Up to now, we passed a "shiftPressed" flag to the individual
selection functions. To be more general replace by a struct
with "shift" and "ctrl" flags.
While doing this:
1) Move the struct into a new statsselection file for better
encapsulation.
2) Change shift to control in the scatter series, since individual
selection of items is usually done with control, not shift.
Shift usually means "select range".
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
Allow the user to select regions of the scatter plot using
a rectangular selection. When shift is pressed, do an
incremental selection.
Unfortunately, the list-selection code is so slow that this
becomes unusable for a large number of selected dives.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
Somewhat improve selection mechanics in the scatter-plot by
allowing additional selections with shift-clicking. When the
dives under the mouse are already selected, then deselect them.
This appears to be a rather common UI idiom in desktop
applications.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
As a visual feedback, show the selected dives in the scatter
plot. React to application-wide selection changes. Currently,
the dive list is deactivated while in statistics mode, but
that may change.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
When clicking on items in a plot, select the corresponding
dives. This can be useful for data validation.
Signed-off-by: Robert C. Helling <helling@atdotde.de>
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
It is crucial to replot the statistics when dives are
added / removed, to avoid stale pointers.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
If we want to make bar charts selectable (when clicking on a
bar select the dives the bar represents), then we must store
the dives behind bars. Therefore, use dive-based bins instead
of count based bins in bar charts and pie charts. This gave
some churn because every structure where a count is stored
has to be changed to store a vector of dives. Try to use
move semantics where possible to avoid duplication of dive
lists.
On a positive note, the count_dives() function of the
binners can now be removed, since it is unused.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
Placing labels at half-integer values gives horrible
rendering artifacts. Therefore, always round to integer
values. The easiest way to do this is right before setting
the position. Introduce a helper function to round QPointF
in such scenarios.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
When reparenting the statistics widget, QtQuick deletes
the rootNode and all the child nodes. It is unclear whether
this is a bug or intended behavior. In any case, it means
that the pointers to QSG nodes in the chart items become
stale.
To avoid this, delete all chart items in the root node's
destructor, before QtQuick can do anything. It is unclear
from which context this is called (render or UI) and whether
this is even valid. In some tests, it seemed to work.
The difficulty is that all the stale pointers to chart items
have to be deleted as well. All in all, the QSG memory
management is a big nuisance and very brittle.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
The updatePaintNode() function, which is run on the render
thread detected a geometry change and initiated recalculation
of the chart layout.
This means that plotAreaChanged() was called in two different
thread contexts, which is questionable. Instead, hook into
the geometryChanged() function and recalculate the chart items
there.
This fixes a rendering bug, because the old code would first
delete unneeded items and then rerender the chart. Thus, old
grid and tick items were still visible.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
While this didn't appear to be needed when dragging the legend with a
mouse, on a touch screen for some reason the drag ended after 30 pixels
either way horizontally (but no apparent limit vertically). By setting
this flag to true, drags on a tablet appear to work as expected.
Signed-off-by: Dirk Hohndel <dirk@hohndel.org>
Most colors were already collected there, but a few were dispersed
throughout the source files.
For future themeability, move the remaining colors to this common
place.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
This is not perfect - the polygon of the confidence area is
calculated even if it is not shown. Oh well.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
Up to now, when the user changed the visibility of chart features
(legend, quartiles, labels, etc.) the whole chart was replot.
Instead, only change the visibility status of these items.
After all, this modularity is one of the things the conversion
to QSG was all about.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
The code was wrong, because it deleted the ChartItems in the
main UI thread, not the render thread. This would delete the
QSG nodes in the UI thread and then crash on mobile.
Therefore refactor this part of the code by adding the
items to be deleted to a list that will be deleted by the
render thread.
As a drop in replacement of std::unique_ptr, implement
a silly ChartItemPtr class, which auto-initializes to null.
This turns the deterministic and easily controlled memory
management into a steaming pile of insanity. Obviously,
this can be made much more elegant, but this has to do for now.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
This one is trivial, since everything is there already:
Replace the QGraphicsSimpleTextItem with a ChartTextItem.
Only few functions have to be renamed.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
All series are converted to QSG. Thus, the pointer to the
QGraphicsView can be removed from the common base class.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
Since there are no disk-segment QSG primitives (one could draw
a triangle fan, but that doesn't seem optimal), this draws
into a pixmap and blits that as a QSG node.
Since this is the only series without axis, it needs a function
that returns the size of the plot area. This didn't exist, so
add it.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
Render the confidence area and the regression line into a pixmap
and show that using a QSGNode.
It is unclear whether it is preferred to do it this way or to
triangulate the confidence area into triangles to be drawn by
the shader.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
So far the items to be recalculated in the drawing thread
had a "dirty" flag and were kept in one array par z-level.
Once the series are implemented in terms of QSGNodes, there
may lots of these items. To make this more efficient when
only one or two of these items change (e.g. highlighting due
to mouseover), keep the dirty items in a linked list.
Of course, this makes the draw first version of the chart
less efficient.
There are more fancy ways of implementing the double-linked
list, but the few ns gained in the render thread are hardly
worth it.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
Render the labels and the title into a pixmap and render
the ticks and the base line using individual QSGNodes.
Attempting to render the ticks likewise into the pixmap
gave horrible results, because (quite obviously) rendering
with QPainter and the QSG shader gives non-matching ticks
and grid lines.
The memory management had to be changed a bit: The ChartItems
were collected in the root QSGNode. However, the axes are added
before the first plotting, so this node might not exist.
Therefore, store the axes in the StatsView object.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
Turn the background grid into QSGNodes. Each grid line is
represented by a QSG line item. An alternative would be
drawing the grid into a QImage and blasting that onto the
screen. It is unclear which one is preferred.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
Currently, the background was drawn as solid color onto
the chart-scene. This is of course incompatible with doing
the grid as QSGNodes. Therefore, make the scene image
transparent and use a QSGRectangle as background color.
We could also simply omit the background and show the
widget's background. However, that would mean setting
the background color in two seperate code paths
(desktop and mobile). I found no way of directly setting
the background of the QQuickItem.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
Slowly converting the QGraphicsScene items to QSGNodes to
avoid full replot of the scene.
This adds a new abstraction for line-nodes. Since the render()
function here is fundamentally different from the pixmap-nodes
we had so far, this has to be made virtual.
Also, move the quartile markers to their own source file,
since the StatsView source file is quite huge already.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
The position of the legend was reset when resizing. This was
OK as long as the legend wasn't movable.
To avoid resetting the position, store the center position
of the legend relatively to the size of the canvas. On
resize restore the center to the same relative size.
To avoid code duplication, move the sanitizing of the
coordinates from the StatsView to the Legend.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
The chart items were drawn in order of creation. To control this,
add a notion of Z-value. In contrast to QGraphicsScene, make
this a small integer value.
To controll order of drawing, a plain QSGNode is created for
every possible Z-Value and items are added to these nodes.
Thus, items are rendered by Z-value and if the Z-value is equal
by order of creation.
Likewise split the list of chart-items into Z-values, so that
items can be quickly unregistered: The items that will be
removed individually will usuall be part of Z-levels with only
few items (e.g. legend, infobox). Z-levels with many items
(notably the series) will always be fully rebuilt.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
A small step in converting from QGraphicsScene to QQuickItem.
This is the second item to be converted (after the legend)
and for now items are drawn in order of creation, which means
that the infobox is on top of the legend. This will have
to be made deterministic in follow-up commits.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
The series were passed a pointer to the QGraphicsScene to add
their item. In the future these items will be replaced by
QSGNodes. To add these, the series need a reference to the StatsView.
Therefore pass it in the constructor. Once everything is
replaces by QSGNodes, remove the QGraphicsScene member.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
Catch mouse move events and move the legend accordingly.
Currently, this is the only item that can be dragged and
therefore there is no need of doing some kind of fancy
interface. Simply keep a pointer to the legend if it is
dragged.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
In order not to waste CPU by constantly rerendering the chart,
we must use these weird OpenGL QSGNode things. The interface
is appallingly low-level and unfriendly.
As a first test, try to convert the legend. Create a wrapper
class that represents a rectangular item with a texture
and that will certainly need some (lots of) optimization.
Make sure that all low-level QSG-objects are only accessed
in the rendering thread. This means that the wrapper has
to maintain a notion of "dirtiness" of the state. I.e.
which part of the QSG-objects have to be modified.
From the low-level wrapper derive a class that draws a rounded
rectangle for every resize. The child class of that must then
paint on the rectangle after every resize.
That looks all not very fortunate, but it displays a
legend and will make it possible to move the legend
without and drawing operations, only shifting around
an OpenGL surface.
The render thread goes through all chart-items and
rerenders them if dirty. Currently, on deletion
of these items, this list is not reset. I.e. currently
it is not supported to remove individual items.
Only the full scene can be cleared!
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
I was coninced that that rather than doing an order of
magnitude estimate of the confidence region it's better
to have the correct concave shapes that indicate the
95% confidence level for the regression line.
It also turned out that the previous expression was
missing a factor of 1/sqrt(n).
Signed-off-by: Robert C. Helling <helling@atdotde.de>
The goodness of fit of a regression line is the percentage
of the variance of the y values that is explained by the
dependence on the x values.
Set the alpha value of the regression line to this goodness
of fit.
Further, set the width of the regression line to a standard
deviation of the values from the regression line valies.
Signed-off-by: Robert C. Helling <helling@atdotde.de>
The old code didn't consider that labels can peak out of
horizontal axes if labels are under ticks.
This commit takes this into account. However, it must be
noted that this is only heuristics: Before setting the
size of the axes, the actual minimum and maximum label are
not known, because we round to "nice" numbers. But the
size of the axis can only be set after knowing the overhang,
leading to a circular dependency. Therefore, the code
currently simply uses the minimum and maximum value of
the data, hoping that the "nice" values will not format
to something significantly larger. We could do a multi-pass
scheme, but let's not for now.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
It turns out that the wrong base class was used for the chart.
QQuickWidget can only be used on desktop, not in a mobile UI.
Therefore, turn this into a QQuickItem and move the container
QQuickWidget into desktop-only code.
Currently, this code is insane: The chart is rendered onto a
QGraphicsScene (as it was before), which is then rendered into
a QImage, which is transformed into a QSGTexture, which is then
projected onto the device. This is performed on every mouse
move event, since these events in general change the position
of the info-box.
The plan is to slowly convert elements such as the info-box into
QQuickItems. Browsing the QtQuick documentation, this will
not be much fun.
Also note that the rendering currently tears, flickers and has
antialiasing artifacts, most likely owing to integer (QImage)
to floating point (QGraphicsScene, QQuickItem) conversion
problems. The data flow is
QGraphicsScene (float) -> QImage (int) -> QQuickItem (float).
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
The coordinates of these were calculated when creating the feature.
This is wrong, because the min/max values of the axes can change
on resize to get "nice" number. Therefore, recalculate after resizing.
This means that the general "LineMarker" class has to be split into
two classes, one for regression lines and one for median/mean
markers.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
Easy enough to implement, but one weirdness:
To get the height of the rotated text, one has to access the
width() member of the boundingRect. I'm not sure if that makes
sense, but so be it.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
Replace by custom implementation, with the ultimate goal to
remove the QtCharts module. This doesn't yet display axis
titles or a grid.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
Signed-off-by: Dirk Hohndel <dirk@hohndel.org>
Since we want to get rid of QtCharts, we have to render our own
title. Simply keep around a QGraphicsSimpleTextItem and put in
the center of the chart. Define the borders to the scene as
constants.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
Signed-off-by: Dirk Hohndel <dirk@hohndel.org>
The chart was passed as argument to the function recalculating
the axis labels. Instead, pass the chart in the constructor of
the axes and save it. This gains us flexibility for the future:
There will be more functions that need to access the chart (e.g.
resizing of the axes).
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
We don't really give a user visible error message which is kind of a problem,
but at least we don't crash anymore.
Signed-off-by: Dirk Hohndel <dirk@hohndel.org>
A steep regression line would shoot out of the chart. Therefore,
clip to the y = minY and y = maxY lines.
QtGraphicsScene has its own clipping routines, but they are
very general, so let's do this trivial case by hand.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
For better visual guidance, format labels as "count (percentage)"
in horizontal bar charts. In vertical bar charts two lines are used
anyway.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
The old ways was to select the chart first, then depending on
the chart choose the binning.
Willem says that it should work the other way round: select
the binning (or operation) and make the charts depend on
that.
I'm not arguing one way or the other, just note that the new
way is much more tricky, because it is easy to get unsupported
combinations. For example, there is no chart where the
first variable is unbinned, but the second axis is binned
or has an operation. This makes things distinctly more tricky
and this code still needs a thorough audit.
Since this is all more tricky, implement a "invalid" chart
state. Ideally that should be never shown to the user, but
let's try to be defensive.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
The StatsView shows the chart described by the StatsState structure.
It is based on a QML ChartView. This should make it possible to
easily port to mobile. It does not include any of the UI around
the chart, viz. the variable and chart selection, etc.
The code checking for the statistical significance of the regression
line was written by Willem.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
Signed-off-by: willemferguson <willemferguson@zoology.up.ac.za>
