diff --git a/include/mulfft.hpp b/include/mulfft.hpp
index d43ef75..ccd7f40 100644
--- a/include/mulfft.hpp
+++ b/include/mulfft.hpp
@@ -155,20 +155,20 @@ template <uint32_t N>
 inline void FMAInFD(std::array<double, N> &res, const std::array<double, N> &a,
                     const std::array<double, N> &b)
 {
-    // for (int i = 0; i < N / 2; i++) {
-    //     res[i] = std::fma(a[i], b[i], res[i]);
-    //     res[i + N / 2] = std::fma(a[i + N / 2], b[i], res[i + N / 2]);
-    // }
-    // for (int i = 0; i < N / 2; i++) {
-    //     res[i + N / 2] = std::fma(a[i], b[i + N / 2], res[i + N / 2]);
-    //     res[i] -= a[i + N / 2] * b[i + N / 2];
-    // }
     for (int i = 0; i < N / 2; i++) {
-        res[i] = std::fma(a[i + N / 2], b[i + N / 2], -res[i]);
-        res[i] = std::fma(a[i], b[i], -res[i]);
-        res[i + N / 2] = std::fma(a[i], b[i + N / 2], res[i + N / 2]);
+        res[i] = std::fma(a[i], b[i], res[i]);
         res[i + N / 2] = std::fma(a[i + N / 2], b[i], res[i + N / 2]);
     }
+    for (int i = 0; i < N / 2; i++) {
+        res[i + N / 2] = std::fma(a[i], b[i + N / 2], res[i + N / 2]);
+        res[i] -= a[i + N / 2] * b[i + N / 2];
+    }
+    // for (int i = 0; i < N / 2; i++) {
+    //     res[i] = std::fma(a[i + N / 2], b[i + N / 2], -res[i]);
+    //     res[i] = std::fma(a[i], b[i], -res[i]);
+    //     res[i + N / 2] = std::fma(a[i], b[i + N / 2], res[i + N / 2]);
+    //     res[i + N / 2] = std::fma(a[i + N / 2], b[i], res[i + N / 2]);
+    // }
 }
 
 template <class P>
diff --git a/include/params.hpp b/include/params.hpp
index 0fe21cd..d4d19a1 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 =
-    std::array<BootstrappingKeyElementFFT<P>,
+    alignas(16) 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 = std::array<
+using KeySwitchingKey = alignas(16) std::array<
     std::array<std::array<TLWE<typename P::targetP>, (1 << P::basebit) - 1>,
                P::t>,
     P::domainP::k * P::domainP::n>;
diff --git a/test/gatebootstrapping.cpp b/test/gatebootstrapping.cpp
index 3693024..53f9b02 100644
--- a/test/gatebootstrapping.cpp
+++ b/test/gatebootstrapping.cpp
@@ -1,3 +1,7 @@
+#ifdef USE_PERF
+#include <gperftools/profiler.h>
+#endif
+
 #include <cassert>
 #include <chrono>
 #include <iostream>
@@ -6,7 +10,7 @@
 
 int main()
 {
-    constexpr uint32_t num_test = 10;
+    constexpr uint32_t num_test = 1000;
     std::random_device seed_gen;
     std::default_random_engine engine(seed_gen());
     std::uniform_int_distribution<uint32_t> binary(0, 1);
@@ -28,11 +32,16 @@ int main()
 
     std::chrono::system_clock::time_point start, end;
     start = std::chrono::system_clock::now();
-
+#ifdef USE_PERF
+    ProfilerStart("gb.prof");
+#endif
     for (int test = 0; test < num_test; test++) {
         TFHEpp::GateBootstrapping<iksP, bkP, bkP::targetP::μ>(bootedtlwe[test],
                                                               tlwe[test], ek);
     }
+#ifdef USE_PERF
+    ProfilerStop();
+#endif
 
     end = std::chrono::system_clock::now();
     double elapsed =