msg162474 - (view) |
Author: Serhiy Storchaka (serhiy.storchaka) *  |
Date: 2012-06-07 13:57 |
In pair to here is a patch than speed up UTF-32 encoding in several times. In addition, it fixes an unsafe check of an integer overflow. Here are the results of benchmarking. See benchmark tools in https://bitbucket.org/storchaka/cpython-stuff repository. On 32-bit Linux, AMD Athlon 64 X2 4600+ @ 2.4GHz: Py2.7 Py3.2 Py3.3 patched 541 (+1032%) 541 (+1032%) 844 (+626%) 6125 encode utf-32le 'A'*10000 543 (+1056%) 541 (+1060%) 844 (+643%) 6275 encode utf-32le '\x80'*10000 544 (+1010%) 542 (+1014%) 843 (+616%) 6037 encode utf-32le '\x80'+'A'*9999 541 (+799%) 542 (+797%) 764 (+537%) 4864 encode utf-32le '\u0100'*10000 544 (+781%) 542 (+784%) 767 (+525%) 4793 encode utf-32le '\u0100'+'A'*9999 544 (+789%) 542 (+792%) 766 (+531%) 4834 encode utf-32le '\u0100'+'\x80'*9999 542 (+799%) 541 (+801%) 764 (+538%) 4874 encode utf-32le '\u8000'*10000 544 (+779%) 542 (+782%) 767 (+523%) 4780 encode utf-32le '\u8000'+'A'*9999 544 (+793%) 542 (+796%) 766 (+534%) 4859 encode utf-32le '\u8000'+'\x80'*9999 544 (+819%) 542 (+823%) 766 (+553%) 5001 encode utf-32le '\u8000'+'\u0100'*9999 430 (+867%) 427 (+874%) 860 (+383%) 4157 encode utf-32le '\U00010000'*10000 543 (+655%) 543 (+655%) 861 (+376%) 4101 encode utf-32le '\U00010000'+'A'*9999 543 (+658%) 543 (+658%) 861 (+378%) 4116 encode utf-32le '\U00010000'+'\x80'*9999 543 (+670%) 543 (+670%) 859 (+387%) 4180 encode utf-32le '\U00010000'+'\u0100'*9999 543 (+666%) 543 (+666%) 860 (+383%) 4158 encode utf-32le '\U00010000'+'\u8000'*9999 541 (+880%) 543 (+876%) 844 (+528%) 5300 encode utf-32be 'A'*10000 541 (+872%) 542 (+870%) 844 (+523%) 5256 encode utf-32be '\x80'*10000 544 (+843%) 542 (+846%) 843 (+509%) 5130 encode utf-32be '\x80'+'A'*9999 541 (+363%) 542 (+362%) 764 (+228%) 2505 encode utf-32be '\u0100'*10000 544 (+366%) 542 (+368%) 766 (+231%) 2534 encode utf-32be '\u0100'+'A'*9999 544 (+363%) 542 (+365%) 766 (+229%) 2519 encode utf-32be '\u0100'+'\x80'*9999 542 (+363%) 541 (+364%) 764 (+228%) 2509 encode utf-32be '\u8000'*10000 544 (+366%) 542 (+368%) 766 (+231%) 2534 encode utf-32be '\u8000'+'A'*9999 544 (+363%) 542 (+364%) 766 (+229%) 2517 encode utf-32be '\u8000'+'\x80'*9999 544 (+372%) 542 (+374%) 766 (+235%) 2568 encode utf-32be '\u8000'+'\u0100'*9999 430 (+428%) 427 (+432%) 860 (+164%) 2270 encode utf-32be '\U00010000'*10000 543 (+317%) 541 (+318%) 861 (+163%) 2262 encode utf-32be '\U00010000'+'A'*9999 543 (+320%) 541 (+321%) 861 (+165%) 2279 encode utf-32be '\U00010000'+'\x80'*9999 543 (+322%) 541 (+323%) 859 (+167%) 2290 encode utf-32be '\U00010000'+'\u0100'*9999 543 (+322%) 541 (+324%) 860 (+167%) 2292 encode utf-32be '\U00010000'+'\u8000'*9999 |
|
|
msg162823 - (view) |
Author: Serhiy Storchaka (serhiy.storchaka) *  |
Date: 2012-06-14 20:30 |
On 32-bit Linux, Intel Atom N570 @ 1.66GHz: Py2.7 Py3.2 Py3.3 patched 214 (+718%) 215 (+714%) 363 (+382%) 1750 encode utf-32le 'A'*10000 214 (+704%) 214 (+704%) 362 (+375%) 1720 encode utf-32le '\x80'*10000 214 (+712%) 215 (+708%) 363 (+379%) 1738 encode utf-32le '\x80'+'A'*9999 214 (+698%) 214 (+698%) 342 (+399%) 1707 encode utf-32le '\u0100'*10000 214 (+688%) 215 (+684%) 343 (+392%) 1686 encode utf-32le '\u0100'+'A'*9999 214 (+699%) 215 (+695%) 342 (+400%) 1710 encode utf-32le '\u0100'+'\x80'*9999 214 (+694%) 214 (+694%) 342 (+397%) 1699 encode utf-32le '\u8000'*10000 214 (+688%) 215 (+685%) 343 (+392%) 1687 encode utf-32le '\u8000'+'A'*9999 214 (+700%) 214 (+700%) 342 (+401%) 1713 encode utf-32le '\u8000'+'\x80'*9999 214 (+682%) 215 (+679%) 342 (+389%) 1674 encode utf-32le '\u8000'+'\u0100'*9999 121 (+2237%) 121 (+2237%) 333 (+749%) 2828 encode utf-32le '\U00010000'*10000 214 (+1108%) 214 (+1108%) 333 (+676%) 2585 encode utf-32le '\U00010000'+'A'*9999 214 (+1112%) 214 (+1112%) 333 (+679%) 2594 encode utf-32le '\U00010000'+'\x80'*9999 214 (+1208%) 214 (+1208%) 333 (+741%) 2799 encode utf-32le '\U00010000'+'\u0100'*9999 214 (+1214%) 215 (+1208%) 333 (+745%) 2813 encode utf-32le '\U00010000'+'\u8000'*9999 214 (+556%) 214 (+556%) 363 (+287%) 1404 encode utf-32be 'A'*10000 214 (+558%) 214 (+558%) 363 (+288%) 1408 encode utf-32be '\x80'*10000 214 (+550%) 214 (+550%) 363 (+283%) 1390 encode utf-32be '\x80'+'A'*9999 214 (+224%) 214 (+224%) 342 (+103%) 693 encode utf-32be '\u0100'*10000 214 (+229%) 214 (+229%) 343 (+105%) 703 encode utf-32be '\u0100'+'A'*9999 214 (+221%) 214 (+221%) 342 (+101%) 688 encode utf-32be '\u0100'+'\x80'*9999 214 (+224%) 214 (+224%) 342 (+103%) 694 encode utf-32be '\u8000'*10000 215 (+227%) 214 (+229%) 343 (+105%) 704 encode utf-32be '\u8000'+'A'*9999 214 (+221%) 214 (+221%) 342 (+101%) 686 encode utf-32be '\u8000'+'\x80'*9999 214 (+222%) 214 (+222%) 341 (+102%) 690 encode utf-32be '\u8000'+'\u0100'*9999 121 (+387%) 121 (+387%) 333 (+77%) 589 encode utf-32be '\U00010000'*10000 214 (+174%) 215 (+173%) 333 (+76%) 587 encode utf-32be '\U00010000'+'A'*9999 214 (+183%) 214 (+183%) 333 (+82%) 606 encode utf-32be '\U00010000'+'\x80'*9999 214 (+184%) 214 (+184%) 333 (+82%) 607 encode utf-32be '\U00010000'+'\u0100'*9999 214 (+183%) 214 (+183%) 333 (+82%) 605 encode utf-32be '\U00010000'+'\u8000'*9999 |
|
|
msg173404 - (view) |
Author: Serhiy Storchaka (serhiy.storchaka) *  |
Date: 2012-10-20 19:05 |
Patch updated to 3.4. Is anyone interested in 7x speedup of UTF-32 encoder? |
|
|
msg205912 - (view) |
Author: Mark Lawrence (BreamoreBoy) * |
Date: 2013-12-11 18:28 |
From http://kmike.ru/python-data-structures/ under heading DATrie "Python wrapper uses utf_32_le codec internally; this codec is currently slow and it is the bottleneck for datrie. There is a ticket with a patch in the CPython bug tracker (http://bugs.python.org/issue15027) that should make this codec fast, so there is a hope datrie will become faster with future Pythons." |
|
|
msg205934 - (view) |
Author: Serhiy Storchaka (serhiy.storchaka) *  |
Date: 2013-12-11 22:17 |
Here is updated patch, synchronized with trunk. UTF-32 encoder now checks surrogates and therefore speedup is less (only up to 5 times). But this compensates regression in 3.4. On 32-bit Linux, Intel Atom N570 @ 1.66GHz: Py3.3 Py3.4 patched 531 (+245%) 489 (+274%) 1831 encode utf-32le 'A'*10000 383 (+158%) 223 (+344%) 990 encode utf-32le '\u0100'*10000 325 (+262%) 229 (+414%) 1177 encode utf-32le '\U00010000'*10000 544 (+166%) 494 (+193%) 1448 encode utf-32be 'A'*10000 384 (+67%) 223 (+188%) 642 encode utf-32be '\u0100'*10000 323 (+108%) 229 (+193%) 671 encode utf-32be '\U00010000'*10000 |
|
|
msg205940 - (view) |
Author: Gregory P. Smith (gregory.p.smith) *  |
Date: 2013-12-11 23:05 |
one comment to address on the review, otherwise after addressing that I believe this is ready to go in for 3.4. |
|
|
msg207292 - (view) |
Author: Roundup Robot (python-dev)  |
Date: 2014-01-04 17:26 |
New changeset b72c5573c5e7 by Serhiy Storchaka in branch 'default': Issue #15027: Rewrite the UTF-32 encoder. It is now 1.6x to 3.5x faster. http://hg.python.org/cpython/rev/b72c5573c5e7 |
|
|
msg207294 - (view) |
Author: Serhiy Storchaka (serhiy.storchaka) *  |
Date: 2014-01-04 17:32 |
Thank you Gregory for your review. |
|
|
msg207302 - (view) |
Author: Larry Hastings (larry) *  |
Date: 2014-01-04 18:41 |
Isn't this a new feature? |
|
|
msg207305 - (view) |
Author: Serhiy Storchaka (serhiy.storchaka) *  |
Date: 2014-01-04 19:59 |
Sorry if I have missed. Should I revert changeset b72c5573c5e7? This patch doesn't introduce new functions and doesn't change behavior. Without this patch the UTF-32 encoder is up to 2.5x slower in 3.4 than in 3.3 (due to ). |
|
|
msg207306 - (view) |
Author: Larry Hastings (larry) *  |
Date: 2014-01-04 20:10 |
Would you describe it as a "bug fix" or a "security fix"? If it's neither of those things, then you need special permission to add it during beta. And given that this patch has the possibility of causing bugs, I'd prefer to not accept it for 3.4. Please revert it for now. If you think it should go in to 3.4, you may ask on python-dev that it be considered and take a poll. (Note that the poll is not binding on me; this is still solely my decision. However if there was an uproar of support for your patch, that would certainly cause me to reconsider.) |
|
|
msg207311 - (view) |
Author: Roundup Robot (python-dev)  |
Date: 2014-01-04 20:51 |
New changeset 1e345924f7ea by Serhiy Storchaka in branch 'default': Reverted changeset b72c5573c5e7 (issue #15027). http://hg.python.org/cpython/rev/1e345924f7ea |
|
|
msg210147 - (view) |
Author: Larry Hastings (larry) *  |
Date: 2014-02-03 16:02 |
BreamoreBoy: why did you remove Arfrever from this issue? |
|
|
msg210148 - (view) |
Author: Charles-François Natali (neologix) *  |
Date: 2014-02-03 16:28 |
> BreamoreBoy: why did you remove Arfrever from this issue? Noisy lists members are sorted by alphabetical order: since Arfrever comes just before BreamoreBoy, I assume his fingers tripped ;-) |
|
|
msg242871 - (view) |
Author: Mark Lawrence (BreamoreBoy) * |
Date: 2015-05-10 23:21 |
As this appears to be a performance improvement only can it go into 3.5 or do we wait for 3.x? |
|
|
msg242954 - (view) |
Author: Serhiy Storchaka (serhiy.storchaka) *  |
Date: 2015-05-12 10:22 |
Can I commit the patch now Larry? |
|
|
msg242981 - (view) |
Author: Larry Hastings (larry) *  |
Date: 2015-05-12 15:43 |
We're still in alpha, so it's fine for 3.5 right now. The cutoff for new features for 3.5 will be May 23. |
|
|
msg243005 - (view) |
Author: Roundup Robot (python-dev)  |
Date: 2015-05-12 20:13 |
New changeset 80cf7723c4cf by Serhiy Storchaka in branch 'default': Issue #15027: The UTF-32 encoder is now 3x to 7x faster. https://hg.python.org/cpython/rev/80cf7723c4cf |
|
|
msg243008 - (view) |
Author: Serhiy Storchaka (serhiy.storchaka) *  |
Date: 2015-05-12 20:26 |
And that's not all... |
|
|
msg243523 - (view) |
Author: Arfrever Frehtes Taifersar Arahesis (Arfrever) *  |
Date: 2015-05-18 19:14 |
In Objects/stringlib/codecs.h in 2 comments U+DC800 should be changed into U+D800 (from definition of Py_UNICODE_IS_SURROGATE) or U+DC80 (from result of b"\x80".decode(errors="surrogateescape")). |
|
|
msg243524 - (view) |
Author: Serhiy Storchaka (serhiy.storchaka) *  |
Date: 2015-05-18 19:22 |
Thank you Arfrever. That was copy-pasted old typo. Fixed in 3d5bf6174c4b and bc6ed8360312. |
|
|