mirror of
https://gitlab.com/libeigen/eigen.git
synced 2025-09-13 18:03:13 +08:00
Generalized the matrix vector product code.
This commit is contained in:
Benoit Steiner
parent
7f2c6ed2fa
commit
2dde63499c
@ -457,11 +457,13 @@ template<> struct gemv_selector<OnTheRight,ColMajor,true>
|
||||
MappedDest(actualDestPtr, dest.size()) = dest;
|
||||
}
|
||||
|
||||
typedef const_blas_data_mapper<LhsScalar,Index,ColMajor> LhsMapper;
|
||||
typedef const_blas_data_mapper<RhsScalar,Index,RowMajor> RhsMapper;
|
||||
general_matrix_vector_product
|
||||
<Index,LhsScalar,ColMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsBlasTraits::NeedToConjugate>::run(
|
||||
<Index,LhsScalar,LhsMapper,ColMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsMapper,RhsBlasTraits::NeedToConjugate>::run(
|
||||
actualLhs.rows(), actualLhs.cols(),
|
||||
actualLhs.data(), actualLhs.outerStride(),
|
||||
actualRhs.data(), actualRhs.innerStride(),
|
||||
LhsMapper(actualLhs.data(), actualLhs.outerStride()),
|
||||
RhsMapper(actualRhs.data(), actualRhs.innerStride()),
|
||||
actualDestPtr, 1,
|
||||
compatibleAlpha);
|
||||
|
||||
@ -516,11 +518,13 @@ template<> struct gemv_selector<OnTheRight,RowMajor,true>
|
||||
Map<typename _ActualRhsType::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs;
|
||||
}
|
||||
|
||||
typedef const_blas_data_mapper<LhsScalar,Index,RowMajor> LhsMapper;
|
||||
typedef const_blas_data_mapper<RhsScalar,Index,ColMajor> RhsMapper;
|
||||
general_matrix_vector_product
|
||||
<Index,LhsScalar,RowMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsBlasTraits::NeedToConjugate>::run(
|
||||
<Index,LhsScalar,LhsMapper,RowMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsMapper,RhsBlasTraits::NeedToConjugate>::run(
|
||||
actualLhs.rows(), actualLhs.cols(),
|
||||
actualLhs.data(), actualLhs.outerStride(),
|
||||
actualRhsPtr, 1,
|
||||
LhsMapper(actualLhs.data(), actualLhs.outerStride()),
|
||||
RhsMapper(actualRhsPtr, 1),
|
||||
dest.data(), dest.innerStride(),
|
||||
actualAlpha);
|
||||
}
|
||||
|
||||
@ -55,10 +55,10 @@ namespace internal {
|
||||
* compared to unaligned loads on a 4 byte boundary.
|
||||
*
|
||||
*/
|
||||
template<typename Index, typename LhsScalar, bool ConjugateLhs, typename RhsScalar, bool ConjugateRhs, int Version>
|
||||
struct general_matrix_vector_product<Index,LhsScalar,ColMajor,ConjugateLhs,RhsScalar,ConjugateRhs,Version>
|
||||
template<typename Index, typename LhsScalar, typename LhsMapper, bool ConjugateLhs, typename RhsScalar, typename RhsMapper, bool ConjugateRhs, int Version>
|
||||
struct general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,ConjugateLhs,RhsScalar,RhsMapper,ConjugateRhs,Version>
|
||||
{
|
||||
typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar;
|
||||
typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar;
|
||||
|
||||
enum {
|
||||
Vectorizable = packet_traits<LhsScalar>::Vectorizable && packet_traits<RhsScalar>::Vectorizable
|
||||
@ -78,17 +78,17 @@ typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket;
|
||||
|
||||
EIGEN_DONT_INLINE static void run(
|
||||
Index rows, Index cols,
|
||||
const LhsScalar* lhs, Index lhsStride,
|
||||
const RhsScalar* rhs, Index rhsIncr,
|
||||
const LhsMapper& lhs,
|
||||
const RhsMapper& rhs,
|
||||
ResScalar* res, Index resIncr,
|
||||
RhsScalar alpha);
|
||||
};
|
||||
|
||||
template<typename Index, typename LhsScalar, bool ConjugateLhs, typename RhsScalar, bool ConjugateRhs, int Version>
|
||||
EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,ColMajor,ConjugateLhs,RhsScalar,ConjugateRhs,Version>::run(
|
||||
template<typename Index, typename LhsScalar, typename LhsMapper, bool ConjugateLhs, typename RhsScalar, typename RhsMapper, bool ConjugateRhs, int Version>
|
||||
EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,ConjugateLhs,RhsScalar,RhsMapper,ConjugateRhs,Version>::run(
|
||||
Index rows, Index cols,
|
||||
const LhsScalar* lhs, Index lhsStride,
|
||||
const RhsScalar* rhs, Index rhsIncr,
|
||||
const LhsMapper& lhs,
|
||||
const RhsMapper& rhs,
|
||||
ResScalar* res, Index resIncr,
|
||||
RhsScalar alpha)
|
||||
{
|
||||
@ -97,14 +97,16 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,ColMajor,Co
|
||||
#ifdef _EIGEN_ACCUMULATE_PACKETS
|
||||
#error _EIGEN_ACCUMULATE_PACKETS has already been defined
|
||||
#endif
|
||||
#define _EIGEN_ACCUMULATE_PACKETS(A0,A13,A2) \
|
||||
#define _EIGEN_ACCUMULATE_PACKETS(Alignment0,Alignment13,Alignment2) \
|
||||
pstore(&res[j], \
|
||||
padd(pload<ResPacket>(&res[j]), \
|
||||
padd( \
|
||||
padd(pcj.pmul(EIGEN_CAT(ploa , A0)<LhsPacket>(&lhs0[j]), ptmp0), \
|
||||
pcj.pmul(EIGEN_CAT(ploa , A13)<LhsPacket>(&lhs1[j]), ptmp1)), \
|
||||
padd(pcj.pmul(EIGEN_CAT(ploa , A2)<LhsPacket>(&lhs2[j]), ptmp2), \
|
||||
pcj.pmul(EIGEN_CAT(ploa , A13)<LhsPacket>(&lhs3[j]), ptmp3)) )))
|
||||
padd(pcj.pmul(lhs0.template load<LhsPacket, Alignment0>(j), ptmp0), \
|
||||
pcj.pmul(lhs1.template load<LhsPacket, Alignment13>(j), ptmp1)), \
|
||||
padd(pcj.pmul(lhs2.template load<LhsPacket, Alignment2>(j), ptmp2), \
|
||||
pcj.pmul(lhs3.template load<LhsPacket, Alignment13>(j), ptmp3)) )))
|
||||
|
||||
typedef typename LhsMapper::VectorMapper LhsScalars;
|
||||
|
||||
conj_helper<LhsScalar,RhsScalar,ConjugateLhs,ConjugateRhs> cj;
|
||||
conj_helper<LhsPacket,RhsPacket,ConjugateLhs,ConjugateRhs> pcj;
|
||||
@ -119,6 +121,8 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,ColMajor,Co
|
||||
// const Index PeelAlignedMask = ResPacketSize*peels-1;
|
||||
const Index size = rows;
|
||||
|
||||
const Index lhsStride = lhs.stride();
|
||||
|
||||
// How many coeffs of the result do we have to skip to be aligned.
|
||||
// Here we assume data are at least aligned on the base scalar type.
|
||||
Index alignedStart = internal::first_aligned(res,size);
|
||||
@ -131,12 +135,12 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,ColMajor,Co
|
||||
: FirstAligned;
|
||||
|
||||
// we cannot assume the first element is aligned because of sub-matrices
|
||||
const Index lhsAlignmentOffset = internal::first_aligned(lhs,size);
|
||||
const Index lhsAlignmentOffset = lhs.firstAligned(size);
|
||||
|
||||
// find how many columns do we have to skip to be aligned with the result (if possible)
|
||||
Index skipColumns = 0;
|
||||
// if the data cannot be aligned (TODO add some compile time tests when possible, e.g. for floats)
|
||||
if( (size_t(lhs)%sizeof(LhsScalar)) || (size_t(res)%sizeof(ResScalar)) )
|
||||
if( (lhsAlignmentOffset < 0) || (size_t(res)%sizeof(ResScalar)) )
|
||||
{
|
||||
alignedSize = 0;
|
||||
alignedStart = 0;
|
||||
@ -149,7 +153,7 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,ColMajor,Co
|
||||
}
|
||||
else if (LhsPacketSize>1)
|
||||
{
|
||||
eigen_internal_assert(size_t(lhs+lhsAlignmentOffset)%sizeof(LhsPacket)==0 || size<LhsPacketSize);
|
||||
// eigen_internal_assert(size_t(firstLhs+lhsAlignmentOffset)%sizeof(LhsPacket)==0 || size<LhsPacketSize);
|
||||
|
||||
while (skipColumns<LhsPacketSize &&
|
||||
alignedStart != ((lhsAlignmentOffset + alignmentStep*skipColumns)%LhsPacketSize))
|
||||
@ -166,10 +170,10 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,ColMajor,Co
|
||||
// note that the skiped columns are processed later.
|
||||
}
|
||||
|
||||
eigen_internal_assert( (alignmentPattern==NoneAligned)
|
||||
/* eigen_internal_assert( (alignmentPattern==NoneAligned)
|
||||
|| (skipColumns + columnsAtOnce >= cols)
|
||||
|| LhsPacketSize > size
|
||||
|| (size_t(lhs+alignedStart+lhsStride*skipColumns)%sizeof(LhsPacket))==0);
|
||||
|| (size_t(firstLhs+alignedStart+lhsStride*skipColumns)%sizeof(LhsPacket))==0);*/
|
||||
}
|
||||
else if(Vectorizable)
|
||||
{
|
||||
@ -178,20 +182,20 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,ColMajor,Co
|
||||
alignmentPattern = AllAligned;
|
||||
}
|
||||
|
||||
Index offset1 = (FirstAligned && alignmentStep==1?3:1);
|
||||
Index offset3 = (FirstAligned && alignmentStep==1?1:3);
|
||||
const Index offset1 = (FirstAligned && alignmentStep==1?3:1);
|
||||
const Index offset3 = (FirstAligned && alignmentStep==1?1:3);
|
||||
|
||||
Index columnBound = ((cols-skipColumns)/columnsAtOnce)*columnsAtOnce + skipColumns;
|
||||
for (Index i=skipColumns; i<columnBound; i+=columnsAtOnce)
|
||||
{
|
||||
RhsPacket ptmp0 = pset1<RhsPacket>(alpha*rhs[i*rhsIncr]),
|
||||
ptmp1 = pset1<RhsPacket>(alpha*rhs[(i+offset1)*rhsIncr]),
|
||||
ptmp2 = pset1<RhsPacket>(alpha*rhs[(i+2)*rhsIncr]),
|
||||
ptmp3 = pset1<RhsPacket>(alpha*rhs[(i+offset3)*rhsIncr]);
|
||||
RhsPacket ptmp0 = pset1<RhsPacket>(alpha*rhs(i, 0)),
|
||||
ptmp1 = pset1<RhsPacket>(alpha*rhs(i+offset1, 0)),
|
||||
ptmp2 = pset1<RhsPacket>(alpha*rhs(i+2, 0)),
|
||||
ptmp3 = pset1<RhsPacket>(alpha*rhs(i+offset3, 0));
|
||||
|
||||
// this helps a lot generating better binary code
|
||||
const LhsScalar *lhs0 = lhs + i*lhsStride, *lhs1 = lhs + (i+offset1)*lhsStride,
|
||||
*lhs2 = lhs + (i+2)*lhsStride, *lhs3 = lhs + (i+offset3)*lhsStride;
|
||||
const LhsScalars lhs0 = lhs.getVectorMapper(0, i+0), lhs1 = lhs.getVectorMapper(0, i+offset1),
|
||||
lhs2 = lhs.getVectorMapper(0, i+2), lhs3 = lhs.getVectorMapper(0, i+offset3);
|
||||
|
||||
if (Vectorizable)
|
||||
{
|
||||
@ -199,10 +203,10 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,ColMajor,Co
|
||||
// process initial unaligned coeffs
|
||||
for (Index j=0; j<alignedStart; ++j)
|
||||
{
|
||||
res[j] = cj.pmadd(lhs0[j], pfirst(ptmp0), res[j]);
|
||||
res[j] = cj.pmadd(lhs1[j], pfirst(ptmp1), res[j]);
|
||||
res[j] = cj.pmadd(lhs2[j], pfirst(ptmp2), res[j]);
|
||||
res[j] = cj.pmadd(lhs3[j], pfirst(ptmp3), res[j]);
|
||||
res[j] = cj.pmadd(lhs0(j), pfirst(ptmp0), res[j]);
|
||||
res[j] = cj.pmadd(lhs1(j), pfirst(ptmp1), res[j]);
|
||||
res[j] = cj.pmadd(lhs2(j), pfirst(ptmp2), res[j]);
|
||||
res[j] = cj.pmadd(lhs3(j), pfirst(ptmp3), res[j]);
|
||||
}
|
||||
|
||||
if (alignedSize>alignedStart)
|
||||
@ -211,11 +215,11 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,ColMajor,Co
|
||||
{
|
||||
case AllAligned:
|
||||
for (Index j = alignedStart; j<alignedSize; j+=ResPacketSize)
|
||||
_EIGEN_ACCUMULATE_PACKETS(d,d,d);
|
||||
_EIGEN_ACCUMULATE_PACKETS(Aligned,Aligned,Aligned);
|
||||
break;
|
||||
case EvenAligned:
|
||||
for (Index j = alignedStart; j<alignedSize; j+=ResPacketSize)
|
||||
_EIGEN_ACCUMULATE_PACKETS(d,du,d);
|
||||
_EIGEN_ACCUMULATE_PACKETS(Aligned,Unaligned,Aligned);
|
||||
break;
|
||||
case FirstAligned:
|
||||
{
|
||||
@ -225,28 +229,28 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,ColMajor,Co
|
||||
LhsPacket A00, A01, A02, A03, A10, A11, A12, A13;
|
||||
ResPacket T0, T1;
|
||||
|
||||
A01 = pload<LhsPacket>(&lhs1[alignedStart-1]);
|
||||
A02 = pload<LhsPacket>(&lhs2[alignedStart-2]);
|
||||
A03 = pload<LhsPacket>(&lhs3[alignedStart-3]);
|
||||
A01 = lhs1.template load<LhsPacket, Aligned>(alignedStart-1);
|
||||
A02 = lhs2.template load<LhsPacket, Aligned>(alignedStart-2);
|
||||
A03 = lhs3.template load<LhsPacket, Aligned>(alignedStart-3);
|
||||
|
||||
for (; j<peeledSize; j+=peels*ResPacketSize)
|
||||
{
|
||||
A11 = pload<LhsPacket>(&lhs1[j-1+LhsPacketSize]); palign<1>(A01,A11);
|
||||
A12 = pload<LhsPacket>(&lhs2[j-2+LhsPacketSize]); palign<2>(A02,A12);
|
||||
A13 = pload<LhsPacket>(&lhs3[j-3+LhsPacketSize]); palign<3>(A03,A13);
|
||||
A11 = lhs1.template load<LhsPacket, Aligned>(j-1+LhsPacketSize); palign<1>(A01,A11);
|
||||
A12 = lhs2.template load<LhsPacket, Aligned>(j-2+LhsPacketSize); palign<2>(A02,A12);
|
||||
A13 = lhs3.template load<LhsPacket, Aligned>(j-3+LhsPacketSize); palign<3>(A03,A13);
|
||||
|
||||
A00 = pload<LhsPacket>(&lhs0[j]);
|
||||
A10 = pload<LhsPacket>(&lhs0[j+LhsPacketSize]);
|
||||
A00 = lhs0.template load<LhsPacket, Aligned>(j);
|
||||
A10 = lhs0.template load<LhsPacket, Aligned>(j+LhsPacketSize);
|
||||
T0 = pcj.pmadd(A00, ptmp0, pload<ResPacket>(&res[j]));
|
||||
T1 = pcj.pmadd(A10, ptmp0, pload<ResPacket>(&res[j+ResPacketSize]));
|
||||
|
||||
T0 = pcj.pmadd(A01, ptmp1, T0);
|
||||
A01 = pload<LhsPacket>(&lhs1[j-1+2*LhsPacketSize]); palign<1>(A11,A01);
|
||||
A01 = lhs1.template load<LhsPacket, Aligned>(j-1+2*LhsPacketSize); palign<1>(A11,A01);
|
||||
T0 = pcj.pmadd(A02, ptmp2, T0);
|
||||
A02 = pload<LhsPacket>(&lhs2[j-2+2*LhsPacketSize]); palign<2>(A12,A02);
|
||||
A02 = lhs2.template load<LhsPacket, Aligned>(j-2+2*LhsPacketSize); palign<2>(A12,A02);
|
||||
T0 = pcj.pmadd(A03, ptmp3, T0);
|
||||
pstore(&res[j],T0);
|
||||
A03 = pload<LhsPacket>(&lhs3[j-3+2*LhsPacketSize]); palign<3>(A13,A03);
|
||||
A03 = lhs3.template load<LhsPacket, Aligned>(j-3+2*LhsPacketSize); palign<3>(A13,A03);
|
||||
T1 = pcj.pmadd(A11, ptmp1, T1);
|
||||
T1 = pcj.pmadd(A12, ptmp2, T1);
|
||||
T1 = pcj.pmadd(A13, ptmp3, T1);
|
||||
@ -254,12 +258,12 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,ColMajor,Co
|
||||
}
|
||||
}
|
||||
for (; j<alignedSize; j+=ResPacketSize)
|
||||
_EIGEN_ACCUMULATE_PACKETS(d,du,du);
|
||||
_EIGEN_ACCUMULATE_PACKETS(Aligned,Unaligned,Unaligned);
|
||||
break;
|
||||
}
|
||||
default:
|
||||
for (Index j = alignedStart; j<alignedSize; j+=ResPacketSize)
|
||||
_EIGEN_ACCUMULATE_PACKETS(du,du,du);
|
||||
_EIGEN_ACCUMULATE_PACKETS(Unaligned,Unaligned,Unaligned);
|
||||
break;
|
||||
}
|
||||
}
|
||||
@ -268,10 +272,10 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,ColMajor,Co
|
||||
/* process remaining coeffs (or all if there is no explicit vectorization) */
|
||||
for (Index j=alignedSize; j<size; ++j)
|
||||
{
|
||||
res[j] = cj.pmadd(lhs0[j], pfirst(ptmp0), res[j]);
|
||||
res[j] = cj.pmadd(lhs1[j], pfirst(ptmp1), res[j]);
|
||||
res[j] = cj.pmadd(lhs2[j], pfirst(ptmp2), res[j]);
|
||||
res[j] = cj.pmadd(lhs3[j], pfirst(ptmp3), res[j]);
|
||||
res[j] = cj.pmadd(lhs0(j), pfirst(ptmp0), res[j]);
|
||||
res[j] = cj.pmadd(lhs1(j), pfirst(ptmp1), res[j]);
|
||||
res[j] = cj.pmadd(lhs2(j), pfirst(ptmp2), res[j]);
|
||||
res[j] = cj.pmadd(lhs3(j), pfirst(ptmp3), res[j]);
|
||||
}
|
||||
}
|
||||
|
||||
@ -282,27 +286,27 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,ColMajor,Co
|
||||
{
|
||||
for (Index k=start; k<end; ++k)
|
||||
{
|
||||
RhsPacket ptmp0 = pset1<RhsPacket>(alpha*rhs[k*rhsIncr]);
|
||||
const LhsScalar* lhs0 = lhs + k*lhsStride;
|
||||
RhsPacket ptmp0 = pset1<RhsPacket>(alpha*rhs(k, 0));
|
||||
const LhsScalars lhs0 = lhs.getVectorMapper(0, k);
|
||||
|
||||
if (Vectorizable)
|
||||
{
|
||||
/* explicit vectorization */
|
||||
// process first unaligned result's coeffs
|
||||
for (Index j=0; j<alignedStart; ++j)
|
||||
res[j] += cj.pmul(lhs0[j], pfirst(ptmp0));
|
||||
res[j] += cj.pmul(lhs0(j), pfirst(ptmp0));
|
||||
// process aligned result's coeffs
|
||||
if ((size_t(lhs0+alignedStart)%sizeof(LhsPacket))==0)
|
||||
if (lhs0.template aligned<LhsPacket>(alignedStart))
|
||||
for (Index i = alignedStart;i<alignedSize;i+=ResPacketSize)
|
||||
pstore(&res[i], pcj.pmadd(pload<LhsPacket>(&lhs0[i]), ptmp0, pload<ResPacket>(&res[i])));
|
||||
pstore(&res[i], pcj.pmadd(lhs0.template load<LhsPacket, Aligned>(i), ptmp0, pload<ResPacket>(&res[i])));
|
||||
else
|
||||
for (Index i = alignedStart;i<alignedSize;i+=ResPacketSize)
|
||||
pstore(&res[i], pcj.pmadd(ploadu<LhsPacket>(&lhs0[i]), ptmp0, pload<ResPacket>(&res[i])));
|
||||
pstore(&res[i], pcj.pmadd(lhs0.template load<LhsPacket, Unaligned>(i), ptmp0, pload<ResPacket>(&res[i])));
|
||||
}
|
||||
|
||||
// process remaining scalars (or all if no explicit vectorization)
|
||||
for (Index i=alignedSize; i<size; ++i)
|
||||
res[i] += cj.pmul(lhs0[i], pfirst(ptmp0));
|
||||
res[i] += cj.pmul(lhs0(i), pfirst(ptmp0));
|
||||
}
|
||||
if (skipColumns)
|
||||
{
|
||||
@ -326,8 +330,8 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,ColMajor,Co
|
||||
* - alpha is always a complex (or converted to a complex)
|
||||
* - no vectorization
|
||||
*/
|
||||
template<typename Index, typename LhsScalar, bool ConjugateLhs, typename RhsScalar, bool ConjugateRhs, int Version>
|
||||
struct general_matrix_vector_product<Index,LhsScalar,RowMajor,ConjugateLhs,RhsScalar,ConjugateRhs,Version>
|
||||
template<typename Index, typename LhsScalar, typename LhsMapper, bool ConjugateLhs, typename RhsScalar, typename RhsMapper, bool ConjugateRhs, int Version>
|
||||
struct general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,ConjugateLhs,RhsScalar,RhsMapper,ConjugateRhs,Version>
|
||||
{
|
||||
typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar;
|
||||
|
||||
@ -349,64 +353,66 @@ typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket;
|
||||
|
||||
EIGEN_DONT_INLINE static void run(
|
||||
Index rows, Index cols,
|
||||
const LhsScalar* lhs, Index lhsStride,
|
||||
const RhsScalar* rhs, Index rhsIncr,
|
||||
const LhsMapper& lhs,
|
||||
const RhsMapper& rhs,
|
||||
ResScalar* res, Index resIncr,
|
||||
ResScalar alpha);
|
||||
};
|
||||
|
||||
template<typename Index, typename LhsScalar, bool ConjugateLhs, typename RhsScalar, bool ConjugateRhs, int Version>
|
||||
EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,RowMajor,ConjugateLhs,RhsScalar,ConjugateRhs,Version>::run(
|
||||
template<typename Index, typename LhsScalar, typename LhsMapper, bool ConjugateLhs, typename RhsScalar, typename RhsMapper, bool ConjugateRhs, int Version>
|
||||
EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,ConjugateLhs,RhsScalar,RhsMapper,ConjugateRhs,Version>::run(
|
||||
Index rows, Index cols,
|
||||
const LhsScalar* lhs, Index lhsStride,
|
||||
const RhsScalar* rhs, Index rhsIncr,
|
||||
const LhsMapper& lhs,
|
||||
const RhsMapper& rhs,
|
||||
ResScalar* res, Index resIncr,
|
||||
ResScalar alpha)
|
||||
{
|
||||
EIGEN_UNUSED_VARIABLE(rhsIncr);
|
||||
eigen_internal_assert(rhsIncr==1);
|
||||
eigen_internal_assert(rhs.stride()==1);
|
||||
|
||||
#ifdef _EIGEN_ACCUMULATE_PACKETS
|
||||
#error _EIGEN_ACCUMULATE_PACKETS has already been defined
|
||||
#endif
|
||||
|
||||
#define _EIGEN_ACCUMULATE_PACKETS(A0,A13,A2) {\
|
||||
RhsPacket b = pload<RhsPacket>(&rhs[j]); \
|
||||
ptmp0 = pcj.pmadd(EIGEN_CAT(ploa,A0) <LhsPacket>(&lhs0[j]), b, ptmp0); \
|
||||
ptmp1 = pcj.pmadd(EIGEN_CAT(ploa,A13)<LhsPacket>(&lhs1[j]), b, ptmp1); \
|
||||
ptmp2 = pcj.pmadd(EIGEN_CAT(ploa,A2) <LhsPacket>(&lhs2[j]), b, ptmp2); \
|
||||
ptmp3 = pcj.pmadd(EIGEN_CAT(ploa,A13)<LhsPacket>(&lhs3[j]), b, ptmp3); }
|
||||
#define _EIGEN_ACCUMULATE_PACKETS(Alignment0,Alignment13,Alignment2) {\
|
||||
RhsPacket b = rhs.getVectorMapper(j, 0).template load<RhsPacket, Aligned>(0); \
|
||||
ptmp0 = pcj.pmadd(lhs0.template load<LhsPacket, Alignment0>(j), b, ptmp0); \
|
||||
ptmp1 = pcj.pmadd(lhs1.template load<LhsPacket, Alignment13>(j), b, ptmp1); \
|
||||
ptmp2 = pcj.pmadd(lhs2.template load<LhsPacket, Alignment2>(j), b, ptmp2); \
|
||||
ptmp3 = pcj.pmadd(lhs3.template load<LhsPacket, Alignment13>(j), b, ptmp3); }
|
||||
|
||||
conj_helper<LhsScalar,RhsScalar,ConjugateLhs,ConjugateRhs> cj;
|
||||
conj_helper<LhsPacket,RhsPacket,ConjugateLhs,ConjugateRhs> pcj;
|
||||
|
||||
typedef typename LhsMapper::VectorMapper LhsScalars;
|
||||
|
||||
enum { AllAligned=0, EvenAligned=1, FirstAligned=2, NoneAligned=3 };
|
||||
const Index rowsAtOnce = 4;
|
||||
const Index peels = 2;
|
||||
const Index RhsPacketAlignedMask = RhsPacketSize-1;
|
||||
const Index LhsPacketAlignedMask = LhsPacketSize-1;
|
||||
// const Index PeelAlignedMask = RhsPacketSize*peels-1;
|
||||
const Index depth = cols;
|
||||
const Index lhsStride = lhs.stride();
|
||||
|
||||
// How many coeffs of the result do we have to skip to be aligned.
|
||||
// Here we assume data are at least aligned on the base scalar type
|
||||
// if that's not the case then vectorization is discarded, see below.
|
||||
Index alignedStart = internal::first_aligned(rhs, depth);
|
||||
Index alignedStart = rhs.firstAligned(depth);
|
||||
Index alignedSize = RhsPacketSize>1 ? alignedStart + ((depth-alignedStart) & ~RhsPacketAlignedMask) : 0;
|
||||
const Index peeledSize = alignedSize - RhsPacketSize*peels - RhsPacketSize + 1;
|
||||
|
||||
const Index alignmentStep = LhsPacketSize>1 ? (LhsPacketSize - lhsStride % LhsPacketSize) & LhsPacketAlignedMask : 0;
|
||||
Index alignmentPattern = alignmentStep==0 ? AllAligned
|
||||
: alignmentStep==(LhsPacketSize/2) ? EvenAligned
|
||||
: FirstAligned;
|
||||
: alignmentStep==(LhsPacketSize/2) ? EvenAligned
|
||||
: FirstAligned;
|
||||
|
||||
// we cannot assume the first element is aligned because of sub-matrices
|
||||
const Index lhsAlignmentOffset = internal::first_aligned(lhs,depth);
|
||||
const Index lhsAlignmentOffset = lhs.firstAligned(depth);
|
||||
const Index rhsAlignmentOffset = rhs.firstAligned(rows);
|
||||
|
||||
// find how many rows do we have to skip to be aligned with rhs (if possible)
|
||||
Index skipRows = 0;
|
||||
// if the data cannot be aligned (TODO add some compile time tests when possible, e.g. for floats)
|
||||
if( (sizeof(LhsScalar)!=sizeof(RhsScalar)) || (size_t(lhs)%sizeof(LhsScalar)) || (size_t(rhs)%sizeof(RhsScalar)) )
|
||||
if( (sizeof(LhsScalar)!=sizeof(RhsScalar)) || (lhsAlignmentOffset < 0) || (rhsAlignmentOffset < 0) )
|
||||
{
|
||||
alignedSize = 0;
|
||||
alignedStart = 0;
|
||||
@ -418,7 +424,7 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,RowMajor,Co
|
||||
}
|
||||
else if (LhsPacketSize>1)
|
||||
{
|
||||
eigen_internal_assert(size_t(lhs+lhsAlignmentOffset)%sizeof(LhsPacket)==0 || depth<LhsPacketSize);
|
||||
// eigen_internal_assert(size_t(firstLhs+lhsAlignmentOffset)%sizeof(LhsPacket)==0 || depth<LhsPacketSize);
|
||||
|
||||
while (skipRows<LhsPacketSize &&
|
||||
alignedStart != ((lhsAlignmentOffset + alignmentStep*skipRows)%LhsPacketSize))
|
||||
@ -434,11 +440,11 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,RowMajor,Co
|
||||
skipRows = (std::min)(skipRows,Index(rows));
|
||||
// note that the skiped columns are processed later.
|
||||
}
|
||||
eigen_internal_assert( alignmentPattern==NoneAligned
|
||||
/* eigen_internal_assert( alignmentPattern==NoneAligned
|
||||
|| LhsPacketSize==1
|
||||
|| (skipRows + rowsAtOnce >= rows)
|
||||
|| LhsPacketSize > depth
|
||||
|| (size_t(lhs+alignedStart+lhsStride*skipRows)%sizeof(LhsPacket))==0);
|
||||
|| (size_t(firstLhs+alignedStart+lhsStride*skipRows)%sizeof(LhsPacket))==0);*/
|
||||
}
|
||||
else if(Vectorizable)
|
||||
{
|
||||
@ -447,8 +453,8 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,RowMajor,Co
|
||||
alignmentPattern = AllAligned;
|
||||
}
|
||||
|
||||
Index offset1 = (FirstAligned && alignmentStep==1?3:1);
|
||||
Index offset3 = (FirstAligned && alignmentStep==1?1:3);
|
||||
const Index offset1 = (FirstAligned && alignmentStep==1?3:1);
|
||||
const Index offset3 = (FirstAligned && alignmentStep==1?1:3);
|
||||
|
||||
Index rowBound = ((rows-skipRows)/rowsAtOnce)*rowsAtOnce + skipRows;
|
||||
for (Index i=skipRows; i<rowBound; i+=rowsAtOnce)
|
||||
@ -457,8 +463,8 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,RowMajor,Co
|
||||
ResScalar tmp1 = ResScalar(0), tmp2 = ResScalar(0), tmp3 = ResScalar(0);
|
||||
|
||||
// this helps the compiler generating good binary code
|
||||
const LhsScalar *lhs0 = lhs + i*lhsStride, *lhs1 = lhs + (i+offset1)*lhsStride,
|
||||
*lhs2 = lhs + (i+2)*lhsStride, *lhs3 = lhs + (i+offset3)*lhsStride;
|
||||
const LhsScalars lhs0 = lhs.getVectorMapper(i+0, 0), lhs1 = lhs.getVectorMapper(i+offset1, 0),
|
||||
lhs2 = lhs.getVectorMapper(i+2, 0), lhs3 = lhs.getVectorMapper(i+offset3, 0);
|
||||
|
||||
if (Vectorizable)
|
||||
{
|
||||
@ -470,9 +476,9 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,RowMajor,Co
|
||||
// FIXME this loop get vectorized by the compiler !
|
||||
for (Index j=0; j<alignedStart; ++j)
|
||||
{
|
||||
RhsScalar b = rhs[j];
|
||||
tmp0 += cj.pmul(lhs0[j],b); tmp1 += cj.pmul(lhs1[j],b);
|
||||
tmp2 += cj.pmul(lhs2[j],b); tmp3 += cj.pmul(lhs3[j],b);
|
||||
RhsScalar b = rhs(j, 0);
|
||||
tmp0 += cj.pmul(lhs0(j),b); tmp1 += cj.pmul(lhs1(j),b);
|
||||
tmp2 += cj.pmul(lhs2(j),b); tmp3 += cj.pmul(lhs3(j),b);
|
||||
}
|
||||
|
||||
if (alignedSize>alignedStart)
|
||||
@ -481,11 +487,11 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,RowMajor,Co
|
||||
{
|
||||
case AllAligned:
|
||||
for (Index j = alignedStart; j<alignedSize; j+=RhsPacketSize)
|
||||
_EIGEN_ACCUMULATE_PACKETS(d,d,d);
|
||||
_EIGEN_ACCUMULATE_PACKETS(Aligned,Aligned,Aligned);
|
||||
break;
|
||||
case EvenAligned:
|
||||
for (Index j = alignedStart; j<alignedSize; j+=RhsPacketSize)
|
||||
_EIGEN_ACCUMULATE_PACKETS(d,du,d);
|
||||
_EIGEN_ACCUMULATE_PACKETS(Aligned,Unaligned,Aligned);
|
||||
break;
|
||||
case FirstAligned:
|
||||
{
|
||||
@ -499,39 +505,39 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,RowMajor,Co
|
||||
* than basic unaligned loads.
|
||||
*/
|
||||
LhsPacket A01, A02, A03, A11, A12, A13;
|
||||
A01 = pload<LhsPacket>(&lhs1[alignedStart-1]);
|
||||
A02 = pload<LhsPacket>(&lhs2[alignedStart-2]);
|
||||
A03 = pload<LhsPacket>(&lhs3[alignedStart-3]);
|
||||
A01 = lhs1.template load<LhsPacket, Aligned>(alignedStart-1);
|
||||
A02 = lhs2.template load<LhsPacket, Aligned>(alignedStart-2);
|
||||
A03 = lhs3.template load<LhsPacket, Aligned>(alignedStart-3);
|
||||
|
||||
for (; j<peeledSize; j+=peels*RhsPacketSize)
|
||||
{
|
||||
RhsPacket b = pload<RhsPacket>(&rhs[j]);
|
||||
A11 = pload<LhsPacket>(&lhs1[j-1+LhsPacketSize]); palign<1>(A01,A11);
|
||||
A12 = pload<LhsPacket>(&lhs2[j-2+LhsPacketSize]); palign<2>(A02,A12);
|
||||
A13 = pload<LhsPacket>(&lhs3[j-3+LhsPacketSize]); palign<3>(A03,A13);
|
||||
RhsPacket b = rhs.getVectorMapper(j, 0).template load<RhsPacket, Aligned>(0);
|
||||
A11 = lhs1.template load<LhsPacket, Aligned>(j-1+LhsPacketSize); palign<1>(A01,A11);
|
||||
A12 = lhs2.template load<LhsPacket, Aligned>(j-2+LhsPacketSize); palign<2>(A02,A12);
|
||||
A13 = lhs3.template load<LhsPacket, Aligned>(j-3+LhsPacketSize); palign<3>(A03,A13);
|
||||
|
||||
ptmp0 = pcj.pmadd(pload<LhsPacket>(&lhs0[j]), b, ptmp0);
|
||||
ptmp0 = pcj.pmadd(lhs0.template load<LhsPacket, Aligned>(j), b, ptmp0);
|
||||
ptmp1 = pcj.pmadd(A01, b, ptmp1);
|
||||
A01 = pload<LhsPacket>(&lhs1[j-1+2*LhsPacketSize]); palign<1>(A11,A01);
|
||||
A01 = lhs1.template load<LhsPacket, Aligned>(j-1+2*LhsPacketSize); palign<1>(A11,A01);
|
||||
ptmp2 = pcj.pmadd(A02, b, ptmp2);
|
||||
A02 = pload<LhsPacket>(&lhs2[j-2+2*LhsPacketSize]); palign<2>(A12,A02);
|
||||
A02 = lhs2.template load<LhsPacket, Aligned>(j-2+2*LhsPacketSize); palign<2>(A12,A02);
|
||||
ptmp3 = pcj.pmadd(A03, b, ptmp3);
|
||||
A03 = pload<LhsPacket>(&lhs3[j-3+2*LhsPacketSize]); palign<3>(A13,A03);
|
||||
A03 = lhs3.template load<LhsPacket, Aligned>(j-3+2*LhsPacketSize); palign<3>(A13,A03);
|
||||
|
||||
b = pload<RhsPacket>(&rhs[j+RhsPacketSize]);
|
||||
ptmp0 = pcj.pmadd(pload<LhsPacket>(&lhs0[j+LhsPacketSize]), b, ptmp0);
|
||||
b = rhs.getVectorMapper(j+RhsPacketSize, 0).template load<RhsPacket, Aligned>(0);
|
||||
ptmp0 = pcj.pmadd(lhs0.template load<LhsPacket, Aligned>(j+LhsPacketSize), b, ptmp0);
|
||||
ptmp1 = pcj.pmadd(A11, b, ptmp1);
|
||||
ptmp2 = pcj.pmadd(A12, b, ptmp2);
|
||||
ptmp3 = pcj.pmadd(A13, b, ptmp3);
|
||||
}
|
||||
}
|
||||
for (; j<alignedSize; j+=RhsPacketSize)
|
||||
_EIGEN_ACCUMULATE_PACKETS(d,du,du);
|
||||
_EIGEN_ACCUMULATE_PACKETS(Aligned,Unaligned,Unaligned);
|
||||
break;
|
||||
}
|
||||
default:
|
||||
for (Index j = alignedStart; j<alignedSize; j+=RhsPacketSize)
|
||||
_EIGEN_ACCUMULATE_PACKETS(du,du,du);
|
||||
_EIGEN_ACCUMULATE_PACKETS(Unaligned,Unaligned,Unaligned);
|
||||
break;
|
||||
}
|
||||
tmp0 += predux(ptmp0);
|
||||
@ -545,9 +551,9 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,RowMajor,Co
|
||||
// FIXME this loop get vectorized by the compiler !
|
||||
for (Index j=alignedSize; j<depth; ++j)
|
||||
{
|
||||
RhsScalar b = rhs[j];
|
||||
tmp0 += cj.pmul(lhs0[j],b); tmp1 += cj.pmul(lhs1[j],b);
|
||||
tmp2 += cj.pmul(lhs2[j],b); tmp3 += cj.pmul(lhs3[j],b);
|
||||
RhsScalar b = rhs(j, 0);
|
||||
tmp0 += cj.pmul(lhs0(j),b); tmp1 += cj.pmul(lhs1(j),b);
|
||||
tmp2 += cj.pmul(lhs2(j),b); tmp3 += cj.pmul(lhs3(j),b);
|
||||
}
|
||||
res[i*resIncr] += alpha*tmp0;
|
||||
res[(i+offset1)*resIncr] += alpha*tmp1;
|
||||
@ -564,28 +570,28 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,RowMajor,Co
|
||||
{
|
||||
EIGEN_ALIGN_DEFAULT ResScalar tmp0 = ResScalar(0);
|
||||
ResPacket ptmp0 = pset1<ResPacket>(tmp0);
|
||||
const LhsScalar* lhs0 = lhs + i*lhsStride;
|
||||
const LhsScalars lhs0 = lhs.getVectorMapper(i, 0);
|
||||
// process first unaligned result's coeffs
|
||||
// FIXME this loop get vectorized by the compiler !
|
||||
for (Index j=0; j<alignedStart; ++j)
|
||||
tmp0 += cj.pmul(lhs0[j], rhs[j]);
|
||||
tmp0 += cj.pmul(lhs0(j), rhs(j, 0));
|
||||
|
||||
if (alignedSize>alignedStart)
|
||||
{
|
||||
// process aligned rhs coeffs
|
||||
if ((size_t(lhs0+alignedStart)%sizeof(LhsPacket))==0)
|
||||
if (lhs0.template aligned<LhsPacket>(alignedStart))
|
||||
for (Index j = alignedStart;j<alignedSize;j+=RhsPacketSize)
|
||||
ptmp0 = pcj.pmadd(pload<LhsPacket>(&lhs0[j]), pload<RhsPacket>(&rhs[j]), ptmp0);
|
||||
ptmp0 = pcj.pmadd(lhs0.template load<LhsPacket, Aligned>(j), rhs.getVectorMapper(j, 0).template load<RhsPacket, Aligned>(0), ptmp0);
|
||||
else
|
||||
for (Index j = alignedStart;j<alignedSize;j+=RhsPacketSize)
|
||||
ptmp0 = pcj.pmadd(ploadu<LhsPacket>(&lhs0[j]), pload<RhsPacket>(&rhs[j]), ptmp0);
|
||||
ptmp0 = pcj.pmadd(lhs0.template load<LhsPacket, Unaligned>(j), rhs.getVectorMapper(j, 0).template load<RhsPacket, Aligned>(0), ptmp0);
|
||||
tmp0 += predux(ptmp0);
|
||||
}
|
||||
|
||||
// process remaining scalars
|
||||
// FIXME this loop get vectorized by the compiler !
|
||||
for (Index j=alignedSize; j<depth; ++j)
|
||||
tmp0 += cj.pmul(lhs0[j], rhs[j]);
|
||||
tmp0 += cj.pmul(lhs0(j), rhs(j, 0));
|
||||
res[i*resIncr] += alpha*tmp0;
|
||||
}
|
||||
if (skipRows)
|
||||
|
||||
@ -51,6 +51,9 @@ EIGEN_DONT_INLINE void triangular_matrix_vector_product<Index,Mode,LhsScalar,Con
|
||||
typedef Map<Matrix<ResScalar,Dynamic,1> > ResMap;
|
||||
ResMap res(_res,rows);
|
||||
|
||||
typedef const_blas_data_mapper<LhsScalar,Index,ColMajor> LhsMapper;
|
||||
typedef const_blas_data_mapper<RhsScalar,Index,RowMajor> RhsMapper;
|
||||
|
||||
for (Index pi=0; pi<size; pi+=PanelWidth)
|
||||
{
|
||||
Index actualPanelWidth = (std::min)(PanelWidth, size-pi);
|
||||
@ -68,19 +71,19 @@ EIGEN_DONT_INLINE void triangular_matrix_vector_product<Index,Mode,LhsScalar,Con
|
||||
if (r>0)
|
||||
{
|
||||
Index s = IsLower ? pi+actualPanelWidth : 0;
|
||||
general_matrix_vector_product<Index,LhsScalar,ColMajor,ConjLhs,RhsScalar,ConjRhs,BuiltIn>::run(
|
||||
general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,ConjLhs,RhsScalar,RhsMapper,ConjRhs,BuiltIn>::run(
|
||||
r, actualPanelWidth,
|
||||
&lhs.coeffRef(s,pi), lhsStride,
|
||||
&rhs.coeffRef(pi), rhsIncr,
|
||||
LhsMapper(&lhs.coeffRef(s,pi), lhsStride),
|
||||
RhsMapper(&rhs.coeffRef(pi), rhsIncr),
|
||||
&res.coeffRef(s), resIncr, alpha);
|
||||
}
|
||||
}
|
||||
if((!IsLower) && cols>size)
|
||||
{
|
||||
general_matrix_vector_product<Index,LhsScalar,ColMajor,ConjLhs,RhsScalar,ConjRhs>::run(
|
||||
general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,ConjLhs,RhsScalar,RhsMapper,ConjRhs>::run(
|
||||
rows, cols-size,
|
||||
&lhs.coeffRef(0,size), lhsStride,
|
||||
&rhs.coeffRef(size), rhsIncr,
|
||||
LhsMapper(&lhs.coeffRef(0,size), lhsStride),
|
||||
RhsMapper(&rhs.coeffRef(size), rhsIncr),
|
||||
_res, resIncr, alpha);
|
||||
}
|
||||
}
|
||||
@ -119,6 +122,9 @@ EIGEN_DONT_INLINE void triangular_matrix_vector_product<Index,Mode,LhsScalar,Con
|
||||
typedef Map<Matrix<ResScalar,Dynamic,1>, 0, InnerStride<> > ResMap;
|
||||
ResMap res(_res,rows,InnerStride<>(resIncr));
|
||||
|
||||
typedef const_blas_data_mapper<LhsScalar,Index,RowMajor> LhsMapper;
|
||||
typedef const_blas_data_mapper<RhsScalar,Index,RowMajor> RhsMapper;
|
||||
|
||||
for (Index pi=0; pi<diagSize; pi+=PanelWidth)
|
||||
{
|
||||
Index actualPanelWidth = (std::min)(PanelWidth, diagSize-pi);
|
||||
@ -136,19 +142,19 @@ EIGEN_DONT_INLINE void triangular_matrix_vector_product<Index,Mode,LhsScalar,Con
|
||||
if (r>0)
|
||||
{
|
||||
Index s = IsLower ? 0 : pi + actualPanelWidth;
|
||||
general_matrix_vector_product<Index,LhsScalar,RowMajor,ConjLhs,RhsScalar,ConjRhs,BuiltIn>::run(
|
||||
general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,ConjLhs,RhsScalar,RhsMapper,ConjRhs,BuiltIn>::run(
|
||||
actualPanelWidth, r,
|
||||
&lhs.coeffRef(pi,s), lhsStride,
|
||||
&rhs.coeffRef(s), rhsIncr,
|
||||
LhsMapper(&lhs.coeffRef(pi,s), lhsStride),
|
||||
RhsMapper(&rhs.coeffRef(s), rhsIncr),
|
||||
&res.coeffRef(pi), resIncr, alpha);
|
||||
}
|
||||
}
|
||||
if(IsLower && rows>diagSize)
|
||||
{
|
||||
general_matrix_vector_product<Index,LhsScalar,RowMajor,ConjLhs,RhsScalar,ConjRhs>::run(
|
||||
general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,ConjLhs,RhsScalar,RhsMapper,ConjRhs>::run(
|
||||
rows-diagSize, cols,
|
||||
&lhs.coeffRef(diagSize,0), lhsStride,
|
||||
&rhs.coeffRef(0), rhsIncr,
|
||||
LhsMapper(&lhs.coeffRef(diagSize,0), lhsStride),
|
||||
RhsMapper(&rhs.coeffRef(0), rhsIncr),
|
||||
&res.coeffRef(diagSize), resIncr, alpha);
|
||||
}
|
||||
}
|
||||
|
||||
@ -37,6 +37,10 @@ struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheLeft, Mode, Con
|
||||
{
|
||||
typedef Map<const Matrix<LhsScalar,Dynamic,Dynamic,RowMajor>, 0, OuterStride<> > LhsMap;
|
||||
const LhsMap lhs(_lhs,size,size,OuterStride<>(lhsStride));
|
||||
|
||||
typedef const_blas_data_mapper<LhsScalar,Index,RowMajor> LhsMapper;
|
||||
typedef const_blas_data_mapper<RhsScalar,Index,ColMajor> RhsMapper;
|
||||
|
||||
typename internal::conditional<
|
||||
Conjugate,
|
||||
const CwiseUnaryOp<typename internal::scalar_conjugate_op<LhsScalar>,LhsMap>,
|
||||
@ -58,10 +62,10 @@ struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheLeft, Mode, Con
|
||||
Index startRow = IsLower ? pi : pi-actualPanelWidth;
|
||||
Index startCol = IsLower ? 0 : pi;
|
||||
|
||||
general_matrix_vector_product<Index,LhsScalar,RowMajor,Conjugate,RhsScalar,false>::run(
|
||||
general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,Conjugate,RhsScalar,RhsMapper,false>::run(
|
||||
actualPanelWidth, r,
|
||||
&lhs.coeffRef(startRow,startCol), lhsStride,
|
||||
rhs + startCol, 1,
|
||||
LhsMapper(&lhs.coeffRef(startRow,startCol), lhsStride),
|
||||
RhsMapper(rhs + startCol, 1),
|
||||
rhs + startRow, 1,
|
||||
RhsScalar(-1));
|
||||
}
|
||||
@ -91,6 +95,8 @@ struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheLeft, Mode, Con
|
||||
{
|
||||
typedef Map<const Matrix<LhsScalar,Dynamic,Dynamic,ColMajor>, 0, OuterStride<> > LhsMap;
|
||||
const LhsMap lhs(_lhs,size,size,OuterStride<>(lhsStride));
|
||||
typedef const_blas_data_mapper<LhsScalar,Index,ColMajor> LhsMapper;
|
||||
typedef const_blas_data_mapper<RhsScalar,Index,ColMajor> RhsMapper;
|
||||
typename internal::conditional<Conjugate,
|
||||
const CwiseUnaryOp<typename internal::scalar_conjugate_op<LhsScalar>,LhsMap>,
|
||||
const LhsMap&
|
||||
@ -122,10 +128,10 @@ struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheLeft, Mode, Con
|
||||
// let's directly call the low level product function because:
|
||||
// 1 - it is faster to compile
|
||||
// 2 - it is slighlty faster at runtime
|
||||
general_matrix_vector_product<Index,LhsScalar,ColMajor,Conjugate,RhsScalar,false>::run(
|
||||
general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,Conjugate,RhsScalar,RhsMapper,false>::run(
|
||||
r, actualPanelWidth,
|
||||
&lhs.coeffRef(endBlock,startBlock), lhsStride,
|
||||
rhs+startBlock, 1,
|
||||
LhsMapper(&lhs.coeffRef(endBlock,startBlock), lhsStride),
|
||||
RhsMapper(rhs+startBlock, 1),
|
||||
rhs+endBlock, 1, RhsScalar(-1));
|
||||
}
|
||||
}
|
||||
|
||||
@ -34,7 +34,9 @@ template<
|
||||
int ResStorageOrder>
|
||||
struct general_matrix_matrix_product;
|
||||
|
||||
template<typename Index, typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs, typename RhsScalar, bool ConjugateRhs, int Version=Specialized>
|
||||
template<typename Index,
|
||||
typename LhsScalar, typename LhsMapper, int LhsStorageOrder, bool ConjugateLhs,
|
||||
typename RhsScalar, typename RhsMapper, bool ConjugateRhs, int Version=Specialized>
|
||||
struct general_matrix_vector_product;
|
||||
|
||||
|
||||
@ -118,13 +120,35 @@ template<typename Scalar> struct get_factor<Scalar,typename NumTraits<Scalar>::R
|
||||
};
|
||||
|
||||
|
||||
template<typename Scalar, typename Index>
|
||||
class BlasVectorMapper {
|
||||
public:
|
||||
EIGEN_ALWAYS_INLINE BlasVectorMapper(Scalar *data) : m_data(data) {}
|
||||
|
||||
EIGEN_ALWAYS_INLINE Scalar operator()(Index i) const {
|
||||
return m_data[i];
|
||||
}
|
||||
template <typename Packet, int AlignmentType>
|
||||
EIGEN_ALWAYS_INLINE Packet load(Index i) const {
|
||||
return ploadt<Packet, AlignmentType>(m_data + i);
|
||||
}
|
||||
|
||||
template <typename Packet>
|
||||
bool aligned(Index i) const {
|
||||
return (size_t(m_data+i)%sizeof(Packet))==0;
|
||||
}
|
||||
|
||||
protected:
|
||||
Scalar* m_data;
|
||||
};
|
||||
|
||||
template<typename Scalar, typename Index, int AlignmentType>
|
||||
class MatrixLinearMapper {
|
||||
class BlasLinearMapper {
|
||||
public:
|
||||
typedef typename packet_traits<Scalar>::type Packet;
|
||||
typedef typename packet_traits<Scalar>::half HalfPacket;
|
||||
|
||||
EIGEN_ALWAYS_INLINE MatrixLinearMapper(Scalar *data) : m_data(data) {}
|
||||
EIGEN_ALWAYS_INLINE BlasLinearMapper(Scalar *data) : m_data(data) {}
|
||||
|
||||
EIGEN_ALWAYS_INLINE void prefetch(int i) const {
|
||||
internal::prefetch(&operator()(i));
|
||||
@ -157,7 +181,8 @@ class blas_data_mapper {
|
||||
typedef typename packet_traits<Scalar>::type Packet;
|
||||
typedef typename packet_traits<Scalar>::half HalfPacket;
|
||||
|
||||
typedef MatrixLinearMapper<Scalar, Index, AlignmentType> LinearMapper;
|
||||
typedef BlasLinearMapper<Scalar, Index, AlignmentType> LinearMapper;
|
||||
typedef BlasVectorMapper<Scalar, Index> VectorMapper;
|
||||
|
||||
EIGEN_ALWAYS_INLINE blas_data_mapper(Scalar* data, Index stride) : m_data(data), m_stride(stride) {}
|
||||
|
||||
@ -170,6 +195,11 @@ class blas_data_mapper {
|
||||
return LinearMapper(&operator()(i, j));
|
||||
}
|
||||
|
||||
EIGEN_ALWAYS_INLINE VectorMapper getVectorMapper(Index i, Index j) const {
|
||||
return VectorMapper(&operator()(i, j));
|
||||
}
|
||||
|
||||
|
||||
EIGEN_DEVICE_FUNC
|
||||
EIGEN_ALWAYS_INLINE Scalar& operator()(Index i, Index j) const {
|
||||
return m_data[StorageOrder==RowMajor ? j + i*m_stride : i + j*m_stride];
|
||||
@ -193,6 +223,15 @@ class blas_data_mapper {
|
||||
return pgather<Scalar, SubPacket>(&operator()(i, j), m_stride);
|
||||
}
|
||||
|
||||
const Index stride() const { return m_stride; }
|
||||
|
||||
Index firstAligned(Index size) const {
|
||||
if (size_t(m_data)%sizeof(Scalar)) {
|
||||
return -1;
|
||||
}
|
||||
return internal::first_aligned(m_data, size);
|
||||
}
|
||||
|
||||
protected:
|
||||
Scalar* EIGEN_RESTRICT m_data;
|
||||
const Index m_stride;
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user