ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 1

Advances in Computer Architecture

Subject Code : 12SCS23 IA Marks : 50
No of Lecture Hrs/Week : 4 Exam hours : 3
Total No of Lecture Hours : 52 Exam Marks : 100

1. Introduction and Review of Fundamentals of Computer Design: Introduction; Classes
computers; Defining computer architecture; Trends in Technology; Trends in power in Integrated
Circuits; Trends in cost; Dependability, Measuring, reporting and summarizing Performance;
Quantitative Principles of computer design; Performance and Price-Performance; Fallacies and
pitfalls; Case studies.

2. Some topics in Pipelining, Instruction Level Parallelism, Its Exploitation and Limits on ILP:
Introduction to pipelining, ILP; Crosscutting issues, fallacies, and pitfalls with respect to
pipelining; Basic concepts and challenges of ILP; Case study of Pentium 4, Fallacies and pitfalls.
Introduction to limits in ILP; Performance and efficiency in advanced multiple-issue processors.

3. Memory Hierarchy Design, Storage Systems: Review of basic concepts; Crosscutting issues in
the design of memory hierarchies; Case study of AMD Opteron memory hierarchy; Fallacies and
pitfalls in the design of memory hierarchies. Introduction to Storage Systems; Advanced topics in
disk storage; Definition and examples of real faults and failures; I/O performance, reliability
measures, and benchmarks; Queuing theory; Crosscutting issues; Designing and evaluating an I/O
system The Internet archive cluster; Case study of NetAA FAS6000 filer; Fallacies and pitfalls.

4. Hardware and Software for VLIW and EPIC Introduction: Exploiting Instruction-Level
Parallelism Statically, Detecting and Enhancing Loop-Level Parallelism, Scheduling and
Structuring Code for Parallelism, Hardware Support for Exposing Parallelism: Predicated
Instructions, Hardware Support for Compiler Speculation, The Intel IA-64 Architecture and
Itanium Processor, Concluding Remarks.

5. Large-Scale Multiprocessors and Scientific Applications: Introduction, Interprocessor
Communication: The Critical Performance Issue, Characteristics of Scientific Applications,
Synchronization: Scaling Up, Performance of Scientific Applications on Shared-Memory
Multiprocessors, Performance Measurement of Parallel Processors with Scientific Applications,
Implementing Cache Coherence, The Custom Cluster Approach: Blue Gene/L, Concluding
Remarks.
6. Computer Arithmetic Introduction, Basic Techniques of Integer Arithmetic, Floating Point,
Floating-Point Multiplication, Floating-Point Addition, Division and Remainder, More on
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 2

Floating-Point Arithmetic, Speeding Up Integer Addition, Speeding Up Integer Multiplication
and Division, Fallacies and Pitfalls.

Text Book:
1. Hennessey and Patterson: Computer Architecture A Quantitative Approach, 4th Edition,
Elsevier, 2007.

Reference Books:
1. Kai Hwang: Advanced Computer Architecture - Parallelism, Scalability, Programmability, 2nd
Edition, Tata McGraw Hill, 2010.






















ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 3


CONTENTS
UNITS PAGE NO.
1 Fundamentals of computer design 4
2 Pipelining, Instruction level parallelism and Limitations 39
3 Multiprocessors and thread level parallelism 80
4 Hardware and software for VLIW and EPIC 144
5 Large scale multiprocessors and scientific application 168
6 Computer Arithmetic 189






















ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 4

UNIT I
FUNDAMENTALS OF COMPUTER DESIGN
Introduction
Today s desktop computers (less than $500 cost) ar e having more
performance, larger memory and storage than a computer bought in 1085 for 1
million dollar. Highest performance microprocessors of today outperform
Supercomputers of less than 10 years ago. The rapid improvement has come both
from advances in the technology used to build computers and innovations made in
the computer design or in other words, the improvement made in the computers
can be attributed to innovations of technology and architecture design.
During the first 25 years of electronic computers, both forces made a
major contribution, delivering performance improvement of about 25% per year.
Microprocessors were evolved during late 1970s and their ability along with
improvements made in the Integrated Circuit (IC) technology y contributed to 35%
performance growth per year.
The virtual elimination of assembly language programming reduced the n eed
for object-code compatibility. The creation of standardized vendor-independent
operating system lowered the cost and risk of bringing out a new architecture.
In the yearly 1980s, the Reduced Instruction Set Computer (RISC) based
machines focused the attention of designers on two critical performance techniques,
the exploitation Instruction Level Parallelism (ILP) and the use of caches. The figu
re 1.1 shows the growth in processor performance since the mid 1980s. The graph
plots performance relative to the VAX-11/780 as measured b y the SPECint
benchmarks. From the figure it is clear that architectural and organizational
enhancements led to 16 years of sustained growth in performance at an annual rate of
over 50%. Since 2002, processor performance improvement has dropped to about 20%
per year due to the following hurdles:
Maximum power dissipation of air-cooled chips
Little ILP left to exploit efficiently
Limitations laid by memory latency
The hurdles signals historic switch from relying solely on ILP to Thread Level
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 5

Parallelism (TLP) and Data Level Parallelism (DLP).

Figure 1.1 The evolution of various classes of computers:
Classes of Computers
1960: Large Main frames (Millions of $ )
(Applications: Business Data processing, large Scientific computin g)
1970: Minicomputers (Scientific laboratories, Time sharing concepts)
1980: Desktop Computers (Ps) in the form of Personal computers and workstations.
(Larger Memory, more computing power, Replaced Time sharing g systems)
1990: Emergence of Internet and WWW, PDAs, emergence of high performance digital
consumer electronics
2000: Cell phones
These changes in computer use have led to three different computing classes each
characterized by different applications, requirements and computing technologies.owth in
processor performance since 1980s


ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 6

Desktop computing
The first and still the largest market in dollar terms is desktop computing. Desktop
computing system cost range from $ 500 (low end) to $ 5000 (high-end
configuration). Throughout this range in price, the desktop market tends to drive to
optimize price- performance. The perf ormance concerned is compute performance
and graphics performance. The combination of performance and price are the
driving factors to the customers and the computer designer. Hence, the newest,
high performance and cost effective processor often appears first in desktop computers.
Servers:
Servers provide large-scale and reliable computing and file services and are
mainly used in the large-scale en terprise computing and web based services. The three
important
Characteristics of servers are:
Dependability: Severs must operate 24x7 hours a week. Failure of server system
is far more catastrophic than a failure of desktop. Enterprise will lose revenue if
the server is unavailable.
Scalability: as the business grows, the server may have to provide more
functionality/ services. Thus ability to scale up the computin g capacity, memory,
storage and I/O bandwidth is crucial.
Throughput: transactions completed per minute or web pages served per second
are crucial for servers.
Embedded Computers
Simple embedded microprocessors are seen in washing machines, printers,
network switches, handheld devices such as cell phones, smart cards video game
devices etc. embedded computers have the widest spread of processing power and
cost. The primary goal is often meeting the performance need at a minimum price
rather than achieving higher performance at a higher price. The other two characteristic
requirements are to minimize the memory and power.
In many embedded applications, the memory can be substantial portion of
the systems cost and it is very important to optimize the memory size in such
cases. The application is expected to fit totally in the memory on the p rocessor
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 7

chip or off chip memory. The importance of memory size translates to an emphasis
on code size which is dictated by the application. Larger memory consumes more
power. All these aspects are considered while choosing or designing processor for the
embedded applications.
Defining Computer Arch itecture: The computer designer has to ascertain the
attributes that are important for a new computer and design the system to maximize
the performance while staying within cost, power and availability constraints. The task
has few important aspects such as Instruction Set design, Functional organization, Logic
design and implementation.
Instruction Set Architecture (ISA)
ISA refers to the actual programmer visible Instruction set. The ISA serves as
boundary between the software and hardware. Th e seven dimensions of the ISA are:
i) Class of ISA: Nearly all ISAs today ar e classified as General-Purpose- Register
architectures. The operands are either Registers or Memory locations.
The two popular versions of this class are:
Register-Memory ISAs : ISA of 80x86, can access memory as part of many
instructions. Load -Store ISA Eg. ISA of MIPS, can access memory only with
Load or Store instructions.
ii) Memory addressing: Byte addressing scheme is most widely used in all
desktop and server computers. Both 80x86 and MIPS use byte addressing. Incase
of MIPS the object must be aligned. An access to an object of s byte at byte address A
is aligned if A mod s =0. 80x86 does not require alignment. Accesses are faster if
operands are aligned.

iii) Addressing modes:Specify the address of a M object apart from register and constant
operands. MIPS Addressing modes:
Register mode addressing
Immediate mode addressing
Displacement mode addressing
80x86 in addition to the above addressing modes supports the additional modes of
addressing:
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 8

i. Register Indirect
ii. Indexed
iii,Based with Scaled index
iv) Types and sizes of operands:
MIPS and x86 support:
8 bit (ASCII character), 16 bit(Unicode character)
32 bit (Integer/word )
64 bit (long integer/ Double word)
32 bit (IEEE-754 floating point)
64 bit (Double precision floating point)
80x86 also supports 80 bit floating point operand.(extended double Precision)
v)Operations:The general category o f operations are:
Data Transfer
Arithmetic operations
Logic operations
Control operations
MIPS ISA: simple & easy to implement
x86 ISA: richer & larger set of operations
vi) Control flow instructions:All ISAs support:
Conditional & Unconditional Branches
Procedure C alls & Returns MIPS 80x86
Conditional Branches tests content of Register Condition code bits
Procedure C all JAL CALLF
Return Address in a R Stack in M
vii) Encoding an ISA
Fixed Length ISA Variable Length ISA
MIPS 32 Bit long 80x86 (1-18 bytes)
Simplifies decoding Takes less space

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 9

Number of Registers and number of Addressing modes have significant impact
on the length of instruction as the register field and addressing mode field can
appear many times in a single instruction.
Trends in Technology
The designer must be aware of the following rapid changes in implementation
technology.
Integrated C ircuit (IC) Logic technology
Memory technology (semiconductor DRAM technology)
Storage o r magnetic disk technology
Network technology

IC Logic technology:
Transistor density increases by about 35%per year. Increase in die size
corresponds to about 10 % to 20% per year. The combined effect is a growth rate
in transistor count on a chip of about 40% to 55% per year. Semiconductor DRAM
technology:cap acity increases by about 40% per year.
Storage Technology:
Before 1990: the storage density increased by about 30% per year.
After 1990: the storage density increased by about 60 % per year.
Disks are still 50 to 100 times cheaper per bit than DRAM.

Network Technology:
Network performance depends both on the performance of the switches and
on the performance of the transmission system. Although the technology improves
continuously, the impact of these improvements can be in discrete leaps.

Performance trends: Bandwidth or throughput is the total amount of work done in given
time.
Latency or response time is the time between the start and the completion of an
event. (for eg. Millisecond for disk access)

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 10




A simple rule of thumb is that bandwidth gro ws by at least the square of the improvement
in latency. Computer designers should make plans accordingly.
IC Processes are characterizes by the f ature sizes.
Feature sizes decreased from 10 microns(1971) to 0.09 microns(2006)
Feature sizes shrink, devices shrink quadr atically.
Shrink in vertical direction makes the operating v oltage of the transistor to
reduce.
Transistor performance improves linearly with decreasing feature size
. Transistor count improves quadratically with a linear improvement in Transistor
performance.
!!! Wire delay scales poorly compared to Transistor performance.
Feature sizes shrink, wires get shorter.
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 11

Signal delay fo r a wire increases in proportion to the product of Resistance and
Capacitance.

Trends in Power in Integrated Circuits
For CMOS chips, the dominant source of energy consumption is due to switching
transistor, also called as Dynamic power and is given b y the following equation.
Power = (1/2)*Capacitive load* Voltage

* Frequency switched dynamic
For mobile devices, energy is the better metric

Energy dynamic = Capacitive load x Voltage 2
For a fix ed task, slowing clock rate (frequency switched) reduces power, but not energy
Capacitive load a function of number of transistors connected to output and technology,
which determines capacitance of wires and transistors
Dropping voltage helps both, so went from 5V down to 1V
To save energy & dynamic power, most CPUs now turn off clock of inactive modules
Distributing the power, removing the heat and preventing hot spots have become
increasingly difficult challenges.
The leakage current flows even when a transistor is off. Therefore static power is
equally important.

Power static= Current static * Voltage

Leakage current increases in processors with smaller transistor sizes
Increasing the number of transistors increases power even if they are turned off
In 2006, goal for leakage is 25% of total power consumption; high performance designs
at 40%
Very low power systems even gate voltage to inactive modules to control loss due to
leakage

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 12

Trends in Cost
The underlying principle that drives the cost down is the learning curve manufacturing
costs decrease over time.
Volume is a second key factor in determining cost. Volume decreases cost since it
increases purchasing manufacturing efficiency. As a rule of thumb, the cost decreases
about 10% for each doubling of volume.
Cost of an Integrated Circuit Although the cost of ICs have dropped exponentially, the
basic process of silicon manufacture is unchanged. A wafer is still tested and chopped
into dies that are
packaged.

Cost of IC = Cost of [die+ testing die+ Packaging and final test] / (Final test yoeld)
Cost of die = Cost of wafer/ (Die per wafer x Die yield)

The number of dies per wafer is approximately the area of the wafer divided by the area
of the die.
Die per wafer = [_ * (Wafer Dia/2)2/Die area]-[_* wafer dia/_(2*Die area)]

The first term is the ratio of wafer area to die area and the second term compensates for
the rectangular dies near the periphery of round wafers(as shown in figure).

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 13




Dependability:
The Infrastructure providers offer Service Level Agreement (SLA) or Service Level
Objectives (SLO) to guarantee that their networking or power services would be
dependable.
Systems alternate between 2 states of service with respect to an SLA:
1. Service accomplishment, where the service is delivered as specified in SLA
2. Service interruption, where the delivered service is different from the SLA
Failure = transition from state 1 to state 2
Restoration = transition from state 2 to state 1

The two main measures of Dependability are Module Reliability and Module
Availability. Module reliability is a measure of continuous service accomplishment (or
time to failure) from a reference initial instant.
1. Mean Time To Failure (MTTF) measures Reliability
2. Failures In Time (FIT) = 1/MTTF, the rate of failures
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 14

Traditionally reported as failures per billion hours of operation
Mean Time To Repair (MTTR) measures Service Interruption
Mean Time Between Failures (MTBF) = MTTF+MTTR
Module availability measures service as alternate between the 2 states of
accomplishment and interruption (number between 0 and 1, e.g. 0.9)
Module availability = MTTF / ( MTTF + MTTR)

Performance:
The Execution time or Response time is defined as the time between the start and
completion of an event. The total amount of work done in a given time is defined as the
Throughput.

The Administrator of a data center may be interested in increasing the Throughput. The
computer user may be interested in reducing the Response time.
Computer user says that computer is faster when a program runs in less time.



The routinely executed programs are the best candidates for evaluating the performance
of the new computers. To evaluate new system the user would simply compare the
execution time of their workloads.

Benchmarks
The real applications are the best choice of benchmarks to evaluate the
performance. However, for many of the cases, the workloads will not be known at the
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 15

time of evaluation. Hence, the benchmark program which resemble the real applications
are chosen. The three types of benchmarks are:
KERNELS, which are small, key pieces of real applications;
Toy Programs: which are 100 line programs from beginning programming
assignments, such Quicksort etc.,
Synthetic Benchmarks: Fake programs invented to try to match the profile and
behavior of real applications such as Dhrystone.
To make the process of evaluation a fair justice, the following points are to be followed.
Source code modifications are not allowed.
Source code modifications are allowed, but are essentially impossible.
Source code modifications are allowed, as long as the modified version produces
the same output.
To increase predictability, collections of benchmark applications, called
benchmark suites, are popular
SPECCPU: popular desktop benchmark suite given by Standard Performance
Evaluation committee (SPEC)
CPU only, split between integer and floating point programs
SPECint2000 has 12 integer, SPECfp2000 has 14 integer programs
SPECCPU2006 announced in Spring 2006.
SPECSFS (NFS file server) and SPECWeb (WebServer) added as server
benchmarks
Transaction Processing Council measures server performance and
costperformance for databases
TPC-C Complex query for Online Transaction Processing
TPC-H models ad hoc decision support
TPC-W a transactional web benchmark
TPC-App application server and web services benchmark

SPEC Ratio: Normalize execution times to reference computer, yielding a ratio
proportional to performance = time on reference computer/time on computer being rated

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 16

If program SPECRatio on Computer A is 1.25 times bigger than Computer B, then

Note : when comparing 2 computers as a ratio, execution times on the reference
computer drop out, so choice of reference computer is irrelevant.

Quantitative Principles of Computer Design
While designing the computer, the advantage of the following points can be exploited to
enhance the performance.
* Parallelism: is one of most important methods for improving performance.
- One of the simplest ways to do this is through pipelining ie, to overlap the
instruction Execution to reduce the total time to complete an instruction
sequence.
- Parallelism can also be exploited at the level of detailed digital design.
- Set- associative caches use multiple banks of memory that are typically searched
n parallel. Carry look ahead which uses parallelism to speed the process of
computing.

* Principle of locality: program tends to reuse data and instructions they have used
recently. The rule of thumb is that program spends 90 % of its execution time in only
10% of the code. With reasonable good accuracy, prediction can be made to find what
instruction and data the program will use in the near future based on its accesses in the
recent past.
* Focus on the common case while making a design trade off, favor the frequent case
over the infrequent case. This principle applies when determining how to spend
resources, since the impact of the improvement is higher if the occurrence is frequent.
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 17

Amdahls Law: Amdahls law is used to find the performance gain that can be obtained
by improving some portion or a functional unit of a computer Amdahls law defines the
speedup that can be gained by using a particular feature.
Speedup is the ratio of performance for entire task without using the enhancement
when possible to the performance for entire task without using the enhancement.
Execution time is the reciprocal of performance. Alternatively, speedup is defined as thee
ratio of execution time for entire task without using the enhancement to the execution
time for entair task using the enhancement when possible.
Speedup from some enhancement depends an two factors:
i. The fraction of the computation time in the original computer that can be converted to
take advantage of the enhancement. Fraction enhanced is always less than or equal to.
Example: If 15 seconds of the execution time of a program that takes 50 seconds in total
can use an enhancement, the fraction is 15/50 or 0.3
ii. The improvement gained by the enhanced execution mode; ie how much faster the task
would run if the enhanced mode were used for the entire program. Speedup enhanced is
the time of the original mode over the time of the enhanced mode and is always greater
then 1.


The Processor performance Equation:
Processor is connected with a clock running at constant rate. These discrete time events
are called clock ticks or clock cycle.
CPU time for a program can be evaluated:
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 18


Example:
A System contains Floating point (FP) and Floating Point Square Root (FPSQR) unit.
FPSQR is responsible for 20% of the execution time. One proposal is to enhance the
FPSQR hardware and speedup this operation by a factor of 15 second alternate is just to
try to make all FP instructions run faster by a factor of 1.6 times faster with the same
effort as required for the fast FPSQR, compare the two design alternative


ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 19






Pipelining: Basic and Intermediate concepts
Pipeline is an implementation technique that exploits parallelism among the instructions
in a sequential instruction stream. Pipeline allows to overlapping the execution of
multiple instructions. A Pipeline is like an assembly line each step or pipeline stage
completes a part of an instructions. Each stage of the pipeline will be operating an a
separate instruction. Instructions enter at one end progress through the stage and exit at
the other end. If the stages are perfectly balance. (assuming ideal conditions), then the
time per instruction on the pipeline processor is
given by the ratio:
Time per instruction on unpipelined machine/ Number of Pipeline stages
Under these conditions, the speedup from pipelining is equal to the number of stage
pipeline. In practice, the pipeline stages are not perfectly balanced and pipeline does
involve some overhead. Therefore, the speedup will be always then practically less than
the number of stages of the pipeline. Pipeline yields a reduction in the average execution
time per instruction. If the processor is assumed to take one (long) clock cycle per
instruction, then pipelining decrease the clock cycle time. If the processor is assumed to
take multiple CPI, then pipelining will aid to reduce the CPI.

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 20

A Simple implementation of a RISC instruction set
Instruction set of implementation in RISC takes at most 5 cycles without pipelining.
The 5 clock cycles are:
1. Instruction fetch (IF) cycle:
Send the content of program count (PC) to memory and fetch the current
instruction from memory to update the PC.


2. Instruction decode / Register fetch cycle (ID):
Decode the instruction and access the register file. Decoding is done in parallel
with reading registers, which is possible because the register specifies are at a fixed
location in a RISC architecture. This corresponds to fixed field decoding. In addition it
involves:
- Perform equality test on the register as they are read for a possible branch.
- Sign-extend the offset field of the instruction in case it is needed.
- Compute the possible branch target address.
3. Execution / Effective address Cycle (EXE)
The ALU operates on the operands prepared in the previous cycle and performs
one of the following function defending on the instruction type.


* Register- Register ALU instruction: ALU performs the operation specified in the
instruction using the values read from the register file.
* Register- Immediate ALU instruction: ALU performs the operation specified in the
instruction using the first value read from the register file and that sign extended
immediate.
4. Memory access (MEM)
For a load instruction, using effective address the memory is read. For a store
instruction memory writes the data from the 2nd register read using effective address.
5. Write back cycle (WB)
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 21

Write the result in to the register file, whether it comes from memory system (for
a LOAD instruction) or from the ALU.


Five stage Pipeline for a RISC processor
Each instruction taken at most 5 clock cycles for the execution
* Instruction fetch cycle (IF)
* Instruction decode / register fetch cycle (ID)
* Execution / Effective address cycle (EX)
* Memory access (MEM)
* Write back cycle (WB)

The execution of the instruction comprising of the above subtask can be pipelined. Each
of the clock cycles from the previous section becomes a pipe stage a cycle in the
pipeline. A new instruction can be started on each clock cycle which results in the
execution pattern shown figure 2.1. Though each instruction takes 5 clock cycles to
complete, during each clock cycle the hardware will initiate a new instruction and will be
executing some part of the five different instructions as illustrated in figure 2.1.

Each stage of the pipeline must be independent of the other stages. Also, two different
operations cant be performed with the same data path resource on the same clock. For
example, a single ALU cannot be used to compute the effective address and perform a
subtract operation during the same clock cycle. An adder is to be provided in the stage 1
to compute new PC value and an ALU in the stage 3 to perform the arithmetic indicatedin
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 22

the instruction (See figure 2.2). Conflict should not arise out of overlap of instructions
using pipeline. In other words, functional unit of each stage need to be independent of
other functional unit. There are three observations due to which the risk of conflict is
reduced.
Separate Instruction and data memories at the level of L1 cache eliminates a
conflict for a single memory that would arise between instruction fetch and data
access.
Register file is accessed during two stages namely ID stage WB. Hardware
should allow to perform maximum two reads one write every clock cycle.
To start a new instruction every cycle, it is necessary to increment and store the
PC every cycle.



Buffers or registers are introduced between successive stages of the pipeline so that at the
end of a clock cycle the results from one stage are stored into a register (see figure 2.3).
During the next clock cycle, the next stage will use the content of these buffers as input.
Figure 2.4 visualizes the pipeline activity.

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 23





Basic Performance issues in Pipelining

Pipelining increases the CPU instruction throughput but, it does not reduce the
executiontime of an individual instruction. In fact, the pipelining increases the execution
time of each instruction due to overhead in the control of the pipeline. Pipeline overhead
arises from the combination of register delays and clock skew. Imbalance among the pipe
stages reduces the performance since the clock can run no faster than the time needed for
the slowest pipeline stage.


ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 24



Pipeline Hazards

Hazards may cause the pipeline to stall. When an instruction is stalled, all the
instructions issued later than the stalled instructions are also stalled. Instructions issued
earlier than the stalled instructions will continue in a normal way. No new instructions
are fetched during the stall. Hazard is situation that prevents the next instruction in the
instruction stream from k executing during its designated clock cycle. Hazards will
reduce the pipeline performance.

Performance with Pipeline stall

A stall causes the pipeline performance to degrade from ideal performance. Performance
improvement from pipelining is obtained from:

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 25



Assume that,
i) cycle time overhead of pipeline is ignored
ii) stages are balanced
With theses assumptions

If all the instructions take the same number of cycles and is equal to the number of
pipeline stages or depth of the pipeline, then,


If there are no pipeline stalls,
Pipeline stall cycles per instruction = zero
Therefore,
Speedup = Depth of the pipeline.

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 26


Types of hazard
Three types hazards are:
1. Structural hazard
2. Data Hazard
3. Control Hazard


Structural hazard
Structural hazard arise from resource conflicts, when the hardware cannot support all
possible combination of instructions simultaneously in overlapped execution. If some
combination of instructions cannot be accommodated because of resource conflicts, the
processor is said to have structural hazard. Structural hazard will arise when some
functional unit is not fully pipelined or when some resource has not been duplicated
enough to allow all combination of instructions in the pipeline to execute. For example, if
memory is shared for data and instruction as a result, when an instruction contains data
memory reference, it will conflict with the instruction reference for a later instruction (as
shown in figure 2.5a). This will cause hazard and pipeline stalls for 1 clock cycle.


ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 27





ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 28



Pipeline stall is commonly called Pipeline bubble or just simply bubble

Data Hazard

Consider the pipelined execution of the following instruction sequence (Timing diagram
shown in figure 2.6)

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 29


DADD instruction produces the value of R1 in WB stage (Clock cycle 5) but the DSUB
instruction reads the value during its ID stage (clock cycle 3). This problem is called Data
Hazard. DSUB may read the wrong value if precautions are not taken. AND instruction
will read the register during clock cycle 4 and will receive the wrong results. The XOR
instruction operates properly, because its register read occurs in clock cycle 6 after
DADD writes in clock cycle 5. The OR instruction also operates without incurring a
hazard because the register file reads are performed in the second half of the cycle
whereas the writes are performed in the first half of the cycle.

Minimizing data hazard by Forwarding
The DADD instruction will produce the value of R! at the end of clock cycle 3. DSUB
instruction requires this value only during the clock cycle 4. If the result can be moved
from the pipeline register where the DADD store it to the point (input of LAU) where
DSUB needs it, then the need for a stall can be avoided. Using a simple hardware
technique called Data Forwarding or Bypassing or short circuiting, data can be made
available from the output of the ALU to the point where it is required (input of LAU) at
the beginning of immediate next clock cycle.
Forwarding works as follows:
i) The output of ALU from EX/MEM and MEM/WB pipeline register is always
feedback to the ALU inputs.
ii) If the Forwarding hardware detects that the previous ALU output serves as the
source for the current ALU operations, control logic selects the forwarded result as the
input rather than the value read from the register file. Forwarded results are required not
only from the immediate previous instruction, but also from an instruction that started 2
cycles earlier. The result of ith instruction Is required to be forwarded to (i+2)th
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 30

instruction also. Forwarding can be generalized to include passing a result directly to the
functional unit that requires it.


Data Hazard requiring stalls

LD R1, 0(R2)
DADD R3, R1, R4
AND R5, R1, R6
OR R7, R1, R8
The pipelined data path for these instructions is shown in the timing diagram (figure 2.7)


The LD instruction gets the data from the memory at the end of cycle 4. even with
forwarding technique, the data from LD instruction can be made available earliest during
clock cycle 5. DADD instruction requires the result of LD instruction at the beginning of
clock cycle 5. DADD instruction requires the result of LD instruction at the beginning of
clock cycle 4. This demands data forwarding of clock cycle 4. This demands data
forwarding in negative time which is not possible. Hence, the situation calls for a pipeline
stall.Result from the LD instruction can be forwarded from the pipeline register to the
and instruction which begins at 2 clock cycles later after the LD instruction. The load
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 31

instruction has a delay or latency that cannot be eliminated by forwarding alone. It is
necessary to stall pipeline by 1 clock cycle. A hardware called Pipeline interlock detects a
hazard and stalls the pipeline until the hazard is cleared. The pipeline interlock helps to
preserve the correct execution pattern by introducing a stall or bubble. The CPI for the
stalled instruction increases by the length of the stall. Figure 2.7 shows the pipeline
before and after the stall. Stall causes the DADD to move 1 clock cycle later in time.
Forwarding to the AND instruction now goes through the register file or forwarding is
not required for the OR instruction. No instruction is started during the clock cycle 4.

Control Hazard

When a branch is executed, it may or may not change the content of PC. If a branch is
taken, the content of PC is changed to target address. If a branch is taken, the content of
PC is not changed

The simple way of dealing with the branches is to redo the fetch of the instruction
following a branch. The first IF cycle is essentially a stall, because, it never performs
useful work. One stall cycle for every branch will yield a performance loss 10% to 30%
depending on the branch frequency

Reducing the Brach Penalties
There are many methods for dealing with the pipeline stalls caused by branch delay
1. Freeze or Flush the pipeline, holding or deleting any instructions after the
ranch until the branch destination is known. It is a simple scheme and branch penalty is
fixed and cannot be reduced by software
2. Treat every branch as not taken, simply allowing the hardware to continue as if
the branch were not to executed. Care must be taken not to change the processor state
until the branch outcome is known.
Instructions were fetched as if the branch were a normal instruction. If the branch
is taken, it is necessary to turn the fetched instruction in to a no-of instruction and restart
the fetch at the target address. Figure 2.8 shows the timing diagram of both the situations.
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 32



3. Treat every branch as taken: As soon as the branch is decoded and target
Address is computed, begin fetching and executing at the target if the branch target is
known before branch outcome, then this scheme gets advantage.
For both predicated taken or predicated not taken scheme, the compiler can improve
performance by organizing the code so that the most frequent path matches the hardware
choice.
4. Delayed branch technique is commonly used in early RISC processors. In a
delayed branch, the execution cycle with a branch delay of one is
Branch instruction
Sequential successor-1
Branch target if taken

The sequential successor is in the branch delay slot and it is executed irrespective of
whether or not the branch is taken. The pipeline behavior with a branch delay is shown in
Figure 2.9. Processor with delayed branch, normally have a single instruction delay.
Compiler has to make the successor instructions valid and useful there are three ways in
which the to delay slot can be filled by the compiler.


ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 33


The limitations on delayed branch arise from
i) Restrictions on the instructions that are scheduled in to delay slots.
ii) Ability to predict at compiler time whether a branch is likely to be taken or
not taken.
The delay slot can be filled from choosing an instruction
a) From before the branch instruction
b) From the target address
c) From fall- through path.
The principle of scheduling the branch delay is shown in fig 2.10




ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 34


What makes pipelining hard to implements?

Dealing with exceptions: Overlapping of instructions makes it more difficult to
know whether an instruction can safely change the state of the CPU. In a pipelined CPU,
an instruction execution extends over several clock cycles. When this instruction is in
execution, the other instruction may raise exception that may force the CPU to abort the
instruction in the pipeline before they complete

Types of exceptions:

The term exception is used to cover the terms interrupt, fault and exception. I/O
device request, page fault, Invoking an OS service from a user program, Integer
arithmetic overflow, memory protection overflow, Hardware malfunctions, Power failure
etc. are the different classes of exception. Individual events have important characteristics
that determine what action is needed corresponding to that exception.
i) Synchronous versus Asynchronous

If the event occurs at the same place every time the program is executed with the
same data and memory allocation, the event is asynchronous. Asynchronous events are
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 35

caused by devices external to the CPU and memory such events are handled after the
completion of the current instruction.
ii) User requested versus coerced:
User requested exceptions are predictable and can always be handled after the
current instruction has completed. Coerced exceptions are caused by some
hardware event that is not under the control of the user program. Coerced
exceptions are harder to implement because they are not predictable
iii) User maskable versus user non maskable :

If an event can be masked by a user task, it is user maskable. Otherwise it is user
non maskable.
iv) Within versus between instructions:
Exception that occur within instruction are usually synchronous, since the
instruction triggers the exception. It is harder to implement exceptions that occur
withininstructions than those between instructions, since the instruction must be
stopped and restarted. Asynchronous exceptions that occurs within instructions
arise from catastrophic situations and always causes program termination.
v) Resume versus terminate:
If the programs execution continues after the interrupt, it is a resuming event
otherwise if is terminating event. It is easier implement exceptions that terminate
execution. 29

Stopping and restarting execution:
The most difficult exception have 2 properties:
1. Exception that occur within instructions
2. They must be restartable
For example, a page fault must be restartable and requires the intervention of OS. Thus
pipeline must be safely shutdown, so that the instruction can be restarted in the correct
state. If the restarted instruction is not a branch, then we will continue to fetch the
sequential successors and begin their execution in the normal fashion. 11) Restarting is
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 36

usually implemented by saving the PC of the instruction at which to restart. Pipeline
control can take the following steps to save the pipeline state safely.
i) Force a trap instruction in to the pipeline on the next IF
ii) Until the trap is taken, turn off all writes for the faulting instruction and for all
instructions that follow in pipeline. This prevents any state changes for instructions that
will not be completed before the exception is handled.
iii) After the exception handling routine receives control, it immediately saves
the PC of the faulting instruction. This value will be used to return from the exception
later.

NOTE:
1. with pipelining multiple exceptions may occur in the same clock cycle because there
are multiple instructions in execution.
2 Handling the exception becomes still more complicated when the instructions are
allowed to execute in out of order fashion.

Operation: send out the [PC] and fetch the instruction from memory in to the Instruction
Register (IR). Increment PC by 4 to address the next sequential instruction.

2. Instruction decode / Register fetch cycle (ID)

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 37


Operation: decode the instruction and access that register file to read the registers
( rs and rt). File to read the register (rs and rt). A & B are the temporary registers.
Operands are kept ready for use in the next cycle.
Decoding is done in concurrent with reading register. MIPS ISA has fixed length
Instructions. Hence, these fields are at fixed locations.

3. Execution/ Effective address cycle (EX)

One of the following operations are performed depending on the instruction type.
* Memory reference:

:


Operation: ALU adds the operands to compute the effective address and places the result
in to the register ALU output.
Register Register ALU instruction:



Operation: The ALU performs the operation specified by the function code on the value
taken from content of register A and register B.
*. Register- Immediate ALU instruction:



ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 38

Operation: the content of register A and register Imm are operated (function Op) and
result is placed in temporary register ALU output.
*. Branch:


























ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 39

UNIT II
INSTRUCTION LEVEL PARALLELISM

The potential overlap among instruction execution is called Instruction Level Parallelism
(ILP) since instructions can be executed in parallel. There are mainly two approaches to
exploit ILP.

i) Hardware based approach: An approach that relies on hardware to help
discover and exploit the parallelism dynamically. Intel Pentium series which
has dominated in the market) uses this approach.

ii) Software based approach: An approach that relies on software technology to
find parallelism statically at compile time. This approach has limited use in
scientific or application specific environment. Static approach of exploiting
ILP is found in Intel Itanium.

Factors of both programs and processors limit the amount of parallelism that can be
exploited among instructions and these limit the performance achievable. The
performance of the pipelined processors is given by:

Pipeline CPI= Ideal Pipeline CPI + Structural stalls + Data hazard stalls + Control stalls

By reducing each of the terms on the right hand side, it is possible to minimize the overall
pipeline CPI.

To exploit the ILP, the primary focus is on Basic Block (BB). The BB is a straight line
code sequence with no branches in except the entry and no branches out except at the
exit. The average size of the BB is very small i.e., about 4 to 6 instructions. The flow
diagram segment of a program is shown below (Figure 3.1). BB1 , BB2 and BB3 are the
Basic Blocks.

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 40

Figure 3.1 Flow diagram segment


The amount of overlap that can be exploited within a Basic Block is likely to be less than
the average size of BB. To further enhance ILP, it is possible to look at ILP across
multiple BB. The simplest and most common way to increase the ILP is to exploit the
parallelism among iterations of a loop (Loop level parallelism). Each iteration of a loop
can overlap with any other iteration.


Data Dependency and Hazard
If two instructions are parallel, they can execute simultaneously in a pipeline of
arbitrary length without causing any stalls, assuming the pipeline has sufficient resources.
If two instructions are dependent, they are not parallel and must be executed in sequential
order.
There are three different types dependences.

Data Dependences (True Data Dependency)
Name Dependences
Control Dependences

Data Dependences
An instruction j is data dependant on instruction i if either of the following holds:
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 41

i) Instruction i produces a result that may be used by instruction j
Eg1: i: L.D F0, 0(R1)
j: ADD.D F4, F0, F2
ith instruction is loading the data into the F0 and jth instruction use F0 as one the
operand. Hence, jth instruction is data dependant on ith instruction.
Eg2: DADD R1, R2, R3
DSUB R4, R1, R5

ii) Instruction j is data dependant on instruction k and instruction k data dependant on
instruction i
Eg: L.D F4, 0(R1)
MUL.D F0, F4, F6
ADD.D F5, F0, F7

Dependences are the property of the programs. A Data value may flow between
instructions either through registers or through memory locations. Detecting the data flow
and dependence that occurs through registers is quite straight forward. Dependences that
flow through the memory locations are more difficult to detect. A data dependence
convey three things.

a) The possibility of the Hazard.
b) The order in which results must be calculated and
c) An upper bound on how much parallelism can possibly exploited.

Name Dependences

A Name Dependence occurs when two instructions use the same Register or Memory
location, but there is no flow of data between the instructions associated with that name.

Two types of Name dependences:
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 42

i) Antidependence: between instruction i and instruction j occurs when instruction j
writes a register or memory location that instruction i reads. he original ordering must be
preserved to ensure that i reads the correct value.
Eg: L.D F0, 0(R1)
DADDUI R1, R1, R3

ii) Output dependence: Output Dependence occurs when instructions i and j write to the
same register or memory location.
Ex: ADD.D F4, F0, F2
SUB.D F4, F3, F5

The ordering between the instructions must be preserved to ensure that the value finally
written corresponds to instruction j.The above instruction can be reordered or can be
executed simultaneously if the name of the register is changed. The renaming can be
easily done either statically by a compiler or dynamically by the hardware.

Data hazard: Hazards are named by the ordering in the program that must be preserved
by the pipeline

RAW (Read After Write): j tries to read a source before i writes it, so j in correctly gets
old value, this hazard is due to true data dependence.

WAW (Write After Write): j tries to write an operand before it is written by i. WAW
hazard arises from output dependence.

WAR (Write After Read): j tries to write a destination before it is read by i, so that I
incorrectly gets the new value. WAR hazard arises from an antidependence and normally
cannot occur in static issue pipeline.

CONTROL DEPENDENCE:
A control dependence determines the ordering of an instruction i with respect to a branch
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 43

instruction,
Ex: if P1 {
S1;
}
if P2 {
S2;
}
S1 is Control dependent on P1 and
S2 is control dependent on P2 but not on P1.
a) An instruction that is control dependent on a branch cannot be moved before the
branch ,so that its execution is no longer controlled by the branch.
b) An instruction that is not control dependent on a branch cannot be moved after the
branch so that its execution is controlled by the branch.


BASIC PIPELINE SCHEDULE AND LOOP UNROLLING

To keep a pipe line full, parallelism among instructions must be exploited by
finding sequence of unrelated instructions that can be overlapped in the pipeline. To
avoid a pipeline stall,a dependent instruction must be separated from the source
instruction by the distance in clock cycles equal to the pipeline latency of that source
instruction. A compilers ability to perform this scheduling depends both on the amount
of ILP available in the program and on the latencies of the functional units in the
pipeline.

The compiler can increase the amount of available ILP by transferring loops.
for(i=1000; i>0 ;i=i-1)
X[i] = X[i] + s;
We see that this loop is parallel by the noticing that body of the each iteration is
independent.

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 44

The first step is to translate the above segment to MIPS assembly language
Loop: L.D F0, 0(R1) : F0=array element
ADD.D F4, F0, F2 : add scalar in F2
S.D F4, 0(R1) : store result
DADDUI R1, R1, #-8 : decrement pointer
: 8 Bytes (per DW)
BNE R1, R2, Loop : branch R1! = R2

Without any Scheduling the loop will execute as follows and takes 9 cycles for each
iteration.
1 Loop: L.D F0, 0(R1) ;F0=vector element
2 stall
3 ADD.D F4, F0, F2 ;add scalar in F2
4 stall
5 stall
6 S.D F4, 0(R1) ;store result
7 DADDUI R1, R1,# -8 ;decrement pointer 8B (DW)
8 stall ;assumes cant forward to branch
9 BNEZ R1, Loop ;branch R1!=zero

We can schedule the loop to obtain only two stalls and reduce the time to 7 cycles:
L.D F0, 0(R1)

DADDUI R1, R1, #-8
ADD.D F4, F0, F2
Stall
Stall
S.D F4, 0(R1)
BNE R1, R2, Loop

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 45

Loop Unrolling can be used to minimize the number of stalls. Unrolling the body of the
loop by our times, the execution of four iteration can be done in 27 clock cycles or 6.75
clock cycles per iteration.

1 Loop: L.D F0,0(R1)
3 ADD.D F4,F0,F2
6 S.D 0(R1),F4 ;drop DSUBUI & BNEZ
7 L.D F6,-8(R1)
9 ADD.D F8,F6,F2
12 S.D -8(R1),F8 ;drop DSUBUI & BNEZ
13 L.D F10,-16(R1)
15 ADD.D F12,F10,F2
18 S.D -16(R1),F12 ;drop DSUBUI & BNEZ
19 L.D F14,-24(R1)
21 ADD.D F16,F14,F2
24 S.D -24(R1),F16
25 DADDUI R1,R1,#-32 :alter to 4*8
26 BNEZ R1,LOOP

Unrolled loop that minimizes the stalls to 14 clock cycles for four iterations is given
below:
1 Loop: L.D F0, 0(R1)
2 L.D F6, -8(R1)
3 L.D F10, -16(R1)
4 L.D F14, -24(R1)
5 ADD.D F4, F0, F2
6 ADD.D F8, F6, F2

7 ADD.D F12, F10, F2
8 ADD.D F16, F14, F2
9 S.D 0(R1), F4
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 46

10 S.D -8(R1), F8
11 S.D -16(R1), F12
12 DSUBUI R1, R1,#32
13 S.D 8(R1), F16 ;8-32 = -24
14 BNEZ R1, LOOP

Summary of Loop unrolling and scheduling
The loop unrolling requires understanding how one instruction depends on another and
how the instructions can be changed or reordered given the dependences:

1. Determine loop unrolling useful by finding that loop iterations were independent
(except for maintenance code)
2. Use different registers to avoid unnecessary constraints forced by using same registers
for different computations
3. Eliminate the extra test and branch instructions and adjust the loop termination and
iteration code
4. Determine that loads and stores in unrolled loop can be interchanged by observing that
loads and stores from different iterations are independent
Transformation requires analyzing memory addresses and finding that they do
not refer to the same address
5. Schedule the code, preserving any dependences needed to yield the same result as the
original code
To reduce the Branch cost, prediction of the outcome of the branch may be done.
The prediction may be done statically at compile time using compiler support or
dynamically using hardware support. Schemes to reduce the impact of control hazard are
discussed below:

Static Branch Prediction

Assume that the branch will not be taken and continue execution down the
sequential instruction stream. If the branch is taken, the instruction that are being fetched
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 47

and decoded must be discarded. Execution continues at the branch target. Discarding
instructions means we must be able to flush instructions in the IF, ID and EXE stages.
Alternately, it is possible that the branch can be predicted as taken. As soon as the
instruction decoded is found as branch, at the earliest, start fetching the instruction from
the target address.

Average misprediction = untaken branch frequency = 34% for SPEC pgms.



The graph shows the misprediction rate for set of SPEC benchmark programs


Dynamic Branch Prediction
With deeper pipelines the branch penalty increases when measured in clock
cycles. Similarly, with multiple issue, the branch penalty increases in terms of
instructions lost. Hence, a simple static prediction scheme is inefficient or may not be
efficient in most of the situations. One approach is to look up the address of the
instruction to see if a branch was taken the last time this instruction was executed, and if
so, to begin fetching new instruction from the target address.
This technique is called Dynamic branch prediction.
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 48

Why does prediction work?

Underlying algorithm has regularities
Data that is being operated on has regularities
Instruction sequence has redundancies that are artifacts of way that
humans/compilers think about problems.
There are a small number of important branches in programs which have
dynamic behavior for which dynamic branch prediction performance will be definitely
better compared to static branch prediction.

Performance = (accuracy, cost of misprediction)
Branch History Table (BHT) is used to dynamically predict the outcome of the
current branch instruction. Lower bits of PC address index table of 1-bit values
Says whether or not branch taken last time
o - No address check

Problem: in a loop, 1-bit BHT will cause two mispredictions (average is 9 iterations
before exit):
End of loop case, when it exits instead of looping as before
First time through loop on next time through code, when it predicts exit instead of
looping

Simple two bit history table will give better performance. The four different states of 2
bit predictor is shown in the state transition diagram.

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 49


Correlating Branch Predictor
It may be possible to improve the prediction accuracy by considering the recent behavior
of other branches rather than just the branch under consideration. Correlating predictors
are two-level predictors. Existing correlating predictors add information about the
behavior of the most recent branches to decide how to predict a given branch.
Idea: record m most recently executed branches as taken or not taken, and use
that pattern to select the proper n-bit branch history table (BHT)
In general, (m,n) predictor means record last m branches to select between 2m
history tables, each with n-bit counters
Thus, old 2-bit BHT is a (0,2) predictor
Global Branch History: m-bit shift register keeping T/NT status of last m
branches.
Each entry in table has m n-bit predictors. In case of (2,2) predictor, behavior of
recent branches selects between four predictions of next branch, updating just that
prediction. The scheme of the table is shown:
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 50


Comparisons of different schemes are shown in the graph.


Tournament predictor is a multi level branch predictor and uses n bit saturating counter
to chose between predictors. The predictors used are global predictor and local predictor.
Advantage of tournament predictor is ability to select the right predictor for a
particular branch which is particularly crucial for integer benchmarks.
A typical tournament predictor will select the global predictor almost 40% of the
time for the SPEC integer benchmarks and less than 15% of the time for the SPEC
FP benchmarks
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 51

Existing tournament predictors use a 2-bit saturating counter per branch to choose
among two different predictors based on which predictor was most effective oin
recent prediction.


Dynamic Branch Prediction Summary
Prediction is becoming important part of execution as it improves the performance of
the pipeline.
Branch History Table: 2 bits for loop accuracy
Correlation: Recently executed branches correlated with next branch
Either different branches (GA)
Or different executions of same branches (PA)
Tournament predictors take insight to next level, by using multiple predictors
usually one based on global information and one based on local information,
and combining them with a selector
In 2006, tournament predictors using 30K bits are in processors like the Power
and Pentium 4
Tomasulu algorithm and Reorder Buffer
Tomasulu idea:
1. Have reservation stations where register renaming is possible
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 52

2. Results are directly forwarded to the reservation station along with the final
registers. This is also called short circuiting or bypassing.

ROB:
1.The instructions are stored sequentially but we have indicators to say if it is speculative
or completed execution.
2. If completed execution and not speculative and reached head of the queue then we
commit it.



Speculating on Branch Outcomes
To optimally exploit ILP (instruction-level parallelism) e.g. with pipelining,
Tomasulo,etc. it is critical to efficiently maintain control dependencies (=branch
dependencies)
Key idea: Speculate on the outcome of branches(=predict) and execute instructions as if
the predictions are correct
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 53

of course, we must proceed in such a manner as to be able to recover if our speculation
turns out wrong

Three components of hardware-based speculation
1. dynamic branch prediction to pick branch outcome
2. speculation to allow instructions to execute before control dependencies are
resolved, i.e., before branch outcomes become known with ability to undo in case
of incorrect speculation
3. dynamic scheduling



Speculating with Tomasulo
Key ideas:
1. separate execution from completion: instructions to execute speculatively but no
instructions update registers or memory until no more speculative
2. therefore, add a final step after an instruction is no longer speculative, called
instruction commit when it is allowed to make register and memory updates
3. allow instructions to execute and complete out of order but force them to commit in
order
4. Add hardware called the reorder buffer (ROB), with registers to hold the result of
an instruction between completion and commit

Tomasulos Algorithm with Speculation: Four Stages
1. Issue: get instruction from Instruction Queue
_ if reservation station and ROB slot free (no structural hazard),
control issues instruction to reservation station and ROB, and sends to reservation
station operand values (or reservation station source for values) as well as
allocated ROB slot number
2. Execution: operate on operands (EX)
_ when both operands ready then execute;if not ready, watch CDB for result
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 54

3. Write result: finish execution (WB)
_ write on CDB to all awaiting units and ROB; mark reservation station available
4. Commit: update register or memory with ROB result
_ when instruction reaches head of ROB and results present, update register with
result or store to memory and remove instruction from ROB
_ if an incorrectly predicted branch reaches the head of ROB, flush the ROB, and
restart at correct successor of branch

ROB Data Structure
ROB entry fields
Instruction type: branch, store, register operation (i.e., ALU or load)
State: indicates if instruction has completed and value is ready
Destination: where result is to be written register number for register operation (i.e.
ALU or load), memory address for store
branch has no destination result


Value: holds the value of instruction result till time to commit
Additional reservation station field
Destination: Corresponding ROB entry number
Example
1. L.D F6, 34(R2)
2. L.D F2, 45(R3
3. MUL.D F0, F2, F4
4. SUB.D F8, F2, F6
5. DIV.D F10, F0, F6
6. ADD.D F6, F8, F2

The position of Reservation stations, ROB and FP registers are indicated below:

Assume latencies load 1 clock, add 2 clocks, multiply 10 clocks, divide 40 clocks
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 55

Show data structures just before MUL.D goes to commit

Reservation Stations




At the time MUL.D is ready to commit only the two L.D instructions have already
committed,though others have completed execution
Actually, the MUL.D is at the head of the ROB the L.D instructions are shown only for
understanding purposes #X represents value field of ROB entry number X
Floating point registers

Reorder Buffer
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 56


Example
Loop: LD F0 0 R1
MULTD F4 F0 F2
SD F4 0 R1
SUBI R1 R1 #8
BNEZ R1 Loop

Assume instructions in the loop have been issued twice
Assume L.D and MUL.D from the first iteration have committed and all other
instructions have completed
Assume effective address for store is computed prior to its issue
Show data structures






Reorder Buffer
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 57


Notes
If a branch is mispredicted, recovery is done by flushing the ROB of all entries that
appear after the mispredicted branch
entries before the branch are allowed to continue
restart the fetch at the correct branch successor
When an instruction commits or is flushed from the ROB then the corresponding slots
become available for subsequent instructions

Advantages of hardware-based speculation:

-able to disambiguate memory references;
-better when hardware-based branch prediction is better than software-based
branch
prediction done at compile time; - maintains a completely precise exception
model even for speculated instructions;
does not require compensation or bookkeeping code;
main disadvantage:
complex and requires substantial hardware resources;
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 58


INSTRUCTION LEVEL PARALLELISM 2
What is ILP?
Instruction Level Parallelism
Number of operations (instructions) that can be performed in parallel
Formally, two instructions are parallel if they can execute simultaneously in a pipeline
of arbitrary depth without causing any stalls assuming that the pipeline has sufficient
resources
Primary techniques used to exploit ILP
Deep pipelines
Multiple issue machines
Basic program blocks tend to have 4-8 instructions between branches
Little ILP within these blocks
Must find ILP between groups of blocks
Example Instruction Sequences
Independent instruction sequence:
lw $10, 12($1)
sub $11, $2, $3
and $12, $4, $5
or $13, $6, $7
add $14, $8, $9
Dependent instruction sequence:

lw $10, 12($1)
sub $11, $2, $10
and $12, $11, $10
or $13, $6, $7
add $14, $8, $13

Finding ILP:
Must deal with groups of basic code blocks
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 59

Common approach: loop-level parallelism
Example:
In MIPS (assume $s0 initialized properly):

for (i=1000; i > 0; i--)
x[i] = x[i] + s;
Loop: lw $t0, 0($s1) # t0 = array element
addu $t0, $t0, $s2 # add scalar in $s2
sw $t0, 0($s1) # store result
addi $s1, $s1, -4 # decrement pointer
bne $s1, $0, Loop # branch $s1 != 0

Loop Unrolling:
Technique used to help scheduling (and performance)
Copy the loop body and schedule instructions from different iterations of the
loop gether
MIPS example (from prev. slide):

Loop: lw $t0, 0($s1) # t0 = array element
addu $t0, $t0, $s2 # add scalar in $s2
sw $t0, 0($s1) # store result
lw $t1, -4($s1)
addu $t1, $t1, $s2
sw $t1, -4($s1)
addi $s1, $s1, -8 # decrement pointer
bne $s1, $0, Loop # branch $s1 != 0

Note the new register & counter adjustment!
Previous example, we unrolled the loop once
This gave us a second copy
Why introduce a new register ($t1)?
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 60

Antidependence (name dependence)
Loop iterations would reuse register $t0
No data overlap between loop iterations!
Compiler RENAMED the register to prevent a dependence
Allows for better instruction scheduling and identification of true dependencies
In general, you can unroll the loop as much as you want
A factor of the loop counter is generally used
Limited advantages to unrolling more than a few times


Loop Unrolling: Performance:
Performance (dis)advantage of unrolling
Assume basic 5-stage pipeline
Recall lw requires a bubble if value used immediately after
For original loop
10 cycles to execute first iteration
16 cycles to execute two iterations
Assuming perfect prediction
For unrolled loop
14 cycles to execute first iteration -- without reordering
Gain from skipping addi, bne
12 cycles to execute first iteration -- with reordering
Put lw together, avoid bubbles after ea

Loop Unrolling: Limitations
Overhead amortization decreases as loop is unrolled more
Increase in code size
Could be bad if ICache miss rate increases
Register pressure
Run out of registers that can be used in renaming process

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 61



Exploiting ILP: Deep Pipelines
Deep Pipelines
Increase pipeline depth beyond 5 stages
Generally allows for higher clock rates
UltraSparc III -- 14 stages
Pentium III -- 12 stages
Pentium IV -- 22 stages
Some versions have almost 30 stages
Core 2 Duo -- 14 stages
AMD Athlon -- 9 stages
AMD Opteron -- 12 stages
Motorola G4e -- 7 stages
IBM PowerPC 970 (G5) -- 14 stages
Increases the number of instructions executing at the same time
Most of the CPUs listed above also issue multiple instructions per cycle
Issues with Deep Pipelines
Branch (Mis-)prediction
Speculation: Guess the outcome of an instruction to remove it as a dependence
to other instructions
Tens to hundreds of instructions in flight
Have to flush some/all if a branch is mispredicted
Memory latencies/configurations
To keep latencies reasonable at high clock rates, need fast caches
Generally smaller caches are faster
Smaller caches have lower hit rates
Techniques like way prediction and prefetching can help lower latencies

Optimal Pipelining Depths
Several papers published on this topic
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 62

Esp. the 29th International Symposium on Computer Architecture (ISCA)
Intel had one pushing the depth to 50 stages
Others have shown ranges between 15 and 40
Most of the variation is due to the intended workload
Exploiting ILP: Multiple Issue Computers
Multiple Issue Computers
Benefit
CPIs go below one, use IPC instead (instructions/cycle)
Example: Issue width = 3 instructions, Clock = 3GHz
Peak rate: 9 billion instructions/second, IPC = 3
For our 5 stage pipeline, 15 instructions in flight at any given time
Multiple Issue types
Static
Most instruction scheduling is done by the compiler
Dynamic (superscalar)
CPU makes most of the scheduling decisions
Challenge: overcoming instruction dependencies
Increased latency for loads
Control hazards become worse
Requires a more ambitious design
Compiler techniques for scheduling
Complex instruction decoding logic

Exploiting ILP:Multiple Issue Computers Static Scheduling
Instruction Issuing
Have to decide which instruction types can issue in a cycle
Issue packet: instructions issued in a single clock cycle
Issue slot: portion of an issue packet
Compiler assumes a large responsibility for hazard checking, scheduling, etc.
Static Multiple Issue
For now, assume a souped-up 5-stage MIPS pipeline that can issue a packet with:
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 63

One slot is an ALU or branch instruction
One slot is a load/store instruction


Static Multiple Issue







ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 64


Static Multiple Issue Scheduling


Static Mult. Issue w/Loop Unrolling

Static Mult. Issue w/Loop Unrolling
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 65


Exploiting ILP:Multiple Issue Computers Dynamic Scheduling
Dynamic Multiple Issue Computers
Superscalar computers
CPU generally manages instruction issuing and ordering
Compiler helps, but CPU dominates
Process
Instructions issue in-order
Instructions can execute out-of-order
Execute once all operands are ready
Instructions commit in-order
Commit refers to when the architectural register file is updated (current completed state
of program
Aside: Data Hazard Refresher
Two instructions (i and j), j follows i in program order
Read after Read (RAR)
Read after Write (RAW)
Type:
Problem:
Write after Read (WAR)
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 66

Type:
Problem:
Write after Write (WAW)
Type: Problem:
Superscalar Processors
Register Renaming
Use more registers than are defined by the architecture
Architectural registers: defined by ISA
Physical registers: total registers
Help with name dependencies
Antidependence
Write after Read hazard
Output dependence
Write after Write hazard

Tomasulos Superscalar Computers
R. M. Tomasulo, An Efficient Algorithm for Exploiting Multiple Arithmetic Units,
IBM J. of Research and Development, Jan. 1967
See also: D. W. Anderson, F. J. Sparacio, and R. M. Tomasulo, The IBM System/360
model 91: Machine philosophy and instruction-handling, IBM J. of Research and
evelopment, Jan. 1967
Allows out-of-order execution
Tracks when operands are available
Minimizes RAW hazards
Introduced renaming for WAW and WAR
hazards
Tomasulos Superscalar Computers


ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 67




Instruction Execution Process
Three parts, arbitrary number of cycles/part
Above does not allow for speculative execution
Issue (aka Dispatch)
If empty reservation station (RS) that matches instruction, send to RS with operands
rom register file and/or know which functional unit will send operand
If no empty RS, stall until one is available

Rename registers as appropriate
Instruction Execution Process
Execute
All branches before instruction must be resolved
Preserves exception behavior
When all operands available for an instruction, send it to functional unit
Monitor common data bus (CDB) to see if result is needed by RS entry
For non-load/store reservation stations
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 68

If multiple instructions ready, have to pick one to send to functional unit
For load/store
Compute address, then place in buffer
Loads can execute once memory is free
Stores must wait for value to be stored, then execute

Write Back
Functional unit places on CDB
Goes to both register file and reservation stations
Use of CDB enables forwarding for RAW hazards
Also introduces a latency between result and use of a value

Reservation Stations
Require 7 fields
Operation to perform on operands (2 operands)
Tags showing which RS/Func. Unit will be producing operand (or zero if operand
available/unnecessary)
Two source operand values
A field for holding memory address calculation data
Initially, immediate field of instruction
Later, effective address
Busy
Indicates that RS and its functional unit are busy
Register file support
Each entry contains a field that identifies which RS/func. unit will be writing into this
entry (or blank/zero if noone will be writing to it) Limitation of Current Machine

Instruction execution requires branches to be resolved
For wide-issue machines, may issue one branch per clock cycle!
Desire:
Predict branch direction to get more ILP
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 69

Eliminate control dependencies
Approach:
Predict branches, utilize speculative instruction execution
Requires mechanisms for fixing machine when speculation is incorrect
Tomasulos w/Hardware Speculation




Tomasulos w/HW Speculation
Key aspects of this design
Separate forwarding (result bypassing) from actual instruction completion
Assuming instructions are executing speculatively
Can pass results to later instructions, but prevents instruction from performing updates
that cant be undone
Once instruction is no longer speculative it can update register file/memory
New step in execution sequence: instruction commit
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 70

Requires instructions to wait until they can commit Commits still happen in order
Reorder Buffer (ROB)

Instructions hang out here before committing
Provides extra registers for RS/RegFile
Is a source for operands
Four fields/entry
Instruction type
Branch, store, or register operation (ALU & load)
Destination field
Register number or store address
Value field
Holds value to write to register or data for store
Ready field
Has instruction finished executing?
Note: store buffers from previous version now in ROB
Instruction Execution Sequence
Issue
Issue instruction if opening in RS & ROB
Send operands to RS from RegFile and/or ROB
Execute
Essentially the same as before
Write Result
Similar to before, but put result into ROB
Commit (next slide)
Committing Instructions
Look at head of ROB
Three types of instructions
Incorrectly predicted branch
Indicates speculation was wrong
Flush ROB
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 71

Execution restarts at proper location Store
Update memory
Remove store from ROB
Everything else
Update registers
Remove instruction from ROB



RUU Superscalar Computers

Modeling tool SimpleScalar implements an RUU style processor
You will be using this tool after Spring Break
Architecture similar to speculative Tomasulos
Register Update Unit (RUU)
Controls instructions scheduling and dispatching to functional units
Stores intermediate source values for instructions
Ensures instruction commit occurs in order!
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 72

Needs to be of appropriate size
Minimum of issue width * number of pipeline stages
Too small of an RUU can be a structural hazard!
Result bus could be a structural hazard

A Real Computer:Intel Pentium 4
Pentium 4 Die Photo


Overview of P4
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 73


Pentium 4 Pipeline
See handout for overview of major steps
Prescott (90nm version of P4) had 31 pipeline stages
Not sure how pipeline is divided up


P4: Trace Cache
Non-traditional instruction cache
Recall x86 ISA
CISC/VLIW: ugly assembly instructions of varying lengths
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 74

Hard for HW to decode
Ended up translating code into RISC-like microoperations to execute
Trace Cache holds sequences of RISC-like micro-ops
Less time decoding, more time executing
Sequence storage similar to normal instruction cache
P4: Branch Handling
BTBs (Branch Target Buffers)
Keep both branch history and branch target addresses
Target address is instruction immediately after branch
Predict if no entry in BTB for branch
Static prediction
If a backwards branch, see how far target is from current; if within a threshold, predict
taken, else predict not taken
If a forward branch, predict not taken
Also some other rules
Front-end BTB is L2 (like) for the trace cache BTB (L1 like)

P4: Execution Core
Tomasulos algorithm-like
Can have up to 126 instructions in-flight
Max of 3 micro-ops sent to core/cycle
Max of 48 loads, 32 stores
Send up to 6 instructions to functional units per cycle via 4 ports
Port 0: Shared between first fast ALU and FP/Media move scheduler
Port 1: Shared between second fast ALU and Complex integer and FP/Media scheduler
Port 2: Load
Port 3: Store
P4: Rapid Execution Engine
Execute 6 micro-ops/cycle
Simple ALUs run at 2x machine clock rate
Can generate 4 simple ALU results/cycle
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 75

Do one load and one store per cycle
Loads involve data speculation
Assume that most loads hit L1 and Data Translation Look-aside Buffer (DTLB)
Get data into execution, while doing address check
Fix if L1 miss occurred

P4: Memory Tricks
Store-to-Load Forwarding
Stores must wait to write until non-speculative
Loads occasionally want data from store location
Check both cache and Store Forwarding Buffer
SFB is where stores are waiting to be written
If hit when comparing load address to SFB address, use SFB data, not cache data
Done on a partial address
Memory Ordering Buffer
Ensures that store-to-load forwarding was correct
If not, must re-execute load
Force forwarding
Mechanism for forwarding in case addresses are misaligned
MOB can tell SFB to forward or not
False forwarding
Fixes partial address match between load and SFB

P4: Specs for Rest of Slides
For one running at 3.2 GHz
From grad arch book
L1 Cache
Int: Load to use - 4 cycles
FP: Load to use - 12 cycles
Can handle up to 8 outstanding load misses
L2 Cache (2 MB)
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 76

18 cycle access time

P4: Branch Prediction





P4: Misspeculation Percentages

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 77


P4: Data Cache Miss Rates



P4: CPI
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 78


P4 vs. AMD Opteron



P4 vs. Opteron: Real Performance

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 79



















ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 80

UNIT III

MULTIPROCESSORS AND THREAD-LEVEL PARALLELISM

We have seen the renewed interest in developing multiprocessors in early 2000:
- The slowdown in uniprocessor performance due to the diminishing returns in exploring
instruction-level parallelism.
- Difficulty to dissipate the heat generated by uniprocessors with high clock rates.
- Demand for high-performance servers where thread-level parallelism is natural.
For all these reasons multiprocessor architectures has become increasingly attractive.

Taxonomy of Parallel Architectures
The idea of using multiple processors both to increase performance and to
improve availability dates back to the earliest electronic computers. About 30 years ago,
Flynn proposed a simple model of categorizing all computers that is still useful today. He
looked at the parallelism in the instruction and data streams called for by the instructions
at the most constrained component of the multiprocessor, and placed all computers in one
of four categories:

1. Single instruction stream, single data stream
(SISD)This category is the uniprocessor.


ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 81


2. Single instruction stream, multiple data streams
(SIMD)The same instruction is executed by multiple processors using different data
streams. Each processor has its own data memory (hence multiple data), but there is a
single instruction memory and control processor, which fetches and dispatches
instructions. Vector architectures are the largest class of processors of this type.


3.Multiple instruction streams, single data stream (MISD)No commercial
multiprocessor of this type has been built to date, but may be in the future. Some special
purpose stream processors approximate a limited form of this (there is only a single data
stream that is operated on by successive functional units).
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 82



4. Multiple instruction streams, multiple data streams (MIMD)Each processor
fetches its own instructions and operates on its own data. The processors are often off-
the-shelf microprocessors. This is a coarse model, as some multiprocessors are hybrids of
these categories. Nonetheless, it is useful to put a framework on the design space.



1. MIMDs offer flexibility. With the correct hardware and software support, MIMDs can
function as single-user multiprocessors focusing on high performance for one application,
as multiprogrammed multiprocessors running many tasks simultaneously, or as some
combination of these functions.
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 83

2. MIMDs can build on the cost/performance advantages of off-the-shelf
microprocessors. In fact, nearly all multiprocessors built today use the same
microprocessors found in workstations and single-processor servers.

With an MIMD, each processor is executing its own instruction stream. In many cases,
each processor executes a different process. Recall from the last chapter, that a process is
an segment of code that may be run independently, and that the state of the process
contains all the information necessary to execute that program on a processor. In a
multiprogrammed environment, where the processors may be running independent tasks,
each process is typically independent of the processes on other processors. It is also
useful to be able to have multiple processors executing a single program and sharing the
code and most of their address space. When multiple processes share code and data in
this way, they are often called threads

. Today, the term thread is often used in a casual way to refer to multiple loci of
execution that may run on different processors, even when they do not share an address
space. To take advantage of an MIMD multiprocessor with n processors, we must usually
have at least n threads or processes to execute. The independent threads are typically
identified by the programmer or created by the compiler. Since the parallelism in this
situation is contained in the threads, it is called thread-level parallelism.
Threads may vary from large-scale, independent processesfor example,
independent programs running in a multiprogrammed fashion on different processors to
parallel iterations of a loop, automatically generated by a compiler and each executing for
perhaps less than a thousand instructions. Although the size of a thread is important in
considering how to exploit thread-level parallelism efficiently, the important qualitative
distinction is that such parallelism is identified at a high-level by the software system and
that the threads consist of hundreds to millions of instructions that may be executed in
parallel. In contrast, instruction level parallelism is identified by primarily by the
hardware, though with software help in some cases, and is found and exploited one
instruction at a time.

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 84

Existing MIMD multiprocessors fall into two classes, depending on the number of
processors involved, which in turn dictate a memory organization and interconnect
strategy. We refer to the multiprocessors by their memory organization, because what
constitutes a small or large number of processors is likely to change over time.
The first group, which we call




Centralized shared memory architectures have at most a few dozen processors in
2000. For multiprocessors with small processor counts, it is possible for the processors to
share a single centralized memory and to interconnect the processors and memory by a
bus. With large caches, the bus and the single memory, possibly with multiple banks, can
satisfy the memory demands of a small number of processors. By replacing a single bus
with multiple buses, or even a switch, a centralized shared memory design can be scaled
to a few dozen processors. Although scaling beyond that is technically conceivable,
sharing a centralized memory, even organized as multiple banks, becomes less attractive
as the number of processors sharing it increases.

Because there is a single main memory that has a symmetric relationship to all
processos and a uniform access time from any processor, these multiprocessors are often
called symmetric (shared-memory) multiprocessors ( SMPs), and this style of architecture
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 85

is sometimes called UMA for uniform memory access. This type of centralized
sharedmemory architecture is currently by far the most popular organization.
The second group consists of multiprocessors with physically distributed memory.
To support larger processor counts, memory must be distributed among the processors
rather than centralized; otherwise the memory system would not be able to support the
bandwidth demands of a larger number of processors without incurring excessively long
access latency. With the rapid increase in processor performance and the associated
increase in a processors memory bandwidth requirements, the scale of multiprocessor for
which distributed memory is preferred over a single, centralized memory continues to
decrease in number (which is another reason not to use small and large scale). Of course,
the larger number of processors raises the need for a high bandwidth interconnect.





Distributing the memory among the nodes has two major benefits. First, it is a
costeffective way to scale the memory bandwidth, if most of the accesses are to the local
memory in the node. Second, it reduces the latency for accesses to the local memory.
These two advantages make distributed memory attractive at smaller processor counts as
processors get ever faster and require more memory bandwidth and lower memory
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 86

latency. The key disadvantage for a distributed memory architecture is that
communicating data between processors becomes somewhat more complex and has
higher latency, at least when there is no contention, because the processors no longer
share a single centralized memory. As we will see shortly, the use of distributed memory
leads to two different paradigms for interprocessor communication. Typically, I/O as well
as memory is distributed among the nodes of the multiprocessor, and the nodes may be
small SMPs (28 processors). Although the use of multiple processors in a node together
with a memory and a network interface is quite useful from the cost-efficiency viewpoint.

Challenges for Parallel Processing
Limited parallelism available in programs
Need new algorithms that can have better parallel performance
Suppose you want to achieve a speedup of 80 with 100 processors. What fraction
of the original computation can be sequential?



Data Communication Models for Multiprocessors
shared memory: access shared address space implicitly via load and store
operations.
message-passing: done by explicitly passing messages among the
processors
can invoke software with Remote Procedure Call (RPC)
often via library, such as MPI: Message Passing Interface
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 87

also called "Synchronous communication" since communication causes
synchronization between 2 processes

Message-Passing Multiprocessor
- The address space can consist of multiple private address spaces that are
logically disjoint and cannot be addressed by a remote processor
- The same physical address on two different processors refers to two different
locations in two different memories.

Multicomputer (cluster):
- can even consist of completely separate computers connected on a LAN.
- cost-effective for applications that require little or no communication





Symmetric Shared-Memory Architectures

Multilevel caches can substantially reduce the memory bandwidth demands of a
processor.
This is extremely
- Cost-effective
- This can work as plug in play by placing the processor and cache sub-
system on a board into the bus backplane.
Developed by
IBM One chip multiprocessor
AMD and INTEL- Two Processor
SUN 8 processor multi core
Symmetric shared memory support caching of
Shared Data
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 88

Private Data

Private data: used by a single processor When a private item is cached, its location is
migrated to the cache Since no other processor uses the data, the program behavior is
identical to that in a uniprocessor.
Shared data: used by multiple processor When shared data are cached, the shared value
may be replicated in multiple caches advantages: reduce access latency and memory
contention induces a new problem: cache coherence.

Cache Coherence
Unfortunately, caching shared data introduces a new problem because the view of
memory held by two different processors is through their individual caches, which,
without any additional precautions, could end up seeing two different values. I.e, If two
different processors have two different values for the same location, this difficulty is
generally referred to as cache coherence problem





Informally:
Any read must return the most recent write
Too strict and too difficult to implement
Better:
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 89

Any write must eventually be seen by a read
All writes are seen in proper order (serialization)
Two rules to ensure this:
If P writes x and then P1 reads it, Ps write will be seen by P1 if the read
and write are sufficiently far apart
Writes to a single location are serialized: seen in one order
Latest write will be seen
Otherwise could see writes in illogical order (could see older
value after a newer value)
The definition contains two different aspects of memory system:
Coherence
Consistency
A memory system is coherent if,
Program order is preserved.
Processor should not continuously read the old data value.
Write to the same location are serialized.
The above three properties are sufficient to ensure coherence,When a written value will
be seen is also important. This issue is defined by memory consistency model. Coherence
and consistency are complementary.
Basic schemes for enforcing coherence
Coherence cache provides:
migration: a data item can be moved to a local cache and used there in a
transparent fashion.
replication for shared data that are being simultaneously read.
both are critical to performance in accessing shared data.
To over come these problems, adopt a hardware solution by introducing a
protocol tomaintain coherent caches named as Cache Coherence Protocols
These protocols are implemented for tracking the state of any sharing of a data block.
Two classes of Protocols
Directory based
Snooping based
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 90


Directory based
Sharing status of a block of physical memory is kept in one location called the
directory.
Directory-based coherence has slightly higher implementation overhead than
snooping.
It can scale to larger processor count.

Snooping
Every cache that has a copy of data also has a copy of the sharing status of the
block.
No centralized state is kept.
Caches are also accessible via some broadcast medium (bus or switch)
Cache controller monitor or snoop on the medium to determine whether or not
they have a copy of a block that is represented on a bus or switch access.

Snooping protocols are popular with multiprocessor and caches attached to single
shared memory as they can use the existing physical connection- bus to memory, to
interrogate the status of the caches. Snoop based cache coherence scheme is implemented
on a shared bus. Any communication medium that broadcasts cache misses to all the
processors.

Basic Snoopy Protocols
Write strategies
Write-through: memory is always up-to-date
Write-back: snoop in caches to find most recent copy
Write Invalidate Protocol
Multiple readers, single writer
Write to shared data: an invalidate is sent to all caches which snoop and
invalidate any copies
Read miss: further read will miss in the cache and fetch a new
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 91

copy of the data.
Write Broadcast/Update Protocol (typically write through)
Write to shared data: broadcast on bus, processors snoop, and update
any copies
Read miss: memory/cache is always up-to-date.
Write serialization: bus serializes requests!
Bus is single point of arbitration

Examples of Basic Snooping Protocols

Write Invalidate










Write Update
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 92




Assume neither cache initially holds X and the value of X in memory is 0

Example Protocol
Snooping coherence protocol is usually implemented by incorporating a
finitestate controller in each node
Logically, think of a separate controller associated with each cache block
That is, snooping operations or cache requests for different blocks can
proceed independently
In implementations, a single controller allows multiple operations to distinct
blocks to proceed in interleaved fashion
that is, one operation may be initiated before another is completed, even
through only one cache access or one bus access is allowed at time

Example Write Back Snoopy Protocol
Invalidation protocol, write-back cache
Snoops every address on bus
If it has a dirty copy of requested block, provides that block in response
to the read request and aborts the memory access
Each memory block is in one state:
Clean in all caches and up-to-date in memory (Shared)
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 93

OR Dirty in exactly one cache (Exclusive)
OR Not in any caches
Each cache block is in one state (track these):
Shared : block can be read
OR Exclusive : cache has only copy, its writeable, and dirty
OR Invalid : block contains no data (in uniprocessor cache too)
Read misses: cause all caches to snoop bus
Writes to clean blocks are treated as misses
Write-Back State Machine CPU
State Transitions for Each Cache Block is as shown below





CPU may read/write hit/miss to the block
May place write/read miss on bus
May receive read/write miss from bus

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 94




Conclusion
End of uniprocessors speedup => Multiprocessors
Parallelism challenges: % parallalizable, long latency to remote memory
Centralized vs. distributed memory
Small MP vs. lower latency, larger BW for Larger MP
Message Passing vs. Shared Address
Uniform access time vs. Non-uniform access time
Snooping cache over shared medium for smaller MP by invalidating other
cached copies on write
Sharing cached data _ Coherence (values returned by a read), Consistency
(when a written value will be returned by a read)
Shared medium serializes writes _ Write consistency

Implementation Complications
Write Races:
Cannot update cache until bus is obtained
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 95

Otherwise, another processor may get bus first,
and then write the same cache block!
Two step process:
Arbitrate for bus
Place miss on bus and complete operation
If miss occurs to block while waiting for bus, handle miss (invalidate
may be needed) and then restart.
Split transaction bus:
Bus transaction is not atomic:
can have multiple outstanding transactions for a block
Multiple misses can interleave, allowing two caches to grab block in the
Exclusive state
Must track and prevent multiple misses for one block
Must support interventions and invalidations

Performance Measurement
Overall cache performance is a combination of
Uniprocessor cache miss traffic
Traffic caused by communication invalidation and subsequent cache
misses
Changing the processor count, cache size, and block size can affect these two
components of miss rate
Uniprocessor miss rate: compulsory, capacity, conflict
Communication miss rate: coherence misses
True sharing misses + false sharing misses

True and False Sharing Miss
True sharing miss
The first write by a PE to a shared cache block causes an invalidation to
establish ownership of that block
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 96

When another PE attempts to read a modified word in that cache block,
a miss occurs and the resultant block is transferred
False sharing miss
Occur when a block a block is invalidate (and a subsequent reference
causes a miss) because some word in the block, other than the one being
read, is written to
The block is shared, but no word in the cache is actually shared, and
this miss would not occur if the block size were a single word
Assume that words x1 and x2 are in the same cache block, which is in the shared
state in the caches of P1 and P2. Assuming the following sequence of events,
identify each miss as a true sharing miss or a false sharing miss.



Example Result
True sharing miss (invalidate P2)
2: False sharing miss
x2 was invalidated by the write of P1, but that value of x1 is not used in
P2
3: False sharing miss
The block containing x1 is marked shared due to the read in P2, but P2
did not read x1. A write miss is required to obtain exclusive access to the block
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 97

4: False sharing miss
5: True sharing miss







Distributed Shared-Memory Architectures
Distributed shared-memory architectures
Separate memory per processor
Local or remote access via memory controller
The physical address space is statically distributed Coherence
Problems
Simple approach: uncacheable
shared data are marked as uncacheable and only private data are
kept in caches
very long latency to access memory for shared data
Alternative: directory for memory blocks
The directory per memory tracks state of every block in every
cache
which caches have a copies of the memory block, dirty vs. clean,
...
Two additional complications
The interconnect cannot be used as a single point of arbitration like the
bus
Because the interconnect is message oriented, many messages must have
explicit responses

To prevent directory becoming the bottleneck, we distribute directory entries with
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 98

memory, each keeping track of which processors have copies of their memory blocks

Directory Protocols
Similar to Snoopy Protocol: Three states
Shared: 1 or more processors have the block cached, and the value in
memory is up-to-date (as well as in all the caches)
Uncached: no processor has a copy of the cache block (not valid in any
cache)
Exclusive: Exactly one processor has a copy of the cache block, and it
has written the block, so the memory copy is out of date
The processor is called the owner of the block
In addition to tracking the state of each cache block, we must track the
processors that have copies of the block when it is shared (usually a bit vector for
each memory block: 1 if processor has copy)
Keep it simple(r):
Writes to non-exclusive data => write miss
Processor blocks until access completes
Assume messages received and acted upon in order sent



local node: the node where a request originates
home node: the node where the memory location and directory entry of an address
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 99

reside
remote node: the node that has a copy of a cache block (exclusive or shared)




Comparing to snooping protocols:
identical states
stimulus is almost identical
write a shared cache block is treated as a write miss (without fetch the
block)
cache block must be in exclusive state when it is written
any shared block must be up to date in memory
write miss: data fetch and selective invalidate operations sent by the directory
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 100

controller (broadcast in snooping protocols)

Directory Operations: Requests and Actions
Message sent to directory causes two actions:
Update the directory
More messages to satisfy request
Block is in Uncached state: the copy in memory is the current value; only
possible requests for that block are:
Read miss: requesting processor sent data from memory &requestor
made only sharing node; state of block made Shared.
Write miss: requesting processor is sent the value & becomes the
Sharing node. The block is made Exclusive to indicate that the only valid copy is
cached. Sharers indicates the identity of the owner.
Block is Shared => the memory value is up-to-date:
Read miss: requesting processor is sent back the data from memory &
requesting processor is added to the sharing set.
Write miss: requesting processor is sent the value. All processors in the
set Sharers are sent invalidate messages, & Sharers is set to identity of
requesting processor. The state of the block is made Exclusive.
Block is Exclusive: current value of the block is held in the cache of the
processor identified by the set Sharers (the owner) => three possible directory requests:
Read miss: owner processor sent data fetch message, causing state of
block in owners cache to transition to Shared and causes owner to send data to
directory, where it is written to memory & sent back to requesting processor.
Identity of requesting processor is added to set Sharers, which still contains the
identity of the processor that was the owner (since it still has a readable copy). State is
shared.
Data write-back: owner processor is replacing the block and hence must
write it back, making memory copy up-to-date (the home directory essentially
becomes the owner), the block is now Uncached, and the Sharer set is empty.
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 101

Write miss: block has a new owner. A message is sent to old owner
causing the cache to send the value of the block to the directory from which it is
sent to the requesting processor, which becomes the new owner. Sharers is set to
identity of new owner, and state of block is made Exclusive.

Synchronization: The Basics
Synchronization mechanisms are typically built with user-level software routines
that rely on hardware supplied synchronization instructions.
Why Synchronize?
Need to know when it is safe for different processes to use shared data
Issues for Synchronization:
Uninterruptable instruction to fetch and update memory (atomic
operation);
User level synchronization operation using this primitive;
For large scale MPs, synchronization can be a bottleneck; techniques to
reduce contention and latency of synchronization

Uninterruptable Instruction to Fetch and Update Memory
Atomic exchange: interchange a value in a register for a value in memory
0 _ synchronization variable is free
1 _ synchronization variable is locked and unavailable
Set register to 1 & swap
New value in register determines success in getting lock
0 if you succeeded in setting the lock (you were first)
1 if other processor had already claimed access
Key is that exchange operation is indivisible
Test-and-set: tests a value and sets it if the value passes the test
Fetch-and-increment: it returns the value of a memory location and atomically
increments it
0 _ synchronization variable is free
Hard to have read & write in 1 instruction: use 2 instead
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 102

Load linked (or load locked) + store conditional
Load linked returns the initial value
Store conditional returns 1 if it succeeds (no other store to same memory
location since preceding load) and 0 otherwise
Example doing atomic swap with LL & SC:
try: mov R3,R4 ; mov exchange value

ll R2,0(R1) ; load linked
sc R3,0(R1) ; store conditional
beqz R3,try ; branch store fails (R3 = 0)
mov R4,R2 ; put load value in R4

Example doing fetch & increment with LL & SC:
try: ll R2,0(R1) ; load linked
addi R2,R2,#1 ; increment (OK if regreg)
sc R2,0(R1) ; store conditional
beqz R2,try ; branch store fails (R2 = 0)


User Level SynchronizationOperation Using this Primitive
Spin locks: processor continuously tries to acquire, spinning around a loop
trying to get the lock
li R2,#1
lockit: exch R2,0(R1) ; atomic exchange
bnez R2,lockit ; already locked?
What about MP with cache coherency?
Want to spin on cache copy to avoid full memory latency
Likely to get cache hits for such variables
Problem: exchange includes a write, which invalidates all other copies; this
generates considerable bus traffic
Solution: start by simply repeatedly reading the variable; when it changes, then
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 103

try exchange (test and test&set):
try: li R2,#1
lockit: lw R3,0(R1) ;load var
bnez R3,lockit ; _ 0 _ not free _ spin
exch R2,0(R1) ; atomic exchange
bnez R2,try ; already locked?

Memory Consistency Models
What is consistency? When must a processor see the new value? e.g.,
seems that P1: A = 0; P2: B = 0;
..... .....
A = 1; B = 1;
L1: if (B == 0) ... L2: if (A == 0) ...
Impossible for both if statements L1 & L2 to be true?
What if write invalidate is delayed & processor continues?
Memory consistency models:
what are the rules for such cases?
Sequential consistency: result of any execution is the same as if the accesses of
each processor were kept in order and the accesses among different processors were
interleaved _ assignments before ifs above
SC: delay all memory accesses until all invalidates done
Schemes faster execution to sequential consistency
Not an issue for most programs; they are synchronized
A program is synchronized if all access to shared data are ordered by
synchronization operations
write (x)
...
release (s) {unlock}
...
acquire (s) {lock}
...
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 104

read(x)
Only those programs willing to be nondeterministic are not synchronized: data
race: outcome f(proc. speed)
Several Relaxed Models for Memory Consistency since most programs are
synchronized; characterized by their attitude towards: RAR, WAR, RAW, WAW to
different addresses


Relaxed Consistency Models: The Basics
Key idea: allow reads and writes to complete out of order, but to use
synchronization operations to enforce ordering, so that a synchronized program behaves
as if the processor were sequentially consistent
By relaxing orderings, may obtain performance advantages
Also specifies range of legal compiler optimizations on shared data
Unless synchronization points are clearly defined and programs are
synchronized, compiler could not interchange read and write of 2 shared data items
because might affect the semantics of the program
3 major sets of relaxed orderings:
1. W_R ordering (all writes completed before next read)
Because retains ordering among writes, many programs that operate under
sequential consistency operate under this model, without additional synchronization.
Called processor consistency
2. W _ W ordering (all writes completed before next write)
3. R _ W and R _ R orderings, a variety of models depending on ordering restrictions and
how synchronization operations enforce ordering
Many complexities in relaxed consistency models; defining precisely what it means for
a write to complete; deciding when processors can see values that it has written




ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 105


















UNIT VI

REVIEW OF MEMORY HIERARCHY

Unlimited amount of fast memory
- Economical solution is memory hierarchy
- Locality
- Cost performance
Principle of locality
- Most programs do not access all code or data uniformly.
Locality occurs
- Time (Temporal locality)
- Space (spatial locality)
Guidelines
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 106

Smaller hardware can be made faster
Different speed and sizes






Goal is provide a memory system with cost per byte than the next lower level
Each level maps addresses from a slower, larger memory to a smaller but faster
memory higher in the hierarchy.
Address mapping
Address checking.
Hence protection scheme for address for scrutinizing addresses are also part of
the memory hierarchy.

Why More on Memory Hierarchy?
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 107



The importance of memory hierarchy has increased with advances in performance of
processors.

Prototype
` When a word is not found in cache
Fetched from memory and placed in cache with the address tag.
Multiple words( block) is fetched for moved for efficiency reasons.
key design
Set associative
Set is a group of block in the cache.
Block is first mapped on to set.
Find mapping
Searching the set
Chosen by the address of the data:
(Block address) MOD(Number of sets in cache)
n-block in a set
Cache replacement is called n-way set associative.
Cache data
- Cache read.
- Cache write.
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 108

Write through: update cache and writes through to update memory.
Both strategies
- Use write buffer.
this allows the cache to proceed as soon as the data is placed in the
buffer rather than wait the full latency to write the data into memory.
Metric
used to measure the benefits is miss rate
No of access that miss
No of accesses
Write back: updates the copy in the cache.
Causes of high miss rates
Three Cs model sorts all misses into three categories
Compulsory: every first access cannot be in cache
Compulsory misses are those that occur if there is an infinite cache
Capacity: cache cannot contain all that blocks that are needed for the program.
As blocks are being discarded and later retrieved.
Conflict: block placement strategy is not fully associative
Block miss if blocks map to its set.

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 109




Miss rate can be a misleading measure for several reasons
So, misses per instruction can be used per memory reference



ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 110



Cache Optimizations
Six basic cache optimizations
1. Larger block size to reduce miss rate:
- To reduce miss rate through spatial locality.
- Increase block size.
- Larger block size reduce compulsory misses.
- But they increase the miss penalty.
2. Bigger caches to reduce miss rate:
- capacity misses can be reduced by increasing the cache capacity.
- Increases larger hit time for larger cache memory and higher cost and power.
3. Higher associativity to reduce miss rate:
- Increase in associativity reduces conflict misses.
4. Multilevel caches to reduce penalty:
- Introduces additional level cache
- Between original cache and memory.
- L1- original cache
L2- added cache.
L1 cache: - small enough
- speed matches with clock cycle time.
L2 cache: - large enough
- capture many access that would go to main memory.
Average access time can be redefined as
Hit timeL1+ Miss rate L1 X ( Hit time L2 + Miss rate L2 X Miss penalty L2)
5. Giving priority to read misses over writes to reduce miss penalty:
- write buffer is a good place to implement this optimization.
- write buffer creates hazards: read after write hazard.
6. Avoiding address translation during indexing of the cache to reduce hit time:
- Caches must cope with the translation of a virtual address from the processor to
a physical address to access memory.
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 111

- common optimization is to use the page offset.
- part that is identical in both virtual and physical addresses- to index the cache.


Advanced Cache Optimizations
Reducing hit time
Small and simple caches
Way prediction
Trace caches
Increasing cache bandwidth
Pipelined caches
Multibanked caches
Nonblocking caches
Reducing Miss Penalty
Critical word first
Merging write buffers
Reducing Miss Rate
Compiler optimizations
Reducing miss penalty or miss rate via parallelism
Hardware prefetching
Compiler prefetching

First Optimization : Small and Simple Caches
Index tag memory and then compare takes time
_ Small cache can help hit time since smaller memory takes less time to index
E.g., L1 caches same size for 3 generations of AMD microprocessors:
K6, Athlon, and Opteron
Also L2 cache small enough to fit on chip with the processor avoids time
penalty of going off chip
Simple _ direct mapping
Can overlap tag check with data transmission since no choice
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 112

Access time estimate for 90 nm using CACTI model 4.0
Median ratios of access time relative to the direct-mapped caches are 1.32,
1.39, and 1.43 for 2-way, 4-way, and 8-way caches




Second Optimization: Way Prediction
How to combine fast hit time of Direct Mapped and have the lower conflict
misses of 2-way SA cache?
Way prediction: keep extra bits in cache to predict the way, or block within
the set, of next cache access.




Multiplexer is set early to select desired block, only 1 tag comparison performed that
clock cycle in parallel with reading the cache data

Miss _ 1st check other blocks for matches in next clock cycle
Accuracy 85%
Drawback: CPU pipeline is hard if hit takes 1 or 2 cycles
- Used for instruction caches vs. data caches

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 113

Third optimization: Trace Cache
Find more instruction level parallelism?
How to avoid translation from x86 to microops?
Trace cache in Pentium 4
1. Dynamic traces of the executed instructions vs. static sequences of instructions
as determined by layout in memory
Built-in branch predictor
2. Cache the micro-ops vs. x86 instructions
Decode/translate from x86 to micro-ops on trace cache miss
+ 1. _ better utilize long blocks (dont exit in middle of block, dont enter
at label in middle of block)
- 1. _ complicated address mapping since addresses no longer aligned to
powerof-
2 multiples of word size
- 1. _ instructions may appear multiple times in multiple dynamic traces
due to different branch outcomes

Fourth optimization: pipelined cache access to increase bandwidth
Pipeline cache access to maintain bandwidth, but higher latency
Instruction cache access pipeline stages:
1: Pentium
2: Pentium Pro through Pentium III
4: Pentium 4
- _ greater penalty on mispredicted branches
- _ more clock cycles between the issue of the load and the use of the data

Fifth optimization: Increasing Cache Bandwidth Non-Blocking Caches
Non-blocking cache or lockup-free cache allow data cache to continue to supply
cache hits during a miss
requires F/E bits on registers or out-of-order execution
requires multi-bank memories
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 114

hit under miss reduces the effective miss penalty by working during miss vs.
ignoring CPU requests
hit under multiple miss or miss under miss may further lower the effective
miss penalty by overlapping multiple misses
Significantly increases the complexity of the cache controller as there
can be multiple outstanding memory accesses
Requires multiple memory banks (otherwise cannot support)
Pentium Pro allows 4 outstanding memory misses

Value of Hit Under Miss for SPEC






FP programs on average: AMAT= 0.68 -> 0.52 -> 0.34 -> 0.26
Int programs on average: AMAT= 0.24 -> 0.20 -> 0.19 -> 0.19
8 KB Data Cache, Direct Mapped, 32B block, 16 cycle miss, SPEC 92
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 115



Sixth optimization: Increasing Cache Bandwidth via Multiple Banks

Rather than treat the cache as a single monolithic block, divide into independent
banks that can support simultaneous accesses
E.g.,T1 (Niagara) L2 has 4 banks
Banking works best when accesses naturally spread themselves across banks _
mapping of addresses to banks affects behavior of memory system
Simple mapping that works well is sequential interleaving
Spread block addresses sequentially across banks
E,g, if there 4 banks, Bank 0 has all blocks whose address modulo 4 is 0;
bank 1 has all blocks whose address modulo 4 is 1;




ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 116


Seventh optimization: Reduce Miss Penalty: Early Restart and Critical
Word First
Dont wait for full block before restarting CPU
Early restartAs soon as the requested word of the block arrives, send
it to the CPU and let the CPU continue execution
Spatial locality _ tend to want next sequential word, so not clear size of
benefit of just early restart
Critical Word FirstRequest the missed word first from memory and
send it to the CPU as soon as it arrives; let the CPU continue execution while
filling the rest of the words in the block
Long blocks more popular today _ Critical Word 1st Widely used




Eight optimization: Merging Write Buffer to Reduce Miss Penalty-

Write buffer to allow processor to continue while waiting to write to memory
If buffer contains modified blocks, the addresses can be checked to see if
address of new data matches the address of a valid write buffer entry
If so, new data are combined with that entry
Increases block size of write for write-through cache of writes to sequential
words, bytes since multiword writes more efficient to memory
The Sun T1 (Niagara) processor, among many others, uses write merging
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 117




Ninth optimization: Reducing Misses by Compiler Optimizations
McFarling [1989] reduced caches misses by 75% on 8KB direct mapped cache,
4 byte blocks in software
Instructions
Reorder procedures in memory so as to reduce conflict misses
Profiling to look at conflicts (using tools they developed)
Data
Merging Arrays: improve spatial locality by single array of compound
elements vs. 2 arrays
Loop Interchange: change nesting of loops to access data in order stored
in memory
Loop Fusion: Combine 2 independent loops that have same looping and
some variables overlap
Blocking: Improve temporal locality by accessing blocks of data
repeatedly vs. going down whole columns or rows

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 118

Merging Arrays Example
/* Before: 2 sequential arrays */
int val[SIZE];
int key[SIZE];
/* After: 1 array of stuctures */
struct merge {
int val;
int key;
};
struct merge merged_array[SIZE];
Reducing conflicts between val & key; improve spatial locality

Conflict misses in caches not FA vs. Blocking size
Lam et al [1991] a blocking factor of 24 had a fifth the misses vs. 48
despite both fit in cache

Tenth optimization Reducing Misses by Hardware Prefetching of
Instructions & Data
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 119

Prefetching relies on having extra memory bandwidth that can be used without
penalty
Instruction Prefetching
Typically, CPU fetches 2 blocks on a miss: the requested block and the
next consecutive block.
Requested block is placed in instruction cache when it returns, and
prefetched block is placed into instruction stream buffer
Data Prefetching
Pentium 4 can prefetch data into L2 cache from up to 8 streams from 8
different 4 KB pages
Prefetching invoked if 2 successive L2 cache misses to a page, if
distance between those cache blocks is < 256 bytes





Eleventh optimization: Reducing Misses by Software Prefetching Data
Data Prefetch
Load data into register (HP PA-RISC loads)
Cache Prefetch: load into cache
(MIPS IV, PowerPC, SPARC v. 9)
Special prefetching instructions cannot cause faults;
a form of speculative execution
Issuing Prefetch Instructions takes time
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 120

Is cost of prefetch issues < savings in reduced misses?
Higher superscalar reduces difficulty of issue bandwidth
The techniques to improve hit time, bandwidth, miss penalty and miss rate generally
affect the other components of the average memory access equation as well as the
complexity of the memory hierarchy.



MEMORY HIERARCHY DESIGN

AMAT and Processor Performance
AMAT = Average Memory Access Time
Miss-oriented Approach to Memory Access
CPIExec includes ALU and Memory instructions
Separating out Memory component entirely
CPIALUOps does not include memory instructions

Summary: Caches
The Principle of Locality:
Program access a relatively small portion of the address space at any instant of
time.
Temporal Locality OR Spatial Locality:
Three Major Categories of Cache Misses:
Compulsory Misses: sad facts of life. Example: cold start misses.
Capacity Misses: increase cache size
Conflict Misses: increase cache size and/or associativity

Where Misses Come From?
Classifying Misses: 3 Cs
Compulsory The first access to a block is not in the cache,
Also called cold start misses or first reference misses.
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 121

(Misses in even an Infinite Cache)
Capacity If the cache cannot contain all the blocks needed during execution
of a program,
Conflict If block-placement strategy is set associative or direct mapped,
conflict misses (in addition to compulsory & capacity misses) will occur because
a block can be discarded and later retrieved if too many blocks map to its set.
(Misses in N-way Associative, Size X Cache)
More recent, 4th C:
Coherence Misses caused by cache coherence




Write Policy:
Write Through: needs a write buffer.
Write Back: control can be complex
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 122


Summary:
The Cache Design Space
Several interacting dimensions
cache size
block size
associativity
replacement policy
write-through vs write-back
The optimal choice is a compromise
Simplicity often wins



Cache Organization?
Assume total cache size not changed

What happens if: Which of 3Cs is obviously affected?
Change Block Size
Change Cache Size
Change Cache Internal Organization
Change Associativity
Change Compiler

Cache Optimization Summary

How to Improve Cache Performance?
Cache optimizations
1. Reduce the miss rate
2. Reduce the miss penalty
3. Reduce the time to hit in the cache
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 123


Cache Optimisation
Why improve Cache performance:




Review: 6 Basic Cache Optimizations
Reducing hit time
1.Address Translation during Cache Indexing
Reducing Miss Penalty
2. Multilevel Caches
3. Giving priority to read misses over write misses
Reducing Miss Rate
4. Larger Block size (Compulsory misses)
5. Larger Cache size (Capacity misses)
6. Higher Associativity (Conflict misses)

11 Advanced Cache Optimizations
Reducing hit time
1. Small and simple caches
2. Way prediction
3. Trace caches
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 124

Increasing cache bandwidth
4. Pipelined caches
5. Multibanked caches
6. Nonblocking caches
Reducing Miss Penalty
7. Critical word first
8. Merging write buffers
Reducing Miss Rate
9.Compiler optimizations
Reducing miss penalty or miss rate via parallelism
10.Hardware prefetching
11.Compiler prefetching


1. Fast Hit times via Small and Simple Caches
Index tag memory and then compare takes time
Small cache can help hit time since smaller memory takes less time to index
E.g., L1 caches same size for 3 generations of AMD icroprocessors:
K6, Athlon, and Opteron
Also L2 cache small enough to fit on chip with the processor avoids time
penalty of going off chip
Simple direct mapping
Can overlap tag check with data transmission since no choice
2. Fast Hit times via Way Prediction
How to combine fast hit time of Direct Mapped and have the lower conflict
misses of 2-way SA cache?
Way prediction: keep extra bits in cache to predict the way, or block within
the set, ofnext cache access.
Multiplexer is set early to select desired block, only 1 tag comparison performed
that clock cycle in parallel with reading the cache data
Miss - 1st check other blocks for matches in next clock cycle
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 125


3. Fast Hit times via Trace Cache
Find more instruction level parallelism?
How avoid translation from x86 to microops?- Trace cache in Pentium 4
1. Dynamic traces of the executed instructions vs. static sequence of instructions
as determined by layout in memory
Built-in branch predictor
2. Cache the micro-ops vs. x86 instructions - Decode/translate from x86 to
micro-ops on trace cache miss
+ 1. better utilize long blocks (dont exit in middle of block, dont enter at label in
middle of block)
- 1. complicated address mapping since addresses no longer aligned to power-of-2
multiples of word size
- 1. instructions may appear multiple times in multiple dynamic traces due to different
branch outcomes

4: Increasing Cache Bandwidth by Pipelining
Pipeline cache access to maintain bandwidth, but higher latency
Instruction cache access pipeline stages:
1: Pentium
2: Pentium Pro through Pentium III
4: Pentium 4
- greater penalty on mispredicted branches
- more clock cycles between the issue of the load and the use of the data


5. Increasing Cache Bandwidth:
Non-Blocking Caches- Reduce Misses/Penalty
Non-blocking cache or lockup-free cache allow data cache to continue to supply
cache hits during a m iss
requires F/E bits on registers or out-of-order execution
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 126

requires multi-bank memories
hit under miss reduces the effective miss penalty by working
during miss vs. ignoring CPU requests
hit under multiple miss or miss under miss may further lower the effective
miss penalty by overlapping multiple misses
Significantly increases the complexity of the cache controller as there
can be multiple outstanding memory accesses
Requires muliple memory banks (otherwise cannot support)
Penium Pro allows 4 outstanding memory misses

6: Increasing Cache Bandwidth via Multiple Banks
Rather than treat the cache as a single monolithic block, divide into independent banks
that can support simultaneous accesses
E.g.,T1 (Niagara) L2 has 4 banks
Banking works best when accesses naturally spread themselves across banks I mapping
of addresses to banks affects behavior of memory system




Simple mapping that works well is sequential interleaving
Spread block addresses sequentially across banks
E,g, if there 4 banks, Bank 0 has all blocks whose address modulo 4 is
0; bank 1 has all blocks whose address modulo 4 is 1; .
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 127


7. Reduce Miss Penalty:
Early Restart and Critical Word First
Dont wait for full block before restarting CPU
Early restartAs soon as the requested word of the block arrives, send it to the CPU
and let the CPU continue execution
Spatial locality - tend to want next sequential word, so not clear size of benefit of just
early restart
Critical Word FirstRequest the missed word first from memory and send it to the
CPU as soon as it arrives; let the CPU continue execution while filling the rest of the
words in the block









ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 128



8. Merging Write Buffer to Reduce Miss Penalty
Write buffer to allow processor to continue while waiting to write to memory
If buffer contains modified blocks, the addresses can be checked to see if address
of new data matches the address of a valid write buffer entry -If so, new data are
combined with that entry
Increases block size of write for write-through cache of writes to sequential
words, bytes since multiword writes more efficient to memory
The Sun T1 (Niagara) processor, among many others, uses write merging

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 129




9. Reducing Misses by Compiler Optimizations
McFarling [1989] reduced caches misses by 75% on 8KB direct mapped cache, 4 byte
blocks in software
Instructions
Reorder procedures in memory so as to reduce conflict misses
Profiling to look at conflicts (using tools they developed)
Data
Merging Arrays: improve spatial locality by single array of compound elements vs. 2
arrays
Loop Interchange: change nesting of loops to access data in order stored in memory
Loop Fusion: Combine 2 independent loops that have same looping and some variables
overlap
Blocking: Improve temporal locality by accessing blocks of data repeatedly vs. going
down whole columns or rows Compiler Optimizations- Reduction comes from software
(no Hw ch.)
Loop Interchange
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 130

Motivation: some programs have nested loops that access data in nonsequential order
Solution: Simply exchanging the nesting of the loops can make the code access the data
in the order it is stored =>
reduce misses by improving spatial locality; reordering maximizes use of data in a cache
block before it is discarded
Loop Interchange Example
/* Before */
for (j = 0; j < 100; j = j+1)
for (i = 0; i < 5000; i = i+1)
x[i][j] = 2 * x[i][j];
/* After */
for (i = 0; i < 5000; i = i+1)
for (j = 0; j < 100; j = j+1)
x[i][j] = 2 * x[i][j];
Blocking
Motivation: multiple arrays, some accessed by rows and some by columns
Storing the arrays row by row (row major order) or column by column (column major
order) does not help: both rows and columns are used in every iteration of the loop (Loop
Interchange cannot help)
Solution: instead of operating on entire rows and columns of an array, blocked
algorithms operate on submatrices or blocks => maximize accesses to the data loaded
into the cache before the data is replaced

Blocking Example
/* Before */
for (i = 0; i < N; i = i+1)
for (j = 0; j < N; j = j+1)
{r = 0;
for (k = 0; k < N; k = k+1){
r = r + y[i][k]*z[k][j];};
x[i][j] = r;
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 131

};
/* After */
for (jj = 0; jj < N; jj = jj+B)
for (kk = 0; kk < N; kk = kk+B)
for (i = 0; i < N; i = i+1)
for (j = jj; j < min(jj+B,N); j = j+1)
{r = 0;
for (k = kk; k < min(kk+B,N); k = k + 1)
r = r + y[i][k]*z[k][j];
x[i][j] = x[i][j] + r;
};
Snapshot of x, y, z when
i=1
White:




White: not yet touched
Light: older access
Dark: newer access Before.
The Age of Accesses to x, y, Z



ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 132





Merging Arrays
Motivation: some programs reference multiple arrays in the same dimension with the
same indices at the same time => these accesses can interfere with each other,leading to
conflict misses
Solution: combine these independent matrices into a single compound array, so that a
single cache block can contain the desired elements Merging Arrays Example

Loop Fusion
Some programs have separate sections of code that access with the same loops,
performing different computations on the common data
Solution:
Fuse the code into a single loop => the data that are fetched into the cache can be used
repeatedly before being swapped out => reducing misses via improved temporal locality

Loop Fusion Example
Summary of Compiler Optimizations- to Reduce Cache Misses (by hand)

10. Reducing Misses by Hardware Prefetching of Instructions & Data Prefetching
relies on having extra memory bandwidth that can be used without penalty
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 133

Instruction Prefetching
Typically, CPU fetches 2 blocks on a miss: the requested block and the next
consecutive block.
Requested block is placed in instruction cache when it returns, and prefetched block is
placed into instruction stream buffer

Data Prefetching
Pentium 4 can prefetch data into L2 cache from up to 8 streams from 8 different 4 KB
pages
Prefetching invoked if 2 successive L2 cache misses to a page,if distance between those
cache blocks is < 256 bytes

11. Reducing Misses by Software Prefetching Data
Data Prefetch
Load data into register (HP PA-RISC loads)
Cache Prefetch: load into cache
(MIPS IV, PowerPC, SPARC v. 9)
Special prefetching instructions cannot cause faults;
a form of speculative execution
Issuing Prefetch Instructions takes time
Is cost of prefetch issues < savings in reduced misses?
Higher superscalar reduces difficulty of issue bandwi

Compiler Optimization vs. Memory Hierarchy Search
Compiler tries to figure out memory hierarchy optimizations
New approach: Auto-tuners 1st run variations of program on computer to find
best combinations of optimizations (blocking, padding, ) and algorithms, then produce
C code to be compiled for that computer
Auto-tuner targeted to numerical method
E.g., PHiPAC (BLAS), Atlas (BLAS), Sparsity (Sparse linear algebra),
Spiral (DSP), FFT-W
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 134


Cache Optimization Summary
Comparison of the 11 techniques





Main Memory Background
Performance of Main Memory:
Latency: Cache Miss Penalty
Access Time: time between request and word arrives
Cycle Time: time between requests
Bandwidth: I/O & Large Block Miss Penalty (L2)
Main Memory is DRAM: Dynamic Random Access Memory
Dynamic since needs to be refreshed periodically (8 ms, 1% time)
Addresses divided into 2 halves (Memory as a 2D matrix):
RAS or Row Access Strobe
CAS or Column Access Strobe
Cache uses SRAM: Static Random Access Memory
No refresh (6 transistors/bit vs. 1 transistor
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 135

Size: DRAM/SRAM - 4-8,
Cost/Cycle time: SRAM/DRAM - 8-16
Main Memory Deep Background
Out-of-Core, In-Core, Core Dump?
Core memory?
Non-volatile, magnetic
Lost to 4 Kbit DRAM (today using 512Mbit DRAM)
Access time 750 ns, cycle time 1500-3000 ns
DRAM logical organization (4 Mbit)
Quest for DRAM Performance
1. Fast Page mode
Add timing signals that allow repeated accesses to row buffer without
nother row access time
Such a buffer comes naturally, as each array will buffer 1024 to 2048
bits for each access
2. Synchronous DRAM (SDRAM)
Add a clock signal to DRAM interface, so that the repeated transfers
would not bear overhead to synchronize with DRAM controller
3. Double Data Rate (DDR SDRAM)
Transfer data on both the rising edge and falling edge of the DRAM
clock signal I doubling the peak data rate
DDR2 lowers power by dropping the voltage from 2.5 to 1.8 volts +
offers higher clock rates: up to 400 MHz
DDR3 drops to 1.5 volts + higher clock rates: up to 800 MHz
4.Improved Bandwidth, not Latency
DRAM name based on Peak Chip Transfers / Sec
DIMM name based on Peak DIMM MBytes / Sec
Need for Error Correction!
Motivation:
Failures/time proportional to number of bits!
As DRAM cells shrink, more vulnerable
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 136

Went through period in which failure rate was low enough without error
correction that people didnt do correction
DRAM banks too large now
Servers always corrected memory systems
Basic idea: add redundancy through parity bits
Common configuration: Random error correction
SEC-DED (single error correct, double error detect)
One example: 64 data bits + 8 parity bits (11% overhead)
Really want to handle failures of physical components as well
Organization is multiple DRAMs/DIMM, multiple DIMMs
Want to recover from failed DRAM and failed DIMM!
Chip kill handle failures width of single DRAM chip

DRAM Technology
Semiconductor Dynamic Random Access Memory
Emphasize on cost per bit and capacity
Multiplex address lines cutting # of address pins in half
Row access strobe (RAS) first, then column access strobe (CAS)
Memory as a 2D matrix rows go to a buffer
Subsequent CAS selects subrow
Use only a single transistor to store a bit
Reading that bit can destroy the information
Refresh each bit periodically (ex. 8 milliseconds) by writing back
Keep refreshing time less than 5% of the total time
DRAM capacity is 4 to 8 times that of SRAM
DIMM: Dual inline memory module
DRAM chips are commonly sold on small boards called DIMMs
DIMMs typically contain 4 to 16 DRAMs
Slowing down in DRAM capacity growth
Four times the capacity every three years, for more than 20 years
New chips only double capacity every two year, since 1998
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 137

DRAM performance is growing at a slower rate
RAS (related to latency): 5% per year
CAS (related to bandwidth): 10%+ per year

RAS improvement
SRAM Technology
Cache uses SRAM: Static Random Access Memory
SRAM uses six transistors per bit to prevent the information from being disturbed when
read
_no need to refresh
SRAM needs only minimal power to retain the charge in
the standby mode _ good for embedded applications
No difference between access time and cycle time for
SRAM
Emphasize on speed and capacity
SRAM address lines are not multiplexed
SRAM speed is 8 to 16x that of DRAM
Improving Memory Performance in a Standard DRAM Chip
Fast page mode: time signals that allow repeated accesses to buffer without another row
access time
Synchronous RAM (SDRAM): add a clock signal to DRAM interface, so that the
repeated transfer would not bear overhead to synchronize with the controller
Asynchronous DRAM involves overhead to sync with controller
Peak speed per memory module 8001200MB/sec in 2001
Double data rate (DDR): transfer data on both the rising edge and falling edge of
DRAM clock signal
Peak speed per memory module 16002400MB/sec in 2001


ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 138






Protection:
Virtual Memory and Virtual Machines
Slide Sources: Based on Computer Architecture by Hennessy/Patterson.
Supplemented from various freely downloadable sources
Security and Privacy
Innovations in Computer Architecture and System software
Protection through Virtual Memory
Protection from Virtual Machines
Architectural requirements
Performance
Protection via Virtual Memory
Processes
Running program
Environment (state) needed to continue running it
Protect Processes from each other
Page based virtual memory including TLB which caches page table
entries Example: Segmentation and paging in 80x86
Processes share hardware without interfering with each other
Provide User Process and Kernel Process
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 139

Readable portion of Processor state:
User supervisor mode bit
Exception enable/disable bit
Memory protection information
System call to transfer to supervisor mode
Return like normal subroutine to user mode
Mechanism to limit memory access
Memory protection
Virtual Memory
Restriction on each page entry in page table
Read, write, execute privileges
Only OS can update page table
TLB entries also have protection field
Bugs in OS
Lead to compromising security
Bugs likely due to huge size of OS code
Protection via Virtual Machines
Virtualization
Goal:
Run multiple instances of different OS on the same hardware
Present a transparent view of one or more environments (M-to-N mapping of M real
resources, N virtual resources)
Protection via Virtual Machines
Virtualization- cont.
Challenges:
Have to split all resources (processor, memory, hard drive, graphics card, networking
card etc.) among the different OS -> virtualize the resources
The OS can not be aware that it is using virtual resources instead of
real resources

Problems with virtualization
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 140

Two components when using virtualization:
Virtual Machine Monitor (VMM)
Virtual Machine(s) (VM)
Para-virtualization:
Operating System has been modified in order to run as a VM
Fully Virtualized:
No modification required of an OS to run as a VM

Virtual Machine Monitor-hypervisor
Isolates the state of each guest OS from each other
Protects itself from guest software
Determines how to map virtual resources to physical resources
Access to privileged state
Address translation
I/O
Exceptions and interrupts
Relatively small code ( compared to an OS)
VMM must run in a higher privilege mode than guest OS
Managing Virtual Memory
Virtual memory offers many of the features required for hardware virtualization
Separates the physical memory onto multiple processes
Each process thinks it has a linear address space of full size
Processor holds a page table translating virtual addresses used by a process and the
according physical memory
Additional information restricts processes from
Reading a page of on another process or
Allow reading but not modifying a memory page or
Do not allow to interpret data in the memory page as instructions and do not try to
execute them
Virtual Memory management thus requires
Mechanisms to limit memory access to protected memory
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 141

At least two modes of execution for instructions
Privileged mode: an instruction is allowed to do what it whatever it wants -> kernel
mode for OS
Non-privileged mode: user-level processes
Intel x86 Architecture: processor supports four levels
Level 0 used by OS
Level 3 used by regular applications
Provide mechanisms to go from non-privileged mode to privileged mode -> system call
Provide a portion of processor state that a user process can read but not modify
E.g. memory protection information
Each guest OS maintains its page tables to do the mapping from virtual address to
physical address
Most simple solution: VMM holds an additional table which maps the physical address
of a guest OS onto the machine address
Introduces a third level of redirection for every memory access
Alternative solution: VMM maintains a shadow page table of each guest OS
Copy of the page table of the OS
Page tables still works with regular physical addresses
Only modifications to the page table are intercepted by the VMM



Protection via Virtual Machines
-some definitions
VMs include all emulation methods to provide a standard software interface
Different ISAs can be used (emulated) on the native machine
When the ISAs match the hardware we call it (operating) system virtual machines
Multiple OSes all share the native hardware

Cost of Processor Virtualisation
VM is much smaller than traditional OS
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 142

Isolation portion is only about 10000 lines for a VMM
Processor bound programs have very little virtualisation overhead
I/O bound jobs have more overhead
ISA emulation is costly

Other benefits of VMs
Managing software
Complete software stack
Old Oses like DOS
Current stable OS
Next OS release
Managing Hardware
Multiple servers avoided
VMs enable hardware sharing
Migration of a running VM to another m/c
For balancing load or evacuate from failing HW
Requirements of a VMM
Guest sw should behave exactly on VM as if on native hw
Guest sw should not be able to change allocation of RT resources directly
Timer interrupts should be virtualised
Two processor modes- system and user
Priveleged subset of instruction available only in system mode




More issues on VM usage
ISA support for virtual machine
IBM360 support
80x86 do no support
Use of virtual memory
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 143

Concept of virtual- real- physical memories
Instead of extra indirection use shadow page table
Virtualising I/Os
More i/o
More diversity
Physical disks to partitioned virtual disks
Network cards time sliced
























ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 144


UNIT IV

HARDWARE AND SOFTWARE FOR VLIW AND EPIC

Loop Level Parallelism- Detection and Enhancement

Static Exploitation of ILP
Use compiler support for increasing parallelism
Supported by hardware
Techniques for eliminating some types of dependences
Applied at compile time (no run time support)
Finding parallelism
Reducing control and data dependencies
Using speculation
Unrolling Loops High-level
for (i=1000; i>0; i=i-1) x[i] = x[i] + s;
C equivalent of unrolling to block four iterations into one:
for (i=250; i>0; i=i-1)
{
x[4*i] = x[4*i] + s;
x[4*i-1] = x[4*i-1] + s;
x[4*i-2] = x[4*i-2] + s;
x[4*i-3] = x[4*i-3] + s;
}
Enhancing Loop-Level Parallelism
Consider the previous running example:
for (i=1000; i>0; i=i-1) x[i] = x[i] + s;
there is no loop-carried dependence where data used in a later iteration depends on
data produced in an earlier one
in other words, all iterations could (conceptually) be executed in parallel
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 145

Contrast with the following loop:
for (i=1; i<=100; i=i+1) { A[i+1] = A[i] + C[i]; /* S1 */
B[i+1] = B[i] + A[i+1]; /* S2 */ }
what are the dependences?
A Loop with Dependences
For the loop:
for (i=1; i<=100; i=i+1) { A[i+1] = A[i] + C[i]; /* S1 */
B[i+1] = B[i] + A[i+1]; /* S2 */ }

what are the dependences?
There are two different dependences:
loop-carried: (prevents parallel operation of iterations)
S1 computes A[i+1] using value of A[i] computed in previous iteration
S2 computes B[i+1] using value of B[i] computed in previous iteration
not loop-carried: (parallel operation of iterations is ok)
S2 uses the value A[i+1] computed by S1 in the same iteration
The loop-carried dependences in this case force successive iterations of the loop to
execute in series. Why?
S1 of iteration i depends on S1 of iteration i-1 which in turn depends on , etc. Another
Loop with Dependences
Generally, loop-carried dependences hinder ILP
if there are no loop-carried dependences all iterations could be executed in parallel
even if there are loop-carried dependences it may be possible to parallelize the loop an
analysis of the dependences is required

For the loop:
for (i=1; i<=100; i=i+1) { A[i] = A[i] + B[i]; /* S1 */
B[i+1] = C[i] + D[i]; /* S2 */ }

what are the dependences?
There is one loop-carried dependence:
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 146

S1 uses the value of B[i] computed in a previous iteration by S2
but this does not force iterations to execute in series. Why?
because S1 of iteration i depends on S2 of iteration i-1, and the chain of
dependences stops here!


Parallelizing Loops with Short Chains of Dependences
Parallelize the loop:
for (i=1; i<=100; i=i+1) { A[i] = A[i] + B[i]; /* S1 */
B[i+1] = C[i] + D[i]; /* S2 */ }
Parallelized code:
A[1] = A[1] + B[1];
for (i=1; i<=99; i=i+1)
{ B[i+1] = C[i] + D[i];
A[i+1] = A[i+1] + B[i+1];
}
B[101] = C[100] + D[100];
the dependence between the two statements in the loop is no longer loop-carried and
iterations of the loop may be executed in parallel

Loop-Carried Dependence Detection: affine array index: a x i+b
To detect loop-carried dependence in a loop, the Greatest Common Divisor (GCD) test
can be used by the compiler, which is based on the following:
If an array element with index: a x i + b is stored and element: c x i + d of the same array
is loaded later where index runs from m to n, a dependence exists if the following two
conditions hold:
1.There are two iteration indices, j and k , m <= j , k <= n (within iteration limits)
2.The loop stores into an array element indexed by:
a x j + b
and later loads from the same array the element indexed by:
c x k + d
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 147

Thus:
a x j + b = c x k + d
The Greatest Common Divisor (GCD) Test
If a loop carried dependence exists, then :
GCD(c, a) must divide (d-b)
The GCD test is sufficient to guarantee no loop carried dependence However there are
cases where GCD test succeeds but no dependence exits because GCD test does not take
loop bounds into account
Example:
for (i=1; i<=100; i=i+1) {
x[2*i+3] = x[2*i] * 5.0;
}
a = 2 b = 3 c = 2 d = 0
GCD(a, c) = 2
d - b = -3
2 does not divide -3 _ No loop carried dependence possible.

Example- Loop Iterations to be Independent
Finding multiple types of dependences
for (i=1; i<=100; i=i+1) {
Y[i] = X[i] / c; /* S1 */
X[i] = X[i] + c; /* S2 */
Z[i] = Y[i] + c; /* S3 */
Y[i] = c - Y[i]; /* S4 */ }
Answer The following dependences exist among the four statements:
1. There are true dependences from S1 to S3 and from S1 to S4 because of Y[i]. These
are not loop carried, so they do not prevent the loop from being considered parallel.
These dependences will force S3 and S4 to wait for S1 to complete.
2. There is an antidependence from S1 to S2, based on X[i].
3. There is an antidependence from S3 to S4 for Y[i].
4. There is an output dependence from S1 to S4, based on Y[i].
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 148


Eliminating false dependencies
The following version of the loop eliminates these false (or pseudo) dependences.
for (i=1; i<=100; i=i+1 {
/* Y renamed to T to remove output dependence */
T[i] = X[i] / c;
/* X renamed to X1 to remove antidependence */
X1[i] = X[i] + c;
/* Y renamed to T to remove antidependence */
Z[i] = T[i] + c;
Y[i] = c - T[i];
}
Drawback of dependence analysis
When objects are referenced via pointers rather than array indices (but see discussion
below)
When array indexing is indirect through another array, which happens with many
representations of sparse arrays
When a dependence may exist for some value of the inputs, but does not exist in
actuality when the code is run since the inputs never take on those values
When an optimization depends on knowing more than just the possibility of a
dependence, but needs to know on which write of a variable does a read of that variable
depend

Points-to analysis
Relies on information from three major sources:
1. Type information, which restricts what a pointer can point to.
2. Information derived when an object is allocated or when the address of an object is
taken, which can be used to restrict what a pointer can point to. For example, if p always
points to an object allocated in a given source line and q never points to that object, then
p and q can never point to the same object.
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 149

3. Information derived from pointer assignments. For example, if p may be assigned the
value of q, then p may point to anything q points to.

Eliminating dependent computations
copy propagation, used to simplify sequences like the following:
DADDUI R1,R2,#4
DADDUI R1,R1,#4
to
DADDUI R1,R2,#8
Tree height reduction
they reduce the height of the tree structure representing a computation, making it wider
but shorter.

Recurrence
Recurrences are expressions whose value on one iteration is given by a function that
depends onthe previous iterations.
sum = sum + x;
sum = sum + x1 + x2 + x3 + x4 + x5;
If unoptimized requires five dependent operations, but it can be rewritten as
sum = ((sum + x1) + (x2 + x3)) + (x4 + x5);
evaluated in only three dependent operations.

Scheduling and Structuring Code for Parallelism
Static Exploitation of ILP
Use compiler support for increasing parallelism
Supported by hardware
Techniques for eliminating some types of dependences
Applied at compile time (no run time support)
Finding parallelism
Reducing control and data dependencies
Using speculation
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 150


Techniques to increase the amount of ILP
For processor issuing more than one instruction on every clock cycle.
Loop unrolling,
software pipelining,
trace scheduling, and
superblock scheduling

Software pipelining
Symbolic loop unrolling
Benefits of loop unrolling with reduced code size
Instructions in loop body selected from different loop iterations
Increase distance between dependent instructions in

Software pipelined loop
Loop: SD F4,16(R1) #store to v[i]
ADDD F4,F0,F2 #add to v[i-1]
LD F0,0(R1) #load v[i-2]
ADDI R1,R1,-8
BNE R1,R2,Loop
5 cycles/iteration (with dynamic scheduling and renaming)
Need startup/cleanup code

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 151





SW pipelining example
Iteration i: L.D F0,0(R1)
ADD.D F4,F0,F2
S.D F4,0(R1)
Iteration i+1: L.D F0,0(R1)
ADD.D F4,F0,F2
S.D F4,0(R1)
Iteration i+2: L.D F0,0(R1)
ADD.D F4,F0,F2
S.D F4,0(R1)

SW pipelined loop with startup and cleanup code

#startup, assume i runs from 0 to n
ADDI R1,R1-16 #point to v[n-2]
LD F0,16(R1) #load v[n]
ADDD F4,F0,F2 #add v[n]
LD F0,8(R1) #load v[n-1]
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 152

#body for (i=2;i<=n-2;i++)
Loop: SD F4,16(R1) #store to v[i]
ADDD F4,F0,F2 #add to v[i-1]
LD F0,0(R1) #load v[i-2]
ADDI R1,R1,-8
BNE R1,R2,Loop
#cleanup
SD F4,8(R1) #store v[1]
ADDD F4,F0,F2 #add v[0]
SD F4,0(R1) #store v[0]
Software pipelining versus unrolling


Performance effects of SW pipelining vs. unrolling
Unrolling reduces loop overhead per iteration
SW pipelining reduces startup-cleanup pipeline overhead

Software pipelining versus unrolling (cont.)
Software pipelining


ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 153



Advantages
Less code space than conventional unrolling
Loop runs at peak speed during steady state
Overhead only at loop initiation and termination
Complements unrolling

Disadvantages
Hard to overlap long latencies
Unrolling combined with SW pipelining
Requires advanced compiler transformations

Global Code Scheduling
Global code scheduling aims to compact a code fragment with internal control structure
into the shortest possible sequence that preserves the data and control dependences.


ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 154






Global code scheduling
aims to compact a code fragment with internal control
structure into the shortest possible sequence
that preserves the data and control dependences
Data dependences are overcome by unrolling
In the case of memory operations, using dependence analysis to determine if two
references refer to the same address.
Finding the shortest possible sequence of dependent instructions- critical path
Reduce the effect of control dependences arising from conditional nonloop branches by
moving code.
Since moving code across branches will often affect the frequency of execution of such
code, effectively using global code motion requires estimates of the relative frequency of
different paths.
if the frequency information is accurate, is likely to lead to faster code. Global code
scheduling- cont.
Global code motion is important since many inner loops contain conditional statements.
Effectively scheduling this code could require that we move the assignments to B and C
to earlier in the execution sequence, before the test of A.

Factors for compiler
Global code scheduling is an extremely complex problem
What are the relative execution frequencies
What is the cost of executing the computation
How will the movement of B change the execution time
Is B the best code fragment that can be moved
What is the cost of the compensation code
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 155





Trace Scheduling:
Focusing on the Critical Path


Code generation sequence

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 156


Trace Scheduling,
Superblocks and Predicated Instructions
For processor issuing more than one instruction on every clock cycle.
Loop unrolling,
software pipelining,
trace scheduling, and
superblock scheduling

Trace Scheduling
Used when
Predicated execution is not supported
Unrolling is insufficient
Best used
If profile information clearly favors one path over the other
Significant overheads are added to the infrequent path
Two steps :
Trace Selection
Trace Compaction
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 157



Trace Selection
Likely sequence of basic blocks that can be put together
Sequence is called a trace
What can you select?
Loop unrolling generates long traces
Static branch prediction forces some straight-line code behavior
Trace Selection

(cont.)
Trace Example
If the shaded portion in previous code was frequent path and it was unrolled 4 times : _
Trace exits are jumps off the frequent path
Trace entrances are returns to the trace
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 158


Trace Compaction
Squeeze these into smaller number of wide instructions
Move operations as early as it can be in a trace
Pack the instructions into as few wide instructions as possible
Simplifies the decisions concerning global code motion
All branches are viewed as jumps into or out of the trace
Bookkeeping
Cost is assumed to be little
Best used in scientific code with extensive loops





Super Blocks for Global Scheduling
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 159

Motivation :
Entries and exits into trace schedule code are complicated
Compiler cannot do a good cost analysis about compensation code
Superblocks
Are like traces
One entry point
Unlike traces
Different exit points
Common in for loops
Single entry and exit points
Code motion across exit only need be considered


Superblock Construction

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 160







Tail duplication
Creates a separate block that corresponds to the portion of trace after the entry
When proceeding as per prediction Take the path of superblock code
When exit from
superblock
Residual loop that handles rest of the iterations

Analysis on Superblocks
Reduces the complexity of bookkeeping and scheduling
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 161

Unlike the trace approach
Can have larger code size though
Assessing the cost of duplication
Compilation process is not simple any more

H/W Support : Conditional Execution
Also known as Predicated Execution
Enhancement to instruction set
Can be used to eliminate branches
All control dependences are converted to data dependences
Instruction refers to a condition
Evaluated as part of the execution
True?
Executed normally
False?
Execution continues as if the instruction were a no-op
Example :
Conditional move between registers

Example
if (A==0)
S = T;
Straightforward Code
BNEZ R1, L;
ADDU R2, R3, R0
L:
Conditional Code
CMOVZ R2, R3, R1
Annulled if R1 is not 0
Conditional Instruction
Can convert control to data dependence
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 162

In vector computing, its called if conversion.
Traditionally, in a pipelined system
Dependence has to be resolved closer to front of pipeline
For conditional execution
Dependence is resolved at end of pipeline, closer to the register write

Another example
A = abs(B)
if (B < 0)
A = -B;
else
A = B;
Two conditional moves
One unconditional and one conditional move
The branch condition has moved into the
instruction
Control dependence becomes data dependence

Limitations of Conditional Moves
Conditional moves are the simplest form of predicated instructions
Useful for short sequences
For large code, this can be inefficient
Introduces many conditional moves
Some architectures support full predication
All instructions, not just moves
Very useful in global scheduling
Can handle nonloop branches nicely
Eg : The whole if portion can be predicated if the frequent path is not taken


ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 163


Assume : Two issues, one to ALU and one to memory; or branch by itself
Wastes a memory operation slot in second cycle
Can incur a data dependence stall if branch is not taken
R9 depends on R8

Predicated Execution
Assume : LWC is predicated load and loads if third operand is not 0








One instruction issue slot is eliminated
On mispredicted branch, predicated instruction will not have any effect
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 164

If sequence following the branch is short, the entire block of the code can be predicated

Some Complications
Exception Behavior
Must not generate exception if the predicate is false
If R10 is zero in the previous example
LW R8, 0(R10) can cause a protection fault
If condition is satisfied
A page fault can still occur
Biggest Issue Decide when to annul an instruction
Can be done during issue
Early in pipeline
Value of condition must be known early, can induce stalls
Can be done before commit
Modern processors do this
Annulled instructions will use functional resources
Register forwarding and such can complicate implementation


Limitations of Predicated Instructions
Annulled instructions still take resources
Fetch and execute atleast
For longer code sequences, benefits of conditional move vs branch is not clear
Only useful when predicate can be evaluated early in the instruction stream
What if there are multiple branches?
Predicate on two values?
Higher cycle count or slower clock rate for predicated instructions
More hardware overhead
MIPS, Alpha, Pentium etc support partial predication
IA-64 has full predication

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 165


Hardware support for Compiler Speculation
H/W Support : Conditional Execution
Also known as Predicated Execution
Enhancement to instruction set
Can be used to eliminate branches
All control dependences are converted to data dependences
Instruction refers to a condition
Evaluated as part of the execution
True?
Executed normally
False?
Execution continues as if the instruction were a no-op
Example :
Conditional move between registers
Example
if (A==0)
S = T;
Straightforward Code
BNEZ R1, L;
ADDU R2, R3, R0
L:
Conditional Code
CMOVZ R2, R3, R1
Annulled if R1 is not 0
Conditional Instruction
Can convert control to data dependence
In vector computing, its called if conversion.
Traditionally, in a pipelined system
Dependence has to be resolved closer to front of pipeline
For conditional execution
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 166

Dependence is resolved at end of pipeline, closer to the register write
Another example
A = abs(B)
if (B < 0)
A = -B;
else
A = B;
Two conditional moves
One unconditional and one conditional move
The branch condition has moved into the
instruction
Control dependence becomes data dependence

Limitations of Conditional Moves
Conditional moves are the simplest form of predicated instructions
Useful for short sequences
For large code, this can be inefficient
Introduces many conditional moves
Some architectures support full predication
All instructions, not just moves
Very useful in global scheduling
Can handle nonloop branches nicely
Eg : The whole if portion can be predicated if the frequent path is not taken Example
Assume : Two issues, one to ALU and one to memory; or branch by itself
Wastes a memory operation slot in second cycle
Can incur a data dependence stall if branch is not taken
R9 depends on R8

Predicated Execution
Assume : LWC is predicated load and loads if third operand is not 0
One instruction issue slot is eliminated
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 167

On mispredicted branch, predicated instruction will not have any effect
If sequence following the branch is short, the entire block of the code can be predicated

Predication
Some Complications
Exception Behavior
Must not generate exception if the predicate is false
If R10 is zero in the previous example
LW R8, 0(R10) can cause a protection fault
If condition is satisfied
A page fault can still occur
Biggest Issue Decide when to annul an instruction
Can be done during issue
-- Early in pipeline
Value of condition must be known early, can induce stalls
Can be done before commit
Modern processors do this
Annulled instructions will use functional resources
Register forwarding and such can complicate implementation

Limitations of Predicated Instructions
Annulled instructions still take resources
Fetch and execute atleast
For longer code sequences, benefits of conditional move vs branch is not clear
Only useful when predicate can be evaluated early in the instruction stream
What if there are multiple branches?
Predicate on two values?
Higher cycle count or slower clock rate for predicated instructions
More hardware overhead
MIPS, Alpha, Pentium etc support partial predication
IA-64 has full predication
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 168

UNIT V

LARGE SCALE MULTIPROCESSORS AND SCIENTIFIC
APPLICATION

Preserve control and data flow, precise interrupts in Predication
Speculative predicated instructions may not throw illegal exceptions
LWC may not throw exception if R10 == 0
LWC may throw recoverable page fault if R10 6= 0
Instruction conversion to nop
Early condition detection may not be possible due to data dependence
Late condition detection incurs stalls and consumes pipeline resources needlessly
Instructions may be dependent on multiple branches
Compiler able to find instruction slots and reorder


Hardware support for speculation
Alternatives for handling speculative exceptions
Hardware and OS ignore exceptions from speculative instructions
Mark speculative instructions and check for exceptions
Additional instructions to check for exceptions and recover
Registers marked with poison bits to catch exceptions upon read
Hardware buffers instruction results until instruction is no longer speculative

Exception classes
Recoverable: exception from speculative instruction may harm performance, but not
preciseness
Unrecoverable: exception from speculative instruction compromises preciseness

Solution I: Ignore exceptions
HW/SW solution
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 169

Instruction causing exception returns undefined value
Value not used if instruction is speculative
Incorrect result if instruction is non-speculative
Compiler generates code to throw regular exception
Rename registers receiving speculative results

Solution I: Ignore exceptions
Example




Solution II: mark speculative instructions



ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 170




R4 marked with poison bit

Use of R4 in SD raises exception if SLD raises exception
Generate exception when result of offending instruction is used for the first time
OS code needs to save poison bits during context switching

Solution IV HW mechanism like a ROB
Instructions are marked as speculative
How many branches speculatively moved
Action (T/NT) assumed by compiler
Usually only one branch
Other functions like a ROB

HW support for Memory Reference Speculation
Moving stores across loads
To avoid address conflict
Special instruction checks for address conflict
Left at original location of load instruction
Acts like a guardian
On speculative load HW saves address
Speculation failed if a stores changes this address before check
nstruction
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 171

Fix-up code re-executes all speculated instructions

IA-64 and Itanium Processor
Introducing The IA-64 Architecture

Itanium and Itanium2 Processor
Slide Sources: Based on Computer Architecture by Hennessy/Patterson.
Supplemented from various freely downloadable sources IA-64 is an EPIC

IA-64 largely depends on software for parallelism
VLIW Very Long Instruction Word
EPIC Explicitly Parallel Instruction Computer
VLIW points
VLIW Overview
RISC technique
Bundles of instructions to be run in parallel
Similar to superscaling
Uses compiler instead of branch prediction hardware


EPIC
EPIC Overview
Builds on VLIW
Redefines instruction format
Instruction coding tells CPU how to process data
Very compiler dependent
Predicated execution
EPIC pros and cons
EPIC Pros:
Compiler has more time to spend with code
Time spent by compiler is a one-time cost
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 172

Reduces circuit complexity





Chip Layout
Itanium Architecture Diagram




ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 173



Itanium Specs
4 Integer ALU's
4 multimedia ALU's
2 Extended Precision FP Units
2 Single Precision FP units
2 Load or Store Units
3 Branch Units
10 Stage 6 Wide Pipeline
32k L1 Cache
96K L2 Cache
4MB L3 Cache(extern)
800Mhz Clock


Intel Itanium
800 MHz
10 stage pipeline
Can issue 6 instructions (2 bundles) per cycle
4 Integer, 4 Floating Point, 4 Multimedia, 2 Memory, 3 Branch Units
32 KB L1, 96 KB L2, 4 MB L3 caches
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 174

2.1 GB/s memory bandwidth






Itanium2 Specs
6 Integer ALU's
6 multimedia ALU's
2 Extended Precision FP Units
2 Single Precision FP units
2 Load and Store Units
3 Branch Units
8 Stage 6 Wide Pipeline
32k L1 Cache
256K L2 Cache
3MB L3 Cache(on die)
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 175

1Ghz Clock initially
Up to 1.66Ghz on Montvale

Itanium2 Improvements
Initially a 180nm process
Increased to 130nm in 2003
Further increased to 90nm in 2007
Improved Thermal Management
Clock Speed increased to 1.0Ghz
Bus Speed Increase from 266Mhz to 400Mhz

L3 cache moved on die
Faster access rate

IA-64 Pipeline Features
Branch Prediction
Predicate Registers allow branches to be turned on or off
Compiler can provide branch prediction hints
Register Rotation
Allows faster loop execution in parallel
Predication Controls Pipeline Stages
Cache Features
L1 Cache
4 way associative
16Kb Instruction
16Kb Data
L2 Cache
Itanium
6 way associative
96 Kb
Itanium2
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 176

8 way associative
256 Kb Initially
256Kb Data and 1Mb Instruction on Montvale!
Cache Features
L3 Cache
Itanium
4 way associative
Accessible through FSB
2-4Mb
Itanium2
2 4 way associative
On Die
3Mb
Up to 24Mb on Montvale chips(12Mb/core)!
Register
Specification
_128, 65-bit General Purpose Registers
_128, 82-bit Floating Point Registers
_128, 64-bit Application Registers
_8, 64-bit Branch Registers
_64, 1-bit Predicate Registers
Register Model
_128 General and Floating Point Registers
_32 always available, 96 on stack
_As functions are called, compiler allocates a specific number of local and
output
registers to use in the function by using register allocation instruction
Alloc.
_Programs renames registers to start from 32 to 127.
_Register Stack Engine (RSE) automatically saves/restores stack to
memory when needed
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 177

_RSE may be designed to utilize unused memory bandwidth to perform
register spill and fill operations in the background


On function call, machine shifts register window such that previous output registers
become new locals starting at r32

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 178





ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 179





ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 180





ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 181






Instruction Encoding
Each instruction includes the opcode and three operands
Each instructions holds the identifier for a corresponding Predicate Register
Each bundle contains 3 independent instructions
Each instruction is 41 bits wide
Each bundle also holds a 5 bit template field

Distributing Responsibility
_ILP Instruction Groups
_Control flow parallelism
Parallel comparison
Multiway branches
_Influencing dynamic events
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 182

Provides an extensive set of hints that the compiler uses to tell the hardware about likely
branch behavior (taken or not taken, amount to fetch at branch target) and memory
operations (in what level of the memory hierarchy to cache data).




_Use predicates to eliminate branches, move instructions across branches
_Conditional execution of an instruction based on predicate register (64 1-bit
predicate registers)
_Predicates are set by compare instructions
_Most instructions can be predicated each instruction code contains predicate
field
_If predicate is true, the instruction updates the computation state; otherwise, it
behaves like a nop

Scheduling and Speculation
Basic block: code with single entry and exit, exit point can be multiway branch
Control Improve ILP by statically move ahead long latency code blocks.
path is a frequent execution path
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 183

Schedule for control paths
Because of branches and loops, only small percentage of code is executed regularly
Analyze dependences in blocks and paths
Compiler can analyze more efficiently - more time, memory, larger view of the program
Compiler can locate and optimize the commonly executed blocks




Control speculation
_ Not all the branches can be removed using predication.
_ Loads have longer latency than most instructions and tend to start timecritical
chains of instructions
_ Constraints on code motion on loads limit parallelism
_ Non-EPIC architectures constrain motion of load instruction
_ IA-64: Speculative loads, can safely schedule load instruction before one or more
prior branches
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 184


Control Speculation
_Exceptions are handled by setting NaT (Not a Thing) in target register
_Check instruction-branch to fix-up code if NaT flag set
_Fix-up code: generated by compiler, handles exceptions
_NaT bit propagates in execution (almost all IA-64 instructions)
_NaT propagation reduces required check points

Speculative Load
_ Load instruction (ld.s) can be moved outside of a basic block even if branch target
is not known
_ Speculative loads does not produce exception - it sets the NaT
_ Check instruction (chk.s) will jump to fix-up code if NaT is set Data Speculation
_ The compiler may not be able to determine the location in memory being referenced
(pointers)
_ Want to move calculations ahead of a possible memory dependency
_ Traditionally, given a store followed by a load, if the compiler cannot determine if the
addresses will be equal, the load cannot be moved ahead of the store.
_ IA-64: allows compiler to schedule a load before one or more stores
_ Use advance load (ld.a) and check (chk.a) to implement
_ ALAT (Advanced Load Address Table) records target register, memory address
accessed, and access size

Data Speculation
1. Allows for loads to be moved ahead of stores even if the compiler is unsure if
addresses are the same
2. A speculative load generates an entry in the ALAT
3. A store removes every entry in the ALAT that have the same address
4. Check instruction will branch to fix-up if the given address is not in the ALAT


ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 185




ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 186


ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 187

Use address field as the key for comparison
If an address cannot be found, run recovery code
ALAT are smaller and simpler implementation than equivalent structures for
superscalars


Register Model
_128 General and Floating Point Registers
_32 always available, 96 on stack
_As functions are called, compiler allocates a specific number of local and
output registers to use in the function by using register allocation instruction Alloc.
_Programs renames registers to start from 32 to 127.
_Register Stack Engine (RSE) automatically saves/restores stack to memory when
needed
_RSE may be designed to utilize unused memory bandwidth to perform register
spill and fill operations in the background





On function call, machine shifts register window such that previous output registers
become new locals starting at r32
Software Pipelining
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 188

_loops generally encompass a large portion of a programs execution time, so its
important to expose as much loop-level parallelism as possible.
_Overlapping one loop iteration with the next can often increase the parallelism.
Software Pipelining



We can implement loops in parallel by resolve some problems.
_Managing the loop count,
_Handling the renaming of registers for the pipeline,
_Finishing the work in progress when the loop ends,
_Starting the pipeline when the loop is entered, and
_Unrolling to expose cross-iteration parallelism.
IA-64 gives hardware support to compilers managing a software pipeline
Facilities for managing loop count, loop termination, and rotating registers
The combination of these loop features and predication enables the compiler to
generate compact code, which performs the essential work of the loop in a highly parallel
form.

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 189

UNIT VI
COMPUTER ARITHMETIC

Introduction, Basic Techniques of Integer Arithmetic: Although computer arithmetic is
sometimes viewed as a specialized part of CPU design, it is very important part. This was
brought home for Intel in 1994 whentheir Pentium chip was discovered to have a bug in
the divide algorithm. This floating-point flaw resulted in a flurry of bad publicity for Intel
and also cost them a lot of money. Intel took a $300 million write-off to cover the cost of
replacing the buggy chips.
In this appendix we will study some basic floating-point algorithms, including the
division algorithm used on the Pentium. Although a tremendous variety of algorithms
have been proposed or use in floating-point accelerators, actual implementations are
usually based on refinements nd variations of the few basic algorithms presented here. In
addition to choosing algorithms for addition, subtraction, multiplication, and division, the
computer architect must make other choices. What precisions should be implemented?
How should exceptions be handled? This appendix will give you the background for
making these and other decisions.
Our discussion of floating point will focus almost exclusively on the IEEE floating-point
standard (IEEE 754) because of its rapidly increasing acceptance. Although floating-
point arithmetic involves manipulating exponents and shifting fractions, the bulk of the
time in floating-point operations is spent operating on fractions using integer algorithms
(but not necessarily sharing the hardware that implements integer instructions). Thus,
after our discussion of floating point, we
will take a more detailed look at integer algorithms.
Some good references on computer arithmetic, in order from least to most detailed, are
Chapter 4 of Patterson and Hennessy [1994]; Chapter 7 of Hamacher, Vranesic, and Zaky
[1984]; Gosling [1980]; and Scott [1985]. Readers who have studied computer arithmetic
before will find most of this section to be review.

Ripple-Carry Addition
Adders are usually implemented by combining multiple copies of simple components.
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 190

The natural components for addition are half adders and full adders
The half adder takes two bits a and b as input and produces a sum bit s and a carry bit c
out as output. Mathematically, s= (a + b ) mod 2 , and c out = ( a + b )/2
The half adder is also called a (2,2) adder, since it takes two inputs and produces two
outputs. The full adder is a (3,2) adder and is defined by or the logic equations

Basic Techniques of Integer Arithmetic
The principal problem in constructing an adder for n -bit numbers out of smaller pieces is
propagating the carries from one piece to the next. The most obvious way to solve this is
with a
ripple-carry adder, consisting of n full adders, as illustrated in Figure H.1. (In the figures
in this appendix, the least-significant bit is always on the right.) The inputs to the adder
are a represents the number output of the I th adder is fed into the c I +1 input of thenext
adder (the (I + 1)-th adder) with the lower-order carry-in c 0 set to 0. Since
the low-order carry-in is wired to 0, the low-order adder could be a half adder. Later,
however, we will see that setting the low-order carry-in bit to 1 is useful for performing
subtraction. In general, the time a circuit takes to produce an output is proportional to the
maximum numb er of logic levels through which a signal travels. However, determining
the exact relationship between logic levels and timings is highly technology dependent.
Therefore, when comparing adders we will simply compare the number of logic levels in
each one. How many levels are there for a ripple-carry adder? It takes two levels to
compute are 32 for integer arithmetic and 53 for doubleprecision floating point. The
ripple-carry adder is the slowest adder, but also the cheapest. It can be built with only n
simple cells, connected in a simple, regular way. Because the ripple-carry adder is
relatively slow compared with the designs discussed in Secti on H.8, you might wonder
why it is used at all. In technologies like CMOS, even though ripple adders take time O(
n ), the constant factor is very small. In such cases short ripple adders are often used as
building blocks in larger adders.

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 191

Ripple-carry adder, consisting of n full adders.The carry-out of one full adder is
connected to the carry-in of the adder for the next most-significant bit. The carries ripple
from the least-significant bit (on the right) to the most-significant bit (on the left).

Radix-2 Multiplication and Division
The simplest multiplier computes the product of two unsigned numbers, one bit at a time,
as illustrated in Figure H.2(a). The numbers to be multiplied are an1an2 a0
and bn1bn2 b0, and they are placed in registers A and B, respectively.
Register P is initially 0. Each multiply step has two parts.
Multiply Step (i) If the least-significant bit of A is 1, then register B, containing bn1bn
2 0, is added to P; otherwise 00 00 is added to P. The sum is placed back
into P.

Each multiplication step consists of adding the contents of P to either B or 0 (depending
on the low-order bit of A), replacing P with the sum, and then shifting both P and A one
bit right. Each division step involves first shifting P and A one bit left, subtracting B from
P, and, if the difference is nonnegative, putting it into P. If the

Floating Point, Floating-Point Multiplication,
Many applications require numbers that arent integers. There are a number of ways that
nonintegers can be represented. One is to use fixed point; that is, use integer arithmetic
and simply imagine the binary point somewhere other than just to the right of the least-
significant digit. Adding two such numbers can be done with an integer add, whereas
multiplication requires some extra shifting. Other representations that have been
proposed involve storing the logarithm of a number and doing multiplication by adding
the logarithms, or using a pair of integers (a,b) to represent the fraction a/b. However,
only one noninteger representation has gained widespread use, and that is floating point.
In this system, a computer word is divided int o two parts, an exponent and a significand.
As an example, an exponent of 3 and significand of 1.5 might represent the number 1.5
23 = 0.1875. The advantages of standardizing a particular representation are
obvious. Numerical analysts can build up high-quality software libraries, computer
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 192

designers can develop techniques for implementing high-performance hardware, and
hardware vendors can build standard accelerators. Given the predominance of the
floating-point representation, it appears unlikely that any other representation will come
into widespread use.
The semantics of floating-point instructions are not as clear-cut as the semantics of the
rest of the instruction set, and in the past the behavior of floating-point operations varied
considerably from one computer family to the next. The variations involved such things
as the number of bits allocated to the exponent and significand, the range of exponents,
how rounding was carried out, and the actions taken on exceptional conditions like
underflow and overflow.
Computer architecture books used to dispense advice on how to deal with all these
details, but fortunately this is no longer necessary. Thats because the computer industry
is rapidly converging on the format specified by IEEE standard 754-1985 (also an
international standard, IEC 559). The advantages of using a standard variant of floating
point are similar to those for using floating point over other noninteger representations.
1. When rounding a halfway result to the nearest floating-point number, it picks the one
that is even.
2. It includes the special values NaN, , and .
3. It uses denormal numbers to represent the result of computations whose value is less
than 1.0
4. It rounds to nearest by default, but it also has three other rounding modes.
5. It has sophisticated facilities for handling exceptions.
To elaborate on (1), note that when operating on two floating-point numbers,
the result is usually a number that cannot be exactly represented as another floating- point
number. For example, in a floating-point system using base 10 and two significant digits,
6.1 0.5 = 3.05. This needs to be rounded to two digits. Should it be rounded to 3.0 or
3.1? In the IEEE standard, such halfway cases are rounded to the number whose low-
order digit is even. That is, 3.05 rounds to 3.0, not 3.1. The standard actually has four
rounding modes. The default is round to nearest, which rounds ties to an even number as
just explained. The other modes are round toward 0, round toward +, and round toward
. We will elaborate on the other differences in following sections
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 193

Special Values and Denormals
Probably the most notable feature of the standard is that by default a computation
continues in the face of exceptional conditions, such as dividing by 0 or taking the square
root of a negative number. For example, the result of taking the square root of a negative
number I s a NaN (Not a Number), a bit pattern that does not represent an ordinary
number. As an example of how NaNs might be useful, consider the code for a zero finder
that takes a function F as an argument and evaluates F at various points to determine a
zero for it. If the zero finder accidentally probes outside the valid values for F, F may
well cause an exception. Writing a zero finder that deals with this case is highly language
and operating-system dependent, because it relies on how the operating system reacts to
exceptions and how this reaction is mapped back into the programming language. In
IEEE arithmetic
it is easy to write a zero finder that handles this situation and runs on many different
systems. After each evaluation of F, it simply checks to see whether F has returned a
NaN; if so, it knows it has probed outside the domain of F. In IEEE arithmetic, if the
input to an operation is a NaN, the output is NaN (e.g., 3 + NaN = NaN). Because of this
rule, writing floating-point subroutines that can accept NaN as an argument rarely
requires any special case checks. For example, suppose that arccos is computed in terms
of arctan, using the formula arccos x = 2 arctan( ). If arctan handles an argument of NaN
properly, arccos will automatically do so too. Thats because if x is a NaN, 1 + x, 1 x,
(1 + x)/(1 x), and will also be NaNs. No checking for NaNs is required.
1 x1 + x
1 x1 + x


While the result of is a NaN, the result of 1/0 is not a NaN, but +which is another
special value. The standard defines arithmetic on infinities (there is both +and )
using rules such as 1/= 0. The formula arccos x = 2 arctan( ) illustrates how infinity
arithmetic can be used. Since arctan x asymptotically approaches /2 as x approaches
, it is natural to define arctan() = /2, in which case arccos(1) will automatically
be computed correctly as 2 arctan() = . The final kind of special values in the standard
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 194

are denormal numbers. In many floating-point systems, if Emin is the smallest exponent,
a number less than 1.0 2Emin cannot be represented, and a floating-point operation
that results in a number less than this is simply flushed to 0. In the IEEE standard, on the
other hand, numbers less than 1.0 2Emin are represented using significands less than
1.
This is called gradual underflow. Thus, as numbers decrease in magnitude below 2Emin,
they gradually lose their significance and are only represented by 0 when all their
significance has been shifted out. For example, in base 10 with four significant figures,
let x = 1.234 10Emin. Then x/10 will be rounded to 0.123 10Emin, having lost a
digit of precision. Similarly x/100 rounds to 0.012 10Emin, and x/1000 to 0.001
10Emin, while x/10000 is finally small enough to be rounded to 0. Denormals make
dealing with small numbers more predictable by maintaining familiar properties such as x
= y x y = 0. For example, in a flushto- zero system (again in base 10 with four
significant digits), if x = 1.256 10Emin and y = 1.234 10Emin, then x y =
0.022 10Emin, which flushes to zero. So even though x y, the computed value of x
y = 0. This never happens with gradual underflow. In this example, x y = 0.022
10Emin is a denormal number, and so the computation of x y is exact.

Representation of Floating-Point Numbers
Let us consider how to represent single-precision numbers in IEEE arithmetic. Single-
precision numbers are stored in 32 bits: 1 for the sign, 8 for the exponent, and 23 for the
fraction. The exponent is a signed number represented using the bias method (see the
subsection Signed Numbers, page H-7) with a bias of 127. The term biased exponent
refers to the unsigned number contained in bits 1 through 8 and unbiased exponent (or
just exponent) means the actual power to
which 2 is to be raised. The fraction represents a number less than 1, but the significand
of the floating-point number is 1 plus the fraction part. In other words, if e is the biased
exponent (value of the exponent field) and f is the value of the fraction field, the number
being represented is 1. f 2e127.

Example What single-precision number does the following 32-bit word represent?
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 195

1 10000001 01000000000000000000000 1 1 x1 + x
H-16 _ Appendix H Computer Arithmetic
Answer Considered as an unsigned number, the exponent field is 129, making the value
of the exponent 129 127 =2. The fraction part is .012 = .25, making the significand
1.25. Thus, this bit pattern represents the number 1.25 22 = 5. The fractional part
of a floating-point number (.25 in the example above) must not be confused with the
significand, which is 1 plus the fractional part. The leading 1 in the significand 1. f does
not appear in the representation; that is, the leading bit is implicit. When performing
arithmetic on IEEE format numbers, the fraction part is usually unpacked, which is to say
the implicit 1 is made explicit.
Figure H.7 summarizes the parameters for single (and other) precisions. It shows the
exponents for single precision to range from 126 to 127; accordingly, the biased
exponents range from 1 to 254. The biased exponents of 0 and 255 are used to represent
special values. This is summarized in Figure H.8. When the biased exponent is 255, a
zero fraction field represents infinity, and a nonzero fraction field represents a NaN.
Thus, there is an entire family of NaNs. When the biased exponent and the fraction field
are 0, then the number represented is 0.
Because of the implicit leading 1, ordinary numbers always have a significand greater
than or equal to 1. Thus, a special convention such as this is required to represent 0.
Denormalized numbers are implemented by having a word with a zero exponent field
represent the number 0. f

Format parameters for the IEEE 754 floating-point standard. The first row gives the
number of bits in the significand. The blanks are unspecified parameters.
outside the range Emin e Emax, then that number has a special interpretation as
indicated in the table.
The primary reason why the IEEE standard, like most other floating-point formats, uses
biased exponents is that it means nonnegative numbers are ordered in the same way as
integers. That is, the magnitude of floating-point numbers can be compared using an
integer comparator. Another (related) advantage is that 0 is represented by a word of all
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 196

0s. The downside of biased exponents is that adding them is slightly awkward, because it
requires that the bias be subtracted from their sum.
The simplest floating-point operation is multiplication, so we discuss it first. A binary
floating-point number x is represented as a significand and an exponent, x = s 2e. The
formula (s1 2e1) (s2 2e2) = (s1 s2) 2e1+e2 shows that a floating-point
multiply algorithm has several parts. The first part multiplies the significands using
ordinary integer multiplication. Because floatingpoint numbers are stored in sign
magnitude form, the multiplier need only deal with unsigned numbers (although we have
seen that Booth recoding handles signed twos complement numbers painlessly). The
second part rounds the result.
If the significands are unsigned p-bit numbers (e.g., p = 24 for single precision), then the
product can have as many as 2p bits and must be rounded to a p-bit number. The third
part computes the new exponent. Because exponents are stored with a bias, this involves
subtracting the bias from the sum of the biased exponents.

Floating-Point Multiplication
Appendix H Computer Arithmetic
10000010
10000011
+ 10000001
10000110
Since this is 134 decimal, it represents an exponent of 134 bias = 134 127 = 7, as
expected.
The interesting part of floating-point multiplication is rounding. Some of the different
cases that can occur are illustrated in Figure H.9. Since the cases are similar in all bases,
the figure uses human-friendly base 10, rather than base 2. In the figure, p = 3, so the
final result must be rounded to three significant digits. The three most-significant digits
are in boldface. The fourth most-significant digit (marked with an arrow) is the round
digit, denoted by r. If the round digit is less than 5, then the bold digits represent the
rounded result. If the round digit is greater than 5 (as in (a)), then 1 must be added to the
least-significant bold digit. If the round digit is exactly 5 (as in (b)), then additional digits
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 197

must be examined to decide between truncation or incrementing by 1. It is only necessary
to know if any digits past 5 are nonzero. In the algorithm below, this will be recorded in a
sticky bit. Comparing (a) and (b) in the figure shows that there are two possible positions
for the round digit (relative to the least-significant digit of the product). Case (c)
illustrates that when adding 1 to the least-significant bold digit, there may be a carry-out.
When this happens, the final significand must be 10.0.
There is a straightforward method of handling rounding using the multiplier of Figure
H.2 (page H-4) together with an extra sticky bit. If p is the number of bits in the
significand, then the A, B, and P registers should be p bits wide. Multiply the two
significands to obtain a 2p-bit product in the (P,A) registers (see Figure H.9 Examples
of rounding a multiplication. Using base 10 and p = 3, parts (a) and (b) illustrate that
the result of a multiplication can have either 2p 1 or 2p digits, and hence the position
where a 1 is added when rounding up (just left of the arrow) can vary. Part (c) shows that
rounding up can cause a carry-out.

Denormals
Checking for underflow is somewhat more complex because of denormals. In single
precision, if the result has an exponent less than 126, that does not necessarily indicate
underflow, because the result might be a denormal number. For example, the product of
(1 264) with (1 265) is 1 2129, and 129 is below the legal exponent
limit. But this result is a valid denormal number, namely, 0.125 2126. In general,
when the unbiased exponent of a product dips below 126, the
resulting product must be shifted right and the exponent incremented until the Rounding
mode Sign of result 0 Sign of result < 0 +1 if r s ++1 if r s 0 Nearest
+1 if r p0 or r s +1 if r p0 or r s
Let S be the magnitude of the preliminary result. Blanks mean that the p most-significant
bits of S are the ctual result bits. If the condition listed is true, add 1 to the pth most-
significant bit of S.
The symbols r and s represent the round and sticky bits, while p0 is the pth
mostsignificant bit of S.
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 198

If this process causes the entire significand to be shifted out, then underflow has
occurred. The precise definition of underflow is somewhat subtlesee Section H.7 for
details. When one of the operands of a multiplication is denormal, its significand will
have leading zeros, and so the product of the significands will also have leading zeros. If
the exponent of the product is less than 126, then the result is denormal, so right-shift
and increment the exponent as before. If the exponent is greater than 126, the result may
be a normalized number. In this case, left-shift the product (while decrementing the
exponent) until either it becomes normalized or the exponent .
Denormal numbers present a major stumbling block to implementing floating-point
multiplication, because they require performing a variable shift in the multiplier, which
wouldnt otherwise be needed. Thus, high-performance, floating-point multipliers often
do not handle denormalized numbers, but instead trap, letting software handle them. A
few practical codes frequently underflow, even when working properly, and these
programs will run quite a bit slower on systems that require denormals to be processed by
a trap handler. So far we havent mentioned how to deal with operands of zero. This can
be handled by either testing both operands before beginning the multiplication or testing
the product afterward. If you test afterward, be sure to handle the case 0 properly:
This results in NaN, not 0. Once you detect that the result is 0, set the biased exponent to
0. Dont forget about the sign. The sign of a product is the XOR of the signs of the
operands, even when the result is 0.


Precision of Multiplication
In the discussion of integer multiplication, we mentioned that designers must decide
whether to deliver the low-order word of the product or the entire product. A similar issue
arises in floating-point multiplication, where the exact product can be rounded to the
precision of the operands or to the next higher precision. In the case of integer
multiplication, none of the standard high-level languages contains a construct that would
generate a single times single gets double instruction. The situation is different for
floating point. Many languages allow assigning the product of two single-precision
variables to a double-precision one, and the construction can also be exploited by
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 199

numerical algorithms. The best-known case is using iterative refinement to solve linear
systems of equations.

Floating-Point Addition, Division and Remainder, More on Floating-Point
Arithmetic

Floating-Point Addition
First, the sum of the binary 6-bit numbers 1.100112 and 1.100012 25: When the
summands are shifted so they have the same exponent, this is 1.10011 + .0000110001
Using a 6-bit adder (and discarding the low-order bits of the second addend) gives
1.10011 + .00001 1.10100 The first discarded bit is 1. This isnt enough to decide
whether to round up. The rest of the discarded bits, 0001, need to be examined. Or
actually, we just need to record whether any of these bits are nonzero, storing this fact in
a sticky bit just as in the multiplication algorithm. So for adding two p-bit numbers, a p-
bit adder is sufficient, as long as the first discarded bit (round) and the OR of the rest of
the bits (sticky) are kept. Then Figure H.11 can be used to determine if a roundup is
necessary, just as with multiplication. In the example above, sticky is 1, so a roundup is
necessary. The final sum is 1.101012. Heres another example: 1.11011 + .0101001 A 6-
bit adder gives 1.11011 + .01010 10.00101 Because of the carry-out on the left, the round
bit isnt the first discarded bit; rather, it is the low-order bit of the sum (1). The discarded
bits, 01, are ORed together to make sticky. Because round and sticky are both 1, the
high-order 6 bits of the sum, 10.00102, must be rounded up for the final answer of
10.00112.
Because the top bits canceled, the first discarded bit (the guard bit) is needed to fill in the
least-significant bit of the sum, which becomes 0.1111102, and the second discarded bit
becomes the round bit. This is analogous to case (1) in the multiplication algorithm (see
page H-19). The round bit of 1 isnt enough to decide whether to round up. Instead, we
need to OR all the remaining bits (0001) into a sticky bit. In this case, sticky is 1, so the
final result must be rounded up to 0.111111. This example shows that if subtraction
causes the most-significant bit to cancel, then one guard bit is needed. It is natural to ask
whether two guard bits are needed for the case when the two most-significant bits cancel.
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 200

The answer is no, because if x and y are so close that the top two bits of x y cancel,
then x y will be exact, so guard bits arent needed at all.To summarize, addition is
more complex than multiplication because,depending on the signs of the operands, it may
actually be a subtraction. If it is an addition, there can be carry-out on the left, as in the
second example. If it is subtraction, there can be cancellation, as in the third example. In
each case, the position of the round bit is different. However, we dont need to compute
the exact sum and then round. We can infer it from the sum of the high-order p bits
together with the round and sticky bits. The rest of this section is devoted to a detailed
discussion of the floating-point addition algorithm. Let a1 and a2 be the two numbers to
be added. The notations ei and si are used for the exponent and significand of the addends
ai. This means that the floating-point inputs have been unpacked and that si has an
explicit leading bit. To add a1 and a2, perform these eight steps.
1. If e1< e2, swap the operands. This ensures that the difference of the exponents satisfies
d = e1 e2 0. Tentatively set the exponent of the result to e1.
2. If the signs of a1 and a2 differ, replace s2 by its twos complement.
3. Place s2 in a p-bit register and shift it d = e1 e2 places to the right (shifting in 1s if
s2 was complemented in the previous step). From the bits shifted out, set g to the most-
significant bit, r to the next most-significant bit, and set sticky to the OR of the rest.
4. Compute a preliminary significand S = s1 + s2 by adding s1 to the p-bit register
containing s2. If the signs of a1 and a2 are different, the most-significant bit of S is 1, and
there was no carry-out, then S is negative. Replace S with its twos complement. This can
only happen when d = 0.
5. Shift S as follows. If the signs of a1 and a2 are the same and there was a carryout in
step 4, shift S right by one, filling in the high-order position with 1 (the carry-out).
Otherwise shift it left until it is normalized. When left-shifting, on the first shift fill in the
low-order position with the g bit. After that, shift in zeros. Adjust the exponent of the
result accordingly.
6. Adjust r and s. If S was shifted right in step 5, set r := low-order bit of S before shifting
and s := g OR r OR s. If there was no shift, set r := g, s := r OR s. If there was a single
left shift, dont change r and s. If there were two or more left shifts, r := 0, s := 0. (In the
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 201

last case, two or more shifts can only happen when a1 and a2 have opposite signs and the
same exponent, in which case the computation s1 + s2 in step 4 will be exact.)
7. Round S using Figure H.11; namely, if a table entry is nonempty, add 1 to the low-
order bit of S. If rounding causes carry-out, shift S right and adjust the exponent. This is
the significand of the result.
8. Compute the sign of the result. If a1 and a2 have the same sign, this is the sign of the
result. If a1 and a2 have different signs, then the sign of the result depends on which of
a1, a2 is negative, whether there was a swap in step 1, and whether S was replaced by its
twos complement in step
H.12.
must twos complement sum, giving S = 0.010.
5. Shift left twice, so S = 1.000, exp = exp 2, or exp = 2.
6. Two left shifts, so r = g = s = 0.
7. No addition required for rounding.
8. Answer is sign S 2exp or sign 1.000 2 2. Get sign from Figure H.12.
Since complement but no swap and sign(a1) is , the sign of sum is +. Thus answer =
1.0002 22.

Speeding Up Addition
Lets estimate how long it takes to perform the algorithm above. Step 2 may require an
addition, step 4 requires one or two additions, and step 7 may require an addition. If it
takes T time units to perform a p-bit add (where p = 24 for single precision, 53 for
double), then it appears the algorithm will take at least 4T time units. But that is too
pessimistic. If step 4 requires two adds, then a1 and a2 have the same exponent and
different signs. But in that case the difference is exact, and so no roundup is required in
step 7. Thus only three additions will ever occur.
Similarly, it appears that a variable shift may be required both in step 3 and step 5. But if
|e1 e2| 1, then step 3 requires a right shift of at most one place, so only step 5 needs
a variable shift. And if |e1 e2| > 1, then step 3 needs a variable shift, but step 5 will
require a left shift of at most one place. So only a single variable shift will be performed.
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 202

Still, the algorithm requires three sequential adds, which, in the case of a 53-bit double-
precision significand, can be rather time consuming.
A number of techniques can speed up addition. One is to use pipelining. The Putting It
All Together section gives examples of how some commercial chips pipeline addition.
Another method (used on the Intel 860 [Kohn and Fu 1989]) is to perform two additions
in parallel. We now explain how this reduces the latency from 3T to T. There are three
cases to consider. First, suppose that both operands have the same sign. We want to
combine the addition operations from steps 4 and 7. The position of the high-order bit of
the sum is not known ahead of time, because the addition in step 4 may or may not cause
a carry-out. Both possibilities are accounted for by having two adders. The first adder
assumes the add in step 4 will not result in a carry-out. Thus the values of r and s can be
computed before the add is actually done. If r and s indicate a roundup is necessary, the
first adder will compute S = s1 + s2 + 1, where the notation +1 means adding 1 at the
position of the least-significant bit of s1. This can be done with a regular adder by setting
the low-order carry-in bit to 1. If r and s indicate no roundup, the adder computes S = s1
+ s2 as usual. One extra detail: when r = 1, s = 0, you will also need to know the low-
order bit of the sum, which can also be computed in advance very quickly. The second
adder covers the possibility that there will be carry-out. The values of r and s and the
position where the roundup 1 is added are different from above, but again they can be
quickly computed in advance. It is not known whether there will be a carry-out until after
the add is actually done, but that doesnt matter. By doing both adds in parallel, one adder
is guaranteed to reduce the correct answer.
The next case is when a1 and a2 have opposite signs, but the same exponent. The sum a1
+ a2 is exact in this case (no roundup is necessary), but the sign isnt known until the add
is completed. So dont compute the twos complement (which requires an add) in step 2,
but instead compute s1 + s2 + 1 and s1 + s2 +1 in parallel. The first sum has the result of
simultaneously complementing s1 and computing the sum, resulting in s2 s1. The
second sum computes s1 s2. One of these will be nonnegative and hence the correct
final answer. Once again, all the additions are done in one step using two adders
operating in parallel. The last case, when a1 and a2 have opposite signs and different
exponents, is more complex. If |e1e2| > 1, the location of the le ading bit of the
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 203

difference is in one of two locations, so there are two cases just as in addition. When
|e1e2| = 1, cancellation is possible and the leading bit could be almost anywhere.
However, only if the leading bit of the difference is in the same position as the leading bit
of s1 could a roundup be necessary. So one adder assumes a roundup, the other assumes
no roundu . Thus the addition of step 4 and the rounding of step 7 can
Denormalized Numbers
Unlike multiplication, for addition very little changes in the preceding description if one
of the inputs is a denormal number. There must be a test to see if the exponent field is 0.
If it is, then when unpacking the significand there will not be a leading 1. By setting the
biased exponent to 1 when unpacking a denormal, the algorithm works unchanged. To
deal with den ormalized outputs, step 5 must be modified slightly. Shift S until it is
normalized, or until the exponent becomes Emin (that is, the biased exponent becomes 1).
If the exponent is Emin and, after rounding, the high-order bit of S is 1, then the result is
a normalized number and should be packed in the usual way, by omitting the 1. If, on the
other hand, the high-order bit is 0, the result is denormal. When the result is unpacked,
the exponent field must be set to 0. discusses the exact rules for detecting underflow.
Incidentally, detecting overflow is very easy. It can only happen if step 5
involves a shift right and the biased exponent at that point is bumped up to 255 in single
precision (or 2047 for double precision), or if this occurs after rounding. In this section,
well discuss floating-point division and remainder.
Iterative Division
We earlier discussed an algorithm for integer division. Converting it into a floating- point
division algorithm is similar to converting the integer multiplication algorithm into
floating point. The formula (s1 2e1) / (s2 2e2) = (s1 / s2) 2e1e2
operands is slightly different from integer division. Load s2 into B and s1 into P. The A
register is not needed to hold the operands. Then the integer algorithm for division (with
the one small change of skipping the very first left shift) can be used, and the result will
be of the form q0.q1To round, simply compute two additional quotient bits
(guard and round) and use the remainder as the sticky bit. The guard digit is necessary
because the first quotient bit might be 0. However, since the numerator and denominator
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 204

are both normalized, it is not possible for the two most-significant quotient bits to be 0.
This algorithm produces one quotient bit in each step.
A different approach to division converges to the quotient at a quadratic rather than a
linear rate. An actual machine that uses this algorithm will be discussed in Section H.10.
First, we will describe the two main iterative algorithms, and then we will discuss the
pros and cons of iteration when compared with the direct algorithms. There is a general
technique fo r constructing iterative algorithms, alled Newtons iteration, shown in Figure
H.13. First, cast the problem in the form of finding the zero of a function. Then, starting
from a guess for the zero, approximate the function by its tangent at that guess and form a
new guess based on where the tangent has a zero. If xi is a guess at a zero, then the
tangent line has the equation
Floating-Point Remainder
For nonnegative integers, integer division and remainder satisfya = (a DIV b)b + a REM
b, 0 a REM b b A floating-point remainder x REM y can be similarly defined as x
= INT(x/y)y + x REM y. How should x/y be converted to an integer? The IEEE remainder
function uses the round-to-even rule. That is, pick n = INT (x/y) so that x/y
n1/2. If two different n satisfy this relation, pick the even one. Then REM is
defined to be x yn. Unlike integers where 0 a REM b b, for floating-point
numbers x REM yy/2. Although this defines REM precisely, it is not a practical
operational definition, because n can be huge. In single precision, n could be as large as
2127/2126 = 2253. There is a natural way to compute REM if a direct division
algorithm is used. Proceed as if you were computing x/y. If x = s12e1 and y = s22e2 and
the divider is as in Figure H.2(b) (page H-4), then load s1 into P and s2 into B. After e1
e2 division steps, the P register will hold a number r of the form x yn satisfying _ 0
r < y. Since the IEEE remainder satisfies REMy/2, REM is equal to either r
or r y. It is only necessary to keep track of the last quotient bit produced, which is
needed to resolve halfway cases. Unfortunately, e1 e2 can be a lot of steps, and
floating-point units typically have a maximum amount of time they are allowed to spend
on one instruction. Thus, it is usually not possible to implement REM directly. None of
the chips discussed in Section H.10 implements REM, but they could by providing a
remainder-step instructionthis is what is done on the Intel 8087 family. A remainder
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 205

step takes as arguments two numbers x and y, and performs divide steps until either the
remainder is in P or n steps have been performed, where n is a small number, such as the
number of steps required for division in the highest-supported precision. Then REM can
be implemented as a software routine that calls the RE M step instruction (e1
e2)/ntimes, initially using x as the numerator, but then replacing it with the
remainder from the previous REM step. REM can be used for computing trigonometric
functions. To simplify things,
imagine that we are working in base 10 with five significant figures, and consider
computing sin x. Suppose that x = 7. Then we can reduce by = 3.1416 and compute
sin(7) = sin(7 2 3.1416) = sin(0.7168) instead. But suppose we wa nt to compute
sin(2.0 105). Then 2 105/3.1416 = 63661.8, which in our five -place system comes
out to be 63662. Since multiplying 3.1416 times 63662 gives 200000.5392, which rounds
to 2.0000 105, argument reduction reduces 2 105 to 0, which is not even close to
being correct. The problem is that our five-place system does not have the precision to do
correct argument reduction. Suppose we had the REM operator. Then we could compute
2 105 REM 3.1416 and get .53920.However, this is still not correct because we
used 3.1416, which is an approximation for . The value of 2 105 REM is
.071513.
Traditionally, there have been two approaches to computing periodic functions with
large arguments. The first is to return an error for their value when x is large. The second
is to store to a very large number of places and do exact argumentreduction. The
REM operator is not much help in either of these situations. There is a third approach that
has been used in some math libraries, such as the Berkeley UNIX 4.3bsd release. In these
libraries, is computed to the nearestfloating-point number. Lets call this machine ,
and denote it by Then when computing sin x, reduce x using x REM . As we
saw in the above example, x REM is quite differen t from x REM when x is
large, so that computing sin x as sin(x REM ) will not give the exact value of sin x.
However, computing trigonometric functions in this fashion has the property that all
familiar identities (such as sin2 x + cos2 x = 1) are true to within a few rounding errors.
Thus, using REM together with machine provides a simple method of computing
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 206

trigonometric functions that is accurate for small arguments and still may be useful for
large arguments.
When REM is used for argument reduction, it is very handy if it also returns the low-
order bits of n (where x REM y = x ny). This is because a practical implementation of
trigonometric functions will reduce by something smaller than 2. For example, it might
use /2, exploiting identities such as sin(x /2) = cos x, sin(x ) = sin x. Then
the low bits of n are needed to choose the correct identity.
Before leaving the subject of floating-point arithmetic, we present a few additional topics.
Fused Multiply-Add
Probably the most common use of floating-point units is performing matrix operations,
and the most frequent matrix operation is multiplying a matrix times a matrix (or vector),
which boils down to computing an inner product, x1 y1 + x2 y2 + . . . + xn yn.
Computing this requires a series of multiply-add combinations. Motivated by this, the
IBM RS/6000 introduced a single instruction that computes ab + c, the fused multiply-
add. Although this requires being able to read three operands in a single instruction, it has
the potential for improving the performance of computing inner products.
The fused multiply-add computes ab + c exactly and then rounds. Although rounding
only once increases the accuracy of inner products somewhat, that is not its primary
motivation. There are two main advantages of rounding once. First, as we saw in the
previous sections, rounding is expensive to implement because it may require an addition.
By rouding only once, an addition operation has been eliminated. Second, the extra
accuracy of fused multiply-add can be used to compute correctly rounded division and
square root when these are not available directly in hardware. Fused multiply-add can
also be used to implement efficient floating-point multiple-precision packages.
The implementation of correctly rounded division using fused multiply-add has many
details, but the main idea is simple. Consider again the example from Section H.6 (page
H-31), which was computing a/b with a = 13, b = 51. Then 1/b rounds to b= .020, and
abrounds to q= .26, which is not the correctly rounded quotient. Applying fused
multiply-add twice will correctly adjust the
More on Floating-Point Arithmetic
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 207

The standard specifies four precisions: single, single extended, double, and double
extended. The properties of these precisions are summarized in Figure H.7 (page H-16).
Implementations are not required to have all four precisions, but are encouraged to
support either the combination of single and single extended or all of single, double, and
double extended. Because of the widespread use of double precision in scientific
computing, double precision is almost always implemented.
Thus the computer designer usually only has to decide whether to support double
extended and, if so, how many bits it should have. The Motorola 68882 and Intel 387
coprocessors implement extended precision using the smallest allowable size of 80 bits
(64 bits of significand). However, many of the more recently designed, high-performance
floating-point chips do not implement 80-bit extended precision. One reason is that the
80-bit width of extended precision is awkward for 64-bit buses and registers. Some new
architectures, such as SPARC V8 and PA-RISC, specify a 128-bit extended (or quad)
precision.
They have established a de facto convention for quad that has 15 bits of exponent and
113 bits of significand. Although most high-level languages do not provide access to
extended precision, it is very useful to writers of mathematical software. As an example,
consider writing a library routine to compute the length of a vector (x,y) in the plane,
namely, . If x is larger than 2Emax/2, then computing this in the obvious way will
overflow. This means that either the allowable exponent range for this subroutine will be
cut in half or a more complex algorithm using scaling will have to be employed. But if
extended precision is available, then the simple algorithm will work. Computing the
length of a vector is a simple task, and it is not difficult to come up with an algorithm that
doesnt overflow. However, there are more complex problems for which extended
precision means the difference between a simple, fast algorithm and a much more
complex one. One o the best examples of this is binary-to-decimal conversion. An
efficient algorithm for binary-to-decimal conversion that makes essential use of extended
precision is very readably presented in Coonen [1984]. This algorithm is also briefly
sketched in Goldberg [1991]. Computing accurate values for transcendental functions is
another example of a problem that is made much easier if extended precision is present.

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 208

Speeding Up Integer Addition
Carry-Lookahead
An n-bit adder is just a combinational circuit. It can therefore be written by a logic
formula whose form is a sum of products and can be computed by a circuit with two
levels of logic. How do you figure out what this circuit looks like? From Equation H.2.1
(page H-3) the formula for the ith sum can be written as H.8.1 si = ai bi ci + ai bi ci + ai
bi ci + ai bi ci
where ci is both the carry-in to the ith adder and the carry-out from the (i1)-st adder.
The problem with this formula is that although we know the values of ai and bithey are
inputs to the circuitwe dont know ci. So our goal is to write ci in terms of ai and bi. To
accomplish this, we first rewrite Equation H.2.2 (page H-3)
as
H.8.2 ci = gi 1+ p i 1c i 1, g i 1= a i 1b i 1, p i 1 = a i 1 + b i 1
Here is the reason for the symbols p and g: If gi1 is true, then ci is certainly true, so a
carry is generated. Thus, g is for generate. If pi1 is true, then if ci1 is true, it is
propagated to ci. Start with Equation H.8.1 and use Equation H.8.2 to replace ci with gi
1 + pi1ci1. Then, use Equation H.8.2 with i 1 in place of i to replace ci1 with ci2,
and so on. This gives the result H.8.3 ci = g i1 + p i1 gi2 + p i1 pi2gi3 + +
p i1 pi2 p1 g0 + p i1 pi2 p1p0c0 An adder that computes carries
using Equation H.8.3 is called a carrylookahead adder, or CLA. A CLA requires one
logic level to form p and g, two levels to form the carries, and two for the sum, for a
grand total of five logic levels.
This is a vast improvement over the 2n levels required for the ripple-carry adder.
Unfortunately, as is evident from Equation H.8.3 or from Figure H.14, a carry-lookahead
adder on n bits requires a fan-in of n + 1 at the OR gate as well as at the rightmost AND
gate. Also, the pn1 signal must drive n AND gates. In addition, the rather irregular
structure and m any long wires of Figure H.14 make it impractical to build a full carry-
lookahead adder when n is large. However, we can use the carry-lookahead idea to build
an adder that has about log2n logic levels (substantially fewer than the 2n required by a
ripple-carry adder) and yet has a simple, regular structure. The idea is to build up the ps
and gs in steps. We have already seen that
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 209

c1 = g0 + c0p0
This says there is a carry-out of the 0th position (c1) either if there is a carry generated in
the 0th position or if there is a carry into the 0th position and the carry propagates.
Similarly,
c2 = G01 + P01c0
H.8 Speeding Up Integer Addition
G01 means there is a carry generated out of the block consisting of the first two bits. P01
means that a carry propagates through this block. P and G have the following logic
equations:
G01 = g1 + p1g0
P01 = p1p0
With these preliminaries out of the way, we can now show the design of a practical CLA.
The adder consists of two parts. The first part computes various values of P and G from
pi and gi, using Equations H.8.5 and H.8.6; the second part uses these P and G values to
compute all the carries via Equation H.8.4. The first part of the design is shown in Figure
H.15. At the top of the diagram, input numbers a7. . . a0 and b7. . . b0 are converted to
ps and gs using cells of type 1.
Then various Ps and Gs are generated by combining cells of type 2 in a binary tree
structure. The second part of the design is shown in Figure H.16. By feeding c0 in at the
bottom of this tree, all the carry bits come out at the top. Each cell must know a pair of
(P,G) values in order to do the conversion, and the value it needs is written inside the
cells. Now compare Figures H.15 and H.16. There is a one-to-one correspondence
between cells, and the value of (P,G) needed by the
carry-generating cells is exactly the value known by the corresponding (P,G)- generating
cells. The combined cell is shown in Figure H.17. The numbers to be added flow into the
top and downward through the tree, combining with c0 at the bottom and flowing back up
the tree to form the carries. Note that one thing is missing from Figure H.17: a small
piece of extra logic to compute c8 for the carry-out of the adder.
The bits in a CLA must pass through about log2 n logic levels, compared with 2n for a
ripple-carry adder. This is a substantial speed improvement, especially for a large n.
Whereas the ripple-carry adder had n cells, however, the CLA has.
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 210

The multiplication and division algorithms presented in Section H.2 are fairly slow,
producing 1 bit per cycle (although that cycle might be a fraction of the CPU instruction
cycle time). In this section we discuss various techniques for higher-performance
multiplication and division, including the division algorithm used in the Pentium chip.

Shifting over Zeros
Although the technique of shifting over zeros is not currently used much, it is instructive
to consider. It is distinguished by the fact that its execution time is operand dependent. Its
lack of use is primarily attributable to its failure to offer enough speedup over bit-at-a-
time algorithms. In addition, pipelining, synchronization with the CPU, and good
compiler optimization are difficult with algorithms that run in variable time. In
multiplication, the idea behind shifting over zeros is to add logic that detects when the
low-order bit of the A register is 0 (seeFigure H.2(a) on page H-4) and, if so, skips the
addition step and proceedsdirectly to the shift stephence the term shifting over zeros.
What about shifting for division? In nonrestoring division, an ALU operation (either an
addition or subtraction) is performed at every step. There appears to be no opportunity for
kipping an operation. But think about division this way: To compute a/b, subtract
multiples of b rom a, and then report how many subtractions were done. At each stage of
the subtraction rocess the remainder must fit into the P register. In the case when the
remainder is a small positive number, you normally subtract b; but suppose instead you
only shifted the remainder and subtracted b the next time. As long as the remainder was
sufficiently small (its high-order bit 0), after shifting it still would fit into the P register,
and no information would be lost. However, this method does require changing the way
we keep track of the number of times b has been subtracted from a. This idea usually
goes under the name of SRT division, for Sweeney Robertson, and Tocher, who
independently proposed algorithms of this nature. The main extra complication of SRT
division is that the quotient bits cannot be determined immediately from the sign of P at
each step, as they can be in ordinary nonrestoring division.
More precisely, to divide a by b where a and b are n-bit numbers, load a and b into the A
and B registers, respectively, of Figure H.2
SRT Division
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 211

1. If B has k leading zeros when expressed using n bits, shift all the registers left k bits.
2. For i = 0, n 1,
(a) If the top three bits of P are equal, set qi = 0 and shift (P,A) one bit left.
(b) If the top three bits of P are not all equal and P is negative, set qi = 1 (also written as
1), shift (P,A) one bit left, and add B.
(c) Otherwise set qi = 1, shift (P,A) one bit left, and subtract B.
End loop
3. If the final remainder is negative, correct the remainder by adding B, and correct the
quotient by subtracting 1 from q0. Finally, the remainder must be shifted k bits right,
where k is the initial shift.
A numerical example is given in Figure H.23. Although we are discuss in integer
division, it helps in explaining the algorithm to imagine the binary point just left of the
most-significant bit. This changes Figure H.23 from 010002/00112 to 0.10002/.00112.
Since the binary point is changed in both the numerator and denominator, the quotient is
not affected. The (P,A) register pair holds the remainder and is a twos complement
number. For example, if P contains 111102 and A = 0, then the remainder is 1.11102 =
1/8. If r is the value of there mainder, then 1 r < 1. Given these preliminaries, we can
now analyze the SRT division algorithm. The first step of the algorithm shifts b so that b
1/2. The rule for which ALU operation to perform is this: If 1/4 r < 1/4 (true
whenever the top three bits of P are equal), then compute 2r by shifting (P,A) left one bit;
else if r < 0 (and hence r < 1/4, since otherwise it would have been eliminated by the
first condition), then compute 2r + b by shifting and then adding, else r 1/4 and subtract
b from 2r. Using b 1/2, it is easy to check that these rules keep 1/2 r < 1/2. For no
restoring division, we only have r b, and we need P to be n + 1 bits wide. But for
SRT division, the bound on r is tighter, namely, 1/2 r < 1/2. Thus, we can save a bit
by eliminating the high-order bit of P (and b and the adder). In particular, the test for
equality of the top three bits of P becomes a test on ust two bits.The algorithm might
change slightly in an implementation of SRT division. After each ALU operation, the P
register can be shifted as many places as necessary to make either r 1/4 or r < 1/4. By
shifting k places, k quotient bits are set equal to zero all at once. For this reason SRT
ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 212

division is sometimes described as one that keeps the remainder normalized to r
1/4.

Notice that the value of the quotient bit computed in a given step is based on which
operation is performed in that step (which in turn depends on the result of the operation
from the previous step). This is in contrast to nonrestoring division, where the quotient bit
computed in the ith step depends on the result of the operation in the same step. This
difference is reflected in the fact that when the final remainder is negative, the last
quotient bit must be adjusted in SRT division,
but not in nonrestoring division. However, the key fact about the quotient bits in SRT
division is that they can include 1. Although Figure H.23 shows the quotient bits being
stored in the low-order bits of A, an actual implementation cant do this because you
cant fit the three values 1, 0, 1 into one bit. Furthermore, the quotient must be
converted to ordinary twos complement in a full adder. A common way to do this is to
accumulate the positive quotient bits in one register and the negative quotient bits in
another, and then subtract the two registers after all the bits are known. Because there is
more than one way to write a number in terms of the digits 1, 0, 1, SRT division is said
to use a redundant quotient representation.

The differences between SRT division and ordinary nonrestoring division can be
summarized as follows:
1. ALU decision rule: In nonrestoring division, it is determined by the sign of P; in SRT,
it is determined by the two most-significant bits of P.
2. Final quotient: In nonrestoring division, it is immediate from the successive signs of P;
in SRT, there are three quotient digits (1, 0, 1), and the final quotient must be computed
in a full n-bit adder.
3. Speed: SRT division will be faster on operands that produce zero quotient bits. The
simple version of the SRT division algorithm given above does not offer enough of a
speedup to be practical in most cases. However, later on in this section we will study
variants of SRT division that are quite practical.

ADVANCED COMPUTER ARCHITECTURE 12SCS23

Dept of CS&E, SJBIT Page 213

Fallacies and Pitfalls
Fallacy Underflows rarely occur in actual floating-point application code.
Although most codes rarely underflow, there are actual codes that underflow frequently.
SDRWAVE [Kahaner 1988], which solves a one-dimensional wave equation, is one such
example. This program underflows quite frequently, even when functioning properly.
Measurements on one machine show that adding hardware support for gradual underflow
would cause SDRWAVE to run about 50% faster.
Fallacy Conversions between integer and floating point are rare.
In fact, in spice they are as frequent as divides. The assumption that conversions are rare
leads to a mistake in the SPARC version 8 instruction set, which does not provide an
instruction to move from integer registers to floating-point registers.
Pitfall Dont increase the speed of a floating-point unit without increasing its memory
bandwidth.
A typical use of a floating-point unit is to add two vectors to produce a third vector. If
these vectors consist of double-precision numbers, then each floating-point add will use
three operands of 64 bits each, or 24 bytes of memory. The memory bandwidth
requirements are even greater if the floating-point unit can perform addition and
multiplication in parallel (as most do).
Pitfall x is not the same as 0 x.
This is a fine point in the IEEE standard that has tripped up some designers. Because
floating-point numbers use the sign magnitude system, there are two zeros, +0 and 0.
The standard says that 0 0 = +0, whereas (0) = 0. Thus x
is not the same as 0 x when x = 0.