portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-10 07:55:35 +00:00
Code Releases Activity

support fcmla in vector_types, untested

Browse Source
This commit is contained in:
nmeyer-ur 2020-05-18 19:55:18 +02:00
parent 032f7dde1a
commit 9f212679f1
3 changed files with 71 additions and 22 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

46
Grid/simd/Grid_a64fx-2.h
View File

@ -388,8 +388,29 @@ struct MultComplex{
typename acle<T>::vt z_v = acle<T>::zero();
// using FCMLA
typename acle<T>::vt r_v = svcmla_x(pg1, z_v, a_v, b_v, 90);
r_v = svcmla_x(pg1, r_v, a_v, b_v, 0);
typename acle<T>::vt r_v = svcmla_x(pg1, z_v, a_v, b_v, 0);
r_v = svcmla_x(pg1, r_v, a_v, b_v, 90);
svst1(pg1, out.v, r_v);
return out;
}
};
struct MultAddComplex{
// Complex a*b+c
template <typename T>
inline vec<T> mac(const vec<T> &a, const vec<T> &b, const vec<T> &c){
vec<T> out;
svbool_t pg1 = acle<T>::pg1();
typename acle<T>::vt a_v = svld1(pg1, a.v);
typename acle<T>::vt b_v = svld1(pg1, b.v);
typename acle<T>::vt c_v = svld1(pg1, c.v);;
// using FCMLA
typename acle<T>::vt r_v = svcmla_x(pg1, c_v, a_v, b_v, 0);
r_v = svcmla_x(pg1, r_v, a_v, b_v, 90);
svst1(pg1, out.v, r_v);
@ -897,15 +918,16 @@ typedef Optimization::Vstream VstreamSIMD;
template <typename S, typename T> using ReduceSIMD = Optimization::Reduce<S,T>;
// Arithmetic operations
typedef Optimization::Sum SumSIMD;
typedef Optimization::Sub SubSIMD;
typedef Optimization::Div DivSIMD;
typedef Optimization::Mult MultSIMD;
typedef Optimization::MultComplex MultComplexSIMD;
typedef Optimization::MultRealPart MultRealPartSIMD;
typedef Optimization::MaddRealPart MaddRealPartSIMD;
typedef Optimization::Conj ConjSIMD;
typedef Optimization::TimesMinusI TimesMinusISIMD;
typedef Optimization::TimesI TimesISIMD;
typedef Optimization::Sum SumSIMD;
typedef Optimization::Sub SubSIMD;
typedef Optimization::Div DivSIMD;
typedef Optimization::Mult MultSIMD;
typedef Optimization::MultComplex MultComplexSIMD;
typedef Optimization::MultAddComplex MultAddComplexSIMD;
typedef Optimization::MultRealPart MultRealPartSIMD;
typedef Optimization::MaddRealPart MaddRealPartSIMD;
typedef Optimization::Conj ConjSIMD;
typedef Optimization::TimesMinusI TimesMinusISIMD;
typedef Optimization::TimesI TimesISIMD;
NAMESPACE_END(Grid)

39
Grid/simd/Grid_a64fx-fixedsize.h
View File

@ -367,6 +367,24 @@ struct MultComplex{
}
};
struct MultAddComplex{
// Complex a*b+c
// Complex float
inline vecf mac(vecf a, vecf b, vecf c){
pred pg1 = acle<float>::pg1();
// using FCMLA
vecf r_v = svcmla_x(pg1, c, a, b, 0);
return svcmla_x(pg1, r_v, a, b, 90);
}
// Complex double
inline vecd mac(vecd a, vecd b, vecd c){
pred pg1 = acle<double>::pg1();
// using FCMLA
vecf r_v = svcmla_x(pg1, c, a, b, 0);
return svcmla_x(pg1, r_v, a, b, 90);
}
};
struct Div{
// Real float
inline vecf operator()(vecf a, vecf b){
@ -772,15 +790,16 @@ typedef Optimization::Vstream VstreamSIMD;
template <typename S, typename T> using ReduceSIMD = Optimization::Reduce<S,T>;
// Arithmetic operations
typedef Optimization::Sum SumSIMD;
typedef Optimization::Sub SubSIMD;
typedef Optimization::Div DivSIMD;
typedef Optimization::Mult MultSIMD;
typedef Optimization::MultComplex MultComplexSIMD;
typedef Optimization::MultRealPart MultRealPartSIMD;
typedef Optimization::MaddRealPart MaddRealPartSIMD;
typedef Optimization::Conj ConjSIMD;
typedef Optimization::TimesMinusI TimesMinusISIMD;
typedef Optimization::TimesI TimesISIMD;
typedef Optimization::Sum SumSIMD;
typedef Optimization::Sub SubSIMD;
typedef Optimization::Div DivSIMD;
typedef Optimization::Mult MultSIMD;
typedef Optimization::MultComplex MultComplexSIMD;
typedef Optimization::MultAddComplex MultAddComplexSIMD;
typedef Optimization::MultRealPart MultRealPartSIMD;
typedef Optimization::MaddRealPart MaddRealPartSIMD;
typedef Optimization::Conj ConjSIMD;
typedef Optimization::TimesMinusI TimesMinusISIMD;
typedef Optimization::TimesI TimesISIMD;
NAMESPACE_END(Grid);

8
Grid/simd/Grid_vector_types.h
View File

@ -298,11 +298,19 @@ public:
///////////////////////////////////////////////
// FIXME -- alias this to an accelerator_inline MAC struct.
#if defined(A64FX) || defined(A64FXFIXEDSIZE) // on A64FX use FCMLA
friend accelerator_inline void mac(Grid_simd *__restrict__ y,
const Grid_simd *__restrict__ a,
const Grid_simd *__restrict__ x) {
y->v = Optimization::MultAddComplex::mac(a->v, x->v, y->v);
};
#else
friend accelerator_inline void mac(Grid_simd *__restrict__ y,
const Grid_simd *__restrict__ a,
const Grid_simd *__restrict__ x) {
*y = (*a) * (*x) + (*y);
};
#endif
friend accelerator_inline void mult(Grid_simd *__restrict__ y,
const Grid_simd *__restrict__ l,