diff options
| -rw-r--r-- | Eigen/src/Core/GenericPacketMath.h | 41 | ||||
| -rw-r--r-- | Eigen/src/Core/arch/AVX/PacketMath.h | 14 | ||||
| -rw-r--r-- | Eigen/src/Core/arch/AVX512/PacketMath.h | 10 | ||||
| -rwxr-xr-x | Eigen/src/Core/arch/AltiVec/PacketMath.h | 15 | ||||
| -rw-r--r-- | Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h | 8 | ||||
| -rw-r--r-- | Eigen/src/Core/arch/NEON/PacketMath.h | 99 | ||||
| -rwxr-xr-x | Eigen/src/Core/arch/SSE/PacketMath.h | 6 | ||||
| -rw-r--r-- | Eigen/src/Core/functors/UnaryFunctors.h | 38 | ||||
| -rw-r--r-- | Eigen/src/plugins/CommonCwiseUnaryOps.h | 43 |
9 files changed, 209 insertions, 65 deletions
diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h index 146d34fb5..42d1eb077 100644 --- a/Eigen/src/Core/GenericPacketMath.h +++ b/Eigen/src/Core/GenericPacketMath.h @@ -44,19 +44,20 @@ struct default_packet_traits enum { HasHalfPacket = 0, - HasAdd = 1, - HasSub = 1, - HasMul = 1, - HasNegate = 1, - HasAbs = 1, - HasArg = 0, - HasAbs2 = 1, - HasMin = 1, - HasMax = 1, - HasConj = 1, + HasAdd = 1, + HasSub = 1, + HasShift = 1, + HasMul = 1, + HasNegate = 1, + HasAbs = 1, + HasArg = 0, + HasAbs2 = 1, + HasMin = 1, + HasMax = 1, + HasConj = 1, HasSetLinear = 1, - HasBlend = 0, - HasReduxp = 1, + HasBlend = 0, + HasReduxp = 1, HasDiv = 0, HasSqrt = 0, @@ -230,17 +231,23 @@ EIGEN_DEVICE_FUNC inline std::complex<RealScalar> ptrue(const std::complex<RealS template <typename Packet> EIGEN_DEVICE_FUNC inline Packet pnot(const Packet& a) { return pxor(ptrue(a), a);} -/** \internal \returns \a a shifted by N bits to the right */ +/** \internal \returns \a a logically shifted by N bits to the right */ template<int N> EIGEN_DEVICE_FUNC inline int -pshiftright(const int& a) { return a >> N; } +parithmetic_shift_right(const int& a) { return a >> N; } template<int N> EIGEN_DEVICE_FUNC inline long int -pshiftright(const long int& a) { return a >> N; } +parithmetic_shift_right(const long int& a) { return a >> N; } + +/** \internal \returns \a a arithmetically shifted by N bits to the right */ +template<int N> EIGEN_DEVICE_FUNC inline int +plogical_shift_right(const int& a) { return static_cast<int>(static_cast<unsigned int>(a) >> N); } +template<int N> EIGEN_DEVICE_FUNC inline long int +plogical_shift_right(const long int& a) { return static_cast<long>(static_cast<unsigned long>(a) >> N); } /** \internal \returns \a a shifted by N bits to the left */ template<int N> EIGEN_DEVICE_FUNC inline int -pshiftleft(const int& a) { return a << N; } +plogical_shift_left(const int& a) { return a << N; } template<int N> EIGEN_DEVICE_FUNC inline long int -pshiftleft(const long int& a) { return a << N; } +plogical_shift_left(const long int& a) { return a << N; } /** \internal \returns the significant and exponent of the underlying floating point numbers * See https://en.cppreference.com/w/cpp/numeric/math/frexp diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index 11c7bcb43..b7d37f92b 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -415,7 +415,17 @@ template<> EIGEN_STRONG_INLINE Packet8f pselect<Packet8f>(const Packet8f& mask, template<> EIGEN_STRONG_INLINE Packet4d pselect<Packet4d>(const Packet4d& mask, const Packet4d& a, const Packet4d& b) { return _mm256_blendv_pd(b,a,mask); } -template<int N> EIGEN_STRONG_INLINE Packet8i pshiftright(Packet8i a) { +template<int N> EIGEN_STRONG_INLINE Packet8i parithmetic_shift_right(Packet8i a) { +#ifdef EIGEN_VECTORIZE_AVX2 + return _mm256_srai_epi32(a, N); +#else + __m128i lo = _mm_srai_epi32(_mm256_extractf128_si256(a, 0), N); + __m128i hi = _mm_srai_epi32(_mm256_extractf128_si256(a, 1), N); + return _mm256_insertf128_si256(_mm256_castsi128_si256(lo), (hi), 1); +#endif +} + +template<int N> EIGEN_STRONG_INLINE Packet8i plogical_shift_right(Packet8i a) { #ifdef EIGEN_VECTORIZE_AVX2 return _mm256_srli_epi32(a, N); #else @@ -425,7 +435,7 @@ template<int N> EIGEN_STRONG_INLINE Packet8i pshiftright(Packet8i a) { #endif } -template<int N> EIGEN_STRONG_INLINE Packet8i pshiftleft(Packet8i a) { +template<int N> EIGEN_STRONG_INLINE Packet8i plogical_shift_left(Packet8i a) { #ifdef EIGEN_VECTORIZE_AVX2 return _mm256_slli_epi32(a, N); #else diff --git a/Eigen/src/Core/arch/AVX512/PacketMath.h b/Eigen/src/Core/arch/AVX512/PacketMath.h index a53f9bc6a..98a8a5b0f 100644 --- a/Eigen/src/Core/arch/AVX512/PacketMath.h +++ b/Eigen/src/Core/arch/AVX512/PacketMath.h @@ -538,7 +538,15 @@ EIGEN_STRONG_INLINE Packet8d pandnot<Packet8d>(const Packet8d& a,const Packet8d& #endif } -template<int N> EIGEN_STRONG_INLINE Packet16i pshiftleft(Packet16i a) { +template<int N> EIGEN_STRONG_INLINE Packet16i parithmetic_shift_right(Packet16i a) { + return _mm512_srai_epi32(a, N); +} + +template<int N> EIGEN_STRONG_INLINE Packet16i plogical_shift_right(Packet16i a) { + return _mm512_srli_epi32(a, N); +} + +template<int N> EIGEN_STRONG_INLINE Packet16i plogical_shift_left(Packet16i a) { return _mm512_slli_epi32(a, N); } diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index 154442cf9..95fa887e3 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -171,10 +171,11 @@ struct packet_traits<int> : default_packet_traits { size = 4, HasHalfPacket = 0, - HasAdd = 1, - HasSub = 1, - HasMul = 1, - HasDiv = 0, + HasAdd = 1, + HasSub = 1, + HasShift = 1, + HasMul = 1, + HasDiv = 0, HasBlend = 1 }; }; @@ -567,9 +568,11 @@ template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) { return vec_abs(a); } template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) { return vec_abs(a); } -template<int N> EIGEN_STRONG_INLINE Packet4i pshiftright(Packet4i a) +template<int N> EIGEN_STRONG_INLINE Packet4i parithmetic_shift_right(Packet4i a) +{ return vec_sra(a,reinterpret_cast<Packet4ui>(pset1<Packet4i>(N))); } +template<int N> EIGEN_STRONG_INLINE Packet4i plogical_shift_right(Packet4i a) { return vec_sr(a,reinterpret_cast<Packet4ui>(pset1<Packet4i>(N))); } -template<int N> EIGEN_STRONG_INLINE Packet4i pshiftleft(Packet4i a) +template<int N> EIGEN_STRONG_INLINE Packet4i plogical_shift_left(Packet4i a) { return vec_sl(a,reinterpret_cast<Packet4ui>(pset1<Packet4i>(N))); } template<> EIGEN_STRONG_INLINE Packet4f pfrexp<Packet4f>(const Packet4f& a, Packet4f& exponent) { diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 2b69512e3..4d9b3b44c 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -25,7 +25,7 @@ pfrexp_float(const Packet& a, Packet& exponent) { const Packet cst_126f = pset1<Packet>(126.0f); const Packet cst_half = pset1<Packet>(0.5f); const Packet cst_inv_mant_mask = pset1frombits<Packet>(~0x7f800000u); - exponent = psub(pcast<PacketI,Packet>(pshiftright<23>(preinterpret<PacketI>(a))), cst_126f); + exponent = psub(pcast<PacketI,Packet>(plogical_shift_right<23>(preinterpret<PacketI>(a))), cst_126f); return por(pand(a, cst_inv_mant_mask), cst_half); } @@ -36,7 +36,7 @@ pldexp_float(Packet a, Packet exponent) const Packet cst_127 = pset1<Packet>(127.f); // return a * 2^exponent PacketI ei = pcast<Packet,PacketI>(padd(exponent, cst_127)); - return pmul(a, preinterpret<Packet>(pshiftleft<23>(ei))); + return pmul(a, preinterpret<Packet>(plogical_shift_left<23>(ei))); } // Natural logarithm @@ -466,8 +466,8 @@ Packet psincos_float(const Packet& _x) // Compute the sign to apply to the polynomial. // sin: sign = second_bit(y_int) xor signbit(_x) // cos: sign = second_bit(y_int+1) - Packet sign_bit = ComputeSine ? pxor(_x, preinterpret<Packet>(pshiftleft<30>(y_int))) - : preinterpret<Packet>(pshiftleft<30>(padd(y_int,csti_1))); + Packet sign_bit = ComputeSine ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int,csti_1))); sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit // Get the polynomial selection mask from the second bit of y_int diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index 93a6d5885..4b162d493 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -139,6 +139,7 @@ struct packet_traits<float> : default_packet_traits HasAdd = 1, HasSub = 1, + HasShift = 1, HasMul = 1, HasNegate = 1, HasAbs = 1, @@ -178,6 +179,7 @@ struct packet_traits<int8_t> : default_packet_traits HasAdd = 1, HasSub = 1, + HasShift = 1, HasMul = 1, HasNegate = 1, HasAbs = 1, @@ -206,6 +208,7 @@ struct packet_traits<uint8_t> : default_packet_traits HasAdd = 1, HasSub = 1, + HasShift = 1, HasMul = 1, HasNegate = 0, HasAbs = 1, @@ -236,6 +239,7 @@ struct packet_traits<int16_t> : default_packet_traits HasAdd = 1, HasSub = 1, + HasShift = 1, HasMul = 1, HasNegate = 1, HasAbs = 1, @@ -264,6 +268,7 @@ struct packet_traits<uint16_t> : default_packet_traits HasAdd = 1, HasSub = 1, + HasShift = 1, HasMul = 1, HasNegate = 0, HasAbs = 0, @@ -294,6 +299,7 @@ struct packet_traits<int32_t> : default_packet_traits HasAdd = 1, HasSub = 1, + HasShift = 1, HasMul = 1, HasNegate = 1, HasAbs = 1, @@ -322,6 +328,7 @@ struct packet_traits<uint32_t> : default_packet_traits HasAdd = 1, HasSub = 1, + HasShift = 1, HasMul = 1, HasNegate = 0, HasAbs = 0, @@ -1454,43 +1461,69 @@ template<> EIGEN_STRONG_INLINE Packet2l pandnot<Packet2l>(const Packet2l& a, con template<> EIGEN_STRONG_INLINE Packet2ul pandnot<Packet2ul>(const Packet2ul& a, const Packet2ul& b) { return vbicq_u64(a,b); } -template<int N> EIGEN_STRONG_INLINE Packet4c pshiftright(Packet4c& a) +template<int N> EIGEN_STRONG_INLINE Packet4c parithmetic_shift_right(Packet4c& a) { return vget_lane_s32(vreinterpret_s32_s8(vshr_n_s8(vreinterpret_s8_s32(vdup_n_s32(a)), N)), 0); } -template<int N> EIGEN_STRONG_INLINE Packet8c pshiftright(Packet8c a) { return vshr_n_s8(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet16c pshiftright(Packet16c a) { return vshrq_n_s8(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet4uc pshiftright(Packet4uc& a) +template<int N> EIGEN_STRONG_INLINE Packet8c parithmetic_shift_right(Packet8c a) { return vshr_n_s8(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet16c parithmetic_shift_right(Packet16c a) { return vshrq_n_s8(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet4uc parithmetic_shift_right(Packet4uc& a) { return vget_lane_u32(vreinterpret_u32_u8(vshr_n_u8(vreinterpret_u8_u32(vdup_n_u32(a)), N)), 0); } -template<int N> EIGEN_STRONG_INLINE Packet8uc pshiftright(Packet8uc a) { return vshr_n_u8(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet16uc pshiftright(Packet16uc a) { return vshrq_n_u8(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet4s pshiftright(Packet4s a) { return vshr_n_s16(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet8s pshiftright(Packet8s a) { return vshrq_n_s16(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet4us pshiftright(Packet4us a) { return vshr_n_u16(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet8us pshiftright(Packet8us a) { return vshrq_n_u16(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet2i pshiftright(Packet2i a) { return vshr_n_s32(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet4i pshiftright(Packet4i a) { return vshrq_n_s32(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet2ui pshiftright(Packet2ui a) { return vshr_n_u32(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet4ui pshiftright(Packet4ui a) { return vshrq_n_u32(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet2l pshiftright(Packet2l a) { return vshrq_n_s64(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet2ul pshiftright(Packet2ul a) { return vshrq_n_u64(a,N); } - -template<int N> EIGEN_STRONG_INLINE Packet4c pshiftleft(Packet4c& a) +template<int N> EIGEN_STRONG_INLINE Packet8uc parithmetic_shift_right(Packet8uc a) { return vshr_n_u8(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet16uc parithmetic_shift_right(Packet16uc a) { return vshrq_n_u8(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet4s parithmetic_shift_right(Packet4s a) { return vshr_n_s16(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet8s parithmetic_shift_right(Packet8s a) { return vshrq_n_s16(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet4us parithmetic_shift_right(Packet4us a) { return vshr_n_u16(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet8us parithmetic_shift_right(Packet8us a) { return vshrq_n_u16(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet2i parithmetic_shift_right(Packet2i a) { return vshr_n_s32(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet4i parithmetic_shift_right(Packet4i a) { return vshrq_n_s32(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet2ui parithmetic_shift_right(Packet2ui a) { return vshr_n_u32(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet4ui parithmetic_shift_right(Packet4ui a) { return vshrq_n_u32(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet2l parithmetic_shift_right(Packet2l a) { return vshrq_n_s64(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet2ul parithmetic_shift_right(Packet2ul a) { return vshrq_n_u64(a,N); } + +template<int N> EIGEN_STRONG_INLINE Packet4c plogical_shift_right(Packet4c& a) +{ return vget_lane_s32(vreinterpret_s32_u8(vshr_n_u8(vreinterpret_u8_s32(vdup_n_s32(a)), N)), 0); } +template<int N> EIGEN_STRONG_INLINE Packet8c plogical_shift_right(Packet8c a) +{ return vreinterpret_s8_u8(vshr_n_u8(vreinterpret_u8_s8(a),N)); } +template<int N> EIGEN_STRONG_INLINE Packet16c plogical_shift_right(Packet16c a) +{ return vreinterpretq_s8_u8(vshrq_n_u8(vreinterpretq_u8_s8(a),N)); } +template<int N> EIGEN_STRONG_INLINE Packet4uc plogical_shift_right(Packet4uc& a) +{ return vget_lane_u32(vreinterpret_u32_s8(vshr_n_s8(vreinterpret_s8_u32(vdup_n_u32(a)), N)), 0); } +template<int N> EIGEN_STRONG_INLINE Packet8uc plogical_shift_right(Packet8uc a) { return vshr_n_u8(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet16uc plogical_shift_right(Packet16uc a) { return vshrq_n_u8(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet4s plogical_shift_right(Packet4s a) +{ return vreinterpret_s16_u16(vshr_n_u16(vreinterpret_u16_s16(a),N)); } +template<int N> EIGEN_STRONG_INLINE Packet8s plogical_shift_right(Packet8s a) +{ return vreinterpretq_s16_u16(vshrq_n_u16(vreinterpretq_u16_s16(a),N)); } +template<int N> EIGEN_STRONG_INLINE Packet4us plogical_shift_right(Packet4us a) { return vshr_n_u16(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet8us plogical_shift_right(Packet8us a) { return vshrq_n_u16(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet2i plogical_shift_right(Packet2i a) +{ return vreinterpret_s32_u32(vshr_n_u32(vreinterpret_u32_s32(a),N)); } +template<int N> EIGEN_STRONG_INLINE Packet4i plogical_shift_right(Packet4i a) +{ return vreinterpretq_s32_u32(vshrq_n_u32(vreinterpretq_u32_s32(a),N)); } +template<int N> EIGEN_STRONG_INLINE Packet2ui plogical_shift_right(Packet2ui a) { return vshr_n_u32(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet4ui plogical_shift_right(Packet4ui a) { return vshrq_n_u32(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet2l plogical_shift_right(Packet2l a) +{ return vreinterpretq_s64_u64(vshrq_n_u64(vreinterpretq_u64_s64(a),N)); } +template<int N> EIGEN_STRONG_INLINE Packet2ul plogical_shift_right(Packet2ul a) { return vshrq_n_u64(a,N); } + +template<int N> EIGEN_STRONG_INLINE Packet4c plogical_shift_left(Packet4c& a) { return vget_lane_s32(vreinterpret_s32_s8(vshl_n_s8(vreinterpret_s8_s32(vdup_n_s32(a)), N)), 0); } -template<int N> EIGEN_STRONG_INLINE Packet8c pshiftleft(Packet8c a) { return vshl_n_s8(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet16c pshiftleft(Packet16c a) { return vshlq_n_s8(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet4uc pshiftleft(Packet4uc& a) +template<int N> EIGEN_STRONG_INLINE Packet8c plogical_shift_left(Packet8c a) { return vshl_n_s8(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet16c plogical_shift_left(Packet16c a) { return vshlq_n_s8(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet4uc plogical_shift_left(Packet4uc& a) { return vget_lane_u32(vreinterpret_u32_u8(vshl_n_u8(vreinterpret_u8_u32(vdup_n_u32(a)), N)), 0); } -template<int N> EIGEN_STRONG_INLINE Packet8uc pshiftleft(Packet8uc a) { return vshl_n_u8(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet16uc pshiftleft(Packet16uc a) { return vshlq_n_u8(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet4s pshiftleft(Packet4s a) { return vshl_n_s16(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet8s pshiftleft(Packet8s a) { return vshlq_n_s16(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet4us pshiftleft(Packet4us a) { return vshl_n_u16(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet8us pshiftleft(Packet8us a) { return vshlq_n_u16(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet2i pshiftleft(Packet2i a) { return vshl_n_s32(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet4i pshiftleft(Packet4i a) { return vshlq_n_s32(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet2ui pshiftleft(Packet2ui a) { return vshl_n_u32(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet4ui pshiftleft(Packet4ui a) { return vshlq_n_u32(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet2l pshiftleft(Packet2l a) { return vshlq_n_s64(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet2ul pshiftleft(Packet2ul a) { return vshlq_n_u64(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet8uc plogical_shift_left(Packet8uc a) { return vshl_n_u8(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet16uc plogical_shift_left(Packet16uc a) { return vshlq_n_u8(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet4s plogical_shift_left(Packet4s a) { return vshl_n_s16(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet8s plogical_shift_left(Packet8s a) { return vshlq_n_s16(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet4us plogical_shift_left(Packet4us a) { return vshl_n_u16(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet8us plogical_shift_left(Packet8us a) { return vshlq_n_u16(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet2i plogical_shift_left(Packet2i a) { return vshl_n_s32(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet4i plogical_shift_left(Packet4i a) { return vshlq_n_s32(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet2ui plogical_shift_left(Packet2ui a) { return vshl_n_u32(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet4ui plogical_shift_left(Packet4ui a) { return vshlq_n_u32(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet2l plogical_shift_left(Packet2l a) { return vshlq_n_s64(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet2ul plogical_shift_left(Packet2ul a) { return vshlq_n_u64(a,N); } template<> EIGEN_STRONG_INLINE Packet2f pload<Packet2f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1_f32(from); } diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index d6a4a5c7f..f624f2c9f 100755 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -165,6 +165,7 @@ template<> struct packet_traits<int> : default_packet_traits AlignedOnScalar = 1, size=4, + HasShift = 1, HasBlend = 1 }; }; @@ -426,8 +427,9 @@ template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, con template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_andnot_pd(b,a); } template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_andnot_si128(b,a); } -template<int N> EIGEN_STRONG_INLINE Packet4i pshiftright(Packet4i a) { return _mm_srli_epi32(a,N); } -template<int N> EIGEN_STRONG_INLINE Packet4i pshiftleft(Packet4i a) { return _mm_slli_epi32(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet4i parithmetic_shift_right(Packet4i a) { return _mm_srai_epi32(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet4i plogical_shift_right(Packet4i a) { return _mm_srli_epi32(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet4i plogical_shift_left(Packet4i a) { return _mm_slli_epi32(a,N); } #ifdef EIGEN_VECTORIZE_SSE4_1 template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a) diff --git a/Eigen/src/Core/functors/UnaryFunctors.h b/Eigen/src/Core/functors/UnaryFunctors.h index a07ddaa30..e34eac296 100644 --- a/Eigen/src/Core/functors/UnaryFunctors.h +++ b/Eigen/src/Core/functors/UnaryFunctors.h @@ -167,6 +167,44 @@ struct functor_traits<scalar_cast_op<Scalar,NewType> > { enum { Cost = is_same<Scalar, NewType>::value ? 0 : NumTraits<NewType>::AddCost, PacketAccess = false }; }; /** \internal + * \brief Template functor to arithmetically shift a scalar right by a number of bits + * + * \sa class CwiseUnaryOp, MatrixBase::shift_right() + */ +template<typename Scalar, int N> +struct scalar_shift_right_op : unary_op_base<Scalar> { + EIGEN_EMPTY_STRUCT_CTOR(scalar_shift_right_op) + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const + { return a >> N; } + template<typename Packet> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const + { return internal::parithmetic_shift_right<N>(a); } +}; +template<typename Scalar, int N> +struct functor_traits<scalar_shift_right_op<Scalar,N> > +{ enum { Cost = NumTraits<Scalar>::AddCost, PacketAccess = packet_traits<Scalar>::HasShift }; }; + +/** \internal + * \brief Template functor to logically shift a scalar left by a number of bits + * + * \sa class CwiseUnaryOp, MatrixBase::shift_left() + */ +template<typename Scalar, int N> +struct scalar_shift_left_op : unary_op_base<Scalar> { + EIGEN_EMPTY_STRUCT_CTOR(scalar_shift_left_op) + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const + { return a << N; } + template<typename Packet> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const + { return internal::plogical_shift_left<N>(a); } +}; +template<typename Scalar, int N> +struct functor_traits<scalar_shift_left_op<Scalar,N> > +{ enum { Cost = NumTraits<Scalar>::AddCost, PacketAccess = packet_traits<Scalar>::HasShift }; }; + +/** \internal * \brief Template functor to extract the real part of a complex * * \sa class CwiseUnaryOp, MatrixBase::real() diff --git a/Eigen/src/plugins/CommonCwiseUnaryOps.h b/Eigen/src/plugins/CommonCwiseUnaryOps.h index 5418dc415..42ff901ca 100644 --- a/Eigen/src/plugins/CommonCwiseUnaryOps.h +++ b/Eigen/src/plugins/CommonCwiseUnaryOps.h @@ -64,6 +64,49 @@ cast() const return typename CastXpr<NewType>::Type(derived()); } +template<int N> struct ShiftRightXpr { + typedef CwiseUnaryOp<internal::scalar_shift_right_op<Scalar, N>, const Derived> Type; +}; + +/// \returns an expression of \c *this with the \a Scalar type arithmetically +/// shifted right by \a N bit positions. +/// +/// The template parameter \a N specifies the number of bit positions to shift. +/// +EIGEN_DOC_UNARY_ADDONS(cast,conversion function) +/// +/// \sa class CwiseUnaryOp +/// +template<int N> +EIGEN_DEVICE_FUNC +typename ShiftRightXpr<N>::Type +shift_right() const +{ + return typename ShiftRightXpr<N>::Type(derived()); +} + + +template<int N> struct ShiftLeftXpr { + typedef CwiseUnaryOp<internal::scalar_shift_left_op<Scalar, N>, const Derived> Type; +}; + +/// \returns an expression of \c *this with the \a Scalar type logically +/// shifted left by \a N bit positions. +/// +/// The template parameter \a N specifies the number of bit positions to shift. +/// +EIGEN_DOC_UNARY_ADDONS(cast,conversion function) +/// +/// \sa class CwiseUnaryOp +/// +template<int N> +EIGEN_DEVICE_FUNC +typename ShiftLeftXpr<N>::Type +shift_left() const +{ + return typename ShiftLeftXpr<N>::Type(derived()); +} + /// \returns an expression of the complex conjugate of \c *this. /// EIGEN_DOC_UNARY_ADDONS(conjugate,complex conjugate) |