aboutsummaryrefslogtreecommitdiffhomepage
path: root/Eigen/src/Core/util
diff options
context:
space:
mode:
authorGravatar Rasmus Munk Larsen <rmlarsen@google.com>2020-02-25 01:07:15 +0000
committerGravatar Rasmus Munk Larsen <rmlarsen@google.com>2020-02-25 01:07:15 +0000
commitd7a42eade60ba64821c7879859837a4ea70ba026 (patch)
tree517be95e3ced351acb0feb741adb0824bc92f3ec /Eigen/src/Core/util
parent6ac37768a9f373484d73c2fe1f058ae7d7557059 (diff)
Revert "do not pick full-packet if it'd result in more operations"
Diffstat (limited to 'Eigen/src/Core/util')
-rw-r--r--Eigen/src/Core/util/XprHelper.h18
1 files changed, 3 insertions, 15 deletions
diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h
index cfceb7b18..22f8bdd6c 100644
--- a/Eigen/src/Core/util/XprHelper.h
+++ b/Eigen/src/Core/util/XprHelper.h
@@ -194,12 +194,8 @@ 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.
+// If we vectorize regardless of alignment, always pick the full-sized packet,
+// if we have enough data to process.
//
// Otherwise, pick the packet size we know is going to work with the alignment
// of the given data. See
@@ -207,15 +203,7 @@ template<typename T> struct unpacket_traits
// information.
#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>
+ bool Stop = Size==Dynamic || Size >= unpacket_traits<PacketType>::size || 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>