Pixman 0.28 Packs In Major Features

[phoronix]

Phoronix: Pixman 0.28 Packs In Major Features

The Pixman rendering library is out with a major new release that presents several interesting new features...

http://www.phoronix.com/vr.php?view=MTIyNDA

[mattst88]

Improvements from the bilinear scaling code are written up here:

http://mattst88.com/blog/2012/07/06/...O_1,75_laptop/

HTML5 stuff looks like it's significantly faster on ARM/iwMMX now.

[madbiologist]

I've had to wait a while for libpixman 0.28.2 to be added to Ubuntu's xorg-edgers PPA, but finally....

Since this pixman release mentions Faster bilinear scaling on iwMMX, Loongson _and_ MMX, as well as Faster scaling in general... , I decided to test it on an ancient Athlon XP 2700+ (has MMX and SSE, but not SSE2). I installed cairo-perf-utils to get cairo-perf-trace but I was unable to uncompress the .lzma files I had downloaded from http://cgit.freedesktop.org/cairo-traces/tree/benchmark so instead I opened Firefox 17.0.1 and ran the Asteroids HTML5 Canvas 2D Rendering and JavaScript Benchmark at http://www.kevs3d.co.uk/dev/asteroidsbench/

Asteroids results on Ubuntu 12.04:

Before (libpixman-1-0 0.24.4-1) - score: 295, FPS: 7
After (libpixman-1-0 0.28.2-0ubuntu1~xedgers~precise1) - score: 320, FPS: 7

So an 8.5% increase. I'm not sure if that is significant or not.

[ssvb]

Before (libpixman-1-0 0.24.4-1) - score: 295, FPS: 7
After (libpixman-1-0 0.28.2-0ubuntu1~xedgers~precise1) - score: 320, FPS: 7

So an 8.5% increase. I'm not sure if that is significant or not.

Considering that this test is about "commonly used operations in HTML5 games: drawImage, drawImage scaling, alpha, composition, shadows and text functions", which is a bit more than just scaling, the improvement you got might be not too bad.

But to get a better idea where the time is spent, it makes sense to profile this use case. For the start, on ARM Cortex-A8 1GHz with firefox17 (compiled to use system pixman and cairo libraries instead of mozilla's internal copies) and "gfx.xrender.enabled" set to "false" in about:config, I'm getting the following scores in the same asteroids benchmark:

ARM Cortex-A8 1GHz:

Benchmark Score: 197
Average FPS: 5

And profiling it with "perf record -p `pidof firefox`" while running the test, results in the following overall statistics:

$ perf report -s dso

    37.92%  libxul.so                  
    21.40%  libcairo.so.2.11200.8      
    16.24%  libpixman-1.so.0.28.2      
     7.51%  libfontconfig.so.1.4.4     
     3.49%  firefox                    
     2.56%  perf-984.map               
     2.07%  libz.so.1.2.7              
     1.86%  libm-2.15.so               
     1.63%  [kernel.kallsyms]          
     1.52%  libc-2.15.so               
     1.14%  ld-2.15.so                 
     0.41%  [unknown]

Taking a look at the distribution of time spent in different libraries, we see that pixman is not the biggest contributor here.

And for just pixman alone, we get the following:

$ perf report | grep libpixman

     4.93%  firefox  libpixman-1.so.0.28.2        [.] pixman_composite_over_8888_8888_asm_neon
     2.98%  firefox  libpixman-1.so.0.28.2        [.] pixman_composite_src_8888_8888_asm_neon
     2.14%  firefox  libpixman-1.so.0.28.2        [.] pixman_composite_src_x888_8888_asm_neon
     2.06%  firefox  libpixman-1.so.0.28.2        [.] pixman_scaled_bilinear_scanline_8888_8_8888_OVER_asm_neon
     0.44%  firefox  libpixman-1.so.0.28.2        [.] pixman_composite_src_n_8888_asm_neon
     0.41%  firefox  libpixman-1.so.0.28.2        [.] pixman_composite_over_n_8_8888_asm_neon
     0.29%  firefox  libpixman-1.so.0.28.2        [.] pixman_image_composite32
     0.29%  firefox  libpixman-1.so.0.28.2        [.] fast_composite_scaled_bilinear_neon_8888_8_8888_pad_OVER
     0.29%  firefox  libpixman-1.so.0.28.2        [.] analyze_extent
     0.27%  firefox  libpixman-1.so.0.28.2        [.] pixman_composite_add_n_8_8888_asm_neon
     0.23%  firefox  libpixman-1.so.0.28.2        [.] pixman_transform_point
     0.17%  firefox  libpixman-1.so.0.28.2        [.] _pixman_image_validate
     0.14%  firefox  libpixman-1.so.0.28.2        [.] compute_transformed_extents

The scaling operations are not particularly heavy here. Mostly it's just simple non-scaled alpha blending and blitting.

If you are also wondering about what is done by libcairo:

$ perf report | grep libcairo

     4.91%  firefox  libcairo.so.2.11200.8        [.] _cairo_tor_scan_converter_generate
     1.11%  firefox  libcairo.so.2.11200.8        [.] _fill_xrgb32_lerp_opaque_spans
     0.72%  firefox  libcairo.so.2.11200.8        [.] cell_list_render_edge
     0.64%  firefox  libcairo.so.2.11200.8        [.] _cairo_spline_decompose_into
     0.60%  firefox  libcairo.so.2.11200.8        [.] cairo_matrix_transform_point
     0.44%  firefox  libcairo.so.2.11200.8        [.] _cairo_tor_scan_converter_add_polygon
     0.35%  firefox  libcairo.so.2.11200.8        [.] .divsi3_skip_div0_test
     0.34%  firefox  libcairo.so.2.11200.8        [.] _cairo_bentley_ottmann_tessellate_polygon
     0.30%  firefox  libcairo.so.2.11200.8        [.] 0x0000e0b4
     0.25%  firefox  libcairo.so.2.11200.8        [.] __udivsi3
     0.25%  firefox  libcairo.so.2.11200.8        [.] outer_join
     0.25%  firefox  libcairo.so.2.11200.8        [.] compute_face
     0.20%  firefox  libcairo.so.2.11200.8        [.] _cairo_matrix_to_pixman_matrix_offset

You can also try a similar profiling exercise on your hardware

[madbiologist]

OK, I had to use sudo to get perf to do anything useful (otherwise it refused to run with an error about kernel symbols). And "perf report -s dso" just results in "perf: Segmentation fault". perf report works though

I think my initial test may have been meaningless. I did not realise that Firefox has internal copies of pixman and cairo and did not recompile it to use the system versions. That probably explains why perf report just shows:

55.14%  DNS Res~ver #37  libxul.so                    [.] 0xf4aa9c
 17.13%  DNS Res~ver #37  libfontconfig.so.1.4.4       [.] 0x72c8
  5.74%  DNS Res~ver #37  [kernel.kallsyms]            [k] 0xc1035df8
  5.46%  DNS Res~ver #37  firefox                      [.] 0xc930
  1.77%  DNS Res~ver #37  perf-2141.map                [.] 0x1235b22d
  1.73%  DNS Res~ver #37  libc-2.15.so                 [.] 0x7d710
  1.34%  DNS Res~ver #37  libfontconfig.so.1.4.4       [.] FcStrCmpIgnoreCase
  1.27%  DNS Res~ver #37  libfontconfig.so.1.4.4       [.] FcConfigSubstituteWit
  0.80%  DNS Res~ver #37  libpthread-2.15.so           [.] pthread_mutex_lock 
  0.34%  DNS Res~ver #37  libxul.so                    [.] vp8_idct_dequant_dc_f
  0.32%  DNS Res~ver #37  libfontconfig.so.1.4.4       [.] FcUtf8ToUcs4
  0.31%  DNS Res~ver #37  firefox                      [.] _start

and why perf report | grep libpixman just shows:

[kernel.kallsyms] with build id dd11b20eac2611858a619b764b7796765d2ad877 not found, continuing without symbols
Failed to open /tmp/perf-2141.map, continuing without symbols
     0.00%  DNS Res~ver #37  libpixman-1.so.0.28.2        [.] pixman_image_composite32

I was originally hoping to benchmark poppler as well, but as I said in my previous post, I was unable to uncompress the .lzma files I had downloaded from the cairo-traces site
I've just worked out that I had to use the "plain" link in the far right column, instead of right-clicking on filename and selecting "Save Link As..." I'll report back how that goes.

[madbiologist]

OK, it still doesn't work. I got the file uncompressed this time, but cairo-perf-trace just gives:

Error: Found no traces in 'cairo-traces:/usr/src/cairo-traces:/usr/share/cairo-traces'.
Have you cloned the cairo-traces repository and uncompressed the traces?
git clone git://anongit.freedesktop.org/cairo-traces
cd cairo-traces && make
Or set the env.var CAIRO_TRACE_DIR to point to your traces?

ls -l /usr/share/cairo-traces gives:

total 42860
-r-------- 1 user user 43886133 Jan 9 00:29 poppler

I tried setting the environment variable (set env CAIRO_TRACE_DIR /usr/share/cairo-traces) anyway but it made no difference. Ubuntu 12.04 here.