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

bug #955 - Implement a rotating kernel alternative in the 3px4 gebp path

Browse Source 
This is substantially faster on ARM, where it's important to minimize the number of loads.

This is specific to the case where all packet types are of size 4. I made my best attempt to minimize how dirty this is... opinions welcome.

Eventually one could have a generic rotated kernel, but it would take some work to get there. Also, on sandy bridge, in my experience, it's not beneficial (even about 1% slower).
This commit is contained in:
Benoit Jacob
2015-02-18 15:03:35 -05:00
parent ee27d50633
commit 9bd8a4bab5
5 changed files with 133 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
Eigen/src/Core/GenericPacketMath.h
View File

@@ -287,6 +287,21 @@ template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Pack
template<typename Packet> EIGEN_DEVICE_FUNC inline Packet preverse(const Packet& a)
{ return a; }
template<size_t offset, typename Packet>
struct protate_impl
{
static Packet run(const Packet& a) { return a; }
};
/** \internal \returns a packet with the coefficients rotated to the right in little-endian convention,
* by the given offset, e.g. for offset == 1:
* (packet[3], packet[2], packet[1], packet[0]) becomes (packet[0], packet[3], packet[2], packet[1])
*/
template<size_t offset, typename Packet> EIGEN_DEVICE_FUNC inline Packet protate(const Packet& a)
{
EIGEN_STATIC_ASSERT(offset < unpacket_traits<Packet>::size, ROTATION_BY_ILLEGAL_OFFSET);
return offset ? protate_impl<offset, Packet>::run(a) : a;
}
/** \internal \returns \a a with real and imaginary part flipped (for complex type only) */
template<typename Packet> EIGEN_DEVICE_FUNC inline Packet pcplxflip(const Packet& a)