From fd1dcb6b45a2c797ad4c4d6cc7678ee70763b4ed Mon Sep 17 00:00:00 2001
From: Antonio Sanchez <cantonios@google.com>
Date: Fri, 20 Nov 2020 08:49:57 -0800
Subject: Fixes duplicate symbol when building blas

Missing inline breaks blas, since symbol generated in
`complex_single.cpp`, `complex_double.cpp`, `single.cpp`, `double.cpp`

Changed rest of inlines to `EIGEN_STRONG_INLINE`.
---
 Eigen/src/Core/arch/NEON/PacketMath.h | 216 +++++++++++++++++-----------------
 1 file changed, 108 insertions(+), 108 deletions(-)

(limited to 'Eigen/src/Core/arch/NEON/PacketMath.h')

diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h
index b13cbe5ec..709cebe4e 100644
--- a/Eigen/src/Core/arch/NEON/PacketMath.h
+++ b/Eigen/src/Core/arch/NEON/PacketMath.h
@@ -87,8 +87,8 @@ typedef uint64x2_t                           Packet2ul;
 // fuctionally equivalent to _mm_shuffle_ps in SSE when interleave
 // == false (i.e. shuffle<false>(m, n, mask) equals _mm_shuffle_ps(m, n, mask)),
 // interleave m and n when interleave == true. Currently used in LU/arch/InverseSize4.h
-// to enable a shared implementation for fast inversion of matrices of size 4. 
-template<bool interleave> 
+// to enable a shared implementation for fast inversion of matrices of size 4.
+template<bool interleave>
 EIGEN_STRONG_INLINE Packet4f shuffle(const Packet4f &m, const Packet4f &n, int mask)
 {
   const float* a = reinterpret_cast<const float*>(&m);
@@ -97,8 +97,8 @@ EIGEN_STRONG_INLINE Packet4f shuffle(const Packet4f &m, const Packet4f &n, int m
   return res;
 }
 
-template<> 
-EIGEN_STRONG_INLINE Packet4f shuffle<true>(const Packet4f &m, const Packet4f &n, int mask) 
+template<>
+EIGEN_STRONG_INLINE Packet4f shuffle<true>(const Packet4f &m, const Packet4f &n, int mask)
 {
   const float* a = reinterpret_cast<const float*>(&m);
   const float* b = reinterpret_cast<const float*>(&n);
@@ -109,7 +109,7 @@ EIGEN_STRONG_INLINE Packet4f shuffle<true>(const Packet4f &m, const Packet4f &n,
 EIGEN_STRONG_INLINE static int eigen_neon_shuffle_mask(int p, int q, int r, int s) {return ((s)<<6|(r)<<4|(q)<<2|(p));}
 
 EIGEN_STRONG_INLINE Packet4f vec4f_swizzle2(const Packet4f& a, const Packet4f& b, int p, int q, int r, int s)
-{ 
+{
   return shuffle<false>(a,b,eigen_neon_shuffle_mask(p, q, r, s));
 }
 EIGEN_STRONG_INLINE Packet4f vec4f_movelh(const Packet4f& a, const Packet4f& b)
@@ -1922,13 +1922,13 @@ template<> EIGEN_STRONG_INLINE void pstoreu<int64_t>(int64_t* to, const Packet2l
 template<> EIGEN_STRONG_INLINE void pstoreu<uint64_t>(uint64_t* to, const Packet2ul& from)
 { EIGEN_DEBUG_UNALIGNED_STORE vst1q_u64(to,from); }
 
-template<> EIGEN_DEVICE_FUNC inline Packet2f pgather<float, Packet2f>(const float* from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet2f pgather<float, Packet2f>(const float* from, Index stride)
 {
   Packet2f res = vld1_dup_f32(from);
   res = vld1_lane_f32(from + 1*stride, res, 1);
   return res;
 }
-template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4f pgather<float, Packet4f>(const float* from, Index stride)
 {
   Packet4f res = vld1q_dup_f32(from);
   res = vld1q_lane_f32(from + 1*stride, res, 1);
@@ -1936,14 +1936,14 @@ template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const floa
   res = vld1q_lane_f32(from + 3*stride, res, 3);
   return res;
 }
-template<> EIGEN_DEVICE_FUNC inline Packet4c pgather<int8_t, Packet4c>(const int8_t* from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4c pgather<int8_t, Packet4c>(const int8_t* from, Index stride)
 {
   Packet4c res;
   for (int i = 0; i != 4; i++)
     reinterpret_cast<int8_t*>(&res)[i] = *(from + i * stride);
   return res;
 }
-template<> EIGEN_DEVICE_FUNC inline Packet8c pgather<int8_t, Packet8c>(const int8_t* from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet8c pgather<int8_t, Packet8c>(const int8_t* from, Index stride)
 {
   Packet8c res = vld1_dup_s8(from);
   res = vld1_lane_s8(from + 1*stride, res, 1);
@@ -1955,7 +1955,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet8c pgather<int8_t, Packet8c>(const int
   res = vld1_lane_s8(from + 7*stride, res, 7);
   return res;
 }
-template<> EIGEN_DEVICE_FUNC inline Packet16c pgather<int8_t, Packet16c>(const int8_t* from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet16c pgather<int8_t, Packet16c>(const int8_t* from, Index stride)
 {
   Packet16c res = vld1q_dup_s8(from);
   res = vld1q_lane_s8(from + 1*stride, res, 1);
@@ -1975,14 +1975,14 @@ template<> EIGEN_DEVICE_FUNC inline Packet16c pgather<int8_t, Packet16c>(const i
   res = vld1q_lane_s8(from + 15*stride, res, 15);
   return res;
 }
-template<> EIGEN_DEVICE_FUNC inline Packet4uc pgather<uint8_t, Packet4uc>(const uint8_t* from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4uc pgather<uint8_t, Packet4uc>(const uint8_t* from, Index stride)
 {
   Packet4uc res;
   for (int i = 0; i != 4; i++)
     reinterpret_cast<uint8_t*>(&res)[i] = *(from + i * stride);
   return res;
 }
-template<> EIGEN_DEVICE_FUNC inline Packet8uc pgather<uint8_t, Packet8uc>(const uint8_t* from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet8uc pgather<uint8_t, Packet8uc>(const uint8_t* from, Index stride)
 {
   Packet8uc res = vld1_dup_u8(from);
   res = vld1_lane_u8(from + 1*stride, res, 1);
@@ -1994,7 +1994,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet8uc pgather<uint8_t, Packet8uc>(const
   res = vld1_lane_u8(from + 7*stride, res, 7);
   return res;
 }
-template<> EIGEN_DEVICE_FUNC inline Packet16uc pgather<uint8_t, Packet16uc>(const uint8_t* from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet16uc pgather<uint8_t, Packet16uc>(const uint8_t* from, Index stride)
 {
   Packet16uc res = vld1q_dup_u8(from);
   res = vld1q_lane_u8(from + 1*stride, res, 1);
@@ -2014,7 +2014,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet16uc pgather<uint8_t, Packet16uc>(cons
   res = vld1q_lane_u8(from + 15*stride, res, 15);
   return res;
 }
-template<> EIGEN_DEVICE_FUNC inline Packet4s pgather<int16_t, Packet4s>(const int16_t* from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4s pgather<int16_t, Packet4s>(const int16_t* from, Index stride)
 {
   Packet4s res = vld1_dup_s16(from);
   res = vld1_lane_s16(from + 1*stride, res, 1);
@@ -2022,7 +2022,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet4s pgather<int16_t, Packet4s>(const in
   res = vld1_lane_s16(from + 3*stride, res, 3);
   return res;
 }
-template<> EIGEN_DEVICE_FUNC inline Packet8s pgather<int16_t, Packet8s>(const int16_t* from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet8s pgather<int16_t, Packet8s>(const int16_t* from, Index stride)
 {
   Packet8s res = vld1q_dup_s16(from);
   res = vld1q_lane_s16(from + 1*stride, res, 1);
@@ -2034,7 +2034,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet8s pgather<int16_t, Packet8s>(const in
   res = vld1q_lane_s16(from + 7*stride, res, 7);
   return res;
 }
-template<> EIGEN_DEVICE_FUNC inline Packet4us pgather<uint16_t, Packet4us>(const uint16_t* from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4us pgather<uint16_t, Packet4us>(const uint16_t* from, Index stride)
 {
   Packet4us res = vld1_dup_u16(from);
   res = vld1_lane_u16(from + 1*stride, res, 1);
@@ -2042,7 +2042,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet4us pgather<uint16_t, Packet4us>(const
   res = vld1_lane_u16(from + 3*stride, res, 3);
   return res;
 }
-template<> EIGEN_DEVICE_FUNC inline Packet8us pgather<uint16_t, Packet8us>(const uint16_t* from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet8us pgather<uint16_t, Packet8us>(const uint16_t* from, Index stride)
 {
   Packet8us res = vld1q_dup_u16(from);
   res = vld1q_lane_u16(from + 1*stride, res, 1);
@@ -2054,13 +2054,13 @@ template<> EIGEN_DEVICE_FUNC inline Packet8us pgather<uint16_t, Packet8us>(const
   res = vld1q_lane_u16(from + 7*stride, res, 7);
   return res;
 }
-template<> EIGEN_DEVICE_FUNC inline Packet2i pgather<int32_t, Packet2i>(const int32_t* from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet2i pgather<int32_t, Packet2i>(const int32_t* from, Index stride)
 {
   Packet2i res = vld1_dup_s32(from);
   res = vld1_lane_s32(from + 1*stride, res, 1);
   return res;
 }
-template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int32_t, Packet4i>(const int32_t* from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4i pgather<int32_t, Packet4i>(const int32_t* from, Index stride)
 {
   Packet4i res = vld1q_dup_s32(from);
   res = vld1q_lane_s32(from + 1*stride, res, 1);
@@ -2068,13 +2068,13 @@ template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int32_t, Packet4i>(const in
   res = vld1q_lane_s32(from + 3*stride, res, 3);
   return res;
 }
-template<> EIGEN_DEVICE_FUNC inline Packet2ui pgather<uint32_t, Packet2ui>(const uint32_t* from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet2ui pgather<uint32_t, Packet2ui>(const uint32_t* from, Index stride)
 {
   Packet2ui res = vld1_dup_u32(from);
   res = vld1_lane_u32(from + 1*stride, res, 1);
   return res;
 }
-template<> EIGEN_DEVICE_FUNC inline Packet4ui pgather<uint32_t, Packet4ui>(const uint32_t* from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4ui pgather<uint32_t, Packet4ui>(const uint32_t* from, Index stride)
 {
   Packet4ui res = vld1q_dup_u32(from);
   res = vld1q_lane_u32(from + 1*stride, res, 1);
@@ -2082,37 +2082,37 @@ template<> EIGEN_DEVICE_FUNC inline Packet4ui pgather<uint32_t, Packet4ui>(const
   res = vld1q_lane_u32(from + 3*stride, res, 3);
   return res;
 }
-template<> EIGEN_DEVICE_FUNC inline Packet2l pgather<int64_t, Packet2l>(const int64_t* from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet2l pgather<int64_t, Packet2l>(const int64_t* from, Index stride)
 {
   Packet2l res = vld1q_dup_s64(from);
   res = vld1q_lane_s64(from + 1*stride, res, 1);
   return res;
 }
-template<> EIGEN_DEVICE_FUNC inline Packet2ul pgather<uint64_t, Packet2ul>(const uint64_t* from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet2ul pgather<uint64_t, Packet2ul>(const uint64_t* from, Index stride)
 {
   Packet2ul res = vld1q_dup_u64(from);
   res = vld1q_lane_u64(from + 1*stride, res, 1);
   return res;
 }
 
-template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet2f>(float* to, const Packet2f& from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pscatter<float, Packet2f>(float* to, const Packet2f& from, Index stride)
 {
   vst1_lane_f32(to + stride*0, from, 0);
   vst1_lane_f32(to + stride*1, from, 1);
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride)
 {
   vst1q_lane_f32(to + stride*0, from, 0);
   vst1q_lane_f32(to + stride*1, from, 1);
   vst1q_lane_f32(to + stride*2, from, 2);
   vst1q_lane_f32(to + stride*3, from, 3);
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<int8_t, Packet4c>(int8_t* to, const Packet4c& from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pscatter<int8_t, Packet4c>(int8_t* to, const Packet4c& from, Index stride)
 {
   for (int i = 0; i != 4; i++)
     *(to + i * stride) = reinterpret_cast<const int8_t*>(&from)[i];
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<int8_t, Packet8c>(int8_t* to, const Packet8c& from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pscatter<int8_t, Packet8c>(int8_t* to, const Packet8c& from, Index stride)
 {
   vst1_lane_s8(to + stride*0, from, 0);
   vst1_lane_s8(to + stride*1, from, 1);
@@ -2123,7 +2123,7 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<int8_t, Packet8c>(int8_t* to,
   vst1_lane_s8(to + stride*6, from, 6);
   vst1_lane_s8(to + stride*7, from, 7);
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<int8_t, Packet16c>(int8_t* to, const Packet16c& from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pscatter<int8_t, Packet16c>(int8_t* to, const Packet16c& from, Index stride)
 {
   vst1q_lane_s8(to + stride*0, from, 0);
   vst1q_lane_s8(to + stride*1, from, 1);
@@ -2142,12 +2142,12 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<int8_t, Packet16c>(int8_t* to,
   vst1q_lane_s8(to + stride*14, from, 14);
   vst1q_lane_s8(to + stride*15, from, 15);
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<uint8_t, Packet4uc>(uint8_t* to, const Packet4uc& from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pscatter<uint8_t, Packet4uc>(uint8_t* to, const Packet4uc& from, Index stride)
 {
   for (int i = 0; i != 4; i++)
     *(to + i * stride) = reinterpret_cast<const uint8_t*>(&from)[i];
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<uint8_t, Packet8uc>(uint8_t* to, const Packet8uc& from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pscatter<uint8_t, Packet8uc>(uint8_t* to, const Packet8uc& from, Index stride)
 {
   vst1_lane_u8(to + stride*0, from, 0);
   vst1_lane_u8(to + stride*1, from, 1);
@@ -2158,7 +2158,7 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<uint8_t, Packet8uc>(uint8_t* t
   vst1_lane_u8(to + stride*6, from, 6);
   vst1_lane_u8(to + stride*7, from, 7);
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<uint8_t, Packet16uc>(uint8_t* to, const Packet16uc& from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pscatter<uint8_t, Packet16uc>(uint8_t* to, const Packet16uc& from, Index stride)
 {
   vst1q_lane_u8(to + stride*0, from, 0);
   vst1q_lane_u8(to + stride*1, from, 1);
@@ -2177,14 +2177,14 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<uint8_t, Packet16uc>(uint8_t*
   vst1q_lane_u8(to + stride*14, from, 14);
   vst1q_lane_u8(to + stride*15, from, 15);
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<int16_t, Packet4s>(int16_t* to, const Packet4s& from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pscatter<int16_t, Packet4s>(int16_t* to, const Packet4s& from, Index stride)
 {
   vst1_lane_s16(to + stride*0, from, 0);
   vst1_lane_s16(to + stride*1, from, 1);
   vst1_lane_s16(to + stride*2, from, 2);
   vst1_lane_s16(to + stride*3, from, 3);
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<int16_t, Packet8s>(int16_t* to, const Packet8s& from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pscatter<int16_t, Packet8s>(int16_t* to, const Packet8s& from, Index stride)
 {
   vst1q_lane_s16(to + stride*0, from, 0);
   vst1q_lane_s16(to + stride*1, from, 1);
@@ -2195,14 +2195,14 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<int16_t, Packet8s>(int16_t* to
   vst1q_lane_s16(to + stride*6, from, 6);
   vst1q_lane_s16(to + stride*7, from, 7);
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<uint16_t, Packet4us>(uint16_t* to, const Packet4us& from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pscatter<uint16_t, Packet4us>(uint16_t* to, const Packet4us& from, Index stride)
 {
   vst1_lane_u16(to + stride*0, from, 0);
   vst1_lane_u16(to + stride*1, from, 1);
   vst1_lane_u16(to + stride*2, from, 2);
   vst1_lane_u16(to + stride*3, from, 3);
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<uint16_t, Packet8us>(uint16_t* to, const Packet8us& from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pscatter<uint16_t, Packet8us>(uint16_t* to, const Packet8us& from, Index stride)
 {
   vst1q_lane_u16(to + stride*0, from, 0);
   vst1q_lane_u16(to + stride*1, from, 1);
@@ -2213,36 +2213,36 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<uint16_t, Packet8us>(uint16_t*
   vst1q_lane_u16(to + stride*6, from, 6);
   vst1q_lane_u16(to + stride*7, from, 7);
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<int32_t, Packet2i>(int32_t* to, const Packet2i& from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pscatter<int32_t, Packet2i>(int32_t* to, const Packet2i& from, Index stride)
 {
   vst1_lane_s32(to + stride*0, from, 0);
   vst1_lane_s32(to + stride*1, from, 1);
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<int32_t, Packet4i>(int32_t* to, const Packet4i& from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pscatter<int32_t, Packet4i>(int32_t* to, const Packet4i& from, Index stride)
 {
   vst1q_lane_s32(to + stride*0, from, 0);
   vst1q_lane_s32(to + stride*1, from, 1);
   vst1q_lane_s32(to + stride*2, from, 2);
   vst1q_lane_s32(to + stride*3, from, 3);
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<uint32_t, Packet2ui>(uint32_t* to, const Packet2ui& from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pscatter<uint32_t, Packet2ui>(uint32_t* to, const Packet2ui& from, Index stride)
 {
   vst1_lane_u32(to + stride*0, from, 0);
   vst1_lane_u32(to + stride*1, from, 1);
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<uint32_t, Packet4ui>(uint32_t* to, const Packet4ui& from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pscatter<uint32_t, Packet4ui>(uint32_t* to, const Packet4ui& from, Index stride)
 {
   vst1q_lane_u32(to + stride*0, from, 0);
   vst1q_lane_u32(to + stride*1, from, 1);
   vst1q_lane_u32(to + stride*2, from, 2);
   vst1q_lane_u32(to + stride*3, from, 3);
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<int64_t, Packet2l>(int64_t* to, const Packet2l& from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pscatter<int64_t, Packet2l>(int64_t* to, const Packet2l& from, Index stride)
 {
   vst1q_lane_s64(to + stride*0, from, 0);
   vst1q_lane_s64(to + stride*1, from, 1);
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<uint64_t, Packet2ul>(uint64_t* to, const Packet2ul& from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pscatter<uint64_t, Packet2ul>(uint64_t* to, const Packet2ul& from, Index stride)
 {
   vst1q_lane_u64(to + stride*0, from, 0);
   vst1q_lane_u64(to + stride*1, from, 1);
@@ -2457,23 +2457,23 @@ template<> EIGEN_STRONG_INLINE int64_t predux<Packet2l>(const Packet2l& a)
 template<> EIGEN_STRONG_INLINE uint64_t predux<Packet2ul>(const Packet2ul& a)
 { return vgetq_lane_u64(a, 0) + vgetq_lane_u64(a, 1); }
 
-template<> EIGEN_DEVICE_FUNC inline Packet4c predux_half_dowto4(const Packet8c& a)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4c predux_half_dowto4(const Packet8c& a)
 {
   return vget_lane_s32(vreinterpret_s32_s8(vadd_s8(a,
       vreinterpret_s8_s32(vrev64_s32(vreinterpret_s32_s8(a))))), 0);
 }
-template<> EIGEN_DEVICE_FUNC inline Packet8c predux_half_dowto4(const Packet16c& a)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet8c predux_half_dowto4(const Packet16c& a)
 { return vadd_s8(vget_high_s8(a), vget_low_s8(a)); }
-template<> EIGEN_DEVICE_FUNC inline Packet4uc predux_half_dowto4(const Packet8uc& a)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4uc predux_half_dowto4(const Packet8uc& a)
 {
   return vget_lane_u32(vreinterpret_u32_u8(vadd_u8(a,
       vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(a))))), 0);
 }
-template<> EIGEN_DEVICE_FUNC inline Packet8uc predux_half_dowto4(const Packet16uc& a)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet8uc predux_half_dowto4(const Packet16uc& a)
 { return vadd_u8(vget_high_u8(a), vget_low_u8(a)); }
-template<> EIGEN_DEVICE_FUNC inline Packet4s predux_half_dowto4(const Packet8s& a)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4s predux_half_dowto4(const Packet8s& a)
 { return vadd_s16(vget_high_s16(a), vget_low_s16(a)); }
-template<> EIGEN_DEVICE_FUNC inline Packet4us predux_half_dowto4(const Packet8us& a)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4us predux_half_dowto4(const Packet8us& a)
 { return vadd_u16(vget_high_u16(a), vget_low_u16(a)); }
 
 // Other reduction functions:
@@ -2752,13 +2752,13 @@ template<> EIGEN_STRONG_INLINE bool predux_any(const Packet4f& x)
   return vget_lane_u32(vpmax_u32(tmp, tmp), 0);
 }
 
-EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet2f, 2>& kernel)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2f, 2>& kernel)
 {
   const float32x2x2_t z = vzip_f32(kernel.packet[0], kernel.packet[1]);
   kernel.packet[0] = z.val[0];
   kernel.packet[1] = z.val[1];
 }
-EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4f, 4>& kernel)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet4f, 4>& kernel)
 {
   const float32x4x2_t tmp1 = vzipq_f32(kernel.packet[0], kernel.packet[1]);
   const float32x4x2_t tmp2 = vzipq_f32(kernel.packet[2], kernel.packet[3]);
@@ -2768,7 +2768,7 @@ EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4f, 4>& kernel)
   kernel.packet[2] = vcombine_f32(vget_low_f32(tmp1.val[1]), vget_low_f32(tmp2.val[1]));
   kernel.packet[3] = vcombine_f32(vget_high_f32(tmp1.val[1]), vget_high_f32(tmp2.val[1]));
 }
-EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4c, 4>& kernel)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet4c, 4>& kernel)
 {
   const int8x8_t a = vreinterpret_s8_s32(vset_lane_s32(kernel.packet[2], vdup_n_s32(kernel.packet[0]), 1));
   const int8x8_t b = vreinterpret_s8_s32(vset_lane_s32(kernel.packet[3], vdup_n_s32(kernel.packet[1]), 1));
@@ -2781,7 +2781,7 @@ EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4c, 4>& kernel)
   kernel.packet[2] = vget_lane_s32(vreinterpret_s32_s16(zip16.val[1]), 0);
   kernel.packet[3] = vget_lane_s32(vreinterpret_s32_s16(zip16.val[1]), 1);
 }
-EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8c, 8>& kernel)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet8c, 8>& kernel)
 {
   int8x8x2_t zip8[4];
   uint16x4x2_t zip16[4];
@@ -2811,7 +2811,7 @@ EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8c, 8>& kernel)
     }
   }
 }
-EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet16c, 16>& kernel)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet16c, 16>& kernel)
 {
   int8x16x2_t zip8[8];
   uint16x8x2_t zip16[8];
@@ -2858,7 +2858,7 @@ EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet16c, 16>& kernel)
     }
   }
 }
-EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4uc, 4>& kernel)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet4uc, 4>& kernel)
 {
   const uint8x8_t a = vreinterpret_u8_u32(vset_lane_u32(kernel.packet[2], vdup_n_u32(kernel.packet[0]), 1));
   const uint8x8_t b = vreinterpret_u8_u32(vset_lane_u32(kernel.packet[3], vdup_n_u32(kernel.packet[1]), 1));
@@ -2871,7 +2871,7 @@ EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4uc, 4>& kernel)
   kernel.packet[2] = vget_lane_u32(vreinterpret_u32_u16(zip16.val[1]), 0);
   kernel.packet[3] = vget_lane_u32(vreinterpret_u32_u16(zip16.val[1]), 1);
 }
-EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8uc, 8>& kernel)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet8uc, 8>& kernel)
 {
   uint8x8x2_t zip8[4];
   uint16x4x2_t zip16[4];
@@ -2901,7 +2901,7 @@ EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8uc, 8>& kernel)
     }
   }
 }
-EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet16uc, 16>& kernel)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet16uc, 16>& kernel)
 {
   uint8x16x2_t zip8[8];
   uint16x8x2_t zip16[8];
@@ -2946,7 +2946,7 @@ EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet16uc, 16>& kernel)
     }
   }
 }
-EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4s, 4>& kernel)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet4s, 4>& kernel)
 {
   const int16x4x2_t zip16_1 = vzip_s16(kernel.packet[0], kernel.packet[1]);
   const int16x4x2_t zip16_2 = vzip_s16(kernel.packet[2], kernel.packet[3]);
@@ -2960,7 +2960,7 @@ EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4s, 4>& kernel)
   kernel.packet[3] = vreinterpret_s16_u32(zip32_2.val[1]);
 }
 
-EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8s, 4>& kernel)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet8s, 4>& kernel)
 {
   const int16x8x2_t zip16_1 = vzipq_s16(kernel.packet[0], kernel.packet[1]);
   const int16x8x2_t zip16_2 = vzipq_s16(kernel.packet[2], kernel.packet[3]);
@@ -2974,7 +2974,7 @@ EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8s, 4>& kernel)
   kernel.packet[3] = vreinterpretq_s16_u32(zip32_2.val[1]);
 }
 
-EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet16uc, 4>& kernel)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet16uc, 4>& kernel)
 {
   const uint8x16x2_t zip8_1 = vzipq_u8(kernel.packet[0], kernel.packet[1]);
   const uint8x16x2_t zip8_2 = vzipq_u8(kernel.packet[2], kernel.packet[3]);
@@ -2988,7 +2988,7 @@ EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet16uc, 4>& kernel)
   kernel.packet[3] = vreinterpretq_u8_u16(zip16_2.val[1]);
 }
 
-EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8s, 8>& kernel)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet8s, 8>& kernel)
 {
   const int16x8x2_t zip16_1 = vzipq_s16(kernel.packet[0], kernel.packet[1]);
   const int16x8x2_t zip16_2 = vzipq_s16(kernel.packet[2], kernel.packet[3]);
@@ -3009,7 +3009,7 @@ EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8s, 8>& kernel)
   kernel.packet[6] = vreinterpretq_s16_u32(vcombine_u32(vget_low_u32(zip32_2.val[1]), vget_low_u32(zip32_4.val[1])));
   kernel.packet[7] = vreinterpretq_s16_u32(vcombine_u32(vget_high_u32(zip32_2.val[1]), vget_high_u32(zip32_4.val[1])));
 }
-EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4us, 4>& kernel)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet4us, 4>& kernel)
 {
   const uint16x4x2_t zip16_1 = vzip_u16(kernel.packet[0], kernel.packet[1]);
   const uint16x4x2_t zip16_2 = vzip_u16(kernel.packet[2], kernel.packet[3]);
@@ -3022,7 +3022,7 @@ EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4us, 4>& kernel)
   kernel.packet[2] = vreinterpret_u16_u32(zip32_2.val[0]);
   kernel.packet[3] = vreinterpret_u16_u32(zip32_2.val[1]);
 }
-EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8us, 8>& kernel)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet8us, 8>& kernel)
 {
   const uint16x8x2_t zip16_1 = vzipq_u16(kernel.packet[0], kernel.packet[1]);
   const uint16x8x2_t zip16_2 = vzipq_u16(kernel.packet[2], kernel.packet[3]);
@@ -3043,13 +3043,13 @@ EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8us, 8>& kernel)
   kernel.packet[6] = vreinterpretq_u16_u32(vcombine_u32(vget_low_u32(zip32_2.val[1]), vget_low_u32(zip32_4.val[1])));
   kernel.packet[7] = vreinterpretq_u16_u32(vcombine_u32(vget_high_u32(zip32_2.val[1]), vget_high_u32(zip32_4.val[1])));
 }
-EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet2i, 2>& kernel)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2i, 2>& kernel)
 {
   const int32x2x2_t z = vzip_s32(kernel.packet[0], kernel.packet[1]);
   kernel.packet[0] = z.val[0];
   kernel.packet[1] = z.val[1];
 }
-EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4i, 4>& kernel)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet4i, 4>& kernel)
 {
   const int32x4x2_t tmp1 = vzipq_s32(kernel.packet[0], kernel.packet[1]);
   const int32x4x2_t tmp2 = vzipq_s32(kernel.packet[2], kernel.packet[3]);
@@ -3059,13 +3059,13 @@ EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4i, 4>& kernel)
   kernel.packet[2] = vcombine_s32(vget_low_s32(tmp1.val[1]), vget_low_s32(tmp2.val[1]));
   kernel.packet[3] = vcombine_s32(vget_high_s32(tmp1.val[1]), vget_high_s32(tmp2.val[1]));
 }
-EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet2ui, 2>& kernel)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2ui, 2>& kernel)
 {
   const uint32x2x2_t z = vzip_u32(kernel.packet[0], kernel.packet[1]);
   kernel.packet[0] = z.val[0];
   kernel.packet[1] = z.val[1];
 }
-EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4ui, 4>& kernel)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet4ui, 4>& kernel)
 {
   const uint32x4x2_t tmp1 = vzipq_u32(kernel.packet[0], kernel.packet[1]);
   const uint32x4x2_t tmp2 = vzipq_u32(kernel.packet[2], kernel.packet[3]);
@@ -3075,7 +3075,7 @@ EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4ui, 4>& kernel)
   kernel.packet[2] = vcombine_u32(vget_low_u32(tmp1.val[1]), vget_low_u32(tmp2.val[1]));
   kernel.packet[3] = vcombine_u32(vget_high_u32(tmp1.val[1]), vget_high_u32(tmp2.val[1]));
 }
-EIGEN_DEVICE_FUNC inline void
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void
 ptranspose(PacketBlock<Packet2l, 2>& kernel)
 {
 #if EIGEN_ARCH_ARM64
@@ -3094,7 +3094,7 @@ ptranspose(PacketBlock<Packet2l, 2>& kernel)
   kernel.packet[1] = vcombine_s64(tmp[0][1], tmp[1][1]);
 #endif
 }
-EIGEN_DEVICE_FUNC inline void
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void
 ptranspose(PacketBlock<Packet2ul, 2>& kernel)
 {
 #if EIGEN_ARCH_ARM64
@@ -3114,37 +3114,37 @@ ptranspose(PacketBlock<Packet2ul, 2>& kernel)
 #endif
 }
 
-template<> EIGEN_DEVICE_FUNC inline Packet2f pselect( const Packet2f& mask, const Packet2f& a, const Packet2f& b)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet2f pselect( const Packet2f& mask, const Packet2f& a, const Packet2f& b)
 { return vbsl_f32(vreinterpret_u32_f32(mask), a, b); }
-template<> EIGEN_DEVICE_FUNC inline Packet4f pselect(const Packet4f& mask, const Packet4f& a, const Packet4f& b)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4f pselect(const Packet4f& mask, const Packet4f& a, const Packet4f& b)
 { return vbslq_f32(vreinterpretq_u32_f32(mask), a, b); }
-template<> EIGEN_DEVICE_FUNC inline Packet8c pselect(const Packet8c& mask, const Packet8c& a, const Packet8c& b)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet8c pselect(const Packet8c& mask, const Packet8c& a, const Packet8c& b)
 { return vbsl_s8(vreinterpret_u8_s8(mask), a, b); }
-template<> EIGEN_DEVICE_FUNC inline Packet16c pselect(const Packet16c& mask, const Packet16c& a, const Packet16c& b)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet16c pselect(const Packet16c& mask, const Packet16c& a, const Packet16c& b)
 { return vbslq_s8(vreinterpretq_u8_s8(mask), a, b); }
-template<> EIGEN_DEVICE_FUNC inline Packet8uc pselect(const Packet8uc& mask, const Packet8uc& a, const Packet8uc& b)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet8uc pselect(const Packet8uc& mask, const Packet8uc& a, const Packet8uc& b)
 { return vbsl_u8(mask, a, b); }
-template<> EIGEN_DEVICE_FUNC inline Packet16uc pselect(const Packet16uc& mask, const Packet16uc& a, const Packet16uc& b)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet16uc pselect(const Packet16uc& mask, const Packet16uc& a, const Packet16uc& b)
 { return vbslq_u8(mask, a, b); }
-template<> EIGEN_DEVICE_FUNC inline Packet4s pselect(const Packet4s& mask, const Packet4s& a, const Packet4s& b)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4s pselect(const Packet4s& mask, const Packet4s& a, const Packet4s& b)
 { return vbsl_s16(vreinterpret_u16_s16(mask), a, b); }
-template<> EIGEN_DEVICE_FUNC inline Packet8s pselect(const Packet8s& mask, const Packet8s& a, const Packet8s& b)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet8s pselect(const Packet8s& mask, const Packet8s& a, const Packet8s& b)
 { return vbslq_s16(vreinterpretq_u16_s16(mask), a, b); }
-template<> EIGEN_DEVICE_FUNC inline Packet4us pselect(const Packet4us& mask, const Packet4us& a, const Packet4us& b)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4us pselect(const Packet4us& mask, const Packet4us& a, const Packet4us& b)
 { return vbsl_u16(mask, a, b); }
-template<> EIGEN_DEVICE_FUNC inline Packet8us pselect(const Packet8us& mask, const Packet8us& a, const Packet8us& b)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet8us pselect(const Packet8us& mask, const Packet8us& a, const Packet8us& b)
 { return vbslq_u16(mask, a, b); }
-template<> EIGEN_DEVICE_FUNC inline Packet2i pselect(const Packet2i& mask, const Packet2i& a, const Packet2i& b)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet2i pselect(const Packet2i& mask, const Packet2i& a, const Packet2i& b)
 { return vbsl_s32(vreinterpret_u32_s32(mask), a, b); }
-template<> EIGEN_DEVICE_FUNC inline Packet4i pselect(const Packet4i& mask, const Packet4i& a, const Packet4i& b)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4i pselect(const Packet4i& mask, const Packet4i& a, const Packet4i& b)
 { return vbslq_s32(vreinterpretq_u32_s32(mask), a, b); }
-template<> EIGEN_DEVICE_FUNC inline Packet2ui pselect(const Packet2ui& mask, const Packet2ui& a, const Packet2ui& b)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet2ui pselect(const Packet2ui& mask, const Packet2ui& a, const Packet2ui& b)
 { return vbsl_u32(mask, a, b); }
-template<> EIGEN_DEVICE_FUNC inline Packet4ui pselect(const Packet4ui& mask, const Packet4ui& a, const Packet4ui& b)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4ui pselect(const Packet4ui& mask, const Packet4ui& a, const Packet4ui& b)
 { return vbslq_u32(mask, a, b); }
-template<> EIGEN_DEVICE_FUNC inline Packet2l pselect(const Packet2l& mask, const Packet2l& a, const Packet2l& b)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet2l pselect(const Packet2l& mask, const Packet2l& a, const Packet2l& b)
 { return vbslq_s64(vreinterpretq_u64_s64(mask), a, b); }
-template<> EIGEN_DEVICE_FUNC inline Packet2ul pselect(const Packet2ul& mask, const Packet2ul& a, const Packet2ul& b)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet2ul pselect(const Packet2ul& mask, const Packet2ul& a, const Packet2ul& b)
 { return vbslq_u64(mask, a, b); }
 
 /**
@@ -3254,7 +3254,7 @@ template<> EIGEN_STRONG_INLINE Packet4f psqrt(const Packet4f& _x){
                                       vcltq_f32(_x, pset1<Packet4f>((std::numeric_limits<float>::min)())));
   // Compute approximate reciprocal sqrt.
   Packet4f x = vrsqrteq_f32(_x);
-  // Do a single step of Newton's iteration. 
+  // Do a single step of Newton's iteration.
   //the number 1.5f was set reference to Quake3's fast inverse square root
   x = vmulq_f32(x, psub(pset1<Packet4f>(1.5f), pmul(half, pmul(x, x))));
   // Flush results for denormals to zero.
@@ -3273,7 +3273,7 @@ template<> EIGEN_STRONG_INLINE Packet2f psqrt(const Packet2f& _x){
   return vreinterpret_f32_u32(vbic_u32(vreinterpret_u32_f32(pmul(_x, x)), denormal_mask));
 }
 
-#else 
+#else
 template<> EIGEN_STRONG_INLINE Packet4f psqrt(const Packet4f& _x){return vsqrtq_f32(_x);}
 template<> EIGEN_STRONG_INLINE Packet2f psqrt(const Packet2f& _x){return vsqrt_f32(_x); }
 #endif
@@ -3441,7 +3441,7 @@ template<> EIGEN_STRONG_INLINE Packet4bf pandnot(const Packet4bf& a,const Packet
   return pandnot<Packet4us>(a, b);
 }
 
-template<> EIGEN_DEVICE_FUNC inline Packet4bf pselect(const Packet4bf& mask, const Packet4bf& a,
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4bf pselect(const Packet4bf& mask, const Packet4bf& a,
                                                       const Packet4bf& b)
 {
   return pselect<Packet4us>(mask, a, b);
@@ -3507,7 +3507,7 @@ template<> EIGEN_STRONG_INLINE Packet4bf preverse<Packet4bf>(const Packet4bf& a)
   return preverse<Packet4us>(a);
 }
 
-EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4bf, 4>& kernel)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet4bf, 4>& kernel)
 {
   PacketBlock<Packet4us, 4> k;
   k.packet[0] = kernel.packet[0];
@@ -3573,7 +3573,7 @@ typedef float64x2_t Packet2d;
 typedef float64x1_t Packet1d;
 
 // fuctionally equivalent to _mm_shuffle_pd in SSE (i.e. shuffle(m, n, mask) equals _mm_shuffle_pd(m,n,mask))
-// Currently used in LU/arch/InverseSize4.h to enable a shared implementation 
+// Currently used in LU/arch/InverseSize4.h to enable a shared implementation
 // for fast inversion of matrices of size 4.
 EIGEN_STRONG_INLINE Packet2d shuffle(const Packet2d& m, const Packet2d& n, int mask)
 {
@@ -3739,7 +3739,7 @@ template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet2d& f
 template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from)
 { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f64(to,from); }
 
-template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet2d pgather<double, Packet2d>(const double* from, Index stride)
 {
   Packet2d res = pset1<Packet2d>(0.0);
   res = vld1q_lane_f64(from + 0*stride, res, 0);
@@ -3747,7 +3747,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const dou
   return res;
 }
 
-template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride)
 {
   vst1q_lane_f64(to + stride*0, from, 0);
   vst1q_lane_f64(to + stride*1, from, 1);
@@ -3791,7 +3791,7 @@ template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a)
 { return vgetq_lane_f64(vpmaxq_f64(a,a), 0); }
 
 
-EIGEN_DEVICE_FUNC inline void
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void
 ptranspose(PacketBlock<Packet2d, 2>& kernel)
 {
   const float64x2_t tmp1 = vzip1q_f64(kernel.packet[0], kernel.packet[1]);
@@ -3801,7 +3801,7 @@ ptranspose(PacketBlock<Packet2d, 2>& kernel)
   kernel.packet[1] = tmp2;
 }
 
-template<> EIGEN_DEVICE_FUNC inline Packet2d pselect( const Packet2d& mask, const Packet2d& a, const Packet2d& b)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet2d pselect( const Packet2d& mask, const Packet2d& a, const Packet2d& b)
 { return vbslq_f64(vreinterpretq_u64_f64(mask), a, b); }
 
 template<> EIGEN_STRONG_INLINE Packet2d pldexp<Packet2d>(const Packet2d& a, const Packet2d& exponent)
@@ -3829,7 +3829,7 @@ template<> EIGEN_STRONG_INLINE Packet2d psqrt(const Packet2d& _x){
                                       vcltq_f64(_x, pset1<Packet2d>((std::numeric_limits<double>::min)())));
   // Compute approximate reciprocal sqrt.
   Packet2d x = vrsqrteq_f64(_x);
-  // Do a single step of Newton's iteration. 
+  // Do a single step of Newton's iteration.
   //the number 1.5f was set reference to Quake3's fast inverse square root
   x = vmulq_f64(x, psub(pset1<Packet2d>(1.5), pmul(half, pmul(x, x))));
   // Do one more Newton's iteration to get more accurate result.
@@ -3838,7 +3838,7 @@ template<> EIGEN_STRONG_INLINE Packet2d psqrt(const Packet2d& _x){
   return vreinterpretq_f64_u64(vbicq_u64(vreinterpretq_u64_f64(pmul(_x, x)), denormal_mask));
 }
 
-#else 
+#else
 template<> EIGEN_STRONG_INLINE Packet2d psqrt(const Packet2d& _x){ return vsqrtq_f64(_x); }
 #endif
 
@@ -3914,7 +3914,7 @@ struct unpacket_traits<Packet8hf> {
 };
 
 template<>
-EIGEN_DEVICE_FUNC Packet4hf predux_half_dowto4<Packet8hf>(const Packet8hf& a) {
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4hf predux_half_dowto4<Packet8hf>(const Packet8hf& a) {
   return vadd_f16(vget_low_f16(a), vget_high_f16(a));
 }
 
@@ -4193,23 +4193,23 @@ EIGEN_STRONG_INLINE Packet8hf ploadquad<Packet8hf>(const Eigen::half* from) {
   return vcombine_f16(lo, hi);
 }
 
-EIGEN_DEVICE_FUNC inline Packet8hf pinsertfirst(const Packet8hf& a, Eigen::half b) { return vsetq_lane_f16(b.x, a, 0); }
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet8hf pinsertfirst(const Packet8hf& a, Eigen::half b) { return vsetq_lane_f16(b.x, a, 0); }
 
-EIGEN_DEVICE_FUNC inline Packet4hf pinsertfirst(const Packet4hf& a, Eigen::half b) { return vset_lane_f16(b.x, a, 0); }
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4hf pinsertfirst(const Packet4hf& a, Eigen::half b) { return vset_lane_f16(b.x, a, 0); }
 
 template <>
-EIGEN_DEVICE_FUNC inline Packet8hf pselect(const Packet8hf& mask, const Packet8hf& a, const Packet8hf& b) {
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet8hf pselect(const Packet8hf& mask, const Packet8hf& a, const Packet8hf& b) {
   return vbslq_f16(vreinterpretq_u16_f16(mask), a, b);
 }
 
 template <>
-EIGEN_DEVICE_FUNC inline Packet4hf pselect(const Packet4hf& mask, const Packet4hf& a, const Packet4hf& b) {
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4hf pselect(const Packet4hf& mask, const Packet4hf& a, const Packet4hf& b) {
   return vbsl_f16(vreinterpret_u16_f16(mask), a, b);
 }
 
-EIGEN_DEVICE_FUNC inline Packet8hf pinsertlast(const Packet8hf& a, Eigen::half b) { return vsetq_lane_f16(b.x, a, 7); }
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet8hf pinsertlast(const Packet8hf& a, Eigen::half b) { return vsetq_lane_f16(b.x, a, 7); }
 
-EIGEN_DEVICE_FUNC inline Packet4hf pinsertlast(const Packet4hf& a, Eigen::half b) { return vset_lane_f16(b.x, a, 3); }
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4hf pinsertlast(const Packet4hf& a, Eigen::half b) { return vset_lane_f16(b.x, a, 3); }
 
 template <>
 EIGEN_STRONG_INLINE void pstore<Eigen::half>(Eigen::half* to, const Packet8hf& from) {
@@ -4232,7 +4232,7 @@ EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half* to, const Packet4hf&
 }
 
 template <>
-EIGEN_DEVICE_FUNC inline Packet8hf pgather<Eigen::half, Packet8hf>(const Eigen::half* from, Index stride) {
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet8hf pgather<Eigen::half, Packet8hf>(const Eigen::half* from, Index stride) {
   Packet8hf res = pset1<Packet8hf>(Eigen::half(0.f));
   res = vsetq_lane_f16(from[0 * stride].x, res, 0);
   res = vsetq_lane_f16(from[1 * stride].x, res, 1);
@@ -4246,7 +4246,7 @@ EIGEN_DEVICE_FUNC inline Packet8hf pgather<Eigen::half, Packet8hf>(const Eigen::
 }
 
 template <>
-EIGEN_DEVICE_FUNC inline Packet4hf pgather<Eigen::half, Packet4hf>(const Eigen::half* from, Index stride) {
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4hf pgather<Eigen::half, Packet4hf>(const Eigen::half* from, Index stride) {
   Packet4hf res = pset1<Packet4hf>(Eigen::half(0.f));
   res = vset_lane_f16(from[0 * stride].x, res, 0);
   res = vset_lane_f16(from[1 * stride].x, res, 1);
@@ -4256,7 +4256,7 @@ EIGEN_DEVICE_FUNC inline Packet4hf pgather<Eigen::half, Packet4hf>(const Eigen::
 }
 
 template <>
-EIGEN_DEVICE_FUNC inline void pscatter<Eigen::half, Packet8hf>(Eigen::half* to, const Packet8hf& from, Index stride) {
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pscatter<Eigen::half, Packet8hf>(Eigen::half* to, const Packet8hf& from, Index stride) {
   to[stride * 0].x = vgetq_lane_f16(from, 0);
   to[stride * 1].x = vgetq_lane_f16(from, 1);
   to[stride * 2].x = vgetq_lane_f16(from, 2);
@@ -4268,7 +4268,7 @@ EIGEN_DEVICE_FUNC inline void pscatter<Eigen::half, Packet8hf>(Eigen::half* to,
 }
 
 template <>
-EIGEN_DEVICE_FUNC inline void pscatter<Eigen::half, Packet4hf>(Eigen::half* to, const Packet4hf& from, Index stride) {
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pscatter<Eigen::half, Packet4hf>(Eigen::half* to, const Packet4hf& from, Index stride) {
   to[stride * 0].x = vget_lane_f16(from, 0);
   to[stride * 1].x = vget_lane_f16(from, 1);
   to[stride * 2].x = vget_lane_f16(from, 2);
@@ -4422,7 +4422,7 @@ EIGEN_STRONG_INLINE Eigen::half predux_max<Packet4hf>(const Packet4hf& a) {
   return h;
 }
 
-EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8hf, 4>& kernel)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet8hf, 4>& kernel)
 {
   EIGEN_ALIGN16 Eigen::half in[4][8];
 
@@ -4451,7 +4451,7 @@ EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8hf, 4>& kernel)
   kernel.packet[3] = pload<Packet8hf>(out[3]);
 }
 
-EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4hf, 4>& kernel) {
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet4hf, 4>& kernel) {
   EIGEN_ALIGN16 float16x4x4_t tmp_x4;
   float16_t* tmp = (float16_t*)&kernel;
   tmp_x4 = vld4_f16(tmp);
@@ -4462,7 +4462,7 @@ EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4hf, 4>& kernel) {
   kernel.packet[3] = tmp_x4.val[3];
 }
 
-EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8hf, 8>& kernel) {
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet8hf, 8>& kernel) {
   float16x8x2_t T_1[4];
 
   T_1[0] = vuzpq_f16(kernel.packet[0], kernel.packet[1]);
-- 
cgit v1.2.3