summaryrefslogtreecommitdiffstats
path: root/benchmarks
Commit message (Collapse)AuthorAgeFilesLines
* Add BreakIterator benchmark to the libcore benchmark suite.Narayan Kamath2015-02-202-1/+73
| | | | | | | | | | Strings come from con.android.markbench.Text. This change also reworks the dependency list for benchmarks.jar to avoid hitting the dex limit (sigh). bug: 17185402 Change-Id: Iffd4f88b84649c86446571fea3868e98beae3175
* Use ICU for relative time formattingTao Bao2015-02-111-0/+68
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Rewrite the DateUtils' relative time formatting APIs (getRelativeTimeSpanString, getRelativeDateTimeString) to use ICU ones. Two APIs that take withPreposition parameter are not changed. Because (a) ICU doesn't provide functionality to format preposition; (b) They are not really computing relative time but instead calling formatDateRange() to get the absolute time/date string. Benchmark results on aosp_hammerhead-userdebug: before: benchmark us linear runtime DateUtils_getRelativeDateTimeString 127.1 ========================== DateUtils_getRelativeDateTimeString_ABBREV 145.0 ============================== DateUtils_getRelativeTimeSpanString 28.0 ===== DateUtils_getRelativeTimeSpanString_ABBREV 27.9 ===== now: benchmark us linear runtime RelativeDateTimeFormatter_getRelativeDateTimeString 119.2 ========================== RelativeDateTimeFormatter_getRelativeDateTimeString_ABBREV 133.8 ============================== RelativeDateTimeFormatter_getRelativeTimeSpanString 24.6 ===== RelativeDateTimeFormatter_getRelativeTimeSpanString_ABBREV 24.7 ===== Bug: 19146457 Bug: 5252772 Change-Id: Ief74608354964a17e42191d7b1a58964f3a9acfd
* Extend ArrayList fastpath to Collections.sort(List<T>, Comparator<T>).Narayan Kamath2015-01-021-7/+32
| | | | | | | | | | | | | | | | | | | | | | | | | | | My original reasons for not extending it were bogus. We can safely cast Object[] to T[] in this case. Also tidy up the benchmark by changing buildList(.. boolean isArrayList) to buildList(.. Class<T extends List<Integer> list. AFTER SortWithComparator_arrayList, arrayListLength=4} 10881.97 ns; σ=64.97 ns @ 3 trials SortWithComparator_arrayList, arrayListLength=16} 33235.73 ns; σ=127.45 ns @ 3 trials SortWithComparator_arrayList, arrayListLength=64} 130317.52 ns; σ=135.31 ns @ 3 trials SortWithComparator_arrayList, arrayListLength=256} 495200.19 ns; σ=1120.86 ns @ 3 trials SortWithComparator_arrayList, arrayListLength=1024} 1990796.02 ns; σ=19151.58 ns @ 6 trials BEFORE SortWithComparator_arrayList, arrayListLength=4} 40030.32 ns; σ=509.60 ns @ 10 trials SortWithComparator_arrayList, arrayListLength=16} 122122.08 ns; σ=182.58 ns @ 3 trials SortWithComparator_arrayList, arrayListLength=64} 457833.87 ns; σ=1311.54 ns @ 3 trials SortWithComparator_arrayList, arrayListLength=256} 1764647.22 ns; σ=35150.74 ns @ 10 trials SortWithComparator_arrayList, arrayListLength=1024} 7127679.90 ns; σ=122546.75 ns @ 10 trials bug: 18821961 Change-Id: I22a18d99598dfacbdc54cf75e38dbc32ab027754
* Add a zero-copy path for Collections.sort() on ArrayList.Narayan Kamath2014-12-301-0/+55
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Given that ArrayLists are array backed, we can sort the underlying array in place. A similar optimisation can be applied to Vector and CopyOnWriteArrayList but vector isn't used often enough to slow everything else down with an instanceof check and nobody should be daft enough to try sorting a CopyOnWriteArrayList. Note that it's possible to apply the same optimisation to Collections.sort(List<T>, Comparator<? super T>) but it involves a larger refactoring to allow us to use Object[] instead of T[] and is accompanied by a general loss of readability. ArrayList performance improves for all array sizes : BEFORE Scenario{vm=app_process, trial=0, benchmark=Sort_arrayList, arrayListLength=4} 39021.26 ns; σ=206.27 ns @ 3 trials Scenario{vm=app_process, trial=0, benchmark=Sort_arrayList, arrayListLength=16} 120903.59 ns; σ=1188.88 ns @ 5 trials Scenario{vm=app_process, trial=0, benchmark=Sort_arrayList, arrayListLength=64} 435484.98 ns; σ=19141.88 ns @ 10 trials Scenario{vm=app_process, trial=0, benchmark=Sort_arrayList, arrayListLength=256} 1772876.81 ns; σ=5542.74 ns @ 3 trials Scenario{vm=app_process, trial=0, benchmark=Sort_arrayList, arrayListLength=1024} 7289431.83 ns; σ=110869.13 ns @ 10 trials AFTER Scenario{vm=app_process, trial=0, benchmark=Sort_arrayList, arrayListLength=4} 9161.79 ns; σ=90.04 ns @ 6 trials Scenario{vm=app_process, trial=0, benchmark=Sort_arrayList, arrayListLength=16} 29988.90 ns; σ=74.05 ns @ 3 trials Scenario{vm=app_process, trial=0, benchmark=Sort_arrayList, arrayListLength=64} 118142.86 ns; σ=1133.52 ns @ 7 trials Scenario{vm=app_process, trial=0, benchmark=Sort_arrayList, arrayListLength=256} 456091.11 ns; σ=9110.18 ns @ 10 trials Scenario{vm=app_process, trial=0, benchmark=Sort_arrayList, arrayListLength=1024} 1851939.39 ns; σ=69542.41 ns @ 10 trials For all other lists, there's no measurable difference for most sizes. length == 4 is harder to be sure about, given the large standard deviations and we'd expect the time spent in instanceof to be larger relative to the rest of what's going on. BEFORE Scenario{vm=app_process, trial=0, benchmark=Sort_vector, arrayListLength=4} 49002.02 ns; σ=1392.58 ns @ 10 trials Scenario{vm=app_process, trial=0, benchmark=Sort_vector, arrayListLength=16} 154967.88 ns; σ=332.29 ns @ 3 trials Scenario{vm=app_process, trial=0, benchmark=Sort_vector, arrayListLength=64} 588207.14 ns; σ=1863.60 ns @ 3 trials Scenario{vm=app_process, trial=0, benchmark=Sort_vector, arrayListLength=256} 2292967.79 ns; σ=4578.14 ns @ 3 trials Scenario{vm=app_process, trial=0, benchmark=Sort_vector, arrayListLength=1024} 9299582.60 ns; σ=179459.24 ns @ 10 trials AFTER Scenario{vm=app_process, trial=0, benchmark=Sort_vector, arrayListLength=4} 49324.98 ns; σ=240.86 ns @ 3 trials Scenario{vm=app_process, trial=0, benchmark=Sort_vector, arrayListLength=16} 155639.91 ns; σ=215.12 ns @ 3 trials Scenario{vm=app_process, trial=0, benchmark=Sort_vector, arrayListLength=64} 581950.81 ns; σ=938.11 ns @ 3 trials Scenario{vm=app_process, trial=0, benchmark=Sort_vector, arrayListLength=256} 2285933.64 ns; σ=13265.42 ns @ 3 trials Scenario{vm=app_process, trial=0, benchmark=Sort_vector, arrayListLength=1024} 8981475.84 ns; σ=431676.48 ns @ 10 trials bug: 18821961 Change-Id: I30a3eaf89ff478663f555c93556167107da10873
* Add benchmark for reflection performance on large classes.Narayan Kamath2014-11-142-0/+2560
| | | | | | | | | | | The R.java file in this change is adapted from an autogenerated frameworks class with android.* references stripped. bug: 18211592 (cherry picked from commit f9362eeb6a8ec6411729a688d8050efffe8baa9d) Change-Id: Ie627dd15e776425a1a7c0109bcf833872db46198
* Removing differences between AOSP and internalNeil Fuller2014-09-041-20/+25
| | | | | | | | | | | These commits were equivalent but not exactly the same: AOSP commit: adb19b9b603ba83ba49b1a10b6301e396132bf3b Internal commit: 18e74061b760389a94439f4b14acb376809a44ce This commit makes the internal files the same as AOSP to make future diffs easier. Change-Id: Ie2dcdd3940244c6dd18334b58037d2e8cca93920
* Dalvik is dead, long live Dalvik!Brian Carlstrom2014-08-051-1/+1
| | | | | | | | | | | | | | | | croot cd libcore repo start dalvik-is-dead-long-live-dalvik . repo sync -c . git rm -r libdvm git add JavaLibrary.mk (after removing libdvm references, adding explict core-libart references) git add Docs.mk (after replacing references to libdvm with libart) git add benchmarks/Android.mk (after adding explict core-libart references) git add Android.mk (after removing dalvik-host target) git commit -a -m 'Dalvik is dead, long live Dalvik!' Bug: 14298175 Change-Id: I977e7b4be8ef1043b70ceb8532ac4947ea1b7ab8
* Avoid Math.min() benchmark being optimized away.Vladimir Marko2014-07-221-1/+1
| | | | | | | | | | | This is a follow-up to https://android-review.googlesource.com/101914 Bug: 11579369 (cherry picked from 3c32023358065ddbad256a9fe7c1385592746d3e) Change-Id: I6193489a8ca9dca6ccfa8c5ffd814baab49e0f63
* Avoid Math/StrictMath benchmarks being optimized away.Vladimir Marko2014-07-221-111/+228
| | | | | | | | | | | | | ART is completely optimizing away abs() for float and double on x86 if the result is unused. We may soon optimize it away on ARM as well. To avoid the benchmark becoming irrelevant modify all the math benchmarks to use the return value. Bug: 11579369 (cherry-picked from 547edf38495b20fa3669520cd3989243d75787ee) Change-Id: I1c2eec3bfcef9e7b5d56522ba8330603ad774eec
* reconcile aosp (86eae8dd1e7405394f0e9efe1252575ef9a9674c) after branching. ↵Ed Heyl2014-07-141-0/+29
| | | | | | Please do not merge. Change-Id: Iae444b6b2840916ef9a4d0962a55a74f650bd5a5
* reconcile aosp (0d589d78c1c2a809429911fc4756856b53404fb9) after branching. ↵Ed Heyl2014-07-141-0/+60
| | | | | | Please do not merge. Change-Id: Ib9fbb179f981f5159c202db15235c2129b81cb5e
* am 42a6d80d: Merge "Expanded arraycopy function to deal with more types and ↵Jeff Hao2014-06-261-8/+68
|\ | | | | | | | | | | | | updated benchmarking." * commit '42a6d80d91d8f33162b8946e6fc05e8b43833ca7': Expanded arraycopy function to deal with more types and updated benchmarking.
| * Merge "Expanded arraycopy function to deal with more types and updated ↵Jeff Hao2014-06-261-8/+68
| |\ | | | | | | | | | benchmarking."
| | * Expanded arraycopy function to deal with more types and updated benchmarking.Emma Meersman2014-06-231-8/+68
| |/ |/| | | | | | | | | | | | | Runs 48% faster on average for short arrays of length 16-128, depending on the type. Defaults to native code for longer arrays. Bug: 7103825 Change-Id: I70ca8f4a379600917e5bc103364af0637f18e03a
| * Dalvik is dead, long live Dalvik! DO NOT MERGEBrian Carlstrom2014-06-181-20/+25
|/ | | | | | | | | | | | | | | | croot cd libcore repo start dalvik-is-dead-long-live-dalvik . repo sync -c . git rm -r libdvm git add JavaLibrary.mk (after removing libdvm references, adding explict core-libart references) git add Docs.mk (after replacing references to libdvm with libart) git add benchmarks/Android.mk (after adding explict core-libart references) git add Android.mk (after removing dalvik-host target) git commit -a -m 'Dalvik is dead, long live Dalvik! DO NOT MERGE' Bug: 14298175 Change-Id: I0e7a3ce47e0b6e3a7b15006a6d17d103d23e8064
* Pass BCP-47 tags and not Locale.toString results to ICU.Elliott Hughes2014-06-162-3/+3
| | | | | | | | | | | | | | | | | | ICU can't handle the new toString forms for scripts etc. and it's also guaranteed to deal with BCP-47 tags correctly. Most of the changes in this patch are required to keep backwards compatibility for getDisplayCountry string etc. in the face of the transformations toLanguageTag performs. A few tests were changed, but for the better. The tagalog -> filipino charlie foxtrot will be dealt with in a follow up change. Co-Authored-By: Narayan Kamath <narayan@google.com> Change-Id: Ia7f26d92a0e38c4bbb1d839c0fbd8ad16a473bf5
* Fix Greek uppercasing.Elliott Hughes2014-04-041-0/+8
| | | | | | | This matches Chrome's behavior. Bug: 13246961 Change-Id: I1954da3116a1759cb650bffed558dced5770e3ec
* Don't build benchmarks if LIBCORE_SKIP_TESTS=true.Narayan Kamath2014-01-241-0/+2
| | | | | | | benchmarks depend on core-tests, so won't build if this flag is set. Change-Id: Ia619bb9d4c30249307c0657e2d5a37ba7f259a58
* CipherInputStream: increase buffers for speedKenny Root2014-01-151-0/+85
| | | | | | | This increases speed of CipherInputStreamBenchmark from 606ms to 28.5ms. Bug: 12246774 Change-Id: If61c0a055fe1b8e87909711b15d0441fcdbe46c7
* Remove HarmonyJSSE SSLContext, SSLSocket and SSLServerSocket.Alex Klyubin2013-12-162-30/+2
| | | | Change-Id: I3c939e9275ba8f1d00342d1f83c6fdaf110f2317
* am eaa0add6: am 48132b98: Merge "Minor cleanup of deepHashCode / deepEquals."Narayan Kamath2013-11-211-0/+148
|\ | | | | | | | | * commit 'eaa0add6aaa503221a530d346a3b4a066bc46449': Minor cleanup of deepHashCode / deepEquals.
| * Minor cleanup of deepHashCode / deepEquals.Narayan Kamath2013-11-211-0/+148
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | These methods used equals() comparisons where identity checks would've sufficient. As expected, this results in a small but measurable performance increase on both art & dalvik. ---------------------- After -------------- ART: arrayLength benchmark us linear runtime 1 DeepHashCode 8.08 = 1 Equals 17.68 = 4 DeepHashCode 8.63 = 4 Equals 18.20 = 16 DeepHashCode 19.59 = 16 Equals 41.19 = 256 DeepHashCode 255.41 = 256 Equals 483.61 == 2048 DeepHashCode 2321.91 ============== 2048 Equals 4932.40 ============================== Dalvik: arrayLength benchmark us linear runtime 1 DeepHashCode 32.7 = 1 Equals 65.3 = 4 DeepHashCode 34.4 = 4 Equals 67.7 = 16 DeepHashCode 76.0 = 16 Equals 142.8 = 256 DeepHashCode 755.6 = 256 Equals 1515.6 === 2048 DeepHashCode 6915.4 =============== 2048 Equals 13786.6 ============================== ------------------------ Before -------------------- Art: arrayLength benchmark us linear runtime 1 DeepHashCode 8.13 = 1 Equals 17.86 = 4 DeepHashCode 8.63 = 4 Equals 18.92 = 16 DeepHashCode 19.59 = 16 Equals 40.89 = 256 DeepHashCode 253.88 = 256 Equals 477.02 == 2048 DeepHashCode 2795.26 ================ 2048 Equals 4962.11 ============================== Dalvik: arrayLength benchmark us linear runtime 1 DeepHashCode 32.8 = 1 Equals 66.7 = 4 DeepHashCode 34.4 = 4 Equals 70.1 = 16 DeepHashCode 74.1 = 16 Equals 147.9 = 256 DeepHashCode 805.0 = 256 Equals 1573.2 === 2048 DeepHashCode 7253.4 =============== 2048 Equals 14349.8 ============================== bug: 9320068 Change-Id: Id73bf38dd0777b2ec9004a2941523d082bbc74c9
* | am 7e528a4a: am bd03ef07: Merge "Improve URI.equals performance."Narayan Kamath2013-11-211-0/+15
|\ \ | |/ | | | | | | * commit '7e528a4abc3001c269142c4e3b8e17427d932ba0': Improve URI.equals performance.
| * Improve URI.equals performance.Narayan Kamath2013-11-201-0/+15
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | The main improvement is that we're vastly reducing the amount of garbage this method creates. Also, get rid of some incorrect checks and add a new benchmark test case. Before vs. After: URI Equals 1540.1 = URI EqualsWithHeavilyEscapedComponent 11921397.1 ============================== URI Equals 1337.2 = URI EqualsWithHeavilyEscapedComponent 588934.3 ============================== Change-Id: I9ddb87b8267f0ac6a3ac85c2989cc0dd305ef6a1
* | A char array copy optimization (libcore).Hiroshi Yamauchi2013-11-051-0/+40
|/ | | | | | | | | | | | | | - Based on measurements, copy char by char for arrays of length <= 64. - With this change, the Ritz MemAllocBench got ~25% faster on Nexus 4 and ~20% faster on host. - This change only handles arraycopy calls in the core libraries and char arrays with the rest future work. Bug: 7103825 (cherry picked from commit a3eb10feba1c3479ad99ef5c5323d011a5365f4b) Change-Id: Ic002ced3bc2f2a682cd4b8cec07ede990edd7463
* Merge "Add a cache to ICU.getBestDateTimePattern."Elliott Hughes2013-10-311-0/+32
|\
| * Add a cache to ICU.getBestDateTimePattern.Elliott Hughes2013-10-311-0/+32
| | | | | | | | | | | | | | Brings the time down from 7ms to 7us for a cache hit on Nexus 4. Bug: 11447043 Change-Id: I7cfc055c264ce49dadc72b5acdc2f2b0b339a5a1
* | Remove unused code from CharsetBenchmark.Narayan Kamath2013-10-301-59/+0
|/ | | | | | | | This in preparation for removing UnsafeBytesequence. Split into a separate change so that it's easier to resurrect if required. Change-Id: I24b8cd6a56a37d4efc23cfea0b59adc0f065f5a2
* Add benchmark for SSL loopbackKenny Root2013-09-291-0/+61
| | | | Change-Id: Ifb05ee34f4398ad61e047d1a7ecfcea9a787b6ca
* icu4c DateIntervalFormat objects are expensive enough that we need to cache ↵Elliott Hughes2013-09-111-0/+56
| | | | | | | | | | | | | them. It takes ~1ms to create a DateIntervalFormat on a z620, and around 8ms on a current mobile device (Nexus 4). Add a small cache of recently-used formatters, using a big lock rather than per-thread caches since this typically only happens on the UI thread anyway, and because all the other frameworks date/time formatting is behind a single lock too. Bug: 10696944 Change-Id: I8ccbe0b31b722a95a08fbc5a119173b7a0f44807
* Move URLConnectionBenchmark to latest mockwebserver APIBrian Carlstrom2013-05-091-10/+13
| | | | | | (cherry picked from commit 50b871d4eb3b42a8209bb4a6d54cb649f56e873f) Change-Id: I67e7ae5de6fc3b568c34bf56706dd5146d9f2b18
* Fix mentions of JSSEKenny Root2013-05-091-0/+1
| | | | | | (cherry picked from commit ee41931d976501d0fb4516bd43919b9564558619) Change-Id: If14d8a4be38286df1e9239add0cc7161559e7fb7
* am 645c1562: Merge "Replace mentions of old JSSE package name"Kenny Root2013-05-031-1/+1
|\ | | | | | | | | * commit '645c15629a8abdb9f3a6e428ccb393eacd933552': Replace mentions of old JSSE package name
| * Replace mentions of old JSSE package nameKenny Root2013-05-031-1/+1
| | | | | | | | Change-Id: I7c90b08ee2b43c73e02512d0c6caaec098967e93
* | Build Caliper microbenchmark tests.Also fix the URLConnectionBenchmark tests ↵Tsu Chiang Chuang2013-03-2873-7/+241
|/ | | | | | to use the external/mockwebserver. Change-Id: I48ec32e94f992fe570a6d729bb38971b3f211188
* Use H instead of k in DateToStringBenchmark.Elliott Hughes2013-03-181-1/+1
| | | | | Bug: 8359981 Change-Id: I01f495f84035e2cc35196c65f0002328d347c4d0
* Add the code.google.com/p/dalvik benchmarks to libcore.Elliott Hughes2013-02-1370-0/+7239
I'll remove them from code.google.com afterwards. Change-Id: I369a301ed0e116c7971caebc607c5a86d06d4ee2
generated by cgit v1.1 at 2025-01-10 09:23:34 +0000