We have to lock Mixer when touching Song's state via Song::startExport()
and Song::stopExport() in ProjectRenderer::run() as the FIFO writer
thread may call Mixer::renderNextBuffer() (which calls Song::doActions())
simultaneously. Fixes a random segfault when exporting project.
This was a new bug as the ProjectRenderer does not operate FIFO-less
anymore.
* auto-save:
don't change the current project when auto-saving
auto-saves every minute and recovers upon crash
don't show WelcomeScreen when importing/loading
Conflicts:
src/gui/MainWindow.cpp
There's no need to start the ProjectRenderer thread with high priority
anymore as the actual rendering is done on the other side of the FIFO.
The ProjectRenderer just waits for new data in the FIFO and encodes
them as they arrive.
Renamed the ProjectRenderer::OutputSettings structure to
ProjectRenderer::EncoderSettings to better reflect its meaning.
Additionally added some basic Doxygen comments.
Until now, Mixer not only was responsible for rendering audio buffers
but also managed writing them to audio backend (through a FIFO) and
handled various quality related parameters.
All this functionality has been moved into the new AudioOutputContext
class. It glues together AudioBackend (formerly called AudioDevice),
global quality settings and the Mixer.
The AudioOutputContext class creates a FifoWriter which calls
Mixer::renderNextBuffer() and writes the output into the BufferFifo of
the AudioOutputContext it belongs to. The BufferFifo is read by the
according AudioBackend which belongs to the AudioOutputContext as well.
The AudioOutputContext also handles resampling in case the AudioBackend
wants the buffers in a different samplerate.
During this rewrite the Mixer class and the according source files have
been renamed from "mixer" to "Mixer". This results in small changes
all over LMMS' code base.
This reverts commit c517f1fa5a.
The commit was not very helpful and introduced new xrun problems.
Instead I'll be rewriting the part of LMMS where Mixer, Mixer's
quality settings, Buffer FIFO and AudioDevice are sticked together.
There's really no need to allocate a buffer each period, push it to the
FifoBuffer and free it when fetching the buffer in
AudioDevice::getNextBuffer(). Instead keep the pointer in FifoBuffer's
pool and reuse it.
Try to adjust latency of PulseAudio according to our settings of mixer
so it does not have such a bad latency anymore.
Furthermore force PulseAudio to play silence instead of rewinding
streams in case of underruns.
There have been some problems with the threading logic in the
AudioPulseAudio backend resulting in an endless loop when quitting LMMS.
Furthermore allocated PulseAudio resources were not freed properly.
Thanks to Armin Kazmi <armin.kazmi@tu-dortmund.de> for pointing out
this issue.
every time auto-save ran, it would change the current project to
"recover.mmp". Now it doesn't do this because Song has
guiSaveProject(), guiSaveProjectAs(), and saveProjectFile().
(the latter is used for auto-save)
Changes in base directory were not tracked by filesystem watcher.
Explicitely add base directory to filesystem watcher and add a special
case to LocalResourceProvider::reloadDirectory().
When customizing background color of a widget via CSS, Qt somehow
falls back to ugly old windows style. This resulted in ugly looking
scrollbars in WelcomeScreen when working at a low screen resolution.
Therefore disable those scrollbars as they're rather useless anyways.
QListView automatically shows three dots for too wide text items.
auto-save time is not configurable yet. saves "recover.mmp" to
WORKING_DIR every 60 seconds. Deletes recover.mmp on successful
close of LMMS. If recover.mmp is found upon start, it loads that
project.
In AudioPulseAudio::streamWriteCallback() we're operating on pcmbuf
via an operation from CPU class. All methods of CPU assume the buffer
being aligned on 16 byte boundaries. However pa_xmalloc() & friends do
not allocate aligned memory which resulted in a crash when calling
CPU::convertToS16() (SSE2 version takes advantage of aligned memory).
Replacing pa_xmalloc()/pa_xfree() with CPU::memAlloc()/CPU::memFree()
fixes this crash.
Closes bug #2890465.
Rewrote implementation of class SideBar to use QToolBar instead of
KMultiTabBar. We can style the SideBar now easily via CSS and do not
have to ship 3rd party KDE code with LMMS. Also the QToolBar based
SideBar integrates much better into the according widget style.
Furthermore renamed SideBar related classes and files to match new
coding style.
Still QString::toAscii() & friends were used in several places causing
problems such as failed access to files with non-ASCII characters in
name.
Closes#2884115.
Functionality from QuickLoadDialog which can be used in similiar dialogs
for selecting resources has been moved into generic and highly flexible
base class ResourceSelectDialog.
Again there's no need for duplicated functionality. Adding a
ResourceAction::defaultTrigger() method allows to separate all
functionality from ResourceBrowser and possibly use it in other places
as well.
There's absolutely no need for duplicating the Action enumeration
in ResourceBrowser as there's already one in ResourceAction class.
Replacing the enumeration with ResourceAction's one makes code clearer
and reduces redundancy.
Added another buffer operation BufMixCoeff allowing to mix a certain
buffer to another at a given amount (coeff). CpuX86 has been extended
by an according SSE implementation.
This should resolve the remaining issues. OSS has auto-complete with
the filesystem. Alsa has an editable combobox for PCM and MIDI-raw.
The DEVICE box for Alsa MIDI-seq should be obsoleted
JobQueues can now operate in JobQueue::Static and JobQueue::Dynamic mode.
In static mode it operates the way it always used to while in dynamic
mode a changing job queue is supported. This is particularly important
for FX mixer sends.
There were also heavy improvements regarding the overall structure
and functionality of MixerWorkerThread and MixerWorkerThread::JobQueue.
There's now a clean distinction between multi-threaded processing
and actual (thread-safe) job queue processing. MixerWorkerThread does
not need to be a friend class of Mixer anymore.
Exit the outer loop of processJobQueue() Only as soon as we went through
the job queue without having processed at least one job.
In MixerWorkerThread::waitForJobs() re-introduced "pause" instruction on
x86 and x86_64 giving better performance on HyperThreading systems.
These improvements however still do not solve all performance and
multithreading issues.
Calling FxMixerView::refreshDisplay causes LMMS to crash when running
in console mode. Therefore explicitely check GUI mode before calling
this function.
Declarations and implementations of MixerWorkerThread and ThreadableJob
have been moved into separate source files.
Furthermore there were some improvements to MixerWorkerThreads.
MixerWorkerThread::processJobQueue() does not return until the job
queue completely has been processed. This way each thread can "help"
to finish processing the queue and does not get back to sleep until
all of the work is done.
Management of the queue is now done via an array of QAtomicPointers.
Items that are non-NULL still need to be processed while NULL-items
were taken from the queue (i.e. in progress or done). Thus we do not
need to deal with ThreadableJob-states within MixerWorkerThread anymore.
The FxMixer now uses job threads to accomplish its mixing. It's
theoretically efficient, but there is a horrible thread bug
preventing the code from working. I've spent 5 hours debugging and
need some external help!
In Mixer, the old C-macro based code has been replaced by an OOP-like
design. Management of job queue now happens via some static member
methods of MixerWorkerThread. All the moved code still needs to be
splitted into some new files but here's a first dirty version.
All objects that are intended to be processed by MixerWorkerThreads
have to inherit ThreadableJob (name of class is subject of change).
One can add jobs to the job queue even if the queue is already being
processed. This is merely important for multithreading with upcoming
FX sends support.
Previously, the extensions were stored in the translations, this i18n
more difficult, additionally, there was a preprocessor statement that
would cause the translated phrase to change! Hence, no translations if
you didn't have ZIP file support. The text is now split into seperate
translations and I added "All Files".
