tzeng@miro.Berkeley.EDU (Ping-San Tzeng) (02/23/90)
I found some confusing performance data about the R4 server and OpenWindows 1.0 (XNEWS) on my SPARC-1 (No GX board). For my own program that draws a lot vectors with "XDrawLines", OpenWindows is about 3 times faster than the R4 server to draw the complete picture on the screen. (measured with a stop watch, not CPU time) However, for "x11perf -line???" (??? is any legal number), the results show that R4 server is about 50% faster than the OpenWindows. Does anyone know why? Ping-san.
keith@EXPO.LCS.MIT.EDU (Keith Packard) (02/24/90)
> I found some confusing performance data about > the R4 server and OpenWindows 1.0 (XNEWS) on my SPARC-1 (No GX board). > > For my own program that draws a lot vectors with "XDrawLines", > OpenWindows is about 3 times faster than the R4 server > to draw the complete picture on the screen. > (measured with a stop watch, not CPU time) > > However, for "x11perf -line???" (??? is any legal number), > the results show that R4 server is about 50% faster than the OpenWindows. > > Does anyone know why? The R4 server was tuned to draw only copy-mode lines with a planemask of all ones. If your application uses some other raster op, or other planemask, the performance of the R4 server will be abysmal. Time constraints prevented me from spending the week necessary to rewrite the line code for faster non-copy mode lines. Also, x11perf attempts to represent the "perfect" client. The performance data it generates are best case numbers. In the case of XDrawLines, x11perf draws hundreds of lines in each request. The R4 server was extensively tuned to run x11perf quickly which could possibly have impacted the performance for clients which draw few objects at a time (I don't think this is the case, but I have no performance data to support this position). So, if you application drew more lines at a time you could expect the performance of the R4 server to improve. Given the wide disparity between the XNEWS server and the R4 server for your application, I suspect the first cause is more likely than the second. Keith Packard MIT X Consortium
joel@decwrl.dec.com (Joel McCormack) (02/24/90)
The obvious conclusion would be that the R4 server performs some other operation verrrry slowly, which eats up all of its advantage, or that the way you draw lines is not precisely what x11perf measures. What rasterop are you using to paint the window? In the released R4 server, almost anything other than GXcopy (and sometimes GXxor) is quite slow. Also, any operation that requires a planemask that is not all 1's is quite slow. - Joel McCormack (decwrl!joel, joel@decwrl.dec.com)
tzeng@MIRO.BERKELEY.EDU (Ping-San Tzeng) (02/24/90)
My program use only copy-mode and a planemask of all ones. On each call of XDrawLines, I draw about 4-20 lines. By the way, I draw my picture in a XView window. Does this matter? (Direct Xlib calls, not pw-ops) Ping-san.
msm@src.dec.com (Mark S. Manasse) (02/24/90)
If I had to guess, I'd say that what you're seeing is the performance of width 1 lines. The X protocol distinguishes width 0 lines (which can be rendered in an arbitrary way) from width 1 lines (which must satisfy certain mathematical properties). The R4 server handles width 1 lines by using the general wide-line code. The XNews server treats width 1 specially. Some suspicions were raised at the X conference that the resulting lines don't meet the specification, but I don't know for sure. Mark
tzeng@miro.Berkeley.EDU (Ping-San Tzeng) (02/24/90)
In article <14643@jumbo.dec.com> msm@src.dec.com (Mark S. Manasse) writes:
=If I had to guess, I'd say that what you're seeing is the performance of
=width 1 lines. ........ The R4 server handles width 1 lines
=by using the general wide-line code. The XNews server treats width 1
=specially.
YES. All lines are of width 1. I think this is the answer.
Thanks!
Ping-san.