Hi Jim,
On 31.01.2014 2:32, Jim Graham wrote:
Hi Anton,
I think I'm getting a little lost in all of the details that I'm only on the periphery of, but it
sounds like if that had been the original performance we had seen then we might not have gone down
the path of using/forcing SW/BufferedImage?
This is a good point. However Sergey mentioned an issue in Netbeans, where switching (for some
reason) a JViewport to a backingstore (a buffered image back buffer) mode leads to a drastic
performance drop on OSX. I didn't yet have chances to get into the details, but this looks just like
the issue in question.
Anyway, in the 2D code there's an obvious bottleneck - reading a texture into SW surface with
glReadPixels by scanlines.
To be clear, this is embedding a Swing hierarchy into an FX scene?
Right.
I forget, is FX forced into SW mode when this is used? Or are we reading out the pixels in Swing
only to put them back into a texture when they get to FX?
It isn't forced. Yes, we put back SW pixels to a texture on the fx side. This is the original,
probably straightforward & simple, design of the swing/fx interop.
I think I mentioned the RSL (Resource Sharing Layer) that Dmitri and Chris had created which
allowed external libraries to get at the texture IDs used by Java2D hw pipes, and at one time we
were using that to run FX hw acceleration, but I think once we went with our own toolkit (Glass),
we had to own our own contexts and then I think RSL broke. I wonder if we can reintroduce
something similar here, or are we using completely different (and incompatible) contexts now
between FX and J2D?
Unfortunately, I can't answer this question now. As I already wrote, we have in mind the "unified
rendering" project, which is aimed at exactly this - exchanging pixels on the gl/d3d level.
But so far we have the sw-based interop. And I'd like to undestand what I'm doing with the sw-based
Retina support. The fix is ready. It has a couple of concerns, 1) your last suggestion to migrate
the "scale" logic from OffscreenHiDPIImage to OffscreenImage (I'm working on it) 2) a correction to
the device detection logic is required.
My understanding is that, unless the fix is absolutely irrelevant (whic I hope it isn't), we should
integrate it into 9/8u20 to support SwingNode in its current implementation on Retina displays.
What do you think?
Regards,
Anton.
...jim
On 1/28/14 7:35 AM, Anton V. Tarasov wrote:
Hi Jim,
What I have so far. I've implemented flipping of the image on the native
side. I didn't find any means by wich GL allows me to flip an image once
it has been rendered to a texture, unless I set up a transform matrix
prior to the render (just like in J2D). So I did that with an SW image
right after I fetched it from the texture.
At first, this works just fine in terms of correctness of the picture I
eventually see on the screen. At second, the perceived performance is
quite not bad. So, the native flip appeared to work pretty fast.
I tried to get some scores. In average, with volatile images it worked
from 3 to 20 times slower than with buffered images. The worst result is
with scrolling and this is perceptibly (when I scroll really fast up and
down, I can see some delay, but it's subtle). All the other scenarios I
tried (text rendering, 2D animations) performed visually really similar
to the buffered case. At least, the results are far from what it was
before, when it worked close to freeze. So, I would say it now looks
acceptable from the first glance.
What is interesting is that blitting a GL surface to an SW surface
spends a good time around glReadPixels, up to 3 times greater. Probably,
we can do something with it as well.
Additionally, I ran the bouncing balls app. It works 20% faster with
buffered images, but consumes 30% more CPU. So, in total, the volatile
version wins here.
And also. The issue with JViewPort (which is forced to a buffered image)
still exists (however, I have a straightforward solution and looking for
a better one). As well as the issue with Nimbus which creates buffers
based on the topmost BufferedImage. (May I use a volatile image as the
topmost buffer? Didn't try yet).
My conclusion is that, with the improvemtn, at least, the user may
consider the volatile mode as an alternative choice.
What are your thoughts?
Thanks,
Anton.
On 25.01.2014 5:50, Jim Graham wrote:
Hi Anton,
I think the main question is how fast is it compared to forcing a
software buffer? It may be slower than a straight read, but is that
slow enough that we need to use a sw buffer instead?
...jim
On 1/24/14 6:46 AM, Anton V. Tarasov wrote:
Hi Jim,
As I wrote in RT-30035, I tried that on the java side (actually, I tried
to emulate the set of operations needed to turn the image over). This
increased the perf by ~10 times, however this was still ~10 slower than
a simple read. But, I think I can try the following as well:
1) to do the turn natively (not sure if it's much faster)
2) to look if OGL is able do the turn in vram.
Thanks,
Anton.
On 24.01.2014 2:08, Jim Graham wrote:
Hi Anton,
Could the upside-down nature of the pixel readback be solved by
reading the entire frame in a single operation and then swapping the
pixels around in our own code? The memory movement may be faster than
the overhead of H calls to glReadPixels. (In the long run, we might
want to teach the rest of our code how to deal with upside-down image
data as well and then we don't have to swap it...)
...jim
On 1/21/14 5:29 AM, Anton V. Tarasov wrote:
Hi all,
Let me please resume the review process.
With the new webrev:
http://cr.openjdk.java.net/~ant/JDK-8029455/webrev.3
I'm addressing the last concern which was about "leaking" the internal
OffscreenHiDPIImage to the public via
RepaintManager.getOffscreenBuffer.
The explanation will follow, but before that I'd like to share the
info
related to the volatile buffer performance issue (which I was talking
about before). I did some investigations of the native side and
figured
out the real source of the performance drop. It's the code where
glReadPixels is called to read _every_ scanline of the image. This
is so
because of the nature of the OGL coordinate space which is upside down
comparing to the j2d space. Please find more details here:
https://javafx-jira.kenai.com/browse/RT-30035?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=380146.
If I'm not mistaken, we can do nothing about it.
So, Swing/Interop can't use a volatile image as a back buffer, and it
should use a buffered image or no back buffer at all. (Otherwise,
performance is not acceptable).
Now, to the fix. What I did is I added a "hidpiEnabled" property to
the
OffscreenHiDPIImage class. When it's "true" (by default) the image
returns its size in layout space (just like in the previous version),
when it's "false" the image returns its size in physical space (just
like an ordinary BufferedImage). In RepaintManager I set the image
"hidpi disabled", thus hiding its layout size from the developer . The
property is taken into account in SunGraphics2D.isHiDPIImage().
Because
an OffscreenHiDPIImage with hidpiEnabled==false is drawn as an
ordinary
image, I draw it via drawImage(img, x, y, width, height) in
RepaintManager.
Why I still use the OffscreenHiDPIImage class instead of a
BufferedImage
is because otherwise I'd have to do pretty the same coding, but it
would
be scattered. The class basically incapsulates the following logic:
- Keeps layout size of the image. (Used in drawImage.)
- Keeps the scale factor. (Used by
SurfaceData/VolatileImage/GraphicsConfig.)
- Overrides SurfaceData.getDefaultScale. The point is that I can't
simply call Graphics2D.scale(s, s) as this won't work when Swing draws
into the image. (SunGraphics2D asks SurfaceData for the scale).
- Overrides VolatileImage to make it return a scaled BI as its backup.
(Used by Nimbus.)
- Overrides GraphicsConfiguration to let it access the BI and its
scale
factor. (Used by Nimbus. This could be implemented otherwise, but not
much better).
No more changes in the current version. I know there're some concerns
related to detecting a device change, but let me address it after we
come to agreement about the approach discussed above.
Thanks,
Anton.
On 18.12.2013 5:03, Jim Graham wrote:
Hi Anton,
javax.swing.RepaintManager.getOffscreenBuffer is a public method that
can now return one of the new HiDPI offscreen images which is a
subclass of BufferedImage. This was what I was worried about in
terms
of one of these internal double buffers leaking to developer
code. If
they test the image using instanceof they will see that it is a
BufferdImage and they may try to dig out the raster and get
confused...
...jim
On 12/17/13 10:21 AM, Anton V. Tarasov wrote:
Hi all,
Please look at the new version:
http://cr.openjdk.java.net/~ant/JDK-8029455/webrev.2
It contains the following changes:
- All the scale related stuff is moved to the new class:
OffScreenHiDPIImage.java
- JViewport and RepaintManager no longer cache buffers.
- JLightweightFrame has new method: createHiDPIImage(w, h).
- JViewport, RepaintManager and AbstractRegionPainter goes the new
path
to create a HiDPI buffered image.
- A new internal property is added: "swing.jlf.hidpiImageEnabled".
False
by default. It makes JLF.createImage(w, h) forward the call to
JLF.createHiDPIImage(w, h). This can be used by a third party
code in
case it creates a buffered image via Component.createImage(w, h) and
uses the image so that it can benefit from being a HiDPI image on a
Retina display.
For instance, SwingSet2 has an animating Bezier curve demo.
Switching
the property on makes the curve auto scale smoothly. Please, look
at the
screenshots:
-- http://cr.openjdk.java.net/~ant/JDK-8029455/RoughtCurve.png
-- http://cr.openjdk.java.net/~ant/JDK-8029455/SmoothCurve.png
- SunGraphics2D now draws a HiDPI buffered image the same way it
draws a
VolatileImage.
- I've removed the copyArea() method from the BufImgSurfaceData, and
modified the original version. The only question I have is: do I
need to
check for "instanceof OffScreenHiDPIImage.SurfaceData" in case when
"transformState == TRANSFORM_TRANSLATESCALE"? If this method is
invoked
with some other SD, and the transform is SCALE, will it do the job
with
the coordinates conversion done?
- I've left the new methods in FramePeer default... May yet we
implement
them in other peers when we really need it?
- CPlatformLWWindow.getGraphicsDevice() checks for an intersection +
scale. This heuristic actually may fail when a Window is moved b/w
three
or four displays so that it intersects them all at some time. JFX
will
set a new scale factor in between and AWT may pick up a wrong
device. I
don't know any simple solution for that. For two monitors this will
work.
Thanks,
Anton.
