1
0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-12 20:27:06 +01:00

Reworking to keep intel compiler happy

This commit is contained in:
Peter Boyle
2015-05-19 21:29:07 +01:00
parent a4b3bc7714
commit b562b50196
8 changed files with 62 additions and 24 deletions

View File

@ -345,20 +345,30 @@ friend inline void vstore(const vComplexD &ret, ComplexD *a){
// REDUCE FIXME must be a cleaner implementation
friend inline ComplexD Reduce(const vComplexD & in)
{
vComplexD v1,v2;
union {
zvec v;
double f[sizeof(zvec)/sizeof(double)];
} conv;
#ifdef SSE4
return ComplexD(in.v[0],in.v[1]);
v1=in;
#endif
#if defined(AVX1) || defined (AVX2)
vComplexD v1;
permute(v1,in,0); // sse 128; paired complex single
v1=v1+in;
return ComplexD(v1.v[0],v1.v[1]);
#endif
#ifdef AVX512
return ComplexD(_mm512_mask_reduce_add_pd(0x55, in.v),_mm512_mask_reduce_add_pd(0xAA, in.v));
permute(v1,in,0); // sse 128; paired complex single
v1=v1+in;
permute(v2,v1,1); // avx 256; quad complex single
v1=v1+v2;
#endif
#ifdef QPX
#error
#endif
conv.v = v1.v;
return ComplexD(conv.f[0],conv.f[1]);
}
// Unary negation

View File

@ -234,26 +234,34 @@ namespace Grid {
}
friend inline ComplexF Reduce(const vComplexF & in)
{
vComplexF v1,v2;
union {
cvec v;
float f[sizeof(cvec)/sizeof(float)];
} conv;
#ifdef SSE4
vComplexF v1;
permute(v1,in,0); // sse 128; paired complex single
v1=v1+in;
return ComplexF(v1.v[0],v1.v[1]);
#endif
#if defined(AVX1) || defined (AVX2)
vComplexF v1,v2;
permute(v1,in,0); // sse 128; paired complex single
v1=v1+in;
permute(v2,v1,1); // avx 256; quad complex single
v1=v1+v2;
return ComplexF(v1.v[0],v1.v[1]);
#endif
#ifdef AVX512
return ComplexF(_mm512_mask_reduce_add_ps(0x5555, in.v),_mm512_mask_reduce_add_ps(0xAAAA, in.v));
permute(v1,in,0); // avx512 octo-complex single
v1=v1+in;
permute(v2,v1,1);
v1=v1+v2;
permute(v2,v1,2);
v1=v1+v2;
#endif
#ifdef QPX
#error
#endif
conv.v = v1.v;
return ComplexF(conv.f[0],conv.f[1]);
}
friend inline vComplexF operator * (const ComplexF &a, vComplexF b){

View File

@ -210,25 +210,33 @@ namespace Grid {
friend inline RealD Reduce(const vRealD & in)
{
vRealD v1,v2;
union {
dvec v;
double f[sizeof(dvec)/sizeof(double)];
} conv;
#ifdef SSE4
vRealD v1;
permute(v1,in,0); // sse 128; paired real double
v1=v1+in;
return RealD(v1.v[0]);
#endif
#if defined(AVX1) || defined (AVX2)
vRealD v1,v2;
permute(v1,in,0); // avx 256; quad double
v1=v1+in;
permute(v2,v1,1);
v1=v1+v2;
return v1.v[0];
#endif
#ifdef AVX512
return _mm512_reduce_add_pd(in.v);
permute(v1,in,0); // avx 512; octo-double
v1=v1+in;
permute(v2,v1,1);
v1=v1+v2;
permute(v2,v1,2);
v1=v1+v2;
#endif
#ifdef QPX
#endif
conv.v=v1.v;
return conv.f[0];
}
// *=,+=,-= operators

View File

@ -243,29 +243,39 @@ friend inline void vstore(const vRealF &ret, float *a){
}
friend inline RealF Reduce(const vRealF & in)
{
#ifdef SSE4
vRealF v1,v2;
union {
fvec v;
float f[sizeof(fvec)/sizeof(double)];
} conv;
#ifdef SSE4
permute(v1,in,0); // sse 128; quad single
v1=v1+in;
permute(v2,v1,1);
v1=v1+v2;
return v1.v[0];
#endif
#if defined(AVX1) || defined (AVX2)
vRealF v1,v2;
permute(v1,in,0); // avx 256; octo-double
v1=v1+in;
permute(v2,v1,1);
v1=v1+v2;
permute(v2,v1,2);
v1=v1+v2;
return v1.v[0];
#endif
#ifdef AVX512
return _mm512_reduce_add_ps(in.v);
permute(v1,in,0); // avx 256; octo-double
v1=v1+in;
permute(v2,v1,1);
v1=v1+v2;
permute(v2,v1,2);
v1=v1+v2;
permute(v2,v1,3);
v1=v1+v2;
#endif
#ifdef QPX
#endif
conv.v=v1.v;
return conv.f[0];
}
// *=,+=,-= operators