2D, Acceleration, And Windows: Aren't All Graphics Cards Equal?

The Limits Of 2D: One Space With Many Windows

Rendering and 2D Acceleration in Windows

If you look aWidth and height are all you needWidth and height are all you needt 2D rendering inside any display window, only two dimensions apply: X and Y, or width and height. What’s missing? Any kind of depth information.

In Windows, 2D graphics get rendered thanks to the GDI (Graphics Device Interface). This interface supports all of the higher-level programming languages, and includes all of the important graphics functions necessary to render 2D graphical objects. Later enhancements, such as GDI+ and Direct2D, don’t play much of a role here because GDI is (and remains) the most important tool for 2D graphical output in applications. The critical drawing functions for pixels, lines, curves, polygons, rectangles, ellipses, and so forth were initially all calculated on the CPU. Thanks to focused specialization in graphics cards, later generations of hardware delivered faster 2D calculations and rendering. This early form of 2D acceleration remains important even today, but no longer takes two-dimensional acceleration as its primary focus. To make the most of graphics performance, we need a third coordinate.

Simple Rendering of Multiple Windows without Hardware Acceleration

The original 2D rendering method that lurks behind overlapping display windows is both simple and straightforward. One needs to know two things: first, the area on screen inside any window that has changed (and thus, must be redrawn), and second, the order in which windows or objects overlay one another (whether it’s visible in whole or in part, or covered by some other window). This type of information offers two-and-a-half dimensions, or layered graphics, where a third coordinate takes a 0 (hidden) or 1 (visible) value as a kind of "helper dimension." That’s why you’ll hear Windows experts talk about 2.5D graphics.

The Z order indicates how windows are layered or organizedThe Z order indicates how windows are layered or organized

After order or visibility issues are decided, visible window contents may be drawn using purely two-dimensional graphics functions. Nevertheless, it’s not only necessary to render display window contents completely, but it’s also necessary to manage numerous types of information and window content displays. What happens, for example, when windows get moved? Whenever another window contains a region that is partially or completely revealed as a consequence, the system graphics routine WM_PAINT must be called with precise information to indicate which rectangular region must be redrawn ("dirty rectangle"). Optimized implementations will reconstruct or redraw this area. Sadly, many implementations instead opt to redraw the whole window, despite ready access to more precise instructions, whether or not it needs to be completely or only partially reproduced. This can have a profound effect on graphics performance. Another disadvantage is the well-known and despised blurring or repeat effect that often occurs when windows get dragged quickly across the screen on a system that lacks hardware acceleration.

Let’s sum up what we’ve reviewed so far. There are individual windows on display, whose two-dimensional contents must be drawn on-screen so they may be viewed. These windows can be moved arbitrarily, so that they can overlap and be partially or completely obscured by other windows. The visible contents from all of these windows need to managed, and to be drawn on-screen with an absolute minimum delay. We also know that the CPU itself, even for very fast processors, can be overwhelmed in performing such complex tasks. What else is there to do, except to offload this onto the graphics card? What this entails, and how it sounds simpler in concept than in actual practice, is what we tackle in the next section.

This thread is closed for comments
156 comments
    Your comment
  • fatkid35
    next, toms finds bigfoot eating a chupicabre! WTF? nice catch guys.
  • pcxt21
    Very nice work! Until that little update I was ready to put my old Matrox Millennium (1996/1997 I think) 2d accelerator back into my gaming rig...
  • For those who still remember Matrox...shouldn't that be included in the test as well? Ancient history shows that it was the best card to be paired with VooDoo when first released...:P
  • one-shot
    LOL. The clip of Steve Ballmer looks like an ad from the late Billy Mays. To think he's the CEO of Microsoft after watching that clip makes me laugh. I'm going to watch it again.
  • chookman
    I actually went to put a Professional Matrox card (G550 PCI-e) in my Windows 7 machinde for a few more displays... alas it seems most of the range doesnt have Windows 7 drivers yet :(
  • killerclick
    Hey, I have GF7050 on my motherboard but I'm not using it!
  • belardo
    The sad thing is, as shown was that Windows 1.0~3.0 were never an operating system, ever. 1.0~2.x we're never really used by anyone, yet Amigas and Macs had full GUI functionality since 85/84... Windows3 was at least usable, but still a single-tasking POS, and yet MS-DOS was able to dominate the computer industry... not because of talent, but because of luck, good moves by Gates & Ballmer and the stupidity of IBM. It kind of worked out anyway thou.

    Think about it, it wasn't until the tail-end of 1995, with the release of Windows95, did MS produce an actual GUI OS for the general public, 10 whole years after Apple's Mac and the Amiga by chicken lips. Atari only copied much of the hardware abilities of Amiga with a variant of a MS-DOS clone.

    Only business bought PC-Clones in the 80s~90s, paying around $1500~4000 per desktop. Since IBM didn't make the OS and the hardware was generic, it was super easy for clones to exists. Home users bought Amigas and Macs and laughed as DOS users... but in the end, they lost to the clones.

    At one time in 1990, the Computer Shopper Mag was about 500 pages (phone book) with at least 500 Clone companies advertising.

    Windows is still the Copy-Cat OS. Its still an ugly beast under the hood. At least Windows7 runs good and looks nice.
  • Raid3r
    Indeed, way to take one for the team..I am one of those 2d workers and greatly appreciate the foot that was used to affirm the position of 2d recognition on these "new" cards. I can't say it enough.
  • micky_lund
    woot for toms...
    catch some more massive companies out, and make them fix up their drivers
  • belardo
    I just finished reading this entire article. A good one too guys.

    While I had my rant about old MS days, I did work in the PC field starting with 3.1. We'd benchmark various video cards with programs that would test lines, boxes, etc. This WAS important for some games like DOOM and Quake which were not "true" 3D cards we have today.

    For every new type of PC build or a clients computer, I would save these generic overall system benchmarks. And I *have* notice different 2D performance abilities of computers and the various cards. All super fast compared to the 90s, but I've seen a GF card perform worse than an older ATI, but also an older ATI work a bit better than a newer one. These are all WinXP and don't have anything to do with the problems of Win7 and/or DX11... Part2 isn't out yet... but I would be curious to see if the problems in Win7 happens in Vista with DX11 installed hmmm.

    ATI has been putting all their work into 3D gaming performance, its good to see that they have put a team to fix their 2D issues. There is a good chance that this is a DX11 issue they were not aware of. Since Nvidia doesn't have any DX11 parts (but a DX10.5) - this "bug" doesn't yet show... but I'm speculating of course. :)

    With the eye-candy of Win7, the 2D performance *IS* important for those doing work, watching videos, etc. No good excuse to miss this, and hopefully ATI will have it resolved in 1-2 months.

    PS: Your memory usage chart of vista vs Win7 shows exactly WHY Vist . That is why Windows7 runs pretty good on a 1GB computer, but Vista still needs at least 3GB for a bottom-end PC. My notebook has Win7rc with 1GB, runs fine.

    Hey, will your results sometime include intel graphics? After this bug-issue is resolves, include overall-2D scores with your graphics Charts. :)
  • notty22
    This is no news to the enthusiasts, ATI drivers. One more reason to go green, Nvidia, :)
  • BartG
    So the Vista bashers now need to rethink why they dont like Vista, its not the OS (even though its not perfect), its the lack of 2D performance from your cards!
  • sohei
    good article
  • supasso
    The new ATI cards throttling down pretty aggressively in 2D AKA "idle" mode to keep the power consumption low. Perhaps that's what happening here.
  • sohei
    notty22 you exaggerate...if ATi has a problem with drivers in 2d, what about nvidia with new drivers ? no over-clocking...you know that?

    pls read more than talk
  • eddieroolz
    Reading it at 2:20am, it was hard to grasp every part of the article but it did arouse my mind a bit :P But good job to you guys!

    Now back to programming in C xD
  • mitch074
    @supasso: even throttled down, current cards far exceed the capabilities of older 2D cards; the problem is that, if an operation is badly optimized, it can slow down even the most powerful system to a crawl, by requiring thousands of CPU cycles and as many sleep() cycles, to perform something that, on said older hardware, would take a few cycles at best.

    One good area to look at here, is the Linux world. Why is it relevant?
    - dixit an AMD engineer, 90% of the Linux driver code is common with Windows'. That includes 2D, video and OpenGL.
    - there are two (well three, but this is not really relevant here) Ati card drivers for R600/700 cards: one (proprietary, very close to Windows) from AMD, one (FOSS, based on AMD documents and reverse engineering) from the Xorg development team.

    Benchmarking those in 2D (say, with X's own benches and GNOME's) shows that the
    FOSS driver is around 4 times (and in some cases, 100x) faster than AMD's driver on purely 2D operations - showing that, indeed, AMD's driver needs a kick in the arse to push windows around.

    Please note that Xorg does also provide purely 2D (EXA+Render) and 2.5D (EXA+Composite) modes. enabling Composite on AMD's driver causes it to slow down across the range, while it actually improves things on the Free driver when used with a compositing ('like Aero', although it actually came out before) window manager - and a GPU that is at least level with a Radeon 9500.

    Please note that the very same thing happens with Nvidia hardware, eventhough Nvidia seems to have done a slightly better job at supporting 2.5D (the difference between Nouveau and the Nvidia blob is less apparent).
  • dimitrik
    Excellent article and about time 2D got some attention. The amount of emphasis on 3D graphics in the majority of technology media is insane given that gamers are hardly a majority among mainstream users. Clearly this extreme emphasis on gaming has finally led to the "core" features being neglected by the vendors!
  • haplo602
    Same but worse situation is with ATI linux drivers. Their 2d part is horribly bad even compared to the OSS X11 drivers. I just hope that by fixing the 2d in Windows, they will get to fix 2d in Linux as well.
  • Wouldn't this benchmark be better suited using Direct2D instead of the long-since-deprecated GDI? You should get performance improvements for all of the graphics cards that way. I did some checking and it seems that at least since Vista, the main GUI/window manager has been using DirectX, not GDI. Someone brought up the issue of vector drawing programs like Inkscape (GTK), but I don't know what those use.

    Here is a 2008 article espousing Direct2D over GDI/GDI+:
    http://www.maximumpc.com/article/news/gdigdi_move_over_microsoft_introduces_direct2d

    @Belardo

    Both Windows 3.x and Mac OS 9 & earlier were cooperative multi-tasking OSes, and thus suffered the same fatal flaw. Windows 9x offered preemptive multi-tasking and memory protection (and management), whereas Mac OS 9 & earlier all had such poor protection that any program could easily read/write any other programs memory space intentionally and accidentally. Prior to the MultiFinder's creation in 1988 (OS 5), there was no multitasking at all, and the Desk Accessories had to be written using drivers to emulate it. That's the reason Mac OS X was created; all the previous versions were such utterly irredeemable, insecure, unreliable pieces of shit, they had to start almost from scratch.

    I don't know what an "actual GUI OS" is in your mind, but it doesn't seem like Mac OS classic qualified by your own standards. OSes can run within OSes, and I don't just mean emulating hardware running a "full OS," but rather the definition of an OS doesn't require that it be capable of running completely on its own. This why some research and student OSes are made to run on top of pre-existing ones.

    The way I see it, Windows 3.x was like a powerful DOS extender (DOS4GW is an example of a dos extender that many games used albeit 32-bit) that added a graphics subsystem, cooperative multi-tasking, and ran in 16-bit protected mode (DOS ran in real mode).
  • This is a rather interesting article. Thanks for the good read. What is strange though is the fact that a lot of companies, including Microsoft is looking at Direct2D to speed up browsers and various apps. I find it strange that MS have not spoken more harshly to the Graphics people of ATI and nVidia.

    One take 2D for granted, yet it is clear that no one at ATI or nVidia does...yet we spend 99.5% of our time in there...

    It also points out some of the underlying problems of MS designs...they have so much raw horsepower to yield, that it is no longer important to focus on optimisation. If it where, this would have been picked up earlier and flagged during the beta stages.
  • notty22
    Straatkat I find it strange that MS have not spoken more harshly to the Graphics people of ATI and nVidia.

    Because its only ATI that has the problem. Even Nvidia's IGP performs satisfactorily. Lack of testing, rushing out a product before its right.
    This is poor. Fermi may be late, but it will be right.
  • Onus
    No HD5xxx for me until this is fixed.
  • scrumworks
    Update (1/26/2010): Nice to know AMD has taken a task to fix those pending issues even those problems might just look bad in benchmarks and not in real use.