1 \documentclass[8pt]{extreport}
4 \usepackage{amsmath} % needed for math align
5 \usepackage{bm} % needed for maths bold face
6 \usepackage{graphicx} % needed for including graphics e.g. EPS, PS
7 \usepackage{fancyhdr} % needed for header
8 %\usepackage{epstopdf} % Converts eps to pdf before including. Do it manually instead.
12 \topmargin -1.5cm % read Lamport p.163
13 \oddsidemargin -0.04cm % read Lamport p.163
14 \evensidemargin -0.04cm % same as oddsidemargin but for left-hand pages
17 %\pagestyle{empty} % Uncomment if don't want page numbers
18 \parskip 8.2pt % sets spacing between paragraphs
19 %\renewcommand{\baselinestretch}{1.5} % Uncomment for 1.5 spacing between lines
20 \parindent 0pt % sets leading space for paragraphs
24 \newcommand{\vect}[1]{\boldsymbol{#1}} % Draw a vector
25 \newcommand{\divg}[1]{\nabla \cdot #1} % divergence
26 \newcommand{\curl}[1]{\nabla \times #1} % curl
27 \newcommand{\grad}[1]{\nabla #1} %gradient
28 \newcommand{\pd}[3][ ]{\frac{\partial^{#1} #2}{\partial #3^{#1}}} %partial derivative
29 %\newcommand{\d}[3][ ]{\frac{d^{#1} #2}{d #3^{#1}}} %full derivative
30 \newcommand{\phasor}[1]{\tilde{#1}} % make a phasor
31 \newcommand{\laplacian}[1]{\nabla^2 {#1}} % The laplacian operator
36 \definecolor{darkgray}{rgb}{0.95,0.95,0.95}
37 \definecolor{darkred}{rgb}{0.75,0,0}
38 \definecolor{darkblue}{rgb}{0,0,0.75}
39 \definecolor{pink}{rgb}{1,0.5,0.5}
41 \lstset{backgroundcolor=\color{darkgray}}
42 %\lstset{numbers=left, numberstyle=\tiny, stepnumber=1, numbersep=5pt}
43 %\lstset{keywordstyle=\color{darkred}\bfseries}
44 %\lstset{commentstyle=\color{darkblue}}
45 %\lstset{stringsyle=\color{red}}
46 %\lstset{showstringspaces=false}
47 %\lstset{basicstyle=\small}
49 \newcommand{\shell}[1]{\texttt{#1}}
50 \newcommand{\code}[1]{\texttt{#1}}
60 \fancyfoot[CO, C]{\thepage}
63 B.Eng. Final Year Project \par
64 {\bf \Large Literature Notes} \par
65 Sam Moore, David Gow \\
66 Faculty of Engineering, Computing and Mathematics, University of Western Australia \\
72 \chapter{Literature Summaries}
74 \section{Postscript Language Reference Manual \cite{plrm}}
76 Adobe's official reference manual for PostScript.
81 \item First version was published BEFORE the IEEE standard and used smaller floats than binary32
82 \item Now uses binary32 floats.
85 \section{Portable Document Format Reference Manual \cite{pdfref17}}
87 Adobe's official reference for PDF.
92 \item Early versions did not use IEEE binary32 but 16-16 exponent/mantissa encodings (Why?)
93 \item Current standard is restricted to binary32
94 \item It specifically says PDF creators must use at most binary32 because higher precision is not supported by Adobe Reader.
97 \section{IEEE Standard for Floating-Point Arithmetic \cite{ieee2008-754}}
99 The IEEE (revised) 754 standard.
105 \item Has been adopted by CPUs
106 \item Standardised floats for programmers --- accomplishes goal of allowing non-numerical experts to write reasonably sophisticated platform independent programs that may perform complex numerical operations
111 \item Adoption by GPUs slower\cite{hillesland2004paranoia}
112 \item It specifies the maximum errors for operations using IEEE types but nothing about internal representations
113 \item Many hardware devices (GPUs and CPUs) use non-IEEE representations internally and simply truncate/round the result
115 \item This isn't so much of a problem when the device uses additional bits but it is misleading when GPUs use less than binary32 and act as if they are using binary32 from the programmer's perspective.
116 \item Devices using {\bf less} bits internally aren't IEEE compliant
118 \item Thus the same program compiled and run on different architectures may give completely different results\cite{HFP}
120 \item The ultimate goal of allowing people to write numerical programs in total ignorance of the hardware is not entirely realised
122 \item This is the sort of thing that makes people want to use a virtual machine, and thus Java
124 \item Objectively I probably shouldn't say that using Java is in itself a failure
126 \item Standards such as PostScript and PDF were slow to adopt IEEE representations
127 \item The OpenVG standard accepts IEEE binary32 in the API but specifically states that hardware may use less than this\cite{rice2008openvg}
134 \section{Portable Document Format (PDF) --- Finally... \cite{cheng2002portable}}
136 This is not spectacularly useful, is basically an advertisement for Adobe software.
140 \item Visual communications has been revolutionised by computing
141 \item BUT there have always been problems in exchanging formats
142 \item Filetypes like text, rich text, IGES, DXF, TIFF, JPEG, GIFF solve problems for particular types of files only
143 \item PDF solves everything for everyone; can include text, images, animation, sound, etc
147 \item Raster Image Process (RIP) --- For printing (presumably also displaying on screen)
148 \item Originally needed to convert to PS then RIP, with PS 3 can now RIP directly.
149 \item Reduced filesize due to compression
150 \item Four major applications - Stoy 1999\cite{stoy1999}
152 \item Download files from internet
154 \item Files for outputing to printers
155 \item Conventional [commercial scale?] printing
157 \item List of various (Adobe) PDF related software
159 \item Includes software for PS that converts to/from PDF
160 \item So PS was obviously pretty popular before PDF
162 \item Can Optimize for screen/printer [not clear how]
163 \item Can compress for size
167 \section{Pixels or Perish \cite{hayes2012pixels}}
169 ``The art of scientific illustration will have to adapt to the new age of online publishing''
170 And therefore, JavaScript libraries ($\text{D}^3$) are the future.
172 The point is that we need to change from thinking about documents as paper to thinking of them as pixels.
173 This kind of makes it related to our paper, because it is the same way we are justifying our project. It does mention precision, but doesn't say we need to get more of it.
175 I get the feeling from this that Web based documents are a whole bunch of completely different design philosophies hacked together
178 This paper uses Metaphors a lot. I never met a phor that didn't over extend itself.
183 \item Drawings/Pictures are ornaments in science but they are not just ornamental
184 \item Processes have changed a lot; eg: photographic plates $\to$ digital images
185 \item ``we are about to turn the page --- if not close the book --- on yet another chapter in publishing history.'' (HO HO HO)
186 \item It would be cool to have animated figures in documents (eg: Population pyramid; changes with time); not just as ``supplements''
187 \item In the beginning, there was PostScript, 1970s and 1980s, John Warnock and Charles Geschke, Adobe Systems
188 \item PS is a language for vector graphics; objects are constructed from geometric primitives rather than a discrete array of pixels
189 \item PS is a complete programming language; an image is also a program; can exploit this to control how images are created based on data (eg: Faces)
190 \item PDF is ``flattened'' PS. No longer programable. Aspires to be ``virtual paper''.
191 \item But why are we using such powerful computing machines just to emulate sheets paper? (the author asks)
194 {\bf Web based Documents}
196 \item HTML, CSS, JavaScript - The Axis of Web Documents
198 \item HTML - Defines document structure
199 \item CSS - Defines presentation of elements in document
200 \item JavaScript - Encodes actions, allows dynamic content (change the HTML/CSS)
202 \item \texttt{<canvas>} will let you draw anything (So in principle don't even need all of HTML/CSS)
204 \item Not device independent
205 \item ``Coordinates can be specified with precision finer than pixel resolution'' {\bf (TODO: Investigate this?)}
206 \item JavaScript operators to draw things on canvas are very similar to the PostScript model
208 \item SVG --- Same structure (Document Object Model (DOM)) as HTML
210 \item ``Noun language''
211 \item Nouns define lines/curves etc, rather than paragraphs/lists
212 \item Also borrows things from PostScript (eg: line caps and joints)
213 \item IS device independent, ``very high precision'' {\bf (TODO: Investigate)}
214 \item JavaScript can be used to interact with SVG too
216 \item $\text{D}^{3}$ (Data Driven Documents) - A JavaScript library
218 \item Idea is to create or modify elements of a DOM document using supplied data
219 \item \url{https://github.com/mbostock/d3/wiki}
221 \item We are in a new Golden Age of data visualisation
222 \item Why do we still use PDFs?
224 \item PDFs are ``owned'' by the author/reader; you download it, store it, you can print it, etc
225 \item HTML documents are normally on websites. They are not self contained. They often rely on remote content from other websites (annoying to download the whole document).
227 \item {\bf Conclusion} Someone should open up PDF to accept things like $\text{D}^{3}$ and other graphics formats (links nicely with \cite{barnes2013embedding})
228 \item Also, Harry Potter reference
232 \section{Embedding and Publishing Interactive, 3D Figures in PDF Files \cite{barnes2013embedding}}
235 \item Linkes well with \cite{hayes2012pixels}; I heard you liked figures so I put a figure in your PDF
236 \item Title pretty much summarises it; similar to \cite{hayes2012pixels} except these guys actually did something practical
239 \section{27 Bits are not enough for 8 digit accuracy \cite{goldberg1967twentyseven}}
241 Proves with maths, that rounding errors mean that you need at least $q$ bits for $p$ decimal digits. $10^p < 2^{q-1}$
244 \item Eg: For 8 decimal digits, since $10^8 < 2^{27}$ would expect to be able to represent with 27 binary digits
245 \item But: Integer part requires digits bits (regardless of fixed or floating point represenetation)
246 \item Trade-off between precision and range
248 \item 9000000.0 $\to$ 9999999.9 needs 24 digits for the integer part $2^{23} = 83886008$
250 \item Floating point zero = smallest possible machine exponent
251 \item Floating point representation:
253 y &= 0.y_1 y_2 \text{...} y_q \times 2^{n}
255 \item Can eliminate a bit by considering whether $n = -e$ for $-e$ the smallest machine exponent (???)
257 \item Get very small numbers with the same precision
258 \item Get large numbers with the extra bit of precision
262 \section{What every computer scientist should know about floating-point arithmetic \cite{goldberg1991whatevery}}
265 \item Book: \emph{Floating Point Computation} by Pat Sterbenz (out of print... in 1991)
266 \item IEEE floating point standard becoming popular (introduced in 1987, this is 1991)
268 \item As well as structure, defines the algorithms for addition, multiplication, division and square root
269 \item Makes things portable because results of operations are the same on all machines (following the standard)
270 \item Alternatives to floating point: Floating slasi and Signed Logarithm (TODO: Look at these, although they will probably not be useful)
273 \item Base $\beta$ and precision $p$ (number of digits to represent with) - powers of the base can be represented exactly.
274 \item Largest and smallest exponents $e_{min}$ and $e_{max}$
275 \item Need bits for exponent and fraction, plus one for sign
276 \item ``Floating point number'' is one that can be represented exactly.
277 \item Representations are not unique! $0.01 \times 10^1 = 1.00 \times 10^{-1}$ Leading digit of one $\implies$ ``normalised''
278 \item Requiring the representation to be normalised makes it unique, {\bf but means it is impossible to represent zero}.
280 \item Represent zero as $1 \times \beta^{e_{min}-1}$ - requires extra bit in the exponent
282 \item {\bf Rounding Error}
284 \item ``Units in the last place'' eg: 0.0314159 compared to 0.0314 has ulp error of 0.159
285 \item If calculation is the nearest floating point number to the result, it will still be as much as 1/2 ulp in error
286 \item Relative error corresponding to 1/2 ulp can vary by a factor of $\beta$ ``wobble''. Written in terms of $\epsilon$
287 \item Maths $\implies$ {\bf Relative error is always bounded by $\epsilon = (\beta/2)\beta^{-p}$}
288 \item Fixed relative error $\implies$ ulp can vary by a factor of $\beta$ . Vice versa
289 \item Larger $\beta \implies$ larger errors
291 \item {\bf Guard Digits}
293 \item In subtraction: Could compute exact difference and then round; this is expensive
294 \item Keep fixed number of digits but shift operand right; discard precision. Lead to relative error up to $\beta - 1$
295 \item Guard digit: Add extra digits before truncating. Leads to relative error of less than $2\epsilon$. This also applies to addition
297 \item {\bf Catastrophic Cancellation} - Operands are subject to rounding errors - multiplication
298 \item {\bf Benign Cancellation} - Subtractions. Error $< 2\epsilon$
299 \item Rearrange formula to avoid catastrophic cancellation
300 \item Historical interest only - speculation on why IBM used $\beta = 16$ for the system/370 - increased range? Avoids shifting
301 \item Precision: IEEE defines extended precision (a lower bound only)
302 \item Discussion of the IEEE standard for operations (TODO: Go over in more detail)
303 \item NaN allow continuing with underflow and Infinity with overflow
304 \item ``Incidentally, some people think that the solution to such anomalies is never to compare floating-point numbers for equality but instead to consider them equal if they are within some error bound E. This is hardly a cure all, because it raises as many questions as it answers.'' - On equality of floating point numbers
312 \section{Compositing Digital Images \cite{porter1984compositing}}
316 Perter and Duff's classic paper "Compositing Digital Images" lays the
317 foundation for digital compositing today. By providing an "alpha channel,"
318 images of arbitrary shapes — and images with soft edges or sub-pixel coverage
319 information — can be overlayed digitally, allowing separate objects to be
320 rasterized separately without a loss in quality.
322 Pixels in digital images are usually represented as 3-tuples containing
323 (red component, green component, blue component). Nominally these values are in
324 the [0-1] range. In the Porter-Duff paper, pixels are stored as $(R,G,B,\alpha)$
325 4-tuples, where alpha is the fractional coverage of each pixel. If the image
326 only covers half of a given pixel, for example, its alpha value would be 0.5.
328 To improve compositing performance, albeit at a possible loss of precision in
329 some implementations, the red, green and blue channels are premultiplied by the
330 alpha channel. This also simplifies the resulting arithmetic by having the
331 colour channels and alpha channels use the same compositing equations.
333 Several binary compositing operations are defined:
343 The paper further provides some additional operations for implementing fades and
344 dissolves, as well as for changing the opacity of individual elements in a
347 The method outlined in this paper is still the standard system for compositing
348 and is implemented almost exactly by modern graphics APIs such as \texttt{OpenGL}. It is
349 all but guaranteed that this is the method we will be using for compositing
350 document elements in our project.
352 {\bf Some issues with the statements made here...}
354 When introducing Bresenham's algorithm below you say modern graphics systems ``will often use Wu's line-drawing algorithm instead, \emph{as it produces antialiased lines}'' (and don't give a citation). Here you say OpenGL uses Porter-Duff compositing ``almost exactly''. But in their introduction they say: ``
358 \section{Bresenham's Algorithm: Algorithm for computer control of a digital plotter \cite{bresenham1965algorithm}}
359 Bresenham's line drawing algorithm is a fast, high quality line rasterization
360 algorithm which is still the basis for most (aliased) line drawing today. The
361 paper, while originally written to describe how to control a particular plotter,
362 is uniquely suited to rasterizing lines for display on a pixel grid.
364 Lines drawn with Bresenham's algorithm must begin and end at integer pixel
365 coordinates, though one can round or truncate the fractional part. In order to
366 avoid multiplication or division in the algorithm's inner loop,
368 The algorithm works by scanning along the long axis of the line, moving along
369 the short axis when the error along that axis exceeds 0.5px. Because error
370 accumulates linearly, this can be achieved by simply adding the per-pixel
371 error (equal to (short axis/long axis)) until it exceeds 0.5, then incrementing
372 the position along the short axis and subtracting 1 from the error accumulator.
374 As this requires nothing but addition, it is very fast, particularly on the
375 older CPUs used in Bresenham's time. Modern graphics systems will often use Wu's
376 line-drawing algorithm instead, as it produces antialiased lines, taking
377 sub-pixel coverage into account. Bresenham himself extended this algorithm to
378 produce Bresenham's circle algorithm. The principles behind the algorithm have
379 also been used to rasterize other shapes, including B\'{e}zier curves.
381 \section{Quad Trees: A Data Structure for Retrieval on Composite Keys \cite{finkel1974quad}}
383 This paper introduces the ``quadtree'' spatial data structure. The quadtree structure is
384 a search tree in which every node has four children representing the north-east, north-west,
385 south-east and south-west quadrants of its space.
387 \section{Xr: Cross-device Rendering for Vector Graphics \cite{worth2003xr}}
389 Xr (now known as Cairo) is an implementation of the PDF v1.4 rendering model,
390 independent of the PDF or PostScript file formats, and is now widely used
391 as a rendering API. In this paper, Worth and Packard describe the PDF v1.4 rendering
392 model, and their PostScript-derived API for it.
394 The PDF v1.4 rendering model is based on the original PostScript model, based around
395 a set of \emph{paths} (and other objects, such as raster images) each made up of lines
396 and B\'{e}zier curves, which are transformed by the ``Current Transformation Matrix.''
397 Paths can be \emph{filled} in a number of ways, allowing for different handling of self-intersecting
398 paths, or can have their outlines \emph{stroked}.
399 Furthermore, paths can be painted with RGB colours and/or patterns derived from either
400 previously rendered objects or external raster images.
401 PDF v1.4 extends this to provide, amongst other features, support for layering paths and
402 objects using Porter-Duff compositing\cite{porter1984compositing}, giving each painted path
403 the option of having an $\alpha$ value and a choice of any of the Porter-Duff compositing
406 The Cairo library approximates the rendering of some objects (particularly curved objects
407 such as splines) with a set of polygons. An \texttt{XrSetTolerance} function allows the user
408 of the library to set an upper bound on the approximation error in fractions of device pixels,
409 providing a trade-off between rendering quality and performance. The library developers found
410 that setting the tolerance to greater than $0.1$ device pixels resulted in errors visible to the
413 \section{Glitz: Hardware Accelerated Image Compositing using OpenGL \cite{nilsson2004glitz}}
415 This paper describes the implementation of an \texttt{OpenGL} based rendering backend for
416 the \texttt{Cairo} library.
418 The paper describes how OpenGL's Porter-Duff compositing is easily suited to the Cairo/PDF v1.4
419 rendering model. Similarly, traditional OpenGL (pre-version 3.0 core) support a matrix stack
420 of the same form as Cairo.
422 The ``Glitz'' backend will emulate support for tiled, non-power-of-two patterns/textures if
423 the hardware does not support it.
425 Glitz can render both triangles and trapezoids (which are formed from pairs of triangles).
426 However, it cannot guarantee that the rasterization is pixel-precise, as OpenGL does not proveide
429 Glitz also supports multi-sample anti-aliasing, convolution filters for raster image reads (implemented
432 Performance was much improved over the software rasterization and over XRender accellerated rendering
433 on all except nVidia hardware. However, nVidia's XRender implementation did slow down significantly when
434 some transformations were applied.
436 In \cite{kilgard2012gpu}, Kilgard mentions that Glitz has been abandoned. He describes it as ''GPU assisted'' rather than GPU accelerated, since it used the XRender (??) extension.
440 \section{Boost Multiprecision Library \cite{boost_multiprecision}}
443 \item ``The Multiprecision Library provides integer, rational and floating-point types in C++ that have more range and precision than C++'s ordinary built-in types.''
444 \item Specify number of digits for precision as a template argument.
445 \item Precision is fixed... {\bf possible approach to project:} Use \verb/boost::mpf_float<N>/ and increase \verb/N/ as more precision is required?
449 % Some hardware related sounding stuff...
451 \section{A CMOS Floating Point Unit \cite{kelley1997acmos}}
453 The paper describes the implentation of a FPU for PowerPC using a particular Hewlett Packard process (HP14B 0.5$\mu$m, 3M, 3.3V).
454 It implements a ``subset of the most commonly used double precision floating point instructions''. The unimplemented operations are compiled for the CPU.
456 The paper gives a description of the architecture and design methods.
457 This appears to be an entry to a student design competition.
459 Standard is IEEE 754, but the multiplier tree is a 64-bit tree instead of a 54 bit tree.
460 `` The primary reason for implementing a larger tree is for future additions of SIMD [Single Instruction Multiple Data (?)] instructions similar to Intel's MMX and Sun's VIS instructions''.
462 HSPICE simulations used to determine transistor sizing.
464 Paper has a block diagram that sort of vaguely makes sense to me.
465 The rest requires more background knowledge.
467 \section{Simply FPU\cite{filiatreault2003simply}}
469 This is a webpage at one degree of seperation from wikipedia.
471 It talks about FPU internals, but mostly focuses on the instruction sets.
472 It includes FPU assembly code examples (!)
474 It is probably not that useful, I don't think we'll end up writing FPU assembly?
476 FPU's typically have 80 bit registers so they can support REAL4, REAL8 and REAL10 (single, double, extended precision).
478 Note: Presumably this is referring to the x86 80 bit floats that David was talking about?
481 \section{Floating Point Package User's Guide \cite{bishop2008floating}}
483 This is a technical report describing floating point VHDL packages \url{http://www.vhdl.org/fphdl/vhdl.html}
485 In theory I know VHDL (cough) so I am interested in looking at this further to see how FPU hardware works.
486 It might be getting a bit sidetracked from the ``document formats'' scope though.
488 The report does talk briefly about the IEEE standard and normalised / denormalised numbers as well.
490 See also: Java Optimized Processor\cite{jop} (it has a VHDL implementation of a FPU).
492 \section{Low-Cost Microarchitectural Support for Improved Floating-Point Accuracy\cite{dieter2007lowcost}}
494 Mentions how GPUs offer very good floating point performance but only for single precision floats. (NOTE: This statement seems to contradict \cite{hillesland2004paranoia}.
496 Has a diagram of a Floating Point adder.
498 Talks about some magical technique called "Native-pair Arithmetic" that somehow makes 32-bit floating point accuracy ``competitive'' with 64-bit floating point numbers.
500 \section{Accurate Floating Point Arithmetic through Hardware Error-Free Transformations \cite{kadric2013accurate}}
502 From the abstract: ``This paper presents a hardware approach to performing ac-
503 curate floating point addition and multiplication using the idea of error-
504 free transformations. Specialized iterative algorithms are implemented
505 for computing arbitrarily accurate sums and dot products.''
507 The references for this look useful.
509 It also mentions VHDL.
511 So whenever hardware papers come up, VHDL gets involved...
512 I guess it's time to try and work out how to use the Opensource VHDL implementations.
514 This is about reduction of error in hardware operations rather than the precision or range of floats.
515 But it is probably still relevant.
517 This has the Fast2Sum algorithm but for the love of god I cannot see how you can compute anything other than $a + b = 0 \forall a,b$ using the algorithm as written in their paper. It references Dekker\cite{dekker1971afloating} and Kahn; will look at them instead.
519 \section{Floating Point Unit from JOP \cite{jop}}
521 This is a 32 bit floating point unit developed for JOP in VHDL.
522 I have been able to successfully compile it and the test program using GHDL\cite{ghdl}.
524 Whilst there are constants (eg: \verb/FP_WIDTH = 32, EXP_WIDTH = 8, FRAC_WIDTH = 23/) defined, the actual implementation mostly uses magic numbers, so
525 some investigation is needed into what, for example, the "52" bits used in the sqrt units are for.
527 \section{GHDL \cite{ghdl}}
529 GHDL is an open source GPL licensed VHDL compiler written in Ada. It had packages in debian up until wheezy when it was removed. However the sourceforge site still provides a \shell{deb} file for wheezy.
531 This reference explains how to use the \shell{ghdl} compiler, but not the VHDL language itself.
533 GHDL is capable of compiling a ``testbench'' - essentially an executable which simulates the design and ensures it meets test conditions.
534 A common technique is using a text file to provide the inputs/outputs of the test. The testbench executable can be supplied an argument to save a \shell{vcd} file which can be viewed in \shell{gtkwave} to see timing diagrams.
536 Sam has successfully compiled the VHDL design for an FPU in JOP\cite{jop} into a ``testbench'' executable which uses standard i/o instead of a regular file.
537 Using unix domain sockets we can execute the FPU as a child process and communicate with it from our document viewing test software. This means we can potentially simulate alternate hardware designs for FPUs and witness the effect they will have on precision in the document viewer.
539 Using \shell{ghdl} the testbench can also be linked as part a C/C++ program and run using a function; however there is still no way to communicate with it other than forking a child process and using a unix domain socket anyway. Also, compiling the VHDL FPU as part of our document viewer would clutter the code repository and probably be highly unportable. The VHDL FPU has been given a seperate repository.
541 \section{On the design of fast IEEE floating-point adders \cite{seidel2001onthe}}
543 This paper gives an overview of the ``Naive'' floating point addition/subtraction algorithm and gives several optimisation techniques:
545 TODO: Actually understand these...
548 \item Use parallel paths (based on exponent)
549 \item Unification of significand result ranges
550 \item Reduction of IEEE rounding modes
551 \item Sign-magnitude computation of a difference
552 \item Compound Addition
553 \item Approximate counting of leading zeroes
554 \item Pre-computation of post-normalization shift
557 They then give an implementation that uses these optimisation techniques including very scary looking block diagrams.
559 They simulated the FPU. Does not mention what simulation method was used directly, but cites another paper (TODO: Look at this. I bet it was VHDL).
561 The paper concludes by summarising the optimisation techniques used by various adders in production (in 2001).
563 This paper does not specifically mention the precision of the operations, but may be useful because a faster adder design might mean you can increase the precision.
565 \section{Re: round32 ( round64 ( X ) ) ?= round32 ( X ) \cite{beebe2011round32}}
567 I included this just for the quote by Nelson H. F. Beebe:
569 ``It is too late now to repair the mistakes of the past that are present
570 in millions of installed systems, but it is good to know that careful
571 research before designing hardware can be helpful.''
573 This is in regards to the problem of double rounding. It provides a reference for a paper that discusses a rounding mode that eliminates the problem, and a software implementation.
575 It shows that the IEEE standard can be fallible!
577 Not sure how to work this into our literature review though.
580 \section{Basic Issues in Floating Point Arithmetic and Error Analysis \cite{demmel1996basic}}
582 These are lecture notes from U.C Berkelye CS267 in 1996.
585 \section{Charles Babbage \cite{dodge_babbage, nature1871babbage}}
587 Tributes to Charles Babbage. Might be interesting for historical background. Don't mention anything about floating point numbers though.
589 \section{GPU Floating-Point Paranoia \cite{hillesland2004paranoia}}
591 This paper discusses floating point representations on GPUs. They have reproduced the program \emph{Paranoia} by William Kahan for characterising floating point behaviour of computers (pre IEEE) for GPUs.
594 There are a few remarks about GPU vendors not being very open about what they do or do not do with
597 Unfortunately we only have the extended abstract, but a pretty good summary of the paper (written by the authors) is at: \url{www.cs.unc.edu/~ibr/projects/paranoia/}
601 ``... [GPUs are often similar to IEEE] However, we have found
602 that GPUs do not adhere to IEEE standards for floating-point op-
603 erations, nor do they give the information necessary to establish
604 bounds on error for these operations ... ''
606 and ``...Our goal is to determine the error bounds on floating-point op-
607 eration results for quickly evolving graphics systems. We have cre-
608 ated a tool to measure the error for four basic floating-point opera-
609 tions: addition, subtraction, multiplication and division.''
611 The implementation is only for windows and uses glut and glew and things.
612 Implement our own version?
614 \section{A floating-point technique for extending the available precision \cite{dekker1971afloating}}
616 This is Dekker's formalisation of the Fast2Sum algorithm originally implemented by Kahn.
619 z &= \text{RN}(x + y) \\
620 w &= \text{RN}(z - x) \\
621 zz &= \text{RN}(y - w) \\
622 \implies z + zz &= x + y
625 There is a version for multiplication.
627 I'm still not quite sure when this is useful. I haven't been able to find an example for $x$ and $y$ where $x + y \neq \text{Fast2Sum}(x, y)$.
629 \section{Handbook of Floating-Point Arithmetic \cite{HFP}}
631 This book is amazingly useful and pretty much says everything there is to know about Floating Point numbers.
632 It is much easier to read than Goldberg or Priest's papers.
634 I'm going to start working through it and compile their test programs.
636 \subsection{A sequence that seems to converge to a wrong limit - pgs 9-10, \cite{HFP}}
639 u_n &= \left\{ \begin{array}{c} u_0 = 2 \\ u_1 = -4 \\ u_n = 111 - \frac{1130}{u_{n-1}} + \frac{3000}{u_{n-1}u_{n-2}}\end{array}\right.
642 The limit of the series should be $6$ but when calculated with IEEE floats it is actually $100$
643 The authors show that the limit is actually $100$ for different starting values, and the error in floating point arithmetic causes the series to go to that limit instead.
647 \includegraphics[width=0.8\textwidth]{figures/handbook1-1.pdf}
648 \caption{Output of Program 1.1 from \emph{Handbook of Floating-Point Arithmetic}\cite{HFP} for various IEEE types}
652 \subsection{Mr Gullible and the Chaotic Bank Society pgs 10-11 \cite{HFP}}
654 This is an example of a sequence involving $e$. Since $e$ cannot be represented exactly with FP, even though the sequence should go to $0$ for $a_0 = e - 1$, the representation of $a_0 \neq e - 1$ so the sequence goes to $\pm \infty$.
656 To eliminate these types of problems we'd need an \emph{exact} representation of all real numbers.
657 For \emph{any} FP representation, regardless of precision (a finite number of digits) there will be numbers that can't be represented exactly hence you could find a similar sequence that would explode.
659 IE: The more precise the representation, the slower things go wrong, but they still go wrong, {\bf even with errorless operations}.
662 \subsection{Rump's example pg 12 \cite {HFP}}
664 This is an example where the calculation of a function $f(a,b)$ is not only totally wrong, it gives completely different results depending on the CPU. Despite the CPU conforming to IEEE.
666 \section{Scalable Vector Graphics (SVG) 1.1 (Second Edition) \cite{svg2011-1.1}}
668 Scalable Vector Graphics (SVG) is a XML language for describing two dimensional graphics. This document is \url{http://www.w3.org/TR/2011/REC-SVG11-20110816/}, the latest version of the standard at the time of writing.
670 \subsubsection{Three types of object}
672 \item Vector graphics shapes (paths)
673 \item Images (embedded bitmaps)
677 \subsubsection{Rendering Model and Precision}
679 SVG uses the ``painter's model''. Paint is applied to regions of the page in an order, covering the paint below it according to rules for alpha blending.
681 ``Implementations of SVG are expected to behave as though they implement a rendering (or imaging) model cor-
682 responding to the one described in this chapter. A real implementation is not required to implement the model in
683 this way, but the result on any device supported by the implementation shall match that described by this model.''
685 SVG uses {\bf single precision floats}. Unlike PDF and PostScript, the standard specifies this as a {\bf minimum} range from \verb/-3.4e+38F/ to \verb/+3.4e+38F/
687 ``It is recommended that higher precision floating point storage and computation be performed on operations
688 such as coordinate system transformations to provide the best possible precision and to prevent round-off errors.''
690 There is also a ``High Quality SVG Viewers'' requirement to use at least {\bf double} precision floats.
692 \section{OpenVG Specification 1.1 \cite{rice2008openvg}}
693 ``OpenVG is an application programming interface (API) for hardware-accelerated two-
694 dimensional vector and raster graphics developed under the auspices of the Khronos
695 Group \url{www.khronos.org}.''
697 Specifically, provides the same drawing functionality required by a high-performance SVG document viewer (SVG Tiny 1.2)
698 TODO: Look at that $\neq$ SVG 1.1
700 It is designed to be similar to OpenGL.
702 \subsubsection{Precision}
703 ``All floating-point values are specified in standard IEEE 754 format. However,
704 implementations may clamp extremely large or small values to a restricted
705 range, and internal processing may be performed with lesser precision. At least
706 16 bits of mantissa, 6 bits of exponent, and a sign bit must be present, allowing
707 values from $\pm 2^{-30}$ to $\pm2^{31}$ to be represented with a fractional precision of at least 1
710 IEEE but with a non standard number of bits.
712 Presumably the decreased precision is for increased efficiency the standard states that example applications are for ebooks.
715 \section{Document Object Model --- pugixml 1.4 manual \cite{pugixmlDOM}}
717 Pugixml is a C++ library for parsing XML documents\cite{kapoulkine2014pugixml}. XML is based on the Document Object Model (DOM); this is explained pretty well by the pugixml manual.
719 The document is the root node of the tree. Each child node has a type. These may
721 \item Have attributes
722 \item Have child nodes of their own
726 In the case of HTML/SVG an XML parser within the browser/viewer creates the DOM tree from the XML (plain text) and then interprets that tree to produce the objects that will be rendered.
728 Example of XML $\to$ DOM tree given at\cite{pugixmlDOM}.
729 Example of XML parsing using pugixml is in \shell{code/src/tests/xml.cpp}
734 \begin{lstlisting}[language=XML,basicstyle=\ttfamily]
735 <?xml version="1.0"?>
736 <mesh name="mesh_root">
737 <!-- here is a mesh node -->
739 <![CDATA[someothertext]]>
741 <node attr1="value1" attr2="value2" />
742 <node attr1="value2">
749 \includegraphics[width=0.6\textwidth]{references/pugixmlDOM-dom_tree.png}
750 \caption{Tree representation of the above listing \cite{pugixmlDOM}}
753 \section{An Algorithm For Shading of Regions on Vector Display Devices \cite{brassel1979analgorithm}}
755 All modern display devices are raster based and therefore this paper is mainly of historical interest. It provides some references for shading on a raster display.
757 The algorithm described will shade an arbitrary simply-connected polygon using one or two sets of parallel lines.
759 The ``traditional'' method is:
761 \item Start with a $N$ vertex polygon, rotate coords by the shading angle
762 \item Determine a bounding rectangle
763 \item For $M$ equally spaced parallel lines, compute the intersections with the boundaries of the polygon
764 \item Rotate coordinates back
765 \item Render the $M$ lines
768 This is pretty much exactly how an artist would shade a pencil drawing. It is $O(M\times N)$.
770 The algorithm in this paper does:
772 \item Rotate polygon coords by shading angle
773 \item Subdivide the polygon into trapezoids (special case triangle)
774 \item Shade the trapezoids independently
775 \item Rotate it all back
777 It is more complicated than it seems. The subdivision requires a sort to be performed on the vertices of the polygon based on their rotated $x$ and $y$ coordinates.
779 \section{An Algorithm For Filling Regions on Graphics Display Devices \cite{lane1983analgorithm}}
781 This gives an algorithm for for polygons (which may have ``holes'').
782 It requires the ability to ``subtract'' fill from a region; this is (as far as I can tell) difficult for vector graphics devices but simple on raster graphics devices, so the paper claims it is oriented to the raster graphics devices.
784 If the polygon is defined by $(x_i, y_i)$ then this algorithm iterates from $i = 2$ and alternates between filling and erasing the triangles $[(x_i, y_i), (x_{i+1}, y_{i+1}), (x_1, y_1)]$. It requires no sorting of the points.
786 The paper provides a proof that the algorithm is correct and is ``optimal in the number of pixel updates required for convex polygons''.
787 In the conclusion it is noted that trapezoids could be used from a fixed line and edge of the polygon, but this is not pixel optimal.
789 This paper doesn't have a very high citation count but it is cited by the NVIDIA article \cite{kilgard2012gpu}.
790 Apparently someone else adapted this algorithm for use with the stencil buffer.
792 \section{GPU-accelerated path rendering \cite{kilgard2012gpu, kilgard300programming}}
794 Vector graphics on the GPU; an NVIDIA extension. \cite{kilgard300programming} is the API.
798 \item The focus has been on 3D acceleration in GPUs; most path rendering is done by the CPU.
799 \item Touch devices allow the screen to be transformed rapidly; CPU rastering of the path becomes inefficient
801 \item The source of the ugly pixelated effects on a smartphone when scaling?
803 \item Especially when combined with increased resolution of these devices
804 \item Standards such as HTML5, SVG, etc, expose path rendering
805 \item Javascript is getting fast enough that we can't blame it anymore (the path rendering is the bottleneck not the JS)
806 \item GPU is more power efficient than the CPU
809 Results show the extension is faster than almost every renderer it was compared with for almost every test image.
811 Comparisons to other attempts:
813 \item Cairo and Glitz \cite{nilsson2004glitz} (abandoned)
814 \ \item Direct2D from Microsoft uses CPU to tesselate trapezoids and then renders these on the GPU
815 \item Skia in Android/Chrome uses CPU but now has Ganesh which is also hybrid CPU/GPU
816 \item Khronos Group created OpenVG\cite{rice2008openvg} with several companies creating hardware units to implement the standard. Performance is not as good as ``what we report''
820 \section{A Multiple Precision Binary Floating Point Library With Correct Rounding \cite{fousse2007mpfr}}
822 This is what is now the GNU MPFR C library; it has packages in debian wheezy.
824 The library was motivated by the lack of existing arbitrary precision libraries which conformed to IEEE rounding standards.
825 Examples include Mathematica, GNU MP (which this library is actually built upon), Maple (which is an exception but buggy).
827 TODO: Read up on IEEE rounding to better understand the first section
831 \item $(s, m, e)$ where $e$ is signed
832 \item Precision $p$ is number of bits in $m$
833 \item $\frac{1}{2} \leq m < 1$
834 \item The leading bit of the mantissa is always $1$ but it is not implied
835 \item There are no denormalised numbers
836 \item Mantissa is stored as an array of ``limbs'' (unsigned integers) as in GMP
839 The paper includes performance comparisons with several other libraries, and a list of literature using the MPFR (the dates indicating that the library has been used reasonably widely before this paper was published).
841 \section{Lecture Notes on the Status of IEEE Standard 754 for Binary Floating-Point Arithmetic \cite{kahan1996ieee754}}
843 11 years after the IEEE 754 standard becomes official, Kahan discusses various misunderstood features of the standard and laments at the failure of compilers and microprocessors to conform to some of these.
845 I am not sure how relevant these complaints are today, but it makes for interesting reading as Kahan is clearly very passionate about the need to conform \emph{exactly} to IEEE 754.
847 Issues considered are: Rounding rules, Exception handling and NaNs (eg: The payload of NaNs is not used in any software Kahan is aware of), Bugs in compilers (mostly Fortran) where an expression contains floats of different precisions (the compiler may attempt to optimise an expression resulting in a failure to round a double to a single), Infinity (which is unsupported by many compilers though it is supported in hardware)...
850 An example is this Fortran compiler ``nasty bug'' where the compiler replaces $s$ with $x$ in line 4 and thus a rounding operation is lost.
851 \begin{lstlisting}[language=Fortran, basicstyle=\ttfamily]
852 real(8) :: x, y % double precision (or extended as real(12))
853 real(4) :: s, t % single precision
854 s = x % s should be rounded
855 t = (s - y) / (...) % Compiler incorrectly replaces s with x in this line
858 \subsection{The Baleful Influence of Benchmarks \cite{kahan1996ieee754} pg 20}
860 This section discusses the tendency to rate hardware (or software) on their speed performance and neglect accuracy.
862 Is this complaint still relevant now when we consider the eagerness to perform calculations on the GPU?
863 ie: Doing the coordinate transforms on the GPU is faster, but less accurate (see our program).
865 Benchmarks need to be well designed when considering accuracy; how do we know an inaccurate computer hasn't gotten the right answer for a benchmark by accident?
867 A proposed benchmark for determining the worst case rounding error is given. This is based around computing the roots to a quadratic equation.
868 A quick scan of the source code for paranoia does not reveal such a test.
870 As we needed benchmarks for CPUs perhaps we need benchmarks for GPUs. The GPU Paranoia paper\cite{hillesland2004paranoia} is the only one I have found so far.
872 \section{Prof W. Kahan's Web Pages \cite{kahanweb}}
874 William Kahan, architect of the IEEE 754-1985 standard, creator of the program ``paranoia'', the ``Kahan Summation Algorithm'' and contributor to the revised standard.
876 Kahan's web page has more examples of errors in floating point arithmetic (and places where things have not conformed to the IEEE 754 standard) than you can poke a stick at.
878 Wikipedia's description is pretty accurate: ``He is an outspoken advocate of better education of the general computing population about floating-point issues, and regularly denounces decisions in the design of computers and programming languages that may impair good floating-point computations.''
880 Kahan's articles are written with almost religious zeal but they are backed up by examples and results, a couple of which I have confirmed.\footnote{One that doesn't work is an example of wierd arithmetic in Excel 2000 when repeated in LibreOffice Calc 4.1.5.3} This is the unpublished work. I haven't read any of the published papers yet.
882 The articles are filled sporadically with lamentation for the decline of experts in numerical analysis (and error) which is somewhat ironic; if there were no IEEE 754 standard meaning any man/woman and his/her dog could write floating point arithmetic and expect it to produce platform independent results\footnote{Even if, as Kahan's articles often point out, platforms aren't actually following IEEE 754} this decline would probably not have happened.
884 These examples would be of real value if the topic of the project were on accuracy of numerical operations. They also explain some of the more bizarre features of IEEE 754 (in a manner attacking those who dismiss these features as being too bizarre to care about of course).
886 It's somewhat surprising he hasn't written anything (that I could see from scanning the list) about the lack of IEEE in PDF/PostScript (until Adobe Reader 6) and further that the precision is only binary32.
888 I kind of feel really guilty saying this, but since our aim is to obtain arbitrary scaling, it will be inevitable that we break from the IEEE 754 standard. Or at least, I (Sam) will. Probably. Also there is no way I will have time to read and understand the thousands of pages that Kahan has written.
890 Therefore we might end up doing some things Kahan would not approve of.
892 Still this is a very valuable reference to go in the overview of floating point arithmetic section.
894 \chapter{General Notes}
896 \section{The DOM Model}
898 A document is modelled as a tree. The root of the tree is the document. This has nodes of varying types. Some nodes may have children, attributes, and data.
900 \section{Floating-Point Numbers\cite{HFP,goldberg1991whatevery,goldberg1992thedesign,priest1991algorithms}}
902 A set of FP numbers is characterised by:
904 \item Radix (base) $\beta \geq 2$
905 \item Precision %$p \req 2$ ``number of sig digits''
906 \item Two ``extremal`` exponents $e_min < 0 < e_max$ (generally, don't have to have the $0$ in there)
909 Numbers are represented by {\bf integers}: $(M, e)$ such that $x = M \times \beta^{e - p + 1}$
911 Require: $|M| \leq \beta^{p}-1$ and $e_min \leq e \leq e_max$.
913 Representations are not unique; set of equivelant representations is a cohort.
915 $\beta^{e-p+1}$ is the quantum, $e-p+1$ is the quantum exponent.
917 Alternate represetnation: $(s, m, e)$ such that $x = (-1)^s \times m \times \beta^{e}$
918 $m$ is the significand, mantissa, or fractional part. Depending on what you read.
924 \section{Rounding Errors}
927 They happen. There is ULP and I don't mean a political party.
929 TODO: Probably say something more insightful. Other than "here is a graph that shows errors and we blame rounding".
931 \subsection{Results of calculatepi}
933 We can calculate pi by numerically solving the integral:
935 \int_0^1 \left(\frac{4}{1+x^2}\right) dx &= \pi
938 Results with Simpson Method:
941 \includegraphics[width=0.8\textwidth]{figures/calculatepi.pdf}
942 \caption{Example of accumulated rounding errors in a numerical calculation}
945 Tests with \verb/calculatepi/ show it's not quite as simple as just blindly replacing all your additions with Fast2Sum from Dekker\cite{dekker1971afloating}.
946 ie: The graph looks exactly the same for single precision. \verb/calculatepi/ obviously also has multiplication ops in it which I didn't change. Will look at after sleep maybe.
950 \bibliographystyle{unsrt}
951 \bibliography{papers}