Allocate only the vertices we need for text.

Allocate only the vertices we need for text.

This restructures the vertex allocation for text rendering to compute
the max number of vertices we would need for a line of text, and then
only allocate that much. If this number exceeds the quad index limit,
then it will allocate for the max number of quads, and reallocate for
the rest later.

Committed: https://skia.googlesource.com/skia/+/73f105345066b1f22f15ba6575cb3800dd9313b8

[jvanverth1, 2014-10-21 20:11:38]

jvanverth@google.com changed reviewers:
+ bsalomon@google.com, robertphillips@google.com

[jvanverth1, 2014-10-21 20:11:38]

jvanverth@google.com changed required reviewers:
+ bsalomon@google.com

[jvanverth1, 2014-10-21 20:11:38]

This appears to be slightly faster and less noisy. Compare before (on Mac OS):

maxrss	loops	min	median	mean	max	stddev	samples   	config	bench
  19M	1717	2.75µs	2.88µs	2.86µs	3.05µs	4%	▁▂▂▂▁█▆▄▄▅	gpu	text_16_pos_AA_BK
  19M	1800	2.55µs	2.57µs	2.57µs	2.58µs	0%	▆█▃▅▄▅▄▅▁▅	gpu	text_16_LCD_88
  19M	1810	2.51µs	2.58µs	2.59µs	2.68µs	3%	▄▇█▅█▄▁▁▂▁	gpu	text_16_LCD_FF
  20M	1785	2.53µs	2.57µs	2.57µs	2.68µs	2%	▂▁▃▁▃▃▄█▃▂	gpu	text_16_LCD_BK
  21M	1747	2.61µs	2.62µs	2.62µs	2.63µs	0%	█▃▃▃▃▁▂▂▃▂	gpu	text_16_AA_88
  21M	1732	2.62µs	2.64µs	2.65µs	2.72µs	1%	█▁▄▂▄▂▃▃▃▂	gpu	text_16_AA_FF
  22M	1752	2.62µs	2.64µs	2.63µs	2.65µs	0%	▇▇▅▄█▃▆▁█▁	gpu	text_16_AA_BK
  22M	1751	2.61µs	2.62µs	2.63µs	2.77µs	2%	█▁▂▁▁▁▁▁▂▁	gpu	text_16_BW_88
  23M	1743	2.63µs	2.74µs	2.78µs	3.1µs	5%	▂▃█▄▃▂▁▁▂▆	gpu	text_16_BW_FF
  24M	1739	2.62µs	2.64µs	2.71µs	3.23µs	7%	▁▁█▁▁▁▃▁▁▁	gpu	text_16_BW_BK
  41M	543	9.04µs	9.05µs	9.05µs	9.08µs	0%	▄▆▃▁▂█▄▃▁▆	gpu	cmap_paint_textToGlyphs
  47M	13	322µs	325µs	328µs	344µs	2%	▁▁▅▅▄▂█▁▂▁	gpu	GM_varied_text_clipped_lcd
 
50M	13	292µs	313µs	330µs	443µs	14%	▁█▂▂▃▄▄▂▁▁	gpu	GM_varied_text_ignorable_clip_lcd
  50M	16	301µs	309µs	310µs	330µs	3%	▃▁█▃▃▂▃▄▂▃	gpu	GM_varied_text_clipped_no_lcd
 
52M	18	228µs	233µs	244µs	327µs	12%	▁▁▁▂▂▂█▁▁▁	gpu	GM_varied_text_ignorable_clip_no_lcd

And after:
maxrss	loops	min	median	mean	max	stddev	samples   	config	bench
  18M	810	2.88µs	2.9µs	2.9µs	2.91µs	0%	▃█▁▅▇▂▆▅▅▃	gpu	text_16_pos_AA_BK
  19M	1842	2.53µs	2.55µs	2.55µs	2.57µs	0%	▁▅▅▆▅█▆▆▅▄	gpu	text_16_LCD_88
  20M	1856	2.53µs	2.55µs	2.56µs	2.73µs	2%	█▁▂▁▂▂▂▁▂▁	gpu	text_16_LCD_FF
  21M	1845	2.54µs	2.56µs	2.56µs	2.57µs	0%	▁▅▄▃▅▄▄▃█▅	gpu	text_16_LCD_BK
  21M	1728	2.66µs	2.68µs	2.68µs	2.71µs	1%	▄▄▃▇▂█▇▂▁▁	gpu	text_16_AA_88
  21M	1727	2.64µs	2.66µs	2.68µs	2.82µs	2%	█▅▁▁▁▂▁▂▂▁	gpu	text_16_AA_FF
  21M	1732	2.62µs	2.64µs	2.63µs	2.65µs	0%	▇▁▄▂▅▃▅▂█▅	gpu	text_16_AA_BK
  23M	1785	2.64µs	2.65µs	2.65µs	2.65µs	0%	█▆▅▅▄▃▄▃▁▇	gpu	text_16_BW_88
  23M	1792	2.63µs	2.64µs	2.64µs	2.65µs	0%	█▁▂▃▁▇▂▂▃▁	gpu	text_16_BW_FF
  23M	1750	2.65µs	2.65µs	2.66µs	2.7µs	1%	█▂▁▁▁▃▂▄▅▂	gpu	text_16_BW_BK
  41M	558	8.95µs	8.99µs	8.99µs	9.04µs	0%	▅▅▂▂▇▄█▄▂▁	gpu	cmap_paint_textToGlyphs
  47M	16	320µs	328µs	331µs	355µs	4%	▃▁▁█▄▁▆▄▂▁	gpu	GM_varied_text_clipped_lcd
 
48M	18	285µs	293µs	292µs	311µs	3%	▂▃▂▁▄▃▁▃█▁	gpu	GM_varied_text_ignorable_clip_lcd
  49M	17	296µs	297µs	304µs	345µs	5%	▁▄▁▁▁▁▁█▁▁	gpu	GM_varied_text_clipped_no_lcd
 
51M	18	232µs	235µs	237µs	244µs	2%	▂▁▂▁▆█▃▂▇▇	gpu	GM_varied_text_ignorable_clip_no_lcd

[jvanverth1, 2014-10-21 20:16:06]

Well, looking again, any perf changes (including the stddev) may be noise. Some
are faster, some are slower, all are about the same.

[mtklein, 2014-10-21 20:26:49]

On 2014/10/21 20:16:06, jvanverth1 wrote:
> Well, looking again, any perf changes (including the stddev) may be noise.
Some
> are faster, some are slower, all are about the same.

For what it's worth, I've been building some little scripts to help this sort of
comparison: https://gist.github.com/mtklein/7fe3b12ef0b6c294c387

I just ran "c --match skp --config nonrendering" and got a print out like this:
                desk_youtube.skp	 975us ->  884us	0.91x
desk_googlespreadsheetdashed.skp	1.75ms -> 1.58ms	0.9x
                   tabl_cnet.skp	 670us ->  594us	0.89x
                 desk_carsvg.skp	 533us ->  469us	0.88x
           tabl_worldjournal.skp	 770us ->  675us	0.88x
    desk_rectangletransition.skp	16.9us -> 14.8us	0.87x
...
desk_youtubetvvideo.skp	25.1us -> 15.3us	0.61x
               tabl_frantzen.skp	 165us -> 96.6us	0.59x
                tabl_mozilla.skp	3.59ms -> 2.09ms	0.58x
             desk_forecastio.skp	 172us -> 96.8us	0.56x
               desk_fontwipe.skp	  16us -> 6.16us	0.38x
            desk_oldinboxapp.skp	14.5us -> 4.45us	0.31x

If you find this handy, perhaps we can check these scripts in.

[mtklein, 2014-10-21 20:39:19]

On my Mac, --match text --config gpu,                      text_16_LCD_FF	2.76us
-> 2.79us	1x
   GM_varied_text_ignorable_clip_lcd	 310us ->  313us	1x
GM_varied_text_ignorable_clip_no_lcd	 245us ->  246us	1x
                       text_16_AA_88	 2.9us -> 2.89us	1x
                       text_16_BW_88	2.87us -> 2.86us	1x
          GM_varied_text_clipped_lcd	 353us ->  350us	0.99x
                       text_16_AA_BK	2.91us -> 2.89us	0.99x
                       text_16_BW_BK	 2.9us -> 2.87us	0.99x
                   text_16_pos_AA_BK	3.06us -> 3.02us	0.99x
                       text_16_BW_FF	 2.9us -> 2.85us	0.98x
       GM_varied_text_clipped_no_lcd	 325us ->  316us	0.97x

[bsalomon, 2014-10-22 13:22:55]

Love the simplification. If it is faster that's great.

https://codereview.chromium.org/663423003/diff/20001/src/gpu/GrBitmapTextCont...
File src/gpu/GrBitmapTextContext.cpp (left):

https://codereview.chromium.org/663423003/diff/20001/src/gpu/GrBitmapTextCont...
src/gpu/GrBitmapTextContext.cpp:457: bool flush =
fDrawTarget->geometryHints(&fMaxVertices, NULL);
Are all callers of geometryHints() gone? I'd love to delete that.

https://codereview.chromium.org/663423003/diff/20001/src/gpu/GrDistanceFieldT...
File src/gpu/GrDistanceFieldTextContext.cpp (right):

https://codereview.chromium.org/663423003/diff/20001/src/gpu/GrDistanceFieldT...
src/gpu/GrDistanceFieldTextContext.cpp:387:
drawTarget->drawState()->setVertexAttribs<gTextVertexWithColorAttribs>(
line wraps in here

https://codereview.chromium.org/663423003/diff/20001/src/gpu/GrTextContext.h
File src/gpu/GrTextContext.h (right):

https://codereview.chromium.org/663423003/diff/20001/src/gpu/GrTextContext.h#...
src/gpu/GrTextContext.h:60: static int MeasureText(SkGlyphCache* cache,
SkDrawCacheProc glyphCacheProc,
comment on return value?

[jvanverth1, 2014-10-23 17:33:44]

https://codereview.chromium.org/663423003/diff/20001/src/gpu/GrBitmapTextCont...
File src/gpu/GrBitmapTextContext.cpp (left):

https://codereview.chromium.org/663423003/diff/20001/src/gpu/GrBitmapTextCont...
src/gpu/GrBitmapTextContext.cpp:457: bool flush =
fDrawTarget->geometryHints(&fMaxVertices, NULL);
On 2014/10/22 13:22:55, bsalomon wrote:
> Are all callers of geometryHints() gone? I'd love to delete that.

It's used in GrInOrderDrawBuffer::willReserveVertexAndIndexSpace(), and that's
it.

https://codereview.chromium.org/663423003/diff/20001/src/gpu/GrDistanceFieldT...
File src/gpu/GrDistanceFieldTextContext.cpp (right):

https://codereview.chromium.org/663423003/diff/20001/src/gpu/GrDistanceFieldT...
src/gpu/GrDistanceFieldTextContext.cpp:387:
drawTarget->drawState()->setVertexAttribs<gTextVertexWithColorAttribs>(
On 2014/10/22 13:22:55, bsalomon wrote:
> line wraps in here

Done.

https://codereview.chromium.org/663423003/diff/20001/src/gpu/GrTextContext.h
File src/gpu/GrTextContext.h (right):

https://codereview.chromium.org/663423003/diff/20001/src/gpu/GrTextContext.h#...
src/gpu/GrTextContext.h:60: static int MeasureText(SkGlyphCache* cache,
SkDrawCacheProc glyphCacheProc,
On 2014/10/22 13:22:55, bsalomon wrote:
> comment on return value?

Done.

[bsalomon, 2014-10-23 18:44:01]

lgtm

[jvanverth1, 2014-10-23 18:48:36]

The CQ bit was checked by jvanverth@google.com

[commit-bot: I haz the power, 2014-10-23 18:48:57]

CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/663423003/40001

[commit-bot: I haz the power, 2014-10-23 18:57:16]

Committed patchset #3 (id:40001) as 73f105345066b1f22f15ba6575cb3800dd9313b8