diff --git a/build_linux64/libamscppnarray.linux64.a b/build_linux64/libamscppnarray.linux64.a
index 04e2098..eb22e83 100644
Binary files a/build_linux64/libamscppnarray.linux64.a and b/build_linux64/libamscppnarray.linux64.a differ
diff --git a/build_linux64/objstore/amscppnarray.o b/build_linux64/objstore/amscppnarray.o
index 20758ff..c482dc4 100644
Binary files a/build_linux64/objstore/amscppnarray.o and b/build_linux64/objstore/amscppnarray.o differ
diff --git a/build_linux64/tests b/build_linux64/tests
index 1beebae..1953c10 100644
Binary files a/build_linux64/tests and b/build_linux64/tests differ
diff --git a/include/amscppnarray/amscppnarray.hpp b/include/amscppnarray/amscppnarray.hpp
index 70ec991..9bce8c9 100644
--- a/include/amscppnarray/amscppnarray.hpp
+++ b/include/amscppnarray/amscppnarray.hpp
@@ -14,6 +14,10 @@ namespace ams
 namespace narray
 {
 
+// Forward declaration of the operator- template function
+template<typename T> class narray;
+template<typename T> narray<T> operator-(const narray<T> &n);
+
 typedef int64_t narray_size_t;
 
 static const int narray_success = 1;
@@ -33,9 +37,14 @@ template<typename T> class narray
 
     narray();
     narray(const narray<T>& other);
+    narray(narray<T>&& other) noexcept;
+
     narray<T>& operator=(const narray<T>& other);
+    narray<T>& operator=(narray<T>&& other) noexcept;
+    
     narray<T>& operator=(const std::vector<T>& other);
     narray(const std::initializer_list<T> initlist);
+    narray<T>& operator=(const std::initializer_list<T> initlist);
     ~narray();
 
     int resize(const narray_size_t newlength);
@@ -49,18 +58,35 @@ template<typename T> class narray
     void setall(const T& val); //sets all elements to val
 
     //Comparators
+    bool operator==(const narray<T>& other) const;
+    bool operator!=(const narray<T>& other) const;
 
     //Operations
-
+    narray<T> operator+(const narray<T>& other) const;
+    narray<T> operator-(const narray<T>& other) const;
+    narray<T> operator*(const narray<T>& other) const;
+    narray<T> operator/(const narray<T>& other) const;
+    narray<T> operator+(const T val) const;
+    narray<T> operator-(const T val) const;
+    narray<T> operator*(const T val) const;
+    narray<T> operator/(const T val) const;
+    template<class U> friend narray<U> operator-(const narray<U> &n);
+    
+    //subarray operations
+
+    //permutations and sorting
 };
 
 
 
 void test_narray1();
+void test_narray2();
+void test_narray3();
 
 }; //end namespace narray
 }; //end namespace ams
 
 #include <amscppnarray/amscppnarray_impl.hpp>
+#include <amscppnarray/amscppnarray_imploper.hpp>
 
 #endif
\ No newline at end of file
diff --git a/include/amscppnarray/amscppnarray_impl.hpp b/include/amscppnarray/amscppnarray_impl.hpp
index b766755..30b6a56 100644
--- a/include/amscppnarray/amscppnarray_impl.hpp
+++ b/include/amscppnarray/amscppnarray_impl.hpp
@@ -8,6 +8,8 @@ namespace narray
 
 template<typename T> narray<T>::narray()
 {
+    //printf("debug: constructor called for %p.\n",this);
+
     length = 0;
     data = NULL;
 
@@ -59,6 +61,8 @@ template<typename T> narray<T>::narray(const narray<T>& other)
     length = 0;
     data = NULL;
 
+    //printf("debug: copy constructor called for\n\t%p\n\t%p\n", this, &other);
+
     if(this!=&other)
     {
         res = this->resize(other.length);
@@ -76,12 +80,48 @@ template<typename T> narray<T>::narray(const narray<T>& other)
     }
 }
 
+template<typename T> narray<T>::narray(narray<T>&& other) noexcept
+{
+    //this is a move operator
+    //I still need to understand when these are called.
+
+    //printf("debug: move constructor called for\n\t%p\n\t%p\n", this, &other);
+
+    //steal the resources from other and reset other to prevent double deletion
+    length = other.length;
+    data = other.data;
+
+    other.length = 0;
+    other.data = NULL;
+
+    return;
+}
+
+template<typename T> narray<T>& narray<T>::operator=(narray<T>&& other) noexcept
+{
+
+    //printf("debug: move= called for\n\t%p\n\t%p\n", this, &other);
+
+    if(this!=&other)
+    {
+        if(data!=NULL) {delete[] data; data=NULL;}
+        
+        length = other.length;
+        data = other.data;
+
+        other.length = 0;
+        other.data = NULL;
+    }
+    return *this;
+}
 
 template<typename T> narray<T>& narray<T>::operator=(const narray<T>& other)
 {
     narray_size_t I;
     int res;
 
+    //printf("debug: operator= called for\n\t%p\n\t%p\n", this, &other);
+
     if(this!=&other)
     {
         res = this->resize(other.length);
@@ -116,7 +156,7 @@ template<typename T> narray<T>& narray<T>::operator=(const std::vector<T>& other
         {
             for(I=0;I<length && I<other.size();I++)
             {
-                data[I] = other.data[I];
+                data[I] = other[I];
             }
         }
     }
@@ -146,8 +186,34 @@ template<typename T> narray<T>::narray(const std::initializer_list<T> initlist)
     }
 }
 
+template<typename T> narray<T>& narray<T>::operator=(const std::initializer_list<T> initlist)
+{
+    narray_size_t I;
+    int res;
+
+    res = this->resize(initlist.size());
+    if(res!=narray_success)
+    {
+        this->length = 0;
+        if(this->data != NULL) {delete[] this->data; this->data=NULL;}
+        return *this;
+    }
+    else
+    {
+        I = 0;
+        for(T elem : initlist)
+        {
+            data[I] = elem;
+            if(I<length) I++;
+        }
+    }
+
+    return *this;
+}
+
 template<typename T> narray<T>::~narray()
 {
+    //printf("debug: destructor called for %p\n",this);
     length = 0;
     if(data!=NULL) {delete[] data; data=NULL;}
     return;
@@ -179,10 +245,27 @@ template<typename T> void narray<T>::clear()
     return;
 }
 
+template<typename T> void narray_setall_tf(narray<T> *q, const T& val, int threadnum, int nthreads)
+{
+    int I0,I1,Is,I;
+
+    Is = q->length/nthreads;
+    I0 = Is*threadnum;
+    I1 = (threadnum>=(nthreads-1))? q->length: Is*(threadnum+1);
+    for(I=I0;I<I1;I++)
+    {
+        q->data[I] = val;
+    }
+
+    return;
+}
+
 template<typename T> void narray<T>::setall(const T& val)
 {
     narray_size_t I;
+    int J;
     int nthreads;
+    std::vector<std::thread*> threads;
 
     if(length<narray_thread_sz)
     {
@@ -191,8 +274,25 @@ template<typename T> void narray<T>::setall(const T& val)
     else
     {
         //threaded operation
-        nthreads = std::thread::hardware_concurrency;
-        
+        nthreads = std::thread::hardware_concurrency();
+        if(nthreads<1) nthreads=1;
+        if(nthreads>narray_max_threads) nthreads = narray_max_threads;
+        threads.resize(nthreads);
+        for(J=0;J<nthreads;J++)
+        {
+            threads[J] = new(std::nothrow) std::thread(
+                &narray_setall_tf<T>,this,val,J,nthreads
+            );
+        }
+        for(J=0;J<nthreads;J++)
+        {
+            if(threads[J]!=NULL)
+            { 
+                threads[J]->join(); 
+                delete threads[J]; 
+                threads[J]= NULL;
+            }
+        }
     }
     return;
 }
diff --git a/include/amscppnarray/amscppnarray_imploper.hpp b/include/amscppnarray/amscppnarray_imploper.hpp
new file mode 100644
index 0000000..9efc485
--- /dev/null
+++ b/include/amscppnarray/amscppnarray_imploper.hpp
@@ -0,0 +1,736 @@
+#ifndef __AMSCPPNARRAY_IMPLOPER_HPP__
+#define __AMSCPPNARRAY_IMPLOPER_HPP__
+
+namespace ams
+{
+namespace narray
+{
+
+    template<typename T> void narray_operator_comp_tf( 
+        const narray<T> *a,
+        const narray<T> *b, 
+        narray<bool> *cmp,
+        int threadnum, int nthreads
+    )
+    {
+        int I0,I1,Is,I;
+    
+        Is = a->length/nthreads;
+        I0 = Is*threadnum;
+        I1 = (threadnum>=(nthreads-1))? a->length: Is*(threadnum+1);
+        cmp->data[threadnum] = 1;
+        for(I=I0;I<I1;I++)
+        {
+            if(a->data[I]!=b->data[I])
+            {
+                cmp->data[threadnum] = 0;
+                break;
+            }
+        }
+        return;
+    }
+
+    template<typename T> bool narray<T>::operator==(const narray<T>& other) const
+    {
+        bool ret = 1;
+        narray<bool> cmp;
+        narray_size_t I;
+        int J;
+        int nthreads;
+        std::vector<std::thread*> threads;
+
+        if(this->length != other.length)
+        {
+            ret = 0;
+            return ret;
+        }
+        if(this->length == 0)
+        {
+            ret = 1;
+            return ret;
+        }
+        if(this->length < narray_thread_sz)
+        {
+            for(I=0;I<this->length;I++)
+            {
+                if(this->data[I]!=other.data[I])
+                {
+                    ret = 0;
+                    break;
+                }
+            }
+        }
+        else
+        {
+            //threaded operation
+            nthreads = std::thread::hardware_concurrency();
+            nthreads = (nthreads<1) ? 1 : nthreads;
+            nthreads = (nthreads>narray_max_threads) ? narray_max_threads : nthreads;
+            //if(nthreads<1) nthreads=1;
+            //if(nthreads>narray_max_threads) nthreads = narray_max_threads;
+            threads.resize(nthreads);
+            cmp.resize(nthreads);
+            for(J=0;J<nthreads;J++)
+            {
+                //printf("debug: launching thread %d\n",J);
+                threads[J] = new(std::nothrow) std::thread(
+                    &narray_operator_comp_tf<T>,
+                    this,
+                    &other,
+                    &cmp,
+                    J,nthreads
+                );
+            }
+            for(J=0;J<nthreads;J++)
+            {
+                if(threads[J]!=NULL)
+                { 
+                    threads[J]->join(); 
+                    delete threads[J]; 
+                    threads[J]= NULL;
+                }
+            }
+
+            ret = 1;
+            for(J=0;J<nthreads;J++)
+            {
+                if(cmp[J]==0)
+                {
+                    ret = 0;
+                    break;
+                }
+            }
+
+            cmp.resize(0);
+        }
+
+        return ret;
+    }
+
+    template<typename T> bool narray<T>::operator!=(const narray<T>& other) const
+    {
+        return !(*this==other);
+    }
+
+    template<typename T> void narray_operator_add_tf( 
+        const narray<T> *a,
+        const narray<T> *b, 
+        narray<T> *out,
+        int threadnum, int nthreads
+    )
+    {
+        int I0,I1,Is,I;
+    
+        Is = a->length/nthreads;
+        I0 = Is*threadnum;
+        I1 = (threadnum>=(nthreads-1))? a->length: Is*(threadnum+1);
+        for(I=I0;I<I1;I++)
+        {
+            out->data[I] = a->data[I] + b->data[I];
+        }
+    
+        return;
+    }
+
+    template<typename T> narray<T> narray<T>::operator+(const narray<T>& other) const
+    {
+        narray<T> ret;
+        narray_size_t I;
+        int J;
+        int nthreads;
+        std::vector<std::thread*> threads;
+    
+        if(this->length!=other.length)
+        {
+            ret.resize(0);
+            printf("narray<T>::operator+:: error - array sizes %ld and %ld are not the same.\n",(long)this->length,(long)other.length);
+        }
+    
+        ret.resize(this->length);
+    
+        if(this->length<narray_thread_sz)
+        {
+            for(I=0;I<this->length;I++) ret.data[I] = this->data[I] + other.data[I];
+        }
+        else
+        {
+            //threaded operation
+            nthreads = std::thread::hardware_concurrency();
+            nthreads = (nthreads<1) ? 1 : nthreads;
+            nthreads = (nthreads>narray_max_threads) ? narray_max_threads : nthreads;
+            //if(nthreads<1) nthreads=1;
+            //if(nthreads>narray_max_threads) nthreads = narray_max_threads;
+            threads.resize(nthreads);
+            for(J=0;J<nthreads;J++)
+            {
+                //printf("debug: launching thread %d\n",J);
+                threads[J] = new(std::nothrow) std::thread(
+                    &narray_operator_add_tf<T>,
+                    this,
+                    &other,
+                    &ret,
+                    J,nthreads
+                );
+            }
+            for(J=0;J<nthreads;J++)
+            {
+                if(threads[J]!=NULL)
+                { 
+                    threads[J]->join(); 
+                    delete threads[J]; 
+                    threads[J]= NULL;
+                }
+            }
+        }
+        return ret;
+    }
+
+    template<typename T> void narray_operator_sub_tf( 
+        const narray<T> *a,
+        const narray<T> *b, 
+        narray<T> *out,
+        int threadnum, int nthreads
+    )
+    {
+        int I0,I1,Is,I;
+    
+        Is = a->length/nthreads;
+        I0 = Is*threadnum;
+        I1 = (threadnum>=(nthreads-1))? a->length: Is*(threadnum+1);
+        for(I=I0;I<I1;I++)
+        {
+            out->data[I] = a->data[I] - b->data[I];
+        }
+    
+        return;
+    }
+
+    template<typename T> narray<T> narray<T>::operator-(const narray<T>& other) const
+    {
+        narray<T> ret;
+        narray_size_t I;
+        int J;
+        int nthreads;
+        std::vector<std::thread*> threads;
+    
+        if(this->length!=other.length)
+        {
+            ret.resize(0);
+            printf("narray<T>::operator-:: error - array sizes %ld and %ld are not the same.\n",(long)this->length,(long)other.length);
+        }
+    
+        ret.resize(this->length);
+    
+        if(this->length<narray_thread_sz)
+        {
+            for(I=0;I<this->length;I++) ret.data[I] = this->data[I] + other.data[I];
+        }
+        else
+        {
+            //threaded operation
+            nthreads = std::thread::hardware_concurrency();
+            nthreads = (nthreads<1) ? 1 : nthreads;
+            nthreads = (nthreads>narray_max_threads) ? narray_max_threads : nthreads;
+            //if(nthreads<1) nthreads=1;
+            //if(nthreads>narray_max_threads) nthreads = narray_max_threads;
+            threads.resize(nthreads);
+            for(J=0;J<nthreads;J++)
+            {
+                threads[J] = new(std::nothrow) std::thread(
+                    &narray_operator_sub_tf<T>,
+                    this,
+                    &other,
+                    &ret,
+                    J,nthreads
+                );
+            }
+            for(J=0;J<nthreads;J++)
+            {
+                if(threads[J]!=NULL)
+                { 
+                    threads[J]->join(); 
+                    delete threads[J]; 
+                    threads[J]= NULL;
+                }
+            }
+        }
+        return ret;
+    }
+
+    template<typename T> void narray_operator_mult_tf( 
+        const narray<T> *a,
+        const narray<T> *b, 
+        narray<T> *out,
+        int threadnum, int nthreads
+    )
+    {
+        int I0,I1,Is,I;
+    
+        Is = a->length/nthreads;
+        I0 = Is*threadnum;
+        I1 = (threadnum>=(nthreads-1))? a->length: Is*(threadnum+1);
+        for(I=I0;I<I1;I++)
+        {
+            out->data[I] = a->data[I] * b->data[I];
+        }
+    
+        return;
+    }
+
+    template<typename T> narray<T> narray<T>::operator*(const narray<T>& other) const
+    {
+        narray<T> ret;
+        narray_size_t I;
+        int J;
+        int nthreads;
+        std::vector<std::thread*> threads;
+    
+        if(this->length!=other.length)
+        {
+            ret.resize(0);
+            printf("narray<T>::operator*:: error - array sizes %ld and %ld are not the same.\n",(long)this->length,(long)other.length);
+        }
+    
+        ret.resize(this->length);
+    
+        if(this->length<narray_thread_sz)
+        {
+            for(I=0;I<this->length;I++) ret.data[I] = this->data[I] * other.data[I];
+        }
+        else
+        {
+            //threaded operation
+            nthreads = std::thread::hardware_concurrency();
+            nthreads = (nthreads<1) ? 1 : nthreads;
+            nthreads = (nthreads>narray_max_threads) ? narray_max_threads : nthreads;
+            //if(nthreads<1) nthreads=1;
+            //if(nthreads>narray_max_threads) nthreads = narray_max_threads;
+            threads.resize(nthreads);
+            for(J=0;J<nthreads;J++)
+            {
+                threads[J] = new(std::nothrow) std::thread(
+                    &narray_operator_mult_tf<T>,
+                    this,
+                    &other,
+                    &ret,
+                    J,nthreads
+                );
+            }
+            for(J=0;J<nthreads;J++)
+            {
+                if(threads[J]!=NULL)
+                { 
+                    threads[J]->join(); 
+                    delete threads[J]; 
+                    threads[J]= NULL;
+                }
+            }
+        }
+        return ret;
+    }
+
+    template<typename T> void narray_operator_div_tf( 
+        const narray<T> *a,
+        const narray<T> *b, 
+        narray<T> *out,
+        int threadnum, int nthreads
+    )
+    {
+        int I0,I1,Is,I;
+    
+        Is = a->length/nthreads;
+        I0 = Is*threadnum;
+        I1 = (threadnum>=(nthreads-1))? a->length: Is*(threadnum+1);
+        for(I=I0;I<I1;I++)
+        {
+            out->data[I] = a->data[I] / b->data[I];
+        }
+    
+        return;
+    }
+
+    template<typename T> narray<T> narray<T>::operator/(const narray<T>& other) const
+    {
+        narray<T> ret;
+        narray_size_t I;
+        int J;
+        int nthreads;
+        std::vector<std::thread*> threads;
+    
+        if(this->length!=other.length)
+        {
+            ret.resize(0);
+            printf("narray<T>::operator/:: error - array sizes %ld and %ld are not the same.\n",(long)this->length,(long)other.length);
+        }
+    
+        ret.resize(this->length);
+    
+        if(this->length<narray_thread_sz)
+        {
+            for(I=0;I<this->length;I++) ret.data[I] = this->data[I] / other.data[I];
+        }
+        else
+        {
+            //threaded operation
+            nthreads = std::thread::hardware_concurrency();
+            nthreads = (nthreads<1) ? 1 : nthreads;
+            nthreads = (nthreads>narray_max_threads) ? narray_max_threads : nthreads;
+            //if(nthreads<1) nthreads=1;
+            //if(nthreads>narray_max_threads) nthreads = narray_max_threads;
+            threads.resize(nthreads);
+            for(J=0;J<nthreads;J++)
+            {
+                threads[J] = new(std::nothrow) std::thread(
+                    &narray_operator_div_tf<T>,
+                    this,
+                    &other,
+                    &ret,
+                    J,nthreads
+                );
+            }
+            for(J=0;J<nthreads;J++)
+            {
+                if(threads[J]!=NULL)
+                { 
+                    threads[J]->join(); 
+                    delete threads[J]; 
+                    threads[J]= NULL;
+                }
+            }
+        }
+        return ret;
+    }
+    
+    template<typename T> void narray_operator_add2_tf( 
+        const narray<T> *a,
+        const T val,
+        narray<T> *out,
+        int threadnum, int nthreads
+    )
+    {
+        int I0,I1,Is,I;
+    
+        Is = a->length/nthreads;
+        I0 = Is*threadnum;
+        I1 = (threadnum>=(nthreads-1))? a->length: Is*(threadnum+1);
+        for(I=I0;I<I1;I++)
+        {
+            out->data[I] = a->data[I] + val;
+        }
+    
+        return;
+    }
+
+    template<typename T> narray<T> narray<T>::operator+(const T val) const
+    {
+        narray<T> ret;
+        narray_size_t I;
+        int J;
+        int nthreads;
+        std::vector<std::thread*> threads;
+    
+        ret.resize(this->length);
+    
+        if(this->length<narray_thread_sz)
+        {
+            for(I=0;I<this->length;I++) ret.data[I] = this->data[I] + val;
+        }
+        else
+        {
+            //threaded operation
+            nthreads = std::thread::hardware_concurrency();
+            nthreads = (nthreads<1) ? 1 : nthreads;
+            nthreads = (nthreads>narray_max_threads) ? narray_max_threads : nthreads;
+            //if(nthreads<1) nthreads=1;
+            //if(nthreads>narray_max_threads) nthreads = narray_max_threads;
+            threads.resize(nthreads);
+            for(J=0;J<nthreads;J++)
+            {
+                threads[J] = new(std::nothrow) std::thread(
+                    &narray_operator_add2_tf<T>,
+                    this,
+                    val,
+                    &ret,
+                    J,nthreads
+                );
+            }
+            for(J=0;J<nthreads;J++)
+            {
+                if(threads[J]!=NULL)
+                { 
+                    threads[J]->join(); 
+                    delete threads[J]; 
+                    threads[J]= NULL;
+                }
+            }
+        }
+        return ret;
+    }
+
+    template<typename T> void narray_operator_sub2_tf( 
+        const narray<T> *a,
+        const T val,
+        narray<T> *out,
+        int threadnum, int nthreads
+    )
+    {
+        int I0,I1,Is,I;
+    
+        Is = a->length/nthreads;
+        I0 = Is*threadnum;
+        I1 = (threadnum>=(nthreads-1))? a->length: Is*(threadnum+1);
+        for(I=I0;I<I1;I++)
+        {
+            out->data[I] = a->data[I] - val;
+        }
+    
+        return;
+    }
+
+    template<typename T> narray<T> narray<T>::operator-(const T val) const
+    {
+        narray<T> ret;
+        narray_size_t I;
+        int J;
+        int nthreads;
+        std::vector<std::thread*> threads;
+    
+        ret.resize(this->length);
+    
+        if(this->length<narray_thread_sz)
+        {
+            for(I=0;I<this->length;I++) ret.data[I] = this->data[I] - val;
+        }
+        else
+        {
+            //threaded operation
+            nthreads = std::thread::hardware_concurrency();
+            nthreads = (nthreads<1) ? 1 : nthreads;
+            nthreads = (nthreads>narray_max_threads) ? narray_max_threads : nthreads;
+            //if(nthreads<1) nthreads=1;
+            //if(nthreads>narray_max_threads) nthreads = narray_max_threads;
+            threads.resize(nthreads);
+            for(J=0;J<nthreads;J++)
+            {
+                threads[J] = new(std::nothrow) std::thread(
+                    &narray_operator_sub2_tf<T>,
+                    this,
+                    val,
+                    &ret,
+                    J,nthreads
+                );
+            }
+            for(J=0;J<nthreads;J++)
+            {
+                if(threads[J]!=NULL)
+                { 
+                    threads[J]->join(); 
+                    delete threads[J]; 
+                    threads[J]= NULL;
+                }
+            }
+        }
+        return ret;
+    }
+
+    template<typename T> void narray_operator_mult2_tf( 
+        const narray<T> *a,
+        const T val,
+        narray<T> *out,
+        int threadnum, int nthreads
+    )
+    {
+        int I0,I1,Is,I;
+    
+        Is = a->length/nthreads;
+        I0 = Is*threadnum;
+        I1 = (threadnum>=(nthreads-1))? a->length: Is*(threadnum+1);
+        for(I=I0;I<I1;I++)
+        {
+            out->data[I] = a->data[I] * val;
+        }
+    
+        return;
+    }
+
+    template<typename T> narray<T> narray<T>::operator*(const T val) const
+    {
+        narray<T> ret;
+        narray_size_t I;
+        int J;
+        int nthreads;
+        std::vector<std::thread*> threads;
+    
+        ret.resize(this->length);
+    
+        if(this->length<narray_thread_sz)
+        {
+            for(I=0;I<this->length;I++) ret.data[I] = this->data[I] * val;
+        }
+        else
+        {
+            //threaded operation
+            nthreads = std::thread::hardware_concurrency();
+            nthreads = (nthreads<1) ? 1 : nthreads;
+            nthreads = (nthreads>narray_max_threads) ? narray_max_threads : nthreads;
+            //if(nthreads<1) nthreads=1;
+            //if(nthreads>narray_max_threads) nthreads = narray_max_threads;
+            threads.resize(nthreads);
+            for(J=0;J<nthreads;J++)
+            {
+                threads[J] = new(std::nothrow) std::thread(
+                    &narray_operator_mult2_tf<T>,
+                    this,
+                    val,
+                    &ret,
+                    J,nthreads
+                );
+            }
+            for(J=0;J<nthreads;J++)
+            {
+                if(threads[J]!=NULL)
+                { 
+                    threads[J]->join(); 
+                    delete threads[J]; 
+                    threads[J]= NULL;
+                }
+            }
+        }
+        return ret;
+    }
+
+    template<typename T> void narray_operator_div2_tf( 
+        const narray<T> *a,
+        const T val,
+        narray<T> *out,
+        int threadnum, int nthreads
+    )
+    {
+        int I0,I1,Is,I;
+    
+        Is = a->length/nthreads;
+        I0 = Is*threadnum;
+        I1 = (threadnum>=(nthreads-1))? a->length: Is*(threadnum+1);
+        for(I=I0;I<I1;I++)
+        {
+            out->data[I] = a->data[I] / val;
+        }
+    
+        return;
+    }
+
+    template<typename T> narray<T> narray<T>::operator/(const T val) const
+    {
+        narray<T> ret;
+        narray_size_t I;
+        int J;
+        int nthreads;
+        std::vector<std::thread*> threads;
+    
+        ret.resize(this->length);
+    
+        if(this->length<narray_thread_sz)
+        {
+            for(I=0;I<this->length;I++) ret.data[I] = this->data[I] / val;
+        }
+        else
+        {
+            //threaded operation
+            nthreads = std::thread::hardware_concurrency();
+            nthreads = (nthreads<1) ? 1 : nthreads;
+            nthreads = (nthreads>narray_max_threads) ? narray_max_threads : nthreads;
+            //if(nthreads<1) nthreads=1;
+            //if(nthreads>narray_max_threads) nthreads = narray_max_threads;
+            threads.resize(nthreads);
+            for(J=0;J<nthreads;J++)
+            {
+                threads[J] = new(std::nothrow) std::thread(
+                    &narray_operator_div2_tf<T>,
+                    this,
+                    val,
+                    &ret,
+                    J,nthreads
+                );
+            }
+            for(J=0;J<nthreads;J++)
+            {
+                if(threads[J]!=NULL)
+                { 
+                    threads[J]->join(); 
+                    delete threads[J]; 
+                    threads[J]= NULL;
+                }
+            }
+        }
+        return ret;
+    }
+
+
+    template<typename T> void narray_operator_negative_tf( 
+        const narray<T> *a,
+        narray<T> *out,
+        int threadnum, int nthreads
+    )
+    {
+        int I0,I1,Is,I;
+    
+        Is = a->length/nthreads;
+        I0 = Is*threadnum;
+        I1 = (threadnum>=(nthreads-1))? a->length: Is*(threadnum+1);
+        for(I=I0;I<I1;I++)
+        {
+            out->data[I] = -a->data[I];
+        }
+    
+        return;
+    }
+
+    template<class T> narray<T> operator-(const narray<T> &n)
+    {
+        narray<T> ret;
+        narray_size_t I;
+        int J;
+        int nthreads;
+        std::vector<std::thread*> threads;
+    
+        ret.resize(n.length);
+    
+        if(n.length<narray_thread_sz)
+        {
+            for(I=0;I<n.length;I++) ret.data[I] = -n.data[I];
+        }
+        else
+        {
+            //threaded operation
+            nthreads = std::thread::hardware_concurrency();
+            nthreads = (nthreads<1) ? 1 : nthreads;
+            nthreads = (nthreads>narray_max_threads) ? narray_max_threads : nthreads;
+            //if(nthreads<1) nthreads=1;
+            //if(nthreads>narray_max_threads) nthreads = narray_max_threads;
+            threads.resize(nthreads);
+            for(J=0;J<nthreads;J++)
+            {
+                threads[J] = new(std::nothrow) std::thread(
+                    &narray_operator_negative_tf<T>,
+                    &n,
+                    &ret,
+                    J,nthreads
+                );
+            }
+            for(J=0;J<nthreads;J++)
+            {
+                if(threads[J]!=NULL)
+                { 
+                    threads[J]->join(); 
+                    delete threads[J]; 
+                    threads[J]= NULL;
+                }
+            }
+        }
+        return ret;
+    }
+
+}; //end namespace narray
+}; //end namespace ams
+
+#endif
\ No newline at end of file
diff --git a/src/amscppnarray/amscppnarray.cpp b/src/amscppnarray/amscppnarray.cpp
index d962e01..7458e9d 100644
--- a/src/amscppnarray/amscppnarray.cpp
+++ b/src/amscppnarray/amscppnarray.cpp
@@ -5,12 +5,97 @@ namespace ams
 namespace narray
 {
 
+
 void test_narray1()
 {
-    
+    int I;
+    narray<float> a;
+    narray<float> *b = NULL;
+    narray<float> tmp;
+
+    b = new(std::nothrow) narray<float>();
+    a.resize(100);
+    b->resize(100000);
+    a.setall(2.5f);
+    b->setall(3.0f);
+
+    for(I=0;I<100000;I++)
+    {
+        if(b->at(I)!=3.0f)
+        {
+            printf("Error: b[%d]=%1.3f, not %1.3f\n",I,b->at(I),3.0f);
+            break;
+        }
+    }
+
+    a.resize(100000);
+
+    tmp = a;
+    a = *b;
+    *b = tmp;
+
+    *b = tmp+a;
+    for(I=0;I<b->length;I++)
+    {
+        if(b->at(I) != tmp[I] + a[I])
+        {
+            printf("addition error in index %d",I);
+        }
+    }
+
+    printf("threaded boolean test: %d\n", (*b==(a+tmp)));
+
+
+
+    delete b;
+
     return;
 }
 
+void test_narray2()
+{
+    int I;
+    narray<float> a,b,c,d;
+
+    a.resize(100000); a.setall(1.0f);
+    b.resize(100000); for(I=0;I<b.length;I++) b[I] = 0.5f*I;
+    c.resize(100000); c.setall(2.0f);
+
+    d = -a*(b+c);
+    for(I=0;I<5;I++)
+    {
+        printf("d[%d]=%1.3f\n",I,d[I]);
+
+    }
+
+    printf("d==-a*(b+c):%d\n",d==(-a*(b+c)));
+    d[50000] = 0.0f;
+    printf("d==-a*(b+c):%d\n",d==(-a*(b+c)));
+    
+}
+
+void test_narray3()
+{
+    int I;
+    narray<float> a,b,c,d;
+
+    a.resize(5); a.setall(1.0f);
+    b = {0,0.25,0.5,0.675,0.8675};
+    c.resize(5); c.setall(2.0f);
+
+    d = -a*(b+c);
+    for(I=0;I<5;I++)
+    {
+        printf("d[%d]=%1.3f\n",I,d[I]);
+
+    }
+
+    printf("d==-a*(b+c):%d\n",d==(-a*(b+c)));
+    d[4] = 0.0f;
+    printf("d==-a*(b+c):%d\n",d==(-a*(b+c)));
+    
+}
+
 
 }; //end namespace narray
 }; //end namespace ams
\ No newline at end of file
diff --git a/src/main.cpp b/src/main.cpp
index a24d585..fc25301 100644
--- a/src/main.cpp
+++ b/src/main.cpp
@@ -4,7 +4,9 @@ int main(int argc, char* argv[])
 {
     int ret = 0;
     printf("ams c++ numeric array library tests.\n");
-    ams::narray::test_narray1();
+    //ams::narray::test_narray1();
+    //ams::narray::test_narray2();
+    ams::narray::test_narray3();
 
     return ret;
 }
\ No newline at end of file