diff --git a/Eigen/src/Core/arch/AVX512/PacketMath.h b/Eigen/src/Core/arch/AVX512/PacketMath.h
index 9e66575a9..9fbb256a1 100644
--- a/Eigen/src/Core/arch/AVX512/PacketMath.h
+++ b/Eigen/src/Core/arch/AVX512/PacketMath.h
@@ -551,7 +551,7 @@ EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet16i& from) {
 template <>
 EIGEN_DEVICE_FUNC inline Packet16f pgather<float, Packet16f>(const float* from,
                                                              Index stride) {
-  Packet16i stride_vector = _mm512_set1_epi32(stride);
+  Packet16i stride_vector = _mm512_set1_epi32(convert_index<int>(stride));
   Packet16i stride_multiplier =
       _mm512_set_epi32(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
   Packet16i indices = _mm512_mullo_epi32(stride_vector, stride_multiplier);
@@ -561,7 +561,7 @@ EIGEN_DEVICE_FUNC inline Packet16f pgather<float, Packet16f>(const float* from,
 template <>
 EIGEN_DEVICE_FUNC inline Packet8d pgather<double, Packet8d>(const double* from,
                                                             Index stride) {
-  Packet8i stride_vector = _mm256_set1_epi32(stride);
+  Packet8i stride_vector = _mm256_set1_epi32(convert_index<int>(stride));
   Packet8i stride_multiplier = _mm256_set_epi32(7, 6, 5, 4, 3, 2, 1, 0);
   Packet8i indices = _mm256_mullo_epi32(stride_vector, stride_multiplier);
 
@@ -572,7 +572,7 @@ template <>
 EIGEN_DEVICE_FUNC inline void pscatter<float, Packet16f>(float* to,
                                                          const Packet16f& from,
                                                          Index stride) {
-  Packet16i stride_vector = _mm512_set1_epi32(stride);
+  Packet16i stride_vector = _mm512_set1_epi32(convert_index<int>(stride));
   Packet16i stride_multiplier =
       _mm512_set_epi32(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
   Packet16i indices = _mm512_mullo_epi32(stride_vector, stride_multiplier);
@@ -582,7 +582,7 @@ template <>
 EIGEN_DEVICE_FUNC inline void pscatter<double, Packet8d>(double* to,
                                                          const Packet8d& from,
                                                          Index stride) {
-  Packet8i stride_vector = _mm256_set1_epi32(stride);
+  Packet8i stride_vector = _mm256_set1_epi32(convert_index<int>(stride));
   Packet8i stride_multiplier = _mm256_set_epi32(7, 6, 5, 4, 3, 2, 1, 0);
   Packet8i indices = _mm256_mullo_epi32(stride_vector, stride_multiplier);
   _mm512_i32scatter_pd(to, indices, from, 8);
@@ -660,8 +660,7 @@ EIGEN_STRONG_INLINE Packet8d pabs(const Packet8d& a) {
 
 #ifdef EIGEN_VECTORIZE_AVX512DQ
 #define EIGEN_INSERT_8f_INTO_16f(OUTPUT, INPUTA, INPUTB) \
-  OUTPUT = _mm512_insertf32x8(OUTPUT, INPUTA, 0);        \
-  OUTPUT = _mm512_insertf32x8(OUTPUT, INPUTB, 1);
+  OUTPUT = _mm512_insertf32x8(_mm512_castps256_ps512(INPUTA), INPUTB, 1);
 #else
 #define EIGEN_INSERT_8f_INTO_16f(OUTPUT, INPUTA, INPUTB)                    \
   OUTPUT = _mm512_insertf32x4(OUTPUT, _mm256_extractf128_ps(INPUTA, 0), 0); \
@@ -855,7 +854,7 @@ template<> EIGEN_STRONG_INLINE Packet8d preduxp<Packet8d>(const Packet8d* vecs)
 
   final_1 = _mm256_add_pd(final_1, _mm256_blend_pd(tmp0, tmp1, 0xC));
 
-  __m512d final_output = _mm512_insertf64x4(final_output, final_0, 0);
+  __m512d final_output = _mm512_castpd256_pd512(final_0);
 
   return _mm512_insertf64x4(final_output, final_1, 1);
 }
diff --git a/Eigen/src/Core/arch/GPU/PacketMathHalf.h b/Eigen/src/Core/arch/GPU/PacketMathHalf.h
index c4feda87d..8787adcde 100644
--- a/Eigen/src/Core/arch/GPU/PacketMathHalf.h
+++ b/Eigen/src/Core/arch/GPU/PacketMathHalf.h
@@ -542,11 +542,15 @@ template<> EIGEN_STRONG_INLINE Packet16h ploadu<Packet16h>(const Eigen::half* fr
 }
 
 template<> EIGEN_STRONG_INLINE void pstore<half>(Eigen::half* to, const Packet16h& from) {
-  _mm256_store_si256((__m256i*)to, from.x);
+  // (void*) -> workaround clang warning:
+  // cast from 'Eigen::half *' to '__m256i *' increases required alignment from 2 to 32
+  _mm256_store_si256((__m256i*)(void*)to, from.x);
 }
 
 template<> EIGEN_STRONG_INLINE void pstoreu<half>(Eigen::half* to, const Packet16h& from) {
-  _mm256_storeu_si256((__m256i*)to, from.x);
+  // (void*) -> workaround clang warning:
+  // cast from 'Eigen::half *' to '__m256i *' increases required alignment from 2 to 32
+  _mm256_storeu_si256((__m256i*)(void*)to, from.x);
 }
 
 template<> EIGEN_STRONG_INLINE Packet16h
diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
index 3ec8eb082..b012691c1 100644
--- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h
+++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
@@ -390,6 +390,7 @@ public:
   typedef typename conditional<Vectorizable,_LhsPacket,LhsScalar>::type LhsPacket;
   typedef typename conditional<Vectorizable,_RhsPacket,RhsScalar>::type RhsPacket;
   typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket;
+  typedef LhsPacket LhsPacket4Packing;
 
   typedef ResPacket AccPacket;
   
@@ -496,6 +497,7 @@ public:
   typedef typename conditional<Vectorizable,_LhsPacket,LhsScalar>::type LhsPacket;
   typedef typename conditional<Vectorizable,_RhsPacket,RhsScalar>::type RhsPacket;
   typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket;
+  typedef LhsPacket LhsPacket4Packing;
 
   typedef ResPacket AccPacket;
 
@@ -626,6 +628,7 @@ public:
   typedef typename packet_traits<Scalar>::type     ScalarPacket;
   typedef DoublePacket<RealPacket> DoublePacketType;
 
+  typedef typename conditional<Vectorizable,ScalarPacket,Scalar>::type LhsPacket4Packing;
   typedef typename conditional<Vectorizable,RealPacket,  Scalar>::type LhsPacket;
   typedef typename conditional<Vectorizable,DoublePacketType,Scalar>::type RhsPacket;
   typedef typename conditional<Vectorizable,ScalarPacket,Scalar>::type ResPacket;
@@ -777,6 +780,7 @@ public:
   typedef typename conditional<Vectorizable,_LhsPacket,LhsScalar>::type LhsPacket;
   typedef typename conditional<Vectorizable,_RhsPacket,RhsScalar>::type RhsPacket;
   typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket;
+  typedef LhsPacket LhsPacket4Packing;
 
   typedef ResPacket AccPacket;
 
@@ -1025,9 +1029,9 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
           ResPacket R0, R1, R2;
           ResPacket alphav = pset1<ResPacket>(alpha);
 
-          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
-          R1 = r0.loadPacket(1 * Traits::ResPacketSize);
-          R2 = r0.loadPacket(2 * Traits::ResPacketSize);
+          R0 = r0.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R1 = r0.template loadPacket<ResPacket>(1 * Traits::ResPacketSize);
+          R2 = r0.template loadPacket<ResPacket>(2 * Traits::ResPacketSize);
           traits.acc(C0, alphav, R0);
           traits.acc(C4, alphav, R1);
           traits.acc(C8, alphav, R2);
@@ -1035,9 +1039,9 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
           r0.storePacket(1 * Traits::ResPacketSize, R1);
           r0.storePacket(2 * Traits::ResPacketSize, R2);
 
-          R0 = r1.loadPacket(0 * Traits::ResPacketSize);
-          R1 = r1.loadPacket(1 * Traits::ResPacketSize);
-          R2 = r1.loadPacket(2 * Traits::ResPacketSize);
+          R0 = r1.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R1 = r1.template loadPacket<ResPacket>(1 * Traits::ResPacketSize);
+          R2 = r1.template loadPacket<ResPacket>(2 * Traits::ResPacketSize);
           traits.acc(C1, alphav, R0);
           traits.acc(C5, alphav, R1);
           traits.acc(C9, alphav, R2);
@@ -1045,9 +1049,9 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
           r1.storePacket(1 * Traits::ResPacketSize, R1);
           r1.storePacket(2 * Traits::ResPacketSize, R2);
 
-          R0 = r2.loadPacket(0 * Traits::ResPacketSize);
-          R1 = r2.loadPacket(1 * Traits::ResPacketSize);
-          R2 = r2.loadPacket(2 * Traits::ResPacketSize);
+          R0 = r2.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R1 = r2.template loadPacket<ResPacket>(1 * Traits::ResPacketSize);
+          R2 = r2.template loadPacket<ResPacket>(2 * Traits::ResPacketSize);
           traits.acc(C2, alphav, R0);
           traits.acc(C6, alphav, R1);
           traits.acc(C10, alphav, R2);
@@ -1055,9 +1059,9 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
           r2.storePacket(1 * Traits::ResPacketSize, R1);
           r2.storePacket(2 * Traits::ResPacketSize, R2);
 
-          R0 = r3.loadPacket(0 * Traits::ResPacketSize);
-          R1 = r3.loadPacket(1 * Traits::ResPacketSize);
-          R2 = r3.loadPacket(2 * Traits::ResPacketSize);
+          R0 = r3.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R1 = r3.template loadPacket<ResPacket>(1 * Traits::ResPacketSize);
+          R2 = r3.template loadPacket<ResPacket>(2 * Traits::ResPacketSize);
           traits.acc(C3, alphav, R0);
           traits.acc(C7, alphav, R1);
           traits.acc(C11, alphav, R2);
@@ -1134,9 +1138,9 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
           ResPacket R0, R1, R2;
           ResPacket alphav = pset1<ResPacket>(alpha);
 
-          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
-          R1 = r0.loadPacket(1 * Traits::ResPacketSize);
-          R2 = r0.loadPacket(2 * Traits::ResPacketSize);
+          R0 = r0.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R1 = r0.template loadPacket<ResPacket>(1 * Traits::ResPacketSize);
+          R2 = r0.template loadPacket<ResPacket>(2 * Traits::ResPacketSize);
           traits.acc(C0, alphav, R0);
           traits.acc(C4, alphav, R1);
           traits.acc(C8, alphav, R2);
@@ -1244,10 +1248,10 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
           ResPacket R0, R1, R2, R3;
           ResPacket alphav = pset1<ResPacket>(alpha);
 
-          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
-          R1 = r0.loadPacket(1 * Traits::ResPacketSize);
-          R2 = r1.loadPacket(0 * Traits::ResPacketSize);
-          R3 = r1.loadPacket(1 * Traits::ResPacketSize);
+          R0 = r0.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R1 = r0.template loadPacket<ResPacket>(1 * Traits::ResPacketSize);
+          R2 = r1.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R3 = r1.template loadPacket<ResPacket>(1 * Traits::ResPacketSize);
           traits.acc(C0, alphav, R0);
           traits.acc(C4, alphav, R1);
           traits.acc(C1, alphav, R2);
@@ -1257,10 +1261,10 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
           r1.storePacket(0 * Traits::ResPacketSize, R2);
           r1.storePacket(1 * Traits::ResPacketSize, R3);
 
-          R0 = r2.loadPacket(0 * Traits::ResPacketSize);
-          R1 = r2.loadPacket(1 * Traits::ResPacketSize);
-          R2 = r3.loadPacket(0 * Traits::ResPacketSize);
-          R3 = r3.loadPacket(1 * Traits::ResPacketSize);
+          R0 = r2.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R1 = r2.template loadPacket<ResPacket>(1 * Traits::ResPacketSize);
+          R2 = r3.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R3 = r3.template loadPacket<ResPacket>(1 * Traits::ResPacketSize);
           traits.acc(C2,  alphav, R0);
           traits.acc(C6,  alphav, R1);
           traits.acc(C3,  alphav, R2);
@@ -1337,8 +1341,8 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
           ResPacket R0, R1;
           ResPacket alphav = pset1<ResPacket>(alpha);
 
-          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
-          R1 = r0.loadPacket(1 * Traits::ResPacketSize);
+          R0 = r0.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R1 = r0.template loadPacket<ResPacket>(1 * Traits::ResPacketSize);
           traits.acc(C0, alphav, R0);
           traits.acc(C4, alphav, R1);
           r0.storePacket(0 * Traits::ResPacketSize, R0);
@@ -1431,15 +1435,15 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
           ResPacket R0, R1;
           ResPacket alphav = pset1<ResPacket>(alpha);
 
-          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
-          R1 = r1.loadPacket(0 * Traits::ResPacketSize);
+          R0 = r0.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R1 = r1.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
           traits.acc(C0, alphav, R0);
           traits.acc(C1,  alphav, R1);
           r0.storePacket(0 * Traits::ResPacketSize, R0);
           r1.storePacket(0 * Traits::ResPacketSize, R1);
 
-          R0 = r2.loadPacket(0 * Traits::ResPacketSize);
-          R1 = r3.loadPacket(0 * Traits::ResPacketSize);
+          R0 = r2.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
+          R1 = r3.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
           traits.acc(C2,  alphav, R0);
           traits.acc(C3,  alphav, R1);
           r2.storePacket(0 * Traits::ResPacketSize, R0);
@@ -1504,7 +1508,7 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
 #undef EIGEN_GEBGP_ONESTEP
           ResPacket R0;
           ResPacket alphav = pset1<ResPacket>(alpha);
-          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
+          R0 = r0.template loadPacket<ResPacket>(0 * Traits::ResPacketSize);
           traits.acc(C0, alphav, R0);
           r0.storePacket(0 * Traits::ResPacketSize, R0);
         }
@@ -1685,19 +1689,18 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
 //
 //  32 33 34 35 ...
 //  36 36 38 39 ...
-template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, bool Conjugate, bool PanelMode>
-struct gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, ColMajor, Conjugate, PanelMode>
+template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, typename Packet, bool Conjugate, bool PanelMode>
+struct gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Packet, ColMajor, Conjugate, PanelMode>
 {
   typedef typename DataMapper::LinearMapper LinearMapper;
   EIGEN_DONT_INLINE void operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride=0, Index offset=0);
 };
 
-template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, bool Conjugate, bool PanelMode>
-EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, ColMajor, Conjugate, PanelMode>
+template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, typename Packet, bool Conjugate, bool PanelMode>
+EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Packet, ColMajor, Conjugate, PanelMode>
   ::operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride, Index offset)
 {
-  typedef typename packet_traits<Scalar>::type Packet;
-  enum { PacketSize = packet_traits<Scalar>::size };
+  enum { PacketSize = unpacket_traits<Packet>::size };
 
   EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK LHS");
   EIGEN_UNUSED_VARIABLE(stride);
@@ -1725,9 +1728,9 @@ EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Co
       for(Index k=0; k<depth; k++)
       {
         Packet A, B, C;
-        A = lhs.loadPacket(i+0*PacketSize, k);
-        B = lhs.loadPacket(i+1*PacketSize, k);
-        C = lhs.loadPacket(i+2*PacketSize, k);
+        A = lhs.template loadPacket<Packet>(i+0*PacketSize, k);
+        B = lhs.template loadPacket<Packet>(i+1*PacketSize, k);
+        C = lhs.template loadPacket<Packet>(i+2*PacketSize, k);
         pstore(blockA+count, cj.pconj(A)); count+=PacketSize;
         pstore(blockA+count, cj.pconj(B)); count+=PacketSize;
         pstore(blockA+count, cj.pconj(C)); count+=PacketSize;
@@ -1745,8 +1748,8 @@ EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Co
       for(Index k=0; k<depth; k++)
       {
         Packet A, B;
-        A = lhs.loadPacket(i+0*PacketSize, k);
-        B = lhs.loadPacket(i+1*PacketSize, k);
+        A = lhs.template loadPacket<Packet>(i+0*PacketSize, k);
+        B = lhs.template loadPacket<Packet>(i+1*PacketSize, k);
         pstore(blockA+count, cj.pconj(A)); count+=PacketSize;
         pstore(blockA+count, cj.pconj(B)); count+=PacketSize;
       }
@@ -1763,7 +1766,7 @@ EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Co
       for(Index k=0; k<depth; k++)
       {
         Packet A;
-        A = lhs.loadPacket(i+0*PacketSize, k);
+        A = lhs.template loadPacket<Packet>(i+0*PacketSize, k);
         pstore(blockA+count, cj.pconj(A));
         count+=PacketSize;
       }
@@ -1793,19 +1796,18 @@ EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Co
   }
 }
 
-template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, bool Conjugate, bool PanelMode>
-struct gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, RowMajor, Conjugate, PanelMode>
+template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, typename Packet, bool Conjugate, bool PanelMode>
+struct gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Packet, RowMajor, Conjugate, PanelMode>
 {
   typedef typename DataMapper::LinearMapper LinearMapper;
   EIGEN_DONT_INLINE void operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride=0, Index offset=0);
 };
 
-template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, bool Conjugate, bool PanelMode>
-EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, RowMajor, Conjugate, PanelMode>
+template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, typename Packet, bool Conjugate, bool PanelMode>
+EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Packet, RowMajor, Conjugate, PanelMode>
   ::operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride, Index offset)
 {
-  typedef typename packet_traits<Scalar>::type Packet;
-  enum { PacketSize = packet_traits<Scalar>::size };
+  enum { PacketSize = unpacket_traits<Packet>::size };
 
   EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK LHS");
   EIGEN_UNUSED_VARIABLE(stride);
@@ -1837,7 +1839,7 @@ EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Ro
           for (Index m = 0; m < pack; m += PacketSize)
           {
             PacketBlock<Packet> kernel;
-            for (int p = 0; p < PacketSize; ++p) kernel.packet[p] = lhs.loadPacket(i+p+m, k);
+            for (int p = 0; p < PacketSize; ++p) kernel.packet[p] = lhs.template loadPacket<Packet>(i+p+m, k);
             ptranspose(kernel);
             for (int p = 0; p < PacketSize; ++p) pstore(blockA+count+m+(pack)*p, cj.pconj(kernel.packet[p]));
           }
@@ -1971,10 +1973,10 @@ EIGEN_DONT_INLINE void gemm_pack_rhs<Scalar, Index, DataMapper, nr, ColMajor, Co
       {
         for(; k<peeled_k; k+=PacketSize) {
           PacketBlock<Packet,(PacketSize%4)==0?4:PacketSize> kernel;
-          kernel.packet[0] = dm0.loadPacket(k);
-          kernel.packet[1%PacketSize] = dm1.loadPacket(k);
-          kernel.packet[2%PacketSize] = dm2.loadPacket(k);
-          kernel.packet[3%PacketSize] = dm3.loadPacket(k);
+          kernel.packet[0           ] = dm0.template loadPacket<Packet>(k);
+          kernel.packet[1%PacketSize] = dm1.template loadPacket<Packet>(k);
+          kernel.packet[2%PacketSize] = dm2.template loadPacket<Packet>(k);
+          kernel.packet[3%PacketSize] = dm3.template loadPacket<Packet>(k);
           ptranspose(kernel);
           pstoreu(blockB+count+0*PacketSize, cj.pconj(kernel.packet[0]));
           pstoreu(blockB+count+1*PacketSize, cj.pconj(kernel.packet[1%PacketSize]));
@@ -2075,7 +2077,7 @@ EIGEN_DONT_INLINE void gemm_pack_rhs<Scalar, Index, DataMapper, nr, RowMajor, Co
       for(Index k=0; k<depth; k++)
       {
         if (PacketSize==4) {
-          Packet A = rhs.loadPacket(k, j2);
+          Packet A = rhs.template loadPacket<Packet>(k, j2);
           pstoreu(blockB+count, cj.pconj(A));
           count += PacketSize;
         } else {
diff --git a/Eigen/src/Core/products/GeneralMatrixMatrix.h b/Eigen/src/Core/products/GeneralMatrixMatrix.h
index bd7b6ff2a..f49abcad5 100644
--- a/Eigen/src/Core/products/GeneralMatrixMatrix.h
+++ b/Eigen/src/Core/products/GeneralMatrixMatrix.h
@@ -75,7 +75,7 @@ static void run(Index rows, Index cols, Index depth,
   Index mc = (std::min)(rows,blocking.mc());  // cache block size along the M direction
   Index nc = (std::min)(cols,blocking.nc());  // cache block size along the N direction
 
-  gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+  gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, LhsStorageOrder> pack_lhs;
   gemm_pack_rhs<RhsScalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs;
   gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp;
 
diff --git a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h
index e436c50a4..ec2825bf0 100644
--- a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h
+++ b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h
@@ -84,7 +84,7 @@ struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,
     ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, sizeA, blocking.blockA());
     ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, blocking.blockB());
 
-    gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, LhsStorageOrder> pack_lhs;
     gemm_pack_rhs<RhsScalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs;
     gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp;
     tribb_kernel<LhsScalar, RhsScalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs, UpLo> sybb;
@@ -110,7 +110,6 @@ struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,
           gebp(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc,
                (std::min)(size,i2), alpha, -1, -1, 0, 0);
 
-
         sybb(_res+resStride*i2 + i2, resStride, blockA, blockB + actual_kc*i2, actual_mc, actual_kc, alpha);
 
         if (UpLo==Upper)
@@ -160,7 +159,7 @@ struct tribb_kernel
       if(UpLo==Upper)
         gebp_kernel(res.getSubMapper(0, j), blockA, actual_b, j, depth, actualBlockSize, alpha,
                     -1, -1, 0, 0);
-
+      
       // selfadjoint micro block
       {
         Index i = j;
@@ -168,6 +167,7 @@ struct tribb_kernel
         // 1 - apply the kernel on the temporary buffer
         gebp_kernel(ResMapper(buffer.data(), BlockSize), blockA+depth*i, actual_b, actualBlockSize, depth, actualBlockSize, alpha,
                     -1, -1, 0, 0);
+
         // 2 - triangular accumulation
         for(Index j1=0; j1<actualBlockSize; ++j1)
         {
diff --git a/Eigen/src/Core/products/SelfadjointMatrixMatrix.h b/Eigen/src/Core/products/SelfadjointMatrixMatrix.h
index da6f82abc..c84c71609 100644
--- a/Eigen/src/Core/products/SelfadjointMatrixMatrix.h
+++ b/Eigen/src/Core/products/SelfadjointMatrixMatrix.h
@@ -352,7 +352,7 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,t
     gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
     symm_pack_lhs<Scalar, Index, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs;
-    gemm_pack_lhs<Scalar, Index, LhsTransposeMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder==RowMajor?ColMajor:RowMajor, true> pack_lhs_transposed;
+    gemm_pack_lhs<Scalar, Index, LhsTransposeMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, LhsStorageOrder==RowMajor?ColMajor:RowMajor, true> pack_lhs_transposed;
 
     for(Index k2=0; k2<size; k2+=kc)
     {
@@ -387,7 +387,7 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,t
       for(Index i2=k2+kc; i2<size; i2+=mc)
       {
         const Index actual_mc = (std::min)(i2+mc,size)-i2;
-        gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder,false>()
+        gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, LhsStorageOrder,false>()
           (blockA, lhs.getSubMapper(i2, k2), actual_kc, actual_mc);
 
         gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, cols, alpha);
@@ -437,7 +437,7 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,f
     ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
 
     gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
-    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, LhsStorageOrder> pack_lhs;
     symm_pack_rhs<Scalar, Index, Traits::nr,RhsStorageOrder> pack_rhs;
 
     for(Index k2=0; k2<size; k2+=kc)
diff --git a/Eigen/src/Core/products/TriangularMatrixMatrix.h b/Eigen/src/Core/products/TriangularMatrixMatrix.h
index f784507e7..85fd744b9 100644
--- a/Eigen/src/Core/products/TriangularMatrixMatrix.h
+++ b/Eigen/src/Core/products/TriangularMatrixMatrix.h
@@ -151,7 +151,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
       triangularBuffer.diagonal().setOnes();
 
     gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
-    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, LhsStorageOrder> pack_lhs;
     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs;
 
     for(Index k2=IsLower ? depth : 0;
@@ -222,7 +222,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
         for(Index i2=start; i2<end; i2+=mc)
         {
           const Index actual_mc = (std::min)(i2+mc,end)-i2;
-          gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr,Traits::LhsProgress, LhsStorageOrder,false>()
+          gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr,Traits::LhsProgress, typename Traits::LhsPacket4Packing, LhsStorageOrder,false>()
             (blockA, lhs.getSubMapper(i2, actual_k2), actual_kc, actual_mc);
 
           gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc,
@@ -299,7 +299,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
       triangularBuffer.diagonal().setOnes();
 
     gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
-    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, LhsStorageOrder> pack_lhs;
     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs;
     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder,false,true> pack_rhs_panel;
 
diff --git a/Eigen/src/Core/products/TriangularSolverMatrix.h b/Eigen/src/Core/products/TriangularSolverMatrix.h
index 223c38b86..8ff2e9d9d 100644
--- a/Eigen/src/Core/products/TriangularSolverMatrix.h
+++ b/Eigen/src/Core/products/TriangularSolverMatrix.h
@@ -76,7 +76,7 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conju
 
     conj_if<Conjugate> conj;
     gebp_kernel<Scalar, Scalar, Index, OtherMapper, Traits::mr, Traits::nr, Conjugate, false> gebp_kernel;
-    gemm_pack_lhs<Scalar, Index, TriMapper, Traits::mr, Traits::LhsProgress, TriStorageOrder> pack_lhs;
+    gemm_pack_lhs<Scalar, Index, TriMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, TriStorageOrder> pack_lhs;
     gemm_pack_rhs<Scalar, Index, OtherMapper, Traits::nr, ColMajor, false, true> pack_rhs;
 
     // the goal here is to subdivise the Rhs panels such that we keep some cache
@@ -229,7 +229,7 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conj
     gebp_kernel<Scalar, Scalar, Index, LhsMapper, Traits::mr, Traits::nr, false, Conjugate> gebp_kernel;
     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs;
     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr, RhsStorageOrder,false,true> pack_rhs_panel;
-    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, ColMajor, false, true> pack_lhs_panel;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, ColMajor, false, true> pack_lhs_panel;
 
     for(Index k2=IsLower ? size : 0;
         IsLower ? k2>0 : k2<size;
diff --git a/Eigen/src/Core/util/BlasUtil.h b/Eigen/src/Core/util/BlasUtil.h
index 705925984..a32630ed7 100755
--- a/Eigen/src/Core/util/BlasUtil.h
+++ b/Eigen/src/Core/util/BlasUtil.h
@@ -24,7 +24,7 @@ struct gebp_kernel;
 template<typename Scalar, typename Index, typename DataMapper, int nr, int StorageOrder, bool Conjugate = false, bool PanelMode=false>
 struct gemm_pack_rhs;
 
-template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, int StorageOrder, bool Conjugate = false, bool PanelMode = false>
+template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, typename Packet, int StorageOrder, bool Conjugate = false, bool PanelMode = false>
 struct gemm_pack_lhs;
 
 template<
@@ -156,11 +156,9 @@ class BlasVectorMapper {
 };
 
 template<typename Scalar, typename Index, int AlignmentType>
-class BlasLinearMapper {
-  public:
-  typedef typename packet_traits<Scalar>::type Packet;
-  typedef typename packet_traits<Scalar>::half HalfPacket;
-
+class BlasLinearMapper
+{
+public:
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE BlasLinearMapper(Scalar *data) : m_data(data) {}
 
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void prefetch(int i) const {
@@ -171,29 +169,25 @@ class BlasLinearMapper {
     return m_data[i];
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet loadPacket(Index i) const {
-    return ploadt<Packet, AlignmentType>(m_data + i);
+  template<typename PacketType>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE PacketType loadPacket(Index i) const {
+    return ploadt<PacketType, AlignmentType>(m_data + i);
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE HalfPacket loadHalfPacket(Index i) const {
-    return ploadt<HalfPacket, AlignmentType>(m_data + i);
+  template<typename PacketType>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void storePacket(Index i, const PacketType &p) const {
+    pstoret<Scalar, PacketType, AlignmentType>(m_data + i, p);
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void storePacket(Index i, const Packet &p) const {
-    pstoret<Scalar, Packet, AlignmentType>(m_data + i, p);
-  }
-
-  protected:
+protected:
   Scalar *m_data;
 };
 
 // Lightweight helper class to access matrix coefficients.
 template<typename Scalar, typename Index, int StorageOrder, int AlignmentType = Unaligned>
-class blas_data_mapper {
-  public:
-  typedef typename packet_traits<Scalar>::type Packet;
-  typedef typename packet_traits<Scalar>::half HalfPacket;
-
+class blas_data_mapper
+{
+public:
   typedef BlasLinearMapper<Scalar, Index, AlignmentType> LinearMapper;
   typedef BlasVectorMapper<Scalar, Index> VectorMapper;
 
@@ -218,8 +212,9 @@ class blas_data_mapper {
     return m_data[StorageOrder==RowMajor ? j + i*m_stride : i + j*m_stride];
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet loadPacket(Index i, Index j) const {
-    return ploadt<Packet, AlignmentType>(&operator()(i, j));
+  template<typename PacketType>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE PacketType loadPacket(Index i, Index j) const {
+    return ploadt<PacketType, AlignmentType>(&operator()(i, j));
   }
 
   template <typename PacketT, int AlignmentT>
@@ -227,10 +222,6 @@ class blas_data_mapper {
     return ploadt<PacketT, AlignmentT>(&operator()(i, j));
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE HalfPacket loadHalfPacket(Index i, Index j) const {
-    return ploadt<HalfPacket, AlignmentType>(&operator()(i, j));
-  }
-
   template<typename SubPacket>
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void scatterPacket(Index i, Index j, const SubPacket &p) const {
     pscatter<Scalar, SubPacket>(&operator()(i, j), p, m_stride);
@@ -251,7 +242,7 @@ class blas_data_mapper {
     return internal::first_default_aligned(m_data, size);
   }
 
-  protected:
+protected:
   Scalar* EIGEN_RESTRICT m_data;
   const Index m_stride;
 };
diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index b4730cff0..c215bfbb2 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -198,7 +198,7 @@ ei_add_test(smallvectors)
 ei_add_test(mapped_matrix)
 ei_add_test(mapstride)
 ei_add_test(mapstaticmethods)
-ei_add_test(array)
+ei_add_test(array_cwise)
 ei_add_test(array_for_matrix)
 ei_add_test(array_replicate)
 ei_add_test(array_reverse)
diff --git a/test/array.cpp b/test/array_cwise.cpp
similarity index 99%
rename from test/array.cpp
rename to test/array_cwise.cpp
index d9c4626c0..84e46665b 100644
--- a/test/array.cpp
+++ b/test/array_cwise.cpp
@@ -453,7 +453,7 @@ template<typename ArrayType> void min_max(const ArrayType& m)
 
 }
 
-EIGEN_DECLARE_TEST(array)
+EIGEN_DECLARE_TEST(array_cwise)
 {
   for(int i = 0; i < g_repeat; i++) {
     CALL_SUBTEST_1( array(Array<float, 1, 1>()) );
diff --git a/test/packetmath.cpp b/test/packetmath.cpp
index 58a1c60bf..2b0dda573 100644
--- a/test/packetmath.cpp
+++ b/test/packetmath.cpp
@@ -148,24 +148,25 @@ template<typename Scalar> void packetmath()
 
   for (int offset=0; offset<PacketSize; ++offset)
   {
+    #define MIN(A,B) (A<B?A:B)
     packets[0] = internal::pload<Packet>(data1);
     packets[1] = internal::pload<Packet>(data1+PacketSize);
          if (offset==0) internal::palign<0>(packets[0], packets[1]);
-    else if (offset==1) internal::palign<1>(packets[0], packets[1]);
-    else if (offset==2) internal::palign<2>(packets[0], packets[1]);
-    else if (offset==3) internal::palign<3>(packets[0], packets[1]);
-    else if (offset==4) internal::palign<4>(packets[0], packets[1]);
-    else if (offset==5) internal::palign<5>(packets[0], packets[1]);
-    else if (offset==6) internal::palign<6>(packets[0], packets[1]);
-    else if (offset==7) internal::palign<7>(packets[0], packets[1]);
-    else if (offset==8) internal::palign<8>(packets[0], packets[1]);
-    else if (offset==9) internal::palign<9>(packets[0], packets[1]);
-    else if (offset==10) internal::palign<10>(packets[0], packets[1]);
-    else if (offset==11) internal::palign<11>(packets[0], packets[1]);
-    else if (offset==12) internal::palign<12>(packets[0], packets[1]);
-    else if (offset==13) internal::palign<13>(packets[0], packets[1]);
-    else if (offset==14) internal::palign<14>(packets[0], packets[1]);
-    else if (offset==15) internal::palign<15>(packets[0], packets[1]);
+    else if (offset==1) internal::palign<MIN(1,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==2) internal::palign<MIN(2,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==3) internal::palign<MIN(3,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==4) internal::palign<MIN(4,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==5) internal::palign<MIN(5,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==6) internal::palign<MIN(6,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==7) internal::palign<MIN(7,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==8) internal::palign<MIN(8,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==9) internal::palign<MIN(9,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==10) internal::palign<MIN(10,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==11) internal::palign<MIN(11,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==12) internal::palign<MIN(12,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==13) internal::palign<MIN(13,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==14) internal::palign<MIN(14,PacketSize-1)>(packets[0], packets[1]);
+    else if (offset==15) internal::palign<MIN(15,PacketSize-1)>(packets[0], packets[1]);
     internal::pstore(data2, packets[0]);
 
     for (int i=0; i<PacketSize; ++i)
diff --git a/test/vectorization_logic.cpp b/test/vectorization_logic.cpp
index 01b55e192..c15f75103 100644
--- a/test/vectorization_logic.cpp
+++ b/test/vectorization_logic.cpp
@@ -22,6 +22,14 @@
 #include "main.h"
 #include <typeinfo>
 
+// Disable "ignoring attributes on template argument"
+// for packet_traits<Packet*>
+// => The only workaround would be to wrap _m128 and the likes
+//    within wrappers.
+#if EIGEN_GNUC_AT_LEAST(6,0)
+    #pragma GCC diagnostic ignored "-Wignored-attributes"
+#endif
+
 using internal::demangle_flags;
 using internal::demangle_traversal;
 using internal::demangle_unrolling;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
index f0f61fade..3b22e43e7 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
@@ -620,7 +620,7 @@ struct TensorContractionEvaluatorBase
     typedef internal::blas_data_mapper<Scalar, Index, ColMajor> OutputMapper;
 
     // Declare GEBP packing and kernel structs
-    internal::gemm_pack_lhs<LhsScalar, Index, typename LhsMapper::SubMapper, mr, Traits::LhsProgress, ColMajor> pack_lhs;
+    internal::gemm_pack_lhs<LhsScalar, Index, typename LhsMapper::SubMapper, mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, ColMajor> pack_lhs;
     internal::gemm_pack_rhs<RhsScalar, Index, typename RhsMapper::SubMapper, nr, ColMajor> pack_rhs;
 
     internal::gebp_kernel<LhsScalar, RhsScalar, Index, OutputMapper, mr, nr, false, false> gebp;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
index b5e186d21..056665749 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
@@ -549,12 +549,12 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
 #define prefetch_lhs(reg, row, col)                            \
     if (!CHECK_LHS_BOUNDARY) {                                 \
       if (col < k_size) {                                      \
-        reg =lhs.template loadPacket<Unaligned>(row, col);     \
+        reg =lhs.template loadPacket<float4,Unaligned>(row, col);     \
       }                                                        \
     } else {                                                   \
       if (col < k_size) {                                      \
         if (row + 3 < m_size) {                                \
-          reg =lhs.template loadPacket<Unaligned>(row, col);   \
+          reg =lhs.template loadPacket<float4,Unaligned>(row, col);   \
         } else if (row + 2 < m_size) {                         \
           reg.x =lhs(row + 0, col);                            \
           reg.y =lhs(row + 1, col);                            \
@@ -584,7 +584,7 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
     if (!CHECK_RHS_BOUNDARY) {
       if ((rhs_vert + 3) < k_size) {
         // just CHECK_RHS_BOUNDARY
-        rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
+        rhs_pf0 = rhs.template loadPacket<float4,Unaligned>(rhs_vert, rhs_horiz0);
       } else if (rhs_vert + 2 < k_size) {
         // just CHECK_RHS_BOUNDARY
         rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
@@ -599,7 +599,7 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
     } else {
       if (rhs_horiz0 < n_size) {
         if ((rhs_vert + 3) < k_size) {
-          rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
+          rhs_pf0 = rhs.template loadPacket<float4,Unaligned>(rhs_vert, rhs_horiz0);
         } else if ((rhs_vert + 2) < k_size) {
           rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
           rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
@@ -799,37 +799,37 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
 
      if (!CHECK_LHS_BOUNDARY) {
       if ((threadIdx.y/4+k+24) < k_size) {
-        lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
-        lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
-        lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
-        lhs_pf3 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+24));
+        lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
+        lhs_pf1 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+        lhs_pf2 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
+        lhs_pf3 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+24));
       } else if ((threadIdx.y/4+k+16) < k_size) {
-        lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
-        lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
-        lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
+        lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
+        lhs_pf1 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+        lhs_pf2 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
       } else if ((threadIdx.y/4+k+8) < k_size) {
-        lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
-        lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+        lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
+        lhs_pf1 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
       } else if ((threadIdx.y/4+k) < k_size) {
-        lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
+        lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
       }
     } else {
       // just CHECK_LHS_BOUNDARY
       if (lhs_vert + 3 < m_size) {
         if ((threadIdx.y/4+k+24) < k_size) {
-          lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
-          lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
-          lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
-          lhs_pf3 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+24));
+          lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
+          lhs_pf1 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+          lhs_pf2 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
+          lhs_pf3 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+24));
         } else if ((threadIdx.y/4+k+16) < k_size) {
-          lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
-          lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
-          lhs_pf2 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
+          lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
+          lhs_pf1 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+          lhs_pf2 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
         } else if ((threadIdx.y/4+k+8) < k_size) {
-          lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
-          lhs_pf1 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+          lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
+          lhs_pf1 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
         } else if ((threadIdx.y/4+k) < k_size) {
-          lhs_pf0 =lhs.template loadPacket<Unaligned>(lhs_vert, (threadIdx.y/4+k));
+          lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
         }
       } else if (lhs_vert + 2 < m_size) {
         if ((threadIdx.y/4+k+24) < k_size) {
@@ -918,8 +918,8 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
     if (!CHECK_RHS_BOUNDARY) {
       if ((rhs_vert + 3) < k_size) {
         // just CHECK_RHS_BOUNDARY
-        rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
-        rhs_pf1 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz1);
+        rhs_pf0 = rhs.template loadPacket<float4,Unaligned>(rhs_vert, rhs_horiz0);
+        rhs_pf1 = rhs.template loadPacket<float4,Unaligned>(rhs_vert, rhs_horiz1);
       } else if (rhs_vert + 2 < k_size) {
         // just CHECK_RHS_BOUNDARY
         rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
@@ -941,8 +941,8 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
       if (rhs_horiz1 < n_size) {
         if ((rhs_vert + 3) < k_size) {
           // just CHECK_RHS_BOUNDARY
-          rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
-          rhs_pf1 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz1);
+          rhs_pf0 = rhs.template loadPacket<float4,Unaligned>(rhs_vert, rhs_horiz0);
+          rhs_pf1 = rhs.template loadPacket<float4,Unaligned>(rhs_vert, rhs_horiz1);
         } else if (rhs_vert + 2 < k_size) {
           // just CHECK_RHS_BOUNDARY
           rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
@@ -963,7 +963,7 @@ EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
       } else if (rhs_horiz0 < n_size) {
         if ((rhs_vert + 3) < k_size) {
           // just CHECK_RHS_BOUNDARY
-          rhs_pf0 = rhs.template loadPacket<Unaligned>(rhs_vert, rhs_horiz0);
+          rhs_pf0 = rhs.template loadPacket<float4,Unaligned>(rhs_vert, rhs_horiz0);
         } else if ((rhs_vert + 2) < k_size) {
           // just CHECK_RHS_BOUNDARY
           rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionMapper.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionMapper.h
index ab320a50d..dbb0f76bb 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionMapper.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionMapper.h
@@ -238,9 +238,6 @@ class BaseTensorContractionMapper : public SimpleTensorContractionMapper<Scalar,
                               const contract_t& k_strides) :
   ParentMapper(tensor, nocontract_strides, ij_strides, contract_strides, k_strides) { }
 
-  typedef typename Tensor::PacketReturnType Packet;
-  typedef typename unpacket_traits<Packet>::half HalfPacket;
-
   template <typename PacketT,int AlignmentType>
   EIGEN_DEVICE_FUNC
   EIGEN_STRONG_INLINE PacketT load(Index i, Index j) const {
@@ -284,27 +281,10 @@ class BaseTensorContractionMapper : public SimpleTensorContractionMapper<Scalar,
     return pload<PacketT>(data);
   }
 
-  template <int AlignmentType>
+  template <typename PacketT,int AlignmentType>
   EIGEN_DEVICE_FUNC
-  EIGEN_STRONG_INLINE Packet loadPacket(Index i, Index j) const {
-    return this->load<Packet,AlignmentType>(i,j);
-  }
-
-  template <int AlignmentType>
-  EIGEN_DEVICE_FUNC
-  EIGEN_STRONG_INLINE HalfPacket loadHalfPacket(Index i, Index j) const {
-    // whole method makes column major assumption
-
-    // don't need to add offsets for now (because operator handles that)
-    const Index half_packet_size = unpacket_traits<HalfPacket>::size;
-    if (half_packet_size == packet_size) {
-      return loadPacket<AlignmentType>(i, j);
-    }
-    EIGEN_ALIGN_MAX Scalar data[half_packet_size];
-    for (Index k = 0; k < half_packet_size; k++) {
-      data[k] = operator()(i + k, j);
-    }
-    return pload<HalfPacket>(data);
+  EIGEN_STRONG_INLINE PacketT loadPacket(Index i, Index j) const {
+    return this->load<PacketT,AlignmentType>(i,j);
   }
 };
 
@@ -314,7 +294,8 @@ template<typename Scalar, typename Index, int side,
          typename nocontract_t, typename contract_t,
          bool inner_dim_contiguous,
          bool inner_dim_reordered, int Alignment, template <class> class MakePointer_>
-class BaseTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, 1, inner_dim_contiguous, inner_dim_reordered, Alignment, MakePointer_> : public SimpleTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, 1, inner_dim_contiguous, Alignment, MakePointer_>
+class BaseTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, 1, inner_dim_contiguous, inner_dim_reordered, Alignment, MakePointer_>
+  : public SimpleTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, 1, inner_dim_contiguous, Alignment, MakePointer_>
 {
  public:
   typedef SimpleTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, 1, inner_dim_contiguous, Alignment, MakePointer_> ParentMapper;
@@ -327,12 +308,11 @@ class BaseTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, con
                               const contract_t& k_strides) :
   ParentMapper(tensor, nocontract_strides, ij_strides, contract_strides, k_strides) { }
 
-  typedef typename Tensor::PacketReturnType Packet;
-  template <int> EIGEN_DEVICE_FUNC
-  EIGEN_STRONG_INLINE Packet loadPacket(Index i, Index j) const {
+  template <typename PacketT,int> EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE PacketT loadPacket(Index i, Index j) const {
     EIGEN_ALIGN_MAX Scalar data[1];
     data[0] = this->m_tensor.coeff(this->computeIndex(i, j));
-    return pload<typename Tensor::PacketReturnType>(data);
+    return pload<PacketT>(data);
   }
   template <typename PacketT,int> EIGEN_DEVICE_FUNC
   EIGEN_STRONG_INLINE PacketT load(Index i, Index j) const {
@@ -340,10 +320,6 @@ class BaseTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, con
     data[0] = this->m_tensor.coeff(this->computeIndex(i, j));
     return pload<PacketT>(data);
   }
-  template <int> EIGEN_DEVICE_FUNC
-  EIGEN_STRONG_INLINE Packet loadHalfPacket(Index i, Index j) const {
-    return loadPacket(i, j);
-  }
 };
 
 
@@ -354,8 +330,6 @@ template<typename Scalar, typename Index, int side,
          bool inner_dim_contiguous, bool inner_dim_reordered, int Alignment, template <class> class MakePointer_=MakePointer>
 class TensorContractionSubMapper {
  public:
-  typedef typename Tensor::PacketReturnType Packet;
-  typedef typename unpacket_traits<Packet>::half HalfPacket;
 
   typedef BaseTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, packet_size, inner_dim_contiguous, inner_dim_reordered, Alignment, MakePointer_> ParentMapper;
   typedef TensorContractionSubMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, packet_size, inner_dim_contiguous, inner_dim_reordered, Alignment, MakePointer_> Self;
@@ -390,17 +364,20 @@ class TensorContractionSubMapper {
     return m_base_mapper(i + m_vert_offset, j + m_horiz_offset);
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet loadPacket(Index i) const {
+  template <typename PacketT>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE PacketT loadPacket(Index i) const {
     if (UseDirectOffsets) {
-      return m_base_mapper.template loadPacket<Alignment>(i, 0);
+      return m_base_mapper.template loadPacket<PacketT,Alignment>(i, 0);
     }
-    return m_base_mapper.template loadPacket<Alignment>(i + m_vert_offset, m_horiz_offset);
+    return m_base_mapper.template loadPacket<PacketT,Alignment>(i + m_vert_offset, m_horiz_offset);
   }
-  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet loadPacket(Index i, Index j) const {
+
+  template <typename PacketT>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE PacketT loadPacket(Index i, Index j) const {
     if (UseDirectOffsets) {
-      return m_base_mapper.template loadPacket<Alignment>(i, j);
+      return m_base_mapper.template loadPacket<PacketT,Alignment>(i, j);
     }
-    return m_base_mapper.template loadPacket<Alignment>(i + m_vert_offset, j + m_horiz_offset);
+    return m_base_mapper.template loadPacket<PacketT,Alignment>(i + m_vert_offset, j + m_horiz_offset);
   }
 
   template <typename PacketT, int AlignmentType>
@@ -411,14 +388,8 @@ class TensorContractionSubMapper {
     return m_base_mapper.template loadPacket<PacketT,AlignmentType>(i + m_vert_offset, j + m_horiz_offset);
   }
 
-  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE HalfPacket loadHalfPacket(Index i) const {
-    if (UseDirectOffsets) {
-      return m_base_mapper.template loadHalfPacket<Alignment>(i, 0);
-    }
-    return m_base_mapper.template loadHalfPacket<Alignment>(i + m_vert_offset, m_horiz_offset);
-  }
-
-  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void storePacket(Index i, const Packet& p) const {
+  template <typename PacketT>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void storePacket(Index i, const PacketT& p) const {
     if (UseDirectOffsets) {
       m_base_mapper.storePacket(i, 0, p);
     }
@@ -434,15 +405,15 @@ class TensorContractionSubMapper {
 
   template <typename PacketT, int AlignmentType>
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE PacketT load(Index i) const {
-    EIGEN_STATIC_ASSERT((internal::is_same<PacketT, Packet>::value), YOU_MADE_A_PROGRAMMING_MISTAKE);
+    EIGEN_STATIC_ASSERT((internal::is_same<PacketT, PacketT>::value), YOU_MADE_A_PROGRAMMING_MISTAKE);
     const int ActualAlignment = (AlignmentType == Aligned) && (Alignment == Aligned) ? Aligned : Unaligned;
     if (UseDirectOffsets) {
-     return m_base_mapper.template loadPacket<ActualAlignment>(i, 0);
+     return m_base_mapper.template loadPacket<PacketT,ActualAlignment>(i, 0);
     }
-    return m_base_mapper.template loadPacket<ActualAlignment>(i + m_vert_offset, m_horiz_offset);
+    return m_base_mapper.template loadPacket<PacketT,ActualAlignment>(i + m_vert_offset, m_horiz_offset);
   }
 
-  template <typename Packet>
+  template <typename PacketT>
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool aligned(Index) const {
     return false;
   }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
index 1d145c4b1..0980854b4 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
@@ -244,7 +244,7 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
         RhsMapper;
     typedef internal::gemm_pack_lhs<LhsScalar, Index,
                                     typename LhsMapper::SubMapper, Traits::mr,
-                                    Traits::LhsProgress, ColMajor>
+                                    Traits::LhsProgress, typename Traits::LhsPacket4Packing, ColMajor>
         LhsPacker;
     typedef internal::gemm_pack_rhs<
         RhsScalar, Index, typename RhsMapper::SubMapper, Traits::nr, ColMajor>
diff --git a/unsupported/Eigen/CXX11/src/util/EmulateArray.h b/unsupported/Eigen/CXX11/src/util/EmulateArray.h
index d5c000e08..39c255791 100644
--- a/unsupported/Eigen/CXX11/src/util/EmulateArray.h
+++ b/unsupported/Eigen/CXX11/src/util/EmulateArray.h
@@ -21,9 +21,9 @@ namespace Eigen {
 template <typename T, size_t n> class array {
  public:
   EIGEN_DEVICE_FUNC
-  EIGEN_STRONG_INLINE T& operator[] (size_t index) { return values[index]; }
+  EIGEN_STRONG_INLINE T& operator[] (size_t index) { eigen_internal_assert(index < size()); return values[index]; }
   EIGEN_DEVICE_FUNC
-  EIGEN_STRONG_INLINE const T& operator[] (size_t index) const { return values[index]; }
+  EIGEN_STRONG_INLINE const T& operator[] (size_t index) const { eigen_internal_assert(index < size()); return values[index]; }
 
   EIGEN_DEVICE_FUNC
   EIGEN_STRONG_INLINE T& at(size_t index) { eigen_assert(index < size()); return values[index]; }
diff --git a/unsupported/Eigen/MPRealSupport b/unsupported/Eigen/MPRealSupport
index 6392bea91..c4ea4ec5f 100644
--- a/unsupported/Eigen/MPRealSupport
+++ b/unsupported/Eigen/MPRealSupport
@@ -162,6 +162,7 @@ int main()
       typedef ResScalar LhsPacket;
       typedef ResScalar RhsPacket;
       typedef ResScalar ResPacket;
+      typedef LhsPacket LhsPacket4Packing;
       
     };
 
diff --git a/unsupported/test/cxx11_tensor_shuffling.cpp b/unsupported/test/cxx11_tensor_shuffling.cpp
index 062dd1c0f..2ec85d2d4 100644
--- a/unsupported/test/cxx11_tensor_shuffling.cpp
+++ b/unsupported/test/cxx11_tensor_shuffling.cpp
@@ -83,10 +83,10 @@ static void test_expr_shuffling()
 
   Tensor<float, 4, DataLayout> result(5, 7, 3, 2);
 
-  array<ptrdiff_t, 4> src_slice_dim({2, 3, 1, 7});
-  array<ptrdiff_t, 4> src_slice_start({0, 0, 0, 0});
-  array<ptrdiff_t, 4> dst_slice_dim({1, 7, 3, 2});
-  array<ptrdiff_t, 4> dst_slice_start({0, 0, 0, 0});
+  array<ptrdiff_t, 4> src_slice_dim{{2, 3, 1, 7}};
+  array<ptrdiff_t, 4> src_slice_start{{0, 0, 0, 0}};
+  array<ptrdiff_t, 4> dst_slice_dim{{1, 7, 3, 2}};
+  array<ptrdiff_t, 4> dst_slice_start{{0, 0, 0, 0}};
 
   for (int i = 0; i < 5; ++i) {
     result.slice(dst_slice_start, dst_slice_dim) =