Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!swrinde!zaphod.mps.ohio-state.edu!sdd.hp.com!ucsd!ucbvax!bloom-beacon!EXPO.LCS.MIT.EDU!keith
From: keith@EXPO.LCS.MIT.EDU (Keith Packard)
Newsgroups: comp.windows.x
Subject: Re: XFillRectangle with Tile Vs Manual tiling
Message-ID: <9011071643.AA27346@xenon.lcs.mit.edu>
Date: 7 Nov 90 16:42:59 GMT
References: <723@uni2a.unige.ch>
Sender: daemon@athena.mit.edu (Mr Background)
Organization: The Internet
Lines: 37


> In both cases I was using a Pixmap to tile a drawable.  I found (much to my
> surprise) that the not-so-obvious way was MUCH faster (one my Sparcstation
> 1) with an area about twelve times the area of the tile.  This seems bizzare
> to me, after all the obvious way requires one trip to the server, while the
> not-so-obvious way requires many trips. 

This shouldn't seem so strange to you.  Tiling operations are very complicated
and cover a wide range of potential usages.  The most common tiling operations
use small tiles (< 32x32) to fill areas on the screen.  The R4 server optimizes
this case, and leaves the other cases to wither with pretty horrible code.  

The problem with other tile sizes is that the server needs to choose to either
emulate bitblt code, and plop down repeated blocks of the tile, or to attempt
something different.  For smaller tiles, emulating bitblt is painfully slow;
the frame buffer is repeatedly modified at the same addresses to lay down those
bits.  Larger tiles could make effective use of the bitblt code without the
performance penalty

So, while it may seem like the server should always choose the 'fastest' way of
doing something, it is also the case that the time spent optimizing various
of these algorithms is better spent doing something else.  You could spend an
arbitrary amount of time optimizing various potential combinations of client
requests; this one has never reached the top of the heap.

Yes, the R5 server will do the tiling case faster than the bitblt case on color
frame buffers.  I fixed your example which uses CopyArea to work correctly on
CFB (removing the CopyPlane call and disabling GraphicsExpose events) and
measured:

	24 seconds	Tiling code
	34 seconds	Bitblt code

I sped this up several months ago so that xphoon would run reasonably fast on
my display.

Besides, if everything went as fast as it could, I'd be out of work, wouldn't I.