MUI Subclassing Tutorial: SciMark2 Port

Grzegorz Kraszewski

The Application

Many programming tutorials tend to bore readers with some useless examples. In this one a "real world" application will be ported to MorphOS and "MUI-fied". The application is SciMark 2. SciMark is yet another CPU/memory benchmark. It performs some typical scientific calculations like Fast Fourier Transform, matrix LU decomposition, sparse matrix multiplication and so on. The benchmark measures mainly CPU speed at floating point calculations, cache efficiency and memory speed. Being written in Java initially, it has been rewritten in C (and in fact in many other languages). The C source is available on the project homepage.

The source uses only pure ANSI C standard, so it is easily compilable on MorphOS using provided Makefile. One has just replace $CC = cc line to $CC = gcc, to match the name of the MorphOS compiler. As a result, a typical shell-based application is obtained. Here are example results for a Pegasos 2 machine with G4 processor:

Not very impressive in fact. This is because no optimizaton flags are passed to the compiler in the makefile. They can be added by inserting a line $CFLAGS = -O3 below the $CC = gcc one. Let's also link with libnix (a statically linked unix environment emulation, see Standard C and C++ Libraries) by adding -noixemul to CFLAGS and LDFLAGS. After rebuilding the program and running it again the results are significantly improved (the program has been compiled with GCC 4.4.4 from the official SDK).

This shows how important is optimization of the code, especially computationally intensive one. Optimized code is more than 4 times faster!

Code Inspection

The original source code is well modularized. Five files: FFT.c, LU.c, MonteCarlo.c, SOR.c and SparseCompRow.c implement the five of single benchmarks. Files array.c and Random.c contain auxiliary functions used in benchmarks. File Stopwatch.c implements time measurement. An important file kernel.c gathers all the above and provides five functions performing complete benchmarks with timing. Finally scimark2.c contains the main() function and implements the shell interface.

A planned MUI interface should allow to run every benchmark separately and run all of them. There is also -large option, which increases memory sizes of calculated problems, so they do not fit into the processor cache. A general rule of porting is that as few files as possible should be modified. The rule makes it easier to upgrade the port when a new version of original program is released. In the case of SciMark, only one file, scimark2.c has to be replaced. An advanced port may also replace Stopwatch.c with code using timer.device directly for improved time measurements accuracy, this is out of scope of this tutorial however.

A closer look at "scimark2.c" reveals that there is a Random object (a structure defined in "Random.h"), which is required for all the benchmarks. In the original code it is created with new_Random_seed() at the program start and disposed with delete_Random() at exit. The best place for it in the MUI-fied version is the instance data area of subclassed Application class. Then it can be created in OM_NEW() of the class and deleted in OM_DISPOSE(). These two methods should be overridden then.

GUI Design

Of course there is no one and only proper GUI design for SciMark. A simple design, using a limited set of MUI classes is shown on the left. There are five buttons for individual benchmarks and one for running all of them. All these buttons are instances of the Text class. On the right there are gadgets for displaying benchmark results. These gadgets belong to Text class too, just having different attributes. The "Large Data" button, of Text class of course, is a toggle one. Surprisingly the status bar (displaying "Ready.") is not an instance of the Text class, but Gauge class. Then it will be able to display a progress bar when running all five tests. Spacing horizontal bars above the "All Benchmarks" button are instances of the Rectangle class. There are also three invisible objects of the Group class. The first is a vertical, main group, being the root object of the window. It contains two sub-groups. The upper one is the table group with two columns and contains all the benchmark buttons and result display gadgets. The lower group contains "Large Data" toggle button and the status bar.

The simplest way to start with GUI design is just to copy the "Hello World" example. Then MUI objects may be added to build_gui() function. The modified example is ready to compile and run. It is not a complete program of course, just a GUI model without any functionality added.

A quick view into the build_gui() function reveals that it does not contain all the GUI code. Code for some subobjects is placed in functions called from the main MUI_NewObject(). Splitting the GUI building function into many subfunctions has a few important advantages:

• Improved code readability and easier modifications. A single MUI_NewObject() call gets longer and longer quickly as the project evolves. Editing a large function spanning over a few screens is uncomfortable. Adding and removing GUI objects in such a function becomes nightmare even with indentation used consequently. On the other hand the function can have 10 or more indentation levels, which makes it hard to read as well.
• Code size reduction. Instead of writing very similar code multiple times, for example buttons with different labels, a subroutine may be called with label as an argument.
• Debugging. It happens sometimes that MUI refuses to create the application object because of some buggy tags or values passed. If the main MUI_NewObject() call is splitted into subfunctions, it is easy to isolate the buggy object by inserting some Printf()-s in subfunctions.

Methods and Attributes

The SciMark GUI just designed, defines six actions of the application. There are five actions of running individual benchmarks and the sixth one of running all the tests and calculate the global result. Actions will be directly mapped to an Application subclass methods. There is also one attribute connected with "Large Data" button, it determines sizes of problems solved by benchmarks. Methods do not need any parameters, so there is no need to define method messages. An attribute may be applicable at init time (in the object constructor), may be also settable (needs OM_SET() method overriding) and gettable (needs OM_GET() method overriding). Our new attribute, named APPA_LargeData in the code only needs to be settable. In the constructor it can be implicitly set to FALSE, as the button is switched off initially. Gettability is not needed, because this attribute will be used only inside the Application subclass.

It is recommended that every subclass in the application is placed in a separate source file. It helps to keep code modularity and also allows for hiding class private data. This requires writing a makefile, but one is needed anyway, as the original SciMark code consists of multiple files. Implementing the design directions discussed above a class header file and class code can be written. The class still does nothing, just implements six empty methods and overrides OM_SET(), OM_NEW() and OM_DISPOSE(). In fact it is a boring template work and as such it has been generated with ChocolateCastle template generator. Unfortunately ChocolateCastle is still beta, so files had to be tweaked manually after generation.

The next step in the application design is to connect methods and attributes with GUI elements using notifications. Notifications must be of course created after both source and target object are created. In the SciMark code they are just set up after executing build_gui(). All the six action buttons have very similar notifications, so only one is shown here:

DoMethod(findobj(OBJ_BUTTON_FFT, App), MUIM_Notify, MUIA_Pressed, FALSE,
 App, 1, APPM_FastFourierTransform);

"Large Data" button has a notification setting the corresponding attribute:

DoMethod(findobj(OBJ_BUTTON_LDATA, App), MUIM_Notify, MUIA_Selected, MUIV_EveryTime,
 App, 3, MUIM_Set, APPA_LargeData, MUIV_TriggerValue);

Overriding the Constructor

The first thing do be done in the constructor is calling the superclass method. There is an exception when an object created can have children objects. These objects are then created before calling superclass, usually with convenient DoSuperNew() call. When the object is ready, its instance data area is created and may be accessed.

Overriding the Destructor

The destructor should just free all resources allocated by the constructor, then call the superclass method.

IPTR ApplicationDispose(Class *cl, Object *obj, Msg msg)
{
  struct ApplicationData *d = INST_DATA(cl, obj);

  if (d->R) Random_delete(d->R);
  return DoSuperMethodA(cl, obj, msg);
}

The only resource to free is the Random structure.