From 65eccee074ed92db07e92f0cdcb30f80861af793 Mon Sep 17 00:00:00 2001 From: mappu Date: Sun, 20 Nov 2016 13:38:15 +1300 Subject: [PATCH] libimagequant (2.8-git-a425e83) GPLv3+ --- COPYRIGHT | 641 +++++++++++++++ blur.c | 132 ++++ blur.h | 4 + libimagequant.c | 1982 +++++++++++++++++++++++++++++++++++++++++++++++ libimagequant.h | 137 ++++ mediancut.c | 478 ++++++++++++ mediancut.h | 2 + mempool.c | 81 ++ mempool.h | 13 + nearest.c | 206 +++++ nearest.h | 8 + pam.c | 276 +++++++ pam.h | 270 +++++++ viter.c | 107 +++ viter.h | 19 + 15 files changed, 4356 insertions(+) create mode 100644 COPYRIGHT create mode 100644 blur.c create mode 100644 blur.h create mode 100644 libimagequant.c create mode 100644 libimagequant.h create mode 100644 mediancut.c create mode 100644 mediancut.h create mode 100644 mempool.c create mode 100644 mempool.h create mode 100644 nearest.c create mode 100644 nearest.h create mode 100644 pam.c create mode 100644 pam.h create mode 100644 viter.c create mode 100644 viter.h diff --git a/COPYRIGHT b/COPYRIGHT new file mode 100644 index 0000000..3514d66 --- /dev/null +++ b/COPYRIGHT @@ -0,0 +1,641 @@ + +libimagequant is derived from code by Jef Poskanzer and Greg Roelofs +licensed under pngquant's original license (at the end of this file), +and contains extensive changes and additions by Kornel Lesiński +licensed under GPL v3. + +- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - + +libimagequant © 2009-2016 by Kornel Lesiński. + + GNU GENERAL PUBLIC LICENSE + Version 3, 29 June 2007 + + Copyright (C) 2007 Free Software Foundation, Inc. + 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. + +- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - + +© 1989, 1991 by Jef Poskanzer. +© 1997, 2000, 2002 by Greg Roelofs. + +Permission to use, copy, modify, and distribute this software and its +documentation for any purpose and without fee is hereby granted, provided +that the above copyright notice appear in all copies and that both that +copyright notice and this permission notice appear in supporting +documentation. This software is provided "as is" without express or +implied warranty. diff --git a/blur.c b/blur.c new file mode 100644 index 0000000..7f0a716 --- /dev/null +++ b/blur.c @@ -0,0 +1,132 @@ +/* +© 2011-2015 by Kornel Lesiński. + +This file is part of libimagequant. + +libimagequant 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. + +libimagequant 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 libimagequant. If not, see . +*/ + +#include "libimagequant.h" +#include "pam.h" +#include "blur.h" + +/* + Blurs image horizontally (width 2*size+1) and writes it transposed to dst (called twice gives 2d blur) + */ +static void transposing_1d_blur(unsigned char *restrict src, unsigned char *restrict dst, unsigned int width, unsigned int height, const unsigned int size) +{ + assert(size > 0); + + for(unsigned int j=0; j < height; j++) { + unsigned char *restrict row = src + j*width; + + // accumulate sum for pixels outside line + unsigned int sum; + sum = row[0]*size; + for(unsigned int i=0; i < size; i++) { + sum += row[i]; + } + + // blur with left side outside line + for(unsigned int i=0; i < size; i++) { + sum -= row[0]; + sum += row[i+size]; + + dst[i*height + j] = sum / (size*2); + } + + for(unsigned int i=size; i < width-size; i++) { + sum -= row[i-size]; + sum += row[i+size]; + + dst[i*height + j] = sum / (size*2); + } + + // blur with right side outside line + for(unsigned int i=width-size; i < width; i++) { + sum -= row[i-size]; + sum += row[width-1]; + + dst[i*height + j] = sum / (size*2); + } + } +} + +/** + * Picks maximum of neighboring pixels (blur + lighten) + */ +LIQ_PRIVATE void liq_max3(unsigned char *src, unsigned char *dst, unsigned int width, unsigned int height) +{ + for(unsigned int j=0; j < height; j++) { + const unsigned char *row = src + j*width, + *prevrow = src + (j > 1 ? j-1 : 0)*width, + *nextrow = src + MIN(height-1,j+1)*width; + + unsigned char prev,curr=row[0],next=row[0]; + + for(unsigned int i=0; i < width-1; i++) { + prev=curr; + curr=next; + next=row[i+1]; + + unsigned char t1 = MAX(prev,next); + unsigned char t2 = MAX(nextrow[i],prevrow[i]); + *dst++ = MAX(curr,MAX(t1,t2)); + } + unsigned char t1 = MAX(curr,next); + unsigned char t2 = MAX(nextrow[width-1],prevrow[width-1]); + *dst++ = MAX(t1,t2); + } +} + +/** + * Picks minimum of neighboring pixels (blur + darken) + */ +LIQ_PRIVATE void liq_min3(unsigned char *src, unsigned char *dst, unsigned int width, unsigned int height) +{ + for(unsigned int j=0; j < height; j++) { + const unsigned char *row = src + j*width, + *prevrow = src + (j > 1 ? j-1 : 0)*width, + *nextrow = src + MIN(height-1,j+1)*width; + + unsigned char prev,curr=row[0],next=row[0]; + + for(unsigned int i=0; i < width-1; i++) { + prev=curr; + curr=next; + next=row[i+1]; + + unsigned char t1 = MIN(prev,next); + unsigned char t2 = MIN(nextrow[i],prevrow[i]); + *dst++ = MIN(curr,MIN(t1,t2)); + } + unsigned char t1 = MIN(curr,next); + unsigned char t2 = MIN(nextrow[width-1],prevrow[width-1]); + *dst++ = MIN(t1,t2); + } +} + +/* + Filters src image and saves it to dst, overwriting tmp in the process. + Image must be width*height pixels high. Size controls radius of box blur. + */ +LIQ_PRIVATE void liq_blur(unsigned char *src, unsigned char *tmp, unsigned char *dst, unsigned int width, unsigned int height, unsigned int size) +{ + assert(size > 0); + if (width < 2*size+1 || height < 2*size+1) { + return; + } + transposing_1d_blur(src, tmp, width, height, size); + transposing_1d_blur(tmp, dst, height, width, size); +} diff --git a/blur.h b/blur.h new file mode 100644 index 0000000..06ae8cb --- /dev/null +++ b/blur.h @@ -0,0 +1,4 @@ + +LIQ_PRIVATE void liq_blur(unsigned char *src, unsigned char *tmp, unsigned char *dst, unsigned int width, unsigned int height, unsigned int size); +LIQ_PRIVATE void liq_max3(unsigned char *src, unsigned char *dst, unsigned int width, unsigned int height); +LIQ_PRIVATE void liq_min3(unsigned char *src, unsigned char *dst, unsigned int width, unsigned int height); diff --git a/libimagequant.c b/libimagequant.c new file mode 100644 index 0000000..c17b93d --- /dev/null +++ b/libimagequant.c @@ -0,0 +1,1982 @@ +/* +** © 2009-2016 by Kornel Lesiński. +** +** This file is part of libimagequant. +** +** libimagequant 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. +** +** libimagequant 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 libimagequant. If not, see . +*/ +/* Copyright (C) 1989, 1991 by Jef Poskanzer. +** Copyright (C) 1997, 2000, 2002 by Greg Roelofs; based on an idea by +** Stefan Schneider. +** +** Permission to use, copy, modify, and distribute this software and its +** documentation for any purpose and without fee is hereby granted, provided +** that the above copyright notice appear in all copies and that both that +** copyright notice and this permission notice appear in supporting +** documentation. This software is provided "as is" without express or +** implied warranty. +*/ + +#include +#include +#include +#include +#include +#include +#include + +#if !(defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199900L) && !(defined(_MSC_VER) && _MSC_VER >= 1800) +#error "This program requires C99, e.g. -std=c99 switch in GCC or it requires MSVC 18.0 or higher." +#error "Ignore torrent of syntax errors that may follow. It's only because compiler is set to use too old C version." +#endif + +#ifdef _OPENMP +#include +#else +#define omp_get_max_threads() 1 +#define omp_get_thread_num() 0 +#endif + +#include "libimagequant.h" + +#include "pam.h" +#include "mediancut.h" +#include "nearest.h" +#include "blur.h" +#include "viter.h" + +#define LIQ_HIGH_MEMORY_LIMIT (1<<26) /* avoid allocating buffers larger than 64MB */ + +// each structure has a pointer as a unique identifier that allows type checking at run time +static const char liq_attr_magic[] = "liq_attr"; +static const char liq_image_magic[] = "liq_image"; +static const char liq_result_magic[] = "liq_result"; +static const char liq_histogram_magic[] = "liq_histogram"; +static const char liq_remapping_result_magic[] = "liq_remapping_result"; +static const char liq_freed_magic[] = "free"; +#define CHECK_STRUCT_TYPE(attr, kind) liq_crash_if_invalid_handle_pointer_given((const liq_attr*)attr, kind ## _magic) +#define CHECK_USER_POINTER(ptr) liq_crash_if_invalid_pointer_given(ptr) + +struct liq_attr { + const char *magic_header; + void* (*malloc)(size_t); + void (*free)(void*); + + double target_mse, max_mse, voronoi_iteration_limit; + float min_opaque_val; + unsigned int max_colors, max_histogram_entries; + unsigned int min_posterization_output /* user setting */, min_posterization_input /* speed setting */; + unsigned int voronoi_iterations, feedback_loop_trials; + bool last_index_transparent, use_contrast_maps, use_dither_map, fast_palette; + unsigned char speed; + + unsigned char progress_stage1, progress_stage2, progress_stage3; + liq_progress_callback_function *progress_callback; + void *progress_callback_user_info; + + liq_log_callback_function *log_callback; + void *log_callback_user_info; + liq_log_flush_callback_function *log_flush_callback; + void *log_flush_callback_user_info; +}; + +struct liq_image { + const char *magic_header; + void* (*malloc)(size_t); + void (*free)(void*); + + f_pixel *f_pixels; + rgba_pixel **rows; + double gamma; + unsigned int width, height; + unsigned char *noise, *edges, *dither_map; + rgba_pixel *pixels, *temp_row; + f_pixel *temp_f_row; + liq_image_get_rgba_row_callback *row_callback; + void *row_callback_user_info; + float min_opaque_val; + f_pixel fixed_colors[256]; + unsigned short fixed_colors_count; + bool free_pixels, free_rows, free_rows_internal; +}; + +typedef struct liq_remapping_result { + const char *magic_header; + void* (*malloc)(size_t); + void (*free)(void*); + + unsigned char *pixels; + colormap *palette; + liq_progress_callback_function *progress_callback; + void *progress_callback_user_info; + + liq_palette int_palette; + double gamma, palette_error; + float dither_level; + bool use_dither_map; unsigned char progress_stage1; +} liq_remapping_result; + +struct liq_result { + const char *magic_header; + void* (*malloc)(size_t); + void (*free)(void*); + + liq_remapping_result *remapping; + colormap *palette; + liq_progress_callback_function *progress_callback; + void *progress_callback_user_info; + + liq_palette int_palette; + float dither_level; + double gamma, palette_error; + int min_posterization_output; + bool use_dither_map, fast_palette; +}; + +struct liq_histogram { + const char *magic_header; + void* (*malloc)(size_t); + void (*free)(void*); + + struct acolorhash_table *acht; + double gamma; + f_pixel fixed_colors[256]; + unsigned short fixed_colors_count; + unsigned short ignorebits; + bool had_image_added; +}; + +static void modify_alpha(liq_image *input_image, rgba_pixel *const row_pixels) LIQ_NONNULL; +static void contrast_maps(liq_image *image) LIQ_NONNULL; +static liq_error finalize_histogram(liq_histogram *input_hist, liq_attr *options, histogram **hist_output) LIQ_NONNULL; +static const rgba_pixel *liq_image_get_row_rgba(liq_image *input_image, unsigned int row) LIQ_NONNULL; +static const f_pixel *liq_image_get_row_f(liq_image *input_image, unsigned int row) LIQ_NONNULL; +static void liq_remapping_result_destroy(liq_remapping_result *result) LIQ_NONNULL; +static liq_error pngquant_quantize(histogram *hist, const liq_attr *options, const int fixed_colors_count, const f_pixel fixed_colors[], const double gamma, bool fixed_result_colors, liq_result **) LIQ_NONNULL; +static liq_error liq_histogram_quantize_internal(liq_histogram *input_hist, liq_attr *attr, bool fixed_result_colors, liq_result **result_output) LIQ_NONNULL; + +LIQ_NONNULL static void liq_verbose_printf(const liq_attr *context, const char *fmt, ...) +{ + if (context->log_callback) { + va_list va; + va_start(va, fmt); + int required_space = vsnprintf(NULL, 0, fmt, va)+1; // +\0 + va_end(va); + + char buf[required_space]; + va_start(va, fmt); + vsnprintf(buf, required_space, fmt, va); + va_end(va); + + context->log_callback(context, buf, context->log_callback_user_info); + } +} + +LIQ_NONNULL inline static void verbose_print(const liq_attr *attr, const char *msg) +{ + if (attr->log_callback) { + attr->log_callback(attr, msg, attr->log_callback_user_info); + } +} + +LIQ_NONNULL static void liq_verbose_printf_flush(liq_attr *attr) +{ + if (attr->log_flush_callback) { + attr->log_flush_callback(attr, attr->log_flush_callback_user_info); + } +} + +LIQ_NONNULL static bool liq_progress(const liq_attr *attr, const float percent) +{ + return attr->progress_callback && !attr->progress_callback(percent, attr->progress_callback_user_info); +} + +LIQ_NONNULL static bool liq_remap_progress(const liq_remapping_result *quant, const float percent) +{ + return quant->progress_callback && !quant->progress_callback(percent, quant->progress_callback_user_info); +} + +#if USE_SSE +inline static bool is_sse_available() +{ +#if (defined(__x86_64__) || defined(__amd64)) + return true; +#else + int a,b,c,d; + cpuid(1, a, b, c, d); + return d & (1<<25); // edx bit 25 is set when SSE is present +#endif +} +#endif + +/* make it clear in backtrace when user-supplied handle points to invalid memory */ +NEVER_INLINE LIQ_EXPORT bool liq_crash_if_invalid_handle_pointer_given(const liq_attr *user_supplied_pointer, const char *const expected_magic_header); +LIQ_EXPORT bool liq_crash_if_invalid_handle_pointer_given(const liq_attr *user_supplied_pointer, const char *const expected_magic_header) +{ + if (!user_supplied_pointer) { + return false; + } + + if (user_supplied_pointer->magic_header == liq_freed_magic) { + fprintf(stderr, "%s used after being freed", expected_magic_header); + // this is not normal error handling, this is programmer error that should crash the program. + // program cannot safely continue if memory has been used after it's been freed. + // abort() is nasty, but security vulnerability may be worse. + abort(); + } + + return user_supplied_pointer->magic_header == expected_magic_header; +} + +NEVER_INLINE LIQ_EXPORT bool liq_crash_if_invalid_pointer_given(const void *pointer); +LIQ_EXPORT bool liq_crash_if_invalid_pointer_given(const void *pointer) +{ + if (!pointer) { + return false; + } + // Force a read from the given (potentially invalid) memory location in order to check early whether this crashes the program or not. + // It doesn't matter what value is read, the code here is just to shut the compiler up about unused read. + char test_access = *((volatile char *)pointer); + return test_access || true; +} + +LIQ_NONNULL static void liq_log_error(const liq_attr *attr, const char *msg) +{ + if (!CHECK_STRUCT_TYPE(attr, liq_attr)) return; + liq_verbose_printf(attr, " error: %s", msg); +} + +static double quality_to_mse(long quality) +{ + if (quality == 0) { + return MAX_DIFF; + } + if (quality == 100) { + return 0; + } + + // curve fudged to be roughly similar to quality of libjpeg + // except lowest 10 for really low number of colors + const double extra_low_quality_fudge = MAX(0,0.016/(0.001+quality) - 0.001); + return extra_low_quality_fudge + 2.5/pow(210.0 + quality, 1.2) * (100.1-quality)/100.0; +} + +static unsigned int mse_to_quality(double mse) +{ + for(int i=100; i > 0; i--) { + if (mse <= quality_to_mse(i) + 0.000001) { // + epsilon for floating point errors + return i; + } + } + return 0; +} + +/** internally MSE is a sum of all channels with pixels 0..1 range, + but other software gives per-RGB-channel MSE for 0..255 range */ +static double mse_to_standard_mse(double mse) { + return mse * 65536.0/6.0; +} + +LIQ_EXPORT LIQ_NONNULL liq_error liq_set_quality(liq_attr* attr, int minimum, int target) +{ + if (!CHECK_STRUCT_TYPE(attr, liq_attr)) return LIQ_INVALID_POINTER; + if (target < 0 || target > 100 || target < minimum || minimum < 0) return LIQ_VALUE_OUT_OF_RANGE; + + attr->target_mse = quality_to_mse(target); + attr->max_mse = quality_to_mse(minimum); + return LIQ_OK; +} + +LIQ_EXPORT LIQ_NONNULL int liq_get_min_quality(const liq_attr *attr) +{ + if (!CHECK_STRUCT_TYPE(attr, liq_attr)) return -1; + return mse_to_quality(attr->max_mse); +} + +LIQ_EXPORT LIQ_NONNULL int liq_get_max_quality(const liq_attr *attr) +{ + if (!CHECK_STRUCT_TYPE(attr, liq_attr)) return -1; + return mse_to_quality(attr->target_mse); +} + + +LIQ_EXPORT LIQ_NONNULL liq_error liq_set_max_colors(liq_attr* attr, int colors) +{ + if (!CHECK_STRUCT_TYPE(attr, liq_attr)) return LIQ_INVALID_POINTER; + if (colors < 2 || colors > 256) return LIQ_VALUE_OUT_OF_RANGE; + + attr->max_colors = colors; + return LIQ_OK; +} + +LIQ_EXPORT LIQ_NONNULL int liq_get_max_colors(const liq_attr *attr) +{ + if (!CHECK_STRUCT_TYPE(attr, liq_attr)) return -1; + + return attr->max_colors; +} + +LIQ_EXPORT LIQ_NONNULL liq_error liq_set_min_posterization(liq_attr *attr, int bits) +{ + if (!CHECK_STRUCT_TYPE(attr, liq_attr)) return LIQ_INVALID_POINTER; + if (bits < 0 || bits > 4) return LIQ_VALUE_OUT_OF_RANGE; + + attr->min_posterization_output = bits; + return LIQ_OK; +} + +LIQ_EXPORT LIQ_NONNULL int liq_get_min_posterization(const liq_attr *attr) +{ + if (!CHECK_STRUCT_TYPE(attr, liq_attr)) return -1; + + return attr->min_posterization_output; +} + +LIQ_EXPORT LIQ_NONNULL liq_error liq_set_speed(liq_attr* attr, int speed) +{ + if (!CHECK_STRUCT_TYPE(attr, liq_attr)) return LIQ_INVALID_POINTER; + if (speed < 1 || speed > 10) return LIQ_VALUE_OUT_OF_RANGE; + + unsigned int iterations = MAX(8-speed, 0); iterations += iterations * iterations/2; + attr->voronoi_iterations = iterations; + attr->voronoi_iteration_limit = 1.0/(double)(1<<(23-speed)); + attr->feedback_loop_trials = MAX(56-9*speed, 0); + + attr->max_histogram_entries = (1<<17) + (1<<18)*(10-speed); + attr->min_posterization_input = (speed >= 8) ? 1 : 0; + attr->fast_palette = (speed >= 7); + attr->use_dither_map = (speed <= (omp_get_max_threads() > 1 ? 7 : 5)); // parallelized dither map might speed up floyd remapping + attr->use_contrast_maps = (speed <= 7) || attr->use_dither_map; + attr->speed = speed; + + attr->progress_stage1 = attr->use_contrast_maps ? 20 : 8; + if (attr->feedback_loop_trials < 2) attr->progress_stage1 += 30; + attr->progress_stage3 = 50 / (1+speed); + attr->progress_stage2 = 100 - attr->progress_stage1 - attr->progress_stage3; + return LIQ_OK; +} + +LIQ_EXPORT LIQ_NONNULL int liq_get_speed(const liq_attr *attr) +{ + if (!CHECK_STRUCT_TYPE(attr, liq_attr)) return -1; + + return attr->speed; +} + +LIQ_EXPORT LIQ_NONNULL liq_error liq_set_output_gamma(liq_result* res, double gamma) +{ + if (!CHECK_STRUCT_TYPE(res, liq_result)) return LIQ_INVALID_POINTER; + if (gamma <= 0 || gamma >= 1.0) return LIQ_VALUE_OUT_OF_RANGE; + + if (res->remapping) { + liq_remapping_result_destroy(res->remapping); + res->remapping = NULL; + } + + res->gamma = gamma; + return LIQ_OK; +} + +LIQ_EXPORT LIQ_NONNULL liq_error liq_set_min_opacity(liq_attr* attr, int min) +{ + if (!CHECK_STRUCT_TYPE(attr, liq_attr)) return LIQ_INVALID_POINTER; + if (min < 0 || min > 255) return LIQ_VALUE_OUT_OF_RANGE; + + attr->min_opaque_val = (double)min/255.0; + return LIQ_OK; +} + +LIQ_EXPORT LIQ_NONNULL int liq_get_min_opacity(const liq_attr *attr) +{ + if (!CHECK_STRUCT_TYPE(attr, liq_attr)) return -1; + + return MIN(255, 256.0 * attr->min_opaque_val); +} + +LIQ_EXPORT LIQ_NONNULL void liq_set_last_index_transparent(liq_attr* attr, int is_last) +{ + if (!CHECK_STRUCT_TYPE(attr, liq_attr)) return; + + attr->last_index_transparent = !!is_last; +} + +LIQ_EXPORT void liq_attr_set_progress_callback(liq_attr *attr, liq_progress_callback_function *callback, void *user_info) +{ + if (!CHECK_STRUCT_TYPE(attr, liq_attr)) return; + + attr->progress_callback = callback; + attr->progress_callback_user_info = user_info; +} + +LIQ_EXPORT void liq_result_set_progress_callback(liq_result *result, liq_progress_callback_function *callback, void *user_info) +{ + if (!CHECK_STRUCT_TYPE(result, liq_result)) return; + + result->progress_callback = callback; + result->progress_callback_user_info = user_info; +} + +LIQ_EXPORT void liq_set_log_callback(liq_attr *attr, liq_log_callback_function *callback, void* user_info) +{ + if (!CHECK_STRUCT_TYPE(attr, liq_attr)) return; + + liq_verbose_printf_flush(attr); + attr->log_callback = callback; + attr->log_callback_user_info = user_info; +} + +LIQ_EXPORT void liq_set_log_flush_callback(liq_attr *attr, liq_log_flush_callback_function *callback, void* user_info) +{ + if (!CHECK_STRUCT_TYPE(attr, liq_attr)) return; + + attr->log_flush_callback = callback; + attr->log_flush_callback_user_info = user_info; +} + +LIQ_EXPORT liq_attr* liq_attr_create() +{ + return liq_attr_create_with_allocator(NULL, NULL); +} + +LIQ_EXPORT LIQ_NONNULL void liq_attr_destroy(liq_attr *attr) +{ + if (!CHECK_STRUCT_TYPE(attr, liq_attr)) { + return; + } + + liq_verbose_printf_flush(attr); + + attr->magic_header = liq_freed_magic; + attr->free(attr); +} + +LIQ_EXPORT LIQ_NONNULL liq_attr* liq_attr_copy(liq_attr *orig) +{ + if (!CHECK_STRUCT_TYPE(orig, liq_attr)) { + return NULL; + } + + liq_attr *attr = orig->malloc(sizeof(liq_attr)); + if (!attr) return NULL; + *attr = *orig; + return attr; +} + +static void *liq_aligned_malloc(size_t size) +{ + unsigned char *ptr = malloc(size + 16); + if (!ptr) { + return NULL; + } + + uintptr_t offset = 16 - ((uintptr_t)ptr & 15); // also reserves 1 byte for ptr[-1] + ptr += offset; + assert(0 == (((uintptr_t)ptr) & 15)); + ptr[-1] = offset ^ 0x59; // store how much pointer was shifted to get the original for free() + return ptr; +} + +LIQ_NONNULL static void liq_aligned_free(void *inptr) +{ + unsigned char *ptr = inptr; + size_t offset = ptr[-1] ^ 0x59; + assert(offset > 0 && offset <= 16); + free(ptr - offset); +} + +LIQ_EXPORT liq_attr* liq_attr_create_with_allocator(void* (*custom_malloc)(size_t), void (*custom_free)(void*)) +{ +#if USE_SSE + if (!is_sse_available()) { + return NULL; + } +#endif + if (!custom_malloc && !custom_free) { + custom_malloc = liq_aligned_malloc; + custom_free = liq_aligned_free; + } else if (!custom_malloc != !custom_free) { + return NULL; // either specify both or none + } + + liq_attr *attr = custom_malloc(sizeof(liq_attr)); + if (!attr) return NULL; + *attr = (liq_attr) { + .magic_header = liq_attr_magic, + .malloc = custom_malloc, + .free = custom_free, + .max_colors = 256, + .min_opaque_val = 1, // whether preserve opaque colors for IE (1.0=no, does not affect alpha) + .last_index_transparent = false, // puts transparent color at last index. This is workaround for blu-ray subtitles. + .target_mse = 0, + .max_mse = MAX_DIFF, + }; + liq_set_speed(attr, 3); + return attr; +} + +LIQ_EXPORT LIQ_NONNULL liq_error liq_image_add_fixed_color(liq_image *img, liq_color color) +{ + if (!CHECK_STRUCT_TYPE(img, liq_image)) return LIQ_INVALID_POINTER; + if (img->fixed_colors_count > 255) return LIQ_BUFFER_TOO_SMALL; + + float gamma_lut[256]; + to_f_set_gamma(gamma_lut, img->gamma); + img->fixed_colors[img->fixed_colors_count++] = to_f(gamma_lut, (rgba_pixel){ + .r = color.r, + .g = color.g, + .b = color.b, + .a = color.a, + }); + return LIQ_OK; +} + +LIQ_NONNULL static liq_error liq_histogram_add_fixed_color_internal(liq_histogram *hist, f_pixel color) +{ + if (hist->fixed_colors_count > 255) return LIQ_BUFFER_TOO_SMALL; + + hist->fixed_colors[hist->fixed_colors_count++] = color; + return LIQ_OK; +} + +LIQ_NONNULL static bool liq_image_use_low_memory(liq_image *img) +{ + img->temp_f_row = img->malloc(sizeof(img->f_pixels[0]) * img->width * omp_get_max_threads()); + return img->temp_f_row != NULL; +} + +LIQ_NONNULL static bool liq_image_should_use_low_memory(liq_image *img, const bool low_memory_hint) +{ + return img->width * img->height > (low_memory_hint ? LIQ_HIGH_MEMORY_LIMIT/8 : LIQ_HIGH_MEMORY_LIMIT) / sizeof(f_pixel); // Watch out for integer overflow +} + +static liq_image *liq_image_create_internal(const liq_attr *attr, rgba_pixel* rows[], liq_image_get_rgba_row_callback *row_callback, void *row_callback_user_info, int width, int height, double gamma) +{ + if (gamma < 0 || gamma > 1.0) { + liq_log_error(attr, "gamma must be >= 0 and <= 1 (try 1/gamma instead)"); + return NULL; + } + + if (!rows && !row_callback) { + liq_log_error(attr, "missing row data"); + return NULL; + } + + liq_image *img = attr->malloc(sizeof(liq_image)); + if (!img) return NULL; + *img = (liq_image){ + .magic_header = liq_image_magic, + .malloc = attr->malloc, + .free = attr->free, + .width = width, .height = height, + .gamma = gamma ? gamma : 0.45455, + .rows = rows, + .row_callback = row_callback, + .row_callback_user_info = row_callback_user_info, + .min_opaque_val = attr->min_opaque_val, + }; + + if (!rows || attr->min_opaque_val < 1.f) { + img->temp_row = attr->malloc(sizeof(img->temp_row[0]) * width * omp_get_max_threads()); + if (!img->temp_row) return NULL; + } + + // if image is huge or converted pixels are not likely to be reused then don't cache converted pixels + if (liq_image_should_use_low_memory(img, !img->temp_row && !attr->use_contrast_maps && !attr->use_dither_map)) { + verbose_print(attr, " conserving memory"); + if (!liq_image_use_low_memory(img)) return NULL; + } + + if (img->min_opaque_val < 1.f) { + verbose_print(attr, " Working around IE6 bug by making image less transparent..."); + } + + return img; +} + +LIQ_EXPORT LIQ_NONNULL liq_error liq_image_set_memory_ownership(liq_image *img, int ownership_flags) +{ + if (!CHECK_STRUCT_TYPE(img, liq_image)) return LIQ_INVALID_POINTER; + if (!img->rows || !ownership_flags || (ownership_flags & ~(LIQ_OWN_ROWS|LIQ_OWN_PIXELS))) { + return LIQ_VALUE_OUT_OF_RANGE; + } + + if (ownership_flags & LIQ_OWN_ROWS) { + if (img->free_rows_internal) return LIQ_VALUE_OUT_OF_RANGE; + img->free_rows = true; + } + + if (ownership_flags & LIQ_OWN_PIXELS) { + img->free_pixels = true; + if (!img->pixels) { + // for simplicity of this API there's no explicit bitmap argument, + // so the row with the lowest address is assumed to be at the start of the bitmap + img->pixels = img->rows[0]; + for(unsigned int i=1; i < img->height; i++) { + img->pixels = MIN(img->pixels, img->rows[i]); + } + } + } + + return LIQ_OK; +} + +LIQ_NONNULL static bool check_image_size(const liq_attr *attr, const int width, const int height) +{ + if (!CHECK_STRUCT_TYPE(attr, liq_attr)) { + return false; + } + + if (width <= 0 || height <= 0) { + liq_log_error(attr, "width and height must be > 0"); + return false; + } + + if (width > INT_MAX/sizeof(rgba_pixel)/height || width > INT_MAX/16/sizeof(f_pixel) || height > INT_MAX/sizeof(size_t)) { + liq_log_error(attr, "image too large"); + return false; + } + return true; +} + +LIQ_EXPORT liq_image *liq_image_create_custom(const liq_attr *attr, liq_image_get_rgba_row_callback *row_callback, void* user_info, int width, int height, double gamma) +{ + if (!check_image_size(attr, width, height)) { + return NULL; + } + return liq_image_create_internal(attr, NULL, row_callback, user_info, width, height, gamma); +} + +LIQ_EXPORT liq_image *liq_image_create_rgba_rows(const liq_attr *attr, void *const rows[], int width, int height, double gamma) +{ + if (!check_image_size(attr, width, height)) { + return NULL; + } + + for(int i=0; i < height; i++) { + if (!CHECK_USER_POINTER(rows+i) || !CHECK_USER_POINTER(rows[i])) { + liq_log_error(attr, "invalid row pointers"); + return NULL; + } + } + return liq_image_create_internal(attr, (rgba_pixel**)rows, NULL, NULL, width, height, gamma); +} + +LIQ_EXPORT LIQ_NONNULL liq_image *liq_image_create_rgba(const liq_attr *attr, const void* bitmap, int width, int height, double gamma) +{ + if (!check_image_size(attr, width, height)) { + return NULL; + } + if (!CHECK_USER_POINTER(bitmap)) { + liq_log_error(attr, "invalid bitmap pointer"); + return NULL; + } + + rgba_pixel *const pixels = (rgba_pixel *const)bitmap; + rgba_pixel **rows = attr->malloc(sizeof(rows[0])*height); + if (!rows) return NULL; + + for(int i=0; i < height; i++) { + rows[i] = pixels + width * i; + } + + liq_image *image = liq_image_create_internal(attr, rows, NULL, NULL, width, height, gamma); + if (!image) { + attr->free(rows); + return NULL; + } + image->free_rows = true; + image->free_rows_internal = true; + return image; +} + +NEVER_INLINE LIQ_EXPORT void liq_executing_user_callback(liq_image_get_rgba_row_callback *callback, liq_color *temp_row, int row, int width, void *user_info); +LIQ_EXPORT void liq_executing_user_callback(liq_image_get_rgba_row_callback *callback, liq_color *temp_row, int row, int width, void *user_info) +{ + assert(callback); + assert(temp_row); + callback(temp_row, row, width, user_info); +} + +LIQ_NONNULL inline static bool liq_image_has_rgba_pixels(const liq_image *img) +{ + if (!CHECK_STRUCT_TYPE(img, liq_image)) { + return false; + } + return img->rows || (img->temp_row && img->row_callback); +} + +LIQ_NONNULL inline static bool liq_image_can_use_rgba_rows(const liq_image *img) +{ + assert(liq_image_has_rgba_pixels(img)); + + const bool iebug = img->min_opaque_val < 1.f; + return (img->rows && !iebug); +} + +LIQ_NONNULL static const rgba_pixel *liq_image_get_row_rgba(liq_image *img, unsigned int row) +{ + if (liq_image_can_use_rgba_rows(img)) { + return img->rows[row]; + } + + assert(img->temp_row); + rgba_pixel *temp_row = img->temp_row + img->width * omp_get_thread_num(); + if (img->rows) { + memcpy(temp_row, img->rows[row], img->width * sizeof(temp_row[0])); + } else { + liq_executing_user_callback(img->row_callback, (liq_color*)temp_row, row, img->width, img->row_callback_user_info); + } + + if (img->min_opaque_val < 1.f) modify_alpha(img, temp_row); + return temp_row; +} + +LIQ_NONNULL static void convert_row_to_f(liq_image *img, f_pixel *row_f_pixels, const unsigned int row, const float gamma_lut[]) +{ + assert(row_f_pixels); + assert(!USE_SSE || 0 == ((uintptr_t)row_f_pixels & 15)); + + const rgba_pixel *const row_pixels = liq_image_get_row_rgba(img, row); + + for(unsigned int col=0; col < img->width; col++) { + row_f_pixels[col] = to_f(gamma_lut, row_pixels[col]); + } +} + +LIQ_NONNULL static const f_pixel *liq_image_get_row_f(liq_image *img, unsigned int row) +{ + if (!img->f_pixels) { + if (img->temp_f_row) { + float gamma_lut[256]; + to_f_set_gamma(gamma_lut, img->gamma); + f_pixel *row_for_thread = img->temp_f_row + img->width * omp_get_thread_num(); + convert_row_to_f(img, row_for_thread, row, gamma_lut); + return row_for_thread; + } + + assert(omp_get_thread_num() == 0); + if (!liq_image_should_use_low_memory(img, false)) { + img->f_pixels = img->malloc(sizeof(img->f_pixels[0]) * img->width * img->height); + } + if (!img->f_pixels) { + if (!liq_image_use_low_memory(img)) return NULL; + return liq_image_get_row_f(img, row); + } + + float gamma_lut[256]; + to_f_set_gamma(gamma_lut, img->gamma); + for(unsigned int i=0; i < img->height; i++) { + convert_row_to_f(img, &img->f_pixels[i*img->width], i, gamma_lut); + } + } + return img->f_pixels + img->width * row; +} + +LIQ_EXPORT LIQ_NONNULL int liq_image_get_width(const liq_image *input_image) +{ + if (!CHECK_STRUCT_TYPE(input_image, liq_image)) return -1; + return input_image->width; +} + +LIQ_EXPORT LIQ_NONNULL int liq_image_get_height(const liq_image *input_image) +{ + if (!CHECK_STRUCT_TYPE(input_image, liq_image)) return -1; + return input_image->height; +} + +typedef void free_func(void*); + +LIQ_NONNULL static free_func *get_default_free_func(liq_image *img) +{ + // When default allocator is used then user-supplied pointers must be freed with free() + if (img->free_rows_internal || img->free != liq_aligned_free) { + return img->free; + } + return free; +} + +LIQ_NONNULL static void liq_image_free_rgba_source(liq_image *input_image) +{ + if (input_image->free_pixels && input_image->pixels) { + get_default_free_func(input_image)(input_image->pixels); + input_image->pixels = NULL; + } + + if (input_image->free_rows && input_image->rows) { + get_default_free_func(input_image)(input_image->rows); + input_image->rows = NULL; + } +} + +LIQ_EXPORT LIQ_NONNULL void liq_image_destroy(liq_image *input_image) +{ + if (!CHECK_STRUCT_TYPE(input_image, liq_image)) return; + + liq_image_free_rgba_source(input_image); + + if (input_image->noise) { + input_image->free(input_image->noise); + } + + if (input_image->edges) { + input_image->free(input_image->edges); + } + + if (input_image->dither_map) { + input_image->free(input_image->dither_map); + } + + if (input_image->f_pixels) { + input_image->free(input_image->f_pixels); + } + + if (input_image->temp_row) { + input_image->free(input_image->temp_row); + } + + if (input_image->temp_f_row) { + input_image->free(input_image->temp_f_row); + } + + input_image->magic_header = liq_freed_magic; + input_image->free(input_image); +} + +LIQ_EXPORT liq_histogram* liq_histogram_create(liq_attr* attr) +{ + if (!CHECK_STRUCT_TYPE(attr, liq_attr)) { + return NULL; + } + + liq_histogram *hist = attr->malloc(sizeof(liq_histogram)); + if (!hist) return NULL; + *hist = (liq_histogram) { + .magic_header = liq_histogram_magic, + .malloc = attr->malloc, + .free = attr->free, + + .ignorebits = MAX(attr->min_posterization_output, attr->min_posterization_input), + }; + return hist; +} + +LIQ_EXPORT LIQ_NONNULL void liq_histogram_destroy(liq_histogram *hist) +{ + if (!CHECK_STRUCT_TYPE(hist, liq_histogram)) return; + hist->magic_header = liq_freed_magic; + + pam_freeacolorhash(hist->acht); + hist->free(hist); +} + +LIQ_EXPORT LIQ_NONNULL liq_result *liq_quantize_image(liq_attr *attr, liq_image *img) +{ + liq_result *res; + if (LIQ_OK != liq_image_quantize(img, attr, &res)) { + return NULL; + } + return res; +} + +LIQ_EXPORT LIQ_NONNULL liq_error liq_image_quantize(liq_image *const img, liq_attr *const attr, liq_result **result_output) +{ + if (!CHECK_STRUCT_TYPE(attr, liq_attr)) return LIQ_INVALID_POINTER; + if (!liq_image_has_rgba_pixels(img)) { + return LIQ_INVALID_POINTER; + } + + liq_histogram *hist = liq_histogram_create(attr); + if (!hist) { + return LIQ_OUT_OF_MEMORY; + } + liq_error err = liq_histogram_add_image(hist, attr, img); + if (LIQ_OK != err) { + return err; + } + + err = liq_histogram_quantize_internal(hist, attr, false, result_output); + liq_histogram_destroy(hist); + + return err; +} + +LIQ_EXPORT LIQ_NONNULL liq_error liq_histogram_quantize(liq_histogram *input_hist, liq_attr *attr, liq_result **result_output) { + return liq_histogram_quantize_internal(input_hist, attr, true, result_output); +} + +LIQ_NONNULL static liq_error liq_histogram_quantize_internal(liq_histogram *input_hist, liq_attr *attr, bool fixed_result_colors, liq_result **result_output) +{ + if (!CHECK_USER_POINTER(result_output)) return LIQ_INVALID_POINTER; + *result_output = NULL; + + if (!CHECK_STRUCT_TYPE(attr, liq_attr)) return LIQ_INVALID_POINTER; + if (!CHECK_STRUCT_TYPE(input_hist, liq_histogram)) return LIQ_INVALID_POINTER; + + if (liq_progress(attr, 0)) return LIQ_ABORTED; + + histogram *hist; + liq_error err = finalize_histogram(input_hist, attr, &hist); + if (err != LIQ_OK) { + return err; + } + + err = pngquant_quantize(hist, attr, input_hist->fixed_colors_count, input_hist->fixed_colors, input_hist->gamma, fixed_result_colors, result_output); + pam_freeacolorhist(hist); + + return err; +} + +LIQ_EXPORT LIQ_NONNULL liq_error liq_set_dithering_level(liq_result *res, float dither_level) +{ + if (!CHECK_STRUCT_TYPE(res, liq_result)) return LIQ_INVALID_POINTER; + + if (res->remapping) { + liq_remapping_result_destroy(res->remapping); + res->remapping = NULL; + } + + if (res->dither_level < 0 || res->dither_level > 1.0f) return LIQ_VALUE_OUT_OF_RANGE; + res->dither_level = dither_level; + return LIQ_OK; +} + +LIQ_NONNULL static liq_remapping_result *liq_remapping_result_create(liq_result *result) +{ + if (!CHECK_STRUCT_TYPE(result, liq_result)) { + return NULL; + } + + liq_remapping_result *res = result->malloc(sizeof(liq_remapping_result)); + if (!res) return NULL; + *res = (liq_remapping_result) { + .magic_header = liq_remapping_result_magic, + .malloc = result->malloc, + .free = result->free, + .dither_level = result->dither_level, + .use_dither_map = result->use_dither_map, + .palette_error = result->palette_error, + .gamma = result->gamma, + .palette = pam_duplicate_colormap(result->palette), + .progress_callback = result->progress_callback, + .progress_callback_user_info = result->progress_callback_user_info, + .progress_stage1 = result->use_dither_map ? 20 : 0, + }; + return res; +} + +LIQ_EXPORT LIQ_NONNULL double liq_get_output_gamma(const liq_result *result) +{ + if (!CHECK_STRUCT_TYPE(result, liq_result)) return -1; + + return result->gamma; +} + +LIQ_NONNULL static void liq_remapping_result_destroy(liq_remapping_result *result) +{ + if (!CHECK_STRUCT_TYPE(result, liq_remapping_result)) return; + + if (result->palette) pam_freecolormap(result->palette); + if (result->pixels) result->free(result->pixels); + + result->magic_header = liq_freed_magic; + result->free(result); +} + +LIQ_EXPORT LIQ_NONNULL void liq_result_destroy(liq_result *res) +{ + if (!CHECK_STRUCT_TYPE(res, liq_result)) return; + + memset(&res->int_palette, 0, sizeof(liq_palette)); + + if (res->remapping) { + memset(&res->remapping->int_palette, 0, sizeof(liq_palette)); + liq_remapping_result_destroy(res->remapping); + } + + pam_freecolormap(res->palette); + + res->magic_header = liq_freed_magic; + res->free(res); +} + + +LIQ_EXPORT LIQ_NONNULL double liq_get_quantization_error(liq_result *result) { + if (!CHECK_STRUCT_TYPE(result, liq_result)) return -1; + + if (result->palette_error >= 0) { + return mse_to_standard_mse(result->palette_error); + } + + return -1; +} + +LIQ_EXPORT LIQ_NONNULL double liq_get_remapping_error(liq_result *result) { + if (!CHECK_STRUCT_TYPE(result, liq_result)) return -1; + + if (result->remapping && result->remapping->palette_error >= 0) { + return mse_to_standard_mse(result->remapping->palette_error); + } + + return -1; +} + +LIQ_EXPORT LIQ_NONNULL int liq_get_quantization_quality(liq_result *result) { + if (!CHECK_STRUCT_TYPE(result, liq_result)) return -1; + + if (result->palette_error >= 0) { + return mse_to_quality(result->palette_error); + } + + return -1; +} + +LIQ_EXPORT LIQ_NONNULL int liq_get_remapping_quality(liq_result *result) { + if (!CHECK_STRUCT_TYPE(result, liq_result)) return -1; + + if (result->remapping && result->remapping->palette_error >= 0) { + return mse_to_quality(result->remapping->palette_error); + } + + return -1; +} + +LIQ_NONNULL static int compare_popularity(const void *ch1, const void *ch2) +{ + const float v1 = ((const colormap_item*)ch1)->popularity; + const float v2 = ((const colormap_item*)ch2)->popularity; + return v1 > v2 ? -1 : 1; +} + +LIQ_NONNULL static void sort_palette_qsort(colormap *map, int start, int nelem) +{ + if (!nelem) return; + qsort(map->palette + start, nelem, sizeof(map->palette[0]), compare_popularity); +} + +#define SWAP_PALETTE(map, a,b) { \ + const colormap_item tmp = (map)->palette[(a)]; \ + (map)->palette[(a)] = (map)->palette[(b)]; \ + (map)->palette[(b)] = tmp; } + +LIQ_NONNULL static void sort_palette(colormap *map, const liq_attr *options) +{ + /* + ** Step 3.5 [GRR]: remap the palette colors so that all entries with + ** the maximal alpha value (i.e., fully opaque) are at the end and can + ** therefore be omitted from the tRNS chunk. + */ + if (options->last_index_transparent) { + for(unsigned int i=0; i < map->colors; i++) { + if (map->palette[i].acolor.a < 1.0/256.0) { + const unsigned int old = i, transparent_dest = map->colors-1; + + SWAP_PALETTE(map, transparent_dest, old); + + /* colors sorted by popularity make pngs slightly more compressible */ + sort_palette_qsort(map, 0, map->colors-1); + return; + } + } + } + + unsigned int non_fixed_colors = 0; + for(unsigned int i = 0; i < map->colors; i++) { + if (map->palette[i].fixed) { + break; + } + non_fixed_colors++; + } + + /* move transparent colors to the beginning to shrink trns chunk */ + unsigned int num_transparent = 0; + for(unsigned int i = 0; i < non_fixed_colors; i++) { + if (map->palette[i].acolor.a < 255.0/256.0) { + // current transparent color is swapped with earlier opaque one + if (i != num_transparent) { + SWAP_PALETTE(map, num_transparent, i); + i--; + } + num_transparent++; + } + } + + liq_verbose_printf(options, " eliminated opaque tRNS-chunk entries...%d entr%s transparent", num_transparent, (num_transparent == 1)? "y" : "ies"); + + /* colors sorted by popularity make pngs slightly more compressible + * opaque and transparent are sorted separately + */ + sort_palette_qsort(map, 0, num_transparent); + sort_palette_qsort(map, num_transparent, non_fixed_colors - num_transparent); + + if (non_fixed_colors > 9 && map->colors > 16) { + SWAP_PALETTE(map, 7, 1); // slightly improves compression + SWAP_PALETTE(map, 8, 2); + SWAP_PALETTE(map, 9, 3); + } +} + +inline static unsigned int posterize_channel(unsigned int color, unsigned int bits) +{ + return (color & ~((1<> (8-bits)); +} + +LIQ_NONNULL static void set_rounded_palette(liq_palette *const dest, colormap *const map, const double gamma, unsigned int posterize) +{ + float gamma_lut[256]; + to_f_set_gamma(gamma_lut, gamma); + + dest->count = map->colors; + for(unsigned int x = 0; x < map->colors; ++x) { + rgba_pixel px = to_rgb(gamma, map->palette[x].acolor); + + px.r = posterize_channel(px.r, posterize); + px.g = posterize_channel(px.g, posterize); + px.b = posterize_channel(px.b, posterize); + px.a = posterize_channel(px.a, posterize); + + map->palette[x].acolor = to_f(gamma_lut, px); /* saves rounding error introduced by to_rgb, which makes remapping & dithering more accurate */ + + if (!px.a && !map->palette[x].fixed) { + px.r = 71; px.g = 112; px.b = 76; + } + + dest->entries[x] = (liq_color){.r=px.r,.g=px.g,.b=px.b,.a=px.a}; + } +} + +LIQ_EXPORT LIQ_NONNULL const liq_palette *liq_get_palette(liq_result *result) +{ + if (!CHECK_STRUCT_TYPE(result, liq_result)) return NULL; + + if (result->remapping && result->remapping->int_palette.count) { + return &result->remapping->int_palette; + } + + if (!result->int_palette.count) { + set_rounded_palette(&result->int_palette, result->palette, result->gamma, result->min_posterization_output); + } + return &result->int_palette; +} + +LIQ_NONNULL static float remap_to_palette(liq_image *const input_image, unsigned char *const *const output_pixels, colormap *const map, const bool fast) +{ + const int rows = input_image->height; + const unsigned int cols = input_image->width; + double remapping_error=0; + + if (!liq_image_get_row_f(input_image, 0)) { // trigger lazy conversion + return -1; + } + + struct nearest_map *const n = nearest_init(map, fast); + + const unsigned int max_threads = omp_get_max_threads(); + viter_state average_color[(VITER_CACHE_LINE_GAP+map->colors) * max_threads]; + viter_init(map, max_threads, average_color); + + #pragma omp parallel for if (rows*cols > 3000) \ + schedule(static) default(none) shared(average_color) reduction(+:remapping_error) + for(int row = 0; row < rows; ++row) { + const f_pixel *const row_pixels = liq_image_get_row_f(input_image, row); + unsigned int last_match=0; + for(unsigned int col = 0; col < cols; ++col) { + float diff; + output_pixels[row][col] = last_match = nearest_search(n, &row_pixels[col], last_match, &diff); + + remapping_error += diff; + viter_update_color(row_pixels[col], 1.0, map, last_match, omp_get_thread_num(), average_color); + } + } + + viter_finalize(map, max_threads, average_color); + + nearest_free(n); + + return remapping_error / (input_image->width * input_image->height); +} + +inline static f_pixel get_dithered_pixel(const float dither_level, const float max_dither_error, const f_pixel thiserr, const f_pixel px) +{ + /* Use Floyd-Steinberg errors to adjust actual color. */ + const float sr = thiserr.r * dither_level, + sg = thiserr.g * dither_level, + sb = thiserr.b * dither_level, + sa = thiserr.a * dither_level; + + float ratio = 1.0; + const float max_overflow = 1.1f; + const float max_underflow = -0.1f; + + // allowing some overflow prevents undithered bands caused by clamping of all channels + if (px.r + sr > max_overflow) ratio = MIN(ratio, (max_overflow -px.r)/sr); + else if (px.r + sr < max_underflow) ratio = MIN(ratio, (max_underflow-px.r)/sr); + if (px.g + sg > max_overflow) ratio = MIN(ratio, (max_overflow -px.g)/sg); + else if (px.g + sg < max_underflow) ratio = MIN(ratio, (max_underflow-px.g)/sg); + if (px.b + sb > max_overflow) ratio = MIN(ratio, (max_overflow -px.b)/sb); + else if (px.b + sb < max_underflow) ratio = MIN(ratio, (max_underflow-px.b)/sb); + + float a = px.a + sa; + if (a > 1.0) { a = 1.0; } + else if (a < 0) { a = 0; } + + // If dithering error is crazy high, don't propagate it that much + // This prevents crazy geen pixels popping out of the blue (or red or black! ;) + const float dither_error = sr*sr + sg*sg + sb*sb + sa*sa; + if (dither_error > max_dither_error) { + ratio *= 0.8; + } else if (dither_error < 2.f/256.f/256.f) { + // don't dither areas that don't have noticeable error — makes file smaller + return px; + } + + return (f_pixel){ + .r=px.r + sr * ratio, + .g=px.g + sg * ratio, + .b=px.b + sb * ratio, + .a=a, + }; +} + +/** + Uses edge/noise map to apply dithering only to flat areas. Dithering on edges creates jagged lines, and noisy areas are "naturally" dithered. + + If output_image_is_remapped is true, only pixels noticeably changed by error diffusion will be written to output image. + */ +LIQ_NONNULL static bool remap_to_palette_floyd(liq_image *input_image, unsigned char *const output_pixels[], liq_remapping_result *quant, const float max_dither_error, const bool output_image_is_remapped) +{ + const unsigned int rows = input_image->height, cols = input_image->width; + const unsigned char *dither_map = quant->use_dither_map ? (input_image->dither_map ? input_image->dither_map : input_image->edges) : NULL; + + const colormap *map = quant->palette; + const colormap_item *acolormap = map->palette; + + /* Initialize Floyd-Steinberg error vectors. */ + f_pixel *restrict thiserr, *restrict nexterr; + const size_t errsize = (cols + 2) * sizeof(*thiserr) * 2; + thiserr = input_image->malloc(errsize); // +2 saves from checking out of bounds access + if (!thiserr) return false; + memset(thiserr, 0, errsize); + nexterr = thiserr + (cols + 2); + + bool ok = true; + struct nearest_map *const n = nearest_init(map, false); + + // response to this value is non-linear and without it any value < 0.8 would give almost no dithering + float base_dithering_level = quant->dither_level; + base_dithering_level = 1.0 - (1.0-base_dithering_level)*(1.0-base_dithering_level); + + if (dither_map) { + base_dithering_level *= 1.0/255.0; // convert byte to float + } + base_dithering_level *= 15.0/16.0; // prevent small errors from accumulating + + bool fs_direction = true; + unsigned int last_match=0; + for (unsigned int row = 0; row < rows; ++row) { + if (liq_remap_progress(quant, quant->progress_stage1 + row * (100.f - quant->progress_stage1) / rows)) { + ok = false; + break; + } + + memset(nexterr, 0, (cols + 2) * sizeof(*nexterr)); + + unsigned int col = (fs_direction) ? 0 : (cols - 1); + const f_pixel *const row_pixels = liq_image_get_row_f(input_image, row); + + do { + float dither_level = base_dithering_level; + if (dither_map) { + dither_level *= dither_map[row*cols + col]; + } + + const f_pixel spx = get_dithered_pixel(dither_level, max_dither_error, thiserr[col + 1], row_pixels[col]); + + const unsigned int guessed_match = output_image_is_remapped ? output_pixels[row][col] : last_match; + output_pixels[row][col] = last_match = nearest_search(n, &spx, guessed_match, NULL); + + const f_pixel output_px = acolormap[last_match].acolor; + f_pixel err = { + .r = (spx.r - output_px.r), + .g = (spx.g - output_px.g), + .b = (spx.b - output_px.b), + .a = (spx.a - output_px.a), + }; + + // If dithering error is crazy high, don't propagate it that much + // This prevents crazy geen pixels popping out of the blue (or red or black! ;) + if (err.r*err.r + err.g*err.g + err.b*err.b + err.a*err.a > max_dither_error) { + err.r *= 0.75; + err.g *= 0.75; + err.b *= 0.75; + err.a *= 0.75; + } + + /* Propagate Floyd-Steinberg error terms. */ + if (fs_direction) { + thiserr[col + 2].a += err.a * (7.f/16.f); + thiserr[col + 2].r += err.r * (7.f/16.f); + thiserr[col + 2].g += err.g * (7.f/16.f); + thiserr[col + 2].b += err.b * (7.f/16.f); + + nexterr[col + 2].a = err.a * (1.f/16.f); + nexterr[col + 2].r = err.r * (1.f/16.f); + nexterr[col + 2].g = err.g * (1.f/16.f); + nexterr[col + 2].b = err.b * (1.f/16.f); + + nexterr[col + 1].a += err.a * (5.f/16.f); + nexterr[col + 1].r += err.r * (5.f/16.f); + nexterr[col + 1].g += err.g * (5.f/16.f); + nexterr[col + 1].b += err.b * (5.f/16.f); + + nexterr[col ].a += err.a * (3.f/16.f); + nexterr[col ].r += err.r * (3.f/16.f); + nexterr[col ].g += err.g * (3.f/16.f); + nexterr[col ].b += err.b * (3.f/16.f); + + } else { + thiserr[col ].a += err.a * (7.f/16.f); + thiserr[col ].r += err.r * (7.f/16.f); + thiserr[col ].g += err.g * (7.f/16.f); + thiserr[col ].b += err.b * (7.f/16.f); + + nexterr[col ].a = err.a * (1.f/16.f); + nexterr[col ].r = err.r * (1.f/16.f); + nexterr[col ].g = err.g * (1.f/16.f); + nexterr[col ].b = err.b * (1.f/16.f); + + nexterr[col + 1].a += err.a * (5.f/16.f); + nexterr[col + 1].r += err.r * (5.f/16.f); + nexterr[col + 1].g += err.g * (5.f/16.f); + nexterr[col + 1].b += err.b * (5.f/16.f); + + nexterr[col + 2].a += err.a * (3.f/16.f); + nexterr[col + 2].r += err.r * (3.f/16.f); + nexterr[col + 2].g += err.g * (3.f/16.f); + nexterr[col + 2].b += err.b * (3.f/16.f); + } + + // remapping is done in zig-zag + if (fs_direction) { + ++col; + if (col >= cols) break; + } else { + if (col <= 0) break; + --col; + } + } while(1); + + f_pixel *const temperr = thiserr; + thiserr = nexterr; + nexterr = temperr; + fs_direction = !fs_direction; + } + + input_image->free(MIN(thiserr, nexterr)); // MIN because pointers were swapped + nearest_free(n); + + return ok; +} + +/* fixed colors are always included in the palette, so it would be wasteful to duplicate them in palette from histogram */ +LIQ_NONNULL static void remove_fixed_colors_from_histogram(histogram *hist, const int fixed_colors_count, const f_pixel fixed_colors[], const float target_mse) +{ + const float max_difference = MAX(target_mse/2.0, 2.0/256.0/256.0); + if (fixed_colors_count) { + for(int j=0; j < hist->size; j++) { + for(unsigned int i=0; i < fixed_colors_count; i++) { + if (colordifference(hist->achv[j].acolor, fixed_colors[i]) < max_difference) { + hist->achv[j] = hist->achv[--hist->size]; // remove color from histogram by overwriting with the last entry + j--; break; // continue searching histogram + } + } + } + } +} + +LIQ_EXPORT LIQ_NONNULL liq_error liq_histogram_add_image(liq_histogram *input_hist, liq_attr *options, liq_image *input_image) +{ + const unsigned int cols = input_image->width, rows = input_image->height; + + if (!input_image->noise && options->use_contrast_maps) { + contrast_maps(input_image); + } + + input_hist->gamma = input_image->gamma; + + for(int i = 0; i < input_image->fixed_colors_count; i++) { + liq_error res = liq_histogram_add_fixed_color_internal(input_hist, input_image->fixed_colors[i]); + if (res != LIQ_OK) { + return res; + } + } + + /* + ** Step 2: attempt to make a histogram of the colors, unclustered. + ** If at first we don't succeed, increase ignorebits to increase color + ** coherence and try again. + */ + + if (liq_progress(options, options->progress_stage1 * 0.4f)) return LIQ_ABORTED; + + const bool all_rows_at_once = liq_image_can_use_rgba_rows(input_image); + + // Usual solution is to start from scratch when limit is exceeded, but that's not possible if it's not + // the first image added + const unsigned int max_histogram_entries = input_hist->had_image_added ? ~0 : options->max_histogram_entries; + do { + if (!input_hist->acht) { + input_hist->acht = pam_allocacolorhash(max_histogram_entries, rows*cols, input_hist->ignorebits, options->malloc, options->free); + } + if (!input_hist->acht) return LIQ_OUT_OF_MEMORY; + + // histogram uses noise contrast map for importance. Color accuracy in noisy areas is not very important. + // noise map does not include edges to avoid ruining anti-aliasing + for(unsigned int row=0; row < rows; row++) { + bool added_ok; + if (all_rows_at_once) { + added_ok = pam_computeacolorhash(input_hist->acht, (const rgba_pixel *const *)input_image->rows, cols, rows, input_image->noise); + if (added_ok) break; + } else { + const rgba_pixel* rows_p[1] = { liq_image_get_row_rgba(input_image, row) }; + added_ok = pam_computeacolorhash(input_hist->acht, rows_p, cols, 1, input_image->noise ? &input_image->noise[row * cols] : NULL); + } + if (!added_ok) { + input_hist->ignorebits++; + liq_verbose_printf(options, " too many colors! Scaling colors to improve clustering... %d", input_hist->ignorebits); + pam_freeacolorhash(input_hist->acht); + input_hist->acht = NULL; + if (liq_progress(options, options->progress_stage1 * 0.6f)) return LIQ_ABORTED; + break; + } + } + } while(!input_hist->acht); + + input_hist->had_image_added = true; + + if (input_image->noise) { + input_image->free(input_image->noise); + input_image->noise = NULL; + } + + if (input_image->free_pixels && input_image->f_pixels) { + liq_image_free_rgba_source(input_image); // bow can free the RGBA source if copy has been made in f_pixels + } + + return LIQ_OK; +} + +LIQ_NONNULL static liq_error finalize_histogram(liq_histogram *input_hist, liq_attr *options, histogram **hist_output) +{ + if (liq_progress(options, options->progress_stage1 * 0.9f)) { + return LIQ_ABORTED; + } + + if (!input_hist->acht) { + return LIQ_BITMAP_NOT_AVAILABLE; + } + + histogram *hist = pam_acolorhashtoacolorhist(input_hist->acht, input_hist->gamma, options->malloc, options->free); + pam_freeacolorhash(input_hist->acht); + input_hist->acht = NULL; + + if (!hist) { + return LIQ_OUT_OF_MEMORY; + } + liq_verbose_printf(options, " made histogram...%d colors found", hist->size); + remove_fixed_colors_from_histogram(hist, input_hist->fixed_colors_count, input_hist->fixed_colors, options->target_mse); + + *hist_output = hist; + return LIQ_OK; +} + +LIQ_NONNULL static void modify_alpha(liq_image *input_image, rgba_pixel *const row_pixels) +{ + /* IE6 makes colors with even slightest transparency completely transparent, + thus to improve situation in IE, make colors that are less than ~10% transparent + completely opaque */ + + const float min_opaque_val = input_image->min_opaque_val; + const float almost_opaque_val = min_opaque_val * 169.f/256.f; + const unsigned int almost_opaque_val_int = (min_opaque_val * 169.f/256.f)*255.f; + + for(unsigned int col = 0; col < input_image->width; col++) { + const rgba_pixel px = row_pixels[col]; + + /* ie bug: to avoid visible step caused by forced opaqueness, linearily raise opaqueness of almost-opaque colors */ + if (px.a >= almost_opaque_val_int) { + float al = px.a / 255.f; + al = almost_opaque_val + (al-almost_opaque_val) * (1.f-almost_opaque_val) / (min_opaque_val-almost_opaque_val); + al *= 256.f; + row_pixels[col].a = al >= 255.f ? 255 : al; + } + } +} + +/** + Builds two maps: + noise - approximation of areas with high-frequency noise, except straight edges. 1=flat, 0=noisy. + edges - noise map including all edges + */ +LIQ_NONNULL static void contrast_maps(liq_image *image) +{ + const unsigned int cols = image->width, rows = image->height; + if (cols < 4 || rows < 4 || (3*cols*rows) > LIQ_HIGH_MEMORY_LIMIT) { + return; + } + + unsigned char *restrict noise = image->noise ? image->noise : image->malloc(cols*rows); + image->noise = NULL; + unsigned char *restrict edges = image->edges ? image->edges : image->malloc(cols*rows); + image->edges = NULL; + + unsigned char *restrict tmp = image->malloc(cols*rows); + + if (!noise || !edges || !tmp) { + image->free(noise); + image->free(edges); + image->free(tmp); + return; + } + + const f_pixel *curr_row, *prev_row, *next_row; + curr_row = prev_row = next_row = liq_image_get_row_f(image, 0); + + for (unsigned int j=0; j < rows; j++) { + prev_row = curr_row; + curr_row = next_row; + next_row = liq_image_get_row_f(image, MIN(rows-1,j+1)); + + f_pixel prev, curr = curr_row[0], next=curr; + for (unsigned int i=0; i < cols; i++) { + prev=curr; + curr=next; + next = curr_row[MIN(cols-1,i+1)]; + + // contrast is difference between pixels neighbouring horizontally and vertically + const float a = fabsf(prev.a+next.a - curr.a*2.f), + r = fabsf(prev.r+next.r - curr.r*2.f), + g = fabsf(prev.g+next.g - curr.g*2.f), + b = fabsf(prev.b+next.b - curr.b*2.f); + + const f_pixel prevl = prev_row[i]; + const f_pixel nextl = next_row[i]; + + const float a1 = fabsf(prevl.a+nextl.a - curr.a*2.f), + r1 = fabsf(prevl.r+nextl.r - curr.r*2.f), + g1 = fabsf(prevl.g+nextl.g - curr.g*2.f), + b1 = fabsf(prevl.b+nextl.b - curr.b*2.f); + + const float horiz = MAX(MAX(a,r),MAX(g,b)); + const float vert = MAX(MAX(a1,r1),MAX(g1,b1)); + const float edge = MAX(horiz,vert); + float z = edge - fabsf(horiz-vert)*.5f; + z = 1.f - MAX(z,MIN(horiz,vert)); + z *= z; // noise is amplified + z *= z; + + z *= 256.f; + noise[j*cols+i] = z < 256 ? z : 255; + z = (1.f-edge)*256.f; + edges[j*cols+i] = z < 256 ? z : 255; + } + } + + // noise areas are shrunk and then expanded to remove thin edges from the map + liq_max3(noise, tmp, cols, rows); + liq_max3(tmp, noise, cols, rows); + + liq_blur(noise, tmp, noise, cols, rows, 3); + + liq_max3(noise, tmp, cols, rows); + + liq_min3(tmp, noise, cols, rows); + liq_min3(noise, tmp, cols, rows); + liq_min3(tmp, noise, cols, rows); + + liq_min3(edges, tmp, cols, rows); + liq_max3(tmp, edges, cols, rows); + for(unsigned int i=0; i < cols*rows; i++) edges[i] = MIN(noise[i], edges[i]); + + image->free(tmp); + + image->noise = noise; + image->edges = edges; +} + +/** + * Builds map of neighbor pixels mapped to the same palette entry + * + * For efficiency/simplicity it mainly looks for same consecutive pixels horizontally + * and peeks 1 pixel above/below. Full 2d algorithm doesn't improve it significantly. + * Correct flood fill doesn't have visually good properties. + */ +LIQ_NONNULL static void update_dither_map(unsigned char *const *const row_pointers, liq_image *input_image) +{ + const unsigned int width = input_image->width; + const unsigned int height = input_image->height; + unsigned char *const edges = input_image->edges; + + for(unsigned int row=0; row < height; row++) { + unsigned char lastpixel = row_pointers[row][0]; + unsigned int lastcol=0; + + for(unsigned int col=1; col < width; col++) { + const unsigned char px = row_pointers[row][col]; + + if (px != lastpixel || col == width-1) { + int neighbor_count = 10 * (col-lastcol); + + unsigned int i=lastcol; + while(i < col) { + if (row > 0) { + unsigned char pixelabove = row_pointers[row-1][i]; + if (pixelabove == lastpixel) neighbor_count += 15; + } + if (row < height-1) { + unsigned char pixelbelow = row_pointers[row+1][i]; + if (pixelbelow == lastpixel) neighbor_count += 15; + } + i++; + } + + while(lastcol <= col) { + int e = edges[row*width + lastcol]; + edges[row*width + lastcol++] = (e+128) * (255.f/(255+128)) * (1.f - 20.f / (20 + neighbor_count)); + } + lastpixel = px; + } + } + } + input_image->dither_map = input_image->edges; + input_image->edges = NULL; +} + +/** + * Palette can be NULL, in which case it creates a new palette from scratch. + */ +static colormap *add_fixed_colors_to_palette(colormap *palette, const int max_colors, const f_pixel fixed_colors[], const int fixed_colors_count, void* (*malloc)(size_t), void (*free)(void*)) +{ + if (!fixed_colors_count) return palette; + + colormap *newpal = pam_colormap(MIN(max_colors, (palette ? palette->colors : 0) + fixed_colors_count), malloc, free); + unsigned int i=0; + if (palette && fixed_colors_count < max_colors) { + unsigned int palette_max = MIN(palette->colors, max_colors - fixed_colors_count); + for(; i < palette_max; i++) { + newpal->palette[i] = palette->palette[i]; + } + } + for(int j=0; j < MIN(max_colors, fixed_colors_count); j++) { + newpal->palette[i++] = (colormap_item){ + .acolor = fixed_colors[j], + .fixed = true, + }; + } + if (palette) pam_freecolormap(palette); + return newpal; +} + +LIQ_NONNULL static void adjust_histogram_callback(hist_item *item, float diff) +{ + item->adjusted_weight = (item->perceptual_weight+item->adjusted_weight) * (sqrtf(1.f+diff)); +} + +/** + Repeats mediancut with different histogram weights to find palette with minimum error. + + feedback_loop_trials controls how long the search will take. < 0 skips the iteration. + */ +static colormap *find_best_palette(histogram *hist, const liq_attr *options, const double max_mse, const f_pixel fixed_colors[], const unsigned int fixed_colors_count, double *palette_error_p) +{ + unsigned int max_colors = options->max_colors; + + // if output is posterized it doesn't make sense to aim for perfrect colors, so increase target_mse + // at this point actual gamma is not set, so very conservative posterization estimate is used + const double target_mse = MIN(max_mse, MAX(options->target_mse, pow((1<min_posterization_output)/1024.0, 2))); + int feedback_loop_trials = options->feedback_loop_trials; + colormap *acolormap = NULL; + double least_error = MAX_DIFF; + double target_mse_overshoot = feedback_loop_trials>0 ? 1.05 : 1.0; + const float total_trials = (float)(feedback_loop_trials>0?feedback_loop_trials:1); + + do { + colormap *newmap; + if (hist->size && fixed_colors_count < max_colors) { + newmap = mediancut(hist, max_colors-fixed_colors_count, target_mse * target_mse_overshoot, MAX(MAX(45.0/65536.0, target_mse), least_error)*1.2, + options->malloc, options->free); + } else { + feedback_loop_trials = 0; + newmap = NULL; + } + newmap = add_fixed_colors_to_palette(newmap, max_colors, fixed_colors, fixed_colors_count, options->malloc, options->free); + if (!newmap) { + return NULL; + } + + if (feedback_loop_trials <= 0) { + return newmap; + } + + // after palette has been created, total error (MSE) is calculated to keep the best palette + // at the same time Voronoi iteration is done to improve the palette + // and histogram weights are adjusted based on remapping error to give more weight to poorly matched colors + + const bool first_run_of_target_mse = !acolormap && target_mse > 0; + double total_error = viter_do_iteration(hist, newmap, first_run_of_target_mse ? NULL : adjust_histogram_callback, !acolormap || options->fast_palette); + + // goal is to increase quality or to reduce number of colors used if quality is good enough + if (!acolormap || total_error < least_error || (total_error <= target_mse && newmap->colors < max_colors)) { + if (acolormap) pam_freecolormap(acolormap); + acolormap = newmap; + + if (total_error < target_mse && total_error > 0) { + // voronoi iteration improves quality above what mediancut aims for + // this compensates for it, making mediancut aim for worse + target_mse_overshoot = MIN(target_mse_overshoot*1.25, target_mse/total_error); + } + + least_error = total_error; + + // if number of colors could be reduced, try to keep it that way + // but allow extra color as a bit of wiggle room in case quality can be improved too + max_colors = MIN(newmap->colors+1, max_colors); + + feedback_loop_trials -= 1; // asymptotic improvement could make it go on forever + } else { + for(unsigned int j=0; j < hist->size; j++) { + hist->achv[j].adjusted_weight = (hist->achv[j].perceptual_weight + hist->achv[j].adjusted_weight)/2.0; + } + + target_mse_overshoot = 1.0; + feedback_loop_trials -= 6; + // if error is really bad, it's unlikely to improve, so end sooner + if (total_error > least_error*4) feedback_loop_trials -= 3; + pam_freecolormap(newmap); + } + + float fraction_done = 1.f-MAX(0.f, feedback_loop_trials/total_trials); + if (liq_progress(options, options->progress_stage1 + fraction_done * options->progress_stage2)) break; + liq_verbose_printf(options, " selecting colors...%d%%", (int)(100.f * fraction_done)); + } + while(feedback_loop_trials > 0); + + *palette_error_p = least_error; + return acolormap; +} + +static colormap *histogram_to_palette(const histogram *hist, const liq_attr *options) { + if (!hist->size) { + return NULL; + } + colormap *acolormap = pam_colormap(hist->size, options->malloc, options->free); + for(unsigned int i=0; i < hist->size; i++) { + acolormap->palette[i].acolor = hist->achv[i].acolor; + acolormap->palette[i].popularity = hist->achv[i].perceptual_weight; + } + return acolormap; +} + +LIQ_NONNULL static liq_error pngquant_quantize(histogram *hist, const liq_attr *options, const int fixed_colors_count, const f_pixel fixed_colors[], const double gamma, bool fixed_result_colors, liq_result **result_output) +{ + colormap *acolormap; + double palette_error = -1; + + assert((verbose_print(options, "SLOW debug checks enabled. Recompile with NDEBUG for normal operation."),1)); + + // no point having perfect match with imperfect colors (ignorebits > 0) + const bool fast_palette = options->fast_palette || hist->ignorebits > 0; + const bool few_input_colors = hist->size+fixed_colors_count <= options->max_colors; + + if (liq_progress(options, options->progress_stage1)) return LIQ_ABORTED; + + // If image has few colors to begin with (and no quality degradation is required) + // then it's possible to skip quantization entirely + if (few_input_colors && options->target_mse == 0) { + acolormap = add_fixed_colors_to_palette(histogram_to_palette(hist, options), options->max_colors, fixed_colors, fixed_colors_count, options->malloc, options->free); + palette_error = 0; + } else { + const double max_mse = options->max_mse * (few_input_colors ? 0.33 : 1.0); // when degrading image that's already paletted, require much higher improvement, since pal2pal often looks bad and there's little gain + acolormap = find_best_palette(hist, options, max_mse, fixed_colors, fixed_colors_count, &palette_error); + if (!acolormap) { + return LIQ_VALUE_OUT_OF_RANGE; + } + + // Voronoi iteration approaches local minimum for the palette + const double iteration_limit = options->voronoi_iteration_limit; + unsigned int iterations = options->voronoi_iterations; + + if (!iterations && palette_error < 0 && max_mse < MAX_DIFF) iterations = 1; // otherwise total error is never calculated and MSE limit won't work + + if (iterations) { + // likely_colormap_index (used and set in viter_do_iteration) can't point to index outside colormap + if (acolormap->colors < 256) for(unsigned int j=0; j < hist->size; j++) { + if (hist->achv[j].tmp.likely_colormap_index >= acolormap->colors) { + hist->achv[j].tmp.likely_colormap_index = 0; // actual value doesn't matter, as the guess is out of date anyway + } + } + + verbose_print(options, " moving colormap towards local minimum"); + + double previous_palette_error = MAX_DIFF; + + for(unsigned int i=0; i < iterations; i++) { + palette_error = viter_do_iteration(hist, acolormap, NULL, i==0 || options->fast_palette); + + if (liq_progress(options, options->progress_stage1 + options->progress_stage2 + (i * options->progress_stage3 * 0.9f) / iterations)) { + break; + } + + if (fabs(previous_palette_error-palette_error) < iteration_limit) { + break; + } + + if (palette_error > max_mse*1.5) { // probably hopeless + if (palette_error > max_mse*3.0) break; // definitely hopeless + i++; + } + + previous_palette_error = palette_error; + } + } + + if (palette_error > max_mse) { + liq_verbose_printf(options, " image degradation MSE=%.3f (Q=%d) exceeded limit of %.3f (%d)", + mse_to_standard_mse(palette_error), mse_to_quality(palette_error), + mse_to_standard_mse(max_mse), mse_to_quality(max_mse)); + pam_freecolormap(acolormap); + return LIQ_QUALITY_TOO_LOW; + } + } + + if (liq_progress(options, options->progress_stage1 + options->progress_stage2 + options->progress_stage3 * 0.95f)) { + pam_freecolormap(acolormap); + return LIQ_ABORTED; + } + + sort_palette(acolormap, options); + + // If palette was created from a multi-image histogram, + // then it shouldn't be optimized for one image during remapping + if (fixed_result_colors) { + for(unsigned int i=0; i < acolormap->colors; i++) { + acolormap->palette[i].fixed = true; + } + } + + liq_result *result = options->malloc(sizeof(liq_result)); + if (!result) return LIQ_OUT_OF_MEMORY; + *result = (liq_result){ + .magic_header = liq_result_magic, + .malloc = options->malloc, + .free = options->free, + .palette = acolormap, + .palette_error = palette_error, + .fast_palette = fast_palette, + .use_dither_map = options->use_dither_map, + .gamma = gamma, + .min_posterization_output = options->min_posterization_output, + }; + *result_output = result; + return LIQ_OK; +} + +LIQ_EXPORT LIQ_NONNULL liq_error liq_write_remapped_image(liq_result *result, liq_image *input_image, void *buffer, size_t buffer_size) +{ + if (!CHECK_STRUCT_TYPE(result, liq_result)) { + return LIQ_INVALID_POINTER; + } + if (!CHECK_STRUCT_TYPE(input_image, liq_image)) { + return LIQ_INVALID_POINTER; + } + if (!CHECK_USER_POINTER(buffer)) { + return LIQ_INVALID_POINTER; + } + + const size_t required_size = input_image->width * input_image->height; + if (buffer_size < required_size) { + return LIQ_BUFFER_TOO_SMALL; + } + + unsigned char *rows[input_image->height]; + unsigned char *buffer_bytes = buffer; + for(unsigned int i=0; i < input_image->height; i++) { + rows[i] = &buffer_bytes[input_image->width * i]; + } + return liq_write_remapped_image_rows(result, input_image, rows); +} + +LIQ_EXPORT LIQ_NONNULL liq_error liq_write_remapped_image_rows(liq_result *quant, liq_image *input_image, unsigned char **row_pointers) +{ + if (!CHECK_STRUCT_TYPE(quant, liq_result)) return LIQ_INVALID_POINTER; + if (!CHECK_STRUCT_TYPE(input_image, liq_image)) return LIQ_INVALID_POINTER; + for(unsigned int i=0; i < input_image->height; i++) { + if (!CHECK_USER_POINTER(row_pointers+i) || !CHECK_USER_POINTER(row_pointers[i])) return LIQ_INVALID_POINTER; + } + + if (quant->remapping) { + liq_remapping_result_destroy(quant->remapping); + } + liq_remapping_result *const result = quant->remapping = liq_remapping_result_create(quant); + if (!result) return LIQ_OUT_OF_MEMORY; + + if (!input_image->edges && !input_image->dither_map && quant->use_dither_map) { + contrast_maps(input_image); + } + + if (liq_remap_progress(result, result->progress_stage1 * 0.25f)) { + return LIQ_ABORTED; + } + + /* + ** Step 4: map the colors in the image to their closest match in the + ** new colormap, and write 'em out. + */ + + float remapping_error = result->palette_error; + if (result->dither_level == 0) { + set_rounded_palette(&result->int_palette, result->palette, result->gamma, quant->min_posterization_output); + remapping_error = remap_to_palette(input_image, row_pointers, result->palette, quant->fast_palette); + } else { + const bool generate_dither_map = result->use_dither_map && (input_image->edges && !input_image->dither_map); + if (generate_dither_map) { + // If dithering (with dither map) is required, this image is used to find areas that require dithering + remapping_error = remap_to_palette(input_image, row_pointers, result->palette, quant->fast_palette); + update_dither_map(row_pointers, input_image); + } + + if (liq_remap_progress(result, result->progress_stage1 * 0.5f)) { + return LIQ_ABORTED; + } + + // remapping above was the last chance to do voronoi iteration, hence the final palette is set after remapping + set_rounded_palette(&result->int_palette, result->palette, result->gamma, quant->min_posterization_output); + + if (!remap_to_palette_floyd(input_image, row_pointers, result, MAX(remapping_error*2.4, 16.f/256.f), generate_dither_map)) { + return LIQ_ABORTED; + } + } + + // remapping error from dithered image is absurd, so always non-dithered value is used + // palette_error includes some perceptual weighting from histogram which is closer correlated with dssim + // so that should be used when possible. + if (result->palette_error < 0) { + result->palette_error = remapping_error; + } + + return LIQ_OK; +} + +LIQ_EXPORT int liq_version() { + return LIQ_VERSION; +} diff --git a/libimagequant.h b/libimagequant.h new file mode 100644 index 0000000..ae503b1 --- /dev/null +++ b/libimagequant.h @@ -0,0 +1,137 @@ +/* + * https://pngquant.org + */ + +#ifndef LIBIMAGEQUANT_H +#define LIBIMAGEQUANT_H + +#ifdef IMAGEQUANT_EXPORTS +#define LIQ_EXPORT __declspec(dllexport) +#endif + +#ifndef LIQ_EXPORT +#define LIQ_EXPORT extern +#endif + +#define LIQ_VERSION 20800 +#define LIQ_VERSION_STRING "2.8.0" + +#ifndef LIQ_PRIVATE +#if defined(__GNUC__) || defined (__llvm__) +#define LIQ_PRIVATE __attribute__((visibility("hidden"))) +#define LIQ_NONNULL __attribute__((nonnull)) +#define LIQ_USERESULT __attribute__((warn_unused_result)) +#else +#define LIQ_PRIVATE +#define LIQ_NONNULL +#define LIQ_USERESULT +#endif +#endif + +#ifdef __cplusplus +extern "C" { +#endif + +#include + +typedef struct liq_attr liq_attr; +typedef struct liq_image liq_image; +typedef struct liq_result liq_result; +typedef struct liq_histogram liq_histogram; + +typedef struct liq_color { + unsigned char r, g, b, a; +} liq_color; + +typedef struct liq_palette { + unsigned int count; + liq_color entries[256]; +} liq_palette; + +typedef enum liq_error { + LIQ_OK = 0, + LIQ_QUALITY_TOO_LOW = 99, + LIQ_VALUE_OUT_OF_RANGE = 100, + LIQ_OUT_OF_MEMORY, + LIQ_ABORTED, + LIQ_BITMAP_NOT_AVAILABLE, + LIQ_BUFFER_TOO_SMALL, + LIQ_INVALID_POINTER, +} liq_error; + +enum liq_ownership {LIQ_OWN_ROWS=4, LIQ_OWN_PIXELS=8}; + +LIQ_EXPORT LIQ_USERESULT liq_attr* liq_attr_create(void); +LIQ_EXPORT LIQ_USERESULT liq_attr* liq_attr_create_with_allocator(void* (*malloc)(size_t), void (*free)(void*)); +LIQ_EXPORT LIQ_USERESULT liq_attr* liq_attr_copy(liq_attr *orig) LIQ_NONNULL; +LIQ_EXPORT void liq_attr_destroy(liq_attr *attr) LIQ_NONNULL; + +LIQ_EXPORT LIQ_USERESULT liq_histogram* liq_histogram_create(liq_attr* attr); +LIQ_EXPORT LIQ_USERESULT liq_error liq_histogram_add_image(liq_histogram *hist, liq_attr *attr, liq_image* image); +LIQ_EXPORT void liq_histogram_destroy(liq_histogram *hist) LIQ_NONNULL; + +LIQ_EXPORT liq_error liq_set_max_colors(liq_attr* attr, int colors) LIQ_NONNULL; +LIQ_EXPORT LIQ_USERESULT int liq_get_max_colors(const liq_attr* attr) LIQ_NONNULL; +LIQ_EXPORT liq_error liq_set_speed(liq_attr* attr, int speed) LIQ_NONNULL; +LIQ_EXPORT LIQ_USERESULT int liq_get_speed(const liq_attr* attr) LIQ_NONNULL; +LIQ_EXPORT liq_error liq_set_min_opacity(liq_attr* attr, int min) LIQ_NONNULL; +LIQ_EXPORT LIQ_USERESULT int liq_get_min_opacity(const liq_attr* attr) LIQ_NONNULL; +LIQ_EXPORT liq_error liq_set_min_posterization(liq_attr* attr, int bits) LIQ_NONNULL; +LIQ_EXPORT LIQ_USERESULT int liq_get_min_posterization(const liq_attr* attr) LIQ_NONNULL; +LIQ_EXPORT liq_error liq_set_quality(liq_attr* attr, int minimum, int maximum) LIQ_NONNULL; +LIQ_EXPORT LIQ_USERESULT int liq_get_min_quality(const liq_attr* attr) LIQ_NONNULL; +LIQ_EXPORT LIQ_USERESULT int liq_get_max_quality(const liq_attr* attr) LIQ_NONNULL; +LIQ_EXPORT void liq_set_last_index_transparent(liq_attr* attr, int is_last) LIQ_NONNULL; + +typedef void liq_log_callback_function(const liq_attr*, const char *message, void* user_info); +typedef void liq_log_flush_callback_function(const liq_attr*, void* user_info); +LIQ_EXPORT void liq_set_log_callback(liq_attr*, liq_log_callback_function*, void* user_info); +LIQ_EXPORT void liq_set_log_flush_callback(liq_attr*, liq_log_flush_callback_function*, void* user_info); + +typedef int liq_progress_callback_function(float progress_percent, void* user_info); +LIQ_EXPORT void liq_attr_set_progress_callback(liq_attr*, liq_progress_callback_function*, void* user_info); +LIQ_EXPORT void liq_result_set_progress_callback(liq_result*, liq_progress_callback_function*, void* user_info); + +// The rows and their data are not modified. The type of `rows` is non-const only due to a bug in C's typesystem design. +LIQ_EXPORT LIQ_USERESULT liq_image *liq_image_create_rgba_rows(const liq_attr *attr, void *const rows[], int width, int height, double gamma) LIQ_NONNULL; +LIQ_EXPORT LIQ_USERESULT liq_image *liq_image_create_rgba(const liq_attr *attr, const void *bitmap, int width, int height, double gamma) LIQ_NONNULL; + +typedef void liq_image_get_rgba_row_callback(liq_color row_out[], int row, int width, void* user_info); +LIQ_EXPORT LIQ_USERESULT liq_image *liq_image_create_custom(const liq_attr *attr, liq_image_get_rgba_row_callback *row_callback, void* user_info, int width, int height, double gamma); + +LIQ_EXPORT liq_error liq_image_set_memory_ownership(liq_image *image, int ownership_flags) LIQ_NONNULL; +LIQ_EXPORT liq_error liq_image_add_fixed_color(liq_image *img, liq_color color) LIQ_NONNULL; +LIQ_EXPORT LIQ_USERESULT int liq_image_get_width(const liq_image *img) LIQ_NONNULL; +LIQ_EXPORT LIQ_USERESULT int liq_image_get_height(const liq_image *img) LIQ_NONNULL; +LIQ_EXPORT void liq_image_destroy(liq_image *img) LIQ_NONNULL; + +LIQ_EXPORT LIQ_USERESULT liq_error liq_histogram_quantize(liq_histogram *const input_hist, liq_attr *const options, liq_result **result_output) LIQ_NONNULL; +LIQ_EXPORT LIQ_USERESULT liq_error liq_image_quantize(liq_image *const input_image, liq_attr *const options, liq_result **result_output) LIQ_NONNULL; + +LIQ_EXPORT liq_error liq_set_dithering_level(liq_result *res, float dither_level) LIQ_NONNULL; +LIQ_EXPORT liq_error liq_set_output_gamma(liq_result* res, double gamma) LIQ_NONNULL; +LIQ_EXPORT LIQ_USERESULT double liq_get_output_gamma(const liq_result *result) LIQ_NONNULL; + +LIQ_EXPORT LIQ_USERESULT const liq_palette *liq_get_palette(liq_result *result) LIQ_NONNULL; + +LIQ_EXPORT liq_error liq_write_remapped_image(liq_result *result, liq_image *input_image, void *buffer, size_t buffer_size) LIQ_NONNULL; +LIQ_EXPORT liq_error liq_write_remapped_image_rows(liq_result *result, liq_image *input_image, unsigned char **row_pointers) LIQ_NONNULL; + +LIQ_EXPORT double liq_get_quantization_error(liq_result *result) LIQ_NONNULL; +LIQ_EXPORT int liq_get_quantization_quality(liq_result *result) LIQ_NONNULL; +LIQ_EXPORT double liq_get_remapping_error(liq_result *result) LIQ_NONNULL; +LIQ_EXPORT int liq_get_remapping_quality(liq_result *result) LIQ_NONNULL; + +LIQ_EXPORT void liq_result_destroy(liq_result *) LIQ_NONNULL; + +LIQ_EXPORT int liq_version(void); + + +// Deprecated +LIQ_EXPORT LIQ_USERESULT liq_result *liq_quantize_image(liq_attr *options, liq_image *input_image) LIQ_NONNULL; + +#ifdef __cplusplus +} +#endif + +#endif diff --git a/mediancut.c b/mediancut.c new file mode 100644 index 0000000..9d93d49 --- /dev/null +++ b/mediancut.c @@ -0,0 +1,478 @@ +/* +** © 2009-2015 by Kornel Lesiński. +** +** This file is part of libimagequant. +** +** libimagequant 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. +** +** libimagequant 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 libimagequant. If not, see . +*/ +/* +** Copyright (C) 1989, 1991 by Jef Poskanzer. +** Copyright (C) 1997, 2000, 2002 by Greg Roelofs; based on an idea by +** Stefan Schneider. +** +** Permission to use, copy, modify, and distribute this software and its +** documentation for any purpose and without fee is hereby granted, provided +** that the above copyright notice appear in all copies and that both that +** copyright notice and this permission notice appear in supporting +** documentation. This software is provided "as is" without express or +** implied warranty. +*/ + +#include +#include + +#include "libimagequant.h" +#include "pam.h" +#include "mediancut.h" + +#define index_of_channel(ch) (offsetof(f_pixel,ch)/sizeof(float)) + +static f_pixel averagepixels(unsigned int clrs, const hist_item achv[], const f_pixel center); + +struct box { + f_pixel color; + f_pixel variance; + double sum, total_error, max_error; + unsigned int ind; + unsigned int colors; +}; + +ALWAYS_INLINE static double variance_diff(double val, const double good_enough); +inline static double variance_diff(double val, const double good_enough) +{ + val *= val; + if (val < good_enough*good_enough) return val*0.25; + return val; +} + +/** Weighted per-channel variance of the box. It's used to decide which channel to split by */ +static f_pixel box_variance(const hist_item achv[], const struct box *box) +{ + f_pixel mean = box->color; + double variancea=0, variancer=0, varianceg=0, varianceb=0; + + for(unsigned int i = 0; i < box->colors; ++i) { + f_pixel px = achv[box->ind + i].acolor; + double weight = achv[box->ind + i].adjusted_weight; + variancea += variance_diff(mean.a - px.a, 2.0/256.0)*weight; + variancer += variance_diff(mean.r - px.r, 1.0/256.0)*weight; + varianceg += variance_diff(mean.g - px.g, 1.0/256.0)*weight; + varianceb += variance_diff(mean.b - px.b, 1.0/256.0)*weight; + } + + return (f_pixel){ + .a = variancea*(4.0/16.0), + .r = variancer*(7.0/16.0), + .g = varianceg*(9.0/16.0), + .b = varianceb*(5.0/16.0), + }; +} + +static double box_max_error(const hist_item achv[], const struct box *box) +{ + f_pixel mean = box->color; + double max_error = 0; + + for(unsigned int i = 0; i < box->colors; ++i) { + const double diff = colordifference(mean, achv[box->ind + i].acolor); + if (diff > max_error) { + max_error = diff; + } + } + return max_error; +} + +ALWAYS_INLINE static double color_weight(f_pixel median, hist_item h); + +static inline void hist_item_swap(hist_item *l, hist_item *r) +{ + if (l != r) { + hist_item t = *l; + *l = *r; + *r = t; + } +} + +ALWAYS_INLINE static unsigned int qsort_pivot(const hist_item *const base, const unsigned int len); +inline static unsigned int qsort_pivot(const hist_item *const base, const unsigned int len) +{ + if (len < 32) { + return len/2; + } + + const unsigned int aidx=8, bidx=len/2, cidx=len-1; + const unsigned int a=base[aidx].tmp.sort_value, b=base[bidx].tmp.sort_value, c=base[cidx].tmp.sort_value; + return (a < b) ? ((b < c) ? bidx : ((a < c) ? cidx : aidx )) + : ((b > c) ? bidx : ((a < c) ? aidx : cidx )); +} + +ALWAYS_INLINE static unsigned int qsort_partition(hist_item *const base, const unsigned int len); +inline static unsigned int qsort_partition(hist_item *const base, const unsigned int len) +{ + unsigned int l = 1, r = len; + if (len >= 8) { + hist_item_swap(&base[0], &base[qsort_pivot(base,len)]); + } + + const unsigned int pivot_value = base[0].tmp.sort_value; + while (l < r) { + if (base[l].tmp.sort_value >= pivot_value) { + l++; + } else { + while(l < --r && base[r].tmp.sort_value <= pivot_value) {} + hist_item_swap(&base[l], &base[r]); + } + } + l--; + hist_item_swap(&base[0], &base[l]); + + return l; +} + +/** quick select algorithm */ +static void hist_item_sort_range(hist_item *base, unsigned int len, unsigned int sort_start) +{ + for(;;) { + const unsigned int l = qsort_partition(base, len), r = l+1; + + if (l > 0 && sort_start < l) { + len = l; + } + else if (r < len && sort_start > r) { + base += r; len -= r; sort_start -= r; + } + else break; + } +} + +/** sorts array to make sum of weights lower than halfvar one side, returns edge between halfvar parts of the set */ +static hist_item *hist_item_sort_halfvar(hist_item *base, unsigned int len, double *const lowervar, const double halfvar) +{ + do { + const unsigned int l = qsort_partition(base, len), r = l+1; + + // check if sum of left side is smaller than half, + // if it is, then it doesn't need to be sorted + unsigned int t = 0; double tmpsum = *lowervar; + while (t <= l && tmpsum < halfvar) tmpsum += base[t++].color_weight; + + if (tmpsum < halfvar) { + *lowervar = tmpsum; + } else { + if (l > 0) { + hist_item *res = hist_item_sort_halfvar(base, l, lowervar, halfvar); + if (res) return res; + } else { + // End of left recursion. This will be executed in order from the first element. + *lowervar += base[0].color_weight; + if (*lowervar > halfvar) return &base[0]; + } + } + + if (len > r) { + base += r; len -= r; // tail-recursive "call" + } else { + *lowervar += base[r].color_weight; + return (*lowervar > halfvar) ? &base[r] : NULL; + } + } while(1); +} + +static f_pixel get_median(const struct box *b, hist_item achv[]); + +typedef struct { + unsigned int chan; float variance; +} channelvariance; + +static int comparevariance(const void *ch1, const void *ch2) +{ + return ((const channelvariance*)ch1)->variance > ((const channelvariance*)ch2)->variance ? -1 : + (((const channelvariance*)ch1)->variance < ((const channelvariance*)ch2)->variance ? 1 : 0); +} + +/** Finds which channels need to be sorted first and preproceses achv for fast sort */ +static double prepare_sort(struct box *b, hist_item achv[]) +{ + /* + ** Sort dimensions by their variance, and then sort colors first by dimension with highest variance + */ + channelvariance channels[4] = { + {index_of_channel(r), b->variance.r}, + {index_of_channel(g), b->variance.g}, + {index_of_channel(b), b->variance.b}, + {index_of_channel(a), b->variance.a}, + }; + + qsort(channels, 4, sizeof(channels[0]), comparevariance); + + for(unsigned int i=0; i < b->colors; i++) { + const float *chans = (const float *)&achv[b->ind + i].acolor; + // Only the first channel really matters. When trying median cut many times + // with different histogram weights, I don't want sort randomness to influence outcome. + achv[b->ind + i].tmp.sort_value = ((unsigned int)(chans[channels[0].chan]*65535.0)<<16) | + (unsigned int)((chans[channels[2].chan] + chans[channels[1].chan]/2.0 + chans[channels[3].chan]/4.0)*65535.0); + } + + const f_pixel median = get_median(b, achv); + + // box will be split to make color_weight of each side even + const unsigned int ind = b->ind, end = ind+b->colors; + double totalvar = 0; + for(unsigned int j=ind; j < end; j++) totalvar += (achv[j].color_weight = color_weight(median, achv[j])); + return totalvar / 2.0; +} + +/** finds median in unsorted set by sorting only minimum required */ +static f_pixel get_median(const struct box *b, hist_item achv[]) +{ + const unsigned int median_start = (b->colors-1)/2; + + hist_item_sort_range(&(achv[b->ind]), b->colors, + median_start); + + if (b->colors&1) return achv[b->ind + median_start].acolor; + + // technically the second color is not guaranteed to be sorted correctly + // but most of the time it is good enough to be useful + return averagepixels(2, &achv[b->ind + median_start], (f_pixel){0.5,0.5,0.5,0.5}); +} + +/* + ** Find the best splittable box. -1 if no boxes are splittable. + */ +static int best_splittable_box(struct box* bv, unsigned int boxes, const double max_mse) +{ + int bi=-1; double maxsum=0; + for(unsigned int i=0; i < boxes; i++) { + if (bv[i].colors < 2) { + continue; + } + + // looks only at max variance, because it's only going to split by it + const double cv = MAX(bv[i].variance.r, MAX(bv[i].variance.g,bv[i].variance.b)); + double thissum = bv[i].sum * MAX(bv[i].variance.a, cv); + + if (bv[i].max_error > max_mse) { + thissum = thissum* bv[i].max_error/max_mse; + } + + if (thissum > maxsum) { + maxsum = thissum; + bi = i; + } + } + return bi; +} + +inline static double color_weight(f_pixel median, hist_item h) +{ + float diff = colordifference(median, h.acolor); + // if color is "good enough", don't split further + if (diff < 1.f/256.f/256.f) diff /= 2.f; + return sqrt(diff) * (sqrt(1.0+h.adjusted_weight)-1.0); +} + +static void set_colormap_from_boxes(colormap *map, struct box* bv, unsigned int boxes, hist_item *achv); +static void adjust_histogram(hist_item *achv, const colormap *map, const struct box* bv, unsigned int boxes); + +static double box_error(const struct box *box, const hist_item achv[]) +{ + f_pixel avg = box->color; + + double total_error=0; + for (unsigned int i = 0; i < box->colors; ++i) { + total_error += colordifference(avg, achv[box->ind + i].acolor) * achv[box->ind + i].perceptual_weight; + } + + return total_error; +} + + +static bool total_box_error_below_target(double target_mse, struct box bv[], unsigned int boxes, const histogram *hist) +{ + target_mse *= hist->total_perceptual_weight; + double total_error=0; + + for(unsigned int i=0; i < boxes; i++) { + // error is (re)calculated lazily + if (bv[i].total_error >= 0) { + total_error += bv[i].total_error; + } + if (total_error > target_mse) return false; + } + + for(unsigned int i=0; i < boxes; i++) { + if (bv[i].total_error < 0) { + bv[i].total_error = box_error(&bv[i], hist->achv); + total_error += bv[i].total_error; + } + if (total_error > target_mse) return false; + } + + return true; +} + +/* + ** Here is the fun part, the median-cut colormap generator. This is based + ** on Paul Heckbert's paper, "Color Image Quantization for Frame Buffer + ** Display," SIGGRAPH 1982 Proceedings, page 297. + */ +LIQ_PRIVATE colormap *mediancut(histogram *hist, unsigned int newcolors, const double target_mse, const double max_mse, void* (*malloc)(size_t), void (*free)(void*)) +{ + hist_item *achv = hist->achv; + struct box bv[newcolors]; + + /* + ** Set up the initial box. + */ + bv[0].ind = 0; + bv[0].colors = hist->size; + bv[0].color = averagepixels(bv[0].colors, &achv[bv[0].ind], (f_pixel){0.5,0.5,0.5,0.5}); + bv[0].variance = box_variance(achv, &bv[0]); + bv[0].max_error = box_max_error(achv, &bv[0]); + bv[0].sum = 0; + bv[0].total_error = -1; + for(unsigned int i=0; i < bv[0].colors; i++) bv[0].sum += achv[i].adjusted_weight; + + unsigned int boxes = 1; + + /* + ** Main loop: split boxes until we have enough. + */ + while (boxes < newcolors) { + + // first splits boxes that exceed quality limit (to have colors for things like odd green pixel), + // later raises the limit to allow large smooth areas/gradients get colors. + const double current_max_mse = max_mse + (boxes/(double)newcolors)*16.0*max_mse; + const int bi = best_splittable_box(bv, boxes, current_max_mse); + if (bi < 0) + break; /* ran out of colors! */ + + unsigned int indx = bv[bi].ind; + unsigned int clrs = bv[bi].colors; + + /* + Classic implementation tries to get even number of colors or pixels in each subdivision. + + Here, instead of popularity I use (sqrt(popularity)*variance) metric. + Each subdivision balances number of pixels (popular colors) and low variance - + boxes can be large if they have similar colors. Later boxes with high variance + will be more likely to be split. + + Median used as expected value gives much better results than mean. + */ + + const double halfvar = prepare_sort(&bv[bi], achv); + double lowervar=0; + + // hist_item_sort_halfvar sorts and sums lowervar at the same time + // returns item to break at …minus one, which does smell like an off-by-one error. + hist_item *break_p = hist_item_sort_halfvar(&achv[indx], clrs, &lowervar, halfvar); + unsigned int break_at = MIN(clrs-1, break_p - &achv[indx] + 1); + + /* + ** Split the box. + */ + double sm = bv[bi].sum; + double lowersum = 0; + for(unsigned int i=0; i < break_at; i++) lowersum += achv[indx + i].adjusted_weight; + + const f_pixel previous_center = bv[bi].color; + bv[bi].colors = break_at; + bv[bi].sum = lowersum; + bv[bi].color = averagepixels(bv[bi].colors, &achv[bv[bi].ind], previous_center); + bv[bi].total_error = -1; + bv[bi].variance = box_variance(achv, &bv[bi]); + bv[bi].max_error = box_max_error(achv, &bv[bi]); + bv[boxes].ind = indx + break_at; + bv[boxes].colors = clrs - break_at; + bv[boxes].sum = sm - lowersum; + bv[boxes].color = averagepixels(bv[boxes].colors, &achv[bv[boxes].ind], previous_center); + bv[boxes].total_error = -1; + bv[boxes].variance = box_variance(achv, &bv[boxes]); + bv[boxes].max_error = box_max_error(achv, &bv[boxes]); + + ++boxes; + + if (total_box_error_below_target(target_mse, bv, boxes, hist)) { + break; + } + } + + colormap *map = pam_colormap(boxes, malloc, free); + set_colormap_from_boxes(map, bv, boxes, achv); + + adjust_histogram(achv, map, bv, boxes); + + return map; +} + +static void set_colormap_from_boxes(colormap *map, struct box* bv, unsigned int boxes, hist_item *achv) +{ + /* + ** Ok, we've got enough boxes. Now choose a representative color for + ** each box. There are a number of possible ways to make this choice. + ** One would be to choose the center of the box; this ignores any structure + ** within the boxes. Another method would be to average all the colors in + ** the box - this is the method specified in Heckbert's paper. + */ + + for(unsigned int bi = 0; bi < boxes; ++bi) { + map->palette[bi].acolor = bv[bi].color; + + /* store total color popularity (perceptual_weight is approximation of it) */ + map->palette[bi].popularity = 0; + for(unsigned int i=bv[bi].ind; i < bv[bi].ind+bv[bi].colors; i++) { + map->palette[bi].popularity += achv[i].perceptual_weight; + } + } +} + +/* increase histogram popularity by difference from the final color (this is used as part of feedback loop) */ +static void adjust_histogram(hist_item *achv, const colormap *map, const struct box* bv, unsigned int boxes) +{ + for(unsigned int bi = 0; bi < boxes; ++bi) { + for(unsigned int i=bv[bi].ind; i < bv[bi].ind+bv[bi].colors; i++) { + achv[i].adjusted_weight *= sqrt(1.0 +colordifference(map->palette[bi].acolor, achv[i].acolor)/2.0); + achv[i].tmp.likely_colormap_index = bi; + } + } +} + +static f_pixel averagepixels(unsigned int clrs, const hist_item achv[], f_pixel center) +{ + double r = 0, g = 0, b = 0, a = 0, sum = 0; + + for(unsigned int i = 0; i < clrs; i++) { + const f_pixel px = achv[i].acolor; + const double weight = achv[i].adjusted_weight; + + sum += weight; + r += px.r * weight; + g += px.g * weight; + b += px.b * weight; + a += px.a * weight; + } + + if (sum) { + a /= sum; + r /= sum; + g /= sum; + b /= sum; + } + + assert(!isnan(r) && !isnan(g) && !isnan(b) && !isnan(a)); + + return (f_pixel){.r=r, .g=g, .b=b, .a=a}; +} diff --git a/mediancut.h b/mediancut.h new file mode 100644 index 0000000..d97696c --- /dev/null +++ b/mediancut.h @@ -0,0 +1,2 @@ + +LIQ_PRIVATE colormap *mediancut(histogram *hist, unsigned int newcolors, const double target_mse, const double max_mse, void* (*malloc)(size_t), void (*free)(void*)); diff --git a/mempool.c b/mempool.c new file mode 100644 index 0000000..cd345a7 --- /dev/null +++ b/mempool.c @@ -0,0 +1,81 @@ +/* +© 2011-2016 by Kornel Lesiński. + +This file is part of libimagequant. + +libimagequant 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. + +libimagequant 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 libimagequant. If not, see . +*/ + +#include "libimagequant.h" +#include "mempool.h" +#include +#include +#include + +#define ALIGN_MASK 15UL +#define MEMPOOL_RESERVED ((sizeof(struct mempool)+ALIGN_MASK) & ~ALIGN_MASK) + +struct mempool { + unsigned int used, size; + void* (*malloc)(size_t); + void (*free)(void*); + struct mempool *next; +}; +LIQ_PRIVATE void* mempool_create(mempool *mptr, const unsigned int size, unsigned int max_size, void* (*malloc)(size_t), void (*free)(void*)) +{ + if (*mptr && ((*mptr)->used+size) <= (*mptr)->size) { + unsigned int prevused = (*mptr)->used; + (*mptr)->used += (size+15UL) & ~0xFUL; + return ((char*)(*mptr)) + prevused; + } + + mempool old = *mptr; + if (!max_size) max_size = (1<<17); + max_size = size+ALIGN_MASK > max_size ? size+ALIGN_MASK : max_size; + + *mptr = malloc(MEMPOOL_RESERVED + max_size); + if (!*mptr) return NULL; + **mptr = (struct mempool){ + .malloc = malloc, + .free = free, + .size = MEMPOOL_RESERVED + max_size, + .used = sizeof(struct mempool), + .next = old, + }; + uintptr_t mptr_used_start = (uintptr_t)(*mptr) + (*mptr)->used; + (*mptr)->used += (ALIGN_MASK + 1 - (mptr_used_start & ALIGN_MASK)) & ALIGN_MASK; // reserve bytes required to make subsequent allocations aligned + assert(!(((uintptr_t)(*mptr) + (*mptr)->used) & ALIGN_MASK)); + + return mempool_alloc(mptr, size, size); +} + +LIQ_PRIVATE void* mempool_alloc(mempool *mptr, const unsigned int size, const unsigned int max_size) +{ + if (((*mptr)->used+size) <= (*mptr)->size) { + unsigned int prevused = (*mptr)->used; + (*mptr)->used += (size + ALIGN_MASK) & ~ALIGN_MASK; + return ((char*)(*mptr)) + prevused; + } + + return mempool_create(mptr, size, max_size, (*mptr)->malloc, (*mptr)->free); +} + +LIQ_PRIVATE void mempool_destroy(mempool m) +{ + while (m) { + mempool next = m->next; + m->free(m); + m = next; + } +} diff --git a/mempool.h b/mempool.h new file mode 100644 index 0000000..0797e7c --- /dev/null +++ b/mempool.h @@ -0,0 +1,13 @@ +#ifndef MEMPOOL_H +#define MEMPOOL_H + +#include + +struct mempool; +typedef struct mempool *mempool; + +LIQ_PRIVATE void* mempool_create(mempool *mptr, const unsigned int size, unsigned int capacity, void* (*malloc)(size_t), void (*free)(void*)); +LIQ_PRIVATE void* mempool_alloc(mempool *mptr, const unsigned int size, const unsigned int capacity); +LIQ_PRIVATE void mempool_destroy(mempool m); + +#endif diff --git a/nearest.c b/nearest.c new file mode 100644 index 0000000..448ab3a --- /dev/null +++ b/nearest.c @@ -0,0 +1,206 @@ +/* +© 2011-2015 by Kornel Lesiński. + +This file is part of libimagequant. + +libimagequant 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. + +libimagequant 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 libimagequant. If not, see . +*/ + +#include "libimagequant.h" +#include "pam.h" +#include "nearest.h" +#include "mempool.h" +#include + +typedef struct vp_sort_tmp { + float distance_squared; + unsigned int idx; +} vp_sort_tmp; + +typedef struct vp_search_tmp { + float distance; + unsigned int idx; + int exclude; +} vp_search_tmp; + +typedef struct vp_node { + struct vp_node *near, *far; + f_pixel vantage_point; + float radius; + unsigned int idx; +} vp_node; + +struct nearest_map { + vp_node *root; + const colormap_item *palette; + float nearest_other_color_dist[256]; + mempool mempool; +}; + +static void vp_search_node(const vp_node *node, const f_pixel *const needle, vp_search_tmp *const best_candidate); + +static int vp_compare_distance(const void *ap, const void *bp) { + float a = ((const vp_sort_tmp*)ap)->distance_squared; + float b = ((const vp_sort_tmp*)bp)->distance_squared; + return a > b ? 1 : -1; +} + +static void vp_sort_indexes_by_distance(const f_pixel vantage_point, vp_sort_tmp *indexes, int num_indexes, const colormap_item items[]) { + for(int i=0; i < num_indexes; i++) { + indexes[i].distance_squared = colordifference(vantage_point, items[indexes[i].idx].acolor); + } + qsort(indexes, num_indexes, sizeof(indexes[0]), vp_compare_distance); +} + +/* + * Usually it should pick farthest point, but picking most popular point seems to make search quicker anyway + */ +static int vp_find_best_vantage_point_index(vp_sort_tmp *indexes, int num_indexes, const colormap_item items[]) { + int best = 0; + float best_popularity = items[indexes[0].idx].popularity; + for(int i = 1; i < num_indexes; i++) { + if (items[indexes[i].idx].popularity > best_popularity) { + best_popularity = items[indexes[i].idx].popularity; + best = i; + } + } + return best; +} + +static vp_node *vp_create_node(mempool *m, vp_sort_tmp *indexes, int num_indexes, const colormap_item items[]) { + if (num_indexes <= 0) { + return NULL; + } + + vp_node *node = mempool_alloc(m, sizeof(node[0]), 0); + + if (num_indexes == 1) { + *node = (vp_node){ + .vantage_point = items[indexes[0].idx].acolor, + .idx = indexes[0].idx, + .radius = MAX_DIFF, + }; + return node; + } + + const int ref = vp_find_best_vantage_point_index(indexes, num_indexes, items); + const int ref_idx = indexes[ref].idx; + + // Removes the `ref_idx` item from remaining items, because it's included in the current node + num_indexes -= 1; + indexes[ref] = indexes[num_indexes]; + + vp_sort_indexes_by_distance(items[ref_idx].acolor, indexes, num_indexes, items); + + // Remaining items are split by the median distance + const int half_idx = num_indexes/2; + + *node = (vp_node){ + .vantage_point = items[ref_idx].acolor, + .idx = ref_idx, + .radius = sqrtf(indexes[half_idx].distance_squared), + }; + node->near = vp_create_node(m, indexes, half_idx, items); + node->far = vp_create_node(m, &indexes[half_idx], num_indexes - half_idx, items); + + return node; +} + +LIQ_PRIVATE struct nearest_map *nearest_init(const colormap *map, const bool fast) { + mempool m = NULL; + struct nearest_map *handle = mempool_create(&m, sizeof(handle[0]), sizeof(handle[0]) + sizeof(vp_node)*map->colors+16, map->malloc, map->free); + + vp_sort_tmp indexes[map->colors]; + + for(unsigned int i=0; i < map->colors; i++) { + indexes[i].idx = i; + } + + vp_node *root = vp_create_node(&m, indexes, map->colors, map->palette); + *handle = (struct nearest_map){ + .root = root, + .palette = map->palette, + .mempool = m, + }; + + for(unsigned int i=0; i < map->colors; i++) { + vp_search_tmp best = { + .distance = MAX_DIFF, + .exclude = i, + }; + vp_search_node(root, &map->palette[i].acolor, &best); + handle->nearest_other_color_dist[i] = best.distance * best.distance / 4.0; // half of squared distance + } + + return handle; +} + +static void vp_search_node(const vp_node *node, const f_pixel *const needle, vp_search_tmp *const best_candidate) { + do { + const float distance = sqrtf(colordifference(node->vantage_point, *needle)); + + if (distance < best_candidate->distance && best_candidate->exclude != node->idx) { + best_candidate->distance = distance; + best_candidate->idx = node->idx; + } + + // Recurse towards most likely candidate first to narrow best candidate's distance as soon as possible + if (distance < node->radius) { + if (node->near) { + vp_search_node(node->near, needle, best_candidate); + } + // The best node (final answer) may be just ouside the radius, but not farther than + // the best distance we know so far. The vp_search_node above should have narrowed + // best_candidate->distance, so this path is rarely taken. + if (node->far && distance >= node->radius - best_candidate->distance) { + node = node->far; // Fast tail recursion + } else { + break; + } + } else { + if (node->far) { + vp_search_node(node->far, needle, best_candidate); + } + if (node->near && distance <= node->radius + best_candidate->distance) { + node = node->near; // Fast tail recursion + } else { + break; + } + } + } while(true); +} + +LIQ_PRIVATE unsigned int nearest_search(const struct nearest_map *handle, const f_pixel *px, const int likely_colormap_index, float *diff) { + const float guess_diff = colordifference(handle->palette[likely_colormap_index].acolor, *px); + if (guess_diff < handle->nearest_other_color_dist[likely_colormap_index]) { + if (diff) *diff = guess_diff; + return likely_colormap_index; + } + + vp_search_tmp best_candidate = { + .distance = sqrtf(guess_diff), + .idx = likely_colormap_index, + .exclude = -1, + }; + vp_search_node(handle->root, px, &best_candidate); + if (diff) { + *diff = best_candidate.distance * best_candidate.distance; + } + return best_candidate.idx; +} + +LIQ_PRIVATE void nearest_free(struct nearest_map *centroids) +{ + mempool_destroy(centroids->mempool); +} diff --git a/nearest.h b/nearest.h new file mode 100644 index 0000000..0a98ca6 --- /dev/null +++ b/nearest.h @@ -0,0 +1,8 @@ +// +// nearest.h +// pngquant +// +struct nearest_map; +LIQ_PRIVATE struct nearest_map *nearest_init(const colormap *palette, const bool fast); +LIQ_PRIVATE unsigned int nearest_search(const struct nearest_map *map, const f_pixel *px, const int palette_index_guess, float *diff); +LIQ_PRIVATE void nearest_free(struct nearest_map *map); diff --git a/pam.c b/pam.c new file mode 100644 index 0000000..aae2927 --- /dev/null +++ b/pam.c @@ -0,0 +1,276 @@ +/* pam.c - pam (portable alpha map) utility library +** +** Copyright (C) 1989, 1991 by Jef Poskanzer. +** Copyright (C) 1997, 2000, 2002 by Greg Roelofs; based on an idea by +** Stefan Schneider. +** © 2009-2016 by Kornel Lesinski. +** +** Permission to use, copy, modify, and distribute this software and its +** documentation for any purpose and without fee is hereby granted, provided +** that the above copyright notice appear in all copies and that both that +** copyright notice and this permission notice appear in supporting +** documentation. This software is provided "as is" without express or +** implied warranty. +*/ + +#include +#include + +#include "libimagequant.h" +#include "pam.h" +#include "mempool.h" + +LIQ_PRIVATE bool pam_computeacolorhash(struct acolorhash_table *acht, const rgba_pixel *const pixels[], unsigned int cols, unsigned int rows, const unsigned char *importance_map) +{ + const unsigned int maxacolors = acht->maxcolors, ignorebits = acht->ignorebits; + const unsigned int channel_mask = 255U>>ignorebits<>ignorebits) ^ 0xFFU; + const unsigned int posterize_mask = channel_mask << 24 | channel_mask << 16 | channel_mask << 8 | channel_mask; + const unsigned int posterize_high_mask = channel_hmask << 24 | channel_hmask << 16 | channel_hmask << 8 | channel_hmask; + struct acolorhist_arr_head *const buckets = acht->buckets; + + unsigned int colors = acht->colors; + const unsigned int hash_size = acht->hash_size; + + const unsigned int stacksize = sizeof(acht->freestack)/sizeof(acht->freestack[0]); + struct acolorhist_arr_item **freestack = acht->freestack; + unsigned int freestackp=acht->freestackp; + + /* Go through the entire image, building a hash table of colors. */ + for(unsigned int row = 0; row < rows; ++row) { + + float boost=1.0; + for(unsigned int col = 0; col < cols; ++col) { + if (importance_map) { + boost = 0.5f+ (double)*importance_map++/255.f; + } + + // RGBA color is casted to long for easier hasing/comparisons + union rgba_as_int px = {pixels[row][col]}; + unsigned int hash; + if (!px.rgba.a) { + // "dirty alpha" has different RGBA values that end up being the same fully transparent color + px.l=0; hash=0; + boost = 10; + } else { + // mask posterizes all 4 channels in one go + px.l = (px.l & posterize_mask) | ((px.l & posterize_high_mask) >> (8-ignorebits)); + // fancier hashing algorithms didn't improve much + hash = px.l % hash_size; + } + + /* head of the hash function stores first 2 colors inline (achl->used = 1..2), + to reduce number of allocations of achl->other_items. + */ + struct acolorhist_arr_head *achl = &buckets[hash]; + if (achl->inline1.color.l == px.l && achl->used) { + achl->inline1.perceptual_weight += boost; + continue; + } + if (achl->used) { + if (achl->used > 1) { + if (achl->inline2.color.l == px.l) { + achl->inline2.perceptual_weight += boost; + continue; + } + // other items are stored as an array (which gets reallocated if needed) + struct acolorhist_arr_item *other_items = achl->other_items; + unsigned int i = 0; + for (; i < achl->used-2; i++) { + if (other_items[i].color.l == px.l) { + other_items[i].perceptual_weight += boost; + goto continue_outer_loop; + } + } + + // the array was allocated with spare items + if (i < achl->capacity) { + other_items[i] = (struct acolorhist_arr_item){ + .color = px, + .perceptual_weight = boost, + }; + achl->used++; + ++colors; + continue; + } + + if (++colors > maxacolors) { + acht->colors = colors; + acht->freestackp = freestackp; + return false; + } + + struct acolorhist_arr_item *new_items; + unsigned int capacity; + if (!other_items) { // there was no array previously, alloc "small" array + capacity = 8; + if (freestackp <= 0) { + // estimate how many colors are going to be + headroom + const size_t mempool_size = ((acht->rows + rows-row) * 2 * colors / (acht->rows + row + 1) + 1024) * sizeof(struct acolorhist_arr_item); + new_items = mempool_alloc(&acht->mempool, sizeof(struct acolorhist_arr_item)*capacity, mempool_size); + } else { + // freestack stores previously freed (reallocated) arrays that can be reused + // (all pesimistically assumed to be capacity = 8) + new_items = freestack[--freestackp]; + } + } else { + // simply reallocs and copies array to larger capacity + capacity = achl->capacity*2 + 16; + if (freestackp < stacksize-1) { + freestack[freestackp++] = other_items; + } + const size_t mempool_size = ((acht->rows + rows-row) * 2 * colors / (acht->rows + row + 1) + 32*capacity) * sizeof(struct acolorhist_arr_item); + new_items = mempool_alloc(&acht->mempool, sizeof(struct acolorhist_arr_item)*capacity, mempool_size); + if (!new_items) return false; + memcpy(new_items, other_items, sizeof(other_items[0])*achl->capacity); + } + + achl->other_items = new_items; + achl->capacity = capacity; + new_items[i] = (struct acolorhist_arr_item){ + .color = px, + .perceptual_weight = boost, + }; + achl->used++; + } else { + // these are elses for first checks whether first and second inline-stored colors are used + achl->inline2.color.l = px.l; + achl->inline2.perceptual_weight = boost; + achl->used = 2; + ++colors; + } + } else { + achl->inline1.color.l = px.l; + achl->inline1.perceptual_weight = boost; + achl->used = 1; + ++colors; + } + + continue_outer_loop:; + } + + } + acht->colors = colors; + acht->cols = cols; + acht->rows += rows; + acht->freestackp = freestackp; + return true; +} + +LIQ_PRIVATE struct acolorhash_table *pam_allocacolorhash(unsigned int maxcolors, unsigned int surface, unsigned int ignorebits, void* (*malloc)(size_t), void (*free)(void*)) +{ + const size_t estimated_colors = MIN(maxcolors, surface/(ignorebits + (surface > 512*512 ? 6 : 5))); + const size_t hash_size = estimated_colors < 66000 ? 6673 : (estimated_colors < 200000 ? 12011 : 24019); + + mempool m = NULL; + const size_t buckets_size = hash_size * sizeof(struct acolorhist_arr_head); + const size_t mempool_size = sizeof(struct acolorhash_table) + buckets_size + estimated_colors * sizeof(struct acolorhist_arr_item); + struct acolorhash_table *t = mempool_create(&m, sizeof(*t) + buckets_size, mempool_size, malloc, free); + if (!t) return NULL; + *t = (struct acolorhash_table){ + .mempool = m, + .hash_size = hash_size, + .maxcolors = maxcolors, + .ignorebits = ignorebits, + }; + memset(t->buckets, 0, buckets_size); + return t; +} + +#define PAM_ADD_TO_HIST(entry) { \ + hist->achv[j].acolor = to_f(gamma_lut, entry.color.rgba); \ + total_weight += hist->achv[j].adjusted_weight = hist->achv[j].perceptual_weight = MIN(entry.perceptual_weight, max_perceptual_weight); \ + ++j; \ +} + +LIQ_PRIVATE histogram *pam_acolorhashtoacolorhist(const struct acolorhash_table *acht, const double gamma, void* (*malloc)(size_t), void (*free)(void*)) +{ + histogram *hist = malloc(sizeof(hist[0])); + if (!hist || !acht) return NULL; + *hist = (histogram){ + .achv = malloc(MAX(1,acht->colors) * sizeof(hist->achv[0])), + .size = acht->colors, + .free = free, + .ignorebits = acht->ignorebits, + }; + if (!hist->achv) return NULL; + + float gamma_lut[256]; + to_f_set_gamma(gamma_lut, gamma); + + /* Limit perceptual weight to 1/10th of the image surface area to prevent + a single color from dominating all others. */ + float max_perceptual_weight = 0.1f * acht->cols * acht->rows; + double total_weight = 0; + + for(unsigned int j=0, i=0; i < acht->hash_size; ++i) { + const struct acolorhist_arr_head *const achl = &acht->buckets[i]; + if (achl->used) { + PAM_ADD_TO_HIST(achl->inline1); + + if (achl->used > 1) { + PAM_ADD_TO_HIST(achl->inline2); + + for(unsigned int k=0; k < achl->used-2; k++) { + PAM_ADD_TO_HIST(achl->other_items[k]); + } + } + } + } + + hist->total_perceptual_weight = total_weight; + return hist; +} + + +LIQ_PRIVATE void pam_freeacolorhash(struct acolorhash_table *acht) +{ + if (acht) { + mempool_destroy(acht->mempool); + } +} + +LIQ_PRIVATE void pam_freeacolorhist(histogram *hist) +{ + hist->free(hist->achv); + hist->free(hist); +} + +LIQ_PRIVATE colormap *pam_colormap(unsigned int colors, void* (*malloc)(size_t), void (*free)(void*)) +{ + assert(colors > 0 && colors < 65536); + + colormap *map; + const size_t colors_size = colors * sizeof(map->palette[0]); + map = malloc(sizeof(colormap) + colors_size); + if (!map) return NULL; + *map = (colormap){ + .malloc = malloc, + .free = free, + .colors = colors, + }; + memset(map->palette, 0, colors_size); + return map; +} + +LIQ_PRIVATE colormap *pam_duplicate_colormap(colormap *map) +{ + colormap *dupe = pam_colormap(map->colors, map->malloc, map->free); + for(unsigned int i=0; i < map->colors; i++) { + dupe->palette[i] = map->palette[i]; + } + return dupe; +} + +LIQ_PRIVATE void pam_freecolormap(colormap *c) +{ + c->free(c); +} + +LIQ_PRIVATE void to_f_set_gamma(float gamma_lut[], const double gamma) +{ + for(int i=0; i < 256; i++) { + gamma_lut[i] = pow((double)i/255.0, internal_gamma/gamma); + } +} + diff --git a/pam.h b/pam.h new file mode 100644 index 0000000..3af703e --- /dev/null +++ b/pam.h @@ -0,0 +1,270 @@ +/* pam.h - pam (portable alpha map) utility library + ** + ** Colormap routines. + ** + ** Copyright (C) 1989, 1991 by Jef Poskanzer. + ** Copyright (C) 1997 by Greg Roelofs. + ** + ** Permission to use, copy, modify, and distribute this software and its + ** documentation for any purpose and without fee is hereby granted, provided + ** that the above copyright notice appear in all copies and that both that + ** copyright notice and this permission notice appear in supporting + ** documentation. This software is provided "as is" without express or + ** implied warranty. + */ + +#ifndef PAM_H +#define PAM_H + +#include +#include +#include +#include + +#ifndef MAX +# define MAX(a,b) ((a) > (b)? (a) : (b)) +# define MIN(a,b) ((a) < (b)? (a) : (b)) +#endif + +#define MAX_DIFF 1e20 + +#ifndef USE_SSE +# if defined(__SSE__) && (defined(__amd64__) || defined(__X86_64__) || defined(_WIN64) || defined(WIN32) || defined(__WIN32__)) +# define USE_SSE 1 +# else +# define USE_SSE 0 +# endif +#endif + +#if USE_SSE +# include +# ifdef _MSC_VER +# include +# define SSE_ALIGN +# else +# define SSE_ALIGN __attribute__ ((aligned (16))) +# if defined(__i386__) && defined(__PIC__) +# define cpuid(func,ax,bx,cx,dx)\ + __asm__ __volatile__ ( \ + "push %%ebx\n" \ + "cpuid\n" \ + "mov %%ebx, %1\n" \ + "pop %%ebx\n" \ + : "=a" (ax), "=r" (bx), "=c" (cx), "=d" (dx) \ + : "a" (func)); +# else +# define cpuid(func,ax,bx,cx,dx)\ + __asm__ __volatile__ ("cpuid":\ + "=a" (ax), "=b" (bx), "=c" (cx), "=d" (dx) : "a" (func)); +# endif +#endif +#else +# define SSE_ALIGN +#endif + +#if defined(__GNUC__) || defined (__llvm__) +#define ALWAYS_INLINE __attribute__((always_inline)) inline +#define NEVER_INLINE __attribute__ ((noinline)) +#elif defined(_MSC_VER) +#define inline __inline +#define restrict __restrict +#define ALWAYS_INLINE __forceinline +#define NEVER_INLINE __declspec(noinline) +#else +#define ALWAYS_INLINE inline +#define NEVER_INLINE +#endif + +/* from pam.h */ + +typedef struct { + unsigned char r, g, b, a; +} rgba_pixel; + +typedef struct { + float a, r, g, b; +} SSE_ALIGN f_pixel; + +static const double internal_gamma = 0.5499; + +LIQ_PRIVATE void to_f_set_gamma(float gamma_lut[], const double gamma); + +/** + Converts 8-bit color to internal gamma and premultiplied alpha. + (premultiplied color space is much better for blending of semitransparent colors) + */ +ALWAYS_INLINE static f_pixel to_f(const float gamma_lut[], const rgba_pixel px); +inline static f_pixel to_f(const float gamma_lut[], const rgba_pixel px) +{ + float a = px.a/255.f; + + return (f_pixel) { + .a = a, + .r = gamma_lut[px.r]*a, + .g = gamma_lut[px.g]*a, + .b = gamma_lut[px.b]*a, + }; +} + +inline static rgba_pixel to_rgb(const float gamma, const f_pixel px) +{ + if (px.a < 1.f/256.f) { + return (rgba_pixel){0,0,0,0}; + } + + float r = px.r / px.a, + g = px.g / px.a, + b = px.b / px.a, + a = px.a; + + r = powf(r, gamma/internal_gamma); + g = powf(g, gamma/internal_gamma); + b = powf(b, gamma/internal_gamma); + + // 256, because numbers are in range 1..255.9999… rounded down + r *= 256.f; + g *= 256.f; + b *= 256.f; + a *= 256.f; + + return (rgba_pixel){ + .r = r>=255.f ? 255 : r, + .g = g>=255.f ? 255 : g, + .b = b>=255.f ? 255 : b, + .a = a>=255.f ? 255 : a, + }; +} + +ALWAYS_INLINE static double colordifference_ch(const double x, const double y, const double alphas); +inline static double colordifference_ch(const double x, const double y, const double alphas) +{ + // maximum of channel blended on white, and blended on black + // premultiplied alpha and backgrounds 0/1 shorten the formula + const double black = x-y, white = black+alphas; + return MAX(black*black, white*white); +} + +ALWAYS_INLINE static float colordifference_stdc(const f_pixel px, const f_pixel py); +inline static float colordifference_stdc(const f_pixel px, const f_pixel py) +{ + // px_b.rgb = px.rgb + 0*(1-px.a) // blend px on black + // px_b.a = px.a + 1*(1-px.a) + // px_w.rgb = px.rgb + 1*(1-px.a) // blend px on white + // px_w.a = px.a + 1*(1-px.a) + + // px_b.rgb = px.rgb // difference same as in opaque RGB + // px_b.a = 1 + // px_w.rgb = px.rgb - px.a // difference simplifies to formula below + // px_w.a = 1 + + // (px.rgb - px.a) - (py.rgb - py.a) + // (px.rgb - py.rgb) + (py.a - px.a) + + const double alphas = py.a-px.a; + return colordifference_ch(px.r, py.r, alphas) + + colordifference_ch(px.g, py.g, alphas) + + colordifference_ch(px.b, py.b, alphas); +} + +ALWAYS_INLINE static float colordifference(f_pixel px, f_pixel py); +inline static float colordifference(f_pixel px, f_pixel py) +{ +#if USE_SSE + const __m128 vpx = _mm_load_ps((const float*)&px); + const __m128 vpy = _mm_load_ps((const float*)&py); + + // y.a - x.a + __m128 alphas = _mm_sub_ss(vpy, vpx); + alphas = _mm_shuffle_ps(alphas,alphas,0); // copy first to all four + + __m128 onblack = _mm_sub_ps(vpx, vpy); // x - y + __m128 onwhite = _mm_add_ps(onblack, alphas); // x - y + (y.a - x.a) + + onblack = _mm_mul_ps(onblack, onblack); + onwhite = _mm_mul_ps(onwhite, onwhite); + const __m128 max = _mm_max_ps(onwhite, onblack); + + // add rgb, not a + const __m128 maxhl = _mm_movehl_ps(max, max); + const __m128 tmp = _mm_add_ps(max, maxhl); + const __m128 sum = _mm_add_ss(maxhl, _mm_shuffle_ps(tmp, tmp, 1)); + + const float res = _mm_cvtss_f32(sum); + assert(fabs(res - colordifference_stdc(px,py)) < 0.001); + return res; +#else + return colordifference_stdc(px,py); +#endif +} + +/* from pamcmap.h */ +union rgba_as_int { + rgba_pixel rgba; + unsigned int l; +}; + +typedef struct { + f_pixel acolor; + float adjusted_weight, // perceptual weight changed to tweak how mediancut selects colors + perceptual_weight; // number of pixels weighted by importance of different areas of the picture + + float color_weight; // these two change every time histogram subset is sorted + union { + unsigned int sort_value; + unsigned char likely_colormap_index; + } tmp; +} hist_item; + +typedef struct { + hist_item *achv; + void (*free)(void*); + double total_perceptual_weight; + unsigned int size; + unsigned int ignorebits; +} histogram; + +typedef struct { + f_pixel acolor; + float popularity; + bool fixed; // if true it's user-supplied and must not be changed (e.g in voronoi iteration) +} colormap_item; + +typedef struct colormap { + unsigned int colors; + void* (*malloc)(size_t); + void (*free)(void*); + colormap_item palette[]; +} colormap; + +struct acolorhist_arr_item { + union rgba_as_int color; + float perceptual_weight; +}; + +struct acolorhist_arr_head { + struct acolorhist_arr_item inline1, inline2; + unsigned int used, capacity; + struct acolorhist_arr_item *other_items; +}; + +struct acolorhash_table { + struct mempool *mempool; + unsigned int ignorebits, maxcolors, colors, cols, rows; + unsigned int hash_size; + unsigned int freestackp; + struct acolorhist_arr_item *freestack[512]; + struct acolorhist_arr_head buckets[]; +}; + +LIQ_PRIVATE void pam_freeacolorhash(struct acolorhash_table *acht); +LIQ_PRIVATE struct acolorhash_table *pam_allocacolorhash(unsigned int maxcolors, unsigned int surface, unsigned int ignorebits, void* (*malloc)(size_t), void (*free)(void*)); +LIQ_PRIVATE histogram *pam_acolorhashtoacolorhist(const struct acolorhash_table *acht, const double gamma, void* (*malloc)(size_t), void (*free)(void*)); +LIQ_PRIVATE bool pam_computeacolorhash(struct acolorhash_table *acht, const rgba_pixel *const pixels[], unsigned int cols, unsigned int rows, const unsigned char *importance_map); + +LIQ_PRIVATE void pam_freeacolorhist(histogram *h); + +LIQ_PRIVATE colormap *pam_colormap(unsigned int colors, void* (*malloc)(size_t), void (*free)(void*)); +LIQ_PRIVATE colormap *pam_duplicate_colormap(colormap *map); +LIQ_PRIVATE void pam_freecolormap(colormap *c); + +#endif diff --git a/viter.c b/viter.c new file mode 100644 index 0000000..f526326 --- /dev/null +++ b/viter.c @@ -0,0 +1,107 @@ +/* +© 2011-2016 by Kornel Lesiński. + +This file is part of libimagequant. + +libimagequant 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. + +libimagequant 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 libimagequant. If not, see . +*/ + +#include "libimagequant.h" +#include "pam.h" +#include "viter.h" +#include "nearest.h" +#include +#include + +#ifdef _OPENMP +#include +#else +#define omp_get_max_threads() 1 +#define omp_get_thread_num() 0 +#endif + +/* + * Voronoi iteration: new palette color is computed from weighted average of colors that map to that palette entry. + */ +LIQ_PRIVATE void viter_init(const colormap *map, const unsigned int max_threads, viter_state average_color[]) +{ + memset(average_color, 0, sizeof(average_color[0])*(VITER_CACHE_LINE_GAP+map->colors)*max_threads); +} + +LIQ_PRIVATE void viter_update_color(const f_pixel acolor, const float value, const colormap *map, unsigned int match, const unsigned int thread, viter_state average_color[]) +{ + match += thread * (VITER_CACHE_LINE_GAP+map->colors); + average_color[match].a += acolor.a * value; + average_color[match].r += acolor.r * value; + average_color[match].g += acolor.g * value; + average_color[match].b += acolor.b * value; + average_color[match].total += value; +} + +LIQ_PRIVATE void viter_finalize(colormap *map, const unsigned int max_threads, const viter_state average_color[]) +{ + for (unsigned int i=0; i < map->colors; i++) { + double a=0, r=0, g=0, b=0, total=0; + + // Aggregate results from all threads + for(unsigned int t=0; t < max_threads; t++) { + const unsigned int offset = (VITER_CACHE_LINE_GAP+map->colors) * t + i; + + a += average_color[offset].a; + r += average_color[offset].r; + g += average_color[offset].g; + b += average_color[offset].b; + total += average_color[offset].total; + } + + if (total && !map->palette[i].fixed) { + map->palette[i].acolor = (f_pixel){ + .a = a / total, + .r = r / total, + .g = g / total, + .b = b / total, + }; + map->palette[i].popularity = total; + } + } +} + +LIQ_PRIVATE double viter_do_iteration(histogram *hist, colormap *const map, viter_callback callback, const bool fast_palette) +{ + const unsigned int max_threads = omp_get_max_threads(); + viter_state average_color[(VITER_CACHE_LINE_GAP+map->colors) * max_threads]; + viter_init(map, max_threads, average_color); + struct nearest_map *const n = nearest_init(map, fast_palette); + hist_item *const achv = hist->achv; + const int hist_size = hist->size; + + double total_diff=0; + #pragma omp parallel for if (hist_size > 3000) \ + schedule(static) default(none) shared(average_color,callback) reduction(+:total_diff) + for(int j=0; j < hist_size; j++) { + float diff; + unsigned int match = nearest_search(n, &achv[j].acolor, achv[j].tmp.likely_colormap_index, &diff); + achv[j].tmp.likely_colormap_index = match; + total_diff += diff * achv[j].perceptual_weight; + + viter_update_color(achv[j].acolor, achv[j].perceptual_weight, map, match, omp_get_thread_num(), average_color); + + if (callback) callback(&achv[j], diff); + } + + nearest_free(n); + viter_finalize(map, max_threads, average_color); + + return total_diff / hist->total_perceptual_weight; +} diff --git a/viter.h b/viter.h new file mode 100644 index 0000000..bbbaaa1 --- /dev/null +++ b/viter.h @@ -0,0 +1,19 @@ + +#ifndef VITER_H +#define VITER_H + +// Spread memory touched by different threads at least 64B apart which I assume is the cache line size. This should avoid memory write contention. +#define VITER_CACHE_LINE_GAP ((64+sizeof(viter_state)-1)/sizeof(viter_state)) + +typedef struct { + double a, r, g, b, total; +} viter_state; + +typedef void (*viter_callback)(hist_item *item, float diff); + +LIQ_PRIVATE void viter_init(const colormap *map, const unsigned int max_threads, viter_state state[]); +LIQ_PRIVATE void viter_update_color(const f_pixel acolor, const float value, const colormap *map, unsigned int match, const unsigned int thread, viter_state average_color[]); +LIQ_PRIVATE void viter_finalize(colormap *map, const unsigned int max_threads, const viter_state state[]); +LIQ_PRIVATE double viter_do_iteration(histogram *hist, colormap *const map, viter_callback callback, const bool fast_palette); + +#endif