| 2adb04f8edb61857a2c53c16ced533df47be67c3 |
|
23-Mar-2018 |
Seigo Nonaka <nona@google.com> |
Discard hyphenation context
During line breaking, the hyphenated pieces are computed with context.
On the other hand, during drawing, the context is not available.
This ends up with cache miss during drawing.
To increase the cache hit and making line break correct, remove context
during line breaking.
Bug: 65024629
Test: atest CtsWidgetTestCases:EditTextTest
CtsWidgetTestCases:TextViewFadingEdgeTest
FrameworksCoreTests:TextViewFallbackLineSpacingTest
FrameworksCoreTests:TextViewTest FrameworksCoreTests:TypefaceTest
CtsGraphicsTestCases:TypefaceTest CtsWidgetTestCases:TextViewTest
CtsTextTestCases FrameworksCoreTests:android.text
CtsWidgetTestCases:TextViewPrecomputedTextTest
Test: minikin_test
Change-Id: I27efa55825aa5060d742020661f831c69697aaf5
/frameworks/minikin/libs/minikin/LineBreakerUtil.h
|
| 43a52c5ee653e75ce3ccf2477a9dcdec6f8d7fa3 |
|
14-Mar-2018 |
Seigo Nonaka <nona@google.com> |
Precompute full hyphenation layout
The previous offset based computation doesn't work well for some cases,
e.g. the final text for drawing may be a substring of the original
string. The offset based matching doesn't work well for such case.
This CL changes
- Store the layout results during measuring hyphenation pieces.
- Change the LayoutPiece key from offset to string.
- This CL increases the heap usage about 23kb for 500 characters text.
This changes the performance behaviors.
android.text.PrecomputedTextMemoryUsageTest:
MemoryUsage_Hyphenation : 28,448 -> 51,836: (+82.2%)
MemoryUsage_Hyphenation_WidthOnly : 8,856 -> 8,856: (+0.0%)
MemoryUsage_NoHyphenation : 27,592 -> 26,386: (-4.4%)
MemoryUsage_NoHyphenation_WidthOnly : 8,000 -> 8,000: (+0.0%)
android.text.PrecomputedTextPerfTest:
create_NoStyled_Hyphenation : 17,162,475 -> 17,695,377: (+3.1%)
create_NoStyled_Hyphenation_WidthOnly : 17,044,784 -> 17,677,423: (+3.7%)
create_NoStyled_NoHyphenation : 7,108,911 -> 7,021,486: (-1.2%)
create_NoStyled_NoHyphenation_WidthOnly : 7,102,481 -> 7,045,453: (-0.8%)
create_Styled_Hyphenation : 12,394,607 -> 12,090,933: (-2.5%)
create_Styled_Hyphenation_WidthOnly : 12,297,095 -> 12,105,491: (-1.6%)
create_Styled_NoHyphenation : 11,886,364 -> 11,835,249: (-0.4%)
create_Styled_NoHyphenation_WidthOnly : 12,018,548 -> 11,871,765: (-1.2%)
android.text.StaticLayoutMultithreadPerfTest:
create_RandomText_Thread_1 : 7,051,921 -> 6,998,143: (-0.8%)
create_RandomText_Thread_2 : 7,112,819 -> 7,032,146: (-1.1%)
create_RandomText_Thread_4 : 7,843,295 -> 7,860,874: (+0.2%)
android.text.StaticLayoutPerfTest:
create_PrecomputedText_Balanced_Hyphenation : 703,563 -> 709,839: (+0.9%)
create_PrecomputedText_Balanced_NoHyphenation: 523,437 -> 527,671: (+0.8%)
create_PrecomputedText_Greedy_Hyphenation : 470,881 -> 477,259: (+1.4%)
create_PrecomputedText_Greedy_NoHyphenation : 471,487 -> 479,772: (+1.8%)
create_RandomText_Balanced_Hyphenation : 17,166,857 -> 17,123,681: (-0.3%)
create_RandomText_Balanced_NoHyphenation : 7,107,289 -> 7,040,572: (-0.9%)
create_RandomText_Greedy_Hyphenation : 7,055,902 -> 7,000,681: (-0.8%)
create_RandomText_Greedy_NoHyphenation : 7,032,325 -> 6,997,115: (-0.5%)
draw_PrecomputedText_NoStyled : 527,505 -> 543,623: (+3.1%)
draw_PrecomputedText_NoStyled_WithoutCache : 526,084 -> 564,742: (+7.3%)
draw_PrecomputedText_Styled : 880,951 -> 838,581: (-4.8%)
draw_PrecomputedText_Styled_WithoutCache : 874,259 -> 826,775: (-5.4%)
draw_RandomText_NoStyled : 549,637 -> 538,162: (-2.1%)
draw_RandomText_NoStyled_WithoutCache : 6,449,481 -> 6,401,486: (-0.7%)
draw_RandomText_Styled : 1,001,350 -> 1,024,683: (+2.3%)
draw_RandomText_Styled_WithoutCache : 2,759,161 -> 2,733,204: (-0.9%)
android.widget.TextViewPrecomputedTextPerfTest:
newLayout_PrecomputedText : 738,105 -> 736,130: (-0.3%)
newLayout_PrecomputedText_Selectable : 17,410,426 -> 17,379,765: (-0.2%)
newLayout_RandomText : 16,565,334 -> 16,495,200: (-0.4%)
newLayout_RandomText_Selectable : 17,607,671 -> 17,482,439: (-0.7%)
onDraw_PrecomputedText : 2,371,858 -> 1,274,921: (-46.2%)
onDraw_PrecomputedText_Selectable : 17,493,221 -> 17,367,238: (-0.7%)
onDraw_RandomText : 17,349,102 -> 17,224,949: (-0.7%)
onDraw_RandomText_Selectable : 18,107,851 -> 18,067,397: (-0.2%)
onMeasure_PrecomputedText : 756,229 -> 752,875: (-0.4%)
onMeasure_PrecomputedText_Selectable : 17,725,005 -> 17,647,842: (-0.4%)
onMeasure_RandomText : 16,636,892 -> 16,435,649: (-1.2%)
onMeasure_RandomText_Selectable : 17,866,544 -> 17,724,819: (-0.8%)
setText_PrecomputedText : 90,499 -> 92,894: (+2.6%)
setText_PrecomputedText_Selectable : 146,906 -> 145,134: (-1.2%)
setText_RandomText : 11,251 -> 11,130: (-1.1%)
setText_RandomText_Selectable : 48,745 -> 48,900: (+0.3%)
Bug: 72998298
Test: atest CtsWidgetTestCases:EditTextTest
CtsWidgetTestCases:TextViewFadingEdgeTest
FrameworksCoreTests:TextViewFallbackLineSpacingTest
FrameworksCoreTests:TextViewTest FrameworksCoreTests:TypefaceTest
CtsGraphicsTestCases:TypefaceTest CtsWidgetTestCases:TextViewTest
CtsTextTestCases FrameworksCoreTests:android.text
CtsWidgetTestCases:TextViewPrecomputedTextTest
Test: minikin_tests
Change-Id: I2290d7e06d9e2f4bc13f60246ce66ba1c1785cdd
/frameworks/minikin/libs/minikin/LineBreakerUtil.h
|
| 8e60b6f3a6bed9ece5608847f7dab7c3359c9a67 |
|
24-Feb-2018 |
Seigo Nonaka <nona@google.com> |
Stop including CR, LF characters in the previous run
Bug: 73625594
Test: bit FrameworksCoreTests:android.text.
Test: atest CtsWidgetTestCases:EditTextTest
CtsWidgetTestCases:TextViewFadingEdgeTest
FrameworksCoreTests:TextViewFallbackLineSpacingTest
FrameworksCoreTests:TextViewTest FrameworksCoreTests:TypefaceTest
CtsGraphicsTestCases:TypefaceTest CtsWidgetTestCases:TextViewTest
CtsTextTestCases
Change-Id: I809296e7edd9fe54d5defd83c7ddc5bb40a1d0e4
/frameworks/minikin/libs/minikin/LineBreakerUtil.h
|
| 7d8c4ed5b7683e16e6205a6b52b1910f8ff6b51b |
|
09-Feb-2018 |
Seigo Nonaka <nona@google.com> |
Remove global mutex lock from minikin
All modules are now free from gMinikinLock.
This increases the performance of parallel text layout in background thread.
Here is a raw performance score on walleye-userdebug
StaticLayout creation time in parallel:
RandomText Thread 1 : 7,348,118 -> 7,327,542: (-0.3%)
RandomText Thread 2 : 13,872,214 -> 7,402,398: (-46.6%)
RandomText Thread 4 : 24,999,986 -> 8,185,263: (-67.3%)
StaticLayout creation time:
MeasuredText Balanced Hyphenation : 774,245 -> 781,534: (+0.9%)
MeasuredText Balanced NoHyphenation: 587,157 -> 574,882: (-2.1%)
MeasuredText Greedy Hyphenation : 535,843 -> 523,630: (-2.3%)
MeasuredText Greedy NoHyphenation : 534,965 -> 522,259: (-2.4%)
RandomText Balanced Hyphenation : 17,886,651 -> 17,814,108: (-0.4%)
RandomText Balanced NoHyphenation : 7,385,288 -> 7,410,588: (+0.3%)
RandomText Greedy Hyphenation : 7,380,350 -> 7,296,407: (-1.1%)
RandomText Greedy NoHyphenation : 7,358,680 -> 7,320,040: (-0.5%)
MeasuredText creation time
NoStyled Hyphenation : 17,587,451 -> 17,544,350: (-0.2%)
NoStyled Hyphenation WidthOnly : 17,112,258 -> 17,092,507: (-0.1%)
NoStyled NoHyphenation : 7,307,064 -> 7,296,519: (-0.1%)
NoStyled NoHyphenation WidthOnly : 6,880,721 -> 6,848,543: (-0.5%)
Styled Hyphenation : 14,753,919 -> 14,659,049: (-0.6%)
Styled Hyphenation WidthOnly : 13,767,908 -> 13,755,623: (-0.1%)
Styled NoHyphenation : 14,368,118 -> 14,337,125: (-0.2%)
Styled NoHyphenation WidthOnly : 13,414,942 -> 13,375,858: (-0.3%)
StaticLayout draw time:
MeasuredText NoStyled : 676,438 -> 678,795: (+0.3%)
MeasuredText NoStyled WithoutCache : 663,041 -> 654,687: (-1.3%)
MeasuredText Styled : 885,574 -> 890,281: (+0.5%)
MeasuredText Styled WithoutCache : 944,259 -> 928,675: (-1.7%)
RandomText NoStyled : 562,361 -> 553,651: (-1.5%)
RandomText NoStyled WithoutCache : 6,678,968 -> 6,696,637: (+0.3%)
RandomText Styled : 2,970,369 -> 2,933,900: (-1.2%)
RandomText Styled WithoutCache : 3,408,554 -> 3,387,260: (-0.6%)
Bug: 37567215
Test: minikin_tests
Test: minikin_stress_tests
Test: atest CtsWidgetTestCases:EditTextTest
CtsWidgetTestCases:TextViewFadingEdgeTest
FrameworksCoreTests:TextViewFallbackLineSpacingTest
FrameworksCoreTests:TextViewTest FrameworksCoreTests:TypefaceTest
CtsGraphicsTestCases:TypefaceTest CtsWidgetTestCases:TextViewTest
CtsTextTestCases CtsGraphicsTestCases:PaintTest
Test: bit FrameworksCoreTests:android.text.
Change-Id: I6aab090be6bfa65f1205b26b77c6c65823a3d94e
/frameworks/minikin/libs/minikin/LineBreakerUtil.h
|
| 5cc07b9fefb63956668a06926cc8a044f65f78dc |
|
20-Jan-2018 |
Seigo Nonaka <nona@google.com> |
Remove unused argument LayoutOverhang
The overhang stuff is not used now. Let's remove them.
Bug: 65024629
Test: minikin_tests
Change-Id: I5b19341824308effe4e7c0920f808f05c357df57
/frameworks/minikin/libs/minikin/LineBreakerUtil.h
|
| c9611626465f9e11817854f554e7b7d0c6cee905 |
|
12-Jan-2018 |
Seigo Nonaka <nona@google.com> |
Store hyphenated pieces in MeasuredText
Do hyphenation calculation at the same time of measuring character
widths. We can do the 93% of line break work beforehand.
Here is a raw performance score:
(walleye-userdebug, w/o patch -> w/ patch, N=30)
Measured Balanced Hyphenation : 11,428,287 -> 703,417 (-93.8%)
Measured Balanced NoHyphenation: 535,660 -> 536,210 (+0.1%)
Random Balanced Hyphenation : 18,903,251 -> 19,046,535 (+0.8%)
Random Balanced NoHyphenation : 7,973,160 -> 7,956,803 (-0.2%)
Measured Greedy Hyphenation : 486,633 -> 486,381 (-0.1%)
Measured Greedy NoHyphenation : 484,716 -> 490,018 (+1.1%)
Random Greedy Hyphenation : 7,966,823 -> 7,921,520 (-0.6%)
Random Greedy NoHyphenation : 7,918,480 -> 7,925,110 (+0.1%)
This CL also removes overhang member variable from MeasuredText since
this is no longer used in line breaker.
Bug: 67504091
Test: bit CtsTextTestCases:*
Test: bit CtsWidgetTestCases:android.widget.cts.TextViewTest
Test: minikin_tests
Change-Id: I94c65980bf1190c9351d50dc0c811a377fb93ee6
/frameworks/minikin/libs/minikin/LineBreakerUtil.h
|
| 366f57fb4efd0ddf5d48cd232cd88d3777517ea2 |
|
09-Jan-2018 |
Seigo Nonaka <nona@google.com> |
Reorganize optimal line breaker
This CL contains:
- Rename LineBreakImpl.{h|cpp} with OptimalLineBreaker.{h|cpp}
- Extract line break candidate enumeration code from LineBreakerImpl
and rename LineBreakImpl with LineBreakOptimizer.
- Fix unexpected hyphenation by ordering hyphenation break point and
desperate break points.
- Refactor line break candidate enumeration code.
Compatibility:
To check the compatibility from current implementation, manually
compared the line break result for the UDHR text (http://unicode.org/udhr/).
Then verified new implementation completely replicate the same output as the
current implementation except for desperate breaks.
Here is the combination verified:
Langauge: ar, zh-CN, zh-TW, nl, fr, de, it, ja, ko, po, pl-PL, pl-BR,
ru, es-ES, th, tr, bg, ca, hr, cs, da, el, he, hi, hu, id,
lv, lt, no, ro, sr-Cyrl, sr-Latn, sk, sl, sv, uk, vi
Font Size: 14px, 15px, 16px, 17px, 18px, 19px
Line Width: 128px, 256px, 512px, 1024px
Performance:
Verified this doesn't cause performance regression.
Here is raw performance score.
StaticLayoutPerfTest(walleye-userdebug, w/o patch -> w/ patch, N=50):
Random Text, No Hyphen : 7,782,872 -> 7,754,140 (-0.37%)
Random Text, Hyphen : 19,028,618 -> 18,942,728 (-0.45%)
measured Text, No Hyphen : 346,625 -> 349,809 (+0.92%)
measured Text, Hyphen : 11,411,894 -> 11,379,360 (-0.29%)
Bug: 71709070
Bug: 65024629
Test: minikin_tests
Test: bit CtsTextTestCases:*
Test: bit FrameworksCoreTests:android.text.StaticLayoutTest
Change-Id: I1933335764405da6d7c82cf1d1af1692e6d10a04
/frameworks/minikin/libs/minikin/LineBreakerUtil.h
|
| d57562ceb31842c7bcdb28c084c863631a844acd |
|
08-Dec-2017 |
Seigo Nonaka <nona@google.com> |
Implement new line breaker for greedy algorithm
The new greedy line breaker is 2x faster than current line breaker.
Current greedy line breaker does lots of unnecessary things.
- Hyphenation is not necessary until the word is too long for the line.
- Cost calculation and optimization for the cost is not necessary. Only
if the break point is inside email or url. This can be done in O(1).
The new line breaker is fully compatible with older line breaker except
for
- Overhang calculation is dropped since it is implemented partially but
not consistent with other line break strategy. Also it is not exposed
as Java API.
- The priority of desperate line breaking and hyphenation is not consistent
and current master implementation is broken. Sometimes the line
breaker exceeds the line width. The new line breaker has explicit
priority for the hyphenation and keep the line width under the limit.
(See Issue 70415793)
Compatibility:
To check the compatibility from current implementation, manually
compared the line break result for the UDHR text (http://unicode.org/udhr/).
Then verified new implementation completely replicate the same output as the
current implementaiton.
Here is the combination verified:
Langauge: ar, zh-CN, zh-TW, nl, fr, de, it, ja, ko, po, pl-PL, pl-BR,
ru, es-ES, th, tr, bg, ca, hr, cs, da, el, he, hi, hu, id,
lv, lt, no, ro, sr-Cyrl, sr-Latn, sk, sl, sv, uk, vi
Font Size: 14px, 15px, 16px, 17px, 18px, 19px
Line Width: 128px, 256px, 512px, 1024px
Performance:
Here is the performance score:
(All greedy algorithm, w/o patch -> w/ patch, median, N=100)
Random, No Style, Hyphenation OFF: 7,854,007 -> 7,780,766 (-0.93%)
Random, No Style, Hyphenation ON : 19,274,298 -> 7,762,346 (-60.0%)
Measured, No Style, Hyphenation OFF: 516,039 -> 424,173 (-18.0%)
Measured, No Style, Hyphenation ON : 11,772,800 -> 482,506 (-96.0%)
I'll remove old greedy line breaker once this is submitted.
Test: CtsTextTestCases:*
Test: CtsGraphicsTestCases:*
Test: CtsWidgetTestCases:*
Test: minikin_test
Bug: 70415793
Bug: 65024629
Bug: 64389125
Change-Id: Iaff9afed4bb6eebecb010e468546239f10852903
/frameworks/minikin/libs/minikin/LineBreakerUtil.h
|