Changes to make Nsimd=1 sComplex lexLattices functional for "where" statements

This commit is contained in:
Peter Boyle
2026-08-19 19:30:34 -04:00
parent 8c9ea4cde2
commit 7b81ca6b52
3 changed files with 53 additions and 47 deletions
+1 -1
View File
@@ -74,7 +74,7 @@ accelerator_inline vobj predicatedWhere(const iobj &predicate,
extract(iftrue, truevals); extract(iftrue, truevals);
extract(iffalse, falsevals); extract(iffalse, falsevals);
extract<vInteger, Integer>(TensorRemove(predicate), mask); extract(TensorRemove(predicate), mask);
for (int s = 0; s < Nsimd; s++) { for (int s = 0; s < Nsimd; s++) {
if (mask[s]) falsevals[s] = truevals[s]; if (mask[s]) falsevals[s] = truevals[s];
+25 -25
View File
@@ -40,16 +40,16 @@ NAMESPACE_BEGIN(Grid);
//Query supporting logical &&, ||, //Query supporting logical &&, ||,
////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////
typedef iScalar<vInteger> vPredicate ; template<class vobj> using vPredicate = iScalar<IntegerPredicate<vobj> > ;
////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////
// compare lattice to lattice // compare lattice to lattice
////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////
template<class vfunctor,class lobj,class robj> template<class vfunctor,class lobj,class robj>
inline Lattice<vPredicate> LLComparison(vfunctor op,const Lattice<lobj> &lhs,const Lattice<robj> &rhs) inline Lattice<vPredicate<lobj> > LLComparison(vfunctor op,const Lattice<lobj> &lhs,const Lattice<robj> &rhs)
{ {
Lattice<vPredicate> ret(rhs.Grid()); Lattice<vPredicate<lobj> > ret(rhs.Grid());
autoView( lhs_v, lhs, CpuRead); autoView( lhs_v, lhs, CpuRead);
autoView( rhs_v, rhs, CpuRead); autoView( rhs_v, rhs, CpuRead);
autoView( ret_v, ret, CpuWrite); autoView( ret_v, ret, CpuWrite);
@@ -62,9 +62,9 @@ inline Lattice<vPredicate> LLComparison(vfunctor op,const Lattice<lobj> &lhs,con
// compare lattice to scalar // compare lattice to scalar
////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////
template<class vfunctor,class lobj,class robj> template<class vfunctor,class lobj,class robj>
inline Lattice<vPredicate> LSComparison(vfunctor op,const Lattice<lobj> &lhs,const robj &rhs) inline Lattice<vPredicate<lobj> > LSComparison(vfunctor op,const Lattice<lobj> &lhs,const robj &rhs)
{ {
Lattice<vPredicate> ret(lhs.Grid()); Lattice<vPredicate<lobj> > ret(lhs.Grid());
autoView( lhs_v, lhs, CpuRead); autoView( lhs_v, lhs, CpuRead);
autoView( ret_v, ret, CpuWrite); autoView( ret_v, ret, CpuWrite);
thread_for( ss, lhs_v.size(), { thread_for( ss, lhs_v.size(), {
@@ -76,9 +76,9 @@ inline Lattice<vPredicate> LSComparison(vfunctor op,const Lattice<lobj> &lhs,con
// compare scalar to lattice // compare scalar to lattice
////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////
template<class vfunctor,class lobj,class robj> template<class vfunctor,class lobj,class robj>
inline Lattice<vPredicate> SLComparison(vfunctor op,const lobj &lhs,const Lattice<robj> &rhs) inline Lattice<vPredicate<robj> > SLComparison(vfunctor op,const lobj &lhs,const Lattice<robj> &rhs)
{ {
Lattice<vPredicate> ret(rhs.Grid()); Lattice<vPredicate<robj> > ret(rhs.Grid());
autoView( rhs_v, rhs, CpuRead); autoView( rhs_v, rhs, CpuRead);
autoView( ret_v, ret, CpuWrite); autoView( ret_v, ret, CpuWrite);
thread_for( ss, rhs_v.size(), { thread_for( ss, rhs_v.size(), {
@@ -92,87 +92,87 @@ inline Lattice<vPredicate> SLComparison(vfunctor op,const lobj &lhs,const Lattic
////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////
// Less than // Less than
template<class lobj,class robj> template<class lobj,class robj>
inline Lattice<vPredicate> operator < (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) { inline Lattice<vPredicate<lobj> > operator < (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
return LLComparison(vlt<lobj,robj>(),lhs,rhs); return LLComparison(vlt<lobj,robj>(),lhs,rhs);
} }
template<class lobj,class robj> template<class lobj,class robj>
inline Lattice<vPredicate> operator < (const Lattice<lobj> & lhs, const robj & rhs) { inline Lattice<vPredicate<lobj> > operator < (const Lattice<lobj> & lhs, const robj & rhs) {
return LSComparison(vlt<lobj,robj>(),lhs,rhs); return LSComparison(vlt<lobj,robj>(),lhs,rhs);
} }
template<class lobj,class robj> template<class lobj,class robj>
inline Lattice<vPredicate> operator < (const lobj & lhs, const Lattice<robj> & rhs) { inline Lattice<vPredicate<robj> > operator < (const lobj & lhs, const Lattice<robj> & rhs) {
return SLComparison(vlt<lobj,robj>(),lhs,rhs); return SLComparison(vlt<lobj,robj>(),lhs,rhs);
} }
// Less than equal // Less than equal
template<class lobj,class robj> template<class lobj,class robj>
inline Lattice<vPredicate> operator <= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) { inline Lattice<vPredicate<lobj> > operator <= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
return LLComparison(vle<lobj,robj>(),lhs,rhs); return LLComparison(vle<lobj,robj>(),lhs,rhs);
} }
template<class lobj,class robj> template<class lobj,class robj>
inline Lattice<vPredicate> operator <= (const Lattice<lobj> & lhs, const robj & rhs) { inline Lattice<vPredicate<lobj> > operator <= (const Lattice<lobj> & lhs, const robj & rhs) {
return LSComparison(vle<lobj,robj>(),lhs,rhs); return LSComparison(vle<lobj,robj>(),lhs,rhs);
} }
template<class lobj,class robj> template<class lobj,class robj>
inline Lattice<vPredicate> operator <= (const lobj & lhs, const Lattice<robj> & rhs) { inline Lattice<vPredicate<robj> > operator <= (const lobj & lhs, const Lattice<robj> & rhs) {
return SLComparison(vle<lobj,robj>(),lhs,rhs); return SLComparison(vle<lobj,robj>(),lhs,rhs);
} }
// Greater than // Greater than
template<class lobj,class robj> template<class lobj,class robj>
inline Lattice<vPredicate> operator > (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) { inline Lattice<vPredicate<lobj> > operator > (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
return LLComparison(vgt<lobj,robj>(),lhs,rhs); return LLComparison(vgt<lobj,robj>(),lhs,rhs);
} }
template<class lobj,class robj> template<class lobj,class robj>
inline Lattice<vPredicate> operator > (const Lattice<lobj> & lhs, const robj & rhs) { inline Lattice<vPredicate<lobj> > operator > (const Lattice<lobj> & lhs, const robj & rhs) {
return LSComparison(vgt<lobj,robj>(),lhs,rhs); return LSComparison(vgt<lobj,robj>(),lhs,rhs);
} }
template<class lobj,class robj> template<class lobj,class robj>
inline Lattice<vPredicate> operator > (const lobj & lhs, const Lattice<robj> & rhs) { inline Lattice<vPredicate<robj> > operator > (const lobj & lhs, const Lattice<robj> & rhs) {
return SLComparison(vgt<lobj,robj>(),lhs,rhs); return SLComparison(vgt<lobj,robj>(),lhs,rhs);
} }
// Greater than equal // Greater than equal
template<class lobj,class robj> template<class lobj,class robj>
inline Lattice<vPredicate> operator >= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) { inline Lattice<vPredicate<lobj> > operator >= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
return LLComparison(vge<lobj,robj>(),lhs,rhs); return LLComparison(vge<lobj,robj>(),lhs,rhs);
} }
template<class lobj,class robj> template<class lobj,class robj>
inline Lattice<vPredicate> operator >= (const Lattice<lobj> & lhs, const robj & rhs) { inline Lattice<vPredicate<lobj> > operator >= (const Lattice<lobj> & lhs, const robj & rhs) {
return LSComparison(vge<lobj,robj>(),lhs,rhs); return LSComparison(vge<lobj,robj>(),lhs,rhs);
} }
template<class lobj,class robj> template<class lobj,class robj>
inline Lattice<vPredicate> operator >= (const lobj & lhs, const Lattice<robj> & rhs) { inline Lattice<vPredicate<robj> > operator >= (const lobj & lhs, const Lattice<robj> & rhs) {
return SLComparison(vge<lobj,robj>(),lhs,rhs); return SLComparison(vge<lobj,robj>(),lhs,rhs);
} }
// equal // equal
template<class lobj,class robj> template<class lobj,class robj>
inline Lattice<vPredicate> operator == (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) { inline Lattice<vPredicate<lobj> > operator == (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
return LLComparison(veq<lobj,robj>(),lhs,rhs); return LLComparison(veq<lobj,robj>(),lhs,rhs);
} }
template<class lobj,class robj> template<class lobj,class robj>
inline Lattice<vPredicate> operator == (const Lattice<lobj> & lhs, const robj & rhs) { inline Lattice<vPredicate<lobj> > operator == (const Lattice<lobj> & lhs, const robj & rhs) {
return LSComparison(veq<lobj,robj>(),lhs,rhs); return LSComparison(veq<lobj,robj>(),lhs,rhs);
} }
template<class lobj,class robj> template<class lobj,class robj>
inline Lattice<vPredicate> operator == (const lobj & lhs, const Lattice<robj> & rhs) { inline Lattice<vPredicate<robj> > operator == (const lobj & lhs, const Lattice<robj> & rhs) {
return SLComparison(veq<lobj,robj>(),lhs,rhs); return SLComparison(veq<lobj,robj>(),lhs,rhs);
} }
// not equal // not equal
template<class lobj,class robj> template<class lobj,class robj>
inline Lattice<vPredicate> operator != (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) { inline Lattice<vPredicate<lobj> > operator != (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
return LLComparison(vne<lobj,robj>(),lhs,rhs); return LLComparison(vne<lobj,robj>(),lhs,rhs);
} }
template<class lobj,class robj> template<class lobj,class robj>
inline Lattice<vPredicate> operator != (const Lattice<lobj> & lhs, const robj & rhs) { inline Lattice<vPredicate<lobj> > operator != (const Lattice<lobj> & lhs, const robj & rhs) {
return LSComparison(vne<lobj,robj>(),lhs,rhs); return LSComparison(vne<lobj,robj>(),lhs,rhs);
} }
template<class lobj,class robj> template<class lobj,class robj>
inline Lattice<vPredicate> operator != (const lobj & lhs, const Lattice<robj> & rhs) { inline Lattice<vPredicate<robj> > operator != (const lobj & lhs, const Lattice<robj> & rhs) {
return SLComparison(vne<lobj,robj>(),lhs,rhs); return SLComparison(vne<lobj,robj>(),lhs,rhs);
} }
NAMESPACE_END(Grid); NAMESPACE_END(Grid);
+27 -21
View File
@@ -39,47 +39,53 @@ NAMESPACE_BEGIN(Grid);
// //
// Should guard this with isGridTensor<> enable if? // Should guard this with isGridTensor<> enable if?
///////////////////////////////////////// /////////////////////////////////////////
//
// Predicate type matching the operand's vectorisation
//
template<class obj> using IntegerPredicate =
typename GridTypeMapper<typename GridTypeMapper<obj>::vector_type>::Integerified;
// //
// Generic list of functors // Generic list of functors
// //
template<class lobj,class robj> class veq { template<class lobj,class robj> class veq {
public: public:
accelerator vInteger operator()(const lobj &lhs, const robj &rhs) accelerator IntegerPredicate<lobj> operator()(const lobj &lhs, const robj &rhs)
{ {
return (lhs) == (rhs); return (lhs) == (rhs);
} }
}; };
template<class lobj,class robj> class vne { template<class lobj,class robj> class vne {
public: public:
accelerator vInteger operator()(const lobj &lhs, const robj &rhs) accelerator IntegerPredicate<lobj> operator()(const lobj &lhs, const robj &rhs)
{ {
return (lhs) != (rhs); return (lhs) != (rhs);
} }
}; };
template<class lobj,class robj> class vlt { template<class lobj,class robj> class vlt {
public: public:
accelerator vInteger operator()(const lobj &lhs, const robj &rhs) accelerator IntegerPredicate<lobj> operator()(const lobj &lhs, const robj &rhs)
{ {
return (lhs) < (rhs); return (lhs) < (rhs);
} }
}; };
template<class lobj,class robj> class vle { template<class lobj,class robj> class vle {
public: public:
accelerator vInteger operator()(const lobj &lhs, const robj &rhs) accelerator IntegerPredicate<lobj> operator()(const lobj &lhs, const robj &rhs)
{ {
return (lhs) <= (rhs); return (lhs) <= (rhs);
} }
}; };
template<class lobj,class robj> class vgt { template<class lobj,class robj> class vgt {
public: public:
accelerator vInteger operator()(const lobj &lhs, const robj &rhs) accelerator IntegerPredicate<lobj> operator()(const lobj &lhs, const robj &rhs)
{ {
return (lhs) > (rhs); return (lhs) > (rhs);
} }
}; };
template<class lobj,class robj> class vge { template<class lobj,class robj> class vge {
public: public:
accelerator vInteger operator()(const lobj &lhs, const robj &rhs) accelerator IntegerPredicate<lobj> operator()(const lobj &lhs, const robj &rhs)
{ {
return (lhs) >= (rhs); return (lhs) >= (rhs);
} }
@@ -133,78 +139,78 @@ NAMESPACE_BEGIN(Grid);
// Integer and real get extra relational functions. // Integer and real get extra relational functions.
////////////////////////////////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////////////////////////////////
template<class sfunctor, class vsimd,IfNotComplex<vsimd> = 0> template<class sfunctor, class vsimd,IfNotComplex<vsimd> = 0>
accelerator_inline vInteger Comparison(sfunctor sop,const vsimd & lhs, const vsimd & rhs) accelerator_inline IntegerPredicate<vsimd> Comparison(sfunctor sop,const vsimd & lhs, const vsimd & rhs)
{ {
typedef typename vsimd::scalar_type scalar; typedef typename vsimd::scalar_type scalar;
ExtractBuffer<scalar> vlhs(vsimd::Nsimd()); // Use functors to reduce this to single implementation ExtractBuffer<scalar> vlhs(vsimd::Nsimd()); // Use functors to reduce this to single implementation
ExtractBuffer<scalar> vrhs(vsimd::Nsimd()); ExtractBuffer<scalar> vrhs(vsimd::Nsimd());
ExtractBuffer<Integer> vpred(vsimd::Nsimd()); ExtractBuffer<Integer> vpred(vsimd::Nsimd());
vInteger ret; IntegerPredicate<vsimd> ret;
extract<vsimd,scalar>(lhs,vlhs); extract<vsimd,scalar>(lhs,vlhs);
extract<vsimd,scalar>(rhs,vrhs); extract<vsimd,scalar>(rhs,vrhs);
for(int s=0;s<vsimd::Nsimd();s++){ for(int s=0;s<vsimd::Nsimd();s++){
vpred[s] = sop(vlhs[s],vrhs[s]); vpred[s] = sop(vlhs[s],vrhs[s]);
} }
merge<vInteger,Integer>(ret,vpred); merge<IntegerPredicate<vsimd>,Integer>(ret,vpred);
return ret; return ret;
} }
template<class sfunctor, class vsimd,IfNotComplex<vsimd> = 0> template<class sfunctor, class vsimd,IfNotComplex<vsimd> = 0>
accelerator_inline vInteger Comparison(sfunctor sop,const vsimd & lhs, const typename vsimd::scalar_type & rhs) accelerator_inline IntegerPredicate<vsimd> Comparison(sfunctor sop,const vsimd & lhs, const typename vsimd::scalar_type & rhs)
{ {
typedef typename vsimd::scalar_type scalar; typedef typename vsimd::scalar_type scalar;
ExtractBuffer<scalar> vlhs(vsimd::Nsimd()); // Use functors to reduce this to single implementation ExtractBuffer<scalar> vlhs(vsimd::Nsimd()); // Use functors to reduce this to single implementation
ExtractBuffer<Integer> vpred(vsimd::Nsimd()); ExtractBuffer<Integer> vpred(vsimd::Nsimd());
vInteger ret; IntegerPredicate<vsimd> ret;
extract<vsimd,scalar>(lhs,vlhs); extract<vsimd,scalar>(lhs,vlhs);
for(int s=0;s<vsimd::Nsimd();s++){ for(int s=0;s<vsimd::Nsimd();s++){
vpred[s] = sop(vlhs[s],rhs); vpred[s] = sop(vlhs[s],rhs);
} }
merge<vInteger,Integer>(ret,vpred); merge<IntegerPredicate<vsimd>,Integer>(ret,vpred);
return ret; return ret;
} }
template<class sfunctor, class vsimd,IfNotComplex<vsimd> = 0> template<class sfunctor, class vsimd,IfNotComplex<vsimd> = 0>
accelerator_inline vInteger Comparison(sfunctor sop,const typename vsimd::scalar_type & lhs, const vsimd & rhs) accelerator_inline IntegerPredicate<vsimd> Comparison(sfunctor sop,const typename vsimd::scalar_type & lhs, const vsimd & rhs)
{ {
typedef typename vsimd::scalar_type scalar; typedef typename vsimd::scalar_type scalar;
ExtractBuffer<scalar> vrhs(vsimd::Nsimd()); // Use functors to reduce this to single implementation ExtractBuffer<scalar> vrhs(vsimd::Nsimd()); // Use functors to reduce this to single implementation
ExtractBuffer<Integer> vpred(vsimd::Nsimd()); ExtractBuffer<Integer> vpred(vsimd::Nsimd());
vInteger ret; IntegerPredicate<vsimd> ret;
extract<vsimd,scalar>(rhs,vrhs); extract<vsimd,scalar>(rhs,vrhs);
for(int s=0;s<vsimd::Nsimd();s++){ for(int s=0;s<vsimd::Nsimd();s++){
vpred[s] = sop(lhs,vrhs[s]); vpred[s] = sop(lhs,vrhs[s]);
} }
merge<vInteger,Integer>(ret,vpred); merge<IntegerPredicate<vsimd>,Integer>(ret,vpred);
return ret; return ret;
} }
#define DECLARE_RELATIONAL_EQ(op,functor) \ #define DECLARE_RELATIONAL_EQ(op,functor) \
template<class vsimd,IfSimd<vsimd> = 0>\ template<class vsimd,IfSimd<vsimd> = 0>\
accelerator_inline vInteger operator op (const vsimd & lhs, const vsimd & rhs)\ accelerator_inline IntegerPredicate<vsimd> operator op (const vsimd & lhs, const vsimd & rhs)\
{\ {\
typedef typename vsimd::scalar_type scalar;\ typedef typename vsimd::scalar_type scalar;\
return Comparison(functor<scalar,scalar>(),lhs,rhs);\ return Comparison(functor<scalar,scalar>(),lhs,rhs);\
}\ }\
template<class vsimd,IfSimd<vsimd> = 0>\ template<class vsimd,IfSimd<vsimd> = 0>\
accelerator_inline vInteger operator op (const vsimd & lhs, const typename vsimd::scalar_type & rhs) \ accelerator_inline IntegerPredicate<vsimd> operator op (const vsimd & lhs, const typename vsimd::scalar_type & rhs) \
{\ {\
typedef typename vsimd::scalar_type scalar;\ typedef typename vsimd::scalar_type scalar;\
return Comparison(functor<scalar,scalar>(),lhs,rhs);\ return Comparison(functor<scalar,scalar>(),lhs,rhs);\
}\ }\
template<class vsimd,IfSimd<vsimd> = 0>\ template<class vsimd,IfSimd<vsimd> = 0>\
accelerator_inline vInteger operator op (const typename vsimd::scalar_type & lhs, const vsimd & rhs) \ accelerator_inline IntegerPredicate<vsimd> operator op (const typename vsimd::scalar_type & lhs, const vsimd & rhs) \
{\ {\
typedef typename vsimd::scalar_type scalar;\ typedef typename vsimd::scalar_type scalar;\
return Comparison(functor<scalar,scalar>(),lhs,rhs);\ return Comparison(functor<scalar,scalar>(),lhs,rhs);\
}\ }\
template<class vsimd>\ template<class vsimd>\
accelerator_inline vInteger operator op(const iScalar<vsimd> &lhs,const typename vsimd::scalar_type &rhs) \ accelerator_inline IntegerPredicate<vsimd> operator op(const iScalar<vsimd> &lhs,const typename vsimd::scalar_type &rhs) \
{ \ { \
return lhs._internal op rhs; \ return lhs._internal op rhs; \
} \ } \
template<class vsimd>\ template<class vsimd>\
accelerator_inline vInteger operator op(const typename vsimd::scalar_type &lhs,const iScalar<vsimd> &rhs) \ accelerator_inline IntegerPredicate<vsimd> operator op(const typename vsimd::scalar_type &lhs,const iScalar<vsimd> &rhs) \
{ \ { \
return lhs op rhs._internal; \ return lhs op rhs._internal; \
} \ } \
@@ -212,7 +218,7 @@ NAMESPACE_BEGIN(Grid);
#define DECLARE_RELATIONAL(op,functor) \ #define DECLARE_RELATIONAL(op,functor) \
DECLARE_RELATIONAL_EQ(op,functor) \ DECLARE_RELATIONAL_EQ(op,functor) \
template<class vsimd>\ template<class vsimd>\
accelerator_inline vInteger operator op(const iScalar<vsimd> &lhs,const iScalar<vsimd> &rhs)\ accelerator_inline IntegerPredicate<vsimd> operator op(const iScalar<vsimd> &lhs,const iScalar<vsimd> &rhs)\
{ \ { \
return lhs._internal op rhs._internal; \ return lhs._internal op rhs._internal; \
} }