├── LICENSE.md ├── README ├── main.c ├── makefile ├── mt8173_perf.txt ├── nanopi2_perf.txt ├── orangepi1plus_perf.txt ├── pi2_perf.txt ├── pi3_perf.txt ├── pi4_32_perf.txt ├── pi4_64_perf.txt ├── rk3288_perf.txt ├── sd410_perf.txt ├── test32.s ├── test64.s └── x86_perf.txt /LICENSE.md: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | ========================== 3 | Version 3, 29 June 2007 4 | ========================== 5 | 6 | > Copyright (C) 2007 Free Software Foundation, Inc. 7 | Everyone is permitted to copy and distribute verbatim copies of this license document, but changing it is not allowed. 8 | 9 | # Preamble 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 | 624 | 625 | # How to Apply These Terms to Your New Programs 626 | 627 | If you develop a new program, and you want it to be of the greatest 628 | possible use to the public, the best way to achieve this is to make it 629 | free software which everyone can redistribute and change under these terms. 630 | 631 | To do so, attach the following notices to the program. It is safest 632 | to attach them to the start of each source file to most effectively 633 | state the exclusion of warranty; and each file should have at least 634 | the "copyright" line and a pointer to where the full notice is found. 635 | 636 | 637 | Copyright (C) 638 | 639 | This program is free software: you can redistribute it and/or modify 640 | it under the terms of the GNU General Public License as published by 641 | the Free Software Foundation, either version 3 of the License, or 642 | (at your option) any later version. 643 | 644 | This program is distributed in the hope that it will be useful, 645 | but WITHOUT ANY WARRANTY; without even the implied warranty of 646 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 647 | GNU General Public License for more details. 648 | 649 | You should have received a copy of the GNU General Public License 650 | along with this program. If not, see . 651 | 652 | Also add information on how to contact you by electronic and paper mail. 653 | 654 | If the program does terminal interaction, make it output a short 655 | notice like this when it starts in an interactive mode: 656 | 657 | Copyright (C) 658 | This program comes with ABSOLUTELY NO WARRANTY; for details type 'show w'. 659 | This is free software, and you are welcome to redistribute it 660 | under certain conditions; type 'show c' for details. 661 | 662 | The hypothetical commands _'show w'_ and _'show c'_ should show the appropriate 663 | parts of the General Public License. Of course, your program's commands 664 | might be different; for a GUI interface, you would use an "about box". 665 | 666 | You should also get your employer (if you work as a programmer) or school, 667 | if any, to sign a "copyright disclaimer" for the program, if necessary. 668 | For more information on this, and how to apply and follow the GNU GPL, see 669 | . 670 | 671 | The GNU General Public License does not permit incorporating your program 672 | into proprietary programs. If your program is a subroutine library, you 673 | may consider it more useful to permit linking proprietary applications with 674 | the library. If this is what you want to do, use the GNU Lesser General 675 | Public License instead of this License. But first, please read 676 | . 677 | 678 | -------------------------------------------------------------------------------- /README: -------------------------------------------------------------------------------- 1 | GCC/SIMD/ASM PERFORMANCE TESTS 2 | 3 | Written by Larry Bank 4 | Project started 3/16/2016 5 | 6 | The gcc_perf project is to test the CPU performance (float/int) on Intel+ARM 7 | devices in 32-bit and 64-bit mode. The purpose is to see what the differences 8 | can be attributed to the execution mode and what can be attributed to the 9 | C compiler (GCC). 10 | 11 | Running the tests: 12 | ------------------ 13 | If you run the test program with no parameters, it will run through 14 | every function and test the performance on small and large buffers. The 15 | numbers represent the time in milliseconds to run thousands of iterations 16 | of each test. 17 | 18 | A single numerical parameter will run a single test: 19 | gcc_perf 2 will run test number 2. To get a list of what functions 20 | are available use a parameter of -1. Here is sample output for 21 | one of the functions: 22 | 23 | Multiply Complex C (bigger than cache) = 5070ms 24 | Multiply Complex SIMD (bigger than cache) = 2604ms 25 | Multiply Complex ASM (bigger than cache) = 2554ms 26 | -------------------------------------------------------------------------------- /main.c: -------------------------------------------------------------------------------- 1 | // 2 | // GCC_PERF 3 | // 4 | // A command line tool to test the performance of 32 and 64-bit ARM platforms 5 | // The purpose of this program is to highlight the performance differences across 6 | // different SoC's and the difference between 32 and 64-bit ARM code generated by GCC. 7 | // A set of 9 simple math functions are written in C, SIMD intrinsics and assembly 8 | // language. The compiler is asked to do its best to auto-vectorize and optimize 9 | // the functions. 10 | // Each test is first validated by comparing the output of the C code to the SIMD 11 | // and ASM versions. Once validated, the function is given a large buffer test 12 | // (larger than the L2 cache) and then a small buffer test (fits in L1 cache). 13 | // The 32-bit SIMD and ASM code use prefetch hints to help the "large" test perform 14 | // better. This has the side affect of slightly hurting the performance of the "small" 15 | // tests. 16 | // 17 | // Written by Larry Bank 18 | // 19 | // Project started 3/16/2016 20 | // 21 | // This program is free software: you can redistribute it and/or modify 22 | // it under the terms of the GNU General Public License as published by 23 | // the Free Software Foundation, either version 3 of the License, or 24 | // (at your option) any later version. 25 | // 26 | // This program is distributed in the hope that it will be useful, 27 | // but WITHOUT ANY WARRANTY; without even the implied warranty of 28 | // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 29 | // GNU General Public License for more details. 30 | // 31 | // You should have received a copy of the GNU General Public License 32 | // along with this program. If not, see . 33 | // 34 | #define _POSIX_C_SOURCE 199309L 35 | #include 36 | #include 37 | #include 38 | #include 39 | #include 40 | #include 41 | #include 42 | #include 43 | #ifdef USE_NEON 44 | #include 45 | #endif 46 | #ifdef USE_RISCV 47 | #include 48 | #endif 49 | #ifdef USE_SSE 50 | #include 51 | #include 52 | #include 53 | #include 54 | #endif 55 | typedef int (*PFNTEST)(void *src, void *dest, int iLen); 56 | 57 | typedef struct tagCOMPLEX 58 | { 59 | float r; 60 | float i; 61 | } COMPLEX; 62 | 63 | typedef struct tagTests 64 | { 65 | char *szTestName; 66 | PFNTEST pCFunc; 67 | PFNTEST pSIMDFunc; 68 | PFNTEST pASMFunc; 69 | bool bFloat; 70 | } TESTS; 71 | 72 | enum { 73 | MODE_VALIDATE=0, 74 | MODE_C_PERF, 75 | MODE_SIMD_PERF, 76 | MODE_ASM_PERF 77 | }; 78 | 79 | // Forward declarations 80 | int MilliTime(); 81 | void RunTest(int iTest, int iMode, int iIterations, float *pFloatArray1, float *pFloatArray2, int32_t *pIntArray1, int32_t *pIntArray2, void *pDest, void *pCompare, bool *bPassed, int iLen); 82 | 83 | int c_compare_pixels(void *in, void *out, int iLen); 84 | int simd_compare_pixels(void *in, void *out, int iLen); 85 | int c_update_vertices(void *in, void *out, int iLen); 86 | int simd_update_vertices(void *in, void *out, int iLen); 87 | int c_combine_masks(void *in, void *out, int iLen); 88 | int simd_combine_masks(void *in, void *out, int iLen); 89 | int c_integer_sum(void *in, void *out, int iLen); 90 | int simd_integer_sum(void *in, void *out, int iLen); 91 | int c_float_sum(void *in, void *out, int iLen); 92 | int simd_float_sum(void *in, void *out, int iLen); 93 | int c_integer_diff(void *in, void *out, int iLen); 94 | int simd_integer_diff(void *in, void *out, int iLen); 95 | int c_float_diff(void *in, void *out, int iLen); 96 | int simd_float_diff(void *in, void *out, int iLen); 97 | int c_float_max(void *in, void *out, int iLen); 98 | int simd_float_max(void *in, void *out, int iLen); 99 | int c_integer_max(void *in, void *out, int iLen); 100 | int simd_integer_max(void *in, void *out, int iLen); 101 | int c_float_accumulate(void *in, void *out, int iLen); 102 | int simd_float_accumulate(void *in, void *out, int iLen); 103 | int c_integer_accumulate(void *in, void *out, int iLen); 104 | int simd_integer_accumulate(void *in, void *out, int iLen); 105 | int c_multiply_complex(void *in, void *out, int iLen); 106 | int simd_multiply_complex(void *in, void *out, int iLen); 107 | int simd_16to32(void *in,void *out, int iLen); 108 | int c_16to32(void *in, void *out, int iLen); 109 | int c_cache_test(void *in, void *out, int iLen); 110 | int simd_cache_test(void *in, void *out, int iLen); 111 | int asm_cache_test(void *in, void *out, int iLen); 112 | 113 | // Using intrinsics in x86 land is sufficient. It's hard to 114 | // beat the compiler with hand written asm code for these 115 | // type of functions 116 | #if defined( USE_NEON ) && !defined(NO_ASM) 117 | extern int asm_integer_sum(void *in, void *out, int iLen); 118 | extern int asm_float_sum(void *in, void *out, int iLen); 119 | extern int asm_integer_diff(void *in, void *out, int iLen); 120 | extern int asm_float_diff(void *in, void *out, int iLen); 121 | extern int asm_float_max(void *in, void *out, int iLen); 122 | extern int asm_integer_max(void *in, void *out, int iLen); 123 | extern int asm_float_accumulate(void *in, void *out, int iLen); 124 | extern int asm_integer_accumulate(void *in, void *out, int iLen); 125 | extern int asm_multiply_complex(void *in, void *out, int iLen); 126 | #else 127 | int asm_integer_sum(void *in, void *out, int iLen) { return 0;}; 128 | int asm_float_sum(void *in, void *out, int iLen) { return 0;}; 129 | int asm_integer_diff(void *in, void *out, int iLen) { return 0;}; 130 | int asm_float_diff(void *in, void *out, int iLen) { return 0;}; 131 | int asm_integer_max(void *in, void *out, int iLen) { return 0;}; 132 | int asm_float_max(void *in, void *out, int iLen) { return 0;}; 133 | int asm_integer_accumulate(void *in, void *out, int iLen) { return 0;}; 134 | int asm_float_accumulate(void *in, void *out, int iLen) { return 0;}; 135 | int asm_multiply_complex(void *in, void *out, int iLen) { return 0;}; 136 | #endif // USE_NEON 137 | int c_writebuf_byte(void *in, void *out, int iLen); 138 | int c_writebuf_byte_coalesced(void *in, void *out, int iLen); 139 | int c_writebuf_short(void *in, void *out, int iLen); 140 | int c_writebuf_word(void *in, void *out, int iLen); 141 | int c_writebuf_long(void *in, void *out, int iLen); 142 | 143 | #define TEST_COUNT 19 144 | // List of functions to test 145 | TESTS testList[TEST_COUNT] = { 146 | {"Cache Test", c_cache_test, simd_cache_test, asm_cache_test, false}, 147 | {"Compare Pixels", c_compare_pixels, simd_compare_pixels, NULL, false}, 148 | {"Update Vertices", c_update_vertices, simd_update_vertices, NULL, false}, 149 | {"Write buffer - byte", c_writebuf_byte, NULL, NULL, false}, 150 | {"Write buffer - byte - coalesced", c_writebuf_byte_coalesced, NULL, NULL, false}, 151 | {"Write buffer - short (16-bits)", c_writebuf_short, NULL, NULL, false}, 152 | {"Write buffer - word (32-bits)", c_writebuf_word, NULL, NULL, false}, 153 | {"Write buffer - long (64-bits)", c_writebuf_long, NULL, NULL, false}, 154 | {"Multiply Complex",c_multiply_complex, simd_multiply_complex, asm_multiply_complex, true}, 155 | {"Combine Masks", c_combine_masks, simd_combine_masks, NULL, false}, 156 | {"Integer Sum",c_integer_sum, simd_integer_sum, asm_integer_sum, false}, 157 | {"Float Sum",c_float_sum, simd_float_sum, asm_float_sum, true}, 158 | {"Integer Difference",c_integer_diff, simd_integer_diff, asm_integer_diff, false}, 159 | {"Float Difference",c_float_diff, simd_float_diff, asm_float_diff, true}, 160 | {"Integer Max",c_integer_max, simd_integer_max, asm_integer_max, false}, 161 | {"Float Max", c_float_max, simd_float_max, asm_float_max, true}, 162 | {"Integer Accumulate",c_integer_accumulate, simd_integer_accumulate, asm_integer_accumulate, false}, 163 | {"Float Accumulate", c_float_accumulate, simd_float_accumulate, asm_float_accumulate, true}, 164 | {"RGB565 to RGB8888", c_16to32, simd_16to32, NULL, false} 165 | }; 166 | 167 | /**************************************************************************** 168 | * * 169 | * FUNCTION : main(int, char**) * 170 | * * 171 | ****************************************************************************/ 172 | int main( int argc, char *argv[ ]) 173 | { 174 | int i, iTest, iTime, iLen, iIterations; 175 | int iStart, iEnd; 176 | int iColor; 177 | void *pFloatMem1, *pFloatMem2, *pIntMem1, *pIntMem2, *pDest, *pCompare; 178 | float *pf1, *pf2; 179 | int32_t *pi1, *pi2; 180 | char *szLabel; 181 | char szColor[32]; 182 | char *szUnColor; 183 | bool bPassed; 184 | bool bUseColor; 185 | char *szCPU; 186 | 187 | printf("GCC_PERF tests 1.0\n"); 188 | printf("Written by Larry Bank\n"); 189 | 190 | if (argc == 2) // run a single test 191 | { 192 | sscanf(argv[1], "%d", &iStart); 193 | iEnd = iStart; 194 | if (iStart < 0 || iStart >= TEST_COUNT) 195 | { 196 | // print error message and list the test names and numbers 197 | printf("Requested test is out of range\n"); 198 | for (iTest=0; iTest= 3) 224 | { // no color 225 | bUseColor = false; 226 | strcpy(szColor,""); 227 | szUnColor = ""; 228 | } 229 | else 230 | { 231 | bUseColor = true; 232 | szUnColor = "\033[0m"; 233 | iColor = 32; 234 | } 235 | 236 | iIterations = 200; 237 | #ifdef USE_SSE 238 | iLen = 0x400000; // 16MB on Intel 239 | #else 240 | iLen = 0x100000; // 4MB on ARM should be enough to not fit in L2 cache 241 | #endif 242 | pFloatMem1 = (void *)malloc(iLen * sizeof(float) + 16); 243 | pFloatMem2 = (void *)malloc(iLen * sizeof(float) + 16); 244 | pIntMem1 = (void *)malloc(iLen * sizeof(int32_t) + 16); 245 | pIntMem2 = (void *)malloc(iLen * sizeof(int32_t) + 16); 246 | pDest = (void *)malloc(iLen * sizeof(float) + 16); 247 | pCompare = (void *)malloc(iLen * sizeof(float) + 16); 248 | 249 | // Prepare some reasonable test data 250 | pf1 = (float *)pFloatMem1; 251 | pi1 = (int32_t *)pIntMem1; 252 | pf2 = (float *)pFloatMem2; 253 | pi2 = (int32_t *)pIntMem2; 254 | for (i=0; i 33) iColor = 32; // toggle between green and yellow for passing tests 325 | } 326 | 327 | free(pFloatMem1); 328 | free(pFloatMem2); 329 | free(pIntMem1); 330 | free(pIntMem2); 331 | free(pDest); 332 | free(pCompare); 333 | return 0; 334 | 335 | } /* main() */ 336 | 337 | /**************************************************************************** 338 | * * 339 | * FUNCTION : MilliTime() * 340 | * returns the relative time in milliseconds * 341 | * * 342 | ****************************************************************************/ 343 | int MilliTime() 344 | { 345 | int iTime; 346 | struct timespec res; 347 | 348 | clock_gettime(CLOCK_MONOTONIC, &res); 349 | iTime = 1000*res.tv_sec + res.tv_nsec/1000000; 350 | 351 | return iTime; 352 | } /* MilliTime() */ 353 | 354 | /************************************************************************************** 355 | * * 356 | * FUNCTION : RunTest(int, int, int, float *, float *, int32_t *, int32_t *, * 357 | * void *, bool **, int) * 358 | * Runs the given test N times. * 359 | * * 360 | **************************************************************************************/ 361 | void RunTest(int iTest, int iMode, int iIterations, float *pFloatArray1, float *pFloatArray2, int32_t *pIntArray1, int32_t *pIntArray2, void *pDest, void *pCompare, bool *bPassed, int iLen) 362 | { 363 | int i = 0; 364 | int iSize; 365 | bool bSIMD, bASM; 366 | void *pSrc1, *pSrc2; 367 | 368 | pSrc1 = (testList[iTest].bFloat) ? (void *)pFloatArray1 : (void *)pIntArray1; 369 | pSrc2 = (testList[iTest].bFloat) ? (void *)pFloatArray2 : (void *)pIntArray2; 370 | 371 | if (iMode == MODE_VALIDATE) 372 | { 373 | *bPassed = bSIMD = bASM = false; 374 | 375 | memcpy(pDest, pSrc2, iLen * sizeof(int32_t)); 376 | iSize = (*testList[iTest].pCFunc)(pSrc1, pDest, iLen); 377 | memcpy(pCompare, pSrc2, iLen * sizeof(int32_t)); 378 | (*testList[iTest].pSIMDFunc)(pSrc1, pCompare, iLen); 379 | if (memcmp(pDest, pCompare, iSize*sizeof(int32_t)) == 0) 380 | { 381 | bSIMD = true; 382 | } 383 | else 384 | { 385 | printf("SIMD failed to compare\n"); 386 | } 387 | #ifdef USE_NEON 388 | memcpy(pCompare, pSrc2, iLen * sizeof(int32_t)); 389 | (*testList[iTest].pASMFunc)(pSrc1, pCompare, iLen); 390 | if (memcmp(pDest, pCompare, iSize*sizeof(int32_t)) == 0) 391 | { 392 | bASM = true; 393 | } 394 | else 395 | { 396 | printf("ASM failed to compare\n"); 397 | } 398 | #else 399 | bASM = true; 400 | #endif // USE_NEON 401 | if (bASM && bSIMD) 402 | { 403 | *bPassed = true; 404 | } 405 | } 406 | else if (iMode == MODE_C_PERF) 407 | { 408 | for (i=0; i> 1); // green 457 | u32 |= (us & 0xf800) >> 8; 458 | u32 |= (us & 0xe000) >> 11; 459 | u32 |= 0xff000000; // alpha 460 | *d++ = u32; 461 | } // for x 462 | } // for y 463 | } 464 | return iLen; 465 | 466 | } /* c_16to32() */ 467 | 468 | int simd_16to32(void *source, void *dest, int iLen) 469 | { 470 | // convert inputs into width/height/pitch info 471 | int iSrcPitch, iDestPitch, iWidth, iHeight; 472 | unsigned char *pSrc, *pDest; 473 | pSrc = (unsigned char *)source; 474 | pDest = (unsigned char *)dest; 475 | iWidth = 128; 476 | iHeight = iLen / 128; 477 | iSrcPitch = iWidth*sizeof(uint16_t); 478 | iDestPitch = iWidth*sizeof(uint32_t); 479 | #ifdef USE_SSE 480 | { 481 | __m128i xmmIn, xmmInNext, xmmOut0, xmmOut1; 482 | __m128i xmmMul5, xmmMul6, xmmRMask, xmmGMask, xmmFF00, xmm00FF; 483 | __m128i xmmR, xmmG, xmmT1, xmmB, xmmT2; 484 | int x, y; 485 | unsigned char *s, *d; 486 | xmmMul5 = _mm_set1_epi32(0x01080108); 487 | xmmMul6 = _mm_set1_epi32(0x20802080); 488 | xmmRMask = _mm_set1_epi32(0xf800f800); 489 | xmmGMask = _mm_set1_epi32(0x07e007e0); 490 | xmmFF00 = _mm_set1_epi32(0xff00ff00); // alpha value 491 | xmm00FF = _mm_set1_epi32(0x00ff00ff); // mask 492 | for (y=0; y= 64) // do 4 writes of 16-bytes at a time 581 | { 582 | _mm_storeu_si128((__m128i*)d, xmm_zero); 583 | _mm_storeu_si128((__m128i*)&d[16], xmm_zero); 584 | _mm_storeu_si128((__m128i*)&d[32], xmm_zero); 585 | _mm_storeu_si128((__m128i*)&d[48], xmm_zero); 586 | d += 64; 587 | n -= 64; 588 | } // while 589 | } 590 | #endif // USE_SSE 591 | #ifdef USE_NEON 592 | { 593 | uint8x16_t xmm_zero = vdupq_n_u8(0); 594 | while (n >= 64) // do 4 writes of 16-bytes at a time 595 | { 596 | vst1q_u8(d, xmm_zero); 597 | vst1q_u8(&d[16], xmm_zero); 598 | vst1q_u8(&d[32], xmm_zero); 599 | vst1q_u8(&d[48], xmm_zero); 600 | d += 64; 601 | n -= 64; 602 | } // while 603 | } 604 | #endif // USE_NEON 605 | // finish stragglers 606 | while (n) 607 | { 608 | *d++ = 0; 609 | n--; 610 | } 611 | return 0; 612 | } /* c_cache_test() */ 613 | 614 | // 615 | // In this version, we'll purposely align the writes on cache boundaries 616 | // 617 | int simd_cache_test(void *in, void *out, int iLen) 618 | { 619 | unsigned char *s = (unsigned char *)in; 620 | unsigned char *d = (unsigned char *)out; 621 | int n = iLen*4; 622 | 623 | if (s) {}; // suppress warning 624 | while (((long)d & 0xf) != 0) // make sure it's 16-byte aligned 625 | { 626 | *d++ = 0; 627 | n--; 628 | } 629 | #ifdef USE_SSE 630 | { 631 | __m128i xmm_zero = _mm_setzero_si128(); 632 | while (n >= 64) // do 4 writes of 16-bytes at a time 633 | { 634 | _mm_storeu_si128((__m128i*)d, xmm_zero); 635 | _mm_storeu_si128((__m128i*)&d[16], xmm_zero); 636 | _mm_storeu_si128((__m128i*)&d[32], xmm_zero); 637 | _mm_storeu_si128((__m128i*)&d[48], xmm_zero); 638 | d += 64; 639 | n -= 64; 640 | } // while 641 | } 642 | #endif // USE_SSE 643 | #ifdef USE_NEON 644 | { 645 | uint8x16_t xmm_zero = vdupq_n_u8(0); 646 | while (n >= 64) // do 4 writes of 16-bytes at a time 647 | { 648 | vst1q_u8(d, xmm_zero); 649 | vst1q_u8(&d[16], xmm_zero); 650 | vst1q_u8(&d[32], xmm_zero); 651 | vst1q_u8(&d[48], xmm_zero); 652 | d += 64; 653 | n -= 64; 654 | } // while 655 | } 656 | #endif // USE_NEON 657 | 658 | // finish stragglers 659 | while (n) 660 | { 661 | *d++ = 0; 662 | n--; 663 | } 664 | return 0; 665 | } /* simd_cache_test() */ 666 | 667 | // 668 | // For this version, we'll use the non-temporal version of the write 669 | // instruction to not pollute the cache 670 | // 671 | int asm_cache_test(void *in, void *out, int iLen) 672 | { 673 | unsigned char *s = (unsigned char *)in; 674 | unsigned char *d = (unsigned char *)out; 675 | int n = iLen*4; 676 | 677 | if (s) {}; // suppress warning 678 | while (((long)d & 0xf) != 0) // make sure it's 16-byte aligned 679 | { 680 | *d++ = 0; 681 | n--; 682 | } 683 | #ifdef USE_SSE 684 | { 685 | __m128i xmm_zero = _mm_setzero_si128(); 686 | while (n >= 64) // do 4 writes of 16-bytes at a time 687 | { 688 | _mm_stream_si128((__m128i*)d, xmm_zero); 689 | _mm_stream_si128((__m128i*)&d[16], xmm_zero); 690 | _mm_stream_si128((__m128i*)&d[32], xmm_zero); 691 | _mm_stream_si128((__m128i*)&d[48], xmm_zero); 692 | d += 64; 693 | n -= 64; 694 | } // while 695 | } 696 | #endif // USE_SSE 697 | #ifdef USE_NEON 698 | { 699 | uint8x16_t xmm_zero = vdupq_n_u8(0); 700 | while (n >= 64) // do 4 writes of 16-bytes at a time 701 | { 702 | vst1q_u8(d, xmm_zero); 703 | vst1q_u8(&d[16], xmm_zero); 704 | vst1q_u8(&d[32], xmm_zero); 705 | vst1q_u8(&d[48], xmm_zero); 706 | d += 64; 707 | n -= 64; 708 | } // while 709 | } 710 | #endif // USE_NEON 711 | // finish stragglers 712 | while (n) 713 | { 714 | *d++ = 0; 715 | n--; 716 | } 717 | return 0; 718 | } /* asm_cache_test() */ 719 | 720 | void initVertexTest(void *in, int iLen) 721 | { 722 | uint32_t *indices; 723 | float *weights, *inputP, *inputN; 724 | int indexCount, i, j; 725 | float fStart = 1.0f; 726 | 727 | indexCount = iLen / 8; // 6/8 float values, 1/8 indices, 1/8 weights 728 | weights = (float *)in; 729 | indices = (uint32_t *)&weights[indexCount]; 730 | inputP = &weights[indexCount*2]; 731 | inputN = &weights[indexCount*5]; 732 | for (i=0, j=0; i 0; iCount -= vl, x+=vl) // compare as many pixelsas possible per iteration 945 | { 946 | vl = __riscv_vsetvl_e8m8(iCount); 947 | vIn0 = __riscv_vle8_v_u8m8(&s8[x], vl); 948 | vIn1 = __riscv_vle8_v_u8m8(&s8[x+iPitch], vl); 949 | vOut0 = __riscv_vle8_v_u8m8(&d8[x], vl); 950 | vOut1 = __riscv_vle8_v_u8m8(&d8[x+iPitch], vl); 951 | not_equal0 = __riscv_vmsne_vv_u8m8_b1(vIn0, vOut0, vl); 952 | not_equal1 = __riscv_vmsne_vv_u8m8_b1(vIn1, vOut1, vl); 953 | not_equal0 = __riscv_vmor_mm_b1(not_equal0, not_equal1, vl); 954 | if (__riscv_vfirst_m_b1(not_equal0, vl)) // non-match 955 | { 956 | u32RowBits |= (1 << i); 957 | iTotalChanged++; 958 | y = dy; iCount = 0; // continue to next tile 959 | } 960 | } // for x 961 | #endif // USE_RISCV 962 | #ifdef USE_SSE 963 | __m128i xmmIn0, xmmIn1, xmmIn2, xmmIn3; 964 | __m128i xmmOut0, xmmOut1, xmmOut2, xmmOut3; 965 | __m128i xmmSum0, xmmSum1; 966 | for (x = 0; x < dx/2; x+=8) // compare pairs of pixels 967 | { // test 2 lines of 8 RGB565 pixels at a time 968 | xmmIn0 = _mm_loadu_si128((__m128i*)&s[x]); 969 | xmmIn1 = _mm_loadu_si128((__m128i*)&s[x+4]); 970 | xmmIn2 = _mm_loadu_si128((__m128i*)&s[x+iPitch4]); 971 | xmmIn3 = _mm_loadu_si128((__m128i*)&s[x+4+iPitch4]); 972 | xmmOut0 = _mm_loadu_si128((__m128i*)&d[x]); 973 | xmmOut1 = _mm_loadu_si128((__m128i*)&d[x+4]); 974 | xmmOut2 = _mm_loadu_si128((__m128i*)&d[x+iPitch4]); 975 | xmmOut3 = _mm_loadu_si128((__m128i*)&d[x+4+iPitch4]); 976 | xmmSum0 = _mm_sad_epu8(xmmIn0, xmmOut0); 977 | xmmSum1 = _mm_sad_epu8(xmmIn1, xmmOut1); 978 | xmmSum0 = _mm_add_epi32(xmmSum0, _mm_sad_epu8(xmmIn2, xmmOut2)); 979 | xmmSum1 = _mm_add_epi32(xmmSum1, _mm_sad_epu8(xmmIn3, xmmOut3)); 980 | xmmSum0 = _mm_add_epi32(xmmSum0, xmmSum1); 981 | xmmSum1 = _mm_cmpeq_epi32(xmmSum0, _mm_setzero_si128()); 982 | if (_mm_movemask_epi8(xmmSum1) != 0xffff) // non-match 983 | { 984 | u32RowBits |= (1 << i); 985 | iTotalChanged++; 986 | y = dy; x = dx; // continue to next tile 987 | } 988 | } // for x 989 | #endif // USE_SSE 990 | #ifdef USE_NEON 991 | uint16x8_t xmmIn0, xmmIn1, xmmIn2, xmmIn3; 992 | uint16x8_t xmmOut0, xmmOut1, xmmOut2, xmmOut3; 993 | uint16x8_t xmmSum0, xmmSum1; 994 | uint16x4_t xmmHalf; 995 | xmmSum0 = xmmSum1 = vdupq_n_u16(0); 996 | for (x = 0; x < dx/2; x+=8) // compare pairs of pixels 997 | { // test 2 lines of 8 RGB565 pixels at a time 998 | xmmIn0 = vld1q_u16((uint16_t*)&s[x]); 999 | xmmIn1 = vld1q_u16((uint16_t*)&s[x+4]); 1000 | xmmIn2 = vld1q_u16((uint16_t*)&s[x+iPitch4]); 1001 | xmmIn3 = vld1q_u16((uint16_t*)&s[x+4+iPitch4]); 1002 | xmmOut0 = vld1q_u16((uint16_t*)&d[x]); 1003 | xmmOut1 = vld1q_u16((uint16_t*)&d[x+4]); 1004 | xmmOut2 = vld1q_u16((uint16_t*)&d[x+iPitch4]); 1005 | xmmOut3 = vld1q_u16((uint16_t*)&d[x+4+iPitch4]); 1006 | xmmSum0 = vabaq_u16(xmmSum0, xmmIn0, xmmOut0); 1007 | xmmSum1 = vabaq_u16(xmmSum1, xmmIn1, xmmOut1); 1008 | xmmSum0 = vabaq_u16(xmmSum0, xmmIn2, xmmOut2); 1009 | xmmSum1 = vabaq_u16(xmmSum1, xmmIn3, xmmOut3); 1010 | xmmSum0 = vaddq_u16(xmmSum0, xmmSum1); 1011 | xmmHalf = vadd_u16(vget_low_u16(xmmSum0), vget_high_u16(xmmSum0)); 1012 | xmmHalf = vpadd_u16(xmmHalf, xmmHalf); // 4 to 2 1013 | if (vget_lane_u32(vreinterpret_u32_u16(xmmHalf), 0) != 0) // non-match 1014 | { 1015 | u32RowBits |= (1 << i); 1016 | iTotalChanged++; 1017 | y = dy; x = dx; // continue to next tile 1018 | } 1019 | } // for x 1020 | #endif // USE_NEON 1021 | s += (iPitch/2); // skip 2 lines 1022 | d += (iPitch/2); 1023 | } // for y 1024 | } // for each tile 1025 | return iTotalChanged; 1026 | } /* simd_compare_pixels() */ 1027 | 1028 | int c_writebuf_byte(void *in, void *out, int iLen) 1029 | { 1030 | int i; 1031 | uint8_t c, *s, *d; 1032 | 1033 | s = (uint8_t *)in; 1034 | d = (uint8_t *)out; 1035 | for (i=0; i max) max = s[i]; 1688 | } 1689 | d[0] = max; // store the max value as element 0 in the output 1690 | return 1; 1691 | } /* c_float_max() */ 1692 | 1693 | // Find the maximum value in an array 1694 | int simd_float_max(void *in, void *out, int iLen) 1695 | { 1696 | int i; 1697 | float *s = (float *)in; 1698 | float *d = (float *)out; 1699 | 1700 | #ifdef USE_NEON 1701 | { 1702 | float32x4_t xmmCur0, xmmCur1, xmmCur2, xmmCur3; 1703 | float32x4_t xmmMax0, xmmMax1, xmmMax2, xmmMax3; 1704 | 1705 | xmmMax0 = vld1q_f32(&s[0]); 1706 | xmmMax1 = vld1q_f32(&s[4]); 1707 | xmmMax2 = vld1q_f32(&s[8]); 1708 | xmmMax3 = vld1q_f32(&s[12]); 1709 | for (i=16; i< iLen-15; i+=16) 1710 | { 1711 | xmmCur0 = vld1q_f32(&s[i]); 1712 | xmmCur1 = vld1q_f32(&s[i+4]); 1713 | xmmCur2 = vld1q_f32(&s[i+8]); 1714 | xmmCur3 = vld1q_f32(&s[i+12]); 1715 | __builtin_prefetch(&s[i+96]); 1716 | xmmMax0 = vmaxq_f32(xmmCur0, xmmMax0); // hold on to the max 4 values 1717 | xmmMax1 = vmaxq_f32(xmmCur1, xmmMax1); 1718 | xmmMax2 = vmaxq_f32(xmmCur2, xmmMax2); 1719 | xmmMax3 = vmaxq_f32(xmmCur3, xmmMax3); 1720 | } 1721 | xmmMax0 = vmaxq_f32(xmmMax0, xmmMax1); 1722 | xmmMax0 = vmaxq_f32(xmmMax0, xmmMax2); 1723 | xmmMax0 = vmaxq_f32(xmmMax0, xmmMax3); 1724 | // now we have the 4 max values, reduce to 1 1725 | xmmCur0 = vextq_f32(xmmMax0, xmmMax0, 2); // compare upper and lower 2 values 1726 | xmmMax0 = vmaxq_f32(xmmCur0, xmmMax0); // keep max 2 1727 | xmmCur0 = vextq_f32(xmmMax0, xmmMax0, 1); // compare last 2 values 1728 | xmmMax0 = vmaxq_f32(xmmCur0, xmmMax0); // keep max 1 1729 | d[0] = vgetq_lane_f32(xmmMax0, 0); 1730 | } 1731 | #endif // USE_NEON 1732 | #ifdef USE_SSE 1733 | { 1734 | __m128 xmmCur0, xmmCur1, xmmCur2, xmmCur3; 1735 | __m128 xmmMax0, xmmMax1, xmmMax2, xmmMax3; 1736 | float kludge[4]; 1737 | 1738 | xmmMax0 = _mm_loadu_ps(&s[0]); 1739 | xmmMax1 = _mm_loadu_ps(&s[4]); 1740 | xmmMax2 = _mm_loadu_ps(&s[8]); 1741 | xmmMax3 = _mm_loadu_ps(&s[12]); 1742 | for (i=16; i< iLen-15; i+=16) 1743 | { 1744 | xmmCur0 = _mm_loadu_ps(&s[i]); 1745 | xmmCur1 = _mm_loadu_ps(&s[i+4]); 1746 | xmmCur2 = _mm_loadu_ps(&s[i+8]); 1747 | xmmCur3 = _mm_loadu_ps(&s[i+12]); 1748 | xmmMax0 = _mm_max_ps(xmmCur0, xmmMax0); // hold on to the max 4 values 1749 | xmmMax1 = _mm_max_ps(xmmCur1, xmmMax1); 1750 | xmmMax2 = _mm_max_ps(xmmCur2, xmmMax2); 1751 | xmmMax3 = _mm_max_ps(xmmCur3, xmmMax3); 1752 | } 1753 | xmmMax0 = _mm_max_ps(xmmMax0, xmmMax1); 1754 | xmmMax0 = _mm_max_ps(xmmMax0, xmmMax2); 1755 | xmmMax0 = _mm_max_ps(xmmMax0, xmmMax3); 1756 | // now we have the 4 max values, reduce to 1 1757 | xmmCur0 = (__m128)_mm_srli_si128((__m128i)xmmMax0, 8); // compare upper and lower 2 values 1758 | xmmMax0 = _mm_max_ps(xmmCur0, xmmMax0); // keep max 2 1759 | xmmCur0 = (__m128)_mm_srli_si128((__m128i)xmmMax0, 4); // compare last 2 values 1760 | xmmMax0 = _mm_max_ps(xmmCur0, xmmMax0); // keep max 1 1761 | _mm_storeu_ps(&kludge[0], xmmMax0); // no float extract in SSE/SSE2/SSE4 1762 | d[0] = kludge[0]; 1763 | } 1764 | #endif // USE_SSE 1765 | 1766 | return 1; 1767 | } /* simd_float_max() */ 1768 | 1769 | // Find the maximum value in an array 1770 | int c_integer_max(void *in, void *out, int iLen) 1771 | { 1772 | int i; 1773 | int32_t *s = (int32_t *)in; 1774 | int32_t *d = (int32_t *)out; 1775 | int32_t max; 1776 | 1777 | max = -10000; 1778 | for (i=0; i max) max = s[i]; 1781 | } 1782 | d[0] = max; // store the max value as element 0 in the output 1783 | return 1; 1784 | } /* c_integer_max() */ 1785 | 1786 | // Find the maximum value in an array 1787 | int simd_integer_max(void *in, void *out, int iLen) 1788 | { 1789 | int i; 1790 | int32_t *s = (int32_t *)in; 1791 | int32_t *d = (int32_t *)out; 1792 | 1793 | #ifdef USE_NEON 1794 | { 1795 | int32x4_t xmmCur0, xmmCur1, xmmCur2, xmmCur3; 1796 | int32x4_t xmmMax0, xmmMax1, xmmMax2, xmmMax3; 1797 | 1798 | xmmMax0 = vld1q_s32(&s[0]); 1799 | xmmMax1 = vld1q_s32(&s[4]); 1800 | xmmMax2 = vld1q_s32(&s[8]); 1801 | xmmMax3 = vld1q_s32(&s[12]); 1802 | for (i=16; i< iLen-15; i+=16) 1803 | { 1804 | xmmCur0 = vld1q_s32(&s[i]); 1805 | xmmCur1 = vld1q_s32(&s[i+4]); 1806 | xmmCur2 = vld1q_s32(&s[i+8]); 1807 | xmmCur3 = vld1q_s32(&s[i+12]); 1808 | __builtin_prefetch(&s[i+96]); 1809 | xmmMax0 = vmaxq_s32(xmmCur0, xmmMax0); // hold on to the max 4 values 1810 | xmmMax1 = vmaxq_s32(xmmCur1, xmmMax1); 1811 | xmmMax2 = vmaxq_s32(xmmCur2, xmmMax2); 1812 | xmmMax3 = vmaxq_s32(xmmCur3, xmmMax3); 1813 | } 1814 | xmmMax0 = vmaxq_s32(xmmMax0, xmmMax1); 1815 | xmmMax0 = vmaxq_s32(xmmMax0, xmmMax2); 1816 | xmmMax0 = vmaxq_s32(xmmMax0, xmmMax3); 1817 | // now we have the 4 max values, reduce to 1 1818 | xmmCur0 = vextq_s32(xmmMax0, xmmMax0, 2); // compare upper and lower 2 values 1819 | xmmMax0 = vmaxq_s32(xmmCur0, xmmMax0); // keep max 2 1820 | xmmCur0 = vextq_s32(xmmMax0, xmmMax0, 1); // compare last 2 values 1821 | xmmMax0 = vmaxq_s32(xmmCur0, xmmMax0); // keep max 1 1822 | d[0] = vgetq_lane_s32(xmmMax0, 0); 1823 | } 1824 | #endif // USE_NEON 1825 | #ifdef USE_SSE 1826 | { 1827 | __m128i xmmCur0, xmmCur1, xmmCur2, xmmCur3; 1828 | __m128i xmmMax0, xmmMax1, xmmMax2, xmmMax3; 1829 | 1830 | xmmMax0 = _mm_loadu_si128((__m128i *)&s[0]); 1831 | xmmMax1 = _mm_loadu_si128((__m128i *)&s[4]); 1832 | xmmMax2 = _mm_loadu_si128((__m128i *)&s[8]); 1833 | xmmMax3 = _mm_loadu_si128((__m128i *)&s[12]); 1834 | for (i=16; i< iLen-15; i+=16) 1835 | { 1836 | xmmCur0 = _mm_loadu_si128((__m128i *)&s[i]); 1837 | xmmCur1 = _mm_loadu_si128((__m128i *)&s[i+4]); 1838 | xmmCur2 = _mm_loadu_si128((__m128i *)&s[i+8]); 1839 | xmmCur3 = _mm_loadu_si128((__m128i *)&s[i+12]); 1840 | xmmMax0 = _mm_max_epi32(xmmCur0, xmmMax0); // hold on to the max 4 values 1841 | xmmMax1 = _mm_max_epi32(xmmCur1, xmmMax1); 1842 | xmmMax2 = _mm_max_epi32(xmmCur2, xmmMax2); 1843 | xmmMax3 = _mm_max_epi32(xmmCur3, xmmMax3); 1844 | } 1845 | xmmMax0 = _mm_max_epi32(xmmMax0, xmmMax1); 1846 | xmmMax0 = _mm_max_epi32(xmmMax0, xmmMax2); 1847 | xmmMax0 = _mm_max_epi32(xmmMax0, xmmMax3); 1848 | // now we have the 4 max values, reduce to 1 1849 | xmmCur0 = _mm_srli_si128(xmmMax0, 8); // compare upper and lower 2 values 1850 | xmmMax0 = _mm_max_epi32(xmmCur0, xmmMax0); // keep max 2 1851 | xmmCur0 = _mm_srli_si128(xmmMax0, 4); // compare last 2 values 1852 | xmmMax0 = _mm_max_epi32(xmmCur0, xmmMax0); // keep max 1 1853 | d[0] = _mm_cvtsi128_si32(xmmMax0); 1854 | } 1855 | #endif // USE_SSE 1856 | 1857 | return 1; 1858 | } /* simd_integer_max() */ 1859 | 1860 | // Calculate the sum of all values in an array 1861 | int c_float_accumulate(void *in, void *out, int iLen) 1862 | { 1863 | int i; 1864 | float *s = (float *)in; 1865 | float *d = (float *)out; 1866 | float sum = 0.0; 1867 | 1868 | for (i=0; i