diff --git a/CMakeLists.txt b/CMakeLists.txt
index 67b0bed8..1d9313b7 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -52,10 +52,13 @@ include_directories("src")
 foreach(F ${PIP_FOLDERS})
 	include_directories("src/${F}")
 	file(GLOB HS "src/${F}/*.h")
+	file(GLOB PHS "src/${F}/*_p.h")
 	file(GLOB CS "src/${F}/*.cpp")
 	list(APPEND HDRS ${HS})
+	list(APPEND PHDRS ${PHS})
 	list(APPEND CPPS ${CS})
 endforeach(F)
+list(REMOVE_ITEM HDRS ${PHDRS})
 
 # Check Bessel functions
 set(CMAKE_REQUIRED_INCLUDES math.h)
@@ -150,7 +153,7 @@ if (FFTW)
 	if (WIN32)
 		list(APPEND LIBS fftw3-3 fftw3f-3 fftw3l-3)
 	else ()
-		list(APPEND LIBS fftw3 fftw3f fftw3l)
+		list(APPEND LIBS fftw3 fftw3f)
 	endif ()
 else ()
 	message(STATUS "Building without fftw3 support")
@@ -183,7 +186,7 @@ if (WIN32)
 	list(APPEND LIBS ws2_32 iphlpapi psapi)
 	list(APPEND CPPS "pip_resource_win.rc")
 	set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DPSAPI_VERSION=1")
-	add_library(pip SHARED ${CPPS} ${HDRS})
+	add_library(pip SHARED ${CPPS} ${HDRS} ${PHDRS})
 	if (${CMAKE_C_COMPILER} STREQUAL "cl")
 		include(GenerateExportHeader)
 		generate_export_header(pip)
diff --git a/main.cpp b/main.cpp
index 794eb8ad..019f70e3 100644
--- a/main.cpp
+++ b/main.cpp
@@ -43,8 +43,9 @@ int asize(ssize_t s, size_t pid_rsize) {
 int main(int argc, char *argv[]) {
 	PIFFTWf fft;
 	PIVector<float> in;
-	for (int i = 0; i < 32; ++i)
+	for (int i = 0; i < 50; ++i)
 		in << i%10;
+	fft.preparePlan(50, PIFFTWf::foReal);
 	PIVector<complexf> out = fft.calcFFT(in);
 	piCout << out;
 	/*for (int i = 0; i < 16; ++i) {
diff --git a/src/math/picrypt.cpp b/src/math/picrypt.cpp
index bed068e1..3559f0d9 100644
--- a/src/math/picrypt.cpp
+++ b/src/math/picrypt.cpp
@@ -35,7 +35,7 @@ PICrypt::PICrypt() {
 	randombytes_buf(key_.data(), key_.size());
 	randombytes_buf(nonce_.data(), nonce_.size());
 #else
-	piCout << "[PICrypt]" << "Warning: PICrypt is disabled, to enable install sodium library and build pip with -DCRYPT=";
+	piCout << "[PICrypt]" << "Warning: PICrypt is disabled, to enable install libsodium-dev library and build pip with -DCRYPT=1";
 #endif
 }
 
diff --git a/src/math/pifft.h b/src/math/pifft.h
index edce2688..80dcc1dc 100644
--- a/src/math/pifft.h
+++ b/src/math/pifft.h
@@ -153,6 +153,7 @@ public:
 
 private:
 	void operator =(const PIFFTW & );
+	PIFFTW(const PIFFTW &);
 	inline void newP(void *& _p) {}
 	inline void deleteP(void *& _p) {}
 
@@ -167,6 +168,8 @@ template<> inline void PIFFTW<float>::preparePlan(int size, FFT_Operation op) {(
 template<> inline void PIFFTW<float>::newP(void *& _p) {_p = new _PIFFTW_P_float_();}
 template<> inline void PIFFTW<float>::deleteP(void *& _p) {if (_p) delete (_PIFFTW_P_float_*)_p; _p = 0;}
 
+typedef PIFFTW<float>   PIFFTWf;
+
 
 template<> inline const PIVector<complex<double> > & PIFFTW<double>::calcFFT(const PIVector<complex<double> > & in) {return ((_PIFFTW_P_double_*)p)->calcFFT(in);}
 template<> inline const PIVector<complex<double> > & PIFFTW<double>::calcFFT(const PIVector<double> & in) {return ((_PIFFTW_P_double_*)p)->calcFFTR(in);}
@@ -175,7 +178,10 @@ template<> inline void PIFFTW<double>::preparePlan(int size, FFT_Operation op) {
 template<> inline void PIFFTW<double>::newP(void *& _p) {_p = new _PIFFTW_P_double_();}
 template<> inline void PIFFTW<double>::deleteP(void *& _p) {if (_p) delete (_PIFFTW_P_double_*)_p; _p = 0;}
 
+typedef PIFFTW<double>  PIFFTWd;
 
+
+#ifdef WINDOWS
 template<> inline const PIVector<complex<ldouble> > & PIFFTW<ldouble>::calcFFT(const PIVector<complex<ldouble> > & in) {return ((_PIFFTW_P_ldouble_*)p)->calcFFT(in);}
 template<> inline const PIVector<complex<ldouble> > & PIFFTW<ldouble>::calcFFT(const PIVector<ldouble> & in) {return ((_PIFFTW_P_ldouble_*)p)->calcFFTR(in);}
 template<> inline const PIVector<complex<ldouble> > & PIFFTW<ldouble>::calcFFTinverse(const PIVector<complex<ldouble> > & in) {return ((_PIFFTW_P_ldouble_*)p)->calcFFTI(in);}
@@ -183,10 +189,7 @@ template<> inline void PIFFTW<ldouble>::preparePlan(int size, FFT_Operation op)
 template<> inline void PIFFTW<ldouble>::newP(void *& _p) {_p = new _PIFFTW_P_ldouble_();}
 template<> inline void PIFFTW<ldouble>::deleteP(void *& _p) {if (_p) delete (_PIFFTW_P_ldouble_*)_p; _p = 0;}
 
-
-typedef PIFFTW<float>   PIFFTWf;
-typedef PIFFTW<double>  PIFFTWd;
 typedef PIFFTW<ldouble> PIFFTWld;
-
+#endif // WINDOWS
 
 #endif // PIFFT_H
diff --git a/src/math/pifft_p.h b/src/math/pifft_p.h
index c43e53a9..a3fe041e 100644
--- a/src/math/pifft_p.h
+++ b/src/math/pifft_p.h
@@ -25,7 +25,7 @@
 
 #include "pivector.h"
 #include "pimutex.h"
-#include <complex.h>
+#include "picout.h"
 #ifdef PIP_FFTW
 #  include "fftw3.h"
 #else
@@ -36,106 +36,136 @@
 #endif
 
 
-static PIMutex __pip_fft_plan_mutex;
-
-
 template <typename T>
 class PIFFTW_Private
 {
 public:
-	explicit PIFFTW_Private() {plan = 0; newPlan(plan);}
-	~PIFFTW_Private() {deletePlan(plan);}
+	explicit PIFFTW_Private() {
+		plan = 0;
+#ifdef PIP_FFTW
+//		fftwf_m
+#else
+		piCout << "[PIFFTW]" << "Warning: PIFFTW is disabled, to enable install libfftw3-dev library and build pip with -DFFTW=1";
+#endif
+		p_makeThreadSafe();
+	}
+	~PIFFTW_Private() {p_destroyPlan(plan);}
 
 	const PIVector<complex<T> > & calcFFT(const PIVector<complex<T> > & in) {
-		result.resize(in.size());
-		__pip_fft_plan_mutex.lock();
-		createPlan_c_1d(plan, in.size_s(), in.data(), result.data(), FFTW_FORWARD, FFTW_ESTIMATE);
-		__pip_fft_plan_mutex.unlock();
-		executePlan(plan);
-		destroyPlan(plan);
-		return result;
+		if (prepare != PlanParams(in.size(), fo_complex)) {
+			p_out.resize(in.size());
+			piCout << "[PIFFTW]" << "creating plan";
+			p_createPlan_c2c_1d(plan, in.size(), in.data(), p_out.data(), FFTW_FORWARD, FFTW_ESTIMATE | FFTW_UNALIGNED);
+			prepare = PlanParams(in.size(), fo_complex);
+		}
+		p_executePlan_c2c(plan, in.data(), p_out.data());
+		return p_out;
 	}
 	const PIVector<complex<T> > & calcFFT(const PIVector<T> & in) {
-		result.resize(in.size());
-		__pip_fft_plan_mutex.lock();
-		createPlan_r2c_1d(plan, in.size_s(), in.data(), result.data(), FFTW_ESTIMATE);
-		__pip_fft_plan_mutex.unlock();
-		executePlan(plan);
-		destroyPlan(plan);
-		return result;
+		if (prepare != PlanParams(in.size(), fo_real)) {
+			p_out.resize(in.size());
+			piCout << "[PIFFTW]" << "creating plan";
+			p_createPlan_r2c_1d(plan, in.size(), in.data(), p_out.data(), FFTW_ESTIMATE | FFTW_UNALIGNED);
+			prepare = PlanParams(in.size(), fo_real);
+		}
+		p_executePlan_r2c(plan, in.data(), p_out.data());
+		return p_out;
 	}
 	const PIVector<complex<T> > & calcFFTinverse(const PIVector<complex<T> > & in) {
-		result.resize(in.size());
-		__pip_fft_plan_mutex.lock();
-		createPlan_c_1d(plan, in.size_s(), in.data(), result.data(), FFTW_BACKWARD, FFTW_ESTIMATE);
-		__pip_fft_plan_mutex.unlock();
-		executePlan(plan);
-		destroyPlan(plan);
-		return result;
+		if (prepare != PlanParams(in.size(), fo_inverse)) {
+			p_out.resize(in.size());
+			piCout << "[PIFFTW]" << "creating plan";
+			p_createPlan_c2c_1d(plan, in.size(), in.data(), p_out.data(), FFTW_BACKWARD, FFTW_ESTIMATE | FFTW_UNALIGNED);
+			prepare = PlanParams(in.size(), fo_inverse);
+		}
+		p_executePlan_c2c(plan, in.data(), p_out.data());
+		return p_out;
 	}
 
 	enum FFT_Operation {fo_real, fo_complex, fo_inverse};
 
 	void preparePlan(int size, int op) {
-		PIVector<complex<T> > in(size), out(size);
-		PIVector<T> inr(size);
-		__pip_fft_plan_mutex.lock();
+		p_inr.clear();
+		p_in.clear();
+		p_out.clear();
 		switch ((FFT_Operation)op) {
-		case fo_real:
-			createPlan_r2c_1d(plan, in.size_s(), inr.data(), out.data(), FFTW_MEASURE);
+			case fo_real:
+				p_inr.resize(size);
+				p_out.resize(size);
+				p_createPlan_r2c_1d(plan, size, p_inr.data(), p_out.data(), FFTW_MEASURE | FFTW_UNALIGNED);
 			break;
-		case fo_complex:
-			createPlan_c_1d(plan, in.size_s(), in.data(), out.data(), FFTW_FORWARD, FFTW_MEASURE);
+			case fo_complex:
+				p_in.resize(size);
+				p_out.resize(size);
+				p_createPlan_c2c_1d(plan, size, p_in.data(), p_out.data(), FFTW_FORWARD, FFTW_MEASURE | FFTW_UNALIGNED);
 			break;
-		case fo_inverse:
-			createPlan_c_1d(plan, in.size_s(), in.data(), out.data(), FFTW_BACKWARD, FFTW_MEASURE);
+			case fo_inverse:
+				p_in.resize(size);
+				p_out.resize(size);
+				p_createPlan_c2c_1d(plan, size, p_in.data(), p_out.data(), FFTW_BACKWARD, FFTW_MEASURE | FFTW_UNALIGNED);
+			break;
+			default:
+				size = 0;
 			break;
 		}
-		__pip_fft_plan_mutex.unlock();
+		prepare = PlanParams(size, (FFT_Operation)op);
 	}
 
-	inline void createPlan_c_1d(void * plan, int size, const void * in, void * out, int dir, int flags) {}
-	inline void createPlan_r2c_1d(void * plan, int size, const void * in, void * out, int flags) {}
-	inline void executePlan(void * plan) {}
-	inline void destroyPlan(void * plan) {}
-	inline void newPlan(void *& plan) {}
-	inline void deletePlan(void *& plan) {}
+	inline void p_createPlan_c2c_1d(void *& plan, int size, const void * in, void * out, int dir, int flags) {}
+	inline void p_createPlan_r2c_1d(void *& plan, int size, const void * in, void * out, int flags) {}
+	inline void p_executePlan_c2c(void * plan, const void * in, void * out) {}
+	inline void p_executePlan_r2c(void * plan, const void * in, void * out) {}
+	inline void p_destroyPlan(void *& plan) {}
+	inline void p_makeThreadSafe() {}
 
-	PIVector<complex<T> > result;
+	struct PlanParams {
+		PlanParams() {size = 0; op = fo_complex;}
+		PlanParams(int size_, FFT_Operation op_) {size = size_; op = op_;}
+		bool isValid() {return size > 0;}
+		bool operator ==(const PlanParams & v) const {return (v.size == size) && (v.op == op);}
+		bool operator !=(const PlanParams & v) const {return !(*this == v);}
+		int size;
+		FFT_Operation op;
+	};
+
+	PIVector<complex<T> > p_in;
+	PIVector<T> p_inr;
+	PIVector<complex<T> > p_out;
 	void * plan;
+	PlanParams prepare;
 };
 
 
 #ifdef PIP_FFTW
+template<> inline void PIFFTW_Private<float>::p_createPlan_c2c_1d(void *& plan, int size, const void * in, void * out, int dir, int flags) {
+	plan = fftwf_plan_dft_1d(size, (fftwf_complex *)in, (fftwf_complex *)out, dir, flags);}
+template<> inline void PIFFTW_Private<float>::p_createPlan_r2c_1d(void *& plan, int size, const void * in, void * out, int flags) {
+	plan = fftwf_plan_dft_r2c_1d(size, (float *)in, (fftwf_complex *)out, flags);}
+template<> inline void PIFFTW_Private<float>::p_executePlan_c2c(void * plan, const void * in, void * out) {fftwf_execute_dft((fftwf_plan)plan, (fftwf_complex *)in, (fftwf_complex *)out);}
+template<> inline void PIFFTW_Private<float>::p_executePlan_r2c(void * plan, const void * in, void * out) {fftwf_execute_dft_r2c((fftwf_plan)plan, (float *)in, (fftwf_complex *)out);}
+template<> inline void PIFFTW_Private<float>::p_destroyPlan(void *& plan) {if (plan) fftwf_destroy_plan((fftwf_plan)plan); plan = 0;}
+template<> inline void PIFFTW_Private<float>::p_makeThreadSafe() {fftwf_make_planner_thread_safe();}
 
-template<> inline void PIFFTW_Private<float>::createPlan_c_1d(void * plan, int size, const void * in, void * out, int dir, int flags) {
-	*(fftwf_plan*)plan = fftwf_plan_dft_1d(size, (fftwf_complex *)in, (fftwf_complex *)out, dir, flags);}
-template<> inline void PIFFTW_Private<float>::createPlan_r2c_1d(void * plan, int size, const void * in, void * out, int flags) {
-	*(fftwf_plan*)plan = fftwf_plan_dft_r2c_1d(size, (float *)in, (fftwf_complex *)out, flags);}
-template<> inline void PIFFTW_Private<float>::executePlan(void * plan) {fftwf_execute(*(fftwf_plan*)plan);}
-template<> inline void PIFFTW_Private<float>::destroyPlan(void * plan) {fftwf_destroy_plan(*(fftwf_plan*)plan);}
-template<> inline void PIFFTW_Private<float>::newPlan(void *& plan) {plan = new fftwf_plan();}
-template<> inline void PIFFTW_Private<float>::deletePlan(void *& plan) {if (plan) delete (fftwf_plan*)plan; plan = 0;}
+template<> inline void PIFFTW_Private<double>::p_createPlan_c2c_1d(void *& plan, int size, const void * in, void * out, int dir, int flags) {
+	plan = fftw_plan_dft_1d(size, (fftw_complex *)in, (fftw_complex *)out, dir, flags);}
+template<> inline void PIFFTW_Private<double>::p_createPlan_r2c_1d(void *& plan, int size, const void * in, void * out, int flags) {
+	plan = fftw_plan_dft_r2c_1d(size, (double *)in, (fftw_complex *)out, flags);}
+template<> inline void PIFFTW_Private<double>::p_executePlan_c2c(void * plan, const void * in, void * out) {fftw_execute_dft((fftw_plan)plan, (fftw_complex *)in, (fftw_complex *)out);}
+template<> inline void PIFFTW_Private<double>::p_executePlan_r2c(void * plan, const void * in, void * out) {fftw_execute_dft_r2c((fftw_plan)plan, (double *)in, (fftw_complex *)out);}
+template<> inline void PIFFTW_Private<double>::p_destroyPlan(void *& plan) {if (plan) fftw_destroy_plan((fftw_plan)plan); plan = 0;}
+template<> inline void PIFFTW_Private<double>::p_makeThreadSafe() {fftw_make_planner_thread_safe();}
 
+#ifdef WINDOWS
+template<> inline void PIFFTW_Private<ldouble>::p_createPlan_c2c_1d(void *& plan, int size, const void * in, void * out, int dir, int flags) {
+	plan = fftwl_plan_dft_1d(size, (fftwl_complex *)in, (fftwl_complex *)out, dir, flags);}
+template<> inline void PIFFTW_Private<ldouble>::p_createPlan_r2c_1d(void *& plan, int size, const void * in, void * out, int flags) {
+	plan = fftwl_plan_dft_r2c_1d(size, (ldouble *)in, (fftwl_complex *)out, flags);}
+template<> inline void PIFFTW_Private<ldouble>::p_executePlan_c2c(void * plan, const void * in, void * out) {fftwl_execute_dft((fftwl_plan)plan, (fftwl_complex *)in, (fftwl_complex *)out);}
+template<> inline void PIFFTW_Private<ldouble>::p_executePlan_r2c(void * plan, const void * in, void * out) {fftwl_execute_dft_r2c((fftwl_plan)plan, (ldouble *)in, (fftwl_complex *)out);}
+template<> inline void PIFFTW_Private<ldouble>::p_destroyPlan(void *& plan) {if (plan) fftwl_destroy_plan((fftwl_plan)plan); plan = 0;}
+template<> inline void PIFFTW_Private<ldouble>::p_makeThreadSafe() {fftwl_make_planner_thread_safe();}
+#endif // WINDOWS
 
-template<> inline void PIFFTW_Private<double>::createPlan_c_1d(void * plan, int size, const void * in, void * out, int dir, int flags) {
-	*(fftw_plan*)plan = fftw_plan_dft_1d(size, (fftw_complex *)in, (fftw_complex *)out, dir, flags);}
-template<> inline void PIFFTW_Private<double>::createPlan_r2c_1d(void * plan, int size, const void * in, void * out, int flags) {
-	*(fftw_plan*)plan = fftw_plan_dft_r2c_1d(size, (double *)in, (fftw_complex *)out, flags);}
-template<> inline void PIFFTW_Private<double>::executePlan(void * plan) {fftw_execute(*(fftw_plan*)plan);}
-template<> inline void PIFFTW_Private<double>::destroyPlan(void * plan) {fftw_destroy_plan(*(fftw_plan*)plan);}
-template<> inline void PIFFTW_Private<double>::newPlan(void *& plan) {plan = new fftw_plan();}
-template<> inline void PIFFTW_Private<double>::deletePlan(void *& plan) {if (plan) delete (fftw_plan*)plan; plan = 0;}
-
-
-template<> inline void PIFFTW_Private<ldouble>::createPlan_c_1d(void * plan, int size, const void * in, void * out, int dir, int flags) {
-	*(fftwl_plan*)plan = fftwl_plan_dft_1d(size, (fftwl_complex *)in, (fftwl_complex *)out, dir, flags);}
-template<> inline void PIFFTW_Private<ldouble>::createPlan_r2c_1d(void * plan, int size, const void * in, void * out, int flags) {
-	*(fftwl_plan*)plan = fftwl_plan_dft_r2c_1d(size, (ldouble *)in, (fftwl_complex *)out, flags);}
-template<> inline void PIFFTW_Private<ldouble>::executePlan(void * plan) {fftwl_execute(*(fftwl_plan*)plan);}
-template<> inline void PIFFTW_Private<ldouble>::destroyPlan(void * plan) {fftwl_destroy_plan(*(fftwl_plan*)plan);}
-template<> inline void PIFFTW_Private<ldouble>::newPlan(void *& plan) {plan = new fftwl_plan();}
-template<> inline void PIFFTW_Private<ldouble>::deletePlan(void *& plan) {if (plan) delete (fftwl_plan*)plan; plan = 0;}
 
 #endif // PIP_FFTW