From 895cfa5ff95288795e6009470cd37c944c3826b2 Mon Sep 17 00:00:00 2001 From: ArrogantGao <45324209+ArrogantGao@users.noreply.github.com> Date: Wed, 5 Jul 2023 21:24:08 +0800 Subject: [PATCH] Added wrapped C cuda code and runable examples (#1) initial --- .gitignore | 2 + Artifacts.toml | 6 + LICENSE | 695 +++++++++++++++++++- Project.toml | 6 + benchmark/benchmark_CUDA_mapreduce.jl | 20 + benchmark/benchmark_CuTropicalGemm.jl | 29 + benchmark/benchmark_CuTropicalGemm_C.sh | 6 + benchmark/benchmark_GemmKerenls_Tropical.jl | 83 +++ src/CuTropicalGEMM.jl | 7 +- src/TropicalGemmFP32.cu | 627 ++++++++++++++++++ src/TropicalGemm_Cuda_wrapper.jl | 15 + test/runtests.jl | 3 +- test/test_FP32_matmul.jl | 32 + 13 files changed, 1508 insertions(+), 23 deletions(-) create mode 100644 Artifacts.toml create mode 100644 benchmark/benchmark_CUDA_mapreduce.jl create mode 100644 benchmark/benchmark_CuTropicalGemm.jl create mode 100644 benchmark/benchmark_CuTropicalGemm_C.sh create mode 100644 benchmark/benchmark_GemmKerenls_Tropical.jl create mode 100644 src/TropicalGemmFP32.cu create mode 100644 src/TropicalGemm_Cuda_wrapper.jl create mode 100644 test/test_FP32_matmul.jl diff --git a/.gitignore b/.gitignore index 0f84bed..cc4cf6e 100644 --- a/.gitignore +++ b/.gitignore @@ -2,3 +2,5 @@ *.jl.cov *.jl.mem /Manifest.toml +lib +.vscode \ No newline at end of file diff --git a/Artifacts.toml b/Artifacts.toml new file mode 100644 index 0000000..473b517 --- /dev/null +++ b/Artifacts.toml @@ -0,0 +1,6 @@ +[CUDA_lib] +git-tree-sha1 = "2918fba865582556e219191a7f393c47c2e822e0" + + [[CUDA_lib.download]] + sha256 = "751bf9d1f2921d4176ffb8ed1ddbd59bb60d6a517e6784bb71d61b62357c0007" + url = "https://gist.github.com/ArrogantGao/c38791f143d36d4b2481ac7e4aa4ecce/raw/2918fba865582556e219191a7f393c47c2e822e0.tar.gz" diff --git a/LICENSE b/LICENSE index 4b712ea..e72bfdd 100644 --- a/LICENSE +++ b/LICENSE @@ -1,21 +1,674 @@ -MIT License - -Copyright (c) 2023 Xuanzhao Gao and contributors - -Permission is hereby granted, free of charge, to any person obtaining a copy -of this software and associated documentation files (the "Software"), to deal -in the Software without restriction, including without limitation the rights -to use, copy, modify, merge, publish, distribute, sublicense, and/or sell -copies of the Software, and to permit persons to whom the Software is -furnished to do so, subject to the following conditions: - -The above copyright notice and this permission notice shall be included in all -copies or substantial portions of the Software. - -THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR -IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, -FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE -AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER -LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, -OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE -SOFTWARE. + GNU GENERAL PUBLIC LICENSE + Version 3, 29 June 2007 + + Copyright (C) 2007 Free Software Foundation, Inc. + Everyone is permitted to copy and distribute verbatim copies + of this license document, but changing it is not allowed. + + Preamble + + The GNU General Public License is a free, copyleft license for +software and other kinds of works. + + The licenses for most software and other practical works are designed +to take away your freedom to share and change the works. By contrast, +the GNU General Public License is intended to guarantee your freedom to +share and change all versions of a program--to make sure it remains free +software for all its users. We, the Free Software Foundation, use the +GNU General Public License for most of our software; it applies also to +any other work released this way by its authors. You can apply it to +your programs, too. + + When we speak of free software, we are referring to freedom, not +price. Our General Public Licenses are designed to make sure that you +have the freedom to distribute copies of free software (and charge for +them if you wish), that you receive source code or can get it if you +want it, that you can change the software or use pieces of it in new +free programs, and that you know you can do these things. + + To protect your rights, we need to prevent others from denying you +these rights or asking you to surrender the rights. Therefore, you have +certain responsibilities if you distribute copies of the software, or if +you modify it: responsibilities to respect the freedom of others. + + For example, if you distribute copies of such a program, whether +gratis or for a fee, you must pass on to the recipients the same +freedoms that you received. You must make sure that they, too, receive +or can get the source code. And you must show them these terms so they +know their rights. + + Developers that use the GNU GPL protect your rights with two steps: +(1) assert copyright on the software, and (2) offer you this License +giving you legal permission to copy, distribute and/or modify it. + + For the developers' and authors' protection, the GPL clearly explains +that there is no warranty for this free software. For both users' and +authors' sake, the GPL requires that modified versions be marked as +changed, so that their problems will not be attributed erroneously to +authors of previous versions. + + Some devices are designed to deny users access to install or run +modified versions of the software inside them, although the manufacturer +can do so. This is fundamentally incompatible with the aim of +protecting users' freedom to change the software. The systematic +pattern of such abuse occurs in the area of products for individuals to +use, which is precisely where it is most unacceptable. Therefore, we +have designed this version of the GPL to prohibit the practice for those +products. If such problems arise substantially in other domains, we +stand ready to extend this provision to those domains in future versions +of the GPL, as needed to protect the freedom of users. + + Finally, every program is threatened constantly by software patents. +States should not allow patents to restrict development and use of +software on general-purpose computers, but in those that do, we wish to +avoid the special danger that patents applied to a free program could +make it effectively proprietary. To prevent this, the GPL assures that +patents cannot be used to render the program non-free. + + The precise terms and conditions for copying, distribution and +modification follow. + + TERMS AND CONDITIONS + + 0. Definitions. + + "This License" refers to version 3 of the GNU General Public License. + + "Copyright" also means copyright-like laws that apply to other kinds of +works, such as semiconductor masks. + + "The Program" refers to any copyrightable work licensed under this +License. Each licensee is addressed as "you". "Licensees" and +"recipients" may be individuals or organizations. + + To "modify" a work means to copy from or adapt all or part of the work +in a fashion requiring copyright permission, other than the making of an +exact copy. The resulting work is called a "modified version" of the +earlier work or a work "based on" the earlier work. + + A "covered work" means either the unmodified Program or a work based +on the Program. + + To "propagate" a work means to do anything with it that, without +permission, would make you directly or secondarily liable for +infringement under applicable copyright law, except executing it on a +computer or modifying a private copy. Propagation includes copying, +distribution (with or without modification), making available to the +public, and in some countries other activities as well. + + To "convey" a work means any kind of propagation that enables other +parties to make or receive copies. Mere interaction with a user through +a computer network, with no transfer of a copy, is not conveying. + + An interactive user interface displays "Appropriate Legal Notices" +to the extent that it includes a convenient and prominently visible +feature that (1) displays an appropriate copyright notice, and (2) +tells the user that there is no warranty for the work (except to the +extent that warranties are provided), that licensees may convey the +work under this License, and how to view a copy of this License. If +the interface presents a list of user commands or options, such as a +menu, a prominent item in the list meets this criterion. + + 1. Source Code. + + The "source code" for a work means the preferred form of the work +for making modifications to it. "Object code" means any non-source +form of a work. + + A "Standard Interface" means an interface that either is an official +standard defined by a recognized standards body, or, in the case of +interfaces specified for a particular programming language, one that +is widely used among developers working in that language. + + The "System Libraries" of an executable work include anything, other +than the work as a whole, that (a) is included in the normal form of +packaging a Major Component, but which is not part of that Major +Component, and (b) serves only to enable use of the work with that +Major Component, or to implement a Standard Interface for which an +implementation is available to the public in source code form. A +"Major Component", in this context, means a major essential component +(kernel, window system, and so on) of the specific operating system +(if any) on which the executable work runs, or a compiler used to +produce the work, or an object code interpreter used to run it. + + The "Corresponding Source" for a work in object code form means all +the source code needed to generate, install, and (for an executable +work) run the object code and to modify the work, including scripts to +control those activities. However, it does not include the work's +System Libraries, or general-purpose tools or generally available free +programs which are used unmodified in performing those activities but +which are not part of the work. For example, Corresponding Source +includes interface definition files associated with source files for +the work, and the source code for shared libraries and dynamically +linked subprograms that the work is specifically designed to require, +such as by intimate data communication or control flow between those +subprograms and other parts of the work. + + The Corresponding Source need not include anything that users +can regenerate automatically from other parts of the Corresponding +Source. + + The Corresponding Source for a work in source code form is that +same work. + + 2. Basic Permissions. + + All rights granted under this License are granted for the term of +copyright on the Program, and are irrevocable provided the stated +conditions are met. This License explicitly affirms your unlimited +permission to run the unmodified Program. The output from running a +covered work is covered by this License only if the output, given its +content, constitutes a covered work. This License acknowledges your +rights of fair use or other equivalent, as provided by copyright law. + + You may make, run and propagate covered works that you do not +convey, without conditions so long as your license otherwise remains +in force. You may convey covered works to others for the sole purpose +of having them make modifications exclusively for you, or provide you +with facilities for running those works, provided that you comply with +the terms of this License in conveying all material for which you do +not control copyright. Those thus making or running the covered works +for you must do so exclusively on your behalf, under your direction +and control, on terms that prohibit them from making any copies of +your copyrighted material outside their relationship with you. + + Conveying under any other circumstances is permitted solely under +the conditions stated below. Sublicensing is not allowed; section 10 +makes it unnecessary. + + 3. Protecting Users' Legal Rights From Anti-Circumvention Law. + + No covered work shall be deemed part of an effective technological +measure under any applicable law fulfilling obligations under article +11 of the WIPO copyright treaty adopted on 20 December 1996, or +similar laws prohibiting or restricting circumvention of such +measures. + + When you convey a covered work, you waive any legal power to forbid +circumvention of technological measures to the extent such circumvention +is effected by exercising rights under this License with respect to +the covered work, and you disclaim any intention to limit operation or +modification of the work as a means of enforcing, against the work's +users, your or third parties' legal rights to forbid circumvention of +technological measures. + + 4. Conveying Verbatim Copies. + + You may convey verbatim copies of the Program's source code as you +receive it, in any medium, provided that you conspicuously and +appropriately publish on each copy an appropriate copyright notice; +keep intact all notices stating that this License and any +non-permissive terms added in accord with section 7 apply to the code; +keep intact all notices of the absence of any warranty; and give all +recipients a copy of this License along with the Program. + + You may charge any price or no price for each copy that you convey, +and you may offer support or warranty protection for a fee. + + 5. Conveying Modified Source Versions. + + You may convey a work based on the Program, or the modifications to +produce it from the Program, in the form of source code under the +terms of section 4, provided that you also meet all of these conditions: + + a) The work must carry prominent notices stating that you modified + it, and giving a relevant date. + + b) The work must carry prominent notices stating that it is + released under this License and any conditions added under section + 7. This requirement modifies the requirement in section 4 to + "keep intact all notices". + + c) You must license the entire work, as a whole, under this + License to anyone who comes into possession of a copy. This + License will therefore apply, along with any applicable section 7 + additional terms, to the whole of the work, and all its parts, + regardless of how they are packaged. This License gives no + permission to license the work in any other way, but it does not + invalidate such permission if you have separately received it. + + d) If the work has interactive user interfaces, each must display + Appropriate Legal Notices; however, if the Program has interactive + interfaces that do not display Appropriate Legal Notices, your + work need not make them do so. + + A compilation of a covered work with other separate and independent +works, which are not by their nature extensions of the covered work, +and which are not combined with it such as to form a larger program, +in or on a volume of a storage or distribution medium, is called an +"aggregate" if the compilation and its resulting copyright are not +used to limit the access or legal rights of the compilation's users +beyond what the individual works permit. Inclusion of a covered work +in an aggregate does not cause this License to apply to the other +parts of the aggregate. + + 6. Conveying Non-Source Forms. + + You may convey a covered work in object code form under the terms +of sections 4 and 5, provided that you also convey the +machine-readable Corresponding Source under the terms of this License, +in one of these ways: + + a) Convey the object code in, or embodied in, a physical product + (including a physical distribution medium), accompanied by the + Corresponding Source fixed on a durable physical medium + customarily used for software interchange. + + b) Convey the object code in, or embodied in, a physical product + (including a physical distribution medium), accompanied by a + written offer, valid for at least three years and valid for as + long as you offer spare parts or customer support for that product + model, to give anyone who possesses the object code either (1) a + copy of the Corresponding Source for all the software in the + product that is covered by this License, on a durable physical + medium customarily used for software interchange, for a price no + more than your reasonable cost of physically performing this + conveying of source, or (2) access to copy the + Corresponding Source from a network server at no charge. + + c) Convey individual copies of the object code with a copy of the + written offer to provide the Corresponding Source. This + alternative is allowed only occasionally and noncommercially, and + only if you received the object code with such an offer, in accord + with subsection 6b. + + d) Convey the object code by offering access from a designated + place (gratis or for a charge), and offer equivalent access to the + Corresponding Source in the same way through the same place at no + further charge. You need not require recipients to copy the + Corresponding Source along with the object code. If the place to + copy the object code is a network server, the Corresponding Source + may be on a different server (operated by you or a third party) + that supports equivalent copying facilities, provided you maintain + clear directions next to the object code saying where to find the + Corresponding Source. Regardless of what server hosts the + Corresponding Source, you remain obligated to ensure that it is + available for as long as needed to satisfy these requirements. + + e) Convey the object code using peer-to-peer transmission, provided + you inform other peers where the object code and Corresponding + Source of the work are being offered to the general public at no + charge under subsection 6d. + + A separable portion of the object code, whose source code is excluded +from the Corresponding Source as a System Library, need not be +included in conveying the object code work. + + A "User Product" is either (1) a "consumer product", which means any +tangible personal property which is normally used for personal, family, +or household purposes, or (2) anything designed or sold for incorporation +into a dwelling. In determining whether a product is a consumer product, +doubtful cases shall be resolved in favor of coverage. For a particular +product received by a particular user, "normally used" refers to a +typical or common use of that class of product, regardless of the status +of the particular user or of the way in which the particular user +actually uses, or expects or is expected to use, the product. A product +is a consumer product regardless of whether the product has substantial +commercial, industrial or non-consumer uses, unless such uses represent +the only significant mode of use of the product. + + "Installation Information" for a User Product means any methods, +procedures, authorization keys, or other information required to install +and execute modified versions of a covered work in that User Product from +a modified version of its Corresponding Source. The information must +suffice to ensure that the continued functioning of the modified object +code is in no case prevented or interfered with solely because +modification has been made. + + If you convey an object code work under this section in, or with, or +specifically for use in, a User Product, and the conveying occurs as +part of a transaction in which the right of possession and use of the +User Product is transferred to the recipient in perpetuity or for a +fixed term (regardless of how the transaction is characterized), the +Corresponding Source conveyed under this section must be accompanied +by the Installation Information. But this requirement does not apply +if neither you nor any third party retains the ability to install +modified object code on the User Product (for example, the work has +been installed in ROM). + + The requirement to provide Installation Information does not include a +requirement to continue to provide support service, warranty, or updates +for a work that has been modified or installed by the recipient, or for +the User Product in which it has been modified or installed. Access to a +network may be denied when the modification itself materially and +adversely affects the operation of the network or violates the rules and +protocols for communication across the network. + + Corresponding Source conveyed, and Installation Information provided, +in accord with this section must be in a format that is publicly +documented (and with an implementation available to the public in +source code form), and must require no special password or key for +unpacking, reading or copying. + + 7. Additional Terms. + + "Additional permissions" are terms that supplement the terms of this +License by making exceptions from one or more of its conditions. +Additional permissions that are applicable to the entire Program shall +be treated as though they were included in this License, to the extent +that they are valid under applicable law. If additional permissions +apply only to part of the Program, that part may be used separately +under those permissions, but the entire Program remains governed by +this License without regard to the additional permissions. + + When you convey a copy of a covered work, you may at your option +remove any additional permissions from that copy, or from any part of +it. (Additional permissions may be written to require their own +removal in certain cases when you modify the work.) You may place +additional permissions on material, added by you to a covered work, +for which you have or can give appropriate copyright permission. + + Notwithstanding any other provision of this License, for material you +add to a covered work, you may (if authorized by the copyright holders of +that material) supplement the terms of this License with terms: + + a) Disclaiming warranty or limiting liability differently from the + terms of sections 15 and 16 of this License; or + + b) Requiring preservation of specified reasonable legal notices or + author attributions in that material or in the Appropriate Legal + Notices displayed by works containing it; or + + c) Prohibiting misrepresentation of the origin of that material, or + requiring that modified versions of such material be marked in + reasonable ways as different from the original version; or + + d) Limiting the use for publicity purposes of names of licensors or + authors of the material; or + + e) Declining to grant rights under trademark law for use of some + trade names, trademarks, or service marks; or + + f) Requiring indemnification of licensors and authors of that + material by anyone who conveys the material (or modified versions of + it) with contractual assumptions of liability to the recipient, for + any liability that these contractual assumptions directly impose on + those licensors and authors. + + All other non-permissive additional terms are considered "further +restrictions" within the meaning of section 10. If the Program as you +received it, or any part of it, contains a notice stating that it is +governed by this License along with a term that is a further +restriction, you may remove that term. If a license document contains +a further restriction but permits relicensing or conveying under this +License, you may add to a covered work material governed by the terms +of that license document, provided that the further restriction does +not survive such relicensing or conveying. + + If you add terms to a covered work in accord with this section, you +must place, in the relevant source files, a statement of the +additional terms that apply to those files, or a notice indicating +where to find the applicable terms. + + Additional terms, permissive or non-permissive, may be stated in the +form of a separately written license, or stated as exceptions; +the above requirements apply either way. + + 8. Termination. + + You may not propagate or modify a covered work except as expressly +provided under this License. Any attempt otherwise to propagate or +modify it is void, and will automatically terminate your rights under +this License (including any patent licenses granted under the third +paragraph of section 11). + + However, if you cease all violation of this License, then your +license from a particular copyright holder is reinstated (a) +provisionally, unless and until the copyright holder explicitly and +finally terminates your license, and (b) permanently, if the copyright +holder fails to notify you of the violation by some reasonable means +prior to 60 days after the cessation. + + Moreover, your license from a particular copyright holder is +reinstated permanently if the copyright holder notifies you of the +violation by some reasonable means, this is the first time you have +received notice of violation of this License (for any work) from that +copyright holder, and you cure the violation prior to 30 days after +your receipt of the notice. + + Termination of your rights under this section does not terminate the +licenses of parties who have received copies or rights from you under +this License. If your rights have been terminated and not permanently +reinstated, you do not qualify to receive new licenses for the same +material under section 10. + + 9. Acceptance Not Required for Having Copies. + + You are not required to accept this License in order to receive or +run a copy of the Program. Ancillary propagation of a covered work +occurring solely as a consequence of using peer-to-peer transmission +to receive a copy likewise does not require acceptance. However, +nothing other than this License grants you permission to propagate or +modify any covered work. These actions infringe copyright if you do +not accept this License. Therefore, by modifying or propagating a +covered work, you indicate your acceptance of this License to do so. + + 10. Automatic Licensing of Downstream Recipients. + + Each time you convey a covered work, the recipient automatically +receives a license from the original licensors, to run, modify and +propagate that work, subject to this License. You are not responsible +for enforcing compliance by third parties with this License. + + An "entity transaction" is a transaction transferring control of an +organization, or substantially all assets of one, or subdividing an +organization, or merging organizations. If propagation of a covered +work results from an entity transaction, each party to that +transaction who receives a copy of the work also receives whatever +licenses to the work the party's predecessor in interest had or could +give under the previous paragraph, plus a right to possession of the +Corresponding Source of the work from the predecessor in interest, if +the predecessor has it or can get it with reasonable efforts. + + You may not impose any further restrictions on the exercise of the +rights granted or affirmed under this License. For example, you may +not impose a license fee, royalty, or other charge for exercise of +rights granted under this License, and you may not initiate litigation +(including a cross-claim or counterclaim in a lawsuit) alleging that +any patent claim is infringed by making, using, selling, offering for +sale, or importing the Program or any portion of it. + + 11. Patents. + + A "contributor" is a copyright holder who authorizes use under this +License of the Program or a work on which the Program is based. The +work thus licensed is called the contributor's "contributor version". + + A contributor's "essential patent claims" are all patent claims +owned or controlled by the contributor, whether already acquired or +hereafter acquired, that would be infringed by some manner, permitted +by this License, of making, using, or selling its contributor version, +but do not include claims that would be infringed only as a +consequence of further modification of the contributor version. For +purposes of this definition, "control" includes the right to grant +patent sublicenses in a manner consistent with the requirements of +this License. + + Each contributor grants you a non-exclusive, worldwide, royalty-free +patent license under the contributor's essential patent claims, to +make, use, sell, offer for sale, import and otherwise run, modify and +propagate the contents of its contributor version. + + In the following three paragraphs, a "patent license" is any express +agreement or commitment, however denominated, not to enforce a patent +(such as an express permission to practice a patent or covenant not to +sue for patent infringement). To "grant" such a patent license to a +party means to make such an agreement or commitment not to enforce a +patent against the party. + + If you convey a covered work, knowingly relying on a patent license, +and the Corresponding Source of the work is not available for anyone +to copy, free of charge and under the terms of this License, through a +publicly available network server or other readily accessible means, +then you must either (1) cause the Corresponding Source to be so +available, or (2) arrange to deprive yourself of the benefit of the +patent license for this particular work, or (3) arrange, in a manner +consistent with the requirements of this License, to extend the patent +license to downstream recipients. "Knowingly relying" means you have +actual knowledge that, but for the patent license, your conveying the +covered work in a country, or your recipient's use of the covered work +in a country, would infringe one or more identifiable patents in that +country that you have reason to believe are valid. + + If, pursuant to or in connection with a single transaction or +arrangement, you convey, or propagate by procuring conveyance of, a +covered work, and grant a patent license to some of the parties +receiving the covered work authorizing them to use, propagate, modify +or convey a specific copy of the covered work, then the patent license +you grant is automatically extended to all recipients of the covered +work and works based on it. + + A patent license is "discriminatory" if it does not include within +the scope of its coverage, prohibits the exercise of, or is +conditioned on the non-exercise of one or more of the rights that are +specifically granted under this License. You may not convey a covered +work if you are a party to an arrangement with a third party that is +in the business of distributing software, under which you make payment +to the third party based on the extent of your activity of conveying +the work, and under which the third party grants, to any of the +parties who would receive the covered work from you, a discriminatory +patent license (a) in connection with copies of the covered work +conveyed by you (or copies made from those copies), or (b) primarily +for and in connection with specific products or compilations that +contain the covered work, unless you entered into that arrangement, +or that patent license was granted, prior to 28 March 2007. + + Nothing in this License shall be construed as excluding or limiting +any implied license or other defenses to infringement that may +otherwise be available to you under applicable patent law. + + 12. No Surrender of Others' Freedom. + + If conditions are imposed on you (whether by court order, agreement or +otherwise) that contradict the conditions of this License, they do not +excuse you from the conditions of this License. If you cannot convey a +covered work so as to satisfy simultaneously your obligations under this +License and any other pertinent obligations, then as a consequence you may +not convey it at all. For example, if you agree to terms that obligate you +to collect a royalty for further conveying from those to whom you convey +the Program, the only way you could satisfy both those terms and this +License would be to refrain entirely from conveying the Program. + + 13. Use with the GNU Affero General Public License. + + Notwithstanding any other provision of this License, you have +permission to link or combine any covered work with a work licensed +under version 3 of the GNU Affero General Public License into a single +combined work, and to convey the resulting work. The terms of this +License will continue to apply to the part which is the covered work, +but the special requirements of the GNU Affero General Public License, +section 13, concerning interaction through a network will apply to the +combination as such. + + 14. Revised Versions of this License. + + The Free Software Foundation may publish revised and/or new versions of +the GNU General Public License from time to time. Such new versions will +be similar in spirit to the present version, but may differ in detail to +address new problems or concerns. + + Each version is given a distinguishing version number. If the +Program specifies that a certain numbered version of the GNU General +Public License "or any later version" applies to it, you have the +option of following the terms and conditions either of that numbered +version or of any later version published by the Free Software +Foundation. If the Program does not specify a version number of the +GNU General Public License, you may choose any version ever published +by the Free Software Foundation. + + If the Program specifies that a proxy can decide which future +versions of the GNU General Public License can be used, that proxy's +public statement of acceptance of a version permanently authorizes you +to choose that version for the Program. + + Later license versions may give you additional or different +permissions. However, no additional obligations are imposed on any +author or copyright holder as a result of your choosing to follow a +later version. + + 15. Disclaimer of Warranty. + + THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY +APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT +HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY +OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, +THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR +PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM +IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF +ALL NECESSARY SERVICING, REPAIR OR CORRECTION. + + 16. Limitation of Liability. + + IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING +WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS +THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY +GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE +USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF +DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD +PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), +EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF +SUCH DAMAGES. + + 17. Interpretation of Sections 15 and 16. + + If the disclaimer of warranty and limitation of liability provided +above cannot be given local legal effect according to their terms, +reviewing courts shall apply local law that most closely approximates +an absolute waiver of all civil liability in connection with the +Program, unless a warranty or assumption of liability accompanies a +copy of the Program in return for a fee. + + END OF TERMS AND CONDITIONS + + How to Apply These Terms to Your New Programs + + If you develop a new program, and you want it to be of the greatest +possible use to the public, the best way to achieve this is to make it +free software which everyone can redistribute and change under these terms. + + To do so, attach the following notices to the program. It is safest +to attach them to the start of each source file to most effectively +state the exclusion of warranty; and each file should have at least +the "copyright" line and a pointer to where the full notice is found. + + + Copyright (C) + + This program is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . + +Also add information on how to contact you by electronic and paper mail. + + If the program does terminal interaction, make it output a short +notice like this when it starts in an interactive mode: + + Copyright (C) + This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. + This is free software, and you are welcome to redistribute it + under certain conditions; type `show c' for details. + +The hypothetical commands `show w' and `show c' should show the appropriate +parts of the General Public License. Of course, your program's commands +might be different; for a GUI interface, you would use an "about box". + + You should also get your employer (if you work as a programmer) or school, +if any, to sign a "copyright disclaimer" for the program, if necessary. +For more information on this, and how to apply and follow the GNU GPL, see +. + + 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 +. \ No newline at end of file diff --git a/Project.toml b/Project.toml index 0f40dd3..f65924a 100644 --- a/Project.toml +++ b/Project.toml @@ -3,6 +3,12 @@ uuid = "c2b282c3-c9c2-431d-80f7-a1a0561ebe55" authors = ["Xuanzhao Gao and contributors"] version = "1.0.0-DEV" +[deps] +ArtifactUtils = "8b73e784-e7d8-4ea5-973d-377fed4e3bce" +Artifacts = "56f22d72-fd6d-98f1-02f0-08ddc0907c33" +CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba" +Libdl = "8f399da3-3557-5675-b5ff-fb832c97cbdb" + [compat] julia = "1" diff --git a/benchmark/benchmark_CUDA_mapreduce.jl b/benchmark/benchmark_CUDA_mapreduce.jl new file mode 100644 index 0000000..50ff682 --- /dev/null +++ b/benchmark/benchmark_CUDA_mapreduce.jl @@ -0,0 +1,20 @@ +using TropicalNumbers, CUDA, BenchmarkTools + +function map_reduce_benchmark(m::T, n::T, k::T) where{T} + A = Tropical.(CUDA.randn(Float32, (m, k))) + B = Tropical.(CUDA.randn(Float32, (k, n))) + C = Tropical.(CUDA.randn(Float32, (k, n))) + + elapsed_time = @belapsed CUDA.@sync begin + $C = $A * $B + end + + work_load = 2 * m * n * k + flops = work_load / elapsed_time / 1e9 + @show m, n, k, elapsed_time, flops + return nothing +end + +map_reduce_benchmark(2560, 2048, 2048) +map_reduce_benchmark(2 * 2560, 2 * 2048, 2 * 2048) +map_reduce_benchmark(4 * 2560, 4 * 2048, 4 * 2048) \ No newline at end of file diff --git a/benchmark/benchmark_CuTropicalGemm.jl b/benchmark/benchmark_CuTropicalGemm.jl new file mode 100644 index 0000000..2e8082a --- /dev/null +++ b/benchmark/benchmark_CuTropicalGemm.jl @@ -0,0 +1,29 @@ +using CUDA +using BenchmarkTools +using CuTropicalGEMM + +function benchmakr_CuTropicalGemmFP32(m::T, n::T, k::T) where{T} + A = rand(Float32, (m * k)) + B = rand(Float32, (k * n)) + C = rand(Float32, (m * n)) + + CuA = CuArray(A) + CuB = CuArray(B) + CuC = CuArray(C) + + # I found hat @belapsed and CUDA.@sync can not properly benchmark the function from .so lib + elapsed_time = @belapsed CUDA.@sync begin + 1 + 1 + CuTropicalGemmMatmulFP32!($m, $n, $k, $CuA, $CuB, $CuC) + 1 + 1 + end + + work_load = 2 * m * n * k + flops = work_load / elapsed_time / 1e9 + @show m, n, k, elapsed_time, flops + return nothing +end + +benchmakr_CuTropicalGemmFP32(2560, 2048, 2048) +benchmakr_CuTropicalGemmFP32(2 * 2560, 2 * 2048, 2 * 2048) +benchmakr_CuTropicalGemmFP32(4 * 2560, 4 * 2048, 4 * 2048) \ No newline at end of file diff --git a/benchmark/benchmark_CuTropicalGemm_C.sh b/benchmark/benchmark_CuTropicalGemm_C.sh new file mode 100644 index 0000000..a6dd5e6 --- /dev/null +++ b/benchmark/benchmark_CuTropicalGemm_C.sh @@ -0,0 +1,6 @@ +#! /bin/bash + +nvcc -arch=sm_80 ../src/TropicalSGemmFP32.cu +./a.out + +rm a.out \ No newline at end of file diff --git a/benchmark/benchmark_GemmKerenls_Tropical.jl b/benchmark/benchmark_GemmKerenls_Tropical.jl new file mode 100644 index 0000000..6b71d6a --- /dev/null +++ b/benchmark/benchmark_GemmKerenls_Tropical.jl @@ -0,0 +1,83 @@ +# notice: this code is used to benchmark the Tropical matmul in GemmKernels.jl, which is not yet released in the latest version and only supported in julia@v1.7 +# to run the code, you need to manually download the latest version repo of GemmKernels.jl and activate the enviroment + +import CUDA +import InteractiveUtils + +using CUDA +using GemmKernels +using LinearAlgebra +using BenchmarkTools +using Test + +CUDA.allowscalar(false) + +function try_tropical(M, N, K) + for (A_type, B_type, CD_type, min_dimension) in [(Float32, Float32, Float32, 128)], + transpose_a = [true, false], + transpose_b = [true, false], + (OP_M, OP_N, OP_K) in [(8, 16, 2)] + + a_h = rand(A_type, (M, K)) / sqrt(A_type(K)) + b_h = rand(B_type, (K, N)) / sqrt(B_type(K)) + c_h = rand(CD_type, (M, N)) + d_h = similar(c_h) + + + # Transpose input if necessary + a_h = transpose_a ? transpose(a_h) : a_h + b_h = transpose_b ? transpose(b_h) : b_h + + a = CuArray(a_h) + b = CuArray(b_h) + c = CuArray(c_h) + d = similar(c) + + conf = GemmKernels.get_config( + gemm_shape = (M = M, N = N, K = K), + block_shape = (M = 64, N = 64, K = 32), + operator = Operator.TropicalFPUOp{OP_M, OP_N, OP_K, CD_type, A_type}, + global_a_layout = transpose_a ? Layout.AlignedRowMajor{A_type} : Layout.AlignedColMajor{A_type}, + global_b_layout = transpose_b ? Layout.AlignedRowMajor{B_type} : Layout.AlignedColMajor{B_type}, + + global_c_layout = Layout.AlignedColMajor{CD_type}, + global_d_layout = Layout.AlignedColMajor{CD_type}, + + is_a_col_major = !transpose_a, + is_b_col_major = !transpose_b, + ) + + n_iter = 1 + elapsed_time = @belapsed CUDA.@sync begin + GemmKernels.matmul($a, $b, $c, $d, $conf; kernel = Kernel.matmul_pipelined) + end + TFlops = (n_iter * M * N * K * 2 / elapsed_time) / 1e9 + @show TFlops, elapsed_time, transpose_a, transpose_b, M, N, K + + + d_c = Array(d) + + # random 1600 points took to test + if transpose_a == transpose_b == false + @testset begin + for _ in 1 : 40 + for _ in 1 : 40 + i = rand(1:M) + j = rand(1:N) + d_h[i, j] = c_h[i, j] + for k in 1 : K + d_h[i, j] = max(a_h[i, k] + b_h[k, j], d_h[i, j]) + end + @test isapprox(d_h[i, j], d_c[i, j]; rtol = sqrt(eps(A_type))) + end + end + end + end + end + return nothing +end + + +try_tropical(2560, 2048, 2048) +try_tropical(2 * 2560, 2 * 2048, 2 * 2048) +try_tropical(4 * 2560, 4 * 2048, 4 * 2048) \ No newline at end of file diff --git a/src/CuTropicalGEMM.jl b/src/CuTropicalGEMM.jl index 0caf629..168c111 100644 --- a/src/CuTropicalGEMM.jl +++ b/src/CuTropicalGEMM.jl @@ -1,5 +1,10 @@ module CuTropicalGEMM -# Write your package code here. +export CuTropicalGemmMatmulFP32! + +using CUDA +using Artifacts + +include("TropicalGemm_Cuda_wrapper.jl") end diff --git a/src/TropicalGemmFP32.cu b/src/TropicalGemmFP32.cu new file mode 100644 index 0000000..c015ad6 --- /dev/null +++ b/src/TropicalGemmFP32.cu @@ -0,0 +1,627 @@ +// This CUDA code is modified based on github repo https://github.com/Yinghan-Li/YHs_Sample, which is under GPL 3.0 License +// To build this code as .so lib, use command +// nvcc -Xcompiler -fPIC --shared TropicalGemmFP32.cu -o TropicalGemmFP32.so + +#include +#include +#include +#include +#include + +void random_init(float *data, size_t size) { + for (size_t i = 0; i < size; ++i) { + data[i] = float(rand()) / RAND_MAX; + } +} + + +bool check(const float *A, + const float *B, + const float *C, + int m, int n, int k) { + for (int i = 0; i < 20; ++i) { + int a = rand() % m; + for (int j = 0; j < 20; ++j) { + int b = rand() % n; + float sum = 0.f; + for (int p = 0; p < k; ++p) { + sum = max(sum, A[a * k + p] + B[b + p * n]); + } + + if (std::fabs(sum - C[a * n + b]) / std::fabs(sum) > 1e-5f) { + printf("C[%d][%d] not match, %f vs %f\n", i, j, sum, C[a * n + b]); + return false; + } + } + } + + return true; +} + +__device__ __forceinline__ +uint32_t smem_u32addr(const void *smem_ptr) { + uint32_t addr; + asm ("{.reg .u64 u64addr;\n" + " cvta.to.shared.u64 u64addr, %1;\n" + " cvt.u32.u64 %0, u64addr;}\n" + : "=r"(addr) + : "l"(smem_ptr) + ); + + return addr; +} + +__device__ __forceinline__ +void ldg32_nc(float ®, const void *ptr, bool guard) { + asm volatile ( + "{.reg .pred p;\n" + " setp.ne.b32 p, %2, 0;\n" +#if __CUDACC_VER_MAJOR__ >= 11 && __CUDACC_VER_MINOR__ >= 4 && \ + __CUDA_ARCH__ >= 750 + " @p ld.global.nc.L2::128B.f32 %0, [%1];}\n" +#else + " @p ld.global.nc.f32 %0, [%1];}\n" +#endif + : "=f"(reg) + : "l"(ptr), "r"((int)guard) + ); +} + +__device__ __forceinline__ +void ldg32_nc_0(float ®, const void *ptr, bool guard) { + asm volatile ( + "{.reg .pred p;\n" + " setp.ne.b32 p, %2, 0;\n" + " @!p mov.b32 %0, 0;\n" +#if __CUDACC_VER_MAJOR__ >= 11 && __CUDACC_VER_MINOR__ >= 4 && \ + __CUDA_ARCH__ >= 750 + " @p ld.global.nc.L2::128B.f32 %0, [%1];}\n" +#else + " @p ld.global.nc.f32 %0, [%1];}\n" +#endif + : "=f"(reg) + : "l"(ptr), "r"((int)guard) + ); +} + +__device__ __forceinline__ +void stg32(const float ®, void *ptr, bool guard) { + asm volatile ( + "{.reg .pred p;\n" + " setp.ne.b32 p, %2, 0;\n" + " @p st.global.f32 [%0], %1;}\n" + : : "l"(ptr), "f"(reg), "r"((int)guard) + ); +} + +__device__ __forceinline__ +void lds128(float ®0, float ®1, + float ®2, float ®3, + const uint32_t &addr) { + asm volatile ( + "ld.shared.v4.f32 {%0, %1, %2, %3}, [%4];\n" + : "=f"(reg0), "=f"(reg1), "=f"(reg2), "=f"(reg3) + : "r"(addr) + ); +} + +__device__ __forceinline__ +void sts32(const float ®, const uint32_t &addr) { + asm volatile ( + "st.shared.f32 [%0], %1;\n" + : : "r"(addr), "f"(reg) + ); +} + +__device__ __forceinline__ +void sts128(const float ®0, const float ®1, + const float ®2, const float ®3, + const uint32_t &addr) { + asm volatile ( + "st.shared.v4.f32 [%0], {%1, %2, %3, %4};\n" + : : "r"(addr), "f"(reg0), "f"(reg1), "f"(reg2), "f"(reg3) + ); +} + +struct StgFrag { + float data[4][4]; + + __device__ __forceinline__ + StgFrag(const float (&C_frag)[8][8], int tile_x, int tile_y) { + #pragma unroll + for (int i = 0; i < 4; ++i) { + #pragma unroll + for (int j = 0; j < 4; ++j) { + data[i][j] = C_frag[tile_y * 4 + i][tile_x * 4 + j]; + } + } + } +}; + +__device__ __noinline__ +void C_tile_wb(StgFrag C_frag, + float *C_stg_ptr, + const float *C_lds_ptr, + uint32_t C_sts_addr, + uint32_t m, + uint32_t n, + uint32_t m_idx, + uint32_t n_idx) { + __syncthreads(); + + #pragma unroll + for (int i = 0; i < 4; ++i) { + sts128(C_frag.data[i][0], + C_frag.data[i][1], + C_frag.data[i][2], + C_frag.data[i][3], + C_sts_addr + i * 8 * sizeof(float4)); + } + + __syncthreads(); + + uint32_t m_guard = m < m_idx ? 0 : m - m_idx; + + #pragma unroll + for (int i = 0; i < 16; ++i) { + stg32(C_lds_ptr[i * 32], + C_stg_ptr + i * n, + i < m_guard && n_idx < n); + } +} + +/* + * matrix A, B and C: row-major + * + * mma block: + * thread block tile: m128n128k8 + * warp tile: m32n64k8 + * thread tile: m8n8k8 + * thread fragment: + * matrixA: 8x1 FP32 + * matrixB: 1x8 FP32 + * + * ---------------------------------------------------------------- + * thread block tile map: + * + * 128 + * --|---------------------| + * B_tile 8| | + * --|---------------------| + * + * A_tile | 8 | | 64 | + * --|---| --|----------|----------| + * | | 32| warp_0 | warp_1 | + * | | --|----------|----------| + * | | | warp_2 | warp_3 | + * 128| | |----------|----------| + * | | | warp_4 | warp_5 | + * | | |----------|----------| + * | | | warp_6 | warp_7 | + * --|---| |----------|----------| + * + * ---------------------------------------------------------------- + * warp tile map: + * + * 'z' thread map to avoid LDS.128 shared memory broadcast limitation. + * + * | 32 || + * B_frag --|---|---|---|---|---|---|---|---||---|---|---|---|---|---|---|---| + * 1|///| | | | | | | ||///| | | | | | | | + * --|---|---|---|---|---|---|---|---||---|---|---|---|---|---|---|---| + * A_frag | 4 | || + * | 1 | || + * --|---|-- |---|---|---|---|---|---|---|---||---|---------------------------| + * |///|4 |t0 |t2 |t4 |t6 |t8 |t10|t12|t14||t0 | | + * |---|-- |---|---|---|---|---|---|---|---||---| | + * | | |t1 |t3 |t5 |t7 |t9 |t11|t13|t15|| | + * 16|---| |---|---|---|---|---|---|---|---|| | + * | | |t16|t18|t20|t22|t24|t26|t28|t30|| | + * |---| |---|---|---|---|---|---|---|---|| | + * | | |t17|t19|t21|t23|t25|t27|t29|t31|| | + * ==|===|=====|===|===|===|===|===|===|===|===||===|============================ + * |///| |t0 | ||t0 | | + * |---| |---| ||---| | + * | | | || | + * |---| | || | + * | | | || | + * |---| | || | + * | | | || | + * |---| |-------------------------------||-------------------------------| + * + */ +__global__ __launch_bounds__(256, 2) +void sgemm_128x128x8_kernel(const float *A, + const float *B, + float *C, + uint32_t m, + uint32_t n, + uint32_t k, + uint32_t A_ldg_step, // k * sizeof(float) + uint32_t B_ldg_step) { // n * sizeof(float) * 8 + /* + * matrix A & B thread block tile shared memory (double buffer) + * matrix A: 132 * 8 * 4Byte/item * double buffer = 4.125KB * 2 + * matrix B: 128 * 8 * 4Byte/item * double buffer = 8KB + * + * for double buffer faster switch, A_smem requires 8KB * 2 shared memory + * and 16KB aligned, B_smem should be 8KB aligned, then the double buffer + * can be switched by only 1 xor instruction: + * (uint32_t &)A_smem ^= 0x2000; + * (uint32_t &)B_smem ^= 0x1000; + */ + __shared__ __align__(16 * 1024) char smem[24 * 1024]; + float *A_smem = reinterpret_cast(smem); + float *B_smem = reinterpret_cast(smem + 16 * 1024); + + // A, B and C register fragment + float A_frag[2][8]; + float B_frag[2][8]; + float C_frag[8][8]; + #pragma unroll + for (int i = 0; i < 8; ++i) { + #pragma unroll + for (int j = 0; j < 8; ++j) { + C_frag[i][j] = 0; + } + } + + const uint32_t lane_id = threadIdx.x % 32; + const uint32_t warp_id = threadIdx.x / 32; + + // 4x8 threads each warp for FFMA + const uint32_t mma_tid_x = (lane_id / 2) % 8; + const uint32_t mma_tid_y = (lane_id / 16) * 2 + (lane_id % 2); + + // A_tile & B_tile ldg pointer + const char *A_ldg_ptr = (const char *)( + A + (blockIdx.y * 128 + threadIdx.x / 8 * 4) * k + threadIdx.x % 8); + const char *B_ldg_ptr = (const char *)( + B + (threadIdx.x / 32) * n + blockIdx.x * 128 + threadIdx.x % 32); + + // A_tile & B_tile sts/lds pointer + // using uint32_t pointer for faster double buffer switch + uint32_t A_sts_addr = smem_u32addr( + A_smem + (threadIdx.x % 8) * 132 + (threadIdx.x / 8) * 4); + uint32_t B_sts_addr = smem_u32addr( + B_smem + (threadIdx.x / 32) * 128 + (threadIdx.x % 32)); + + uint32_t A_lds_addr = smem_u32addr( + A_smem + (warp_id / 2) * 32 + mma_tid_y * 4); + uint32_t B_lds_addr = smem_u32addr( + B_smem + (warp_id % 2) * 64 + mma_tid_x * 4); + + // ldg_guard to avoid LDG out of bound + uint32_t A_ldg_guard = 0; + #pragma unroll + for (int i = 0; i < 4; ++i) { + int m_idx = blockIdx.y * 128 + threadIdx.x / 8 * 4 + i; + if (m_idx < m) { + A_ldg_guard |= (1u << i); + } + } + + uint32_t B_ldg_guard = 0; + #pragma unroll + for (int i = 0; i < 4; ++i) { + int n_idx = blockIdx.x * 128 + threadIdx.x % 32 + i * 32; + if (n_idx < n) { + B_ldg_guard |= (1u << i); + } + } + + float A_ldg_reg[4]; + float B_ldg_reg[4]; + + // 1'st A&B tile loaded before the k_tile loop + uint32_t k_tiles = (k + 7) / 8 - 1; + + // load 1'st tile to shared memory + { + uint32_t first_k_tile = k - k_tiles * 8; + + #pragma unroll + for (int i = 0; i < 4; ++i) { + bool guard = (A_ldg_guard & (1u << i)) != 0 && + threadIdx.x % 8 < first_k_tile; + ldg32_nc_0(A_ldg_reg[i], + A_ldg_ptr + i * A_ldg_step, + guard); + } + + sts128(A_ldg_reg[0], A_ldg_reg[1], A_ldg_reg[2], A_ldg_reg[3], + A_sts_addr); + + #pragma unroll + for (int i = 0; i < 4; ++i) { + bool guard = (B_ldg_guard & (1u << i)) != 0 && + threadIdx.x / 32 < first_k_tile; + ldg32_nc_0(B_ldg_reg[i], + B_ldg_ptr + i * 32 * sizeof(float), + guard); + } + + #pragma unroll + for (int i = 0; i < 4; ++i) { + sts32(B_ldg_reg[i], B_sts_addr + i * 32 * sizeof(float)); + } + + __syncthreads(); + + // switch double buffer + A_sts_addr ^= 0x2000; + B_sts_addr ^= 0x1000; + + // ldg pointer for next tile + A_ldg_ptr += first_k_tile * sizeof(float); + B_ldg_ptr += n * first_k_tile * sizeof(float); + } + + // load 1'st fragment + lds128(A_frag[0][0], A_frag[0][1], A_frag[0][2], A_frag[0][3], + A_lds_addr); + lds128(A_frag[0][4], A_frag[0][5], A_frag[0][6], A_frag[0][7], + A_lds_addr + 16 * sizeof(float)); + lds128(B_frag[0][0], B_frag[0][1], B_frag[0][2], B_frag[0][3], + B_lds_addr); + lds128(B_frag[0][4], B_frag[0][5], B_frag[0][6], B_frag[0][7], + B_lds_addr + 32 * sizeof(float)); + + // k_tiles loop + for (; k_tiles > 0; --k_tiles) { + #pragma unroll + for (int k_frag = 0; k_frag < 8; ++k_frag) { + // store next A&B tile to shared memory + if (k_frag == 7) { + sts128(A_ldg_reg[0], A_ldg_reg[1], A_ldg_reg[2], A_ldg_reg[3], + A_sts_addr); + #pragma unroll + for (int i = 0; i < 4; ++i) { + sts32(B_ldg_reg[i], B_sts_addr + i * 32 * sizeof(float)); + } + + __syncthreads(); + + // switch double buffer + A_lds_addr ^= 0x2000; + B_lds_addr ^= 0x1000; + A_sts_addr ^= 0x2000; + B_sts_addr ^= 0x1000; + + // ldg pointer for next tile + A_ldg_ptr += 8 * sizeof(float); + B_ldg_ptr += B_ldg_step; + } + + // load next A&B fragment from shared memory to register + lds128(A_frag[(k_frag + 1) % 2][0], + A_frag[(k_frag + 1) % 2][1], + A_frag[(k_frag + 1) % 2][2], + A_frag[(k_frag + 1) % 2][3], + A_lds_addr + (k_frag + 1) % 8 * 132 * sizeof(float)); + lds128(A_frag[(k_frag + 1) % 2][4], + A_frag[(k_frag + 1) % 2][5], + A_frag[(k_frag + 1) % 2][6], + A_frag[(k_frag + 1) % 2][7], + A_lds_addr + ((k_frag + 1) % 8 * 132 + 16) * sizeof(float)); + lds128(B_frag[(k_frag + 1) % 2][0], + B_frag[(k_frag + 1) % 2][1], + B_frag[(k_frag + 1) % 2][2], + B_frag[(k_frag + 1) % 2][3], + B_lds_addr + (k_frag + 1) % 8 * 128 * sizeof(float)); + lds128(B_frag[(k_frag + 1) % 2][4], + B_frag[(k_frag + 1) % 2][5], + B_frag[(k_frag + 1) % 2][6], + B_frag[(k_frag + 1) % 2][7], + B_lds_addr + ((k_frag + 1) % 8 * 128 + 32) * sizeof(float)); + + // load next A&B tile + if (k_frag == 0) { + #pragma unroll + for (int i = 0; i < 4; ++i) { + ldg32_nc(A_ldg_reg[i], + A_ldg_ptr + i * A_ldg_step, + (A_ldg_guard & (1u << i)) != 0); + } + + #pragma unroll + for (int i = 0; i < 4; ++i) { + ldg32_nc(B_ldg_reg[i], + B_ldg_ptr + i * 32 * sizeof(float), + (B_ldg_guard & (1u << i)) != 0); + } + } + + // FFMA loop + #pragma unroll + for (int i = 0; i < 8; ++i) { + #pragma unroll + for (int j = 0; j < 8; ++j) { + C_frag[i][j] = max(C_frag[i][j], A_frag[k_frag % 2][i] + B_frag[k_frag % 2][j]); + } + } + } + } + + // FFMA for the last tile + #pragma unroll + for (int k_frag = 0; k_frag < 8; ++k_frag) { + if (k_frag < 7) { + // load next A&B fragment from shared memory to register + lds128(A_frag[(k_frag + 1) % 2][0], + A_frag[(k_frag + 1) % 2][1], + A_frag[(k_frag + 1) % 2][2], + A_frag[(k_frag + 1) % 2][3], + A_lds_addr + (k_frag + 1) % 8 * 132 * sizeof(float)); + lds128(A_frag[(k_frag + 1) % 2][4], + A_frag[(k_frag + 1) % 2][5], + A_frag[(k_frag + 1) % 2][6], + A_frag[(k_frag + 1) % 2][7], + A_lds_addr + ((k_frag + 1) % 8 * 132 + 16) * sizeof(float)); + lds128(B_frag[(k_frag + 1) % 2][0], + B_frag[(k_frag + 1) % 2][1], + B_frag[(k_frag + 1) % 2][2], + B_frag[(k_frag + 1) % 2][3], + B_lds_addr + (k_frag + 1) % 8 * 128 * sizeof(float)); + lds128(B_frag[(k_frag + 1) % 2][4], + B_frag[(k_frag + 1) % 2][5], + B_frag[(k_frag + 1) % 2][6], + B_frag[(k_frag + 1) % 2][7], + B_lds_addr + ((k_frag + 1) % 8 * 128 + 32) * sizeof(float)); + } + + // FFMA loop + #pragma unroll + for (int i = 0; i < 8; ++i) { + #pragma unroll + for (int j = 0; j < 8; ++j) { + C_frag[i][j] = max(C_frag[i][j], A_frag[k_frag % 2][i] + B_frag[k_frag % 2][j]); + } + } + } + + // C_tile write back, reuse A&B tile shared memory buffer + uint32_t C_sts_addr = smem_u32addr((float4 *)(smem + warp_id * 2048) + + mma_tid_y * 4 * 8 + mma_tid_x); + const float *C_lds_ptr = (float *)(smem + warp_id * 2048) + lane_id; + + uint32_t m_idx = blockIdx.y * 128 + warp_id / 2 * 32; + uint32_t n_idx = blockIdx.x * 128 + warp_id % 2 * 64 + lane_id; + + float *C_stg_ptr = C + m_idx * n + n_idx; + + if (m_idx >= m) { + return; + } else if (m_idx + 32 <= m) { + uint32_t n_guard = n < n_idx ? 0 : n - n_idx; + + #pragma unroll + for (int i = 0; i < 2; ++i) { + #pragma unroll + for (int j = 0; j < 2; ++j) { + __syncthreads(); + + #pragma unroll + for (int p = 0; p < 4; ++p) { + sts128(C_frag[i * 4 + p][j * 4], + C_frag[i * 4 + p][j * 4 + 1], + C_frag[i * 4 + p][j * 4 + 2], + C_frag[i * 4 + p][j * 4 + 3], + C_sts_addr + p * 8 * sizeof(float4)); + } + + __syncthreads(); + + #pragma unroll + for (int p = 0; p < 16; ++p) { + stg32(C_lds_ptr[p * 32], + C_stg_ptr + (i * 16 + p) * n + j * 32, + j * 32 < n_guard); + } + } + } + } else { + #pragma unroll + for (int i = 0; i < 2; ++i) { + #pragma unroll + for (int j = 0; j < 2; ++j) { + StgFrag stg_frag(C_frag, j, i); + + C_tile_wb(stg_frag, + C_stg_ptr + i * 16 * n + j * 32, + C_lds_ptr, + C_sts_addr, + m, + n, + m_idx + i * 16, + n_idx + j * 32); + } + } + } +} + +extern "C" +void TropicalMatmul(int m, int n, int k, float *d_A, float *d_B, float *d_C){ + + dim3 grid((n + 127) / 128, (m + 127) / 128); + sgemm_128x128x8_kernel<<>>(d_A, d_B, d_C, m, n, k, k * sizeof(float), n * sizeof(float) * 8); + +} + +extern "C" +void MatrixPrint(int m, int k, float *A){ + // float *cpu_A; + // cudaMallocHost(&cpu_A, m * k * sizeof(float)); + // cudaMemcpy(A, cpu_A, m * k * sizeof(float), cudaMemcpyDefault); + + for(int i = 0; i < m; ++i){ + for(int j = 0; j < k; ++j){ + // printf("%f\n", cpu_A[i * k + j]); + printf("%f\n", A[i * k + j]); + } + } + + // cudaFreeHost(cpu_A); +} + +int main() { + int m = 4 * 2560; + int n = 4 * 2000; + int k = 4 * 2048; + int n_iter = 10; + + float *h_A, *h_B, *h_C; + cudaMallocHost(&h_A, m * k * sizeof(float)); + cudaMallocHost(&h_B, k * n * sizeof(float)); + cudaMallocHost(&h_C, m * n * sizeof(float)); + random_init(h_A, m * k); + random_init(h_B, k * n); + + float *d_A, *d_B, *d_C; + cudaMalloc(&d_A, m * k * sizeof(float)); + cudaMalloc(&d_B, k * n * sizeof(float)); + cudaMalloc(&d_C, m * n * sizeof(float)); + + cudaMemcpy(d_A, h_A, m * k * sizeof(float), cudaMemcpyDefault); + cudaMemcpy(d_B, h_B, k * n * sizeof(float), cudaMemcpyDefault); + + cudaEvent_t start, end; + cudaEventCreate(&start); + cudaEventCreate(&end); + + // dim3 grid((n + 127) / 128, (m + 127) / 128); + + // warmup + // sgemm_128x128x8_kernel<<>>( + // d_A, d_B, d_C, m, n, k, k * sizeof(float), n * sizeof(float) * 8); + + cudaEventRecord(start); + for (int i = 0; i < n_iter; ++i) { + // sgemm_128x128x8_kernel<<>>(d_A, d_B, d_C, m, n, k, k * sizeof(float), n * sizeof(float) * 8); + TropicalMatmul(m, n, k, d_A, d_B, d_C); + } + cudaEventRecord(end); + cudaEventSynchronize(end); + + float ms; + cudaEventElapsedTime(&ms, start, end); + cudaEventDestroy(start); + cudaEventDestroy(end); + + long workload = n_iter * long(m) * n * k * 2; + double gflops = (double(workload) / 1e9) / (double(ms) / 1e3); + printf("M: %d, N: %d, K: %d\n", M, N, K); + printf("TropicalSGemm: %f GFLOPS\n", gflops); + printf("time: %f ms\n", ms); + cudaMemcpy(h_C, d_C, m * n * sizeof(float), cudaMemcpyDefault); + + bool chk = check(h_A, h_B, h_C, m, n, k); + printf("Matrix_C check: %s\n", chk ? "OK" : "Failed"); + + cudaFree(d_A); + cudaFree(d_B); + cudaFree(d_C); + cudaFreeHost(h_A); + cudaFreeHost(h_B); + cudaFreeHost(h_C); +} diff --git a/src/TropicalGemm_Cuda_wrapper.jl b/src/TropicalGemm_Cuda_wrapper.jl new file mode 100644 index 0000000..b476de6 --- /dev/null +++ b/src/TropicalGemm_Cuda_wrapper.jl @@ -0,0 +1,15 @@ +export CuTropicalGemmMatmulFP32! + +const FP32_lib = joinpath(artifact"CUDA_lib", "TropicalGemmFP32.so") + + +# A wrapper for FP32 Tropical GEMM +function CuTropicalGemmMatmulFP32!(m::Integer, n::Integer, k::Integer, A::CuArray{Float32}, B::CuArray{Float32}, C::CuArray{Float32}) + A_ptr = pointer(A); + B_ptr = pointer(B); + C_ptr = pointer(C); + + @ccall FP32_lib.TropicalMatmul(m::Cint, n::Cint, k::Cint, A_ptr::CuPtr{Cfloat}, B_ptr::CuPtr{Cfloat}, C_ptr::CuPtr{Cfloat})::Cvoid + + return nothing +end \ No newline at end of file diff --git a/test/runtests.jl b/test/runtests.jl index 5684d84..2089f76 100644 --- a/test/runtests.jl +++ b/test/runtests.jl @@ -1,6 +1,7 @@ using CuTropicalGEMM using Test +using CUDA @testset "CuTropicalGEMM.jl" begin - # Write your tests here. + include("test_FP32_matmul.jl") end diff --git a/test/test_FP32_matmul.jl b/test/test_FP32_matmul.jl new file mode 100644 index 0000000..3742321 --- /dev/null +++ b/test/test_FP32_matmul.jl @@ -0,0 +1,32 @@ +# this function will return the (i, j) element of the Tropical result of C + A * B +function TropicalDirectMatmul_ij(i::Integer, j::Integer, m::Integer, n::Integer, k::Integer, A::Array{Float32}, B::Array{Float32}, C::Array{Float32}) + sum = zero(Float32) + for l in 1:k + sum = max(A[(i - 1) * k + l] + B[(l - 1) * n + j], sum) + end + return sum +end + +@testset "FP32 matmul 1024 * 1024 * 1024, random 1600 terms" begin + m, n, k = 1024, 1024, 1024 + + A = rand(Float32, (m * k)) + B = rand(Float32, (k * n)) + C = rand(Float32, (m * n)) + + CuA = CuArray(A) + CuB = CuArray(B) + CuC = CuArray(C) + + CuTropicalGemmMatmulFP32!(m, n, k, CuA, CuB, CuC) + + C_result = Array(CuC) + + for _ in 1:40 + for _ in 1:40 + a = rand(1:m) + b = rand(1:n) + @test TropicalDirectMatmul_ij(a, b, m, n, k, A, B, C) ≈ C_result[(a - 1) * n + b] + end + end +end \ No newline at end of file