GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-08-05 11:40:40 +08:00
Code Issues Packages Projects Releases Wiki Activity

WIP

Browse Source
This commit is contained in:
Everton Constantino 2021-04-22 18:11:53 +00:00
parent 43ce8e9d2d
commit 82a7715b01
2 changed files with 107 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

107
Eigen/src/Core/arch/NEON/Kernels.h
View File

@ -14,6 +14,63 @@ namespace Eigen {
namespace internal { namespace internal {
template<int CPU, typename Scalar, typename ResScalar, typename DataMapper>
struct Accumulator<0, CPU, Scalar, ResScalar, DataMapper, 4, 1>
{
using LinearMapper = typename DataMapper::LinearMapper;
using AccPacket = typename packet_traits<Scalar>::type;
using ResPacket = typename packet_traits<ResScalar>::type;
AccPacket _acc;
EIGEN_STRONG_INLINE void zero()
{
_acc = pset1<AccPacket>(0);
}
template<typename ResPacket_>
EIGEN_STRONG_INLINE void scale(ResScalar alpha, const ResPacket_& pAlpha)
{
_acc *= pAlpha;
}
EIGEN_STRONG_INLINE void store(const DataMapper& dest, Index row, Index col)
{
LinearMapper r0 = dest.getLinearMapper(row, col + 0);
r0.storePacket(0, r0.template loadPacket<ResPacket>(0) + _acc);
}
};
template<int CPU, typename Scalar, typename ResScalar, typename DataMapper>
struct Accumulator<0, CPU, Scalar, ResScalar, DataMapper, 1, 4>
{
using LinearMapper = typename DataMapper::LinearMapper;
using AccPacket = typename packet_traits<Scalar>::type;
using ResPacket = typename packet_traits<ResScalar>::type;
AccPacket _acc;
EIGEN_STRONG_INLINE void zero()
{
_acc = pset1<AccPacket>(0);
}
template<typename ResPacket_>
EIGEN_STRONG_INLINE void scale(ResScalar alpha, const ResPacket_& pAlpha)
{
_acc *= pAlpha;
}
EIGEN_STRONG_INLINE void store(const DataMapper& dest, Index row, Index col)
{
dest(row, col + 0) = _acc[0];
dest(row, col + 1) = _acc[1];
dest(row, col + 2) = _acc[2];
dest(row, col + 3) = _acc[3];
}
};
template<int CPU, typename Scalar, typename ResScalar, typename DataMapper> template<int CPU, typename Scalar, typename ResScalar, typename DataMapper>
struct Accumulator<0, CPU, Scalar, ResScalar, DataMapper, 4, 4> struct Accumulator<0, CPU, Scalar, ResScalar, DataMapper, 4, 4>
{ {
@ -65,7 +122,7 @@ struct MicroKernel<0, CPU, Index, LhsScalar, LhsPackMap, RhsScalar, RhsPackMap,
using LhsPacket = typename packet_traits<LhsScalar>::type; using LhsPacket = typename packet_traits<LhsScalar>::type;
using RhsPacket = typename packet_traits<RhsScalar>::type; using RhsPacket = typename packet_traits<RhsScalar>::type;
asm __volatile__("#BEGIN_NEON_MICROKERNEL_3x1\n\t"); asm __volatile__("#BEGIN_NEON_MICROKERNEL_4x1x4\n\t");
LhsPacket pLhs = pload<LhsPacket>(lhsPackMap.pCur); LhsPacket pLhs = pload<LhsPacket>(lhsPackMap.pCur);
RhsPacket pRhs = pload<RhsPacket>(rhsPackMap.pCur); RhsPacket pRhs = pload<RhsPacket>(rhsPackMap.pCur);
RhsPacket pRhs0 = pset1<RhsPacket>(pRhs[0]); RhsPacket pRhs0 = pset1<RhsPacket>(pRhs[0]);
@ -80,7 +137,53 @@ struct MicroKernel<0, CPU, Index, LhsScalar, LhsPackMap, RhsScalar, RhsPackMap,
lhsPackMap.advance(4*1); lhsPackMap.advance(4*1);
rhsPackMap.advance(1*4); rhsPackMap.advance(1*4);
asm __volatile__("#END_NEON_MICROKERNEL_3x1\n\t"); asm __volatile__("#END_NEON_MICROKERNEL_4x1x4\n\t");
};
};
template<int CPU, typename Index, typename LhsScalar, typename LhsPackMap, typename RhsScalar, typename RhsPackMap, typename AccScalar, typename ResScalar, typename Accumulator>
struct MicroKernel<0, CPU, Index, LhsScalar, LhsPackMap, RhsScalar, RhsPackMap, AccScalar, ResScalar, Accumulator, 4, 1, 1>
{
EIGEN_STRONG_INLINE void operator()(LhsPackMap& lhsPackMap,
RhsPackMap& rhsPackMap,
Index rowIdx, Index colIdx, Index depthIdx,
Accumulator& acc)
{
using LhsPacket = typename packet_traits<LhsScalar>::type;
asm __volatile__("#BEGIN_NEON_MICROKERNEL_4x1x1\n\t");
LhsPacket pLhs = pload<LhsPacket>(lhsPackMap.pCur);
RhsScalar rhs = *rhsPackMap.pCur;
acc._acc += pLhs*rhs;
lhsPackMap.advance(4*1);
rhsPackMap.advance(1);
asm __volatile__("#END_NEON_MICROKERNEL_4x1x1\n\t");
};
};
template<int CPU, typename Index, typename LhsScalar, typename LhsPackMap, typename RhsScalar, typename RhsPackMap, typename AccScalar, typename ResScalar, typename Accumulator>
struct MicroKernel<0, CPU, Index, LhsScalar, LhsPackMap, RhsScalar, RhsPackMap, AccScalar, ResScalar, Accumulator, 1, 1, 4>
{
EIGEN_STRONG_INLINE void operator()(LhsPackMap& lhsPackMap,
RhsPackMap& rhsPackMap,
Index rowIdx, Index colIdx, Index depthIdx,
Accumulator& acc)
{
using RhsPacket = typename packet_traits<RhsScalar>::type;
asm __volatile__("#BEGIN_NEON_MICROKERNEL_1x1x4\n\t");
RhsPacket pRhs = pload<RhsPacket>(rhsPackMap.pCur);
LhsScalar lhs = *lhsPackMap.pCur;
acc._acc += pRhs*lhs;
lhsPackMap.advance(1);
rhsPackMap.advance(4*1);
asm __volatile__("#END_NEON_MICROKERNEL_1x1x4\n\t");
}; };
}; };

2
Eigen/src/Core/arch/NEON/MatrixProduct.h
View File

@ -928,6 +928,7 @@ EIGEN_STRONG_INLINE void gemm(const DataMapper& res, const LhsScalar* blockA, co
} }
std::cout << std::endl; std::cout << std::endl;
#endif #endif
asm __volatile__("#BEGING_GEBP\n\t");
RhsLoopStruct<Architecture, CPU, Index, LhsScalar, LhsPackMap, RhsScalar, RhsPackMap, AccScalar, ResScalar, ResPacket, DataMapper, SHAPES_COUNT<0, 0, LhsScalar, RhsScalar>-1> rhsLS; RhsLoopStruct<Architecture, CPU, Index, LhsScalar, LhsPackMap, RhsScalar, RhsPackMap, AccScalar, ResScalar, ResPacket, DataMapper, SHAPES_COUNT<0, 0, LhsScalar, RhsScalar>-1> rhsLS;
LhsPackMap lhsPackMap(blockA, depth, strideA, offsetA); LhsPackMap lhsPackMap(blockA, depth, strideA, offsetA);
RhsPackMap rhsPackMap(blockB, depth, strideB, offsetB); RhsPackMap rhsPackMap(blockB, depth, strideB, offsetB);
@ -935,6 +936,7 @@ EIGEN_STRONG_INLINE void gemm(const DataMapper& res, const LhsScalar* blockA, co
ResPacket pAlpha = pset1<ResPacket>(alpha); ResPacket pAlpha = pset1<ResPacket>(alpha);
rhsLS(0, res, rows, depth, cols, alpha, pAlpha, lhsPackMap, rhsPackMap); rhsLS(0, res, rows, depth, cols, alpha, pAlpha, lhsPackMap, rhsPackMap);
asm __volatile__("#END_GEBP\n\t");
} }
template<typename Index, typename DataMapper, int nr, bool Conjugate, bool PanelMode> template<typename Index, typename DataMapper, int nr, bool Conjugate, bool PanelMode>