Embedded Systems Design Europe - February 2008 - (Page 33)

feature most of your time is spent. As an example, the screen shots above show the MATLAB proﬁler report for a block-matching algorithm used for motion estimation. Looking at the proﬁle, it is quickly evident that most of the run-time is spent in the costFunctionMAD function. As one might expect, a look inside this function reveals that the hot-spot is the computation of mean absolute difference between two blocks. Once your performance bottlenecks are identiﬁed, you want to carefully review that code and look into ways of optimizing it. There are several common techniques at your disposal for improving MATLAB performance. These include: 1. pre-deﬁning your arrays, 2. vectorizing your code, and 3. using compiled MEX-ﬁles. These techniques are applicable only for code you have written. There are no general techniques to accelerate MATLAB’s core functions such as ﬁlter and fft. These core MATLAB functions are considered “built-in” functions. They are implemented using fast compiled libraries (analogous to MEX-ﬁles) and are already highly optimized. When arrays are growing inside of a loop, unnecessary time is spent allocating memory and copying data. In general you can save considerable time by pre-allocating memory. This can be done by adding calls to the zeros functions before your loops. It is also possible to allocate structure arrays using the struct function in conjunction with the repmat function. The universal rule for writing fast MATLAB is to vectorize your code. This is done by replacing loops and scalar operations by vectored operations. Several techniques for vectorizing your code can be found in The Mathworks’ code vectorization guide. In the block-matching algorithm presented above, the loop that computes the mean absolute difference computation can be vectorized into one line using the sum function. The transformation shown on page 34 speeds up the overall run time by almost 2x. Vectorization lends itself well to problems that are fairly regular, that is, when the same computation can be consistently applied to large sets of data. Vectorization does not work well for algorithms that have an irregular structure or have many exceptions. In these cases vectorization is often cumbersome and non-intuitive. It can also force you to write less readable code, which makes your algorithm harder to understand, maintain, and debug. In some cases, vectorization may not be possible or may actually slow your code www.embedded.com/europe | embedded systems design europe | JANUARY – FEBRUARY 2008 33 032-033-034_ESDE.indd 33 8/02/08 10:37:40 http://www.dvsinc.com http://www.dvsinc.com http://www.embedded.com/europe

Table of Contents Feed for the Digital Edition of Embedded Systems Design Europe - February 2008

Embedded Systems Design Europe - February 2008 Contents ARTEMIS and ENIAC Get Parlimentary Approval Product Teardown Videos Come On Screen Esterel and Abslnt Integrate Products Microsoft Opens Windows to Networked Embedded Applications Trango Embeds Virtualization Tool in Cavium's Multicore CPUs MindTree - ADI Develop Security DVR Platform NXP Extends Deal with ARM to Cover MCUs Automotive and Embedded to Dominate DATE 08 ZigBee Spec Gets Smart On Energy Updated Card Spec Provides for Power-On Boot The Basics of Embedded Multitasking On a PIC Cover Feature: The Art of FPGA Construction Is Symmetric Multiprocessing For You? Accelerating MATLAB Using MEX-Files ARM Provides the Microcontroller Solution Embedded World Advertising Contracts

Embedded Systems Design Europe - February 2008

For optimal viewing of this digital publication, please enable JavaScript and then refresh the page. If you would like to try to load the digital publication without using Flash Player detection, please click here.