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

24
Eigen/src/Core/util/XprHelper.h
View File

@@ -188,8 +188,32 @@ template<typename T> struct packet_traits;
template<typename T> struct unpacket_traits;
// If we vectorize regardless of alignment, pick the full-sized packet if:
//
// * The size is large enough;
// * Picking it will result in less operations than picking the half size.
// Consider the case where the size is 12, the full packet is 8, and the
// half packet is 4. If we pick the full packet we'd have 1 + 4 operations,
// but only 3 operations if we pick the half-packet.
//
// The reason why we only do this with EIGEN_UNALIGNED_VECTORIZE is that if
// we chose packets which do not divide the data size exactly we're going to
// be left with some possibly unaligned data at the end.
#if EIGEN_UNALIGNED_VECTORIZE
template<int Size, typename PacketType,
bool Stop =
Size==Dynamic ||
(Size >= unpacket_traits<PacketType>::size &&
// If the packet size is 1 we're always good -- it will always divide things perfectly.
// We have this check since otherwise 1/2 would be 0 in the division below.
(unpacket_traits<PacketType>::size == 1 ||
(Size/unpacket_traits<PacketType>::size + Size%unpacket_traits<PacketType>::size) <=
(Size/(unpacket_traits<PacketType>::size/2) + Size%(unpacket_traits<PacketType>::size/2)))) ||
is_same<PacketType,typename unpacket_traits<PacketType>::half>::value>
#else
template<int Size, typename PacketType,
bool Stop = Size==Dynamic || (Size%unpacket_traits<PacketType>::size)==0 || is_same<PacketType,typename unpacket_traits<PacketType>::half>::value>
#endif
struct find_best_packet_helper;
template< int Size, typename PacketType>