Implement vectorized versions of log1p and expm1 in Eigen using Kahan's formulas, and change the scalar implementations to properly handle infinite arguments.

Depending on instruction set, significant speedups are observed for the vectorized path:
log1p wall time is reduced 60-93% (2.5x - 15x speedup)
expm1 wall time is reduced 0-85% (1x - 7x speedup)

The scalar path is slower by 20-30% due to the extra branch needed to handle +infinity correctly.

Full benchmarks measured on Intel(R) Xeon(R) Gold 6154 here: https://bitbucket.org/snippets/rmlarsen/MXBkpM
This commit is contained in:
Rasmus Munk Larsen
2019-08-12 13:53:28 -07:00
parent d55d392e7b
commit a3298b22ec
9 changed files with 95 additions and 6 deletions

View File

@@ -604,11 +604,13 @@ template<typename Scalar,typename Packet> void packetmath_real()
CHECK_CWISE1_IF(PacketTraits::HasSqrt, Scalar(1)/std::sqrt, internal::prsqrt);
CHECK_CWISE1_IF(PacketTraits::HasLog, std::log, internal::plog);
#if EIGEN_HAS_C99_MATH && (__cplusplus > 199711L)
CHECK_CWISE1_IF(PacketTraits::HasExpm1, std::expm1, internal::pexpm1);
CHECK_CWISE1_IF(PacketTraits::HasLog1p, std::log1p, internal::plog1p);
CHECK_CWISE1_IF(internal::packet_traits<Scalar>::HasLGamma, std::lgamma, internal::plgamma);
CHECK_CWISE1_IF(internal::packet_traits<Scalar>::HasErf, std::erf, internal::perf);
CHECK_CWISE1_IF(internal::packet_traits<Scalar>::HasErfc, std::erfc, internal::perfc);
data1[0] = std::numeric_limits<Scalar>::infinity();
data1[1] = std::numeric_limits<Scalar>::denorm_min();
CHECK_CWISE1_IF(PacketTraits::HasExpm1, std::expm1, internal::pexpm1);
CHECK_CWISE1_IF(PacketTraits::HasLog1p, std::log1p, internal::plog1p);
#endif
if(PacketSize>=2)