The following are excerpts from an interesting Usenet discussion about the differences in convention between using column vectors or row vectors. This drives the 4x4 matrix representation as well, since vector-matrix multiplication is represented with the matrix on the left for column vectors, and on the right for row vectors. The 4x4 matrices in each case are the transpose of the other convention.
Put another way, if you're using row vectors, your 4x4 translation components are in the bottom row. If you're using column vectors, your 4x4 translation components are in the right column.
Generally, in 3D graphics, you have to keep on your toes, since this convention varies from system to system.
From: email@example.com (Mark Segal)
Subject: Re: Row major or column major matrices?
Date: 26 Jan 1993 03:35:10 GMT
I'm surprised that this simple topic has generated so many posts. I'm the one responsible for the 'column-major ordering' used in OpenGL, so I'll try to explain what's going on to try to put this discussion to rest.
First, there are two issues that seem to be confused. One issue is how matrices are stored in memory, and the other is whether one treats vectors as rows of coordinates or as columns.
I'll dispense with the second issue first. Recent mathematical treatments of linear algebra and related fields invariably treat vectors as columns (there are some technical reasons for this). For some reason, this has not been the case in computer graphics, where vectors were written as rows, thus transposing everything from standard mathematical usage. When I wrote the OpenGL spec, I decided to do my part to right this heinous inconsistency. Thus the spec is written with vectors treated as columns, with a matrix correspondingly applied on the left of a column.
The one difficulty was compatibility with the current GL, where vectors had been written as rows. So I come up with this subterfuge: say that matrices in OpenGL are stored in column major order. The point is that you could rewrite the spec with everything transposed (with vectors written as rows), and everything would be exactly the same as it was, including the row major ordering of matrices.
So we've come to the second issue: how matrices are stored in memory. If you treat vectors as columns, then in OpenGL matrices are stored column major. If you treat vectors as rows, then matrices are stored (as they are in the current GL) row major. That's all there is to it. OpenGL documentation treats vectors as columns, so it will be simplest for most people who use it to adopt this convention. But it won't affect any code you may have written to calculate matrices for the current GL in any way.
To say without qualification that 'OpenGL stores matrices this way' is technically nonsense. OpenGL does nothing. OpenGL DOCUMENTATION assumes that vectors are columns therefore implying that matrices are given column major, but you are free to take the alternate interpretation if you wish. It is also nonsense to say that the hardware (or an OpenGL implementation) 'premultiples' or 'postmultiplies' (I could never remember which was which anyway). OpenGL applies a matrix (representing a transformation) to a series of coordinates (representing a vector) using the rules of matrix multiplication. You may view this as performing the multiplication c' = Mc, where c and c' are columns, or as r' = rM(transpose) where r and r' are rows (and r=c(transpose) and r'=c'(transpose) ). The two formulas represent the same calculation.
So, in spite of this lengthy explanation, I don't think that there is actually any issue here, since there is no change from the current GL, and since a programmer is free to view matrices and vectors as she pleases.
Once upon a time (like, last week) I wrote:
Someone was telling me that OpenGL allows only post-concatenation of matrices. This seemed bass-ackwards, and after a few minutes reconfirming in my slow turtle brain that you'd normally want pre-concatenation of matrices, I read over the OpenGL manual page carefully. Aha, they store matrices in column-major form:a0 a4 a8 a12 a typical 1 0 0 -2 a1 a5 a9 a13 translation 0 1 0 5 a2 a6 a10 a14 matrix --> 0 0 1 3 a3 a7 a11 a15 0 0 0 1
Just out of curiosity, I looked at a few computer graphics books and whatnot and how they presented 3D matrices.
I received a large amount of mail about this topic (thanks, everyone), with much of it having unexpected angles on the question. There were a fair number of responses which said that column major is what mathematicians use (and that computer graphics people are finally seeing the light), while others said they learned row major order in mathematics, engineering, and physics classes. Go figure. Actually, Ron Levine's answer has an authoritative sound to it, and he says that the type of notation used depends on the field of mathematics.
As Ron also points out, a source of confusion arose from my propagation of the term "column major", which is what OpenGL calls the form shown above. My question was meant to address the pencil-and-paper notation of matrices, not the internal representation in the machine. As Mark Segal (who designed OpenGL) writes, there is _no_ difference between IRIS GL and OpenGL on the way matrices are passed via the API's or used internally, so there should be no porting problems (other than rearranging the neuro-electrical pathways in the programmer's head!). The data elements 0 through 15 are in the same order for a "row major" matrix as for a "column major" matrix. To beat it to death: the OpenGL matrix stores translation values in elements 12,13,14, which is exactly where they are stored in IRIS GL. There may be other systems in which the translation elements are stored in locations 3,7, and 11, but this is a separate computer language issue (can you say "FORTRAN"?). As Ron Levine notes, postfix vs. prefix is probably a better way to talk about these representations, though then we have to remember that "prefix == translation vector in column".
Anyway, here's a new list of who uses what. I started to list everyone who replied and anyone they mentioned, but it got too long and pointless. The ratio seems to be 2:1 of row major (postfix) to column major (prefix) users.
Row Major users:
Column Major users:
Here are various quotes of interest. If you have too much time on your hands, write me and I'll send you the entire file of mail message replies. Some related "this way or that" issues that people mentioned:
Interesting to us amateur psychologists, there seems to be an inferiority complex in the computer graphics community, e.g. "mathematicians use column major, so we should, too". Reading the replies over, there's a fair mix of people who learned row major in their engineering and physics and even math classes. At this point most computer graphics people seem to use row major, and I like Robin Forrest's reason of how this began. So, I don't believe there's currently a "right" way, but rather it's worth conforming to the usage of whoever you're talking with (see Jon Leech's quote as a good example of this). It'll be interesting to see how this sorts itself out in a decade or two. I suspect Gavin Bell's quote is probably right, since the F&VD&F&H book, OpenGL, Phigs, and PEX are major influences in our field.
I wish I could have found a relevant quote from Major Major Major Major in Catch-22, but no such luck. Anyway, here we go (in some vaguely structured order):
"`Row-major' and `column-major', as I have understood the terms for years, refer to how the matrices are stored in memory, not to how they are used to represent transformations. That is, the terms distinguish between the Fortran convention, in which the first subscript varies most rapidly as memory address increases, and the C convention, in which the second subscript varies most rapidly as memory address increases. Although your note refers to how matrices are stored, I think you are really concerned with quite a different distinction, namely, whether the matrix represents a linear transformation as a left multiplier of column vectors or as a right multiplier of row vectors. Mathematicians sometimes distinguish these two conventions as the `pre-fix' and `post-fix' conventions.
Mathematicians have long been split into two camps, between pre-fix and post-fix preferences. Geometers and analysts have tended more to the pre-fix convention--thus you learned in calculus that f(x) denotes the value of a function f for a domain element x. Then, if you compose two functions, (gf)(x) means g(f(x)), that is, first apply f then apply g. This corresponds to interpreting a matrix as a left multiplier of column vectors.
But some schools of abstract algebraists logicians, and some authors of computer graphics books, have insisted on the post-fix (x)f or xf. I always work in the pre-fix convention when I have a choice, but this is only because in the past most of my math professors and books have used it and so it is a little more natural to me. I especially prefer it in computer graphics, which is inspired more by geometry than by abstract algebra.
But the two conventions are completely equivalent, and one should have no difficulty in using either. Of course it is important to know which convention is used by any particular system with which you are working. And of course, the matrices may be stored in row-major or column-major order for either convention of writing functional expressions; that distinction tends to depend on the programming language."firstname.lastname@example.org (Ron Levine)
"I plead guilty. As an author of the OpenGL specification, I'm the one responsible for this foolishness. I explained it in a message in comp.graphics.opengl a couple of months ago (you can probably find it if you look through your archives), but since you seemed to have figured out the salient points, I'll just explain my reasons.
I have always been irritated with computer graphics treatments using rows to represent vectors, since this is at odds with what is done throughout the mathematics world, where vectors are represented by columns (there are some good reasons for this, but I'll spare you). I wanted to do my part to remedy this disturbing situation (it had nothing to do with Foley and van Dam; I wasn't even aware until recently that they had adopted a new convention). So I wrote everything in the spec with vectors as columns, with matrices and application order corresponding. The only snag was compatibility with existing GL programs, so I used the subterfuge of saying that the matrices are stored in column major order. This was kind of a dirty trick, but I didn't think it would seriously bother anyone.
As you note, of course, there is no actual difference. You can rewrite everything the old way, and then your matrices are in row major order. I'm sure it's not much consolation, but many numerical analysts and FORTRAN afficianados actually prefer column major order."email@example.com (Mark Segal)
"I have no strong preference for either, but it would make my life much easier if real-world systems could choose one and stick with it! I recently had days of unnecessary skullsweat grafting OpenGL evaluator onto a utility library that used the opposite ordering."jbulf@balsa.Berkeley.EDU (Jeff Bulf, aka Dr. Memory)
"Funny thing you know, one of the other summer students here are the ACGC is writing a matrix transformation library, and nicked some of his matrix formulas from the new GL reference manual. Guess what? They did some really bizarre things (especially the perspective :-) until I told him they in row major form, not column major like we were used to!"firstname.lastname@example.org.OZ.AU (Kendall Bennett)
"I was also shocked to learn that OpenGL decided to use post multiplication of matrices. Not because I learned pre-multiplication, but because it is opposite of what IRIS GL uses. [...]
I learned the 'NEW' way of using matrices. The reason that Foley decided to use post-multiplication of matrices in the second edition, is because it allows the notation to be written in the same manner as that used by mathematicians.
If you look in linear algebra books, they all use post-multiplication notation, and if you show a math person the 'OLD' way they will look at you like a leper... At least this is the opinion I formed from my teacher (who was Dr. Foley)."email@example.com (Wayne Wooten)
"I learned Row Major at Computervision and have been confused by column major ever since. Now the only question I want answered is whether the X offset is the fourth element of the data structure.
I'm always amused at people saying the everyone does it this way or that. Matrices are a fairly recent invention (in the 40's?). And different disciplines use row & column matrices arbitrarily, at least in my limited understanding. It's so arbitrary."Jan "YON" Hardenbergh (firstname.lastname@example.org)
[Jan is one of the designers of PEX. - EAH]
"The translation-matrix for me looks like:1 0 0 0 -2 1 0 0 5 0 1 0 3 0 0 1Christian Linhart <email@example.com>
"The question of row versus column vectors is historical. Steve Coons used row vectors for his influential early papers on transformations (University of Michigan Summer Courses in the mid 60's). I asked him why in 1967 and he said it was because it was easier for the stenographer to type row vectors!
Regarding left and right hand co-ordinate systems, Larry Roberts who introduced 4x4 matrix transforms to graphics used RH co-ordinates and anyone brought up with his papers or Coons papers uses RH systems. Ivan Sutherland, on the other hand, believed that z should be positive into the screen and hence used LH co-ordinates as did his pupils. Sutherland's treatment of perspective also diverged from that devised by Roberts, hence the two schools of thought there."Robin Forrest (firstname.lastname@example.org)
"Actually, I have no clear preference, and I'll explain why. In all my education (math and physics, B.S Physics UCLA 1986), only row-major matrices were used, so I naturally think of that as the standard. However, I have an idea as to why so many people are switching to column-major notation. It's so that matrices can be post-multiplied instead of pre-multiplied! Post- multiplication notation accords better with the compound operators of C and C++. If you're writing a matrix class in C++, and you want to multiply one matrix by another, it seems natural to define the operator `*=' as the multiplication operator, such that:A *= B means A = A * B.
I was writing such a class for a simple 3-D graphics program, and I quickly realised that the most common multiplication I wanted to do was PRE- multiplication. I therefore said `Screw convention', and defined `*=' so thatA *= B means A = B * A.
So, my solution was to change the standard C relationship between `*=' and `*'. Apparently, others decided it was better to change the very operation of matrix multiplication to imply a column-major order, so that `A *= B' would still mean `A = A*B'."email@example.com (John Scott Peter XXXIII)
"`I am the major model of a modern Row Major General'..... I just saw the Pirates of Penzance last week. Even Gilbert and Sullivan were Row Major fanatics, if I got the quote correct 8-)"Brian Corrie <Brian.Corrie@cs.anu.edu.au>
"I go for what you call "row major". My "pappy's knee" wasCS374 -- R.F. Sproull -- Introduction to Computer Graphics, Fall 1976
at Caltech, straight from the (then new) book. Together with Sutherland about the most lucid lectures I ever heard (only heard Feynman speak once, and to a semi-technical audience). I later had the privilege of brief involvement with N&S vol. II while at Xerox Pasadena (our laser printers produced all the camera-ready copy)."Alan Wm Paeth <firstname.lastname@example.org>
"I learned linear algebra before computer graphics. The linear algebra books normally write vectors as column vectors."Kevin.Wu@Eng.Sun.COM (Kevin Wu)
"I personally prefer row major... however, computer hardware that I've worked on recently preferred column major. The reason for this was the presence of an auto-increment load/store assembly instruction which allowed the matrix routines to work faster if the data was presented to them in column major form."reid.judd@East.Sun.COM (Reid Judd - Sun NC Development Center)
[This answer (and a few others received) discusses the storage of row vs. column, not the presentation on paper of row vs. column, a different issue. "Row major" is definitely an overloaded term. - EAH]
"Avast there, ye scurvy scum! Doré uses column vectors and LIKES IT!!"email@example.com (Steve Hollasch)
"The reason/justification/excuse for the change in matrices is to switch to the way the mathematicians have been doing it all along. They have a far longer history than we do in this affine transform jazz."Sam Uselton (firstname.lastname@example.org)
"I'm a row major type person. From my engineering and graphics background that is all I usually see. I learned my first matrix stuff in a Linear Algebra class in college. [...]
I don't really see any difference between the two methods other than I wish they would pick one and stick with it!!! By the way, engineers, structural mechanics, fluids guys, and aerodynamicists use row major from what I've been taught and seen."Mike Goza (email@example.com)
"Characterizing the matrices is not the clearest way IMO. I prefer stating that points are either row vectors or column vectors.
Jim Blinn taught me to use row vectors in 1982 and Al Barr taught me to use column vectors in 1984. I prefer column vectors these days, although I'll go either way as needed. In teaching an undergraduate graphics course last semester I used row vectors, since I was handing out copies of Blinn papers and didn't want to confuse them too much.
I'm less sanguine about left-handed coordinate systems, which are Tools of the Devil (as anyone doing physically based modelling should think)."Jon Leech <firstname.lastname@example.org>
"I got started using row-major matrices for transforms when we used mimeographed copies of the manuscript for the 1st edition of Newman and Sproull in Prof. Bert Herzog's class at the Univ. of Michigan! That should date me pretty well. :-)
The system I work on now uses column-major matrices, and to make things more interesting, homogeneous coordinates are expressed as (w,x,y,z) instead of (x,y,z,w). That mixes things up pretty well in those transform matrices and keeps me on my toes...
I hate to try to classify people by their preferences to use row-major versus column-major form. Maybe engineers tend to use row-major, while mathematicians tend to use the transpose; don't know about physicists. Conventions do seem to tend to cluster around various subfields. Computer graphics, maybe because it is so cross-disciplinary, is definitely in a state of flux on transform order conventions.
The row-major convention is good for thinking about transforms as operators. Column-major form is best for thinking about transforms as functions, e.g.,v' = T(v)
for a transform T that maps vector v to vector v'; then transform concatenation is just function composition:v' = T2(T1(v))
May you always compose your matrices in the "right" order..."Ron Capelli <email@example.com>
"It's even more surprising since the current version of GL normally pre-concatenates. I can imagine some hard to find bugs when people port old GL code in which they generate their own matrices from scratch. [not true - EAH]
I'm not too upset though, cause column vectors and post-concatenation is what I learned as an engineer, and so it seems more natural to me. Note, that pre-concatenation with column vectors or post-concatenation with row vectors is useful. In this one instance I prefer PHIGS, which has a direct mechanism for doing both pre- & post-concatenation of matrices ... you can post-concatenate in GL but only indirectly and inefficiently.
According to the GL manual, GL uses row vectors and pre-multiplication for purely historical reasons related to the efficiency of early hardware."ledwards@leland.Stanford.EDU (Laurence James Edwards)
"In 10 years we will probably remember the good old days when the translations were at the bottom of the matrix..."firstname.lastname@example.org (Gavin Bell)