Generalized predux4 to support AVX512 packets, and renamed it predux_half.

Disabled the implementation of pabs for avx512 since the corresponding intrinsics are not shipped with gcc
This commit is contained in:
Benoit Steiner 2016-02-01 14:35:51 -08:00
parent c1a42c2d0d
commit 85b6d82b49
3 changed files with 8 additions and 6 deletions

View File

@ -314,7 +314,7 @@ template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Pack
*/ */
template<typename Packet> EIGEN_DEVICE_FUNC inline template<typename Packet> EIGEN_DEVICE_FUNC inline
typename conditional<(unpacket_traits<Packet>::size%8)==0,typename unpacket_traits<Packet>::half,Packet>::type typename conditional<(unpacket_traits<Packet>::size%8)==0,typename unpacket_traits<Packet>::half,Packet>::type
predux4(const Packet& a) predux_half(const Packet& a)
{ return a; } { return a; }
/** \internal \returns the product of the elements of \a a*/ /** \internal \returns the product of the elements of \a a*/

View File

@ -401,7 +401,7 @@ template<> EIGEN_STRONG_INLINE double predux<Packet4d>(const Packet4d& a)
return pfirst(_mm256_hadd_pd(tmp0,tmp0)); return pfirst(_mm256_hadd_pd(tmp0,tmp0));
} }
template<> EIGEN_STRONG_INLINE Packet4f predux4<Packet8f>(const Packet8f& a) template<> EIGEN_STRONG_INLINE Packet4f predux_half<Packet8f>(const Packet8f& a)
{ {
return _mm_add_ps(_mm256_castps256_ps128(a),_mm256_extractf128_ps(a,1)); return _mm_add_ps(_mm256_castps256_ps128(a),_mm256_extractf128_ps(a,1));
} }

View File

@ -633,11 +633,13 @@ template<> EIGEN_STRONG_INLINE Packet8d preverse(const Packet8d& a)
template<> EIGEN_STRONG_INLINE Packet16f pabs(const Packet16f& a) template<> EIGEN_STRONG_INLINE Packet16f pabs(const Packet16f& a)
{ {
return _mm512_abs_ps(a); assert(false && "to be implemented");
// return _mm512_abs_ps(a);
} }
template<> EIGEN_STRONG_INLINE Packet8d pabs(const Packet8d& a) template<> EIGEN_STRONG_INLINE Packet8d pabs(const Packet8d& a)
{ {
return _mm512_abs_pd(a); assert(false && "to be implemented");
// return _mm512_abs_pd(a);
} }
template<> EIGEN_STRONG_INLINE Packet16f preduxp<Packet16f>(const Packet16f* vecs) template<> EIGEN_STRONG_INLINE Packet16f preduxp<Packet16f>(const Packet16f* vecs)
@ -679,7 +681,7 @@ EIGEN_STRONG_INLINE double predux<Packet8d>(const Packet8d& a) {
} }
template <> template <>
EIGEN_STRONG_INLINE Packet8f predux4<Packet16f>(const Packet16f& a) { EIGEN_STRONG_INLINE Packet8f predux_half<Packet16f>(const Packet16f& a) {
#ifdef EIGEN_VECTORIZE_AVX512DQ #ifdef EIGEN_VECTORIZE_AVX512DQ
Packet8f lane0 = _mm512_extractf32x8_ps(a, 0); Packet8f lane0 = _mm512_extractf32x8_ps(a, 0);
Packet8f lane1 = _mm512_extractf32x8_ps(a, 1); Packet8f lane1 = _mm512_extractf32x8_ps(a, 1);
@ -695,7 +697,7 @@ EIGEN_STRONG_INLINE Packet8f predux4<Packet16f>(const Packet16f& a) {
#endif #endif
} }
template <> template <>
EIGEN_STRONG_INLINE Packet4d predux4<Packet8d>(const Packet8d& a) { EIGEN_STRONG_INLINE Packet4d predux_half<Packet8d>(const Packet8d& a) {
Packet4d lane0 = _mm512_extractf64x4_pd(a, 0); Packet4d lane0 = _mm512_extractf64x4_pd(a, 0);
Packet4d lane1 = _mm512_extractf64x4_pd(a, 1); Packet4d lane1 = _mm512_extractf64x4_pd(a, 1);
Packet4d res = padd(lane0, lane1); Packet4d res = padd(lane0, lane1);