 megaflops, 32 million memory, 83 vector instruction, 195 scalar
instruction
» VP-2600 : 5 gigaflops
Pipelining and Vector Processing 35

Superscalar operation

• Pipelining enables multiple instructions to be

executed concurrently by dividing the execution of
an instruction into several stages:
– Instructions enter the pipeline in strict program order.
– If the pipeline does not stall, one instruction enters the pipeline and
one instruction completes execution in one clock cycle.
– Maximum throughput of a pipelined processor is one instruction
per clock cycle.
• An alternative approach is to equip the processor
with multiple processing units to handle several
instructions in parallel in each stage.

35
Pipelining and Vector Processing 36

Superscalar operation (contd..)

• If a processor has multiple processing units then

several instructions can start execution in the same
clock cycle.
– Processor is said to use “multiple issue”.
• These processors are capable of achieving
instruction execution throughput of more than one
instruction per cycle.
• These processors are known as “superscalar
processors”.

36
Pipelining and Vector Processing 37

Superscalar operation (contd..)

Instruction fetch unit is capable of reading two
instructions at a time and storing them in the
F : Instruction instruction queue.
fetch unit
Instruction queue

Dispatch unit fetches Processor has two execution units:

and retrieves up to Integer and Floating Point
two instructions at
a time from the
front of the queue. Floating-
point
unit
Dispatch
unit W : Write
results

Integer
If there is one integer and one unit
floating point instruction, and no
hazards, then both instructions are
dispatched in the same clock cycle.
37
Pipelining and Vector Processing 38

Superscalar operation (contd..)

• Various hazards cause a even greater deterioration in

performance in case of a superscalar processor.
• Compiler can avoid many hazards by careful
ordering of instructions:
– For example, the compiler should try to interleave floating-point
and integer instructions.
– Dispatch unit can then dispatch two instructions in most clock
cycles, and keep both integer and floating point units busy most of
the time.
• If the compiler can order instructions in such a way
that the available hardware units can be kept busy
most of the time, high performance can be achieved.

38
Pipelining and Vector Processing 39

Superscalar operation (contd..)

Clock cycle 1 2 3 4 5 6 7
I1 (Fadd) F1 D1 E1A E1B E1C W1

I2 (Add) F2 D2 E2 W2

I3 (Fsub) F3 D3 E3 E3 E3 W3

I4 (Sub) F4 D4 E4 W4
•Instructions in the floating-point unit take three cycles to execute.
•Floating-point unit is organized as a three-stage pipeline.
•Instructions in the integer unit take one cycle to execute.
•Integer unit is organized as a single-stage pipeline.
•Clock cycle 1:
- Instructions I1 (floating point) and I2 (integer) are fetched.
•Clock cycle 2:
- Instructions I1 and I2 are decoded and dispatched, I3 is fetched.

39
Pipelining and Vector Processing 40

Superscalar operation (contd..)

Clock cycle 1 2 3 4 5 6 7
I1 (Fadd) F1 D1 E1A E1B E1C W1

I2 (Add) F2 D2 E2 W2

I3 (Fsub) F3 D3 E3 E3 E3 W3

I4 (Sub) F4 D4 E4 W4
•Clock cycle 3:
- I1 and I2 begin execution, I2 completes execution. I3 is dispatched to floating
- point unit and I4 is dispatched to integer unit.
Clock cycle 4:
- I1 continues execution, I3 begins execution, I2 completes Write stage,
I4 completes execution.
Clock cycle 5:
- I1 completes execution, I3 continues execution, and I4 completes Write.
Order of completion is I2, I4, I1, I3

40
Pipelining and Vector Processing 41

• Array Processors
– Performs computations on large arrays of data

Vector processing : Adder/Multiplier pipeline

Array processing :array processor
– Array Processing
» Attached array processor :
• Auxiliary processor attached to a general purpose computer
» SIMD array processor :
• Computer with multiple processing units operating in parallel
– Vector C = A + B ci = ai + bi

PE 1 M1

M aster control
unit
PE 2 M2
General-purpose Input-Output Attached array
com puter interface Processor
PE 3 M3

High-speed m em ory to-

Main m em ory Local m em ory
m em ory bus

M ain m em ory
PE n Mn
Pipelining and Vector Processing 42 Parallel Processing
COMPUTER ARCHITECTURES FOR PARALLEL
PROCESSING
Von-Neuman SISD Superscalar processors
based
Superpipelined processors

VLIW

MISD Nonexistence

SIMD Array processors

Systolic arrays
Dataflow
Associative processors

MIMD Shared-memory multiprocessors

Reduction
Bus based
Crossbar switch based
Multistage IN based

Message-passing multicomputers

Hypercube
Mesh
Reconfigurable
Pipelining and Vector Processing 43 Parallel Processing

TYPES OF SIMD COMPUTERS

Array Processors
- The control unit broadcasts instructions to all PEs,
and all active PEs execute the same instructions
- ILLIAC IV, GF-11, Connection Machine, DAP, MPP

Systolic Arrays
- Regular arrangement of a large number of
very simple processors constructed on
VLSI circuits
- CMU Warp, Purdue CHiP

Associative Processors
- Content addressing
- Data transformation operations over many sets
of arguments with a single instruction
- STARAN, PEPE