arnaucube/poulpy
1
0
Fork 0
mirror of https://github.com/arnaucube/poulpy.git synced 2026-02-10 05:06:44 +01:00
Code Issues Packages Projects Releases Wiki Activity

Update READMEs & add SECURITY.md (#78)

Browse Source
This commit is contained in:
Jean-Philippe Bossuat
2025-08-20 20:52:59 +02:00
committed by GitHub
parent 3b94ab047e
commit ccd94e36cc
256 changed files with 1982 additions and 31869 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
poulpy-backend/Cargo.toml
View File

@@ -10,7 +10,7 @@ homepage = "https://github.com/phantomzone-org/poulpy"
documentation = "https://docs.rs/poulpy"
[dependencies]
poulpy-hal = "0.1.2"
poulpy-hal = {path="../poulpy-hal"}
rug = {workspace = true}
criterion = {workspace = true}
itertools = {workspace = true}

45
poulpy-backend/README.md
View File

@@ -1,15 +1,38 @@
# 🐙 Poulpy-Backend
**Poulpy-Backend** is a Rust crate that provides concrete implementations of **`poulpy-hal`**. This crate is used to instantiate projects implemented with **`poulpy-hal`**, **`poulpy-core`** and/or **`poulpy-schemes`**.
## spqlios-arithmetic
## Backends
### WSL/Ubuntu
To use this crate you need to build spqlios-arithmetic, which is provided a as a git submodule:
1) Initialize the sub-module
2) $ cd backend/spqlios-arithmetic
3) mdkir build
4) cd build
5) cmake ..
6) make
### cpu-spqlios
### Others
Steps 3 to 6 might change depending of your platform. See [spqlios-arithmetic/wiki/build](https://github.com/tfhe/spqlios-arithmetic/wiki/build) for additional information and build options.
This module provides a CPU AVX2 accelerated backend through C bindings over [**spqlios-arithmetic**](https://github.com/tfhe/spqlios-arithmetic).
- Currently supported: `FFT64` backend
- Planned: `NTT120` backend
### Build Notes
This backend is built and compiled automatically and has been tested on wsl/ubuntu.
- `cmake` is invoked automatically by the build script (`build.rs`) when compiling the crate.
- No manual setup is required beyond having a standard Rust toolchain.
- Build options can be changed in `/build/cpu_spqlios.rs`
- Automatic build of cpu-spqlios/spqlios-arithmetic can be disabled in `build.rs`.
Spqlios-arithmetic is windows/mac compatible but building for those platforms is slightly different (see [spqlios-arithmetic/wiki/build](https://github.com/tfhe/spqlios-arithmetic/wiki/build)) and has not been tested in Poulpy.
### Example
```rust
use poulpy_backend::cpu_spqlios::FFT64;
use poulpy_hal::{api::ModuleNew, layouts::Module};
let log_n: usize = 10;
let module = Module<FFT64> = Module<FFT64>::new(1<<log_n);
```
## Contributors
To add a backend, implement the open extension traits from **`poulpy-hal/oep`** for a struct that implements the `Backend` trait.
This will automatically make your backend compatible with the API of **`poulpy-hal`**, **`poulpy-core`** and **`poulpy-schemes`**.

1
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic Submodule

Submodule poulpy-backend/src/cpu_spqlios/spqlios-arithmetic added at de62af3507

14
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/.clang-format
View File

@@ -1,14 +0,0 @@
# Use the Google style in this project.
BasedOnStyle: Google
# Some folks prefer to write "int& foo" while others prefer "int &foo". The
# Google Style Guide only asks for consistency within a project, we chose
# "int& foo" for this project:
DerivePointerAlignment: false
PointerAlignment: Left
# The Google Style Guide only asks for consistency w.r.t. "east const" vs.
# "const west" alignment of cv-qualifiers. In this project we use "east const".
QualifierAlignment: Left
ColumnLimit: 120

20
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/.github/workflows/auto-release.yml vendored
View File

@@ -1,20 +0,0 @@
name: Auto-Release
on:
workflow_dispatch:
push:
branches: [ "main" ]
jobs:
build:
name: Auto-Release
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v3
with:
fetch-depth: 3
# sparse-checkout: manifest.yaml scripts/auto-release.sh
- run:
${{github.workspace}}/scripts/auto-release.sh

6
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/.gitignore vendored
View File

@@ -1,6 +0,0 @@
cmake-build-*
.idea
build
.vscode
.*.sh

69
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/CMakeLists.txt
View File

@@ -1,69 +0,0 @@
cmake_minimum_required(VERSION 3.8)
project(spqlios)
# read the current version from the manifest file
file(READ "manifest.yaml" manifest)
string(REGEX MATCH "version: +(([0-9]+)\\.([0-9]+)\\.([0-9]+))" SPQLIOS_VERSION_BLAH ${manifest})
#message(STATUS "Version: ${SPQLIOS_VERSION_BLAH}")
set(SPQLIOS_VERSION ${CMAKE_MATCH_1})
set(SPQLIOS_VERSION_MAJOR ${CMAKE_MATCH_2})
set(SPQLIOS_VERSION_MINOR ${CMAKE_MATCH_3})
set(SPQLIOS_VERSION_PATCH ${CMAKE_MATCH_4})
message(STATUS "Compiling spqlios-fft version: ${SPQLIOS_VERSION_MAJOR}.${SPQLIOS_VERSION_MINOR}.${SPQLIOS_VERSION_PATCH}")
#set(ENABLE_SPQLIOS_F128 ON CACHE BOOL "Enable float128 via libquadmath")
set(WARNING_PARANOID ON CACHE BOOL "Treat all warnings as errors")
set(ENABLE_TESTING ON CACHE BOOL "Compiles unittests and integration tests")
set(DEVMODE_INSTALL OFF CACHE BOOL "Install private headers and testlib (mainly for CI)")
if (NOT CMAKE_BUILD_TYPE OR CMAKE_BUILD_TYPE STREQUAL "")
set(CMAKE_BUILD_TYPE "Release" CACHE STRING "Build type: Release or Debug" FORCE)
endif()
message(STATUS "Build type: ${CMAKE_BUILD_TYPE}")
if (WARNING_PARANOID)
add_compile_options(-Wall -Werror -Wno-unused-command-line-argument)
endif()
message(STATUS "CMAKE_HOST_SYSTEM_NAME: ${CMAKE_HOST_SYSTEM_NAME}")
message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}")
message(STATUS "CMAKE_SYSTEM_NAME: ${CMAKE_SYSTEM_NAME}")
if (CMAKE_SYSTEM_PROCESSOR MATCHES "(x86)|(X86)|(amd64)|(AMD64)")
set(X86 ON)
set(AARCH64 OFF)
else ()
set(X86 OFF)
# set(ENABLE_SPQLIOS_F128 OFF) # float128 are only supported for x86 targets
endif ()
if (CMAKE_SYSTEM_PROCESSOR MATCHES "(aarch64)|(arm64)")
set(AARCH64 ON)
endif ()
if (CMAKE_SYSTEM_NAME MATCHES "(Windows)|(MSYS)")
set(WIN32 ON)
endif ()
if (WIN32)
#overrides for win32
set(X86 OFF)
set(AARCH64 OFF)
set(X86_WIN32 ON)
else()
set(X86_WIN32 OFF)
set(WIN32 OFF)
endif (WIN32)
message(STATUS "--> WIN32: ${WIN32}")
message(STATUS "--> X86_WIN32: ${X86_WIN32}")
message(STATUS "--> X86_LINUX: ${X86}")
message(STATUS "--> AARCH64: ${AARCH64}")
# compiles the main library in spqlios
add_subdirectory(spqlios)
# compiles and activates unittests and itests
if (${ENABLE_TESTING})
enable_testing()
add_subdirectory(test)
endif()

77
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/CONTRIBUTING.md
View File

@@ -1,77 +0,0 @@
# Contributing to SPQlios-fft
The spqlios-fft team encourages contributions.
We encourage users to fix bugs, improve the documentation, write tests and to enhance the code, or ask for new features.
We encourage researchers to contribute with implementations of their FFT or NTT algorithms.
In the following we are trying to give some guidance on how to contribute effectively.
## Communication ##
Communication in the spqlios-fft project happens mainly on [GitHub](https://github.com/tfhe/spqlios-fft/issues).
All communications are public, so please make sure to maintain professional behaviour in
all published comments. See [Code of Conduct](https://www.contributor-covenant.org/version/2/1/code_of_conduct/) for
guidelines.
## Reporting Bugs or Requesting features ##
Bug should be filed at [https://github.com/tfhe/spqlios-fft/issues](https://github.com/tfhe/spqlios-fft/issues).
Features can also be requested there, in this case, please ensure that the features you request are self-contained,
easy to define, and generic enough to be used in different use-cases. Please provide an example of use-cases if
possible.
## Setting up topic branches and generating pull requests
This section applies to people that already have write access to the repository. Specific instructions for pull-requests
from public forks will be given later.
To implement some changes, please follow these steps:
- Create a "topic branch". Usually, the branch name should be `username/small-title`
or better `username/issuenumber-small-title` where `issuenumber` is the number of
the github issue number that is tackled.
- Push any needed commits to your branch. Make sure it compiles in `CMAKE_BUILD_TYPE=Debug` and `=Release`, with `-DWARNING_PARANOID=ON`.
- When the branch is nearly ready for review, please open a pull request, and add the label `check-on-arm`
- Do as many commits as necessary until all CI checks pass and all PR comments have been resolved.
> _During the process, you may optionnally use `git rebase -i` to clean up your commit history. If you elect to do so,
please at the very least make sure that nobody else is working or has forked from your branch: the conflicts it would generate
and the human hours to fix them are not worth it. `Git merge` remains the preferred option._
- Finally, when all reviews are positive and all CI checks pass, you may merge your branch via the github webpage.
### Keep your pull requests limited to a single issue
Pull requests should be as small/atomic as possible.
### Coding Conventions
* Please make sure that your code is formatted according to the `.clang-format` file and
that all files end with a newline character.
* Please make sure that all the functions declared in the public api have relevant doxygen comments.
Preferably, functions in the private apis should also contain a brief doxygen description.
### Versions and History
* **Stable API** The project uses semantic versioning on the functions that are listed as `stable` in the documentation. A version has
the form `x.y.z`
* a patch release that increments `z` does not modify the stable API.
* a minor release that increments `y` adds a new feature to the stable API.
* In the unlikely case where we need to change or remove a feature, we will trigger a major release that
increments `x`.
> _If any, we will mark those features as deprecated at least six months before the major release._
* **Experimental API** Features that are not part of the stable section in the documentation are experimental features: you may test them at
your own risk,
but keep in mind that semantic versioning does not apply to them.
> _If you have a use-case that uses an experimental feature, we encourage
> you to tell us about it, so that this feature reaches to the stable section faster!_
* **Version history** The current version is reported in `manifest.yaml`, any change of version comes up with a tag on the main branch, and the history between releases is summarized in `Changelog.md`. It is the main source of truth for anyone who wishes to
get insight about
the history of the repository (not the commit graph).
> Note: _The commit graph of git is for git's internal use only. Its main purpose is to reduce potential merge conflicts to a minimum, even in scenario where multiple features are developped in parallel: it may therefore be non-linear. If, as humans, we like to see a linear history, please read `Changelog.md` instead!_

18
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/Changelog.md
View File

@@ -1,18 +0,0 @@
# Changelog
All notable changes to this project will be documented in this file.
this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
## [2.0.0] - 2024-08-21
- Initial release of the `vec_znx` (except convolution products), `vec_rnx` and `zn` apis.
- Hardware acceleration available: AVX2 (most parts)
- APIs are documented in the wiki and are in "beta mode": during the 2.x -> 3.x transition, functions whose API is satisfactory in test projects will pass in "stable mode".
## [1.0.0] - 2023-07-18
- Initial release of the double precision fft on the reim and cplx backends
- Coeffs-space conversions cplx <-> znx32 and tnx32
- FFT-space conversions cplx <-> reim4 layouts
- FFT-space multiplications on the cplx, reim and reim4 layouts.
- In this first release, the only platform supported is linux x86_64 (generic C code, and avx2/fma). It compiles on arm64, but without any acceleration.

201
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/LICENSE
View File

@@ -1,201 +0,0 @@
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 [yyyy] [name of copyright owner]
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.

65
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/README.md
View File

@@ -1,65 +0,0 @@
# SPQlios library
The SPQlios library provides fast arithmetic for Fully Homomorphic Encryption, and other lattice constructions that arise in post quantum cryptography.
<img src="docs/api-full.svg">
Namely, it is divided into 4 sections:
* The low-level DFT section support FFT over 64-bit floats, as well as NTT modulo one fixed 120-bit modulus. It is an upgrade of the original spqlios-fft module embedded in the TFHE library since 2016. The DFT section exposes the traditional DFT, inverse-DFT, and coefficient-wise multiplications in DFT space.
* The VEC_ZNX section exposes fast algebra over vectors of small integer polynomial modulo $X^N+1$. It proposed in particular efficient (prepared) vector-matrix products, scalar-vector products, convolution products, and element-wise products, operations that naturally occurs on gadget-decomposed Ring-LWE coordinates.
* The RNX section is a simpler variant of VEC_ZNX, to represent single polynomials modulo $X^N+1$ (over the reals or over the torus) when the coefficient precision fits on 64-bit doubles. The small vector-matrix API of the RNX section is particularly adapted to reproducing the fastest CGGI-based bootstrappings.
* The ZN section focuses over vector and matrix algebra over scalars (used by scalar LWE, or scalar key-switches, but also on non-ring schemes like Frodo, FrodoPIR, and SimplePIR).
### A high value target for hardware accelerations
SPQlios is more than a library, it is also a good target for hardware developers.
On one hand, the arithmetic operations that are defined in the library have a clear standalone mathematical definition. And at the same time, the amount of work in each operations is sufficiently large so that meaningful functions only require a few of these.
This makes the SPQlios API a high value target for hardware acceleration, that targets FHE.
### SPQLios is not an FHE library, but a huge enabler
SPQlios itself is not an FHE library: there is no ciphertext, plaintext or key. It is a mathematical library that exposes efficient algebra over polynomials. Using the functions exposed, it is possible to quickly build efficient FHE libraries, with support for the main schemes based on Ring-LWE: BFV, BGV, CGGI, DM, CKKS.
## Dependencies
The SPQLIOS-FFT library is a C library that can be compiled with a standard C compiler, and depends only on libc and libm. The API
interface can be used in a regular C code, and any other language via classical foreign APIs.
The unittests and integration tests are in an optional part of the code, and are written in C++. These tests rely on
[```benchmark```](https://github.com/google/benchmark), and [```gtest```](https://github.com/google/googletest) libraries, and therefore require a C++17 compiler.
Currently, the project has been tested with the gcc,g++ >= 11.3.0 compiler under Linux (x86_64). In the future, we plan to
extend the compatibility to other compilers, platforms and operating systems.
## Installation
The library uses a classical ```cmake``` build mechanism: use ```cmake``` to create a ```build``` folder in the top level directory and run ```make``` from inside it. This assumes that the standard tool ```cmake``` is already installed on the system, and an up-to-date c++ compiler (i.e. g++ >=11.3.0) as well.
It will compile the shared library in optimized mode, and ```make install``` install it to the desired prefix folder (by default ```/usr/local/lib```).
If you want to choose additional compile options (i.e. other installation folder, debug mode, tests), you need to run cmake manually and pass the desired options:
```
mkdir build
cd build
cmake ../src -CMAKE_INSTALL_PREFIX=/usr/
make
```
The available options are the following:
| Variable Name | values |
| -------------------- | ------------------------------------------------------------ |
| CMAKE_INSTALL_PREFIX | */usr/local* installation folder (libs go in lib/ and headers in include/) |
| WARNING_PARANOID | All warnings are shown and treated as errors. Off by default |
| ENABLE_TESTING | Compiles unit tests and integration tests |
------
<img src="docs/logo-sandboxaq-black.svg">
<img src="docs/logo-inpher1.png">

416
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/docs/api-full.svg
View File

File diff suppressed because one or more lines are too long
Side by Side

Before

Width:  |  Height:  |  Size: 550 KiB

BIN
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/docs/logo-inpher1.png
View File

Binary file not shown.
Side by Side

Before

Width:  |  Height:  |  Size: 24 KiB

BIN
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/docs/logo-inpher2.png
View File

Binary file not shown.
Side by Side

Before

Width:  |  Height:  |  Size: 24 KiB

139
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/docs/logo-sandboxaq-black.svg
View File

@@ -1,139 +0,0 @@
<?xml version="1.0" encoding="UTF-8" standalone="no"?>
<!-- Generator: Adobe Illustrator 24.2.1, SVG Export Plug-In . SVG Version: 6.00 Build 0) -->
<svg
version="1.1"
id="Layer_1"
x="0px"
y="0px"
viewBox="0 0 270 49.4"
style="enable-background:new 0 0 270 49.4;"
xml:space="preserve"
sodipodi:docname="logo-sandboxaq-black.svg"
inkscape:version="1.3.2 (1:1.3.2+202311252150+091e20ef0f)"
xmlns:inkscape="http://www.inkscape.org/namespaces/inkscape"
xmlns:sodipodi="http://sodipodi.sourceforge.net/DTD/sodipodi-0.dtd"
xmlns="http://www.w3.org/2000/svg"
xmlns:svg="http://www.w3.org/2000/svg"><defs
id="defs9839">
</defs><sodipodi:namedview
id="namedview9837"
pagecolor="#ffffff"
bordercolor="#000000"
borderopacity="0.25"
inkscape:showpageshadow="2"
inkscape:pageopacity="0.0"
inkscape:pagecheckerboard="0"
inkscape:deskcolor="#d1d1d1"
showgrid="false"
inkscape:zoom="1.194332"
inkscape:cx="135.64068"
inkscape:cy="25.118645"
inkscape:window-width="804"
inkscape:window-height="436"
inkscape:window-x="190"
inkscape:window-y="27"
inkscape:window-maximized="0"
inkscape:current-layer="Layer_1" />
<style
type="text/css"
id="style9786">
.st0{fill:#EBB028;}
.st1{fill:#FFFFFF;}
</style>
<text
transform="matrix(1 0 0 1 393.832 -491.944)"
class="st1"
style="font-family:'Satoshi-Medium'; font-size:86.2078px;"
id="text9788">SANDBOX </text>
<text
transform="matrix(1 0 0 1 896.332 -491.944)"
class="st1"
style="font-family:'Satoshi-Black'; font-size:86.2078px;"
id="text9790">AQ</text>
<g
id="g9808">
<g
id="g9800">
<g
id="g9798">
<path
class="st0"
d="m 8.9,9.7 v 3.9 l 29.6,17.1 v 2.7 c 0,1.2 -0.6,2.3 -1.6,2.9 L 31,39.8 v -4 L 1.4,18.6 V 15.9 C 1.4,14.7 2,13.6 3.1,13 Z"
id="path9792" />
<path
class="st0"
d="M 18.3,45.1 3.1,36.3 C 2.1,35.7 1.4,34.6 1.4,33.4 V 26 L 28,41.4 21.5,45.1 c -0.9,0.6 -2.2,0.6 -3.2,0 z"
id="path9794" />
<path
class="st0"
d="m 21.6,4.3 15.2,8.8 c 1,0.6 1.7,1.7 1.7,2.9 v 7.5 L 11.8,8 18.3,4.3 c 1,-0.6 2.3,-0.6 3.3,0 z"
id="path9796" />
</g>
</g>
<g
id="g9806">
<polygon
class="st0"
points="248.1,23.2 248.1,30 251.4,33.8 257.3,33.8 "
id="polygon9802" />
<path
class="st0"
d="m 246.9,31 -0.1,-0.1 h -0.1 c -0.2,0 -0.4,0 -0.6,0 -3.5,0 -5.7,-2.6 -5.7,-6.7 0,-4.1 2.2,-6.7 5.7,-6.7 3.5,0 5.7,2.6 5.7,6.7 0,0.3 0,0.6 0,0.9 l 3.6,4.2 c 0.7,-1.5 1,-3.2 1,-5.1 0,-6.5 -4.2,-11 -10.3,-11 -6.1,0 -10.3,4.5 -10.3,11 0,6.5 4.2,11 10.3,11 1.2,0 2.3,-0.2 3.4,-0.5 l 0.5,-0.2 z"
id="path9804" />
</g>
</g><g
id="g9824"
style="fill:#1a1a1a">
<path
class="st1"
d="m 58.7,13.2 c 4.6,0 7.4,2.5 7.4,6.5 h -4.6 c 0,-1.5 -1.1,-2.4 -2.9,-2.4 -1.9,0 -3.1,0.9 -3.1,2.3 0,1.3 0.7,1.9 2.2,2.2 l 3.2,0.7 c 3.8,0.8 5.6,2.6 5.6,5.9 0,4.1 -3.2,6.8 -8.1,6.8 -4.7,0 -7.8,-2.6 -7.8,-6.5 h 4.6 c 0,1.6 1.1,2.4 3.2,2.4 2.1,0 3.4,-0.8 3.4,-2.2 0,-1.2 -0.5,-1.8 -2,-2.1 l -3.2,-0.7 c -3.8,-0.8 -5.7,-2.9 -5.7,-6.4 0,-3.7 3.2,-6.5 7.8,-6.5 z"
id="path9810"
style="fill:#1a1a1a" />
<path
class="st1"
d="M 70.4,34.9 78,13.6 h 4.5 l 7.6,21.3 h -4.9 l -1.5,-4.5 h -6.9 l -1.5,4.5 z m 7.7,-8.4 h 4.2 L 80.8,22 c -0.2,-0.7 -0.5,-1.6 -0.6,-2.1 -0.1,0.5 -0.3,1.3 -0.6,2.1 z"
id="path9812"
style="fill:#1a1a1a" />
<path
class="st1"
d="M 95.3,34.9 V 13.6 h 4.6 l 9,13.5 V 13.6 h 4.6 v 21.3 h -4.6 l -9,-13.5 v 13.5 z"
id="path9814"
style="fill:#1a1a1a" />
<path
class="st1"
d="M 120.7,34.9 V 13.6 h 8 c 6.2,0 10.6,4.4 10.6,10.7 0,6.2 -4.2,10.6 -10.3,10.6 z m 4.7,-17 v 12.6 h 3.2 c 3.7,0 5.8,-2.3 5.8,-6.3 0,-4 -2.3,-6.4 -6.1,-6.4 h -2.9 z"
id="path9816"
style="fill:#1a1a1a" />
<path
class="st1"
d="m 145.4,13.6 h 8.8 c 4.3,0 6.9,2.2 6.9,5.9 0,2.3 -1,3.9 -3,4.8 2.1,0.7 3.2,2.3 3.2,4.7 0,3.8 -2.5,5.9 -7.1,5.9 h -8.8 z m 4.7,4.1 v 4.6 h 3.7 c 1.7,0 2.6,-0.8 2.6,-2.4 0,-1.5 -0.9,-2.3 -2.6,-2.3 h -3.7 z m 0,8.5 v 4.6 h 3.9 c 1.7,0 2.6,-0.8 2.6,-2.4 0,-1.4 -0.9,-2.2 -2.6,-2.2 z"
id="path9818"
style="fill:#1a1a1a" />
<path
class="st1"
d="m 176.5,35.2 c -6.1,0 -10.4,-4.5 -10.4,-11 0,-6.5 4.3,-11 10.4,-11 6.2,0 10.4,4.5 10.4,11 0,6.5 -4.2,11 -10.4,11 z m 0.1,-17.5 c -3.4,0 -5.5,2.4 -5.5,6.5 0,4.1 2.1,6.5 5.5,6.5 3.4,0 5.5,-2.5 5.5,-6.5 0,-4 -2.1,-6.5 -5.5,-6.5 z"
id="path9820"
style="fill:#1a1a1a" />
<path
class="st1"
d="m 190.4,13.6 h 5.5 l 1.8,2.8 c 0.8,1.2 1.5,2.5 2.5,4.3 l 4.3,-7 h 5.4 l -6.7,10.6 6.7,10.6 h -5.5 L 203,32.7 c -1.1,-1.7 -1.8,-3 -2.8,-4.9 l -4.6,7.1 h -5.5 l 7.1,-10.6 z"
id="path9822"
style="fill:#1a1a1a" />
</g><path
class="st0"
d="m 229,34.9 h 4.7 L 226,13.6 h -4.3 L 214,34.8 h 4.6 l 1.6,-4.5 h 7.1 z m -5.1,-14.6 c 0,0 0,0 0,0 0,-0.1 0,-0.1 0,0 l 2.2,6.2 h -4.4 z"
id="path9826" /><g
id="g9832">
<path
class="st1"
d="m 259.5,11.2 h 3.9 v 1 h -1.3 v 3.1 h -1.3 v -3.1 h -1.3 z m 4.5,0 h 1.7 l 0.6,2.5 0.6,-2.5 h 1.7 v 4.1 h -1 v -3.1 l -0.8,3.1 h -0.9 l -0.8,-3.1 v 3.1 h -1 v -4.1 z"
id="path9830" />
</g>
</svg>
Side by Side

Before

Width:  |  Height:  |  Size: 5.0 KiB

133
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/docs/logo-sandboxaq-white.svg
View File

@@ -1,133 +0,0 @@
<?xml version="1.0" encoding="UTF-8" standalone="no"?>
<!-- Generator: Adobe Illustrator 24.2.1, SVG Export Plug-In . SVG Version: 6.00 Build 0) -->
<svg
version="1.1"
id="Layer_1"
x="0px"
y="0px"
viewBox="0 0 270 49.4"
style="enable-background:new 0 0 270 49.4;"
xml:space="preserve"
sodipodi:docname="logo-sandboxaq-white.svg"
inkscape:version="1.2.2 (1:1.2.2+202212051551+b0a8486541)"
xmlns:inkscape="http://www.inkscape.org/namespaces/inkscape"
xmlns:sodipodi="http://sodipodi.sourceforge.net/DTD/sodipodi-0.dtd"
xmlns="http://www.w3.org/2000/svg"
xmlns:svg="http://www.w3.org/2000/svg"><defs
id="defs9839" /><sodipodi:namedview
id="namedview9837"
pagecolor="#ffffff"
bordercolor="#000000"
borderopacity="0.25"
inkscape:showpageshadow="2"
inkscape:pageopacity="0.0"
inkscape:pagecheckerboard="0"
inkscape:deskcolor="#d1d1d1"
showgrid="false"
inkscape:zoom="2.3886639"
inkscape:cx="135.22204"
inkscape:cy="25.327967"
inkscape:window-width="1072"
inkscape:window-height="688"
inkscape:window-x="0"
inkscape:window-y="0"
inkscape:window-maximized="1"
inkscape:current-layer="Layer_1" />
<style
type="text/css"
id="style9786">
.st0{fill:#EBB028;}
.st1{fill:#FFFFFF;}
</style>
<text
transform="matrix(1 0 0 1 393.832 -491.944)"
class="st1"
style="font-family:'Satoshi-Medium'; font-size:86.2078px;"
id="text9788">SANDBOX </text>
<text
transform="matrix(1 0 0 1 896.332 -491.944)"
class="st1"
style="font-family:'Satoshi-Black'; font-size:86.2078px;"
id="text9790">AQ</text>
<g
id="g9834">
<g
id="g9828">
<g
id="g9808">
<g
id="g9800">
<g
id="g9798">
<path
class="st0"
d="M8.9,9.7v3.9l29.6,17.1v2.7c0,1.2-0.6,2.3-1.6,2.9L31,39.8v-4L1.4,18.6v-2.7c0-1.2,0.6-2.3,1.7-2.9 L8.9,9.7z"
id="path9792" />
<path
class="st0"
d="M18.3,45.1L3.1,36.3c-1-0.6-1.7-1.7-1.7-2.9V26L28,41.4l-6.5,3.7C20.6,45.7,19.3,45.7,18.3,45.1z"
id="path9794" />
<path
class="st0"
d="M21.6,4.3l15.2,8.8c1,0.6,1.7,1.7,1.7,2.9v7.5L11.8,8l6.5-3.7C19.3,3.7,20.6,3.7,21.6,4.3z"
id="path9796" />
</g>
</g>
<g
id="g9806">
<polygon
class="st0"
points="248.1,23.2 248.1,30 251.4,33.8 257.3,33.8 "
id="polygon9802" />
<path
class="st0"
d="M246.9,31l-0.1-0.1l-0.1,0c-0.2,0-0.4,0-0.6,0c-3.5,0-5.7-2.6-5.7-6.7c0-4.1,2.2-6.7,5.7-6.7 s5.7,2.6,5.7,6.7c0,0.3,0,0.6,0,0.9l3.6,4.2c0.7-1.5,1-3.2,1-5.1c0-6.5-4.2-11-10.3-11c-6.1,0-10.3,4.5-10.3,11s4.2,11,10.3,11 c1.2,0,2.3-0.2,3.4-0.5l0.5-0.2L246.9,31z"
id="path9804" />
</g>
</g>
<g
id="g9824">
<path
class="st1"
d="M58.7,13.2c4.6,0,7.4,2.5,7.4,6.5h-4.6c0-1.5-1.1-2.4-2.9-2.4c-1.9,0-3.1,0.9-3.1,2.3c0,1.3,0.7,1.9,2.2,2.2 l3.2,0.7c3.8,0.8,5.6,2.6,5.6,5.9c0,4.1-3.2,6.8-8.1,6.8c-4.7,0-7.8-2.6-7.8-6.5h4.6c0,1.6,1.1,2.4,3.2,2.4 c2.1,0,3.4-0.8,3.4-2.2c0-1.2-0.5-1.8-2-2.1l-3.2-0.7c-3.8-0.8-5.7-2.9-5.7-6.4C50.9,16,54.1,13.2,58.7,13.2z"
id="path9810" />
<path
class="st1"
d="M70.4,34.9L78,13.6h4.5l7.6,21.3h-4.9l-1.5-4.5h-6.9l-1.5,4.5H70.4z M78.1,26.5h4.2L80.8,22 c-0.2-0.7-0.5-1.6-0.6-2.1c-0.1,0.5-0.3,1.3-0.6,2.1L78.1,26.5z"
id="path9812" />
<path
class="st1"
d="M95.3,34.9V13.6h4.6l9,13.5V13.6h4.6v21.3h-4.6l-9-13.5v13.5H95.3z"
id="path9814" />
<path
class="st1"
d="M120.7,34.9V13.6h8c6.2,0,10.6,4.4,10.6,10.7c0,6.2-4.2,10.6-10.3,10.6H120.7z M125.4,17.9v12.6h3.2 c3.7,0,5.8-2.3,5.8-6.3c0-4-2.3-6.4-6.1-6.4H125.4z"
id="path9816" />
<path
class="st1"
d="M145.4,13.6h8.8c4.3,0,6.9,2.2,6.9,5.9c0,2.3-1,3.9-3,4.8c2.1,0.7,3.2,2.3,3.2,4.7c0,3.8-2.5,5.9-7.1,5.9 h-8.8V13.6z M150.1,17.7v4.6h3.7c1.7,0,2.6-0.8,2.6-2.4c0-1.5-0.9-2.3-2.6-2.3H150.1z M150.1,26.2v4.6h3.9c1.7,0,2.6-0.8,2.6-2.4 c0-1.4-0.9-2.2-2.6-2.2H150.1z"
id="path9818" />
<path
class="st1"
d="M176.5,35.2c-6.1,0-10.4-4.5-10.4-11s4.3-11,10.4-11c6.2,0,10.4,4.5,10.4,11S182.7,35.2,176.5,35.2z M176.6,17.7c-3.4,0-5.5,2.4-5.5,6.5c0,4.1,2.1,6.5,5.5,6.5c3.4,0,5.5-2.5,5.5-6.5C182.1,20.2,180,17.7,176.6,17.7z"
id="path9820" />
<path
class="st1"
d="M190.4,13.6h5.5l1.8,2.8c0.8,1.2,1.5,2.5,2.5,4.3l4.3-7h5.4l-6.7,10.6l6.7,10.6h-5.5l-1.4-2.2 c-1.1-1.7-1.8-3-2.8-4.9l-4.6,7.1h-5.5l7.1-10.6L190.4,13.6z"
id="path9822" />
</g>
<path
class="st0"
d="M229,34.9h4.7L226,13.6h-4.3l-7.7,21.2h4.6l1.6-4.5h7.1L229,34.9z M223.9,20.3 C223.9,20.3,223.9,20.3,223.9,20.3C223.9,20.2,223.9,20.2,223.9,20.3l2.2,6.2h-4.4L223.9,20.3z"
id="path9826" />
</g>
<g
id="g9832">
<path
class="st1"
d="M259.5,11.2h3.9v1h-1.3v3.1h-1.3v-3.1h-1.3V11.2L259.5,11.2z M264,11.2h1.7l0.6,2.5l0.6-2.5h1.7v4.1h-1v-3.1 l-0.8,3.1h-0.9l-0.8-3.1v3.1h-1V11.2L264,11.2z"
id="path9830" />
</g>
</g>
</svg>
Side by Side

Before

Width:  |  Height:  |  Size: 4.7 KiB

2
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/manifest.yaml
View File

@@ -1,2 +0,0 @@
library: spqlios-fft
version: 2.0.0

27
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/scripts/auto-release.sh
View File

@@ -1,27 +0,0 @@
#!/bin/sh
# this script generates one tag if there is a version change in manifest.yaml
cd `dirname $0`/..
if [ "v$1" = "v-y" ]; then
echo "production mode!";
fi
changes=`git diff HEAD~1..HEAD -- manifest.yaml | grep 'version:'`
oldversion=$(echo "$changes" | grep '^-version:' | cut '-d ' -f2)
version=$(echo "$changes" | grep '^+version:' | cut '-d ' -f2)
echo "Versions: $oldversion --> $version"
if [ "v$oldversion" = "v$version" ]; then
echo "Same version - nothing to do"; exit 0;
fi
if [ "v$1" = "v-y" ]; then
git config user.name github-actions
git config user.email github-actions@github.com
git tag -a "v$version" -m "Version $version"
git push origin "v$version"
else
cat <<EOF
# the script would do:
git tag -a "v$version" -m "Version $version"
git push origin "v$version"
EOF
fi

102
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/scripts/ci-pkg
View File

@@ -1,102 +0,0 @@
#!/bin/sh
# ONLY USE A PREFIX YOU ARE CONFIDENT YOU CAN WIPE OUT ENTIRELY
CI_INSTALL_PREFIX=/opt/spqlios
CI_REPO_URL=https://spq-dav.algonics.net/ci
WORKDIR=`pwd`
if [ "x$DESTDIR" = "x" ]; then
DESTDIR=/
else
mkdir -p $DESTDIR
DESTDIR=`realpath $DESTDIR`
fi
DIR=`dirname "$0"`
cd $DIR/..
DIR=`pwd`
FULL_UNAME=`uname -a | tr '[A-Z]' '[a-z]'`
HOST=`echo $FULL_UNAME | sed 's/ .*//'`
ARCH=none
case "$HOST" in
*linux*)
DISTRIB=`lsb_release -c | awk '{print $2}' | tr '[A-Z]' '[a-z]'`
HOST=linux-$DISTRIB
;;
*darwin*)
HOST=darwin
;;
*mingw*|*msys*)
DISTRIB=`echo $MSYSTEM | tr '[A-Z]' '[a-z]'`
HOST=msys64-$DISTRIB
;;
*)
echo "Host unknown: $HOST";
exit 1
esac
case "$FULL_UNAME" in
*x86_64*)
ARCH=x86_64
;;
*aarch64*)
ARCH=aarch64
;;
*arm64*)
ARCH=arm64
;;
*)
echo "Architecture unknown: $FULL_UNAME";
exit 1
esac
UNAME="$HOST-$ARCH"
CMH=
if [ -d lib/spqlios/.git ]; then
CMH=`git submodule status | sed 's/\(..........\).*/\1/'`
else
CMH=`git rev-parse HEAD | sed 's/\(..........\).*/\1/'`
fi
FNAME=spqlios-arithmetic-$CMH-$UNAME.tar.gz
cat <<EOF
================= CI MINI-PACKAGER ==================
Work Dir: WORKDIR=$WORKDIR
Spq Dir: DIR=$DIR
Install Root: DESTDIR=$DESTDIR
Install Prefix: CI_INSTALL_PREFIX=$CI_INSTALL_PREFIX
Archive Name: FNAME=$FNAME
CI WebDav: CI_REPO_URL=$CI_REPO_URL
=====================================================
EOF
if [ "x$1" = "xcreate" ]; then
rm -rf dist
cmake -B build -S . -DCMAKE_INSTALL_PREFIX="$CI_INSTALL_PREFIX" -DCMAKE_BUILD_TYPE=Release -DENABLE_TESTING=ON -DWARNING_PARANOID=ON -DDEVMODE_INSTALL=ON || exit 1
cmake --build build || exit 1
rm -rf "$DIR/dist" 2>/dev/null
rm -f "$DIR/$FNAME" 2>/dev/null
DESTDIR="$DIR/dist" cmake --install build || exit 1
if [ -d "$DIR/dist$CI_INSTALL_PREFIX" ]; then
tar -C "$DIR/dist" -cvzf "$DIR/$FNAME" .
else
# fix since msys can mess up the paths
REAL_DEST=`find "$DIR/dist" -type d -exec test -d "{}$CI_INSTALL_PREFIX" \; -print`
echo "REAL_DEST: $REAL_DEST"
[ -d "$REAL_DEST$CI_INSTALL_PREFIX" ] && tar -C "$REAL_DEST" -cvzf "$DIR/$FNAME" .
fi
[ -f "$DIR/$FNAME" ] || { echo "failed to create $DIR/$FNAME"; exit 1; }
[ "x$CI_CREDS" = "x" ] && { echo "CI_CREDS is not set: not uploading"; exit 1; }
curl -u "$CI_CREDS" -T "$DIR/$FNAME" "$CI_REPO_URL/$FNAME"
fi
if [ "x$1" = "xinstall" ]; then
[ "x$CI_CREDS" = "x" ] && { echo "CI_CREDS is not set: not downloading"; exit 1; }
# cleaning
rm -rf "$DESTDIR$CI_INSTALL_PREFIX"/* 2>/dev/null
rm -f "$DIR/$FNAME" 2>/dev/null
# downloading
curl -u "$CI_CREDS" -o "$DIR/$FNAME" "$CI_REPO_URL/$FNAME"
[ -f "$DIR/$FNAME" ] || { echo "failed to download $DIR/$FNAME"; exit 0; }
# installing
mkdir -p $DESTDIR
tar -C "$DESTDIR" -xvzf "$DIR/$FNAME"
exit 0
fi

181
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/scripts/prepare-release
View File

@@ -1,181 +0,0 @@
#!/usr/bin/perl
##
## This script will help update manifest.yaml and Changelog.md before a release
## Any merge to master that changes the version line in manifest.yaml
## is considered as a new release.
##
## When ready to make a release, please run ./scripts/prepare-release
## and commit push the final result!
use File::Basename;
use Cwd 'abs_path';
# find its way to the root of git's repository
my $scriptsdirname = dirname(abs_path(__FILE__));
chdir "$scriptsdirname/..";
print "✓ Entering directory:".`pwd`;
# ensures that the current branch is ahead of origin/main
my $diff= `git diff`;
chop $diff;
if ($diff =~ /./) {
die("ERROR: Please commit all the changes before calling the prepare-release script.");
} else {
print("✓ All changes are comitted.\n");
}
system("git fetch origin");
my $vcount = `git rev-list --left-right --count origin/main...HEAD`;
$vcount =~ /^([0-9]+)[ \t]*([0-9]+)$/;
if ($2>0) {
die("ERROR: the current HEAD is not ahead of origin/main\n. Please use git merge origin/main.");
} else {
print("✓ Current HEAD is up to date with origin/main.\n");
}
mkdir ".changes";
my $currentbranch = `git rev-parse --abbrev-ref HEAD`;
chop $currentbranch;
$currentbranch =~ s/[^a-zA-Z._-]+/-/g;
my $changefile=".changes/$currentbranch.md";
my $origmanifestfile=".changes/$currentbranch--manifest.yaml";
my $origchangelogfile=".changes/$currentbranch--Changelog.md";
my $exit_code=system("wget -O $origmanifestfile https://raw.githubusercontent.com/tfhe/spqlios-fft/main/manifest.yaml");
if ($exit_code!=0 or ! -f $origmanifestfile) {
die("ERROR: failed to download manifest.yaml");
}
$exit_code=system("wget -O $origchangelogfile https://raw.githubusercontent.com/tfhe/spqlios-fft/main/Changelog.md");
if ($exit_code!=0 or ! -f $origchangelogfile) {
die("ERROR: failed to download Changelog.md");
}
# read the current version (from origin/main manifest)
my $vmajor = 0;
my $vminor = 0;
my $vpatch = 0;
my $versionline = `grep '^version: ' $origmanifestfile | cut -d" " -f2`;
chop $versionline;
if (not $versionline =~ /^([0-9]+)\.([0-9]+)\.([0-9]+)$/) {
die("ERROR: invalid version in manifest file: $versionline\n");
} else {
$vmajor = int($1);
$vminor = int($2);
$vpatch = int($3);
}
print "Version in manifest file: $vmajor.$vminor.$vpatch\n";
if (not -f $changefile) {
## create a changes file
open F,">$changefile";
print F "# Changefile for branch $currentbranch\n\n";
print F "## Type of release (major,minor,patch)?\n\n";
print F "releasetype: patch\n\n";
print F "## What has changed (please edit)?\n\n";
print F "- This has changed.\n";
close F;
}
system("editor $changefile");
# compute the new version
my $nvmajor;
my $nvminor;
my $nvpatch;
my $changelog;
my $recordchangelog=0;
open F,"$changefile";
while ($line=<F>) {
chop $line;
if ($recordchangelog) {
($line =~ /^$/) and next;
$changelog .= "$line\n";
next;
}
if ($line =~ /^releasetype *: *patch *$/) {
$nvmajor=$vmajor;
$nvminor=$vminor;
$nvpatch=$vpatch+1;
}
if ($line =~ /^releasetype *: *minor *$/) {
$nvmajor=$vmajor;
$nvminor=$vminor+1;
$nvpatch=0;
}
if ($line =~ /^releasetype *: *major *$/) {
$nvmajor=$vmajor+1;
$nvminor=0;
$nvpatch=0;
}
if ($line =~ /^## What has changed/) {
$recordchangelog=1;
}
}
close F;
print "New version: $nvmajor.$nvminor.$nvpatch\n";
print "Changes:\n$changelog";
# updating manifest.yaml
open F,"manifest.yaml";
open G,">.changes/manifest.yaml";
while ($line=<F>) {
if ($line =~ /^version *: */) {
print G "version: $nvmajor.$nvminor.$nvpatch\n";
next;
}
print G $line;
}
close F;
close G;
# updating Changelog.md
open F,"$origchangelogfile";
open G,">.changes/Changelog.md";
print G <<EOF
# Changelog
All notable changes to this project will be documented in this file.
this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
EOF
;
print G "## [$nvmajor.$nvminor.$nvpatch] - ".`date '+%Y-%m-%d'`."\n";
print G "$changelog\n";
my $skip_section=1;
while ($line=<F>) {
if ($line =~ /^## +\[([0-9]+)\.([0-9]+)\.([0-9]+)\] +/) {
if ($1>$nvmajor) {
die("ERROR: found larger version $1.$2.$3 in the Changelog.md\n");
} elsif ($1<$nvmajor) {
$skip_section=0;
} elsif ($2>$nvminor) {
die("ERROR: found larger version $1.$2.$3 in the Changelog.md\n");
} elsif ($2<$nvminor) {
$skip_section=0;
} elsif ($3>$nvpatch) {
die("ERROR: found larger version $1.$2.$3 in the Changelog.md\n");
} elsif ($2<$nvpatch) {
$skip_section=0;
} else {
$skip_section=1;
}
}
($skip_section) and next;
print G $line;
}
close F;
close G;
print "-------------------------------------\n";
print "THIS WILL BE UPDATED:\n";
print "-------------------------------------\n";
system("diff -u manifest.yaml .changes/manifest.yaml");
system("diff -u Changelog.md .changes/Changelog.md");
print "-------------------------------------\n";
print "To proceed: press <enter> otherwise <CTRL+C>\n";
my $bla;
$bla=<STDIN>;
system("cp -vf .changes/manifest.yaml manifest.yaml");
system("cp -vf .changes/Changelog.md Changelog.md");
system("git commit -a -m \"Update version and changelog.\"");
system("git push");
print("✓ Changes have been committed and pushed!\n");
print("✓ A new release will be created when this branch is merged to main.\n");

223
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/CMakeLists.txt
View File

@@ -1,223 +0,0 @@
enable_language(ASM)
# C source files that are compiled for all targets (i.e. reference code)
set(SRCS_GENERIC
commons.c
commons_private.c
coeffs/coeffs_arithmetic.c
arithmetic/vec_znx.c
arithmetic/vec_znx_dft.c
arithmetic/vector_matrix_product.c
cplx/cplx_common.c
cplx/cplx_conversions.c
cplx/cplx_fft_asserts.c
cplx/cplx_fft_ref.c
cplx/cplx_fftvec_ref.c
cplx/cplx_ifft_ref.c
cplx/spqlios_cplx_fft.c
reim4/reim4_arithmetic_ref.c
reim4/reim4_fftvec_addmul_ref.c
reim4/reim4_fftvec_conv_ref.c
reim/reim_conversions.c
reim/reim_fft_ifft.c
reim/reim_fft_ref.c
reim/reim_fftvec_ref.c
reim/reim_ifft_ref.c
reim/reim_ifft_ref.c
reim/reim_to_tnx_ref.c
q120/q120_ntt.c
q120/q120_arithmetic_ref.c
q120/q120_arithmetic_simple.c
arithmetic/scalar_vector_product.c
arithmetic/vec_znx_big.c
arithmetic/znx_small.c
arithmetic/module_api.c
arithmetic/zn_vmp_int8_ref.c
arithmetic/zn_vmp_int16_ref.c
arithmetic/zn_vmp_int32_ref.c
arithmetic/zn_vmp_ref.c
arithmetic/zn_api.c
arithmetic/zn_conversions_ref.c
arithmetic/zn_approxdecomp_ref.c
arithmetic/vec_rnx_api.c
arithmetic/vec_rnx_conversions_ref.c
arithmetic/vec_rnx_svp_ref.c
reim/reim_execute.c
cplx/cplx_execute.c
reim4/reim4_execute.c
arithmetic/vec_rnx_arithmetic.c
arithmetic/vec_rnx_approxdecomp_ref.c
arithmetic/vec_rnx_vmp_ref.c
)
# C or assembly source files compiled only on x86 targets
set(SRCS_X86
)
# C or assembly source files compiled only on aarch64 targets
set(SRCS_AARCH64
cplx/cplx_fallbacks_aarch64.c
reim/reim_fallbacks_aarch64.c
reim4/reim4_fallbacks_aarch64.c
q120/q120_fallbacks_aarch64.c
reim/reim_fft_neon.c
)
# C or assembly source files compiled only on x86: avx, avx2, fma targets
set(SRCS_FMA_C
arithmetic/vector_matrix_product_avx.c
cplx/cplx_conversions_avx2_fma.c
cplx/cplx_fft_avx2_fma.c
cplx/cplx_fft_sse.c
cplx/cplx_fftvec_avx2_fma.c
cplx/cplx_ifft_avx2_fma.c
reim4/reim4_arithmetic_avx2.c
reim4/reim4_fftvec_conv_fma.c
reim4/reim4_fftvec_addmul_fma.c
reim/reim_conversions_avx.c
reim/reim_fft4_avx_fma.c
reim/reim_fft8_avx_fma.c
reim/reim_ifft4_avx_fma.c
reim/reim_ifft8_avx_fma.c
reim/reim_fft_avx2.c
reim/reim_ifft_avx2.c
reim/reim_to_tnx_avx.c
reim/reim_fftvec_fma.c
)
set(SRCS_FMA_ASM
cplx/cplx_fft16_avx_fma.s
cplx/cplx_ifft16_avx_fma.s
reim/reim_fft16_avx_fma.s
reim/reim_ifft16_avx_fma.s
)
set(SRCS_FMA_WIN32_ASM
cplx/cplx_fft16_avx_fma_win32.s
cplx/cplx_ifft16_avx_fma_win32.s
reim/reim_fft16_avx_fma_win32.s
reim/reim_ifft16_avx_fma_win32.s
)
set_source_files_properties(${SRCS_FMA_C} PROPERTIES COMPILE_OPTIONS "-mfma;-mavx;-mavx2")
set_source_files_properties(${SRCS_FMA_ASM} PROPERTIES COMPILE_OPTIONS "-mfma;-mavx;-mavx2")
# C or assembly source files compiled only on x86: avx512f/vl/dq + fma targets
set(SRCS_AVX512
cplx/cplx_fft_avx512.c
)
set_source_files_properties(${SRCS_AVX512} PROPERTIES COMPILE_OPTIONS "-mfma;-mavx512f;-mavx512vl;-mavx512dq")
# C or assembly source files compiled only on x86: avx2 + bmi targets
set(SRCS_AVX2
arithmetic/vec_znx_avx.c
coeffs/coeffs_arithmetic_avx.c
arithmetic/vec_znx_dft_avx2.c
arithmetic/zn_vmp_int8_avx.c
arithmetic/zn_vmp_int16_avx.c
arithmetic/zn_vmp_int32_avx.c
q120/q120_arithmetic_avx2.c
q120/q120_ntt_avx2.c
arithmetic/vec_rnx_arithmetic_avx.c
arithmetic/vec_rnx_approxdecomp_avx.c
arithmetic/vec_rnx_vmp_avx.c
)
set_source_files_properties(${SRCS_AVX2} PROPERTIES COMPILE_OPTIONS "-mbmi2;-mavx2")
# C source files on float128 via libquadmath on x86 targets targets
set(SRCS_F128
cplx_f128/cplx_fft_f128.c
cplx_f128/cplx_fft_f128.h
)
# H header files containing the public API (these headers are installed)
set(HEADERSPUBLIC
commons.h
arithmetic/vec_znx_arithmetic.h
arithmetic/vec_rnx_arithmetic.h
arithmetic/zn_arithmetic.h
cplx/cplx_fft.h
reim/reim_fft.h
q120/q120_common.h
q120/q120_arithmetic.h
q120/q120_ntt.h
)
# H header files containing the private API (these headers are used internally)
set(HEADERSPRIVATE
commons_private.h
cplx/cplx_fft_internal.h
cplx/cplx_fft_private.h
reim4/reim4_arithmetic.h
reim4/reim4_fftvec_internal.h
reim4/reim4_fftvec_private.h
reim4/reim4_fftvec_public.h
reim/reim_fft_internal.h
reim/reim_fft_private.h
q120/q120_arithmetic_private.h
q120/q120_ntt_private.h
arithmetic/vec_znx_arithmetic.h
arithmetic/vec_rnx_arithmetic_private.h
arithmetic/vec_rnx_arithmetic_plugin.h
arithmetic/zn_arithmetic_private.h
arithmetic/zn_arithmetic_plugin.h
coeffs/coeffs_arithmetic.h
reim/reim_fft_core_template.h
)
set(SPQLIOSSOURCES
${SRCS_GENERIC}
${HEADERSPUBLIC}
${HEADERSPRIVATE}
)
if (${X86})
set(SPQLIOSSOURCES ${SPQLIOSSOURCES}
${SRCS_X86}
${SRCS_FMA_C}
${SRCS_FMA_ASM}
${SRCS_AVX2}
${SRCS_AVX512}
)
elseif (${X86_WIN32})
set(SPQLIOSSOURCES ${SPQLIOSSOURCES}
#${SRCS_X86}
${SRCS_FMA_C}
${SRCS_FMA_WIN32_ASM}
${SRCS_AVX2}
${SRCS_AVX512}
)
elseif (${AARCH64})
set(SPQLIOSSOURCES ${SPQLIOSSOURCES}
${SRCS_AARCH64}
)
endif ()
set(SPQLIOSLIBDEP
m # libmath depencency for cosinus/sinus functions
)
if (ENABLE_SPQLIOS_F128)
find_library(quadmath REQUIRED NAMES quadmath)
set(SPQLIOSSOURCES ${SPQLIOSSOURCES} ${SRCS_F128})
set(SPQLIOSLIBDEP ${SPQLIOSLIBDEP} quadmath)
endif (ENABLE_SPQLIOS_F128)
add_library(libspqlios-static STATIC ${SPQLIOSSOURCES})
add_library(libspqlios SHARED ${SPQLIOSSOURCES})
set_property(TARGET libspqlios-static PROPERTY POSITION_INDEPENDENT_CODE ON)
set_property(TARGET libspqlios PROPERTY OUTPUT_NAME spqlios)
set_property(TARGET libspqlios-static PROPERTY OUTPUT_NAME spqlios)
set_property(TARGET libspqlios PROPERTY POSITION_INDEPENDENT_CODE ON)
set_property(TARGET libspqlios PROPERTY SOVERSION ${SPQLIOS_VERSION_MAJOR})
set_property(TARGET libspqlios PROPERTY VERSION ${SPQLIOS_VERSION})
if (NOT APPLE)
target_link_options(libspqlios-static PUBLIC -Wl,--no-undefined)
target_link_options(libspqlios PUBLIC -Wl,--no-undefined)
endif()
target_link_libraries(libspqlios ${SPQLIOSLIBDEP})
target_link_libraries(libspqlios-static ${SPQLIOSLIBDEP})
install(TARGETS libspqlios-static)
install(TARGETS libspqlios)
# install the public headers only
foreach (file ${HEADERSPUBLIC})
get_filename_component(dir ${file} DIRECTORY)
install(FILES ${file} DESTINATION include/spqlios/${dir})
endforeach ()

172
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/module_api.c
View File

@@ -1,172 +0,0 @@
#include <string.h>
#include "vec_znx_arithmetic_private.h"
static void fill_generic_virtual_table(MODULE* module) {
// TODO add default ref handler here
module->func.vec_znx_zero = vec_znx_zero_ref;
module->func.vec_znx_copy = vec_znx_copy_ref;
module->func.vec_znx_negate = vec_znx_negate_ref;
module->func.vec_znx_add = vec_znx_add_ref;
module->func.vec_znx_sub = vec_znx_sub_ref;
module->func.vec_znx_rotate = vec_znx_rotate_ref;
module->func.vec_znx_mul_xp_minus_one = vec_znx_mul_xp_minus_one_ref;
module->func.vec_znx_automorphism = vec_znx_automorphism_ref;
module->func.vec_znx_normalize_base2k = vec_znx_normalize_base2k_ref;
module->func.vec_znx_normalize_base2k_tmp_bytes = vec_znx_normalize_base2k_tmp_bytes_ref;
if (CPU_SUPPORTS("avx2")) {
// TODO add avx handlers here
module->func.vec_znx_negate = vec_znx_negate_avx;
module->func.vec_znx_add = vec_znx_add_avx;
module->func.vec_znx_sub = vec_znx_sub_avx;
}
}
static void fill_fft64_virtual_table(MODULE* module) {
// TODO add default ref handler here
// module->func.vec_znx_dft = ...;
module->func.vec_znx_big_normalize_base2k = fft64_vec_znx_big_normalize_base2k;
module->func.vec_znx_big_normalize_base2k_tmp_bytes = fft64_vec_znx_big_normalize_base2k_tmp_bytes;
module->func.vec_znx_big_range_normalize_base2k = fft64_vec_znx_big_range_normalize_base2k;
module->func.vec_znx_big_range_normalize_base2k_tmp_bytes = fft64_vec_znx_big_range_normalize_base2k_tmp_bytes;
module->func.vec_znx_dft = fft64_vec_znx_dft;
module->func.vec_znx_idft = fft64_vec_znx_idft;
module->func.vec_dft_add = fft64_vec_dft_add;
module->func.vec_dft_sub = fft64_vec_dft_sub;
module->func.vec_znx_idft_tmp_bytes = fft64_vec_znx_idft_tmp_bytes;
module->func.vec_znx_idft_tmp_a = fft64_vec_znx_idft_tmp_a;
module->func.vec_znx_big_add = fft64_vec_znx_big_add;
module->func.vec_znx_big_add_small = fft64_vec_znx_big_add_small;
module->func.vec_znx_big_add_small2 = fft64_vec_znx_big_add_small2;
module->func.vec_znx_big_sub = fft64_vec_znx_big_sub;
module->func.vec_znx_big_sub_small_a = fft64_vec_znx_big_sub_small_a;
module->func.vec_znx_big_sub_small_b = fft64_vec_znx_big_sub_small_b;
module->func.vec_znx_big_sub_small2 = fft64_vec_znx_big_sub_small2;
module->func.vec_znx_big_rotate = fft64_vec_znx_big_rotate;
module->func.vec_znx_big_automorphism = fft64_vec_znx_big_automorphism;
module->func.svp_prepare = fft64_svp_prepare_ref;
module->func.svp_apply_dft = fft64_svp_apply_dft_ref;
module->func.svp_apply_dft_to_dft = fft64_svp_apply_dft_to_dft_ref;
module->func.znx_small_single_product = fft64_znx_small_single_product;
module->func.znx_small_single_product_tmp_bytes = fft64_znx_small_single_product_tmp_bytes;
module->func.vmp_prepare_contiguous = fft64_vmp_prepare_contiguous_ref;
module->func.vmp_prepare_tmp_bytes = fft64_vmp_prepare_tmp_bytes;
module->func.vmp_apply_dft = fft64_vmp_apply_dft_ref;
module->func.vmp_apply_dft_add = fft64_vmp_apply_dft_add_ref;
module->func.vmp_apply_dft_tmp_bytes = fft64_vmp_apply_dft_tmp_bytes;
module->func.vmp_apply_dft_to_dft = fft64_vmp_apply_dft_to_dft_ref;
module->func.vmp_apply_dft_to_dft_add = fft64_vmp_apply_dft_to_dft_add_ref;
module->func.vmp_apply_dft_to_dft_tmp_bytes = fft64_vmp_apply_dft_to_dft_tmp_bytes;
module->func.bytes_of_vec_znx_dft = fft64_bytes_of_vec_znx_dft;
module->func.bytes_of_vec_znx_big = fft64_bytes_of_vec_znx_big;
module->func.bytes_of_svp_ppol = fft64_bytes_of_svp_ppol;
module->func.bytes_of_vmp_pmat = fft64_bytes_of_vmp_pmat;
if (CPU_SUPPORTS("avx2")) {
// TODO add avx handlers here
// TODO: enable when avx implementation is done
module->func.vmp_prepare_contiguous = fft64_vmp_prepare_contiguous_avx;
module->func.vmp_apply_dft = fft64_vmp_apply_dft_avx;
module->func.vmp_apply_dft_add = fft64_vmp_apply_dft_add_avx;
module->func.vmp_apply_dft_to_dft = fft64_vmp_apply_dft_to_dft_avx;
module->func.vmp_apply_dft_to_dft_add = fft64_vmp_apply_dft_to_dft_add_avx;
}
}
static void fill_ntt120_virtual_table(MODULE* module) {
// TODO add default ref handler here
// module->func.vec_znx_dft = ...;
if (CPU_SUPPORTS("avx2")) {
// TODO add avx handlers here
module->func.vec_znx_dft = ntt120_vec_znx_dft_avx;
module->func.vec_znx_idft = ntt120_vec_znx_idft_avx;
module->func.vec_znx_idft_tmp_bytes = ntt120_vec_znx_idft_tmp_bytes_avx;
module->func.vec_znx_idft_tmp_a = ntt120_vec_znx_idft_tmp_a_avx;
}
}
static void fill_virtual_table(MODULE* module) {
fill_generic_virtual_table(module);
switch (module->module_type) {
case FFT64:
fill_fft64_virtual_table(module);
break;
case NTT120:
fill_ntt120_virtual_table(module);
break;
default:
NOT_SUPPORTED(); // invalid type
}
}
static void fill_fft64_precomp(MODULE* module) {
// fill any necessary precomp stuff
module->mod.fft64.p_conv = new_reim_from_znx64_precomp(module->m, 50);
module->mod.fft64.p_fft = new_reim_fft_precomp(module->m, 0);
module->mod.fft64.p_reim_to_znx = new_reim_to_znx64_precomp(module->m, module->m, 63);
module->mod.fft64.p_ifft = new_reim_ifft_precomp(module->m, 0);
module->mod.fft64.p_addmul = new_reim_fftvec_addmul_precomp(module->m);
module->mod.fft64.mul_fft = new_reim_fftvec_mul_precomp(module->m);
module->mod.fft64.add_fft = new_reim_fftvec_add_precomp(module->m);
module->mod.fft64.sub_fft = new_reim_fftvec_sub_precomp(module->m);
}
static void fill_ntt120_precomp(MODULE* module) {
// fill any necessary precomp stuff
if (CPU_SUPPORTS("avx2")) {
module->mod.q120.p_ntt = q120_new_ntt_bb_precomp(module->nn);
module->mod.q120.p_intt = q120_new_intt_bb_precomp(module->nn);
}
}
static void fill_module_precomp(MODULE* module) {
switch (module->module_type) {
case FFT64:
fill_fft64_precomp(module);
break;
case NTT120:
fill_ntt120_precomp(module);
break;
default:
NOT_SUPPORTED(); // invalid type
}
}
static void fill_module(MODULE* module, uint64_t nn, MODULE_TYPE mtype) {
// init to zero to ensure that any non-initialized field bug is detected
// by at least a "proper" segfault
memset(module, 0, sizeof(MODULE));
module->module_type = mtype;
module->nn = nn;
module->m = nn >> 1;
fill_module_precomp(module);
fill_virtual_table(module);
}
EXPORT MODULE* new_module_info(uint64_t N, MODULE_TYPE mtype) {
MODULE* m = (MODULE*)malloc(sizeof(MODULE));
fill_module(m, N, mtype);
return m;
}
EXPORT void delete_module_info(MODULE* mod) {
switch (mod->module_type) {
case FFT64:
free(mod->mod.fft64.p_conv);
free(mod->mod.fft64.p_fft);
free(mod->mod.fft64.p_ifft);
free(mod->mod.fft64.p_reim_to_znx);
free(mod->mod.fft64.mul_fft);
free(mod->mod.fft64.p_addmul);
break;
case NTT120:
if (CPU_SUPPORTS("avx2")) {
q120_del_ntt_bb_precomp(mod->mod.q120.p_ntt);
q120_del_intt_bb_precomp(mod->mod.q120.p_intt);
}
break;
default:
break;
}
free(mod);
}
EXPORT uint64_t module_get_n(const MODULE* module) { return module->nn; }

102
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/scalar_vector_product.c
View File

@@ -1,102 +0,0 @@
#include <string.h>
#include "vec_znx_arithmetic_private.h"
EXPORT uint64_t bytes_of_svp_ppol(const MODULE* module) { return module->func.bytes_of_svp_ppol(module); }
EXPORT uint64_t fft64_bytes_of_svp_ppol(const MODULE* module) { return module->nn * sizeof(double); }
EXPORT SVP_PPOL* new_svp_ppol(const MODULE* module) { return spqlios_alloc(bytes_of_svp_ppol(module)); }
EXPORT void delete_svp_ppol(SVP_PPOL* ppol) { spqlios_free(ppol); }
// public wrappers
EXPORT void svp_prepare(const MODULE* module, // N
SVP_PPOL* ppol, // output
const int64_t* pol // a
) {
module->func.svp_prepare(module, ppol, pol);
}
/** @brief prepares a svp polynomial */
EXPORT void fft64_svp_prepare_ref(const MODULE* module, // N
SVP_PPOL* ppol, // output
const int64_t* pol // a
) {
reim_from_znx64(module->mod.fft64.p_conv, ppol, pol);
reim_fft(module->mod.fft64.p_fft, (double*)ppol);
}
EXPORT void svp_apply_dft(const MODULE* module, // N
const VEC_ZNX_DFT* res, uint64_t res_size, // output
const SVP_PPOL* ppol, // prepared pol
const int64_t* a, uint64_t a_size, uint64_t a_sl) {
module->func.svp_apply_dft(module, // N
res,
res_size, // output
ppol, // prepared pol
a, a_size, a_sl);
}
EXPORT void svp_apply_dft_to_dft(const MODULE* module, // N
const VEC_ZNX_DFT* res, uint64_t res_size,
uint64_t res_cols, // output
const SVP_PPOL* ppol, // prepared pol
const VEC_ZNX_DFT* a, uint64_t a_size, uint64_t a_cols) {
module->func.svp_apply_dft_to_dft(module, // N
res, res_size, res_cols, // output
ppol, a, a_size, a_cols // prepared pol
);
}
// result = ppol * a
EXPORT void fft64_svp_apply_dft_ref(const MODULE* module, // N
const VEC_ZNX_DFT* res, uint64_t res_size, // output
const SVP_PPOL* ppol, // prepared pol
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
) {
const uint64_t nn = module->nn;
double* const dres = (double*)res;
double* const dppol = (double*)ppol;
const uint64_t auto_end_idx = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < auto_end_idx; ++i) {
const int64_t* a_ptr = a + i * a_sl;
double* const res_ptr = dres + i * nn;
// copy the polynomial to res, apply fft in place, call fftvec_mul in place.
reim_from_znx64(module->mod.fft64.p_conv, res_ptr, a_ptr);
reim_fft(module->mod.fft64.p_fft, res_ptr);
reim_fftvec_mul(module->mod.fft64.mul_fft, res_ptr, res_ptr, dppol);
}
// then extend with zeros
memset(dres + auto_end_idx * nn, 0, (res_size - auto_end_idx) * nn * sizeof(double));
}
// result = ppol * a
EXPORT void fft64_svp_apply_dft_to_dft_ref(const MODULE* module, // N
const VEC_ZNX_DFT* res, uint64_t res_size,
uint64_t res_cols, // output
const SVP_PPOL* ppol, // prepared pol
const VEC_ZNX_DFT* a, uint64_t a_size,
uint64_t a_cols // a
) {
const uint64_t nn = module->nn;
const uint64_t res_sl = nn * res_cols;
const uint64_t a_sl = nn * a_cols;
double* const dres = (double*)res;
double* const da = (double*)a;
double* const dppol = (double*)ppol;
const uint64_t auto_end_idx = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < auto_end_idx; ++i) {
const double* a_ptr = da + i * a_sl;
double* const res_ptr = dres + i * res_sl;
reim_fftvec_mul(module->mod.fft64.mul_fft, res_ptr, a_ptr, dppol);
}
// then extend with zeros
for (uint64_t i = auto_end_idx; i < res_size; i++) {
memset(dres + i * res_sl, 0, nn * sizeof(double));
}
}

344
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/vec_rnx_api.c
View File

@@ -1,344 +0,0 @@
#include <string.h>
#include "vec_rnx_arithmetic_private.h"
void fft64_init_rnx_module_precomp(MOD_RNX* module) {
// Add here initialization of items that are in the precomp
const uint64_t m = module->m;
module->precomp.fft64.p_fft = new_reim_fft_precomp(m, 0);
module->precomp.fft64.p_ifft = new_reim_ifft_precomp(m, 0);
module->precomp.fft64.p_fftvec_add = new_reim_fftvec_add_precomp(m);
module->precomp.fft64.p_fftvec_mul = new_reim_fftvec_mul_precomp(m);
module->precomp.fft64.p_fftvec_addmul = new_reim_fftvec_addmul_precomp(m);
}
void fft64_finalize_rnx_module_precomp(MOD_RNX* module) {
// Add here deleters for items that are in the precomp
delete_reim_fft_precomp(module->precomp.fft64.p_fft);
delete_reim_ifft_precomp(module->precomp.fft64.p_ifft);
delete_reim_fftvec_add_precomp(module->precomp.fft64.p_fftvec_add);
delete_reim_fftvec_mul_precomp(module->precomp.fft64.p_fftvec_mul);
delete_reim_fftvec_addmul_precomp(module->precomp.fft64.p_fftvec_addmul);
}
void fft64_init_rnx_module_vtable(MOD_RNX* module) {
// Add function pointers here
module->vtable.vec_rnx_add = vec_rnx_add_ref;
module->vtable.vec_rnx_zero = vec_rnx_zero_ref;
module->vtable.vec_rnx_copy = vec_rnx_copy_ref;
module->vtable.vec_rnx_negate = vec_rnx_negate_ref;
module->vtable.vec_rnx_sub = vec_rnx_sub_ref;
module->vtable.vec_rnx_rotate = vec_rnx_rotate_ref;
module->vtable.vec_rnx_automorphism = vec_rnx_automorphism_ref;
module->vtable.vec_rnx_mul_xp_minus_one = vec_rnx_mul_xp_minus_one_ref;
module->vtable.rnx_vmp_apply_dft_to_dft_tmp_bytes = fft64_rnx_vmp_apply_dft_to_dft_tmp_bytes_ref;
module->vtable.rnx_vmp_apply_dft_to_dft = fft64_rnx_vmp_apply_dft_to_dft_ref;
module->vtable.rnx_vmp_apply_tmp_a_tmp_bytes = fft64_rnx_vmp_apply_tmp_a_tmp_bytes_ref;
module->vtable.rnx_vmp_apply_tmp_a = fft64_rnx_vmp_apply_tmp_a_ref;
module->vtable.rnx_vmp_prepare_tmp_bytes = fft64_rnx_vmp_prepare_tmp_bytes_ref;
module->vtable.rnx_vmp_prepare_contiguous = fft64_rnx_vmp_prepare_contiguous_ref;
module->vtable.rnx_vmp_prepare_dblptr = fft64_rnx_vmp_prepare_dblptr_ref;
module->vtable.rnx_vmp_prepare_row = fft64_rnx_vmp_prepare_row_ref;
module->vtable.bytes_of_rnx_vmp_pmat = fft64_bytes_of_rnx_vmp_pmat;
module->vtable.rnx_approxdecomp_from_tnxdbl = rnx_approxdecomp_from_tnxdbl_ref;
module->vtable.vec_rnx_to_znx32 = vec_rnx_to_znx32_ref;
module->vtable.vec_rnx_from_znx32 = vec_rnx_from_znx32_ref;
module->vtable.vec_rnx_to_tnx32 = vec_rnx_to_tnx32_ref;
module->vtable.vec_rnx_from_tnx32 = vec_rnx_from_tnx32_ref;
module->vtable.vec_rnx_to_tnxdbl = vec_rnx_to_tnxdbl_ref;
module->vtable.bytes_of_rnx_svp_ppol = fft64_bytes_of_rnx_svp_ppol;
module->vtable.rnx_svp_prepare = fft64_rnx_svp_prepare_ref;
module->vtable.rnx_svp_apply = fft64_rnx_svp_apply_ref;
// Add optimized function pointers here
if (CPU_SUPPORTS("avx")) {
module->vtable.vec_rnx_add = vec_rnx_add_avx;
module->vtable.vec_rnx_sub = vec_rnx_sub_avx;
module->vtable.vec_rnx_negate = vec_rnx_negate_avx;
module->vtable.rnx_vmp_apply_dft_to_dft_tmp_bytes = fft64_rnx_vmp_apply_dft_to_dft_tmp_bytes_avx;
module->vtable.rnx_vmp_apply_dft_to_dft = fft64_rnx_vmp_apply_dft_to_dft_avx;
module->vtable.rnx_vmp_apply_tmp_a_tmp_bytes = fft64_rnx_vmp_apply_tmp_a_tmp_bytes_avx;
module->vtable.rnx_vmp_apply_tmp_a = fft64_rnx_vmp_apply_tmp_a_avx;
module->vtable.rnx_vmp_prepare_tmp_bytes = fft64_rnx_vmp_prepare_tmp_bytes_avx;
module->vtable.rnx_vmp_prepare_contiguous = fft64_rnx_vmp_prepare_contiguous_avx;
module->vtable.rnx_vmp_prepare_dblptr = fft64_rnx_vmp_prepare_dblptr_avx;
module->vtable.rnx_vmp_prepare_row = fft64_rnx_vmp_prepare_row_avx;
module->vtable.rnx_approxdecomp_from_tnxdbl = rnx_approxdecomp_from_tnxdbl_avx;
}
}
void init_rnx_module_info(MOD_RNX* module, //
uint64_t n, RNX_MODULE_TYPE mtype) {
memset(module, 0, sizeof(MOD_RNX));
module->n = n;
module->m = n >> 1;
module->mtype = mtype;
switch (mtype) {
case FFT64:
fft64_init_rnx_module_precomp(module);
fft64_init_rnx_module_vtable(module);
break;
default:
NOT_SUPPORTED(); // unknown mtype
}
}
void finalize_rnx_module_info(MOD_RNX* module) {
if (module->custom) module->custom_deleter(module->custom);
switch (module->mtype) {
case FFT64:
fft64_finalize_rnx_module_precomp(module);
// fft64_finalize_rnx_module_vtable(module); // nothing to finalize
break;
default:
NOT_SUPPORTED(); // unknown mtype
}
}
EXPORT MOD_RNX* new_rnx_module_info(uint64_t nn, RNX_MODULE_TYPE mtype) {
MOD_RNX* res = (MOD_RNX*)malloc(sizeof(MOD_RNX));
init_rnx_module_info(res, nn, mtype);
return res;
}
EXPORT void delete_rnx_module_info(MOD_RNX* module_info) {
finalize_rnx_module_info(module_info);
free(module_info);
}
EXPORT uint64_t rnx_module_get_n(const MOD_RNX* module) { return module->n; }
/** @brief allocates a prepared matrix (release with delete_rnx_vmp_pmat) */
EXPORT RNX_VMP_PMAT* new_rnx_vmp_pmat(const MOD_RNX* module, // N
uint64_t nrows, uint64_t ncols) { // dimensions
return (RNX_VMP_PMAT*)spqlios_alloc(bytes_of_rnx_vmp_pmat(module, nrows, ncols));
}
EXPORT void delete_rnx_vmp_pmat(RNX_VMP_PMAT* ptr) { spqlios_free(ptr); }
//////////////// wrappers //////////////////
/** @brief sets res = a + b */
EXPORT void vec_rnx_add( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl, // a
const double* b, uint64_t b_size, uint64_t b_sl // b
) {
module->vtable.vec_rnx_add(module, res, res_size, res_sl, a, a_size, a_sl, b, b_size, b_sl);
}
/** @brief sets res = 0 */
EXPORT void vec_rnx_zero( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl // res
) {
module->vtable.vec_rnx_zero(module, res, res_size, res_sl);
}
/** @brief sets res = a */
EXPORT void vec_rnx_copy( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
) {
module->vtable.vec_rnx_copy(module, res, res_size, res_sl, a, a_size, a_sl);
}
/** @brief sets res = -a */
EXPORT void vec_rnx_negate( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
) {
module->vtable.vec_rnx_negate(module, res, res_size, res_sl, a, a_size, a_sl);
}
/** @brief sets res = a - b */
EXPORT void vec_rnx_sub( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl, // a
const double* b, uint64_t b_size, uint64_t b_sl // b
) {
module->vtable.vec_rnx_sub(module, res, res_size, res_sl, a, a_size, a_sl, b, b_size, b_sl);
}
/** @brief sets res = a . X^p */
EXPORT void vec_rnx_rotate( //
const MOD_RNX* module, // N
const int64_t p, // rotation value
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
) {
module->vtable.vec_rnx_rotate(module, p, res, res_size, res_sl, a, a_size, a_sl);
}
/** @brief sets res = a(X^p) */
EXPORT void vec_rnx_automorphism( //
const MOD_RNX* module, // N
int64_t p, // X -> X^p
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
) {
module->vtable.vec_rnx_automorphism(module, p, res, res_size, res_sl, a, a_size, a_sl);
}
EXPORT void vec_rnx_mul_xp_minus_one( //
const MOD_RNX* module, // N
const int64_t p, // rotation value
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
) {
module->vtable.vec_rnx_mul_xp_minus_one(module, p, res, res_size, res_sl, a, a_size, a_sl);
}
/** @brief number of bytes in a RNX_VMP_PMAT (for manual allocation) */
EXPORT uint64_t bytes_of_rnx_vmp_pmat(const MOD_RNX* module, // N
uint64_t nrows, uint64_t ncols) { // dimensions
return module->vtable.bytes_of_rnx_vmp_pmat(module, nrows, ncols);
}
/** @brief prepares a vmp matrix (contiguous row-major version) */
EXPORT void rnx_vmp_prepare_contiguous( //
const MOD_RNX* module, // N
RNX_VMP_PMAT* pmat, // output
const double* a, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
) {
module->vtable.rnx_vmp_prepare_contiguous(module, pmat, a, nrows, ncols, tmp_space);
}
/** @brief prepares a vmp matrix (mat[row]+col*N points to the item) */
EXPORT void rnx_vmp_prepare_dblptr( //
const MOD_RNX* module, // N
RNX_VMP_PMAT* pmat, // output
const double** a, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
) {
module->vtable.rnx_vmp_prepare_dblptr(module, pmat, a, nrows, ncols, tmp_space);
}
/** @brief prepares the ith-row of a vmp matrix with nrows and ncols */
EXPORT void rnx_vmp_prepare_row( //
const MOD_RNX* module, // N
RNX_VMP_PMAT* pmat, // output
const double* a, uint64_t row_i, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
) {
module->vtable.rnx_vmp_prepare_row(module, pmat, a, row_i, nrows, ncols, tmp_space);
}
/** @brief number of scratch bytes necessary to prepare a matrix */
EXPORT uint64_t rnx_vmp_prepare_tmp_bytes(const MOD_RNX* module) {
return module->vtable.rnx_vmp_prepare_tmp_bytes(module);
}
/** @brief applies a vmp product res = a x pmat */
EXPORT void rnx_vmp_apply_tmp_a( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
double* tmpa, uint64_t a_size, uint64_t a_sl, // a (will be overwritten)
const RNX_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space
) {
module->vtable.rnx_vmp_apply_tmp_a(module, res, res_size, res_sl, tmpa, a_size, a_sl, pmat, nrows, ncols, tmp_space);
}
EXPORT uint64_t rnx_vmp_apply_tmp_a_tmp_bytes( //
const MOD_RNX* module, // N
uint64_t res_size, // res size
uint64_t a_size, // a size
uint64_t nrows, uint64_t ncols // prep matrix dims
) {
return module->vtable.rnx_vmp_apply_tmp_a_tmp_bytes(module, res_size, a_size, nrows, ncols);
}
/** @brief minimal size of the tmp_space */
EXPORT void rnx_vmp_apply_dft_to_dft( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a_dft, uint64_t a_size, uint64_t a_sl, // a
const RNX_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space (a_size*sizeof(reim4) bytes)
) {
module->vtable.rnx_vmp_apply_dft_to_dft(module, res, res_size, res_sl, a_dft, a_size, a_sl, pmat, nrows, ncols,
tmp_space);
}
/** @brief minimal size of the tmp_space */
EXPORT uint64_t rnx_vmp_apply_dft_to_dft_tmp_bytes( //
const MOD_RNX* module, // N
uint64_t res_size, // res
uint64_t a_size, // a
uint64_t nrows, uint64_t ncols // prep matrix
) {
return module->vtable.rnx_vmp_apply_dft_to_dft_tmp_bytes(module, res_size, a_size, nrows, ncols);
}
EXPORT uint64_t bytes_of_rnx_svp_ppol(const MOD_RNX* module) { return module->vtable.bytes_of_rnx_svp_ppol(module); }
EXPORT void rnx_svp_prepare(const MOD_RNX* module, // N
RNX_SVP_PPOL* ppol, // output
const double* pol // a
) {
module->vtable.rnx_svp_prepare(module, ppol, pol);
}
EXPORT void rnx_svp_apply( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // output
const RNX_SVP_PPOL* ppol, // prepared pol
const double* a, uint64_t a_size, uint64_t a_sl // a
) {
module->vtable.rnx_svp_apply(module, // N
res, res_size, res_sl, // output
ppol, // prepared pol
a, a_size, a_sl);
}
EXPORT void rnx_approxdecomp_from_tnxdbl( //
const MOD_RNX* module, // N
const TNXDBL_APPROXDECOMP_GADGET* gadget, // output base 2^K
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a) { // a
module->vtable.rnx_approxdecomp_from_tnxdbl(module, gadget, res, res_size, res_sl, a);
}
EXPORT void vec_rnx_to_znx32( //
const MOD_RNX* module, // N
int32_t* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
) {
module->vtable.vec_rnx_to_znx32(module, res, res_size, res_sl, a, a_size, a_sl);
}
EXPORT void vec_rnx_from_znx32( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const int32_t* a, uint64_t a_size, uint64_t a_sl // a
) {
module->vtable.vec_rnx_from_znx32(module, res, res_size, res_sl, a, a_size, a_sl);
}
EXPORT void vec_rnx_to_tnx32( //
const MOD_RNX* module, // N
int32_t* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
) {
module->vtable.vec_rnx_to_tnx32(module, res, res_size, res_sl, a, a_size, a_sl);
}
EXPORT void vec_rnx_from_tnx32( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const int32_t* a, uint64_t a_size, uint64_t a_sl // a
) {
module->vtable.vec_rnx_from_tnx32(module, res, res_size, res_sl, a, a_size, a_sl);
}
EXPORT void vec_rnx_to_tnxdbl( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
) {
module->vtable.vec_rnx_to_tnxdbl(module, res, res_size, res_sl, a, a_size, a_sl);
}

59
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/vec_rnx_approxdecomp_avx.c
View File

@@ -1,59 +0,0 @@
#include <memory.h>
#include "immintrin.h"
#include "vec_rnx_arithmetic_private.h"
/** @brief sets res = gadget_decompose(a) */
EXPORT void rnx_approxdecomp_from_tnxdbl_avx( //
const MOD_RNX* module, // N
const TNXDBL_APPROXDECOMP_GADGET* gadget, // output base 2^K
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a // a
) {
const uint64_t nn = module->n;
if (nn < 4) return rnx_approxdecomp_from_tnxdbl_ref(module, gadget, res, res_size, res_sl, a);
const uint64_t ell = gadget->ell;
const __m256i k = _mm256_set1_epi64x(gadget->k);
const __m256d add_cst = _mm256_set1_pd(gadget->add_cst);
const __m256i and_mask = _mm256_set1_epi64x(gadget->and_mask);
const __m256i or_mask = _mm256_set1_epi64x(gadget->or_mask);
const __m256d sub_cst = _mm256_set1_pd(gadget->sub_cst);
const uint64_t msize = res_size <= ell ? res_size : ell;
// gadget decompose column by column
if (msize == ell) {
// this is the main scenario when msize == ell
double* const last_r = res + (msize - 1) * res_sl;
for (uint64_t j = 0; j < nn; j += 4) {
double* rr = last_r + j;
const double* aa = a + j;
__m256d t_dbl = _mm256_add_pd(_mm256_loadu_pd(aa), add_cst);
__m256i t_int = _mm256_castpd_si256(t_dbl);
do {
__m256i u_int = _mm256_or_si256(_mm256_and_si256(t_int, and_mask), or_mask);
_mm256_storeu_pd(rr, _mm256_sub_pd(_mm256_castsi256_pd(u_int), sub_cst));
t_int = _mm256_srlv_epi64(t_int, k);
rr -= res_sl;
} while (rr >= res);
}
} else if (msize > 0) {
// otherwise, if msize < ell: there is one additional rshift
const __m256i first_rsh = _mm256_set1_epi64x((ell - msize) * gadget->k);
double* const last_r = res + (msize - 1) * res_sl;
for (uint64_t j = 0; j < nn; j += 4) {
double* rr = last_r + j;
const double* aa = a + j;
__m256d t_dbl = _mm256_add_pd(_mm256_loadu_pd(aa), add_cst);
__m256i t_int = _mm256_srlv_epi64(_mm256_castpd_si256(t_dbl), first_rsh);
do {
__m256i u_int = _mm256_or_si256(_mm256_and_si256(t_int, and_mask), or_mask);
_mm256_storeu_pd(rr, _mm256_sub_pd(_mm256_castsi256_pd(u_int), sub_cst));
t_int = _mm256_srlv_epi64(t_int, k);
rr -= res_sl;
} while (rr >= res);
}
}
// zero-out the last slices (if any)
for (uint64_t i = msize; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(double));
}
}

75
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/vec_rnx_approxdecomp_ref.c
View File

@@ -1,75 +0,0 @@
#include <memory.h>
#include "vec_rnx_arithmetic_private.h"
typedef union di {
double dv;
uint64_t uv;
} di_t;
/** @brief new gadget: delete with delete_tnxdbl_approxdecomp_gadget */
EXPORT TNXDBL_APPROXDECOMP_GADGET* new_tnxdbl_approxdecomp_gadget( //
const MOD_RNX* module, // N
uint64_t k, uint64_t ell // base 2^K and size
) {
if (k * ell > 50) return spqlios_error("gadget requires a too large fp precision");
TNXDBL_APPROXDECOMP_GADGET* res = spqlios_alloc(sizeof(TNXDBL_APPROXDECOMP_GADGET));
res->k = k;
res->ell = ell;
// double add_cst; // double(3.2^(51-ell.K) + 1/2.(sum 2^(-iK)) for i=[0,ell[)
union di add_cst;
add_cst.dv = UINT64_C(3) << (51 - ell * k);
for (uint64_t i = 0; i < ell; ++i) {
add_cst.uv |= UINT64_C(1) << ((i + 1) * k - 1);
}
res->add_cst = add_cst.dv;
// uint64_t and_mask; // uint64(2^(K)-1)
res->and_mask = (UINT64_C(1) << k) - 1;
// uint64_t or_mask; // double(2^52)
union di or_mask;
or_mask.dv = (UINT64_C(1) << 52);
res->or_mask = or_mask.uv;
// double sub_cst; // double(2^52 + 2^(K-1))
res->sub_cst = ((UINT64_C(1) << 52) + (UINT64_C(1) << (k - 1)));
return res;
}
EXPORT void delete_tnxdbl_approxdecomp_gadget(TNXDBL_APPROXDECOMP_GADGET* gadget) { spqlios_free(gadget); }
/** @brief sets res = gadget_decompose(a) */
EXPORT void rnx_approxdecomp_from_tnxdbl_ref( //
const MOD_RNX* module, // N
const TNXDBL_APPROXDECOMP_GADGET* gadget, // output base 2^K
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a // a
) {
const uint64_t nn = module->n;
const uint64_t k = gadget->k;
const uint64_t ell = gadget->ell;
const double add_cst = gadget->add_cst;
const uint64_t and_mask = gadget->and_mask;
const uint64_t or_mask = gadget->or_mask;
const double sub_cst = gadget->sub_cst;
const uint64_t msize = res_size <= ell ? res_size : ell;
const uint64_t first_rsh = (ell - msize) * k;
// gadget decompose column by column
if (msize > 0) {
double* const last_r = res + (msize - 1) * res_sl;
for (uint64_t j = 0; j < nn; ++j) {
double* rr = last_r + j;
di_t t = {.dv = a[j] + add_cst};
if (msize < ell) t.uv >>= first_rsh;
do {
di_t u;
u.uv = (t.uv & and_mask) | or_mask;
*rr = u.dv - sub_cst;
t.uv >>= k;
rr -= res_sl;
} while (rr >= res);
}
}
// zero-out the last slices (if any)
for (uint64_t i = msize; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(double));
}
}

223
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/vec_rnx_arithmetic.c
View File

@@ -1,223 +0,0 @@
#include <string.h>
#include "../coeffs/coeffs_arithmetic.h"
#include "vec_rnx_arithmetic_private.h"
void rnx_add_ref(uint64_t nn, double* res, const double* a, const double* b) {
for (uint64_t i = 0; i < nn; ++i) {
res[i] = a[i] + b[i];
}
}
void rnx_sub_ref(uint64_t nn, double* res, const double* a, const double* b) {
for (uint64_t i = 0; i < nn; ++i) {
res[i] = a[i] - b[i];
}
}
void rnx_negate_ref(uint64_t nn, double* res, const double* a) {
for (uint64_t i = 0; i < nn; ++i) {
res[i] = -a[i];
}
}
/** @brief sets res = a + b */
EXPORT void vec_rnx_add_ref( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl, // a
const double* b, uint64_t b_size, uint64_t b_sl // b
) {
const uint64_t nn = module->n;
if (a_size < b_size) {
const uint64_t msize = res_size < a_size ? res_size : a_size;
const uint64_t nsize = res_size < b_size ? res_size : b_size;
for (uint64_t i = 0; i < msize; ++i) {
rnx_add_ref(nn, res + i * res_sl, a + i * a_sl, b + i * b_sl);
}
for (uint64_t i = msize; i < nsize; ++i) {
memcpy(res + i * res_sl, b + i * b_sl, nn * sizeof(double));
}
for (uint64_t i = nsize; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(double));
}
} else {
const uint64_t msize = res_size < b_size ? res_size : b_size;
const uint64_t nsize = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < msize; ++i) {
rnx_add_ref(nn, res + i * res_sl, a + i * a_sl, b + i * b_sl);
}
for (uint64_t i = msize; i < nsize; ++i) {
memcpy(res + i * res_sl, a + i * a_sl, nn * sizeof(double));
}
for (uint64_t i = nsize; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(double));
}
}
}
/** @brief sets res = 0 */
EXPORT void vec_rnx_zero_ref( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl // res
) {
const uint64_t nn = module->n;
for (uint64_t i = 0; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(double));
}
}
/** @brief sets res = a */
EXPORT void vec_rnx_copy_ref( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
) {
const uint64_t nn = module->n;
const uint64_t rot_end_idx = res_size < a_size ? res_size : a_size;
// rotate up to the smallest dimension
for (uint64_t i = 0; i < rot_end_idx; ++i) {
double* res_ptr = res + i * res_sl;
const double* a_ptr = a + i * a_sl;
memcpy(res_ptr, a_ptr, nn * sizeof(double));
}
// then extend with zeros
for (uint64_t i = rot_end_idx; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(double));
}
}
/** @brief sets res = -a */
EXPORT void vec_rnx_negate_ref( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
) {
const uint64_t nn = module->n;
const uint64_t rot_end_idx = res_size < a_size ? res_size : a_size;
// rotate up to the smallest dimension
for (uint64_t i = 0; i < rot_end_idx; ++i) {
double* res_ptr = res + i * res_sl;
const double* a_ptr = a + i * a_sl;
rnx_negate_ref(nn, res_ptr, a_ptr);
}
// then extend with zeros
for (uint64_t i = rot_end_idx; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(double));
}
}
/** @brief sets res = a - b */
EXPORT void vec_rnx_sub_ref( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl, // a
const double* b, uint64_t b_size, uint64_t b_sl // b
) {
const uint64_t nn = module->n;
if (a_size < b_size) {
const uint64_t msize = res_size < a_size ? res_size : a_size;
const uint64_t nsize = res_size < b_size ? res_size : b_size;
for (uint64_t i = 0; i < msize; ++i) {
rnx_sub_ref(nn, res + i * res_sl, a + i * a_sl, b + i * b_sl);
}
for (uint64_t i = msize; i < nsize; ++i) {
rnx_negate_ref(nn, res + i * res_sl, b + i * b_sl);
}
for (uint64_t i = nsize; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(double));
}
} else {
const uint64_t msize = res_size < b_size ? res_size : b_size;
const uint64_t nsize = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < msize; ++i) {
rnx_sub_ref(nn, res + i * res_sl, a + i * a_sl, b + i * b_sl);
}
for (uint64_t i = msize; i < nsize; ++i) {
memcpy(res + i * res_sl, a + i * a_sl, nn * sizeof(double));
}
for (uint64_t i = nsize; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(double));
}
}
}
/** @brief sets res = a . X^p */
EXPORT void vec_rnx_rotate_ref( //
const MOD_RNX* module, // N
const int64_t p, // rotation value
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
) {
const uint64_t nn = module->n;
const uint64_t rot_end_idx = res_size < a_size ? res_size : a_size;
// rotate up to the smallest dimension
for (uint64_t i = 0; i < rot_end_idx; ++i) {
double* res_ptr = res + i * res_sl;
const double* a_ptr = a + i * a_sl;
if (res_ptr == a_ptr) {
rnx_rotate_inplace_f64(nn, p, res_ptr);
} else {
rnx_rotate_f64(nn, p, res_ptr, a_ptr);
}
}
// then extend with zeros
for (uint64_t i = rot_end_idx; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(double));
}
}
/** @brief sets res = a(X^p) */
EXPORT void vec_rnx_automorphism_ref( //
const MOD_RNX* module, // N
int64_t p, // X -> X^p
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
) {
const uint64_t nn = module->n;
const uint64_t rot_end_idx = res_size < a_size ? res_size : a_size;
// rotate up to the smallest dimension
for (uint64_t i = 0; i < rot_end_idx; ++i) {
double* res_ptr = res + i * res_sl;
const double* a_ptr = a + i * a_sl;
if (res_ptr == a_ptr) {
rnx_automorphism_inplace_f64(nn, p, res_ptr);
} else {
rnx_automorphism_f64(nn, p, res_ptr, a_ptr);
}
}
// then extend with zeros
for (uint64_t i = rot_end_idx; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(double));
}
}
/** @brief sets res = a . (X^p - 1) */
EXPORT void vec_rnx_mul_xp_minus_one_ref( //
const MOD_RNX* module, // N
const int64_t p, // rotation value
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
) {
const uint64_t nn = module->n;
const uint64_t rot_end_idx = res_size < a_size ? res_size : a_size;
// rotate up to the smallest dimension
for (uint64_t i = 0; i < rot_end_idx; ++i) {
double* res_ptr = res + i * res_sl;
const double* a_ptr = a + i * a_sl;
if (res_ptr == a_ptr) {
rnx_mul_xp_minus_one_inplace_f64(nn, p, res_ptr);
} else {
rnx_mul_xp_minus_one_f64(nn, p, res_ptr, a_ptr);
}
}
// then extend with zeros
for (uint64_t i = rot_end_idx; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(double));
}
}

356
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/vec_rnx_arithmetic.h
View File

@@ -1,356 +0,0 @@
#ifndef SPQLIOS_VEC_RNX_ARITHMETIC_H
#define SPQLIOS_VEC_RNX_ARITHMETIC_H
#include <stdint.h>
#include "../commons.h"
/**
* We support the following module families:
* - FFT64:
* the overall precision should fit at all times over 52 bits.
*/
typedef enum rnx_module_type_t { FFT64 } RNX_MODULE_TYPE;
/** @brief opaque structure that describes the modules (RnX,ZnX,TnX) and the hardware */
typedef struct rnx_module_info_t MOD_RNX;
/**
* @brief obtain a module info for ring dimension N
* the module-info knows about:
* - the dimension N (or the complex dimension m=N/2)
* - any moduleuted fft or ntt items
* - the hardware (avx, arm64, x86, ...)
*/
EXPORT MOD_RNX* new_rnx_module_info(uint64_t N, RNX_MODULE_TYPE mode);
EXPORT void delete_rnx_module_info(MOD_RNX* module_info);
EXPORT uint64_t rnx_module_get_n(const MOD_RNX* module);
// basic arithmetic
/** @brief sets res = 0 */
EXPORT void vec_rnx_zero( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl // res
);
/** @brief sets res = a */
EXPORT void vec_rnx_copy( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
);
/** @brief sets res = -a */
EXPORT void vec_rnx_negate( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
);
/** @brief sets res = a + b */
EXPORT void vec_rnx_add( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl, // a
const double* b, uint64_t b_size, uint64_t b_sl // b
);
/** @brief sets res = a - b */
EXPORT void vec_rnx_sub( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl, // a
const double* b, uint64_t b_size, uint64_t b_sl // b
);
/** @brief sets res = a . X^p */
EXPORT void vec_rnx_rotate( //
const MOD_RNX* module, // N
const int64_t p, // rotation value
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
);
/** @brief sets res = a . (X^p - 1) */
EXPORT void vec_rnx_mul_xp_minus_one( //
const MOD_RNX* module, // N
const int64_t p, // rotation value
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
);
/** @brief sets res = a(X^p) */
EXPORT void vec_rnx_automorphism( //
const MOD_RNX* module, // N
int64_t p, // X -> X^p
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
);
///////////////////////////////////////////////////////////////////
// conversions //
///////////////////////////////////////////////////////////////////
EXPORT void vec_rnx_to_znx32( //
const MOD_RNX* module, // N
int32_t* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
);
EXPORT void vec_rnx_from_znx32( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const int32_t* a, uint64_t a_size, uint64_t a_sl // a
);
EXPORT void vec_rnx_to_tnx32( //
const MOD_RNX* module, // N
int32_t* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
);
EXPORT void vec_rnx_from_tnx32( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const int32_t* a, uint64_t a_size, uint64_t a_sl // a
);
EXPORT void vec_rnx_to_tnx32x2( //
const MOD_RNX* module, // N
int32_t* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
);
EXPORT void vec_rnx_from_tnx32x2( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const int32_t* a, uint64_t a_size, uint64_t a_sl // a
);
EXPORT void vec_rnx_to_tnxdbl( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
);
///////////////////////////////////////////////////////////////////
// isolated products (n.log(n), but not particularly optimized //
///////////////////////////////////////////////////////////////////
/** @brief res = a * b : small polynomial product */
EXPORT void rnx_small_single_product( //
const MOD_RNX* module, // N
double* res, // output
const double* a, // a
const double* b, // b
uint8_t* tmp); // scratch space
EXPORT uint64_t rnx_small_single_product_tmp_bytes(const MOD_RNX* module);
/** @brief res = a * b centermod 1: small polynomial product */
EXPORT void tnxdbl_small_single_product( //
const MOD_RNX* module, // N
double* torus_res, // output
const double* int_a, // a
const double* torus_b, // b
uint8_t* tmp); // scratch space
EXPORT uint64_t tnxdbl_small_single_product_tmp_bytes(const MOD_RNX* module);
/** @brief res = a * b: small polynomial product */
EXPORT void znx32_small_single_product( //
const MOD_RNX* module, // N
int32_t* int_res, // output
const int32_t* int_a, // a
const int32_t* int_b, // b
uint8_t* tmp); // scratch space
EXPORT uint64_t znx32_small_single_product_tmp_bytes(const MOD_RNX* module);
/** @brief res = a * b centermod 1: small polynomial product */
EXPORT void tnx32_small_single_product( //
const MOD_RNX* module, // N
int32_t* torus_res, // output
const int32_t* int_a, // a
const int32_t* torus_b, // b
uint8_t* tmp); // scratch space
EXPORT uint64_t tnx32_small_single_product_tmp_bytes(const MOD_RNX* module);
///////////////////////////////////////////////////////////////////
// prepared gadget decompositions (optimized) //
///////////////////////////////////////////////////////////////////
// decompose from tnx32
typedef struct tnx32_approxdecomp_gadget_t TNX32_APPROXDECOMP_GADGET;
/** @brief new gadget: delete with delete_tnx32_approxdecomp_gadget */
EXPORT TNX32_APPROXDECOMP_GADGET* new_tnx32_approxdecomp_gadget( //
const MOD_RNX* module, // N
uint64_t k, uint64_t ell // base 2^K and size
);
EXPORT void delete_tnx32_approxdecomp_gadget(const MOD_RNX* module);
/** @brief sets res = gadget_decompose(a) */
EXPORT void rnx_approxdecomp_from_tnx32( //
const MOD_RNX* module, // N
const TNX32_APPROXDECOMP_GADGET* gadget, // output base 2^K
double* res, uint64_t res_size, uint64_t res_sl, // res
const int32_t* a // a
);
// decompose from tnx32x2
typedef struct tnx32x2_approxdecomp_gadget_t TNX32X2_APPROXDECOMP_GADGET;
/** @brief new gadget: delete with delete_tnx32x2_approxdecomp_gadget */
EXPORT TNX32X2_APPROXDECOMP_GADGET* new_tnx32x2_approxdecomp_gadget(const MOD_RNX* module, uint64_t ka, uint64_t ella,
uint64_t kb, uint64_t ellb);
EXPORT void delete_tnx32x2_approxdecomp_gadget(const MOD_RNX* module);
/** @brief sets res = gadget_decompose(a) */
EXPORT void rnx_approxdecomp_from_tnx32x2( //
const MOD_RNX* module, // N
const TNX32X2_APPROXDECOMP_GADGET* gadget, // output base 2^K
double* res, uint64_t res_size, uint64_t res_sl, // res
const int32_t* a // a
);
// decompose from tnxdbl
typedef struct tnxdbl_approxdecomp_gadget_t TNXDBL_APPROXDECOMP_GADGET;
/** @brief new gadget: delete with delete_tnxdbl_approxdecomp_gadget */
EXPORT TNXDBL_APPROXDECOMP_GADGET* new_tnxdbl_approxdecomp_gadget( //
const MOD_RNX* module, // N
uint64_t k, uint64_t ell // base 2^K and size
);
EXPORT void delete_tnxdbl_approxdecomp_gadget(TNXDBL_APPROXDECOMP_GADGET* gadget);
/** @brief sets res = gadget_decompose(a) */
EXPORT void rnx_approxdecomp_from_tnxdbl( //
const MOD_RNX* module, // N
const TNXDBL_APPROXDECOMP_GADGET* gadget, // output base 2^K
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a); // a
///////////////////////////////////////////////////////////////////
// prepared scalar-vector product (optimized) //
///////////////////////////////////////////////////////////////////
/** @brief opaque type that represents a polynomial of RnX prepared for a scalar-vector product */
typedef struct rnx_svp_ppol_t RNX_SVP_PPOL;
/** @brief number of bytes in a RNX_VMP_PMAT (for manual allocation) */
EXPORT uint64_t bytes_of_rnx_svp_ppol(const MOD_RNX* module); // N
/** @brief allocates a prepared vector (release with delete_rnx_svp_ppol) */
EXPORT RNX_SVP_PPOL* new_rnx_svp_ppol(const MOD_RNX* module); // N
/** @brief frees memory for a prepared vector */
EXPORT void delete_rnx_svp_ppol(RNX_SVP_PPOL* res);
/** @brief prepares a svp polynomial */
EXPORT void rnx_svp_prepare(const MOD_RNX* module, // N
RNX_SVP_PPOL* ppol, // output
const double* pol // a
);
/** @brief apply a svp product, result = ppol * a, presented in DFT space */
EXPORT void rnx_svp_apply( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // output
const RNX_SVP_PPOL* ppol, // prepared pol
const double* a, uint64_t a_size, uint64_t a_sl // a
);
///////////////////////////////////////////////////////////////////
// prepared vector-matrix product (optimized) //
///////////////////////////////////////////////////////////////////
typedef struct rnx_vmp_pmat_t RNX_VMP_PMAT;
/** @brief number of bytes in a RNX_VMP_PMAT (for manual allocation) */
EXPORT uint64_t bytes_of_rnx_vmp_pmat(const MOD_RNX* module, // N
uint64_t nrows, uint64_t ncols); // dimensions
/** @brief allocates a prepared matrix (release with delete_rnx_vmp_pmat) */
EXPORT RNX_VMP_PMAT* new_rnx_vmp_pmat(const MOD_RNX* module, // N
uint64_t nrows, uint64_t ncols); // dimensions
EXPORT void delete_rnx_vmp_pmat(RNX_VMP_PMAT* ptr);
/** @brief prepares a vmp matrix (contiguous row-major version) */
EXPORT void rnx_vmp_prepare_contiguous( //
const MOD_RNX* module, // N
RNX_VMP_PMAT* pmat, // output
const double* a, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
);
/** @brief prepares a vmp matrix (mat[row]+col*N points to the item) */
EXPORT void rnx_vmp_prepare_dblptr( //
const MOD_RNX* module, // N
RNX_VMP_PMAT* pmat, // output
const double** a, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
);
/** @brief prepares the ith-row of a vmp matrix with nrows and ncols */
EXPORT void rnx_vmp_prepare_row( //
const MOD_RNX* module, // N
RNX_VMP_PMAT* pmat, // output
const double* a, uint64_t row_i, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
);
/** @brief number of scratch bytes necessary to prepare a matrix */
EXPORT uint64_t rnx_vmp_prepare_tmp_bytes(const MOD_RNX* module);
/** @brief applies a vmp product res = a x pmat */
EXPORT void rnx_vmp_apply_tmp_a( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
double* tmpa, uint64_t a_size, uint64_t a_sl, // a (will be overwritten)
const RNX_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space
);
EXPORT uint64_t rnx_vmp_apply_tmp_a_tmp_bytes( //
const MOD_RNX* module, // N
uint64_t res_size, // res size
uint64_t a_size, // a size
uint64_t nrows, uint64_t ncols // prep matrix dims
);
/** @brief minimal size of the tmp_space */
EXPORT void rnx_vmp_apply_dft_to_dft( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a_dft, uint64_t a_size, uint64_t a_sl, // a
const RNX_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space (a_size*sizeof(reim4) bytes)
);
/** @brief minimal size of the tmp_space */
EXPORT uint64_t rnx_vmp_apply_dft_to_dft_tmp_bytes( //
const MOD_RNX* module, // N
uint64_t res_size, // res
uint64_t a_size, // a
uint64_t nrows, uint64_t ncols // prep matrix
);
/** @brief sets res = DFT(a) */
EXPORT void vec_rnx_dft(const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
);
/** @brief sets res = iDFT(a_dft) -- idft is not normalized */
EXPORT void vec_rnx_idft(const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a_dft, uint64_t a_size, uint64_t a_sl // a
);
#endif // SPQLIOS_VEC_RNX_ARITHMETIC_H

189
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/vec_rnx_arithmetic_avx.c
View File

@@ -1,189 +0,0 @@
#include <immintrin.h>
#include <string.h>
#include "vec_rnx_arithmetic_private.h"
void rnx_add_avx(uint64_t nn, double* res, const double* a, const double* b) {
if (nn < 8) {
if (nn == 4) {
_mm256_storeu_pd(res, _mm256_add_pd(_mm256_loadu_pd(a), _mm256_loadu_pd(b)));
} else if (nn == 2) {
_mm_storeu_pd(res, _mm_add_pd(_mm_loadu_pd(a), _mm_loadu_pd(b)));
} else if (nn == 1) {
*res = *a + *b;
} else {
NOT_SUPPORTED(); // not a power of 2
}
return;
}
// general case: nn >= 8
__m256d x0, x1, x2, x3, x4, x5;
const double* aa = a;
const double* bb = b;
double* rr = res;
double* const rrend = res + nn;
do {
x0 = _mm256_loadu_pd(aa);
x1 = _mm256_loadu_pd(aa + 4);
x2 = _mm256_loadu_pd(bb);
x3 = _mm256_loadu_pd(bb + 4);
x4 = _mm256_add_pd(x0, x2);
x5 = _mm256_add_pd(x1, x3);
_mm256_storeu_pd(rr, x4);
_mm256_storeu_pd(rr + 4, x5);
aa += 8;
bb += 8;
rr += 8;
} while (rr < rrend);
}
void rnx_sub_avx(uint64_t nn, double* res, const double* a, const double* b) {
if (nn < 8) {
if (nn == 4) {
_mm256_storeu_pd(res, _mm256_sub_pd(_mm256_loadu_pd(a), _mm256_loadu_pd(b)));
} else if (nn == 2) {
_mm_storeu_pd(res, _mm_sub_pd(_mm_loadu_pd(a), _mm_loadu_pd(b)));
} else if (nn == 1) {
*res = *a - *b;
} else {
NOT_SUPPORTED(); // not a power of 2
}
return;
}
// general case: nn >= 8
__m256d x0, x1, x2, x3, x4, x5;
const double* aa = a;
const double* bb = b;
double* rr = res;
double* const rrend = res + nn;
do {
x0 = _mm256_loadu_pd(aa);
x1 = _mm256_loadu_pd(aa + 4);
x2 = _mm256_loadu_pd(bb);
x3 = _mm256_loadu_pd(bb + 4);
x4 = _mm256_sub_pd(x0, x2);
x5 = _mm256_sub_pd(x1, x3);
_mm256_storeu_pd(rr, x4);
_mm256_storeu_pd(rr + 4, x5);
aa += 8;
bb += 8;
rr += 8;
} while (rr < rrend);
}
void rnx_negate_avx(uint64_t nn, double* res, const double* b) {
if (nn < 8) {
if (nn == 4) {
_mm256_storeu_pd(res, _mm256_sub_pd(_mm256_set1_pd(0), _mm256_loadu_pd(b)));
} else if (nn == 2) {
_mm_storeu_pd(res, _mm_sub_pd(_mm_set1_pd(0), _mm_loadu_pd(b)));
} else if (nn == 1) {
*res = -*b;
} else {
NOT_SUPPORTED(); // not a power of 2
}
return;
}
// general case: nn >= 8
__m256d x2, x3, x4, x5;
const __m256d ZERO = _mm256_set1_pd(0);
const double* bb = b;
double* rr = res;
double* const rrend = res + nn;
do {
x2 = _mm256_loadu_pd(bb);
x3 = _mm256_loadu_pd(bb + 4);
x4 = _mm256_sub_pd(ZERO, x2);
x5 = _mm256_sub_pd(ZERO, x3);
_mm256_storeu_pd(rr, x4);
_mm256_storeu_pd(rr + 4, x5);
bb += 8;
rr += 8;
} while (rr < rrend);
}
/** @brief sets res = a + b */
EXPORT void vec_rnx_add_avx( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl, // a
const double* b, uint64_t b_size, uint64_t b_sl // b
) {
const uint64_t nn = module->n;
if (a_size < b_size) {
const uint64_t msize = res_size < a_size ? res_size : a_size;
const uint64_t nsize = res_size < b_size ? res_size : b_size;
for (uint64_t i = 0; i < msize; ++i) {
rnx_add_avx(nn, res + i * res_sl, a + i * a_sl, b + i * b_sl);
}
for (uint64_t i = msize; i < nsize; ++i) {
memcpy(res + i * res_sl, b + i * b_sl, nn * sizeof(double));
}
for (uint64_t i = nsize; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(double));
}
} else {
const uint64_t msize = res_size < b_size ? res_size : b_size;
const uint64_t nsize = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < msize; ++i) {
rnx_add_avx(nn, res + i * res_sl, a + i * a_sl, b + i * b_sl);
}
for (uint64_t i = msize; i < nsize; ++i) {
memcpy(res + i * res_sl, a + i * a_sl, nn * sizeof(double));
}
for (uint64_t i = nsize; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(double));
}
}
}
/** @brief sets res = -a */
EXPORT void vec_rnx_negate_avx( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
) {
const uint64_t nn = module->n;
const uint64_t msize = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < msize; ++i) {
rnx_negate_avx(nn, res + i * res_sl, a + i * a_sl);
}
for (uint64_t i = msize; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(double));
}
}
/** @brief sets res = a - b */
EXPORT void vec_rnx_sub_avx( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl, // a
const double* b, uint64_t b_size, uint64_t b_sl // b
) {
const uint64_t nn = module->n;
if (a_size < b_size) {
const uint64_t msize = res_size < a_size ? res_size : a_size;
const uint64_t nsize = res_size < b_size ? res_size : b_size;
for (uint64_t i = 0; i < msize; ++i) {
rnx_sub_avx(nn, res + i * res_sl, a + i * a_sl, b + i * b_sl);
}
for (uint64_t i = msize; i < nsize; ++i) {
rnx_negate_avx(nn, res + i * res_sl, b + i * b_sl);
}
for (uint64_t i = nsize; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(double));
}
} else {
const uint64_t msize = res_size < b_size ? res_size : b_size;
const uint64_t nsize = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < msize; ++i) {
rnx_sub_avx(nn, res + i * res_sl, a + i * a_sl, b + i * b_sl);
}
for (uint64_t i = msize; i < nsize; ++i) {
memcpy(res + i * res_sl, a + i * a_sl, nn * sizeof(double));
}
for (uint64_t i = nsize; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(double));
}
}
}

92
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/vec_rnx_arithmetic_plugin.h
View File

@@ -1,92 +0,0 @@
#ifndef SPQLIOS_VEC_RNX_ARITHMETIC_PLUGIN_H
#define SPQLIOS_VEC_RNX_ARITHMETIC_PLUGIN_H
#include "vec_rnx_arithmetic.h"
typedef typeof(vec_rnx_zero) VEC_RNX_ZERO_F;
typedef typeof(vec_rnx_copy) VEC_RNX_COPY_F;
typedef typeof(vec_rnx_negate) VEC_RNX_NEGATE_F;
typedef typeof(vec_rnx_add) VEC_RNX_ADD_F;
typedef typeof(vec_rnx_sub) VEC_RNX_SUB_F;
typedef typeof(vec_rnx_rotate) VEC_RNX_ROTATE_F;
typedef typeof(vec_rnx_mul_xp_minus_one) VEC_RNX_MUL_XP_MINUS_ONE_F;
typedef typeof(vec_rnx_automorphism) VEC_RNX_AUTOMORPHISM_F;
typedef typeof(vec_rnx_to_znx32) VEC_RNX_TO_ZNX32_F;
typedef typeof(vec_rnx_from_znx32) VEC_RNX_FROM_ZNX32_F;
typedef typeof(vec_rnx_to_tnx32) VEC_RNX_TO_TNX32_F;
typedef typeof(vec_rnx_from_tnx32) VEC_RNX_FROM_TNX32_F;
typedef typeof(vec_rnx_to_tnx32x2) VEC_RNX_TO_TNX32X2_F;
typedef typeof(vec_rnx_from_tnx32x2) VEC_RNX_FROM_TNX32X2_F;
typedef typeof(vec_rnx_to_tnxdbl) VEC_RNX_TO_TNXDBL_F;
// typedef typeof(vec_rnx_from_tnxdbl) VEC_RNX_FROM_TNXDBL_F;
typedef typeof(rnx_small_single_product) RNX_SMALL_SINGLE_PRODUCT_F;
typedef typeof(rnx_small_single_product_tmp_bytes) RNX_SMALL_SINGLE_PRODUCT_TMP_BYTES_F;
typedef typeof(tnxdbl_small_single_product) TNXDBL_SMALL_SINGLE_PRODUCT_F;
typedef typeof(tnxdbl_small_single_product_tmp_bytes) TNXDBL_SMALL_SINGLE_PRODUCT_TMP_BYTES_F;
typedef typeof(znx32_small_single_product) ZNX32_SMALL_SINGLE_PRODUCT_F;
typedef typeof(znx32_small_single_product_tmp_bytes) ZNX32_SMALL_SINGLE_PRODUCT_TMP_BYTES_F;
typedef typeof(tnx32_small_single_product) TNX32_SMALL_SINGLE_PRODUCT_F;
typedef typeof(tnx32_small_single_product_tmp_bytes) TNX32_SMALL_SINGLE_PRODUCT_TMP_BYTES_F;
typedef typeof(rnx_approxdecomp_from_tnx32) RNX_APPROXDECOMP_FROM_TNX32_F;
typedef typeof(rnx_approxdecomp_from_tnx32x2) RNX_APPROXDECOMP_FROM_TNX32X2_F;
typedef typeof(rnx_approxdecomp_from_tnxdbl) RNX_APPROXDECOMP_FROM_TNXDBL_F;
typedef typeof(bytes_of_rnx_svp_ppol) BYTES_OF_RNX_SVP_PPOL_F;
typedef typeof(rnx_svp_prepare) RNX_SVP_PREPARE_F;
typedef typeof(rnx_svp_apply) RNX_SVP_APPLY_F;
typedef typeof(bytes_of_rnx_vmp_pmat) BYTES_OF_RNX_VMP_PMAT_F;
typedef typeof(rnx_vmp_prepare_contiguous) RNX_VMP_PREPARE_CONTIGUOUS_F;
typedef typeof(rnx_vmp_prepare_dblptr) RNX_VMP_PREPARE_DBLPTR_F;
typedef typeof(rnx_vmp_prepare_row) RNX_VMP_PREPARE_ROW_F;
typedef typeof(rnx_vmp_prepare_tmp_bytes) RNX_VMP_PREPARE_TMP_BYTES_F;
typedef typeof(rnx_vmp_apply_tmp_a) RNX_VMP_APPLY_TMP_A_F;
typedef typeof(rnx_vmp_apply_tmp_a_tmp_bytes) RNX_VMP_APPLY_TMP_A_TMP_BYTES_F;
typedef typeof(rnx_vmp_apply_dft_to_dft) RNX_VMP_APPLY_DFT_TO_DFT_F;
typedef typeof(rnx_vmp_apply_dft_to_dft_tmp_bytes) RNX_VMP_APPLY_DFT_TO_DFT_TMP_BYTES_F;
typedef typeof(vec_rnx_dft) VEC_RNX_DFT_F;
typedef typeof(vec_rnx_idft) VEC_RNX_IDFT_F;
typedef struct rnx_module_vtable_t RNX_MODULE_VTABLE;
struct rnx_module_vtable_t {
VEC_RNX_ZERO_F* vec_rnx_zero;
VEC_RNX_COPY_F* vec_rnx_copy;
VEC_RNX_NEGATE_F* vec_rnx_negate;
VEC_RNX_ADD_F* vec_rnx_add;
VEC_RNX_SUB_F* vec_rnx_sub;
VEC_RNX_ROTATE_F* vec_rnx_rotate;
VEC_RNX_MUL_XP_MINUS_ONE_F* vec_rnx_mul_xp_minus_one;
VEC_RNX_AUTOMORPHISM_F* vec_rnx_automorphism;
VEC_RNX_TO_ZNX32_F* vec_rnx_to_znx32;
VEC_RNX_FROM_ZNX32_F* vec_rnx_from_znx32;
VEC_RNX_TO_TNX32_F* vec_rnx_to_tnx32;
VEC_RNX_FROM_TNX32_F* vec_rnx_from_tnx32;
VEC_RNX_TO_TNX32X2_F* vec_rnx_to_tnx32x2;
VEC_RNX_FROM_TNX32X2_F* vec_rnx_from_tnx32x2;
VEC_RNX_TO_TNXDBL_F* vec_rnx_to_tnxdbl;
RNX_SMALL_SINGLE_PRODUCT_F* rnx_small_single_product;
RNX_SMALL_SINGLE_PRODUCT_TMP_BYTES_F* rnx_small_single_product_tmp_bytes;
TNXDBL_SMALL_SINGLE_PRODUCT_F* tnxdbl_small_single_product;
TNXDBL_SMALL_SINGLE_PRODUCT_TMP_BYTES_F* tnxdbl_small_single_product_tmp_bytes;
ZNX32_SMALL_SINGLE_PRODUCT_F* znx32_small_single_product;
ZNX32_SMALL_SINGLE_PRODUCT_TMP_BYTES_F* znx32_small_single_product_tmp_bytes;
TNX32_SMALL_SINGLE_PRODUCT_F* tnx32_small_single_product;
TNX32_SMALL_SINGLE_PRODUCT_TMP_BYTES_F* tnx32_small_single_product_tmp_bytes;
RNX_APPROXDECOMP_FROM_TNX32_F* rnx_approxdecomp_from_tnx32;
RNX_APPROXDECOMP_FROM_TNX32X2_F* rnx_approxdecomp_from_tnx32x2;
RNX_APPROXDECOMP_FROM_TNXDBL_F* rnx_approxdecomp_from_tnxdbl;
BYTES_OF_RNX_SVP_PPOL_F* bytes_of_rnx_svp_ppol;
RNX_SVP_PREPARE_F* rnx_svp_prepare;
RNX_SVP_APPLY_F* rnx_svp_apply;
BYTES_OF_RNX_VMP_PMAT_F* bytes_of_rnx_vmp_pmat;
RNX_VMP_PREPARE_CONTIGUOUS_F* rnx_vmp_prepare_contiguous;
RNX_VMP_PREPARE_DBLPTR_F* rnx_vmp_prepare_dblptr;
RNX_VMP_PREPARE_ROW_F* rnx_vmp_prepare_row;
RNX_VMP_PREPARE_TMP_BYTES_F* rnx_vmp_prepare_tmp_bytes;
RNX_VMP_APPLY_TMP_A_F* rnx_vmp_apply_tmp_a;
RNX_VMP_APPLY_TMP_A_TMP_BYTES_F* rnx_vmp_apply_tmp_a_tmp_bytes;
RNX_VMP_APPLY_DFT_TO_DFT_F* rnx_vmp_apply_dft_to_dft;
RNX_VMP_APPLY_DFT_TO_DFT_TMP_BYTES_F* rnx_vmp_apply_dft_to_dft_tmp_bytes;
VEC_RNX_DFT_F* vec_rnx_dft;
VEC_RNX_IDFT_F* vec_rnx_idft;
};
#endif // SPQLIOS_VEC_RNX_ARITHMETIC_PLUGIN_H

309
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/vec_rnx_arithmetic_private.h
View File

@@ -1,309 +0,0 @@
#ifndef SPQLIOS_VEC_RNX_ARITHMETIC_PRIVATE_H
#define SPQLIOS_VEC_RNX_ARITHMETIC_PRIVATE_H
#include "../commons_private.h"
#include "../reim/reim_fft.h"
#include "vec_rnx_arithmetic.h"
#include "vec_rnx_arithmetic_plugin.h"
typedef struct fft64_rnx_module_precomp_t FFT64_RNX_MODULE_PRECOMP;
struct fft64_rnx_module_precomp_t {
REIM_FFT_PRECOMP* p_fft;
REIM_IFFT_PRECOMP* p_ifft;
REIM_FFTVEC_ADD_PRECOMP* p_fftvec_add;
REIM_FFTVEC_MUL_PRECOMP* p_fftvec_mul;
REIM_FFTVEC_ADDMUL_PRECOMP* p_fftvec_addmul;
};
typedef union rnx_module_precomp_t RNX_MODULE_PRECOMP;
union rnx_module_precomp_t {
FFT64_RNX_MODULE_PRECOMP fft64;
};
void fft64_init_rnx_module_precomp(MOD_RNX* module);
void fft64_finalize_rnx_module_precomp(MOD_RNX* module);
/** @brief opaque structure that describes the modules (RnX,ZnX,TnX) and the hardware */
struct rnx_module_info_t {
uint64_t n;
uint64_t m;
RNX_MODULE_TYPE mtype;
RNX_MODULE_VTABLE vtable;
RNX_MODULE_PRECOMP precomp;
void* custom;
void (*custom_deleter)(void*);
};
void init_rnx_module_info(MOD_RNX* module, //
uint64_t, RNX_MODULE_TYPE mtype);
void finalize_rnx_module_info(MOD_RNX* module);
void fft64_init_rnx_module_vtable(MOD_RNX* module);
///////////////////////////////////////////////////////////////////
// prepared gadget decompositions (optimized) //
///////////////////////////////////////////////////////////////////
struct tnx32_approxdec_gadget_t {
uint64_t k;
uint64_t ell;
int32_t add_cst; // 1/2.(sum 2^-(i+1)K)
int32_t rshift_base; // 32 - K
int64_t and_mask; // 2^K-1
int64_t or_mask; // double(2^52)
double sub_cst; // double(2^52 + 2^(K-1))
uint8_t rshifts[8]; // 32 - (i+1).K
};
struct tnx32x2_approxdec_gadget_t {
// TODO
};
struct tnxdbl_approxdecomp_gadget_t {
uint64_t k;
uint64_t ell;
double add_cst; // double(3.2^(51-ell.K) + 1/2.(sum 2^(-iK)) for i=[0,ell[)
uint64_t and_mask; // uint64(2^(K)-1)
uint64_t or_mask; // double(2^52)
double sub_cst; // double(2^52 + 2^(K-1))
};
EXPORT void vec_rnx_add_ref( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl, // a
const double* b, uint64_t b_size, uint64_t b_sl // b
);
EXPORT void vec_rnx_add_avx( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl, // a
const double* b, uint64_t b_size, uint64_t b_sl // b
);
/** @brief sets res = 0 */
EXPORT void vec_rnx_zero_ref( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl // res
);
/** @brief sets res = a */
EXPORT void vec_rnx_copy_ref( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
);
/** @brief sets res = -a */
EXPORT void vec_rnx_negate_ref( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
);
/** @brief sets res = -a */
EXPORT void vec_rnx_negate_avx( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
);
/** @brief sets res = a - b */
EXPORT void vec_rnx_sub_ref( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl, // a
const double* b, uint64_t b_size, uint64_t b_sl // b
);
/** @brief sets res = a - b */
EXPORT void vec_rnx_sub_avx( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl, // a
const double* b, uint64_t b_size, uint64_t b_sl // b
);
/** @brief sets res = a . X^p */
EXPORT void vec_rnx_rotate_ref( //
const MOD_RNX* module, // N
const int64_t p, // rotation value
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
);
/** @brief sets res = a(X^p) */
EXPORT void vec_rnx_automorphism_ref( //
const MOD_RNX* module, // N
int64_t p, // X -> X^p
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
);
/** @brief number of bytes in a RNX_VMP_PMAT (for manual allocation) */
EXPORT uint64_t fft64_bytes_of_rnx_vmp_pmat(const MOD_RNX* module, // N
uint64_t nrows, uint64_t ncols);
EXPORT void fft64_rnx_vmp_apply_dft_to_dft_ref( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a_dft, uint64_t a_size, uint64_t a_sl, // a
const RNX_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space (a_size*sizeof(reim4) bytes)
);
EXPORT void fft64_rnx_vmp_apply_dft_to_dft_avx( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a_dft, uint64_t a_size, uint64_t a_sl, // a
const RNX_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space (a_size*sizeof(reim4) bytes)
);
EXPORT uint64_t fft64_rnx_vmp_apply_dft_to_dft_tmp_bytes_ref( //
const MOD_RNX* module, // N
uint64_t res_size, // res
uint64_t a_size, // a
uint64_t nrows, uint64_t ncols // prep matrix
);
EXPORT uint64_t fft64_rnx_vmp_apply_dft_to_dft_tmp_bytes_avx( //
const MOD_RNX* module, // N
uint64_t res_size, // res
uint64_t a_size, // a
uint64_t nrows, uint64_t ncols // prep matrix
);
EXPORT void fft64_rnx_vmp_prepare_contiguous_ref( //
const MOD_RNX* module, // N
RNX_VMP_PMAT* pmat, // output
const double* mat, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
);
EXPORT void fft64_rnx_vmp_prepare_contiguous_avx( //
const MOD_RNX* module, // N
RNX_VMP_PMAT* pmat, // output
const double* mat, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
);
EXPORT void fft64_rnx_vmp_prepare_dblptr_ref( //
const MOD_RNX* module, // N
RNX_VMP_PMAT* pmat, // output
const double** mat, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
);
EXPORT void fft64_rnx_vmp_prepare_dblptr_avx( //
const MOD_RNX* module, // N
RNX_VMP_PMAT* pmat, // output
const double** mat, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
);
EXPORT void fft64_rnx_vmp_prepare_row_ref( //
const MOD_RNX* module, // N
RNX_VMP_PMAT* pmat, // output
const double* mat, uint64_t row_i, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
);
EXPORT void fft64_rnx_vmp_prepare_row_avx( //
const MOD_RNX* module, // N
RNX_VMP_PMAT* pmat, // output
const double* mat, uint64_t row_i, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
);
EXPORT uint64_t fft64_rnx_vmp_prepare_tmp_bytes_ref(const MOD_RNX* module);
EXPORT uint64_t fft64_rnx_vmp_prepare_tmp_bytes_avx(const MOD_RNX* module);
EXPORT void fft64_rnx_vmp_apply_tmp_a_ref( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res (addr must be != a)
double* tmpa, uint64_t a_size, uint64_t a_sl, // a (will be overwritten)
const RNX_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space
);
EXPORT void fft64_rnx_vmp_apply_tmp_a_avx( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res (addr must be != a)
double* tmpa, uint64_t a_size, uint64_t a_sl, // a (will be overwritten)
const RNX_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space
);
EXPORT uint64_t fft64_rnx_vmp_apply_tmp_a_tmp_bytes_ref( //
const MOD_RNX* module, // N
uint64_t res_size, // res
uint64_t a_size, // a
uint64_t nrows, uint64_t ncols // prep matrix
);
EXPORT uint64_t fft64_rnx_vmp_apply_tmp_a_tmp_bytes_avx( //
const MOD_RNX* module, // N
uint64_t res_size, // res
uint64_t a_size, // a
uint64_t nrows, uint64_t ncols // prep matrix
);
/// gadget decompositions
/** @brief sets res = gadget_decompose(a) */
EXPORT void rnx_approxdecomp_from_tnxdbl_ref( //
const MOD_RNX* module, // N
const TNXDBL_APPROXDECOMP_GADGET* gadget, // output base 2^K
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a); // a
EXPORT void rnx_approxdecomp_from_tnxdbl_avx( //
const MOD_RNX* module, // N
const TNXDBL_APPROXDECOMP_GADGET* gadget, // output base 2^K
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a); // a
EXPORT void vec_rnx_mul_xp_minus_one_ref( //
const MOD_RNX* module, // N
const int64_t p, // rotation value
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
);
EXPORT void vec_rnx_to_znx32_ref( //
const MOD_RNX* module, // N
int32_t* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
);
EXPORT void vec_rnx_from_znx32_ref( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const int32_t* a, uint64_t a_size, uint64_t a_sl // a
);
EXPORT void vec_rnx_to_tnx32_ref( //
const MOD_RNX* module, // N
int32_t* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
);
EXPORT void vec_rnx_from_tnx32_ref( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const int32_t* a, uint64_t a_size, uint64_t a_sl // a
);
EXPORT void vec_rnx_to_tnxdbl_ref( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
);
EXPORT uint64_t fft64_bytes_of_rnx_svp_ppol(const MOD_RNX* module); // N
/** @brief prepares a svp polynomial */
EXPORT void fft64_rnx_svp_prepare_ref(const MOD_RNX* module, // N
RNX_SVP_PPOL* ppol, // output
const double* pol // a
);
/** @brief apply a svp product, result = ppol * a, presented in DFT space */
EXPORT void fft64_rnx_svp_apply_ref( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // output
const RNX_SVP_PPOL* ppol, // prepared pol
const double* a, uint64_t a_size, uint64_t a_sl // a
);
#endif // SPQLIOS_VEC_RNX_ARITHMETIC_PRIVATE_H

91
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/vec_rnx_conversions_ref.c
View File

@@ -1,91 +0,0 @@
#include <memory.h>
#include "vec_rnx_arithmetic_private.h"
#include "zn_arithmetic_private.h"
EXPORT void vec_rnx_to_znx32_ref( //
const MOD_RNX* module, // N
int32_t* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
) {
const uint64_t nn = module->n;
const uint64_t msize = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < msize; ++i) {
dbl_round_to_i32_ref(NULL, res + i * res_sl, nn, a + i * a_sl, nn);
}
for (uint64_t i = msize; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(int32_t));
}
}
EXPORT void vec_rnx_from_znx32_ref( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const int32_t* a, uint64_t a_size, uint64_t a_sl // a
) {
const uint64_t nn = module->n;
const uint64_t msize = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < msize; ++i) {
i32_to_dbl_ref(NULL, res + i * res_sl, nn, a + i * a_sl, nn);
}
for (uint64_t i = msize; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(int32_t));
}
}
EXPORT void vec_rnx_to_tnx32_ref( //
const MOD_RNX* module, // N
int32_t* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
) {
const uint64_t nn = module->n;
const uint64_t msize = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < msize; ++i) {
dbl_to_tn32_ref(NULL, res + i * res_sl, nn, a + i * a_sl, nn);
}
for (uint64_t i = msize; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(int32_t));
}
}
EXPORT void vec_rnx_from_tnx32_ref( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const int32_t* a, uint64_t a_size, uint64_t a_sl // a
) {
const uint64_t nn = module->n;
const uint64_t msize = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < msize; ++i) {
tn32_to_dbl_ref(NULL, res + i * res_sl, nn, a + i * a_sl, nn);
}
for (uint64_t i = msize; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(int32_t));
}
}
static void dbl_to_tndbl_ref( //
const void* UNUSED, // N
double* res, uint64_t res_size, // res
const double* a, uint64_t a_size // a
) {
static const double OFF_CST = INT64_C(3) << 51;
const uint64_t msize = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < msize; ++i) {
double ai = a[i] + OFF_CST;
res[i] = a[i] - (ai - OFF_CST);
}
memset(res + msize, 0, (res_size - msize) * sizeof(double));
}
EXPORT void vec_rnx_to_tnxdbl_ref( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a, uint64_t a_size, uint64_t a_sl // a
) {
const uint64_t nn = module->n;
const uint64_t msize = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < msize; ++i) {
dbl_to_tndbl_ref(NULL, res + i * res_sl, nn, a + i * a_sl, nn);
}
for (uint64_t i = msize; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(int32_t));
}
}

47
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/vec_rnx_svp_ref.c
View File

@@ -1,47 +0,0 @@
#include <string.h>
#include "../coeffs/coeffs_arithmetic.h"
#include "vec_rnx_arithmetic_private.h"
EXPORT uint64_t fft64_bytes_of_rnx_svp_ppol(const MOD_RNX* module) { return module->n * sizeof(double); }
EXPORT RNX_SVP_PPOL* new_rnx_svp_ppol(const MOD_RNX* module) { return spqlios_alloc(bytes_of_rnx_svp_ppol(module)); }
EXPORT void delete_rnx_svp_ppol(RNX_SVP_PPOL* ppol) { spqlios_free(ppol); }
/** @brief prepares a svp polynomial */
EXPORT void fft64_rnx_svp_prepare_ref(const MOD_RNX* module, // N
RNX_SVP_PPOL* ppol, // output
const double* pol // a
) {
double* const dppol = (double*)ppol;
rnx_divide_by_m_ref(module->n, module->m, dppol, pol);
reim_fft(module->precomp.fft64.p_fft, dppol);
}
EXPORT void fft64_rnx_svp_apply_ref( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // output
const RNX_SVP_PPOL* ppol, // prepared pol
const double* a, uint64_t a_size, uint64_t a_sl // a
) {
const uint64_t nn = module->n;
double* const dppol = (double*)ppol;
const uint64_t auto_end_idx = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < auto_end_idx; ++i) {
const double* a_ptr = a + i * a_sl;
double* const res_ptr = res + i * res_sl;
// copy the polynomial to res, apply fft in place, call fftvec
// _mul, apply ifft in place.
memcpy(res_ptr, a_ptr, nn * sizeof(double));
reim_fft(module->precomp.fft64.p_fft, (double*)res_ptr);
reim_fftvec_mul(module->precomp.fft64.p_fftvec_mul, res_ptr, res_ptr, dppol);
reim_ifft(module->precomp.fft64.p_ifft, res_ptr);
}
// then extend with zeros
for (uint64_t i = auto_end_idx; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(double));
}
}

254
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/vec_rnx_vmp_avx.c
View File

@@ -1,254 +0,0 @@
#include <assert.h>
#include <immintrin.h>
#include <string.h>
#include "../coeffs/coeffs_arithmetic.h"
#include "../reim/reim_fft.h"
#include "../reim4/reim4_arithmetic.h"
#include "vec_rnx_arithmetic_private.h"
/** @brief prepares a vmp matrix (contiguous row-major version) */
EXPORT void fft64_rnx_vmp_prepare_contiguous_avx( //
const MOD_RNX* module, // N
RNX_VMP_PMAT* pmat, // output
const double* mat, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
) {
// there is an edge case if nn < 8
const uint64_t nn = module->n;
const uint64_t m = module->m;
double* const dtmp = (double*)tmp_space;
double* const output_mat = (double*)pmat;
double* start_addr = (double*)pmat;
uint64_t offset = nrows * ncols * 8;
if (nn >= 8) {
for (uint64_t row_i = 0; row_i < nrows; row_i++) {
for (uint64_t col_i = 0; col_i < ncols; col_i++) {
rnx_divide_by_m_avx(nn, m, dtmp, mat + (row_i * ncols + col_i) * nn);
reim_fft(module->precomp.fft64.p_fft, dtmp);
if (col_i == (ncols - 1) && (ncols % 2 == 1)) {
// special case: last column out of an odd column number
start_addr = output_mat + col_i * nrows * 8 // col == ncols-1
+ row_i * 8;
} else {
// general case: columns go by pair
start_addr = output_mat + (col_i / 2) * (2 * nrows) * 8 // second: col pair index
+ row_i * 2 * 8 // third: row index
+ (col_i % 2) * 8;
}
for (uint64_t blk_i = 0; blk_i < m / 4; blk_i++) {
// extract blk from tmp and save it
reim4_extract_1blk_from_reim_avx(m, blk_i, start_addr + blk_i * offset, dtmp);
}
}
}
} else {
for (uint64_t row_i = 0; row_i < nrows; row_i++) {
for (uint64_t col_i = 0; col_i < ncols; col_i++) {
double* res = output_mat + (col_i * nrows + row_i) * nn;
rnx_divide_by_m_avx(nn, m, res, mat + (row_i * ncols + col_i) * nn);
reim_fft(module->precomp.fft64.p_fft, res);
}
}
}
}
/** @brief prepares a vmp matrix (mat[row]+col*N points to the item) */
EXPORT void fft64_rnx_vmp_prepare_dblptr_avx( //
const MOD_RNX* module, // N
RNX_VMP_PMAT* pmat, // output
const double** mat, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
) {
for (uint64_t row_i = 0; row_i < nrows; row_i++) {
fft64_rnx_vmp_prepare_row_avx(module, pmat, mat[row_i], row_i, nrows, ncols, tmp_space);
}
}
/** @brief prepares the ith-row of a vmp matrix with nrows and ncols */
EXPORT void fft64_rnx_vmp_prepare_row_avx( //
const MOD_RNX* module, // N
RNX_VMP_PMAT* pmat, // output
const double* row, uint64_t row_i, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
) {
// there is an edge case if nn < 8
const uint64_t nn = module->n;
const uint64_t m = module->m;
double* const dtmp = (double*)tmp_space;
double* const output_mat = (double*)pmat;
double* start_addr = (double*)pmat;
uint64_t offset = nrows * ncols * 8;
if (nn >= 8) {
for (uint64_t col_i = 0; col_i < ncols; col_i++) {
rnx_divide_by_m_avx(nn, m, dtmp, row + col_i * nn);
reim_fft(module->precomp.fft64.p_fft, dtmp);
if (col_i == (ncols - 1) && (ncols % 2 == 1)) {
// special case: last column out of an odd column number
start_addr = output_mat + col_i * nrows * 8 // col == ncols-1
+ row_i * 8;
} else {
// general case: columns go by pair
start_addr = output_mat + (col_i / 2) * (2 * nrows) * 8 // second: col pair index
+ row_i * 2 * 8 // third: row index
+ (col_i % 2) * 8;
}
for (uint64_t blk_i = 0; blk_i < m / 4; blk_i++) {
// extract blk from tmp and save it
reim4_extract_1blk_from_reim_avx(m, blk_i, start_addr + blk_i * offset, dtmp);
}
}
} else {
for (uint64_t col_i = 0; col_i < ncols; col_i++) {
double* res = output_mat + (col_i * nrows + row_i) * nn;
rnx_divide_by_m_avx(nn, m, res, row + col_i * nn);
reim_fft(module->precomp.fft64.p_fft, res);
}
}
}
/** @brief minimal size of the tmp_space */
EXPORT void fft64_rnx_vmp_apply_dft_to_dft_avx( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a_dft, uint64_t a_size, uint64_t a_sl, // a
const RNX_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space (a_size*sizeof(reim4) bytes)
) {
const uint64_t m = module->m;
const uint64_t nn = module->n;
double* mat2cols_output = (double*)tmp_space; // 128 bytes
double* extracted_blk = (double*)tmp_space + 16; // 64*min(nrows,a_size) bytes
double* mat_input = (double*)pmat;
const uint64_t row_max = nrows < a_size ? nrows : a_size;
const uint64_t col_max = ncols < res_size ? ncols : res_size;
if (row_max > 0 && col_max > 0) {
if (nn >= 8) {
// let's do some prefetching of the GSW key, since on some cpus,
// it helps
const uint64_t ms4 = m >> 2; // m/4
const uint64_t gsw_iter_doubles = 8 * nrows * ncols;
const uint64_t pref_doubles = 1200;
const double* gsw_pref_ptr = mat_input;
const double* const gsw_ptr_end = mat_input + ms4 * gsw_iter_doubles;
const double* gsw_pref_ptr_target = mat_input + pref_doubles;
for (; gsw_pref_ptr < gsw_pref_ptr_target; gsw_pref_ptr += 8) {
__builtin_prefetch(gsw_pref_ptr, 0, _MM_HINT_T0);
}
const double* mat_blk_start;
uint64_t blk_i;
for (blk_i = 0, mat_blk_start = mat_input; blk_i < ms4; blk_i++, mat_blk_start += gsw_iter_doubles) {
// prefetch the next iteration
if (gsw_pref_ptr_target < gsw_ptr_end) {
gsw_pref_ptr_target += gsw_iter_doubles;
if (gsw_pref_ptr_target > gsw_ptr_end) gsw_pref_ptr_target = gsw_ptr_end;
for (; gsw_pref_ptr < gsw_pref_ptr_target; gsw_pref_ptr += 8) {
__builtin_prefetch(gsw_pref_ptr, 0, _MM_HINT_T0);
}
}
reim4_extract_1blk_from_contiguous_reim_sl_avx(m, a_sl, row_max, blk_i, extracted_blk, a_dft);
// apply mat2cols
for (uint64_t col_i = 0; col_i < col_max - 1; col_i += 2) {
uint64_t col_offset = col_i * (8 * nrows);
reim4_vec_mat2cols_product_avx2(row_max, mat2cols_output, extracted_blk, mat_blk_start + col_offset);
reim4_save_1blk_to_reim_avx(m, blk_i, res + col_i * res_sl, mat2cols_output);
reim4_save_1blk_to_reim_avx(m, blk_i, res + (col_i + 1) * res_sl, mat2cols_output + 8);
}
// check if col_max is odd, then special case
if (col_max % 2 == 1) {
uint64_t last_col = col_max - 1;
uint64_t col_offset = last_col * (8 * nrows);
// the last column is alone in the pmat: vec_mat1col
if (ncols == col_max) {
reim4_vec_mat1col_product_avx2(row_max, mat2cols_output, extracted_blk, mat_blk_start + col_offset);
} else {
// the last column is part of a colpair in the pmat: vec_mat2cols and ignore the second position
reim4_vec_mat2cols_product_avx2(row_max, mat2cols_output, extracted_blk, mat_blk_start + col_offset);
}
reim4_save_1blk_to_reim_avx(m, blk_i, res + last_col * res_sl, mat2cols_output);
}
}
} else {
const double* in;
uint64_t in_sl;
if (res == a_dft) {
// it is in place: copy the input vector
in = (double*)tmp_space;
in_sl = nn;
// vec_rnx_copy(module, (double*)tmp_space, row_max, nn, a_dft, row_max, a_sl);
for (uint64_t row_i = 0; row_i < row_max; row_i++) {
memcpy((double*)tmp_space + row_i * nn, a_dft + row_i * a_sl, nn * sizeof(double));
}
} else {
// it is out of place: do the product directly
in = a_dft;
in_sl = a_sl;
}
for (uint64_t col_i = 0; col_i < col_max; col_i++) {
double* pmat_col = mat_input + col_i * nrows * nn;
{
reim_fftvec_mul(module->precomp.fft64.p_fftvec_mul, //
res + col_i * res_sl, //
in, //
pmat_col);
}
for (uint64_t row_i = 1; row_i < row_max; row_i++) {
reim_fftvec_addmul(module->precomp.fft64.p_fftvec_addmul, //
res + col_i * res_sl, //
in + row_i * in_sl, //
pmat_col + row_i * nn);
}
}
}
}
// zero out remaining bytes (if any)
for (uint64_t i = col_max; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(double));
}
}
/** @brief applies a vmp product res = a x pmat */
EXPORT void fft64_rnx_vmp_apply_tmp_a_avx( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res (addr must be != a)
double* tmpa, uint64_t a_size, uint64_t a_sl, // a (will be overwritten)
const RNX_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space
) {
const uint64_t nn = module->n;
const uint64_t rows = nrows < a_size ? nrows : a_size;
const uint64_t cols = ncols < res_size ? ncols : res_size;
// fft is done in place on the input (tmpa is destroyed)
for (uint64_t i = 0; i < rows; ++i) {
reim_fft(module->precomp.fft64.p_fft, tmpa + i * a_sl);
}
fft64_rnx_vmp_apply_dft_to_dft_avx(module, //
res, cols, res_sl, //
tmpa, rows, a_sl, //
pmat, nrows, ncols, //
tmp_space);
// ifft is done in place on the output
for (uint64_t i = 0; i < cols; ++i) {
reim_ifft(module->precomp.fft64.p_ifft, res + i * res_sl);
}
// zero out the remaining positions
for (uint64_t i = cols; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(double));
}
}

309
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/vec_rnx_vmp_ref.c
View File

@@ -1,309 +0,0 @@
#include <assert.h>
#include <string.h>
#include "../coeffs/coeffs_arithmetic.h"
#include "../reim/reim_fft.h"
#include "../reim4/reim4_arithmetic.h"
#include "vec_rnx_arithmetic_private.h"
/** @brief number of bytes in a RNX_VMP_PMAT (for manual allocation) */
EXPORT uint64_t fft64_bytes_of_rnx_vmp_pmat(const MOD_RNX* module, // N
uint64_t nrows, uint64_t ncols) { // dimensions
return nrows * ncols * module->n * sizeof(double);
}
/** @brief prepares a vmp matrix (contiguous row-major version) */
EXPORT void fft64_rnx_vmp_prepare_contiguous_ref( //
const MOD_RNX* module, // N
RNX_VMP_PMAT* pmat, // output
const double* mat, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
) {
// there is an edge case if nn < 8
const uint64_t nn = module->n;
const uint64_t m = module->m;
double* const dtmp = (double*)tmp_space;
double* const output_mat = (double*)pmat;
double* start_addr = (double*)pmat;
uint64_t offset = nrows * ncols * 8;
if (nn >= 8) {
for (uint64_t row_i = 0; row_i < nrows; row_i++) {
for (uint64_t col_i = 0; col_i < ncols; col_i++) {
rnx_divide_by_m_ref(nn, m, dtmp, mat + (row_i * ncols + col_i) * nn);
reim_fft(module->precomp.fft64.p_fft, dtmp);
if (col_i == (ncols - 1) && (ncols % 2 == 1)) {
// special case: last column out of an odd column number
start_addr = output_mat + col_i * nrows * 8 // col == ncols-1
+ row_i * 8;
} else {
// general case: columns go by pair
start_addr = output_mat + (col_i / 2) * (2 * nrows) * 8 // second: col pair index
+ row_i * 2 * 8 // third: row index
+ (col_i % 2) * 8;
}
for (uint64_t blk_i = 0; blk_i < m / 4; blk_i++) {
// extract blk from tmp and save it
reim4_extract_1blk_from_reim_ref(m, blk_i, start_addr + blk_i * offset, dtmp);
}
}
}
} else {
for (uint64_t row_i = 0; row_i < nrows; row_i++) {
for (uint64_t col_i = 0; col_i < ncols; col_i++) {
double* res = output_mat + (col_i * nrows + row_i) * nn;
rnx_divide_by_m_ref(nn, m, res, mat + (row_i * ncols + col_i) * nn);
reim_fft(module->precomp.fft64.p_fft, res);
}
}
}
}
/** @brief prepares a vmp matrix (mat[row]+col*N points to the item) */
EXPORT void fft64_rnx_vmp_prepare_dblptr_ref( //
const MOD_RNX* module, // N
RNX_VMP_PMAT* pmat, // output
const double** mat, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
) {
for (uint64_t row_i = 0; row_i < nrows; row_i++) {
fft64_rnx_vmp_prepare_row_ref(module, pmat, mat[row_i], row_i, nrows, ncols, tmp_space);
}
}
/** @brief prepares the ith-row of a vmp matrix with nrows and ncols */
EXPORT void fft64_rnx_vmp_prepare_row_ref( //
const MOD_RNX* module, // N
RNX_VMP_PMAT* pmat, // output
const double* row, uint64_t row_i, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
) {
// there is an edge case if nn < 8
const uint64_t nn = module->n;
const uint64_t m = module->m;
double* const dtmp = (double*)tmp_space;
double* const output_mat = (double*)pmat;
double* start_addr = (double*)pmat;
uint64_t offset = nrows * ncols * 8;
if (nn >= 8) {
for (uint64_t col_i = 0; col_i < ncols; col_i++) {
rnx_divide_by_m_ref(nn, m, dtmp, row + col_i * nn);
reim_fft(module->precomp.fft64.p_fft, dtmp);
if (col_i == (ncols - 1) && (ncols % 2 == 1)) {
// special case: last column out of an odd column number
start_addr = output_mat + col_i * nrows * 8 // col == ncols-1
+ row_i * 8;
} else {
// general case: columns go by pair
start_addr = output_mat + (col_i / 2) * (2 * nrows) * 8 // second: col pair index
+ row_i * 2 * 8 // third: row index
+ (col_i % 2) * 8;
}
for (uint64_t blk_i = 0; blk_i < m / 4; blk_i++) {
// extract blk from tmp and save it
reim4_extract_1blk_from_reim_ref(m, blk_i, start_addr + blk_i * offset, dtmp);
}
}
} else {
for (uint64_t col_i = 0; col_i < ncols; col_i++) {
double* res = output_mat + (col_i * nrows + row_i) * nn;
rnx_divide_by_m_ref(nn, m, res, row + col_i * nn);
reim_fft(module->precomp.fft64.p_fft, res);
}
}
}
/** @brief number of scratch bytes necessary to prepare a matrix */
EXPORT uint64_t fft64_rnx_vmp_prepare_tmp_bytes_ref(const MOD_RNX* module) {
const uint64_t nn = module->n;
return nn * sizeof(int64_t);
}
/** @brief minimal size of the tmp_space */
EXPORT void fft64_rnx_vmp_apply_dft_to_dft_ref( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res
const double* a_dft, uint64_t a_size, uint64_t a_sl, // a
const RNX_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space (a_size*sizeof(reim4) bytes)
) {
const uint64_t m = module->m;
const uint64_t nn = module->n;
double* mat2cols_output = (double*)tmp_space; // 128 bytes
double* extracted_blk = (double*)tmp_space + 16; // 64*min(nrows,a_size) bytes
double* mat_input = (double*)pmat;
const uint64_t row_max = nrows < a_size ? nrows : a_size;
const uint64_t col_max = ncols < res_size ? ncols : res_size;
if (row_max > 0 && col_max > 0) {
if (nn >= 8) {
for (uint64_t blk_i = 0; blk_i < m / 4; blk_i++) {
double* mat_blk_start = mat_input + blk_i * (8 * nrows * ncols);
reim4_extract_1blk_from_contiguous_reim_sl_ref(m, a_sl, row_max, blk_i, extracted_blk, a_dft);
// apply mat2cols
for (uint64_t col_i = 0; col_i < col_max - 1; col_i += 2) {
uint64_t col_offset = col_i * (8 * nrows);
reim4_vec_mat2cols_product_ref(row_max, mat2cols_output, extracted_blk, mat_blk_start + col_offset);
reim4_save_1blk_to_reim_ref(m, blk_i, res + col_i * res_sl, mat2cols_output);
reim4_save_1blk_to_reim_ref(m, blk_i, res + (col_i + 1) * res_sl, mat2cols_output + 8);
}
// check if col_max is odd, then special case
if (col_max % 2 == 1) {
uint64_t last_col = col_max - 1;
uint64_t col_offset = last_col * (8 * nrows);
// the last column is alone in the pmat: vec_mat1col
if (ncols == col_max) {
reim4_vec_mat1col_product_ref(row_max, mat2cols_output, extracted_blk, mat_blk_start + col_offset);
} else {
// the last column is part of a colpair in the pmat: vec_mat2cols and ignore the second position
reim4_vec_mat2cols_product_ref(row_max, mat2cols_output, extracted_blk, mat_blk_start + col_offset);
}
reim4_save_1blk_to_reim_ref(m, blk_i, res + last_col * res_sl, mat2cols_output);
}
}
} else {
const double* in;
uint64_t in_sl;
if (res == a_dft) {
// it is in place: copy the input vector
in = (double*)tmp_space;
in_sl = nn;
// vec_rnx_copy(module, (double*)tmp_space, row_max, nn, a_dft, row_max, a_sl);
for (uint64_t row_i = 0; row_i < row_max; row_i++) {
memcpy((double*)tmp_space + row_i * nn, a_dft + row_i * a_sl, nn * sizeof(double));
}
} else {
// it is out of place: do the product directly
in = a_dft;
in_sl = a_sl;
}
for (uint64_t col_i = 0; col_i < col_max; col_i++) {
double* pmat_col = mat_input + col_i * nrows * nn;
{
reim_fftvec_mul(module->precomp.fft64.p_fftvec_mul, //
res + col_i * res_sl, //
in, //
pmat_col);
}
for (uint64_t row_i = 1; row_i < row_max; row_i++) {
reim_fftvec_addmul(module->precomp.fft64.p_fftvec_addmul, //
res + col_i * res_sl, //
in + row_i * in_sl, //
pmat_col + row_i * nn);
}
}
}
}
// zero out remaining bytes (if any)
for (uint64_t i = col_max; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(double));
}
}
/** @brief applies a vmp product res = a x pmat */
EXPORT void fft64_rnx_vmp_apply_tmp_a_ref( //
const MOD_RNX* module, // N
double* res, uint64_t res_size, uint64_t res_sl, // res (addr must be != a)
double* tmpa, uint64_t a_size, uint64_t a_sl, // a (will be overwritten)
const RNX_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space
) {
const uint64_t nn = module->n;
const uint64_t rows = nrows < a_size ? nrows : a_size;
const uint64_t cols = ncols < res_size ? ncols : res_size;
// fft is done in place on the input (tmpa is destroyed)
for (uint64_t i = 0; i < rows; ++i) {
reim_fft(module->precomp.fft64.p_fft, tmpa + i * a_sl);
}
fft64_rnx_vmp_apply_dft_to_dft_ref(module, //
res, cols, res_sl, //
tmpa, rows, a_sl, //
pmat, nrows, ncols, //
tmp_space);
// ifft is done in place on the output
for (uint64_t i = 0; i < cols; ++i) {
reim_ifft(module->precomp.fft64.p_ifft, res + i * res_sl);
}
// zero out the remaining positions
for (uint64_t i = cols; i < res_size; ++i) {
memset(res + i * res_sl, 0, nn * sizeof(double));
}
}
/** @brief minimal size of the tmp_space */
EXPORT uint64_t fft64_rnx_vmp_apply_dft_to_dft_tmp_bytes_ref( //
const MOD_RNX* module, // N
uint64_t res_size, // res
uint64_t a_size, // a
uint64_t nrows, uint64_t ncols // prep matrix
) {
const uint64_t row_max = nrows < a_size ? nrows : a_size;
return (128) + (64 * row_max);
}
#ifdef __APPLE__
EXPORT uint64_t fft64_rnx_vmp_apply_tmp_a_tmp_bytes_ref( //
const MOD_RNX* module, // N
uint64_t res_size, // res
uint64_t a_size, // a
uint64_t nrows, uint64_t ncols // prep matrix
) {
return fft64_rnx_vmp_apply_dft_to_dft_tmp_bytes_ref(module, res_size, a_size, nrows, ncols);
}
#else
EXPORT uint64_t fft64_rnx_vmp_apply_tmp_a_tmp_bytes_ref( //
const MOD_RNX* module, // N
uint64_t res_size, // res
uint64_t a_size, // a
uint64_t nrows, uint64_t ncols // prep matrix
) __attribute((alias("fft64_rnx_vmp_apply_dft_to_dft_tmp_bytes_ref")));
#endif
// avx aliases that need to be defined in the same .c file
/** @brief number of scratch bytes necessary to prepare a matrix */
#ifdef __APPLE__
#pragma weak fft64_rnx_vmp_prepare_tmp_bytes_avx = fft64_rnx_vmp_prepare_tmp_bytes_ref
#else
EXPORT uint64_t fft64_rnx_vmp_prepare_tmp_bytes_avx(const MOD_RNX* module)
__attribute((alias("fft64_rnx_vmp_prepare_tmp_bytes_ref")));
#endif
/** @brief minimal size of the tmp_space */
#ifdef __APPLE__
#pragma weak fft64_rnx_vmp_apply_dft_to_dft_tmp_bytes_avx = fft64_rnx_vmp_apply_dft_to_dft_tmp_bytes_ref
#else
EXPORT uint64_t fft64_rnx_vmp_apply_dft_to_dft_tmp_bytes_avx( //
const MOD_RNX* module, // N
uint64_t res_size, // res
uint64_t a_size, // a
uint64_t nrows, uint64_t ncols // prep matrix
) __attribute((alias("fft64_rnx_vmp_apply_dft_to_dft_tmp_bytes_ref")));
#endif
#ifdef __APPLE__
#pragma weak fft64_rnx_vmp_apply_tmp_a_tmp_bytes_avx = fft64_rnx_vmp_apply_dft_to_dft_tmp_bytes_ref
#else
EXPORT uint64_t fft64_rnx_vmp_apply_tmp_a_tmp_bytes_avx( //
const MOD_RNX* module, // N
uint64_t res_size, // res
uint64_t a_size, // a
uint64_t nrows, uint64_t ncols // prep matrix
) __attribute((alias("fft64_rnx_vmp_apply_dft_to_dft_tmp_bytes_ref")));
#endif
// wrappers

369
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/vec_znx.c
View File

@@ -1,369 +0,0 @@
#include <assert.h>
#include <math.h>
#include <stdint.h>
#include <stdlib.h>
#include <string.h>
#include "../coeffs/coeffs_arithmetic.h"
#include "../q120/q120_arithmetic.h"
#include "../q120/q120_ntt.h"
#include "../reim/reim_fft_internal.h"
#include "../reim4/reim4_arithmetic.h"
#include "vec_znx_arithmetic.h"
#include "vec_znx_arithmetic_private.h"
// general function (virtual dispatch)
EXPORT void vec_znx_add(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
) {
module->func.vec_znx_add(module, // N
res, res_size, res_sl, // res
a, a_size, a_sl, // a
b, b_size, b_sl // b
);
}
EXPORT void vec_znx_sub(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
) {
module->func.vec_znx_sub(module, // N
res, res_size, res_sl, // res
a, a_size, a_sl, // a
b, b_size, b_sl // b
);
}
EXPORT void vec_znx_rotate(const MODULE* module, // N
const int64_t p, // rotation value
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
) {
module->func.vec_znx_rotate(module, // N
p, // p
res, res_size, res_sl, // res
a, a_size, a_sl // a
);
}
EXPORT void vec_znx_mul_xp_minus_one(const MODULE* module, // N
const int64_t p, // p
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
) {
module->func.vec_znx_mul_xp_minus_one(module, // N
p, // p
res, res_size, res_sl, // res
a, a_size, a_sl // a
);
}
EXPORT void vec_znx_automorphism(const MODULE* module, // N
const int64_t p, // X->X^p
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
) {
module->func.vec_znx_automorphism(module, // N
p, // p
res, res_size, res_sl, // res
a, a_size, a_sl // a
);
}
EXPORT void vec_znx_normalize_base2k(const MODULE* module, uint64_t nn, // N
uint64_t log2_base2k, // output base 2^K
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
uint8_t* tmp_space // scratch space of size >= N
) {
module->func.vec_znx_normalize_base2k(module, nn, // N
log2_base2k, // log2_base2k
res, res_size, res_sl, // res
a, a_size, a_sl, // a
tmp_space);
}
EXPORT uint64_t vec_znx_normalize_base2k_tmp_bytes(const MODULE* module, uint64_t nn // N
) {
return module->func.vec_znx_normalize_base2k_tmp_bytes(module, nn // N
);
}
// specialized function (ref)
EXPORT void vec_znx_add_ref(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
) {
const uint64_t nn = module->nn;
if (a_size <= b_size) {
const uint64_t sum_idx = res_size < a_size ? res_size : a_size;
const uint64_t copy_idx = res_size < b_size ? res_size : b_size;
// add up to the smallest dimension
for (uint64_t i = 0; i < sum_idx; ++i) {
znx_add_i64_ref(nn, res + i * res_sl, a + i * a_sl, b + i * b_sl);
}
// then copy to the largest dimension
for (uint64_t i = sum_idx; i < copy_idx; ++i) {
znx_copy_i64_ref(nn, res + i * res_sl, b + i * b_sl);
}
// then extend with zeros
for (uint64_t i = copy_idx; i < res_size; ++i) {
znx_zero_i64_ref(nn, res + i * res_sl);
}
} else {
const uint64_t sum_idx = res_size < b_size ? res_size : b_size;
const uint64_t copy_idx = res_size < a_size ? res_size : a_size;
// add up to the smallest dimension
for (uint64_t i = 0; i < sum_idx; ++i) {
znx_add_i64_ref(nn, res + i * res_sl, a + i * a_sl, b + i * b_sl);
}
// then copy to the largest dimension
for (uint64_t i = sum_idx; i < copy_idx; ++i) {
znx_copy_i64_ref(nn, res + i * res_sl, a + i * a_sl);
}
// then extend with zeros
for (uint64_t i = copy_idx; i < res_size; ++i) {
znx_zero_i64_ref(nn, res + i * res_sl);
}
}
}
EXPORT void vec_znx_sub_ref(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
) {
const uint64_t nn = module->nn;
if (a_size <= b_size) {
const uint64_t sub_idx = res_size < a_size ? res_size : a_size;
const uint64_t copy_idx = res_size < b_size ? res_size : b_size;
// subtract up to the smallest dimension
for (uint64_t i = 0; i < sub_idx; ++i) {
znx_sub_i64_ref(nn, res + i * res_sl, a + i * a_sl, b + i * b_sl);
}
// then negate to the largest dimension
for (uint64_t i = sub_idx; i < copy_idx; ++i) {
znx_negate_i64_ref(nn, res + i * res_sl, b + i * b_sl);
}
// then extend with zeros
for (uint64_t i = copy_idx; i < res_size; ++i) {
znx_zero_i64_ref(nn, res + i * res_sl);
}
} else {
const uint64_t sub_idx = res_size < b_size ? res_size : b_size;
const uint64_t copy_idx = res_size < a_size ? res_size : a_size;
// subtract up to the smallest dimension
for (uint64_t i = 0; i < sub_idx; ++i) {
znx_sub_i64_ref(nn, res + i * res_sl, a + i * a_sl, b + i * b_sl);
}
// then copy to the largest dimension
for (uint64_t i = sub_idx; i < copy_idx; ++i) {
znx_copy_i64_ref(nn, res + i * res_sl, a + i * a_sl);
}
// then extend with zeros
for (uint64_t i = copy_idx; i < res_size; ++i) {
znx_zero_i64_ref(nn, res + i * res_sl);
}
}
}
EXPORT void vec_znx_rotate_ref(const MODULE* module, // N
const int64_t p, // rotation value
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
) {
const uint64_t nn = module->nn;
const uint64_t rot_end_idx = res_size < a_size ? res_size : a_size;
// rotate up to the smallest dimension
for (uint64_t i = 0; i < rot_end_idx; ++i) {
int64_t* res_ptr = res + i * res_sl;
const int64_t* a_ptr = a + i * a_sl;
if (res_ptr == a_ptr) {
znx_rotate_inplace_i64(nn, p, res_ptr);
} else {
znx_rotate_i64(nn, p, res_ptr, a_ptr);
}
}
// then extend with zeros
for (uint64_t i = rot_end_idx; i < res_size; ++i) {
znx_zero_i64_ref(nn, res + i * res_sl);
}
}
EXPORT void vec_znx_mul_xp_minus_one_ref(const MODULE* module, // N
const int64_t p, // p
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
) {
const uint64_t nn = module->nn;
const uint64_t rot_end_idx = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < rot_end_idx; ++i) {
int64_t* res_ptr = res + i * res_sl;
const int64_t* a_ptr = a + i * a_sl;
if (res_ptr == a_ptr) {
znx_mul_xp_minus_one_inplace_i64(nn, p, res_ptr);
} else {
znx_mul_xp_minus_one_i64(nn, p, res_ptr, a_ptr);
}
}
// then extend with zeros
for (uint64_t i = rot_end_idx; i < res_size; ++i) {
znx_zero_i64_ref(nn, res + i * res_sl);
}
}
EXPORT void vec_znx_automorphism_ref(const MODULE* module, // N
const int64_t p, // X->X^p
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
) {
const uint64_t nn = module->nn;
const uint64_t auto_end_idx = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < auto_end_idx; ++i) {
int64_t* res_ptr = res + i * res_sl;
const int64_t* a_ptr = a + i * a_sl;
if (res_ptr == a_ptr) {
znx_automorphism_inplace_i64(nn, p, res_ptr);
} else {
znx_automorphism_i64(nn, p, res_ptr, a_ptr);
}
}
// then extend with zeros
for (uint64_t i = auto_end_idx; i < res_size; ++i) {
znx_zero_i64_ref(nn, res + i * res_sl);
}
}
EXPORT void vec_znx_normalize_base2k_ref(const MODULE* module, uint64_t nn, // N
uint64_t log2_base2k, // output base 2^K
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
uint8_t* tmp_space // scratch space of size >= N
) {
// use MSB limb of res for carry propagation
int64_t* cout = (int64_t*)tmp_space;
int64_t* cin = 0x0;
// propagate carry until first limb of res
int64_t i = a_size - 1;
for (; i >= res_size; --i) {
znx_normalize(nn, log2_base2k, 0x0, cout, a + i * a_sl, cin);
cin = cout;
}
// propagate carry and normalize
for (; i >= 1; --i) {
znx_normalize(nn, log2_base2k, res + i * res_sl, cout, a + i * a_sl, cin);
cin = cout;
}
// normalize last limb
znx_normalize(nn, log2_base2k, res, 0x0, a, cin);
// extend result with zeros
for (uint64_t i = a_size; i < res_size; ++i) {
znx_zero_i64_ref(nn, res + i * res_sl);
}
}
EXPORT uint64_t vec_znx_normalize_base2k_tmp_bytes_ref(const MODULE* module, uint64_t nn // N
) {
return nn * sizeof(int64_t);
}
// alias have to be defined in this unit: do not move
#ifdef __APPLE__
EXPORT uint64_t fft64_vec_znx_big_range_normalize_base2k_tmp_bytes( //
const MODULE* module, // N
uint64_t nn
) {
return vec_znx_normalize_base2k_tmp_bytes_ref(module, nn);
}
EXPORT uint64_t fft64_vec_znx_big_normalize_base2k_tmp_bytes( //
const MODULE* module, // N
uint64_t nn
) {
return vec_znx_normalize_base2k_tmp_bytes_ref(module, nn);
}
#else
EXPORT uint64_t fft64_vec_znx_big_normalize_base2k_tmp_bytes( //
const MODULE* module, // N
uint64_t nn
) __attribute((alias("vec_znx_normalize_base2k_tmp_bytes_ref")));
EXPORT uint64_t fft64_vec_znx_big_range_normalize_base2k_tmp_bytes( //
const MODULE* module, // N
uint64_t nn
) __attribute((alias("vec_znx_normalize_base2k_tmp_bytes_ref")));
#endif
/** @brief sets res = 0 */
EXPORT void vec_znx_zero(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl // res
) {
module->func.vec_znx_zero(module, res, res_size, res_sl);
}
/** @brief sets res = a */
EXPORT void vec_znx_copy(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
) {
module->func.vec_znx_copy(module, res, res_size, res_sl, a, a_size, a_sl);
}
/** @brief sets res = a */
EXPORT void vec_znx_negate(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
) {
module->func.vec_znx_negate(module, res, res_size, res_sl, a, a_size, a_sl);
}
EXPORT void vec_znx_zero_ref(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl // res
) {
uint64_t nn = module->nn;
for (uint64_t i = 0; i < res_size; ++i) {
znx_zero_i64_ref(nn, res + i * res_sl);
}
}
EXPORT void vec_znx_copy_ref(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
) {
uint64_t nn = module->nn;
uint64_t smin = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < smin; ++i) {
znx_copy_i64_ref(nn, res + i * res_sl, a + i * a_sl);
}
for (uint64_t i = smin; i < res_size; ++i) {
znx_zero_i64_ref(nn, res + i * res_sl);
}
}
EXPORT void vec_znx_negate_ref(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
) {
uint64_t nn = module->nn;
uint64_t smin = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < smin; ++i) {
znx_negate_i64_ref(nn, res + i * res_sl, a + i * a_sl);
}
for (uint64_t i = smin; i < res_size; ++i) {
znx_zero_i64_ref(nn, res + i * res_sl);
}
}

370
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/vec_znx_arithmetic.h
View File

@@ -1,370 +0,0 @@
#ifndef SPQLIOS_VEC_ZNX_ARITHMETIC_H
#define SPQLIOS_VEC_ZNX_ARITHMETIC_H
#include <stdint.h>
#include "../commons.h"
#include "../reim/reim_fft.h"
/**
* We support the following module families:
* - FFT64:
* all the polynomials should fit at all times over 52 bits.
* for FHE implementations, the recommended limb-sizes are
* between K=10 and 20, which is good for low multiplicative depths.
* - NTT120:
* all the polynomials should fit at all times over 119 bits.
* for FHE implementations, the recommended limb-sizes are
* between K=20 and 40, which is good for large multiplicative depths.
*/
typedef enum module_type_t { FFT64, NTT120 } MODULE_TYPE;
/** @brief opaque structure that describr the modules (ZnX,TnX) and the hardware */
typedef struct module_info_t MODULE;
/** @brief opaque type that represents a prepared matrix */
typedef struct vmp_pmat_t VMP_PMAT;
/** @brief opaque type that represents a vector of znx in DFT space */
typedef struct vec_znx_dft_t VEC_ZNX_DFT;
/** @brief opaque type that represents a vector of znx in large coeffs space */
typedef struct vec_znx_bigcoeff_t VEC_ZNX_BIG;
/** @brief opaque type that represents a prepared scalar vector product */
typedef struct svp_ppol_t SVP_PPOL;
/** @brief opaque type that represents a prepared left convolution vector product */
typedef struct cnv_pvec_l_t CNV_PVEC_L;
/** @brief opaque type that represents a prepared right convolution vector product */
typedef struct cnv_pvec_r_t CNV_PVEC_R;
/** @brief bytes needed for a vec_znx in DFT space */
EXPORT uint64_t bytes_of_vec_znx_dft(const MODULE* module, // N
uint64_t size);
/** @brief allocates a vec_znx in DFT space */
EXPORT VEC_ZNX_DFT* new_vec_znx_dft(const MODULE* module, // N
uint64_t size);
/** @brief frees memory from a vec_znx in DFT space */
EXPORT void delete_vec_znx_dft(VEC_ZNX_DFT* res);
/** @brief bytes needed for a vec_znx_big */
EXPORT uint64_t bytes_of_vec_znx_big(const MODULE* module, // N
uint64_t size);
/** @brief allocates a vec_znx_big */
EXPORT VEC_ZNX_BIG* new_vec_znx_big(const MODULE* module, // N
uint64_t size);
/** @brief frees memory from a vec_znx_big */
EXPORT void delete_vec_znx_big(VEC_ZNX_BIG* res);
/** @brief bytes needed for a prepared vector */
EXPORT uint64_t bytes_of_svp_ppol(const MODULE* module); // N
/** @brief allocates a prepared vector */
EXPORT SVP_PPOL* new_svp_ppol(const MODULE* module); // N
/** @brief frees memory for a prepared vector */
EXPORT void delete_svp_ppol(SVP_PPOL* res);
/** @brief bytes needed for a prepared matrix */
EXPORT uint64_t bytes_of_vmp_pmat(const MODULE* module, // N
uint64_t nrows, uint64_t ncols);
/** @brief allocates a prepared matrix */
EXPORT VMP_PMAT* new_vmp_pmat(const MODULE* module, // N
uint64_t nrows, uint64_t ncols);
/** @brief frees memory for a prepared matrix */
EXPORT void delete_vmp_pmat(VMP_PMAT* res);
/**
* @brief obtain a module info for ring dimension N
* the module-info knows about:
* - the dimension N (or the complex dimension m=N/2)
* - any moduleuted fft or ntt items
* - the hardware (avx, arm64, x86, ...)
*/
EXPORT MODULE* new_module_info(uint64_t N, MODULE_TYPE mode);
EXPORT void delete_module_info(MODULE* module_info);
EXPORT uint64_t module_get_n(const MODULE* module);
/** @brief sets res = 0 */
EXPORT void vec_znx_zero(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl // res
);
/** @brief sets res = a */
EXPORT void vec_znx_copy(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
);
/** @brief sets res = a */
EXPORT void vec_znx_negate(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
);
/** @brief sets res = a + b */
EXPORT void vec_znx_add(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
);
/** @brief sets res = a - b */
EXPORT void vec_znx_sub(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
);
/** @brief sets res = k-normalize-reduce(a) */
EXPORT void vec_znx_normalize_base2k(const MODULE* module, // MODULE
uint64_t nn, // N
uint64_t log2_base2k, // output base 2^K
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
uint8_t* tmp_space // scratch space (size >= N)
);
/** @brief returns the minimal byte length of scratch space for vec_znx_normalize_base2k */
EXPORT uint64_t vec_znx_normalize_base2k_tmp_bytes(const MODULE* module, uint64_t nn // N
);
/** @brief sets res = a . X^p */
EXPORT void vec_znx_rotate(const MODULE* module, // N
const int64_t p, // rotation value
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
);
/** @brief sets res = a * (X^{p} - 1) */
EXPORT void vec_znx_mul_xp_minus_one(const MODULE* module, // N
const int64_t p, // rotation value
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
);
/** @brief sets res = a(X^p) */
EXPORT void vec_znx_automorphism(const MODULE* module, // N
const int64_t p, // X-X^p
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
);
/** @brief sets res = 0 */
EXPORT void vec_dft_zero(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size // res
);
/** @brief sets res = a+b */
EXPORT void vec_dft_add(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const VEC_ZNX_DFT* a, uint64_t a_size, // a
const VEC_ZNX_DFT* b, uint64_t b_size // b
);
/** @brief sets res = a-b */
EXPORT void vec_dft_sub(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const VEC_ZNX_DFT* a, uint64_t a_size, // a
const VEC_ZNX_DFT* b, uint64_t b_size // b
);
/** @brief sets res = DFT(a) */
EXPORT void vec_znx_dft(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
);
/** @brief sets res = iDFT(a_dft) -- output in big coeffs space */
EXPORT void vec_znx_idft(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_DFT* a_dft, uint64_t a_size, // a
uint8_t* tmp // scratch space
);
/** @brief tmp bytes required for vec_znx_idft */
EXPORT uint64_t vec_znx_idft_tmp_bytes(const MODULE* module, uint64_t nn);
/**
* @brief sets res = iDFT(a_dft) -- output in big coeffs space
*
* @note a_dft is overwritten
*/
EXPORT void vec_znx_idft_tmp_a(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
VEC_ZNX_DFT* a_dft, uint64_t a_size // a is overwritten
);
/** @brief sets res = a+b */
EXPORT void vec_znx_big_add(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size, // a
const VEC_ZNX_BIG* b, uint64_t b_size // b
);
/** @brief sets res = a+b */
EXPORT void vec_znx_big_add_small(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
);
EXPORT void vec_znx_big_add_small2(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
);
/** @brief sets res = a-b */
EXPORT void vec_znx_big_sub(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size, // a
const VEC_ZNX_BIG* b, uint64_t b_size // b
);
EXPORT void vec_znx_big_sub_small_b(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
);
EXPORT void vec_znx_big_sub_small_a(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const VEC_ZNX_BIG* b, uint64_t b_size // b
);
EXPORT void vec_znx_big_sub_small2(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
);
/** @brief sets res = k-normalize(a) -- output in int64 coeffs space */
EXPORT void vec_znx_big_normalize_base2k(const MODULE* module, // MODULE
uint64_t nn, // N
uint64_t log2_base2k, // base-2^k
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const VEC_ZNX_BIG* a, uint64_t a_size, // a
uint8_t* tmp_space // temp space
);
/** @brief returns the minimal byte length of scratch space for vec_znx_big_normalize_base2k */
EXPORT uint64_t vec_znx_big_normalize_base2k_tmp_bytes(const MODULE* module, uint64_t nn // N
);
/** @brief sets res = k-normalize(a.subrange) -- output in int64 coeffs space */
EXPORT void vec_znx_big_range_normalize_base2k( //
const MODULE* module, // MODULE
uint64_t nn,
uint64_t log2_base2k, // base-2^k
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const VEC_ZNX_BIG* a, uint64_t a_range_begin, uint64_t a_range_xend, uint64_t a_range_step, // range
uint8_t* tmp_space // temp space
);
/** @brief returns the minimal byte length of scratch space for vec_znx_big_range_normalize_base2k */
EXPORT uint64_t vec_znx_big_range_normalize_base2k_tmp_bytes( //
const MODULE* module, uint64_t nn // N
);
/** @brief sets res = a . X^p */
EXPORT void vec_znx_big_rotate(const MODULE* module, // N
int64_t p, // rotation value
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size // a
);
/** @brief sets res = a(X^p) */
EXPORT void vec_znx_big_automorphism(const MODULE* module, // N
int64_t p, // X-X^p
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size // a
);
/** @brief apply a svp product, result = ppol * a, presented in DFT space */
EXPORT void svp_apply_dft(const MODULE* module, // N
const VEC_ZNX_DFT* res, uint64_t res_size, // output
const SVP_PPOL* ppol, // prepared pol
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
);
/** @brief apply a svp product, result = ppol * a, presented in DFT space */
EXPORT void svp_apply_dft_to_dft(const MODULE* module, // N
const VEC_ZNX_DFT* res, uint64_t res_size,
uint64_t res_cols, // output
const SVP_PPOL* ppol, // prepared pol
const VEC_ZNX_DFT* a, uint64_t a_size, uint64_t a_cols // a
);
/** @brief prepares a svp polynomial */
EXPORT void svp_prepare(const MODULE* module, // N
SVP_PPOL* ppol, // output
const int64_t* pol // a
);
/** @brief res = a * b : small integer polynomial product */
EXPORT void znx_small_single_product(const MODULE* module, // N
int64_t* res, // output
const int64_t* a, // a
const int64_t* b, // b
uint8_t* tmp);
/** @brief tmp bytes required for znx_small_single_product */
EXPORT uint64_t znx_small_single_product_tmp_bytes(const MODULE* module, uint64_t nn);
/** @brief minimal scratch space byte-size required for the vmp_prepare function */
EXPORT uint64_t vmp_prepare_tmp_bytes(const MODULE* module, uint64_t nn, // N
uint64_t nrows, uint64_t ncols);
/** @brief prepares a vmp matrix (contiguous row-major version) */
EXPORT void vmp_prepare_contiguous(const MODULE* module, // N
VMP_PMAT* pmat, // output
const int64_t* mat, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
);
/** @brief applies a vmp product (result in DFT space) adds to res inplace */
EXPORT void vmp_apply_dft_add(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols, uint64_t pmat_scale, // prep matrix
uint8_t* tmp_space // scratch space
);
/** @brief applies a vmp product (result in DFT space) */
EXPORT void vmp_apply_dft(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space
);
/** @brief minimal size of the tmp_space */
EXPORT uint64_t vmp_apply_dft_tmp_bytes(const MODULE* module, uint64_t nn, // N
uint64_t res_size, // res
uint64_t a_size, // a
uint64_t nrows, uint64_t ncols // prep matrix
);
/** @brief applies vmp product */
EXPORT void vmp_apply_dft_to_dft(const MODULE* module, // N
VEC_ZNX_DFT* res, const uint64_t res_size, // res
const VEC_ZNX_DFT* a_dft, uint64_t a_size, // a
const VMP_PMAT* pmat, const uint64_t nrows,
const uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space (a_size*sizeof(reim4) bytes)
);
/** @brief applies vmp product and adds to res inplace */
EXPORT void vmp_apply_dft_to_dft_add(const MODULE* module, // N
VEC_ZNX_DFT* res, const uint64_t res_size, // res
const VEC_ZNX_DFT* a_dft, uint64_t a_size, // a
const VMP_PMAT* pmat, const uint64_t nrows, const uint64_t ncols,
const uint64_t pmat_scale, // prep matrix
uint8_t* tmp_space // scratch space (a_size*sizeof(reim4) bytes)
);
/** @brief minimal size of the tmp_space */
EXPORT uint64_t vmp_apply_dft_to_dft_tmp_bytes(const MODULE* module, uint64_t nn, // N
uint64_t res_size, // res
uint64_t a_size, // a
uint64_t nrows, uint64_t ncols // prep matrix
);
#endif // SPQLIOS_VEC_ZNX_ARITHMETIC_H

563
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/vec_znx_arithmetic_private.h
View File

@@ -1,563 +0,0 @@
#ifndef SPQLIOS_VEC_ZNX_ARITHMETIC_PRIVATE_H
#define SPQLIOS_VEC_ZNX_ARITHMETIC_PRIVATE_H
#include "../commons_private.h"
#include "../q120/q120_ntt.h"
#include "vec_znx_arithmetic.h"
/**
* Layouts families:
*
* fft64:
* K: <= 20, N: <= 65536, ell: <= 200
* vec<ZnX> normalized: represented by int64
* vec<ZnX> large: represented by int64 (expect <=52 bits)
* vec<ZnX> DFT: represented by double (reim_fft space)
* On AVX2 inftastructure, PMAT, LCNV, RCNV use a special reim4_fft space
*
* ntt120:
* K: <= 50, N: <= 65536, ell: <= 80
* vec<ZnX> normalized: represented by int64
* vec<ZnX> large: represented by int128 (expect <=120 bits)
* vec<ZnX> DFT: represented by int64x4 (ntt120 space)
* On AVX2 inftastructure, PMAT, LCNV, RCNV use a special ntt120 space
*
* ntt104:
* K: <= 40, N: <= 65536, ell: <= 80
* vec<ZnX> normalized: represented by int64
* vec<ZnX> large: represented by int128 (expect <=120 bits)
* vec<ZnX> DFT: represented by int64x4 (ntt120 space)
* On AVX512 inftastructure, PMAT, LCNV, RCNV use a special ntt104 space
*/
struct fft64_module_info_t {
// pre-computation for reim_fft
REIM_FFT_PRECOMP* p_fft;
// pre-computation for add_fft
REIM_FFTVEC_ADD_PRECOMP* add_fft;
// pre-computation for add_fft
REIM_FFTVEC_SUB_PRECOMP* sub_fft;
// pre-computation for mul_fft
REIM_FFTVEC_MUL_PRECOMP* mul_fft;
// pre-computation for reim_from_znx6
REIM_FROM_ZNX64_PRECOMP* p_conv;
// pre-computation for reim_tp_znx6
REIM_TO_ZNX64_PRECOMP* p_reim_to_znx;
// pre-computation for reim_fft
REIM_IFFT_PRECOMP* p_ifft;
// pre-computation for reim_fftvec_addmul
REIM_FFTVEC_ADDMUL_PRECOMP* p_addmul;
};
struct q120_module_info_t {
// pre-computation for q120b to q120b ntt
q120_ntt_precomp* p_ntt;
// pre-computation for q120b to q120b intt
q120_ntt_precomp* p_intt;
};
// TODO add function types here
typedef typeof(vec_znx_zero) VEC_ZNX_ZERO_F;
typedef typeof(vec_znx_copy) VEC_ZNX_COPY_F;
typedef typeof(vec_znx_negate) VEC_ZNX_NEGATE_F;
typedef typeof(vec_znx_add) VEC_ZNX_ADD_F;
typedef typeof(vec_znx_dft) VEC_ZNX_DFT_F;
typedef typeof(vec_dft_add) VEC_DFT_ADD_F;
typedef typeof(vec_dft_sub) VEC_DFT_SUB_F;
typedef typeof(vec_znx_idft) VEC_ZNX_IDFT_F;
typedef typeof(vec_znx_idft_tmp_bytes) VEC_ZNX_IDFT_TMP_BYTES_F;
typedef typeof(vec_znx_idft_tmp_a) VEC_ZNX_IDFT_TMP_A_F;
typedef typeof(vec_znx_sub) VEC_ZNX_SUB_F;
typedef typeof(vec_znx_rotate) VEC_ZNX_ROTATE_F;
typedef typeof(vec_znx_mul_xp_minus_one) VEC_ZNX_MUL_XP_MINUS_ONE_F;
typedef typeof(vec_znx_automorphism) VEC_ZNX_AUTOMORPHISM_F;
typedef typeof(vec_znx_normalize_base2k) VEC_ZNX_NORMALIZE_BASE2K_F;
typedef typeof(vec_znx_normalize_base2k_tmp_bytes) VEC_ZNX_NORMALIZE_BASE2K_TMP_BYTES_F;
typedef typeof(vec_znx_big_normalize_base2k) VEC_ZNX_BIG_NORMALIZE_BASE2K_F;
typedef typeof(vec_znx_big_normalize_base2k_tmp_bytes) VEC_ZNX_BIG_NORMALIZE_BASE2K_TMP_BYTES_F;
typedef typeof(vec_znx_big_range_normalize_base2k) VEC_ZNX_BIG_RANGE_NORMALIZE_BASE2K_F;
typedef typeof(vec_znx_big_range_normalize_base2k_tmp_bytes) VEC_ZNX_BIG_RANGE_NORMALIZE_BASE2K_TMP_BYTES_F;
typedef typeof(vec_znx_big_add) VEC_ZNX_BIG_ADD_F;
typedef typeof(vec_znx_big_add_small) VEC_ZNX_BIG_ADD_SMALL_F;
typedef typeof(vec_znx_big_add_small2) VEC_ZNX_BIG_ADD_SMALL2_F;
typedef typeof(vec_znx_big_sub) VEC_ZNX_BIG_SUB_F;
typedef typeof(vec_znx_big_sub_small_a) VEC_ZNX_BIG_SUB_SMALL_A_F;
typedef typeof(vec_znx_big_sub_small_b) VEC_ZNX_BIG_SUB_SMALL_B_F;
typedef typeof(vec_znx_big_sub_small2) VEC_ZNX_BIG_SUB_SMALL2_F;
typedef typeof(vec_znx_big_rotate) VEC_ZNX_BIG_ROTATE_F;
typedef typeof(vec_znx_big_automorphism) VEC_ZNX_BIG_AUTOMORPHISM_F;
typedef typeof(svp_prepare) SVP_PREPARE;
typedef typeof(svp_apply_dft) SVP_APPLY_DFT_F;
typedef typeof(svp_apply_dft_to_dft) SVP_APPLY_DFT_TO_DFT_F;
typedef typeof(znx_small_single_product) ZNX_SMALL_SINGLE_PRODUCT_F;
typedef typeof(znx_small_single_product_tmp_bytes) ZNX_SMALL_SINGLE_PRODUCT_TMP_BYTES_F;
typedef typeof(vmp_prepare_contiguous) VMP_PREPARE_CONTIGUOUS_F;
typedef typeof(vmp_prepare_tmp_bytes) VMP_PREPARE_TMP_BYTES_F;
typedef typeof(vmp_apply_dft) VMP_APPLY_DFT_F;
typedef typeof(vmp_apply_dft_add) VMP_APPLY_DFT_ADD_F;
typedef typeof(vmp_apply_dft_tmp_bytes) VMP_APPLY_DFT_TMP_BYTES_F;
typedef typeof(vmp_apply_dft_to_dft) VMP_APPLY_DFT_TO_DFT_F;
typedef typeof(vmp_apply_dft_to_dft_add) VMP_APPLY_DFT_TO_DFT_ADD_F;
typedef typeof(vmp_apply_dft_to_dft_tmp_bytes) VMP_APPLY_DFT_TO_DFT_TMP_BYTES_F;
typedef typeof(bytes_of_vec_znx_dft) BYTES_OF_VEC_ZNX_DFT_F;
typedef typeof(bytes_of_vec_znx_big) BYTES_OF_VEC_ZNX_BIG_F;
typedef typeof(bytes_of_svp_ppol) BYTES_OF_SVP_PPOL_F;
typedef typeof(bytes_of_vmp_pmat) BYTES_OF_VMP_PMAT_F;
struct module_virtual_functions_t {
// TODO add functions here
VEC_ZNX_ZERO_F* vec_znx_zero;
VEC_ZNX_COPY_F* vec_znx_copy;
VEC_ZNX_NEGATE_F* vec_znx_negate;
VEC_ZNX_ADD_F* vec_znx_add;
VEC_ZNX_DFT_F* vec_znx_dft;
VEC_DFT_ADD_F* vec_dft_add;
VEC_DFT_SUB_F* vec_dft_sub;
VEC_ZNX_IDFT_F* vec_znx_idft;
VEC_ZNX_IDFT_TMP_BYTES_F* vec_znx_idft_tmp_bytes;
VEC_ZNX_IDFT_TMP_A_F* vec_znx_idft_tmp_a;
VEC_ZNX_SUB_F* vec_znx_sub;
VEC_ZNX_ROTATE_F* vec_znx_rotate;
VEC_ZNX_MUL_XP_MINUS_ONE_F* vec_znx_mul_xp_minus_one;
VEC_ZNX_AUTOMORPHISM_F* vec_znx_automorphism;
VEC_ZNX_NORMALIZE_BASE2K_F* vec_znx_normalize_base2k;
VEC_ZNX_NORMALIZE_BASE2K_TMP_BYTES_F* vec_znx_normalize_base2k_tmp_bytes;
VEC_ZNX_BIG_NORMALIZE_BASE2K_F* vec_znx_big_normalize_base2k;
VEC_ZNX_BIG_NORMALIZE_BASE2K_TMP_BYTES_F* vec_znx_big_normalize_base2k_tmp_bytes;
VEC_ZNX_BIG_RANGE_NORMALIZE_BASE2K_F* vec_znx_big_range_normalize_base2k;
VEC_ZNX_BIG_RANGE_NORMALIZE_BASE2K_TMP_BYTES_F* vec_znx_big_range_normalize_base2k_tmp_bytes;
VEC_ZNX_BIG_ADD_F* vec_znx_big_add;
VEC_ZNX_BIG_ADD_SMALL_F* vec_znx_big_add_small;
VEC_ZNX_BIG_ADD_SMALL2_F* vec_znx_big_add_small2;
VEC_ZNX_BIG_SUB_F* vec_znx_big_sub;
VEC_ZNX_BIG_SUB_SMALL_A_F* vec_znx_big_sub_small_a;
VEC_ZNX_BIG_SUB_SMALL_B_F* vec_znx_big_sub_small_b;
VEC_ZNX_BIG_SUB_SMALL2_F* vec_znx_big_sub_small2;
VEC_ZNX_BIG_ROTATE_F* vec_znx_big_rotate;
VEC_ZNX_BIG_AUTOMORPHISM_F* vec_znx_big_automorphism;
SVP_PREPARE* svp_prepare;
SVP_APPLY_DFT_F* svp_apply_dft;
SVP_APPLY_DFT_TO_DFT_F* svp_apply_dft_to_dft;
ZNX_SMALL_SINGLE_PRODUCT_F* znx_small_single_product;
ZNX_SMALL_SINGLE_PRODUCT_TMP_BYTES_F* znx_small_single_product_tmp_bytes;
VMP_PREPARE_CONTIGUOUS_F* vmp_prepare_contiguous;
VMP_PREPARE_TMP_BYTES_F* vmp_prepare_tmp_bytes;
VMP_APPLY_DFT_F* vmp_apply_dft;
VMP_APPLY_DFT_ADD_F* vmp_apply_dft_add;
VMP_APPLY_DFT_TMP_BYTES_F* vmp_apply_dft_tmp_bytes;
VMP_APPLY_DFT_TO_DFT_F* vmp_apply_dft_to_dft;
VMP_APPLY_DFT_TO_DFT_ADD_F* vmp_apply_dft_to_dft_add;
VMP_APPLY_DFT_TO_DFT_TMP_BYTES_F* vmp_apply_dft_to_dft_tmp_bytes;
BYTES_OF_VEC_ZNX_DFT_F* bytes_of_vec_znx_dft;
BYTES_OF_VEC_ZNX_BIG_F* bytes_of_vec_znx_big;
BYTES_OF_SVP_PPOL_F* bytes_of_svp_ppol;
BYTES_OF_VMP_PMAT_F* bytes_of_vmp_pmat;
};
union backend_module_info_t {
struct fft64_module_info_t fft64;
struct q120_module_info_t q120;
};
struct module_info_t {
// generic parameters
MODULE_TYPE module_type;
uint64_t nn;
uint64_t m;
// backend_dependent functions
union backend_module_info_t mod;
// virtual functions
struct module_virtual_functions_t func;
};
EXPORT uint64_t fft64_bytes_of_vec_znx_dft(const MODULE* module, // N
uint64_t size);
EXPORT uint64_t fft64_bytes_of_vec_znx_big(const MODULE* module, // N
uint64_t size);
EXPORT uint64_t fft64_bytes_of_svp_ppol(const MODULE* module); // N
EXPORT uint64_t fft64_bytes_of_vmp_pmat(const MODULE* module, // N
uint64_t nrows, uint64_t ncols);
EXPORT void vec_znx_zero_ref(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl // res
);
EXPORT void vec_znx_copy_ref(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
);
EXPORT void vec_znx_negate_ref(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
);
EXPORT void vec_znx_negate_avx(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
);
EXPORT void vec_znx_add_ref(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
);
EXPORT void vec_znx_add_avx(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
);
EXPORT void vec_znx_sub_ref(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
);
EXPORT void vec_znx_sub_avx(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
);
EXPORT void vec_znx_normalize_base2k_ref(const MODULE* module, uint64_t nn, // N
uint64_t log2_base2k, // output base 2^K
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // inp
uint8_t* tmp_space // scratch space
);
EXPORT uint64_t vec_znx_normalize_base2k_tmp_bytes_ref(const MODULE* module, uint64_t nn // N
);
EXPORT void vec_znx_rotate_ref(const MODULE* module, // N
const int64_t p, // rotation value
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
);
EXPORT void vec_znx_mul_xp_minus_one_ref(const MODULE* module, // N
const int64_t p, // rotation value
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
);
EXPORT void vec_znx_automorphism_ref(const MODULE* module, // N
const int64_t p, // X->X^p
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
);
EXPORT void vmp_prepare_ref(const MODULE* module, // N
VMP_PMAT* pmat, // output
const int64_t* mat, uint64_t nrows, uint64_t ncols // a
);
EXPORT void vmp_apply_dft_ref(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols // prep matrix
);
EXPORT void vec_dft_zero_ref(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size // res
);
EXPORT void vec_dft_add_ref(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const VEC_ZNX_DFT* a, uint64_t a_size, // a
const VEC_ZNX_DFT* b, uint64_t b_size // b
);
EXPORT void vec_dft_sub_ref(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const VEC_ZNX_DFT* a, uint64_t a_size, // a
const VEC_ZNX_DFT* b, uint64_t b_size // b
);
EXPORT void vec_dft_ref(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
);
EXPORT void vec_idft_ref(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_DFT* a_dft, uint64_t a_size);
EXPORT void vec_znx_big_normalize_ref(const MODULE* module, // MODULE
uint64_t nn, // N
uint64_t k, // base-2^k
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const VEC_ZNX_BIG* a, uint64_t a_size // a
);
/** @brief apply a svp product, result = ppol * a, presented in DFT space */
EXPORT void fft64_svp_apply_dft_ref(const MODULE* module, // N
const VEC_ZNX_DFT* res, uint64_t res_size, // output
const SVP_PPOL* ppol, // prepared pol
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
);
/** @brief apply a svp product, result = ppol * a, presented in DFT space */
EXPORT void fft64_svp_apply_dft_to_dft_ref(const MODULE* module, // N
const VEC_ZNX_DFT* res, uint64_t res_size,
uint64_t res_cols, // output
const SVP_PPOL* ppol, // prepared pol
const VEC_ZNX_DFT* a, uint64_t a_size,
uint64_t a_cols // a
);
/** @brief sets res = k-normalize(a) -- output in int64 coeffs space */
EXPORT void fft64_vec_znx_big_normalize_base2k(const MODULE* module, // MODULE
uint64_t nn, // N
uint64_t k, // base-2^k
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const VEC_ZNX_BIG* a, uint64_t a_size, // a
uint8_t* tmp_space // temp space
);
/** @brief returns the minimal byte length of scratch space for vec_znx_big_normalize_base2k */
EXPORT uint64_t fft64_vec_znx_big_normalize_base2k_tmp_bytes(const MODULE* module, uint64_t nn // N
);
/** @brief sets res = k-normalize(a.subrange) -- output in int64 coeffs space */
EXPORT void fft64_vec_znx_big_range_normalize_base2k(const MODULE* module, // MODULE
uint64_t nn,
uint64_t log2_base2k, // base-2^k
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const VEC_ZNX_BIG* a, uint64_t a_range_begin, // a
uint64_t a_range_xend, uint64_t a_range_step, // range
uint8_t* tmp_space // temp space
);
/** @brief returns the minimal byte length of scratch space for vec_znx_big_range_normalize_base2k */
EXPORT uint64_t fft64_vec_znx_big_range_normalize_base2k_tmp_bytes(const MODULE* module, uint64_t nn // N
);
EXPORT void fft64_vec_znx_dft(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
);
EXPORT void fft64_vec_dft_add(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const VEC_ZNX_DFT* a, uint64_t a_size, // a
const VEC_ZNX_DFT* b, uint64_t b_size // b
);
EXPORT void fft64_vec_dft_sub(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const VEC_ZNX_DFT* a, uint64_t a_size, // a
const VEC_ZNX_DFT* b, uint64_t b_size // b
);
EXPORT void fft64_vec_znx_idft(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_DFT* a_dft, uint64_t a_size, // a
uint8_t* tmp // scratch space
);
EXPORT uint64_t fft64_vec_znx_idft_tmp_bytes(const MODULE* module, uint64_t nn);
EXPORT void fft64_vec_znx_idft_tmp_a(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
VEC_ZNX_DFT* a_dft, uint64_t a_size // a is overwritten
);
EXPORT void ntt120_vec_znx_dft_avx(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
);
/** */
EXPORT void ntt120_vec_znx_idft_avx(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_DFT* a_dft, uint64_t a_size, // a
uint8_t* tmp // scratch space
);
EXPORT uint64_t ntt120_vec_znx_idft_tmp_bytes_avx(const MODULE* module, uint64_t nn);
EXPORT void ntt120_vec_znx_idft_tmp_a_avx(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
VEC_ZNX_DFT* a_dft, uint64_t a_size // a is overwritten
);
// big additions/subtractions
/** @brief sets res = a+b */
EXPORT void fft64_vec_znx_big_add(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size, // a
const VEC_ZNX_BIG* b, uint64_t b_size // b
);
/** @brief sets res = a+b */
EXPORT void fft64_vec_znx_big_add_small(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
);
EXPORT void fft64_vec_znx_big_add_small2(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
);
/** @brief sets res = a-b */
EXPORT void fft64_vec_znx_big_sub(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size, // a
const VEC_ZNX_BIG* b, uint64_t b_size // b
);
EXPORT void fft64_vec_znx_big_sub_small_b(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
);
EXPORT void fft64_vec_znx_big_sub_small_a(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const VEC_ZNX_BIG* b, uint64_t b_size // b
);
EXPORT void fft64_vec_znx_big_sub_small2(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
);
/** @brief sets res = a . X^p */
EXPORT void fft64_vec_znx_big_rotate(const MODULE* module, // N
int64_t p, // rotation value
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size // a
);
/** @brief sets res = a(X^p) */
EXPORT void fft64_vec_znx_big_automorphism(const MODULE* module, // N
int64_t p, // X-X^p
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size // a
);
/** @brief prepares a svp polynomial */
EXPORT void fft64_svp_prepare_ref(const MODULE* module, // N
SVP_PPOL* ppol, // output
const int64_t* pol // a
);
/** @brief res = a * b : small integer polynomial product */
EXPORT void fft64_znx_small_single_product(const MODULE* module, // N
int64_t* res, // output
const int64_t* a, // a
const int64_t* b, // b
uint8_t* tmp);
/** @brief tmp bytes required for znx_small_single_product */
EXPORT uint64_t fft64_znx_small_single_product_tmp_bytes(const MODULE* module, uint64_t nn);
/** @brief prepares a vmp matrix (contiguous row-major version) */
EXPORT void fft64_vmp_prepare_contiguous_ref(const MODULE* module, // N
VMP_PMAT* pmat, // output
const int64_t* mat, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
);
/** @brief prepares a vmp matrix (contiguous row-major version) */
EXPORT void fft64_vmp_prepare_contiguous_avx(const MODULE* module, // N
VMP_PMAT* pmat, // output
const int64_t* mat, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
);
/** @brief minimal scratch space byte-size required for the vmp_prepare function */
EXPORT uint64_t fft64_vmp_prepare_tmp_bytes(const MODULE* module, uint64_t nn, // N
uint64_t nrows, uint64_t ncols);
/** @brief applies a vmp product (result in DFT space) and adds to res inplace */
EXPORT void fft64_vmp_apply_dft_add_ref(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols,
uint64_t pmat_scale, // prep matrix
uint8_t* tmp_space // scratch space
);
/** @brief applies a vmp product (result in DFT space) */
EXPORT void fft64_vmp_apply_dft_ref(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space
);
/** @brief applies a vmp product (result in DFT space) */
EXPORT void fft64_vmp_apply_dft_avx(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space
);
/** @brief applies a vmp product (result in DFT space) and adds to res inplace*/
EXPORT void fft64_vmp_apply_dft_add_avx(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols,
uint64_t pmat_scale, // prep matrix
uint8_t* tmp_space // scratch space
);
/** @brief this inner function could be very handy */
EXPORT void fft64_vmp_apply_dft_to_dft_ref(const MODULE* module, // N
VEC_ZNX_DFT* res, const uint64_t res_size, // res
const VEC_ZNX_DFT* a_dft, uint64_t a_size, // a
const VMP_PMAT* pmat, const uint64_t nrows,
const uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space (a_size*sizeof(reim4) bytes)
);
/** @brief applies rmp product and adds to res inplace */
EXPORT void fft64_vmp_apply_dft_to_dft_add_ref(const MODULE* module, // N
VEC_ZNX_DFT* res, const uint64_t res_size, // res
const VEC_ZNX_DFT* a_dft, uint64_t a_size, // a
const VMP_PMAT* pmat, const uint64_t nrows, const uint64_t ncols,
uint64_t pmat_scale, // prep matrix
uint8_t* tmp_space // scratch space (a_size*sizeof(reim4) bytes)
);
/** @brief this inner function could be very handy */
EXPORT void fft64_vmp_apply_dft_to_dft_avx(const MODULE* module, // N
VEC_ZNX_DFT* res, const uint64_t res_size, // res
const VEC_ZNX_DFT* a_dft, uint64_t a_size, // a
const VMP_PMAT* pmat, const uint64_t nrows,
const uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space (a_size*sizeof(reim4) bytes)
);
/** @brief applies rmp product and adds to res inplace */
EXPORT void fft64_vmp_apply_dft_to_dft_add_avx(const MODULE* module, // N
VEC_ZNX_DFT* res, const uint64_t res_size, // res
const VEC_ZNX_DFT* a_dft, uint64_t a_size, // a
const VMP_PMAT* pmat, const uint64_t nrows, const uint64_t ncols,
uint64_t pmat_scale, // prep matrix
uint8_t* tmp_space // scratch space (a_size*sizeof(reim4) bytes)
);
/** @brief minimal size of the tmp_space */
EXPORT uint64_t fft64_vmp_apply_dft_tmp_bytes(const MODULE* module, uint64_t nn, // N
uint64_t res_size, // res
uint64_t a_size, // a
uint64_t nrows, uint64_t ncols // prep matrix
);
/** @brief minimal size of the tmp_space */
EXPORT uint64_t fft64_vmp_apply_dft_to_dft_tmp_bytes(const MODULE* module, uint64_t nn, // N
uint64_t res_size, // res
uint64_t a_size, // a
uint64_t nrows, uint64_t ncols // prep matrix
);
#endif // SPQLIOS_VEC_ZNX_ARITHMETIC_PRIVATE_H

103
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/vec_znx_avx.c
View File

@@ -1,103 +0,0 @@
#include <string.h>
#include "../coeffs/coeffs_arithmetic.h"
#include "../reim4/reim4_arithmetic.h"
#include "vec_znx_arithmetic_private.h"
// specialized function (ref)
// Note: these functions do not have an avx variant.
#define znx_copy_i64_avx znx_copy_i64_ref
#define znx_zero_i64_avx znx_zero_i64_ref
EXPORT void vec_znx_add_avx(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
) {
const uint64_t nn = module->nn;
if (a_size <= b_size) {
const uint64_t sum_idx = res_size < a_size ? res_size : a_size;
const uint64_t copy_idx = res_size < b_size ? res_size : b_size;
// add up to the smallest dimension
for (uint64_t i = 0; i < sum_idx; ++i) {
znx_add_i64_avx(nn, res + i * res_sl, a + i * a_sl, b + i * b_sl);
}
// then copy to the largest dimension
for (uint64_t i = sum_idx; i < copy_idx; ++i) {
znx_copy_i64_avx(nn, res + i * res_sl, b + i * b_sl);
}
// then extend with zeros
for (uint64_t i = copy_idx; i < res_size; ++i) {
znx_zero_i64_avx(nn, res + i * res_sl);
}
} else {
const uint64_t sum_idx = res_size < b_size ? res_size : b_size;
const uint64_t copy_idx = res_size < a_size ? res_size : a_size;
// add up to the smallest dimension
for (uint64_t i = 0; i < sum_idx; ++i) {
znx_add_i64_avx(nn, res + i * res_sl, a + i * a_sl, b + i * b_sl);
}
// then copy to the largest dimension
for (uint64_t i = sum_idx; i < copy_idx; ++i) {
znx_copy_i64_avx(nn, res + i * res_sl, a + i * a_sl);
}
// then extend with zeros
for (uint64_t i = copy_idx; i < res_size; ++i) {
znx_zero_i64_avx(nn, res + i * res_sl);
}
}
}
EXPORT void vec_znx_sub_avx(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
) {
const uint64_t nn = module->nn;
if (a_size <= b_size) {
const uint64_t sub_idx = res_size < a_size ? res_size : a_size;
const uint64_t copy_idx = res_size < b_size ? res_size : b_size;
// subtract up to the smallest dimension
for (uint64_t i = 0; i < sub_idx; ++i) {
znx_sub_i64_avx(nn, res + i * res_sl, a + i * a_sl, b + i * b_sl);
}
// then negate to the largest dimension
for (uint64_t i = sub_idx; i < copy_idx; ++i) {
znx_negate_i64_avx(nn, res + i * res_sl, b + i * b_sl);
}
// then extend with zeros
for (uint64_t i = copy_idx; i < res_size; ++i) {
znx_zero_i64_avx(nn, res + i * res_sl);
}
} else {
const uint64_t sub_idx = res_size < b_size ? res_size : b_size;
const uint64_t copy_idx = res_size < a_size ? res_size : a_size;
// subtract up to the smallest dimension
for (uint64_t i = 0; i < sub_idx; ++i) {
znx_sub_i64_avx(nn, res + i * res_sl, a + i * a_sl, b + i * b_sl);
}
// then copy to the largest dimension
for (uint64_t i = sub_idx; i < copy_idx; ++i) {
znx_copy_i64_avx(nn, res + i * res_sl, a + i * a_sl);
}
// then extend with zeros
for (uint64_t i = copy_idx; i < res_size; ++i) {
znx_zero_i64_avx(nn, res + i * res_sl);
}
}
}
EXPORT void vec_znx_negate_avx(const MODULE* module, // N
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
) {
uint64_t nn = module->nn;
uint64_t smin = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < smin; ++i) {
znx_negate_i64_avx(nn, res + i * res_sl, a + i * a_sl);
}
for (uint64_t i = smin; i < res_size; ++i) {
znx_zero_i64_ref(nn, res + i * res_sl);
}
}

278
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/vec_znx_big.c
View File

@@ -1,278 +0,0 @@
#include "vec_znx_arithmetic_private.h"
EXPORT uint64_t bytes_of_vec_znx_big(const MODULE* module, // N
uint64_t size) {
return module->func.bytes_of_vec_znx_big(module, size);
}
// public wrappers
/** @brief sets res = a+b */
EXPORT void vec_znx_big_add(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size, // a
const VEC_ZNX_BIG* b, uint64_t b_size // b
) {
module->func.vec_znx_big_add(module, res, res_size, a, a_size, b, b_size);
}
/** @brief sets res = a+b */
EXPORT void vec_znx_big_add_small(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
) {
module->func.vec_znx_big_add_small(module, res, res_size, a, a_size, b, b_size, b_sl);
}
EXPORT void vec_znx_big_add_small2(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
) {
module->func.vec_znx_big_add_small2(module, res, res_size, a, a_size, a_sl, b, b_size, b_sl);
}
/** @brief sets res = a-b */
EXPORT void vec_znx_big_sub(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size, // a
const VEC_ZNX_BIG* b, uint64_t b_size // b
) {
module->func.vec_znx_big_sub(module, res, res_size, a, a_size, b, b_size);
}
EXPORT void vec_znx_big_sub_small_b(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
) {
module->func.vec_znx_big_sub_small_b(module, res, res_size, a, a_size, b, b_size, b_sl);
}
EXPORT void vec_znx_big_sub_small_a(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const VEC_ZNX_BIG* b, uint64_t b_size // b
) {
module->func.vec_znx_big_sub_small_a(module, res, res_size, a, a_size, a_sl, b, b_size);
}
EXPORT void vec_znx_big_sub_small2(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
) {
module->func.vec_znx_big_sub_small2(module, res, res_size, a, a_size, a_sl, b, b_size, b_sl);
}
/** @brief sets res = a . X^p */
EXPORT void vec_znx_big_rotate(const MODULE* module, // N
int64_t p, // rotation value
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size // a
) {
module->func.vec_znx_big_rotate(module, p, res, res_size, a, a_size);
}
/** @brief sets res = a(X^p) */
EXPORT void vec_znx_big_automorphism(const MODULE* module, // N
int64_t p, // X-X^p
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size // a
) {
module->func.vec_znx_big_automorphism(module, p, res, res_size, a, a_size);
}
// private wrappers
EXPORT uint64_t fft64_bytes_of_vec_znx_big(const MODULE* module, // N
uint64_t size) {
return module->nn * size * sizeof(double);
}
EXPORT VEC_ZNX_BIG* new_vec_znx_big(const MODULE* module, // N
uint64_t size) {
return spqlios_alloc(bytes_of_vec_znx_big(module, size));
}
EXPORT void delete_vec_znx_big(VEC_ZNX_BIG* res) { spqlios_free(res); }
/** @brief sets res = a+b */
EXPORT void fft64_vec_znx_big_add(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size, // a
const VEC_ZNX_BIG* b, uint64_t b_size // b
) {
const uint64_t n = module->nn;
vec_znx_add(module, //
(int64_t*)res, res_size, n, //
(int64_t*)a, a_size, n, //
(int64_t*)b, b_size, n);
}
/** @brief sets res = a+b */
EXPORT void fft64_vec_znx_big_add_small(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
) {
const uint64_t n = module->nn;
vec_znx_add(module, //
(int64_t*)res, res_size, n, //
(int64_t*)a, a_size, n, //
b, b_size, b_sl);
}
EXPORT void fft64_vec_znx_big_add_small2(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
) {
const uint64_t n = module->nn;
vec_znx_add(module, //
(int64_t*)res, res_size, n, //
a, a_size, a_sl, //
b, b_size, b_sl);
}
/** @brief sets res = a-b */
EXPORT void fft64_vec_znx_big_sub(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size, // a
const VEC_ZNX_BIG* b, uint64_t b_size // b
) {
const uint64_t n = module->nn;
vec_znx_sub(module, //
(int64_t*)res, res_size, n, //
(int64_t*)a, a_size, n, //
(int64_t*)b, b_size, n);
}
EXPORT void fft64_vec_znx_big_sub_small_b(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
) {
const uint64_t n = module->nn;
vec_znx_sub(module, //
(int64_t*)res, res_size, n, //
(int64_t*)a, a_size, //
n, b, b_size, b_sl);
}
EXPORT void fft64_vec_znx_big_sub_small_a(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const VEC_ZNX_BIG* b, uint64_t b_size // b
) {
const uint64_t n = module->nn;
vec_znx_sub(module, //
(int64_t*)res, res_size, n, //
a, a_size, a_sl, //
(int64_t*)b, b_size, n);
}
EXPORT void fft64_vec_znx_big_sub_small2(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const int64_t* b, uint64_t b_size, uint64_t b_sl // b
) {
const uint64_t n = module->nn;
vec_znx_sub(module, //
(int64_t*)res, res_size, //
n, a, a_size, //
a_sl, b, b_size, b_sl);
}
/** @brief sets res = a . X^p */
EXPORT void fft64_vec_znx_big_rotate(const MODULE* module, // N
int64_t p, // rotation value
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size // a
) {
uint64_t nn = module->nn;
vec_znx_rotate(module, p, (int64_t*)res, res_size, nn, (int64_t*)a, a_size, nn);
}
/** @brief sets res = a(X^p) */
EXPORT void fft64_vec_znx_big_automorphism(const MODULE* module, // N
int64_t p, // X-X^p
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_BIG* a, uint64_t a_size // a
) {
uint64_t nn = module->nn;
vec_znx_automorphism(module, p, (int64_t*)res, res_size, nn, (int64_t*)a, a_size, nn);
}
EXPORT void vec_znx_big_normalize_base2k(const MODULE* module, // MODULE
uint64_t nn, // N
uint64_t k, // base-2^k
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const VEC_ZNX_BIG* a, uint64_t a_size, // a
uint8_t* tmp_space // temp space
) {
module->func.vec_znx_big_normalize_base2k(module, // MODULE
nn, // N
k, // base-2^k
res, res_size, res_sl, // res
a, a_size, // a
tmp_space);
}
EXPORT uint64_t vec_znx_big_normalize_base2k_tmp_bytes(const MODULE* module, uint64_t nn // N
) {
return module->func.vec_znx_big_normalize_base2k_tmp_bytes(module, nn // N
);
}
/** @brief sets res = k-normalize(a.subrange) -- output in int64 coeffs space */
EXPORT void vec_znx_big_range_normalize_base2k( //
const MODULE* module, // MODULE
uint64_t nn, // N
uint64_t log2_base2k, // base-2^k
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const VEC_ZNX_BIG* a, uint64_t a_range_begin, uint64_t a_range_xend, uint64_t a_range_step, // range
uint8_t* tmp_space // temp space
) {
module->func.vec_znx_big_range_normalize_base2k(module, nn, log2_base2k, res, res_size, res_sl, a, a_range_begin,
a_range_xend, a_range_step, tmp_space);
}
/** @brief returns the minimal byte length of scratch space for vec_znx_big_range_normalize_base2k */
EXPORT uint64_t vec_znx_big_range_normalize_base2k_tmp_bytes( //
const MODULE* module, // MODULE
uint64_t nn // N
) {
return module->func.vec_znx_big_range_normalize_base2k_tmp_bytes(module, nn);
}
EXPORT void fft64_vec_znx_big_normalize_base2k(const MODULE* module, // MODULE
uint64_t nn, // N
uint64_t k, // base-2^k
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const VEC_ZNX_BIG* a, uint64_t a_size, // a
uint8_t* tmp_space) {
uint64_t a_sl = nn;
module->func.vec_znx_normalize_base2k(module, // N
nn,
k, // log2_base2k
res, res_size, res_sl, // res
(int64_t*)a, a_size, a_sl, // a
tmp_space);
}
EXPORT void fft64_vec_znx_big_range_normalize_base2k( //
const MODULE* module, // MODULE
uint64_t nn, // N
uint64_t k, // base-2^k
int64_t* res, uint64_t res_size, uint64_t res_sl, // res
const VEC_ZNX_BIG* a, uint64_t a_begin, uint64_t a_end, uint64_t a_step, // a
uint8_t* tmp_space) {
// convert the range indexes to int64[] slices
const int64_t* a_st = ((int64_t*)a) + nn * a_begin;
const uint64_t a_size = (a_end + a_step - 1 - a_begin) / a_step;
const uint64_t a_sl = nn * a_step;
// forward the call
module->func.vec_znx_normalize_base2k(module, // MODULE
nn, // N
k, // log2_base2k
res, res_size, res_sl, // res
a_st, a_size, a_sl, // a
tmp_space);
}

214
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/vec_znx_dft.c
View File

@@ -1,214 +0,0 @@
#include <string.h>
#include "../q120/q120_arithmetic.h"
#include "vec_znx_arithmetic_private.h"
EXPORT void vec_znx_dft(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
) {
return module->func.vec_znx_dft(module, res, res_size, a, a_size, a_sl);
}
EXPORT void vec_dft_add(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const VEC_ZNX_DFT* a, uint64_t a_size, // a
const VEC_ZNX_DFT* b, uint64_t b_size // b
) {
return module->func.vec_dft_add(module, res, res_size, a, a_size, b, b_size);
}
EXPORT void vec_dft_sub(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const VEC_ZNX_DFT* a, uint64_t a_size, // a
const VEC_ZNX_DFT* b, uint64_t b_size // b
) {
return module->func.vec_dft_sub(module, res, res_size, a, a_size, b, b_size);
}
EXPORT void vec_znx_idft(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_DFT* a_dft, uint64_t a_size, // a
uint8_t* tmp // scratch space
) {
return module->func.vec_znx_idft(module, res, res_size, a_dft, a_size, tmp);
}
EXPORT uint64_t vec_znx_idft_tmp_bytes(const MODULE* module, uint64_t nn) { return module->func.vec_znx_idft_tmp_bytes(module, nn); }
EXPORT void vec_znx_idft_tmp_a(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
VEC_ZNX_DFT* a_dft, uint64_t a_size // a is overwritten
) {
return module->func.vec_znx_idft_tmp_a(module, res, res_size, a_dft, a_size);
}
EXPORT uint64_t bytes_of_vec_znx_dft(const MODULE* module, // N
uint64_t size) {
return module->func.bytes_of_vec_znx_dft(module, size);
}
// fft64 backend
EXPORT uint64_t fft64_bytes_of_vec_znx_dft(const MODULE* module, // N
uint64_t size) {
return module->nn * size * sizeof(double);
}
EXPORT VEC_ZNX_DFT* new_vec_znx_dft(const MODULE* module, // N
uint64_t size) {
return spqlios_alloc(bytes_of_vec_znx_dft(module, size));
}
EXPORT void delete_vec_znx_dft(VEC_ZNX_DFT* res) { spqlios_free(res); }
EXPORT void fft64_vec_znx_dft(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
) {
const uint64_t smin = res_size < a_size ? res_size : a_size;
const uint64_t nn = module->nn;
for (uint64_t i = 0; i < smin; i++) {
reim_from_znx64(module->mod.fft64.p_conv, ((double*)res) + i * nn, a + i * a_sl);
reim_fft(module->mod.fft64.p_fft, ((double*)res) + i * nn);
}
// fill up remaining part with 0's
double* const dres = (double*)res;
memset(dres + smin * nn, 0, (res_size - smin) * nn * sizeof(double));
}
EXPORT void fft64_vec_dft_add(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const VEC_ZNX_DFT* a, uint64_t a_size, // a
const VEC_ZNX_DFT* b, uint64_t b_size // b
) {
const uint64_t smin0 = a_size < b_size ? a_size : b_size;
const uint64_t smin = res_size < smin0 ? res_size : smin0;
const uint64_t nn = module->nn;
for (uint64_t i = 0; i < smin; i++) {
reim_fftvec_add(module->mod.fft64.add_fft, ((double*)res) + i * nn, ((double*)a) + i * nn, ((double*)b) + i * nn);
}
// fill remain `res` part with 0's
double* const dres = (double*)res;
memset(dres + smin * nn, 0, (res_size - smin) * nn * sizeof(double));
}
EXPORT void fft64_vec_dft_sub(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const VEC_ZNX_DFT* a, uint64_t a_size, // a
const VEC_ZNX_DFT* b, uint64_t b_size // b
) {
const uint64_t smin0 = a_size < b_size ? a_size : b_size;
const uint64_t smin = res_size < smin0 ? res_size : smin0;
const uint64_t nn = module->nn;
for (uint64_t i = 0; i < smin; i++) {
reim_fftvec_sub(module->mod.fft64.sub_fft, ((double*)res) + i * nn, ((double*)a) + i * nn, ((double*)b) + i * nn);
}
// fill remain `res` part with 0's
double* const dres = (double*)res;
memset(dres + smin * nn, 0, (res_size - smin) * nn * sizeof(double));
}
EXPORT void fft64_vec_znx_idft(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_DFT* a_dft, uint64_t a_size, // a
uint8_t* tmp // unused
) {
const uint64_t nn = module->nn;
const uint64_t smin = res_size < a_size ? res_size : a_size;
if ((double*)res != (double*)a_dft) {
memcpy(res, a_dft, smin * nn * sizeof(double));
}
for (uint64_t i = 0; i < smin; i++) {
reim_ifft(module->mod.fft64.p_ifft, ((double*)res) + i * nn);
reim_to_znx64(module->mod.fft64.p_reim_to_znx, ((int64_t*)res) + i * nn, ((int64_t*)res) + i * nn);
}
// fill up remaining part with 0's
int64_t* const dres = (int64_t*)res;
memset(dres + smin * nn, 0, (res_size - smin) * nn * sizeof(double));
}
EXPORT uint64_t fft64_vec_znx_idft_tmp_bytes(const MODULE* module, uint64_t nn) { return 0; }
EXPORT void fft64_vec_znx_idft_tmp_a(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
VEC_ZNX_DFT* a_dft, uint64_t a_size // a is overwritten
) {
const uint64_t nn = module->nn;
const uint64_t smin = res_size < a_size ? res_size : a_size;
int64_t* const tres = (int64_t*)res;
double* const ta = (double*)a_dft;
for (uint64_t i = 0; i < smin; i++) {
reim_ifft(module->mod.fft64.p_ifft, ta + i * nn);
reim_to_znx64(module->mod.fft64.p_reim_to_znx, tres + i * nn, ta + i * nn);
}
// fill up remaining part with 0's
memset(tres + smin * nn, 0, (res_size - smin) * nn * sizeof(double));
}
// ntt120 backend
EXPORT void ntt120_vec_znx_dft_avx(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl // a
) {
const uint64_t nn = module->nn;
const uint64_t smin = res_size < a_size ? res_size : a_size;
int64_t* tres = (int64_t*)res;
for (uint64_t i = 0; i < smin; i++) {
q120_b_from_znx64_simple(nn, (q120b*)(tres + i * nn * 4), a + i * a_sl);
q120_ntt_bb_avx2(module->mod.q120.p_ntt, (q120b*)(tres + i * nn * 4));
}
// fill up remaining part with 0's
memset(tres + smin * nn * 4, 0, (res_size - smin) * nn * 4 * sizeof(int64_t));
}
EXPORT void ntt120_vec_znx_idft_avx(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
const VEC_ZNX_DFT* a_dft, uint64_t a_size, // a
uint8_t* tmp) {
const uint64_t nn = module->nn;
const uint64_t smin = res_size < a_size ? res_size : a_size;
__int128_t* const tres = (__int128_t*)res;
const int64_t* const ta = (int64_t*)a_dft;
for (uint64_t i = 0; i < smin; i++) {
memcpy(tmp, ta + i * nn * 4, nn * 4 * sizeof(uint64_t));
q120_intt_bb_avx2(module->mod.q120.p_intt, (q120b*)tmp);
q120_b_to_znx128_simple(nn, tres + i * nn, (q120b*)tmp);
}
// fill up remaining part with 0's
memset(tres + smin * nn, 0, (res_size - smin) * nn * sizeof(*tres));
}
EXPORT uint64_t ntt120_vec_znx_idft_tmp_bytes_avx(const MODULE* module, uint64_t nn) { return nn * 4 * sizeof(uint64_t); }
EXPORT void ntt120_vec_znx_idft_tmp_a_avx(const MODULE* module, // N
VEC_ZNX_BIG* res, uint64_t res_size, // res
VEC_ZNX_DFT* a_dft, uint64_t a_size // a is overwritten
) {
const uint64_t nn = module->nn;
const uint64_t smin = res_size < a_size ? res_size : a_size;
__int128_t* const tres = (__int128_t*)res;
int64_t* const ta = (int64_t*)a_dft;
for (uint64_t i = 0; i < smin; i++) {
q120_intt_bb_avx2(module->mod.q120.p_intt, (q120b*)(ta + i * nn * 4));
q120_b_to_znx128_simple(nn, tres + i * nn, (q120b*)(ta + i * nn * 4));
}
// fill up remaining part with 0's
memset(tres + smin * nn, 0, (res_size - smin) * nn * sizeof(*tres));
}

1
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/vec_znx_dft_avx2.c
View File

@@ -1 +0,0 @@
#include "vec_znx_arithmetic_private.h"

369
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/vector_matrix_product.c
View File

@@ -1,369 +0,0 @@
#include <assert.h>
#include <string.h>
#include "../reim4/reim4_arithmetic.h"
#include "vec_znx_arithmetic_private.h"
EXPORT uint64_t bytes_of_vmp_pmat(const MODULE* module, // N
uint64_t nrows, uint64_t ncols // dimensions
) {
return module->func.bytes_of_vmp_pmat(module, nrows, ncols);
}
// fft64
EXPORT uint64_t fft64_bytes_of_vmp_pmat(const MODULE* module, // N
uint64_t nrows, uint64_t ncols // dimensions
) {
return module->nn * nrows * ncols * sizeof(double);
}
EXPORT VMP_PMAT* new_vmp_pmat(const MODULE* module, // N
uint64_t nrows, uint64_t ncols // dimensions
) {
return spqlios_alloc(bytes_of_vmp_pmat(module, nrows, ncols));
}
EXPORT void delete_vmp_pmat(VMP_PMAT* res) { spqlios_free(res); }
/** @brief prepares a vmp matrix (contiguous row-major version) */
EXPORT void vmp_prepare_contiguous(const MODULE* module, // N
VMP_PMAT* pmat, // output
const int64_t* mat, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
) {
module->func.vmp_prepare_contiguous(module, pmat, mat, nrows, ncols, tmp_space);
}
/** @brief minimal scratch space byte-size required for the vmp_prepare function */
EXPORT uint64_t vmp_prepare_tmp_bytes(const MODULE* module, uint64_t nn, // N
uint64_t nrows, uint64_t ncols) {
return module->func.vmp_prepare_tmp_bytes(module, nn, nrows, ncols);
}
EXPORT double* get_blk_addr(uint64_t row_i, uint64_t col_i, uint64_t nrows, uint64_t ncols, const VMP_PMAT* pmat) {
double* output_mat = (double*)pmat;
if (col_i == (ncols - 1) && (ncols % 2 == 1)) {
// special case: last column out of an odd column number
return output_mat + col_i * nrows * 8 // col == ncols-1
+ row_i * 8;
} else {
// general case: columns go by pair
return output_mat + (col_i / 2) * (2 * nrows) * 8 // second: col pair index
+ row_i * 2 * 8 // third: row index
+ (col_i % 2) * 8;
}
}
void fft64_store_svp_ppol_into_vmp_pmat_row_blk_ref(uint64_t nn, uint64_t m, const SVP_PPOL* svp_ppol, uint64_t row_i,
uint64_t col_i, uint64_t nrows, uint64_t ncols, VMP_PMAT* pmat) {
double* start_addr = get_blk_addr(row_i, col_i, nrows, ncols, pmat);
uint64_t offset = nrows * ncols * 8;
for (uint64_t blk_i = 0; blk_i < m / 4; blk_i++) {
reim4_extract_1blk_from_reim_ref(m, blk_i, start_addr + blk_i * offset, (double*)svp_ppol);
}
}
/** @brief prepares a vmp matrix (contiguous row-major version) */
EXPORT void fft64_vmp_prepare_contiguous_ref(const MODULE* module, // N
VMP_PMAT* pmat, // output
const int64_t* mat, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
) {
// there is an edge case if nn < 8
const uint64_t nn = module->nn;
const uint64_t m = module->m;
if (nn >= 8) {
for (uint64_t row_i = 0; row_i < nrows; row_i++) {
for (uint64_t col_i = 0; col_i < ncols; col_i++) {
reim_from_znx64(module->mod.fft64.p_conv, (SVP_PPOL*)tmp_space, mat + (row_i * ncols + col_i) * nn);
reim_fft(module->mod.fft64.p_fft, (double*)tmp_space);
fft64_store_svp_ppol_into_vmp_pmat_row_blk_ref(nn, m, (SVP_PPOL*)tmp_space, row_i, col_i, nrows, ncols, pmat);
}
}
} else {
for (uint64_t row_i = 0; row_i < nrows; row_i++) {
for (uint64_t col_i = 0; col_i < ncols; col_i++) {
double* res = (double*)pmat + (col_i * nrows + row_i) * nn;
reim_from_znx64(module->mod.fft64.p_conv, (SVP_PPOL*)res, mat + (row_i * ncols + col_i) * nn);
reim_fft(module->mod.fft64.p_fft, res);
}
}
}
}
/** @brief minimal scratch space byte-size required for the vmp_prepare function */
EXPORT uint64_t fft64_vmp_prepare_tmp_bytes(const MODULE* module, uint64_t nn, // N
uint64_t nrows, uint64_t ncols) {
return nn * sizeof(int64_t);
}
/** @brief applies a vmp product (result in DFT space) and adds to res inplace */
EXPORT void fft64_vmp_apply_dft_add_ref(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols,
uint64_t pmat_scale, // prep matrix
uint8_t* tmp_space // scratch space
) {
const uint64_t nn = module->nn;
const uint64_t rows = nrows < a_size ? nrows : a_size;
VEC_ZNX_DFT* a_dft = (VEC_ZNX_DFT*)tmp_space;
uint8_t* new_tmp_space = (uint8_t*)tmp_space + rows * nn * sizeof(double);
fft64_vec_znx_dft(module, a_dft, rows, a, a_size, a_sl);
fft64_vmp_apply_dft_to_dft_add_ref(module, res, res_size, a_dft, a_size, pmat, nrows, ncols, pmat_scale,
new_tmp_space);
}
/** @brief applies a vmp product (result in DFT space) */
EXPORT void fft64_vmp_apply_dft_ref(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space
) {
const uint64_t nn = module->nn;
const uint64_t rows = nrows < a_size ? nrows : a_size;
VEC_ZNX_DFT* a_dft = (VEC_ZNX_DFT*)tmp_space;
uint8_t* new_tmp_space = (uint8_t*)tmp_space + rows * nn * sizeof(double);
fft64_vec_znx_dft(module, a_dft, rows, a, a_size, a_sl);
fft64_vmp_apply_dft_to_dft_ref(module, res, res_size, a_dft, a_size, pmat, nrows, ncols, new_tmp_space);
}
/** @brief like fft64_vmp_apply_dft_to_dft_ref but adds in place */
EXPORT void fft64_vmp_apply_dft_to_dft_add_ref(const MODULE* module, // N
VEC_ZNX_DFT* res, const uint64_t res_size, // res
const VEC_ZNX_DFT* a_dft, uint64_t a_size, // a
const VMP_PMAT* pmat, const uint64_t nrows, const uint64_t ncols,
uint64_t pmat_scale, // prep matrix
uint8_t* tmp_space // scratch space (a_size*sizeof(reim4) bytes)
) {
const uint64_t m = module->m;
const uint64_t nn = module->nn;
assert(nn >= 8);
double* mat2cols_output = (double*)tmp_space; // 128 bytes
double* extracted_blk = (double*)tmp_space + 16; // 64*min(nrows,a_size) bytes
double* mat_input = (double*)pmat;
double* vec_input = (double*)a_dft;
double* vec_output = (double*)res;
// const uint64_t row_max0 = res_size < a_size ? res_size: a_size;
const uint64_t row_max = nrows < a_size ? nrows : a_size;
const uint64_t col_max = ncols < res_size ? ncols : res_size;
if (nn >= 8) {
for (uint64_t blk_i = 0; blk_i < m / 4; blk_i++) {
double* mat_blk_start = mat_input + blk_i * (8 * nrows * ncols);
reim4_extract_1blk_from_contiguous_reim_ref(m, row_max, blk_i, (double*)extracted_blk, (double*)a_dft);
if (pmat_scale % 2 == 0) {
// apply mat2cols
for (uint64_t col_res = 0, col_pmat = pmat_scale; col_pmat < col_max - 1; col_res += 2, col_pmat += 2) {
uint64_t col_offset = col_pmat * (8 * nrows);
reim4_vec_mat2cols_product_ref(row_max, mat2cols_output, extracted_blk, mat_blk_start + col_offset);
reim4_add_1blk_to_reim_ref(m, blk_i, vec_output + col_res * nn, mat2cols_output);
reim4_add_1blk_to_reim_ref(m, blk_i, vec_output + (col_res + 1) * nn, mat2cols_output + 8);
}
} else {
uint64_t col_offset = (pmat_scale - 1) * (8 * nrows);
reim4_vec_mat2cols_product_ref(row_max, mat2cols_output, extracted_blk, mat_blk_start + col_offset);
reim4_add_1blk_to_reim_ref(m, blk_i, vec_output, mat2cols_output + 8);
// apply mat2cols
for (uint64_t col_res = 1, col_pmat = pmat_scale + 1; col_pmat < col_max - 1; col_res += 2, col_pmat += 2) {
uint64_t col_offset = col_pmat * (8 * nrows);
reim4_vec_mat2cols_product_ref(row_max, mat2cols_output, extracted_blk, mat_blk_start + col_offset);
reim4_add_1blk_to_reim_ref(m, blk_i, vec_output + col_res * nn, mat2cols_output);
reim4_add_1blk_to_reim_ref(m, blk_i, vec_output + (col_res + 1) * nn, mat2cols_output + 8);
}
}
// check if col_max is odd, then special case
if (col_max % 2 == 1) {
uint64_t last_col = col_max - 1;
uint64_t col_offset = last_col * (8 * nrows);
if (last_col >= pmat_scale) {
// the last column is alone in the pmat: vec_mat1col
if (ncols == col_max) {
reim4_vec_mat1col_product_ref(row_max, mat2cols_output, extracted_blk, mat_blk_start + col_offset);
} else {
// the last column is part of a colpair in the pmat: vec_mat2cols and ignore the second position
reim4_vec_mat2cols_product_ref(row_max, mat2cols_output, extracted_blk, mat_blk_start + col_offset);
}
reim4_add_1blk_to_reim_ref(m, blk_i, vec_output + (last_col - pmat_scale) * nn, mat2cols_output);
}
}
}
} else {
for (uint64_t col_res = 0, col_pmat = pmat_scale; col_pmat < col_max; col_res += 1, col_pmat += 1) {
double* pmat_col = mat_input + col_pmat * nrows * nn;
for (uint64_t row_i = 0; row_i < row_max; row_i++) {
reim_fftvec_addmul(module->mod.fft64.p_addmul, vec_output + col_res * nn, vec_input + row_i * nn,
pmat_col + row_i * nn);
}
}
}
// zero out remaining bytes
memset(vec_output + col_max * nn, 0, (res_size - col_max) * nn * sizeof(double));
}
/** @brief this inner function could be very handy */
EXPORT void fft64_vmp_apply_dft_to_dft_ref(const MODULE* module, // N
VEC_ZNX_DFT* res, const uint64_t res_size, // res
const VEC_ZNX_DFT* a_dft, uint64_t a_size, // a
const VMP_PMAT* pmat, const uint64_t nrows,
const uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space (a_size*sizeof(reim4) bytes)
) {
const uint64_t m = module->m;
const uint64_t nn = module->nn;
assert(nn >= 8);
double* mat2cols_output = (double*)tmp_space; // 128 bytes
double* extracted_blk = (double*)tmp_space + 16; // 64*min(nrows,a_size) bytes
double* mat_input = (double*)pmat;
double* vec_input = (double*)a_dft;
double* vec_output = (double*)res;
// const uint64_t row_max0 = res_size < a_size ? res_size: a_size;
const uint64_t row_max = nrows < a_size ? nrows : a_size;
const uint64_t col_max = ncols < res_size ? ncols : res_size;
if (nn >= 8) {
for (uint64_t blk_i = 0; blk_i < m / 4; blk_i++) {
double* mat_blk_start = mat_input + blk_i * (8 * nrows * ncols);
reim4_extract_1blk_from_contiguous_reim_ref(m, row_max, blk_i, (double*)extracted_blk, (double*)a_dft);
// apply mat2cols
for (uint64_t col_i = 0; col_i < col_max - 1; col_i += 2) {
uint64_t col_offset = col_i * (8 * nrows);
reim4_vec_mat2cols_product_ref(row_max, mat2cols_output, extracted_blk, mat_blk_start + col_offset);
reim4_save_1blk_to_reim_ref(m, blk_i, vec_output + col_i * nn, mat2cols_output);
reim4_save_1blk_to_reim_ref(m, blk_i, vec_output + (col_i + 1) * nn, mat2cols_output + 8);
}
// check if col_max is odd, then special case
if (col_max % 2 == 1) {
uint64_t last_col = col_max - 1;
uint64_t col_offset = last_col * (8 * nrows);
// the last column is alone in the pmat: vec_mat1col
if (ncols == col_max) {
reim4_vec_mat1col_product_ref(row_max, mat2cols_output, extracted_blk, mat_blk_start + col_offset);
} else {
// the last column is part of a colpair in the pmat: vec_mat2cols and ignore the second position
reim4_vec_mat2cols_product_ref(row_max, mat2cols_output, extracted_blk, mat_blk_start + col_offset);
}
reim4_save_1blk_to_reim_ref(m, blk_i, vec_output + last_col * nn, mat2cols_output);
}
}
} else {
for (uint64_t col_i = 0; col_i < col_max; col_i++) {
double* pmat_col = mat_input + col_i * nrows * nn;
for (uint64_t row_i = 0; row_i < 1; row_i++) {
reim_fftvec_mul(module->mod.fft64.mul_fft, vec_output + col_i * nn, vec_input + row_i * nn,
pmat_col + row_i * nn);
}
for (uint64_t row_i = 1; row_i < row_max; row_i++) {
reim_fftvec_addmul(module->mod.fft64.p_addmul, vec_output + col_i * nn, vec_input + row_i * nn,
pmat_col + row_i * nn);
}
}
}
// zero out remaining bytes
memset(vec_output + col_max * nn, 0, (res_size - col_max) * nn * sizeof(double));
}
/** @brief minimal size of the tmp_space */
EXPORT uint64_t fft64_vmp_apply_dft_tmp_bytes(const MODULE* module, uint64_t nn, // N
uint64_t res_size, // res
uint64_t a_size, // a
uint64_t nrows, uint64_t ncols // prep matrix
) {
const uint64_t row_max = nrows < a_size ? nrows : a_size;
return (row_max * nn * sizeof(double)) + (128) + (64 * row_max);
}
/** @brief minimal size of the tmp_space */
EXPORT uint64_t fft64_vmp_apply_dft_to_dft_tmp_bytes(const MODULE* module, uint64_t nn, // N
uint64_t res_size, // res
uint64_t a_size, // a
uint64_t nrows, uint64_t ncols // prep matrix
) {
const uint64_t row_max = nrows < a_size ? nrows : a_size;
return (128) + (64 * row_max);
}
EXPORT void vmp_apply_dft_to_dft(const MODULE* module, // N
VEC_ZNX_DFT* res, const uint64_t res_size, // res
const VEC_ZNX_DFT* a_dft, uint64_t a_size, // a
const VMP_PMAT* pmat, const uint64_t nrows,
const uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space (a_size*sizeof(reim4) bytes)
) {
module->func.vmp_apply_dft_to_dft(module, res, res_size, a_dft, a_size, pmat, nrows, ncols, tmp_space);
}
EXPORT void vmp_apply_dft_to_dft_add(const MODULE* module, // N
VEC_ZNX_DFT* res, const uint64_t res_size, // res
const VEC_ZNX_DFT* a_dft, uint64_t a_size, // a
const VMP_PMAT* pmat, const uint64_t nrows, const uint64_t ncols,
uint64_t pmat_scale, // prep matrix
uint8_t* tmp_space // scratch space (a_size*sizeof(reim4) bytes)
) {
module->func.vmp_apply_dft_to_dft_add(module, res, res_size, a_dft, a_size, pmat, nrows, ncols, pmat_scale,
tmp_space);
}
EXPORT uint64_t vmp_apply_dft_to_dft_tmp_bytes(const MODULE* module, uint64_t nn, // N
uint64_t res_size, // res
uint64_t a_size, // a
uint64_t nrows, uint64_t ncols // prep matrix
) {
return module->func.vmp_apply_dft_to_dft_tmp_bytes(module, nn, res_size, a_size, nrows, ncols);
}
/** @brief applies a vmp product (result in DFT space) adds to res inplace */
EXPORT void vmp_apply_dft_add(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols, uint64_t pmat_scale, // prep matrix
uint8_t* tmp_space // scratch space
) {
module->func.vmp_apply_dft_add(module, res, res_size, a, a_size, a_sl, pmat, nrows, ncols, pmat_scale, tmp_space);
}
/** @brief applies a vmp product (result in DFT space) */
EXPORT void vmp_apply_dft(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space
) {
module->func.vmp_apply_dft(module, res, res_size, a, a_size, a_sl, pmat, nrows, ncols, tmp_space);
}
/** @brief minimal size of the tmp_space */
EXPORT uint64_t vmp_apply_dft_tmp_bytes(const MODULE* module, uint64_t nn, // N
uint64_t res_size, // res
uint64_t a_size, // a
uint64_t nrows, uint64_t ncols // prep matrix
) {
return module->func.vmp_apply_dft_tmp_bytes(module, nn, res_size, a_size, nrows, ncols);
}

244
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/vector_matrix_product_avx.c
View File

@@ -1,244 +0,0 @@
#include <string.h>
#include "../reim4/reim4_arithmetic.h"
#include "vec_znx_arithmetic_private.h"
/** @brief prepares a vmp matrix (contiguous row-major version) */
EXPORT void fft64_vmp_prepare_contiguous_avx(const MODULE* module, // N
VMP_PMAT* pmat, // output
const int64_t* mat, uint64_t nrows, uint64_t ncols, // a
uint8_t* tmp_space // scratch space
) {
// there is an edge case if nn < 8
const uint64_t nn = module->nn;
const uint64_t m = module->m;
double* output_mat = (double*)pmat;
double* start_addr = (double*)pmat;
uint64_t offset = nrows * ncols * 8;
if (nn >= 8) {
for (uint64_t row_i = 0; row_i < nrows; row_i++) {
for (uint64_t col_i = 0; col_i < ncols; col_i++) {
reim_from_znx64(module->mod.fft64.p_conv, (SVP_PPOL*)tmp_space, mat + (row_i * ncols + col_i) * nn);
reim_fft(module->mod.fft64.p_fft, (double*)tmp_space);
if (col_i == (ncols - 1) && (ncols % 2 == 1)) {
// special case: last column out of an odd column number
start_addr = output_mat + col_i * nrows * 8 // col == ncols-1
+ row_i * 8;
} else {
// general case: columns go by pair
start_addr = output_mat + (col_i / 2) * (2 * nrows) * 8 // second: col pair index
+ row_i * 2 * 8 // third: row index
+ (col_i % 2) * 8;
}
for (uint64_t blk_i = 0; blk_i < m / 4; blk_i++) {
// extract blk from tmp and save it
reim4_extract_1blk_from_reim_avx(m, blk_i, start_addr + blk_i * offset, (double*)tmp_space);
}
}
}
} else {
for (uint64_t row_i = 0; row_i < nrows; row_i++) {
for (uint64_t col_i = 0; col_i < ncols; col_i++) {
double* res = (double*)pmat + (col_i * nrows + row_i) * nn;
reim_from_znx64(module->mod.fft64.p_conv, (SVP_PPOL*)res, mat + (row_i * ncols + col_i) * nn);
reim_fft(module->mod.fft64.p_fft, res);
}
}
}
}
double* get_blk_addr(int row, int col, int nrows, int ncols, VMP_PMAT* pmat);
void fft64_store_svp_ppol_into_vmp_pmat_row_blk_avx(uint64_t nn, uint64_t m, const SVP_PPOL* svp_ppol, uint64_t row_i,
uint64_t col_i, uint64_t nrows, uint64_t ncols, VMP_PMAT* pmat) {
double* start_addr = get_blk_addr(row_i, col_i, nrows, ncols, pmat);
uint64_t offset = nrows * ncols * 8;
for (uint64_t blk_i = 0; blk_i < m / 4; blk_i++) {
reim4_extract_1blk_from_reim_avx(m, blk_i, start_addr + blk_i * offset, (double*)svp_ppol);
}
}
/** @brief applies a vmp product (result in DFT space) abd adds to res inplace */
EXPORT void fft64_vmp_apply_dft_add_avx(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols,
uint64_t pmat_scale, // prep matrix
uint8_t* tmp_space // scratch space
) {
const uint64_t nn = module->nn;
const uint64_t rows = nrows < a_size ? nrows : a_size;
VEC_ZNX_DFT* a_dft = (VEC_ZNX_DFT*)tmp_space;
uint8_t* new_tmp_space = (uint8_t*)tmp_space + rows * nn * sizeof(double);
fft64_vec_znx_dft(module, a_dft, rows, a, a_size, a_sl);
fft64_vmp_apply_dft_to_dft_add_avx(module, res, res_size, a_dft, a_size, pmat, nrows, ncols, pmat_scale,
new_tmp_space);
}
/** @brief applies a vmp product (result in DFT space) */
EXPORT void fft64_vmp_apply_dft_avx(const MODULE* module, // N
VEC_ZNX_DFT* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size, uint64_t a_sl, // a
const VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space
) {
const uint64_t nn = module->nn;
const uint64_t rows = nrows < a_size ? nrows : a_size;
VEC_ZNX_DFT* a_dft = (VEC_ZNX_DFT*)tmp_space;
uint8_t* new_tmp_space = (uint8_t*)tmp_space + rows * nn * sizeof(double);
fft64_vec_znx_dft(module, a_dft, rows, a, a_size, a_sl);
fft64_vmp_apply_dft_to_dft_avx(module, res, res_size, a_dft, a_size, pmat, nrows, ncols, new_tmp_space);
}
EXPORT void fft64_vmp_apply_dft_to_dft_add_avx(const MODULE* module, // N
VEC_ZNX_DFT* res, const uint64_t res_size, // res
const VEC_ZNX_DFT* a_dft, uint64_t a_size, // a
const VMP_PMAT* pmat, const uint64_t nrows, const uint64_t ncols,
uint64_t pmat_scale, // prep matrix
uint8_t* tmp_space // scratch space (a_size*sizeof(reim4) bytes)
) {
const uint64_t m = module->m;
const uint64_t nn = module->nn;
double* mat2cols_output = (double*)tmp_space; // 128 bytes
double* extracted_blk = (double*)tmp_space + 16; // 64*min(nrows,a_size) bytes
double* mat_input = (double*)pmat;
double* vec_input = (double*)a_dft;
double* vec_output = (double*)res;
const uint64_t row_max = nrows < a_size ? nrows : a_size;
const uint64_t col_max = ncols < res_size ? ncols : res_size;
if (nn >= 8) {
for (uint64_t blk_i = 0; blk_i < m / 4; blk_i++) {
double* mat_blk_start = mat_input + blk_i * (8 * nrows * ncols);
reim4_extract_1blk_from_contiguous_reim_avx(m, row_max, blk_i, (double*)extracted_blk, (double*)a_dft);
if (pmat_scale % 2 == 0) {
for (uint64_t col_res = 0, col_pmat = pmat_scale; col_pmat < col_max - 1; col_res += 2, col_pmat += 2) {
uint64_t col_offset = col_pmat * (8 * nrows);
reim4_vec_mat2cols_product_avx2(row_max, mat2cols_output, extracted_blk, mat_blk_start + col_offset);
reim4_add_1blk_to_reim_avx(m, blk_i, vec_output + col_res * nn, mat2cols_output);
reim4_add_1blk_to_reim_avx(m, blk_i, vec_output + (col_res + 1) * nn, mat2cols_output + 8);
}
} else {
uint64_t col_offset = (pmat_scale - 1) * (8 * nrows);
reim4_vec_mat2cols_product_avx2(row_max, mat2cols_output, extracted_blk, mat_blk_start + col_offset);
reim4_add_1blk_to_reim_avx(m, blk_i, vec_output, mat2cols_output + 8);
for (uint64_t col_res = 1, col_pmat = pmat_scale + 1; col_pmat < col_max - 1; col_res += 2, col_pmat += 2) {
uint64_t col_offset = col_pmat * (8 * nrows);
reim4_vec_mat2cols_product_avx2(row_max, mat2cols_output, extracted_blk, mat_blk_start + col_offset);
reim4_add_1blk_to_reim_avx(m, blk_i, vec_output + col_res * nn, mat2cols_output);
reim4_add_1blk_to_reim_avx(m, blk_i, vec_output + (col_res + 1) * nn, mat2cols_output + 8);
}
}
// check if col_max is odd, then special case
if (col_max % 2 == 1) {
uint64_t last_col = col_max - 1;
uint64_t col_offset = last_col * (8 * nrows);
if (last_col >= pmat_scale) {
// the last column is alone in the pmat: vec_mat1col
if (ncols == col_max)
reim4_vec_mat1col_product_avx2(row_max, mat2cols_output, extracted_blk, mat_blk_start + col_offset);
else {
// the last column is part of a colpair in the pmat: vec_mat2cols and ignore the second position
reim4_vec_mat2cols_product_avx2(row_max, mat2cols_output, extracted_blk, mat_blk_start + col_offset);
}
reim4_add_1blk_to_reim_avx(m, blk_i, vec_output + (last_col - pmat_scale) * nn, mat2cols_output);
}
}
}
} else {
for (uint64_t col_res = 0, col_pmat = pmat_scale; col_pmat < col_max; col_res += 1, col_pmat += 1) {
double* pmat_col = mat_input + col_pmat * nrows * nn;
for (uint64_t row_i = 0; row_i < row_max; row_i++) {
reim_fftvec_addmul(module->mod.fft64.p_addmul, vec_output + col_res * nn, vec_input + row_i * nn,
pmat_col + row_i * nn);
}
}
}
// zero out remaining bytes
memset(vec_output + col_max * nn, 0, (res_size - col_max) * nn * sizeof(double));
}
/** @brief this inner function could be very handy */
EXPORT void fft64_vmp_apply_dft_to_dft_avx(const MODULE* module, // N
VEC_ZNX_DFT* res, const uint64_t res_size, // res
const VEC_ZNX_DFT* a_dft, uint64_t a_size, // a
const VMP_PMAT* pmat, const uint64_t nrows,
const uint64_t ncols, // prep matrix
uint8_t* tmp_space // scratch space (a_size*sizeof(reim4) bytes)
) {
const uint64_t m = module->m;
const uint64_t nn = module->nn;
double* mat2cols_output = (double*)tmp_space; // 128 bytes
double* extracted_blk = (double*)tmp_space + 16; // 64*min(nrows,a_size) bytes
double* mat_input = (double*)pmat;
double* vec_input = (double*)a_dft;
double* vec_output = (double*)res;
const uint64_t row_max = nrows < a_size ? nrows : a_size;
const uint64_t col_max = ncols < res_size ? ncols : res_size;
if (nn >= 8) {
for (uint64_t blk_i = 0; blk_i < m / 4; blk_i++) {
double* mat_blk_start = mat_input + blk_i * (8 * nrows * ncols);
reim4_extract_1blk_from_contiguous_reim_avx(m, row_max, blk_i, (double*)extracted_blk, (double*)a_dft);
// apply mat2cols
for (uint64_t col_i = 0; col_i < col_max - 1; col_i += 2) {
uint64_t col_offset = col_i * (8 * nrows);
reim4_vec_mat2cols_product_avx2(row_max, mat2cols_output, extracted_blk, mat_blk_start + col_offset);
reim4_save_1blk_to_reim_avx(m, blk_i, vec_output + col_i * nn, mat2cols_output);
reim4_save_1blk_to_reim_avx(m, blk_i, vec_output + (col_i + 1) * nn, mat2cols_output + 8);
}
// check if col_max is odd, then special case
if (col_max % 2 == 1) {
uint64_t last_col = col_max - 1;
uint64_t col_offset = last_col * (8 * nrows);
// the last column is alone in the pmat: vec_mat1col
if (ncols == col_max)
reim4_vec_mat1col_product_avx2(row_max, mat2cols_output, extracted_blk, mat_blk_start + col_offset);
else {
// the last column is part of a colpair in the pmat: vec_mat2cols and ignore the second position
reim4_vec_mat2cols_product_avx2(row_max, mat2cols_output, extracted_blk, mat_blk_start + col_offset);
}
reim4_save_1blk_to_reim_avx(m, blk_i, vec_output + last_col * nn, mat2cols_output);
}
}
} else {
for (uint64_t col_i = 0; col_i < col_max; col_i++) {
double* pmat_col = mat_input + col_i * nrows * nn;
for (uint64_t row_i = 0; row_i < 1; row_i++) {
reim_fftvec_mul(module->mod.fft64.mul_fft, vec_output + col_i * nn, vec_input + row_i * nn,
pmat_col + row_i * nn);
}
for (uint64_t row_i = 1; row_i < row_max; row_i++) {
reim_fftvec_addmul(module->mod.fft64.p_addmul, vec_output + col_i * nn, vec_input + row_i * nn,
pmat_col + row_i * nn);
}
}
}
// zero out remaining bytes
memset(vec_output + col_max * nn, 0, (res_size - col_max) * nn * sizeof(double));
}

185
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/zn_api.c
View File

@@ -1,185 +0,0 @@
#include <string.h>
#include "zn_arithmetic_private.h"
void default_init_z_module_precomp(MOD_Z* module) {
// Add here initialization of items that are in the precomp
}
void default_finalize_z_module_precomp(MOD_Z* module) {
// Add here deleters for items that are in the precomp
}
void default_init_z_module_vtable(MOD_Z* module) {
// Add function pointers here
module->vtable.i8_approxdecomp_from_tndbl = default_i8_approxdecomp_from_tndbl_ref;
module->vtable.i16_approxdecomp_from_tndbl = default_i16_approxdecomp_from_tndbl_ref;
module->vtable.i32_approxdecomp_from_tndbl = default_i32_approxdecomp_from_tndbl_ref;
module->vtable.zn32_vmp_prepare_contiguous = default_zn32_vmp_prepare_contiguous_ref;
module->vtable.zn32_vmp_prepare_dblptr = default_zn32_vmp_prepare_dblptr_ref;
module->vtable.zn32_vmp_prepare_row = default_zn32_vmp_prepare_row_ref;
module->vtable.zn32_vmp_apply_i8 = default_zn32_vmp_apply_i8_ref;
module->vtable.zn32_vmp_apply_i16 = default_zn32_vmp_apply_i16_ref;
module->vtable.zn32_vmp_apply_i32 = default_zn32_vmp_apply_i32_ref;
module->vtable.dbl_to_tn32 = dbl_to_tn32_ref;
module->vtable.tn32_to_dbl = tn32_to_dbl_ref;
module->vtable.dbl_round_to_i32 = dbl_round_to_i32_ref;
module->vtable.i32_to_dbl = i32_to_dbl_ref;
module->vtable.dbl_round_to_i64 = dbl_round_to_i64_ref;
module->vtable.i64_to_dbl = i64_to_dbl_ref;
// Add optimized function pointers here
if (CPU_SUPPORTS("avx")) {
module->vtable.zn32_vmp_apply_i8 = default_zn32_vmp_apply_i8_avx;
module->vtable.zn32_vmp_apply_i16 = default_zn32_vmp_apply_i16_avx;
module->vtable.zn32_vmp_apply_i32 = default_zn32_vmp_apply_i32_avx;
}
}
void init_z_module_info(MOD_Z* module, //
Z_MODULE_TYPE mtype) {
memset(module, 0, sizeof(MOD_Z));
module->mtype = mtype;
switch (mtype) {
case DEFAULT:
default_init_z_module_precomp(module);
default_init_z_module_vtable(module);
break;
default:
NOT_SUPPORTED(); // unknown mtype
}
}
void finalize_z_module_info(MOD_Z* module) {
if (module->custom) module->custom_deleter(module->custom);
switch (module->mtype) {
case DEFAULT:
default_finalize_z_module_precomp(module);
// fft64_finalize_rnx_module_vtable(module); // nothing to finalize
break;
default:
NOT_SUPPORTED(); // unknown mtype
}
}
EXPORT MOD_Z* new_z_module_info(Z_MODULE_TYPE mtype) {
MOD_Z* res = (MOD_Z*)malloc(sizeof(MOD_Z));
init_z_module_info(res, mtype);
return res;
}
EXPORT void delete_z_module_info(MOD_Z* module_info) {
finalize_z_module_info(module_info);
free(module_info);
}
//////////////// wrappers //////////////////
/** @brief sets res = gadget_decompose(a) (int8_t* output) */
EXPORT void i8_approxdecomp_from_tndbl(const MOD_Z* module, // N
const TNDBL_APPROXDECOMP_GADGET* gadget, // gadget
int8_t* res, uint64_t res_size, // res (in general, size ell.a_size)
const double* a, uint64_t a_size) { // a
module->vtable.i8_approxdecomp_from_tndbl(module, gadget, res, res_size, a, a_size);
}
/** @brief sets res = gadget_decompose(a) (int16_t* output) */
EXPORT void i16_approxdecomp_from_tndbl(const MOD_Z* module, // N
const TNDBL_APPROXDECOMP_GADGET* gadget, // gadget
int16_t* res, uint64_t res_size, // res (in general, size ell.a_size)
const double* a, uint64_t a_size) { // a
module->vtable.i16_approxdecomp_from_tndbl(module, gadget, res, res_size, a, a_size);
}
/** @brief sets res = gadget_decompose(a) (int32_t* output) */
EXPORT void i32_approxdecomp_from_tndbl(const MOD_Z* module, // N
const TNDBL_APPROXDECOMP_GADGET* gadget, // gadget
int32_t* res, uint64_t res_size, // res (in general, size ell.a_size)
const double* a, uint64_t a_size) { // a
module->vtable.i32_approxdecomp_from_tndbl(module, gadget, res, res_size, a, a_size);
}
/** @brief prepares a vmp matrix (contiguous row-major version) */
EXPORT void zn32_vmp_prepare_contiguous(const MOD_Z* module,
ZN32_VMP_PMAT* pmat, // output
const int32_t* mat, uint64_t nrows, uint64_t ncols) { // a
module->vtable.zn32_vmp_prepare_contiguous(module, pmat, mat, nrows, ncols);
}
/** @brief prepares a vmp matrix (mat[row]+col*N points to the item) */
EXPORT void zn32_vmp_prepare_dblptr(const MOD_Z* module,
ZN32_VMP_PMAT* pmat, // output
const int32_t** mat, uint64_t nrows, uint64_t ncols) { // a
module->vtable.zn32_vmp_prepare_dblptr(module, pmat, mat, nrows, ncols);
}
/** @brief prepares the ith-row of a vmp matrix with nrows and ncols */
EXPORT void zn32_vmp_prepare_row(const MOD_Z* module,
ZN32_VMP_PMAT* pmat, // output
const int32_t* row, uint64_t row_i, uint64_t nrows, uint64_t ncols) { // a
module->vtable.zn32_vmp_prepare_row(module, pmat, row, row_i, nrows, ncols);
}
/** @brief applies a vmp product (int32_t* input) */
EXPORT void zn32_vmp_apply_i32(const MOD_Z* module, int32_t* res, uint64_t res_size, const int32_t* a, uint64_t a_size,
const ZN32_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols) {
module->vtable.zn32_vmp_apply_i32(module, res, res_size, a, a_size, pmat, nrows, ncols);
}
/** @brief applies a vmp product (int16_t* input) */
EXPORT void zn32_vmp_apply_i16(const MOD_Z* module, int32_t* res, uint64_t res_size, const int16_t* a, uint64_t a_size,
const ZN32_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols) {
module->vtable.zn32_vmp_apply_i16(module, res, res_size, a, a_size, pmat, nrows, ncols);
}
/** @brief applies a vmp product (int8_t* input) */
EXPORT void zn32_vmp_apply_i8(const MOD_Z* module, int32_t* res, uint64_t res_size, const int8_t* a, uint64_t a_size,
const ZN32_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols) {
module->vtable.zn32_vmp_apply_i8(module, res, res_size, a, a_size, pmat, nrows, ncols);
}
/** reduction mod 1, output in torus32 space */
EXPORT void dbl_to_tn32(const MOD_Z* module, //
int32_t* res, uint64_t res_size, // res
const double* a, uint64_t a_size // a
) {
module->vtable.dbl_to_tn32(module, res, res_size, a, a_size);
}
/** real centerlift mod 1, output in double space */
EXPORT void tn32_to_dbl(const MOD_Z* module, //
double* res, uint64_t res_size, // res
const int32_t* a, uint64_t a_size // a
) {
module->vtable.tn32_to_dbl(module, res, res_size, a, a_size);
}
/** round to the nearest int, output in i32 space */
EXPORT void dbl_round_to_i32(const MOD_Z* module, //
int32_t* res, uint64_t res_size, // res
const double* a, uint64_t a_size // a
) {
module->vtable.dbl_round_to_i32(module, res, res_size, a, a_size);
}
/** small int (int32 space) to double */
EXPORT void i32_to_dbl(const MOD_Z* module, //
double* res, uint64_t res_size, // res
const int32_t* a, uint64_t a_size // a
) {
module->vtable.i32_to_dbl(module, res, res_size, a, a_size);
}
/** round to the nearest int, output in int64 space */
EXPORT void dbl_round_to_i64(const MOD_Z* module, //
int64_t* res, uint64_t res_size, // res
const double* a, uint64_t a_size // a
) {
module->vtable.dbl_round_to_i64(module, res, res_size, a, a_size);
}
/** small int (int64 space, <= 2^50) to double */
EXPORT void i64_to_dbl(const MOD_Z* module, //
double* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size // a
) {
module->vtable.i64_to_dbl(module, res, res_size, a, a_size);
}

81
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/zn_approxdecomp_ref.c
View File

@@ -1,81 +0,0 @@
#include <memory.h>
#include "zn_arithmetic_private.h"
EXPORT TNDBL_APPROXDECOMP_GADGET* new_tndbl_approxdecomp_gadget(const MOD_Z* module, //
uint64_t k, uint64_t ell) {
if (k * ell > 50) {
return spqlios_error("approx decomposition requested is too precise for doubles");
}
if (k < 1) {
return spqlios_error("approx decomposition supports k>=1");
}
TNDBL_APPROXDECOMP_GADGET* res = malloc(sizeof(TNDBL_APPROXDECOMP_GADGET));
memset(res, 0, sizeof(TNDBL_APPROXDECOMP_GADGET));
res->k = k;
res->ell = ell;
double add_cst = INT64_C(3) << (51 - k * ell);
for (uint64_t i = 0; i < ell; ++i) {
add_cst += pow(2., -(double)(i * k + 1));
}
res->add_cst = add_cst;
res->and_mask = (UINT64_C(1) << k) - 1;
res->sub_cst = UINT64_C(1) << (k - 1);
for (uint64_t i = 0; i < ell; ++i) res->rshifts[i] = (ell - 1 - i) * k;
return res;
}
EXPORT void delete_tndbl_approxdecomp_gadget(TNDBL_APPROXDECOMP_GADGET* ptr) { free(ptr); }
EXPORT int default_init_tndbl_approxdecomp_gadget(const MOD_Z* module, //
TNDBL_APPROXDECOMP_GADGET* res, //
uint64_t k, uint64_t ell) {
return 0;
}
typedef union {
double dv;
uint64_t uv;
} du_t;
#define IMPL_ixx_approxdecomp_from_tndbl_ref(ITYPE) \
if (res_size != a_size * gadget->ell) NOT_IMPLEMENTED(); \
const uint64_t ell = gadget->ell; \
const double add_cst = gadget->add_cst; \
const uint8_t* const rshifts = gadget->rshifts; \
const ITYPE and_mask = gadget->and_mask; \
const ITYPE sub_cst = gadget->sub_cst; \
ITYPE* rr = res; \
const double* aa = a; \
const double* aaend = a + a_size; \
while (aa < aaend) { \
du_t t = {.dv = *aa + add_cst}; \
for (uint64_t i = 0; i < ell; ++i) { \
ITYPE v = (ITYPE)(t.uv >> rshifts[i]); \
*rr = (v & and_mask) - sub_cst; \
++rr; \
} \
++aa; \
}
/** @brief sets res = gadget_decompose(a) (int8_t* output) */
EXPORT void default_i8_approxdecomp_from_tndbl_ref(const MOD_Z* module, // N
const TNDBL_APPROXDECOMP_GADGET* gadget, // gadget
int8_t* res, uint64_t res_size, // res (in general, size ell.a_size)
const double* a, uint64_t a_size //
){IMPL_ixx_approxdecomp_from_tndbl_ref(int8_t)}
/** @brief sets res = gadget_decompose(a) (int16_t* output) */
EXPORT void default_i16_approxdecomp_from_tndbl_ref(const MOD_Z* module, // N
const TNDBL_APPROXDECOMP_GADGET* gadget, // gadget
int16_t* res, uint64_t res_size, // res
const double* a, uint64_t a_size // a
){IMPL_ixx_approxdecomp_from_tndbl_ref(int16_t)}
/** @brief sets res = gadget_decompose(a) (int32_t* output) */
EXPORT void default_i32_approxdecomp_from_tndbl_ref(const MOD_Z* module, // N
const TNDBL_APPROXDECOMP_GADGET* gadget, // gadget
int32_t* res, uint64_t res_size, // res
const double* a, uint64_t a_size // a
) {
IMPL_ixx_approxdecomp_from_tndbl_ref(int32_t)
}

147
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/zn_arithmetic.h
View File

@@ -1,147 +0,0 @@
#ifndef SPQLIOS_ZN_ARITHMETIC_H
#define SPQLIOS_ZN_ARITHMETIC_H
#include <stdint.h>
#include "../commons.h"
typedef enum z_module_type_t { DEFAULT } Z_MODULE_TYPE;
/** @brief opaque structure that describes the module and the hardware */
typedef struct z_module_info_t MOD_Z;
/**
* @brief obtain a module info for ring dimension N
* the module-info knows about:
* - the dimension N (or the complex dimension m=N/2)
* - any moduleuted fft or ntt items
* - the hardware (avx, arm64, x86, ...)
*/
EXPORT MOD_Z* new_z_module_info(Z_MODULE_TYPE mode);
EXPORT void delete_z_module_info(MOD_Z* module_info);
typedef struct tndbl_approxdecomp_gadget_t TNDBL_APPROXDECOMP_GADGET;
EXPORT TNDBL_APPROXDECOMP_GADGET* new_tndbl_approxdecomp_gadget(const MOD_Z* module, //
uint64_t k,
uint64_t ell); // base 2^k, and size
EXPORT void delete_tndbl_approxdecomp_gadget(TNDBL_APPROXDECOMP_GADGET* ptr);
/** @brief sets res = gadget_decompose(a) (int8_t* output) */
EXPORT void i8_approxdecomp_from_tndbl(const MOD_Z* module, // N
const TNDBL_APPROXDECOMP_GADGET* gadget, // gadget
int8_t* res, uint64_t res_size, // res (in general, size ell.a_size)
const double* a, uint64_t a_size); // a
/** @brief sets res = gadget_decompose(a) (int16_t* output) */
EXPORT void i16_approxdecomp_from_tndbl(const MOD_Z* module, // N
const TNDBL_APPROXDECOMP_GADGET* gadget, // gadget
int16_t* res, uint64_t res_size, // res (in general, size ell.a_size)
const double* a, uint64_t a_size); // a
/** @brief sets res = gadget_decompose(a) (int32_t* output) */
EXPORT void i32_approxdecomp_from_tndbl(const MOD_Z* module, // N
const TNDBL_APPROXDECOMP_GADGET* gadget, // gadget
int32_t* res, uint64_t res_size, // res (in general, size ell.a_size)
const double* a, uint64_t a_size); // a
/** @brief opaque type that represents a prepared matrix */
typedef struct zn32_vmp_pmat_t ZN32_VMP_PMAT;
/** @brief size in bytes of a prepared matrix (for custom allocation) */
EXPORT uint64_t bytes_of_zn32_vmp_pmat(const MOD_Z* module, // N
uint64_t nrows, uint64_t ncols); // dimensions
/** @brief allocates a prepared matrix (release with delete_zn32_vmp_pmat) */
EXPORT ZN32_VMP_PMAT* new_zn32_vmp_pmat(const MOD_Z* module, // N
uint64_t nrows, uint64_t ncols); // dimensions
/** @brief deletes a prepared matrix (release with free) */
EXPORT void delete_zn32_vmp_pmat(ZN32_VMP_PMAT* ptr); // dimensions
/** @brief prepares a vmp matrix (contiguous row-major version) */
EXPORT void zn32_vmp_prepare_contiguous( //
const MOD_Z* module,
ZN32_VMP_PMAT* pmat, // output
const int32_t* mat, uint64_t nrows, uint64_t ncols); // a
/** @brief prepares a vmp matrix (mat[row]+col*N points to the item) */
EXPORT void zn32_vmp_prepare_dblptr( //
const MOD_Z* module,
ZN32_VMP_PMAT* pmat, // output
const int32_t** mat, uint64_t nrows, uint64_t ncols); // a
/** @brief prepares a vmp matrix (mat[row]+col*N points to the item) */
EXPORT void zn32_vmp_prepare_row( //
const MOD_Z* module,
ZN32_VMP_PMAT* pmat, // output
const int32_t* row, uint64_t row_i, uint64_t nrows, uint64_t ncols); // a
/** @brief applies a vmp product (int32_t* input) */
EXPORT void zn32_vmp_apply_i32( //
const MOD_Z* module, //
int32_t* res, uint64_t res_size, // res
const int32_t* a, uint64_t a_size, // a
const ZN32_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols); // prep matrix
/** @brief applies a vmp product (int16_t* input) */
EXPORT void zn32_vmp_apply_i16( //
const MOD_Z* module, //
int32_t* res, uint64_t res_size, // res
const int16_t* a, uint64_t a_size, // a
const ZN32_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols); // prep matrix
/** @brief applies a vmp product (int8_t* input) */
EXPORT void zn32_vmp_apply_i8( //
const MOD_Z* module, //
int32_t* res, uint64_t res_size, // res
const int8_t* a, uint64_t a_size, // a
const ZN32_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols); // prep matrix
// explicit conversions
/** reduction mod 1, output in torus32 space */
EXPORT void dbl_to_tn32(const MOD_Z* module, //
int32_t* res, uint64_t res_size, // res
const double* a, uint64_t a_size // a
);
/** real centerlift mod 1, output in double space */
EXPORT void tn32_to_dbl(const MOD_Z* module, //
double* res, uint64_t res_size, // res
const int32_t* a, uint64_t a_size // a
);
/** round to the nearest int, output in i32 space.
* WARNING: ||a||_inf must be <= 2^18 in this function
*/
EXPORT void dbl_round_to_i32(const MOD_Z* module, //
int32_t* res, uint64_t res_size, // res
const double* a, uint64_t a_size // a
);
/** small int (int32 space) to double
* WARNING: ||a||_inf must be <= 2^18 in this function
*/
EXPORT void i32_to_dbl(const MOD_Z* module, //
double* res, uint64_t res_size, // res
const int32_t* a, uint64_t a_size // a
);
/** round to the nearest int, output in int64 space
* WARNING: ||a||_inf must be <= 2^50 in this function
*/
EXPORT void dbl_round_to_i64(const MOD_Z* module, //
int64_t* res, uint64_t res_size, // res
const double* a, uint64_t a_size // a
);
/** small int (int64 space, <= 2^50) to double
* WARNING: ||a||_inf must be <= 2^50 in this function
*/
EXPORT void i64_to_dbl(const MOD_Z* module, //
double* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size // a
);
#endif // SPQLIOS_ZN_ARITHMETIC_H

43
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/zn_arithmetic_plugin.h
View File

@@ -1,43 +0,0 @@
#ifndef SPQLIOS_ZN_ARITHMETIC_PLUGIN_H
#define SPQLIOS_ZN_ARITHMETIC_PLUGIN_H
#include "zn_arithmetic.h"
typedef typeof(i8_approxdecomp_from_tndbl) I8_APPROXDECOMP_FROM_TNDBL_F;
typedef typeof(i16_approxdecomp_from_tndbl) I16_APPROXDECOMP_FROM_TNDBL_F;
typedef typeof(i32_approxdecomp_from_tndbl) I32_APPROXDECOMP_FROM_TNDBL_F;
typedef typeof(bytes_of_zn32_vmp_pmat) BYTES_OF_ZN32_VMP_PMAT_F;
typedef typeof(zn32_vmp_prepare_contiguous) ZN32_VMP_PREPARE_CONTIGUOUS_F;
typedef typeof(zn32_vmp_prepare_dblptr) ZN32_VMP_PREPARE_DBLPTR_F;
typedef typeof(zn32_vmp_prepare_row) ZN32_VMP_PREPARE_ROW_F;
typedef typeof(zn32_vmp_apply_i32) ZN32_VMP_APPLY_I32_F;
typedef typeof(zn32_vmp_apply_i16) ZN32_VMP_APPLY_I16_F;
typedef typeof(zn32_vmp_apply_i8) ZN32_VMP_APPLY_I8_F;
typedef typeof(dbl_to_tn32) DBL_TO_TN32_F;
typedef typeof(tn32_to_dbl) TN32_TO_DBL_F;
typedef typeof(dbl_round_to_i32) DBL_ROUND_TO_I32_F;
typedef typeof(i32_to_dbl) I32_TO_DBL_F;
typedef typeof(dbl_round_to_i64) DBL_ROUND_TO_I64_F;
typedef typeof(i64_to_dbl) I64_TO_DBL_F;
typedef struct z_module_vtable_t Z_MODULE_VTABLE;
struct z_module_vtable_t {
I8_APPROXDECOMP_FROM_TNDBL_F* i8_approxdecomp_from_tndbl;
I16_APPROXDECOMP_FROM_TNDBL_F* i16_approxdecomp_from_tndbl;
I32_APPROXDECOMP_FROM_TNDBL_F* i32_approxdecomp_from_tndbl;
BYTES_OF_ZN32_VMP_PMAT_F* bytes_of_zn32_vmp_pmat;
ZN32_VMP_PREPARE_CONTIGUOUS_F* zn32_vmp_prepare_contiguous;
ZN32_VMP_PREPARE_DBLPTR_F* zn32_vmp_prepare_dblptr;
ZN32_VMP_PREPARE_ROW_F* zn32_vmp_prepare_row;
ZN32_VMP_APPLY_I32_F* zn32_vmp_apply_i32;
ZN32_VMP_APPLY_I16_F* zn32_vmp_apply_i16;
ZN32_VMP_APPLY_I8_F* zn32_vmp_apply_i8;
DBL_TO_TN32_F* dbl_to_tn32;
TN32_TO_DBL_F* tn32_to_dbl;
DBL_ROUND_TO_I32_F* dbl_round_to_i32;
I32_TO_DBL_F* i32_to_dbl;
DBL_ROUND_TO_I64_F* dbl_round_to_i64;
I64_TO_DBL_F* i64_to_dbl;
};
#endif // SPQLIOS_ZN_ARITHMETIC_PLUGIN_H

164
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/zn_arithmetic_private.h
View File

@@ -1,164 +0,0 @@
#ifndef SPQLIOS_ZN_ARITHMETIC_PRIVATE_H
#define SPQLIOS_ZN_ARITHMETIC_PRIVATE_H
#include "../commons_private.h"
#include "zn_arithmetic.h"
#include "zn_arithmetic_plugin.h"
typedef struct main_z_module_precomp_t MAIN_Z_MODULE_PRECOMP;
struct main_z_module_precomp_t {
// TODO
};
typedef union z_module_precomp_t Z_MODULE_PRECOMP;
union z_module_precomp_t {
MAIN_Z_MODULE_PRECOMP main;
};
void main_init_z_module_precomp(MOD_Z* module);
void main_finalize_z_module_precomp(MOD_Z* module);
/** @brief opaque structure that describes the modules (RnX,ZnX,TnX) and the hardware */
struct z_module_info_t {
Z_MODULE_TYPE mtype;
Z_MODULE_VTABLE vtable;
Z_MODULE_PRECOMP precomp;
void* custom;
void (*custom_deleter)(void*);
};
void init_z_module_info(MOD_Z* module, Z_MODULE_TYPE mtype);
void main_init_z_module_vtable(MOD_Z* module);
struct tndbl_approxdecomp_gadget_t {
uint64_t k;
uint64_t ell;
double add_cst; // 3.2^51-(K.ell) + 1/2.(sum 2^-(i+1)K)
int64_t and_mask; // (2^K)-1
int64_t sub_cst; // 2^(K-1)
uint8_t rshifts[64]; // 2^(ell-1-i).K for i in [0:ell-1]
};
/** @brief sets res = gadget_decompose(a) (int8_t* output) */
EXPORT void default_i8_approxdecomp_from_tndbl_ref(const MOD_Z* module, // N
const TNDBL_APPROXDECOMP_GADGET* gadget, // gadget
int8_t* res, uint64_t res_size, // res (in general, size ell.a_size)
const double* a, uint64_t a_size); // a
/** @brief sets res = gadget_decompose(a) (int16_t* output) */
EXPORT void default_i16_approxdecomp_from_tndbl_ref(const MOD_Z* module, // N
const TNDBL_APPROXDECOMP_GADGET* gadget, // gadget
int16_t* res,
uint64_t res_size, // res (in general, size ell.a_size)
const double* a, uint64_t a_size); // a
/** @brief sets res = gadget_decompose(a) (int32_t* output) */
EXPORT void default_i32_approxdecomp_from_tndbl_ref(const MOD_Z* module, // N
const TNDBL_APPROXDECOMP_GADGET* gadget, // gadget
int32_t* res,
uint64_t res_size, // res (in general, size ell.a_size)
const double* a, uint64_t a_size); // a
/** @brief prepares a vmp matrix (contiguous row-major version) */
EXPORT void default_zn32_vmp_prepare_contiguous_ref( //
const MOD_Z* module,
ZN32_VMP_PMAT* pmat, // output
const int32_t* mat, uint64_t nrows, uint64_t ncols // a
);
/** @brief prepares a vmp matrix (mat[row]+col*N points to the item) */
EXPORT void default_zn32_vmp_prepare_dblptr_ref( //
const MOD_Z* module,
ZN32_VMP_PMAT* pmat, // output
const int32_t** mat, uint64_t nrows, uint64_t ncols // a
);
/** @brief prepares the ith-row of a vmp matrix with nrows and ncols */
EXPORT void default_zn32_vmp_prepare_row_ref( //
const MOD_Z* module,
ZN32_VMP_PMAT* pmat, // output
const int32_t* row, uint64_t row_i, uint64_t nrows, uint64_t ncols // a
);
/** @brief applies a vmp product (int32_t* input) */
EXPORT void default_zn32_vmp_apply_i32_ref( //
const MOD_Z* module, //
int32_t* res, uint64_t res_size, // res
const int32_t* a, uint64_t a_size, // a
const ZN32_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols); // prep matrix
/** @brief applies a vmp product (int16_t* input) */
EXPORT void default_zn32_vmp_apply_i16_ref( //
const MOD_Z* module, // N
int32_t* res, uint64_t res_size, // res
const int16_t* a, uint64_t a_size, // a
const ZN32_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols); // prep matrix
/** @brief applies a vmp product (int8_t* input) */
EXPORT void default_zn32_vmp_apply_i8_ref( //
const MOD_Z* module, // N
int32_t* res, uint64_t res_size, // res
const int8_t* a, uint64_t a_size, // a
const ZN32_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols); // prep matrix
/** @brief applies a vmp product (int32_t* input) */
EXPORT void default_zn32_vmp_apply_i32_avx( //
const MOD_Z* module, //
int32_t* res, uint64_t res_size, // res
const int32_t* a, uint64_t a_size, // a
const ZN32_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols); // prep matrix
/** @brief applies a vmp product (int16_t* input) */
EXPORT void default_zn32_vmp_apply_i16_avx( //
const MOD_Z* module, // N
int32_t* res, uint64_t res_size, // res
const int16_t* a, uint64_t a_size, // a
const ZN32_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols); // prep matrix
/** @brief applies a vmp product (int8_t* input) */
EXPORT void default_zn32_vmp_apply_i8_avx( //
const MOD_Z* module, // N
int32_t* res, uint64_t res_size, // res
const int8_t* a, uint64_t a_size, // a
const ZN32_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols); // prep matrix
// explicit conversions
/** reduction mod 1, output in torus32 space */
EXPORT void dbl_to_tn32_ref(const MOD_Z* module, //
int32_t* res, uint64_t res_size, // res
const double* a, uint64_t a_size // a
);
/** real centerlift mod 1, output in double space */
EXPORT void tn32_to_dbl_ref(const MOD_Z* module, //
double* res, uint64_t res_size, // res
const int32_t* a, uint64_t a_size // a
);
/** round to the nearest int, output in i32 space */
EXPORT void dbl_round_to_i32_ref(const MOD_Z* module, //
int32_t* res, uint64_t res_size, // res
const double* a, uint64_t a_size // a
);
/** small int (int32 space) to double */
EXPORT void i32_to_dbl_ref(const MOD_Z* module, //
double* res, uint64_t res_size, // res
const int32_t* a, uint64_t a_size // a
);
/** round to the nearest int, output in int64 space */
EXPORT void dbl_round_to_i64_ref(const MOD_Z* module, //
int64_t* res, uint64_t res_size, // res
const double* a, uint64_t a_size // a
);
/** small int (int64 space) to double */
EXPORT void i64_to_dbl_ref(const MOD_Z* module, //
double* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size // a
);
#endif // SPQLIOS_ZN_ARITHMETIC_PRIVATE_H

108
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/zn_conversions_ref.c
View File

@@ -1,108 +0,0 @@
#include <memory.h>
#include "zn_arithmetic_private.h"
typedef union {
double dv;
int64_t s64v;
int32_t s32v;
uint64_t u64v;
uint32_t u32v;
} di_t;
/** reduction mod 1, output in torus32 space */
EXPORT void dbl_to_tn32_ref(const MOD_Z* module, //
int32_t* res, uint64_t res_size, // res
const double* a, uint64_t a_size // a
) {
static const double ADD_CST = 0.5 + (double)(INT64_C(3) << (51 - 32));
static const int32_t XOR_CST = (INT32_C(1) << 31);
const uint64_t msize = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < msize; ++i) {
di_t t = {.dv = a[i] + ADD_CST};
res[i] = t.s32v ^ XOR_CST;
}
memset(res + msize, 0, (res_size - msize) * sizeof(int32_t));
}
/** real centerlift mod 1, output in double space */
EXPORT void tn32_to_dbl_ref(const MOD_Z* module, //
double* res, uint64_t res_size, // res
const int32_t* a, uint64_t a_size // a
) {
static const uint32_t XOR_CST = (UINT32_C(1) << 31);
static const di_t OR_CST = {.dv = (double)(INT64_C(1) << (52 - 32))};
static const double SUB_CST = 0.5 + (double)(INT64_C(1) << (52 - 32));
const uint64_t msize = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < msize; ++i) {
uint32_t ai = a[i] ^ XOR_CST;
di_t t = {.u64v = OR_CST.u64v | (uint64_t)ai};
res[i] = t.dv - SUB_CST;
}
memset(res + msize, 0, (res_size - msize) * sizeof(double));
}
/** round to the nearest int, output in i32 space */
EXPORT void dbl_round_to_i32_ref(const MOD_Z* module, //
int32_t* res, uint64_t res_size, // res
const double* a, uint64_t a_size // a
) {
static const double ADD_CST = (double)((INT64_C(3) << (51)) + (INT64_C(1) << (31)));
static const int32_t XOR_CST = INT32_C(1) << 31;
const uint64_t msize = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < msize; ++i) {
di_t t = {.dv = a[i] + ADD_CST};
res[i] = t.s32v ^ XOR_CST;
}
memset(res + msize, 0, (res_size - msize) * sizeof(int32_t));
}
/** small int (int32 space) to double */
EXPORT void i32_to_dbl_ref(const MOD_Z* module, //
double* res, uint64_t res_size, // res
const int32_t* a, uint64_t a_size // a
) {
static const uint32_t XOR_CST = (UINT32_C(1) << 31);
static const di_t OR_CST = {.dv = (double)(INT64_C(1) << 52)};
static const double SUB_CST = (double)((INT64_C(1) << 52) + (INT64_C(1) << 31));
const uint64_t msize = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < msize; ++i) {
uint32_t ai = a[i] ^ XOR_CST;
di_t t = {.u64v = OR_CST.u64v | (uint64_t)ai};
res[i] = t.dv - SUB_CST;
}
memset(res + msize, 0, (res_size - msize) * sizeof(double));
}
/** round to the nearest int, output in int64 space */
EXPORT void dbl_round_to_i64_ref(const MOD_Z* module, //
int64_t* res, uint64_t res_size, // res
const double* a, uint64_t a_size // a
) {
static const double ADD_CST = (double)(INT64_C(3) << (51));
static const int64_t AND_CST = (INT64_C(1) << 52) - 1;
static const int64_t SUB_CST = INT64_C(1) << 51;
const uint64_t msize = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < msize; ++i) {
di_t t = {.dv = a[i] + ADD_CST};
res[i] = (t.s64v & AND_CST) - SUB_CST;
}
memset(res + msize, 0, (res_size - msize) * sizeof(int64_t));
}
/** small int (int64 space) to double */
EXPORT void i64_to_dbl_ref(const MOD_Z* module, //
double* res, uint64_t res_size, // res
const int64_t* a, uint64_t a_size // a
) {
static const uint64_t ADD_CST = UINT64_C(1) << 51;
static const uint64_t AND_CST = (UINT64_C(1) << 52) - 1;
static const di_t OR_CST = {.dv = (INT64_C(1) << 52)};
static const double SUB_CST = INT64_C(3) << 51;
const uint64_t msize = res_size < a_size ? res_size : a_size;
for (uint64_t i = 0; i < msize; ++i) {
di_t t = {.u64v = ((a[i] + ADD_CST) & AND_CST) | OR_CST.u64v};
res[i] = t.dv - SUB_CST;
}
memset(res + msize, 0, (res_size - msize) * sizeof(double));
}

4
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/zn_vmp_int16_avx.c
View File

@@ -1,4 +0,0 @@
#define INTTYPE int16_t
#define INTSN i16
#include "zn_vmp_int32_avx.c"

4
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/zn_vmp_int16_ref.c
View File

@@ -1,4 +0,0 @@
#define INTTYPE int16_t
#define INTSN i16
#include "zn_vmp_int32_ref.c"

223
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/zn_vmp_int32_avx.c
View File

@@ -1,223 +0,0 @@
// This file is actually a template: it will be compiled multiple times with
// different INTTYPES
#ifndef INTTYPE
#define INTTYPE int32_t
#define INTSN i32
#endif
#include <immintrin.h>
#include <memory.h>
#include "zn_arithmetic_private.h"
#define concat_inner(aa, bb, cc) aa##_##bb##_##cc
#define concat(aa, bb, cc) concat_inner(aa, bb, cc)
#define zn32_vec_fn(cc) concat(zn32_vec, INTSN, cc)
static void zn32_vec_mat32cols_avx_prefetch(uint64_t nrows, int32_t* res, const INTTYPE* a, const int32_t* b) {
if (nrows == 0) {
memset(res, 0, 32 * sizeof(int32_t));
return;
}
const int32_t* bb = b;
const int32_t* pref_bb = b;
const uint64_t pref_iters = 128;
const uint64_t pref_start = pref_iters < nrows ? pref_iters : nrows;
const uint64_t pref_last = pref_iters > nrows ? 0 : nrows - pref_iters;
// let's do some prefetching of the GSW key, since on some cpus,
// it helps
for (uint64_t i = 0; i < pref_start; ++i) {
__builtin_prefetch(pref_bb, 0, _MM_HINT_T0);
__builtin_prefetch(pref_bb + 16, 0, _MM_HINT_T0);
pref_bb += 32;
}
// we do the first iteration
__m256i x = _mm256_set1_epi32(a[0]);
__m256i r0 = _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb)));
__m256i r1 = _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb + 8)));
__m256i r2 = _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb + 16)));
__m256i r3 = _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb + 24)));
bb += 32;
uint64_t row = 1;
for (; //
row < pref_last; //
++row, bb += 32) {
// prefetch the next iteration
__builtin_prefetch(pref_bb, 0, _MM_HINT_T0);
__builtin_prefetch(pref_bb + 16, 0, _MM_HINT_T0);
pref_bb += 32;
INTTYPE ai = a[row];
if (ai == 0) continue;
x = _mm256_set1_epi32(ai);
r0 = _mm256_add_epi32(r0, _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb))));
r1 = _mm256_add_epi32(r1, _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb + 8))));
r2 = _mm256_add_epi32(r2, _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb + 16))));
r3 = _mm256_add_epi32(r3, _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb + 24))));
}
for (; //
row < nrows; //
++row, bb += 32) {
INTTYPE ai = a[row];
if (ai == 0) continue;
x = _mm256_set1_epi32(ai);
r0 = _mm256_add_epi32(r0, _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb))));
r1 = _mm256_add_epi32(r1, _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb + 8))));
r2 = _mm256_add_epi32(r2, _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb + 16))));
r3 = _mm256_add_epi32(r3, _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb + 24))));
}
_mm256_storeu_si256((__m256i*)(res), r0);
_mm256_storeu_si256((__m256i*)(res + 8), r1);
_mm256_storeu_si256((__m256i*)(res + 16), r2);
_mm256_storeu_si256((__m256i*)(res + 24), r3);
}
void zn32_vec_fn(mat32cols_avx)(uint64_t nrows, int32_t* res, const INTTYPE* a, const int32_t* b, uint64_t b_sl) {
if (nrows == 0) {
memset(res, 0, 32 * sizeof(int32_t));
return;
}
const INTTYPE* aa = a;
const INTTYPE* const aaend = a + nrows;
const int32_t* bb = b;
__m256i x = _mm256_set1_epi32(*aa);
__m256i r0 = _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb)));
__m256i r1 = _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb + 8)));
__m256i r2 = _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb + 16)));
__m256i r3 = _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb + 24)));
bb += b_sl;
++aa;
for (; //
aa < aaend; //
bb += b_sl, ++aa) {
INTTYPE ai = *aa;
if (ai == 0) continue;
x = _mm256_set1_epi32(ai);
r0 = _mm256_add_epi32(r0, _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb))));
r1 = _mm256_add_epi32(r1, _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb + 8))));
r2 = _mm256_add_epi32(r2, _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb + 16))));
r3 = _mm256_add_epi32(r3, _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb + 24))));
}
_mm256_storeu_si256((__m256i*)(res), r0);
_mm256_storeu_si256((__m256i*)(res + 8), r1);
_mm256_storeu_si256((__m256i*)(res + 16), r2);
_mm256_storeu_si256((__m256i*)(res + 24), r3);
}
void zn32_vec_fn(mat24cols_avx)(uint64_t nrows, int32_t* res, const INTTYPE* a, const int32_t* b, uint64_t b_sl) {
if (nrows == 0) {
memset(res, 0, 24 * sizeof(int32_t));
return;
}
const INTTYPE* aa = a;
const INTTYPE* const aaend = a + nrows;
const int32_t* bb = b;
__m256i x = _mm256_set1_epi32(*aa);
__m256i r0 = _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb)));
__m256i r1 = _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb + 8)));
__m256i r2 = _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb + 16)));
bb += b_sl;
++aa;
for (; //
aa < aaend; //
bb += b_sl, ++aa) {
INTTYPE ai = *aa;
if (ai == 0) continue;
x = _mm256_set1_epi32(ai);
r0 = _mm256_add_epi32(r0, _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb))));
r1 = _mm256_add_epi32(r1, _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb + 8))));
r2 = _mm256_add_epi32(r2, _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb + 16))));
}
_mm256_storeu_si256((__m256i*)(res), r0);
_mm256_storeu_si256((__m256i*)(res + 8), r1);
_mm256_storeu_si256((__m256i*)(res + 16), r2);
}
void zn32_vec_fn(mat16cols_avx)(uint64_t nrows, int32_t* res, const INTTYPE* a, const int32_t* b, uint64_t b_sl) {
if (nrows == 0) {
memset(res, 0, 16 * sizeof(int32_t));
return;
}
const INTTYPE* aa = a;
const INTTYPE* const aaend = a + nrows;
const int32_t* bb = b;
__m256i x = _mm256_set1_epi32(*aa);
__m256i r0 = _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb)));
__m256i r1 = _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb + 8)));
bb += b_sl;
++aa;
for (; //
aa < aaend; //
bb += b_sl, ++aa) {
INTTYPE ai = *aa;
if (ai == 0) continue;
x = _mm256_set1_epi32(ai);
r0 = _mm256_add_epi32(r0, _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb))));
r1 = _mm256_add_epi32(r1, _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb + 8))));
}
_mm256_storeu_si256((__m256i*)(res), r0);
_mm256_storeu_si256((__m256i*)(res + 8), r1);
}
void zn32_vec_fn(mat8cols_avx)(uint64_t nrows, int32_t* res, const INTTYPE* a, const int32_t* b, uint64_t b_sl) {
if (nrows == 0) {
memset(res, 0, 8 * sizeof(int32_t));
return;
}
const INTTYPE* aa = a;
const INTTYPE* const aaend = a + nrows;
const int32_t* bb = b;
__m256i x = _mm256_set1_epi32(*aa);
__m256i r0 = _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb)));
bb += b_sl;
++aa;
for (; //
aa < aaend; //
bb += b_sl, ++aa) {
INTTYPE ai = *aa;
if (ai == 0) continue;
x = _mm256_set1_epi32(ai);
r0 = _mm256_add_epi32(r0, _mm256_mullo_epi32(x, _mm256_loadu_si256((__m256i*)(bb))));
}
_mm256_storeu_si256((__m256i*)(res), r0);
}
typedef void (*vm_f)(uint64_t nrows, //
int32_t* res, //
const INTTYPE* a, //
const int32_t* b, uint64_t b_sl //
);
static const vm_f zn32_vec_mat8kcols_avx[4] = { //
zn32_vec_fn(mat8cols_avx), //
zn32_vec_fn(mat16cols_avx), //
zn32_vec_fn(mat24cols_avx), //
zn32_vec_fn(mat32cols_avx)};
/** @brief applies a vmp product (int32_t* input) */
EXPORT void concat(default_zn32_vmp_apply, INTSN, avx)( //
const MOD_Z* module, //
int32_t* res, uint64_t res_size, //
const INTTYPE* a, uint64_t a_size, //
const ZN32_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols) {
const uint64_t rows = a_size < nrows ? a_size : nrows;
const uint64_t cols = res_size < ncols ? res_size : ncols;
const uint64_t ncolblk = cols >> 5;
const uint64_t ncolrem = cols & 31;
// copy the first full blocks
const uint64_t full_blk_size = nrows * 32;
const int32_t* mat = (int32_t*)pmat;
int32_t* rr = res;
for (uint64_t blk = 0; //
blk < ncolblk; //
++blk, mat += full_blk_size, rr += 32) {
zn32_vec_mat32cols_avx_prefetch(rows, rr, a, mat);
}
// last block
if (ncolrem) {
uint64_t orig_rem = ncols - (ncolblk << 5);
uint64_t b_sl = orig_rem >= 32 ? 32 : orig_rem;
int32_t tmp[32];
zn32_vec_mat8kcols_avx[(ncolrem - 1) >> 3](rows, tmp, a, mat, b_sl);
memcpy(rr, tmp, ncolrem * sizeof(int32_t));
}
// trailing bytes
memset(res + cols, 0, (res_size - cols) * sizeof(int32_t));
}

88
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/zn_vmp_int32_ref.c
View File

@@ -1,88 +0,0 @@
// This file is actually a template: it will be compiled multiple times with
// different INTTYPES
#ifndef INTTYPE
#define INTTYPE int32_t
#define INTSN i32
#endif
#include <memory.h>
#include "zn_arithmetic_private.h"
#define concat_inner(aa, bb, cc) aa##_##bb##_##cc
#define concat(aa, bb, cc) concat_inner(aa, bb, cc)
#define zn32_vec_fn(cc) concat(zn32_vec, INTSN, cc)
// the ref version shares the same implementation for each fixed column size
// optimized implementations may do something different.
static __always_inline void IMPL_zn32_vec_matcols_ref(
const uint64_t NCOLS, // fixed number of columns
uint64_t nrows, // nrows of b
int32_t* res, // result: size NCOLS, only the first min(b_sl, NCOLS) are relevant
const INTTYPE* a, // a: nrows-sized vector
const int32_t* b, uint64_t b_sl // b: nrows * min(b_sl, NCOLS) matrix
) {
memset(res, 0, NCOLS * sizeof(int32_t));
for (uint64_t row = 0; row < nrows; ++row) {
int32_t ai = a[row];
const int32_t* bb = b + row * b_sl;
for (uint64_t i = 0; i < NCOLS; ++i) {
res[i] += ai * bb[i];
}
}
}
void zn32_vec_fn(mat32cols_ref)(uint64_t nrows, int32_t* res, const INTTYPE* a, const int32_t* b, uint64_t b_sl) {
IMPL_zn32_vec_matcols_ref(32, nrows, res, a, b, b_sl);
}
void zn32_vec_fn(mat24cols_ref)(uint64_t nrows, int32_t* res, const INTTYPE* a, const int32_t* b, uint64_t b_sl) {
IMPL_zn32_vec_matcols_ref(24, nrows, res, a, b, b_sl);
}
void zn32_vec_fn(mat16cols_ref)(uint64_t nrows, int32_t* res, const INTTYPE* a, const int32_t* b, uint64_t b_sl) {
IMPL_zn32_vec_matcols_ref(16, nrows, res, a, b, b_sl);
}
void zn32_vec_fn(mat8cols_ref)(uint64_t nrows, int32_t* res, const INTTYPE* a, const int32_t* b, uint64_t b_sl) {
IMPL_zn32_vec_matcols_ref(8, nrows, res, a, b, b_sl);
}
typedef void (*vm_f)(uint64_t nrows, //
int32_t* res, //
const INTTYPE* a, //
const int32_t* b, uint64_t b_sl //
);
static const vm_f zn32_vec_mat8kcols_ref[4] = { //
zn32_vec_fn(mat8cols_ref), //
zn32_vec_fn(mat16cols_ref), //
zn32_vec_fn(mat24cols_ref), //
zn32_vec_fn(mat32cols_ref)};
/** @brief applies a vmp product (int32_t* input) */
EXPORT void concat(default_zn32_vmp_apply, INTSN, ref)( //
const MOD_Z* module, //
int32_t* res, uint64_t res_size, //
const INTTYPE* a, uint64_t a_size, //
const ZN32_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols) {
const uint64_t rows = a_size < nrows ? a_size : nrows;
const uint64_t cols = res_size < ncols ? res_size : ncols;
const uint64_t ncolblk = cols >> 5;
const uint64_t ncolrem = cols & 31;
// copy the first full blocks
const uint32_t full_blk_size = nrows * 32;
const int32_t* mat = (int32_t*)pmat;
int32_t* rr = res;
for (uint64_t blk = 0; //
blk < ncolblk; //
++blk, mat += full_blk_size, rr += 32) {
zn32_vec_fn(mat32cols_ref)(rows, rr, a, mat, 32);
}
// last block
if (ncolrem) {
uint64_t orig_rem = ncols - (ncolblk << 5);
uint64_t b_sl = orig_rem >= 32 ? 32 : orig_rem;
int32_t tmp[32];
zn32_vec_mat8kcols_ref[(ncolrem - 1) >> 3](rows, tmp, a, mat, b_sl);
memcpy(rr, tmp, ncolrem * sizeof(int32_t));
}
// trailing bytes
memset(res + cols, 0, (res_size - cols) * sizeof(int32_t));
}

4
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/zn_vmp_int8_avx.c
View File

@@ -1,4 +0,0 @@
#define INTTYPE int8_t
#define INTSN i8
#include "zn_vmp_int32_avx.c"

4
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/zn_vmp_int8_ref.c
View File

@@ -1,4 +0,0 @@
#define INTTYPE int8_t
#define INTSN i8
#include "zn_vmp_int32_ref.c"

185
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/zn_vmp_ref.c
View File

@@ -1,185 +0,0 @@
#include <memory.h>
#include "zn_arithmetic_private.h"
/** @brief size in bytes of a prepared matrix (for custom allocation) */
EXPORT uint64_t bytes_of_zn32_vmp_pmat(const MOD_Z* module, // N
uint64_t nrows, uint64_t ncols // dimensions
) {
return (nrows * ncols + 7) * sizeof(int32_t);
}
/** @brief allocates a prepared matrix (release with delete_zn32_vmp_pmat) */
EXPORT ZN32_VMP_PMAT* new_zn32_vmp_pmat(const MOD_Z* module, // N
uint64_t nrows, uint64_t ncols) {
return (ZN32_VMP_PMAT*)spqlios_alloc(bytes_of_zn32_vmp_pmat(module, nrows, ncols));
}
/** @brief deletes a prepared matrix (release with free) */
EXPORT void delete_zn32_vmp_pmat(ZN32_VMP_PMAT* ptr) { spqlios_free(ptr); }
/** @brief prepares a vmp matrix (contiguous row-major version) */
EXPORT void default_zn32_vmp_prepare_contiguous_ref( //
const MOD_Z* module,
ZN32_VMP_PMAT* pmat, // output
const int32_t* mat, uint64_t nrows, uint64_t ncols // a
) {
int32_t* const out = (int32_t*)pmat;
const uint64_t nblk = ncols >> 5;
const uint64_t ncols_rem = ncols & 31;
const uint64_t final_elems = (8 - nrows * ncols) & 7;
for (uint64_t blk = 0; blk < nblk; ++blk) {
int32_t* outblk = out + blk * nrows * 32;
const int32_t* srcblk = mat + blk * 32;
for (uint64_t row = 0; row < nrows; ++row) {
int32_t* dest = outblk + row * 32;
const int32_t* src = srcblk + row * ncols;
for (uint64_t i = 0; i < 32; ++i) {
dest[i] = src[i];
}
}
}
// copy the last block if any
if (ncols_rem) {
int32_t* outblk = out + nblk * nrows * 32;
const int32_t* srcblk = mat + nblk * 32;
for (uint64_t row = 0; row < nrows; ++row) {
int32_t* dest = outblk + row * ncols_rem;
const int32_t* src = srcblk + row * ncols;
for (uint64_t i = 0; i < ncols_rem; ++i) {
dest[i] = src[i];
}
}
}
// zero-out the final elements that may be accessed
if (final_elems) {
int32_t* f = out + nrows * ncols;
for (uint64_t i = 0; i < final_elems; ++i) {
f[i] = 0;
}
}
}
/** @brief prepares a vmp matrix (mat[row]+col*N points to the item) */
EXPORT void default_zn32_vmp_prepare_dblptr_ref( //
const MOD_Z* module,
ZN32_VMP_PMAT* pmat, // output
const int32_t** mat, uint64_t nrows, uint64_t ncols // a
) {
for (uint64_t row_i = 0; row_i < nrows; ++row_i) {
default_zn32_vmp_prepare_row_ref(module, pmat, mat[row_i], row_i, nrows, ncols);
}
}
/** @brief prepares the ith-row of a vmp matrix with nrows and ncols */
EXPORT void default_zn32_vmp_prepare_row_ref( //
const MOD_Z* module,
ZN32_VMP_PMAT* pmat, // output
const int32_t* row, uint64_t row_i, uint64_t nrows, uint64_t ncols // a
) {
int32_t* const out = (int32_t*)pmat;
const uint64_t nblk = ncols >> 5;
const uint64_t ncols_rem = ncols & 31;
const uint64_t final_elems = (row_i == nrows - 1) && (8 - nrows * ncols) & 7;
for (uint64_t blk = 0; blk < nblk; ++blk) {
int32_t* outblk = out + blk * nrows * 32;
int32_t* dest = outblk + row_i * 32;
const int32_t* src = row + blk * 32;
for (uint64_t i = 0; i < 32; ++i) {
dest[i] = src[i];
}
}
// copy the last block if any
if (ncols_rem) {
int32_t* outblk = out + nblk * nrows * 32;
int32_t* dest = outblk + row_i * ncols_rem;
const int32_t* src = row + nblk * 32;
for (uint64_t i = 0; i < ncols_rem; ++i) {
dest[i] = src[i];
}
}
// zero-out the final elements that may be accessed
if (final_elems) {
int32_t* f = out + nrows * ncols;
for (uint64_t i = 0; i < final_elems; ++i) {
f[i] = 0;
}
}
}
#if 0
#define IMPL_zn32_vec_ixxx_matyyycols_ref(NCOLS) \
memset(res, 0, NCOLS * sizeof(int32_t)); \
for (uint64_t row = 0; row < nrows; ++row) { \
int32_t ai = a[row]; \
const int32_t* bb = b + row * b_sl; \
for (uint64_t i = 0; i < NCOLS; ++i) { \
res[i] += ai * bb[i]; \
} \
}
#define IMPL_zn32_vec_ixxx_mat8cols_ref() IMPL_zn32_vec_ixxx_matyyycols_ref(8)
#define IMPL_zn32_vec_ixxx_mat16cols_ref() IMPL_zn32_vec_ixxx_matyyycols_ref(16)
#define IMPL_zn32_vec_ixxx_mat24cols_ref() IMPL_zn32_vec_ixxx_matyyycols_ref(24)
#define IMPL_zn32_vec_ixxx_mat32cols_ref() IMPL_zn32_vec_ixxx_matyyycols_ref(32)
void zn32_vec_i8_mat32cols_ref(uint64_t nrows, int32_t* res, const int8_t* a, const int32_t* b, uint64_t b_sl) {
IMPL_zn32_vec_ixxx_mat32cols_ref()
}
void zn32_vec_i16_mat32cols_ref(uint64_t nrows, int32_t* res, const int16_t* a, const int32_t* b, uint64_t b_sl) {
IMPL_zn32_vec_ixxx_mat32cols_ref()
}
void zn32_vec_i32_mat32cols_ref(uint64_t nrows, int32_t* res, const int32_t* a, const int32_t* b, uint64_t b_sl) {
IMPL_zn32_vec_ixxx_mat32cols_ref()
}
void zn32_vec_i32_mat24cols_ref(uint64_t nrows, int32_t* res, const int32_t* a, const int32_t* b, uint64_t b_sl) {
IMPL_zn32_vec_ixxx_mat24cols_ref()
}
void zn32_vec_i32_mat16cols_ref(uint64_t nrows, int32_t* res, const int32_t* a, const int32_t* b, uint64_t b_sl) {
IMPL_zn32_vec_ixxx_mat16cols_ref()
}
void zn32_vec_i32_mat8cols_ref(uint64_t nrows, int32_t* res, const int32_t* a, const int32_t* b, uint64_t b_sl) {
IMPL_zn32_vec_ixxx_mat8cols_ref()
}
typedef void (*zn32_vec_i32_mat8kcols_ref_f)(uint64_t nrows, //
int32_t* res, //
const int32_t* a, //
const int32_t* b, uint64_t b_sl //
);
zn32_vec_i32_mat8kcols_ref_f zn32_vec_i32_mat8kcols_ref[4] = { //
zn32_vec_i32_mat8cols_ref, zn32_vec_i32_mat16cols_ref, //
zn32_vec_i32_mat24cols_ref, zn32_vec_i32_mat32cols_ref};
/** @brief applies a vmp product (int32_t* input) */
EXPORT void default_zn32_vmp_apply_i32_ref(const MOD_Z* module, //
int32_t* res, uint64_t res_size, //
const int32_t* a, uint64_t a_size, //
const ZN32_VMP_PMAT* pmat, uint64_t nrows, uint64_t ncols) {
const uint64_t rows = a_size < nrows ? a_size : nrows;
const uint64_t cols = res_size < ncols ? res_size : ncols;
const uint64_t ncolblk = cols >> 5;
const uint64_t ncolrem = cols & 31;
// copy the first full blocks
const uint32_t full_blk_size = nrows * 32;
const int32_t* mat = (int32_t*)pmat;
int32_t* rr = res;
for (uint64_t blk = 0; //
blk < ncolblk; //
++blk, mat += full_blk_size, rr += 32) {
zn32_vec_i32_mat32cols_ref(rows, rr, a, mat, 32);
}
// last block
if (ncolrem) {
uint64_t orig_rem = ncols - (ncolblk << 5);
uint64_t b_sl = orig_rem >= 32 ? 32 : orig_rem;
int32_t tmp[32];
zn32_vec_i32_mat8kcols_ref[(ncolrem - 1) >> 3](rows, tmp, a, mat, b_sl);
memcpy(rr, tmp, ncolrem * sizeof(int32_t));
}
// trailing bytes
memset(res + cols, 0, (res_size - cols) * sizeof(int32_t));
}
#endif

38
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/arithmetic/znx_small.c
View File

@@ -1,38 +0,0 @@
#include "vec_znx_arithmetic_private.h"
/** @brief res = a * b : small integer polynomial product */
EXPORT void fft64_znx_small_single_product(const MODULE* module, // N
int64_t* res, // output
const int64_t* a, // a
const int64_t* b, // b
uint8_t* tmp) {
const uint64_t nn = module->nn;
double* const ffta = (double*)tmp;
double* const fftb = ((double*)tmp) + nn;
reim_from_znx64(module->mod.fft64.p_conv, ffta, a);
reim_from_znx64(module->mod.fft64.p_conv, fftb, b);
reim_fft(module->mod.fft64.p_fft, ffta);
reim_fft(module->mod.fft64.p_fft, fftb);
reim_fftvec_mul_simple(module->m, ffta, ffta, fftb);
reim_ifft(module->mod.fft64.p_ifft, ffta);
reim_to_znx64(module->mod.fft64.p_reim_to_znx, res, ffta);
}
/** @brief tmp bytes required for znx_small_single_product */
EXPORT uint64_t fft64_znx_small_single_product_tmp_bytes(const MODULE* module, uint64_t nn) {
return 2 * nn * sizeof(double);
}
/** @brief res = a * b : small integer polynomial product */
EXPORT void znx_small_single_product(const MODULE* module, // N
int64_t* res, // output
const int64_t* a, // a
const int64_t* b, // b
uint8_t* tmp) {
module->func.znx_small_single_product(module, res, a, b, tmp);
}
/** @brief tmp bytes required for znx_small_single_product */
EXPORT uint64_t znx_small_single_product_tmp_bytes(const MODULE* module, uint64_t nn) {
return module->func.znx_small_single_product_tmp_bytes(module, nn);
}

524
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/coeffs/coeffs_arithmetic.c
View File

@@ -1,524 +0,0 @@
#include "coeffs_arithmetic.h"
#include <assert.h>
#include <memory.h>
/** res = a + b */
EXPORT void znx_add_i64_ref(uint64_t nn, int64_t* res, const int64_t* a, const int64_t* b) {
for (uint64_t i = 0; i < nn; ++i) {
res[i] = a[i] + b[i];
}
}
/** res = a - b */
EXPORT void znx_sub_i64_ref(uint64_t nn, int64_t* res, const int64_t* a, const int64_t* b) {
for (uint64_t i = 0; i < nn; ++i) {
res[i] = a[i] - b[i];
}
}
EXPORT void znx_negate_i64_ref(uint64_t nn, int64_t* res, const int64_t* a) {
for (uint64_t i = 0; i < nn; ++i) {
res[i] = -a[i];
}
}
EXPORT void znx_copy_i64_ref(uint64_t nn, int64_t* res, const int64_t* a) { memcpy(res, a, nn * sizeof(int64_t)); }
EXPORT void znx_zero_i64_ref(uint64_t nn, int64_t* res) { memset(res, 0, nn * sizeof(int64_t)); }
EXPORT void rnx_divide_by_m_ref(uint64_t n, double m, double* res, const double* a) {
const double invm = 1. / m;
for (uint64_t i = 0; i < n; ++i) {
res[i] = a[i] * invm;
}
}
EXPORT void rnx_rotate_f64(uint64_t nn, int64_t p, double* res, const double* in) {
uint64_t a = (-p) & (2 * nn - 1); // a= (-p) (pos)mod (2*nn)
if (a < nn) { // rotate to the left
uint64_t nma = nn - a;
// rotate first half
for (uint64_t j = 0; j < nma; j++) {
res[j] = in[j + a];
}
for (uint64_t j = nma; j < nn; j++) {
res[j] = -in[j - nma];
}
} else {
a -= nn;
uint64_t nma = nn - a;
for (uint64_t j = 0; j < nma; j++) {
res[j] = -in[j + a];
}
for (uint64_t j = nma; j < nn; j++) {
// rotate first half
res[j] = in[j - nma];
}
}
}
EXPORT void znx_rotate_i64(uint64_t nn, int64_t p, int64_t* res, const int64_t* in) {
uint64_t a = (-p) & (2 * nn - 1); // a= (-p) (pos)mod (2*nn)
if (a < nn) { // rotate to the left
uint64_t nma = nn - a;
// rotate first half
for (uint64_t j = 0; j < nma; j++) {
res[j] = in[j + a];
}
for (uint64_t j = nma; j < nn; j++) {
res[j] = -in[j - nma];
}
} else {
a -= nn;
uint64_t nma = nn - a;
for (uint64_t j = 0; j < nma; j++) {
res[j] = -in[j + a];
}
for (uint64_t j = nma; j < nn; j++) {
// rotate first half
res[j] = in[j - nma];
}
}
}
EXPORT void rnx_mul_xp_minus_one_f64(uint64_t nn, int64_t p, double* res, const double* in) {
uint64_t a = (-p) & (2 * nn - 1); // a= (-p) (pos)mod (2*nn)
if (a < nn) { // rotate to the left
uint64_t nma = nn - a;
// rotate first half
for (uint64_t j = 0; j < nma; j++) {
res[j] = in[j + a] - in[j];
}
for (uint64_t j = nma; j < nn; j++) {
res[j] = -in[j - nma] - in[j];
}
} else {
a -= nn;
uint64_t nma = nn - a;
for (uint64_t j = 0; j < nma; j++) {
res[j] = -in[j + a] - in[j];
}
for (uint64_t j = nma; j < nn; j++) {
// rotate first half
res[j] = in[j - nma] - in[j];
}
}
}
EXPORT void znx_mul_xp_minus_one_i64(uint64_t nn, int64_t p, int64_t* res, const int64_t* in) {
uint64_t a = (-p) & (2 * nn - 1); // a= (-p) (pos)mod (2*nn)
if (a < nn) { // rotate to the left
uint64_t nma = nn - a;
// rotate first half
for (uint64_t j = 0; j < nma; j++) {
res[j] = in[j + a] - in[j];
}
for (uint64_t j = nma; j < nn; j++) {
res[j] = -in[j - nma] - in[j];
}
} else {
a -= nn;
uint64_t nma = nn - a;
for (uint64_t j = 0; j < nma; j++) {
res[j] = -in[j + a] - in[j];
}
for (uint64_t j = nma; j < nn; j++) {
// rotate first half
res[j] = in[j - nma] - in[j];
}
}
}
EXPORT void znx_mul_xp_minus_one_inplace_i64(uint64_t nn, int64_t p, int64_t* res) {
const uint64_t _2mn = 2 * nn - 1;
const uint64_t _mn = nn - 1;
uint64_t nb_modif = 0;
uint64_t j_start = 0;
while (nb_modif < nn) {
// follow the cycle that start with j_start
uint64_t j = j_start;
int64_t tmp1 = res[j];
do {
// find where the value should go, and with which sign
uint64_t new_j = (j + p) & _2mn; // mod 2n to get the position and sign
uint64_t new_j_n = new_j & _mn; // mod n to get just the position
// exchange this position with tmp1 (and take care of the sign)
int64_t tmp2 = res[new_j_n];
res[new_j_n] = ((new_j < nn) ? tmp1 : -tmp1) - res[new_j_n];
tmp1 = tmp2;
// move to the new location, and store the number of items modified
++nb_modif;
j = new_j_n;
} while (j != j_start);
// move to the start of the next cycle:
// we need to find an index that has not been touched yet, and pick it as next j_start.
// in practice, it is enough to do +1, because the group of rotations is cyclic and 1 is a generator.
++j_start;
}
}
// 0 < p < 2nn
EXPORT void rnx_automorphism_f64(uint64_t nn, int64_t p, double* res, const double* in) {
res[0] = in[0];
uint64_t a = 0;
uint64_t _2mn = 2 * nn - 1;
for (uint64_t i = 1; i < nn; i++) {
a = (a + p) & _2mn; // i*p mod 2n
if (a < nn) {
res[a] = in[i]; // res[ip mod 2n] = res[i]
} else {
res[a - nn] = -in[i];
}
}
}
EXPORT void znx_automorphism_i64(uint64_t nn, int64_t p, int64_t* res, const int64_t* in) {
res[0] = in[0];
uint64_t a = 0;
uint64_t _2mn = 2 * nn - 1;
for (uint64_t i = 1; i < nn; i++) {
a = (a + p) & _2mn;
if (a < nn) {
res[a] = in[i]; // res[ip mod 2n] = res[i]
} else {
res[a - nn] = -in[i];
}
}
}
EXPORT void rnx_rotate_inplace_f64(uint64_t nn, int64_t p, double* res) {
const uint64_t _2mn = 2 * nn - 1;
const uint64_t _mn = nn - 1;
uint64_t nb_modif = 0;
uint64_t j_start = 0;
while (nb_modif < nn) {
// follow the cycle that start with j_start
uint64_t j = j_start;
double tmp1 = res[j];
do {
// find where the value should go, and with which sign
uint64_t new_j = (j + p) & _2mn; // mod 2n to get the position and sign
uint64_t new_j_n = new_j & _mn; // mod n to get just the position
// exchange this position with tmp1 (and take care of the sign)
double tmp2 = res[new_j_n];
res[new_j_n] = (new_j < nn) ? tmp1 : -tmp1;
tmp1 = tmp2;
// move to the new location, and store the number of items modified
++nb_modif;
j = new_j_n;
} while (j != j_start);
// move to the start of the next cycle:
// we need to find an index that has not been touched yet, and pick it as next j_start.
// in practice, it is enough to do +1, because the group of rotations is cyclic and 1 is a generator.
++j_start;
}
}
EXPORT void znx_rotate_inplace_i64(uint64_t nn, int64_t p, int64_t* res) {
const uint64_t _2mn = 2 * nn - 1;
const uint64_t _mn = nn - 1;
uint64_t nb_modif = 0;
uint64_t j_start = 0;
while (nb_modif < nn) {
// follow the cycle that start with j_start
uint64_t j = j_start;
int64_t tmp1 = res[j];
do {
// find where the value should go, and with which sign
uint64_t new_j = (j + p) & _2mn; // mod 2n to get the position and sign
uint64_t new_j_n = new_j & _mn; // mod n to get just the position
// exchange this position with tmp1 (and take care of the sign)
int64_t tmp2 = res[new_j_n];
res[new_j_n] = (new_j < nn) ? tmp1 : -tmp1;
tmp1 = tmp2;
// move to the new location, and store the number of items modified
++nb_modif;
j = new_j_n;
} while (j != j_start);
// move to the start of the next cycle:
// we need to find an index that has not been touched yet, and pick it as next j_start.
// in practice, it is enough to do +1, because the group of rotations is cyclic and 1 is a generator.
++j_start;
}
}
EXPORT void rnx_mul_xp_minus_one_inplace_f64(uint64_t nn, int64_t p, double* res) {
const uint64_t _2mn = 2 * nn - 1;
const uint64_t _mn = nn - 1;
uint64_t nb_modif = 0;
uint64_t j_start = 0;
while (nb_modif < nn) {
// follow the cycle that start with j_start
uint64_t j = j_start;
double tmp1 = res[j];
do {
// find where the value should go, and with which sign
uint64_t new_j = (j + p) & _2mn; // mod 2n to get the position and sign
uint64_t new_j_n = new_j & _mn; // mod n to get just the position
// exchange this position with tmp1 (and take care of the sign)
double tmp2 = res[new_j_n];
res[new_j_n] = ((new_j < nn) ? tmp1 : -tmp1) - res[new_j_n];
tmp1 = tmp2;
// move to the new location, and store the number of items modified
++nb_modif;
j = new_j_n;
} while (j != j_start);
// move to the start of the next cycle:
// we need to find an index that has not been touched yet, and pick it as next j_start.
// in practice, it is enough to do +1, because the group of rotations is cyclic and 1 is a generator.
++j_start;
}
}
__always_inline int64_t get_base_k_digit(const int64_t x, const uint64_t base_k) {
return (x << (64 - base_k)) >> (64 - base_k);
}
__always_inline int64_t get_base_k_carry(const int64_t x, const int64_t digit, const uint64_t base_k) {
return (x - digit) >> base_k;
}
EXPORT void znx_normalize(uint64_t nn, uint64_t base_k, int64_t* out, int64_t* carry_out, const int64_t* in,
const int64_t* carry_in) {
assert(in);
if (out != 0) {
if (carry_in != 0x0 && carry_out != 0x0) {
// with carry in and carry out is computed
for (uint64_t i = 0; i < nn; ++i) {
const int64_t x = in[i];
const int64_t cin = carry_in[i];
int64_t digit = get_base_k_digit(x, base_k);
int64_t carry = get_base_k_carry(x, digit, base_k);
int64_t digit_plus_cin = digit + cin;
int64_t y = get_base_k_digit(digit_plus_cin, base_k);
int64_t cout = carry + get_base_k_carry(digit_plus_cin, y, base_k);
out[i] = y;
carry_out[i] = cout;
}
} else if (carry_in != 0) {
// with carry in and carry out is dropped
for (uint64_t i = 0; i < nn; ++i) {
const int64_t x = in[i];
const int64_t cin = carry_in[i];
int64_t digit = get_base_k_digit(x, base_k);
int64_t digit_plus_cin = digit + cin;
int64_t y = get_base_k_digit(digit_plus_cin, base_k);
out[i] = y;
}
} else if (carry_out != 0) {
// no carry in and carry out is computed
for (uint64_t i = 0; i < nn; ++i) {
const int64_t x = in[i];
int64_t y = get_base_k_digit(x, base_k);
int64_t cout = get_base_k_carry(x, y, base_k);
out[i] = y;
carry_out[i] = cout;
}
} else {
// no carry in and carry out is dropped
for (uint64_t i = 0; i < nn; ++i) {
out[i] = get_base_k_digit(in[i], base_k);
}
}
} else {
assert(carry_out);
if (carry_in != 0x0) {
// with carry in and carry out is computed
for (uint64_t i = 0; i < nn; ++i) {
const int64_t x = in[i];
const int64_t cin = carry_in[i];
int64_t digit = get_base_k_digit(x, base_k);
int64_t carry = get_base_k_carry(x, digit, base_k);
int64_t digit_plus_cin = digit + cin;
int64_t y = get_base_k_digit(digit_plus_cin, base_k);
int64_t cout = carry + get_base_k_carry(digit_plus_cin, y, base_k);
carry_out[i] = cout;
}
} else {
// no carry in and carry out is computed
for (uint64_t i = 0; i < nn; ++i) {
const int64_t x = in[i];
int64_t y = get_base_k_digit(x, base_k);
int64_t cout = get_base_k_carry(x, y, base_k);
carry_out[i] = cout;
}
}
}
}
void znx_automorphism_inplace_i64(uint64_t nn, int64_t p, int64_t* res) {
const uint64_t _2mn = 2 * nn - 1;
const uint64_t _mn = nn - 1;
const uint64_t m = nn >> 1;
// reduce p mod 2n
p &= _2mn;
// uint64_t vp = p & _2mn;
/// uint64_t target_modifs = m >> 1;
// we proceed by increasing binary valuation
for (uint64_t binval = 1, vp = p & _2mn, orb_size = m; binval < nn;
binval <<= 1, vp = (vp << 1) & _2mn, orb_size >>= 1) {
// In this loop, we are going to treat the orbit of indexes = binval mod 2.binval.
// At the beginning of this loop we have:
// vp = binval * p mod 2n
// target_modif = m / binval (i.e. order of the orbit binval % 2.binval)
// first, handle the orders 1 and 2.
// if p*binval == binval % 2n: we're done!
if (vp == binval) return;
// if p*binval == -binval % 2n: nega-mirror the orbit and all the sub-orbits and exit!
if (((vp + binval) & _2mn) == 0) {
for (uint64_t j = binval; j < m; j += binval) {
int64_t tmp = res[j];
res[j] = -res[nn - j];
res[nn - j] = -tmp;
}
res[m] = -res[m];
return;
}
// if p*binval == binval + n % 2n: negate the orbit and exit
if (((vp - binval) & _mn) == 0) {
for (uint64_t j = binval; j < nn; j += 2 * binval) {
res[j] = -res[j];
}
return;
}
// if p*binval == n - binval % 2n: mirror the orbit and continue!
if (((vp + binval) & _mn) == 0) {
for (uint64_t j = binval; j < m; j += 2 * binval) {
int64_t tmp = res[j];
res[j] = res[nn - j];
res[nn - j] = tmp;
}
continue;
}
// otherwise we will follow the orbit cycles,
// starting from binval and -binval in parallel
uint64_t j_start = binval;
uint64_t nb_modif = 0;
while (nb_modif < orb_size) {
// follow the cycle that start with j_start
uint64_t j = j_start;
int64_t tmp1 = res[j];
int64_t tmp2 = res[nn - j];
do {
// find where the value should go, and with which sign
uint64_t new_j = (j * p) & _2mn; // mod 2n to get the position and sign
uint64_t new_j_n = new_j & _mn; // mod n to get just the position
// exchange this position with tmp1 (and take care of the sign)
int64_t tmp1a = res[new_j_n];
int64_t tmp2a = res[nn - new_j_n];
if (new_j < nn) {
res[new_j_n] = tmp1;
res[nn - new_j_n] = tmp2;
} else {
res[new_j_n] = -tmp1;
res[nn - new_j_n] = -tmp2;
}
tmp1 = tmp1a;
tmp2 = tmp2a;
// move to the new location, and store the number of items modified
nb_modif += 2;
j = new_j_n;
} while (j != j_start);
// move to the start of the next cycle:
// we need to find an index that has not been touched yet, and pick it as next j_start.
// in practice, it is enough to do *5, because 5 is a generator.
j_start = (5 * j_start) & _mn;
}
}
}
void rnx_automorphism_inplace_f64(uint64_t nn, int64_t p, double* res) {
const uint64_t _2mn = 2 * nn - 1;
const uint64_t _mn = nn - 1;
const uint64_t m = nn >> 1;
// reduce p mod 2n
p &= _2mn;
// uint64_t vp = p & _2mn;
/// uint64_t target_modifs = m >> 1;
// we proceed by increasing binary valuation
for (uint64_t binval = 1, vp = p & _2mn, orb_size = m; binval < nn;
binval <<= 1, vp = (vp << 1) & _2mn, orb_size >>= 1) {
// In this loop, we are going to treat the orbit of indexes = binval mod 2.binval.
// At the beginning of this loop we have:
// vp = binval * p mod 2n
// target_modif = m / binval (i.e. order of the orbit binval % 2.binval)
// first, handle the orders 1 and 2.
// if p*binval == binval % 2n: we're done!
if (vp == binval) return;
// if p*binval == -binval % 2n: nega-mirror the orbit and all the sub-orbits and exit!
if (((vp + binval) & _2mn) == 0) {
for (uint64_t j = binval; j < m; j += binval) {
double tmp = res[j];
res[j] = -res[nn - j];
res[nn - j] = -tmp;
}
res[m] = -res[m];
return;
}
// if p*binval == binval + n % 2n: negate the orbit and exit
if (((vp - binval) & _mn) == 0) {
for (uint64_t j = binval; j < nn; j += 2 * binval) {
res[j] = -res[j];
}
return;
}
// if p*binval == n - binval % 2n: mirror the orbit and continue!
if (((vp + binval) & _mn) == 0) {
for (uint64_t j = binval; j < m; j += 2 * binval) {
double tmp = res[j];
res[j] = res[nn - j];
res[nn - j] = tmp;
}
continue;
}
// otherwise we will follow the orbit cycles,
// starting from binval and -binval in parallel
uint64_t j_start = binval;
uint64_t nb_modif = 0;
while (nb_modif < orb_size) {
// follow the cycle that start with j_start
uint64_t j = j_start;
double tmp1 = res[j];
double tmp2 = res[nn - j];
do {
// find where the value should go, and with which sign
uint64_t new_j = (j * p) & _2mn; // mod 2n to get the position and sign
uint64_t new_j_n = new_j & _mn; // mod n to get just the position
// exchange this position with tmp1 (and take care of the sign)
double tmp1a = res[new_j_n];
double tmp2a = res[nn - new_j_n];
if (new_j < nn) {
res[new_j_n] = tmp1;
res[nn - new_j_n] = tmp2;
} else {
res[new_j_n] = -tmp1;
res[nn - new_j_n] = -tmp2;
}
tmp1 = tmp1a;
tmp2 = tmp2a;
// move to the new location, and store the number of items modified
nb_modif += 2;
j = new_j_n;
} while (j != j_start);
// move to the start of the next cycle:
// we need to find an index that has not been touched yet, and pick it as next j_start.
// in practice, it is enough to do *5, because 5 is a generator.
j_start = (5 * j_start) & _mn;
}
}
}

79
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/coeffs/coeffs_arithmetic.h
View File

@@ -1,79 +0,0 @@
#ifndef SPQLIOS_COEFFS_ARITHMETIC_H
#define SPQLIOS_COEFFS_ARITHMETIC_H
#include "../commons.h"
/** res = a + b */
EXPORT void znx_add_i64_ref(uint64_t nn, int64_t* res, const int64_t* a, const int64_t* b);
EXPORT void znx_add_i64_avx(uint64_t nn, int64_t* res, const int64_t* a, const int64_t* b);
/** res = a - b */
EXPORT void znx_sub_i64_ref(uint64_t nn, int64_t* res, const int64_t* a, const int64_t* b);
EXPORT void znx_sub_i64_avx(uint64_t nn, int64_t* res, const int64_t* a, const int64_t* b);
/** res = -a */
EXPORT void znx_negate_i64_ref(uint64_t nn, int64_t* res, const int64_t* a);
EXPORT void znx_negate_i64_avx(uint64_t nn, int64_t* res, const int64_t* a);
/** res = a */
EXPORT void znx_copy_i64_ref(uint64_t nn, int64_t* res, const int64_t* a);
/** res = 0 */
EXPORT void znx_zero_i64_ref(uint64_t nn, int64_t* res);
/** res = a / m where m is a power of 2 */
EXPORT void rnx_divide_by_m_ref(uint64_t nn, double m, double* res, const double* a);
EXPORT void rnx_divide_by_m_avx(uint64_t nn, double m, double* res, const double* a);
/**
* @param res = X^p *in mod X^nn +1
* @param nn the ring dimension
* @param p a power for the rotation -2nn <= p <= 2nn
* @param in is a rnx/znx vector of dimension nn
*/
EXPORT void rnx_rotate_f64(uint64_t nn, int64_t p, double* res, const double* in);
EXPORT void znx_rotate_i64(uint64_t nn, int64_t p, int64_t* res, const int64_t* in);
EXPORT void rnx_rotate_inplace_f64(uint64_t nn, int64_t p, double* res);
EXPORT void znx_rotate_inplace_i64(uint64_t nn, int64_t p, int64_t* res);
/**
* @brief res(X) = in(X^p)
* @param nn the ring dimension
* @param p is odd integer and must be between 0 < p < 2nn
* @param in is a rnx/znx vector of dimension nn
*/
EXPORT void rnx_automorphism_f64(uint64_t nn, int64_t p, double* res, const double* in);
EXPORT void znx_automorphism_i64(uint64_t nn, int64_t p, int64_t* res, const int64_t* in);
EXPORT void rnx_automorphism_inplace_f64(uint64_t nn, int64_t p, double* res);
EXPORT void znx_automorphism_inplace_i64(uint64_t nn, int64_t p, int64_t* res);
/**
* @brief res = (X^p-1).in
* @param nn the ring dimension
* @param p must be between -2nn <= p <= 2nn
* @param in is a rnx/znx vector of dimension nn
*/
EXPORT void rnx_mul_xp_minus_one_f64(uint64_t nn, int64_t p, double* res, const double* in);
EXPORT void znx_mul_xp_minus_one_i64(uint64_t nn, int64_t p, int64_t* res, const int64_t* in);
EXPORT void rnx_mul_xp_minus_one_inplace_f64(uint64_t nn, int64_t p, double* res);
EXPORT void znx_mul_xp_minus_one_inplace_i64(uint64_t nn, int64_t p, int64_t* res);
/**
* @brief Normalize input plus carry mod-2^k. The following
* equality holds @c {in + carry_in == out + carry_out . 2^k}.
*
* @c in must be in [-2^62 .. 2^62]
*
* @c out is in [ -2^(base_k-1), 2^(base_k-1) [.
*
* @c carry_in and @carry_out have at most 64+1-k bits.
*
* Null @c carry_in or @c carry_out are ignored.
*
* @param[in] nn the ring dimension
* @param[in] base_k the base k
* @param out output normalized znx
* @param carry_out output carry znx
* @param[in] in input znx
* @param[in] carry_in input carry znx
*/
EXPORT void znx_normalize(uint64_t nn, uint64_t base_k, int64_t* out, int64_t* carry_out, const int64_t* in,
const int64_t* carry_in);
#endif // SPQLIOS_COEFFS_ARITHMETIC_H

124
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/coeffs/coeffs_arithmetic_avx.c
View File

@@ -1,124 +0,0 @@
#include <immintrin.h>
#include "../commons_private.h"
#include "coeffs_arithmetic.h"
// res = a + b. dimension n must be a power of 2
EXPORT void znx_add_i64_avx(uint64_t nn, int64_t* res, const int64_t* a, const int64_t* b) {
if (nn <= 2) {
if (nn == 1) {
res[0] = a[0] + b[0];
} else {
_mm_storeu_si128((__m128i*)res, //
_mm_add_epi64( //
_mm_loadu_si128((__m128i*)a), //
_mm_loadu_si128((__m128i*)b)));
}
} else {
const __m256i* aa = (__m256i*)a;
const __m256i* bb = (__m256i*)b;
__m256i* rr = (__m256i*)res;
__m256i* const rrend = (__m256i*)(res + nn);
do {
_mm256_storeu_si256(rr, //
_mm256_add_epi64( //
_mm256_loadu_si256(aa), //
_mm256_loadu_si256(bb)));
++rr;
++aa;
++bb;
} while (rr < rrend);
}
}
// res = a - b. dimension n must be a power of 2
EXPORT void znx_sub_i64_avx(uint64_t nn, int64_t* res, const int64_t* a, const int64_t* b) {
if (nn <= 2) {
if (nn == 1) {
res[0] = a[0] - b[0];
} else {
_mm_storeu_si128((__m128i*)res, //
_mm_sub_epi64( //
_mm_loadu_si128((__m128i*)a), //
_mm_loadu_si128((__m128i*)b)));
}
} else {
const __m256i* aa = (__m256i*)a;
const __m256i* bb = (__m256i*)b;
__m256i* rr = (__m256i*)res;
__m256i* const rrend = (__m256i*)(res + nn);
do {
_mm256_storeu_si256(rr, //
_mm256_sub_epi64( //
_mm256_loadu_si256(aa), //
_mm256_loadu_si256(bb)));
++rr;
++aa;
++bb;
} while (rr < rrend);
}
}
EXPORT void znx_negate_i64_avx(uint64_t nn, int64_t* res, const int64_t* a) {
if (nn <= 2) {
if (nn == 1) {
res[0] = -a[0];
} else {
_mm_storeu_si128((__m128i*)res, //
_mm_sub_epi64( //
_mm_set1_epi64x(0), //
_mm_loadu_si128((__m128i*)a)));
}
} else {
const __m256i* aa = (__m256i*)a;
__m256i* rr = (__m256i*)res;
__m256i* const rrend = (__m256i*)(res + nn);
do {
_mm256_storeu_si256(rr, //
_mm256_sub_epi64( //
_mm256_set1_epi64x(0), //
_mm256_loadu_si256(aa)));
++rr;
++aa;
} while (rr < rrend);
}
}
EXPORT void rnx_divide_by_m_avx(uint64_t n, double m, double* res, const double* a) {
// TODO: see if there is a faster way of dividing by a power of 2?
const double invm = 1. / m;
if (n < 8) {
switch (n) {
case 1:
*res = *a * invm;
break;
case 2:
_mm_storeu_pd(res, //
_mm_mul_pd(_mm_loadu_pd(a), //
_mm_set1_pd(invm)));
break;
case 4:
_mm256_storeu_pd(res, //
_mm256_mul_pd(_mm256_loadu_pd(a), //
_mm256_set1_pd(invm)));
break;
default:
NOT_SUPPORTED(); // non-power of 2
}
return;
}
const __m256d invm256 = _mm256_set1_pd(invm);
double* rr = res;
const double* aa = a;
const double* const aaend = a + n;
do {
_mm256_storeu_pd(rr, //
_mm256_mul_pd(_mm256_loadu_pd(aa), //
invm256));
_mm256_storeu_pd(rr + 4, //
_mm256_mul_pd(_mm256_loadu_pd(aa + 4), //
invm256));
rr += 8;
aa += 8;
} while (aa < aaend);
}

165
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/commons.c
View File

@@ -1,165 +0,0 @@
#include "commons.h"
#include <math.h>
#include <stdio.h>
#include <stdlib.h>
EXPORT void* UNDEFINED_p_ii(int32_t n, int32_t m) { UNDEFINED(); }
EXPORT void* UNDEFINED_p_uu(uint32_t n, uint32_t m) { UNDEFINED(); }
EXPORT double* UNDEFINED_dp_pi(const void* p, int32_t n) { UNDEFINED(); }
EXPORT void* UNDEFINED_vp_pi(const void* p, int32_t n) { UNDEFINED(); }
EXPORT void* UNDEFINED_vp_pu(const void* p, uint32_t n) { UNDEFINED(); }
EXPORT void UNDEFINED_v_vpdp(const void* p, double* a) { UNDEFINED(); }
EXPORT void UNDEFINED_v_vpvp(const void* p, void* a) { UNDEFINED(); }
EXPORT double* NOT_IMPLEMENTED_dp_i(int32_t n) { NOT_IMPLEMENTED(); }
EXPORT void* NOT_IMPLEMENTED_vp_i(int32_t n) { NOT_IMPLEMENTED(); }
EXPORT void* NOT_IMPLEMENTED_vp_u(uint32_t n) { NOT_IMPLEMENTED(); }
EXPORT void NOT_IMPLEMENTED_v_dp(double* a) { NOT_IMPLEMENTED(); }
EXPORT void NOT_IMPLEMENTED_v_vp(void* p) { NOT_IMPLEMENTED(); }
EXPORT void NOT_IMPLEMENTED_v_idpdpdp(int32_t n, double* a, const double* b, const double* c) { NOT_IMPLEMENTED(); }
EXPORT void NOT_IMPLEMENTED_v_uvpcvpcvp(uint32_t n, void* r, const void* a, const void* b) { NOT_IMPLEMENTED(); }
EXPORT void NOT_IMPLEMENTED_v_uvpvpcvp(uint32_t n, void* a, void* b, const void* o) { NOT_IMPLEMENTED(); }
#ifdef _WIN32
#define __always_inline inline __attribute((always_inline))
#endif
void internal_accurate_sincos(double* rcos, double* rsin, double x) {
double _4_x_over_pi = 4 * x / M_PI;
int64_t int_part = ((int64_t)rint(_4_x_over_pi)) & 7;
double frac_part = _4_x_over_pi - (double)(int_part);
double frac_x = M_PI * frac_part / 4.;
// compute the taylor series
double cosp = 1.;
double sinp = 0.;
double powx = 1.;
int64_t nn = 0;
while (fabs(powx) > 1e-20) {
++nn;
powx = powx * frac_x / (double)(nn); // x^n/n!
switch (nn & 3) {
case 0:
cosp += powx;
break;
case 1:
sinp += powx;
break;
case 2:
cosp -= powx;
break;
case 3:
sinp -= powx;
break;
default:
abort(); // impossible
}
}
// final multiplication
switch (int_part) {
case 0:
*rcos = cosp;
*rsin = sinp;
break;
case 1:
*rcos = M_SQRT1_2 * (cosp - sinp);
*rsin = M_SQRT1_2 * (cosp + sinp);
break;
case 2:
*rcos = -sinp;
*rsin = cosp;
break;
case 3:
*rcos = -M_SQRT1_2 * (cosp + sinp);
*rsin = M_SQRT1_2 * (cosp - sinp);
break;
case 4:
*rcos = -cosp;
*rsin = -sinp;
break;
case 5:
*rcos = -M_SQRT1_2 * (cosp - sinp);
*rsin = -M_SQRT1_2 * (cosp + sinp);
break;
case 6:
*rcos = sinp;
*rsin = -cosp;
break;
case 7:
*rcos = M_SQRT1_2 * (cosp + sinp);
*rsin = -M_SQRT1_2 * (cosp - sinp);
break;
default:
abort(); // impossible
}
if (fabs(cos(x) - *rcos) > 1e-10 || fabs(sin(x) - *rsin) > 1e-10) {
printf("cos(%.17lf) =? %.17lf instead of %.17lf\n", x, *rcos, cos(x));
printf("sin(%.17lf) =? %.17lf instead of %.17lf\n", x, *rsin, sin(x));
printf("fracx = %.17lf\n", frac_x);
printf("cosp = %.17lf\n", cosp);
printf("sinp = %.17lf\n", sinp);
printf("nn = %d\n", (int)(nn));
}
}
double internal_accurate_cos(double x) {
double rcos, rsin;
internal_accurate_sincos(&rcos, &rsin, x);
return rcos;
}
double internal_accurate_sin(double x) {
double rcos, rsin;
internal_accurate_sincos(&rcos, &rsin, x);
return rsin;
}
EXPORT void spqlios_debug_free(void* addr) { free((uint8_t*)addr - 64); }
EXPORT void* spqlios_debug_alloc(uint64_t size) { return (uint8_t*)malloc(size + 64) + 64; }
EXPORT void spqlios_free(void* addr) {
#ifndef NDEBUG
// in debug mode, we deallocated with spqlios_debug_free()
spqlios_debug_free(addr);
#else
// in release mode, the function will free aligned memory
#ifdef _WIN32
_aligned_free(addr);
#else
free(addr);
#endif
#endif
}
EXPORT void* spqlios_alloc(uint64_t size) {
#ifndef NDEBUG
// in debug mode, the function will not necessarily have any particular alignment
// it will also ensure that memory can only be deallocated with spqlios_free()
return spqlios_debug_alloc(size);
#else
// in release mode, the function will return 64-bytes aligned memory
#ifdef _WIN32
void* reps = _aligned_malloc((size + 63) & (UINT64_C(-64)), 64);
#else
void* reps = aligned_alloc(64, (size + 63) & (UINT64_C(-64)));
#endif
if (reps == 0) FATAL_ERROR("Out of memory");
return reps;
#endif
}
EXPORT void* spqlios_alloc_custom_align(uint64_t align, uint64_t size) {
#ifndef NDEBUG
// in debug mode, the function will not necessarily have any particular alignment
// it will also ensure that memory can only be deallocated with spqlios_free()
return spqlios_debug_alloc(size);
#else
// in release mode, the function will return aligned memory
#ifdef _WIN32
void* reps = _aligned_malloc(size, align);
#else
void* reps = aligned_alloc(align, size);
#endif
if (reps == 0) FATAL_ERROR("Out of memory");
return reps;
#endif
}

77
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/commons.h
View File

@@ -1,77 +0,0 @@
#ifndef SPQLIOS_COMMONS_H
#define SPQLIOS_COMMONS_H
#ifdef __cplusplus
#include <cstdint>
#include <cstdio>
#include <cstdlib>
#define EXPORT extern "C"
#define EXPORT_DECL extern "C"
#else
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#define EXPORT
#define EXPORT_DECL extern
#define nullptr 0x0;
#endif
#define UNDEFINED() \
{ \
fprintf(stderr, "UNDEFINED!!!\n"); \
abort(); \
}
#define NOT_IMPLEMENTED() \
{ \
fprintf(stderr, "NOT IMPLEMENTED!!!\n"); \
abort(); \
}
#define FATAL_ERROR(MESSAGE) \
{ \
fprintf(stderr, "ERROR: %s\n", (MESSAGE)); \
abort(); \
}
EXPORT void* UNDEFINED_p_ii(int32_t n, int32_t m);
EXPORT void* UNDEFINED_p_uu(uint32_t n, uint32_t m);
EXPORT double* UNDEFINED_dp_pi(const void* p, int32_t n);
EXPORT void* UNDEFINED_vp_pi(const void* p, int32_t n);
EXPORT void* UNDEFINED_vp_pu(const void* p, uint32_t n);
EXPORT void UNDEFINED_v_vpdp(const void* p, double* a);
EXPORT void UNDEFINED_v_vpvp(const void* p, void* a);
EXPORT double* NOT_IMPLEMENTED_dp_i(int32_t n);
EXPORT void* NOT_IMPLEMENTED_vp_i(int32_t n);
EXPORT void* NOT_IMPLEMENTED_vp_u(uint32_t n);
EXPORT void NOT_IMPLEMENTED_v_dp(double* a);
EXPORT void NOT_IMPLEMENTED_v_vp(void* p);
EXPORT void NOT_IMPLEMENTED_v_idpdpdp(int32_t n, double* a, const double* b, const double* c);
EXPORT void NOT_IMPLEMENTED_v_uvpcvpcvp(uint32_t n, void* r, const void* a, const void* b);
EXPORT void NOT_IMPLEMENTED_v_uvpvpcvp(uint32_t n, void* a, void* b, const void* o);
// windows
#if defined(_WIN32) || defined(__APPLE__)
#define __always_inline inline __attribute((always_inline))
#endif
EXPORT void spqlios_free(void* address);
EXPORT void* spqlios_alloc(uint64_t size);
EXPORT void* spqlios_alloc_custom_align(uint64_t align, uint64_t size);
#define USE_LIBM_SIN_COS
#ifndef USE_LIBM_SIN_COS
// if at some point, we want to remove the libm dependency, we can
// consider this:
EXPORT double internal_accurate_cos(double x);
EXPORT double internal_accurate_sin(double x);
EXPORT void internal_accurate_sincos(double* rcos, double* rsin, double x);
#define m_accurate_cos internal_accurate_cos
#define m_accurate_sin internal_accurate_sin
#else
// let's use libm sin and cos
#define m_accurate_cos cos
#define m_accurate_sin sin
#endif
#endif // SPQLIOS_COMMONS_H

55
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/commons_private.c
View File

@@ -1,55 +0,0 @@
#include "commons_private.h"
#include <stdio.h>
#include <stdlib.h>
#include "commons.h"
EXPORT void* spqlios_error(const char* error) {
fputs(error, stderr);
abort();
return nullptr;
}
EXPORT void* spqlios_keep_or_free(void* ptr, void* ptr2) {
if (!ptr2) {
free(ptr);
}
return ptr2;
}
EXPORT uint32_t log2m(uint32_t m) {
uint32_t a = m - 1;
if (m & a) FATAL_ERROR("m must be a power of two");
a = (a & 0x55555555u) + ((a >> 1) & 0x55555555u);
a = (a & 0x33333333u) + ((a >> 2) & 0x33333333u);
a = (a & 0x0F0F0F0Fu) + ((a >> 4) & 0x0F0F0F0Fu);
a = (a & 0x00FF00FFu) + ((a >> 8) & 0x00FF00FFu);
return (a & 0x0000FFFFu) + ((a >> 16) & 0x0000FFFFu);
}
EXPORT uint64_t is_not_pow2_double(void* doublevalue) { return (*(uint64_t*)doublevalue) & 0x7FFFFFFFFFFFFUL; }
uint32_t revbits(uint32_t nbits, uint32_t value) {
uint32_t res = 0;
for (uint32_t i = 0; i < nbits; ++i) {
res = (res << 1) + (value & 1);
value >>= 1;
}
return res;
}
/**
* @brief this computes the sequence: 0,1/2,1/4,3/4,1/8,5/8,3/8,7/8,...
* essentially: the bits of (i+1) in lsb order on the basis (1/2^k) mod 1*/
double fracrevbits(uint32_t i) {
if (i == 0) return 0;
if (i == 1) return 0.5;
if (i % 2 == 0)
return fracrevbits(i / 2) / 2.;
else
return fracrevbits((i - 1) / 2) / 2. + 0.5;
}
uint64_t ceilto64b(uint64_t size) { return (size + UINT64_C(63)) & (UINT64_C(-64)); }
uint64_t ceilto32b(uint64_t size) { return (size + UINT64_C(31)) & (UINT64_C(-32)); }

72
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/commons_private.h
View File

@@ -1,72 +0,0 @@
#ifndef SPQLIOS_COMMONS_PRIVATE_H
#define SPQLIOS_COMMONS_PRIVATE_H
#include "commons.h"
#ifdef __cplusplus
#include <cmath>
#include <cstdio>
#include <cstdlib>
#else
#include <math.h>
#include <stdio.h>
#include <stdlib.h>
#define nullptr 0x0;
#endif
/** @brief log2 of a power of two (UB if m is not a power of two) */
EXPORT uint32_t log2m(uint32_t m);
/** @brief checks if the doublevalue is a power of two */
EXPORT uint64_t is_not_pow2_double(void* doublevalue);
#define UNDEFINED() \
{ \
fprintf(stderr, "UNDEFINED!!!\n"); \
abort(); \
}
#define NOT_IMPLEMENTED() \
{ \
fprintf(stderr, "NOT IMPLEMENTED!!!\n"); \
abort(); \
}
#define NOT_SUPPORTED() \
{ \
fprintf(stderr, "NOT SUPPORTED!!!\n"); \
abort(); \
}
#define FATAL_ERROR(MESSAGE) \
{ \
fprintf(stderr, "ERROR: %s\n", (MESSAGE)); \
abort(); \
}
#define STATIC_ASSERT(condition) (void)sizeof(char[-1 + 2 * !!(condition)])
/** @brief reports the error and returns nullptr */
EXPORT void* spqlios_error(const char* error);
/** @brief if ptr2 is not null, returns ptr, otherwise free ptr and return null */
EXPORT void* spqlios_keep_or_free(void* ptr, void* ptr2);
#ifdef __x86_64__
#define CPU_SUPPORTS __builtin_cpu_supports
#else
// TODO for now, we do not have any optimization for non x86 targets
#define CPU_SUPPORTS(xxxx) 0
#endif
/** @brief returns the n bits of value in reversed order */
EXPORT uint32_t revbits(uint32_t nbits, uint32_t value);
/**
* @brief this computes the sequence: 0,1/2,1/4,3/4,1/8,5/8,3/8,7/8,...
* essentially: the bits of (i+1) in lsb order on the basis (1/2^k) mod 1*/
EXPORT double fracrevbits(uint32_t i);
/** @brief smallest multiple of 64 higher or equal to size */
EXPORT uint64_t ceilto64b(uint64_t size);
/** @brief smallest multiple of 32 higher or equal to size */
EXPORT uint64_t ceilto32b(uint64_t size);
#endif // SPQLIOS_COMMONS_PRIVATE_H

22
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/cplx/README.md
View File

@@ -1,22 +0,0 @@
In this folder, we deal with the full complex FFT in `C[X] mod X^M-i`.
One complex is represented by two consecutive doubles `(real,imag)`
Note that a real polynomial sum_{j=0}^{N-1} p_j.X^j mod X^N+1
corresponds to the complex polynomial of half degree `M=N/2`:
`sum_{j=0}^{M-1} (p_{j} + i.p_{j+M}) X^j mod X^M-i`
For a complex polynomial A(X) sum c_i X^i of degree M-1
or a real polynomial sum a_i X^i of degree N
coefficient space:
a_0,a_M,a_1,a_{M+1},...,a_{M-1},a_{2M-1}
or equivalently
Re(c_0),Im(c_0),Re(c_1),Im(c_1),...Re(c_{M-1}),Im(c_{M-1})
eval space:
c(omega_{0}),...,c(omega_{M-1})
where
omega_j = omega^{1+rev_{2N}(j)}
and omega = exp(i.pi/N)
rev_{2N}(j) is the number that has the log2(2N) bits of j in reverse order.

80
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/cplx/cplx_common.c
View File

@@ -1,80 +0,0 @@
#include "cplx_fft_internal.h"
void cplx_set(CPLX r, const CPLX a) {
r[0] = a[0];
r[1] = a[1];
}
void cplx_neg(CPLX r, const CPLX a) {
r[0] = -a[0];
r[1] = -a[1];
}
void cplx_add(CPLX r, const CPLX a, const CPLX b) {
r[0] = a[0] + b[0];
r[1] = a[1] + b[1];
}
void cplx_sub(CPLX r, const CPLX a, const CPLX b) {
r[0] = a[0] - b[0];
r[1] = a[1] - b[1];
}
void cplx_mul(CPLX r, const CPLX a, const CPLX b) {
double re = a[0] * b[0] - a[1] * b[1];
r[1] = a[0] * b[1] + a[1] * b[0];
r[0] = re;
}
/**
* @brief splits 2h evaluations of one polynomials into 2 times h evaluations of even/odd polynomial
* Input: Q_0(y),...,Q_{h-1}(y),Q_0(-y),...,Q_{h-1}(-y)
* Output: P_0(z),...,P_{h-1}(z),P_h(z),...,P_{2h-1}(z)
* where Q_i(X)=P_i(X^2)+X.P_{h+i}(X^2) and y^2 = z
* @param h number of "coefficients" h >= 1
* @param data 2h complex coefficients interleaved and 256b aligned
* @param powom y represented as (yre,yim)
*/
EXPORT void cplx_split_fft_ref(int32_t h, CPLX* data, const CPLX powom) {
CPLX* d0 = data;
CPLX* d1 = data + h;
for (uint64_t i = 0; i < h; ++i) {
CPLX diff;
cplx_sub(diff, d0[i], d1[i]);
cplx_add(d0[i], d0[i], d1[i]);
cplx_mul(d1[i], diff, powom);
}
}
/**
* @brief Do two layers of itwiddle (i.e. split).
* Input/output: d0,d1,d2,d3 of length h
* Algo:
* itwiddle(d0,d1,om[0]),itwiddle(d2,d3,i.om[0])
* itwiddle(d0,d2,om[1]),itwiddle(d1,d3,om[1])
* @param h number of "coefficients" h >= 1
* @param data 4h complex coefficients interleaved and 256b aligned
* @param powom om[0] (re,im) and om[1] where om[1]=om[0]^2
*/
EXPORT void cplx_bisplit_fft_ref(int32_t h, CPLX* data, const CPLX powom[2]) {
CPLX* d0 = data;
CPLX* d2 = data + 2 * h;
const CPLX* om0 = powom;
CPLX iom0;
iom0[0] = powom[0][1];
iom0[1] = -powom[0][0];
const CPLX* om1 = powom + 1;
cplx_split_fft_ref(h, d0, *om0);
cplx_split_fft_ref(h, d2, iom0);
cplx_split_fft_ref(2 * h, d0, *om1);
}
/**
* Input: Q(y),Q(-y)
* Output: P_0(z),P_1(z)
* where Q(X)=P_0(X^2)+X.P_1(X^2) and y^2 = z
* @param data 2 complexes coefficients interleaved and 256b aligned
* @param powom (z,-z) interleaved: (zre,zim,-zre,-zim)
*/
void split_fft_last_ref(CPLX* data, const CPLX powom) {
CPLX diff;
cplx_sub(diff, data[0], data[1]);
cplx_add(data[0], data[0], data[1]);
cplx_mul(data[1], diff, powom);
}

158
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/cplx/cplx_conversions.c
View File

@@ -1,158 +0,0 @@
#include <errno.h>
#include <string.h>
#include "../commons_private.h"
#include "cplx_fft_internal.h"
#include "cplx_fft_private.h"
EXPORT void cplx_from_znx32_ref(const CPLX_FROM_ZNX32_PRECOMP* precomp, void* r, const int32_t* x) {
const uint32_t m = precomp->m;
const int32_t* inre = x;
const int32_t* inim = x + m;
CPLX* out = r;
for (uint32_t i = 0; i < m; ++i) {
out[i][0] = (double)inre[i];
out[i][1] = (double)inim[i];
}
}
EXPORT void cplx_from_tnx32_ref(const CPLX_FROM_TNX32_PRECOMP* precomp, void* r, const int32_t* x) {
static const double _2p32 = 1. / (INT64_C(1) << 32);
const uint32_t m = precomp->m;
const int32_t* inre = x;
const int32_t* inim = x + m;
CPLX* out = r;
for (uint32_t i = 0; i < m; ++i) {
out[i][0] = ((double)inre[i]) * _2p32;
out[i][1] = ((double)inim[i]) * _2p32;
}
}
EXPORT void cplx_to_tnx32_ref(const CPLX_TO_TNX32_PRECOMP* precomp, int32_t* r, const void* x) {
static const double _2p32 = (INT64_C(1) << 32);
const uint32_t m = precomp->m;
double factor = _2p32 / precomp->divisor;
int32_t* outre = r;
int32_t* outim = r + m;
const CPLX* in = x;
// Note: this formula will only work if abs(in) < 2^32
for (uint32_t i = 0; i < m; ++i) {
outre[i] = (int32_t)(int64_t)(rint(in[i][0] * factor));
outim[i] = (int32_t)(int64_t)(rint(in[i][1] * factor));
}
}
void* init_cplx_from_znx32_precomp(CPLX_FROM_ZNX32_PRECOMP* res, uint32_t m) {
res->m = m;
if (CPU_SUPPORTS("avx2")) {
if (m >= 8) {
res->function = cplx_from_znx32_avx2_fma;
} else {
res->function = cplx_from_znx32_ref;
}
} else {
res->function = cplx_from_znx32_ref;
}
return res;
}
CPLX_FROM_ZNX32_PRECOMP* new_cplx_from_znx32_precomp(uint32_t m) {
CPLX_FROM_ZNX32_PRECOMP* res = malloc(sizeof(CPLX_FROM_ZNX32_PRECOMP));
if (!res) return spqlios_error(strerror(errno));
return spqlios_keep_or_free(res, init_cplx_from_znx32_precomp(res, m));
}
void* init_cplx_from_tnx32_precomp(CPLX_FROM_TNX32_PRECOMP* res, uint32_t m) {
res->m = m;
if (CPU_SUPPORTS("avx2")) {
if (m >= 8) {
res->function = cplx_from_tnx32_avx2_fma;
} else {
res->function = cplx_from_tnx32_ref;
}
} else {
res->function = cplx_from_tnx32_ref;
}
return res;
}
CPLX_FROM_TNX32_PRECOMP* new_cplx_from_tnx32_precomp(uint32_t m) {
CPLX_FROM_TNX32_PRECOMP* res = malloc(sizeof(CPLX_FROM_TNX32_PRECOMP));
if (!res) return spqlios_error(strerror(errno));
return spqlios_keep_or_free(res, init_cplx_from_tnx32_precomp(res, m));
}
void* init_cplx_to_tnx32_precomp(CPLX_TO_TNX32_PRECOMP* res, uint32_t m, double divisor, uint32_t log2overhead) {
if (is_not_pow2_double(&divisor)) return spqlios_error("divisor must be a power of 2");
if (m & (m - 1)) return spqlios_error("m must be a power of 2");
if (log2overhead > 52) return spqlios_error("log2overhead is too large");
res->m = m;
res->divisor = divisor;
if (CPU_SUPPORTS("avx2")) {
if (log2overhead <= 18) {
if (m >= 8) {
res->function = cplx_to_tnx32_avx2_fma;
} else {
res->function = cplx_to_tnx32_ref;
}
} else {
res->function = cplx_to_tnx32_ref;
}
} else {
res->function = cplx_to_tnx32_ref;
}
return res;
}
EXPORT CPLX_TO_TNX32_PRECOMP* new_cplx_to_tnx32_precomp(uint32_t m, double divisor, uint32_t log2overhead) {
CPLX_TO_TNX32_PRECOMP* res = malloc(sizeof(CPLX_TO_TNX32_PRECOMP));
if (!res) return spqlios_error(strerror(errno));
return spqlios_keep_or_free(res, init_cplx_to_tnx32_precomp(res, m, divisor, log2overhead));
}
/**
* @brief Simpler API for the znx32 to cplx conversion.
* For each dimension, the precomputed tables for this dimension are generated automatically the first time.
* It is advised to do one dry-run call per desired dimension before using in a multithread environment */
EXPORT void cplx_from_znx32_simple(uint32_t m, void* r, const int32_t* x) {
// not checking for log2bound which is not relevant here
static CPLX_FROM_ZNX32_PRECOMP precomp[32];
CPLX_FROM_ZNX32_PRECOMP* p = precomp + log2m(m);
if (!p->function) {
if (!init_cplx_from_znx32_precomp(p, m)) abort();
}
p->function(p, r, x);
}
/**
* @brief Simpler API for the tnx32 to cplx conversion.
* For each dimension, the precomputed tables for this dimension are generated automatically the first time.
* It is advised to do one dry-run call per desired dimension before using in a multithread environment */
EXPORT void cplx_from_tnx32_simple(uint32_t m, void* r, const int32_t* x) {
static CPLX_FROM_TNX32_PRECOMP precomp[32];
CPLX_FROM_TNX32_PRECOMP* p = precomp + log2m(m);
if (!p->function) {
if (!init_cplx_from_tnx32_precomp(p, m)) abort();
}
p->function(p, r, x);
}
/**
* @brief Simpler API for the cplx to tnx32 conversion.
* For each dimension, the precomputed tables for this dimension are generated automatically the first time.
* It is advised to do one dry-run call per desired dimension before using in a multithread environment */
EXPORT void cplx_to_tnx32_simple(uint32_t m, double divisor, uint32_t log2overhead, int32_t* r, const void* x) {
struct LAST_CPLX_TO_TNX32_PRECOMP {
CPLX_TO_TNX32_PRECOMP p;
double last_divisor;
double last_log2over;
};
static __thread struct LAST_CPLX_TO_TNX32_PRECOMP precomp[32];
struct LAST_CPLX_TO_TNX32_PRECOMP* p = precomp + log2m(m);
if (!p->p.function || divisor != p->last_divisor || log2overhead != p->last_log2over) {
memset(p, 0, sizeof(*p));
if (!init_cplx_to_tnx32_precomp(&p->p, m, divisor, log2overhead)) abort();
p->last_divisor = divisor;
p->last_log2over = log2overhead;
}
p->p.function(&p->p, r, x);
}

104
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/cplx/cplx_conversions_avx2_fma.c
View File

@@ -1,104 +0,0 @@
#include <immintrin.h>
#include <stdio.h>
#include "cplx_fft_internal.h"
#include "cplx_fft_private.h"
typedef int32_t I8MEM[8];
typedef double D4MEM[4];
__always_inline void cplx_from_any_fma(uint64_t m, void* r, const int32_t* x, const __m256i C, const __m256d R) {
const __m256i S = _mm256_set1_epi32(0x80000000);
const I8MEM* inre = (I8MEM*)(x);
const I8MEM* inim = (I8MEM*)(x + m);
D4MEM* out = (D4MEM*)r;
const uint64_t ms8 = m / 8;
for (uint32_t i = 0; i < ms8; ++i) {
__m256i rea = _mm256_loadu_si256((__m256i*)inre[0]);
__m256i ima = _mm256_loadu_si256((__m256i*)inim[0]);
rea = _mm256_add_epi32(rea, S);
ima = _mm256_add_epi32(ima, S);
__m256i tmpa = _mm256_unpacklo_epi32(rea, ima);
__m256i tmpc = _mm256_unpackhi_epi32(rea, ima);
__m256i cpla = _mm256_permute2x128_si256(tmpa, tmpc, 0x20);
__m256i cplc = _mm256_permute2x128_si256(tmpa, tmpc, 0x31);
tmpa = _mm256_unpacklo_epi32(cpla, C);
__m256i tmpb = _mm256_unpackhi_epi32(cpla, C);
tmpc = _mm256_unpacklo_epi32(cplc, C);
__m256i tmpd = _mm256_unpackhi_epi32(cplc, C);
cpla = _mm256_permute2x128_si256(tmpa, tmpb, 0x20);
__m256i cplb = _mm256_permute2x128_si256(tmpa, tmpb, 0x31);
cplc = _mm256_permute2x128_si256(tmpc, tmpd, 0x20);
__m256i cpld = _mm256_permute2x128_si256(tmpc, tmpd, 0x31);
__m256d dcpla = _mm256_sub_pd(_mm256_castsi256_pd(cpla), R);
__m256d dcplb = _mm256_sub_pd(_mm256_castsi256_pd(cplb), R);
__m256d dcplc = _mm256_sub_pd(_mm256_castsi256_pd(cplc), R);
__m256d dcpld = _mm256_sub_pd(_mm256_castsi256_pd(cpld), R);
_mm256_storeu_pd(out[0], dcpla);
_mm256_storeu_pd(out[1], dcplb);
_mm256_storeu_pd(out[2], dcplc);
_mm256_storeu_pd(out[3], dcpld);
inre += 1;
inim += 1;
out += 4;
}
}
EXPORT void cplx_from_znx32_avx2_fma(const CPLX_FROM_ZNX32_PRECOMP* precomp, void* r, const int32_t* x) {
// note: the hex code of 2^31 + 2^52 is 0x4330000080000000
const __m256i C = _mm256_set1_epi32(0x43300000);
const __m256d R = _mm256_set1_pd((INT64_C(1) << 31) + (INT64_C(1) << 52));
// double XX = INT64_C(1) + (INT64_C(1)<<31) + (INT64_C(1)<<52);
// printf("\n\n%016lx\n", *(uint64_t*)&XX);
// abort();
const uint64_t m = precomp->m;
cplx_from_any_fma(m, r, x, C, R);
}
EXPORT void cplx_from_tnx32_avx2_fma(const CPLX_FROM_TNX32_PRECOMP* precomp, void* r, const int32_t* x) {
// note: the hex code of 2^-1 + 2^30 is 0x4130000080000000
const __m256i C = _mm256_set1_epi32(0x41300000);
const __m256d R = _mm256_set1_pd(0.5 + (INT64_C(1) << 20));
// double XX = (double)(INT64_C(1) + (INT64_C(1)<<31) + (INT64_C(1)<<52))/(INT64_C(1)<<32);
// printf("\n\n%016lx\n", *(uint64_t*)&XX);
// abort();
const uint64_t m = precomp->m;
cplx_from_any_fma(m, r, x, C, R);
}
EXPORT void cplx_to_tnx32_avx2_fma(const CPLX_TO_TNX32_PRECOMP* precomp, int32_t* r, const void* x) {
const __m256d R = _mm256_set1_pd((0.5 + (INT64_C(3) << 19)) * precomp->divisor);
const __m256i MASK = _mm256_set1_epi64x(0xFFFFFFFFUL);
const __m256i S = _mm256_set1_epi32(0x80000000);
// const __m256i IDX = _mm256_set_epi32(0,4,1,5,2,6,3,7);
const __m256i IDX = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
const uint64_t m = precomp->m;
const uint64_t ms8 = m / 8;
I8MEM* outre = (I8MEM*)r;
I8MEM* outim = (I8MEM*)(r + m);
const D4MEM* in = x;
// Note: this formula will only work if abs(in) < 2^32
for (uint32_t i = 0; i < ms8; ++i) {
__m256d cpla = _mm256_loadu_pd(in[0]);
__m256d cplb = _mm256_loadu_pd(in[1]);
__m256d cplc = _mm256_loadu_pd(in[2]);
__m256d cpld = _mm256_loadu_pd(in[3]);
__m256i icpla = _mm256_castpd_si256(_mm256_add_pd(cpla, R));
__m256i icplb = _mm256_castpd_si256(_mm256_add_pd(cplb, R));
__m256i icplc = _mm256_castpd_si256(_mm256_add_pd(cplc, R));
__m256i icpld = _mm256_castpd_si256(_mm256_add_pd(cpld, R));
icpla = _mm256_or_si256(_mm256_and_si256(icpla, MASK), _mm256_slli_epi64(icplb, 32));
icplc = _mm256_or_si256(_mm256_and_si256(icplc, MASK), _mm256_slli_epi64(icpld, 32));
icpla = _mm256_xor_si256(icpla, S);
icplc = _mm256_xor_si256(icplc, S);
__m256i re = _mm256_unpacklo_epi64(icpla, icplc);
__m256i im = _mm256_unpackhi_epi64(icpla, icplc);
re = _mm256_permutevar8x32_epi32(re, IDX);
im = _mm256_permutevar8x32_epi32(im, IDX);
_mm256_storeu_si256((__m256i*)outre[0], re);
_mm256_storeu_si256((__m256i*)outim[0], im);
outre += 1;
outim += 1;
in += 4;
}
}

18
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/cplx/cplx_execute.c
View File

@@ -1,18 +0,0 @@
#include "cplx_fft_internal.h"
#include "cplx_fft_private.h"
EXPORT void cplx_from_znx32(const CPLX_FROM_ZNX32_PRECOMP* tables, void* r, const int32_t* a) {
tables->function(tables, r, a);
}
EXPORT void cplx_from_tnx32(const CPLX_FROM_TNX32_PRECOMP* tables, void* r, const int32_t* a) {
tables->function(tables, r, a);
}
EXPORT void cplx_to_tnx32(const CPLX_TO_TNX32_PRECOMP* tables, int32_t* r, const void* a) {
tables->function(tables, r, a);
}
EXPORT void cplx_fftvec_mul(const CPLX_FFTVEC_MUL_PRECOMP* tables, void* r, const void* a, const void* b) {
tables->function(tables, r, a, b);
}
EXPORT void cplx_fftvec_addmul(const CPLX_FFTVEC_ADDMUL_PRECOMP* tables, void* r, const void* a, const void* b) {
tables->function(tables, r, a, b);
}

43
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/cplx/cplx_fallbacks_aarch64.c
View File

@@ -1,43 +0,0 @@
#include "cplx_fft_internal.h"
#include "cplx_fft_private.h"
EXPORT void cplx_fftvec_addmul_fma(const CPLX_FFTVEC_ADDMUL_PRECOMP* tables, void* r, const void* a, const void* b) {
UNDEFINED(); // not defined for non x86 targets
}
EXPORT void cplx_fftvec_mul_fma(const CPLX_FFTVEC_MUL_PRECOMP* tables, void* r, const void* a, const void* b) {
UNDEFINED();
}
EXPORT void cplx_fftvec_addmul_sse(const CPLX_FFTVEC_ADDMUL_PRECOMP* precomp, void* r, const void* a, const void* b) {
UNDEFINED();
}
EXPORT void cplx_fftvec_addmul_avx512(const CPLX_FFTVEC_ADDMUL_PRECOMP* precomp, void* r, const void* a,
const void* b) {
UNDEFINED();
}
EXPORT void cplx_fft16_avx_fma(void* data, const void* omega) { UNDEFINED(); }
EXPORT void cplx_ifft16_avx_fma(void* data, const void* omega) { UNDEFINED(); }
EXPORT void cplx_from_znx32_avx2_fma(const CPLX_FROM_ZNX32_PRECOMP* precomp, void* r, const int32_t* x) { UNDEFINED(); }
EXPORT void cplx_from_tnx32_avx2_fma(const CPLX_FROM_TNX32_PRECOMP* precomp, void* r, const int32_t* x) { UNDEFINED(); }
EXPORT void cplx_to_tnx32_avx2_fma(const CPLX_TO_TNX32_PRECOMP* precomp, int32_t* x, const void* c) { UNDEFINED(); }
EXPORT void cplx_fft_avx2_fma(const CPLX_FFT_PRECOMP* tables, void* data){UNDEFINED()} EXPORT
void cplx_ifft_avx2_fma(const CPLX_IFFT_PRECOMP* itables, void* data){UNDEFINED()} EXPORT
void cplx_fftvec_twiddle_fma(const CPLX_FFTVEC_TWIDDLE_PRECOMP* tables, void* a, void* b, const void* om){
UNDEFINED()} EXPORT void cplx_fftvec_twiddle_avx512(const CPLX_FFTVEC_TWIDDLE_PRECOMP* tables, void* a, void* b,
const void* om){UNDEFINED()} EXPORT
void cplx_fftvec_bitwiddle_fma(const CPLX_FFTVEC_BITWIDDLE_PRECOMP* tables, void* a, uint64_t slice,
const void* om){UNDEFINED()} EXPORT
void cplx_fftvec_bitwiddle_avx512(const CPLX_FFTVEC_BITWIDDLE_PRECOMP* tables, void* a, uint64_t slice,
const void* om){UNDEFINED()}
// DEPRECATED?
EXPORT void cplx_fftvec_add_fma(uint32_t m, void* r, const void* a, const void* b){UNDEFINED()} EXPORT
void cplx_fftvec_sub2_to_fma(uint32_t m, void* r, const void* a, const void* b){UNDEFINED()} EXPORT
void cplx_fftvec_copy_fma(uint32_t m, void* r, const void* a) {
UNDEFINED()
}
// executors
// EXPORT void cplx_ifft(const CPLX_IFFT_PRECOMP* itables, void* data) {
// itables->function(itables, data);
//}
// EXPORT void cplx_fft(const CPLX_FFT_PRECOMP* tables, void* data) { tables->function(tables, data); }

221
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/cplx/cplx_fft.h
View File

@@ -1,221 +0,0 @@
#ifndef SPQLIOS_CPLX_FFT_H
#define SPQLIOS_CPLX_FFT_H
#include "../commons.h"
typedef struct cplx_fft_precomp CPLX_FFT_PRECOMP;
typedef struct cplx_ifft_precomp CPLX_IFFT_PRECOMP;
typedef struct cplx_mul_precomp CPLX_FFTVEC_MUL_PRECOMP;
typedef struct cplx_addmul_precomp CPLX_FFTVEC_ADDMUL_PRECOMP;
typedef struct cplx_from_znx32_precomp CPLX_FROM_ZNX32_PRECOMP;
typedef struct cplx_from_tnx32_precomp CPLX_FROM_TNX32_PRECOMP;
typedef struct cplx_to_tnx32_precomp CPLX_TO_TNX32_PRECOMP;
typedef struct cplx_to_znx32_precomp CPLX_TO_ZNX32_PRECOMP;
typedef struct cplx_from_rnx64_precomp CPLX_FROM_RNX64_PRECOMP;
typedef struct cplx_to_rnx64_precomp CPLX_TO_RNX64_PRECOMP;
typedef struct cplx_round_to_rnx64_precomp CPLX_ROUND_TO_RNX64_PRECOMP;
/**
* @brief precomputes fft tables.
* The FFT tables contains a constant section that is required for efficient FFT operations in dimension nn.
* The resulting pointer is to be passed as "tables" argument to any call to the fft function.
* The user can optionnally allocate zero or more computation buffers, which are scratch spaces that are contiguous to
* the constant tables in memory, and allow for more efficient operations. It is the user's responsibility to ensure
* that each of those buffers are never used simultaneously by two ffts on different threads at the same time. The fft
* table must be deleted by delete_fft_precomp after its last usage.
*/
EXPORT CPLX_FFT_PRECOMP* new_cplx_fft_precomp(uint32_t m, uint32_t num_buffers);
/**
* @brief gets the address of a fft buffer allocated during new_fft_precomp.
* This buffer can be used as data pointer in subsequent calls to fft,
* and does not need to be released afterwards.
*/
EXPORT void* cplx_fft_precomp_get_buffer(const CPLX_FFT_PRECOMP* tables, uint32_t buffer_index);
/**
* @brief allocates a new fft buffer.
* This buffer can be used as data pointer in subsequent calls to fft,
* and must be deleted afterwards by calling delete_fft_buffer.
*/
EXPORT void* new_cplx_fft_buffer(uint32_t m);
/**
* @brief allocates a new fft buffer.
* This buffer can be used as data pointer in subsequent calls to fft,
* and must be deleted afterwards by calling delete_fft_buffer.
*/
EXPORT void delete_cplx_fft_buffer(void* buffer);
/**
* @brief deallocates a fft table and all its built-in buffers.
*/
#define delete_cplx_fft_precomp free
/**
* @brief computes a direct fft in-place over data.
*/
EXPORT void cplx_fft(const CPLX_FFT_PRECOMP* tables, void* data);
EXPORT CPLX_IFFT_PRECOMP* new_cplx_ifft_precomp(uint32_t m, uint32_t num_buffers);
EXPORT void* cplx_ifft_precomp_get_buffer(const CPLX_IFFT_PRECOMP* tables, uint32_t buffer_index);
EXPORT void cplx_ifft(const CPLX_IFFT_PRECOMP* tables, void* data);
#define delete_cplx_ifft_precomp free
EXPORT CPLX_FFTVEC_MUL_PRECOMP* new_cplx_fftvec_mul_precomp(uint32_t m);
EXPORT void cplx_fftvec_mul(const CPLX_FFTVEC_MUL_PRECOMP* tables, void* r, const void* a, const void* b);
#define delete_cplx_fftvec_mul_precomp free
EXPORT CPLX_FFTVEC_ADDMUL_PRECOMP* new_cplx_fftvec_addmul_precomp(uint32_t m);
EXPORT void cplx_fftvec_addmul(const CPLX_FFTVEC_ADDMUL_PRECOMP* tables, void* r, const void* a, const void* b);
#define delete_cplx_fftvec_addmul_precomp free
/**
* @brief prepares a conversion from ZnX to the cplx layout.
* All the coefficients must be strictly lower than 2^log2bound in absolute value. Any attempt to use
* this function on a larger coefficient is undefined behaviour. The resulting precomputed data must
* be freed with `new_cplx_from_znx32_precomp`
* @param m the target complex dimension m from C[X] mod X^m-i. Note that the inputs have n=2m
* int32 coefficients in natural order modulo X^n+1
* @param log2bound bound on the input coefficients. Must be between 0 and 32
*/
EXPORT CPLX_FROM_ZNX32_PRECOMP* new_cplx_from_znx32_precomp(uint32_t m);
/**
* @brief converts from ZnX to the cplx layout.
* @param tables precomputed data obtained by new_cplx_from_znx32_precomp.
* @param r resulting array of m complexes coefficients mod X^m-i
* @param x input array of n bounded integer coefficients mod X^n+1
*/
EXPORT void cplx_from_znx32(const CPLX_FROM_ZNX32_PRECOMP* tables, void* r, const int32_t* a);
/** @brief frees a precomputed conversion data initialized with new_cplx_from_znx32_precomp. */
#define delete_cplx_from_znx32_precomp free
/**
* @brief prepares a conversion from TnX to the cplx layout.
* @param m the target complex dimension m from C[X] mod X^m-i. Note that the inputs have n=2m
* torus32 coefficients. The resulting precomputed data must
* be freed with `delete_cplx_from_tnx32_precomp`
*/
EXPORT CPLX_FROM_TNX32_PRECOMP* new_cplx_from_tnx32_precomp(uint32_t m);
/**
* @brief converts from TnX to the cplx layout.
* @param tables precomputed data obtained by new_cplx_from_tnx32_precomp.
* @param r resulting array of m complexes coefficients mod X^m-i
* @param x input array of n torus32 coefficients mod X^n+1
*/
EXPORT void cplx_from_tnx32(const CPLX_FROM_TNX32_PRECOMP* tables, void* r, const int32_t* a);
/** @brief frees a precomputed conversion data initialized with new_cplx_from_tnx32_precomp. */
#define delete_cplx_from_tnx32_precomp free
/**
* @brief prepares a rescale and conversion from the cplx layout to TnX.
* @param m the target complex dimension m from C[X] mod X^m-i. Note that the outputs have n=2m
* torus32 coefficients.
* @param divisor must be a power of two. The inputs are rescaled by divisor before being reduced modulo 1.
* Remember that the output of an iFFT must be divided by m.
* @param log2overhead all inputs absolute values must be within divisor.2^log2overhead.
* For any inputs outside of these bounds, the conversion is undefined behaviour.
* The maximum supported log2overhead is 52, and the algorithm is faster for log2overhead=18.
*/
EXPORT CPLX_TO_TNX32_PRECOMP* new_cplx_to_tnx32_precomp(uint32_t m, double divisor, uint32_t log2overhead);
/**
* @brief rescale, converts and reduce mod 1 from cplx layout to torus32.
* @param tables precomputed data obtained by new_cplx_from_tnx32_precomp.
* @param r resulting array of n torus32 coefficients mod X^n+1
* @param x input array of m cplx coefficients mod X^m-i
*/
EXPORT void cplx_to_tnx32(const CPLX_TO_TNX32_PRECOMP* tables, int32_t* r, const void* a);
#define delete_cplx_to_tnx32_precomp free
EXPORT CPLX_TO_ZNX32_PRECOMP* new_cplx_to_znx32_precomp(uint32_t m, double divisor);
EXPORT void cplx_to_znx32(const CPLX_TO_ZNX32_PRECOMP* precomp, int32_t* r, const void* x);
#define delete_cplx_to_znx32_simple free
EXPORT CPLX_FROM_RNX64_PRECOMP* new_cplx_from_rnx64_simple(uint32_t m);
EXPORT void cplx_from_rnx64(const CPLX_FROM_RNX64_PRECOMP* precomp, void* r, const double* x);
#define delete_cplx_from_rnx64_simple free
EXPORT CPLX_TO_RNX64_PRECOMP* new_cplx_to_rnx64(uint32_t m, double divisor);
EXPORT void cplx_to_rnx64(const CPLX_TO_RNX64_PRECOMP* precomp, double* r, const void* x);
#define delete_cplx_round_to_rnx64_simple free
EXPORT CPLX_ROUND_TO_RNX64_PRECOMP* new_cplx_round_to_rnx64(uint32_t m, double divisor, uint32_t log2bound);
EXPORT void cplx_round_to_rnx64(const CPLX_ROUND_TO_RNX64_PRECOMP* precomp, double* r, const void* x);
#define delete_cplx_round_to_rnx64_simple free
/**
* @brief Simpler API for the fft function.
* For each dimension, the precomputed tables for this dimension are generated automatically.
* It is advised to do one dry-run per desired dimension before using in a multithread environment */
EXPORT void cplx_fft_simple(uint32_t m, void* data);
/**
* @brief Simpler API for the ifft function.
* For each dimension, the precomputed tables for this dimension are generated automatically the first time.
* It is advised to do one dry-run call per desired dimension in the main thread before using in a multithread
* environment */
EXPORT void cplx_ifft_simple(uint32_t m, void* data);
/**
* @brief Simpler API for the fftvec multiplication function.
* For each dimension, the precomputed tables for this dimension are generated automatically the first time.
* It is advised to do one dry-run call per desired dimension before using in a multithread environment */
EXPORT void cplx_fftvec_mul_simple(uint32_t m, void* r, const void* a, const void* b);
/**
* @brief Simpler API for the fftvec addmul function.
* For each dimension, the precomputed tables for this dimension are generated automatically the first time.
* It is advised to do one dry-run call per desired dimension before using in a multithread environment */
EXPORT void cplx_fftvec_addmul_simple(uint32_t m, void* r, const void* a, const void* b);
/**
* @brief Simpler API for the znx32 to cplx conversion.
* For each dimension, the precomputed tables for this dimension are generated automatically the first time.
* It is advised to do one dry-run call per desired dimension before using in a multithread environment */
EXPORT void cplx_from_znx32_simple(uint32_t m, void* r, const int32_t* x);
/**
* @brief Simpler API for the tnx32 to cplx conversion.
* For each dimension, the precomputed tables for this dimension are generated automatically the first time.
* It is advised to do one dry-run call per desired dimension before using in a multithread environment */
EXPORT void cplx_from_tnx32_simple(uint32_t m, void* r, const int32_t* x);
/**
* @brief Simpler API for the cplx to tnx32 conversion.
* For each dimension, the precomputed tables for this dimension are generated automatically the first time.
* It is advised to do one dry-run call per desired dimension before using in a multithread environment */
EXPORT void cplx_to_tnx32_simple(uint32_t m, double divisor, uint32_t log2overhead, int32_t* r, const void* x);
/**
* @brief converts, divides and round from cplx to znx32 (simple API)
* @param m the complex dimension
* @param divisor the divisor: a power of two, often m after an ifft
* @param r the result: must be a double array of size 2m. r must be distinct from x
* @param x the input: must hold m complex numbers.
*/
EXPORT void cplx_to_znx32_simple(uint32_t m, double divisor, int32_t* r, const void* x);
/**
* @brief converts from rnx64 to cplx (simple API)
* The bound on the output is assumed to be within ]2^-31,2^31[.
* Any coefficient that would fall outside this range is undefined behaviour.
* @param m the complex dimension
* @param r the result: must be an array of m complex numbers. r must be distinct from x
* @param x the input: must be an array of 2m doubles.
*/
EXPORT void cplx_from_rnx64_simple(uint32_t m, void* r, const double* x);
/**
* @brief converts, divides from cplx to rnx64 (simple API)
* @param m the complex dimension
* @param divisor the divisor: a power of two, often m after an ifft
* @param r the result: must be a double array of size 2m. r must be distinct from x
* @param x the input: must hold m complex numbers.
*/
EXPORT void cplx_to_rnx64_simple(uint32_t m, double divisor, double* r, const void* x);
/**
* @brief converts, divides and round to integer from cplx to rnx32 (simple API)
* @param m the complex dimension
* @param divisor the divisor: a power of two, often m after an ifft
* @param log2bound a guarantee on the log2bound of the output. log2bound<=48 will use a more efficient algorithm.
* @param r the result: must be a double array of size 2m. r must be distinct from x
* @param x the input: must hold m complex numbers.
*/
EXPORT void cplx_round_to_rnx64_simple(uint32_t m, double divisor, uint32_t log2bound, double* r, const void* x);
#endif // SPQLIOS_CPLX_FFT_H

156
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/cplx/cplx_fft16_avx_fma.s
View File

@@ -1,156 +0,0 @@
# shifted FFT over X^16-i
# 1st argument (rdi) contains 16 complexes
# 2nd argument (rsi) contains: 8 complexes
# omega,alpha,beta,j.beta,gamma,j.gamma,k.gamma,kj.gamma
# alpha = sqrt(omega), beta = sqrt(alpha), gamma = sqrt(beta)
# j = sqrt(i), k=sqrt(j)
.globl cplx_fft16_avx_fma
cplx_fft16_avx_fma:
vmovupd (%rdi),%ymm8
vmovupd 0x20(%rdi),%ymm9
vmovupd 0x40(%rdi),%ymm10
vmovupd 0x60(%rdi),%ymm11
vmovupd 0x80(%rdi),%ymm12
vmovupd 0xa0(%rdi),%ymm13
vmovupd 0xc0(%rdi),%ymm14
vmovupd 0xe0(%rdi),%ymm15
.first_pass:
vmovupd (%rsi),%xmm0 /* omri */
vinsertf128 $1, %xmm0, %ymm0, %ymm0 /* omriri */
vshufpd $15, %ymm0, %ymm0, %ymm1 /* ymm1: omiiii */
vshufpd $0, %ymm0, %ymm0, %ymm0 /* ymm0: omrrrr */
vshufpd $5, %ymm12, %ymm12, %ymm4
vshufpd $5, %ymm13, %ymm13, %ymm5
vshufpd $5, %ymm14, %ymm14, %ymm6
vshufpd $5, %ymm15, %ymm15, %ymm7
vmulpd %ymm4,%ymm1,%ymm4
vmulpd %ymm5,%ymm1,%ymm5
vmulpd %ymm6,%ymm1,%ymm6
vmulpd %ymm7,%ymm1,%ymm7
vfmaddsub231pd %ymm12, %ymm0, %ymm4 # ymm4 = (ymm0 * ymm12) +/- ymm4
vfmaddsub231pd %ymm13, %ymm0, %ymm5
vfmaddsub231pd %ymm14, %ymm0, %ymm6
vfmaddsub231pd %ymm15, %ymm0, %ymm7
vsubpd %ymm4,%ymm8,%ymm12
vsubpd %ymm5,%ymm9,%ymm13
vsubpd %ymm6,%ymm10,%ymm14
vsubpd %ymm7,%ymm11,%ymm15
vaddpd %ymm4,%ymm8,%ymm8
vaddpd %ymm5,%ymm9,%ymm9
vaddpd %ymm6,%ymm10,%ymm10
vaddpd %ymm7,%ymm11,%ymm11
.second_pass:
vmovupd 16(%rsi),%xmm0 /* omri */
vinsertf128 $1, %xmm0, %ymm0, %ymm0 /* omriri */
vshufpd $15, %ymm0, %ymm0, %ymm1 /* ymm1: omiiii */
vshufpd $0, %ymm0, %ymm0, %ymm0 /* ymm0: omrrrr */
vshufpd $5, %ymm10, %ymm10, %ymm4
vshufpd $5, %ymm11, %ymm11, %ymm5
vshufpd $5, %ymm14, %ymm14, %ymm6
vshufpd $5, %ymm15, %ymm15, %ymm7
vmulpd %ymm4,%ymm1,%ymm4
vmulpd %ymm5,%ymm1,%ymm5
vmulpd %ymm6,%ymm0,%ymm6
vmulpd %ymm7,%ymm0,%ymm7
vfmaddsub231pd %ymm10, %ymm0, %ymm4 # ymm4 = (ymm0 * ymm10) +/- ymm4
vfmaddsub231pd %ymm11, %ymm0, %ymm5
vfmsubadd231pd %ymm14, %ymm1, %ymm6
vfmsubadd231pd %ymm15, %ymm1, %ymm7
vsubpd %ymm4,%ymm8,%ymm10
vsubpd %ymm5,%ymm9,%ymm11
vaddpd %ymm6,%ymm12,%ymm14
vaddpd %ymm7,%ymm13,%ymm15
vaddpd %ymm4,%ymm8,%ymm8
vaddpd %ymm5,%ymm9,%ymm9
vsubpd %ymm6,%ymm12,%ymm12
vsubpd %ymm7,%ymm13,%ymm13
.third_pass:
vmovupd 32(%rsi),%xmm0 /* gamma */
vmovupd 48(%rsi),%xmm2 /* delta */
vinsertf128 $1, %xmm0, %ymm0, %ymm0
vinsertf128 $1, %xmm2, %ymm2, %ymm2
vshufpd $15, %ymm0, %ymm0, %ymm1 /* ymm1: gama.iiii */
vshufpd $15, %ymm2, %ymm2, %ymm3 /* ymm3: delta.iiii */
vshufpd $0, %ymm0, %ymm0, %ymm0 /* ymm0: gama.rrrr */
vshufpd $0, %ymm2, %ymm2, %ymm2 /* ymm2: delta.rrrr */
vshufpd $5, %ymm9, %ymm9, %ymm4
vshufpd $5, %ymm11, %ymm11, %ymm5
vshufpd $5, %ymm13, %ymm13, %ymm6
vshufpd $5, %ymm15, %ymm15, %ymm7
vmulpd %ymm4,%ymm1,%ymm4
vmulpd %ymm5,%ymm0,%ymm5
vmulpd %ymm6,%ymm3,%ymm6
vmulpd %ymm7,%ymm2,%ymm7
vfmaddsub231pd %ymm9, %ymm0, %ymm4 # ymm4 = (ymm0 * ymm10) +/- ymm4
vfmsubadd231pd %ymm11, %ymm1, %ymm5
vfmaddsub231pd %ymm13, %ymm2, %ymm6
vfmsubadd231pd %ymm15, %ymm3, %ymm7
vsubpd %ymm4,%ymm8,%ymm9
vaddpd %ymm5,%ymm10,%ymm11
vsubpd %ymm6,%ymm12,%ymm13
vaddpd %ymm7,%ymm14,%ymm15
vaddpd %ymm4,%ymm8,%ymm8
vsubpd %ymm5,%ymm10,%ymm10
vaddpd %ymm6,%ymm12,%ymm12
vsubpd %ymm7,%ymm14,%ymm14
.fourth_pass:
vmovupd 64(%rsi),%ymm0 /* gamma */
vmovupd 96(%rsi),%ymm2 /* delta */
vshufpd $15, %ymm0, %ymm0, %ymm1 /* ymm1: gama.iiii */
vshufpd $15, %ymm2, %ymm2, %ymm3 /* ymm3: delta.iiii */
vshufpd $0, %ymm0, %ymm0, %ymm0 /* ymm0: gama.rrrr */
vshufpd $0, %ymm2, %ymm2, %ymm2 /* ymm2: delta.rrrr */
vperm2f128 $0x31,%ymm10,%ymm8,%ymm4 # ymm4 contains c1,c5 -- x gamma
vperm2f128 $0x31,%ymm11,%ymm9,%ymm5 # ymm5 contains c3,c7 -- x igamma
vperm2f128 $0x31,%ymm14,%ymm12,%ymm6 # ymm6 contains c9,c13 -- x delta
vperm2f128 $0x31,%ymm15,%ymm13,%ymm7 # ymm7 contains c11,c15 -- x idelta
vperm2f128 $0x20,%ymm10,%ymm8,%ymm8 # ymm8 contains c0,c4
vperm2f128 $0x20,%ymm11,%ymm9,%ymm9 # ymm9 contains c2,c6
vperm2f128 $0x20,%ymm14,%ymm12,%ymm10 # ymm10 contains c8,c12
vperm2f128 $0x20,%ymm15,%ymm13,%ymm11 # ymm11 contains c10,c14
vshufpd $5, %ymm4, %ymm4, %ymm12
vshufpd $5, %ymm5, %ymm5, %ymm13
vshufpd $5, %ymm6, %ymm6, %ymm14
vshufpd $5, %ymm7, %ymm7, %ymm15
vmulpd %ymm12,%ymm1,%ymm12
vmulpd %ymm13,%ymm0,%ymm13
vmulpd %ymm14,%ymm3,%ymm14
vmulpd %ymm15,%ymm2,%ymm15
vfmaddsub231pd %ymm4, %ymm0, %ymm12 # ymm12 = (ymm0 * ymm4) +/- ymm12
vfmsubadd231pd %ymm5, %ymm1, %ymm13
vfmaddsub231pd %ymm6, %ymm2, %ymm14
vfmsubadd231pd %ymm7, %ymm3, %ymm15
vsubpd %ymm12,%ymm8,%ymm4
vaddpd %ymm13,%ymm9,%ymm5
vsubpd %ymm14,%ymm10,%ymm6
vaddpd %ymm15,%ymm11,%ymm7
vaddpd %ymm12,%ymm8,%ymm8
vsubpd %ymm13,%ymm9,%ymm9
vaddpd %ymm14,%ymm10,%ymm10
vsubpd %ymm15,%ymm11,%ymm11
vperm2f128 $0x20,%ymm6,%ymm10,%ymm12 # ymm4 contains c1,c5 -- x gamma
vperm2f128 $0x20,%ymm7,%ymm11,%ymm13 # ymm5 contains c3,c7 -- x igamma
vperm2f128 $0x31,%ymm6,%ymm10,%ymm14 # ymm6 contains c9,c13 -- x delta
vperm2f128 $0x31,%ymm7,%ymm11,%ymm15 # ymm7 contains c11,c15 -- x idelta
vperm2f128 $0x31,%ymm4,%ymm8,%ymm10 # ymm10 contains c8,c12
vperm2f128 $0x31,%ymm5,%ymm9,%ymm11 # ymm11 contains c10,c14
vperm2f128 $0x20,%ymm4,%ymm8,%ymm8 # ymm8 contains c0,c4
vperm2f128 $0x20,%ymm5,%ymm9,%ymm9 # ymm9 contains c2,c6
.save_and_return:
vmovupd %ymm8,(%rdi)
vmovupd %ymm9,0x20(%rdi)
vmovupd %ymm10,0x40(%rdi)
vmovupd %ymm11,0x60(%rdi)
vmovupd %ymm12,0x80(%rdi)
vmovupd %ymm13,0xa0(%rdi)
vmovupd %ymm14,0xc0(%rdi)
vmovupd %ymm15,0xe0(%rdi)
ret
.size cplx_fft16_avx_fma, .-cplx_fft16_avx_fma
.section .note.GNU-stack,"",@progbits

190
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/cplx/cplx_fft16_avx_fma_win32.s
View File

@@ -1,190 +0,0 @@
.text
.p2align 4
.globl cplx_fft16_avx_fma
.def cplx_fft16_avx_fma; .scl 2; .type 32; .endef
cplx_fft16_avx_fma:
pushq %rdi
pushq %rsi
movq %rcx,%rdi
movq %rdx,%rsi
subq $0x100,%rsp
movdqu %xmm6,(%rsp)
movdqu %xmm7,0x10(%rsp)
movdqu %xmm8,0x20(%rsp)
movdqu %xmm9,0x30(%rsp)
movdqu %xmm10,0x40(%rsp)
movdqu %xmm11,0x50(%rsp)
movdqu %xmm12,0x60(%rsp)
movdqu %xmm13,0x70(%rsp)
movdqu %xmm14,0x80(%rsp)
movdqu %xmm15,0x90(%rsp)
callq cplx_fft16_avx_fma_amd64
movdqu (%rsp),%xmm6
movdqu 0x10(%rsp),%xmm7
movdqu 0x20(%rsp),%xmm8
movdqu 0x30(%rsp),%xmm9
movdqu 0x40(%rsp),%xmm10
movdqu 0x50(%rsp),%xmm11
movdqu 0x60(%rsp),%xmm12
movdqu 0x70(%rsp),%xmm13
movdqu 0x80(%rsp),%xmm14
movdqu 0x90(%rsp),%xmm15
addq $0x100,%rsp
popq %rsi
popq %rdi
retq
# shifted FFT over X^16-i
# 1st argument (rdi) contains 16 complexes
# 2nd argument (rsi) contains: 8 complexes
# omega,alpha,beta,j.beta,gamma,j.gamma,k.gamma,kj.gamma
# alpha = sqrt(omega), beta = sqrt(alpha), gamma = sqrt(beta)
# j = sqrt(i), k=sqrt(j)
cplx_fft16_avx_fma_amd64:
vmovupd (%rdi),%ymm8
vmovupd 0x20(%rdi),%ymm9
vmovupd 0x40(%rdi),%ymm10
vmovupd 0x60(%rdi),%ymm11
vmovupd 0x80(%rdi),%ymm12
vmovupd 0xa0(%rdi),%ymm13
vmovupd 0xc0(%rdi),%ymm14
vmovupd 0xe0(%rdi),%ymm15
.first_pass:
vmovupd (%rsi),%xmm0 /* omri */
vinsertf128 $1, %xmm0, %ymm0, %ymm0 /* omriri */
vshufpd $15, %ymm0, %ymm0, %ymm1 /* ymm1: omiiii */
vshufpd $0, %ymm0, %ymm0, %ymm0 /* ymm0: omrrrr */
vshufpd $5, %ymm12, %ymm12, %ymm4
vshufpd $5, %ymm13, %ymm13, %ymm5
vshufpd $5, %ymm14, %ymm14, %ymm6
vshufpd $5, %ymm15, %ymm15, %ymm7
vmulpd %ymm4,%ymm1,%ymm4
vmulpd %ymm5,%ymm1,%ymm5
vmulpd %ymm6,%ymm1,%ymm6
vmulpd %ymm7,%ymm1,%ymm7
vfmaddsub231pd %ymm12, %ymm0, %ymm4 # ymm4 = (ymm0 * ymm12) +/- ymm4
vfmaddsub231pd %ymm13, %ymm0, %ymm5
vfmaddsub231pd %ymm14, %ymm0, %ymm6
vfmaddsub231pd %ymm15, %ymm0, %ymm7
vsubpd %ymm4,%ymm8,%ymm12
vsubpd %ymm5,%ymm9,%ymm13
vsubpd %ymm6,%ymm10,%ymm14
vsubpd %ymm7,%ymm11,%ymm15
vaddpd %ymm4,%ymm8,%ymm8
vaddpd %ymm5,%ymm9,%ymm9
vaddpd %ymm6,%ymm10,%ymm10
vaddpd %ymm7,%ymm11,%ymm11
.second_pass:
vmovupd 16(%rsi),%xmm0 /* omri */
vinsertf128 $1, %xmm0, %ymm0, %ymm0 /* omriri */
vshufpd $15, %ymm0, %ymm0, %ymm1 /* ymm1: omiiii */
vshufpd $0, %ymm0, %ymm0, %ymm0 /* ymm0: omrrrr */
vshufpd $5, %ymm10, %ymm10, %ymm4
vshufpd $5, %ymm11, %ymm11, %ymm5
vshufpd $5, %ymm14, %ymm14, %ymm6
vshufpd $5, %ymm15, %ymm15, %ymm7
vmulpd %ymm4,%ymm1,%ymm4
vmulpd %ymm5,%ymm1,%ymm5
vmulpd %ymm6,%ymm0,%ymm6
vmulpd %ymm7,%ymm0,%ymm7
vfmaddsub231pd %ymm10, %ymm0, %ymm4 # ymm4 = (ymm0 * ymm10) +/- ymm4
vfmaddsub231pd %ymm11, %ymm0, %ymm5
vfmsubadd231pd %ymm14, %ymm1, %ymm6
vfmsubadd231pd %ymm15, %ymm1, %ymm7
vsubpd %ymm4,%ymm8,%ymm10
vsubpd %ymm5,%ymm9,%ymm11
vaddpd %ymm6,%ymm12,%ymm14
vaddpd %ymm7,%ymm13,%ymm15
vaddpd %ymm4,%ymm8,%ymm8
vaddpd %ymm5,%ymm9,%ymm9
vsubpd %ymm6,%ymm12,%ymm12
vsubpd %ymm7,%ymm13,%ymm13
.third_pass:
vmovupd 32(%rsi),%xmm0 /* gamma */
vmovupd 48(%rsi),%xmm2 /* delta */
vinsertf128 $1, %xmm0, %ymm0, %ymm0
vinsertf128 $1, %xmm2, %ymm2, %ymm2
vshufpd $15, %ymm0, %ymm0, %ymm1 /* ymm1: gama.iiii */
vshufpd $15, %ymm2, %ymm2, %ymm3 /* ymm3: delta.iiii */
vshufpd $0, %ymm0, %ymm0, %ymm0 /* ymm0: gama.rrrr */
vshufpd $0, %ymm2, %ymm2, %ymm2 /* ymm2: delta.rrrr */
vshufpd $5, %ymm9, %ymm9, %ymm4
vshufpd $5, %ymm11, %ymm11, %ymm5
vshufpd $5, %ymm13, %ymm13, %ymm6
vshufpd $5, %ymm15, %ymm15, %ymm7
vmulpd %ymm4,%ymm1,%ymm4
vmulpd %ymm5,%ymm0,%ymm5
vmulpd %ymm6,%ymm3,%ymm6
vmulpd %ymm7,%ymm2,%ymm7
vfmaddsub231pd %ymm9, %ymm0, %ymm4 # ymm4 = (ymm0 * ymm10) +/- ymm4
vfmsubadd231pd %ymm11, %ymm1, %ymm5
vfmaddsub231pd %ymm13, %ymm2, %ymm6
vfmsubadd231pd %ymm15, %ymm3, %ymm7
vsubpd %ymm4,%ymm8,%ymm9
vaddpd %ymm5,%ymm10,%ymm11
vsubpd %ymm6,%ymm12,%ymm13
vaddpd %ymm7,%ymm14,%ymm15
vaddpd %ymm4,%ymm8,%ymm8
vsubpd %ymm5,%ymm10,%ymm10
vaddpd %ymm6,%ymm12,%ymm12
vsubpd %ymm7,%ymm14,%ymm14
.fourth_pass:
vmovupd 64(%rsi),%ymm0 /* gamma */
vmovupd 96(%rsi),%ymm2 /* delta */
vshufpd $15, %ymm0, %ymm0, %ymm1 /* ymm1: gama.iiii */
vshufpd $15, %ymm2, %ymm2, %ymm3 /* ymm3: delta.iiii */
vshufpd $0, %ymm0, %ymm0, %ymm0 /* ymm0: gama.rrrr */
vshufpd $0, %ymm2, %ymm2, %ymm2 /* ymm2: delta.rrrr */
vperm2f128 $0x31,%ymm10,%ymm8,%ymm4 # ymm4 contains c1,c5 -- x gamma
vperm2f128 $0x31,%ymm11,%ymm9,%ymm5 # ymm5 contains c3,c7 -- x igamma
vperm2f128 $0x31,%ymm14,%ymm12,%ymm6 # ymm6 contains c9,c13 -- x delta
vperm2f128 $0x31,%ymm15,%ymm13,%ymm7 # ymm7 contains c11,c15 -- x idelta
vperm2f128 $0x20,%ymm10,%ymm8,%ymm8 # ymm8 contains c0,c4
vperm2f128 $0x20,%ymm11,%ymm9,%ymm9 # ymm9 contains c2,c6
vperm2f128 $0x20,%ymm14,%ymm12,%ymm10 # ymm10 contains c8,c12
vperm2f128 $0x20,%ymm15,%ymm13,%ymm11 # ymm11 contains c10,c14
vshufpd $5, %ymm4, %ymm4, %ymm12
vshufpd $5, %ymm5, %ymm5, %ymm13
vshufpd $5, %ymm6, %ymm6, %ymm14
vshufpd $5, %ymm7, %ymm7, %ymm15
vmulpd %ymm12,%ymm1,%ymm12
vmulpd %ymm13,%ymm0,%ymm13
vmulpd %ymm14,%ymm3,%ymm14
vmulpd %ymm15,%ymm2,%ymm15
vfmaddsub231pd %ymm4, %ymm0, %ymm12 # ymm12 = (ymm0 * ymm4) +/- ymm12
vfmsubadd231pd %ymm5, %ymm1, %ymm13
vfmaddsub231pd %ymm6, %ymm2, %ymm14
vfmsubadd231pd %ymm7, %ymm3, %ymm15
vsubpd %ymm12,%ymm8,%ymm4
vaddpd %ymm13,%ymm9,%ymm5
vsubpd %ymm14,%ymm10,%ymm6
vaddpd %ymm15,%ymm11,%ymm7
vaddpd %ymm12,%ymm8,%ymm8
vsubpd %ymm13,%ymm9,%ymm9
vaddpd %ymm14,%ymm10,%ymm10
vsubpd %ymm15,%ymm11,%ymm11
vperm2f128 $0x20,%ymm6,%ymm10,%ymm12 # ymm4 contains c1,c5 -- x gamma
vperm2f128 $0x20,%ymm7,%ymm11,%ymm13 # ymm5 contains c3,c7 -- x igamma
vperm2f128 $0x31,%ymm6,%ymm10,%ymm14 # ymm6 contains c9,c13 -- x delta
vperm2f128 $0x31,%ymm7,%ymm11,%ymm15 # ymm7 contains c11,c15 -- x idelta
vperm2f128 $0x31,%ymm4,%ymm8,%ymm10 # ymm10 contains c8,c12
vperm2f128 $0x31,%ymm5,%ymm9,%ymm11 # ymm11 contains c10,c14
vperm2f128 $0x20,%ymm4,%ymm8,%ymm8 # ymm8 contains c0,c4
vperm2f128 $0x20,%ymm5,%ymm9,%ymm9 # ymm9 contains c2,c6
.save_and_return:
vmovupd %ymm8,(%rdi)
vmovupd %ymm9,0x20(%rdi)
vmovupd %ymm10,0x40(%rdi)
vmovupd %ymm11,0x60(%rdi)
vmovupd %ymm12,0x80(%rdi)
vmovupd %ymm13,0xa0(%rdi)
vmovupd %ymm14,0xc0(%rdi)
vmovupd %ymm15,0xe0(%rdi)
ret

8
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/cplx/cplx_fft_asserts.c
View File

@@ -1,8 +0,0 @@
#include "../commons_private.h"
#include "cplx_fft_private.h"
__always_inline void my_asserts() {
STATIC_ASSERT(sizeof(FFT_FUNCTION) == 8);
STATIC_ASSERT(sizeof(CPLX_FFT_PRECOMP) == 40);
STATIC_ASSERT(sizeof(CPLX_IFFT_PRECOMP) == 40);
}

266
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/cplx/cplx_fft_avx2_fma.c
View File

@@ -1,266 +0,0 @@
#include <immintrin.h>
#include <string.h>
#include "cplx_fft_internal.h"
#include "cplx_fft_private.h"
typedef double D4MEM[4];
/**
* @brief complex fft via bfs strategy (for m between 2 and 8)
* @param dat the data to run the algorithm on
* @param omg precomputed tables (must have been filled with fill_omega)
* @param m ring dimension of the FFT (modulo X^m-i)
*/
void cplx_fft_avx2_fma_bfs_2(D4MEM* dat, const D4MEM** omg, uint32_t m) {
double* data = (double*)dat;
int32_t _2nblock = m >> 1; // = h in ref code
D4MEM* const finaldd = (D4MEM*)(data + 2 * m);
while (_2nblock >= 2) {
int32_t nblock = _2nblock >> 1; // =h/2 in ref code
D4MEM* dd = (D4MEM*)data;
do {
const __m256d om = _mm256_load_pd(*omg[0]);
const __m256d omim = _mm256_addsub_pd(_mm256_setzero_pd(), _mm256_unpackhi_pd(om, om));
const __m256d omre = _mm256_unpacklo_pd(om, om);
D4MEM* const ddend = (dd + nblock);
D4MEM* ddmid = ddend;
do {
const __m256d b = _mm256_loadu_pd(ddmid[0]);
const __m256d t1 = _mm256_mul_pd(b, omre);
const __m256d barb = _mm256_shuffle_pd(b, b, 5);
const __m256d t2 = _mm256_fmadd_pd(barb, omim, t1);
const __m256d a = _mm256_loadu_pd(dd[0]);
const __m256d newa = _mm256_add_pd(a, t2);
const __m256d newb = _mm256_sub_pd(a, t2);
_mm256_storeu_pd(dd[0], newa);
_mm256_storeu_pd(ddmid[0], newb);
dd += 1;
ddmid += 1;
} while (dd < ddend);
dd += nblock;
*omg += 1;
} while (dd < finaldd);
_2nblock >>= 1;
}
// last iteration when _2nblock == 1
{
D4MEM* dd = (D4MEM*)data;
do {
const __m256d om = _mm256_load_pd(*omg[0]);
const __m256d omre = _mm256_unpacklo_pd(om, om);
const __m256d omim = _mm256_unpackhi_pd(om, om);
const __m256d ab = _mm256_loadu_pd(dd[0]);
const __m256d bb = _mm256_permute4x64_pd(ab, 0b11101110);
const __m256d bbbar = _mm256_permute4x64_pd(ab, 0b10111011);
const __m256d t1 = _mm256_mul_pd(bbbar, omim);
const __m256d t2 = _mm256_fmaddsub_pd(bb, omre, t1);
const __m256d aa = _mm256_permute4x64_pd(ab, 0b01000100);
const __m256d newab = _mm256_add_pd(aa, t2);
_mm256_storeu_pd(dd[0], newab);
dd += 1;
*omg += 1;
} while (dd < finaldd);
}
}
__always_inline void cplx_twiddle_fft_avx2(int32_t h, D4MEM* data, const void* omg) {
const __m256d om = _mm256_loadu_pd(omg);
const __m256d omim = _mm256_unpackhi_pd(om, om);
const __m256d omre = _mm256_unpacklo_pd(om, om);
D4MEM* d0 = data;
D4MEM* const ddend = d0 + (h >> 1);
D4MEM* d1 = ddend;
do {
const __m256d b = _mm256_loadu_pd(d1[0]);
const __m256d barb = _mm256_shuffle_pd(b, b, 5);
const __m256d t1 = _mm256_mul_pd(barb, omim);
const __m256d t2 = _mm256_fmaddsub_pd(b, omre, t1);
const __m256d a = _mm256_loadu_pd(d0[0]);
const __m256d newa = _mm256_add_pd(a, t2);
const __m256d newb = _mm256_sub_pd(a, t2);
_mm256_storeu_pd(d0[0], newa);
_mm256_storeu_pd(d1[0], newb);
d0 += 1;
d1 += 1;
} while (d0 < ddend);
}
__always_inline void cplx_bitwiddle_fft_avx2(int32_t h, void* data, const void* powom) {
const __m256d omx = _mm256_loadu_pd(powom);
const __m256d oma = _mm256_permute2f128_pd(omx, omx, 0x00);
const __m256d omb = _mm256_permute2f128_pd(omx, omx, 0x11);
const __m256d omaim = _mm256_unpackhi_pd(oma, oma);
const __m256d omare = _mm256_unpacklo_pd(oma, oma);
const __m256d ombim = _mm256_unpackhi_pd(omb, omb);
const __m256d ombre = _mm256_unpacklo_pd(omb, omb);
D4MEM* d0 = (D4MEM*)data;
D4MEM* const ddend = d0 + (h >> 1);
D4MEM* d1 = ddend;
D4MEM* d2 = d0 + h;
D4MEM* d3 = d1 + h;
__m256d reg0, reg1, reg2, reg3, tmp0, tmp1;
do {
reg0 = _mm256_loadu_pd(d0[0]);
reg1 = _mm256_loadu_pd(d1[0]);
reg2 = _mm256_loadu_pd(d2[0]);
reg3 = _mm256_loadu_pd(d3[0]);
tmp0 = _mm256_shuffle_pd(reg2, reg2, 5);
tmp1 = _mm256_shuffle_pd(reg3, reg3, 5);
tmp0 = _mm256_mul_pd(tmp0, omaim);
tmp1 = _mm256_mul_pd(tmp1, omaim);
tmp0 = _mm256_fmaddsub_pd(reg2, omare, tmp0);
tmp1 = _mm256_fmaddsub_pd(reg3, omare, tmp1);
reg2 = _mm256_sub_pd(reg0, tmp0);
reg3 = _mm256_sub_pd(reg1, tmp1);
reg0 = _mm256_add_pd(reg0, tmp0);
reg1 = _mm256_add_pd(reg1, tmp1);
//--------------------------------------
tmp0 = _mm256_shuffle_pd(reg1, reg1, 5);
tmp1 = _mm256_shuffle_pd(reg3, reg3, 5);
tmp0 = _mm256_mul_pd(tmp0, ombim); //(r,i)
tmp1 = _mm256_mul_pd(tmp1, ombre); //(-i,r)
tmp0 = _mm256_fmaddsub_pd(reg1, ombre, tmp0);
tmp1 = _mm256_fmsubadd_pd(reg3, ombim, tmp1);
reg1 = _mm256_sub_pd(reg0, tmp0);
reg3 = _mm256_add_pd(reg2, tmp1);
reg0 = _mm256_add_pd(reg0, tmp0);
reg2 = _mm256_sub_pd(reg2, tmp1);
/////
_mm256_storeu_pd(d0[0], reg0);
_mm256_storeu_pd(d1[0], reg1);
_mm256_storeu_pd(d2[0], reg2);
_mm256_storeu_pd(d3[0], reg3);
d0 += 1;
d1 += 1;
d2 += 1;
d3 += 1;
} while (d0 < ddend);
}
/**
* @brief complex fft via bfs strategy (for m >= 16)
* @param dat the data to run the algorithm on
* @param omg precomputed tables (must have been filled with fill_omega)
* @param m ring dimension of the FFT (modulo X^m-i)
*/
void cplx_fft_avx2_fma_bfs_16(D4MEM* dat, const D4MEM** omg, uint32_t m) {
double* data = (double*)dat;
D4MEM* const finaldd = (D4MEM*)(data + 2 * m);
uint32_t mm = m;
uint32_t log2m = _mm_popcnt_u32(m - 1); // log2(m)
if (log2m % 2 == 1) {
uint32_t h = mm >> 1;
cplx_twiddle_fft_avx2(h, dat, **omg);
*omg += 1;
mm >>= 1;
}
while (mm > 16) {
uint32_t h = mm / 4;
for (CPLX* d = (CPLX*)data; d < (CPLX*)finaldd; d += mm) {
cplx_bitwiddle_fft_avx2(h, d, (CPLX*)*omg);
*omg += 1;
}
mm = h;
}
{
D4MEM* dd = (D4MEM*)data;
do {
cplx_fft16_avx_fma(dd, *omg);
dd += 8;
*omg += 4;
} while (dd < finaldd);
_mm256_zeroupper();
}
/*
int32_t _2nblock = m >> 1; // = h in ref code
while (_2nblock >= 16) {
int32_t nblock = _2nblock >> 1; // =h/2 in ref code
D4MEM* dd = (D4MEM*)data;
do {
const __m256d om = _mm256_load_pd(*omg[0]);
const __m256d omim = _mm256_addsub_pd(_mm256_setzero_pd(), _mm256_unpackhi_pd(om, om));
const __m256d omre = _mm256_unpacklo_pd(om, om);
D4MEM* const ddend = (dd + nblock);
D4MEM* ddmid = ddend;
do {
const __m256d b = _mm256_loadu_pd(ddmid[0]);
const __m256d t1 = _mm256_mul_pd(b, omre);
const __m256d barb = _mm256_shuffle_pd(b, b, 5);
const __m256d t2 = _mm256_fmadd_pd(barb, omim, t1);
const __m256d a = _mm256_loadu_pd(dd[0]);
const __m256d newa = _mm256_add_pd(a, t2);
const __m256d newb = _mm256_sub_pd(a, t2);
_mm256_storeu_pd(dd[0], newa);
_mm256_storeu_pd(ddmid[0], newb);
dd += 1;
ddmid += 1;
} while (dd < ddend);
dd += nblock;
*omg += 1;
} while (dd < finaldd);
_2nblock >>= 1;
}
// last iteration when _2nblock == 8
{
D4MEM* dd = (D4MEM*)data;
do {
cplx_fft16_avx_fma(dd, *omg);
dd += 8;
*omg += 4;
} while (dd < finaldd);
_mm256_zeroupper();
}
*/
}
/**
* @brief complex fft via dfs recursion (for m >= 16)
* @param dat the data to run the algorithm on
* @param omg precomputed tables (must have been filled with fill_omega)
* @param m ring dimension of the FFT (modulo X^m-i)
*/
void cplx_fft_avx2_fma_rec_16(D4MEM* dat, const D4MEM** omg, uint32_t m) {
if (m <= 8) return cplx_fft_avx2_fma_bfs_2(dat, omg, m);
if (m <= 2048) return cplx_fft_avx2_fma_bfs_16(dat, omg, m);
double* data = (double*)dat;
int32_t _2nblock = m >> 1; // = h in ref code
int32_t nblock = _2nblock >> 1; // =h/2 in ref code
D4MEM* dd = (D4MEM*)data;
const __m256d om = _mm256_load_pd(*omg[0]);
const __m256d omim = _mm256_addsub_pd(_mm256_setzero_pd(), _mm256_unpackhi_pd(om, om));
const __m256d omre = _mm256_unpacklo_pd(om, om);
D4MEM* const ddend = (dd + nblock);
D4MEM* ddmid = ddend;
do {
const __m256d b = _mm256_loadu_pd(ddmid[0]);
const __m256d t1 = _mm256_mul_pd(b, omre);
const __m256d barb = _mm256_shuffle_pd(b, b, 5);
const __m256d t2 = _mm256_fmadd_pd(barb, omim, t1);
const __m256d a = _mm256_loadu_pd(dd[0]);
const __m256d newa = _mm256_add_pd(a, t2);
const __m256d newb = _mm256_sub_pd(a, t2);
_mm256_storeu_pd(dd[0], newa);
_mm256_storeu_pd(ddmid[0], newb);
dd += 1;
ddmid += 1;
} while (dd < ddend);
*omg += 1;
cplx_fft_avx2_fma_rec_16(dat, omg, _2nblock);
cplx_fft_avx2_fma_rec_16(ddend, omg, _2nblock);
}
/**
* @brief complex fft via best strategy (for m>=1)
* @param dat the data to run the algorithm on: m complex numbers
* @param omg precomputed tables (must have been filled with fill_omega)
* @param m ring dimension of the FFT (modulo X^m-i)
*/
EXPORT void cplx_fft_avx2_fma(const CPLX_FFT_PRECOMP* precomp, void* d) {
const uint32_t m = precomp->m;
const D4MEM* omg = (D4MEM*)precomp->powomegas;
if (m <= 1) return;
if (m <= 8) return cplx_fft_avx2_fma_bfs_2(d, &omg, m);
if (m <= 2048) return cplx_fft_avx2_fma_bfs_16(d, &omg, m);
cplx_fft_avx2_fma_rec_16(d, &omg, m);
}

451
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/cplx/cplx_fft_avx512.c
View File

@@ -1,451 +0,0 @@
#include <immintrin.h>
#include "cplx_fft_internal.h"
#include "cplx_fft_private.h"
typedef double D2MEM[2];
typedef double D4MEM[4];
typedef double D8MEM[8];
EXPORT void cplx_fftvec_addmul_avx512(const CPLX_FFTVEC_ADDMUL_PRECOMP* precomp, void* r, const void* a,
const void* b) {
const uint32_t m = precomp->m;
const D8MEM* aa = (D8MEM*)a;
const D8MEM* bb = (D8MEM*)b;
D8MEM* rr = (D8MEM*)r;
const D8MEM* const aend = aa + (m >> 2);
do {
/*
BEGIN_TEMPLATE
const __m512d ari% = _mm512_loadu_pd(aa[%]);
const __m512d bri% = _mm512_loadu_pd(bb[%]);
const __m512d rri% = _mm512_loadu_pd(rr[%]);
const __m512d bir% = _mm512_shuffle_pd(bri%,bri%, 0b01010101);
const __m512d aii% = _mm512_shuffle_pd(ari%,ari%, 0b11111111);
const __m512d pro% = _mm512_fmaddsub_pd(aii%,bir%,rri%);
const __m512d arr% = _mm512_shuffle_pd(ari%,ari%, 0b00000000);
const __m512d res% = _mm512_fmaddsub_pd(arr%,bri%,pro%);
_mm512_storeu_pd(rr[%],res%);
rr += @; // ONCE
aa += @; // ONCE
bb += @; // ONCE
END_TEMPLATE
*/
// BEGIN_INTERLEAVE 2
const __m512d ari0 = _mm512_loadu_pd(aa[0]);
const __m512d ari1 = _mm512_loadu_pd(aa[1]);
const __m512d bri0 = _mm512_loadu_pd(bb[0]);
const __m512d bri1 = _mm512_loadu_pd(bb[1]);
const __m512d rri0 = _mm512_loadu_pd(rr[0]);
const __m512d rri1 = _mm512_loadu_pd(rr[1]);
const __m512d bir0 = _mm512_shuffle_pd(bri0, bri0, 0b01010101);
const __m512d bir1 = _mm512_shuffle_pd(bri1, bri1, 0b01010101);
const __m512d aii0 = _mm512_shuffle_pd(ari0, ari0, 0b11111111);
const __m512d aii1 = _mm512_shuffle_pd(ari1, ari1, 0b11111111);
const __m512d pro0 = _mm512_fmaddsub_pd(aii0, bir0, rri0);
const __m512d pro1 = _mm512_fmaddsub_pd(aii1, bir1, rri1);
const __m512d arr0 = _mm512_shuffle_pd(ari0, ari0, 0b00000000);
const __m512d arr1 = _mm512_shuffle_pd(ari1, ari1, 0b00000000);
const __m512d res0 = _mm512_fmaddsub_pd(arr0, bri0, pro0);
const __m512d res1 = _mm512_fmaddsub_pd(arr1, bri1, pro1);
_mm512_storeu_pd(rr[0], res0);
_mm512_storeu_pd(rr[1], res1);
rr += 2; // ONCE
aa += 2; // ONCE
bb += 2; // ONCE
// END_INTERLEAVE
} while (aa < aend);
}
#if 0
EXPORT void cplx_fftvec_mul_fma(uint32_t m, void* r, const void* a, const void* b) {
const double(*aa)[4] = (double(*)[4])a;
const double(*bb)[4] = (double(*)[4])b;
double(*rr)[4] = (double(*)[4])r;
const double(*const aend)[4] = aa + (m >> 1);
do {
/*
BEGIN_TEMPLATE
const __m256d ari% = _mm256_loadu_pd(aa[%]);
const __m256d bri% = _mm256_loadu_pd(bb[%]);
const __m256d bir% = _mm256_shuffle_pd(bri%,bri%, 5); // conj of b
const __m256d aii% = _mm256_shuffle_pd(ari%,ari%, 15); // im of a
const __m256d pro% = _mm256_mul_pd(aii%,bir%);
const __m256d arr% = _mm256_shuffle_pd(ari%,ari%, 0); // rr of a
const __m256d res% = _mm256_fmaddsub_pd(arr%,bri%,pro%);
_mm256_storeu_pd(rr[%],res%);
END_TEMPLATE
*/
// BEGIN_INTERLEAVE 4
const __m256d ari0 = _mm256_loadu_pd(aa[0]);
const __m256d ari1 = _mm256_loadu_pd(aa[1]);
const __m256d ari2 = _mm256_loadu_pd(aa[2]);
const __m256d ari3 = _mm256_loadu_pd(aa[3]);
const __m256d bri0 = _mm256_loadu_pd(bb[0]);
const __m256d bri1 = _mm256_loadu_pd(bb[1]);
const __m256d bri2 = _mm256_loadu_pd(bb[2]);
const __m256d bri3 = _mm256_loadu_pd(bb[3]);
const __m256d bir0 = _mm256_shuffle_pd(bri0,bri0, 5); // conj of b
const __m256d bir1 = _mm256_shuffle_pd(bri1,bri1, 5); // conj of b
const __m256d bir2 = _mm256_shuffle_pd(bri2,bri2, 5); // conj of b
const __m256d bir3 = _mm256_shuffle_pd(bri3,bri3, 5); // conj of b
const __m256d aii0 = _mm256_shuffle_pd(ari0,ari0, 15); // im of a
const __m256d aii1 = _mm256_shuffle_pd(ari1,ari1, 15); // im of a
const __m256d aii2 = _mm256_shuffle_pd(ari2,ari2, 15); // im of a
const __m256d aii3 = _mm256_shuffle_pd(ari3,ari3, 15); // im of a
const __m256d pro0 = _mm256_mul_pd(aii0,bir0);
const __m256d pro1 = _mm256_mul_pd(aii1,bir1);
const __m256d pro2 = _mm256_mul_pd(aii2,bir2);
const __m256d pro3 = _mm256_mul_pd(aii3,bir3);
const __m256d arr0 = _mm256_shuffle_pd(ari0,ari0, 0); // rr of a
const __m256d arr1 = _mm256_shuffle_pd(ari1,ari1, 0); // rr of a
const __m256d arr2 = _mm256_shuffle_pd(ari2,ari2, 0); // rr of a
const __m256d arr3 = _mm256_shuffle_pd(ari3,ari3, 0); // rr of a
const __m256d res0 = _mm256_fmaddsub_pd(arr0,bri0,pro0);
const __m256d res1 = _mm256_fmaddsub_pd(arr1,bri1,pro1);
const __m256d res2 = _mm256_fmaddsub_pd(arr2,bri2,pro2);
const __m256d res3 = _mm256_fmaddsub_pd(arr3,bri3,pro3);
_mm256_storeu_pd(rr[0],res0);
_mm256_storeu_pd(rr[1],res1);
_mm256_storeu_pd(rr[2],res2);
_mm256_storeu_pd(rr[3],res3);
// END_INTERLEAVE
rr += 4;
aa += 4;
bb += 4;
} while (aa < aend);
}
EXPORT void cplx_fftvec_add_fma(uint32_t m, void* r, const void* a, const void* b) {
const double(*aa)[4] = (double(*)[4])a;
const double(*bb)[4] = (double(*)[4])b;
double(*rr)[4] = (double(*)[4])r;
const double(*const aend)[4] = aa + (m >> 1);
do {
/*
BEGIN_TEMPLATE
const __m256d ari% = _mm256_loadu_pd(aa[%]);
const __m256d bri% = _mm256_loadu_pd(bb[%]);
const __m256d res% = _mm256_add_pd(ari%,bri%);
_mm256_storeu_pd(rr[%],res%);
END_TEMPLATE
*/
// BEGIN_INTERLEAVE 4
const __m256d ari0 = _mm256_loadu_pd(aa[0]);
const __m256d ari1 = _mm256_loadu_pd(aa[1]);
const __m256d ari2 = _mm256_loadu_pd(aa[2]);
const __m256d ari3 = _mm256_loadu_pd(aa[3]);
const __m256d bri0 = _mm256_loadu_pd(bb[0]);
const __m256d bri1 = _mm256_loadu_pd(bb[1]);
const __m256d bri2 = _mm256_loadu_pd(bb[2]);
const __m256d bri3 = _mm256_loadu_pd(bb[3]);
const __m256d res0 = _mm256_add_pd(ari0,bri0);
const __m256d res1 = _mm256_add_pd(ari1,bri1);
const __m256d res2 = _mm256_add_pd(ari2,bri2);
const __m256d res3 = _mm256_add_pd(ari3,bri3);
_mm256_storeu_pd(rr[0],res0);
_mm256_storeu_pd(rr[1],res1);
_mm256_storeu_pd(rr[2],res2);
_mm256_storeu_pd(rr[3],res3);
// END_INTERLEAVE
rr += 4;
aa += 4;
bb += 4;
} while (aa < aend);
}
EXPORT void cplx_fftvec_sub2_to_fma(uint32_t m, void* r, const void* a, const void* b) {
const double(*aa)[4] = (double(*)[4])a;
const double(*bb)[4] = (double(*)[4])b;
double(*rr)[4] = (double(*)[4])r;
const double(*const aend)[4] = aa + (m >> 1);
do {
/*
BEGIN_TEMPLATE
const __m256d ari% = _mm256_loadu_pd(aa[%]);
const __m256d bri% = _mm256_loadu_pd(bb[%]);
const __m256d sum% = _mm256_add_pd(ari%,bri%);
const __m256d rri% = _mm256_loadu_pd(rr[%]);
const __m256d res% = _mm256_sub_pd(rri%,sum%);
_mm256_storeu_pd(rr[%],res%);
END_TEMPLATE
*/
// BEGIN_INTERLEAVE 4
const __m256d ari0 = _mm256_loadu_pd(aa[0]);
const __m256d ari1 = _mm256_loadu_pd(aa[1]);
const __m256d ari2 = _mm256_loadu_pd(aa[2]);
const __m256d ari3 = _mm256_loadu_pd(aa[3]);
const __m256d bri0 = _mm256_loadu_pd(bb[0]);
const __m256d bri1 = _mm256_loadu_pd(bb[1]);
const __m256d bri2 = _mm256_loadu_pd(bb[2]);
const __m256d bri3 = _mm256_loadu_pd(bb[3]);
const __m256d sum0 = _mm256_add_pd(ari0,bri0);
const __m256d sum1 = _mm256_add_pd(ari1,bri1);
const __m256d sum2 = _mm256_add_pd(ari2,bri2);
const __m256d sum3 = _mm256_add_pd(ari3,bri3);
const __m256d rri0 = _mm256_loadu_pd(rr[0]);
const __m256d rri1 = _mm256_loadu_pd(rr[1]);
const __m256d rri2 = _mm256_loadu_pd(rr[2]);
const __m256d rri3 = _mm256_loadu_pd(rr[3]);
const __m256d res0 = _mm256_sub_pd(rri0,sum0);
const __m256d res1 = _mm256_sub_pd(rri1,sum1);
const __m256d res2 = _mm256_sub_pd(rri2,sum2);
const __m256d res3 = _mm256_sub_pd(rri3,sum3);
_mm256_storeu_pd(rr[0],res0);
_mm256_storeu_pd(rr[1],res1);
_mm256_storeu_pd(rr[2],res2);
_mm256_storeu_pd(rr[3],res3);
// END_INTERLEAVE
rr += 4;
aa += 4;
bb += 4;
} while (aa < aend);
}
EXPORT void cplx_fftvec_copy_fma(uint32_t m, void* r, const void* a) {
const double(*aa)[4] = (double(*)[4])a;
double(*rr)[4] = (double(*)[4])r;
const double(*const aend)[4] = aa + (m >> 1);
do {
/*
BEGIN_TEMPLATE
const __m256d ari% = _mm256_loadu_pd(aa[%]);
_mm256_storeu_pd(rr[%],ari%);
END_TEMPLATE
*/
// BEGIN_INTERLEAVE 4
const __m256d ari0 = _mm256_loadu_pd(aa[0]);
const __m256d ari1 = _mm256_loadu_pd(aa[1]);
const __m256d ari2 = _mm256_loadu_pd(aa[2]);
const __m256d ari3 = _mm256_loadu_pd(aa[3]);
_mm256_storeu_pd(rr[0],ari0);
_mm256_storeu_pd(rr[1],ari1);
_mm256_storeu_pd(rr[2],ari2);
_mm256_storeu_pd(rr[3],ari3);
// END_INTERLEAVE
rr += 4;
aa += 4;
} while (aa < aend);
}
EXPORT void cplx_fftvec_twiddle_fma(uint32_t m, void* a, void* b, const void* omg) {
double(*aa)[4] = (double(*)[4])a;
double(*bb)[4] = (double(*)[4])b;
const double(*const aend)[4] = aa + (m >> 1);
const __m256d om = _mm256_loadu_pd(omg);
const __m256d omrr = _mm256_shuffle_pd(om, om, 0);
const __m256d omii = _mm256_shuffle_pd(om, om, 15);
do {
/*
BEGIN_TEMPLATE
const __m256d bri% = _mm256_loadu_pd(bb[%]);
const __m256d bir% = _mm256_shuffle_pd(bri%,bri%,5);
__m256d p% = _mm256_mul_pd(bir%,omii);
p% = _mm256_fmaddsub_pd(bri%,omrr,p%);
const __m256d ari% = _mm256_loadu_pd(aa[%]);
_mm256_storeu_pd(aa[%],_mm256_add_pd(ari%,p%));
_mm256_storeu_pd(bb[%],_mm256_sub_pd(ari%,p%));
END_TEMPLATE
*/
// BEGIN_INTERLEAVE 4
const __m256d bri0 = _mm256_loadu_pd(bb[0]);
const __m256d bri1 = _mm256_loadu_pd(bb[1]);
const __m256d bri2 = _mm256_loadu_pd(bb[2]);
const __m256d bri3 = _mm256_loadu_pd(bb[3]);
const __m256d bir0 = _mm256_shuffle_pd(bri0,bri0,5);
const __m256d bir1 = _mm256_shuffle_pd(bri1,bri1,5);
const __m256d bir2 = _mm256_shuffle_pd(bri2,bri2,5);
const __m256d bir3 = _mm256_shuffle_pd(bri3,bri3,5);
__m256d p0 = _mm256_mul_pd(bir0,omii);
__m256d p1 = _mm256_mul_pd(bir1,omii);
__m256d p2 = _mm256_mul_pd(bir2,omii);
__m256d p3 = _mm256_mul_pd(bir3,omii);
p0 = _mm256_fmaddsub_pd(bri0,omrr,p0);
p1 = _mm256_fmaddsub_pd(bri1,omrr,p1);
p2 = _mm256_fmaddsub_pd(bri2,omrr,p2);
p3 = _mm256_fmaddsub_pd(bri3,omrr,p3);
const __m256d ari0 = _mm256_loadu_pd(aa[0]);
const __m256d ari1 = _mm256_loadu_pd(aa[1]);
const __m256d ari2 = _mm256_loadu_pd(aa[2]);
const __m256d ari3 = _mm256_loadu_pd(aa[3]);
_mm256_storeu_pd(aa[0],_mm256_add_pd(ari0,p0));
_mm256_storeu_pd(aa[1],_mm256_add_pd(ari1,p1));
_mm256_storeu_pd(aa[2],_mm256_add_pd(ari2,p2));
_mm256_storeu_pd(aa[3],_mm256_add_pd(ari3,p3));
_mm256_storeu_pd(bb[0],_mm256_sub_pd(ari0,p0));
_mm256_storeu_pd(bb[1],_mm256_sub_pd(ari1,p1));
_mm256_storeu_pd(bb[2],_mm256_sub_pd(ari2,p2));
_mm256_storeu_pd(bb[3],_mm256_sub_pd(ari3,p3));
// END_INTERLEAVE
bb += 4;
aa += 4;
} while (aa < aend);
}
#endif
EXPORT void cplx_fftvec_twiddle_avx512(const CPLX_FFTVEC_TWIDDLE_PRECOMP* precomp, void* a, void* b, const void* omg) {
const uint32_t m = precomp->m;
D8MEM* aa = (D8MEM*)a;
D8MEM* bb = (D8MEM*)b;
D8MEM* const aend = aa + (m >> 2);
const __m512d om = _mm512_broadcast_f64x4(_mm256_loadu_pd(omg));
const __m512d omrr = _mm512_shuffle_pd(om, om, 0b00000000);
const __m512d omii = _mm512_shuffle_pd(om, om, 0b11111111);
do {
/*
BEGIN_TEMPLATE
const __m512d bri% = _mm512_loadu_pd(bb[%]);
const __m512d bir% = _mm512_shuffle_pd(bri%,bri%,0b10011001);
__m512d p% = _mm512_mul_pd(bir%,omii);
p% = _mm512_fmaddsub_pd(bri%,omrr,p%);
const __m512d ari% = _mm512_loadu_pd(aa[%]);
_mm512_storeu_pd(aa[%],_mm512_add_pd(ari%,p%));
_mm512_storeu_pd(bb[%],_mm512_sub_pd(ari%,p%));
bb += @; // ONCE
aa += @; // ONCE
END_TEMPLATE
*/
// BEGIN_INTERLEAVE 4
const __m512d bri0 = _mm512_loadu_pd(bb[0]);
const __m512d bri1 = _mm512_loadu_pd(bb[1]);
const __m512d bri2 = _mm512_loadu_pd(bb[2]);
const __m512d bri3 = _mm512_loadu_pd(bb[3]);
const __m512d bir0 = _mm512_shuffle_pd(bri0, bri0, 0b10011001);
const __m512d bir1 = _mm512_shuffle_pd(bri1, bri1, 0b10011001);
const __m512d bir2 = _mm512_shuffle_pd(bri2, bri2, 0b10011001);
const __m512d bir3 = _mm512_shuffle_pd(bri3, bri3, 0b10011001);
__m512d p0 = _mm512_mul_pd(bir0, omii);
__m512d p1 = _mm512_mul_pd(bir1, omii);
__m512d p2 = _mm512_mul_pd(bir2, omii);
__m512d p3 = _mm512_mul_pd(bir3, omii);
p0 = _mm512_fmaddsub_pd(bri0, omrr, p0);
p1 = _mm512_fmaddsub_pd(bri1, omrr, p1);
p2 = _mm512_fmaddsub_pd(bri2, omrr, p2);
p3 = _mm512_fmaddsub_pd(bri3, omrr, p3);
const __m512d ari0 = _mm512_loadu_pd(aa[0]);
const __m512d ari1 = _mm512_loadu_pd(aa[1]);
const __m512d ari2 = _mm512_loadu_pd(aa[2]);
const __m512d ari3 = _mm512_loadu_pd(aa[3]);
_mm512_storeu_pd(aa[0], _mm512_add_pd(ari0, p0));
_mm512_storeu_pd(aa[1], _mm512_add_pd(ari1, p1));
_mm512_storeu_pd(aa[2], _mm512_add_pd(ari2, p2));
_mm512_storeu_pd(aa[3], _mm512_add_pd(ari3, p3));
_mm512_storeu_pd(bb[0], _mm512_sub_pd(ari0, p0));
_mm512_storeu_pd(bb[1], _mm512_sub_pd(ari1, p1));
_mm512_storeu_pd(bb[2], _mm512_sub_pd(ari2, p2));
_mm512_storeu_pd(bb[3], _mm512_sub_pd(ari3, p3));
bb += 4; // ONCE
aa += 4; // ONCE
// END_INTERLEAVE
} while (aa < aend);
}
EXPORT void cplx_fftvec_bitwiddle_avx512(const CPLX_FFTVEC_BITWIDDLE_PRECOMP* precomp, void* a, uint64_t slicea,
const void* omg) {
const uint32_t m = precomp->m;
const uint64_t OFFSET = slicea / sizeof(D8MEM);
D8MEM* aa = (D8MEM*)a;
const D8MEM* aend = aa + (m >> 2);
const __m512d om = _mm512_broadcast_f64x4(_mm256_loadu_pd(omg));
const __m512d om1rr = _mm512_shuffle_pd(om, om, 0);
const __m512d om1ii = _mm512_shuffle_pd(om, om, 15);
const __m512d om2rr = _mm512_shuffle_pd(om, om, 0);
const __m512d om2ii = _mm512_shuffle_pd(om, om, 0);
const __m512d om3rr = _mm512_shuffle_pd(om, om, 15);
const __m512d om3ii = _mm512_shuffle_pd(om, om, 15);
do {
/*
BEGIN_TEMPLATE
__m512d ari% = _mm512_loadu_pd(aa[%]);
__m512d bri% = _mm512_loadu_pd((aa+OFFSET)[%]);
__m512d cri% = _mm512_loadu_pd((aa+2*OFFSET)[%]);
__m512d dri% = _mm512_loadu_pd((aa+3*OFFSET)[%]);
__m512d pa% = _mm512_shuffle_pd(cri%,cri%,5);
__m512d pb% = _mm512_shuffle_pd(dri%,dri%,5);
pa% = _mm512_mul_pd(pa%,om1ii);
pb% = _mm512_mul_pd(pb%,om1ii);
pa% = _mm512_fmaddsub_pd(cri%,om1rr,pa%);
pb% = _mm512_fmaddsub_pd(dri%,om1rr,pb%);
cri% = _mm512_sub_pd(ari%,pa%);
dri% = _mm512_sub_pd(bri%,pb%);
ari% = _mm512_add_pd(ari%,pa%);
bri% = _mm512_add_pd(bri%,pb%);
pa% = _mm512_shuffle_pd(bri%,bri%,5);
pb% = _mm512_shuffle_pd(dri%,dri%,5);
pa% = _mm512_mul_pd(pa%,om2ii);
pb% = _mm512_mul_pd(pb%,om3ii);
pa% = _mm512_fmaddsub_pd(bri%,om2rr,pa%);
pb% = _mm512_fmaddsub_pd(dri%,om3rr,pb%);
bri% = _mm512_sub_pd(ari%,pa%);
dri% = _mm512_sub_pd(cri%,pb%);
ari% = _mm512_add_pd(ari%,pa%);
cri% = _mm512_add_pd(cri%,pb%);
_mm512_storeu_pd(aa[%], ari%);
_mm512_storeu_pd((aa+OFFSET)[%],bri%);
_mm512_storeu_pd((aa+2*OFFSET)[%],cri%);
_mm512_storeu_pd((aa+3*OFFSET)[%],dri%);
aa += @; // ONCE
END_TEMPLATE
*/
// BEGIN_INTERLEAVE 2
__m512d ari0 = _mm512_loadu_pd(aa[0]);
__m512d ari1 = _mm512_loadu_pd(aa[1]);
__m512d bri0 = _mm512_loadu_pd((aa + OFFSET)[0]);
__m512d bri1 = _mm512_loadu_pd((aa + OFFSET)[1]);
__m512d cri0 = _mm512_loadu_pd((aa + 2 * OFFSET)[0]);
__m512d cri1 = _mm512_loadu_pd((aa + 2 * OFFSET)[1]);
__m512d dri0 = _mm512_loadu_pd((aa + 3 * OFFSET)[0]);
__m512d dri1 = _mm512_loadu_pd((aa + 3 * OFFSET)[1]);
__m512d pa0 = _mm512_shuffle_pd(cri0, cri0, 5);
__m512d pa1 = _mm512_shuffle_pd(cri1, cri1, 5);
__m512d pb0 = _mm512_shuffle_pd(dri0, dri0, 5);
__m512d pb1 = _mm512_shuffle_pd(dri1, dri1, 5);
pa0 = _mm512_mul_pd(pa0, om1ii);
pa1 = _mm512_mul_pd(pa1, om1ii);
pb0 = _mm512_mul_pd(pb0, om1ii);
pb1 = _mm512_mul_pd(pb1, om1ii);
pa0 = _mm512_fmaddsub_pd(cri0, om1rr, pa0);
pa1 = _mm512_fmaddsub_pd(cri1, om1rr, pa1);
pb0 = _mm512_fmaddsub_pd(dri0, om1rr, pb0);
pb1 = _mm512_fmaddsub_pd(dri1, om1rr, pb1);
cri0 = _mm512_sub_pd(ari0, pa0);
cri1 = _mm512_sub_pd(ari1, pa1);
dri0 = _mm512_sub_pd(bri0, pb0);
dri1 = _mm512_sub_pd(bri1, pb1);
ari0 = _mm512_add_pd(ari0, pa0);
ari1 = _mm512_add_pd(ari1, pa1);
bri0 = _mm512_add_pd(bri0, pb0);
bri1 = _mm512_add_pd(bri1, pb1);
pa0 = _mm512_shuffle_pd(bri0, bri0, 5);
pa1 = _mm512_shuffle_pd(bri1, bri1, 5);
pb0 = _mm512_shuffle_pd(dri0, dri0, 5);
pb1 = _mm512_shuffle_pd(dri1, dri1, 5);
pa0 = _mm512_mul_pd(pa0, om2ii);
pa1 = _mm512_mul_pd(pa1, om2ii);
pb0 = _mm512_mul_pd(pb0, om3ii);
pb1 = _mm512_mul_pd(pb1, om3ii);
pa0 = _mm512_fmaddsub_pd(bri0, om2rr, pa0);
pa1 = _mm512_fmaddsub_pd(bri1, om2rr, pa1);
pb0 = _mm512_fmaddsub_pd(dri0, om3rr, pb0);
pb1 = _mm512_fmaddsub_pd(dri1, om3rr, pb1);
bri0 = _mm512_sub_pd(ari0, pa0);
bri1 = _mm512_sub_pd(ari1, pa1);
dri0 = _mm512_sub_pd(cri0, pb0);
dri1 = _mm512_sub_pd(cri1, pb1);
ari0 = _mm512_add_pd(ari0, pa0);
ari1 = _mm512_add_pd(ari1, pa1);
cri0 = _mm512_add_pd(cri0, pb0);
cri1 = _mm512_add_pd(cri1, pb1);
_mm512_storeu_pd(aa[0], ari0);
_mm512_storeu_pd(aa[1], ari1);
_mm512_storeu_pd((aa + OFFSET)[0], bri0);
_mm512_storeu_pd((aa + OFFSET)[1], bri1);
_mm512_storeu_pd((aa + 2 * OFFSET)[0], cri0);
_mm512_storeu_pd((aa + 2 * OFFSET)[1], cri1);
_mm512_storeu_pd((aa + 3 * OFFSET)[0], dri0);
_mm512_storeu_pd((aa + 3 * OFFSET)[1], dri1);
aa += 2; // ONCE
// END_INTERLEAVE
} while (aa < aend);
_mm256_zeroupper();
}

123
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/cplx/cplx_fft_internal.h
View File

@@ -1,123 +0,0 @@
#ifndef SPQLIOS_CPLX_FFT_INTERNAL_H
#define SPQLIOS_CPLX_FFT_INTERNAL_H
#include "cplx_fft.h"
/** @brief a complex number contains two doubles real,imag */
typedef double CPLX[2];
EXPORT void cplx_set(CPLX r, const CPLX a);
EXPORT void cplx_neg(CPLX r, const CPLX a);
EXPORT void cplx_add(CPLX r, const CPLX a, const CPLX b);
EXPORT void cplx_sub(CPLX r, const CPLX a, const CPLX b);
EXPORT void cplx_mul(CPLX r, const CPLX a, const CPLX b);
/**
* @brief splits 2h evaluations of one polynomials into 2 times h evaluations of even/odd polynomial
* Input: Q_0(y),...,Q_{h-1}(y),Q_0(-y),...,Q_{h-1}(-y)
* Output: P_0(z),...,P_{h-1}(z),P_h(z),...,P_{2h-1}(z)
* where Q_i(X)=P_i(X^2)+X.P_{h+i}(X^2) and y^2 = z
* @param h number of "coefficients" h >= 1
* @param data 2h complex coefficients interleaved and 256b aligned
* @param powom y represented as (yre,yim)
*/
EXPORT void cplx_split_fft_ref(int32_t h, CPLX* data, const CPLX powom);
EXPORT void cplx_bisplit_fft_ref(int32_t h, CPLX* data, const CPLX powom[2]);
/**
* Input: Q(y),Q(-y)
* Output: P_0(z),P_1(z)
* where Q(X)=P_0(X^2)+X.P_1(X^2) and y^2 = z
* @param data 2 complexes coefficients interleaved and 256b aligned
* @param powom (z,-z) interleaved: (zre,zim,-zre,-zim)
*/
EXPORT void split_fft_last_ref(CPLX* data, const CPLX powom);
EXPORT void cplx_ifft_naive(const uint32_t m, const double entry_pwr, CPLX* data);
EXPORT void cplx_ifft16_avx_fma(void* data, const void* omega);
EXPORT void cplx_ifft16_ref(void* data, const void* omega);
/**
* @brief compute the ifft evaluations of P in place
* ifft(data) = ifft_rec(data, i);
* function ifft_rec(data, omega) {
* if #data = 1: return data
* let s = sqrt(omega) w. re(s)>0
* let (u,v) = data
* return split_fft([ifft_rec(u, s), ifft_rec(v, -s)],s)
* }
* @param itables precomputed tables (contains all the powers of omega in the order they are used)
* @param data vector of m complexes (coeffs as input, evals as output)
*/
EXPORT void cplx_ifft_ref(const CPLX_IFFT_PRECOMP* itables, void* data);
EXPORT void cplx_ifft_avx2_fma(const CPLX_IFFT_PRECOMP* itables, void* data);
EXPORT void cplx_fft_naive(const uint32_t m, const double entry_pwr, CPLX* data);
EXPORT void cplx_fft16_avx_fma(void* data, const void* omega);
EXPORT void cplx_fft16_ref(void* data, const void* omega);
/**
* @brief compute the fft evaluations of P in place
* fft(data) = fft_rec(data, i);
* function fft_rec(data, omega) {
* if #data = 1: return data
* let s = sqrt(omega) w. re(s)>0
* let (u,v) = merge_fft(data, s)
* return [fft_rec(u, s), fft_rec(v, -s)]
* }
* @param tables precomputed tables (contains all the powers of omega in the order they are used)
* @param data vector of m complexes (coeffs as input, evals as output)
*/
EXPORT void cplx_fft_ref(const CPLX_FFT_PRECOMP* tables, void* data);
EXPORT void cplx_fft_avx2_fma(const CPLX_FFT_PRECOMP* tables, void* data);
/**
* @brief merges 2 times h evaluations of even/odd polynomials into 2h evaluations of a sigle polynomial
* Input: P_0(z),...,P_{h-1}(z),P_h(z),...,P_{2h-1}(z)
* Output: Q_0(y),...,Q_{h-1}(y),Q_0(-y),...,Q_{h-1}(-y)
* where Q_i(X)=P_i(X^2)+X.P_{h+i}(X^2) and y^2 = z
* @param h number of "coefficients" h >= 1
* @param data 2h complex coefficients interleaved and 256b aligned
* @param powom y represented as (yre,yim)
*/
EXPORT void cplx_twiddle_fft_ref(int32_t h, CPLX* data, const CPLX powom);
EXPORT void citwiddle(CPLX a, CPLX b, const CPLX om);
EXPORT void ctwiddle(CPLX a, CPLX b, const CPLX om);
EXPORT void invctwiddle(CPLX a, CPLX b, const CPLX ombar);
EXPORT void invcitwiddle(CPLX a, CPLX b, const CPLX ombar);
// CONVERSIONS
/** @brief r = x from ZnX (coeffs as signed int32_t's ) to double */
EXPORT void cplx_from_znx32_ref(const CPLX_FROM_ZNX32_PRECOMP* precomp, void* r, const int32_t* x);
EXPORT void cplx_from_znx32_avx2_fma(const CPLX_FROM_ZNX32_PRECOMP* precomp, void* r, const int32_t* x);
/** @brief r = x to ZnX (coeffs as signed int32_t's ) to double */
EXPORT void cplx_to_znx32_ref(const CPLX_TO_ZNX32_PRECOMP* precomp, int32_t* r, const void* x);
EXPORT void cplx_to_znx32_avx2_fma(const CPLX_TO_ZNX32_PRECOMP* precomp, int32_t* r, const void* x);
/** @brief r = x mod 1 from TnX (coeffs as signed int32_t's) to double */
EXPORT void cplx_from_tnx32_ref(const CPLX_FROM_TNX32_PRECOMP* precomp, void* r, const int32_t* x);
EXPORT void cplx_from_tnx32_avx2_fma(const CPLX_FROM_TNX32_PRECOMP* precomp, void* r, const int32_t* x);
/** @brief r = x mod 1 from TnX (coeffs as signed int32_t's) */
EXPORT void cplx_to_tnx32_ref(const CPLX_TO_TNX32_PRECOMP* precomp, int32_t* x, const void* c);
EXPORT void cplx_to_tnx32_avx2_fma(const CPLX_TO_TNX32_PRECOMP* precomp, int32_t* x, const void* c);
/** @brief r = x from RnX (coeffs as doubles ) to double */
EXPORT void cplx_from_rnx64_ref(const CPLX_FROM_RNX64_PRECOMP* precomp, void* r, const double* x);
EXPORT void cplx_from_rnx64_avx2_fma(const CPLX_FROM_RNX64_PRECOMP* precomp, void* r, const double* x);
/** @brief r = x to RnX (coeffs as doubles ) to double */
EXPORT void cplx_to_rnx64_ref(const CPLX_TO_RNX64_PRECOMP* precomp, double* r, const void* x);
EXPORT void cplx_to_rnx64_avx2_fma(const CPLX_TO_RNX64_PRECOMP* precomp, double* r, const void* x);
/** @brief r = x to integers in RnX (coeffs as doubles ) to double */
EXPORT void cplx_round_to_rnx64_ref(const CPLX_ROUND_TO_RNX64_PRECOMP* precomp, double* r, const void* x);
EXPORT void cplx_round_to_rnx64_avx2_fma(const CPLX_ROUND_TO_RNX64_PRECOMP* precomp, double* r, const void* x);
// fftvec operations
/** @brief element-wise addmul r += ab */
EXPORT void cplx_fftvec_addmul_ref(const CPLX_FFTVEC_ADDMUL_PRECOMP* precomp, void* r, const void* a, const void* b);
EXPORT void cplx_fftvec_addmul_fma(const CPLX_FFTVEC_ADDMUL_PRECOMP* tables, void* r, const void* a, const void* b);
EXPORT void cplx_fftvec_addmul_sse(const CPLX_FFTVEC_ADDMUL_PRECOMP* precomp, void* r, const void* a, const void* b);
EXPORT void cplx_fftvec_addmul_avx512(const CPLX_FFTVEC_ADDMUL_PRECOMP* precomp, void* r, const void* a, const void* b);
/** @brief element-wise mul r = ab */
EXPORT void cplx_fftvec_mul_ref(const CPLX_FFTVEC_MUL_PRECOMP* tables, void* r, const void* a, const void* b);
EXPORT void cplx_fftvec_mul_fma(const CPLX_FFTVEC_MUL_PRECOMP* tables, void* r, const void* a, const void* b);
#endif // SPQLIOS_CPLX_FFT_INTERNAL_H

109
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/cplx/cplx_fft_private.h
View File

@@ -1,109 +0,0 @@
#ifndef SPQLIOS_CPLX_FFT_PRIVATE_H
#define SPQLIOS_CPLX_FFT_PRIVATE_H
#include "cplx_fft.h"
typedef struct cplx_twiddle_precomp CPLX_FFTVEC_TWIDDLE_PRECOMP;
typedef struct cplx_bitwiddle_precomp CPLX_FFTVEC_BITWIDDLE_PRECOMP;
typedef void (*IFFT_FUNCTION)(const CPLX_IFFT_PRECOMP*, void*);
typedef void (*FFT_FUNCTION)(const CPLX_FFT_PRECOMP*, void*);
// conversions
typedef void (*FROM_ZNX32_FUNCTION)(const CPLX_FROM_ZNX32_PRECOMP*, void*, const int32_t*);
typedef void (*TO_ZNX32_FUNCTION)(const CPLX_FROM_ZNX32_PRECOMP*, int32_t*, const void*);
typedef void (*FROM_TNX32_FUNCTION)(const CPLX_FROM_TNX32_PRECOMP*, void*, const int32_t*);
typedef void (*TO_TNX32_FUNCTION)(const CPLX_TO_TNX32_PRECOMP*, int32_t*, const void*);
typedef void (*FROM_RNX64_FUNCTION)(const CPLX_FROM_RNX64_PRECOMP* precomp, void* r, const double* x);
typedef void (*TO_RNX64_FUNCTION)(const CPLX_TO_RNX64_PRECOMP* precomp, double* r, const void* x);
typedef void (*ROUND_TO_RNX64_FUNCTION)(const CPLX_ROUND_TO_RNX64_PRECOMP* precomp, double* r, const void* x);
// fftvec operations
typedef void (*FFTVEC_MUL_FUNCTION)(const CPLX_FFTVEC_MUL_PRECOMP*, void*, const void*, const void*);
typedef void (*FFTVEC_ADDMUL_FUNCTION)(const CPLX_FFTVEC_ADDMUL_PRECOMP*, void*, const void*, const void*);
typedef void (*FFTVEC_TWIDDLE_FUNCTION)(const CPLX_FFTVEC_TWIDDLE_PRECOMP*, void*, const void*, const void*);
typedef void (*FFTVEC_BITWIDDLE_FUNCTION)(const CPLX_FFTVEC_BITWIDDLE_PRECOMP*, void*, uint64_t, const void*);
struct cplx_ifft_precomp {
IFFT_FUNCTION function;
int64_t m;
uint64_t buf_size;
double* powomegas;
void* aligned_buffers;
};
struct cplx_fft_precomp {
FFT_FUNCTION function;
int64_t m;
uint64_t buf_size;
double* powomegas;
void* aligned_buffers;
};
struct cplx_from_znx32_precomp {
FROM_ZNX32_FUNCTION function;
int64_t m;
};
struct cplx_to_znx32_precomp {
TO_ZNX32_FUNCTION function;
int64_t m;
double divisor;
};
struct cplx_from_tnx32_precomp {
FROM_TNX32_FUNCTION function;
int64_t m;
};
struct cplx_to_tnx32_precomp {
TO_TNX32_FUNCTION function;
int64_t m;
double divisor;
};
struct cplx_from_rnx64_precomp {
FROM_RNX64_FUNCTION function;
int64_t m;
};
struct cplx_to_rnx64_precomp {
TO_RNX64_FUNCTION function;
int64_t m;
double divisor;
};
struct cplx_round_to_rnx64_precomp {
ROUND_TO_RNX64_FUNCTION function;
int64_t m;
double divisor;
uint32_t log2bound;
};
typedef struct cplx_mul_precomp {
FFTVEC_MUL_FUNCTION function;
int64_t m;
} CPLX_FFTVEC_MUL_PRECOMP;
typedef struct cplx_addmul_precomp {
FFTVEC_ADDMUL_FUNCTION function;
int64_t m;
} CPLX_FFTVEC_ADDMUL_PRECOMP;
struct cplx_twiddle_precomp {
FFTVEC_TWIDDLE_FUNCTION function;
int64_t m;
};
struct cplx_bitwiddle_precomp {
FFTVEC_BITWIDDLE_FUNCTION function;
int64_t m;
};
EXPORT void cplx_fftvec_twiddle_fma(const CPLX_FFTVEC_TWIDDLE_PRECOMP* tables, void* a, void* b, const void* om);
EXPORT void cplx_fftvec_twiddle_avx512(const CPLX_FFTVEC_TWIDDLE_PRECOMP* tables, void* a, void* b, const void* om);
EXPORT void cplx_fftvec_bitwiddle_fma(const CPLX_FFTVEC_BITWIDDLE_PRECOMP* tables, void* a, uint64_t slice,
const void* om);
EXPORT void cplx_fftvec_bitwiddle_avx512(const CPLX_FFTVEC_BITWIDDLE_PRECOMP* tables, void* a, uint64_t slice,
const void* om);
#endif // SPQLIOS_CPLX_FFT_PRIVATE_H

367
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/cplx/cplx_fft_ref.c
View File

@@ -1,367 +0,0 @@
#include <memory.h>
#include <stdio.h>
#include "../commons_private.h"
#include "cplx_fft_internal.h"
#include "cplx_fft_private.h"
/** @brief (a,b) <- (a+omega.b,a-omega.b) */
void ctwiddle(CPLX a, CPLX b, const CPLX om) {
double re = om[0] * b[0] - om[1] * b[1];
double im = om[0] * b[1] + om[1] * b[0];
b[0] = a[0] - re;
b[1] = a[1] - im;
a[0] += re;
a[1] += im;
}
/** @brief (a,b) <- (a+i.omega.b,a-i.omega.b) */
void citwiddle(CPLX a, CPLX b, const CPLX om) {
double re = -om[1] * b[0] - om[0] * b[1];
double im = -om[1] * b[1] + om[0] * b[0];
b[0] = a[0] - re;
b[1] = a[1] - im;
a[0] += re;
a[1] += im;
}
/**
* @brief FFT modulo X^16-omega^2 (in registers)
* @param data contains 16 complexes
* @param omega 8 complexes in this order:
* omega,alpha,beta,j.beta,gamma,j.gamma,k.gamma,kj.gamma
* alpha = sqrt(omega), beta = sqrt(alpha), gamma = sqrt(beta)
* j = sqrt(i), k=sqrt(j)
*/
void cplx_fft16_ref(void* data, const void* omega) {
CPLX* d = data;
const CPLX* om = omega;
// first pass
for (uint64_t i = 0; i < 8; ++i) {
ctwiddle(d[0 + i], d[8 + i], om[0]);
}
//
ctwiddle(d[0], d[4], om[1]);
ctwiddle(d[1], d[5], om[1]);
ctwiddle(d[2], d[6], om[1]);
ctwiddle(d[3], d[7], om[1]);
citwiddle(d[8], d[12], om[1]);
citwiddle(d[9], d[13], om[1]);
citwiddle(d[10], d[14], om[1]);
citwiddle(d[11], d[15], om[1]);
//
ctwiddle(d[0], d[2], om[2]);
ctwiddle(d[1], d[3], om[2]);
citwiddle(d[4], d[6], om[2]);
citwiddle(d[5], d[7], om[2]);
ctwiddle(d[8], d[10], om[3]);
ctwiddle(d[9], d[11], om[3]);
citwiddle(d[12], d[14], om[3]);
citwiddle(d[13], d[15], om[3]);
//
ctwiddle(d[0], d[1], om[4]);
citwiddle(d[2], d[3], om[4]);
ctwiddle(d[4], d[5], om[5]);
citwiddle(d[6], d[7], om[5]);
ctwiddle(d[8], d[9], om[6]);
citwiddle(d[10], d[11], om[6]);
ctwiddle(d[12], d[13], om[7]);
citwiddle(d[14], d[15], om[7]);
}
double cos_2pix(double x) { return m_accurate_cos(2 * M_PI * x); }
double sin_2pix(double x) { return m_accurate_sin(2 * M_PI * x); }
void cplx_set_e2pix(CPLX res, double x) {
res[0] = cos_2pix(x);
res[1] = sin_2pix(x);
}
void cplx_fft16_precomp(const double entry_pwr, CPLX** omg) {
static const double j_pow = 1. / 8.;
static const double k_pow = 1. / 16.;
const double pom = entry_pwr / 2.;
const double pom_2 = entry_pwr / 4.;
const double pom_4 = entry_pwr / 8.;
const double pom_8 = entry_pwr / 16.;
cplx_set_e2pix((*omg)[0], pom);
cplx_set_e2pix((*omg)[1], pom_2);
cplx_set_e2pix((*omg)[2], pom_4);
cplx_set_e2pix((*omg)[3], pom_4 + j_pow);
cplx_set_e2pix((*omg)[4], pom_8);
cplx_set_e2pix((*omg)[5], pom_8 + j_pow);
cplx_set_e2pix((*omg)[6], pom_8 + k_pow);
cplx_set_e2pix((*omg)[7], pom_8 + j_pow + k_pow);
*omg += 8;
}
/**
* @brief h twiddles-fft on the same omega
* (also called merge-fft)merges 2 times h evaluations of even/odd polynomials into 2h evaluations of a sigle polynomial
* Input: P_0(z),...,P_{h-1}(z),P_h(z),...,P_{2h-1}(z)
* Output: Q_0(y),...,Q_{h-1}(y),Q_0(-y),...,Q_{h-1}(-y)
* where Q_i(X)=P_i(X^2)+X.P_{h+i}(X^2) and y^2 = z
* @param h number of "coefficients" h >= 1
* @param data 2h complex coefficients interleaved and 256b aligned
* @param powom y represented as (yre,yim)
*/
void cplx_twiddle_fft_ref(int32_t h, CPLX* data, const CPLX powom) {
CPLX* d0 = data;
CPLX* d1 = data + h;
for (uint64_t i = 0; i < h; ++i) {
ctwiddle(d0[i], d1[i], powom);
}
}
void cplx_bitwiddle_fft_ref(int32_t h, CPLX* data, const CPLX powom[2]) {
CPLX* d0 = data;
CPLX* d1 = data + h;
CPLX* d2 = data + 2 * h;
CPLX* d3 = data + 3 * h;
for (uint64_t i = 0; i < h; ++i) {
ctwiddle(d0[i], d2[i], powom[0]);
ctwiddle(d1[i], d3[i], powom[0]);
}
for (uint64_t i = 0; i < h; ++i) {
ctwiddle(d0[i], d1[i], powom[1]);
citwiddle(d2[i], d3[i], powom[1]);
}
}
/**
* Input: P_0(z),P_1(z)
* Output: Q(y),Q(-y)
* where Q(X)=P_0(X^2)+X.P_1(X^2) and y^2 = z
* @param data 2 complexes coefficients interleaved and 256b aligned
* @param powom (z,-z) interleaved: (zre,zim,-zre,-zim)
*/
void merge_fft_last_ref(CPLX* data, const CPLX powom) {
CPLX prod;
cplx_mul(prod, data[1], powom);
cplx_sub(data[1], data[0], prod);
cplx_add(data[0], data[0], prod);
}
void cplx_fft_ref_bfs_2(CPLX* dat, const CPLX** omg, uint32_t m) {
CPLX* data = (CPLX*)dat;
CPLX* const dend = data + m;
for (int32_t h = m / 2; h >= 2; h >>= 1) {
for (CPLX* d = data; d < dend; d += 2 * h) {
if (memcmp((*omg)[0], (*omg)[1], 8) != 0) abort();
cplx_twiddle_fft_ref(h, d, **omg);
*omg += 2;
}
#if 0
printf("after merge %d: ", h);
for (uint64_t ii=0; ii<nn/2; ++ii) {
printf("%.6lf %.6lf ",data[ii][0],data[ii][1]);
}
printf("\n");
#endif
}
for (CPLX* d = data; d < dend; d += 2) {
// TODO see if encoding changes
if ((*omg)[0][0] != -(*omg)[1][0]) abort();
if ((*omg)[0][1] != -(*omg)[1][1]) abort();
merge_fft_last_ref(d, **omg);
*omg += 2;
}
#if 0
printf("after last: ");
for (uint64_t ii=0; ii<nn/2; ++ii) {
printf("%.6lf %.6lf ",data[ii][0],data[ii][1]);
}
printf("\n");
#endif
}
void cplx_fft_ref_bfs_16(CPLX* dat, const CPLX** omg, uint32_t m) {
CPLX* data = (CPLX*)dat;
CPLX* const dend = data + m;
uint32_t mm = m;
uint32_t log2m = log2(m);
if (log2m % 2 == 1) {
cplx_twiddle_fft_ref(mm / 2, data, **omg);
*omg += 2;
mm >>= 1;
}
while (mm > 16) {
uint32_t h = mm / 4;
for (CPLX* d = data; d < dend; d += mm) {
cplx_bitwiddle_fft_ref(h, d, *omg);
*omg += 2;
}
mm = h;
}
for (CPLX* d = data; d < dend; d += 16) {
cplx_fft16_ref(d, *omg);
*omg += 8;
}
#if 0
printf("after last: ");
for (uint64_t ii=0; ii<nn/2; ++ii) {
printf("%.6lf %.6lf ",data[ii][0],data[ii][1]);
}
printf("\n");
#endif
}
/** @brief fft modulo X^m-exp(i.2pi.entry+pwr) -- reference code */
void cplx_fft_naive(const uint32_t m, const double entry_pwr, CPLX* data) {
if (m == 1) return;
const double pom = entry_pwr / 2.;
const uint32_t h = m / 2;
// apply the twiddle factors
CPLX cpom;
cplx_set_e2pix(cpom, pom);
for (uint64_t i = 0; i < h; ++i) {
ctwiddle(data[i], data[i + h], cpom);
}
// do the recursive calls
cplx_fft_naive(h, pom, data);
cplx_fft_naive(h, pom + 0.5, data + h);
}
/** @brief fills omega for cplx_fft_bfs_16 modulo X^m-exp(i.2.pi.entry_pwr) */
void fill_cplx_fft_omegas_bfs_16(const double entry_pwr, CPLX** omg, uint32_t m) {
uint32_t mm = m;
uint32_t log2m = log2(m);
double ss = entry_pwr;
if (log2m % 2 == 1) {
uint32_t h = mm / 2;
double pom = ss / 2.;
for (uint32_t i = 0; i < m / mm; i++) {
cplx_set_e2pix(omg[0][0], pom + fracrevbits(i) / 2.);
cplx_set(omg[0][1], omg[0][0]);
*omg += 2;
}
mm = h;
ss = pom;
}
while (mm > 16) {
double pom = ss / 4.;
uint32_t h = mm / 4;
for (uint32_t i = 0; i < m / mm; i++) {
double om = pom + fracrevbits(i) / 4.;
cplx_set_e2pix(omg[0][0], 2. * om);
cplx_set_e2pix(omg[0][1], om);
*omg += 2;
}
mm = h;
ss = pom;
}
{
// mm=16
for (uint32_t i = 0; i < m / 16; i++) {
cplx_fft16_precomp(ss + fracrevbits(i), omg);
}
}
}
/** @brief fills omega for cplx_fft_bfs_2 modulo X^m-exp(i.2.pi.entry_pwr) */
void fill_cplx_fft_omegas_bfs_2(const double entry_pwr, CPLX** omg, uint32_t m) {
double pom = entry_pwr / 2.;
for (int32_t h = m / 2; h >= 2; h >>= 1) {
for (uint32_t i = 0; i < m / (2 * h); i++) {
cplx_set_e2pix(omg[0][0], pom + fracrevbits(i) / 2.);
cplx_set(omg[0][1], omg[0][0]);
*omg += 2;
}
pom /= 2;
}
{
// h=1
for (uint32_t i = 0; i < m / 2; i++) {
cplx_set_e2pix((*omg)[0], pom + fracrevbits(i) / 2.);
cplx_neg((*omg)[1], (*omg)[0]);
*omg += 2;
}
}
}
/** @brief fills omega for cplx_fft_rec modulo X^m-exp(i.2.pi.entry_pwr) */
void fill_cplx_fft_omegas_rec_16(const double entry_pwr, CPLX** omg, uint32_t m) {
// note that the cases below are for recursive calls only!
// externally, this function shall only be called with m>=4096
if (m == 1) return;
if (m <= 8) return fill_cplx_fft_omegas_bfs_2(entry_pwr, omg, m);
if (m <= 2048) return fill_cplx_fft_omegas_bfs_16(entry_pwr, omg, m);
double pom = entry_pwr / 2.;
cplx_set_e2pix((*omg)[0], pom);
cplx_set_e2pix((*omg)[1], pom);
*omg += 2;
fill_cplx_fft_omegas_rec_16(pom, omg, m / 2);
fill_cplx_fft_omegas_rec_16(pom + 0.5, omg, m / 2);
}
void cplx_fft_ref_rec_16(CPLX* dat, const CPLX** omg, uint32_t m) {
if (m == 1) return;
if (m <= 8) return cplx_fft_ref_bfs_2(dat, omg, m);
if (m <= 2048) return cplx_fft_ref_bfs_16(dat, omg, m);
const uint32_t h = m / 2;
if (memcmp((*omg)[0], (*omg)[1], 8) != 0) abort();
cplx_twiddle_fft_ref(h, dat, **omg);
*omg += 2;
cplx_fft_ref_rec_16(dat, omg, h);
cplx_fft_ref_rec_16(dat + h, omg, h);
}
void cplx_fft_ref(const CPLX_FFT_PRECOMP* precomp, void* d) {
CPLX* data = (CPLX*)d;
const int32_t m = precomp->m;
const CPLX* omg = (CPLX*)precomp->powomegas;
if (m == 1) return;
if (m <= 8) return cplx_fft_ref_bfs_2(data, &omg, m);
if (m <= 2048) return cplx_fft_ref_bfs_16(data, &omg, m);
cplx_fft_ref_rec_16(data, &omg, m);
}
EXPORT CPLX_FFT_PRECOMP* new_cplx_fft_precomp(uint32_t m, uint32_t num_buffers) {
const uint64_t OMG_SPACE = ceilto64b((2 * m) * sizeof(CPLX));
const uint64_t BUF_SIZE = ceilto64b(m * sizeof(CPLX));
void* reps = malloc(sizeof(CPLX_FFT_PRECOMP) + 63 // padding
+ OMG_SPACE // tables //TODO 16?
+ num_buffers * BUF_SIZE // buffers
);
uint64_t aligned_addr = ceilto64b((uint64_t)(reps) + sizeof(CPLX_FFT_PRECOMP));
CPLX_FFT_PRECOMP* r = (CPLX_FFT_PRECOMP*)reps;
r->m = m;
r->buf_size = BUF_SIZE;
r->powomegas = (double*)aligned_addr;
r->aligned_buffers = (void*)(aligned_addr + OMG_SPACE);
// fill in powomegas
CPLX* omg = (CPLX*)r->powomegas;
if (m <= 8) {
fill_cplx_fft_omegas_bfs_2(0.25, &omg, m);
} else if (m <= 2048) {
fill_cplx_fft_omegas_bfs_16(0.25, &omg, m);
} else {
fill_cplx_fft_omegas_rec_16(0.25, &omg, m);
}
if (((uint64_t)omg) - aligned_addr > OMG_SPACE) abort();
// dispatch the right implementation
{
if (m <= 4) {
// currently, we do not have any acceletated
// implementation for m<=4
r->function = cplx_fft_ref;
} else if (CPU_SUPPORTS("fma")) {
r->function = cplx_fft_avx2_fma;
} else {
r->function = cplx_fft_ref;
}
}
return reps;
}
EXPORT void* cplx_fft_precomp_get_buffer(const CPLX_FFT_PRECOMP* tables, uint32_t buffer_index) {
return (uint8_t*)tables->aligned_buffers + buffer_index * tables->buf_size;
}
EXPORT void cplx_fft_simple(uint32_t m, void* data) {
static CPLX_FFT_PRECOMP* p[31] = {0};
CPLX_FFT_PRECOMP** f = p + log2m(m);
if (!*f) *f = new_cplx_fft_precomp(m, 0);
(*f)->function(*f, data);
}
EXPORT void cplx_fft(const CPLX_FFT_PRECOMP* tables, void* data) { tables->function(tables, data); }

309
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/cplx/cplx_fft_sse.c
View File

@@ -1,309 +0,0 @@
#include <immintrin.h>
#include "cplx_fft_internal.h"
#include "cplx_fft_private.h"
typedef double D2MEM[2];
EXPORT void cplx_fftvec_addmul_sse(const CPLX_FFTVEC_ADDMUL_PRECOMP* precomp, void* r, const void* a, const void* b) {
const uint32_t m = precomp->m;
const D2MEM* aa = (D2MEM*)a;
const D2MEM* bb = (D2MEM*)b;
D2MEM* rr = (D2MEM*)r;
const D2MEM* const aend = aa + m;
do {
/*
BEGIN_TEMPLATE
const __m128d ari% = _mm_loadu_pd(aa[%]);
const __m128d bri% = _mm_loadu_pd(bb[%]);
const __m128d rri% = _mm_loadu_pd(rr[%]);
const __m128d bir% = _mm_shuffle_pd(bri%,bri%, 5);
const __m128d aii% = _mm_shuffle_pd(ari%,ari%, 15);
const __m128d pro% = _mm_fmaddsub_pd(aii%,bir%,rri%);
const __m128d arr% = _mm_shuffle_pd(ari%,ari%, 0);
const __m128d res% = _mm_fmaddsub_pd(arr%,bri%,pro%);
_mm_storeu_pd(rr[%],res%);
rr += @; // ONCE
aa += @; // ONCE
bb += @; // ONCE
END_TEMPLATE
*/
// BEGIN_INTERLEAVE 2
const __m128d ari0 = _mm_loadu_pd(aa[0]);
const __m128d ari1 = _mm_loadu_pd(aa[1]);
const __m128d bri0 = _mm_loadu_pd(bb[0]);
const __m128d bri1 = _mm_loadu_pd(bb[1]);
const __m128d rri0 = _mm_loadu_pd(rr[0]);
const __m128d rri1 = _mm_loadu_pd(rr[1]);
const __m128d bir0 = _mm_shuffle_pd(bri0, bri0, 0b01);
const __m128d bir1 = _mm_shuffle_pd(bri1, bri1, 0b01);
const __m128d aii0 = _mm_shuffle_pd(ari0, ari0, 0b11);
const __m128d aii1 = _mm_shuffle_pd(ari1, ari1, 0b11);
const __m128d pro0 = _mm_fmaddsub_pd(aii0, bir0, rri0);
const __m128d pro1 = _mm_fmaddsub_pd(aii1, bir1, rri1);
const __m128d arr0 = _mm_shuffle_pd(ari0, ari0, 0b00);
const __m128d arr1 = _mm_shuffle_pd(ari1, ari1, 0b00);
const __m128d res0 = _mm_fmaddsub_pd(arr0, bri0, pro0);
const __m128d res1 = _mm_fmaddsub_pd(arr1, bri1, pro1);
_mm_storeu_pd(rr[0], res0);
_mm_storeu_pd(rr[1], res1);
rr += 2; // ONCE
aa += 2; // ONCE
bb += 2; // ONCE
// END_INTERLEAVE
} while (aa < aend);
}
#if 0
EXPORT void cplx_fftvec_mul_fma(const CPLX_FFTVEC_MUL_PRECOMP* precomp, void* r, const void* a, const void* b) {
const uint32_t m = precomp->m;
const double(*aa)[4] = (double(*)[4])a;
const double(*bb)[4] = (double(*)[4])b;
double(*rr)[4] = (double(*)[4])r;
const double(*const aend)[4] = aa + (m >> 1);
do {
/*
BEGIN_TEMPLATE
const __m256d ari% = _mm256_loadu_pd(aa[%]);
const __m256d bri% = _mm256_loadu_pd(bb[%]);
const __m256d bir% = _mm256_shuffle_pd(bri%,bri%, 5); // conj of b
const __m256d aii% = _mm256_shuffle_pd(ari%,ari%, 15); // im of a
const __m256d pro% = _mm256_mul_pd(aii%,bir%);
const __m256d arr% = _mm256_shuffle_pd(ari%,ari%, 0); // rr of a
const __m256d res% = _mm256_fmaddsub_pd(arr%,bri%,pro%);
_mm256_storeu_pd(rr[%],res%);
END_TEMPLATE
*/
// BEGIN_INTERLEAVE 4
const __m256d ari0 = _mm256_loadu_pd(aa[0]);
const __m256d ari1 = _mm256_loadu_pd(aa[1]);
const __m256d ari2 = _mm256_loadu_pd(aa[2]);
const __m256d ari3 = _mm256_loadu_pd(aa[3]);
const __m256d bri0 = _mm256_loadu_pd(bb[0]);
const __m256d bri1 = _mm256_loadu_pd(bb[1]);
const __m256d bri2 = _mm256_loadu_pd(bb[2]);
const __m256d bri3 = _mm256_loadu_pd(bb[3]);
const __m256d bir0 = _mm256_shuffle_pd(bri0,bri0, 5); // conj of b
const __m256d bir1 = _mm256_shuffle_pd(bri1,bri1, 5); // conj of b
const __m256d bir2 = _mm256_shuffle_pd(bri2,bri2, 5); // conj of b
const __m256d bir3 = _mm256_shuffle_pd(bri3,bri3, 5); // conj of b
const __m256d aii0 = _mm256_shuffle_pd(ari0,ari0, 15); // im of a
const __m256d aii1 = _mm256_shuffle_pd(ari1,ari1, 15); // im of a
const __m256d aii2 = _mm256_shuffle_pd(ari2,ari2, 15); // im of a
const __m256d aii3 = _mm256_shuffle_pd(ari3,ari3, 15); // im of a
const __m256d pro0 = _mm256_mul_pd(aii0,bir0);
const __m256d pro1 = _mm256_mul_pd(aii1,bir1);
const __m256d pro2 = _mm256_mul_pd(aii2,bir2);
const __m256d pro3 = _mm256_mul_pd(aii3,bir3);
const __m256d arr0 = _mm256_shuffle_pd(ari0,ari0, 0); // rr of a
const __m256d arr1 = _mm256_shuffle_pd(ari1,ari1, 0); // rr of a
const __m256d arr2 = _mm256_shuffle_pd(ari2,ari2, 0); // rr of a
const __m256d arr3 = _mm256_shuffle_pd(ari3,ari3, 0); // rr of a
const __m256d res0 = _mm256_fmaddsub_pd(arr0,bri0,pro0);
const __m256d res1 = _mm256_fmaddsub_pd(arr1,bri1,pro1);
const __m256d res2 = _mm256_fmaddsub_pd(arr2,bri2,pro2);
const __m256d res3 = _mm256_fmaddsub_pd(arr3,bri3,pro3);
_mm256_storeu_pd(rr[0],res0);
_mm256_storeu_pd(rr[1],res1);
_mm256_storeu_pd(rr[2],res2);
_mm256_storeu_pd(rr[3],res3);
// END_INTERLEAVE
rr += 4;
aa += 4;
bb += 4;
} while (aa < aend);
}
EXPORT void cplx_fftvec_add_fma(uint32_t m, void* r, const void* a, const void* b) {
const double(*aa)[4] = (double(*)[4])a;
const double(*bb)[4] = (double(*)[4])b;
double(*rr)[4] = (double(*)[4])r;
const double(*const aend)[4] = aa + (m >> 1);
do {
/*
BEGIN_TEMPLATE
const __m256d ari% = _mm256_loadu_pd(aa[%]);
const __m256d bri% = _mm256_loadu_pd(bb[%]);
const __m256d res% = _mm256_add_pd(ari%,bri%);
_mm256_storeu_pd(rr[%],res%);
END_TEMPLATE
*/
// BEGIN_INTERLEAVE 4
const __m256d ari0 = _mm256_loadu_pd(aa[0]);
const __m256d ari1 = _mm256_loadu_pd(aa[1]);
const __m256d ari2 = _mm256_loadu_pd(aa[2]);
const __m256d ari3 = _mm256_loadu_pd(aa[3]);
const __m256d bri0 = _mm256_loadu_pd(bb[0]);
const __m256d bri1 = _mm256_loadu_pd(bb[1]);
const __m256d bri2 = _mm256_loadu_pd(bb[2]);
const __m256d bri3 = _mm256_loadu_pd(bb[3]);
const __m256d res0 = _mm256_add_pd(ari0,bri0);
const __m256d res1 = _mm256_add_pd(ari1,bri1);
const __m256d res2 = _mm256_add_pd(ari2,bri2);
const __m256d res3 = _mm256_add_pd(ari3,bri3);
_mm256_storeu_pd(rr[0],res0);
_mm256_storeu_pd(rr[1],res1);
_mm256_storeu_pd(rr[2],res2);
_mm256_storeu_pd(rr[3],res3);
// END_INTERLEAVE
rr += 4;
aa += 4;
bb += 4;
} while (aa < aend);
}
EXPORT void cplx_fftvec_sub2_to_fma(uint32_t m, void* r, const void* a, const void* b) {
const double(*aa)[4] = (double(*)[4])a;
const double(*bb)[4] = (double(*)[4])b;
double(*rr)[4] = (double(*)[4])r;
const double(*const aend)[4] = aa + (m >> 1);
do {
/*
BEGIN_TEMPLATE
const __m256d ari% = _mm256_loadu_pd(aa[%]);
const __m256d bri% = _mm256_loadu_pd(bb[%]);
const __m256d sum% = _mm256_add_pd(ari%,bri%);
const __m256d rri% = _mm256_loadu_pd(rr[%]);
const __m256d res% = _mm256_sub_pd(rri%,sum%);
_mm256_storeu_pd(rr[%],res%);
END_TEMPLATE
*/
// BEGIN_INTERLEAVE 4
const __m256d ari0 = _mm256_loadu_pd(aa[0]);
const __m256d ari1 = _mm256_loadu_pd(aa[1]);
const __m256d ari2 = _mm256_loadu_pd(aa[2]);
const __m256d ari3 = _mm256_loadu_pd(aa[3]);
const __m256d bri0 = _mm256_loadu_pd(bb[0]);
const __m256d bri1 = _mm256_loadu_pd(bb[1]);
const __m256d bri2 = _mm256_loadu_pd(bb[2]);
const __m256d bri3 = _mm256_loadu_pd(bb[3]);
const __m256d sum0 = _mm256_add_pd(ari0,bri0);
const __m256d sum1 = _mm256_add_pd(ari1,bri1);
const __m256d sum2 = _mm256_add_pd(ari2,bri2);
const __m256d sum3 = _mm256_add_pd(ari3,bri3);
const __m256d rri0 = _mm256_loadu_pd(rr[0]);
const __m256d rri1 = _mm256_loadu_pd(rr[1]);
const __m256d rri2 = _mm256_loadu_pd(rr[2]);
const __m256d rri3 = _mm256_loadu_pd(rr[3]);
const __m256d res0 = _mm256_sub_pd(rri0,sum0);
const __m256d res1 = _mm256_sub_pd(rri1,sum1);
const __m256d res2 = _mm256_sub_pd(rri2,sum2);
const __m256d res3 = _mm256_sub_pd(rri3,sum3);
_mm256_storeu_pd(rr[0],res0);
_mm256_storeu_pd(rr[1],res1);
_mm256_storeu_pd(rr[2],res2);
_mm256_storeu_pd(rr[3],res3);
// END_INTERLEAVE
rr += 4;
aa += 4;
bb += 4;
} while (aa < aend);
}
EXPORT void cplx_fftvec_copy_fma(uint32_t m, void* r, const void* a) {
const double(*aa)[4] = (double(*)[4])a;
double(*rr)[4] = (double(*)[4])r;
const double(*const aend)[4] = aa + (m >> 1);
do {
/*
BEGIN_TEMPLATE
const __m256d ari% = _mm256_loadu_pd(aa[%]);
_mm256_storeu_pd(rr[%],ari%);
END_TEMPLATE
*/
// BEGIN_INTERLEAVE 4
const __m256d ari0 = _mm256_loadu_pd(aa[0]);
const __m256d ari1 = _mm256_loadu_pd(aa[1]);
const __m256d ari2 = _mm256_loadu_pd(aa[2]);
const __m256d ari3 = _mm256_loadu_pd(aa[3]);
_mm256_storeu_pd(rr[0],ari0);
_mm256_storeu_pd(rr[1],ari1);
_mm256_storeu_pd(rr[2],ari2);
_mm256_storeu_pd(rr[3],ari3);
// END_INTERLEAVE
rr += 4;
aa += 4;
} while (aa < aend);
}
EXPORT void cplx_fftvec_twiddle_fma(uint32_t m, void* a, void* b, const void* omg) {
double(*aa)[4] = (double(*)[4])a;
double(*bb)[4] = (double(*)[4])b;
const double(*const aend)[4] = aa + (m >> 1);
const __m256d om = _mm256_loadu_pd(omg);
const __m256d omrr = _mm256_shuffle_pd(om, om, 0);
const __m256d omii = _mm256_shuffle_pd(om, om, 15);
do {
/*
BEGIN_TEMPLATE
const __m256d bri% = _mm256_loadu_pd(bb[%]);
const __m256d bir% = _mm256_shuffle_pd(bri%,bri%,5);
__m256d p% = _mm256_mul_pd(bir%,omii);
p% = _mm256_fmaddsub_pd(bri%,omrr,p%);
const __m256d ari% = _mm256_loadu_pd(aa[%]);
_mm256_storeu_pd(aa[%],_mm256_add_pd(ari%,p%));
_mm256_storeu_pd(bb[%],_mm256_sub_pd(ari%,p%));
END_TEMPLATE
*/
// BEGIN_INTERLEAVE 4
const __m256d bri0 = _mm256_loadu_pd(bb[0]);
const __m256d bri1 = _mm256_loadu_pd(bb[1]);
const __m256d bri2 = _mm256_loadu_pd(bb[2]);
const __m256d bri3 = _mm256_loadu_pd(bb[3]);
const __m256d bir0 = _mm256_shuffle_pd(bri0,bri0,5);
const __m256d bir1 = _mm256_shuffle_pd(bri1,bri1,5);
const __m256d bir2 = _mm256_shuffle_pd(bri2,bri2,5);
const __m256d bir3 = _mm256_shuffle_pd(bri3,bri3,5);
__m256d p0 = _mm256_mul_pd(bir0,omii);
__m256d p1 = _mm256_mul_pd(bir1,omii);
__m256d p2 = _mm256_mul_pd(bir2,omii);
__m256d p3 = _mm256_mul_pd(bir3,omii);
p0 = _mm256_fmaddsub_pd(bri0,omrr,p0);
p1 = _mm256_fmaddsub_pd(bri1,omrr,p1);
p2 = _mm256_fmaddsub_pd(bri2,omrr,p2);
p3 = _mm256_fmaddsub_pd(bri3,omrr,p3);
const __m256d ari0 = _mm256_loadu_pd(aa[0]);
const __m256d ari1 = _mm256_loadu_pd(aa[1]);
const __m256d ari2 = _mm256_loadu_pd(aa[2]);
const __m256d ari3 = _mm256_loadu_pd(aa[3]);
_mm256_storeu_pd(aa[0],_mm256_add_pd(ari0,p0));
_mm256_storeu_pd(aa[1],_mm256_add_pd(ari1,p1));
_mm256_storeu_pd(aa[2],_mm256_add_pd(ari2,p2));
_mm256_storeu_pd(aa[3],_mm256_add_pd(ari3,p3));
_mm256_storeu_pd(bb[0],_mm256_sub_pd(ari0,p0));
_mm256_storeu_pd(bb[1],_mm256_sub_pd(ari1,p1));
_mm256_storeu_pd(bb[2],_mm256_sub_pd(ari2,p2));
_mm256_storeu_pd(bb[3],_mm256_sub_pd(ari3,p3));
// END_INTERLEAVE
bb += 4;
aa += 4;
} while (aa < aend);
}
EXPORT void cplx_fftvec_innerprod_avx2_fma(const CPLX_FFTVEC_INNERPROD_PRECOMP* precomp, const int32_t ellbar,
const uint64_t lda, const uint64_t ldb,
void* r, const void* a, const void* b) {
const uint32_t m = precomp->m;
const uint32_t blk = precomp->blk;
const uint32_t nblocks = precomp->nblocks;
const CPLX* aa = (CPLX*)a;
const CPLX* bb = (CPLX*)b;
CPLX* rr = (CPLX*)r;
const uint64_t ldda = lda >> 4; // in CPLX
const uint64_t lddb = ldb >> 4;
if (m==0) {
memset(r, 0, m*sizeof(CPLX));
return;
}
for (uint32_t k=0; k<nblocks; ++k) {
const uint64_t offset = k*blk;
const CPLX* aaa = aa+offset;
const CPLX* bbb = bb+offset;
CPLX *rrr = rr+offset;
cplx_fftvec_mul_fma(&precomp->mul_func, rrr, aaa, bbb);
for (int32_t i=1; i<ellbar; ++i) {
cplx_fftvec_addmul_fma(&precomp->addmul_func, rrr, aaa + i * ldda, bbb + i * lddb);
}
}
}
#endif

387
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/cplx/cplx_fftvec_avx2_fma.c
View File

@@ -1,387 +0,0 @@
#include <immintrin.h>
#include <string.h>
#include "cplx_fft_internal.h"
#include "cplx_fft_private.h"
typedef double D4MEM[4];
EXPORT void cplx_fftvec_mul_fma(const CPLX_FFTVEC_MUL_PRECOMP* precomp, void* r, const void* a, const void* b) {
const uint32_t m = precomp->m;
const double(*aa)[4] = (double(*)[4])a;
const double(*bb)[4] = (double(*)[4])b;
double(*rr)[4] = (double(*)[4])r;
const double(*const aend)[4] = aa + (m >> 1);
do {
/*
BEGIN_TEMPLATE
const __m256d ari% = _mm256_loadu_pd(aa[%]);
const __m256d bri% = _mm256_loadu_pd(bb[%]);
const __m256d bir% = _mm256_shuffle_pd(bri%,bri%, 5); // conj of b
const __m256d aii% = _mm256_shuffle_pd(ari%,ari%, 15); // im of a
const __m256d pro% = _mm256_mul_pd(aii%,bir%);
const __m256d arr% = _mm256_shuffle_pd(ari%,ari%, 0); // rr of a
const __m256d res% = _mm256_fmaddsub_pd(arr%,bri%,pro%);
_mm256_storeu_pd(rr[%],res%);
rr += @; // ONCE
aa += @; // ONCE
bb += @; // ONCE
END_TEMPLATE
*/
// BEGIN_INTERLEAVE 4
// This block is automatically generated from the template above
// by the interleave.pl script. Please do not edit by hand
const __m256d ari0 = _mm256_loadu_pd(aa[0]);
const __m256d ari1 = _mm256_loadu_pd(aa[1]);
const __m256d ari2 = _mm256_loadu_pd(aa[2]);
const __m256d ari3 = _mm256_loadu_pd(aa[3]);
const __m256d bri0 = _mm256_loadu_pd(bb[0]);
const __m256d bri1 = _mm256_loadu_pd(bb[1]);
const __m256d bri2 = _mm256_loadu_pd(bb[2]);
const __m256d bri3 = _mm256_loadu_pd(bb[3]);
const __m256d bir0 = _mm256_shuffle_pd(bri0, bri0, 5); // conj of b
const __m256d bir1 = _mm256_shuffle_pd(bri1, bri1, 5); // conj of b
const __m256d bir2 = _mm256_shuffle_pd(bri2, bri2, 5); // conj of b
const __m256d bir3 = _mm256_shuffle_pd(bri3, bri3, 5); // conj of b
const __m256d aii0 = _mm256_shuffle_pd(ari0, ari0, 15); // im of a
const __m256d aii1 = _mm256_shuffle_pd(ari1, ari1, 15); // im of a
const __m256d aii2 = _mm256_shuffle_pd(ari2, ari2, 15); // im of a
const __m256d aii3 = _mm256_shuffle_pd(ari3, ari3, 15); // im of a
const __m256d pro0 = _mm256_mul_pd(aii0, bir0);
const __m256d pro1 = _mm256_mul_pd(aii1, bir1);
const __m256d pro2 = _mm256_mul_pd(aii2, bir2);
const __m256d pro3 = _mm256_mul_pd(aii3, bir3);
const __m256d arr0 = _mm256_shuffle_pd(ari0, ari0, 0); // rr of a
const __m256d arr1 = _mm256_shuffle_pd(ari1, ari1, 0); // rr of a
const __m256d arr2 = _mm256_shuffle_pd(ari2, ari2, 0); // rr of a
const __m256d arr3 = _mm256_shuffle_pd(ari3, ari3, 0); // rr of a
const __m256d res0 = _mm256_fmaddsub_pd(arr0, bri0, pro0);
const __m256d res1 = _mm256_fmaddsub_pd(arr1, bri1, pro1);
const __m256d res2 = _mm256_fmaddsub_pd(arr2, bri2, pro2);
const __m256d res3 = _mm256_fmaddsub_pd(arr3, bri3, pro3);
_mm256_storeu_pd(rr[0], res0);
_mm256_storeu_pd(rr[1], res1);
_mm256_storeu_pd(rr[2], res2);
_mm256_storeu_pd(rr[3], res3);
rr += 4; // ONCE
aa += 4; // ONCE
bb += 4; // ONCE
// END_INTERLEAVE
} while (aa < aend);
}
EXPORT void cplx_fftvec_addmul_fma(const CPLX_FFTVEC_ADDMUL_PRECOMP* precomp, void* r, const void* a, const void* b) {
const uint32_t m = precomp->m;
const double(*aa)[4] = (double(*)[4])a;
const double(*bb)[4] = (double(*)[4])b;
double(*rr)[4] = (double(*)[4])r;
const double(*const aend)[4] = aa + (m >> 1);
do {
/*
BEGIN_TEMPLATE
const __m256d ari% = _mm256_loadu_pd(aa[%]);
const __m256d bri% = _mm256_loadu_pd(bb[%]);
const __m256d rri% = _mm256_loadu_pd(rr[%]);
const __m256d bir% = _mm256_shuffle_pd(bri%,bri%, 5);
const __m256d aii% = _mm256_shuffle_pd(ari%,ari%, 15);
const __m256d pro% = _mm256_fmaddsub_pd(aii%,bir%,rri%);
const __m256d arr% = _mm256_shuffle_pd(ari%,ari%, 0);
const __m256d res% = _mm256_fmaddsub_pd(arr%,bri%,pro%);
_mm256_storeu_pd(rr[%],res%);
rr += @; // ONCE
aa += @; // ONCE
bb += @; // ONCE
END_TEMPLATE
*/
// BEGIN_INTERLEAVE 2
// This block is automatically generated from the template above
// by the interleave.pl script. Please do not edit by hand
const __m256d ari0 = _mm256_loadu_pd(aa[0]);
const __m256d ari1 = _mm256_loadu_pd(aa[1]);
const __m256d bri0 = _mm256_loadu_pd(bb[0]);
const __m256d bri1 = _mm256_loadu_pd(bb[1]);
const __m256d rri0 = _mm256_loadu_pd(rr[0]);
const __m256d rri1 = _mm256_loadu_pd(rr[1]);
const __m256d bir0 = _mm256_shuffle_pd(bri0, bri0, 5);
const __m256d bir1 = _mm256_shuffle_pd(bri1, bri1, 5);
const __m256d aii0 = _mm256_shuffle_pd(ari0, ari0, 15);
const __m256d aii1 = _mm256_shuffle_pd(ari1, ari1, 15);
const __m256d pro0 = _mm256_fmaddsub_pd(aii0, bir0, rri0);
const __m256d pro1 = _mm256_fmaddsub_pd(aii1, bir1, rri1);
const __m256d arr0 = _mm256_shuffle_pd(ari0, ari0, 0);
const __m256d arr1 = _mm256_shuffle_pd(ari1, ari1, 0);
const __m256d res0 = _mm256_fmaddsub_pd(arr0, bri0, pro0);
const __m256d res1 = _mm256_fmaddsub_pd(arr1, bri1, pro1);
_mm256_storeu_pd(rr[0], res0);
_mm256_storeu_pd(rr[1], res1);
rr += 2; // ONCE
aa += 2; // ONCE
bb += 2; // ONCE
// END_INTERLEAVE
} while (aa < aend);
}
EXPORT void cplx_fftvec_bitwiddle_fma(const CPLX_FFTVEC_BITWIDDLE_PRECOMP* precomp, void* a, uint64_t slicea,
const void* omg) {
const uint32_t m = precomp->m;
const uint64_t OFFSET = slicea / sizeof(D4MEM);
D4MEM* aa = (D4MEM*)a;
const double(*const aend)[4] = aa + (m >> 1);
const __m256d om = _mm256_loadu_pd(omg);
const __m256d om1rr = _mm256_shuffle_pd(om, om, 0);
const __m256d om1ii = _mm256_shuffle_pd(om, om, 15);
const __m256d om2rr = _mm256_shuffle_pd(om, om, 0);
const __m256d om2ii = _mm256_shuffle_pd(om, om, 0);
const __m256d om3rr = _mm256_shuffle_pd(om, om, 15);
const __m256d om3ii = _mm256_shuffle_pd(om, om, 15);
do {
/*
BEGIN_TEMPLATE
__m256d ari% = _mm256_loadu_pd(aa[%]);
__m256d bri% = _mm256_loadu_pd((aa+OFFSET)[%]);
__m256d cri% = _mm256_loadu_pd((aa+2*OFFSET)[%]);
__m256d dri% = _mm256_loadu_pd((aa+3*OFFSET)[%]);
__m256d pa% = _mm256_shuffle_pd(cri%,cri%,5);
__m256d pb% = _mm256_shuffle_pd(dri%,dri%,5);
pa% = _mm256_mul_pd(pa%,om1ii);
pb% = _mm256_mul_pd(pb%,om1ii);
pa% = _mm256_fmaddsub_pd(cri%,om1rr,pa%);
pb% = _mm256_fmaddsub_pd(dri%,om1rr,pb%);
cri% = _mm256_sub_pd(ari%,pa%);
dri% = _mm256_sub_pd(bri%,pb%);
ari% = _mm256_add_pd(ari%,pa%);
bri% = _mm256_add_pd(bri%,pb%);
pa% = _mm256_shuffle_pd(bri%,bri%,5);
pb% = _mm256_shuffle_pd(dri%,dri%,5);
pa% = _mm256_mul_pd(pa%,om2ii);
pb% = _mm256_mul_pd(pb%,om3ii);
pa% = _mm256_fmaddsub_pd(bri%,om2rr,pa%);
pb% = _mm256_fmaddsub_pd(dri%,om3rr,pb%);
bri% = _mm256_sub_pd(ari%,pa%);
dri% = _mm256_sub_pd(cri%,pb%);
ari% = _mm256_add_pd(ari%,pa%);
cri% = _mm256_add_pd(cri%,pb%);
_mm256_storeu_pd(aa[%], ari%);
_mm256_storeu_pd((aa+OFFSET)[%],bri%);
_mm256_storeu_pd((aa+2*OFFSET)[%],cri%);
_mm256_storeu_pd((aa+3*OFFSET)[%],dri%);
aa += @; // ONCE
END_TEMPLATE
*/
// BEGIN_INTERLEAVE 1
// This block is automatically generated from the template above
// by the interleave.pl script. Please do not edit by hand
__m256d ari0 = _mm256_loadu_pd(aa[0]);
__m256d bri0 = _mm256_loadu_pd((aa + OFFSET)[0]);
__m256d cri0 = _mm256_loadu_pd((aa + 2 * OFFSET)[0]);
__m256d dri0 = _mm256_loadu_pd((aa + 3 * OFFSET)[0]);
__m256d pa0 = _mm256_shuffle_pd(cri0, cri0, 5);
__m256d pb0 = _mm256_shuffle_pd(dri0, dri0, 5);
pa0 = _mm256_mul_pd(pa0, om1ii);
pb0 = _mm256_mul_pd(pb0, om1ii);
pa0 = _mm256_fmaddsub_pd(cri0, om1rr, pa0);
pb0 = _mm256_fmaddsub_pd(dri0, om1rr, pb0);
cri0 = _mm256_sub_pd(ari0, pa0);
dri0 = _mm256_sub_pd(bri0, pb0);
ari0 = _mm256_add_pd(ari0, pa0);
bri0 = _mm256_add_pd(bri0, pb0);
pa0 = _mm256_shuffle_pd(bri0, bri0, 5);
pb0 = _mm256_shuffle_pd(dri0, dri0, 5);
pa0 = _mm256_mul_pd(pa0, om2ii);
pb0 = _mm256_mul_pd(pb0, om3ii);
pa0 = _mm256_fmaddsub_pd(bri0, om2rr, pa0);
pb0 = _mm256_fmaddsub_pd(dri0, om3rr, pb0);
bri0 = _mm256_sub_pd(ari0, pa0);
dri0 = _mm256_sub_pd(cri0, pb0);
ari0 = _mm256_add_pd(ari0, pa0);
cri0 = _mm256_add_pd(cri0, pb0);
_mm256_storeu_pd(aa[0], ari0);
_mm256_storeu_pd((aa + OFFSET)[0], bri0);
_mm256_storeu_pd((aa + 2 * OFFSET)[0], cri0);
_mm256_storeu_pd((aa + 3 * OFFSET)[0], dri0);
aa += 1; // ONCE
// END_INTERLEAVE
} while (aa < aend);
}
EXPORT void cplx_fftvec_sub2_to_fma(uint32_t m, void* r, const void* a, const void* b) {
const double(*aa)[4] = (double(*)[4])a;
const double(*bb)[4] = (double(*)[4])b;
double(*rr)[4] = (double(*)[4])r;
const double(*const aend)[4] = aa + (m >> 1);
do {
/*
BEGIN_TEMPLATE
const __m256d ari% = _mm256_loadu_pd(aa[%]);
const __m256d bri% = _mm256_loadu_pd(bb[%]);
const __m256d sum% = _mm256_add_pd(ari%,bri%);
const __m256d rri% = _mm256_loadu_pd(rr[%]);
const __m256d res% = _mm256_sub_pd(rri%,sum%);
_mm256_storeu_pd(rr[%],res%);
END_TEMPLATE
*/
// BEGIN_INTERLEAVE 4
// This block is automatically generated from the template above
// by the interleave.pl script. Please do not edit by hand
const __m256d ari0 = _mm256_loadu_pd(aa[0]);
const __m256d ari1 = _mm256_loadu_pd(aa[1]);
const __m256d ari2 = _mm256_loadu_pd(aa[2]);
const __m256d ari3 = _mm256_loadu_pd(aa[3]);
const __m256d bri0 = _mm256_loadu_pd(bb[0]);
const __m256d bri1 = _mm256_loadu_pd(bb[1]);
const __m256d bri2 = _mm256_loadu_pd(bb[2]);
const __m256d bri3 = _mm256_loadu_pd(bb[3]);
const __m256d sum0 = _mm256_add_pd(ari0, bri0);
const __m256d sum1 = _mm256_add_pd(ari1, bri1);
const __m256d sum2 = _mm256_add_pd(ari2, bri2);
const __m256d sum3 = _mm256_add_pd(ari3, bri3);
const __m256d rri0 = _mm256_loadu_pd(rr[0]);
const __m256d rri1 = _mm256_loadu_pd(rr[1]);
const __m256d rri2 = _mm256_loadu_pd(rr[2]);
const __m256d rri3 = _mm256_loadu_pd(rr[3]);
const __m256d res0 = _mm256_sub_pd(rri0, sum0);
const __m256d res1 = _mm256_sub_pd(rri1, sum1);
const __m256d res2 = _mm256_sub_pd(rri2, sum2);
const __m256d res3 = _mm256_sub_pd(rri3, sum3);
_mm256_storeu_pd(rr[0], res0);
_mm256_storeu_pd(rr[1], res1);
_mm256_storeu_pd(rr[2], res2);
_mm256_storeu_pd(rr[3], res3);
// END_INTERLEAVE
rr += 4;
aa += 4;
bb += 4;
} while (aa < aend);
}
EXPORT void cplx_fftvec_add_fma(uint32_t m, void* r, const void* a, const void* b) {
const double(*aa)[4] = (double(*)[4])a;
const double(*bb)[4] = (double(*)[4])b;
double(*rr)[4] = (double(*)[4])r;
const double(*const aend)[4] = aa + (m >> 1);
do {
/*
BEGIN_TEMPLATE
const __m256d ari% = _mm256_loadu_pd(aa[%]);
const __m256d bri% = _mm256_loadu_pd(bb[%]);
const __m256d res% = _mm256_add_pd(ari%,bri%);
_mm256_storeu_pd(rr[%],res%);
rr += @; // ONCE
aa += @; // ONCE
bb += @; // ONCE
END_TEMPLATE
*/
// BEGIN_INTERLEAVE 4
// This block is automatically generated from the template above
// by the interleave.pl script. Please do not edit by hand
const __m256d ari0 = _mm256_loadu_pd(aa[0]);
const __m256d ari1 = _mm256_loadu_pd(aa[1]);
const __m256d ari2 = _mm256_loadu_pd(aa[2]);
const __m256d ari3 = _mm256_loadu_pd(aa[3]);
const __m256d bri0 = _mm256_loadu_pd(bb[0]);
const __m256d bri1 = _mm256_loadu_pd(bb[1]);
const __m256d bri2 = _mm256_loadu_pd(bb[2]);
const __m256d bri3 = _mm256_loadu_pd(bb[3]);
const __m256d res0 = _mm256_add_pd(ari0, bri0);
const __m256d res1 = _mm256_add_pd(ari1, bri1);
const __m256d res2 = _mm256_add_pd(ari2, bri2);
const __m256d res3 = _mm256_add_pd(ari3, bri3);
_mm256_storeu_pd(rr[0], res0);
_mm256_storeu_pd(rr[1], res1);
_mm256_storeu_pd(rr[2], res2);
_mm256_storeu_pd(rr[3], res3);
rr += 4; // ONCE
aa += 4; // ONCE
bb += 4; // ONCE
// END_INTERLEAVE
} while (aa < aend);
}
EXPORT void cplx_fftvec_twiddle_fma(const CPLX_FFTVEC_TWIDDLE_PRECOMP* precomp, void* a, void* b, const void* omg) {
const uint32_t m = precomp->m;
double(*aa)[4] = (double(*)[4])a;
double(*bb)[4] = (double(*)[4])b;
const double(*const aend)[4] = aa + (m >> 1);
const __m256d om = _mm256_loadu_pd(omg);
const __m256d omrr = _mm256_shuffle_pd(om, om, 0);
const __m256d omii = _mm256_shuffle_pd(om, om, 15);
do {
/*
BEGIN_TEMPLATE
const __m256d bri% = _mm256_loadu_pd(bb[%]);
const __m256d bir% = _mm256_shuffle_pd(bri%,bri%,5);
__m256d p% = _mm256_mul_pd(bir%,omii);
p% = _mm256_fmaddsub_pd(bri%,omrr,p%);
const __m256d ari% = _mm256_loadu_pd(aa[%]);
_mm256_storeu_pd(aa[%],_mm256_add_pd(ari%,p%));
_mm256_storeu_pd(bb[%],_mm256_sub_pd(ari%,p%));
bb += @; // ONCE
aa += @; // ONCE
END_TEMPLATE
*/
// BEGIN_INTERLEAVE 4
// This block is automatically generated from the template above
// by the interleave.pl script. Please do not edit by hand
const __m256d bri0 = _mm256_loadu_pd(bb[0]);
const __m256d bri1 = _mm256_loadu_pd(bb[1]);
const __m256d bri2 = _mm256_loadu_pd(bb[2]);
const __m256d bri3 = _mm256_loadu_pd(bb[3]);
const __m256d bir0 = _mm256_shuffle_pd(bri0, bri0, 5);
const __m256d bir1 = _mm256_shuffle_pd(bri1, bri1, 5);
const __m256d bir2 = _mm256_shuffle_pd(bri2, bri2, 5);
const __m256d bir3 = _mm256_shuffle_pd(bri3, bri3, 5);
__m256d p0 = _mm256_mul_pd(bir0, omii);
__m256d p1 = _mm256_mul_pd(bir1, omii);
__m256d p2 = _mm256_mul_pd(bir2, omii);
__m256d p3 = _mm256_mul_pd(bir3, omii);
p0 = _mm256_fmaddsub_pd(bri0, omrr, p0);
p1 = _mm256_fmaddsub_pd(bri1, omrr, p1);
p2 = _mm256_fmaddsub_pd(bri2, omrr, p2);
p3 = _mm256_fmaddsub_pd(bri3, omrr, p3);
const __m256d ari0 = _mm256_loadu_pd(aa[0]);
const __m256d ari1 = _mm256_loadu_pd(aa[1]);
const __m256d ari2 = _mm256_loadu_pd(aa[2]);
const __m256d ari3 = _mm256_loadu_pd(aa[3]);
_mm256_storeu_pd(aa[0], _mm256_add_pd(ari0, p0));
_mm256_storeu_pd(aa[1], _mm256_add_pd(ari1, p1));
_mm256_storeu_pd(aa[2], _mm256_add_pd(ari2, p2));
_mm256_storeu_pd(aa[3], _mm256_add_pd(ari3, p3));
_mm256_storeu_pd(bb[0], _mm256_sub_pd(ari0, p0));
_mm256_storeu_pd(bb[1], _mm256_sub_pd(ari1, p1));
_mm256_storeu_pd(bb[2], _mm256_sub_pd(ari2, p2));
_mm256_storeu_pd(bb[3], _mm256_sub_pd(ari3, p3));
bb += 4; // ONCE
aa += 4; // ONCE
// END_INTERLEAVE
} while (aa < aend);
}
EXPORT void cplx_fftvec_copy_fma(uint32_t m, void* r, const void* a) {
const double(*aa)[4] = (double(*)[4])a;
double(*rr)[4] = (double(*)[4])r;
const double(*const aend)[4] = aa + (m >> 1);
do {
/*
BEGIN_TEMPLATE
const __m256d ari% = _mm256_loadu_pd(aa[%]);
_mm256_storeu_pd(rr[%],ari%);
rr += @; // ONCE
aa += @; // ONCE
END_TEMPLATE
*/
// BEGIN_INTERLEAVE 4
// This block is automatically generated from the template above
// by the interleave.pl script. Please do not edit by hand
const __m256d ari0 = _mm256_loadu_pd(aa[0]);
const __m256d ari1 = _mm256_loadu_pd(aa[1]);
const __m256d ari2 = _mm256_loadu_pd(aa[2]);
const __m256d ari3 = _mm256_loadu_pd(aa[3]);
_mm256_storeu_pd(rr[0], ari0);
_mm256_storeu_pd(rr[1], ari1);
_mm256_storeu_pd(rr[2], ari2);
_mm256_storeu_pd(rr[3], ari3);
rr += 4; // ONCE
aa += 4; // ONCE
// END_INTERLEAVE
} while (aa < aend);
}

85
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/cplx/cplx_fftvec_ref.c
View File

@@ -1,85 +0,0 @@
#include <string.h>
#include "../commons_private.h"
#include "cplx_fft_internal.h"
#include "cplx_fft_private.h"
EXPORT void cplx_fftvec_addmul_ref(const CPLX_FFTVEC_ADDMUL_PRECOMP* precomp, void* r, const void* a, const void* b) {
const uint32_t m = precomp->m;
const CPLX* aa = (CPLX*)a;
const CPLX* bb = (CPLX*)b;
CPLX* rr = (CPLX*)r;
for (uint32_t i = 0; i < m; ++i) {
const double re = aa[i][0] * bb[i][0] - aa[i][1] * bb[i][1];
const double im = aa[i][0] * bb[i][1] + aa[i][1] * bb[i][0];
rr[i][0] += re;
rr[i][1] += im;
}
}
EXPORT void cplx_fftvec_mul_ref(const CPLX_FFTVEC_MUL_PRECOMP* precomp, void* r, const void* a, const void* b) {
const uint32_t m = precomp->m;
const CPLX* aa = (CPLX*)a;
const CPLX* bb = (CPLX*)b;
CPLX* rr = (CPLX*)r;
for (uint32_t i = 0; i < m; ++i) {
const double re = aa[i][0] * bb[i][0] - aa[i][1] * bb[i][1];
const double im = aa[i][0] * bb[i][1] + aa[i][1] * bb[i][0];
rr[i][0] = re;
rr[i][1] = im;
}
}
EXPORT void* init_cplx_fftvec_addmul_precomp(CPLX_FFTVEC_ADDMUL_PRECOMP* r, uint32_t m) {
if (m & (m - 1)) return spqlios_error("m must be a power of two");
r->m = m;
if (m <= 4) {
r->function = cplx_fftvec_addmul_ref;
} else if (CPU_SUPPORTS("fma")) {
r->function = cplx_fftvec_addmul_fma;
} else {
r->function = cplx_fftvec_addmul_ref;
}
return r;
}
EXPORT void* init_cplx_fftvec_mul_precomp(CPLX_FFTVEC_MUL_PRECOMP* r, uint32_t m) {
if (m & (m - 1)) return spqlios_error("m must be a power of two");
r->m = m;
if (m <= 4) {
r->function = cplx_fftvec_mul_ref;
} else if (CPU_SUPPORTS("fma")) {
r->function = cplx_fftvec_mul_fma;
} else {
r->function = cplx_fftvec_mul_ref;
}
return r;
}
EXPORT CPLX_FFTVEC_ADDMUL_PRECOMP* new_cplx_fftvec_addmul_precomp(uint32_t m) {
CPLX_FFTVEC_ADDMUL_PRECOMP* r = malloc(sizeof(CPLX_FFTVEC_MUL_PRECOMP));
return spqlios_keep_or_free(r, init_cplx_fftvec_addmul_precomp(r, m));
}
EXPORT CPLX_FFTVEC_MUL_PRECOMP* new_cplx_fftvec_mul_precomp(uint32_t m) {
CPLX_FFTVEC_MUL_PRECOMP* r = malloc(sizeof(CPLX_FFTVEC_MUL_PRECOMP));
return spqlios_keep_or_free(r, init_cplx_fftvec_mul_precomp(r, m));
}
EXPORT void cplx_fftvec_mul_simple(uint32_t m, void* r, const void* a, const void* b) {
static CPLX_FFTVEC_MUL_PRECOMP p[31] = {0};
CPLX_FFTVEC_MUL_PRECOMP* f = p + log2m(m);
if (!f->function) {
if (!init_cplx_fftvec_mul_precomp(f, m)) abort();
}
f->function(f, r, a, b);
}
EXPORT void cplx_fftvec_addmul_simple(uint32_t m, void* r, const void* a, const void* b) {
static CPLX_FFTVEC_ADDMUL_PRECOMP p[31] = {0};
CPLX_FFTVEC_ADDMUL_PRECOMP* f = p + log2m(m);
if (!f->function) {
if (!init_cplx_fftvec_addmul_precomp(f, m)) abort();
}
f->function(f, r, a, b);
}

157
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/cplx/cplx_ifft16_avx_fma.s
View File

@@ -1,157 +0,0 @@
# shifted FFT over X^16-i
# 1st argument (rdi) contains 16 complexes
# 2nd argument (rsi) contains: 8 complexes
# omega,alpha,beta,j.beta,gamma,j.gamma,k.gamma,kj.gamma
# alpha = sqrt(omega), beta = sqrt(alpha), gamma = sqrt(beta)
# j = sqrt(i), k=sqrt(j)
.globl cplx_ifft16_avx_fma
cplx_ifft16_avx_fma:
vmovupd (%rdi),%ymm8 # load data into registers %ymm8 -> %ymm15
vmovupd 0x20(%rdi),%ymm9
vmovupd 0x40(%rdi),%ymm10
vmovupd 0x60(%rdi),%ymm11
vmovupd 0x80(%rdi),%ymm12
vmovupd 0xa0(%rdi),%ymm13
vmovupd 0xc0(%rdi),%ymm14
vmovupd 0xe0(%rdi),%ymm15
.fourth_pass:
vmovupd 0(%rsi),%ymm0 /* gamma */
vmovupd 32(%rsi),%ymm2 /* delta */
vshufpd $15, %ymm0, %ymm0, %ymm1 /* ymm1: gama.iiii */
vshufpd $15, %ymm2, %ymm2, %ymm3 /* ymm3: delta.iiii */
vshufpd $0, %ymm0, %ymm0, %ymm0 /* ymm0: gama.rrrr */
vshufpd $0, %ymm2, %ymm2, %ymm2 /* ymm2: delta.rrrr */
vperm2f128 $0x31,%ymm10,%ymm8,%ymm4 # ymm4 contains c1,c5
vperm2f128 $0x31,%ymm11,%ymm9,%ymm5 # ymm5 contains c3,c7
vperm2f128 $0x31,%ymm14,%ymm12,%ymm6 # ymm6 contains c9,c13
vperm2f128 $0x31,%ymm15,%ymm13,%ymm7 # ymm7 contains c11,c15
vperm2f128 $0x20,%ymm10,%ymm8,%ymm8 # ymm8 contains c0,c4
vperm2f128 $0x20,%ymm11,%ymm9,%ymm9 # ymm9 contains c2,c6
vperm2f128 $0x20,%ymm14,%ymm12,%ymm10 # ymm10 contains c8,c12
vperm2f128 $0x20,%ymm15,%ymm13,%ymm11 # ymm11 contains c10,c14
vsubpd %ymm4,%ymm8,%ymm12 # tw: to mul by gamma
vsubpd %ymm5,%ymm9,%ymm13 # itw: to mul by i.gamma
vsubpd %ymm6,%ymm10,%ymm14 # tw: to mul by delta
vsubpd %ymm7,%ymm11,%ymm15 # itw: to mul by i.delta
vaddpd %ymm4,%ymm8,%ymm8
vaddpd %ymm5,%ymm9,%ymm9
vaddpd %ymm6,%ymm10,%ymm10
vaddpd %ymm7,%ymm11,%ymm11
vshufpd $5, %ymm12, %ymm12, %ymm4
vshufpd $5, %ymm13, %ymm13, %ymm5
vshufpd $5, %ymm14, %ymm14, %ymm6
vshufpd $5, %ymm15, %ymm15, %ymm7
vmulpd %ymm4,%ymm1,%ymm4
vmulpd %ymm5,%ymm0,%ymm5
vmulpd %ymm6,%ymm3,%ymm6
vmulpd %ymm7,%ymm2,%ymm7
vfmaddsub231pd %ymm12, %ymm0, %ymm4 # ymm4 = (ymm0 * ymm12) +/- ymm4
vfmsubadd231pd %ymm13, %ymm1, %ymm5
vfmaddsub231pd %ymm14, %ymm2, %ymm6
vfmsubadd231pd %ymm15, %ymm3, %ymm7
vperm2f128 $0x20,%ymm6,%ymm10,%ymm12 # ymm4 contains c1,c5 -- x gamma
vperm2f128 $0x20,%ymm7,%ymm11,%ymm13 # ymm5 contains c3,c7 -- x igamma
vperm2f128 $0x31,%ymm6,%ymm10,%ymm14 # ymm6 contains c9,c13 -- x delta
vperm2f128 $0x31,%ymm7,%ymm11,%ymm15 # ymm7 contains c11,c15 -- x idelta
vperm2f128 $0x31,%ymm4,%ymm8,%ymm10 # ymm10 contains c8,c12
vperm2f128 $0x31,%ymm5,%ymm9,%ymm11 # ymm11 contains c10,c14
vperm2f128 $0x20,%ymm4,%ymm8,%ymm8 # ymm8 contains c0,c4
vperm2f128 $0x20,%ymm5,%ymm9,%ymm9 # ymm9 contains c2,c6
.third_pass:
vmovupd 64(%rsi),%xmm0 /* gamma */
vmovupd 80(%rsi),%xmm2 /* delta */
vinsertf128 $1, %xmm0, %ymm0, %ymm0
vinsertf128 $1, %xmm2, %ymm2, %ymm2
vshufpd $15, %ymm0, %ymm0, %ymm1 /* ymm1: gama.iiii */
vshufpd $15, %ymm2, %ymm2, %ymm3 /* ymm3: delta.iiii */
vshufpd $0, %ymm0, %ymm0, %ymm0 /* ymm0: gama.rrrr */
vshufpd $0, %ymm2, %ymm2, %ymm2 /* ymm2: delta.rrrr */
vsubpd %ymm9,%ymm8,%ymm4
vsubpd %ymm11,%ymm10,%ymm5
vsubpd %ymm13,%ymm12,%ymm6
vsubpd %ymm15,%ymm14,%ymm7
vaddpd %ymm9,%ymm8,%ymm8
vaddpd %ymm11,%ymm10,%ymm10
vaddpd %ymm13,%ymm12,%ymm12
vaddpd %ymm15,%ymm14,%ymm14
vshufpd $5, %ymm4, %ymm4, %ymm9
vshufpd $5, %ymm5, %ymm5, %ymm11
vshufpd $5, %ymm6, %ymm6, %ymm13
vshufpd $5, %ymm7, %ymm7, %ymm15
vmulpd %ymm9,%ymm1,%ymm9
vmulpd %ymm11,%ymm0,%ymm11
vmulpd %ymm13,%ymm3,%ymm13
vmulpd %ymm15,%ymm2,%ymm15
vfmaddsub231pd %ymm4, %ymm0, %ymm9 # ymm9 = (ymm0 * ymm4) +/- ymm9
vfmsubadd231pd %ymm5, %ymm1, %ymm11
vfmaddsub231pd %ymm6, %ymm2, %ymm13
vfmsubadd231pd %ymm7, %ymm3, %ymm15
.second_pass:
vmovupd 96(%rsi),%xmm0 /* omri */
vinsertf128 $1, %xmm0, %ymm0, %ymm0 /* omriri */
vshufpd $15, %ymm0, %ymm0, %ymm1 /* ymm1: omiiii */
vshufpd $0, %ymm0, %ymm0, %ymm0 /* ymm0: omrrrr */
vsubpd %ymm10,%ymm8,%ymm4
vsubpd %ymm11,%ymm9,%ymm5
vsubpd %ymm14,%ymm12,%ymm6
vsubpd %ymm15,%ymm13,%ymm7
vaddpd %ymm10,%ymm8,%ymm8
vaddpd %ymm11,%ymm9,%ymm9
vaddpd %ymm14,%ymm12,%ymm12
vaddpd %ymm15,%ymm13,%ymm13
vshufpd $5, %ymm4, %ymm4, %ymm10
vshufpd $5, %ymm5, %ymm5, %ymm11
vshufpd $5, %ymm6, %ymm6, %ymm14
vshufpd $5, %ymm7, %ymm7, %ymm15
vmulpd %ymm10,%ymm1,%ymm10
vmulpd %ymm11,%ymm1,%ymm11
vmulpd %ymm14,%ymm0,%ymm14
vmulpd %ymm15,%ymm0,%ymm15
vfmaddsub231pd %ymm4, %ymm0, %ymm10 # ymm10 = (ymm0 * ymm4) +/- ymm10
vfmaddsub231pd %ymm5, %ymm0, %ymm11
vfmsubadd231pd %ymm6, %ymm1, %ymm14
vfmsubadd231pd %ymm7, %ymm1, %ymm15
.first_pass:
vmovupd 112(%rsi),%xmm0 /* omri */
vinsertf128 $1, %xmm0, %ymm0, %ymm0 /* omriri */
vshufpd $15, %ymm0, %ymm0, %ymm1 /* ymm1: omiiii */
vshufpd $0, %ymm0, %ymm0, %ymm0 /* ymm0: omrrrr */
vsubpd %ymm12,%ymm8,%ymm4
vsubpd %ymm13,%ymm9,%ymm5
vsubpd %ymm14,%ymm10,%ymm6
vsubpd %ymm15,%ymm11,%ymm7
vaddpd %ymm12,%ymm8,%ymm8
vaddpd %ymm13,%ymm9,%ymm9
vaddpd %ymm14,%ymm10,%ymm10
vaddpd %ymm15,%ymm11,%ymm11
vshufpd $5, %ymm4, %ymm4, %ymm12
vshufpd $5, %ymm5, %ymm5, %ymm13
vshufpd $5, %ymm6, %ymm6, %ymm14
vshufpd $5, %ymm7, %ymm7, %ymm15
vmulpd %ymm12,%ymm1,%ymm12
vmulpd %ymm13,%ymm1,%ymm13
vmulpd %ymm14,%ymm1,%ymm14
vmulpd %ymm15,%ymm1,%ymm15
vfmaddsub231pd %ymm4, %ymm0, %ymm12 # ymm12 = (ymm0 * ymm4) +/- ymm12
vfmaddsub231pd %ymm5, %ymm0, %ymm13
vfmaddsub231pd %ymm6, %ymm0, %ymm14
vfmaddsub231pd %ymm7, %ymm0, %ymm15
.save_and_return:
vmovupd %ymm8,(%rdi)
vmovupd %ymm9,0x20(%rdi)
vmovupd %ymm10,0x40(%rdi)
vmovupd %ymm11,0x60(%rdi)
vmovupd %ymm12,0x80(%rdi)
vmovupd %ymm13,0xa0(%rdi)
vmovupd %ymm14,0xc0(%rdi)
vmovupd %ymm15,0xe0(%rdi)
ret
.size cplx_ifft16_avx_fma, .-cplx_ifft16_avx_fma
.section .note.GNU-stack,"",@progbits

192
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/cplx/cplx_ifft16_avx_fma_win32.s
View File

@@ -1,192 +0,0 @@
.text
.p2align 4
.globl cplx_ifft16_avx_fma
.def cplx_ifft16_avx_fma; .scl 2; .type 32; .endef
cplx_ifft16_avx_fma:
pushq %rdi
pushq %rsi
movq %rcx,%rdi
movq %rdx,%rsi
subq $0x100,%rsp
movdqu %xmm6,(%rsp)
movdqu %xmm7,0x10(%rsp)
movdqu %xmm8,0x20(%rsp)
movdqu %xmm9,0x30(%rsp)
movdqu %xmm10,0x40(%rsp)
movdqu %xmm11,0x50(%rsp)
movdqu %xmm12,0x60(%rsp)
movdqu %xmm13,0x70(%rsp)
movdqu %xmm14,0x80(%rsp)
movdqu %xmm15,0x90(%rsp)
callq cplx_ifft16_avx_fma_amd64
movdqu (%rsp),%xmm6
movdqu 0x10(%rsp),%xmm7
movdqu 0x20(%rsp),%xmm8
movdqu 0x30(%rsp),%xmm9
movdqu 0x40(%rsp),%xmm10
movdqu 0x50(%rsp),%xmm11
movdqu 0x60(%rsp),%xmm12
movdqu 0x70(%rsp),%xmm13
movdqu 0x80(%rsp),%xmm14
movdqu 0x90(%rsp),%xmm15
addq $0x100,%rsp
popq %rsi
popq %rdi
retq
# shifted FFT over X^16-i
# 1st argument (rdi) contains 16 complexes
# 2nd argument (rsi) contains: 8 complexes
# omega,alpha,beta,j.beta,gamma,j.gamma,k.gamma,kj.gamma
# alpha = sqrt(omega), beta = sqrt(alpha), gamma = sqrt(beta)
# j = sqrt(i), k=sqrt(j)
cplx_ifft16_avx_fma_amd64:
vmovupd (%rdi),%ymm8 # load data into registers %ymm8 -> %ymm15
vmovupd 0x20(%rdi),%ymm9
vmovupd 0x40(%rdi),%ymm10
vmovupd 0x60(%rdi),%ymm11
vmovupd 0x80(%rdi),%ymm12
vmovupd 0xa0(%rdi),%ymm13
vmovupd 0xc0(%rdi),%ymm14
vmovupd 0xe0(%rdi),%ymm15
.fourth_pass:
vmovupd 0(%rsi),%ymm0 /* gamma */
vmovupd 32(%rsi),%ymm2 /* delta */
vshufpd $15, %ymm0, %ymm0, %ymm1 /* ymm1: gama.iiii */
vshufpd $15, %ymm2, %ymm2, %ymm3 /* ymm3: delta.iiii */
vshufpd $0, %ymm0, %ymm0, %ymm0 /* ymm0: gama.rrrr */
vshufpd $0, %ymm2, %ymm2, %ymm2 /* ymm2: delta.rrrr */
vperm2f128 $0x31,%ymm10,%ymm8,%ymm4 # ymm4 contains c1,c5
vperm2f128 $0x31,%ymm11,%ymm9,%ymm5 # ymm5 contains c3,c7
vperm2f128 $0x31,%ymm14,%ymm12,%ymm6 # ymm6 contains c9,c13
vperm2f128 $0x31,%ymm15,%ymm13,%ymm7 # ymm7 contains c11,c15
vperm2f128 $0x20,%ymm10,%ymm8,%ymm8 # ymm8 contains c0,c4
vperm2f128 $0x20,%ymm11,%ymm9,%ymm9 # ymm9 contains c2,c6
vperm2f128 $0x20,%ymm14,%ymm12,%ymm10 # ymm10 contains c8,c12
vperm2f128 $0x20,%ymm15,%ymm13,%ymm11 # ymm11 contains c10,c14
vsubpd %ymm4,%ymm8,%ymm12 # tw: to mul by gamma
vsubpd %ymm5,%ymm9,%ymm13 # itw: to mul by i.gamma
vsubpd %ymm6,%ymm10,%ymm14 # tw: to mul by delta
vsubpd %ymm7,%ymm11,%ymm15 # itw: to mul by i.delta
vaddpd %ymm4,%ymm8,%ymm8
vaddpd %ymm5,%ymm9,%ymm9
vaddpd %ymm6,%ymm10,%ymm10
vaddpd %ymm7,%ymm11,%ymm11
vshufpd $5, %ymm12, %ymm12, %ymm4
vshufpd $5, %ymm13, %ymm13, %ymm5
vshufpd $5, %ymm14, %ymm14, %ymm6
vshufpd $5, %ymm15, %ymm15, %ymm7
vmulpd %ymm4,%ymm1,%ymm4
vmulpd %ymm5,%ymm0,%ymm5
vmulpd %ymm6,%ymm3,%ymm6
vmulpd %ymm7,%ymm2,%ymm7
vfmaddsub231pd %ymm12, %ymm0, %ymm4 # ymm4 = (ymm0 * ymm12) +/- ymm4
vfmsubadd231pd %ymm13, %ymm1, %ymm5
vfmaddsub231pd %ymm14, %ymm2, %ymm6
vfmsubadd231pd %ymm15, %ymm3, %ymm7
vperm2f128 $0x20,%ymm6,%ymm10,%ymm12 # ymm4 contains c1,c5 -- x gamma
vperm2f128 $0x20,%ymm7,%ymm11,%ymm13 # ymm5 contains c3,c7 -- x igamma
vperm2f128 $0x31,%ymm6,%ymm10,%ymm14 # ymm6 contains c9,c13 -- x delta
vperm2f128 $0x31,%ymm7,%ymm11,%ymm15 # ymm7 contains c11,c15 -- x idelta
vperm2f128 $0x31,%ymm4,%ymm8,%ymm10 # ymm10 contains c8,c12
vperm2f128 $0x31,%ymm5,%ymm9,%ymm11 # ymm11 contains c10,c14
vperm2f128 $0x20,%ymm4,%ymm8,%ymm8 # ymm8 contains c0,c4
vperm2f128 $0x20,%ymm5,%ymm9,%ymm9 # ymm9 contains c2,c6
.third_pass:
vmovupd 64(%rsi),%xmm0 /* gamma */
vmovupd 80(%rsi),%xmm2 /* delta */
vinsertf128 $1, %xmm0, %ymm0, %ymm0
vinsertf128 $1, %xmm2, %ymm2, %ymm2
vshufpd $15, %ymm0, %ymm0, %ymm1 /* ymm1: gama.iiii */
vshufpd $15, %ymm2, %ymm2, %ymm3 /* ymm3: delta.iiii */
vshufpd $0, %ymm0, %ymm0, %ymm0 /* ymm0: gama.rrrr */
vshufpd $0, %ymm2, %ymm2, %ymm2 /* ymm2: delta.rrrr */
vsubpd %ymm9,%ymm8,%ymm4
vsubpd %ymm11,%ymm10,%ymm5
vsubpd %ymm13,%ymm12,%ymm6
vsubpd %ymm15,%ymm14,%ymm7
vaddpd %ymm9,%ymm8,%ymm8
vaddpd %ymm11,%ymm10,%ymm10
vaddpd %ymm13,%ymm12,%ymm12
vaddpd %ymm15,%ymm14,%ymm14
vshufpd $5, %ymm4, %ymm4, %ymm9
vshufpd $5, %ymm5, %ymm5, %ymm11
vshufpd $5, %ymm6, %ymm6, %ymm13
vshufpd $5, %ymm7, %ymm7, %ymm15
vmulpd %ymm9,%ymm1,%ymm9
vmulpd %ymm11,%ymm0,%ymm11
vmulpd %ymm13,%ymm3,%ymm13
vmulpd %ymm15,%ymm2,%ymm15
vfmaddsub231pd %ymm4, %ymm0, %ymm9 # ymm9 = (ymm0 * ymm4) +/- ymm9
vfmsubadd231pd %ymm5, %ymm1, %ymm11
vfmaddsub231pd %ymm6, %ymm2, %ymm13
vfmsubadd231pd %ymm7, %ymm3, %ymm15
.second_pass:
vmovupd 96(%rsi),%xmm0 /* omri */
vinsertf128 $1, %xmm0, %ymm0, %ymm0 /* omriri */
vshufpd $15, %ymm0, %ymm0, %ymm1 /* ymm1: omiiii */
vshufpd $0, %ymm0, %ymm0, %ymm0 /* ymm0: omrrrr */
vsubpd %ymm10,%ymm8,%ymm4
vsubpd %ymm11,%ymm9,%ymm5
vsubpd %ymm14,%ymm12,%ymm6
vsubpd %ymm15,%ymm13,%ymm7
vaddpd %ymm10,%ymm8,%ymm8
vaddpd %ymm11,%ymm9,%ymm9
vaddpd %ymm14,%ymm12,%ymm12
vaddpd %ymm15,%ymm13,%ymm13
vshufpd $5, %ymm4, %ymm4, %ymm10
vshufpd $5, %ymm5, %ymm5, %ymm11
vshufpd $5, %ymm6, %ymm6, %ymm14
vshufpd $5, %ymm7, %ymm7, %ymm15
vmulpd %ymm10,%ymm1,%ymm10
vmulpd %ymm11,%ymm1,%ymm11
vmulpd %ymm14,%ymm0,%ymm14
vmulpd %ymm15,%ymm0,%ymm15
vfmaddsub231pd %ymm4, %ymm0, %ymm10 # ymm10 = (ymm0 * ymm4) +/- ymm10
vfmaddsub231pd %ymm5, %ymm0, %ymm11
vfmsubadd231pd %ymm6, %ymm1, %ymm14
vfmsubadd231pd %ymm7, %ymm1, %ymm15
.first_pass:
vmovupd 112(%rsi),%xmm0 /* omri */
vinsertf128 $1, %xmm0, %ymm0, %ymm0 /* omriri */
vshufpd $15, %ymm0, %ymm0, %ymm1 /* ymm1: omiiii */
vshufpd $0, %ymm0, %ymm0, %ymm0 /* ymm0: omrrrr */
vsubpd %ymm12,%ymm8,%ymm4
vsubpd %ymm13,%ymm9,%ymm5
vsubpd %ymm14,%ymm10,%ymm6
vsubpd %ymm15,%ymm11,%ymm7
vaddpd %ymm12,%ymm8,%ymm8
vaddpd %ymm13,%ymm9,%ymm9
vaddpd %ymm14,%ymm10,%ymm10
vaddpd %ymm15,%ymm11,%ymm11
vshufpd $5, %ymm4, %ymm4, %ymm12
vshufpd $5, %ymm5, %ymm5, %ymm13
vshufpd $5, %ymm6, %ymm6, %ymm14
vshufpd $5, %ymm7, %ymm7, %ymm15
vmulpd %ymm12,%ymm1,%ymm12
vmulpd %ymm13,%ymm1,%ymm13
vmulpd %ymm14,%ymm1,%ymm14
vmulpd %ymm15,%ymm1,%ymm15
vfmaddsub231pd %ymm4, %ymm0, %ymm12 # ymm12 = (ymm0 * ymm4) +/- ymm12
vfmaddsub231pd %ymm5, %ymm0, %ymm13
vfmaddsub231pd %ymm6, %ymm0, %ymm14
vfmaddsub231pd %ymm7, %ymm0, %ymm15
.save_and_return:
vmovupd %ymm8,(%rdi)
vmovupd %ymm9,0x20(%rdi)
vmovupd %ymm10,0x40(%rdi)
vmovupd %ymm11,0x60(%rdi)
vmovupd %ymm12,0x80(%rdi)
vmovupd %ymm13,0xa0(%rdi)
vmovupd %ymm14,0xc0(%rdi)
vmovupd %ymm15,0xe0(%rdi)
ret

267
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/cplx/cplx_ifft_avx2_fma.c
View File

@@ -1,267 +0,0 @@
#include <immintrin.h>
#include <math.h>
#include <string.h>
#include "cplx_fft_internal.h"
#include "cplx_fft_private.h"
typedef double D4MEM[4];
typedef double D2MEM[2];
/**
* @brief complex ifft via bfs strategy (for m between 2 and 8)
* @param dat the data to run the algorithm on
* @param omg precomputed tables (must have been filled with fill_omega)
* @param m ring dimension of the FFT (modulo X^m-i)
*/
void cplx_ifft_avx2_fma_bfs_2(D4MEM* dat, const D2MEM** omga, uint32_t m) {
double* data = (double*)dat;
D4MEM* const finaldd = (D4MEM*)(data + 2 * m);
{
// loop with h = 1
// we do not do any particular optimization in this loop,
// since this function is only called for small dimensions
D4MEM* dd = (D4MEM*)data;
do {
/*
BEGIN_TEMPLATE
const __m256d ab% = _mm256_loadu_pd(dd[0+2*%]);
const __m256d cd% = _mm256_loadu_pd(dd[1+2*%]);
const __m256d ac% = _mm256_permute2f128_pd(ab%, cd%, 0b100000);
const __m256d bd% = _mm256_permute2f128_pd(ab%, cd%, 0b110001);
const __m256d sum% = _mm256_add_pd(ac%, bd%);
const __m256d diff% = _mm256_sub_pd(ac%, bd%);
const __m256d diffbar% = _mm256_shuffle_pd(diff%, diff%, 5);
const __m256d om% = _mm256_load_pd((*omg)[0+%]);
const __m256d omre% = _mm256_unpacklo_pd(om%, om%);
const __m256d omim% = _mm256_unpackhi_pd(om%, om%);
const __m256d t1% = _mm256_mul_pd(diffbar%, omim%);
const __m256d t2% = _mm256_fmaddsub_pd(diff%, omre%, t1%);
const __m256d newab% = _mm256_permute2f128_pd(sum%, t2%, 0b100000);
const __m256d newcd% = _mm256_permute2f128_pd(sum%, t2%, 0b110001);
_mm256_storeu_pd(dd[0+2*%], newab%);
_mm256_storeu_pd(dd[1+2*%], newcd%);
dd += 2*@;
*omg += 2*@;
END_TEMPLATE
*/
// BEGIN_INTERLEAVE 1
const __m256d ab0 = _mm256_loadu_pd(dd[0 + 2 * 0]);
const __m256d cd0 = _mm256_loadu_pd(dd[1 + 2 * 0]);
const __m256d ac0 = _mm256_permute2f128_pd(ab0, cd0, 0b100000);
const __m256d bd0 = _mm256_permute2f128_pd(ab0, cd0, 0b110001);
const __m256d sum0 = _mm256_add_pd(ac0, bd0);
const __m256d diff0 = _mm256_sub_pd(ac0, bd0);
const __m256d diffbar0 = _mm256_shuffle_pd(diff0, diff0, 5);
const __m256d om0 = _mm256_load_pd((*omga)[0 + 0]);
const __m256d omre0 = _mm256_unpacklo_pd(om0, om0);
const __m256d omim0 = _mm256_unpackhi_pd(om0, om0);
const __m256d t10 = _mm256_mul_pd(diffbar0, omim0);
const __m256d t20 = _mm256_fmaddsub_pd(diff0, omre0, t10);
const __m256d newab0 = _mm256_permute2f128_pd(sum0, t20, 0b100000);
const __m256d newcd0 = _mm256_permute2f128_pd(sum0, t20, 0b110001);
_mm256_storeu_pd(dd[0 + 2 * 0], newab0);
_mm256_storeu_pd(dd[1 + 2 * 0], newcd0);
dd += 2 * 1;
*omga += 2 * 1;
// END_INTERLEAVE
} while (dd < finaldd);
#if 0
printf("c after first: ");
for (uint64_t ii=0; ii<nn/2; ++ii) {
printf("%.6lf %.6lf ",ddata[ii][0],ddata[ii][1]);
}
printf("\n");
#endif
}
// general case
const uint32_t ms2 = m >> 1;
for (uint32_t _2nblock = 2; _2nblock <= ms2; _2nblock <<= 1) {
// _2nblock = h in ref code
uint32_t nblock = _2nblock >> 1; // =h/2 in ref code
D4MEM* dd = (D4MEM*)data;
do {
const __m256d om = _mm256_load_pd((*omga)[0]);
const __m256d omre = _mm256_unpacklo_pd(om, om);
const __m256d omim = _mm256_addsub_pd(_mm256_setzero_pd(), _mm256_unpackhi_pd(om, om));
D4MEM* const ddend = (dd + nblock);
D4MEM* ddmid = ddend;
do {
const __m256d a = _mm256_loadu_pd(dd[0]);
const __m256d b = _mm256_loadu_pd(ddmid[0]);
const __m256d newa = _mm256_add_pd(a, b);
_mm256_storeu_pd(dd[0], newa);
const __m256d diff = _mm256_sub_pd(a, b);
const __m256d t1 = _mm256_mul_pd(diff, omre);
const __m256d bardiff = _mm256_shuffle_pd(diff, diff, 5);
const __m256d t2 = _mm256_fmadd_pd(bardiff, omim, t1);
_mm256_storeu_pd(ddmid[0], t2);
dd += 1;
ddmid += 1;
} while (dd < ddend);
dd += nblock;
*omga += 2;
} while (dd < finaldd);
}
}
/**
* @brief complex fft via bfs strategy (for m >= 16)
* @param dat the data to run the algorithm on
* @param omg precomputed tables (must have been filled with fill_omega)
* @param m ring dimension of the FFT (modulo X^m-i)
*/
void cplx_ifft_avx2_fma_bfs_16(D4MEM* dat, const D2MEM** omga, uint32_t m) {
double* data = (double*)dat;
D4MEM* const finaldd = (D4MEM*)(data + 2 * m);
// base iteration when h = _2nblock == 8
{
D4MEM* dd = (D4MEM*)data;
do {
cplx_ifft16_avx_fma(dd, *omga);
dd += 8;
*omga += 8;
} while (dd < finaldd);
}
// general case
const uint32_t log2m = _mm_popcnt_u32(m - 1); //_popcnt32(m-1); //log2(m);
uint32_t h = 16;
if (log2m % 2 == 1) {
uint32_t nblock = h >> 1; // =h/2 in ref code
D4MEM* dd = (D4MEM*)data;
do {
const __m128d om1 = _mm_loadu_pd((*omga)[0]);
const __m256d om = _mm256_set_m128d(om1, om1);
const __m256d omre = _mm256_unpacklo_pd(om, om);
const __m256d omim = _mm256_addsub_pd(_mm256_setzero_pd(), _mm256_unpackhi_pd(om, om));
D4MEM* const ddend = (dd + nblock);
D4MEM* ddmid = ddend;
do {
const __m256d a = _mm256_loadu_pd(dd[0]);
const __m256d b = _mm256_loadu_pd(ddmid[0]);
const __m256d newa = _mm256_add_pd(a, b);
_mm256_storeu_pd(dd[0], newa);
const __m256d diff = _mm256_sub_pd(a, b);
const __m256d t1 = _mm256_mul_pd(diff, omre);
const __m256d bardiff = _mm256_shuffle_pd(diff, diff, 5);
const __m256d t2 = _mm256_fmadd_pd(bardiff, omim, t1);
_mm256_storeu_pd(ddmid[0], t2);
dd += 1;
ddmid += 1;
} while (dd < ddend);
dd += nblock;
*omga += 1;
} while (dd < finaldd);
h = 32;
}
for (; h < m; h <<= 2) {
// _2nblock = h in ref code
uint32_t nblock = h >> 1; // =h/2 in ref code
D4MEM* dd0 = (D4MEM*)data;
do {
const __m128d om1 = _mm_loadu_pd((*omga)[0]);
const __m128d al1 = _mm_loadu_pd((*omga)[1]);
const __m256d om = _mm256_set_m128d(om1, om1);
const __m256d al = _mm256_set_m128d(al1, al1);
const __m256d omre = _mm256_unpacklo_pd(om, om);
const __m256d omim = _mm256_unpackhi_pd(om, om);
const __m256d alre = _mm256_unpacklo_pd(al, al);
const __m256d alim = _mm256_unpackhi_pd(al, al);
D4MEM* const ddend = (dd0 + nblock);
D4MEM* dd1 = ddend;
D4MEM* dd2 = dd1 + nblock;
D4MEM* dd3 = dd2 + nblock;
do {
__m256d u0 = _mm256_loadu_pd(dd0[0]);
__m256d u1 = _mm256_loadu_pd(dd1[0]);
__m256d u2 = _mm256_loadu_pd(dd2[0]);
__m256d u3 = _mm256_loadu_pd(dd3[0]);
__m256d u4 = _mm256_add_pd(u0, u1);
__m256d u5 = _mm256_sub_pd(u0, u1);
__m256d u6 = _mm256_add_pd(u2, u3);
__m256d u7 = _mm256_sub_pd(u2, u3);
u0 = _mm256_shuffle_pd(u5, u5, 5);
u2 = _mm256_shuffle_pd(u7, u7, 5);
u1 = _mm256_mul_pd(u0, omim);
u3 = _mm256_mul_pd(u2, omre);
u5 = _mm256_fmaddsub_pd(u5, omre, u1);
u7 = _mm256_fmsubadd_pd(u7, omim, u3);
//////
u0 = _mm256_add_pd(u4, u6);
u1 = _mm256_add_pd(u5, u7);
u2 = _mm256_sub_pd(u4, u6);
u3 = _mm256_sub_pd(u5, u7);
u4 = _mm256_shuffle_pd(u2, u2, 5);
u5 = _mm256_shuffle_pd(u3, u3, 5);
u6 = _mm256_mul_pd(u4, alim);
u7 = _mm256_mul_pd(u5, alim);
u2 = _mm256_fmaddsub_pd(u2, alre, u6);
u3 = _mm256_fmaddsub_pd(u3, alre, u7);
///////
_mm256_storeu_pd(dd0[0], u0);
_mm256_storeu_pd(dd1[0], u1);
_mm256_storeu_pd(dd2[0], u2);
_mm256_storeu_pd(dd3[0], u3);
dd0 += 1;
dd1 += 1;
dd2 += 1;
dd3 += 1;
} while (dd0 < ddend);
dd0 += 3 * nblock;
*omga += 2;
} while (dd0 < finaldd);
}
}
/**
* @brief complex ifft via dfs recursion (for m >= 16)
* @param dat the data to run the algorithm on
* @param omg precomputed tables (must have been filled with fill_omega)
* @param m ring dimension of the FFT (modulo X^m-i)
*/
void cplx_ifft_avx2_fma_rec_16(D4MEM* dat, const D2MEM** omga, uint32_t m) {
if (m <= 8) return cplx_ifft_avx2_fma_bfs_2(dat, omga, m);
if (m <= 2048) return cplx_ifft_avx2_fma_bfs_16(dat, omga, m);
const uint32_t _2nblock = m >> 1; // = h in ref code
const uint32_t nblock = _2nblock >> 1; // =h/2 in ref code
cplx_ifft_avx2_fma_rec_16(dat, omga, _2nblock);
cplx_ifft_avx2_fma_rec_16(dat + nblock, omga, _2nblock);
{
// final iteration
D4MEM* dd = dat;
const __m256d om = _mm256_load_pd((*omga)[0]);
const __m256d omre = _mm256_unpacklo_pd(om, om);
const __m256d omim = _mm256_addsub_pd(_mm256_setzero_pd(), _mm256_unpackhi_pd(om, om));
D4MEM* const ddend = (dd + nblock);
D4MEM* ddmid = ddend;
do {
const __m256d a = _mm256_loadu_pd(dd[0]);
const __m256d b = _mm256_loadu_pd(ddmid[0]);
const __m256d newa = _mm256_add_pd(a, b);
_mm256_storeu_pd(dd[0], newa);
const __m256d diff = _mm256_sub_pd(a, b);
const __m256d t1 = _mm256_mul_pd(diff, omre);
const __m256d bardiff = _mm256_shuffle_pd(diff, diff, 5);
const __m256d t2 = _mm256_fmadd_pd(bardiff, omim, t1);
_mm256_storeu_pd(ddmid[0], t2);
dd += 1;
ddmid += 1;
} while (dd < ddend);
*omga += 2;
}
}
/**
* @brief complex ifft via best strategy (for m>=1)
* @param dat the data to run the algorithm on: m complex numbers
* @param omg precomputed tables (must have been filled with fill_omega)
* @param m ring dimension of the FFT (modulo X^m-i)
*/
EXPORT void cplx_ifft_avx2_fma(const CPLX_IFFT_PRECOMP* precomp, void* d) {
const uint32_t m = precomp->m;
const D2MEM* omg = (D2MEM*)precomp->powomegas;
if (m <= 1) return;
if (m <= 8) return cplx_ifft_avx2_fma_bfs_2(d, &omg, m);
if (m <= 2048) return cplx_ifft_avx2_fma_bfs_16(d, &omg, m);
cplx_ifft_avx2_fma_rec_16(d, &omg, m);
}

312
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/cplx/cplx_ifft_ref.c
View File

@@ -1,312 +0,0 @@
#include <memory.h>
#include "../commons_private.h"
#include "cplx_fft.h"
#include "cplx_fft_internal.h"
#include "cplx_fft_private.h"
/** @brief (a,b) <- (a+b,omegabar.(a-b)) */
void invctwiddle(CPLX a, CPLX b, const CPLX ombar) {
double diffre = a[0] - b[0];
double diffim = a[1] - b[1];
a[0] = a[0] + b[0];
a[1] = a[1] + b[1];
b[0] = diffre * ombar[0] - diffim * ombar[1];
b[1] = diffre * ombar[1] + diffim * ombar[0];
}
/** @brief (a,b) <- (a+b,-i.omegabar(a-b)) */
void invcitwiddle(CPLX a, CPLX b, const CPLX ombar) {
double diffre = a[0] - b[0];
double diffim = a[1] - b[1];
a[0] = a[0] + b[0];
a[1] = a[1] + b[1];
//-i(x+iy)=-ix+y
b[0] = diffre * ombar[1] + diffim * ombar[0];
b[1] = -diffre * ombar[0] + diffim * ombar[1];
}
/** @brief exp(-i.2pi.x) */
void cplx_set_e2pimx(CPLX res, double x) {
res[0] = m_accurate_cos(2 * M_PI * x);
res[1] = -m_accurate_sin(2 * M_PI * x);
}
/**
* @brief this computes the sequence: 0,1/2,1/4,3/4,1/8,5/8,3/8,7/8,...
* essentially: the bits of (i+1) in lsb order on the basis (1/2^k) mod 1*/
double fracrevbits(uint32_t i);
/** @brief fft modulo X^m-exp(i.2pi.entry+pwr) -- reference code */
void cplx_ifft_naive(const uint32_t m, const double entry_pwr, CPLX* data) {
if (m == 1) return;
const double pom = entry_pwr / 2.;
const uint32_t h = m / 2;
CPLX cpom;
cplx_set_e2pimx(cpom, pom);
// do the recursive calls
cplx_ifft_naive(h, pom, data);
cplx_ifft_naive(h, pom + 0.5, data + h);
// apply the inverse twiddle factors
for (uint64_t i = 0; i < h; ++i) {
invctwiddle(data[i], data[i + h], cpom);
}
}
void cplx_ifft16_precomp(const double entry_pwr, CPLX** omg) {
static const double j_pow = 1. / 8.;
static const double k_pow = 1. / 16.;
const double pom = entry_pwr / 2.;
const double pom_2 = entry_pwr / 4.;
const double pom_4 = entry_pwr / 8.;
const double pom_8 = entry_pwr / 16.;
cplx_set_e2pimx((*omg)[0], pom_8);
cplx_set_e2pimx((*omg)[1], pom_8 + j_pow);
cplx_set_e2pimx((*omg)[2], pom_8 + k_pow);
cplx_set_e2pimx((*omg)[3], pom_8 + j_pow + k_pow);
cplx_set_e2pimx((*omg)[4], pom_4);
cplx_set_e2pimx((*omg)[5], pom_4 + j_pow);
cplx_set_e2pimx((*omg)[6], pom_2);
cplx_set_e2pimx((*omg)[7], pom);
*omg += 8;
}
/**
* @brief iFFT modulo X^16-omega^2 (in registers)
* @param data contains 16 complexes
* @param omegabar 8 complexes in this order:
* gammabar,jb.gammabar,kb.gammabar,kbjb.gammabar,
* betabar,jb.betabar,alphabar,omegabar
* alpha = sqrt(omega), beta = sqrt(alpha), gamma = sqrt(beta)
* jb = sqrt(ib), kb=sqrt(jb)
*/
void cplx_ifft16_ref(void* data, const void* omegabar) {
CPLX* d = data;
const CPLX* om = omegabar;
// fourth pass inverse
invctwiddle(d[0], d[1], om[0]);
invcitwiddle(d[2], d[3], om[0]);
invctwiddle(d[4], d[5], om[1]);
invcitwiddle(d[6], d[7], om[1]);
invctwiddle(d[8], d[9], om[2]);
invcitwiddle(d[10], d[11], om[2]);
invctwiddle(d[12], d[13], om[3]);
invcitwiddle(d[14], d[15], om[3]);
// third pass inverse
invctwiddle(d[0], d[2], om[4]);
invctwiddle(d[1], d[3], om[4]);
invcitwiddle(d[4], d[6], om[4]);
invcitwiddle(d[5], d[7], om[4]);
invctwiddle(d[8], d[10], om[5]);
invctwiddle(d[9], d[11], om[5]);
invcitwiddle(d[12], d[14], om[5]);
invcitwiddle(d[13], d[15], om[5]);
// second pass inverse
invctwiddle(d[0], d[4], om[6]);
invctwiddle(d[1], d[5], om[6]);
invctwiddle(d[2], d[6], om[6]);
invctwiddle(d[3], d[7], om[6]);
invcitwiddle(d[8], d[12], om[6]);
invcitwiddle(d[9], d[13], om[6]);
invcitwiddle(d[10], d[14], om[6]);
invcitwiddle(d[11], d[15], om[6]);
// first pass
for (uint64_t i = 0; i < 8; ++i) {
invctwiddle(d[0 + i], d[8 + i], om[7]);
}
}
void cplx_ifft_ref_bfs_2(CPLX* dat, const CPLX** omg, uint32_t m) {
CPLX* const dend = dat + m;
for (CPLX* d = dat; d < dend; d += 2) {
split_fft_last_ref(d, (*omg)[0]);
*omg += 1;
}
#if 0
printf("after first: ");
for (uint64_t ii=0; ii<nn/2; ++ii) {
printf("%.6lf %.6lf ",data[ii][0],data[ii][1]);
}
printf("\n");
#endif
int32_t Ms2 = m / 2;
for (int32_t h = 2; h <= Ms2; h <<= 1) {
for (CPLX* d = dat; d < dend; d += 2 * h) {
if (memcmp((*omg)[0], (*omg)[1], 8) != 0) abort();
cplx_split_fft_ref(h, d, **omg);
*omg += 2;
}
#if 0
printf("after split %d: ", h);
for (uint64_t ii=0; ii<nn/2; ++ii) {
printf("%.6lf %.6lf ",data[ii][0],data[ii][1]);
}
printf("\n");
#endif
}
}
void cplx_ifft_ref_bfs_16(CPLX* dat, const CPLX** omg, uint32_t m) {
const uint64_t log2m = log2(m);
CPLX* const dend = dat + m;
// h=1,2,4,8 use the 16-dim macroblock
for (CPLX* d = dat; d < dend; d += 16) {
cplx_ifft16_ref(d, *omg);
*omg += 8;
}
#if 0
printf("after first: ");
for (uint64_t ii=0; ii<nn/2; ++ii) {
printf("%.6lf %.6lf ",data[ii][0],data[ii][1]);
}
printf("\n");
#endif
int32_t h = 16;
if (log2m % 2 != 0) {
// if parity needs it, uses one regular twiddle
for (CPLX* d = dat; d < dend; d += 2 * h) {
cplx_split_fft_ref(h, d, **omg);
*omg += 1;
}
h = 32;
}
// h=16,...,2*floor(Ms2/2) use the bitwiddle
for (; h < m; h <<= 2) {
for (CPLX* d = dat; d < dend; d += 4 * h) {
cplx_bisplit_fft_ref(h, d, *omg);
*omg += 2;
}
#if 0
printf("after split %d: ", h);
for (uint64_t ii=0; ii<nn/2; ++ii) {
printf("%.6lf %.6lf ",data[ii][0],data[ii][1]);
}
printf("\n");
#endif
}
}
void fill_cplx_ifft_omegas_bfs_16(const double entry_pwr, CPLX** omg, uint32_t m) {
const uint64_t log2m = log2(m);
double pwr = entry_pwr * 16. / m;
{
// h=8
for (uint32_t i = 0; i < m / 16; i++) {
cplx_ifft16_precomp(pwr + fracrevbits(i), omg);
}
}
int32_t h = 16;
if (log2m % 2 != 0) {
// if parity needs it, uses one regular twiddle
for (uint32_t i = 0; i < m / (2 * h); i++) {
cplx_set_e2pimx(omg[0][0], pwr + fracrevbits(i) / 2.);
*omg += 1;
}
pwr *= 2.;
h = 32;
}
for (; h < m; h <<= 2) {
for (uint32_t i = 0; i < m / (2 * h); i += 2) {
cplx_set_e2pimx(omg[0][0], pwr + fracrevbits(i) / 2.);
cplx_set_e2pimx(omg[0][1], 2. * pwr + fracrevbits(i));
*omg += 2;
}
pwr *= 4.;
}
}
void fill_cplx_ifft_omegas_bfs_2(const double entry_pwr, CPLX** omg, uint32_t m) {
double pom = entry_pwr / m;
{
// h=1
for (uint32_t i = 0; i < m / 2; i++) {
cplx_set_e2pimx((*omg)[0], pom + fracrevbits(i) / 2.);
*omg += 1; // optim function reads by 1
}
}
for (int32_t h = 2; h <= m / 2; h <<= 1) {
pom *= 2;
for (uint32_t i = 0; i < m / (2 * h); i++) {
cplx_set_e2pimx(omg[0][0], pom + fracrevbits(i) / 2.);
cplx_set(omg[0][1], omg[0][0]);
*omg += 2;
}
}
}
void fill_cplx_ifft_omegas_rec_16(const double entry_pwr, CPLX** omg, uint32_t m) {
if (m == 1) return;
if (m <= 8) return fill_cplx_ifft_omegas_bfs_2(entry_pwr, omg, m);
if (m <= 2048) return fill_cplx_ifft_omegas_bfs_16(entry_pwr, omg, m);
double pom = entry_pwr / 2.;
uint32_t h = m / 2;
fill_cplx_ifft_omegas_rec_16(pom, omg, h);
fill_cplx_ifft_omegas_rec_16(pom + 0.5, omg, h);
cplx_set_e2pimx((*omg)[0], pom);
cplx_set((*omg)[1], (*omg)[0]);
*omg += 2;
}
void cplx_ifft_ref_rec_16(CPLX* dat, const CPLX** omg, uint32_t m) {
if (m == 1) return;
if (m <= 8) return cplx_ifft_ref_bfs_2(dat, omg, m);
if (m <= 2048) return cplx_ifft_ref_bfs_16(dat, omg, m);
const uint32_t h = m / 2;
cplx_ifft_ref_rec_16(dat, omg, h);
cplx_ifft_ref_rec_16(dat + h, omg, h);
if (memcmp((*omg)[0], (*omg)[1], 8) != 0) abort();
cplx_split_fft_ref(h, dat, **omg);
*omg += 2;
}
EXPORT void cplx_ifft_ref(const CPLX_IFFT_PRECOMP* precomp, void* d) {
CPLX* data = (CPLX*)d;
const int32_t m = precomp->m;
const CPLX* omg = (CPLX*)precomp->powomegas;
if (m == 1) return;
if (m <= 8) return cplx_ifft_ref_bfs_2(data, &omg, m);
if (m <= 2048) return cplx_ifft_ref_bfs_16(data, &omg, m);
cplx_ifft_ref_rec_16(data, &omg, m);
}
EXPORT CPLX_IFFT_PRECOMP* new_cplx_ifft_precomp(uint32_t m, uint32_t num_buffers) {
const uint64_t OMG_SPACE = ceilto64b(2 * m * sizeof(CPLX));
const uint64_t BUF_SIZE = ceilto64b(m * sizeof(CPLX));
void* reps = malloc(sizeof(CPLX_IFFT_PRECOMP) + 63 // padding
+ OMG_SPACE // tables
+ num_buffers * BUF_SIZE // buffers
);
uint64_t aligned_addr = ceilto64b((uint64_t)reps + sizeof(CPLX_IFFT_PRECOMP));
CPLX_IFFT_PRECOMP* r = (CPLX_IFFT_PRECOMP*)reps;
r->m = m;
r->buf_size = BUF_SIZE;
r->powomegas = (double*)aligned_addr;
r->aligned_buffers = (void*)(aligned_addr + OMG_SPACE);
// fill in powomegas
CPLX* omg = (CPLX*)r->powomegas;
fill_cplx_ifft_omegas_rec_16(0.25, &omg, m);
if (((uint64_t)omg) - aligned_addr > OMG_SPACE) abort();
{
if (m <= 4) {
// currently, we do not have any acceletated
// implementation for m<=4
r->function = cplx_ifft_ref;
} else if (CPU_SUPPORTS("fma")) {
r->function = cplx_ifft_avx2_fma;
} else {
r->function = cplx_ifft_ref;
}
}
return reps;
}
EXPORT void* cplx_ifft_precomp_get_buffer(const CPLX_IFFT_PRECOMP* itables, uint32_t buffer_index) {
return (uint8_t*)itables->aligned_buffers + buffer_index * itables->buf_size;
}
EXPORT void cplx_ifft(const CPLX_IFFT_PRECOMP* itables, void* data) { itables->function(itables, data); }
EXPORT void cplx_ifft_simple(uint32_t m, void* data) {
static CPLX_IFFT_PRECOMP* p[31] = {0};
CPLX_IFFT_PRECOMP** f = p + log2m(m);
if (!*f) *f = new_cplx_ifft_precomp(m, 0);
(*f)->function(*f, data);
}

0
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/cplx/spqlios_cplx_fft.c
View File

136
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/ext/neon_accel/macrof.h
View File

@@ -1,136 +0,0 @@
/*
* This file is extracted from the implementation of the FFT on Arm64/Neon
* available in https://github.com/cothan/Falcon-Arm (neon/macrof.h).
* =============================================================================
* Copyright (c) 2022 by Cryptographic Engineering Research Group (CERG)
* ECE Department, George Mason University
* Fairfax, VA, U.S.A.
* @author: Duc Tri Nguyen dnguye69@gmu.edu, cothannguyen@gmail.com
* Licensed under the Apache License, Version 2.0 (the "License");
* =============================================================================
*
* This 64-bit Floating point NEON macro x1 has not been modified and is provided as is.
*/
#ifndef MACROF_H
#define MACROF_H
#include <arm_neon.h>
// c <= addr x1
#define vload(c, addr) c = vld1q_f64(addr);
// c <= addr interleave 2
#define vload2(c, addr) c = vld2q_f64(addr);
// c <= addr interleave 4
#define vload4(c, addr) c = vld4q_f64(addr);
#define vstore(addr, c) vst1q_f64(addr, c);
// addr <= c
#define vstore2(addr, c) vst2q_f64(addr, c);
// addr <= c
#define vstore4(addr, c) vst4q_f64(addr, c);
// c <= addr x2
#define vloadx2(c, addr) c = vld1q_f64_x2(addr);
// c <= addr x3
#define vloadx3(c, addr) c = vld1q_f64_x3(addr);
// addr <= c
#define vstorex2(addr, c) vst1q_f64_x2(addr, c);
// c = a - b
#define vfsub(c, a, b) c = vsubq_f64(a, b);
// c = a + b
#define vfadd(c, a, b) c = vaddq_f64(a, b);
// c = a * b
#define vfmul(c, a, b) c = vmulq_f64(a, b);
// c = a * n (n is constant)
#define vfmuln(c, a, n) c = vmulq_n_f64(a, n);
// Swap from a|b to b|a
#define vswap(c, a) c = vextq_f64(a, a, 1);
// c = a * b[i]
#define vfmul_lane(c, a, b, i) c = vmulq_laneq_f64(a, b, i);
// c = 1/a
#define vfinv(c, a) c = vdivq_f64(vdupq_n_f64(1.0), a);
// c = -a
#define vfneg(c, a) c = vnegq_f64(a);
#define transpose_f64(a, b, t, ia, ib, it) \
t.val[it] = a.val[ia]; \
a.val[ia] = vzip1q_f64(a.val[ia], b.val[ib]); \
b.val[ib] = vzip2q_f64(t.val[it], b.val[ib]);
/*
* c = a + jb
* c[0] = a[0] - b[1]
* c[1] = a[1] + b[0]
*/
#define vfcaddj(c, a, b) c = vcaddq_rot90_f64(a, b);
/*
* c = a - jb
* c[0] = a[0] + b[1]
* c[1] = a[1] - b[0]
*/
#define vfcsubj(c, a, b) c = vcaddq_rot270_f64(a, b);
// c[0] = c[0] + b[0]*a[0], c[1] = c[1] + b[1]*a[0]
#define vfcmla(c, a, b) c = vcmlaq_f64(c, a, b);
// c[0] = c[0] - b[1]*a[1], c[1] = c[1] + b[0]*a[1]
#define vfcmla_90(c, a, b) c = vcmlaq_rot90_f64(c, a, b);
// c[0] = c[0] - b[0]*a[0], c[1] = c[1] - b[1]*a[0]
#define vfcmla_180(c, a, b) c = vcmlaq_rot180_f64(c, a, b);
// c[0] = c[0] + b[1]*a[1], c[1] = c[1] - b[0]*a[1]
#define vfcmla_270(c, a, b) c = vcmlaq_rot270_f64(c, a, b);
/*
* Complex MUL: c = a*b
* c[0] = a[0]*b[0] - a[1]*b[1]
* c[1] = a[0]*b[1] + a[1]*b[0]
*/
#define FPC_CMUL(c, a, b) \
c = vmulq_laneq_f64(b, a, 0); \
c = vcmlaq_rot90_f64(c, a, b);
/*
* Complex MUL: c = a * conjugate(b) = a * (b[0], -b[1])
* c[0] = b[0]*a[0] + b[1]*a[1]
* c[1] = + b[0]*a[1] - b[1]*a[0]
*/
#define FPC_CMUL_CONJ(c, a, b) \
c = vmulq_laneq_f64(a, b, 0); \
c = vcmlaq_rot270_f64(c, b, a);
#if FMA == 1
// d = c + a *b
#define vfmla(d, c, a, b) d = vfmaq_f64(c, a, b);
// d = c - a * b
#define vfmls(d, c, a, b) d = vfmsq_f64(c, a, b);
// d = c + a * b[i]
#define vfmla_lane(d, c, a, b, i) d = vfmaq_laneq_f64(c, a, b, i);
// d = c - a * b[i]
#define vfmls_lane(d, c, a, b, i) d = vfmsq_laneq_f64(c, a, b, i);
#else
// d = c + a *b
#define vfmla(d, c, a, b) d = vaddq_f64(c, vmulq_f64(a, b));
// d = c - a *b
#define vfmls(d, c, a, b) d = vsubq_f64(c, vmulq_f64(a, b));
// d = c + a * b[i]
#define vfmla_lane(d, c, a, b, i) d = vaddq_f64(c, vmulq_laneq_f64(a, b, i));
#define vfmls_lane(d, c, a, b, i) d = vsubq_f64(c, vmulq_laneq_f64(a, b, i));
#endif
#endif

420
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/ext/neon_accel/macrofx4.h
View File

@@ -1,420 +0,0 @@
/*
* This file is extracted from the implementation of the FFT on Arm64/Neon
* available in https://github.com/cothan/Falcon-Arm (neon/macrof.h).
* =============================================================================
* Copyright (c) 2022 by Cryptographic Engineering Research Group (CERG)
* ECE Department, George Mason University
* Fairfax, VA, U.S.A.
* @author: Duc Tri Nguyen dnguye69@gmu.edu, cothannguyen@gmail.com
* Licensed under the Apache License, Version 2.0 (the "License");
* =============================================================================
*
* This 64-bit Floating point NEON macro x4 has not been modified and is provided as is.
*/
#ifndef MACROFX4_H
#define MACROFX4_H
#include <arm_neon.h>
#include "macrof.h"
#define vloadx4(c, addr) c = vld1q_f64_x4(addr);
#define vstorex4(addr, c) vst1q_f64_x4(addr, c);
#define vfdupx4(c, constant) \
c.val[0] = vdupq_n_f64(constant); \
c.val[1] = vdupq_n_f64(constant); \
c.val[2] = vdupq_n_f64(constant); \
c.val[3] = vdupq_n_f64(constant);
#define vfnegx4(c, a) \
c.val[0] = vnegq_f64(a.val[0]); \
c.val[1] = vnegq_f64(a.val[1]); \
c.val[2] = vnegq_f64(a.val[2]); \
c.val[3] = vnegq_f64(a.val[3]);
#define vfmulnx4(c, a, n) \
c.val[0] = vmulq_n_f64(a.val[0], n); \
c.val[1] = vmulq_n_f64(a.val[1], n); \
c.val[2] = vmulq_n_f64(a.val[2], n); \
c.val[3] = vmulq_n_f64(a.val[3], n);
// c = a - b
#define vfsubx4(c, a, b) \
c.val[0] = vsubq_f64(a.val[0], b.val[0]); \
c.val[1] = vsubq_f64(a.val[1], b.val[1]); \
c.val[2] = vsubq_f64(a.val[2], b.val[2]); \
c.val[3] = vsubq_f64(a.val[3], b.val[3]);
// c = a + b
#define vfaddx4(c, a, b) \
c.val[0] = vaddq_f64(a.val[0], b.val[0]); \
c.val[1] = vaddq_f64(a.val[1], b.val[1]); \
c.val[2] = vaddq_f64(a.val[2], b.val[2]); \
c.val[3] = vaddq_f64(a.val[3], b.val[3]);
#define vfmulx4(c, a, b) \
c.val[0] = vmulq_f64(a.val[0], b.val[0]); \
c.val[1] = vmulq_f64(a.val[1], b.val[1]); \
c.val[2] = vmulq_f64(a.val[2], b.val[2]); \
c.val[3] = vmulq_f64(a.val[3], b.val[3]);
#define vfmulx4_i(c, a, b) \
c.val[0] = vmulq_f64(a.val[0], b); \
c.val[1] = vmulq_f64(a.val[1], b); \
c.val[2] = vmulq_f64(a.val[2], b); \
c.val[3] = vmulq_f64(a.val[3], b);
#define vfinvx4(c, a) \
c.val[0] = vdivq_f64(vdupq_n_f64(1.0), a.val[0]); \
c.val[1] = vdivq_f64(vdupq_n_f64(1.0), a.val[1]); \
c.val[2] = vdivq_f64(vdupq_n_f64(1.0), a.val[2]); \
c.val[3] = vdivq_f64(vdupq_n_f64(1.0), a.val[3]);
#define vfcvtx4(c, a) \
c.val[0] = vcvtq_f64_s64(a.val[0]); \
c.val[1] = vcvtq_f64_s64(a.val[1]); \
c.val[2] = vcvtq_f64_s64(a.val[2]); \
c.val[3] = vcvtq_f64_s64(a.val[3]);
#define vfmlax4(d, c, a, b) \
vfmla(d.val[0], c.val[0], a.val[0], b.val[0]); \
vfmla(d.val[1], c.val[1], a.val[1], b.val[1]); \
vfmla(d.val[2], c.val[2], a.val[2], b.val[2]); \
vfmla(d.val[3], c.val[3], a.val[3], b.val[3]);
#define vfmlsx4(d, c, a, b) \
vfmls(d.val[0], c.val[0], a.val[0], b.val[0]); \
vfmls(d.val[1], c.val[1], a.val[1], b.val[1]); \
vfmls(d.val[2], c.val[2], a.val[2], b.val[2]); \
vfmls(d.val[3], c.val[3], a.val[3], b.val[3]);
#define vfrintx4(c, a) \
c.val[0] = vcvtnq_s64_f64(a.val[0]); \
c.val[1] = vcvtnq_s64_f64(a.val[1]); \
c.val[2] = vcvtnq_s64_f64(a.val[2]); \
c.val[3] = vcvtnq_s64_f64(a.val[3]);
/*
* Wrapper for FFT, split/merge and poly_float.c
*/
#define FPC_MUL(d_re, d_im, a_re, a_im, b_re, b_im) \
vfmul(d_re, a_re, b_re); \
vfmls(d_re, d_re, a_im, b_im); \
vfmul(d_im, a_re, b_im); \
vfmla(d_im, d_im, a_im, b_re);
#define FPC_MULx2(d_re, d_im, a_re, a_im, b_re, b_im) \
vfmul(d_re.val[0], a_re.val[0], b_re.val[0]); \
vfmls(d_re.val[0], d_re.val[0], a_im.val[0], b_im.val[0]); \
vfmul(d_re.val[1], a_re.val[1], b_re.val[1]); \
vfmls(d_re.val[1], d_re.val[1], a_im.val[1], b_im.val[1]); \
vfmul(d_im.val[0], a_re.val[0], b_im.val[0]); \
vfmla(d_im.val[0], d_im.val[0], a_im.val[0], b_re.val[0]); \
vfmul(d_im.val[1], a_re.val[1], b_im.val[1]); \
vfmla(d_im.val[1], d_im.val[1], a_im.val[1], b_re.val[1]);
#define FPC_MULx4(d_re, d_im, a_re, a_im, b_re, b_im) \
vfmul(d_re.val[0], a_re.val[0], b_re.val[0]); \
vfmls(d_re.val[0], d_re.val[0], a_im.val[0], b_im.val[0]); \
vfmul(d_re.val[1], a_re.val[1], b_re.val[1]); \
vfmls(d_re.val[1], d_re.val[1], a_im.val[1], b_im.val[1]); \
vfmul(d_re.val[2], a_re.val[2], b_re.val[2]); \
vfmls(d_re.val[2], d_re.val[2], a_im.val[2], b_im.val[2]); \
vfmul(d_re.val[3], a_re.val[3], b_re.val[3]); \
vfmls(d_re.val[3], d_re.val[3], a_im.val[3], b_im.val[3]); \
vfmul(d_im.val[0], a_re.val[0], b_im.val[0]); \
vfmla(d_im.val[0], d_im.val[0], a_im.val[0], b_re.val[0]); \
vfmul(d_im.val[1], a_re.val[1], b_im.val[1]); \
vfmla(d_im.val[1], d_im.val[1], a_im.val[1], b_re.val[1]); \
vfmul(d_im.val[2], a_re.val[2], b_im.val[2]); \
vfmla(d_im.val[2], d_im.val[2], a_im.val[2], b_re.val[2]); \
vfmul(d_im.val[3], a_re.val[3], b_im.val[3]); \
vfmla(d_im.val[3], d_im.val[3], a_im.val[3], b_re.val[3]);
#define FPC_MLA(d_re, d_im, a_re, a_im, b_re, b_im) \
vfmla(d_re, d_re, a_re, b_re); \
vfmls(d_re, d_re, a_im, b_im); \
vfmla(d_im, d_im, a_re, b_im); \
vfmla(d_im, d_im, a_im, b_re);
#define FPC_MLAx2(d_re, d_im, a_re, a_im, b_re, b_im) \
vfmla(d_re.val[0], d_re.val[0], a_re.val[0], b_re.val[0]); \
vfmls(d_re.val[0], d_re.val[0], a_im.val[0], b_im.val[0]); \
vfmla(d_re.val[1], d_re.val[1], a_re.val[1], b_re.val[1]); \
vfmls(d_re.val[1], d_re.val[1], a_im.val[1], b_im.val[1]); \
vfmla(d_im.val[0], d_im.val[0], a_re.val[0], b_im.val[0]); \
vfmla(d_im.val[0], d_im.val[0], a_im.val[0], b_re.val[0]); \
vfmla(d_im.val[1], d_im.val[1], a_re.val[1], b_im.val[1]); \
vfmla(d_im.val[1], d_im.val[1], a_im.val[1], b_re.val[1]);
#define FPC_MLAx4(d_re, d_im, a_re, a_im, b_re, b_im) \
vfmla(d_re.val[0], d_re.val[0], a_re.val[0], b_re.val[0]); \
vfmls(d_re.val[0], d_re.val[0], a_im.val[0], b_im.val[0]); \
vfmla(d_re.val[1], d_re.val[1], a_re.val[1], b_re.val[1]); \
vfmls(d_re.val[1], d_re.val[1], a_im.val[1], b_im.val[1]); \
vfmla(d_re.val[2], d_re.val[2], a_re.val[2], b_re.val[2]); \
vfmls(d_re.val[2], d_re.val[2], a_im.val[2], b_im.val[2]); \
vfmla(d_re.val[3], d_re.val[3], a_re.val[3], b_re.val[3]); \
vfmls(d_re.val[3], d_re.val[3], a_im.val[3], b_im.val[3]); \
vfmla(d_im.val[0], d_im.val[0], a_re.val[0], b_im.val[0]); \
vfmla(d_im.val[0], d_im.val[0], a_im.val[0], b_re.val[0]); \
vfmla(d_im.val[1], d_im.val[1], a_re.val[1], b_im.val[1]); \
vfmla(d_im.val[1], d_im.val[1], a_im.val[1], b_re.val[1]); \
vfmla(d_im.val[2], d_im.val[2], a_re.val[2], b_im.val[2]); \
vfmla(d_im.val[2], d_im.val[2], a_im.val[2], b_re.val[2]); \
vfmla(d_im.val[3], d_im.val[3], a_re.val[3], b_im.val[3]); \
vfmla(d_im.val[3], d_im.val[3], a_im.val[3], b_re.val[3]);
#define FPC_MUL_CONJx4(d_re, d_im, a_re, a_im, b_re, b_im) \
vfmul(d_re.val[0], b_im.val[0], a_im.val[0]); \
vfmla(d_re.val[0], d_re.val[0], a_re.val[0], b_re.val[0]); \
vfmul(d_re.val[1], b_im.val[1], a_im.val[1]); \
vfmla(d_re.val[1], d_re.val[1], a_re.val[1], b_re.val[1]); \
vfmul(d_re.val[2], b_im.val[2], a_im.val[2]); \
vfmla(d_re.val[2], d_re.val[2], a_re.val[2], b_re.val[2]); \
vfmul(d_re.val[3], b_im.val[3], a_im.val[3]); \
vfmla(d_re.val[3], d_re.val[3], a_re.val[3], b_re.val[3]); \
vfmul(d_im.val[0], b_re.val[0], a_im.val[0]); \
vfmls(d_im.val[0], d_im.val[0], a_re.val[0], b_im.val[0]); \
vfmul(d_im.val[1], b_re.val[1], a_im.val[1]); \
vfmls(d_im.val[1], d_im.val[1], a_re.val[1], b_im.val[1]); \
vfmul(d_im.val[2], b_re.val[2], a_im.val[2]); \
vfmls(d_im.val[2], d_im.val[2], a_re.val[2], b_im.val[2]); \
vfmul(d_im.val[3], b_re.val[3], a_im.val[3]); \
vfmls(d_im.val[3], d_im.val[3], a_re.val[3], b_im.val[3]);
#define FPC_MLA_CONJx4(d_re, d_im, a_re, a_im, b_re, b_im) \
vfmla(d_re.val[0], d_re.val[0], b_im.val[0], a_im.val[0]); \
vfmla(d_re.val[0], d_re.val[0], a_re.val[0], b_re.val[0]); \
vfmla(d_re.val[1], d_re.val[1], b_im.val[1], a_im.val[1]); \
vfmla(d_re.val[1], d_re.val[1], a_re.val[1], b_re.val[1]); \
vfmla(d_re.val[2], d_re.val[2], b_im.val[2], a_im.val[2]); \
vfmla(d_re.val[2], d_re.val[2], a_re.val[2], b_re.val[2]); \
vfmla(d_re.val[3], d_re.val[3], b_im.val[3], a_im.val[3]); \
vfmla(d_re.val[3], d_re.val[3], a_re.val[3], b_re.val[3]); \
vfmla(d_im.val[0], d_im.val[0], b_re.val[0], a_im.val[0]); \
vfmls(d_im.val[0], d_im.val[0], a_re.val[0], b_im.val[0]); \
vfmla(d_im.val[1], d_im.val[1], b_re.val[1], a_im.val[1]); \
vfmls(d_im.val[1], d_im.val[1], a_re.val[1], b_im.val[1]); \
vfmla(d_im.val[2], d_im.val[2], b_re.val[2], a_im.val[2]); \
vfmls(d_im.val[2], d_im.val[2], a_re.val[2], b_im.val[2]); \
vfmla(d_im.val[3], d_im.val[3], b_re.val[3], a_im.val[3]); \
vfmls(d_im.val[3], d_im.val[3], a_re.val[3], b_im.val[3]);
#define FPC_MUL_LANE(d_re, d_im, a_re, a_im, b_re_im) \
vfmul_lane(d_re, a_re, b_re_im, 0); \
vfmls_lane(d_re, d_re, a_im, b_re_im, 1); \
vfmul_lane(d_im, a_re, b_re_im, 1); \
vfmla_lane(d_im, d_im, a_im, b_re_im, 0);
#define FPC_MUL_LANEx4(d_re, d_im, a_re, a_im, b_re_im) \
vfmul_lane(d_re.val[0], a_re.val[0], b_re_im, 0); \
vfmls_lane(d_re.val[0], d_re.val[0], a_im.val[0], b_re_im, 1); \
vfmul_lane(d_re.val[1], a_re.val[1], b_re_im, 0); \
vfmls_lane(d_re.val[1], d_re.val[1], a_im.val[1], b_re_im, 1); \
vfmul_lane(d_re.val[2], a_re.val[2], b_re_im, 0); \
vfmls_lane(d_re.val[2], d_re.val[2], a_im.val[2], b_re_im, 1); \
vfmul_lane(d_re.val[3], a_re.val[3], b_re_im, 0); \
vfmls_lane(d_re.val[3], d_re.val[3], a_im.val[3], b_re_im, 1); \
vfmul_lane(d_im.val[0], a_re.val[0], b_re_im, 1); \
vfmla_lane(d_im.val[0], d_im.val[0], a_im.val[0], b_re_im, 0); \
vfmul_lane(d_im.val[1], a_re.val[1], b_re_im, 1); \
vfmla_lane(d_im.val[1], d_im.val[1], a_im.val[1], b_re_im, 0); \
vfmul_lane(d_im.val[2], a_re.val[2], b_re_im, 1); \
vfmla_lane(d_im.val[2], d_im.val[2], a_im.val[2], b_re_im, 0); \
vfmul_lane(d_im.val[3], a_re.val[3], b_re_im, 1); \
vfmla_lane(d_im.val[3], d_im.val[3], a_im.val[3], b_re_im, 0);
#define FWD_TOP(t_re, t_im, b_re, b_im, zeta_re, zeta_im) FPC_MUL(t_re, t_im, b_re, b_im, zeta_re, zeta_im);
#define FWD_TOP_LANE(t_re, t_im, b_re, b_im, zeta) FPC_MUL_LANE(t_re, t_im, b_re, b_im, zeta);
#define FWD_TOP_LANEx4(t_re, t_im, b_re, b_im, zeta) FPC_MUL_LANEx4(t_re, t_im, b_re, b_im, zeta);
/*
* FPC
*/
#define FPC_SUB(d_re, d_im, a_re, a_im, b_re, b_im) \
d_re = vsubq_f64(a_re, b_re); \
d_im = vsubq_f64(a_im, b_im);
#define FPC_SUBx4(d_re, d_im, a_re, a_im, b_re, b_im) \
d_re.val[0] = vsubq_f64(a_re.val[0], b_re.val[0]); \
d_im.val[0] = vsubq_f64(a_im.val[0], b_im.val[0]); \
d_re.val[1] = vsubq_f64(a_re.val[1], b_re.val[1]); \
d_im.val[1] = vsubq_f64(a_im.val[1], b_im.val[1]); \
d_re.val[2] = vsubq_f64(a_re.val[2], b_re.val[2]); \
d_im.val[2] = vsubq_f64(a_im.val[2], b_im.val[2]); \
d_re.val[3] = vsubq_f64(a_re.val[3], b_re.val[3]); \
d_im.val[3] = vsubq_f64(a_im.val[3], b_im.val[3]);
#define FPC_ADD(d_re, d_im, a_re, a_im, b_re, b_im) \
d_re = vaddq_f64(a_re, b_re); \
d_im = vaddq_f64(a_im, b_im);
#define FPC_ADDx4(d_re, d_im, a_re, a_im, b_re, b_im) \
d_re.val[0] = vaddq_f64(a_re.val[0], b_re.val[0]); \
d_im.val[0] = vaddq_f64(a_im.val[0], b_im.val[0]); \
d_re.val[1] = vaddq_f64(a_re.val[1], b_re.val[1]); \
d_im.val[1] = vaddq_f64(a_im.val[1], b_im.val[1]); \
d_re.val[2] = vaddq_f64(a_re.val[2], b_re.val[2]); \
d_im.val[2] = vaddq_f64(a_im.val[2], b_im.val[2]); \
d_re.val[3] = vaddq_f64(a_re.val[3], b_re.val[3]); \
d_im.val[3] = vaddq_f64(a_im.val[3], b_im.val[3]);
#define FWD_BOT(a_re, a_im, b_re, b_im, t_re, t_im) \
FPC_SUB(b_re, b_im, a_re, a_im, t_re, t_im); \
FPC_ADD(a_re, a_im, a_re, a_im, t_re, t_im);
#define FWD_BOTx4(a_re, a_im, b_re, b_im, t_re, t_im) \
FPC_SUBx4(b_re, b_im, a_re, a_im, t_re, t_im); \
FPC_ADDx4(a_re, a_im, a_re, a_im, t_re, t_im);
/*
* FPC_J
*/
#define FPC_ADDJ(d_re, d_im, a_re, a_im, b_re, b_im) \
d_re = vsubq_f64(a_re, b_im); \
d_im = vaddq_f64(a_im, b_re);
#define FPC_ADDJx4(d_re, d_im, a_re, a_im, b_re, b_im) \
d_re.val[0] = vsubq_f64(a_re.val[0], b_im.val[0]); \
d_im.val[0] = vaddq_f64(a_im.val[0], b_re.val[0]); \
d_re.val[1] = vsubq_f64(a_re.val[1], b_im.val[1]); \
d_im.val[1] = vaddq_f64(a_im.val[1], b_re.val[1]); \
d_re.val[2] = vsubq_f64(a_re.val[2], b_im.val[2]); \
d_im.val[2] = vaddq_f64(a_im.val[2], b_re.val[2]); \
d_re.val[3] = vsubq_f64(a_re.val[3], b_im.val[3]); \
d_im.val[3] = vaddq_f64(a_im.val[3], b_re.val[3]);
#define FPC_SUBJ(d_re, d_im, a_re, a_im, b_re, b_im) \
d_re = vaddq_f64(a_re, b_im); \
d_im = vsubq_f64(a_im, b_re);
#define FPC_SUBJx4(d_re, d_im, a_re, a_im, b_re, b_im) \
d_re.val[0] = vaddq_f64(a_re.val[0], b_im.val[0]); \
d_im.val[0] = vsubq_f64(a_im.val[0], b_re.val[0]); \
d_re.val[1] = vaddq_f64(a_re.val[1], b_im.val[1]); \
d_im.val[1] = vsubq_f64(a_im.val[1], b_re.val[1]); \
d_re.val[2] = vaddq_f64(a_re.val[2], b_im.val[2]); \
d_im.val[2] = vsubq_f64(a_im.val[2], b_re.val[2]); \
d_re.val[3] = vaddq_f64(a_re.val[3], b_im.val[3]); \
d_im.val[3] = vsubq_f64(a_im.val[3], b_re.val[3]);
#define FWD_BOTJ(a_re, a_im, b_re, b_im, t_re, t_im) \
FPC_SUBJ(b_re, b_im, a_re, a_im, t_re, t_im); \
FPC_ADDJ(a_re, a_im, a_re, a_im, t_re, t_im);
#define FWD_BOTJx4(a_re, a_im, b_re, b_im, t_re, t_im) \
FPC_SUBJx4(b_re, b_im, a_re, a_im, t_re, t_im); \
FPC_ADDJx4(a_re, a_im, a_re, a_im, t_re, t_im);
//============== Inverse FFT
/*
* FPC_J
* a * conj(b)
* Original (without swap):
* d_re = b_im * a_im + a_re * b_re;
* d_im = b_re * a_im - a_re * b_im;
*/
#define FPC_MUL_BOTJ_LANE(d_re, d_im, a_re, a_im, b_re_im) \
vfmul_lane(d_re, a_re, b_re_im, 0); \
vfmla_lane(d_re, d_re, a_im, b_re_im, 1); \
vfmul_lane(d_im, a_im, b_re_im, 0); \
vfmls_lane(d_im, d_im, a_re, b_re_im, 1);
#define FPC_MUL_BOTJ_LANEx4(d_re, d_im, a_re, a_im, b_re_im) \
vfmul_lane(d_re.val[0], a_re.val[0], b_re_im, 0); \
vfmla_lane(d_re.val[0], d_re.val[0], a_im.val[0], b_re_im, 1); \
vfmul_lane(d_im.val[0], a_im.val[0], b_re_im, 0); \
vfmls_lane(d_im.val[0], d_im.val[0], a_re.val[0], b_re_im, 1); \
vfmul_lane(d_re.val[1], a_re.val[1], b_re_im, 0); \
vfmla_lane(d_re.val[1], d_re.val[1], a_im.val[1], b_re_im, 1); \
vfmul_lane(d_im.val[1], a_im.val[1], b_re_im, 0); \
vfmls_lane(d_im.val[1], d_im.val[1], a_re.val[1], b_re_im, 1); \
vfmul_lane(d_re.val[2], a_re.val[2], b_re_im, 0); \
vfmla_lane(d_re.val[2], d_re.val[2], a_im.val[2], b_re_im, 1); \
vfmul_lane(d_im.val[2], a_im.val[2], b_re_im, 0); \
vfmls_lane(d_im.val[2], d_im.val[2], a_re.val[2], b_re_im, 1); \
vfmul_lane(d_re.val[3], a_re.val[3], b_re_im, 0); \
vfmla_lane(d_re.val[3], d_re.val[3], a_im.val[3], b_re_im, 1); \
vfmul_lane(d_im.val[3], a_im.val[3], b_re_im, 0); \
vfmls_lane(d_im.val[3], d_im.val[3], a_re.val[3], b_re_im, 1);
#define FPC_MUL_BOTJ(d_re, d_im, a_re, a_im, b_re, b_im) \
vfmul(d_re, b_im, a_im); \
vfmla(d_re, d_re, a_re, b_re); \
vfmul(d_im, b_re, a_im); \
vfmls(d_im, d_im, a_re, b_im);
#define INV_TOPJ(t_re, t_im, a_re, a_im, b_re, b_im) \
FPC_SUB(t_re, t_im, a_re, a_im, b_re, b_im); \
FPC_ADD(a_re, a_im, a_re, a_im, b_re, b_im);
#define INV_TOPJx4(t_re, t_im, a_re, a_im, b_re, b_im) \
FPC_SUBx4(t_re, t_im, a_re, a_im, b_re, b_im); \
FPC_ADDx4(a_re, a_im, a_re, a_im, b_re, b_im);
#define INV_BOTJ(b_re, b_im, t_re, t_im, zeta_re, zeta_im) FPC_MUL_BOTJ(b_re, b_im, t_re, t_im, zeta_re, zeta_im);
#define INV_BOTJ_LANE(b_re, b_im, t_re, t_im, zeta) FPC_MUL_BOTJ_LANE(b_re, b_im, t_re, t_im, zeta);
#define INV_BOTJ_LANEx4(b_re, b_im, t_re, t_im, zeta) FPC_MUL_BOTJ_LANEx4(b_re, b_im, t_re, t_im, zeta);
/*
* FPC_Jm
* a * -conj(b)
* d_re = a_re * b_im - a_im * b_re;
* d_im = a_im * b_im + a_re * b_re;
*/
#define FPC_MUL_BOTJm_LANE(d_re, d_im, a_re, a_im, b_re_im) \
vfmul_lane(d_re, a_re, b_re_im, 1); \
vfmls_lane(d_re, d_re, a_im, b_re_im, 0); \
vfmul_lane(d_im, a_re, b_re_im, 0); \
vfmla_lane(d_im, d_im, a_im, b_re_im, 1);
#define FPC_MUL_BOTJm_LANEx4(d_re, d_im, a_re, a_im, b_re_im) \
vfmul_lane(d_re.val[0], a_re.val[0], b_re_im, 1); \
vfmls_lane(d_re.val[0], d_re.val[0], a_im.val[0], b_re_im, 0); \
vfmul_lane(d_im.val[0], a_re.val[0], b_re_im, 0); \
vfmla_lane(d_im.val[0], d_im.val[0], a_im.val[0], b_re_im, 1); \
vfmul_lane(d_re.val[1], a_re.val[1], b_re_im, 1); \
vfmls_lane(d_re.val[1], d_re.val[1], a_im.val[1], b_re_im, 0); \
vfmul_lane(d_im.val[1], a_re.val[1], b_re_im, 0); \
vfmla_lane(d_im.val[1], d_im.val[1], a_im.val[1], b_re_im, 1); \
vfmul_lane(d_re.val[2], a_re.val[2], b_re_im, 1); \
vfmls_lane(d_re.val[2], d_re.val[2], a_im.val[2], b_re_im, 0); \
vfmul_lane(d_im.val[2], a_re.val[2], b_re_im, 0); \
vfmla_lane(d_im.val[2], d_im.val[2], a_im.val[2], b_re_im, 1); \
vfmul_lane(d_re.val[3], a_re.val[3], b_re_im, 1); \
vfmls_lane(d_re.val[3], d_re.val[3], a_im.val[3], b_re_im, 0); \
vfmul_lane(d_im.val[3], a_re.val[3], b_re_im, 0); \
vfmla_lane(d_im.val[3], d_im.val[3], a_im.val[3], b_re_im, 1);
#define FPC_MUL_BOTJm(d_re, d_im, a_re, a_im, b_re, b_im) \
vfmul(d_re, a_re, b_im); \
vfmls(d_re, d_re, a_im, b_re); \
vfmul(d_im, a_im, b_im); \
vfmla(d_im, d_im, a_re, b_re);
#define INV_TOPJm(t_re, t_im, a_re, a_im, b_re, b_im) \
FPC_SUB(t_re, t_im, b_re, b_im, a_re, a_im); \
FPC_ADD(a_re, a_im, a_re, a_im, b_re, b_im);
#define INV_TOPJmx4(t_re, t_im, a_re, a_im, b_re, b_im) \
FPC_SUBx4(t_re, t_im, b_re, b_im, a_re, a_im); \
FPC_ADDx4(a_re, a_im, a_re, a_im, b_re, b_im);
#define INV_BOTJm(b_re, b_im, t_re, t_im, zeta_re, zeta_im) FPC_MUL_BOTJm(b_re, b_im, t_re, t_im, zeta_re, zeta_im);
#define INV_BOTJm_LANE(b_re, b_im, t_re, t_im, zeta) FPC_MUL_BOTJm_LANE(b_re, b_im, t_re, t_im, zeta);
#define INV_BOTJm_LANEx4(b_re, b_im, t_re, t_im, zeta) FPC_MUL_BOTJm_LANEx4(b_re, b_im, t_re, t_im, zeta);
#endif

115
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/q120/q120_arithmetic.h
View File

@@ -1,115 +0,0 @@
#ifndef SPQLIOS_Q120_ARITHMETIC_H
#define SPQLIOS_Q120_ARITHMETIC_H
#include <stdint.h>
#include "../commons.h"
#include "q120_common.h"
typedef struct _q120_mat1col_product_baa_precomp q120_mat1col_product_baa_precomp;
typedef struct _q120_mat1col_product_bbb_precomp q120_mat1col_product_bbb_precomp;
typedef struct _q120_mat1col_product_bbc_precomp q120_mat1col_product_bbc_precomp;
EXPORT q120_mat1col_product_baa_precomp* q120_new_vec_mat1col_product_baa_precomp();
EXPORT void q120_delete_vec_mat1col_product_baa_precomp(q120_mat1col_product_baa_precomp*);
EXPORT q120_mat1col_product_bbb_precomp* q120_new_vec_mat1col_product_bbb_precomp();
EXPORT void q120_delete_vec_mat1col_product_bbb_precomp(q120_mat1col_product_bbb_precomp*);
EXPORT q120_mat1col_product_bbc_precomp* q120_new_vec_mat1col_product_bbc_precomp();
EXPORT void q120_delete_vec_mat1col_product_bbc_precomp(q120_mat1col_product_bbc_precomp*);
// ell < 10000
EXPORT void q120_vec_mat1col_product_baa_ref(q120_mat1col_product_baa_precomp*, const uint64_t ell, q120b* const res,
const q120a* const x, const q120a* const y);
EXPORT void q120_vec_mat1col_product_bbb_ref(q120_mat1col_product_bbb_precomp*, const uint64_t ell, q120b* const res,
const q120b* const x, const q120b* const y);
EXPORT void q120_vec_mat1col_product_bbc_ref(q120_mat1col_product_bbc_precomp*, const uint64_t ell, q120b* const res,
const q120b* const x, const q120c* const y);
EXPORT void q120_vec_mat1col_product_baa_avx2(q120_mat1col_product_baa_precomp*, const uint64_t ell, q120b* const res,
const q120a* const x, const q120a* const y);
EXPORT void q120_vec_mat1col_product_bbb_avx2(q120_mat1col_product_bbb_precomp*, const uint64_t ell, q120b* const res,
const q120b* const x, const q120b* const y);
EXPORT void q120_vec_mat1col_product_bbc_avx2(q120_mat1col_product_bbc_precomp*, const uint64_t ell, q120b* const res,
const q120b* const x, const q120c* const y);
EXPORT void q120x2_vec_mat1col_product_bbc_ref(q120_mat1col_product_bbc_precomp* precomp, const uint64_t ell,
q120b* const res, const q120b* const x, const q120c* const y);
EXPORT void q120x2_vec_mat1col_product_bbc_avx2(q120_mat1col_product_bbc_precomp* precomp, const uint64_t ell,
q120b* const res, const q120b* const x, const q120c* const y);
EXPORT void q120x2_vec_mat2cols_product_bbc_ref(q120_mat1col_product_bbc_precomp* precomp, const uint64_t ell,
q120b* const res, const q120b* const x, const q120c* const y);
EXPORT void q120x2_vec_mat2cols_product_bbc_avx2(q120_mat1col_product_bbc_precomp* precomp, const uint64_t ell,
q120b* const res, const q120b* const x, const q120c* const y);
/**
* @brief extract 1 q120x2 block from one q120 ntt vectors
* @param nn the size of each vector
* @param blk the block id to extract (<nn/2)
* @param dst the output: nrows q120x2's dst[i] = src[i](blk)
* @param src the input: nrows q120 ntt vecs's
*/
EXPORT void q120x2_extract_1blk_from_q120b_ref(uint64_t nn, uint64_t blk,
q120x2b* const dst, // 8 doubles
const q120b* const src // a reim vector
);
EXPORT void q120x2_extract_1blk_from_q120c_ref(uint64_t nn, uint64_t blk,
q120x2c* const dst, // 8 doubles
const q120c* const src // a reim vector
);
EXPORT void q120x2_extract_1blk_from_q120b_avx(uint64_t nn, uint64_t blk,
q120x2b* const dst, // 8 doubles
const q120b* const src // a reim vector
);
EXPORT void q120x2_extract_1blk_from_q120c_avx(uint64_t nn, uint64_t blk,
q120x2c* const dst, // 8 doubles
const q120c* const src // a reim vector
);
/**
* @brief extract 1 reim4 block from nrows reim vectors of m complexes
* @param nn the size of each q120
* @param nrows the number of q120 (ntt) vectors
* @param blk the block id to extract (<m/4)
* @param dst the output: nrows q120x2's dst[i] = src[i](blk)
* @param src the input: nrows q120 ntt vectors
*/
EXPORT void q120x2_extract_1blk_from_contiguous_q120b_ref(
uint64_t nn, uint64_t nrows, uint64_t blk,
q120x2b* const dst, // nrows * 2 q120
const q120b* const src // a contiguous array of nrows q120b vectors
);
EXPORT void q120x2_extract_1blk_from_contiguous_q120b_avx(
uint64_t nn, uint64_t nrows, uint64_t blk,
q120x2b* const dst, // nrows * 2 q120
const q120b* const src // a contiguous array of nrows q120b vectors
);
/**
* @brief saves 1 single q120x2 block in a q120 vectors of size nn
* @param nn the size of the output q120
* @param blk the block id to save (<nn/2)
* @param dest the output q120b vector: dst(blk) = src
* @param src the input q120x2b
*/
EXPORT void q120x2b_save_1blk_to_q120b_ref(uint64_t nn, uint64_t blk,
q120b* dest, // 1 reim vector of length m
const q120x2b* src // 8 doubles
);
EXPORT void q120x2b_save_1blk_to_q120b_avx(uint64_t nn, uint64_t blk,
q120b* dest, // 1 reim vector of length m
const q120x2b* src // 8 doubles
);
EXPORT void q120_add_bbb_simple(uint64_t nn, q120b* const res, const q120b* const x, const q120b* const y);
EXPORT void q120_add_ccc_simple(uint64_t nn, q120c* const res, const q120c* const x, const q120c* const y);
EXPORT void q120_c_from_b_simple(uint64_t nn, q120c* const res, const q120b* const x);
EXPORT void q120_b_from_znx64_simple(uint64_t nn, q120b* const res, const int64_t* const x);
EXPORT void q120_c_from_znx64_simple(uint64_t nn, q120c* const res, const int64_t* const x);
EXPORT void q120_b_to_znx128_simple(uint64_t nn, __int128_t* const res, const q120b* const x);
#endif // SPQLIOS_Q120_ARITHMETIC_H

567
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/q120/q120_arithmetic_avx2.c
View File

@@ -1,567 +0,0 @@
#include <immintrin.h>
#include "q120_arithmetic.h"
#include "q120_arithmetic_private.h"
EXPORT void q120_vec_mat1col_product_baa_avx2(q120_mat1col_product_baa_precomp* precomp, const uint64_t ell,
q120b* const res, const q120a* const x, const q120a* const y) {
/**
* Algorithm:
* - res = acc1 + acc2 . ((2^H) % Q)
* - acc1 is the sum of H LSB of products x[i].y[i]
* - acc2 is the sum of 64-H MSB of products x[i]].y[i]
* - for l < 10k acc1 will have H + log2(10000) and acc2 64 - H + log2(10000) bits
* - final sum has max(H, 64 - H + bit_size((2^H) % Q)) + log2(10000) + 1 bits
*/
const uint64_t H = precomp->h;
const __m256i MASK = _mm256_set1_epi64x((UINT64_C(1) << H) - 1);
__m256i acc1 = _mm256_setzero_si256();
__m256i acc2 = _mm256_setzero_si256();
const __m256i* x_ptr = (__m256i*)x;
const __m256i* y_ptr = (__m256i*)y;
for (uint64_t i = 0; i < ell; ++i) {
__m256i a = _mm256_loadu_si256(x_ptr);
__m256i b = _mm256_loadu_si256(y_ptr);
__m256i t = _mm256_mul_epu32(a, b);
acc1 = _mm256_add_epi64(acc1, _mm256_and_si256(t, MASK));
acc2 = _mm256_add_epi64(acc2, _mm256_srli_epi64(t, H));
x_ptr++;
y_ptr++;
}
const __m256i H_POW_RED = _mm256_loadu_si256((__m256i*)precomp->h_pow_red);
__m256i t = _mm256_add_epi64(acc1, _mm256_mul_epu32(acc2, H_POW_RED));
_mm256_storeu_si256((__m256i*)res, t);
}
EXPORT void q120_vec_mat1col_product_bbb_avx2(q120_mat1col_product_bbb_precomp* precomp, const uint64_t ell,
q120b* const res, const q120b* const x, const q120b* const y) {
/**
* Algorithm:
* 1. Split x_i and y_i in 2 32-bit parts and compute the cross-products:
* - x_i = xl_i + xh_i . 2^32
* - y_i = yl_i + yh_i . 2^32
* - A_i = xl_i . yl_i
* - B_i = xl_i . yh_i
* - C_i = xh_i . yl_i
* - D_i = xh_i . yh_i
* - we have x_i . y_i == A_i + (B_i + C_i) . 2^32 + D_i . 2^64
* 2. Split A_i, B_i, C_i and D_i into 2 32-bit parts
* - A_i = Al_i + Ah_i . 2^32
* - B_i = Bl_i + Bh_i . 2^32
* - C_i = Cl_i + Ch_i . 2^32
* - D_i = Dl_i + Dh_i . 2^32
* 3. Compute the sums:
* - S1 = \sum Al_i
* - S2 = \sum (Ah_i + Bl_i + Cl_i)
* - S3 = \sum (Bh_i + Ch_i + Dl_i)
* - S4 = \sum Dh_i
* - here S1, S4 have 32 + log2(ell) bits and S2, S3 have 32 + log2(ell) +
* log2(3) bits
* - for ell == 10000 S2, S3 have < 47 bits
* 4. Split S1, S2, S3 and S4 in 2 24-bit parts (24 = ceil(47/2))
* - S1 = S1l + S1h . 2^24
* - S2 = S2l + S2h . 2^24
* - S3 = S3l + S3h . 2^24
* - S4 = S4l + S4h . 2^24
* 5. Compute final result as:
* - \sum x_i . y_i = S1l + S1h . 2^24
* + S2l . 2^32 + S2h . 2^(32+24)
* + S3l . 2^64 + S3h . 2^(64 + 24)
* + S4l . 2^96 + S4l . 2^(96+24)
* - here the powers of 2 are reduced modulo the primes Q before
* multiplications
* - the result will be on 24 + 3 + bit size of primes Q
*/
const uint64_t H1 = 32;
const __m256i MASK1 = _mm256_set1_epi64x((UINT64_C(1) << H1) - 1);
__m256i s1 = _mm256_setzero_si256();
__m256i s2 = _mm256_setzero_si256();
__m256i s3 = _mm256_setzero_si256();
__m256i s4 = _mm256_setzero_si256();
const __m256i* x_ptr = (__m256i*)x;
const __m256i* y_ptr = (__m256i*)y;
for (uint64_t i = 0; i < ell; ++i) {
__m256i x = _mm256_loadu_si256(x_ptr);
__m256i xl = _mm256_and_si256(x, MASK1);
__m256i xh = _mm256_srli_epi64(x, H1);
__m256i y = _mm256_loadu_si256(y_ptr);
__m256i yl = _mm256_and_si256(y, MASK1);
__m256i yh = _mm256_srli_epi64(y, H1);
__m256i a = _mm256_mul_epu32(xl, yl);
__m256i b = _mm256_mul_epu32(xl, yh);
__m256i c = _mm256_mul_epu32(xh, yl);
__m256i d = _mm256_mul_epu32(xh, yh);
s1 = _mm256_add_epi64(s1, _mm256_and_si256(a, MASK1));
s2 = _mm256_add_epi64(s2, _mm256_srli_epi64(a, H1));
s2 = _mm256_add_epi64(s2, _mm256_and_si256(b, MASK1));
s2 = _mm256_add_epi64(s2, _mm256_and_si256(c, MASK1));
s3 = _mm256_add_epi64(s3, _mm256_srli_epi64(b, H1));
s3 = _mm256_add_epi64(s3, _mm256_srli_epi64(c, H1));
s3 = _mm256_add_epi64(s3, _mm256_and_si256(d, MASK1));
s4 = _mm256_add_epi64(s4, _mm256_srli_epi64(d, H1));
x_ptr++;
y_ptr++;
}
const uint64_t H2 = precomp->h;
const __m256i MASK2 = _mm256_set1_epi64x((UINT64_C(1) << H2) - 1);
const __m256i S1H_POW_RED = _mm256_loadu_si256((__m256i*)precomp->s1h_pow_red);
__m256i s1l = _mm256_and_si256(s1, MASK2);
__m256i s1h = _mm256_srli_epi64(s1, H2);
__m256i t = _mm256_add_epi64(s1l, _mm256_mul_epu32(s1h, S1H_POW_RED));
const __m256i S2L_POW_RED = _mm256_loadu_si256((__m256i*)precomp->s2l_pow_red);
const __m256i S2H_POW_RED = _mm256_loadu_si256((__m256i*)precomp->s2h_pow_red);
__m256i s2l = _mm256_and_si256(s2, MASK2);
__m256i s2h = _mm256_srli_epi64(s2, H2);
t = _mm256_add_epi64(t, _mm256_mul_epu32(s2l, S2L_POW_RED));
t = _mm256_add_epi64(t, _mm256_mul_epu32(s2h, S2H_POW_RED));
const __m256i S3L_POW_RED = _mm256_loadu_si256((__m256i*)precomp->s3l_pow_red);
const __m256i S3H_POW_RED = _mm256_loadu_si256((__m256i*)precomp->s3h_pow_red);
__m256i s3l = _mm256_and_si256(s3, MASK2);
__m256i s3h = _mm256_srli_epi64(s3, H2);
t = _mm256_add_epi64(t, _mm256_mul_epu32(s3l, S3L_POW_RED));
t = _mm256_add_epi64(t, _mm256_mul_epu32(s3h, S3H_POW_RED));
const __m256i S4L_POW_RED = _mm256_loadu_si256((__m256i*)precomp->s4l_pow_red);
const __m256i S4H_POW_RED = _mm256_loadu_si256((__m256i*)precomp->s4h_pow_red);
__m256i s4l = _mm256_and_si256(s4, MASK2);
__m256i s4h = _mm256_srli_epi64(s4, H2);
t = _mm256_add_epi64(t, _mm256_mul_epu32(s4l, S4L_POW_RED));
t = _mm256_add_epi64(t, _mm256_mul_epu32(s4h, S4H_POW_RED));
_mm256_storeu_si256((__m256i*)res, t);
}
EXPORT void q120_vec_mat1col_product_bbc_avx2(q120_mat1col_product_bbc_precomp* precomp, const uint64_t ell,
q120b* const res, const q120b* const x, const q120c* const y) {
/**
* Algorithm:
* 0. We have
* - y0_i == y_i % Q and y1_i == (y_i . 2^32) % Q
* 1. Split x_i in 2 32-bit parts and compute the cross-products:
* - x_i = xl_i + xh_i . 2^32
* - A_i = xl_i . y1_i
* - B_i = xh_i . y2_i
* - we have x_i . y_i == A_i + B_i
* 2. Split A_i and B_i into 2 32-bit parts
* - A_i = Al_i + Ah_i . 2^32
* - B_i = Bl_i + Bh_i . 2^32
* 3. Compute the sums:
* - S1 = \sum Al_i + Bl_i
* - S2 = \sum Ah_i + Bh_i
* - here S1 and S2 have 32 + log2(ell) bits
* - for ell == 10000 S1, S2 have < 46 bits
* 4. Split S2 in 27-bit and 19-bit parts (27+19 == 46)
* - S2 = S2l + S2h . 2^27
* 5. Compute final result as:
* - \sum x_i . y_i = S1 + S2l . 2^32 + S2h . 2^(32+27)
* - here the powers of 2 are reduced modulo the primes Q before
* multiplications
* - the result will be on < 52 bits
*/
const uint64_t H1 = 32;
const __m256i MASK1 = _mm256_set1_epi64x((UINT64_C(1) << H1) - 1);
__m256i s1 = _mm256_setzero_si256();
__m256i s2 = _mm256_setzero_si256();
const __m256i* x_ptr = (__m256i*)x;
const __m256i* y_ptr = (__m256i*)y;
for (uint64_t i = 0; i < ell; ++i) {
__m256i x = _mm256_loadu_si256(x_ptr);
__m256i xl = _mm256_and_si256(x, MASK1);
__m256i xh = _mm256_srli_epi64(x, H1);
__m256i y = _mm256_loadu_si256(y_ptr);
__m256i y0 = _mm256_and_si256(y, MASK1);
__m256i y1 = _mm256_srli_epi64(y, H1);
__m256i a = _mm256_mul_epu32(xl, y0);
__m256i b = _mm256_mul_epu32(xh, y1);
s1 = _mm256_add_epi64(s1, _mm256_and_si256(a, MASK1));
s1 = _mm256_add_epi64(s1, _mm256_and_si256(b, MASK1));
s2 = _mm256_add_epi64(s2, _mm256_srli_epi64(a, H1));
s2 = _mm256_add_epi64(s2, _mm256_srli_epi64(b, H1));
x_ptr++;
y_ptr++;
}
const uint64_t H2 = precomp->h;
const __m256i MASK2 = _mm256_set1_epi64x((UINT64_C(1) << H2) - 1);
__m256i t = s1;
const __m256i S2L_POW_RED = _mm256_loadu_si256((__m256i*)precomp->s2l_pow_red);
const __m256i S2H_POW_RED = _mm256_loadu_si256((__m256i*)precomp->s2h_pow_red);
__m256i s2l = _mm256_and_si256(s2, MASK2);
__m256i s2h = _mm256_srli_epi64(s2, H2);
t = _mm256_add_epi64(t, _mm256_mul_epu32(s2l, S2L_POW_RED));
t = _mm256_add_epi64(t, _mm256_mul_epu32(s2h, S2H_POW_RED));
_mm256_storeu_si256((__m256i*)res, t);
}
/**
* @deprecated keeping this one for history only.
* There is a slight register starvation condition on the q120x2_vec_mat2cols
* strategy below sounds better.
*/
EXPORT void q120x2_vec_mat2cols_product_bbc_avx2_old(q120_mat1col_product_bbc_precomp* precomp, const uint64_t ell,
q120b* const res, const q120b* const x, const q120c* const y) {
__m256i s0 = _mm256_setzero_si256(); // col 1a
__m256i s1 = _mm256_setzero_si256();
__m256i s2 = _mm256_setzero_si256(); // col 1b
__m256i s3 = _mm256_setzero_si256();
__m256i s4 = _mm256_setzero_si256(); // col 2a
__m256i s5 = _mm256_setzero_si256();
__m256i s6 = _mm256_setzero_si256(); // col 2b
__m256i s7 = _mm256_setzero_si256();
__m256i s8, s9, s10, s11;
__m256i s12, s13, s14, s15;
const __m256i* x_ptr = (__m256i*)x;
const __m256i* y_ptr = (__m256i*)y;
__m256i* res_ptr = (__m256i*)res;
for (uint64_t i = 0; i < ell; ++i) {
s8 = _mm256_loadu_si256(x_ptr);
s9 = _mm256_loadu_si256(x_ptr + 1);
s10 = _mm256_srli_epi64(s8, 32);
s11 = _mm256_srli_epi64(s9, 32);
s12 = _mm256_loadu_si256(y_ptr);
s13 = _mm256_loadu_si256(y_ptr + 1);
s14 = _mm256_srli_epi64(s12, 32);
s15 = _mm256_srli_epi64(s13, 32);
s12 = _mm256_mul_epu32(s8, s12); // -> s0,s1
s13 = _mm256_mul_epu32(s9, s13); // -> s2,s3
s14 = _mm256_mul_epu32(s10, s14); // -> s0,s1
s15 = _mm256_mul_epu32(s11, s15); // -> s2,s3
s10 = _mm256_slli_epi64(s12, 32); // -> s0
s11 = _mm256_slli_epi64(s13, 32); // -> s2
s12 = _mm256_srli_epi64(s12, 32); // -> s1
s13 = _mm256_srli_epi64(s13, 32); // -> s3
s10 = _mm256_srli_epi64(s10, 32); // -> s0
s11 = _mm256_srli_epi64(s11, 32); // -> s2
s0 = _mm256_add_epi64(s0, s10);
s1 = _mm256_add_epi64(s1, s12);
s2 = _mm256_add_epi64(s2, s11);
s3 = _mm256_add_epi64(s3, s13);
s10 = _mm256_slli_epi64(s14, 32); // -> s0
s11 = _mm256_slli_epi64(s15, 32); // -> s2
s14 = _mm256_srli_epi64(s14, 32); // -> s1
s15 = _mm256_srli_epi64(s15, 32); // -> s3
s10 = _mm256_srli_epi64(s10, 32); // -> s0
s11 = _mm256_srli_epi64(s11, 32); // -> s2
s0 = _mm256_add_epi64(s0, s10);
s1 = _mm256_add_epi64(s1, s14);
s2 = _mm256_add_epi64(s2, s11);
s3 = _mm256_add_epi64(s3, s15);
// deal with the second column
// s8,s9 are still in place!
s10 = _mm256_srli_epi64(s8, 32);
s11 = _mm256_srli_epi64(s9, 32);
s12 = _mm256_loadu_si256(y_ptr + 2);
s13 = _mm256_loadu_si256(y_ptr + 3);
s14 = _mm256_srli_epi64(s12, 32);
s15 = _mm256_srli_epi64(s13, 32);
s12 = _mm256_mul_epu32(s8, s12); // -> s4,s5
s13 = _mm256_mul_epu32(s9, s13); // -> s6,s7
s14 = _mm256_mul_epu32(s10, s14); // -> s4,s5
s15 = _mm256_mul_epu32(s11, s15); // -> s6,s7
s10 = _mm256_slli_epi64(s12, 32); // -> s4
s11 = _mm256_slli_epi64(s13, 32); // -> s6
s12 = _mm256_srli_epi64(s12, 32); // -> s5
s13 = _mm256_srli_epi64(s13, 32); // -> s7
s10 = _mm256_srli_epi64(s10, 32); // -> s4
s11 = _mm256_srli_epi64(s11, 32); // -> s6
s4 = _mm256_add_epi64(s4, s10);
s5 = _mm256_add_epi64(s5, s12);
s6 = _mm256_add_epi64(s6, s11);
s7 = _mm256_add_epi64(s7, s13);
s10 = _mm256_slli_epi64(s14, 32); // -> s4
s11 = _mm256_slli_epi64(s15, 32); // -> s6
s14 = _mm256_srli_epi64(s14, 32); // -> s5
s15 = _mm256_srli_epi64(s15, 32); // -> s7
s10 = _mm256_srli_epi64(s10, 32); // -> s4
s11 = _mm256_srli_epi64(s11, 32); // -> s6
s4 = _mm256_add_epi64(s4, s10);
s5 = _mm256_add_epi64(s5, s14);
s6 = _mm256_add_epi64(s6, s11);
s7 = _mm256_add_epi64(s7, s15);
x_ptr += 2;
y_ptr += 4;
}
// final reduction
const uint64_t H2 = precomp->h;
s8 = _mm256_set1_epi64x((UINT64_C(1) << H2) - 1); // MASK2
s9 = _mm256_loadu_si256((__m256i*)precomp->s2l_pow_red); // S2L_POW_RED
s10 = _mm256_loadu_si256((__m256i*)precomp->s2h_pow_red); // S2H_POW_RED
//--- s0,s1
s11 = _mm256_and_si256(s1, s8);
s12 = _mm256_srli_epi64(s1, H2);
s13 = _mm256_mul_epu32(s11, s9);
s14 = _mm256_mul_epu32(s12, s10);
s0 = _mm256_add_epi64(s0, s13);
s0 = _mm256_add_epi64(s0, s14);
_mm256_storeu_si256(res_ptr + 0, s0);
//--- s2,s3
s11 = _mm256_and_si256(s3, s8);
s12 = _mm256_srli_epi64(s3, H2);
s13 = _mm256_mul_epu32(s11, s9);
s14 = _mm256_mul_epu32(s12, s10);
s2 = _mm256_add_epi64(s2, s13);
s2 = _mm256_add_epi64(s2, s14);
_mm256_storeu_si256(res_ptr + 1, s2);
//--- s4,s5
s11 = _mm256_and_si256(s5, s8);
s12 = _mm256_srli_epi64(s5, H2);
s13 = _mm256_mul_epu32(s11, s9);
s14 = _mm256_mul_epu32(s12, s10);
s4 = _mm256_add_epi64(s4, s13);
s4 = _mm256_add_epi64(s4, s14);
_mm256_storeu_si256(res_ptr + 2, s4);
//--- s6,s7
s11 = _mm256_and_si256(s7, s8);
s12 = _mm256_srli_epi64(s7, H2);
s13 = _mm256_mul_epu32(s11, s9);
s14 = _mm256_mul_epu32(s12, s10);
s6 = _mm256_add_epi64(s6, s13);
s6 = _mm256_add_epi64(s6, s14);
_mm256_storeu_si256(res_ptr + 3, s6);
}
EXPORT void q120x2_vec_mat2cols_product_bbc_avx2(q120_mat1col_product_bbc_precomp* precomp, const uint64_t ell,
q120b* const res, const q120b* const x, const q120c* const y) {
__m256i s0 = _mm256_setzero_si256(); // col 1a
__m256i s1 = _mm256_setzero_si256();
__m256i s2 = _mm256_setzero_si256(); // col 1b
__m256i s3 = _mm256_setzero_si256();
__m256i s4 = _mm256_setzero_si256(); // col 2a
__m256i s5 = _mm256_setzero_si256();
__m256i s6 = _mm256_setzero_si256(); // col 2b
__m256i s7 = _mm256_setzero_si256();
__m256i s8, s9, s10, s11;
__m256i s12, s13, s14, s15;
s11 = _mm256_set1_epi64x(0xFFFFFFFFUL);
const __m256i* x_ptr = (__m256i*)x;
const __m256i* y_ptr = (__m256i*)y;
__m256i* res_ptr = (__m256i*)res;
for (uint64_t i = 0; i < ell; ++i) {
// treat item a
s8 = _mm256_loadu_si256(x_ptr);
s9 = _mm256_srli_epi64(s8, 32);
s12 = _mm256_loadu_si256(y_ptr);
s13 = _mm256_loadu_si256(y_ptr + 2);
s14 = _mm256_srli_epi64(s12, 32);
s15 = _mm256_srli_epi64(s13, 32);
s12 = _mm256_mul_epu32(s8, s12); // c1a -> s0,s1
s13 = _mm256_mul_epu32(s8, s13); // c2a -> s4,s5
s14 = _mm256_mul_epu32(s9, s14); // c1a -> s0,s1
s15 = _mm256_mul_epu32(s9, s15); // c2a -> s4,s5
s8 = _mm256_and_si256(s12, s11); // -> s0
s9 = _mm256_and_si256(s13, s11); // -> s4
s12 = _mm256_srli_epi64(s12, 32); // -> s1
s13 = _mm256_srli_epi64(s13, 32); // -> s5
s0 = _mm256_add_epi64(s0, s8);
s1 = _mm256_add_epi64(s1, s12);
s4 = _mm256_add_epi64(s4, s9);
s5 = _mm256_add_epi64(s5, s13);
s8 = _mm256_and_si256(s14, s11); // -> s0
s9 = _mm256_and_si256(s15, s11); // -> s4
s14 = _mm256_srli_epi64(s14, 32); // -> s1
s15 = _mm256_srli_epi64(s15, 32); // -> s5
s0 = _mm256_add_epi64(s0, s8);
s1 = _mm256_add_epi64(s1, s14);
s4 = _mm256_add_epi64(s4, s9);
s5 = _mm256_add_epi64(s5, s15);
// treat item b
s8 = _mm256_loadu_si256(x_ptr + 1);
s9 = _mm256_srli_epi64(s8, 32);
s12 = _mm256_loadu_si256(y_ptr + 1);
s13 = _mm256_loadu_si256(y_ptr + 3);
s14 = _mm256_srli_epi64(s12, 32);
s15 = _mm256_srli_epi64(s13, 32);
s12 = _mm256_mul_epu32(s8, s12); // c1b -> s2,s3
s13 = _mm256_mul_epu32(s8, s13); // c2b -> s6,s7
s14 = _mm256_mul_epu32(s9, s14); // c1b -> s2,s3
s15 = _mm256_mul_epu32(s9, s15); // c2b -> s6,s7
s8 = _mm256_and_si256(s12, s11); // -> s2
s9 = _mm256_and_si256(s13, s11); // -> s6
s12 = _mm256_srli_epi64(s12, 32); // -> s3
s13 = _mm256_srli_epi64(s13, 32); // -> s7
s2 = _mm256_add_epi64(s2, s8);
s3 = _mm256_add_epi64(s3, s12);
s6 = _mm256_add_epi64(s6, s9);
s7 = _mm256_add_epi64(s7, s13);
s8 = _mm256_and_si256(s14, s11); // -> s2
s9 = _mm256_and_si256(s15, s11); // -> s6
s14 = _mm256_srli_epi64(s14, 32); // -> s3
s15 = _mm256_srli_epi64(s15, 32); // -> s7
s2 = _mm256_add_epi64(s2, s8);
s3 = _mm256_add_epi64(s3, s14);
s6 = _mm256_add_epi64(s6, s9);
s7 = _mm256_add_epi64(s7, s15);
x_ptr += 2;
y_ptr += 4;
}
// final reduction
const uint64_t H2 = precomp->h;
s8 = _mm256_set1_epi64x((UINT64_C(1) << H2) - 1); // MASK2
s9 = _mm256_loadu_si256((__m256i*)precomp->s2l_pow_red); // S2L_POW_RED
s10 = _mm256_loadu_si256((__m256i*)precomp->s2h_pow_red); // S2H_POW_RED
//--- s0,s1
s11 = _mm256_and_si256(s1, s8);
s12 = _mm256_srli_epi64(s1, H2);
s13 = _mm256_mul_epu32(s11, s9);
s14 = _mm256_mul_epu32(s12, s10);
s0 = _mm256_add_epi64(s0, s13);
s0 = _mm256_add_epi64(s0, s14);
_mm256_storeu_si256(res_ptr + 0, s0);
//--- s2,s3
s11 = _mm256_and_si256(s3, s8);
s12 = _mm256_srli_epi64(s3, H2);
s13 = _mm256_mul_epu32(s11, s9);
s14 = _mm256_mul_epu32(s12, s10);
s2 = _mm256_add_epi64(s2, s13);
s2 = _mm256_add_epi64(s2, s14);
_mm256_storeu_si256(res_ptr + 1, s2);
//--- s4,s5
s11 = _mm256_and_si256(s5, s8);
s12 = _mm256_srli_epi64(s5, H2);
s13 = _mm256_mul_epu32(s11, s9);
s14 = _mm256_mul_epu32(s12, s10);
s4 = _mm256_add_epi64(s4, s13);
s4 = _mm256_add_epi64(s4, s14);
_mm256_storeu_si256(res_ptr + 2, s4);
//--- s6,s7
s11 = _mm256_and_si256(s7, s8);
s12 = _mm256_srli_epi64(s7, H2);
s13 = _mm256_mul_epu32(s11, s9);
s14 = _mm256_mul_epu32(s12, s10);
s6 = _mm256_add_epi64(s6, s13);
s6 = _mm256_add_epi64(s6, s14);
_mm256_storeu_si256(res_ptr + 3, s6);
}
EXPORT void q120x2_vec_mat1col_product_bbc_avx2(q120_mat1col_product_bbc_precomp* precomp, const uint64_t ell,
q120b* const res, const q120b* const x, const q120c* const y) {
__m256i s0 = _mm256_setzero_si256(); // col 1a
__m256i s1 = _mm256_setzero_si256();
__m256i s2 = _mm256_setzero_si256(); // col 1b
__m256i s3 = _mm256_setzero_si256();
__m256i s4 = _mm256_set1_epi64x(0xFFFFFFFFUL);
__m256i s8, s9, s10, s11;
__m256i s12, s13, s14, s15;
const __m256i* x_ptr = (__m256i*)x;
const __m256i* y_ptr = (__m256i*)y;
__m256i* res_ptr = (__m256i*)res;
for (uint64_t i = 0; i < ell; ++i) {
s8 = _mm256_loadu_si256(x_ptr);
s9 = _mm256_loadu_si256(x_ptr + 1);
s10 = _mm256_srli_epi64(s8, 32);
s11 = _mm256_srli_epi64(s9, 32);
s12 = _mm256_loadu_si256(y_ptr);
s13 = _mm256_loadu_si256(y_ptr + 1);
s14 = _mm256_srli_epi64(s12, 32);
s15 = _mm256_srli_epi64(s13, 32);
s12 = _mm256_mul_epu32(s8, s12); // -> s0,s1
s13 = _mm256_mul_epu32(s9, s13); // -> s2,s3
s14 = _mm256_mul_epu32(s10, s14); // -> s0,s1
s15 = _mm256_mul_epu32(s11, s15); // -> s2,s3
s8 = _mm256_and_si256(s12, s4); // -> s0
s9 = _mm256_and_si256(s13, s4); // -> s2
s10 = _mm256_and_si256(s14, s4); // -> s0
s11 = _mm256_and_si256(s15, s4); // -> s2
s12 = _mm256_srli_epi64(s12, 32); // -> s1
s13 = _mm256_srli_epi64(s13, 32); // -> s3
s14 = _mm256_srli_epi64(s14, 32); // -> s1
s15 = _mm256_srli_epi64(s15, 32); // -> s3
s0 = _mm256_add_epi64(s0, s8);
s1 = _mm256_add_epi64(s1, s12);
s2 = _mm256_add_epi64(s2, s9);
s3 = _mm256_add_epi64(s3, s13);
s0 = _mm256_add_epi64(s0, s10);
s1 = _mm256_add_epi64(s1, s14);
s2 = _mm256_add_epi64(s2, s11);
s3 = _mm256_add_epi64(s3, s15);
x_ptr += 2;
y_ptr += 2;
}
// final reduction
const uint64_t H2 = precomp->h;
s8 = _mm256_set1_epi64x((UINT64_C(1) << H2) - 1); // MASK2
s9 = _mm256_loadu_si256((__m256i*)precomp->s2l_pow_red); // S2L_POW_RED
s10 = _mm256_loadu_si256((__m256i*)precomp->s2h_pow_red); // S2H_POW_RED
//--- s0,s1
s11 = _mm256_and_si256(s1, s8);
s12 = _mm256_srli_epi64(s1, H2);
s13 = _mm256_mul_epu32(s11, s9);
s14 = _mm256_mul_epu32(s12, s10);
s0 = _mm256_add_epi64(s0, s13);
s0 = _mm256_add_epi64(s0, s14);
_mm256_storeu_si256(res_ptr + 0, s0);
//--- s2,s3
s11 = _mm256_and_si256(s3, s8);
s12 = _mm256_srli_epi64(s3, H2);
s13 = _mm256_mul_epu32(s11, s9);
s14 = _mm256_mul_epu32(s12, s10);
s2 = _mm256_add_epi64(s2, s13);
s2 = _mm256_add_epi64(s2, s14);
_mm256_storeu_si256(res_ptr + 1, s2);
}

37
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/q120/q120_arithmetic_private.h
View File

@@ -1,37 +0,0 @@
#ifndef SPQLIOS_Q120_ARITHMETIC_DEF_H
#define SPQLIOS_Q120_ARITHMETIC_DEF_H
#include <stdint.h>
typedef struct _q120_mat1col_product_baa_precomp {
uint64_t h;
uint64_t h_pow_red[4];
#ifndef NDEBUG
double res_bit_size;
#endif
} q120_mat1col_product_baa_precomp;
typedef struct _q120_mat1col_product_bbb_precomp {
uint64_t h;
uint64_t s1h_pow_red[4];
uint64_t s2l_pow_red[4];
uint64_t s2h_pow_red[4];
uint64_t s3l_pow_red[4];
uint64_t s3h_pow_red[4];
uint64_t s4l_pow_red[4];
uint64_t s4h_pow_red[4];
#ifndef NDEBUG
double res_bit_size;
#endif
} q120_mat1col_product_bbb_precomp;
typedef struct _q120_mat1col_product_bbc_precomp {
uint64_t h;
uint64_t s2l_pow_red[4];
uint64_t s2h_pow_red[4];
#ifndef NDEBUG
double res_bit_size;
#endif
} q120_mat1col_product_bbc_precomp;
#endif // SPQLIOS_Q120_ARITHMETIC_DEF_H

506
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/q120/q120_arithmetic_ref.c
View File

@@ -1,506 +0,0 @@
#include <assert.h>
#include <math.h>
#include <stdint.h>
#include "q120_arithmetic.h"
#include "q120_arithmetic_private.h"
#include "q120_common.h"
#define MODQ(val, q) ((val) % (q))
double comp_bit_size_red(const uint64_t h, const uint64_t qs[4]) {
assert(h < 128);
double h_pow2_bs = 0;
for (uint64_t k = 0; k < 4; ++k) {
double t = log2((double)MODQ((__uint128_t)1 << h, qs[k]));
if (t > h_pow2_bs) h_pow2_bs = t;
}
return h_pow2_bs;
}
double comp_bit_size_sum(const uint64_t n, const double* const bs) {
double s = 0;
for (uint64_t i = 0; i < n; ++i) {
s += pow(2, bs[i]);
}
return log2(s);
}
void vec_mat1col_product_baa_precomp(q120_mat1col_product_baa_precomp* precomp) {
uint64_t qs[4] = {Q1, Q2, Q3, Q4};
double min_res_bs = 1000;
uint64_t min_h = -1;
double ell_bs = log2((double)MAX_ELL);
for (uint64_t h = 1; h < 64; ++h) {
double h_pow2_bs = comp_bit_size_red(h, qs);
const double bs[] = {h + ell_bs, 64 - h + ell_bs + h_pow2_bs};
const double res_bs = comp_bit_size_sum(2, bs);
if (min_res_bs > res_bs) {
min_res_bs = res_bs;
min_h = h;
}
}
assert(min_res_bs < 64);
precomp->h = min_h;
for (uint64_t k = 0; k < 4; ++k) {
precomp->h_pow_red[k] = MODQ(UINT64_C(1) << precomp->h, qs[k]);
}
#ifndef NDEBUG
precomp->res_bit_size = min_res_bs;
#endif
// printf("AA %lu %lf\n", min_h, min_res_bs);
}
EXPORT q120_mat1col_product_baa_precomp* q120_new_vec_mat1col_product_baa_precomp() {
q120_mat1col_product_baa_precomp* res = malloc(sizeof(q120_mat1col_product_baa_precomp));
vec_mat1col_product_baa_precomp(res);
return res;
}
EXPORT void q120_delete_vec_mat1col_product_baa_precomp(q120_mat1col_product_baa_precomp* addr) { free(addr); }
EXPORT void q120_vec_mat1col_product_baa_ref(q120_mat1col_product_baa_precomp* precomp, const uint64_t ell,
q120b* const res, const q120a* const x, const q120a* const y) {
/**
* Algorithm:
* - res = acc1 + acc2 . ((2^H) % Q)
* - acc1 is the sum of H LSB of products x[i].y[i]
* - acc2 is the sum of 64-H MSB of products x[i]].y[i]
* - for l < 10k acc1 will have H + log2(10000) and acc2 64 - H + log2(10000) bits
* - final sum has max(H, 64 - H + bit_size((2^H) % Q)) + log2(10000) + 1 bits
*/
const uint64_t H = precomp->h;
const uint64_t MASK = (UINT64_C(1) << H) - 1;
uint64_t acc1[4] = {0, 0, 0, 0}; // accumulate H least significant bits of product
uint64_t acc2[4] = {0, 0, 0, 0}; // accumulate 64 - H most significan bits of product
const uint64_t* const x_ptr = (uint64_t*)x;
const uint64_t* const y_ptr = (uint64_t*)y;
for (uint64_t i = 0; i < 4 * ell; i += 4) {
for (uint64_t j = 0; j < 4; ++j) {
uint64_t t = x_ptr[i + j] * y_ptr[i + j];
acc1[j] += t & MASK;
acc2[j] += t >> H;
}
}
uint64_t* const res_ptr = (uint64_t*)res;
for (uint64_t j = 0; j < 4; ++j) {
res_ptr[j] = acc1[j] + acc2[j] * precomp->h_pow_red[j];
assert(log2(res_ptr[j]) < precomp->res_bit_size);
}
}
void vec_mat1col_product_bbb_precomp(q120_mat1col_product_bbb_precomp* precomp) {
uint64_t qs[4] = {Q1, Q2, Q3, Q4};
double ell_bs = log2((double)MAX_ELL);
double min_res_bs = 1000;
uint64_t min_h = -1;
const double s1_bs = 32 + ell_bs;
const double s2_bs = 32 + ell_bs + log2(3);
const double s3_bs = 32 + ell_bs + log2(3);
const double s4_bs = 32 + ell_bs;
for (uint64_t h = 16; h < 32; ++h) {
const double s1l_bs = h;
const double s1h_bs = (s1_bs - h) + comp_bit_size_red(h, qs);
const double s2l_bs = h + comp_bit_size_red(32, qs);
const double s2h_bs = (s2_bs - h) + comp_bit_size_red(32 + h, qs);
const double s3l_bs = h + comp_bit_size_red(64, qs);
const double s3h_bs = (s3_bs - h) + comp_bit_size_red(64 + h, qs);
const double s4l_bs = h + comp_bit_size_red(96, qs);
const double s4h_bs = (s4_bs - h) + comp_bit_size_red(96 + h, qs);
const double bs[] = {s1l_bs, s1h_bs, s2l_bs, s2h_bs, s3l_bs, s3h_bs, s4l_bs, s4h_bs};
const double res_bs = comp_bit_size_sum(8, bs);
if (min_res_bs > res_bs) {
min_res_bs = res_bs;
min_h = h;
}
}
assert(min_res_bs < 64);
precomp->h = min_h;
for (uint64_t k = 0; k < 4; ++k) {
precomp->s1h_pow_red[k] = UINT64_C(1) << precomp->h; // 2^24
precomp->s2l_pow_red[k] = MODQ(UINT64_C(1) << 32, qs[k]); // 2^32
precomp->s2h_pow_red[k] = MODQ(precomp->s2l_pow_red[k] * precomp->s1h_pow_red[k], qs[k]); // 2^(32+24)
precomp->s3l_pow_red[k] = MODQ(precomp->s2l_pow_red[k] * precomp->s2l_pow_red[k], qs[k]); // 2^64 = 2^(32+32)
precomp->s3h_pow_red[k] = MODQ(precomp->s3l_pow_red[k] * precomp->s1h_pow_red[k], qs[k]); // 2^(64+24)
precomp->s4l_pow_red[k] = MODQ(precomp->s3l_pow_red[k] * precomp->s2l_pow_red[k], qs[k]); // 2^96 = 2^(64+32)
precomp->s4h_pow_red[k] = MODQ(precomp->s4l_pow_red[k] * precomp->s1h_pow_red[k], qs[k]); // 2^(96+24)
}
// printf("AA %lu %lf\n", min_h, min_res_bs);
#ifndef NDEBUG
precomp->res_bit_size = min_res_bs;
#endif
}
EXPORT q120_mat1col_product_bbb_precomp* q120_new_vec_mat1col_product_bbb_precomp() {
q120_mat1col_product_bbb_precomp* res = malloc(sizeof(q120_mat1col_product_bbb_precomp));
vec_mat1col_product_bbb_precomp(res);
return res;
}
EXPORT void q120_delete_vec_mat1col_product_bbb_precomp(q120_mat1col_product_bbb_precomp* addr) { free(addr); }
EXPORT void q120_vec_mat1col_product_bbb_ref(q120_mat1col_product_bbb_precomp* precomp, const uint64_t ell,
q120b* const res, const q120b* const x, const q120b* const y) {
/**
* Algorithm:
* 1. Split x_i and y_i in 2 32-bit parts and compute the cross-products:
* - x_i = xl_i + xh_i . 2^32
* - y_i = yl_i + yh_i . 2^32
* - A_i = xl_i . yl_i
* - B_i = xl_i . yh_i
* - C_i = xh_i . yl_i
* - D_i = xh_i . yh_i
* - we have x_i . y_i == A_i + (B_i + C_i) . 2^32 + D_i . 2^64
* 2. Split A_i, B_i, C_i and D_i into 2 32-bit parts
* - A_i = Al_i + Ah_i . 2^32
* - B_i = Bl_i + Bh_i . 2^32
* - C_i = Cl_i + Ch_i . 2^32
* - D_i = Dl_i + Dh_i . 2^32
* 3. Compute the sums:
* - S1 = \sum Al_i
* - S2 = \sum (Ah_i + Bl_i + Cl_i)
* - S3 = \sum (Bh_i + Ch_i + Dl_i)
* - S4 = \sum Dh_i
* - here S1, S4 have 32 + log2(ell) bits and S2, S3 have 32 + log2(ell) +
* log2(3) bits
* - for ell == 10000 S2, S3 have < 47 bits
* 4. Split S1, S2, S3 and S4 in 2 24-bit parts (24 = ceil(47/2))
* - S1 = S1l + S1h . 2^24
* - S2 = S2l + S2h . 2^24
* - S3 = S3l + S3h . 2^24
* - S4 = S4l + S4h . 2^24
* 5. Compute final result as:
* - \sum x_i . y_i = S1l + S1h . 2^24
* + S2l . 2^32 + S2h . 2^(32+24)
* + S3l . 2^64 + S3h . 2^(64 + 24)
* + S4l . 2^96 + S4l . 2^(96+24)
* - here the powers of 2 are reduced modulo the primes Q before
* multiplications
* - the result will be on 24 + 3 + bit size of primes Q
*/
const uint64_t H1 = 32;
const uint64_t MASK1 = (UINT64_C(1) << H1) - 1;
uint64_t s1[4] = {0, 0, 0, 0};
uint64_t s2[4] = {0, 0, 0, 0};
uint64_t s3[4] = {0, 0, 0, 0};
uint64_t s4[4] = {0, 0, 0, 0};
const uint64_t* const x_ptr = (uint64_t*)x;
const uint64_t* const y_ptr = (uint64_t*)y;
for (uint64_t i = 0; i < 4 * ell; i += 4) {
for (uint64_t j = 0; j < 4; ++j) {
const uint64_t xl = x_ptr[i + j] & MASK1;
const uint64_t xh = x_ptr[i + j] >> H1;
const uint64_t yl = y_ptr[i + j] & MASK1;
const uint64_t yh = y_ptr[i + j] >> H1;
const uint64_t a = xl * yl;
const uint64_t al = a & MASK1;
const uint64_t ah = a >> H1;
const uint64_t b = xl * yh;
const uint64_t bl = b & MASK1;
const uint64_t bh = b >> H1;
const uint64_t c = xh * yl;
const uint64_t cl = c & MASK1;
const uint64_t ch = c >> H1;
const uint64_t d = xh * yh;
const uint64_t dl = d & MASK1;
const uint64_t dh = d >> H1;
s1[j] += al;
s2[j] += ah + bl + cl;
s3[j] += bh + ch + dl;
s4[j] += dh;
}
}
const uint64_t H2 = precomp->h;
const uint64_t MASK2 = (UINT64_C(1) << H2) - 1;
uint64_t* const res_ptr = (uint64_t*)res;
for (uint64_t j = 0; j < 4; ++j) {
const uint64_t s1l = s1[j] & MASK2;
const uint64_t s1h = s1[j] >> H2;
const uint64_t s2l = s2[j] & MASK2;
const uint64_t s2h = s2[j] >> H2;
const uint64_t s3l = s3[j] & MASK2;
const uint64_t s3h = s3[j] >> H2;
const uint64_t s4l = s4[j] & MASK2;
const uint64_t s4h = s4[j] >> H2;
uint64_t t = s1l;
t += s1h * precomp->s1h_pow_red[j];
t += s2l * precomp->s2l_pow_red[j];
t += s2h * precomp->s2h_pow_red[j];
t += s3l * precomp->s3l_pow_red[j];
t += s3h * precomp->s3h_pow_red[j];
t += s4l * precomp->s4l_pow_red[j];
t += s4h * precomp->s4h_pow_red[j];
res_ptr[j] = t;
assert(log2(res_ptr[j]) < precomp->res_bit_size);
}
}
void vec_mat1col_product_bbc_precomp(q120_mat1col_product_bbc_precomp* precomp) {
uint64_t qs[4] = {Q1, Q2, Q3, Q4};
double min_res_bs = 1000;
uint64_t min_h = -1;
double pow2_32_bs = comp_bit_size_red(32, qs);
double ell_bs = log2((double)MAX_ELL);
double s1_bs = 32 + ell_bs;
for (uint64_t h = 16; h < 32; ++h) {
double s2l_bs = pow2_32_bs + h;
double s2h_bs = s1_bs - h + comp_bit_size_red(32 + h, qs);
const double bs[] = {s1_bs, s2l_bs, s2h_bs};
const double res_bs = comp_bit_size_sum(3, bs);
if (min_res_bs > res_bs) {
min_res_bs = res_bs;
min_h = h;
}
}
assert(min_res_bs < 64);
precomp->h = min_h;
for (uint64_t k = 0; k < 4; ++k) {
precomp->s2l_pow_red[k] = MODQ(UINT64_C(1) << 32, qs[k]);
precomp->s2h_pow_red[k] = MODQ(UINT64_C(1) << (32 + precomp->h), qs[k]);
}
#ifndef NDEBUG
precomp->res_bit_size = min_res_bs;
#endif
// printf("AA %lu %lf\n", min_h, min_res_bs);
}
EXPORT q120_mat1col_product_bbc_precomp* q120_new_vec_mat1col_product_bbc_precomp() {
q120_mat1col_product_bbc_precomp* res = malloc(sizeof(q120_mat1col_product_bbc_precomp));
vec_mat1col_product_bbc_precomp(res);
return res;
}
EXPORT void q120_delete_vec_mat1col_product_bbc_precomp(q120_mat1col_product_bbc_precomp* addr) { free(addr); }
EXPORT void q120_vec_mat1col_product_bbc_ref_old(q120_mat1col_product_bbc_precomp* precomp, const uint64_t ell,
q120b* const res, const q120b* const x, const q120c* const y) {
/**
* Algorithm:
* 0. We have
* - y0_i == y_i % Q and y1_i == (y_i . 2^32) % Q
* 1. Split x_i in 2 32-bit parts and compute the cross-products:
* - x_i = xl_i + xh_i . 2^32
* - A_i = xl_i . y1_i
* - B_i = xh_i . y2_i
* - we have x_i . y_i == A_i + B_i
* 2. Split A_i and B_i into 2 32-bit parts
* - A_i = Al_i + Ah_i . 2^32
* - B_i = Bl_i + Bh_i . 2^32
* 3. Compute the sums:
* - S1 = \sum Al_i + Bl_i
* - S2 = \sum Ah_i + Bh_i
* - here S1 and S2 have 32 + log2(ell) bits
* - for ell == 10000 S1, S2 have < 46 bits
* 4. Split S2 in 27-bit and 19-bit parts (27+19 == 46)
* - S2 = S2l + S2h . 2^27
* 5. Compute final result as:
* - \sum x_i . y_i = S1 + S2l . 2^32 + S2h . 2^(32+27)
* - here the powers of 2 are reduced modulo the primes Q before
* multiplications
* - the result will be on < 52 bits
*/
const uint64_t H1 = 32;
const uint64_t MASK1 = (UINT64_C(1) << H1) - 1;
uint64_t s1[4] = {0, 0, 0, 0};
uint64_t s2[4] = {0, 0, 0, 0};
const uint64_t* const x_ptr = (uint64_t*)x;
const uint32_t* const y_ptr = (uint32_t*)y;
for (uint64_t i = 0; i < 4 * ell; i += 4) {
for (uint64_t j = 0; j < 4; ++j) {
const uint64_t xl = x_ptr[i + j] & MASK1;
const uint64_t xh = x_ptr[i + j] >> H1;
const uint64_t y0 = y_ptr[2 * (i + j)];
const uint64_t y1 = y_ptr[2 * (i + j) + 1];
const uint64_t a = xl * y0;
const uint64_t al = a & MASK1;
const uint64_t ah = a >> H1;
const uint64_t b = xh * y1;
const uint64_t bl = b & MASK1;
const uint64_t bh = b >> H1;
s1[j] += al + bl;
s2[j] += ah + bh;
}
}
const uint64_t H2 = precomp->h;
const uint64_t MASK2 = (UINT64_C(1) << H2) - 1;
uint64_t* const res_ptr = (uint64_t*)res;
for (uint64_t k = 0; k < 4; ++k) {
const uint64_t s2l = s2[k] & MASK2;
const uint64_t s2h = s2[k] >> H2;
uint64_t t = s1[k];
t += s2l * precomp->s2l_pow_red[k];
t += s2h * precomp->s2h_pow_red[k];
res_ptr[k] = t;
assert(log2(res_ptr[k]) < precomp->res_bit_size);
}
}
static __always_inline void accum_mul_q120_bc(uint64_t res[8], //
const uint32_t x_layb[8], const uint32_t y_layc[8]) {
for (uint64_t i = 0; i < 4; ++i) {
static const uint64_t MASK32 = 0xFFFFFFFFUL;
uint64_t x_lo = x_layb[2 * i];
uint64_t x_hi = x_layb[2 * i + 1];
uint64_t y_lo = y_layc[2 * i];
uint64_t y_hi = y_layc[2 * i + 1];
uint64_t xy_lo = x_lo * y_lo;
uint64_t xy_hi = x_hi * y_hi;
res[2 * i] += (xy_lo & MASK32) + (xy_hi & MASK32);
res[2 * i + 1] += (xy_lo >> 32) + (xy_hi >> 32);
}
}
static __always_inline void accum_to_q120b(uint64_t res[4], //
const uint64_t s[8], const q120_mat1col_product_bbc_precomp* precomp) {
const uint64_t H2 = precomp->h;
const uint64_t MASK2 = (UINT64_C(1) << H2) - 1;
for (uint64_t k = 0; k < 4; ++k) {
const uint64_t s2l = s[2 * k + 1] & MASK2;
const uint64_t s2h = s[2 * k + 1] >> H2;
uint64_t t = s[2 * k];
t += s2l * precomp->s2l_pow_red[k];
t += s2h * precomp->s2h_pow_red[k];
res[k] = t;
assert(log2(res[k]) < precomp->res_bit_size);
}
}
EXPORT void q120_vec_mat1col_product_bbc_ref(q120_mat1col_product_bbc_precomp* precomp, const uint64_t ell,
q120b* const res, const q120b* const x, const q120c* const y) {
uint64_t s[8] = {0, 0, 0, 0, 0, 0, 0, 0};
const uint32_t(*const x_ptr)[8] = (const uint32_t(*const)[8])x;
const uint32_t(*const y_ptr)[8] = (const uint32_t(*const)[8])y;
for (uint64_t i = 0; i < ell; i++) {
accum_mul_q120_bc(s, x_ptr[i], y_ptr[i]);
}
accum_to_q120b((uint64_t*)res, s, precomp);
}
EXPORT void q120x2_vec_mat1col_product_bbc_ref(q120_mat1col_product_bbc_precomp* precomp, const uint64_t ell,
q120b* const res, const q120b* const x, const q120c* const y) {
uint64_t s[2][16] = {0};
const uint32_t(*const x_ptr)[2][8] = (const uint32_t(*const)[2][8])x;
const uint32_t(*const y_ptr)[2][8] = (const uint32_t(*const)[2][8])y;
uint64_t(*re)[4] = (uint64_t(*)[4])res;
for (uint64_t i = 0; i < ell; i++) {
accum_mul_q120_bc(s[0], x_ptr[i][0], y_ptr[i][0]);
accum_mul_q120_bc(s[1], x_ptr[i][1], y_ptr[i][1]);
}
accum_to_q120b(re[0], s[0], precomp);
accum_to_q120b(re[1], s[1], precomp);
}
EXPORT void q120x2_vec_mat2cols_product_bbc_ref(q120_mat1col_product_bbc_precomp* precomp, const uint64_t ell,
q120b* const res, const q120b* const x, const q120c* const y) {
uint64_t s[4][16] = {0};
const uint32_t(*const x_ptr)[2][8] = (const uint32_t(*const)[2][8])x;
const uint32_t(*const y_ptr)[4][8] = (const uint32_t(*const)[4][8])y;
uint64_t(*re)[4] = (uint64_t(*)[4])res;
for (uint64_t i = 0; i < ell; i++) {
accum_mul_q120_bc(s[0], x_ptr[i][0], y_ptr[i][0]);
accum_mul_q120_bc(s[1], x_ptr[i][1], y_ptr[i][1]);
accum_mul_q120_bc(s[2], x_ptr[i][0], y_ptr[i][2]);
accum_mul_q120_bc(s[3], x_ptr[i][1], y_ptr[i][3]);
}
accum_to_q120b(re[0], s[0], precomp);
accum_to_q120b(re[1], s[1], precomp);
accum_to_q120b(re[2], s[2], precomp);
accum_to_q120b(re[3], s[3], precomp);
}
EXPORT void q120x2_extract_1blk_from_q120b_ref(uint64_t nn, uint64_t blk,
q120x2b* const dst, // 8 doubles
const q120b* const src // a q120b vector
) {
const uint64_t* in = (uint64_t*)src;
uint64_t* out = (uint64_t*)dst;
for (uint64_t i = 0; i < 8; ++i) {
out[i] = in[8 * blk + i];
}
}
// function on layout c is the exact same as on layout b
#ifdef __APPLE__
#pragma weak q120x2_extract_1blk_from_q120c_ref = q120x2_extract_1blk_from_q120b_ref
#else
EXPORT void q120x2_extract_1blk_from_q120c_ref(uint64_t nn, uint64_t blk,
q120x2c* const dst, // 8 doubles
const q120c* const src // a q120c vector
) __attribute__((alias("q120x2_extract_1blk_from_q120b_ref")));
#endif
EXPORT void q120x2_extract_1blk_from_contiguous_q120b_ref(
uint64_t nn, uint64_t nrows, uint64_t blk,
q120x2b* const dst, // nrows * 2 q120
const q120b* const src // a contiguous array of nrows q120b vectors
) {
const uint64_t* in = (uint64_t*)src;
uint64_t* out = (uint64_t*)dst;
for (uint64_t row = 0; row < nrows; ++row) {
for (uint64_t i = 0; i < 8; ++i) {
out[i] = in[8 * blk + i];
}
in += 4 * nn;
out += 8;
}
}
EXPORT void q120x2b_save_1blk_to_q120b_ref(uint64_t nn, uint64_t blk,
q120b* dest, // 1 reim vector of length m
const q120x2b* src // 8 doubles
) {
const uint64_t* in = (uint64_t*)src;
uint64_t* out = (uint64_t*)dest;
for (uint64_t i = 0; i < 8; ++i) {
out[8 * blk + i] = in[i];
}
}

111
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/q120/q120_arithmetic_simple.c
View File

@@ -1,111 +0,0 @@
#include <assert.h>
#include <math.h>
#include <stdint.h>
#include "q120_arithmetic.h"
#include "q120_common.h"
EXPORT void q120_add_bbb_simple(uint64_t nn, q120b* const res, const q120b* const x, const q120b* const y) {
const uint64_t* x_u64 = (uint64_t*)x;
const uint64_t* y_u64 = (uint64_t*)y;
uint64_t* res_u64 = (uint64_t*)res;
for (uint64_t i = 0; i < 4 * nn; i += 4) {
res_u64[i + 0] = x_u64[i + 0] % ((uint64_t)Q1 << 33) + y_u64[i + 0] % ((uint64_t)Q1 << 33);
res_u64[i + 1] = x_u64[i + 1] % ((uint64_t)Q2 << 33) + y_u64[i + 1] % ((uint64_t)Q2 << 33);
res_u64[i + 2] = x_u64[i + 2] % ((uint64_t)Q3 << 33) + y_u64[i + 2] % ((uint64_t)Q3 << 33);
res_u64[i + 3] = x_u64[i + 3] % ((uint64_t)Q4 << 33) + y_u64[i + 3] % ((uint64_t)Q4 << 33);
}
}
EXPORT void q120_add_ccc_simple(uint64_t nn, q120c* const res, const q120c* const x, const q120c* const y) {
const uint32_t* x_u32 = (uint32_t*)x;
const uint32_t* y_u32 = (uint32_t*)y;
uint32_t* res_u32 = (uint32_t*)res;
for (uint64_t i = 0; i < 8 * nn; i += 8) {
res_u32[i + 0] = (uint32_t)(((uint64_t)x_u32[i + 0] + (uint64_t)y_u32[i + 0]) % Q1);
res_u32[i + 1] = (uint32_t)(((uint64_t)x_u32[i + 1] + (uint64_t)y_u32[i + 1]) % Q1);
res_u32[i + 2] = (uint32_t)(((uint64_t)x_u32[i + 2] + (uint64_t)y_u32[i + 2]) % Q2);
res_u32[i + 3] = (uint32_t)(((uint64_t)x_u32[i + 3] + (uint64_t)y_u32[i + 3]) % Q2);
res_u32[i + 4] = (uint32_t)(((uint64_t)x_u32[i + 4] + (uint64_t)y_u32[i + 4]) % Q3);
res_u32[i + 5] = (uint32_t)(((uint64_t)x_u32[i + 5] + (uint64_t)y_u32[i + 5]) % Q3);
res_u32[i + 6] = (uint32_t)(((uint64_t)x_u32[i + 6] + (uint64_t)y_u32[i + 6]) % Q4);
res_u32[i + 7] = (uint32_t)(((uint64_t)x_u32[i + 7] + (uint64_t)y_u32[i + 7]) % Q4);
}
}
EXPORT void q120_c_from_b_simple(uint64_t nn, q120c* const res, const q120b* const x) {
const uint64_t* x_u64 = (uint64_t*)x;
uint32_t* res_u32 = (uint32_t*)res;
for (uint64_t i = 0, j = 0; i < 4 * nn; i += 4, j += 8) {
res_u32[j + 0] = x_u64[i + 0] % Q1;
res_u32[j + 1] = ((uint64_t)res_u32[j + 0] << 32) % Q1;
res_u32[j + 2] = x_u64[i + 1] % Q2;
res_u32[j + 3] = ((uint64_t)res_u32[j + 2] << 32) % Q2;
res_u32[j + 4] = x_u64[i + 2] % Q3;
res_u32[j + 5] = ((uint64_t)res_u32[j + 4] << 32) % Q3;
res_u32[j + 6] = x_u64[i + 3] % Q4;
res_u32[j + 7] = ((uint64_t)res_u32[j + 6] << 32) % Q4;
}
}
EXPORT void q120_b_from_znx64_simple(uint64_t nn, q120b* const res, const int64_t* const x) {
static const int64_t MASK_HI = INT64_C(0x8000000000000000);
static const int64_t MASK_LO = ~MASK_HI;
static const uint64_t OQ[4] = {
(Q1 - (UINT64_C(0x8000000000000000) % Q1)),
(Q2 - (UINT64_C(0x8000000000000000) % Q2)),
(Q3 - (UINT64_C(0x8000000000000000) % Q3)),
(Q4 - (UINT64_C(0x8000000000000000) % Q4)),
};
uint64_t* res_u64 = (uint64_t*)res;
for (uint64_t i = 0, j = 0; j < nn; i += 4, ++j) {
uint64_t xj_lo = x[j] & MASK_LO;
uint64_t xj_hi = x[j] & MASK_HI;
res_u64[i + 0] = xj_lo + (xj_hi ? OQ[0] : 0);
res_u64[i + 1] = xj_lo + (xj_hi ? OQ[1] : 0);
res_u64[i + 2] = xj_lo + (xj_hi ? OQ[2] : 0);
res_u64[i + 3] = xj_lo + (xj_hi ? OQ[3] : 0);
}
}
static int64_t posmod(int64_t x, int64_t q) {
int64_t t = x % q;
if (t < 0)
return t + q;
else
return t;
}
EXPORT void q120_c_from_znx64_simple(uint64_t nn, q120c* const res, const int64_t* const x) {
uint32_t* res_u32 = (uint32_t*)res;
for (uint64_t i = 0, j = 0; j < nn; i += 8, ++j) {
res_u32[i + 0] = posmod(x[j], Q1);
res_u32[i + 1] = ((uint64_t)res_u32[i + 0] << 32) % Q1;
res_u32[i + 2] = posmod(x[j], Q2);
res_u32[i + 3] = ((uint64_t)res_u32[i + 2] << 32) % Q2;
res_u32[i + 4] = posmod(x[j], Q3);
res_u32[i + 5] = ((uint64_t)res_u32[i + 4] << 32) % Q3;
res_u32[i + 6] = posmod(x[j], Q4);
res_u32[i + 7] = ((uint64_t)res_u32[i + 6] << 32) % Q4;
;
}
}
EXPORT void q120_b_to_znx128_simple(uint64_t nn, __int128_t* const res, const q120b* const x) {
static const __int128_t Q = (__int128_t)Q1 * Q2 * Q3 * Q4;
static const __int128_t Qm1 = (__int128_t)Q2 * Q3 * Q4;
static const __int128_t Qm2 = (__int128_t)Q1 * Q3 * Q4;
static const __int128_t Qm3 = (__int128_t)Q1 * Q2 * Q4;
static const __int128_t Qm4 = (__int128_t)Q1 * Q2 * Q3;
const uint64_t* x_u64 = (uint64_t*)x;
for (uint64_t i = 0, j = 0; j < nn; i += 4, ++j) {
__int128_t tmp = 0;
tmp += (((x_u64[i + 0] % Q1) * Q1_CRT_CST) % Q1) * Qm1;
tmp += (((x_u64[i + 1] % Q2) * Q2_CRT_CST) % Q2) * Qm2;
tmp += (((x_u64[i + 2] % Q3) * Q3_CRT_CST) % Q3) * Qm3;
tmp += (((x_u64[i + 3] % Q4) * Q4_CRT_CST) % Q4) * Qm4;
tmp %= Q;
res[j] = (tmp >= (Q + 1) / 2) ? tmp - Q : tmp;
}
}

94
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/q120/q120_common.h
View File

@@ -1,94 +0,0 @@
#ifndef SPQLIOS_Q120_COMMON_H
#define SPQLIOS_Q120_COMMON_H
#include <stdint.h>
#if !defined(SPQLIOS_Q120_USE_29_BIT_PRIMES) && !defined(SPQLIOS_Q120_USE_30_BIT_PRIMES) && \
!defined(SPQLIOS_Q120_USE_31_BIT_PRIMES)
#define SPQLIOS_Q120_USE_30_BIT_PRIMES
#endif
/**
* 29-bit primes and 2*2^16 roots of unity
*/
#ifdef SPQLIOS_Q120_USE_29_BIT_PRIMES
#define Q1 ((1u << 29) - 2 * (1u << 17) + 1)
#define OMEGA1 78289835
#define Q1_CRT_CST 301701286 // (Q2*Q3*Q4)^-1 mod Q1
#define Q2 ((1u << 29) - 5 * (1u << 17) + 1)
#define OMEGA2 178519192
#define Q2_CRT_CST 536020447 // (Q1*Q3*Q4)^-1 mod Q2
#define Q3 ((1u << 29) - 26 * (1u << 17) + 1)
#define OMEGA3 483889678
#define Q3_CRT_CST 86367873 // (Q1*Q2*Q4)^-1 mod Q3
#define Q4 ((1u << 29) - 35 * (1u << 17) + 1)
#define OMEGA4 239808033
#define Q4_CRT_CST 147030781 // (Q1*Q2*Q3)^-1 mod Q4
#endif
/**
* 30-bit primes and 2*2^16 roots of unity
*/
#ifdef SPQLIOS_Q120_USE_30_BIT_PRIMES
#define Q1 ((1u << 30) - 2 * (1u << 17) + 1)
#define OMEGA1 1070907127
#define Q1_CRT_CST 43599465 // (Q2*Q3*Q4)^-1 mod Q1
#define Q2 ((1u << 30) - 17 * (1u << 17) + 1)
#define OMEGA2 315046632
#define Q2_CRT_CST 292938863 // (Q1*Q3*Q4)^-1 mod Q2
#define Q3 ((1u << 30) - 23 * (1u << 17) + 1)
#define OMEGA3 309185662
#define Q3_CRT_CST 594011630 // (Q1*Q2*Q4)^-1 mod Q3
#define Q4 ((1u << 30) - 42 * (1u << 17) + 1)
#define OMEGA4 846468380
#define Q4_CRT_CST 140177212 // (Q1*Q2*Q3)^-1 mod Q4
#endif
/**
* 31-bit primes and 2*2^16 roots of unity
*/
#ifdef SPQLIOS_Q120_USE_31_BIT_PRIMES
#define Q1 ((1u << 31) - 1 * (1u << 17) + 1)
#define OMEGA1 1615402923
#define Q1_CRT_CST 1811422063 // (Q2*Q3*Q4)^-1 mod Q1
#define Q2 ((1u << 31) - 4 * (1u << 17) + 1)
#define OMEGA2 1137738560
#define Q2_CRT_CST 2093150204 // (Q1*Q3*Q4)^-1 mod Q2
#define Q3 ((1u << 31) - 11 * (1u << 17) + 1)
#define OMEGA3 154880552
#define Q3_CRT_CST 164149010 // (Q1*Q2*Q4)^-1 mod Q3
#define Q4 ((1u << 31) - 23 * (1u << 17) + 1)
#define OMEGA4 558784885
#define Q4_CRT_CST 225197446 // (Q1*Q2*Q3)^-1 mod Q4
#endif
static const uint32_t PRIMES_VEC[4] = {Q1, Q2, Q3, Q4};
static const uint32_t OMEGAS_VEC[4] = {OMEGA1, OMEGA2, OMEGA3, OMEGA4};
#define MAX_ELL 10000
// each number x mod Q120 is represented by uint64_t[4] with (non-unique) values (x mod q1, x mod q2,x mod q3,x mod q4),
// each between [0 and 2^32-1]
typedef struct _q120a q120a;
// each number x mod Q120 is represented by uint64_t[4] with (non-unique) values (x mod q1, x mod q2,x mod q3,x mod q4),
// each between [0 and 2^64-1]
typedef struct _q120b q120b;
// each number x mod Q120 is represented by uint32_t[8] with values (x mod q1, 2^32x mod q1, x mod q2, 2^32.x mod q2, x
// mod q3, 2^32.x mod q3, x mod q4, 2^32.x mod q4) each between [0 and 2^32-1]
typedef struct _q120c q120c;
typedef struct _q120x2b q120x2b;
typedef struct _q120x2c q120x2c;
#endif // SPQLIOS_Q120_COMMON_H

5
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/q120/q120_fallbacks_aarch64.c
View File

@@ -1,5 +0,0 @@
#include "q120_ntt_private.h"
EXPORT void q120_ntt_bb_avx2(const q120_ntt_precomp* const precomp, q120b* const data) { UNDEFINED(); }
EXPORT void q120_intt_bb_avx2(const q120_ntt_precomp* const precomp, q120b* const data) { UNDEFINED(); }

340
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/q120/q120_ntt.c
View File

@@ -1,340 +0,0 @@
#include <assert.h>
#include <inttypes.h>
#include <math.h>
#include <stdint.h>
#include "q120_ntt_private.h"
q120_ntt_precomp* new_precomp(const uint64_t n) {
q120_ntt_precomp* precomp = malloc(sizeof(*precomp));
precomp->n = n;
assert(n && !(n & (n - 1)) && n <= (1 << 16)); // n is a power of 2 smaller than 2^16
const uint64_t logN = ceil(log2(n));
precomp->level_metadata = malloc((logN + 2) * sizeof(*precomp->level_metadata));
precomp->powomega = spqlios_alloc_custom_align(32, 4 * 2 * n * sizeof(*(precomp->powomega)));
return precomp;
}
uint32_t modq_pow(const uint32_t x, const int64_t n, const uint32_t q) {
uint64_t np = (n % (q - 1) + q - 1) % (q - 1);
uint64_t val_pow = x;
uint64_t res = 1;
while (np != 0) {
if (np & 1) res = (res * val_pow) % q;
val_pow = (val_pow * val_pow) % q;
np >>= 1;
}
return res;
}
void fill_omegas(const uint64_t n, uint32_t omegas[4]) {
for (uint64_t k = 0; k < 4; ++k) {
omegas[k] = modq_pow(OMEGAS_VEC[k], (1 << 16) / n, PRIMES_VEC[k]);
}
#ifndef NDEBUG
const uint64_t logQ = ceil(log2(PRIMES_VEC[0]));
for (int k = 1; k < 4; ++k) {
if (logQ != ceil(log2(PRIMES_VEC[k]))) {
fprintf(stderr, "The 4 primes must have the same bit-size\n");
exit(-1);
}
}
// check if each omega is a 2.n primitive root of unity
for (uint64_t k = 0; k < 4; ++k) {
assert(modq_pow(omegas[k], 2 * n, PRIMES_VEC[k]) == 1);
for (uint64_t i = 1; i < 2 * n; ++i) {
assert(modq_pow(omegas[k], i, PRIMES_VEC[k]) != 1);
}
}
if (logQ > 31) {
fprintf(stderr, "Modulus q bit-size is larger than 30 bit\n");
exit(-1);
}
#endif
}
uint64_t fill_reduction_meta(const uint64_t bs_start, q120_ntt_reduc_step_precomp* reduc_metadata) {
// fill reduction metadata
uint64_t bs_after_reduc = -1;
{
uint64_t min_h = -1;
for (uint64_t h = bs_start / 2; h < bs_start; ++h) {
uint64_t t = 0;
for (uint64_t k = 0; k < 4; ++k) {
const uint64_t t1 = bs_start - h + (uint64_t)ceil(log2((UINT64_C(1) << h) % PRIMES_VEC[k]));
const uint64_t t2 = UINT64_C(1) + ((t1 > h) ? t1 : h);
if (t < t2) t = t2;
}
if (t < bs_after_reduc) {
min_h = h;
bs_after_reduc = t;
}
}
reduc_metadata->h = min_h;
reduc_metadata->mask = (UINT64_C(1) << min_h) - 1;
for (uint64_t k = 0; k < 4; ++k) {
reduc_metadata->modulo_red_cst[k] = (UINT64_C(1) << min_h) % PRIMES_VEC[k];
}
assert(bs_after_reduc < 64);
}
return bs_after_reduc;
}
uint64_t round_up_half_n(const uint64_t n) { return (n + 1) / 2; }
EXPORT q120_ntt_precomp* q120_new_ntt_bb_precomp(const uint64_t n) {
uint32_t omega_vec[4];
fill_omegas(n, omega_vec);
const uint64_t logQ = ceil(log2(PRIMES_VEC[0]));
q120_ntt_precomp* precomp = new_precomp(n);
uint64_t bs = precomp->input_bit_size = 64;
LOG("NTT parameters:\n");
LOG("\tsize = %" PRIu64 "\n", n)
LOG("\tlogQ = %" PRIu64 "\n", logQ);
LOG("\tinput bit-size = %" PRIu64 "\n", bs);
if (n == 1) return precomp;
// fill reduction metadata
uint64_t bs_after_reduc = fill_reduction_meta(bs, &(precomp->reduc_metadata));
// forward metadata
q120_ntt_step_precomp* level_metadata_ptr = precomp->level_metadata;
// first level a_k.omega^k
{
const uint64_t half_bs = (bs + 1) / 2;
level_metadata_ptr->half_bs = half_bs;
level_metadata_ptr->mask = (UINT64_C(1) << half_bs) - UINT64_C(1);
level_metadata_ptr->bs = bs = half_bs + logQ + 1;
LOG("\tlevel %6" PRIu64 " output bit-size = %" PRIu64 " (a_k.omega^k) \n", n, bs);
level_metadata_ptr++;
}
for (uint64_t nn = n; nn >= 4; nn /= 2) {
level_metadata_ptr->reduce = (bs == 64);
if (level_metadata_ptr->reduce) {
bs = bs_after_reduc;
LOG("\treduce output bit-size = %" PRIu64 "\n", bs);
}
for (int k = 0; k < 4; ++k) level_metadata_ptr->q2bs[k] = (uint64_t)PRIMES_VEC[k] << (bs - logQ);
double bs_1 = bs + 1; // bit-size of term a+b or a-b
const uint64_t half_bs = round_up_half_n(bs_1);
uint64_t bs_2 = half_bs + logQ + 1; // bit-size of term (a-b).omega^k
bs = (bs_1 > bs_2) ? bs_1 : bs_2;
assert(bs <= 64);
level_metadata_ptr->bs = bs;
level_metadata_ptr->half_bs = half_bs;
level_metadata_ptr->mask = (UINT64_C(1) << half_bs) - UINT64_C(1);
level_metadata_ptr++;
LOG("\tlevel %6" PRIu64 " output bit-size = %" PRIu64 "\n", nn / 2, bs);
}
// last level (a-b, a+b)
{
level_metadata_ptr->reduce = (bs == 64);
if (level_metadata_ptr->reduce) {
bs = bs_after_reduc;
LOG("\treduce output bit-size = %" PRIu64 "\n", bs);
}
for (int k = 0; k < 4; ++k) level_metadata_ptr->q2bs[k] = ((uint64_t)PRIMES_VEC[k] << (bs - logQ));
level_metadata_ptr->bs = ++bs;
level_metadata_ptr->half_bs = level_metadata_ptr->mask = UINT64_C(0); // not used
LOG("\tlevel %6" PRIu64 " output bit-size = %" PRIu64 "\n", UINT64_C(1), bs);
}
precomp->output_bit_size = bs;
// omega powers
uint64_t* powomega = malloc(sizeof(*powomega) * 2 * n);
for (uint64_t k = 0; k < 4; ++k) {
const uint64_t q = PRIMES_VEC[k];
for (uint64_t i = 0; i < 2 * n; ++i) {
powomega[i] = modq_pow(omega_vec[k], i, q);
}
uint64_t* powomega_ptr = precomp->powomega + k;
level_metadata_ptr = precomp->level_metadata;
{
// const uint64_t hpow = UINT64_C(1) << level_metadata_ptr->half_bs;
for (uint64_t i = 0; i < n; ++i) {
uint64_t t = powomega[i];
uint64_t t1 = (t << level_metadata_ptr->half_bs) % q;
powomega_ptr[4 * i] = (t1 << 32) + t;
}
powomega_ptr += 4 * n;
level_metadata_ptr++;
}
for (uint64_t nn = n; nn >= 4; nn /= 2) {
const uint64_t halfnn = nn / 2;
const uint64_t m = n / halfnn;
// const uint64_t hpow = UINT64_C(1) << level_metadata_ptr->half_bs;
for (uint64_t i = 1; i < halfnn; ++i) {
uint64_t t = powomega[i * m];
uint64_t t1 = (t << level_metadata_ptr->half_bs) % q;
powomega_ptr[4 * (i - 1)] = (t1 << 32) + t;
}
powomega_ptr += 4 * (halfnn - 1);
level_metadata_ptr++;
}
}
free(powomega);
return precomp;
}
EXPORT q120_ntt_precomp* q120_new_intt_bb_precomp(const uint64_t n) {
uint32_t omega_vec[4];
fill_omegas(n, omega_vec);
const uint64_t logQ = ceil(log2(PRIMES_VEC[0]));
q120_ntt_precomp* precomp = new_precomp(n);
uint64_t bs = precomp->input_bit_size = 64;
LOG("iNTT parameters:\n");
LOG("\tsize = %" PRIu64 "\n", n)
LOG("\tlogQ = %" PRIu64 "\n", logQ);
LOG("\tinput bit-size = %" PRIu64 "\n", bs);
if (n == 1) return precomp;
// fill reduction metadata
uint64_t bs_after_reduc = fill_reduction_meta(bs, &(precomp->reduc_metadata));
// backward metadata
q120_ntt_step_precomp* level_metadata_ptr = precomp->level_metadata;
// first level (a+b, a-b) adds 1-bit
{
level_metadata_ptr->reduce = (bs == 64);
if (level_metadata_ptr->reduce) {
bs = bs_after_reduc;
LOG("\treduce output bit-size = %" PRIu64 "\n", bs);
}
for (int k = 0; k < 4; ++k) level_metadata_ptr->q2bs[k] = (uint64_t)PRIMES_VEC[k] << (bs - logQ);
level_metadata_ptr->bs = ++bs;
level_metadata_ptr->half_bs = level_metadata_ptr->mask = UINT64_C(0); // not used
LOG("\tlevel %6" PRIu64 " output bit-size = %" PRIu64 "\n", UINT64_C(1), bs);
level_metadata_ptr++;
}
for (uint64_t nn = 4; nn <= n; nn *= 2) {
level_metadata_ptr->reduce = (bs == 64);
if (level_metadata_ptr->reduce) {
bs = bs_after_reduc;
LOG("\treduce output bit-size = %" PRIu64 "\n", bs);
}
const uint64_t half_bs = round_up_half_n(bs);
const uint64_t bs_mult = half_bs + logQ + 1; // bit-size of term b.omega^k
bs = 1 + ((bs > bs_mult) ? bs : bs_mult); // bit-size of a+b.omega^k or a-b.omega^k
assert(bs <= 64);
level_metadata_ptr->bs = bs;
level_metadata_ptr->half_bs = half_bs;
level_metadata_ptr->mask = (UINT64_C(1) << half_bs) - UINT64_C(1);
for (int k = 0; k < 4; ++k) level_metadata_ptr->q2bs[k] = (uint64_t)PRIMES_VEC[k] << (bs_mult - logQ);
level_metadata_ptr++;
LOG("\tlevel %6" PRIu64 " output bit-size = %" PRIu64 "\n", nn / 2, bs);
}
// last level a_k.omega^k
{
level_metadata_ptr->reduce = (bs == 64);
if (level_metadata_ptr->reduce) {
bs = bs_after_reduc;
LOG("\treduce output bit-size = %" PRIu64 "\n", bs);
}
const uint64_t half_bs = round_up_half_n(bs);
bs = half_bs + logQ + 1; // bit-size of term a.omega^k
assert(bs <= 64);
level_metadata_ptr->bs = bs;
level_metadata_ptr->half_bs = half_bs;
level_metadata_ptr->mask = (UINT64_C(1) << half_bs) - UINT64_C(1);
for (int k = 0; k < 4; ++k) level_metadata_ptr->q2bs[k] = (uint64_t)PRIMES_VEC[k] << (bs - logQ);
LOG("\tlevel %6" PRIu64 " output bit-size = %" PRIu64 "\n", n, bs);
}
// omega powers
uint32_t* powomegabar = malloc(sizeof(*powomegabar) * 2 * n);
for (int k = 0; k < 4; ++k) {
const uint64_t q = PRIMES_VEC[k];
for (uint64_t i = 0; i < 2 * n; ++i) {
powomegabar[i] = modq_pow(omega_vec[k], -i, q);
}
uint64_t* powomega_ptr = precomp->powomega + k;
level_metadata_ptr = precomp->level_metadata + 1;
for (uint64_t nn = 4; nn <= n; nn *= 2) {
const uint64_t halfnn = nn / 2;
const uint64_t m = n / halfnn;
for (uint64_t i = 1; i < halfnn; ++i) {
uint64_t t = powomegabar[i * m];
uint64_t t1 = (t << level_metadata_ptr->half_bs) % q;
powomega_ptr[4 * (i - 1)] = (t1 << 32) + t;
}
powomega_ptr += 4 * (halfnn - 1);
level_metadata_ptr++;
}
{
const uint64_t invNmod = modq_pow(n, -1, q);
for (uint64_t i = 0; i < n; ++i) {
uint64_t t = (powomegabar[i] * invNmod) % q;
uint64_t t1 = (t << level_metadata_ptr->half_bs) % q;
powomega_ptr[4 * i] = (t1 << 32) + t;
}
}
}
free(powomegabar);
return precomp;
}
void del_precomp(q120_ntt_precomp* precomp) {
spqlios_free(precomp->powomega);
free(precomp->level_metadata);
free(precomp);
}
EXPORT void q120_del_ntt_bb_precomp(q120_ntt_precomp* precomp) { del_precomp(precomp); }
EXPORT void q120_del_intt_bb_precomp(q120_ntt_precomp* precomp) { del_precomp(precomp); }

25
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/q120/q120_ntt.h
View File

@@ -1,25 +0,0 @@
#ifndef SPQLIOS_Q120_NTT_H
#define SPQLIOS_Q120_NTT_H
#include "../commons.h"
#include "q120_common.h"
typedef struct _q120_ntt_precomp q120_ntt_precomp;
EXPORT q120_ntt_precomp* q120_new_ntt_bb_precomp(const uint64_t n);
EXPORT void q120_del_ntt_bb_precomp(q120_ntt_precomp* precomp);
EXPORT q120_ntt_precomp* q120_new_intt_bb_precomp(const uint64_t n);
EXPORT void q120_del_intt_bb_precomp(q120_ntt_precomp* precomp);
/**
* @brief computes a direct ntt in-place over data.
*/
EXPORT void q120_ntt_bb_avx2(const q120_ntt_precomp* const precomp, q120b* const data);
/**
* @brief computes an inverse ntt in-place over data.
*/
EXPORT void q120_intt_bb_avx2(const q120_ntt_precomp* const precomp, q120b* const data);
#endif // SPQLIOS_Q120_NTT_H

479
poulpy-backend/src/cpu_spqlios/spqlios-arithmetic/spqlios/q120/q120_ntt_avx2.c
View File

@@ -1,479 +0,0 @@
#include <assert.h>
#include <immintrin.h>
#include <inttypes.h>
#include <math.h>
#include <stdint.h>
#include <stdio.h>
#include "q120_common.h"
#include "q120_ntt_private.h"
// at which level to switch from computations by level to computations by block
#define CHANGE_MODE_N 1024
__always_inline __m256i split_precompmul_si256(__m256i inp, __m256i powomega, const uint64_t h, const __m256i mask) {
const __m256i inp_low = _mm256_and_si256(inp, mask);
const __m256i t1 = _mm256_mul_epu32(inp_low, powomega);
const __m256i inp_high = _mm256_srli_epi64(inp, h);
const __m256i powomega_high = _mm256_srli_epi64(powomega, 32);
const __m256i t2 = _mm256_mul_epu32(inp_high, powomega_high);
return _mm256_add_epi64(t1, t2);
}
__always_inline __m256i modq_red(const __m256i x, const uint64_t h, const __m256i mask, const __m256i _2_pow_h_modq) {
__m256i xh = _mm256_srli_epi64(x, h);
__m256i xl = _mm256_and_si256(x, mask);
__m256i xh_1 = _mm256_mul_epu32(xh, _2_pow_h_modq);
return _mm256_add_epi64(xl, xh_1);
}
void print_data(const uint64_t n, const uint64_t* const data, const uint64_t q) {
for (uint64_t i = 0; i < n; i++) {
printf("%" PRIu64 " ", *(data + i) % q);
}
printf("\n");
}
double max_bit_size(const void* const begin, const void* const end) {
double bs = 0;
const uint64_t* data = (uint64_t*)begin;
while (data != end) {
double t = log2(*(data++));
if (bs < t) {
bs = t;
}
}
return bs;
}
void ntt_iter_first(__m256i* const begin, const __m256i* const end, const q120_ntt_step_precomp* const itData,
const __m256i* powomega) {
const uint64_t h = itData->half_bs;
const __m256i vmask = _mm256_set1_epi64x(itData->mask);
__m256i* data = begin;
while (data < end) {
__m256i x = _mm256_loadu_si256(data);
__m256i po = _mm256_loadu_si256(powomega);
__m256i r = split_precompmul_si256(x, po, h, vmask);
_mm256_storeu_si256(data, r);
data++;
powomega++;
}
}
void ntt_iter(const uint64_t nn, __m256i* const begin, const __m256i* const end,
const q120_ntt_step_precomp* const itData, const __m256i* const powomega) {
assert(nn % 2 == 0);
const uint64_t halfnn = nn / 2;
const __m256i vq2bs = _mm256_loadu_si256((__m256i*)itData->q2bs);
const __m256i vmask = _mm256_set1_epi64x(itData->mask);
__m256i* data = begin;
while (data < end) {
__m256i* ptr1 = data;
__m256i* ptr2 = data + halfnn;
const __m256i a = _mm256_loadu_si256(ptr1);
const __m256i b = _mm256_loadu_si256(ptr2);
const __m256i ap = _mm256_add_epi64(a, b);
_mm256_storeu_si256(ptr1, ap);
const __m256i bp = _mm256_sub_epi64(_mm256_add_epi64(a, vq2bs), b);
_mm256_storeu_si256(ptr2, bp);
ptr1++;
ptr2++;
const __m256i* po_ptr = powomega;
for (uint64_t i = 1; i < halfnn; ++i) {
__m256i a = _mm256_loadu_si256(ptr1);
__m256i b = _mm256_loadu_si256(ptr2);
__m256i ap = _mm256_add_epi64(a, b);
_mm256_storeu_si256(ptr1, ap);
__m256i b1 = _mm256_sub_epi64(_mm256_add_epi64(a, vq2bs), b);
__m256i po = _mm256_loadu_si256(po_ptr);
__m256i bp = split_precompmul_si256(b1, po, itData->half_bs, vmask);
_mm256_storeu_si256(ptr2, bp);
ptr1++;
ptr2++;
po_ptr++;
}
data += nn;
}
}
void ntt_iter_red(const uint64_t nn, __m256i* const begin, const __m256i* const end,
const q120_ntt_step_precomp* const itData, const __m256i* const powomega,
const q120_ntt_reduc_step_precomp* const reduc_precomp) {
assert(nn % 2 == 0);
const uint64_t halfnn = nn / 2;
const __m256i vq2bs = _mm256_loadu_si256((__m256i*)itData->q2bs);
const __m256i vmask = _mm256_set1_epi64x(itData->mask);
const __m256i reduc_mask = _mm256_set1_epi64x(reduc_precomp->mask);
const __m256i reduc_cst = _mm256_loadu_si256((__m256i*)reduc_precomp->modulo_red_cst);
__m256i* data = begin;
while (data < end) {
__m256i* ptr1 = data;
__m256i* ptr2 = data + halfnn;
__m256i a = _mm256_loadu_si256(ptr1);
__m256i b = _mm256_loadu_si256(ptr2);
a = modq_red(a, reduc_precomp->h, reduc_mask, reduc_cst);
b = modq_red(b, reduc_precomp->h, reduc_mask, reduc_cst);
const __m256i ap = _mm256_add_epi64(a, b);
_mm256_storeu_si256(ptr1, ap);
const __m256i bp = _mm256_sub_epi64(_mm256_add_epi64(a, vq2bs), b);
_mm256_storeu_si256(ptr2, bp);
ptr1++;
ptr2++;
const __m256i* po_ptr = powomega;
for (uint64_t i = 1; i < halfnn; ++i) {
__m256i a = _mm256_loadu_si256(ptr1);
__m256i b = _mm256_loadu_si256(ptr2);
a = modq_red(a, reduc_precomp->h, reduc_mask, reduc_cst);
b = modq_red(b, reduc_precomp->h, reduc_mask, reduc_cst);
__m256i ap = _mm256_add_epi64(a, b);
_mm256_storeu_si256(ptr1, ap);
__m256i bp = _mm256_sub_epi64(_mm256_add_epi64(a, vq2bs), b);
__m256i po = _mm256_loadu_si256(po_ptr);
bp = split_precompmul_si256(bp, po, itData->half_bs, vmask);
_mm256_storeu_si256(ptr2, bp);
ptr1++;
ptr2++;
po_ptr++;
}
data += nn;
}
}
EXPORT void q120_ntt_bb_avx2(const q120_ntt_precomp* const precomp, q120b* const data_ptr) {
// assert((size_t)data_ptr % 32 == 0); // alignment check
const uint64_t n = precomp->n;
if (n == 1) return;
const q120_ntt_step_precomp* itData = precomp->level_metadata;
const __m256i* powomega = (__m256i*)precomp->powomega;
__m256i* const begin = (__m256i*)data_ptr;
const __m256i* const end = ((__m256i*)data_ptr) + n;
if (CHECK_BOUNDS) {
double bs __attribute__((unused)) = max_bit_size((void*)begin, (void*)end);
LOG("Input %lf %" PRIu64 "\n", bs, precomp->input_bit_size);
assert(bs <= precomp->input_bit_size);
}
// first iteration a_k.omega^k
ntt_iter_first(begin, end, itData, powomega);
if (CHECK_BOUNDS) {
double bs __attribute__((unused)) = max_bit_size((void*)begin, (void*)end);
LOG("Iter %3" PRIu64 " - %lf %" PRIu64 "\n", n, bs, itData->bs);
assert(bs < itData->bs);
}
powomega += n;
itData++;
const uint64_t split_nn = (CHANGE_MODE_N > n) ? n : CHANGE_MODE_N;
// const uint64_t split_nn = 2;
// computations by level
for (uint64_t nn = n; nn > split_nn; nn /= 2) {
const uint64_t halfnn = nn / 2;
if (itData->reduce) {
ntt_iter_red(nn, begin, end, itData, powomega, &precomp->reduc_metadata);
} else {
ntt_iter(nn, begin, end, itData, powomega);
}
if (CHECK_BOUNDS) {
double bs __attribute__((unused)) = max_bit_size((void*)begin, (void*)end);
LOG("Iter %3" PRIu64 " - %lf %" PRIu64 " %c\n", nn / 2, bs, itData->bs, itData->reduce ? '*' : ' ');
assert(bs < itData->bs);
}
powomega += halfnn - 1;
itData++;
}
// computations by memory block
if (split_nn >= 2) {
const q120_ntt_step_precomp* itData1 = itData;
const __m256i* powomega1 = powomega;
for (__m256i* it = begin; it < end; it += split_nn) {
__m256i* const begin1 = it;
const __m256i* const end1 = it + split_nn;
itData = itData1;
powomega = powomega1;
for (uint64_t nn = split_nn; nn >= 2; nn /= 2) {
const uint64_t halfnn = nn / 2;
if (itData->reduce) {
ntt_iter_red(nn, begin1, end1, itData, powomega, &precomp->reduc_metadata);
} else {
ntt_iter(nn, begin1, end1, itData, powomega);
}
if (CHECK_BOUNDS) {
double bs __attribute__((unused)) = max_bit_size((uint64_t*)begin1, (uint64_t*)end1);
// LOG("Iter %3lu - %lf %lu\n", nn / 2, bs, itData->bs);
assert(bs < itData->bs);
}
powomega += halfnn - 1;
itData++;
}
}
}
if (CHECK_BOUNDS) {
double bs __attribute__((unused)) = max_bit_size((void*)begin, (void*)end);
LOG("Iter %3" PRIu64 " - %lf %" PRIu64 "\n", UINT64_C(1), bs, precomp->output_bit_size);
assert(bs < precomp->output_bit_size);
}
}
void intt_iter(const uint64_t nn, __m256i* const begin, const __m256i* const end,
const q120_ntt_step_precomp* const itData, const __m256i* const powomega) {
assert(nn % 2 == 0);
const uint64_t halfnn = nn / 2;
const __m256i vq2bs = _mm256_loadu_si256((__m256i*)itData->q2bs);
const __m256i vmask = _mm256_set1_epi64x(itData->mask);
__m256i* data = begin;
while (data < end) {
__m256i* ptr1 = data;
__m256i* ptr2 = data + halfnn;
const __m256i a = _mm256_loadu_si256(ptr1);
const __m256i b = _mm256_loadu_si256(ptr2);
const __m256i ap = _mm256_add_epi64(a, b);
_mm256_storeu_si256(ptr1, ap);
const __m256i bp = _mm256_sub_epi64(_mm256_add_epi64(a, vq2bs), b);
_mm256_storeu_si256(ptr2, bp);
ptr1++;
ptr2++;
const __m256i* po_ptr = powomega;
for (uint64_t i = 1; i < halfnn; ++i) {
__m256i a = _mm256_loadu_si256(ptr1);
__m256i b = _mm256_loadu_si256(ptr2);
__m256i po = _mm256_loadu_si256(po_ptr);
__m256i bo = split_precompmul_si256(b, po, itData->half_bs, vmask);
__m256i ap = _mm256_add_epi64(a, bo);
_mm256_storeu_si256(ptr1, ap);
__m256i bp = _mm256_sub_epi64(_mm256_add_epi64(a, vq2bs), bo);
_mm256_storeu_si256(ptr2, bp);
ptr1++;
ptr2++;
po_ptr++;
}
data += nn;
}
}
void intt_iter_red(const uint64_t nn, __m256i* const begin, const __m256i* const end,
const q120_ntt_step_precomp* const itData, const __m256i* const powomega,
const q120_ntt_reduc_step_precomp* const reduc_precomp) {
assert(nn % 2 == 0);
const uint64_t halfnn = nn / 2;
const __m256i vq2bs = _mm256_loadu_si256((__m256i*)itData->q2bs);
const __m256i vmask = _mm256_set1_epi64x(itData->mask);
const __m256i reduc_mask = _mm256_set1_epi64x(reduc_precomp->mask);
const __m256i reduc_cst = _mm256_loadu_si256((__m256i*)reduc_precomp->modulo_red_cst);
__m256i* data = begin;
while (data < end) {
__m256i* ptr1 = data;
__m256i* ptr2 = data + halfnn;
__m256i a = _mm256_loadu_si256(ptr1);
__m256i b = _mm256_loadu_si256(ptr2);
a = modq_red(a, reduc_precomp->h, reduc_mask, reduc_cst);
b = modq_red(b, reduc_precomp->h, reduc_mask, reduc_cst);
const __m256i ap = _mm256_add_epi64(a, b);
_mm256_storeu_si256(ptr1, ap);
const __m256i bp = _mm256_sub_epi64(_mm256_add_epi64(a, vq2bs), b);
_mm256_storeu_si256(ptr2, bp);
ptr1++;
ptr2++;
const __m256i* po_ptr = powomega;
for (uint64_t i = 1; i < halfnn; ++i) {
__m256i a = _mm256_loadu_si256(ptr1);
__m256i b = _mm256_loadu_si256(ptr2);
a = modq_red(a, reduc_precomp->h, reduc_mask, reduc_cst);
b = modq_red(b, reduc_precomp->h, reduc_mask, reduc_cst);
__m256i po = _mm256_loadu_si256(po_ptr);
__m256i bo = split_precompmul_si256(b, po, itData->half_bs, vmask);
__m256i ap = _mm256_add_epi64(a, bo);
_mm256_storeu_si256(ptr1, ap);
__m256i bp = _mm256_sub_epi64(_mm256_add_epi64(a, vq2bs), bo);
_mm256_storeu_si256(ptr2, bp);
ptr1++;
ptr2++;
po_ptr++;
}
data += nn;
}
}
void ntt_iter_first_red(__m256i* const begin, const __m256i* const end, const q120_ntt_step_precomp* const itData,
const __m256i* powomega, const q120_ntt_reduc_step_precomp* const reduc_precomp) {
const uint64_t h = itData->half_bs;
const __m256i vmask = _mm256_set1_epi64x(itData->mask);
const __m256i reduc_mask = _mm256_set1_epi64x(reduc_precomp->mask);
const __m256i reduc_cst = _mm256_loadu_si256((__m256i*)reduc_precomp->modulo_red_cst);
__m256i* data = begin;
while (data < end) {
__m256i x = _mm256_loadu_si256(data);
x = modq_red(x, reduc_precomp->h, reduc_mask, reduc_cst);
__m256i po = _mm256_loadu_si256(powomega);
__m256i r = split_precompmul_si256(x, po, h, vmask);
_mm256_storeu_si256(data, r);
data++;
powomega++;
}
}
EXPORT void q120_intt_bb_avx2(const q120_ntt_precomp* const precomp, q120b* const data_ptr) {
// assert((size_t)data_ptr % 32 == 0); // alignment check
const uint64_t n = precomp->n;
if (n == 1) return;
const q120_ntt_step_precomp* itData = precomp->level_metadata;
const __m256i* powomega = (__m256i*)precomp->powomega;
__m256i* const begin = (__m256i*)data_ptr;
const __m256i* const end = ((__m256i*)data_ptr) + n;
if (CHECK_BOUNDS) {
double bs __attribute__((unused)) = max_bit_size((void*)begin, (void*)end);
LOG("Input %lf %" PRIu64 "\n", bs, precomp->input_bit_size);
assert(bs <= precomp->input_bit_size);
}
const uint64_t split_nn = (CHANGE_MODE_N > n) ? n : CHANGE_MODE_N;
// computations by memory block
if (split_nn >= 2) {
const q120_ntt_step_precomp* itData1 = itData;
const __m256i* powomega1 = powomega;
for (__m256i* it = begin; it < end; it += split_nn) {
__m256i* const begin1 = it;
const __m256i* const end1 = it + split_nn;
itData = itData1;
powomega = powomega1;
for (uint64_t nn = 2; nn <= split_nn; nn *= 2) {
const uint64_t halfnn = nn / 2;
if (itData->reduce) {
intt_iter_red(nn, begin1, end1, itData, powomega, &precomp->reduc_metadata);
} else {
intt_iter(nn, begin1, end1, itData, powomega);
}
if (CHECK_BOUNDS) {
double bs __attribute__((unused)) = max_bit_size((uint64_t*)begin1, (uint64_t*)end1);
// LOG("Iter %3lu - %lf %lu\n", nn / 2, bs, itData->bs);
assert(bs < itData->bs);
}
powomega += halfnn - 1;
itData++;
}
}
}
// computations by level
// for (uint64_t nn = 2; nn <= n; nn *= 2) {
for (uint64_t nn = 2 * split_nn; nn <= n; nn *= 2) {
const uint64_t halfnn = nn / 2;
if (itData->reduce) {
intt_iter_red(nn, begin, end, itData, powomega, &precomp->reduc_metadata);
} else {
intt_iter(nn, begin, end, itData, powomega);
}
if (CHECK_BOUNDS) {
double bs __attribute__((unused)) = max_bit_size((void*)begin, (void*)end);
LOG("Iter %3" PRIu64 " - %lf %" PRIu64 " %c\n", nn / 2, bs, itData->bs, itData->reduce ? '*' : ' ');
assert(bs < itData->bs);
}
powomega += halfnn - 1;
itData++;
}
// last iteration a_k . omega^k . n^-1
if (itData->reduce) {
ntt_iter_first_red(begin, end, itData, powomega, &precomp->reduc_metadata);
} else {
ntt_iter_first(begin, end, itData, powomega);
}
if (CHECK_BOUNDS) {
double bs __attribute__((unused)) = max_bit_size((void*)begin, (void*)end);
LOG("Iter %3" PRIu64 " - %lf %" PRIu64 "\n", n, bs, itData->bs);
assert(bs < itData->bs);
}
}

Some files were not shown because too many files have changed in this diff Show More