Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
Floating Point 101
Search
kida
February 06, 2013
Programming
7
310
Floating Point 101
A very very basic introduction to FP.
With some inaccuracies.
kida
February 06, 2013
Tweet
Share
More Decks by kida
See All by kida
Cognitive Supervision for Laser Phonomicrosurgery
kida
0
43
Towards Cognitive Supervision in robot-assisted surgery
kida
0
190
Other Decks in Programming
See All in Programming
Rancher と Terraform
fufuhu
2
240
基礎から学ぶ大画面対応(Learning Large-Screen Support from the Ground Up)
tomoya0x00
0
410
RDoc meets YARD
okuramasafumi
4
170
旅行プランAIエージェント開発の裏側
ippo012
2
890
Putting The Genie in the Bottle - A Crash Course on running LLMs on Android
iurysza
0
140
AI時代のUIはどこへ行く?
yusukebe
17
8.7k
プロポーザル駆動学習 / Proposal-Driven Learning
mackey0225
2
1.2k
rage against annotate_predecessor
junk0612
0
160
アセットのコンパイルについて
ojun9
0
120
AIでLINEスタンプを作ってみた
eycjur
1
230
Azure SRE Agentで運用は楽になるのか?
kkamegawa
0
2k
Go言語での実装を通して学ぶLLMファインチューニングの仕組み / fukuokago22-llm-peft
monochromegane
0
120
Featured
See All Featured
The Success of Rails: Ensuring Growth for the Next 100 Years
eileencodes
46
7.6k
GitHub's CSS Performance
jonrohan
1032
460k
Automating Front-end Workflow
addyosmani
1370
200k
Unsuck your backbone
ammeep
671
58k
4 Signs Your Business is Dying
shpigford
184
22k
Dealing with People You Can't Stand - Big Design 2015
cassininazir
367
27k
実際に使うSQLの書き方 徹底解説 / pgcon21j-tutorial
soudai
PRO
188
55k
Building an army of robots
kneath
306
46k
The World Runs on Bad Software
bkeepers
PRO
70
11k
Templates, Plugins, & Blocks: Oh My! Creating the theme that thinks of everything
marktimemedia
31
2.5k
Practical Orchestrator
shlominoach
190
11k
Facilitating Awesome Meetings
lara
55
6.5k
Transcript
FLOATING 101 POINT
FLOATING 100.999998 POINT
engineers we are
researchers we are
3.14159265358979 3238462643383279 5028841971693993 7510582097494459 2307816406286208 NUMBERS WE PLAY WITH ALL
DAY LONG
well, sometimes even at night. (yawn).
So, what is a floating point?
A floating point is ± D 1 .D 2 D
3 ···D n x Be
A floating point is sign ± D 1 .D 2
D 3 ···D n x Be
A floating point is significand ± D 1 .D 2
D 3 ···D n x Be
A floating point is base ± D 1 .D 2
D 3 ···D n x Be
A floating point is exponent ± D 1 .D 2
D 3 ···D n x Be
A floating point represents ± (D 1 + D 2
* B-1 + D 3 * B-2 + … + D n * B(n-1)) * Be
For example + 3.14 x 100 = (3 + 1*0.1
+ 4*0.01)*1 = 3.14
The point can float ! + 3.14 x 10-1 =
0.314
The point can float ! + 3.14 x 10+1 =
31.4
What if B = 2 ? + 1.00 x 2+2
= 4.0
Like machines do. http://grouper.ieee.org/groups/754/
Normalization of floating point
Multiple representations + 0.01 x 22 = 1.0 + 0.10
x 21 = 1.0 + 1.00 x 20 = 1.0
Normalized representation + 0.01 x 22 = 1.0 + 0.10
x 21 = 1.0 + 1.00 x 20 = 1.0
Normalized representation + (1.)000 x 20 1 is omitted
Normalized representation + (1.)000 x 20 there's room for an
extra digit!
Excess-127 representation -127 → 0 -126 → +1 … -1
→ +126 0 → +127
#include <float.h> FLT_MIN, FLT_MAX, ... #include <math.h> M_PI, M_E, NAN,
INFINITY, ...
Why no exact representation for 0.1?
FLOATING POINT REAL NUMBERS is used to represent
FLOATING POINT RATIONAL NUMBERS denotes a (finite) subset of
0.1 cannot be expressed as a power of 2 +
??? x 2??
+ 00 x 20 1 It's also a matter of
precision
+ 01 x 20 1 1.25 It's also a matter
of precision
+ 10 x 20 1 1.25 1.5 It's also a
matter of precision
+ 11 x 20 1 1.25 1.5 1.75 It's also
a matter of precision
+ 11 x 20 π/2 It's also a matter of
precision
+ 11 x 20 π/2 It's also a matter of
precision
+ 00 x 21 1 1.25 1.5 1.75 2.0 Not
just a matter of precision or basis...
+ 01 x 21 1 1.25 1.5 1.75 2.0 2.5
Not just a matter of precision or basis...
+ 10 x 21 1 1.25 1.5 1.75 2.0 2.5
3.0 Not just a matter of precision or basis...
Like death and taxes rounding errors are a fact of
life. http://wiki.octave.org/FAQ
+ 110 x 21 Operands that differ greatly + 100
x 2-2
+ 110000 x 21 Operands that differ greatly + 000101
x 21
+ 110000 x 21 Operands that differ greatly + 000101
x 21 = 110
None
Operands that are really close + 111 x 21 -
110 x 21 = 001 x 21
Operands that are really close + 111 x 21 -
110 x 21 = 100 x 2-2
None
Fixed point representation + 100.001010 = 22 + 2-3+ 2-5
= 4.15625
POINT WHAT'S THE WITH FLOATING
FP ARITHMETIC IS FAST Embedded in HW.
Single precision up to ~10+38. FP REPRESENTS A WIDE RANGE
HE APPROVES FP
Anyway, errors still there.
Okay, what about increasing the number of digits use decimal
representations estimating errors think before you type
More digits, please! double (52 significant bits) long double (112
significant bits) arbitrary precision * * language support needed
Use decimal representations! decimal (C# only) BigDecimal (Java only) std::decimal
(C++, coming soon)* * after IEEE-754 2008
Estimate the error of your algo rel_err = fabs(f –
fp) / f
Use float to represent time float time; while (true) time
+= 0.20;
Use float to represent time float time; while (true) time
+= 0.20; This is BAD. And you should feel BAD.
Compare float numbers (a == b)
Compare float numbers (a == b) fabs(a -b) <= FLT_EPSILON
Compare float numbers (a == b) fabs(a -b) <= FLT_EPSILON
fabs(a - b) <= max(fabs(a),fabs(b)) * pc
There is no silver bullet.
Use libraries (when available).
Vector addition (naive) float t[SIZE]; float result; for (i =
0; i < SIZE; ++i) result += t[i];
RESCUE GNU GSL TO THE
None
that's all folks! @lorisfichera – https://kid-a.github.com References and source code
available at https://github.com/kid-a/floating-point-seminar Credits Font: Yanone Kaffeesatz (http://www.yanone.de/typedesign/kaffeesatz/)