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

Fix incorrect NEON native fp16 multiplication.

Browse Source
This commit is contained in:
Antonio Sánchez
2022-12-19 20:46:44 +00:00
committed by Rasmus Munk Larsen
parent dd85d26946
commit bb6675caf7
3 changed files with 40 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
Eigen/src/Core/arch/NEON/GeneralBlockPanelKernel.h
View File

@@ -2,7 +2,7 @@
namespace Eigen {
namespace internal {
#if EIGEN_ARCH_ARM && EIGEN_COMP_CLANG
// Clang seems to excessively spill registers in the GEBP kernel on 32-bit arm.
@@ -218,7 +218,9 @@ struct gebp_traits <half,half,false,false,Architecture::NEON>
EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, RhsPacket& dest) const
{
loadRhs(b,dest);
// If LHS is a Packet8h, we cannot correctly mimic a ploadquad of the RHS
// using a single scalar value.
eigen_assert(false && "Cannot loadRhsQuad for a scalar RHS.");
}
EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<0>&) const

11
Eigen/src/Core/products/GeneralBlockPanelKernel.h
View File

@@ -751,6 +751,9 @@ void loadQuadToDoublePacket(const Scalar* b, DoublePacket<RealPacket>& dest,
template<typename Packet> struct unpacket_traits<DoublePacket<Packet> > {
typedef DoublePacket<typename unpacket_traits<Packet>::half> half;
enum{
size = 2 * unpacket_traits<Packet>::size
};
};
// template<typename Packet>
// DoublePacket<Packet> pmadd(const DoublePacket<Packet> &a, const DoublePacket<Packet> &b)
@@ -2490,7 +2493,13 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
// nr (which is currently 4) for the return type.
const int SResPacketHalfSize = unpacket_traits<typename unpacket_traits<SResPacket>::half>::size;
const int SResPacketQuarterSize = unpacket_traits<typename unpacket_traits<typename unpacket_traits<SResPacket>::half>::half>::size;
if ((SwappedTraits::LhsProgress % 4) == 0 &&
// The following code assumes we can load SRhsPacket in such a way that
// it multiplies blocks of 4 elements in SLhsPacket. This is not the
// case for some customized kernels (i.e. NEON fp16). If the assumption
// fails, drop down to the scalar path.
constexpr bool kCanLoadSRhsQuad = (unpacket_traits<SLhsPacket>::size < 4) || (unpacket_traits<SRhsPacket>::size % (unpacket_traits<SLhsPacket>::size / 4)) == 0;
if (kCanLoadSRhsQuad &&
(SwappedTraits::LhsProgress % 4) == 0 &&
(SwappedTraits::LhsProgress<=16) &&
(SwappedTraits::LhsProgress!=8 || SResPacketHalfSize==nr) &&
(SwappedTraits::LhsProgress!=16 || SResPacketQuarterSize==nr))