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

Re-enable Arm Neon Eigen::half packets of size 8

Browse Source 
- Add predux_half_dowto4
- Remove explicit casts in Half.h to match the behaviour of BFloat16.h
- Enable more packetmath tests for Eigen::half
This commit is contained in:
David Tellenbach
2020-11-18 23:02:21 +00:00
parent 17268b155d
commit 11e4056f6b
3 changed files with 27 additions and 50 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
Eigen/src/Core/arch/NEON/PacketMath.h
View File

@@ -3849,16 +3849,15 @@ template<> EIGEN_STRONG_INLINE Packet2d psqrt(const Packet2d& _x){ return vsqrtq
typedef float16x4_t Packet4hf;
typedef float16x8_t Packet8hf;
// TODO(tellenbach): Enable packets of size 8 as soon as the GEBP can handle them
template <>
struct packet_traits<Eigen::half> : default_packet_traits {
typedef Packet4hf type;
typedef Packet8hf type;
typedef Packet4hf half;
enum {
Vectorizable = 1,
AlignedOnScalar = 1,
size = 4,
HasHalfPacket = 0,
size = 8,
HasHalfPacket = 1,
HasCmp = 1,
HasCast = 1,
@@ -3904,7 +3903,7 @@ struct unpacket_traits<Packet4hf> {
template <>
struct unpacket_traits<Packet8hf> {
typedef Eigen::half type;
typedef Packet8hf half;
typedef Packet4hf half;
enum {
size = 8,
alignment = Aligned16,
@@ -3914,6 +3913,11 @@ struct unpacket_traits<Packet8hf> {
};
};
template<>
EIGEN_DEVICE_FUNC Packet4hf predux_half_dowto4<Packet8hf>(const Packet8hf& a) {
return vadd_f16(vget_low_f16(a), vget_high_f16(a));
}
template <>
EIGEN_STRONG_INLINE Packet8hf pset1<Packet8hf>(const Eigen::half& from) {
return vdupq_n_f16(from.x);
@@ -4418,7 +4422,8 @@ EIGEN_STRONG_INLINE Eigen::half predux_max<Packet4hf>(const Packet4hf& a) {
return h;
}
EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8hf, 4>& kernel) {
EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8hf, 4>& kernel)
{
EIGEN_ALIGN16 Eigen::half in[4][8];
pstore<Eigen::half>(in[0], kernel.packet[0]);
@@ -4432,11 +4437,11 @@ EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8hf, 4>& kernel) {
for (int i = 0; i < 4; ++i) {
EIGEN_UNROLL_LOOP
for (int j = 0; j < 4; ++j) {
out[i][j] = in[j][2*i];
out[i][j] = in[j][2 * i];
}
EIGEN_UNROLL_LOOP
for (int j = 0; j < 4; ++j) {
out[i][j+4] = in[j][2*i+1];
out[i][j + 4] = in[j][2 * i + 1];
}
}