Constructing the gm tests and benches causes many calls to font loads.
This is visible as profiling samples in fontconfig and freetype on Linux
for all profiling runs of nanobench. This complicates analysis of
test-cases that are suspected of being slow due to font-related issues.
Move the font loading to GM::onOnceBeforeDraw and Benchmark::onPreDraw.
This way the code is not executed if the testcase does not match the
nanobench --match filter. This way the samples in font-related code are
more easy to identify as legitimate occurances caused by the testcase.
This should not cause differences in timings, because:
* Benchmark::preDraw / onPreDraw is defined to be run outside the timer
* GM::runAsBench is not enabled for any of the modified testcases. Also
nanobench untimed warmup round should run the onOnceBeforeDraw.
(and there are other GM::runAsBench gms already doing loading in
onOnceBeforeDraw).
Changes the behavior:
In TextBench:
Before, the test would report two different gms with the same name if
the color emoji font was not loaded successfully.
After, the test always reports all tests as individual names.
Generally:
The errors from loading fonts now print inbetween each testcase, as
opposed to printing during construction phase. Sample output:
( 143/145 MB 1872) 14.7ms 8888 gm quadclosepathResource /fonts/Funkster.ttf not a valid font.
( 160/160 MB 1831) 575µs 8888 gm surfacenewResource /fonts/Funkster.ttf not a valid font.
( 163/165 MB 1816) 12.5ms 8888 gm linepathResource /fonts/Funkster.ttf not a valid font.
( 263/411 MB 1493) 118ms 8888 gm typefacestyles_kerningResource /fonts/Funkster.ttf not a valid font.
( 374/411 MB 1231) 7.16ms 565 gm getpostextpathResource /fonts/Funkster.ttf not a valid font.
( 323/411 MB 1179) 4.92ms 565 gm stringartResource /fonts/Funkster.ttf not a valid font.
( 347/493 MB 917) 191ms 565 gm patch_gridResource /fonts/Funkster.ttf not a valid font.
( 375/493 MB 857) 23.9ms gpu gm clipdrawdrawCannot render path (0)
( 393/493 MB 706) 2.91ms unit test ParsePath------ png error IEND: CRC error
( 394/493 MB 584) 166ms gpu gm hairmodesResource /fonts/Funkster.ttf not a valid font.
Resource /fonts/Funkster.ttf not a valid font.
Resource /fonts/Funkster.ttf not a valid font.
...
Review URL: https://codereview.chromium.org/1144023002
Make GrResourceCache performance less sensitive to key length change.
The memcmp in GrResourceKey is called when SkTDynamicHash jumps the
slots to find the hash by a index. Avoid most of the memcmps by
comparing the hash first.
This is important because small changes in key data length can cause
big performance regressions. The theory is that key length change causes
different hash values. These hash values might trigger memcmps that
originally weren't there, causing the regression.
Adds few specialized benches to grresourcecache_add to test different
key lengths. The tests are run only on release, because on debug the
SkTDynamicHash validation takes too long, and adding many such delays
to development test runs would be unproductive. On release the tests
are quite fast.
Effect of this patch to the added tests on amd64:
grresourcecache_find_10 738us -> 768us 1.04x
grresourcecache_find_2 472us -> 476us 1.01x
grresourcecache_find_25 841us -> 845us 1x
grresourcecache_find_4 565us -> 531us 0.94x
grresourcecache_find_54 1.18ms -> 1.1ms 0.93x
grresourcecache_find_5 834us -> 749us 0.9x
grresourcecache_find_3 620us -> 542us 0.87x
grresourcecache_add_25 2.74ms -> 2.24ms 0.82x
grresourcecache_add_56 3.23ms -> 2.56ms 0.79x
grresourcecache_add_54 3.34ms -> 2.62ms 0.78x
grresourcecache_add_5 2.68ms -> 2.1ms 0.78x
grresourcecache_add_10 2.7ms -> 2.11ms 0.78x
grresourcecache_add_2 1.85ms -> 1.41ms 0.76x
grresourcecache_add 1.84ms -> 1.4ms 0.76x
grresourcecache_add_4 1.99ms -> 1.49ms 0.75x
grresourcecache_add_3 2.11ms -> 1.55ms 0.73x
grresourcecache_add_55 39ms -> 13.9ms 0.36x
grresourcecache_find_55 23.2ms -> 6.21ms 0.27x
On arm64 the results are similar.
On arm_v7_neon, the results lack the discontinuity at 55:
grresourcecache_add 4.06ms -> 4.26ms 1.05x
grresourcecache_add_2 4.05ms -> 4.23ms 1.05x
grresourcecache_find 1.28ms -> 1.3ms 1.02x
grresourcecache_find_56 3.35ms -> 3.32ms 0.99x
grresourcecache_find_2 1.31ms -> 1.29ms 0.99x
grresourcecache_find_54 3.28ms -> 3.24ms 0.99x
grresourcecache_add_5 6.38ms -> 6.26ms 0.98x
grresourcecache_add_55 8.44ms -> 8.24ms 0.98x
grresourcecache_add_25 7.03ms -> 6.86ms 0.98x
grresourcecache_find_25 2.7ms -> 2.59ms 0.96x
grresourcecache_find_4 1.45ms -> 1.38ms 0.95x
grresourcecache_find_10 2.52ms -> 2.39ms 0.95x
grresourcecache_find_55 3.54ms -> 3.33ms 0.94x
grresourcecache_find_5 2.5ms -> 2.32ms 0.93x
grresourcecache_find_3 1.57ms -> 1.43ms 0.91x
The extremely slow case, 55, is postulated to be due to the index jump
collisions running the memcmp. This is not visible on arm_v7_neon probably due
to hash function producing different results for 32 bit architectures.
This change is needed for extending path cache key in Gr
NV_path_rendering codepath. Extending is needed in order to add dashed
paths to the path cache.
Review URL: https://codereview.chromium.org/1132723003
I'm thinking of using this in perf with something like:
ratio(fill(filter("test=foo")), fill(filter("test=control")))
Does that make sense to you?
Not sure that this is really a good control bench on all bots,
but I propose we just run it a bit and find out if it needs work.
BUG=skia:
Review URL: https://codereview.chromium.org/1129823003
The benches for N <= 10 get around 2x faster on my N7 and N9. I believe this
is because of the reduced function-call-then-function-pointer-call overhead on
the N7, and additionally because it seems autovectorization beats our NEON code
for small N on the N9.
My desktop is unchanged, though that's probably because N=10 lies well within a
region where memset's performance is essentially constant: N=100 takes only
about 2x as long as N=1 and N=10, which perform nearly identically.
BUG=skia:
Review URL: https://codereview.chromium.org/1073863002
The colorfilter is applied to a single (paint's) color, so the bench does not
measure the filter at all, but simply the blit of a color.
BUG=skia:
TBR=
Review URL: https://codereview.chromium.org/1055383002
Now that all SkCodecs can rewind (assuming the stream is rewindable),
we do not need to special case it.
Pointed out by Derek in the code review that added this.
TBR=djsollen
Review URL: https://codereview.chromium.org/1058633002
CodecBench:
Add new class for timing using SkCodec.
DecodingBench:
Include creating a decoder inside the loop. This is to have a better
comparison against SkCodec. SkCodec's factory function does not
necessarily read the same amount as SkImageDecoder's, so in order to
have a meaningful comparison, read the entire stream from the
beginning. Also for comparison, create a new SkStream from the
SkData each time.
Add a debugging check to make sure we have an SkImageDecoder.
Add include guards.
nanobench.cpp:
Decode using SkCodec.
When decoding using SkImageDecoder, exclude benches where we decoded
to a different color type than requested. SkImageDecoder may decide to
decode to a different type, in which case the name is misleading.
TODOs:
Now that we ignore color types that do not match the desired
color type, we should add Index8. This also means calling the more
complex version of getPixels so CodecBench can support kIndex8.
BUG=skia:3257
Review URL: https://codereview.chromium.org/1044363002
The primary feature this delivers is SkNf and SkNd for arbitrary power-of-two N. Non-specialized types or types larger than 128 bits should now Just Work (and we can drop in a specialization to make them faster). Sk4s is now just a typedef for SkNf<4, SkScalar>; Sk4d is SkNf<4, double>, Sk2f SkNf<2, float>, etc.
This also makes implementing new specializations easier and more encapsulated. We're now using template specialization, which means the specialized versions don't have to leak out so much from SkNx_sse.h and SkNx_neon.h.
This design leaves us room to grow up, e.g to SkNf<8, SkScalar> == Sk8s, and to grown down too, to things like SkNi<8, uint16_t> == Sk8h.
To simplify things, I've stripped away most APIs (swizzles, casts, reinterpret_casts) that no one's using yet. I will happily add them back if they seem useful.
You shouldn't feel bad about using any of the typedef Sk4s, Sk4f, Sk4d, Sk2s, Sk2f, Sk2d, Sk4i, etc. Here's how you should feel:
- Sk4f, Sk4s, Sk2d: feel awesome
- Sk2f, Sk2s, Sk4d: feel pretty good
No public API changes.
TBR=reed@google.com
BUG=skia:3592
Review URL: https://codereview.chromium.org/1048593002
Need to land SK_SUPPORT_LEGACY_SCALAR_MAPPOINTS in chrome to suppress Affine
version which causes slight differences (which will need to be rebaselined)
BUG=skia:
Review URL: https://codereview.chromium.org/1045493002
Duplicate code from the HWUI backends for DM and nanobench
moves into a single place, saving a hundred lines or more of
cut-and-paste.
There's some indication that this increases the incidence of
SkCanvas "Unable to find device for layer." warnings, but no
clear degradation in test results.
R=djsollen@google.com,mtklein@google.com
BUG=skia:3589
Review URL: https://codereview.chromium.org/1036303002
Add and test trunc(), which is what get() used to be before rounding.
Using trunc() is a ~40% speedup on our linear gradient bench.
#neon #floats
BUG=skia:3592
#n5
#n9
CQ_INCLUDE_TRYBOTS=client.skia.android:Test-Android-Nexus5-Adreno330-Arm7-Debug-Trybot;client.skia.android:Test-Android-Nexus9-TegraK1-Arm64-Release-Trybot
Review URL: https://codereview.chromium.org/1032243002
Am I going nuts or can we get this down to just adds and converts in the loop?
#floats #n9
BUG=skia:3592
CQ_INCLUDE_TRYBOTS=client.skia.android:Test-Android-Nexus9-TegraK1-Arm64-Release-Trybot
Review URL: https://codereview.chromium.org/1008973004
There is no reason to require the 4 SkPMFloats (registers) to be adjacent.
The only potential win in loads and stores comes from the SkPMColors being adjacent.
Makes no difference to existing bench.
BUG=skia:
Review URL: https://codereview.chromium.org/1035583002
RotatedRectBench was asking for its base layer size, which may
not be what it expects with odd canvas modes (particularly proxies).
Most benchmarks are not so sophisticated; they hard-wire their
size and just use that (expected) value.
R=mtklein@google.com,djsollen@google.com
BUG=skia:3566
Review URL: https://codereview.chromium.org/1015013004
Initial experiments did show that the 256 tile size fixed the hd2000 win7
nanobot failures. However it did not have any effect on other bots, so this
change is to move back to the larger tile size on all bots expect for the
hd2000.
BUG=skia:
Review URL: https://codereview.chromium.org/1022083002
