portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-04-26 05:35:55 +01:00
Code Releases Activity

enable fcmla in tensor arithmetics; fixed-size works, VLA does not compile

Browse Source
This commit is contained in:
nmeyer-ur 2020-05-21 19:39:07 +02:00
parent a65ce237c1
commit 046b1cbbc0
2 changed files with 24 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
Grid/simd/Grid_a64fx-2.h
View File

@ -324,6 +324,18 @@ struct Sub{
};
struct Mult{
template <typename T>
inline vec<T> operator()(vec<T> a, vec<T> b, 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);
typename acle<T>::vt r_v = svmad_x(pg1, b_v, c_v, a_v);
svst1(pg1, out.v, r_v);
return out;
}
template <typename T>
inline vec<T> operator()(vec<T> a, vec<T> b){
vec<T> out;
@ -396,7 +408,7 @@ struct MultComplex{
struct MultAddComplex{
// Complex a*b+c
template <typename T>
inline void mac(const vec<T> &a, const vec<T> b, const vec<T> c){
inline vec<T> operator()(vec<T> a, vec<T> b, vec<T> c){
vec<T> out;
svbool_t pg1 = acle<T>::pg1();
typename acle<T>::vt a_v = svld1(pg1, a.v);
@ -407,7 +419,9 @@ struct MultAddComplex{
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, a.v, r_v);
svst1(pg1, out.v, r_v);
return out;
}
};

16
Grid/simd/Grid_a64fx-fixedsize.h
View File

@ -295,14 +295,14 @@ struct Sub{
struct Mult{
// Real float fma
inline void mac(vecf &a, vecf b, vecf c){
inline void operator()(vecf a, vecf b, vecf c){
pred pg1 = acle<float>::pg1();
a = svmad_x(pg1, b, c, a);
return svmad_x(pg1, b, c, a);
}
// Real double fma
inline void mac(vecd &a, vecd b, vecd c){
inline void operator()(vecd a, vecd b, vecd c){
pred pg1 = acle<double>::pg1();
a = svmad_x(pg1, b, c, a);
return svmad_x(pg1, b, c, a);
}
// Real float
inline vecf operator()(vecf a, vecf b){
@ -376,18 +376,18 @@ struct MultComplex{
struct MultAddComplex{
// Complex a*b+c
// Complex float
inline void mac(vecf &a, vecf b, vecf c){
inline vecf operator()(vecf a, vecf b, vecf c){
pred pg1 = acle<float>::pg1();
// using FCMLA
vecf r_v = svcmla_x(pg1, c, a, b, 0);
a = svcmla_x(pg1, r_v, a, b, 90);
return svcmla_x(pg1, r_v, a, b, 90);
}
// Complex double
inline void mac(vecd &a, vecd b, vecd c){
inline vecd operator()(vecd a, vecd b, vecd c){
pred pg1 = acle<double>::pg1();
// using FCMLA
vecd r_v = svcmla_x(pg1, c, a, b, 0);
a = svcmla_x(pg1, r_v, a, b, 90);
return svcmla_x(pg1, r_v, a, b, 90);
}
};