Archive for the ‘General’ Category

GTK+ performance without RENDER

Wednesday, November 2nd, 2005

Federico, cworth and I came up with a patch to Cairo that greatly improves the performance of GTK+ applications on X servers without the RENDER extension.

In this case the profiler didn’t really help much, it just showed that all of the time was spent in the X server. The real progress was made when Federico just made a guess on IRC, asking about how gdk_window_clear_backing_rect() goes through to cairo, and how the RENDER fallback code worked.

I think we were very lucky here. It’s hard to know whether it’s worth investigating performance on X servers without RENDER. While some people care a lot about these configurations, it is tempting to just let them deal with it. That’s the “open source” way, right? However, I found it disappointing to see how easy Federico’s steps were to recreate the problem, how obviously broken it was, and how simple the fix seems to be.

Sysprof fun

Wednesday, October 26th, 2005

I got sysprof up and running tonight along with Federico‘s latest patches to Pango. Just looking at the shaper, it’s getting much harder to find things to optimize. Still, we had a great time trying. Sysprof is really fun to use, and it was fun hanging out on gimpnet #performance.

My test case was pango-language-profile.c from Federico’s last blog post, using --lang=es. I spent some time on pangocairo-fcfont.c where pango_font_get_glyph_extents was sitting at 7% or so. Here was the initial output from sysprof:

Using gcc-4.0.2 and -Os, the memcpy() was added by the compiler to copy some PangoRectangle structures. I expanded those to just copy the individual fields, and the time basically went to 0. While I suspect this was just a side-effect of -Os, expanding the copy is easy and I don’t think it hurts the readability of the code. Maybe we should consider it.

The more interesting optimization was with the hash table used to cache the glyph extents calculated through Cairo. It starts at a size of 11 and grows by primes, using the glyph index as the hash.

My proposal is to instead use a fixed-size hash with a large fixed-size hash table, and instead of growing by primes, just give it a power-of-two size. I figure characters in any given sentence will mostly be from the same alphabet, clumped in some area of the unicode space. I’m also assuming that many languages start on 256-glyph aligned chunks or so, but I figure this is less important. For my patch, I chose 1024 as a reasonable size for the array per font, giving lots of room without being much of a memory hog over the original. Here’s the result:

The patch to Federico’s latest thing can be found over here: new-cache.diff. In case you’re having trouble, my current cvs diff for pango is here: my-pango.diff.

Using OProfile

Wednesday, October 12th, 2005

Sysprof is great, and the tree view UI makes it quick and easy to narrow down performance issues. The call stack information is invaluable, and being able to quickly sort the times is awesome.

OProfile, like sysprof, is a kernel-level profiler which I’ve been using for working on cairo. While it lacks sysprof’s UI, the command line interface can be handy, and the resulting data is very raw and believable.

You basically use it like this:

  # Clear the data from the last run first
  opcontrol --reset
  opcontrol --start
  # Run your benchmark here
  opcontrol --stop

The opcontrol commands are run as root, while the benchmark can run anywhere. OProfile gathers data for the entire system.

There are four commands I found useful for gathering information from OProfile. First, it can break down the samples by binary:

# opreport --threshold 4.0
  samples|      %|
   807298 38.1359
   446257 21.0807
   364723 17.2291
   148549  7.0173
    92573  4.3731

You can then pick a library, and see the top hit symbols for that library:

# opreport --symbols --debug-info \
  --threshold 3.0 /.../ 
samples  %        linenr info         symbol name
151514   18.7680  pango-script.c:98   pango_script_for_unichar
81120    10.0483  break.c:447         pango_default_break
76167     9.4348  pango-script.c:214  get_pair_index
59039     7.3132  fribidi.c:516       fribidi_analyse_string
30140     3.7334  glyphstring.c:148   pango_glyph_string_extents_range

OProfile can even narrow down which parts of the code were hit in that symbol, although the measurements here are usually too small to be reliable:

# opreport --symbols --debug-info --threshold 1.0 \
    --details --include-symbols get_pair_index \
vma      samples  %        linenr info         symbol name
0001a7b6 76167    100.000  pango-script.c:214  get_pair_index
0001a7b6 189       0.2481  pango-script.c:214
0001a7de 1093      1.4350  pango-script.c:220
0001a7e0 58535    76.8509  pango-script.c:222
0001a7e3 1121      1.4718  pango-script.c:222

OProfile can also generate annotated source code based on this information. The annotated code very visually shows which code paths get hit.

# opannotate --source --output-dir=/root/oprofile/source \

               :static void
               :compact_list (TypeLink *list)
   341  0.0422 :{ /* compact_list total:   7755  0.9606 */
  2003  0.2481 :  if (list->next)
  1389  0.1721 :    for (list = list->next; list; list = list->next)
  2648  0.3280 :      if (RL_TYPE (list->prev) == RL_TYPE (list)
               :          && RL_LEVEL (list->prev) == RL_LEVEL (list))
   141  0.0175 :        list = merge_with_prev (list);
  1233  0.1527 :}

All of the data here is from a simple text measuring benchmark under pango 1.10.1. I had to build pango without inlining to show which functions were really doing the work. This slows the benchmark down, making it more difficult to evaluate changes. Only use the non-inlined results as a guide.

Steps to enlightenment:

  1. Run your benchmark a bunch of times, record timing information
  2. Compile without inlining (I used “-Wall -g -Os -fno-inline”)
  3. Run your benchmark and gather profiler output
  4. Make some performance-enhancing modifications and test them
  5. Run your benchmark and gather profiler output, estimate speedup
  6. Recompile with normal CFLAGS
  7. Run your benchmark a bunch of times, record timing information, compare with step 1
  8. Post the patch and brag about it in your blog