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

Add truncation op

Browse Source
This commit is contained in:
Charles Schlosser
2024-04-29 23:45:49 +00:00
committed by Rasmus Munk Larsen
parent d5524fc57b
commit fb95e90f7f
23 changed files with 333 additions and 267 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
Eigen/src/Core/arch/AVX/PacketMath.h
View File

@@ -124,11 +124,7 @@ struct packet_traits<float> : default_packet_traits {
HasRsqrt = 1,
HasTanh = EIGEN_FAST_MATH,
HasErf = EIGEN_FAST_MATH,
HasBlend = 1,
HasRound = 1,
HasFloor = 1,
HasCeil = 1,
HasRint = 1
HasBlend = 1
};
};
template <>
@@ -151,11 +147,7 @@ struct packet_traits<double> : default_packet_traits {
HasSqrt = 1,
HasRsqrt = 1,
HasATan = 1,
HasBlend = 1,
HasRound = 1,
HasFloor = 1,
HasCeil = 1,
HasRint = 1
HasBlend = 1
};
};
@@ -192,10 +184,6 @@ struct packet_traits<Eigen::half> : default_packet_traits {
HasTanh = EIGEN_FAST_MATH,
HasErf = EIGEN_FAST_MATH,
HasBlend = 0,
HasRound = 1,
HasFloor = 1,
HasCeil = 1,
HasRint = 1,
HasBessel = 1,
HasNdtri = 1
};
@@ -235,10 +223,6 @@ struct packet_traits<bfloat16> : default_packet_traits {
HasTanh = EIGEN_FAST_MATH,
HasErf = EIGEN_FAST_MATH,
HasBlend = 0,
HasRound = 1,
HasFloor = 1,
HasCeil = 1,
HasRint = 1,
HasBessel = 1,
HasNdtri = 1
};
@@ -1257,6 +1241,15 @@ EIGEN_STRONG_INLINE Packet4d pfloor<Packet4d>(const Packet4d& a) {
return _mm256_floor_pd(a);
}
template <>
EIGEN_STRONG_INLINE Packet8f ptrunc<Packet8f>(const Packet8f& a) {
return _mm256_round_ps(a, _MM_FROUND_TRUNC);
}
template <>
EIGEN_STRONG_INLINE Packet4d ptrunc<Packet4d>(const Packet4d& a) {
return _mm256_round_pd(a, _MM_FROUND_TRUNC);
}
template <>
EIGEN_STRONG_INLINE Packet8i ptrue<Packet8i>(const Packet8i& a) {
#ifdef EIGEN_VECTORIZE_AVX2
@@ -2311,6 +2304,11 @@ EIGEN_STRONG_INLINE Packet8h pfloor<Packet8h>(const Packet8h& a) {
return float2half(pfloor<Packet8f>(half2float(a)));
}
template <>
EIGEN_STRONG_INLINE Packet8h ptrunc<Packet8h>(const Packet8h& a) {
return float2half(ptrunc<Packet8f>(half2float(a)));
}
template <>
EIGEN_STRONG_INLINE Packet8h pcmp_eq(const Packet8h& a, const Packet8h& b) {
return Pack16To8(pcmp_eq(half2float(a), half2float(b)));
@@ -2686,6 +2684,11 @@ EIGEN_STRONG_INLINE Packet8bf pfloor<Packet8bf>(const Packet8bf& a) {
return F32ToBf16(pfloor<Packet8f>(Bf16ToF32(a)));
}
template <>
EIGEN_STRONG_INLINE Packet8bf ptrunc<Packet8bf>(const Packet8bf& a) {
return F32ToBf16(ptrunc<Packet8f>(Bf16ToF32(a)));
}
template <>
EIGEN_STRONG_INLINE Packet8bf pcmp_eq(const Packet8bf& a, const Packet8bf& b) {
return Pack16To8(pcmp_eq(Bf16ToF32(a), Bf16ToF32(b)));