This PR attempts a number of improvements to the sample rendering (in the song editor, automation editor, AudioFileProcessor, SlicerT, etc) in LMMS:
Thumbnails: Samples are aggregated into a set of thumbnails of multiple resolutions. The smallest larger thumbnail is then chosen for rendering during sample drawing. Each set of thumbnails is stored with its sample file metadata in an unordered_map, so that duplicate samples will use the same set of thumbnails.
Partial rendering: additionally, this PR only renders the portions of the sample clips visible to the viewer to save rendering time.
* Experimental sample thumbnail
* Rename some classes and type aliases. Make some type declarations explicit
* Use a combination of audioFile name and shared_ptrs to track samples; refactor some loops
* That weird line at the end of the sample is now gone
* Small changes to the code; Add comments
* Add missing word to comment; Fix typo
* Track `SharedSampleThumbnailList`s instead
* Major refactor; implement thumbnailing for SlicerT, AFP and Automation editor
* Code clean up, renames and documenting
* Add the namespace lmms comments
* More code updates and documentation
* Fix error in comment
* Comment out `qDebug`s
* Fix formatting
* Use alternative initialization of `SampleThumbnailVisualizeParameters`
* Remove commented code
* Fix style and simplify code
* Use auto
* Draw lines using floating point
* Merge the classes into one nested class
+ Replace while loop with std::find_if
* Fix comparison of different signedness
* Include memory header
* Fix a logic error when selecting samples; Rename a const
* Fix more issues with signedness
* Fix sample drawing error in `visualizeOriginal`
* Only render regions in view of the sample
* Allow partial repaints of sample clips
* Remove unused variable
* Limit most of the painting to the visible region
* Revert back to using rect() in some places
* Partial rendering for AutomationEditor
* Don't redraw painted regions; allowHighResolution; remove `visualizeOriginal`; Remove s_sampleThumbnailCacheMap
* Add s_sampleThumbnailCacheMap back for testing convenience
* Minor change to the way `thumbnailSizeDivisor` is calculated
* Extend update region by an amount
* forgot about this
* Adapt to master; Redesign VisualizeParameters; Don't rely entirely on needsUpdate()
* Don't try to preserve painted regions
* Allow for a bit more thumbnails; Fix incorrect rendering when vertically scrolling
* Fix missing include statement
* Remove the unused variable
* Code optimization; Remove RMS member from Bit; Rename viewRect to drawRect
* More code optimizations
* Fix formatting
* Use begin instead of cbegin
* Improve generation of thumbnails
* Improve expressiveness of the draw code
* Add support for reversing
* Fix drawing code
* Fix draw code (part 2)
* Apply more fixes and simplifications
* Undo some out of scope changes
* Remove SampleWaveform
* Improve documentation
* Use size_t for some counters
* Change width parameter to be size_t
* Remove temporary aggregated peak variable
* Bump up AggregationPerZoomStep to 10
* Zoom out only requested range of thumbnail instead of clipping it separately
* Rename targetSampleWidth to targetThumbnailWidth
* Handle reversing for AFP; Iterate in reverse instead of reversing the entire thumbnail
* Change names to be more accurate
* Improve implementation of sample thumbnail cache map
* Move AggregationPerZoomStep back down to 2, do not cap smallest thumbnail width to display width
To improve performance with especially long
samples (~20 minutes)
* Simplify sample thumbnail cache handling in constructor
* Call drawLines instead of drawLine in a loop
QPainter::drawLine calls drawLines with a line count of 1. Therefore, calling drawLine in a loop means we are simply calling drawLines a lot more times than necessary.
* Bump up AggregationPerZoomStep to 10 again
Thought using 2 would help performance (when rendering). Maybe it does, but it's still quite slow when rendering a bunch of thumbnails at once.
* Fix off-by-one error when constructing Thumbnail from buffer
* Fix crash when viewport is not in bounds
* Apply performance improvements
Performance in the zoomOut function was bad because of the dynamic memory allocation. Huge chunks of memory were being allocated quite often, casuing a ton of cache misses and all around slower performance. To fix this, all the necessary downsampling is now done within the for loop and handled one pixel after another, instead of all at once.
To further avoid allocations in the draw call, the change to use drawLines has been reverted.
We now pass VisualizeParameters by value (it is only 64 bytes, which will fit quite nicely in a cache line, which is more benefical than reaching for that data by reference to some other part of the code).
After applying these changes, we are now practically only bounded by the painting done by Qt (according to profiling done by Valgrind). Proper use of OpenGL could resolve this issue, which should finally make the performance quite optimal in variety of situations.
* Apply minor changes
Update copyright and unused functions. Move in newly created thumbnail into the cache instead of copying it.
* Use C++20's designated initializers
* Create param right before visualizing
* Fix regressions with reversing
* Fix incorrect rendering in AFP and SlicerT
* Move MaxSampleThumbnailCacheSize and AggregationPerZoomStep into implementation file
* Remove static keyword
* Remove getter and setter for peak data
---------
Co-authored-by: Sotonye Atemie <sakertooth@gmail.com>
* add templates for common geometric constants
* oops missed one
* LD_2PI -> LD_PI
i re-added the wrong constant ffs
* CamelCase names and also verify compilation without -DLMMS_MINIMAL
* C++20 stuff
Updated to account for `<numbers>` and C++20:
- Marked all `lmms_constants.h` constants with an exact equivalent in `<numbers>` as deprecated
- Removed all `lmms_constants.h` constants where no variant is currently in use
- Using `inline constexpr`
- Using `std::floating_point` concept instead of `typename`
* add lmms::numbers namespace
* Remove panning_constants.h
Moves the four constants in panning_constants.h into panning.h, then
removes panning.h.
* Use std::exp(n) instead of powf(numbers::e, n)
* Use C++ std math functions
Co-authored-by: Dalton Messmer <messmer.dalton@gmail.com>
* Use overloaded std math functions
An attempt to fix compiler warnings on some platforms
* Remove uses of __USE_XOPEN
And also update two functions I missed from the previous commit
* Missed a few
* Fix ANOTHER std math function use
Of course there's another one
---------
Co-authored-by: Dalton Messmer <messmer.dalton@gmail.com>
* Fix rendering for Vectorscope
The rendering of the Vectorscope is broken on Wayland if the size of the
Vectorscope is increased. This is caused by using a `QImage` to render
the scope traces which is then scaled up.
Introduce a new way to paint the vector scope (goniometer) which simply
paints lines or points that progressively get dimmer and which does not
make use of a QImage anymore.
It supports the following features:
* Log mode
* Zooming
* Rendering the drawing performance
* Selecting a different color for the traces
It does not support:
* HQ Mode: The new implementation provides a performance that's
equivalent to Non-HQ mode and look similar or better than the HQ mode.
* Blurring of old data
* Persistence: Might be implemented by using a factor for the dimming
Rendering of the samples/trances uses the composition mode "Plus" so
that overlapping elements will appear like adding brightness. Painting
the grid and lines is done using the normal composition mode "Source
Over" so that it simply replaces existing pixels.
Painting of the lines/points and the grids and lines is done in a
"signal space", i.e. a transform where elements in the interval of
[-1, 1] feel "natural". The text is painted in "widget space".
* Remove old implementation
Remove HQ mode and persistence. Also remove the legacy option again.
This removes the models, loading, saving and the GUI controls.
Remove all unnecessary members from `VectorView`, adjust the
constructor. Move the implementation of `paintLinesMode` into
`paintEvent`. Remove methods `paintLegacyMode` and `paintLinesMode`.
* Move colors into VectorView
Move the colors out of `VecControls` into `VectorView` as they are
related to presentation.
* Remove friend relationship to VectorView
Remove a friend relationship to `VectorView` from `VecControls` by
introducing const getters for the models.
* Remove VectorView::m_visible
Remove the unnecessary member `m_visible` from `VectorView`. It was not
initialized and only written to but never read.
* Make Vectorscope themeable
Make the Vectorscope themeable by introducing Qt properties for the
relevant colors.
The default theme gets the values from the code whereas the classic
theme gets a trace with amber color.
* Rename m_colorFG
Rename `m_colorFG` to `m_colorTrace`. Adjust the Qt property
accordingly.
Remove local variable `traceColor` from paint method and use member
`m_colorTrace` directly.
* Remove m_colorOutline
Remove unused member `m_colorOutline`.
* Fix horizontal lines on silence
Fix the horizontal lines that are rendered on silence. They seem to be
produced when rendering lines that start and end at the same point.
Therefore we only draw a point if the current and last point are the
same.
* Add some margin to the VectorView
Add some margin to the rendering of the `VectorView` so that the circle
does not touch the bounary of the widget.
* Clean up the layout of the Vectorscope
Clean up the layout of the Vectorscope. The checkboxes are not put on
top of the vector view anymore but are organized in a horizontal layout
beneath it. This gives a much tidier look.
* Maximize button for resizable instruments
Show the maximize button for resizable instruments.
Most other changes have the character of refactorings and code
reorganizations.
Remove the negation in the if condition for resizable instruments to
make the code better readable.
Only manipulate the system menu if the instrument is not resizable.
Add a TODO to the special code that sets a size.
* Fix rendering of maximized sub windows
In `SubWindow::paintEvent` don't paint anything if the sub window is
maximized . Otherwise some gradients are visible behind the maximized
child content.
In `SubWindow::adjustTitleBar` hide the title label and the buttons if the
sub window is maximized. Always show the title and close button if not
maximized. This is needed to reset the state correctly after
maximization.
* Add SubWindow::addTitleButton
Add the helper method `SubWindow::addTitleButton` to reduce code
repetition in the constructor.
* Only disable the minimize button
Disable the minimize button by taking the current flags and removing
the minimize button hint from them instead of giving a list which might
become incomplete in the future. So only do what we want to do.
* Remove dependency on MdiArea
Remove a dependency on the `MdiArea` when checking if the sub window is
the active one. Query its own window state to find out if it is active.
* Clear Qt::MSWindowsFixedSizeDialogHint
Clear the `Qt::MSWindowsFixedSizeDialogHint` flag for resizable
instruments (symmetric to the `else` case).
* Update the sub window title bar of exchanged instruments
Update the title bar of an instrument's sub window if the model changes, e.g. if an instrument is exchanged via drag & drop.
The main fix is to call the new method `updateSubWindowState` in `InstrumentTrackWindow::modelChanged`. It contains mostly the code that was previously executed in the constructor of `InstrumentTrackWindow`. The constructor now simply calls this method after it has put the constructed instance into a sub window.
With the current implementation the sub window needs to be explicitly triggered to update its title bar once the flags have been adjusted in `updateSubWindowState`. This is done with the new public method `SubWindow::updateTitleBar`. Please note that such an explicit update is not needed if the instrument windows are managed by a `QMdiSubWindow` instead of a `SubWindow`. This means that the implementation of `SubWindow` is still missing something that `QMdiSubWindow` does. However, debugging also showed that setting the window flags of the sub window does not seem to lead to an event that could be caught in `SubWindow::changeEvent`. This was found out by simply dumping the event types of all events that arrive in that method and exchanging an instrument.
The method `updateSubWindowState` uses the added method `findSubWindowInParents` to find the sub window it is contained in. The latter method should be considered to be moved into a templated helper class because it might be useful in other contexts as well.
## Technical details
If you want to experiment with using QMdiSubWindows then simply add the following method to `MainWindow` (right next to `addWindowedWidget`):
```
QMdiSubWindow* MainWindow::addQMdiSubWindow(QWidget *w, Qt::WindowFlags windowFlags)
{
// wrap the widget in our own *custom* window that patches some errors in QMdiSubWindow
auto win = new QMdiSubWindow(m_workspace->viewport(), windowFlags);
win->setAttribute(Qt::WA_DeleteOnClose);
win->setWidget(w);
m_workspace->addSubWindow(win);
return win;
}
```
Then call that method instead of `addWindowedWidget` in the constructor of `InstrumentTrackWindow`:
```
QMdiSubWindow* subWin = getGUI()->mainWindow()->addQMdiSubWindow( this );
```
You can then comment out the cast and the call of `updateTitleBar` in `updateSubWindowState` and everything will still work.
* Update the system menu
Show or hide the "Size" and "Maximize" entries in the system menu
depending on whether the instrument view is resizable or not.
* Show non-resizable instruments as normal
Show the sub windows of non-resizable instruments as normal if the sub
window is maximized because it was previously used with a resizable
instrument.
* Fix typo
* Rename updateSubWindowState
Rename `updateSubWindowState` to `updateSubWindow`.
* Compile in C++20 mode
* Fix implicit lambda captures of `this` by `[=]`
Those implicit captures were deprecated in C++20
* Silence MSVC atomic std::shared_ptr warning
Unfortunately std::atomic<std::shared_ptr> (P0718R2) is not supported by
GCC until GCC 12 and still is not supported by Clang or Apple Clang, so
it looks like we will continue using std::atomic_load for the time being
* Use C++20 in RemoteVstPlugin
* Simplification
* Add comment
* Fix bitwise operations between different enumeration types
* Revert "Fix bitwise operations between different enumeration types"
This reverts commit d45792cd72.
* Use a helper function to combine keys and modifiers
* Remove AnalyzeTemporaryDtors from .clang-tidy
AnalyzeTemporaryDtors was deprecated in clang-tidy 16 and fully removed
in clang-tidy 18
* Use C++20 in .clang-format
* Use bitwise OR
Prevents issues if any enum flags in `args` have bits in common
* replace std::pow with better performing equivalents
* revert one instance where I swapped to fastPow10f
* Negative slope instead of multiplying -1
Co-authored-by: saker <sakertooth@gmail.com>
---------
Co-authored-by: saker <sakertooth@gmail.com>
Historically, the PeakController has had issues like attack/decay knobs
acting like on/off switches, and audio artifacts like buzzing or
clicking. This patch aims to address those issues.
The PeakController previously used lerp (linear interpolation) when
looping through the sample buffer, which was in updateValueBuffer. This
lerp utilized attack/decay values from control knobs. This is not the
correct place to utilize attack/decay because the only temporal data
available to the function is the frame and sample size. Therefore the
coefficient should simply be the sample rate instead.
Between each sample, processImpl would set m_lastSample to the RMS
without any sort of lerp. This resulted in m_lastSample producing
stair-like patterns over time, rather than a smooth line.
For context, m_lastSample is used to set the value of whatever is
connected to the PeakController. The basic lerp formula is:
m_lastSample = m_lastSample + ((1 - attack) * (RMS - m_lastSample))
This is useful because an attack of 0 sets m_lastSample to RMS, whereas
an attack of 1 would set m_lastSample to m_lastSample. This means our
attack/decay knobs can be used on a range from "snap to the next value
immediately" to "never stray from the last value".
* Remove attack/decay from PeakController frame lerp.
* Set frame lerp coefficient to 100.0 / sample_rate to fix buzzing.
* Add lerp between samples for PeakController to fix stairstep bug.
* The newly added lerp utilizes (1 - attack or decay) as the
coefficient, which means the knobs actually do something now.
* Remove QTextCodec
QTextCodec was removed from Qt6 and is only available through the
Qt5Compat module.
QTextCodec was only used by the HydrogenImport plugin when importing old
Hydrogen files that were saved using TinyXML before it supported UTF-8.
HydrogenImport would use QTextCodec to try to get the current encoding
from the locale, and then use that as a best guess for interpreting the
XML data in the unspecified encoding it was saved in. None of this was
ever reliable, since the encoding of the computer that saved the
Hydrogen file might not be the same as the computer running LMMS and
importing that file.
There is no good solution here, so I decided to simply assume the old
Hydrogen files are UTF-8 encoded. The worst that can happen is someone's
ancient Hydrogen files containing non-ASCII text of some random encoding
becomes mojibake'd after importing into LMMS, which is something that
already could have happened.
* Clean up a little
Remove the member `PluginView::m_isResizable` and it's associated method `setResizable`. Turn `isResizable` into a virtual method.
The reasoning is that whether or not an effect can be resized depends on its implementation. Therefore does not make sense to have a method like `setResizable`. If the underlying implementation does not support resizing then it would not make sense to call `setResizable(true)`. So `isResizable` now describes the underlying ability of a plugin to resize. It's then up to the clients of that method to decide how to treat the result of `isResizable`, i.e. if they want to make use of the ability to resize or not.
Put the elements of the `TrackOperationsWidget` into layouts. These are:
* The grip that can be used to move tracks
* The gear icon that opens the operations menu
* The mute button
* The solo button
The grip that can be used to move tracks around is extracted into its own class called `TrackGrip`. This has several advantages:
* It can be put into a layout.
* It can render itself at arbitrary sizes by simply repeating its pattern pixmap.
* It can be used in a much more object-oriented way because it emits signals when it is grabbed and released.
* It is responsible for locally updating its cursor state.
The default cursor of the grip now is an open hand which indicates to the users that it can be grabbed. While being grabbed the cursor now is a closed hand.
## Technical details
The class `TrackOperationsWidget` now holds an instance of `TrackGrip` and provides a getter to retrieve it. This getter is used by `TrackView` to connect to the two new signals `grabbed` and `released`. The method `TrackOperationsWidget::paintEvent` now only paints the background as it does not need to paint the grip anymore.
The `TrackView` now handles the grabbing and release of the grip in `TrackView::onTrackGripGrabbed` and `TrackView::onTrackGripReleased`. Because the events and cursor states are now handled by `TrackGrip` this code could be removed from `TrackView::mousePressEvent`.
There was a comment in `TrackView` which indicated that the `TrackOperationsWidget` had to be updated when the track is moved and released because it would hide some elements during the move. The comment and the corresponding code was removed because the operations widget does not hide its elements during moves (this was already the state before the changes made by this commit).
Adjust the style sheets of the classic and default themes with regards to the `QPushButton` that's used to show the gear menu in the `TrackOperationsWidget`. The `>` has been removed because the `QPushButton` is not a direct decendent of the `TrackOperationsWidget` anymore.
### Wrapping of `PixmapButton` in `QWidget`
The PixmapButtons that are used in `TrackOperationsWidget` current have to be wrapped into a `QWidget`. This is necessary due to some strange effect where the PixmapButtons are resized to a size that's larger than their minimum/fixed size when the method `show` is called in `TrackContainerView::realignTracks`. Specifically, with the default theme the buttons are resized from their minimum size of (16, 14) to (26, 26). This then makes them behave not as expected in layouts.
The resizing is not done for QWidgets. Therefore we wrap the PixmapButton in a QWidget which is set to a fixed size that will be able to show the active and inactive pixmap. We can then use the QWidget in layouts without any disturbances.
The resizing only seems to affect the track view hierarchy and is triggered by Qt's internal mechanisms. For example the buttons in the mixer view do not seem to be affected.
If you want to debug this simply override "PixmapButton::resizeEvent" and trigger a break point in there, e.g. whenever the new size is not (16, 14).
### More layout-friendly PixmapButton
Make the `PixmapButton` more friendly for layouts by implementing `minimumSizeHint`. It returns a size that accommodate to show the active and the inactive pixmap.
Also make `sizeHint` return the minimum size hint. The previous implementation would have made layouts jump when the pixmap is toggled with pixmaps of different sizes.
## Fix rendering of maximized sub windows
### Adjustments in paintEvent
In `SubWindow::paintEvent` don't paint anything if the sub window is
maximized . Otherwise some gradients are visible behind the maximized
child content.
### Adjustments in adjustTitleBar
In `SubWindow::adjustTitleBar` hide the title label and the buttons if the
sub window is maximized. Always show the title and close button if not
maximized. This is needed to reset the state correctly after
maximization.
Remove some calls to `isMaximized` where we already know that the sub
window is not maximized, i.e. where these calls would always return
`false`.
One adjustment would have resulted in a call to `setHidden(false)`. This
was changed to `setVisible(true)` to get rid of the double negation.
The other `false` would have gotten in an or statement and thus could be
removed completely.
### Add method addTitleButton
Add the helper method `SubWindow::addTitleButton` to reduce code
repetition in the constructor.
### Other changes
Remove a dependency on the `MdiArea` when checking if the sub window is
the active one. Query its own window state to find out if it is active.
When calling `setWindowFlags` in the constructor only adjust the existing
flags with the changes that we actually want to do. It was ensured that
all other flags that have been set before still apply with this change.
