Pascal Getreuer, January 2008

Contents
1 The Proler 2 Array Preallocation 3 JIT Acceleration 4 Vectorization 4.1 Vectorized Computations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Vectorized Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 Inlining Simple Functions 6 Referencing Operations 7 Numerical Integration 7.1 One-Dimensional Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2 Multidimensional Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 Signal Processing 9 Miscellaneous Tricks 9.1 Clip a value without using if statements . . 9.2 Convert any array into a column vector . . 9.3 Find the min/max of a matrix or N-d array 9.4 Flood lling . . . . . . . . . . . . . . . . . . 9.5 Vectorized use of set on GUI objects . . . . 10 Further Reading 2 4 6 8 8 9 13 15 18 19 20 22 25 25 25 25 26 26 27

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

Introduction
Matlab code is interpreted during runtime. Languages like C and Fortran are faster because they are rst compiled into the computers native language. Interpreting in realtime enables better platform independence, greater language exibility, and interactive debugging, but at the cost of slower speed, increased overhead, and limited low-level control. This article discusses strategies for improving the speed of Matlab code. Keep in mind that speed depends heavily on the computation platform and that the fastest method on one system is not necessarily the fastest on another. Second, Matlab has gone through many version and continues to evolve: a slow method may become fast in a newer version. This article provides methods that are generally fast, but makes no claim on what is fastest.

Caution!
Learn the language rst. Optimization requires comfort with the syntax and features of the language. This article is not a tutorial on MATLAB. Use comments. Optimized code tends to be terse and cryptic. Help others and yourself by remembering to comment. Dont optimize code before its time. Is optimization worth the eort? If the code will soon be revised or extended, it will be rewritten anyway. Only optimize where necessary. Focus your eorts on bottlenecks.

The Proler

Matlab 5.0 (R10) and newer versions include a tool called the proler that helps identify bottlenecks in a program. Use it with the profile command: profile on Turn the proler on profile off Turn it back o profile clear Clear prole statistics profile report View the results from the proler For example, consider proling the following function: .....................................................................................................
function result = example1(Count) for k = 1:Count result(k) = sin(k/50); if result(k) < 0.9 result(k) = gammaln(k); end end

..................................................................................................... To analyze the eciency this function, rst enable and clear the proler, run the function, and then view the prole report:
>> profile on, profile clear >> example1(5000); >> profile report

There is some overhead when running code for the rst time; to remove this overhead from the timing, run example1 once before proling. The profiler report command shows a report. Depending on the system, proler results may dier from this example. .....................................................................................................

MATLAB Prole Report: Summary

Report generated 30-Jul-2004 16:57:01 Total recorded time: Number of M-functions: Clock precision: 3.09 s 4 0.016 s

Function List

Name example1 gammaln profile profreport

Time 3.09 0.73 0.00 0.00 100.0% 23.7% 0.0% 0.0%

Time 1 3562 1 1

Time/call 3.094000 0.000206 0.000000 0.000000

Self time 2.36 0.73 0.00 0.00 76.3% 23.7% 0.0% 0.0%

Location ~/example1.m ../toolbox/matlab/specfun/gammaln.m ../toolbox/matlab/general/profile.m ../toolbox/matlab/general/profreport.m

Clicking the example1 link gives more details: Lines where the most time was spent Line Number 4 7 6 Totals Code
result(k) = sin(k/50); result(k) = gammaln(k); if result(k) < -0.9

Calls 5000 721 5000

Total Time 2.11 s 0.84 s 0.14 s 3.09 s

% Time 68% 27% 5% 100%

The most time-consuming lines are displayed, along with time, time percentage, and line number. The most costly lines are the computations on lines 4 and 7. Another helpful section of the prole report is M-Lint Results, which gives analysis from the M-Lint code analyzer. Potential problems and suggested improvements are listed here. M-Lint results Line number Message 4 result might be growing inside a loop. Consider preallocating for speed. 7 result might be growing inside a loop. Consider preallocating for speed. For this example, M-Lint recommends using preallocation to improve lines 4 and 7 (preallocation is discussed in the next section). You can also run M-Lint on its own using the mlint command. The proler has limited time resolution, so to prole a piece of code that runs too quickly, run it multiple times with a loop:
>> profile on, profile clear >> for k = 1:100, example; end >> profile report

Adjust the number of loop iterations so that the time it takes to run is noticeable. More iterations yields better time resolution.

2
>> a = 2 a = 2

Array Preallocation

Matlabs matrix variables have the ability to dynamically augment rows and columns. For example,

>> a(2,6) = 1 a = 2 0 0 0 0 0 0 0 0 0 0 1

Matlab automatically resizes the matrix. Internally, the matrix data memory must be reallocated with larger size. If a matrix is resized repeatedlylike within a loopthis overhead can be signicant. To avoid frequent reallocations, preallocate the matrix with the zeros command. Consider the code: .....................................................................................................
a(1) = 1; b(1) = 0; for k = 2:8000 a(k) = 0.99803 * a(k 1) 0.06279 * b(k 1); b(k) = 0.06279 * a(k 1) + 0.99803 * b(k 1); end

..................................................................................................... The proler timed this code to take 0.47 seconds. After the for loop, both arrays are row vectors of length 8000, thus to preallocate, create empty a and b row vectors each with 8000 elements. .....................................................................................................
a = zeros(1,8000); b = zeros(1,8000); a(1) = 1; b(1) = 0; % Preallocation

for k = 2:8000 a(k) = 0.99803 * a(k 1) 0.06279 * b(k 1); b(k) = 0.06279 * a(k 1) + 0.99803 * b(k 1); end

..................................................................................................... With this modication, the code takes only 0.14 seconds (over three times faster). Preallocation is often easy to do, in this case it was only necessary to determine the right preallocation size and add two lines.

What if the nal array size can vary? Here is an example: .....................................................................................................
a = zeros(1,10000); count = 0; for k = 1:10000 v = exp(rand*rand); if v > 0.5 % Conditionally add to array count = count + 1; a(count) = v; end end a = a(1:count); % Trim the results to keep elements 1 to count % Preallocate

..................................................................................................... The average run time of this program is 0.42 seconds without preallocation and 0.18 seconds with it. Preallocation can also be done with cell arrays, using the cell command to create the desired size. Using preallocation with a frequently resizing cell array is even more benecial than with double arrays.

JIT Acceleration

Matlab 6.5 (R13) and later feature the Just-In-Time (JIT) Accelerator, improving the speed of M-code functions and scripts, particularly self-contained loops. By knowing a few things about the accelerator, you can improve its performance. The JIT Accelerator is enabled by default. To disable it, type feature accel off in the console, and feature accel on to enable it again. As of Matlab 7.5 (R2007b), only a subset of the Matlab language is supported for acceleration. Upon encountering an unsupported feature, acceleration processing is interrupted to handle it using nonaccelerated evaluation. The greatest acceleration improvements are gained when signicant contiguous portions of code are supported.
Data types: Code must use supported data types for acceleration: double (both real and complex), logical, char, int8, uint8, int16, uint16, int32, uint32. Code that uses other data types (for example, cell, struct, and function handles) is not accelerated. Sparse arrays are also not accelerated. Array shapes: Array shapes of any size with 3 or fewer dimensions are supported. Changing the shape or data type of an array interrupts acceleration. Function calls: Only calls to built-in functions are accelerated. Calling M-le and MEX functions interrupts acceleration. (See also page 13 on inlining simple functions.) Conditionals and loops: The conditional statements if, elseif, while, and switch are supported if the conditional expression evaluates to a scalar. Loops are accelerated if all the code within the loop is supported. Additionally, for statements must have scalar loop indices.

In future releases, JIT Accelerator may support more of the language and reduce these restrictions.

For example, the following code is supported for acceleration:

function B = bilateral(A,sd,sr,R) % The bilateral image denoising filter B = zeros(size(A)); for i = 1:size(A,1) for j = 1:size(A,2) zsum = 0; for m = R:R if i+m >= 1 && i+m <= size(A,1) for n = R:R if j+n >= 1 && j+n <= size(A,2) z = exp((A(i+m,j+n) A(i,j))2/(2*sd2))... * exp((m2 + n2)/(2*sr2)); zsum = zsum + z; B(i,j) = B(i,j) + z*A(i+m,j+n); end end end end B(i,j) = B(i,j)/zsum; end end

For a 128 128 input image and R = 3, the run time is 53.3 seconds without acceleration and 0.68 seconds with acceleration. Multithreaded Computation Matlab 7.4 (R2007a) introduced multithreaded computation for multi-core processors. Multithreaded computation accelerates some per-element functions when applied to large arrays (for example, .^, sin, exp) and certain linear algebra functions in the BLAS library. Multithreaded computation is disabled by default. To enable it, select File Preferences General Multithreading and select Enable multithreaded computation. As of Matlab 7.5 (R2007b), it seems that most Matlab code gains no improvement from multithreaded computation. In fact, some code is slightly slower with it enabled. (When comparing, beware that there is overhead when running code for the rst time; run the test code twice and time the second run.) Hopefully, the multithreaded computation feature will be developed further in future releases.

Vectorization

A computation is vectorized by taking advantage of vector operations. A variety of programming situations can be vectorized, and often improving speed to 10 times faster or even better. Vectorization is one of the most general and eective techniques for writing fast M-code.

4.1 Vectorized Computations

Many standard Matlab functions are vectorized: they can operate on an array as if the function had been applied individually to every element.
>> sqrt([1,4;9,16]) ans = 1 3 >> abs([0,1,2,5,6,7]) ans = 2 4 0 1 2 5 6 7

Consider the following function: .....................................................................................................

function d = minDistance(x,y,z) % Find the min distance between a set of points and the origin nPoints = length(x); d = zeros(nPoints,1);

% Preallocate

for k = 1:nPoints % Compute distance for every point d(k) = sqrt(x(k)2 + y(k)2 + z(k)2); end d = min(d); % Get the minimum distance

..................................................................................................... For every point, the distance between it and the origin is computed and stored in d. The minimum distance is then found with min. To vectorize the distance computation, replace the for loop with vector operations: .....................................................................................................
function d = minDistance(x,y,z) % Find the min distance between a set of points and the origin d = sqrt(x.2 + y.2 + z.2); d = min(d); % Compute distance for every point % Get the minimum distance

..................................................................................................... The modied code performs the distance computation with vector operations. The x, y and z arrays are rst squared using the per-element power operator, .^ (the per-element operators for multiplication and division are .* and ./). The squared components are added with vector addition. Finally, the 8

square root of the vector sum is computed per element, yielding an array of distances. (A further improvement: it is equivalent to compute d = sqrt(min(x.^2 + y.^2 + z.^2))). The rst version of the minDistance program takes 0.73 seconds on 50000 points. The vectorized version takes less than 0.04 seconds, more than 18 times faster. Some useful functions for vectorizing computations: min, max, repmat, meshgrid, sum, cumsum, diff, prod, cumprod, accumarray, filter

4.2 Vectorized Logic

The previous section shows how to vectorize pure computation. Bottleneck code often involves conditional logic. Like computations, Matlabs logic operators are vectorized:
>> [1,5,3] < [2,2,4] ans = 1

Two arrays are compared per-element. Logic operations return logical arrays with binary values. How is this useful? Matlab has a few powerful functions for operating on logical arrays:
>> find([1,5,3] < [2,2,4]) % Find indices of nonzero elements ans = 1

3 % True if any element of a vector is nonzero % (or percolumn for a matrix)

>> any([1,5,3] < [2,2,4]) ans = 1 >> all([1,5,3] < [2,2,4]) ans = 0

% True if all elements of a vector are nonzero % (or percolumn for a matrix)

Vectorized logic also works on arrays.

>> find(eye(3) == 1) ans = 1 5 9 % (element locations given as indices) % Find which elements == 1 in the 3x3 identity matrix

By default, find returns element locations as indices (see page 15 for indices vs. subscripts).

Example 1: Vector Normalization

To normalize a single vector v to unit length, one can use v = v/norm(v). However, to use norm for set of vectors v(:,1), v(:,2), . . . requires computing v(:,k)/norm(v(:,k)) in a loop. Alternatively,
vMag = sqrt(sum(v.2)); v = v./vMag(ones(1,size(v,1)),:);

The speed gain over a for loop is more signicant as the number of vectors increases and as the number of dimensions decreases. For several thousand vectors of length 3, the vectorized approach is around 10 times faster.

Example 2: Removing elements

The situation often arises where array elements must be removed on some per-element condition. For example, this code removes all NaN and innite elements from an array x:
i = find(isnan(x) | isinf(x)); x(i) = []; % Find bad elements in x % and delete them

Alternatively,
i = find(isnan(x) & isinf(x)); x = x(i); % Find elements that are not NaN and not infinite % Keep those elements

Both of these solutions can be further streamlined by using logical indexing:

x(isnan(x) | isinf(x)) = []; % Delete bad elements

or
x = x(isnan(x) & isinf(x)); % Keep good elements

Example 3: Piecewise functions

The sinc function has a piecewise denition, sinc(x) = sin(x)/x, 1, x=0 x=0

This code uses find with vectorized computation to handle the two cases separately: .....................................................................................................
function y = sinc(x) % Computes the sinc function perelement for a set of x values. y = ones(size(x)); i = find(x = 0); y(i) = sin(x(i)) ./ x(i); % Set y to all ones, sinc(0) = 1 % Find nonzero x values % Compute sinc where x = 0

..................................................................................................... An interesting alternative is y = (sin(x) + (x == 0))./(x + (x == 0)). 10

Example 4: Drawing images with meshgrid

The meshgrid function takes two input vectors and converts them to matrices by replicating the rst over the rows and the second over the columns.
>> [x,y] = meshgrid(1:5,1:3) x = 1 1 1 2 2 2 3 3 3 4 4 4 5 5 5 y = 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3

The matrices above work like a map for a width 5, height 3 image. For each pixel, the x-location can be read from x and the y-location from y. This may seem like a gratuitous use of memory as x and y simply record the column and row positions, but this is useful. For example, to draw an ellipse,
% Create x and y for a width 150, height 100 image [x,y] = meshgrid(1:150,1:100); % Ellipse with origin (60,50) of size 15 x 40 Img = sqrt(((x60).2 / 152) + ((y50).2 / 402)) > 1; % Plot the image imagesc(Img); colormap(copper); axis image, axis off

Drawing lines is just a change in the formula.

[x,y] = meshgrid(1:150,1:100); % The line y = x*0.8 + 20 Img = (abs((x*0.8 + 20) y) > 1); imagesc(Img); colormap(copper); axis image, axis off

Polar functions can be drawn by rst converting x and y variables with the cart2pol function.
[x,y] = meshgrid(1:150,1:100); [th,r] = cart2pol(x 75,y 50); % Spiral centered at (75,50) Img = sin(r/3 + th); imagesc(Img); colormap(hot); axis image, axis off

% Convert to polar

11

Example 5: Polynomial interpolation

Given n points x1 , x2 , . . . xn and corresponding function values y1 , y2 , . . . yn , the coecients c0 , c1 , . . . cn1 of the interpolating polynomial of degree n 1 can be found by solving cn1 x1 n1 x1 n2 x1 2 x1 1 y1 x2 n1 x2 n2 x2 2 x2 1 cn2 y2 . = . . . . . . . . . . . n1 n2 2 yn c0 xn xn xn xn 1 .....................................................................................................
function c = polyint(x,y) % Given a set of points and function values x and y, % computes the interpolating polynomial. x = x(:); y = y(:); n = length(x); % Construct the lefthand side matrix xMatrix = repmat(x, 1, n); powMatrix = repmat(n1:1:0, n, 1); A = xMatrix . powMatrix; c = A\y; % Make sure x and y are both column vectors % n = Number of points

% Make an n by n matrix with x on every column % Make another n by n matrix of exponents % Compute the powers % Solve matrix equation for coefficients

..................................................................................................... The strategy to construct the left-hand side matrix is to rst make two nn matrices of bases and exponents and then put them together using the per element power operator, .^ . The repmat function (replicate matrix) is used to make the base matrix xMatrix and the exponent matrix powMatrix. xMatrix = x(1) x(1) x(2) x(2) . . . x(n) x(n) x(1) x(2) . . . x(n) powMatrix = n1 n2 n1 n2 . . . n1 n2 0 0 . . . 0

The xMatrix is made by repeating the column vector x over the columns n times. The powMatrix is a row vector with elements n 1, n 2, n 3, . . . , 0 repeated down the rows n times. The two matrices could also have been created with [powMatrix, xMatrix] = meshgrid(n-1:-1:0, x). This function is only an example; use the standard polyfit function for serious use.

12

Inlining Simple Functions

Every time an M-le function is called, the Matlab interpreter incurs some overhead. Additionally, many M-le functions begin with conditional code that checks the input arguments for errors or determines the mode of operation. Of course, this overhead is negligible for a single function call. It should only be considered when the function being called is an M-le, the function itself is simple, that is, implemented with only a few lines, and called frequently from within a loop. For example, this code calls the M-le function median repeatedly to perform median ltering: .....................................................................................................
% Apply the median filter of size 5 to signal x y = zeros(size(x)); % Preallocate for k = 3:length(x)2 y(k) = median(x(k2:k+2)); end

..................................................................................................... Given a 2500-sample array for x, the overall run time is 0.42 seconds. Inlining a function means to replace a call to the function with the function code itself. Beware that inlining should not be confused with Matlabs inline function datatype. Studying median.m (type edit median on the console) reveals that most of the work is done using the built-in sort function. The median call can be inlined as .....................................................................................................
% Apply the median filter of size 5 to signal x y = zeros(size(x)); % Preallocate for k = 3:length(x)2 tmp = sort(x(k2:k+2)); y(k) = tmp(3); % y(k) = median(x(k2:k+2)); end

..................................................................................................... Now the overall run time is 0.047 seconds, nearly 9 times faster. Furthermore, by inlining median, it can be specically tailored to evaluating 5-sample medians. But this is only an example; if the Signal Processing Toolbox is available, y = medfilt1(x,5) is much faster. A surprising number of Matlabs functions are implemented as M-les, of which many can be inlined in just a few lines. In Matlab 5.3 (R11), the following functions are worth inlining: conv, cross, fft2, fliplr, flipud, ifft, ifft2, ifftn, ind2sub, ismember, linspace, logspace, mean, median, meshgrid, poly, polyval, repmat, roots, rot90, setdiff, setxor, sortrows, std, sub2ind, union, unique, var 13

The list above is for Matlab R11; some of these functions have since become built-in. Use the which command or try edit function name to determine whether a function is implemented as an M-le. For example, in Matlab R11 and earlier, ifft was implemented by simply calling fft and conj. If x is a one-dimensional array, y = ifft(x) can be inlined with y = conj(fft(conj(x)))/length(x). Another example: b = unique(a) can be inlined with
b = sort(a(:)); b( b((1:end1)') == b((2:end)') ) = [];

While repmat has the generality to operate on matrices, it is often only necessary to tile a vector or just a scalar. To repeat a column vector y over the columns n times,
A = y(:,ones(1,n)); % Equivalent to A = repmat(y,1,n);

To repeat a row vector x over the rows m times,

A = x(ones(1,m),:); % Equivalent to A = repmat(x,m,1);

To repeat a scalar s into an mn matrix,

A = s(ones(m,n)); % Equivalent to A = repmat(s,m,n);

This method avoids the overhead of calling an M-le function. It is never slower than repmat (critics should note that repmat.m itself uses this method to construct mind and nind). For constructing matrices with constant value, there are other ecient methods, for example, s+zeros(m,n). Warning: Dont go overboard. Inlining functions is only benecial when the function is simple and when it is called frequently. Doing it unnecessarily obfuscates the code.

14

Referencing Operations

Referencing in Matlab is varied and powerful enough to deserve a section of discussion. Good understanding of referencing enables vectorizing a broader range of programming situations.

Subscripts vs. Indices

Subscripts are the most common method used to refer to matrix elements, for example, A(3,9) refers to row 3, column 9. Indices are an alternative referencing method. Consider a 10 10 matrix A. Internally, Matlab stores the matrix data linearly as a one-dimensional, 100-element array of data in column-major order. 1 2 3 . . . 10 11 21 12 22 13 23 . . . . . . 20 30 81 91 82 92 83 93 . . . . . . 90 100

An index refers to an elements position in this one-dimensional array. Using an index, A(83) also refers to the element on row 3, column 9. It is faster to scan down columns than over rows. The column-major order means that elements of a column are closer together in memory than elements of a row, which promotes cache eciency. Conversion between subscripts and indices can be done with the sub2ind and ind2sub functions. However, because these are M-le functions rather than fast built-in operations, it is more ecient to compute conversions directly. For a two-dimensional matrix A of size MN, the conversions between subscript (i,j) and index (index) are A(i,j) A(i + (j-1)*M) A(index) A(rem(index-1,M)+1, floor((index-1)/M)+1) Indexing extends to N-D matrices as well, with indices increasing rst through the columns, then through the rows, through the third dimension, and so on. Subscript notation extends intuitively, A(. . . , dim4, dim3, row, col).

Vectorized Subscripts
It is useful to work with submatrices rather than individual elements. This is done with a vector of indices or subscripts. If A is a two-dimensional matrix, a vector subscript reference has the syntax A(rowv, colv) where rowv is a vector of rows with M elements and colv is a vector of columns with N elements. Both may be of any length and their elements may be in any order. If either is a matrix, it is reshaped to a vector. There is no dierence between using row vectors or column vectors in vector subscripts.

15

On the right-hand side of an operation (that is, on the right side of the =), a vector subscript reference returns a submatrix of elements of size MN: A(rowv(1), colv(1)) A(rowv(1), colv(2)) A(rowv(1), colv(3)) A(rowv(1), colv(N)) A(rowv(2), colv(1)) A(rowv(2), colv(2)) A(rowv(2), colv(3)) A(rowv(2), colv(N)) . . . . . . A(rowv(M), colv(1)) A(rowv(M), colv(2)) A(rowv(M), colv(3)) A(rowv(M), colv(N)) If the vector subscripted matrix is on the left-hand side, the right-hand side result must MN or scalar size. If any elements in the destination reference are repeated, for example, this ambiguous assignment of A(1,2) and A(2,2), A([1,2],[2,2]) = [1,2;3,4] A(1, 2) A(1, 2) A(2, 2) A(2, 2) = 1 3 2 4

it is the value in the source array with the greater index that dominates.
>> A = zeros(2); A([1,2],[2,2]) = [1,2;3,4] A = 0 0 2 4

Vector subscript references extend intuitively in higher dimensions.

Vector Indices
Multiple elements can also be referenced with vector indices. A(indexv) where indexv is an array of indices. On the right-hand side, a vector index reference returns a matrix the same size as indexv. For example, if indexv is a 3 4 matrix, A(indexv) is the 3 4 matrix A(indexv(1, 1)) A(indexv(1, 2)) A(indexv(1, 3)) A(indexv(1, 4)) A(indexv) = A(indexv(2, 1)) A(indexv(2, 2)) A(indexv(2, 3)) A(indexv(2, 4)) A(indexv(3, 1)) A(indexv(3, 2)) A(indexv(3, 3)) A(indexv(3, 4)) While vector subscripts are limited to referring to block-shaped submatrices, vector indices can refer to any shape. If a vector index reference is on the left-hand side, the right-hand side must return a matrix of the same size as indexv or a scalar. As with vector subscripts, ambiguous duplicate assignments use the value with greater source index.
>> A = zeros(2); A = 0 0 3 4 A([3,4,3,4]) = [1,2,3,4]

16

Reference Wildcards
Using the wildcard, :, in a subscript refers to an entire row or column. For example, A(:,1) refers to every row in column onethe entire rst column. This can be combined with vector subscripts, A([2,4],:) refers to the second and fourth rows. When the wildcard is used in a vector index, the entire matrix is referenced. On the right-hand side, this always returns a column vector. A(:) = column vector This is frequently useful: for example, if a function input must be a row vector, the users input can be quickly reshaped into row vector with A(:). (make a column vector and transpose to a row vector). A(:) on the left-hand side assigns all the elements of A, but does not change its size. For example, A(:) = 8 changes all elements of matrix A to 8.

Logical Indexing
Given a logical array mask with the same size as A,
A(mask)

refers to all elements of A where the corresponding element of mask is 1 (true). It is equivalent to A(find(mask)). A common usage of logical indexing is to refer to elements based on a per-element condition, for example,
A(abs(A) < 1e3) = 0

sets all elements with magnitude less than 103 to zero. Logical indexing is also useful to select nonrectangular sets of elements, for example,
A(logical(eye(size(A))))

references the diagonal elements of A. Note that for a right-hand side reference, it is faster to use diag(A) to get the diagonal of A.

Deleting Submatrices with [ ]

Elements in a matrix can be deleted by assigning the empty matrix. For example, A([3,5]) = [ ] deletes the third and fth element from A. If this is done with index references, the matrix is reshaped into a row vector. It is also possible to delete with subscripts if all but one subscript are the wildcard. A(2,:) deletes the second row. Deletions like A(2,1) = [ ] or even A(2,1:end) = [ ] are illegal. = []

17

Numerical Integration
b

Numerical approximation of integrals is usually done with quadrature formulas, which have the form f (x) dx
a k

wk f (xk ),

where the xk are called the nodes or abscissas and wk are the associated weights. Simpsons rule is
b

f (x) dx
a

h [f (a) + 4f (a + h) + f (b)] , 3

h=

ba . 2
h 4h h 3, 3 , 3.

Simpsons rule is a quadrature formula with nodes a, a + h, b and node weights

Matlab oers quad and quadl for quadrature in one dimension. These functions are robust and precise, however, they are not very ecient. Both use an adaptive renement procedure to reduce the number of function calls. However, as quad and quadl are recursive M-le functions, the algorithmic overhead is signicant. Furthermore, the adaptive algorithm gains little from vectorization. If an application requires approximating an integral whose integrand can be eciently vectorized, using nonadaptive quadrature may improve speed. .....................................................................................................
% Approximate Fourier series coefficients for exp(sin(x)6) for frequencies 20 to 20 for n = 20:20 c(n + 21) = quad(inline('exp(sin(x).6).*exp(i*x*n)','x','n'),0,pi,1e4,[],n); end

..................................................................................................... This code runs in 5.16 seconds. In place of quad, using Simpsons composite rule with N = 199 nodes yields results with comparable accuracy and allows for vectorized computation. Since the integrals are all over the same interval, the nodes and weights only need to be constructed once. .....................................................................................................
N = 199; h = pi/(N1); x = (0:h:pi).'; w = ones(1,N); w(2:2:N1) = 4; % Nodes % Weights

w(3:2:N2) = 2;

w = w*h/3;

for n = 20:20 c(n + 21) = w * ( exp(sin(x).6).*exp(i*x*n) ); end

..................................................................................................... This version of the code runs in 0.02 seconds (200 times faster). The quadrature is performed by the dot product multiplication with w. It can be further optimized by replacing the for loop with one vector-matrix multiply:
[n,x] = meshgrid(20:20, 0:h:pi); c = w * ( exp(sin(x).6).*exp(i*x.*n) );

18

7.1 One-Dimensional Integration

b a

f (x) dx is approximated by composite Simpsons rule with

= = = = (b a)/(N1); (a:h:b).'; ones(1,N); w(2:2:N1) = 4; w(3:2:N2) = 2; w = w*h/3; w * f(x); % Approximately evaluate the integral

h x w I

where N is an odd integer specifying the number of nodes. A good higher-order choice is composite 4th -order Gauss-Lobatto [3], based on the approximation
1 1 1 1 1 f (x) dx 6 f (1) + 5 f ( 5 ) + 5 f ( 5 ) + 1 f (1). 6 6 6

N h d x w I

= = = = = =

max(3*round((N1)/3),3) + 1; % Adjust N to the closest valid choice (b a)/(N1); (3/sqrt(5) 1)*h/2; (a:h:b).'; x(2:3:N2) = x(2:3:N2) d; x(3:3:N1) = x(3:3:N1) + d; ones(1,N); w(4:3:N3) = 2; w([2:3:N2,3:3:N1]) = 5; w = w*h/4; w * f(x); % Approximately evaluate the integral

The number of nodes N must be such that (N 1)/3 is an integer. If not, the rst line adjusts N to the closest valid choice. It is usually more accurate than Simpsons rule when f has six continuous derivatives, f C 6 (a, b). A disadvantage of this nonadaptive approach is that the accuracy of the result is only indirectly controlled by the parameter N. To guarantee a desired accuracy, either use a generously large value for N or if possible determine the error bounds [5, 6] Simpsons rule error 4 th -order Gauss-Lobatto error (b a)h4 max f (4) () provided f C 4 (a, b), ab 180 27(b a)h6 max f (6) () provided f C 6 (a, b), ab 56000

ba where h = N 1 . Note that these bounds are valid only when the integrand is suciently dierentiable: f must have four continuous derivatives for the Simpsons rule error bound, and six continuous derivatives for Gauss-Lobatto.

For most purposes, composite Simpsons rule is a sucient default choice. Depending on the integrand, other choices can improve accuracy:
Use higher-order quadrature formulas if the integrand has many continuous derivatives. Use lower-order if the integrand function is not smooth. Use the substitution u =
1 1x

or Gauss-Laguerre quadrature for innite-domain integrals like

. 0

19

7.2 Multidimensional Integration

An approach for evaluating double integrals of the form a c f (x, y) dy dx is to apply one-dimensional b quadrature to the outer integral a F (x) dx and then for each x use one-dimensional quadrature over d the inner dimension to approximate F (x) = c f (x, y) dy. The following code does this with composite Simpsons rule with NxNy nodes:
%%% Construct Simpson nodes and weights over x %%% h = (b a)/(Nx1); x = (a:h:b).'; wx = ones(1,Nx); wx(2:2:Nx1) = 4; wx(3:2:Nx2) = 2; %%% Construct Simpson nodes and weights over y %%% h = (d c)/(Ny1); y = (c:h:d).'; wy = ones(1,Ny); wy(2:2:Ny1) = 4; wy(3:2:Ny2) = 2; %%% Combine for twodimensional integration %%% [x,y] = meshgrid(x,y); x = x(:); y = y(:); w = wy.'*wx; w = w(:).'; I = w * f(x,y); % Approximately evaluate the integral
b d

wx = w*h/3;

wy = w*h/3;

Similarly for three-dimensional integrals, the weights are combined with

[x,y,z] = meshgrid(x,y,z); w = wy.'*wx; w = w(:)*wz; x = x(:); y = y(:); w = w(:).'; z = z(:);

When the integration region is complicated or of high dimension, Monte Carlo integration techniques are appropriate. The disadvantage is that an N -point Monte Carlo quadrature has error on the order O( 1 ), so many points are necessary even for moderate accuracy. Nevertheless, the basic Monte Carlo N idea is straightforward and of practical value. Suppose that N points, x1 , x2 , . . . , xN , are uniformly randomly selected in a multidimensional volume . Then f dV

dV N

f (xn ).
n=1

To integrate a complicated volume W that is dicult to sample uniformly, nd an easier volume that contains W and can be sampled [4]. Then f dV =
W

f W dV

dV N

f (xn )W (xn ),
n=1

W (x) =

1, 0,

x W, x W. /

W (x) is the indicator function of W : W (x) = 1 when x is within volume W and W (x) = 0 otherwise. Multiplying the integrand by W sets contributions from outside of W to zero. For example, consider nding the center of mass of the shape W dened by cos 2 and x2 + y 2 4. Given the integrals M =
W

x2 + y 2 x y

dA, Mx = 20

x dA, and My =

y dA, the center of

mass is ( Mx , My ). The region is contained in the rectangle dened by 2 x 2 and 2 y 2. M The following code estimates M , Mx , and My :
%%% Uniformly randomly sample points (x,y) in %%% x = 4*rand(N,1) 2; y = 4*rand(N,1) 2; %%% Restrict the points to region W %%% i = (cos(2*sqrt(x.2 + y.2)).*x <= y & x.2 + y.2 <= 4); x = x(i); y = y(i); %%% Approximately evaluate the integrals %%% area = 4*4; % The area of rectangle M = (area/N) * length(x); Mx = (area/N) * sum(x); My = (area/N) * sum(y);
2

-2 -2

Region W sampled with N = 1500. The center of mass is (0.5, 0.7).

More generally, if W is a two-dimensional region contained in the rectangle dened by a x b and c y d, the following code approximates W f dA:
x y i x = = = = a + (ba)*rand(N,1); c + (dc)*rand(N,1); logical(indicatorW(x,y)); x(i); y = y(i);

area = (ba)*(dc); I = (area/N) * sum(f(x,y));

% Approximately evaluate the integral

where indicatorW(x,y) is the indicator function W (x, y) for region W . For renements and variations of Monte Carlo integration, see for example [1].

21

Signal Processing

Even without the Signal Processing Toolbox, Matlab is quite capable in signal processing computations. This section lists code snippets to perform several common operations eciently.

Locating zero-crossings and extrema

To obtain the indices where signal x crosses zero:
i = find(diff(sign(x))); % The kth zerocrossing lies between x(i(k)) and x(i(k)+1)

Linear interpolation can be used for subsample estimates of zero-crossing locations:

i = find(diff(sign(x))); i = i x(i)./(x(i+1) x(i)); % Linear interpolation

Since local maximum and minimum points of a signal have zero derivative, their locations can be estimated from the zero-crossings of diff(x), provided the signal is sampled with suciently ne resolution. For a coarsely sampled signal, a better estimate is
iMax = find(sign(x(2:end1)x(1:end2)) + sign(x(2:end1)x(3:end)) > iMin = find(sign(x(2:end1)x(1:end2)) + sign(x(2:end1)x(3:end)) < ... 0) + 1; ... 0) + 1;

Moving-average lter
To compute an N-sample moving average of x with zero padding:
y = filter(ones(N,1)/N,1,x);

For large N, it is faster to use

y = cumsum(x)/N; y(N+1:end) = y(N+1:end) y(1:endN);

FFT-based convolution
This line performs FFT-based circular convolution, equivalent to y = filter(b,1,x) except near the boundaries, provided that length(b) < length(x):
y = ifft(fft(b,length(x)).*fft(x));

For FFT-based zero-padded convolution, equivalent to y = filter(b,1,x),

N = length(x)+length(b)1; y = ifft(fft(b,N).*fft(x,N)); y = y(1:length(x));

In both code snippets above, y will be complex-valued, even if x and b are both real. To force y to be real, follow the computation with y = real(y). If you have the Signal Processing Toolbox, it is faster to use fftfilt for FFT-based, zero-padded ltering. 22

Noncausal ltering and other boundary extensions

For its intended use, the filter command is limited to causal lters, that is, lters that do not involve future values to the right of the current sample. Furthermore, filter is limited to zero-padded boundary extension; data beyond the array boundaries are assumed zero as if the data were padded with zeros. For two-tap lters, noncausal ltering and other boundary extensions are possible through filters fourth initial condition argument. Given a boundary extension value padLeft for x(0), the lter yn = b1 xn + b2 xn1 may be implemented as
y = filter(b,1,x,padLeft*b(2));

Similarly, given a boundary extension value padRight for x(end+1), the lter yn = b1 xn+1 + b2 xn is implemented as
y = filter(b,1,[x(2:end),padRight],x(1)*b(2));

Choices for padLeft and padRight for various boundary extensions are Boundary extension Periodic Whole-sample symmetric Half-sample symmetric Antisymmetric padLeft x(end) x(2) x(1) 2*x(1)-x(2) padRight x(1) x(end-1) x(end) 2*x(end)-x(end-1)

It is in principle possible to use a similar approach for longer lters, but ironically, computing the initial conditions itself requires ltering. To implement noncausal ltering and ltering with other boundary handling, it is usually fastest to pad the input signal, apply filter, and then truncate the result.

Upsampling and Downsampling

Upsample x (insert zeros) by factor p:
y = zeros(length(x)*pp+1,1); y(1:p:length(x)*p) = x; % For trailing zeros, use y = zeros(length(x)*p,1);

Downsample x by factor p, where 1 q p:

y = x(q:p:end);

23

Haar Wavelet
This code performs K stages of the Haar wavelet transform on the input data x to produce wavelet coecients y. The input array x must have length divisible by 2K . Forward transform:
y = x(:); N = length(y); for k = 1:K tmp = y(1:2:N) + y(2:2:N); y([1:N/2,N/2+1:N]) = ... [tmp;y(1:2:N) 0.5*tmp]/sqrt(2); N = N/2; end

Inverse transform:
x = y(:); N = length(x)*pow2(K); for k = 1:K N = N*2; tmp = x(N/2+1:N) + 0.5*x(1:N/2); x([1:2:N,2:2:N]) = ... [tmp;x(1:N/2) tmp]*sqrt(2); end

Daubechies 4-Tap Wavelet

This code implements the Daubechies 4-tap wavelet in lifting scheme form [2]. The input array x must have length divisible by 2K . Filtering is done with symmetric boundary handling. Forward transform:
U = 0.25*[2sqrt(3),sqrt(3)]; ScaleS = (sqrt(3) 1)/sqrt(2); ScaleD = (sqrt(3) + 1)/sqrt(2); y = x(:); N = length(y); for k = 1:K N = N/2; y1 = y(1:2:2*N); y2 = y(2:2:2*N) + sqrt(3)*y1; y1 = y1 + filter(U,1,... y2([2:N,max(N1,1)]),y2(1)*U(2)); y(1:2*N) = [ScaleS*... (y2 y1([min(2,N),1:N1]));ScaleD*y1]; end

Inverse transform:
U = 0.25*[2sqrt(3),sqrt(3)]; ScaleS = (sqrt(3) 1)/sqrt(2); ScaleD = (sqrt(3) + 1)/sqrt(2); x = y(:); N = length(x)*pow2(K); for k y1 y2 y1 = 1:K = x(N+1:2*N)/ScaleD; = x(1:N)/ScaleS + y1([min(2,N),1:N1]); = y1 filter(U,1,... y2([2:N,max(N1,1)]),y2(1)*U(2)); x([1:2:2*N,2:2:2*N]) = [y1;y2 sqrt(3)*y1]; N = 2*N; end

To use periodic boundary handling rather than symmetric boundary handling, change appearances of [2:N,max(N-1,1)] to [2:N,1] and [min(2,N),1:N-1] to [N,1:N-1].

24

Miscellaneous Tricks

9.1 Clip a value without using if statements

To clip (or clamp, bound, or saturate) a value to within a range, the straightforward implementation is
if x < x = elseif x = end lowerBound lowerBound; x > upperBound upperBound;

Unfortunately, this is slow. A faster method is to use the min and max functions:
x = max(x,lowerBound); x = min(x,upperBound); % Clip elements from below, force x >= lowerBound % Clip elements from above, force x <= upperBound

Moreover, it works per-element if x a matrix of any size.

9.2 Convert any array into a column vector

It is often useful to force an array to be a column vector, for example, when writing a function expecting a column vector as an input. This simple trick will convert the input array of any shape and number of dimensions to a column vector, even if the input is already a column vector.
x = x(:); % convert x to a column vector

By following this operation with a transpose ., it is possible to convert an array to a row vector.

9.3 Find the min/max of a matrix or N-d array

Given a matrix input (with no other inputs), the min and max functions operate along the columns, nding the extreme element in each column. To nd the extreme element over the entire matrix, one way for a two-dimensional matrix is min(min(A)). A better method that works regardless of the number of dimensions and determines the extreme elements location is
% Find the minimum element in A % The minimum value is MinValue, the index is MinIndex MinSub = ind2sub(size(A), MinIndex); % Convert MinIndex to subscripts [MinValue, MinIndex] = min(A(:));

The minimum element is A(MinIndex) or A(MinSub(1), MinSub(2), . . .) as a subscript reference. (Similarly, replace min with max for maximum value.)

25

9.4 Flood lling

Flood lling, like the bucket tool in image editors, can be elegantly written as a recursive function: .....................................................................................................
function I = flood1(I,c,x,y) % Flood fills image I from point (x,y) with color c. c2 = I(y,x); I(y,x) = c; if if if if x x y y > < > < 1 size(I,2) 1 size(I,1) & & & & I(y,x1) I(y,x+1) I(y1,x) I(y+1,x) == == == == c2 c2 c2 c2 & & & & I(y,x1) I(y,x+1) I(y1,x) I(y+1,x) = = = = c, c, c, c, I I I I = = = = flood1(I,c,x1,y); flood1(I,c,x+1,y); flood1(I,c,x,y1); flood1(I,c,x,y+1); end end end end

..................................................................................................... Being a highly recursive function, this is inecient in Matlab. The following code is faster: .....................................................................................................
function I = flood2(I,c,x,y) % Flood fills image I from point (x,y) with color c. LastFlood = zeros(size(I)); Flood = LastFlood; Flood(y,x) = 1; Mask = (I == I(y,x)); FloodFilter = [0,1,0; 1,1,1; 0,1,0]; while any(LastFlood(:) = Flood(:)) LastFlood = Flood; Flood = conv2(Flood,FloodFilter,'same') & Mask; end I(find(Flood)) = c;

..................................................................................................... The key is the conv2 two-dimensional convolution function. Flood lling a 4040-pixel region takes 1.168 seconds with flood1 and 0.067 seconds with flood2.

9.5 Vectorized use of set on GUI objects

While not a speed improvement, a trick for cleaner code is vectorized use of set on GUI objects. A serious graphical user interface (GUI) can have dozens of objects like text labels, buttons, and sliders. For example, to dene three edit boxes with white text background and left text alignment:
uicontrol('Units', 'normalized', 'Position', [0.1,0.9,0.7,0.05], ... 'HorizontalAlignment', 'left', 'Style', 'edit', 'BackgroundColor', [1,1,1]); uicontrol('Units', 'normalized', 'Position', [0.1,0.8,0.7,0.05], ... 'HorizontalAlignment', 'left', 'Style', 'edit', 'BackgroundColor', [1,1,1]); uicontrol('Units', 'normalized', 'Position', [0.1,0.7,0.7,0.05], ... 'HorizontalAlignment', 'left', 'Style', 'edit', 'BackgroundColor', [1,1,1]);

26

A vectorized call to set reduces this to

h(1) = uicontrol('Units', 'normalized', 'Position', [0.1,0.9,0.7,0.05]); h(2) = uicontrol('Units', 'normalized', 'Position', [0.1,0.8,0.7,0.05]); h(3) = uicontrol('Units', 'normalized', 'Position', [0.1,0.7,0.7,0.05]); set(h, 'HorizontalAlignment', 'left', 'Style', 'edit','BackgroundColor', [1,1,1]);

10

Further Reading

For more reading on vectorization, see the MathWorks vectorization guide: http://www.mathworks.com/support/tech-notes/1100/1109.shtml For further details and examples on JIT Acceleration, see http://www.mathworks.com/access/helpdesk_r13/help/techdoc/matlab_prog/ch7_perf.html For a thorough guide on ecient array manipulation, see MATLAB array manipulation tips and tricks: http://home.online.no/~pjacklam/matlab/doc/mtt For numerical methods with Matlab, see Numerical Computing with MATLAB : http://www.mathworks.com/moler

MEX
In a coding situation that cannot be optimized any further, keep in mind that the Matlab language is intended primarily for easy prototyping rather than large-scale high-speed computation. In some cases, an appropriate solution is Matlab Executable (MEX) external interface functions. With a C or Fortran compiler, it is possible to produce fast MEX functions that can be called within Matlab just like M-le functions. The typical speed improvement over equivalent M-code is easily ten-fold. Installations of Matlab include an example of MEX under directory <MATLAB>/extern/examples/mex. In this directory, a function yprime is written as M-code (yprime.m) and equivalent C (yprime.c) and Fortran (yprime.f) MEX source code. For more information, see the MathWorks MEX-les guide http://www.mathworks.com/support/tech-notes/1600/1605.html For information on compiling MEX les with the GNU compiler collection (GCC), see http://gnumex.sourceforge.net Beware, MEX is an ambitious alternative to M-code. Writing MEX-les requires solid understanding of Matlab and comfort with C or Fortran, and takes more time to develop than M-code.

27

Matlab Compiler
The Matlab Compiler generates C and C++ code from M-code, producing MEX-les and standalone executables. For product information and documentation, see http://www.mathworks.com/access/helpdesk/help/toolbox/compiler

References
[1] A. Bielajew. The Fundamentals of the Monte Carlo Method for Neutral and Charged Particle Transport. University of Michigan, class notes. Available online at http://www-personal.engin.umich.edu/~bielajew/MCBook/book.pdf [2] I. Daubechies and W. Sweldens. Factoring Wavelet Transforms into Lifting Steps. Sep. 1996. [3] W. Gander and W. Gautschi. Adaptive QuadratureRevisited, BIT, vol. 40, pp. 84-101, 2000. [4] W. Press, B. Flannery, S. Teukolsky and W. Vetterling. Numerical Recipes. Cambridge University Press, 1986. [5] E. Weisstein. Lobatto Quadrature. From MathWorldA Wolfram Web Resource. http://mathworld.wolfram.com/LobattoQuadrature.html [6] E. Weisstein. Simpsons Rule. From MathWorldA Wolfram Web Resource. http://mathworld.wolfram.com/SimpsonsRule.html

28