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

81
Eigen/src/Core/arch/SSE/PacketMath.h
View File

@@ -198,12 +198,6 @@ struct packet_traits<float> : default_packet_traits {
HasTanh = EIGEN_FAST_MATH,
HasErf = EIGEN_FAST_MATH,
HasBlend = 1,
HasCeil = 1,
HasFloor = 1,
#ifdef EIGEN_VECTORIZE_SSE4_1
HasRound = 1,
#endif
HasRint = 1,
HasSign = 0 // The manually vectorized version is slightly slower for SSE.
};
};
@@ -225,13 +219,7 @@ struct packet_traits<double> : default_packet_traits {
HasSqrt = 1,
HasRsqrt = 1,
HasATan = 1,
HasBlend = 1,
HasFloor = 1,
HasCeil = 1,
#ifdef EIGEN_VECTORIZE_SSE4_1
HasRound = 1,
#endif
HasRint = 1
HasBlend = 1
};
};
template <>
@@ -1309,73 +1297,14 @@ template <>
EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) {
return _mm_floor_pd(a);
}
#else
template <>
EIGEN_STRONG_INLINE Packet4f print(const Packet4f& a) {
// Adds and subtracts signum(a) * 2^23 to force rounding.
const Packet4f limit = pset1<Packet4f>(static_cast<float>(1 << 23));
const Packet4f abs_a = pabs(a);
Packet4f r = padd(abs_a, limit);
// Don't compile-away addition and subtraction.
EIGEN_OPTIMIZATION_BARRIER(r);
r = psub(r, limit);
// If greater than limit, simply return a. Otherwise, account for sign.
r = pselect(pcmp_lt(abs_a, limit), pselect(pcmp_lt(a, pzero(a)), pnegate(r), r), a);
return r;
}
template <>
EIGEN_STRONG_INLINE Packet2d print(const Packet2d& a) {
// Adds and subtracts signum(a) * 2^52 to force rounding.
const Packet2d limit = pset1<Packet2d>(static_cast<double>(1ull << 52));
const Packet2d abs_a = pabs(a);
Packet2d r = padd(abs_a, limit);
// Don't compile-away addition and subtraction.
EIGEN_OPTIMIZATION_BARRIER(r);
r = psub(r, limit);
// If greater than limit, simply return a. Otherwise, account for sign.
r = pselect(pcmp_lt(abs_a, limit), pselect(pcmp_lt(a, pzero(a)), pnegate(r), r), a);
return r;
EIGEN_STRONG_INLINE Packet4f ptrunc<Packet4f>(const Packet4f& a) {
return _mm_round_ps(a, _MM_FROUND_TRUNC);
}
template <>
EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) {
const Packet4f cst_1 = pset1<Packet4f>(1.0f);
Packet4f tmp = print<Packet4f>(a);
// If greater, subtract one.
Packet4f mask = _mm_cmpgt_ps(tmp, a);
mask = pand(mask, cst_1);
return psub(tmp, mask);
}
template <>
EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) {
const Packet2d cst_1 = pset1<Packet2d>(1.0);
Packet2d tmp = print<Packet2d>(a);
// If greater, subtract one.
Packet2d mask = _mm_cmpgt_pd(tmp, a);
mask = pand(mask, cst_1);
return psub(tmp, mask);
}
template <>
EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const Packet4f& a) {
const Packet4f cst_1 = pset1<Packet4f>(1.0f);
Packet4f tmp = print<Packet4f>(a);
// If smaller, add one.
Packet4f mask = _mm_cmplt_ps(tmp, a);
mask = pand(mask, cst_1);
return padd(tmp, mask);
}
template <>
EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const Packet2d& a) {
const Packet2d cst_1 = pset1<Packet2d>(1.0);
Packet2d tmp = print<Packet2d>(a);
// If smaller, add one.
Packet2d mask = _mm_cmplt_pd(tmp, a);
mask = pand(mask, cst_1);
return padd(tmp, mask);
EIGEN_STRONG_INLINE Packet2d ptrunc<Packet2d>(const Packet2d& a) {
return _mm_round_pd(a, _MM_FROUND_TRUNC);
}
#endif