view Plugins/Input/console/design.txt @ 137:b8d4c1faa6d7 trunk

[svn] Import WMA decoder into SVN.

Game_Music_Emu Design Notes


Managing Complexity
-------------------
Complexity has been a factor in most library decisions. Many features
have been passed by due to the complexity they would add. Once past a
certain level, complexity prevents mentally grasping the library in its
entirety, at which point more defects will occur and be hard to find.

I chose 16-bit signed samples because it seems to be the most common
format. Supporting multiple formats would add too much complexity to be
worth it. Other formats can be obtained via conversion.

I've kept interfaces fairly lean, leaving many possible features
untapped but easy to add if necessary. For example the classic emulators
could have volume and frequency equalization adjusted separately for
each channel, since they each have an associated Blip_Synth.

Source files of 400 lines or less seem to be the best size to limit
complexity. In a few cases there is no reasonable way to split longer
files, or there is benefit from having the source together in one file.


Library Configuration
---------------------
Library optimizations can be configured through macros defined in
config.h. By default, the library is configured to be most likely to
compile and work on any platform, rather than be most optimal with
increased chance of problems. It's easier to track down optimiztation
problems if the library can first be shown to work correctly without
them.


Flexibility through indirection
-------------------------------
I've tried to allow the most flexibility of modules by using indirection
to allow extension by the user. This keeps each module simple and more
focused on its unique task.

The classic emulators use Multi_Buffer, which potentially allows a
separate Blip_Buffer for each channel. This keeps emulators free of
typical code to allow output in mono, stereo, panning, etc.

All emulators use a reader object to access file data, allowing it to be
stored in a regular file, compressed archive, memory, or generated
on-the-fly. Again, the library can be kept free of the particulars of
file access and changes required to support new formats.


Choice of pre-ISO (ARM) C++
---------------------------
The library started out as an unreleased NSF player written using ISO
C++. The code was clean enough that I decided to release it as a player
library. Before release I evaluated its use of C++ features to determine
the important ones.

Namespaces and exceptions weren't essential, so I compared a version of
the library with and without them. I decided that I could do without and
get better compatibility with older compilers or newer ones with buggy
namespace and exception implementations.

Templates are the worst area for most compilers, due to their inherent
complexity, but they are too useful to avoid entirely. I've used them
sparingly and in ways that compilers are more likely to work with.

Sticking to ARM C++ has helped keep the library simpler to understand.


Platform-specific optimization
------------------------------
Performance profiling doesn't shown any big bottlenecks that warrant
heavy platform-specific optimization. The main bottlenecks are CPU
emulation, Blip_Buffer synthesis and sample reading, Super NES DSP, Sega
Genesis FM, and Fir_Resampler.

Further optimization of the CPU emulators can probably only be achieved
by writing them in assembly or using dynamic recompilation. Blip_Buffer
might benefit somewhat from vector instructions. Sega Genesis FM
synthesis can probably be made twice as fast by someone who fully
understands its operation (I am almost clueless about its internals).
The Super NES DSP might have some room for optimization. Fir_Resampler
would benefit greatly from vector operations.

Most of the above optimizations add more complexity than they are worth.
All that seems worthwhile is optimization of Sega Genesis FM emulation
and Fir_Resampler.


Preventing Bugs
---------------
I've done many things to reduce the opportunity for defects. A general
principle is to write code so that defects will be as visible as
possible. I've used several techniques to achieve this.

I put assertions at key points where defects seem likely or where
corruption due to a defect is likely to be visible. I've also put
assetions where violations of the interface are likely. In emulators
where I am unsure of exact hardware operation in a particular case, I
output a debug-only message noting that this has occurred; many times I
haven't implemented a hardware feature because nothing uses it. I've
made code brittle where there is no clear reason flexibility; code
written to handle every possibility sacrifices quality and reliability
to handle vaguely defined situations.


Miscellaneous
-------------
I don't like naming header files with a ".hpp" suffix for some reason.
They aren't as visually distinct from ".cpp" source files. The ".cpp"
suffix is useful to inform the compiler that it's a C++ file rather than
a C file, but I don't know of significant practical benefits the ".hpp"
suffix gives over ".h" (I used to use *no* suffix for header files, but
this does cause problems).

When implementation has to be put in a header file, it's as far near the
end as possible to prevent distraction from the public interface.


CPU Cores
---------
I've spent lots of time coming up with techniques to optimize the CPU
cores. Some of the most important: execute multiple instructions during
an emulation call, keep state in local variables to allow register
assignment, optimize state representation for most common instructions,
defer status flag calculation until actually needed, read program code
directly without a call to the memory read function, always pre-fetch
the operand byte before decoding instruction, and emulate instructions
using common blocks of code.

I've successfully used Nes_Cpu in a simple NES emulator, and I'd like to
make all the CPU emulators suitable for use in emulators. It seems a
waste for them to be used only for the small amount of emulation
necessary for game music files.

I debugged the CPU cores by writing a test shell that ran them in
parallel with other CPU cores and compared all memory accesses and
processor states at each step. This provided good value at little cost.

The CPU mapping page size is adjustable to allow the best tradeoff
between memory/cache usage and handler granularity. The interface allows
code to be somewhat independent of the page size.

I optimize program memory accesses to to direct reads rather than calls
to the memory read function. My assumption is that it would be difficult
to get useful code out of hardware I/O addresses, so no software will
intentionally execute out of I/O space. Since the page size can be
changed easily, most program memory mapping schemes can be accommodated.
This greatly reduces memory access function calls.


Sub-Libraries
-------------
I've also released the sound cores as individual libraries to reduce
complexity for emulator authors who just want a single sound core. These
authors will be using the sound cores directly, while users of this
music emulator library won't even see them, so documentation and demos
can be specific to each library.


Documentation
-------------
I started out with separate documentation in HTML and found that it
wasn't going to be easy to maintain. I switched to putting descriptions
of function behavior in header files before the function declarations.
This has worked well so far.

I think the concrete executable demo code helps the most when someone is
using the library for the first time, since it shows a complete program
and provides a framework for using the library. By recording to a sound
file, I can keep the code portable, and the result can be listened to or
examined closely, which is important for something real-time like sound.


I want to write some tutorials to complement the demo code, desribing
the basic framework and operation of the modules.