mirror/eigen
2
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-03-07 18:27:40 +08:00
Code Issues Packages Projects Releases Wiki Activity

add psqrt ops support packet2f/packet4f for NEON

Browse Source
This commit is contained in:
Guoqiang QI 2020-08-21 03:17:15 +00:00 committed by Rasmus Munk Larsen
parent 1b1082334b
commit 8bb0febaf9
1 changed files with 41 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

42
Eigen/src/Core/arch/NEON/PacketMath.h
View File

@ -143,7 +143,7 @@ struct packet_traits<float> : default_packet_traits
HasCos = EIGEN_FAST_MATH,
HasLog = 1,
HasExp = 1,
HasSqrt = 0,
HasSqrt = 1,
HasTanh = EIGEN_FAST_MATH,
HasErf = EIGEN_FAST_MATH
};
@ -3218,6 +3218,46 @@ template<> EIGEN_STRONG_INLINE Packet4ui psqrt(const Packet4ui& a) {
return res;
}
#if EIGEN_FAST_MATH
/* Functions for sqrt support packet2f/packet4f.*/
// The EIGEN_FAST_MATH version uses the vrsqrte_f32 approximation and one step
// of Newton's method, at a cost of 1-2 bits of precision as opposed to the
// exact solution. It does not handle +inf, or denormalized numbers correctly.
// The main advantage of this approach is not just speed, but also the fact that
// it can be inlined and pipelined with other computations, further reducing its
// effective latency. This is similar to Quake3's fast inverse square root.
// For more details see: http://www.beyond3d.com/content/articles/8/
template<> EIGEN_STRONG_INLINE Packet4f psqrt(const Packet4f& _x){
Packet4f half = vmulq_n_f32(_x, 0.5f);
Packet4ui denormal_mask = vandq_u32(vcgeq_f32(_x, vdupq_n_f32(0.0f)),
vcltq_f32(_x, pset1<Packet4f>((std::numeric_limits<float>::min)())));
// Compute approximate reciprocal sqrt.
Packet4f x = vrsqrteq_f32(_x);
// Do a single step of Newton's iteration.
//the number 1.5f was set reference to Quake3's fast inverse square root
x = vmulq_f32(x, psub(pset1<Packet4f>(1.5f), pmul(half, pmul(x, x))));
// Flush results for denormals to zero.
return vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(pmul(_x, x)), denormal_mask));
}
template<> EIGEN_STRONG_INLINE Packet2f psqrt(const Packet2f& _x){
Packet2f half = vmul_n_f32(_x, 0.5f);
Packet2ui denormal_mask = vand_u32(vcge_f32(_x, vdup_n_f32(0.0f)),
vclt_f32(_x, pset1<Packet2f>((std::numeric_limits<float>::min)())));
// Compute approximate reciprocal sqrt.
Packet2f x = vrsqrte_f32(_x);
// Do a single step of Newton's iteration.
x = vmul_f32(x, psub(pset1<Packet2f>(1.5f), pmul(half, pmul(x, x))));
// Flush results for denormals to zero.
return vreinterpret_f32_u32(vbic_u32(vreinterpret_u32_f32(pmul(_x, x)), denormal_mask));
}
#else
template<> EIGEN_STRONG_INLINE Packet4f psqrt(const Packet4f& _x){return vsqrtq_f32(_x);}
template<> EIGEN_STRONG_INLINE Packet2f psqrt(const Packet2f& _x){return vsqrt_f32(_x); }
#endif
//---------- bfloat16 ----------
// TODO: Add support for native armv8.6-a bfloat16_t