├── GroupBy.cuh
├── License.txt
├── Makefile
├── allocator.cuh
├── bfs_gpu_opt.cuh
├── comm.h
├── expander.cu
├── expander.cuh
├── gen_script.bash
├── get_time.bash
├── gpu_ibfs.cu
├── gpu_ibfs.cuh
├── graph.cu
├── graph.cuh
├── graph.h
├── graph_reader.cuh
├── graph_reader.h
├── groupby.cu
├── ibfs.cu
├── ibfs.h
├── inspector.cu
├── inspector.cuh
├── launch.bash
├── main.cpp
├── main.cu
├── proc_profil.bash
├── proc_time.bash
├── readme.md
├── reporter.cpp
├── run.bash
├── scan.cuh
├── traversal.cu
├── tuple_text_to_binary_csr
    ├── Makefile
    └── tuple_text_to_bin.cpp
├── util.h
├── validate.h
├── write_result.cuh
└── wtime.h


/GroupBy.cuh:
--------------------------------------------------------------------------------
  1 | #ifndef __GROUPBY__
  2 | #define __GROUPBY__
  3 | 
  4 | #include "comm.h"
  5 | 
  6 | template<typename vertex_t, typename index_t>
  7 | int GroupBy(
  8 | 	index_t num_bfs,
  9 | 	vertex_t *source_list,
 10 | 	vertex_t *new_list,
 11 | 	csr_graph ggraph
 12 | ){
 13 | 	index_t *hubmap=new index_t[ggraph->vert_count];
 14 | 	index_t *huboff=new index_t[ggraph->vert_count];
 15 | 	memset(hubmap,0,sizeof(index_t)*ggraph->vert_count);
 16 | 	memset(huboff,0,sizeof(index_t)*ggraph->vert_count);
 17 | 
 18 | 	/*counting eligible instances*/
 19 | 	for(index_t i=0;i<num_bfs;i++)
 20 | 	{
 21 | 		index_t beg=ggraph->beg_pos[source_list[i]];
 22 | 		index_t end=ggraph->beg_pos[source_list[i]+1];
 23 | 		
 24 | 		/*out-degree <= sdegree*/
 25 | 		if(end-beg>128) continue; 
 26 | 		
 27 | 		for(index_t j=beg;j<end;j++)
 28 | 		{
 29 | 			/*has hub neighbor*/
 30 | 			if(ggraph->beg_pos[ggraph->adj_list[j]+1]-
 31 | 				ggraph->beg_pos[ggraph->adj_list[j]] > 128)
 32 | 			{
 33 | 				hubmap[ggraph->adj_list[j]]++;
 34 | 				break;
 35 | 			}
 36 | 		}
 37 | 	}
 38 | 	
 39 | 	/*prefix-sum counted data*/
 40 | 	huboff[0]=0;
 41 | 	for(index_t i=1;i<ggraph->vert_count;i++)
 42 | 		huboff[i]=huboff[i-1]+hubmap[i-1];
 43 | 	index_t remainoff=huboff[ggraph->vert_count-1]+hubmap[ggraph->vert_count-1];
 44 | 	
 45 | 	
 46 | 	index_t test=0;
 47 | 	/*formulating group*/
 48 | 	for(index_t i=0;i<num_bfs;i++)
 49 | 	{
 50 | 		index_t beg=ggraph->beg_pos[source_list[i]];
 51 | 		index_t end=ggraph->beg_pos[source_list[i]+1];
 52 | 		
 53 | 		/*non-eligible instance*/
 54 | 		if(end-beg>128)
 55 | 		{
 56 | 			new_list[remainoff]=source_list[i];
 57 | 			remainoff++;
 58 | 			test++;
 59 | 			continue;
 60 | 			
 61 | 		}
 62 | 		
 63 | 		/*eligible instance*/
 64 | 		bool isEligible=false;
 65 | 		for(index_t j=beg;j<end;j++)
 66 | 		{
 67 | 			/*has hub neighbor*/
 68 | 			if(ggraph->beg_pos[ggraph->adj_list[j]+1]-
 69 | 				ggraph->beg_pos[ggraph->adj_list[j]] > 128)
 70 | 			{
 71 | 				new_list[huboff[ggraph->adj_list[j]]]=source_list[i];
 72 | 				huboff[ggraph->adj_list[j]]++;
 73 | 				isEligible=true;
 74 | 				break;
 75 | 			}
 76 | 		}
 77 | 		
 78 | 		if(!isEligible)
 79 | 		{
 80 | 			new_list[remainoff]=source_list[i];
 81 | 			remainoff++;
 82 | 			test++;
 83 | 		}
 84 | 	}
 85 | 	
 86 | 	std::ofstream myfile("origin_source.dat");
 87 | 	std::ofstream newfile("new_source.dat");
 88 | 	for(index_t i=0;i<num_bfs;i++)
 89 | 	{
 90 | 		myfile<<source_list[i]<<"\n";
 91 | 		newfile<<new_list[i]<<"\n";
 92 | 	}
 93 | 	myfile.close();
 94 | 	newfile.close();
 95 | 
 96 | 
 97 | 	index_t remain=num_bfs;
 98 | 	index_t grouped=0;
 99 | 	std::cout<<"Eligible-BFS(by-source-remain): ";
100 | 	for(index_t i=0;i<ggraph->vert_count;i++)
101 | 	{
102 | 		if(hubmap[i])
103 | 		{
104 | 			std::cout<<hubmap[i]<<" ";
105 | 			grouped+=hubmap[i];
106 | 		}
107 | 	}
108 | 	remain-=grouped;
109 | 	std::cout<<remain<<"\n";
110 | 	
111 | 	std::cout<<"Get test: "<<test<<"\n";
112 | 	delete[] hubmap;
113 | 	delete[] huboff;
114 | 	return 0;
115 | }
116 | 
117 | #endif
118 | 


--------------------------------------------------------------------------------
/License.txt:
--------------------------------------------------------------------------------
  1 | GNU GENERAL PUBLIC LICENSE
  2 | 
  3 | Version 3, 29 June 2007
  4 | 
  5 | Copyright © 2007 Free Software Foundation, Inc. <http://fsf.org/>
  6 | 
  7 | Everyone is permitted to copy and distribute verbatim copies of this license document, but changing it is not allowed.
  8 | 
  9 | Preamble
 10 | 
 11 | The GNU General Public License is a free, copyleft license for software and other kinds of works.
 12 | 
 13 | 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.
 14 | 
 15 | 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.
 16 | 
 17 | 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.
 18 | 
 19 | 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.
 20 | 
 21 | 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.
 22 | 
 23 | 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.
 24 | 
 25 | 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.
 26 | 
 27 | 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.
 28 | 
 29 | The precise terms and conditions for copying, distribution and modification follow.
 30 | 
 31 | TERMS AND CONDITIONS
 32 | 
 33 | 0. Definitions.
 34 | 
 35 | “This License” refers to version 3 of the GNU General Public License.
 36 | 
 37 | “Copyright” also means copyright-like laws that apply to other kinds of works, such as semiconductor masks.
 38 | 
 39 | “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.
 40 | 
 41 | 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.
 42 | 
 43 | A “covered work” means either the unmodified Program or a work based on the Program.
 44 | 
 45 | 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.
 46 | 
 47 | 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.
 48 | 
 49 | 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.
 50 | 
 51 | 1. Source Code.
 52 | 
 53 | 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.
 54 | 
 55 | 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.
 56 | 
 57 | 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.
 58 | 
 59 | 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.
 60 | 
 61 | The Corresponding Source need not include anything that users can regenerate automatically from other parts of the Corresponding Source.
 62 | 
 63 | The Corresponding Source for a work in source code form is that same work.
 64 | 
 65 | 2. Basic Permissions.
 66 | 
 67 | 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.
 68 | 
 69 | 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.
 70 | 
 71 | Conveying under any other circumstances is permitted solely under the conditions stated below. Sublicensing is not allowed; section 10 makes it unnecessary.
 72 | 
 73 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law.
 74 | 
 75 | 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.
 76 | 
 77 | 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.
 78 | 
 79 | 4. Conveying Verbatim Copies.
 80 | 
 81 | 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.
 82 | 
 83 | 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.
 84 | 
 85 | 5. Conveying Modified Source Versions.
 86 | 
 87 | 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:
 88 | 
 89 | a) The work must carry prominent notices stating that you modified it, and giving a relevant date.
 90 | 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”.
 91 | 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.
 92 | 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.
 93 | 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.
 94 | 
 95 | 6. Conveying Non-Source Forms.
 96 | 
 97 | 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:
 98 | 
 99 | 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.
100 | 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.
101 | 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.
102 | 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.
103 | 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.
104 | 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.
105 | 
106 | 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.
107 | 
108 | “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.
109 | 
110 | 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).
111 | 
112 | 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.
113 | 
114 | 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.
115 | 
116 | 7. Additional Terms.
117 | 
118 | “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.
119 | 
120 | 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.
121 | 
122 | 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:
123 | 
124 | a) Disclaiming warranty or limiting liability differently from the terms of sections 15 and 16 of this License; or
125 | 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
126 | 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
127 | d) Limiting the use for publicity purposes of names of licensors or authors of the material; or
128 | e) Declining to grant rights under trademark law for use of some trade names, trademarks, or service marks; or
129 | 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.
130 | 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.
131 | 
132 | 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.
133 | 
134 | 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.
135 | 
136 | 8. Termination.
137 | 
138 | 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).
139 | 
140 | 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.
141 | 
142 | 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.
143 | 
144 | 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.
145 | 
146 | 9. Acceptance Not Required for Having Copies.
147 | 
148 | 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.
149 | 
150 | 10. Automatic Licensing of Downstream Recipients.
151 | 
152 | 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.
153 | 
154 | 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.
155 | 
156 | 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.
157 | 
158 | 11. Patents.
159 | 
160 | 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”.
161 | 
162 | 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.
163 | 
164 | 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.
165 | 
166 | 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.
167 | 
168 | 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.
169 | 
170 | 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.
171 | 
172 | 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.
173 | 
174 | 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.
175 | 
176 | 12. No Surrender of Others' Freedom.
177 | 
178 | 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.
179 | 
180 | 13. Use with the GNU Affero General Public License.
181 | 
182 | 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.
183 | 
184 | 14. Revised Versions of this License.
185 | 
186 | 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.
187 | 
188 | 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.
189 | 
190 | 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.
191 | 
192 | 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.
193 | 
194 | 15. Disclaimer of Warranty.
195 | 
196 | 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.
197 | 
198 | 16. Limitation of Liability.
199 | 
200 | 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.
201 | 
202 | 17. Interpretation of Sections 15 and 16.
203 | 
204 | 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.
205 | 
206 | END OF TERMS AND CONDITIONS
207 | 
208 | How to Apply These Terms to Your New Programs
209 | 
210 | If you develop a new program, and you want it to be of the greatest possible use to the public, the best way to achieve this is to make it free software which everyone can redistribute and change under these terms.
211 | 
212 | To do so, attach the following notices to the program. It is safest to attach them to the start of each source file to most effectively state the exclusion of warranty; and each file should have at least the “copyright” line and a pointer to where the full notice is found.
213 | 
214 |     <one line to give the program's name and a brief idea of what it does.>
215 |     Copyright (C) <year>  <name of author>
216 | 
217 |     This program is free software: you can redistribute it and/or modify
218 |     it under the terms of the GNU General Public License as published by
219 |     the Free Software Foundation, either version 3 of the License, or
220 |     (at your option) any later version.
221 | 
222 |     This program is distributed in the hope that it will be useful,
223 |     but WITHOUT ANY WARRANTY; without even the implied warranty of
224 |     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
225 |     GNU General Public License for more details.
226 | 
227 |     You should have received a copy of the GNU General Public License
228 |     along with this program.  If not, see <http://www.gnu.org/licenses/>.
229 | Also add information on how to contact you by electronic and paper mail.
230 | 
231 | If the program does terminal interaction, make it output a short notice like this when it starts in an interactive mode:
232 | 
233 |     <program>  Copyright (C) <year>  <name of author>
234 |     This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
235 |     This is free software, and you are welcome to redistribute it
236 |     under certain conditions; type `show c' for details.
237 | The hypothetical commands `show w' and `show c' should show the appropriate parts of the General Public License. Of course, your program's commands might be different; for a GUI interface, you would use an “about box”.
238 | 
239 | You should also get your employer (if you work as a programmer) or school, if any, to sign a “copyright disclaimer” for the program, if necessary. For more information on this, and how to apply and follow the GNU GPL, see <http://www.gnu.org/licenses/>.
240 | 
241 | The GNU General Public License does not permit incorporating your program into proprietary programs. If your program is a subroutine library, you may consider it more useful to permit linking proprietary applications with the library. If this is what you want to do, use the GNU Lesser General Public License instead of this License. But first, please read <http://www.gnu.org/philosophy/why-not-lgpl.html>.
242 | 


--------------------------------------------------------------------------------
/Makefile:
--------------------------------------------------------------------------------
 1 | EXE=gpu-ibfs
 2 | 
 3 | COMMFLAGS=-O3 --compiler-options -Wall -Xptxas -v
 4 | CUCC= "$(shell which nvcc)"
 5 | 
 6 | CUFLAGS= -arch=sm_35  ${COMMFLAGS}#-Xptxas -dlcm=cg#disable l1 cache
 7 | CUFLAGS+= -ccbin=g++ -Xcompiler -fopenmp
 8 | 
 9 | 
10 | MPC    = "$(shell which mpicxx)"
11 | MPCFLAGS  = -Wall -I"$(shell dirname $(CUCC))/../include" -L"$(shell dirname $(CUCC))/../lib64" -lcudart -fopenmp
12 | 
13 | 
14 | ifeq ($(enable_monitor), 1)
15 | 	CUFLAGS+= -DENABLE_MONITORING
16 | endif
17 | 
18 | ifeq ($(enable_check), 1)
19 | 	CUFLAGS+= -DENABLE_CHECKING
20 | endif
21 | 
22 | 
23 | ifeq ($(enable_groupby), 1)
24 | 	CUFLAGS+= -DGROUPBY
25 | endif
26 | 
27 | 
28 | OBJS=  	main.o \
29 | 		ibfs.o \
30 | 		reporter.o 
31 | 
32 | DEPS= 	Makefile \
33 | 		expander.cuh \
34 | 		inspector.cuh \
35 | 		comm.h \
36 | 		graph.cuh \
37 | 		bfs_gpu_opt.cuh \
38 | 		wtime.h \
39 | 		validate.h \
40 | 		scan.cuh \
41 | 		allocator.cuh 
42 | 
43 | %.o:%.cpp $(DEPS)
44 | 	${MPC} -c  ${MPCFLAGS} $< -o $@
45 | 
46 | %.o:%.cu $(DEPS)
47 | 	${CUCC} -c  ${CUFLAGS} $< -o $@
48 | 
49 | ${EXE}:${OBJS}
50 | 	${MPC} ${OBJS} $(MPCFLAGS) -o ${EXE}
51 | 
52 | clean:
53 | 	rm -rf *.o ${EXE}
54 | 


--------------------------------------------------------------------------------
/allocator.cuh:
--------------------------------------------------------------------------------
 1 | #include "comm.h"
 2 | 
 3 | template <typename vertex_t, typename index_t, typename depth_t>
 4 | int graph<vertex_t, index_t, depth_t>::alloc_array()
 5 | {
 6 | 	cudaSetDevice(gpu_id);	
 7 | 	//long cpu_bytes= 0;
 8 | 	long gpu_bytes= 0;
 9 | 	
10 | 	const size_t cat_index_sz 	= sizeof(index_t)*BLKS_NUM*THDS_NUM;
11 | 	//const size_t blk_sz 				= sizeof(index_t)*BLKS_NUM*num_agg_bfs;
12 | 	const size_t merge_depth_sz	= sizeof(depth_t)*vert_count*num_agg_bfs;
13 | 	const size_t edge_sz 				= sizeof(vertex_t)*edge_count;
14 | 	const size_t vert_sz 				= sizeof(vertex_t)*vert_count;
15 | 	const size_t index_sz 			= sizeof(index_t)*vert_count;
16 | 	const size_t comp_sz 				= sizeof(comp_t)*vert_count;
17 | 	const size_t src_sz					= sizeof(vertex_t)*src_count;
18 | 	gpu_bytes+= edge_sz+(index_sz*2);
19 | 	gpu_bytes+=(cat_index_sz*6);
20 | 	gpu_bytes+=(comp_sz*2)+src_sz+merge_depth_sz;
21 | 	gpu_bytes+=(vert_sz*3);
22 | 	std::cout<<"GPU space: "<<gpu_bytes<<" byte(s)\n";
23 | 
24 | 	ggraph->edge_count=edge_count;
25 | 	ggraph->vert_count=vert_count;
26 | //	H_ERR(cudaSetDeviceFlags(cudaDeviceMapHost));
27 | 	H_ERR(cudaMalloc((void **)&ggraph->adj_card_d, 	index_sz));
28 | 	H_ERR(cudaMalloc((void **)&ggraph->beg_pos_d,	index_sz));
29 | 	H_ERR(cudaMalloc((void **)&ggraph->adj_list_d, 	edge_sz));
30 | 	H_ERR(cudaMalloc((void **)&ggraph->depth_merge, merge_depth_sz));
31 | 	H_ERR(cudaMemcpy(ggraph->beg_pos_d,ggraph->beg_pos,index_sz, 
32 | 								cudaMemcpyHostToDevice));
33 | 	H_ERR(cudaMemcpy(ggraph->adj_card_d,ggraph->adj_card,index_sz, 
34 | 								cudaMemcpyHostToDevice));
35 | 	H_ERR(cudaMemcpy(ggraph->adj_list_d,ggraph->adj_list,edge_sz,
36 | 								cudaMemcpyHostToDevice));
37 | 		
38 | 	H_ERR(cudaMalloc((void **)&ggraph->ex_sml_off, cat_index_sz));
39 | 	H_ERR(cudaMalloc((void **)&ggraph->ex_mid_off, cat_index_sz));
40 | 	H_ERR(cudaMalloc((void **)&ggraph->ex_lrg_off, cat_index_sz));
41 | 	H_ERR(cudaMalloc((void **)&ggraph->ex_cat_sml_sz,	cat_index_sz));
42 | 	H_ERR(cudaMalloc((void **)&ggraph->ex_cat_mid_sz,	cat_index_sz));
43 | 	H_ERR(cudaMalloc((void **)&ggraph->ex_cat_lrg_sz,	cat_index_sz));
44 | 	
45 | 	H_ERR(cudaMalloc((void **)&ggraph->ex_sml_q,	vert_sz));
46 | 	H_ERR(cudaMalloc((void **)&ggraph->ex_mid_q,	vert_sz));
47 | 	H_ERR(cudaMalloc((void **)&ggraph->ex_lrg_q,	vert_sz));
48 | 		
49 | 	H_ERR(cudaMalloc((void **)&ggraph->src_list,src_sz));
50 | 	H_ERR(cudaMemcpy(ggraph->src_list,src_list,src_sz,cudaMemcpyHostToDevice));
51 | 	H_ERR(cudaMalloc((void **)&ggraph->depth_comp_last,comp_sz));
52 | 	H_ERR(cudaMalloc((void **)&ggraph->depth_comp_curr,comp_sz));
53 | 	cudaStreamCreate(&ggraph->gstream);
54 | 	
55 | 	for(index_t i=0;i<num_agg_bfs;i++){
56 | 		//+----------------------
57 | 		//|FOR CLASSIFICATION
58 | 		//+----------------------------
59 | 	//	H_ERR(cudaMallocHost((void **)&gdata[i]->ex_sml_sz,sizeof(index_t)));
60 | 	//	H_ERR(cudaMallocHost((void **)&gdata[i]->ex_mid_sz,sizeof(index_t)));
61 | 	//	H_ERR(cudaMallocHost((void **)&gdata[i]->ex_lrg_sz,sizeof(index_t)));
62 | 		H_ERR(cudaHostAlloc((void **)&gdata[i]->ex_sml_sz,
63 | 					sizeof(index_t),cudaHostAllocMapped));
64 | 		H_ERR(cudaHostAlloc((void **)&gdata[i]->ex_mid_sz,
65 | 					sizeof(index_t),cudaHostAllocMapped));
66 | 		H_ERR(cudaHostAlloc((void **)&gdata[i]->ex_lrg_sz,
67 | 					sizeof(index_t),cudaHostAllocMapped));
68 | 		H_ERR(cudaHostGetDevicePointer((void **)&gdata[i]->ex_sml_sz_d,
69 | 					(void*)gdata[i]->ex_sml_sz,0));
70 | 		H_ERR(cudaHostGetDevicePointer((void **)&gdata[i]->ex_mid_sz_d,
71 | 					(void*)gdata[i]->ex_mid_sz,0));
72 | 		H_ERR(cudaHostGetDevicePointer((void **)&gdata[i]->ex_lrg_sz_d,
73 | 					(void*)gdata[i]->ex_lrg_sz,0));
74 | 		
75 | 		
76 | 		gdata[i]->stream=(cudaStream_t *)malloc(sizeof(cudaStream_t)*Q_CARD);
77 | 		for(index_t j=0;j<Q_CARD; j++) cudaStreamCreate(&gdata[i]->stream[j]);
78 | 	}
79 | 	return 0;
80 | }
81 | 
82 | 


--------------------------------------------------------------------------------
/bfs_gpu_opt.cuh:
--------------------------------------------------------------------------------
  1 | #include "scan.cuh"
  2 | #include "expander.cuh"
  3 | #include "inspector.cuh"
  4 | #include "wtime.h"
  5 | #include "validate.h"
  6 | #include "ibfs.h"
  7 | 
  8 | template <typename vertex_t, typename index_t, typename depth_t> 
  9 | bool cont_bfs
 10 | (
 11 | 	index_t level, index_t sw_level,
 12 | 	index_t *qsz,
 13 | 	index_t *last_ct,
 14 | 	index_t num_agg_bfs,
 15 | 	tdata *gdata
 16 | ){
 17 | 	bool cont_traverse=false;
 18 | 	qsz[0]=gdata[0]->ex_sml_sz[0]
 19 | 					+gdata[0]->ex_mid_sz[0]
 20 | 					+gdata[0]->ex_lrg_sz[0];
 21 | 	
 22 | 	if(!ENABLE_BTUP){
 23 | 		if(qsz[0]!=0){cont_traverse=true;}
 24 | 	}else{
 25 | 		if(last_ct[0]!=qsz[0]){
 26 | 			cont_traverse=true;
 27 | 			last_ct[0]=qsz[0];
 28 | 		}else if((qsz[0]!=0)&&((level==sw_level+1)||(level==sw_level+2)||(level==sw_level+3)||(level==sw_level+4))){
 29 | 			cont_traverse=true;
 30 | 			last_ct[0]=qsz[0];
 31 | 		}
 32 | 	}
 33 | //	std::cout<<"----\n";
 34 | //	std::cout<<last_ct[0]<<"\t"<<qsz[0]<<"\n";
 35 | //	std::cout<<sw_level<<"\t"<<level<<"\n";
 36 | 	return cont_traverse;
 37 | }
 38 | 
 39 | template <typename vertex_t, typename index_t, typename depth_t> 
 40 | void ibfs
 41 | (  
 42 | 	vertex_t 	*src_v,
 43 | 	index_t		src_grp_off,
 44 | 	tdata			*gdata,
 45 | 	csr_graph ggraph,
 46 | 	index_t		num_agg_bfs,
 47 | 	index_t		vert_count,
 48 | 	index_t		sw_level,
 49 | 	index_t		*last_ct,
 50 | 	depth_t		&level,
 51 | 	const index_t sml_shed,
 52 | 	const index_t lrg_shed
 53 | )
 54 | {
 55 | 	init_bfs<vertex_t,index_t,depth_t>
 56 | 	(
 57 | 		src_v,
 58 | 		src_grp_off,
 59 | 		gdata,
 60 | 		ggraph,
 61 | 		num_agg_bfs,
 62 | 		sml_shed,
 63 | 		lrg_shed
 64 | 	);
 65 | 	cudaDeviceSynchronize();
 66 | 
 67 | 	#ifdef ENABLE_MONITORING
 68 | 	double tm_insp;
 69 | 	double tm_expd;
 70 | 	double tm_step;
 71 | 	double tm_expand	= 0.0;
 72 | 	double tm_inspect	= 0.0;
 73 | //	depth_t *d_depth;
 74 | //	cudaMallocHost((void**)&d_depth,sizeof(depth_t)*vert_count);
 75 | //	cudaMemcpy(d_depth,ggraph->depth_merge, sizeof(depth_t)*vert_count,cudaMemcpyDeviceToHost);
 76 | //	for(int i=0;i<vert_count;i++)
 77 | //		if(d_depth[i]==0)std::cout<<"level 0: "<<i<<"\n";
 78 | //	level=0;
 79 | //	std::stringstream ss;
 80 | //	std::string filename[num_agg_bfs];
 81 | //	std::ofstream ofile;
 82 | //	for(index_t i=0;i<num_agg_bfs;i++){
 83 | //		ss.str("");ss.clear();
 84 | //		ss<<"console-"<<src_v[src_grp_off*num_agg_bfs+i]<<".log";filename[i]=ss.str();
 85 | //	}
 86 | //	for(index_t i=0;i<num_agg_bfs;i++){
 87 | //		ofile.open(filename[i].c_str(), std::fstream::out|std::fstream::app);
 88 | //		ofile		<<"@Level-"<<(int)level<<"-nextq-insp-expd-step: "
 89 | //						<<gdata[0]->ex_sml_sz[0]
 90 | //						+gdata[0]->ex_mid_sz[0]
 91 | //						+gdata[0]->ex_lrg_sz[0]<<" "
 92 | //						<<tm_insp<<" "<<tm_expd<<" "<<tm_step<<"\n";
 93 | //		ofile.close();
 94 | //	}
 95 | //	ofile.close();
 96 | 	#endif
 97 | 	
 98 | 	index_t qsz[num_agg_bfs];
 99 | 	for(level=1;;level++){
100 | 	//	if(level>sw_level+1)
101 | 			if(!cont_bfs<vertex_t,index_t,depth_t>
102 | 				(level, sw_level, qsz,last_ct,num_agg_bfs,gdata))break;
103 | 		
104 | 		//----------------------	
105 | 		//Expand ex_qs and mark frontier in depth_d	
106 | 		//------------------------------------------------
107 | 		#ifdef ENABLE_MONITORING
108 | 		tm_step=wtime();	
109 | 		tm_expd=wtime();	
110 | 		#endif
111 | 		expander
112 | 		<vertex_t, index_t, depth_t,comp_t>
113 | 		(
114 | 			gdata,
115 | 			ggraph,
116 | 			num_agg_bfs,
117 | 			sw_level,
118 | 			level
119 | 		);
120 | 		cudaDeviceSynchronize();
121 | 
122 | 		#ifdef ENABLE_MONITORING
123 | 		tm_expd=wtime()-tm_expd;	
124 | 		tm_insp=wtime();	
125 | 		#endif
126 | 		//-----------------------
127 | 		//Generate ex_qs from depth_d
128 | 		//---------------------------------
129 | 		inspector
130 | 		<vertex_t, index_t, depth_t,comp_t>
131 | 		(
132 | 			gdata,
133 | 			ggraph,
134 | 			num_agg_bfs,
135 | 			sw_level,
136 | 			level,
137 | 			vert_count,
138 | 			sml_shed,
139 | 			lrg_shed
140 | 		);
141 | 		cudaDeviceSynchronize();
142 | 		#ifdef ENABLE_MONITORING
143 | 		tm_insp=wtime()-tm_insp;
144 | 		tm_step=wtime()-tm_step;
145 | 	
146 |     std::cout<<"@Level-"<<(int)level;
147 |     if(level<sw_level) std::cout<<"-td-";
148 |     else if(level==sw_level) std::cout<<"-sw-";
149 |     else std::cout<<"-bu-";
150 |     std::cout<<"nextq-insp-expd-step: "
151 |               <<qsz[0]<<" "<<tm_insp<<" "<<tm_expd<<" "<<tm_step<<"\n";	
152 | 	//	for(index_t i=0;i<num_agg_bfs;i++){
153 | 	//	//	if(level<=sw_level+1) qsz[i]=0;
154 | 	//		ofile.open(filename[i].c_str(), std::fstream::out|std::fstream::app);
155 | 	//		ofile		<<"@Level-"<<(int)level<<"-nextq-insp-expd-step: "
156 | 	//						<<gdata[0]->ex_sml_sz[0]
157 | 	//						+gdata[0]->ex_mid_sz[0]
158 | 	//						+gdata[0]->ex_lrg_sz[0]<<" "
159 | 	//						<<tm_insp<<" "<<tm_expd<<" "<<tm_step<<"\n";
160 | 	//		ofile.close();
161 | 	//	}
162 | 	//	ofile.close();
163 | 		tm_expand		+= tm_expd;
164 | 		tm_inspect	+= tm_insp;
165 | 		#endif
166 | 	}
167 | 	#ifdef ENABLE_MONITORING
168 | //	for(index_t i=0;i<num_agg_bfs;i++){
169 | //		ofile.open(filename[i].c_str(), std::fstream::out|std::fstream::app);
170 | //		ofile<<"Total-insp-expd "<<tm_inspect<<" "<<tm_expand<<"\n";
171 | //		ofile.close();
172 | //	}
173 | //	ofile.close();
174 | 	std::cout<<"Break-fq-sz: "<<qsz[0]<<"\n";
175 | 	std::cout<<"Total-insp-expd "<<tm_inspect<<" "<<tm_expand<<"\n";
176 | 	#endif
177 | }
178 | 
179 | /////////////////////
180 | //CALLING FUNCTION FROM CPU
181 | ///////////////////////////
182 | template<typename vertex_t, typename index_t, typename depth_t>
183 | int graph<vertex_t, index_t, depth_t>::
184 | bfs_gpu_coalescing_mem()
185 | {
186 | 	cudaSetDevice(gpu_id);
187 | 	depth_t *temp;
188 | //	index_t agg_tr_edges, agg_tr_v;
189 | 	double tm_bfs;
190 | //	double average_teps	= 0.0;
191 | //	double curr_teps	= 0.0;
192 | //	index_t validate_count = 0;
193 | 	
194 | 	cudaMallocHost((void **)&temp,sizeof(depth_t)*vert_count*num_agg_bfs);
195 | 	for(index_t i=0;i<vert_count*num_agg_bfs;i++) temp[i]=INFTY;
196 | 	
197 | 	index_t *last_ct=new index_t[num_agg_bfs];
198 | 	index_t i;
199 | 	for(i=0; i<src_count/num_agg_bfs; i++){
200 | 	//	std::cout<<"Test "<<i+1<<"\n";
201 | 	//	std::cout<<"Started from: "<<src_list[i]<<"\n";
202 | 		ENABLE_BTUP		=false;
203 | 		SW_LEVEL			=false;
204 | 		agg_tr_edges	=0;
205 | 		cudaMemcpy(ggraph->depth_merge,temp,sizeof(depth_t)*vert_count*num_agg_bfs,
206 | 					cudaMemcpyHostToDevice);
207 | 		cudaMemset(ggraph->depth_comp_last,0,sizeof(comp_t)*vert_count);
208 | 		cudaMemset(ggraph->depth_comp_curr,0,sizeof(comp_t)*vert_count);
209 | 		last_ct[0]=-1;
210 | 
211 | 		depth_t level=0;
212 | 		tm_bfs=wtime();
213 | 		ibfs<vertex_t, index_t, depth_t>
214 | 		(
215 | 			src_list,
216 | 			i,
217 | 			gdata,
218 | 			ggraph,
219 | 			num_agg_bfs,
220 | 			vert_count,
221 | 			sw_level,
222 | 			last_ct,
223 | 			level,
224 | 			sml_shed,
225 | 			lrg_shed
226 | 		 );
227 | 		tm_bfs=wtime()-tm_bfs;
228 | 		reporter(tm_bfs, my_id, i);
229 | 	//	std::cout<<"Traversal-time: "<<tm_bfs<<" second(s)\n";
230 | 		
231 | 	//		if(level>2){
232 | 	//			validate_count ++;
233 | 	//			if(cudaMemcpy(depth, ggraph->depth_merge, 
234 | 	//								sizeof(depth_t)*vert_count, 
235 | 	//								cudaMemcpyDeviceToHost))
236 | 	//				std::cout<<"copy result error\n";
237 | 
238 | 	//			int ret = validate<index_t, vertex_t, depth_t>
239 | 	//				(depth, adj_list, adj_card, beg_pos, vert_count);
240 | 
241 | 	//			std::cout<<"\nBFS result validation: "<<
242 | 	//						((ret == 0 )? "CORRECT (":"WRONG (")<<ret<<")\n";
243 | 	//			report<vertex_t, index_t, depth_t>
244 | 	//				(agg_tr_edges, agg_tr_v, depth, adj_card, vert_count);
245 | 	//			curr_teps	= agg_tr_edges/(1000000000*tm_bfs);
246 | 	//			average_teps= (curr_teps+average_teps*(validate_count-1))
247 | 	//								/validate_count;
248 | 
249 | 	//			std::cout<<"Traversed vertices: "<< agg_tr_v<<"\t\t\t"
250 | 	//					 <<"Traversed edges: "<<agg_tr_edges<<"\n"
251 | 	//					 <<"Traversed time(s) :"<<tm_bfs<<"\t\t"
252 | 	//					 <<"Current TEPS (Billion): "<<curr_teps<<"\n"
253 | 	//					 <<"Average TEPS (Billion): "<<average_teps<<"\n";
254 | 	//		}else{
255 | 	//			std::cout<<"Traverse depth is "<<(int)level;
256 | 	//		}
257 | 	//		std::cout<<"\n====================================\n";
258 | 		}
259 | 	
260 | //	std::cout<<"Final Average TEPS (Billion): "<<average_teps<<"\n";
261 | 	return 0;
262 | }
263 | 


--------------------------------------------------------------------------------
/comm.h:
--------------------------------------------------------------------------------
  1 | #ifndef	COMM_HEADER
  2 | #define	COMM_HEADER
  3 | #include <stdio.h>
  4 | 
  5 | static void HandleError( cudaError_t err,
  6 |                          const char *file,
  7 |                          int line ) {
  8 |     if (err != cudaSuccess) {
  9 |         printf( "%s in %s at line %d\n", \
 10 |         cudaGetErrorString( err ),
 11 |                 file, line );
 12 |         exit( EXIT_FAILURE );
 13 |     }
 14 | }
 15 | #define H_ERR( err ) \
 16 |   (HandleError( err, __FILE__, __LINE__ ))
 17 | 
 18 | 
 19 | 
 20 | //////////////////////////////////////////////////
 21 | //SCALE*THDS_NUMS*sizeof(int) should be 
 22 | //limited by the size of the shared memory
 23 | /////////////////////////////////////////////////
 24 | //////////////////////////////////////////////////
 25 | #define	THDS_NUM			256	
 26 | #define	BLKS_NUM			256	
 27 | #define OMP_THDS			24
 28 | #define	V_NON_INC			-1
 29 | #define V_INI_HUB			-2
 30 | 
 31 | #define VALIDATE_TIMES		1
 32 | enum ex_q_t
 33 | {
 34 |   SML_Q,
 35 |   MID_Q,
 36 |   LRG_Q,
 37 |   NONE
 38 | };
 39 | 
 40 | //--------------------------
 41 | typedef 	int 						index_t;
 42 | typedef		int							vertex_t;
 43 | typedef 	unsigned char 	depth_t;
 44 | typedef 	uint4	comp_t;
 45 | //--------------------------------
 46 | 
 47 | typedef struct CSR_Graph{
 48 | 	vertex_t*	adj_list;
 49 | 	index_t*	adj_card;
 50 | 	index_t*	beg_pos;
 51 | 	
 52 | 	vertex_t*	adj_list_d;
 53 | 	index_t*	adj_card_d;
 54 | 	index_t*	beg_pos_d;
 55 | 	vertex_t*	src_list;	
 56 | 	
 57 | 	depth_t*	depth_merge;
 58 | 	
 59 | 	//for direction-switching
 60 | 	vertex_t*	ex_sml_q;
 61 | 	vertex_t*	ex_mid_q;
 62 | 	vertex_t*	ex_lrg_q;
 63 | 	index_t*	ex_cat_sml_sz;
 64 | 	index_t*	ex_cat_mid_sz;
 65 | 	index_t*	ex_cat_lrg_sz;
 66 | 	index_t*	ex_sml_off;
 67 | 	index_t*	ex_mid_off;
 68 | 	index_t*	ex_lrg_off;
 69 | 	
 70 | 	comp_t		*depth_comp_last;
 71 | 	comp_t		*depth_comp_curr;
 72 | 	index_t		vert_count;
 73 | 	index_t		edge_count;
 74 | 
 75 | 	cudaStream_t gstream;
 76 | } *csr_graph;
 77 | 
 78 | typedef struct Traverse{
 79 | 	depth_t* 	depth_d;
 80 | 	vertex_t*	parent_d;
 81 | 	index_t		ex_q_sz;
 82 | //	vertex_t*	ex_sml_q;
 83 | //	vertex_t*	ex_mid_q;
 84 | //	vertex_t*	ex_lrg_q;
 85 | 	
 86 | 	index_t	*ex_sml_sz;
 87 | 	index_t	*ex_mid_sz;
 88 | 	index_t	*ex_lrg_sz;
 89 | 	index_t	*ex_sml_sz_d;
 90 | 	index_t	*ex_mid_sz_d;
 91 | 	index_t	*ex_lrg_sz_d;
 92 | 	
 93 | ///	index_t*	ex_cat_sml_sz;
 94 | ///	index_t*	ex_cat_mid_sz;
 95 | ///	index_t*	ex_cat_lrg_sz;
 96 | ///	index_t*	ex_sml_off;
 97 | ///	index_t*	ex_mid_off;
 98 | ///	index_t*	ex_lrg_off;
 99 | 
100 | 	index_t*	tr_edges_c_d;
101 | 	index_t*	tr_edges_c_h;
102 | 
103 | 	cudaStream_t	*stream;
104 | } *tdata;
105 | 
106 | 
107 | #define	VIS				0x02
108 | #define UNVIS			0x00
109 | #define FRT				0x01
110 | #define	SET_VIS(a)		((a)=0x02)
111 | #define	SET_FRT(a)		((a)=0x01)
112 | #define	IS_FRT(a)		((a)==0x01)
113 | #define	IS_VIS(a)		((a)==0x02)
114 | #define	IS_UNVIS(a)		((a)==0x00)
115 | 
116 | //----------------------------------
117 | //GLOBAL VARIABLES
118 | //---------------------------------
119 | //--------------------------------
120 | #define INFTY			255	
121 | #endif
122 | 
123 | #ifndef EXTERN
124 | #define EXTERN
125 | //-------------------
126 | 
127 | index_t ENABLE_BTUP;
128 | index_t SW_LEVEL;
129 | index_t agg_tr_edges;//already traversed edges
130 | index_t EDGES_C;//total edges in the graph
131 | 
132 | //texture <vertex_t,	1, cudaReadModeElementType> tex_adj_list[4];
133 | //size_t  tex_adj_off[4];
134 | 
135 | //texture <index_t, 	1, cudaReadModeElementType> tex_card;
136 | //texture <index_t, 	1, cudaReadModeElementType> tex_strt;
137 | //texture <depth_t, 	1, cudaReadModeElementType> tex_depth;
138 | //
139 | //texture <vertex_t,	1, cudaReadModeElementType> tex_sml_exq;
140 | //texture <vertex_t,	1, cudaReadModeElementType> tex_mid_exq;
141 | //texture <vertex_t,	1, cudaReadModeElementType> tex_lrg_exq;
142 | 
143 | __device__	index_t		hub_vert[16];
144 | __device__ 	depth_t		hub_stat[16];
145 | __device__	index_t		hub_card[16];
146 | 
147 | #define HUB_SZ			16	
148 | #define HUB_BU_SZ		16//should be 1.25 of HUB_SZ
149 | 								//since there is no status
150 | 								//array in __shared__ mem
151 | #define HUB_CRITERIA	0	
152 | #define	Q_CARD	3
153 | 
154 | #endif
155 | 


--------------------------------------------------------------------------------
/expander.cu:
--------------------------------------------------------------------------------
  1 | /*
  2 |  * Copyright 2016 The George Washington University
  3 |  * Written by Hang Liu 
  4 |  * Directed by Prof. Howie Huang
  5 |  *
  6 |  * https://www.seas.gwu.edu/~howie/
  7 |  * Contact: iheartgraph@gmail.com
  8 |  *
  9 |  * 
 10 |  * Please cite the following paper:
 11 |  * 
 12 |  * Hang Liu, H. Howie Huang and Yang Hu. 2016. iBFS: Concurrent Breadth-First Search on GPUs. Proceedings of the 2016 International Conference on Management of Data. ACM. 
 13 |  *
 14 |  * This file is part of iBFS.
 15 |  *
 16 |  * iBFS is free software: you can redistribute it and/or modify
 17 |  * it under the terms of the GNU General Public License as published by
 18 |  * the Free Software Foundation, either version 3 of the License, or
 19 |  * (at your option) any later version.
 20 |  *
 21 |  * iBFS is distributed in the hope that it will be useful,
 22 |  * but WITHOUT ANY WARRANTY; without even the implied warranty of
 23 |  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the 
 24 |  * GNU General Public License for more details.
 25 |  *
 26 |  * You should have received a copy of the GNU General Public License
 27 |  * along with iBFS.  If not, see <http://www.gnu.org/licenses/>.
 28 |  */
 29 | 
 30 | //Hang Dec/10/2013
 31 | //Hang Mar/05/2015
 32 | #include "gpu_ibfs.cuh"
 33 | #include "util.h"
 34 | 
 35 | __device__ void __sync_warp(int predicate){
 36 | 	while((!__all(predicate))){
 37 | 		;
 38 | 	}
 39 | }
 40 | 
 41 | //This kernel is executed by one thread
 42 | __global__ void init_expand_sort
 43 | (	
 44 | 	vertex_t 	*src_list,
 45 | 	index_t 	*beg_pos_d,
 46 | 	index_t		src_grp_off,
 47 | 	index_t		joint_count,
 48 | 	index_t		bit_count,
 49 | 	index_t 	concurr_count,
 50 | 	comp_t		*depth_comp_last,
 51 | 	comp_t		*depth_comp_curr,
 52 | 	index_t		*ex_lrg_sz_d,
 53 | 	vertex_t	*ex_lrg_q_d
 54 | )
 55 | {
 56 | 	index_t tid=threadIdx.x+blockIdx.x*blockDim.x;
 57 | 	const index_t GRNTY=blockDim.x*gridDim.x;
 58 | 
 59 | 	while(tid<concurr_count){
 60 | 		index_t inst_off=tid/bit_count;
 61 | 		vertex_t vert_id=src_list[tid+src_grp_off];
 62 | 		index_t	card=beg_pos_d[vert_id+1]-beg_pos_d[vert_id];
 63 | 		if(!card){return;}
 64 | 		
 65 | 		//put all src into big queue 
 66 | 		//-since we have enough threads 
 67 | 		ex_lrg_q_d[tid]=vert_id;
 68 | 	
 69 | 		comp_t status=(make_uint4(1))<<(tid%bit_count);
 70 | 		depth_comp_last[vert_id*joint_count+inst_off]=status;
 71 | 		depth_comp_curr[vert_id*joint_count+inst_off]=status;
 72 | 
 73 | 		tid+=GRNTY;
 74 | 	}
 75 | 	return ;
 76 | }
 77 | 
 78 | void 
 79 | gpu_ibfs::
 80 | init_bfs(index_t src_grp_off){
 81 | 	init_expand_sort
 82 | 	<<<BLKS_NUM,THDS_NUM>>>
 83 | 	(
 84 | 		src_list_d,
 85 | 		beg_pos_d,
 86 | 		src_grp_off,
 87 | 		joint_count,
 88 | 		bit_count,
 89 | 		concurr_count,
 90 | 		depth_comp_last,
 91 | 		depth_comp_curr,
 92 | 		ex_lrg_sz_d,
 93 | 		ex_lrg_q_d
 94 | 	);
 95 | }
 96 | 
 97 | //+----------------------
 98 | //|for ex_sml_q_d expansion
 99 | //+---------------------------
100 | __global__ void td_expand_thd
101 | (	
102 | 	comp_t			*depth_comp_last,
103 | 	comp_t			*depth_comp_curr,
104 | 	depth_t			curr_level,
105 | 	index_t			ex_sml_sz,
106 | 	const vertex_t* __restrict__ ex_q_d,
107 | 	const index_t*  __restrict__ beg_pos_d,
108 | 	const vertex_t* __restrict__ adj_list_d
109 | )
110 | {
111 | 	const index_t q_sz		= ex_sml_sz;
112 | 	const index_t GRNLTY	= blockDim.x * gridDim.x;
113 | 	index_t tid				= threadIdx.x+blockIdx.x*blockDim.x;
114 | 
115 | 	//used for prefetching
116 | 	vertex_t 	ex_ver_curr;
117 | 	index_t 	end_curr;
118 | 	index_t 	beg_curr;
119 | 	vertex_t 	adj_vert;
120 | 	comp_t 	vert_depth, adj_depth;
121 | 
122 | 	while(tid<q_sz)
123 | 	{
124 | 		ex_ver_curr	= ex_q_d[tid];
125 | 		beg_curr		= beg_pos_d[ex_ver_curr];
126 | 		end_curr		= beg_pos_d[ex_ver_curr+1];
127 | 		vert_depth	= depth_comp_last[ex_ver_curr];
128 | 		while(beg_curr<end_curr)
129 | 		{
130 | 			adj_vert=adj_list_d[beg_curr];
131 | 			adj_depth=depth_comp_curr[adj_vert];
132 | 
133 | 			if((~(adj_depth.x))&vert_depth.x)
134 | 				atomicOr(&(depth_comp_curr[adj_vert].x),vert_depth.x);
135 | 
136 | 			if((~(adj_depth.y))&vert_depth.y)
137 | 				atomicOr(&(depth_comp_curr[adj_vert].y),vert_depth.y);
138 | 
139 | 			if((~(adj_depth.z))&vert_depth.z)
140 | 				atomicOr(&(depth_comp_curr[adj_vert].z),vert_depth.z);
141 | 
142 | 			if((~(adj_depth.w))&vert_depth.w)
143 | 				atomicOr(&(depth_comp_curr[adj_vert].w),vert_depth.w);
144 | 			beg_curr++;
145 | 		}
146 | 		tid += GRNLTY;
147 | 	}
148 | }
149 | 
150 | //+-----------------------
151 | //|ex_mid_q_d expansion
152 | //+------------------------------
153 | __global__ void td_expand_warp
154 | (	
155 | 	comp_t			*depth_comp_last,
156 | 	comp_t			*depth_comp_curr,
157 | 	depth_t			curr_level,
158 | 	index_t			ex_mid_sz,
159 | 	const vertex_t* __restrict__ ex_q_d,
160 | 	const index_t*  __restrict__ beg_pos_d,
161 | 	const vertex_t* __restrict__ adj_list_d
162 | )
163 | {
164 | 	const index_t q_sz		= ex_mid_sz;
165 | 	const index_t vec_sz	= 32;
166 | 	const index_t TID			= threadIdx.x+blockIdx.x*blockDim.x;
167 | 	const index_t lane_s	= TID & (vec_sz-1);
168 | 	const index_t GRNLTY	= (blockDim.x*gridDim.x)/vec_sz;
169 | 	index_t	vec_id				= TID/vec_sz;
170 | 
171 | 	vertex_t 	ex_ver_curr,adj_vert;
172 | 	index_t 	beg_curr, end_curr;
173 | 	comp_t 	vert_depth, adj_depth;
174 | 	
175 | 	while(vec_id<q_sz)
176 | 	{
177 | 		ex_ver_curr	= ex_q_d[vec_id];
178 | 		beg_curr		= beg_pos_d[ex_ver_curr];
179 | 		end_curr		= beg_pos_d[ex_ver_curr+1];
180 | 		vert_depth	= depth_comp_last[ex_ver_curr];
181 | 		index_t lane= beg_curr+lane_s;
182 | 		while(lane<end_curr)
183 | 		{
184 | 			adj_vert=adj_list_d[lane];
185 | 			adj_depth=depth_comp_curr[adj_vert];
186 | 			if((~(adj_depth.x))&vert_depth.x)
187 | 				atomicOr(&(depth_comp_curr[adj_vert].x),vert_depth.x);
188 | 
189 | 			if((~(adj_depth.y))&vert_depth.y)
190 | 				atomicOr(&(depth_comp_curr[adj_vert].y),vert_depth.y);
191 | 
192 | 			if((~(adj_depth.z))&vert_depth.z)
193 | 				atomicOr(&(depth_comp_curr[adj_vert].z),vert_depth.z);
194 | 
195 | 			if((~(adj_depth.w))&vert_depth.w)
196 | 				atomicOr(&(depth_comp_curr[adj_vert].w),vert_depth.w);
197 | 			lane+=vec_sz;
198 | 		}
199 | 		vec_id += GRNLTY;
200 | 	}
201 | }
202 | 
203 | //+-----------------------
204 | //|ex_lrg_q_d expansion
205 | //+------------------------------
206 | __global__ void td_expand_cta
207 | (	
208 | 	comp_t			*depth_comp_last,
209 | 	comp_t			*depth_comp_curr,
210 | 	depth_t			curr_level,
211 | 	index_t			ex_lrg_sz,
212 | 	const vertex_t* __restrict__ ex_q_d,
213 | 	const index_t*  __restrict__ beg_pos_d,
214 | 	const vertex_t* __restrict__ adj_list_d
215 | )
216 | {
217 | 	const index_t q_sz	= ex_lrg_sz;
218 | 	index_t	vec_id		= blockIdx.x;
219 | 
220 | 	vertex_t 	ex_ver_curr,adj_vert;
221 | 	index_t 	end_curr;
222 | 	index_t 	beg_curr;
223 | 	comp_t 	vert_depth, adj_depth;
224 | 	
225 | 	while(vec_id<q_sz)
226 | 	{
227 | 		ex_ver_curr	= ex_q_d[vec_id];
228 | 		beg_curr		= beg_pos_d[ex_ver_curr];
229 | 		end_curr		= beg_pos_d[ex_ver_curr+1];
230 | 		vert_depth	= depth_comp_last[ex_ver_curr];
231 | 		index_t lane= beg_curr+threadIdx.x;
232 | 		while(lane<end_curr)
233 | 		{
234 | 			adj_vert=adj_list_d[lane];
235 | 			adj_depth=depth_comp_curr[adj_vert];
236 | 
237 | 			if((~(adj_depth.x))&vert_depth.x)
238 | 				atomicOr(&(depth_comp_curr[adj_vert].x),vert_depth.x);
239 | 
240 | 			if((~(adj_depth.y))&vert_depth.y)
241 | 				atomicOr(&(depth_comp_curr[adj_vert].y),vert_depth.y);
242 | 
243 | 			if((~(adj_depth.z))&vert_depth.z)
244 | 				atomicOr(&(depth_comp_curr[adj_vert].z),vert_depth.z);
245 | 
246 | 			if((~(adj_depth.w))&vert_depth.w)
247 | 				atomicOr(&(depth_comp_curr[adj_vert].w),vert_depth.w);
248 | 			lane+=blockDim.x;
249 | 		}
250 | 		vec_id += gridDim.x;
251 | 	}
252 | }
253 | 
254 | 
255 | //+------------------------
256 | //|ex_mid_q_d expansion
257 | //+------------------------------
258 | __global__ void sw_expand_warp
259 | (	
260 | 	comp_t			*depth_comp_last,
261 | 	comp_t			*depth_comp_curr,
262 | 	depth_t			curr_level,
263 | 	index_t			ex_sz,
264 | 	const vertex_t* __restrict__ ex_q_d,
265 | 	const index_t*  __restrict__ beg_pos_d,
266 | 	const vertex_t* __restrict__ adj_list_d
267 | )
268 | {
269 | 	const index_t q_sz		= ex_sz;
270 | 	const index_t vec_sz	= 32;
271 | 	const index_t TID			= threadIdx.x+blockIdx.x*blockDim.x;
272 | 	const index_t lane_s	= TID & (vec_sz-1);
273 | 	const index_t GRNLTY	= (blockDim.x*gridDim.x)/vec_sz;
274 | 	index_t	vec_id				= TID/vec_sz;
275 | 
276 | 	vertex_t 	ex_ver_curr;
277 | 	index_t 	end_curr, end_curr_revised;
278 | 	index_t 	beg_curr;
279 | 	unsigned int vert_depth_up, vert_depth_low;
280 | 	comp_t 		vert_depth;
281 | 
282 | 	while(vec_id<q_sz)
283 | 	{
284 | 		ex_ver_curr	= ex_q_d[vec_id];
285 | 		beg_curr		= beg_pos_d[ex_ver_curr];
286 | 		end_curr		= beg_pos_d[ex_ver_curr+1];
287 | 		vert_depth	= depth_comp_curr[ex_ver_curr];
288 | 		index_t lane= beg_curr+lane_s;
289 | 		
290 | 		//make sure all 32 threads can get into the voting loop 
291 | 		if((end_curr-beg_curr)&(vec_sz-1))
292 |       end_curr_revised = beg_curr+((((end_curr-beg_curr)>>5)+1)<<5);
293 |     else 
294 |       end_curr_revised = end_curr;
295 | 
296 | 		while(lane<end_curr_revised)
297 | 		{
298 | 			if(__any((~vert_depth)==0)) break;
299 | 			if(lane<end_curr)
300 | 				vert_depth|=depth_comp_last[adj_list_d[lane]];
301 | 			
302 | 			for(int i=16;i>=1;i>>=1){
303 | 				vert_depth.x|=__shfl_xor((int)vert_depth.x,i,32);
304 | 				vert_depth.y|=__shfl_xor((int)vert_depth.y,i,32);
305 | 				vert_depth.z|=__shfl_xor((int)vert_depth.z,i,32);
306 | 				vert_depth.w|=__shfl_xor((int)vert_depth.w,i,32);
307 | 			}
308 | 
309 | 			lane+=vec_sz;
310 | 		}
311 | 		
312 | 		if(lane_s==0) depth_comp_curr[ex_ver_curr]=vert_depth;
313 | 		vec_id += GRNLTY;
314 | 	}
315 | }
316 | 
317 | //+------------------------
318 | //|ex_mid_q_d expansion
319 | //+------------------------------
320 | __global__ void sw_expand_thd
321 | (	
322 | 	comp_t			*depth_comp_last,
323 | 	comp_t			*depth_comp_curr,
324 | 	depth_t			curr_level,
325 | 	index_t			ex_sz,
326 | 	const vertex_t* __restrict__ ex_q_d,
327 | 	const index_t*  __restrict__ beg_pos_d,
328 | 	const vertex_t* __restrict__ adj_list_d
329 | )
330 | {
331 | 	const index_t q_sz		= ex_sz;
332 | 	index_t tid			= threadIdx.x+blockIdx.x*blockDim.x;
333 | 	const index_t GRNLTY	= blockDim.x*gridDim.x;
334 | 
335 | 	vertex_t 	ex_ver_curr;
336 | 	index_t 	beg_curr, end_curr;
337 | 	comp_t 	adj_depth_curr, vert_depth;
338 | 	
339 | 	while(tid<q_sz)
340 | 	{
341 | 		ex_ver_curr	= ex_q_d[tid];
342 | 		end_curr		= beg_pos_d[ex_ver_curr+1];
343 | 		beg_curr		= beg_pos_d[ex_ver_curr];
344 | 		index_t lane= beg_curr;
345 | 		vert_depth=depth_comp_curr[ex_ver_curr];
346 | 		while(lane<end_curr)
347 | 		{
348 | 			if((~vert_depth)==0) break;
349 | 			adj_depth_curr=depth_comp_last[adj_list_d[lane]];
350 | 			if(((~vert_depth)&adj_depth_curr) != 0) 
351 | 				vert_depth|=adj_depth_curr;			
352 | 			lane++;
353 | 		}
354 | 
355 | 		if((vert_depth^depth_comp_curr[ex_ver_curr]) != 0) 
356 | 			depth_comp_curr[ex_ver_curr]=vert_depth;
357 | 		tid += GRNLTY;
358 | 	}
359 | }
360 | 
361 | //+-----------------
362 | //|CLFY_EXPAND_SORT
363 | //+----------------------
364 | void 
365 | gpu_ibfs::
366 | td_expand(depth_t	curr_level)
367 | {
368 | 	td_expand_thd
369 | 	<<<BLKS_NUM, THDS_NUM, 0, stream[0]>>>
370 | 	(
371 | 		depth_comp_last,
372 | 		depth_comp_curr,
373 | 		curr_level,
374 | 		ex_sml_sz[0],
375 | 		ex_sml_q_d,
376 | 		beg_pos_d,
377 | 		adj_list_d
378 | 	);
379 | 
380 | 	td_expand_warp
381 | 	<<<BLKS_NUM, THDS_NUM, 0, stream[1]>>>
382 | 	(
383 | 		depth_comp_last,
384 | 		depth_comp_curr,
385 | 		curr_level,
386 | 		ex_mid_sz[0],
387 | 		ex_mid_q_d,
388 | 		beg_pos_d,
389 | 		adj_list_d
390 | 	);
391 | 
392 | 	
393 | 	td_expand_cta
394 | 	<<<BLKS_NUM, THDS_NUM, 0, stream[2]>>>
395 | 	(
396 | 		depth_comp_last,
397 | 		depth_comp_curr,
398 | 		curr_level,
399 | 		ex_lrg_sz[0],
400 | 		ex_lrg_q_d,
401 | 		beg_pos_d,
402 | 		adj_list_d
403 | 	);
404 | 
405 | 	for(index_t i=0;i<Q_CARD; i++)
406 | 		cudaStreamSynchronize(stream[i]);
407 | }
408 | 
409 | //+----------------------
410 | //|CLFY_EXPAND_SORT
411 | //+----------------------
412 | void 
413 | gpu_ibfs::
414 | bu_expand(depth_t	curr_level)
415 | {
416 | 	sw_expand_thd
417 | 	<<<BLKS_NUM, THDS_NUM, 0, stream[0]>>>
418 | 	(
419 | 		depth_comp_last,
420 | 		depth_comp_curr,
421 | 		curr_level,
422 | 		ex_sml_sz[0],
423 | 		ex_sml_q_d,
424 | 		beg_pos_d,
425 | 		adj_list_d
426 | 	);
427 | 		
428 | 	sw_expand_thd
429 | 	<<<BLKS_NUM, THDS_NUM, 0, stream[1]>>>
430 | 	(
431 | 		depth_comp_last,
432 | 		depth_comp_curr,
433 | 		curr_level,
434 | 		ex_mid_sz[0],
435 | 		ex_mid_q_d,
436 | 		beg_pos_d,
437 | 		adj_list_d
438 | 	);
439 | 	
440 | 	sw_expand_thd
441 | 	<<<BLKS_NUM, THDS_NUM, 0, stream[2]>>>
442 | 	(
443 | 		depth_comp_last,
444 | 		depth_comp_curr,
445 | 		curr_level,
446 | 		ex_lrg_sz[0],
447 | 		ex_lrg_q_d,
448 | 		beg_pos_d,
449 | 		adj_list_d
450 | 	);
451 | 
452 | 	for(index_t i=0;i<Q_CARD; i++)
453 | 		cudaStreamSynchronize(stream[i]);
454 | }
455 | 
456 | //+----------------------
457 | //|CLFY_EXPAND_SORT
458 | //+----------------------
459 | void 
460 | gpu_ibfs::
461 | sw_expand(depth_t curr_level)
462 | {
463 | 	sw_expand_thd
464 | 	<<<BLKS_NUM<<1, THDS_NUM, 0, stream[0]>>>
465 | 	(
466 | 		depth_comp_last,
467 | 		depth_comp_curr,
468 | 		curr_level,
469 | 		ex_sml_sz[0],
470 | 		ex_sml_q_d,
471 | 		beg_pos_d,
472 | 		adj_list_d
473 | 	);
474 | 		
475 | 	sw_expand_warp
476 | 	<<<BLKS_NUM<<1, THDS_NUM, 0, stream[1]>>>
477 | 	(
478 | 		depth_comp_last,
479 | 		depth_comp_curr,
480 | 		curr_level,
481 | 		ex_mid_sz[0],
482 | 		ex_mid_q_d,
483 | 		beg_pos_d,
484 | 		adj_list_d
485 | 	);
486 | 	
487 | 	sw_expand_warp
488 | 	<<<BLKS_NUM<<1, THDS_NUM, 0, stream[2]>>>
489 | 	(
490 | 		depth_comp_last,
491 | 		depth_comp_curr,
492 | 		curr_level,
493 | 		ex_lrg_sz[0],
494 | 		ex_lrg_q_d,
495 | 		beg_pos_d,
496 | 		adj_list_d
497 | 	);
498 | 
499 | 	for(index_t i=0;i<Q_CARD; i++)
500 | 		cudaStreamSynchronize(stream[i]);
501 | }
502 | 
503 | void 
504 | gpu_ibfs::
505 | expander(depth_t level)
506 | {
507 | 	if(level<=sw_level)	
508 | 		td_expand(level);
509 | 	else if((level==sw_level+1))
510 | 		sw_expand(level);
511 | 	else{//also use merged fq
512 | 		//differ sw_expand of no imbalance issue
513 | 		bu_expand(level);
514 | 	}
515 | }
516 | 


--------------------------------------------------------------------------------
/expander.cuh:
--------------------------------------------------------------------------------
  1 | //Hang Dec/10/2013
  2 | //Hang Mar/05/2015
  3 | 
  4 | #include "comm.h"
  5 | 
  6 | __device__ void __sync_warp(int predicate){
  7 | 	while((!__all(predicate))){
  8 | 		;
  9 | 	}
 10 | }
 11 | 
 12 | //This kernel is executed by one thread
 13 | template<typename vertex_t, 
 14 | 		typename index_t,
 15 | 		typename depth_t>
 16 | __global__ void init_expand_sort
 17 | (	
 18 | 	vertex_t 	*src_list,
 19 | 	index_t		src_grp_off,
 20 | 	depth_t		*depth_merge,
 21 | 	comp_t		*depth_comp_last,
 22 | 	comp_t		*depth_comp_curr,
 23 | 	index_t		num_agg_bfs,
 24 | 	vertex_t	*parent_d,
 25 | 	index_t		*adj_card_d,
 26 | 	index_t		*ex_sml_sz_d,
 27 | 	index_t		*ex_mid_sz_d,
 28 | 	index_t		*ex_lrg_sz_d,
 29 | 	vertex_t	*ex_sml_q,
 30 | 	vertex_t	*ex_mid_q,
 31 | 	vertex_t	*ex_lrg_q,
 32 | 	index_t		vert_count,
 33 | 	index_t		sml_shed,
 34 | 	index_t		lrg_shed
 35 | )
 36 | {
 37 | //	parent_d[src_v]	=-1;
 38 | 	if(threadIdx.x==num_agg_bfs){
 39 | 		ex_sml_sz_d[0]	= 0;
 40 | 		ex_mid_sz_d[0]	= 0;
 41 | 		ex_lrg_sz_d[0]	= num_agg_bfs;
 42 | 	}
 43 | 
 44 | 	if(threadIdx.x<num_agg_bfs){
 45 | 		vertex_t src_v=src_list[num_agg_bfs*src_grp_off+threadIdx.x];
 46 | 	//	depth_merge[src_v*num_agg_bfs+threadIdx.x]= 0;
 47 | 	//	
 48 | 	//	index_t	card	= adj_card_d[src_v];
 49 | 	//	if(!card) return;
 50 | 		
 51 | 		//put all src into big queue since we have enough threads 
 52 | 		//initally
 53 | 		ex_lrg_q[threadIdx.x]=src_v;
 54 | 		
 55 | 		unsigned int status=1;
 56 | 		if(threadIdx.x<32){
 57 | 			status<<=threadIdx.x;
 58 | 			depth_comp_last[src_v].x=status;
 59 | 			depth_comp_curr[src_v].x=status;
 60 | 		}else if(threadIdx.x<64){
 61 | 			status<<=(threadIdx.x-32);
 62 | 			depth_comp_last[src_v].y=status;
 63 | 			depth_comp_curr[src_v].y=status;
 64 | 		}else if(threadIdx.x<96){
 65 | 			status<<=(threadIdx.x-64);
 66 | 			depth_comp_last[src_v].z=status;
 67 | 			depth_comp_curr[src_v].z=status;
 68 | 		}else if(threadIdx.x<128){
 69 | 			status<<=(threadIdx.x-96);
 70 | 			depth_comp_last[src_v].w=status;
 71 | 			depth_comp_curr[src_v].w=status;
 72 | 		}
 73 | 	}
 74 | 	
 75 | 	return ;
 76 | }
 77 | 
 78 | template<typename vertex_t,
 79 | 				typename index_t,
 80 | 				typename depth_t>
 81 | void init_bfs(
 82 | 	vertex_t	*src_v, 
 83 | 	index_t		src_grp_off,
 84 | 	tdata 		*gdata,
 85 | 	csr_graph ggraph,
 86 | 	index_t 	num_agg_bfs,
 87 | 	index_t		sml_shed,
 88 | 	index_t		lrg_shed
 89 | ){
 90 | 
 91 | 	init_expand_sort
 92 | 	<vertex_t, index_t, depth_t>
 93 | 	<<<1,THDS_NUM>>>
 94 | 	(
 95 | 		ggraph->src_list,
 96 | 		src_grp_off,
 97 | 		ggraph->depth_merge,
 98 | 		ggraph->depth_comp_last,
 99 | 		ggraph->depth_comp_curr,
100 | 		num_agg_bfs,
101 | 		gdata[0]->parent_d,
102 | 		ggraph->adj_card_d,
103 | 		gdata[0]->ex_sml_sz_d,
104 | 		gdata[0]->ex_mid_sz_d,
105 | 		gdata[0]->ex_lrg_sz_d,
106 | 		ggraph->ex_sml_q,
107 | 		ggraph->ex_mid_q,
108 | 		ggraph->ex_lrg_q,
109 | 		ggraph->vert_count,
110 | 		sml_shed,
111 | 		lrg_shed
112 | 	);
113 | 
114 | //cudaDeviceSynchronize();
115 | //	index_t vert_count=ggraph->vert_count;
116 | //		comp_t *d_status_curr,*d_status_last;
117 | //		cudaMallocHost((void **)&d_status_curr, sizeof(comp_t)*vert_count);
118 | //		cudaMallocHost((void **)&d_status_last, sizeof(comp_t)*vert_count);
119 | //		cudaMemcpy(d_status_curr, ggraph->depth_comp_curr,sizeof(comp_t)*vert_count,cudaMemcpyDeviceToHost);
120 | //		cudaMemcpy(d_status_last, ggraph->depth_comp_last,sizeof(comp_t)*vert_count,cudaMemcpyDeviceToHost);
121 | //		
122 | //		int ft_count=0; int diff=0;
123 | //		for(int i=0;i<vert_count;i++){
124 | //			if(d_status_curr[i]!=0) diff++;
125 | //			if(d_status_curr[i]!=0xffffffffffffffff && ggraph->adj_card[i]>0) ft_count++;
126 | //		}
127 | //		std::cout<<"ft_count-diff "<<ft_count<<" "<<diff<<"\n";
128 | //		cudaFreeHost(d_status_last);
129 | //		cudaFreeHost(d_status_curr);
130 | //	
131 | }
132 | 
133 | //+----------------------
134 | //|for ex_sml_q expansion
135 | //+---------------------------
136 | template<typename vertex_t, 
137 | 		typename index_t,
138 | 		typename depth_t,
139 | 		typename comp_t>
140 | __global__ void td_expand_thd
141 | (	
142 | 	comp_t			*depth_comp_last,
143 | 	comp_t			*depth_comp_curr,
144 | 	index_t			num_agg_bfs,
145 | 	vertex_t		*parent_d,
146 | 	depth_t			curr_level,
147 | 	index_t			ex_sml_sz,
148 | 	const vertex_t* __restrict__ ex_q_d,
149 | 	const index_t*	__restrict__ adj_card_d,
150 | 	const index_t*  __restrict__ beg_pos_d,
151 | 	const vertex_t* __restrict__ adj_list_d
152 | )
153 | {
154 | 	const index_t q_sz		= ex_sml_sz;
155 | 	const index_t GRNLTY	= blockDim.x * gridDim.x;
156 | 	index_t tid				= threadIdx.x+blockIdx.x*blockDim.x;
157 | 
158 | 	//used for prefetching
159 | 	vertex_t 	ex_ver_curr;
160 | 	index_t 	card_curr;
161 | 	index_t 	beg_pos_curr;
162 | 	vertex_t 	adj_vert;
163 | 	comp_t vert_depth,adj_depth;
164 | 	unsigned int vert_depth_vec[4];
165 | 
166 | 	while(tid<q_sz)
167 | 	{
168 | 		ex_ver_curr	= ex_q_d[tid];
169 | 		card_curr		= adj_card_d[ex_ver_curr];
170 | 		beg_pos_curr= beg_pos_d[ex_ver_curr];
171 | 		index_t lane= beg_pos_curr;
172 | 		card_curr+= beg_pos_curr;
173 | 		vert_depth=depth_comp_last[ex_ver_curr];
174 | 		vert_depth_vec[0]=vert_depth.x;
175 | 		vert_depth_vec[1]=vert_depth.y;
176 | 		vert_depth_vec[2]=vert_depth.z;
177 | 		vert_depth_vec[3]=vert_depth.w;
178 | 		
179 | 		while(lane<card_curr)
180 | 		{
181 | 			adj_vert=adj_list_d[lane];
182 | 			adj_depth=depth_comp_curr[adj_vert];
183 | 
184 | 			if((~(adj_depth.x))&vert_depth_vec[0])
185 | 				atomicOr(&(depth_comp_curr[adj_vert].x),vert_depth_vec[0]);
186 | 
187 | 			if((~(adj_depth.y))&vert_depth_vec[1])
188 | 				atomicOr(&(depth_comp_curr[adj_vert].y),vert_depth_vec[1]);
189 | 
190 | 			if((~(adj_depth.z))&vert_depth_vec[2])
191 | 				atomicOr(&(depth_comp_curr[adj_vert].z),vert_depth_vec[2]);
192 | 
193 | 			if((~(adj_depth.w))&vert_depth_vec[3])
194 | 				atomicOr(&(depth_comp_curr[adj_vert].w),vert_depth_vec[3]);
195 | 			lane++;
196 | 		}
197 | 		tid += GRNLTY;
198 | 	}
199 | }
200 | 
201 | //+-----------------------
202 | //|ex_mid_q expansion
203 | //+------------------------------
204 | template<typename vertex_t, 
205 | 		typename index_t,
206 | 		typename depth_t,
207 | 		typename comp_t>
208 | __global__ void td_expand_warp
209 | (	
210 | 	comp_t			*depth_comp_last,
211 | 	comp_t			*depth_comp_curr,
212 | 	index_t			num_agg_bfs,
213 | 	vertex_t		*parent_d,
214 | 	depth_t			curr_level,
215 | 	index_t			ex_mid_sz,
216 | 	const vertex_t* __restrict__ ex_q_d,
217 | 	const index_t*	__restrict__ adj_card_d,
218 | 	const index_t*  __restrict__ beg_pos_d,
219 | 	const vertex_t* __restrict__ adj_list_d
220 | )
221 | {
222 | 	const index_t q_sz		= ex_mid_sz;
223 | 	const index_t vec_sz	= 32;
224 | 	const index_t TID			= threadIdx.x+blockIdx.x*blockDim.x;
225 | 	const index_t lane_s	= TID & (vec_sz-1);
226 | 	const index_t GRNLTY	= (blockDim.x*gridDim.x)/vec_sz;
227 | 	index_t	vec_id				= TID/vec_sz;
228 | 
229 | 	vertex_t 	ex_ver_curr,adj_vert;
230 | 	index_t 	card_curr;
231 | 	index_t 	beg_pos_curr;
232 | 	comp_t 	vert_depth,adj_depth;
233 | 	unsigned int vert_depth_vec[4];
234 | 	
235 | 	while(vec_id<q_sz)
236 | 	{
237 | 		ex_ver_curr	= ex_q_d[vec_id];
238 | 		beg_pos_curr= beg_pos_d[ex_ver_curr];
239 | 		card_curr		= adj_card_d[ex_ver_curr];
240 | 		vert_depth=depth_comp_last[ex_ver_curr];
241 | 		vert_depth_vec[0]=vert_depth.x;
242 | 		vert_depth_vec[1]=vert_depth.y;
243 | 		vert_depth_vec[2]=vert_depth.z;
244 | 		vert_depth_vec[3]=vert_depth.w;
245 | 		index_t lane= beg_pos_curr+lane_s;
246 | 		card_curr+= beg_pos_curr;
247 | 		while(lane<card_curr)
248 | 		{
249 | 			adj_vert=adj_list_d[lane];
250 | 			adj_depth=depth_comp_curr[adj_vert];
251 | 
252 | 			if((~(adj_depth.x))&vert_depth_vec[0])
253 | 				atomicOr(&(depth_comp_curr[adj_vert].x),vert_depth_vec[0]);
254 | 
255 | 			if((~(adj_depth.y))&vert_depth_vec[1])
256 | 				atomicOr(&(depth_comp_curr[adj_vert].y),vert_depth_vec[1]);
257 | 
258 | 			if((~(adj_depth.z))&vert_depth_vec[2])
259 | 				atomicOr(&(depth_comp_curr[adj_vert].z),vert_depth_vec[2]);
260 | 
261 | 			if((~(adj_depth.w))&vert_depth_vec[3])
262 | 				atomicOr(&(depth_comp_curr[adj_vert].w),vert_depth_vec[3]);
263 | 			lane+=vec_sz;
264 | 		}
265 | 		vec_id += GRNLTY;
266 | 	}
267 | }
268 | 
269 | //+-----------------------
270 | //|ex_lrg_q expansion
271 | //+------------------------------
272 | template<typename vertex_t, 
273 | 		typename index_t,
274 | 		typename depth_t,
275 | 		typename comp_t>
276 | __global__ void td_expand_cta
277 | (	
278 | 	comp_t			*depth_comp_last,
279 | 	comp_t			*depth_comp_curr,
280 | 	index_t			num_agg_bfs,
281 | 	vertex_t		*parent_d,
282 | 	depth_t			curr_level,
283 | 	index_t			ex_lrg_sz,
284 | 	const vertex_t* __restrict__ ex_q_d,
285 | 	const index_t*	__restrict__ adj_card_d,
286 | 	const index_t*  __restrict__ beg_pos_d,
287 | 	const vertex_t* __restrict__ adj_list_d
288 | )
289 | {
290 | 	const index_t q_sz	= ex_lrg_sz;
291 | 	index_t	vec_id		= blockIdx.x;
292 | 
293 | 	vertex_t 	ex_ver_curr,adj_vert;
294 | 	index_t 	card_curr;
295 | 	index_t 	beg_pos_curr;
296 | 	comp_t 	vert_depth,adj_depth;
297 | 	unsigned int vert_depth_vec[4];
298 | 	
299 | 	while(vec_id<q_sz)
300 | 	{
301 | 		ex_ver_curr	= ex_q_d[vec_id];
302 | 		beg_pos_curr= beg_pos_d[ex_ver_curr];
303 | 		card_curr		= adj_card_d[ex_ver_curr];
304 | 		vert_depth=depth_comp_last[ex_ver_curr];
305 | 		vert_depth_vec[0]=vert_depth.x;
306 | 		vert_depth_vec[1]=vert_depth.y;
307 | 		vert_depth_vec[2]=vert_depth.z;
308 | 		vert_depth_vec[3]=vert_depth.w;
309 | 		index_t lane= beg_pos_curr+threadIdx.x;
310 | 		card_curr+= beg_pos_curr;
311 | 		while(lane<card_curr)
312 | 		{
313 | 			adj_vert=adj_list_d[lane];
314 | 			adj_depth=depth_comp_curr[adj_vert];
315 | 
316 | 			if((~(adj_depth.x))&vert_depth_vec[0])
317 | 				atomicOr(&(depth_comp_curr[adj_vert].x),vert_depth_vec[0]);
318 | 
319 | 			if((~(adj_depth.y))&vert_depth_vec[1])
320 | 				atomicOr(&(depth_comp_curr[adj_vert].y),vert_depth_vec[1]);
321 | 
322 | 			if((~(adj_depth.z))&vert_depth_vec[2])
323 | 				atomicOr(&(depth_comp_curr[adj_vert].z),vert_depth_vec[2]);
324 | 
325 | 			if((~(adj_depth.w))&vert_depth_vec[3])
326 | 				atomicOr(&(depth_comp_curr[adj_vert].w),vert_depth_vec[3]);
327 | 			lane+=blockDim.x;
328 | 		}
329 | 		vec_id += gridDim.x;
330 | 	}
331 | }
332 | 
333 | 
334 | //+------------------------
335 | //|ex_mid_q expansion
336 | //+------------------------------
337 | template<typename vertex_t, 
338 | 		typename index_t,
339 | 		typename depth_t,
340 | 		typename comp_t>
341 | __global__ void sw_expand_warp
342 | (	
343 | 	comp_t			*depth_comp_last,
344 | 	comp_t			*depth_comp_curr,
345 | 	index_t			num_agg_bfs,
346 | 	vertex_t		*parent_d,
347 | 	depth_t			curr_level,
348 | 	index_t			ex_sz,
349 | 	const vertex_t* __restrict__ ex_q_d,
350 | 	const index_t*	__restrict__ adj_card_d,
351 | 	const index_t*  __restrict__ beg_pos_d,
352 | 	const vertex_t* __restrict__ adj_list_d
353 | )
354 | {
355 | 	const index_t q_sz		= ex_sz;
356 | 	const index_t vec_sz	= 32;
357 | 	const index_t TID			= threadIdx.x+blockIdx.x*blockDim.x;
358 | 	const index_t lane_s	= TID & (vec_sz-1);
359 | 	const index_t GRNLTY	= (blockDim.x*gridDim.x)/vec_sz;
360 | 	index_t	vec_id				= TID/vec_sz;
361 | 
362 | 	vertex_t 	ex_ver_curr;
363 | 	index_t 	card_curr, card_curr_revised;
364 | 	index_t 	beg_pos_curr;
365 | 	unsigned int vert_depth_vec[4];
366 | 	comp_t 		vert_depth;
367 | 
368 | 	while(vec_id<q_sz)
369 | 	{
370 | 		ex_ver_curr	= ex_q_d[vec_id];
371 | 		beg_pos_curr= beg_pos_d[ex_ver_curr];
372 | 		card_curr		= adj_card_d[ex_ver_curr];
373 | 		vert_depth=depth_comp_curr[ex_ver_curr];
374 | 		vert_depth_vec[0]=vert_depth.x;
375 | 		vert_depth_vec[1]=vert_depth.y;
376 | 		vert_depth_vec[2]=vert_depth.z;
377 | 		vert_depth_vec[3]=vert_depth.w;
378 | 		index_t lane= beg_pos_curr+lane_s;
379 | 		
380 | 		//make sure all 32 threads can get into the voting loop 
381 | 		if(card_curr%vec_sz)
382 |       card_curr_revised = (((card_curr>>5)+1)<<5);
383 |     else 
384 |       card_curr_revised = card_curr;
385 | 
386 | 		card_curr+= beg_pos_curr;
387 | 		card_curr_revised+=beg_pos_curr;
388 | 		while(lane<card_curr_revised)
389 | 		{
390 | 			/*if(	vert_depth_vec[0]==0xffffffff &&
391 | 					vert_depth_vec[1]==0xffffffff &&
392 | 					vert_depth_vec[2]==0xffffffff &&
393 | 					vert_depth_vec[3]==0xffffffff
394 | 					) break;*/
395 | 			if(lane<card_curr){
396 | 				vert_depth=depth_comp_last[adj_list_d[lane]];
397 | 				
398 | 				vert_depth_vec[0]|=vert_depth.x;
399 | 				vert_depth_vec[1]|=vert_depth.y;
400 | 				vert_depth_vec[2]|=vert_depth.z;
401 | 				vert_depth_vec[3]|=vert_depth.w;
402 | 			}
403 | 
404 | 			for(int i=16;i>=1;i>>=1){
405 | 				vert_depth_vec[0]|=__shfl_xor((int)vert_depth_vec[0],i,32);
406 | 				vert_depth_vec[1]|=__shfl_xor((int)vert_depth_vec[1],i,32);
407 | 				vert_depth_vec[2]|=__shfl_xor((int)vert_depth_vec[2],i,32);
408 | 				vert_depth_vec[3]|=__shfl_xor((int)vert_depth_vec[3],i,32);
409 | 			}
410 | 			lane+=vec_sz;
411 | 		}
412 | 		
413 | 		if(lane_s==0){
414 | 			depth_comp_curr[ex_ver_curr].x=vert_depth_vec[0];
415 | 			depth_comp_curr[ex_ver_curr].y=vert_depth_vec[1];
416 | 			depth_comp_curr[ex_ver_curr].z=vert_depth_vec[2];
417 | 			depth_comp_curr[ex_ver_curr].w=vert_depth_vec[3];
418 | 		}
419 | 		
420 | 		vec_id += GRNLTY;
421 | 	}
422 | }
423 | 
424 | //+------------------------
425 | //|ex_mid_q expansion
426 | //+------------------------------
427 | template<typename vertex_t, 
428 | 		typename index_t,
429 | 		typename depth_t,
430 | 		typename comp_t>
431 | __global__ void sw_expand_thd
432 | (	
433 | 	comp_t			*depth_comp_last,
434 | 	comp_t			*depth_comp_curr,
435 | 	index_t			num_agg_bfs,
436 | 	vertex_t		*parent_d,
437 | 	depth_t			curr_level,
438 | 	index_t			ex_sz,
439 | 	const vertex_t* __restrict__ ex_q_d,
440 | 	const index_t*	__restrict__ adj_card_d,
441 | 	const index_t*  __restrict__ beg_pos_d,
442 | 	const vertex_t* __restrict__ adj_list_d
443 | )
444 | {
445 | 	const index_t q_sz		= ex_sz;
446 | 	index_t tid			= threadIdx.x+blockIdx.x*blockDim.x;
447 | 	const index_t GRNLTY	= blockDim.x*gridDim.x;
448 | 
449 | 	vertex_t 	ex_ver_curr;
450 | 	index_t 	card_curr;
451 | 	index_t 	beg_pos_curr;
452 | 	comp_t 		 vert_depth;
453 | 	unsigned int vert_depth_vec[4];
454 | 	
455 | 	while(tid<q_sz)
456 | 	{
457 | 		ex_ver_curr	= ex_q_d[tid];
458 | 		card_curr		= adj_card_d[ex_ver_curr];
459 | 		beg_pos_curr= beg_pos_d[ex_ver_curr];
460 | 		index_t lane= beg_pos_curr;
461 | 		card_curr+= beg_pos_curr;
462 | 		vert_depth=depth_comp_curr[ex_ver_curr];
463 | 		vert_depth_vec[0]=vert_depth.x;
464 | 		vert_depth_vec[1]=vert_depth.y;
465 | 		vert_depth_vec[2]=vert_depth.z;
466 | 		vert_depth_vec[3]=vert_depth.w;
467 | 		while(lane<card_curr)
468 | 		{
469 | 			/*if(	vert_depth_vec[0]==0xffffffff &&
470 | 					vert_depth_vec[1]==0xffffffff &&
471 | 					vert_depth_vec[2]==0xffffffff &&
472 | 					vert_depth_vec[3]==0xffffffff
473 | 					) break;*/
474 | 			vert_depth=depth_comp_last[adj_list_d[lane]];
475 | 			
476 | 			vert_depth_vec[0]|=vert_depth.x;
477 | 			vert_depth_vec[1]|=vert_depth.y;
478 | 			vert_depth_vec[2]|=vert_depth.z;
479 | 			vert_depth_vec[3]|=vert_depth.w;
480 | 			lane++;
481 | 		}
482 | 		
483 | 		depth_comp_curr[ex_ver_curr].x=vert_depth_vec[0];
484 | 		depth_comp_curr[ex_ver_curr].y=vert_depth_vec[1];
485 | 		depth_comp_curr[ex_ver_curr].z=vert_depth_vec[2];
486 | 		depth_comp_curr[ex_ver_curr].w=vert_depth_vec[3];
487 | 		tid += GRNLTY;
488 | 	}
489 | }
490 | 
491 | //+-----------------
492 | //|CLFY_EXPAND_SORT
493 | //+----------------------
494 | template<typename vertex_t, 
495 | 		typename index_t,
496 | 		typename depth_t>
497 | void td_expand
498 | (
499 | 	tdata	*gdata,
500 | 	csr_graph ggraph,
501 | 	index_t	num_agg_bfs,
502 | 	depth_t	curr_level
503 | )
504 | {
505 | 	td_expand_thd<vertex_t,index_t,depth_t,comp_t>
506 | 	<<<BLKS_NUM, THDS_NUM, 0, gdata[0]->stream[0]>>>
507 | 	(
508 | 		ggraph->depth_comp_last,
509 | 		ggraph->depth_comp_curr,
510 | 		num_agg_bfs,
511 | 		gdata[0]->parent_d,
512 | 		curr_level,
513 | 		gdata[0]->ex_sml_sz[0],
514 | 		ggraph->ex_sml_q,
515 | 		ggraph->adj_card_d,
516 | 		ggraph->beg_pos_d,
517 | 		ggraph->adj_list_d
518 | 	);
519 | 
520 | 	td_expand_warp<vertex_t,index_t,depth_t,comp_t>
521 | 	<<<BLKS_NUM, THDS_NUM, 0, gdata[0]->stream[1]>>>
522 | 	(
523 | 		ggraph->depth_comp_last,
524 | 		ggraph->depth_comp_curr,
525 | 		num_agg_bfs,
526 | 		gdata[0]->parent_d,
527 | 		curr_level,
528 | 		gdata[0]->ex_mid_sz[0],
529 | 		ggraph->ex_mid_q,
530 | 		ggraph->adj_card_d,
531 | 		ggraph->beg_pos_d,
532 | 		ggraph->adj_list_d
533 | 	);
534 | 
535 | 	
536 | 	td_expand_cta<vertex_t,index_t,depth_t,comp_t>
537 | 	<<<BLKS_NUM, THDS_NUM, 0, gdata[0]->stream[2]>>>
538 | 	(
539 | 		ggraph->depth_comp_last,
540 | 		ggraph->depth_comp_curr,
541 | 		num_agg_bfs,
542 | 		gdata[0]->parent_d,
543 | 		curr_level,
544 | 		gdata[0]->ex_lrg_sz[0],
545 | 		ggraph->ex_lrg_q,
546 | 		ggraph->adj_card_d,
547 | 		ggraph->beg_pos_d,
548 | 		ggraph->adj_list_d
549 | 	);
550 | 
551 | 	for(index_t i=0;i<Q_CARD; i++)
552 | 		cudaStreamSynchronize(gdata[0]->stream[i]);
553 | }
554 | 
555 | //+----------------------
556 | //|CLFY_EXPAND_SORT
557 | //+----------------------
558 | template<typename vertex_t, 
559 | 		typename index_t,
560 | 		typename depth_t>
561 | void bu_expand
562 | (	
563 | 	tdata	*gdata,
564 | 	csr_graph ggraph,
565 | 	index_t	num_agg_bfs,
566 | 	depth_t	curr_level
567 | )
568 | {
569 | 	sw_expand_thd<vertex_t,index_t,depth_t,comp_t>
570 | 	<<<BLKS_NUM, THDS_NUM, 0, gdata[0]->stream[0]>>>
571 | 	(
572 | 		ggraph->depth_comp_last,
573 | 		ggraph->depth_comp_curr,
574 | 		num_agg_bfs,
575 | 		gdata[0]->parent_d,
576 | 		curr_level,
577 | 		gdata[0]->ex_sml_sz[0],
578 | 		ggraph->ex_sml_q,
579 | 		ggraph->adj_card_d,
580 | 		ggraph->beg_pos_d,
581 | 		ggraph->adj_list_d
582 | 	);
583 | 		
584 | 	sw_expand_thd<vertex_t,index_t,depth_t,comp_t>
585 | 	<<<BLKS_NUM, THDS_NUM, 0, gdata[0]->stream[1]>>>
586 | 	(
587 | 		ggraph->depth_comp_last,
588 | 		ggraph->depth_comp_curr,
589 | 		num_agg_bfs,
590 | 		gdata[0]->parent_d,
591 | 		curr_level,
592 | 		gdata[0]->ex_mid_sz[0],
593 | 		ggraph->ex_mid_q,
594 | 		ggraph->adj_card_d,
595 | 		ggraph->beg_pos_d,
596 | 		ggraph->adj_list_d
597 | 	);
598 | 	
599 | 	sw_expand_thd<vertex_t,index_t,depth_t,comp_t>
600 | 	<<<BLKS_NUM, THDS_NUM, 0, gdata[0]->stream[2]>>>
601 | 	(
602 | 		ggraph->depth_comp_last,
603 | 		ggraph->depth_comp_curr,
604 | 		num_agg_bfs,
605 | 		gdata[0]->parent_d,
606 | 		curr_level,
607 | 		gdata[0]->ex_lrg_sz[0],
608 | 		ggraph->ex_lrg_q,
609 | 		ggraph->adj_card_d,
610 | 		ggraph->beg_pos_d,
611 | 		ggraph->adj_list_d
612 | 	);
613 | 
614 | 	for(index_t i=0;i<Q_CARD; i++)
615 | 		cudaStreamSynchronize(gdata[0]->stream[i]);
616 | }
617 | 
618 | //+----------------------
619 | //|CLFY_EXPAND_SORT
620 | //+----------------------
621 | template<typename vertex_t, 
622 | 		typename index_t,
623 | 		typename depth_t,
624 | 		typename comp_t>
625 | void sw_expand
626 | (	
627 | 	tdata	*gdata,
628 | 	csr_graph ggraph,
629 | 	index_t	num_agg_bfs,
630 | 	depth_t	curr_level
631 | )
632 | {
633 | 	sw_expand_thd<vertex_t,index_t,depth_t,comp_t>
634 | 	<<<BLKS_NUM, THDS_NUM, 0, gdata[0]->stream[0]>>>
635 | 	(
636 | 		ggraph->depth_comp_last,
637 | 		ggraph->depth_comp_curr,
638 | 		num_agg_bfs,
639 | 		gdata[0]->parent_d,
640 | 		curr_level,
641 | 		gdata[0]->ex_sml_sz[0],
642 | 		ggraph->ex_sml_q,
643 | 		ggraph->adj_card_d,
644 | 		ggraph->beg_pos_d,
645 | 		ggraph->adj_list_d
646 | 	);
647 | 		
648 | 	sw_expand_warp<vertex_t,index_t,depth_t,comp_t>
649 | 	<<<BLKS_NUM, THDS_NUM, 0, gdata[0]->stream[1]>>>
650 | 	(
651 | 		ggraph->depth_comp_last,
652 | 		ggraph->depth_comp_curr,
653 | 		num_agg_bfs,
654 | 		gdata[0]->parent_d,
655 | 		curr_level,
656 | 		gdata[0]->ex_mid_sz[0],
657 | 		ggraph->ex_mid_q,
658 | 		ggraph->adj_card_d,
659 | 		ggraph->beg_pos_d,
660 | 		ggraph->adj_list_d
661 | 	);
662 | 	
663 | 	sw_expand_warp<vertex_t,index_t,depth_t,comp_t>
664 | 	<<<BLKS_NUM, THDS_NUM, 0, gdata[0]->stream[2]>>>
665 | 	(
666 | 		ggraph->depth_comp_last,
667 | 		ggraph->depth_comp_curr,
668 | 		num_agg_bfs,
669 | 		gdata[0]->parent_d,
670 | 		curr_level,
671 | 		gdata[0]->ex_lrg_sz[0],
672 | 		ggraph->ex_lrg_q,
673 | 		ggraph->adj_card_d,
674 | 		ggraph->beg_pos_d,
675 | 		ggraph->adj_list_d
676 | 	);
677 | 
678 | 	for(index_t i=0;i<Q_CARD; i++)
679 | 		cudaStreamSynchronize(gdata[0]->stream[i]);
680 | }
681 | 
682 | 
683 | template<typename vertex_t, 
684 | 		typename index_t,
685 | 		typename depth_t,
686 | 		typename comp_t>
687 | void expander(	
688 | 	tdata	*gdata,
689 | 	csr_graph ggraph,
690 | 	index_t	num_agg_bfs,
691 | 	index_t	sw_level,
692 | 	depth_t	level
693 | )
694 | {
695 | 	if(level<=sw_level)	
696 | 		td_expand
697 | 		<vertex_t, index_t, depth_t>
698 | 		(
699 | 			gdata,
700 | 			ggraph,
701 | 			num_agg_bfs,
702 | 			level
703 | 		);
704 | 	else if((level==sw_level+1||level==sw_level+2||level==sw_level+3))
705 | 		sw_expand
706 | 		<vertex_t, index_t, depth_t,comp_t>
707 | 		(
708 | 			gdata,
709 | 			ggraph,
710 | 			num_agg_bfs,
711 | 			level
712 | 		);
713 | 	else{//also use merged fq
714 | 		bu_expand//but do not consider imbalance issue 
715 | 		<vertex_t, index_t, depth_t>
716 | 		(
717 | 			gdata,
718 | 			ggraph,
719 | 			num_agg_bfs,
720 | 			level
721 | 		);
722 | 	}
723 | }
724 | 


--------------------------------------------------------------------------------
/gen_script.bash:
--------------------------------------------------------------------------------
 1 | #!/bin/bash
 2 | 
 3 | declare -a name=(fb hw kg1 ok pk rd rm tw wk)
 4 | for file in ${name[@]}
 5 | do
 6 | 	for ((i=1;i<=32;i=i*2))
 7 | 	do
 8 | 		#cp jobScript"$i"gpu-fb jobScript"$i"gpu-$file
 9 | 		sed -i "s/TG-CCR120025/TG-CTS100002/g" jobScript"$i"gpu-$file
10 | 		#sed -i "s/"$i"gpu/-"$i"gpu-$file/g" jobScript"$i"gpu
11 | 	done
12 | done
13 | 
14 | 


--------------------------------------------------------------------------------
/get_time.bash:
--------------------------------------------------------------------------------
 1 | declare -a data=(rm)
 2 | for arr in ${data[@]}
 3 | do
 4 | 	echo "Processing $arr ... "
 5 | 	for ((i=1;i<=32;i=i*2))
 6 | 	do
 7 | 		file=$(find ./ -maxdepth 1 -iname "traversal-"$i"gpu-$arr.o*")
 8 | 
 9 | 		if [ -f $file ]
10 | 		then
11 | 			tm=$(grep "Traversal-iter" $file | awk -v tm=0 -v count=0 '{tm+=$2;count++}END{print tm}')	
12 | 			echo $tm
13 | 		fi
14 | 	done
15 | 	echo "----------------"
16 | done
17 | 


--------------------------------------------------------------------------------
/gpu_ibfs.cu:
--------------------------------------------------------------------------------
  1 | /*
  2 |  * Copyright 2016 The George Washington University
  3 |  * Written by Hang Liu 
  4 |  * Directed by Prof. Howie Huang
  5 |  *
  6 |  * https://www.seas.gwu.edu/~howie/
  7 |  * Contact: iheartgraph@gmail.com
  8 |  *
  9 |  * 
 10 |  * Please cite the following paper:
 11 |  * 
 12 |  * Hang Liu, H. Howie Huang and Yang Hu. 2016. iBFS: Concurrent Breadth-First Search on GPUs. Proceedings of the 2016 International Conference on Management of Data. ACM. 
 13 |  *
 14 |  * This file is part of iBFS.
 15 |  *
 16 |  * iBFS is free software: you can redistribute it and/or modify
 17 |  * it under the terms of the GNU General Public License as published by
 18 |  * the Free Software Foundation, either version 3 of the License, or
 19 |  * (at your option) any later version.
 20 |  *
 21 |  * iBFS is distributed in the hope that it will be useful,
 22 |  * but WITHOUT ANY WARRANTY; without even the implied warranty of
 23 |  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the 
 24 |  * GNU General Public License for more details.
 25 |  *
 26 |  * You should have received a copy of the GNU General Public License
 27 |  * along with iBFS.  If not, see <http://www.gnu.org/licenses/>.
 28 |  */
 29 | 
 30 | #include "scan.cuh"
 31 | #include "wtime.h"
 32 | #include "validate.h"
 33 | #include "gpu_ibfs.cuh"
 34 | 
 35 | //constructor
 36 | gpu_ibfs::
 37 | gpu_ibfs(
 38 | 	const graph* g,
 39 | 	index_t concurr_count,
 40 | 	depth_t sw_level,
 41 | 	index_t gpu_count,
 42 | 	index_t sml_shed,
 43 | 	index_t lrg_shed)
 44 | 	:g(g),sw_level(sw_level),gpu_count(gpu_count),
 45 | 	sml_shed(sml_shed),lrg_shed(lrg_shed),concurr_count(concurr_count)
 46 | {
 47 | 	bit_count = sizeof(comp_t)<<3;
 48 | 	joint_count = (index_t)ceil((concurr_count*1.0)/bit_count);
 49 | 	
 50 | 	std::cout<<"joint_count vs bit_count: "<<joint_count<<" "<<bit_count<<"\n";
 51 | 	cudaSetDevice(0);	
 52 | 	long gpu_bytes= 0;
 53 | 	
 54 | 	const size_t cat_index_sz 	= sizeof(index_t)*BLKS_NUM*THDS_NUM;
 55 | 	const size_t edge_sz 				= sizeof(vertex_t)*g->edge_count;
 56 | 	const size_t vert_sz 				= sizeof(vertex_t)*g->vert_count;
 57 | 	const	size_t index_sz 			= sizeof(index_t)*g->vert_count;
 58 | 	const	size_t comp_sz 				= sizeof(comp_t)*g->vert_count*joint_count;
 59 | 	const size_t src_sz					= sizeof(vertex_t)*g->src_count;
 60 | 
 61 | 	H_ERR(cudaMalloc((void **)&beg_pos_d,		index_sz));
 62 | 	H_ERR(cudaMalloc((void **)&adj_list_d, 	edge_sz));
 63 | 	gpu_bytes	+= edge_sz+(index_sz);
 64 | 		
 65 | 	H_ERR(cudaMalloc((void **)&cat_sml_off_d, cat_index_sz));
 66 | 	H_ERR(cudaMalloc((void **)&cat_mid_off_d, cat_index_sz));
 67 | 	H_ERR(cudaMalloc((void **)&cat_lrg_off_d, cat_index_sz));
 68 | 	H_ERR(cudaMalloc((void **)&cat_sml_sz_d,	cat_index_sz));
 69 | 	H_ERR(cudaMalloc((void **)&cat_mid_sz_d,	cat_index_sz));
 70 | 	H_ERR(cudaMalloc((void **)&cat_lrg_sz_d,	cat_index_sz));
 71 | 	gpu_bytes+=(cat_index_sz*6);
 72 | 	
 73 | 	H_ERR(cudaMalloc((void **)&ex_sml_q_d,	vert_sz));
 74 | 	H_ERR(cudaMalloc((void **)&ex_mid_q_d,	vert_sz));
 75 | 	H_ERR(cudaMalloc((void **)&ex_lrg_q_d,	vert_sz));
 76 | 	gpu_bytes+=(vert_sz*3);
 77 | 		
 78 | 	H_ERR(cudaMalloc((void **)&src_list_d,src_sz));
 79 | 	H_ERR(cudaMalloc((void **)&depth_comp_last,comp_sz));
 80 | 	H_ERR(cudaMalloc((void **)&depth_comp_curr,comp_sz));
 81 | 	cudaStreamCreate(&gstream);
 82 | 	
 83 | 	gpu_bytes+=(comp_sz*2)+src_sz;
 84 | 	
 85 | 	H_ERR(cudaHostAlloc((void **)&is_done,
 86 | 			sizeof(bool),cudaHostAllocMapped));
 87 | 	H_ERR(cudaHostGetDevicePointer((void **)&is_done_d,
 88 | 			(void*)is_done,0));
 89 | 	
 90 | 	//+----------------------
 91 | 	//|FOR CLASSIFICATION
 92 | 	//+----------------------------
 93 | 	H_ERR(cudaHostAlloc((void **)&ex_sml_sz,
 94 | 				sizeof(index_t),cudaHostAllocMapped));
 95 | 	H_ERR(cudaHostAlloc((void **)&ex_mid_sz,
 96 | 				sizeof(index_t),cudaHostAllocMapped));
 97 | 	H_ERR(cudaHostAlloc((void **)&ex_lrg_sz,
 98 | 				sizeof(index_t),cudaHostAllocMapped));
 99 | 	H_ERR(cudaHostGetDevicePointer((void **)&ex_sml_sz_d,
100 | 				(void*)ex_sml_sz,0));
101 | 	H_ERR(cudaHostGetDevicePointer((void **)&ex_mid_sz_d,
102 | 				(void*)ex_mid_sz,0));
103 | 	H_ERR(cudaHostGetDevicePointer((void **)&ex_lrg_sz_d,
104 | 				(void*)ex_lrg_sz,0));
105 | 	
106 | 	stream=(cudaStream_t *)malloc(sizeof(cudaStream_t)*Q_CARD);
107 | 	for(index_t j=0;j<Q_CARD; j++) cudaStreamCreate(&stream[j]);
108 | 	std::cout<<"GPU space: "<<gpu_bytes<<" byte(s)\n";
109 | 
110 | 	//copy graph data + source list to GPU
111 | 	H_ERR(cudaMemcpy(beg_pos_d,g->beg_pos,index_sz, 
112 | 			cudaMemcpyHostToDevice));
113 | 	H_ERR(cudaMemcpy(adj_list_d,g->csr,edge_sz,
114 | 			cudaMemcpyHostToDevice));
115 | 	H_ERR(cudaMemcpy(src_list_d,g->src_list,src_sz,
116 | 			cudaMemcpyHostToDevice));
117 | }
118 | 
119 | 


--------------------------------------------------------------------------------
/gpu_ibfs.cuh:
--------------------------------------------------------------------------------
  1 | /*
  2 |  * Copyright 2016 The George Washington University
  3 |  * Written by Hang Liu 
  4 |  * Directed by Prof. Howie Huang
  5 |  *
  6 |  * https://www.seas.gwu.edu/~howie/
  7 |  * Contact: iheartgraph@gmail.com
  8 |  *
  9 |  * 
 10 |  * Please cite the following paper:
 11 |  * 
 12 |  * Hang Liu, H. Howie Huang and Yang Hu. 2016. iBFS: Concurrent Breadth-First Search on GPUs. Proceedings of the 2016 International Conference on Management of Data. ACM. 
 13 |  *
 14 |  * This file is part of iBFS.
 15 |  *
 16 |  * iBFS is free software: you can redistribute it and/or modify
 17 |  * it under the terms of the GNU General Public License as published by
 18 |  * the Free Software Foundation, either version 3 of the License, or
 19 |  * (at your option) any later version.
 20 |  *
 21 |  * iBFS is distributed in the hope that it will be useful,
 22 |  * but WITHOUT ANY WARRANTY; without even the implied warranty of
 23 |  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the 
 24 |  * GNU General Public License for more details.
 25 |  *
 26 |  * You should have received a copy of the GNU General Public License
 27 |  * along with iBFS.  If not, see <http://www.gnu.org/licenses/>.
 28 |  */
 29 | 
 30 | //Graph format: Json based format: [src_id, src_weigh,[[connected_ver_0, edge_weight],[connected_ver_1, edge_weight],[connected_ver_2, edge_weight]]]
 31 | //Storage format: 
 32 | //struct{
 33 | //		int: src_ver
 34 | //		Arr: [ver_0|ver_1|ver_2|...]
 35 | //		Int: num_conn_ver
 36 | //	}
 37 | #ifndef	GPU_IBFS_H
 38 | #define	GPU_IBFS_H
 39 | 
 40 | #include <fstream>
 41 | #include <string>
 42 | #include <iostream>
 43 | #include <sstream>
 44 | #include <queue>
 45 | #include "util.h"
 46 | #include "wtime.h"
 47 | #include "graph.h"
 48 | 
 49 | class gpu_ibfs{
 50 | 	
 51 | 	//variable
 52 | public:
 53 | 	const graph *g;
 54 | 	
 55 | 	vertex_t* src_list_d;
 56 | 	vertex_t*	adj_list_d;
 57 | 	index_t*	beg_pos_d;
 58 | 	vertex_t*	ex_sml_q_d;
 59 | 	vertex_t*	ex_mid_q_d;
 60 | 	vertex_t*	ex_lrg_q_d;
 61 | 	index_t*	cat_sml_sz_d;
 62 | 	index_t*	cat_mid_sz_d;
 63 | 	index_t*	cat_lrg_sz_d;
 64 | 	index_t*	cat_sml_off_d;
 65 | 	index_t*	cat_mid_off_d;
 66 | 	index_t*	cat_lrg_off_d;
 67 | 	
 68 | 	index_t	*ex_sml_sz;
 69 | 	index_t	*ex_mid_sz;
 70 | 	index_t	*ex_lrg_sz;
 71 | 	index_t	*ex_sml_sz_d;
 72 | 	index_t	*ex_mid_sz_d;
 73 | 	index_t	*ex_lrg_sz_d;
 74 | 	
 75 | 	comp_t	*depth_comp_last;
 76 | 	comp_t	*depth_comp_curr;
 77 | 	cudaStream_t *stream, gstream;
 78 | 	
 79 | 	index_t	sml_shed;
 80 | 	index_t	lrg_shed;
 81 | 	index_t	gpu_count;
 82 | 	index_t	*gpu_ranger;
 83 | 	index_t joint_count;
 84 | 	index_t bit_count;
 85 | 	index_t concurr_count;
 86 | 	index_t	sw_level;
 87 | 	bool *is_done;
 88 | 	bool *is_done_d;
 89 | 
 90 | 	//constructor
 91 | public:
 92 | 	gpu_ibfs(){};
 93 | 	gpu_ibfs(	
 94 | 		const graph *g,
 95 | 		index_t concurr_count,
 96 | 		depth_t			sw_level,
 97 | 		index_t			gpu_count,
 98 | 		index_t			sml_shed,
 99 | 		index_t			lrg_shed);
100 | 	~gpu_ibfs(){};
101 | 
102 | 	//functions
103 | public:
104 | 	int write_result();
105 | 	
106 | 	void traversal(index_t src_grp_off);
107 | 	void init_bfs(index_t src_grp_off);
108 | 	void debugger(index_t);
109 | 
110 | 	//expander series
111 | 	void expander(depth_t level);
112 | 	void bu_expand(depth_t level);
113 | 	void td_expand(depth_t level);
114 | 	void sw_expand(depth_t level);
115 | 
116 | 	//inspector series
117 | 	void inspector(depth_t level);
118 | 	void td_inspect(depth_t level);
119 | 	void sw_inspect(depth_t level);
120 | };
121 | #endif
122 | 


--------------------------------------------------------------------------------
/graph.cu:
--------------------------------------------------------------------------------
  1 | /*
  2 |  * Copyright 2016 The George Washington University
  3 |  * Written by Hang Liu 
  4 |  * Directed by Prof. Howie Huang
  5 |  *
  6 |  * https://www.seas.gwu.edu/~howie/
  7 |  * Contact: iheartgraph@gmail.com
  8 |  *
  9 |  * 
 10 |  * Please cite the following paper:
 11 |  * 
 12 |  * Hang Liu, H. Howie Huang and Yang Hu. 2016. iBFS: Concurrent Breadth-First Search on GPUs. Proceedings of the 2016 International Conference on Management of Data. ACM. 
 13 |  *
 14 |  * This file is part of iBFS.
 15 |  *
 16 |  * iBFS is free software: you can redistribute it and/or modify
 17 |  * it under the terms of the GNU General Public License as published by
 18 |  * the Free Software Foundation, either version 3 of the License, or
 19 |  * (at your option) any later version.
 20 |  *
 21 |  * iBFS is distributed in the hope that it will be useful,
 22 |  * but WITHOUT ANY WARRANTY; without even the implied warranty of
 23 |  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the 
 24 |  * GNU General Public License for more details.
 25 |  *
 26 |  * You should have received a copy of the GNU General Public License
 27 |  * along with iBFS.  If not, see <http://www.gnu.org/licenses/>.
 28 |  */
 29 | 
 30 | #include "graph.h"
 31 | 
 32 | inline bool deep_than_two(
 33 | 	vertex_t src,
 34 | 	depth_t *depth,
 35 | 	index_t *beg_pos,
 36 | 	vertex_t *csr
 37 | ){
 38 | 	depth[src]=0;
 39 | 	
 40 | 	//two level DFS
 41 | 	index_t beg=beg_pos[src];
 42 | 	index_t end=beg_pos[src+1];
 43 | 	for(;beg<end;++beg)
 44 | 	{
 45 | 		vertex_t nebr=csr[beg];
 46 | 		if(depth[nebr] == INFTY)
 47 | 		{
 48 | 			depth[nebr]=1;
 49 | 			index_t nbeg=beg_pos[nebr];
 50 | 			index_t nend=beg_pos[nebr+1];
 51 | 			for(;nbeg<nend;++nbeg)
 52 | 			{
 53 | 				vertex_t nnebr=csr[nbeg];
 54 | 				if(depth[nnebr]==INFTY)
 55 | 					return true;
 56 | 			}
 57 | 		}
 58 | 	}
 59 | 	return false;
 60 | }
 61 | 
 62 | 
 63 | graph::graph(
 64 | 		const char *beg_file,
 65 | 		const char *csr_file,
 66 | 		const index_t src_count
 67 | ):src_count(src_count)
 68 | {
 69 | 	double tm=wtime();
 70 | 	
 71 | 	typedef uint64_t index_tt;
 72 | 	typedef uint64_t vertex_tt;
 73 | 
 74 | 	//typedef uint32_t index_tt;
 75 | 	//typedef uint32_t vertex_tt;
 76 | 
 77 | 	src_list = new vertex_t[src_count];
 78 | 	vert_count=fsize(beg_file)/sizeof(index_tt) - 1;
 79 | 	edge_count=fsize(csr_file)/sizeof(vertex_tt);
 80 | 	FILE *file=fopen(beg_file, "rb");
 81 | 	if(file==NULL)
 82 | 	{
 83 | 		std::cout<<beg_file<<" cannot open\n";
 84 | 		exit(-1);
 85 | 	}
 86 | 
 87 | 	index_tt *tmp_beg_pos = new index_tt[vert_count+1];
 88 | 	index_tt ret=fread(tmp_beg_pos, sizeof(index_tt), vert_count+1, file);
 89 | 	assert(ret==vert_count+1);
 90 | 	fclose(file);
 91 | 
 92 | 	file=fopen(csr_file, "rb");
 93 | 	if(file==NULL)
 94 | 	{
 95 | 		std::cout<<csr_file<<" cannot open\n";
 96 | 		exit(-1);
 97 | 	}
 98 | 
 99 | 	vertex_tt *tmp_csr = new vertex_tt[edge_count];
100 | 	ret=fread(tmp_csr, sizeof(vertex_tt), edge_count, file);
101 | 	assert(ret==edge_count);
102 | 	fclose(file);
103 | 	
104 | 	//converting to uint32_t
105 | 	beg_pos = new index_t[vert_count+1];
106 | 	csr	= new vertex_t[edge_count];
107 | 	
108 | 	for(index_t i=0;i<vert_count+1;++i)
109 | 		beg_pos[i]=(index_t)tmp_beg_pos[i];
110 | 
111 | 	for(index_t i=0;i<edge_count;++i)
112 | 		csr[i]=(vertex_t)tmp_csr[i];
113 | 	
114 | 	delete[] tmp_beg_pos;
115 | 	delete[] tmp_csr;
116 | 
117 | 	std::cout<<"Graph load (success): "<<vert_count<<" verts, "
118 | 		<<edge_count<<" edges "<<wtime()-tm<<" second(s)\n";
119 | }
120 | 
121 | void
122 | graph::
123 | gen_src()
124 | {
125 | 	depth_t *depth=new depth_t[vert_count];
126 | 	for(index_t i=0;i<src_count;++i)
127 | 	{
128 | 		src_list[i] = rand()%vert_count;
129 | 		memset(depth, INFTY, sizeof(depth_t)*vert_count);
130 | 
131 | 		//no orphan
132 | 		if(!deep_than_two(src_list[i],depth,beg_pos,csr))
133 | 		{
134 | 			--i;
135 | 			continue;
136 | 		}
137 | 
138 | 		//no duplication
139 | 		for(index_t j=0;j<i;++j)
140 | 			if(src_list[i]==src_list[j])
141 | 			{
142 | 				--i;
143 | 				break;
144 | 			}
145 | 	}
146 | 
147 | 	delete[] depth;
148 | }
149 | 


--------------------------------------------------------------------------------
/graph.cuh:
--------------------------------------------------------------------------------
  1 | //Graph format: 
  2 | //Simplified json format: 
  3 | //src degree dest0 dest1 ...
  4 | 
  5 | 
  6 | //#include "graph.h"
  7 | #include <fstream>
  8 | #include <omp.h>
  9 | #include "GroupBy.cuh"
 10 | 
 11 | #define FILE_NOT_EXIST		1
 12 | #define FILE_EXIST				0
 13 | 
 14 | template<typename index_t>
 15 | void progress(
 16 | 		index_t &report,
 17 | 		index_t beg_time,
 18 | 		int			tid,
 19 | 		index_t prc_line
 20 | ){
 21 | 	if(prc_line>report && tid==0){
 22 | 			std::cout<<prc_line<<" lines\t"
 23 | 					<<"time: "<<wtime()-beg_time<<" seconds\n";
 24 | 		report<<=1;
 25 | 	}
 26 | }
 27 | 
 28 | template<typename vertex_t, typename index_t>
 29 | bool deep_than_two(
 30 | 	vertex_t src,
 31 | 	vertex_t *currq,
 32 | 	vertex_t *nextq,
 33 | 	depth_t *depth,
 34 | 	index_t *adj_card,
 35 | 	index_t *beg_pos,
 36 | 	vertex_t *adj_list
 37 | ){
 38 | 	
 39 | 	index_t curr_sz=1;
 40 | 	index_t next_sz=0;
 41 | 	currq[0]=src;
 42 | 	depth[src]=0;
 43 | 	//sequential check if traversal larger than 2
 44 | 	for(index_t level=0;;level++)
 45 | 	{
 46 | 		next_sz=0;
 47 | 		for(index_t i=0;i<curr_sz;i++)
 48 | 		{
 49 | 			vertex_t frontier=currq[i];
 50 | 			index_t adj_off=beg_pos[frontier];
 51 | 			for(index_t j=0;j<adj_card[frontier];j++)
 52 | 			{
 53 | 				vertex_t neighbor=adj_list[adj_off+j];
 54 | 				if(depth[neighbor]==INFTY)
 55 | 				{
 56 | 					nextq[next_sz]=neighbor;
 57 | 					depth[neighbor]=0;
 58 | 					next_sz++;
 59 | 				}
 60 | 			}
 61 | 		}
 62 | 		
 63 | 		if(level==1)
 64 | 			if(next_sz!=0) return false;
 65 | 			else return true;
 66 | 		
 67 | 		memcpy(currq,nextq,sizeof(vertex_t)*next_sz);
 68 | 		curr_sz=next_sz;
 69 | 	}
 70 | }
 71 | 
 72 | 
 73 | template<	typename vertex_t,
 74 | 			typename index_t>
 75 | void arrange_src
 76 | (
 77 | 	int world_sz,
 78 | 	int my_id,
 79 | 	vertex_t *glb_src_list, 
 80 | 	index_t glb_src_count, 
 81 | 	vertex_t* &src_list,
 82 | 	index_t &src_count){
 83 | 	
 84 | 	src_count=glb_src_count/world_sz;
 85 | 	if(glb_src_count != src_count * world_sz){
 86 | 		std::cout<<"Work not even!\n";
 87 | 		exit(-1);
 88 | 	}
 89 | 
 90 | 	src_list=new vertex_t[src_count];
 91 | 	int src_off=my_id*src_count;
 92 | 
 93 | 	memcpy(src_list,glb_src_list+src_off, 
 94 | 	sizeof(vertex_t)*src_count);
 95 | }
 96 | 
 97 | 
 98 | 
 99 | template<	typename vertex_t,
100 | 			typename index_t>
101 | int read_src_v(	
102 | 	vertex_t* 	&glb_src_list,
103 | 	index_t			&glb_src_count,
104 | 	std::string filename
105 | )
106 | {
107 | 	std::stringstream ss;
108 | 	std::string str;
109 | 	std::ifstream myfile(filename.c_str());
110 | 	
111 | 	//count num source in input.dat
112 | 	//one line is a source
113 | 	if(!myfile.good()) return FILE_NOT_EXIST;
114 | 	glb_src_count=0;
115 | 	while(std::getline(myfile, str))
116 | 		glb_src_count++;
117 | 	myfile.close();
118 | 	//std::cout<<"Read input.dat: "<<glb_src_count<<"\n";
119 | 
120 | 	glb_src_list	= new vertex_t[glb_src_count];
121 | 	index_t ptr	= 0;
122 | 	myfile.open(filename.c_str());
123 | 	while(std::getline(myfile, str)){
124 | 		ss.str("");
125 | 		ss.clear();
126 | 		ss<<str;
127 | 		ss>>glb_src_list[ptr];
128 | 		ptr++;
129 | 	}
130 | 	myfile.close();
131 | 
132 | 	return FILE_EXIST;
133 | }
134 | 
135 | //this function aims to get the vertex list from ss
136 | template <typename vertex_t, typename index_t>
137 | int proc_json(
138 | 	std::string kron_addr, 
139 | 	index_t			vert_count,
140 | 	index_t			edge_count,
141 | 	vertex_t*		adj_list,
142 | 	index_t*		adj_card,
143 | 	index_t*		beg_pos)
144 | {
145 | 	
146 | 	double beg_time;
147 | 	std::stringstream ss;
148 | 	std::string		sline;
149 | 	index_t i;
150 | 	index_t *card_sum		= new index_t[OMP_THDS];
151 | 	beg_time=wtime();
152 | 	//read card and formulate beg_pos 
153 | 	std::cout<<"\n\nReading cardinality ... \n";
154 | 	#pragma omp parallel \
155 | 	private(i) num_threads(OMP_THDS)
156 | 	{
157 | 		index_t report=1;
158 | 		index_t prc_line=0;
159 | 		vertex_t beg_vert,end_vert,vert_id;
160 | 		std::string line;
161 | 		std::stringstream ss;
162 | 		int tid=omp_get_thread_num();
163 | 		ss.str("");
164 | 		ss.clear();
165 | 
166 | 		//+---------
167 | 		//Assuming big file is partioned into small 
168 | 		//	files continuously and named sequentially.
169 | 		//+----
170 | 		//example 1,2,3,4,5 partioned into three files
171 | 		//file-0: 1,2; file-1: 3,4; file-2: 5;
172 | 		//+---------------------------------------
173 | 		ss<<kron_addr<<"/kron."<<tid<<".dat";
174 | 		std::string filename=ss.str();
175 | 		card_sum[tid]=0;
176 | 		std::ifstream file(filename.c_str());
177 | 		i=0;beg_vert=0;end_vert=0;
178 | 		if(file.is_open()){
179 | 			while(std::getline(file, line)){
180 | 				progress<index_t>(report,beg_time,tid,prc_line);
181 | 				ss.str("");
182 | 				ss.clear();
183 | 				ss<<line;
184 | 				ss>>vert_id;
185 | 				ss>>adj_card[vert_id];
186 | 				card_sum[tid]+=adj_card[vert_id];
187 | 				if(prc_line==0) beg_vert=vert_id;
188 | 				prc_line++;
189 | 			}
190 | 			end_vert=vert_id;
191 | 			file.close();
192 | 		}else{std::cout<<"card file wrong\n";exit(-4);}
193 | 		#pragma omp barrier
194 | 		
195 | 		beg_pos[beg_vert]=0;
196 | 		for(i=0;i<tid;i++)
197 | 			beg_pos[beg_vert]	+= card_sum[i];
198 | 
199 | 		for(i=beg_vert+1;i<=end_vert;i++)
200 | 			beg_pos[i]=beg_pos[i-1]+adj_card[i-1];
201 | 	}
202 | 	
203 | 	delete[] card_sum;
204 | 	std::cout<<"\n\nReading adj_list ... \n";
205 | 	#pragma omp parallel \
206 | 	private(i) num_threads(OMP_THDS)
207 | 	{
208 | 		index_t report=1;
209 | 		index_t prc_line=0;
210 | 		index_t temp;
211 | 		vertex_t vert_id;
212 | 		std::string line;
213 | 		std::stringstream ss;
214 | 		int tid=omp_get_thread_num();
215 | 		ss.str("");
216 | 		ss.clear();
217 | 		
218 | 		ss<<kron_addr<<"/kron."<<tid<<".dat";
219 | 		std::string filename=ss.str();
220 | 		std::ifstream file(filename.c_str());
221 | 		i=0;
222 | 		if(file.is_open()){
223 | 			while(std::getline(file, line)){
224 | 				progress<index_t>(report,beg_time,tid,prc_line);
225 | 				//top-down
226 | 				ss.str("");
227 | 				ss.clear();
228 | 				ss<<line;
229 | 				ss>>vert_id;
230 | 				ss>>temp;//card
231 | 				if(temp!=beg_pos[vert_id+1]-beg_pos[vert_id]){
232 | 					std::cout<<tid<<" Wrong 1\n"
233 | 					<<"Vert: "<<vert_id<<"\n"
234 | 					<<"Expect: "<<beg_pos[vert_id+1]-beg_pos[vert_id]
235 | 					<<"\tGot: "<<temp<<"\n";exit(-1);}
236 | 				for(i=0;i<beg_pos[vert_id+1]-beg_pos[vert_id];i++)
237 | 					ss>>adj_list[beg_pos[vert_id]+i];
238 | 				prc_line++;
239 | 			}
240 | 			file.close();
241 | 		}else{std::cout<<"kron file wrong\n";exit(-4);}
242 | 	}
243 | 	return 0;
244 | }
245 | 
246 | template <typename vertex_t, typename index_t, typename depth_t>
247 | graph<vertex_t, index_t, depth_t>
248 | ::graph(
249 | 	graph_reader *g,
250 | 		index_t			sw_level,
251 | 	index_t			vert_count,
252 | 	index_t			edge_count,
253 | 	index_t			num_groups,
254 | 	index_t			num_agg_bfs,
255 | 	int world_sz, int my_id,
256 | 	index_t			gpu_id,
257 | 	index_t			sml_shed,
258 | 	index_t			lrg_shed
259 | ):gpu_id(gpu_id),vert_count(vert_count),edge_count(edge_count),
260 | 	num_agg_bfs(num_agg_bfs),sml_shed(sml_shed),lrg_shed(lrg_shed),
261 | 	sw_level(sw_level), world_sz(world_sz), my_id(my_id)
262 | {
263 | 	
264 | 	index_t glb_src_count;
265 | 	vertex_t *glb_src_list;
266 | 
267 | 	//traversal data
268 | 	gdata = new tdata[num_agg_bfs];
269 | 	for(index_t i=0;i<num_agg_bfs;i++)
270 | 		gdata[i]=new Traverse;
271 | 	
272 | 	//gpu graph
273 | 	ggraph=new CSR_Graph;
274 | 	ggraph->vert_count=g->vert_count;
275 | 	ggraph->edge_count=g->edge_count;
276 | 	ggraph->adj_list=g->csr;
277 | 	ggraph->adj_card=new index_t[vert_count];
278 | 	ggraph->beg_pos=g->beg_pos;
279 | 	EDGES_C = edge_count;//bad programming style, related to direction-switch
280 | 
281 | 	
282 | 	for(vertex_t i = 0; i < g->vert_count; i++)
283 | 		ggraph->adj_card[i] = g->beg_pos[i+1] - g->beg_pos[i];
284 | 
285 | 	double timer=wtime();
286 | 	std::cout<<"Loading edges ... \n";
287 | 	//proc_json<vertex_t, index_t>(
288 | 	//		kron_addr,
289 | 	//		vert_count,
290 | 	//		edge_count,
291 | 	//		ggraph->adj_list,
292 | 	//		ggraph->adj_card,
293 | 	//		ggraph->beg_pos);
294 | 	std::cout<<"Loading edges time: "<<wtime()-timer<<" second(s)\n";
295 | 	std::cout<<"Assuming source vertices are stored in input.dat\n";
296 | 	if(true){
297 | 		std::cout<<"input.dat is not detected.\n"
298 | 		<<"Generating source vertices with rand()%vert_count\n";
299 | 	
300 | 		vertex_t *currq=new vertex_t[vert_count];
301 | 		vertex_t *nextq=new vertex_t[vert_count];
302 | 		depth_t *depth=new depth_t[vert_count];
303 | 		
304 | 		//generate glb_src_list
305 | 		glb_src_count = num_groups*num_agg_bfs;
306 | 		std::cout<<glb_src_count<<"\n";
307 | 		glb_src_list 	= new vertex_t[glb_src_count];
308 | 		//std::cout<<"haha\n";
309 | 		for(index_t i = 0; i<glb_src_count; i++){
310 | 			glb_src_list[i] = rand()%vert_count;
311 | 			
312 | 			//Not orphan
313 | 			//And level more than two
314 | 			memset(depth,INFTY,sizeof(depth_t)*vert_count);
315 | 			if(deep_than_two<vertex_t,index_t>
316 | 				(glb_src_list[i],currq,nextq,depth,
317 | 				ggraph->adj_card,ggraph->beg_pos,ggraph->adj_list)){
318 | 				--i;continue;
319 | 			}
320 | 
321 | 			//Not repeated
322 | 			for(index_t j = 0; j< i; j++)
323 | 				if(glb_src_list[j] == glb_src_list[i]){
324 | 					--i;continue;
325 | 				}
326 | 		}
327 | 
328 | 		delete[] currq;
329 | 		delete[] nextq;
330 | 		delete[] depth;
331 | 	};
332 | 	
333 | 	#ifdef GROUPBY
334 | 	std::cout<<"before\n";
335 | 	vertex_t *glb_new_list=new vertex_t[glb_src_count];
336 | 	std::cout<<"after\n";
337 | 	GroupBy<vertex_t,index_t>(glb_src_count,glb_src_list,glb_new_list,ggraph);	
338 | 	memcpy(glb_src_list,glb_new_list,sizeof(vertex_t)*glb_src_count);
339 | 	#endif
340 | 	
341 | 	arrange_src<vertex_t, index_t>(
342 | 	world_sz, my_id, glb_src_list, glb_src_count, src_list, src_count);
343 | 	std::cout<<"Global-source-count vs src_count: "<<glb_src_count<<" vs "<<src_count<<"\n";
344 | 
345 | 	//alloc depth and parent array
346 | 	if(cudaMallocHost((void **)&depth, sizeof(depth_t)*vert_count))
347 | 		std::cout<<"Alloc depth wrong\n";
348 | 	parent = new vertex_t[vert_count];
349 | 
350 | 	#ifdef DEBUG_CONSTRUCTOR
351 | 	std::cout<<"\n";
352 | 	for(int j=0;j<num_lines;j++){
353 | 		std::cout<<"Vertex "<<vertex_list[j]->src_ver
354 | 			<<"("<<vertex_list[j]->num_conn_ver
355 | 			<<","<<vertex_list[j]->depth<<"): ";
356 | 		for(int i=0;i<vertex_list[j]->num_conn_ver;i++)
357 | 			std::cout<<vertex_list[j]->conn_ver_list[i]<<",";
358 | 		std::cout<<"\n";
359 | 	}
360 | 	#endif
361 | 
362 | }
363 | 


--------------------------------------------------------------------------------
/graph.h:
--------------------------------------------------------------------------------
 1 | //Graph format: Json based format: [src_id, src_weigh,[[connected_ver_0, edge_weight],[connected_ver_1, edge_weight],[connected_ver_2, edge_weight]]]
 2 | //Storage format: 
 3 | //struct{
 4 | //		int: src_ver
 5 | //		Arr: [ver_0|ver_1|ver_2|...]
 6 | //		Int: num_conn_ver
 7 | //	}
 8 | #ifndef	GRAPH_H
 9 | #define	GRAPH_H
10 | 
11 | #include <fstream>
12 | #include <string>
13 | #include <iostream>
14 | #include <sstream>
15 | #include <queue>
16 | #include "comm.h"
17 | #include "graph_reader.h"
18 | 
19 | template 
20 | <	typename vertex_t, 
21 | 	typename index_t, 
22 | 	typename depth_t>
23 | class graph{
24 | 	
25 | 	//variable
26 | public:
27 | 	vertex_t	*src_list;
28 | 	index_t		src_count;
29 | 	index_t 	edge_count;
30 | 	index_t 	vert_count;
31 | 	depth_t		*depth;
32 | 	vertex_t	*parent;
33 | 	
34 | 	csr_graph ggraph;
35 | 	tdata 		*gdata;
36 | 	index_t		sml_shed;
37 | 	index_t		lrg_shed;
38 | 	index_t		gpu_id;
39 | 	index_t		*gpu_ranger;
40 | 	index_t 	num_agg_bfs;
41 | 	index_t		sw_level;
42 | 	
43 | 	int world_sz, my_id;
44 | 	//constructor
45 | public:
46 | 	graph() {};
47 | 	graph(	
48 | 		graph_reader * g,
49 | 			index_t			sw_level,
50 | 		index_t			vert_count,
51 | 		index_t			edge_count,
52 | 		index_t			num_groups,
53 | 		index_t			num_agg_bfs,
54 | 		int world_sz, int my_id,
55 | 		index_t			gpu_id,
56 | 		index_t			sml_shed,
57 | 		index_t			lrg_shed);
58 | 
59 | 	//functions
60 | public:
61 | 	int write_result();
62 | 	int bfs_gpu_coalescing_mem();
63 | 	int alloc_array();
64 | 
65 | };
66 | 
67 | #include "bfs_gpu_opt.cuh"
68 | #include "graph.cuh"
69 | #include "write_result.cuh"
70 | #include "allocator.cuh"
71 | #endif
72 | //included in 	main.cpp
73 | //		bfs_seq.cpp
74 | //		bfs.cu
75 | 


--------------------------------------------------------------------------------
/graph_reader.cuh:
--------------------------------------------------------------------------------
  1 | /*
  2 |  * Copyright 2016 The George Washington University
  3 |  * Written by Hang Liu 
  4 |  * Directed by Prof. Howie Huang
  5 |  *
  6 |  * https://www.seas.gwu.edu/~howie/
  7 |  * Contact: iheartgraph@gmail.com
  8 |  *
  9 |  * 
 10 |  * Please cite the following paper:
 11 |  * 
 12 |  * Hang Liu, H. Howie Huang and Yang Hu. 2016. iBFS: Concurrent Breadth-First Search on GPUs. Proceedings of the 2016 International Conference on Management of Data. ACM. 
 13 |  *
 14 |  * This file is part of iBFS.
 15 |  *
 16 |  * iBFS is free software: you can redistribute it and/or modify
 17 |  * it under the terms of the GNU General Public License as published by
 18 |  * the Free Software Foundation, either version 3 of the License, or
 19 |  * (at your option) any later version.
 20 |  *
 21 |  * iBFS is distributed in the hope that it will be useful,
 22 |  * but WITHOUT ANY WARRANTY; without even the implied warranty of
 23 |  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the 
 24 |  * GNU General Public License for more details.
 25 |  *
 26 |  * You should have received a copy of the GNU General Public License
 27 |  * along with iBFS.  If not, see <http://www.gnu.org/licenses/>.
 28 |  */
 29 | 
 30 | //#include "graph_reader.h"
 31 | 
 32 | inline bool deep_than_two(
 33 | 	vertex_t src,
 34 | 	depth_t *depth,
 35 | 	index_t *beg_pos,
 36 | 	vertex_t *csr
 37 | ){
 38 | 	depth[src]=0;
 39 | 	
 40 | 	//two level DFS
 41 | 	index_t beg=beg_pos[src];
 42 | 	index_t end=beg_pos[src+1];
 43 | 	for(;beg<end;++beg)
 44 | 	{
 45 | 		vertex_t nebr=csr[beg];
 46 | 		if(depth[nebr] == INFTY)
 47 | 		{
 48 | 			depth[nebr]=1;
 49 | 			index_t nbeg=beg_pos[nebr];
 50 | 			index_t nend=beg_pos[nebr+1];
 51 | 			for(;nbeg<nend;++nbeg)
 52 | 			{
 53 | 				vertex_t nnebr=csr[nbeg];
 54 | 				if(depth[nnebr]==INFTY)
 55 | 					return true;
 56 | 			}
 57 | 		}
 58 | 	}
 59 | 	return false;
 60 | }
 61 | 
 62 | 
 63 | graph_reader::graph_reader(
 64 | 		const char *beg_file,
 65 | 		const char *csr_file,
 66 | 		const index_t src_count
 67 | ):src_count(src_count)
 68 | {
 69 | 	double tm=wtime();
 70 | 	
 71 | 	typedef uint64_t index_tt;
 72 | 	typedef uint64_t vertex_tt;
 73 | 
 74 | 	//typedef uint32_t index_tt;
 75 | 	//typedef uint32_t vertex_tt;
 76 | 
 77 | 	src_list = new vertex_t[src_count];
 78 | 	vert_count=fsize(beg_file)/sizeof(index_tt) - 1;
 79 | 	edge_count=fsize(csr_file)/sizeof(vertex_tt);
 80 | 	FILE *file=fopen(beg_file, "rb");
 81 | 	if(file==NULL)
 82 | 	{
 83 | 		std::cout<<beg_file<<" cannot open\n";
 84 | 		exit(-1);
 85 | 	}
 86 | 
 87 | 	index_tt *tmp_beg_pos = new index_tt[vert_count+1];
 88 | 	index_tt ret=fread(tmp_beg_pos, sizeof(index_tt), vert_count+1, file);
 89 | 	assert(ret==vert_count+1);
 90 | 	fclose(file);
 91 | 
 92 | 	file=fopen(csr_file, "rb");
 93 | 	if(file==NULL)
 94 | 	{
 95 | 		std::cout<<csr_file<<" cannot open\n";
 96 | 		exit(-1);
 97 | 	}
 98 | 
 99 | 	vertex_tt *tmp_csr = new vertex_tt[edge_count];
100 | 	ret=fread(tmp_csr, sizeof(vertex_tt), edge_count, file);
101 | 	assert(ret==edge_count);
102 | 	fclose(file);
103 | 	
104 | 	//converting to uint32_t
105 | 	beg_pos = new index_t[vert_count+1];
106 | 	csr	= new vertex_t[edge_count];
107 | 	
108 | 	for(index_t i=0;i<vert_count+1;++i)
109 | 		beg_pos[i]=(index_t)tmp_beg_pos[i];
110 | 
111 | 	for(index_t i=0;i<edge_count;++i)
112 | 		csr[i]=(vertex_t)tmp_csr[i];
113 | 	
114 | 	delete[] tmp_beg_pos;
115 | 	delete[] tmp_csr;
116 | 
117 | 	std::cout<<"Graph load (success): "<<vert_count<<" verts, "
118 | 		<<edge_count<<" edges "<<wtime()-tm<<" second(s)\n";
119 | }
120 | 
121 | void
122 | graph_reader::
123 | gen_src()
124 | {
125 | 	depth_t *depth=new depth_t[vert_count];
126 | 	for(index_t i=0;i<src_count;++i)
127 | 	{
128 | 		src_list[i] = rand()%vert_count;
129 | 		memset(depth, INFTY, sizeof(depth_t)*vert_count);
130 | 
131 | 		//no orphan
132 | 		if(!deep_than_two(src_list[i],depth,beg_pos,csr))
133 | 		{
134 | 			--i;
135 | 			continue;
136 | 		}
137 | 
138 | 		//no duplication
139 | 		for(index_t j=0;j<i;++j)
140 | 			if(src_list[i]==src_list[j])
141 | 			{
142 | 				--i;
143 | 				break;
144 | 			}
145 | 	}
146 | 
147 | 	delete[] depth;
148 | }
149 | 


--------------------------------------------------------------------------------
/graph_reader.h:
--------------------------------------------------------------------------------
 1 | /*
 2 |  * Copyright 2016 The George Washington University
 3 |  * Written by Hang Liu 
 4 |  * Directed by Prof. Howie Huang
 5 |  *
 6 |  * https://www.seas.gwu.edu/~howie/
 7 |  * Contact: iheartgraph@gmail.com
 8 |  *
 9 |  * 
10 |  * Please cite the following paper:
11 |  * 
12 |  * Hang Liu, H. Howie Huang and Yang Hu. 2016. iBFS: Concurrent Breadth-First Search on GPUs. Proceedings of the 2016 International Conference on Management of Data. ACM. 
13 |  *
14 |  * This file is part of iBFS.
15 |  *
16 |  * iBFS is free software: you can redistribute it and/or modify
17 |  * it under the terms of the GNU General Public License as published by
18 |  * the Free Software Foundation, either version 3 of the License, or
19 |  * (at your option) any later version.
20 |  *
21 |  * iBFS is distributed in the hope that it will be useful,
22 |  * but WITHOUT ANY WARRANTY; without even the implied warranty of
23 |  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the 
24 |  * GNU General Public License for more details.
25 |  *
26 |  * You should have received a copy of the GNU General Public License
27 |  * along with iBFS.  If not, see <http://www.gnu.org/licenses/>.
28 |  */
29 | 
30 | #ifndef __GRAPH_READER_H__
31 | #define __GRAPH_READER_H__
32 | #include "comm.h"
33 | #include <iostream>
34 | #include <stdio.h>
35 | #include <assert.h>
36 | #include <string.h>
37 | #include "wtime.h"
38 | #include <stdint.h>
39 | #include <sys/stat.h>
40 | #include <stdio.h>
41 | 
42 | inline off_t fsize(const char *filename) {
43 | 	struct stat st; 
44 | 	if (stat(filename, &st) == 0)
45 | 		return st.st_size;
46 | 	return -1; 
47 | }
48 | 
49 | 
50 | class graph_reader
51 | {
52 | 	public:
53 | 		index_t *beg_pos;
54 | 		vertex_t *csr;
55 | 		index_t vert_count;
56 | 		index_t edge_count;
57 | 		vertex_t *src_list;
58 | 		index_t src_count;
59 | 
60 | 	public:
61 | 		graph_reader(){};
62 | 		~graph_reader(){};
63 | 		graph_reader(const char *beg_file, 
64 | 				const char *csr_file,
65 | 				index_t src_count);
66 | 
67 | 		void gen_src();
68 | 		void groupby();
69 | };
70 | #include "graph_reader.cuh"
71 | #endif
72 | 


--------------------------------------------------------------------------------
/groupby.cu:
--------------------------------------------------------------------------------
  1 | /*
  2 |  * Copyright 2016 The George Washington University
  3 |  * Written by Hang Liu 
  4 |  * Directed by Prof. Howie Huang
  5 |  *
  6 |  * https://www.seas.gwu.edu/~howie/
  7 |  * Contact: iheartgraph@gmail.com
  8 |  *
  9 |  * 
 10 |  * Please cite the following paper:
 11 |  * 
 12 |  * Hang Liu, H. Howie Huang and Yang Hu. 2016. iBFS: Concurrent Breadth-First Search on GPUs. Proceedings of the 2016 International Conference on Management of Data. ACM. 
 13 |  *
 14 |  * This file is part of iBFS.
 15 |  *
 16 |  * iBFS is free software: you can redistribute it and/or modify
 17 |  * it under the terms of the GNU General Public License as published by
 18 |  * the Free Software Foundation, either version 3 of the License, or
 19 |  * (at your option) any later version.
 20 |  *
 21 |  * iBFS is distributed in the hope that it will be useful,
 22 |  * but WITHOUT ANY WARRANTY; without even the implied warranty of
 23 |  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the 
 24 |  * GNU General Public License for more details.
 25 |  *
 26 |  * You should have received a copy of the GNU General Public License
 27 |  * along with iBFS.  If not, see <http://www.gnu.org/licenses/>.
 28 |  */
 29 | 
 30 | #include "util.h"
 31 | #include "graph.h"
 32 | #include <iostream>
 33 | #include <fstream>
 34 | 
 35 | 
 36 | void
 37 | graph::
 38 | groupby()
 39 | {
 40 | 	vertex_t *new_list = new index_t[src_count];
 41 | 	index_t *hubmap=new index_t[vert_count];
 42 | 	index_t *huboff=new index_t[vert_count];
 43 | 	memset(hubmap,0,sizeof(index_t)*vert_count);
 44 | 	memset(huboff,0,sizeof(index_t)*vert_count);
 45 | 	
 46 | 	//-divide source into connected to hub
 47 | 	//-not connected to half by "remainoff"
 48 | 	/*counting eligible instances*/
 49 | 	for(index_t i=0;i<src_count;i++)
 50 | 	{
 51 | 		index_t beg=beg_pos[src_list[i]];
 52 | 		index_t end=beg_pos[src_list[i]+1];
 53 | 		
 54 | 		/*out-degree <= sdegree*/
 55 | 		if(end-beg>128) continue; 
 56 | 		for(index_t j=beg;j<end;j++)
 57 | 		{
 58 | 			/*has hub neighbor*/
 59 | 			if(beg_pos[csr[j]+1]-
 60 | 				beg_pos[csr[j]] > 128)
 61 | 			{
 62 | 				hubmap[csr[j]]++;
 63 | 				break;
 64 | 			}
 65 | 		}
 66 | 	}
 67 | 	
 68 | 	/*prefix-sum counted data*/
 69 | 	huboff[0]=0;
 70 | 	for(index_t i=1;i<vert_count;i++)
 71 | 		huboff[i]=huboff[i-1]+hubmap[i-1];
 72 | 	index_t remainoff=huboff[vert_count-1]
 73 | 			+hubmap[vert_count-1];
 74 | 	
 75 | 	index_t test=0;
 76 | 	/*formulating group*/
 77 | 	for(index_t i=0;i<src_count;i++)
 78 | 	{
 79 | 		index_t beg=beg_pos[src_list[i]];
 80 | 		index_t end=beg_pos[src_list[i]+1];
 81 | 		
 82 | 		//no hub neighbor, put it after "remainoff"
 83 | 		/*non-eligible instance*/
 84 | 		if(end-beg>128)
 85 | 		{
 86 | 			new_list[remainoff]=src_list[i];
 87 | 			remainoff++;
 88 | 			test++;
 89 | 			continue;
 90 | 		}
 91 | 		
 92 | 		/*eligible instance*/
 93 | 		bool isEligible=false;
 94 | 		for(index_t j=beg;j<end;j++)
 95 | 		{
 96 | 			//-put it in the corresponding position of 
 97 | 			//-prefix sum
 98 | 			/*has hub neighbor*/
 99 | 			if(beg_pos[csr[j]+1]-
100 | 				beg_pos[csr[j]] > 128)
101 | 			{
102 | 				new_list[huboff[csr[j]]]=src_list[i];
103 | 				huboff[csr[j]]++;
104 | 				isEligible=true;
105 | 				break;
106 | 			}
107 | 		}
108 | 		
109 | 		if(!isEligible)
110 | 		{
111 | 			new_list[remainoff]=src_list[i];
112 | 			remainoff++;
113 | 			test++;
114 | 		}
115 | 	}
116 | 
117 | 	#ifdef ENABLE_MONITORING
118 | 	std::ofstream myfile("origin_src.dat");
119 | 	std::ofstream newfile("new_src.dat");
120 | 	for(index_t i=0;i<src_count;i++)
121 | 	{
122 | 		myfile<<src_list[i]<<"\n";
123 | 		newfile<<new_list[i]<<"\n";
124 | 	}
125 | 	myfile.close();
126 | 	newfile.close();
127 | 
128 | 	index_t remain=src_count;
129 | 	index_t grouped=0;
130 | 	std::cout<<"Eligible-BFS(by-src-remain): ";
131 | 	for(index_t i=0;i<vert_count;i++)
132 | 	{
133 | 		if(hubmap[i])
134 | 		{
135 | 			std::cout<<hubmap[i]<<" ";
136 | 			grouped+=hubmap[i];
137 | 		}
138 | 	}
139 | 	remain-=grouped;
140 | 	std::cout<<remain<<"\n";
141 | 	std::cout<<"Get test: "<<test<<"\n";
142 | 	#endif
143 | 	
144 | 	delete[] src_list;
145 | 	delete[] hubmap;
146 | 	delete[] huboff;
147 | 	src_list=new_list;
148 | 	return;
149 | }
150 | 


--------------------------------------------------------------------------------
/ibfs.cu:
--------------------------------------------------------------------------------
  1 | //Graph format: Json based format
  2 | //Storage format: 
  3 | //struct{
  4 | //		int: src_ver
  5 | //		Arr: [ver_0|ver_1|ver_2|...]
  6 | //		Int: num_conn_ver
  7 | //	}
  8 | #include "graph.h"
  9 | #include "graph_reader.h"
 10 | #include <sstream>
 11 | #include <iostream>
 12 | #include <fstream>
 13 | #include "ibfs.h"
 14 | 
 15 | //template<typename index_t>
 16 | //void proc_desc(
 17 | //	std::string desc_str,
 18 | //	std::string &kron_addr,
 19 | //	index_t	&sw_level,
 20 | //	index_t &vert_count,
 21 | //	index_t	&edge_count
 22 | //){
 23 | //	//proc desc file
 24 | //	std::ifstream descfile(desc_str.c_str());
 25 | //	std::stringstream ss;
 26 | //	std::string line;
 27 | //	index_t	count=0;
 28 | //
 29 | //	if(descfile.is_open()){
 30 | //		while(std::getline(descfile, line)){
 31 | //			count++;
 32 | //			switch(count){
 33 | //				case 5:
 34 | //					ss.str("");
 35 | //					ss.clear();
 36 | //					ss<<line;
 37 | //					ss>>vert_count;ss>>edge_count;break;
 38 | //				case 11:
 39 | //					ss.str("");
 40 | //					ss.clear();
 41 | //					ss<<line;
 42 | //					ss>>kron_addr;ss>>sw_level;break;
 43 | //				default:
 44 | //					break;
 45 | //			}
 46 | //		}
 47 | //	}else{std::cout<<"File desc.dat missing\n"; exit(-1);}
 48 | //	std::cout<<"gpu vert count: "<<vert_count<<"\t"
 49 | //					 <<"gpu edge count: "<<edge_count<<"\n";
 50 | //}
 51 | 
 52 | int ibfs(int world_sz, int my_id, int gpu_id, int args, char *argv[]) {
 53 | 	//typedef int vertex_t;
 54 | 	//typedef int index_t;
 55 | 	//typedef 	unsigned char 	depth_t;
 56 | 	//typedef 	uint4 comp_t;
 57 | 	
 58 | 	std::cout<<"Input: /path/to/exe /path/to/beg "
 59 | 				<<"/path/to/csr concurrent-count " 
 60 | 				<<"switch-level bfs_count\n";
 61 | 	if(args != 6){std::cout<<"Wrong input\n";exit(-1);}
 62 | 	
 63 | 	const char *beg_file=argv[1];
 64 | 	const char *csr_file=argv[2];
 65 | 	const int concurrent_count=atoi(argv[3]);
 66 | 	const depth_t sw_level=atoi(argv[4]);
 67 | 	const index_t src_count=atoi(argv[5]);
 68 | 	graph_reader *g=new graph_reader(beg_file, csr_file, src_count);
 69 | 
 70 | 	cudaSetDeviceFlags(cudaDeviceMapHost);
 71 | 	const index_t sml_shed 	= 32; 
 72 | 	const index_t lrg_shed 	= 512;
 73 | 	index_t gpu_count		= 1;
 74 | 
 75 | 	int num_groups = src_count / concurrent_count;
 76 | 	int num_agg_bfs = concurrent_count;
 77 | 	
 78 | 	graph<vertex_t, index_t, depth_t> *graph_d 
 79 | 		= new graph<vertex_t, index_t, depth_t>
 80 | 				(
 81 | 				 	g,
 82 | 					sw_level,
 83 | 					g->vert_count,
 84 | 					g->edge_count,
 85 | 					num_groups,
 86 | 					num_agg_bfs,
 87 | 					world_sz, my_id,
 88 | 					gpu_id,
 89 | 					sml_shed, 
 90 | 					lrg_shed);
 91 | 
 92 | //	index_t step_sz = graph_d->vert_count/(THDS_NUM*BLKS_NUM);
 93 | //	if((step_sz%16 != 0)||(graph_d->vert_count%(THDS_NUM*BLKS_NUM))){
 94 | //		std::cout<<"Graph vertices sz fails " 
 95 | //						 <<"to be exactly divided by thread count\n";
 96 | //		return -1;
 97 | //	}
 98 | 	
 99 | 	std::cout<<"Alloc GPU space\n";
100 | 	graph_d->alloc_array();
101 | 	std::cout<<"In gpu bfs\n";
102 | 	graph_d->bfs_gpu_coalescing_mem();
103 | //	graph_d->write_result();
104 | 	return 0;
105 | }
106 | 


--------------------------------------------------------------------------------
/ibfs.h:
--------------------------------------------------------------------------------
1 | #ifndef IBFSH
2 | #define IBFSH
3 | 
4 | typedef int index_t;
5 | int ibfs(int world_sz, int my_id, int gpu_count, int args, char *argv[]); 
6 | int reporter(double tm, int my_id, index_t iter);
7 | 
8 | #endif
9 | 


--------------------------------------------------------------------------------
/inspector.cu:
--------------------------------------------------------------------------------
  1 | /*
  2 |  * Copyright 2016 The George Washington University
  3 |  * Written by Hang Liu 
  4 |  * Directed by Prof. Howie Huang
  5 |  *
  6 |  * https://www.seas.gwu.edu/~howie/
  7 |  * Contact: iheartgraph@gmail.com
  8 |  *
  9 |  * 
 10 |  * Please cite the following paper:
 11 |  * 
 12 |  * Hang Liu, H. Howie Huang and Yang Hu. 2016. iBFS: Concurrent Breadth-First Search on GPUs. Proceedings of the 2016 International Conference on Management of Data. ACM. 
 13 |  *
 14 |  * This file is part of iBFS.
 15 |  *
 16 |  * iBFS is free software: you can redistribute it and/or modify
 17 |  * it under the terms of the GNU General Public License as published by
 18 |  * the Free Software Foundation, either version 3 of the License, or
 19 |  * (at your option) any later version.
 20 |  *
 21 |  * iBFS is distributed in the hope that it will be useful,
 22 |  * but WITHOUT ANY WARRANTY; without even the implied warranty of
 23 |  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the 
 24 |  * GNU General Public License for more details.
 25 |  *
 26 |  * You should have received a copy of the GNU General Public License
 27 |  * along with iBFS.  If not, see <http://www.gnu.org/licenses/>.
 28 |  */
 29 | 
 30 | //Dec/15/2013
 31 | //Feb/16/2015
 32 | //Mar/4/2015
 33 | #include "util.h"
 34 | #include "gpu_ibfs.cuh"
 35 | #include "scan.cuh"
 36 | 
 37 | __global__ void extract_depth(
 38 | 	comp_t *depth_comp_last,
 39 | 	comp_t *depth_comp_curr,
 40 | 	bool 			*is_done_d,
 41 | 	depth_t	curr_level,
 42 | 	index_t	vert_count
 43 | ){
 44 | 	
 45 | 	//+-
 46 | 	//|Diff of comp_last and comp_curr = just visited verts
 47 | 	//+--------------------------------------
 48 | 
 49 | 	index_t tid=threadIdx.x+blockIdx.x*blockDim.x;
 50 | 	index_t wid=tid>>5;
 51 | //	const index_t vec_sz=32;
 52 | //	const index_t LANE=tid%32;
 53 | 	const index_t GRNTY=blockDim.x*gridDim.x;
 54 | 	const index_t WGRNTY=GRNTY>>5;
 55 | //	const depth_t LEVEL=curr_level;
 56 | 	const index_t COMP_COUNT=vert_count;
 57 | 	comp_t comp_last, comp_curr;
 58 | //	unsigned int recv_up, recv_low;
 59 | 	bool is_done=true;
 60 | 
 61 | 	while(tid<COMP_COUNT){
 62 | 		comp_last=depth_comp_last[tid];
 63 | 		comp_curr=depth_comp_curr[tid];
 64 | 		if(comp_last != comp_curr){
 65 | 			depth_comp_last[tid]=comp_curr;
 66 | 			if(is_done) is_done=false;
 67 | 		}
 68 | 
 69 | 	//	for(int i=0;i<vec_sz;i++){
 70 | 	//		recv_up=change>>32;
 71 | 	//		recv_low=change&0xffffffff;
 72 | 	//		recv_up=__shfl((int)recv_up,i);
 73 | 	//		recv_low=__shfl((int)recv_low,i);
 74 | 
 75 | 	//		if(recv_up&(1<<LANE)) 
 76 | 	//			depth_merge[wid*vec_sz*num_agg_bfs+i*num_agg_bfs+vec_sz+LANE]=LEVEL;
 77 | 	//		if(recv_low&(1<<LANE)) 
 78 | 	//			depth_merge[wid*vec_sz*num_agg_bfs+i*num_agg_bfs+LANE]=LEVEL;
 79 | 	//	}
 80 | 	//	__syncthreads();
 81 | 		tid += GRNTY;
 82 | 		wid += WGRNTY;
 83 | 	}
 84 | //is_done_d[0]=false;
 85 | 	if(is_done==false) is_done_d[0]=false;
 86 | }
 87 | 
 88 | __global__ void td_frontier_count(
 89 | 	index_t		*beg_pos_d,
 90 | 	index_t	 	*cat_sml_sz_d,
 91 | 	index_t	 	*cat_mid_sz_d,
 92 | 	index_t	 	*cat_lrg_sz_d,
 93 | 	comp_t 		*depth_comp_last,
 94 | 	comp_t 		*depth_comp_curr,
 95 | 	depth_t	 	curr_level,
 96 | 	index_t	 	vert_count,
 97 | 	const index_t sml_shed,
 98 | 	const index_t lrg_shed
 99 | )
100 | {
101 | 	//+-
102 | 	//|Diff of comp_last and comp_curr = just visited verts
103 | 	//+--------------------------------------
104 | 	index_t TID_ST=threadIdx.x+blockIdx.x*blockDim.x;
105 | 	index_t tid=TID_ST;
106 | 	index_t wid=TID_ST>>5;
107 | //	const index_t vec_sz=32;
108 | //	const index_t LANE=tid%32;
109 | 	const index_t GRNTY=blockDim.x*gridDim.x;
110 | 	const index_t WGRNTY=GRNTY>>5;
111 | //	const depth_t LEVEL=curr_level;
112 | 	const index_t COMP_COUNT=vert_count;
113 | 	comp_t comp_last, comp_curr;
114 | 	index_t card_curr;
115 | //	unsigned int recv_up, recv_low;
116 | 	
117 | 	index_t ex_sml_ct= 0;
118 | 	index_t ex_mid_ct= 0;
119 | 	index_t ex_lrg_ct= 0;
120 | 
121 | 	while(tid<COMP_COUNT){
122 | 		comp_last=depth_comp_last[tid];
123 | 		comp_curr=depth_comp_curr[tid];
124 | 		if((comp_last^comp_curr) != 0){
125 | 			
126 | 			//depth_comp_last[tid]=comp_curr;
127 | 			//CANNOT CHANGE NOW!!!
128 | 			//NEED diff FOR GATHERING
129 | 
130 | 			card_curr=beg_pos_d[tid+1]-beg_pos_d[tid];
131 | 			if(card_curr>0)
132 | 				if(card_curr<sml_shed) ex_sml_ct++;
133 | 				else if(card_curr>lrg_shed) ex_lrg_ct++;
134 | 				else ex_mid_ct++;
135 | 		}
136 | 
137 | 	//	for(int i=0;i<vec_sz;i++){
138 | 	//		recv_up=change>>32;
139 | 	//		recv_low=change&0xffffffff;
140 | 	//		recv_up=__shfl((int)recv_up,i);
141 | 	//		recv_low=__shfl((int)recv_low,i);
142 | 
143 | 	//		if(recv_up&(1<<LANE)) 
144 | 	//			depth_merge[wid*vec_sz*num_agg_bfs+i*num_agg_bfs+vec_sz+LANE]=LEVEL;
145 | 	//		if(recv_low&(1<<LANE)) 
146 | 	//			depth_merge[wid*vec_sz*num_agg_bfs+i*num_agg_bfs+LANE]=LEVEL;
147 | 	//	}
148 | 		__syncthreads();
149 | 		tid += GRNTY;
150 | 		wid += WGRNTY;
151 | 	}
152 | 	
153 | 	cat_sml_sz_d[TID_ST]= ex_sml_ct;
154 |  	cat_mid_sz_d[TID_ST]= ex_mid_ct;
155 |  	cat_lrg_sz_d[TID_ST]= ex_lrg_ct;
156 | }
157 | 
158 | __global__ void sw_frontier_count(
159 | 	index_t		*beg_pos_d,
160 | 	index_t	 *cat_sml_sz_d,
161 | 	index_t	 *cat_mid_sz_d,
162 | 	index_t	 *cat_lrg_sz_d,
163 | 	comp_t 	*depth_comp_curr,
164 | 	depth_t	 curr_level,
165 | 	index_t	 vert_count,
166 | 	const index_t sml_shed,
167 | 	const index_t lrg_shed
168 | 
169 | )
170 | {
171 | 	const index_t TID_ST	= threadIdx.x +blockIdx.x*blockDim.x;
172 | 	const index_t NUM_VER	= vert_count;	
173 | 	const index_t GRNTY		= blockDim.x*gridDim.x;
174 | 
175 | 	index_t card_curr,card_next;
176 | 	comp_t depth_curr,depth_next;
177 | 	index_t ex_sml_ct	= 0;
178 | 	index_t ex_mid_ct	= 0;
179 | 	index_t ex_lrg_ct	= 0;
180 | 	index_t tid				= TID_ST;
181 | 	
182 | 	//Figure out its own start and end pos
183 | 	//We want each thread to inspect a continuous block
184 | 	index_t step_sz = NUM_VER/GRNTY;
185 | 	if(step_sz<16){
186 | 		//small problem size
187 | 		const index_t REMAINDER = NUM_VER-step_sz*GRNTY;
188 | 		if(TID_ST<REMAINDER) step_sz ++;
189 | 		const index_t beg_pos=step_sz*TID_ST+ 
190 | 					(TID_ST >= REMAINDER ? REMAINDER:0);
191 | 		const index_t end_pos=beg_pos+step_sz;
192 | 		tid=beg_pos;
193 | 		if(step_sz){
194 | 			card_curr		= beg_pos_d[beg_pos+1]-beg_pos_d[beg_pos];
195 | 			depth_curr	= depth_comp_curr[beg_pos];
196 | //			depth_comp_last[beg_pos]=depth_curr;
197 | 		}
198 | 
199 | 		while(tid<end_pos){
200 | 			tid++;
201 | 			if(tid<end_pos){
202 | 				card_next		= beg_pos_d[tid+1]-beg_pos_d[tid];
203 | 				depth_next	= depth_comp_curr[tid];
204 | //				depth_comp_last[tid]=depth_next;
205 | 			}
206 | 				
207 | 			if(((~depth_curr)!=0) && (card_curr>0)){
208 | 				if(card_curr<sml_shed) ex_sml_ct++;
209 | 				else if(card_curr>lrg_shed) ex_lrg_ct++;
210 | 				else ex_mid_ct++;
211 | 			}
212 | 			
213 | 			depth_curr=depth_next;
214 | 			card_curr	= card_next;
215 | 		}
216 | 	}else{
217 | 		//big problem size
218 | 		//we want each thread to get 16x indices to check.
219 | 		if(step_sz&0xf) step_sz=((step_sz>>4)+1)<<4;
220 | 		if(NUM_VER-step_sz*GRNTY>0) step_sz<<=1;	
221 | 		__shared__ index_t	beg_pos_s[THDS_NUM];
222 | 		__shared__ depth_t	mark_s[THDS_NUM<<4];
223 | 		index_t	beg_pos_own			= step_sz*TID_ST;
224 | 		beg_pos_s[threadIdx.x] 	= beg_pos_own;
225 | 		const index_t TRIES 		= step_sz>>4;
226 | 		const index_t lane_id		= threadIdx.x%32;
227 | 		const index_t	warp_id		= threadIdx.x>>5;
228 | 		const index_t	thread_off= threadIdx.x<<4;
229 | 		index_t tries = 0;
230 | 		index_t	load_ptr;
231 | 		index_t	proc_vert;
232 | 		
233 | 		//+--------------------
234 | 		//in shared memory data representation
235 | 		//---------------------------------------
236 | 		//			0000|1111
237 | 		//					 ||||
238 | 		//					 |||+-------- INFTY ? 
239 | 		//					 ||+--------- SML   ?
240 | 		//					 |+---------- MID   ?
241 | 		//					 +----------- LRG   ?
242 | 		//FURTHER OPTIMIZATION CAN BE CARRIED OUT
243 | 		//BY PACKING MORE HERE.
244 | 		//-----------------------------------------
245 | 		while(tries<TRIES){
246 | 			//warp stride loading
247 | 			for(int i=0; i<32; i+=2){
248 | 				proc_vert=(warp_id<<5)+i+(lane_id>>4);
249 | 				depth_curr.x=0;depth_curr.y=0;depth_curr.y=0;depth_curr.w=0;
250 | 				depth_curr=~depth_curr;
251 | 				if(proc_vert<NUM_VER)
252 | 				{
253 | 					load_ptr=(tries<<4)+beg_pos_s[proc_vert]+(lane_id%16);
254 | 					if(load_ptr<NUM_VER)
255 | 					{
256 | 						depth_curr=depth_comp_curr[load_ptr];
257 | 						card_curr=beg_pos_d[load_ptr+1]-beg_pos_d[load_ptr];
258 | 					//	depth_comp_last[load_ptr]=depth_curr;
259 | 					}else card_curr=0;//out-of-boundary, set it orphan
260 | 				}else card_curr=0;//out-of-boundary, set it orphan
261 | 				mark_s[(proc_vert<<4)+(lane_id%16)]=0x00;
262 | 
263 | 				//construct hub-cache
264 | //				if(depth_curr == LEVEL){
265 | //					hub_ptr	= load_ptr & (HUB_BU_SZ-1);
266 | //					if(card_curr > hub_card[hub_ptr]){
267 | //						hub_vert[hub_ptr] = load_ptr;
268 | //						hub_card[hub_ptr] = card_curr;
269 | //					}
270 | //				}else
271 | 				if(((~depth_curr)!=0) && (card_curr>0))
272 | 					if(card_curr<sml_shed)
273 | 						mark_s[(proc_vert<<4)+(lane_id%16)]=0x03;//0011
274 | 					else if(card_curr>lrg_shed)
275 | 						mark_s[(proc_vert<<4)+(lane_id%16)]=0x09;//1001
276 | 					else mark_s[(proc_vert<<4)+(lane_id%16)]=0x05;//0101
277 | 			}
278 | 			__syncthreads();
279 | 
280 | 			//thread stride checking
281 | 			for(int i=0; i<16; i++)
282 | 				if(mark_s[thread_off+i]&0x02){
283 | 					ex_sml_ct++;
284 | 				}else if(mark_s[thread_off+i]&0x08){
285 | 					ex_lrg_ct++;
286 | 				}else if(mark_s[thread_off+i]&0x04){
287 | 					ex_mid_ct++;
288 | 				}
289 | 			++tries;
290 | 			__syncthreads();
291 | 		}
292 | 	}
293 | 	
294 | 	cat_sml_sz_d[TID_ST]= ex_sml_ct;
295 | 	cat_mid_sz_d[TID_ST]= ex_mid_ct;
296 | 	cat_lrg_sz_d[TID_ST]= ex_lrg_ct;
297 | }
298 | 
299 | __global__ void td_frontier_gather(
300 | 	index_t		*beg_pos_d,
301 | 	index_t	 	*ex_sml_q_d,
302 | 	index_t	 	*ex_mid_q_d,
303 | 	index_t	 	*ex_lrg_q_d,
304 | 	index_t	 	*cat_sml_off_d,
305 | 	index_t	 	*cat_mid_off_d,
306 | 	index_t	 	*cat_lrg_off_d,
307 | 	comp_t 		*depth_comp_last,
308 | 	comp_t 		*depth_comp_curr,
309 | 	depth_t	 	curr_level,
310 | 	index_t	 	vert_count,
311 | 	const index_t sml_shed,
312 | 	const index_t lrg_shed
313 | )
314 | {
315 | 	index_t TID_ST=threadIdx.x+blockIdx.x*blockDim.x;
316 | 	index_t tid=TID_ST;
317 | 	const index_t GRNLTY=blockDim.x*gridDim.x;
318 | 	const index_t COMP_COUNT=vert_count;
319 | 	comp_t comp_last, comp_curr;
320 | 	index_t card_curr;
321 | 	
322 | 	index_t ex_sml_ptr	= cat_sml_off_d[TID_ST];
323 | 	index_t ex_mid_ptr	= cat_mid_off_d[TID_ST];
324 | 	index_t ex_lrg_ptr	= cat_lrg_off_d[TID_ST];
325 | 	
326 | 	while(tid<COMP_COUNT){
327 | 		comp_last=depth_comp_last[tid];
328 | 		comp_curr=depth_comp_curr[tid];
329 | 		comp_t change=comp_curr-comp_last;
330 | 		if(change!=0){
331 | 			depth_comp_last[tid]=comp_curr;
332 | 			card_curr=beg_pos_d[tid+1]-beg_pos_d[tid];
333 | 			if(card_curr>0)
334 | 				if(card_curr<sml_shed)
335 | 				{
336 | 					ex_sml_q_d[ex_sml_ptr]=tid;
337 | 					ex_sml_ptr++;
338 | 				}
339 | 				else if(card_curr>lrg_shed)
340 | 				{
341 | 					ex_lrg_q_d[ex_lrg_ptr]=tid;
342 | 					ex_lrg_ptr++;
343 | 				}
344 | 				else
345 | 				{
346 | 					ex_mid_q_d[ex_mid_ptr]=tid;
347 | 					ex_mid_ptr++;
348 | 				}
349 | 		}
350 | 		tid+= GRNLTY;
351 | 	}
352 | }
353 | 
354 | 
355 | __global__ void sw_frontier_gather(
356 | 	index_t		*beg_pos_d,
357 | 	index_t	 	*ex_sml_q_d,
358 | 	index_t	 	*ex_mid_q_d,
359 | 	index_t	 	*ex_lrg_q_d,
360 | 	index_t	 	*cat_sml_off_d,
361 | 	index_t	 	*cat_mid_off_d,
362 | 	index_t	 	*cat_lrg_off_d,
363 | 	comp_t 	*depth_comp_curr,
364 | 	depth_t	 curr_level,
365 | 	index_t	 vert_count,
366 | 	const index_t sml_shed,
367 | 	const index_t lrg_shed
368 | )
369 | {
370 | 	const index_t TID_ST	= threadIdx.x +blockIdx.x*blockDim.x;
371 | 	const index_t NUM_VER	= vert_count;	
372 | 	const index_t GRNTY		= blockDim.x*gridDim.x;
373 | 
374 | 	index_t ex_sml_ptr	= cat_sml_off_d[TID_ST];
375 | 	index_t ex_mid_ptr	= cat_mid_off_d[TID_ST];
376 | 	index_t ex_lrg_ptr	= cat_lrg_off_d[TID_ST];
377 | 
378 | 	index_t card_curr, card_next;
379 | 	comp_t depth_curr,depth_next;
380 | 	index_t	tid,tid_next;
381 | 	
382 | 	//Figure out its own start and end pos
383 | 	//We want each thread to inspect a continuous block
384 | 	index_t step_sz = NUM_VER/GRNTY;
385 | 	if(step_sz<16){
386 | 		//small problem size
387 | 		const index_t REMAINDER = NUM_VER-step_sz*GRNTY;
388 | 		if(TID_ST<REMAINDER) step_sz ++;
389 | 		const index_t beg_pos=step_sz*TID_ST+ 
390 | 					(TID_ST >= REMAINDER ? REMAINDER:0);
391 | 		const index_t end_pos=beg_pos+step_sz;
392 | 		
393 | 		tid=beg_pos;tid_next=beg_pos;
394 | 		if(step_sz){
395 | 			card_curr		= beg_pos_d[beg_pos+1]-beg_pos_d[beg_pos];
396 | 			depth_curr	= depth_comp_curr[beg_pos];
397 | 			tid_next++;
398 | 		}
399 | 			
400 | 		while(tid<end_pos){
401 | 			if(tid_next < end_pos){
402 | 				card_next		= beg_pos_d[tid_next+1]-beg_pos_d[tid_next];
403 | 				depth_next	= depth_comp_curr[tid_next];
404 | 			}
405 | 			
406 | 			if((card_curr>0)&&((~depth_curr)!=0)){
407 | 				if(card_curr<sml_shed)
408 | 				{
409 | 					ex_sml_q_d[ex_sml_ptr]=tid;
410 | 					ex_sml_ptr++;
411 | 				}
412 | 				else if(card_curr>lrg_shed)
413 | 				{
414 | 					ex_lrg_q_d[ex_lrg_ptr]=tid;
415 | 					ex_lrg_ptr++;
416 | 				}
417 | 				else
418 | 				{
419 | 					ex_mid_q_d[ex_mid_ptr]=tid;
420 | 					ex_mid_ptr++;
421 | 				}
422 | 			}
423 | 			tid=tid_next;
424 | 			tid_next++;
425 | 			card_curr	= card_next;
426 | 			depth_curr=depth_next;
427 | 		}
428 | 	}else{
429 | 		//big problem size
430 | 		//We want each thread to get 16x indices to check.
431 | 		if(step_sz&0xf) step_sz=((step_sz>>4)+1)<<4;
432 | 		if(NUM_VER-step_sz*GRNTY>0) step_sz<<=1;	
433 | 		__shared__ index_t	beg_pos_s[THDS_NUM];
434 | 		__shared__ depth_t	mark_s[THDS_NUM<<4];
435 | 		index_t	beg_pos_own			= step_sz*TID_ST;
436 | 		beg_pos_s[threadIdx.x] 	= beg_pos_own;
437 | 		const index_t TRIES 		= step_sz>>4;
438 | 		const index_t lane_id		= threadIdx.x%32;
439 | 		const index_t	warp_id		= threadIdx.x>>5;
440 | 		const index_t	thread_off= threadIdx.x<<4;
441 | 		index_t tries = 0;
442 | 		index_t	load_ptr;
443 | 		index_t	proc_vert;
444 | 
445 | 		//+--------------------
446 | 		//in shared memory data representation
447 | 		//---------------------------------------
448 | 		//			0000|1111
449 | 		//					 ||||
450 | 		//					 |||+-------- INFTY ? 
451 | 		//					 ||+--------- SML   ?
452 | 		//					 |+---------- MID   ?
453 | 		//					 +----------- LRG   ?
454 | 		//FURTHER OPTIMIZATION CAN BE CARRIED OUT
455 | 		//BY PACKING MORE HERE.
456 | 		//-----------------------------------------
457 | 		while(tries<TRIES){
458 | 			//warp stride loading
459 | 			for(int i=0; i<32; i+=2){
460 | 				proc_vert=(warp_id<<5)+i+(lane_id>>4);
461 | 				depth_curr.x=0;depth_curr.y=0;depth_curr.y=0;depth_curr.w=0;
462 | 				depth_curr=~depth_curr;
463 | 				if(proc_vert<NUM_VER)
464 | 				{
465 | 					load_ptr=(tries<<4)+beg_pos_s[proc_vert]+(lane_id%16);
466 | 					if(load_ptr<NUM_VER)
467 | 					{
468 | 						depth_curr=depth_comp_curr[load_ptr];
469 | 						card_curr=beg_pos_d[load_ptr+1]-beg_pos_d[load_ptr];
470 | 					}else card_curr=0;//out-of-boundary, set it orphan
471 | 				}else card_curr=0;//out-of-boundary, set it orphan
472 | 				mark_s[(proc_vert<<4)+(lane_id%16)]=0x00;
473 | 
474 | 				if(((~depth_curr)!=0)&&(card_curr>0))
475 | 					if(card_curr<sml_shed)
476 | 						mark_s[(proc_vert<<4)+(lane_id%16)]=0x03;//0011
477 | 					else if(card_curr>lrg_shed)
478 | 						mark_s[(proc_vert<<4)+(lane_id%16)]=0x09;//1001
479 | 					else mark_s[(proc_vert<<4)+(lane_id%16)]=0x05;//0101
480 | 			}
481 | 			__syncthreads();
482 | 
483 | 			//thread stride checking
484 | 			for(int i=0; i<16; i++)
485 | 				if(mark_s[thread_off+i]&0x02){
486 | 					ex_sml_q_d[ex_sml_ptr]=beg_pos_own+(tries<<4)+i;
487 | 					ex_sml_ptr++;
488 | 				}else if(mark_s[thread_off+i]&0x08){
489 | 					ex_lrg_q_d[ex_lrg_ptr]=beg_pos_own+(tries<<4)+i;
490 | 					ex_lrg_ptr++;
491 | 				}else if(mark_s[thread_off+i]&0x04){
492 | 					ex_mid_q_d[ex_mid_ptr]=beg_pos_own+(tries<<4)+i;
493 | 					ex_mid_ptr++;
494 | 				}
495 | 
496 | 			++tries;
497 | 			__syncthreads();
498 | 		}
499 | 	}
500 | }
501 | 
502 | 
503 | //+--------------------------------------------
504 | //|CLASSIFIED BASED STORAGE FOR EX_QUEUE
505 | //+---------------------------------------------------
506 | void 
507 | gpu_ibfs::
508 | td_inspect(depth_t curr_level)
509 | {
510 | 	td_frontier_count	
511 | 	<<<BLKS_NUM,THDS_NUM>>>(
512 | 		beg_pos_d,
513 | 		cat_sml_sz_d,
514 | 		cat_mid_sz_d,
515 | 		cat_lrg_sz_d,
516 | 		depth_comp_last,
517 | 		depth_comp_curr,
518 | 		curr_level,
519 | 		g->vert_count,
520 | 		sml_shed,
521 | 		lrg_shed
522 | 	);
523 | 	cudaDeviceSynchronize();
524 | 	
525 | 	insp_scan
526 |  	(
527 |  		cat_sml_sz_d,
528 |  		cat_sml_off_d,
529 |  		THDS_NUM*BLKS_NUM,
530 |  		BLKS_NUM>>1,
531 |  		THDS_NUM>>1,
532 |  		ex_sml_sz_d,
533 |  		stream[0]
534 |  	);
535 |  	insp_scan
536 |  	(
537 |  		cat_mid_sz_d,
538 |  		cat_mid_off_d,
539 |  		THDS_NUM*BLKS_NUM,
540 |  		BLKS_NUM>>1,
541 |  		THDS_NUM>>1,
542 |  		ex_mid_sz_d,
543 |  		stream[1]
544 |  	);
545 |  	insp_scan
546 |  	(
547 |  		cat_lrg_sz_d,
548 |  		cat_lrg_off_d,
549 |  		THDS_NUM*BLKS_NUM,
550 |  		BLKS_NUM>>1,
551 |  		THDS_NUM>>1,
552 |  		ex_lrg_sz_d,
553 |  		stream[2]
554 |  	);
555 |  	
556 | 
557 | 	cudaStreamSynchronize(stream[0]);
558 | 	cudaStreamSynchronize(stream[1]);
559 | 	cudaStreamSynchronize(stream[2]);
560 | 	cudaDeviceSynchronize();
561 | 	
562 | 	td_frontier_gather	
563 | 	<<<BLKS_NUM,THDS_NUM>>>(
564 | 		beg_pos_d,
565 | 		ex_sml_q_d,
566 | 		ex_mid_q_d,
567 | 		ex_lrg_q_d,
568 | 		cat_sml_off_d,
569 | 		cat_mid_off_d,
570 | 		cat_lrg_off_d,
571 | 		depth_comp_last,
572 | 		depth_comp_curr,
573 | 		curr_level,
574 | 		g->vert_count,
575 | 		sml_shed,
576 | 		lrg_shed
577 | 	);
578 | }
579 | 
580 | void 
581 | gpu_ibfs::
582 | sw_inspect(depth_t curr_level)
583 | {
584 | 	is_done[0]=true;
585 | 	extract_depth
586 | 	<<<BLKS_NUM<<2,THDS_NUM<<2,0,gstream>>>
587 | 	(
588 | 		depth_comp_last,
589 | 		depth_comp_curr,
590 | 		is_done_d,
591 | 		curr_level,
592 | 		g->vert_count
593 | 	);
594 | 	
595 | 	sw_frontier_count
596 |  	<<<BLKS_NUM,THDS_NUM>>>(
597 | 		beg_pos_d,
598 |  		cat_sml_sz_d,
599 |  		cat_mid_sz_d,
600 |  		cat_lrg_sz_d,
601 |  		depth_comp_curr,
602 |  		curr_level,
603 |  		g->vert_count,
604 |  		sml_shed,
605 |  		lrg_shed
606 | 	);
607 | 	
608 | 	cudaDeviceSynchronize();
609 | 	
610 | 	insp_scan
611 |  	(
612 |  		cat_sml_sz_d,
613 |  		cat_sml_off_d,
614 |  		THDS_NUM*BLKS_NUM,
615 |  		BLKS_NUM>>1,
616 |  		THDS_NUM>>1,
617 |  		ex_sml_sz_d,
618 |  		stream[0]
619 |  	);
620 |  	insp_scan
621 |  	(
622 |  		cat_mid_sz_d,
623 |  		cat_mid_off_d,
624 |  		THDS_NUM*BLKS_NUM,
625 |  		BLKS_NUM>>1,
626 |  		THDS_NUM>>1,
627 |  		ex_mid_sz_d,
628 |  		stream[1]
629 |  	);
630 |  	insp_scan
631 |  	(
632 |  		cat_lrg_sz_d,
633 |  		cat_lrg_off_d,
634 |  		THDS_NUM*BLKS_NUM,
635 |  		BLKS_NUM>>1,
636 |  		THDS_NUM>>1,
637 |  		ex_lrg_sz_d,
638 |  		stream[2]
639 |  	);
640 | 
641 | 	cudaStreamSynchronize(stream[0]);
642 | 	cudaStreamSynchronize(stream[1]);
643 | 	cudaStreamSynchronize(stream[2]);
644 | 	cudaDeviceSynchronize();
645 | 	
646 | 	sw_frontier_gather
647 | 	<<<BLKS_NUM, THDS_NUM>>>(
648 | 		beg_pos_d,
649 | 		ex_sml_q_d,
650 | 		ex_mid_q_d,
651 | 		ex_lrg_q_d,
652 | 		cat_sml_off_d,
653 | 		cat_mid_off_d,
654 | 		cat_lrg_off_d,
655 |  		depth_comp_curr,
656 | 		curr_level,
657 | 		g->vert_count,
658 | 		sml_shed,
659 | 		lrg_shed
660 | 	);
661 | 	cudaStreamSynchronize(gstream);
662 | }
663 | 
664 | void gpu_ibfs::
665 | inspector(depth_t level)
666 | {
667 | 	if(level<sw_level)
668 | 		td_inspect(level);
669 | 	else 
670 | 		sw_inspect(level);
671 | }
672 | 


--------------------------------------------------------------------------------
/inspector.cuh:
--------------------------------------------------------------------------------
  1 | //Dec/15/2013
  2 | //Feb/16/2015
  3 | //Mar/4/2015
  4 | #include "comm.h"
  5 | 
  6 | 
  7 | template<typename depth_t, typename comp_t, typename index_t>
  8 | __global__ void extract_depth(
  9 | 	comp_t *depth_comp_last,
 10 | 	comp_t *depth_comp_curr,
 11 | 	depth_t *depth_merge,
 12 | 	index_t num_agg_bfs,
 13 | 	depth_t	curr_level,
 14 | 	index_t	vert_count
 15 | ){
 16 | 	
 17 | 	//+-
 18 | 	//|Diff of comp_last and comp_curr = just visited verts
 19 | 	//+--------------------------------------
 20 | 	index_t tid=threadIdx.x+blockIdx.x*blockDim.x;
 21 | 	index_t wid=tid>>5;
 22 | 	const index_t vec_sz=32;
 23 | 	const index_t LANE=tid%32;
 24 | 	const index_t GRNTY=blockDim.x*gridDim.x;
 25 | 	const index_t WGRNTY=GRNTY>>5;
 26 | 	const depth_t LEVEL=curr_level;
 27 | 	const index_t COMP_COUNT=vert_count;
 28 | 	comp_t comp_last, comp_curr;
 29 | //	unsigned int recv_up, recv_low;
 30 | 	unsigned int change[4], recv;
 31 | 		
 32 | 	while(tid<COMP_COUNT){
 33 | 		comp_last=depth_comp_last[tid];
 34 | 		comp_curr=depth_comp_curr[tid];
 35 | 		change[0]=comp_curr.x-comp_last.x;
 36 | 		change[1]=comp_curr.y-comp_last.y;
 37 | 		change[2]=comp_curr.z-comp_last.z;
 38 | 		change[3]=comp_curr.w-comp_last.w;
 39 | 		
 40 | 		if(change[0] || change[1] || change[2] || change[3]){
 41 | 			depth_comp_last[tid]=comp_curr;
 42 | 
 43 | 			for(int i=0;i<vec_sz;i++){
 44 | 				for(int j=0;j<4;j++){
 45 | 					recv=change[j];
 46 | 					recv=__shfl((int)recv,i);
 47 | 					if(recv&(1<<LANE)) 
 48 | 						depth_merge[wid*vec_sz*num_agg_bfs+i*num_agg_bfs+j*vec_sz+LANE]=LEVEL;
 49 | 				}
 50 | 			}
 51 | 		}
 52 | 		__syncthreads();
 53 | 		tid += GRNTY;
 54 | 		wid += WGRNTY;
 55 | 	}
 56 | }
 57 | 
 58 | 
 59 | template
 60 | <	typename vertex_t,
 61 | 	typename index_t,
 62 | 	typename depth_t>
 63 | __global__ void td_frontier_count 
 64 | (
 65 | 	index_t	 	*ex_cat_sml_sz,
 66 | 	index_t	 	*ex_cat_mid_sz,
 67 | 	index_t	 	*ex_cat_lrg_sz,
 68 | 	comp_t 		*depth_comp_last,
 69 | 	comp_t 		*depth_comp_curr,
 70 | 	depth_t 	*depth_merge,
 71 | 	index_t		*adj_card_d,
 72 | 	index_t		num_agg_bfs,
 73 | 	depth_t	 	curr_level,
 74 | 	index_t	 	vert_count,
 75 | 	const index_t sml_shed,
 76 | 	const index_t lrg_shed
 77 | )
 78 | {
 79 | 	//+-
 80 | 	//|Diff of comp_last and comp_curr = just visited verts
 81 | 	//+--------------------------------------
 82 | 	index_t TID_ST=threadIdx.x+blockIdx.x*blockDim.x;
 83 | 	index_t tid=TID_ST;
 84 | 	index_t wid=TID_ST>>5;
 85 | 	const index_t vec_sz=32;
 86 | 	const index_t LANE=tid%32;
 87 | 	const index_t GRNTY=blockDim.x*gridDim.x;
 88 | 	const index_t WGRNTY=GRNTY>>5;
 89 | 	const depth_t LEVEL=curr_level;
 90 | 	const index_t COMP_COUNT=vert_count;
 91 | 	comp_t comp_last, comp_curr;
 92 | 	index_t card_curr;
 93 | 	unsigned int recv;
 94 | 	
 95 | 	index_t ex_sml_ct= 0;
 96 | 	index_t ex_mid_ct= 0;
 97 | 	index_t ex_lrg_ct= 0;
 98 | 	unsigned int change[4];
 99 | 
100 | 	while(tid<COMP_COUNT){
101 | 		comp_last=depth_comp_last[tid];
102 | 		comp_curr=depth_comp_curr[tid];
103 | 		change[0]=comp_curr.x-comp_last.x;
104 | 		change[1]=comp_curr.y-comp_last.y;
105 | 		change[2]=comp_curr.z-comp_last.z;
106 | 		change[3]=comp_curr.w-comp_last.w;
107 | 		
108 | 		if(change[0] || change[1] || change[2] || change[3]){
109 | 			
110 | 			//depth_comp_last[tid]=comp_curr;
111 | 			//CANNOT CHANGE NOW!!!
112 | 			//NEED diff FOR GATHERING
113 | 			for(int i=0;i<vec_sz;i++){
114 | 				for(int j=0;j<4;j++){
115 | 					recv=change[j];
116 | 					recv=__shfl((int)recv,i);
117 | 					if(recv&(1<<LANE)) 
118 | 						depth_merge[wid*vec_sz*num_agg_bfs+i*num_agg_bfs+j*vec_sz+LANE]=LEVEL;
119 | 				}
120 | 			}
121 | 
122 | 			card_curr=adj_card_d[tid];
123 | 			if(card_curr>0)
124 | 				if(card_curr<sml_shed) ex_sml_ct++;
125 | 				else if(card_curr>lrg_shed) ex_lrg_ct++;
126 | 				else ex_mid_ct++;
127 | 		}
128 | 
129 | 		__syncthreads();
130 | 		tid += GRNTY;
131 | 		wid += WGRNTY;
132 | 	}
133 | 	
134 | 	ex_cat_sml_sz[TID_ST]= ex_sml_ct;
135 |  	ex_cat_mid_sz[TID_ST]= ex_mid_ct;
136 |  	ex_cat_lrg_sz[TID_ST]= ex_lrg_ct;
137 | }
138 | 
139 | template <typename vertex_t, typename index_t,
140 | 			typename depth_t>
141 | __global__ void sw_frontier_count 
142 | (
143 | 	index_t	 *ex_cat_sml_sz,
144 | 	index_t	 *ex_cat_mid_sz,
145 | 	index_t	 *ex_cat_lrg_sz,
146 | 	comp_t 	*depth_comp_curr,
147 | 	index_t		*adj_card_d,
148 | 	index_t		num_agg_bfs,
149 | 	depth_t	 curr_level,
150 | 	index_t	 vert_count,
151 | 	const index_t sml_shed,
152 | 	const index_t lrg_shed
153 | )
154 | {
155 | 	const index_t TID_ST	= threadIdx.x +blockIdx.x*blockDim.x;
156 | 	const index_t NUM_VER	= vert_count;	
157 | 	const index_t GRNTY		= blockDim.x*gridDim.x;
158 | 
159 | 	index_t card_curr,card_next;
160 | 	comp_t depth_curr,depth_next;
161 | 	index_t ex_sml_ct	= 0;
162 | 	index_t ex_mid_ct	= 0;
163 | 	index_t ex_lrg_ct	= 0;
164 | 	index_t tid				= TID_ST;
165 | 	
166 | 	//Figure out its own start and end pos
167 | 	//We want each thread to inspect a continuous block
168 | 	index_t step_sz = NUM_VER/GRNTY;
169 | 	if(step_sz<16){
170 | 		//small problem size
171 | 		const index_t REMAINDER = NUM_VER-step_sz*GRNTY;
172 | 		if(TID_ST<REMAINDER) step_sz ++;
173 | 		const index_t beg_pos=step_sz*TID_ST+ 
174 | 					(TID_ST >= REMAINDER ? REMAINDER:0);
175 | 		const index_t end_pos=beg_pos+step_sz;
176 | 		tid=beg_pos;
177 | 		if(step_sz){
178 | 			card_curr		= adj_card_d[beg_pos];
179 | 			depth_curr	= depth_comp_curr[beg_pos];
180 | //			depth_comp_last[beg_pos]=depth_curr;
181 | 		}
182 | 
183 | 		while(tid<end_pos){
184 | 			tid++;
185 | 			if(tid<end_pos){
186 | 				card_next		= adj_card_d[tid];
187 | 				depth_next	= depth_comp_curr[tid];
188 | //				depth_comp_last[tid]=depth_next;
189 | 			}
190 | 				
191 | 			if((card_curr>0)&&(
192 | 				depth_curr.x!=0xffffffff ||
193 | 				depth_curr.y!=0xffffffff ||
194 | 				depth_curr.z!=0xffffffff ||
195 | 				depth_curr.w!=0xffffffff 
196 | 				)){
197 | 				if(card_curr<sml_shed) ex_sml_ct++;
198 | 				else if(card_curr>lrg_shed) ex_lrg_ct++;
199 | 				else ex_mid_ct++;
200 | 			}
201 | 			
202 | 			depth_curr=depth_next;
203 | 			card_curr	= card_next;
204 | 		}
205 | 	}else{
206 | 		//big problem size
207 | 		//we want each thread to get 16x indices to check.
208 | 		if(step_sz&0xf) step_sz=((step_sz>>4)+1)<<4;
209 | 		if(NUM_VER-step_sz*GRNTY>0) step_sz<<=1;	
210 | 		__shared__ index_t	beg_pos_s[THDS_NUM];
211 | 		__shared__ depth_t	mark_s[THDS_NUM<<4];
212 | 		index_t	beg_pos_own			= step_sz*TID_ST;
213 | 		beg_pos_s[threadIdx.x] 	= beg_pos_own;
214 | 		const index_t TRIES 		= step_sz>>4;
215 | 		const index_t lane_id		= threadIdx.x%32;
216 | 		const index_t	warp_id		= threadIdx.x>>5;
217 | 		const index_t	thread_off= threadIdx.x<<4;
218 | 		index_t tries = 0;
219 | 		index_t	load_ptr;
220 | 		index_t	proc_vert;
221 | 
222 | 		//+--------------------
223 | 		//in shared memory data representation
224 | 		//---------------------------------------
225 | 		//			0000|1111
226 | 		//					 ||||
227 | 		//					 |||+-------- INFTY ? 
228 | 		//					 ||+--------- SML   ?
229 | 		//					 |+---------- MID   ?
230 | 		//					 +----------- LRG   ?
231 | 		//FURTHER OPTIMIZATION CAN BE CARRIED OUT
232 | 		//BY PACKING MORE HERE.
233 | 		//-----------------------------------------
234 | 		while(tries<TRIES){
235 | 			//warp stride loading
236 | 			for(int i=0; i<32; i+=2){
237 | 				proc_vert=(warp_id<<5)+i+(lane_id>>4);
238 | 				depth_curr.x=0xffffffff;
239 | 				depth_curr.y=0xffffffff;
240 | 				depth_curr.z=0xffffffff;
241 | 				depth_curr.w=0xffffffff;
242 | 				if(proc_vert<NUM_VER)
243 | 				{
244 | 					load_ptr=(tries<<4)+beg_pos_s[proc_vert]+(lane_id%16);
245 | 					if(load_ptr<NUM_VER)
246 | 					{
247 | 						depth_curr=depth_comp_curr[load_ptr];
248 | 						card_curr=adj_card_d[load_ptr];
249 | 					//	depth_comp_last[load_ptr]=depth_curr;
250 | 					}else card_curr=0;//out-of-boundary, set it orphan
251 | 				}else card_curr=0;//out-of-boundary, set it orphan
252 | 				mark_s[(proc_vert<<4)+(lane_id%16)]=0x00;
253 | 
254 | 				//construct hub-cache
255 | //				if(depth_curr == LEVEL){
256 | //					hub_ptr	= load_ptr & (HUB_BU_SZ-1);
257 | //					if(card_curr > hub_card[hub_ptr]){
258 | //						hub_vert[hub_ptr] = load_ptr;
259 | //						hub_card[hub_ptr] = card_curr;
260 | //					}
261 | //				}else
262 | 				if((card_curr>0) &&(
263 | 				depth_curr.x!=0xffffffff ||
264 | 				depth_curr.y!=0xffffffff ||
265 | 				depth_curr.z!=0xffffffff ||
266 | 				depth_curr.w!=0xffffffff))
267 | 					if(card_curr<sml_shed)
268 | 						mark_s[(proc_vert<<4)+(lane_id%16)]=0x03;//0011
269 | 					else if(card_curr>lrg_shed)
270 | 						mark_s[(proc_vert<<4)+(lane_id%16)]=0x09;//1001
271 | 					else mark_s[(proc_vert<<4)+(lane_id%16)]=0x05;//0101
272 | 			}
273 | 			__syncthreads();
274 | 
275 | 			//thread stride checking
276 | 			for(int i=0; i<16; i++)
277 | 				if(mark_s[thread_off+i]&0x02){
278 | 					ex_sml_ct++;
279 | 				}else if(mark_s[thread_off+i]&0x08){
280 | 					ex_lrg_ct++;
281 | 				}else if(mark_s[thread_off+i]&0x04){
282 | 					ex_mid_ct++;
283 | 				}
284 | 			++tries;
285 | 			__syncthreads();
286 | 		}
287 | 	}
288 | 	
289 | 	ex_cat_sml_sz[TID_ST]= ex_sml_ct;
290 | 	ex_cat_mid_sz[TID_ST]= ex_mid_ct;
291 | 	ex_cat_lrg_sz[TID_ST]= ex_lrg_ct;
292 | }
293 | 
294 | template <typename vertex_t, typename index_t,
295 | 			typename depth_t>
296 | __global__ void td_frontier_gather 
297 | (
298 | 	index_t	 	*ex_sml_q,
299 | 	index_t	 	*ex_mid_q,
300 | 	index_t	 	*ex_lrg_q,
301 | 	index_t	 	*ex_sml_off,
302 | 	index_t	 	*ex_mid_off,
303 | 	index_t	 	*ex_lrg_off,
304 | 	comp_t 		*depth_comp_last,
305 | 	comp_t 		*depth_comp_curr,
306 | 	depth_t 	*depth_merge,
307 | 	index_t		*adj_card_d,
308 | 	index_t		num_agg_bfs,
309 | 	depth_t	 	curr_level,
310 | 	index_t	 	vert_count,
311 | 	const index_t sml_shed,
312 | 	const index_t lrg_shed
313 | )
314 | {
315 | 	index_t TID_ST=threadIdx.x+blockIdx.x*blockDim.x;
316 | 	index_t tid=TID_ST;
317 | 	const index_t GRNLTY=blockDim.x*gridDim.x;
318 | 	const index_t COMP_COUNT=vert_count;
319 | 	comp_t comp_last, comp_curr;
320 | 	index_t card_curr;
321 | 	
322 | 	index_t ex_sml_ptr	= ex_sml_off[TID_ST];
323 | 	index_t ex_mid_ptr	= ex_mid_off[TID_ST];
324 | 	index_t ex_lrg_ptr	= ex_lrg_off[TID_ST];
325 | 	unsigned int change[4];
326 | 
327 | 	
328 | 	while(tid<COMP_COUNT){
329 | 		comp_last=depth_comp_last[tid];
330 | 		comp_curr=depth_comp_curr[tid];
331 | 		change[0]=comp_curr.x-comp_last.x;
332 | 		change[1]=comp_curr.y-comp_last.y;
333 | 		change[2]=comp_curr.z-comp_last.z;
334 | 		change[3]=comp_curr.w-comp_last.w;
335 | 		
336 | 		if(change[0] || change[1] || change[2] || change[3]){		
337 | //			depth_comp_last[tid].x=comp_curr.x;
338 | //			depth_comp_last[tid].y=comp_curr.y;
339 | //			depth_comp_last[tid].z=comp_curr.z;
340 | //			depth_comp_last[tid].w=comp_curr.w;
341 | 			depth_comp_last[tid]=comp_curr;
342 | 			card_curr=adj_card_d[tid];
343 | 			if(card_curr>0)
344 | 				if(card_curr<sml_shed)
345 | 				{
346 | 					ex_sml_q[ex_sml_ptr]=tid;
347 | 					ex_sml_ptr++;
348 | 				}
349 | 				else if(card_curr>lrg_shed)
350 | 				{
351 | 					ex_lrg_q[ex_lrg_ptr]=tid;
352 | 					ex_lrg_ptr++;
353 | 				}
354 | 				else
355 | 				{
356 | 					ex_mid_q[ex_mid_ptr]=tid;
357 | 					ex_mid_ptr++;
358 | 				}
359 | 		}
360 | 		tid+= GRNLTY;
361 | 	}
362 | }
363 | 
364 | 
365 | template <typename vertex_t, typename index_t,
366 | 			typename depth_t>
367 | __global__ void sw_frontier_gather 
368 | (
369 | 	index_t	 	*ex_sml_q,
370 | 	index_t	 	*ex_mid_q,
371 | 	index_t	 	*ex_lrg_q,
372 | 	index_t	 	*ex_sml_off,
373 | 	index_t	 	*ex_mid_off,
374 | 	index_t	 	*ex_lrg_off,
375 | 	comp_t 	*depth_comp_curr,
376 | 	index_t		*adj_card_d,
377 | 	index_t		num_agg_bfs,
378 | 	depth_t	 curr_level,
379 | 	index_t	 vert_count,
380 | 	const index_t sml_shed,
381 | 	const index_t lrg_shed
382 | )
383 | {
384 | 	const index_t TID_ST	= threadIdx.x +blockIdx.x*blockDim.x;
385 | 	const index_t NUM_VER	= vert_count;	
386 | 	const index_t GRNTY		= blockDim.x*gridDim.x;
387 | 
388 | 	index_t ex_sml_ptr	= ex_sml_off[TID_ST];
389 | 	index_t ex_mid_ptr	= ex_mid_off[TID_ST];
390 | 	index_t ex_lrg_ptr	= ex_lrg_off[TID_ST];
391 | 
392 | 	index_t card_curr, card_next;
393 | 	comp_t depth_curr,depth_next;
394 | 	index_t	tid,tid_next;
395 | 	
396 | 	//Figure out its own start and end pos
397 | 	//We want each thread to inspect a continuous block
398 | 	index_t step_sz = NUM_VER/GRNTY;
399 | 	if(step_sz<16){
400 | 		//small problem size
401 | 		const index_t REMAINDER = NUM_VER-step_sz*GRNTY;
402 | 		if(TID_ST<REMAINDER) step_sz ++;
403 | 		const index_t beg_pos=step_sz*TID_ST+ 
404 | 					(TID_ST >= REMAINDER ? REMAINDER:0);
405 | 		const index_t end_pos=beg_pos+step_sz;
406 | 		
407 | 		tid=beg_pos;tid_next=beg_pos;
408 | 		if(step_sz){
409 | 			card_curr		= adj_card_d[beg_pos];
410 | 			depth_curr	= depth_comp_curr[beg_pos];
411 | 			tid_next++;
412 | 		}
413 | 			
414 | 		while(tid<end_pos){
415 | 			if(tid_next < end_pos){
416 | 				card_next		= adj_card_d[tid_next];
417 | 				depth_next	= depth_comp_curr[tid_next];
418 | 			}
419 | 			
420 | 			if((card_curr>0)&&(
421 | 				depth_curr.x!=0xffffffff ||
422 | 				depth_curr.y!=0xffffffff ||
423 | 				depth_curr.z!=0xffffffff ||
424 | 				depth_curr.w!=0xffffffff 
425 | 				)){
426 | 				if(card_curr<sml_shed)
427 | 				{
428 | 					ex_sml_q[ex_sml_ptr]=tid;
429 | 					ex_sml_ptr++;
430 | 				}
431 | 				else if(card_curr>lrg_shed)
432 | 				{
433 | 					ex_lrg_q[ex_lrg_ptr]=tid;
434 | 					ex_lrg_ptr++;
435 | 				}
436 | 				else
437 | 				{
438 | 					ex_mid_q[ex_mid_ptr]=tid;
439 | 					ex_mid_ptr++;
440 | 				}
441 | 			}
442 | 			tid=tid_next;
443 | 			tid_next++;
444 | 			card_curr	= card_next;
445 | 			depth_curr=depth_next;
446 | 		}
447 | 	}else{
448 | 		//big problem size
449 | 		//We want each thread to get 16x indices to check.
450 | 		if(step_sz&0xf) step_sz=((step_sz>>4)+1)<<4;
451 | 		if(NUM_VER-step_sz*GRNTY>0) step_sz<<=1;	
452 | 		__shared__ index_t	beg_pos_s[THDS_NUM];
453 | 		__shared__ depth_t	mark_s[THDS_NUM<<4];
454 | 		index_t	beg_pos_own			= step_sz*TID_ST;
455 | 		beg_pos_s[threadIdx.x] 	= beg_pos_own;
456 | 		const index_t TRIES 		= step_sz>>4;
457 | 		const index_t lane_id		= threadIdx.x%32;
458 | 		const index_t	warp_id		= threadIdx.x>>5;
459 | 		const index_t	thread_off= threadIdx.x<<4;
460 | 		index_t tries = 0;
461 | 		index_t	load_ptr;
462 | 		index_t	proc_vert;
463 | 
464 | 		//+--------------------
465 | 		//in shared memory data representation
466 | 		//---------------------------------------
467 | 		//			0000|1111
468 | 		//					 ||||
469 | 		//					 |||+-------- INFTY ? 
470 | 		//					 ||+--------- SML   ?
471 | 		//					 |+---------- MID   ?
472 | 		//					 +----------- LRG   ?
473 | 		//FURTHER OPTIMIZATION CAN BE CARRIED OUT
474 | 		//BY PACKING MORE HERE.
475 | 		//-----------------------------------------
476 | 		while(tries<TRIES){
477 | 			//warp stride loading
478 | 			for(int i=0; i<32; i+=2){
479 | 				proc_vert=(warp_id<<5)+i+(lane_id>>4);
480 | 				depth_curr.x=0xffffffff;
481 | 				depth_curr.y=0xffffffff;
482 | 				depth_curr.z=0xffffffff;
483 | 				depth_curr.w=0xffffffff;
484 | 				if(proc_vert<NUM_VER)
485 | 				{
486 | 					load_ptr=(tries<<4)+beg_pos_s[proc_vert]+(lane_id%16);
487 | 					if(load_ptr<NUM_VER)
488 | 					{
489 | 						depth_curr=depth_comp_curr[load_ptr];
490 | 						card_curr=adj_card_d[load_ptr];
491 | 					}else card_curr=0;//out-of-boundary, set it orphan
492 | 				}else card_curr=0;//out-of-boundary, set it orphan
493 | 				mark_s[(proc_vert<<4)+(lane_id%16)]=0x00;
494 | 
495 | 				if((card_curr>0) &&(
496 | 				depth_curr.x!=0xffffffff ||
497 | 				depth_curr.y!=0xffffffff ||
498 | 				depth_curr.z!=0xffffffff ||
499 | 				depth_curr.w!=0xffffffff))
500 | 					if(card_curr<sml_shed)
501 | 						mark_s[(proc_vert<<4)+(lane_id%16)]=0x03;//0011
502 | 					else if(card_curr>lrg_shed)
503 | 						mark_s[(proc_vert<<4)+(lane_id%16)]=0x09;//1001
504 | 					else mark_s[(proc_vert<<4)+(lane_id%16)]=0x05;//0101
505 | 			}
506 | 			__syncthreads();
507 | 
508 | 			//thread stride checking
509 | 			for(int i=0; i<16; i++)
510 | 				if(mark_s[thread_off+i]&0x02){
511 | 					ex_sml_q[ex_sml_ptr]=beg_pos_own+(tries<<4)+i;
512 | 					ex_sml_ptr++;
513 | 				}else if(mark_s[thread_off+i]&0x08){
514 | 					ex_lrg_q[ex_lrg_ptr]=beg_pos_own+(tries<<4)+i;
515 | 					ex_lrg_ptr++;
516 | 				}else if(mark_s[thread_off+i]&0x04){
517 | 					ex_mid_q[ex_mid_ptr]=beg_pos_own+(tries<<4)+i;
518 | 					ex_mid_ptr++;
519 | 				}
520 | 
521 | 			++tries;
522 | 			__syncthreads();
523 | 		}
524 | 	}
525 | }
526 | 
527 | 
528 | //+--------------------------------------------
529 | //|CLASSIFIED BASED STORAGE FOR EX_QUEUE
530 | //+---------------------------------------------------
531 | template< typename vertex_t, 
532 | 			typename index_t,
533 | 			typename depth_t>
534 | void td_inspect
535 | (
536 | 	tdata *gdata,
537 | 	csr_graph ggraph,
538 | 	index_t		num_agg_bfs,
539 | 	index_t	 sw_level,
540 | 	depth_t	 curr_level,
541 | 	index_t  vert_count,//num ver in the graph
542 | 	const index_t sml_shed,
543 | 	const index_t lrg_shed
544 | )
545 | {
546 | 	td_frontier_count	
547 | 	<vertex_t, index_t, depth_t>
548 | 	<<<BLKS_NUM,THDS_NUM>>>
549 | 	(
550 | 		ggraph->ex_cat_sml_sz,
551 | 		ggraph->ex_cat_mid_sz,
552 | 		ggraph->ex_cat_lrg_sz,
553 | 		ggraph->depth_comp_last,
554 | 		ggraph->depth_comp_curr,
555 | 		ggraph->depth_merge,
556 | 		ggraph->adj_card_d,
557 | 		num_agg_bfs,
558 | 		curr_level,
559 | 		vert_count,
560 | 		sml_shed,
561 | 		lrg_shed
562 | 	);
563 | 	cudaDeviceSynchronize();
564 | 	
565 | 	insp_scan
566 |  	<vertex_t, index_t>
567 |  	(
568 |  		ggraph->ex_cat_sml_sz,
569 |  		ggraph->ex_sml_off,
570 |  		THDS_NUM*BLKS_NUM,
571 |  		BLKS_NUM>>1,
572 |  		THDS_NUM>>1,
573 |  		gdata[0]->ex_sml_sz_d,
574 |  		gdata[0]->stream[0]
575 |  	);
576 |  	insp_scan
577 |  	<vertex_t, index_t>
578 |  	(
579 |  		ggraph->ex_cat_mid_sz,
580 |  		ggraph->ex_mid_off,
581 |  		THDS_NUM*BLKS_NUM,
582 |  		BLKS_NUM>>1,
583 |  		THDS_NUM>>1,
584 |  		gdata[0]->ex_mid_sz_d,
585 |  		gdata[0]->stream[1]
586 |  	);
587 |  	insp_scan
588 |  	<vertex_t, index_t>
589 |  	(
590 |  		ggraph->ex_cat_lrg_sz,
591 |  		ggraph->ex_lrg_off,
592 |  		THDS_NUM*BLKS_NUM,
593 |  		BLKS_NUM>>1,
594 |  		THDS_NUM>>1,
595 |  		gdata[0]->ex_lrg_sz_d,
596 |  		gdata[0]->stream[2]
597 |  	);
598 |  	
599 | 
600 | 	cudaStreamSynchronize(gdata[0]->stream[0]);
601 | 	cudaStreamSynchronize(gdata[0]->stream[1]);
602 | 	cudaStreamSynchronize(gdata[0]->stream[2]);
603 | 	cudaDeviceSynchronize();
604 | 	
605 | 	td_frontier_gather	
606 | 	<vertex_t, index_t, depth_t>
607 | 	<<<BLKS_NUM,THDS_NUM>>>
608 | 	(
609 | 		ggraph->ex_sml_q,
610 | 		ggraph->ex_mid_q,
611 | 		ggraph->ex_lrg_q,
612 | 		ggraph->ex_sml_off,
613 | 		ggraph->ex_mid_off,
614 | 		ggraph->ex_lrg_off,
615 | 		ggraph->depth_comp_last,
616 | 		ggraph->depth_comp_curr,
617 | 		ggraph->depth_merge,
618 | 		ggraph->adj_card_d,
619 | 		num_agg_bfs,
620 | 		curr_level,
621 | 		vert_count,
622 | 		sml_shed,
623 | 		lrg_shed
624 | 	);
625 | 
626 | }
627 | 
628 | template< typename vertex_t, 
629 | 			typename index_t,
630 | 			typename depth_t>
631 | void sw_inspect
632 | (
633 | 	tdata *gdata,
634 | 	csr_graph ggraph,
635 | 	index_t		num_agg_bfs,
636 | 	depth_t	 curr_level,
637 | 	index_t  vert_count,//num ver in the graph
638 | 	const index_t sml_shed,
639 | 	const index_t lrg_shed
640 | )
641 | {
642 | 	ENABLE_BTUP=true;
643 | 	extract_depth
644 | 	<depth_t, comp_t, index_t>
645 | 	<<<BLKS_NUM<<2,THDS_NUM<<2,0,ggraph->gstream>>>
646 | 	(
647 | 		ggraph->depth_comp_last,
648 | 		ggraph->depth_comp_curr,
649 | 		ggraph->depth_merge,
650 | 		num_agg_bfs,
651 | 		curr_level,
652 | 		vert_count
653 | 	);
654 | 	
655 | 	//ONLY COMPRESS ONE TIME
656 | //	if(ENABLE_BTUP==false){
657 | //		ENABLE_BTUP=true;
658 | //		comp_depth
659 | //		<depth_t, comp_t, index_t>
660 | //		<<<BLKS_NUM<<2,THDS_NUM<<2>>>
661 | //		(
662 | //			ggraph->depth_merge,
663 | //			ggraph->depth_comp_last,
664 | //			ggraph->depth_comp_curr,
665 | //			curr_level,
666 | //			0,
667 | //			vert_count
668 | //		);
669 | //		cudaDeviceSynchronize();
670 | //	}
671 | 	
672 | 	sw_frontier_count
673 |  	<vertex_t,index_t,depth_t>
674 |  	<<<BLKS_NUM,THDS_NUM>>>
675 |  	(
676 |  		ggraph->ex_cat_sml_sz,
677 |  		ggraph->ex_cat_mid_sz,
678 |  		ggraph->ex_cat_lrg_sz,
679 |  		ggraph->depth_comp_curr,
680 |  		ggraph->adj_card_d,
681 |  		num_agg_bfs,
682 |  		curr_level,
683 |  		vert_count,
684 |  		sml_shed,
685 |  		lrg_shed
686 |  	);
687 | 	
688 | 	cudaDeviceSynchronize();
689 | 	
690 | 	insp_scan
691 |  	<vertex_t, index_t>
692 |  	(
693 |  		ggraph->ex_cat_sml_sz,
694 |  		ggraph->ex_sml_off,
695 |  		THDS_NUM*BLKS_NUM,
696 |  		BLKS_NUM>>1,
697 |  		THDS_NUM>>1,
698 |  		gdata[0]->ex_sml_sz_d,
699 |  		gdata[0]->stream[0]
700 |  	);
701 |  	insp_scan
702 |  	<vertex_t, index_t>
703 |  	(
704 |  		ggraph->ex_cat_mid_sz,
705 |  		ggraph->ex_mid_off,
706 |  		THDS_NUM*BLKS_NUM,
707 |  		BLKS_NUM>>1,
708 |  		THDS_NUM>>1,
709 |  		gdata[0]->ex_mid_sz_d,
710 |  		gdata[0]->stream[1]
711 |  	);
712 |  	insp_scan
713 |  	<vertex_t, index_t>
714 |  	(
715 |  		ggraph->ex_cat_lrg_sz,
716 |  		ggraph->ex_lrg_off,
717 |  		THDS_NUM*BLKS_NUM,
718 |  		BLKS_NUM>>1,
719 |  		THDS_NUM>>1,
720 |  		gdata[0]->ex_lrg_sz_d,
721 |  		gdata[0]->stream[2]
722 |  	);
723 |  	
724 | 
725 | 	cudaStreamSynchronize(gdata[0]->stream[0]);
726 | 	cudaStreamSynchronize(gdata[0]->stream[1]);
727 | 	cudaStreamSynchronize(gdata[0]->stream[2]);
728 | 	cudaDeviceSynchronize();
729 | 	
730 | 	sw_frontier_gather
731 | 	<vertex_t, index_t, depth_t>
732 | 	<<<BLKS_NUM, THDS_NUM>>>
733 | 	(
734 | 		ggraph->ex_sml_q,
735 | 		ggraph->ex_mid_q,
736 | 		ggraph->ex_lrg_q,
737 | 		ggraph->ex_sml_off,
738 | 		ggraph->ex_mid_off,
739 | 		ggraph->ex_lrg_off,
740 |  		ggraph->depth_comp_curr,
741 | 		ggraph->adj_card_d,
742 | 		num_agg_bfs,
743 | 		curr_level,
744 | 		vert_count,
745 | 		sml_shed,
746 | 		lrg_shed
747 | 	);
748 | 	cudaStreamSynchronize(ggraph->gstream);
749 | }
750 | 
751 | 
752 | template<typename vertex_t, 
753 | 			typename index_t,
754 | 			typename depth_t,
755 | 			typename comp_t>
756 | void inspector
757 | (
758 | 	tdata *gdata,
759 | 	csr_graph ggraph,
760 | 	index_t		num_agg_bfs,
761 | 	index_t		sw_level,
762 | 	depth_t	 	level,
763 | 	index_t  	vert_count,
764 | 	const index_t sml_shed,
765 | 	const index_t lrg_shed
766 | )
767 | {
768 | //		comp_t *d_status_curr,*d_status_last;
769 | //		cudaMallocHost((void **)&d_status_curr, sizeof(comp_t)*vert_count);
770 | //		cudaMallocHost((void **)&d_status_last, sizeof(comp_t)*vert_count);
771 | //		cudaMemcpy(d_status_curr, ggraph->depth_comp_curr,sizeof(comp_t)*vert_count,cudaMemcpyDeviceToHost);
772 | //		cudaMemcpy(d_status_last, ggraph->depth_comp_last,sizeof(comp_t)*vert_count,cudaMemcpyDeviceToHost);
773 | //		
774 | //		int ft_count=0; int diff=0;
775 | //		for(int i=0;i<vert_count;i++){
776 | //			if(d_status_curr[i]!=d_status_last[i]) diff++;
777 | //			if(d_status_curr[i]!=0xffffffffffffffff && ggraph->adj_card[i]>0) ft_count++;
778 | //		}
779 | //		std::cout<<"ft_count-diff "<<ft_count<<" "<<diff<<"\n";
780 | //		cudaFreeHost(d_status_last);
781 | //		cudaFreeHost(d_status_curr);
782 | //	
783 | 	
784 | 	if(level<sw_level)
785 | 		td_inspect
786 | 		<vertex_t, index_t, depth_t>
787 | 		(
788 | 			gdata,
789 | 			ggraph,
790 | 			num_agg_bfs,
791 | 			sw_level,
792 | 			level,
793 | 			vert_count,
794 | 			sml_shed,
795 | 			lrg_shed
796 | 		);
797 | 	else{
798 | 		
799 | 		sw_inspect
800 | 		<vertex_t, index_t, depth_t>
801 | 		(
802 | 			gdata,
803 | 			ggraph,
804 | 			num_agg_bfs,
805 | 			level,
806 | 			vert_count,
807 | 			sml_shed,
808 | 			lrg_shed
809 | 		);
810 | 	}
811 | }
812 | 


--------------------------------------------------------------------------------
/launch.bash:
--------------------------------------------------------------------------------
 1 | #declare -a name=(fb ok kg1 rd hw tw pk wk)
 2 | declare -a name=(rm)
 3 | 
 4 | for ((gcount=1;gcount<=32;gcount=gcount*2))
 5 | do
 6 | 	for file in ${name[@]}
 7 | 	do
 8 | 		file_inst=jobScript"$gcount"gpu-$file
 9 | 		if [ -f $file_inst ]
10 | 		then
11 | 			sbatch $file_inst 
12 | 		fi
13 | 	done
14 | done
15 | 


--------------------------------------------------------------------------------
/main.cpp:
--------------------------------------------------------------------------------
 1 | #include "ibfs.h"
 2 | #include <mpi.h>
 3 | #include <cuda.h>
 4 | #include <cuda_runtime.h>
 5 | 
 6 | #define USE_GPU_ID		4 
 7 | int main(int args, char *argv[]) {
 8 | 	
 9 | 	MPI_Init(&args,&argv);
10 | 	int world_sz;
11 | 	MPI_Comm_size(MPI_COMM_WORLD, &world_sz);
12 | 	
13 | 	int my_id;
14 | 	MPI_Comm_rank(MPI_COMM_WORLD, &my_id);
15 | 	const index_t gpu_id 	= USE_GPU_ID;
16 | 	ibfs(world_sz, my_id, gpu_id, args, argv);
17 | 
18 | 	MPI_Finalize();
19 | 	return 0;
20 | }
21 | 
22 | 


--------------------------------------------------------------------------------
/main.cu:
--------------------------------------------------------------------------------
 1 | /*
 2 |  * Copyright 2016 The George Washington University
 3 |  * Written by Hang Liu 
 4 |  * Directed by Prof. Howie Huang
 5 |  *
 6 |  * https://www.seas.gwu.edu/~howie/
 7 |  * Contact: iheartgraph@gmail.com
 8 |  *
 9 |  * 
10 |  * Please cite the following paper:
11 |  * 
12 |  * Hang Liu, H. Howie Huang and Yang Hu. 2016. iBFS: Concurrent Breadth-First Search on GPUs. Proceedings of the 2016 International Conference on Management of Data. ACM. 
13 |  *
14 |  * This file is part of iBFS.
15 |  *
16 |  * iBFS is free software: you can redistribute it and/or modify
17 |  * it under the terms of the GNU General Public License as published by
18 |  * the Free Software Foundation, either version 3 of the License, or
19 |  * (at your option) any later version.
20 |  *
21 |  * iBFS is distributed in the hope that it will be useful,
22 |  * but WITHOUT ANY WARRANTY; without even the implied warranty of
23 |  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the 
24 |  * GNU General Public License for more details.
25 |  *
26 |  * You should have received a copy of the GNU General Public License
27 |  * along with iBFS.  If not, see <http://www.gnu.org/licenses/>.
28 |  */
29 | 
30 | #include "gpu_ibfs.cuh"
31 | #include "graph.h"
32 | #include <sstream>
33 | #include <iostream>
34 | #include <fstream>
35 | 
36 | int main(int args, char *argv[]) {
37 | 	typedef long vertex_t;
38 | 	typedef long index_t;
39 | 	typedef unsigned char depth_t;
40 | 	typedef unsigned long long int comp_t;
41 | 	std::cout<<"Input: /path/to/exe /path/to/beg "
42 | 				<<"/path/to/csr concurrent-count " 
43 | 				<<"switch-level bfs_count\n";
44 | 	if(args != 6){std::cout<<"Wrong input\n";exit(-1);}
45 | 	
46 | 	const char *beg_file=argv[1];
47 | 	const char *csr_file=argv[2];
48 | 	const int concurr_count=atoi(argv[3]);
49 | 	const depth_t sw_level=atoi(argv[4]);
50 | 	const index_t src_count=atoi(argv[5]);
51 | 	graph *g=new graph(beg_file, csr_file, src_count);
52 | 	g->gen_src();
53 | 	g->groupby();
54 | 
55 | 	cudaSetDeviceFlags(cudaDeviceMapHost);
56 | 	const index_t sml_shed 	= 32; 
57 | 	const index_t lrg_shed 	= 512;
58 | 	index_t gpu_count	= 1;
59 | 	
60 | 	gpu_ibfs *ibfs_inst = new gpu_ibfs 
61 | 		((const graph *)g, 
62 | 			concurr_count, 
63 | 			sw_level,
64 | 			gpu_count, 
65 | 			sml_shed, 
66 | 			lrg_shed);
67 | 	
68 | 
69 | 	//keep track of orphans
70 | 	comp_t *tmp_depth;
71 | 	H_ERR(cudaMallocHost((void **)&tmp_depth, 
72 | 			sizeof(comp_t)*g->vert_count*ibfs_inst->joint_count));
73 | 	
74 | 	std::cout<<"Start bfs ...\n";
75 | 	double tm_ibfs=wtime();
76 | 	for(index_t i=0; i<src_count/concurr_count; i++)
77 | 	{		
78 | 		index_t grp_beg=i*concurr_count;
79 | 		cudaMemcpy(ibfs_inst->depth_comp_last,tmp_depth,
80 | 			sizeof(comp_t)*g->vert_count*ibfs_inst->joint_count, 
81 | 			cudaMemcpyHostToDevice);
82 | 		cudaMemcpy(ibfs_inst->depth_comp_curr,tmp_depth,
83 | 			sizeof(comp_t)*g->vert_count*ibfs_inst->joint_count, 
84 | 			cudaMemcpyHostToDevice);
85 | 	
86 | 		double tm_ibfs=wtime();
87 | 		ibfs_inst->traversal(grp_beg);
88 | 		std::cout<<"Traversal-time: "<<wtime()-tm_ibfs<<" second(s)\n";
89 | 		//graph_d->write_result();
90 | 	}
91 | 	std::cout<<"Total time: "<<wtime()-tm_ibfs<<"\n";
92 | 	return 0;
93 | }
94 | 


--------------------------------------------------------------------------------
/proc_profil.bash:
--------------------------------------------------------------------------------
 1 | #!/bin/bash
 2 | 
 3 | declare -a file=(desc-fb-24  desc-fr-24  desc-hw-24  desc-kg1-24  desc-kg2-24 desc-lj-24  desc-ok-24  desc-pk-24  desc-rd-24 desc-rm-24  desc-tw-24 desc-wk-24)
 4 | #declare -a kernel=(THD_expand_sort WAP_expand_sort CTA_expand_sort THD_bu_expand_sort WAP_bu_expand_sort CTA_bu_expand_sort)
 5 | declare -a kernel=(sw_expand_warp sw_expand_thd)
 6 | declare -a metric=(gld_transactions gst_transactions)
 7 | 
 8 | kens=""
 9 | count=0
10 | for ken in ${kernel[@]}
11 | do
12 | 	if [ $count -eq 0 ]
13 | 	then
14 | 		kens=$ken
15 | 	else
16 | 		kens=$kens"|"$ken
17 | 	fi
18 | 	count=$((count+1))
19 | done
20 | echo $kens
21 | 
22 | for met in ${metric[@]}
23 | do
24 | 	echo $met
25 | 	for fle in ${file[@]}
26 | 	do
27 | 		grep -E "$kens|$met" $fle.3.6.gld.gst.count | grep -A 1 "Kernel" | sed -e '/^--$/d'|awk -v av=0 -v count=0 '{if(NR%2==0){av+=$NF; count++;}}END{printf "%.2f\n",av/count}' 
28 | 	done
29 | done
30 | 


--------------------------------------------------------------------------------
/proc_time.bash:
--------------------------------------------------------------------------------
 1 | #declare -a file=(desc-fb-24  desc-fr-24  desc-hw-24  desc-kg0-24  desc-kg1-24  desc-kg2-24 desc-lj-24  desc-ok-24  desc-pk-24  desc-rd-24 desc-rm-24  desc-tw-24 desc-wk-24)
 2 | #declare -a file=(desc-hw-24 desc-lj-24 desc-tw-24 desc-wk-24 desc-pk-24)
 3 | declare -a file=(desc-hw-24)
 4 | 
 5 | #echo "#Share diff hub"
 6 | #for name in ${file[@]}
 7 | #do
 8 | #	#echo $name
 9 | #	i=1
10 | #	for share in 4 16 64 128
11 | #	do
12 | #		fqsz=$(grep "Break-fq-sz" $name-diff-outdegree-"$share".log | awk -F " " -v fq=10000000 '{if(fq>$2)fq=$2}END{print fq}')
13 | #		array[$i]=$(grep -E "Break-fq-sz: $fqsz|Traversal-" $name-diff-outdegree-"$share".log | grep -A 1 Break | sed -e "/--/d;/Break/d" | awk -F " " -v count=0 -v time=0 '{time+=$2;count++}END{print time/count}')
14 | #		i=$((i+1))	
15 | #	done
16 | #	printf "%s," "$name"
17 | #	printf "%s," "${array[@]}"
18 | #	echo
19 | #done
20 | 
21 | 
22 | 
23 | echo "#Share same hub"
24 | for name in ${file[@]}
25 | do
26 | 	#echo $name
27 | 	i=1
28 | 	for ((num_groups=1;num_groups<=1048576;num_groups=num_groups*2)) 
29 | 	#for share in 16
30 | 	#for num_groups in 1 2 3 4 5 6 7 8 9 10 50 100 200 300 400 500 600 700 800 900 1000
31 | 	do
32 | 		fqsz=$(grep "Break-fq-sz" "$name"."$num_groups".nogroupby.log | awk -F " " -v fq=10000000 '{if(fq>$2)fq=$2}END{print fq}')
33 | 		array[$i]=$(grep -E "Break-fq-sz: $fqsz|Traversal-" "$name"."$num_groups".nogroupby.log | grep -A 1 Break | sed -e "/--/d;/Break/d" | awk -F " " -v count=0 -v time=0 '{time+=$2;count++}END{if(count!=0){print time/count}}')
34 | 		i=$((i+1))	
35 | 	done
36 | 	printf "%s\n" "$name"
37 | 	printf "%s\n" "${array[@]}"
38 | 	echo
39 | done
40 | 


--------------------------------------------------------------------------------
/readme.md:
--------------------------------------------------------------------------------
 1 | ----
 2 | Software requirement
 3 | -----
 4 | gcc 4.4.7 or 4.8.5
 5 | 
 6 | CUDA 5.5, 6.0, 6.5, 7.0, 7.5 (tested and works)
 7 | 
 8 | MPI: MPICH version 3.3 or higher, corresponding OpenMPI should also work
 9 | 
10 | 
11 | ----
12 | Hardware
13 | ------
14 | GPU: C2050, C2070, K20, K40 (tested)
15 | 
16 | ----
17 | Compile
18 | -----
19 | 
20 | make
21 | 
22 | ----
23 | Execute
24 | ------
25 | Type: "./gpu_ibfs.bin" it will show you what is needed.
26 | 
27 | **Atention**, please make *concurrent-count* as **128**
28 | 
29 | You could use the code from "tuple_text_to_binary_csr" folder to convert a edge list (text formated) graph into CSR files (binary), e.g., if you have a text file called "test.txt", it will convert it into "test.txt_beg_pos.bin" and "test.txt_csr.bin". You will need these two files to run gpu_ibfs.
30 | 
31 | ----
32 | Converter: edge tuples to CSR
33 | ----
34 | - Compile: make
35 | - To execute: type "./text_to_bin", it will show you what is needed
36 | - Basically, you could download a tuple list file from [snap](https://snap.stanford.edu/data/). Afterwards, you could use this converter to convert the edge list into CSR format. 
37 | 
38 | **For example**:
39 | 
40 | - Download https://snap.stanford.edu/data/com-Orkut.html file. **unzip** it. 
41 | - **./text_to_bin.bin soc-orkut.mtx 1 2(the number may change due to how many lines are not edges)**
42 | - You will get *soc-orkut.mtx_beg_pos.bin* and *soc-orkut.mtx_csr.bin*. 
43 | - You could use these two files to run enterprise.
44 | 
45 | ----
46 | Applications of iBFS
47 | ---------
48 | - Centrality computation, e.g., Betweenness centrality, Closeness centrality.
49 | - Multi-source shortest path, All-pairs shortest path.
50 | - Reachiability index construction.
51 | - **In short, any applications, which need to run multiple traversals on the same graph, should benefit**
52 | 
53 | ----
54 | Reference
55 | -------
56 | 
57 | [SC '15] Enterprise: Breadth-First Graph Traversal on GPUs [[PDF](http://hang-liu.com/publication/enterprise_sc15.pdf)] [[Slides](http://hang-liu.com/publication/enterprise_slides.pdf)] [[Blog](http://hang-liu.com/enterprise_blog.html)]
58 | 
59 | [SIGMOD '16] iBFS: Concurrent Breadth-First Search on GPUs [[PDF](http://hang-liu.com/publication/ibfs.pdf)] [[Slides](http://hang-liu.com/publication/ibfs_slides.pdf)] [[Poster](http://hang-liu.com/publication/ibfs_poster.pdf)]
60 | 


--------------------------------------------------------------------------------
/reporter.cpp:
--------------------------------------------------------------------------------
 1 | #include <mpi.h>
 2 | #include "ibfs.h"
 3 | #include <iostream>
 4 | 
 5 | int reporter(double tm, int my_id, index_t iter)
 6 | {
 7 | 	MPI_Barrier(MPI_COMM_WORLD);
 8 | 	double max_tm;
 9 | 	MPI_Reduce(&tm, &max_tm, 1, MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
10 | 	if(my_id==0)std::cout<<"Traversal-iter-"<<iter<<": "<<max_tm<<" second(s)\n";
11 | 	MPI_Barrier(MPI_COMM_WORLD);
12 | 	return 0;
13 | }
14 | 


--------------------------------------------------------------------------------
/run.bash:
--------------------------------------------------------------------------------
 1 | #declare -a file=(desc-fb-24  desc-fr-24  desc-hw-24  desc-kg0-24  desc-kg1-24  desc-kg2-24 desc-lj-24  desc-ok-24  desc-pk-24  desc-rd-24 desc-rm-24  desc-tw-24 desc-wk-24)
 2 | #declare -a file=(desc-hw-24 desc-lj-24 desc-tw-24 desc-wk-24)
 3 | #declare -a file=(desc-hw-24 desc-tw-24)
 4 | declare -a file=(desc-hw-24)
 5 | 
 6 | for name in ${file[@]}
 7 | do
 8 | 	echo $name
 9 | 	for ((i=16;i<=1048576;i=i*2)) 
10 | 	do
11 | 		stdbuf -o 0 ./ibfs-5.0 ../desc-files/$name haha.input $i | tee $name.$i.nogroupby.log
12 | 	done
13 | done
14 | 


--------------------------------------------------------------------------------
/scan.cuh:
--------------------------------------------------------------------------------
  1 | 
  2 | #define NUM_BANKS           32
  3 | #define LOG_NUM_BANKS       5
  4 | 
  5 | #define CONFLICT_FREE_OFFSET(n) \
  6 |     ((n) >>LOG_NUM_BANKS)
  7 | 
  8 | //USED FOR EXPANSION
  9 | //template <typename data_t, typename index_t>
 10 | //__global__ void __pre_scan(	
 11 | //								data_t 	*scan_ind_d,
 12 | //								data_t	*adj_card_d,
 13 | //								data_t 	*scan_out_d,
 14 | //								data_t 	*blk_sum,
 15 | //								index_t num_dat,
 16 | //								const index_t THD_NUM
 17 | //		   				   )
 18 | //{
 19 | //	const data_t tile_sz	= THD_NUM<<1;
 20 | //	const index_t lane		= threadIdx.x<<1;
 21 | //	index_t tid				= threadIdx.x+blockIdx.x*blockDim.x;
 22 | //	index_t proc_tile		= blockIdx.x;
 23 | //
 24 | //	index_t offset			= 1;
 25 | //	index_t tid_strip;
 26 | //	index_t idct_a, idct_b;
 27 | //	const index_t padding	= CONFLICT_FREE_OFFSET(tile_sz - 1);
 28 | //	const index_t NUM_THS	= gridDim.x*blockDim.x;
 29 | //
 30 | //	//conflict free
 31 | //	const index_t off_a		= CONFLICT_FREE_OFFSET(lane);
 32 | //	const index_t off_b		= CONFLICT_FREE_OFFSET(lane+1);
 33 | //	index_t num_tiles		= num_dat/tile_sz;
 34 | //	if(num_dat%tile_sz)
 35 | //		num_tiles++;
 36 | //
 37 | //	//prefetching danger
 38 | //	extern __shared__ data_t s_mem[];
 39 | //	//__shared__ data_t s_mem[280];
 40 | //	
 41 | //	while(proc_tile < num_tiles)
 42 | //	{
 43 | //		tid_strip = tid<<1;
 44 | //		if(tid_strip < num_dat)
 45 | //		{
 46 | //			idct_a	= scan_ind_d[tid_strip];
 47 | //			s_mem[lane+off_a]= adj_card_d[idct_a];
 48 | //		}else{
 49 | //			s_mem[lane+off_a]= 0;
 50 | //		}
 51 | //
 52 | //		if((tid_strip + 1) < num_dat)
 53 | //		{
 54 | //			idct_b	= scan_ind_d[tid_strip + 1];
 55 | //			s_mem[lane + 1 + off_b]	= adj_card_d[idct_b];
 56 | //		}else{
 57 | //			s_mem[lane + 1 + off_b] = 0;
 58 | //		}
 59 | //		__syncthreads();
 60 | //
 61 | //		//up sweep
 62 | //		for(index_t j = THD_NUM;j > 0;j>>=1)
 63 | //		{   
 64 | //			if(threadIdx.x < j)
 65 | //			{
 66 | //				index_t ai	= offset*lane +offset - 1;
 67 | //				index_t bi	= ai + offset;
 68 | //				ai		   += CONFLICT_FREE_OFFSET(ai);
 69 | //				bi		   += CONFLICT_FREE_OFFSET(bi);
 70 | //				s_mem[bi]  += s_mem[ai];
 71 | //			}
 72 | //			offset <<=1;
 73 | //			__syncthreads();
 74 | //		}   
 75 | //		__syncthreads();
 76 | //
 77 | //		//write the block sum
 78 | //		if(blockDim.x > 1)//save sml scan job
 79 | //		{
 80 | //			if(threadIdx.x == 0)
 81 | //			{
 82 | //				blk_sum[tid/THD_NUM]	= 
 83 | //							s_mem[tile_sz -1 + padding];
 84 | //				s_mem[tile_sz-1 + padding]		= 0;
 85 | //			}
 86 | //			__syncthreads();
 87 | //		}
 88 | //
 89 | //		//down sweep
 90 | //		for(index_t j=1; j < tile_sz; j <<=1)
 91 | //		{
 92 | //			offset	>>=	1;
 93 | //			if(threadIdx.x < j)
 94 | //			{
 95 | //				index_t ai	= lane*offset + offset - 1;
 96 | //				index_t bi	= ai + offset;
 97 | //				ai		   += CONFLICT_FREE_OFFSET(ai);
 98 | //				bi		   += CONFLICT_FREE_OFFSET(bi);
 99 | //				index_t t 	= s_mem[ai];
100 | //				s_mem[ai]	= s_mem[bi];
101 | //				s_mem[bi]  += t;
102 | //			}
103 | //			__syncthreads();
104 | //		}
105 | //		__syncthreads();
106 | //
107 | //		//store data
108 | //		if(tid_strip < num_dat)
109 | //		{
110 | //			scan_out_d[tid_strip]	= s_mem[lane+off_a]; 
111 | //		}
112 | //		
113 | //		if((tid_strip + 1) < num_dat)
114 | //		{
115 | //			scan_out_d[tid_strip+1] 	= s_mem[lane+1+off_b]; 
116 | //		}
117 | //
118 | //		tid			+= NUM_THS;
119 | //		proc_tile	+= gridDim.x;
120 | //	}
121 | //}
122 | //
123 | ////USED FOR lrg_scan
124 | //template <typename data_t, typename index_t>
125 | //__global__ void __spine_scan(
126 | //								data_t 	*blk_sum, 
127 | //								data_t 	*grd_sum,
128 | //								index_t num_dat,
129 | //								const index_t THD_NUM
130 | //		   				   )
131 | //{
132 | //	extern __shared__ data_t s_mem[];
133 | //
134 | //	const data_t tile_sz	= THD_NUM<<1;
135 | //	const index_t lane		= threadIdx.x<<1;
136 | //	const index_t GRNTY= blockDim.x*gridDim.x;
137 | //
138 | //	const index_t padding	= CONFLICT_FREE_OFFSET(tile_sz - 1);
139 | //	const index_t off_a		= CONFLICT_FREE_OFFSET(lane);
140 | //	const index_t off_b		= CONFLICT_FREE_OFFSET(lane+1);
141 | //	
142 | //	index_t tid				= threadIdx.x+blockIdx.x*blockDim.x;
143 | //	index_t offset			= 1;
144 | //	index_t	tid_strip		= tid<<1;
145 | //	index_t proc_tile		= blockIdx.x;
146 | //		
147 | //	index_t num_tiles		= num_dat/tile_sz;
148 | //	if(num_dat%tile_sz)
149 | //		num_tiles++;
150 | //
151 | //	while(proc_tile < num_tiles)
152 | //	{	
153 | //		tid_strip		= tid<<1;
154 | //		
155 | //		if(tid_strip < num_dat)
156 | //			s_mem[lane+off_a] = blk_sum[tid_strip];
157 | //		else
158 | //			s_mem[lane+off_a]	= 0;
159 | //
160 | //		if(tid_strip < num_dat)
161 | //			s_mem[lane+1+off_b]	= blk_sum[tid_strip + 1];
162 | //		else
163 | //			s_mem[lane+1+off_b] = 0;
164 | //	
165 | //		__syncthreads();
166 | //
167 | //		//up sweep
168 | //		for(index_t j = THD_NUM;j > 0;j>>=1)
169 | //		{   
170 | //			if(threadIdx.x < j)
171 | //			{
172 | //				index_t ai	= offset*lane +offset - 1;
173 | //				index_t bi	= ai + offset;
174 | //
175 | //				ai		   += CONFLICT_FREE_OFFSET(ai);
176 | //				bi		   += CONFLICT_FREE_OFFSET(bi);
177 | //				s_mem[bi] += s_mem[ai];
178 | //			}
179 | //			offset <<=1;
180 | //			__syncthreads();
181 | //		}   
182 | //		__syncthreads();
183 | //
184 | //		//write the block sum
185 | //		if(threadIdx.x == 0)
186 | //		{
187 | //			grd_sum[tid/THD_NUM]= 
188 | //						s_mem[tile_sz-1+padding];
189 | //			s_mem[tile_sz-1+padding]= 0;
190 | //		}
191 | //		__syncthreads();
192 | //
193 | //		//down sweep
194 | //		for(index_t j=1; j < tile_sz; j <<=1)
195 | //		{
196 | //			offset	>>=	1;
197 | //			if(threadIdx.x < j)
198 | //			{
199 | //				index_t ai	= lane*offset + offset - 1;
200 | //				index_t bi	= ai + offset;
201 | //				ai		   += CONFLICT_FREE_OFFSET(ai);
202 | //				bi		   += CONFLICT_FREE_OFFSET(bi);
203 | //				index_t t 	= s_mem[ai];
204 | //				s_mem[ai]	= s_mem[bi];
205 | //				s_mem[bi]  += t;
206 | //			}
207 | //			__syncthreads();
208 | //		}
209 | //		__syncthreads();
210 | //		if(tid_strip < num_dat)
211 | //			blk_sum[tid_strip]		= s_mem[lane+off_a];
212 | //		if(tid_strip + 1< num_dat)
213 | //			blk_sum[tid_strip+1]	= s_mem[lane+1+off_b];
214 | //		
215 | //		tid			+= GRNTY;
216 | //		proc_tile	+= gridDim.x;
217 | //	}
218 | //}
219 | //
220 | ///////////////////////////////////////////////////////
221 | ////post_scan for large problem
222 | ////////////////////////////////////////////////////////
223 | //template<typename data_t, typename index_t>
224 | //__global__ void __post_scan(
225 | //								data_t 	*scan_out_d,
226 | //								data_t 	*blk_sum,
227 | //								data_t 	*grd_sum,
228 | //								index_t	num_dat,
229 | //								index_t num_grd,
230 | //								const index_t THD_NUM
231 | //							)
232 | //{
233 | //	extern __shared__ data_t s_mem[];
234 | //	
235 | //	const data_t tile_sz	= THD_NUM<<1;
236 | //	const index_t lane		= threadIdx.x<<1;
237 | //	index_t tid				= threadIdx.x+blockIdx.x*blockDim.x;
238 | //	index_t offset			= 1;
239 | //	
240 | //	const index_t padding	= CONFLICT_FREE_OFFSET(tile_sz - 1);
241 | //	
242 | //	const index_t off_a		= CONFLICT_FREE_OFFSET(lane);
243 | //	const index_t off_b		= CONFLICT_FREE_OFFSET(lane+1);
244 | //	/////////////////////////////////////////////////////////
245 | //	//GRID SCAN
246 | //	/////////////////////////////////////////////////////////
247 | //	if(lane < num_grd)
248 | //		s_mem[lane+off_a]=grd_sum[lane];
249 | //	else
250 | //		s_mem[lane+off_a]=0;
251 | //
252 | //	if(lane+1 < num_grd)
253 | //		s_mem[lane+1+off_b]=grd_sum[lane+1];
254 | //	else
255 | //		s_mem[lane+1+off_b]=0;
256 | //	
257 | //	__syncthreads();
258 | //	
259 | //	//up sweep
260 | //	for(index_t j = THD_NUM;j > 0;j>>=1)
261 | //	{   
262 | //		if(threadIdx.x < j)
263 | //		{
264 | //			index_t ai	= offset*lane +offset - 1;
265 | //			index_t bi	= ai + offset;
266 | //
267 | //			ai		   += CONFLICT_FREE_OFFSET(ai);
268 | //			bi		   += CONFLICT_FREE_OFFSET(bi);
269 | //			s_mem[bi] += s_mem[ai];
270 | //		}
271 | //		offset <<=1;
272 | //		__syncthreads();
273 | //	}   
274 | //	__syncthreads();
275 | //
276 | //	//write the block sum
277 | //	if(threadIdx.x == 0)
278 | //	{
279 | //		if(!blockIdx.x)
280 | //			in_q_sz_d	= s_mem[tile_sz-1+padding];
281 | //		
282 | //		s_mem[tile_sz-1+padding]= 0;
283 | //	}
284 | //	__syncthreads();
285 | //
286 | //	//down sweep
287 | //	for(index_t j=1; j < tile_sz; j <<=1)
288 | //	{
289 | //		offset	>>=	1;
290 | //		if(threadIdx.x < j)
291 | //		{
292 | //			index_t ai	= lane*offset + offset - 1;
293 | //			index_t bi	= ai + offset;
294 | //			ai		   += CONFLICT_FREE_OFFSET(ai);
295 | //			bi		   += CONFLICT_FREE_OFFSET(bi);
296 | //			index_t t 	= s_mem[ai];
297 | //			s_mem[ai]	= s_mem[bi];
298 | //			s_mem[bi]  += t;
299 | //		}
300 | //		__syncthreads();
301 | //	}
302 | //	__syncthreads();
303 | //
304 | //	const index_t P_DATA	= gridDim.x*(blockDim.x<<1);
305 | //	tid 		<<= 1;
306 | //	index_t blk_idx	= blockIdx.x;
307 | //	
308 | //	index_t grd_idx_orig= 0;
309 | //	index_t grd_idx		= grd_idx_orig;
310 | //			//should add the offset but we eliminate 
311 | //			//since CONFLICT_FREE_OFFSET(0) = 0;
312 | //	
313 | //	while (tid<num_dat)
314 | //	{
315 | //		scan_out_d[tid]		= 	scan_out_d[tid]+ 
316 | //								blk_sum[blk_idx]+ 
317 | //								s_mem[grd_idx];	
318 | //		if(tid+1<num_dat)
319 | //			scan_out_d[tid+1]=scan_out_d[tid+1]+ 
320 | //								blk_sum[blk_idx]+ 
321 | //								s_mem[grd_idx];
322 | //
323 | //		tid		+= P_DATA;
324 | //
325 | //		blk_idx	+= gridDim.x;
326 | //		if(!((blk_idx -blockIdx.x)%(gridDim.x<<1)))
327 | //		{
328 | //			grd_idx_orig++;
329 | //			grd_idx	= grd_idx_orig + 
330 | //				CONFLICT_FREE_OFFSET(grd_idx_orig);
331 | //		}
332 | //	}
333 | //}
334 | //
335 | //
336 | ////mid problem
337 | //template<typename data_t, typename index_t>
338 | //__global__ void __post_scan(
339 | //								data_t 	*scan_out_d,
340 | //								data_t 	*blk_sum,
341 | //								index_t	num_dat,
342 | //								index_t num_blk,
343 | //								const index_t THD_NUM
344 | //							)
345 | //{
346 | //	extern __shared__ data_t s_mem[];
347 | //	
348 | //	const data_t tile_sz	= THD_NUM<<1;
349 | //	const index_t lane		= threadIdx.x<<1;
350 | //	index_t tid				= threadIdx.x+blockIdx.x*blockDim.x;
351 | //	index_t offset			= 1;
352 | //	const index_t padding	= CONFLICT_FREE_OFFSET(tile_sz - 1);
353 | //	
354 | //	const index_t off_a		= CONFLICT_FREE_OFFSET(lane);
355 | //	const index_t off_b		= CONFLICT_FREE_OFFSET(lane+1);
356 | //	/////////////////////////////////////////////////////////
357 | //	//GRID SCAN
358 | //	/////////////////////////////////////////////////////////
359 | //	if(lane < num_blk)
360 | //	{
361 | //		s_mem[lane+off_a]=blk_sum[lane];
362 | //	}else{
363 | //		s_mem[lane+off_a]=0;
364 | //	}
365 | //
366 | //	if(lane+1 < num_blk)
367 | //	{
368 | //		s_mem[lane+1+off_b]=blk_sum[lane+1];
369 | //	}else{
370 | //		s_mem[lane+1+off_b]=0;
371 | //	}
372 | //	__syncthreads();
373 | //	
374 | //	//up sweep
375 | //	for(index_t j = THD_NUM;j > 0;j>>=1)
376 | //	{   
377 | //		if(threadIdx.x < j)
378 | //		{
379 | //			index_t ai	= offset*lane +offset - 1;
380 | //			index_t bi	= ai + offset;
381 | //
382 | //			ai		   += CONFLICT_FREE_OFFSET(ai);
383 | //			bi		   += CONFLICT_FREE_OFFSET(bi);
384 | //			s_mem[bi] += s_mem[ai];
385 | //		}
386 | //		offset <<=1;
387 | //		__syncthreads();
388 | //	}   
389 | //	__syncthreads();
390 | //
391 | //	//write the block sum
392 | //	if(!threadIdx.x)
393 | //	{
394 | //		if(!blockIdx.x)
395 | //			in_q_sz_d	= s_mem[tile_sz-1+padding]; 
396 | //		
397 | //		s_mem[tile_sz+padding-1]	= 0;
398 | //	}
399 | //	__syncthreads();
400 | //
401 | //	//down sweep
402 | //	for(index_t j=1; j < tile_sz; j <<=1)
403 | //	{
404 | //		offset	>>=	1;
405 | //		if(threadIdx.x < j)
406 | //		{
407 | //			index_t ai	= lane*offset + offset - 1;
408 | //			index_t bi	= ai + offset;
409 | //			ai		   += CONFLICT_FREE_OFFSET(ai);
410 | //			bi		   += CONFLICT_FREE_OFFSET(bi);
411 | //			index_t t 	= s_mem[ai];
412 | //			s_mem[ai]	= s_mem[bi];
413 | //			s_mem[bi]  += t;
414 | //		}
415 | //		__syncthreads();
416 | //	}
417 | //	__syncthreads();
418 | //
419 | //	tid 				<<= 1;
420 | //	index_t blk_idx_orig= blockIdx.x;
421 | //	index_t blk_idx		= blk_idx_orig + 
422 | //				CONFLICT_FREE_OFFSET(blk_idx_orig);
423 | //	const index_t P_DATA= blockDim.x*(gridDim.x<<1);
424 | //	
425 | //	while (tid<num_dat)
426 | //	{
427 | //		scan_out_d[tid]		+=s_mem[blk_idx];	
428 | //		
429 | //		if(tid+1<num_dat)
430 | //			scan_out_d[tid+1]+= s_mem[blk_idx];
431 | //
432 | //		tid			+= P_DATA;
433 | //		blk_idx_orig+= gridDim.x;
434 | //		blk_idx		= blk_idx_orig +
435 | //					CONFLICT_FREE_OFFSET(blk_idx_orig);
436 | ////-------------------------------------------------
437 | ////KEEP THIS AS WELL
438 | ////this time we substract too many CONFLICT_FREE_OFF 
439 | ////		blk_idx	-= CONFLICT_FREE_OFFSET(blk_idx);
440 | ////		blk_idx	+= gridDim.x;
441 | ////		blk_idx	+= CONFLICT_FREE_OFFSET(blk_idx);
442 | ////
443 | ////-------------------------------------------------
444 | ////KEEP IT HERE:
445 | ////	if this way of counting blk_idx, we count 
446 | ////the previous padding twice!!!!
447 | ////		blk_idx	= gridDim.x;
448 | ////		blk_idx	+= CONFLICT_FREE_OFFSET(blk_idx);
449 | ////---------------------------------------------------
450 | //	}
451 | //}
452 | 
453 | 
454 | //USED FOR INSPECTION
455 | template <typename data_t, typename index_t>
456 | __global__ void __insp_pre_scan(	
457 | 								data_t 	*scan_in_d, 
458 | 								data_t 	*scan_out_d,
459 | 								data_t 	*blk_sum,
460 | 								index_t num_dat,
461 | 								const index_t THD_NUM
462 | 		   				   )
463 | {
464 | 	const data_t tile_sz	= THD_NUM<<1;
465 | 	const index_t lane		= threadIdx.x<<1;
466 | 	index_t tid				= threadIdx.x+blockIdx.x*blockDim.x;
467 | 	index_t offset			= 1;
468 | 	index_t	tid_strip		= tid<<1;
469 | 	const index_t padding	= CONFLICT_FREE_OFFSET(tile_sz - 1 );
470 | 	
471 | 	//conflict free
472 | 	const index_t off_a		= CONFLICT_FREE_OFFSET(lane);
473 | 	const index_t off_b		= CONFLICT_FREE_OFFSET(lane+1);
474 | 	const index_t GRNTY = blockDim.x*gridDim.x;
475 | 
476 | 
477 | 	//prefetching danger
478 | 	extern __shared__ data_t s_mem[];
479 | 	
480 | 	
481 | 	while(tid_strip < num_dat)
482 | 	{
483 | 		s_mem[lane+off_a] 		= scan_in_d[tid_strip];
484 | 		s_mem[lane + 1 + off_b]	= scan_in_d[tid_strip + 1];
485 | 		__syncthreads();
486 | 		
487 | 		//up sweep
488 | 		for(index_t j = THD_NUM;j > 0;j>>=1)
489 | 		{   
490 | 			if(threadIdx.x < j)
491 | 			{
492 | 				index_t ai	= offset*lane +offset - 1;
493 | 				index_t bi	= ai + offset;
494 | 				ai		   += CONFLICT_FREE_OFFSET(ai);
495 | 				bi		   += CONFLICT_FREE_OFFSET(bi);
496 | 				s_mem[bi]  += s_mem[ai];
497 | 			}
498 | 			offset <<=1;
499 | 			__syncthreads();
500 | 		}   
501 | 		__syncthreads();
502 | 
503 | 		//write the block sum
504 | 		if(threadIdx.x == 0)
505 | 		{
506 | 			blk_sum[tid/THD_NUM]	= 
507 | 							s_mem[tile_sz -1 + padding];
508 | 			s_mem[tile_sz-1 + padding]		= 0;
509 | 		}
510 | 		__syncthreads();
511 | 
512 | 		//down sweep
513 | 		for(index_t j=1; j < tile_sz; j <<=1)
514 | 		{
515 | 			offset	>>=	1;
516 | 			if(threadIdx.x < j)
517 | 			{
518 | 				index_t ai	= lane*offset + offset - 1;
519 | 				index_t bi	= ai + offset;
520 | 				ai		   += CONFLICT_FREE_OFFSET(ai);
521 | 				bi		   += CONFLICT_FREE_OFFSET(bi);
522 | 				index_t t 	= s_mem[ai];
523 | 				s_mem[ai]	= s_mem[bi];
524 | 				s_mem[bi]  += t;
525 | 			}
526 | 			__syncthreads();
527 | 		}
528 | 		__syncthreads();
529 | 		scan_out_d[tid_strip]	= s_mem[lane + off_a];
530 | 		scan_out_d[tid_strip+1]	= s_mem[lane + 1 + off_b];
531 | 		
532 | 		tid += GRNTY;
533 | 		tid_strip = tid<<1;
534 | 	}
535 | }
536 | 
537 | ////////////////////////////////////////////////////////
538 | //post_scan for inspection
539 | ////////////////////////////////////////////////////////
540 | template<typename data_t, typename index_t>
541 | __global__ void __insp_post_scan(
542 | 								index_t	*ex_q_sz_d,
543 | 								data_t 	*scan_out_d,
544 | 								data_t 	*blk_sum,
545 | 								index_t	num_dat,
546 | 								index_t num_blk,
547 | 								const index_t THD_NUM
548 | 							)
549 | {
550 | 	extern __shared__ data_t s_mem[];
551 | 	
552 | 	const data_t tile_sz	= THD_NUM<<1;
553 | 	const index_t lane		= threadIdx.x<<1;
554 | 	index_t tid				= threadIdx.x+blockIdx.x*blockDim.x;
555 | 	index_t offset			= 1;
556 | 	const index_t padding	= CONFLICT_FREE_OFFSET(tile_sz - 1);
557 | 	const index_t off_a		= CONFLICT_FREE_OFFSET(lane);
558 | 	const index_t off_b		= CONFLICT_FREE_OFFSET(lane+1);
559 | 	/////////////////////////////////////////////////////////
560 | 	//GRID SCAN
561 | 	/////////////////////////////////////////////////////////
562 | 	if(lane < num_blk)
563 | 	{
564 | 		s_mem[lane+off_a]=blk_sum[lane];
565 | 	}else{
566 | 		s_mem[lane+off_a]=0;
567 | 	}
568 | 
569 | 	if(lane+1 < num_blk)
570 | 	{
571 | 		s_mem[lane+1+off_b]=blk_sum[lane+1];
572 | 	}else{
573 | 		s_mem[lane+1+off_b]=0;
574 | 	}
575 | 	__syncthreads();
576 | 	
577 | 	//up sweep
578 | 	for(index_t j = THD_NUM;j > 0;j>>=1)
579 | 	{   
580 | 		if(threadIdx.x < j)
581 | 		{
582 | 			index_t ai	= offset*lane +offset - 1;
583 | 			index_t bi	= ai + offset;
584 | 
585 | 			ai		   += CONFLICT_FREE_OFFSET(ai);
586 | 			bi		   += CONFLICT_FREE_OFFSET(bi);
587 | 			s_mem[bi] += s_mem[ai];
588 | 		}
589 | 		offset <<=1;
590 | 		__syncthreads();
591 | 	}   
592 | 	__syncthreads();
593 | 
594 | 	//write the block sum
595 | 	if(!threadIdx.x)
596 | 	{
597 | 	
598 | 	if(!blockIdx.x)
599 | 	{
600 | 			ex_q_sz_d[0]=s_mem[tile_sz-1+padding];
601 | //		switch(q_t)
602 | //		{
603 | //			case SML_Q:
604 | //				ex_sml_sz_d	= s_mem[tile_sz-1+padding];
605 | //				break;
606 | //			case MID_Q:
607 | //				ex_mid_sz_d	= s_mem[tile_sz-1+padding];
608 | //				break;
609 | //			case LRG_Q:
610 | //				ex_lrg_sz_d	= s_mem[tile_sz-1+padding];
611 | //				break;
612 | //			default:
613 | //				break;
614 | //		}
615 | 	}
616 | 		s_mem[tile_sz-1+padding]= 0;
617 | 	}
618 | 	__syncthreads();
619 | 
620 | 	//down sweep
621 | 	for(index_t j=1; j < tile_sz; j <<=1)
622 | 	{
623 | 		offset	>>=	1;
624 | 		if(threadIdx.x < j)
625 | 		{
626 | 			index_t ai	= lane*offset + offset - 1;
627 | 			index_t bi	= ai + offset;
628 | 			ai		   += CONFLICT_FREE_OFFSET(ai);
629 | 			bi		   += CONFLICT_FREE_OFFSET(bi);
630 | 			index_t t 	= s_mem[ai];
631 | 			s_mem[ai]	= s_mem[bi];
632 | 			s_mem[bi]  += t;
633 | 		}
634 | 		__syncthreads();
635 | 	}
636 | 	__syncthreads();
637 | 
638 | 	tid 			<<= 1;
639 | 	index_t blk_idx_orig= blockIdx.x;
640 | 	index_t blk_idx		= blk_idx_orig + 
641 | 				CONFLICT_FREE_OFFSET(blk_idx_orig);
642 | 	const index_t P_DATA= gridDim.x*blockDim.x*2;
643 | 
644 | 	while (tid<num_dat)
645 | 	{
646 | 		scan_out_d[tid]		= 	scan_out_d[tid] + 
647 | 								s_mem[blk_idx];	
648 | 		scan_out_d[tid + 1]	= 	scan_out_d[tid + 1] + 
649 | 								s_mem[blk_idx];
650 | 		tid			+= P_DATA;
651 | 		blk_idx_orig+= gridDim.x;
652 | 		blk_idx		= blk_idx_orig +
653 | 					CONFLICT_FREE_OFFSET(blk_idx_orig);
654 | 	}
655 | }
656 | 
657 | //template<typename data_t, typename index_t>
658 | //__host__ void lrg_scan(
659 | //						data_t 			*scan_ind_d,
660 | //						data_t			*adj_card_d,
661 | //						//TODO requires scan_in_d to be 
662 | //						//		exact times of 
663 | //						//		THD_NUM*2
664 | //						data_t 			*scan_out_d,
665 | //						const index_t	BLK_NUM,
666 | //						const index_t 	THD_NUM,
667 | //						cudaStream_t 	&stream
668 | //		   		   		)
669 | //{
670 | //	const index_t num_dat = ex_q_sz;
671 | //	data_t *blk_sum;
672 | //	data_t *grd_sum;
673 | //	index_t num_grd;
674 | //	
675 | //	const size_t sz = sizeof(data_t);
676 | //	const index_t padding	= 
677 | //					CONFLICT_FREE_OFFSET((THD_NUM<<1) -1);
678 | //	index_t num_blk	= num_dat/(THD_NUM<<1);
679 | //	if (num_dat % (THD_NUM<<1))
680 | //		num_blk	++;
681 | //	
682 | //	num_grd = num_blk/(THD_NUM<<1);
683 | //	if(num_blk%(THD_NUM<<1))
684 | //		num_grd ++;
685 | //
686 | //	cudaMalloc((void **)&blk_sum,
687 | //			sizeof(data_t)*num_blk);
688 | //
689 | //	cudaMalloc((void **)&grd_sum,
690 | //			sizeof(data_t)*num_grd);
691 | //if(num_grd>(THD_NUM<<1))
692 | //	std::cout<<"Out of Range\n";
693 | //	__pre_scan<data_t, index_t>
694 | //	<<<BLK_NUM, THD_NUM, (padding+(THD_NUM<<1))*sz, stream>>>
695 | //	(
696 | //		scan_ind_d,
697 | //		adj_card_d,
698 | //		scan_out_d,
699 | //		blk_sum,
700 | //		num_dat,
701 | //		THD_NUM
702 | //	);
703 | //	cudaThreadSynchronize();
704 | //	__spine_scan<data_t, index_t>
705 | //	<<<BLK_NUM, THD_NUM, (padding+(THD_NUM<<1))*sz, stream>>>
706 | //	(
707 | //		blk_sum,
708 | //		grd_sum,
709 | //		num_blk,
710 | //		THD_NUM
711 | //	);
712 | //	cudaThreadSynchronize();
713 | //	__post_scan<data_t, index_t>
714 | //	<<<BLK_NUM, THD_NUM, (padding+(THD_NUM<<1))*sz, stream>>>
715 | //	(
716 | //		scan_out_d,
717 | //		blk_sum,
718 | //		grd_sum,
719 | //		num_dat,
720 | //		num_grd,
721 | //		THD_NUM
722 | //	);
723 | //	cudaThreadSynchronize();
724 | //}
725 | //
726 | //template<typename data_t, typename index_t>
727 | //__host__ void mid_scan(
728 | //						data_t 			*scan_ind_d,
729 | //						data_t			*adj_card_d,
730 | //						data_t 			*scan_out_d,
731 | //						const index_t	BLK_NUM,
732 | //						const index_t 	THD_NUM,
733 | //						cudaStream_t 	&stream
734 | //		   		   		)
735 | //{
736 | //	const index_t num_dat = ex_q_sz;
737 | //	data_t *blk_sum;
738 | //	const size_t sz = sizeof(data_t);
739 | //	const index_t padding	= 
740 | //					CONFLICT_FREE_OFFSET((THD_NUM<<1) -1);
741 | //	index_t num_blk	= num_dat/(THD_NUM<<1);
742 | //	if(num_blk % (THD_NUM<<1))
743 | //		num_blk ++;
744 | ////	std::cout<<"mid problem, num dat: "<<num_dat<<"\n";
745 | //	cudaMalloc((void **)&blk_sum,
746 | //			sizeof(data_t)*num_blk);
747 | //	
748 | //	__pre_scan<data_t, index_t>
749 | //	<<<BLK_NUM, THD_NUM, (padding+(THD_NUM<<1))*sz, stream>>>
750 | //	(
751 | //		scan_ind_d,
752 | //		adj_card_d,
753 | //		scan_out_d,
754 | //		blk_sum,
755 | //		num_dat,
756 | //		THD_NUM
757 | //	);	
758 | //	
759 | //	cudaThreadSynchronize();
760 | //	__post_scan<data_t, index_t>
761 | //	<<<BLK_NUM, THD_NUM, (padding+(THD_NUM<<1))*sz, stream>>>
762 | //	(
763 | //		scan_out_d,
764 | //		blk_sum,
765 | //		num_dat,
766 | //		num_blk,
767 | //		THD_NUM
768 | //	);	
769 | //}
770 | //
771 | //template<typename data_t, typename index_t>
772 | //__global__ void sml_scan(
773 | //						data_t 			*scan_ind_d,
774 | //						data_t 			*adj_card_d,
775 | //						data_t 			*scan_out_d,
776 | //						const index_t 	THD_NUM
777 | //		   		   		)
778 | //{
779 | //	
780 | //	const 	index_t	num_dat	= ex_q_sz_d;
781 | //	const data_t tile_sz	= THD_NUM<<1;
782 | //	const index_t lane		= threadIdx.x<<1;
783 | //	index_t offset			= 1;
784 | //	index_t idct_a, idct_b;
785 | //	const index_t padding	= 
786 | //				CONFLICT_FREE_OFFSET((THD_NUM<<1)-1);
787 | //	
788 | //	//conflict free
789 | //	const index_t off_a		= CONFLICT_FREE_OFFSET(lane);
790 | //	const index_t off_b		= CONFLICT_FREE_OFFSET(lane+1);
791 | //
792 | //	//prefetching danger
793 | //	extern __shared__ data_t s_mem[];
794 | //	
795 | //	if(lane < num_dat)
796 | //	{
797 | //		idct_a	= scan_ind_d[lane];
798 | //		s_mem[lane+off_a]	= adj_card_d[idct_a]; 
799 | //	}else{
800 | //		s_mem[lane+off_a]	= 0;	
801 | //	}
802 | //
803 | //	if((lane + 1) < num_dat)
804 | //	{
805 | //		idct_b	= scan_ind_d[lane + 1];
806 | //		s_mem[lane+1+off_b]	= adj_card_d[idct_b]; 
807 | //	}else{
808 | //		s_mem[lane+1+off_b]	= 0;	
809 | //	}
810 | //	__syncthreads();
811 | //	
812 | //	//up sweep
813 | //	for(index_t j = THD_NUM;j > 0;j>>=1)
814 | //	{   
815 | //		if(threadIdx.x < j)
816 | //		{
817 | //			index_t ai	= offset*lane +offset - 1;
818 | //			index_t bi	= ai + offset;
819 | //			ai		   += CONFLICT_FREE_OFFSET(ai);
820 | //			bi		   += CONFLICT_FREE_OFFSET(bi);
821 | //			s_mem[bi]  += s_mem[ai];
822 | //		}
823 | //		offset <<=1;
824 | //		__syncthreads();
825 | //	}   
826 | //	__syncthreads();
827 | //
828 | //	if(!threadIdx.x)
829 | //	{
830 | //		in_q_sz_d	= s_mem[(THD_NUM<<1)+ padding -1];
831 | //		s_mem[(THD_NUM<<1) + padding -1]	= 0;
832 | //	}
833 | //	__syncthreads();
834 | //	
835 | //	//down sweep
836 | //	for(index_t j=1; j < tile_sz; j <<=1)
837 | //	{
838 | //		offset	>>=	1;
839 | //		if(threadIdx.x < j)
840 | //		{
841 | //			index_t ai	= lane*offset + offset - 1;
842 | //			index_t bi	= ai + offset;
843 | //			ai		   += CONFLICT_FREE_OFFSET(ai);
844 | //			bi		   += CONFLICT_FREE_OFFSET(bi);
845 | //			index_t t 	= s_mem[ai];
846 | //			s_mem[ai]	= s_mem[bi];
847 | //			s_mem[bi]  += t;
848 | //		}
849 | //		__syncthreads();
850 | //	}
851 | //	__syncthreads();
852 | //	
853 | //	if(lane < num_dat)
854 | //		scan_out_d[lane]	= s_mem[lane + off_a];
855 | //	
856 | //	if((lane+1) < num_dat)
857 | //		scan_out_d[lane +1]	= s_mem[lane + 1 + off_b];
858 | //}
859 | 
860 | //-----------------------------------------
861 | //For inspection, 
862 | //exact threads number of data to scan
863 | //---------------------------------------
864 | template<typename data_t, typename index_t>
865 | __host__ void insp_scan(
866 | 						data_t 			*scan_in_d,
867 | 						//TODO requires scan_in_d to be 
868 | 						//		exact times of 
869 | 						//		THD_NUM*2
870 | 						data_t 			*scan_out_d,
871 | 						const index_t	num_dat,
872 | 						const index_t	BLK_NUM,
873 | 						const index_t	THD_NUM,
874 | 						index_t				*ex_q_sz_d,
875 | 						cudaStream_t 	&stream
876 | 		   		   		)
877 | {
878 | 	data_t *blk_sum;
879 | 	const size_t sz = sizeof(data_t);
880 | 	const index_t num_blk=THD_NUM<<1;
881 | 	const index_t padding=CONFLICT_FREE_OFFSET((THD_NUM<<1)-1);
882 | 	cudaMalloc((void **)&blk_sum,sz*num_blk);
883 | 	
884 | 	__insp_pre_scan<data_t, index_t>
885 | 	<<<BLK_NUM, THD_NUM, (padding+(THD_NUM<<1))*sz, stream>>>
886 | 	(
887 | 		scan_in_d,
888 | 		scan_out_d,
889 | 		blk_sum,
890 | 		num_dat,
891 | 		THD_NUM
892 | 	);	
893 | 	cudaThreadSynchronize();
894 | 	__insp_post_scan<data_t, index_t>
895 | 	<<<BLK_NUM, THD_NUM, (padding+(THD_NUM<<1))*sz, stream>>>
896 | 	(
897 | 		ex_q_sz_d,
898 | 		scan_out_d,
899 | 		blk_sum,
900 | 		num_dat,
901 | 		num_blk,
902 | 		THD_NUM
903 | 	);	
904 | }
905 | 


--------------------------------------------------------------------------------
/traversal.cu:
--------------------------------------------------------------------------------
  1 | /*
  2 |  * Copyright 2016 The George Washington University
  3 |  * Written by Hang Liu 
  4 |  * Directed by Prof. Howie Huang
  5 |  *
  6 |  * https://www.seas.gwu.edu/~howie/
  7 |  * Contact: iheartgraph@gmail.com
  8 |  *
  9 |  * 
 10 |  * Please cite the following paper:
 11 |  * 
 12 |  * Hang Liu, H. Howie Huang and Yang Hu. 2016. iBFS: Concurrent Breadth-First Search on GPUs. Proceedings of the 2016 International Conference on Management of Data. ACM. 
 13 |  *
 14 |  * This file is part of iBFS.
 15 |  *
 16 |  * iBFS is free software: you can redistribute it and/or modify
 17 |  * it under the terms of the GNU General Public License as published by
 18 |  * the Free Software Foundation, either version 3 of the License, or
 19 |  * (at your option) any later version.
 20 |  *
 21 |  * iBFS is distributed in the hope that it will be useful,
 22 |  * but WITHOUT ANY WARRANTY; without even the implied warranty of
 23 |  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the 
 24 |  * GNU General Public License for more details.
 25 |  *
 26 |  * You should have received a copy of the GNU General Public License
 27 |  * along with iBFS.  If not, see <http://www.gnu.org/licenses/>.
 28 |  */
 29 | 
 30 | #include "gpu_ibfs.cuh"
 31 | 
 32 | void gpu_ibfs::
 33 | debugger(index_t grp_beg)
 34 | {
 35 | 	comp_t *last = new comp_t[g->vert_count*joint_count];
 36 | 	comp_t *curr = new comp_t[g->vert_count*joint_count];
 37 | 	index_t *front_count=new index_t[concurr_count];
 38 | 	
 39 | 	H_ERR(cudaMemcpy(last, depth_comp_last, 
 40 | 		sizeof(comp_t)*g->vert_count*joint_count, cudaMemcpyDeviceToHost));
 41 | 	H_ERR(cudaMemcpy(curr, depth_comp_curr, 
 42 | 		sizeof(comp_t)*g->vert_count*joint_count, cudaMemcpyDeviceToHost));
 43 | 	
 44 | 	//ASSUMING one var can hold all bits
 45 | 	for(index_t i=0;i<concurr_count;++i)
 46 | 		front_count[i] = 0;
 47 | 	
 48 | 	for(index_t i=0;i<g->vert_count;++i)
 49 | 	{
 50 | 		if(last[i]!=curr[i])
 51 | 		{
 52 | 			comp_t curr_var=curr[i];
 53 | 			comp_t last_var=last[i];
 54 | 			for(index_t j=0;j<concurr_count;++j)
 55 | 			{
 56 | 				if((curr_var&(make_uint4(1)<<j))!=(last_var&(make_uint4(1)<<j)))
 57 | 					++front_count[j];
 58 | 			}
 59 | 		}
 60 | 	}
 61 | 	
 62 | 	char filename[256];
 63 | 	for(index_t i=0;i<concurr_count;++i)
 64 | 	{
 65 | 		sprintf(filename, "%d.log",g->src_list[i+grp_beg]);
 66 | 		FILE *file=fopen(filename, "a");
 67 | 		fprintf(file, "%d\n",front_count[i]);
 68 | 		fclose(file);
 69 | 	}
 70 | 
 71 | 	delete[] front_count;
 72 | 	delete[] last;
 73 | 	delete[] curr;
 74 | }
 75 | 
 76 | 
 77 | //traversal
 78 | void gpu_ibfs::
 79 | traversal(index_t	grp_beg){
 80 | 	cudaSetDevice(0);
 81 | 	ex_sml_sz[0]	= 0;
 82 | 	ex_mid_sz[0]	= 0;
 83 | 	ex_lrg_sz[0]	= concurr_count;
 84 | 	depth_t level=0;
 85 | 	
 86 | 	init_bfs(grp_beg);
 87 | 	cudaDeviceSynchronize();
 88 | 
 89 | 	#ifdef ENABLE_MONITORING
 90 | 	std::cout<<"Vert-ranger: "
 91 | 		<<g->src_list[grp_beg]<<"~"
 92 | 		<<g->src_list[grp_beg+concurr_count-1]<<"\n";
 93 | 	double tm_insp;
 94 | 	double tm_expd;
 95 | 	double tm_expand	= 0.0;
 96 | 	double tm_inspect	= 0.0;
 97 | 	debugger(grp_beg);
 98 | 	#endif
 99 | 	
100 | 	is_done[0]=false;	
101 | 	for(level=1;;level++){
102 | 		if(is_done[0])break;
103 | 		
104 | 		//----------------------	
105 | 		//Expand ex_qs and mark frontier in depth_d	
106 | 		//------------------------------------------------
107 | 		#ifdef ENABLE_MONITORING
108 | 		tm_expd=wtime();	
109 | 		#endif
110 | 		expander(level);
111 | 		cudaDeviceSynchronize();
112 | 
113 | 		#ifdef ENABLE_MONITORING
114 | 		tm_expd=wtime()-tm_expd;	
115 | 		debugger(grp_beg);
116 | 		tm_insp=wtime();	
117 | 		#endif
118 | 		//-----------------------
119 | 		//Generate ex_qs from depth_d
120 | 		//---------------------------------
121 | 		inspector(level);
122 | 		cudaDeviceSynchronize();
123 | 		#ifdef ENABLE_MONITORING
124 | 		tm_insp=wtime()-tm_insp;
125 | 		tm_expand		+= tm_expd;
126 | 		tm_inspect	+= tm_insp;
127 | 		std::cout<<"@level-"<<(int)level<<"-insp-expd: "
128 | 			<<tm_insp<<" "<<tm_expd<<"\n";
129 | 		#endif
130 | 	}
131 | 	
132 | 	#ifdef ENABLE_MONITORING
133 | 	debugger(grp_beg);
134 | 	std::cout<<"insp-expd: "<<tm_inspect<<" "<<tm_expand<<"\n";
135 | 	#endif
136 | }
137 | 


--------------------------------------------------------------------------------
/tuple_text_to_binary_csr/Makefile:
--------------------------------------------------------------------------------
 1 | exe = text_to_bin.bin
 2 | 
 3 | cc = "$(shell which g++)" 
 4 | flags = -O3
 5 | 
 6 | objs = $(patsubst %.cpp,%.o,$(wildcard ../../lib/*.cpp)) \
 7 | 			$(patsubst %.cpp,%.o,$(wildcard *.cpp))
 8 | 
 9 | deps = $(wildcard ../../lib/*.h) \
10 | 				$(wildcard *.h) \
11 | 				Makefile
12 | 
13 | %.o:%.cpp $(deps)
14 | 	$(cc) -c $< -o $@ $(flags)
15 | 
16 | $(exe):$(objs)
17 | 	$(cc) $(objs) -o $(exe) $(flags)
18 | 
19 | clean:
20 | 	rm -rf $(exe) $(objs)
21 | 


--------------------------------------------------------------------------------
/tuple_text_to_binary_csr/tuple_text_to_bin.cpp:
--------------------------------------------------------------------------------
  1 | /*
  2 |  * Copyright 2016 The George Washington University
  3 |  * Written by Hang Liu 
  4 |  * Directed by Prof. Howie Huang
  5 |  *
  6 |  * https://www.seas.gwu.edu/~howie/
  7 |  * Contact: iheartgraph@gmail.com
  8 |  *
  9 |  * 
 10 |  * Please cite the following paper:
 11 |  * 
 12 |  * Hang Liu, H. Howie Huang and Yang Hu. 2016. iBFS: Concurrent Breadth-First Search on GPUs. Proceedings of the 2016 International Conference on Management of Data. ACM. 
 13 |  *
 14 |  * This file is part of iBFS.
 15 |  *
 16 |  * iBFS is free software: you can redistribute it and/or modify
 17 |  * it under the terms of the GNU General Public License as published by
 18 |  * the Free Software Foundation, either version 3 of the License, or
 19 |  * (at your option) any later version.
 20 |  *
 21 |  * iBFS is distributed in the hope that it will be useful,
 22 |  * but WITHOUT ANY WARRANTY; without even the implied warranty of
 23 |  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the 
 24 |  * GNU General Public License for more details.
 25 |  *
 26 |  * You should have received a copy of the GNU General Public License
 27 |  * along with iBFS.  If not, see <http://www.gnu.org/licenses/>.
 28 |  */
 29 | 
 30 | #include <iostream>
 31 | #include <unistd.h>
 32 | #include <stdlib.h>
 33 | #include <stdio.h>
 34 | #include <errno.h>
 35 | #include <sys/mman.h>
 36 | #include <sys/types.h>
 37 | #include <sys/stat.h>
 38 | #include <fcntl.h>
 39 | #include <string.h>
 40 | #include <assert.h>
 41 | 
 42 | #define INFTY int(1<<30)
 43 | using namespace std;
 44 | 
 45 | typedef long int vertex_t;
 46 | typedef long int index_t;
 47 | 
 48 | inline off_t fsize(const char *filename) {
 49 |     struct stat st; 
 50 |     if (stat(filename, &st) == 0)
 51 |         return st.st_size;
 52 |     return -1; 
 53 | }
 54 | 
 55 | 		
 56 | int main(int argc, char** argv){
 57 | 	int fd,i;
 58 | 	char* ss_head;
 59 | 	char* ss;
 60 | 	
 61 | 	std::cout<<"Input: ./exe tuple_file(text) "
 62 | 			<<"reverse_the_edge(1 reverse, 0 not reverse) lines_to_skip\n";
 63 | 	if(argc<4){printf("Wrong input\n");exit(-1);}
 64 | 	
 65 | 	size_t file_size = fsize(argv[1]);
 66 | 	int is_reverse=(atol(argv[2])==1);	
 67 | 	long skip_head=atol(argv[3]);
 68 | 	
 69 | 
 70 | 	fd=open(argv[1],O_CREAT|O_RDWR,00666 );
 71 | 	if(fd == -1)
 72 | 	{
 73 | 		printf("%s open error\n", argv[1]);
 74 | 		perror("open");
 75 | 		exit(-1);
 76 | 	}
 77 | 
 78 | 	ss_head = (char*)mmap(NULL,file_size,PROT_READ|PROT_WRITE,MAP_PRIVATE,fd,0);
 79 | 	assert(ss_head != MAP_FAILED);
 80 | 	size_t head_offset=0;
 81 | 	int skip_count = 0;
 82 | 	while(true)
 83 | 	{
 84 | 		if(skip_count == skip_head) break;
 85 | 		if(head_offset == file_size &&
 86 | 				skip_count < skip_head)
 87 | 		{
 88 | 			std::cout<<"Eorr: skip more lines than the file has\n\n\n";
 89 | 			exit(-1);
 90 | 		}
 91 | 
 92 | 		head_offset++;
 93 | 		if(ss_head[head_offset]=='\n')
 94 | 		{
 95 | 			head_offset++;
 96 | 			skip_count++;
 97 | 			if(skip_count == skip_head) break;
 98 | 		}
 99 | 	}
100 | 
101 | 	ss = &ss_head[head_offset];
102 | 	file_size -= head_offset;
103 | 
104 | 	size_t curr=0;
105 | 	size_t next=0;
106 | 
107 | 	//step 1. vert_count,edge_count,
108 | 	size_t edge_count=0;
109 | 	size_t vert_count;
110 | 	vertex_t v_max = 0;
111 | 	vertex_t v_min = INFTY;//as infinity
112 | 	vertex_t a;
113 | 	while(next<file_size){
114 | 		char* sss=ss+curr;
115 | 		a = atol(sss);
116 | 		if(v_max<a){
117 | 			v_max = a;
118 | 		}
119 | 		if(v_min>a){
120 | 			v_min = a;
121 | 		}
122 | 
123 | 		while((ss[next]!=' ')&&(ss[next]!='\n')&&(ss[next]!='\t')){
124 | 			next++;
125 | 		}
126 | 		while((ss[next]==' ')||(ss[next]=='\n')||(ss[next]=='\t')){
127 | 			next++;
128 | 		}
129 | 		curr = next;
130 | 		
131 | 		//one vertex is counted once
132 | 		edge_count++;
133 | 	}
134 | 	
135 | 	const index_t line_count=edge_count>>1;
136 | 	if(!is_reverse) edge_count >>=1;
137 | 	
138 | 	vert_count = v_max - v_min + 1;
139 | 	assert(v_min<INFTY);
140 | 	cout<<"edge count: "<<edge_count<<endl;
141 | 	cout<<"max vertex id: "<<v_max<<endl;
142 | 	cout<<"min vertex id: "<<v_min<<endl;
143 | 
144 | 	cout<<"edge count: "<<edge_count<<endl;
145 | 	cout<<"vert count: "<<vert_count<<endl;
146 | 	
147 | 	//step 2. each file size
148 | 	char filename[256];
149 | 	sprintf(filename,"%s_csr.bin",argv[1]);
150 | 	int fd4 = open(filename,O_CREAT|O_RDWR,00666 );
151 | 	ftruncate(fd4, edge_count*sizeof(vertex_t));
152 | 	vertex_t* adj = (vertex_t*)mmap(NULL,edge_count*sizeof(vertex_t),PROT_READ|PROT_WRITE,MAP_SHARED,fd4,0);
153 | 	assert(adj != MAP_FAILED);
154 | 
155 | 	//added by Hang to generate a weight file
156 | 	sprintf(filename,"%s_weight.bin",argv[1]);
157 | 	int fd6 = open(filename,O_CREAT|O_RDWR,00666 );
158 | 	ftruncate(fd6, edge_count*sizeof(vertex_t));
159 | 	index_t* weight= (vertex_t*)mmap(NULL,edge_count*sizeof(vertex_t),PROT_READ|PROT_WRITE,MAP_SHARED,fd6,0);
160 | 	assert(weight != MAP_FAILED);
161 | 	//-End 
162 | 
163 | 	sprintf(filename,"%s_head.bin",argv[1]);
164 | 	int fd5 = open(filename,O_CREAT|O_RDWR,00666 );
165 | 	ftruncate(fd5, edge_count*sizeof(vertex_t));
166 | 	vertex_t* head = (vertex_t*)mmap(NULL,edge_count*sizeof(vertex_t),PROT_READ|PROT_WRITE,MAP_SHARED,fd5,0);
167 | 	assert(head != MAP_FAILED);
168 | 
169 | 	sprintf(filename,"%s_deg.bin",argv[1]);
170 | 	int fd2 = open(filename,O_CREAT|O_RDWR,00666 );
171 | 	ftruncate(fd2, vert_count*sizeof(index_t));
172 | 	index_t* degree = (index_t*)mmap(NULL,vert_count*sizeof(index_t),PROT_READ|PROT_WRITE,MAP_SHARED,fd2,0);
173 | 	assert(degree != MAP_FAILED);
174 | 	
175 | 	sprintf(filename,"%s_beg_pos.bin",argv[1]);
176 | 	int fd3 = open(filename,O_CREAT|O_RDWR,00666 );
177 | 	ftruncate(fd3, (vert_count+1)*sizeof(index_t));
178 | 	index_t* begin  = (index_t*)mmap(NULL,(vert_count+1)*sizeof(index_t),PROT_READ|PROT_WRITE,MAP_SHARED,fd3,0);
179 | 	assert(begin != MAP_FAILED);
180 | 	
181 | 	//step 3. write degree
182 | 	for(int i=0; i<vert_count;i++){
183 | 		degree[i]=0;
184 | 	}
185 | 
186 | 	vertex_t index, dest;
187 | 	size_t offset =0;
188 | 	curr=0;
189 | 	next=0;
190 | 
191 | 	printf("Getting degree...\n");
192 | 	while(offset<line_count){
193 | 		char* sss=ss+curr;
194 | 		index = atol(sss)-v_min;
195 | 		while((ss[next]!=' ')&&(ss[next]!='\n')&&(ss[next]!='\t')){
196 | 			next++;
197 | 		}
198 | 		while((ss[next]==' ')||(ss[next]=='\n')||(ss[next]=='\t')){
199 | 			next++;
200 | 		}
201 | 		curr = next;
202 | 
203 | 		char* sss1=ss+curr;
204 | 		dest=atol(sss1)-v_min;
205 | 
206 | 		while((ss[next]!=' ')&&(ss[next]!='\n')&&(ss[next]!='\t')){
207 | 			next++;
208 | 		}
209 | 		while((ss[next]==' ')||(ss[next]=='\n')||(ss[next]=='\t')){
210 | 			next++;
211 | 		}
212 | 		curr = next;
213 | 		degree[index]++;
214 | 		if(is_reverse) degree[dest]++;
215 | //		cout<<index<<" "<<degree[index]<<endl;
216 | 
217 | 		offset++;
218 | 	}
219 | //	exit(-1);
220 | 	begin[0]=0;
221 | 	begin[vert_count]=edge_count;
222 | 
223 | 	printf("Calculate beg_pos ...\n");
224 | 	for(size_t i=1; i<vert_count; i++){
225 | 		begin[i] = begin[i-1] + degree[i-1];
226 | //		cout<<begin[i]<<" "<<degree[i]<<endl;
227 | 		degree [i-1] = 0;
228 | 	}
229 | 	degree[vert_count-1] = 0;
230 | 	//step 4: write adjacent list 
231 | 	vertex_t v_id;
232 | 	offset =0;
233 | 	next = 0;
234 | 	curr = 0;
235 | 	
236 | 	printf("Constructing CSR...\n");
237 | 	while(offset<line_count){
238 | 		char* sss=ss+curr;
239 | 		index = atol(sss)-v_min;
240 | 		while((ss[next]!=' ')&&(ss[next]!='\n')&&(ss[next]!='\t')){
241 | 			next++;
242 | 		}
243 | 		while((ss[next]==' ')||(ss[next]=='\n')||(ss[next]=='\t')){
244 | 			next++;
245 | 		}
246 | 		curr = next;
247 | 
248 | 		char* sss1=ss+curr;
249 | 		v_id = atol(sss1)-v_min;
250 | 		adj[begin[index]+degree[index]] = v_id;
251 | 		if(is_reverse) adj[begin[v_id]+degree[v_id]] = index;
252 | 		
253 | 		//Added by Hang
254 | 		int rand_weight=(rand()%63+1);
255 | 		weight[begin[index]+degree[index]] = rand_weight;
256 | 		if(is_reverse)
257 | 			weight[begin[v_id]+degree[v_id]] = rand_weight;
258 | 		//-End
259 | 	
260 | 		head[begin[index]+degree[index]]= index;
261 | 		while((ss[next]!=' ')&&(ss[next]!='\n')&&(ss[next]!='\t')){
262 | 			next++;
263 | 		}
264 | 		while((ss[next]==' ')||(ss[next]=='\n')||(ss[next]=='\t')){
265 | 			next++;
266 | 		}
267 | 		curr = next;
268 | 		degree[index]++;
269 | 		if(is_reverse) degree[v_id]++;
270 | 
271 | 		offset++;
272 | 	}
273 | 	
274 | 	//step 5
275 | 	//print output as a test
276 | //	for(size_t i=0; i<vert_count; i++){
277 | 	for(size_t i=0; i<(vert_count<8?vert_count:8); i++){
278 | 		cout<<i<<" "<<begin[i+1]-begin[i]<<" ";
279 | 		for(index_t j=begin[i]; j<begin[i+1]; j++){
280 | 			cout<<adj[j]<<" ";
281 | 		}
282 | //		if(degree[i]>0){
283 | 			cout<<endl;
284 | //		}
285 | 	}
286 | 
287 | //	for(int i=0; i<edge_count; i++){
288 | //	for(int i=0; i<64; i++){
289 | //		cout<<head[i]<<"	"<<adj[i]<<endl;
290 | //	}
291 | 
292 | 	munmap( ss,sizeof(char)*file_size );
293 | 	
294 | 	//-Added by Hang
295 | 	munmap( weight,sizeof(vertex_t)*edge_count );
296 | 	//-End
297 | 
298 | 	munmap( adj,sizeof(vertex_t)*edge_count );
299 | 	munmap( head,sizeof(vertex_t)*edge_count );
300 | 	munmap( begin,sizeof(index_t)*vert_count+1 );
301 | 	munmap( degree,sizeof(index_t)*vert_count );
302 | 	close(fd2);
303 | 	close(fd3);
304 | 	close(fd4);
305 | 	close(fd5);
306 | 	
307 | 	//-Added by Hang
308 | 	close(fd6);
309 | 	//-End
310 | }
311 | 


--------------------------------------------------------------------------------
/util.h:
--------------------------------------------------------------------------------
  1 | /*
  2 |  * Copyright 2016 The George Washington University
  3 |  * Written by Hang Liu 
  4 |  * Directed by Prof. Howie Huang
  5 |  *
  6 |  * https://www.seas.gwu.edu/~howie/
  7 |  * Contact: iheartgraph@gmail.com
  8 |  *
  9 |  * 
 10 |  * Please cite the following paper:
 11 |  * 
 12 |  * Hang Liu, H. Howie Huang and Yang Hu. 2016. iBFS: Concurrent Breadth-First Search on GPUs. Proceedings of the 2016 International Conference on Management of Data. ACM. 
 13 |  *
 14 |  * This file is part of iBFS.
 15 |  *
 16 |  * iBFS is free software: you can redistribute it and/or modify
 17 |  * it under the terms of the GNU General Public License as published by
 18 |  * the Free Software Foundation, either version 3 of the License, or
 19 |  * (at your option) any later version.
 20 |  *
 21 |  * iBFS is distributed in the hope that it will be useful,
 22 |  * but WITHOUT ANY WARRANTY; without even the implied warranty of
 23 |  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the 
 24 |  * GNU General Public License for more details.
 25 |  *
 26 |  * You should have received a copy of the GNU General Public License
 27 |  * along with iBFS.  If not, see <http://www.gnu.org/licenses/>.
 28 |  */
 29 | 
 30 | #ifndef __UTIL_H__
 31 | #define	__UTIL_H__
 32 | #include <stdint.h>
 33 | #include <sys/stat.h>
 34 | #include <stdio.h>
 35 | 
 36 | inline off_t fsize(const char *filename) {
 37 | 	struct stat st; 
 38 | 	if (stat(filename, &st) == 0)
 39 | 		return st.st_size;
 40 | 	return -1; 
 41 | }
 42 | 
 43 | static void HandleError( cudaError_t err,
 44 |                          const char *file,
 45 |                          int line ) {
 46 |     if (err != cudaSuccess) {
 47 |         printf( "%s in %s at line %d\n", \
 48 |         cudaGetErrorString( err ),
 49 |                 file, line );
 50 |         exit( EXIT_FAILURE );
 51 |     }
 52 | }
 53 | #define H_ERR( err ) \
 54 |   (HandleError( err, __FILE__, __LINE__ ))
 55 | 
 56 | //////////////////////////////////////////////////
 57 | //SCALE*THDS_NUMS*sizeof(int) should be 
 58 | //limited by the size of the shared memory
 59 | /////////////////////////////////////////////////
 60 | //////////////////////////////////////////////////
 61 | #define	THDS_NUM			256	
 62 | #define	BLKS_NUM			256	
 63 | #define OMP_THDS			24
 64 | #define	V_NON_INC			-1
 65 | #define V_INI_HUB			-2
 66 | 
 67 | #define VALIDATE_TIMES		1
 68 | #define NUM_SRC		 		128	
 69 | #define INFTY					255	
 70 | #define	Q_CARD				3
 71 | enum ex_q_t
 72 | {
 73 |   SML_Q,
 74 |   MID_Q,
 75 |   LRG_Q,
 76 |   NONE
 77 | };
 78 | 
 79 | //--------------------------
 80 | typedef 	int 						index_t;
 81 | typedef		int							vertex_t;
 82 | typedef 	unsigned char 	depth_t;
 83 | typedef 	uint4	comp_t;
 84 | //--------------------------------
 85 | 
 86 | //operator overloading
 87 | inline __host__ __device__ uint4 operator-(uint4 a, uint4 b)
 88 | {
 89 |   return make_uint4(a.x - b.x, a.y - b.y, a.z - b.z,  a.w - b.w);
 90 | }
 91 | 
 92 | inline __host__ __device__ uint4 operator|=(uint4 a, uint4 b)
 93 | {
 94 |   return make_uint4(a.x |= b.x, a.y |= b.y, a.z |= b.z,  a.w |= b.w);
 95 | }
 96 | 
 97 | inline __host__ __device__ uint4 operator|(uint4 a, uint4 b)
 98 | {
 99 |   return make_uint4(a.x | b.x, a.y | b.y, a.z | b.z,  a.w | b.w);
100 | }
101 | 
102 | inline __host__ __device__ uint4 operator^(uint4 a, uint4 b)
103 | {
104 |   return make_uint4(a.x ^ b.x, a.y ^ b.y, a.z ^ b.z,  a.w ^ b.w);
105 | }
106 | 
107 | inline __host__ __device__ uint4 operator~(uint4 a)
108 | {
109 |   return make_uint4(~a.x, ~a.y, ~a.z, ~a.w);
110 | }
111 | 
112 | inline __host__ __device__ uint4 make_uint4(int a)
113 | {
114 |   return make_uint4(a, 0, 0, 0);
115 | }
116 | 
117 | inline __host__ __device__ uint4 operator<<(uint4 a, int b)
118 | {
119 |   if(b<32) a.x=(a.x<<b);
120 | 	else
121 | 	{
122 | 		if(b<64) a.y=a.x<<(b-32);
123 | 		else if(b<96) a.z=a.x<<(b-64);
124 | 		else a.w=a.x<<(b-96);
125 | 		
126 | 		a.x=0;
127 | 	}
128 | 
129 | 	return a;
130 | }
131 | 
132 | inline __host__ __device__ uint4 operator&(uint4 a, uint4 b)
133 | {
134 | 	return make_uint4(a.x & b.x, a.y & b.y, a.z & b.z, a.w & b.w);
135 | }
136 | 
137 | inline __host__ __device__ int operator!=(uint4 a, unsigned int b)
138 | {
139 | 	if(	a.x != b ||
140 | 			a.y != b ||
141 | 			a.z != b ||
142 | 			a.w != b ) return true;
143 | 	else return false;
144 | }
145 | 
146 | 
147 | inline __host__ __device__ int operator!=(uint4 a, uint4 b)
148 | {
149 | 	if(	a.x != b.x ||
150 | 			a.y != b.y ||
151 | 			a.z != b.z ||
152 | 			a.w != b.w ) return true;
153 | 	else return false;
154 | }
155 | 
156 | inline __host__ __device__ int operator==(uint4 a, unsigned int b)
157 | {
158 | 	if(a.x == b && 
159 | 		 a.y == b &&
160 | 		 a.z == b &&
161 | 		 a.w == b) return true;
162 | 	else return false;
163 | }
164 | 
165 | #endif
166 | 


--------------------------------------------------------------------------------
/validate.h:
--------------------------------------------------------------------------------
 1 | #include "comm.h"
 2 | template < 	typename index_t,
 3 | 			typename vertex_t,
 4 | 			typename depth_t> 
 5 | index_t validate(	
 6 | 		depth_t 	*depth_h, 
 7 | 		vertex_t	*adj_list,
 8 | 		index_t		*adj_card,
 9 | 		index_t		*beg_pos,
10 | 		index_t		vert_count)
11 | {
12 | 	
13 | 	//divide the adjacency vertices of this vertex into two parts
14 | 	//-parent part: all the vertices connected to the parent part 
15 | 	//				should have smaller or equal to base_depth
16 | 	//-children part: all vertices connected to the children part
17 | 	//				should have larger or equal to base_depth
18 | 	//depth difference should be no more than 1
19 | 	depth_t base_depth;
20 | 	char err_type;
21 | 	for(int i = 0; i< vert_count; i++){
22 | 		base_depth = depth_h[i];
23 | 		if(base_depth ==INFTY) continue;
24 | 
25 | 		if(!base_depth){//src vertex
26 | 			for(int j = beg_pos[i]; j< beg_pos[i]+adj_card[i]; j++)
27 | 				if((depth_h[adj_list[j]]!= 1) && (adj_list[j]!=i)){
28 | 					std::cout<<"Vert: "<<adj_list[j]
29 | 								<<" depth: "<<(int)depth_h[adj_list[j]]<<"\n";
30 | 					return -1;		
31 | 				}
32 | 		}else{
33 | 			
34 | 			//neighbor should consist both parents and children
35 | 			err_type = 0x00;//0x00000011
36 | 			//												||
37 | 			//												|+----child bit
38 | 			//												+-----parent bit
39 | 			//
40 | 			for(int j = beg_pos[i]; j<beg_pos[i]+adj_card[i]; j++){
41 | 				if(depth_h[adj_list[j]] == base_depth - 1)
42 | 					if(!(err_type & 0x02)) err_type |=0x02;
43 | 				else if(depth_h[adj_list[j]] == base_depth + 1)
44 | 					if(!(err_type & 0x01)) err_type |=0x01;
45 | 				else if(depth_h[adj_list[j]] != base_depth)
46 | 					return -2;
47 | 			}
48 | 			
49 | 			if(err_type == 0x01){
50 | 				std::cout<<"Vert: "<<i<<" err: "<<(int)err_type<<"\n";
51 | 				return -3;
52 | 			}	
53 | 		}
54 | 	}
55 | 	return 0;
56 | }
57 | 
58 | 
59 | template< 	typename vertex_t,
60 | 			typename index_t,
61 | 			typename depth_t>
62 | void report(	
63 | 				index_t &agg_tr_edges,
64 | 				index_t &agg_tr_v,
65 | 				depth_t	*depth_h,
66 | 				index_t	*adj_card,
67 | 				index_t	vert_count)
68 | {
69 | 	agg_tr_edges	= 0;
70 | 	agg_tr_v		= 0;
71 | 	for(index_t i = 0; i< vert_count; i++)
72 | 		if(depth_h[i] != INFTY){
73 | 			agg_tr_v ++;
74 | 			agg_tr_edges += adj_card[i];
75 | 		}
76 | }
77 | 


--------------------------------------------------------------------------------
/write_result.cuh:
--------------------------------------------------------------------------------
 1 | #include "graph.h"
 2 | #include <stdlib.h>
 3 | #include <time.h>
 4 | #include <sstream>
 5 | template<typename index_t, typename vertex_t, typename depth_t>
 6 | int graph<index_t, vertex_t, depth_t>::
 7 | write_result()
 8 | {
 9 | 	std::stringstream ss;
10 | 	srand(time(NULL));
11 | 	std::ofstream result_file;
12 | 	std::string file_str="bfs_result.";
13 | 	ss<<rand()%8959;
14 | 	file_str.append(ss.str());
15 | 	file_str.append(".log");
16 | 	result_file.open(file_str.c_str());
17 | 	
18 | 	for(depth_t i=0;;i++){
19 | 		index_t counter=0;
20 | 		result_file<<"Level "<<(int)i<<"(remaining  vertices) :";
21 | 		for(index_t j=0;j<vert_count;j++)
22 | 			if(depth[j]==i){
23 | 				counter==0 ? :result_file<<",";
24 | 				result_file<<j;
25 | 				counter++;
26 | 			}
27 | 		
28 | 		result_file<<"----------Total: "<<counter;
29 | 		if(!counter) 
30 | 			break;
31 | 		result_file<<"\n";
32 | 	}
33 | 
34 | 	result_file.close();
35 | 	return 0;
36 | }
37 | 


--------------------------------------------------------------------------------
/wtime.h:
--------------------------------------------------------------------------------
 1 | #ifndef __WTIME_H_
 2 | #define __WTIME_H_
 3 | 
 4 | #include <sys/time.h>
 5 | #include <stdlib.h>
 6 | 
 7 | double wtime()
 8 | {
 9 | 	double time[2];	
10 | 	struct timeval time1;
11 | 	gettimeofday(&time1, NULL);
12 | 
13 | 	time[0]=time1.tv_sec;
14 | 	time[1]=time1.tv_usec;
15 | 
16 | 	return time[0] + time[1]*1.0e-6;
17 | }
18 | 
19 | 
20 | #endif
21 | 


--------------------------------------------------------------------------------