mirror/eigen
2
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-01-24 14:45:14 +08:00
Code Issues Packages Projects Releases Wiki Activity

optimize linear vectorization both in Assign and Sum (optimal amortized perf)

Browse Source
This commit is contained in:
Gael Guennebaud 2008-06-23 15:50:28 +00:00
parent ea1990ef3d
commit ac9aa47bbc
2 changed files with 24 additions and 28 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
Eigen/src/Core/Assign.h
View File

@ -307,12 +307,17 @@ struct ei_assign_impl<Derived1, Derived2, LinearVectorization, NoUnrolling>
int index = 0;
// do the vectorizable part of the assignment
for ( ; index<alignedSize ; index+=packetSize)
int row = 0;
int col = 0;
while (index<alignedSize)
{
// FIXME the following is not really efficient
const int row = rowMajor ? index/innerSize : index%innerSize;
const int col = rowMajor ? index%innerSize : index/innerSize;
dst.template writePacket<Aligned>(row, col, src.template packet<Aligned>(row, col));
int start = rowMajor ? col : row;
int end = std::min(innerSize, start + alignedSize-index);
for ( ; (rowMajor ? col : row)<end; (rowMajor ? col : row)+=packetSize)
dst.template writePacket<Aligned>(row, col, src.template packet<Aligned>(row, col));
index += (rowMajor ? col : row) - start;
row = rowMajor ? index/innerSize : index%innerSize;
col = rowMajor ? index%innerSize : index/innerSize;
}
// now we must do the rest without vectorization.
@ -380,7 +385,7 @@ struct ei_assign_impl<Derived1, Derived2, SliceVectorization, NoUnrolling>
const int innerSize = rowMajor ? dst.cols() : dst.rows();
const int outerSize = rowMajor ? dst.rows() : dst.cols();
const int alignedInnerSize = (innerSize/packetSize)*packetSize;
for(int i = 0; i < outerSize; i++)
{
// do the vectorizable part of the assignment

35
Eigen/src/Core/Sum.h
View File

@ -54,7 +54,7 @@ public:
Unrolling = Cost <= UnrollingLimit
? CompleteUnrolling
: NoUnrolling
};
};
};
/***************************************************************************
@ -62,7 +62,7 @@ public:
***************************************************************************/
/*** no vectorization ***/
template<typename Derived, int Start, int Length>
struct ei_sum_novec_unroller
{
@ -194,32 +194,23 @@ struct ei_sum_impl<Derived, LinearVectorization, NoUnrolling>
// do the vectorizable part of the sum
if(size >= packetSize)
{
asm("#begin");
PacketScalar packet_res;
packet_res = mat.template packet<Aligned>(0, 0);
int index;
if(Derived::IsVectorAtCompileTime)
int row = 0;
int col = 0;
int index = packetSize;
while (index<alignedSize)
{
for(index = packetSize; index<alignedSize ; index+=packetSize)
{
const int row = Derived::RowsAtCompileTime==1 ? 0 : index;
const int col = Derived::RowsAtCompileTime==1 ? index : 0;
row = rowMajor ? index/innerSize : index%innerSize;
col = rowMajor ? index%innerSize : index/innerSize;
int start = rowMajor ? col : row;
int end = std::min(innerSize, start+alignedSize-index);
if (end<start) getchar();
for ( ; (rowMajor ? col : row)<end; (rowMajor ? col : row)+=packetSize)
packet_res = ei_padd(packet_res, mat.template packet<Aligned>(row, col));
}
}
else
{
for(index = packetSize; index<alignedSize ; index+=packetSize)
{
// FIXME the following is not really efficient
const int row = rowMajor ? index/innerSize : index%innerSize;
const int col = rowMajor ? index%innerSize : index/innerSize;
packet_res = ei_padd(packet_res, mat.template packet<Aligned>(row, col));
}
index += (rowMajor ? col : row) - start;
}
res = ei_predux(packet_res);
asm("#end");
// now we must do the rest without vectorization.
if(alignedSize == size) return res;