moses-smt/mosesdecoder
1
1
Fork 0
mirror of https://github.com/moses-smt/mosesdecoder.git synced 2024-09-11 11:25:40 +03:00
Code Issues Packages Projects Releases Wiki Activity

Extract kenlm into a submodule, referencing lazy

Browse Source
This commit is contained in:
Kenneth Heafield 2012-09-24 18:30:51 +01:00
parent 0580c1ae9f
commit 4faab6c68f
101 changed files with 14 additions and 12413 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
.gitmodules vendored
View File

@ -1,3 +1,6 @@
[submodule "regression-testing/tests"]
path = regression-testing/tests
url = ../moses-regression-tests.git
[submodule "lazy"]
path = lazy
url = git@github.com:kpu/lazy

4
Jamroot
View File

@ -101,9 +101,9 @@ project : requirements
;
#Add directories here if you want their incidental targets too (i.e. tests).
build-projects util lm mert moses-cmd/src moses-chart-cmd/src scripts regression-testing contrib/relent-filter/src ;
build-projects lazy mert moses-cmd/src moses-chart-cmd/src scripts regression-testing contrib/relent-filter/src ;
alias programs : lm//query lm//build_binary lm//kenlm_max_order moses-chart-cmd/src//moses_chart moses-cmd/src//programs OnDiskPt//CreateOnDiskPt OnDiskPt//queryOnDiskPt mert//programs contrib/server//mosesserver misc//programs symal phrase-extract phrase-extract//lexical-reordering phrase-extract//extract-ghkm phrase-extract//pcfg-extract phrase-extract//pcfg-score biconcor contrib/relent-filter/src//calcDivergence ;
alias programs : lazy/lm//programs moses-chart-cmd/src//moses_chart moses-cmd/src//programs OnDiskPt//CreateOnDiskPt OnDiskPt//queryOnDiskPt mert//programs contrib/server//mosesserver misc//programs symal phrase-extract phrase-extract//lexical-reordering phrase-extract//extract-ghkm phrase-extract//pcfg-extract phrase-extract//pcfg-score biconcor contrib/relent-filter/src//calcDivergence ;
install-bin-libs programs ;
install-headers headers-base : [ path.glob-tree biconcor contrib lm mert misc moses-chart-cmd moses-cmd OnDiskPt phrase-extract symal util : *.hh *.h ] : . ;

6
jam-files/sanity.jam
View File

@ -63,7 +63,7 @@ requirements = ;
FORCE-STATIC = [ option.get "static" : : "yes" ] ;
if $(FORCE-STATIC) {
requirements += <runtime-link>static ;
requirements += <link>static <runtime-link>static ;
}
#Determine if a library can be compiled statically.
@ -152,7 +152,7 @@ rule boost ( min-version ) {
boost-lib program_options PROGRAM_OPTIONS_DYN_LINK ;
boost-lib unit_test_framework TEST_DYN_LINK ;
boost-lib iostreams IOSTREAMS_DYN_LINK ;
boost-lib filesystem FILE_SYSTEM_DYN_LINK : boost_system ;
boost-lib filesystem FILE_SYSTEM_DYN_LINK ;
}
#Link normally to a library, but sometimes static isn't installed so fall back to dynamic.
@ -282,3 +282,5 @@ if [ option.get "sanity-test" : : "yes" ] {
EXIT "Bad" : 1 ;
}
}
use-project /top : . ;

1
lazy Submodule

@ -0,0 +1 @@
Subproject commit 5b7033200a7172b2fd4e04c56ea6f24d43f40d04

674
lm/COPYING
View File

@ -1,674 +0,0 @@
GNU GENERAL PUBLIC LICENSE
Version 3, 29 June 2007
Copyright (C) 2007 Free Software Foundation, Inc. <http://fsf.org/>
Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed.
Preamble
The GNU General Public License is a free, copyleft license for
software and other kinds of works.
The licenses for most software and other practical works are designed
to take away your freedom to share and change the works. By contrast,
the GNU General Public License is intended to guarantee your freedom to
share and change all versions of a program--to make sure it remains free
software for all its users. We, the Free Software Foundation, use the
GNU General Public License for most of our software; it applies also to
any other work released this way by its authors. You can apply it to
your programs, too.
When we speak of free software, we are referring to freedom, not
price. Our General Public Licenses are designed to make sure that you
have the freedom to distribute copies of free software (and charge for
them if you wish), that you receive source code or can get it if you
want it, that you can change the software or use pieces of it in new
free programs, and that you know you can do these things.
To protect your rights, we need to prevent others from denying you
these rights or asking you to surrender the rights. Therefore, you have
certain responsibilities if you distribute copies of the software, or if
you modify it: responsibilities to respect the freedom of others.
For example, if you distribute copies of such a program, whether
gratis or for a fee, you must pass on to the recipients the same
freedoms that you received. You must make sure that they, too, receive
or can get the source code. And you must show them these terms so they
know their rights.
Developers that use the GNU GPL protect your rights with two steps:
(1) assert copyright on the software, and (2) offer you this License
giving you legal permission to copy, distribute and/or modify it.
For the developers' and authors' protection, the GPL clearly explains
that there is no warranty for this free software. For both users' and
authors' sake, the GPL requires that modified versions be marked as
changed, so that their problems will not be attributed erroneously to
authors of previous versions.
Some devices are designed to deny users access to install or run
modified versions of the software inside them, although the manufacturer
can do so. This is fundamentally incompatible with the aim of
protecting users' freedom to change the software. The systematic
pattern of such abuse occurs in the area of products for individuals to
use, which is precisely where it is most unacceptable. Therefore, we
have designed this version of the GPL to prohibit the practice for those
products. If such problems arise substantially in other domains, we
stand ready to extend this provision to those domains in future versions
of the GPL, as needed to protect the freedom of users.
Finally, every program is threatened constantly by software patents.
States should not allow patents to restrict development and use of
software on general-purpose computers, but in those that do, we wish to
avoid the special danger that patents applied to a free program could
make it effectively proprietary. To prevent this, the GPL assures that
patents cannot be used to render the program non-free.
The precise terms and conditions for copying, distribution and
modification follow.
TERMS AND CONDITIONS
0. Definitions.
"This License" refers to version 3 of the GNU General Public License.
"Copyright" also means copyright-like laws that apply to other kinds of
works, such as semiconductor masks.
"The Program" refers to any copyrightable work licensed under this
License. Each licensee is addressed as "you". "Licensees" and
"recipients" may be individuals or organizations.
To "modify" a work means to copy from or adapt all or part of the work
in a fashion requiring copyright permission, other than the making of an
exact copy. The resulting work is called a "modified version" of the
earlier work or a work "based on" the earlier work.
A "covered work" means either the unmodified Program or a work based
on the Program.
To "propagate" a work means to do anything with it that, without
permission, would make you directly or secondarily liable for
infringement under applicable copyright law, except executing it on a
computer or modifying a private copy. Propagation includes copying,
distribution (with or without modification), making available to the
public, and in some countries other activities as well.
To "convey" a work means any kind of propagation that enables other
parties to make or receive copies. Mere interaction with a user through
a computer network, with no transfer of a copy, is not conveying.
An interactive user interface displays "Appropriate Legal Notices"
to the extent that it includes a convenient and prominently visible
feature that (1) displays an appropriate copyright notice, and (2)
tells the user that there is no warranty for the work (except to the
extent that warranties are provided), that licensees may convey the
work under this License, and how to view a copy of this License. If
the interface presents a list of user commands or options, such as a
menu, a prominent item in the list meets this criterion.
1. Source Code.
The "source code" for a work means the preferred form of the work
for making modifications to it. "Object code" means any non-source
form of a work.
A "Standard Interface" means an interface that either is an official
standard defined by a recognized standards body, or, in the case of
interfaces specified for a particular programming language, one that
is widely used among developers working in that language.
The "System Libraries" of an executable work include anything, other
than the work as a whole, that (a) is included in the normal form of
packaging a Major Component, but which is not part of that Major
Component, and (b) serves only to enable use of the work with that
Major Component, or to implement a Standard Interface for which an
implementation is available to the public in source code form. A
"Major Component", in this context, means a major essential component
(kernel, window system, and so on) of the specific operating system
(if any) on which the executable work runs, or a compiler used to
produce the work, or an object code interpreter used to run it.
The "Corresponding Source" for a work in object code form means all
the source code needed to generate, install, and (for an executable
work) run the object code and to modify the work, including scripts to
control those activities. However, it does not include the work's
System Libraries, or general-purpose tools or generally available free
programs which are used unmodified in performing those activities but
which are not part of the work. For example, Corresponding Source
includes interface definition files associated with source files for
the work, and the source code for shared libraries and dynamically
linked subprograms that the work is specifically designed to require,
such as by intimate data communication or control flow between those
subprograms and other parts of the work.
The Corresponding Source need not include anything that users
can regenerate automatically from other parts of the Corresponding
Source.
The Corresponding Source for a work in source code form is that
same work.
2. Basic Permissions.
All rights granted under this License are granted for the term of
copyright on the Program, and are irrevocable provided the stated
conditions are met. This License explicitly affirms your unlimited
permission to run the unmodified Program. The output from running a
covered work is covered by this License only if the output, given its
content, constitutes a covered work. This License acknowledges your
rights of fair use or other equivalent, as provided by copyright law.
You may make, run and propagate covered works that you do not
convey, without conditions so long as your license otherwise remains
in force. You may convey covered works to others for the sole purpose
of having them make modifications exclusively for you, or provide you
with facilities for running those works, provided that you comply with
the terms of this License in conveying all material for which you do
not control copyright. Those thus making or running the covered works
for you must do so exclusively on your behalf, under your direction
and control, on terms that prohibit them from making any copies of
your copyrighted material outside their relationship with you.
Conveying under any other circumstances is permitted solely under
the conditions stated below. Sublicensing is not allowed; section 10
makes it unnecessary.
3. Protecting Users' Legal Rights From Anti-Circumvention Law.
No covered work shall be deemed part of an effective technological
measure under any applicable law fulfilling obligations under article
11 of the WIPO copyright treaty adopted on 20 December 1996, or
similar laws prohibiting or restricting circumvention of such
measures.
When you convey a covered work, you waive any legal power to forbid
circumvention of technological measures to the extent such circumvention
is effected by exercising rights under this License with respect to
the covered work, and you disclaim any intention to limit operation or
modification of the work as a means of enforcing, against the work's
users, your or third parties' legal rights to forbid circumvention of
technological measures.
4. Conveying Verbatim Copies.
You may convey verbatim copies of the Program's source code as you
receive it, in any medium, provided that you conspicuously and
appropriately publish on each copy an appropriate copyright notice;
keep intact all notices stating that this License and any
non-permissive terms added in accord with section 7 apply to the code;
keep intact all notices of the absence of any warranty; and give all
recipients a copy of this License along with the Program.
You may charge any price or no price for each copy that you convey,
and you may offer support or warranty protection for a fee.
5. Conveying Modified Source Versions.
You may convey a work based on the Program, or the modifications to
produce it from the Program, in the form of source code under the
terms of section 4, provided that you also meet all of these conditions:
a) The work must carry prominent notices stating that you modified
it, and giving a relevant date.
b) The work must carry prominent notices stating that it is
released under this License and any conditions added under section
7. This requirement modifies the requirement in section 4 to
"keep intact all notices".
c) You must license the entire work, as a whole, under this
License to anyone who comes into possession of a copy. This
License will therefore apply, along with any applicable section 7
additional terms, to the whole of the work, and all its parts,
regardless of how they are packaged. This License gives no
permission to license the work in any other way, but it does not
invalidate such permission if you have separately received it.
d) If the work has interactive user interfaces, each must display
Appropriate Legal Notices; however, if the Program has interactive
interfaces that do not display Appropriate Legal Notices, your
work need not make them do so.
A compilation of a covered work with other separate and independent
works, which are not by their nature extensions of the covered work,
and which are not combined with it such as to form a larger program,
in or on a volume of a storage or distribution medium, is called an
"aggregate" if the compilation and its resulting copyright are not
used to limit the access or legal rights of the compilation's users
beyond what the individual works permit. Inclusion of a covered work
in an aggregate does not cause this License to apply to the other
parts of the aggregate.
6. Conveying Non-Source Forms.
You may convey a covered work in object code form under the terms
of sections 4 and 5, provided that you also convey the
machine-readable Corresponding Source under the terms of this License,
in one of these ways:
a) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by the
Corresponding Source fixed on a durable physical medium
customarily used for software interchange.
b) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by a
written offer, valid for at least three years and valid for as
long as you offer spare parts or customer support for that product
model, to give anyone who possesses the object code either (1) a
copy of the Corresponding Source for all the software in the
product that is covered by this License, on a durable physical
medium customarily used for software interchange, for a price no
more than your reasonable cost of physically performing this
conveying of source, or (2) access to copy the
Corresponding Source from a network server at no charge.
c) Convey individual copies of the object code with a copy of the
written offer to provide the Corresponding Source. This
alternative is allowed only occasionally and noncommercially, and
only if you received the object code with such an offer, in accord
with subsection 6b.
d) Convey the object code by offering access from a designated
place (gratis or for a charge), and offer equivalent access to the
Corresponding Source in the same way through the same place at no
further charge. You need not require recipients to copy the
Corresponding Source along with the object code. If the place to
copy the object code is a network server, the Corresponding Source
may be on a different server (operated by you or a third party)
that supports equivalent copying facilities, provided you maintain
clear directions next to the object code saying where to find the
Corresponding Source. Regardless of what server hosts the
Corresponding Source, you remain obligated to ensure that it is
available for as long as needed to satisfy these requirements.
e) Convey the object code using peer-to-peer transmission, provided
you inform other peers where the object code and Corresponding
Source of the work are being offered to the general public at no
charge under subsection 6d.
A separable portion of the object code, whose source code is excluded
from the Corresponding Source as a System Library, need not be
included in conveying the object code work.
A "User Product" is either (1) a "consumer product", which means any
tangible personal property which is normally used for personal, family,
or household purposes, or (2) anything designed or sold for incorporation
into a dwelling. In determining whether a product is a consumer product,
doubtful cases shall be resolved in favor of coverage. For a particular
product received by a particular user, "normally used" refers to a
typical or common use of that class of product, regardless of the status
of the particular user or of the way in which the particular user
actually uses, or expects or is expected to use, the product. A product
is a consumer product regardless of whether the product has substantial
commercial, industrial or non-consumer uses, unless such uses represent
the only significant mode of use of the product.
"Installation Information" for a User Product means any methods,
procedures, authorization keys, or other information required to install
and execute modified versions of a covered work in that User Product from
a modified version of its Corresponding Source. The information must
suffice to ensure that the continued functioning of the modified object
code is in no case prevented or interfered with solely because
modification has been made.
If you convey an object code work under this section in, or with, or
specifically for use in, a User Product, and the conveying occurs as
part of a transaction in which the right of possession and use of the
User Product is transferred to the recipient in perpetuity or for a
fixed term (regardless of how the transaction is characterized), the
Corresponding Source conveyed under this section must be accompanied
by the Installation Information. But this requirement does not apply
if neither you nor any third party retains the ability to install
modified object code on the User Product (for example, the work has
been installed in ROM).
The requirement to provide Installation Information does not include a
requirement to continue to provide support service, warranty, or updates
for a work that has been modified or installed by the recipient, or for
the User Product in which it has been modified or installed. Access to a
network may be denied when the modification itself materially and
adversely affects the operation of the network or violates the rules and
protocols for communication across the network.
Corresponding Source conveyed, and Installation Information provided,
in accord with this section must be in a format that is publicly
documented (and with an implementation available to the public in
source code form), and must require no special password or key for
unpacking, reading or copying.
7. Additional Terms.
"Additional permissions" are terms that supplement the terms of this
License by making exceptions from one or more of its conditions.
Additional permissions that are applicable to the entire Program shall
be treated as though they were included in this License, to the extent
that they are valid under applicable law. If additional permissions
apply only to part of the Program, that part may be used separately
under those permissions, but the entire Program remains governed by
this License without regard to the additional permissions.
When you convey a copy of a covered work, you may at your option
remove any additional permissions from that copy, or from any part of
it. (Additional permissions may be written to require their own
removal in certain cases when you modify the work.) You may place
additional permissions on material, added by you to a covered work,
for which you have or can give appropriate copyright permission.
Notwithstanding any other provision of this License, for material you
add to a covered work, you may (if authorized by the copyright holders of
that material) supplement the terms of this License with terms:
a) Disclaiming warranty or limiting liability differently from the
terms of sections 15 and 16 of this License; or
b) Requiring preservation of specified reasonable legal notices or
author attributions in that material or in the Appropriate Legal
Notices displayed by works containing it; or
c) Prohibiting misrepresentation of the origin of that material, or
requiring that modified versions of such material be marked in
reasonable ways as different from the original version; or
d) Limiting the use for publicity purposes of names of licensors or
authors of the material; or
e) Declining to grant rights under trademark law for use of some
trade names, trademarks, or service marks; or
f) Requiring indemnification of licensors and authors of that
material by anyone who conveys the material (or modified versions of
it) with contractual assumptions of liability to the recipient, for
any liability that these contractual assumptions directly impose on
those licensors and authors.
All other non-permissive additional terms are considered "further
restrictions" within the meaning of section 10. If the Program as you
received it, or any part of it, contains a notice stating that it is
governed by this License along with a term that is a further
restriction, you may remove that term. If a license document contains
a further restriction but permits relicensing or conveying under this
License, you may add to a covered work material governed by the terms
of that license document, provided that the further restriction does
not survive such relicensing or conveying.
If you add terms to a covered work in accord with this section, you
must place, in the relevant source files, a statement of the
additional terms that apply to those files, or a notice indicating
where to find the applicable terms.
Additional terms, permissive or non-permissive, may be stated in the
form of a separately written license, or stated as exceptions;
the above requirements apply either way.
8. Termination.
You may not propagate or modify a covered work except as expressly
provided under this License. Any attempt otherwise to propagate or
modify it is void, and will automatically terminate your rights under
this License (including any patent licenses granted under the third
paragraph of section 11).
However, if you cease all violation of this License, then your
license from a particular copyright holder is reinstated (a)
provisionally, unless and until the copyright holder explicitly and
finally terminates your license, and (b) permanently, if the copyright
holder fails to notify you of the violation by some reasonable means
prior to 60 days after the cessation.
Moreover, your license from a particular copyright holder is
reinstated permanently if the copyright holder notifies you of the
violation by some reasonable means, this is the first time you have
received notice of violation of this License (for any work) from that
copyright holder, and you cure the violation prior to 30 days after
your receipt of the notice.
Termination of your rights under this section does not terminate the
licenses of parties who have received copies or rights from you under
this License. If your rights have been terminated and not permanently
reinstated, you do not qualify to receive new licenses for the same
material under section 10.
9. Acceptance Not Required for Having Copies.
You are not required to accept this License in order to receive or
run a copy of the Program. Ancillary propagation of a covered work
occurring solely as a consequence of using peer-to-peer transmission
to receive a copy likewise does not require acceptance. However,
nothing other than this License grants you permission to propagate or
modify any covered work. These actions infringe copyright if you do
not accept this License. Therefore, by modifying or propagating a
covered work, you indicate your acceptance of this License to do so.
10. Automatic Licensing of Downstream Recipients.
Each time you convey a covered work, the recipient automatically
receives a license from the original licensors, to run, modify and
propagate that work, subject to this License. You are not responsible
for enforcing compliance by third parties with this License.
An "entity transaction" is a transaction transferring control of an
organization, or substantially all assets of one, or subdividing an
organization, or merging organizations. If propagation of a covered
work results from an entity transaction, each party to that
transaction who receives a copy of the work also receives whatever
licenses to the work the party's predecessor in interest had or could
give under the previous paragraph, plus a right to possession of the
Corresponding Source of the work from the predecessor in interest, if
the predecessor has it or can get it with reasonable efforts.
You may not impose any further restrictions on the exercise of the
rights granted or affirmed under this License. For example, you may
not impose a license fee, royalty, or other charge for exercise of
rights granted under this License, and you may not initiate litigation
(including a cross-claim or counterclaim in a lawsuit) alleging that
any patent claim is infringed by making, using, selling, offering for
sale, or importing the Program or any portion of it.
11. Patents.
A "contributor" is a copyright holder who authorizes use under this
License of the Program or a work on which the Program is based. The
work thus licensed is called the contributor's "contributor version".
A contributor's "essential patent claims" are all patent claims
owned or controlled by the contributor, whether already acquired or
hereafter acquired, that would be infringed by some manner, permitted
by this License, of making, using, or selling its contributor version,
but do not include claims that would be infringed only as a
consequence of further modification of the contributor version. For
purposes of this definition, "control" includes the right to grant
patent sublicenses in a manner consistent with the requirements of
this License.
Each contributor grants you a non-exclusive, worldwide, royalty-free
patent license under the contributor's essential patent claims, to
make, use, sell, offer for sale, import and otherwise run, modify and
propagate the contents of its contributor version.
In the following three paragraphs, a "patent license" is any express
agreement or commitment, however denominated, not to enforce a patent
(such as an express permission to practice a patent or covenant not to
sue for patent infringement). To "grant" such a patent license to a
party means to make such an agreement or commitment not to enforce a
patent against the party.
If you convey a covered work, knowingly relying on a patent license,
and the Corresponding Source of the work is not available for anyone
to copy, free of charge and under the terms of this License, through a
publicly available network server or other readily accessible means,
then you must either (1) cause the Corresponding Source to be so
available, or (2) arrange to deprive yourself of the benefit of the
patent license for this particular work, or (3) arrange, in a manner
consistent with the requirements of this License, to extend the patent
license to downstream recipients. "Knowingly relying" means you have
actual knowledge that, but for the patent license, your conveying the
covered work in a country, or your recipient's use of the covered work
in a country, would infringe one or more identifiable patents in that
country that you have reason to believe are valid.
If, pursuant to or in connection with a single transaction or
arrangement, you convey, or propagate by procuring conveyance of, a
covered work, and grant a patent license to some of the parties
receiving the covered work authorizing them to use, propagate, modify
or convey a specific copy of the covered work, then the patent license
you grant is automatically extended to all recipients of the covered
work and works based on it.
A patent license is "discriminatory" if it does not include within
the scope of its coverage, prohibits the exercise of, or is
conditioned on the non-exercise of one or more of the rights that are
specifically granted under this License. You may not convey a covered
work if you are a party to an arrangement with a third party that is
in the business of distributing software, under which you make payment
to the third party based on the extent of your activity of conveying
the work, and under which the third party grants, to any of the
parties who would receive the covered work from you, a discriminatory
patent license (a) in connection with copies of the covered work
conveyed by you (or copies made from those copies), or (b) primarily
for and in connection with specific products or compilations that
contain the covered work, unless you entered into that arrangement,
or that patent license was granted, prior to 28 March 2007.
Nothing in this License shall be construed as excluding or limiting
any implied license or other defenses to infringement that may
otherwise be available to you under applicable patent law.
12. No Surrender of Others' Freedom.
If conditions are imposed on you (whether by court order, agreement or
otherwise) that contradict the conditions of this License, they do not
excuse you from the conditions of this License. If you cannot convey a
covered work so as to satisfy simultaneously your obligations under this
License and any other pertinent obligations, then as a consequence you may
not convey it at all. For example, if you agree to terms that obligate you
to collect a royalty for further conveying from those to whom you convey
the Program, the only way you could satisfy both those terms and this
License would be to refrain entirely from conveying the Program.
13. Use with the GNU Affero General Public License.
Notwithstanding any other provision of this License, you have
permission to link or combine any covered work with a work licensed
under version 3 of the GNU Affero General Public License into a single
combined work, and to convey the resulting work. The terms of this
License will continue to apply to the part which is the covered work,
but the special requirements of the GNU Affero General Public License,
section 13, concerning interaction through a network will apply to the
combination as such.
14. Revised Versions of this License.
The Free Software Foundation may publish revised and/or new versions of
the GNU General Public License from time to time. Such new versions will
be similar in spirit to the present version, but may differ in detail to
address new problems or concerns.
Each version is given a distinguishing version number. If the
Program specifies that a certain numbered version of the GNU General
Public License "or any later version" applies to it, you have the
option of following the terms and conditions either of that numbered
version or of any later version published by the Free Software
Foundation. If the Program does not specify a version number of the
GNU General Public License, you may choose any version ever published
by the Free Software Foundation.
If the Program specifies that a proxy can decide which future
versions of the GNU General Public License can be used, that proxy's
public statement of acceptance of a version permanently authorizes you
to choose that version for the Program.
Later license versions may give you additional or different
permissions. However, no additional obligations are imposed on any
author or copyright holder as a result of your choosing to follow a
later version.
15. Disclaimer of Warranty.
THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
16. Limitation of Liability.
IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
SUCH DAMAGES.
17. Interpretation of Sections 15 and 16.
If the disclaimer of warranty and limitation of liability provided
above cannot be given local legal effect according to their terms,
reviewing courts shall apply local law that most closely approximates
an absolute waiver of all civil liability in connection with the
Program, unless a warranty or assumption of liability accompanies a
copy of the Program in return for a fee.
END OF TERMS AND CONDITIONS
How to Apply These Terms to Your New Programs
If you develop a new program, and you want it to be of the greatest
possible use to the public, the best way to achieve this is to make it
free software which everyone can redistribute and change under these terms.
To do so, attach the following notices to the program. It is safest
to attach them to the start of each source file to most effectively
state the exclusion of warranty; and each file should have at least
the "copyright" line and a pointer to where the full notice is found.
<one line to give the program's name and a brief idea of what it does.>
Copyright (C) <year> <name of author>
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <http://www.gnu.org/licenses/>.
Also add information on how to contact you by electronic and paper mail.
If the program does terminal interaction, make it output a short
notice like this when it starts in an interactive mode:
<program> Copyright (C) <year> <name of author>
This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
This is free software, and you are welcome to redistribute it
under certain conditions; type `show c' for details.
The hypothetical commands `show w' and `show c' should show the appropriate
parts of the General Public License. Of course, your program's commands
might be different; for a GUI interface, you would use an "about box".
You should also get your employer (if you work as a programmer) or school,
if any, to sign a "copyright disclaimer" for the program, if necessary.
For more information on this, and how to apply and follow the GNU GPL, see
<http://www.gnu.org/licenses/>.
The GNU General Public License does not permit incorporating your program
into proprietary programs. If your program is a subroutine library, you
may consider it more useful to permit linking proprietary applications with
the library. If this is what you want to do, use the GNU Lesser General
Public License instead of this License. But first, please read
<http://www.gnu.org/philosophy/why-not-lgpl.html>.

165
lm/COPYING.LESSER
View File

@ -1,165 +0,0 @@
GNU LESSER GENERAL PUBLIC LICENSE
Version 3, 29 June 2007
Copyright (C) 2007 Free Software Foundation, Inc. <http://fsf.org/>
Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed.
This version of the GNU Lesser General Public License incorporates
the terms and conditions of version 3 of the GNU General Public
License, supplemented by the additional permissions listed below.
0. Additional Definitions.
As used herein, "this License" refers to version 3 of the GNU Lesser
General Public License, and the "GNU GPL" refers to version 3 of the GNU
General Public License.
"The Library" refers to a covered work governed by this License,
other than an Application or a Combined Work as defined below.
An "Application" is any work that makes use of an interface provided
by the Library, but which is not otherwise based on the Library.
Defining a subclass of a class defined by the Library is deemed a mode
of using an interface provided by the Library.
A "Combined Work" is a work produced by combining or linking an
Application with the Library. The particular version of the Library
with which the Combined Work was made is also called the "Linked
Version".
The "Minimal Corresponding Source" for a Combined Work means the
Corresponding Source for the Combined Work, excluding any source code
for portions of the Combined Work that, considered in isolation, are
based on the Application, and not on the Linked Version.
The "Corresponding Application Code" for a Combined Work means the
object code and/or source code for the Application, including any data
and utility programs needed for reproducing the Combined Work from the
Application, but excluding the System Libraries of the Combined Work.
1. Exception to Section 3 of the GNU GPL.
You may convey a covered work under sections 3 and 4 of this License
without being bound by section 3 of the GNU GPL.
2. Conveying Modified Versions.
If you modify a copy of the Library, and, in your modifications, a
facility refers to a function or data to be supplied by an Application
that uses the facility (other than as an argument passed when the
facility is invoked), then you may convey a copy of the modified
version:
a) under this License, provided that you make a good faith effort to
ensure that, in the event an Application does not supply the
function or data, the facility still operates, and performs
whatever part of its purpose remains meaningful, or
b) under the GNU GPL, with none of the additional permissions of
this License applicable to that copy.
3. Object Code Incorporating Material from Library Header Files.
The object code form of an Application may incorporate material from
a header file that is part of the Library. You may convey such object
code under terms of your choice, provided that, if the incorporated
material is not limited to numerical parameters, data structure
layouts and accessors, or small macros, inline functions and templates
(ten or fewer lines in length), you do both of the following:
a) Give prominent notice with each copy of the object code that the
Library is used in it and that the Library and its use are
covered by this License.
b) Accompany the object code with a copy of the GNU GPL and this license
document.
4. Combined Works.
You may convey a Combined Work under terms of your choice that,
taken together, effectively do not restrict modification of the
portions of the Library contained in the Combined Work and reverse
engineering for debugging such modifications, if you also do each of
the following:
a) Give prominent notice with each copy of the Combined Work that
the Library is used in it and that the Library and its use are
covered by this License.
b) Accompany the Combined Work with a copy of the GNU GPL and this license
document.
c) For a Combined Work that displays copyright notices during
execution, include the copyright notice for the Library among
these notices, as well as a reference directing the user to the
copies of the GNU GPL and this license document.
d) Do one of the following:
0) Convey the Minimal Corresponding Source under the terms of this
License, and the Corresponding Application Code in a form
suitable for, and under terms that permit, the user to
recombine or relink the Application with a modified version of
the Linked Version to produce a modified Combined Work, in the
manner specified by section 6 of the GNU GPL for conveying
Corresponding Source.
1) Use a suitable shared library mechanism for linking with the
Library. A suitable mechanism is one that (a) uses at run time
a copy of the Library already present on the user's computer
system, and (b) will operate properly with a modified version
of the Library that is interface-compatible with the Linked
Version.
e) Provide Installation Information, but only if you would otherwise
be required to provide such information under section 6 of the
GNU GPL, and only to the extent that such information is
necessary to install and execute a modified version of the
Combined Work produced by recombining or relinking the
Application with a modified version of the Linked Version. (If
you use option 4d0, the Installation Information must accompany
the Minimal Corresponding Source and Corresponding Application
Code. If you use option 4d1, you must provide the Installation
Information in the manner specified by section 6 of the GNU GPL
for conveying Corresponding Source.)
5. Combined Libraries.
You may place library facilities that are a work based on the
Library side by side in a single library together with other library
facilities that are not Applications and are not covered by this
License, and convey such a combined library under terms of your
choice, if you do both of the following:
a) Accompany the combined library with a copy of the same work based
on the Library, uncombined with any other library facilities,
conveyed under the terms of this License.
b) Give prominent notice with the combined library that part of it
is a work based on the Library, and explaining where to find the
accompanying uncombined form of the same work.
6. Revised Versions of the GNU Lesser General Public License.
The Free Software Foundation may publish revised and/or new versions
of the GNU Lesser General Public License from time to time. Such new
versions will be similar in spirit to the present version, but may
differ in detail to address new problems or concerns.
Each version is given a distinguishing version number. If the
Library as you received it specifies that a certain numbered version
of the GNU Lesser General Public License "or any later version"
applies to it, you have the option of following the terms and
conditions either of that published version or of any later version
published by the Free Software Foundation. If the Library as you
received it does not specify a version number of the GNU Lesser
General Public License, you may choose any version of the GNU Lesser
General Public License ever published by the Free Software Foundation.
If the Library as you received it specifies that a proxy can decide
whether future versions of the GNU Lesser General Public License shall
apply, that proxy's public statement of acceptance of any version is
permanent authorization for you to choose that version for the
Library.

20
lm/Jamfile
View File

@ -1,20 +0,0 @@
# If you need higher order, change this option
# Having this limit means that State can be
# (KENLM_MAX_ORDER - 1) * sizeof(float) bytes instead of
# sizeof(float*) + (KENLM_MAX_ORDER - 1) * sizeof(float) + malloc overhead
max-order = [ option.get "max-kenlm-order" : 6 : 6 ] ;
if ( $(max-order) != 6 ) {
echo "Setting KenLM maximum n-gram order to $(max-order)" ;
}
max-order = <define>KENLM_MAX_ORDER=$(max-order) ;
lib kenlm : bhiksha.cc binary_format.cc config.cc lm_exception.cc model.cc quantize.cc read_arpa.cc search_hashed.cc search_trie.cc trie.cc trie_sort.cc value_build.cc virtual_interface.cc vocab.cc ../util//kenutil : <include>.. $(max-order) : : <include>.. <library>../util//kenutil $(max-order) ;
import testing ;
run left_test.cc ../util//kenutil kenlm ..//boost_unit_test_framework : : test.arpa ;
run model_test.cc ../util//kenutil kenlm ..//boost_unit_test_framework : : test.arpa test_nounk.arpa ;
exe query : ngram_query.cc kenlm ../util//kenutil ;
exe build_binary : build_binary.cc kenlm ../util//kenutil ;
exe kenlm_max_order : max_order.cc : $(max-order) ;

12
lm/LICENSE
View File

@ -1,12 +0,0 @@
Avenue code is free software: you can redistribute it and/or modify
it under the terms of the GNU Lesser General Public License as published
by the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
Avenue code is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public License
along with Avenue code. If not, see <http://www.gnu.org/licenses/>.

44
lm/README
View File

@ -1,44 +0,0 @@
Language model inference code by Kenneth Heafield <kenlm at kheafield.com>
THE GIT REPOSITORY https://github.com/kpu/kenlm IS WHERE ACTIVE DEVELOPMENT HAPPENS. IT MAY RETURN SILENTLY WRONG ANSWERS OR BE SILENTLY BINARY-INCOMPATIBLE WITH STABLE RELEASES.
The website http://kheafield.com/code/kenlm/ has more documentation. If you're a decoder developer, please download the latest version from there instead of copying from another decoder.
Two data structures are supported: probing and trie. Probing is a probing hash table with keys that ere 64-bit hashes of n-grams and floats as values. Trie is a fairly standard trie but with bit-level packing so it uses the minimum number of bits to store word indices and pointers. The trie node entries are sorted by word index. Probing is the fastest and uses the most memory. Trie uses the least memory and a bit slower.
With trie, resident memory is 58% of IRST's smallest version and 21% of SRI's compact version. Simultaneously, trie CPU's use is 81% of IRST's fastest version and 84% of SRI's fast version. KenLM's probing hash table implementation goes even faster at the expense of using more memory. See http://kheafield.com/code/kenlm/benchmark/.
Binary format via mmap is supported. Run ./build_binary to make one then pass the binary file name to the appropriate Model constructor.
PLATFORMS
murmur_hash.cc and bit_packing.hh perform unaligned reads and writes that make the code architecture-dependent.
It has been sucessfully tested on x86_64, x86, and PPC64.
ARM support is reportedly working, at least on the iphone, but I cannot test this.
Runs on Linux, OS X, Cygwin, and MinGW.
Hideo Okuma and Tomoyuki Yoshimura from NICT contributed ports to ARM and MinGW. Hieu Hoang is working on a native Windows port.
DECODER DEVELOPERS
- I recommend copying the code and distributing it with your decoder. However, please send improvements upstream as indicated in CONTRIBUTORS.
- It does not depend on Boost or ICU. If you use ICU, define HAVE_ICU in util/have.hh (uncomment the line) to avoid a name conflict. Defining HAVE_BOOST will let you hash StringPiece.
- Most people have zlib. If you don't want to depend on that, comment out #define HAVE_ZLIB in util/have.hh. This will disable loading gzipped ARPA files.
- There are two build systems: compile.sh and Jamroot+Jamfile. They're pretty simple and are intended to be reimplemented in your build system.
- Use either the interface in lm/model.hh or lm/virtual_interface.hh. Interface documentation is in comments of lm/virtual_interface.hh and lm/model.hh.
- There are several possible data structures in model.hh. Use RecognizeBinary in binary_format.hh to determine which one a user has provided. You probably already implement feature functions as an abstract virtual base class with several children. I suggest you co-opt this existing virtual dispatch by templatizing the language model feature implementation on the KenLM model identified by RecognizeBinary. This is the strategy used in Moses and cdec.
- See lm/config.hh for tuning options.
CONTRIBUTORS
Contributions to KenLM are welcome. Please base your contributions on https://github.com/kpu/kenlm and send pull requests (or I might give you commit access). Downstream copies in Moses and cdec are maintained by overwriting them so do not make changes there.
The name was Hieu Hoang's idea, not mine.

95
lm/bhiksha.cc
View File

@ -1,95 +0,0 @@
#include "lm/bhiksha.hh"
#include "lm/config.hh"
#include "util/file.hh"
#include "util/exception.hh"
#include <limits>
namespace lm {
namespace ngram {
namespace trie {
DontBhiksha::DontBhiksha(const void * /*base*/, uint64_t /*max_offset*/, uint64_t max_next, const Config &/*config*/) :
next_(util::BitsMask::ByMax(max_next)) {}
const uint8_t kArrayBhikshaVersion = 0;
// TODO: put this in binary file header instead when I change the binary file format again.
void ArrayBhiksha::UpdateConfigFromBinary(int fd, Config &config) {
uint8_t version;
uint8_t configured_bits;
util::ReadOrThrow(fd, &version, 1);
util::ReadOrThrow(fd, &configured_bits, 1);
if (version != kArrayBhikshaVersion) UTIL_THROW(FormatLoadException, "This file has sorted array compression version " << (unsigned) version << " but the code expects version " << (unsigned)kArrayBhikshaVersion);
config.pointer_bhiksha_bits = configured_bits;
}
namespace {
// Find argmin_{chopped \in [0, RequiredBits(max_next)]} ChoppedDelta(max_offset)
uint8_t ChopBits(uint64_t max_offset, uint64_t max_next, const Config &config) {
uint8_t required = util::RequiredBits(max_next);
uint8_t best_chop = 0;
int64_t lowest_change = std::numeric_limits<int64_t>::max();
// There are probably faster ways but I don't care because this is only done once per order at construction time.
for (uint8_t chop = 0; chop <= std::min(required, config.pointer_bhiksha_bits); ++chop) {
int64_t change = (max_next >> (required - chop)) * 64 /* table cost in bits */
- max_offset * static_cast<int64_t>(chop); /* savings in bits*/
if (change < lowest_change) {
lowest_change = change;
best_chop = chop;
}
}
return best_chop;
}
std::size_t ArrayCount(uint64_t max_offset, uint64_t max_next, const Config &config) {
uint8_t required = util::RequiredBits(max_next);
uint8_t chopping = ChopBits(max_offset, max_next, config);
return (max_next >> (required - chopping)) + 1 /* we store 0 too */;
}
} // namespace
std::size_t ArrayBhiksha::Size(uint64_t max_offset, uint64_t max_next, const Config &config) {
return sizeof(uint64_t) * (1 /* header */ + ArrayCount(max_offset, max_next, config)) + 7 /* 8-byte alignment */;
}
uint8_t ArrayBhiksha::InlineBits(uint64_t max_offset, uint64_t max_next, const Config &config) {
return util::RequiredBits(max_next) - ChopBits(max_offset, max_next, config);
}
namespace {
void *AlignTo8(void *from) {
uint8_t *val = reinterpret_cast<uint8_t*>(from);
std::size_t remainder = reinterpret_cast<std::size_t>(val) & 7;
if (!remainder) return val;
return val + 8 - remainder;
}
} // namespace
ArrayBhiksha::ArrayBhiksha(void *base, uint64_t max_offset, uint64_t max_next, const Config &config)
: next_inline_(util::BitsMask::ByBits(InlineBits(max_offset, max_next, config))),
offset_begin_(reinterpret_cast<const uint64_t*>(AlignTo8(base)) + 1 /* 8-byte header */),
offset_end_(offset_begin_ + ArrayCount(max_offset, max_next, config)),
write_to_(reinterpret_cast<uint64_t*>(AlignTo8(base)) + 1 /* 8-byte header */ + 1 /* first entry is 0 */),
original_base_(base) {}
void ArrayBhiksha::FinishedLoading(const Config &config) {
// *offset_begin_ = 0 but without a const_cast.
*(write_to_ - (write_to_ - offset_begin_)) = 0;
if (write_to_ != offset_end_) UTIL_THROW(util::Exception, "Did not get all the array entries that were expected.");
uint8_t *head_write = reinterpret_cast<uint8_t*>(original_base_);
*(head_write++) = kArrayBhikshaVersion;
*(head_write++) = config.pointer_bhiksha_bits;
}
void ArrayBhiksha::LoadedBinary() {
}
} // namespace trie
} // namespace ngram
} // namespace lm

115
lm/bhiksha.hh
View File

@ -1,115 +0,0 @@
/* Simple implementation of
* @inproceedings{bhikshacompression,
* author={Bhiksha Raj and Ed Whittaker},
* year={2003},
* title={Lossless Compression of Language Model Structure and Word Identifiers},
* booktitle={Proceedings of IEEE International Conference on Acoustics, Speech and Signal Processing},
* pages={388--391},
* }
*
* Currently only used for next pointers.
*/
#ifndef LM_BHIKSHA__
#define LM_BHIKSHA__
#include <stdint.h>
#include <assert.h>
#include "lm/model_type.hh"
#include "lm/trie.hh"
#include "util/bit_packing.hh"
#include "util/sorted_uniform.hh"
namespace lm {
namespace ngram {
struct Config;
namespace trie {
class DontBhiksha {
public:
static const ModelType kModelTypeAdd = static_cast<ModelType>(0);
static void UpdateConfigFromBinary(int /*fd*/, Config &/*config*/) {}
static std::size_t Size(uint64_t /*max_offset*/, uint64_t /*max_next*/, const Config &/*config*/) { return 0; }
static uint8_t InlineBits(uint64_t /*max_offset*/, uint64_t max_next, const Config &/*config*/) {
return util::RequiredBits(max_next);
}
DontBhiksha(const void *base, uint64_t max_offset, uint64_t max_next, const Config &config);
void ReadNext(const void *base, uint64_t bit_offset, uint64_t /*index*/, uint8_t total_bits, NodeRange &out) const {
out.begin = util::ReadInt57(base, bit_offset, next_.bits, next_.mask);
out.end = util::ReadInt57(base, bit_offset + total_bits, next_.bits, next_.mask);
//assert(out.end >= out.begin);
}
void WriteNext(void *base, uint64_t bit_offset, uint64_t /*index*/, uint64_t value) {
util::WriteInt57(base, bit_offset, next_.bits, value);
}
void FinishedLoading(const Config &/*config*/) {}
void LoadedBinary() {}
uint8_t InlineBits() const { return next_.bits; }
private:
util::BitsMask next_;
};
class ArrayBhiksha {
public:
static const ModelType kModelTypeAdd = kArrayAdd;
static void UpdateConfigFromBinary(int fd, Config &config);
static std::size_t Size(uint64_t max_offset, uint64_t max_next, const Config &config);
static uint8_t InlineBits(uint64_t max_offset, uint64_t max_next, const Config &config);
ArrayBhiksha(void *base, uint64_t max_offset, uint64_t max_value, const Config &config);
void ReadNext(const void *base, uint64_t bit_offset, uint64_t index, uint8_t total_bits, NodeRange &out) const {
const uint64_t *begin_it = util::BinaryBelow(util::IdentityAccessor<uint64_t>(), offset_begin_, offset_end_, index);
const uint64_t *end_it;
for (end_it = begin_it; (end_it < offset_end_) && (*end_it <= index + 1); ++end_it) {}
--end_it;
out.begin = ((begin_it - offset_begin_) << next_inline_.bits) |
util::ReadInt57(base, bit_offset, next_inline_.bits, next_inline_.mask);
out.end = ((end_it - offset_begin_) << next_inline_.bits) |
util::ReadInt57(base, bit_offset + total_bits, next_inline_.bits, next_inline_.mask);
//assert(out.end >= out.begin);
}
void WriteNext(void *base, uint64_t bit_offset, uint64_t index, uint64_t value) {
uint64_t encode = value >> next_inline_.bits;
for (; write_to_ <= offset_begin_ + encode; ++write_to_) *write_to_ = index;
util::WriteInt57(base, bit_offset, next_inline_.bits, value & next_inline_.mask);
}
void FinishedLoading(const Config &config);
void LoadedBinary();
uint8_t InlineBits() const { return next_inline_.bits; }
private:
const util::BitsMask next_inline_;
const uint64_t *const offset_begin_;
const uint64_t *const offset_end_;
uint64_t *write_to_;
void *original_base_;
};
} // namespace trie
} // namespace ngram
} // namespace lm
#endif // LM_BHIKSHA__

241
lm/binary_format.cc
View File

@ -1,241 +0,0 @@
#include "lm/binary_format.hh"
#include "lm/lm_exception.hh"
#include "util/file.hh"
#include "util/file_piece.hh"
#include <cstddef>
#include <cstring>
#include <limits>
#include <string>
#include <stdint.h>
namespace lm {
namespace ngram {
namespace {
const char kMagicBeforeVersion[] = "mmap lm http://kheafield.com/code format version";
const char kMagicBytes[] = "mmap lm http://kheafield.com/code format version 5\n\0";
// This must be shorter than kMagicBytes and indicates an incomplete binary file (i.e. build failed).
const char kMagicIncomplete[] = "mmap lm http://kheafield.com/code incomplete\n";
const long int kMagicVersion = 5;
// Old binary files built on 32-bit machines have this header.
// TODO: eliminate with next binary release.
struct OldSanity {
char magic[sizeof(kMagicBytes)];
float zero_f, one_f, minus_half_f;
WordIndex one_word_index, max_word_index;
uint64_t one_uint64;
void SetToReference() {
std::memset(this, 0, sizeof(OldSanity));
std::memcpy(magic, kMagicBytes, sizeof(magic));
zero_f = 0.0; one_f = 1.0; minus_half_f = -0.5;
one_word_index = 1;
max_word_index = std::numeric_limits<WordIndex>::max();
one_uint64 = 1;
}
};
// Test values aligned to 8 bytes.
struct Sanity {
char magic[ALIGN8(sizeof(kMagicBytes))];
float zero_f, one_f, minus_half_f;
WordIndex one_word_index, max_word_index, padding_to_8;
uint64_t one_uint64;
void SetToReference() {
std::memset(this, 0, sizeof(Sanity));
std::memcpy(magic, kMagicBytes, sizeof(kMagicBytes));
zero_f = 0.0; one_f = 1.0; minus_half_f = -0.5;
one_word_index = 1;
max_word_index = std::numeric_limits<WordIndex>::max();
padding_to_8 = 0;
one_uint64 = 1;
}
};
const char *kModelNames[6] = {"probing hash tables", "probing hash tables with rest costs", "trie", "trie with quantization", "trie with array-compressed pointers", "trie with quantization and array-compressed pointers"};
std::size_t TotalHeaderSize(unsigned char order) {
return ALIGN8(sizeof(Sanity) + sizeof(FixedWidthParameters) + sizeof(uint64_t) * order);
}
void WriteHeader(void *to, const Parameters &params) {
Sanity header = Sanity();
header.SetToReference();
std::memcpy(to, &header, sizeof(Sanity));
char *out = reinterpret_cast<char*>(to) + sizeof(Sanity);
*reinterpret_cast<FixedWidthParameters*>(out) = params.fixed;
out += sizeof(FixedWidthParameters);
uint64_t *counts = reinterpret_cast<uint64_t*>(out);
for (std::size_t i = 0; i < params.counts.size(); ++i) {
counts[i] = params.counts[i];
}
}
} // namespace
uint8_t *SetupJustVocab(const Config &config, uint8_t order, std::size_t memory_size, Backing &backing) {
if (config.write_mmap) {
std::size_t total = TotalHeaderSize(order) + memory_size;
backing.vocab.reset(util::MapZeroedWrite(config.write_mmap, total, backing.file), total, util::scoped_memory::MMAP_ALLOCATED);
strncpy(reinterpret_cast<char*>(backing.vocab.get()), kMagicIncomplete, TotalHeaderSize(order));
return reinterpret_cast<uint8_t*>(backing.vocab.get()) + TotalHeaderSize(order);
} else {
util::MapAnonymous(memory_size, backing.vocab);
return reinterpret_cast<uint8_t*>(backing.vocab.get());
}
}
uint8_t *GrowForSearch(const Config &config, std::size_t vocab_pad, std::size_t memory_size, Backing &backing) {
std::size_t adjusted_vocab = backing.vocab.size() + vocab_pad;
if (config.write_mmap) {
// Grow the file to accomodate the search, using zeros.
try {
util::ResizeOrThrow(backing.file.get(), adjusted_vocab + memory_size);
} catch (util::ErrnoException &e) {
e << " for file " << config.write_mmap;
throw e;
}
if (config.write_method == Config::WRITE_AFTER) {
util::MapAnonymous(memory_size, backing.search);
return reinterpret_cast<uint8_t*>(backing.search.get());
}
// mmap it now.
// We're skipping over the header and vocab for the search space mmap. mmap likes page aligned offsets, so some arithmetic to round the offset down.
std::size_t page_size = util::SizePage();
std::size_t alignment_cruft = adjusted_vocab % page_size;
backing.search.reset(util::MapOrThrow(alignment_cruft + memory_size, true, util::kFileFlags, false, backing.file.get(), adjusted_vocab - alignment_cruft), alignment_cruft + memory_size, util::scoped_memory::MMAP_ALLOCATED);
return reinterpret_cast<uint8_t*>(backing.search.get()) + alignment_cruft;
} else {
util::MapAnonymous(memory_size, backing.search);
return reinterpret_cast<uint8_t*>(backing.search.get());
}
}
void FinishFile(const Config &config, ModelType model_type, unsigned int search_version, const std::vector<uint64_t> &counts, std::size_t vocab_pad, Backing &backing) {
if (!config.write_mmap) return;
util::SyncOrThrow(backing.vocab.get(), backing.vocab.size());
switch (config.write_method) {
case Config::WRITE_MMAP:
util::SyncOrThrow(backing.search.get(), backing.search.size());
break;
case Config::WRITE_AFTER:
util::SeekOrThrow(backing.file.get(), backing.vocab.size() + vocab_pad);
util::WriteOrThrow(backing.file.get(), backing.search.get(), backing.search.size());
util::FSyncOrThrow(backing.file.get());
break;
}
// header and vocab share the same mmap. The header is written here because we know the counts.
Parameters params = Parameters();
params.counts = counts;
params.fixed.order = counts.size();
params.fixed.probing_multiplier = config.probing_multiplier;
params.fixed.model_type = model_type;
params.fixed.has_vocabulary = config.include_vocab;
params.fixed.search_version = search_version;
WriteHeader(backing.vocab.get(), params);
}
namespace detail {
bool IsBinaryFormat(int fd) {
const uint64_t size = util::SizeFile(fd);
if (size == util::kBadSize || (size <= static_cast<uint64_t>(sizeof(Sanity)))) return false;
// Try reading the header.
util::scoped_memory memory;
try {
util::MapRead(util::LAZY, fd, 0, sizeof(Sanity), memory);
} catch (const util::Exception &e) {
return false;
}
Sanity reference_header = Sanity();
reference_header.SetToReference();
if (!memcmp(memory.get(), &reference_header, sizeof(Sanity))) return true;
if (!memcmp(memory.get(), kMagicIncomplete, strlen(kMagicIncomplete))) {
UTIL_THROW(FormatLoadException, "This binary file did not finish building");
}
if (!memcmp(memory.get(), kMagicBeforeVersion, strlen(kMagicBeforeVersion))) {
char *end_ptr;
const char *begin_version = static_cast<const char*>(memory.get()) + strlen(kMagicBeforeVersion);
long int version = strtol(begin_version, &end_ptr, 10);
if ((end_ptr != begin_version) && version != kMagicVersion) {
UTIL_THROW(FormatLoadException, "Binary file has version " << version << " but this implementation expects version " << kMagicVersion << " so you'll have to use the ARPA to rebuild your binary");
}
OldSanity old_sanity = OldSanity();
old_sanity.SetToReference();
UTIL_THROW_IF(!memcmp(memory.get(), &old_sanity, sizeof(OldSanity)), FormatLoadException, "Looks like this is an old 32-bit format. The old 32-bit format has been removed so that 64-bit and 32-bit files are exchangeable.");
UTIL_THROW(FormatLoadException, "File looks like it should be loaded with mmap, but the test values don't match. Try rebuilding the binary format LM using the same code revision, compiler, and architecture");
}
return false;
}
void ReadHeader(int fd, Parameters &out) {
util::SeekOrThrow(fd, sizeof(Sanity));
util::ReadOrThrow(fd, &out.fixed, sizeof(out.fixed));
if (out.fixed.probing_multiplier < 1.0)
UTIL_THROW(FormatLoadException, "Binary format claims to have a probing multiplier of " << out.fixed.probing_multiplier << " which is < 1.0.");
out.counts.resize(static_cast<std::size_t>(out.fixed.order));
if (out.fixed.order) util::ReadOrThrow(fd, &*out.counts.begin(), sizeof(uint64_t) * out.fixed.order);
}
void MatchCheck(ModelType model_type, unsigned int search_version, const Parameters &params) {
if (params.fixed.model_type != model_type) {
if (static_cast<unsigned int>(params.fixed.model_type) >= (sizeof(kModelNames) / sizeof(const char *)))
UTIL_THROW(FormatLoadException, "The binary file claims to be model type " << static_cast<unsigned int>(params.fixed.model_type) << " but this is not implemented for in this inference code.");
UTIL_THROW(FormatLoadException, "The binary file was built for " << kModelNames[params.fixed.model_type] << " but the inference code is trying to load " << kModelNames[model_type]);
}
UTIL_THROW_IF(search_version != params.fixed.search_version, FormatLoadException, "The binary file has " << kModelNames[params.fixed.model_type] << " version " << params.fixed.search_version << " but this code expects " << kModelNames[params.fixed.model_type] << " version " << search_version);
}
void SeekPastHeader(int fd, const Parameters &params) {
util::SeekOrThrow(fd, TotalHeaderSize(params.counts.size()));
}
uint8_t *SetupBinary(const Config &config, const Parameters &params, std::size_t memory_size, Backing &backing) {
const uint64_t file_size = util::SizeFile(backing.file.get());
// The header is smaller than a page, so we have to map the whole header as well.
std::size_t total_map = TotalHeaderSize(params.counts.size()) + memory_size;
if (file_size != util::kBadSize && static_cast<uint64_t>(file_size) < total_map)
UTIL_THROW(FormatLoadException, "Binary file has size " << file_size << " but the headers say it should be at least " << total_map);
util::MapRead(config.load_method, backing.file.get(), 0, total_map, backing.search);
if (config.enumerate_vocab && !params.fixed.has_vocabulary)
UTIL_THROW(FormatLoadException, "The decoder requested all the vocabulary strings, but this binary file does not have them. You may need to rebuild the binary file with an updated version of build_binary.");
// Seek to vocabulary words
util::SeekOrThrow(backing.file.get(), total_map);
return reinterpret_cast<uint8_t*>(backing.search.get()) + TotalHeaderSize(params.counts.size());
}
void ComplainAboutARPA(const Config &config, ModelType model_type) {
if (config.write_mmap || !config.messages) return;
if (config.arpa_complain == Config::ALL) {
*config.messages << "Loading the LM will be faster if you build a binary file." << std::endl;
} else if (config.arpa_complain == Config::EXPENSIVE && model_type == TRIE_SORTED) {
*config.messages << "Building " << kModelNames[model_type] << " from ARPA is expensive. Save time by building a binary format." << std::endl;
}
}
} // namespace detail
bool RecognizeBinary(const char *file, ModelType &recognized) {
util::scoped_fd fd(util::OpenReadOrThrow(file));
if (!detail::IsBinaryFormat(fd.get())) return false;
Parameters params;
detail::ReadHeader(fd.get(), params);
recognized = params.fixed.model_type;
return true;
}
} // namespace ngram
} // namespace lm

108
lm/binary_format.hh
View File

@ -1,108 +0,0 @@
#ifndef LM_BINARY_FORMAT__
#define LM_BINARY_FORMAT__
#include "lm/config.hh"
#include "lm/model_type.hh"
#include "lm/read_arpa.hh"
#include "util/file_piece.hh"
#include "util/mmap.hh"
#include "util/scoped.hh"
#include <cstddef>
#include <vector>
#include <stdint.h>
namespace lm {
namespace ngram {
/*Inspect a file to determine if it is a binary lm. If not, return false.
* If so, return true and set recognized to the type. This is the only API in
* this header designed for use by decoder authors.
*/
bool RecognizeBinary(const char *file, ModelType &recognized);
struct FixedWidthParameters {
unsigned char order;
float probing_multiplier;
// What type of model is this?
ModelType model_type;
// Does the end of the file have the actual strings in the vocabulary?
bool has_vocabulary;
unsigned int search_version;
};
// This is a macro instead of an inline function so constants can be assigned using it.
#define ALIGN8(a) ((std::ptrdiff_t(((a)-1)/8)+1)*8)
// Parameters stored in the header of a binary file.
struct Parameters {
FixedWidthParameters fixed;
std::vector<uint64_t> counts;
};
struct Backing {
// File behind memory, if any.
util::scoped_fd file;
// Vocabulary lookup table. Not to be confused with the vocab words themselves.
util::scoped_memory vocab;
// Raw block of memory backing the language model data structures
util::scoped_memory search;
};
// Create just enough of a binary file to write vocabulary to it.
uint8_t *SetupJustVocab(const Config &config, uint8_t order, std::size_t memory_size, Backing &backing);
// Grow the binary file for the search data structure and set backing.search, returning the memory address where the search data structure should begin.
uint8_t *GrowForSearch(const Config &config, std::size_t vocab_pad, std::size_t memory_size, Backing &backing);
// Write header to binary file. This is done last to prevent incomplete files
// from loading.
void FinishFile(const Config &config, ModelType model_type, unsigned int search_version, const std::vector<uint64_t> &counts, std::size_t vocab_pad, Backing &backing);
namespace detail {
bool IsBinaryFormat(int fd);
void ReadHeader(int fd, Parameters &params);
void MatchCheck(ModelType model_type, unsigned int search_version, const Parameters &params);
void SeekPastHeader(int fd, const Parameters &params);
uint8_t *SetupBinary(const Config &config, const Parameters &params, std::size_t memory_size, Backing &backing);
void ComplainAboutARPA(const Config &config, ModelType model_type);
} // namespace detail
template <class To> void LoadLM(const char *file, const Config &config, To &to) {
Backing &backing = to.MutableBacking();
backing.file.reset(util::OpenReadOrThrow(file));
try {
if (detail::IsBinaryFormat(backing.file.get())) {
Parameters params;
detail::ReadHeader(backing.file.get(), params);
detail::MatchCheck(To::kModelType, To::kVersion, params);
// Replace the run-time configured probing_multiplier with the one in the file.
Config new_config(config);
new_config.probing_multiplier = params.fixed.probing_multiplier;
detail::SeekPastHeader(backing.file.get(), params);
To::UpdateConfigFromBinary(backing.file.get(), params.counts, new_config);
std::size_t memory_size = To::Size(params.counts, new_config);
uint8_t *start = detail::SetupBinary(new_config, params, memory_size, backing);
to.InitializeFromBinary(start, params, new_config, backing.file.get());
} else {
detail::ComplainAboutARPA(config, To::kModelType);
to.InitializeFromARPA(file, config);
}
} catch (util::Exception &e) {
e << " File: " << file;
throw;
}
}
} // namespace ngram
} // namespace lm
#endif // LM_BINARY_FORMAT__

43
lm/blank.hh
View File

@ -1,43 +0,0 @@
#ifndef LM_BLANK__
#define LM_BLANK__
#include <limits>
#include <stdint.h>
#include <math.h>
namespace lm {
namespace ngram {
/* Suppose "foo bar" appears with zero backoff but there is no trigram
* beginning with these words. Then, when scoring "foo bar", the model could
* return out_state containing "bar" or even null context if "bar" also has no
* backoff and is never followed by another word. Then the backoff is set to
* kNoExtensionBackoff. If the n-gram might be extended, then out_state must
* contain the full n-gram, in which case kExtensionBackoff is set. In any
* case, if an n-gram has non-zero backoff, the full state is returned so
* backoff can be properly charged.
* These differ only in sign bit because the backoff is in fact zero in either
* case.
*/
const float kNoExtensionBackoff = -0.0;
const float kExtensionBackoff = 0.0;
const uint64_t kNoExtensionQuant = 0;
const uint64_t kExtensionQuant = 1;
inline void SetExtension(float &backoff) {
if (backoff == kNoExtensionBackoff) backoff = kExtensionBackoff;
}
// This compiles down nicely.
inline bool HasExtension(const float &backoff) {
typedef union { float f; uint32_t i; } UnionValue;
UnionValue compare, interpret;
compare.f = kNoExtensionBackoff;
interpret.f = backoff;
return compare.i != interpret.i;
}
} // namespace ngram
} // namespace lm
#endif // LM_BLANK__

253
lm/build_binary.cc
View File

@ -1,253 +0,0 @@
#include "lm/model.hh"
#include "util/file_piece.hh"
#include <cstdlib>
#include <exception>
#include <iostream>
#include <iomanip>
#include <math.h>
#include <stdlib.h>
#ifdef WIN32
#include "util/getopt.hh"
#endif
namespace lm {
namespace ngram {
namespace {
void Usage(const char *name) {
std::cerr << "Usage: " << name << " [-u log10_unknown_probability] [-s] [-i] [-w mmap|after] [-p probing_multiplier] [-t trie_temporary] [-m trie_building_megabytes] [-q bits] [-b bits] [-a bits] [type] input.arpa [output.mmap]\n\n"
"-u sets the log10 probability for <unk> if the ARPA file does not have one.\n"
" Default is -100. The ARPA file will always take precedence.\n"
"-s allows models to be built even if they do not have <s> and </s>.\n"
"-i allows buggy models from IRSTLM by mapping positive log probability to 0.\n"
"-w mmap|after determines how writing is done.\n"
" mmap maps the binary file and writes to it. Default for trie.\n"
" after allocates anonymous memory, builds, and writes. Default for probing.\n"
"-r \"order1.arpa order2 order3 order4\" adds lower-order rest costs from these\n"
" model files. order1.arpa must be an ARPA file. All others may be ARPA or\n"
" the same data structure as being built. All files must have the same\n"
" vocabulary. For probing, the unigrams must be in the same order.\n\n"
"type is either probing or trie. Default is probing.\n\n"
"probing uses a probing hash table. It is the fastest but uses the most memory.\n"
"-p sets the space multiplier and must be >1.0. The default is 1.5.\n\n"
"trie is a straightforward trie with bit-level packing. It uses the least\n"
"memory and is still faster than SRI or IRST. Building the trie format uses an\n"
"on-disk sort to save memory.\n"
"-t is the temporary directory prefix. Default is the output file name.\n"
"-m limits memory use for sorting. Measured in MB. Default is 1024MB.\n"
"-q turns quantization on and sets the number of bits (e.g. -q 8).\n"
"-b sets backoff quantization bits. Requires -q and defaults to that value.\n"
"-a compresses pointers using an array of offsets. The parameter is the\n"
" maximum number of bits encoded by the array. Memory is minimized subject\n"
" to the maximum, so pick 255 to minimize memory.\n\n"
"Get a memory estimate by passing an ARPA file without an output file name.\n";
exit(1);
}
// I could really use boost::lexical_cast right about now.
float ParseFloat(const char *from) {
char *end;
float ret = strtod(from, &end);
if (*end) throw util::ParseNumberException(from);
return ret;
}
unsigned long int ParseUInt(const char *from) {
char *end;
unsigned long int ret = strtoul(from, &end, 10);
if (*end) throw util::ParseNumberException(from);
return ret;
}
uint8_t ParseBitCount(const char *from) {
unsigned long val = ParseUInt(from);
if (val > 25) {
util::ParseNumberException e(from);
e << " bit counts are limited to 25.";
}
return val;
}
void ParseFileList(const char *from, std::vector<std::string> &to) {
to.clear();
while (true) {
const char *i;
for (i = from; *i && *i != ' '; ++i) {}
to.push_back(std::string(from, i - from));
if (!*i) break;
from = i + 1;
}
}
void ShowSizes(const char *file, const lm::ngram::Config &config) {
std::vector<uint64_t> counts;
util::FilePiece f(file);
lm::ReadARPACounts(f, counts);
std::size_t sizes[6];
sizes[0] = ProbingModel::Size(counts, config);
sizes[1] = RestProbingModel::Size(counts, config);
sizes[2] = TrieModel::Size(counts, config);
sizes[3] = QuantTrieModel::Size(counts, config);
sizes[4] = ArrayTrieModel::Size(counts, config);
sizes[5] = QuantArrayTrieModel::Size(counts, config);
std::size_t max_length = *std::max_element(sizes, sizes + sizeof(sizes) / sizeof(size_t));
std::size_t min_length = *std::min_element(sizes, sizes + sizeof(sizes) / sizeof(size_t));
std::size_t divide;
char prefix;
if (min_length < (1 << 10) * 10) {
prefix = ' ';
divide = 1;
} else if (min_length < (1 << 20) * 10) {
prefix = 'k';
divide = 1 << 10;
} else if (min_length < (1ULL << 30) * 10) {
prefix = 'M';
divide = 1 << 20;
} else {
prefix = 'G';
divide = 1 << 30;
}
long int length = std::max<long int>(2, static_cast<long int>(ceil(log10((double) max_length / divide))));
std::cout << "Memory estimate:\ntype ";
// right align bytes.
for (long int i = 0; i < length - 2; ++i) std::cout << ' ';
std::cout << prefix << "B\n"
"probing " << std::setw(length) << (sizes[0] / divide) << " assuming -p " << config.probing_multiplier << "\n"
"probing " << std::setw(length) << (sizes[1] / divide) << " assuming -r models -p " << config.probing_multiplier << "\n"
"trie " << std::setw(length) << (sizes[2] / divide) << " without quantization\n"
"trie " << std::setw(length) << (sizes[3] / divide) << " assuming -q " << (unsigned)config.prob_bits << " -b " << (unsigned)config.backoff_bits << " quantization \n"
"trie " << std::setw(length) << (sizes[4] / divide) << " assuming -a " << (unsigned)config.pointer_bhiksha_bits << " array pointer compression\n"
"trie " << std::setw(length) << (sizes[5] / divide) << " assuming -a " << (unsigned)config.pointer_bhiksha_bits << " -q " << (unsigned)config.prob_bits << " -b " << (unsigned)config.backoff_bits<< " array pointer compression and quantization\n";
}
void ProbingQuantizationUnsupported() {
std::cerr << "Quantization is only implemented in the trie data structure." << std::endl;
exit(1);
}
} // namespace ngram
} // namespace lm
} // namespace
int main(int argc, char *argv[]) {
using namespace lm::ngram;
try {
bool quantize = false, set_backoff_bits = false, bhiksha = false, set_write_method = false, rest = false;
lm::ngram::Config config;
int opt;
while ((opt = getopt(argc, argv, "q:b:a:u:p:t:m:w:sir:")) != -1) {
switch(opt) {
case 'q':
config.prob_bits = ParseBitCount(optarg);
if (!set_backoff_bits) config.backoff_bits = config.prob_bits;
quantize = true;
break;
case 'b':
config.backoff_bits = ParseBitCount(optarg);
set_backoff_bits = true;
break;
case 'a':
config.pointer_bhiksha_bits = ParseBitCount(optarg);
bhiksha = true;
break;
case 'u':
config.unknown_missing_logprob = ParseFloat(optarg);
break;
case 'p':
config.probing_multiplier = ParseFloat(optarg);
break;
case 't':
config.temporary_directory_prefix = optarg;
break;
case 'm':
config.building_memory = ParseUInt(optarg) * 1048576;
break;
case 'w':
set_write_method = true;
if (!strcmp(optarg, "mmap")) {
config.write_method = Config::WRITE_MMAP;
} else if (!strcmp(optarg, "after")) {
config.write_method = Config::WRITE_AFTER;
} else {
Usage(argv[0]);
}
break;
case 's':
config.sentence_marker_missing = lm::SILENT;
break;
case 'i':
config.positive_log_probability = lm::SILENT;
break;
case 'r':
rest = true;
ParseFileList(optarg, config.rest_lower_files);
config.rest_function = Config::REST_LOWER;
break;
default:
Usage(argv[0]);
}
}
if (!quantize && set_backoff_bits) {
std::cerr << "You specified backoff quantization (-b) but not probability quantization (-q)" << std::endl;
abort();
}
if (optind + 1 == argc) {
ShowSizes(argv[optind], config);
return 0;
}
const char *model_type;
const char *from_file;
if (optind + 2 == argc) {
model_type = "probing";
from_file = argv[optind];
config.write_mmap = argv[optind + 1];
} else if (optind + 3 == argc) {
model_type = argv[optind];
from_file = argv[optind + 1];
config.write_mmap = argv[optind + 2];
} else {
Usage(argv[0]);
}
if (!strcmp(model_type, "probing")) {
if (!set_write_method) config.write_method = Config::WRITE_AFTER;
if (quantize || set_backoff_bits) ProbingQuantizationUnsupported();
if (rest) {
RestProbingModel(from_file, config);
} else {
ProbingModel(from_file, config);
}
} else if (!strcmp(model_type, "trie")) {
if (rest) {
std::cerr << "Rest + trie is not supported yet." << std::endl;
return 1;
}
if (!set_write_method) config.write_method = Config::WRITE_MMAP;
if (quantize) {
if (bhiksha) {
QuantArrayTrieModel(from_file, config);
} else {
QuantTrieModel(from_file, config);
}
} else {
if (bhiksha) {
ArrayTrieModel(from_file, config);
} else {
TrieModel(from_file, config);
}
}
} else {
Usage(argv[0]);
}
}
catch (const std::exception &e) {
std::cerr << e.what() << std::endl;
std::cerr << "ERROR" << std::endl;
return 1;
}
std::cerr << "SUCCESS" << std::endl;
return 0;
}

3
lm/clean.sh
View File

@ -1,3 +0,0 @@
#!/bin/bash
cd "$(dirname "$0")/.."
rm -rf {lm,util}/*.o lm/query lm/build_binary {lm,util}/*_test lm/test.binary* lm/test.arpa?????? util/file_piece.cc.gz

16
lm/compile.sh
View File

@ -1,16 +0,0 @@
#!/bin/bash
#This is just an example compilation. You should integrate these files into your build system. I can provide boost jam if you want.
#If your code uses ICU, edit util/string_piece.hh and uncomment #define USE_ICU
#I use zlib by default. If you don't want to depend on zlib, remove #define USE_ZLIB from util/file_piece.hh
#don't need to use if compiling with moses Makefiles already
cd "$(dirname "$0")/.."
set -e
for i in util/{bit_packing,ersatz_progress,exception,file_piece,murmur_hash,file,mmap} lm/{bhiksha,binary_format,config,lm_exception,model,quantize,read_arpa,search_hashed,search_trie,trie,trie_sort,virtual_interface,vocab}; do
g++ -I. -O3 -DNDEBUG $CXXFLAGS -c $i.cc -o $i.o
done
g++ -I. -O3 -DNDEBUG $CXXFLAGS lm/build_binary.cc {lm,util}/*.o -lz -o lm/build_binary
g++ -I. -O3 -DNDEBUG $CXXFLAGS lm/ngram_query.cc {lm,util}/*.o -lz -o lm/query

29
lm/config.cc
View File

@ -1,29 +0,0 @@
#include "lm/config.hh"
#include <iostream>
namespace lm {
namespace ngram {
Config::Config() :
messages(&std::cerr),
enumerate_vocab(NULL),
unknown_missing(COMPLAIN),
sentence_marker_missing(THROW_UP),
positive_log_probability(THROW_UP),
unknown_missing_logprob(-100.0),
probing_multiplier(1.5),
building_memory(1073741824ULL), // 1 GB
temporary_directory_prefix(NULL),
arpa_complain(ALL),
write_mmap(NULL),
write_method(WRITE_AFTER),
include_vocab(true),
rest_function(REST_MAX),
prob_bits(8),
backoff_bits(8),
pointer_bhiksha_bits(22),
load_method(util::POPULATE_OR_READ) {}
} // namespace ngram
} // namespace lm

120
lm/config.hh
View File

@ -1,120 +0,0 @@
#ifndef LM_CONFIG__
#define LM_CONFIG__
#include "lm/lm_exception.hh"
#include "util/mmap.hh"
#include <iosfwd>
#include <string>
#include <vector>
/* Configuration for ngram model. Separate header to reduce pollution. */
namespace lm {
class EnumerateVocab;
namespace ngram {
struct Config {
// EFFECTIVE FOR BOTH ARPA AND BINARY READS
// Where to log messages including the progress bar. Set to NULL for
// silence.
std::ostream *messages;
// This will be called with every string in the vocabulary. See
// enumerate_vocab.hh for more detail. Config does not take ownership; you
// are still responsible for deleting it (or stack allocating).
EnumerateVocab *enumerate_vocab;
// ONLY EFFECTIVE WHEN READING ARPA
// What to do when <unk> isn't in the provided model.
WarningAction unknown_missing;
// What to do when <s> or </s> is missing from the model.
// If THROW_UP, the exception will be of type util::SpecialWordMissingException.
WarningAction sentence_marker_missing;
// What to do with a positive log probability. For COMPLAIN and SILENT, map
// to 0.
WarningAction positive_log_probability;
// The probability to substitute for <unk> if it's missing from the model.
// No effect if the model has <unk> or unknown_missing == THROW_UP.
float unknown_missing_logprob;
// Size multiplier for probing hash table. Must be > 1. Space is linear in
// this. Time is probing_multiplier / (probing_multiplier - 1). No effect
// for sorted variant.
// If you find yourself setting this to a low number, consider using the
// TrieModel which has lower memory consumption.
float probing_multiplier;
// Amount of memory to use for building. The actual memory usage will be
// higher since this just sets sort buffer size. Only applies to trie
// models.
std::size_t building_memory;
// Template for temporary directory appropriate for passing to mkdtemp.
// The characters XXXXXX are appended before passing to mkdtemp. Only
// applies to trie. If NULL, defaults to write_mmap. If that's NULL,
// defaults to input file name.
const char *temporary_directory_prefix;
// Level of complaining to do when loading from ARPA instead of binary format.
enum ARPALoadComplain {ALL, EXPENSIVE, NONE};
ARPALoadComplain arpa_complain;
// While loading an ARPA file, also write out this binary format file. Set
// to NULL to disable.
const char *write_mmap;
enum WriteMethod {
WRITE_MMAP, // Map the file directly.
WRITE_AFTER // Write after we're done.
};
WriteMethod write_method;
// Include the vocab in the binary file? Only effective if write_mmap != NULL.
bool include_vocab;
// Left rest options. Only used when the model includes rest costs.
enum RestFunction {
REST_MAX, // Maximum of any score to the left
REST_LOWER, // Use lower-order files given below.
};
RestFunction rest_function;
// Only used for REST_LOWER.
std::vector<std::string> rest_lower_files;
// Quantization options. Only effective for QuantTrieModel. One value is
// reserved for each of prob and backoff, so 2^bits - 1 buckets will be used
// to quantize (and one of the remaining backoffs will be 0).
uint8_t prob_bits, backoff_bits;
// Bhiksha compression (simple form). Only works with trie.
uint8_t pointer_bhiksha_bits;
// ONLY EFFECTIVE WHEN READING BINARY
// How to get the giant array into memory: lazy mmap, populate, read etc.
// See util/mmap.hh for details of MapMethod.
util::LoadMethod load_method;
// Set defaults.
Config();
};
} /* namespace ngram */ } /* namespace lm */
#endif // LM_CONFIG__

28
lm/enumerate_vocab.hh
View File

@ -1,28 +0,0 @@
#ifndef LM_ENUMERATE_VOCAB__
#define LM_ENUMERATE_VOCAB__
#include "lm/word_index.hh"
#include "util/string_piece.hh"
namespace lm {
/* If you need the actual strings in the vocabulary, inherit from this class
* and implement Add. Then put a pointer in Config.enumerate_vocab; it does
* not take ownership. Add is called once per vocab word. index starts at 0
* and increases by 1 each time. This is only used by the Model constructor;
* the pointer is not retained by the class.
*/
class EnumerateVocab {
public:
virtual ~EnumerateVocab() {}
virtual void Add(WordIndex index, const StringPiece &str) = 0;
protected:
EnumerateVocab() {}
};
} // namespace lm
#endif // LM_ENUMERATE_VOCAB__

64
lm/facade.hh
View File

@ -1,64 +0,0 @@
#ifndef LM_FACADE__
#define LM_FACADE__
#include "lm/virtual_interface.hh"
#include "util/string_piece.hh"
#include <string>
namespace lm {
namespace base {
// Common model interface that depends on knowing the specific classes.
// Curiously recurring template pattern.
template <class Child, class StateT, class VocabularyT> class ModelFacade : public Model {
public:
typedef StateT State;
typedef VocabularyT Vocabulary;
// Default Score function calls FullScore. Model can override this.
float Score(const State &in_state, const WordIndex new_word, State &out_state) const {
return static_cast<const Child*>(this)->FullScore(in_state, new_word, out_state).prob;
}
/* Translate from void* to State */
FullScoreReturn FullScore(const void *in_state, const WordIndex new_word, void *out_state) const {
return static_cast<const Child*>(this)->FullScore(
*reinterpret_cast<const State*>(in_state),
new_word,
*reinterpret_cast<State*>(out_state));
}
float Score(const void *in_state, const WordIndex new_word, void *out_state) const {
return static_cast<const Child*>(this)->Score(
*reinterpret_cast<const State*>(in_state),
new_word,
*reinterpret_cast<State*>(out_state));
}
const State &BeginSentenceState() const { return begin_sentence_; }
const State &NullContextState() const { return null_context_; }
const Vocabulary &GetVocabulary() const { return *static_cast<const Vocabulary*>(&BaseVocabulary()); }
protected:
ModelFacade() : Model(sizeof(State)) {}
virtual ~ModelFacade() {}
// begin_sentence and null_context can disappear after. vocab should stay.
void Init(const State &begin_sentence, const State &null_context, const Vocabulary &vocab, unsigned char order) {
begin_sentence_ = begin_sentence;
null_context_ = null_context;
begin_sentence_memory_ = &begin_sentence_;
null_context_memory_ = &null_context_;
base_vocab_ = &vocab;
order_ = order;
}
private:
State begin_sentence_, null_context_;
};
} // mamespace base
} // namespace lm
#endif // LM_FACADE__

211
lm/left.hh
View File

@ -1,211 +0,0 @@
/* Efficient left and right language model state for sentence fragments.
* Intended usage:
* Store ChartState with every chart entry.
* To do a rule application:
* 1. Make a ChartState object for your new entry.
* 2. Construct RuleScore.
* 3. Going from left to right, call Terminal or NonTerminal.
* For terminals, just pass the vocab id.
* For non-terminals, pass that non-terminal's ChartState.
* If your decoder expects scores inclusive of subtree scores (i.e. you
* label entries with the highest-scoring path), pass the non-terminal's
* score as prob.
* If your decoder expects relative scores and will walk the chart later,
* pass prob = 0.0.
* In other words, the only effect of prob is that it gets added to the
* returned log probability.
* 4. Call Finish. It returns the log probability.
*
* There's a couple more details:
* Do not pass <s> to Terminal as it is formally not a word in the sentence,
* only context. Instead, call BeginSentence. If called, it should be the
* first call after RuleScore is constructed (since <s> is always the
* leftmost).
*
* If the leftmost RHS is a non-terminal, it's faster to call BeginNonTerminal.
*
* Hashing and sorting comparison operators are provided. All state objects
* are POD. If you intend to use memcmp on raw state objects, you must call
* ZeroRemaining first, as the value of array entries beyond length is
* otherwise undefined.
*
* Usage is of course not limited to chart decoding. Anything that generates
* sentence fragments missing left context could benefit. For example, a
* phrase-based decoder could pre-score phrases, storing ChartState with each
* phrase, even if hypotheses are generated left-to-right.
*/
#ifndef LM_LEFT__
#define LM_LEFT__
#include "lm/state.hh"
#include "lm/return.hh"
#include "util/murmur_hash.hh"
#include <algorithm>
namespace lm {
namespace ngram {
template <class M> class RuleScore {
public:
explicit RuleScore(const M &model, ChartState &out) : model_(model), out_(out), left_done_(false), prob_(0.0) {
out.left.length = 0;
out.right.length = 0;
}
void BeginSentence() {
out_.right = model_.BeginSentenceState();
// out_.left is empty.
left_done_ = true;
}
void Terminal(WordIndex word) {
State copy(out_.right);
FullScoreReturn ret(model_.FullScore(copy, word, out_.right));
if (left_done_) { prob_ += ret.prob; return; }
if (ret.independent_left) {
prob_ += ret.prob;
left_done_ = true;
return;
}
out_.left.pointers[out_.left.length++] = ret.extend_left;
prob_ += ret.rest;
if (out_.right.length != copy.length + 1)
left_done_ = true;
}
// Faster version of NonTerminal for the case where the rule begins with a non-terminal.
void BeginNonTerminal(const ChartState &in, float prob = 0.0) {
prob_ = prob;
out_ = in;
left_done_ = in.left.full;
}
void NonTerminal(const ChartState &in, float prob = 0.0) {
prob_ += prob;
if (!in.left.length) {
if (in.left.full) {
for (const float *i = out_.right.backoff; i < out_.right.backoff + out_.right.length; ++i) prob_ += *i;
left_done_ = true;
out_.right = in.right;
}
return;
}
if (!out_.right.length) {
out_.right = in.right;
if (left_done_) {
prob_ += model_.UnRest(in.left.pointers, in.left.pointers + in.left.length, 1);
return;
}
if (out_.left.length) {
left_done_ = true;
} else {
out_.left = in.left;
left_done_ = in.left.full;
}
return;
}
float backoffs[KENLM_MAX_ORDER - 1], backoffs2[KENLM_MAX_ORDER - 1];
float *back = backoffs, *back2 = backoffs2;
unsigned char next_use = out_.right.length;
// First word
if (ExtendLeft(in, next_use, 1, out_.right.backoff, back)) return;
// Words after the first, so extending a bigram to begin with
for (unsigned char extend_length = 2; extend_length <= in.left.length; ++extend_length) {
if (ExtendLeft(in, next_use, extend_length, back, back2)) return;
std::swap(back, back2);
}
if (in.left.full) {
for (const float *i = back; i != back + next_use; ++i) prob_ += *i;
left_done_ = true;
out_.right = in.right;
return;
}
// Right state was minimized, so it's already independent of the new words to the left.
if (in.right.length < in.left.length) {
out_.right = in.right;
return;
}
// Shift exisiting words down.
for (WordIndex *i = out_.right.words + next_use - 1; i >= out_.right.words; --i) {
*(i + in.right.length) = *i;
}
// Add words from in.right.
std::copy(in.right.words, in.right.words + in.right.length, out_.right.words);
// Assemble backoff composed on the existing state's backoff followed by the new state's backoff.
std::copy(in.right.backoff, in.right.backoff + in.right.length, out_.right.backoff);
std::copy(back, back + next_use, out_.right.backoff + in.right.length);
out_.right.length = in.right.length + next_use;
}
float Finish() {
// A N-1-gram might extend left and right but we should still set full to true because it's an N-1-gram.
out_.left.full = left_done_ || (out_.left.length == model_.Order() - 1);
return prob_;
}
void Reset() {
prob_ = 0.0;
left_done_ = false;
out_.left.length = 0;
out_.right.length = 0;
}
private:
bool ExtendLeft(const ChartState &in, unsigned char &next_use, unsigned char extend_length, const float *back_in, float *back_out) {
ProcessRet(model_.ExtendLeft(
out_.right.words, out_.right.words + next_use, // Words to extend into
back_in, // Backoffs to use
in.left.pointers[extend_length - 1], extend_length, // Words to be extended
back_out, // Backoffs for the next score
next_use)); // Length of n-gram to use in next scoring.
if (next_use != out_.right.length) {
left_done_ = true;
if (!next_use) {
// Early exit.
out_.right = in.right;
prob_ += model_.UnRest(in.left.pointers + extend_length, in.left.pointers + in.left.length, extend_length + 1);
return true;
}
}
// Continue scoring.
return false;
}
void ProcessRet(const FullScoreReturn &ret) {
if (left_done_) {
prob_ += ret.prob;
return;
}
if (ret.independent_left) {
prob_ += ret.prob;
left_done_ = true;
return;
}
out_.left.pointers[out_.left.length++] = ret.extend_left;
prob_ += ret.rest;
}
const M &model_;
ChartState &out_;
bool left_done_;
float prob_;
};
} // namespace ngram
} // namespace lm
#endif // LM_LEFT__

397
lm/left_test.cc
View File

@ -1,397 +0,0 @@
#include "lm/left.hh"
#include "lm/model.hh"
#include "util/tokenize_piece.hh"
#include <vector>
#define BOOST_TEST_MODULE LeftTest
#include <boost/test/unit_test.hpp>
#include <boost/test/floating_point_comparison.hpp>
namespace lm {
namespace ngram {
namespace {
#define Term(word) score.Terminal(m.GetVocabulary().Index(word));
#define VCheck(word, value) BOOST_CHECK_EQUAL(m.GetVocabulary().Index(word), value);
// Apparently some Boost versions use templates and are pretty strict about types matching.
#define SLOPPY_CHECK_CLOSE(ref, value, tol) BOOST_CHECK_CLOSE(static_cast<double>(ref), static_cast<double>(value), static_cast<double>(tol));
template <class M> void Short(const M &m) {
ChartState base;
{
RuleScore<M> score(m, base);
Term("more");
Term("loin");
SLOPPY_CHECK_CLOSE(-1.206319 - 0.3561665, score.Finish(), 0.001);
}
BOOST_CHECK(base.left.full);
BOOST_CHECK_EQUAL(2, base.left.length);
BOOST_CHECK_EQUAL(1, base.right.length);
VCheck("loin", base.right.words[0]);
ChartState more_left;
{
RuleScore<M> score(m, more_left);
Term("little");
score.NonTerminal(base, -1.206319 - 0.3561665);
// p(little more loin | null context)
SLOPPY_CHECK_CLOSE(-1.56538, score.Finish(), 0.001);
}
BOOST_CHECK_EQUAL(3, more_left.left.length);
BOOST_CHECK_EQUAL(1, more_left.right.length);
VCheck("loin", more_left.right.words[0]);
BOOST_CHECK(more_left.left.full);
ChartState shorter;
{
RuleScore<M> score(m, shorter);
Term("to");
score.NonTerminal(base, -1.206319 - 0.3561665);
SLOPPY_CHECK_CLOSE(-0.30103 - 1.687872 - 1.206319 - 0.3561665, score.Finish(), 0.01);
}
BOOST_CHECK_EQUAL(1, shorter.left.length);
BOOST_CHECK_EQUAL(1, shorter.right.length);
VCheck("loin", shorter.right.words[0]);
BOOST_CHECK(shorter.left.full);
}
template <class M> void Charge(const M &m) {
ChartState base;
{
RuleScore<M> score(m, base);
Term("on");
Term("more");
SLOPPY_CHECK_CLOSE(-1.509559 -0.4771212 -1.206319, score.Finish(), 0.001);
}
BOOST_CHECK_EQUAL(1, base.left.length);
BOOST_CHECK_EQUAL(1, base.right.length);
VCheck("more", base.right.words[0]);
BOOST_CHECK(base.left.full);
ChartState extend;
{
RuleScore<M> score(m, extend);
Term("looking");
score.NonTerminal(base, -1.509559 -0.4771212 -1.206319);
SLOPPY_CHECK_CLOSE(-3.91039, score.Finish(), 0.001);
}
BOOST_CHECK_EQUAL(2, extend.left.length);
BOOST_CHECK_EQUAL(1, extend.right.length);
VCheck("more", extend.right.words[0]);
BOOST_CHECK(extend.left.full);
ChartState tobos;
{
RuleScore<M> score(m, tobos);
score.BeginSentence();
score.NonTerminal(extend, -3.91039);
SLOPPY_CHECK_CLOSE(-3.471169, score.Finish(), 0.001);
}
BOOST_CHECK_EQUAL(0, tobos.left.length);
BOOST_CHECK_EQUAL(1, tobos.right.length);
}
template <class M> float LeftToRight(const M &m, const std::vector<WordIndex> &words, bool begin_sentence = false) {
float ret = 0.0;
State right = begin_sentence ? m.BeginSentenceState() : m.NullContextState();
for (std::vector<WordIndex>::const_iterator i = words.begin(); i != words.end(); ++i) {
State copy(right);
ret += m.Score(copy, *i, right);
}
return ret;
}
template <class M> float RightToLeft(const M &m, const std::vector<WordIndex> &words, bool begin_sentence = false) {
float ret = 0.0;
ChartState state;
state.left.length = 0;
state.right.length = 0;
state.left.full = false;
for (std::vector<WordIndex>::const_reverse_iterator i = words.rbegin(); i != words.rend(); ++i) {
ChartState copy(state);
RuleScore<M> score(m, state);
score.Terminal(*i);
score.NonTerminal(copy, ret);
ret = score.Finish();
}
if (begin_sentence) {
ChartState copy(state);
RuleScore<M> score(m, state);
score.BeginSentence();
score.NonTerminal(copy, ret);
ret = score.Finish();
}
return ret;
}
template <class M> float TreeMiddle(const M &m, const std::vector<WordIndex> &words, bool begin_sentence = false) {
std::vector<std::pair<ChartState, float> > states(words.size());
for (unsigned int i = 0; i < words.size(); ++i) {
RuleScore<M> score(m, states[i].first);
score.Terminal(words[i]);
states[i].second = score.Finish();
}
while (states.size() > 1) {
std::vector<std::pair<ChartState, float> > upper((states.size() + 1) / 2);
for (unsigned int i = 0; i < states.size() / 2; ++i) {
RuleScore<M> score(m, upper[i].first);
score.NonTerminal(states[i*2].first, states[i*2].second);
score.NonTerminal(states[i*2+1].first, states[i*2+1].second);
upper[i].second = score.Finish();
}
if (states.size() % 2) {
upper.back() = states.back();
}
std::swap(states, upper);
}
if (states.empty()) return 0.0;
if (begin_sentence) {
ChartState ignored;
RuleScore<M> score(m, ignored);
score.BeginSentence();
score.NonTerminal(states.front().first, states.front().second);
return score.Finish();
} else {
return states.front().second;
}
}
template <class M> void LookupVocab(const M &m, const StringPiece &str, std::vector<WordIndex> &out) {
out.clear();
for (util::TokenIter<util::SingleCharacter, true> i(str, ' '); i; ++i) {
out.push_back(m.GetVocabulary().Index(*i));
}
}
#define TEXT_TEST(str) \
LookupVocab(m, str, words); \
expect = LeftToRight(m, words, rest); \
SLOPPY_CHECK_CLOSE(expect, RightToLeft(m, words, rest), 0.001); \
SLOPPY_CHECK_CLOSE(expect, TreeMiddle(m, words, rest), 0.001); \
// Build sentences, or parts thereof, from right to left.
template <class M> void GrowBig(const M &m, bool rest = false) {
std::vector<WordIndex> words;
float expect;
TEXT_TEST("in biarritz watching considering looking . on a little more loin also would consider higher to look good unknown the screening foo bar , unknown however unknown </s>");
TEXT_TEST("on a little more loin also would consider higher to look good unknown the screening foo bar , unknown however unknown </s>");
TEXT_TEST("on a little more loin also would consider higher to look good");
TEXT_TEST("more loin also would consider higher to look good");
TEXT_TEST("more loin also would consider higher to look");
TEXT_TEST("also would consider higher to look");
TEXT_TEST("also would consider higher");
TEXT_TEST("would consider higher to look");
TEXT_TEST("consider higher to look");
TEXT_TEST("consider higher to");
TEXT_TEST("consider higher");
}
template <class M> void GrowSmall(const M &m, bool rest = false) {
std::vector<WordIndex> words;
float expect;
TEXT_TEST("in biarritz watching considering looking . </s>");
TEXT_TEST("in biarritz watching considering looking .");
TEXT_TEST("in biarritz");
}
template <class M> void AlsoWouldConsiderHigher(const M &m) {
ChartState also;
{
RuleScore<M> score(m, also);
score.Terminal(m.GetVocabulary().Index("also"));
SLOPPY_CHECK_CLOSE(-1.687872, score.Finish(), 0.001);
}
ChartState would;
{
RuleScore<M> score(m, would);
score.Terminal(m.GetVocabulary().Index("would"));
SLOPPY_CHECK_CLOSE(-1.687872, score.Finish(), 0.001);
}
ChartState combine_also_would;
{
RuleScore<M> score(m, combine_also_would);
score.NonTerminal(also, -1.687872);
score.NonTerminal(would, -1.687872);
SLOPPY_CHECK_CLOSE(-1.687872 - 2.0, score.Finish(), 0.001);
}
BOOST_CHECK_EQUAL(2, combine_also_would.right.length);
ChartState also_would;
{
RuleScore<M> score(m, also_would);
score.Terminal(m.GetVocabulary().Index("also"));
score.Terminal(m.GetVocabulary().Index("would"));
SLOPPY_CHECK_CLOSE(-1.687872 - 2.0, score.Finish(), 0.001);
}
BOOST_CHECK_EQUAL(2, also_would.right.length);
ChartState consider;
{
RuleScore<M> score(m, consider);
score.Terminal(m.GetVocabulary().Index("consider"));
SLOPPY_CHECK_CLOSE(-1.687872, score.Finish(), 0.001);
}
BOOST_CHECK_EQUAL(1, consider.left.length);
BOOST_CHECK_EQUAL(1, consider.right.length);
BOOST_CHECK(!consider.left.full);
ChartState higher;
float higher_score;
{
RuleScore<M> score(m, higher);
score.Terminal(m.GetVocabulary().Index("higher"));
higher_score = score.Finish();
}
SLOPPY_CHECK_CLOSE(-1.509559, higher_score, 0.001);
BOOST_CHECK_EQUAL(1, higher.left.length);
BOOST_CHECK_EQUAL(1, higher.right.length);
BOOST_CHECK(!higher.left.full);
VCheck("higher", higher.right.words[0]);
SLOPPY_CHECK_CLOSE(-0.30103, higher.right.backoff[0], 0.001);
ChartState consider_higher;
{
RuleScore<M> score(m, consider_higher);
score.NonTerminal(consider, -1.687872);
score.NonTerminal(higher, higher_score);
SLOPPY_CHECK_CLOSE(-1.509559 - 1.687872 - 0.30103, score.Finish(), 0.001);
}
BOOST_CHECK_EQUAL(2, consider_higher.left.length);
BOOST_CHECK(!consider_higher.left.full);
ChartState full;
{
RuleScore<M> score(m, full);
score.NonTerminal(combine_also_would, -1.687872 - 2.0);
score.NonTerminal(consider_higher, -1.509559 - 1.687872 - 0.30103);
SLOPPY_CHECK_CLOSE(-10.6879, score.Finish(), 0.001);
}
BOOST_CHECK_EQUAL(4, full.right.length);
}
#define CHECK_SCORE(str, val) \
{ \
float got = val; \
std::vector<WordIndex> indices; \
LookupVocab(m, str, indices); \
SLOPPY_CHECK_CLOSE(LeftToRight(m, indices), got, 0.001); \
}
template <class M> void FullGrow(const M &m) {
std::vector<WordIndex> words;
LookupVocab(m, "in biarritz watching considering looking . </s>", words);
ChartState lexical[7];
float lexical_scores[7];
for (unsigned int i = 0; i < 7; ++i) {
RuleScore<M> score(m, lexical[i]);
score.Terminal(words[i]);
lexical_scores[i] = score.Finish();
}
CHECK_SCORE("in", lexical_scores[0]);
CHECK_SCORE("biarritz", lexical_scores[1]);
CHECK_SCORE("watching", lexical_scores[2]);
CHECK_SCORE("</s>", lexical_scores[6]);
ChartState l1[4];
float l1_scores[4];
{
RuleScore<M> score(m, l1[0]);
score.NonTerminal(lexical[0], lexical_scores[0]);
score.NonTerminal(lexical[1], lexical_scores[1]);
CHECK_SCORE("in biarritz", l1_scores[0] = score.Finish());
}
{
RuleScore<M> score(m, l1[1]);
score.NonTerminal(lexical[2], lexical_scores[2]);
score.NonTerminal(lexical[3], lexical_scores[3]);
CHECK_SCORE("watching considering", l1_scores[1] = score.Finish());
}
{
RuleScore<M> score(m, l1[2]);
score.NonTerminal(lexical[4], lexical_scores[4]);
score.NonTerminal(lexical[5], lexical_scores[5]);
CHECK_SCORE("looking .", l1_scores[2] = score.Finish());
}
BOOST_CHECK_EQUAL(l1[2].left.length, 1);
l1[3] = lexical[6];
l1_scores[3] = lexical_scores[6];
ChartState l2[2];
float l2_scores[2];
{
RuleScore<M> score(m, l2[0]);
score.NonTerminal(l1[0], l1_scores[0]);
score.NonTerminal(l1[1], l1_scores[1]);
CHECK_SCORE("in biarritz watching considering", l2_scores[0] = score.Finish());
}
{
RuleScore<M> score(m, l2[1]);
score.NonTerminal(l1[2], l1_scores[2]);
score.NonTerminal(l1[3], l1_scores[3]);
CHECK_SCORE("looking . </s>", l2_scores[1] = score.Finish());
}
BOOST_CHECK_EQUAL(l2[1].left.length, 1);
BOOST_CHECK(l2[1].left.full);
ChartState top;
{
RuleScore<M> score(m, top);
score.NonTerminal(l2[0], l2_scores[0]);
score.NonTerminal(l2[1], l2_scores[1]);
CHECK_SCORE("in biarritz watching considering looking . </s>", score.Finish());
}
}
const char *FileLocation() {
if (boost::unit_test::framework::master_test_suite().argc < 2) {
return "test.arpa";
}
return boost::unit_test::framework::master_test_suite().argv[1];
}
template <class M> void Everything() {
Config config;
config.messages = NULL;
M m(FileLocation(), config);
Short(m);
Charge(m);
GrowBig(m);
AlsoWouldConsiderHigher(m);
GrowSmall(m);
FullGrow(m);
}
BOOST_AUTO_TEST_CASE(ProbingAll) {
Everything<Model>();
}
BOOST_AUTO_TEST_CASE(TrieAll) {
Everything<TrieModel>();
}
BOOST_AUTO_TEST_CASE(QuantTrieAll) {
Everything<QuantTrieModel>();
}
BOOST_AUTO_TEST_CASE(ArrayQuantTrieAll) {
Everything<QuantArrayTrieModel>();
}
BOOST_AUTO_TEST_CASE(ArrayTrieAll) {
Everything<ArrayTrieModel>();
}
BOOST_AUTO_TEST_CASE(RestProbing) {
Config config;
config.messages = NULL;
RestProbingModel m(FileLocation(), config);
GrowBig(m, true);
}
} // namespace
} // namespace ngram
} // namespace lm

23
lm/lm_exception.cc
View File

@ -1,23 +0,0 @@
#include "lm/lm_exception.hh"
#include<errno.h>
#include<stdio.h>
namespace lm {
ConfigException::ConfigException() throw() {}
ConfigException::~ConfigException() throw() {}
LoadException::LoadException() throw() {}
LoadException::~LoadException() throw() {}
FormatLoadException::FormatLoadException() throw() {}
FormatLoadException::~FormatLoadException() throw() {}
VocabLoadException::VocabLoadException() throw() {}
VocabLoadException::~VocabLoadException() throw() {}
SpecialWordMissingException::SpecialWordMissingException() throw() {}
SpecialWordMissingException::~SpecialWordMissingException() throw() {}
} // namespace lm

50
lm/lm_exception.hh
View File

@ -1,50 +0,0 @@
#ifndef LM_LM_EXCEPTION__
#define LM_LM_EXCEPTION__
// Named to avoid conflict with util/exception.hh.
#include "util/exception.hh"
#include "util/string_piece.hh"
#include <exception>
#include <string>
namespace lm {
typedef enum {THROW_UP, COMPLAIN, SILENT} WarningAction;
class ConfigException : public util::Exception {
public:
ConfigException() throw();
~ConfigException() throw();
};
class LoadException : public util::Exception {
public:
virtual ~LoadException() throw();
protected:
LoadException() throw();
};
class FormatLoadException : public LoadException {
public:
FormatLoadException() throw();
~FormatLoadException() throw();
};
class VocabLoadException : public LoadException {
public:
virtual ~VocabLoadException() throw();
VocabLoadException() throw();
};
class SpecialWordMissingException : public VocabLoadException {
public:
explicit SpecialWordMissingException() throw();
~SpecialWordMissingException() throw();
};
} // namespace lm
#endif // LM_LM_EXCEPTION

5
lm/max_order.cc
View File

@ -1,5 +0,0 @@
#include <iostream>
int main(int argc, char *argv[]) {
std::cerr << "KenLM was compiled with a maximum supported n-gram order set to " << KENLM_MAX_ORDER << "." << std::endl;
}

294
lm/model.cc
View File

@ -1,294 +0,0 @@
#include "lm/model.hh"
#include "lm/blank.hh"
#include "lm/lm_exception.hh"
#include "lm/search_hashed.hh"
#include "lm/search_trie.hh"
#include "lm/read_arpa.hh"
#include "util/murmur_hash.hh"
#include <algorithm>
#include <functional>
#include <numeric>
#include <cmath>
namespace lm {
namespace ngram {
namespace detail {
template <class Search, class VocabularyT> const ModelType GenericModel<Search, VocabularyT>::kModelType = Search::kModelType;
template <class Search, class VocabularyT> size_t GenericModel<Search, VocabularyT>::Size(const std::vector<uint64_t> &counts, const Config &config) {
return VocabularyT::Size(counts[0], config) + Search::Size(counts, config);
}
template <class Search, class VocabularyT> void GenericModel<Search, VocabularyT>::SetupMemory(void *base, const std::vector<uint64_t> &counts, const Config &config) {
uint8_t *start = static_cast<uint8_t*>(base);
size_t allocated = VocabularyT::Size(counts[0], config);
vocab_.SetupMemory(start, allocated, counts[0], config);
start += allocated;
start = search_.SetupMemory(start, counts, config);
if (static_cast<std::size_t>(start - static_cast<uint8_t*>(base)) != Size(counts, config)) UTIL_THROW(FormatLoadException, "The data structures took " << (start - static_cast<uint8_t*>(base)) << " but Size says they should take " << Size(counts, config));
}
template <class Search, class VocabularyT> GenericModel<Search, VocabularyT>::GenericModel(const char *file, const Config &config) {
LoadLM(file, config, *this);
// g++ prints warnings unless these are fully initialized.
State begin_sentence = State();
begin_sentence.length = 1;
begin_sentence.words[0] = vocab_.BeginSentence();
typename Search::Node ignored_node;
bool ignored_independent_left;
uint64_t ignored_extend_left;
begin_sentence.backoff[0] = search_.LookupUnigram(begin_sentence.words[0], ignored_node, ignored_independent_left, ignored_extend_left).Backoff();
State null_context = State();
null_context.length = 0;
P::Init(begin_sentence, null_context, vocab_, search_.Order());
}
template <class Search, class VocabularyT> void GenericModel<Search, VocabularyT>::InitializeFromBinary(void *start, const Parameters &params, const Config &config, int fd) {
UTIL_THROW_IF(params.counts.size() > KENLM_MAX_ORDER, FormatLoadException, "This model has order " << params.counts.size() << ". Re-compile (use -a), passing a number at least this large to bjam's --max-kenlm-order flag.");
SetupMemory(start, params.counts, config);
vocab_.LoadedBinary(params.fixed.has_vocabulary, fd, config.enumerate_vocab);
search_.LoadedBinary();
}
template <class Search, class VocabularyT> void GenericModel<Search, VocabularyT>::InitializeFromARPA(const char *file, const Config &config) {
// Backing file is the ARPA. Steal it so we can make the backing file the mmap output if any.
util::FilePiece f(backing_.file.release(), file, config.messages);
try {
std::vector<uint64_t> counts;
// File counts do not include pruned trigrams that extend to quadgrams etc. These will be fixed by search_.
ReadARPACounts(f, counts);
UTIL_THROW_IF(counts.size() > KENLM_MAX_ORDER, FormatLoadException, "This model has order " << counts.size() << ". Re-compile (use -a), passing a number at least this large to bjam's --max-kenlm-order flag.");
if (counts.size() < 2) UTIL_THROW(FormatLoadException, "This ngram implementation assumes at least a bigram model.");
if (config.probing_multiplier <= 1.0) UTIL_THROW(ConfigException, "probing multiplier must be > 1.0");
std::size_t vocab_size = VocabularyT::Size(counts[0], config);
// Setup the binary file for writing the vocab lookup table. The search_ is responsible for growing the binary file to its needs.
vocab_.SetupMemory(SetupJustVocab(config, counts.size(), vocab_size, backing_), vocab_size, counts[0], config);
if (config.write_mmap) {
WriteWordsWrapper wrap(config.enumerate_vocab);
vocab_.ConfigureEnumerate(&wrap, counts[0]);
search_.InitializeFromARPA(file, f, counts, config, vocab_, backing_);
wrap.Write(backing_.file.get());
} else {
vocab_.ConfigureEnumerate(config.enumerate_vocab, counts[0]);
search_.InitializeFromARPA(file, f, counts, config, vocab_, backing_);
}
if (!vocab_.SawUnk()) {
assert(config.unknown_missing != THROW_UP);
// Default probabilities for unknown.
search_.UnknownUnigram().backoff = 0.0;
search_.UnknownUnigram().prob = config.unknown_missing_logprob;
}
FinishFile(config, kModelType, kVersion, counts, vocab_.UnkCountChangePadding(), backing_);
} catch (util::Exception &e) {
e << " Byte: " << f.Offset();
throw;
}
}
template <class Search, class VocabularyT> void GenericModel<Search, VocabularyT>::UpdateConfigFromBinary(int fd, const std::vector<uint64_t> &counts, Config &config) {
util::AdvanceOrThrow(fd, VocabularyT::Size(counts[0], config));
Search::UpdateConfigFromBinary(fd, counts, config);
}
template <class Search, class VocabularyT> FullScoreReturn GenericModel<Search, VocabularyT>::FullScore(const State &in_state, const WordIndex new_word, State &out_state) const {
FullScoreReturn ret = ScoreExceptBackoff(in_state.words, in_state.words + in_state.length, new_word, out_state);
for (const float *i = in_state.backoff + ret.ngram_length - 1; i < in_state.backoff + in_state.length; ++i) {
ret.prob += *i;
}
return ret;
}
template <class Search, class VocabularyT> FullScoreReturn GenericModel<Search, VocabularyT>::FullScoreForgotState(const WordIndex *context_rbegin, const WordIndex *context_rend, const WordIndex new_word, State &out_state) const {
context_rend = std::min(context_rend, context_rbegin + P::Order() - 1);
FullScoreReturn ret = ScoreExceptBackoff(context_rbegin, context_rend, new_word, out_state);
// Add the backoff weights for n-grams of order start to (context_rend - context_rbegin).
unsigned char start = ret.ngram_length;
if (context_rend - context_rbegin < static_cast<std::ptrdiff_t>(start)) return ret;
bool independent_left;
uint64_t extend_left;
typename Search::Node node;
if (start <= 1) {
ret.prob += search_.LookupUnigram(*context_rbegin, node, independent_left, extend_left).Backoff();
start = 2;
} else if (!search_.FastMakeNode(context_rbegin, context_rbegin + start - 1, node)) {
return ret;
}
// i is the order of the backoff we're looking for.
unsigned char order_minus_2 = start - 2;
for (const WordIndex *i = context_rbegin + start - 1; i < context_rend; ++i, ++order_minus_2) {
typename Search::MiddlePointer p(search_.LookupMiddle(order_minus_2, *i, node, independent_left, extend_left));
if (!p.Found()) break;
ret.prob += p.Backoff();
}
return ret;
}
template <class Search, class VocabularyT> void GenericModel<Search, VocabularyT>::GetState(const WordIndex *context_rbegin, const WordIndex *context_rend, State &out_state) const {
// Generate a state from context.
context_rend = std::min(context_rend, context_rbegin + P::Order() - 1);
if (context_rend == context_rbegin) {
out_state.length = 0;
return;
}
typename Search::Node node;
bool independent_left;
uint64_t extend_left;
out_state.backoff[0] = search_.LookupUnigram(*context_rbegin, node, independent_left, extend_left).Backoff();
out_state.length = HasExtension(out_state.backoff[0]) ? 1 : 0;
float *backoff_out = out_state.backoff + 1;
unsigned char order_minus_2 = 0;
for (const WordIndex *i = context_rbegin + 1; i < context_rend; ++i, ++backoff_out, ++order_minus_2) {
typename Search::MiddlePointer p(search_.LookupMiddle(order_minus_2, *i, node, independent_left, extend_left));
if (!p.Found()) {
std::copy(context_rbegin, context_rbegin + out_state.length, out_state.words);
return;
}
*backoff_out = p.Backoff();
if (HasExtension(*backoff_out)) out_state.length = i - context_rbegin + 1;
}
std::copy(context_rbegin, context_rbegin + out_state.length, out_state.words);
}
template <class Search, class VocabularyT> FullScoreReturn GenericModel<Search, VocabularyT>::ExtendLeft(
const WordIndex *add_rbegin, const WordIndex *add_rend,
const float *backoff_in,
uint64_t extend_pointer,
unsigned char extend_length,
float *backoff_out,
unsigned char &next_use) const {
FullScoreReturn ret;
typename Search::Node node;
if (extend_length == 1) {
typename Search::UnigramPointer ptr(search_.LookupUnigram(static_cast<WordIndex>(extend_pointer), node, ret.independent_left, ret.extend_left));
ret.rest = ptr.Rest();
ret.prob = ptr.Prob();
assert(!ret.independent_left);
} else {
typename Search::MiddlePointer ptr(search_.Unpack(extend_pointer, extend_length, node));
ret.rest = ptr.Rest();
ret.prob = ptr.Prob();
ret.extend_left = extend_pointer;
// If this function is called, then it does depend on left words.
ret.independent_left = false;
}
float subtract_me = ret.rest;
ret.ngram_length = extend_length;
next_use = extend_length;
ResumeScore(add_rbegin, add_rend, extend_length - 1, node, backoff_out, next_use, ret);
next_use -= extend_length;
// Charge backoffs.
for (const float *b = backoff_in + ret.ngram_length - extend_length; b < backoff_in + (add_rend - add_rbegin); ++b) ret.prob += *b;
ret.prob -= subtract_me;
ret.rest -= subtract_me;
return ret;
}
namespace {
// Do a paraonoid copy of history, assuming new_word has already been copied
// (hence the -1). out_state.length could be zero so I avoided using
// std::copy.
void CopyRemainingHistory(const WordIndex *from, State &out_state) {
WordIndex *out = out_state.words + 1;
const WordIndex *in_end = from + static_cast<ptrdiff_t>(out_state.length) - 1;
for (const WordIndex *in = from; in < in_end; ++in, ++out) *out = *in;
}
} // namespace
/* Ugly optimized function. Produce a score excluding backoff.
* The search goes in increasing order of ngram length.
* Context goes backward, so context_begin is the word immediately preceeding
* new_word.
*/
template <class Search, class VocabularyT> FullScoreReturn GenericModel<Search, VocabularyT>::ScoreExceptBackoff(
const WordIndex *const context_rbegin,
const WordIndex *const context_rend,
const WordIndex new_word,
State &out_state) const {
FullScoreReturn ret;
// ret.ngram_length contains the last known non-blank ngram length.
ret.ngram_length = 1;
typename Search::Node node;
typename Search::UnigramPointer uni(search_.LookupUnigram(new_word, node, ret.independent_left, ret.extend_left));
out_state.backoff[0] = uni.Backoff();
ret.prob = uni.Prob();
ret.rest = uni.Rest();
// This is the length of the context that should be used for continuation to the right.
out_state.length = HasExtension(out_state.backoff[0]) ? 1 : 0;
// We'll write the word anyway since it will probably be used and does no harm being there.
out_state.words[0] = new_word;
if (context_rbegin == context_rend) return ret;
ResumeScore(context_rbegin, context_rend, 0, node, out_state.backoff + 1, out_state.length, ret);
CopyRemainingHistory(context_rbegin, out_state);
return ret;
}
template <class Search, class VocabularyT> void GenericModel<Search, VocabularyT>::ResumeScore(const WordIndex *hist_iter, const WordIndex *const context_rend, unsigned char order_minus_2, typename Search::Node &node, float *backoff_out, unsigned char &next_use, FullScoreReturn &ret) const {
for (; ; ++order_minus_2, ++hist_iter, ++backoff_out) {
if (hist_iter == context_rend) return;
if (ret.independent_left) return;
if (order_minus_2 == P::Order() - 2) break;
typename Search::MiddlePointer pointer(search_.LookupMiddle(order_minus_2, *hist_iter, node, ret.independent_left, ret.extend_left));
if (!pointer.Found()) return;
*backoff_out = pointer.Backoff();
ret.prob = pointer.Prob();
ret.rest = pointer.Rest();
ret.ngram_length = order_minus_2 + 2;
if (HasExtension(*backoff_out)) {
next_use = ret.ngram_length;
}
}
ret.independent_left = true;
typename Search::LongestPointer longest(search_.LookupLongest(*hist_iter, node));
if (longest.Found()) {
ret.prob = longest.Prob();
ret.rest = ret.prob;
// There is no blank in longest_.
ret.ngram_length = P::Order();
}
}
template <class Search, class VocabularyT> float GenericModel<Search, VocabularyT>::InternalUnRest(const uint64_t *pointers_begin, const uint64_t *pointers_end, unsigned char first_length) const {
float ret;
typename Search::Node node;
if (first_length == 1) {
if (pointers_begin >= pointers_end) return 0.0;
bool independent_left;
uint64_t extend_left;
typename Search::UnigramPointer ptr(search_.LookupUnigram(static_cast<WordIndex>(*pointers_begin), node, independent_left, extend_left));
ret = ptr.Prob() - ptr.Rest();
++first_length;
++pointers_begin;
} else {
ret = 0.0;
}
for (const uint64_t *i = pointers_begin; i < pointers_end; ++i, ++first_length) {
typename Search::MiddlePointer ptr(search_.Unpack(*i, first_length, node));
ret += ptr.Prob() - ptr.Rest();
}
return ret;
}
template class GenericModel<HashedSearch<BackoffValue>, ProbingVocabulary>;
template class GenericModel<HashedSearch<RestValue>, ProbingVocabulary>;
template class GenericModel<trie::TrieSearch<DontQuantize, trie::DontBhiksha>, SortedVocabulary>;
template class GenericModel<trie::TrieSearch<DontQuantize, trie::ArrayBhiksha>, SortedVocabulary>;
template class GenericModel<trie::TrieSearch<SeparatelyQuantize, trie::DontBhiksha>, SortedVocabulary>;
template class GenericModel<trie::TrieSearch<SeparatelyQuantize, trie::ArrayBhiksha>, SortedVocabulary>;
} // namespace detail
} // namespace ngram
} // namespace lm

159
lm/model.hh
View File

@ -1,159 +0,0 @@
#ifndef LM_MODEL__
#define LM_MODEL__
#include "lm/bhiksha.hh"
#include "lm/binary_format.hh"
#include "lm/config.hh"
#include "lm/facade.hh"
#include "lm/quantize.hh"
#include "lm/search_hashed.hh"
#include "lm/search_trie.hh"
#include "lm/state.hh"
#include "lm/value.hh"
#include "lm/vocab.hh"
#include "lm/weights.hh"
#include "util/murmur_hash.hh"
#include <algorithm>
#include <vector>
#include <string.h>
namespace util { class FilePiece; }
namespace lm {
namespace ngram {
namespace detail {
// Should return the same results as SRI.
// ModelFacade typedefs Vocabulary so we use VocabularyT to avoid naming conflicts.
template <class Search, class VocabularyT> class GenericModel : public base::ModelFacade<GenericModel<Search, VocabularyT>, State, VocabularyT> {
private:
typedef base::ModelFacade<GenericModel<Search, VocabularyT>, State, VocabularyT> P;
public:
// This is the model type returned by RecognizeBinary.
static const ModelType kModelType;
static const unsigned int kVersion = Search::kVersion;
/* Get the size of memory that will be mapped given ngram counts. This
* does not include small non-mapped control structures, such as this class
* itself.
*/
static size_t Size(const std::vector<uint64_t> &counts, const Config &config = Config());
/* Load the model from a file. It may be an ARPA or binary file. Binary
* files must have the format expected by this class or you'll get an
* exception. So TrieModel can only load ARPA or binary created by
* TrieModel. To classify binary files, call RecognizeBinary in
* lm/binary_format.hh.
*/
explicit GenericModel(const char *file, const Config &config = Config());
/* Score p(new_word | in_state) and incorporate new_word into out_state.
* Note that in_state and out_state must be different references:
* &in_state != &out_state.
*/
FullScoreReturn FullScore(const State &in_state, const WordIndex new_word, State &out_state) const;
/* Slower call without in_state. Try to remember state, but sometimes it
* would cost too much memory or your decoder isn't setup properly.
* To use this function, make an array of WordIndex containing the context
* vocabulary ids in reverse order. Then, pass the bounds of the array:
* [context_rbegin, context_rend). The new_word is not part of the context
* array unless you intend to repeat words.
*/
FullScoreReturn FullScoreForgotState(const WordIndex *context_rbegin, const WordIndex *context_rend, const WordIndex new_word, State &out_state) const;
/* Get the state for a context. Don't use this if you can avoid it. Use
* BeginSentenceState or EmptyContextState and extend from those. If
* you're only going to use this state to call FullScore once, use
* FullScoreForgotState.
* To use this function, make an array of WordIndex containing the context
* vocabulary ids in reverse order. Then, pass the bounds of the array:
* [context_rbegin, context_rend).
*/
void GetState(const WordIndex *context_rbegin, const WordIndex *context_rend, State &out_state) const;
/* More efficient version of FullScore where a partial n-gram has already
* been scored.
* NOTE: THE RETURNED .rest AND .prob ARE RELATIVE TO THE .rest RETURNED BEFORE.
*/
FullScoreReturn ExtendLeft(
// Additional context in reverse order. This will update add_rend to
const WordIndex *add_rbegin, const WordIndex *add_rend,
// Backoff weights to use.
const float *backoff_in,
// extend_left returned by a previous query.
uint64_t extend_pointer,
// Length of n-gram that the pointer corresponds to.
unsigned char extend_length,
// Where to write additional backoffs for [extend_length + 1, min(Order() - 1, return.ngram_length)]
float *backoff_out,
// Amount of additional content that should be considered by the next call.
unsigned char &next_use) const;
/* Return probabilities minus rest costs for an array of pointers. The
* first length should be the length of the n-gram to which pointers_begin
* points.
*/
float UnRest(const uint64_t *pointers_begin, const uint64_t *pointers_end, unsigned char first_length) const {
// Compiler should optimize this if away.
return Search::kDifferentRest ? InternalUnRest(pointers_begin, pointers_end, first_length) : 0.0;
}
private:
friend void lm::ngram::LoadLM<>(const char *file, const Config &config, GenericModel<Search, VocabularyT> &to);
static void UpdateConfigFromBinary(int fd, const std::vector<uint64_t> &counts, Config &config);
FullScoreReturn ScoreExceptBackoff(const WordIndex *const context_rbegin, const WordIndex *const context_rend, const WordIndex new_word, State &out_state) const;
// Score bigrams and above. Do not include backoff.
void ResumeScore(const WordIndex *context_rbegin, const WordIndex *const context_rend, unsigned char starting_order_minus_2, typename Search::Node &node, float *backoff_out, unsigned char &next_use, FullScoreReturn &ret) const;
// Appears after Size in the cc file.
void SetupMemory(void *start, const std::vector<uint64_t> &counts, const Config &config);
void InitializeFromBinary(void *start, const Parameters &params, const Config &config, int fd);
void InitializeFromARPA(const char *file, const Config &config);
float InternalUnRest(const uint64_t *pointers_begin, const uint64_t *pointers_end, unsigned char first_length) const;
Backing &MutableBacking() { return backing_; }
Backing backing_;
VocabularyT vocab_;
Search search_;
};
} // namespace detail
// Instead of typedef, inherit. This allows the Model etc to be forward declared.
// Oh the joys of C and C++.
#define LM_COMMA() ,
#define LM_NAME_MODEL(name, from)\
class name : public from {\
public:\
name(const char *file, const Config &config = Config()) : from(file, config) {}\
};
LM_NAME_MODEL(ProbingModel, detail::GenericModel<detail::HashedSearch<BackoffValue> LM_COMMA() ProbingVocabulary>);
LM_NAME_MODEL(RestProbingModel, detail::GenericModel<detail::HashedSearch<RestValue> LM_COMMA() ProbingVocabulary>);
LM_NAME_MODEL(TrieModel, detail::GenericModel<trie::TrieSearch<DontQuantize LM_COMMA() trie::DontBhiksha> LM_COMMA() SortedVocabulary>);
LM_NAME_MODEL(ArrayTrieModel, detail::GenericModel<trie::TrieSearch<DontQuantize LM_COMMA() trie::ArrayBhiksha> LM_COMMA() SortedVocabulary>);
LM_NAME_MODEL(QuantTrieModel, detail::GenericModel<trie::TrieSearch<SeparatelyQuantize LM_COMMA() trie::DontBhiksha> LM_COMMA() SortedVocabulary>);
LM_NAME_MODEL(QuantArrayTrieModel, detail::GenericModel<trie::TrieSearch<SeparatelyQuantize LM_COMMA() trie::ArrayBhiksha> LM_COMMA() SortedVocabulary>);
// Default implementation. No real reason for it to be the default.
typedef ::lm::ngram::ProbingVocabulary Vocabulary;
typedef ProbingModel Model;
} // namespace ngram
} // namespace lm
#endif // LM_MODEL__

438
lm/model_test.cc
View File

@ -1,438 +0,0 @@
#include "lm/model.hh"
#include <stdlib.h>
#define BOOST_TEST_MODULE ModelTest
#include <boost/test/unit_test.hpp>
#include <boost/test/floating_point_comparison.hpp>
// Apparently some Boost versions use templates and are pretty strict about types matching.
#define SLOPPY_CHECK_CLOSE(ref, value, tol) BOOST_CHECK_CLOSE(static_cast<double>(ref), static_cast<double>(value), static_cast<double>(tol));
namespace lm {
namespace ngram {
std::ostream &operator<<(std::ostream &o, const State &state) {
o << "State length " << static_cast<unsigned int>(state.length) << ':';
for (const WordIndex *i = state.words; i < state.words + state.length; ++i) {
o << ' ' << *i;
}
return o;
}
namespace {
const char *TestLocation() {
if (boost::unit_test::framework::master_test_suite().argc < 2) {
return "test.arpa";
}
return boost::unit_test::framework::master_test_suite().argv[1];
}
const char *TestNoUnkLocation() {
if (boost::unit_test::framework::master_test_suite().argc < 3) {
return "test_nounk.arpa";
}
return boost::unit_test::framework::master_test_suite().argv[2];
}
template <class Model> State GetState(const Model &model, const char *word, const State &in) {
WordIndex context[in.length + 1];
context[0] = model.GetVocabulary().Index(word);
std::copy(in.words, in.words + in.length, context + 1);
State ret;
model.GetState(context, context + in.length + 1, ret);
return ret;
}
#define StartTest(word, ngram, score, indep_left) \
ret = model.FullScore( \
state, \
model.GetVocabulary().Index(word), \
out);\
SLOPPY_CHECK_CLOSE(score, ret.prob, 0.001); \
BOOST_CHECK_EQUAL(static_cast<unsigned int>(ngram), ret.ngram_length); \
BOOST_CHECK_GE(std::min<unsigned char>(ngram, 5 - 1), out.length); \
BOOST_CHECK_EQUAL(indep_left, ret.independent_left); \
BOOST_CHECK_EQUAL(out, GetState(model, word, state));
#define AppendTest(word, ngram, score, indep_left) \
StartTest(word, ngram, score, indep_left) \
state = out;
template <class M> void Starters(const M &model) {
FullScoreReturn ret;
Model::State state(model.BeginSentenceState());
Model::State out;
StartTest("looking", 2, -0.4846522, true);
// , probability plus <s> backoff
StartTest(",", 1, -1.383514 + -0.4149733, true);
// <unk> probability plus <s> backoff
StartTest("this_is_not_found", 1, -1.995635 + -0.4149733, true);
}
template <class M> void Continuation(const M &model) {
FullScoreReturn ret;
Model::State state(model.BeginSentenceState());
Model::State out;
AppendTest("looking", 2, -0.484652, true);
AppendTest("on", 3, -0.348837, true);
AppendTest("a", 4, -0.0155266, true);
AppendTest("little", 5, -0.00306122, true);
State preserve = state;
AppendTest("the", 1, -4.04005, true);
AppendTest("biarritz", 1, -1.9889, true);
AppendTest("not_found", 1, -2.29666, true);
AppendTest("more", 1, -1.20632 - 20.0, true);
AppendTest(".", 2, -0.51363, true);
AppendTest("</s>", 3, -0.0191651, true);
BOOST_CHECK_EQUAL(0, state.length);
state = preserve;
AppendTest("more", 5, -0.00181395, true);
BOOST_CHECK_EQUAL(4, state.length);
AppendTest("loin", 5, -0.0432557, true);
BOOST_CHECK_EQUAL(1, state.length);
}
template <class M> void Blanks(const M &model) {
FullScoreReturn ret;
State state(model.NullContextState());
State out;
AppendTest("also", 1, -1.687872, false);
AppendTest("would", 2, -2, true);
AppendTest("consider", 3, -3, true);
State preserve = state;
AppendTest("higher", 4, -4, true);
AppendTest("looking", 5, -5, true);
BOOST_CHECK_EQUAL(1, state.length);
state = preserve;
// also would consider not_found
AppendTest("not_found", 1, -1.995635 - 7.0 - 0.30103, true);
state = model.NullContextState();
// higher looking is a blank.
AppendTest("higher", 1, -1.509559, false);
AppendTest("looking", 2, -1.285941 - 0.30103, false);
State higher_looking = state;
BOOST_CHECK_EQUAL(1, state.length);
AppendTest("not_found", 1, -1.995635 - 0.4771212, true);
state = higher_looking;
// higher looking consider
AppendTest("consider", 1, -1.687872 - 0.4771212, true);
state = model.NullContextState();
AppendTest("would", 1, -1.687872, false);
BOOST_CHECK_EQUAL(1, state.length);
AppendTest("consider", 2, -1.687872 -0.30103, false);
BOOST_CHECK_EQUAL(2, state.length);
AppendTest("higher", 3, -1.509559 - 0.30103, false);
BOOST_CHECK_EQUAL(3, state.length);
AppendTest("looking", 4, -1.285941 - 0.30103, false);
}
template <class M> void Unknowns(const M &model) {
FullScoreReturn ret;
State state(model.NullContextState());
State out;
AppendTest("not_found", 1, -1.995635, false);
State preserve = state;
AppendTest("not_found2", 2, -15.0, true);
AppendTest("not_found3", 2, -15.0 - 2.0, true);
state = preserve;
AppendTest("however", 2, -4, true);
AppendTest("not_found3", 3, -6, true);
}
template <class M> void MinimalState(const M &model) {
FullScoreReturn ret;
State state(model.NullContextState());
State out;
AppendTest("baz", 1, -6.535897, true);
BOOST_CHECK_EQUAL(0, state.length);
state = model.NullContextState();
AppendTest("foo", 1, -3.141592, true);
BOOST_CHECK_EQUAL(1, state.length);
AppendTest("bar", 2, -6.0, true);
// Has to include the backoff weight.
BOOST_CHECK_EQUAL(1, state.length);
AppendTest("bar", 1, -2.718281 + 3.0, true);
BOOST_CHECK_EQUAL(1, state.length);
state = model.NullContextState();
AppendTest("to", 1, -1.687872, false);
AppendTest("look", 2, -0.2922095, true);
BOOST_CHECK_EQUAL(2, state.length);
AppendTest("good", 3, -7, true);
}
template <class M> void ExtendLeftTest(const M &model) {
State right;
FullScoreReturn little(model.FullScore(model.NullContextState(), model.GetVocabulary().Index("little"), right));
const float kLittleProb = -1.285941;
SLOPPY_CHECK_CLOSE(kLittleProb, little.prob, 0.001);
unsigned char next_use;
float backoff_out[4];
FullScoreReturn extend_none(model.ExtendLeft(NULL, NULL, NULL, little.extend_left, 1, NULL, next_use));
BOOST_CHECK_EQUAL(0, next_use);
BOOST_CHECK_EQUAL(little.extend_left, extend_none.extend_left);
SLOPPY_CHECK_CLOSE(little.prob - little.rest, extend_none.prob, 0.001);
BOOST_CHECK_EQUAL(1, extend_none.ngram_length);
const WordIndex a = model.GetVocabulary().Index("a");
float backoff_in = 3.14;
// a little
FullScoreReturn extend_a(model.ExtendLeft(&a, &a + 1, &backoff_in, little.extend_left, 1, backoff_out, next_use));
BOOST_CHECK_EQUAL(1, next_use);
SLOPPY_CHECK_CLOSE(-0.69897, backoff_out[0], 0.001);
SLOPPY_CHECK_CLOSE(-0.09132547 - little.rest, extend_a.prob, 0.001);
BOOST_CHECK_EQUAL(2, extend_a.ngram_length);
BOOST_CHECK(!extend_a.independent_left);
const WordIndex on = model.GetVocabulary().Index("on");
FullScoreReturn extend_on(model.ExtendLeft(&on, &on + 1, &backoff_in, extend_a.extend_left, 2, backoff_out, next_use));
BOOST_CHECK_EQUAL(1, next_use);
SLOPPY_CHECK_CLOSE(-0.4771212, backoff_out[0], 0.001);
SLOPPY_CHECK_CLOSE(-0.0283603 - (extend_a.rest + little.rest), extend_on.prob, 0.001);
BOOST_CHECK_EQUAL(3, extend_on.ngram_length);
BOOST_CHECK(!extend_on.independent_left);
const WordIndex both[2] = {a, on};
float backoff_in_arr[4];
FullScoreReturn extend_both(model.ExtendLeft(both, both + 2, backoff_in_arr, little.extend_left, 1, backoff_out, next_use));
BOOST_CHECK_EQUAL(2, next_use);
SLOPPY_CHECK_CLOSE(-0.69897, backoff_out[0], 0.001);
SLOPPY_CHECK_CLOSE(-0.4771212, backoff_out[1], 0.001);
SLOPPY_CHECK_CLOSE(-0.0283603 - little.rest, extend_both.prob, 0.001);
BOOST_CHECK_EQUAL(3, extend_both.ngram_length);
BOOST_CHECK(!extend_both.independent_left);
BOOST_CHECK_EQUAL(extend_on.extend_left, extend_both.extend_left);
}
#define StatelessTest(word, provide, ngram, score) \
ret = model.FullScoreForgotState(indices + num_words - word, indices + num_words - word + provide, indices[num_words - word - 1], state); \
SLOPPY_CHECK_CLOSE(score, ret.prob, 0.001); \
BOOST_CHECK_EQUAL(static_cast<unsigned int>(ngram), ret.ngram_length); \
model.GetState(indices + num_words - word, indices + num_words - word + provide, before); \
ret = model.FullScore(before, indices[num_words - word - 1], out); \
BOOST_CHECK(state == out); \
SLOPPY_CHECK_CLOSE(score, ret.prob, 0.001); \
BOOST_CHECK_EQUAL(static_cast<unsigned int>(ngram), ret.ngram_length);
template <class M> void Stateless(const M &model) {
const char *words[] = {"<s>", "looking", "on", "a", "little", "the", "biarritz", "not_found", "more", ".", "</s>"};
const size_t num_words = sizeof(words) / sizeof(const char*);
// Silience "array subscript is above array bounds" when extracting end pointer.
WordIndex indices[num_words + 1];
for (unsigned int i = 0; i < num_words; ++i) {
indices[num_words - 1 - i] = model.GetVocabulary().Index(words[i]);
}
FullScoreReturn ret;
State state, out, before;
ret = model.FullScoreForgotState(indices + num_words - 1, indices + num_words, indices[num_words - 2], state);
SLOPPY_CHECK_CLOSE(-0.484652, ret.prob, 0.001);
StatelessTest(1, 1, 2, -0.484652);
// looking
StatelessTest(1, 2, 2, -0.484652);
// on
AppendTest("on", 3, -0.348837, true);
StatelessTest(2, 3, 3, -0.348837);
StatelessTest(2, 2, 3, -0.348837);
StatelessTest(2, 1, 2, -0.4638903);
// a
StatelessTest(3, 4, 4, -0.0155266);
// little
AppendTest("little", 5, -0.00306122, true);
StatelessTest(4, 5, 5, -0.00306122);
// the
AppendTest("the", 1, -4.04005, true);
StatelessTest(5, 5, 1, -4.04005);
// No context of the.
StatelessTest(5, 0, 1, -1.687872);
// biarritz
StatelessTest(6, 1, 1, -1.9889);
// not found
StatelessTest(7, 1, 1, -2.29666);
StatelessTest(7, 0, 1, -1.995635);
WordIndex unk[1];
unk[0] = 0;
model.GetState(unk, unk + 1, state);
BOOST_CHECK_EQUAL(1, state.length);
BOOST_CHECK_EQUAL(static_cast<WordIndex>(0), state.words[0]);
}
template <class M> void NoUnkCheck(const M &model) {
WordIndex unk_index = 0;
State state;
FullScoreReturn ret = model.FullScoreForgotState(&unk_index, &unk_index + 1, unk_index, state);
SLOPPY_CHECK_CLOSE(-100.0, ret.prob, 0.001);
}
template <class M> void Everything(const M &m) {
Starters(m);
Continuation(m);
Blanks(m);
Unknowns(m);
MinimalState(m);
ExtendLeftTest(m);
Stateless(m);
}
class ExpectEnumerateVocab : public EnumerateVocab {
public:
ExpectEnumerateVocab() {}
void Add(WordIndex index, const StringPiece &str) {
BOOST_CHECK_EQUAL(seen.size(), index);
seen.push_back(std::string(str.data(), str.length()));
}
void Check(const base::Vocabulary &vocab) {
BOOST_CHECK_EQUAL(37ULL, seen.size());
BOOST_REQUIRE(!seen.empty());
BOOST_CHECK_EQUAL("<unk>", seen[0]);
for (WordIndex i = 0; i < seen.size(); ++i) {
BOOST_CHECK_EQUAL(i, vocab.Index(seen[i]));
}
}
void Clear() {
seen.clear();
}
std::vector<std::string> seen;
};
template <class ModelT> void LoadingTest() {
Config config;
config.arpa_complain = Config::NONE;
config.messages = NULL;
config.probing_multiplier = 2.0;
{
ExpectEnumerateVocab enumerate;
config.enumerate_vocab = &enumerate;
ModelT m(TestLocation(), config);
enumerate.Check(m.GetVocabulary());
BOOST_CHECK_EQUAL((WordIndex)37, m.GetVocabulary().Bound());
Everything(m);
}
{
ExpectEnumerateVocab enumerate;
config.enumerate_vocab = &enumerate;
ModelT m(TestNoUnkLocation(), config);
enumerate.Check(m.GetVocabulary());
BOOST_CHECK_EQUAL((WordIndex)37, m.GetVocabulary().Bound());
NoUnkCheck(m);
}
}
BOOST_AUTO_TEST_CASE(probing) {
LoadingTest<Model>();
}
BOOST_AUTO_TEST_CASE(trie) {
LoadingTest<TrieModel>();
}
BOOST_AUTO_TEST_CASE(quant_trie) {
LoadingTest<QuantTrieModel>();
}
BOOST_AUTO_TEST_CASE(bhiksha_trie) {
LoadingTest<ArrayTrieModel>();
}
BOOST_AUTO_TEST_CASE(quant_bhiksha_trie) {
LoadingTest<QuantArrayTrieModel>();
}
template <class ModelT> void BinaryTest() {
Config config;
config.write_mmap = "test.binary";
config.messages = NULL;
ExpectEnumerateVocab enumerate;
config.enumerate_vocab = &enumerate;
{
ModelT copy_model(TestLocation(), config);
enumerate.Check(copy_model.GetVocabulary());
enumerate.Clear();
Everything(copy_model);
}
config.write_mmap = NULL;
ModelType type;
BOOST_REQUIRE(RecognizeBinary("test.binary", type));
BOOST_CHECK_EQUAL(ModelT::kModelType, type);
{
ModelT binary("test.binary", config);
enumerate.Check(binary.GetVocabulary());
Everything(binary);
}
unlink("test.binary");
// Now test without <unk>.
config.write_mmap = "test_nounk.binary";
config.messages = NULL;
enumerate.Clear();
{
ModelT copy_model(TestNoUnkLocation(), config);
enumerate.Check(copy_model.GetVocabulary());
enumerate.Clear();
NoUnkCheck(copy_model);
}
config.write_mmap = NULL;
{
ModelT binary(TestNoUnkLocation(), config);
enumerate.Check(binary.GetVocabulary());
NoUnkCheck(binary);
}
unlink("test_nounk.binary");
}
BOOST_AUTO_TEST_CASE(write_and_read_probing) {
BinaryTest<ProbingModel>();
}
BOOST_AUTO_TEST_CASE(write_and_read_rest_probing) {
BinaryTest<RestProbingModel>();
}
BOOST_AUTO_TEST_CASE(write_and_read_trie) {
BinaryTest<TrieModel>();
}
BOOST_AUTO_TEST_CASE(write_and_read_quant_trie) {
BinaryTest<QuantTrieModel>();
}
BOOST_AUTO_TEST_CASE(write_and_read_array_trie) {
BinaryTest<ArrayTrieModel>();
}
BOOST_AUTO_TEST_CASE(write_and_read_quant_array_trie) {
BinaryTest<QuantArrayTrieModel>();
}
BOOST_AUTO_TEST_CASE(rest_max) {
Config config;
config.arpa_complain = Config::NONE;
config.messages = NULL;
RestProbingModel model(TestLocation(), config);
State state, out;
FullScoreReturn ret(model.FullScore(model.NullContextState(), model.GetVocabulary().Index("."), state));
SLOPPY_CHECK_CLOSE(-0.2705918, ret.rest, 0.001);
SLOPPY_CHECK_CLOSE(-0.01916512, model.FullScore(state, model.GetVocabulary().EndSentence(), out).rest, 0.001);
}
} // namespace
} // namespace ngram
} // namespace lm

23
lm/model_type.hh
View File

@ -1,23 +0,0 @@
#ifndef LM_MODEL_TYPE__
#define LM_MODEL_TYPE__
namespace lm {
namespace ngram {
/* Not the best numbering system, but it grew this way for historical reasons
* and I want to preserve existing binary files. */
typedef enum {PROBING=0, REST_PROBING=1, TRIE=2, QUANT_TRIE=3, ARRAY_TRIE=4, QUANT_ARRAY_TRIE=5} ModelType;
// Historical names.
const ModelType HASH_PROBING = PROBING;
const ModelType TRIE_SORTED = TRIE;
const ModelType QUANT_TRIE_SORTED = QUANT_TRIE;
const ModelType ARRAY_TRIE_SORTED = ARRAY_TRIE;
const ModelType QUANT_ARRAY_TRIE_SORTED = QUANT_ARRAY_TRIE;
const static ModelType kQuantAdd = static_cast<ModelType>(QUANT_TRIE - TRIE);
const static ModelType kArrayAdd = static_cast<ModelType>(ARRAY_TRIE - TRIE);
} // namespace ngram
} // namespace lm
#endif // LM_MODEL_TYPE__

47
lm/ngram_query.cc
View File

@ -1,47 +0,0 @@
#include "lm/ngram_query.hh"
int main(int argc, char *argv[]) {
if (!(argc == 2 || (argc == 3 && !strcmp(argv[2], "null")))) {
std::cerr << "Usage: " << argv[0] << " lm_file [null]" << std::endl;
std::cerr << "Input is wrapped in <s> and </s> unless null is passed." << std::endl;
return 1;
}
try {
bool sentence_context = (argc == 2);
using namespace lm::ngram;
ModelType model_type;
if (RecognizeBinary(argv[1], model_type)) {
switch(model_type) {
case PROBING:
Query<lm::ngram::ProbingModel>(argv[1], sentence_context, std::cin, std::cout);
break;
case REST_PROBING:
Query<lm::ngram::RestProbingModel>(argv[1], sentence_context, std::cin, std::cout);
break;
case TRIE:
Query<TrieModel>(argv[1], sentence_context, std::cin, std::cout);
break;
case QUANT_TRIE:
Query<QuantTrieModel>(argv[1], sentence_context, std::cin, std::cout);
break;
case ARRAY_TRIE:
Query<ArrayTrieModel>(argv[1], sentence_context, std::cin, std::cout);
break;
case QUANT_ARRAY_TRIE:
Query<QuantArrayTrieModel>(argv[1], sentence_context, std::cin, std::cout);
break;
default:
std::cerr << "Unrecognized kenlm model type " << model_type << std::endl;
abort();
}
} else {
Query<ProbingModel>(argv[1], sentence_context, std::cin, std::cout);
}
std::cerr << "Total time including destruction:\n";
util::PrintUsage(std::cerr);
} catch (const std::exception &e) {
std::cerr << e.what() << std::endl;
return 1;
}
return 0;
}

72
lm/ngram_query.hh
View File

@ -1,72 +0,0 @@
#ifndef LM_NGRAM_QUERY__
#define LM_NGRAM_QUERY__
#include "lm/enumerate_vocab.hh"
#include "lm/model.hh"
#include "util/usage.hh"
#include <cstdlib>
#include <iostream>
#include <ostream>
#include <istream>
#include <string>
namespace lm {
namespace ngram {
template <class Model> void Query(const Model &model, bool sentence_context, std::istream &in_stream, std::ostream &out_stream) {
std::cerr << "Loading statistics:\n";
util::PrintUsage(std::cerr);
typename Model::State state, out;
lm::FullScoreReturn ret;
std::string word;
while (in_stream) {
state = sentence_context ? model.BeginSentenceState() : model.NullContextState();
float total = 0.0;
bool got = false;
unsigned int oov = 0;
while (in_stream >> word) {
got = true;
lm::WordIndex vocab = model.GetVocabulary().Index(word);
if (vocab == 0) ++oov;
ret = model.FullScore(state, vocab, out);
total += ret.prob;
out_stream << word << '=' << vocab << ' ' << static_cast<unsigned int>(ret.ngram_length) << ' ' << ret.prob << '\t';
state = out;
char c;
while (true) {
c = in_stream.get();
if (!in_stream) break;
if (c == '\n') break;
if (!isspace(c)) {
in_stream.unget();
break;
}
}
if (c == '\n') break;
}
if (!got && !in_stream) break;
if (sentence_context) {
ret = model.FullScore(state, model.GetVocabulary().EndSentence(), out);
total += ret.prob;
out_stream << "</s>=" << model.GetVocabulary().EndSentence() << ' ' << static_cast<unsigned int>(ret.ngram_length) << ' ' << ret.prob << '\t';
}
out_stream << "Total: " << total << " OOV: " << oov << '\n';
}
std::cerr << "After queries:\n";
util::PrintUsage(std::cerr);
}
template <class M> void Query(const char *file, bool sentence_context, std::istream &in_stream, std::ostream &out_stream) {
Config config;
M model(file, config);
Query(model, sentence_context, in_stream, out_stream);
}
} // namespace ngram
} // namespace lm
#endif // LM_NGRAM_QUERY__

93
lm/quantize.cc
View File

@ -1,93 +0,0 @@
/* Quantize into bins of equal size as described in
* M. Federico and N. Bertoldi. 2006. How many bits are needed
* to store probabilities for phrase-based translation? In Proc.
* of the Workshop on Statistical Machine Translation, pages
* 94–101, New York City, June. Association for Computa-
* tional Linguistics.
*/
#include "lm/quantize.hh"
#include "lm/binary_format.hh"
#include "lm/lm_exception.hh"
#include "util/file.hh"
#include <algorithm>
#include <numeric>
namespace lm {
namespace ngram {
namespace {
void MakeBins(std::vector<float> &values, float *centers, uint32_t bins) {
std::sort(values.begin(), values.end());
std::vector<float>::const_iterator start = values.begin(), finish;
for (uint32_t i = 0; i < bins; ++i, ++centers, start = finish) {
finish = values.begin() + ((values.size() * static_cast<uint64_t>(i + 1)) / bins);
if (finish == start) {
// zero length bucket.
*centers = i ? *(centers - 1) : -std::numeric_limits<float>::infinity();
} else {
*centers = std::accumulate(start, finish, 0.0) / static_cast<float>(finish - start);
}
}
}
const char kSeparatelyQuantizeVersion = 2;
} // namespace
void SeparatelyQuantize::UpdateConfigFromBinary(int fd, const std::vector<uint64_t> &/*counts*/, Config &config) {
char version;
util::ReadOrThrow(fd, &version, 1);
util::ReadOrThrow(fd, &config.prob_bits, 1);
util::ReadOrThrow(fd, &config.backoff_bits, 1);
if (version != kSeparatelyQuantizeVersion) UTIL_THROW(FormatLoadException, "This file has quantization version " << (unsigned)version << " but the code expects version " << (unsigned)kSeparatelyQuantizeVersion);
util::AdvanceOrThrow(fd, -3);
}
void SeparatelyQuantize::SetupMemory(void *base, unsigned char order, const Config &config) {
prob_bits_ = config.prob_bits;
backoff_bits_ = config.backoff_bits;
// We need the reserved values.
if (config.prob_bits == 0) UTIL_THROW(ConfigException, "You can't quantize probability to zero");
if (config.backoff_bits == 0) UTIL_THROW(ConfigException, "You can't quantize backoff to zero");
if (config.prob_bits > 25) UTIL_THROW(ConfigException, "For efficiency reasons, quantizing probability supports at most 25 bits. Currently you have requested " << static_cast<unsigned>(config.prob_bits) << " bits.");
if (config.backoff_bits > 25) UTIL_THROW(ConfigException, "For efficiency reasons, quantizing backoff supports at most 25 bits. Currently you have requested " << static_cast<unsigned>(config.backoff_bits) << " bits.");
// Reserve 8 byte header for bit counts.
actual_base_ = static_cast<uint8_t*>(base);
float *start = reinterpret_cast<float*>(actual_base_ + 8);
for (unsigned char i = 0; i < order - 2; ++i) {
tables_[i][0] = Bins(prob_bits_, start);
start += (1ULL << prob_bits_);
tables_[i][1] = Bins(backoff_bits_, start);
start += (1ULL << backoff_bits_);
}
longest_ = tables_[order - 2][0] = Bins(prob_bits_, start);
}
void SeparatelyQuantize::Train(uint8_t order, std::vector<float> &prob, std::vector<float> &backoff) {
TrainProb(order, prob);
// Backoff
float *centers = tables_[order - 2][1].Populate();
*(centers++) = kNoExtensionBackoff;
*(centers++) = kExtensionBackoff;
MakeBins(backoff, centers, (1ULL << backoff_bits_) - 2);
}
void SeparatelyQuantize::TrainProb(uint8_t order, std::vector<float> &prob) {
float *centers = tables_[order - 2][0].Populate();
MakeBins(prob, centers, (1ULL << prob_bits_));
}
void SeparatelyQuantize::FinishedLoading(const Config &config) {
uint8_t *actual_base = actual_base_;
*(actual_base++) = kSeparatelyQuantizeVersion; // version
*(actual_base++) = config.prob_bits;
*(actual_base++) = config.backoff_bits;
}
} // namespace ngram
} // namespace lm

231
lm/quantize.hh
View File

@ -1,231 +0,0 @@
#ifndef LM_QUANTIZE_H__
#define LM_QUANTIZE_H__
#include "lm/blank.hh"
#include "lm/config.hh"
#include "lm/model_type.hh"
#include "util/bit_packing.hh"
#include <algorithm>
#include <vector>
#include <stdint.h>
#include <iostream>
namespace lm {
namespace ngram {
struct Config;
/* Store values directly and don't quantize. */
class DontQuantize {
public:
static const ModelType kModelTypeAdd = static_cast<ModelType>(0);
static void UpdateConfigFromBinary(int, const std::vector<uint64_t> &, Config &) {}
static std::size_t Size(uint8_t /*order*/, const Config &/*config*/) { return 0; }
static uint8_t MiddleBits(const Config &/*config*/) { return 63; }
static uint8_t LongestBits(const Config &/*config*/) { return 31; }
class MiddlePointer {
public:
MiddlePointer(const DontQuantize & /*quant*/, unsigned char /*order_minus_2*/, util::BitAddress address) : address_(address) {}
MiddlePointer() : address_(NULL, 0) {}
bool Found() const {
return address_.base != NULL;
}
float Prob() const {
return util::ReadNonPositiveFloat31(address_.base, address_.offset);
}
float Backoff() const {
return util::ReadFloat32(address_.base, address_.offset + 31);
}
float Rest() const { return Prob(); }
void Write(float prob, float backoff) {
util::WriteNonPositiveFloat31(address_.base, address_.offset, prob);
util::WriteFloat32(address_.base, address_.offset + 31, backoff);
}
private:
util::BitAddress address_;
};
class LongestPointer {
public:
explicit LongestPointer(const DontQuantize &/*quant*/, util::BitAddress address) : address_(address) {}
LongestPointer() : address_(NULL, 0) {}
bool Found() const {
return address_.base != NULL;
}
float Prob() const {
return util::ReadNonPositiveFloat31(address_.base, address_.offset);
}
void Write(float prob) {
util::WriteNonPositiveFloat31(address_.base, address_.offset, prob);
}
private:
util::BitAddress address_;
};
DontQuantize() {}
void SetupMemory(void * /*start*/, unsigned char /*order*/, const Config & /*config*/) {}
static const bool kTrain = false;
// These should never be called because kTrain is false.
void Train(uint8_t /*order*/, std::vector<float> &/*prob*/, std::vector<float> &/*backoff*/) {}
void TrainProb(uint8_t, std::vector<float> &/*prob*/) {}
void FinishedLoading(const Config &) {}
};
class SeparatelyQuantize {
private:
class Bins {
public:
// Sigh C++ default constructor
Bins() {}
Bins(uint8_t bits, float *begin) : begin_(begin), end_(begin_ + (1ULL << bits)), bits_(bits), mask_((1ULL << bits) - 1) {}
float *Populate() { return begin_; }
uint64_t EncodeProb(float value) const {
return Encode(value, 0);
}
uint64_t EncodeBackoff(float value) const {
if (value == 0.0) {
return HasExtension(value) ? kExtensionQuant : kNoExtensionQuant;
}
return Encode(value, 2);
}
float Decode(std::size_t off) const { return begin_[off]; }
uint8_t Bits() const { return bits_; }
uint64_t Mask() const { return mask_; }
private:
uint64_t Encode(float value, size_t reserved) const {
const float *above = std::lower_bound(static_cast<const float*>(begin_) + reserved, end_, value);
if (above == begin_ + reserved) return reserved;
if (above == end_) return end_ - begin_ - 1;
return above - begin_ - (value - *(above - 1) < *above - value);
}
float *begin_;
const float *end_;
uint8_t bits_;
uint64_t mask_;
};
public:
static const ModelType kModelTypeAdd = kQuantAdd;
static void UpdateConfigFromBinary(int fd, const std::vector<uint64_t> &counts, Config &config);
static std::size_t Size(uint8_t order, const Config &config) {
size_t longest_table = (static_cast<size_t>(1) << static_cast<size_t>(config.prob_bits)) * sizeof(float);
size_t middle_table = (static_cast<size_t>(1) << static_cast<size_t>(config.backoff_bits)) * sizeof(float) + longest_table;
// unigrams are currently not quantized so no need for a table.
return (order - 2) * middle_table + longest_table + /* for the bit counts and alignment padding) */ 8;
}
static uint8_t MiddleBits(const Config &config) { return config.prob_bits + config.backoff_bits; }
static uint8_t LongestBits(const Config &config) { return config.prob_bits; }
class MiddlePointer {
public:
MiddlePointer(const SeparatelyQuantize &quant, unsigned char order_minus_2, const util::BitAddress &address) : bins_(quant.GetTables(order_minus_2)), address_(address) {}
MiddlePointer() : address_(NULL, 0) {}
bool Found() const { return address_.base != NULL; }
float Prob() const {
return ProbBins().Decode(util::ReadInt25(address_.base, address_.offset + BackoffBins().Bits(), ProbBins().Bits(), ProbBins().Mask()));
}
float Backoff() const {
return BackoffBins().Decode(util::ReadInt25(address_.base, address_.offset, BackoffBins().Bits(), BackoffBins().Mask()));
}
float Rest() const { return Prob(); }
void Write(float prob, float backoff) const {
util::WriteInt57(address_.base, address_.offset, ProbBins().Bits() + BackoffBins().Bits(),
(ProbBins().EncodeProb(prob) << BackoffBins().Bits()) | BackoffBins().EncodeBackoff(backoff));
}
private:
const Bins &ProbBins() const { return bins_[0]; }
const Bins &BackoffBins() const { return bins_[1]; }
const Bins *bins_;
util::BitAddress address_;
};
class LongestPointer {
public:
LongestPointer(const SeparatelyQuantize &quant, const util::BitAddress &address) : table_(&quant.LongestTable()), address_(address) {}
LongestPointer() : address_(NULL, 0) {}
bool Found() const { return address_.base != NULL; }
void Write(float prob) const {
util::WriteInt25(address_.base, address_.offset, table_->Bits(), table_->EncodeProb(prob));
}
float Prob() const {
return table_->Decode(util::ReadInt25(address_.base, address_.offset, table_->Bits(), table_->Mask()));
}
private:
const Bins *table_;
util::BitAddress address_;
};
SeparatelyQuantize() {}
void SetupMemory(void *start, unsigned char order, const Config &config);
static const bool kTrain = true;
// Assumes 0.0 is removed from backoff.
void Train(uint8_t order, std::vector<float> &prob, std::vector<float> &backoff);
// Train just probabilities (for longest order).
void TrainProb(uint8_t order, std::vector<float> &prob);
void FinishedLoading(const Config &config);
const Bins *GetTables(unsigned char order_minus_2) const { return tables_[order_minus_2]; }
const Bins &LongestTable() const { return longest_; }
private:
Bins tables_[KENLM_MAX_ORDER - 1][2];
Bins longest_;
uint8_t *actual_base_;
uint8_t prob_bits_, backoff_bits_;
};
} // namespace ngram
} // namespace lm
#endif // LM_QUANTIZE_H__

148
lm/read_arpa.cc
View File

@ -1,148 +0,0 @@
#include "lm/read_arpa.hh"
#include "lm/blank.hh"
#include <cstdlib>
#include <iostream>
#include <vector>
#include <ctype.h>
#include <math.h>
#include <string.h>
#include <stdint.h>
#ifdef WIN32
#include <float.h>
#endif
namespace lm {
// 1 for '\t', '\n', and ' '. This is stricter than isspace.
const bool kARPASpaces[256] = {0,0,0,0,0,0,0,0,0,1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
namespace {
bool IsEntirelyWhiteSpace(const StringPiece &line) {
for (size_t i = 0; i < static_cast<size_t>(line.size()); ++i) {
if (!isspace(line.data()[i])) return false;
}
return true;
}
const char kBinaryMagic[] = "mmap lm http://kheafield.com/code";
} // namespace
void ReadARPACounts(util::FilePiece &in, std::vector<uint64_t> &number) {
number.clear();
StringPiece line;
while (IsEntirelyWhiteSpace(line = in.ReadLine())) {}
if (line != "\\data\\") {
if ((line.size() >= 2) && (line.data()[0] == 0x1f) && (static_cast<unsigned char>(line.data()[1]) == 0x8b)) {
UTIL_THROW(FormatLoadException, "Looks like a gzip file. If this is an ARPA file, pipe " << in.FileName() << " through zcat. If this already in binary format, you need to decompress it because mmap doesn't work on top of gzip.");
}
if (static_cast<size_t>(line.size()) >= strlen(kBinaryMagic) && StringPiece(line.data(), strlen(kBinaryMagic)) == kBinaryMagic)
UTIL_THROW(FormatLoadException, "This looks like a binary file but got sent to the ARPA parser. Did you compress the binary file or pass a binary file where only ARPA files are accepted?");
UTIL_THROW_IF(line.size() >= 4 && StringPiece(line.data(), 4) == "blmt", FormatLoadException, "This looks like an IRSTLM binary file. Did you forget to pass --text yes to compile-lm?");
UTIL_THROW_IF(line == "iARPA", FormatLoadException, "This looks like an IRSTLM iARPA file. You need an ARPA file. Run\n compile-lm --text yes " << in.FileName() << " " << in.FileName() << ".arpa\nfirst.");
UTIL_THROW(FormatLoadException, "first non-empty line was \"" << line << "\" not \\data\\.");
}
while (!IsEntirelyWhiteSpace(line = in.ReadLine())) {
if (line.size() < 6 || strncmp(line.data(), "ngram ", 6)) UTIL_THROW(FormatLoadException, "count line \"" << line << "\"doesn't begin with \"ngram \"");
// So strtol doesn't go off the end of line.
std::string remaining(line.data() + 6, line.size() - 6);
char *end_ptr;
unsigned long int length = std::strtol(remaining.c_str(), &end_ptr, 10);
if ((end_ptr == remaining.c_str()) || (length - 1 != number.size())) UTIL_THROW(FormatLoadException, "ngram count lengths should be consecutive starting with 1: " << line);
if (*end_ptr != '=') UTIL_THROW(FormatLoadException, "Expected = immediately following the first number in the count line " << line);
++end_ptr;
const char *start = end_ptr;
long int count = std::strtol(start, &end_ptr, 10);
if (count < 0) UTIL_THROW(FormatLoadException, "Negative n-gram count " << count);
if (start == end_ptr) UTIL_THROW(FormatLoadException, "Couldn't parse n-gram count from " << line);
number.push_back(count);
}
}
void ReadNGramHeader(util::FilePiece &in, unsigned int length) {
StringPiece line;
while (IsEntirelyWhiteSpace(line = in.ReadLine())) {}
std::stringstream expected;
expected << '\\' << length << "-grams:";
if (line != expected.str()) UTIL_THROW(FormatLoadException, "Was expecting n-gram header " << expected.str() << " but got " << line << " instead");
}
void ReadBackoff(util::FilePiece &in, Prob &/*weights*/) {
switch (in.get()) {
case '\t':
{
float got = in.ReadFloat();
if (got != 0.0)
UTIL_THROW(FormatLoadException, "Non-zero backoff " << got << " provided for an n-gram that should have no backoff");
}
break;
case '\n':
break;
default:
UTIL_THROW(FormatLoadException, "Expected tab or newline for backoff");
}
}
void ReadBackoff(util::FilePiece &in, float &backoff) {
// Always make zero negative.
// Negative zero means that no (n+1)-gram has this n-gram as context.
// Therefore the hypothesis state can be shorter. Of course, many n-grams
// are context for (n+1)-grams. An algorithm in the data structure will go
// back and set the backoff to positive zero in these cases.
switch (in.get()) {
case '\t':
backoff = in.ReadFloat();
if (backoff == ngram::kExtensionBackoff) backoff = ngram::kNoExtensionBackoff;
{
#ifdef WIN32
int float_class = _fpclass(backoff);
UTIL_THROW_IF(float_class == _FPCLASS_SNAN || float_class == _FPCLASS_QNAN || float_class == _FPCLASS_NINF || float_class == _FPCLASS_PINF, FormatLoadException, "Bad backoff " << backoff);
#else
int float_class = fpclassify(backoff);
UTIL_THROW_IF(float_class == FP_NAN || float_class == FP_INFINITE, FormatLoadException, "Bad backoff " << backoff);
#endif
}
UTIL_THROW_IF(in.get() != '\n', FormatLoadException, "Expected newline after backoff");
break;
case '\n':
backoff = ngram::kNoExtensionBackoff;
break;
default:
UTIL_THROW(FormatLoadException, "Expected tab or newline for backoff");
}
}
void ReadEnd(util::FilePiece &in) {
StringPiece line;
do {
line = in.ReadLine();
} while (IsEntirelyWhiteSpace(line));
if (line != "\\end\\") UTIL_THROW(FormatLoadException, "Expected \\end\\ but the ARPA file has " << line);
try {
while (true) {
line = in.ReadLine();
if (!IsEntirelyWhiteSpace(line)) UTIL_THROW(FormatLoadException, "Trailing line " << line);
}
} catch (const util::EndOfFileException &e) {}
}
void PositiveProbWarn::Warn(float prob) {
switch (action_) {
case THROW_UP:
UTIL_THROW(FormatLoadException, "Positive log probability " << prob << " in the model. This is a bug in IRSTLM; you can set config.positive_log_probability = SILENT or pass -i to build_binary to substitute 0.0 for the log probability. Error");
case COMPLAIN:
std::cerr << "There's a positive log probability " << prob << " in the APRA file, probably because of a bug in IRSTLM. This and subsequent entires will be mapepd to 0 log probability." << std::endl;
action_ = SILENT;
break;
case SILENT:
break;
}
}
} // namespace lm

90
lm/read_arpa.hh
View File

@ -1,90 +0,0 @@
#ifndef LM_READ_ARPA__
#define LM_READ_ARPA__
#include "lm/lm_exception.hh"
#include "lm/word_index.hh"
#include "lm/weights.hh"
#include "util/file_piece.hh"
#include <cstddef>
#include <iosfwd>
#include <vector>
namespace lm {
void ReadARPACounts(util::FilePiece &in, std::vector<uint64_t> &number);
void ReadNGramHeader(util::FilePiece &in, unsigned int length);
void ReadBackoff(util::FilePiece &in, Prob &weights);
void ReadBackoff(util::FilePiece &in, float &backoff);
inline void ReadBackoff(util::FilePiece &in, ProbBackoff &weights) {
ReadBackoff(in, weights.backoff);
}
inline void ReadBackoff(util::FilePiece &in, RestWeights &weights) {
ReadBackoff(in, weights.backoff);
}
void ReadEnd(util::FilePiece &in);
extern const bool kARPASpaces[256];
// Positive log probability warning.
class PositiveProbWarn {
public:
PositiveProbWarn() : action_(THROW_UP) {}
explicit PositiveProbWarn(WarningAction action) : action_(action) {}
void Warn(float prob);
private:
WarningAction action_;
};
template <class Voc, class Weights> void Read1Gram(util::FilePiece &f, Voc &vocab, Weights *unigrams, PositiveProbWarn &warn) {
try {
float prob = f.ReadFloat();
if (prob > 0.0) {
warn.Warn(prob);
prob = 0.0;
}
if (f.get() != '\t') UTIL_THROW(FormatLoadException, "Expected tab after probability");
Weights &value = unigrams[vocab.Insert(f.ReadDelimited(kARPASpaces))];
value.prob = prob;
ReadBackoff(f, value);
} catch(util::Exception &e) {
e << " in the 1-gram at byte " << f.Offset();
throw;
}
}
// Return true if a positive log probability came out.
template <class Voc, class Weights> void Read1Grams(util::FilePiece &f, std::size_t count, Voc &vocab, Weights *unigrams, PositiveProbWarn &warn) {
ReadNGramHeader(f, 1);
for (std::size_t i = 0; i < count; ++i) {
Read1Gram(f, vocab, unigrams, warn);
}
vocab.FinishedLoading(unigrams);
}
// Return true if a positive log probability came out.
template <class Voc, class Weights> void ReadNGram(util::FilePiece &f, const unsigned char n, const Voc &vocab, WordIndex *const reverse_indices, Weights &weights, PositiveProbWarn &warn) {
try {
weights.prob = f.ReadFloat();
if (weights.prob > 0.0) {
warn.Warn(weights.prob);
weights.prob = 0.0;
}
for (WordIndex *vocab_out = reverse_indices + n - 1; vocab_out >= reverse_indices; --vocab_out) {
*vocab_out = vocab.Index(f.ReadDelimited(kARPASpaces));
}
ReadBackoff(f, weights);
} catch(util::Exception &e) {
e << " in the " << static_cast<unsigned int>(n) << "-gram at byte " << f.Offset();
throw;
}
}
} // namespace lm
#endif // LM_READ_ARPA__

42
lm/return.hh
View File

@ -1,42 +0,0 @@
#ifndef LM_RETURN__
#define LM_RETURN__
#include <stdint.h>
namespace lm {
/* Structure returned by scoring routines. */
struct FullScoreReturn {
// log10 probability
float prob;
/* The length of n-gram matched. Do not use this for recombination.
* Consider a model containing only the following n-grams:
* -1 foo
* -3.14 bar
* -2.718 baz -5
* -6 foo bar
*
* If you score ``bar'' then ngram_length is 1 and recombination state is the
* empty string because bar has zero backoff and does not extend to the
* right.
* If you score ``foo'' then ngram_length is 1 and recombination state is
* ``foo''.
*
* Ideally, keep output states around and compare them. Failing that,
* get out_state.ValidLength() and use that length for recombination.
*/
unsigned char ngram_length;
/* Left extension information. If independent_left is set, then prob is
* independent of words to the left (up to additional backoff). Otherwise,
* extend_left indicates how to efficiently extend further to the left.
*/
bool independent_left;
uint64_t extend_left; // Defined only if independent_left
// Rest cost for extension to the left.
float rest;
};
} // namespace lm
#endif // LM_RETURN__

294
lm/search_hashed.cc
View File

@ -1,294 +0,0 @@
#include "lm/search_hashed.hh"
#include "lm/binary_format.hh"
#include "lm/blank.hh"
#include "lm/lm_exception.hh"
#include "lm/model.hh"
#include "lm/read_arpa.hh"
#include "lm/value.hh"
#include "lm/vocab.hh"
#include "util/bit_packing.hh"
#include "util/file_piece.hh"
#include <string>
namespace lm {
namespace ngram {
class ProbingModel;
namespace {
/* These are passed to ReadNGrams so that n-grams with zero backoff that appear as context will still be used in state. */
template <class Middle> class ActivateLowerMiddle {
public:
explicit ActivateLowerMiddle(Middle &middle) : modify_(middle) {}
void operator()(const WordIndex *vocab_ids, const unsigned int n) {
uint64_t hash = static_cast<WordIndex>(vocab_ids[1]);
for (const WordIndex *i = vocab_ids + 2; i < vocab_ids + n; ++i) {
hash = detail::CombineWordHash(hash, *i);
}
typename Middle::MutableIterator i;
// TODO: somehow get text of n-gram for this error message.
if (!modify_.UnsafeMutableFind(hash, i))
UTIL_THROW(FormatLoadException, "The context of every " << n << "-gram should appear as a " << (n-1) << "-gram");
SetExtension(i->value.backoff);
}
private:
Middle &modify_;
};
template <class Weights> class ActivateUnigram {
public:
explicit ActivateUnigram(Weights *unigram) : modify_(unigram) {}
void operator()(const WordIndex *vocab_ids, const unsigned int /*n*/) {
// assert(n == 2);
SetExtension(modify_[vocab_ids[1]].backoff);
}
private:
Weights *modify_;
};
// Find the lower order entry, inserting blanks along the way as necessary.
template <class Value> void FindLower(
const std::vector<uint64_t> &keys,
typename Value::Weights &unigram,
std::vector<util::ProbingHashTable<typename Value::ProbingEntry, util::IdentityHash> > &middle,
std::vector<typename Value::Weights *> &between) {
typename util::ProbingHashTable<typename Value::ProbingEntry, util::IdentityHash>::MutableIterator iter;
typename Value::ProbingEntry entry;
// Backoff will always be 0.0. We'll get the probability and rest in another pass.
entry.value.backoff = kNoExtensionBackoff;
// Go back and find the longest right-aligned entry, informing it that it extends left. Normally this will match immediately, but sometimes SRI is dumb.
for (int lower = keys.size() - 2; ; --lower) {
if (lower == -1) {
between.push_back(&unigram);
return;
}
entry.key = keys[lower];
bool found = middle[lower].FindOrInsert(entry, iter);
between.push_back(&iter->value);
if (found) return;
}
}
// Between usually has single entry, the value to adjust. But sometimes SRI stupidly pruned entries so it has unitialized blank values to be set here.
template <class Added, class Build> void AdjustLower(
const Added &added,
const Build &build,
std::vector<typename Build::Value::Weights *> &between,
const unsigned int n,
const std::vector<WordIndex> &vocab_ids,
typename Build::Value::Weights *unigrams,
std::vector<util::ProbingHashTable<typename Build::Value::ProbingEntry, util::IdentityHash> > &middle) {
typedef typename Build::Value Value;
if (between.size() == 1) {
build.MarkExtends(*between.front(), added);
return;
}
typedef util::ProbingHashTable<typename Value::ProbingEntry, util::IdentityHash> Middle;
float prob = -fabs(between.back()->prob);
// Order of the n-gram on which probabilities are based.
unsigned char basis = n - between.size();
assert(basis != 0);
typename Build::Value::Weights **change = &between.back();
// Skip the basis.
--change;
if (basis == 1) {
// Hallucinate a bigram based on a unigram's backoff and a unigram probability.
float &backoff = unigrams[vocab_ids[1]].backoff;
SetExtension(backoff);
prob += backoff;
(*change)->prob = prob;
build.SetRest(&*vocab_ids.begin(), 2, **change);
basis = 2;
--change;
}
uint64_t backoff_hash = static_cast<uint64_t>(vocab_ids[1]);
for (unsigned char i = 2; i <= basis; ++i) {
backoff_hash = detail::CombineWordHash(backoff_hash, vocab_ids[i]);
}
for (; basis < n - 1; ++basis, --change) {
typename Middle::MutableIterator gotit;
if (middle[basis - 2].UnsafeMutableFind(backoff_hash, gotit)) {
float &backoff = gotit->value.backoff;
SetExtension(backoff);
prob += backoff;
}
(*change)->prob = prob;
build.SetRest(&*vocab_ids.begin(), basis + 1, **change);
backoff_hash = detail::CombineWordHash(backoff_hash, vocab_ids[basis+1]);
}
typename std::vector<typename Value::Weights *>::const_iterator i(between.begin());
build.MarkExtends(**i, added);
const typename Value::Weights *longer = *i;
// Everything has probability but is not marked as extending.
for (++i; i != between.end(); ++i) {
build.MarkExtends(**i, *longer);
longer = *i;
}
}
// Continue marking lower entries even they know that they extend left. This is used for upper/lower bounds.
template <class Build> void MarkLower(
const std::vector<uint64_t> &keys,
const Build &build,
typename Build::Value::Weights &unigram,
std::vector<util::ProbingHashTable<typename Build::Value::ProbingEntry, util::IdentityHash> > &middle,
int start_order,
const typename Build::Value::Weights &longer) {
if (start_order == 0) return;
typename util::ProbingHashTable<typename Build::Value::ProbingEntry, util::IdentityHash>::MutableIterator iter;
// Hopefully the compiler will realize that if MarkExtends always returns false, it can simplify this code.
for (int even_lower = start_order - 2 /* index in middle */; ; --even_lower) {
if (even_lower == -1) {
build.MarkExtends(unigram, longer);
return;
}
middle[even_lower].UnsafeMutableFind(keys[even_lower], iter);
if (!build.MarkExtends(iter->value, longer)) return;
}
}
template <class Build, class Activate, class Store> void ReadNGrams(
util::FilePiece &f,
const unsigned int n,
const size_t count,
const ProbingVocabulary &vocab,
const Build &build,
typename Build::Value::Weights *unigrams,
std::vector<util::ProbingHashTable<typename Build::Value::ProbingEntry, util::IdentityHash> > &middle,
Activate activate,
Store &store,
PositiveProbWarn &warn) {
typedef typename Build::Value Value;
typedef util::ProbingHashTable<typename Value::ProbingEntry, util::IdentityHash> Middle;
assert(n >= 2);
ReadNGramHeader(f, n);
// Both vocab_ids and keys are non-empty because n >= 2.
// vocab ids of words in reverse order.
std::vector<WordIndex> vocab_ids(n);
std::vector<uint64_t> keys(n-1);
typename Store::Entry entry;
std::vector<typename Value::Weights *> between;
for (size_t i = 0; i < count; ++i) {
ReadNGram(f, n, vocab, &*vocab_ids.begin(), entry.value, warn);
build.SetRest(&*vocab_ids.begin(), n, entry.value);
keys[0] = detail::CombineWordHash(static_cast<uint64_t>(vocab_ids.front()), vocab_ids[1]);
for (unsigned int h = 1; h < n - 1; ++h) {
keys[h] = detail::CombineWordHash(keys[h-1], vocab_ids[h+1]);
}
// Initially the sign bit is on, indicating it does not extend left. Most already have this but there might +0.0.
util::SetSign(entry.value.prob);
entry.key = keys[n-2];
store.Insert(entry);
between.clear();
FindLower<Value>(keys, unigrams[vocab_ids.front()], middle, between);
AdjustLower<typename Store::Entry::Value, Build>(entry.value, build, between, n, vocab_ids, unigrams, middle);
if (Build::kMarkEvenLower) MarkLower<Build>(keys, build, unigrams[vocab_ids.front()], middle, n - between.size() - 1, *between.back());
activate(&*vocab_ids.begin(), n);
}
store.FinishedInserting();
}
} // namespace
namespace detail {
template <class Value> uint8_t *HashedSearch<Value>::SetupMemory(uint8_t *start, const std::vector<uint64_t> &counts, const Config &config) {
std::size_t allocated = Unigram::Size(counts[0]);
unigram_ = Unigram(start, counts[0], allocated);
start += allocated;
for (unsigned int n = 2; n < counts.size(); ++n) {
allocated = Middle::Size(counts[n - 1], config.probing_multiplier);
middle_.push_back(Middle(start, allocated));
start += allocated;
}
allocated = Longest::Size(counts.back(), config.probing_multiplier);
longest_ = Longest(start, allocated);
start += allocated;
return start;
}
template <class Value> void HashedSearch<Value>::InitializeFromARPA(const char * /*file*/, util::FilePiece &f, const std::vector<uint64_t> &counts, const Config &config, ProbingVocabulary &vocab, Backing &backing) {
// TODO: fix sorted.
SetupMemory(GrowForSearch(config, vocab.UnkCountChangePadding(), Size(counts, config), backing), counts, config);
PositiveProbWarn warn(config.positive_log_probability);
Read1Grams(f, counts[0], vocab, unigram_.Raw(), warn);
CheckSpecials(config, vocab);
DispatchBuild(f, counts, config, vocab, warn);
}
template <> void HashedSearch<BackoffValue>::DispatchBuild(util::FilePiece &f, const std::vector<uint64_t> &counts, const Config &config, const ProbingVocabulary &vocab, PositiveProbWarn &warn) {
NoRestBuild build;
ApplyBuild(f, counts, config, vocab, warn, build);
}
template <> void HashedSearch<RestValue>::DispatchBuild(util::FilePiece &f, const std::vector<uint64_t> &counts, const Config &config, const ProbingVocabulary &vocab, PositiveProbWarn &warn) {
switch (config.rest_function) {
case Config::REST_MAX:
{
MaxRestBuild build;
ApplyBuild(f, counts, config, vocab, warn, build);
}
break;
case Config::REST_LOWER:
{
LowerRestBuild<ProbingModel> build(config, counts.size(), vocab);
ApplyBuild(f, counts, config, vocab, warn, build);
}
break;
}
}
template <class Value> template <class Build> void HashedSearch<Value>::ApplyBuild(util::FilePiece &f, const std::vector<uint64_t> &counts, const Config &config, const ProbingVocabulary &vocab, PositiveProbWarn &warn, const Build &build) {
for (WordIndex i = 0; i < counts[0]; ++i) {
build.SetRest(&i, (unsigned int)1, unigram_.Raw()[i]);
}
try {
if (counts.size() > 2) {
ReadNGrams<Build, ActivateUnigram<typename Value::Weights>, Middle>(
f, 2, counts[1], vocab, build, unigram_.Raw(), middle_, ActivateUnigram<typename Value::Weights>(unigram_.Raw()), middle_[0], warn);
}
for (unsigned int n = 3; n < counts.size(); ++n) {
ReadNGrams<Build, ActivateLowerMiddle<Middle>, Middle>(
f, n, counts[n-1], vocab, build, unigram_.Raw(), middle_, ActivateLowerMiddle<Middle>(middle_[n-3]), middle_[n-2], warn);
}
if (counts.size() > 2) {
ReadNGrams<Build, ActivateLowerMiddle<Middle>, Longest>(
f, counts.size(), counts[counts.size() - 1], vocab, build, unigram_.Raw(), middle_, ActivateLowerMiddle<Middle>(middle_.back()), longest_, warn);
} else {
ReadNGrams<Build, ActivateUnigram<typename Value::Weights>, Longest>(
f, counts.size(), counts[counts.size() - 1], vocab, build, unigram_.Raw(), middle_, ActivateUnigram<typename Value::Weights>(unigram_.Raw()), longest_, warn);
}
} catch (util::ProbingSizeException &e) {
UTIL_THROW(util::ProbingSizeException, "Avoid pruning n-grams like \"bar baz quux\" when \"foo bar baz quux\" is still in the model. KenLM will work when this pruning happens, but the probing model assumes these events are rare enough that using blank space in the probing hash table will cover all of them. Increase probing_multiplier (-p to build_binary) to add more blank spaces.\n");
}
ReadEnd(f);
}
template <class Value> void HashedSearch<Value>::LoadedBinary() {
unigram_.LoadedBinary();
for (typename std::vector<Middle>::iterator i = middle_.begin(); i != middle_.end(); ++i) {
i->LoadedBinary();
}
longest_.LoadedBinary();
}
template class HashedSearch<BackoffValue>;
template class HashedSearch<RestValue>;
} // namespace detail
} // namespace ngram
} // namespace lm

201
lm/search_hashed.hh
View File

@ -1,201 +0,0 @@
#ifndef LM_SEARCH_HASHED__
#define LM_SEARCH_HASHED__
#include "lm/model_type.hh"
#include "lm/config.hh"
#include "lm/read_arpa.hh"
#include "lm/return.hh"
#include "lm/weights.hh"
#include "util/bit_packing.hh"
#include "util/probing_hash_table.hh"
#include <algorithm>
#include <iostream>
#include <vector>
namespace util { class FilePiece; }
namespace lm {
namespace ngram {
struct Backing;
class ProbingVocabulary;
namespace detail {
inline uint64_t CombineWordHash(uint64_t current, const WordIndex next) {
uint64_t ret = (current * 8978948897894561157ULL) ^ (static_cast<uint64_t>(1 + next) * 17894857484156487943ULL);
return ret;
}
#pragma pack(push)
#pragma pack(4)
struct ProbEntry {
uint64_t key;
Prob value;
typedef uint64_t Key;
typedef Prob Value;
uint64_t GetKey() const {
return key;
}
};
#pragma pack(pop)
class LongestPointer {
public:
explicit LongestPointer(const float &to) : to_(&to) {}
LongestPointer() : to_(NULL) {}
bool Found() const {
return to_ != NULL;
}
float Prob() const {
return *to_;
}
private:
const float *to_;
};
template <class Value> class HashedSearch {
public:
typedef uint64_t Node;
typedef typename Value::ProbingProxy UnigramPointer;
typedef typename Value::ProbingProxy MiddlePointer;
typedef ::lm::ngram::detail::LongestPointer LongestPointer;
static const ModelType kModelType = Value::kProbingModelType;
static const bool kDifferentRest = Value::kDifferentRest;
static const unsigned int kVersion = 0;
// TODO: move probing_multiplier here with next binary file format update.
static void UpdateConfigFromBinary(int, const std::vector<uint64_t> &, Config &) {}
static std::size_t Size(const std::vector<uint64_t> &counts, const Config &config) {
std::size_t ret = Unigram::Size(counts[0]);
for (unsigned char n = 1; n < counts.size() - 1; ++n) {
ret += Middle::Size(counts[n], config.probing_multiplier);
}
return ret + Longest::Size(counts.back(), config.probing_multiplier);
}
uint8_t *SetupMemory(uint8_t *start, const std::vector<uint64_t> &counts, const Config &config);
void InitializeFromARPA(const char *file, util::FilePiece &f, const std::vector<uint64_t> &counts, const Config &config, ProbingVocabulary &vocab, Backing &backing);
void LoadedBinary();
unsigned char Order() const {
return middle_.size() + 2;
}
typename Value::Weights &UnknownUnigram() { return unigram_.Unknown(); }
UnigramPointer LookupUnigram(WordIndex word, Node &next, bool &independent_left, uint64_t &extend_left) const {
extend_left = static_cast<uint64_t>(word);
next = extend_left;
UnigramPointer ret(unigram_.Lookup(word));
independent_left = ret.IndependentLeft();
return ret;
}
#pragma GCC diagnostic ignored "-Wuninitialized"
MiddlePointer Unpack(uint64_t extend_pointer, unsigned char extend_length, Node &node) const {
node = extend_pointer;
typename Middle::ConstIterator found;
bool got = middle_[extend_length - 2].Find(extend_pointer, found);
assert(got);
(void)got;
return MiddlePointer(found->value);
}
MiddlePointer LookupMiddle(unsigned char order_minus_2, WordIndex word, Node &node, bool &independent_left, uint64_t &extend_pointer) const {
node = CombineWordHash(node, word);
typename Middle::ConstIterator found;
if (!middle_[order_minus_2].Find(node, found)) {
independent_left = true;
return MiddlePointer();
}
extend_pointer = node;
MiddlePointer ret(found->value);
independent_left = ret.IndependentLeft();
return ret;
}
LongestPointer LookupLongest(WordIndex word, const Node &node) const {
// Sign bit is always on because longest n-grams do not extend left.
typename Longest::ConstIterator found;
if (!longest_.Find(CombineWordHash(node, word), found)) return LongestPointer();
return LongestPointer(found->value.prob);
}
// Generate a node without necessarily checking that it actually exists.
// Optionally return false if it's know to not exist.
bool FastMakeNode(const WordIndex *begin, const WordIndex *end, Node &node) const {
assert(begin != end);
node = static_cast<Node>(*begin);
for (const WordIndex *i = begin + 1; i < end; ++i) {
node = CombineWordHash(node, *i);
}
return true;
}
private:
// Interpret config's rest cost build policy and pass the right template argument to ApplyBuild.
void DispatchBuild(util::FilePiece &f, const std::vector<uint64_t> &counts, const Config &config, const ProbingVocabulary &vocab, PositiveProbWarn &warn);
template <class Build> void ApplyBuild(util::FilePiece &f, const std::vector<uint64_t> &counts, const Config &config, const ProbingVocabulary &vocab, PositiveProbWarn &warn, const Build &build);
class Unigram {
public:
Unigram() {}
Unigram(void *start, uint64_t count, std::size_t /*allocated*/) :
unigram_(static_cast<typename Value::Weights*>(start))
#ifdef DEBUG
, count_(count)
#endif
{}
static std::size_t Size(uint64_t count) {
return (count + 1) * sizeof(ProbBackoff); // +1 for hallucinate <unk>
}
const typename Value::Weights &Lookup(WordIndex index) const {
#ifdef DEBUG
assert(index < count_);
#endif
return unigram_[index];
}
typename Value::Weights &Unknown() { return unigram_[0]; }
void LoadedBinary() {}
// For building.
typename Value::Weights *Raw() { return unigram_; }
private:
typename Value::Weights *unigram_;
#ifdef DEBUG
uint64_t count_;
#endif
};
Unigram unigram_;
typedef util::ProbingHashTable<typename Value::ProbingEntry, util::IdentityHash> Middle;
std::vector<Middle> middle_;
typedef util::ProbingHashTable<ProbEntry, util::IdentityHash> Longest;
Longest longest_;
};
} // namespace detail
} // namespace ngram
} // namespace lm
#endif // LM_SEARCH_HASHED__

610
lm/search_trie.cc
View File

@ -1,610 +0,0 @@
/* This is where the trie is built. It's on-disk. */
#include "lm/search_trie.hh"
#include "lm/bhiksha.hh"
#include "lm/binary_format.hh"
#include "lm/blank.hh"
#include "lm/lm_exception.hh"
#include "lm/quantize.hh"
#include "lm/trie.hh"
#include "lm/trie_sort.hh"
#include "lm/vocab.hh"
#include "lm/weights.hh"
#include "lm/word_index.hh"
#include "util/ersatz_progress.hh"
#include "util/mmap.hh"
#include "util/proxy_iterator.hh"
#include "util/scoped.hh"
#include "util/sized_iterator.hh"
#include <algorithm>
#include <cstring>
#include <cstdio>
#include <cstdlib>
#include <queue>
#include <limits>
#include <numeric>
#include <vector>
#if defined(_WIN32) || defined(_WIN64)
#include <windows.h>
#endif
namespace lm {
namespace ngram {
namespace trie {
namespace {
void ReadOrThrow(FILE *from, void *data, size_t size) {
UTIL_THROW_IF(1 != std::fread(data, size, 1, from), util::ErrnoException, "Short read");
}
int Compare(unsigned char order, const void *first_void, const void *second_void) {
const WordIndex *first = reinterpret_cast<const WordIndex*>(first_void), *second = reinterpret_cast<const WordIndex*>(second_void);
const WordIndex *end = first + order;
for (; first != end; ++first, ++second) {
if (*first < *second) return -1;
if (*first > *second) return 1;
}
return 0;
}
struct ProbPointer {
unsigned char array;
uint64_t index;
};
// Array of n-grams and float indices.
class BackoffMessages {
public:
void Init(std::size_t entry_size) {
current_ = NULL;
allocated_ = NULL;
entry_size_ = entry_size;
}
void Add(const WordIndex *to, ProbPointer index) {
while (current_ + entry_size_ > allocated_) {
std::size_t allocated_size = allocated_ - (uint8_t*)backing_.get();
Resize(std::max<std::size_t>(allocated_size * 2, entry_size_));
}
memcpy(current_, to, entry_size_ - sizeof(ProbPointer));
*reinterpret_cast<ProbPointer*>(current_ + entry_size_ - sizeof(ProbPointer)) = index;
current_ += entry_size_;
}
void Apply(float *const *const base, FILE *unigrams) {
FinishedAdding();
if (current_ == allocated_) return;
rewind(unigrams);
ProbBackoff weights;
WordIndex unigram = 0;
ReadOrThrow(unigrams, &weights, sizeof(weights));
for (; current_ != allocated_; current_ += entry_size_) {
const WordIndex &cur_word = *reinterpret_cast<const WordIndex*>(current_);
for (; unigram < cur_word; ++unigram) {
ReadOrThrow(unigrams, &weights, sizeof(weights));
}
if (!HasExtension(weights.backoff)) {
weights.backoff = kExtensionBackoff;
UTIL_THROW_IF(fseek(unigrams, -sizeof(weights), SEEK_CUR), util::ErrnoException, "Seeking backwards to denote unigram extension failed.");
WriteOrThrow(unigrams, &weights, sizeof(weights));
}
const ProbPointer &write_to = *reinterpret_cast<const ProbPointer*>(current_ + sizeof(WordIndex));
base[write_to.array][write_to.index] += weights.backoff;
}
backing_.reset();
}
void Apply(float *const *const base, RecordReader &reader) {
FinishedAdding();
if (current_ == allocated_) return;
// We'll also use the same buffer to record messages to blanks that they extend.
WordIndex *extend_out = reinterpret_cast<WordIndex*>(current_);
const unsigned char order = (entry_size_ - sizeof(ProbPointer)) / sizeof(WordIndex);
for (reader.Rewind(); reader && (current_ != allocated_); ) {
switch (Compare(order, reader.Data(), current_)) {
case -1:
++reader;
break;
case 1:
// Message but nobody to receive it. Write it down at the beginning of the buffer so we can inform this blank that it extends.
for (const WordIndex *w = reinterpret_cast<const WordIndex *>(current_); w != reinterpret_cast<const WordIndex *>(current_) + order; ++w, ++extend_out) *extend_out = *w;
current_ += entry_size_;
break;
case 0:
float &backoff = reinterpret_cast<ProbBackoff*>((uint8_t*)reader.Data() + order * sizeof(WordIndex))->backoff;
if (!HasExtension(backoff)) {
backoff = kExtensionBackoff;
reader.Overwrite(&backoff, sizeof(float));
} else {
const ProbPointer &write_to = *reinterpret_cast<const ProbPointer*>(current_ + entry_size_ - sizeof(ProbPointer));
base[write_to.array][write_to.index] += backoff;
}
current_ += entry_size_;
break;
}
}
// Now this is a list of blanks that extend right.
entry_size_ = sizeof(WordIndex) * order;
Resize(sizeof(WordIndex) * (extend_out - (const WordIndex*)backing_.get()));
current_ = (uint8_t*)backing_.get();
}
// Call after Apply
bool Extends(unsigned char order, const WordIndex *words) {
if (current_ == allocated_) return false;
assert(order * sizeof(WordIndex) == entry_size_);
while (true) {
switch(Compare(order, words, current_)) {
case 1:
current_ += entry_size_;
if (current_ == allocated_) return false;
break;
case -1:
return false;
case 0:
return true;
}
}
}
private:
void FinishedAdding() {
Resize(current_ - (uint8_t*)backing_.get());
// Sort requests in same order as files.
std::sort(
util::SizedIterator(util::SizedProxy(backing_.get(), entry_size_)),
util::SizedIterator(util::SizedProxy(current_, entry_size_)),
util::SizedCompare<EntryCompare>(EntryCompare((entry_size_ - sizeof(ProbPointer)) / sizeof(WordIndex))));
current_ = (uint8_t*)backing_.get();
}
void Resize(std::size_t to) {
std::size_t current = current_ - (uint8_t*)backing_.get();
backing_.call_realloc(to);
current_ = (uint8_t*)backing_.get() + current;
allocated_ = (uint8_t*)backing_.get() + to;
}
util::scoped_malloc backing_;
uint8_t *current_, *allocated_;
std::size_t entry_size_;
};
const float kBadProb = std::numeric_limits<float>::infinity();
class SRISucks {
public:
SRISucks() {
for (BackoffMessages *i = messages_; i != messages_ + KENLM_MAX_ORDER - 1; ++i)
i->Init(sizeof(ProbPointer) + sizeof(WordIndex) * (i - messages_ + 1));
}
void Send(unsigned char begin, unsigned char order, const WordIndex *to, float prob_basis) {
assert(prob_basis != kBadProb);
ProbPointer pointer;
pointer.array = order - 1;
pointer.index = values_[order - 1].size();
for (unsigned char i = begin; i < order; ++i) {
messages_[i - 1].Add(to, pointer);
}
values_[order - 1].push_back(prob_basis);
}
void ObtainBackoffs(unsigned char total_order, FILE *unigram_file, RecordReader *reader) {
for (unsigned char i = 0; i < KENLM_MAX_ORDER - 1; ++i) {
it_[i] = values_[i].empty() ? NULL : &*values_[i].begin();
}
messages_[0].Apply(it_, unigram_file);
BackoffMessages *messages = messages_ + 1;
const RecordReader *end = reader + total_order - 2 /* exclude unigrams and longest order */;
for (; reader != end; ++messages, ++reader) {
messages->Apply(it_, *reader);
}
}
ProbBackoff GetBlank(unsigned char total_order, unsigned char order, const WordIndex *indices) {
assert(order > 1);
ProbBackoff ret;
ret.prob = *(it_[order - 1]++);
ret.backoff = ((order != total_order - 1) && messages_[order - 1].Extends(order, indices)) ? kExtensionBackoff : kNoExtensionBackoff;
return ret;
}
const std::vector<float> &Values(unsigned char order) const {
return values_[order - 1];
}
private:
// This used to be one array. Then I needed to separate it by order for quantization to work.
std::vector<float> values_[KENLM_MAX_ORDER - 1];
BackoffMessages messages_[KENLM_MAX_ORDER - 1];
float *it_[KENLM_MAX_ORDER - 1];
};
class FindBlanks {
public:
FindBlanks(unsigned char order, const ProbBackoff *unigrams, SRISucks &messages)
: counts_(order), unigrams_(unigrams), sri_(messages) {}
float UnigramProb(WordIndex index) const {
return unigrams_[index].prob;
}
void Unigram(WordIndex /*index*/) {
++counts_[0];
}
void MiddleBlank(const unsigned char order, const WordIndex *indices, unsigned char lower, float prob_basis) {
sri_.Send(lower, order, indices + 1, prob_basis);
++counts_[order - 1];
}
void Middle(const unsigned char order, const void * /*data*/) {
++counts_[order - 1];
}
void Longest(const void * /*data*/) {
++counts_.back();
}
// Unigrams wrote one past.
void Cleanup() {
--counts_[0];
}
const std::vector<uint64_t> &Counts() const {
return counts_;
}
private:
std::vector<uint64_t> counts_;
const ProbBackoff *unigrams_;
SRISucks &sri_;
};
// Phase to actually write n-grams to the trie.
template <class Quant, class Bhiksha> class WriteEntries {
public:
WriteEntries(RecordReader *contexts, const Quant &quant, UnigramValue *unigrams, BitPackedMiddle<Bhiksha> *middle, BitPackedLongest &longest, unsigned char order, SRISucks &sri) :
contexts_(contexts),
quant_(quant),
unigrams_(unigrams),
middle_(middle),
longest_(longest),
bigram_pack_((order == 2) ? static_cast<BitPacked&>(longest_) : static_cast<BitPacked&>(*middle_)),
order_(order),
sri_(sri) {}
float UnigramProb(WordIndex index) const { return unigrams_[index].weights.prob; }
void Unigram(WordIndex word) {
unigrams_[word].next = bigram_pack_.InsertIndex();
}
void MiddleBlank(const unsigned char order, const WordIndex *indices, unsigned char /*lower*/, float /*prob_base*/) {
ProbBackoff weights = sri_.GetBlank(order_, order, indices);
typename Quant::MiddlePointer(quant_, order - 2, middle_[order - 2].Insert(indices[order - 1])).Write(weights.prob, weights.backoff);
}
void Middle(const unsigned char order, const void *data) {
RecordReader &context = contexts_[order - 1];
const WordIndex *words = reinterpret_cast<const WordIndex*>(data);
ProbBackoff weights = *reinterpret_cast<const ProbBackoff*>(words + order);
if (context && !memcmp(data, context.Data(), sizeof(WordIndex) * order)) {
SetExtension(weights.backoff);
++context;
}
typename Quant::MiddlePointer(quant_, order - 2, middle_[order - 2].Insert(words[order - 1])).Write(weights.prob, weights.backoff);
}
void Longest(const void *data) {
const WordIndex *words = reinterpret_cast<const WordIndex*>(data);
typename Quant::LongestPointer(quant_, longest_.Insert(words[order_ - 1])).Write(reinterpret_cast<const Prob*>(words + order_)->prob);
}
void Cleanup() {}
private:
RecordReader *contexts_;
const Quant &quant_;
UnigramValue *const unigrams_;
BitPackedMiddle<Bhiksha> *const middle_;
BitPackedLongest &longest_;
BitPacked &bigram_pack_;
const unsigned char order_;
SRISucks &sri_;
};
struct Gram {
Gram(const WordIndex *in_begin, unsigned char order) : begin(in_begin), end(in_begin + order) {}
const WordIndex *begin, *end;
// For queue, this is the direction we want.
bool operator<(const Gram &other) const {
return std::lexicographical_compare(other.begin, other.end, begin, end);
}
};
template <class Doing> class BlankManager {
public:
BlankManager(unsigned char total_order, Doing &doing) : total_order_(total_order), been_length_(0), doing_(doing) {
for (float *i = basis_; i != basis_ + KENLM_MAX_ORDER - 1; ++i) *i = kBadProb;
}
void Visit(const WordIndex *to, unsigned char length, float prob) {
basis_[length - 1] = prob;
unsigned char overlap = std::min<unsigned char>(length - 1, been_length_);
const WordIndex *cur;
WordIndex *pre;
for (cur = to, pre = been_; cur != to + overlap; ++cur, ++pre) {
if (*pre != *cur) break;
}
if (cur == to + length - 1) {
*pre = *cur;
been_length_ = length;
return;
}
// There are blanks to insert starting with order blank.
unsigned char blank = cur - to + 1;
UTIL_THROW_IF(blank == 1, FormatLoadException, "Missing a unigram that appears as context.");
const float *lower_basis;
for (lower_basis = basis_ + blank - 2; *lower_basis == kBadProb; --lower_basis) {}
unsigned char based_on = lower_basis - basis_ + 1;
for (; cur != to + length - 1; ++blank, ++cur, ++pre) {
assert(*lower_basis != kBadProb);
doing_.MiddleBlank(blank, to, based_on, *lower_basis);
*pre = *cur;
// Mark that the probability is a blank so it shouldn't be used as the basis for a later n-gram.
basis_[blank - 1] = kBadProb;
}
*pre = *cur;
been_length_ = length;
}
private:
const unsigned char total_order_;
WordIndex been_[KENLM_MAX_ORDER];
unsigned char been_length_;
float basis_[KENLM_MAX_ORDER];
Doing &doing_;
};
template <class Doing> void RecursiveInsert(const unsigned char total_order, const WordIndex unigram_count, RecordReader *input, std::ostream *progress_out, const char *message, Doing &doing) {
util::ErsatzProgress progress(unigram_count + 1, progress_out, message);
WordIndex unigram = 0;
std::priority_queue<Gram> grams;
grams.push(Gram(&unigram, 1));
for (unsigned char i = 2; i <= total_order; ++i) {
if (input[i-2]) grams.push(Gram(reinterpret_cast<const WordIndex*>(input[i-2].Data()), i));
}
BlankManager<Doing> blank(total_order, doing);
while (true) {
Gram top = grams.top();
grams.pop();
unsigned char order = top.end - top.begin;
if (order == 1) {
blank.Visit(&unigram, 1, doing.UnigramProb(unigram));
doing.Unigram(unigram);
progress.Set(unigram);
if (++unigram == unigram_count + 1) break;
grams.push(top);
} else {
if (order == total_order) {
blank.Visit(top.begin, order, reinterpret_cast<const Prob*>(top.end)->prob);
doing.Longest(top.begin);
} else {
blank.Visit(top.begin, order, reinterpret_cast<const ProbBackoff*>(top.end)->prob);
doing.Middle(order, top.begin);
}
RecordReader &reader = input[order - 2];
if (++reader) grams.push(top);
}
}
assert(grams.empty());
doing.Cleanup();
}
void SanityCheckCounts(const std::vector<uint64_t> &initial, const std::vector<uint64_t> &fixed) {
if (fixed[0] != initial[0]) UTIL_THROW(util::Exception, "Unigram count should be constant but initial is " << initial[0] << " and recounted is " << fixed[0]);
if (fixed.back() != initial.back()) UTIL_THROW(util::Exception, "Longest count should be constant but it changed from " << initial.back() << " to " << fixed.back());
for (unsigned char i = 0; i < initial.size(); ++i) {
if (fixed[i] < initial[i]) UTIL_THROW(util::Exception, "Counts came out lower than expected. This shouldn't happen");
}
}
template <class Quant> void TrainQuantizer(uint8_t order, uint64_t count, const std::vector<float> &additional, RecordReader &reader, util::ErsatzProgress &progress, Quant &quant) {
std::vector<float> probs(additional), backoffs;
probs.reserve(count + additional.size());
backoffs.reserve(count);
for (reader.Rewind(); reader; ++reader) {
const ProbBackoff &weights = *reinterpret_cast<const ProbBackoff*>(reinterpret_cast<const uint8_t*>(reader.Data()) + sizeof(WordIndex) * order);
probs.push_back(weights.prob);
if (weights.backoff != 0.0) backoffs.push_back(weights.backoff);
++progress;
}
quant.Train(order, probs, backoffs);
}
template <class Quant> void TrainProbQuantizer(uint8_t order, uint64_t count, RecordReader &reader, util::ErsatzProgress &progress, Quant &quant) {
std::vector<float> probs, backoffs;
probs.reserve(count);
for (reader.Rewind(); reader; ++reader) {
const Prob &weights = *reinterpret_cast<const Prob*>(reinterpret_cast<const uint8_t*>(reader.Data()) + sizeof(WordIndex) * order);
probs.push_back(weights.prob);
++progress;
}
quant.TrainProb(order, probs);
}
void PopulateUnigramWeights(FILE *file, WordIndex unigram_count, RecordReader &contexts, UnigramValue *unigrams) {
// Fill unigram probabilities.
try {
rewind(file);
for (WordIndex i = 0; i < unigram_count; ++i) {
ReadOrThrow(file, &unigrams[i].weights, sizeof(ProbBackoff));
if (contexts && *reinterpret_cast<const WordIndex*>(contexts.Data()) == i) {
SetExtension(unigrams[i].weights.backoff);
++contexts;
}
}
} catch (util::Exception &e) {
e << " while re-reading unigram probabilities";
throw;
}
}
} // namespace
template <class Quant, class Bhiksha> void BuildTrie(SortedFiles &files, std::vector<uint64_t> &counts, const Config &config, TrieSearch<Quant, Bhiksha> &out, Quant &quant, const SortedVocabulary &vocab, Backing &backing) {
RecordReader inputs[KENLM_MAX_ORDER - 1];
RecordReader contexts[KENLM_MAX_ORDER - 1];
for (unsigned char i = 2; i <= counts.size(); ++i) {
inputs[i-2].Init(files.Full(i), i * sizeof(WordIndex) + (i == counts.size() ? sizeof(Prob) : sizeof(ProbBackoff)));
contexts[i-2].Init(files.Context(i), (i-1) * sizeof(WordIndex));
}
SRISucks sri;
std::vector<uint64_t> fixed_counts;
util::scoped_FILE unigram_file;
util::scoped_fd unigram_fd(files.StealUnigram());
{
util::scoped_memory unigrams;
MapRead(util::POPULATE_OR_READ, unigram_fd.get(), 0, counts[0] * sizeof(ProbBackoff), unigrams);
FindBlanks finder(counts.size(), reinterpret_cast<const ProbBackoff*>(unigrams.get()), sri);
RecursiveInsert(counts.size(), counts[0], inputs, config.messages, "Identifying n-grams omitted by SRI", finder);
fixed_counts = finder.Counts();
}
unigram_file.reset(util::FDOpenOrThrow(unigram_fd));
for (const RecordReader *i = inputs; i != inputs + counts.size() - 2; ++i) {
if (*i) UTIL_THROW(FormatLoadException, "There's a bug in the trie implementation: the " << (i - inputs + 2) << "-gram table did not complete reading");
}
SanityCheckCounts(counts, fixed_counts);
counts = fixed_counts;
sri.ObtainBackoffs(counts.size(), unigram_file.get(), inputs);
out.SetupMemory(GrowForSearch(config, vocab.UnkCountChangePadding(), TrieSearch<Quant, Bhiksha>::Size(fixed_counts, config), backing), fixed_counts, config);
for (unsigned char i = 2; i <= counts.size(); ++i) {
inputs[i-2].Rewind();
}
if (Quant::kTrain) {
util::ErsatzProgress progress(std::accumulate(counts.begin() + 1, counts.end(), 0), config.messages, "Quantizing");
for (unsigned char i = 2; i < counts.size(); ++i) {
TrainQuantizer(i, counts[i-1], sri.Values(i), inputs[i-2], progress, quant);
}
TrainProbQuantizer(counts.size(), counts.back(), inputs[counts.size() - 2], progress, quant);
quant.FinishedLoading(config);
}
UnigramValue *unigrams = out.unigram_.Raw();
PopulateUnigramWeights(unigram_file.get(), counts[0], contexts[0], unigrams);
unigram_file.reset();
for (unsigned char i = 2; i <= counts.size(); ++i) {
inputs[i-2].Rewind();
}
// Fill entries except unigram probabilities.
{
WriteEntries<Quant, Bhiksha> writer(contexts, quant, unigrams, out.middle_begin_, out.longest_, counts.size(), sri);
RecursiveInsert(counts.size(), counts[0], inputs, config.messages, "Writing trie", writer);
}
// Do not disable this error message or else too little state will be returned. Both WriteEntries::Middle and returning state based on found n-grams will need to be fixed to handle this situation.
for (unsigned char order = 2; order <= counts.size(); ++order) {
const RecordReader &context = contexts[order - 2];
if (context) {
FormatLoadException e;
e << "A " << static_cast<unsigned int>(order) << "-gram has context";
const WordIndex *ctx = reinterpret_cast<const WordIndex*>(context.Data());
for (const WordIndex *i = ctx; i != ctx + order - 1; ++i) {
e << ' ' << *i;
}
e << " so this context must appear in the model as a " << static_cast<unsigned int>(order - 1) << "-gram but it does not";
throw e;
}
}
/* Set ending offsets so the last entry will be sized properly */
// Last entry for unigrams was already set.
if (out.middle_begin_ != out.middle_end_) {
for (typename TrieSearch<Quant, Bhiksha>::Middle *i = out.middle_begin_; i != out.middle_end_ - 1; ++i) {
i->FinishedLoading((i+1)->InsertIndex(), config);
}
(out.middle_end_ - 1)->FinishedLoading(out.longest_.InsertIndex(), config);
}
}
template <class Quant, class Bhiksha> uint8_t *TrieSearch<Quant, Bhiksha>::SetupMemory(uint8_t *start, const std::vector<uint64_t> &counts, const Config &config) {
quant_.SetupMemory(start, counts.size(), config);
start += Quant::Size(counts.size(), config);
unigram_.Init(start);
start += Unigram::Size(counts[0]);
FreeMiddles();
middle_begin_ = static_cast<Middle*>(malloc(sizeof(Middle) * (counts.size() - 2)));
middle_end_ = middle_begin_ + (counts.size() - 2);
std::vector<uint8_t*> middle_starts(counts.size() - 2);
for (unsigned char i = 2; i < counts.size(); ++i) {
middle_starts[i-2] = start;
start += Middle::Size(Quant::MiddleBits(config), counts[i-1], counts[0], counts[i], config);
}
// Crazy backwards thing so we initialize using pointers to ones that have already been initialized
for (unsigned char i = counts.size() - 1; i >= 2; --i) {
new (middle_begin_ + i - 2) Middle(
middle_starts[i-2],
quant_.MiddleBits(config),
counts[i-1],
counts[0],
counts[i],
(i == counts.size() - 1) ? static_cast<const BitPacked&>(longest_) : static_cast<const BitPacked &>(middle_begin_[i-1]),
config);
}
longest_.Init(start, quant_.LongestBits(config), counts[0]);
return start + Longest::Size(Quant::LongestBits(config), counts.back(), counts[0]);
}
template <class Quant, class Bhiksha> void TrieSearch<Quant, Bhiksha>::LoadedBinary() {
unigram_.LoadedBinary();
for (Middle *i = middle_begin_; i != middle_end_; ++i) {
i->LoadedBinary();
}
longest_.LoadedBinary();
}
template <class Quant, class Bhiksha> void TrieSearch<Quant, Bhiksha>::InitializeFromARPA(const char *file, util::FilePiece &f, std::vector<uint64_t> &counts, const Config &config, SortedVocabulary &vocab, Backing &backing) {
std::string temporary_prefix;
if (config.temporary_directory_prefix) {
temporary_prefix = config.temporary_directory_prefix;
} else if (config.write_mmap) {
temporary_prefix = config.write_mmap;
} else {
temporary_prefix = file;
}
// At least 1MB sorting memory.
SortedFiles sorted(config, f, counts, std::max<size_t>(config.building_memory, 1048576), temporary_prefix, vocab);
BuildTrie(sorted, counts, config, *this, quant_, vocab, backing);
}
template class TrieSearch<DontQuantize, DontBhiksha>;
template class TrieSearch<DontQuantize, ArrayBhiksha>;
template class TrieSearch<SeparatelyQuantize, DontBhiksha>;
template class TrieSearch<SeparatelyQuantize, ArrayBhiksha>;
} // namespace trie
} // namespace ngram
} // namespace lm

130
lm/search_trie.hh
View File

@ -1,130 +0,0 @@
#ifndef LM_SEARCH_TRIE__
#define LM_SEARCH_TRIE__
#include "lm/config.hh"
#include "lm/model_type.hh"
#include "lm/return.hh"
#include "lm/trie.hh"
#include "lm/weights.hh"
#include "util/file.hh"
#include "util/file_piece.hh"
#include <vector>
#include <assert.h>
namespace lm {
namespace ngram {
struct Backing;
class SortedVocabulary;
namespace trie {
template <class Quant, class Bhiksha> class TrieSearch;
class SortedFiles;
template <class Quant, class Bhiksha> void BuildTrie(SortedFiles &files, std::vector<uint64_t> &counts, const Config &config, TrieSearch<Quant, Bhiksha> &out, Quant &quant, const SortedVocabulary &vocab, Backing &backing);
template <class Quant, class Bhiksha> class TrieSearch {
public:
typedef NodeRange Node;
typedef ::lm::ngram::trie::UnigramPointer UnigramPointer;
typedef typename Quant::MiddlePointer MiddlePointer;
typedef typename Quant::LongestPointer LongestPointer;
static const bool kDifferentRest = false;
static const ModelType kModelType = static_cast<ModelType>(TRIE_SORTED + Quant::kModelTypeAdd + Bhiksha::kModelTypeAdd);
static const unsigned int kVersion = 1;
static void UpdateConfigFromBinary(int fd, const std::vector<uint64_t> &counts, Config &config) {
Quant::UpdateConfigFromBinary(fd, counts, config);
util::AdvanceOrThrow(fd, Quant::Size(counts.size(), config) + Unigram::Size(counts[0]));
Bhiksha::UpdateConfigFromBinary(fd, config);
}
static std::size_t Size(const std::vector<uint64_t> &counts, const Config &config) {
std::size_t ret = Quant::Size(counts.size(), config) + Unigram::Size(counts[0]);
for (unsigned char i = 1; i < counts.size() - 1; ++i) {
ret += Middle::Size(Quant::MiddleBits(config), counts[i], counts[0], counts[i+1], config);
}
return ret + Longest::Size(Quant::LongestBits(config), counts.back(), counts[0]);
}
TrieSearch() : middle_begin_(NULL), middle_end_(NULL) {}
~TrieSearch() { FreeMiddles(); }
uint8_t *SetupMemory(uint8_t *start, const std::vector<uint64_t> &counts, const Config &config);
void LoadedBinary();
void InitializeFromARPA(const char *file, util::FilePiece &f, std::vector<uint64_t> &counts, const Config &config, SortedVocabulary &vocab, Backing &backing);
unsigned char Order() const {
return middle_end_ - middle_begin_ + 2;
}
ProbBackoff &UnknownUnigram() { return unigram_.Unknown(); }
UnigramPointer LookupUnigram(WordIndex word, Node &next, bool &independent_left, uint64_t &extend_left) const {
extend_left = static_cast<uint64_t>(word);
UnigramPointer ret(unigram_.Find(word, next));
independent_left = (next.begin == next.end);
return ret;
}
MiddlePointer Unpack(uint64_t extend_pointer, unsigned char extend_length, Node &node) const {
return MiddlePointer(quant_, extend_length - 2, middle_begin_[extend_length - 2].ReadEntry(extend_pointer, node));
}
MiddlePointer LookupMiddle(unsigned char order_minus_2, WordIndex word, Node &node, bool &independent_left, uint64_t &extend_left) const {
util::BitAddress address(middle_begin_[order_minus_2].Find(word, node, extend_left));
independent_left = (address.base == NULL) || (node.begin == node.end);
return MiddlePointer(quant_, order_minus_2, address);
}
LongestPointer LookupLongest(WordIndex word, const Node &node) const {
return LongestPointer(quant_, longest_.Find(word, node));
}
bool FastMakeNode(const WordIndex *begin, const WordIndex *end, Node &node) const {
assert(begin != end);
bool independent_left;
uint64_t ignored;
LookupUnigram(*begin, node, independent_left, ignored);
for (const WordIndex *i = begin + 1; i < end; ++i) {
if (independent_left || !LookupMiddle(i - begin - 1, *i, node, independent_left, ignored).Found()) return false;
}
return true;
}
private:
friend void BuildTrie<Quant, Bhiksha>(SortedFiles &files, std::vector<uint64_t> &counts, const Config &config, TrieSearch<Quant, Bhiksha> &out, Quant &quant, const SortedVocabulary &vocab, Backing &backing);
// Middles are managed manually so we can delay construction and they don't have to be copyable.
void FreeMiddles() {
for (const Middle *i = middle_begin_; i != middle_end_; ++i) {
i->~Middle();
}
free(middle_begin_);
}
typedef trie::BitPackedMiddle<Bhiksha> Middle;
typedef trie::BitPackedLongest Longest;
Longest longest_;
Middle *middle_begin_, *middle_end_;
Quant quant_;
typedef ::lm::ngram::trie::Unigram Unigram;
Unigram unigram_;
};
} // namespace trie
} // namespace ngram
} // namespace lm
#endif // LM_SEARCH_TRIE__

122
lm/state.hh
View File

@ -1,122 +0,0 @@
#ifndef LM_STATE__
#define LM_STATE__
#include "lm/word_index.hh"
#include "util/murmur_hash.hh"
#include <string.h>
namespace lm {
namespace ngram {
// This is a POD but if you want memcmp to return the same as operator==, call
// ZeroRemaining first.
class State {
public:
bool operator==(const State &other) const {
if (length != other.length) return false;
return !memcmp(words, other.words, length * sizeof(WordIndex));
}
// Three way comparison function.
int Compare(const State &other) const {
if (length != other.length) return length < other.length ? -1 : 1;
return memcmp(words, other.words, length * sizeof(WordIndex));
}
bool operator<(const State &other) const {
if (length != other.length) return length < other.length;
return memcmp(words, other.words, length * sizeof(WordIndex)) < 0;
}
// Call this before using raw memcmp.
void ZeroRemaining() {
for (unsigned char i = length; i < KENLM_MAX_ORDER - 1; ++i) {
words[i] = 0;
backoff[i] = 0.0;
}
}
unsigned char Length() const { return length; }
// You shouldn't need to touch anything below this line, but the members are public so FullState will qualify as a POD.
// This order minimizes total size of the struct if WordIndex is 64 bit, float is 32 bit, and alignment of 64 bit integers is 64 bit.
WordIndex words[KENLM_MAX_ORDER - 1];
float backoff[KENLM_MAX_ORDER - 1];
unsigned char length;
};
inline uint64_t hash_value(const State &state, uint64_t seed = 0) {
return util::MurmurHashNative(state.words, sizeof(WordIndex) * state.length, seed);
}
struct Left {
bool operator==(const Left &other) const {
return
(length == other.length) &&
pointers[length - 1] == other.pointers[length - 1] &&
full == other.full;
}
int Compare(const Left &other) const {
if (length < other.length) return -1;
if (length > other.length) return 1;
if (pointers[length - 1] > other.pointers[length - 1]) return 1;
if (pointers[length - 1] < other.pointers[length - 1]) return -1;
return (int)full - (int)other.full;
}
bool operator<(const Left &other) const {
return Compare(other) == -1;
}
void ZeroRemaining() {
for (uint64_t * i = pointers + length; i < pointers + KENLM_MAX_ORDER - 1; ++i)
*i = 0;
}
uint64_t pointers[KENLM_MAX_ORDER - 1];
unsigned char length;
bool full;
};
inline uint64_t hash_value(const Left &left) {
unsigned char add[2];
add[0] = left.length;
add[1] = left.full;
return util::MurmurHashNative(add, 2, left.length ? left.pointers[left.length - 1] : 0);
}
struct ChartState {
bool operator==(const ChartState &other) {
return (right == other.right) && (left == other.left);
}
int Compare(const ChartState &other) const {
int lres = left.Compare(other.left);
if (lres) return lres;
return right.Compare(other.right);
}
bool operator<(const ChartState &other) const {
return Compare(other) == -1;
}
void ZeroRemaining() {
left.ZeroRemaining();
right.ZeroRemaining();
}
Left left;
State right;
};
inline uint64_t hash_value(const ChartState &state) {
return hash_value(state.right, hash_value(state.left));
}
} // namespace ngram
} // namespace lm
#endif // LM_STATE__

124
lm/test.arpa
View File

@ -1,124 +0,0 @@
\data\
ngram 1=37
ngram 2=47
ngram 3=11
ngram 4=6
ngram 5=4
\1-grams:
-1.383514 , -0.30103
-1.139057 . -0.845098
-1.029493 </s>
-99 <s> -0.4149733
-1.995635 <unk> -20
-1.285941 a -0.69897
-1.687872 also -0.30103
-1.687872 beyond -0.30103
-1.687872 biarritz -0.30103
-1.687872 call -0.30103
-1.687872 concerns -0.30103
-1.687872 consider -0.30103
-1.687872 considering -0.30103
-1.687872 for -0.30103
-1.509559 higher -0.30103
-1.687872 however -0.30103
-1.687872 i -0.30103
-1.687872 immediate -0.30103
-1.687872 in -0.30103
-1.687872 is -0.30103
-1.285941 little -0.69897
-1.383514 loin -0.30103
-1.687872 look -0.30103
-1.285941 looking -0.4771212
-1.206319 more -0.544068
-1.509559 on -0.4771212
-1.509559 screening -0.4771212
-1.687872 small -0.30103
-1.687872 the -0.30103
-1.687872 to -0.30103
-1.687872 watch -0.30103
-1.687872 watching -0.30103
-1.687872 what -0.30103
-1.687872 would -0.30103
-3.141592 foo
-2.718281 bar 3.0
-6.535897 baz -0.0
\2-grams:
-0.6925742 , .
-0.7522095 , however
-0.7522095 , is
-0.0602359 . </s>
-0.4846522 <s> looking -0.4771214
-1.051485 <s> screening
-1.07153 <s> the
-1.07153 <s> watching
-1.07153 <s> what
-0.09132547 a little -0.69897
-0.2922095 also call
-0.2922095 beyond immediate
-0.2705918 biarritz .
-0.2922095 call for
-0.2922095 concerns in
-0.2922095 consider watch
-0.2922095 considering consider
-0.2834328 for ,
-0.5511513 higher more
-0.5845945 higher small
-0.2834328 however ,
-0.2922095 i would
-0.2922095 immediate concerns
-0.2922095 in biarritz
-0.2922095 is to
-0.09021038 little more -0.1998621
-0.7273645 loin ,
-0.6925742 loin .
-0.6708385 loin </s>
-0.2922095 look beyond
-0.4638903 looking higher
-0.4638903 looking on -0.4771212
-0.5136299 more . -0.4771212
-0.3561665 more loin
-0.1649931 on a -0.4771213
-0.1649931 screening a -0.4771213
-0.2705918 small .
-0.287799 the screening
-0.2922095 to look
-0.2622373 watch </s>
-0.2922095 watching considering
-0.2922095 what i
-0.2922095 would also
-2 also would -6
-15 <unk> <unk> -2
-4 <unk> however -1
-6 foo bar
\3-grams:
-0.01916512 more . </s>
-0.0283603 on a little -0.4771212
-0.0283603 screening a little -0.4771212
-0.01660496 a little more -0.09409451
-0.3488368 <s> looking higher
-0.3488368 <s> looking on -0.4771212
-0.1892331 little more loin
-0.04835128 looking on a -0.4771212
-3 also would consider -7
-6 <unk> however <unk> -12
-7 to look good
\4-grams:
-0.009249173 looking on a little -0.4771212
-0.005464747 on a little more -0.4771212
-0.005464747 screening a little more
-0.1453306 a little more loin
-0.01552657 <s> looking on a -0.4771212
-4 also would consider higher -8
\5-grams:
-0.003061223 <s> looking on a little
-0.001813953 looking on a little more
-0.0432557 on a little more loin
-5 also would consider higher looking
\end\

10
lm/test.sh
View File

@ -1,10 +0,0 @@
#!/bin/bash
#Run tests. Requires Boost.
cd "$(dirname "$0")/.."
set -e
lm/compile.sh
for i in util/{bit_packing,file_piece,joint_sort,key_value_packing,probing_hash_table,sorted_uniform,tokenize_piece}_test lm/{model,left}_test; do
g++ -I. -O3 $CXXFLAGS $i.cc {lm,util}/*.o -lboost_test_exec_monitor -lz -o $i
pushd $(dirname $i) >/dev/null && ./$(basename $i) || echo "$i failed"; popd >/dev/null
done

120
lm/test_nounk.arpa
View File

@ -1,120 +0,0 @@
\data\
ngram 1=36
ngram 2=45
ngram 3=10
ngram 4=6
ngram 5=4
\1-grams:
-1.383514 , -0.30103
-1.139057 . -0.845098
-1.029493 </s>
-99 <s> -0.4149733
-1.285941 a -0.69897
-1.687872 also -0.30103
-1.687872 beyond -0.30103
-1.687872 biarritz -0.30103
-1.687872 call -0.30103
-1.687872 concerns -0.30103
-1.687872 consider -0.30103
-1.687872 considering -0.30103
-1.687872 for -0.30103
-1.509559 higher -0.30103
-1.687872 however -0.30103
-1.687872 i -0.30103
-1.687872 immediate -0.30103
-1.687872 in -0.30103
-1.687872 is -0.30103
-1.285941 little -0.69897
-1.383514 loin -0.30103
-1.687872 look -0.30103
-1.285941 looking -0.4771212
-1.206319 more -0.544068
-1.509559 on -0.4771212
-1.509559 screening -0.4771212
-1.687872 small -0.30103
-1.687872 the -0.30103
-1.687872 to -0.30103
-1.687872 watch -0.30103
-1.687872 watching -0.30103
-1.687872 what -0.30103
-1.687872 would -0.30103
-3.141592 foo
-2.718281 bar 3.0
-6.535897 baz -0.0
\2-grams:
-0.6925742 , .
-0.7522095 , however
-0.7522095 , is
-0.0602359 . </s>
-0.4846522 <s> looking -0.4771214
-1.051485 <s> screening
-1.07153 <s> the
-1.07153 <s> watching
-1.07153 <s> what
-0.09132547 a little -0.69897
-0.2922095 also call
-0.2922095 beyond immediate
-0.2705918 biarritz .
-0.2922095 call for
-0.2922095 concerns in
-0.2922095 consider watch
-0.2922095 considering consider
-0.2834328 for ,
-0.5511513 higher more
-0.5845945 higher small
-0.2834328 however ,
-0.2922095 i would
-0.2922095 immediate concerns
-0.2922095 in biarritz
-0.2922095 is to
-0.09021038 little more -0.1998621
-0.7273645 loin ,
-0.6925742 loin .
-0.6708385 loin </s>
-0.2922095 look beyond
-0.4638903 looking higher
-0.4638903 looking on -0.4771212
-0.5136299 more . -0.4771212
-0.3561665 more loin
-0.1649931 on a -0.4771213
-0.1649931 screening a -0.4771213
-0.2705918 small .
-0.287799 the screening
-0.2922095 to look
-0.2622373 watch </s>
-0.2922095 watching considering
-0.2922095 what i
-0.2922095 would also
-2 also would -6
-6 foo bar
\3-grams:
-0.01916512 more . </s>
-0.0283603 on a little -0.4771212
-0.0283603 screening a little -0.4771212
-0.01660496 a little more -0.09409451
-0.3488368 <s> looking higher
-0.3488368 <s> looking on -0.4771212
-0.1892331 little more loin
-0.04835128 looking on a -0.4771212
-3 also would consider -7
-7 to look good
\4-grams:
-0.009249173 looking on a little -0.4771212
-0.005464747 on a little more -0.4771212
-0.005464747 screening a little more
-0.1453306 a little more loin
-0.01552657 <s> looking on a -0.4771212
-4 also would consider higher -8
\5-grams:
-0.003061223 <s> looking on a little
-0.001813953 looking on a little more
-0.0432557 on a little more loin
-5 also would consider higher looking
\end\

128
lm/trie.cc
View File

@ -1,128 +0,0 @@
#include "lm/trie.hh"
#include "lm/bhiksha.hh"
#include "util/bit_packing.hh"
#include "util/exception.hh"
#include "util/sorted_uniform.hh"
#include <assert.h>
namespace lm {
namespace ngram {
namespace trie {
namespace {
class KeyAccessor {
public:
KeyAccessor(const void *base, uint64_t key_mask, uint8_t key_bits, uint8_t total_bits)
: base_(reinterpret_cast<const uint8_t*>(base)), key_mask_(key_mask), key_bits_(key_bits), total_bits_(total_bits) {}
typedef uint64_t Key;
Key operator()(uint64_t index) const {
return util::ReadInt57(base_, index * static_cast<uint64_t>(total_bits_), key_bits_, key_mask_);
}
private:
const uint8_t *const base_;
const WordIndex key_mask_;
const uint8_t key_bits_, total_bits_;
};
bool FindBitPacked(const void *base, uint64_t key_mask, uint8_t key_bits, uint8_t total_bits, uint64_t begin_index, uint64_t end_index, const uint64_t max_vocab, const uint64_t key, uint64_t &at_index) {
KeyAccessor accessor(base, key_mask, key_bits, total_bits);
if (!util::BoundedSortedUniformFind<uint64_t, KeyAccessor, util::PivotSelect<sizeof(WordIndex)>::T>(accessor, begin_index - 1, (uint64_t)0, end_index, max_vocab, key, at_index)) return false;
return true;
}
} // namespace
std::size_t BitPacked::BaseSize(uint64_t entries, uint64_t max_vocab, uint8_t remaining_bits) {
uint8_t total_bits = util::RequiredBits(max_vocab) + remaining_bits;
// Extra entry for next pointer at the end.
// +7 then / 8 to round up bits and convert to bytes
// +sizeof(uint64_t) so that ReadInt57 etc don't go segfault.
// Note that this waste is O(order), not O(number of ngrams).
return ((1 + entries) * total_bits + 7) / 8 + sizeof(uint64_t);
}
void BitPacked::BaseInit(void *base, uint64_t max_vocab, uint8_t remaining_bits) {
util::BitPackingSanity();
word_bits_ = util::RequiredBits(max_vocab);
word_mask_ = (1ULL << word_bits_) - 1ULL;
if (word_bits_ > 57) UTIL_THROW(util::Exception, "Sorry, word indices more than " << (1ULL << 57) << " are not implemented. Edit util/bit_packing.hh and fix the bit packing functions.");
total_bits_ = word_bits_ + remaining_bits;
base_ = static_cast<uint8_t*>(base);
insert_index_ = 0;
max_vocab_ = max_vocab;
}
template <class Bhiksha> std::size_t BitPackedMiddle<Bhiksha>::Size(uint8_t quant_bits, uint64_t entries, uint64_t max_vocab, uint64_t max_ptr, const Config &config) {
return Bhiksha::Size(entries + 1, max_ptr, config) + BaseSize(entries, max_vocab, quant_bits + Bhiksha::InlineBits(entries + 1, max_ptr, config));
}
template <class Bhiksha> BitPackedMiddle<Bhiksha>::BitPackedMiddle(void *base, uint8_t quant_bits, uint64_t entries, uint64_t max_vocab, uint64_t max_next, const BitPacked &next_source, const Config &config) :
BitPacked(),
quant_bits_(quant_bits),
// If the offset of the method changes, also change TrieSearch::UpdateConfigFromBinary.
bhiksha_(base, entries + 1, max_next, config),
next_source_(&next_source) {
if (entries + 1 >= (1ULL << 57) || (max_next >= (1ULL << 57))) UTIL_THROW(util::Exception, "Sorry, this does not support more than " << (1ULL << 57) << " n-grams of a particular order. Edit util/bit_packing.hh and fix the bit packing functions.");
BaseInit(reinterpret_cast<uint8_t*>(base) + Bhiksha::Size(entries + 1, max_next, config), max_vocab, quant_bits_ + bhiksha_.InlineBits());
}
template <class Bhiksha> util::BitAddress BitPackedMiddle<Bhiksha>::Insert(WordIndex word) {
assert(word <= word_mask_);
uint64_t at_pointer = insert_index_ * total_bits_;
util::WriteInt57(base_, at_pointer, word_bits_, word);
at_pointer += word_bits_;
util::BitAddress ret(base_, at_pointer);
at_pointer += quant_bits_;
uint64_t next = next_source_->InsertIndex();
bhiksha_.WriteNext(base_, at_pointer, insert_index_, next);
++insert_index_;
return ret;
}
template <class Bhiksha> util::BitAddress BitPackedMiddle<Bhiksha>::Find(WordIndex word, NodeRange &range, uint64_t &pointer) const {
uint64_t at_pointer;
if (!FindBitPacked(base_, word_mask_, word_bits_, total_bits_, range.begin, range.end, max_vocab_, word, at_pointer)) {
return util::BitAddress(NULL, 0);
}
pointer = at_pointer;
at_pointer *= total_bits_;
at_pointer += word_bits_;
bhiksha_.ReadNext(base_, at_pointer + quant_bits_, pointer, total_bits_, range);
return util::BitAddress(base_, at_pointer);
}
template <class Bhiksha> void BitPackedMiddle<Bhiksha>::FinishedLoading(uint64_t next_end, const Config &config) {
uint64_t last_next_write = (insert_index_ + 1) * total_bits_ - bhiksha_.InlineBits();
bhiksha_.WriteNext(base_, last_next_write, insert_index_ + 1, next_end);
bhiksha_.FinishedLoading(config);
}
util::BitAddress BitPackedLongest::Insert(WordIndex index) {
assert(index <= word_mask_);
uint64_t at_pointer = insert_index_ * total_bits_;
util::WriteInt57(base_, at_pointer, word_bits_, index);
at_pointer += word_bits_;
++insert_index_;
return util::BitAddress(base_, at_pointer);
}
util::BitAddress BitPackedLongest::Find(WordIndex word, const NodeRange &range) const {
uint64_t at_pointer;
if (!FindBitPacked(base_, word_mask_, word_bits_, total_bits_, range.begin, range.end, max_vocab_, word, at_pointer)) return util::BitAddress(NULL, 0);
at_pointer = at_pointer * total_bits_ + word_bits_;
return util::BitAddress(base_, at_pointer);
}
template class BitPackedMiddle<DontBhiksha>;
template class BitPackedMiddle<ArrayBhiksha>;
} // namespace trie
} // namespace ngram
} // namespace lm

155
lm/trie.hh
View File

@ -1,155 +0,0 @@
#ifndef LM_TRIE__
#define LM_TRIE__
#include "lm/weights.hh"
#include "lm/word_index.hh"
#include "util/bit_packing.hh"
#include <cstddef>
#include <stdint.h>
namespace lm {
namespace ngram {
struct Config;
namespace trie {
struct NodeRange {
uint64_t begin, end;
};
// TODO: if the number of unigrams is a concern, also bit pack these records.
struct UnigramValue {
ProbBackoff weights;
uint64_t next;
uint64_t Next() const { return next; }
};
class UnigramPointer {
public:
explicit UnigramPointer(const ProbBackoff &to) : to_(&to) {}
UnigramPointer() : to_(NULL) {}
bool Found() const { return to_ != NULL; }
float Prob() const { return to_->prob; }
float Backoff() const { return to_->backoff; }
float Rest() const { return Prob(); }
private:
const ProbBackoff *to_;
};
class Unigram {
public:
Unigram() {}
void Init(void *start) {
unigram_ = static_cast<UnigramValue*>(start);
}
static std::size_t Size(uint64_t count) {
// +1 in case unknown doesn't appear. +1 for the final next.
return (count + 2) * sizeof(UnigramValue);
}
const ProbBackoff &Lookup(WordIndex index) const { return unigram_[index].weights; }
ProbBackoff &Unknown() { return unigram_[0].weights; }
UnigramValue *Raw() {
return unigram_;
}
void LoadedBinary() {}
UnigramPointer Find(WordIndex word, NodeRange &next) const {
UnigramValue *val = unigram_ + word;
next.begin = val->next;
next.end = (val+1)->next;
return UnigramPointer(val->weights);
}
private:
UnigramValue *unigram_;
};
class BitPacked {
public:
BitPacked() {}
uint64_t InsertIndex() const {
return insert_index_;
}
protected:
static std::size_t BaseSize(uint64_t entries, uint64_t max_vocab, uint8_t remaining_bits);
void BaseInit(void *base, uint64_t max_vocab, uint8_t remaining_bits);
uint8_t word_bits_;
uint8_t total_bits_;
uint64_t word_mask_;
uint8_t *base_;
uint64_t insert_index_, max_vocab_;
};
template <class Bhiksha> class BitPackedMiddle : public BitPacked {
public:
static std::size_t Size(uint8_t quant_bits, uint64_t entries, uint64_t max_vocab, uint64_t max_next, const Config &config);
// next_source need not be initialized.
BitPackedMiddle(void *base, uint8_t quant_bits, uint64_t entries, uint64_t max_vocab, uint64_t max_next, const BitPacked &next_source, const Config &config);
util::BitAddress Insert(WordIndex word);
void FinishedLoading(uint64_t next_end, const Config &config);
void LoadedBinary() { bhiksha_.LoadedBinary(); }
util::BitAddress Find(WordIndex word, NodeRange &range, uint64_t &pointer) const;
util::BitAddress ReadEntry(uint64_t pointer, NodeRange &range) {
uint64_t addr = pointer * total_bits_;
addr += word_bits_;
bhiksha_.ReadNext(base_, addr + quant_bits_, pointer, total_bits_, range);
return util::BitAddress(base_, addr);
}
private:
uint8_t quant_bits_;
Bhiksha bhiksha_;
const BitPacked *next_source_;
};
class BitPackedLongest : public BitPacked {
public:
static std::size_t Size(uint8_t quant_bits, uint64_t entries, uint64_t max_vocab) {
return BaseSize(entries, max_vocab, quant_bits);
}
BitPackedLongest() {}
void Init(void *base, uint8_t quant_bits, uint64_t max_vocab) {
BaseInit(base, max_vocab, quant_bits);
}
void LoadedBinary() {}
util::BitAddress Insert(WordIndex word);
util::BitAddress Find(WordIndex word, const NodeRange &node) const;
private:
uint8_t quant_bits_;
};
} // namespace trie
} // namespace ngram
} // namespace lm
#endif // LM_TRIE__

298
lm/trie_sort.cc
View File

@ -1,298 +0,0 @@
#include "lm/trie_sort.hh"
#include "lm/config.hh"
#include "lm/lm_exception.hh"
#include "lm/read_arpa.hh"
#include "lm/vocab.hh"
#include "lm/weights.hh"
#include "lm/word_index.hh"
#include "util/file_piece.hh"
#include "util/mmap.hh"
#include "util/proxy_iterator.hh"
#include "util/sized_iterator.hh"
#include <algorithm>
#include <cstring>
#include <cstdio>
#include <cstdlib>
#include <deque>
#include <limits>
#include <vector>
namespace lm {
namespace ngram {
namespace trie {
void WriteOrThrow(FILE *to, const void *data, size_t size) {
assert(size);
if (1 != std::fwrite(data, size, 1, to)) UTIL_THROW(util::ErrnoException, "Short write; requested size " << size);
}
namespace {
typedef util::SizedIterator NGramIter;
// Proxy for an entry except there is some extra cruft between the entries. This is used to sort (n-1)-grams using the same memory as the sorted n-grams.
class PartialViewProxy {
public:
PartialViewProxy() : attention_size_(0), inner_() {}
PartialViewProxy(void *ptr, std::size_t block_size, std::size_t attention_size) : attention_size_(attention_size), inner_(ptr, block_size) {}
operator std::string() const {
return std::string(reinterpret_cast<const char*>(inner_.Data()), attention_size_);
}
PartialViewProxy &operator=(const PartialViewProxy &from) {
memcpy(inner_.Data(), from.inner_.Data(), attention_size_);
return *this;
}
PartialViewProxy &operator=(const std::string &from) {
memcpy(inner_.Data(), from.data(), attention_size_);
return *this;
}
const void *Data() const { return inner_.Data(); }
void *Data() { return inner_.Data(); }
private:
friend class util::ProxyIterator<PartialViewProxy>;
typedef std::string value_type;
const std::size_t attention_size_;
typedef util::SizedInnerIterator InnerIterator;
InnerIterator &Inner() { return inner_; }
const InnerIterator &Inner() const { return inner_; }
InnerIterator inner_;
};
typedef util::ProxyIterator<PartialViewProxy> PartialIter;
FILE *DiskFlush(const void *mem_begin, const void *mem_end, const util::TempMaker &maker) {
util::scoped_fd file(maker.Make());
util::WriteOrThrow(file.get(), mem_begin, (uint8_t*)mem_end - (uint8_t*)mem_begin);
return util::FDOpenOrThrow(file);
}
FILE *WriteContextFile(uint8_t *begin, uint8_t *end, const util::TempMaker &maker, std::size_t entry_size, unsigned char order) {
const size_t context_size = sizeof(WordIndex) * (order - 1);
// Sort just the contexts using the same memory.
PartialIter context_begin(PartialViewProxy(begin + sizeof(WordIndex), entry_size, context_size));
PartialIter context_end(PartialViewProxy(end + sizeof(WordIndex), entry_size, context_size));
#if defined(_WIN32) || defined(_WIN64)
std::stable_sort
#else
std::sort
#endif
(context_begin, context_end, util::SizedCompare<EntryCompare, PartialViewProxy>(EntryCompare(order - 1)));
util::scoped_FILE out(maker.MakeFile());
// Write out to file and uniqueify at the same time. Could have used unique_copy if there was an appropriate OutputIterator.
if (context_begin == context_end) return out.release();
PartialIter i(context_begin);
WriteOrThrow(out.get(), i->Data(), context_size);
const void *previous = i->Data();
++i;
for (; i != context_end; ++i) {
if (memcmp(previous, i->Data(), context_size)) {
WriteOrThrow(out.get(), i->Data(), context_size);
previous = i->Data();
}
}
return out.release();
}
struct ThrowCombine {
void operator()(std::size_t /*entry_size*/, const void * /*first*/, const void * /*second*/, FILE * /*out*/) const {
UTIL_THROW(FormatLoadException, "Duplicate n-gram detected.");
}
};
// Useful for context files that just contain records with no value.
struct FirstCombine {
void operator()(std::size_t entry_size, const void *first, const void * /*second*/, FILE *out) const {
WriteOrThrow(out, first, entry_size);
}
};
template <class Combine> FILE *MergeSortedFiles(FILE *first_file, FILE *second_file, const util::TempMaker &maker, std::size_t weights_size, unsigned char order, const Combine &combine) {
std::size_t entry_size = sizeof(WordIndex) * order + weights_size;
RecordReader first, second;
first.Init(first_file, entry_size);
second.Init(second_file, entry_size);
util::scoped_FILE out_file(maker.MakeFile());
EntryCompare less(order);
while (first && second) {
if (less(first.Data(), second.Data())) {
WriteOrThrow(out_file.get(), first.Data(), entry_size);
++first;
} else if (less(second.Data(), first.Data())) {
WriteOrThrow(out_file.get(), second.Data(), entry_size);
++second;
} else {
combine(entry_size, first.Data(), second.Data(), out_file.get());
++first; ++second;
}
}
for (RecordReader &remains = (first ? first : second); remains; ++remains) {
WriteOrThrow(out_file.get(), remains.Data(), entry_size);
}
return out_file.release();
}
} // namespace
void RecordReader::Init(FILE *file, std::size_t entry_size) {
entry_size_ = entry_size;
data_.reset(malloc(entry_size));
UTIL_THROW_IF(!data_.get(), util::ErrnoException, "Failed to malloc read buffer");
file_ = file;
if (file) {
rewind(file);
remains_ = true;
++*this;
} else {
remains_ = false;
}
}
void RecordReader::Overwrite(const void *start, std::size_t amount) {
long internal = (uint8_t*)start - (uint8_t*)data_.get();
UTIL_THROW_IF(fseek(file_, internal - entry_size_, SEEK_CUR), util::ErrnoException, "Couldn't seek backwards for revision");
WriteOrThrow(file_, start, amount);
long forward = entry_size_ - internal - amount;
#if !defined(_WIN32) && !defined(_WIN64)
if (forward)
#endif
UTIL_THROW_IF(fseek(file_, forward, SEEK_CUR), util::ErrnoException, "Couldn't seek forwards past revision");
}
void RecordReader::Rewind() {
if (file_) {
rewind(file_);
remains_ = true;
++*this;
} else {
remains_ = false;
}
}
SortedFiles::SortedFiles(const Config &config, util::FilePiece &f, std::vector<uint64_t> &counts, size_t buffer, const std::string &file_prefix, SortedVocabulary &vocab) {
util::TempMaker maker(file_prefix);
PositiveProbWarn warn(config.positive_log_probability);
unigram_.reset(maker.Make());
{
// In case <unk> appears.
size_t size_out = (counts[0] + 1) * sizeof(ProbBackoff);
util::scoped_mmap unigram_mmap(util::MapZeroedWrite(unigram_.get(), size_out), size_out);
Read1Grams(f, counts[0], vocab, reinterpret_cast<ProbBackoff*>(unigram_mmap.get()), warn);
CheckSpecials(config, vocab);
if (!vocab.SawUnk()) ++counts[0];
}
// Only use as much buffer as we need.
size_t buffer_use = 0;
for (unsigned int order = 2; order < counts.size(); ++order) {
buffer_use = std::max<size_t>(buffer_use, static_cast<size_t>((sizeof(WordIndex) * order + 2 * sizeof(float)) * counts[order - 1]));
}
buffer_use = std::max<size_t>(buffer_use, static_cast<size_t>((sizeof(WordIndex) * counts.size() + sizeof(float)) * counts.back()));
buffer = std::min<size_t>(buffer, buffer_use);
util::scoped_malloc mem;
mem.reset(malloc(buffer));
if (!mem.get()) UTIL_THROW(util::ErrnoException, "malloc failed for sort buffer size " << buffer);
for (unsigned char order = 2; order <= counts.size(); ++order) {
ConvertToSorted(f, vocab, counts, maker, order, warn, mem.get(), buffer);
}
ReadEnd(f);
}
namespace {
class Closer {
public:
explicit Closer(std::deque<FILE*> &files) : files_(files) {}
~Closer() {
for (std::deque<FILE*>::iterator i = files_.begin(); i != files_.end(); ++i) {
util::scoped_FILE deleter(*i);
}
}
void PopFront() {
util::scoped_FILE deleter(files_.front());
files_.pop_front();
}
private:
std::deque<FILE*> &files_;
};
} // namespace
void SortedFiles::ConvertToSorted(util::FilePiece &f, const SortedVocabulary &vocab, const std::vector<uint64_t> &counts, const util::TempMaker &maker, unsigned char order, PositiveProbWarn &warn, void *mem, std::size_t mem_size) {
ReadNGramHeader(f, order);
const size_t count = counts[order - 1];
// Size of weights. Does it include backoff?
const size_t words_size = sizeof(WordIndex) * order;
const size_t weights_size = sizeof(float) + ((order == counts.size()) ? 0 : sizeof(float));
const size_t entry_size = words_size + weights_size;
const size_t batch_size = std::min(count, mem_size / entry_size);
uint8_t *const begin = reinterpret_cast<uint8_t*>(mem);
std::deque<FILE*> files, contexts;
Closer files_closer(files), contexts_closer(contexts);
for (std::size_t batch = 0, done = 0; done < count; ++batch) {
uint8_t *out = begin;
uint8_t *out_end = out + std::min(count - done, batch_size) * entry_size;
if (order == counts.size()) {
for (; out != out_end; out += entry_size) {
ReadNGram(f, order, vocab, reinterpret_cast<WordIndex*>(out), *reinterpret_cast<Prob*>(out + words_size), warn);
}
} else {
for (; out != out_end; out += entry_size) {
ReadNGram(f, order, vocab, reinterpret_cast<WordIndex*>(out), *reinterpret_cast<ProbBackoff*>(out + words_size), warn);
}
}
// Sort full records by full n-gram.
util::SizedProxy proxy_begin(begin, entry_size), proxy_end(out_end, entry_size);
// parallel_sort uses too much RAM. TODO: figure out why windows sort doesn't like my proxies.
#if defined(_WIN32) || defined(_WIN64)
std::stable_sort
#else
std::sort
#endif
(NGramIter(proxy_begin), NGramIter(proxy_end), util::SizedCompare<EntryCompare>(EntryCompare(order)));
files.push_back(DiskFlush(begin, out_end, maker));
contexts.push_back(WriteContextFile(begin, out_end, maker, entry_size, order));
done += (out_end - begin) / entry_size;
}
// All individual files created. Merge them.
while (files.size() > 1) {
files.push_back(MergeSortedFiles(files[0], files[1], maker, weights_size, order, ThrowCombine()));
files_closer.PopFront();
files_closer.PopFront();
contexts.push_back(MergeSortedFiles(contexts[0], contexts[1], maker, 0, order - 1, FirstCombine()));
contexts_closer.PopFront();
contexts_closer.PopFront();
}
if (!files.empty()) {
// Steal from closers.
full_[order - 2].reset(files.front());
files.pop_front();
context_[order - 2].reset(contexts.front());
contexts.pop_front();
}
}
} // namespace trie
} // namespace ngram
} // namespace lm

116
lm/trie_sort.hh
View File

@ -1,116 +0,0 @@
// Step of trie builder: create sorted files.
#ifndef LM_TRIE_SORT__
#define LM_TRIE_SORT__
#include "lm/word_index.hh"
#include "util/file.hh"
#include "util/scoped.hh"
#include <cstddef>
#include <functional>
#include <string>
#include <vector>
#include <stdint.h>
namespace util {
class FilePiece;
class TempMaker;
} // namespace util
namespace lm {
class PositiveProbWarn;
namespace ngram {
class SortedVocabulary;
struct Config;
namespace trie {
void WriteOrThrow(FILE *to, const void *data, size_t size);
class EntryCompare : public std::binary_function<const void*, const void*, bool> {
public:
explicit EntryCompare(unsigned char order) : order_(order) {}
bool operator()(const void *first_void, const void *second_void) const {
const WordIndex *first = static_cast<const WordIndex*>(first_void);
const WordIndex *second = static_cast<const WordIndex*>(second_void);
const WordIndex *end = first + order_;
for (; first != end; ++first, ++second) {
if (*first < *second) return true;
if (*first > *second) return false;
}
return false;
}
private:
unsigned char order_;
};
class RecordReader {
public:
RecordReader() : remains_(true) {}
void Init(FILE *file, std::size_t entry_size);
void *Data() { return data_.get(); }
const void *Data() const { return data_.get(); }
RecordReader &operator++() {
std::size_t ret = fread(data_.get(), entry_size_, 1, file_);
if (!ret) {
UTIL_THROW_IF(!feof(file_), util::ErrnoException, "Error reading temporary file");
remains_ = false;
}
return *this;
}
operator bool() const { return remains_; }
void Rewind();
std::size_t EntrySize() const { return entry_size_; }
void Overwrite(const void *start, std::size_t amount);
private:
FILE *file_;
util::scoped_malloc data_;
bool remains_;
std::size_t entry_size_;
};
class SortedFiles {
public:
// Build from ARPA
SortedFiles(const Config &config, util::FilePiece &f, std::vector<uint64_t> &counts, std::size_t buffer, const std::string &file_prefix, SortedVocabulary &vocab);
int StealUnigram() {
return unigram_.release();
}
FILE *Full(unsigned char order) {
return full_[order - 2].get();
}
FILE *Context(unsigned char of_order) {
return context_[of_order - 2].get();
}
private:
void ConvertToSorted(util::FilePiece &f, const SortedVocabulary &vocab, const std::vector<uint64_t> &counts, const util::TempMaker &maker, unsigned char order, PositiveProbWarn &warn, void *mem, std::size_t mem_size);
util::scoped_fd unigram_;
util::scoped_FILE full_[KENLM_MAX_ORDER - 1], context_[KENLM_MAX_ORDER - 1];
};
} // namespace trie
} // namespace ngram
} // namespace lm
#endif // LM_TRIE_SORT__

157
lm/value.hh
View File

@ -1,157 +0,0 @@
#ifndef LM_VALUE__
#define LM_VALUE__
#include "lm/model_type.hh"
#include "lm/value_build.hh"
#include "lm/weights.hh"
#include "util/bit_packing.hh"
#include <stdint.h>
namespace lm {
namespace ngram {
// Template proxy for probing unigrams and middle.
template <class Weights> class GenericProbingProxy {
public:
explicit GenericProbingProxy(const Weights &to) : to_(&to) {}
GenericProbingProxy() : to_(0) {}
bool Found() const { return to_ != 0; }
float Prob() const {
util::FloatEnc enc;
enc.f = to_->prob;
enc.i |= util::kSignBit;
return enc.f;
}
float Backoff() const { return to_->backoff; }
bool IndependentLeft() const {
util::FloatEnc enc;
enc.f = to_->prob;
return enc.i & util::kSignBit;
}
protected:
const Weights *to_;
};
// Basic proxy for trie unigrams.
template <class Weights> class GenericTrieUnigramProxy {
public:
explicit GenericTrieUnigramProxy(const Weights &to) : to_(&to) {}
GenericTrieUnigramProxy() : to_(0) {}
bool Found() const { return to_ != 0; }
float Prob() const { return to_->prob; }
float Backoff() const { return to_->backoff; }
float Rest() const { return Prob(); }
protected:
const Weights *to_;
};
struct BackoffValue {
typedef ProbBackoff Weights;
static const ModelType kProbingModelType = PROBING;
class ProbingProxy : public GenericProbingProxy<Weights> {
public:
explicit ProbingProxy(const Weights &to) : GenericProbingProxy<Weights>(to) {}
ProbingProxy() {}
float Rest() const { return Prob(); }
};
class TrieUnigramProxy : public GenericTrieUnigramProxy<Weights> {
public:
explicit TrieUnigramProxy(const Weights &to) : GenericTrieUnigramProxy<Weights>(to) {}
TrieUnigramProxy() {}
float Rest() const { return Prob(); }
};
struct ProbingEntry {
typedef uint64_t Key;
typedef Weights Value;
uint64_t key;
ProbBackoff value;
uint64_t GetKey() const { return key; }
};
struct TrieUnigramValue {
Weights weights;
uint64_t next;
uint64_t Next() const { return next; }
};
const static bool kDifferentRest = false;
template <class Model, class C> void Callback(const Config &, unsigned int, typename Model::Vocabulary &, C &callback) {
NoRestBuild build;
callback(build);
}
};
struct RestValue {
typedef RestWeights Weights;
static const ModelType kProbingModelType = REST_PROBING;
class ProbingProxy : public GenericProbingProxy<RestWeights> {
public:
explicit ProbingProxy(const Weights &to) : GenericProbingProxy<RestWeights>(to) {}
ProbingProxy() {}
float Rest() const { return to_->rest; }
};
class TrieUnigramProxy : public GenericTrieUnigramProxy<Weights> {
public:
explicit TrieUnigramProxy(const Weights &to) : GenericTrieUnigramProxy<Weights>(to) {}
TrieUnigramProxy() {}
float Rest() const { return to_->rest; }
};
// gcc 4.1 doesn't properly back dependent types :-(.
#pragma pack(push)
#pragma pack(4)
struct ProbingEntry {
typedef uint64_t Key;
typedef Weights Value;
Key key;
Value value;
Key GetKey() const { return key; }
};
struct TrieUnigramValue {
Weights weights;
uint64_t next;
uint64_t Next() const { return next; }
};
#pragma pack(pop)
const static bool kDifferentRest = true;
template <class Model, class C> void Callback(const Config &config, unsigned int order, typename Model::Vocabulary &vocab, C &callback) {
switch (config.rest_function) {
case Config::REST_MAX:
{
MaxRestBuild build;
callback(build);
}
break;
case Config::REST_LOWER:
{
LowerRestBuild<Model> build(config, order, vocab);
callback(build);
}
break;
}
}
};
} // namespace ngram
} // namespace lm
#endif // LM_VALUE__

58
lm/value_build.cc
View File

@ -1,58 +0,0 @@
#include "lm/value_build.hh"
#include "lm/model.hh"
#include "lm/read_arpa.hh"
namespace lm {
namespace ngram {
template <class Model> LowerRestBuild<Model>::LowerRestBuild(const Config &config, unsigned int order, const typename Model::Vocabulary &vocab) {
UTIL_THROW_IF(config.rest_lower_files.size() != order - 1, ConfigException, "This model has order " << order << " so there should be " << (order - 1) << " lower-order models for rest cost purposes.");
Config for_lower = config;
for_lower.rest_lower_files.clear();
// Unigram models aren't supported, so this is a custom loader.
// TODO: optimize the unigram loading?
{
util::FilePiece uni(config.rest_lower_files[0].c_str());
std::vector<uint64_t> number;
ReadARPACounts(uni, number);
UTIL_THROW_IF(number.size() != 1, FormatLoadException, "Expected the unigram model to have order 1, not " << number.size());
ReadNGramHeader(uni, 1);
unigrams_.resize(number[0]);
unigrams_[0] = config.unknown_missing_logprob;
PositiveProbWarn warn;
for (uint64_t i = 0; i < number[0]; ++i) {
WordIndex w;
Prob entry;
ReadNGram(uni, 1, vocab, &w, entry, warn);
unigrams_[w] = entry.prob;
}
}
try {
for (unsigned int i = 2; i < order; ++i) {
models_.push_back(new Model(config.rest_lower_files[i - 1].c_str(), for_lower));
UTIL_THROW_IF(models_.back()->Order() != i, FormatLoadException, "Lower order file " << config.rest_lower_files[i-1] << " should have order " << i);
}
} catch (...) {
for (typename std::vector<const Model*>::const_iterator i = models_.begin(); i != models_.end(); ++i) {
delete *i;
}
models_.clear();
throw;
}
// TODO: force/check same vocab.
}
template <class Model> LowerRestBuild<Model>::~LowerRestBuild() {
for (typename std::vector<const Model*>::const_iterator i = models_.begin(); i != models_.end(); ++i) {
delete *i;
}
}
template class LowerRestBuild<ProbingModel>;
} // namespace ngram
} // namespace lm

97
lm/value_build.hh
View File

@ -1,97 +0,0 @@
#ifndef LM_VALUE_BUILD__
#define LM_VALUE_BUILD__
#include "lm/weights.hh"
#include "lm/word_index.hh"
#include "util/bit_packing.hh"
#include <vector>
namespace lm {
namespace ngram {
struct Config;
struct BackoffValue;
struct RestValue;
class NoRestBuild {
public:
typedef BackoffValue Value;
NoRestBuild() {}
void SetRest(const WordIndex *, unsigned int, const Prob &/*prob*/) const {}
void SetRest(const WordIndex *, unsigned int, const ProbBackoff &) const {}
template <class Second> bool MarkExtends(ProbBackoff &weights, const Second &) const {
util::UnsetSign(weights.prob);
return false;
}
// Probing doesn't need to go back to unigram.
const static bool kMarkEvenLower = false;
};
class MaxRestBuild {
public:
typedef RestValue Value;
MaxRestBuild() {}
void SetRest(const WordIndex *, unsigned int, const Prob &/*prob*/) const {}
void SetRest(const WordIndex *, unsigned int, RestWeights &weights) const {
weights.rest = weights.prob;
util::SetSign(weights.rest);
}
bool MarkExtends(RestWeights &weights, const RestWeights &to) const {
util::UnsetSign(weights.prob);
if (weights.rest >= to.rest) return false;
weights.rest = to.rest;
return true;
}
bool MarkExtends(RestWeights &weights, const Prob &to) const {
util::UnsetSign(weights.prob);
if (weights.rest >= to.prob) return false;
weights.rest = to.prob;
return true;
}
// Probing does need to go back to unigram.
const static bool kMarkEvenLower = true;
};
template <class Model> class LowerRestBuild {
public:
typedef RestValue Value;
LowerRestBuild(const Config &config, unsigned int order, const typename Model::Vocabulary &vocab);
~LowerRestBuild();
void SetRest(const WordIndex *, unsigned int, const Prob &/*prob*/) const {}
void SetRest(const WordIndex *vocab_ids, unsigned int n, RestWeights &weights) const {
typename Model::State ignored;
if (n == 1) {
weights.rest = unigrams_[*vocab_ids];
} else {
weights.rest = models_[n-2]->FullScoreForgotState(vocab_ids + 1, vocab_ids + n, *vocab_ids, ignored).prob;
}
}
template <class Second> bool MarkExtends(RestWeights &weights, const Second &) const {
util::UnsetSign(weights.prob);
return false;
}
const static bool kMarkEvenLower = false;
std::vector<float> unigrams_;
std::vector<const Model*> models_;
};
} // namespace ngram
} // namespace lm
#endif // LM_VALUE_BUILD__

19
lm/virtual_interface.cc
View File

@ -1,19 +0,0 @@
#include "lm/virtual_interface.hh"
#include "lm/lm_exception.hh"
namespace lm {
namespace base {
Vocabulary::~Vocabulary() {}
void Vocabulary::SetSpecial(WordIndex begin_sentence, WordIndex end_sentence, WordIndex not_found) {
begin_sentence_ = begin_sentence;
end_sentence_ = end_sentence;
not_found_ = not_found;
}
Model::~Model() {}
} // namespace base
} // namespace lm

154
lm/virtual_interface.hh
View File

@ -1,154 +0,0 @@
#ifndef LM_VIRTUAL_INTERFACE__
#define LM_VIRTUAL_INTERFACE__
#include "lm/return.hh"
#include "lm/word_index.hh"
#include "util/string_piece.hh"
#include <string>
namespace lm {
namespace base {
template <class T, class U, class V> class ModelFacade;
/* Vocabulary interface. Call Index(string) and get a word index for use in
* calling Model. It provides faster convenience functions for <s>, </s>, and
* <unk> although you can also find these using Index.
*
* Some models do not load the mapping from index to string. If you need this,
* check if the model Vocabulary class implements such a function and access it
* directly.
*
* The Vocabulary object is always owned by the Model and can be retrieved from
* the Model using BaseVocabulary() for this abstract interface or
* GetVocabulary() for the actual implementation (in which case you'll need the
* actual implementation of the Model too).
*/
class Vocabulary {
public:
virtual ~Vocabulary();
WordIndex BeginSentence() const { return begin_sentence_; }
WordIndex EndSentence() const { return end_sentence_; }
WordIndex NotFound() const { return not_found_; }
/* Most implementations allow StringPiece lookups and need only override
* Index(StringPiece). SRI requires null termination and overrides all
* three methods.
*/
virtual WordIndex Index(const StringPiece &str) const = 0;
virtual WordIndex Index(const std::string &str) const {
return Index(StringPiece(str));
}
virtual WordIndex Index(const char *str) const {
return Index(StringPiece(str));
}
protected:
// Call SetSpecial afterward.
Vocabulary() {}
Vocabulary(WordIndex begin_sentence, WordIndex end_sentence, WordIndex not_found) {
SetSpecial(begin_sentence, end_sentence, not_found);
}
void SetSpecial(WordIndex begin_sentence, WordIndex end_sentence, WordIndex not_found);
WordIndex begin_sentence_, end_sentence_, not_found_;
private:
// Disable copy constructors. They're private and undefined.
// Ersatz boost::noncopyable.
Vocabulary(const Vocabulary &);
Vocabulary &operator=(const Vocabulary &);
};
/* There are two ways to access a Model.
*
*
* OPTION 1: Access the Model directly (e.g. lm::ngram::Model in model.hh).
*
* Every Model implements the scoring function:
* float Score(
* const Model::State &in_state,
* const WordIndex new_word,
* Model::State &out_state) const;
*
* It can also return the length of n-gram matched by the model:
* FullScoreReturn FullScore(
* const Model::State &in_state,
* const WordIndex new_word,
* Model::State &out_state) const;
*
*
* There are also accessor functions:
* const State &BeginSentenceState() const;
* const State &NullContextState() const;
* const Vocabulary &GetVocabulary() const;
* unsigned int Order() const;
*
* NB: In case you're wondering why the model implementation looks like it's
* missing these methods, see facade.hh.
*
* This is the fastest way to use a model and presents a normal State class to
* be included in a hypothesis state structure.
*
*
* OPTION 2: Use the virtual interface below.
*
* The virtual interface allow you to decide which Model to use at runtime
* without templatizing everything on the Model type. However, each Model has
* its own State class, so a single State cannot be efficiently provided (it
* would require using the maximum memory of any Model's State or memory
* allocation with each lookup). This means you become responsible for
* allocating memory with size StateSize() and passing it to the Score or
* FullScore functions provided here.
*
* For example, cdec has a std::string containing the entire state of a
* hypothesis. It can reserve StateSize bytes in this string for the model
* state.
*
* All the State objects are POD, so it's ok to use raw memory for storing
* State.
* in_state and out_state must not have the same address.
*/
class Model {
public:
virtual ~Model();
size_t StateSize() const { return state_size_; }
const void *BeginSentenceMemory() const { return begin_sentence_memory_; }
const void *NullContextMemory() const { return null_context_memory_; }
// Requires in_state != out_state
virtual float Score(const void *in_state, const WordIndex new_word, void *out_state) const = 0;
// Requires in_state != out_state
virtual FullScoreReturn FullScore(const void *in_state, const WordIndex new_word, void *out_state) const = 0;
unsigned char Order() const { return order_; }
const Vocabulary &BaseVocabulary() const { return *base_vocab_; }
private:
template <class T, class U, class V> friend class ModelFacade;
explicit Model(size_t state_size) : state_size_(state_size) {}
const size_t state_size_;
const void *begin_sentence_memory_, *null_context_memory_;
const Vocabulary *base_vocab_;
unsigned char order_;
// Disable copy constructors. They're private and undefined.
// Ersatz boost::noncopyable.
Model(const Model &);
Model &operator=(const Model &);
};
} // mamespace base
} // namespace lm
#endif // LM_VIRTUAL_INTERFACE__

239
lm/vocab.cc
View File

@ -1,239 +0,0 @@
#include "lm/vocab.hh"
#include "lm/binary_format.hh"
#include "lm/enumerate_vocab.hh"
#include "lm/lm_exception.hh"
#include "lm/config.hh"
#include "lm/weights.hh"
#include "util/exception.hh"
#include "util/file.hh"
#include "util/joint_sort.hh"
#include "util/murmur_hash.hh"
#include "util/probing_hash_table.hh"
#include <string>
#include <string.h>
namespace lm {
namespace ngram {
namespace detail {
uint64_t HashForVocab(const char *str, std::size_t len) {
// This proved faster than Boost's hash in speed trials: total load time Murmur 67090000, Boost 72210000
// Chose to use 64A instead of native so binary format will be portable across 64 and 32 bit.
return util::MurmurHash64A(str, len, 0);
}
} // namespace detail
namespace {
// Normally static initialization is a bad idea but MurmurHash is pure arithmetic, so this is ok.
const uint64_t kUnknownHash = detail::HashForVocab("<unk>", 5);
// Sadly some LMs have <UNK>.
const uint64_t kUnknownCapHash = detail::HashForVocab("<UNK>", 5);
void ReadWords(int fd, EnumerateVocab *enumerate, WordIndex expected_count) {
// Check that we're at the right place by reading <unk> which is always first.
char check_unk[6];
util::ReadOrThrow(fd, check_unk, 6);
UTIL_THROW_IF(
memcmp(check_unk, "<unk>", 6),
FormatLoadException,
"Vocabulary words are in the wrong place. This could be because the binary file was built with stale gcc and old kenlm. Stale gcc, including the gcc distributed with RedHat and OS X, has a bug that ignores pragma pack for template-dependent types. New kenlm works around this, so you'll save memory but have to rebuild any binary files using the probing data structure.");
if (!enumerate) return;
enumerate->Add(0, "<unk>");
// Read all the words after unk.
const std::size_t kInitialRead = 16384;
std::string buf;
buf.reserve(kInitialRead + 100);
buf.resize(kInitialRead);
WordIndex index = 1; // Read <unk> already.
while (true) {
std::size_t got = util::ReadOrEOF(fd, &buf[0], kInitialRead);
if (got == 0) break;
buf.resize(got);
while (buf[buf.size() - 1]) {
char next_char;
util::ReadOrThrow(fd, &next_char, 1);
buf.push_back(next_char);
}
// Ok now we have null terminated strings.
for (const char *i = buf.data(); i != buf.data() + buf.size();) {
std::size_t length = strlen(i);
enumerate->Add(index++, StringPiece(i, length));
i += length + 1 /* null byte */;
}
}
UTIL_THROW_IF(expected_count != index, FormatLoadException, "The binary file has the wrong number of words at the end. This could be caused by a truncated binary file.");
}
} // namespace
WriteWordsWrapper::WriteWordsWrapper(EnumerateVocab *inner) : inner_(inner) {}
WriteWordsWrapper::~WriteWordsWrapper() {}
void WriteWordsWrapper::Add(WordIndex index, const StringPiece &str) {
if (inner_) inner_->Add(index, str);
buffer_.append(str.data(), str.size());
buffer_.push_back(0);
}
void WriteWordsWrapper::Write(int fd) {
util::SeekEnd(fd);
util::WriteOrThrow(fd, buffer_.data(), buffer_.size());
}
SortedVocabulary::SortedVocabulary() : begin_(NULL), end_(NULL), enumerate_(NULL) {}
std::size_t SortedVocabulary::Size(std::size_t entries, const Config &/*config*/) {
// Lead with the number of entries.
return sizeof(uint64_t) + sizeof(uint64_t) * entries;
}
void SortedVocabulary::SetupMemory(void *start, std::size_t allocated, std::size_t entries, const Config &config) {
assert(allocated >= Size(entries, config));
// Leave space for number of entries.
begin_ = reinterpret_cast<uint64_t*>(start) + 1;
end_ = begin_;
saw_unk_ = false;
}
void SortedVocabulary::ConfigureEnumerate(EnumerateVocab *to, std::size_t max_entries) {
enumerate_ = to;
if (enumerate_) {
enumerate_->Add(0, "<unk>");
strings_to_enumerate_.resize(max_entries);
}
}
WordIndex SortedVocabulary::Insert(const StringPiece &str) {
uint64_t hashed = detail::HashForVocab(str);
if (hashed == kUnknownHash || hashed == kUnknownCapHash) {
saw_unk_ = true;
return 0;
}
*end_ = hashed;
if (enumerate_) {
strings_to_enumerate_[end_ - begin_].assign(str.data(), str.size());
}
++end_;
// This is 1 + the offset where it was inserted to make room for unk.
return end_ - begin_;
}
void SortedVocabulary::FinishedLoading(ProbBackoff *reorder_vocab) {
if (enumerate_) {
if (!strings_to_enumerate_.empty()) {
util::PairedIterator<ProbBackoff*, std::string*> values(reorder_vocab + 1, &*strings_to_enumerate_.begin());
util::JointSort(begin_, end_, values);
}
for (WordIndex i = 0; i < static_cast<WordIndex>(end_ - begin_); ++i) {
// <unk> strikes again: +1 here.
enumerate_->Add(i + 1, strings_to_enumerate_[i]);
}
strings_to_enumerate_.clear();
} else {
util::JointSort(begin_, end_, reorder_vocab + 1);
}
SetSpecial(Index("<s>"), Index("</s>"), 0);
// Save size. Excludes UNK.
*(reinterpret_cast<uint64_t*>(begin_) - 1) = end_ - begin_;
// Includes UNK.
bound_ = end_ - begin_ + 1;
}
void SortedVocabulary::LoadedBinary(bool have_words, int fd, EnumerateVocab *to) {
end_ = begin_ + *(reinterpret_cast<const uint64_t*>(begin_) - 1);
SetSpecial(Index("<s>"), Index("</s>"), 0);
bound_ = end_ - begin_ + 1;
if (have_words) ReadWords(fd, to, bound_);
}
namespace {
const unsigned int kProbingVocabularyVersion = 0;
} // namespace
namespace detail {
struct ProbingVocabularyHeader {
// Lowest unused vocab id. This is also the number of words, including <unk>.
unsigned int version;
WordIndex bound;
};
} // namespace detail
ProbingVocabulary::ProbingVocabulary() : enumerate_(NULL) {}
std::size_t ProbingVocabulary::Size(std::size_t entries, const Config &config) {
return ALIGN8(sizeof(detail::ProbingVocabularyHeader)) + Lookup::Size(entries, config.probing_multiplier);
}
void ProbingVocabulary::SetupMemory(void *start, std::size_t allocated, std::size_t /*entries*/, const Config &/*config*/) {
header_ = static_cast<detail::ProbingVocabularyHeader*>(start);
lookup_ = Lookup(static_cast<uint8_t*>(start) + ALIGN8(sizeof(detail::ProbingVocabularyHeader)), allocated);
bound_ = 1;
saw_unk_ = false;
}
void ProbingVocabulary::ConfigureEnumerate(EnumerateVocab *to, std::size_t /*max_entries*/) {
enumerate_ = to;
if (enumerate_) {
enumerate_->Add(0, "<unk>");
}
}
WordIndex ProbingVocabulary::Insert(const StringPiece &str) {
uint64_t hashed = detail::HashForVocab(str);
// Prevent unknown from going into the table.
if (hashed == kUnknownHash || hashed == kUnknownCapHash) {
saw_unk_ = true;
return 0;
} else {
if (enumerate_) enumerate_->Add(bound_, str);
lookup_.Insert(ProbingVocabuaryEntry::Make(hashed, bound_));
return bound_++;
}
}
void ProbingVocabulary::InternalFinishedLoading() {
lookup_.FinishedInserting();
header_->bound = bound_;
header_->version = kProbingVocabularyVersion;
SetSpecial(Index("<s>"), Index("</s>"), 0);
}
void ProbingVocabulary::LoadedBinary(bool have_words, int fd, EnumerateVocab *to) {
UTIL_THROW_IF(header_->version != kProbingVocabularyVersion, FormatLoadException, "The binary file has probing version " << header_->version << " but the code expects version " << kProbingVocabularyVersion << ". Please rerun build_binary using the same version of the code.");
lookup_.LoadedBinary();
bound_ = header_->bound;
SetSpecial(Index("<s>"), Index("</s>"), 0);
if (have_words) ReadWords(fd, to, bound_);
}
void MissingUnknown(const Config &config) throw(SpecialWordMissingException) {
switch(config.unknown_missing) {
case SILENT:
return;
case COMPLAIN:
if (config.messages) *config.messages << "The ARPA file is missing <unk>. Substituting log10 probability " << config.unknown_missing_logprob << "." << std::endl;
break;
case THROW_UP:
UTIL_THROW(SpecialWordMissingException, "The ARPA file is missing <unk> and the model is configured to throw an exception.");
}
}
void MissingSentenceMarker(const Config &config, const char *str) throw(SpecialWordMissingException) {
switch (config.sentence_marker_missing) {
case SILENT:
return;
case COMPLAIN:
if (config.messages) *config.messages << "Missing special word " << str << "; will treat it as <unk>.";
break;
case THROW_UP:
UTIL_THROW(SpecialWordMissingException, "The ARPA file is missing " << str << " and the model is configured to reject these models. Run build_binary -s to disable this check.");
}
}
} // namespace ngram
} // namespace lm

182
lm/vocab.hh
View File

@ -1,182 +0,0 @@
#ifndef LM_VOCAB__
#define LM_VOCAB__
#include "lm/enumerate_vocab.hh"
#include "lm/lm_exception.hh"
#include "lm/virtual_interface.hh"
#include "util/probing_hash_table.hh"
#include "util/sorted_uniform.hh"
#include "util/string_piece.hh"
#include <limits>
#include <string>
#include <vector>
namespace lm {
struct ProbBackoff;
class EnumerateVocab;
namespace ngram {
struct Config;
namespace detail {
uint64_t HashForVocab(const char *str, std::size_t len);
inline uint64_t HashForVocab(const StringPiece &str) {
return HashForVocab(str.data(), str.length());
}
class ProbingVocabularyHeader;
} // namespace detail
class WriteWordsWrapper : public EnumerateVocab {
public:
WriteWordsWrapper(EnumerateVocab *inner);
~WriteWordsWrapper();
void Add(WordIndex index, const StringPiece &str);
void Write(int fd);
private:
EnumerateVocab *inner_;
std::string buffer_;
};
// Vocabulary based on sorted uniform find storing only uint64_t values and using their offsets as indices.
class SortedVocabulary : public base::Vocabulary {
public:
SortedVocabulary();
WordIndex Index(const StringPiece &str) const {
const uint64_t *found;
if (util::BoundedSortedUniformFind<const uint64_t*, util::IdentityAccessor<uint64_t>, util::Pivot64>(
util::IdentityAccessor<uint64_t>(),
begin_ - 1, 0,
end_, std::numeric_limits<uint64_t>::max(),
detail::HashForVocab(str), found)) {
return found - begin_ + 1; // +1 because <unk> is 0 and does not appear in the lookup table.
} else {
return 0;
}
}
// Size for purposes of file writing
static size_t Size(std::size_t entries, const Config &config);
// Vocab words are [0, Bound()) Only valid after FinishedLoading/LoadedBinary.
WordIndex Bound() const { return bound_; }
// Everything else is for populating. I'm too lazy to hide and friend these, but you'll only get a const reference anyway.
void SetupMemory(void *start, std::size_t allocated, std::size_t entries, const Config &config);
void ConfigureEnumerate(EnumerateVocab *to, std::size_t max_entries);
WordIndex Insert(const StringPiece &str);
// Reorders reorder_vocab so that the IDs are sorted.
void FinishedLoading(ProbBackoff *reorder_vocab);
// Trie stores the correct counts including <unk> in the header. If this was previously sized based on a count exluding <unk>, padding with 8 bytes will make it the correct size based on a count including <unk>.
std::size_t UnkCountChangePadding() const { return SawUnk() ? 0 : sizeof(uint64_t); }
bool SawUnk() const { return saw_unk_; }
void LoadedBinary(bool have_words, int fd, EnumerateVocab *to);
private:
uint64_t *begin_, *end_;
WordIndex bound_;
WordIndex highest_value_;
bool saw_unk_;
EnumerateVocab *enumerate_;
// Actual strings. Used only when loading from ARPA and enumerate_ != NULL
std::vector<std::string> strings_to_enumerate_;
};
#pragma pack(push)
#pragma pack(4)
struct ProbingVocabuaryEntry {
uint64_t key;
WordIndex value;
typedef uint64_t Key;
uint64_t GetKey() const {
return key;
}
static ProbingVocabuaryEntry Make(uint64_t key, WordIndex value) {
ProbingVocabuaryEntry ret;
ret.key = key;
ret.value = value;
return ret;
}
};
#pragma pack(pop)
// Vocabulary storing a map from uint64_t to WordIndex.
class ProbingVocabulary : public base::Vocabulary {
public:
ProbingVocabulary();
WordIndex Index(const StringPiece &str) const {
Lookup::ConstIterator i;
return lookup_.Find(detail::HashForVocab(str), i) ? i->value : 0;
}
static size_t Size(std::size_t entries, const Config &config);
// Vocab words are [0, Bound()).
WordIndex Bound() const { return bound_; }
// Everything else is for populating. I'm too lazy to hide and friend these, but you'll only get a const reference anyway.
void SetupMemory(void *start, std::size_t allocated, std::size_t entries, const Config &config);
void ConfigureEnumerate(EnumerateVocab *to, std::size_t max_entries);
WordIndex Insert(const StringPiece &str);
template <class Weights> void FinishedLoading(Weights * /*reorder_vocab*/) {
InternalFinishedLoading();
}
std::size_t UnkCountChangePadding() const { return 0; }
bool SawUnk() const { return saw_unk_; }
void LoadedBinary(bool have_words, int fd, EnumerateVocab *to);
private:
void InternalFinishedLoading();
typedef util::ProbingHashTable<ProbingVocabuaryEntry, util::IdentityHash> Lookup;
Lookup lookup_;
WordIndex bound_;
bool saw_unk_;
EnumerateVocab *enumerate_;
detail::ProbingVocabularyHeader *header_;
};
void MissingUnknown(const Config &config) throw(SpecialWordMissingException);
void MissingSentenceMarker(const Config &config, const char *str) throw(SpecialWordMissingException);
template <class Vocab> void CheckSpecials(const Config &config, const Vocab &vocab) throw(SpecialWordMissingException) {
if (!vocab.SawUnk()) MissingUnknown(config);
if (vocab.BeginSentence() == vocab.NotFound()) MissingSentenceMarker(config, "<s>");
if (vocab.EndSentence() == vocab.NotFound()) MissingSentenceMarker(config, "</s>");
}
} // namespace ngram
} // namespace lm
#endif // LM_VOCAB__

22
lm/weights.hh
View File

@ -1,22 +0,0 @@
#ifndef LM_WEIGHTS__
#define LM_WEIGHTS__
// Weights for n-grams. Probability and possibly a backoff.
namespace lm {
struct Prob {
float prob;
};
// No inheritance so this will be a POD.
struct ProbBackoff {
float prob;
float backoff;
};
struct RestWeights {
float prob;
float backoff;
float rest;
};
} // namespace lm
#endif // LM_WEIGHTS__

11
lm/word_index.hh
View File

@ -1,11 +0,0 @@
// Separate header because this is used often.
#ifndef LM_WORD_INDEX__
#define LM_WORD_INDEX__
namespace lm {
typedef unsigned int WordIndex;
} // namespace lm
typedef lm::WordIndex LMWordIndex;
#endif

2
mert/Jamfile
View File

@ -49,7 +49,7 @@ SentenceLevelScorer.cpp
Permutation.cpp
PermutationScorer.cpp
StatisticsBasedScorer.cpp
../util//kenutil m ..//z ;
../lazy/util//kenutil m ..//z ;
exe mert : mert.cpp mert_lib ../moses/src//ThreadPool ;

2
moses/src/DynSAInclude/Jamfile
View File

@ -1 +1 @@
lib dynsa : [ glob *.cpp ] ../../../util//kenutil : : : <include>. ;
lib dynsa : [ glob *.cpp ] ../../../lazy/util//kenutil : : : <include>. ;

2
moses/src/Jamfile
View File

@ -1,4 +1,4 @@
alias headers : ../../util//kenutil : : : <include>. ;
alias headers : ../../lazy/util//kenutil : : : <include>. ;
alias ThreadPool : ThreadPool.cpp ;

2
moses/src/LM/Jamfile
View File

@ -74,7 +74,7 @@ obj Factory.o : Factory.cpp ..//headers $(dependencies) : <include>../DynSAInclu
#Top-level LM library. If you've added a file that doesn't depend on external
#libraries, put it here.
lib LM : Base.cpp Factory.o Implementation.cpp Joint.cpp Ken.cpp MultiFactor.cpp Remote.cpp SingleFactor.cpp ORLM.o
../../../lm//kenlm ..//headers $(dependencies) ;
../../../lazy/lm//kenlm ..//headers $(dependencies) ;
#Everything below is a kludge to force rebuilding if different --with options
#are passed. Could have used features like <srilm>on but getting these to

674
util/COPYING
View File

@ -1,674 +0,0 @@
GNU GENERAL PUBLIC LICENSE
Version 3, 29 June 2007
Copyright (C) 2007 Free Software Foundation, Inc. <http://fsf.org/>
Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed.
Preamble
The GNU General Public License is a free, copyleft license for
software and other kinds of works.
The licenses for most software and other practical works are designed
to take away your freedom to share and change the works. By contrast,
the GNU General Public License is intended to guarantee your freedom to
share and change all versions of a program--to make sure it remains free
software for all its users. We, the Free Software Foundation, use the
GNU General Public License for most of our software; it applies also to
any other work released this way by its authors. You can apply it to
your programs, too.
When we speak of free software, we are referring to freedom, not
price. Our General Public Licenses are designed to make sure that you
have the freedom to distribute copies of free software (and charge for
them if you wish), that you receive source code or can get it if you
want it, that you can change the software or use pieces of it in new
free programs, and that you know you can do these things.
To protect your rights, we need to prevent others from denying you
these rights or asking you to surrender the rights. Therefore, you have
certain responsibilities if you distribute copies of the software, or if
you modify it: responsibilities to respect the freedom of others.
For example, if you distribute copies of such a program, whether
gratis or for a fee, you must pass on to the recipients the same
freedoms that you received. You must make sure that they, too, receive
or can get the source code. And you must show them these terms so they
know their rights.
Developers that use the GNU GPL protect your rights with two steps:
(1) assert copyright on the software, and (2) offer you this License
giving you legal permission to copy, distribute and/or modify it.
For the developers' and authors' protection, the GPL clearly explains
that there is no warranty for this free software. For both users' and
authors' sake, the GPL requires that modified versions be marked as
changed, so that their problems will not be attributed erroneously to
authors of previous versions.
Some devices are designed to deny users access to install or run
modified versions of the software inside them, although the manufacturer
can do so. This is fundamentally incompatible with the aim of
protecting users' freedom to change the software. The systematic
pattern of such abuse occurs in the area of products for individuals to
use, which is precisely where it is most unacceptable. Therefore, we
have designed this version of the GPL to prohibit the practice for those
products. If such problems arise substantially in other domains, we
stand ready to extend this provision to those domains in future versions
of the GPL, as needed to protect the freedom of users.
Finally, every program is threatened constantly by software patents.
States should not allow patents to restrict development and use of
software on general-purpose computers, but in those that do, we wish to
avoid the special danger that patents applied to a free program could
make it effectively proprietary. To prevent this, the GPL assures that
patents cannot be used to render the program non-free.
The precise terms and conditions for copying, distribution and
modification follow.
TERMS AND CONDITIONS
0. Definitions.
"This License" refers to version 3 of the GNU General Public License.
"Copyright" also means copyright-like laws that apply to other kinds of
works, such as semiconductor masks.
"The Program" refers to any copyrightable work licensed under this
License. Each licensee is addressed as "you". "Licensees" and
"recipients" may be individuals or organizations.
To "modify" a work means to copy from or adapt all or part of the work
in a fashion requiring copyright permission, other than the making of an
exact copy. The resulting work is called a "modified version" of the
earlier work or a work "based on" the earlier work.
A "covered work" means either the unmodified Program or a work based
on the Program.
To "propagate" a work means to do anything with it that, without
permission, would make you directly or secondarily liable for
infringement under applicable copyright law, except executing it on a
computer or modifying a private copy. Propagation includes copying,
distribution (with or without modification), making available to the
public, and in some countries other activities as well.
To "convey" a work means any kind of propagation that enables other
parties to make or receive copies. Mere interaction with a user through
a computer network, with no transfer of a copy, is not conveying.
An interactive user interface displays "Appropriate Legal Notices"
to the extent that it includes a convenient and prominently visible
feature that (1) displays an appropriate copyright notice, and (2)
tells the user that there is no warranty for the work (except to the
extent that warranties are provided), that licensees may convey the
work under this License, and how to view a copy of this License. If
the interface presents a list of user commands or options, such as a
menu, a prominent item in the list meets this criterion.
1. Source Code.
The "source code" for a work means the preferred form of the work
for making modifications to it. "Object code" means any non-source
form of a work.
A "Standard Interface" means an interface that either is an official
standard defined by a recognized standards body, or, in the case of
interfaces specified for a particular programming language, one that
is widely used among developers working in that language.
The "System Libraries" of an executable work include anything, other
than the work as a whole, that (a) is included in the normal form of
packaging a Major Component, but which is not part of that Major
Component, and (b) serves only to enable use of the work with that
Major Component, or to implement a Standard Interface for which an
implementation is available to the public in source code form. A
"Major Component", in this context, means a major essential component
(kernel, window system, and so on) of the specific operating system
(if any) on which the executable work runs, or a compiler used to
produce the work, or an object code interpreter used to run it.
The "Corresponding Source" for a work in object code form means all
the source code needed to generate, install, and (for an executable
work) run the object code and to modify the work, including scripts to
control those activities. However, it does not include the work's
System Libraries, or general-purpose tools or generally available free
programs which are used unmodified in performing those activities but
which are not part of the work. For example, Corresponding Source
includes interface definition files associated with source files for
the work, and the source code for shared libraries and dynamically
linked subprograms that the work is specifically designed to require,
such as by intimate data communication or control flow between those
subprograms and other parts of the work.
The Corresponding Source need not include anything that users
can regenerate automatically from other parts of the Corresponding
Source.
The Corresponding Source for a work in source code form is that
same work.
2. Basic Permissions.
All rights granted under this License are granted for the term of
copyright on the Program, and are irrevocable provided the stated
conditions are met. This License explicitly affirms your unlimited
permission to run the unmodified Program. The output from running a
covered work is covered by this License only if the output, given its
content, constitutes a covered work. This License acknowledges your
rights of fair use or other equivalent, as provided by copyright law.
You may make, run and propagate covered works that you do not
convey, without conditions so long as your license otherwise remains
in force. You may convey covered works to others for the sole purpose
of having them make modifications exclusively for you, or provide you
with facilities for running those works, provided that you comply with
the terms of this License in conveying all material for which you do
not control copyright. Those thus making or running the covered works
for you must do so exclusively on your behalf, under your direction
and control, on terms that prohibit them from making any copies of
your copyrighted material outside their relationship with you.
Conveying under any other circumstances is permitted solely under
the conditions stated below. Sublicensing is not allowed; section 10
makes it unnecessary.
3. Protecting Users' Legal Rights From Anti-Circumvention Law.
No covered work shall be deemed part of an effective technological
measure under any applicable law fulfilling obligations under article
11 of the WIPO copyright treaty adopted on 20 December 1996, or
similar laws prohibiting or restricting circumvention of such
measures.
When you convey a covered work, you waive any legal power to forbid
circumvention of technological measures to the extent such circumvention
is effected by exercising rights under this License with respect to
the covered work, and you disclaim any intention to limit operation or
modification of the work as a means of enforcing, against the work's
users, your or third parties' legal rights to forbid circumvention of
technological measures.
4. Conveying Verbatim Copies.
You may convey verbatim copies of the Program's source code as you
receive it, in any medium, provided that you conspicuously and
appropriately publish on each copy an appropriate copyright notice;
keep intact all notices stating that this License and any
non-permissive terms added in accord with section 7 apply to the code;
keep intact all notices of the absence of any warranty; and give all
recipients a copy of this License along with the Program.
You may charge any price or no price for each copy that you convey,
and you may offer support or warranty protection for a fee.
5. Conveying Modified Source Versions.
You may convey a work based on the Program, or the modifications to
produce it from the Program, in the form of source code under the
terms of section 4, provided that you also meet all of these conditions:
a) The work must carry prominent notices stating that you modified
it, and giving a relevant date.
b) The work must carry prominent notices stating that it is
released under this License and any conditions added under section
7. This requirement modifies the requirement in section 4 to
"keep intact all notices".
c) You must license the entire work, as a whole, under this
License to anyone who comes into possession of a copy. This
License will therefore apply, along with any applicable section 7
additional terms, to the whole of the work, and all its parts,
regardless of how they are packaged. This License gives no
permission to license the work in any other way, but it does not
invalidate such permission if you have separately received it.
d) If the work has interactive user interfaces, each must display
Appropriate Legal Notices; however, if the Program has interactive
interfaces that do not display Appropriate Legal Notices, your
work need not make them do so.
A compilation of a covered work with other separate and independent
works, which are not by their nature extensions of the covered work,
and which are not combined with it such as to form a larger program,
in or on a volume of a storage or distribution medium, is called an
"aggregate" if the compilation and its resulting copyright are not
used to limit the access or legal rights of the compilation's users
beyond what the individual works permit. Inclusion of a covered work
in an aggregate does not cause this License to apply to the other
parts of the aggregate.
6. Conveying Non-Source Forms.
You may convey a covered work in object code form under the terms
of sections 4 and 5, provided that you also convey the
machine-readable Corresponding Source under the terms of this License,
in one of these ways:
a) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by the
Corresponding Source fixed on a durable physical medium
customarily used for software interchange.
b) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by a
written offer, valid for at least three years and valid for as
long as you offer spare parts or customer support for that product
model, to give anyone who possesses the object code either (1) a
copy of the Corresponding Source for all the software in the
product that is covered by this License, on a durable physical
medium customarily used for software interchange, for a price no
more than your reasonable cost of physically performing this
conveying of source, or (2) access to copy the
Corresponding Source from a network server at no charge.
c) Convey individual copies of the object code with a copy of the
written offer to provide the Corresponding Source. This
alternative is allowed only occasionally and noncommercially, and
only if you received the object code with such an offer, in accord
with subsection 6b.
d) Convey the object code by offering access from a designated
place (gratis or for a charge), and offer equivalent access to the
Corresponding Source in the same way through the same place at no
further charge. You need not require recipients to copy the
Corresponding Source along with the object code. If the place to
copy the object code is a network server, the Corresponding Source
may be on a different server (operated by you or a third party)
that supports equivalent copying facilities, provided you maintain
clear directions next to the object code saying where to find the
Corresponding Source. Regardless of what server hosts the
Corresponding Source, you remain obligated to ensure that it is
available for as long as needed to satisfy these requirements.
e) Convey the object code using peer-to-peer transmission, provided
you inform other peers where the object code and Corresponding
Source of the work are being offered to the general public at no
charge under subsection 6d.
A separable portion of the object code, whose source code is excluded
from the Corresponding Source as a System Library, need not be
included in conveying the object code work.
A "User Product" is either (1) a "consumer product", which means any
tangible personal property which is normally used for personal, family,
or household purposes, or (2) anything designed or sold for incorporation
into a dwelling. In determining whether a product is a consumer product,
doubtful cases shall be resolved in favor of coverage. For a particular
product received by a particular user, "normally used" refers to a
typical or common use of that class of product, regardless of the status
of the particular user or of the way in which the particular user
actually uses, or expects or is expected to use, the product. A product
is a consumer product regardless of whether the product has substantial
commercial, industrial or non-consumer uses, unless such uses represent
the only significant mode of use of the product.
"Installation Information" for a User Product means any methods,
procedures, authorization keys, or other information required to install
and execute modified versions of a covered work in that User Product from
a modified version of its Corresponding Source. The information must
suffice to ensure that the continued functioning of the modified object
code is in no case prevented or interfered with solely because
modification has been made.
If you convey an object code work under this section in, or with, or
specifically for use in, a User Product, and the conveying occurs as
part of a transaction in which the right of possession and use of the
User Product is transferred to the recipient in perpetuity or for a
fixed term (regardless of how the transaction is characterized), the
Corresponding Source conveyed under this section must be accompanied
by the Installation Information. But this requirement does not apply
if neither you nor any third party retains the ability to install
modified object code on the User Product (for example, the work has
been installed in ROM).
The requirement to provide Installation Information does not include a
requirement to continue to provide support service, warranty, or updates
for a work that has been modified or installed by the recipient, or for
the User Product in which it has been modified or installed. Access to a
network may be denied when the modification itself materially and
adversely affects the operation of the network or violates the rules and
protocols for communication across the network.
Corresponding Source conveyed, and Installation Information provided,
in accord with this section must be in a format that is publicly
documented (and with an implementation available to the public in
source code form), and must require no special password or key for
unpacking, reading or copying.
7. Additional Terms.
"Additional permissions" are terms that supplement the terms of this
License by making exceptions from one or more of its conditions.
Additional permissions that are applicable to the entire Program shall
be treated as though they were included in this License, to the extent
that they are valid under applicable law. If additional permissions
apply only to part of the Program, that part may be used separately
under those permissions, but the entire Program remains governed by
this License without regard to the additional permissions.
When you convey a copy of a covered work, you may at your option
remove any additional permissions from that copy, or from any part of
it. (Additional permissions may be written to require their own
removal in certain cases when you modify the work.) You may place
additional permissions on material, added by you to a covered work,
for which you have or can give appropriate copyright permission.
Notwithstanding any other provision of this License, for material you
add to a covered work, you may (if authorized by the copyright holders of
that material) supplement the terms of this License with terms:
a) Disclaiming warranty or limiting liability differently from the
terms of sections 15 and 16 of this License; or
b) Requiring preservation of specified reasonable legal notices or
author attributions in that material or in the Appropriate Legal
Notices displayed by works containing it; or
c) Prohibiting misrepresentation of the origin of that material, or
requiring that modified versions of such material be marked in
reasonable ways as different from the original version; or
d) Limiting the use for publicity purposes of names of licensors or
authors of the material; or
e) Declining to grant rights under trademark law for use of some
trade names, trademarks, or service marks; or
f) Requiring indemnification of licensors and authors of that
material by anyone who conveys the material (or modified versions of
it) with contractual assumptions of liability to the recipient, for
any liability that these contractual assumptions directly impose on
those licensors and authors.
All other non-permissive additional terms are considered "further
restrictions" within the meaning of section 10. If the Program as you
received it, or any part of it, contains a notice stating that it is
governed by this License along with a term that is a further
restriction, you may remove that term. If a license document contains
a further restriction but permits relicensing or conveying under this
License, you may add to a covered work material governed by the terms
of that license document, provided that the further restriction does
not survive such relicensing or conveying.
If you add terms to a covered work in accord with this section, you
must place, in the relevant source files, a statement of the
additional terms that apply to those files, or a notice indicating
where to find the applicable terms.
Additional terms, permissive or non-permissive, may be stated in the
form of a separately written license, or stated as exceptions;
the above requirements apply either way.
8. Termination.
You may not propagate or modify a covered work except as expressly
provided under this License. Any attempt otherwise to propagate or
modify it is void, and will automatically terminate your rights under
this License (including any patent licenses granted under the third
paragraph of section 11).
However, if you cease all violation of this License, then your
license from a particular copyright holder is reinstated (a)
provisionally, unless and until the copyright holder explicitly and
finally terminates your license, and (b) permanently, if the copyright
holder fails to notify you of the violation by some reasonable means
prior to 60 days after the cessation.
Moreover, your license from a particular copyright holder is
reinstated permanently if the copyright holder notifies you of the
violation by some reasonable means, this is the first time you have
received notice of violation of this License (for any work) from that
copyright holder, and you cure the violation prior to 30 days after
your receipt of the notice.
Termination of your rights under this section does not terminate the
licenses of parties who have received copies or rights from you under
this License. If your rights have been terminated and not permanently
reinstated, you do not qualify to receive new licenses for the same
material under section 10.
9. Acceptance Not Required for Having Copies.
You are not required to accept this License in order to receive or
run a copy of the Program. Ancillary propagation of a covered work
occurring solely as a consequence of using peer-to-peer transmission
to receive a copy likewise does not require acceptance. However,
nothing other than this License grants you permission to propagate or
modify any covered work. These actions infringe copyright if you do
not accept this License. Therefore, by modifying or propagating a
covered work, you indicate your acceptance of this License to do so.
10. Automatic Licensing of Downstream Recipients.
Each time you convey a covered work, the recipient automatically
receives a license from the original licensors, to run, modify and
propagate that work, subject to this License. You are not responsible
for enforcing compliance by third parties with this License.
An "entity transaction" is a transaction transferring control of an
organization, or substantially all assets of one, or subdividing an
organization, or merging organizations. If propagation of a covered
work results from an entity transaction, each party to that
transaction who receives a copy of the work also receives whatever
licenses to the work the party's predecessor in interest had or could
give under the previous paragraph, plus a right to possession of the
Corresponding Source of the work from the predecessor in interest, if
the predecessor has it or can get it with reasonable efforts.
You may not impose any further restrictions on the exercise of the
rights granted or affirmed under this License. For example, you may
not impose a license fee, royalty, or other charge for exercise of
rights granted under this License, and you may not initiate litigation
(including a cross-claim or counterclaim in a lawsuit) alleging that
any patent claim is infringed by making, using, selling, offering for
sale, or importing the Program or any portion of it.
11. Patents.
A "contributor" is a copyright holder who authorizes use under this
License of the Program or a work on which the Program is based. The
work thus licensed is called the contributor's "contributor version".
A contributor's "essential patent claims" are all patent claims
owned or controlled by the contributor, whether already acquired or
hereafter acquired, that would be infringed by some manner, permitted
by this License, of making, using, or selling its contributor version,
but do not include claims that would be infringed only as a
consequence of further modification of the contributor version. For
purposes of this definition, "control" includes the right to grant
patent sublicenses in a manner consistent with the requirements of
this License.
Each contributor grants you a non-exclusive, worldwide, royalty-free
patent license under the contributor's essential patent claims, to
make, use, sell, offer for sale, import and otherwise run, modify and
propagate the contents of its contributor version.
In the following three paragraphs, a "patent license" is any express
agreement or commitment, however denominated, not to enforce a patent
(such as an express permission to practice a patent or covenant not to
sue for patent infringement). To "grant" such a patent license to a
party means to make such an agreement or commitment not to enforce a
patent against the party.
If you convey a covered work, knowingly relying on a patent license,
and the Corresponding Source of the work is not available for anyone
to copy, free of charge and under the terms of this License, through a
publicly available network server or other readily accessible means,
then you must either (1) cause the Corresponding Source to be so
available, or (2) arrange to deprive yourself of the benefit of the
patent license for this particular work, or (3) arrange, in a manner
consistent with the requirements of this License, to extend the patent
license to downstream recipients. "Knowingly relying" means you have
actual knowledge that, but for the patent license, your conveying the
covered work in a country, or your recipient's use of the covered work
in a country, would infringe one or more identifiable patents in that
country that you have reason to believe are valid.
If, pursuant to or in connection with a single transaction or
arrangement, you convey, or propagate by procuring conveyance of, a
covered work, and grant a patent license to some of the parties
receiving the covered work authorizing them to use, propagate, modify
or convey a specific copy of the covered work, then the patent license
you grant is automatically extended to all recipients of the covered
work and works based on it.
A patent license is "discriminatory" if it does not include within
the scope of its coverage, prohibits the exercise of, or is
conditioned on the non-exercise of one or more of the rights that are
specifically granted under this License. You may not convey a covered
work if you are a party to an arrangement with a third party that is
in the business of distributing software, under which you make payment
to the third party based on the extent of your activity of conveying
the work, and under which the third party grants, to any of the
parties who would receive the covered work from you, a discriminatory
patent license (a) in connection with copies of the covered work
conveyed by you (or copies made from those copies), or (b) primarily
for and in connection with specific products or compilations that
contain the covered work, unless you entered into that arrangement,
or that patent license was granted, prior to 28 March 2007.
Nothing in this License shall be construed as excluding or limiting
any implied license or other defenses to infringement that may
otherwise be available to you under applicable patent law.
12. No Surrender of Others' Freedom.
If conditions are imposed on you (whether by court order, agreement or
otherwise) that contradict the conditions of this License, they do not
excuse you from the conditions of this License. If you cannot convey a
covered work so as to satisfy simultaneously your obligations under this
License and any other pertinent obligations, then as a consequence you may
not convey it at all. For example, if you agree to terms that obligate you
to collect a royalty for further conveying from those to whom you convey
the Program, the only way you could satisfy both those terms and this
License would be to refrain entirely from conveying the Program.
13. Use with the GNU Affero General Public License.
Notwithstanding any other provision of this License, you have
permission to link or combine any covered work with a work licensed
under version 3 of the GNU Affero General Public License into a single
combined work, and to convey the resulting work. The terms of this
License will continue to apply to the part which is the covered work,
but the special requirements of the GNU Affero General Public License,
section 13, concerning interaction through a network will apply to the
combination as such.
14. Revised Versions of this License.
The Free Software Foundation may publish revised and/or new versions of
the GNU General Public License from time to time. Such new versions will
be similar in spirit to the present version, but may differ in detail to
address new problems or concerns.
Each version is given a distinguishing version number. If the
Program specifies that a certain numbered version of the GNU General
Public License "or any later version" applies to it, you have the
option of following the terms and conditions either of that numbered
version or of any later version published by the Free Software
Foundation. If the Program does not specify a version number of the
GNU General Public License, you may choose any version ever published
by the Free Software Foundation.
If the Program specifies that a proxy can decide which future
versions of the GNU General Public License can be used, that proxy's
public statement of acceptance of a version permanently authorizes you
to choose that version for the Program.
Later license versions may give you additional or different
permissions. However, no additional obligations are imposed on any
author or copyright holder as a result of your choosing to follow a
later version.
15. Disclaimer of Warranty.
THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
16. Limitation of Liability.
IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
SUCH DAMAGES.
17. Interpretation of Sections 15 and 16.
If the disclaimer of warranty and limitation of liability provided
above cannot be given local legal effect according to their terms,
reviewing courts shall apply local law that most closely approximates
an absolute waiver of all civil liability in connection with the
Program, unless a warranty or assumption of liability accompanies a
copy of the Program in return for a fee.
END OF TERMS AND CONDITIONS
How to Apply These Terms to Your New Programs
If you develop a new program, and you want it to be of the greatest
possible use to the public, the best way to achieve this is to make it
free software which everyone can redistribute and change under these terms.
To do so, attach the following notices to the program. It is safest
to attach them to the start of each source file to most effectively
state the exclusion of warranty; and each file should have at least
the "copyright" line and a pointer to where the full notice is found.
<one line to give the program's name and a brief idea of what it does.>
Copyright (C) <year> <name of author>
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <http://www.gnu.org/licenses/>.
Also add information on how to contact you by electronic and paper mail.
If the program does terminal interaction, make it output a short
notice like this when it starts in an interactive mode:
<program> Copyright (C) <year> <name of author>
This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
This is free software, and you are welcome to redistribute it
under certain conditions; type `show c' for details.
The hypothetical commands `show w' and `show c' should show the appropriate
parts of the General Public License. Of course, your program's commands
might be different; for a GUI interface, you would use an "about box".
You should also get your employer (if you work as a programmer) or school,
if any, to sign a "copyright disclaimer" for the program, if necessary.
For more information on this, and how to apply and follow the GNU GPL, see
<http://www.gnu.org/licenses/>.
The GNU General Public License does not permit incorporating your program
into proprietary programs. If your program is a subroutine library, you
may consider it more useful to permit linking proprietary applications with
the library. If this is what you want to do, use the GNU Lesser General
Public License instead of this License. But first, please read
<http://www.gnu.org/philosophy/why-not-lgpl.html>.

165
util/COPYING.LESSER
View File

@ -1,165 +0,0 @@
GNU LESSER GENERAL PUBLIC LICENSE
Version 3, 29 June 2007
Copyright (C) 2007 Free Software Foundation, Inc. <http://fsf.org/>
Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed.
This version of the GNU Lesser General Public License incorporates
the terms and conditions of version 3 of the GNU General Public
License, supplemented by the additional permissions listed below.
0. Additional Definitions.
As used herein, "this License" refers to version 3 of the GNU Lesser
General Public License, and the "GNU GPL" refers to version 3 of the GNU
General Public License.
"The Library" refers to a covered work governed by this License,
other than an Application or a Combined Work as defined below.
An "Application" is any work that makes use of an interface provided
by the Library, but which is not otherwise based on the Library.
Defining a subclass of a class defined by the Library is deemed a mode
of using an interface provided by the Library.
A "Combined Work" is a work produced by combining or linking an
Application with the Library. The particular version of the Library
with which the Combined Work was made is also called the "Linked
Version".
The "Minimal Corresponding Source" for a Combined Work means the
Corresponding Source for the Combined Work, excluding any source code
for portions of the Combined Work that, considered in isolation, are
based on the Application, and not on the Linked Version.
The "Corresponding Application Code" for a Combined Work means the
object code and/or source code for the Application, including any data
and utility programs needed for reproducing the Combined Work from the
Application, but excluding the System Libraries of the Combined Work.
1. Exception to Section 3 of the GNU GPL.
You may convey a covered work under sections 3 and 4 of this License
without being bound by section 3 of the GNU GPL.
2. Conveying Modified Versions.
If you modify a copy of the Library, and, in your modifications, a
facility refers to a function or data to be supplied by an Application
that uses the facility (other than as an argument passed when the
facility is invoked), then you may convey a copy of the modified
version:
a) under this License, provided that you make a good faith effort to
ensure that, in the event an Application does not supply the
function or data, the facility still operates, and performs
whatever part of its purpose remains meaningful, or
b) under the GNU GPL, with none of the additional permissions of
this License applicable to that copy.
3. Object Code Incorporating Material from Library Header Files.
The object code form of an Application may incorporate material from
a header file that is part of the Library. You may convey such object
code under terms of your choice, provided that, if the incorporated
material is not limited to numerical parameters, data structure
layouts and accessors, or small macros, inline functions and templates
(ten or fewer lines in length), you do both of the following:
a) Give prominent notice with each copy of the object code that the
Library is used in it and that the Library and its use are
covered by this License.
b) Accompany the object code with a copy of the GNU GPL and this license
document.
4. Combined Works.
You may convey a Combined Work under terms of your choice that,
taken together, effectively do not restrict modification of the
portions of the Library contained in the Combined Work and reverse
engineering for debugging such modifications, if you also do each of
the following:
a) Give prominent notice with each copy of the Combined Work that
the Library is used in it and that the Library and its use are
covered by this License.
b) Accompany the Combined Work with a copy of the GNU GPL and this license
document.
c) For a Combined Work that displays copyright notices during
execution, include the copyright notice for the Library among
these notices, as well as a reference directing the user to the
copies of the GNU GPL and this license document.
d) Do one of the following:
0) Convey the Minimal Corresponding Source under the terms of this
License, and the Corresponding Application Code in a form
suitable for, and under terms that permit, the user to
recombine or relink the Application with a modified version of
the Linked Version to produce a modified Combined Work, in the
manner specified by section 6 of the GNU GPL for conveying
Corresponding Source.
1) Use a suitable shared library mechanism for linking with the
Library. A suitable mechanism is one that (a) uses at run time
a copy of the Library already present on the user's computer
system, and (b) will operate properly with a modified version
of the Library that is interface-compatible with the Linked
Version.
e) Provide Installation Information, but only if you would otherwise
be required to provide such information under section 6 of the
GNU GPL, and only to the extent that such information is
necessary to install and execute a modified version of the
Combined Work produced by recombining or relinking the
Application with a modified version of the Linked Version. (If
you use option 4d0, the Installation Information must accompany
the Minimal Corresponding Source and Corresponding Application
Code. If you use option 4d1, you must provide the Installation
Information in the manner specified by section 6 of the GNU GPL
for conveying Corresponding Source.)
5. Combined Libraries.
You may place library facilities that are a work based on the
Library side by side in a single library together with other library
facilities that are not Applications and are not covered by this
License, and convey such a combined library under terms of your
choice, if you do both of the following:
a) Accompany the combined library with a copy of the same work based
on the Library, uncombined with any other library facilities,
conveyed under the terms of this License.
b) Give prominent notice with the combined library that part of it
is a work based on the Library, and explaining where to find the
accompanying uncombined form of the same work.
6. Revised Versions of the GNU Lesser General Public License.
The Free Software Foundation may publish revised and/or new versions
of the GNU Lesser General Public License from time to time. Such new
versions will be similar in spirit to the present version, but may
differ in detail to address new problems or concerns.
Each version is given a distinguishing version number. If the
Library as you received it specifies that a certain numbered version
of the GNU Lesser General Public License "or any later version"
applies to it, you have the option of following the terms and
conditions either of that published version or of any later version
published by the Free Software Foundation. If the Library as you
received it does not specify a version number of the GNU Lesser
General Public License, you may choose any version of the GNU Lesser
General Public License ever published by the Free Software Foundation.
If the Library as you received it specifies that a proxy can decide
whether future versions of the GNU Lesser General Public License shall
apply, that proxy's public statement of acceptance of any version is
permanent authorization for you to choose that version for the
Library.

10
util/Jamfile
View File

@ -1,10 +0,0 @@
lib kenutil : bit_packing.cc ersatz_progress.cc exception.cc file.cc file_piece.cc mmap.cc murmur_hash.cc usage.cc ..//z : <include>.. : : <include>.. ;
import testing ;
unit-test bit_packing_test : bit_packing_test.cc kenutil ..//boost_unit_test_framework ;
run file_piece_test.cc kenutil ..//boost_unit_test_framework : : file_piece.cc ;
unit-test joint_sort_test : joint_sort_test.cc kenutil ..//boost_unit_test_framework ;
unit-test probing_hash_table_test : probing_hash_table_test.cc kenutil ..//boost_unit_test_framework ;
unit-test sorted_uniform_test : sorted_uniform_test.cc kenutil ..//boost_unit_test_framework ;
unit-test tokenize_piece_test : tokenize_piece_test.cc kenutil ..//boost_unit_test_framework ;

20
util/LICENSE
View File

@ -1,20 +0,0 @@
Most of the code here is licensed under the LGPL. There are exceptions which have their own licenses, listed below. See comments in those files for more details.
murmur_hash.cc is under the MIT license.
string_piece.hh and util/string_piece.cc are Google code and contains its own license.
file.cc contains a modified implementation of mkstemp under the LGPL.
For the rest:
Avenue code is free software: you can redistribute it and/or modify
it under the terms of the GNU Lesser General Public License as published
by the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
Avenue code is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public License
along with Avenue code. If not, see <http://www.gnu.org/licenses/>.

40
util/bit_packing.cc
View File

@ -1,40 +0,0 @@
#include "util/bit_packing.hh"
#include "util/exception.hh"
#include <string.h>
namespace util {
namespace {
template <bool> struct StaticCheck {};
template <> struct StaticCheck<true> { typedef bool StaticAssertionPassed; };
// If your float isn't 4 bytes, we're hosed.
typedef StaticCheck<sizeof(float) == 4>::StaticAssertionPassed FloatSize;
} // namespace
uint8_t RequiredBits(uint64_t max_value) {
if (!max_value) return 0;
uint8_t ret = 1;
while (max_value >>= 1) ++ret;
return ret;
}
void BitPackingSanity() {
const FloatEnc neg1 = { -1.0 }, pos1 = { 1.0 };
if ((neg1.i ^ pos1.i) != 0x80000000) UTIL_THROW(Exception, "Sign bit is not 0x80000000");
char mem[57+8];
memset(mem, 0, sizeof(mem));
const uint64_t test57 = 0x123456789abcdefULL;
for (uint64_t b = 0; b < 57 * 8; b += 57) {
WriteInt57(mem, b, 57, test57);
}
for (uint64_t b = 0; b < 57 * 8; b += 57) {
if (test57 != ReadInt57(mem, b, 57, (1ULL << 57) - 1))
UTIL_THROW(Exception, "The bit packing routines are failing for your architecture. Please send a bug report with your architecture, operating system, and compiler.");
}
// TODO: more checks.
}
} // namespace util

186
util/bit_packing.hh
View File

@ -1,186 +0,0 @@
#ifndef UTIL_BIT_PACKING__
#define UTIL_BIT_PACKING__
/* Bit-level packing routines
*
* WARNING WARNING WARNING:
* The write functions assume that memory is zero initially. This makes them
* faster and is the appropriate case for mmapped language model construction.
* These routines assume that unaligned access to uint64_t is fast. This is
* the case on x86_64. I'm not sure how fast unaligned 64-bit access is on
* x86 but my target audience is large language models for which 64-bit is
* necessary.
*
* Call the BitPackingSanity function to sanity check. Calling once suffices,
* but it may be called multiple times when that's inconvenient.
*
* ARM and MinGW ports contributed by Hideo Okuma and Tomoyuki Yoshimura at
* NICT.
*/
#include <assert.h>
#ifdef __APPLE__
#include <architecture/byte_order.h>
#elif __linux__
#include <endian.h>
#elif !defined(_WIN32) && !defined(_WIN64)
#include <arpa/nameser_compat.h>
#endif
#include <stdint.h>
#include <string.h>
namespace util {
// Fun fact: __BYTE_ORDER is wrong on Solaris Sparc, but the version without __ is correct.
#if BYTE_ORDER == LITTLE_ENDIAN
inline uint8_t BitPackShift(uint8_t bit, uint8_t /*length*/) {
return bit;
}
#elif BYTE_ORDER == BIG_ENDIAN
inline uint8_t BitPackShift(uint8_t bit, uint8_t length) {
return 64 - length - bit;
}
#else
#error "Bit packing code isn't written for your byte order."
#endif
inline uint64_t ReadOff(const void *base, uint64_t bit_off) {
#if defined(__arm) || defined(__arm__)
const uint8_t *base_off = reinterpret_cast<const uint8_t*>(base) + (bit_off >> 3);
uint64_t value64;
memcpy(&value64, base_off, sizeof(value64));
return value64;
#else
return *reinterpret_cast<const uint64_t*>(reinterpret_cast<const uint8_t*>(base) + (bit_off >> 3));
#endif
}
/* Pack integers up to 57 bits using their least significant digits.
* The length is specified using mask:
* Assumes mask == (1 << length) - 1 where length <= 57.
*/
inline uint64_t ReadInt57(const void *base, uint64_t bit_off, uint8_t length, uint64_t mask) {
return (ReadOff(base, bit_off) >> BitPackShift(bit_off & 7, length)) & mask;
}
/* Assumes value < (1 << length) and length <= 57.
* Assumes the memory is zero initially.
*/
inline void WriteInt57(void *base, uint64_t bit_off, uint8_t length, uint64_t value) {
#if defined(__arm) || defined(__arm__)
uint8_t *base_off = reinterpret_cast<uint8_t*>(base) + (bit_off >> 3);
uint64_t value64;
memcpy(&value64, base_off, sizeof(value64));
value64 |= (value << BitPackShift(bit_off & 7, length));
memcpy(base_off, &value64, sizeof(value64));
#else
*reinterpret_cast<uint64_t*>(reinterpret_cast<uint8_t*>(base) + (bit_off >> 3)) |=
(value << BitPackShift(bit_off & 7, length));
#endif
}
/* Same caveats as above, but for a 25 bit limit. */
inline uint32_t ReadInt25(const void *base, uint64_t bit_off, uint8_t length, uint32_t mask) {
#if defined(__arm) || defined(__arm__)
const uint8_t *base_off = reinterpret_cast<const uint8_t*>(base) + (bit_off >> 3);
uint32_t value32;
memcpy(&value32, base_off, sizeof(value32));
return (value32 >> BitPackShift(bit_off & 7, length)) & mask;
#else
return (*reinterpret_cast<const uint32_t*>(reinterpret_cast<const uint8_t*>(base) + (bit_off >> 3)) >> BitPackShift(bit_off & 7, length)) & mask;
#endif
}
inline void WriteInt25(void *base, uint64_t bit_off, uint8_t length, uint32_t value) {
#if defined(__arm) || defined(__arm__)
uint8_t *base_off = reinterpret_cast<uint8_t*>(base) + (bit_off >> 3);
uint32_t value32;
memcpy(&value32, base_off, sizeof(value32));
value32 |= (value << BitPackShift(bit_off & 7, length));
memcpy(base_off, &value32, sizeof(value32));
#else
*reinterpret_cast<uint32_t*>(reinterpret_cast<uint8_t*>(base) + (bit_off >> 3)) |=
(value << BitPackShift(bit_off & 7, length));
#endif
}
typedef union { float f; uint32_t i; } FloatEnc;
inline float ReadFloat32(const void *base, uint64_t bit_off) {
FloatEnc encoded;
encoded.i = ReadOff(base, bit_off) >> BitPackShift(bit_off & 7, 32);
return encoded.f;
}
inline void WriteFloat32(void *base, uint64_t bit_off, float value) {
FloatEnc encoded;
encoded.f = value;
WriteInt57(base, bit_off, 32, encoded.i);
}
const uint32_t kSignBit = 0x80000000;
inline void SetSign(float &to) {
FloatEnc enc;
enc.f = to;
enc.i |= kSignBit;
to = enc.f;
}
inline void UnsetSign(float &to) {
FloatEnc enc;
enc.f = to;
enc.i &= ~kSignBit;
to = enc.f;
}
inline float ReadNonPositiveFloat31(const void *base, uint64_t bit_off) {
FloatEnc encoded;
encoded.i = ReadOff(base, bit_off) >> BitPackShift(bit_off & 7, 31);
// Sign bit set means negative.
encoded.i |= kSignBit;
return encoded.f;
}
inline void WriteNonPositiveFloat31(void *base, uint64_t bit_off, float value) {
FloatEnc encoded;
encoded.f = value;
encoded.i &= ~kSignBit;
WriteInt57(base, bit_off, 31, encoded.i);
}
void BitPackingSanity();
// Return bits required to store integers upto max_value. Not the most
// efficient implementation, but this is only called a few times to size tries.
uint8_t RequiredBits(uint64_t max_value);
struct BitsMask {
static BitsMask ByMax(uint64_t max_value) {
BitsMask ret;
ret.FromMax(max_value);
return ret;
}
static BitsMask ByBits(uint8_t bits) {
BitsMask ret;
ret.bits = bits;
ret.mask = (1ULL << bits) - 1;
return ret;
}
void FromMax(uint64_t max_value) {
bits = RequiredBits(max_value);
mask = (1ULL << bits) - 1;
}
uint8_t bits;
uint64_t mask;
};
struct BitAddress {
BitAddress(void *in_base, uint64_t in_offset) : base(in_base), offset(in_offset) {}
void *base;
uint64_t offset;
};
} // namespace util
#endif // UTIL_BIT_PACKING__

59
util/bit_packing_test.cc
View File

@ -1,59 +0,0 @@
#include "util/bit_packing.hh"
#define BOOST_TEST_MODULE BitPackingTest
#include <boost/test/unit_test.hpp>
#include <string.h>
namespace util {
namespace {
const uint64_t test57 = 0x123456789abcdefULL;
const uint32_t test25 = 0x1234567;
BOOST_AUTO_TEST_CASE(ZeroBit57) {
char mem[16];
memset(mem, 0, sizeof(mem));
WriteInt57(mem, 0, 57, test57);
BOOST_CHECK_EQUAL(test57, ReadInt57(mem, 0, 57, (1ULL << 57) - 1));
}
BOOST_AUTO_TEST_CASE(EachBit57) {
char mem[16];
for (uint8_t b = 0; b < 8; ++b) {
memset(mem, 0, sizeof(mem));
WriteInt57(mem, b, 57, test57);
BOOST_CHECK_EQUAL(test57, ReadInt57(mem, b, 57, (1ULL << 57) - 1));
}
}
BOOST_AUTO_TEST_CASE(Consecutive57) {
char mem[57+8];
memset(mem, 0, sizeof(mem));
for (uint64_t b = 0; b < 57 * 8; b += 57) {
WriteInt57(mem, b, 57, test57);
BOOST_CHECK_EQUAL(test57, ReadInt57(mem, b, 57, (1ULL << 57) - 1));
}
for (uint64_t b = 0; b < 57 * 8; b += 57) {
BOOST_CHECK_EQUAL(test57, ReadInt57(mem, b, 57, (1ULL << 57) - 1));
}
}
BOOST_AUTO_TEST_CASE(Consecutive25) {
char mem[25+8];
memset(mem, 0, sizeof(mem));
for (uint64_t b = 0; b < 25 * 8; b += 25) {
WriteInt25(mem, b, 25, test25);
BOOST_CHECK_EQUAL(test25, ReadInt25(mem, b, 25, (1ULL << 25) - 1));
}
for (uint64_t b = 0; b < 25 * 8; b += 25) {
BOOST_CHECK_EQUAL(test25, ReadInt25(mem, b, 25, (1ULL << 25) - 1));
}
}
BOOST_AUTO_TEST_CASE(Sanity) {
BitPackingSanity();
}
} // namespace
} // namespace util

21
util/check.hh
View File

@ -1,21 +0,0 @@
/* People have been abusing assert by assuming it will always execute. To
* rememdy the situation, asserts were replaced with CHECK. These should then
* be manually replaced with assert (when used correctly) or UTIL_THROW (for
* runtime checks).
*/
#ifndef UTIL_CHECK__
#define UTIL_CHECK__
#include <stdlib.h>
#include <iostream>
#include <cassert>
#define CHECK(Condition) do { \
if (!(Condition)) { \
std::cerr << "Check " << #Condition << " failed in " << __FILE__ << ":" << __LINE__ << std::endl; \
abort(); \
} \
} while (0) // swallow ;
#endif // UTIL_CHECK__

45
util/ersatz_progress.cc
View File

@ -1,45 +0,0 @@
#include "util/ersatz_progress.hh"
#include <algorithm>
#include <ostream>
#include <limits>
#include <string>
namespace util {
namespace { const unsigned char kWidth = 100; }
ErsatzProgress::ErsatzProgress() : current_(0), next_(std::numeric_limits<std::size_t>::max()), complete_(next_), out_(NULL) {}
ErsatzProgress::~ErsatzProgress() {
if (out_) Finished();
}
ErsatzProgress::ErsatzProgress(std::size_t complete, std::ostream *to, const std::string &message)
: current_(0), next_(complete / kWidth), complete_(complete), stones_written_(0), out_(to) {
if (!out_) {
next_ = std::numeric_limits<std::size_t>::max();
return;
}
if (!message.empty()) *out_ << message << '\n';
*out_ << "----5---10---15---20---25---30---35---40---45---50---55---60---65---70---75---80---85---90---95--100\n";
}
void ErsatzProgress::Milestone() {
if (!out_) { current_ = 0; return; }
if (!complete_) return;
unsigned char stone = std::min(static_cast<std::size_t>(kWidth), (current_ * kWidth) / complete_);
for (; stones_written_ < stone; ++stones_written_) {
(*out_) << '*';
}
if (stone == kWidth) {
(*out_) << std::endl;
next_ = std::numeric_limits<std::size_t>::max();
out_ = NULL;
} else {
next_ = std::max(next_, (stone * complete_) / kWidth);
}
}
} // namespace util

54
util/ersatz_progress.hh
View File

@ -1,54 +0,0 @@
#ifndef UTIL_ERSATZ_PROGRESS__
#define UTIL_ERSATZ_PROGRESS__
#include <iostream>
#include <string>
// Ersatz version of boost::progress so core language model doesn't depend on
// boost. Also adds option to print nothing.
namespace util {
class ErsatzProgress {
public:
// No output.
ErsatzProgress();
// Null means no output. The null value is useful for passing along the ostream pointer from another caller.
explicit ErsatzProgress(std::size_t complete, std::ostream *to = &std::cerr, const std::string &message = "");
~ErsatzProgress();
ErsatzProgress &operator++() {
if (++current_ >= next_) Milestone();
return *this;
}
ErsatzProgress &operator+=(std::size_t amount) {
if ((current_ += amount) >= next_) Milestone();
return *this;
}
void Set(std::size_t to) {
if ((current_ = to) >= next_) Milestone();
Milestone();
}
void Finished() {
Set(complete_);
}
private:
void Milestone();
std::size_t current_, next_, complete_;
unsigned char stones_written_;
std::ostream *out_;
// noncopyable
ErsatzProgress(const ErsatzProgress &other);
ErsatzProgress &operator=(const ErsatzProgress &other);
};
} // namespace util
#endif // UTIL_ERSATZ_PROGRESS__

87
util/exception.cc
View File

@ -1,87 +0,0 @@
#include "util/exception.hh"
#ifdef __GXX_RTTI
#include <typeinfo>
#endif
#include <errno.h>
#include <string.h>
namespace util {
Exception::Exception() throw() {}
Exception::~Exception() throw() {}
Exception::Exception(const Exception &from) : std::exception() {
stream_ << from.stream_.str();
}
Exception &Exception::operator=(const Exception &from) {
stream_ << from.stream_.str();
return *this;
}
const char *Exception::what() const throw() {
text_ = stream_.str();
return text_.c_str();
}
void Exception::SetLocation(const char *file, unsigned int line, const char *func, const char *child_name, const char *condition) {
/* The child class might have set some text, but we want this to come first.
* Another option would be passing this information to the constructor, but
* then child classes would have to accept constructor arguments and pass
* them down.
*/
text_ = stream_.str();
stream_.str("");
stream_ << file << ':' << line;
if (func) stream_ << " in " << func << " threw ";
if (child_name) {
stream_ << child_name;
} else {
#ifdef __GXX_RTTI
stream_ << typeid(this).name();
#else
stream_ << "an exception";
#endif
}
if (condition) stream_ << " because `" << condition;
stream_ << "'.\n";
stream_ << text_;
}
namespace {
// The XOPEN version.
const char *HandleStrerror(int ret, const char *buf) {
if (!ret) return buf;
return NULL;
}
// The GNU version.
const char *HandleStrerror(const char *ret, const char * /*buf*/) {
return ret;
}
} // namespace
ErrnoException::ErrnoException() throw() : errno_(errno) {
char buf[200];
buf[0] = 0;
#if defined(sun) || defined(_WIN32) || defined(_WIN64)
const char *add = strerror(errno);
#else
const char *add = HandleStrerror(strerror_r(errno, buf, 200), buf);
#endif
if (add) {
*this << add << ' ';
}
}
ErrnoException::~ErrnoException() throw() {}
EndOfFileException::EndOfFileException() throw() {
*this << "End of file";
}
EndOfFileException::~EndOfFileException() throw() {}
} // namespace util

116
util/exception.hh
View File

@ -1,116 +0,0 @@
#ifndef UTIL_EXCEPTION__
#define UTIL_EXCEPTION__
#include <exception>
#include <sstream>
#include <string>
namespace util {
template <class Except, class Data> typename Except::template ExceptionTag<Except&>::Identity operator<<(Except &e, const Data &data);
class Exception : public std::exception {
public:
Exception() throw();
virtual ~Exception() throw();
Exception(const Exception &from);
Exception &operator=(const Exception &from);
// Not threadsafe, but probably doesn't matter. FWIW, Boost's exception guidance implies that what() isn't threadsafe.
const char *what() const throw();
// For use by the UTIL_THROW macros.
void SetLocation(
const char *file,
unsigned int line,
const char *func,
const char *child_name,
const char *condition);
private:
template <class Except, class Data> friend typename Except::template ExceptionTag<Except&>::Identity operator<<(Except &e, const Data &data);
// This helps restrict operator<< defined below.
template <class T> struct ExceptionTag {
typedef T Identity;
};
std::stringstream stream_;
mutable std::string text_;
};
/* This implements the normal operator<< for Exception and all its children.
* SNIFAE means it only applies to Exception. Think of this as an ersatz
* boost::enable_if.
*/
template <class Except, class Data> typename Except::template ExceptionTag<Except&>::Identity operator<<(Except &e, const Data &data) {
e.stream_ << data;
return e;
}
#ifdef __GNUC__
#define UTIL_FUNC_NAME __PRETTY_FUNCTION__
#else
#ifdef _WIN32
#define UTIL_FUNC_NAME __FUNCTION__
#else
#define UTIL_FUNC_NAME NULL
#endif
#endif
#define UTIL_SET_LOCATION(UTIL_e, child, condition) do { \
(UTIL_e).SetLocation(__FILE__, __LINE__, UTIL_FUNC_NAME, (child), (condition)); \
} while (0)
/* Create an instance of Exception, add the message Modify, and throw it.
* Modify is appended to the what() message and can contain << for ostream
* operations.
*
* do .. while kludge to swallow trailing ; character
* http://gcc.gnu.org/onlinedocs/cpp/Swallowing-the-Semicolon.html .
*/
#define UTIL_THROW(Exception, Modify) do { \
Exception UTIL_e; \
UTIL_SET_LOCATION(UTIL_e, #Exception, NULL); \
UTIL_e << Modify; \
throw UTIL_e; \
} while (0)
#define UTIL_THROW_VAR(Var, Modify) do { \
Exception &UTIL_e = (Var); \
UTIL_SET_LOCATION(UTIL_e, NULL, NULL); \
UTIL_e << Modify; \
throw UTIL_e; \
} while (0)
#define UTIL_THROW_IF(Condition, Exception, Modify) do { \
if (Condition) { \
Exception UTIL_e; \
UTIL_SET_LOCATION(UTIL_e, #Exception, #Condition); \
UTIL_e << Modify; \
throw UTIL_e; \
} \
} while (0)
class ErrnoException : public Exception {
public:
ErrnoException() throw();
virtual ~ErrnoException() throw();
int Error() const throw() { return errno_; }
private:
int errno_;
};
class EndOfFileException : public Exception {
public:
EndOfFileException() throw();
~EndOfFileException() throw();
};
} // namespace util
#endif // UTIL_EXCEPTION__

294
util/file.cc
View File

@ -1,294 +0,0 @@
#include "util/file.hh"
#include "util/exception.hh"
#include <cstdlib>
#include <cstdio>
#include <iostream>
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <stdint.h>
#if defined(_WIN32) || defined(_WIN64)
#include <windows.h>
#include <io.h>
#else
#include <unistd.h>
#endif
namespace util {
scoped_fd::~scoped_fd() {
if (fd_ != -1 && close(fd_)) {
std::cerr << "Could not close file " << fd_ << std::endl;
std::abort();
}
}
scoped_FILE::~scoped_FILE() {
if (file_ && std::fclose(file_)) {
std::cerr << "Could not close file " << std::endl;
std::abort();
}
}
int OpenReadOrThrow(const char *name) {
int ret;
#if defined(_WIN32) || defined(_WIN64)
UTIL_THROW_IF(-1 == (ret = _open(name, _O_BINARY | _O_RDONLY)), ErrnoException, "while opening " << name);
#else
UTIL_THROW_IF(-1 == (ret = open(name, O_RDONLY)), ErrnoException, "while opening " << name);
#endif
return ret;
}
int CreateOrThrow(const char *name) {
int ret;
#if defined(_WIN32) || defined(_WIN64)
UTIL_THROW_IF(-1 == (ret = _open(name, _O_CREAT | _O_TRUNC | _O_RDWR, _S_IREAD | _S_IWRITE)), ErrnoException, "while creating " << name);
#else
UTIL_THROW_IF(-1 == (ret = open(name, O_CREAT | O_TRUNC | O_RDWR, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH)), ErrnoException, "while creating " << name);
#endif
return ret;
}
uint64_t SizeFile(int fd) {
#if defined(_WIN32) || defined(_WIN64)
__int64 ret = _filelengthi64(fd);
return (ret == -1) ? kBadSize : ret;
#else
struct stat sb;
if (fstat(fd, &sb) == -1 || (!sb.st_size && !S_ISREG(sb.st_mode))) return kBadSize;
return sb.st_size;
#endif
}
void ResizeOrThrow(int fd, uint64_t to) {
#if defined(_WIN32) || defined(_WIN64)
UTIL_THROW_IF(_chsize_s(fd, to), ErrnoException, "Resizing to " << to << " bytes failed");
#else
UTIL_THROW_IF(ftruncate(fd, to), ErrnoException, "Resizing to " << to << " bytes failed");
#endif
}
#ifdef WIN32
typedef int ssize_t;
#endif
void ReadOrThrow(int fd, void *to_void, std::size_t amount) {
uint8_t *to = static_cast<uint8_t*>(to_void);
while (amount) {
ssize_t ret = read(fd, to, amount);
UTIL_THROW_IF(ret == -1, ErrnoException, "Reading " << amount << " from fd " << fd << " failed.");
UTIL_THROW_IF(ret == 0, EndOfFileException, "Hit EOF in fd " << fd << " but there should be " << amount << " more bytes to read.");
amount -= ret;
to += ret;
}
}
std::size_t ReadOrEOF(int fd, void *to_void, std::size_t amount) {
uint8_t *to = static_cast<uint8_t*>(to_void);
std::size_t remaining = amount;
while (remaining) {
ssize_t ret = read(fd, to, remaining);
UTIL_THROW_IF(ret == -1, ErrnoException, "Reading " << remaining << " from fd " << fd << " failed.");
if (!ret) return amount - remaining;
remaining -= ret;
to += ret;
}
return amount;
}
void WriteOrThrow(int fd, const void *data_void, std::size_t size) {
const uint8_t *data = static_cast<const uint8_t*>(data_void);
while (size) {
ssize_t ret = write(fd, data, size);
if (ret < 1) UTIL_THROW(util::ErrnoException, "Write failed");
data += ret;
size -= ret;
}
}
void FSyncOrThrow(int fd) {
// Apparently windows doesn't have fsync?
#if !defined(_WIN32) && !defined(_WIN64)
UTIL_THROW_IF(-1 == fsync(fd), ErrnoException, "Sync of " << fd << " failed.");
#endif
}
namespace {
void InternalSeek(int fd, off_t off, int whence) {
UTIL_THROW_IF((off_t)-1 == lseek(fd, off, whence), ErrnoException, "Seek failed");
}
} // namespace
void SeekOrThrow(int fd, uint64_t off) {
InternalSeek(fd, off, SEEK_SET);
}
void AdvanceOrThrow(int fd, int64_t off) {
InternalSeek(fd, off, SEEK_CUR);
}
void SeekEnd(int fd) {
InternalSeek(fd, 0, SEEK_END);
}
std::FILE *FDOpenOrThrow(scoped_fd &file) {
std::FILE *ret = fdopen(file.get(), "r+b");
if (!ret) UTIL_THROW(util::ErrnoException, "Could not fdopen");
file.release();
return ret;
}
TempMaker::TempMaker(const std::string &prefix) : base_(prefix) {
base_ += "XXXXXX";
}
// Sigh. Windows temporary file creation is full of race conditions.
#if defined(_WIN32) || defined(_WIN64)
/* mkstemp extracted from libc/sysdeps/posix/tempname.c. Copyright
(C) 1991-1999, 2000, 2001, 2006 Free Software Foundation, Inc.
The GNU C Library is free software; you can redistribute it and/or
modify it under the terms of the GNU Lesser General Public
License as published by the Free Software Foundation; either
version 2.1 of the License, or (at your option) any later version. */
/* This has been modified from the original version to rename the function and
* set the Windows temporary flag. */
static const char letters[] =
"abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789";
/* Generate a temporary file name based on TMPL. TMPL must match the
rules for mk[s]temp (i.e. end in "XXXXXX"). The name constructed
does not exist at the time of the call to mkstemp. TMPL is
overwritten with the result. */
int
mkstemp_and_unlink(char *tmpl, bool and_unlink)
{
int len;
char *XXXXXX;
static unsigned long long value;
unsigned long long random_time_bits;
unsigned int count;
int fd = -1;
int save_errno = errno;
/* A lower bound on the number of temporary files to attempt to
generate. The maximum total number of temporary file names that
can exist for a given template is 62**6. It should never be
necessary to try all these combinations. Instead if a reasonable
number of names is tried (we define reasonable as 62**3) fail to
give the system administrator the chance to remove the problems. */
#define ATTEMPTS_MIN (62 * 62 * 62)
/* The number of times to attempt to generate a temporary file. To
conform to POSIX, this must be no smaller than TMP_MAX. */
#if ATTEMPTS_MIN < TMP_MAX
unsigned int attempts = TMP_MAX;
#else
unsigned int attempts = ATTEMPTS_MIN;
#endif
len = strlen (tmpl);
if (len < 6 || strcmp (&tmpl[len - 6], "XXXXXX"))
{
errno = EINVAL;
return -1;
}
/* This is where the Xs start. */
XXXXXX = &tmpl[len - 6];
/* Get some more or less random data. */
{
SYSTEMTIME stNow;
FILETIME ftNow;
// get system time
GetSystemTime(&stNow);
stNow.wMilliseconds = 500;
if (!SystemTimeToFileTime(&stNow, &ftNow))
{
errno = -1;
return -1;
}
random_time_bits = (((unsigned long long)ftNow.dwHighDateTime << 32)
| (unsigned long long)ftNow.dwLowDateTime);
}
value += random_time_bits ^ (unsigned long long)GetCurrentThreadId ();
for (count = 0; count < attempts; value += 7777, ++count)
{
unsigned long long v = value;
/* Fill in the random bits. */
XXXXXX[0] = letters[v % 62];
v /= 62;
XXXXXX[1] = letters[v % 62];
v /= 62;
XXXXXX[2] = letters[v % 62];
v /= 62;
XXXXXX[3] = letters[v % 62];
v /= 62;
XXXXXX[4] = letters[v % 62];
v /= 62;
XXXXXX[5] = letters[v % 62];
/* Modified for windows and to unlink */
// fd = open (tmpl, O_RDWR | O_CREAT | O_EXCL, _S_IREAD | _S_IWRITE);
int flags = _O_RDWR | _O_CREAT | _O_EXCL | _O_BINARY;
if (and_unlink) flags |= _O_TEMPORARY;
fd = _open (tmpl, flags, _S_IREAD | _S_IWRITE);
if (fd >= 0)
{
errno = save_errno;
return fd;
}
else if (errno != EEXIST)
return -1;
}
/* We got out of the loop because we ran out of combinations to try. */
errno = EEXIST;
return -1;
}
#else
int
mkstemp_and_unlink(char *tmpl, bool and_unlink) {
int ret = mkstemp(tmpl);
if (ret != -1 && and_unlink) {
UTIL_THROW_IF(unlink(tmpl), util::ErrnoException, "Failed to delete " << tmpl);
}
return ret;
}
#endif
int TempMaker::Make() const {
std::string name(base_);
name.push_back(0);
int ret;
UTIL_THROW_IF(-1 == (ret = mkstemp_and_unlink(&name[0], true)), util::ErrnoException, "Failed to make a temporary based on " << base_);
return ret;
}
std::FILE *TempMaker::MakeFile() const {
util::scoped_fd file(Make());
return FDOpenOrThrow(file);
}
std::string TempMaker::Name(scoped_fd &opened) const {
std::string name(base_);
name.push_back(0);
int fd;
UTIL_THROW_IF(-1 == (fd = mkstemp_and_unlink(&name[0], false)), util::ErrnoException, "Failed to make a temporary based on " << base_);
opened.reset(fd);
return name;
}
} // namespace util

110
util/file.hh
View File

@ -1,110 +0,0 @@
#ifndef UTIL_FILE__
#define UTIL_FILE__
#include <cstddef>
#include <cstdio>
#include <string>
#include <stdint.h>
namespace util {
class scoped_fd {
public:
scoped_fd() : fd_(-1) {}
explicit scoped_fd(int fd) : fd_(fd) {}
~scoped_fd();
void reset(int to) {
scoped_fd other(fd_);
fd_ = to;
}
int get() const { return fd_; }
int operator*() const { return fd_; }
int release() {
int ret = fd_;
fd_ = -1;
return ret;
}
operator bool() { return fd_ != -1; }
private:
int fd_;
scoped_fd(const scoped_fd &);
scoped_fd &operator=(const scoped_fd &);
};
class scoped_FILE {
public:
explicit scoped_FILE(std::FILE *file = NULL) : file_(file) {}
~scoped_FILE();
std::FILE *get() { return file_; }
const std::FILE *get() const { return file_; }
void reset(std::FILE *to = NULL) {
scoped_FILE other(file_);
file_ = to;
}
std::FILE *release() {
std::FILE *ret = file_;
file_ = NULL;
return ret;
}
private:
std::FILE *file_;
};
// Open for read only.
int OpenReadOrThrow(const char *name);
// Create file if it doesn't exist, truncate if it does. Opened for write.
int CreateOrThrow(const char *name);
// Return value for SizeFile when it can't size properly.
const uint64_t kBadSize = (uint64_t)-1;
uint64_t SizeFile(int fd);
void ResizeOrThrow(int fd, uint64_t to);
void ReadOrThrow(int fd, void *to, std::size_t size);
std::size_t ReadOrEOF(int fd, void *to_void, std::size_t amount);
void WriteOrThrow(int fd, const void *data_void, std::size_t size);
void FSyncOrThrow(int fd);
// Seeking
void SeekOrThrow(int fd, uint64_t off);
void AdvanceOrThrow(int fd, int64_t off);
void SeekEnd(int fd);
std::FILE *FDOpenOrThrow(scoped_fd &file);
class TempMaker {
public:
explicit TempMaker(const std::string &prefix);
// These will already be unlinked for you.
int Make() const;
std::FILE *MakeFile() const;
// This will force you to close the fd instead of leaving it open.
std::string Name(scoped_fd &opened) const;
private:
std::string base_;
};
} // namespace util
#endif // UTIL_FILE__

312
util/file_piece.cc
View File

@ -1,312 +0,0 @@
#include "util/file_piece.hh"
#include "util/exception.hh"
#include "util/file.hh"
#include "util/mmap.hh"
#ifdef WIN32
#include <io.h>
#endif // WIN32
#include <iostream>
#include <string>
#include <limits>
#include <assert.h>
#include <ctype.h>
#include <fcntl.h>
#include <stdlib.h>
#include <sys/types.h>
#include <sys/stat.h>
namespace util {
ParseNumberException::ParseNumberException(StringPiece value) throw() {
*this << "Could not parse \"" << value << "\" into a number";
}
#ifdef HAVE_ZLIB
GZException::GZException(gzFile file) {
int num;
*this << gzerror(file, &num) << " from zlib";
}
#endif // HAVE_ZLIB
// Sigh this is the only way I could come up with to do a _const_ bool. It has ' ', '\f', '\n', '\r', '\t', and '\v' (same as isspace on C locale).
const bool kSpaces[256] = {0,0,0,0,0,0,0,0,0,1,1,1,1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
FilePiece::FilePiece(const char *name, std::ostream *show_progress, std::size_t min_buffer) :
file_(OpenReadOrThrow(name)), total_size_(SizeFile(file_.get())), page_(SizePage()),
progress_(total_size_, total_size_ == kBadSize ? NULL : show_progress, std::string("Reading ") + name) {
Initialize(name, show_progress, min_buffer);
}
FilePiece::FilePiece(int fd, const char *name, std::ostream *show_progress, std::size_t min_buffer) :
file_(fd), total_size_(SizeFile(file_.get())), page_(SizePage()),
progress_(total_size_, total_size_ == kBadSize ? NULL : show_progress, std::string("Reading ") + name) {
Initialize(name, show_progress, min_buffer);
}
FilePiece::~FilePiece() {
#ifdef HAVE_ZLIB
if (gz_file_) {
// zlib took ownership
file_.release();
int ret;
if (Z_OK != (ret = gzclose(gz_file_))) {
std::cerr << "could not close file " << file_name_ << " using zlib" << std::endl;
abort();
}
}
#endif
}
StringPiece FilePiece::ReadLine(char delim) {
std::size_t skip = 0;
while (true) {
for (const char *i = position_ + skip; i < position_end_; ++i) {
if (*i == delim) {
StringPiece ret(position_, i - position_);
position_ = i + 1;
return ret;
}
}
if (at_end_) {
if (position_ == position_end_) Shift();
return Consume(position_end_);
}
skip = position_end_ - position_;
Shift();
}
}
float FilePiece::ReadFloat() {
return ReadNumber<float>();
}
double FilePiece::ReadDouble() {
return ReadNumber<double>();
}
long int FilePiece::ReadLong() {
return ReadNumber<long int>();
}
unsigned long int FilePiece::ReadULong() {
return ReadNumber<unsigned long int>();
}
void FilePiece::Initialize(const char *name, std::ostream *show_progress, std::size_t min_buffer) {
#ifdef HAVE_ZLIB
gz_file_ = NULL;
#endif
file_name_ = name;
default_map_size_ = page_ * std::max<std::size_t>((min_buffer / page_ + 1), 2);
position_ = NULL;
position_end_ = NULL;
mapped_offset_ = 0;
at_end_ = false;
if (total_size_ == kBadSize) {
// So the assertion passes.
fallback_to_read_ = false;
if (show_progress)
*show_progress << "File " << name << " isn't normal. Using slower read() instead of mmap(). No progress bar." << std::endl;
TransitionToRead();
} else {
fallback_to_read_ = false;
}
Shift();
// gzip detect.
if ((position_end_ - position_) > 2 && *position_ == 0x1f && static_cast<unsigned char>(*(position_ + 1)) == 0x8b) {
#ifndef HAVE_ZLIB
UTIL_THROW(GZException, "Looks like a gzip file but support was not compiled in.");
#endif
if (!fallback_to_read_) {
at_end_ = false;
TransitionToRead();
}
}
}
namespace {
void ParseNumber(const char *begin, char *&end, float &out) {
#if defined(sun) || defined(WIN32)
out = static_cast<float>(strtod(begin, &end));
#else
out = strtof(begin, &end);
#endif
}
void ParseNumber(const char *begin, char *&end, double &out) {
out = strtod(begin, &end);
}
void ParseNumber(const char *begin, char *&end, long int &out) {
out = strtol(begin, &end, 10);
}
void ParseNumber(const char *begin, char *&end, unsigned long int &out) {
out = strtoul(begin, &end, 10);
}
} // namespace
template <class T> T FilePiece::ReadNumber() {
SkipSpaces();
while (last_space_ < position_) {
if (at_end_) {
// Hallucinate a null off the end of the file.
std::string buffer(position_, position_end_);
char *end;
T ret;
ParseNumber(buffer.c_str(), end, ret);
if (buffer.c_str() == end) throw ParseNumberException(buffer);
position_ += end - buffer.c_str();
return ret;
}
Shift();
}
char *end;
T ret;
ParseNumber(position_, end, ret);
if (end == position_) throw ParseNumberException(ReadDelimited());
position_ = end;
return ret;
}
const char *FilePiece::FindDelimiterOrEOF(const bool *delim) {
std::size_t skip = 0;
while (true) {
for (const char *i = position_ + skip; i < position_end_; ++i) {
if (delim[static_cast<unsigned char>(*i)]) return i;
}
if (at_end_) {
if (position_ == position_end_) Shift();
return position_end_;
}
skip = position_end_ - position_;
Shift();
}
}
void FilePiece::Shift() {
if (at_end_) {
progress_.Finished();
throw EndOfFileException();
}
uint64_t desired_begin = position_ - data_.begin() + mapped_offset_;
if (!fallback_to_read_) MMapShift(desired_begin);
// Notice an mmap failure might set the fallback.
if (fallback_to_read_) ReadShift();
for (last_space_ = position_end_ - 1; last_space_ >= position_; --last_space_) {
if (isspace(*last_space_)) break;
}
}
void FilePiece::MMapShift(uint64_t desired_begin) {
// Use mmap.
uint64_t ignore = desired_begin % page_;
// Duplicate request for Shift means give more data.
if (position_ == data_.begin() + ignore) {
default_map_size_ *= 2;
}
// Local version so that in case of failure it doesn't overwrite the class variable.
uint64_t mapped_offset = desired_begin - ignore;
uint64_t mapped_size;
if (default_map_size_ >= static_cast<std::size_t>(total_size_ - mapped_offset)) {
at_end_ = true;
mapped_size = total_size_ - mapped_offset;
} else {
mapped_size = default_map_size_;
}
// Forcibly clear the existing mmap first.
data_.reset();
try {
MapRead(POPULATE_OR_LAZY, *file_, mapped_offset, mapped_size, data_);
} catch (const util::ErrnoException &e) {
if (desired_begin) {
SeekOrThrow(*file_, desired_begin);
}
// The mmap was scheduled to end the file, but now we're going to read it.
at_end_ = false;
TransitionToRead();
return;
}
mapped_offset_ = mapped_offset;
position_ = data_.begin() + ignore;
position_end_ = data_.begin() + mapped_size;
progress_.Set(desired_begin);
}
void FilePiece::TransitionToRead() {
assert(!fallback_to_read_);
fallback_to_read_ = true;
data_.reset();
data_.reset(malloc(default_map_size_), default_map_size_, scoped_memory::MALLOC_ALLOCATED);
UTIL_THROW_IF(!data_.get(), ErrnoException, "malloc failed for " << default_map_size_);
position_ = data_.begin();
position_end_ = position_;
#ifdef HAVE_ZLIB
assert(!gz_file_);
gz_file_ = gzdopen(file_.get(), "r");
UTIL_THROW_IF(!gz_file_, GZException, "zlib failed to open " << file_name_);
#endif
}
#ifdef WIN32
typedef int ssize_t;
#endif
void FilePiece::ReadShift() {
assert(fallback_to_read_);
// Bytes [data_.begin(), position_) have been consumed.
// Bytes [position_, position_end_) have been read into the buffer.
// Start at the beginning of the buffer if there's nothing useful in it.
if (position_ == position_end_) {
mapped_offset_ += (position_end_ - data_.begin());
position_ = data_.begin();
position_end_ = position_;
}
std::size_t already_read = position_end_ - data_.begin();
if (already_read == default_map_size_) {
if (position_ == data_.begin()) {
// Buffer too small.
std::size_t valid_length = position_end_ - position_;
default_map_size_ *= 2;
data_.call_realloc(default_map_size_);
UTIL_THROW_IF(!data_.get(), ErrnoException, "realloc failed for " << default_map_size_);
position_ = data_.begin();
position_end_ = position_ + valid_length;
} else {
size_t moving = position_end_ - position_;
memmove(data_.get(), position_, moving);
position_ = data_.begin();
position_end_ = position_ + moving;
already_read = moving;
}
}
ssize_t read_return;
#ifdef HAVE_ZLIB
read_return = gzread(gz_file_, static_cast<char*>(data_.get()) + already_read, default_map_size_ - already_read);
if (read_return == -1) throw GZException(gz_file_);
if (total_size_ != kBadSize) {
// Just get the position, don't actually seek. Apparently this is how you do it. . .
off_t ret = lseek(file_.get(), 0, SEEK_CUR);
if (ret != -1) progress_.Set(ret);
}
#else
read_return = read(file_.get(), static_cast<char*>(data_.get()) + already_read, default_map_size_ - already_read);
UTIL_THROW_IF(read_return == -1, ErrnoException, "read failed");
progress_.Set(mapped_offset_);
#endif
if (read_return == 0) {
at_end_ = true;
}
position_end_ += read_return;
}
} // namespace util

132
util/file_piece.hh
View File

@ -1,132 +0,0 @@
#ifndef UTIL_FILE_PIECE__
#define UTIL_FILE_PIECE__
#include "util/ersatz_progress.hh"
#include "util/exception.hh"
#include "util/file.hh"
#include "util/have.hh"
#include "util/mmap.hh"
#include "util/string_piece.hh"
#include <cstddef>
#include <string>
#include <stdint.h>
#ifdef HAVE_ZLIB
#include <zlib.h>
#endif
namespace util {
class ParseNumberException : public Exception {
public:
explicit ParseNumberException(StringPiece value) throw();
~ParseNumberException() throw() {}
};
class GZException : public Exception {
public:
#ifdef HAVE_ZLIB
explicit GZException(gzFile file);
#endif
GZException() throw() {}
~GZException() throw() {}
};
extern const bool kSpaces[256];
// Memory backing the returned StringPiece may vanish on the next call.
class FilePiece {
public:
// 32 MB default.
explicit FilePiece(const char *file, std::ostream *show_progress = NULL, std::size_t min_buffer = 33554432);
// Takes ownership of fd. name is used for messages.
explicit FilePiece(int fd, const char *name, std::ostream *show_progress = NULL, std::size_t min_buffer = 33554432);
~FilePiece();
char get() {
if (position_ == position_end_) {
Shift();
if (at_end_) throw EndOfFileException();
}
return *(position_++);
}
// Leaves the delimiter, if any, to be returned by get(). Delimiters defined by isspace().
StringPiece ReadDelimited(const bool *delim = kSpaces) {
SkipSpaces(delim);
return Consume(FindDelimiterOrEOF(delim));
}
// Unlike ReadDelimited, this includes leading spaces and consumes the delimiter.
// It is similar to getline in that way.
StringPiece ReadLine(char delim = '\n');
float ReadFloat();
double ReadDouble();
long int ReadLong();
unsigned long int ReadULong();
// Skip spaces defined by isspace.
void SkipSpaces(const bool *delim = kSpaces) {
for (; ; ++position_) {
if (position_ == position_end_) Shift();
if (!delim[static_cast<unsigned char>(*position_)]) return;
}
}
uint64_t Offset() const {
return position_ - data_.begin() + mapped_offset_;
}
const std::string &FileName() const { return file_name_; }
private:
void Initialize(const char *name, std::ostream *show_progress, std::size_t min_buffer);
template <class T> T ReadNumber();
StringPiece Consume(const char *to) {
StringPiece ret(position_, to - position_);
position_ = to;
return ret;
}
const char *FindDelimiterOrEOF(const bool *delim = kSpaces);
void Shift();
// Backends to Shift().
void MMapShift(uint64_t desired_begin);
void TransitionToRead();
void ReadShift();
const char *position_, *last_space_, *position_end_;
scoped_fd file_;
const uint64_t total_size_;
const uint64_t page_;
std::size_t default_map_size_;
uint64_t mapped_offset_;
// Order matters: file_ should always be destroyed after this.
scoped_memory data_;
bool at_end_;
bool fallback_to_read_;
ErsatzProgress progress_;
std::string file_name_;
#ifdef HAVE_ZLIB
gzFile gz_file_;
#endif // HAVE_ZLIB
};
} // namespace util
#endif // UTIL_FILE_PIECE__

123
util/file_piece_test.cc
View File

@ -1,123 +0,0 @@
// Tests might fail if you have creative characters in your path. Sue me.
#include "util/file_piece.hh"
#include "util/scoped.hh"
#define BOOST_TEST_MODULE FilePieceTest
#include <boost/test/unit_test.hpp>
#include <fstream>
#include <iostream>
#include <stdio.h>
#include <sys/types.h>
#include <sys/stat.h>
namespace util {
namespace {
std::string FileLocation() {
if (boost::unit_test::framework::master_test_suite().argc < 2) {
return "file_piece.cc";
}
std::string ret(boost::unit_test::framework::master_test_suite().argv[1]);
return ret;
}
/* mmap implementation */
BOOST_AUTO_TEST_CASE(MMapReadLine) {
std::fstream ref(FileLocation().c_str(), std::ios::in);
FilePiece test(FileLocation().c_str(), NULL, 1);
std::string ref_line;
while (getline(ref, ref_line)) {
StringPiece test_line(test.ReadLine());
// I submitted a bug report to ICU: http://bugs.icu-project.org/trac/ticket/7924
if (!test_line.empty() || !ref_line.empty()) {
BOOST_CHECK_EQUAL(ref_line, test_line);
}
}
BOOST_CHECK_THROW(test.get(), EndOfFileException);
}
#ifndef __APPLE__
/* Apple isn't happy with the popen, fileno, dup. And I don't want to
* reimplement popen. This is an issue with the test.
*/
/* read() implementation */
BOOST_AUTO_TEST_CASE(StreamReadLine) {
std::fstream ref(FileLocation().c_str(), std::ios::in);
std::string popen_args = "cat \"";
popen_args += FileLocation();
popen_args += '"';
FILE *catter = popen(popen_args.c_str(), "r");
BOOST_REQUIRE(catter);
FilePiece test(dup(fileno(catter)), "file_piece.cc", NULL, 1);
std::string ref_line;
while (getline(ref, ref_line)) {
StringPiece test_line(test.ReadLine());
// I submitted a bug report to ICU: http://bugs.icu-project.org/trac/ticket/7924
if (!test_line.empty() || !ref_line.empty()) {
BOOST_CHECK_EQUAL(ref_line, test_line);
}
}
BOOST_CHECK_THROW(test.get(), EndOfFileException);
BOOST_REQUIRE(!pclose(catter));
}
#endif // __APPLE__
#ifdef HAVE_ZLIB
// gzip file
BOOST_AUTO_TEST_CASE(PlainZipReadLine) {
std::string location(FileLocation());
std::fstream ref(location.c_str(), std::ios::in);
std::string command("gzip <\"");
command += location + "\" >\"" + location + "\".gz";
BOOST_REQUIRE_EQUAL(0, system(command.c_str()));
FilePiece test((location + ".gz").c_str(), NULL, 1);
unlink((location + ".gz").c_str());
std::string ref_line;
while (getline(ref, ref_line)) {
StringPiece test_line(test.ReadLine());
// I submitted a bug report to ICU: http://bugs.icu-project.org/trac/ticket/7924
if (!test_line.empty() || !ref_line.empty()) {
BOOST_CHECK_EQUAL(ref_line, test_line);
}
}
BOOST_CHECK_THROW(test.get(), EndOfFileException);
}
// gzip stream. Apple doesn't like popen, fileno, dup. This is an issue with
// the test.
#ifndef __APPLE__
BOOST_AUTO_TEST_CASE(StreamZipReadLine) {
std::fstream ref(FileLocation().c_str(), std::ios::in);
std::string command("gzip <\"");
command += FileLocation() + "\"";
FILE * catter = popen(command.c_str(), "r");
BOOST_REQUIRE(catter);
FilePiece test(dup(fileno(catter)), "file_piece.cc.gz", NULL, 1);
std::string ref_line;
while (getline(ref, ref_line)) {
StringPiece test_line(test.ReadLine());
// I submitted a bug report to ICU: http://bugs.icu-project.org/trac/ticket/7924
if (!test_line.empty() || !ref_line.empty()) {
BOOST_CHECK_EQUAL(ref_line, test_line);
}
}
BOOST_CHECK_THROW(test.get(), EndOfFileException);
BOOST_REQUIRE(!pclose(catter));
}
#endif // __APPLE__
#endif // HAVE_ZLIB
} // namespace
} // namespace util

78
util/getopt.c
View File

@ -1,78 +0,0 @@
/*
POSIX getopt for Windows
AT&T Public License
Code given out at the 1985 UNIFORUM conference in Dallas.
*/
#ifndef __GNUC__
#include "getopt.hh"
#include <stdio.h>
#include <string.h>
#define NULL 0
#define EOF (-1)
#define ERR(s, c) if(opterr){\
char errbuf[2];\
errbuf[0] = c; errbuf[1] = '\n';\
fputs(argv[0], stderr);\
fputs(s, stderr);\
fputc(c, stderr);}
//(void) write(2, argv[0], (unsigned)strlen(argv[0]));\
//(void) write(2, s, (unsigned)strlen(s));\
//(void) write(2, errbuf, 2);}
int opterr = 1;
int optind = 1;
int optopt;
char *optarg;
int
getopt(argc, argv, opts)
int argc;
char **argv, *opts;
{
static int sp = 1;
register int c;
register char *cp;
if(sp == 1)
if(optind >= argc ||
argv[optind][0] != '-' || argv[optind][1] == '\0')
return(EOF);
else if(strcmp(argv[optind], "--") == NULL) {
optind++;
return(EOF);
}
optopt = c = argv[optind][sp];
if(c == ':' || (cp=strchr(opts, c)) == NULL) {
ERR(": illegal option -- ", c);
if(argv[optind][++sp] == '\0') {
optind++;
sp = 1;
}
return('?');
}
if(*++cp == ':') {
if(argv[optind][sp+1] != '\0')
optarg = &argv[optind++][sp+1];
else if(++optind >= argc) {
ERR(": option requires an argument -- ", c);
sp = 1;
return('?');
} else
optarg = argv[optind++];
sp = 1;
} else {
if(argv[optind][++sp] == '\0') {
sp = 1;
optind++;
}
optarg = NULL;
}
return(c);
}
#endif /* __GNUC__ */

33
util/getopt.hh
View File

@ -1,33 +0,0 @@
/*
POSIX getopt for Windows
AT&T Public License
Code given out at the 1985 UNIFORUM conference in Dallas.
*/
#ifdef __GNUC__
#include <getopt.h>
#endif
#ifndef __GNUC__
#ifndef _WINGETOPT_H_
#define _WINGETOPT_H_
#ifdef __cplusplus
extern "C" {
#endif
extern int opterr;
extern int optind;
extern int optopt;
extern char *optarg;
extern int getopt(int argc, char **argv, char *opts);
#ifdef __cplusplus
}
#endif
#endif /* _GETOPT_H_ */
#endif /* __GNUC__ */

23
util/have.hh
View File

@ -1,23 +0,0 @@
/* Optional packages. You might want to integrate this with your build system e.g. config.h from ./configure. */
#ifndef UTIL_HAVE__
#define UTIL_HAVE__
#ifndef HAVE_ZLIB
#if !defined(_WIN32) && !defined(_WIN64)
#define HAVE_ZLIB
#endif
#endif
#ifndef HAVE_ICU
//#define HAVE_ICU
#endif
#ifndef HAVE_BOOST
//#define HAVE_BOOST
#endif
#ifndef HAVE_THREADS
//#define HAVE_THREADS
#endif
#endif // UTIL_HAVE__

151
util/joint_sort.hh
View File

@ -1,151 +0,0 @@
#ifndef UTIL_JOINT_SORT__
#define UTIL_JOINT_SORT__
/* A terrifying amount of C++ to coax std::sort into soring one range while
* also permuting another range the same way.
*/
#include "util/proxy_iterator.hh"
#include <algorithm>
#include <functional>
#include <iostream>
namespace util {
namespace detail {
template <class KeyIter, class ValueIter> class JointProxy;
template <class KeyIter, class ValueIter> class JointIter {
public:
JointIter() {}
JointIter(const KeyIter &key_iter, const ValueIter &value_iter) : key_(key_iter), value_(value_iter) {}
bool operator==(const JointIter<KeyIter, ValueIter> &other) const { return key_ == other.key_; }
bool operator<(const JointIter<KeyIter, ValueIter> &other) const { return (key_ < other.key_); }
std::ptrdiff_t operator-(const JointIter<KeyIter, ValueIter> &other) const { return key_ - other.key_; }
JointIter<KeyIter, ValueIter> &operator+=(std::ptrdiff_t amount) {
key_ += amount;
value_ += amount;
return *this;
}
void swap(const JointIter &other) {
std::swap(key_, other.key_);
std::swap(value_, other.value_);
}
private:
friend class JointProxy<KeyIter, ValueIter>;
KeyIter key_;
ValueIter value_;
};
template <class KeyIter, class ValueIter> class JointProxy {
private:
typedef JointIter<KeyIter, ValueIter> InnerIterator;
public:
typedef struct {
typename std::iterator_traits<KeyIter>::value_type key;
typename std::iterator_traits<ValueIter>::value_type value;
const typename std::iterator_traits<KeyIter>::value_type &GetKey() const { return key; }
} value_type;
JointProxy(const KeyIter &key_iter, const ValueIter &value_iter) : inner_(key_iter, value_iter) {}
JointProxy(const JointProxy<KeyIter, ValueIter> &other) : inner_(other.inner_) {}
operator const value_type() const {
value_type ret;
ret.key = *inner_.key_;
ret.value = *inner_.value_;
return ret;
}
JointProxy &operator=(const JointProxy &other) {
*inner_.key_ = *other.inner_.key_;
*inner_.value_ = *other.inner_.value_;
return *this;
}
JointProxy &operator=(const value_type &other) {
*inner_.key_ = other.key;
*inner_.value_ = other.value;
return *this;
}
typename std::iterator_traits<KeyIter>::reference GetKey() const {
return *(inner_.key_);
}
void swap(JointProxy<KeyIter, ValueIter> &other) {
std::swap(*inner_.key_, *other.inner_.key_);
std::swap(*inner_.value_, *other.inner_.value_);
}
private:
friend class ProxyIterator<JointProxy<KeyIter, ValueIter> >;
InnerIterator &Inner() { return inner_; }
const InnerIterator &Inner() const { return inner_; }
InnerIterator inner_;
};
template <class Proxy, class Less> class LessWrapper : public std::binary_function<const typename Proxy::value_type &, const typename Proxy::value_type &, bool> {
public:
explicit LessWrapper(const Less &less) : less_(less) {}
bool operator()(const Proxy &left, const Proxy &right) const {
return less_(left.GetKey(), right.GetKey());
}
bool operator()(const Proxy &left, const typename Proxy::value_type &right) const {
return less_(left.GetKey(), right.GetKey());
}
bool operator()(const typename Proxy::value_type &left, const Proxy &right) const {
return less_(left.GetKey(), right.GetKey());
}
bool operator()(const typename Proxy::value_type &left, const typename Proxy::value_type &right) const {
return less_(left.GetKey(), right.GetKey());
}
private:
const Less less_;
};
} // namespace detail
template <class KeyIter, class ValueIter> class PairedIterator : public ProxyIterator<detail::JointProxy<KeyIter, ValueIter> > {
public:
PairedIterator(const KeyIter &key, const ValueIter &value) :
ProxyIterator<detail::JointProxy<KeyIter, ValueIter> >(detail::JointProxy<KeyIter, ValueIter>(key, value)) {}
};
template <class KeyIter, class ValueIter, class Less> void JointSort(const KeyIter &key_begin, const KeyIter &key_end, const ValueIter &value_begin, const Less &less) {
ProxyIterator<detail::JointProxy<KeyIter, ValueIter> > full_begin(detail::JointProxy<KeyIter, ValueIter>(key_begin, value_begin));
detail::LessWrapper<detail::JointProxy<KeyIter, ValueIter>, Less> less_wrap(less);
std::sort(full_begin, full_begin + (key_end - key_begin), less_wrap);
}
template <class KeyIter, class ValueIter> void JointSort(const KeyIter &key_begin, const KeyIter &key_end, const ValueIter &value_begin) {
JointSort(key_begin, key_end, value_begin, std::less<typename std::iterator_traits<KeyIter>::value_type>());
}
} // namespace util
namespace std {
template <class KeyIter, class ValueIter> void swap(util::detail::JointIter<KeyIter, ValueIter> &left, util::detail::JointIter<KeyIter, ValueIter> &right) {
left.swap(right);
}
template <class KeyIter, class ValueIter> void swap(util::detail::JointProxy<KeyIter, ValueIter> &left, util::detail::JointProxy<KeyIter, ValueIter> &right) {
left.swap(right);
}
} // namespace std
#endif // UTIL_JOINT_SORT__

50
util/joint_sort_test.cc
View File

@ -1,50 +0,0 @@
#include "util/joint_sort.hh"
#define BOOST_TEST_MODULE JointSortTest
#include <boost/test/unit_test.hpp>
namespace util { namespace {
BOOST_AUTO_TEST_CASE(just_flip) {
char keys[2];
int values[2];
keys[0] = 1; values[0] = 327;
keys[1] = 0; values[1] = 87897;
JointSort<char *, int *>(keys + 0, keys + 2, values + 0);
BOOST_CHECK_EQUAL(0, keys[0]);
BOOST_CHECK_EQUAL(87897, values[0]);
BOOST_CHECK_EQUAL(1, keys[1]);
BOOST_CHECK_EQUAL(327, values[1]);
}
BOOST_AUTO_TEST_CASE(three) {
char keys[3];
int values[3];
keys[0] = 1; values[0] = 327;
keys[1] = 2; values[1] = 87897;
keys[2] = 0; values[2] = 10;
JointSort<char *, int *>(keys + 0, keys + 3, values + 0);
BOOST_CHECK_EQUAL(0, keys[0]);
BOOST_CHECK_EQUAL(1, keys[1]);
BOOST_CHECK_EQUAL(2, keys[2]);
}
BOOST_AUTO_TEST_CASE(char_int) {
char keys[4];
int values[4];
keys[0] = 3; values[0] = 327;
keys[1] = 1; values[1] = 87897;
keys[2] = 2; values[2] = 10;
keys[3] = 0; values[3] = 24347;
JointSort<char *, int *>(keys + 0, keys + 4, values + 0);
BOOST_CHECK_EQUAL(0, keys[0]);
BOOST_CHECK_EQUAL(24347, values[0]);
BOOST_CHECK_EQUAL(1, keys[1]);
BOOST_CHECK_EQUAL(87897, values[1]);
BOOST_CHECK_EQUAL(2, keys[2]);
BOOST_CHECK_EQUAL(10, values[2]);
BOOST_CHECK_EQUAL(3, keys[3]);
BOOST_CHECK_EQUAL(327, values[3]);
}
}} // namespace anonymous util

184
util/mmap.cc
View File

@ -1,184 +0,0 @@
/* Memory mapping wrappers.
* ARM and MinGW ports contributed by Hideo Okuma and Tomoyuki Yoshimura at
* NICT.
*/
#include "util/mmap.hh"
#include "util/exception.hh"
#include "util/file.hh"
#include <iostream>
#include <assert.h>
#include <fcntl.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <stdlib.h>
#if defined(_WIN32) || defined(_WIN64)
#include <windows.h>
#include <io.h>
#else
#include <sys/mman.h>
#include <unistd.h>
#endif
namespace util {
long SizePage() {
#if defined(_WIN32) || defined(_WIN64)
SYSTEM_INFO si;
GetSystemInfo(&si);
return si.dwAllocationGranularity;
#else
return sysconf(_SC_PAGE_SIZE);
#endif
}
void SyncOrThrow(void *start, size_t length) {
#if defined(_WIN32) || defined(_WIN64)
UTIL_THROW_IF(!::FlushViewOfFile(start, length), ErrnoException, "Failed to sync mmap");
#else
UTIL_THROW_IF(msync(start, length, MS_SYNC), ErrnoException, "Failed to sync mmap");
#endif
}
void UnmapOrThrow(void *start, size_t length) {
#if defined(_WIN32) || defined(_WIN64)
UTIL_THROW_IF(!::UnmapViewOfFile(start), ErrnoException, "Failed to unmap a file");
#else
UTIL_THROW_IF(munmap(start, length), ErrnoException, "munmap failed");
#endif
}
scoped_mmap::~scoped_mmap() {
if (data_ != (void*)-1) {
try {
// Thanks Denis Filimonov for pointing out NFS likes msync first.
SyncOrThrow(data_, size_);
UnmapOrThrow(data_, size_);
} catch (const util::ErrnoException &e) {
std::cerr << e.what();
abort();
}
}
}
void scoped_memory::reset(void *data, std::size_t size, Alloc source) {
switch(source_) {
case MMAP_ALLOCATED:
scoped_mmap(data_, size_);
break;
case ARRAY_ALLOCATED:
delete [] reinterpret_cast<char*>(data_);
break;
case MALLOC_ALLOCATED:
free(data_);
break;
case NONE_ALLOCATED:
break;
}
data_ = data;
size_ = size;
source_ = source;
}
void scoped_memory::call_realloc(std::size_t size) {
assert(source_ == MALLOC_ALLOCATED || source_ == NONE_ALLOCATED);
void *new_data = realloc(data_, size);
if (!new_data) {
reset();
} else {
reset(new_data, size, MALLOC_ALLOCATED);
}
}
void *MapOrThrow(std::size_t size, bool for_write, int flags, bool prefault, int fd, uint64_t offset) {
#ifdef MAP_POPULATE // Linux specific
if (prefault) {
flags |= MAP_POPULATE;
}
#endif
#if defined(_WIN32) || defined(_WIN64)
int protectC = for_write ? PAGE_READWRITE : PAGE_READONLY;
int protectM = for_write ? FILE_MAP_WRITE : FILE_MAP_READ;
uint64_t total_size = size + offset;
HANDLE hMapping = CreateFileMapping((HANDLE)_get_osfhandle(fd), NULL, protectC, total_size >> 32, static_cast<DWORD>(total_size), NULL);
UTIL_THROW_IF(!hMapping, ErrnoException, "CreateFileMapping failed");
LPVOID ret = MapViewOfFile(hMapping, protectM, offset >> 32, offset, size);
CloseHandle(hMapping);
UTIL_THROW_IF(!ret, ErrnoException, "MapViewOfFile failed");
#else
int protect = for_write ? (PROT_READ | PROT_WRITE) : PROT_READ;
void *ret = mmap(NULL, size, protect, flags, fd, offset);
UTIL_THROW_IF(ret == MAP_FAILED, ErrnoException, "mmap failed for size " << size << " at offset " << offset);
#endif
return ret;
}
const int kFileFlags =
#if defined(_WIN32) || defined(_WIN64)
0 // MapOrThrow ignores flags on windows
#elif defined(MAP_FILE)
MAP_FILE | MAP_SHARED
#else
MAP_SHARED
#endif
;
void MapRead(LoadMethod method, int fd, uint64_t offset, std::size_t size, scoped_memory &out) {
switch (method) {
case LAZY:
out.reset(MapOrThrow(size, false, kFileFlags, false, fd, offset), size, scoped_memory::MMAP_ALLOCATED);
break;
case POPULATE_OR_LAZY:
#ifdef MAP_POPULATE
case POPULATE_OR_READ:
#endif
out.reset(MapOrThrow(size, false, kFileFlags, true, fd, offset), size, scoped_memory::MMAP_ALLOCATED);
break;
#ifndef MAP_POPULATE
case POPULATE_OR_READ:
#endif
case READ:
out.reset(malloc(size), size, scoped_memory::MALLOC_ALLOCATED);
if (!out.get()) UTIL_THROW(util::ErrnoException, "Allocating " << size << " bytes with malloc");
SeekOrThrow(fd, offset);
ReadOrThrow(fd, out.get(), size);
break;
}
}
// Allocates zeroed memory in to.
void MapAnonymous(std::size_t size, util::scoped_memory &to) {
to.reset();
#if defined(_WIN32) || defined(_WIN64)
to.reset(calloc(1, size), size, scoped_memory::MALLOC_ALLOCATED);
#else
to.reset(MapOrThrow(size, true,
# if defined(MAP_ANONYMOUS)
MAP_ANONYMOUS | MAP_PRIVATE // Linux
# else
MAP_ANON | MAP_PRIVATE // BSD
# endif
, false, -1, 0), size, scoped_memory::MMAP_ALLOCATED);
#endif
}
void *MapZeroedWrite(int fd, std::size_t size) {
ResizeOrThrow(fd, 0);
ResizeOrThrow(fd, size);
return MapOrThrow(size, true, kFileFlags, false, fd, 0);
}
void *MapZeroedWrite(const char *name, std::size_t size, scoped_fd &file) {
file.reset(CreateOrThrow(name));
try {
return MapZeroedWrite(file.get(), size);
} catch (ErrnoException &e) {
e << " in file " << name;
throw;
}
}
} // namespace util

114
util/mmap.hh
View File

@ -1,114 +0,0 @@
#ifndef UTIL_MMAP__
#define UTIL_MMAP__
// Utilities for mmaped files.
#include <cstddef>
#include <stdint.h>
#include <sys/types.h>
namespace util {
class scoped_fd;
long SizePage();
// (void*)-1 is MAP_FAILED; this is done to avoid including the mmap header here.
class scoped_mmap {
public:
scoped_mmap() : data_((void*)-1), size_(0) {}
scoped_mmap(void *data, std::size_t size) : data_(data), size_(size) {}
~scoped_mmap();
void *get() const { return data_; }
const uint8_t *begin() const { return reinterpret_cast<uint8_t*>(data_); }
const uint8_t *end() const { return reinterpret_cast<uint8_t*>(data_) + size_; }
std::size_t size() const { return size_; }
void reset(void *data, std::size_t size) {
scoped_mmap other(data_, size_);
data_ = data;
size_ = size;
}
void reset() {
reset((void*)-1, 0);
}
private:
void *data_;
std::size_t size_;
scoped_mmap(const scoped_mmap &);
scoped_mmap &operator=(const scoped_mmap &);
};
/* For when the memory might come from mmap, new char[], or malloc. Uses NULL
* and 0 for blanks even though mmap signals errors with (void*)-1). The reset
* function checks that blank for mmap.
*/
class scoped_memory {
public:
typedef enum {MMAP_ALLOCATED, ARRAY_ALLOCATED, MALLOC_ALLOCATED, NONE_ALLOCATED} Alloc;
scoped_memory() : data_(NULL), size_(0), source_(NONE_ALLOCATED) {}
~scoped_memory() { reset(); }
void *get() const { return data_; }
const char *begin() const { return reinterpret_cast<char*>(data_); }
const char *end() const { return reinterpret_cast<char*>(data_) + size_; }
std::size_t size() const { return size_; }
Alloc source() const { return source_; }
void reset() { reset(NULL, 0, NONE_ALLOCATED); }
void reset(void *data, std::size_t size, Alloc from);
// realloc allows the current data to escape hence the need for this call
// If realloc fails, destroys the original too and get() returns NULL.
void call_realloc(std::size_t to);
private:
void *data_;
std::size_t size_;
Alloc source_;
scoped_memory(const scoped_memory &);
scoped_memory &operator=(const scoped_memory &);
};
typedef enum {
// mmap with no prepopulate
LAZY,
// On linux, pass MAP_POPULATE to mmap.
POPULATE_OR_LAZY,
// Populate on Linux. malloc and read on non-Linux.
POPULATE_OR_READ,
// malloc and read.
READ
} LoadMethod;
extern const int kFileFlags;
// Wrapper around mmap to check it worked and hide some platform macros.
void *MapOrThrow(std::size_t size, bool for_write, int flags, bool prefault, int fd, uint64_t offset = 0);
void MapRead(LoadMethod method, int fd, uint64_t offset, std::size_t size, scoped_memory &out);
void MapAnonymous(std::size_t size, scoped_memory &to);
// Open file name with mmap of size bytes, all of which are initially zero.
void *MapZeroedWrite(int fd, std::size_t size);
void *MapZeroedWrite(const char *name, std::size_t size, scoped_fd &file);
// msync wrapper
void SyncOrThrow(void *start, size_t length);
} // namespace util
#endif // UTIL_MMAP__

168
util/murmur_hash.cc
View File

@ -1,168 +0,0 @@
/* Downloaded from http://sites.google.com/site/murmurhash/ which says "All
* code is released to the public domain. For business purposes, Murmurhash is
* under the MIT license."
* This is modified from the original:
* ULL tag on 0xc6a4a7935bd1e995 so this will compile on 32-bit.
* length changed to unsigned int.
* placed in namespace util
* add MurmurHashNative
* default option = 0 for seed
* ARM port from NICT
*/
#include "util/murmur_hash.hh"
#include <string.h>
namespace util {
//-----------------------------------------------------------------------------
// MurmurHash2, 64-bit versions, by Austin Appleby
// The same caveats as 32-bit MurmurHash2 apply here - beware of alignment
// and endian-ness issues if used across multiple platforms.
// 64-bit hash for 64-bit platforms
uint64_t MurmurHash64A ( const void * key, std::size_t len, uint64_t seed )
{
const uint64_t m = 0xc6a4a7935bd1e995ULL;
const int r = 47;
uint64_t h = seed ^ (len * m);
#if defined(__arm) || defined(__arm__)
const size_t ksize = sizeof(uint64_t);
const unsigned char * data = (const unsigned char *)key;
const unsigned char * end = data + (std::size_t)(len/8) * ksize;
#else
const uint64_t * data = (const uint64_t *)key;
const uint64_t * end = data + (len/8);
#endif
while(data != end)
{
#if defined(__arm) || defined(__arm__)
uint64_t k;
memcpy(&k, data, ksize);
data += ksize;
#else
uint64_t k = *data++;
#endif
k *= m;
k ^= k >> r;
k *= m;
h ^= k;
h *= m;
}
const unsigned char * data2 = (const unsigned char*)data;
switch(len & 7)
{
case 7: h ^= uint64_t(data2[6]) << 48;
case 6: h ^= uint64_t(data2[5]) << 40;
case 5: h ^= uint64_t(data2[4]) << 32;
case 4: h ^= uint64_t(data2[3]) << 24;
case 3: h ^= uint64_t(data2[2]) << 16;
case 2: h ^= uint64_t(data2[1]) << 8;
case 1: h ^= uint64_t(data2[0]);
h *= m;
};
h ^= h >> r;
h *= m;
h ^= h >> r;
return h;
}
// 64-bit hash for 32-bit platforms
uint64_t MurmurHash64B ( const void * key, std::size_t len, uint64_t seed )
{
const unsigned int m = 0x5bd1e995;
const int r = 24;
unsigned int h1 = seed ^ len;
unsigned int h2 = 0;
#if defined(__arm) || defined(__arm__)
size_t ksize = sizeof(unsigned int);
const unsigned char * data = (const unsigned char *)key;
#else
const unsigned int * data = (const unsigned int *)key;
#endif
unsigned int k1, k2;
while(len >= 8)
{
#if defined(__arm) || defined(__arm__)
memcpy(&k1, data, ksize);
data += ksize;
memcpy(&k2, data, ksize);
data += ksize;
#else
k1 = *data++;
k2 = *data++;
#endif
k1 *= m; k1 ^= k1 >> r; k1 *= m;
h1 *= m; h1 ^= k1;
len -= 4;
k2 *= m; k2 ^= k2 >> r; k2 *= m;
h2 *= m; h2 ^= k2;
len -= 4;
}
if(len >= 4)
{
#if defined(__arm) || defined(__arm__)
memcpy(&k1, data, ksize);
data += ksize;
#else
k1 = *data++;
#endif
k1 *= m; k1 ^= k1 >> r; k1 *= m;
h1 *= m; h1 ^= k1;
len -= 4;
}
switch(len)
{
case 3: h2 ^= ((unsigned char*)data)[2] << 16;
case 2: h2 ^= ((unsigned char*)data)[1] << 8;
case 1: h2 ^= ((unsigned char*)data)[0];
h2 *= m;
};
h1 ^= h2 >> 18; h1 *= m;
h2 ^= h1 >> 22; h2 *= m;
h1 ^= h2 >> 17; h1 *= m;
h2 ^= h1 >> 19; h2 *= m;
uint64_t h = h1;
h = (h << 32) | h2;
return h;
}
// Trick to test for 64-bit architecture at compile time.
namespace {
template <unsigned L> inline uint64_t MurmurHashNativeBackend(const void * key, std::size_t len, uint64_t seed) {
return MurmurHash64A(key, len, seed);
}
template <> inline uint64_t MurmurHashNativeBackend<4>(const void * key, std::size_t len, uint64_t seed) {
return MurmurHash64B(key, len, seed);
}
} // namespace
uint64_t MurmurHashNative(const void * key, std::size_t len, uint64_t seed) {
return MurmurHashNativeBackend<sizeof(void*)>(key, len, seed);
}
} // namespace util

14
util/murmur_hash.hh
View File

@ -1,14 +0,0 @@
#ifndef UTIL_MURMUR_HASH__
#define UTIL_MURMUR_HASH__
#include <cstddef>
#include <stdint.h>
namespace util {
uint64_t MurmurHash64A(const void * key, std::size_t len, uint64_t seed = 0);
uint64_t MurmurHash64B(const void * key, std::size_t len, uint64_t seed = 0);
uint64_t MurmurHashNative(const void * key, std::size_t len, uint64_t seed = 0);
} // namespace util
#endif // UTIL_MURMUR_HASH__

143
util/probing_hash_table.hh
View File

@ -1,143 +0,0 @@
#ifndef UTIL_PROBING_HASH_TABLE__
#define UTIL_PROBING_HASH_TABLE__
#include "util/exception.hh"
#include <algorithm>
#include <cstddef>
#include <functional>
#include <assert.h>
namespace util {
/* Thrown when table grows too large */
class ProbingSizeException : public Exception {
public:
ProbingSizeException() throw() {}
~ProbingSizeException() throw() {}
};
// std::identity is an SGI extension :-(
struct IdentityHash {
template <class T> T operator()(T arg) const { return arg; }
};
/* Non-standard hash table
* Buckets must be set at the beginning and must be greater than maximum number
* of elements, else it throws ProbingSizeException.
* Memory management and initialization is externalized to make it easier to
* serialize these to disk and load them quickly.
* Uses linear probing to find value.
* Only insert and lookup operations.
*/
template <class EntryT, class HashT, class EqualT = std::equal_to<typename EntryT::Key> > class ProbingHashTable {
public:
typedef EntryT Entry;
typedef typename Entry::Key Key;
typedef const Entry *ConstIterator;
typedef Entry *MutableIterator;
typedef HashT Hash;
typedef EqualT Equal;
public:
static std::size_t Size(std::size_t entries, float multiplier) {
std::size_t buckets = std::max(entries + 1, static_cast<std::size_t>(multiplier * static_cast<float>(entries)));
return buckets * sizeof(Entry);
}
// Must be assigned to later.
ProbingHashTable() : entries_(0)
#ifdef DEBUG
, initialized_(false)
#endif
{}
ProbingHashTable(void *start, std::size_t allocated, const Key &invalid = Key(), const Hash &hash_func = Hash(), const Equal &equal_func = Equal())
: begin_(reinterpret_cast<MutableIterator>(start)),
buckets_(allocated / sizeof(Entry)),
end_(begin_ + buckets_),
invalid_(invalid),
hash_(hash_func),
equal_(equal_func),
entries_(0)
#ifdef DEBUG
, initialized_(true)
#endif
{}
template <class T> MutableIterator Insert(const T &t) {
#ifdef DEBUG
assert(initialized_);
#endif
UTIL_THROW_IF(++entries_ >= buckets_, ProbingSizeException, "Hash table with " << buckets_ << " buckets is full.");
for (MutableIterator i(begin_ + (hash_(t.GetKey()) % buckets_));;) {
if (equal_(i->GetKey(), invalid_)) { *i = t; return i; }
if (++i == end_) { i = begin_; }
}
}
// Return true if the value was found (and not inserted). This is consistent with Find but the opposite if hash_map!
template <class T> bool FindOrInsert(const T &t, MutableIterator &out) {
#ifdef DEBUG
assert(initialized_);
#endif
for (MutableIterator i(begin_ + (hash_(t.GetKey()) % buckets_));;) {
Key got(i->GetKey());
if (equal_(got, t.GetKey())) { out = i; return true; }
if (equal_(got, invalid_)) {
UTIL_THROW_IF(++entries_ >= buckets_, ProbingSizeException, "Hash table with " << buckets_ << " buckets is full.");
*i = t;
out = i;
return false;
}
if (++i == end_) i = begin_;
}
}
void FinishedInserting() {}
void LoadedBinary() {}
// Don't change anything related to GetKey,
template <class Key> bool UnsafeMutableFind(const Key key, MutableIterator &out) {
#ifdef DEBUG
assert(initialized_);
#endif
for (MutableIterator i(begin_ + (hash_(key) % buckets_));;) {
Key got(i->GetKey());
if (equal_(got, key)) { out = i; return true; }
if (equal_(got, invalid_)) return false;
if (++i == end_) i = begin_;
}
}
template <class Key> bool Find(const Key key, ConstIterator &out) const {
#ifdef DEBUG
assert(initialized_);
#endif
for (ConstIterator i(begin_ + (hash_(key) % buckets_));;) {
Key got(i->GetKey());
if (equal_(got, key)) { out = i; return true; }
if (equal_(got, invalid_)) return false;
if (++i == end_) i = begin_;
}
}
private:
MutableIterator begin_;
std::size_t buckets_;
MutableIterator end_;
Key invalid_;
Hash hash_;
Equal equal_;
std::size_t entries_;
#ifdef DEBUG
bool initialized_;
#endif
};
} // namespace util
#endif // UTIL_PROBING_HASH_TABLE__

50
util/probing_hash_table_test.cc
View File

@ -1,50 +0,0 @@
#include "util/probing_hash_table.hh"
#define BOOST_TEST_MODULE ProbingHashTableTest
#include <boost/test/unit_test.hpp>
#include <boost/scoped_array.hpp>
#include <boost/functional/hash.hpp>
#include <stdio.h>
#include <string.h>
#include <stdint.h>
namespace util {
namespace {
struct Entry {
unsigned char key;
typedef unsigned char Key;
unsigned char GetKey() const {
return key;
}
uint64_t GetValue() const {
return value;
}
uint64_t value;
};
typedef ProbingHashTable<Entry, boost::hash<unsigned char> > Table;
BOOST_AUTO_TEST_CASE(simple) {
size_t size = Table::Size(10, 1.2);
boost::scoped_array<char> mem(new char[size]);
memset(mem.get(), 0, size);
Table table(mem.get(), size);
const Entry *i = NULL;
BOOST_CHECK(!table.Find(2, i));
Entry to_ins;
to_ins.key = 3;
to_ins.value = 328920;
table.Insert(to_ins);
BOOST_REQUIRE(table.Find(3, i));
BOOST_CHECK_EQUAL(3, i->GetKey());
BOOST_CHECK_EQUAL(static_cast<uint64_t>(328920), i->GetValue());
BOOST_CHECK(!table.Find(2, i));
}
} // namespace
} // namespace util

96
util/proxy_iterator.hh
View File

@ -1,96 +0,0 @@
#ifndef UTIL_PROXY_ITERATOR__
#define UTIL_PROXY_ITERATOR__
#include <cstddef>
#include <iterator>
/* This is a RandomAccessIterator that uses a proxy to access the underlying
* data. Useful for packing data at bit offsets but still using STL
* algorithms.
*
* Normally I would use boost::iterator_facade but some people are too lazy to
* install boost and still want to use my language model. It's amazing how
* many operators an iterator has.
*
* The Proxy needs to provide:
* class InnerIterator;
* InnerIterator &Inner();
* const InnerIterator &Inner() const;
*
* InnerIterator has to implement:
* operator==(InnerIterator)
* operator<(InnerIterator)
* operator+=(std::ptrdiff_t)
* operator-(InnerIterator)
* and of course whatever Proxy needs to dereference it.
*
* It's also a good idea to specialize std::swap for Proxy.
*/
namespace util {
template <class Proxy> class ProxyIterator {
private:
// Self.
typedef ProxyIterator<Proxy> S;
typedef typename Proxy::InnerIterator InnerIterator;
public:
typedef std::random_access_iterator_tag iterator_category;
typedef typename Proxy::value_type value_type;
typedef std::ptrdiff_t difference_type;
typedef Proxy reference;
typedef Proxy * pointer;
ProxyIterator() {}
// For cast from non const to const.
template <class AlternateProxy> ProxyIterator(const ProxyIterator<AlternateProxy> &in) : p_(*in) {}
explicit ProxyIterator(const Proxy &p) : p_(p) {}
// p_'s operator= does value copying, but here we want iterator copying.
S &operator=(const S &other) {
I() = other.I();
return *this;
}
bool operator==(const S &other) const { return I() == other.I(); }
bool operator!=(const S &other) const { return !(*this == other); }
bool operator<(const S &other) const { return I() < other.I(); }
bool operator>(const S &other) const { return other < *this; }
bool operator<=(const S &other) const { return !(*this > other); }
bool operator>=(const S &other) const { return !(*this < other); }
S &operator++() { return *this += 1; }
S operator++(int) { S ret(*this); ++*this; return ret; }
S &operator+=(std::ptrdiff_t amount) { I() += amount; return *this; }
S operator+(std::ptrdiff_t amount) const { S ret(*this); ret += amount; return ret; }
S &operator--() { return *this -= 1; }
S operator--(int) { S ret(*this); --*this; return ret; }
S &operator-=(std::ptrdiff_t amount) { I() += (-amount); return *this; }
S operator-(std::ptrdiff_t amount) const { S ret(*this); ret -= amount; return ret; }
std::ptrdiff_t operator-(const S &other) const { return I() - other.I(); }
Proxy operator*() { return p_; }
const Proxy operator*() const { return p_; }
Proxy *operator->() { return &p_; }
const Proxy *operator->() const { return &p_; }
Proxy operator[](std::ptrdiff_t amount) const { return *(*this + amount); }
const InnerIterator &Inner() { return p_.Inner(); }
private:
InnerIterator &I() { return p_.Inner(); }
const InnerIterator &I() const { return p_.Inner(); }
Proxy p_;
};
template <class Proxy> ProxyIterator<Proxy> operator+(std::ptrdiff_t amount, const ProxyIterator<Proxy> &it) {
return it + amount;
}
} // namespace util
#endif // UTIL_PROXY_ITERATOR__

97
util/scoped.hh
View File

@ -1,97 +0,0 @@
#ifndef UTIL_SCOPED__
#define UTIL_SCOPED__
#include "util/exception.hh"
/* Other scoped objects in the style of scoped_ptr. */
#include <cstddef>
#include <cstdlib>
namespace util {
template <class T, class R, R (*Free)(T*)> class scoped_thing {
public:
explicit scoped_thing(T *c = static_cast<T*>(0)) : c_(c) {}
~scoped_thing() { if (c_) Free(c_); }
void reset(T *c) {
if (c_) Free(c_);
c_ = c;
}
T &operator*() { return *c_; }
const T&operator*() const { return *c_; }
T &operator->() { return *c_; }
const T&operator->() const { return *c_; }
T *get() { return c_; }
const T *get() const { return c_; }
private:
T *c_;
scoped_thing(const scoped_thing &);
scoped_thing &operator=(const scoped_thing &);
};
class scoped_malloc {
public:
scoped_malloc() : p_(NULL) {}
scoped_malloc(void *p) : p_(p) {}
~scoped_malloc() { std::free(p_); }
void reset(void *p = NULL) {
scoped_malloc other(p_);
p_ = p;
}
void call_realloc(std::size_t to) {
void *ret;
UTIL_THROW_IF(!(ret = std::realloc(p_, to)) && to, util::ErrnoException, "realloc to " << to << " bytes failed.");
p_ = ret;
}
void *get() { return p_; }
const void *get() const { return p_; }
private:
void *p_;
scoped_malloc(const scoped_malloc &);
scoped_malloc &operator=(const scoped_malloc &);
};
// Hat tip to boost.
template <class T> class scoped_array {
public:
explicit scoped_array(T *content = NULL) : c_(content) {}
~scoped_array() { delete [] c_; }
T *get() { return c_; }
const T* get() const { return c_; }
T &operator*() { return *c_; }
const T&operator*() const { return *c_; }
T &operator->() { return *c_; }
const T&operator->() const { return *c_; }
T &operator[](std::size_t idx) { return c_[idx]; }
const T &operator[](std::size_t idx) const { return c_[idx]; }
void reset(T *to = NULL) {
scoped_array<T> other(c_);
c_ = to;
}
private:
T *c_;
};
} // namespace util
#endif // UTIL_SCOPED__

107
util/sized_iterator.hh
View File

@ -1,107 +0,0 @@
#ifndef UTIL_SIZED_ITERATOR__
#define UTIL_SIZED_ITERATOR__
#include "util/proxy_iterator.hh"
#include <functional>
#include <string>
#include <stdint.h>
#include <string.h>
namespace util {
class SizedInnerIterator {
public:
SizedInnerIterator() {}
SizedInnerIterator(void *ptr, std::size_t size) : ptr_(static_cast<uint8_t*>(ptr)), size_(size) {}
bool operator==(const SizedInnerIterator &other) const {
return ptr_ == other.ptr_;
}
bool operator<(const SizedInnerIterator &other) const {
return ptr_ < other.ptr_;
}
SizedInnerIterator &operator+=(std::ptrdiff_t amount) {
ptr_ += amount * size_;
return *this;
}
std::ptrdiff_t operator-(const SizedInnerIterator &other) const {
return (ptr_ - other.ptr_) / size_;
}
const void *Data() const { return ptr_; }
void *Data() { return ptr_; }
std::size_t EntrySize() const { return size_; }
private:
uint8_t *ptr_;
std::size_t size_;
};
class SizedProxy {
public:
SizedProxy() {}
SizedProxy(void *ptr, std::size_t size) : inner_(ptr, size) {}
operator std::string() const {
return std::string(reinterpret_cast<const char*>(inner_.Data()), inner_.EntrySize());
}
SizedProxy &operator=(const SizedProxy &from) {
memcpy(inner_.Data(), from.inner_.Data(), inner_.EntrySize());
return *this;
}
SizedProxy &operator=(const std::string &from) {
memcpy(inner_.Data(), from.data(), inner_.EntrySize());
return *this;
}
const void *Data() const { return inner_.Data(); }
void *Data() { return inner_.Data(); }
private:
friend class util::ProxyIterator<SizedProxy>;
typedef std::string value_type;
typedef SizedInnerIterator InnerIterator;
InnerIterator &Inner() { return inner_; }
const InnerIterator &Inner() const { return inner_; }
InnerIterator inner_;
};
typedef ProxyIterator<SizedProxy> SizedIterator;
inline SizedIterator SizedIt(void *ptr, std::size_t size) { return SizedIterator(SizedProxy(ptr, size)); }
// Useful wrapper for a comparison function i.e. sort.
template <class Delegate, class Proxy = SizedProxy> class SizedCompare : public std::binary_function<const Proxy &, const Proxy &, bool> {
public:
explicit SizedCompare(const Delegate &delegate = Delegate()) : delegate_(delegate) {}
bool operator()(const Proxy &first, const Proxy &second) const {
return delegate_(first.Data(), second.Data());
}
bool operator()(const Proxy &first, const std::string &second) const {
return delegate_(first.Data(), second.data());
}
bool operator()(const std::string &first, const Proxy &second) const {
return delegate_(first.data(), second.Data());
}
bool operator()(const std::string &first, const std::string &second) const {
return delegate_(first.data(), second.data());
}
const Delegate &GetDelegate() const { return delegate_; }
private:
const Delegate delegate_;
};
} // namespace util
#endif // UTIL_SIZED_ITERATOR__

127
util/sorted_uniform.hh
View File

@ -1,127 +0,0 @@
#ifndef UTIL_SORTED_UNIFORM__
#define UTIL_SORTED_UNIFORM__
#include <algorithm>
#include <cstddef>
#include <assert.h>
#include <stdint.h>
namespace util {
template <class T> class IdentityAccessor {
public:
typedef T Key;
T operator()(const T *in) const { return *in; }
};
struct Pivot64 {
static inline std::size_t Calc(uint64_t off, uint64_t range, std::size_t width) {
std::size_t ret = static_cast<std::size_t>(static_cast<float>(off) / static_cast<float>(range) * static_cast<float>(width));
// Cap for floating point rounding
return (ret < width) ? ret : width - 1;
}
};
// Use when off * width is <2^64. This is guaranteed when each of them is actually a 32-bit value.
struct Pivot32 {
static inline std::size_t Calc(uint64_t off, uint64_t range, uint64_t width) {
return static_cast<std::size_t>((off * width) / (range + 1));
}
};
// Usage: PivotSelect<sizeof(DataType)>::T
template <unsigned> struct PivotSelect;
template <> struct PivotSelect<8> { typedef Pivot64 T; };
template <> struct PivotSelect<4> { typedef Pivot32 T; };
template <> struct PivotSelect<2> { typedef Pivot32 T; };
/* Binary search. */
template <class Iterator, class Accessor> bool BinaryFind(
const Accessor &accessor,
Iterator begin,
Iterator end,
const typename Accessor::Key key, Iterator &out) {
while (end > begin) {
Iterator pivot(begin + (end - begin) / 2);
typename Accessor::Key mid(accessor(pivot));
if (mid < key) {
begin = pivot + 1;
} else if (mid > key) {
end = pivot;
} else {
out = pivot;
return true;
}
}
return false;
}
// Search the range [before_it + 1, after_it - 1] for key.
// Preconditions:
// before_v <= key <= after_v
// before_v <= all values in the range [before_it + 1, after_it - 1] <= after_v
// range is sorted.
template <class Iterator, class Accessor, class Pivot> bool BoundedSortedUniformFind(
const Accessor &accessor,
Iterator before_it, typename Accessor::Key before_v,
Iterator after_it, typename Accessor::Key after_v,
const typename Accessor::Key key, Iterator &out) {
while (after_it - before_it > 1) {
Iterator pivot(before_it + (1 + Pivot::Calc(key - before_v, after_v - before_v, after_it - before_it - 1)));
typename Accessor::Key mid(accessor(pivot));
if (mid < key) {
before_it = pivot;
before_v = mid;
} else if (mid > key) {
after_it = pivot;
after_v = mid;
} else {
out = pivot;
return true;
}
}
return false;
}
template <class Iterator, class Accessor, class Pivot> bool SortedUniformFind(const Accessor &accessor, Iterator begin, Iterator end, const typename Accessor::Key key, Iterator &out) {
if (begin == end) return false;
typename Accessor::Key below(accessor(begin));
if (key <= below) {
if (key == below) { out = begin; return true; }
return false;
}
// Make the range [begin, end].
--end;
typename Accessor::Key above(accessor(end));
if (key >= above) {
if (key == above) { out = end; return true; }
return false;
}
return BoundedSortedUniformFind<Iterator, Accessor, Pivot>(accessor, begin, below, end, above, key, out);
}
// May return begin - 1.
template <class Iterator, class Accessor> Iterator BinaryBelow(
const Accessor &accessor,
Iterator begin,
Iterator end,
const typename Accessor::Key key) {
while (end > begin) {
Iterator pivot(begin + (end - begin) / 2);
typename Accessor::Key mid(accessor(pivot));
if (mid < key) {
begin = pivot + 1;
} else if (mid > key) {
end = pivot;
} else {
for (++pivot; (pivot < end) && accessor(pivot) == mid; ++pivot) {}
return pivot - 1;
}
}
return begin - 1;
}
} // namespace util
#endif // UTIL_SORTED_UNIFORM__

127
util/sorted_uniform_test.cc
View File

@ -1,127 +0,0 @@
#include "util/sorted_uniform.hh"
#include <boost/random/mersenne_twister.hpp>
#include <boost/random/uniform_int.hpp>
#include <boost/random/variate_generator.hpp>
#include <boost/scoped_array.hpp>
#include <boost/unordered_map.hpp>
#define BOOST_TEST_MODULE SortedUniformTest
#include <boost/test/unit_test.hpp>
#include <algorithm>
#include <limits>
#include <vector>
namespace util {
namespace {
template <class KeyT, class ValueT> struct Entry {
typedef KeyT Key;
typedef ValueT Value;
Key key;
Value value;
Key GetKey() const {
return key;
}
Value GetValue() const {
return value;
}
bool operator<(const Entry<Key,Value> &other) const {
return key < other.key;
}
};
template <class KeyT> struct Accessor {
typedef KeyT Key;
template <class Value> Key operator()(const Entry<Key, Value> *entry) const {
return entry->GetKey();
}
};
template <class Key, class Value> void Check(const Entry<Key, Value> *begin, const Entry<Key, Value> *end, const boost::unordered_map<Key, Value> &reference, const Key key) {
typename boost::unordered_map<Key, Value>::const_iterator ref = reference.find(key);
typedef const Entry<Key, Value> *It;
// g++ can't tell that require will crash and burn.
It i = NULL;
bool ret = SortedUniformFind<It, Accessor<Key>, Pivot64>(Accessor<Key>(), begin, end, key, i);
if (ref == reference.end()) {
BOOST_CHECK(!ret);
} else {
BOOST_REQUIRE(ret);
BOOST_CHECK_EQUAL(ref->second, i->GetValue());
}
}
BOOST_AUTO_TEST_CASE(empty) {
typedef const Entry<uint64_t, float> T;
const T *i;
bool ret = SortedUniformFind<const T*, Accessor<uint64_t>, Pivot64>(Accessor<uint64_t>(), (const T*)NULL, (const T*)NULL, (uint64_t)10, i);
BOOST_CHECK(!ret);
}
template <class Key> void RandomTest(Key upper, size_t entries, size_t queries) {
typedef unsigned char Value;
boost::mt19937 rng;
boost::uniform_int<Key> range_key(0, upper);
boost::uniform_int<Value> range_value(0, 255);
boost::variate_generator<boost::mt19937&, boost::uniform_int<Key> > gen_key(rng, range_key);
boost::variate_generator<boost::mt19937&, boost::uniform_int<unsigned char> > gen_value(rng, range_value);
typedef Entry<Key, Value> Ent;
std::vector<Ent> backing;
boost::unordered_map<Key, unsigned char> reference;
Ent ent;
for (size_t i = 0; i < entries; ++i) {
Key key = gen_key();
unsigned char value = gen_value();
if (reference.insert(std::make_pair(key, value)).second) {
ent.key = key;
ent.value = value;
backing.push_back(ent);
}
}
std::sort(backing.begin(), backing.end());
// Random queries.
for (size_t i = 0; i < queries; ++i) {
const Key key = gen_key();
Check<Key, unsigned char>(&*backing.begin(), &*backing.end(), reference, key);
}
typename boost::unordered_map<Key, unsigned char>::const_iterator it = reference.begin();
for (size_t i = 0; (i < queries) && (it != reference.end()); ++i, ++it) {
Check<Key, unsigned char>(&*backing.begin(), &*backing.end(), reference, it->second);
}
}
BOOST_AUTO_TEST_CASE(basic) {
RandomTest<uint8_t>(11, 10, 200);
}
BOOST_AUTO_TEST_CASE(tiny_dense_random) {
RandomTest<uint8_t>(11, 50, 200);
}
BOOST_AUTO_TEST_CASE(small_dense_random) {
RandomTest<uint8_t>(100, 100, 200);
}
BOOST_AUTO_TEST_CASE(small_sparse_random) {
RandomTest<uint8_t>(200, 15, 200);
}
BOOST_AUTO_TEST_CASE(medium_sparse_random) {
RandomTest<uint16_t>(32000, 1000, 2000);
}
BOOST_AUTO_TEST_CASE(sparse_random) {
RandomTest<uint64_t>(std::numeric_limits<uint64_t>::max(), 100000, 2000);
}
} // namespace
} // namespace util

288
util/string_piece.hh
View File

@ -1,288 +0,0 @@
/* If you use ICU in your program, then compile with -DHAVE_ICU -licui18n. If
* you don't use ICU, then this will use the Google implementation from Chrome.
* This has been modified from the original version to let you choose.
*/
// Copyright 2008, Google Inc.
// All rights reserved.
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are
// met:
//
// * Redistributions of source code must retain the above copyright
// notice, this list of conditions and the following disclaimer.
// * Redistributions in binary form must reproduce the above
// copyright notice, this list of conditions and the following disclaimer
// in the documentation and/or other materials provided with the
// distribution.
// * Neither the name of Google Inc. nor the names of its
// contributors may be used to endorse or promote products derived from
// this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
// Copied from strings/stringpiece.h with modifications
//
// A string-like object that points to a sized piece of memory.
//
// Functions or methods may use const StringPiece& parameters to accept either
// a "const char*" or a "string" value that will be implicitly converted to
// a StringPiece. The implicit conversion means that it is often appropriate
// to include this .h file in other files rather than forward-declaring
// StringPiece as would be appropriate for most other Google classes.
//
// Systematic usage of StringPiece is encouraged as it will reduce unnecessary
// conversions from "const char*" to "string" and back again.
//
#ifndef BASE_STRING_PIECE_H__
#define BASE_STRING_PIECE_H__
#include "util/have.hh"
#ifdef HAVE_BOOST
#include <boost/functional/hash/hash.hpp>
#endif // HAVE_BOOST
#include <cstring>
#include <iosfwd>
#include <ostream>
#ifdef HAVE_ICU
#include <unicode/stringpiece.h>
#include <unicode/uversion.h>
// Old versions of ICU don't define operator== and operator!=.
#if (U_ICU_VERSION_MAJOR_NUM < 4) || ((U_ICU_VERSION_MAJOR_NUM == 4) && (U_ICU_VERSION_MINOR_NUM < 4))
#warning You are using an old version of ICU. Consider upgrading to ICU >= 4.6.
inline bool operator==(const StringPiece& x, const StringPiece& y) {
if (x.size() != y.size())
return false;
return std::memcmp(x.data(), y.data(), x.size()) == 0;
}
inline bool operator!=(const StringPiece& x, const StringPiece& y) {
return !(x == y);
}
#endif // old version of ICU
U_NAMESPACE_BEGIN
#else
#include <algorithm>
#include <cstddef>
#include <string>
#include <string.h>
#ifdef WIN32
#undef max
#undef min
#endif
class StringPiece {
public:
typedef size_t size_type;
private:
const char* ptr_;
size_type length_;
public:
// We provide non-explicit singleton constructors so users can pass
// in a "const char*" or a "string" wherever a "StringPiece" is
// expected.
StringPiece() : ptr_(NULL), length_(0) { }
StringPiece(const char* str)
: ptr_(str), length_((str == NULL) ? 0 : strlen(str)) { }
StringPiece(const std::string& str)
: ptr_(str.data()), length_(str.size()) { }
StringPiece(const char* offset, size_type len)
: ptr_(offset), length_(len) { }
// data() may return a pointer to a buffer with embedded NULs, and the
// returned buffer may or may not be null terminated. Therefore it is
// typically a mistake to pass data() to a routine that expects a NUL
// terminated string.
const char* data() const { return ptr_; }
size_type size() const { return length_; }
size_type length() const { return length_; }
bool empty() const { return length_ == 0; }
void clear() { ptr_ = NULL; length_ = 0; }
void set(const char* data, size_type len) { ptr_ = data; length_ = len; }
void set(const char* str) {
ptr_ = str;
length_ = str ? strlen(str) : 0;
}
void set(const void* data, size_type len) {
ptr_ = reinterpret_cast<const char*>(data);
length_ = len;
}
char operator[](size_type i) const { return ptr_[i]; }
void remove_prefix(size_type n) {
ptr_ += n;
length_ -= n;
}
void remove_suffix(size_type n) {
length_ -= n;
}
int compare(const StringPiece& x) const {
int r = wordmemcmp(ptr_, x.ptr_, std::min(length_, x.length_));
if (r == 0) {
if (length_ < x.length_) r = -1;
else if (length_ > x.length_) r = +1;
}
return r;
}
std::string as_string() const {
// std::string doesn't like to take a NULL pointer even with a 0 size.
return std::string(!empty() ? data() : "", size());
}
void CopyToString(std::string* target) const;
void AppendToString(std::string* target) const;
// Does "this" start with "x"
bool starts_with(const StringPiece& x) const {
return ((length_ >= x.length_) &&
(wordmemcmp(ptr_, x.ptr_, x.length_) == 0));
}
// Does "this" end with "x"
bool ends_with(const StringPiece& x) const {
return ((length_ >= x.length_) &&
(wordmemcmp(ptr_ + (length_-x.length_), x.ptr_, x.length_) == 0));
}
// standard STL container boilerplate
typedef char value_type;
typedef const char* pointer;
typedef const char& reference;
typedef const char& const_reference;
typedef ptrdiff_t difference_type;
static const size_type npos;
typedef const char* const_iterator;
typedef const char* iterator;
typedef std::reverse_iterator<const_iterator> const_reverse_iterator;
typedef std::reverse_iterator<iterator> reverse_iterator;
iterator begin() const { return ptr_; }
iterator end() const { return ptr_ + length_; }
const_reverse_iterator rbegin() const {
return const_reverse_iterator(ptr_ + length_);
}
const_reverse_iterator rend() const {
return const_reverse_iterator(ptr_);
}
size_type max_size() const { return length_; }
size_type capacity() const { return length_; }
size_type copy(char* buf, size_type n, size_type pos = 0) const;
size_type find(const StringPiece& s, size_type pos = 0) const;
size_type find(char c, size_type pos = 0) const;
size_type rfind(const StringPiece& s, size_type pos = npos) const;
size_type rfind(char c, size_type pos = npos) const;
size_type find_first_of(const StringPiece& s, size_type pos = 0) const;
size_type find_first_of(char c, size_type pos = 0) const {
return find(c, pos);
}
size_type find_first_not_of(const StringPiece& s, size_type pos = 0) const;
size_type find_first_not_of(char c, size_type pos = 0) const;
size_type find_last_of(const StringPiece& s, size_type pos = npos) const;
size_type find_last_of(char c, size_type pos = npos) const {
return rfind(c, pos);
}
size_type find_last_not_of(const StringPiece& s, size_type pos = npos) const;
size_type find_last_not_of(char c, size_type pos = npos) const;
StringPiece substr(size_type pos, size_type n = npos) const;
static int wordmemcmp(const char* p, const char* p2, size_type N) {
return memcmp(p, p2, N);
}
};
inline bool operator==(const StringPiece& x, const StringPiece& y) {
if (x.size() != y.size())
return false;
return std::memcmp(x.data(), y.data(), x.size()) == 0;
}
inline bool operator!=(const StringPiece& x, const StringPiece& y) {
return !(x == y);
}
#endif // HAVE_ICU undefined
inline bool operator<(const StringPiece& x, const StringPiece& y) {
const int r = std::memcmp(x.data(), y.data(),
std::min(x.size(), y.size()));
return ((r < 0) || ((r == 0) && (x.size() < y.size())));
}
inline bool operator>(const StringPiece& x, const StringPiece& y) {
return y < x;
}
inline bool operator<=(const StringPiece& x, const StringPiece& y) {
return !(x > y);
}
inline bool operator>=(const StringPiece& x, const StringPiece& y) {
return !(x < y);
}
// allow StringPiece to be logged (needed for unit testing).
inline std::ostream& operator<<(std::ostream& o, const StringPiece& piece) {
return o.write(piece.data(), static_cast<std::streamsize>(piece.size()));
}
#ifdef HAVE_BOOST
inline size_t hash_value(const StringPiece &str) {
return boost::hash_range(str.data(), str.data() + str.length());
}
/* Support for lookup of StringPiece in boost::unordered_map<std::string> */
struct StringPieceCompatibleHash : public std::unary_function<const StringPiece &, size_t> {
size_t operator()(const StringPiece &str) const {
return hash_value(str);
}
};
struct StringPieceCompatibleEquals : public std::binary_function<const StringPiece &, const std::string &, bool> {
bool operator()(const StringPiece &first, const StringPiece &second) const {
return first == second;
}
};
template <class T> typename T::const_iterator FindStringPiece(const T &t, const StringPiece &key) {
return t.find(key, StringPieceCompatibleHash(), StringPieceCompatibleEquals());
}
template <class T> typename T::iterator FindStringPiece(T &t, const StringPiece &key) {
return t.find(key, StringPieceCompatibleHash(), StringPieceCompatibleEquals());
}
#endif
#ifdef HAVE_ICU
U_NAMESPACE_END
#endif
#endif // BASE_STRING_PIECE_H__

108
util/tokenize_piece.hh
View File

@ -1,108 +0,0 @@
#ifndef UTIL_TOKENIZE_PIECE__
#define UTIL_TOKENIZE_PIECE__
#include "util/exception.hh"
#include "util/string_piece.hh"
#include <boost/iterator/iterator_facade.hpp>
#include <algorithm>
#include <iostream>
namespace util {
// Thrown on dereference when out of tokens to parse
class OutOfTokens : public Exception {
public:
OutOfTokens() throw() {}
~OutOfTokens() throw() {}
};
class SingleCharacter {
public:
explicit SingleCharacter(char delim) : delim_(delim) {}
StringPiece Find(const StringPiece &in) const {
return StringPiece(std::find(in.data(), in.data() + in.size(), delim_), 1);
}
private:
char delim_;
};
class MultiCharacter {
public:
explicit MultiCharacter(const StringPiece &delimiter) : delimiter_(delimiter) {}
StringPiece Find(const StringPiece &in) const {
return StringPiece(std::search(in.data(), in.data() + in.size(), delimiter_.data(), delimiter_.data() + delimiter_.size()), delimiter_.size());
}
private:
StringPiece delimiter_;
};
class AnyCharacter {
public:
explicit AnyCharacter(const StringPiece &chars) : chars_(chars) {}
StringPiece Find(const StringPiece &in) const {
return StringPiece(std::find_first_of(in.data(), in.data() + in.size(), chars_.data(), chars_.data() + chars_.size()), 1);
}
private:
StringPiece chars_;
};
template <class Find, bool SkipEmpty = false> class TokenIter : public boost::iterator_facade<TokenIter<Find, SkipEmpty>, const StringPiece, boost::forward_traversal_tag> {
public:
TokenIter() {}
template <class Construct> TokenIter(const StringPiece &str, const Construct &construct) : after_(str), finder_(construct) {
increment();
}
bool operator!() const {
return current_.data() == 0;
}
operator bool() const {
return current_.data() != 0;
}
static TokenIter<Find> end() {
return TokenIter<Find>();
}
private:
friend class boost::iterator_core_access;
void increment() {
do {
StringPiece found(finder_.Find(after_));
current_ = StringPiece(after_.data(), found.data() - after_.data());
if (found.data() == after_.data() + after_.size()) {
after_ = StringPiece(NULL, 0);
} else {
after_ = StringPiece(found.data() + found.size(), after_.data() - found.data() + after_.size() - found.size());
}
} while (SkipEmpty && current_.data() && current_.empty()); // Compiler should optimize this away if SkipEmpty is false.
}
bool equal(const TokenIter<Find> &other) const {
return after_.data() == other.after_.data();
}
const StringPiece &dereference() const {
UTIL_THROW_IF(!current_.data(), OutOfTokens, "Ran out of tokens");
return current_;
}
StringPiece current_;
StringPiece after_;
Find finder_;
};
} // namespace util
#endif // UTIL_TOKENIZE_PIECE__

48
util/tokenize_piece_test.cc
View File

@ -1,48 +0,0 @@
#include "util/tokenize_piece.hh"
#include "util/string_piece.hh"
#define BOOST_TEST_MODULE TokenIteratorTest
#include <boost/test/unit_test.hpp>
#include <iostream>
namespace util {
namespace {
BOOST_AUTO_TEST_CASE(pipe_pipe_none) {
const char str[] = "nodelimit at all";
TokenIter<MultiCharacter> it(str, MultiCharacter("|||"));
BOOST_REQUIRE(it);
BOOST_CHECK_EQUAL(StringPiece(str), *it);
++it;
BOOST_CHECK(!it);
}
BOOST_AUTO_TEST_CASE(pipe_pipe_two) {
const char str[] = "|||";
TokenIter<MultiCharacter> it(str, MultiCharacter("|||"));
BOOST_REQUIRE(it);
BOOST_CHECK_EQUAL(StringPiece(), *it);
++it;
BOOST_REQUIRE(it);
BOOST_CHECK_EQUAL(StringPiece(), *it);
++it;
BOOST_CHECK(!it);
}
BOOST_AUTO_TEST_CASE(remove_empty) {
const char str[] = "|||";
TokenIter<MultiCharacter, true> it(str, MultiCharacter("|||"));
BOOST_CHECK(!it);
}
BOOST_AUTO_TEST_CASE(remove_empty_keep) {
const char str[] = " |||";
TokenIter<MultiCharacter, true> it(str, MultiCharacter("|||"));
BOOST_REQUIRE(it);
BOOST_CHECK_EQUAL(StringPiece(" "), *it);
++it;
BOOST_CHECK(!it);
}
} // namespace
} // namespace util

46
util/usage.cc
View File

@ -1,46 +0,0 @@
#include "util/usage.hh"
#include <fstream>
#include <ostream>
#include <string.h>
#include <ctype.h>
#if !defined(_WIN32) && !defined(_WIN64)
#include <sys/resource.h>
#include <sys/time.h>
#endif
namespace util {
namespace {
#if !defined(_WIN32) && !defined(_WIN64)
float FloatSec(const struct timeval &tv) {
return static_cast<float>(tv.tv_sec) + (static_cast<float>(tv.tv_usec) / 1000000.0);
}
#endif
} // namespace
void PrintUsage(std::ostream &out) {
#if !defined(_WIN32) && !defined(_WIN64)
struct rusage usage;
if (getrusage(RUSAGE_SELF, &usage)) {
perror("getrusage");
return;
}
out << "user\t" << FloatSec(usage.ru_utime) << "\nsys\t" << FloatSec(usage.ru_stime) << '\n';
// Linux doesn't set memory usage :-(.
std::ifstream status("/proc/self/status", std::ios::in);
std::string line;
while (getline(status, line)) {
if (!strncmp(line.c_str(), "VmRSS:\t", 7)) {
out << "VmRSS: " << (line.c_str() + 7) << '\n';
break;
} else if (!strncmp(line.c_str(), "VmPeak:\t", 8)) {
out << "VmPeak: " << (line.c_str() + 8) << '\n';
}
}
#endif
}
} // namespace util

8
util/usage.hh
View File

@ -1,8 +0,0 @@
#ifndef UTIL_USAGE__
#define UTIL_USAGE__
#include <iosfwd>
namespace util {
void PrintUsage(std::ostream &to);
} // namespace util
#endif // UTIL_USAGE__

Some files were not shown because too many files have changed in this diff Show More