Speed up Matlab Xin Sheng Zhou Matlab MATrix LABoratory Started - PowerPoint PPT Presentation

Speed up Matlab Xin Sheng Zhou

Matlab  MATrix LABoratory  Started in the late 1970s  Matrix manipulation  Plotting of data  Implementation of algorithms  Creation of user interface  Interfacing with other languages, such as C and Fortran  Slower, but reducing programming effort

Java  Java based  Appeared in 1995  Run on Java Virtual Machine  Compile once, run everywhere  Inherited slower than C/C++  Currently (February 2012), microbenchmarks show Java 7 is approximately 1.5 times slower than C

Code 1

Preallocating arrays

Preallocating arrays  Looking for larger memory  Moving data

Preallocating arrays  Looking for larger memory  Moving data

Preallocating arrays

Preallocating a non-double matrix  A = int8(zeros(100));  A = zeros(100, 'int8');

Changing a variable’s data type or dimension  Another case: assigning real and complex numbers

Using appropriate logical operators  Large if…else… block, consider switch…case…

Code 2

Vectorizing loops

Profiling  Measure where a program spends time

Case study  Low density parity check decoder  Block decoding  Block length: 10 6  Generate parity check matrix  Decoding

Resource used for block length 10^6 and 10^7

LDPC Code  H*c T =0  For code ½, block length 10 6  H: Matrix (0.5*10 6 ,10 6 )  C: Row vector, size 10 6  Encoding: c=s*G  s: row vector, size 0.5*10 6  G: matrix, (0.5*10 6 ,10 6 )  Get G from H is not easy  Instead, generate parity check bits p=A -1 Bs  H=[A B], c=[p s]  Matlab’s matrix multiplication and inversion cannot be used since it is based on double value and encoding is based on GF(2)  Solution: Custom matrix inversion and multiplication on GF(2)  Inversion is slow, and A -1 is required to be saved  When block length > 10 5 , saving is also slow  Since H is sparse, sparse matrix format should be used. However, A -1 is not sparse

Other ways…  Parallel computing  Multi-threaded  GPU  Multi-machine  MEX file  Matlab Compiler

Summary  Preallocating arrays  Vectorizing loops  Profiling

Read more …