├── LICENSE ├── README.md └── src ├── ann.asm ├── build.sh ├── clean.sh ├── fops.asm ├── linux.asm ├── matrix.asm ├── mse.asm ├── sigmoid.asm ├── test_ann.asm ├── tostring.asm └── vectors.asm /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 | # AsmANN (Assembly language Artificial Neural Network) 2 | 3 | This repository contains different experiments with building an Artificial Neural Network in x86 64-bit Assembly language. 4 | 5 | Currently it contains only a minimalist ANN. 6 | 7 | ## Building 8 | 9 | The code is intended for a Linux operating system using nasm. OS-specific code is in [src/linux.asm](src/linux.asm) . If you want to compile this on another operating system, you need to change this file. I have a video on my Youtube channel on how to implement these functions on Windows. 10 | 11 | To compile the code, you need *nasm* and *ld*. There is a [src/build.sh](src/build.sh) file that can be used to automate the build. You should edit it to set the currect paths to your nasm and ld. If you don't have nasm on your OS, I have a video about compiling it from source without root access. 12 | 13 | ## Running 14 | 15 | After compiling, you will get the executable file *test_ann*. You should execute this file to get the network to train. I have a video about it here: https://youtu.be/AYuyN8vvkAM 16 | 17 | ## Other resources 18 | 19 | Checkout my assembly language playlist on Youtube: https://www.youtube.com/playlist?list=PL7-u-wmV6bWqmPScAQzyfqLT05xKvUUEC 20 | And my artificial intelligence playlist: https://www.youtube.com/playlist?list=PL7-u-wmV6bWopRa1Qchpq9s0PPJyn2MKl 21 | 22 | -------------------------------------------------------------------------------- /src/ann.asm: -------------------------------------------------------------------------------- 1 | %use masm 2 | section .text 3 | global layerForward 4 | global outputLoss 5 | global layerBackward 6 | 7 | extern vecMulMatLines 8 | extern vecSigmoid 9 | extern vecSigmoid_d 10 | extern vecMSE 11 | extern vecMulVecOuter 12 | extern vecMulScalar 13 | extern vecMulVecHadamard 14 | extern vecDiff 15 | extern vecConsoleWrite 16 | extern matConsoleWrite 17 | 18 | ; Forward pass of a single layer 19 | ; Input: 20 | ; rdi = input vector 21 | ; rsi = Weights matrix 22 | ; rcx = vector size = line size (number of columns) 23 | ; rdx = number of lines 24 | ; r8 = output vector 25 | ; Output: 26 | ; memory [r8 ... r8+rcx] = result = sigmoid(W*x) 27 | layerForward: 28 | PUSH rdi 29 | PUSH rsi 30 | 31 | CALL vecMulMatLines 32 | MOV rdi, r8 33 | MOV rsi, r8 34 | XCHG rcx, rdx 35 | CALL vecSigmoid 36 | XCHG rcx, rdx 37 | 38 | POP rsi 39 | POP rdi 40 | RET 41 | 42 | ; Computes the loss using the Mean Squared Error (MSE) 43 | ; Input: 44 | ; rdi = output 45 | ; rsi = expected 46 | ; rcx = vector size 47 | ; Output: 48 | ; eax = result = MSE loss 49 | outputLoss: 50 | JMP vecMSE 51 | 52 | ; Updates the Weights matrix using backpropagation 53 | ; Input: 54 | ; rdi = output 55 | ; rsi = expected 56 | ; rcx = output size 57 | ; eax = eta (learning rate) 58 | ; r9 = input 59 | ; rdx = input size 60 | ; r10 = weights (size = rcx lines, r10 columns) 61 | layerBackward: 62 | PUSH rdi 63 | PUSH rcx 64 | PUSH rdx 65 | PUSH r8 66 | PUSH rax 67 | PUSH rsi 68 | 69 | PUSH rsi 70 | MOV rsi, tmpOut 71 | CALL vecSigmoid_d ; tmpOut=SigmoidDeriv(output) 72 | POP rsi 73 | 74 | CALL vecDiff ; output=output-expected 75 | 76 | PUSH rsi 77 | MOV rsi, tmpOut 78 | CALL vecMulVecHadamard ; output=(out-expected)*SigmoidDeriv() 79 | POP rsi 80 | 81 | CALL vecMulScalar ; output=eta*(out-expected)*SigmoidDeriv() 82 | 83 | MOV rsi, rdi 84 | MOV rdi, r9 85 | XCHG rcx,rdx 86 | MOV r8, tmpW 87 | ; rsi = output 88 | CALL vecMulVecOuter ; tmpW=weights adjustment 89 | 90 | MOV rax, rdx 91 | MUL rcx ; rax = cols*lines 92 | MOV rcx,rax 93 | MOV rdi, r10 94 | MOV rsi, tmpW 95 | CALL vecDiff ; weights get updated 96 | 97 | POP rsi 98 | POP rax 99 | POP r8 100 | POP rdx 101 | POP rcx 102 | POP rdi 103 | RET 104 | 105 | debugTmpOut: 106 | PUSH rdi 107 | PUSH rcx 108 | MOV rdi, tmpOut 109 | MOV rcx, 2 110 | call vecConsoleWrite 111 | POP rcx 112 | POP rdi 113 | RET 114 | 115 | debugTmpW: 116 | PUSH rdi 117 | PUSH rcx 118 | PUSH rdx 119 | MOV rdi, tmpW 120 | MOV rcx, 3 121 | MOV rdx, 2 122 | call matConsoleWrite 123 | POP rdx 124 | POP rcx 125 | POP rdi 126 | RET 127 | 128 | section .data 129 | 130 | tmpOut: DD 100 dup(0) ; Temporary space to store a copy of the output vector 131 | 132 | tmpW: DD 100 dup(0) ; Temporary space to store weight changes 133 | -------------------------------------------------------------------------------- /src/build.sh: -------------------------------------------------------------------------------- 1 | #!/bin/sh 2 | 3 | AS=/home/user/programs/bin/nasm 4 | LINK=ld 5 | 6 | $AS -f elf64 -O0 -g -l linux.lst -o linux.o linux.asm 7 | $AS -f elf64 -O0 -g -l tostring.lst -o tostring.o tostring.asm 8 | $AS -f elf64 -O0 -g -l vectors.lst -o vectors.o vectors.asm 9 | $AS -f elf64 -O0 -g -l fops.lst -o fops.o fops.asm 10 | $AS -f elf64 -O0 -g -l sigmoid.lst -o sigmoid.o sigmoid.asm 11 | $AS -f elf64 -O0 -g -l matrix.lst -o matrix.o matrix.asm 12 | $AS -f elf64 -O0 -g -l mse.lst -o mse.o mse.asm 13 | $AS -f elf64 -O0 -g -l ann.lst -o ann.o ann.asm 14 | $AS -f elf64 -O0 -g -l test_ann.lst -o test_ann.o test_ann.asm 15 | 16 | $LINK -b elf64-x86-64 \ 17 | vectors.o linux.o tostring.o matrix.o fops.o sigmoid.o mse.o ann.o test_ann.o \ 18 | -o test_ann --entry main -z stack-size=1000 19 | -------------------------------------------------------------------------------- /src/clean.sh: -------------------------------------------------------------------------------- 1 | #!/bin/sh 2 | 3 | rm -f *.o 4 | rm -f *.lst 5 | rm -f test_ann 6 | -------------------------------------------------------------------------------- /src/fops.asm: -------------------------------------------------------------------------------- 1 | %use masm 2 | section .text 3 | 4 | global f32_square 5 | global f32_sqrt 6 | global f32_exp 7 | global f32_pow 8 | 9 | ; Compute the square value of a fp32 number 10 | ; Input: 11 | ; rdi = address of input number X 12 | ; rsi = address of output number 13 | ; Output: 14 | ; DWORD PTR [rsi] = X*X 15 | f32_square: 16 | FLD DWORD PTR [rdi] 17 | FMUL st0,st0 18 | FSTP DWORD PTR [rsi] 19 | RET 20 | 21 | ; Compute the square root of a fp32 number 22 | ; Input: 23 | ; rdi = address of input number X 24 | ; rsi = address of output number 25 | ; Output: 26 | ; DWORD PTR [rsi] = sqrt(X) 27 | f32_sqrt: 28 | FLD DWORD PTR [rdi] 29 | FSQRT 30 | FSTP DWORD PTR [rsi] 31 | RET 32 | 33 | ; Compute the e^X where X is an fp32 number 34 | ; Input: 35 | ; rdi = address of input number X 36 | ; rsi = address of output number 37 | ; Output: 38 | ; DWORD PTR [rsi] = e^X = 2^(X*log2(e)) 39 | ; https://stackoverflow.com/questions/48713712/calculating-expx-in-x86-assembly 40 | ;f32_exp: 41 | ; FLD DWORD PTR [rdi] ;st0 = x 42 | ; FLDL2E ;st0 = log2(e), st1 = x 43 | ; FMULP st1,st0 ;st0 = x*log2(e) 44 | ; FLD1 ;st0 = 1, st1=x*log2(e) 45 | ; FSCALE ;st0 = 2^int(x*log2(e)), st1=x*log2(e) 46 | ; FXCH ;st0 = x*log2(e), st1=2^int(x*log2(e)) 47 | ; FLD1 ;st0 = 1, st1=x*log2(e), st2=2^int(x*log2(e)) 48 | ; FXCH ;st0 = x*log2(e), st1=1, st2=2^int(x*log2(e)) 49 | ; FPREM ;st0 = fract(x*log2(e)), st1=1, st2=2^int(x*log2(e)) 50 | ; F2XM1 ;st0 = 2^(fract(x*log2(e))) - 1, st1=1, st2=2^int(x*log2(e)) 51 | ; FADDP st1,st0 ;st0 = 2^(fract(x*log2(e))), st1 = 2^int(x*log2(e)) 52 | ; FMULP st1,st0 ;st0 = 2^(int(x*log2(e)) + fract(x*log2(e))) = 2^(x*log2(e)) 53 | ; FSTP DWORD PTR [rsi] 54 | ; RET 55 | 56 | ; Compute the X^Y where X,Y are fp32 numbers 57 | ; Input: 58 | ; rdi = address of input number X 59 | ; r8 = address of input number Y 60 | ; rsi = address of output number 61 | ; Output: 62 | ; DWORD PTR [rsi] = X^Y = 2^(Y*log2(X)) 63 | ; https://www.madwizard.org/programming/snippets?id=36 64 | ; https://stackoverflow.com/questions/4638473/how-to-powreal-real-in-x86 65 | f32_pow: 66 | FLD DWORD PTR [r8] ;st0=Y 67 | FLD DWORD PTR [rdi] ;st0=X, st1=Y 68 | FYL2X ;st0=Y*log2(X) 69 | FLD1 ;st0=1, st1=Y*log2(X) 70 | FLD st1 ;st0=Y*log2(X), st1=1, st2=Y*log2(X) 71 | FPREM ;st0=fract(Y*log2(X)), st1=1, st2=Y*log2(X) 72 | F2XM1 ;st0=2^fract(Y*log2(X))-1, st1=1, st2=Y*log2(X) 73 | FADDP ;st0=2^fract(Y*log2(X)), st1=Y*log2(X) 74 | FSCALE ;st0=2^(fract(Y*log2(X))+int(Y*log2(X)))=2^(Y*log2(X)), st1=Y*log2(X) 75 | FXCH st1 ;st0=Y*log2(X), st1=2^(Y*log2(X)) 76 | FSTP st0 ;st0=2^(Y*log2(X)) 77 | FSTP DWORD PTR [rsi] 78 | RET 79 | 80 | ; Compute the e^X where X is an fp32 number 81 | ; Input: 82 | ; rdi = address of input number X 83 | ; rsi = address of output number 84 | ; Output: 85 | ; DWORD PTR [rsi] = e^X = 2^(X*log2(e)) 86 | ; - based on the pow function above 87 | f32_exp: 88 | FLD DWORD PTR [rdi] ;st0 = x 89 | FLDL2E ;st0 = log2(e), st1 = x 90 | FMULP st1,st0 ;st0 = x*log2(e) 91 | FLD1 ;st0=1, st1=x*log2(e) 92 | FLD st1 ;st0=x*log2(e), st1=1, st2=x*log2(e) 93 | FPREM ;st0=fract(x*log2(e)), st1=1, st2=x*log2(e) 94 | F2XM1 ;st0=2^fract(x*log2(e))-1, st1=1, st2=x*log2(e) 95 | FADDP ;st0=2^fract(x*log2(e)), st1=x*log2(e) 96 | FSCALE ;st0=2^(fract(x*log2(e))+int(x*log2(e)))=2^(x*log2(e)), st1=x*log2(e) 97 | FXCH st1 ;st0=x*log2(e), st1=2^(x*log2(e)) 98 | FSTP st0 ;st0=2^(x*log2(e)) 99 | FSTP DWORD PTR [rsi] 100 | RET 101 | -------------------------------------------------------------------------------- /src/linux.asm: -------------------------------------------------------------------------------- 1 | %use masm 2 | section .text 3 | global osInit 4 | global osConsoleWrite 5 | global osExit 6 | 7 | ; Initialization of OS related functionality 8 | ; Input: None 9 | ; Output: None 10 | osInit: 11 | RET 12 | 13 | ; Write a message to console 14 | ; Input: 15 | ; RSI - Message 16 | ; RDX - Length 17 | osConsoleWrite: 18 | PUSH rdi 19 | PUSH rax 20 | PUSH rbx 21 | PUSH rcx 22 | PUSH rdx 23 | 24 | MOV rdi, 1 ; file handle 1=STDOUT 25 | MOV rax, 1 ; syscall 1=write 26 | SYSCALL 27 | 28 | POP rdx 29 | POP rcx 30 | POP rbx 31 | POP rax 32 | POP rdi 33 | RET 34 | 35 | ; Exits back to the operating system 36 | osExit: 37 | XOR rdi,rdi ; exit code 38 | MOV rax,60 ; syscall 60=exit 39 | SYSCALL 40 | RET ; this will never execute, after the syscall the program is ended 41 | -------------------------------------------------------------------------------- /src/matrix.asm: -------------------------------------------------------------------------------- 1 | %use masm 2 | section .text 3 | global matInit 4 | global matConsoleWrite 5 | 6 | extern osConsoleWrite 7 | extern float32ToString 8 | 9 | ; Initialize matrix with float32 number 10 | ; Input: 11 | ; eax = float32 number 12 | ; rcx = line size (number of columns) 13 | ; rdx = column size (number of lines) 14 | ; rdi = matrix 15 | matInit: 16 | PUSH rcx 17 | PUSH rdi 18 | PUSH rdx 19 | 20 | PUSH rax 21 | MOV rax,rcx 22 | MUL rdx 23 | MOV rcx,rax ; rcx=rcx*rdx 24 | POP rax 25 | 26 | REP STOSD 27 | 28 | POP rdx 29 | POP rdi 30 | POP rcx 31 | RET 32 | 33 | 34 | ; Write a matrix to console 35 | ; Input: 36 | ; rdi = matrix 37 | ; rcx = line size (number of columns) 38 | ; rdx = column size (number of lines) 39 | matConsoleWrite: 40 | PUSH rax 41 | PUSH rcx 42 | PUSH rsi 43 | PUSH rdi 44 | PUSH r8 45 | PUSH rdx 46 | 47 | _mcw_loop: 48 | PUSH rcx ; Will be decremented in the loop 49 | PUSH rdx ; Will be changed for osConsoleWrite 50 | 51 | _mcw_loop_disp_line: 52 | MOV eax, DWORD PTR [rdi] 53 | MOV r8, matfp32Buff 54 | CALL float32ToString 55 | 56 | MOV rsi, matfp32Buff 57 | MOV rdx, matfp32BuffSz 58 | CALL osConsoleWrite 59 | 60 | ADD rdi,4 61 | LOOP _mcw_loop_disp_line ; End loop for display line 62 | 63 | MOV rsi, matnl 64 | MOV rdx, matnlSz 65 | CALL osConsoleWrite 66 | 67 | POP rdx ; Restore rcx to be line size 68 | POP rcx ; Restore rdx to current line 69 | DEC rdx 70 | JZ _mcw_done 71 | JMP _mcw_loop 72 | 73 | _mcw_done: 74 | POP rdx 75 | POP r8 76 | POP rdi 77 | POP rsi 78 | POP rcx 79 | POP rax 80 | RET 81 | 82 | section .data 83 | 84 | matfp32Buff: DB "XXXXXXXXXXXXXXXXXXXXX " 85 | matfp32BuffSz EQU $-matfp32Buff 86 | 87 | matnl: DB 13,10 88 | matnlSz EQU $-matnl 89 | 90 | vec_tmp: DD 0 91 | -------------------------------------------------------------------------------- /src/mse.asm: -------------------------------------------------------------------------------- 1 | %use masm 2 | section .text 3 | global vecMSE 4 | 5 | ; Compute the Mean Square Error (MSE) between two vectors 6 | ; Input: 7 | ; rdi = vector1 8 | ; rsi = vector2 9 | ; rcx = vector size 10 | ; Output: 11 | ; eax = result 12 | vecMSE: 13 | PUSH rcx 14 | PUSH rdi 15 | PUSH rsi 16 | 17 | MOV DWORD PTR [tmp], ecx 18 | FILD DWORD PTR [tmp] ; st0=N 19 | 20 | FLDZ ; st0=0, st1=N 21 | 22 | _vmse_loop: 23 | FLD DWORD PTR [rdi] ; st0=Ai, st1=0, st2=N 24 | FSUB DWORD PTR [rsi] ; st0=Ai-Bi, st1=0, st2=N 25 | FMUL st0,st0 ; st0=(Ai-Bi)^2, st1=0, st2=N 26 | FADDP ; st0=sum((Ai-Bi)^2), st1=N 27 | 28 | ADD rdi,4 29 | ADD rsi,4 30 | LOOP _vmse_loop 31 | 32 | ; st0=sum((Ai-Bi)^2), st1=N 33 | FXCH ; st0=N, st1=sum((Ai-Bi)^2) 34 | FDIVP ; st0=sum((Ai-Bi)^2) / N 35 | 36 | FSTP DWORD PTR [tmp] 37 | MOV eax, DWORD PTR [tmp] 38 | 39 | POP rsi 40 | POP rdi 41 | POP rcx 42 | RET 43 | 44 | section .data 45 | 46 | tmp: DD 0 47 | -------------------------------------------------------------------------------- /src/sigmoid.asm: -------------------------------------------------------------------------------- 1 | %use masm 2 | section .text 3 | 4 | global f32_sigmoid 5 | global f32_sigmoid_d 6 | 7 | ; Compute the sigmoid(X) where X is an fp32 number 8 | ; Input: 9 | ; rdi = address of input number X 10 | ; rsi = address of output number 11 | ; Output: 12 | ; DWORD PTR [rsi] = sigmoid(X)=e^X/(1+e^X) 13 | ; - based on the f32_exp function in fops.asm 14 | f32_sigmoid: 15 | FLD DWORD PTR [rdi] ;st0 = x 16 | 17 | FLDL2E ;st0 = log2(e), st1 = x 18 | FMULP st1,st0 ;st0 = x*log2(e) 19 | FLD1 ;st0=1, st1=x*log2(e) 20 | FLD st1 ;st0=x*log2(e), st1=1, st2=x*log2(e) 21 | FPREM ;st0=fract(x*log2(e)), st1=1, st2=x*log2(e) 22 | F2XM1 ;st0=2^fract(x*log2(e))-1, st1=1, st2=x*log2(e) 23 | FADDP ;st0=2^fract(x*log2(e)), st1=x*log2(e) 24 | FSCALE ;st0=2^(fract(x*log2(e))+int(x*log2(e)))=2^(x*log2(e))=e^x, st1=x*log2(e) 25 | FXCH st1 ;st0=x*log2(e), st1=e^x 26 | FSTP st0 ;st0=e^x 27 | 28 | FLD1 ;st0=1, st1=e^x 29 | FLD st1 ;st0=e^x, st1=1, st2=e^x 30 | FADDP ;st0=1+e^x, st1=e^x 31 | FDIVP st1,st0 ;st0=e^x/(1+e^x) 32 | 33 | FSTP DWORD PTR [rsi] 34 | RET 35 | 36 | ; Compute the derivative sigmoid(X) where X is an fp32 number 37 | ; Input: 38 | ; rdi = address of input sigmoid(X); this must already be computed with f32_sigmoid 39 | ; rsi = address of output sigmoid_d(X) 40 | ; Output: 41 | ; DWORD PTR [rsi] = sigmoid_d(X)=sigmoid(X)(1-sigmoid(X)) 42 | ; 43 | f32_sigmoid_d: 44 | FLD DWORD PTR [rdi] ;st0 = s(x) 45 | 46 | FLD1 ;st0=1, st1=s(x) 47 | FLD st1 ;st0=s(x), st1=1, st2=s(x) 48 | FCHS ;st0=-s(x), st1=1, st2=s(x) 49 | FADDP ;st0=1-s(x), st1=s(x) 50 | FMULP ;st0=s(x)(1-s(x)) 51 | 52 | FSTP DWORD PTR [rsi] 53 | RET 54 | -------------------------------------------------------------------------------- /src/test_ann.asm: -------------------------------------------------------------------------------- 1 | %use masm 2 | section .text 3 | global main 4 | extern osInit 5 | extern osConsoleWrite 6 | extern osExit 7 | extern vecInit 8 | extern vecConsoleWrite 9 | extern vecMulScalar 10 | extern vecMulVecDot 11 | extern vecMulVecHadamard 12 | extern float32ToString 13 | extern vecMulMatLines 14 | extern vecMulMatColumns 15 | extern matInit 16 | extern matConsoleWrite 17 | extern vecMSE 18 | extern layerForward 19 | extern outputLoss 20 | extern layerBackward 21 | extern uint32ToString 22 | 23 | main: ; Main entry point. Will run tests for the vector functions 24 | 25 | CALL osInit 26 | 27 | ; Print all input/expected output data 28 | MOV rsi, msgInput 29 | MOV rdx, msgInputSz 30 | CALL osConsoleWrite 31 | MOV rdi, in_data 32 | MOV rcx, in_cols 33 | MOV rdx, in_rows 34 | CALL matConsoleWrite 35 | 36 | MOV rsi, msgExpectedOutput 37 | MOV rdx, msgExpectedOutputSz 38 | CALL osConsoleWrite 39 | MOV rdi, expected_data 40 | MOV rcx, expected_cols 41 | MOV rdx, expected_rows 42 | CALL matConsoleWrite 43 | 44 | MOV rbx,0 ; Epoch number 45 | 46 | ; EPOCH START 47 | epoch_start: 48 | ; Display current epoch number 49 | MOV rsi, msgEpoch 50 | MOV rdx, msgEpochSz 51 | CALL osConsoleWrite 52 | MOV rax, rbx 53 | MOV r8, uint32buff 54 | CALL uint32ToString 55 | MOV rsi, uint32buff 56 | MOV rdx, uint32buffSz 57 | CALL osConsoleWrite 58 | 59 | data_start: 60 | XOR r14,r14 ; data index 61 | MOV r12, in_data 62 | MOV r13, expected_data 63 | 64 | data_loop: 65 | ; Display input 66 | MOV rsi, msgInput 67 | MOV rdx, msgInputSz 68 | CALL osConsoleWrite 69 | MOV rdi, r12 70 | MOV rcx, in_cols 71 | CALL vecConsoleWrite 72 | 73 | ; Display expected output 74 | MOV rsi, msgExpectedOutput 75 | MOV rdx, msgExpectedOutputSz 76 | CALL osConsoleWrite 77 | MOV rdi, r13 78 | MOV rcx, expected_cols 79 | CALL vecConsoleWrite 80 | 81 | ; Run the forward step 82 | MOV rdi, r12 83 | MOV rsi, weights_data 84 | MOV rcx, in_cols 85 | MOV rdx, out_size 86 | MOV r8, out_data 87 | CALL layerForward 88 | 89 | ; Display predicted output 90 | MOV rsi, msgOutput 91 | MOV rdx, msgOutputSz 92 | CALL osConsoleWrite 93 | MOV rdi, out_data 94 | MOV rcx, out_size 95 | CALL vecConsoleWrite 96 | 97 | ; Compute loss 98 | MOV rdi, out_data 99 | MOV rsi, r13 100 | MOV rcx, out_size 101 | CALL outputLoss 102 | 103 | ; Display loss 104 | MOV rsi, msgLoss 105 | MOV rdx, msgLossSz 106 | CALL osConsoleWrite 107 | MOV r8, f32buff 108 | CALL float32ToString 109 | MOV rsi, f32buff 110 | MOV rdx, f32buffSz 111 | CALL osConsoleWrite 112 | 113 | MOV rdi, out_data 114 | MOV rsi, r13 115 | MOV rcx, out_size 116 | MOV eax, __?float32?__(0.1) 117 | MOV r9, r12 118 | MOV rdx, in_cols 119 | MOV r10, weights_data 120 | call layerBackward 121 | 122 | ; Display weights 123 | MOV rsi, msgWeights 124 | MOV rdx, msgWeightsSz 125 | CALL osConsoleWrite 126 | MOV rdi, weights_data 127 | MOV rcx, weights_cols 128 | MOV rdx, weights_rows 129 | CALL matConsoleWrite 130 | 131 | ADD r12, in_line_size 132 | ADD r13, expected_line_size 133 | INC r14 134 | CMP r14, in_rows 135 | JNZ data_loop 136 | 137 | INC rbx 138 | CMP rbx, 1000 ; number of epochs 139 | JNZ epoch_start 140 | 141 | CALL osExit 142 | RET ; this will never be executed 143 | 144 | section .data 145 | 146 | msgInput: DB 13,10,"Input: ",13,10 147 | msgInputSz EQU $-msgInput 148 | 149 | msgExpectedOutput: DB 13,10,"Expected Output: ",13,10 150 | msgExpectedOutputSz EQU $-msgExpectedOutput 151 | 152 | msgOutput: DB 13,10,"Output: ",13,10 153 | msgOutputSz EQU $-msgOutput 154 | 155 | msgLoss: DB 13,10,"Loss: " 156 | msgLossSz EQU $-msgLoss 157 | 158 | msgWeights: DB 13,10,"Weights: ",13,10 159 | msgWeightsSz EQU $-msgWeights 160 | 161 | msgEpoch: DB 13,10,"Epoch: " 162 | msgEpochSz EQU $-msgEpoch 163 | 164 | in_data: 165 | ; X1 X2 | BIAS 166 | DD 0.0, 0.0 , 1.0 167 | DD 0.0, 1.0 , 1.0 168 | DD 1.0, 0.0 , 1.0 169 | DD 1.0, 1.0 , 1.0 170 | in_size EQU ($-in_data)/4 171 | in_rows EQU 4 172 | in_cols EQU 3 173 | in_line_size EQU in_cols*4 174 | 175 | expected_data: 176 | ; 0 1 177 | DD 1.0, 0.0 178 | DD 0.0, 1.0 179 | DD 0.0, 1.0 180 | DD 0.0, 1.0 181 | expected_size EQU ($-expected_data)/4 182 | expected_rows EQU 4 183 | expected_cols EQU 2 184 | expected_line_size EQU expected_cols*4 185 | 186 | out_data: DD expected_cols dup(0) 187 | out_size EQU ($-out_data)/4 188 | 189 | weights_data: 190 | DD 6 dup(0) 191 | weights_size EQU ($-weights_data)/4 192 | weights_rows EQU 2 193 | weights_cols EQU 3 194 | 195 | f32buff: DB "XXXXXXXXXXXXXXXXXXXXX",13,10 196 | f32buffSz EQU $-f32buff 197 | 198 | uint32buff: DB "XXXXXXXXXX",13,10 199 | uint32buffSz EQU $-uint32buff 200 | -------------------------------------------------------------------------------- /src/tostring.asm: -------------------------------------------------------------------------------- 1 | %use masm 2 | section .text 3 | extern uint4ToHexString 4 | extern uint32ToHexString 5 | extern uint32ToString ; this is an alias to uint32ToString10 6 | extern uint32ToString9 7 | extern uint32ToString10 8 | extern uint32ToString_internal 9 | extern float32ToString 10 | 11 | ; Converts an int 4 (possible values 0-15) to a hex digit 12 | ; Input: 13 | ; AL=number 14 | ; R8=string position 15 | ; Output: 16 | ; AL=hex digit ('0'-'F') 17 | ; R8=R8+1 18 | uint4ToHexString: ; digit in al, position in r8, al is modified 19 | AND al,0Fh 20 | CMP al,0Ah 21 | JC _uint4ToHexString_0 22 | ADD al,'A'-10 23 | JMP _uint4ToHexString_disp 24 | _uint4ToHexString_0: 25 | ADD al,'0' 26 | _uint4ToHexString_disp: 27 | MOV byte ptr [r8], al 28 | INC r8 29 | RET 30 | 31 | ; Converts an uint32 to a hex string 32 | ; Input: 33 | ; EAX = dword uint32 value 34 | ; R8 = string position 35 | ; Output: 36 | ; R8 = R8 + 8 37 | uint32ToHexString: ; dword in eax, position in r8, uses stack to save/restore registers 38 | PUSH rax ; save RAX 39 | PUSH rcx ; save RCX 40 | 41 | MOV CL, 32 42 | PUSH rax ; working rax 43 | _uint32ToHexString_for_digit: 44 | POP rax 45 | PUSH rax 46 | SUB cl,4 47 | SHR eax, cl 48 | CALL uint4ToHexString ; will increment R8 49 | CMP cl,0 50 | JNZ _uint32ToHexString_for_digit 51 | POP rax ; clear the working rax 52 | 53 | POP rcx ; restore rcx 54 | POP rax ; restore RAX 55 | RET 56 | 57 | ; Converts an uint32 to a decimal string. 58 | ; The resulting string will be prefixed with zeros. 59 | ; Similar to printf("%010d"). 60 | ; Input: 61 | ; EAX = dword uint32 62 | ; R8 = string position 63 | ; Output: 64 | ; R8 = R8 + 10 65 | uint32ToString: 66 | uint32ToString10: 67 | PUSH rsi 68 | MOV rsi, conv_table_int 69 | 70 | CALL uint32ToString_internal 71 | 72 | POP rsi 73 | RET 74 | 75 | ; Similar to uint32ToString, but will only use 9 digits. 76 | ; Similar to printf("%09d") 77 | ; Input: 78 | ; EAX = dword uint32 79 | ; R8 = string position 80 | ; Output: 81 | ; R8 = R8 + 9 82 | uint32ToString9: 83 | PUSH rsi 84 | MOV rsi, conv_table_int 85 | ADD rsi, 4 86 | 87 | CALL uint32ToString_internal 88 | 89 | POP rsi 90 | RET 91 | 92 | ; Internal function called by uint32ToString and uint32ToString9 93 | ; Input: 94 | ; EAX = dword uint32 95 | ; R8 = string position 96 | ; RSI = conv_table_int position 97 | ; Output: 98 | ; R8 = R8 + sz 99 | uint32ToString_internal: 100 | PUSH rax 101 | PUSH rbx 102 | PUSH rcx 103 | PUSH rdx 104 | PUSH rsi 105 | _uint32ToString_loop: 106 | XOR edx,edx 107 | MOV ecx, DWORD PTR [rsi] 108 | CMP ecx,0 109 | jz _uint32ToString_done 110 | DIV ecx ; divide edx:eax to ecx => eax = quotient, edx = reminder 111 | ADD al, '0' 112 | MOV byte ptr [r8], al 113 | INC r8 114 | MOV eax, edx 115 | ADD rsi,4 116 | JMP _uint32ToString_loop 117 | _uint32ToString_done: 118 | POP rsi 119 | POP rdx 120 | POP rcx 121 | POP rbx 122 | POP rax 123 | RET 124 | 125 | ; Converts a float32 (IEEE 754 standard) to a decimal string. 126 | ; The resulting string will occupy 21 bytes (example: +0000009999.990234375). 127 | ; Input: 128 | ; EAX = float32 129 | ; R8 = string position 130 | ; Output: 131 | ; R8 = R8 + 20 132 | float32ToString: ; float32 dword in eax, position in r8 133 | PUSH rax 134 | PUSH rbx 135 | PUSH rcx 136 | PUSH rsi 137 | PUSH rdi 138 | 139 | TEST eax, 80000000h 140 | JNZ _f32s_neg 141 | MOV BYTE PTR [r8], '+' 142 | JMP _f32s_exp 143 | _f32s_neg: 144 | MOV BYTE PTR [r8], '-' 145 | 146 | _f32s_exp: 147 | INC r8 148 | PUSH rax 149 | MOV cl, 23 ; separate exponent 150 | SHR rax, cl 151 | AND rax, 0FFh 152 | MOV rcx,rax ; exponent now in rcx 153 | ;CALL uint32ToString 154 | POP rax 155 | 156 | AND rax,7FFFFFh ; separate mantissa 157 | OR rax,800000h ; add missing first 1 bit 158 | SHL rax, 9 ; upper part of rax=integer, eax=decimals 159 | 160 | CMP rcx,127 161 | JZ _f32s_exp_done 162 | JC _f32s_exp_neg 163 | SUB rcx, 127 ; rcx = exponent - 127, is positive 164 | SHL rax, cl ; multiply with exponent 165 | JMP _f32s_exp_done 166 | _f32s_exp_neg: 167 | ADD rcx, 128 168 | NEG rcx 169 | AND rcx, 07Fh 170 | SUB rcx, 1 ; rcx = abs(exponent - 127), exponent was negative 171 | SHR rax, cl ; multiply with exponent 172 | _f32s_exp_done: 173 | 174 | PUSH rax 175 | XOR rbx,rbx 176 | SHLD rbx,rax,32 177 | MOV rax,rbx 178 | CALL uint32ToString10 179 | MOV BYTE PTR [r8], "." 180 | INC r8 181 | POP rbx 182 | 183 | XOR rax,rax 184 | MOV rsi,conv_table_f32 185 | MOV rcx, 32 186 | MOV rdi,80000000h 187 | _f32s_frac: 188 | TEST rbx,rdi 189 | JZ _f32s_frac_skip 190 | ADD eax, DWORD PTR [rsi] 191 | _f32s_frac_skip: 192 | ADD rsi,4 193 | SHR rdi,1 194 | LOOP _f32s_frac 195 | 196 | CALL uint32ToString9 197 | 198 | POP rdi 199 | POP rsi 200 | POP rcx 201 | POP rbx 202 | POP rax 203 | 204 | RET 205 | 206 | 207 | 208 | section .data 209 | 210 | conv_table_int: 211 | DD 1000000000, 100000000, 10000000, 1000000, 100000, 10000, 1000, 100, 10, 1, 0 212 | 213 | conv_table_f32: 214 | ; truncated 215 | ;DD 500000000, 250000000, 125000000, 62500000, 31250000, 15625000, 7812500, 3906250 216 | ;DD 1953125, 976562, 488281, 244140, 122070, 61035, 30517, 15258 217 | ;DD 7629, 3814, 1907, 953, 476, 238, 119, 59 218 | ;DD 29, 14, 7, 3, 1, 0, 0, 0 219 | 220 | ; rounded 221 | DD 500000000, 250000000, 125000000, 62500000, 31250000, 15625000, 7812500, 3906250 222 | DD 1953125, 976563, 488281, 244141, 122070, 61035, 30518, 15259 223 | DD 7629, 3815, 1907, 954, 477, 238, 119, 60 224 | DD 30, 15, 8, 4, 2, 1, 0, 0 225 | 226 | 227 | -------------------------------------------------------------------------------- /src/vectors.asm: -------------------------------------------------------------------------------- 1 | %use masm 2 | section .text 3 | global vecInit 4 | global vecConsoleWrite 5 | global vecMulScalar 6 | global vecMulVecDot 7 | global vecMulVecHadamard 8 | global vecMulMatLines 9 | global vecMulMatColumns 10 | global vecCopy 11 | global vecSigmoid 12 | global vecSigmoid_d 13 | global vecDiff 14 | global vecAdd 15 | global vecMulVecOuter 16 | 17 | extern osConsoleWrite 18 | extern float32ToString 19 | 20 | extern f32_sigmoid 21 | extern f32_sigmoid_d 22 | 23 | ; Initialize vector with float32 number 24 | ; Input: 25 | ; eax = float32 number 26 | ; rcx = vector size 27 | ; rdi = vector 28 | vecInit: 29 | PUSH rcx 30 | PUSH rdi 31 | REP STOSD 32 | POP rdi 33 | POP rcx 34 | RET 35 | 36 | 37 | ; Multiply the vector with a scalar value 38 | ; Input: 39 | ; eax = float32 number 40 | ; rdi = vector (this will be overwritten) 41 | ; rcx = vector size, must be greater than zero 42 | ; Output: 43 | ; rdi = vector * eax 44 | vecMulScalar: 45 | PUSH rax 46 | PUSH rcx 47 | PUSH rdi 48 | MOV DWORD PTR [vec_tmp], eax ; store to temp location for passing to FP unit 49 | 50 | _vms_loop: 51 | FLD DWORD PTR [vec_tmp] 52 | FMUL DWORD PTR [rdi] 53 | FSTP DWORD PTR [rdi] 54 | 55 | ADD rdi,4 56 | LOOP _vms_loop 57 | 58 | POP rdi 59 | POP rcx 60 | POP rax 61 | RET 62 | 63 | ; Multiply the vector with another vector and compute the dot product 64 | ; Input: 65 | ; rdi = vector1 66 | ; rsi = vector2 67 | ; rcx = vector size 68 | ; Output: 69 | ; eax = result 70 | vecMulVecDot: 71 | PUSH rcx 72 | PUSH rdi 73 | PUSH rsi 74 | 75 | FLDZ 76 | 77 | _vmvd_loop: 78 | FLD DWORD PTR [rdi] 79 | FMUL DWORD PTR [rsi] 80 | FADDP 81 | 82 | ADD rdi,4 83 | ADD rsi,4 84 | LOOP _vmvd_loop 85 | 86 | FSTP DWORD PTR [vec_tmp] 87 | MOV eax, DWORD PTR [vec_tmp] 88 | 89 | POP rsi 90 | POP rdi 91 | POP rcx 92 | RET 93 | 94 | ; Multiply the vector with another vector and compute the Hadamard product 95 | ; Input: 96 | ; rdi = vector1 (this will be overwritten) 97 | ; rsi = vector2 98 | ; rcx = vector size 99 | ; Output: 100 | ; rdi = vector1*vector2 101 | vecMulVecHadamard: 102 | PUSH rcx 103 | PUSH rdi 104 | PUSH rsi 105 | 106 | _vmvh_loop: 107 | FLD DWORD PTR [rdi] 108 | FMUL DWORD PTR [rsi] 109 | FSTP DWORD PTR [rdi] 110 | 111 | ADD rdi,4 112 | ADD rsi,4 113 | LOOP _vmvh_loop 114 | 115 | POP rsi 116 | POP rdi 117 | POP rcx 118 | RET 119 | 120 | 121 | ; Multiply a matrix with a vector, line by line 122 | ; Each matrix line is considered a vector and the dot product is computed 123 | ; Input: 124 | ; rdi = vector1 125 | ; rsi = matrix 126 | ; rcx = vector size = line size (number of columns) 127 | ; rdx = number of lines 128 | ; r8 = output vector 129 | ; Output: 130 | ; memory [r8 ... r8+rcx] = result 131 | vecMulMatLines: 132 | PUSH rcx 133 | PUSH rsi 134 | PUSH r8 135 | PUSH rax 136 | PUSH rbx 137 | PUSH rdx 138 | 139 | MOV rbx,rcx 140 | SHL rbx, 2 ; Line size = rcx * 4 141 | 142 | _vmml_loop: 143 | CMP rdx,0 144 | JZ _vmml_done 145 | 146 | CALL vecMulVecDot ; line * vector 147 | MOV DWORD PTR [r8], eax ; result in [r8] 148 | 149 | ADD rsi,rbx ; next line 150 | ADD r8, 4 151 | DEC rdx 152 | JMP _vmml_loop 153 | 154 | _vmml_done: 155 | 156 | POP rdx 157 | POP rbx 158 | POP rax 159 | POP r8 160 | POP rsi 161 | POP rcx 162 | RET 163 | 164 | 165 | ; Multiply a matrix with a vector, column by column (this is the regular vector matrix product) 166 | ; Each matrix column is considered a vector and the dot product is computed 167 | ; Input: 168 | ; rdi = vector1 169 | ; rsi = matrix 170 | ; rcx = vector size = column size (number of lines) 171 | ; rdx = number of columns (line size) 172 | ; r8 = output vector 173 | ; Output: 174 | ; memory [r8 ... r8+rcx] = result 175 | vecMulMatColumns: 176 | PUSH rcx 177 | PUSH rsi 178 | PUSH r8 179 | PUSH rax 180 | PUSH rbx 181 | PUSH rdx 182 | 183 | MOV rbx,rdx 184 | SHL rbx, 2 ; Line size = rdx * 4 185 | 186 | _vmvc_loop: 187 | CMP rdx,0 188 | JZ _vmvc_done 189 | 190 | ; CALL vecMulVecDot ; this cannot be called, instead we need to implement the code here 191 | 192 | PUSH rcx 193 | PUSH rdi 194 | PUSH rsi 195 | 196 | FLDZ 197 | 198 | _vmvc_loop1: 199 | FLD DWORD PTR [rdi] 200 | FMUL DWORD PTR [rsi] 201 | FADDP 202 | 203 | ADD rdi,4 204 | ADD rsi,rbx 205 | LOOP _vmvc_loop1 206 | 207 | FSTP DWORD PTR [r8] 208 | 209 | POP rsi 210 | POP rdi 211 | POP rcx 212 | 213 | ADD rsi, 4 ; next column 214 | ADD r8, 4 215 | DEC rdx 216 | JMP _vmvc_loop 217 | 218 | _vmvc_done: 219 | POP rdx 220 | POP rbx 221 | POP rax 222 | POP r8 223 | POP rsi 224 | POP rcx 225 | RET 226 | 227 | ; Write a vector to console 228 | ; Input: 229 | ; rdi = vector 230 | ; rcx = vector size 231 | vecConsoleWrite: 232 | PUSH rax 233 | PUSH rcx 234 | PUSH rsi 235 | PUSH rdi 236 | PUSH r8 237 | PUSH rdx 238 | 239 | _vcw_loop: 240 | MOV eax, DWORD PTR [rdi] 241 | MOV r8, vecfp32Buff 242 | CALL float32ToString 243 | 244 | MOV rsi, vecfp32Buff 245 | MOV rdx, vecfp32BuffSz 246 | CALL osConsoleWrite 247 | 248 | ADD rdi,4 249 | LOOP _vcw_loop 250 | 251 | POP rdx 252 | POP r8 253 | POP rdi 254 | POP rsi 255 | POP rcx 256 | POP rax 257 | RET 258 | 259 | ; Copy vector1 to vector 2 260 | ; Input: 261 | ; rdi = vector1 262 | ; rsi = vector2 263 | ; rcx = vector size 264 | vecCopy: 265 | PUSH rcx 266 | PUSH rsi 267 | PUSH rdi 268 | 269 | XCHG rsi,rdi 270 | REP MOVSD 271 | 272 | POP rdi 273 | POP rsi 274 | POP rcx 275 | RET 276 | 277 | ; Apply the sigmoid function on vector1 and store in vector 2 278 | ; Input: 279 | ; rdi = vector1 280 | ; rsi = vector2 281 | ; rcx = vector size 282 | vecSigmoid: 283 | PUSH rcx 284 | PUSH rsi 285 | PUSH rdi 286 | 287 | _vc_loop: 288 | call f32_sigmoid 289 | ADD rsi,4 290 | ADD rdi,4 291 | LOOP _vc_loop 292 | 293 | POP rdi 294 | POP rsi 295 | POP rcx 296 | RET 297 | 298 | ; Apply the sigmoid derivation function on vector1 and store in vector 2 299 | ; Input: 300 | ; rdi = vector1 301 | ; rsi = vector2 302 | ; rcx = vector size 303 | vecSigmoid_d: 304 | PUSH rcx 305 | PUSH rsi 306 | PUSH rdi 307 | 308 | _vcd_loop: 309 | call f32_sigmoid_d 310 | ADD rsi,4 311 | ADD rdi,4 312 | LOOP _vcd_loop 313 | 314 | POP rdi 315 | POP rsi 316 | POP rcx 317 | RET 318 | 319 | ; Add two vectors 320 | ; Input: 321 | ; rdi = vector1 (this will be overwritten) 322 | ; rsi = vector2 323 | ; rcx = vector size 324 | ; Output: 325 | ; rdi = vector1+vector2 326 | vecAdd: 327 | PUSH rcx 328 | PUSH rdi 329 | PUSH rsi 330 | 331 | _va_loop: 332 | FLD DWORD PTR [rdi] 333 | FADD DWORD PTR [rsi] 334 | FSTP DWORD PTR [rdi] 335 | 336 | ADD rdi,4 337 | ADD rsi,4 338 | LOOP _va_loop 339 | 340 | POP rsi 341 | POP rdi 342 | POP rcx 343 | RET 344 | 345 | ; Subtract two vectors 346 | ; Input: 347 | ; rdi = vector1 (this will be overwritten) 348 | ; rsi = vector2 349 | ; rcx = vector size 350 | ; Output: 351 | ; rdi = vector1-vector2 352 | vecDiff: 353 | PUSH rcx 354 | PUSH rdi 355 | PUSH rsi 356 | 357 | _vd_loop: 358 | FLD DWORD PTR [rdi] 359 | FSUB DWORD PTR [rsi] 360 | FSTP DWORD PTR [rdi] 361 | 362 | ADD rdi,4 363 | ADD rsi,4 364 | LOOP _vd_loop 365 | 366 | POP rsi 367 | POP rdi 368 | POP rcx 369 | RET 370 | 371 | ; Multiply one vector with another vector and compute the outer product 372 | ; The result is a matrix 373 | ; Input: 374 | ; rdi = vector1 375 | ; rsi = vector2 376 | ; rcx = vector1 size = line size (number of columns) 377 | ; rdx = vector2 size = column size (number of lines) 378 | ; r8 = output matrix (must have space vec1_size x vec2_size) 379 | ; Output: 380 | ; [r8] = vector1*vector2 381 | vecMulVecOuter: 382 | PUSH rcx 383 | PUSH rdi 384 | PUSH rsi 385 | PUSH rdx 386 | 387 | _vmvc_loop_v2: 388 | PUSH rcx 389 | PUSH rdi 390 | FLD DWORD PTR [rsi] 391 | 392 | _vmvc_loop_v1: 393 | FLD st0 394 | FMUL DWORD PTR [rdi] 395 | FSTP DWORD PTR [r8] 396 | 397 | ADD rdi,4 398 | ADD r8,4 399 | LOOP _vmvc_loop_v1 400 | 401 | FSTP st0 402 | POP rdi 403 | POP rcx 404 | ADD rsi,4 405 | DEC rdx 406 | jnz _vmvc_loop_v2 407 | 408 | POP rdx 409 | POP rsi 410 | POP rdi 411 | POP rcx 412 | RET 413 | 414 | 415 | section .data 416 | 417 | vecfp32Buff: DB "XXXXXXXXXXXXXXXXXXXXX",13,10 418 | vecfp32BuffSz EQU $-vecfp32Buff 419 | 420 | vec_tmp: DD 0 421 | --------------------------------------------------------------------------------