Flicker when moving/resizing window

Question

I have developed an application to display jpeg images. It can display 4 images, one in each quadrant of the screen. It uses 4 windows for that. The windows have no border (frame) nor titlebar. When a new image is loaded, the window size is adjusted for the new image and then the image is displayed.

Especially when the window is made larger, there is often a flicker. With my eyes to slits, it seems that the old contents is moved when resizing before the new contents is displayed.

I consulted many resources and used all tricks:

• the window has only style CS_DBLCLKS (no CS_HREDRAW or CS_VREDRAW);

• the background brush is NULL;

• WM_ERASEBKGND returns 1;

• WM_NCPAINT returns 0;

• WM_NCCALCSIZE tells to align to the side not moved (can you tell it to discard the client area?);

• WM_WINDOWPOSCHANGING returns 0;

• SetWindowPos has flags SWP_NOCOPYBITS | SWP_DEFERERASE | SWP_NOREDRAW | SWP_NOSENDCHANGING.

Still, the flicker (or the contents move) occurrs when resizing the window. What I want is to:

• SetWindowPos to new size and position;

• InvalidateRect (hWnd, NULL, FALSE);

• UpdateWindow(hWnd);

without any painting, background erasing or content moving until WM_PAINT.

WM_PAINT does:

hDC= BeginPaint (hWnd, &ps);
hMemDC= CreateCompatibleDC (hDC);
hOldBitmap = SelectObject (hMemDC, hNewBitmap);
BitBlt (hDC,...,hMemDC,0,0,SRCCOPY);
SelectObject (hMemDC, hOldBitmap);
DeleteDC (hMemDC);
EndPaint (hWnd, &ps);

Can anyone tell me if/where I make a mistake that causes the old content of the window to be moved?

Hardware etc: Windows 7 on HP Elitebook Core7 64 bits with NVIDIA Quadro K1000m driver 9.18.13.3265 (updated to 341.44).

---

UPDATE (Jul '17)

I have seen the behavior of the pogram also on another Windows computer (Windows 8/10). It does not seem to be the NVIDIA display driver.

The behavior is the most visible when resizing a window tiled to the centre of the screen (right bottom = w/2, h/2) to the left or left upper corner (0, 0).

I may have problems with the calculations for the WM_NCCALCSIZE message to tell Windows not to do anything. Could anyone give an example calculation for my purpose? See also How do I force windows NOT to redraw anything in my dialog when the user is resizing my dialog?

Answer 1

You have an impressive list of anti-flickering tricks :-)

I don't know if this is of importance (since it depends on how your tool windows are created and espacially if they are child windows to a common parent): Try setting window style WS_CLIPCHILDREN in the parent window of the tool windows (if there is one).

If not set the parent window will erase it's (entire) background and then forward the paint messages to the child windows which will cause flickering. If WS_CLIPCHILDREN is set the parent window does nothing to the client area occupied by child windows. As a result the area of child windows isn't drawn twice and there is no chance for flickering.

Answer 2

This is a theory more than an answer:

By default, in modern Windows, your window is just a texture on the video card, and the desktop window manager is mapping that to a rectangle on the screen. You've seem to have done everything necessary to make sure that texture gets updated in one fell swoop.

But when you resize the window, perhaps the desktop compositor immediately updates its geometry, causing the (still unchanged) texture to be appear in the new position on the screen. It's only later, when you do the paint, that the texture is updated.

You can test this theory by temporarily turning off desktop compositing. On Windows 7, you navigate to System Properties, choose the Advanced tab, under Performance choose Settings..., on the Visual Effects tab, deselect the "Enable desktop composition" setting. Then try to reproduce the problem. If it goes away, then that supports (but doesn't absolutely prove) my theory.

(Remember to re-enable compositing, since that's how most of your users will be running most of the time.)

If the theory is true, it seems the goal is to get to the paint as soon possible after the window resize. If the window of time is small enough, then both could happen within a monitor refresh cycle and there would be no flicker.

Ironically, your efforts to eliminate flicker may be working against you here, since you've intentionally suppressed the invalidation and redraw that would normally result from SetWindowPos until you do it manually at a later step.

A debugging tip: Try introducing delays (e.g., Sleep(1000);) at key points in the process so you can see whether the resize and redraw are actually rendering on screen as two distinct steps.

Answer 3

You have indeed got a nice set of tricks.

First I can suggest a few variants on the existing tricks that might help especially on XP/Vista/7, and second I want to mention where the persistent flicker you see on Win8/10 is likely coming from and some tricks to reduce that.

First, despite advice to the contrary in other OP posts, you may find that if you set the CS_HREDRAW|CS_VREDRAW window styles that you avoided before, it actually eliminates the BitBlt that is done inside the internal SetWindowPos Windows does during window resizing (but---and this is the confusing part---on Windows 8/10 you will still see flicker from another source...more on that below).

If you don't want to include CS_HREDRAW|CS_VREDRAW, you can also intercept WM_WINDOWPOSCHANGING (first passing it onto DefWindowProc) and set WINDOWPOS.flags |= SWP_NOCOPYBITS, which disables the BitBlt inside the internal call to SetWindowPos() that Windows makes during window resizing.

Or, you could add to your WM_NCCALCSIZE trick by having it return a set of values that will tell Windows to just BitBlt 1 pixel on top of itself so that even if the BitBlt does happen, it doesn't do anything visibly:

case WM_NCCALCSIZE:
{
    RECT ocr; // old client rect
    if (wParam)
    {
        NCCALCSIZE_PARAMS *np = (NCCALCSIZE_PARAMS *)lParam;
        // np->rgrc[0] is new window rect
        // np->rgrc[1] is old window rect
        // np->rgrc[2] is old client rect
        ocr = np->rgrc[2];
    }
    else
    {
        RECT *r = (RECT *)lParam;
        // *r is window rect
    }

    // first give Windoze a crack at it
    lRet = DefWindowProc(hWnd, uMsg, wParam, lParam);

    if (wParam)
    {
        NCCALCSIZE_PARAMS *np = (NCCALCSIZE_PARAMS *)lParam;

        // np->rgrc[0] is new client rect computed
        // np->rgrc[1] is going to be dst blit rect
        // np->rgrc[2] is going to be src blit rect
        // 
        ncr = np->rgrc[0];
        RECT &dst = np->rgrc[1];
        RECT &src = np->rgrc[2];

        // FYI DefWindowProc gives us new client rect that
        // - in y
        //   - shares bottom edge if user dragging top border
        //   - shares top edge if user dragging bottom border
        // - in x
        //   - shares left edge if user dragging right border
        //   - shares right edge if user dragging left border
        //
        src = ocr;
        dst = ncr; 

        // - so src and dst may have different size
        // - ms dox are totally unclear about what this means
        // https://learn.microsoft.com/en-us/windows/desktop/
        //         winmsg/wm-nccalcsize
        // https://learn.microsoft.com/en-us/windows/desktop/
        //         api/winuser/ns-winuser-tagnccalcsize_params
        // - they just say "src is clipped by dst"
        // - they don't say how src and dst align for blt
        // - resolve ambiguity

        // essentially disable BitBlt by making src/dst same
        // make it 1 px to avoid waste in case Windoze still does it

        dst.right  = dst.left + 1;
        dst.bottom = dst.top  + 1;

        src.right  = dst.left + 1;
        src.bottom = dst.top  + 1;

        lRet = WVR_VALIDRECTS;
    }
    else // wParam == 0: Windoze wants us to map a single rect w->c
    {
        RECT *r = (RECT *)lParam;
        // *r is client rect
    }

    return lRet;

So that's all very nice, but why does Windows 8/10 look so bad?

Apps under Windows 8/10 Aero don't draw directly to the screen, but rather draw to offscreen buffers that are then composited by the evil DWM.exe window manager. DWM actually adds another layer of BitBlt-type behavior on top of the existing legacy XP/Vista/7 BitBlt behavior.

And the DWM blit behavior is even more crazy because they don't just copy the client area, but they actually replicate pixels at the edges of your old client area to make the new one. Unfortunately, making DWM not do its blit is much harder than just passing some extra flags.

I don't have a 100% solution, but I hope the above info will help, and please see this Q&A for a timing trick you can use to reduce the chances of the DWM-layer blit happening to your app:

How to smooth ugly jitter/flicker/jumping when resizing windows, especially dragging left/top border (Win 7-10; bg, bitblt and DWM)?

Answer 4

Just in addition to your list of tricks. I have the same problem with flicker on my Dell XPS notebook while resizing window using left/top edge. I've tried all the tricks you mentioned. As far as I understand window border is drawn in GPU and the window content is prepared in GDI subsystem and transferred to video memory for window composition (DWM introduced in Windows 8.1). I tried to remove GDI rendering completely (setting WS_EX_NOREDIRECTIONBITMAP style) which makes window without any content, and then create content surface directly using DXGI subsystem (using CreateSwapChainForComposition, there are few examples on how to do this). But the problem remains. There is still a lag between rendering a window frame and resizing/displaying content surface.

However it may solve your problem, as you don't have the border and you will not notice this lag. But you will have full control over window repaint and it will be made on GPU side.