Pitfalls of Object Oriented Programming

(1)

Tony Albrecht – Technical Consultant Developer Services

Sony Computer Entertainment Europe Research & Development Division

Pitfalls of Object Oriented Programming

(2)

Slide 2

• A quick look at Object Oriented (OO) programming

• A common example

• Optimisation of that example

• Summary

What I will be covering

(3)

Slide 3

• What is OO programming?

– a programming paradigm that uses "objects" – data structures

consisting of datafields and methods together with their interactions – to design applications and computer programs.

(Wikipedia)

• Includes features such as

– Data abstraction – Encapsulation – Polymorphism – Inheritance

Object Oriented (OO) Programming

(4)

Slide 4

• OO programming allows you to think about problems in terms of objects and their

interactions.

• Each object is (ideally) self contained

– Contains its own code and data.

– Defines an interface to its code and data.

• Each object can be perceived as a „black box‟.

What’s OOP for?

(5)

Slide 5

• If objects are self contained then they can be

– Reused.

– Maintained without side effects.

– Used without understanding internal implementation/representation.

• This is good, yes?

Objects

(6)

Slide 6

• Well, yes

• And no.

• First some history.

Are Objects Good?

(7)

Slide 7

A Brief History of C++

C++ development started

1979 2009

(8)

Slide 8

A Brief History of C++

Named “C++”

1979 1983 2009

(9)

Slide 9

A Brief History of C++

First Commercial release

1979 1985 2009

(10)

Slide 10

A Brief History of C++

Release of v2.0

1979 1989 2009

(11)

Slide 11

A Brief History of C++

Release of v2.0

1979 1989 2009

Added

• multiple inheritance,

• abstract classes,

• static member functions,

• const member functions

• protected members.

(12)

Slide 12

A Brief History of C++

Standardised

1979 1998 2009

(13)

Slide 13

A Brief History of C++

Updated

1979 2003 2009

(14)

Slide 14

A Brief History of C++

C++0x

1979 2009 ?

(15)

Slide 15

• Many more features have been added to C++

• CPUs have become much faster.

• Transition to multiple cores

• Memory has become faster.

So what has changed since 1979?

http://www.vintagecomputing.com

(16)

Slide 16

CPU performance

Computer architecture: a quantitative approach

By John L. Hennessy, David A. Patterson, Andrea C. Arpaci-Dusseau

(17)

Slide 17

CPU/Memory performance

Computer architecture: a quantitative approach

By John L. Hennessy, David A. Patterson, Andrea C. Arpaci-Dusseau

(18)

Slide 18

• One of the biggest changes is that memory access speeds are far slower (relatively)

– 1980: RAM latency ~ 1 cycle

– 2009: RAM latency ~ 400+ cycles

• What can you do in 400 cycles?

What has changed since 1979?

(19)

Slide 19

• OO classes encapsulate code and data.

• So, an instantiated object will generally contain all data associated with it.

What has this to do with OO?

(20)

Slide 20

• With modern HW (particularly consoles), excessive encapsulation is BAD.

• Data flow should be fundamental to your design (Data Oriented Design)

My Claim

(21)

Slide 21

• Base Object class

– Contains general data

• Node

– Container class

• Modifier

– Updates transforms

• Drawable/Cube

– Renders objects

Consider a simple OO Scene Tree

(22)

Slide 22

• Each object

– Maintains bounding sphere for culling – Has transform (local and world)

– Dirty flag (optimisation) – Pointer to Parent

Object

(23)

Slide 23

Objects

Class Definition Memory Layout

Each square is 4 bytes

(24)

Slide 24

• Each Node is an object, plus

– Has a container of other objects – Has a visibility flag.

Nodes

(25)

Slide 25

Nodes

Class Definition Memory Layout

(26)

Slide 26

Consider the following code…

• Update the world transform and world

space bounding sphere for each object.

(27)

Slide 27

Consider the following code…

• Leaf nodes (objects) return transformed

bounding spheres

(28)

Slide 28

Consider the following code…

• Leaf nodes (objects) return transformed bounding spheres

What‟s wrong with this code?

(29)

Slide 29

Consider the following code…

• Leaf nodes (objects) return transformed bounding spheres

If m_Dirty=false then we get branch misprediction which costs 23 or 24 cycles.

(30)

Slide 30

Consider the following code…

• Leaf nodes (objects) return transformed bounding spheres

Calculation of the world bounding sphere

takes only 12 cycles.

(31)

Slide 31

Consider the following code…

• Leaf nodes (objects) return transformed bounding spheres

So using a dirty flag here is actually

slower than not using one (in the case where it is false)

(32)

Slide 32

Lets illustrate cache usage

Main Memory Each cache line is L2 Cache 128 bytes

(33)

Slide 33

Cache usage

Main Memory

parentTransform is already in the cache (somewhere)

L2 Cache

(34)

Slide 34

Cache usage

Main Memory

Assume this is a 128byte

boundary (start of cacheline) L2 Cache

(35)

Slide 35

Cache usage

Main Memory

Load m_Transform into cache

L2 Cache

(36)

Slide 36

Cache usage

Main Memory

m_WorldTransform is stored via cache (write-back)

L2 Cache

(37)

Slide 37

Cache usage

Main Memory

Next it loads m_Objects L2 Cache

(38)

Slide 38

Cache usage

Main Memory

Then a pointer is pulled from somewhere else (Memory

managed by std::vector)

L2 Cache

(39)

Slide 39

Cache usage

Main Memory

vtbl ptr loaded into Cache

L2 Cache

(40)

Slide 40

Cache usage

Main Memory

Look up virtual function

L2 Cache

(41)

Slide 41

Cache usage

Main Memory

Then branch to that code (load in instructions)

L2 Cache

(42)

Slide 42

Cache usage

Main Memory

New code checks dirty flag then sets world bounding sphere

L2 Cache

(43)

Slide 43

Cache usage

Main Memory

Node‟s World Bounding Sphere is then Expanded

L2 Cache

(44)

Slide 44

Cache usage

Main Memory

Then the next Object is processed

L2 Cache

(45)

Slide 45

Cache usage

Main Memory

First object costs at least 7 cache misses

L2 Cache

(46)

Slide 46

Cache usage

Main Memory

Subsequent objects cost at least 2 cache misses each

L2 Cache

(47)

Slide 47

• 11,111 nodes/objects in a tree 5 levels deep

• Every node being transformed

• Hierarchical culling of tree

• Render method is empty

The Test

(48)

Slide 48

Performance

This is the time taken just to

traverse the tree!

(49)

Slide 49

Why is it so slow?

~22ms

(50)

Slide 50

Look at GetWorldBoundingSphere()

(51)

Slide 51

Samples can be a little misleading at the source code level

(52)

Slide 52

if(!m_Dirty) comparison

(53)

Slide 53

Stalls due to the load 2 instructions earlier

(54)

Slide 54

Similarly with the matrix multiply

(55)

Slide 55

Some rough calculations

Branch Mispredictions: 50,421 @ 23 cycles each ~= 0.36ms

(56)

Slide 56

Some rough calculations

Branch Mispredictions: 50,421 @ 23 cycles each ~= 0.36ms L2 Cache misses: 36,345 @ 400 cycles each ~= 4.54ms

(57)

Slide 57

• From Tuner, ~ 3 L2 cache misses per object

– These cache misses are mostly sequential (more than 1 fetch from main memory can happen at once)

– Code/cache miss/code/cache miss/code…

(58)

Slide 58

• How can we fix it?

• And still keep the same functionality and interface?

Slow memory is the problem here

(59)

Slide 59

• Use homogenous, sequential sets of data

The first step

(60)

Slide 60

Homogeneous Sequential Data

(61)

Slide 61

• Use custom allocators

– Minimal impact on existing code

• Allocate contiguous

– Nodes – Matrices

– Bounding spheres

Generating Contiguous Data

(62)

Slide 62

Performance

19.6ms -> 12.9ms 35% faster just by

moving things around in memory!

(63)

Slide 63

• Process data in order

• Use implicit structure for hierarchy

– Minimise to and fro from nodes.

• Group logic to optimally use what is already in cache.

• Remove regularly called virtuals.

What next?

(64)

Slide 64

Hierarchy

Node We start with

a parent Node

(65)

Slide 65

Hierarchy

Node

Node Which has children Node nodes

(66)

Slide 66

Hierarchy

Node

Node Node

And they have a parent

(67)

Slide 67

Hierarchy

Node

Node Node

Node Node Node Node

And they have children

(68)

Slide 68

Hierarchy

Node

Node Node

Node Node Node Node

And they all have parents

(69)

Slide 69

Hierarchy

Node

Node Node

Node Node Node Node

A lot of this information can be

inferred

Node Node Node Node Node Node Node Node

(70)

Slide 70

Hierarchy

Node

Node Node

Use a set of arrays, one per hierarchy

level

(71)

Slide 71

Hierarchy

Node

Node Node

Parent has 2 children

:2

:4 :4

Children have 4 children

(72)

Slide 72

Hierarchy

Node

Node Node

Ensure nodes and their data are contiguous in

memory

:2

:4 :4

(73)

Slide 73

• Make the processing global rather than local

– Pull the updates out of the objects.

• No more virtuals

– Easier to understand too – all code in one

place.

(74)

Slide 74

• OO version

– Update transform top down and expand WBS bottom up

Need to change some things…

(75)

Slide 75

Node

Node Node

Node Node Node Node

Update transform

(76)

Slide 76

Node

Node Node

Node Node Node Node

Update transform

(77)

Slide 77

Node

Node Node

Node Node Node Node

Update transform and world bounding sphere

(78)

Slide 78

Node

Node Node

Node Node Node Node

Add bounding sphere of child

(79)

Slide 79

Node

Node Node

Node Node Node Node

(80)

Slide 80

Node

Node Node

Node Node Node Node

(81)

Slide 81

Node

Node Node

Node Node Node Node

(82)

Slide 82

Node

Node Node

Node Node Node Node

(83)

Slide 83

• Hierarchical bounding spheres pass info up

• Transforms cascade down

• Data use and code is „striped‟.

– Processing is alternating

(84)

Slide 84

• To do this with a „flat‟ hierarchy, break it into 2 passes

– Update the transforms and bounding spheres(from top down)

– Expand bounding spheres (bottom up)

Conversion to linear

(85)

Slide 85

Transform and BS updates

Node

Node Node

For each node at each level (top down) {

multiply world transform by parent‟s transform wbs by world transform }

:2

:4 :4

(86)

Slide 86

Update bounding sphere hierarchies

Node

Node Node

For each node at each level (bottom up) {

add wbs to parent‟s }

:2

:4 :4

For each node at each level (bottom up) {

add wbs to parent‟s

cull wbs against frustum }

(87)

Slide 87

Update Transform and Bounding Sphere

How many children nodes to process

(88)

Slide 88

Update Transform and Bounding Sphere

For each child, update transform and bounding sphere

(89)

Slide 89

Update Transform and Bounding Sphere

Note the contiguous arrays

(90)

Slide 90

So, what’s happening in the cache?

Parent Children

Parent Data

Childrens‟ Data Children‟s node data not needed

Unified L2 Cache

(91)

Slide 91

Load parent and its transform

Parent

Parent Data

Childrens‟ Data

Unified L2 Cache

(92)

Slide 92

Load child transform and set world transform

Parent

Parent Data

Childrens‟ Data

Unified L2 Cache

(93)

Slide 93

Load child BS and set WBS

Parent

Parent Data

Childrens‟ Data

Unified L2 Cache

(94)

Slide 94

Load child BS and set WBS

Parent

Parent Data

Childrens‟ Data

Next child is calculated with

no extra cache misses ! Unified L2 Cache

(95)

Slide 95

Load child BS and set WBS

Parent

Parent Data

Childrens‟ Data

The next 2 children incur 2

cache misses in total Unified L2 Cache

(96)

Slide 96

Prefetching

Parent

Parent Data

Childrens‟ Data

Because all data is linear, we can predict what memory will be needed in ~400 cycles and prefetch

Unified L2 Cache

(97)

Slide 97

• Tuner scans show about 1.7 cache misses per node.

• But, these misses are much more frequent

– Code/cache miss/cache miss/code

– Less stalling

(98)

Slide 98

Performance

19.6 -> 12.9 -> 4.8ms

(99)

Slide 99

• Data accesses are now predictable

• Can use prefetch (dcbt) to warm the cache

– Data streams can be tricky

– Many reasons for stream termination – Easier to just use dcbt blindly

• (look ahead x number of iterations)

Prefetching

(100)

Slide 100

• Prefetch a predetermined number of iterations ahead

• Ignore incorrect prefetches

Prefetching example

(101)

Slide 101

Performance

19.6 -> 12.9 -> 4.8 -> 3.3ms

(102)

Slide 102

• This example makes very heavy use of the cache

• This can affect other threads‟ use of the cache

– Multiple threads with heavy cache use may thrash the cache

A Warning on Prefetching

(103)

Slide 103

The old scan

~22ms

(104)

Slide 104

The new scan

~16.6ms

(105)

Slide 105

Up close

~16.6ms

(106)

Slide 106

Looking at the code (samples)

(107)

Slide 107

Performance counters

Branch mispredictions: 2,867 (cf. 47,000) L2 cache misses: 16,064 (cf 36,000)

(108)

Slide 108

• Just reorganising data locations was a win

• Data + code reorganisation= dramatic improvement.

• + prefetching equals even more WIN.

In Summary

(109)

Slide 109

• Be careful not to design yourself into a corner

• Consider data in your design

– Can you decouple data from objects?

– …code from objects?

• Be aware of what the compiler and HW are doing

OO is not necessarily EVIL

(110)

Slide 110

• Optimise for data first, then code.

– Memory access is probably going to be your biggest bottleneck

• Simplify systems

– KISS

– Easier to optimise, easier to parallelise

Its all about the memory

(111)

Slide 111

• Keep code and data homogenous

– Avoid introducing variations

– Don‟t test for exceptions – sort by them.

• Not everything needs to be an object

– If you must have a pattern, then consider using Managers

Homogeneity

(112)

Slide 112

• You are writing a GAME

– You have control over the input data

– Don‟t be afraid to preformat it – drastically if need be.

• Design for specifics, not generics (generally).

Remember

(113)

Slide 113

• Better performance

• Better realisation of code optimisations

• Often simpler code

• More parallelisable code

Data Oriented Design Delivers

(114)

Slide 114