diff --git a/README.md b/README.md
index 9c0b42aa..7f4f7bce 100644
--- a/README.md
+++ b/README.md
@@ -1,5 +1,9 @@
-# Detailed installation instructions (64-bit Ubuntu 16 or 18):
+# Installation instructions:
+In principle you should be able to install this on Windows / OS X / Linux using conda and 'conda install -c marcus-o s4'.
+Install in a new environment to avoid incompatabilities.
+I don't have a large test base, so I would appreciate if you let me know if it works (marcus.ossiander at gmail) or not, then I'll try to help).
 
+# Detailed installation instructions (64-bit Ubuntu 16 or 18):
 ## Key steps:
 
 ```
diff --git a/S4/RNP/Eigensystems.h b/S4/RNP/Eigensystems.h
index eee2560c..f5950b1d 100644
--- a/S4/RNP/Eigensystems.h
+++ b/S4/RNP/Eigensystems.h
@@ -1,6 +1,10 @@
 #ifndef _RNP_EIGENSYSTEMS_H_
 #define _RNP_EIGENSYSTEMS_H_
 
+#ifdef _MSC_VER 
+#include <algorithm>
+#endif
+
 #include <cstddef>
 #include <complex>
 
diff --git a/S4/S4.cpp b/S4/S4.cpp
index d0ac9dfe..6297aa1a 100644
--- a/S4/S4.cpp
+++ b/S4/S4.cpp
@@ -61,7 +61,7 @@ namespace bs = boost::serialization;
 
 
 void* S4_malloc(size_t size){ // for debugging
-	void* ret = malloc_aligned(size, 16);
+	void* ret = malloc_aligned(size, 128);
 	// memset(ret, 0x0, size);
 	return ret;
 }
@@ -2939,7 +2939,8 @@ int Simulation_GetField(Simulation *S, const double r[3], double fE[6], double f
 	}
 	std::complex<double> *work = ab + n4;
 
-	RNP::TBLAS::Copy(n4, Lsoln->ab,1, ab,1);
+	memcpy(ab, Lsoln->ab, sizeof(std::complex<double>) * n4);
+	//RNP::TBLAS::Copy(n4, Lsoln->ab,1, ab,1);
 	//RNP::IO::PrintVector(n4, ab, 1);
 	TranslateAmplitudes(S->n_G, Lbands->q, L->thickness, dz, ab);
 	std::complex<double> efield[3], hfield[3];
@@ -3051,7 +3052,8 @@ int Simulation_GetFieldPlane(Simulation *S, int nxy[2], double zz, double *E, do
 	}
 	std::complex<double> *work = ab + n4;
 
-	RNP::TBLAS::Copy(n4, Lsoln->ab,1, ab,1);
+	memcpy(ab, Lsoln->ab, sizeof(std::complex<double>) * n4);
+	//RNP::TBLAS::Copy(n4, Lsoln->ab,1, ab,1);
 	//RNP::IO::PrintVector(n4, ab, 1);
 	TranslateAmplitudes(S->n_G, Lbands->q, L->thickness, dz, ab);
 	size_t snxy[2] = { nxy[0], nxy[1] };
diff --git a/S4/config.h b/S4/config.h
index 87dbdc38..53b55b4b 100644
--- a/S4/config.h
+++ b/S4/config.h
@@ -8,3 +8,7 @@
 #define PACKAGE_URL ""
 #define PACKAGE_VERSION "1.1.1"
 #define VERSION "1.1.1"
+
+#ifdef _MSC_VER 
+#define strcasecmp _stricmp
+#endif
\ No newline at end of file
diff --git a/S4/fmm/fft_iface.cpp b/S4/fmm/fft_iface.cpp
index 187a145e..e354e02d 100644
--- a/S4/fmm/fft_iface.cpp
+++ b/S4/fmm/fft_iface.cpp
@@ -20,7 +20,7 @@
 #include "fft_iface.h"
 #include <cstdlib>
 
-#ifdef HAVE_LIBFFTW3
+#ifdef HAVE_FFTW3
 #include <fftw3.h>
 #else
 #include <kiss_fft.h>
@@ -46,7 +46,7 @@ int fft_next_fast_size(int n){
 }
 
 std::complex<double> *fft_alloc_complex(size_t n){
-#ifdef HAVE_LIBFFTW3
+#ifdef HAVE_FFTW3
 	return (std::complex<double>*)(fftw_complex*)fftw_malloc(sizeof(fftw_complex) * n);
 #else
 	return (std::complex<double>*)KISS_FFT_MALLOC(sizeof(std::complex<double>) * n);
@@ -54,7 +54,7 @@ std::complex<double> *fft_alloc_complex(size_t n){
 }
 
 void fft_free(void *p){
-#ifdef HAVE_LIBFFTW3
+#ifdef HAVE_FFTW3
 	fftw_free(p);
 #else
 	KISS_FFT_FREE(p);
@@ -62,7 +62,7 @@ void fft_free(void *p){
 }
 
 struct tag_fft_plan{
-#ifdef HAVE_LIBFFTW3
+#ifdef HAVE_FFTW3
 	fftw_plan plan;
 #else
 	kiss_fftnd_cfg cfg;
@@ -76,7 +76,7 @@ fft_plan fft_plan_dft_2d(
 	int sign
 ){
 	fft_plan plan = NULL;
-#ifdef HAVE_LIBFFTW3
+#ifdef HAVE_FFTW3
 # ifdef HAVE_LIBPTHREAD
 	pthread_mutex_lock(&mutex);
 # endif
@@ -103,7 +103,7 @@ fft_plan fft_plan_dft_2d(
 }
 
 void fft_plan_exec(const fft_plan plan){
-#ifdef HAVE_LIBFFTW3
+#ifdef HAVE_FFTW3
 	fftw_execute(plan->plan);
 #else
 	kiss_fftnd(plan->cfg, (const kiss_fft_cpx *)plan->in, (kiss_fft_cpx *)plan->out);
@@ -112,7 +112,7 @@ void fft_plan_exec(const fft_plan plan){
 
 void fft_plan_destroy(fft_plan plan){
 	if(NULL == plan){ return; }
-#ifdef HAVE_LIBFFTW3
+#ifdef HAVE_FFTW3
 # ifdef HAVE_LIBPTHREAD
 	pthread_mutex_lock(&mutex);
 # endif
@@ -135,7 +135,7 @@ void fft_init(){
 }
 
 void fft_destroy(){
-#ifdef HAVE_LIBFFTW3
+#ifdef HAVE_FFTW3
 	fftw_cleanup();
 #endif
 #ifdef HAVE_LIBPTHREAD
diff --git a/S4/fmm/fmm_FFT.cpp b/S4/fmm/fmm_FFT.cpp
index e54e0b8f..c9ee7388 100644
--- a/S4/fmm/fmm_FFT.cpp
+++ b/S4/fmm/fmm_FFT.cpp
@@ -32,8 +32,8 @@
 #include "fmm.h"
 
 #include <limits>
-#include <kiss_fft.h>
-#include <tools/kiss_fftnd.h>
+//#include <kiss_fft.h>
+//#include <tools/kiss_fftnd.h>
 #include "fft_iface.h"
 
 int FMMGetEpsilon_FFT(const Simulation *S, const Layer *L, const int n, std::complex<double> *Epsilon2, std::complex<double> *Epsilon_inv){
diff --git a/S4/fmm/fmm_PolBasisJones.cpp b/S4/fmm/fmm_PolBasisJones.cpp
index e9389ad6..0a95d685 100644
--- a/S4/fmm/fmm_PolBasisJones.cpp
+++ b/S4/fmm/fmm_PolBasisJones.cpp
@@ -19,6 +19,7 @@
 
 #include <config.h>
 
+#define _USE_MATH_DEFINES
 #include <cmath>
 #include <S4.h>
 #include "../RNP/TBLAS.h"
diff --git a/S4/fmm/fmm_PolBasisVL.cpp b/S4/fmm/fmm_PolBasisVL.cpp
index c84f8b50..de9e2afa 100644
--- a/S4/fmm/fmm_PolBasisVL.cpp
+++ b/S4/fmm/fmm_PolBasisVL.cpp
@@ -238,7 +238,7 @@ int FMMGetEpsilon_PolBasisVL(const Simulation *S, const Layer *L, const int n, s
 		}
 		
 
-		fft_plan plan = fft_plan_dft_2d(ngrid, Ffrom, Fto, -1);
+		fft_plan plan = fft_plan_dft_2d(ngrid, Ffrom, Fto, 1);
 
 		// We fill in the quarter blocks of F in Fortran order
 		for(int w = 0; w < 4; ++w){
diff --git a/S4/main_python.c b/S4/main_python.c
index 376f6bbc..8f9a366a 100644
--- a/S4/main_python.c
+++ b/S4/main_python.c
@@ -30,7 +30,7 @@
 #include <string.h>
 #include <stdarg.h>
 
-#ifdef WIN32
+#ifdef _WIN32
 #define _USE_MATH_DEFINES
 #endif
 
@@ -2169,7 +2169,7 @@ static PyObject *S4_NewInterpolator(PyObject *self, PyObject *args, PyObject *kw
 static PyObject *S4_SolveInParallel(PyObject *Self, PyObject *args, PyObject *kwds)
 {
 	static char *kwlist[] = {"Layer", "Simulations", NULL};
-	const char *layerName;
+	//const char *layerName;
 	//S4_solve_in
 	Py_RETURN_NONE;
 }
diff --git a/S4/numalloc.c b/S4/numalloc.c
index c753b532..94bc9cc5 100644
--- a/S4/numalloc.c
+++ b/S4/numalloc.c
@@ -20,10 +20,10 @@
 #include <stdlib.h>
 #include <stdio.h>
 
-#ifdef WIN32
+#ifdef _WIN32
 # include <malloc.h>
-void * _aligned_malloc(size_t size, size_t alignment);
-void _aligned_free(void *ptr);
+// void * _aligned_malloc(size_t size, size_t alignment);
+// void _aligned_free(void *ptr);
 #else
 #include <inttypes.h>
 typedef uintptr_t malloc_aligned_ULONG_PTR;
@@ -33,7 +33,7 @@ typedef uintptr_t malloc_aligned_ULONG_PTR;
 //        The actual size of allocation will be greater than this size.
 // alignment : the alignment boundary
 void *malloc_aligned(size_t size, size_t alignment){
-#ifdef WIN32
+#ifdef _WIN32
 	return (void*)_aligned_malloc(size, alignment);
 #else
 	void *pa, *ptr;
@@ -51,7 +51,7 @@ void *malloc_aligned(size_t size, size_t alignment){
 }
 
 void free_aligned(void *ptr){
-#ifdef WIN32
+#ifdef _WIN32
 	_aligned_free(ptr);
 #else
 	if(ptr){
diff --git a/S4/pattern/predicates.c b/S4/pattern/predicates.c
index e0a65339..e7d1a193 100644
--- a/S4/pattern/predicates.c
+++ b/S4/pattern/predicates.c
@@ -116,7 +116,9 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <math.h>
+#ifndef _MSC_VER 
 #include <sys/time.h>
+#endif
 
 /* On some machines, the exact arithmetic routines might be defeated by the  */
 /*   use of internal extended precision floating-point registers.  Sometimes */
diff --git a/S4/rcwa.cpp b/S4/rcwa.cpp
index 1c7a8f8d..7c9ce5ce 100644
--- a/S4/rcwa.cpp
+++ b/S4/rcwa.cpp
@@ -50,7 +50,7 @@
 #include <numalloc.h>
 
 static inline void* rcwa_malloc(size_t size){
-	void *ret = malloc_aligned(size, 16);
+	void *ret = malloc_aligned(size, 128);
 	//memset(ret, 0x0, size);
 	return ret;
 }
diff --git a/S4/sort.c b/S4/sort.c
index 00b0c369..a82dc5ae 100644
--- a/S4/sort.c
+++ b/S4/sort.c
@@ -209,7 +209,9 @@ sort (void *const pbase, size_t total_elems, size_t size,
      of the array to sort, and END_PTR points at the very last element in
      the array (*not* one beyond it!). */
 
+#ifndef _MSC_VER
 #define min(x, y) ((x) < (y) ? (x) : (y))
+#endif
 
   {
     char *const end_ptr = &base_ptr[size * (total_elems - 1)];
diff --git a/conda_recipe/s4/Makefile b/conda_recipe/s4/Makefile
new file mode 100644
index 00000000..ed9c37f0
--- /dev/null
+++ b/conda_recipe/s4/Makefile
@@ -0,0 +1,182 @@
+CONDA_INC = $(CONDA_PREFIX)/lib/
+CONDA_LIB = $(CONDA_PREFIX)/lib/
+
+# BLAS_LIB = -L$(CONDA_LIB) -lmkl_core_dll -lmkl_rt
+# LAPACK_LIB = -L$(CONDA_LIB) -lmkl_core_dll -lmkl_rt
+
+BLAS_LIB = -L$(CONDA_LIB) -lopenblas
+LAPACK_LIB = -L$(CONDA_LIB) -lopenblas
+
+# FFTW3_INC = -I$(CONDA_INC)/include/
+# FFTW3_LIB = -L$(CONDA_LIB) -lfftw3
+
+PTHREAD_INC = -DHAVE_UNISTD_H
+# PTHREAD_LIB = -lpthread
+
+CHOLMOD_INC = -I$(CONDA_INC)
+CHOLMOD_LIB = -L$(CONDA_LIB) -lcholmod -lamd -lcolamd -lcamd -lccolamd
+
+BOOST_INC = -I$(CONDA_PREFIX)/include
+BOOST_LIBS = -L$(CONDA_LIB) -lboost_serialization
+
+# Specify custom compilers if needed
+CXX = g++
+CC  = gcc
+
+#CFLAGS += -O3 -fPIC
+CFLAGS = -Wall -O3 -m64 -msse3 -msse2 -msse -fPIC
+
+OPTFLAGS = -O3
+
+OBJDIR = ./build
+S4_BINNAME = $(OBJDIR)/S4
+S4_LIBNAME = $(OBJDIR)/libS4.a
+S4r_LIBNAME = $(OBJDIR)/libS4r.a
+
+CPPFLAGS = -Wall -I. -IS4 -IS4/RNP -IS4/kiss_fft 
+ 
+ifdef BOOST_INC
+	CPPFLAGS += $(BOOST_INC) $(BOOST_LIBS)
+endif
+
+ifdef BLAS_LIB
+CPPFLAGS += -DHAVE_BLAS
+endif
+
+ifdef LAPACK_LIB
+CPPFLAGS += -DHAVE_LAPACK
+endif
+
+ifdef FFTW3_LIB
+CPPFLAGS += -DHAVE_FFTW3 $(FFTW3_INC)
+endif
+
+ifdef PTHREAD_LIB
+CPPFLAGS += -DHAVE_LIBPTHREAD $(PTHREAD_INC)
+endif
+
+ifdef CHOLMOD_LIB
+CPPFLAGS += -DHAVE_LIBCHOLMOD $(CHOLMOD_INC)
+endif
+
+ifdef MPI_LIB
+CPPFLAGS += -DHAVE_MPI $(MPI_INC)
+endif
+
+LIBS = $(BLAS_LIB) $(LAPACK_LIB) $(FFTW3_LIB) $(PTHREAD_LIB) $(CHOLMOD_LIB) $(MPI_LIB) $(BOOST_LIBS)
+
+#### Compilation targets
+
+all: $(S4_LIBNAME)
+
+objdir:
+	mkdir -p $(OBJDIR)
+	mkdir -p $(OBJDIR)/S4k
+	mkdir -p $(OBJDIR)/S4r
+	mkdir -p $(OBJDIR)/modules
+	
+S4_LIBOBJS = \
+	$(OBJDIR)/S4k/S4.o \
+	$(OBJDIR)/S4k/rcwa.o \
+	$(OBJDIR)/S4k/fmm_common.o \
+	$(OBJDIR)/S4k/fmm_FFT.o \
+	$(OBJDIR)/S4k/fmm_kottke.o \
+	$(OBJDIR)/S4k/fmm_closed.o \
+	$(OBJDIR)/S4k/fmm_PolBasisNV.o \
+	$(OBJDIR)/S4k/fmm_PolBasisVL.o \
+	$(OBJDIR)/S4k/fmm_PolBasisJones.o \
+	$(OBJDIR)/S4k/fmm_experimental.o \
+	$(OBJDIR)/S4k/fft_iface.o \
+	$(OBJDIR)/S4k/pattern.o \
+	$(OBJDIR)/S4k/intersection.o \
+	$(OBJDIR)/S4k/predicates.o \
+	$(OBJDIR)/S4k/numalloc.o \
+	$(OBJDIR)/S4k/gsel.o \
+	$(OBJDIR)/S4k/sort.o \
+	$(OBJDIR)/S4k/kiss_fft.o \
+	$(OBJDIR)/S4k/kiss_fftnd.o \
+	$(OBJDIR)/S4k/SpectrumSampler.o \
+	$(OBJDIR)/S4k/cubature.o \
+	$(OBJDIR)/S4k/Interpolator.o \
+	$(OBJDIR)/S4k/convert.o
+
+S4r_LIBOBJS = \
+	$(OBJDIR)/S4r/Material.o \
+	$(OBJDIR)/S4r/LatticeGridRect.o \
+	$(OBJDIR)/S4r/LatticeGridArb.o \
+	$(OBJDIR)/S4r/POFF2Mesh.o \
+	$(OBJDIR)/S4r/PeriodicMesh.o \
+	$(OBJDIR)/S4r/Shape.o \
+	$(OBJDIR)/S4r/Simulation.o \
+	$(OBJDIR)/S4r/Layer.o \
+	$(OBJDIR)/S4r/Pseudoinverse.o \
+	$(OBJDIR)/S4r/Eigensystems.o \
+	$(OBJDIR)/S4r/IRA.o \
+	$(OBJDIR)/S4r/intersection.o \
+	$(OBJDIR)/S4r/predicates.o \
+	$(OBJDIR)/S4r/periodic_off2.o
+
+ifndef LAPACK_LIB
+  S4_LIBOBJS += $(OBJDIR)/S4k/Eigensystems.o
+endif
+
+$(S4_LIBNAME): objdir $(S4_LIBOBJS)
+	$(AR) crvs $@ $(S4_LIBOBJS)
+
+$(OBJDIR)/S4k/S4.o: S4/S4.cpp
+	$(CXX) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+$(OBJDIR)/S4k/rcwa.o: S4/rcwa.cpp
+	$(CXX) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+$(OBJDIR)/S4k/fmm_common.o: S4/fmm/fmm_common.cpp
+	$(CXX) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+$(OBJDIR)/S4k/fmm_FFT.o: S4/fmm/fmm_FFT.cpp
+	$(CXX) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+$(OBJDIR)/S4k/fmm_kottke.o: S4/fmm/fmm_kottke.cpp
+	$(CXX) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+$(OBJDIR)/S4k/fmm_closed.o: S4/fmm/fmm_closed.cpp
+	$(CXX) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+$(OBJDIR)/S4k/fmm_PolBasisNV.o: S4/fmm/fmm_PolBasisNV.cpp
+	$(CXX) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+$(OBJDIR)/S4k/fmm_PolBasisVL.o: S4/fmm/fmm_PolBasisVL.cpp
+	$(CXX) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+$(OBJDIR)/S4k/fmm_PolBasisJones.o: S4/fmm/fmm_PolBasisJones.cpp
+	$(CXX) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+$(OBJDIR)/S4k/fmm_experimental.o: S4/fmm/fmm_experimental.cpp
+	$(CXX) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+$(OBJDIR)/S4k/fft_iface.o: S4/fmm/fft_iface.cpp
+	$(CXX) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+$(OBJDIR)/S4k/pattern.o: S4/pattern/pattern.c
+	$(CC) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+$(OBJDIR)/S4k/intersection.o: S4/pattern/intersection.c
+	$(CC) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+$(OBJDIR)/S4k/predicates.o: S4/pattern/predicates.c
+	$(CC) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+$(OBJDIR)/S4k/numalloc.o: S4/numalloc.c
+	$(CC) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+$(OBJDIR)/S4k/gsel.o: S4/gsel.c
+	$(CC) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+$(OBJDIR)/S4k/sort.o: S4/sort.c
+	$(CC) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+$(OBJDIR)/S4k/kiss_fft.o: S4/kiss_fft/kiss_fft.c
+	$(CC) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+$(OBJDIR)/S4k/kiss_fftnd.o: S4/kiss_fft/tools/kiss_fftnd.c
+	$(CC) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+$(OBJDIR)/S4k/SpectrumSampler.o: S4/SpectrumSampler.c
+	$(CC) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+$(OBJDIR)/S4k/cubature.o: S4/cubature.c
+	$(CC) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+$(OBJDIR)/S4k/Interpolator.o: S4/Interpolator.c
+	$(CC) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+$(OBJDIR)/S4k/convert.o: S4/convert.c
+	$(CC) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+$(OBJDIR)/S4k/Eigensystems.o: S4/RNP/Eigensystems.cpp
+	$(CXX) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
+
+#### Python extension
+
+S4_pyext: objdir $(S4_LIBNAME)
+	sh conda_recipe/s4/gensetup.py.sh $(OBJDIR) $(S4_LIBNAME) "$(LIBS)" $(CONDA_LIB)
+	$(PYTHON) setup.py install
+
+clean:
+	rm -rf $(OBJDIR)
diff --git a/conda_recipe/s4/bld.bat b/conda_recipe/s4/bld.bat
new file mode 100644
index 00000000..7dd00058
--- /dev/null
+++ b/conda_recipe/s4/bld.bat
@@ -0,0 +1,130 @@
+@echo on
+
+setlocal EnableDelayedExpansion
+
+::if "%PY_VER%"=="2.7" (
+::    set MSVC_VER=9.0
+::    set LIB_VER=90
+::) else if "%PY_VER%"=="3.4" (
+::    set MSVC_VER=10.0
+::    set LIB_VER=100
+::) else (
+::    set MSVC_VER=14.0
+::    set LIB_VER=140
+::)
+
+set DISTUTILS_USE_SDK=1
+::CALL "C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\vcvarsall.bat" amd64
+CALL "C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Auxiliary\Build/vcvarsall.bat" amd64 -vcvars_ver=14.0
+@echo on
+
+set "BLAS_PATH=%CONDA_PREFIX%/Library/lib/"
+set "BLAS_LIB=mkl_rt"
+set "LAPACK_PATH=%CONDA_PREFIX%/Library/lib/"
+set "LAPACK_LIB=mkl_rt"
+
+set "FFTW3_INC=-I%CONDA_PREFIX%/Library/include/fftw"
+set "FFTW3_PATH=%CONDA_PREFIX%/Library/lib/"
+set "FFTW3_LIB=mkl_rt"
+
+::set "PTHREAD_PATH=%CONDA_PREFIX%/Library/lib/"
+::set "PTHREAD_LIB=pthread"
+
+set "CHOLMOD_INC=-I%CONDA_PREFIX%\Library\include\suitesparse"
+set "CHOLMOD_PATH=%CONDA_PREFIX%\Library\lib\"
+set "CHOLMOD_LIB=cholmod amd colamd camd ccolamd suitesparseconfig metis"
+
+set "BOOST_INC=-I%CONDA_PREFIX%\Library\include\"
+set "BOOST_PATH=%CONDA_PREFIX%\Library\lib\"
+set "BOOST_LIB="
+
+:: Specify custom compilers if needed
+set "CXX=cl"
+set "CC=cl"
+
+set "OBJDIR=build"
+set "S4_LIBNAME=%OBJDIR%/S4.lib"
+
+set "CPPFLAGS=-O2 -W2 -EHs -LD -MD -DCPU86 -I. -IS4 -IS4/RNP -IS4/kiss_fft"
+::enable debug output
+::set "CPPFLAGS=%CPPFLAGS% -DENABLE_S4_TRACE"
+set "CPPFLAGS=%CPPFLAGS% %BOOST_INC%"
+IF DEFINED BLAS_LIB (set "CPPFLAGS=%CPPFLAGS% -DHAVE_BLAS")
+IF DEFINED LAPACK_LIB (set "CPPFLAGS=%CPPFLAGS% -DHAVE_LAPACK")
+IF DEFINED FFTW3_LIB (set "CPPFLAGS=%CPPFLAGS% -DHAVE_FFTW3 %FFTW3_INC%")
+IF DEFINED PTHREAD_LIB (set "CPPFLAGS=%CPPFLAGS% -DHAVE_LIBPTHREAD %PTHREAD_INC%")
+IF DEFINED CHOLMOD_LIB (set "CPPFLAGS=%CPPFLAGS% -DHAVE_LIBCHOLMOD %CHOLMOD_INC%")
+
+set "LIBS=%BLAS_LIB% %LAPACK_LIB% %FFTW3_LIB% %PTHREAD_LIB% %CHOLMOD_LIB% %BOOST_LIB%"
+set "LIBPATHS=%BLAS_PATH% %LAPACK_PATH% %FFTW3_PATH% %PTHREAD_PATH% %CHOLMOD_PATH% %BOOST_PATH%"
+
+:: Make a build folder and change to it.
+rmdir /Q /s %OBJDIR%
+mkdir %OBJDIR%
+cd %OBJDIR%
+mkdir S4k
+cd ..
+%CXX% -c %CPPFLAGS% S4/S4.cpp -Fo%OBJDIR%/S4k/S4.obj
+%CXX% -c %CPPFLAGS% S4/rcwa.cpp -Fo%OBJDIR%/S4k/rcwa.obj
+%CXX% -c %CPPFLAGS% S4/fmm/fmm_common.cpp -Fo%OBJDIR%/S4k/fmm_common.obj
+%CXX% -c %CPPFLAGS% S4/fmm/fmm_FFT.cpp -Fo%OBJDIR%/S4k/fmm_FFT.obj
+%CXX% -c %CPPFLAGS% S4/fmm/fmm_kottke.cpp -Fo%OBJDIR%/S4k/fmm_kottke.obj
+%CXX% -c %CPPFLAGS% S4/fmm/fmm_closed.cpp -Fo%OBJDIR%/S4k/fmm_closed.obj
+%CXX% -c %CPPFLAGS% S4/fmm/fmm_PolBasisNV.cpp -Fo%OBJDIR%/S4k/fmm_PolBasisNV.obj
+%CXX% -c %CPPFLAGS% S4/fmm/fmm_PolBasisVL.cpp -Fo%OBJDIR%/S4k/fmm_PolBasisVL.obj
+%CXX% -c %CPPFLAGS% S4/fmm/fmm_PolBasisJones.cpp -Fo%OBJDIR%/S4k/fmm_PolBasisJones.obj
+%CXX% -c %CPPFLAGS% S4/fmm/fmm_experimental.cpp -Fo%OBJDIR%/S4k/fmm_experimental.obj
+%CXX% -c %CPPFLAGS% S4/fmm/fft_iface.cpp -Fo%OBJDIR%/S4k/fft_iface.obj
+%CC% -c %CPPFLAGS% S4/pattern/pattern.c -Fo%OBJDIR%/S4k/pattern.obj
+%CC% -c %CPPFLAGS% S4/pattern/intersection.c -Fo%OBJDIR%/S4k/intersection.obj
+%CC% -c %CPPFLAGS% S4/pattern/predicates.c -Fo%OBJDIR%/S4k/predicates.obj
+%CC% -c %CPPFLAGS% S4/numalloc.c -Fo%OBJDIR%/S4k/numalloc.obj
+%CC% -c %CPPFLAGS% S4/gsel.c -Fo%OBJDIR%/S4k/gsel.obj
+%CC% -c %CPPFLAGS% S4/sort.c -Fo%OBJDIR%/S4k/sort.obj
+IF DEFINED FFTW3_LIB (echo "using fftw") ELSE (%CC% -c %CPPFLAGS% S4/kiss_fft/kiss_fft.c -Fo%OBJDIR%/S4k/kiss_fft.obj)
+IF DEFINED FFTW3_LIB (echo "using fftw") ELSE (%CC% -c %CPPFLAGS% S4/kiss_fft/tools/kiss_fftnd.c -Fo%OBJDIR%/S4k/kiss_fftnd.obj)
+%CC% -c %CPPFLAGS% S4/SpectrumSampler.c -Fo%OBJDIR%/S4k/SpectrumSampler.obj
+::%CC% -c %CPPFLAGS% S4/cubature.c -Fo%OBJDIR%/S4k/cubature.obj
+%CC% -c %CPPFLAGS% S4/Interpolator.c -Fo%OBJDIR%/S4k/Interpolator.obj
+::%CC% -c %CPPFLAGS% S4/convert.c -Fo%OBJDIR%/S4k/convert.obj
+IF DEFINED LAPACK_LIB (echo "using lapack") ELSE (%CXX% -c %CPPFLAGS% S4/RNP/Eigensystems.cpp -Fo%OBJDIR%/S4k/Eigensystems.obj)
+
+lib.exe %OBJDIR%/S4k/* /out:%S4_LIBNAME%
+
+
+set "LIBS=%LIBS:\=/%"
+set "LIBPATHS=%LIBPATHS:\=/%"
+set "OBJDIR=%OBJDIR:\=/%"
+set "S4_LIBNAME=%S4_LIBNAME:\=/%"
+
+(echo:from distutils.core import setup, Extension
+echo:import numpy as np
+echo:
+echo:lib_dirs = ['%OBJDIR%']
+echo:lib_dirs.extend^([libpath for libpath in '%LIBPATHS%'.split^(^)]^)
+echo:libs = ['S4']
+echo:libs.extend^([lib for lib in '%LIBS%'.split^(^)]^)
+echo:include_dirs = [np.get_include^(^)]
+echo:extra_compile_args = ["-LD", "-MD"]
+echo:extra_link_args = []
+::echo:extra_link_args = ["/DEBUG", "/verbose:lib"]
+echo:print^(lib_dirs^)
+echo:print^(libs^)
+echo:print^(include_dirs^)
+echo:S4module = Extension^('S4',
+echo:	sources = ['S4/main_python.c'],
+echo:	libraries = libs,
+echo:	library_dirs = lib_dirs,
+echo:	include_dirs = include_dirs,
+echo:	extra_objects = ['%S4_LIBNAME%'],
+echo:   extra_compile_args = extra_compile_args,
+echo:   extra_link_args = extra_link_args
+echo:^)
+echo:
+echo:setup^(name = 'S4',
+echo:	version = '1.1',
+echo:	description = 'Stanford Stratified Structure Solver ^(S4^): Fourier Modal Method',
+echo:	ext_modules = [S4module]
+echo:^))>setup.py
+
+IF DEFINED PYTHON ("%PYTHON%" setup.py install) ELSE (python setup.py install)
\ No newline at end of file
diff --git a/conda_recipe/s4/build.sh b/conda_recipe/s4/build.sh
new file mode 100644
index 00000000..55382e3d
--- /dev/null
+++ b/conda_recipe/s4/build.sh
@@ -0,0 +1,3 @@
+#!/bin/bash
+make -f conda_recipe/s4/Makefile clean
+make -f conda_recipe/s4/Makefile S4_pyext
\ No newline at end of file
diff --git a/conda_recipe/s4/conda_build_config.yaml b/conda_recipe/s4/conda_build_config.yaml
new file mode 100644
index 00000000..07b65fe8
--- /dev/null
+++ b/conda_recipe/s4/conda_build_config.yaml
@@ -0,0 +1,5 @@
+python:
+  - 3.7
+  - 3.8
+numpy:
+  - 1.18
\ No newline at end of file
diff --git a/conda_recipe/s4/gensetup.py.sh b/conda_recipe/s4/gensetup.py.sh
new file mode 100644
index 00000000..88d9a459
--- /dev/null
+++ b/conda_recipe/s4/gensetup.py.sh
@@ -0,0 +1,39 @@
+#!/bin/bash
+
+OBJDIR="$1"
+LIBFILE="$2"
+LIBS="$3"
+BOOST_LIBS="$4"
+
+echo "LIBFILE: $LIBFILE"
+
+cat <<SETUPPY > setup.py
+from distutils.core import setup, Extension
+import numpy as np
+#import os
+#os.environ["CC"] = "g++"
+#os.environ["CXX"] = "g++"
+
+libs = ['S4', 'stdc++']
+lib_dirs = ['$OBJDIR', '$BOOST_LIBS']
+libs.extend([lib[2::] for lib in '$LIBS'.split()])
+include_dirs = ['$BOOST_PREFIX/include', np.get_include()]
+extra_link_args = ['$LIBFILE']
+
+S4module = Extension('S4',
+	sources = ['S4/main_python.c'],
+	libraries = libs,
+	library_dirs = lib_dirs,
+    include_dirs = include_dirs,
+    extra_objects = ['$LIBFILE'],
+	# extra_link_args = extra_link_args,
+    runtime_library_dirs=['$BOOST_PREFIX/lib'],
+	extra_compile_args=['-std=gnu99'] 
+)
+
+setup(name = 'S4',
+	version = '1.1',
+	description = 'Stanford Stratified Structure Solver (S4): Fourier Modal Method',
+	ext_modules = [S4module]
+)
+SETUPPY
diff --git a/conda_recipe/s4/meta.yaml b/conda_recipe/s4/meta.yaml
new file mode 100644
index 00000000..f6dd4fd6
--- /dev/null
+++ b/conda_recipe/s4/meta.yaml
@@ -0,0 +1,67 @@
+
+package:
+  name: s4
+  version: 1.1
+
+source:
+  path: ../../
+
+build:
+  number: 5
+
+requirements:
+  build:
+    - {{ compiler('c') }}
+    - {{ compiler('cxx') }}
+    - python
+    - setuptools
+    - numpy
+    - libboost
+    - suitesparse
+    - mkl-devel [win]
+    - openblas [unix]
+
+  host:
+    - python
+    - numpy
+    - libboost
+    - suitesparse
+    - mkl-devel [win]
+    - openblas [unix]
+
+  run:
+    - python
+    - {{ pin_compatible('numpy') }}
+    - {{ pin_compatible('libboost') }}
+    - {{ pin_compatible('suitesparse') }}
+    - {{ pin_compatible('mkl') }} [win]
+    - {{ pin_compatible('openblas') }} [unix]
+
+test:
+  imports:
+    - S4
+
+about:
+  home: https://web.stanford.edu/group/fan/S4/index.html
+  license: GPL-2.0
+  license_family: GPL
+  # license_file: LICENSE.txt
+  summary: 'S4 (Stanford Stratified Structure Solver) Conda Package'
+
+  # The remaining entries in this section are optional, but recommended.
+  description: |
+    These are conda packages of the below software. 
+    Victor Liu wrote the initial code but is not responsible for these packages.
+    A program for computing electromagnetic fields in periodic, layered
+    structures, developed by Victor Liu (victorliu@alumni.stanford.edu) of the
+    Fan group in the Stanford Electrical Engineering Department.
+    See the S4 manual, in doc/index.html, for a complete
+    description of the package and its user interface, as well as
+    installation instructions, the license and copyright, contact
+    addresses, and other important information.
+  doc_url: https://web.stanford.edu/group/fan/S4/index.html
+  dev_url: https://web.stanford.edu/group/fan/S4/index.html
+
+extra:
+  recipe-maintainers:
+    - marcus-o
\ No newline at end of file
diff --git a/examples/C_api/spec.awk b/examples/C_api/spec.awk
old mode 100755
new mode 100644