devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Select AVX2 even if the data size is not a multiple of 8

Browse Source
This commit is contained in:
Francesco Mazzoli
2021-11-29 21:13:24 +00:00
committed by Rasmus Munk Larsen
parent eef33946b7
commit eb85b97339
2 changed files with 55 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
test/vectorization_logic.cpp
View File

@@ -258,7 +258,38 @@ struct vectorization_logic
VERIFY(test_redux(VectorX(10),
LinearVectorizedTraversal,NoUnrolling));
// Some static checks for packet-picking -- see
// <https://gitlab.com/libeigen/eigen/merge_requests/46#note_271497656> for context.
// Any multiple of the packet size itself will result in the normal packet
STATIC_CHECK((
internal::is_same<typename internal::find_best_packet<Scalar, PacketSize>::type, PacketType>::value
));
STATIC_CHECK((
internal::is_same<typename internal::find_best_packet<Scalar, PacketSize*2>::type, PacketType>::value
));
STATIC_CHECK((
internal::is_same<typename internal::find_best_packet<Scalar, PacketSize*5>::type, PacketType>::value
));
// Moreover, situations where the size is _not_ a multiple but picking the full packet
// is convenient will also work, but only with unaligned vectorize
STATIC_CHECK((
!(EIGEN_UNALIGNED_VECTORIZE || PacketSize == HalfPacketSize) ||
internal::is_same<typename internal::find_best_packet<Scalar, PacketSize*5+1>::type, PacketType>::value
));
STATIC_CHECK((
!(EIGEN_UNALIGNED_VECTORIZE || PacketSize == HalfPacketSize) ||
internal::is_same<typename internal::find_best_packet<Scalar, PacketSize*5+2>::type, PacketType>::value
));
// In situations where the picking the full-packet would be detrimental the half-packet
// is chosen.
STATIC_CHECK((
!(PacketSize > 2) ||
internal::is_same<typename internal::find_best_packet<Scalar, PacketSize*2-1>::type, HalfPacketType>::value
));
}
};
template<typename Scalar> struct vectorization_logic<Scalar,false>