Thursday, September 12, 2013

Light Logging Facilities for Android/iOS/Linux/MacOS/Windows

I don't like to reinvent the wheel, and there are many logging codes out there, but still I couldn't find what I wanted. Those were all either too complex or not exactly what I wanted. That is why I took my old macros and added some ideas taken from this good tutorial: http://www.drdobbs.com/cpp/logging-in-c/201804215.

The result of a few hours is this: https://github.com/carlonluca/LightLogger. I'm using it on Windows/Linux/Mac OS/iOS/Android. This is an overview:
  1. Simple: include the header and build.
  2. Logging to Android results in logs sent to logcat (consider having a look at logcat-colorize, pretty project: https://bitbucket.org/brunobraga/logcat-colorize).
  3. Supports the usual feature of log levels.
  4. Supports coloring of logs: coloring is implemented using ANSI Escape sequences on Linux and MacOS. On Windows I disabled it, but you can enable if you're using something like cygwin or similar. For iOS I use XcodeColors, a great project: https://github.com/robbiehanson/XcodeColors. So a specific implementation is reserved for that platform.
  5. Selecting debug levels with a macro at build-time enables/disables logs. Compiler optimizations should, in most cases, simply strip logs entirely from the binary, resulting in no/minimal overhead.
  6. It is possible to reimplement the "sink" of the logs by reimplementing an output.
  7. I preferred the printf way of formatting logs to the stream implementation. Anyway I tried to provide both. The usage of a "null sink" when logs are disabled should, together with compiler optimizations, make the overhead minimal.
  8. Each log is flushed to avoid issues related to buffering. This might increase the overhead, but it is simple to remove it.
  9. On Windows/Linux/iOS a stack trace function is also available to show the current call stack.
  10. Should be entirely thread-safe.
  11. Each log can be associated to a tag; I commonly use this with grep to filter logs by module.
There a still things I don't like about this approach, like being impossible to use in C sources, needing Objective-C++ instead of simple Objective-C and so on. Still someone may find it useful, so I uploaded to github: https://github.com/carlonluca/LightLogger.

How to Use

Just include in your sources and you're done. Most useful functions are those level-based:

inline bool log_info_t_v(const char* log_tag, const char* format, va_list args);
inline bool log_info_t(const char* log_tag, const char* format, ...);
inline bool log_info_v(const char* format, va_list args);
inline bool log_info(const char* format, ...);

These functions work differently according to the platform: on Android send INFO logs to logcat, on iOS print colored text to Xcode (and thus the XcodeColors plugin is needed if you keep colors enabled), on Windows simply print text and on Mac OS/Linux print text with ANSI colors to the shell. I use the return type to do something like:

if (!condition)
   return log_warn("Ooops, something bad happened, returning failure.");

These functions are "wrappers" for the LC_Log template. You choose a delegate and call:

LC_Log(...).printf(...);
LC_Log(...) << "Some stream based " << "text.";

or use the default logger:

LC_LogDef(...).printf(...);
LC_LogDef(...) << "Some stream based " << "text.";

For specific cases I also needed to write text with specific attributes in the past, so I added something like:

inline bool log_formatted_t_v(
   const char* log_tag,
   LC_LogAttrib a,
   LC_LogColor c,
   const char* format,
   va_list args
   );
inline bool log_formatted_t(
   const char* log_tag,
   LC_LogAttrib a,
   LC_LogColor c,
   const char* format,
   ...
   );
inline bool log_formatted_v(
   const char* format,
   va_list args
   );
inline bool log_formatted(
   LC_LogAttrib a,
   LC_LogColor c,
   const char* format,
   ...
   );
inline bool log_formatted(
   LC_LogColor c,
   const char* format,
   ...
   );

There are a few macros I use to configure for each project: COLORING_ENABLED to enable colors, BUILD_LOG_LEVEL_* to set the logging verbosity and XCODE_COLORING_ENABLED to enable/disable XcodeColors support.

In github you'll find Qt, iOS/XCode and Android sample projects.

How it Works

Pretty simple: wrapper functions like log_info(...) use the template class LC_Log to print the string. The LC_Log delegates actual log call to the class of type T, which can be implemented according to the needs. Delegates I currently implemented are:

  1. LC_Output2Std: outputs to standard output, adding ANSI escape codes.
  2. LC_Output2FILE: write the logs to file (no escape codes here).
  3. LC_OutputAndroid: implements logging to logcat.
  4. LC_Output2XCodeColors: implements logging using XcodeColors format.
Hope you can find this useful. If you improve this, share your work! If you find issues, report them! Bye! ;-)

Thursday, August 15, 2013

A Sample to Use for Debugging and Testing Performance of the QtMultimedia Backend

In the github project for PiOmxTextures I recently uploaded also a sample player (tools/POCPlayer) that can be used for debugging and testing the performance. This is supposed to be a sample for the performance of the QtMultimedia backend and Qt itself on the Pi. It is entirely Qt/QML based, with the GUI written entirely in QML/Javascript. It shows the common controls for a media player and the service for metadata extraction. This is a demo of the current state:

How to Build

This is simple: it is a standard Qt application, with C++/Qt and QML code. A simple build of Qt version >= 5.1 should be sufficient. There is no interaction whatsoever with the plugin during build time.

Limitations

The current version can handle images and videos. Unfortunately it seems to perform pretty good in 1080p with 720p video, but 1080p videos are quite difficult to handle when other animations are running concurrently. Animations not involving video perform really good.
What you see in the video is running on the eglfs plugin: running on some more interesting layers like Wayland would require some changes in how EGL images are handled in the QtMultimedia plugin and PiOmxTextures. Unfortunately, it seems the QtWayland module is not currently perfectly working, so a hard work might be needed.

Thursday, August 8, 2013

Raspberry Pi Wheezy Image With OpenMAX-based Multimedia Backend

Please take into consideration that this firmware is based on an old firmware and an old PiOmxTextures version.

As I see still someone has troubles building PiOmxTextures, I built a new image based on the Raspbian wheezy image released the 26th of July (http://www.raspberrypi.org/downloads). This version of the image also contains Qt version 5.1.2 (35fe76d9c343f98ec55c7cacd8c402e90d91ef38) and an OpenMAX-based Qt Multimedia plugin (PiOmxTextures). The available Qt modules are:
  • qtbase
  • qtscript
  • qtxmlpatterns
  • qtjsbackend
  • qtdeclarative
  • qtsensors
  • qtquick1
  • qtgraphicaleffects
  • qtwebkit
  • qtwayland
  • qtsvg
  • qtserialport
  • qttolls
  • qttranslations
  • qtquickcontrols
  • qtmultimedia
  • qtimageformats
The mediaservice plugin installed is the openmaxil plugin (available here: https://github.com/carlonluca/pi). It is based on ffmpeg version 2.0 and that version of ffmpeg is statically built inside the PiOmxTextures library available in /usr/lib.
Note that the wayland module and the wayland and xkbcommon libs are available in the image, but I never tested those. Also, PiOmxTextures won't work with wayland but only using the eglfs platform plugin.

How to Test

Of course all this is experimental and if you seriously intend to use in production I suggest you start working on the sources in https://github.com/carlonluca/pi. I placed a build of the POCPlayer (sources are available in the same repository under tools/POCPlayer) in the home directory of the pi user. You can run it and see how it works:

./POCPlayer [video_file]

By pressing the "l" button, you can see the possible options. The POC player is experimental as well and it is entirely implemented in QML.
You can help improve and fix both PiOmxTextures lib and the openmaxil plugin by simply replacing libPiOmxTextures in /usr/lib and/or libopenmaxilmediaplayer.so in /usr/local/Qt-rasp-5.1.2/plugins/mediaservice.

How to Use

Refer to Raspberry Pi Wheezy Image With Qt and Wayland Support for instructions on how to use. The procedure is almost identical. Then, you can simply build your QtMultimedia-based application and test.

Where to Download

Unfortunately the download is over 1GB. You'll have to download from this ed2k link. You can download it with this excellent torrent. Thanks to the author.

Please let me know if you experience troubles downloading or with the image itself. Bye!

Sunday, May 26, 2013

Cross-building ICU for Applications on Embedded Devices

There are some pretty common libraries that come very handy in some situations when developing for common embedded systems like Android and iOS. One of this is the ICU library, which is useful when you need to work on code page conversion, collation, transliteration etc...

ICU is available as library for some systems, but it is quite large (libicudata for instance is more than 23MB alone). It is possible to reduce the size considerably reading this, and rebuilding.

ICU is perfectly portable on Linux, MacOS, Android, iOS, Linux Embedded etc... The process of cross-building is very simple, but still it took me a couple of hours to build for Linux, MacOS, iOS device and simulator and Android. So, I found some scripts around and fixed those bit (maybe the originals were a little outdated). I therefore write here a couple of notes on how to do it quickly (tested on 51.1).

Download the sources

I commonly download the sources from the repo directly:

export MY_DIR=some building directory
cd $MY_DIR
svn export http://source.icu-project.org/repos/icu/icu/tags/release-51-2 icu-51.2


You might want now to modify uconfig.h or data to avoid including data which is useless for your application: http://userguide.icu-project.org/icudata.

Build for Android

Cross-building ICU requires to build it first for the system where the cross-build is run, then for the target system. So, if we're using Linux when building for Android, let's first build ICU for Linux:

cd $MY_DIR
mkdir build_icu_linux
cd build_icu_linux


and use this script to build from there (change the variables and build options according to your needs):

export ICU_SOURCES=$MY_DIR/icu-51.2
export CPPFLAGS="-O3 -fno-short-wchar -DU_USING_ICU_NAMESPACE=1 -fno-short-enums \
-DU_HAVE_NL_LANGINFO_CODESET=0 -D__STDC_INT64__ -DU_TIMEZONE=0 \
-DUCONFIG_NO_LEGACY_CONVERSION=1 -DUCONFIG_NO_BREAK_ITERATION=1 \
-DUCONFIG_NO_COLLATION=1 -DUCONFIG_NO_FORMATTING=1 -DUCONFIG_NO_TRANSLITERATION=0 \
-DUCONFIG_NO_REGULAR_EXPRESSIONS=1"

sh $ICU_SOURCES/source/runConfigureICU Linux --prefix=$PWD/icu_build --enable-extras=no \
--enable-strict=no -enable-static --enable-shared=no --enable-tests=no \
--enable-samples=no --enable-dyload=no
make -j4
make install


Now you can build for Android:

cd $MY_DIR
mkdir build_icu_android
cd build_icu_android


and use this script:

export ICU_SOURCES=$MY_DIR/icu-51.2
export ANDROIDVER=8
export AR=/usr/bin/ar
export BASE=$MY_DIR
export HOST_ICU=$BASE/build_icu_android
export ICU_CROSS_BUILD=$BASE/build_icu_linux
export NDK_STANDARD_ROOT=your toolchain root
export CPPFLAGS="-I$NDK_STANDARD_ROOT/sysroot/usr/include/ \
-O3 -fno-short-wchar -DU_USING_ICU_NAMESPACE=1 -fno-short-enums \
-DU_HAVE_NL_LANGINFO_CODESET=0 -D__STDC_INT64__ -DU_TIMEZONE=0 \
-DUCONFIG_NO_LEGACY_CONVERSION=1 -DUCONFIG_NO_BREAK_ITERATION=1 \
-DUCONFIG_NO_COLLATION=1 -DUCONFIG_NO_FORMATTING=1 -DUCONFIG_NO_TRANSLITERATION=0 \
-DUCONFIG_NO_REGULAR_EXPRESSIONS=1"
export LDFLAGS="-lc -lstdc++ -Wl,-rpath-link=$NDK_STANDARD_ROOT/sysroot/usr/lib/"

export PATH=$PATH:$NDK_STANDARD_ROOT/bin

$ICU_SOURCES/source/configure --with-cross-build=$ICU_CROSS_BUILD \
--enable-extras=no --enable-strict=no -enable-static --enable-shared=no \
--enable-tests=no --enable-samples=no --enable-dyload=no \
--host=arm-linux-androideabi --prefix=$PWD/icu_build
make -j4
make install


In the icu_build directory you should have all you need to build your new application.
Note that I didn't use the NDK here, but the standard toolchain that results from the make-standalone-toolchain.sh script in the NDK.
Also note that part of ICU is already in /system/lib in some Android devices but I don't think there is any guarantee that it will be in every device (not part of the standard Android interface) and don't know exactly what is included inside that build.

Building for iOS

The same approach can be applied to cross-build for iOS. First, I built for MacOS in this case, and then for iOS device and simulator.

cd $MY_DIR
mkdir build_icu_mac
cd build_icu_mac


The script is similar to the Linux one:

ICUSRC_PATH=$MY_DIR/icu-51.2
export CPPFLAGS="-O3 -DU_USING_ICU_NAMESPACE=1 -fno-short-enums \
-DU_HAVE_NL_LANGINFO_CODESET=0 -D__STDC_INT64__ -DU_TIMEZONE=0 \
-DUCONFIG_NO_LEGACY_CONVERSION=1 -DUCONFIG_NO_BREAK_ITERATION=1 \
-DUCONFIG_NO_COLLATION=1 -DUCONFIG_NO_FORMATTING=1 -DUCONFIG_NO_TRANSLITERATION=0 \
-DUCONFIG_NO_REGULAR_EXPRESSIONS=1"

sh $ICUSRC_PATH/source/runConfigureICU MacOSX --prefix=$PWD/icu_build --enable-extras=no \
--enable-strict=no -enable-static --enable-shared=no --enable-tests=no \
--enable-samples=no --enable-dyload=no
make -j4
make install


Then I built for the iOS simulator in:

cd $MY_DIR
mkdir build_icu_simulator
cd build_icu_simulator


with this script:

DEVROOT=/Applications/Xcode.app/Contents/Developer/Platforms/iPhoneSimulator.platform/Developer
SDKROOT=$DEVROOT/SDKs/iPhoneSimulator6.1.sdk
SYSROOT=$SDKROOT

ICU_SOURCES=$MY_DIR/icu-51.2
ICU_FLAGS="-I$ICU_PATH/source/common/ -I$ICU_PATH/source/tools/tzcode/ -O3 \
-DU_USING_ICU_NAMESPACE=1 -fno-short-enums -DU_HAVE_NL_LANGINFO_CODESET=0 \
-D__STDC_INT64__ -DU_TIMEZONE=0 -DUCONFIG_NO_LEGACY_CONVERSION=1 \
-DUCONFIG_NO_BREAK_ITERATION=1 -DUCONFIG_NO_COLLATION=1 -DUCONFIG_NO_FORMATTING=1 \
-DUCONFIG_NO_TRANSLITERATION=0 -DUCONFIG_NO_REGULAR_EXPRESSIONS=1"

export CPPFLAGS="-I$SDKROOT/usr/include/ -I$SDKROOT/usr/include/ -I./include/ \
-miphoneos-version-min=2.2 $ICU_FLAGS -pipe -arch i386 -no-cpp-precomp \
-isysroot $SDKROOT"

export LDFLAGS="-arch i386 -L$SDKROOT/usr/lib/ -isysroot $SDKROOT \
-Wl,-dead_strip -miphoneos-version-min=2.0"

sh $ICU_PATH/source/configure --host=i686-apple-darwin11 --enable-static --disable-shared \
-with-cross-build=/Users/luca/tmp/icu_build_mac --prefix=$PWD/icu_build
make -j4
make install


It works similarly for the arm device itself:

cd $MY_DIR
mkdir build_icu_device
cd build_icu_device


and this is the script to build:

DEVROOT=/Applications/Xcode.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer
SDKROOT=$DEVROOT/SDKs/iPhoneOS6.1.sdk
SYSROOT=$SDKROOT

ICU_PATH=$MY_DIR/icu-51.2
ICU_FLAGS="-I$ICU_PATH/source/common/ -I$ICU_PATH/source/tools/tzcode/ -O3 \
-fno-short-wchar -DU_USING_ICU_NAMESPACE=1 -fno-short-enums \
-DU_HAVE_NL_LANGINFO_CODESET=0 -D__STDC_INT64__ -DU_TIMEZONE=0 \
-DUCONFIG_NO_LEGACY_CONVERSION=1 -DUCONFIG_NO_BREAK_ITERATION=1 \
-DUCONFIG_NO_COLLATION=1 -DUCONFIG_NO_FORMATTING=1 -DUCONFIG_NO_TRANSLITERATION=0 \
-DUCONFIG_NO_REGULAR_EXPRESSIONS=1"

export CPPFLAGS="-I$SDKROOT/usr/include/ -I$SDKROOT/usr/include/ -I./include/ \
-miphoneos-version-min=2.2 $ICU_FLAGS -pipe -no-cpp-precomp -isysroot $SDKROOT"

export CC="$DEVROOT/usr/llvm-gcc-4.2/bin/arm-apple-darwin10-llvm-gcc-4.2"
export CXX="$DEVROOT/usr/llvm-gcc-4.2/bin/arm-apple-darwin10-llvm-g++-4.2"

export LDFLAGS="-L$SDKROOT/usr/lib/ -isysroot $SDKROOT -Wl,-dead_strip \
-miphoneos-version-min=2.0"

sh $ICU_PATH/source/configure --host=arm-apple-darwin --enable-static \
--disable-shared -with-cross-build=$MY_DIR/icu_build_mac --prefix=$PWD/icu_build
make -j4
make install


Post a comment if you find something wrong!
Of course the same approach might work similarly for other embedded devices with a proper toolchain :-)
Not too difficult, but still might speed up your work! ;-)

Wednesday, April 24, 2013

Hardware Accelerated QtMultimedia Backend for Raspberry Pi and OpenGL Shaders on Video

EDIT: This article is completely outdated. Refer to newer posts for properly working builds, instructions on how to build etc... In particular: "Using POT Builds" and "Build Procedure for PiOmxTextures".

In some previous posts I developed a custom QML component to render video in a QML scene using hardware accelerated decoding capabilities and rendering without passing on the ARM side. This resulted in good performance of the Raspberry Pi even with 1080p high profile h264 videos.

Many bugs need to be fixed, code should be refactored a little but still it shows it is possible and that it works good. So I decided to move the following step: modifying Qt to make it possible to use the "standard" QtMultimedia module to access the same decoding/rendering implementation. This would make it possible to better integrate with Qt and allow users to recompile without changing anything on their implementation.

The QtMultimedia module uses gstreamer on Linux to provide multimedia capabilities: gstreamer is unfortunately not hardware accelerated on the Pi unless you use something like gst-omx.

Thus, I started to look at the QtMultimedia module sources in Qt5 and found out (as I was hoping), that the Qt guys have done, as usual, a very good job in designing the concept, providing the classic plugin structure also for multimedia backends. Unfortunately, also as usual, not much documentation is provided on how to implement a new backend, but it is not that difficult anyway by looking at the other implementations.

Design

At the end, I came up with a structure like this: I implemented a new QtMultimedia backend providing the MediaPlayer and VideoOutput minimal functionalities leveraging a "library-version" of the PiOmxTextures sample code which in turn uses a "bundled" version of omxplayer implemented using the OpenMAX texture render component as a sink for the video.

As said, Qt guys have done a good job! I didn't have to change almost nothing of the Qt implementation; all the implementation is inside the plugin (apart from a minimal modification on the texture mapping, for some reason it was upside-down and inverted).

Results

The result is pretty good, I don't see many differences from the previous custom QML component (the decoding and rendering code is the same and the QML component is implemented using the same exact principle, so nothing really changed).
I'm only beginning to play a little bit with this, I just tried a couple of things. In the video you can see the "standard" qmlvideo and qmlvideofx examples provided with the Qt sources.

How to Build

Clone the repo somewhere, then use the prepare_openmaxil_backend.sh script in tools. It will compile PiOmxTextures as a shared lib and will place everything you need in the openmaxil_backend directory. Copy that directory recursively into your Qt source tree in qt_source_tree/qtmultimedia/src/plugins naming it simply openmaxil.

Some changes are needed to the Qt tree to make it compile the new backend automatically instead of the gstreamer backend, for the texture mapping and to make the "standard" qmlvideo and qmlvideofx examples work. No real modification to the code is needed: sufficient to instantiate the QQuickView those examples use with a specific class definition. This is needed. and to provide the plugin the instance of the QQuickWindow containing the media player.
These changes can be applied with a patch to the qtmultimedia tree using the patch in the tools in git. Then build the qtmultimedia module with:

path_to_qmake/qmake "CONFIG+=raspberry"

You'll find all you need here: https://github.com/carlonluca/pi.

How to Use

After you have built the plugin, you can simply use the "standard" Qt API for MediaPlayer and VideoOutput. Only restriction is that the plugin needs to access a QQuickView to access the renderer thread of the Qt Scene Graph. This might be an issue, but I've not found another solution to this yet.

What you have to do is to simply provide your application the QQuickView by using the exact class, which must be included in your application:


class RPiQuickView
{
public:
   static QQuickView* getSingleInstance();
};

Q_DECL_EXPORT QQuickView* RPiQuickView::getSingleInstance() {
   static QQuickView instance;
   return &instance;
}


This is needed because the plugin will look for the RPiQuickView::getSingleInstance() symbol, which should be found after the dynamic linker has linked to plugin the the executable. Also, you'll need to add -rdynamic to the LFLAGS of your application, so we ensure that the linker will add the symbol to the symbol table.

This is what I added to the qmlvideo and qmlvideofx examples to make those work. This is of course not elegant, but still I couldn't find a better way in reasonable time.

Of cuorse, you'll have to copy the Qt libraries that are built to your Pi, together with libPiOmxTextures.so (unless you build it statically) and the ffmpeg libraries (do not use the ffmpeg libs you have in your Pi, it is likely those won't work; use those compiled by the compile_ffmpeg.sh script in tools.

What Remains to Be Done

Most the calls are not implemented, just the minimal to get video on the screen. Also, still the audio implementation is missing (but OMX_MediaProcessor class should be ready to play audio as well) and only the QtQuick side is taken into consideration: I've never had the time to look at the widget implementation.

In case you find bugs, try to report an issue on github. If I'll find the time I'll answer.

Edit 6.25.2013

Instantiation of the QQuickView using the RPiQuickView class is no more needed from 30e24106c5dd7a5998d49d7093baef49f332b1d2. I tested this revision with Qt 5.1.1 and everything seems to keep working correctly.

Friday, April 19, 2013

Accelerated Video Decoding and Rendering in QML with Necessitas Qt 4 for Android

Ok, I did this a couple of months ago but I just realized it might be of help to someone who is currently using Necessitas Qt4 for some project and still cannot use Qt5.
This is a sample code which shows how to create a custom QML component in the Qt4 Necessitas porting to use hardware acceleration on any Android devices with API level at least 11. The result is pretty good, you can check the demo I uploaded on youtube a couple of months ago (the third application shown is the one which is implemented over Qt 4):



The description says it all: "The third sample code uses a custom QML component written in C++ using a Qt 4.8.2 port for Android (Necessitas). Regular QML animations are then applied to the custom component. The semi-transparent image is a regular QML Image element with alpha set to 0.5."

The code is available now here on github: https://github.com/carlonluca/TextureStreaming. The project TextureStreaming can be opened with Qt Creator and run on a device (assuming API level constraint is met).

Take into consideration that Qt guys are working on the QtMultimedia backend for Android, and I think it should be available in Qt 5.2. You might want to try that also: http://qt.gitorious.org/qt/qtmultimedia/trees/dev/src/plugins/android.

How it Works

As you can see from the code, a custom QML component is used and placed in the QML scene. That component instantiates some Java classes through JNI glue code and use the Android standard Media Player to start decoding video and playing audio. The sink is set to be a SurfaceTexture instance, which provides the OpenGL texture that the custom QML component renders in the QML scene. Result is pretty good.