Adding an accelerated graphics driver to Qt/Embedded

Qt/Embedded has the capacity to make use of hardware accelerators. To do so for a PCI or AGP driver, you need to perform the following steps:

1) Define an accelerated descendant of QLinuxFbScreen. This should implement QVoodooScreen::connect to map its registers. Use qt_probe_bus to get a pointer to PCI config space. This is where you should check that you're being pointed at the right device (using PCI device/manufacturer ID information). Then use PCI config space to locate your device's accelerator registers in physical memory and mmap the appropriate region in from /dev/mem. There is no need to map the framebuffer, QLinuxFbScreen will do this for you. Return FALSE if a problem occurs at any point. QVoodooScreen::initDevice will be called only by the QWS server and is guaranteed to be called before any drawing is done (and so is a good place to set registers to known states). connect will be called by every connecting client.

2) Define an accelerated descendant of QGfxRaster. This is where the actual drawing code goes. Anything not implemented by hardware can be passed back to QGfxRaster to do in software. Use the optype variable to make sure that accelerated and unaccelerated operations are synchronised (if you start drawing via software into an area where the hardware accelerator is still drawing then your drawing operations will appear to be in the wrong order). optype is stored in shared memory and is set to 0 by unaccelerated operations; accelerated operations should set it to 1. When a software graphics operation is requested and optype is 1, QGfxRaster::sync() is called; you should provide your own implementation of this which waits for the graphics engine to go idle. lastop is also available for optimisation and is stored in the shared space - this will not be set by the software-only QGfx and can be used to store the type of your last operation (e.g. drawing a rectangle) so that part of the setup for the next operation can be avoided when a lot of the same operations are performed in sequence. All drawing operations should be protected via a QWSDisplay::grab() before any registers, lastop or optype are accessed, and ungrabbed() at the end. This prevents two applications trying to access the accelerator at once and possibly locking up the machine. It's possible that your source data is not on the graphics card so you should check in such cases and fall back to software if necessary. Note that QGfxRaster supports some features not directly supported by QPainter (for instance, alpha channels in 32-bit data and stretchBlt's). These features are used by Qt; stretchBlt speeds up QPixmap::xForm and drawPixmap into a transformed QPainter, alpha channel acceleration is supported for 32-bit pixmaps.

3) If you wish, define an accelerated descendant of QScreenCursor. restoreUnder,saveUnder,drawCursor and draw should be defined as null operations. Implement set, move, show and hide. 4k is left for your cursor at the end of the visible part of the framebuffer (i.e. at (width*height*depth)/8 )

4) Implement initCursor and createGfx in your QScreen descendant. Implement useOffscreen and return true if you can make use of offscreen graphics memory.

5) Implement a small function qt_get_screen_mychip, which simply returns a new QMychipScreen

6) Add your driver to the DriverTable table in qgfxraster_qws.cpp - e.g.

{ "MyChip", qt_get_screen_mychip,1 },

The first parameter is the name used with QWS_DISPLAY to request your accelerated driver.

7) To run with your new driver, export QWS_DISPLAY=MyChip (optionally MyChip:/dev/fb to request a different Linux framebuffer than /dev/fb0) run the program

If your driver is not PCI or AGP you'll need to inherit QScreen instead of QLinuxFbScreen and implement similar functionality to QLinuxFbScreen, but otherwise the process should be similar. The most complete example driver is qgfxmach64_qws.cpp; qgfxvoodoo_qws.cpp may provide a smaller and more easy to understand driver.

Copyright © 2005 TrolltechTrademarks
Qt version 2.3.10