Overview
Examples
Screenshots
Comparisons
Applications
Download
Documentation
Tutorials
Bazaar
Status & Roadmap
FAQ
Authors & License
Forums
Funding Ultimate++
Search on this site
Search in forums












SourceForge.net Logo
Home » Developing U++ » U++ Developers corner » Optimizing DrawImage across platforms
Re: MILESTONE: gtk3 replaces gtk2 as default linux backend [message #53652 is a reply to message #53647] Sat, 18 April 2020 19:21 Go to previous messageGo to next message
Tom1
Messages: 1212
Registered: March 2007
Senior Contributor
mirek wrote on Sat, 18 April 2020 17:48
As for windows numbers, looks like Image in Win32 was overengineered, one of optimizations was that the first time it is painted, it paints with SetSurface, only for subsequent paints actually move Image to system handles.

I have now exluded all those things whith #if 0 - looks like it improved at least current RectTracker performance quite nicely...

Mirek


void ImageSysData::Paint(SystemDraw& w, int x, int y, const Rect& src, Color c)


Is this the one (#if 0) you are referring to?

This caused a slowdown of 15-20 % for me. In this case I'm running on Core i7 with integrated Intel HD Graphics 4600 or something...

BR, Tom
Re: MILESTONE: gtk3 replaces gtk2 as default linux backend [message #53659 is a reply to message #53652] Sun, 19 April 2020 14:28 Go to previous messageGo to next message
mirek is currently offline  mirek
Messages: 13975
Registered: November 2005
Ultimate Member
Tom1 wrote on Sat, 18 April 2020 19:21
mirek wrote on Sat, 18 April 2020 17:48
As for windows numbers, looks like Image in Win32 was overengineered, one of optimizations was that the first time it is painted, it paints with SetSurface, only for subsequent paints actually move Image to system handles.

I have now exluded all those things whith #if 0 - looks like it improved at least current RectTracker performance quite nicely...

Mirek


void ImageSysData::Paint(SystemDraw& w, int x, int y, const Rect& src, Color c)


Is this the one (#if 0) you are referring to?

This caused a slowdown of 15-20 % for me. In this case I'm running on Core i7 with integrated Intel HD Graphics 4600 or something...

BR, Tom


Feel free to experiment with it... Smile And report results.

Mirek
Re: MILESTONE: gtk3 replaces gtk2 as default linux backend [message #53663 is a reply to message #53659] Sun, 19 April 2020 22:37 Go to previous messageGo to next message
Tom1
Messages: 1212
Registered: March 2007
Senior Contributor
Hi Mirek,

A 4K maximized window Paint from readily available Image/ImageBuffer Size(3840, 2035) comparison:

- Direct SetSurface: 5800 us
- Original DrawImage (1st>2nd>3rd...etc): 10900 us > 35000 us > 5100 us ...
- #if 0 variant (1st>2nd...etc): 18100 us > 6400 us ...
- Original without SetSurface optimization (1st>2nd...etc): 37000 us > 5000 us ...

This is still on Windows 10, Intel Core i7 with integrated Intel HD 4600.

So, SetSurface is always about 5800 us, which is a good all-around solution. #if 0 variant does not really help here. Original code with or without SetSurface optimization only starts to deliver after quite a few rounds when initialization penalty of 35-37 milliseconds is payed back with small gains like 700-800 us per round compared to direct SetSurface.

Maybe computers with better GPUs deliver better with these optimizations. (I don't know this as I do not have such hardware.) Anyway, with typical business setup this is not the case as high end GPUs only come with gaming rigs.

In the end I will probably stick with SetSurface for predictable performance.

On the Linux/GTK3 front the Ctrl::GetPrimaryScreenArea().GetSize(); (used for cache management) still eats half of the time. I think it should be cached in a per-monitor way.

Best regards,

Tom
Re: MILESTONE: gtk3 replaces gtk2 as default linux backend [message #53666 is a reply to message #53663] Mon, 20 April 2020 10:31 Go to previous messageGo to next message
mirek is currently offline  mirek
Messages: 13975
Registered: November 2005
Ultimate Member
Worth testing:

#include <CtrlLib/CtrlLib.h>

using namespace Upp;

GUI_APP_MAIN
{
	DDUMP(GetDeviceCaps(GetDC(NULL), SHADEBLENDCAPS) & SB_PIXEL_ALPHA);
}


This should detect whether AlphaBlend is HW accelerated on the platform.... It is 0 on my computer with RX580...

It would also be interesting to test how SW emulation fares, simply by changing

	if(0 && fnAlphaBlend() && IsNull(c) && !ImageFallBack &&


In general, I think it might be a good idea to "reactivate" SetSurface, but maybe we can do that with some Image hinting system? I do not really like the idea of GetKind anymore, it is from old days when 1024x768 bitmap was considered huge...

gtk3: Have you tested:

static Size sz = Ctrl::GetPrimaryScreenArea().GetSize();

?

Mirek

[Updated on: Mon, 20 April 2020 11:24]

Report message to a moderator

Re: MILESTONE: gtk3 replaces gtk2 as default linux backend [message #53668 is a reply to message #53666] Mon, 20 April 2020 11:13 Go to previous messageGo to next message
mirek is currently offline  mirek
Messages: 13975
Registered: November 2005
Ultimate Member
BTW, my test in win32 show:

SetSurface about 2x faster than call to AlphaBlend (which does not seem to be HW accelerated based on GetDeviceCaps).

However, AlphaBlend (alone) is 2x faster than SW emulation...

Also time to scan Image for test its kind (GetKind) takes about the same time as AlphaBlend function....

[Updated on: Mon, 20 April 2020 11:23]

Report message to a moderator

Re: MILESTONE: gtk3 replaces gtk2 as default linux backend [message #53671 is a reply to message #53666] Mon, 20 April 2020 12:57 Go to previous messageGo to next message
Tom1
Messages: 1212
Registered: March 2007
Senior Contributor
mirek wrote on Mon, 20 April 2020 11:31
Worth testing:

#include <CtrlLib/CtrlLib.h>

using namespace Upp;

GUI_APP_MAIN
{
	DDUMP(GetDeviceCaps(GetDC(NULL), SHADEBLENDCAPS) & SB_PIXEL_ALPHA);
}


This should detect whether AlphaBlend is HW accelerated on the platform.... It is 0 on my computer with RX580...

It would also be interesting to test how SW emulation fares, simply by changing

	if(0 && fnAlphaBlend() && IsNull(c) && !ImageFallBack &&


In general, I think it might be a good idea to "reactivate" SetSurface, but maybe we can do that with some Image hinting system? I do not really like the idea of GetKind anymore, it is from old days when 1024x768 bitmap was considered huge...

gtk3: Have you tested:

static Size sz = Ctrl::GetPrimaryScreenArea().GetSize();

?

Mirek


Hi

1. DUMP gives me: GetDeviceCaps(GetDC(NULL), SHADEBLENDCAPS) & SB_PIXEL_ALPHA = 0

2. SW emulation (with "if(0 && ..." ) yields 28-31 milliseconds on a maximized 4k window.

3. On the Linux/GTK3 dept. "static Size sz = Ctrl::GetPrimaryScreenArea().GetSize();" does not work because first call yields zero size... The following calls return the monitor size. The initialization of "static Size sz" should be linked with a valid non-zero size returned.

Best regards,

Tom
Re: MILESTONE: gtk3 replaces gtk2 as default linux backend [message #53673 is a reply to message #53671] Mon, 20 April 2020 13:13 Go to previous messageGo to next message
mirek is currently offline  mirek
Messages: 13975
Registered: November 2005
Ultimate Member
Quote:

3. On the Linux/GTK3 dept. "static Size sz = Ctrl::GetPrimaryScreenArea().GetSize();" does not work because first call yields zero size... The following calls return the monitor size. The initialization of "static Size sz" should be linked with a valid non-zero size returned.
Tom


And then it works? Smile

Anyway, I am now trying to make some sense / resolution for all the facts presented.

So it looks like while AlphaBlend is NOT HW accelerated, it is still good to use it anyway, as SW emulation is way slower. Now opaque images so far seem to be best handled with SetSurface. However, testing that Image is opaque takes up several ms. OTOH, using SetSurface has a nice advantage that usually large image does not have to go into the cache...

What is the source of large opaque images in your case? Painter or ImageDraw or something else?

Mirek
Re: MILESTONE: gtk3 replaces gtk2 as default linux backend [message #53674 is a reply to message #53673] Mon, 20 April 2020 14:11 Go to previous messageGo to next message
Tom1
Messages: 1212
Registered: March 2007
Senior Contributor
Hi,

This works:

void SystemDraw::SysDrawImageOp(int x, int y, const Image& img, Color color)
{
...
	static Size sz;
	if(!sz.cx) sz = Ctrl::GetPrimaryScreenArea().GetSize();
	cache.Shrink(4 * sz.cx * sz.cy, 1000); // Cache must be after Paint because of PaintOnly!
}


(Possibly not perfect on multi monitor environments if not started on largest monitor.)

Quote:
What is the source of large opaque images in your case? Painter or ImageDraw or something else?


It's nearly always Painter -- or something similar (e.g. custom rendering) but the result is always in ImageBuffer.

Best regards,

Tom
Re: MILESTONE: gtk3 replaces gtk2 as default linux backend [message #53676 is a reply to message #53674] Mon, 20 April 2020 15:11 Go to previous messageGo to next message
mirek is currently offline  mirek
Messages: 13975
Registered: November 2005
Ultimate Member
Tom1 wrote on Mon, 20 April 2020 14:11

It's nearly always Painter -- or something similar (e.g. custom rendering) but the result is always in ImageBuffer.


I am rather asking because of GetKind. At this point, it appears that it would be nice to have information about whether Image is opaque or not (and use SetSurface or AlphaBlend based on this info). With BufferPainter, I can get this info based on Clear. With generic ImageBuffer, test will be needed, unless client code uses SetKind....

Mirek
Re: MILESTONE: gtk3 replaces gtk2 as default linux backend [message #53677 is a reply to message #53676] Mon, 20 April 2020 15:25 Go to previous messageGo to next message
Tom1
Messages: 1212
Registered: March 2007
Senior Contributor
For custom rendering I can use SetKind(IMAGE_OPAQUE); if that's what it takes to make it fast. Smile

Best regards,

Tom

Re: MILESTONE: gtk3 replaces gtk2 as default linux backend [message #53678 is a reply to message #53677] Mon, 20 April 2020 15:34 Go to previous messageGo to next message
Tom1
Messages: 1212
Registered: March 2007
Senior Contributor
How about putting it readily in constructor:
ImageBuffer(Size sz, int kind=IMAGE_UNKNOWN)

and default to something that always works, while not optimal for all uses.

And also in:
ImageBuffer::Create(Size sz, int kind=IMAGE_UNKNOWN)...

Mostly I know exactly what is going to be found in the buffer anyway...

Best regards,

Tom
Re: MILESTONE: gtk3 replaces gtk2 as default linux backend [message #53683 is a reply to message #53678] Mon, 20 April 2020 17:55 Go to previous messageGo to next message
mirek is currently offline  mirek
Messages: 13975
Registered: November 2005
Ultimate Member
I have comitted intended changes for Win32. Some of Image kinds are deprecated and SetKind(IMAGE_OPAQUE) has synonyme Opaque.

Please check whether this is acceptable...

Mirek
Re: MILESTONE: gtk3 replaces gtk2 as default linux backend [message #53684 is a reply to message #53683] Mon, 20 April 2020 20:37 Go to previous messageGo to next message
Tom1
Messages: 1212
Registered: March 2007
Senior Contributor
Hi Mirek,

Win32: Current solution looks good! Probably as good as it can get with GDI.

Linux/GTK3: While SetSurface scores 12 ms times, DrawImage now yields about 5-6 ms results after initial run at 13-14 ms (due to caching).

(There is one RTIMESTOP("cairo_paint"); left in the code.)

I think this has now been taken pretty much as far as it goes with these backends, unless GTK3 still has some optimization available for opaque images.

Thanks and best regards,

Tom
Re: MILESTONE: gtk3 replaces gtk2 as default linux backend [message #53685 is a reply to message #53684] Mon, 20 April 2020 20:59 Go to previous messageGo to next message
Tom1
Messages: 1212
Registered: March 2007
Senior Contributor
Hi,

One Win32 question: Is it really necessary to cache the image if it is actually going to be drawn as a colored rect or via SetSurface?

Best regards,

Tom

[EDIT] Did some testing and this is faster than using SetSurface after caching:

void SystemDraw::SysDrawImageOp(int x, int y, const Image& img, const Rect& src, Color color)
{
	GuiLock __;
	if(img.GetLength() == 0)
		return;
	LLOG("SysDrawImageOp " << img.GetSerialId() << ' ' << img.GetSize());

// Insert this optimization here:
	int kind = img.GetKindNoScan();
	if(kind == IMAGE_OPAQUE && !IsNull(color)) {
		Size sz=img.GetSize();
		DrawRect(x, y, sz.cx, sz.cy, color);
		return;
	}
	if(kind == IMAGE_OPAQUE && (GetDeviceCaps(GetHandle(), RASTERCAPS) & RC_DIBTODEV)) {
		LTIMING("Image Opaque direct set");
		Size sz=img.GetSize();
		SetSurface(*this, x, y, sz.cx, sz.cy, ~img);
		return;
	}
// End of insertion

	ImageSysDataMaker m;
	...


Or is there some other reason to cache the image before display?

[Updated on: Mon, 20 April 2020 22:49]

Report message to a moderator

Re: MILESTONE: gtk3 replaces gtk2 as default linux backend [message #53687 is a reply to message #53684] Tue, 21 April 2020 16:20 Go to previous messageGo to next message
mirek is currently offline  mirek
Messages: 13975
Registered: November 2005
Ultimate Member
Tom1 wrote on Mon, 20 April 2020 20:37


I think this has now been taken pretty much as far as it goes with these backends, unless GTK3 still has some optimization available for opaque images.



Actually, it has cairo_surface_create_similar_image.

I have just finished experimenting with this feature. It is very confusing: It works really well, bringing DrawImage to microseconds time (probably using XRender path), however it has really nasty feature that it kills the drawing performance after DrawImage.

My guess is that it flips cairo into "GPU" mode and then when you want to draw e.g. rectangles, it starts copying memory between GPU and CPU for each element. Or something like that....

For future reference, before I revert them for now, here are changes I have tried:

struct ImageSysData {
	Image            img;
	cairo_surface_t *surface = NULL;
	
	void Init(const Image& m, cairo_surface_t *other);
	~ImageSysData();
};

cairo_surface_t *CreateCairoSurface(const Image& img, cairo_surface_t *other)
{
	Size isz = img.GetSize();
	cairo_format_t fmt = CAIRO_FORMAT_ARGB32;
	cairo_surface_t *surface = other && 0 ? cairo_surface_create_similar_image(other, fmt, isz.cx, isz.cy)
	                                 : cairo_image_surface_create(fmt, isz.cx, isz.cy);
	cairo_surface_flush(surface);
	byte *a = (byte *)cairo_image_surface_get_data(surface);
	int stride = cairo_format_stride_for_width(fmt, isz.cx);
	for(int yy = 0; yy < isz.cy; yy++) {
		Copy((RGBA *)a, img[yy], isz.cx);
		a += stride;
	}
	cairo_surface_mark_dirty(surface);
	return surface;
}

cairo_surface_t *CreateCairoSurface(const Image& img)
{
	return CreateCairoSurface(img, NULL);
}

void ImageSysData::Init(const Image& m, cairo_surface_t *other)
{
	img = m;
	surface = CreateCairoSurface(m, other);
	SysImageRealized(img);
}

ImageSysData::~ImageSysData()
{
	SysImageReleased(img);
	cairo_surface_destroy(surface);
}

struct ImageSysDataMaker : LRUCache<ImageSysData, int64>::Maker {
	Image img;
	cairo_surface_t *other;

	virtual int64  Key() const                      { return img.GetSerialId(); }
	virtual int    Make(ImageSysData& object) const { object.Init(img, other); return img.GetLength(); }
};

void SystemDraw::SysDrawImageOp(int x, int y, const Image& img, Color color)
{
	GuiLock __;
	FlushText();
	if(img.GetLength() == 0)
		return;
	LLOG("SysDrawImageOp " << img.GetSerialId() << ' ' << x << ", " << y << ", "<< img.GetSize());
	ImageSysDataMaker m;
	static LRUCache<ImageSysData, int64> cache;
	static int Rsz;
	Rsz += img.GetLength();
	if(Rsz > 200 * 200) { // we do not want to do this for each small image painted...
		Rsz = 0;
		cache.Remove([](const ImageSysData& object) {
			return object.img.GetRefCount() == 1;
		});
	}
	LLOG("SysImage cache pixels " << cache.GetSize() << ", count " << cache.GetCount());
	m.img = img;
	m.other = cairo_get_target(cr);
	ImageSysData& sd = cache.Get(m);
	if(!IsNull(color)) {
		SetColor(color);
		cairo_mask_surface(cr, sd.surface, x, y);
	}
	else {
		RTIMESTOP("cairo_paint");
		if(img.GetKindNoScan() == IMAGE_OPAQUE)
			cairo_set_operator(cr, CAIRO_OPERATOR_SOURCE);
		cairo_set_source_surface(cr, sd.surface, x, y);
		cairo_paint(cr);
		cairo_set_operator(cr, CAIRO_OPERATOR_OVER);
	}
	static Size ssz;
	if(ssz.cx == 0)
		ssz = Ctrl::GetVirtualScreenArea().GetSize();
	cache.Shrink(4 * ssz.cx * ssz.cy, 1000); // Cache must be after Paint because of PaintOnly!
}


Re: MILESTONE: gtk3 replaces gtk2 as default linux backend [message #53688 is a reply to message #53687] Tue, 21 April 2020 18:03 Go to previous messageGo to next message
Tom1
Messages: 1212
Registered: March 2007
Senior Contributor
Hi,

Nice! I mean the latest GTK3 DrawImage optimization... Tried it and it feels very appealing and would perform great with these Painter generated scenes. I'm just wondering how well it would work with a typical GUI application workload, if the GUI was first rendered with Painter and then painted to window with this new efficient DrawImage solution. I also wonder if rectangles could possibly be drawn more efficiently using DrawImage with opaque color, since this uses an image mask and should therefore not need to copy it back to CPU accessible memory. I could not benchmark this properly here, but it might be something to look at.

Best regards,

Tom
Re: MILESTONE: gtk3 replaces gtk2 as default linux backend [message #53689 is a reply to message #53687] Tue, 21 April 2020 18:28 Go to previous messageGo to next message
mirek is currently offline  mirek
Messages: 13975
Registered: November 2005
Ultimate Member
Ha, I kept digging and in the end found that them problem is caused by InvertColor in DrawDragFrame (probably it is not possible to implement difference operator with Xrandr, so cairo needs to copy memory chunks from/to GPU).

As using that is not mandatory, e.g. alternating white and black dots work as well, I have removed InvertColor and the whole thing started to work as expected. Gtk3 for repeated image draw are now in microseconds range.

Mirek
Re: MILESTONE: gtk3 replaces gtk2 as default linux backend [message #53691 is a reply to message #53689] Tue, 21 April 2020 19:25 Go to previous messageGo to next message
Tom1
Messages: 1212
Registered: March 2007
Senior Contributor
Mirek,

Absolutely brilliant! It works just great! Smile

I'm working on a 'SpriteCtrl' and after this improvement it is actually faster with full Paint than with a partial ViewDraw update!!!

I only wish Windows had something similar. As GDI in Windows 10 probably only lives virtually in a layer of software feeding the GPU, I'm getting a strong feeling that we should gradually start looking at Direct2D as it is the MS recommended 2D graphics solution for new applications.

Thanks and best regards,

Tom

Re: MILESTONE: gtk3 replaces gtk2 as default linux backend [message #53693 is a reply to message #53691] Tue, 21 April 2020 22:19 Go to previous messageGo to next message
Didier is currently offline  Didier
Messages: 680
Registered: November 2008
Location: France
Contributor
Hi,

All this is very interresting.

I just compiled RectTracker projet with svn.14195 (3 weeks old) : I wanted to see the performance difference with new solution
... and it didn't work at all !

After updating to svn.14346, it works fine again Smile

I also compiled my current project in which I use my GraphCtrl widget ... and the drag/scroll is completely broken : While draging, I get a completely white widget.
When scrolling is finished, image commes back again.

My usage of Localloop seems to be broken Confused

Goiing back to svn.14195 for now

I am under linu, latest Fedora, with an old GPU (Nvidia GT9600)

[Updated on: Tue, 21 April 2020 22:28]

Report message to a moderator

Re: MILESTONE: gtk3 replaces gtk2 as default linux backend [message #53699 is a reply to message #53689] Wed, 22 April 2020 14:08 Go to previous messageGo to previous message
Tom1
Messages: 1212
Registered: March 2007
Senior Contributor
Hi,

Here's a completing feature for RectTracker for tracking line drawing (TrackLine) and free-hand polyline (TrackRoute):

class LineTracker: public RectTracker{
public:
	int mode;
	Vector<Point> route;

	Vector<Point> TrackLine(const Point& p){
		mode=1;
		route.Clear();
		route.Add(p);
		Track(Rect(p,p),0,0);
		return clone(route);
	}
	
	Vector<Point> TrackRoute(const Point& p){
		mode=2;
		route.Clear();
		route.Add(p);
		Track(Rect(p,p),0,0);
		return clone(route);
	}

	void MouseMove(Point p, dword){
		if(mode==2 || route.GetCount()<2){
			route.Add(p);
		}
		else route[1]=p;
		Refresh();
		sync(Rect(route[0],route[route.GetCount()-1]));
	}

	void Paint(Draw& w){
		w.DrawImage(0, 0, master_image);
		w.Clip(clip & GetMaster().GetSize());
		int style=width;
		if(pattern!=DRAWDRAGRECT_SOLID){
			Color color2 = IsDark(color) ? White() : Black();
			w.DrawPolyline(route,width,color2);
			switch(pattern){
				case DRAWDRAGRECT_NORMAL:
					style=PEN_DOT;
					break;
				case DRAWDRAGRECT_DASHED:
					style=PEN_DASHDOT;
					break;
			}
		}
		w.DrawPolyline(route,style,color);
		w.End();
	}
	
	LineTracker(Ctrl &master): RectTracker(master){
		MinSize(Size(-100000,-100000));
		mode=1;
	}
};


Feel free to include it in U++.

To test the functionality, you can try:
	void LeftDown(Point p, dword flags){
		LineTracker tracker(*this);
		tracker.Solid();
		tracker.Width(3);
		tracker.SetColor(Blue());

		Vector<Point> res;

		if(flags&K_CTRL) res = tracker.TrackRoute(p);
		else res = tracker.TrackLine(p);
		// ...
	}

It proved easy enough to inherit from RectTracker, so I made it a separate class after all.

Best regards,

Tom

[Updated on: Wed, 22 April 2020 21:11]

Report message to a moderator

Previous Topic: Rapsberry PI - cpu dilemma
Next Topic: U++ on Telegram
Goto Forum:
  


Current Time: Fri Mar 29 14:05:11 CET 2024

Total time taken to generate the page: 0.01832 seconds