From 041dde839290035978f2df85f82d2f0845745bbd Mon Sep 17 00:00:00 2001
From: nindanaoto <matsuoka.kotaro@gmail.com>
Date: Fri, 12 Jul 2024 17:11:45 +0000
Subject: [PATCH] fixed README and removed alignas from param.hpp

---
 README.md          | 9 +++++++++
 include/mulfft.hpp | 6 ++++++
 include/params.hpp | 4 ++--
 3 files changed, 17 insertions(+), 2 deletions(-)

diff --git a/README.md b/README.md
index 345a479..fc2f198 100644
--- a/README.md
+++ b/README.md
@@ -41,6 +41,15 @@ We assume to install MKL by [this procedure](https://www.intel.com/content/www/u
 
 Add `-DUSE_MKL` to the CMake option to use MKL
 
+### FFTW3 API
+To use this, you have to also add `-DUSE_FFTW3`.
+
+### Native API
+Instead of FFTW3 API, I also added native API version. This will be enabled if `-DUSE_FFTW3` is not specified with `-DUSE_MKL`.
+
+## concrete-fft
+concrete-fft is the pure rust FFT library developed by Zama.ai. This can be enabled by `-DUSE_CONCRETE_FFT`.
+
 ## spqliox_aarch64
 spqliox_aarch64 is the FFT library for aarch64 forked from SPQLIOS.
 This is slightly faster than FFTW3(average 1ms).
diff --git a/include/mulfft.hpp b/include/mulfft.hpp
index c9c0235..b2cc84e 100644
--- a/include/mulfft.hpp
+++ b/include/mulfft.hpp
@@ -149,6 +149,9 @@ template <uint32_t N>
 inline void MulInFD(std::array<double, N> &res, const std::array<double, N> &a,
                     const std::array<double, N> &b)
 {
+    std::assume_aligned<16>(res.data());
+    std::assume_aligned<16>(a.data());
+    std::assume_aligned<16>(b.data());
     #ifdef USE_INTERLEAVED_FORMAT
     for(int i = 0; i < N / 2; i++){
         const std::complex tmp = std::complex(a[2*i], a[2*i+1]) * std::complex(b[2*i], b[2*i+1]);
@@ -171,6 +174,9 @@ template <uint32_t N>
 inline void FMAInFD(std::array<double, N> &res, const std::array<double, N> &a,
                     const std::array<double, N> &b)
 {
+    std::assume_aligned<16>(res.data());
+    std::assume_aligned<16>(a.data());
+    std::assume_aligned<16>(b.data());
     #ifdef USE_INTERLEAVED_FORMAT
     for(int i = 0; i < N / 2; i++){
         std::complex tmp = std::complex(a[2*i], a[2*i+1]) * std::complex(b[2*i], b[2*i+1]);
diff --git a/include/params.hpp b/include/params.hpp
index d4d19a1..0fe21cd 100644
--- a/include/params.hpp
+++ b/include/params.hpp
@@ -133,7 +133,7 @@ using BootstrappingKey =
     std::array<BootstrappingKeyElement<P>, P::domainP::k * P::domainP::n>;
 template <class P>
 using BootstrappingKeyFFT =
-    alignas(16) std::array<BootstrappingKeyElementFFT<P>,
+    std::array<BootstrappingKeyElementFFT<P>,
                P::domainP::k * P::domainP::n / P::Addends>;
 template <class P>
 using BootstrappingKeyNTT =
@@ -143,7 +143,7 @@ using BootstrappingKeyRAINTT =
     std::array<TRGSWRAINTT<typename P::targetP>, P::domainP::k * P::domainP::n>;
 
 template <class P>
-using KeySwitchingKey = alignas(16) std::array<
+using KeySwitchingKey = std::array<
     std::array<std::array<TLWE<typename P::targetP>, (1 << P::basebit) - 1>,
                P::t>,
     P::domainP::k * P::domainP::n>;