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

Update HVX floating-point reduction to support V79 architecture.

Browse Source 
libeigen/eigen!2124
This commit is contained in:
YJ Chang
2026-02-04 16:39:51 +00:00
committed by Antonio Sánchez
parent ddfc68d399
commit c648296368
1 changed files with 8 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
Eigen/src/Core/arch/HVX/PacketMath.h
View File

@@ -764,11 +764,19 @@ EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet32f, 32>& kernel) {
template <HVXPacketSize T>
EIGEN_STRONG_INLINE float predux_hvx(const HVXPacket<T>& a) {
const Index packet_size = unpacket_traits<HVXPacket<T>>::size;
#if __HVX_ARCH__ >= 79
HVX_Vector vsum = Q6_Vsf_vadd_VsfVsf(a.Get(), Q6_V_vror_VR(a.Get(), sizeof(float)));
for (int i = 2; i < packet_size; i <<= 1) {
vsum = Q6_Vsf_vadd_VsfVsf(vsum, Q6_V_vror_VR(vsum, i * sizeof(float)));
}
return pfirst(HVXPacket<T>::Create(vsum));
#else
HVX_Vector vsum = Q6_Vqf32_vadd_VsfVsf(a.Get(), Q6_V_vror_VR(a.Get(), sizeof(float)));
for (int i = 2; i < packet_size; i <<= 1) {
vsum = Q6_Vqf32_vadd_Vqf32Vqf32(vsum, Q6_V_vror_VR(vsum, i * sizeof(float)));
}
return pfirst(HVXPacket<T>::Create(Q6_Vsf_equals_Vqf32(vsum)));
#endif
}
template <>
EIGEN_STRONG_INLINE float predux<Packet32f>(const Packet32f& a) {