But simply ignore when building outside of Travis.
Of course since we are building Android in a container, we need to first pass
the environment variable to the container...
Signed-off-by: Dirk Hohndel <dirk@hohndel.org>
These come originally from https://github.com/travis-ci/travis-build and are
available when running on Travis, but not when running inside a Docker
container on Travis as we do in order to build for Android.
The goal is to provide a quasi heart-beat on STDOUT during very long running
commands - without this the wget to download Qt often times out, so that's
where we are going to use this.
Signed-off-by: Dirk Hohndel <dirk@hohndel.org>
Move the find-moved-images functions into a new translation unit
and present the user with the identified matches before applying
them.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
In the last commits, the canonical-to-local filename map was made
independent from the image hashes and the location of moved images
was based on filename not hashes. The hashes are now in principle
unused (except for conversion of old-style local filename lookups).
Therefore, remove the hashes in this commit. This makes addition
of images distinctly faster.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
Users might have edited their pictures. Therefore, instead of identifying
pictures by the hash of the file-content, use the file path. The match
between original and new filename is graded by a score. Currently, this
is the number of path components that match, starting from the filename.
Camparison is case-insensitive.
After having identified the matching images, write the caches so that they
are saved even if the user doesn't cleanly quit the application.
Since the new code uses significantly less resources, it can be run in a
single background thread. Thus, the multi-threading can be simplified.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
The connection canonical filename to local filename was done via
two maps:
1) canonical filename -> hash
2) hash -> local filename
But the local filename was always queried from the canonical filename.
Therefore, directly index the former with the latter.
On startup, convert the old map to the new one.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
This fixes a bug introduced in fbe1144eaf:
For an empty log, in DivePictureModel::updateDivePictures()
beginResetModel() would be called without a corresponding endResetModel().
It is unclear whether this can ever be hit, because in the no-dives
case, at least in the desktop version no profile is shown.
Note, that this makes the check double-unnecessary.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
Some OSTC 2 and OSTC Plus variants show 'OSTC+ xxxxx' as BLE name and we
recognized this as OSTC 3 (but that one doesn't support BLE). With this
we recognize these models as OSTC 2 (which is identical from a download
perspective to the OSTC Plus) and both of those support BLE.
Signed-off-by: Dirk Hohndel <dirk@hohndel.org>
When we split a dive in two, we keep the dive computer ID for the dive,
but we should update the actual _time_ of the split dive to match the
split.
And when we look for "are these the exact same dives", we should check
not only that the dive computer dive ID matches, but also that the dive
computer time matches, so that we don't consider two parts of a dive
that has been split to be obviously the same dive.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Dirk Hohndel <dirk@hohndel.org>
The dive splitting was completely wrong, because we checked the time of
the previous sample by doing
sample[i - 1].time.seconds
which is entirely wrong. The 'sample' variable is the *current* sample,
so the time of the previous sample is simply
sample[-1].time.seconds
Alternatively, we could have started from the first sample, and done
dc->sample[i - 1].time.seconds
but mixing the two concepts up just gets you a random sample pointer
that is likely not a valid sample at all, and obviously does not have
the right time at all.
As a result, dive splitting was pretty much random. Sometimes it worked
purely by mistake, because the rest of the logic was right (ie we _had_
found the right point where we reached the surface in the dive etc, the
"previous sample time" was simply used to decide if the surface interval
was sufficient to split the dive up).
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Dirk Hohndel <dirk@hohndel.org>
Closes#1229
[Dirk Hohndel: minor white space adjustment]
Signed-off-by: Martin Měřinský <mermar@centrum.cz>
Signed-off-by: Dirk Hohndel <dirk@hohndel.org>
move #include prefs-macros from SettingsObjectWrapper.h to SettingsObjectWrapper.cpp
include dive.h directly (only part of prefs-macros.h used) in preference classes
Signed-off-by: Jan Iversen <jani@apache.org>
In DivePictureModel, rowDDEnd and rowDDStart specify the range of
pictures in the profile plot. Obviously, these have to be adjusted
when pictures are deleted.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
The DivePictureModel signals were connected int the setProfileState()
function of ProfileWidget2. If this function was called repeatedly
(e.g. after adding a dive), multiple connections were made, which led
to multiply delivered signals. This in turn led to confusion, when
deleting pictures (i.e. multiple pictures were deleted).
Instead, connect once in ProfileWidget2's constructor. It is unclear
whether these signals should be disconnected temporarily if no pictures
are shown. In any case, the weren't, so this shouldn't introduce any
regression.
Fixes#1441
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
The SVG icons for failed / still-loading pictures were rendered with an
alpha channel. This lead to strange behavior when hovering over the
icon in the profile plot: When hitting a "hole" the icon would be
minimized again.
Therefore, render the SVGs onto a white background.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
The clipboard fails if we attempt to copy more than 1MB of data. But the
data buffer used is shared between all transactions 'in flight' and we
cannot tell what else is currently using that buffer. Limiting ourselves
to 500k of text for the logfiles seems reasonable and hopefully makes it
more likely that the transaction will succeed (sadly, Qt doesn't tell us
if it failed).
Signed-off-by: Dirk Hohndel <dirk@hohndel.org>
On some devices (e.g., a Sony Xperia phone) the GenericDataLocation is
not app writeable. Instead of just giving app, try a few other default
locations as well (and since all of these are actually string lists,
try all of the options that Qt gives us).
Reasonably, we should only set the libdivecomputer logfile name if we
found a writeable location.
Signed-off-by: Dirk Hohndel <dirk@hohndel.org>
Adding Cressi Giotto, Newton and Drake to the list of devices
that can be selected on Android devices.
Signed-off-by: Stephen Goodall <stephen.goodall88@googlemail.com>
The info text from the download process wasn't rendered correctly.
maximumWidth ended being a recursive reference and as a result the text
would render as very narrow and super-tall field.
Signed-off-by: Dirk Hohndel <dirk@hohndel.org>
This got disabled as unintended (I hope) side effect of commit
807571a588 ("core: update deviceData default from qml").
Signed-off-by: Dirk Hohndel <dirk@hohndel.org>
Adding Cressi Leonardo to the list of devices that can be selected
on Android devices.
Signed-off-by: Stephen Goodall <stephen.goodall88@googlemail.com>
When the tooltip is hidden QInputMethodEvent is handled
by QLineEdit and works properly.
When the tooltip is visible DiveLocationListView gains
focus and receives QInputMethodEvent, without handling
them properly.
This patch just forwards the event back to the QLineEdit
object.
Fix#1405
Signed-off-by: Murillo Bernardes <mfbernardes@gmail.com>
Qt 5.11 adds useful warnings when code attempts to use anchors within
Layouts and even tells you how to fix things.
Signed-off-by: Dirk Hohndel <dirk@hohndel.org>
In order to trigger the redraw of an edited dive we need to make sure
the model realizes that it has been updated. So far the only way to make
sure this happens reliably appears to be to remove the item and
re-insert it. Seems weird, but with this the bug of not redrawing the
profile after an edit appears fixed.
Fixes#1419
Signed-off-by: Dirk Hohndel <dirk@hohndel.org>
The old trGettext() was not thread-safe and the returned C-strings
could be freed in the case of empty translations strings. Therefore:
1) Introduce a mutex protecting access to the cache.
2) Never change existing entries, even if the translation string is empty.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
There were a handfull instances of the kind
1) gettextFromC::instance()->tr(...)
2) gettextFromC::instance()->trGettext(...)
1) is pointless, as tr is a static function.
All instances of 2) were likewise pointless, because trGettext()
returns a C-string, which was then immediately converted to a
QString.
Thus, replace both constructs by gettextFromC::tr(...).
After this change there was only one user of gettextFromC::instance()
left, viz. the C-interface funtion trGettext(). Therefore, remove
gettextFromC::instance() and do all the caching / translating
directly in the global trGettext().
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
