Add fft-bench

.gitmodules

@@ -10,3 +10,6 @@
 [submodule "thirdparties/concrete-fft"]
 	path = thirdparties/concrete-fft
 	url = https://github.com/virtualsecureplatform/concrete-fft.git
+[submodule "thirdparties/concrete-ntt"]
+	path = thirdparties/concrete-ntt
+	url = https://github.com/virtualsecureplatform/concrete-ntt.git

Dockerfile-ubuntu2404

@@ -0,0 +1,10 @@
+FROM ubuntu:24.04
+
+LABEL maintainer="nindanaoto(Kotaro MATSUOKA) <[email protected]>"
+
+# install build dependencies
+RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get upgrade -y
+RUN DEBIAN_FRONTEND=noninteractive apt-get install -y build-essential g++ libomp-dev cmake git ninja-build libfftw3-dev wget curl
+RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | bash -s -- -y 
+ENV PATH="/root/.cargo/bin:${PATH}"
+RUN wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB && apt-key add GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB && echo "deb https://apt.repos.intel.com/oneapi all main" | tee /etc/apt/sources.list.d/oneAPI.list && apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y intel-oneapi-mkl-devel && rm GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB

docker-fft-bench.bash

@@ -0,0 +1,3 @@
+#!/bin/bash
+docker build -t tfhepp -f Dockerfile-ubuntu2404 .
+docker run -v $PWD:/TFHEpp tfhepp bash -c "cd /TFHEpp && bash ./fft-bench.bash"

fft-bench.bash

@@ -0,0 +1,36 @@
+#!/bin/bash
+# This script is used to benchmark the FFT performance of TFHEpp
+echo "Benchmarking HomNAND using different FFTs" | tee /tmp/log.txt
+cmake . -G Ninja -B /tmp/build -DENABLE_TEST=ON
+cd /tmp/build
+ninja
+echo "Benchmarking SPQLIOS" | tee -a /tmp/log.txt
+./test/nand | tee -a /tmp/log.txt
+
+cd /TFHEpp
+rm -rf /tmp/build
+cmake . -G Ninja -B /tmp/build -DENABLE_TEST=ON  -DUSE_CONCRETE_FFT=ON
+cd /tmp/build
+ninja
+ninja
+echo "Benchmarking concrete-fft" | tee -a /tmp/log.txt
+./test/nand | tee -a /tmp/log.txt
+
+cd /TFHEpp
+rm -rf /tmp/build
+source /opt/intel/oneapi/setvars.sh
+cmake . -G Ninja -B /tmp/build -DENABLE_TEST=ON -DUSE_MKL=ON
+cd /tmp/build
+ninja
+echo "Benchmarking MKL" | tee -a /tmp/log.txt
+./test/nand | tee -a /tmp/log.txt
+
+cd /TFHEpp
+rm -rf /tmp/build
+cmake . -G Ninja -B /tmp/build -DENABLE_TEST=ON -DUSE_FFTW3=ON
+cd /tmp/build
+ninja
+echo "Benchmarking FFTW3" | tee -a /tmp/log.txt
+./test/nand | tee -a /tmp/log.txt
+
+cat /tmp/log.txt

include/mulfft.hpp

@@ -152,10 +152,10 @@ inline void TwistIFFTUInt(PolynomialInFD<P> &res, const Polynomial<P> &a)
         if constexpr (std::is_same_v<typename P::T, uint32_t>)
             fftplvl1.execute_reverse_uint(res.data(), a.data());
         // if constexpr (std::is_same_v<typename P::T, uint64_t>)
-        // fftplvl1.execute_reverse_torus64(res.data(), a.data());
+            // fftplvl1.execute_reverse_torus64(res.data(), a.data());
     }
     // else if constexpr (std::is_same_v<typename P::T, uint64_t>)
-    // fftplvl2.execute_reverse_torus64(res.data(), a.data());
+        // fftplvl2.execute_reverse_torus64(res.data(), a.data());
     else
         static_assert(false_v<typename P::T>, "Undefined TwistIFFT!");
 }
@@ -316,7 +316,8 @@ inline void PolyMul(Polynomial<P> &res, const Polynomial<P> &a,
 }
 
 template <class P>
-inline void PolyMulRescaleUnsigned(Polynomial<P> &res, const Polynomial<P> &a,
+inline void PolyMulRescaleUnsigned(Polynomial<P> &res,
+                                   const Polynomial<P> &a,
                                    const Polynomial<P> &b)
 {
     // if constexpr (std::is_same_v<typename P::T, uint32_t>) {
@@ -349,20 +350,16 @@ inline void PolyMulNaive(Polynomial<P> &res, const Polynomial<P> &a,
 }
 
 template <class P>
-inline void PolyMulNaieveRescaleUnsigned(Polynomial<P> &res,
-                                         const Polynomial<P> &a,
-                                         const Polynomial<P> &b)
+inline void PolyMulNaieveRescaleUnsigned(Polynomial<P> &res, const Polynomial<P> &a,
+                                 const Polynomial<P> &b)
 {
     for (int i = 0; i < P::n; i++) {
         __int128_t ri = 0;
         for (int j = 0; j <= i; j++)
-            ri += static_cast<__int128_t>(a[j]) *
-                  static_cast<__int128_t>(b[i - j]);
+            ri += static_cast<__int128_t>(a[j]) * static_cast<__int128_t>(b[i - j]);
         for (int j = i + 1; j < P::n; j++)
-            ri -= static_cast<__int128_t>(a[j]) *
-                  static_cast<__int128_t>(b[P::n + i - j]);
-        // res[i] = static_cast<typename P::T>((ri) >>
-        // (std::numeric_limits<typename P::T>::digits - 3));
+            ri -= static_cast<__int128_t>(a[j]) * static_cast<__int128_t>(b[P::n + i - j]);
+        // res[i] = static_cast<typename P::T>((ri) >> (std::numeric_limits<typename P::T>::digits - 3));
         res[i] = static_cast<typename P::T>((ri) >> 29);
     }
 }

thirdparties/fftw/fft_processor_fftw.cpp

@@ -42,6 +42,20 @@ void FFT_Processor_FFTW::execute_reverse_int(double *res, const int32_t *a)
     }
 }
 
+void FFT_Processor_FFTW::execute_reverse_uint(double *res, const uint32_t *a)
+{
+    for (int i = 0; i < Ns2; i++) {
+        auto tmp = twist[i] * std::complex((double)a[i], (double)a[Ns2 + i]);
+        inbuf[i][0] = tmp.real();
+        inbuf[i][1] = tmp.imag();
+    }
+    fftw_execute_dft(plan_forward, inbuf, outbuf);
+    for (int i = 0; i < Ns2; i++) {
+        res[i] = outbuf[i][0];
+        res[i + Ns2] = outbuf[i][1];
+    }
+}
+
 void FFT_Processor_FFTW::execute_reverse_torus32(double *res, const uint32_t *a)
 {
     execute_reverse_int(res, (int32_t *)a);
@@ -89,8 +103,8 @@ void FFT_Processor_FFTW::execute_direct_torus32_rescale(uint32_t *res,
     for (int i = 0; i < Ns2; i++) {
         auto res_tmp = std::complex<double>(outbuf[i][0], outbuf[i][1]) *
                        std::conj(twist[i]);
-        res[i] = CAST_DOUBLE_TO_UINT32(res_tmp.real() / (Δ / 4));
-        res[i + Ns2] = CAST_DOUBLE_TO_UINT32(res_tmp.imag() / (Δ / 4));
+        res[i] = CAST_DOUBLE_TO_UINT32(res_tmp.real() / Δ);
+        res[i + Ns2] = CAST_DOUBLE_TO_UINT32(res_tmp.imag() / Δ);
     }
 }
 
@@ -139,7 +153,7 @@ void FFT_Processor_FFTW::execute_direct_torus64_rescale(uint64_t *res,
         tmp[i] = res_tmp.real();
         tmp[i + Ns2] = res_tmp.imag();
     }
-    for (int i = 0; i < N; i++) res[i] = uint64_t(std::round(tmp[i] / (Δ / 4)));
+    for (int i = 0; i < N; i++) res[i] = uint64_t(std::round(tmp[i] / Δ));
 }
 
 FFT_Processor_FFTW::~FFT_Processor_FFTW()

thirdparties/fftw/fft_processor_fftw.h

@@ -27,6 +27,8 @@ class FFT_Processor_FFTW {
 
     void execute_reverse_int(double *res, const int32_t *a);
 
+    void execute_reverse_uint(double *res, const uint32_t *a);
+
     void execute_reverse_torus32(double *res, const uint32_t *a);
 
     void execute_direct_torus32(uint32_t *res, const double *a);

thirdparties/mkl/fft_processor_mkl.hpp

@@ -41,6 +41,16 @@ class FFT_Processor_MKL {
         DftiComputeForward(desc_handle, res);
     }
 
+    void execute_reverse_uint(double *res, const uint32_t *a)
+    {
+        for (int i = 0; i < Ns2; i++) {
+            auto tmp = twist[i] * std::complex((double)a[i], (double)a[Ns2 + i]);
+            res[2*i] = tmp.real();
+            res[2*i+1] = tmp.imag();
+        }
+        DftiComputeForward(desc_handle, res);
+    }
+
     void execute_reverse_torus32(double *res, const uint32_t *a)
     {
         execute_reverse_int(res, (int32_t *)a);