portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-17 07:17:06 +01:00
Code Releases Activity

introduce AddTimesI and SubTimesI; slight benefit in operators, but < 1%; breaks all other impls

Browse Source
This commit is contained in:
nmeyer-ur
2020-06-03 15:20:13 +02:00
parent 5ee3ea2144
commit 9872c76825
4 changed files with 256 additions and 78 deletions
Download Patch File Download Diff File Expand all files Collapse all files

41
Grid/simd/Grid_a64fx-fixedsize.h
View File

@ -443,8 +443,8 @@ struct TimesMinusI{
};
// alternative implementation using fcadd
// this is not optimal because we have op1 = op2 + TimesMinusI(op3) etc
// ideally we have AddTimesMinusI(op1,op2,op3)
// this is not optimal because we have op1 = op2 + TimesMinusI(op3) = op2 - TimesI(op3) etc
// but ideally we have op1 = SubTimesI(op2,op3)
//
// makes performance worse in Benchmark_wilson using MPI
// increases halogtime and gathertime
@ -467,6 +467,34 @@ struct TimesMinusI{
};
*/
// SVE only, fcadd returns a +- i*b
// a + i * b
struct AddTimesI{
// Complex float
inline vecf operator()(vecf a, vecf b){
pred pg1 = acle<float>::pg1();
return svcadd_x(pg1, a, b, 90);
}
// Complex double
inline vecd operator()(vecd a, vecd b){
pred pg1 = acle<double>::pg1();
return svcadd_x(pg1, a, b, 90);
}
};
// a - i * b
struct SubTimesI{
// Complex float
inline vecf operator()(vecf a, vecf b){
pred pg1 = acle<float>::pg1();
return svcadd_x(pg1, a, b, 270);
}
// Complex double
inline vecd operator()(vecd a, vecd b){
pred pg1 = acle<double>::pg1();
return svcadd_x(pg1, a, b, 270);
}
};
struct TimesI{
// Complex float
inline vecf operator()(vecf a, vecf b){
@ -493,7 +521,7 @@ struct TimesI{
// alternative implementation using fcadd
// this is not optimal because we have op1 = op2 + TimesI(op3) etc
// ideally we have AddTimesI(op1,op2,op3)
// ideally we have op1 = AddTimesI(op2,op3)
//
// makes performance worse in Benchmark_wilson using MPI
// increases halogtime and gathertime
@ -800,7 +828,7 @@ typedef veci SIMD_Itype; // Integer type
// prefetch utilities
inline void v_prefetch0(int size, const char *ptr){};
/* PF 256 worse than PF 64
/* PF 256
inline void prefetch_HINT_T0(const char *ptr){
static int64_t last_ptr;
int64_t vptr = reinterpret_cast<std::intptr_t>(ptr) & 0x7fffffffffffff00ll;
@ -812,7 +840,7 @@ inline void prefetch_HINT_T0(const char *ptr){
}
};
*/
/* beneficial for operators?
/* PF 64
inline void prefetch_HINT_T0(const char *ptr){
pred pg1 = Optimization::acle<double>::pg1();
svprfd(pg1, ptr, SV_PLDL1STRM);
@ -839,5 +867,8 @@ typedef Optimization::MaddRealPart MaddRealPartSIMD;
typedef Optimization::Conj ConjSIMD;
typedef Optimization::TimesMinusI TimesMinusISIMD;
typedef Optimization::TimesI TimesISIMD;
typedef Optimization::AddTimesI AddTimesISIMD;
typedef Optimization::SubTimesI SubTimesISIMD;
NAMESPACE_END(Grid);

43
Grid/simd/Grid_vector_types.h
View File

@ -298,7 +298,7 @@ public:
// FIXME -- alias this to an accelerator_inline MAC struct.
// FIXME VLA build error
// specialize mac for A64FX
#if defined(A64FX) || defined(A64FXFIXEDSIZE)
friend accelerator_inline void mac(Grid_simd *__restrict__ y,
const Grid_simd *__restrict__ a,
@ -894,6 +894,47 @@ accelerator_inline Grid_simd<S, V> timesI(const Grid_simd<S, V> &in) {
return in;
}
// -----------------------------------------------------------------------------
// SVE only
///////////////////////
// AddTimesI
///////////////////////
template <class S, class V, IfComplex<S> = 0>
accelerator_inline void addTimesI(Grid_simd<S, V> &ret, const Grid_simd<S, V> &in1, const Grid_simd<S, V> &in2) {
ret.v = binary<V>(in1.v, in2.v, AddTimesISIMD());
}
template <class S, class V, IfComplex<S> = 0>
accelerator_inline Grid_simd<S, V> addTimesI(const Grid_simd<S, V> &in1, const Grid_simd<S, V> &in2) {
Grid_simd<S, V> ret;
ret = addTimesI(in1, in2);
return ret;
}
template <class S, class V, IfNotComplex<S> = 0>
accelerator_inline Grid_simd<S, V> addTimesI(const Grid_simd<S, V> &in1, const Grid_simd<S, V> &in2) {
return in1;
}
///////////////////////
// SubTimesI
///////////////////////
template <class S, class V, IfComplex<S> = 0>
accelerator_inline void subTimesI(Grid_simd<S, V> &ret, const Grid_simd<S, V> &in1, const Grid_simd<S, V> &in2) {
ret.v = binary<V>(in1.v, in2.v, SubTimesISIMD());
}
template <class S, class V, IfComplex<S> = 0>
accelerator_inline Grid_simd<S, V> subTimesI(const Grid_simd<S, V> &in1, const Grid_simd<S, V> &in2) {
Grid_simd<S, V> ret;
ret = subTimesI(in1, in2);
return ret;
}
template <class S, class V, IfNotComplex<S> = 0>
accelerator_inline Grid_simd<S, V> subTimesI(const Grid_simd<S, V> &in1, const Grid_simd<S, V> &in2) {
return in1;
}
// end SVE
// -----------------------------------------------------------------------------
/////////////////////
// Inner, outer
/////////////////////