We're developing a GIS app aimed at Windows store. Our app follows the Direct3D+XAML pattern and uses the SwapChainBackgroundPanel scheme. We call the ID2D1RenderTarget::DrawText method to display city/town/station names but find the text rendering performance is terrible. For example, when rendering 5000 strings, we see an unacceptable performance drop in our app, sweep zooming, panning no longer follow the input pointer (be it finger or mouse), UI is retarded too. Any ideas to improve the performance? thanks. Here is our code to call the ID2D1RenderTarget::DrawText method:
[source lang="cpp"]void drawString(const wchar_t *str, float x, float y, float dx, float dy, float theta)
{
ComPtr<ID2D1SolidColorBrush> brush;
D2D1_RECT_F rectf;
float r, g, b, a;
int num_char;
float x_offset, y_offset;
r = m_constantBufferData.penColor.r;
g = m_constantBufferData.penColor.g;
b = m_constantBufferData.penColor.b;
a = m_constantBufferData.penColor.a;
DX::ThrowIfFailed(m_d2dContext->CreateSolidColorBrush(D2D1::ColorF(r,g,b,a), &brush));
Avoid creating resource when rendering (avoid CreateSolidColorBrush in a rendering loop).
Also DrawText is absolutely not performance oriented to draw dynamic strings, d2d and dwrite has to calculate the whole vertex buffers of these strings at runtime. If you have 5000 dynamic strings, I would use a SpriteFont bitmap approach instead. If these strings are mostly static, use TextLayout instead whenever it is possible (the vertex layout is precalculated at creation time).
Oh, also, D2D is a batch API, a batch is done between a pair of Begin/End. Calling DrawText between each Begin/End is going to hurt a *lot* the batching (well, you will have no batching). The recommended usage is perform mostly a single Begin/End for the whole rendering of a render pass, and perform all your drawing between this Begin/End pair.
If the strings don't change every frame, pre-create DirectWrite text-layouts for them, and only update those layouts when the strings change (you can of course still change their matrix each frame).
If it's still slow, cache static strings to quads and draw them with D3D.
There are three things that might slow you down here, text-layout, text-rendering, and device state changes.
When a text layout is constructed, a lot of things happen behind the scenes, allocating storage and finding font data to put characters with correct spacing between them and new-lines and what not. This is probably a significant bottle-neck, possibly much worse than the others.
When D2D draws text it renders vector-fonts with antialiasing, which is probably pretty fast, though not as fast as copying pixels.
Finally, each DrawText likely does a load of calls to the underlying D3D-device, like setting up shaders and render-states, possibly constructing lots of geometry and updating GPU-memory buffers on the fly etc. Doing this 5000 times per frame can cause significant performance loss.
I wrote a font-wrapper library using DirectWrite, and one test I wrote shows that it takes about 10 ms to construct 1000 text-layouts of short small strings. Setting up render-states and updating vertex-buffers 1000 times takes maybe 3-4 ms (D2D likely does a lot more than update simple quads, so it might be slower). Rendering the strings takes maybe 0.5 ms (simpler rendering than D2D, using only copy).
Exact numbers ofcourse depend on your hardware, and this is on Windows 7 so perhaps things have improved to Windows 8, but relative costs are likely somewhere in that area.
