Difference between revisions of "Tips and tricks"
m (→Changing BLAS: Should not write as if all Linux workstations used apt-get) |
|||
Line 176: | Line 176: | ||
===Changing BLAS=== | ===Changing BLAS=== | ||
− | |||
Many Octave functions are wrappers to optimized numerical libraries, notably BLAS and ATLAS. It is possible to achieve impressive performance gains by simply using a library tuned to your platform. One example is using OpenBLAS to replace the default BLAS implementation ([http://www.stat.cmu.edu/~nmv/2013/07/09/for-faster-r-use-openblas-instead-better-than-atlas-trivial-to-switch-to-on-ubuntu/ further details]). | Many Octave functions are wrappers to optimized numerical libraries, notably BLAS and ATLAS. It is possible to achieve impressive performance gains by simply using a library tuned to your platform. One example is using OpenBLAS to replace the default BLAS implementation ([http://www.stat.cmu.edu/~nmv/2013/07/09/for-faster-r-use-openblas-instead-better-than-atlas-trivial-to-switch-to-on-ubuntu/ further details]). | ||
− | On Linux, this just takes a few commands | + | On some Linux distributions, this just takes a few commands. For instance, on Ubuntu, it usually suffices to run |
− | sudo apt-get install libopenblas-base libatlas3gf-base | + | sudo apt-get install libopenblas-base libatlas3gf-base |
− | + | followed by | |
− | sudo update-alternatives --config libblas.so.3 | + | sudo update-alternatives --config libblas.so.3 |
− | and | + | and then selecting the openblas option. |
Revision as of 19:29, 17 September 2016
Preferences
Sometimes, Octave defaults are not the best for someone's specific use. To change the defaults, use the following on the .octaverc file.
Changing default figure size
The default size of a figure may be appropriate for simple figures but not so much when using subplot
for example. This can be changed though.
Code: change default figure size |
set (0, 'DefaultFigurePosition', [1 get(0, "screensize")(4:-1:3) get(0, "DefaultFigurePosition")(4)]); |
The value of DefaultFigurePosition
must be a four element vector with the x and y coordinates for the figure, followed by its width and height. The code above sets the default image to be placed at the top of the monitor, with the width of the monitor and the same height previously set as default.
Changing default font for axes
To display tex characters such as '\alpha' or '\lambda', a TrueType font ("arial" for instance) is better for the gnuplot backend [1]
Code: change default axes font name |
set (0, "DefaultAxesFontName", "Arial") |
It can then be necessary to change the font size as well
Code: change default axes font size |
set(0, 'DefaultAxesFontSize', 10) |
Shorten help message
To get rid of the long help message with the link to the Octave homepage place this in your startup file usually at ~/.octaverc (if it does not exist, create it). See the documentation for more information.
suppress_verbose_help_message(1)
Tiny helper functions
This is a list of tiny helper functions (the equivalent of e.g., shell aliases), the kind one would have on its .octaverc file.
Reload 'octave.rc' after 'clear'
When using clear
, one may accidentally remove functions (alias) or other variables set on the octave.rc file. This can fixed by shadowing the clear
function with the following:
Code: reload octave.rc after clear |
function clear (varargin) args = sprintf (', "%s"', varargin{:}); evalin ("caller", ['builtin ("clear"' args ')']); source ("~/.octaverc"); endfunction |
The problem with this approach is if there's path manipulation on the octave.rc file, such as addpath
. A workaround is needed for each case since it is not possible to obtain a reliable list of what's in Octave load path. But basically should be to undo what the file does, before source ("~/.octaverc")
.
If there's a pkg unload all
on it, this would also unload all packages. The following adjustment will keep the packages loaded
Code: reload octave.rc after clear but keep packages loaded |
function clear (varargin) args = sprintf (', "%s"', varargin{:}); evalin ("caller", ['builtin ("clear"' args ')']); pkglist = pkg ("list"); loadedpkg = cell (0); for ii = 1:numel (pkglist) if (pkglist{ii}.loaded) loadedpkg{end+1} = pkglist{ii}.name; endif endfor source ("~/.octaverc"); if (numel (loadedpkg) != 0) pkg ("load", loadedpkg{:}); endif endfunction |
replace help with man
If you use octave too much, you'll find yourself trying to use help
instead of man
on bash. This function will fix that so you can use man
in your octave instance (you can also do the opposite, create a help
alias in bash but man
has fewwer characters).
Code: alias to help |
function man (name) help (char (name)) endfunction |
C++
Real matrix operations
This is a table of matrix operations commonly performed in Octave and their equivalents in C++ when using the octave libraries.
Operation | Octave | C++ |
add | A+B | A+B |
subtract | A-B | A-B |
matrix multiplication | A*B | A*B |
element multiplication | A.*B | product(A,B) |
element division | A./B | quotient(A,B) |
transpose* | A' | A.transpose() |
select element m,n of A** | A(m,n) | A(m-1,n-1) |
select row N of A** | A(N,:) | A.row(N-1) |
select column N of A** | A(:,N) | A.column(N-1) |
extract submatrix of A | A(a:b,c:d) | A.extract(a-1,c-1,b-1,d-1) |
absolute value of A | abs(A) | A.abs() |
comparison to scalar*** | A>2 | mx_el_gt(A,2) |
A<2 | mx_el_lt(A,2) | |
A==2 | mx_el_eq(A,2) | |
A~=2 | mx_el_ne(A,2) | |
A>=2 | mx_el_ge(A,2) | |
A<=2 | mx_el_le(A,2) | |
matrix of zeros | A=zeros(m,n) | A.fill(0.0) |
matrix of ones | A=ones(m,n) | A.fill(1.0) |
identity matrix | eye(N) | identity_matrix(N,N) |
inverse of A | inv(A) | A.inverse() |
pseudoinverse of A | pinv(A) | A.pseudo_inverse() |
diagonal elements of A | diag(A) | A.diag() |
column vector | A(:) | ColumnVector(A.reshape (dim_vector(A.length()))) |
row vector | A(:)' | RowVector(A.reshape (dim_vector(A.length()))) |
check for Inf or NaN | any(~isfinite(A)) | A.any_element_is_inf_or_nan() |
stack two matrices vertically | A=[B;C] | B.stack(C) |
uniform random matrix | rand(a,b) | octave_rand::distribution("uniform"); octave_rand::matrix(a,b) |
normal random matrix | randn(a,b) | octave_rand::distribution("normal"); octave_rand::matrix(a,b) |
sum squares of columns | sumsq(A) | A.sumsq() |
sum along columns | sum(A,1) | A.sum(0) |
sum along rows | sum(A,2) | A.sum(1) |
product along columns | prod(A,1) | A.prod(0) |
product along rows | prod(A,2) | A.prod(1) |
cumsum along columns | cumsum(A,1) | A.cumsum(0) |
cumsum along rows | cumsum(A,2) | A.cumsum(1) |
cumproduct along columns | cumprod(A,1) | A.cumprod(0) |
cumproduct along rows | cumprod(A,2) | A.cumprod(1) |
number of rows | size(A,1) | A.rows() |
number of columns | size(A,2) | A.cols() |
Notes:
- Transpose, addition, and multiplication operations also apply to RowVector, ComplexRowVector, ColumnVector, and ComplexColumnVector data types when the dimensions are in agreement.
- The difference is due to the fact that arrays are zero-based in C++, but one-based in Octave.
- The names of Octave internal functions, such as mx_el_gt, are not documented and are subject to change. Functions such as mx_el_gt may eventually be available at both the scripting level and in C++ under more common names such as gt.
Complex Matrix Operations
Operation | Octave | C++ |
conjugate tranpose | A' | A.hermitian() |
General
Vectorizing Tricks
You can easily fill a vector with an index:
for i=1:n, x(i) = i; end
x = 1:n;
This works for expressions on the index by wrapping the index in an expression:
for i=1:n, x(i) = sin(2*pi*i*f/r); end
x = sin(2*pi*(1:n)*f/r);
You can also work with other vectors this way:
for i=1:n, x(i) = sin(2*pi*y(i)*f/r); end
x = sin(2*pi*y*f/r);
Conditionals in the for loop are a little bit tricky. We need to create an index vector for the true condition, and another for the false condition, then calculate the two independently.
for i=1:n, if y(i)<1, x(i)=y(i); else x(i) = 2*y(i); endif
idx = y < 1; x(idx) = y(idx); x(!idx) = 2*y(!idx);
FIXME: add the following
- examples from matrices
- tricks with sort and cumsum (e.g., hist, lookup)
- counter-examples such as a tridiagonal solver
- sparse matrix tricks
- tricks relying on fortran indexing
Other references
- MATLAB array manipulation tips and tricks by Peter Acklam: http://home.online.no/~pjacklam/matlab/doc/mtt/index.html
- The MathWorks: Code Vectorization Guide: http://www.mathworks.com/support/tech-notes/1100/1109.html
Changing BLAS
Many Octave functions are wrappers to optimized numerical libraries, notably BLAS and ATLAS. It is possible to achieve impressive performance gains by simply using a library tuned to your platform. One example is using OpenBLAS to replace the default BLAS implementation (further details).
On some Linux distributions, this just takes a few commands. For instance, on Ubuntu, it usually suffices to run
sudo apt-get install libopenblas-base libatlas3gf-base
followed by
sudo update-alternatives --config libblas.so.3
and then selecting the openblas option.