Mercurial > octave-nkf
view PROJECTS @ 5540:cda6a105ae9a before-ov-branch
[project @ 2005-11-17 05:47:13 by jwe]
| author | jwe |
|---|---|
| date | Thu, 17 Nov 2005 05:47:13 +0000 |
| parents | b4cfbb0ec8c4 |
| children | 9761b7d24e9e |
line wrap: on
line source
<html> <pre> Octave PROJECTS -*- text -*- =============== Check with maintainers@octave.org for a possibly more current copy. Also, if you start working steadily on a project, please let maintainers@octave.org know. We might have information that could help you; we'd also like to send you the GNU coding standards. This list is not exclusive -- there are many other things that might be good projects, but it might instead be something we already have, so check with maintainers@octave.org before you start. --------- Numerical: --------- * Improve logm, and sqrtm. * Improve complex mapper functions. See W. Kahan, ``Branch Cuts for Complex Elementary Functions, or Much Ado About Nothing's Sign Bit'' (in The State of the Art in Numerical Analysis, eds. Iserles and Powell, Clarendon Press, Oxford, 1987) for explicit trigonometric formulae. * Make functions like gamma() return the right IEEE Inf or NaN values for extreme args or other undefined cases. * Handle complex values in fread and fwrite. * Support for lp_solve for linear programming problems. * Free QP solver. * Free NLP solver. * Fix CollocWt to handle Laguerre polynomials. Make it easy to extend it to other polynomial types. * Make filter faster (perhaps by calling BLAS functions). * Add optional arguments to colloc so that it's not restricted to Legendre polynomials. * Fix eig to also be able to solve the generalized eigenvalue problem, and to solve for eigenvalues and eigenvectors without performing a balancing step first. * Move rand, eye, xpow, xdiv, etc., functions to the matrix classes. * Use octave_allocator for memory management in Array classes once g++ supports static member templates. * Implement the following functions: -- ppval -- cross -- dot * When constructing NLConst (and other) objects, make sure that there are sufficient checks to ensure that the dimensions all conform. * Allow parameters to be passed through the call to fsolve() to the user-supplied function for Matlab compatibility. Don't place an upper limit on the number of arguments. * Check matrix classes for proper handling of empty matrices. * Make operations with empty matrices produce empty matrices, for compatibility with Matlab (but only if Matlab 5 still does things this way). For example: [1, 2] * [] ==> []. * Improve design of ODE, DAE, classes. * Extend meaning of .* to include v .* M or M .* v (where v is a column vector with the same number of rows as M) to scale rows of M by elements of v. Similarly, if w is a row vector with as many columns as M, then either w .* M or M .* w scales the columns of M. * Given two vectors x and y of length m and n, implement a function outer (x, y, f) that returns an m-by-n matrix with entries f (x(i), y(j)). If f is omitted, multiplication is the default. Should probably work for any vectors, not just if x is a column vector and y is a row vector. * Make it possible to solve b = L \ x efficiently, either by providing an explicit function call, or by automatically determining that L is triangular. If it is done automatically, provide some means for determining whether Octave has actually detected that the matrix is triangular. * The polyfit function uses the economy QR factorization, but even that can take a lot of time for large datasets. Consider an option to compute the result with inv (A' * A) * A' * y or some other faster method.d Possibly just switch to this method if the dataset is larger than some value. * Make QR more memory efficient for large matrices when not all the columns of Q are required (apparently this is not handled by the lapack code yet). * Consider making the behavior of the / and \ operators for non-square systems compatible with Matlab. Currently, they return the minimum norm solution from DGELSS, which behaves differently. --------------- Sparse Matrices: --------------- * QR factorization functions, also for use in lssolve functions. Write svds function based on this. Write sprank function based on svds. * Once dmperm is implemented, perhaps use the technique to detect permuted triangular matrices with permutations in both rows and cols. * Implement fourth argument to the sprand and sprandn that the leading brand implements. * Sparse logical indexing in idx_vector class so that something like "a=sprandn(1e6,1e6,1e-6); a(a<1) = 0" won't cause a memory overflow. * Write the rest of the sparse docs * The algo in TOMS 582 is perfect for symrcm function. However, this is under the ACM license and can't be used in a GPL program. An alternative is that PETSC is GPL compatiable and has a symrcm implemented from the original SPARSPAK. Its not clear that this is legal to me as I have found no clarification of the original license of SPARSPAK. As PETSC has had this code for over 10 years without problem, we can perhaps assume that there is no issues. Maybe need to contact PETSC people or the SPARSPAK people at uni of waterloo to check issues. * Make spalloc(r,c,n) actually create an empty sparse with n non-zero elements? This allows something like sm = spalloc (r,c,n) for j=1:c for i=1:r tmp = foo (i,j); if (tmp != 0.) sm (i,j) = tmp; endif endfor endfor actually make sense. Otherwise the above will cause massive amounts of memory reallocation. The fact is that this doesn't make sense in any case as the assign function makes another copy of the sparse matrix. So although spalloc might easily be made to have the correct behaviour, the first assign will cause the matrix to be resized !!! There seems to be no simple way to treat this but a complete rewrite of the sparse assignment functions... * Port the sparse testing code into the DejaGNU testing code. Add tests for spkron, matrix_type, luinc.. * Other missing Functions - eigs - dmperm Use Tim Davis' BTF code when available - symmmd Superseded by symamd - colmmd Superseded by colamd - sprandsym - etreeplot - treeplot - gplot - treelayout - cholinc - condest - normest - bicg - bicgstab - cgs - gmres - lsqr - minres - pcg Can this be taken from o-forge? Is it compatiable - pcr Can and should this be taken from o-forge? - qmr - symmlq - spaugment -------- Graphics: -------- * Make plotting with plplot work. * Fix interface with gnuplot to wait for gnuplot to output any text before octave returns a prompt. Possible by implementing two way communication between gnuplot and Octave. * Handle gnuplot ranges correctly for parametric modes (accept 3 ranges for 2d plots and 5 ranges for 3d plots). * Make gsave (and possibly gload) work. Implement gsave by having it also alter the plot command to not use temporary files (perhaps with some user-specified template for naming them) and then sending a `save' command to gnuplot. * It would be nice to be able to check that a plot is currently being displayed. * Implement clf, gcf, get, set, orient, print, close, etc. in Matlab-compatible ways. * Make it possible to check the current graphics terminal type. * If possible, pass data to gnuplot without using temporary files. * If possible, pass binary data to gnuplot to speed things up. * If using gnuplot, consider setting a smaller default for the `zero' value (e.g., set zero sqrt (realmin) or something). ------- Strings: ------- * Improve performance of string functions, particularly for searching and replacing. * Provide some regex matching functions. * Convert string functions to work on string arrays. * Make find work for strings. * Consider making octave_print_internal() print some sort of text representation for unprintable characters instead of sending them directly to the terminal. (But don't do this for fprintf!) * Consider changing the default value of `string_fill_char' from SPC to NUL. * Consider making ["test", []] ==> "test", for compatibility with Matlab, at least when some set of preferences are set. ---------------- Other Data Types: ---------------- * Template functions for mixed-type ops. ------------------------ Graphical User Interface: ------------------------ * In an X11 or other windowing environment, allow the user to pop up windows for menus and other purposes. A good place to start might be Tk, as long as Tcl is avoided. * Add a way to handle events, like alarms, mouse clicks, etc. ------------ Input/Output: ------------ * Make fread and fwrite work for complex data. Iostreams based versions of these functions would also be nice, and if you are working on them, it would be good to support other size specifications (integer*2, etc.). * Make fread and fopen look in LOADPATH for files. * Make load and save look for <file>.mat if only given <file>. Potential sticky points: - For load, if both foo and foo.mat exist, should it prefer foo or foo.mat? Should the preference depend on the arguments to load? I think it should only prefer .mat files if the -mat-binary option is supplied, or if the file foo.mat exists but the file foo does not. - For save, should it prefer to create foo or foo.mat? Should the preference depend on the arguments to save? Should the default_save_format imply a default preference? I think it should only create .mat files if it is writing Matlab compatible files. * Move some pr-output stuff to liboctave. * Make the cutoff point for changing to packed storage a user-preference variable with default value 8192. * Make it possible to load other image formats (ppm, pbm, etc. would probably be best since there are already filters to convert to these formats from others.) * Complain if there is not enough disk space available (I think there is simply not enough error checking in the code that handles writing data). * Make it possible to tie arbitrary input and output streams together, similar to the way iostreams can be tied together. ----------- Interpreter: ----------- * Allow customization of the debug prompt. * For the keyboard function, parse return (or quit) more intelligently so that something like debug> x = 1; return will work as expected. * Handle multi-line input at the keyboard/debug prompt correctly. * Fix the parser so that if (expr) 'this is a string' end is parsed as IF expr STRING END. * Consider grouping all preference variables in a structure instead of further polluting the namespace. Maybe `Octave_options.xxx'? * Rewrite functions in input.cc to do the right thing in all cases of using readline, forced interactive behavior, echoing of input, etc. * Consider making linspace() and logspace() return the value corresponding to the first argument if the number of requested points is 1. * Consider allowing an arbitrary property list to be attached to any variable. This could be a more general way to handle the help string that can currently be added with `document'. * Allow more command line options to be accessible as built-in variables (--echo-commands, etc.). * Allow `octave -c STRING' to execute commands from STRING then exit. * Make the interpreter run faster. * Make it possible to disable or enable all warnings on an individual basis from the command line or via some built-in structure variable. * Warn about complex comparisons? Could just use double_value() or matrix_value() instead of explicit conversions to real types. For this to really be useful, some additional information must be available to point to the location of the code that triggers the warning. * Allow arbitrary lower bounds for array indexing. * Improve performance of recursive function calls. * Improve the way ignore_function_time_stamp works to allow selecting by individual directories or functions. * Add a command-line option to tell Octave to just do syntax checking and not execute statements. * Is it necessary for do_binary_op and do_unary_op to be friends of the tree_constant class. * Clean up symtab and variable stuff. * Input stream class for parser files -- must manage buffers for flex and context for global variable settings. * make parser do more semantic checking, continue after errors when compiling functions, etc. * Make LEXICAL_ERROR have a value that is the error message for parse_error() to print? * Add a run-time alias mechanism that would allow things like alias fun function_with_a_very_long_name so that `function_with_a_very_long_name' could be invoked as `fun'. * What should is_global() return when called for built-in variables? * Allow local changes to variables to be written more compactly than is currently possible with unwind_protect. For example, function f () local prefer_column_vectors = something; ... endfunction would be equivalent to function f () save_prefer_column_vectors = prefer_column_vectors; unwind_protect prefer_column_vectors = something; ... unwind_protect_cleanup prefer_column_vectors = save_prefer_column_vectors; end_unwind_protect endfunction * Fix all function files to check for bogus inputs (wrong number or types of input arguments, wrong number of output arguments). * Reduce the memory and time required to parse very large matrix lists. * Handle options for built-in functions more consistently. * Too much time is spent allocating and freeing memory. What can be done to improve performance? * Error output from Fortran code is ugly. Something should be done to make it look better. * It would be nice if output from the Fortran routines could be passed through the pager. * Attempt to recognize common subexpressions in the parser. * Handle arrays with more than two dimensions. * Consider making it possible to specify an empty matrix with a syntax like [](e1, e2). Of course at least one of the expressions must be zero... * Is Matrix::fortran_vec() really necessary? * print_usage() should set error_state in most cases? * Add a command that works like bash's `builtin' command. * Handle end-of-line comments correctly in parse trees for use with the type command. * Clean up eye, eval, feval, keyboard, input, ones, zeros. * It would be nice to have an interactive debugger. * Make whos report total memory used by variables (and functions?). * Rewrite whos and the symbol_record_info class. Write a built-in function that gives all the basic information, then write who and whos as M-files. * Make whos work for structure elements: prot type rows cols name ==== ==== ==== ==== ==== wd matrix m n struct.field * Allow who to return information as a list of strings. * On systems that support matherr(), make it possible for users to enable the printing of warning messages. * Make it possible to mark variables and functions as read-only. * Provide a built-in function for applying a scalar function to an array. Be sure to note in the manual that this is not the preferred way to write a function that can handle vector/matrix arguments because there is a significant overhead for function calls. If you are really looking to make a function work for vector/matrix arguments and you want it to run fast, you should write it in terms of the existing vector/matrix operators as much as possible. * Make it possible to write a function that gets a reference to a matrix in memory and change one or more elements without generating a second copy of the data. * Use nanosleep instead of usleep if it is available? Apparently nanosleep is to be preferred over usleep on Solaris systems. ------- History: ------- * Add an option to allow saving input from script files in the history list. * The history command should accept two numeric arguments to indicate a range of history entries to display, save or read. * Add an option to include information about the Octave session in the history list. Possibly a time/date stamp and the current Octave line number, appended as a comment (users should probably be able to control the format). * Avoid writing the history file if the history list has not changed. * Avoid permission errors if the history file cannot be opened for writing. * Fix history problems -- core dump if multiple processes are writing to the same history file? ------------------------------ Configuration and Installation: ------------------------------ * Add an --enable-pathsearch option to configure to make it possible to configure and run without kpathsea. * Makefile changes: -- eliminate for loops -- define shell commands or eliminate them -- verify distclean -- consolidate targets * Make it possible to configure so that installed binaries and shared libraries are stripped. * Create a docs-only distribution? ------------------------------ Documentation and On-Line Help: ------------------------------ * Document new features. -- history-search-{back,for}ward. -- Other stuff mentioned in the NEWS file. * Improve the Texinfo Documentation for the interpreter. It would be useful to have lots more examples, to not have so many forward references, and to not have very many simple lists of functions. * The docs should mention something about efficiency and that using array operations is almost always a good idea for speed. * Texinfo documentation for the C++ classes. * Support multiple info files, perhaps allowing one or more in each directory in the LOADPATH, so that local collections of M-files could be documented with Info. * Improve help messages for operators and keywords in help.cc. * Make index entries more consistent to improve behavior of `help -i'. * Make `help -i' try to find a whole word match first. * Allow help for local additions to be accessible with help -i. * Clean up help stuff. * Demo files. * As the number of m-files with octave grows perhaps a 'Contents.m' file for each toolbox (directory) would be appropriate so one knows exactly what functions are in a toolbox with a quick look. It would be best to generate information for each function directly from the M-files, so that the information doesn't have to be duplicated, and will remain current if the M-files change. It would also be best to do as much of this as possible in an M-file, though I wouldn't mind adding some basic support for listing the names of all the directories in the LOADPATH, and the names of all the M-files in a given directory if that is needed. Also make it possible to recursively search for Contents files: help dir -- Contents from dir help dir// -- Contents from dir and all its subdirectories help dir1/dir2 -- Contents from dir2 which is under dir1 * Some sort of blurb (2-3 pages) describing Octave in a reasonably coherent way, where to get it etc., perhaps formatted pretty, i.e. not just text. Maybe start with the latest Announce file. ----- Tests: ----- * Improved set of tests: -- Tests for various functions. Would be nice to have a test file corresponding to every function. -- Tests for element by element operators: + - .* ./ .\ .^ | & < <= == >= > != ! -- Tests for boolean operators: && || -- Tests for other operators: * / \ ' .' -- Tests from bug reports. -- Tests for indexed assignment. Need to consider the following: o fortran-style indexing o zero-one indexing o assignment of empty matrix as well as values o resizing * Tests for all internal functions. ----------- Programming: ----------- * Better error messages for missing operators? * Eliminate duplicate enums in pt-exp.cc, pt-const.cc, and ov.cc. * Handle octave_print_internal() stuff at the liboctave level. Then the octave_value classes could just call on the print() methods for the underlying classes. * As much as possible, eliminate explicit checks for the types of octave_value objects so that user-defined types will automatically do the right thing in more cases. * Only include config.h in files that actually need it, instead of including it in every .cc file. Unfortunately, this might not be so easy to figure out. * GNU coding standards: -- Add a `Makefile' target to the Makefiles. -- Comments on #else and #endif preprocessor commands. -- Change error message format to match standards everywhere. * Use STL stuff instead of libg++ lists, maps, and stacks. * Eliminate more global variables. * Move procstream to liboctave. * Use references and classes in more places. * Share more code among the various *_options functions. ------------- Miscellaneous: ------------- * Implement some functions for interprocess communication: bind, accept, connect, gethostbyname, etc. * The installation process should also install octave.el. This needs to detect the appropriate Emacs binary to use to byte-compile the .el file. Following GNU Emacs philosophy, installation would be into $(prefix)/share/emacs/site-lisp by default, but it should be selectable. * The ability to transparently handle very large files: Juhana K Kouhia <kouhia@nic.funet.fi> wrote: If I have a one-dimensional signal data with the size 400 Mbytes, then what are my choices to operate with it: * I have to split the data * Octave has a virtual memory on its own and I don't have to worry about the splitting. If I split the data, then my easily programmed processing programs will become hard to program. If possible, I would like to have the virtual memory system in Octave i.e. the all big files, the user see as one big array or such. There could be several user selectable models to do the virtual memory depending on what kind of data the user have (1d, 2d) and in what order they are processed (stream or random access). Perhaps this can be done entirely with a library of M-files. * An interface to gdb. Michael Smolsky <fnsiguc@weizmann.weizmann.ac.il> wrote: I was thinking about a tool, which could be very useful for me in my numerical simulation work. It is an interconnection between gdb and octave. We are often managing very large arrays of data in our fortran or c codes, which might be studied with the help of octave at the algorithm development stages. Assume you're coding, say, wave equation. And want to debug the code. It would be great to pick some array from the memory of the code you're develloping, fft it and see the image as a log-log plot of the spectral density. I'm facing similar problems now. To avoid high c-development cost, I develop in matlab/octave, and then rewrite into c. It might be so much easier, if I could off-load a c array right from the debugger into octave, study it, and, perhaps, change some [many] values with a convenient matlab/octave syntax, similar to a(:,50:250)=zeros(100,200), and then store it back into the memory of my c code. * Add a definition to lgrind so that it supports Octave. (See http://www.tex.ac.uk/tex-archive/support/lgrind/ for more information about lgrind.) ------ Always: ------ * Squash bugs. --30-- </pre> </html>