Difference between revisions of "JWE Project Ideas"
|Line 20:||Line 20:|
=== Syntax, semantics, and data types ===
=== Syntax, semantics, and data types ===
==== Function handle refactoring ====
==== Function handle refactoring ====
Latest revision as of 15:03, 28 April 2022
2021-11-19: This page is out of date -- jwe
Language and functions
Make a list here, pointing to individual bug reports?
Load/save for classdef
See also general load/save issues.
Improve / simplify implementation
Although the basic features that are implemented now appear to mostly work, the implementation seems overly complicated, making it difficult to debug and modify. There seems to be quite a bit of room for improvement here.
Syntax, semantics, and data types
Matlab-compatible argument validation blocks
New language feature, syntax is accepted by parser now but argument validation is not performed.
Function handle refactoring
- Load/save for all types of function handles and all data formats (ascii, binary, hdf5, mat5)
- Use std::shared_ptr for function objects instead of bare pointer to octave_function.
Matlab now uses "" to create string objects that behave differently from Octave double-quoted strings. We could start by creating a compatible string class, then hooking it up to the "" syntax. No matter what, the transition will be difficult because Matlab's "" strings still treat "\n" as two characters (backslash and n) rather than a single character (newline).
Other new data types
Andrew Janke has implementations of these classes (FIX: link to repos here)
- datetime, duration, calendarDuration
single / integer valued ranges
This is a compatibility issue.
- Directories are not properly removed from load path (FIX: link to bug report here)
- Should we really have ADD_PKG and DEL_PKG files? If so, how can we make them safe?
Eliminate special matrix types
Although the special range, diagonal matrix, and permutation matrix data types in Octave require less memory than storing full matrices, they tend to cause trouble when people expect full compatibility or exactly the same results when performing arithmetic on Ranges vs. Matrices. Now that we have broadcasting operators, the need for diagonal matrices is not as great.
Special case FOR loop limits
Currently, "for i = 1:N ..." uses a Range object for the "1:N" loop bounds. If we eliminate Ranges as a special space-saving type, then we should handle this syntax as a special case. Even if we don't eliminate Ranges, that might be a good idea, as we could handle "for i = 1:Inf ..." easily without having to worry about how to deal with that in an ordinary Range object vs. FOR loop bounds.
The semantics for local functions in scripts is different from the way Octave currently handles functions that are defined in script files.
+DIR directories in the loadpath; related to classdef
Octave already searches for files in package directories and understands the PKG.fcn syntax and functionality. The big missing piece is implementation of the "import" functionality and handling it efficiently and in a way that is compatible with Matlab.
Are there better ways to use templates to handle function calls rather than using macros to define a set of functions for array/array, array/scalar, and scalar/array ops as in DEFMXBINOP in mx-inlines.cc?
Sparse matrix issues
Broadcasting does not work for sparse matrices. This seems like a big missing feature.
Octave currently skips structural zeros for most (all?) sparse matrix operations. Matlab returns a sparse matrix filled with NaNs for something like "sprand (5, 5, 0.1) .^ NaN". Should we go for full compatibility? Mathematical correctness? Traditional behavior of sparse matrix libraries? It seems no one really agrees on what is correct or best. Maybe compatibility should win?
In an assignment like Sparse_object(idx) = GrB_object(idx), Octave does not attempt to apply a conversion operator to transform the RHS type to the LHS type. Is this also a problem for assignments of objects with conversion operators to full matrix objects?
graph and digraph
Would it be difficult to provide these objects?
Communication with interpreter
Currently, communication between the GUI and the interpreter mostly happens when the interpreter is otherwise idle and waiting for user input at the command prompt and the implementation is somewhat complicated. We need to determine whether this is the best we can do, or if there is some other implementation that would be more flexible and reliable.
The implementation of the GUI command window for Unix-like systems is a completely separate implementations from the one used on Windows systems. There should be only one, and the GUI should be completely in charge of user input and output. This will probably require implementing some kind of simple output pager internally instead of using an external program, but overall user interaction could be improved.
GUI code editor
Make it possible to use external editors such as Emacs, vim, or others with the GUI in addition to Octave's built-in code editor
Generating EPS or PDF versions of figures needs improvement.
- Scaling plot data values/ranges to fit in single-precision OpenGL values
- Performance issues
- Lack of WYSIWYG
With the rest of the GUI using Qt widgets, we should eliminate the FLTK plotting widget. It duplicates functionality and requires additional effort to maintain. Maybe we no longer need the octave-cli binary (the one that is not linked with Qt libraries)?
Qt toolkit threading
It seems likely that the locking of the gh_manager object is insufficient or even incorrect in some cases.
classdef graphics objects
This is a large project, but one that will likely have to be tackled at some point.
We need to handle UTF-8 (or whatever) characters properly in all parts of Octave. Try to do this in a Matlab-compatible way.
Load / Save
- Make the load and save commands compatible with Matlab's HDF5-based file format. Matlab users expect this and we need something like this to support large arrays anyway. As much as possible, the initial implementation should be written in Octave's scripting language and the proposed Low-level interface to HDF5 functions so that it can easily be updated and patched as needed while we are still working out the details. Only later should we consider translating performance-critical parts to C++, and then, only if really necessary.
- Phase out Octave's own text and binary formats. Too much effort is required to maintain the code to support all the various formats.
Low-level interface to HDF5 functions
Create a thin wrapper for the HDF5 library. As much as possible, make it compatible with the Matlab interface to HDF5. However, we may support newer functions (as of 2020/10/30, the list of Matlab functions appears to correspond to an older version of the library than is presently available in the HDF5 library itself) and support for legacy functions has a low priority.
Also as of 2020/10/30, jwe is working on this project. Help is welcome!
RandStream and Other RNG issues
This is likely a large project, but it would be nice to have updated, compatible interfaces.
Implement mxMakeReal and mxMakeComplex functions.
A proof-of-concept implementation was done several years ago by a Google Summer of Code student. It was never complete and little work has been done since. It also depends on an old version of LLVM. In addition to LLVM, we should consider the JIT library features of GCC.
This is probably the most difficult item (at least for me) since it will require fairly advanced knowledge of compiler infrastructure and Octave internals.
This feature might be nice to have but it has a low priority.
Should we support this feature? Should we refactor the implementation of array objects to make this job easier?
who -file option
Should just read file and list info, not create dummy scope. Likewise for whos function.
Maintenance and packaging
General code quality
- Use C++11 features where possible.
- Better and more complete use of C++ namespaces.
- Better use of C++ features. Especially standard library features as their implementation becomes more widely available. For example, we might be able to simplify some things in Octave by using the C++17 filesystem and special functions libraries, if they provide results that are at least as good what we are using now.
- Eliminate C preprocessor macros where possible
- added_static must go! (not sure about this now)
- Should not expose symbol_record in call_stack functions if possible
- Remove unused symbol_table/scope/record functions
- Use const in more parse tree functions
- Do recursive functions work properly with load/save now?
- Use enums for options internally (typically to replace bool values)
- Audit global variables and eliminate them where possible
We really should be tagging the functions that we wish to export from shared libraries.
Dispatch types for functions
Search for "classes:" in sources to find the few current examples.
min/max nargin values
Should we do this, and allow the interpreter to automatically error when a function is given too few/many arguments?
Move some core toolboxes (communications, control systems, image processing, optimization, signal processing, and statistics), to core Octave so development is managed along with Octave. Core Octave developers are already responsible for these packages anyway, and users don't seem to understand why they need to install them separately. Core parts of the ordinary differential equations package have already been moved to Octave.
- Docs for call stack with examples and illustrations
- Docs for lexer and parser with examples and illustrations
- Docs for fcn_info object
- Docs for load_path object
- Docs for classdef internals
- Docs for Qt graphics toolkit internals
- Docs for Qt GUI and communication with interpreter
- Improve other Doxygen docs for internals to make it easier for new contributors to understand the Octave code base.
Eliminate the following msys packages. Some might be removed entirely if they are unnecessary for running Octave or building Octave Forge packages. Otherwise, we should be building them from source as we do all other tools and libraries that are distributed with Octave. The difficulty is that although the msys packges are typically based on old versions of these packages, they sometimes have fixes that are needed to allow them to run properly on Windows systems. Note also that we distribute a termcap library, but the msys version of less depends on the msys termcap library.