├── .gitignore ├── LICENSE ├── README.md ├── cpp ├── .gitignore └── how_to_build_cpp_static_lib │ ├── README.md │ └── test_staticlib │ ├── .gitignore │ ├── CMakeLists.txt │ ├── liba │ ├── func.cc │ ├── func.h │ ├── funca.cc │ ├── funca.h │ └── global.cc │ ├── libb │ ├── func.cc │ ├── func.h │ ├── funcb.cc │ └── funcb.h │ ├── libc │ ├── funcc.cc │ └── funcc.h │ ├── test.cc │ └── test_a_inst.cc ├── cuda-check ├── README.md └── check.cu ├── cuda-simd └── README.md ├── elementwise └── README.md ├── gemm └── README.md ├── gemv └── README.md ├── nsight-compute └── README.md ├── reduce └── README.md ├── spmm └── README.md └── spmv └── README.md /.gitignore: -------------------------------------------------------------------------------- 1 | *.so 2 | *.a 3 | *.dylib 4 | *.dll 5 | *.lib 6 | .DS_Store 7 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # CUDA高频面试题汇总/C++笔记/CUDA笔记 📔📕📗 2 | 3 | 12 | 13 | CUDA-Learn-Note: CUDA 笔记 / 高频面试题汇总 / C++笔记，个人笔记，更新随缘: sgemm、sgemv、warp reduce、block reduce、dot、elementwise、softmax、layernorm、rmsnorm、histogram、relu、sigmoid ... 14 | 15 | ## 0x00 前言 16 | 前段时间参加了一些`大模型`面试，大部分都要手撕CUDA，因此也整体复习了一遍CUDA优化相关的内容，整理了一些高频题的基本写法，保存在这里也便于日后自己复习。当然，有些代码不一定是最优化解，比如GEMM，想要在面试短短的30分钟内写一个好的`GEMM` Kernel，是有些难度的。印象比较深刻的是，其中有一场面试2个多小时，一个小时问项目，剩下一个小时在写GEMM，虽然写的kernel很一般，但是印象还挺深刻的。[代码文件](./cuda-check/check.cu) 17 | TIPS: 仓库整理的代码为方便自己复习回顾，不喜欢的请自动跳过哈。 18 | 19 | ## 0x01 高频面试题汇总简介 20 |
21 | 22 | 相关kernel如下。也就是不到1000行代码，建议背下来，我个人是喜欢背记，背的过程中基本就慢慢理解所有细节。当然，每个人的学习方法都不一样哈，自己觉得舒服就行。 23 | 24 | - [x] [sgemm naive, sgemm + block-tile + k-tile + vec4](#sgemm) 25 | - [x] [sgemv k32/k128/k16 kernel](#sgemv) 26 | - [x] [warp/block reduce sum/max](#warpreduce) 27 | - [x] [block all reduce + vec4](#blockallreduce) 28 | - [x] [dot product, dot product + vec4](#dot) 29 | - [x] [elementwise, elementwise + vec4](#elementwise) 30 | - [x] [histogram, histogram + vec4](#histogram) 31 | - [x] [softmax, softmax + vec4 (grid level memory fence)](#softmax) 32 | - [x] [safe softmax, safe softmax + vec4](#safesoftmax) 33 | - [x] [sigmoid, sigmoid + vec4](#sigmoid) 34 | - [x] [relu, relu + vec4](#relu) 35 | - [x] [layer_norm, layer_norm + vec4](#layernorm) 36 | - [x] [rms_norm, rms_norm + vec4](#rmsnorm) 37 | - [x] [nms](#NMS) 38 | - [ ] sgemm + double buffer 39 | - [ ] sgemm + fp16 40 | - [ ] ... 41 | 42 | 43 | 题内话，大模型相关的岗位，手撕CUDA的概率非常大，leetcode反而写的少，就前段时间个人的经验，基本是4:1的比例，还是建议好好复习下CUDA。当然，这些只是最简单的kernel实现，比如flash_attn，FMHA这些优化手段，就不在这里写了，面试中基本都会问到。后边有空再补档一些文章吧。 44 | 45 | ## 0x02 sgemm naive, sgemm + block-tile + k-tile + vec4 ([©️back👆🏻](#kernellist)) 46 |
47 | 48 | ```c++ 49 | #include 50 | #include 51 | #include 52 | #include 53 | #include 54 | #include 55 | 56 | #define WARP_SIZE 32 57 | #define INT4(value) (reinterpret_cast(&(value))[0]) 58 | #define FLOAT4(value) (reinterpret_cast(&(value))[0]) 59 | 60 | // SGEMM: Block Tile + K Tile, with smem 61 | // Block Tile (BM, BN) + K Tile (BK=32) 62 | // grid((N + BN - 1) / BN, (M + BM - 1) / BM), block(BN, BM) 63 | // a: MxK, b: KxN, c: MxN, compute: c = a * b, all row major 64 | __global__ void sgemm(float* a, float* b, float* c, int M, int N, int K) { 65 | // [1] Block Tile: 32x32的block处理c上一块32x32的元素计算 66 | // [2] K Tile: 使用共享内存，并将K分块为BK大小的块 67 | constexpr int BM = 32; 68 | constexpr int BN = 32; 69 | constexpr int BK = 32; 70 | __shared__ float s_a[BM][BK], s_b[BK][BN]; 71 | 72 | int bx = blockIdx.x; 73 | int by = blockIdx.y; 74 | int tx = threadIdx.x; 75 | int ty = threadIdx.y; 76 | int tid = threadIdx.y * blockDim.x + tx; // tid within the block 77 | // load values to shared memory, 32x32 threads working together 78 | // to fetch data along the row direction of a and b both for s_a 79 | // and s_b 32x32x4x2=8KB, we use 32x32 threads within block to 80 | // load 32x32 elements from global memory to shared memory, namely, 81 | // each thread will load 1 element. 82 | int load_smem_a_m = tid / 32; // 0~31, tid / 32, tid / BM, threadIdx.y 83 | int load_smem_a_k = tid % 32; // 0~31, tid % 32, tid % BK, threadIdx.x 84 | int load_smem_b_k = tid / 32; // 0~31, tid / 32, tid / BK, threadIdx.y 85 | int load_smem_b_n = tid % 32; // 0~31, tid % 32, tid % BN, threadIdx.x 86 | int load_gmem_a_m = by * BM + load_smem_a_m; // global row of a and c 87 | int load_gmem_b_n = bx * BN + load_smem_b_n; // global col of b and c 88 | // if (load_gmem_a_m >= M || load_gmem_b_n >= N) return; 89 | 90 | float sum = 0.f; 91 | for (int bk = 0; bk < (K + BK - 1) / BK; ++bk) { 92 | int load_gmem_a_k = bk * BK + load_smem_a_k; 93 | int load_gmem_a_addr = load_gmem_a_m * K + load_gmem_a_k; 94 | s_a[load_smem_a_m][load_smem_a_k] = a[load_gmem_a_addr]; 95 | int load_gmem_b_k = bk * BK + load_smem_b_k; 96 | int load_gmem_b_addr = load_gmem_b_k * N + load_gmem_b_n; 97 | s_b[load_smem_b_k][load_smem_b_n] = b[load_gmem_b_addr]; 98 | __syncthreads(); 99 | #pragma unroll 100 | for (int k = 0; k < BK; ++k) { 101 | int comp_smem_a_m = load_smem_a_m; 102 | int comp_smem_b_n = load_smem_b_n; 103 | sum += s_a[comp_smem_a_m][k] * s_b[k][comp_smem_b_n]; 104 | } 105 | __syncthreads(); 106 | } 107 | int store_gmem_c_m = load_gmem_a_m; 108 | int store_gmem_c_n = load_gmem_b_n; 109 | int store_gmem_c_addr = store_gmem_c_m * N + store_gmem_c_n; 110 | c[store_gmem_c_addr] = sum; 111 | } 112 | 113 | // SGEMM: Block Tile + Thread Tile + K Tile + Vec4, with smem 114 | // BK:TILE_K=8 BM=BN=128 115 | // TM=TN=8 增加计算密度 BM/TM=16 BN/TN=16 116 | // dim3 blockDim(BN/TN, BM/TM); 117 | // dim3 gridDim((N + BN - 1) / BN, (M + BM - 1) / BM) 118 | __global__ void sgemm_thread_tile_vec4( 119 | float* a, float* b, float* c, int M, int N, int K) { 120 | // [1] Block Tile: 一个16x16的block处理C上大小为128X128的一个目标块 121 | // [2] Thread Tile: 每个thread负责计算TM*TN(8*8)个元素，增加计算密度 122 | // [3] K Tile: 将K分块，每块BK大小，迭代(K+BK-1/BK)次， 123 | // 每次计算TM*TN个元素各自的部分乘累加 124 | // [4] Vectorize: 减少load和store指令，使用float4 125 | constexpr int BM = 128; 126 | constexpr int BN = 128; 127 | constexpr int BK = 8; 128 | constexpr int TM = 8; 129 | constexpr int TN = 8; 130 | 131 | int bx = blockIdx.x; 132 | int by = blockIdx.y; 133 | int tx = threadIdx.x; 134 | int ty = threadIdx.y; 135 | int tid = threadIdx.y * blockDim.x + tx; // tid within the block 136 | __shared__ float s_a[BM][BK], s_b[BK][BN]; // 2*128*8*4=8KB 137 | 138 | // 0. 先计算shared memory中的索引 139 | // tid和需要加载的smem s_a[BM][BK] 之间的索引关系 BM=128 BK=8 按行读取 A行主序 140 | // 对于s_a每行8个数据，每个线程读取4个，需要2个线程；总共128行，需要128x2刚好256线程 141 | int load_smem_a_m = tid / 2; // tid/2 (128/8)*(128/8)=256 threads per block, tid/2->[0,128), BM=128 0~127 142 | int load_smem_a_k = (tid % 2 == 0) ? 0 : 4; // (tid%2 == 0) ? 0 : 4, col of s_a 0,4 143 | // tid和需要加载的smem s_b[BK][BN] 之间的索引关系 BK=8 BN=128 按行读取 B行主序 144 | // 对于s_b每行128个数据，每个线程读4个数据，需要32个线程；总共8行，需要32x8=256个线程 145 | int load_smem_b_k = tid / 32; // tid/32, row of s_b 256/32=8 行 0~7 146 | int load_smem_b_n = (tid % 32) * 4; // (tid % 32) * 4, col of s_b 0,4,...,124 147 | // 1. 再计算全局内存中的索引 148 | // 要加载到s_a中的元素对应到A全局内存中的行数 每个block负责出C中大小为BM*BN的块 149 | int load_gmem_a_m = by * BM + load_smem_a_m; // global row of a and c 150 | int load_gmem_b_n = bx * BN + load_smem_b_n; // global col of b and c 151 | 152 | float r_c[TM][TN] = {0.0}; // 8x8 153 | // 2. 先对K进行分块，每块BK大小 154 | for (int bk = 0; bk < (K + BK - 1) / BK; ++bk) { 155 | // 加载数据到共享内存smem s_a BM*BK 128*8 vectorize float4 156 | int load_gmem_a_k = bk * BK + load_smem_a_k; // global col of a 157 | int load_gmem_a_addr = load_gmem_a_m * K + load_gmem_a_k; 158 | FLOAT4(s_a[load_smem_a_m][load_smem_a_k]) = FLOAT4(a[load_gmem_a_addr]); 159 | // 加载数据到共享内存smem s_b BK*BN 8*128 vectorize float4 160 | int load_gmem_b_k = bk * BK + load_smem_b_k; // global row of b 161 | int load_gmem_b_addr = load_gmem_b_k * N + load_gmem_b_n; 162 | FLOAT4(s_b[load_smem_b_k][load_smem_b_n]) = FLOAT4(b[load_gmem_b_addr]); 163 | __syncthreads(); 164 | #pragma unroll 165 | for (int k = 0; k < BK; k++) { 166 | // 3. 每个线程负责计算BM*BN(12x128)中的TM*TN(8x8)个元素 167 | #pragma unroll 168 | for (int m = 0; m < TM; m++) { 169 | #pragma unroll 170 | for (int n = 0; n < TN; n++) { 171 | // k from 0~7，0 ~ BK, ty and tx range from 0 to 15, 16x8=128 172 | int comp_smem_a_m = ty * TM + m; // 128*8 128/TM(8)=16 M方向 16线程 173 | int comp_smem_b_n = tx * TN + n; // 8*128 128/TN(8)=16 N方向 16线程 174 | r_c[m][n] += s_a[comp_smem_a_m][k] * s_b[k][comp_smem_b_n]; 175 | } 176 | } 177 | } 178 | __syncthreads(); 179 | } 180 | 181 | #pragma unroll 182 | for (int m = 0; m < TM; ++m) { 183 | int store_gmem_c_m = by * BM + ty * TM + m; 184 | #pragma unroll 185 | for (int n = 0; n < TN; n += 4) { 186 | int store_gmem_c_n = bx * BN + tx * TN + n; 187 | int store_gmem_c_addr = store_gmem_c_m * N + store_gmem_c_n; 188 | FLOAT4(c[store_gmem_c_addr]) = FLOAT4(r_c[m][n]); 189 | } 190 | } 191 | } 192 | ``` 193 | 这里gemm的实现比较简单，只使用了CUDA Cores，并且只实现Block Tile + K Tile以及Block Tile + K Tile+Thread Tile+向量化的版本。主要在于如何加载gmem中的数据到smem，也就是把全局内存中的数据索引mapping到共享内存中的。核心思维：把一个block中的线程id按照线性来理解，然后把这个线性的id和全局内存索引以及共享内存索引进行匹配。比如Block Tile + K Tile的实现，block内一共32x32个Threads，需要加载到smem的数据也是32x32，那么，最简单的做法，只需要每个线程加载一个互不重复数据即可。NOTE，本文的gemm kernel修改自：[紫气东来：CUDA（三）：通用矩阵乘法：从入门到熟练](https://zhuanlan.zhihu.com/p/657632577) 194 | 195 | 196 | ## 0x03 warp/block reduce sum/max ([©️back👆🏻](#kernellist)) 197 |
198 | 199 | ```C++ 200 | // Warp Reduce Sum 201 | template 202 | __device__ __forceinline__ float warp_reduce_sum(float val) { 203 | #pragma unroll 204 | for (int mask = kWarpSize >> 1; mask >= 1; mask >>= 1) { 205 | val += __shfl_xor_sync(0xffffffff, val, mask); 206 | } 207 | return val; 208 | } 209 | 210 | // Warp Reduce Max 211 | template 212 | __device__ __forceinline__ float warp_reduce_max(float val) { 213 | #pragma unroll 214 | for (int mask = kWarpSize >> 1; mask >= 1; mask >>= 1) { 215 | val = fmaxf(val, __shfl_xor_sync(0xffffffff, val, mask)); 216 | } 217 | return val; 218 | } 219 | 220 | // Block reduce sum/max/min device helper for Layer/RMS Norm/Softmax etc. 221 | // grid 1D block 1D, grid(N/128), block(128) 222 | template 223 | __device__ __forceinline__ float block_reduce_sum(float val) { 224 | // always <= 32 warps per block (limited by 1024 threads per block) 225 | constexpr int NUM_WARPS = (NUM_THREADS + WARP_SIZE - 1) / WARP_SIZE; 226 | int warp = threadIdx.x / WARP_SIZE; 227 | int lane = threadIdx.x % WARP_SIZE; 228 | static __shared__ float shared[NUM_WARPS]; 229 | 230 | val = warp_reduce_sum(val); 231 | if (lane == 0) shared[warp] = val; 232 | __syncthreads(); 233 | val = (lane < NUM_WARPS) ? shared[lane] : 0.0f; 234 | val = warp_reduce_sum(val); 235 | return val; 236 | } 237 | 238 | template 239 | __device__ __forceinline__ float block_reduce_max(float val) { 240 | // always <= 32 warps per block (limited by 1024 threads per block) 241 | constexpr int NUM_WARPS = (NUM_THREADS + WARP_SIZE - 1) / WARP_SIZE; 242 | int warp = threadIdx.x / WARP_SIZE; 243 | int lane = threadIdx.x % WARP_SIZE; 244 | static __shared__ float shared[NUM_WARPS]; 245 | 246 | val = warp_reduce_max(val); 247 | if (lane == 0) shared[warp] = val; 248 | __syncthreads(); 249 | val = (lane < NUM_WARPS) ? shared[lane] : -FLT_MAX; 250 | val = warp_reduce_max(val); 251 | return val; 252 | } 253 | ``` 254 | warp reduce几乎已经成为大部分reduce kernel的标准写法了，比如vLLM中，就是这种经典的写法。所以，先搞懂warp reduce（也就是搞懂各种warp functions的用法），再去写其他kernel，思路就会容易很多。需要注意的是，warp函数处理的是寄存器上的数据，也就是说，此时，没必要先加载数据到smem，再进行reduce，直接加载到寄存器即可（以前犯过这个小错误...）。Warp Functions建议参考：[jhang：CUDA编程入门之Warp-Level Primitives](https://zhuanlan.zhihu.com/p/572820783) 255 | 256 | ## 0x04 block all reduce + vec4 ([©️back👆🏻](#kernellist)) 257 |
258 | 259 | ```c++ 260 | // Block All Reduce Sum 261 | // grid(N/128), block(128) 262 | // a: Nx1, y=sum(a) 263 | template 264 | __global__ void block_all_reduce_sum(float* a, float* y, int N) { 265 | int tid = threadIdx.x; 266 | int idx = blockIdx.x * NUM_THREADS + tid; 267 | constexpr int NUM_WARPS = (NUM_THREADS + WARP_SIZE - 1) / WARP_SIZE; 268 | __shared__ float reduce_smem[NUM_WARPS]; 269 | // keep the data in register is enougth for warp operaion. 270 | float sum = (idx < N) ? a[idx] : 0.0f; 271 | int warp = tid / WARP_SIZE; 272 | int lane = tid % WARP_SIZE; 273 | // perform warp sync reduce. 274 | sum = warp_reduce_sum(sum); 275 | // warp leaders store the data to shared memory. 276 | if (lane == 0) reduce_smem[warp] = sum; 277 | __syncthreads(); // make sure the data is in shared memory. 278 | // the first warp compute the final sum. 279 | sum = (lane < NUM_WARPS) ? reduce_smem[lane] : 0.0f; 280 | if (warp == 0) sum = warp_reduce_sum(sum); 281 | if (tid == 0) atomicAdd(y, sum); 282 | } 283 | 284 | // Block All Reduce Sum + float4 285 | // grid(N/128), block(128/4) 286 | // a: Nx1, y=sum(a) 287 | template 288 | __global__ void block_all_reduce_sum_vec4(float* a, float* y, int N) { 289 | int tid = threadIdx.x; 290 | int idx = (blockIdx.x * NUM_THREADS + tid) * 4; 291 | constexpr int NUM_WARPS = (NUM_THREADS + WARP_SIZE - 1) / WARP_SIZE; 292 | __shared__ float reduce_smem[NUM_WARPS]; 293 | 294 | float4 reg_a = FLOAT4(a[idx]); 295 | // keep the data in register is enougth for warp operaion. 296 | float sum = (idx < N) ? (reg_a.x + reg_a.y + reg_a.z + reg_a.w) : 0.0f; 297 | int warp = tid / WARP_SIZE; 298 | int lane = tid % WARP_SIZE; 299 | // perform warp sync reduce. 300 | sum = warp_reduce_sum(sum); 301 | // warp leaders store the data to shared memory. 302 | if (lane == 0) reduce_smem[warp] = sum; 303 | __syncthreads(); // make sure the data is in shared memory. 304 | // the first warp compute the final sum. 305 | sum = (lane < NUM_WARPS) ? reduce_smem[lane] : 0.0f; 306 | if (warp == 0) sum = warp_reduce_sum(sum); 307 | if (tid == 0) atomicAdd(y, sum); 308 | } 309 | ``` 310 | block all reduce是在warp reduce的基础上进行的，reduce_smem这部分的共享内存申请无法避免，这是用来同步每个warp之间得到局部结果。注意，最后，还需要atomicAdd做一个block级别的原子操作，以得到全局的和。float4向量化优化访存，可以减缓WarpScheduler发送指令的压力。 311 | 312 | ## 0x05 sgemv k32/k128/k16 kernel ([©️back👆🏻](#kernellist)) 313 |
314 | 315 | ```C++ 316 | // SGEMV: Warp SGEMV K32 317 | // 假设K为32的倍数，每个warp负责一行 318 | // grid(M/4), block(32,4) blockDim.x=32=K, blockDim.y=4 319 | // a: MxK, x: Kx1, y: Mx1, compute: y = a * x 320 | __global__ void sgemv_k32(float* a, float* x, float* y, int M, int K) { 321 | int tx = threadIdx.x; // 0~31 322 | int ty = threadIdx.y; // 0~4 323 | int bx = blockIdx.x; // 0~M/4 324 | int lane = tx % WARP_SIZE; // 0~31 325 | int m = bx * blockDim.y + ty; // (0~M/4) * 4 + (0~3) 326 | if (m < M) { 327 | float sum = 0.0f; 328 | int NUM_WARPS = (K + WARP_SIZE - 1) / WARP_SIZE; 329 | #pragma unroll 330 | for (int w = 0; w < NUM_WARPS; ++w) { 331 | // 若NUM_WARPS>=2，先将当前行的数据累加到第一个warp中 332 | int k = w * WARP_SIZE + lane; 333 | sum += a[m * K + k] * x[k]; 334 | } 335 | sum = warp_reduce_sum(sum); 336 | if (lane == 0) y[m] = sum; 337 | } 338 | } 339 | 340 | // SGEMV: Warp SGEMV K128 + Vec4 341 | // 假设K为128的倍数 float4 342 | // grid(M/4), block(32,4) blockDim.x=32=K, blockDim.y=4 343 | // a: MxK, x: Kx1, y: Mx1, compute: y = a * x 344 | __global__ void sgemv_k128(float* a, float* x, float* y, int M, int K) { 345 | // 每个线程负责4个元素，一个warp覆盖128个元素 346 | int tx = threadIdx.x; // 0~31 347 | int ty = threadIdx.y; // 0~3 348 | int bx = blockIdx.x; // 0~M/4 349 | int lane = tx % WARP_SIZE; // 0~31 350 | int m = blockDim.y * bx + ty; // (0~M/4) * 4 + (0~3) 351 | 352 | if (m < M) { 353 | float sum = 0.0f; 354 | // process 4*WARP_SIZE elements per warp. 355 | int NUM_WARPS = (((K + WARP_SIZE - 1) / WARP_SIZE) + 4 - 1) / 4; 356 | #pragma unroll 357 | for (int w = 0; w < NUM_WARPS; ++w) { 358 | int k = (w * WARP_SIZE + lane) * 4; 359 | float4 reg_x = FLOAT4(x[k]); 360 | float4 reg_a = FLOAT4(a[m * K + k]); 361 | sum += (reg_a.x * reg_x.x + reg_a.y * reg_x.y 362 | + reg_a.z * reg_x.z + reg_a.w * reg_x.w); 363 | } 364 | sum = warp_reduce_sum(sum); 365 | if(lane == 0) y[m] = sum; 366 | } 367 | } 368 | 369 | // SGEMV: Warp SGEMV K16 370 | // 假设K为16 < 32,每个warp负责2行，每行有16个元素 371 | // NUM_THREADS=128, NUM_WARPS=NUM_THREADS/WARP_SIZE; 372 | // NUM_ROWS=NUM_WARPS * ROW_PER_WARP, grid(M/NUM_ROWS), block(32,NUM_WARPS) 373 | // a: MxK, x: Kx1, y: Mx1, compute: y = a * x 374 | template 375 | __global__ void sgemv_k16(float* A, float* x, float* y, int M, int K) { 376 | constexpr int K_WARP_SIZE = (WARP_SIZE + ROW_PER_WARP - 1) / ROW_PER_WARP; 377 | int tx = threadIdx.x; // 0~31 378 | int ty = threadIdx.y; // 0~NUM_WARPS 379 | int bx = blockIdx.x; // 0~M/NUM_ROWS (NUM_ROWS=NUM_WARPS * ROW_PER_WARP) 380 | int lane = tx % WARP_SIZE; // 0~31 381 | int k = lane % K_WARP_SIZE; // 0~15 382 | // gloabl row of a: MxK and y:Mx1, blockDim.y=NUM_WARPS 383 | int m = (blockDim.y * bx + ty) * ROW_PER_WARP + lane / K_WARP_SIZE; 384 | if (m < M) { 385 | float sum = A[m * K + k] * x[k]; 386 | sum = warp_reduce_sum(sum); 387 | // 注意是k == 0，而不是lane == 0 388 | if(k == 0) y[m] = sum; 389 | } 390 | } 391 | ``` 392 | 估计有些大佬倒立都能写sgemv的各种优化版了，核心思路其实也是基于warp reduce，考虑K的不同情况进行优化。本文的sgemv kernel修改自：[有了琦琦的棍子：深入浅出GPU优化系列：gemv优化](https://zhuanlan.zhihu.com/p/494144694) 393 | 394 | ## 0x06 dot product, dot product + vec4 ([©️back👆🏻](#kernellist)) 395 |
396 | 397 | ```c++ 398 | // Dot Product 399 | // grid(N/128), block(128) 400 | // a: Nx1, b: Nx1, y=sum(elementwise_mul(a,b)) 401 | template 402 | __global__ void dot(float* a, float* b, float* y, int N) { 403 | int tid = threadIdx.x; 404 | int idx = blockIdx.x * NUM_THREADS + tid; 405 | constexpr int NUM_WARPS = (NUM_THREADS + WARP_SIZE - 1) / WARP_SIZE; 406 | __shared__ float reduce_smem[NUM_WARPS]; 407 | 408 | // keep the data in register is enougth for warp operaion. 409 | float prod = (idx < N) ? a[idx] * b[idx] : 0.0f; 410 | int warp = tid / WARP_SIZE; 411 | int lane = tid % WARP_SIZE; 412 | // perform warp sync reduce. 413 | prod = warp_reduce_sum(prod); 414 | // warp leaders store the data to shared memory. 415 | if (lane == 0) reduce_smem[warp] = prod; 416 | __syncthreads(); // make sure the data is in shared memory. 417 | // the first warp compute the final sum. 418 | prod = (lane < NUM_WARPS) ? reduce_smem[lane] : 0.0f; 419 | if (warp == 0) prod = warp_reduce_sum(prod); 420 | if (tid == 0) atomicAdd(y, prod); 421 | } 422 | 423 | // Dot Product + Vec4 424 | // grid(N/128), block(128/4) 425 | // a: Nx1, b: Nx1, y=sum(elementwise_mul(a,b)) 426 | template 427 | __global__ void dot_vec4(float* a, float* b, float* y, int N) { 428 | int tid = threadIdx.x; 429 | int idx = (blockIdx.x * NUM_THREADS + tid) * 4; 430 | constexpr int NUM_WARPS = (NUM_THREADS + WARP_SIZE - 1) / WARP_SIZE; 431 | __shared__ float reduce_smem[NUM_WARPS]; 432 | 433 | float4 reg_a = FLOAT4(a[idx]); 434 | float4 reg_b = FLOAT4(b[idx]); 435 | float prod = (idx < N) ? (reg_a.x * reg_b.x + reg_a.y * reg_b.y 436 | + reg_a.z * reg_b.z + reg_a.w * reg_b.w) : 0.0f; 437 | int warp = tid / WARP_SIZE; 438 | int lane = tid % WARP_SIZE; 439 | // perform warp sync reduce. 440 | prod = warp_reduce_sum(prod); 441 | // warp leaders store the data to shared memory. 442 | if (lane == 0) reduce_smem[warp] = prod; 443 | __syncthreads(); // make sure the data is in shared memory. 444 | // the first warp compute the final sum. 445 | prod = (lane < NUM_WARPS) ? reduce_smem[lane] : 0.0f; 446 | if (warp == 0) prod = warp_reduce_sum(prod); 447 | if (tid == 0) atomicAdd(y, prod); 448 | } 449 | ``` 450 | dot product kernel的核心就是block reduce，不多说了。 451 | 452 | ## 0x07 elementwise, elementwise + vec4 ([©️back👆🏻](#kernellist)) 453 |
454 | 455 | ```c++ 456 | // ElementWise Add 457 | // grid(N/128), block(128) 458 | // a: Nx1, b: Nx1, c: Nx1, c = elementwise_add(a, b) 459 | __global__ void elementwise_add(float* a, float* b, float* c, int N) { 460 | int idx = blockIdx.x * blockDim.x + threadIdx.x; 461 | if (idx < N) c[idx] = a[idx] + b[idx]; 462 | } 463 | 464 | // ElementWise Add + Vec4 465 | // grid(N/128), block(128/4) 466 | // a: Nx1, b: Nx1, c: Nx1, c = elementwise_add(a, b) 467 | __global__ void elementwise_add_vec4(float* a, float* b, float* c, int N) { 468 | int idx = 4 * (blockIdx.x * blockDim.x + threadIdx.x); 469 | if (idx < N) { 470 | float4 reg_a = FLOAT4(a[idx]); 471 | float4 reg_b = FLOAT4(b[idx]); 472 | float4 reg_c; 473 | reg_c.x = reg_a.x + reg_b.x; 474 | reg_c.y = reg_a.y + reg_b.y; 475 | reg_c.z = reg_a.z + reg_b.z; 476 | reg_c.w = reg_a.w + reg_b.w; 477 | FLOAT4(c[idx]) = reg_c; 478 | } 479 | } 480 | ``` 481 | elementwise可以考虑加点向量化进行访存优化。 482 | 483 | ## 0x08 histogram, histogram + vec4 484 |
485 | 486 | ```c++ 487 | // Histogram 488 | // grid(N/128), block(128) 489 | // a: Nx1, y: count histogram 490 | __global__ void histogram(int* a, int* y, int N) { 491 | int idx = blockIdx.x * blockDim.x + threadIdx.x; 492 | if (idx < N) atomicAdd(&(y[a[idx]]), 1); 493 | } 494 | 495 | // Histogram + Vec4 496 | // grid(N/128), block(128/4) 497 | // a: Nx1, y: count histogram 498 | __global__ void histogram_vec4(int* a, int* y, int N) { 499 | int idx = 4 * (blockIdx.x * blockDim.x + threadIdx.x); 500 | if (idx < N) { 501 | int4 reg_a = INT4(a[idx]); 502 | atomicAdd(&(y[reg_a.x]), 1); 503 | atomicAdd(&(y[reg_a.y]), 1); 504 | atomicAdd(&(y[reg_a.z]), 1); 505 | atomicAdd(&(y[reg_a.w]), 1); 506 | } 507 | } 508 | ``` 509 | 统计频数直方图，很简单，两行代码搞定。 510 | 511 | ## 0x09 softmax, softmax + vec4 (grid level memory fence) ([©️back👆🏻](#kernellist)) 512 |
513 | 514 | ```c++ 515 | // Softmax x: N, y: N 516 | // grid(N/128), block(K=128) 517 | template 518 | __global__ void softmax(float* x, float* y, float* total, int N) { 519 | const int tid = threadIdx.x; 520 | const int idx = blockIdx.x * blockDim.x + tid; 521 | constexpr int NUM_WARPS = (NUM_THREADS + WARP_SIZE - 1) / WARP_SIZE; 522 | __shared__ float reduce_smem[NUM_WARPS]; 523 | 524 | float sum = (idx < N) ? expf(x[idx]) : 0.0f; 525 | int warp = tid / WARP_SIZE; 526 | int lane = tid % WARP_SIZE; 527 | sum = warp_reduce_sum(sum); 528 | if (lane == 0) reduce_smem[warp] = sum; 529 | __syncthreads(); 530 | // compute the final sum in each warp 531 | sum = (lane < NUM_WARPS) ? reduce_smem[lane] : 0.0f; 532 | sum = warp_reduce_sum(sum); // sum(e^x_0,...,e^x_n-1) 533 | // get the total sum of all blocks. 534 | if (tid == 0) atomicAdd(total, sum); 535 | __threadfence(); // grid level memory fence 注意这里需要网格级别的内存同步 536 | // e^x_i/sum(e^x_0,...,e^x_n-1) 537 | if (idx < N) y[idx] = block_smem[tid] / (*total); 538 | } 539 | 540 | // Softmax x: N, y: N 541 | // grid(N/128), block(K=128) 542 | template 543 | __global__ void softmax_v2(float* x, float* y, float* total, int N) { 544 | const int tid = threadIdx.x; 545 | const int idx = blockIdx.x * blockDim.x + tid; 546 | 547 | float exp_val = (idx < N) ? expf(x[idx]) : 0.0f; 548 | float sum = block_reduce_sum(exp_val); 549 | // get the total sum of all blocks. 550 | if (tid == 0) atomicAdd(total, sum); 551 | __threadfence(); // grid level memory fence 注意这里需要网格级别的内存同步 552 | // e^x_i/sum(e^x_0,...,e^x_n-1) 553 | if (idx < N) y[idx] = exp_val / (*total); 554 | } 555 | 556 | // Softmax Vec4 x: N, y: N 557 | // grid(N/128), block(128/4) 558 | template 559 | __global__ void softmax_v2_vec4(float* x, float* y, float* total, int N) { 560 | const int tid = threadIdx.x; 561 | const int idx = (blockIdx.x * blockDim.x + tid) * 4; 562 | 563 | float4 reg_x = FLOAT4(x[idx]); 564 | float4 reg_exp; 565 | reg_exp.x = (idx < N) ? expf(reg_x.x) : 0.0f; 566 | reg_exp.y = (idx < N) ? expf(reg_x.y) : 0.0f; 567 | reg_exp.z = (idx < N) ? expf(reg_x.z) : 0.0f; 568 | reg_exp.w = (idx < N) ? expf(reg_x.w) : 0.0f; 569 | float exp_val = (reg_exp.x + reg_exp.y + reg_exp.z + reg_exp.w); 570 | float sum = block_reduce_sum(exp_val); 571 | // get the total sum of all blocks. 572 | if (tid == 0) atomicAdd(total, sum); 573 | __threadfence(); // grid level memory fence 注意这里需要网格级别的内存同步 574 | // e^x_i/sum(e^x_0,...,e^x_n-1) 575 | if (idx < N) { 576 | float4 reg_y; 577 | reg_y.x = reg_exp.x / (*total); 578 | reg_y.y = reg_exp.y / (*total); 579 | reg_y.z = reg_exp.z / (*total); 580 | reg_y.w = reg_exp.w / (*total); 581 | FLOAT4(y[idx]) = reg_y; 582 | } 583 | } 584 | ``` 585 | softmax稍微要注意的就是内存同步的问题，这里，你需要做一个网格级别的同步，而不能仅仅是block级别，否则拿不到全局的exp sum作为分母项。因此使用 __threadfence 这个网格及内存同步操作。不过效率我还没测过，实在要高效的话，可能得整成FA2那样的 1-pass + online softmax的实现。不过，如果是面试的话，就不要太为难自己了...，但是FA1/FA2的论文很经典，强烈建议多读几遍。 586 | 587 | ## 0x0a safe softmax, safe softmax + vec4 ([©️back👆🏻](#kernellist)) 588 |
589 | 590 | ```c++ 591 | // Safe Softmax x: N, y: N 592 | // grid(N/128), block(K=128) 593 | template 594 | __global__ void softmax_safe(float* x, float* y, float* total, int N) { 595 | const int tid = threadIdx.x; 596 | const int idx = blockIdx.x * blockDim.x + tid; 597 | 598 | float ori_val = (idx < N) ? x[idx] : -FLT_MAX; 599 | float max_val = block_reduce_max(ori_val); 600 | float exp_val = (idx < N) ? expf(ori_val - max_val) : 0.0f; 601 | float sum = block_reduce_sum(exp_val); 602 | // get the total sum of all blocks. 603 | if (tid == 0) atomicAdd(total, sum); 604 | __threadfence(); // grid level memory fence 605 | // e^x_i/sum(e^x_0,...,e^x_n-1) 606 | if (idx < N) y[idx] = exp_val / (*total); 607 | } 608 | ``` 609 | 对比softmax减去一个max值防止数值溢出，比如float16。 610 | 611 | ## 0x0b sigmoid, sigmoid + vec4 ([©️back👆🏻](#kernellist)) 612 |
613 | 614 | ```c++ 615 | // Sigmoid x: N, y: N y=1/(1+exp(-x)) 616 | // grid(N/128), block(K=128) 617 | __global__ void sigmoid(float* x, float* y, int N) { 618 | int idx = blockIdx.x * blockDim.x + threadIdx.x; 619 | if (idx < N) y[idx] = 1.0f / (1.0f + expf(-x[idx])); 620 | } 621 | 622 | // Sigmoid x: N, y: N y=1/(1+exp(-x)) Vec4 623 | // grid(N/128), block(128/4) 624 | __global__ void sigmoid_vec4(float* x, float* y, int N) { 625 | int idx = (blockIdx.x * blockDim.x + threadIdx.x) * 4; 626 | if (idx < N) { 627 | float4 reg_x = FLOAT4(x[idx]); 628 | float4 reg_y; 629 | reg_y.x = 1.0f / (1.0f + expf(-reg_x.x)); 630 | reg_y.y = 1.0f / (1.0f + expf(-reg_x.y)); 631 | reg_y.z = 1.0f / (1.0f + expf(-reg_x.z)); 632 | reg_y.w = 1.0f / (1.0f + expf(-reg_x.w)); 633 | FLOAT4(y[idx]) = reg_y; 634 | } 635 | } 636 | ``` 637 | 638 | ## 0x0c relu, relu + vec4 ([©️back👆🏻](#kernellist)) 639 |
640 | 641 | ```c++ 642 | // Relu x: N, y: N y=max(0,x) 643 | // grid(N/128), block(K=128) 644 | __global__ void relu(float* x, float* y, int N) { 645 | int idx = blockIdx.x * blockDim.x + threadIdx.x; 646 | if (idx < N) y[idx] = fmaxf(0.0f, x[idx]); 647 | } 648 | 649 | // Relu x: N, y: N y=max(0,x) Vec4 650 | // grid(N/128/4), block(128/4) 651 | __global__ void relu_vec4(float* x, float* y, int N) { 652 | int idx = (blockIdx.x * blockDim.x + threadIdx.x) * 4; 653 | if (idx < N) { 654 | float4 reg_x = FLOAT4(x[idx]); 655 | float4 reg_y; 656 | reg_y.x = fmaxf(0.0f, reg_x.x); 657 | reg_y.y = fmaxf(0.0f, reg_x.y); 658 | reg_y.z = fmaxf(0.0f, reg_x.z); 659 | reg_y.w = fmaxf(0.0f, reg_x.w); 660 | FLOAT4(y[idx]) = reg_y; 661 | } 662 | } 663 | ``` 664 | 665 | ## 0x0d layer_norm, layer_norm + vec4 ([©️back👆🏻](#kernellist)) 666 |
667 | 668 | ```c++ 669 | // Layer Norm: x: NxK(K=128<1024), y': NxK, y'=x-mean(x)/std(x) each row 670 | // mean(x) = sum(x)/K, 1/std(x) = rsqrtf( sum( (x-mean(x))^2 )/K ) each row 671 | // grid(N*K/K), block(K<1024) N=batch_size*seq_len, K=hidden_size 672 | // y=y'*g + b (g: scale, b: bias) 673 | template 674 | __global__ void layer_norm(float* x, float* y, float g, float b, int N, int K) { 675 | int tid = threadIdx.x; // 0..K-1 676 | int bid = blockIdx.x; // 0..N-1 677 | int idx = bid * blockDim.x + threadIdx.x; 678 | const float epsilon = 1e-5f; 679 | 680 | __shared__ float s_mean; // shared within block 681 | __shared__ float s_variance; // shared within block 682 | float value = (idx < N * K) ? x[idx] : 0.0f; // load once only 683 | float sum = block_reduce_sum(value); 684 | if (tid == 0) s_mean = sum / (float) K; 685 | // wait for s_mean in shared memory to be ready for all threads 686 | __syncthreads(); 687 | float variance = (value - s_mean) * (value - s_mean); 688 | variance = block_reduce_sum(variance); 689 | if (tid == 0) s_variance = rsqrtf(variance / (float) K + epsilon); 690 | // wait for s_variance in shared memory to be ready for all threads 691 | __syncthreads(); 692 | if (idx < N * K) y[idx] = ((value - s_mean) * s_variance) * g + b; 693 | } 694 | 695 | // Layer Norm Vec4: x: NxK(K=128<1024), y': NxK, y'=x-mean(x)/std(x) each row 696 | // mean(x) = sum(x)/K, 1/std(x) = rsqrtf( sum( (x-mean(x))^2 )/K ) each row 697 | // grid(N*K/K), block(K/4<1024) N=batch_size*seq_len, K=hidden_size 698 | // y=y'*g + b (g: scale, b: bias) 699 | template 700 | __global__ void layer_norm_vec4(float* x, float* y, float g, float b, int N, int K) { 701 | int tid = threadIdx.x; // 0..K-1 702 | int bid = blockIdx.x; // 0..N-1 703 | int idx = (bid * blockDim.x + threadIdx.x) * 4; 704 | const float epsilon = 1e-5f; 705 | 706 | __shared__ float s_mean; // shared within block 707 | __shared__ float s_variance; // shared within block 708 | float4 reg_x = FLOAT4(x[idx]) 709 | float value = (idx < N * K) ? (reg_x.x + reg_x.y 710 | + reg_x.z + reg_x.w) : 0.0f; 711 | float sum = block_reduce_sum(value); 712 | if (tid == 0) s_mean = sum / (float) K; 713 | // wait for s_mean in shared memory to be ready for all threads 714 | __syncthreads(); 715 | float4 reg_x_hat; 716 | reg_x_hat.x = reg_x.x - s_mean; 717 | reg_x_hat.y = reg_x.y - s_mean; 718 | reg_x_hat.z = reg_x.z - s_mean; 719 | reg_x_hat.w = reg_x.w - s_mean; 720 | float variance = reg_x_hat.x * reg_x_hat.x + reg_x_hat.y * reg_x_hat.y 721 | + reg_x_hat.z * reg_x_hat.z + reg_x_hat.w * reg_x_hat.w; 722 | variance = block_reduce_sum(variance); 723 | if (tid == 0) s_variance = rsqrtf(variance / (float) K + epsilon); 724 | // wait for s_variance in shared memory to be ready for all threads 725 | __syncthreads(); 726 | float4 reg_y; 727 | reg_y.x = reg_x_hat.x * s_variance * g + b; 728 | reg_y.y = reg_x_hat.y * s_variance * g + b; 729 | reg_y.z = reg_x_hat.z * s_variance * g + b; 730 | reg_y.w = reg_x_hat.w * s_variance * g + b; 731 | if (idx < N * K) FLOAT4(y[idx]) = reg_y; 732 | } 733 | ``` 734 | layer norm实现的核心同样也是block reduce和warp reduce，然后再整点向量化... 735 | 736 | ## 0x0e rms_norm, rms_norm + vec4 ([©️back👆🏻](#kernellist)) 737 |
738 | 739 | ```c++ 740 | // RMS Norm: x: NxK(K=128<1024), y': NxK, y'=x/rms(x) each row 741 | // 1/rms(x) = rsqrtf( sum(x^2)/K ) each row 742 | // grid(N*K/K), block(K<1024) N=batch_size*seq_len, K=hidden_size 743 | // y=y'*g (g: scale) 744 | template 745 | __global__ void rms_norm(float* x, float* y, float g, int N, int K) { 746 | int tid = threadIdx.x; // 0..K-1 747 | int bid = blockIdx.x; // 0..N-1 748 | int idx = bid * blockDim.x + threadIdx.x; 749 | const float epsilon = 1e-5f; 750 | 751 | __shared__ float s_variance; // shared within block 752 | float value = (idx < N * K) ? x[idx] : 0.0f; // load once only 753 | float variance = value * value; 754 | variance = block_reduce_sum(variance); 755 | if (tid == 0) s_variance = rsqrtf(variance / (float) K + epsilon); 756 | // wait for s_variance in shared memory to be ready for all threads 757 | __syncthreads(); 758 | if (idx < N * K) y[idx] = (value * s_variance) * g; 759 | } 760 | 761 | // RMS Norm Vec4: x: NxK(K=128<1024), y': NxK, y'=x/rms(x) each row 762 | // 1/rms(x) = rsqrtf( sum(x^2)/K ) each row 763 | // grid(N*K/K), block(K/4<1024) N=batch_size*seq_len, K=hidden_size 764 | // y=y'*g (g: scale) 765 | template 766 | __global__ void rms_norm_vec4(float* x, float* y, float g, int N, int K) { 767 | int tid = threadIdx.x; // 0..K-1 768 | int bid = blockIdx.x; // 0..N-1 769 | int idx = (bid * blockDim.x + threadIdx.x) * 4; 770 | const float epsilon = 1e-5f; 771 | 772 | __shared__ float s_variance; // shared within block 773 | float4 reg_x = FLOAT4(x[idx]); 774 | float variance = (idx < N * K) ? (reg_x.x * reg_x.x + reg_x.y * reg_x.y 775 | + reg_x.z * reg_x.z + reg_x.w * reg_x.w) : 0.0f; 776 | variance = block_reduce_sum(variance); 777 | if (tid == 0) s_variance = rsqrtf(variance / (float) K + epsilon); 778 | // wait for s_variance in shared memory to be ready for all threads 779 | __syncthreads(); 780 | float4 reg_y; 781 | reg_y.x = reg_x.x * s_variance * g; 782 | reg_y.y = reg_x.y * s_variance * g; 783 | reg_y.z = reg_x.z * s_variance * g; 784 | reg_y.w = reg_x.w * s_variance * g; 785 | if (idx < N * K) FLOAT4(y[idx]) = reg_y; 786 | } 787 | ``` 788 | rms norm实现的核心同样也是block reduce和warp reduce...，然后再加点float4向量化什么的。 789 | 790 | ## 0x0d NMS ([©️back👆🏻](#kernellist)) 791 |
792 | 793 | ```c++ 794 | struct Box { 795 | float x1, y1, x2, y2, score; 796 | float area() const {return (std::abs(x2 - x1 + 1)) * std::abs(y2 - y1 + 1); } 797 | float iou_of(const Box& other) const{ 798 | float inner_x1 = x1 > other.x1 ? x1 : other.x1; 799 | float inner_y1 = y1 > other.y1 ? y1 : other.y1; 800 | float inner_x2 = x2 < other.x2 ? x2 : other.x2; 801 | float inner_y2 = y2 < other.y2 ? y2 : other.y2; 802 | float inner_h = inner_y2 - inner_y1 + 1.0f; 803 | float inner_w = inner_x2 - inner_x1 + 1.0f; 804 | float inner_area = inner_h * inner_w; 805 | return (inner_area / (area() + tbox.area() - inner_area)); 806 | } 807 | } 808 | void hard_nms(std::vector &input, std::vector &output, float iou_threshold){ 809 | if (input.empty()) return; 810 | std::sort(input.begin(), input.end(),[](Box& a, Box& b) { return a.score > b.score; }); 811 | int box_num = input.size(); 812 | std::vector merged(box_num, 0); 813 | for (int i = 0; i < box_num; ++i) { 814 | if (merged[i]) continue; 815 | merged[i] = 1; 816 | for (int j = i + 1; j < box_num; ++j) { 817 | if (merged[j]) continue; 818 | float iou = input[i].iou_of(input[j]); 819 | if (iou > iou_threshold) merged[j] = 1; 820 | } 821 | output.push_back(input[i]); 822 | } 823 | } 824 | ``` 825 | CV相关的经常会要手撕NMS，也记录下。 826 | 827 | ## 0x0f 总结 ([©️back👆🏻](#kernellist)) 828 | 可以发现，大部分kernel的基本写法都是依赖warp reduce和block reduce的，基本上只要熟练应用warp functions各种场景的写法，应该问题不大；softmax需要考虑网格级同步的问题，或者online softmax以及FlashAttention；sgemm的优化是个很大的课题，不是案例中写的这么简单，但是入门的话，基本就是tiling的思想以及如何做索引之间的mapping；sgemv的优化则主要考虑K不同的值（因为M为1了），比如K=16,64,128等情况下，如何按照warp来处理；relu、sigmoid等都是elementwise的操作，很好实现，可以再考虑加点向量化优化访存；layer norm和rms norm在数学上其实也是挺清晰简单的，落实到cuda kernel时，只要按照逐个token来处理，headdim没有超过1024的情况下（一个block最多可以放1024个threads），可以放到一个block处理，这样并行化就很好写。当然，核心还是warp reduce和block reduce；NMS是乱入的，没有CUDA版本，别问了... 829 | 830 | ## ©️License 831 | GNU General Public License v3.0 832 | 833 | ## 🎉Contribute 834 | 🌟如果觉得有用，不妨给个🌟👆🏻Star支持一下吧~ 835 | -------------------------------------------------------------------------------- /cpp/.gitignore: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/alexngng/CUDA-Learn-Note/87e410107e837407ab1bafb95f71020e6b927d0c/cpp/.gitignore -------------------------------------------------------------------------------- /cpp/how_to_build_cpp_static_lib/README.md: -------------------------------------------------------------------------------- 1 | ## 编译命令 2 | 3 | ```bash 4 | git clone https://github.com/DefTruth/simd-notebook.git 5 | cd cpp/how_to_build_cpp_static_lib/test_staticlib 6 | mkdir build && cd build 7 | cmake .. 8 | make -j 9 | ``` 10 | 11 | -------------------------------------------------------------------------------- /cpp/how_to_build_cpp_static_lib/test_staticlib/.gitignore: -------------------------------------------------------------------------------- 1 | build 2 | test 3 | .DS_Store 4 | -------------------------------------------------------------------------------- /cpp/how_to_build_cpp_static_lib/test_staticlib/CMakeLists.txt: -------------------------------------------------------------------------------- 1 | PROJECT(staticlib_demo C CXX) 2 | CMAKE_MINIMUM_REQUIRED(VERSION 3.12) 3 | 4 | include_directories(${PROJECT_SOURCE_DIR}) 5 | 6 | set(LIB_ADD_SRCS 7 | ${PROJECT_SOURCE_DIR}/liba/func.cc 8 | ${PROJECT_SOURCE_DIR}/liba/funca.cc 9 | ${PROJECT_SOURCE_DIR}/liba/global.cc 10 | ${PROJECT_SOURCE_DIR}/libb/func.cc 11 | ${PROJECT_SOURCE_DIR}/libb/funcb.cc) 12 | 13 | set(LIB_OTHER_ADD_SRCS 14 | ${PROJECT_SOURCE_DIR}/libc/funcc.cc) 15 | 16 | add_library(addfunc_static STATIC ${LIB_ADD_SRCS}) # 编译静态库 17 | add_library(addfunc_shared SHARED ${LIB_ADD_SRCS}) # 编译动态库 18 | add_library(other_addfunc_static_link_static STATIC ${LIB_OTHER_ADD_SRCS}) 19 | add_library(other_addfunc_static_link_shared STATIC ${LIB_OTHER_ADD_SRCS}) 20 | add_library(other_addfunc_shared_link_static SHARED ${LIB_OTHER_ADD_SRCS}) 21 | add_library(other_addfunc_shared_link_shared SHARED ${LIB_OTHER_ADD_SRCS}) 22 | 23 | target_link_libraries(other_addfunc_static_link_static addfunc_static) # 编译静态库链接静态库 24 | target_link_libraries(other_addfunc_static_link_shared addfunc_shared) # 编译静态库链接动态库 25 | target_link_libraries(other_addfunc_shared_link_static addfunc_static) # 编译动态库链接静态库 26 | target_link_libraries(other_addfunc_shared_link_shared addfunc_shared) # 编译动态库链接动态库 27 | 28 | add_executable(test_static ${PROJECT_SOURCE_DIR}/test.cc) 29 | add_executable(test_shared ${PROJECT_SOURCE_DIR}/test.cc) 30 | target_link_libraries(test_static addfunc_static) 31 | set_target_properties(test_static PROPERTIES LINK_FLAGS 32 | "-Wl,--whole-archive libaddfunc_static.a -Wl,-no-whole-archive") 33 | 34 | target_link_libraries(test_shared addfunc_shared) 35 | 36 | add_executable(test_a_inst_static ${PROJECT_SOURCE_DIR}/test_a_inst.cc) 37 | add_executable(test_a_inst_shared ${PROJECT_SOURCE_DIR}/test_a_inst.cc) 38 | target_link_libraries(test_a_inst_static addfunc_static) 39 | target_link_libraries(test_a_inst_shared addfunc_shared) 40 | 41 | -------------------------------------------------------------------------------- /cpp/how_to_build_cpp_static_lib/test_staticlib/liba/func.cc: -------------------------------------------------------------------------------- 1 | #include "liba/func.h" 2 | int AddFuncA(int a, int b) {return a + b;} 3 | -------------------------------------------------------------------------------- /cpp/how_to_build_cpp_static_lib/test_staticlib/liba/func.h: -------------------------------------------------------------------------------- 1 | #pragma once 2 | 3 | int AddFuncA(int a, int b); 4 | -------------------------------------------------------------------------------- /cpp/how_to_build_cpp_static_lib/test_staticlib/liba/funca.cc: -------------------------------------------------------------------------------- 1 | #include "liba/funca.h" 2 | int AddFuncAV2(int a, int b) {return a + b;} 3 | -------------------------------------------------------------------------------- /cpp/how_to_build_cpp_static_lib/test_staticlib/liba/funca.h: -------------------------------------------------------------------------------- 1 | #pragma once 2 | 3 | int AddFuncAV2(int a, int b); 4 | -------------------------------------------------------------------------------- /cpp/how_to_build_cpp_static_lib/test_staticlib/liba/global.cc: -------------------------------------------------------------------------------- 1 | #include 2 | 3 | class ACls { 4 | public: 5 | ACls() { 6 | std::cout << "Create an ACls instance and do some things!" << std::endl; 7 | } 8 | }; 9 | 10 | // create a global instance. 11 | ACls* a_inst = new ACls(); -------------------------------------------------------------------------------- /cpp/how_to_build_cpp_static_lib/test_staticlib/libb/func.cc: -------------------------------------------------------------------------------- 1 | #include "libb/func.h" 2 | float AddFuncB(float a, float b) {return a + b;} 3 | -------------------------------------------------------------------------------- /cpp/how_to_build_cpp_static_lib/test_staticlib/libb/func.h: -------------------------------------------------------------------------------- 1 | #pragma once 2 | 3 | float AddFuncB(float a, float b); 4 | -------------------------------------------------------------------------------- /cpp/how_to_build_cpp_static_lib/test_staticlib/libb/funcb.cc: -------------------------------------------------------------------------------- 1 | #include "libb/funcb.h" 2 | float AddFuncBV2(float a, float b) {return a + b;} 3 | -------------------------------------------------------------------------------- /cpp/how_to_build_cpp_static_lib/test_staticlib/libb/funcb.h: -------------------------------------------------------------------------------- 1 | #pragma once 2 | 3 | float AddFuncBV2(float a, float b); 4 | -------------------------------------------------------------------------------- /cpp/how_to_build_cpp_static_lib/test_staticlib/libc/funcc.cc: -------------------------------------------------------------------------------- 1 | #include "libc/funcc.h" 2 | #include "liba/func.h" 3 | int AddFuncC(int a, int b) {return AddFuncA(a, a) + AddFuncA(b, b);} 4 | -------------------------------------------------------------------------------- /cpp/how_to_build_cpp_static_lib/test_staticlib/libc/funcc.h: -------------------------------------------------------------------------------- 1 | #pragma once 2 | int AddFuncC(int a, int b); 3 | -------------------------------------------------------------------------------- /cpp/how_to_build_cpp_static_lib/test_staticlib/test.cc: -------------------------------------------------------------------------------- 1 | #include "liba/func.h" 2 | #include "libb/func.h" 3 | #include 4 | 5 | int main(int argc, char* argv[]) { 6 | std::cout << "AddFuncA(2, 3): " << AddFuncA(2, 3) << std::endl; 7 | std::cout << "AddFuncB(2.0f, 3.0f): " << AddFuncB(2.0f, 3.0f) << std::endl; 8 | } 9 | -------------------------------------------------------------------------------- /cpp/how_to_build_cpp_static_lib/test_staticlib/test_a_inst.cc: -------------------------------------------------------------------------------- 1 | #include "liba/func.h" 2 | #include "libb/func.h" 3 | #include 4 | 5 | class ACls; 6 | extern ACls* a_inst; 7 | 8 | int main(int argc, char* argv[]) { 9 | std::cout << "a_inst: " << a_inst << std::endl; 10 | std::cout << "AddFuncA(2, 3): " << AddFuncA(2, 3) << std::endl; 11 | std::cout << "AddFuncB(2.0f, 3.0f): " << AddFuncB(2.0f, 3.0f) << std::endl; 12 | } 13 | -------------------------------------------------------------------------------- /cuda-check/README.md: -------------------------------------------------------------------------------- 1 | ## CUDA 高频面试题汇总 2 | 3 | 前段时间参加了一些面试，大部分都要手撕CUDA，因此也整体复习了一遍CUDA优化相关的内容，整理了一些高频题的基本写法，保存在这里也便于日后自己复习，具体见[高频面试题汇总](./check.cu)。当然，有些代码不一定是最优化解，比如GEMM，想要在面试短短的30分钟内写一个好的GEMM Kernel，那实在是太难了，普通人能写个shared memory + block-tile + k-tile 的版本的很不错了。相关kernel如下： 4 | 5 | - sgemm naive, sgemm + block-tile + k-tile + vec4 6 | - sgemv k32/k128/k16 kernel 7 | - warp/block reduce sum/max, block all reduce + vec4 8 | - dot product, dot product + vec4 9 | - elementwise, elementwise + vec4 10 | - histogram, histogram + vec4 11 | - softmax, softmax + vec4 (grid level memory fence) 12 | - safe softmax, safe softmax + vec4 13 | - sigmoid, sigmoid + vec4 14 | - relu, relu + vec4 15 | - layer_norm, layer_norm + vec4 16 | - rms_norm, rms_norm + vec4 17 | - .... 18 | 19 | 不定期更新... -------------------------------------------------------------------------------- /cuda-check/check.cu: -------------------------------------------------------------------------------- 1 | // sgemm 2 | // sgemv 3 | // warp/block reduce 4 | // dot product 5 | // elementwise 6 | // hist 7 | // softmax 8 | // safe softmax 9 | // sigmoid 10 | // relu 11 | // layer_norm 12 | // rms_norm 13 | // iou 14 | // nms 15 | #include 16 | #include 17 | #include 18 | #include 19 | #include 20 | #include 21 | 22 | #define WARP_SIZE 32 23 | #define INT4(value) (reinterpret_cast(&(value))[0]) 24 | #define FLOAT4(value) (reinterpret_cast(&(value))[0]) 25 | 26 | // SGEMM: Block Tile + K Tile, with smem 27 | // Block Tile (BM, BN) + K Tile (BK=32) 28 | // grid((N + BN - 1) / BN, (M + BM - 1) / BM), block(BN, BM) 29 | // a: MxK, b: KxN, c: MxN, compute: c = a * b, all row major 30 | __global__ void sgemm(float* a, float* b, float* c, int M, int N, int K) { 31 | // [1] Block Tile: 32x32的block处理c上一块32x32的元素计算 32 | // [2] K Tile: 使用共享内存，并将K分块为BK大小的块 33 | constexpr int BM = 32; 34 | constexpr int BN = 32; 35 | constexpr int BK = 32; 36 | __shared__ float s_a[BM][BK], s_b[BK][BN]; 37 | 38 | int bx = blockIdx.x; 39 | int by = blockIdx.y; 40 | int tx = threadIdx.x; 41 | int ty = threadIdx.y; 42 | int tid = threadIdx.y * blockDim.x + tx; // tid within the block 43 | // load values to shared memory, 32x32 threads working together 44 | // to fetch data along the row direction of a and b both for s_a 45 | // and s_b 32x32x4x2=8KB, we use 32x32 threads within block to 46 | // load 32x32 elements from global memory to shared memory, namely, 47 | // each thread will load 1 element. 48 | int load_smem_a_m = tid / 32; // 0~31, tid / 32, tid / BM, threadIdx.y 49 | int load_smem_a_k = tid % 32; // 0~31, tid % 32, tid % BK, threadIdx.x 50 | int load_smem_b_k = tid / 32; // 0~31, tid / 32, tid / BK, threadIdx.y 51 | int load_smem_b_n = tid % 32; // 0~31, tid % 32, tid % BN, threadIdx.x 52 | int load_gmem_a_m = by * BM + load_smem_a_m; // global row of a and c 53 | int load_gmem_b_n = bx * BN + load_smem_b_n; // global col of b and c 54 | // if (load_gmem_a_m >= M || load_gmem_b_n >= N) return; 55 | 56 | float sum = 0.f; 57 | for (int bk = 0; bk < (K + BK - 1) / BK; ++bk) { 58 | int load_gmem_a_k = bk * BK + load_smem_a_k; 59 | int load_gmem_a_addr = load_gmem_a_m * K + load_gmem_a_k; 60 | s_a[load_smem_a_m][load_smem_a_k] = a[load_gmem_a_addr]; 61 | int load_gmem_b_k = bk * BK + load_smem_b_k; 62 | int load_gmem_b_addr = load_gmem_b_k * N + load_gmem_b_n; 63 | s_b[load_smem_b_k][load_smem_b_n] = b[load_gmem_b_addr]; 64 | __syncthreads(); 65 | #pragma unroll 66 | for (int k = 0; k < BK; ++k) { 67 | int comp_smem_a_m = load_smem_a_m; 68 | int comp_smem_b_n = load_smem_b_n; 69 | sum += s_a[comp_smem_a_m][k] * s_b[k][comp_smem_b_n]; 70 | } 71 | __syncthreads(); 72 | } 73 | int store_gmem_c_m = load_gmem_a_m; 74 | int store_gmem_c_n = load_gmem_b_n; 75 | int store_gmem_c_addr = store_gmem_c_m * N + store_gmem_c_n; 76 | c[store_gmem_c_addr] = sum; 77 | } 78 | 79 | // SGEMM: Block Tile + Thread Tile + K Tile + Vec4, with smem 80 | // BK:TILE_K=8 BM=BN=128 81 | // TM=TN=8 增加计算密度 BM/TM=16 BN/TN=16 82 | // dim3 blockDim(BN/TN, BM/TM); 83 | // dim3 gridDim((N + BN - 1) / BN, (M + BM - 1) / BM) 84 | __global__ void sgemm_thread_tile_vec4( 85 | float* a, float* b, float* c, int M, int N, int K) { 86 | // [1] Block Tile: 一个16x16的block处理C上大小为128X128的一个目标块 87 | // [2] Thread Tile: 每个thread负责计算TM*TN(8*8)个元素，增加计算密度 88 | // [3] K Tile: 将K分块，每块BK大小，迭代(K+BK-1/BK)次， 89 | // 每次计算TM*TN个元素各自的部分乘累加 90 | // [4] Vectorize: 减少load和store指令，使用float4 91 | constexpr int BM = 128; 92 | constexpr int BN = 128; 93 | constexpr int BK = 8; 94 | constexpr int TM = 8; 95 | constexpr int TN = 8; 96 | 97 | int bx = blockIdx.x; 98 | int by = blockIdx.y; 99 | int tx = threadIdx.x; 100 | int ty = threadIdx.y; 101 | int tid = threadIdx.y * blockDim.x + tx; // tid within the block 102 | __shared__ float s_a[BM][BK], s_b[BK][BN]; // 2*128*8*4=8KB 103 | 104 | // 0. 先计算shared memory中的索引 105 | // tid和需要加载的smem s_a[BM][BK] 之间的索引关系 BM=128 BK=8 按行读取 A行主序 106 | // 对于s_a每行8个数据，每个线程读取4个，需要2个线程；总共128行，需要128x2刚好256线程 107 | int load_smem_a_m = tid / 2; // tid/2 (128/8)*(128/8)=256 threads per block, tid/2->[0,128), BM=128 0~127 108 | int load_smem_a_k = (tid % 2 == 0) ? 0 : 4; // (tid%2 == 0) ? 0 : 4, col of s_a 0,4 109 | // tid和需要加载的smem s_b[BK][BN] 之间的索引关系 BK=8 BN=128 按行读取 B行主序 110 | // 对于s_b每行128个数据，每个线程读4个数据，需要32个线程；总共8行，需要32x8=256个线程 111 | int load_smem_b_k = tid / 32; // tid/32, row of s_b 256/32=8 行 0~7 112 | int load_smem_b_n = (tid % 32) * 4; // (tid % 32) * 4, col of s_b 0,4,...,124 113 | // 1. 再计算全局内存中的索引 114 | // 要加载到s_a中的元素对应到A全局内存中的行数 每个block负责出C中大小为BM*BN的块 115 | int load_gmem_a_m = by * BM + load_smem_a_m; // global row of a and c 116 | int load_gmem_b_n = bx * BN + load_smem_b_n; // global col of b and c 117 | 118 | float r_c[TM][TN] = {0.0}; // 8x8 119 | // 2. 先对K进行分块，每块BK大小 120 | for (int bk = 0; bk < (K + BK - 1) / BK; ++bk) { 121 | // 加载数据到共享内存smem s_a BM*BK 128*8 vectorize float4 122 | int load_gmem_a_k = bk * BK + load_smem_a_k; // global col of a 123 | int load_gmem_a_addr = load_gmem_a_m * K + load_gmem_a_k; 124 | FLOAT4(s_a[load_smem_a_m][load_smem_a_k]) = FLOAT4(a[load_gmem_a_addr]); 125 | // 加载数据到共享内存smem s_b BK*BN 8*128 vectorize float4 126 | int load_gmem_b_k = bk * BK + load_smem_b_k; // global row of b 127 | int load_gmem_b_addr = load_gmem_b_k * N + load_gmem_b_n; 128 | FLOAT4(s_b[load_smem_b_k][load_smem_b_n]) = FLOAT4(b[load_gmem_b_addr]); 129 | __syncthreads(); 130 | #pragma unroll 131 | for (int k = 0; k < BK; k++) { 132 | // 3. 每个线程负责计算BM*BN(12x128)中的TM*TN(8x8)个元素 133 | #pragma unroll 134 | for (int m = 0; m < TM; m++) { 135 | #pragma unroll 136 | for (int n = 0; n < TN; n++) { 137 | // k from 0~7，0 ~ BK, ty and tx range from 0 to 15, 16x8=128 138 | int comp_smem_a_m = ty * TM + m; // 128*8 128/TM(8)=16 M方向 16线程 139 | int comp_smem_b_n = tx * TN + n; // 8*128 128/TN(8)=16 N方向 16线程 140 | r_c[m][n] += s_a[comp_smem_a_m][k] * s_b[k][comp_smem_b_n]; 141 | } 142 | } 143 | } 144 | __syncthreads(); 145 | } 146 | 147 | #pragma unroll 148 | for (int m = 0; m < TM; ++m) { 149 | int store_gmem_c_m = by * BM + ty * TM + m; 150 | #pragma unroll 151 | for (int n = 0; n < TN; n += 4) { 152 | int store_gmem_c_n = bx * BN + tx * TN + n; 153 | int store_gmem_c_addr = store_gmem_c_m * N + store_gmem_c_n; 154 | FLOAT4(c[store_gmem_c_addr]) = FLOAT4(r_c[m][n]); 155 | } 156 | } 157 | } 158 | 159 | // Warp Reduce Sum 160 | template 161 | __device__ __forceinline__ float warp_reduce_sum(float val) { 162 | #pragma unroll 163 | for (int mask = kWarpSize >> 1; mask >= 1; mask >>= 1) { 164 | val += __shfl_xor_sync(0xffffffff, val, mask); 165 | } 166 | return val; 167 | } 168 | 169 | // Warp Reduce Max 170 | template 171 | __device__ __forceinline__ float warp_reduce_max(float val) { 172 | #pragma unroll 173 | for (int mask = kWarpSize >> 1; mask >= 1; mask >>= 1) { 174 | val = fmaxf(val, __shfl_xor_sync(0xffffffff, val, mask)); 175 | } 176 | return val; 177 | } 178 | 179 | // Block reduce sum/max/min device helper for Layer/RMS Norm/Softmax etc. 180 | // grid 1D block 1D, grid(N/128), block(128) 181 | template 182 | __device__ __forceinline__ float block_reduce_sum(float val) { 183 | // always <= 32 warps per block (limited by 1024 threads per block) 184 | constexpr int NUM_WARPS = (NUM_THREADS + WARP_SIZE - 1) / WARP_SIZE; 185 | int warp = threadIdx.x / WARP_SIZE; 186 | int lane = threadIdx.x % WARP_SIZE; 187 | static __shared__ float shared[NUM_WARPS]; 188 | 189 | val = warp_reduce_sum(val); 190 | if (lane == 0) shared[warp] = val; 191 | __syncthreads(); 192 | val = (lane < NUM_WARPS) ? shared[lane] : 0.0f; 193 | val = warp_reduce_sum(val); 194 | return val; 195 | } 196 | 197 | template 198 | __device__ __forceinline__ float block_reduce_max(float val) { 199 | // always <= 32 warps per block (limited by 1024 threads per block) 200 | constexpr int NUM_WARPS = (NUM_THREADS + WARP_SIZE - 1) / WARP_SIZE; 201 | int warp = threadIdx.x / WARP_SIZE; 202 | int lane = threadIdx.x % WARP_SIZE; 203 | static __shared__ float shared[NUM_WARPS]; 204 | 205 | val = warp_reduce_max(val); 206 | if (lane == 0) shared[warp] = val; 207 | __syncthreads(); 208 | val = (lane < NUM_WARPS) ? shared[lane] : -FLT_MAX; 209 | val = warp_reduce_max(val); 210 | return val; 211 | } 212 | 213 | // SGEMV: Warp SGEMV K32 214 | // 假设K为32的倍数，每个warp负责一行 215 | // grid(M/4), block(32,4) blockDim.x=32=K, blockDim.y=4 216 | // a: MxK, x: Kx1, y: Mx1, compute: y = a * x 217 | __global__ void sgemv_k32(float* a, float* x, float* y, int M, int K) { 218 | int tx = threadIdx.x; // 0~31 219 | int ty = threadIdx.y; // 0~4 220 | int bx = blockIdx.x; // 0~M/4 221 | int lane = tx % WARP_SIZE; // 0~31 222 | int m = bx * blockDim.y + ty; // (0~M/4) * 4 + (0~3) 223 | if (m < M) { 224 | float sum = 0.0f; 225 | int NUM_WARPS = (K + WARP_SIZE - 1) / WARP_SIZE; 226 | #pragma unroll 227 | for (int w = 0; w < NUM_WARPS; ++w) { 228 | // 若NUM_WARPS>=2，先将当前行的数据累加到第一个warp中 229 | int k = w * WARP_SIZE + lane; 230 | sum += a[m * K + k] * x[k]; 231 | } 232 | sum = warp_reduce_sum(sum); 233 | if (lane == 0) y[m] = sum; 234 | } 235 | } 236 | 237 | // SGEMV: Warp SGEMV K128 + Vec4 238 | // 假设K为128的倍数 float4 239 | // grid(M/4), block(32,4) blockDim.x=32=K, blockDim.y=4 240 | // a: MxK, x: Kx1, y: Mx1, compute: y = a * x 241 | __global__ void sgemv_k128(float* a, float* x, float* y, int M, int K) { 242 | // 每个线程负责4个元素，一个warp覆盖128个元素 243 | int tx = threadIdx.x; // 0~31 244 | int ty = threadIdx.y; // 0~3 245 | int bx = blockIdx.x; // 0~M/4 246 | int lane = tx % WARP_SIZE; // 0~31 247 | int m = blockDim.y * bx + ty; // (0~M/4) * 4 + (0~3) 248 | 249 | if (m < M) { 250 | float sum = 0.0f; 251 | // process 4*WARP_SIZE elements per warp. 252 | int NUM_WARPS = (((K + WARP_SIZE - 1) / WARP_SIZE) + 4 - 1) / 4; 253 | #pragma unroll 254 | for (int w = 0; w < NUM_WARPS; ++w) { 255 | int k = (w * WARP_SIZE + lane) * 4; 256 | float4 reg_x = FLOAT4(x[k]); 257 | float4 reg_a = FLOAT4(a[m * K + k]); 258 | sum += (reg_a.x * reg_x.x + reg_a.y * reg_x.y 259 | + reg_a.z * reg_x.z + reg_a.w * reg_x.w); 260 | } 261 | sum = warp_reduce_sum(sum); 262 | if(lane == 0) y[m] = sum; 263 | } 264 | } 265 | 266 | // SGEMV: Warp SGEMV K16 267 | // 假设K为16 < 32,每个warp负责2行，每行有16个元素 268 | // NUM_THREADS=128, NUM_WARPS=NUM_THREADS/WARP_SIZE; 269 | // NUM_ROWS=NUM_WARPS * ROW_PER_WARP, grid(M/NUM_ROWS), block(32,NUM_WARPS) 270 | // a: MxK, x: Kx1, y: Mx1, compute: y = a * x 271 | template 272 | __global__ void sgemv_k16(float* A, float* x, float* y, int M, int K) { 273 | constexpr int K_WARP_SIZE = (WARP_SIZE + ROW_PER_WARP - 1) / ROW_PER_WARP; 274 | int tx = threadIdx.x; // 0~31 275 | int ty = threadIdx.y; // 0~NUM_WARPS 276 | int bx = blockIdx.x; // 0~M/NUM_ROWS (NUM_ROWS=NUM_WARPS * ROW_PER_WARP) 277 | int lane = tx % WARP_SIZE; // 0~31 278 | int k = lane % K_WARP_SIZE; // 0~15 279 | // gloabl row of a: MxK and y:Mx1, blockDim.y=NUM_WARPS 280 | int m = (blockDim.y * bx + ty) * ROW_PER_WARP + lane / K_WARP_SIZE; 281 | if (m < M) { 282 | float sum = A[m * K + k] * x[k]; 283 | sum = warp_reduce_sum(sum); 284 | // 注意是k == 0，而不是lane == 0 285 | if(k == 0) y[m] = sum; 286 | } 287 | } 288 | 289 | // Block All Reduce Sum 290 | // grid(N/128), block(128) 291 | // a: Nx1, y=sum(a) 292 | template 293 | __global__ void block_all_reduce_sum(float* a, float* y, int N) { 294 | int tid = threadIdx.x; 295 | int idx = blockIdx.x * NUM_THREADS + tid; 296 | constexpr int NUM_WARPS = (NUM_THREADS + WARP_SIZE - 1) / WARP_SIZE; 297 | __shared__ float reduce_smem[NUM_WARPS]; 298 | // keep the data in register is enougth for warp operaion. 299 | float sum = (idx < N) ? a[idx] : 0.0f; 300 | int warp = tid / WARP_SIZE; 301 | int lane = tid % WARP_SIZE; 302 | // perform warp sync reduce. 303 | sum = warp_reduce_sum(sum); 304 | // warp leaders store the data to shared memory. 305 | if (lane == 0) reduce_smem[warp] = sum; 306 | __syncthreads(); // make sure the data is in shared memory. 307 | // the first warp compute the final sum. 308 | sum = (lane < NUM_WARPS) ? reduce_smem[lane] : 0.0f; 309 | if (warp == 0) sum = warp_reduce_sum(sum); 310 | if (tid == 0) atomicAdd(y, sum); 311 | } 312 | 313 | // Block All Reduce Sum + float4 314 | // grid(N/128), block(128/4) 315 | // a: Nx1, y=sum(a) 316 | template 317 | __global__ void block_all_reduce_sum_vec4(float* a, float* y, int N) { 318 | int tid = threadIdx.x; 319 | int idx = (blockIdx.x * NUM_THREADS + tid) * 4; 320 | constexpr int NUM_WARPS = (NUM_THREADS + WARP_SIZE - 1) / WARP_SIZE; 321 | __shared__ float reduce_smem[NUM_WARPS]; 322 | 323 | float4 reg_a = FLOAT4(a[idx]); 324 | // keep the data in register is enougth for warp operaion. 325 | float sum = (idx < N) ? (reg_a.x + reg_a.y + reg_a.z + reg_a.w) : 0.0f; 326 | int warp = tid / WARP_SIZE; 327 | int lane = tid % WARP_SIZE; 328 | // perform warp sync reduce. 329 | sum = warp_reduce_sum(sum); 330 | // warp leaders store the data to shared memory. 331 | if (lane == 0) reduce_smem[warp] = sum; 332 | __syncthreads(); // make sure the data is in shared memory. 333 | // the first warp compute the final sum. 334 | sum = (lane < NUM_WARPS) ? reduce_smem[lane] : 0.0f; 335 | if (warp == 0) sum = warp_reduce_sum(sum); 336 | if (tid == 0) atomicAdd(y, sum); 337 | } 338 | 339 | // Dot Product 340 | // grid(N/128), block(128) 341 | // a: Nx1, b: Nx1, y=sum(elementwise_mul(a,b)) 342 | template 343 | __global__ void dot(float* a, float* b, float* y, int N) { 344 | int tid = threadIdx.x; 345 | int idx = blockIdx.x * NUM_THREADS + tid; 346 | constexpr int NUM_WARPS = (NUM_THREADS + WARP_SIZE - 1) / WARP_SIZE; 347 | __shared__ float reduce_smem[NUM_WARPS]; 348 | 349 | // keep the data in register is enougth for warp operaion. 350 | float prod = (idx < N) ? a[idx] * b[idx] : 0.0f; 351 | int warp = tid / WARP_SIZE; 352 | int lane = tid % WARP_SIZE; 353 | // perform warp sync reduce. 354 | prod = warp_reduce_sum(prod); 355 | // warp leaders store the data to shared memory. 356 | if (lane == 0) reduce_smem[warp] = prod; 357 | __syncthreads(); // make sure the data is in shared memory. 358 | // the first warp compute the final sum. 359 | prod = (lane < NUM_WARPS) ? reduce_smem[lane] : 0.0f; 360 | if (warp == 0) prod = warp_reduce_sum(prod); 361 | if (tid == 0) atomicAdd(y, prod); 362 | } 363 | 364 | // Dot Product + Vec4 365 | // grid(N/128), block(128/4) 366 | // a: Nx1, b: Nx1, y=sum(elementwise_mul(a,b)) 367 | template 368 | __global__ void dot_vec4(float* a, float* b, float* y, int N) { 369 | int tid = threadIdx.x; 370 | int idx = (blockIdx.x * NUM_THREADS + tid) * 4; 371 | constexpr int NUM_WARPS = (NUM_THREADS + WARP_SIZE - 1) / WARP_SIZE; 372 | __shared__ float reduce_smem[NUM_WARPS]; 373 | 374 | float4 reg_a = FLOAT4(a[idx]); 375 | float4 reg_b = FLOAT4(b[idx]); 376 | float prod = (idx < N) ? (reg_a.x * reg_b.x + reg_a.y * reg_b.y 377 | + reg_a.z * reg_b.z + reg_a.w * reg_b.w) : 0.0f; 378 | int warp = tid / WARP_SIZE; 379 | int lane = tid % WARP_SIZE; 380 | // perform warp sync reduce. 381 | prod = warp_reduce_sum(prod); 382 | // warp leaders store the data to shared memory. 383 | if (lane == 0) reduce_smem[warp] = prod; 384 | __syncthreads(); // make sure the data is in shared memory. 385 | // the first warp compute the final sum. 386 | prod = (lane < NUM_WARPS) ? reduce_smem[lane] : 0.0f; 387 | if (warp == 0) prod = warp_reduce_sum(prod); 388 | if (tid == 0) atomicAdd(y, prod); 389 | } 390 | 391 | // Histogram 392 | // grid(N/128), block(128) 393 | // a: Nx1, y: count histogram 394 | __global__ void histogram(int* a, int* y, int N) { 395 | int idx = blockIdx.x * blockDim.x + threadIdx.x; 396 | if (idx < N) atomicAdd(&(y[a[idx]]), 1); 397 | } 398 | 399 | // Histogram + Vec4 400 | // grid(N/128), block(128/4) 401 | // a: Nx1, y: count histogram 402 | __global__ void histogram_vec4(int* a, int* y, int N) { 403 | int idx = 4 * (blockIdx.x * blockDim.x + threadIdx.x); 404 | if (idx < N) { 405 | int4 reg_a = INT4(a[idx]); 406 | atomicAdd(&(y[reg_a.x]), 1); 407 | atomicAdd(&(y[reg_a.y]), 1); 408 | atomicAdd(&(y[reg_a.z]), 1); 409 | atomicAdd(&(y[reg_a.w]), 1); 410 | } 411 | } 412 | 413 | // ElementWise Add 414 | // grid(N/128), block(128) 415 | // a: Nx1, b: Nx1, c: Nx1, c = elementwise_add(a, b) 416 | __global__ void elementwise_add(float* a, float* b, float* c, int N) { 417 | int idx = blockIdx.x * blockDim.x + threadIdx.x; 418 | if (idx < N) c[idx] = a[idx] + b[idx]; 419 | } 420 | 421 | // ElementWise Add + Vec4 422 | // grid(N/128), block(128/4) 423 | // a: Nx1, b: Nx1, c: Nx1, c = elementwise_add(a, b) 424 | __global__ void elementwise_add_vec4(float* a, float* b, float* c, int N) { 425 | int idx = 4 * (blockIdx.x * blockDim.x + threadIdx.x); 426 | if (idx < N) { 427 | float4 reg_a = FLOAT4(a[idx]); 428 | float4 reg_b = FLOAT4(b[idx]); 429 | float4 reg_c; 430 | reg_c.x = reg_a.x + reg_b.x; 431 | reg_c.y = reg_a.y + reg_b.y; 432 | reg_c.z = reg_a.z + reg_b.z; 433 | reg_c.w = reg_a.w + reg_b.w; 434 | FLOAT4(c[idx]) = reg_c; 435 | } 436 | } 437 | 438 | // Softmax x: N, y: N 439 | // grid(N/128), block(K=128) 440 | template 441 | __global__ void softmax(float* x, float* y, float* total, int N) { 442 | const int tid = threadIdx.x; 443 | const int idx = blockIdx.x * blockDim.x + tid; 444 | constexpr int NUM_WARPS = (NUM_THREADS + WARP_SIZE - 1) / WARP_SIZE; 445 | __shared__ float reduce_smem[NUM_WARPS]; 446 | 447 | float sum = (idx < N) ? expf(x[idx]) : 0.0f; 448 | int warp = tid / WARP_SIZE; 449 | int lane = tid % WARP_SIZE; 450 | sum = warp_reduce_sum(sum); 451 | if (lane == 0) reduce_smem[warp] = sum; 452 | __syncthreads(); 453 | // compute the final sum in each warp 454 | sum = (lane < NUM_WARPS) ? reduce_smem[lane] : 0.0f; 455 | sum = warp_reduce_sum(sum); // sum(e^x_0,...,e^x_n-1) 456 | // get the total sum of all blocks. 457 | if (tid == 0) atomicAdd(total, sum); 458 | __threadfence(); // grid level memory fence 459 | // e^x_i/sum(e^x_0,...,e^x_n-1) 460 | if (idx < N) y[idx] = block_smem[tid] / (*total); 461 | } 462 | 463 | // Softmax x: N, y: N 464 | // grid(N/128), block(K=128) 465 | template 466 | __global__ void softmax_v2(float* x, float* y, float* total, int N) { 467 | const int tid = threadIdx.x; 468 | const int idx = blockIdx.x * blockDim.x + tid; 469 | 470 | float exp_val = (idx < N) ? expf(x[idx]) : 0.0f; 471 | float sum = block_reduce_sum(exp_val); 472 | // get the total sum of all blocks. 473 | if (tid == 0) atomicAdd(total, sum); 474 | __threadfence(); // grid level memory fence 475 | // e^x_i/sum(e^x_0,...,e^x_n-1) 476 | if (idx < N) y[idx] = exp_val / (*total); 477 | } 478 | 479 | // Softmax Vec4 x: N, y: N 480 | // grid(N/128), block(128/4) 481 | template 482 | __global__ void softmax_v2_vec4(float* x, float* y, float* total, int N) { 483 | const int tid = threadIdx.x; 484 | const int idx = (blockIdx.x * blockDim.x + tid) * 4; 485 | 486 | float4 reg_x = FLOAT4(x[idx]); 487 | float4 reg_exp; 488 | reg_exp.x = (idx < N) ? expf(reg_x.x) : 0.0f; 489 | reg_exp.y = (idx < N) ? expf(reg_x.y) : 0.0f; 490 | reg_exp.z = (idx < N) ? expf(reg_x.z) : 0.0f; 491 | reg_exp.w = (idx < N) ? expf(reg_x.w) : 0.0f; 492 | float exp_val = (reg_exp.x + reg_exp.y + reg_exp.z + reg_exp.w); 493 | float sum = block_reduce_sum(exp_val); 494 | // get the total sum of all blocks. 495 | if (tid == 0) atomicAdd(total, sum); 496 | __threadfence(); // grid level memory fence 497 | // e^x_i/sum(e^x_0,...,e^x_n-1) 498 | if (idx < N) { 499 | float4 reg_y; 500 | reg_y.x = reg_exp.x / (*total); 501 | reg_y.y = reg_exp.y / (*total); 502 | reg_y.z = reg_exp.z / (*total); 503 | reg_y.w = reg_exp.w / (*total); 504 | FLOAT4(y[idx]) = reg_y; 505 | } 506 | } 507 | 508 | // Safe Softmax x: N, y: N 509 | // grid(N/128), block(K=128) 510 | template 511 | __global__ void softmax_safe(float* x, float* y, float* total, int N) { 512 | const int tid = threadIdx.x; 513 | const int idx = blockIdx.x * blockDim.x + tid; 514 | 515 | float ori_val = (idx < N) ? x[idx] : -FLT_MAX; 516 | float max_val = block_reduce_max(ori_val); 517 | float exp_val = (idx < N) ? expf(ori_val - max_val) : 0.0f; 518 | float sum = block_reduce_sum(exp_val); 519 | // get the total sum of all blocks. 520 | if (tid == 0) atomicAdd(total, sum); 521 | __threadfence(); // grid level memory fence 522 | // e^x_i/sum(e^x_0,...,e^x_n-1) 523 | if (idx < N) y[idx] = exp_val / (*total); 524 | } 525 | 526 | // Sigmoid x: N, y: N y=1/(1+exp(-x)) 527 | // grid(N/128), block(K=128) 528 | __global__ void sigmoid(float* x, float* y, int N) { 529 | int idx = blockIdx.x * blockDim.x + threadIdx.x; 530 | if (idx < N) y[idx] = 1.0f / (1.0f + expf(-x[idx])); 531 | } 532 | 533 | // Sigmoid x: N, y: N y=1/(1+exp(-x)) Vec4 534 | // grid(N/128), block(128/4) 535 | __global__ void sigmoid_vec4(float* x, float* y, int N) { 536 | int idx = (blockIdx.x * blockDim.x + threadIdx.x) * 4; 537 | if (idx < N) { 538 | float4 reg_x = FLOAT4(x[idx]); 539 | float4 reg_y; 540 | reg_y.x = 1.0f / (1.0f + expf(-reg_x.x)); 541 | reg_y.y = 1.0f / (1.0f + expf(-reg_x.y)); 542 | reg_y.z = 1.0f / (1.0f + expf(-reg_x.z)); 543 | reg_y.w = 1.0f / (1.0f + expf(-reg_x.w)); 544 | FLOAT4(y[idx]) = reg_y; 545 | } 546 | } 547 | 548 | // Relu x: N, y: N y=max(0,x) 549 | // grid(N/128), block(K=128) 550 | __global__ void relu(float* x, float* y, int N) { 551 | int idx = blockIdx.x * blockDim.x + threadIdx.x; 552 | if (idx < N) y[idx] = fmaxf(0.0f, x[idx]); 553 | } 554 | 555 | // Relu x: N, y: N y=max(0,x) Vec4 556 | // grid(N/128/4), block(128/4) 557 | __global__ void relu_vec4(float* x, float* y, int N) { 558 | int idx = (blockIdx.x * blockDim.x + threadIdx.x) * 4; 559 | if (idx < N) { 560 | float4 reg_x = FLOAT4(x[idx]); 561 | float4 reg_y; 562 | reg_y.x = fmaxf(0.0f, reg_x.x); 563 | reg_y.y = fmaxf(0.0f, reg_x.y); 564 | reg_y.z = fmaxf(0.0f, reg_x.z); 565 | reg_y.w = fmaxf(0.0f, reg_x.w); 566 | FLOAT4(y[idx]) = reg_y; 567 | } 568 | } 569 | 570 | // RMS Norm: x: NxK(K=128<1024), y': NxK, y'=x/rms(x) each row 571 | // 1/rms(x) = rsqrtf( sum(x^2)/K ) each row 572 | // grid(N*K/K), block(K<1024) N=batch_size*seq_len, K=hidden_size 573 | // y=y'*g (g: scale) 574 | template 575 | __global__ void rms_norm(float* x, float* y, float g, int N, int K) { 576 | int tid = threadIdx.x; // 0..K-1 577 | int bid = blockIdx.x; // 0..N-1 578 | int idx = bid * blockDim.x + threadIdx.x; 579 | const float epsilon = 1e-5f; 580 | 581 | __shared__ float s_variance; // shared within block 582 | float value = (idx < N * K) ? x[idx] : 0.0f; // load once only 583 | float variance = value * value; 584 | variance = block_reduce_sum(variance); 585 | if (tid == 0) s_variance = rsqrtf(variance / (float) K + epsilon); 586 | // wait for s_variance in shared memory to be ready for all threads 587 | __syncthreads(); 588 | if (idx < N * K) y[idx] = (value * s_variance) * g; 589 | } 590 | 591 | // RMS Norm Vec4: x: NxK(K=128<1024), y': NxK, y'=x/rms(x) each row 592 | // 1/rms(x) = rsqrtf( sum(x^2)/K ) each row 593 | // grid(N*K/K), block(K/4<1024) N=batch_size*seq_len, K=hidden_size 594 | // y=y'*g (g: scale) 595 | template 596 | __global__ void rms_norm_vec4(float* x, float* y, float g, int N, int K) { 597 | int tid = threadIdx.x; // 0..K-1 598 | int bid = blockIdx.x; // 0..N-1 599 | int idx = (bid * blockDim.x + threadIdx.x) * 4; 600 | const float epsilon = 1e-5f; 601 | 602 | __shared__ float s_variance; // shared within block 603 | float4 reg_x = FLOAT4(x[idx]); 604 | float variance = (idx < N * K) ? (reg_x.x * reg_x.x + reg_x.y * reg_x.y 605 | + reg_x.z * reg_x.z + reg_x.w * reg_x.w) : 0.0f; 606 | variance = block_reduce_sum(variance); 607 | if (tid == 0) s_variance = rsqrtf(variance / (float) K + epsilon); 608 | // wait for s_variance in shared memory to be ready for all threads 609 | __syncthreads(); 610 | float4 reg_y; 611 | reg_y.x = reg_x.x * s_variance * g; 612 | reg_y.y = reg_x.y * s_variance * g; 613 | reg_y.z = reg_x.z * s_variance * g; 614 | reg_y.w = reg_x.w * s_variance * g; 615 | if (idx < N * K) FLOAT4(y[idx]) = reg_y; 616 | } 617 | 618 | // Layer Norm: x: NxK(K=128<1024), y': NxK, y'=x-mean(x)/std(x) each row 619 | // mean(x) = sum(x)/K, 1/std(x) = rsqrtf( sum( (x-mean(x))^2 )/K ) each row 620 | // grid(N*K/K), block(K<1024) N=batch_size*seq_len, K=hidden_size 621 | // y=y'*g + b (g: scale, b: bias) 622 | template 623 | __global__ void layer_norm(float* x, float* y, float g, float b, int N, int K) { 624 | int tid = threadIdx.x; // 0..K-1 625 | int bid = blockIdx.x; // 0..N-1 626 | int idx = bid * blockDim.x + threadIdx.x; 627 | const float epsilon = 1e-5f; 628 | 629 | __shared__ float s_mean; // shared within block 630 | __shared__ float s_variance; // shared within block 631 | float value = (idx < N * K) ? x[idx] : 0.0f; // load once only 632 | float sum = block_reduce_sum(value); 633 | if (tid == 0) s_mean = sum / (float) K; 634 | // wait for s_mean in shared memory to be ready for all threads 635 | __syncthreads(); 636 | float variance = (value - s_mean) * (value - s_mean); 637 | variance = block_reduce_sum(variance); 638 | if (tid == 0) s_variance = rsqrtf(variance / (float) K + epsilon); 639 | // wait for s_variance in shared memory to be ready for all threads 640 | __syncthreads(); 641 | if (idx < N * K) y[idx] = ((value - s_mean) * s_variance) * g + b; 642 | } 643 | 644 | // Layer Norm Vec4: x: NxK(K=128<1024), y': NxK, y'=x-mean(x)/std(x) each row 645 | // mean(x) = sum(x)/K, 1/std(x) = rsqrtf( sum( (x-mean(x))^2 )/K ) each row 646 | // grid(N*K/K), block(K/4<1024) N=batch_size*seq_len, K=hidden_size 647 | // y=y'*g + b (g: scale, b: bias) 648 | template 649 | __global__ void layer_norm_vec4(float* x, float* y, float g, float b, int N, int K) { 650 | int tid = threadIdx.x; // 0..K-1 651 | int bid = blockIdx.x; // 0..N-1 652 | int idx = (bid * blockDim.x + threadIdx.x) * 4; 653 | const float epsilon = 1e-5f; 654 | 655 | __shared__ float s_mean; // shared within block 656 | __shared__ float s_variance; // shared within block 657 | float4 reg_x = FLOAT4(x[idx]) 658 | float value = (idx < N * K) ? (reg_x.x + reg_x.y 659 | + reg_x.z + reg_x.w) : 0.0f; 660 | float sum = block_reduce_sum(value); 661 | if (tid == 0) s_mean = sum / (float) K; 662 | // wait for s_mean in shared memory to be ready for all threads 663 | __syncthreads(); 664 | float4 reg_x_hat; 665 | reg_x_hat.x = reg_x.x - s_mean; 666 | reg_x_hat.y = reg_x.y - s_mean; 667 | reg_x_hat.z = reg_x.z - s_mean; 668 | reg_x_hat.w = reg_x.w - s_mean; 669 | float variance = reg_x_hat.x * reg_x_hat.x + reg_x_hat.y * reg_x_hat.y 670 | + reg_x_hat.z * reg_x_hat.z + reg_x_hat.w * reg_x_hat.w; 671 | variance = block_reduce_sum(variance); 672 | if (tid == 0) s_variance = rsqrtf(variance / (float) K + epsilon); 673 | // wait for s_variance in shared memory to be ready for all threads 674 | __syncthreads(); 675 | float4 reg_y; 676 | reg_y.x = reg_x_hat.x * s_variance * g + b; 677 | reg_y.y = reg_x_hat.y * s_variance * g + b; 678 | reg_y.z = reg_x_hat.z * s_variance * g + b; 679 | reg_y.w = reg_x_hat.w * s_variance * g + b; 680 | if (idx < N * K) FLOAT4(y[idx]) = reg_y; 681 | } 682 | -------------------------------------------------------------------------------- /cuda-simd/README.md: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/alexngng/CUDA-Learn-Note/87e410107e837407ab1bafb95f71020e6b927d0c/cuda-simd/README.md -------------------------------------------------------------------------------- /elementwise/README.md: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/alexngng/CUDA-Learn-Note/87e410107e837407ab1bafb95f71020e6b927d0c/elementwise/README.md -------------------------------------------------------------------------------- /gemm/README.md: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/alexngng/CUDA-Learn-Note/87e410107e837407ab1bafb95f71020e6b927d0c/gemm/README.md -------------------------------------------------------------------------------- /gemv/README.md: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/alexngng/CUDA-Learn-Note/87e410107e837407ab1bafb95f71020e6b927d0c/gemv/README.md -------------------------------------------------------------------------------- /nsight-compute/README.md: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/alexngng/CUDA-Learn-Note/87e410107e837407ab1bafb95f71020e6b927d0c/nsight-compute/README.md -------------------------------------------------------------------------------- /reduce/README.md: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/alexngng/CUDA-Learn-Note/87e410107e837407ab1bafb95f71020e6b927d0c/reduce/README.md -------------------------------------------------------------------------------- /spmm/README.md: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/alexngng/CUDA-Learn-Note/87e410107e837407ab1bafb95f71020e6b927d0c/spmm/README.md -------------------------------------------------------------------------------- /spmv/README.md: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/alexngng/CUDA-Learn-Note/87e410107e837407ab1bafb95f71020e6b927d0c/spmv/README.md --------------------------------------------------------------------------------