mirror/eigen
2
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-03-07 18:27:40 +08:00
Code Issues Packages Projects Releases Wiki Activity

add "slice vectorization" of redux (eg. m.block().minCoeff() is now

Browse Source 
vectorized)
This commit is contained in:
Gael Guennebaud 2009-03-09 23:16:39 +00:00
parent c087373968
commit 8aa5aa269a
1 changed files with 54 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

61
Eigen/src/Core/Redux.h
View File

@ -39,18 +39,26 @@ struct ei_redux_traits
{
private:
enum {
PacketSize = ei_packet_traits<typename Derived::Scalar>::size
PacketSize = ei_packet_traits<typename Derived::Scalar>::size,
InnerMaxSize = int(Derived::Flags)&RowMajorBit
? Derived::MaxColsAtCompileTime
: Derived::MaxRowsAtCompileTime
};
enum {
MightVectorize = (int(Derived::Flags)&ActualPacketAccessBit)
&& (ei_functor_traits<Func>::PacketAccess),
MayLinearVectorize = MightVectorize && (int(Derived::Flags)&LinearAccessBit),
MaySliceVectorize = MightVectorize && int(InnerMaxSize)>=3*PacketSize
};
public:
enum {
Vectorization = (int(Derived::Flags)&ActualPacketAccessBit)
&& (int(Derived::Flags)&LinearAccessBit)
&& (ei_functor_traits<Func>::PacketAccess)
? LinearVectorization
: NoVectorization
Vectorization = int(MayLinearVectorize) ? int(LinearVectorization)
: int(MaySliceVectorize) ? int(SliceVectorization)
: int(NoVectorization)
};
private:
enum {
Cost = Derived::SizeAtCompileTime * Derived::CoeffReadCost
@ -227,6 +235,45 @@ struct ei_redux_impl<Func, Derived, LinearVectorization, NoUnrolling>
}
};
template<typename Func, typename Derived>
struct ei_redux_impl<Func, Derived, SliceVectorization, NoUnrolling>
{
typedef typename Derived::Scalar Scalar;
typedef typename ei_packet_traits<Scalar>::type PacketScalar;
static Scalar run(const Derived& mat, const Func& func)
{
const int innerSize = mat.innerSize();
const int outerSize = mat.outerSize();
enum {
packetSize = ei_packet_traits<Scalar>::size,
isRowMajor = Derived::Flags&RowMajorBit?1:0
};
const int packetedInnerSize = ((innerSize)/packetSize)*packetSize;
Scalar res;
if(packetedInnerSize)
{
PacketScalar packet_res = mat.template packet<Unaligned>(0,0);
for(int j=0; j<outerSize; ++j)
for(int i=0; i<packetedInnerSize; i+=int(packetSize))
packet_res = func.packetOp(packet_res, mat.template packet<Unaligned>
(isRowMajor?j:i, isRowMajor?i:j));
res = func.predux(packet_res);
for(int j=0; j<outerSize; ++j)
for(int i=packetedInnerSize; i<innerSize; ++i)
res = func(res, mat.coeff(isRowMajor?j:i, isRowMajor?i:j));
}
else // too small to vectorize anything.
// since this is dynamic-size hence inefficient anyway for such small sizes, don't try to optimize.
{
res = ei_redux_impl<Func, Derived, NoVectorization, NoUnrolling>::run(mat, func);
}
return res;
}
};
template<typename Func, typename Derived>
struct ei_redux_impl<Func, Derived, LinearVectorization, CompleteUnrolling>
{