1 files changed, 85 insertions, 0 deletions
diff --git a/Documentation/fb/cyblafb/whycyblafb b/Documentation/fb/cyblafb/whycyblafb
new file mode 100644
index 0000000..a123bc1
--- /dev/null
+++ b/Documentation/fb/cyblafb/whycyblafb
@@ -0,0 +1,85 @@
+I tried the following framebuffer drivers:
+
+	- TRIDENTFB is full of bugs. Acceleration is broken for Blade3D
+	  graphics cores like the cyberblade/i1. It claims to support a great
+	  number of devices, but documentation for most of these devices is
+	  unfortunately not available. There is _no_ reason to use tridentfb
+	  for cyberblade/i1 + CRT users. VESAFB is faster, and the one
+	  advantage, mode switching, is broken in tridentfb.
+
+	- VESAFB is used by many distributions as a standard. Vesafb does
+	  not support mode switching. VESAFB is a bit faster than the working
+	  configurations of TRIDENTFB, but it is still too slow, even if you
+	  use ypan.
+
+	- EPIAFB (you'll find it on sourceforge) supports the Cyberblade/i1
+	  graphics core, but it still has serious bugs and developement seems
+	  to have stopped. This is the one driver with TV-out support. If you
+	  do need this feature, try epiafb.
+
+None of these drivers was a real option for me.
+
+I believe that is unreasonable to change code that announces to support 20
+devices if I only have more or less sufficient documentation for exactly one
+of these. The risk of breaking device foo while fixing device bar is too high.
+
+So I decided to start CyBlaFB as a stripped down tridentfb.
+
+All code specific to other Trident chips has been removed. After that there
+were a lot of cosmetic changes to increase the readability of the code. All
+register names were changed to those mnemonics used in the datasheet. Function
+and macro names were changed if they hindered easy understanding of the code.
+
+After that I debugged the code and implemented some new features. I'll try to
+give a little summary of the main changes:
+
+	- calculation of vertical and horizontal timings was fixed
+
+	- video signal quality has been improved dramatically
+
+	- acceleration:
+
+		- fillrect and copyarea were fixed and reenabled
+
+		- color expanding imageblit was newly implemented, color
+		  imageblit (only used to draw the penguine) still uses the
+		  generic code.
+
+		- init of the acceleration engine was improved and moved to a
+		  place where it really works ...
+
+		- sync function has a timeout now and tries to reset and
+		  reinit the accel engine if necessary
+
+		- fewer slow copyarea calls when doing ypan scrolling by using
+		  undocumented bit d21 of screen start address stored in
+		  CR2B[5]. BIOS does use it also, so this should be safe.
+
+	- cyblafb rejects any attempt to set modes that would cause vclk
+	  values above reasonable 230 MHz. 32bit modes use a clock
+	  multiplicator of 2, so fbset does show the correct values for
+	  pixclock but not for vclk in this case. The fbset limit is 115 MHz
+	  for 32 bpp modes.
+
+	- cyblafb rejects modes known to be broken or unimplemented (all
+	  interlaced modes, all doublescan modes for now)
+
+	- cyblafb now works independant of the video mode in effect at startup
+	  time (tridentfb does not init all needed registers to reasonable
+	  values)
+
+	- switching between video modes does work reliably now
+
+	- the first video mode now is the one selected on startup using the
+	  vga=???? mechanism or any of
+		- 640x480, 800x600, 1024x768, 1280x1024
+		- 8, 16, 24 or 32 bpp
+		- refresh between 50 Hz and 85 Hz, 1 Hz steps (1280x1024-32
+		  is limited to 63Hz)
+
+	- pci retry and pci burst mode are settable (try to disable if you
+	  experience latency problems)
+
+	- built as a module cyblafb might be unloaded and reloaded using
+	  the vfb module and con2vt or might be used together with vesafb
+