Swapchain->Present locks up

SwapChain::Present exhibits asynchronous behaviour and queues blit/flip operations rather than performing them immediately. Depending on your swap mode (you seem to be using blit rather than flip) the rate at which the queue gets processed with vsync disabled will vary, as will the maximum length of the queue.

You're submitting frames at roughly 200/s and they're being consumed by the device at about 70/s; so it's most likely that you're saturating the presentation queue and the Present function is then blocking until there's some free space. It looks like you might be able to use DXGI_PRESENT_DO_NOT_WAIT though that may be specific to DX11 threaded model, rather than behaving like DX9's do not wait flag.

Jans.

Thanks for the insight Jansic! What you're saying is making sense. I looked at the flag but it has a small note:

"Direct3D 11: [color=#2A2A2A]This enumeration value is supported starting with Windows Developer Preview." In other words not possible on windows 7

According to what you are saying the problem should be gone if i either render more sprites (lets say 12000 sprites) or enable vsync (not correctly working currently, need to look into that) or put a ~10ms Sleep in my code. I will test this a bit later tonight.

Also i remember trying "DXGI_PRESENT_DO_NOT_SEQUENCE" But this caused no backbuffer flipping -> I only see the very first frame i draw.

I am still working on this problem. And i am kinda getting near the solution. I actually had GPU-Z running on my 2nd monitor and i detected that my GPU (HD5770) didnt switch to 3d clocksettings. This is ofcourse pretty strange. I tried messing with some value's in my code on the directx part, but nothing changed. (not that i could change much there anyway).

So i went to my window handler and copied settings over from the hieroglyph 3.0 engine.

New settings:


WNDCLASSEX wc;

// Get the instance of this application.
hinstance = GetModuleHandle(NULL);

// Give the application a name.
applicationName = L"Default Name";

// Setup the windows class with default settings.
wc.style = CS_HREDRAW | CS_VREDRAW;
wc.lpfnWndProc = WndProc;
wc.cbClsExtra = 0;
wc.cbWndExtra = 0;
wc.hInstance = hinstance;
wc.hIcon = LoadIcon(NULL, IDI_WINLOGO);
wc.hIconSm = wc.hIcon;
wc.hCursor = LoadCursor(NULL, IDC_ARROW);
wc.hbrBackground = (HBRUSH)GetStockObject(BLACK_BRUSH);
wc.lpszMenuName = NULL;
wc.lpszClassName = applicationName;
wc.cbSize = sizeof(WNDCLASSEX);

// Register the window class.
RegisterClassEx(&wc);

DWORD windowStyle;

// Setup the screen settings depending on whether it is running in full screen or in windowed mode.
if(isFullScreen)
{
windowStyle = WS_OVERLAPPEDWINDOW | WS_VISIBLE;
}
else
{
windowStyle = WS_POPUP | WS_VISIBLE;
}

// Place the window in the middle of the screen.
int posX = (GetSystemMetrics(SM_CXSCREEN) - screenWidth) / 2;
int posY = (GetSystemMetrics(SM_CYSCREEN) - screenHeight) / 2;

// Create the window with the screen settings and get the handle to it.
hwnd = CreateWindowEx(WS_EX_APPWINDOW, applicationName, applicationName,
windowStyle,
posX, posY, screenWidth, screenHeight, NULL, NULL, hinstance, NULL);

// Bring the window up on the screen and set it as main focus.
ShowWindow(hwnd, SW_SHOWNORMAL);
UpdateWindow(hwnd);


This fixed the GPU not going to 3D mode and it actually makes the the amount of time the presentbuffer takes to run to 65 ms. But it's still the same as in:

0.0003 seconds
0.0004 s
0.0650 s
0.0640 s
0.0003 s
etc etc etc

So sometimes its super fast, or 'as expected' and sometimes it locks up for ~4 vsyncs?


I just remembed i also added this piece of code:

IDXGIDevice1* idxgiDevice = NULL;
device->QueryInterface(__uuidof(IDXGIDevice1), (void**)&idxgiDevice);
idxgiDevice->SetMaximumFrameLatency(1);



Here is a graph without SetMaximumFrameLatency(1): http://puu.sh/da2r
As you can see, at the start its not as much of a problem but later one it locks up more

Here is a graph WITH SetMaximumFrameLatency(1): http://puu.sh/da38
It seems to me that it locks up for ~0.064 seconds for a duration of 15 frames.

Note: the "fps" label is incorrect and its the amount of seconds it takes for PresentBuffer to complete.
Note 2: All these tests are without vsync.

Question is, is there anything i need to change to creating the window? Or maybe a different setting. I can't run the hieroglyph engine myself as i have to install VS2010 C++ express, which i cant currently do.

Edit:
Interesting, i did a deviceContext->Flush() call after calling deviceContext->DrawIndexed(...) and it actually depends on how much i draw how long it takes. So why i am getting this bad performance, rendering only 1000 vertices with 1500 indices..

Fixed. It was actually just the amount of pixels i had to draw was to much. I was wanting to draw 4000 x 440000 pixels. Which is to much for the GPU to handle. When moving back to much smaller (and normal sized) sprites the performance was as expected.