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

Implemented the pgather/pscatter packet primitives for the arm/NEON architecture

Browse Source
This commit is contained in:
Benoit Steiner
2014-04-17 12:28:01 -07:00
parent d936ddc3d1
commit 6d6df90c9a
2 changed files with 50 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
Eigen/src/Core/arch/NEON/PacketMath.h
View File

@@ -221,6 +221,40 @@ template<> EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet4i& f
template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f32(to, from); }
template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_s32(to, from); }
template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, int stride)
{
Packet4f res;
res = vsetq_lane_f32(from[0*stride], res, 0);
res = vsetq_lane_f32(from[1*stride], res, 1);
res = vsetq_lane_f32(from[2*stride], res, 2);
res = vsetq_lane_f32(from[3*stride], res, 3);
return res;
}
template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* from, int stride)
{
Packet4i res;
res = vsetq_lane_s32(from[0*stride], res, 0);
res = vsetq_lane_s32(from[1*stride], res, 1);
res = vsetq_lane_s32(from[2*stride], res, 2);
res = vsetq_lane_s32(from[3*stride], res, 3);
return res;
}
template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, int stride)
{
to[stride*0] = vgetq_lane_f32(from, 0);
to[stride*1] = vgetq_lane_f32(from, 1);
to[stride*2] = vgetq_lane_f32(from, 2);
to[stride*3] = vgetq_lane_f32(from, 3);
}
template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const Packet4i& from, int stride)
{
to[stride*0] = vgetq_lane_s32(from, 0);
to[stride*1] = vgetq_lane_s32(from, 1);
to[stride*2] = vgetq_lane_s32(from, 2);
to[stride*3] = vgetq_lane_s32(from, 3);
}
template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { EIGEN_ARM_PREFETCH(addr); }
template<> EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) { EIGEN_ARM_PREFETCH(addr); }