| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
on initial makeCurrent when offscreen
makeCurrent shall skip SetNSViewCmd for offscreen, i.e. refine criteria of nsViewChanged.
Previous term enforced SetNSViewCmd on initial call as lastNSViewDescr was null.
Expand first term to require an actual non null NSView.
contextMadeCurrent must avoid blocking to wait for completion of our SetNSViewCmd on AppKit.
AWT has procedures running on AppKit under certain situations,
where it issues a feedback flush on AWTEDT (from Appkit) blocking.
This in turn deadlocks our SetNSViewCmd waiting on the AppKit,
as we are blocking the AWTEDT waiting for same command.
Further avoiding other potential deadlocks, by adding a 500ms timeout.
Also clearing the lastSetNSViewCmd field post wait, regardless,
which avoid repeatitive SetNSViewCmd issuance on timeout.
Note that the SetNSViewCmd, we failed to wait for eventually gets executed.
|
|
|
|
| |
36ca7245653b1a0897f2070b9acbe0f0898f5949
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
makeCurrent() call
Command SetNSViewCmd sets NSOpenGLContext's NSView via [NSOpenGLContext setView:]
on the main-thread as enforced since XCode 11 using SDK macosx10.15, see Bug 1398.
This command is injected into OSX's main-thread @ NSOpenGLImpl.makeCurrent(long) only if required,
i.e. issued only for a newly bound NSView and skipped for surface-less or offscreen 'surfaces'.
This operation must be performed w/o blocking other tasks locking the NativeSurface on main-thread to complete.
Since [NSOpenGLContext setView:] acquires the CGLContext lock on the main-thread,
it can't be locked by the calling thread until this task has been completed.
Command issuer NSOpenGLImpl.makeCurrent(long) will not acquire the CGLContext lock if this command is pending.
contextMadeCurrent(true) cures the potential unlocked CGLContext by issuing
a whole GLContext.release() and GLContext.makeCurrent() cycle while waiting for this command to be completed in-between.
This GLContext cycle also ensures an unlocked NativeSurface.getLock() in-between,
allowing potentially blocked other tasks on the main-thread to complete and hence this queued command to execute.
Notable test provoking critical multithreading issues is com.jogamp.opengl.test.junit.jogl.demos.es2.newt.TestGearsES2NewtCanvasSWT.
Notable test exposing issues with an unlocked CGLContext is com.jogamp.opengl.test.junit.jogl.glsl.TestGLSLShaderState02NEWT.
|
|
|
|
| |
This prepares proper release of the acquired NativeSurface lock to cure the missing CGLContext lock, see followup commit.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
default test behavior
SWT and OSX's UI TK have their strict threading policy we require to comply with, e.g. see Bug 1398 lately.
It doesn't help using our own MainThread vehicle to move the unit test on the OS main thread,
as this removes potential causes of deadlocks - which we intend to find and resolve.
This patch removed using MainThread altogether from our ant unit testing recipe
as well from our manual test scripts.
Unit tests are no more executed on the 'main thread'.
SWT tests are patched to comply with SWT's UI threading policy.
We also catch violations within NewtCanvasSWT and our SWT GLCanvas
to provide same behavior on all platforms.
|
|\ |
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
blocking
Set NSOpenGLContext's NSView via [NSOpenGLContext setView:]
on the main-thread as enforced since XCode 11 using SDL macosx10.15, using Runnable SetNSViewCmd.
This operation must be performed async w/o blocking to allow
other tasks locking the NativeSurface on main-thread to complete.
Further, since [NSOpenGLContext setView:] acquired the CGLContext lock,
it can't be locked until this task has been completed.
Worst case scenario for a late [NSOpenGLContext setView:] issuance
might be corrupt initial frame(s) displayed.
Since all concurrent locking is performed within JOGL,
the unlocked CGLContext window risk is only academic.
However, if native 3rd party toolkits take share control,
we might have a situation.
+++
SetNSViewCmd is issued @ makeCurrent() now as opposed to createContext(..)
and associateDrawable(true). The latter was actually late as well,
as it also happened after makeCurrent when updating the drawable
association. It also missed setting a null NSView when detached!
release() will also set a null NSView if called after associateDrawable(false).
SetNSViewCmd will only be issued if the NSView has been changed,
i.e. first makeCurrent() or changing the drawable.
If issued, makeCurrent() will not lock the underlying CGLContext
and hence allow SetNSViewCmd to perform - see above.
+++
NSViewDescriptor class structure replaces the less convenient method 'getNSViewHandle(..)',
exposing all collected drawable characteristics as fields.
NSViewDescriptor also respects a ProxySurface's OPT_UPSTREAM_SURFACELESS mode,
which results in not using any underlying NSView - similar to OPT_UPSTREAM_WINDOW_INVISIBLE.
This change ensures that all surfaceless GL operations will not use any NSView.
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
11 _and_ its default SDK 'macosx10.15'
This patch demonstrates that using the SDK 'macosx10.11' does not cause the crash @ -[NSOpenGLContext setView:].
SDK 'macosx10.15' enforces Apple's own Cargo Cult of 'main-thread' by throwing a SIGILL signal (or SIGABRT)
- essentially an exception.
This surely renders our code officially invalid due to this policy,
i.e. we are not allowed to issue [* setView] on any non main-thread.
+++
The crash occurs independently of used Java version on Java 8 - 11,
as well as on JogAmp 2.3.2 - current master tip.
+++
The initial remedy to issue said action on the main-thread in a blocking/wait
manner has the risk to deadlock, due to
1) [NSOpenGLContext setView:] itself using a mutex (Thanks to Ken Harris's analysis)
and
(2) in case where we are 'thread hopping':
- [main-thread] Event like 'window ready' -> kick off action on EDT-thread *blocking*
- [EDT-thread] Create stuff incl OpenGLContext -> kick off setView on main-thread *blocking*
This has to be further investigated.
This crash finally has been reliably reproduced now.
|
| |
| |
| |
| | |
+ show JOGL version
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
latest tip using XCode 11
Changes to test
- Using own 'NSApplicationMain' entry to avoid the 'missing Info.plist' message!
- Configurable CLASSPATH and LIBPATH at compile time.
Note that the java.library.path is now hardcoded as well.
- Don't close stderr in test, just fflush
- Don't close the JVM after launch via 'die(env)', let it run.
- Java: Add GLEventListener RedSquareES2 to see something in action @ 30fps
- Java: GLCanvase visible bounds
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
using XCode 11 and JOGL 2.3.2 as well as current tip
This change imports '[email protected]' patch as
reported and provided in Bug 1398.
I can not execute this patch properly,
as received a error message regarding missing Info.plist at start.
|
| |
| |
| |
| | |
in concurrent async test case.
|
|/
|
|
| |
variant.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Thanks to Phil Jordan, who is debating certain build issues of JOGL with me,
we figured that certain generated header files ended up in the wrong build folder.
E.g. on GNU/Linux, jogamp_nativewindow_windows_GDI.h ended up in build-x86_64/nativewindow/gensrc/native/drm/
This happens due to the new java11 way of generating header files using javac.
Per default, all referenced files are being compiled and header files were generated
in this case for all of them using the given '-h location' argument.
Since nativewindow generates headers for the new drm functionality first (where available),
all headers of all included referenced classes ended up in this drm build location.
Subsequent javac was skipped 'thanks' to the already existing class files ;-)
This patch disables implicit class file creation when explicitly generating headers only
using the javac argument '-implicit:none'.
This is done for the whole JOGL project (nativewindow, jogl and newt).
Other JogAmp modules should not be concerned due to the lack of overspecific
platform parts ..
|
|
|
|
|
|
|
|
|
|
| |
libraries for testing
Used for testing Bug 1422 on Windows via qemu-kvm
using Mesa 19.3.2 <https://github.com/pal1000/mesa-dist-win/releases>.
Here I dropped the native libs to C:\temp and moved the C:\Windows\System32\OpenGL32.dll
out of sight by renaming it.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
native DPI toolkit aware platforms (Linux, Windows)
NEWT + NewtCanvasAWT:
Maybe create "interface ScalableSurface.Upstream {
void pixelScaleChangeNotify(final float[] curPixelScale, final float[] minPixelScale, final float[] maxPixelScale); }"
to allow downstream to notify upstream ScalableSurface implementations like NEWT's Window to act accordingly.
+++
AWT GLCanvas: Add remark where to add the potential pixel scale.
|
|
|
|
|
|
| |
fullscreen mode
WindowDriver OSX: useParent(..) -> useParentLocation(..) and add !isFullscreen() as a criteria to use the parent derived location.
|
|
|
|
| |
input if not visible.
|
|
|
|
|
|
|
|
| |
NewtCanvasSWT.SWTNativeWindow's surfaceSize in pixel units
shall only return scaled-up windowUnits using SWTAccessor.deviceZoomScaleUp(..) for !OSX
and potentially auto scaled-up pixelUnits to passthrough (OSX).
See detailed API doc to NewtCanvasSWT.newtScaleUp(..)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
compensation like 'DPIUtil.useCairoAutoScale()'
We can't use DPIUtil's 'autoScaleUp(..)' method on non-native DPI scaling platforms
as it uses a scale-factor of 1f if the higher toolkit compensates, i.e. 'DPIUtil.useCairoAutoScale()'.
Since NEWT uses X11 and GDI directly, which are not DPI scale-aware,
we have to drop the semnatics of 'DPIUtil.useCairoAutoScale()'
and merely use the actual 'deviceZoom'.
This was proposed by Marcel Au in the first place.
At least I understand these semantics by now.
+++
Additionally NewtCanvasSWT.SWTNativeWindow needs to return the 'deviceZoomScaleUp(..)'
values for returning its size in window- and pixel-units (surface).
|
|
|
|
|
|
|
| |
toString(..), added toSimpleString(..)
This also revised commit f56adf14deadd4ee8f434ea1293e27bcafdf2a90 Window API addition of 'append(StringBuilder)',
which has been renamed to 'toString(StringBuilder)' as aligned to out other classes.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Bug 1422 shows that it seems to be desired to emulate DPI scaling where
the native toolkit does not implmement the same.
On GTK, DPIUtil.mapDPIToZoom (int dpi) reads:
double zoom = (double) dpi * 100 / 96;
int roundedZoom = (int) Math.round (zoom);
return roundedZoom;
While having dpi calculated as:
dpi = 96 * GDK.gdk_monitor_get_scale_factor(monitor);
Well, this seems to exist to allow 96 dpi fixed layout to
'look' OK on high-dpi screens.
However, you get in trouble if you layout high-dpi aware,
i.e. using percentages etc.
There is one exception: If DPIUtil.useCairoAutoScale() is true, scalingFactor is 1f
and hence the scaling emulation dropped.
'DPIUtil.setUseCairoAutoScale((sx[0]*100) == scaleFactor || OS.isGNOME);'
|
|
|
|
|
|
| |
new Rectangle instance (mutable)
Also return 'this' for setter methods for chaining.
|
|
|
|
|
| |
Otherwise one would want to pause the Animator instance for the hidden GLWindow,
otherwise such animator with zero visible drawables will become a CPU hog.
|
|
|
|
|
|
|
| |
off-viewport position
Ensure it stays out of sight by moving it to 2x width/height of viewport.
Otherwise one could see the child window moving from lower-right to upper-left ;-)
|
|
|
|
|
|
|
| |
instances if 'below notification threshold'
'below notification threshold' here is simply being a child SWT Control of like a Composition or SashForm etc
where these events won't get propagated.
|
|
|
|
|
|
| |
remove all SWT listener on dispose.
Additionally print more details about the newtChild's state in DEBUG mode.
|
|
|
|
| |
building custom efficient presentations
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
'Fake invisible child window' is implemented by simply moving the window out of sight (viewport).
- orderOut0 needs to use '[mWin orderWindow: NSWindowBelow relativeTo:..' parentWindow
instead of '[mWin orderBack:..', otherwise the whole parent application gets invisible w/ SWT ;-)
- NewtNSWindow may also needs to use parent's Screen instance if moved offscreen,
as the own Screen is invalid (zero size) in this case.
- WindowDriver: Adding special treatment for 'Fake invisible child window' (tagged as such):
-- reconfigureWindowImpl: setWindowClientTopLeftPointAndSize0(..) will be called
using the viewport's max position -> out of sight.
-- screenPositionChanged: ignore the 'new' position
-- sizeChanged: ignore the 'new' size
This sensitive NEWT change set shall benefit other toolkits being used as parentWindow
besides SWT, as this behavior is the same across MacOS.
|
|
|
|
|
|
|
|
|
|
| |
d92dc518eb891f2d125a8136efd6ed603d74a6e9
We also cannot use 'mWin orderWindow: NSWindowOut relativeTo:..]' as it also removes the child-NSWindow from its parent like 'orderOut'.
Hence only use 'orderBack' to keep the relationship inplace.
Fake invisible child window is in progress,
i.e. moving it out of the overal viewport (all screens).
|
| |
|
|
|
|
| |
reconfig: Only orderOut w/ valid window-handle
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
commenting on child-window orderOut
Actual small change is to have child-NSWindow to use '[myWindow orderWindow: NSWindowAbove relativeTo:..'
instead of 'orderFront' in creation and use the simple 'orderFront' to set a top-level NSWindow visible.
Adding comment why we can't use 'orderOut' on child-NSWindow setting it invisible,
this is due to OSX 10.7 changes and testing detaching the child-window from its parent
causes havoc w/ SWT at least.
Hence we only issue 'mWin orderWindow: NSWindowOut relativeTo:..]' and the result is
having the child-NSWindow below the application.
This in turn will make it visible again when moving the application around,
as this child-NSWindow will no more follow the position.
Suggestion is to have this 'fake invisible' child-NSWindow to be moved
out of the overal viewport (all screens).
|
|
|
|
| |
independent UI interaction coding
|
|
|
|
|
|
| |
deadlocks on OSX and Windows
Essentially same code path as NewtCanvasSWT
|
|
|
|
|
|
| |
(its a property!)
Missed the SWT headless unit tests altogether ;-)
|
|
|
|
| |
pattern
|
|
|
|
|
| |
Show and Hide handling resolves TabFolder layout,
i.e. hiding the 'hidden' and showing the current tab.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
MacOS (fixes NewtCanvasSWT on SWT positioning)
Newt's OSX Window consist out of NSView wrapped up within its own NSWindow.
It's position is being set via its NSWindow's client-area position on screen (frame),
which we derive from NSView's client-area position.
When NEWT reparents into a new 'window',
on OSX it uses the parent's NSView and its NSWindow
to attach its own NSView and NSWindow as a subview and childwindow.
SWT's OSX implementation uses NSView's for each Compositor,
but an individual NSWindow is only established for the Shell (Window).
An oversight in Nativewindow and NEWT's coordinate translation:
'top-left view <-> top-left screen'
by missing the 'view <-> window' translation caused this whole issue.
The oversight occured as NEWT's 'view <-> window' translation
had no impact due to its 1-view to 1-window mapping.
Fixing the coordinate translation resolves the mess
for SWT and for potential other toolkits on OSX.
NewtCanvasSWT behaves same on OSX as on X11 etc finally.
|
| |
|
|
|
|
|
|
|
|
|
| |
b91c680fb93a03720ff9fcb39cf052cfe8d40e76
See commit b91c680fb93a03720ff9fcb39cf052cfe8d40e76
To support static libraries JEP 178, we have to provide JNI_OnLoad_<libname> etc.
Hence change and add function entries accordingly.
|
|
|
|
|
|
| |
This also demonstrates that on non MacOS, SWT's scale-factor is artificially imposed
on the actual pixel-units, as SWT's getClientArea() is simply returning:
DPIUtil.autoScaleDown(getClientAreaInPixels())
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
getClientArea() on MacOS produces a 'difficult' result regarding the position,
which usually is returned as zero.
Using a zero position issues the bug w/ SashForm, where the offset doesn't seems
to be covered by the native NSView nor an SWT parent Composition.
Then using the getLocation() as is (i.e. the view's frame position)
may also cause issues with the TabFolder, as it includes the tab's trimming.
Here the native NSView 's position includes the tab's trimming,
gladly the parent (TabFolder or a Composition)'s clientArea includes this offset.
Therefor, as a testbed - on OSX, getClientArea2(..) returns
- position: getLocation() - getParent().getClientArea().position
- size: getSize()
This at least works OK'sh using
- no special layout parent
- TabFolder
- SashForm
++++
Unit test TestGLCanvasSWTNewtCanvasSWTPosInTabs: Adding 'addComposite' to test matrix.
'addComposite' wraps our GLCanvas or NewtCanvasSWT into a Composite instead of
adding it directly into the layouting parent.
It demonstrates an issue with the new test 'test32_NewtCanvasSWTTabSashGLWComp',
i.e. the NewtCanvasSWT is shown on the left as the SashForm's offset is being dropped.
Summary:
- No more issues with High-DPI pixelScale observed!
- GLCanvas is being most well layouted, no issues in tests
- NewtCanvasSWT may show severe positioning issues -> test32_NewtCanvasSWTTabSashGLWComp
- NewtCanvasSWT always shows a small positioning offset into the lower-right corner w/ overlapping
- NewtCanvasSWT overall positioning is not perfectly understood
- NewtCanvasSWT misses to hide the NEWT child when changing tabs in TabFolder
|
| |
|
| |
|
|
|
|
| |
'setNEWTChild(..)'
|
|
|
|
| |
position
|
|
|
|
| |
Due to issues on MacOS, added the src.zip for MacOS as well.
|
| |
|
|
|
|
|
|
| |
TestGLCanvasSWTNewtCanvasSWTPosInTabs (2/2)
Complete merging unit tests.
|