Collapsed revision

* Add packet up "pones". Write pnot(a) as pxor(pones(a), a).
* Collapsed revision
* Simplify a bit.
* Undo useless diffs.
* Fix typo.
This commit is contained in:
Rasmus Munk Larsen
2019-01-09 16:34:23 -08:00
parent cb3c059fa4
commit e15bb785ad
9 changed files with 104 additions and 53 deletions

View File

@@ -82,6 +82,9 @@ template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, con
#endif
}
template<> EIGEN_STRONG_INLINE Packet2cf pones <Packet2cf>(const Packet2cf& a) { return Packet2cf(pones(a.v)); }
template<> EIGEN_STRONG_INLINE Packet2cf pnot <Packet2cf>(const Packet2cf& a) { return Packet2cf(pnot(a.v)); }
template<> EIGEN_STRONG_INLINE Packet2cf pand <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_and_ps(a.v,b.v)); }
template<> EIGEN_STRONG_INLINE Packet2cf por <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_or_ps(a.v,b.v)); }
template<> EIGEN_STRONG_INLINE Packet2cf pxor <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_xor_ps(a.v,b.v)); }
@@ -305,6 +308,8 @@ template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, con
#endif
}
template<> EIGEN_STRONG_INLINE Packet1cd pones <Packet1cd>(const Packet1cd& a) { return Packet1cd(pones(a.v)); }
template<> EIGEN_STRONG_INLINE Packet1cd pnot <Packet1cd>(const Packet1cd& a) { return Packet1cd(pnot(a.v)); }
template<> EIGEN_STRONG_INLINE Packet1cd pand <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_and_pd(a.v,b.v)); }
template<> EIGEN_STRONG_INLINE Packet1cd por <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_or_pd(a.v,b.v)); }
template<> EIGEN_STRONG_INLINE Packet1cd pxor <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_xor_pd(a.v,b.v)); }
@@ -442,15 +447,13 @@ ptranspose(PacketBlock<Packet2cf,2>& kernel) {
template<> EIGEN_STRONG_INLINE Packet2cf pcmp_eq(const Packet2cf& a, const Packet2cf& b)
{
__m128 eq = _mm_cmpeq_ps(a.v, b.v);
__m128 real_and_imag_equal = _mm_and_ps(eq, vec4f_swizzle1(eq, 1, 0, 3, 2));
return Packet2cf(real_and_imag_equal);
return Packet2cf(pand(eq, vec4f_swizzle1(eq, 1, 0, 3, 2)));
}
template<> EIGEN_STRONG_INLINE Packet1cd pcmp_eq(const Packet1cd& a, const Packet1cd& b)
{
__m128d eq = _mm_cmpeq_pd(a.v, b.v);
__m128d real_and_imag_equal = _mm_and_pd(eq, vec2d_swizzle1(eq, 1, 0));
return Packet1cd(real_and_imag_equal);
return Packet1cd(pand(eq, vec2d_swizzle1(eq, 1, 0)));
}
template<> EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, const Packet2cf& thenPacket, const Packet2cf& elsePacket) {

View File

@@ -378,6 +378,17 @@ template<> EIGEN_STRONG_INLINE Packet4i pcmp_eq(const Packet4i& a, const Packet4
template<> EIGEN_STRONG_INLINE Packet2d pcmp_eq(const Packet2d& a, const Packet2d& b) { return _mm_cmpeq_pd(a,b); }
template<> EIGEN_STRONG_INLINE Packet4f pcmp_lt_or_nan(const Packet4f& a, const Packet4f& b) { return _mm_cmpnge_ps(a,b); }
template<> EIGEN_STRONG_INLINE Packet4i pones<Packet4i>(const Packet4i& a) { return _mm_cmpeq_epi32(a, a); }
template<> EIGEN_STRONG_INLINE Packet4f
pones<Packet4f>(const Packet4f& a) {
Packet4i b = _mm_castps_si128(a);
return _mm_castsi128_ps(_mm_cmpeq_epi32(b, b));
}
template<> EIGEN_STRONG_INLINE Packet2d
pones<Packet2d>(const Packet2d& a) {
Packet4i b = _mm_castpd_si128(a);
return _mm_castsi128_pd(_mm_cmpeq_epi32(b, b));
}
template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_and_ps(a,b); }
template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_and_pd(a,b); }