add an optimized "apply in place a rotation in the plane",

and make Jacobi and SelfAdjointEigenSolver use it
=> ~ x1.75 speedup for JacobiSVD and x2 for SelfAdjointEigenSolver
This commit is contained in:
Gael Guennebaud 2009-08-13 11:42:02 +02:00
parent 1d80f561ad
commit 1b257a7620
4 changed files with 136 additions and 27 deletions

View File

@ -194,6 +194,7 @@ namespace Eigen {
#include "src/Core/products/TriangularMatrixVector.h"
#include "src/Core/products/TriangularMatrixMatrix.h"
#include "src/Core/products/TriangularSolverMatrix.h"
#include "src/Core/products/RotationInThePlane.h"
#include "src/Core/BandMatrix.h"
} // namespace Eigen

View File

@ -0,0 +1,127 @@
// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2009 Gael Guennebaud <g.gael@free.fr>
//
// Eigen is free software; you can redistribute it and/or
// modify it under the terms of the GNU Lesser General Public
// License as published by the Free Software Foundation; either
// version 3 of the License, or (at your option) any later version.
//
// Alternatively, you can redistribute it and/or
// modify it under the terms of the GNU General Public License as
// published by the Free Software Foundation; either version 2 of
// the License, or (at your option) any later version.
//
// Eigen is distributed in the hope that it will be useful, but WITHOUT ANY
// WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
// FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License or the
// GNU General Public License for more details.
//
// You should have received a copy of the GNU Lesser General Public
// License and a copy of the GNU General Public License along with
// Eigen. If not, see <http://www.gnu.org/licenses/>.
#ifndef EIGEN_ROTATION_IN_THE_PLANE_H
#define EIGEN_ROTATION_IN_THE_PLANE_H
/**********************************************************************
* This file implement ...
**********************************************************************/
template<typename Scalar, int Incr>
struct ei_apply_rotation_in_the_plane_selector;
template<typename VectorX, typename VectorY>
void ei_apply_rotation_in_the_plane(VectorX& x, VectorY& y, typename VectorX::Scalar c, typename VectorY::Scalar s)
{
ei_assert(x.size() == y.size());
int size = x.size();
int incrx = size ==1 ? 1 : &x.coeffRef(1) - &x.coeffRef(0);
int incry = size ==1 ? 1 : &y.coeffRef(1) - &y.coeffRef(0);
if (incrx==1 && incry==1)
ei_apply_rotation_in_the_plane_selector<typename VectorX::Scalar,1>
::run(&x.coeffRef(0), &y.coeffRef(0), x.size(), c, s, 1, 1);
else
ei_apply_rotation_in_the_plane_selector<typename VectorX::Scalar,Dynamic>
::run(&x.coeffRef(0), &y.coeffRef(0), x.size(), c, s, incrx, incry);
}
template<typename Scalar>
struct ei_apply_rotation_in_the_plane_selector<Scalar,Dynamic>
{
static EIGEN_DONT_INLINE void run(Scalar* x, Scalar* y, int size, Scalar c, Scalar s, int incrx, int incry)
{
for(int i=0; i<size; ++i)
{
Scalar xi = *x;
Scalar yi = *y;
*x = c * xi - s * yi;
*y = s * xi + c * yi;
x += incrx;
y += incry;
}
}
};
// both vectors are sequentially stored in memory => vectorization
template<typename Scalar>
struct ei_apply_rotation_in_the_plane_selector<Scalar,1>
{
static EIGEN_DONT_INLINE void run(Scalar* x, Scalar* y, int size, Scalar c, Scalar s, int, int)
{
typedef typename ei_packet_traits<Scalar>::type Packet;
enum { PacketSize = ei_packet_traits<Scalar>::size, Peeling = 2 };
int alignedStart = ei_alignmentOffset(y, size);
int alignedEnd = alignedStart + ((size-alignedStart)/(Peeling*PacketSize))*(Peeling*PacketSize);
const Packet pc = ei_pset1(c);
const Packet ps = ei_pset1(s);
for(int i=0; i<alignedStart; ++i)
{
Scalar xi = x[i];
Scalar yi = y[i];
x[i] = c * xi - s * yi;
y[i] = s * xi + c * yi;
}
Scalar* px = x + alignedStart;
Scalar* py = y + alignedStart;
if(ei_alignmentOffset(x, size)==alignedStart)
for(int i=alignedStart; i<alignedEnd; i+=PacketSize)
{
Packet xi = ei_pload(px);
Packet yi = ei_pload(py);
ei_pstore(px, ei_psub(ei_pmul(pc,xi),ei_pmul(ps,yi)));
ei_pstore(py, ei_padd(ei_pmul(ps,xi),ei_pmul(pc,yi)));
px += PacketSize;
py += PacketSize;
}
else
for(int i=alignedStart; i<alignedEnd; i+=Peeling*PacketSize)
{
Packet xi = ei_ploadu(px);
Packet xi1 = ei_ploadu(px+PacketSize);
Packet yi = ei_pload (py);
Packet yi1 = ei_pload (py+PacketSize);
ei_pstoreu(px, ei_psub(ei_pmul(pc,xi),ei_pmul(ps,yi)));
ei_pstoreu(px+PacketSize, ei_psub(ei_pmul(pc,xi1),ei_pmul(ps,yi1)));
ei_pstore (py, ei_padd(ei_pmul(ps,xi),ei_pmul(pc,yi)));
ei_pstore (py+PacketSize, ei_padd(ei_pmul(ps,xi1),ei_pmul(pc,yi1)));
px += Peeling*PacketSize;
py += Peeling*PacketSize;
}
for(int i=alignedEnd; i<size; ++i)
{
Scalar xi = x[i];
Scalar yi = y[i];
x[i] = c * xi - s * yi;
y[i] = s * xi + c * yi;
}
}
};
#endif // EIGEN_ROTATION_IN_THE_PLANE_H

View File

@ -28,23 +28,17 @@
template<typename Derived>
void MatrixBase<Derived>::applyJacobiOnTheLeft(int p, int q, Scalar c, Scalar s)
{
for(int i = 0; i < cols(); ++i)
{
Scalar tmp = coeff(p,i);
coeffRef(p,i) = c * tmp - s * coeff(q,i);
coeffRef(q,i) = s * tmp + c * coeff(q,i);
}
RowXpr x(row(p));
RowXpr y(row(q));
ei_apply_rotation_in_the_plane(x, y, c, s);
}
template<typename Derived>
void MatrixBase<Derived>::applyJacobiOnTheRight(int p, int q, Scalar c, Scalar s)
{
for(int i = 0; i < rows(); ++i)
{
Scalar tmp = coeff(i,p);
coeffRef(i,p) = c * tmp - s * coeff(i,q);
coeffRef(i,q) = s * tmp + c * coeff(i,q);
}
ColXpr x(col(p));
ColXpr y(col(q));
ei_apply_rotation_in_the_plane(x, y, c, s);
}
template<typename Scalar>

View File

@ -378,23 +378,10 @@ static void ei_tridiagonal_qr_step(RealScalar* diag, RealScalar* subdiag, int st
if (matrixQ)
{
#ifdef EIGEN_DEFAULT_TO_ROW_MAJOR
ei_apply_rotation_in_the_plane_selector<Scalar,Dynamic>::run(matrixQ+k, matrixQ+k+1, n, c, s, n, n);
#else
int kn = k*n;
int kn1 = (k+1)*n;
ei_apply_rotation_in_the_plane_selector<Scalar,1>::run(matrixQ+k*n, matrixQ+k*n+n, n, c, s, 1, 1);
#endif
// let's do the product manually to avoid the need of temporaries...
for (int i=0; i<n; ++i)
{
#ifdef EIGEN_DEFAULT_TO_ROW_MAJOR
Scalar matrixQ_i_k = matrixQ[i*n+k];
matrixQ[i*n+k] = c * matrixQ_i_k - s * matrixQ[i*n+k+1];
matrixQ[i*n+k+1] = s * matrixQ_i_k + c * matrixQ[i*n+k+1];
#else
Scalar matrixQ_i_k = matrixQ[i+kn];
matrixQ[i+kn] = c * matrixQ_i_k - s * matrixQ[i+kn1];
matrixQ[i+kn1] = s * matrixQ_i_k + c * matrixQ[i+kn1];
#endif
}
}
}
}