Fix non-VSX PowerPC build

This commit is contained in:
Chip Kerchner 2022-08-08 18:18:17 +00:00 committed by Rasmus Munk Larsen
parent 84a9d6fac9
commit 9e0afe0f02
4 changed files with 50 additions and 16 deletions

View File

@ -350,7 +350,7 @@ using std::ptrdiff_t;
#include "src/Core/CoreIterators.h"
#include "src/Core/ConditionEstimator.h"
#if defined(EIGEN_VECTORIZE_ALTIVEC) || defined(EIGEN_VECTORIZE_VSX)
#if defined(EIGEN_VECTORIZE_VSX)
#include "src/Core/arch/AltiVec/MatrixProduct.h"
#elif defined EIGEN_VECTORIZE_NEON
#include "src/Core/arch/NEON/GeneralBlockPanelKernel.h"

View File

@ -102,6 +102,7 @@ template<> struct packet_traits<std::complex<float> > : default_packet_traits
HasAbs2 = 0,
HasMin = 0,
HasMax = 0,
HasSqrt = 1,
#ifdef __VSX__
HasBlend = 1,
#endif
@ -370,6 +371,7 @@ template<> struct packet_traits<std::complex<double> > : default_packet_traits
HasAbs2 = 0,
HasMin = 0,
HasMax = 0,
HasSqrt = 1,
HasSetLinear = 0
};
};

View File

@ -42,16 +42,14 @@ Packet4f pcos<Packet4f>(const Packet4f& _x)
return pcos_float(_x);
}
#ifdef __VSX__
#ifndef EIGEN_COMP_CLANG
template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
Packet4f prsqrt<Packet4f>(const Packet4f& x)
{
return vec_rsqrt(x);
}
#endif
#ifdef __VSX__
#ifndef EIGEN_COMP_CLANG
template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
Packet2d prsqrt<Packet2d>(const Packet2d& x)
{
@ -76,6 +74,26 @@ Packet2d pexp<Packet2d>(const Packet2d& _x)
{
return pexp_double(_x);
}
template<> EIGEN_STRONG_INLINE Packet8bf psqrt<Packet8bf> (const Packet8bf& a){
BF16_TO_F32_UNARY_OP_WRAPPER(psqrt<Packet4f>, a);
}
#ifndef EIGEN_COMP_CLANG
template<> EIGEN_STRONG_INLINE Packet8bf prsqrt<Packet8bf> (const Packet8bf& a){
BF16_TO_F32_UNARY_OP_WRAPPER(prsqrt<Packet4f>, a);
}
#endif
#else
template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
Packet4f psqrt<Packet4f>(const Packet4f& x)
{
Packet4f a;
for (Index i = 0; i < packet_traits<float>::size; i++) {
a[i] = numext::sqrt(x[i]);
}
return a;
}
#endif
// Hyperbolic Tangent function.

View File

@ -177,16 +177,19 @@ struct packet_traits<float> : default_packet_traits {
#else
HasRsqrt = 0,
#endif
HasTanh = EIGEN_FAST_MATH,
HasErf = EIGEN_FAST_MATH,
HasRint = 1,
#else
HasSqrt = 0,
HasRsqrt = 0,
HasTanh = 0,
HasErf = 0,
HasRint = 0,
#endif
HasTanh = EIGEN_FAST_MATH,
HasErf = EIGEN_FAST_MATH,
HasRound = 1,
HasFloor = 1,
HasCeil = 1,
HasRint = 1,
HasNegate = 1,
HasBlend = 1
};
@ -219,16 +222,17 @@ struct packet_traits<bfloat16> : default_packet_traits {
#else
HasRsqrt = 0,
#endif
HasRint = 1,
#else
HasSqrt = 0,
HasRsqrt = 0,
HasTanh = EIGEN_FAST_MATH,
HasErf = EIGEN_FAST_MATH,
HasRint = 0,
#endif
HasTanh = 0,
HasErf = 0,
HasRound = 1,
HasFloor = 1,
HasCeil = 1,
HasRint = 1,
HasNegate = 1,
HasBlend = 1
};
@ -1042,9 +1046,11 @@ template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i&
template<> EIGEN_STRONG_INLINE Packet8s pmadd(const Packet8s& a, const Packet8s& b, const Packet8s& c) { return vec_madd(a,b,c); }
template<> EIGEN_STRONG_INLINE Packet8us pmadd(const Packet8us& a, const Packet8us& b, const Packet8us& c) { return vec_madd(a,b,c); }
#ifdef __VSX__
template<> EIGEN_STRONG_INLINE Packet4f pmsub(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vec_msub(a,b,c); }
template<> EIGEN_STRONG_INLINE Packet4f pnmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vec_nmsub(a,b,c); }
template<> EIGEN_STRONG_INLINE Packet4f pnmsub(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vec_nmadd(a,b,c); }
#endif
template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b)
{
@ -1089,19 +1095,29 @@ template<> EIGEN_STRONG_INLINE Packet4f pcmp_lt_or_nan(const Packet4f& a, const
return vec_nor(c,c);
}
#ifdef __VSX__
template<> EIGEN_STRONG_INLINE Packet4i pcmp_le(const Packet4i& a, const Packet4i& b) { return reinterpret_cast<Packet4i>(vec_cmple(a,b)); }
#endif
template<> EIGEN_STRONG_INLINE Packet4i pcmp_lt(const Packet4i& a, const Packet4i& b) { return reinterpret_cast<Packet4i>(vec_cmplt(a,b)); }
template<> EIGEN_STRONG_INLINE Packet4i pcmp_eq(const Packet4i& a, const Packet4i& b) { return reinterpret_cast<Packet4i>(vec_cmpeq(a,b)); }
#ifdef __VSX__
template<> EIGEN_STRONG_INLINE Packet8s pcmp_le(const Packet8s& a, const Packet8s& b) { return reinterpret_cast<Packet8s>(vec_cmple(a,b)); }
#endif
template<> EIGEN_STRONG_INLINE Packet8s pcmp_lt(const Packet8s& a, const Packet8s& b) { return reinterpret_cast<Packet8s>(vec_cmplt(a,b)); }
template<> EIGEN_STRONG_INLINE Packet8s pcmp_eq(const Packet8s& a, const Packet8s& b) { return reinterpret_cast<Packet8s>(vec_cmpeq(a,b)); }
#ifdef __VSX__
template<> EIGEN_STRONG_INLINE Packet8us pcmp_le(const Packet8us& a, const Packet8us& b) { return reinterpret_cast<Packet8us>(vec_cmple(a,b)); }
#endif
template<> EIGEN_STRONG_INLINE Packet8us pcmp_lt(const Packet8us& a, const Packet8us& b) { return reinterpret_cast<Packet8us>(vec_cmplt(a,b)); }
template<> EIGEN_STRONG_INLINE Packet8us pcmp_eq(const Packet8us& a, const Packet8us& b) { return reinterpret_cast<Packet8us>(vec_cmpeq(a,b)); }
#ifdef __VSX__
template<> EIGEN_STRONG_INLINE Packet16c pcmp_le(const Packet16c& a, const Packet16c& b) { return reinterpret_cast<Packet16c>(vec_cmple(a,b)); }
#endif
template<> EIGEN_STRONG_INLINE Packet16c pcmp_lt(const Packet16c& a, const Packet16c& b) { return reinterpret_cast<Packet16c>(vec_cmplt(a,b)); }
template<> EIGEN_STRONG_INLINE Packet16c pcmp_eq(const Packet16c& a, const Packet16c& b) { return reinterpret_cast<Packet16c>(vec_cmpeq(a,b)); }
#ifdef __VSX__
template<> EIGEN_STRONG_INLINE Packet16uc pcmp_le(const Packet16uc& a, const Packet16uc& b) { return reinterpret_cast<Packet16uc>(vec_cmple(a,b)); }
#endif
template<> EIGEN_STRONG_INLINE Packet16uc pcmp_lt(const Packet16uc& a, const Packet16uc& b) { return reinterpret_cast<Packet16uc>(vec_cmplt(a,b)); }
template<> EIGEN_STRONG_INLINE Packet16uc pcmp_eq(const Packet16uc& a, const Packet16uc& b) { return reinterpret_cast<Packet16uc>(vec_cmpeq(a,b)); }
@ -1154,6 +1170,7 @@ template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a)
}
template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const Packet4f& a) { return vec_ceil(a); }
template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) { return vec_floor(a); }
#ifdef __VSX__
template<> EIGEN_STRONG_INLINE Packet4f print<Packet4f>(const Packet4f& a)
{
Packet4f res;
@ -1164,6 +1181,7 @@ template<> EIGEN_STRONG_INLINE Packet4f print<Packet4f>(const Packet4f& a)
return res;
}
#endif
template<typename Packet> EIGEN_STRONG_INLINE Packet ploadu_common(const __UNPACK_TYPE__(Packet)* from)
{
@ -1706,12 +1724,6 @@ template<> EIGEN_STRONG_INLINE Packet8bf psub<Packet8bf>(const Packet8bf& a, con
BF16_TO_F32_BINARY_OP_WRAPPER(psub<Packet4f>, a, b);
}
template<> EIGEN_STRONG_INLINE Packet8bf psqrt<Packet8bf> (const Packet8bf& a){
BF16_TO_F32_UNARY_OP_WRAPPER(vec_sqrt, a);
}
template<> EIGEN_STRONG_INLINE Packet8bf prsqrt<Packet8bf> (const Packet8bf& a){
BF16_TO_F32_UNARY_OP_WRAPPER(prsqrt<Packet4f>, a);
}
template<> EIGEN_STRONG_INLINE Packet8bf pexp<Packet8bf> (const Packet8bf& a){
BF16_TO_F32_UNARY_OP_WRAPPER(pexp_float, a);
}
@ -1755,9 +1767,11 @@ template<> EIGEN_STRONG_INLINE Packet8bf pceil<Packet8bf> (const Packet8bf& a){
template<> EIGEN_STRONG_INLINE Packet8bf pround<Packet8bf> (const Packet8bf& a){
BF16_TO_F32_UNARY_OP_WRAPPER(pround<Packet4f>, a);
}
#ifdef __VSX__
template<> EIGEN_STRONG_INLINE Packet8bf print<Packet8bf> (const Packet8bf& a){
BF16_TO_F32_UNARY_OP_WRAPPER(print<Packet4f>, a);
}
#endif
template<> EIGEN_STRONG_INLINE Packet8bf pmadd(const Packet8bf& a, const Packet8bf& b, const Packet8bf& c) {
Packet4f a_even = Bf16ToF32Even(a);
Packet4f a_odd = Bf16ToF32Odd(a);