1 © 2012 The MathWorks, Inc. Speeding up MATLAB Applications

2 Agenda  Leveraging the power of vector and matrix operations in MATLAB ® - Demonstration: Preallocation and vectorization  How does MATLAB ® store and provide access to its’ variables? - Demonstration: Row- vs. Column-major indexing - Demonstration: Logical indexing  Addressing bottlenecks - Demonstration: Profiling code in MATLAB ®  Memory considerations - Demonstration: Copy-on-write  Summary and Q&A

3 Example: Block Processing Images  Evaluate function at grid points  Reevaluate function over larger blocks  Compare the results  Evaluate code performance

4 Summary of Example  Used built-in timing functions >> tic >> toc  Used Code Analyzer to find suboptimal code  Preallocated arrays  Vectorized code

5 Effect of Not Preallocating Memory >> x = 4 >> x(2) = 7 >> x(3) = 12 0x0000 0x0008 0x0010 0x0018 0x0020 0x0028 0x0030 0x0038 0x0000 0x0008 0x0010 0x0018 0x0020 0x0028 0x0030 0x0038 0x0000 0x0008 0x0010 0x0018 0x0020 0x0028 0x0030 0x X(3) = 12X(2) = 7 Resizing Arrays is Expensive

6 Benefit of Preallocation >> x = zeros(3,1) >> x(1) = 4 >> x(2) = 7 >> x(3) = 12 0x0000 0x0008 0x0010 0x0018 0x0020 0x0028 0x0030 0x x0000 0x0008 0x0010 0x0018 0x0020 0x0028 0x0030 0x x0000 0x0008 0x0010 0x0018 0x0020 0x0028 0x0030 0x0038 0x0000 0x0008 0x0010 0x0018 0x0020 0x0028 0x0030 0x Reduced Memory Operations

7 Agenda  Leveraging the power of vector and matrix operations in MATLAB ® - Demonstration: Preallocation and vectorization  How does MATLAB ® store and provide access to its’ variables? - Demonstration: Row- vs. Column-major indexing - Demonstration: Logical indexing  Addressing bottlenecks - Demonstration: Profiling code in MATLAB ®  Memory considerations - Demonstration: Copy-on-write  Summary and Q&A

8 Data Storage of MATLAB Arrays What is the fastest way to process MATLAB matrices with for loops? (i.e. de-vectorize) a) Down the columns b) Along the rows c) Doesn't matter 0x0000 0x0008 0x0010 0x0018 0x0020 0x0028 0x0030 0x0038 0x0040 0x0048 0x0050 0x0058 0x0060 0x0068 See the June 2007 article in “The MathWorks News and Notes”: >> x = magic(3) x = Column-Major Memory Storage

9 Indexing into MATLAB Arrays  Subscripted –Access elements by rows and columns  Linear –Access elements with a single number  Logical –Access elements with logical operations or mask ,11,21,3 2,12,22,3 3,13,23,3 Linear indexing Subscripted indexing ind2sub sub2ind

10 Speed and Memory Usage  Balance vectorization and memory usage –Use bsxfun instead of functions such as repmat –Reduce size of arrays to smaller blocks for block processing  Consider using sparse matrices –Less Memory: Store only nonzero elements and their indices –Faster: Eliminate operations on zero elements –Blog Post - Creating Sparse Finite Element Matrices element-matrices-in-matlab/ element-matrices-in-matlab/

11 MATLAB Underlying Technologies  Commercial libraries –BLAS:Basic Linear Algebra Subroutines (multithreaded) –LAPACK: Linear Algebra Package –etc.  JIT/Accelerator –Improves looping –Generates on-the-fly multithreaded code –Continually improving

12 Other Best Practices  Minimize dynamically changing path >> addpath(…) >> fullfile(…)  Use the functional load syntax >> x = load('myvars.mat') x = a: 5 b: 'hello'  Minimize changing variable class >> x = 1; >> xnew = 'hello'; instead of cd(…) instead of load('myvars.mat') instead of x = 'hello';

13 Agenda  Leveraging the power of vector and matrix operations in MATLAB ® - Demonstration: Preallocation and vectorization  How does MATLAB ® store and provide access to its’ variables? - Demonstration: Row- vs. Column-major indexing - Demonstration: Logical indexing  Addressing bottlenecks - Demonstration: Profiling code in MATLAB ®  Memory considerations - Demonstration: Copy-on-write  Summary and Q&A

14 Example: Fitting Data  Load data from multiple files  Extract a specific test  Fit a spline to the data  Write results to Microsoft Excel

15 Summary of Example  Used profiler to analyze code  Targeted significant bottlenecks  Reduced file I/O  Reused figure

16 Acceleration using MEX  Call C or Fortran code directly from MATLAB –Integrate existing code using MEX API –Auto-generate C-based MEX files from MATLAB code using MATLAB Coder  Speed-up factor will vary –May see speedup for state-based for-loops –May not see a speedup when MATLAB code is  Using multithreaded computations  Using optimized libraries (BLAS, FFTW, etc.) c = myFcn(a,b) myFcn.c void mexFunction( int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[]) { /* more C code... */ }

17 Agenda  Leveraging the power of vector and matrix operations in MATLAB ® - Demonstration: Preallocation and vectorization  How does MATLAB ® store and provide access to its’ variables? - Demonstration: Row- vs. Column-major indexing - Demonstration: Logical indexing  Addressing bottlenecks - Demonstration: Profiling code in MATLAB ®  Memory considerations - Demonstration: Copy-on-write  Summary and Q&A

18 MATLAB and Memory Lazy copying (copy-on-write): C{1} = rand(3000,3000); %72Mb C{2} = rand(3000,3000); %72Mb C_new = C; C{1}(1,1) = 2; Arrays are stored in contiguous memory %72Mb %0 Mb

19 In-place Operations When does MATLAB do calculations “in-place”? function y = myfunc(x) y = sin(2*x.^2+3*x+4); function x = myfuncInPlace(x) x = sin(2*x.^2+3*x+4); x = randn(3e7,1); y = myfunc(x); x = myfuncInPlace(x); % In-place x = myfunc(x); % Copy

20 Agenda  Leveraging the power of vector and matrix operations in MATLAB ® - Demonstration: Preallocation and vectorization  How does MATLAB ® store and provide access to its’ variables? - Demonstration: Row- vs. Column-major indexing - Demonstration: Logical indexing  Addressing bottlenecks - Demonstration: Profiling code in MATLAB ®  Memory considerations - Demonstration: Copy-on-write  Summary and Q&A

21 Summary  Consider performance benefit of vector and matrix operations in MATLAB  Preallocate memory and look at your indexing.  Analyze your code for bottlenecks and address most critical items  Never write hard-to-read code unless you have to.

22 Summary (continued…)  MATLAB passes data to functions by value using lazy copy.

23 Sample of Other Performance Resources  MATLAB documentation MATLAB  Programming Fundamentals  Performance  Memory Management Guide  The Art of MATLAB, Loren Shure’s blog blogs.mathworks.com/loren/  MATLAB Answers