From 7904d77bd899577cb155db01bf887b7f9b1ffc9d Mon Sep 17 00:00:00 2001
From: Christian Zimmermann <christian.zimmermann@ur.de>
Date: Tue, 18 Oct 2022 00:30:05 +0200
Subject: [PATCH] EPos + EFor (for extensions, simd vectorization, etc)

---
 src/include/base/base.h              |   1 +
 src/include/base/types.h             |   4 -
 src/include/ranges/xpr/for.cc.h      |  37 ++++++++
 src/include/ranges/xpr/for.h         |  27 +++++-
 src/include/ranges/xpr/pos_type.cc.h | 107 +++++++++++++++++++++-
 src/include/ranges/xpr/pos_type.h    | 131 ++++++++++++++++++++++-----
 6 files changed, 277 insertions(+), 30 deletions(-)

diff --git a/src/include/base/base.h b/src/include/base/base.h
index 523dbeb..5897c99 100644
--- a/src/include/base/base.h
+++ b/src/include/base/base.h
@@ -15,6 +15,7 @@
 
 #include <cstdlib>
 #include <iostream>
+#include <utility>
 
 #include "macros.h"
 #include "assert.h"
diff --git a/src/include/base/types.h b/src/include/base/types.h
index 511bd90..d34aae7 100644
--- a/src/include/base/types.h
+++ b/src/include/base/types.h
@@ -81,10 +81,6 @@ namespace CNORXZ
     template <typename T>
     class Allocator;
 
-    // definition: ranges/xpr/pos_type.h
-    template <class PosT>
-    class CPosInterface;
-
     // definition: ranges/xpr/vpos_type.h
     class VPosBase;
 
diff --git a/src/include/ranges/xpr/for.cc.h b/src/include/ranges/xpr/for.cc.h
index 69192fb..8776bb4 100644
--- a/src/include/ranges/xpr/for.cc.h
+++ b/src/include/ranges/xpr/for.cc.h
@@ -184,6 +184,43 @@ namespace CNORXZ
     }
 
 
+    /************
+     *   EFor   *
+     ************/
+
+    template <SizeT N, SizeT L, class PosT, class Xpr>
+    constexpr EFor<N,L,PosT,Xpr>::EFor(const IndexId<L>& id, const PosT& step, const Xpr& xpr) :
+	mId(id),
+	mStep(step),
+	mXpr(xpr),
+	mExt(mXpr.RootSteps(mId))
+    {
+	static_assert(is_pos_type<PosT>::value, "got non-pos type");
+    }
+
+    template <SizeT N, SizeT L, class PosT, class Xpr>
+    template <class PosT1, class PosT2>
+    constexpr SizeT EFor<N,L,PosT,Xpr>::operator()(const PosT1& mlast, const PosT2& last) const
+    {
+	auto mpos = mkEPos<N>(mlast, mStep);
+	auto pos = mkEPos<N>(last, mExt);
+	return mXpr(mpos, pos);
+    }
+	
+    template <SizeT N, SizeT L, class PosT, class Xpr>
+    constexpr SizeT EFor<N,L,PosT,Xpr>::operator()() const
+    {
+	auto mpos = mkEPos<N>(SPos<0>(), mStep);
+	auto pos = mkEPos<N>(SPos<0>(), mExt);
+	return mXpr(mpos, pos);
+    }
+
+    template <SizeT N, SizeT L, class PosT, class Xpr>
+    template <SizeT I>
+    constexpr decltype(auto) EFor<N,L,PosT,Xpr>::rootSteps(const IndexId<I>& id) const
+    {
+	return mXpr.rootSteps(id);
+    }
 }
 
 #endif
diff --git a/src/include/ranges/xpr/for.h b/src/include/ranges/xpr/for.h
index 8056b8a..40a061b 100644
--- a/src/include/ranges/xpr/for.h
+++ b/src/include/ranges/xpr/for.h
@@ -87,7 +87,7 @@ namespace CNORXZ
 	
     private:
 	SizeT mSize = 0;
-	IndexId<L> mId = 0;
+	IndexId<L> mId;
 	Xpr mXpr;
 	PosT mStep;
 	typedef decltype(mXpr.rootSteps(mId)) XPosT;
@@ -95,7 +95,30 @@ namespace CNORXZ
 	
     };
 
-    
+    // Extension For (Vectorization)
+    template <SizeT N, SizeT L, class PosT, class Xpr>
+    class EFor : public XprInterface<EFor<N,L,PosT,Xpr>>
+    {
+    public:
+	DEFAULT_MEMBERS(EFor);
+
+	constexpr EFor(const IndexId<L>& id, const PosT& step, const Xpr& xpr);
+
+	template <class PosT1, class PosT2>
+	constexpr SizeT operator()(const PosT1& mlast, const PosT2& last) const;
+
+	constexpr SizeT operator()() const;
+
+	template <SizeT I>
+	constexpr decltype(auto) rootSteps(const IndexId<I>& id) const;
+
+    private:
+	IndexId<L> mId;
+	Xpr mXpr;
+	PosT mStep;
+	typedef decltype(mXpr.rootSteps(mId)) XPosT;
+	XPosT mExt;
+    };
     
 }
 
diff --git a/src/include/ranges/xpr/pos_type.cc.h b/src/include/ranges/xpr/pos_type.cc.h
index af53b56..270d750 100644
--- a/src/include/ranges/xpr/pos_type.cc.h
+++ b/src/include/ranges/xpr/pos_type.cc.h
@@ -97,12 +97,22 @@ namespace CNORXZ
     {
 	return UPos(mExt + in.val());
     }
+
+    constexpr SPos<0> UPos::operator*(const SPos<0>& a) const
+    {
+	return SPos<0>();
+    }
     
     template <class PosT>
     constexpr UPos UPos::operator*(const PosT& in) const
     {
 	return UPos(mExt * in.val());
     }
+
+    constexpr SPos<0> UPos::operator()(const SPos<0>& a) const
+    {
+	return SPos<0>();
+    }
 	
     template <class PosT>
     constexpr UPos UPos::operator()(const PosT& in) const
@@ -477,14 +487,107 @@ namespace CNORXZ
 	    VPosRef<PosT> b(&a);
 	    return DPos(mP->vexec( &b ));
 	}
-    }
-	
+    }	
+
     template <class PosT>
     inline DPos DPosRef::extend(const PosT& a) const
     {
 	return DPos(mP->vextend( a ));
     }
     
+    /************
+     *   EPos   *
+     ************/
+
+    template <class BPosT, class... OPosTs>
+    constexpr EPos<BPosT,OPosTs...>::EPos(const BPosT& b, const OPosTs&... os) :
+	BPosT(b),
+	mP(os...)
+    {}
+
+    template <class BPosT, class... OPosTs>
+    constexpr EPos<BPosT,OPosTs...>::EPos(BPosT&& b, OPosTs&&... os) :
+	BPosT(b),
+	mP(os...)
+    {}
+
+    template <class BPosT, class... OPosTs>
+    constexpr decltype(auto) EPos<BPosT,OPosTs...>::val() const
+    {
+	return ival(std::index_sequence_for<OPosTs...>{});
+    }
+
+    template <class BPosT, class... OPosTs>
+    constexpr decltype(auto) EPos<BPosT,OPosTs...>::next() const
+    {
+	return inext(std::index_sequence_for<OPosTs...>{});
+    }
+
+    template <class BPosT, class... OPosTs>
+    template <SizeT... Is>
+    constexpr decltype(auto) EPos<BPosT,OPosTs...>::ival(std::index_sequence<Is...> is) const
+    {
+	if constexpr(is_static_pos_type<EPos<BPosT,OPosTs...>>::value){
+	    return std::index_sequence<std::get<Is>(mP).val()...>{};
+	}
+	else {
+	    return Arr<SizeT,is.size()> { std::get<Is>(mP).val()... };
+	}
+    }
+
+    template <class BPosT, class... OPosTs>
+    template <SizeT... Is>
+    constexpr decltype(auto) EPos<BPosT,OPosTs...>::inext(std::index_sequence<Is...> is) const
+    {
+	typedef EPos<decltype(next()),decltype(std::get<Is>(mP).next())...> OEPosT;
+	return OEPosT(next(), std::get<Is>(mP).next()...);
+    }
+
+    /*********************************
+     *   Traits and Helper-Classes   *
+     *********************************/
+
+    template <SizeT N, class BPosT, class OPosT>
+    decltype(auto) mkEPos(const BPosT& a, const OPosT& b)
+    {
+	return mkiEPos(a, b, std::make_index_sequence<N>{});
+    }
+
+    template <SizeT N, class BPosT, class OPosT, SizeT... Is>
+    decltype(auto) mkiEPos(const BPosT& a, const OPosT& b, std::index_sequence<Is...> is)
+    {
+	return EPos<BPosT,decltype(b*SPos<Is>())...>(a, b*SPos<Is>()...);
+    }
+    
+    template <class OPosT1, class OPosT2, class... OPosTs>
+    constexpr bool pos_type_consecutive_2<OPosT1,OPosT2,OPosTs...>::eval()
+    {
+	if constexpr(is_static_pos_type<OPosT1>::value and is_static_pos_type<OPosT2>::value){
+	    if constexpr(sizeof...(OPosTs) != 0){
+		return OPosT1().val() < OPosT2().val() and pos_type_consecutive<OPosT2,OPosTs...>::value;
+	    }
+	    else {
+		return OPosT1().val() < OPosT2().val();
+	    }
+	}
+	return false;
+    }
+
+    template <class OPosT1, class... OPosTs>
+    constexpr bool pos_type_same<OPosT1,OPosTs...>::eval()
+    {
+	if constexpr(is_static_pos_type<OPosT1>::value){
+	    if constexpr(sizeof...(OPosTs) != 0){
+		return OPosT1().val() == 0 and pos_type_same<OPosTs...>::value;
+	    }
+	    else {
+		return OPosT1().val() == 0;
+	    }
+	}
+	else {
+	    return false;
+	}
+    }
 }
 
 #endif
diff --git a/src/include/ranges/xpr/pos_type.h b/src/include/ranges/xpr/pos_type.h
index 28f8a42..d879405 100644
--- a/src/include/ranges/xpr/pos_type.h
+++ b/src/include/ranges/xpr/pos_type.h
@@ -7,27 +7,6 @@
 namespace CNORXZ
 {
     
-    template <class T>
-    struct is_pos_type { CXZ_CVAL_FALSE; };
-
-    template <class T>
-    struct is_scalar_pos_type { CXZ_CVAL_FALSE; };
-
-    template <class T>
-    struct is_static_pos_type { CXZ_CVAL_FALSE; };
-
-    template <class PosT>
-    struct pos_depth
-    {
-	pos_depth()
-	{
-	    static_assert(is_pos_type<PosT>::value,
-			  "pos_depth is only defined for pos types");
-	}
-
-	static constexpr SizeT value = 1;
-    };
-    
     template <SizeT N>
     class SPos
     {
@@ -71,9 +50,13 @@ namespace CNORXZ
 	template <class PosT>
 	constexpr UPos operator+(const PosT& a) const;
 
+	constexpr SPos<0> operator*(const SPos<0>& a) const;
+
         template <class PosT>
 	constexpr UPos operator*(const PosT& a) const;
 
+	constexpr SPos<0> operator()(const SPos<0>& a) const;
+
 	template <class PosT>
 	constexpr UPos operator()(const PosT& a) const;
 
@@ -234,27 +217,131 @@ namespace CNORXZ
 	template <class PosT>
 	inline DPos extend(const PosT& a) const;
     };
+
+    // for common call of extension vector elements
+    // BPos = base pos type
+    // OPos = offset pos type (i.e. step/ext * for index of lowest for layer)
+    // currently EPos must be used only at the lowest for layer
+    template <class BPosT, class... OPosTs>
+    class EPos : public BPosT
+    {
+    private:
+	Tuple<OPosTs...> mP;
+
+	template <SizeT... Is>
+	constexpr decltype(auto) ival(std::index_sequence<Is...> is) const;
+
+	template <SizeT... Is>
+	constexpr decltype(auto) inext(std::index_sequence<Is...> is) const;
+	
+    public:
+	DEFAULT_MEMBERS(EPos);
+
+	constexpr EPos(const BPosT& b, const OPosTs&... os);
+	constexpr EPos(BPosT&& b, OPosTs&&... os);
+
+	constexpr decltype(auto) val() const;
+	constexpr decltype(auto) next() const;
+    };
+
+    /*********************************
+     *   Traits and Helper-Classes   *
+     *********************************/
     
+    template <class T>
+    struct is_pos_type { CXZ_CVAL_FALSE; };
+
+    template <class T>
+    struct is_scalar_pos_type { CXZ_CVAL_FALSE; };
+
+    template <class T>
+    struct is_static_pos_type { CXZ_CVAL_FALSE; };
+
+    template <class T>
+    struct is_epos_type { CXZ_CVAL_FALSE; };
+    
+    template <class PosT>
+    struct pos_depth
+    {
+	static constexpr SizeT value = is_pos_type<PosT>::value ? 1 : 0;
+    };
+    
+    template <class OPosT1, class OPosT2, class... OPosTs>
+    struct pos_type_consecutive_2
+    {
+    private:
+	static constexpr bool eval();
+	
+    public:
+	static constexpr bool value = eval();
+    };
+
+    template <class OPosT1, class... OPosTs>
+    struct pos_type_consecutive
+    {
+	static constexpr bool value = sizeof...(OPosTs) == 0 ? is_pos_type<OPosT1>::value : pos_type_consecutive_2<OPosT1,OPosTs...>::value;
+    };
+
+    template <class OPosT1, class... OPosTs>
+    struct pos_type_same
+    {
+    private:
+	static constexpr bool eval();
+	
+    public:
+	static constexpr bool value = eval();
+    };
+    
+    template <class PosT> struct pos_type_is_consecutive { CXZ_CVAL_FALSE; };
+
+    template <SizeT N, class BPosT, class OPosT>
+    decltype(auto) mkEPos(const BPosT& a, const OPosT& b);
+
+    template <SizeT N, class BPosT, class OPosT, SizeT... Is>
+    decltype(auto) mkiEPos(const BPosT& a, const OPosT& b, std::index_sequence<Is...> is);
+
+    /**************************************************
+     *   Traits and Helper-Classes: Specializations   *
+     **************************************************/
+
     template <SizeT N> struct is_pos_type<SPos<N>> { CXZ_CVAL_TRUE; };
     template <SizeT N> struct is_scalar_pos_type<SPos<N>> { CXZ_CVAL_TRUE; };
     template <SizeT N> struct is_static_pos_type<SPos<N>> { CXZ_CVAL_TRUE; };
+
     template <> struct is_pos_type<UPos> { CXZ_CVAL_TRUE; };
     template <> struct is_scalar_pos_type<UPos> { CXZ_CVAL_TRUE; };
+
     template <> struct is_pos_type<FPos> { CXZ_CVAL_TRUE; };
     template <> struct is_scalar_pos_type<FPos> { CXZ_CVAL_TRUE; };
+
     template <SizeT N, SizeT... Ms> struct is_pos_type<SFPos<N,Ms...>> { CXZ_CVAL_TRUE; };
     template <SizeT N, SizeT... Ms> struct is_scalar_pos_type<SFPos<N,Ms...>> { CXZ_CVAL_TRUE; };
     template <SizeT N, SizeT... Ms> struct is_static_pos_type<SFPos<N,Ms...>> { CXZ_CVAL_TRUE; };
+
     template <class BPosT, class NPosT> struct is_pos_type<MPos<BPosT,NPosT>> { CXZ_CVAL_TRUE; };
+
     template <> struct is_pos_type<DPos> { CXZ_CVAL_TRUE; };
     template <> struct is_pos_type<DPosRef> { CXZ_CVAL_TRUE; };
 
+    template <class BPosT, class... OPosTs> struct is_pos_type<EPos<BPosT,OPosTs...>> { CXZ_CVAL_TRUE; };
+    template <class BPosT, class... OPosTs> struct is_scalar_pos_type<EPos<BPosT,OPosTs...>>
+    { static constexpr bool value = is_scalar_pos_type<BPosT>::value; };
+    template <class BPosT, class... OPosTs> struct is_static_pos_type<EPos<BPosT,OPosTs...>>
+    { static constexpr bool value = is_static_pos_type<BPosT>::value and (is_static_pos_type<OPosTs>::value and ...); };
+    template <class BPosT, class... OPosTs> struct is_epos_type<EPos<BPosT,OPosTs...>> { CXZ_CVAL_TRUE; };
+    
     template <class BPosT, class NPosT>
     struct pos_depth<MPos<BPosT,NPosT>>
     {
 	static constexpr SizeT value = pos_depth<NPosT>::value + 1;
     };
-    
+
+    template <class BPosT, class... OPosTs>
+    struct pos_type_is_consecutive<EPos<BPosT,OPosTs...>>
+    {
+	static constexpr bool value = pos_type_is_consecutive<OPosTs...>::value;
+    };
+
 } // end namespace CNORXZInternal