Windows resizing artifacts regression

[knopp]

flutter/engine#24428, has caused regression in certain situations.

Bit of background first. As far as I can tell, the resizing with direct composition enabled was originally broken, because every second call to CreateSurface failed. I believe this happens because when recreating the surface, context is still in use, which means a IDCompositionTarget instance is still alive. I haven't tried this with UWP, but I was able to get direct composition surface (EGL_DIRECT_COMPOSITION_ANGLE) recreated as expected by calling ClearContext() right before DestroySurface() in AngleSurfaceManager::ResizeSurface.

The PR fixed it by only calling CreateSurface once and then resizing existing surface. This is done though calling eglPostSubBufferNV, which as a side effects resizes underlying surface. The issue here is that eglPostSubBufferNV also causes Swapchain->Present to be called two additional times during every resize step, with wrong surface size. This is not obvious for windows that have border. DWM simply seems to eat these two extra Present calls.

It does however cause problem for windows that had border removed through WM_NCCALCSIZE . It seems that whatever magic DWM does to synchronize swapchain swap with window resize, it gives up once WM_NCCALCSIZE is used. In this case it manifests in those two extra Present calls properly executed (including waiting for vsync), which results in content wobbling around during resizing.

The ideal solution here would most likely be to introduce a way in angle to force surface resize, without having additional swaps during resize. The resizing was changed lately to ensure that we only do one swap per resize step (with correct size).

It would be also good to know if maybe we're missing some ClearContext() calls, or perhaps it is by design and the ClearContext() call should be placed in ResizeSurface().

If proper solution (surface resize in angle) is a no go, than perhaps, at least temporarily, we could revert this for non UWP target.

cc @clarkezone

[clarkezone]

I will take a look.

[knopp]

I'm attaching videos of best effort frameless resizing before and after flutter/engine#24428:

before-pr.mov

after-pr.mov

[knopp]

And stacktraces responsible for two extra Present calls that cause these glitches

rx::SwapChain11::present(rx::DisplayD3D * displayD3D, int x, int y, int width, int height) Line 894 
rx::SwapChain11::swapRect(rx::DisplayD3D * displayD3D, int x, int y, int width, int height) Line 780
rx::SurfaceD3D::swapRect(rx::DisplayD3D * displayD3D, int x, int y, int width, int height) Line 302
rx::SurfaceD3D::postSubBuffer(const gl::Context * context, int x, int y, int width, int height) Line 373 
egl::Surface::postSubBuffer(const gl::Context * context, int x, int y, int width, int height) Line 337 
egl::PostSubBufferNV(egl::Thread * thread, egl::Display * display, egl::Surface * eglSurface, int x, int y, int width, int height) 
EGL_PostSubBufferNV(void * dpy, void * surface, int x, int y, int width, int height) Line 1099 
flutter::AngleSurfaceManager::ResizeSurface(std::variant<HWND__ *> * render_target, int width, int height) Line 268 

rx::SwapChain11::present(rx::DisplayD3D * displayD3D, int x, int y, int width, int height) Line 894 
rx::SwapChain11::swapRect(rx::DisplayD3D * displayD3D, int x, int y, int width, int height) Line 780 
rx::SwapChain11::resetOffscreenColorBuffer(rx::DisplayD3D * displayD3D, int backbufferWidth, int backbufferHeight) Line 356 
rx::SwapChain11::resetOffscreenBuffers(rx::DisplayD3D * displayD3D, int backbufferWidth, int backbufferHeight) Line 165
rx::SwapChain11::resize(rx::DisplayD3D * displayD3D, int backbufferWidth, int backbufferHeight) Line 535 
rx::SurfaceD3D::resizeSwapChain(rx::DisplayD3D * displayD3D, int backbufferWidth, int backbufferHeight) Line 234 
rx::SurfaceD3D::checkForOutOfDateSwapChain(rx::DisplayD3D * displayD3D) Line 353 
rx::SurfaceD3D::swapRect(rx::DisplayD3D * displayD3D, int x, int y, int width, int height) Line 313
rx::SurfaceD3D::postSubBuffer(const gl::Context * context, int x, int y, int width, int height) Line 373
egl::Surface::postSubBuffer(const gl::Context * context, int x, int y, int width, int height) Line 337 
egl::PostSubBufferNV(egl::Thread * thread, egl::Display * display, egl::Surface * eglSurface, int x, int y, int width, int height) Line 245 
EGL_PostSubBufferNV(void * dpy, void * surface, int x, int y, int width, int height) Line 1099 
flutter::AngleSurfaceManager::ResizeSurface(std::variant<HWND__ *> * render_target, int width, int height) Line 268 

[clarkezone]

I have repro'd the issue and am working on a fix that address both the regression and the UWP scenario.

[rydmike]

Thanks for the link to my tweet https://twitter.com/RydMike/status/1368694991355338763?s=20

The flashy video "after-pr.mov" looks pretty identical to the behavior I observed too, so I'm (pretty) convinced it is this same issue. Good to know it is already registered and worked on. 👍🏻

[clarkezone]

NP. Out of interest is your repro case involving WM_NCCALCSIZE in any way? If no it suggests that the regression is more extensive than previously realized and I'll consider getting a temporary workaround in place pending the full fix.