devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Special function implementations for half/bfloat16 packets.

Browse Source 
Current implementations fail to consider half-float packets, only
half-float scalars.  Added specializations for packets on AVX, AVX512 and
NEON.  Added tests to `special_packetmath`.

The current `special_functions` tests would fail for half and bfloat16 due to
lack of precision. The NEON tests also fail with precision issues and
due to different handling of `sqrt(inf)`, so special functions bessel, ndtri
have been disabled.

Tested with AVX, AVX512.
This commit is contained in:
Antonio Sanchez
2020-12-02 14:00:57 -08:00
parent 305b8bd277
commit e2f21465fe
17 changed files with 418 additions and 159 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
Eigen/src/Core/arch/NEON/PacketMath.h
View File

@@ -192,7 +192,9 @@ struct packet_traits<float> : default_packet_traits
HasExp = 1,
HasSqrt = 1,
HasTanh = EIGEN_FAST_MATH,
HasErf = EIGEN_FAST_MATH
HasErf = EIGEN_FAST_MATH,
HasBessel = 0, // Issues with accuracy.
HasNdtri = 0
};
};
@@ -3321,7 +3323,9 @@ template<> struct packet_traits<bfloat16> : default_packet_traits
HasExp = 1,
HasSqrt = 0,
HasTanh = EIGEN_FAST_MATH,
HasErf = EIGEN_FAST_MATH
HasErf = EIGEN_FAST_MATH,
HasBessel = 0, // Issues with accuracy.
HasNdtri = 0,
};
};
@@ -3887,7 +3891,10 @@ struct packet_traits<Eigen::half> : default_packet_traits {
HasCos = 0,
HasLog = 0,
HasExp = 0,
HasSqrt = 1
HasSqrt = 1,
HasErf = EIGEN_FAST_MATH,
HasBessel = 0, // Issues with accuracy.
HasNdtri = 0,
};
};