From 13dcf1d26ef81a3fcaf15dfd52757fc4b0357b7c Mon Sep 17 00:00:00 2001 From: Aaron Patterson Date: Mon, 6 Jul 2026 15:12:35 -0700 Subject: [PATCH 1/4] initial BLAKE3 import --- README.md | 3 + Rakefile | 2 +- digest.gemspec | 3 + ext/digest/blake3/LICENSE_A2 | 202 +++++++++ ext/digest/blake3/LICENSE_A2LLVM | 219 ++++++++++ ext/digest/blake3/LICENSE_CC0 | 121 ++++++ ext/digest/blake3/blake3.c | 651 ++++++++++++++++++++++++++++ ext/digest/blake3/blake3.h | 86 ++++ ext/digest/blake3/blake3_dispatch.c | 332 ++++++++++++++ ext/digest/blake3/blake3_impl.h | 333 ++++++++++++++ ext/digest/blake3/blake3_portable.c | 160 +++++++ ext/digest/blake3/blake3init.c | 82 ++++ ext/digest/blake3/extconf.rb | 27 ++ test/digest/test_digest.rb | 102 ++++- 14 files changed, 2321 insertions(+), 2 deletions(-) create mode 100644 ext/digest/blake3/LICENSE_A2 create mode 100644 ext/digest/blake3/LICENSE_A2LLVM create mode 100644 ext/digest/blake3/LICENSE_CC0 create mode 100644 ext/digest/blake3/blake3.c create mode 100644 ext/digest/blake3/blake3.h create mode 100644 ext/digest/blake3/blake3_dispatch.c create mode 100644 ext/digest/blake3/blake3_impl.h create mode 100644 ext/digest/blake3/blake3_portable.c create mode 100644 ext/digest/blake3/blake3init.c create mode 100644 ext/digest/blake3/extconf.rb diff --git a/README.md b/README.md index ae25561..70d2434 100644 --- a/README.md +++ b/README.md @@ -78,6 +78,9 @@ SHA2 family:: SHA512 SHA384 SHA256 +BLAKE3:: + As Digest::BLAKE3, producing a 256-bit digest. + See https://github.com/BLAKE3-team/BLAKE3. ``` The latest versions of the FIPS publications can be found here: http://csrc.nist.gov/publications/PubsFIPS.html. diff --git a/Rakefile b/Rakefile index bfaa0c5..c0265be 100644 --- a/Rakefile +++ b/Rakefile @@ -24,7 +24,7 @@ if RUBY_ENGINE == "jruby" else require "rake/extensiontask" Rake::ExtensionTask.new("digest") - %w[bubblebabble md5 rmd160 sha1 sha2 crc32].each do |ext| + %w[bubblebabble md5 rmd160 sha1 sha2 crc32 blake3].each do |ext| Rake::ExtensionTask.new("digest/#{ext}") end end diff --git a/digest.gemspec b/digest.gemspec index 4a01c5f..033779f 100644 --- a/digest.gemspec +++ b/digest.gemspec @@ -37,6 +37,9 @@ Gem::Specification.new do |spec| spec.extensions = Dir["ext/digest/**/extconf.rb"] spec.files += Dir["ext/digest/**/{*.{rb,c,h,sh},depend}"] + # Vendored third-party license files (e.g. BLAKE3's LICENSE_*), which + # carry no filename extension and are missed by the glob above. + spec.files += Dir["ext/digest/**/LICENSE*"] spec.require_paths = %w[lib] end diff --git a/ext/digest/blake3/LICENSE_A2 b/ext/digest/blake3/LICENSE_A2 new file mode 100644 index 0000000..d512ca9 --- /dev/null +++ b/ext/digest/blake3/LICENSE_A2 @@ -0,0 +1,202 @@ + + Apache License + Version 2.0, January 2004 + http://www.apache.org/licenses/ + + TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION + + 1. Definitions. + + "License" shall mean the terms and conditions for use, reproduction, + and distribution as defined by Sections 1 through 9 of this document. + + "Licensor" shall mean the copyright owner or entity authorized by + the copyright owner that is granting the License. + + "Legal Entity" shall mean the union of the acting entity and all + other entities that control, are controlled by, or are under common + control with that entity. For the purposes of this definition, + "control" means (i) the power, direct or indirect, to cause the + direction or management of such entity, whether by contract or + otherwise, or (ii) ownership of fifty percent (50%) or more of the + outstanding shares, or (iii) beneficial ownership of such entity. + + "You" (or "Your") shall mean an individual or Legal Entity + exercising permissions granted by this License. + + "Source" form shall mean the preferred form for making modifications, + including but not limited to software source code, documentation + source, and configuration files. + + "Object" form shall mean any form resulting from mechanical + transformation or translation of a Source form, including but + not limited to compiled object code, generated documentation, + and conversions to other media types. + + "Work" shall mean the work of authorship, whether in Source or + Object form, made available under the License, as indicated by a + copyright notice that is included in or attached to the work + (an example is provided in the Appendix below). + + "Derivative Works" shall mean any work, whether in Source or Object + form, that is based on (or derived from) the Work and for which the + editorial revisions, annotations, elaborations, or other modifications + represent, as a whole, an original work of authorship. For the purposes + of this License, Derivative Works shall not include works that remain + separable from, or merely link (or bind by name) to the interfaces of, + the Work and Derivative Works thereof. + + "Contribution" shall mean any work of authorship, including + the original version of the Work and any modifications or additions + to that Work or Derivative Works thereof, that is intentionally + submitted to Licensor for inclusion in the Work by the copyright owner + or by an individual or Legal Entity authorized to submit on behalf of + the copyright owner. For the purposes of this definition, "submitted" + means any form of electronic, verbal, or written communication sent + to the Licensor or its representatives, including but not limited to + communication on electronic mailing lists, source code control systems, + and issue tracking systems that are managed by, or on behalf of, the + Licensor for the purpose of discussing and improving the Work, but + excluding communication that is conspicuously marked or otherwise + designated in writing by the copyright owner as "Not a Contribution." + + "Contributor" shall mean Licensor and any individual or Legal Entity + on behalf of whom a Contribution has been received by Licensor and + subsequently incorporated within the Work. + + 2. Grant of Copyright License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + copyright license to reproduce, prepare Derivative Works of, + publicly display, publicly perform, sublicense, and distribute the + Work and such Derivative Works in Source or Object form. + + 3. Grant of Patent License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + (except as stated in this section) patent license to make, have made, + use, offer to sell, sell, import, and otherwise transfer the Work, + where such license applies only to those patent claims licensable + by such Contributor that are necessarily infringed by their + Contribution(s) alone or by combination of their Contribution(s) + with the Work to which such Contribution(s) was submitted. If You + institute patent litigation against any entity (including a + cross-claim or counterclaim in a lawsuit) alleging that the Work + or a Contribution incorporated within the Work constitutes direct + or contributory patent infringement, then any patent licenses + granted to You under this License for that Work shall terminate + as of the date such litigation is filed. + + 4. Redistribution. You may reproduce and distribute copies of the + Work or Derivative Works thereof in any medium, with or without + modifications, and in Source or Object form, provided that You + meet the following conditions: + + (a) You must give any other recipients of the Work or + Derivative Works a copy of this License; and + + (b) You must cause any modified files to carry prominent notices + stating that You changed the files; and + + (c) You must retain, in the Source form of any Derivative Works + that You distribute, all copyright, patent, trademark, and + attribution notices from the Source form of the Work, + excluding those notices that do not pertain to any part of + the Derivative Works; and + + (d) If the Work includes a "NOTICE" text file as part of its + distribution, then any Derivative Works that You distribute must + include a readable copy of the attribution notices contained + within such NOTICE file, excluding those notices that do not + pertain to any part of the Derivative Works, in at least one + of the following places: within a NOTICE text file distributed + as part of the Derivative Works; within the Source form or + documentation, if provided along with the Derivative Works; or, + within a display generated by the Derivative Works, if and + wherever such third-party notices normally appear. The contents + of the NOTICE file are for informational purposes only and + do not modify the License. You may add Your own attribution + notices within Derivative Works that You distribute, alongside + or as an addendum to the NOTICE text from the Work, provided + that such additional attribution notices cannot be construed + as modifying the License. + + You may add Your own copyright statement to Your modifications and + may provide additional or different license terms and conditions + for use, reproduction, or distribution of Your modifications, or + for any such Derivative Works as a whole, provided Your use, + reproduction, and distribution of the Work otherwise complies with + the conditions stated in this License. + + 5. Submission of Contributions. Unless You explicitly state otherwise, + any Contribution intentionally submitted for inclusion in the Work + by You to the Licensor shall be under the terms and conditions of + this License, without any additional terms or conditions. + Notwithstanding the above, nothing herein shall supersede or modify + the terms of any separate license agreement you may have executed + with Licensor regarding such Contributions. + + 6. Trademarks. This License does not grant permission to use the trade + names, trademarks, service marks, or product names of the Licensor, + except as required for reasonable and customary use in describing the + origin of the Work and reproducing the content of the NOTICE file. + + 7. Disclaimer of Warranty. Unless required by applicable law or + agreed to in writing, Licensor provides the Work (and each + Contributor provides its Contributions) on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or + implied, including, without limitation, any warranties or conditions + of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A + PARTICULAR PURPOSE. You are solely responsible for determining the + appropriateness of using or redistributing the Work and assume any + risks associated with Your exercise of permissions under this License. + + 8. Limitation of Liability. In no event and under no legal theory, + whether in tort (including negligence), contract, or otherwise, + unless required by applicable law (such as deliberate and grossly + negligent acts) or agreed to in writing, shall any Contributor be + liable to You for damages, including any direct, indirect, special, + incidental, or consequential damages of any character arising as a + result of this License or out of the use or inability to use the + Work (including but not limited to damages for loss of goodwill, + work stoppage, computer failure or malfunction, or any and all + other commercial damages or losses), even if such Contributor + has been advised of the possibility of such damages. + + 9. Accepting Warranty or Additional Liability. While redistributing + the Work or Derivative Works thereof, You may choose to offer, + and charge a fee for, acceptance of support, warranty, indemnity, + or other liability obligations and/or rights consistent with this + License. However, in accepting such obligations, You may act only + on Your own behalf and on Your sole responsibility, not on behalf + of any other Contributor, and only if You agree to indemnify, + defend, and hold each Contributor harmless for any liability + incurred by, or claims asserted against, such Contributor by reason + of your accepting any such warranty or additional liability. + + END OF TERMS AND CONDITIONS + + APPENDIX: How to apply the Apache License to your work. + + To apply the Apache License to your work, attach the following + boilerplate notice, with the fields enclosed by brackets "[]" + replaced with your own identifying information. (Don't include + the brackets!) The text should be enclosed in the appropriate + comment syntax for the file format. We also recommend that a + file or class name and description of purpose be included on the + same "printed page" as the copyright notice for easier + identification within third-party archives. + + Copyright 2019 Jack O'Connor and Samuel Neves + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. diff --git a/ext/digest/blake3/LICENSE_A2LLVM b/ext/digest/blake3/LICENSE_A2LLVM new file mode 100644 index 0000000..8d11bec --- /dev/null +++ b/ext/digest/blake3/LICENSE_A2LLVM @@ -0,0 +1,219 @@ + Apache License + Version 2.0, January 2004 + http://www.apache.org/licenses/ + + TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION + + 1. Definitions. + + "License" shall mean the terms and conditions for use, reproduction, + and distribution as defined by Sections 1 through 9 of this document. + + "Licensor" shall mean the copyright owner or entity authorized by + the copyright owner that is granting the License. + + "Legal Entity" shall mean the union of the acting entity and all + other entities that control, are controlled by, or are under common + control with that entity. For the purposes of this definition, + "control" means (i) the power, direct or indirect, to cause the + direction or management of such entity, whether by contract or + otherwise, or (ii) ownership of fifty percent (50%) or more of the + outstanding shares, or (iii) beneficial ownership of such entity. + + "You" (or "Your") shall mean an individual or Legal Entity + exercising permissions granted by this License. + + "Source" form shall mean the preferred form for making modifications, + including but not limited to software source code, documentation + source, and configuration files. + + "Object" form shall mean any form resulting from mechanical + transformation or translation of a Source form, including but + not limited to compiled object code, generated documentation, + and conversions to other media types. + + "Work" shall mean the work of authorship, whether in Source or + Object form, made available under the License, as indicated by a + copyright notice that is included in or attached to the work + (an example is provided in the Appendix below). + + "Derivative Works" shall mean any work, whether in Source or Object + form, that is based on (or derived from) the Work and for which the + editorial revisions, annotations, elaborations, or other modifications + represent, as a whole, an original work of authorship. For the purposes + of this License, Derivative Works shall not include works that remain + separable from, or merely link (or bind by name) to the interfaces of, + the Work and Derivative Works thereof. + + "Contribution" shall mean any work of authorship, including + the original version of the Work and any modifications or additions + to that Work or Derivative Works thereof, that is intentionally + submitted to Licensor for inclusion in the Work by the copyright owner + or by an individual or Legal Entity authorized to submit on behalf of + the copyright owner. For the purposes of this definition, "submitted" + means any form of electronic, verbal, or written communication sent + to the Licensor or its representatives, including but not limited to + communication on electronic mailing lists, source code control systems, + and issue tracking systems that are managed by, or on behalf of, the + Licensor for the purpose of discussing and improving the Work, but + excluding communication that is conspicuously marked or otherwise + designated in writing by the copyright owner as "Not a Contribution." + + "Contributor" shall mean Licensor and any individual or Legal Entity + on behalf of whom a Contribution has been received by Licensor and + subsequently incorporated within the Work. + + 2. Grant of Copyright License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + copyright license to reproduce, prepare Derivative Works of, + publicly display, publicly perform, sublicense, and distribute the + Work and such Derivative Works in Source or Object form. + + 3. Grant of Patent License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + (except as stated in this section) patent license to make, have made, + use, offer to sell, sell, import, and otherwise transfer the Work, + where such license applies only to those patent claims licensable + by such Contributor that are necessarily infringed by their + Contribution(s) alone or by combination of their Contribution(s) + with the Work to which such Contribution(s) was submitted. If You + institute patent litigation against any entity (including a + cross-claim or counterclaim in a lawsuit) alleging that the Work + or a Contribution incorporated within the Work constitutes direct + or contributory patent infringement, then any patent licenses + granted to You under this License for that Work shall terminate + as of the date such litigation is filed. + + 4. Redistribution. You may reproduce and distribute copies of the + Work or Derivative Works thereof in any medium, with or without + modifications, and in Source or Object form, provided that You + meet the following conditions: + + (a) You must give any other recipients of the Work or + Derivative Works a copy of this License; and + + (b) You must cause any modified files to carry prominent notices + stating that You changed the files; and + + (c) You must retain, in the Source form of any Derivative Works + that You distribute, all copyright, patent, trademark, and + attribution notices from the Source form of the Work, + excluding those notices that do not pertain to any part of + the Derivative Works; and + + (d) If the Work includes a "NOTICE" text file as part of its + distribution, then any Derivative Works that You distribute must + include a readable copy of the attribution notices contained + within such NOTICE file, excluding those notices that do not + pertain to any part of the Derivative Works, in at least one + of the following places: within a NOTICE text file distributed + as part of the Derivative Works; within the Source form or + documentation, if provided along with the Derivative Works; or, + within a display generated by the Derivative Works, if and + wherever such third-party notices normally appear. The contents + of the NOTICE file are for informational purposes only and + do not modify the License. You may add Your own attribution + notices within Derivative Works that You distribute, alongside + or as an addendum to the NOTICE text from the Work, provided + that such additional attribution notices cannot be construed + as modifying the License. + + You may add Your own copyright statement to Your modifications and + may provide additional or different license terms and conditions + for use, reproduction, or distribution of Your modifications, or + for any such Derivative Works as a whole, provided Your use, + reproduction, and distribution of the Work otherwise complies with + the conditions stated in this License. + + 5. Submission of Contributions. Unless You explicitly state otherwise, + any Contribution intentionally submitted for inclusion in the Work + by You to the Licensor shall be under the terms and conditions of + this License, without any additional terms or conditions. + Notwithstanding the above, nothing herein shall supersede or modify + the terms of any separate license agreement you may have executed + with Licensor regarding such Contributions. + + 6. Trademarks. This License does not grant permission to use the trade + names, trademarks, service marks, or product names of the Licensor, + except as required for reasonable and customary use in describing the + origin of the Work and reproducing the content of the NOTICE file. + + 7. Disclaimer of Warranty. Unless required by applicable law or + agreed to in writing, Licensor provides the Work (and each + Contributor provides its Contributions) on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or + implied, including, without limitation, any warranties or conditions + of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A + PARTICULAR PURPOSE. You are solely responsible for determining the + appropriateness of using or redistributing the Work and assume any + risks associated with Your exercise of permissions under this License. + + 8. Limitation of Liability. In no event and under no legal theory, + whether in tort (including negligence), contract, or otherwise, + unless required by applicable law (such as deliberate and grossly + negligent acts) or agreed to in writing, shall any Contributor be + liable to You for damages, including any direct, indirect, special, + incidental, or consequential damages of any character arising as a + result of this License or out of the use or inability to use the + Work (including but not limited to damages for loss of goodwill, + work stoppage, computer failure or malfunction, or any and all + other commercial damages or losses), even if such Contributor + has been advised of the possibility of such damages. + + 9. Accepting Warranty or Additional Liability. While redistributing + the Work or Derivative Works thereof, You may choose to offer, + and charge a fee for, acceptance of support, warranty, indemnity, + or other liability obligations and/or rights consistent with this + License. However, in accepting such obligations, You may act only + on Your own behalf and on Your sole responsibility, not on behalf + of any other Contributor, and only if You agree to indemnify, + defend, and hold each Contributor harmless for any liability + incurred by, or claims asserted against, such Contributor by reason + of your accepting any such warranty or additional liability. + + END OF TERMS AND CONDITIONS + + APPENDIX: How to apply the Apache License to your work. + + To apply the Apache License to your work, attach the following + boilerplate notice, with the fields enclosed by brackets "[]" + replaced with your own identifying information. (Don't include + the brackets!) The text should be enclosed in the appropriate + comment syntax for the file format. We also recommend that a + file or class name and description of purpose be included on the + same "printed page" as the copyright notice for easier + identification within third-party archives. + + Copyright 2019 Jack O'Connor and Samuel Neves + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. + + +---- LLVM Exceptions to the Apache 2.0 License ---- + +As an exception, if, as a result of your compiling your source code, portions +of this Software are embedded into an Object form of such source code, you +may redistribute such embedded portions in such Object form without complying +with the conditions of Sections 4(a), 4(b) and 4(d) of the License. + +In addition, if you combine or link compiled forms of this Software with +software that is licensed under the GPLv2 ("Combined Software") and if a +court of competent jurisdiction determines that the patent provision (Section +3), the indemnity provision (Section 9) or other Section of the License +conflicts with the conditions of the GPLv2, you may retroactively and +prospectively choose to deem waived or otherwise exclude such Section(s) of +the License, but only in their entirety and only with respect to the Combined +Software. + diff --git a/ext/digest/blake3/LICENSE_CC0 b/ext/digest/blake3/LICENSE_CC0 new file mode 100644 index 0000000..0e259d4 --- /dev/null +++ b/ext/digest/blake3/LICENSE_CC0 @@ -0,0 +1,121 @@ +Creative Commons Legal Code + +CC0 1.0 Universal + + CREATIVE COMMONS CORPORATION IS NOT A LAW FIRM AND DOES NOT PROVIDE + LEGAL SERVICES. DISTRIBUTION OF THIS DOCUMENT DOES NOT CREATE AN + ATTORNEY-CLIENT RELATIONSHIP. CREATIVE COMMONS PROVIDES THIS + INFORMATION ON AN "AS-IS" BASIS. CREATIVE COMMONS MAKES NO WARRANTIES + REGARDING THE USE OF THIS DOCUMENT OR THE INFORMATION OR WORKS + PROVIDED HEREUNDER, AND DISCLAIMS LIABILITY FOR DAMAGES RESULTING FROM + THE USE OF THIS DOCUMENT OR THE INFORMATION OR WORKS PROVIDED + HEREUNDER. + +Statement of Purpose + +The laws of most jurisdictions throughout the world automatically confer +exclusive Copyright and Related Rights (defined below) upon the creator +and subsequent owner(s) (each and all, an "owner") of an original work of +authorship and/or a database (each, a "Work"). + +Certain owners wish to permanently relinquish those rights to a Work for +the purpose of contributing to a commons of creative, cultural and +scientific works ("Commons") that the public can reliably and without fear +of later claims of infringement build upon, modify, incorporate in other +works, reuse and redistribute as freely as possible in any form whatsoever +and for any purposes, including without limitation commercial purposes. +These owners may contribute to the Commons to promote the ideal of a free +culture and the further production of creative, cultural and scientific +works, or to gain reputation or greater distribution for their Work in +part through the use and efforts of others. + +For these and/or other purposes and motivations, and without any +expectation of additional consideration or compensation, the person +associating CC0 with a Work (the "Affirmer"), to the extent that he or she +is an owner of Copyright and Related Rights in the Work, voluntarily +elects to apply CC0 to the Work and publicly distribute the Work under its +terms, with knowledge of his or her Copyright and Related Rights in the +Work and the meaning and intended legal effect of CC0 on those rights. + +1. Copyright and Related Rights. A Work made available under CC0 may be +protected by copyright and related or neighboring rights ("Copyright and +Related Rights"). Copyright and Related Rights include, but are not +limited to, the following: + + i. the right to reproduce, adapt, distribute, perform, display, + communicate, and translate a Work; + ii. moral rights retained by the original author(s) and/or performer(s); +iii. publicity and privacy rights pertaining to a person's image or + likeness depicted in a Work; + iv. rights protecting against unfair competition in regards to a Work, + subject to the limitations in paragraph 4(a), below; + v. rights protecting the extraction, dissemination, use and reuse of data + in a Work; + vi. database rights (such as those arising under Directive 96/9/EC of the + European Parliament and of the Council of 11 March 1996 on the legal + protection of databases, and under any national implementation + thereof, including any amended or successor version of such + directive); and +vii. other similar, equivalent or corresponding rights throughout the + world based on applicable law or treaty, and any national + implementations thereof. + +2. Waiver. To the greatest extent permitted by, but not in contravention +of, applicable law, Affirmer hereby overtly, fully, permanently, +irrevocably and unconditionally waives, abandons, and surrenders all of +Affirmer's Copyright and Related Rights and associated claims and causes +of action, whether now known or unknown (including existing as well as +future claims and causes of action), in the Work (i) in all territories +worldwide, (ii) for the maximum duration provided by applicable law or +treaty (including future time extensions), (iii) in any current or future +medium and for any number of copies, and (iv) for any purpose whatsoever, +including without limitation commercial, advertising or promotional +purposes (the "Waiver"). Affirmer makes the Waiver for the benefit of each +member of the public at large and to the detriment of Affirmer's heirs and +successors, fully intending that such Waiver shall not be subject to +revocation, rescission, cancellation, termination, or any other legal or +equitable action to disrupt the quiet enjoyment of the Work by the public +as contemplated by Affirmer's express Statement of Purpose. + +3. Public License Fallback. Should any part of the Waiver for any reason +be judged legally invalid or ineffective under applicable law, then the +Waiver shall be preserved to the maximum extent permitted taking into +account Affirmer's express Statement of Purpose. In addition, to the +extent the Waiver is so judged Affirmer hereby grants to each affected +person a royalty-free, non transferable, non sublicensable, non exclusive, +irrevocable and unconditional license to exercise Affirmer's Copyright and +Related Rights in the Work (i) in all territories worldwide, (ii) for the +maximum duration provided by applicable law or treaty (including future +time extensions), (iii) in any current or future medium and for any number +of copies, and (iv) for any purpose whatsoever, including without +limitation commercial, advertising or promotional purposes (the +"License"). The License shall be deemed effective as of the date CC0 was +applied by Affirmer to the Work. Should any part of the License for any +reason be judged legally invalid or ineffective under applicable law, such +partial invalidity or ineffectiveness shall not invalidate the remainder +of the License, and in such case Affirmer hereby affirms that he or she +will not (i) exercise any of his or her remaining Copyright and Related +Rights in the Work or (ii) assert any associated claims and causes of +action with respect to the Work, in either case contrary to Affirmer's +express Statement of Purpose. + +4. Limitations and Disclaimers. + + a. No trademark or patent rights held by Affirmer are waived, abandoned, + surrendered, licensed or otherwise affected by this document. + b. Affirmer offers the Work as-is and makes no representations or + warranties of any kind concerning the Work, express, implied, + statutory or otherwise, including without limitation warranties of + title, merchantability, fitness for a particular purpose, non + infringement, or the absence of latent or other defects, accuracy, or + the present or absence of errors, whether or not discoverable, all to + the greatest extent permissible under applicable law. + c. Affirmer disclaims responsibility for clearing rights of other persons + that may apply to the Work or any use thereof, including without + limitation any person's Copyright and Related Rights in the Work. + Further, Affirmer disclaims responsibility for obtaining any necessary + consents, permissions or other rights required for any use of the + Work. + d. Affirmer understands and acknowledges that Creative Commons is not a + party to this document and has no duty or obligation with respect to + this CC0 or use of the Work. diff --git a/ext/digest/blake3/blake3.c b/ext/digest/blake3/blake3.c new file mode 100644 index 0000000..00f91f4 --- /dev/null +++ b/ext/digest/blake3/blake3.c @@ -0,0 +1,651 @@ +#include +#include +#include +#include + +#include "blake3.h" +#include "blake3_impl.h" + +const char *blake3_version(void) { return BLAKE3_VERSION_STRING; } + +INLINE void chunk_state_init(blake3_chunk_state *self, const uint32_t key[8], + uint8_t flags) { + memcpy(self->cv, key, BLAKE3_KEY_LEN); + self->chunk_counter = 0; + memset(self->buf, 0, BLAKE3_BLOCK_LEN); + self->buf_len = 0; + self->blocks_compressed = 0; + self->flags = flags; +} + +INLINE void chunk_state_reset(blake3_chunk_state *self, const uint32_t key[8], + uint64_t chunk_counter) { + memcpy(self->cv, key, BLAKE3_KEY_LEN); + self->chunk_counter = chunk_counter; + self->blocks_compressed = 0; + memset(self->buf, 0, BLAKE3_BLOCK_LEN); + self->buf_len = 0; +} + +INLINE size_t chunk_state_len(const blake3_chunk_state *self) { + return (BLAKE3_BLOCK_LEN * (size_t)self->blocks_compressed) + + ((size_t)self->buf_len); +} + +INLINE size_t chunk_state_fill_buf(blake3_chunk_state *self, + const uint8_t *input, size_t input_len) { + size_t take = BLAKE3_BLOCK_LEN - ((size_t)self->buf_len); + if (take > input_len) { + take = input_len; + } + uint8_t *dest = self->buf + ((size_t)self->buf_len); + memcpy(dest, input, take); + self->buf_len += (uint8_t)take; + return take; +} + +INLINE uint8_t chunk_state_maybe_start_flag(const blake3_chunk_state *self) { + if (self->blocks_compressed == 0) { + return CHUNK_START; + } else { + return 0; + } +} + +typedef struct { + uint32_t input_cv[8]; + uint64_t counter; + uint8_t block[BLAKE3_BLOCK_LEN]; + uint8_t block_len; + uint8_t flags; +} output_t; + +INLINE output_t make_output(const uint32_t input_cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags) { + output_t ret; + memcpy(ret.input_cv, input_cv, 32); + memcpy(ret.block, block, BLAKE3_BLOCK_LEN); + ret.block_len = block_len; + ret.counter = counter; + ret.flags = flags; + return ret; +} + +// Chaining values within a given chunk (specifically the compress_in_place +// interface) are represented as words. This avoids unnecessary bytes<->words +// conversion overhead in the portable implementation. However, the hash_many +// interface handles both user input and parent node blocks, so it accepts +// bytes. For that reason, chaining values in the CV stack are represented as +// bytes. +INLINE void output_chaining_value(const output_t *self, uint8_t cv[32]) { + uint32_t cv_words[8]; + memcpy(cv_words, self->input_cv, 32); + blake3_compress_in_place(cv_words, self->block, self->block_len, + self->counter, self->flags); + store_cv_words(cv, cv_words); +} + +INLINE void output_root_bytes(const output_t *self, uint64_t seek, uint8_t *out, + size_t out_len) { + if (out_len == 0) { + return; + } + uint64_t output_block_counter = seek / 64; + size_t offset_within_block = seek % 64; + uint8_t wide_buf[64]; + if(offset_within_block) { + blake3_compress_xof(self->input_cv, self->block, self->block_len, output_block_counter, self->flags | ROOT, wide_buf); + const size_t available_bytes = 64 - offset_within_block; + const size_t bytes = out_len > available_bytes ? available_bytes : out_len; + memcpy(out, wide_buf + offset_within_block, bytes); + out += bytes; + out_len -= bytes; + output_block_counter += 1; + } + if(out_len / 64) { + blake3_xof_many(self->input_cv, self->block, self->block_len, output_block_counter, self->flags | ROOT, out, out_len / 64); + } + output_block_counter += out_len / 64; + out += out_len & -64; + out_len -= out_len & -64; + if(out_len) { + blake3_compress_xof(self->input_cv, self->block, self->block_len, output_block_counter, self->flags | ROOT, wide_buf); + memcpy(out, wide_buf, out_len); + } +} + +INLINE void chunk_state_update(blake3_chunk_state *self, const uint8_t *input, + size_t input_len) { + if (self->buf_len > 0) { + size_t take = chunk_state_fill_buf(self, input, input_len); + input += take; + input_len -= take; + if (input_len > 0) { + blake3_compress_in_place( + self->cv, self->buf, BLAKE3_BLOCK_LEN, self->chunk_counter, + self->flags | chunk_state_maybe_start_flag(self)); + self->blocks_compressed += 1; + self->buf_len = 0; + memset(self->buf, 0, BLAKE3_BLOCK_LEN); + } + } + + while (input_len > BLAKE3_BLOCK_LEN) { + blake3_compress_in_place(self->cv, input, BLAKE3_BLOCK_LEN, + self->chunk_counter, + self->flags | chunk_state_maybe_start_flag(self)); + self->blocks_compressed += 1; + input += BLAKE3_BLOCK_LEN; + input_len -= BLAKE3_BLOCK_LEN; + } + + chunk_state_fill_buf(self, input, input_len); +} + +INLINE output_t chunk_state_output(const blake3_chunk_state *self) { + uint8_t block_flags = + self->flags | chunk_state_maybe_start_flag(self) | CHUNK_END; + return make_output(self->cv, self->buf, self->buf_len, self->chunk_counter, + block_flags); +} + +INLINE output_t parent_output(const uint8_t block[BLAKE3_BLOCK_LEN], + const uint32_t key[8], uint8_t flags) { + return make_output(key, block, BLAKE3_BLOCK_LEN, 0, flags | PARENT); +} + +// Given some input larger than one chunk, return the number of bytes that +// should go in the left subtree. This is the largest power-of-2 number of +// chunks that leaves at least 1 byte for the right subtree. +INLINE size_t left_subtree_len(size_t input_len) { + // Subtract 1 to reserve at least one byte for the right side. input_len + // should always be greater than BLAKE3_CHUNK_LEN. + size_t full_chunks = (input_len - 1) / BLAKE3_CHUNK_LEN; + return round_down_to_power_of_2(full_chunks) * BLAKE3_CHUNK_LEN; +} + +// Use SIMD parallelism to hash up to MAX_SIMD_DEGREE chunks at the same time +// on a single thread. Write out the chunk chaining values and return the +// number of chunks hashed. These chunks are never the root and never empty; +// those cases use a different codepath. +INLINE size_t compress_chunks_parallel(const uint8_t *input, size_t input_len, + const uint32_t key[8], + uint64_t chunk_counter, uint8_t flags, + uint8_t *out) { +#if defined(BLAKE3_TESTING) + assert(0 < input_len); + assert(input_len <= MAX_SIMD_DEGREE * BLAKE3_CHUNK_LEN); +#endif + + const uint8_t *chunks_array[MAX_SIMD_DEGREE]; + size_t input_position = 0; + size_t chunks_array_len = 0; + while (input_len - input_position >= BLAKE3_CHUNK_LEN) { + chunks_array[chunks_array_len] = &input[input_position]; + input_position += BLAKE3_CHUNK_LEN; + chunks_array_len += 1; + } + + blake3_hash_many(chunks_array, chunks_array_len, + BLAKE3_CHUNK_LEN / BLAKE3_BLOCK_LEN, key, chunk_counter, + true, flags, CHUNK_START, CHUNK_END, out); + + // Hash the remaining partial chunk, if there is one. Note that the empty + // chunk (meaning the empty message) is a different codepath. + if (input_len > input_position) { + uint64_t counter = chunk_counter + (uint64_t)chunks_array_len; + blake3_chunk_state chunk_state; + chunk_state_init(&chunk_state, key, flags); + chunk_state.chunk_counter = counter; + chunk_state_update(&chunk_state, &input[input_position], + input_len - input_position); + output_t output = chunk_state_output(&chunk_state); + output_chaining_value(&output, &out[chunks_array_len * BLAKE3_OUT_LEN]); + return chunks_array_len + 1; + } else { + return chunks_array_len; + } +} + +// Use SIMD parallelism to hash up to MAX_SIMD_DEGREE parents at the same time +// on a single thread. Write out the parent chaining values and return the +// number of parents hashed. (If there's an odd input chaining value left over, +// return it as an additional output.) These parents are never the root and +// never empty; those cases use a different codepath. +INLINE size_t compress_parents_parallel(const uint8_t *child_chaining_values, + size_t num_chaining_values, + const uint32_t key[8], uint8_t flags, + uint8_t *out) { +#if defined(BLAKE3_TESTING) + assert(2 <= num_chaining_values); + assert(num_chaining_values <= 2 * MAX_SIMD_DEGREE_OR_2); +#endif + + const uint8_t *parents_array[MAX_SIMD_DEGREE_OR_2]; + size_t parents_array_len = 0; + while (num_chaining_values - (2 * parents_array_len) >= 2) { + parents_array[parents_array_len] = + &child_chaining_values[2 * parents_array_len * BLAKE3_OUT_LEN]; + parents_array_len += 1; + } + + blake3_hash_many(parents_array, parents_array_len, 1, key, + 0, // Parents always use counter 0. + false, flags | PARENT, + 0, // Parents have no start flags. + 0, // Parents have no end flags. + out); + + // If there's an odd child left over, it becomes an output. + if (num_chaining_values > 2 * parents_array_len) { + memcpy(&out[parents_array_len * BLAKE3_OUT_LEN], + &child_chaining_values[2 * parents_array_len * BLAKE3_OUT_LEN], + BLAKE3_OUT_LEN); + return parents_array_len + 1; + } else { + return parents_array_len; + } +} + +// The wide helper function returns (writes out) an array of chaining values +// and returns the length of that array. The number of chaining values returned +// is the dynamically detected SIMD degree, at most MAX_SIMD_DEGREE. Or fewer, +// if the input is shorter than that many chunks. The reason for maintaining a +// wide array of chaining values going back up the tree, is to allow the +// implementation to hash as many parents in parallel as possible. +// +// As a special case when the SIMD degree is 1, this function will still return +// at least 2 outputs. This guarantees that this function doesn't perform the +// root compression. (If it did, it would use the wrong flags, and also we +// wouldn't be able to implement extendable output.) Note that this function is +// not used when the whole input is only 1 chunk long; that's a different +// codepath. +// +// Why not just have the caller split the input on the first update(), instead +// of implementing this special rule? Because we don't want to limit SIMD or +// multi-threading parallelism for that update(). +size_t blake3_compress_subtree_wide(const uint8_t *input, size_t input_len, + const uint32_t key[8], + uint64_t chunk_counter, uint8_t flags, + uint8_t *out, bool use_tbb) { + // Note that the single chunk case does *not* bump the SIMD degree up to 2 + // when it is 1. If this implementation adds multi-threading in the future, + // this gives us the option of multi-threading even the 2-chunk case, which + // can help performance on smaller platforms. + if (input_len <= blake3_simd_degree() * BLAKE3_CHUNK_LEN) { + return compress_chunks_parallel(input, input_len, key, chunk_counter, flags, + out); + } + + // With more than simd_degree chunks, we need to recurse. Start by dividing + // the input into left and right subtrees. (Note that this is only optimal + // as long as the SIMD degree is a power of 2. If we ever get a SIMD degree + // of 3 or something, we'll need a more complicated strategy.) + size_t left_input_len = left_subtree_len(input_len); + size_t right_input_len = input_len - left_input_len; + const uint8_t *right_input = &input[left_input_len]; + uint64_t right_chunk_counter = + chunk_counter + (uint64_t)(left_input_len / BLAKE3_CHUNK_LEN); + + // Make space for the child outputs. Here we use MAX_SIMD_DEGREE_OR_2 to + // account for the special case of returning 2 outputs when the SIMD degree + // is 1. + uint8_t cv_array[2 * MAX_SIMD_DEGREE_OR_2 * BLAKE3_OUT_LEN]; + size_t degree = blake3_simd_degree(); + if (left_input_len > BLAKE3_CHUNK_LEN && degree == 1) { + // The special case: We always use a degree of at least two, to make + // sure there are two outputs. Except, as noted above, at the chunk + // level, where we allow degree=1. (Note that the 1-chunk-input case is + // a different codepath.) + degree = 2; + } + uint8_t *right_cvs = &cv_array[degree * BLAKE3_OUT_LEN]; + + // Recurse! + size_t left_n = SIZE_MAX; + size_t right_n = SIZE_MAX; + +#if defined(BLAKE3_USE_TBB) + blake3_compress_subtree_wide_join_tbb( + key, flags, use_tbb, + // left-hand side + input, left_input_len, chunk_counter, cv_array, &left_n, + // right-hand side + right_input, right_input_len, right_chunk_counter, right_cvs, &right_n); +#else + left_n = blake3_compress_subtree_wide( + input, left_input_len, key, chunk_counter, flags, cv_array, use_tbb); + right_n = blake3_compress_subtree_wide(right_input, right_input_len, key, + right_chunk_counter, flags, right_cvs, + use_tbb); +#endif // BLAKE3_USE_TBB + + // The special case again. If simd_degree=1, then we'll have left_n=1 and + // right_n=1. Rather than compressing them into a single output, return + // them directly, to make sure we always have at least two outputs. + if (left_n == 1) { + memcpy(out, cv_array, 2 * BLAKE3_OUT_LEN); + return 2; + } + + // Otherwise, do one layer of parent node compression. + size_t num_chaining_values = left_n + right_n; + return compress_parents_parallel(cv_array, num_chaining_values, key, flags, + out); +} + +// Hash a subtree with compress_subtree_wide(), and then condense the resulting +// list of chaining values down to a single parent node. Don't compress that +// last parent node, however. Instead, return its message bytes (the +// concatenated chaining values of its children). This is necessary when the +// first call to update() supplies a complete subtree, because the topmost +// parent node of that subtree could end up being the root. It's also necessary +// for extended output in the general case. +// +// As with compress_subtree_wide(), this function is not used on inputs of 1 +// chunk or less. That's a different codepath. +INLINE void +compress_subtree_to_parent_node(const uint8_t *input, size_t input_len, + const uint32_t key[8], uint64_t chunk_counter, + uint8_t flags, uint8_t out[2 * BLAKE3_OUT_LEN], + bool use_tbb) { +#if defined(BLAKE3_TESTING) + assert(input_len > BLAKE3_CHUNK_LEN); +#endif + + uint8_t cv_array[MAX_SIMD_DEGREE_OR_2 * BLAKE3_OUT_LEN]; + size_t num_cvs = blake3_compress_subtree_wide(input, input_len, key, + chunk_counter, flags, cv_array, use_tbb); + assert(num_cvs <= MAX_SIMD_DEGREE_OR_2); + // The following loop never executes when MAX_SIMD_DEGREE_OR_2 is 2, because + // as we just asserted, num_cvs will always be <=2 in that case. But GCC + // (particularly GCC 8.5) can't tell that it never executes, and if NDEBUG is + // set then it emits incorrect warnings here. We tried a few different + // hacks to silence these, but in the end our hacks just produced different + // warnings (see https://github.com/BLAKE3-team/BLAKE3/pull/380). Out of + // desperation, we ifdef out this entire loop when we know it's not needed. +#if MAX_SIMD_DEGREE_OR_2 > 2 + // If MAX_SIMD_DEGREE_OR_2 is greater than 2 and there's enough input, + // compress_subtree_wide() returns more than 2 chaining values. Condense + // them into 2 by forming parent nodes repeatedly. + uint8_t out_array[MAX_SIMD_DEGREE_OR_2 * BLAKE3_OUT_LEN / 2]; + while (num_cvs > 2) { + num_cvs = + compress_parents_parallel(cv_array, num_cvs, key, flags, out_array); + memcpy(cv_array, out_array, num_cvs * BLAKE3_OUT_LEN); + } +#endif + memcpy(out, cv_array, 2 * BLAKE3_OUT_LEN); +} + +INLINE void hasher_init_base(blake3_hasher *self, const uint32_t key[8], + uint8_t flags) { + memcpy(self->key, key, BLAKE3_KEY_LEN); + chunk_state_init(&self->chunk, key, flags); + self->cv_stack_len = 0; +} + +void blake3_hasher_init(blake3_hasher *self) { hasher_init_base(self, IV, 0); } + +void blake3_hasher_init_keyed(blake3_hasher *self, + const uint8_t key[BLAKE3_KEY_LEN]) { + uint32_t key_words[8]; + load_key_words(key, key_words); + hasher_init_base(self, key_words, KEYED_HASH); +} + +void blake3_hasher_init_derive_key_raw(blake3_hasher *self, const void *context, + size_t context_len) { + blake3_hasher context_hasher; + hasher_init_base(&context_hasher, IV, DERIVE_KEY_CONTEXT); + blake3_hasher_update(&context_hasher, context, context_len); + uint8_t context_key[BLAKE3_KEY_LEN]; + blake3_hasher_finalize(&context_hasher, context_key, BLAKE3_KEY_LEN); + uint32_t context_key_words[8]; + load_key_words(context_key, context_key_words); + hasher_init_base(self, context_key_words, DERIVE_KEY_MATERIAL); +} + +void blake3_hasher_init_derive_key(blake3_hasher *self, const char *context) { + blake3_hasher_init_derive_key_raw(self, context, strlen(context)); +} + +// As described in hasher_push_cv() below, we do "lazy merging", delaying +// merges until right before the next CV is about to be added. This is +// different from the reference implementation. Another difference is that we +// aren't always merging 1 chunk at a time. Instead, each CV might represent +// any power-of-two number of chunks, as long as the smaller-above-larger stack +// order is maintained. Instead of the "count the trailing 0-bits" algorithm +// described in the spec, we use a "count the total number of 1-bits" variant +// that doesn't require us to retain the subtree size of the CV on top of the +// stack. The principle is the same: each CV that should remain in the stack is +// represented by a 1-bit in the total number of chunks (or bytes) so far. +INLINE void hasher_merge_cv_stack(blake3_hasher *self, uint64_t total_len) { + size_t post_merge_stack_len = (size_t)popcnt(total_len); + while (self->cv_stack_len > post_merge_stack_len) { + uint8_t *parent_node = + &self->cv_stack[(self->cv_stack_len - 2) * BLAKE3_OUT_LEN]; + output_t output = parent_output(parent_node, self->key, self->chunk.flags); + output_chaining_value(&output, parent_node); + self->cv_stack_len -= 1; + } +} + +// In reference_impl.rs, we merge the new CV with existing CVs from the stack +// before pushing it. We can do that because we know more input is coming, so +// we know none of the merges are root. +// +// This setting is different. We want to feed as much input as possible to +// compress_subtree_wide(), without setting aside anything for the chunk_state. +// If the user gives us 64 KiB, we want to parallelize over all 64 KiB at once +// as a single subtree, if at all possible. +// +// This leads to two problems: +// 1) This 64 KiB input might be the only call that ever gets made to update. +// In this case, the root node of the 64 KiB subtree would be the root node +// of the whole tree, and it would need to be ROOT finalized. We can't +// compress it until we know. +// 2) This 64 KiB input might complete a larger tree, whose root node is +// similarly going to be the root of the whole tree. For example, maybe +// we have 196 KiB (that is, 128 + 64) hashed so far. We can't compress the +// node at the root of the 256 KiB subtree until we know how to finalize it. +// +// The second problem is solved with "lazy merging". That is, when we're about +// to add a CV to the stack, we don't merge it with anything first, as the +// reference impl does. Instead we do merges using the *previous* CV that was +// added, which is sitting on top of the stack, and we put the new CV +// (unmerged) on top of the stack afterwards. This guarantees that we never +// merge the root node until finalize(). +// +// Solving the first problem requires an additional tool, +// compress_subtree_to_parent_node(). That function always returns the top +// *two* chaining values of the subtree it's compressing. We then do lazy +// merging with each of them separately, so that the second CV will always +// remain unmerged. (That also helps us support extendable output when we're +// hashing an input all-at-once.) +INLINE void hasher_push_cv(blake3_hasher *self, uint8_t new_cv[BLAKE3_OUT_LEN], + uint64_t chunk_counter) { + hasher_merge_cv_stack(self, chunk_counter); + memcpy(&self->cv_stack[self->cv_stack_len * BLAKE3_OUT_LEN], new_cv, + BLAKE3_OUT_LEN); + self->cv_stack_len += 1; +} + +INLINE void blake3_hasher_update_base(blake3_hasher *self, const void *input, + size_t input_len, bool use_tbb) { + // Explicitly checking for zero avoids causing UB by passing a null pointer + // to memcpy. This comes up in practice with things like: + // std::vector v; + // blake3_hasher_update(&hasher, v.data(), v.size()); + if (input_len == 0) { + return; + } + + const uint8_t *input_bytes = (const uint8_t *)input; + + // If we have some partial chunk bytes in the internal chunk_state, we need + // to finish that chunk first. + if (chunk_state_len(&self->chunk) > 0) { + size_t take = BLAKE3_CHUNK_LEN - chunk_state_len(&self->chunk); + if (take > input_len) { + take = input_len; + } + chunk_state_update(&self->chunk, input_bytes, take); + input_bytes += take; + input_len -= take; + // If we've filled the current chunk and there's more coming, finalize this + // chunk and proceed. In this case we know it's not the root. + if (input_len > 0) { + output_t output = chunk_state_output(&self->chunk); + uint8_t chunk_cv[32]; + output_chaining_value(&output, chunk_cv); + hasher_push_cv(self, chunk_cv, self->chunk.chunk_counter); + chunk_state_reset(&self->chunk, self->key, self->chunk.chunk_counter + 1); + } else { + return; + } + } + + // Now the chunk_state is clear, and we have more input. If there's more than + // a single chunk (so, definitely not the root chunk), hash the largest whole + // subtree we can, with the full benefits of SIMD (and maybe in the future, + // multi-threading) parallelism. Two restrictions: + // - The subtree has to be a power-of-2 number of chunks. Only subtrees along + // the right edge can be incomplete, and we don't know where the right edge + // is going to be until we get to finalize(). + // - The subtree must evenly divide the total number of chunks up until this + // point (if total is not 0). If the current incomplete subtree is only + // waiting for 1 more chunk, we can't hash a subtree of 4 chunks. We have + // to complete the current subtree first. + // Because we might need to break up the input to form powers of 2, or to + // evenly divide what we already have, this part runs in a loop. + while (input_len > BLAKE3_CHUNK_LEN) { + size_t subtree_len = round_down_to_power_of_2(input_len); + uint64_t count_so_far = self->chunk.chunk_counter * BLAKE3_CHUNK_LEN; + // Shrink the subtree_len until it evenly divides the count so far. We know + // that subtree_len itself is a power of 2, so we can use a bitmasking + // trick instead of an actual remainder operation. (Note that if the caller + // consistently passes power-of-2 inputs of the same size, as is hopefully + // typical, this loop condition will always fail, and subtree_len will + // always be the full length of the input.) + // + // An aside: We don't have to shrink subtree_len quite this much. For + // example, if count_so_far is 1, we could pass 2 chunks to + // compress_subtree_to_parent_node. Since we'll get 2 CVs back, we'll still + // get the right answer in the end, and we might get to use 2-way SIMD + // parallelism. The problem with this optimization, is that it gets us + // stuck always hashing 2 chunks. The total number of chunks will remain + // odd, and we'll never graduate to higher degrees of parallelism. See + // https://github.com/BLAKE3-team/BLAKE3/issues/69. + while ((((uint64_t)(subtree_len - 1)) & count_so_far) != 0) { + subtree_len /= 2; + } + // The shrunken subtree_len might now be 1 chunk long. If so, hash that one + // chunk by itself. Otherwise, compress the subtree into a pair of CVs. + uint64_t subtree_chunks = subtree_len / BLAKE3_CHUNK_LEN; + if (subtree_len <= BLAKE3_CHUNK_LEN) { + blake3_chunk_state chunk_state; + chunk_state_init(&chunk_state, self->key, self->chunk.flags); + chunk_state.chunk_counter = self->chunk.chunk_counter; + chunk_state_update(&chunk_state, input_bytes, subtree_len); + output_t output = chunk_state_output(&chunk_state); + uint8_t cv[BLAKE3_OUT_LEN]; + output_chaining_value(&output, cv); + hasher_push_cv(self, cv, chunk_state.chunk_counter); + } else { + // This is the high-performance happy path, though getting here depends + // on the caller giving us a long enough input. + uint8_t cv_pair[2 * BLAKE3_OUT_LEN]; + compress_subtree_to_parent_node(input_bytes, subtree_len, self->key, + self->chunk.chunk_counter, + self->chunk.flags, cv_pair, use_tbb); + hasher_push_cv(self, cv_pair, self->chunk.chunk_counter); + hasher_push_cv(self, &cv_pair[BLAKE3_OUT_LEN], + self->chunk.chunk_counter + (subtree_chunks / 2)); + } + self->chunk.chunk_counter += subtree_chunks; + input_bytes += subtree_len; + input_len -= subtree_len; + } + + // If there's any remaining input less than a full chunk, add it to the chunk + // state. In that case, also do a final merge loop to make sure the subtree + // stack doesn't contain any unmerged pairs. The remaining input means we + // know these merges are non-root. This merge loop isn't strictly necessary + // here, because hasher_push_chunk_cv already does its own merge loop, but it + // simplifies blake3_hasher_finalize below. + if (input_len > 0) { + chunk_state_update(&self->chunk, input_bytes, input_len); + hasher_merge_cv_stack(self, self->chunk.chunk_counter); + } +} + +void blake3_hasher_update(blake3_hasher *self, const void *input, + size_t input_len) { + bool use_tbb = false; + blake3_hasher_update_base(self, input, input_len, use_tbb); +} + +#if defined(BLAKE3_USE_TBB) +void blake3_hasher_update_tbb(blake3_hasher *self, const void *input, + size_t input_len) { + bool use_tbb = true; + blake3_hasher_update_base(self, input, input_len, use_tbb); +} +#endif // BLAKE3_USE_TBB + +void blake3_hasher_finalize(const blake3_hasher *self, uint8_t *out, + size_t out_len) { + blake3_hasher_finalize_seek(self, 0, out, out_len); +} + +void blake3_hasher_finalize_seek(const blake3_hasher *self, uint64_t seek, + uint8_t *out, size_t out_len) { + // Explicitly checking for zero avoids causing UB by passing a null pointer + // to memcpy. This comes up in practice with things like: + // std::vector v; + // blake3_hasher_finalize(&hasher, v.data(), v.size()); + if (out_len == 0) { + return; + } + + // If the subtree stack is empty, then the current chunk is the root. + if (self->cv_stack_len == 0) { + output_t output = chunk_state_output(&self->chunk); + output_root_bytes(&output, seek, out, out_len); + return; + } + // If there are any bytes in the chunk state, finalize that chunk and do a + // roll-up merge between that chunk hash and every subtree in the stack. In + // this case, the extra merge loop at the end of blake3_hasher_update + // guarantees that none of the subtrees in the stack need to be merged with + // each other first. Otherwise, if there are no bytes in the chunk state, + // then the top of the stack is a chunk hash, and we start the merge from + // that. + output_t output; + size_t cvs_remaining; + if (chunk_state_len(&self->chunk) > 0) { + cvs_remaining = self->cv_stack_len; + output = chunk_state_output(&self->chunk); + } else { + // There are always at least 2 CVs in the stack in this case. + cvs_remaining = self->cv_stack_len - 2; + output = parent_output(&self->cv_stack[cvs_remaining * 32], self->key, + self->chunk.flags); + } + while (cvs_remaining > 0) { + cvs_remaining -= 1; + uint8_t parent_block[BLAKE3_BLOCK_LEN]; + memcpy(parent_block, &self->cv_stack[cvs_remaining * 32], 32); + output_chaining_value(&output, &parent_block[32]); + output = parent_output(parent_block, self->key, self->chunk.flags); + } + output_root_bytes(&output, seek, out, out_len); +} + +void blake3_hasher_reset(blake3_hasher *self) { + chunk_state_reset(&self->chunk, self->key, 0); + self->cv_stack_len = 0; +} diff --git a/ext/digest/blake3/blake3.h b/ext/digest/blake3/blake3.h new file mode 100644 index 0000000..423154f --- /dev/null +++ b/ext/digest/blake3/blake3.h @@ -0,0 +1,86 @@ +#ifndef BLAKE3_H +#define BLAKE3_H + +#include +#include + +#if !defined(BLAKE3_API) +# if defined(_WIN32) || defined(__CYGWIN__) +# if defined(BLAKE3_DLL) +# if defined(BLAKE3_DLL_EXPORTS) +# define BLAKE3_API __declspec(dllexport) +# else +# define BLAKE3_API __declspec(dllimport) +# endif +# define BLAKE3_PRIVATE +# else +# define BLAKE3_API +# define BLAKE3_PRIVATE +# endif +# elif __GNUC__ >= 4 +# define BLAKE3_API __attribute__((visibility("default"))) +# define BLAKE3_PRIVATE __attribute__((visibility("hidden"))) +# else +# define BLAKE3_API +# define BLAKE3_PRIVATE +# endif +#endif + +#ifdef __cplusplus +extern "C" { +#endif + +#define BLAKE3_VERSION_STRING "1.8.5" +#define BLAKE3_KEY_LEN 32 +#define BLAKE3_OUT_LEN 32 +#define BLAKE3_BLOCK_LEN 64 +#define BLAKE3_CHUNK_LEN 1024 +#define BLAKE3_MAX_DEPTH 54 + +// This struct is a private implementation detail. It has to be here because +// it's part of the blake3_hasher structure defined below. +typedef struct { + uint32_t cv[8]; + uint64_t chunk_counter; + uint8_t buf[BLAKE3_BLOCK_LEN]; + uint8_t buf_len; + uint8_t blocks_compressed; + uint8_t flags; +} blake3_chunk_state; + +typedef struct { + uint32_t key[8]; + blake3_chunk_state chunk; + uint8_t cv_stack_len; + // The stack size is MAX_DEPTH + 1 because we do lazy merging. For example, + // with 7 chunks, we have 3 entries in the stack. Adding an 8th chunk + // requires a 4th entry, rather than merging everything down to 1, because we + // don't know whether more input is coming. This is different from how the + // reference implementation does things. + uint8_t cv_stack[(BLAKE3_MAX_DEPTH + 1) * BLAKE3_OUT_LEN]; +} blake3_hasher; + +BLAKE3_API const char *blake3_version(void); +BLAKE3_API void blake3_hasher_init(blake3_hasher *self); +BLAKE3_API void blake3_hasher_init_keyed(blake3_hasher *self, + const uint8_t key[BLAKE3_KEY_LEN]); +BLAKE3_API void blake3_hasher_init_derive_key(blake3_hasher *self, const char *context); +BLAKE3_API void blake3_hasher_init_derive_key_raw(blake3_hasher *self, const void *context, + size_t context_len); +BLAKE3_API void blake3_hasher_update(blake3_hasher *self, const void *input, + size_t input_len); +#if defined(BLAKE3_USE_TBB) +BLAKE3_API void blake3_hasher_update_tbb(blake3_hasher *self, const void *input, + size_t input_len); +#endif // BLAKE3_USE_TBB +BLAKE3_API void blake3_hasher_finalize(const blake3_hasher *self, uint8_t *out, + size_t out_len); +BLAKE3_API void blake3_hasher_finalize_seek(const blake3_hasher *self, uint64_t seek, + uint8_t *out, size_t out_len); +BLAKE3_API void blake3_hasher_reset(blake3_hasher *self); + +#ifdef __cplusplus +} +#endif + +#endif /* BLAKE3_H */ diff --git a/ext/digest/blake3/blake3_dispatch.c b/ext/digest/blake3/blake3_dispatch.c new file mode 100644 index 0000000..14dfbbe --- /dev/null +++ b/ext/digest/blake3/blake3_dispatch.c @@ -0,0 +1,332 @@ +#include +#include +#include + +#include "blake3_impl.h" + +#if defined(_MSC_VER) +#include +#endif + +#if defined(IS_X86) +#if defined(_MSC_VER) +#include +#elif defined(__GNUC__) +#include +#else +#undef IS_X86 /* Unimplemented! */ +#endif +#endif + +#if !defined(BLAKE3_ATOMICS) +#if defined(__has_include) +#if __has_include() && !defined(_MSC_VER) +#define BLAKE3_ATOMICS 1 +#else +#define BLAKE3_ATOMICS 0 +#endif /* __has_include() && !defined(_MSC_VER) */ +#else +#define BLAKE3_ATOMICS 0 +#endif /* defined(__has_include) */ +#endif /* BLAKE3_ATOMICS */ + +#if BLAKE3_ATOMICS +#define ATOMIC_INT _Atomic int +#define ATOMIC_LOAD(x) x +#define ATOMIC_STORE(x, y) x = y +#elif defined(_MSC_VER) +#define ATOMIC_INT LONG +#define ATOMIC_LOAD(x) InterlockedOr(&x, 0) +#define ATOMIC_STORE(x, y) InterlockedExchange(&x, y) +#else +#define ATOMIC_INT int +#define ATOMIC_LOAD(x) x +#define ATOMIC_STORE(x, y) x = y +#endif + +#define MAYBE_UNUSED(x) (void)((x)) + +#if defined(IS_X86) +static uint64_t xgetbv(void) { +#if defined(_MSC_VER) + return _xgetbv(0); +#else + uint32_t eax = 0, edx = 0; + __asm__ __volatile__("xgetbv\n" : "=a"(eax), "=d"(edx) : "c"(0)); + return ((uint64_t)edx << 32) | eax; +#endif +} + +static void cpuid(uint32_t out[4], uint32_t id) { +#if defined(_MSC_VER) + __cpuid((int *)out, id); +#elif defined(__i386__) || defined(_M_IX86) + __asm__ __volatile__("movl %%ebx, %1\n" + "cpuid\n" + "xchgl %1, %%ebx\n" + : "=a"(out[0]), "=r"(out[1]), "=c"(out[2]), "=d"(out[3]) + : "a"(id)); +#else + __asm__ __volatile__("cpuid\n" + : "=a"(out[0]), "=b"(out[1]), "=c"(out[2]), "=d"(out[3]) + : "a"(id)); +#endif +} + +static void cpuidex(uint32_t out[4], uint32_t id, uint32_t sid) { +#if defined(_MSC_VER) + __cpuidex((int *)out, id, sid); +#elif defined(__i386__) || defined(_M_IX86) + __asm__ __volatile__("movl %%ebx, %1\n" + "cpuid\n" + "xchgl %1, %%ebx\n" + : "=a"(out[0]), "=r"(out[1]), "=c"(out[2]), "=d"(out[3]) + : "a"(id), "c"(sid)); +#else + __asm__ __volatile__("cpuid\n" + : "=a"(out[0]), "=b"(out[1]), "=c"(out[2]), "=d"(out[3]) + : "a"(id), "c"(sid)); +#endif +} + + +enum cpu_feature { + SSE2 = 1 << 0, + SSSE3 = 1 << 1, + SSE41 = 1 << 2, + AVX = 1 << 3, + AVX2 = 1 << 4, + AVX512F = 1 << 5, + AVX512VL = 1 << 6, + /* ... */ + UNDEFINED = 1 << 30 +}; + +#if !defined(BLAKE3_TESTING) +static /* Allow the variable to be controlled manually for testing */ +#endif + ATOMIC_INT g_cpu_features = UNDEFINED; + +#if !defined(BLAKE3_TESTING) +static +#endif + enum cpu_feature + get_cpu_features(void) { + + /* If TSAN detects a data race here, try compiling with -DBLAKE3_ATOMICS=1 */ + enum cpu_feature features = ATOMIC_LOAD(g_cpu_features); + if (features != UNDEFINED) { + return features; + } else { +#if defined(IS_X86) + uint32_t regs[4] = {0}; + uint32_t *eax = ®s[0], *ebx = ®s[1], *ecx = ®s[2], *edx = ®s[3]; + (void)edx; + features = 0; + cpuid(regs, 0); + const int max_id = *eax; + cpuid(regs, 1); +#if defined(__amd64__) || defined(_M_X64) + features |= SSE2; +#else + if (*edx & (1UL << 26)) + features |= SSE2; +#endif + if (*ecx & (1UL << 9)) + features |= SSSE3; + if (*ecx & (1UL << 19)) + features |= SSE41; + + if (*ecx & (1UL << 27)) { // OSXSAVE + const uint64_t mask = xgetbv(); + if ((mask & 6) == 6) { // SSE and AVX states + if (*ecx & (1UL << 28)) + features |= AVX; + if (max_id >= 7) { + cpuidex(regs, 7, 0); + if (*ebx & (1UL << 5)) + features |= AVX2; + if ((mask & 224) == 224) { // Opmask, ZMM_Hi256, Hi16_Zmm + if (*ebx & (1UL << 31)) + features |= AVX512VL; + if (*ebx & (1UL << 16)) + features |= AVX512F; + } + } + } + } + ATOMIC_STORE(g_cpu_features, features); + return features; +#else + /* How to detect NEON? */ + return 0; +#endif + } +} +#endif + +void blake3_compress_in_place(uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags) { +#if defined(IS_X86) + const enum cpu_feature features = get_cpu_features(); + MAYBE_UNUSED(features); +#if !defined(BLAKE3_NO_AVX512) + if (features & AVX512VL) { + blake3_compress_in_place_avx512(cv, block, block_len, counter, flags); + return; + } +#endif +#if !defined(BLAKE3_NO_SSE41) + if (features & SSE41) { + blake3_compress_in_place_sse41(cv, block, block_len, counter, flags); + return; + } +#endif +#if !defined(BLAKE3_NO_SSE2) + if (features & SSE2) { + blake3_compress_in_place_sse2(cv, block, block_len, counter, flags); + return; + } +#endif +#endif + blake3_compress_in_place_portable(cv, block, block_len, counter, flags); +} + +void blake3_compress_xof(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, uint8_t flags, + uint8_t out[64]) { +#if defined(IS_X86) + const enum cpu_feature features = get_cpu_features(); + MAYBE_UNUSED(features); +#if !defined(BLAKE3_NO_AVX512) + if (features & AVX512VL) { + blake3_compress_xof_avx512(cv, block, block_len, counter, flags, out); + return; + } +#endif +#if !defined(BLAKE3_NO_SSE41) + if (features & SSE41) { + blake3_compress_xof_sse41(cv, block, block_len, counter, flags, out); + return; + } +#endif +#if !defined(BLAKE3_NO_SSE2) + if (features & SSE2) { + blake3_compress_xof_sse2(cv, block, block_len, counter, flags, out); + return; + } +#endif +#endif + blake3_compress_xof_portable(cv, block, block_len, counter, flags, out); +} + + +void blake3_xof_many(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, uint8_t flags, + uint8_t out[64], size_t outblocks) { + if (outblocks == 0) { + // The current assembly implementation always outputs at least 1 block. + return; + } +#if defined(IS_X86) + const enum cpu_feature features = get_cpu_features(); + MAYBE_UNUSED(features); +#if !defined(_WIN32) && !defined(__CYGWIN__) && !defined(BLAKE3_NO_AVX512) + if (features & AVX512VL) { + blake3_xof_many_avx512(cv, block, block_len, counter, flags, out, outblocks); + return; + } +#endif +#endif + for(size_t i = 0; i < outblocks; ++i) { + blake3_compress_xof(cv, block, block_len, counter + i, flags, out + 64*i); + } +} + +void blake3_hash_many(const uint8_t *const *inputs, size_t num_inputs, + size_t blocks, const uint32_t key[8], uint64_t counter, + bool increment_counter, uint8_t flags, + uint8_t flags_start, uint8_t flags_end, uint8_t *out) { +#if defined(IS_X86) + const enum cpu_feature features = get_cpu_features(); + MAYBE_UNUSED(features); +#if !defined(BLAKE3_NO_AVX512) + if ((features & (AVX512F|AVX512VL)) == (AVX512F|AVX512VL)) { + blake3_hash_many_avx512(inputs, num_inputs, blocks, key, counter, + increment_counter, flags, flags_start, flags_end, + out); + return; + } +#endif +#if !defined(BLAKE3_NO_AVX2) + if (features & AVX2) { + blake3_hash_many_avx2(inputs, num_inputs, blocks, key, counter, + increment_counter, flags, flags_start, flags_end, + out); + return; + } +#endif +#if !defined(BLAKE3_NO_SSE41) + if (features & SSE41) { + blake3_hash_many_sse41(inputs, num_inputs, blocks, key, counter, + increment_counter, flags, flags_start, flags_end, + out); + return; + } +#endif +#if !defined(BLAKE3_NO_SSE2) + if (features & SSE2) { + blake3_hash_many_sse2(inputs, num_inputs, blocks, key, counter, + increment_counter, flags, flags_start, flags_end, + out); + return; + } +#endif +#endif + +#if BLAKE3_USE_NEON == 1 + blake3_hash_many_neon(inputs, num_inputs, blocks, key, counter, + increment_counter, flags, flags_start, flags_end, out); + return; +#endif + + blake3_hash_many_portable(inputs, num_inputs, blocks, key, counter, + increment_counter, flags, flags_start, flags_end, + out); +} + +// The dynamically detected SIMD degree of the current platform. +size_t blake3_simd_degree(void) { +#if defined(IS_X86) + const enum cpu_feature features = get_cpu_features(); + MAYBE_UNUSED(features); +#if !defined(BLAKE3_NO_AVX512) + if ((features & (AVX512F|AVX512VL)) == (AVX512F|AVX512VL)) { + return 16; + } +#endif +#if !defined(BLAKE3_NO_AVX2) + if (features & AVX2) { + return 8; + } +#endif +#if !defined(BLAKE3_NO_SSE41) + if (features & SSE41) { + return 4; + } +#endif +#if !defined(BLAKE3_NO_SSE2) + if (features & SSE2) { + return 4; + } +#endif +#endif +#if BLAKE3_USE_NEON == 1 + return 4; +#endif + return 1; +} diff --git a/ext/digest/blake3/blake3_impl.h b/ext/digest/blake3/blake3_impl.h new file mode 100644 index 0000000..88e71e4 --- /dev/null +++ b/ext/digest/blake3/blake3_impl.h @@ -0,0 +1,333 @@ +#ifndef BLAKE3_IMPL_H +#define BLAKE3_IMPL_H + +#include +#include +#include +#include +#include + +#include "blake3.h" + +#ifdef __cplusplus +extern "C" { +#endif + +// internal flags +enum blake3_flags { + CHUNK_START = 1 << 0, + CHUNK_END = 1 << 1, + PARENT = 1 << 2, + ROOT = 1 << 3, + KEYED_HASH = 1 << 4, + DERIVE_KEY_CONTEXT = 1 << 5, + DERIVE_KEY_MATERIAL = 1 << 6, +}; + +// This C implementation tries to support recent versions of GCC, Clang, and +// MSVC. +#if defined(_MSC_VER) +#define INLINE static __forceinline +#else +#define INLINE static inline __attribute__((always_inline)) +#endif + +#ifdef __cplusplus +#define NOEXCEPT noexcept +#else +#define NOEXCEPT +#endif + +#if (defined(__x86_64__) || defined(_M_X64)) && !defined(_M_ARM64EC) +#define IS_X86 +#define IS_X86_64 +#endif + +#if defined(__i386__) || defined(_M_IX86) +#define IS_X86 +#define IS_X86_32 +#endif + +#if defined(__aarch64__) || defined(_M_ARM64) || defined(_M_ARM64EC) +#define IS_AARCH64 +#endif + +#if defined(IS_X86) +#if defined(_MSC_VER) +#include +#endif +#endif + +#if !defined(BLAKE3_USE_NEON) + // If BLAKE3_USE_NEON not manually set, autodetect based on AArch64ness + #if defined(IS_AARCH64) + #if defined(__ARM_BIG_ENDIAN) + #define BLAKE3_USE_NEON 0 + #else + #define BLAKE3_USE_NEON 1 + #endif + #else + #define BLAKE3_USE_NEON 0 + #endif +#endif + +#if defined(IS_X86) +#define MAX_SIMD_DEGREE 16 +#elif BLAKE3_USE_NEON == 1 +#define MAX_SIMD_DEGREE 4 +#else +#define MAX_SIMD_DEGREE 1 +#endif + +// There are some places where we want a static size that's equal to the +// MAX_SIMD_DEGREE, but also at least 2. +#define MAX_SIMD_DEGREE_OR_2 (MAX_SIMD_DEGREE > 2 ? MAX_SIMD_DEGREE : 2) + +static const uint32_t IV[8] = {0x6A09E667UL, 0xBB67AE85UL, 0x3C6EF372UL, + 0xA54FF53AUL, 0x510E527FUL, 0x9B05688CUL, + 0x1F83D9ABUL, 0x5BE0CD19UL}; + +static const uint8_t MSG_SCHEDULE[7][16] = { + {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15}, + {2, 6, 3, 10, 7, 0, 4, 13, 1, 11, 12, 5, 9, 14, 15, 8}, + {3, 4, 10, 12, 13, 2, 7, 14, 6, 5, 9, 0, 11, 15, 8, 1}, + {10, 7, 12, 9, 14, 3, 13, 15, 4, 0, 11, 2, 5, 8, 1, 6}, + {12, 13, 9, 11, 15, 10, 14, 8, 7, 2, 5, 3, 0, 1, 6, 4}, + {9, 14, 11, 5, 8, 12, 15, 1, 13, 3, 0, 10, 2, 6, 4, 7}, + {11, 15, 5, 0, 1, 9, 8, 6, 14, 10, 2, 12, 3, 4, 7, 13}, +}; + +/* Find index of the highest set bit */ +/* x is assumed to be nonzero. */ +static unsigned int highest_one(uint64_t x) { +#if defined(__GNUC__) || defined(__clang__) + return 63 ^ (unsigned int)__builtin_clzll(x); +#elif defined(_MSC_VER) && defined(IS_X86_64) + unsigned long index; + _BitScanReverse64(&index, x); + return index; +#elif defined(_MSC_VER) && defined(IS_X86_32) + if(x >> 32) { + unsigned long index; + _BitScanReverse(&index, (unsigned long)(x >> 32)); + return 32 + index; + } else { + unsigned long index; + _BitScanReverse(&index, (unsigned long)x); + return index; + } +#else + unsigned int c = 0; + if(x & 0xffffffff00000000ULL) { x >>= 32; c += 32; } + if(x & 0x00000000ffff0000ULL) { x >>= 16; c += 16; } + if(x & 0x000000000000ff00ULL) { x >>= 8; c += 8; } + if(x & 0x00000000000000f0ULL) { x >>= 4; c += 4; } + if(x & 0x000000000000000cULL) { x >>= 2; c += 2; } + if(x & 0x0000000000000002ULL) { c += 1; } + return c; +#endif +} + +// Count the number of 1 bits. +INLINE unsigned int popcnt(uint64_t x) { +#if defined(__GNUC__) || defined(__clang__) + return (unsigned int)__builtin_popcountll(x); +#else + unsigned int count = 0; + while (x != 0) { + count += 1; + x &= x - 1; + } + return count; +#endif +} + +// Largest power of two less than or equal to x. As a special case, returns 1 +// when x is 0. +INLINE uint64_t round_down_to_power_of_2(uint64_t x) { + return 1ULL << highest_one(x | 1); +} + +INLINE uint32_t counter_low(uint64_t counter) { return (uint32_t)counter; } + +INLINE uint32_t counter_high(uint64_t counter) { + return (uint32_t)(counter >> 32); +} + +INLINE uint32_t load32(const void *src) { + const uint8_t *p = (const uint8_t *)src; + return ((uint32_t)(p[0]) << 0) | ((uint32_t)(p[1]) << 8) | + ((uint32_t)(p[2]) << 16) | ((uint32_t)(p[3]) << 24); +} + +INLINE void load_key_words(const uint8_t key[BLAKE3_KEY_LEN], + uint32_t key_words[8]) { + key_words[0] = load32(&key[0 * 4]); + key_words[1] = load32(&key[1 * 4]); + key_words[2] = load32(&key[2 * 4]); + key_words[3] = load32(&key[3 * 4]); + key_words[4] = load32(&key[4 * 4]); + key_words[5] = load32(&key[5 * 4]); + key_words[6] = load32(&key[6 * 4]); + key_words[7] = load32(&key[7 * 4]); +} + +INLINE void load_block_words(const uint8_t block[BLAKE3_BLOCK_LEN], + uint32_t block_words[16]) { + for (size_t i = 0; i < 16; i++) { + block_words[i] = load32(&block[i * 4]); + } +} + +INLINE void store32(void *dst, uint32_t w) { + uint8_t *p = (uint8_t *)dst; + p[0] = (uint8_t)(w >> 0); + p[1] = (uint8_t)(w >> 8); + p[2] = (uint8_t)(w >> 16); + p[3] = (uint8_t)(w >> 24); +} + +INLINE void store_cv_words(uint8_t bytes_out[32], uint32_t cv_words[8]) { + store32(&bytes_out[0 * 4], cv_words[0]); + store32(&bytes_out[1 * 4], cv_words[1]); + store32(&bytes_out[2 * 4], cv_words[2]); + store32(&bytes_out[3 * 4], cv_words[3]); + store32(&bytes_out[4 * 4], cv_words[4]); + store32(&bytes_out[5 * 4], cv_words[5]); + store32(&bytes_out[6 * 4], cv_words[6]); + store32(&bytes_out[7 * 4], cv_words[7]); +} + +void blake3_compress_in_place(uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags); + +void blake3_compress_xof(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, uint8_t flags, + uint8_t out[64]); + +void blake3_xof_many(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, uint8_t flags, + uint8_t out[64], size_t outblocks); + +void blake3_hash_many(const uint8_t *const *inputs, size_t num_inputs, + size_t blocks, const uint32_t key[8], uint64_t counter, + bool increment_counter, uint8_t flags, + uint8_t flags_start, uint8_t flags_end, uint8_t *out); + +size_t blake3_simd_degree(void); + +BLAKE3_PRIVATE size_t blake3_compress_subtree_wide(const uint8_t *input, size_t input_len, + const uint32_t key[8], + uint64_t chunk_counter, uint8_t flags, + uint8_t *out, bool use_tbb); + +#if defined(BLAKE3_USE_TBB) +BLAKE3_PRIVATE void blake3_compress_subtree_wide_join_tbb( + // shared params + const uint32_t key[8], uint8_t flags, bool use_tbb, + // left-hand side params + const uint8_t *l_input, size_t l_input_len, uint64_t l_chunk_counter, + uint8_t *l_cvs, size_t *l_n, + // right-hand side params + const uint8_t *r_input, size_t r_input_len, uint64_t r_chunk_counter, + uint8_t *r_cvs, size_t *r_n) NOEXCEPT; +#endif + +// Declarations for implementation-specific functions. +void blake3_compress_in_place_portable(uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags); + +void blake3_compress_xof_portable(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags, uint8_t out[64]); + +void blake3_hash_many_portable(const uint8_t *const *inputs, size_t num_inputs, + size_t blocks, const uint32_t key[8], + uint64_t counter, bool increment_counter, + uint8_t flags, uint8_t flags_start, + uint8_t flags_end, uint8_t *out); + +#if defined(IS_X86) +#if !defined(BLAKE3_NO_SSE2) +void blake3_compress_in_place_sse2(uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags); +void blake3_compress_xof_sse2(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags, uint8_t out[64]); +void blake3_hash_many_sse2(const uint8_t *const *inputs, size_t num_inputs, + size_t blocks, const uint32_t key[8], + uint64_t counter, bool increment_counter, + uint8_t flags, uint8_t flags_start, + uint8_t flags_end, uint8_t *out); +#endif +#if !defined(BLAKE3_NO_SSE41) +void blake3_compress_in_place_sse41(uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags); +void blake3_compress_xof_sse41(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags, uint8_t out[64]); +void blake3_hash_many_sse41(const uint8_t *const *inputs, size_t num_inputs, + size_t blocks, const uint32_t key[8], + uint64_t counter, bool increment_counter, + uint8_t flags, uint8_t flags_start, + uint8_t flags_end, uint8_t *out); +#endif +#if !defined(BLAKE3_NO_AVX2) +void blake3_hash_many_avx2(const uint8_t *const *inputs, size_t num_inputs, + size_t blocks, const uint32_t key[8], + uint64_t counter, bool increment_counter, + uint8_t flags, uint8_t flags_start, + uint8_t flags_end, uint8_t *out); +#endif +#if !defined(BLAKE3_NO_AVX512) +void blake3_compress_in_place_avx512(uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags); + +void blake3_compress_xof_avx512(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags, uint8_t out[64]); + +void blake3_hash_many_avx512(const uint8_t *const *inputs, size_t num_inputs, + size_t blocks, const uint32_t key[8], + uint64_t counter, bool increment_counter, + uint8_t flags, uint8_t flags_start, + uint8_t flags_end, uint8_t *out); + +#if !defined(_WIN32) && !defined(__CYGWIN__) +void blake3_xof_many_avx512(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, uint8_t flags, + uint8_t* out, size_t outblocks); +#endif +#endif +#endif + +#if BLAKE3_USE_NEON == 1 +void blake3_hash_many_neon(const uint8_t *const *inputs, size_t num_inputs, + size_t blocks, const uint32_t key[8], + uint64_t counter, bool increment_counter, + uint8_t flags, uint8_t flags_start, + uint8_t flags_end, uint8_t *out); +#endif + +#ifdef __cplusplus +} +#endif + +#endif /* BLAKE3_IMPL_H */ diff --git a/ext/digest/blake3/blake3_portable.c b/ext/digest/blake3/blake3_portable.c new file mode 100644 index 0000000..062dd1b --- /dev/null +++ b/ext/digest/blake3/blake3_portable.c @@ -0,0 +1,160 @@ +#include "blake3_impl.h" +#include + +INLINE uint32_t rotr32(uint32_t w, uint32_t c) { + return (w >> c) | (w << (32 - c)); +} + +INLINE void g(uint32_t *state, size_t a, size_t b, size_t c, size_t d, + uint32_t x, uint32_t y) { + state[a] = state[a] + state[b] + x; + state[d] = rotr32(state[d] ^ state[a], 16); + state[c] = state[c] + state[d]; + state[b] = rotr32(state[b] ^ state[c], 12); + state[a] = state[a] + state[b] + y; + state[d] = rotr32(state[d] ^ state[a], 8); + state[c] = state[c] + state[d]; + state[b] = rotr32(state[b] ^ state[c], 7); +} + +INLINE void round_fn(uint32_t state[16], const uint32_t *msg, size_t round) { + // Select the message schedule based on the round. + const uint8_t *schedule = MSG_SCHEDULE[round]; + + // Mix the columns. + g(state, 0, 4, 8, 12, msg[schedule[0]], msg[schedule[1]]); + g(state, 1, 5, 9, 13, msg[schedule[2]], msg[schedule[3]]); + g(state, 2, 6, 10, 14, msg[schedule[4]], msg[schedule[5]]); + g(state, 3, 7, 11, 15, msg[schedule[6]], msg[schedule[7]]); + + // Mix the rows. + g(state, 0, 5, 10, 15, msg[schedule[8]], msg[schedule[9]]); + g(state, 1, 6, 11, 12, msg[schedule[10]], msg[schedule[11]]); + g(state, 2, 7, 8, 13, msg[schedule[12]], msg[schedule[13]]); + g(state, 3, 4, 9, 14, msg[schedule[14]], msg[schedule[15]]); +} + +INLINE void compress_pre(uint32_t state[16], const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, uint8_t flags) { + uint32_t block_words[16]; + block_words[0] = load32(block + 4 * 0); + block_words[1] = load32(block + 4 * 1); + block_words[2] = load32(block + 4 * 2); + block_words[3] = load32(block + 4 * 3); + block_words[4] = load32(block + 4 * 4); + block_words[5] = load32(block + 4 * 5); + block_words[6] = load32(block + 4 * 6); + block_words[7] = load32(block + 4 * 7); + block_words[8] = load32(block + 4 * 8); + block_words[9] = load32(block + 4 * 9); + block_words[10] = load32(block + 4 * 10); + block_words[11] = load32(block + 4 * 11); + block_words[12] = load32(block + 4 * 12); + block_words[13] = load32(block + 4 * 13); + block_words[14] = load32(block + 4 * 14); + block_words[15] = load32(block + 4 * 15); + + state[0] = cv[0]; + state[1] = cv[1]; + state[2] = cv[2]; + state[3] = cv[3]; + state[4] = cv[4]; + state[5] = cv[5]; + state[6] = cv[6]; + state[7] = cv[7]; + state[8] = IV[0]; + state[9] = IV[1]; + state[10] = IV[2]; + state[11] = IV[3]; + state[12] = counter_low(counter); + state[13] = counter_high(counter); + state[14] = (uint32_t)block_len; + state[15] = (uint32_t)flags; + + round_fn(state, &block_words[0], 0); + round_fn(state, &block_words[0], 1); + round_fn(state, &block_words[0], 2); + round_fn(state, &block_words[0], 3); + round_fn(state, &block_words[0], 4); + round_fn(state, &block_words[0], 5); + round_fn(state, &block_words[0], 6); +} + +void blake3_compress_in_place_portable(uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags) { + uint32_t state[16]; + compress_pre(state, cv, block, block_len, counter, flags); + cv[0] = state[0] ^ state[8]; + cv[1] = state[1] ^ state[9]; + cv[2] = state[2] ^ state[10]; + cv[3] = state[3] ^ state[11]; + cv[4] = state[4] ^ state[12]; + cv[5] = state[5] ^ state[13]; + cv[6] = state[6] ^ state[14]; + cv[7] = state[7] ^ state[15]; +} + +void blake3_compress_xof_portable(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags, uint8_t out[64]) { + uint32_t state[16]; + compress_pre(state, cv, block, block_len, counter, flags); + + store32(&out[0 * 4], state[0] ^ state[8]); + store32(&out[1 * 4], state[1] ^ state[9]); + store32(&out[2 * 4], state[2] ^ state[10]); + store32(&out[3 * 4], state[3] ^ state[11]); + store32(&out[4 * 4], state[4] ^ state[12]); + store32(&out[5 * 4], state[5] ^ state[13]); + store32(&out[6 * 4], state[6] ^ state[14]); + store32(&out[7 * 4], state[7] ^ state[15]); + store32(&out[8 * 4], state[8] ^ cv[0]); + store32(&out[9 * 4], state[9] ^ cv[1]); + store32(&out[10 * 4], state[10] ^ cv[2]); + store32(&out[11 * 4], state[11] ^ cv[3]); + store32(&out[12 * 4], state[12] ^ cv[4]); + store32(&out[13 * 4], state[13] ^ cv[5]); + store32(&out[14 * 4], state[14] ^ cv[6]); + store32(&out[15 * 4], state[15] ^ cv[7]); +} + +INLINE void hash_one_portable(const uint8_t *input, size_t blocks, + const uint32_t key[8], uint64_t counter, + uint8_t flags, uint8_t flags_start, + uint8_t flags_end, uint8_t out[BLAKE3_OUT_LEN]) { + uint32_t cv[8]; + memcpy(cv, key, BLAKE3_KEY_LEN); + uint8_t block_flags = flags | flags_start; + while (blocks > 0) { + if (blocks == 1) { + block_flags |= flags_end; + } + blake3_compress_in_place_portable(cv, input, BLAKE3_BLOCK_LEN, counter, + block_flags); + input = &input[BLAKE3_BLOCK_LEN]; + blocks -= 1; + block_flags = flags; + } + store_cv_words(out, cv); +} + +void blake3_hash_many_portable(const uint8_t *const *inputs, size_t num_inputs, + size_t blocks, const uint32_t key[8], + uint64_t counter, bool increment_counter, + uint8_t flags, uint8_t flags_start, + uint8_t flags_end, uint8_t *out) { + while (num_inputs > 0) { + hash_one_portable(inputs[0], blocks, key, counter, flags, flags_start, + flags_end, out); + if (increment_counter) { + counter += 1; + } + inputs += 1; + num_inputs -= 1; + out = &out[BLAKE3_OUT_LEN]; + } +} diff --git a/ext/digest/blake3/blake3init.c b/ext/digest/blake3/blake3init.c new file mode 100644 index 0000000..639107a --- /dev/null +++ b/ext/digest/blake3/blake3init.c @@ -0,0 +1,82 @@ +/* BLAKE3 binding for the Ruby digest framework. */ + +#include +#include "../digest.h" +#include "blake3.h" + +/* + * Thin adapters mapping BLAKE3's public API onto the signatures the + * digest framework expects (see rb_digest_metadata_t in ../digest.h). + * The signatures line up directly, so no function-pointer casts are + * needed. + */ + +static int +rb_digest_BLAKE3_init(void *ctx) +{ + blake3_hasher_init((blake3_hasher *)ctx); + return 1; +} + +static void +rb_digest_BLAKE3_update(void *ctx, unsigned char *ptr, size_t size) +{ + blake3_hasher_update((blake3_hasher *)ctx, ptr, size); +} + +static int +rb_digest_BLAKE3_finish(void *ctx, unsigned char *ptr) +{ + blake3_hasher_finalize((blake3_hasher *)ctx, ptr, BLAKE3_OUT_LEN); + return 1; +} + +static const rb_digest_metadata_t blake3 = { + RUBY_DIGEST_API_VERSION, + BLAKE3_OUT_LEN, + BLAKE3_BLOCK_LEN, + sizeof(blake3_hasher), + rb_digest_BLAKE3_init, + rb_digest_BLAKE3_update, + rb_digest_BLAKE3_finish, +}; + +/* + * Document-class: Digest::BLAKE3 < Digest::Base + * A class for calculating message digests using the BLAKE3 hash function, + * by Jack O'Connor, Jean-Philippe Aumasson, Samuel Neves, and Zooko + * Wilcox-O'Hearn. + * + * BLAKE3 produces a digest of 256 bits (32 bytes) by default. + * + * == Examples + * require 'digest' + * + * # Compute a complete digest + * Digest::BLAKE3.hexdigest 'abc' #=> "6437b3ac38465133ffb63b75273a8db548c558465d79db03fd359c6cd5bd9d85" + * + * # Compute digest by chunks + * blake3 = Digest::BLAKE3.new # =># + * blake3.update "ab" + * blake3 << "c" # alias for #update + * blake3.hexdigest # => "6437b3ac..." + * + * # Use the same object to compute another digest + * blake3.reset + * blake3 << "message" + * blake3.hexdigest + */ +void +Init_blake3(void) +{ + VALUE mDigest, cDigest_Base, cDigest_BLAKE3; + +#if 0 + mDigest = rb_define_module("Digest"); /* let rdoc know */ +#endif + mDigest = rb_digest_namespace(); + cDigest_Base = rb_const_get(mDigest, rb_intern_const("Base")); + + cDigest_BLAKE3 = rb_define_class_under(mDigest, "BLAKE3", cDigest_Base); + rb_iv_set(cDigest_BLAKE3, "metadata", rb_digest_make_metadata(&blake3)); +} diff --git a/ext/digest/blake3/extconf.rb b/ext/digest/blake3/extconf.rb new file mode 100644 index 0000000..3ee48ce --- /dev/null +++ b/ext/digest/blake3/extconf.rb @@ -0,0 +1,27 @@ +# -*- coding: us-ascii -*- +# frozen_string_literal: false + +require "mkmf" + +# Build the portable (no-SIMD) BLAKE3 implementation: only blake3.c, +# blake3_dispatch.c and blake3_portable.c are compiled. The SIMD +# backends (SSE2/SSE4.1/AVX2/AVX512/NEON) are intentionally left out to +# keep the build simple and dependency-free. Each instruction set has +# to be disabled explicitly, otherwise blake3_dispatch.c references +# symbols from the omitted backends. In particular NEON auto-enables on +# AArch64 (see blake3_impl.h), so BLAKE3_USE_NEON must be forced off too. +$defs << "-DBLAKE3_NO_SSE2" +$defs << "-DBLAKE3_NO_SSE41" +$defs << "-DBLAKE3_NO_AVX2" +$defs << "-DBLAKE3_NO_AVX512" +$defs << "-DBLAKE3_USE_NEON=0" + +$objs = %w[blake3init blake3 blake3_dispatch blake3_portable].map do |o| + "#{o}.#{$OBJEXT}" +end + +have_header("sys/cdefs.h") + +$preload = %w[digest] + +create_makefile("digest/blake3") diff --git a/test/digest/test_digest.rb b/test/digest/test_digest.rb index 8a39a7b..c2b98f2 100644 --- a/test/digest/test_digest.rb +++ b/test/digest/test_digest.rb @@ -6,7 +6,7 @@ require 'tempfile' require 'digest' -%w[digest/md5 digest/rmd160 digest/sha1 digest/sha2 digest/bubblebabble digest/crc32].each do |lib| +%w[digest/md5 digest/rmd160 digest/sha1 digest/sha2 digest/bubblebabble digest/crc32 digest/blake3].each do |lib| begin require lib rescue LoadError @@ -279,6 +279,106 @@ def test_initialize_copy_into_frozen_raises end end + class TestBLAKE3 < Test::Unit::TestCase + include TestDigest + ALGO = Digest::BLAKE3 + DATA = { + Data1 => "6437b3ac38465133ffb63b75273a8db548c558465d79db03fd359c6cd5bd9d85", + Data2 => "c19012cc2aaf0dc3d8e5c45a1b79114d2df42abb2a410bf54be09e891af06ff8", + } + + # Input byte i has value (i % 251), matching the BLAKE3 + # test_vectors.json fixture. The expected value is the first 32 bytes + # (the default digest length) of each case's extended output hash. + TEST_VECTORS = { + 0 => "af1349b9f5f9a1a6a0404dea36dcc9499bcb25c9adc112b7cc9a93cae41f3262", + 1 => "2d3adedff11b61f14c886e35afa036736dcd87a74d27b5c1510225d0f592e213", + 64 => "4eed7141ea4a5cd4b788606bd23f46e212af9cacebacdc7d1f4c6dc7f2511b98", + 65 => "de1e5fa0be70df6d2be8fffd0e99ceaa8eb6e8c93a63f2d8d1c30ecb6b263dee", + 1023 => "10108970eeda3eb932baac1428c7a2163b0e924c9a9e25b35bba72b28f70bd11", + 1024 => "42214739f095a406f3fc83deb889744ac00df831c10daa55189b5d121c855af7", + 1025 => "d00278ae47eb27b34faecf67b4fe263f82d5412916c1ffd97c8cb7fb814b8444", + 2048 => "e776b6028c7cd22a4d0ba182a8bf62205d2ef576467e838ed6f2529b85fba24a", + 3072 => "b98cb0ff3623be03326b373de6b9095218513e64f1ee2edd2525c7ad1e5cffd2", + 4096 => "015094013f57a5277b59d8475c0501042c0b642e531b0a1c8f58d2163229e969", + 8192 => "aae792484c8efe4f19e2ca7d371d8c467ffb10748d8a5a1ae579948f718a2a63", + 16384 => "f875d6646de28985646f34ee13be9a576fd515f76b5b0a26bb324735041ddde4", + 102400 => "bc3e3d41a1146b069abffad3c0d44860cf664390afce4d9661f7902e7943e085", + } + + # https://github.com/BLAKE3-team/BLAKE3/blob/93a431c78a52d7ccf0f366f106467f5070e6075e/test_vectors/src/lib.rs#L65-L73 + def paint_test_input(len) + len.times.map { |i| (i % 251).chr }.join + end + + def test_digest_length + assert_equal 32, Digest::BLAKE3.new.digest_length + end + + def test_block_length + assert_equal 64, Digest::BLAKE3.new.block_length + end + + def test_empty_string + assert_equal "af1349b9f5f9a1a6a0404dea36dcc9499bcb25c9adc112b7cc9a93cae41f3262", + Digest::BLAKE3.hexdigest("") + end + + def test_known_vectors + # Exercises single-chunk, chunk-boundary and multi-chunk tree hashing + # against the upstream BLAKE3 test vectors. + TEST_VECTORS.each do |len, expected| + assert_equal expected, Digest::BLAKE3.hexdigest(paint_test_input(len)), + "BLAKE3 of #{len}-byte test input" + end + end + + def test_incremental_equals_one_shot + # Feed a multi-chunk input in awkward, unaligned pieces and compare + # against the one-shot digest for the same bytes. + input = paint_test_input(4097) + inc = Digest::BLAKE3.new + off = 0 + [1, 63, 64, 65, 900, 1024, 1080].each do |n| + inc << input[off, n] + off += n + end + inc << input[off..] + assert_equal Digest::BLAKE3.hexdigest(input), inc.hexdigest + end + + def test_clone_mid_stream_independence + d = Digest::BLAKE3.new + d << "ab" + copy = d.clone + d << "c" + copy << "c" + assert_equal d.hexdigest, copy.hexdigest + assert_equal Digest::BLAKE3.hexdigest("abc"), copy.hexdigest + end + + def test_reset + d = Digest::BLAKE3.new + d << "some other data" + d.reset + d << "abc" + assert_equal Digest::BLAKE3.hexdigest("abc"), d.hexdigest + end + + def test_digest_bang_resets_state + d = Digest::BLAKE3.new + d << "abc" + d.hexdigest! + assert_equal Digest::BLAKE3.hexdigest(""), d.hexdigest + end + + def test_initialize_copy_into_frozen_raises + dest = Digest::BLAKE3.allocate + dest.freeze + assert_raise(FrozenError) { dest.send(:initialize_copy, Digest::BLAKE3.new) } + end + end if defined?(Digest::BLAKE3) + class TestBase < Test::Unit::TestCase def test_base bug3810 = '[ruby-core:32231]' From 6865b37d30caea760408a94a105af837d068b72c Mon Sep 17 00:00:00 2001 From: Aaron Patterson Date: Tue, 7 Jul 2026 09:19:01 -0700 Subject: [PATCH 2/4] Add SIMD support for various platforms We need extconf to detect which SIMD features the compiler and processor support, then we can compile in SIMD support for BLAKE3 Co-Authored-By: Jean Boussier --- ext/digest/blake3/extconf.rb | 86 +++++++++++++++++++++++++++++------- 1 file changed, 71 insertions(+), 15 deletions(-) diff --git a/ext/digest/blake3/extconf.rb b/ext/digest/blake3/extconf.rb index 3ee48ce..09548ec 100644 --- a/ext/digest/blake3/extconf.rb +++ b/ext/digest/blake3/extconf.rb @@ -3,23 +3,79 @@ require "mkmf" -# Build the portable (no-SIMD) BLAKE3 implementation: only blake3.c, -# blake3_dispatch.c and blake3_portable.c are compiled. The SIMD -# backends (SSE2/SSE4.1/AVX2/AVX512/NEON) are intentionally left out to -# keep the build simple and dependency-free. Each instruction set has -# to be disabled explicitly, otherwise blake3_dispatch.c references -# symbols from the omitted backends. In particular NEON auto-enables on -# AArch64 (see blake3_impl.h), so BLAKE3_USE_NEON must be forced off too. -$defs << "-DBLAKE3_NO_SSE2" -$defs << "-DBLAKE3_NO_SSE41" -$defs << "-DBLAKE3_NO_AVX2" -$defs << "-DBLAKE3_NO_AVX512" -$defs << "-DBLAKE3_USE_NEON=0" - -$objs = %w[blake3init blake3 blake3_dispatch blake3_portable].map do |o| - "#{o}.#{$OBJEXT}" +# BLAKE3 build configuration. +# +# The binding is always built from the portable C code (blake3.c, +# blake3_dispatch.c, blake3_portable.c). On architectures BLAKE3 ships +# optimized backends for, we additionally compile the SIMD implementations +# and let blake3_dispatch.c select the fastest one supported by the CPU the +# program is actually running on (via CPUID on x86). This means a binary +# built on an AVX-512-capable machine still runs correctly on an older CPU. +# +# Each SIMD translation unit must be compiled with its own instruction-set +# flag, which a single global $CFLAGS can't express, so we emit one explicit +# object rule per backend at the end (see below). Any x86 backend we do NOT +# compile is disabled with -DBLAKE3_NO_; that macro is honoured both by +# the dispatcher and by sibling backends (e.g. blake3_avx2.c falls back to +# SSE4.1 helpers), keeping the set of referenced symbols consistent. + +objs = %w[blake3init blake3 blake3_dispatch blake3_portable] + +# Extra per-object compiler flags, keyed by object basename. +simd_cflags = {} + +# Probe used to confirm the compiler both accepts +flag+ and can compile the +# intrinsics the backend relies on. +def blake3_have_isa?(name, flag, snippet) + checking_for("#{name} intrinsics (#{flag})") do + try_compile(snippet, flag) + end end +case RbConfig::CONFIG["host_cpu"] +when /\A(x86_64|amd64|x64)\z/i + # Try to detect which SIMD features this x86 machine and compiler has + x86_backends = [ + ["blake3_sse2", "SSE2", "-msse2", + "#include \nint main(void){ volatile __m128i x = _mm_setzero_si128(); (void)x; return 0; }\n", + "BLAKE3_NO_SSE2"], + ["blake3_sse41", "SSE4.1", "-msse4.1", + "#include \nint main(void){ volatile __m128i x = _mm_setzero_si128(); return _mm_testz_si128(x, x); }\n", + "BLAKE3_NO_SSE41"], + ["blake3_avx2", "AVX2", "-mavx2", + "#include \nint main(void){ volatile __m256i x = _mm256_setzero_si256(); (void)x; return 0; }\n", + "BLAKE3_NO_AVX2"], + ["blake3_avx512", "AVX-512", "-mavx512f -mavx512vl", + "#include \nint main(void){ volatile __m512i x = _mm512_setzero_si512(); (void)x; return 0; }\n", + "BLAKE3_NO_AVX512"], + ] + + x86_backends.each do |obj, name, flag, snippet, no_macro| + if blake3_have_isa?(name, flag, snippet) + objs << obj + simd_cflags[obj] = flag + else + $defs << "-D#{no_macro}" + end + end +when /\A(aarch64|arm64)\z/i + # NEON is part of the AArch64 baseline, so no runtime detection or special + # compiler flag is needed. Leave BLAKE3_USE_NEON to auto-detect, which is + # 1 on little-endian AArch64 (see blake3_impl.h). + objs << "blake3_neon" +else + # No optimized backend wired up for this architecture (e.g. 32-bit x86, + # ppc): build portable-only. Disabling every x86 ISA keeps the dispatcher + # from referencing backends we didn't compile, and NEON is forced off. + $defs << "-DBLAKE3_NO_SSE2" + $defs << "-DBLAKE3_NO_SSE41" + $defs << "-DBLAKE3_NO_AVX2" + $defs << "-DBLAKE3_NO_AVX512" + $defs << "-DBLAKE3_USE_NEON=0" +end + +$objs = objs.map { |o| "#{o}.#{$OBJEXT}" } + have_header("sys/cdefs.h") $preload = %w[digest] From 30ae414c1c171c6464178e4ea6e05d7b7242d0d5 Mon Sep 17 00:00:00 2001 From: Aaron Patterson Date: Tue, 7 Jul 2026 13:50:25 -0700 Subject: [PATCH 3/4] oops --- ext/digest/blake3/blake3_avx2.c | 326 +++++++ ext/digest/blake3/blake3_avx512.c | 1388 +++++++++++++++++++++++++++++ ext/digest/blake3/blake3_neon.c | 367 ++++++++ ext/digest/blake3/blake3_sse2.c | 566 ++++++++++++ ext/digest/blake3/blake3_sse41.c | 560 ++++++++++++ 5 files changed, 3207 insertions(+) create mode 100644 ext/digest/blake3/blake3_avx2.c create mode 100644 ext/digest/blake3/blake3_avx512.c create mode 100644 ext/digest/blake3/blake3_neon.c create mode 100644 ext/digest/blake3/blake3_sse2.c create mode 100644 ext/digest/blake3/blake3_sse41.c diff --git a/ext/digest/blake3/blake3_avx2.c b/ext/digest/blake3/blake3_avx2.c new file mode 100644 index 0000000..381e7c4 --- /dev/null +++ b/ext/digest/blake3/blake3_avx2.c @@ -0,0 +1,326 @@ +#include "blake3_impl.h" + +#include + +#define DEGREE 8 + +INLINE __m256i loadu(const uint8_t src[32]) { + return _mm256_loadu_si256((const __m256i *)src); +} + +INLINE void storeu(__m256i src, uint8_t dest[16]) { + _mm256_storeu_si256((__m256i *)dest, src); +} + +INLINE __m256i addv(__m256i a, __m256i b) { return _mm256_add_epi32(a, b); } + +// Note that clang-format doesn't like the name "xor" for some reason. +INLINE __m256i xorv(__m256i a, __m256i b) { return _mm256_xor_si256(a, b); } + +INLINE __m256i set1(uint32_t x) { return _mm256_set1_epi32((int32_t)x); } + +INLINE __m256i rot16(__m256i x) { + return _mm256_shuffle_epi8( + x, _mm256_set_epi8(13, 12, 15, 14, 9, 8, 11, 10, 5, 4, 7, 6, 1, 0, 3, 2, + 13, 12, 15, 14, 9, 8, 11, 10, 5, 4, 7, 6, 1, 0, 3, 2)); +} + +INLINE __m256i rot12(__m256i x) { + return _mm256_or_si256(_mm256_srli_epi32(x, 12), _mm256_slli_epi32(x, 32 - 12)); +} + +INLINE __m256i rot8(__m256i x) { + return _mm256_shuffle_epi8( + x, _mm256_set_epi8(12, 15, 14, 13, 8, 11, 10, 9, 4, 7, 6, 5, 0, 3, 2, 1, + 12, 15, 14, 13, 8, 11, 10, 9, 4, 7, 6, 5, 0, 3, 2, 1)); +} + +INLINE __m256i rot7(__m256i x) { + return _mm256_or_si256(_mm256_srli_epi32(x, 7), _mm256_slli_epi32(x, 32 - 7)); +} + +INLINE void round_fn(__m256i v[16], __m256i m[16], size_t r) { + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][0]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][2]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][4]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][6]]); + v[0] = addv(v[0], v[4]); + v[1] = addv(v[1], v[5]); + v[2] = addv(v[2], v[6]); + v[3] = addv(v[3], v[7]); + v[12] = xorv(v[12], v[0]); + v[13] = xorv(v[13], v[1]); + v[14] = xorv(v[14], v[2]); + v[15] = xorv(v[15], v[3]); + v[12] = rot16(v[12]); + v[13] = rot16(v[13]); + v[14] = rot16(v[14]); + v[15] = rot16(v[15]); + v[8] = addv(v[8], v[12]); + v[9] = addv(v[9], v[13]); + v[10] = addv(v[10], v[14]); + v[11] = addv(v[11], v[15]); + v[4] = xorv(v[4], v[8]); + v[5] = xorv(v[5], v[9]); + v[6] = xorv(v[6], v[10]); + v[7] = xorv(v[7], v[11]); + v[4] = rot12(v[4]); + v[5] = rot12(v[5]); + v[6] = rot12(v[6]); + v[7] = rot12(v[7]); + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][1]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][3]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][5]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][7]]); + v[0] = addv(v[0], v[4]); + v[1] = addv(v[1], v[5]); + v[2] = addv(v[2], v[6]); + v[3] = addv(v[3], v[7]); + v[12] = xorv(v[12], v[0]); + v[13] = xorv(v[13], v[1]); + v[14] = xorv(v[14], v[2]); + v[15] = xorv(v[15], v[3]); + v[12] = rot8(v[12]); + v[13] = rot8(v[13]); + v[14] = rot8(v[14]); + v[15] = rot8(v[15]); + v[8] = addv(v[8], v[12]); + v[9] = addv(v[9], v[13]); + v[10] = addv(v[10], v[14]); + v[11] = addv(v[11], v[15]); + v[4] = xorv(v[4], v[8]); + v[5] = xorv(v[5], v[9]); + v[6] = xorv(v[6], v[10]); + v[7] = xorv(v[7], v[11]); + v[4] = rot7(v[4]); + v[5] = rot7(v[5]); + v[6] = rot7(v[6]); + v[7] = rot7(v[7]); + + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][8]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][10]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][12]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][14]]); + v[0] = addv(v[0], v[5]); + v[1] = addv(v[1], v[6]); + v[2] = addv(v[2], v[7]); + v[3] = addv(v[3], v[4]); + v[15] = xorv(v[15], v[0]); + v[12] = xorv(v[12], v[1]); + v[13] = xorv(v[13], v[2]); + v[14] = xorv(v[14], v[3]); + v[15] = rot16(v[15]); + v[12] = rot16(v[12]); + v[13] = rot16(v[13]); + v[14] = rot16(v[14]); + v[10] = addv(v[10], v[15]); + v[11] = addv(v[11], v[12]); + v[8] = addv(v[8], v[13]); + v[9] = addv(v[9], v[14]); + v[5] = xorv(v[5], v[10]); + v[6] = xorv(v[6], v[11]); + v[7] = xorv(v[7], v[8]); + v[4] = xorv(v[4], v[9]); + v[5] = rot12(v[5]); + v[6] = rot12(v[6]); + v[7] = rot12(v[7]); + v[4] = rot12(v[4]); + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][9]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][11]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][13]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][15]]); + v[0] = addv(v[0], v[5]); + v[1] = addv(v[1], v[6]); + v[2] = addv(v[2], v[7]); + v[3] = addv(v[3], v[4]); + v[15] = xorv(v[15], v[0]); + v[12] = xorv(v[12], v[1]); + v[13] = xorv(v[13], v[2]); + v[14] = xorv(v[14], v[3]); + v[15] = rot8(v[15]); + v[12] = rot8(v[12]); + v[13] = rot8(v[13]); + v[14] = rot8(v[14]); + v[10] = addv(v[10], v[15]); + v[11] = addv(v[11], v[12]); + v[8] = addv(v[8], v[13]); + v[9] = addv(v[9], v[14]); + v[5] = xorv(v[5], v[10]); + v[6] = xorv(v[6], v[11]); + v[7] = xorv(v[7], v[8]); + v[4] = xorv(v[4], v[9]); + v[5] = rot7(v[5]); + v[6] = rot7(v[6]); + v[7] = rot7(v[7]); + v[4] = rot7(v[4]); +} + +INLINE void transpose_vecs(__m256i vecs[DEGREE]) { + // Interleave 32-bit lanes. The low unpack is lanes 00/11/44/55, and the high + // is 22/33/66/77. + __m256i ab_0145 = _mm256_unpacklo_epi32(vecs[0], vecs[1]); + __m256i ab_2367 = _mm256_unpackhi_epi32(vecs[0], vecs[1]); + __m256i cd_0145 = _mm256_unpacklo_epi32(vecs[2], vecs[3]); + __m256i cd_2367 = _mm256_unpackhi_epi32(vecs[2], vecs[3]); + __m256i ef_0145 = _mm256_unpacklo_epi32(vecs[4], vecs[5]); + __m256i ef_2367 = _mm256_unpackhi_epi32(vecs[4], vecs[5]); + __m256i gh_0145 = _mm256_unpacklo_epi32(vecs[6], vecs[7]); + __m256i gh_2367 = _mm256_unpackhi_epi32(vecs[6], vecs[7]); + + // Interleave 64-bit lanes. The low unpack is lanes 00/22 and the high is + // 11/33. + __m256i abcd_04 = _mm256_unpacklo_epi64(ab_0145, cd_0145); + __m256i abcd_15 = _mm256_unpackhi_epi64(ab_0145, cd_0145); + __m256i abcd_26 = _mm256_unpacklo_epi64(ab_2367, cd_2367); + __m256i abcd_37 = _mm256_unpackhi_epi64(ab_2367, cd_2367); + __m256i efgh_04 = _mm256_unpacklo_epi64(ef_0145, gh_0145); + __m256i efgh_15 = _mm256_unpackhi_epi64(ef_0145, gh_0145); + __m256i efgh_26 = _mm256_unpacklo_epi64(ef_2367, gh_2367); + __m256i efgh_37 = _mm256_unpackhi_epi64(ef_2367, gh_2367); + + // Interleave 128-bit lanes. + vecs[0] = _mm256_permute2x128_si256(abcd_04, efgh_04, 0x20); + vecs[1] = _mm256_permute2x128_si256(abcd_15, efgh_15, 0x20); + vecs[2] = _mm256_permute2x128_si256(abcd_26, efgh_26, 0x20); + vecs[3] = _mm256_permute2x128_si256(abcd_37, efgh_37, 0x20); + vecs[4] = _mm256_permute2x128_si256(abcd_04, efgh_04, 0x31); + vecs[5] = _mm256_permute2x128_si256(abcd_15, efgh_15, 0x31); + vecs[6] = _mm256_permute2x128_si256(abcd_26, efgh_26, 0x31); + vecs[7] = _mm256_permute2x128_si256(abcd_37, efgh_37, 0x31); +} + +INLINE void transpose_msg_vecs(const uint8_t *const *inputs, + size_t block_offset, __m256i out[16]) { + out[0] = loadu(&inputs[0][block_offset + 0 * sizeof(__m256i)]); + out[1] = loadu(&inputs[1][block_offset + 0 * sizeof(__m256i)]); + out[2] = loadu(&inputs[2][block_offset + 0 * sizeof(__m256i)]); + out[3] = loadu(&inputs[3][block_offset + 0 * sizeof(__m256i)]); + out[4] = loadu(&inputs[4][block_offset + 0 * sizeof(__m256i)]); + out[5] = loadu(&inputs[5][block_offset + 0 * sizeof(__m256i)]); + out[6] = loadu(&inputs[6][block_offset + 0 * sizeof(__m256i)]); + out[7] = loadu(&inputs[7][block_offset + 0 * sizeof(__m256i)]); + out[8] = loadu(&inputs[0][block_offset + 1 * sizeof(__m256i)]); + out[9] = loadu(&inputs[1][block_offset + 1 * sizeof(__m256i)]); + out[10] = loadu(&inputs[2][block_offset + 1 * sizeof(__m256i)]); + out[11] = loadu(&inputs[3][block_offset + 1 * sizeof(__m256i)]); + out[12] = loadu(&inputs[4][block_offset + 1 * sizeof(__m256i)]); + out[13] = loadu(&inputs[5][block_offset + 1 * sizeof(__m256i)]); + out[14] = loadu(&inputs[6][block_offset + 1 * sizeof(__m256i)]); + out[15] = loadu(&inputs[7][block_offset + 1 * sizeof(__m256i)]); + for (size_t i = 0; i < 8; ++i) { + _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0); + } + transpose_vecs(&out[0]); + transpose_vecs(&out[8]); +} + +INLINE void load_counters(uint64_t counter, bool increment_counter, + __m256i *out_lo, __m256i *out_hi) { + const __m256i mask = _mm256_set1_epi32(-(int32_t)increment_counter); + const __m256i add0 = _mm256_set_epi32(7, 6, 5, 4, 3, 2, 1, 0); + const __m256i add1 = _mm256_and_si256(mask, add0); + __m256i l = _mm256_add_epi32(_mm256_set1_epi32((int32_t)counter), add1); + __m256i carry = _mm256_cmpgt_epi32(_mm256_xor_si256(add1, _mm256_set1_epi32(0x80000000)), + _mm256_xor_si256( l, _mm256_set1_epi32(0x80000000))); + __m256i h = _mm256_sub_epi32(_mm256_set1_epi32((int32_t)(counter >> 32)), carry); + *out_lo = l; + *out_hi = h; +} + +static +void blake3_hash8_avx2(const uint8_t *const *inputs, size_t blocks, + const uint32_t key[8], uint64_t counter, + bool increment_counter, uint8_t flags, + uint8_t flags_start, uint8_t flags_end, uint8_t *out) { + __m256i h_vecs[8] = { + set1(key[0]), set1(key[1]), set1(key[2]), set1(key[3]), + set1(key[4]), set1(key[5]), set1(key[6]), set1(key[7]), + }; + __m256i counter_low_vec, counter_high_vec; + load_counters(counter, increment_counter, &counter_low_vec, + &counter_high_vec); + uint8_t block_flags = flags | flags_start; + + for (size_t block = 0; block < blocks; block++) { + if (block + 1 == blocks) { + block_flags |= flags_end; + } + __m256i block_len_vec = set1(BLAKE3_BLOCK_LEN); + __m256i block_flags_vec = set1(block_flags); + __m256i msg_vecs[16]; + transpose_msg_vecs(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs); + + __m256i v[16] = { + h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3], + h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7], + set1(IV[0]), set1(IV[1]), set1(IV[2]), set1(IV[3]), + counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec, + }; + round_fn(v, msg_vecs, 0); + round_fn(v, msg_vecs, 1); + round_fn(v, msg_vecs, 2); + round_fn(v, msg_vecs, 3); + round_fn(v, msg_vecs, 4); + round_fn(v, msg_vecs, 5); + round_fn(v, msg_vecs, 6); + h_vecs[0] = xorv(v[0], v[8]); + h_vecs[1] = xorv(v[1], v[9]); + h_vecs[2] = xorv(v[2], v[10]); + h_vecs[3] = xorv(v[3], v[11]); + h_vecs[4] = xorv(v[4], v[12]); + h_vecs[5] = xorv(v[5], v[13]); + h_vecs[6] = xorv(v[6], v[14]); + h_vecs[7] = xorv(v[7], v[15]); + + block_flags = flags; + } + + transpose_vecs(h_vecs); + storeu(h_vecs[0], &out[0 * sizeof(__m256i)]); + storeu(h_vecs[1], &out[1 * sizeof(__m256i)]); + storeu(h_vecs[2], &out[2 * sizeof(__m256i)]); + storeu(h_vecs[3], &out[3 * sizeof(__m256i)]); + storeu(h_vecs[4], &out[4 * sizeof(__m256i)]); + storeu(h_vecs[5], &out[5 * sizeof(__m256i)]); + storeu(h_vecs[6], &out[6 * sizeof(__m256i)]); + storeu(h_vecs[7], &out[7 * sizeof(__m256i)]); +} + +#if !defined(BLAKE3_NO_SSE41) +void blake3_hash_many_sse41(const uint8_t *const *inputs, size_t num_inputs, + size_t blocks, const uint32_t key[8], + uint64_t counter, bool increment_counter, + uint8_t flags, uint8_t flags_start, + uint8_t flags_end, uint8_t *out); +#else +void blake3_hash_many_portable(const uint8_t *const *inputs, size_t num_inputs, + size_t blocks, const uint32_t key[8], + uint64_t counter, bool increment_counter, + uint8_t flags, uint8_t flags_start, + uint8_t flags_end, uint8_t *out); +#endif + +void blake3_hash_many_avx2(const uint8_t *const *inputs, size_t num_inputs, + size_t blocks, const uint32_t key[8], + uint64_t counter, bool increment_counter, + uint8_t flags, uint8_t flags_start, + uint8_t flags_end, uint8_t *out) { + while (num_inputs >= DEGREE) { + blake3_hash8_avx2(inputs, blocks, key, counter, increment_counter, flags, + flags_start, flags_end, out); + if (increment_counter) { + counter += DEGREE; + } + inputs += DEGREE; + num_inputs -= DEGREE; + out = &out[DEGREE * BLAKE3_OUT_LEN]; + } +#if !defined(BLAKE3_NO_SSE41) + blake3_hash_many_sse41(inputs, num_inputs, blocks, key, counter, + increment_counter, flags, flags_start, flags_end, out); +#else + blake3_hash_many_portable(inputs, num_inputs, blocks, key, counter, + increment_counter, flags, flags_start, flags_end, + out); +#endif +} diff --git a/ext/digest/blake3/blake3_avx512.c b/ext/digest/blake3/blake3_avx512.c new file mode 100644 index 0000000..f88a32d --- /dev/null +++ b/ext/digest/blake3/blake3_avx512.c @@ -0,0 +1,1388 @@ +#include "blake3_impl.h" + +#include + +#define _mm_shuffle_ps2(a, b, c) \ + (_mm_castps_si128( \ + _mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), (c)))) + +INLINE __m128i loadu_128(const uint8_t src[16]) { + return _mm_loadu_si128((void*)src); +} + +INLINE __m256i loadu_256(const uint8_t src[32]) { + return _mm256_loadu_si256((void*)src); +} + +INLINE __m512i loadu_512(const uint8_t src[64]) { + return _mm512_loadu_si512((void*)src); +} + +INLINE void storeu_128(__m128i src, uint8_t dest[16]) { + _mm_storeu_si128((void*)dest, src); +} + +INLINE void storeu_256(__m256i src, uint8_t dest[16]) { + _mm256_storeu_si256((void*)dest, src); +} + +INLINE void storeu_512(__m512i src, uint8_t dest[16]) { + _mm512_storeu_si512((void*)dest, src); +} + +INLINE __m128i add_128(__m128i a, __m128i b) { return _mm_add_epi32(a, b); } + +INLINE __m256i add_256(__m256i a, __m256i b) { return _mm256_add_epi32(a, b); } + +INLINE __m512i add_512(__m512i a, __m512i b) { return _mm512_add_epi32(a, b); } + +INLINE __m128i xor_128(__m128i a, __m128i b) { return _mm_xor_si128(a, b); } + +INLINE __m256i xor_256(__m256i a, __m256i b) { return _mm256_xor_si256(a, b); } + +INLINE __m512i xor_512(__m512i a, __m512i b) { return _mm512_xor_si512(a, b); } + +INLINE __m128i set1_128(uint32_t x) { return _mm_set1_epi32((int32_t)x); } + +INLINE __m256i set1_256(uint32_t x) { return _mm256_set1_epi32((int32_t)x); } + +INLINE __m512i set1_512(uint32_t x) { return _mm512_set1_epi32((int32_t)x); } + +INLINE __m128i set4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) { + return _mm_setr_epi32((int32_t)a, (int32_t)b, (int32_t)c, (int32_t)d); +} + +INLINE __m128i rot16_128(__m128i x) { return _mm_ror_epi32(x, 16); } + +INLINE __m256i rot16_256(__m256i x) { return _mm256_ror_epi32(x, 16); } + +INLINE __m512i rot16_512(__m512i x) { return _mm512_ror_epi32(x, 16); } + +INLINE __m128i rot12_128(__m128i x) { return _mm_ror_epi32(x, 12); } + +INLINE __m256i rot12_256(__m256i x) { return _mm256_ror_epi32(x, 12); } + +INLINE __m512i rot12_512(__m512i x) { return _mm512_ror_epi32(x, 12); } + +INLINE __m128i rot8_128(__m128i x) { return _mm_ror_epi32(x, 8); } + +INLINE __m256i rot8_256(__m256i x) { return _mm256_ror_epi32(x, 8); } + +INLINE __m512i rot8_512(__m512i x) { return _mm512_ror_epi32(x, 8); } + +INLINE __m128i rot7_128(__m128i x) { return _mm_ror_epi32(x, 7); } + +INLINE __m256i rot7_256(__m256i x) { return _mm256_ror_epi32(x, 7); } + +INLINE __m512i rot7_512(__m512i x) { return _mm512_ror_epi32(x, 7); } + +/* + * ---------------------------------------------------------------------------- + * compress_avx512 + * ---------------------------------------------------------------------------- + */ + +INLINE void g1(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3, + __m128i m) { + *row0 = add_128(add_128(*row0, m), *row1); + *row3 = xor_128(*row3, *row0); + *row3 = rot16_128(*row3); + *row2 = add_128(*row2, *row3); + *row1 = xor_128(*row1, *row2); + *row1 = rot12_128(*row1); +} + +INLINE void g2(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3, + __m128i m) { + *row0 = add_128(add_128(*row0, m), *row1); + *row3 = xor_128(*row3, *row0); + *row3 = rot8_128(*row3); + *row2 = add_128(*row2, *row3); + *row1 = xor_128(*row1, *row2); + *row1 = rot7_128(*row1); +} + +// Note the optimization here of leaving row1 as the unrotated row, rather than +// row0. All the message loads below are adjusted to compensate for this. See +// discussion at https://github.com/sneves/blake2-avx2/pull/4 +INLINE void diagonalize(__m128i *row0, __m128i *row2, __m128i *row3) { + *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(2, 1, 0, 3)); + *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2)); + *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(0, 3, 2, 1)); +} + +INLINE void undiagonalize(__m128i *row0, __m128i *row2, __m128i *row3) { + *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(0, 3, 2, 1)); + *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2)); + *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(2, 1, 0, 3)); +} + +INLINE void compress_pre(__m128i rows[4], const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, uint8_t flags) { + rows[0] = loadu_128((uint8_t *)&cv[0]); + rows[1] = loadu_128((uint8_t *)&cv[4]); + rows[2] = set4(IV[0], IV[1], IV[2], IV[3]); + rows[3] = set4(counter_low(counter), counter_high(counter), + (uint32_t)block_len, (uint32_t)flags); + + __m128i m0 = loadu_128(&block[sizeof(__m128i) * 0]); + __m128i m1 = loadu_128(&block[sizeof(__m128i) * 1]); + __m128i m2 = loadu_128(&block[sizeof(__m128i) * 2]); + __m128i m3 = loadu_128(&block[sizeof(__m128i) * 3]); + + __m128i t0, t1, t2, t3, tt; + + // Round 1. The first round permutes the message words from the original + // input order, into the groups that get mixed in parallel. + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(2, 0, 2, 0)); // 6 4 2 0 + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 3, 1)); // 7 5 3 1 + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(2, 0, 2, 0)); // 14 12 10 8 + t2 = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2, 1, 0, 3)); // 12 10 8 14 + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 1, 3, 1)); // 15 13 11 9 + t3 = _mm_shuffle_epi32(t3, _MM_SHUFFLE(2, 1, 0, 3)); // 13 11 9 15 + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 2. This round and all following rounds apply a fixed permutation + // to the message words from the round before. + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 3 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 4 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 5 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 6 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 7 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); +} + +void blake3_compress_xof_avx512(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags, uint8_t out[64]) { + __m128i rows[4]; + compress_pre(rows, cv, block, block_len, counter, flags); + storeu_128(xor_128(rows[0], rows[2]), &out[0]); + storeu_128(xor_128(rows[1], rows[3]), &out[16]); + storeu_128(xor_128(rows[2], loadu_128((uint8_t *)&cv[0])), &out[32]); + storeu_128(xor_128(rows[3], loadu_128((uint8_t *)&cv[4])), &out[48]); +} + +void blake3_compress_in_place_avx512(uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags) { + __m128i rows[4]; + compress_pre(rows, cv, block, block_len, counter, flags); + storeu_128(xor_128(rows[0], rows[2]), (uint8_t *)&cv[0]); + storeu_128(xor_128(rows[1], rows[3]), (uint8_t *)&cv[4]); +} + +/* + * ---------------------------------------------------------------------------- + * hash4_avx512 + * ---------------------------------------------------------------------------- + */ + +INLINE void round_fn4(__m128i v[16], __m128i m[16], size_t r) { + v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][0]]); + v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][2]]); + v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][4]]); + v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][6]]); + v[0] = add_128(v[0], v[4]); + v[1] = add_128(v[1], v[5]); + v[2] = add_128(v[2], v[6]); + v[3] = add_128(v[3], v[7]); + v[12] = xor_128(v[12], v[0]); + v[13] = xor_128(v[13], v[1]); + v[14] = xor_128(v[14], v[2]); + v[15] = xor_128(v[15], v[3]); + v[12] = rot16_128(v[12]); + v[13] = rot16_128(v[13]); + v[14] = rot16_128(v[14]); + v[15] = rot16_128(v[15]); + v[8] = add_128(v[8], v[12]); + v[9] = add_128(v[9], v[13]); + v[10] = add_128(v[10], v[14]); + v[11] = add_128(v[11], v[15]); + v[4] = xor_128(v[4], v[8]); + v[5] = xor_128(v[5], v[9]); + v[6] = xor_128(v[6], v[10]); + v[7] = xor_128(v[7], v[11]); + v[4] = rot12_128(v[4]); + v[5] = rot12_128(v[5]); + v[6] = rot12_128(v[6]); + v[7] = rot12_128(v[7]); + v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][1]]); + v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][3]]); + v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][5]]); + v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][7]]); + v[0] = add_128(v[0], v[4]); + v[1] = add_128(v[1], v[5]); + v[2] = add_128(v[2], v[6]); + v[3] = add_128(v[3], v[7]); + v[12] = xor_128(v[12], v[0]); + v[13] = xor_128(v[13], v[1]); + v[14] = xor_128(v[14], v[2]); + v[15] = xor_128(v[15], v[3]); + v[12] = rot8_128(v[12]); + v[13] = rot8_128(v[13]); + v[14] = rot8_128(v[14]); + v[15] = rot8_128(v[15]); + v[8] = add_128(v[8], v[12]); + v[9] = add_128(v[9], v[13]); + v[10] = add_128(v[10], v[14]); + v[11] = add_128(v[11], v[15]); + v[4] = xor_128(v[4], v[8]); + v[5] = xor_128(v[5], v[9]); + v[6] = xor_128(v[6], v[10]); + v[7] = xor_128(v[7], v[11]); + v[4] = rot7_128(v[4]); + v[5] = rot7_128(v[5]); + v[6] = rot7_128(v[6]); + v[7] = rot7_128(v[7]); + + v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][8]]); + v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][10]]); + v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][12]]); + v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][14]]); + v[0] = add_128(v[0], v[5]); + v[1] = add_128(v[1], v[6]); + v[2] = add_128(v[2], v[7]); + v[3] = add_128(v[3], v[4]); + v[15] = xor_128(v[15], v[0]); + v[12] = xor_128(v[12], v[1]); + v[13] = xor_128(v[13], v[2]); + v[14] = xor_128(v[14], v[3]); + v[15] = rot16_128(v[15]); + v[12] = rot16_128(v[12]); + v[13] = rot16_128(v[13]); + v[14] = rot16_128(v[14]); + v[10] = add_128(v[10], v[15]); + v[11] = add_128(v[11], v[12]); + v[8] = add_128(v[8], v[13]); + v[9] = add_128(v[9], v[14]); + v[5] = xor_128(v[5], v[10]); + v[6] = xor_128(v[6], v[11]); + v[7] = xor_128(v[7], v[8]); + v[4] = xor_128(v[4], v[9]); + v[5] = rot12_128(v[5]); + v[6] = rot12_128(v[6]); + v[7] = rot12_128(v[7]); + v[4] = rot12_128(v[4]); + v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][9]]); + v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][11]]); + v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][13]]); + v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][15]]); + v[0] = add_128(v[0], v[5]); + v[1] = add_128(v[1], v[6]); + v[2] = add_128(v[2], v[7]); + v[3] = add_128(v[3], v[4]); + v[15] = xor_128(v[15], v[0]); + v[12] = xor_128(v[12], v[1]); + v[13] = xor_128(v[13], v[2]); + v[14] = xor_128(v[14], v[3]); + v[15] = rot8_128(v[15]); + v[12] = rot8_128(v[12]); + v[13] = rot8_128(v[13]); + v[14] = rot8_128(v[14]); + v[10] = add_128(v[10], v[15]); + v[11] = add_128(v[11], v[12]); + v[8] = add_128(v[8], v[13]); + v[9] = add_128(v[9], v[14]); + v[5] = xor_128(v[5], v[10]); + v[6] = xor_128(v[6], v[11]); + v[7] = xor_128(v[7], v[8]); + v[4] = xor_128(v[4], v[9]); + v[5] = rot7_128(v[5]); + v[6] = rot7_128(v[6]); + v[7] = rot7_128(v[7]); + v[4] = rot7_128(v[4]); +} + +INLINE void transpose_vecs_128(__m128i vecs[4]) { + // Interleave 32-bit lanes. The low unpack is lanes 00/11 and the high is + // 22/33. Note that this doesn't split the vector into two lanes, as the + // AVX2 counterparts do. + __m128i ab_01 = _mm_unpacklo_epi32(vecs[0], vecs[1]); + __m128i ab_23 = _mm_unpackhi_epi32(vecs[0], vecs[1]); + __m128i cd_01 = _mm_unpacklo_epi32(vecs[2], vecs[3]); + __m128i cd_23 = _mm_unpackhi_epi32(vecs[2], vecs[3]); + + // Interleave 64-bit lanes. + __m128i abcd_0 = _mm_unpacklo_epi64(ab_01, cd_01); + __m128i abcd_1 = _mm_unpackhi_epi64(ab_01, cd_01); + __m128i abcd_2 = _mm_unpacklo_epi64(ab_23, cd_23); + __m128i abcd_3 = _mm_unpackhi_epi64(ab_23, cd_23); + + vecs[0] = abcd_0; + vecs[1] = abcd_1; + vecs[2] = abcd_2; + vecs[3] = abcd_3; +} + +INLINE void transpose_msg_vecs4(const uint8_t *const *inputs, + size_t block_offset, __m128i out[16]) { + out[0] = loadu_128(&inputs[0][block_offset + 0 * sizeof(__m128i)]); + out[1] = loadu_128(&inputs[1][block_offset + 0 * sizeof(__m128i)]); + out[2] = loadu_128(&inputs[2][block_offset + 0 * sizeof(__m128i)]); + out[3] = loadu_128(&inputs[3][block_offset + 0 * sizeof(__m128i)]); + out[4] = loadu_128(&inputs[0][block_offset + 1 * sizeof(__m128i)]); + out[5] = loadu_128(&inputs[1][block_offset + 1 * sizeof(__m128i)]); + out[6] = loadu_128(&inputs[2][block_offset + 1 * sizeof(__m128i)]); + out[7] = loadu_128(&inputs[3][block_offset + 1 * sizeof(__m128i)]); + out[8] = loadu_128(&inputs[0][block_offset + 2 * sizeof(__m128i)]); + out[9] = loadu_128(&inputs[1][block_offset + 2 * sizeof(__m128i)]); + out[10] = loadu_128(&inputs[2][block_offset + 2 * sizeof(__m128i)]); + out[11] = loadu_128(&inputs[3][block_offset + 2 * sizeof(__m128i)]); + out[12] = loadu_128(&inputs[0][block_offset + 3 * sizeof(__m128i)]); + out[13] = loadu_128(&inputs[1][block_offset + 3 * sizeof(__m128i)]); + out[14] = loadu_128(&inputs[2][block_offset + 3 * sizeof(__m128i)]); + out[15] = loadu_128(&inputs[3][block_offset + 3 * sizeof(__m128i)]); + for (size_t i = 0; i < 4; ++i) { + _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0); + } + transpose_vecs_128(&out[0]); + transpose_vecs_128(&out[4]); + transpose_vecs_128(&out[8]); + transpose_vecs_128(&out[12]); +} + +INLINE void load_counters4(uint64_t counter, bool increment_counter, + __m128i *out_lo, __m128i *out_hi) { + uint64_t mask = (increment_counter ? ~0 : 0); + __m256i mask_vec = _mm256_set1_epi64x(mask); + __m256i deltas = _mm256_setr_epi64x(0, 1, 2, 3); + deltas = _mm256_and_si256(mask_vec, deltas); + __m256i counters = + _mm256_add_epi64(_mm256_set1_epi64x((int64_t)counter), deltas); + *out_lo = _mm256_cvtepi64_epi32(counters); + *out_hi = _mm256_cvtepi64_epi32(_mm256_srli_epi64(counters, 32)); +} + +static +void blake3_hash4_avx512(const uint8_t *const *inputs, size_t blocks, + const uint32_t key[8], uint64_t counter, + bool increment_counter, uint8_t flags, + uint8_t flags_start, uint8_t flags_end, uint8_t *out) { + __m128i h_vecs[8] = { + set1_128(key[0]), set1_128(key[1]), set1_128(key[2]), set1_128(key[3]), + set1_128(key[4]), set1_128(key[5]), set1_128(key[6]), set1_128(key[7]), + }; + __m128i counter_low_vec, counter_high_vec; + load_counters4(counter, increment_counter, &counter_low_vec, + &counter_high_vec); + uint8_t block_flags = flags | flags_start; + + for (size_t block = 0; block < blocks; block++) { + if (block + 1 == blocks) { + block_flags |= flags_end; + } + __m128i block_len_vec = set1_128(BLAKE3_BLOCK_LEN); + __m128i block_flags_vec = set1_128(block_flags); + __m128i msg_vecs[16]; + transpose_msg_vecs4(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs); + + __m128i v[16] = { + h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3], + h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7], + set1_128(IV[0]), set1_128(IV[1]), set1_128(IV[2]), set1_128(IV[3]), + counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec, + }; + round_fn4(v, msg_vecs, 0); + round_fn4(v, msg_vecs, 1); + round_fn4(v, msg_vecs, 2); + round_fn4(v, msg_vecs, 3); + round_fn4(v, msg_vecs, 4); + round_fn4(v, msg_vecs, 5); + round_fn4(v, msg_vecs, 6); + h_vecs[0] = xor_128(v[0], v[8]); + h_vecs[1] = xor_128(v[1], v[9]); + h_vecs[2] = xor_128(v[2], v[10]); + h_vecs[3] = xor_128(v[3], v[11]); + h_vecs[4] = xor_128(v[4], v[12]); + h_vecs[5] = xor_128(v[5], v[13]); + h_vecs[6] = xor_128(v[6], v[14]); + h_vecs[7] = xor_128(v[7], v[15]); + + block_flags = flags; + } + + transpose_vecs_128(&h_vecs[0]); + transpose_vecs_128(&h_vecs[4]); + // The first four vecs now contain the first half of each output, and the + // second four vecs contain the second half of each output. + storeu_128(h_vecs[0], &out[0 * sizeof(__m128i)]); + storeu_128(h_vecs[4], &out[1 * sizeof(__m128i)]); + storeu_128(h_vecs[1], &out[2 * sizeof(__m128i)]); + storeu_128(h_vecs[5], &out[3 * sizeof(__m128i)]); + storeu_128(h_vecs[2], &out[4 * sizeof(__m128i)]); + storeu_128(h_vecs[6], &out[5 * sizeof(__m128i)]); + storeu_128(h_vecs[3], &out[6 * sizeof(__m128i)]); + storeu_128(h_vecs[7], &out[7 * sizeof(__m128i)]); +} + +static +void blake3_xof4_avx512(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, uint8_t flags, + uint8_t out[4 * 64]) { + __m128i h_vecs[8] = { + set1_128(cv[0]), set1_128(cv[1]), set1_128(cv[2]), set1_128(cv[3]), + set1_128(cv[4]), set1_128(cv[5]), set1_128(cv[6]), set1_128(cv[7]), + }; + uint32_t block_words[16]; + load_block_words(block, block_words); + __m128i msg_vecs[16]; + for (size_t i = 0; i < 16; i++) { + msg_vecs[i] = set1_128(block_words[i]); + } + __m128i counter_low_vec, counter_high_vec; + load_counters4(counter, true, &counter_low_vec, &counter_high_vec); + __m128i block_len_vec = set1_128(block_len); + __m128i block_flags_vec = set1_128(flags); + __m128i v[16] = { + h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3], + h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7], + set1_128(IV[0]), set1_128(IV[1]), set1_128(IV[2]), set1_128(IV[3]), + counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec, + }; + round_fn4(v, msg_vecs, 0); + round_fn4(v, msg_vecs, 1); + round_fn4(v, msg_vecs, 2); + round_fn4(v, msg_vecs, 3); + round_fn4(v, msg_vecs, 4); + round_fn4(v, msg_vecs, 5); + round_fn4(v, msg_vecs, 6); + for (size_t i = 0; i < 8; i++) { + v[i] = xor_128(v[i], v[i+8]); + v[i+8] = xor_128(v[i+8], h_vecs[i]); + } + transpose_vecs_128(&v[0]); + transpose_vecs_128(&v[4]); + transpose_vecs_128(&v[8]); + transpose_vecs_128(&v[12]); + for (size_t i = 0; i < 4; i++) { + storeu_128(v[i+ 0], &out[(4*i+0) * sizeof(__m128i)]); + storeu_128(v[i+ 4], &out[(4*i+1) * sizeof(__m128i)]); + storeu_128(v[i+ 8], &out[(4*i+2) * sizeof(__m128i)]); + storeu_128(v[i+12], &out[(4*i+3) * sizeof(__m128i)]); + } +} + +/* + * ---------------------------------------------------------------------------- + * hash8_avx512 + * ---------------------------------------------------------------------------- + */ + +INLINE void round_fn8(__m256i v[16], __m256i m[16], size_t r) { + v[0] = add_256(v[0], m[(size_t)MSG_SCHEDULE[r][0]]); + v[1] = add_256(v[1], m[(size_t)MSG_SCHEDULE[r][2]]); + v[2] = add_256(v[2], m[(size_t)MSG_SCHEDULE[r][4]]); + v[3] = add_256(v[3], m[(size_t)MSG_SCHEDULE[r][6]]); + v[0] = add_256(v[0], v[4]); + v[1] = add_256(v[1], v[5]); + v[2] = add_256(v[2], v[6]); + v[3] = add_256(v[3], v[7]); + v[12] = xor_256(v[12], v[0]); + v[13] = xor_256(v[13], v[1]); + v[14] = xor_256(v[14], v[2]); + v[15] = xor_256(v[15], v[3]); + v[12] = rot16_256(v[12]); + v[13] = rot16_256(v[13]); + v[14] = rot16_256(v[14]); + v[15] = rot16_256(v[15]); + v[8] = add_256(v[8], v[12]); + v[9] = add_256(v[9], v[13]); + v[10] = add_256(v[10], v[14]); + v[11] = add_256(v[11], v[15]); + v[4] = xor_256(v[4], v[8]); + v[5] = xor_256(v[5], v[9]); + v[6] = xor_256(v[6], v[10]); + v[7] = xor_256(v[7], v[11]); + v[4] = rot12_256(v[4]); + v[5] = rot12_256(v[5]); + v[6] = rot12_256(v[6]); + v[7] = rot12_256(v[7]); + v[0] = add_256(v[0], m[(size_t)MSG_SCHEDULE[r][1]]); + v[1] = add_256(v[1], m[(size_t)MSG_SCHEDULE[r][3]]); + v[2] = add_256(v[2], m[(size_t)MSG_SCHEDULE[r][5]]); + v[3] = add_256(v[3], m[(size_t)MSG_SCHEDULE[r][7]]); + v[0] = add_256(v[0], v[4]); + v[1] = add_256(v[1], v[5]); + v[2] = add_256(v[2], v[6]); + v[3] = add_256(v[3], v[7]); + v[12] = xor_256(v[12], v[0]); + v[13] = xor_256(v[13], v[1]); + v[14] = xor_256(v[14], v[2]); + v[15] = xor_256(v[15], v[3]); + v[12] = rot8_256(v[12]); + v[13] = rot8_256(v[13]); + v[14] = rot8_256(v[14]); + v[15] = rot8_256(v[15]); + v[8] = add_256(v[8], v[12]); + v[9] = add_256(v[9], v[13]); + v[10] = add_256(v[10], v[14]); + v[11] = add_256(v[11], v[15]); + v[4] = xor_256(v[4], v[8]); + v[5] = xor_256(v[5], v[9]); + v[6] = xor_256(v[6], v[10]); + v[7] = xor_256(v[7], v[11]); + v[4] = rot7_256(v[4]); + v[5] = rot7_256(v[5]); + v[6] = rot7_256(v[6]); + v[7] = rot7_256(v[7]); + + v[0] = add_256(v[0], m[(size_t)MSG_SCHEDULE[r][8]]); + v[1] = add_256(v[1], m[(size_t)MSG_SCHEDULE[r][10]]); + v[2] = add_256(v[2], m[(size_t)MSG_SCHEDULE[r][12]]); + v[3] = add_256(v[3], m[(size_t)MSG_SCHEDULE[r][14]]); + v[0] = add_256(v[0], v[5]); + v[1] = add_256(v[1], v[6]); + v[2] = add_256(v[2], v[7]); + v[3] = add_256(v[3], v[4]); + v[15] = xor_256(v[15], v[0]); + v[12] = xor_256(v[12], v[1]); + v[13] = xor_256(v[13], v[2]); + v[14] = xor_256(v[14], v[3]); + v[15] = rot16_256(v[15]); + v[12] = rot16_256(v[12]); + v[13] = rot16_256(v[13]); + v[14] = rot16_256(v[14]); + v[10] = add_256(v[10], v[15]); + v[11] = add_256(v[11], v[12]); + v[8] = add_256(v[8], v[13]); + v[9] = add_256(v[9], v[14]); + v[5] = xor_256(v[5], v[10]); + v[6] = xor_256(v[6], v[11]); + v[7] = xor_256(v[7], v[8]); + v[4] = xor_256(v[4], v[9]); + v[5] = rot12_256(v[5]); + v[6] = rot12_256(v[6]); + v[7] = rot12_256(v[7]); + v[4] = rot12_256(v[4]); + v[0] = add_256(v[0], m[(size_t)MSG_SCHEDULE[r][9]]); + v[1] = add_256(v[1], m[(size_t)MSG_SCHEDULE[r][11]]); + v[2] = add_256(v[2], m[(size_t)MSG_SCHEDULE[r][13]]); + v[3] = add_256(v[3], m[(size_t)MSG_SCHEDULE[r][15]]); + v[0] = add_256(v[0], v[5]); + v[1] = add_256(v[1], v[6]); + v[2] = add_256(v[2], v[7]); + v[3] = add_256(v[3], v[4]); + v[15] = xor_256(v[15], v[0]); + v[12] = xor_256(v[12], v[1]); + v[13] = xor_256(v[13], v[2]); + v[14] = xor_256(v[14], v[3]); + v[15] = rot8_256(v[15]); + v[12] = rot8_256(v[12]); + v[13] = rot8_256(v[13]); + v[14] = rot8_256(v[14]); + v[10] = add_256(v[10], v[15]); + v[11] = add_256(v[11], v[12]); + v[8] = add_256(v[8], v[13]); + v[9] = add_256(v[9], v[14]); + v[5] = xor_256(v[5], v[10]); + v[6] = xor_256(v[6], v[11]); + v[7] = xor_256(v[7], v[8]); + v[4] = xor_256(v[4], v[9]); + v[5] = rot7_256(v[5]); + v[6] = rot7_256(v[6]); + v[7] = rot7_256(v[7]); + v[4] = rot7_256(v[4]); +} + +INLINE void transpose_vecs_256(__m256i vecs[8]) { + // Interleave 32-bit lanes. The low unpack is lanes 00/11/44/55, and the high + // is 22/33/66/77. + __m256i ab_0145 = _mm256_unpacklo_epi32(vecs[0], vecs[1]); + __m256i ab_2367 = _mm256_unpackhi_epi32(vecs[0], vecs[1]); + __m256i cd_0145 = _mm256_unpacklo_epi32(vecs[2], vecs[3]); + __m256i cd_2367 = _mm256_unpackhi_epi32(vecs[2], vecs[3]); + __m256i ef_0145 = _mm256_unpacklo_epi32(vecs[4], vecs[5]); + __m256i ef_2367 = _mm256_unpackhi_epi32(vecs[4], vecs[5]); + __m256i gh_0145 = _mm256_unpacklo_epi32(vecs[6], vecs[7]); + __m256i gh_2367 = _mm256_unpackhi_epi32(vecs[6], vecs[7]); + + // Interleave 64-bit lanes. The low unpack is lanes 00/22 and the high is + // 11/33. + __m256i abcd_04 = _mm256_unpacklo_epi64(ab_0145, cd_0145); + __m256i abcd_15 = _mm256_unpackhi_epi64(ab_0145, cd_0145); + __m256i abcd_26 = _mm256_unpacklo_epi64(ab_2367, cd_2367); + __m256i abcd_37 = _mm256_unpackhi_epi64(ab_2367, cd_2367); + __m256i efgh_04 = _mm256_unpacklo_epi64(ef_0145, gh_0145); + __m256i efgh_15 = _mm256_unpackhi_epi64(ef_0145, gh_0145); + __m256i efgh_26 = _mm256_unpacklo_epi64(ef_2367, gh_2367); + __m256i efgh_37 = _mm256_unpackhi_epi64(ef_2367, gh_2367); + + // Interleave 128-bit lanes. + vecs[0] = _mm256_permute2x128_si256(abcd_04, efgh_04, 0x20); + vecs[1] = _mm256_permute2x128_si256(abcd_15, efgh_15, 0x20); + vecs[2] = _mm256_permute2x128_si256(abcd_26, efgh_26, 0x20); + vecs[3] = _mm256_permute2x128_si256(abcd_37, efgh_37, 0x20); + vecs[4] = _mm256_permute2x128_si256(abcd_04, efgh_04, 0x31); + vecs[5] = _mm256_permute2x128_si256(abcd_15, efgh_15, 0x31); + vecs[6] = _mm256_permute2x128_si256(abcd_26, efgh_26, 0x31); + vecs[7] = _mm256_permute2x128_si256(abcd_37, efgh_37, 0x31); +} + +INLINE void transpose_msg_vecs8(const uint8_t *const *inputs, + size_t block_offset, __m256i out[16]) { + out[0] = loadu_256(&inputs[0][block_offset + 0 * sizeof(__m256i)]); + out[1] = loadu_256(&inputs[1][block_offset + 0 * sizeof(__m256i)]); + out[2] = loadu_256(&inputs[2][block_offset + 0 * sizeof(__m256i)]); + out[3] = loadu_256(&inputs[3][block_offset + 0 * sizeof(__m256i)]); + out[4] = loadu_256(&inputs[4][block_offset + 0 * sizeof(__m256i)]); + out[5] = loadu_256(&inputs[5][block_offset + 0 * sizeof(__m256i)]); + out[6] = loadu_256(&inputs[6][block_offset + 0 * sizeof(__m256i)]); + out[7] = loadu_256(&inputs[7][block_offset + 0 * sizeof(__m256i)]); + out[8] = loadu_256(&inputs[0][block_offset + 1 * sizeof(__m256i)]); + out[9] = loadu_256(&inputs[1][block_offset + 1 * sizeof(__m256i)]); + out[10] = loadu_256(&inputs[2][block_offset + 1 * sizeof(__m256i)]); + out[11] = loadu_256(&inputs[3][block_offset + 1 * sizeof(__m256i)]); + out[12] = loadu_256(&inputs[4][block_offset + 1 * sizeof(__m256i)]); + out[13] = loadu_256(&inputs[5][block_offset + 1 * sizeof(__m256i)]); + out[14] = loadu_256(&inputs[6][block_offset + 1 * sizeof(__m256i)]); + out[15] = loadu_256(&inputs[7][block_offset + 1 * sizeof(__m256i)]); + for (size_t i = 0; i < 8; ++i) { + _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0); + } + transpose_vecs_256(&out[0]); + transpose_vecs_256(&out[8]); +} + +INLINE void load_counters8(uint64_t counter, bool increment_counter, + __m256i *out_lo, __m256i *out_hi) { + uint64_t mask = (increment_counter ? ~0 : 0); + __m512i mask_vec = _mm512_set1_epi64(mask); + __m512i deltas = _mm512_setr_epi64(0, 1, 2, 3, 4, 5, 6, 7); + deltas = _mm512_and_si512(mask_vec, deltas); + __m512i counters = + _mm512_add_epi64(_mm512_set1_epi64((int64_t)counter), deltas); + *out_lo = _mm512_cvtepi64_epi32(counters); + *out_hi = _mm512_cvtepi64_epi32(_mm512_srli_epi64(counters, 32)); +} + +static +void blake3_hash8_avx512(const uint8_t *const *inputs, size_t blocks, + const uint32_t key[8], uint64_t counter, + bool increment_counter, uint8_t flags, + uint8_t flags_start, uint8_t flags_end, uint8_t *out) { + __m256i h_vecs[8] = { + set1_256(key[0]), set1_256(key[1]), set1_256(key[2]), set1_256(key[3]), + set1_256(key[4]), set1_256(key[5]), set1_256(key[6]), set1_256(key[7]), + }; + __m256i counter_low_vec, counter_high_vec; + load_counters8(counter, increment_counter, &counter_low_vec, + &counter_high_vec); + uint8_t block_flags = flags | flags_start; + + for (size_t block = 0; block < blocks; block++) { + if (block + 1 == blocks) { + block_flags |= flags_end; + } + __m256i block_len_vec = set1_256(BLAKE3_BLOCK_LEN); + __m256i block_flags_vec = set1_256(block_flags); + __m256i msg_vecs[16]; + transpose_msg_vecs8(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs); + + __m256i v[16] = { + h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3], + h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7], + set1_256(IV[0]), set1_256(IV[1]), set1_256(IV[2]), set1_256(IV[3]), + counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec, + }; + round_fn8(v, msg_vecs, 0); + round_fn8(v, msg_vecs, 1); + round_fn8(v, msg_vecs, 2); + round_fn8(v, msg_vecs, 3); + round_fn8(v, msg_vecs, 4); + round_fn8(v, msg_vecs, 5); + round_fn8(v, msg_vecs, 6); + h_vecs[0] = xor_256(v[0], v[8]); + h_vecs[1] = xor_256(v[1], v[9]); + h_vecs[2] = xor_256(v[2], v[10]); + h_vecs[3] = xor_256(v[3], v[11]); + h_vecs[4] = xor_256(v[4], v[12]); + h_vecs[5] = xor_256(v[5], v[13]); + h_vecs[6] = xor_256(v[6], v[14]); + h_vecs[7] = xor_256(v[7], v[15]); + + block_flags = flags; + } + + transpose_vecs_256(h_vecs); + storeu_256(h_vecs[0], &out[0 * sizeof(__m256i)]); + storeu_256(h_vecs[1], &out[1 * sizeof(__m256i)]); + storeu_256(h_vecs[2], &out[2 * sizeof(__m256i)]); + storeu_256(h_vecs[3], &out[3 * sizeof(__m256i)]); + storeu_256(h_vecs[4], &out[4 * sizeof(__m256i)]); + storeu_256(h_vecs[5], &out[5 * sizeof(__m256i)]); + storeu_256(h_vecs[6], &out[6 * sizeof(__m256i)]); + storeu_256(h_vecs[7], &out[7 * sizeof(__m256i)]); +} + +static +void blake3_xof8_avx512(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, uint8_t flags, + uint8_t out[8 * 64]) { + __m256i h_vecs[8] = { + set1_256(cv[0]), set1_256(cv[1]), set1_256(cv[2]), set1_256(cv[3]), + set1_256(cv[4]), set1_256(cv[5]), set1_256(cv[6]), set1_256(cv[7]), + }; + uint32_t block_words[16]; + load_block_words(block, block_words); + __m256i msg_vecs[16]; + for (size_t i = 0; i < 16; i++) { + msg_vecs[i] = set1_256(block_words[i]); + } + __m256i counter_low_vec, counter_high_vec; + load_counters8(counter, true, &counter_low_vec, &counter_high_vec); + __m256i block_len_vec = set1_256(block_len); + __m256i block_flags_vec = set1_256(flags); + __m256i v[16] = { + h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3], + h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7], + set1_256(IV[0]), set1_256(IV[1]), set1_256(IV[2]), set1_256(IV[3]), + counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec, + }; + round_fn8(v, msg_vecs, 0); + round_fn8(v, msg_vecs, 1); + round_fn8(v, msg_vecs, 2); + round_fn8(v, msg_vecs, 3); + round_fn8(v, msg_vecs, 4); + round_fn8(v, msg_vecs, 5); + round_fn8(v, msg_vecs, 6); + for (size_t i = 0; i < 8; i++) { + v[i] = xor_256(v[i], v[i+8]); + v[i+8] = xor_256(v[i+8], h_vecs[i]); + } + transpose_vecs_256(&v[0]); + transpose_vecs_256(&v[8]); + for (size_t i = 0; i < 8; i++) { + storeu_256(v[i+0], &out[(2*i+0) * sizeof(__m256i)]); + storeu_256(v[i+8], &out[(2*i+1) * sizeof(__m256i)]); + } +} + +/* + * ---------------------------------------------------------------------------- + * hash16_avx512 + * ---------------------------------------------------------------------------- + */ + +INLINE void round_fn16(__m512i v[16], __m512i m[16], size_t r) { + v[0] = add_512(v[0], m[(size_t)MSG_SCHEDULE[r][0]]); + v[1] = add_512(v[1], m[(size_t)MSG_SCHEDULE[r][2]]); + v[2] = add_512(v[2], m[(size_t)MSG_SCHEDULE[r][4]]); + v[3] = add_512(v[3], m[(size_t)MSG_SCHEDULE[r][6]]); + v[0] = add_512(v[0], v[4]); + v[1] = add_512(v[1], v[5]); + v[2] = add_512(v[2], v[6]); + v[3] = add_512(v[3], v[7]); + v[12] = xor_512(v[12], v[0]); + v[13] = xor_512(v[13], v[1]); + v[14] = xor_512(v[14], v[2]); + v[15] = xor_512(v[15], v[3]); + v[12] = rot16_512(v[12]); + v[13] = rot16_512(v[13]); + v[14] = rot16_512(v[14]); + v[15] = rot16_512(v[15]); + v[8] = add_512(v[8], v[12]); + v[9] = add_512(v[9], v[13]); + v[10] = add_512(v[10], v[14]); + v[11] = add_512(v[11], v[15]); + v[4] = xor_512(v[4], v[8]); + v[5] = xor_512(v[5], v[9]); + v[6] = xor_512(v[6], v[10]); + v[7] = xor_512(v[7], v[11]); + v[4] = rot12_512(v[4]); + v[5] = rot12_512(v[5]); + v[6] = rot12_512(v[6]); + v[7] = rot12_512(v[7]); + v[0] = add_512(v[0], m[(size_t)MSG_SCHEDULE[r][1]]); + v[1] = add_512(v[1], m[(size_t)MSG_SCHEDULE[r][3]]); + v[2] = add_512(v[2], m[(size_t)MSG_SCHEDULE[r][5]]); + v[3] = add_512(v[3], m[(size_t)MSG_SCHEDULE[r][7]]); + v[0] = add_512(v[0], v[4]); + v[1] = add_512(v[1], v[5]); + v[2] = add_512(v[2], v[6]); + v[3] = add_512(v[3], v[7]); + v[12] = xor_512(v[12], v[0]); + v[13] = xor_512(v[13], v[1]); + v[14] = xor_512(v[14], v[2]); + v[15] = xor_512(v[15], v[3]); + v[12] = rot8_512(v[12]); + v[13] = rot8_512(v[13]); + v[14] = rot8_512(v[14]); + v[15] = rot8_512(v[15]); + v[8] = add_512(v[8], v[12]); + v[9] = add_512(v[9], v[13]); + v[10] = add_512(v[10], v[14]); + v[11] = add_512(v[11], v[15]); + v[4] = xor_512(v[4], v[8]); + v[5] = xor_512(v[5], v[9]); + v[6] = xor_512(v[6], v[10]); + v[7] = xor_512(v[7], v[11]); + v[4] = rot7_512(v[4]); + v[5] = rot7_512(v[5]); + v[6] = rot7_512(v[6]); + v[7] = rot7_512(v[7]); + + v[0] = add_512(v[0], m[(size_t)MSG_SCHEDULE[r][8]]); + v[1] = add_512(v[1], m[(size_t)MSG_SCHEDULE[r][10]]); + v[2] = add_512(v[2], m[(size_t)MSG_SCHEDULE[r][12]]); + v[3] = add_512(v[3], m[(size_t)MSG_SCHEDULE[r][14]]); + v[0] = add_512(v[0], v[5]); + v[1] = add_512(v[1], v[6]); + v[2] = add_512(v[2], v[7]); + v[3] = add_512(v[3], v[4]); + v[15] = xor_512(v[15], v[0]); + v[12] = xor_512(v[12], v[1]); + v[13] = xor_512(v[13], v[2]); + v[14] = xor_512(v[14], v[3]); + v[15] = rot16_512(v[15]); + v[12] = rot16_512(v[12]); + v[13] = rot16_512(v[13]); + v[14] = rot16_512(v[14]); + v[10] = add_512(v[10], v[15]); + v[11] = add_512(v[11], v[12]); + v[8] = add_512(v[8], v[13]); + v[9] = add_512(v[9], v[14]); + v[5] = xor_512(v[5], v[10]); + v[6] = xor_512(v[6], v[11]); + v[7] = xor_512(v[7], v[8]); + v[4] = xor_512(v[4], v[9]); + v[5] = rot12_512(v[5]); + v[6] = rot12_512(v[6]); + v[7] = rot12_512(v[7]); + v[4] = rot12_512(v[4]); + v[0] = add_512(v[0], m[(size_t)MSG_SCHEDULE[r][9]]); + v[1] = add_512(v[1], m[(size_t)MSG_SCHEDULE[r][11]]); + v[2] = add_512(v[2], m[(size_t)MSG_SCHEDULE[r][13]]); + v[3] = add_512(v[3], m[(size_t)MSG_SCHEDULE[r][15]]); + v[0] = add_512(v[0], v[5]); + v[1] = add_512(v[1], v[6]); + v[2] = add_512(v[2], v[7]); + v[3] = add_512(v[3], v[4]); + v[15] = xor_512(v[15], v[0]); + v[12] = xor_512(v[12], v[1]); + v[13] = xor_512(v[13], v[2]); + v[14] = xor_512(v[14], v[3]); + v[15] = rot8_512(v[15]); + v[12] = rot8_512(v[12]); + v[13] = rot8_512(v[13]); + v[14] = rot8_512(v[14]); + v[10] = add_512(v[10], v[15]); + v[11] = add_512(v[11], v[12]); + v[8] = add_512(v[8], v[13]); + v[9] = add_512(v[9], v[14]); + v[5] = xor_512(v[5], v[10]); + v[6] = xor_512(v[6], v[11]); + v[7] = xor_512(v[7], v[8]); + v[4] = xor_512(v[4], v[9]); + v[5] = rot7_512(v[5]); + v[6] = rot7_512(v[6]); + v[7] = rot7_512(v[7]); + v[4] = rot7_512(v[4]); +} + +// 0b10001000, or lanes a0/a2/b0/b2 in little-endian order +#define LO_IMM8 0x88 + +INLINE __m512i unpack_lo_128(__m512i a, __m512i b) { + return _mm512_shuffle_i32x4(a, b, LO_IMM8); +} + +// 0b11011101, or lanes a1/a3/b1/b3 in little-endian order +#define HI_IMM8 0xdd + +INLINE __m512i unpack_hi_128(__m512i a, __m512i b) { + return _mm512_shuffle_i32x4(a, b, HI_IMM8); +} + +INLINE void transpose_vecs_512(__m512i vecs[16]) { + // Interleave 32-bit lanes. The _0 unpack is lanes + // 0/0/1/1/4/4/5/5/8/8/9/9/12/12/13/13, and the _2 unpack is lanes + // 2/2/3/3/6/6/7/7/10/10/11/11/14/14/15/15. + __m512i ab_0 = _mm512_unpacklo_epi32(vecs[0], vecs[1]); + __m512i ab_2 = _mm512_unpackhi_epi32(vecs[0], vecs[1]); + __m512i cd_0 = _mm512_unpacklo_epi32(vecs[2], vecs[3]); + __m512i cd_2 = _mm512_unpackhi_epi32(vecs[2], vecs[3]); + __m512i ef_0 = _mm512_unpacklo_epi32(vecs[4], vecs[5]); + __m512i ef_2 = _mm512_unpackhi_epi32(vecs[4], vecs[5]); + __m512i gh_0 = _mm512_unpacklo_epi32(vecs[6], vecs[7]); + __m512i gh_2 = _mm512_unpackhi_epi32(vecs[6], vecs[7]); + __m512i ij_0 = _mm512_unpacklo_epi32(vecs[8], vecs[9]); + __m512i ij_2 = _mm512_unpackhi_epi32(vecs[8], vecs[9]); + __m512i kl_0 = _mm512_unpacklo_epi32(vecs[10], vecs[11]); + __m512i kl_2 = _mm512_unpackhi_epi32(vecs[10], vecs[11]); + __m512i mn_0 = _mm512_unpacklo_epi32(vecs[12], vecs[13]); + __m512i mn_2 = _mm512_unpackhi_epi32(vecs[12], vecs[13]); + __m512i op_0 = _mm512_unpacklo_epi32(vecs[14], vecs[15]); + __m512i op_2 = _mm512_unpackhi_epi32(vecs[14], vecs[15]); + + // Interleave 64-bit lanes. The _0 unpack is lanes + // 0/0/0/0/4/4/4/4/8/8/8/8/12/12/12/12, the _1 unpack is lanes + // 1/1/1/1/5/5/5/5/9/9/9/9/13/13/13/13, the _2 unpack is lanes + // 2/2/2/2/6/6/6/6/10/10/10/10/14/14/14/14, and the _3 unpack is lanes + // 3/3/3/3/7/7/7/7/11/11/11/11/15/15/15/15. + __m512i abcd_0 = _mm512_unpacklo_epi64(ab_0, cd_0); + __m512i abcd_1 = _mm512_unpackhi_epi64(ab_0, cd_0); + __m512i abcd_2 = _mm512_unpacklo_epi64(ab_2, cd_2); + __m512i abcd_3 = _mm512_unpackhi_epi64(ab_2, cd_2); + __m512i efgh_0 = _mm512_unpacklo_epi64(ef_0, gh_0); + __m512i efgh_1 = _mm512_unpackhi_epi64(ef_0, gh_0); + __m512i efgh_2 = _mm512_unpacklo_epi64(ef_2, gh_2); + __m512i efgh_3 = _mm512_unpackhi_epi64(ef_2, gh_2); + __m512i ijkl_0 = _mm512_unpacklo_epi64(ij_0, kl_0); + __m512i ijkl_1 = _mm512_unpackhi_epi64(ij_0, kl_0); + __m512i ijkl_2 = _mm512_unpacklo_epi64(ij_2, kl_2); + __m512i ijkl_3 = _mm512_unpackhi_epi64(ij_2, kl_2); + __m512i mnop_0 = _mm512_unpacklo_epi64(mn_0, op_0); + __m512i mnop_1 = _mm512_unpackhi_epi64(mn_0, op_0); + __m512i mnop_2 = _mm512_unpacklo_epi64(mn_2, op_2); + __m512i mnop_3 = _mm512_unpackhi_epi64(mn_2, op_2); + + // Interleave 128-bit lanes. The _0 unpack is + // 0/0/0/0/8/8/8/8/0/0/0/0/8/8/8/8, the _1 unpack is + // 1/1/1/1/9/9/9/9/1/1/1/1/9/9/9/9, and so on. + __m512i abcdefgh_0 = unpack_lo_128(abcd_0, efgh_0); + __m512i abcdefgh_1 = unpack_lo_128(abcd_1, efgh_1); + __m512i abcdefgh_2 = unpack_lo_128(abcd_2, efgh_2); + __m512i abcdefgh_3 = unpack_lo_128(abcd_3, efgh_3); + __m512i abcdefgh_4 = unpack_hi_128(abcd_0, efgh_0); + __m512i abcdefgh_5 = unpack_hi_128(abcd_1, efgh_1); + __m512i abcdefgh_6 = unpack_hi_128(abcd_2, efgh_2); + __m512i abcdefgh_7 = unpack_hi_128(abcd_3, efgh_3); + __m512i ijklmnop_0 = unpack_lo_128(ijkl_0, mnop_0); + __m512i ijklmnop_1 = unpack_lo_128(ijkl_1, mnop_1); + __m512i ijklmnop_2 = unpack_lo_128(ijkl_2, mnop_2); + __m512i ijklmnop_3 = unpack_lo_128(ijkl_3, mnop_3); + __m512i ijklmnop_4 = unpack_hi_128(ijkl_0, mnop_0); + __m512i ijklmnop_5 = unpack_hi_128(ijkl_1, mnop_1); + __m512i ijklmnop_6 = unpack_hi_128(ijkl_2, mnop_2); + __m512i ijklmnop_7 = unpack_hi_128(ijkl_3, mnop_3); + + // Interleave 128-bit lanes again for the final outputs. + vecs[0] = unpack_lo_128(abcdefgh_0, ijklmnop_0); + vecs[1] = unpack_lo_128(abcdefgh_1, ijklmnop_1); + vecs[2] = unpack_lo_128(abcdefgh_2, ijklmnop_2); + vecs[3] = unpack_lo_128(abcdefgh_3, ijklmnop_3); + vecs[4] = unpack_lo_128(abcdefgh_4, ijklmnop_4); + vecs[5] = unpack_lo_128(abcdefgh_5, ijklmnop_5); + vecs[6] = unpack_lo_128(abcdefgh_6, ijklmnop_6); + vecs[7] = unpack_lo_128(abcdefgh_7, ijklmnop_7); + vecs[8] = unpack_hi_128(abcdefgh_0, ijklmnop_0); + vecs[9] = unpack_hi_128(abcdefgh_1, ijklmnop_1); + vecs[10] = unpack_hi_128(abcdefgh_2, ijklmnop_2); + vecs[11] = unpack_hi_128(abcdefgh_3, ijklmnop_3); + vecs[12] = unpack_hi_128(abcdefgh_4, ijklmnop_4); + vecs[13] = unpack_hi_128(abcdefgh_5, ijklmnop_5); + vecs[14] = unpack_hi_128(abcdefgh_6, ijklmnop_6); + vecs[15] = unpack_hi_128(abcdefgh_7, ijklmnop_7); +} + +INLINE void transpose_msg_vecs16(const uint8_t *const *inputs, + size_t block_offset, __m512i out[16]) { + out[0] = loadu_512(&inputs[0][block_offset]); + out[1] = loadu_512(&inputs[1][block_offset]); + out[2] = loadu_512(&inputs[2][block_offset]); + out[3] = loadu_512(&inputs[3][block_offset]); + out[4] = loadu_512(&inputs[4][block_offset]); + out[5] = loadu_512(&inputs[5][block_offset]); + out[6] = loadu_512(&inputs[6][block_offset]); + out[7] = loadu_512(&inputs[7][block_offset]); + out[8] = loadu_512(&inputs[8][block_offset]); + out[9] = loadu_512(&inputs[9][block_offset]); + out[10] = loadu_512(&inputs[10][block_offset]); + out[11] = loadu_512(&inputs[11][block_offset]); + out[12] = loadu_512(&inputs[12][block_offset]); + out[13] = loadu_512(&inputs[13][block_offset]); + out[14] = loadu_512(&inputs[14][block_offset]); + out[15] = loadu_512(&inputs[15][block_offset]); + for (size_t i = 0; i < 16; ++i) { + _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0); + } + transpose_vecs_512(out); +} + +INLINE void load_counters16(uint64_t counter, bool increment_counter, + __m512i *out_lo, __m512i *out_hi) { + const __m512i mask = _mm512_set1_epi32(-(int32_t)increment_counter); + const __m512i deltas = _mm512_set_epi32(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0); + const __m512i masked_deltas = _mm512_and_si512(deltas, mask); + const __m512i low_words = _mm512_add_epi32( + _mm512_set1_epi32((int32_t)counter), + masked_deltas); + // The carry bit is 1 if the high bit of the word was 1 before addition and is + // 0 after. + // NOTE: It would be a bit more natural to use _mm512_cmp_epu32_mask to + // compute the carry bits here, and originally we did, but that intrinsic is + // broken under GCC 5.4. See https://github.com/BLAKE3-team/BLAKE3/issues/271. + const __m512i carries = _mm512_srli_epi32( + _mm512_andnot_si512( + low_words, // 0 after (gets inverted by andnot) + _mm512_set1_epi32((int32_t)counter)), // and 1 before + 31); + const __m512i high_words = _mm512_add_epi32( + _mm512_set1_epi32((int32_t)(counter >> 32)), + carries); + *out_lo = low_words; + *out_hi = high_words; +} + +static +void blake3_hash16_avx512(const uint8_t *const *inputs, size_t blocks, + const uint32_t key[8], uint64_t counter, + bool increment_counter, uint8_t flags, + uint8_t flags_start, uint8_t flags_end, + uint8_t *out) { + __m512i h_vecs[8] = { + set1_512(key[0]), set1_512(key[1]), set1_512(key[2]), set1_512(key[3]), + set1_512(key[4]), set1_512(key[5]), set1_512(key[6]), set1_512(key[7]), + }; + __m512i counter_low_vec, counter_high_vec; + load_counters16(counter, increment_counter, &counter_low_vec, + &counter_high_vec); + uint8_t block_flags = flags | flags_start; + + for (size_t block = 0; block < blocks; block++) { + if (block + 1 == blocks) { + block_flags |= flags_end; + } + __m512i block_len_vec = set1_512(BLAKE3_BLOCK_LEN); + __m512i block_flags_vec = set1_512(block_flags); + __m512i msg_vecs[16]; + transpose_msg_vecs16(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs); + + __m512i v[16] = { + h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3], + h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7], + set1_512(IV[0]), set1_512(IV[1]), set1_512(IV[2]), set1_512(IV[3]), + counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec, + }; + round_fn16(v, msg_vecs, 0); + round_fn16(v, msg_vecs, 1); + round_fn16(v, msg_vecs, 2); + round_fn16(v, msg_vecs, 3); + round_fn16(v, msg_vecs, 4); + round_fn16(v, msg_vecs, 5); + round_fn16(v, msg_vecs, 6); + h_vecs[0] = xor_512(v[0], v[8]); + h_vecs[1] = xor_512(v[1], v[9]); + h_vecs[2] = xor_512(v[2], v[10]); + h_vecs[3] = xor_512(v[3], v[11]); + h_vecs[4] = xor_512(v[4], v[12]); + h_vecs[5] = xor_512(v[5], v[13]); + h_vecs[6] = xor_512(v[6], v[14]); + h_vecs[7] = xor_512(v[7], v[15]); + + block_flags = flags; + } + + // transpose_vecs_512 operates on a 16x16 matrix of words, but we only have 8 + // state vectors. Pad the matrix with zeros. After transposition, store the + // lower half of each vector. + __m512i padded[16] = { + h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3], + h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7], + set1_512(0), set1_512(0), set1_512(0), set1_512(0), + set1_512(0), set1_512(0), set1_512(0), set1_512(0), + }; + transpose_vecs_512(padded); + _mm256_mask_storeu_epi32(&out[0 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[0])); + _mm256_mask_storeu_epi32(&out[1 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[1])); + _mm256_mask_storeu_epi32(&out[2 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[2])); + _mm256_mask_storeu_epi32(&out[3 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[3])); + _mm256_mask_storeu_epi32(&out[4 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[4])); + _mm256_mask_storeu_epi32(&out[5 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[5])); + _mm256_mask_storeu_epi32(&out[6 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[6])); + _mm256_mask_storeu_epi32(&out[7 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[7])); + _mm256_mask_storeu_epi32(&out[8 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[8])); + _mm256_mask_storeu_epi32(&out[9 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[9])); + _mm256_mask_storeu_epi32(&out[10 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[10])); + _mm256_mask_storeu_epi32(&out[11 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[11])); + _mm256_mask_storeu_epi32(&out[12 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[12])); + _mm256_mask_storeu_epi32(&out[13 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[13])); + _mm256_mask_storeu_epi32(&out[14 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[14])); + _mm256_mask_storeu_epi32(&out[15 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[15])); +} + +static +void blake3_xof16_avx512(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, uint8_t flags, + uint8_t out[16 * 64]) { + __m512i h_vecs[8] = { + set1_512(cv[0]), set1_512(cv[1]), set1_512(cv[2]), set1_512(cv[3]), + set1_512(cv[4]), set1_512(cv[5]), set1_512(cv[6]), set1_512(cv[7]), + }; + uint32_t block_words[16]; + load_block_words(block, block_words); + __m512i msg_vecs[16]; + for (size_t i = 0; i < 16; i++) { + msg_vecs[i] = set1_512(block_words[i]); + } + __m512i counter_low_vec, counter_high_vec; + load_counters16(counter, true, &counter_low_vec, &counter_high_vec); + __m512i block_len_vec = set1_512(block_len); + __m512i block_flags_vec = set1_512(flags); + __m512i v[16] = { + h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3], + h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7], + set1_512(IV[0]), set1_512(IV[1]), set1_512(IV[2]), set1_512(IV[3]), + counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec, + }; + round_fn16(v, msg_vecs, 0); + round_fn16(v, msg_vecs, 1); + round_fn16(v, msg_vecs, 2); + round_fn16(v, msg_vecs, 3); + round_fn16(v, msg_vecs, 4); + round_fn16(v, msg_vecs, 5); + round_fn16(v, msg_vecs, 6); + for (size_t i = 0; i < 8; i++) { + v[i] = xor_512(v[i], v[i+8]); + v[i+8] = xor_512(v[i+8], h_vecs[i]); + } + transpose_vecs_512(&v[0]); + for (size_t i = 0; i < 16; i++) { + storeu_512(v[i], &out[i * sizeof(__m512i)]); + } +} + +/* + * ---------------------------------------------------------------------------- + * hash_many_avx512 + * ---------------------------------------------------------------------------- + */ + +INLINE void hash_one_avx512(const uint8_t *input, size_t blocks, + const uint32_t key[8], uint64_t counter, + uint8_t flags, uint8_t flags_start, + uint8_t flags_end, uint8_t out[BLAKE3_OUT_LEN]) { + uint32_t cv[8]; + memcpy(cv, key, BLAKE3_KEY_LEN); + uint8_t block_flags = flags | flags_start; + while (blocks > 0) { + if (blocks == 1) { + block_flags |= flags_end; + } + blake3_compress_in_place_avx512(cv, input, BLAKE3_BLOCK_LEN, counter, + block_flags); + input = &input[BLAKE3_BLOCK_LEN]; + blocks -= 1; + block_flags = flags; + } + memcpy(out, cv, BLAKE3_OUT_LEN); +} + +void blake3_hash_many_avx512(const uint8_t *const *inputs, size_t num_inputs, + size_t blocks, const uint32_t key[8], + uint64_t counter, bool increment_counter, + uint8_t flags, uint8_t flags_start, + uint8_t flags_end, uint8_t *out) { + while (num_inputs >= 16) { + blake3_hash16_avx512(inputs, blocks, key, counter, increment_counter, flags, + flags_start, flags_end, out); + if (increment_counter) { + counter += 16; + } + inputs += 16; + num_inputs -= 16; + out = &out[16 * BLAKE3_OUT_LEN]; + } + while (num_inputs >= 8) { + blake3_hash8_avx512(inputs, blocks, key, counter, increment_counter, flags, + flags_start, flags_end, out); + if (increment_counter) { + counter += 8; + } + inputs += 8; + num_inputs -= 8; + out = &out[8 * BLAKE3_OUT_LEN]; + } + while (num_inputs >= 4) { + blake3_hash4_avx512(inputs, blocks, key, counter, increment_counter, flags, + flags_start, flags_end, out); + if (increment_counter) { + counter += 4; + } + inputs += 4; + num_inputs -= 4; + out = &out[4 * BLAKE3_OUT_LEN]; + } + while (num_inputs > 0) { + hash_one_avx512(inputs[0], blocks, key, counter, flags, flags_start, + flags_end, out); + if (increment_counter) { + counter += 1; + } + inputs += 1; + num_inputs -= 1; + out = &out[BLAKE3_OUT_LEN]; + } +} + +void blake3_xof_many_avx512(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, uint8_t flags, + uint8_t* out, size_t outblocks) { + while (outblocks >= 16) { + blake3_xof16_avx512(cv, block, block_len, counter, flags, out); + counter += 16; + outblocks -= 16; + out += 16 * BLAKE3_BLOCK_LEN; + } + while (outblocks >= 8) { + blake3_xof8_avx512(cv, block, block_len, counter, flags, out); + counter += 8; + outblocks -= 8; + out += 8 * BLAKE3_BLOCK_LEN; + } + while (outblocks >= 4) { + blake3_xof4_avx512(cv, block, block_len, counter, flags, out); + counter += 4; + outblocks -= 4; + out += 4 * BLAKE3_BLOCK_LEN; + } + while (outblocks > 0) { + blake3_compress_xof_avx512(cv, block, block_len, counter, flags, out); + counter += 1; + outblocks -= 1; + out += BLAKE3_BLOCK_LEN; + } +} diff --git a/ext/digest/blake3/blake3_neon.c b/ext/digest/blake3/blake3_neon.c new file mode 100644 index 0000000..794ea80 --- /dev/null +++ b/ext/digest/blake3/blake3_neon.c @@ -0,0 +1,367 @@ +#include "blake3_impl.h" + +#include + +#ifdef __ARM_BIG_ENDIAN +#error "This implementation only supports little-endian ARM." +// It might be that all we need for big-endian support here is to get the loads +// and stores right, but step zero would be finding a way to test it in CI. +#endif + +INLINE uint32x4_t loadu_128(const uint8_t src[16]) { + // vld1q_u32 has alignment requirements. Don't use it. + return vreinterpretq_u32_u8(vld1q_u8(src)); +} + +INLINE void storeu_128(uint32x4_t src, uint8_t dest[16]) { + // vst1q_u32 has alignment requirements. Don't use it. + vst1q_u8(dest, vreinterpretq_u8_u32(src)); +} + +INLINE uint32x4_t add_128(uint32x4_t a, uint32x4_t b) { + return vaddq_u32(a, b); +} + +INLINE uint32x4_t xor_128(uint32x4_t a, uint32x4_t b) { + return veorq_u32(a, b); +} + +INLINE uint32x4_t set1_128(uint32_t x) { return vld1q_dup_u32(&x); } + +INLINE uint32x4_t set4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) { + uint32_t array[4] = {a, b, c, d}; + return vld1q_u32(array); +} + +INLINE uint32x4_t rot16_128(uint32x4_t x) { + // The straightforward implementation would be two shifts and an or, but that's + // slower on microarchitectures we've tested. See + // https://github.com/BLAKE3-team/BLAKE3/pull/319. + // return vorrq_u32(vshrq_n_u32(x, 16), vshlq_n_u32(x, 32 - 16)); + return vreinterpretq_u32_u16(vrev32q_u16(vreinterpretq_u16_u32(x))); +} + +INLINE uint32x4_t rot12_128(uint32x4_t x) { + // See comment in rot16_128. + // return vorrq_u32(vshrq_n_u32(x, 12), vshlq_n_u32(x, 32 - 12)); + return vsriq_n_u32(vshlq_n_u32(x, 32-12), x, 12); +} + +INLINE uint32x4_t rot8_128(uint32x4_t x) { + // See comment in rot16_128. + // return vorrq_u32(vshrq_n_u32(x, 8), vshlq_n_u32(x, 32 - 8)); +#if defined(__clang__) + return vreinterpretq_u32_u8(__builtin_shufflevector(vreinterpretq_u8_u32(x), vreinterpretq_u8_u32(x), 1,2,3,0,5,6,7,4,9,10,11,8,13,14,15,12)); +#elif __GNUC__ * 10000 + __GNUC_MINOR__ * 100 >=40700 + static const uint8x16_t r8 = {1,2,3,0,5,6,7,4,9,10,11,8,13,14,15,12}; + return vreinterpretq_u32_u8(__builtin_shuffle(vreinterpretq_u8_u32(x), vreinterpretq_u8_u32(x), r8)); +#else + return vsriq_n_u32(vshlq_n_u32(x, 32-8), x, 8); +#endif +} + +INLINE uint32x4_t rot7_128(uint32x4_t x) { + // See comment in rot16_128. + // return vorrq_u32(vshrq_n_u32(x, 7), vshlq_n_u32(x, 32 - 7)); + return vsriq_n_u32(vshlq_n_u32(x, 32-7), x, 7); +} + +// TODO: compress_neon + +// TODO: hash2_neon + +/* + * ---------------------------------------------------------------------------- + * hash4_neon + * ---------------------------------------------------------------------------- + */ + +INLINE void round_fn4(uint32x4_t v[16], uint32x4_t m[16], size_t r) { + v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][0]]); + v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][2]]); + v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][4]]); + v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][6]]); + v[0] = add_128(v[0], v[4]); + v[1] = add_128(v[1], v[5]); + v[2] = add_128(v[2], v[6]); + v[3] = add_128(v[3], v[7]); + v[12] = xor_128(v[12], v[0]); + v[13] = xor_128(v[13], v[1]); + v[14] = xor_128(v[14], v[2]); + v[15] = xor_128(v[15], v[3]); + v[12] = rot16_128(v[12]); + v[13] = rot16_128(v[13]); + v[14] = rot16_128(v[14]); + v[15] = rot16_128(v[15]); + v[8] = add_128(v[8], v[12]); + v[9] = add_128(v[9], v[13]); + v[10] = add_128(v[10], v[14]); + v[11] = add_128(v[11], v[15]); + v[4] = xor_128(v[4], v[8]); + v[5] = xor_128(v[5], v[9]); + v[6] = xor_128(v[6], v[10]); + v[7] = xor_128(v[7], v[11]); + v[4] = rot12_128(v[4]); + v[5] = rot12_128(v[5]); + v[6] = rot12_128(v[6]); + v[7] = rot12_128(v[7]); + v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][1]]); + v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][3]]); + v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][5]]); + v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][7]]); + v[0] = add_128(v[0], v[4]); + v[1] = add_128(v[1], v[5]); + v[2] = add_128(v[2], v[6]); + v[3] = add_128(v[3], v[7]); + v[12] = xor_128(v[12], v[0]); + v[13] = xor_128(v[13], v[1]); + v[14] = xor_128(v[14], v[2]); + v[15] = xor_128(v[15], v[3]); + v[12] = rot8_128(v[12]); + v[13] = rot8_128(v[13]); + v[14] = rot8_128(v[14]); + v[15] = rot8_128(v[15]); + v[8] = add_128(v[8], v[12]); + v[9] = add_128(v[9], v[13]); + v[10] = add_128(v[10], v[14]); + v[11] = add_128(v[11], v[15]); + v[4] = xor_128(v[4], v[8]); + v[5] = xor_128(v[5], v[9]); + v[6] = xor_128(v[6], v[10]); + v[7] = xor_128(v[7], v[11]); + v[4] = rot7_128(v[4]); + v[5] = rot7_128(v[5]); + v[6] = rot7_128(v[6]); + v[7] = rot7_128(v[7]); + + v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][8]]); + v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][10]]); + v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][12]]); + v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][14]]); + v[0] = add_128(v[0], v[5]); + v[1] = add_128(v[1], v[6]); + v[2] = add_128(v[2], v[7]); + v[3] = add_128(v[3], v[4]); + v[15] = xor_128(v[15], v[0]); + v[12] = xor_128(v[12], v[1]); + v[13] = xor_128(v[13], v[2]); + v[14] = xor_128(v[14], v[3]); + v[15] = rot16_128(v[15]); + v[12] = rot16_128(v[12]); + v[13] = rot16_128(v[13]); + v[14] = rot16_128(v[14]); + v[10] = add_128(v[10], v[15]); + v[11] = add_128(v[11], v[12]); + v[8] = add_128(v[8], v[13]); + v[9] = add_128(v[9], v[14]); + v[5] = xor_128(v[5], v[10]); + v[6] = xor_128(v[6], v[11]); + v[7] = xor_128(v[7], v[8]); + v[4] = xor_128(v[4], v[9]); + v[5] = rot12_128(v[5]); + v[6] = rot12_128(v[6]); + v[7] = rot12_128(v[7]); + v[4] = rot12_128(v[4]); + v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][9]]); + v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][11]]); + v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][13]]); + v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][15]]); + v[0] = add_128(v[0], v[5]); + v[1] = add_128(v[1], v[6]); + v[2] = add_128(v[2], v[7]); + v[3] = add_128(v[3], v[4]); + v[15] = xor_128(v[15], v[0]); + v[12] = xor_128(v[12], v[1]); + v[13] = xor_128(v[13], v[2]); + v[14] = xor_128(v[14], v[3]); + v[15] = rot8_128(v[15]); + v[12] = rot8_128(v[12]); + v[13] = rot8_128(v[13]); + v[14] = rot8_128(v[14]); + v[10] = add_128(v[10], v[15]); + v[11] = add_128(v[11], v[12]); + v[8] = add_128(v[8], v[13]); + v[9] = add_128(v[9], v[14]); + v[5] = xor_128(v[5], v[10]); + v[6] = xor_128(v[6], v[11]); + v[7] = xor_128(v[7], v[8]); + v[4] = xor_128(v[4], v[9]); + v[5] = rot7_128(v[5]); + v[6] = rot7_128(v[6]); + v[7] = rot7_128(v[7]); + v[4] = rot7_128(v[4]); +} + +INLINE void transpose_vecs_128(uint32x4_t vecs[4]) { + // Individually transpose the four 2x2 sub-matrices in each corner. + uint32x4x2_t rows01 = vtrnq_u32(vecs[0], vecs[1]); + uint32x4x2_t rows23 = vtrnq_u32(vecs[2], vecs[3]); + + // Swap the top-right and bottom-left 2x2s (which just got transposed). + vecs[0] = + vcombine_u32(vget_low_u32(rows01.val[0]), vget_low_u32(rows23.val[0])); + vecs[1] = + vcombine_u32(vget_low_u32(rows01.val[1]), vget_low_u32(rows23.val[1])); + vecs[2] = + vcombine_u32(vget_high_u32(rows01.val[0]), vget_high_u32(rows23.val[0])); + vecs[3] = + vcombine_u32(vget_high_u32(rows01.val[1]), vget_high_u32(rows23.val[1])); +} + +INLINE void transpose_msg_vecs4(const uint8_t *const *inputs, + size_t block_offset, uint32x4_t out[16]) { + out[0] = loadu_128(&inputs[0][block_offset + 0 * sizeof(uint32x4_t)]); + out[1] = loadu_128(&inputs[1][block_offset + 0 * sizeof(uint32x4_t)]); + out[2] = loadu_128(&inputs[2][block_offset + 0 * sizeof(uint32x4_t)]); + out[3] = loadu_128(&inputs[3][block_offset + 0 * sizeof(uint32x4_t)]); + out[4] = loadu_128(&inputs[0][block_offset + 1 * sizeof(uint32x4_t)]); + out[5] = loadu_128(&inputs[1][block_offset + 1 * sizeof(uint32x4_t)]); + out[6] = loadu_128(&inputs[2][block_offset + 1 * sizeof(uint32x4_t)]); + out[7] = loadu_128(&inputs[3][block_offset + 1 * sizeof(uint32x4_t)]); + out[8] = loadu_128(&inputs[0][block_offset + 2 * sizeof(uint32x4_t)]); + out[9] = loadu_128(&inputs[1][block_offset + 2 * sizeof(uint32x4_t)]); + out[10] = loadu_128(&inputs[2][block_offset + 2 * sizeof(uint32x4_t)]); + out[11] = loadu_128(&inputs[3][block_offset + 2 * sizeof(uint32x4_t)]); + out[12] = loadu_128(&inputs[0][block_offset + 3 * sizeof(uint32x4_t)]); + out[13] = loadu_128(&inputs[1][block_offset + 3 * sizeof(uint32x4_t)]); + out[14] = loadu_128(&inputs[2][block_offset + 3 * sizeof(uint32x4_t)]); + out[15] = loadu_128(&inputs[3][block_offset + 3 * sizeof(uint32x4_t)]); + transpose_vecs_128(&out[0]); + transpose_vecs_128(&out[4]); + transpose_vecs_128(&out[8]); + transpose_vecs_128(&out[12]); +} + +INLINE void load_counters4(uint64_t counter, bool increment_counter, + uint32x4_t *out_low, uint32x4_t *out_high) { + uint64_t mask = (increment_counter ? ~0 : 0); + *out_low = set4( + counter_low(counter + (mask & 0)), counter_low(counter + (mask & 1)), + counter_low(counter + (mask & 2)), counter_low(counter + (mask & 3))); + *out_high = set4( + counter_high(counter + (mask & 0)), counter_high(counter + (mask & 1)), + counter_high(counter + (mask & 2)), counter_high(counter + (mask & 3))); +} + +static void blake3_hash4_neon(const uint8_t *const *inputs, size_t blocks, + const uint32_t key[8], uint64_t counter, + bool increment_counter, uint8_t flags, + uint8_t flags_start, uint8_t flags_end, + uint8_t *out) { + uint32x4_t h_vecs[8] = { + set1_128(key[0]), set1_128(key[1]), set1_128(key[2]), set1_128(key[3]), + set1_128(key[4]), set1_128(key[5]), set1_128(key[6]), set1_128(key[7]), + }; + uint32x4_t counter_low_vec, counter_high_vec; + load_counters4(counter, increment_counter, &counter_low_vec, + &counter_high_vec); + uint8_t block_flags = flags | flags_start; + + for (size_t block = 0; block < blocks; block++) { + if (block + 1 == blocks) { + block_flags |= flags_end; + } + uint32x4_t block_len_vec = set1_128(BLAKE3_BLOCK_LEN); + uint32x4_t block_flags_vec = set1_128(block_flags); + uint32x4_t msg_vecs[16]; + transpose_msg_vecs4(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs); + + uint32x4_t v[16] = { + h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3], + h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7], + set1_128(IV[0]), set1_128(IV[1]), set1_128(IV[2]), set1_128(IV[3]), + counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec, + }; + round_fn4(v, msg_vecs, 0); + round_fn4(v, msg_vecs, 1); + round_fn4(v, msg_vecs, 2); + round_fn4(v, msg_vecs, 3); + round_fn4(v, msg_vecs, 4); + round_fn4(v, msg_vecs, 5); + round_fn4(v, msg_vecs, 6); + h_vecs[0] = xor_128(v[0], v[8]); + h_vecs[1] = xor_128(v[1], v[9]); + h_vecs[2] = xor_128(v[2], v[10]); + h_vecs[3] = xor_128(v[3], v[11]); + h_vecs[4] = xor_128(v[4], v[12]); + h_vecs[5] = xor_128(v[5], v[13]); + h_vecs[6] = xor_128(v[6], v[14]); + h_vecs[7] = xor_128(v[7], v[15]); + + block_flags = flags; + } + + transpose_vecs_128(&h_vecs[0]); + transpose_vecs_128(&h_vecs[4]); + // The first four vecs now contain the first half of each output, and the + // second four vecs contain the second half of each output. + storeu_128(h_vecs[0], &out[0 * sizeof(uint32x4_t)]); + storeu_128(h_vecs[4], &out[1 * sizeof(uint32x4_t)]); + storeu_128(h_vecs[1], &out[2 * sizeof(uint32x4_t)]); + storeu_128(h_vecs[5], &out[3 * sizeof(uint32x4_t)]); + storeu_128(h_vecs[2], &out[4 * sizeof(uint32x4_t)]); + storeu_128(h_vecs[6], &out[5 * sizeof(uint32x4_t)]); + storeu_128(h_vecs[3], &out[6 * sizeof(uint32x4_t)]); + storeu_128(h_vecs[7], &out[7 * sizeof(uint32x4_t)]); +} + +/* + * ---------------------------------------------------------------------------- + * hash_many_neon + * ---------------------------------------------------------------------------- + */ + +void blake3_compress_in_place_portable(uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags); + +INLINE void hash_one_neon(const uint8_t *input, size_t blocks, + const uint32_t key[8], uint64_t counter, + uint8_t flags, uint8_t flags_start, uint8_t flags_end, + uint8_t out[BLAKE3_OUT_LEN]) { + uint32_t cv[8]; + memcpy(cv, key, BLAKE3_KEY_LEN); + uint8_t block_flags = flags | flags_start; + while (blocks > 0) { + if (blocks == 1) { + block_flags |= flags_end; + } + // TODO: Implement compress_neon. However note that according to + // https://github.com/BLAKE2/BLAKE2/commit/7965d3e6e1b4193438b8d3a656787587d2579227, + // compress_neon might not be any faster than compress_portable. + blake3_compress_in_place_portable(cv, input, BLAKE3_BLOCK_LEN, counter, + block_flags); + input = &input[BLAKE3_BLOCK_LEN]; + blocks -= 1; + block_flags = flags; + } + memcpy(out, cv, BLAKE3_OUT_LEN); +} + +void blake3_hash_many_neon(const uint8_t *const *inputs, size_t num_inputs, + size_t blocks, const uint32_t key[8], + uint64_t counter, bool increment_counter, + uint8_t flags, uint8_t flags_start, + uint8_t flags_end, uint8_t *out) { + while (num_inputs >= 4) { + blake3_hash4_neon(inputs, blocks, key, counter, increment_counter, flags, + flags_start, flags_end, out); + if (increment_counter) { + counter += 4; + } + inputs += 4; + num_inputs -= 4; + out = &out[4 * BLAKE3_OUT_LEN]; + } + while (num_inputs > 0) { + hash_one_neon(inputs[0], blocks, key, counter, flags, flags_start, + flags_end, out); + if (increment_counter) { + counter += 1; + } + inputs += 1; + num_inputs -= 1; + out = &out[BLAKE3_OUT_LEN]; + } +} diff --git a/ext/digest/blake3/blake3_sse2.c b/ext/digest/blake3/blake3_sse2.c new file mode 100644 index 0000000..691e1c6 --- /dev/null +++ b/ext/digest/blake3/blake3_sse2.c @@ -0,0 +1,566 @@ +#include "blake3_impl.h" + +#include + +#define DEGREE 4 + +#define _mm_shuffle_ps2(a, b, c) \ + (_mm_castps_si128( \ + _mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), (c)))) + +INLINE __m128i loadu(const uint8_t src[16]) { + return _mm_loadu_si128((const __m128i *)src); +} + +INLINE void storeu(__m128i src, uint8_t dest[16]) { + _mm_storeu_si128((__m128i *)dest, src); +} + +INLINE __m128i addv(__m128i a, __m128i b) { return _mm_add_epi32(a, b); } + +// Note that clang-format doesn't like the name "xor" for some reason. +INLINE __m128i xorv(__m128i a, __m128i b) { return _mm_xor_si128(a, b); } + +INLINE __m128i set1(uint32_t x) { return _mm_set1_epi32((int32_t)x); } + +INLINE __m128i set4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) { + return _mm_setr_epi32((int32_t)a, (int32_t)b, (int32_t)c, (int32_t)d); +} + +INLINE __m128i rot16(__m128i x) { + return _mm_shufflehi_epi16(_mm_shufflelo_epi16(x, 0xB1), 0xB1); +} + +INLINE __m128i rot12(__m128i x) { + return xorv(_mm_srli_epi32(x, 12), _mm_slli_epi32(x, 32 - 12)); +} + +INLINE __m128i rot8(__m128i x) { + return xorv(_mm_srli_epi32(x, 8), _mm_slli_epi32(x, 32 - 8)); +} + +INLINE __m128i rot7(__m128i x) { + return xorv(_mm_srli_epi32(x, 7), _mm_slli_epi32(x, 32 - 7)); +} + +INLINE void g1(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3, + __m128i m) { + *row0 = addv(addv(*row0, m), *row1); + *row3 = xorv(*row3, *row0); + *row3 = rot16(*row3); + *row2 = addv(*row2, *row3); + *row1 = xorv(*row1, *row2); + *row1 = rot12(*row1); +} + +INLINE void g2(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3, + __m128i m) { + *row0 = addv(addv(*row0, m), *row1); + *row3 = xorv(*row3, *row0); + *row3 = rot8(*row3); + *row2 = addv(*row2, *row3); + *row1 = xorv(*row1, *row2); + *row1 = rot7(*row1); +} + +// Note the optimization here of leaving row1 as the unrotated row, rather than +// row0. All the message loads below are adjusted to compensate for this. See +// discussion at https://github.com/sneves/blake2-avx2/pull/4 +INLINE void diagonalize(__m128i *row0, __m128i *row2, __m128i *row3) { + *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(2, 1, 0, 3)); + *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2)); + *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(0, 3, 2, 1)); +} + +INLINE void undiagonalize(__m128i *row0, __m128i *row2, __m128i *row3) { + *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(0, 3, 2, 1)); + *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2)); + *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(2, 1, 0, 3)); +} + +INLINE __m128i blend_epi16(__m128i a, __m128i b, const int16_t imm8) { + const __m128i bits = _mm_set_epi16(0x80, 0x40, 0x20, 0x10, 0x08, 0x04, 0x02, 0x01); + __m128i mask = _mm_set1_epi16(imm8); + mask = _mm_and_si128(mask, bits); + mask = _mm_cmpeq_epi16(mask, bits); + return _mm_or_si128(_mm_and_si128(mask, b), _mm_andnot_si128(mask, a)); +} + +INLINE void compress_pre(__m128i rows[4], const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, uint8_t flags) { + rows[0] = loadu((uint8_t *)&cv[0]); + rows[1] = loadu((uint8_t *)&cv[4]); + rows[2] = set4(IV[0], IV[1], IV[2], IV[3]); + rows[3] = set4(counter_low(counter), counter_high(counter), + (uint32_t)block_len, (uint32_t)flags); + + __m128i m0 = loadu(&block[sizeof(__m128i) * 0]); + __m128i m1 = loadu(&block[sizeof(__m128i) * 1]); + __m128i m2 = loadu(&block[sizeof(__m128i) * 2]); + __m128i m3 = loadu(&block[sizeof(__m128i) * 3]); + + __m128i t0, t1, t2, t3, tt; + + // Round 1. The first round permutes the message words from the original + // input order, into the groups that get mixed in parallel. + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(2, 0, 2, 0)); // 6 4 2 0 + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 3, 1)); // 7 5 3 1 + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(2, 0, 2, 0)); // 14 12 10 8 + t2 = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2, 1, 0, 3)); // 12 10 8 14 + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 1, 3, 1)); // 15 13 11 9 + t3 = _mm_shuffle_epi32(t3, _MM_SHUFFLE(2, 1, 0, 3)); // 13 11 9 15 + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 2. This round and all following rounds apply a fixed permutation + // to the message words from the round before. + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 3 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 4 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 5 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 6 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 7 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); +} + +void blake3_compress_in_place_sse2(uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags) { + __m128i rows[4]; + compress_pre(rows, cv, block, block_len, counter, flags); + storeu(xorv(rows[0], rows[2]), (uint8_t *)&cv[0]); + storeu(xorv(rows[1], rows[3]), (uint8_t *)&cv[4]); +} + +void blake3_compress_xof_sse2(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags, uint8_t out[64]) { + __m128i rows[4]; + compress_pre(rows, cv, block, block_len, counter, flags); + storeu(xorv(rows[0], rows[2]), &out[0]); + storeu(xorv(rows[1], rows[3]), &out[16]); + storeu(xorv(rows[2], loadu((uint8_t *)&cv[0])), &out[32]); + storeu(xorv(rows[3], loadu((uint8_t *)&cv[4])), &out[48]); +} + +INLINE void round_fn(__m128i v[16], __m128i m[16], size_t r) { + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][0]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][2]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][4]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][6]]); + v[0] = addv(v[0], v[4]); + v[1] = addv(v[1], v[5]); + v[2] = addv(v[2], v[6]); + v[3] = addv(v[3], v[7]); + v[12] = xorv(v[12], v[0]); + v[13] = xorv(v[13], v[1]); + v[14] = xorv(v[14], v[2]); + v[15] = xorv(v[15], v[3]); + v[12] = rot16(v[12]); + v[13] = rot16(v[13]); + v[14] = rot16(v[14]); + v[15] = rot16(v[15]); + v[8] = addv(v[8], v[12]); + v[9] = addv(v[9], v[13]); + v[10] = addv(v[10], v[14]); + v[11] = addv(v[11], v[15]); + v[4] = xorv(v[4], v[8]); + v[5] = xorv(v[5], v[9]); + v[6] = xorv(v[6], v[10]); + v[7] = xorv(v[7], v[11]); + v[4] = rot12(v[4]); + v[5] = rot12(v[5]); + v[6] = rot12(v[6]); + v[7] = rot12(v[7]); + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][1]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][3]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][5]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][7]]); + v[0] = addv(v[0], v[4]); + v[1] = addv(v[1], v[5]); + v[2] = addv(v[2], v[6]); + v[3] = addv(v[3], v[7]); + v[12] = xorv(v[12], v[0]); + v[13] = xorv(v[13], v[1]); + v[14] = xorv(v[14], v[2]); + v[15] = xorv(v[15], v[3]); + v[12] = rot8(v[12]); + v[13] = rot8(v[13]); + v[14] = rot8(v[14]); + v[15] = rot8(v[15]); + v[8] = addv(v[8], v[12]); + v[9] = addv(v[9], v[13]); + v[10] = addv(v[10], v[14]); + v[11] = addv(v[11], v[15]); + v[4] = xorv(v[4], v[8]); + v[5] = xorv(v[5], v[9]); + v[6] = xorv(v[6], v[10]); + v[7] = xorv(v[7], v[11]); + v[4] = rot7(v[4]); + v[5] = rot7(v[5]); + v[6] = rot7(v[6]); + v[7] = rot7(v[7]); + + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][8]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][10]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][12]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][14]]); + v[0] = addv(v[0], v[5]); + v[1] = addv(v[1], v[6]); + v[2] = addv(v[2], v[7]); + v[3] = addv(v[3], v[4]); + v[15] = xorv(v[15], v[0]); + v[12] = xorv(v[12], v[1]); + v[13] = xorv(v[13], v[2]); + v[14] = xorv(v[14], v[3]); + v[15] = rot16(v[15]); + v[12] = rot16(v[12]); + v[13] = rot16(v[13]); + v[14] = rot16(v[14]); + v[10] = addv(v[10], v[15]); + v[11] = addv(v[11], v[12]); + v[8] = addv(v[8], v[13]); + v[9] = addv(v[9], v[14]); + v[5] = xorv(v[5], v[10]); + v[6] = xorv(v[6], v[11]); + v[7] = xorv(v[7], v[8]); + v[4] = xorv(v[4], v[9]); + v[5] = rot12(v[5]); + v[6] = rot12(v[6]); + v[7] = rot12(v[7]); + v[4] = rot12(v[4]); + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][9]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][11]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][13]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][15]]); + v[0] = addv(v[0], v[5]); + v[1] = addv(v[1], v[6]); + v[2] = addv(v[2], v[7]); + v[3] = addv(v[3], v[4]); + v[15] = xorv(v[15], v[0]); + v[12] = xorv(v[12], v[1]); + v[13] = xorv(v[13], v[2]); + v[14] = xorv(v[14], v[3]); + v[15] = rot8(v[15]); + v[12] = rot8(v[12]); + v[13] = rot8(v[13]); + v[14] = rot8(v[14]); + v[10] = addv(v[10], v[15]); + v[11] = addv(v[11], v[12]); + v[8] = addv(v[8], v[13]); + v[9] = addv(v[9], v[14]); + v[5] = xorv(v[5], v[10]); + v[6] = xorv(v[6], v[11]); + v[7] = xorv(v[7], v[8]); + v[4] = xorv(v[4], v[9]); + v[5] = rot7(v[5]); + v[6] = rot7(v[6]); + v[7] = rot7(v[7]); + v[4] = rot7(v[4]); +} + +INLINE void transpose_vecs(__m128i vecs[DEGREE]) { + // Interleave 32-bit lanes. The low unpack is lanes 00/11 and the high is + // 22/33. Note that this doesn't split the vector into two lanes, as the + // AVX2 counterparts do. + __m128i ab_01 = _mm_unpacklo_epi32(vecs[0], vecs[1]); + __m128i ab_23 = _mm_unpackhi_epi32(vecs[0], vecs[1]); + __m128i cd_01 = _mm_unpacklo_epi32(vecs[2], vecs[3]); + __m128i cd_23 = _mm_unpackhi_epi32(vecs[2], vecs[3]); + + // Interleave 64-bit lanes. + __m128i abcd_0 = _mm_unpacklo_epi64(ab_01, cd_01); + __m128i abcd_1 = _mm_unpackhi_epi64(ab_01, cd_01); + __m128i abcd_2 = _mm_unpacklo_epi64(ab_23, cd_23); + __m128i abcd_3 = _mm_unpackhi_epi64(ab_23, cd_23); + + vecs[0] = abcd_0; + vecs[1] = abcd_1; + vecs[2] = abcd_2; + vecs[3] = abcd_3; +} + +INLINE void transpose_msg_vecs(const uint8_t *const *inputs, + size_t block_offset, __m128i out[16]) { + out[0] = loadu(&inputs[0][block_offset + 0 * sizeof(__m128i)]); + out[1] = loadu(&inputs[1][block_offset + 0 * sizeof(__m128i)]); + out[2] = loadu(&inputs[2][block_offset + 0 * sizeof(__m128i)]); + out[3] = loadu(&inputs[3][block_offset + 0 * sizeof(__m128i)]); + out[4] = loadu(&inputs[0][block_offset + 1 * sizeof(__m128i)]); + out[5] = loadu(&inputs[1][block_offset + 1 * sizeof(__m128i)]); + out[6] = loadu(&inputs[2][block_offset + 1 * sizeof(__m128i)]); + out[7] = loadu(&inputs[3][block_offset + 1 * sizeof(__m128i)]); + out[8] = loadu(&inputs[0][block_offset + 2 * sizeof(__m128i)]); + out[9] = loadu(&inputs[1][block_offset + 2 * sizeof(__m128i)]); + out[10] = loadu(&inputs[2][block_offset + 2 * sizeof(__m128i)]); + out[11] = loadu(&inputs[3][block_offset + 2 * sizeof(__m128i)]); + out[12] = loadu(&inputs[0][block_offset + 3 * sizeof(__m128i)]); + out[13] = loadu(&inputs[1][block_offset + 3 * sizeof(__m128i)]); + out[14] = loadu(&inputs[2][block_offset + 3 * sizeof(__m128i)]); + out[15] = loadu(&inputs[3][block_offset + 3 * sizeof(__m128i)]); + for (size_t i = 0; i < 4; ++i) { + _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0); + } + transpose_vecs(&out[0]); + transpose_vecs(&out[4]); + transpose_vecs(&out[8]); + transpose_vecs(&out[12]); +} + +INLINE void load_counters(uint64_t counter, bool increment_counter, + __m128i *out_lo, __m128i *out_hi) { + const __m128i mask = _mm_set1_epi32(-(int32_t)increment_counter); + const __m128i add0 = _mm_set_epi32(3, 2, 1, 0); + const __m128i add1 = _mm_and_si128(mask, add0); + __m128i l = _mm_add_epi32(_mm_set1_epi32((int32_t)counter), add1); + __m128i carry = _mm_cmpgt_epi32(_mm_xor_si128(add1, _mm_set1_epi32(0x80000000)), + _mm_xor_si128( l, _mm_set1_epi32(0x80000000))); + __m128i h = _mm_sub_epi32(_mm_set1_epi32((int32_t)(counter >> 32)), carry); + *out_lo = l; + *out_hi = h; +} + +static +void blake3_hash4_sse2(const uint8_t *const *inputs, size_t blocks, + const uint32_t key[8], uint64_t counter, + bool increment_counter, uint8_t flags, + uint8_t flags_start, uint8_t flags_end, uint8_t *out) { + __m128i h_vecs[8] = { + set1(key[0]), set1(key[1]), set1(key[2]), set1(key[3]), + set1(key[4]), set1(key[5]), set1(key[6]), set1(key[7]), + }; + __m128i counter_low_vec, counter_high_vec; + load_counters(counter, increment_counter, &counter_low_vec, + &counter_high_vec); + uint8_t block_flags = flags | flags_start; + + for (size_t block = 0; block < blocks; block++) { + if (block + 1 == blocks) { + block_flags |= flags_end; + } + __m128i block_len_vec = set1(BLAKE3_BLOCK_LEN); + __m128i block_flags_vec = set1(block_flags); + __m128i msg_vecs[16]; + transpose_msg_vecs(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs); + + __m128i v[16] = { + h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3], + h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7], + set1(IV[0]), set1(IV[1]), set1(IV[2]), set1(IV[3]), + counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec, + }; + round_fn(v, msg_vecs, 0); + round_fn(v, msg_vecs, 1); + round_fn(v, msg_vecs, 2); + round_fn(v, msg_vecs, 3); + round_fn(v, msg_vecs, 4); + round_fn(v, msg_vecs, 5); + round_fn(v, msg_vecs, 6); + h_vecs[0] = xorv(v[0], v[8]); + h_vecs[1] = xorv(v[1], v[9]); + h_vecs[2] = xorv(v[2], v[10]); + h_vecs[3] = xorv(v[3], v[11]); + h_vecs[4] = xorv(v[4], v[12]); + h_vecs[5] = xorv(v[5], v[13]); + h_vecs[6] = xorv(v[6], v[14]); + h_vecs[7] = xorv(v[7], v[15]); + + block_flags = flags; + } + + transpose_vecs(&h_vecs[0]); + transpose_vecs(&h_vecs[4]); + // The first four vecs now contain the first half of each output, and the + // second four vecs contain the second half of each output. + storeu(h_vecs[0], &out[0 * sizeof(__m128i)]); + storeu(h_vecs[4], &out[1 * sizeof(__m128i)]); + storeu(h_vecs[1], &out[2 * sizeof(__m128i)]); + storeu(h_vecs[5], &out[3 * sizeof(__m128i)]); + storeu(h_vecs[2], &out[4 * sizeof(__m128i)]); + storeu(h_vecs[6], &out[5 * sizeof(__m128i)]); + storeu(h_vecs[3], &out[6 * sizeof(__m128i)]); + storeu(h_vecs[7], &out[7 * sizeof(__m128i)]); +} + +INLINE void hash_one_sse2(const uint8_t *input, size_t blocks, + const uint32_t key[8], uint64_t counter, + uint8_t flags, uint8_t flags_start, + uint8_t flags_end, uint8_t out[BLAKE3_OUT_LEN]) { + uint32_t cv[8]; + memcpy(cv, key, BLAKE3_KEY_LEN); + uint8_t block_flags = flags | flags_start; + while (blocks > 0) { + if (blocks == 1) { + block_flags |= flags_end; + } + blake3_compress_in_place_sse2(cv, input, BLAKE3_BLOCK_LEN, counter, + block_flags); + input = &input[BLAKE3_BLOCK_LEN]; + blocks -= 1; + block_flags = flags; + } + memcpy(out, cv, BLAKE3_OUT_LEN); +} + +void blake3_hash_many_sse2(const uint8_t *const *inputs, size_t num_inputs, + size_t blocks, const uint32_t key[8], + uint64_t counter, bool increment_counter, + uint8_t flags, uint8_t flags_start, + uint8_t flags_end, uint8_t *out) { + while (num_inputs >= DEGREE) { + blake3_hash4_sse2(inputs, blocks, key, counter, increment_counter, flags, + flags_start, flags_end, out); + if (increment_counter) { + counter += DEGREE; + } + inputs += DEGREE; + num_inputs -= DEGREE; + out = &out[DEGREE * BLAKE3_OUT_LEN]; + } + while (num_inputs > 0) { + hash_one_sse2(inputs[0], blocks, key, counter, flags, flags_start, + flags_end, out); + if (increment_counter) { + counter += 1; + } + inputs += 1; + num_inputs -= 1; + out = &out[BLAKE3_OUT_LEN]; + } +} diff --git a/ext/digest/blake3/blake3_sse41.c b/ext/digest/blake3/blake3_sse41.c new file mode 100644 index 0000000..4653a85 --- /dev/null +++ b/ext/digest/blake3/blake3_sse41.c @@ -0,0 +1,560 @@ +#include "blake3_impl.h" + +#include + +#define DEGREE 4 + +#define _mm_shuffle_ps2(a, b, c) \ + (_mm_castps_si128( \ + _mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), (c)))) + +INLINE __m128i loadu(const uint8_t src[16]) { + return _mm_loadu_si128((const __m128i *)src); +} + +INLINE void storeu(__m128i src, uint8_t dest[16]) { + _mm_storeu_si128((__m128i *)dest, src); +} + +INLINE __m128i addv(__m128i a, __m128i b) { return _mm_add_epi32(a, b); } + +// Note that clang-format doesn't like the name "xor" for some reason. +INLINE __m128i xorv(__m128i a, __m128i b) { return _mm_xor_si128(a, b); } + +INLINE __m128i set1(uint32_t x) { return _mm_set1_epi32((int32_t)x); } + +INLINE __m128i set4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) { + return _mm_setr_epi32((int32_t)a, (int32_t)b, (int32_t)c, (int32_t)d); +} + +INLINE __m128i rot16(__m128i x) { + return _mm_shuffle_epi8( + x, _mm_set_epi8(13, 12, 15, 14, 9, 8, 11, 10, 5, 4, 7, 6, 1, 0, 3, 2)); +} + +INLINE __m128i rot12(__m128i x) { + return xorv(_mm_srli_epi32(x, 12), _mm_slli_epi32(x, 32 - 12)); +} + +INLINE __m128i rot8(__m128i x) { + return _mm_shuffle_epi8( + x, _mm_set_epi8(12, 15, 14, 13, 8, 11, 10, 9, 4, 7, 6, 5, 0, 3, 2, 1)); +} + +INLINE __m128i rot7(__m128i x) { + return xorv(_mm_srli_epi32(x, 7), _mm_slli_epi32(x, 32 - 7)); +} + +INLINE void g1(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3, + __m128i m) { + *row0 = addv(addv(*row0, m), *row1); + *row3 = xorv(*row3, *row0); + *row3 = rot16(*row3); + *row2 = addv(*row2, *row3); + *row1 = xorv(*row1, *row2); + *row1 = rot12(*row1); +} + +INLINE void g2(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3, + __m128i m) { + *row0 = addv(addv(*row0, m), *row1); + *row3 = xorv(*row3, *row0); + *row3 = rot8(*row3); + *row2 = addv(*row2, *row3); + *row1 = xorv(*row1, *row2); + *row1 = rot7(*row1); +} + +// Note the optimization here of leaving row1 as the unrotated row, rather than +// row0. All the message loads below are adjusted to compensate for this. See +// discussion at https://github.com/sneves/blake2-avx2/pull/4 +INLINE void diagonalize(__m128i *row0, __m128i *row2, __m128i *row3) { + *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(2, 1, 0, 3)); + *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2)); + *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(0, 3, 2, 1)); +} + +INLINE void undiagonalize(__m128i *row0, __m128i *row2, __m128i *row3) { + *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(0, 3, 2, 1)); + *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2)); + *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(2, 1, 0, 3)); +} + +INLINE void compress_pre(__m128i rows[4], const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, uint8_t flags) { + rows[0] = loadu((uint8_t *)&cv[0]); + rows[1] = loadu((uint8_t *)&cv[4]); + rows[2] = set4(IV[0], IV[1], IV[2], IV[3]); + rows[3] = set4(counter_low(counter), counter_high(counter), + (uint32_t)block_len, (uint32_t)flags); + + __m128i m0 = loadu(&block[sizeof(__m128i) * 0]); + __m128i m1 = loadu(&block[sizeof(__m128i) * 1]); + __m128i m2 = loadu(&block[sizeof(__m128i) * 2]); + __m128i m3 = loadu(&block[sizeof(__m128i) * 3]); + + __m128i t0, t1, t2, t3, tt; + + // Round 1. The first round permutes the message words from the original + // input order, into the groups that get mixed in parallel. + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(2, 0, 2, 0)); // 6 4 2 0 + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 3, 1)); // 7 5 3 1 + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(2, 0, 2, 0)); // 14 12 10 8 + t2 = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2, 1, 0, 3)); // 12 10 8 14 + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 1, 3, 1)); // 15 13 11 9 + t3 = _mm_shuffle_epi32(t3, _MM_SHUFFLE(2, 1, 0, 3)); // 13 11 9 15 + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 2. This round and all following rounds apply a fixed permutation + // to the message words from the round before. + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 3 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 4 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 5 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 6 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 7 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); +} + +void blake3_compress_in_place_sse41(uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags) { + __m128i rows[4]; + compress_pre(rows, cv, block, block_len, counter, flags); + storeu(xorv(rows[0], rows[2]), (uint8_t *)&cv[0]); + storeu(xorv(rows[1], rows[3]), (uint8_t *)&cv[4]); +} + +void blake3_compress_xof_sse41(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags, uint8_t out[64]) { + __m128i rows[4]; + compress_pre(rows, cv, block, block_len, counter, flags); + storeu(xorv(rows[0], rows[2]), &out[0]); + storeu(xorv(rows[1], rows[3]), &out[16]); + storeu(xorv(rows[2], loadu((uint8_t *)&cv[0])), &out[32]); + storeu(xorv(rows[3], loadu((uint8_t *)&cv[4])), &out[48]); +} + +INLINE void round_fn(__m128i v[16], __m128i m[16], size_t r) { + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][0]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][2]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][4]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][6]]); + v[0] = addv(v[0], v[4]); + v[1] = addv(v[1], v[5]); + v[2] = addv(v[2], v[6]); + v[3] = addv(v[3], v[7]); + v[12] = xorv(v[12], v[0]); + v[13] = xorv(v[13], v[1]); + v[14] = xorv(v[14], v[2]); + v[15] = xorv(v[15], v[3]); + v[12] = rot16(v[12]); + v[13] = rot16(v[13]); + v[14] = rot16(v[14]); + v[15] = rot16(v[15]); + v[8] = addv(v[8], v[12]); + v[9] = addv(v[9], v[13]); + v[10] = addv(v[10], v[14]); + v[11] = addv(v[11], v[15]); + v[4] = xorv(v[4], v[8]); + v[5] = xorv(v[5], v[9]); + v[6] = xorv(v[6], v[10]); + v[7] = xorv(v[7], v[11]); + v[4] = rot12(v[4]); + v[5] = rot12(v[5]); + v[6] = rot12(v[6]); + v[7] = rot12(v[7]); + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][1]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][3]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][5]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][7]]); + v[0] = addv(v[0], v[4]); + v[1] = addv(v[1], v[5]); + v[2] = addv(v[2], v[6]); + v[3] = addv(v[3], v[7]); + v[12] = xorv(v[12], v[0]); + v[13] = xorv(v[13], v[1]); + v[14] = xorv(v[14], v[2]); + v[15] = xorv(v[15], v[3]); + v[12] = rot8(v[12]); + v[13] = rot8(v[13]); + v[14] = rot8(v[14]); + v[15] = rot8(v[15]); + v[8] = addv(v[8], v[12]); + v[9] = addv(v[9], v[13]); + v[10] = addv(v[10], v[14]); + v[11] = addv(v[11], v[15]); + v[4] = xorv(v[4], v[8]); + v[5] = xorv(v[5], v[9]); + v[6] = xorv(v[6], v[10]); + v[7] = xorv(v[7], v[11]); + v[4] = rot7(v[4]); + v[5] = rot7(v[5]); + v[6] = rot7(v[6]); + v[7] = rot7(v[7]); + + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][8]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][10]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][12]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][14]]); + v[0] = addv(v[0], v[5]); + v[1] = addv(v[1], v[6]); + v[2] = addv(v[2], v[7]); + v[3] = addv(v[3], v[4]); + v[15] = xorv(v[15], v[0]); + v[12] = xorv(v[12], v[1]); + v[13] = xorv(v[13], v[2]); + v[14] = xorv(v[14], v[3]); + v[15] = rot16(v[15]); + v[12] = rot16(v[12]); + v[13] = rot16(v[13]); + v[14] = rot16(v[14]); + v[10] = addv(v[10], v[15]); + v[11] = addv(v[11], v[12]); + v[8] = addv(v[8], v[13]); + v[9] = addv(v[9], v[14]); + v[5] = xorv(v[5], v[10]); + v[6] = xorv(v[6], v[11]); + v[7] = xorv(v[7], v[8]); + v[4] = xorv(v[4], v[9]); + v[5] = rot12(v[5]); + v[6] = rot12(v[6]); + v[7] = rot12(v[7]); + v[4] = rot12(v[4]); + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][9]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][11]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][13]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][15]]); + v[0] = addv(v[0], v[5]); + v[1] = addv(v[1], v[6]); + v[2] = addv(v[2], v[7]); + v[3] = addv(v[3], v[4]); + v[15] = xorv(v[15], v[0]); + v[12] = xorv(v[12], v[1]); + v[13] = xorv(v[13], v[2]); + v[14] = xorv(v[14], v[3]); + v[15] = rot8(v[15]); + v[12] = rot8(v[12]); + v[13] = rot8(v[13]); + v[14] = rot8(v[14]); + v[10] = addv(v[10], v[15]); + v[11] = addv(v[11], v[12]); + v[8] = addv(v[8], v[13]); + v[9] = addv(v[9], v[14]); + v[5] = xorv(v[5], v[10]); + v[6] = xorv(v[6], v[11]); + v[7] = xorv(v[7], v[8]); + v[4] = xorv(v[4], v[9]); + v[5] = rot7(v[5]); + v[6] = rot7(v[6]); + v[7] = rot7(v[7]); + v[4] = rot7(v[4]); +} + +INLINE void transpose_vecs(__m128i vecs[DEGREE]) { + // Interleave 32-bit lanes. The low unpack is lanes 00/11 and the high is + // 22/33. Note that this doesn't split the vector into two lanes, as the + // AVX2 counterparts do. + __m128i ab_01 = _mm_unpacklo_epi32(vecs[0], vecs[1]); + __m128i ab_23 = _mm_unpackhi_epi32(vecs[0], vecs[1]); + __m128i cd_01 = _mm_unpacklo_epi32(vecs[2], vecs[3]); + __m128i cd_23 = _mm_unpackhi_epi32(vecs[2], vecs[3]); + + // Interleave 64-bit lanes. + __m128i abcd_0 = _mm_unpacklo_epi64(ab_01, cd_01); + __m128i abcd_1 = _mm_unpackhi_epi64(ab_01, cd_01); + __m128i abcd_2 = _mm_unpacklo_epi64(ab_23, cd_23); + __m128i abcd_3 = _mm_unpackhi_epi64(ab_23, cd_23); + + vecs[0] = abcd_0; + vecs[1] = abcd_1; + vecs[2] = abcd_2; + vecs[3] = abcd_3; +} + +INLINE void transpose_msg_vecs(const uint8_t *const *inputs, + size_t block_offset, __m128i out[16]) { + out[0] = loadu(&inputs[0][block_offset + 0 * sizeof(__m128i)]); + out[1] = loadu(&inputs[1][block_offset + 0 * sizeof(__m128i)]); + out[2] = loadu(&inputs[2][block_offset + 0 * sizeof(__m128i)]); + out[3] = loadu(&inputs[3][block_offset + 0 * sizeof(__m128i)]); + out[4] = loadu(&inputs[0][block_offset + 1 * sizeof(__m128i)]); + out[5] = loadu(&inputs[1][block_offset + 1 * sizeof(__m128i)]); + out[6] = loadu(&inputs[2][block_offset + 1 * sizeof(__m128i)]); + out[7] = loadu(&inputs[3][block_offset + 1 * sizeof(__m128i)]); + out[8] = loadu(&inputs[0][block_offset + 2 * sizeof(__m128i)]); + out[9] = loadu(&inputs[1][block_offset + 2 * sizeof(__m128i)]); + out[10] = loadu(&inputs[2][block_offset + 2 * sizeof(__m128i)]); + out[11] = loadu(&inputs[3][block_offset + 2 * sizeof(__m128i)]); + out[12] = loadu(&inputs[0][block_offset + 3 * sizeof(__m128i)]); + out[13] = loadu(&inputs[1][block_offset + 3 * sizeof(__m128i)]); + out[14] = loadu(&inputs[2][block_offset + 3 * sizeof(__m128i)]); + out[15] = loadu(&inputs[3][block_offset + 3 * sizeof(__m128i)]); + for (size_t i = 0; i < 4; ++i) { + _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0); + } + transpose_vecs(&out[0]); + transpose_vecs(&out[4]); + transpose_vecs(&out[8]); + transpose_vecs(&out[12]); +} + +INLINE void load_counters(uint64_t counter, bool increment_counter, + __m128i *out_lo, __m128i *out_hi) { + const __m128i mask = _mm_set1_epi32(-(int32_t)increment_counter); + const __m128i add0 = _mm_set_epi32(3, 2, 1, 0); + const __m128i add1 = _mm_and_si128(mask, add0); + __m128i l = _mm_add_epi32(_mm_set1_epi32((int32_t)counter), add1); + __m128i carry = _mm_cmpgt_epi32(_mm_xor_si128(add1, _mm_set1_epi32(0x80000000)), + _mm_xor_si128( l, _mm_set1_epi32(0x80000000))); + __m128i h = _mm_sub_epi32(_mm_set1_epi32((int32_t)(counter >> 32)), carry); + *out_lo = l; + *out_hi = h; +} + +static +void blake3_hash4_sse41(const uint8_t *const *inputs, size_t blocks, + const uint32_t key[8], uint64_t counter, + bool increment_counter, uint8_t flags, + uint8_t flags_start, uint8_t flags_end, uint8_t *out) { + __m128i h_vecs[8] = { + set1(key[0]), set1(key[1]), set1(key[2]), set1(key[3]), + set1(key[4]), set1(key[5]), set1(key[6]), set1(key[7]), + }; + __m128i counter_low_vec, counter_high_vec; + load_counters(counter, increment_counter, &counter_low_vec, + &counter_high_vec); + uint8_t block_flags = flags | flags_start; + + for (size_t block = 0; block < blocks; block++) { + if (block + 1 == blocks) { + block_flags |= flags_end; + } + __m128i block_len_vec = set1(BLAKE3_BLOCK_LEN); + __m128i block_flags_vec = set1(block_flags); + __m128i msg_vecs[16]; + transpose_msg_vecs(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs); + + __m128i v[16] = { + h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3], + h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7], + set1(IV[0]), set1(IV[1]), set1(IV[2]), set1(IV[3]), + counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec, + }; + round_fn(v, msg_vecs, 0); + round_fn(v, msg_vecs, 1); + round_fn(v, msg_vecs, 2); + round_fn(v, msg_vecs, 3); + round_fn(v, msg_vecs, 4); + round_fn(v, msg_vecs, 5); + round_fn(v, msg_vecs, 6); + h_vecs[0] = xorv(v[0], v[8]); + h_vecs[1] = xorv(v[1], v[9]); + h_vecs[2] = xorv(v[2], v[10]); + h_vecs[3] = xorv(v[3], v[11]); + h_vecs[4] = xorv(v[4], v[12]); + h_vecs[5] = xorv(v[5], v[13]); + h_vecs[6] = xorv(v[6], v[14]); + h_vecs[7] = xorv(v[7], v[15]); + + block_flags = flags; + } + + transpose_vecs(&h_vecs[0]); + transpose_vecs(&h_vecs[4]); + // The first four vecs now contain the first half of each output, and the + // second four vecs contain the second half of each output. + storeu(h_vecs[0], &out[0 * sizeof(__m128i)]); + storeu(h_vecs[4], &out[1 * sizeof(__m128i)]); + storeu(h_vecs[1], &out[2 * sizeof(__m128i)]); + storeu(h_vecs[5], &out[3 * sizeof(__m128i)]); + storeu(h_vecs[2], &out[4 * sizeof(__m128i)]); + storeu(h_vecs[6], &out[5 * sizeof(__m128i)]); + storeu(h_vecs[3], &out[6 * sizeof(__m128i)]); + storeu(h_vecs[7], &out[7 * sizeof(__m128i)]); +} + +INLINE void hash_one_sse41(const uint8_t *input, size_t blocks, + const uint32_t key[8], uint64_t counter, + uint8_t flags, uint8_t flags_start, + uint8_t flags_end, uint8_t out[BLAKE3_OUT_LEN]) { + uint32_t cv[8]; + memcpy(cv, key, BLAKE3_KEY_LEN); + uint8_t block_flags = flags | flags_start; + while (blocks > 0) { + if (blocks == 1) { + block_flags |= flags_end; + } + blake3_compress_in_place_sse41(cv, input, BLAKE3_BLOCK_LEN, counter, + block_flags); + input = &input[BLAKE3_BLOCK_LEN]; + blocks -= 1; + block_flags = flags; + } + memcpy(out, cv, BLAKE3_OUT_LEN); +} + +void blake3_hash_many_sse41(const uint8_t *const *inputs, size_t num_inputs, + size_t blocks, const uint32_t key[8], + uint64_t counter, bool increment_counter, + uint8_t flags, uint8_t flags_start, + uint8_t flags_end, uint8_t *out) { + while (num_inputs >= DEGREE) { + blake3_hash4_sse41(inputs, blocks, key, counter, increment_counter, flags, + flags_start, flags_end, out); + if (increment_counter) { + counter += DEGREE; + } + inputs += DEGREE; + num_inputs -= DEGREE; + out = &out[DEGREE * BLAKE3_OUT_LEN]; + } + while (num_inputs > 0) { + hash_one_sse41(inputs[0], blocks, key, counter, flags, flags_start, + flags_end, out); + if (increment_counter) { + counter += 1; + } + inputs += 1; + num_inputs -= 1; + out = &out[BLAKE3_OUT_LEN]; + } +} From d3180312b260c26c18c1b91f02268147dce3e6a6 Mon Sep 17 00:00:00 2001 From: Aaron Patterson Date: Tue, 7 Jul 2026 13:55:43 -0700 Subject: [PATCH 4/4] try this? --- ext/digest/blake3/extconf.rb | 19 +++++++++++++++++++ 1 file changed, 19 insertions(+) diff --git a/ext/digest/blake3/extconf.rb b/ext/digest/blake3/extconf.rb index 09548ec..df844d0 100644 --- a/ext/digest/blake3/extconf.rb +++ b/ext/digest/blake3/extconf.rb @@ -81,3 +81,22 @@ def blake3_have_isa?(name, flag, snippet) $preload = %w[digest] create_makefile("digest/blake3") + +# Emit one explicit compile rule per SIMD backend so each gets its own +# instruction-set flag. mkmf's implicit .c.o rule compiles every object with +# the same $(CFLAGS), which can't express e.g. -mavx2 for one file only; an +# explicit rule with a recipe overrides that implicit rule for these targets. +# The recipe mirrors mkmf's .c.o rule with the extra flag inserted after +# $(CFLAGS). +unless simd_cflags.empty? + File.open("Makefile", "a") do |mf| + mf.puts + mf.puts "# Per-file instruction-set flags for the BLAKE3 SIMD backends." + simd_cflags.each do |obj, flag| + target = "#{obj}.#{$OBJEXT}" + mf.puts "#{target}: $(srcdir)/#{obj}.c" + mf.puts "\t$(ECHO) compiling $(<)" + mf.puts "\t$(Q) $(CC) $(INCFLAGS) $(CPPFLAGS) $(CFLAGS) #{flag} $(COUTFLAG)$@ -c $(CSRCFLAG)$(srcdir)/#{obj}.c" + end + end +end