nojhan/paradiseo
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

independent ParadisEO-GPU package

Browse source 
git-svn-id: svn://scm.gforge.inria.fr/svnroot/paradiseo@2589 331e1502-861f-0410-8da2-ba01fb791d7f
This commit is contained in:
boufaras 2012-01-09 10:29:30 +00:00
commit 95787e3a24
9 changed files with 1173 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

200
branches/ParadisEO-GPU/paradiseo-gpu/src/eval/moGPUEvalByCpy.h Normal file
View file

@ -0,0 +1,200 @@
/*
<moGPUEvalByCpy.h>
Copyright (C) DOLPHIN Project-Team, INRIA Lille - Nord Europe, 2006-2010
Karima Boufaras, Thé Van LUONG
This software is governed by the CeCILL license under French law and
abiding by the rules of distribution of free software. You can use,
modify and/ or redistribute the software under the terms of the CeCILL
license as circulated by CEA, CNRS and INRIA at the following URL
"http://www.cecill.info".
As a counterpart to the access to the source code and rights to copy,
modify and redistribute granted by the license, users are provided only
with a limited warranty and the software's author, the holder of the
economic rights, and the successive licensors have only limited liability.
In this respect, the user's attention is drawn to the risks associated
with loading, using, modifying and/or developing or reproducing the
software by the user in light of its specific status of free software,
that may mean that it is complicated to manipulate, and that also
therefore means that it is reserved for developers and experienced
professionals having in-depth computer knowledge. Users are therefore
encouraged to load and test the software's suitability as regards their
requirements in conditions enabling the security of their systems and/or
data to be ensured and, more generally, to use and operate it in the
same conditions as regards security.
The fact that you are presently reading this means that you have had
knowledge of the CeCILL license and that you accept its terms.
ParadisEO WebSite : http://paradiseo.gforge.inria.fr
Contact: paradiseo-help@lists.gforge.inria.fr
*/
#ifndef __moGPUEvalByCpy_H
#define __moGPUEvalByCpy_H
#include <eval/moGPUEval.h>
#include <eval/moGPUKernelEvalByCpy.h>
#include <performance/moGPUTimer.h>
/**
* class for the parallel evaluation of neighborhood
*/
template<class Neighbor, class Eval>
class moGPUEvalByCpy: public moGPUEval<Neighbor> {
public:
/**
* Define type of a solution corresponding to Neighbor
**/
typedef typename Neighbor::EOT EOT;
/**
* Define vector type of vector corresponding to Solution
**/
typedef typename EOT::ElemType T;
/**
* Define type of a fitness corresponding to Solution
**/
typedef typename EOT::Fitness Fitness;
using moGPUEval<Neighbor>::neighborhoodSize;
using moGPUEval<Neighbor>::host_FitnessArray;
using moGPUEval<Neighbor>::device_FitnessArray;
using moGPUEval<Neighbor>::device_solution;
using moGPUEval<Neighbor>::NEW_kernel_Dim;
using moGPUEval<Neighbor>::NEW_BLOCK_SIZE;
using moGPUEval<Neighbor>::mutex;
/**
* Constructor
* @param _neighborhoodSize the size of the neighborhood
* @param _eval how to evaluate a neighbor
*/
moGPUEvalByCpy(unsigned int _neighborhoodSize, Eval & _eval) :
moGPUEval<Neighbor> (_neighborhoodSize), eval(_eval) {
}
/**
* Compute fitness for all solution neighbors in device
* @param _sol the solution that generate the neighborhood to evaluate parallely
* @param _cpySolution Launch kernel with local copy option of solution in each thread
* @param _withCalibration an automatic configuration of kernel to launch( NB_THREAD BY BLOCK & NB_BLOCK BY KERNEL ),default (1)
*/
void neighborhoodEval(EOT & _sol, bool _cpySolution, bool _withCalibration) {
if (_cpySolution) {
unsigned size=_sol.size();
if (!mutex) {
//Allocate the space for solution in the global memory of device
cudaMalloc((void**) &device_solution.vect, size * sizeof(T));
if (_withCalibration)
calibration(_sol);
mutex = true;
}
// Get Current solution fitness
Fitness fitness = _sol.fitness();
//Copy the solution vector from the host to device
cudaMemcpy(device_solution.vect, _sol.vect, size * sizeof(T),
cudaMemcpyHostToDevice);
//Launch the Kernel to compute all neighbors fitness with local copy of solution
moGPUKernelEvalByCpy<T,Fitness,Eval><<<NEW_kernel_Dim,NEW_BLOCK_SIZE >>>(eval,device_solution.vect,device_FitnessArray,fitness,neighborhoodSize);
//Copy the result from device to host
cudaMemcpy(host_FitnessArray, device_FitnessArray, neighborhoodSize
* sizeof(Fitness), cudaMemcpyDeviceToHost);
} else
cout << "It's evaluation by copy set cpySolution to true" << endl;
}
/**
* Compute the best combination of number of block by grid and number of thread within block
* @param _sol the solution that generate the neighborhood to evaluate parallely
*/
virtual void calibration(EOT & _sol) {
unsigned size = _sol.size();
Fitness fitness = _sol.fitness();
unsigned NB_THREAD[6] = { 16, 32, 64, 128, 256, 512 };
double mean_time[7] = { 0, 0, 0, 0, 0, 0 };
unsigned i = 0;
double best_time = RAND_MAX;
unsigned tmp_kernel_Dim;
#ifndef BLOCK_SIZE
do {
tmp_kernel_Dim = neighborhoodSize / NB_THREAD[i]
+ ((neighborhoodSize % NB_THREAD[i] == 0) ? 0 : 1);
for (unsigned k = 0; k < 5; k++) {
cudaMemcpy(device_solution.vect, _sol.vect, size * sizeof(T),
cudaMemcpyHostToDevice);
moGPUTimer timer;
timer.start();
moGPUKernelEvalByCpy<T,Fitness,Eval><<<tmp_kernel_Dim,NB_THREAD[i] >>>(eval,device_solution.vect,device_FitnessArray,fitness,neighborhoodSize);
timer.stop();
mean_time[i] += (timer.getTime());
}
if (best_time >= (mean_time[i] / 5)) {
best_time = mean_time[i] / 5;
NEW_BLOCK_SIZE = NB_THREAD[i];
NEW_kernel_Dim = tmp_kernel_Dim;
}
i++;
} while (i < 6);
#else
tmp_kernel_Dim =NEW_kernel_Dim;
for (unsigned k = 0; k < 5; k++) {
cudaMemcpy(device_solution.vect, _sol.vect, size * sizeof(T),
cudaMemcpyHostToDevice);
moGPUTimer timer;
timer.start();
moGPUKernelEvalByCpy<T,Fitness,Eval><<<tmp_kernel_Dim,BLOCK_SIZE >>>(eval,device_solution.vect,device_FitnessArray,fitness,neighborhoodSize);
timer.stop();
mean_time[6] += (timer.getTime());
}
if (best_time >= (mean_time[6] / 5))
best_time = mean_time[6] / 5;
do {
tmp_kernel_Dim = neighborhoodSize / NB_THREAD[i]
+ ((neighborhoodSize % NB_THREAD[i] == 0) ? 0 : 1);
for (unsigned k = 0; k < 5; k++) {
cudaMemcpy(device_solution.vect, _sol.vect, size * sizeof(T),
cudaMemcpyHostToDevice);
moGPUTimer timer;
timer.start();
moGPUKernelEvalByCpy<T,Fitness,Eval><<<tmp_kernel_Dim,NB_THREAD[i] >>>(eval,device_solution.vect,device_FitnessArray,fitness,neighborhoodSize);
timer.stop();
mean_time[i] += (timer.getTime());
}
if (best_time >= (mean_time[i] / 5)) {
best_time = mean_time[i] / 5;
NEW_BLOCK_SIZE = NB_THREAD[i];
NEW_kernel_Dim = tmp_kernel_Dim;
}
i++;
}while (i < 6);
#endif
}
protected:
Eval & eval;
};
#endif

199
branches/ParadisEO-GPU/paradiseo-gpu/src/eval/moGPUEvalByModif.h Normal file
View file

@ -0,0 +1,199 @@
/*
<moGPUEvalByModif.h>
Copyright (C) DOLPHIN Project-Team, INRIA Lille - Nord Europe, 2006-2010
Karima Boufaras, Thé Van LUONG
This software is governed by the CeCILL license under French law and
abiding by the rules of distribution of free software. You can use,
modify and/ or redistribute the software under the terms of the CeCILL
license as circulated by CEA, CNRS and INRIA at the following URL
"http://www.cecill.info".
As a counterpart to the access to the source code and rights to copy,
modify and redistribute granted by the license, users are provided only
with a limited warranty and the software's author, the holder of the
economic rights, and the successive licensors have only limited liability.
In this respect, the user's attention is drawn to the risks associated
with loading, using, modifying and/or developing or reproducing the
software by the user in light of its specific status of free software,
that may mean that it is complicated to manipulate, and that also
therefore means that it is reserved for developers and experienced
professionals having in-depth computer knowledge. Users are therefore
encouraged to load and test the software's suitability as regards their
requirements in conditions enabling the security of their systems and/or
data to be ensured and, more generally, to use and operate it in the
same conditions as regards security.
The fact that you are presently reading this means that you have had
knowledge of the CeCILL license and that you accept its terms.
ParadisEO WebSite : http://paradiseo.gforge.inria.fr
Contact: paradiseo-help@lists.gforge.inria.fr
*/
#ifndef __moGPUEvalByModif_H
#define __moGPUEvalByModif_H
#include <eval/moGPUKernelEvalByModif.h>
#include <eval/moGPUEval.h>
#include <performance/moGPUTimer.h>
/**
* class for the parallel evaluation of neighborhood
*/
template<class Neighbor, class Eval>
class moGPUEvalByModif: public moGPUEval<Neighbor> {
public:
/**
* Define type of a solution corresponding to Neighbor
**/
typedef typename Neighbor::EOT EOT;
/**
* Define vector type of vector corresponding to Solution
**/
typedef typename EOT::ElemType T;
/**
* Define type of a fitness corresponding to Solution
**/
typedef typename EOT::Fitness Fitness;
using moGPUEval<Neighbor>::neighborhoodSize;
using moGPUEval<Neighbor>::host_FitnessArray;
using moGPUEval<Neighbor>::device_FitnessArray;
using moGPUEval<Neighbor>::device_solution;
using moGPUEval<Neighbor>::NEW_kernel_Dim;
using moGPUEval<Neighbor>::NEW_BLOCK_SIZE;
using moGPUEval<Neighbor>::mutex;
/**
* Constructor
* @param _neighborhoodSize the size of the neighborhood
* @param _eval the incremental evaluation
*/
moGPUEvalByModif(unsigned int _neighborhoodSize, Eval & _eval) :
moGPUEval<Neighbor> (_neighborhoodSize), eval(_eval) {
}
/**
* Compute fitness for all solution neighbors in device
* @param _sol the solution that generate the neighborhood to evaluate parallely
* @param _cpySolution Launch kernel with local copy option of solution in each thread
* @param _withCalibration an automatic configuration of kernel to launch( NB_THREAD BY BLOCK & NB_BLOCK BY KERNEL ),default (1)
*/
void neighborhoodEval(EOT & _sol, bool _cpySolution, bool _withCalibration) {
if (!_cpySolution) {
unsigned size = _sol.size();
if (!mutex) {
//Allocate the space for solution in the global memory of device
cudaMalloc((void**) &device_solution.vect, size * sizeof(T));
if (_withCalibration)
calibration(_sol);
mutex = true;
}
// Get Current solution fitness
Fitness fitness = _sol.fitness();
//Copy the solution vector from the host to device
cudaMemcpy(device_solution.vect, _sol.vect, size * sizeof(T),
cudaMemcpyHostToDevice);
//Launch the Kernel to compute all neighbors fitness
moGPUKernelEvalByModif<T,Fitness,Eval><<<NEW_kernel_Dim,NEW_BLOCK_SIZE>>>(eval,device_solution.vect,device_FitnessArray,fitness,neighborhoodSize);
//Copy the result from device to host
cudaMemcpy(host_FitnessArray, device_FitnessArray, neighborhoodSize
* sizeof(Fitness), cudaMemcpyDeviceToHost);
} else
cout << "It's evaluation by Modif set cpySolution to false"
<< endl;
}
/**
* Compute the best combination of number of block by grid and number of thread within block
* @param _sol the solution that generate the neighborhood to evaluate parallely
*/
virtual void calibration(EOT & _sol) {
unsigned size = _sol.size();
Fitness fitness = _sol.fitness();
unsigned NB_THREAD[6] = { 16, 32, 64, 128, 256, 512 };
double mean_time[7] = { 0, 0, 0, 0, 0, 0 };
unsigned i = 0;
double best_time = RAND_MAX;
unsigned tmp_kernel_Dim;
#ifndef BLOCK_SIZE
do {
tmp_kernel_Dim = neighborhoodSize / NB_THREAD[i]
+ ((neighborhoodSize % NB_THREAD[i] == 0) ? 0 : 1);
for (unsigned k = 0; k < 5; k++) {
cudaMemcpy(device_solution.vect, _sol.vect, size * sizeof(T),
cudaMemcpyHostToDevice);
moGPUTimer timer;
timer.start();
moGPUKernelEvalByModif<T,Fitness,Eval><<<tmp_kernel_Dim,NB_THREAD[i]>>>(eval,device_solution.vect,device_FitnessArray,fitness,neighborhoodSize);
timer.stop();
mean_time[i] += (timer.getTime());
}
if (best_time >= (mean_time[i] / 5)) {
best_time = mean_time[i] / 5;
NEW_BLOCK_SIZE = NB_THREAD[i];
NEW_kernel_Dim = tmp_kernel_Dim;
}
i++;
} while (i < 6);
#else
tmp_kernel_Dim =NEW_kernel_Dim;
for (unsigned k = 0; k < 5; k++) {
cudaMemcpy(device_solution.vect, _sol.vect, size * sizeof(T),
cudaMemcpyHostToDevice);
moGPUTimer timer;
timer.start();
moGPUKernelEvalByModif<T,Fitness,Eval><<<tmp_kernel_Dim,BLOCK_SIZE>>>(eval,device_solution.vect,device_FitnessArray,fitness,neighborhoodSize);
timer.stop();
mean_time[6] += (timer.getTime());
}
if (best_time >= (mean_time[6] / 5))
best_time = mean_time[6] / 5;
do {
tmp_kernel_Dim = neighborhoodSize / NB_THREAD[i]
+ ((neighborhoodSize % NB_THREAD[i] == 0) ? 0 : 1);
for (unsigned k = 0; k < 5; k++) {
cudaMemcpy(device_solution.vect, _sol.vect, size * sizeof(T),
cudaMemcpyHostToDevice);
moGPUTimer timer;
timer.start();
moGPUKernelEvalByModif<T,Fitness,Eval><<<tmp_kernel_Dim,NB_THREAD[i]>>>(eval,device_solution.vect,device_FitnessArray,fitness,neighborhoodSize);
timer.stop();
mean_time[i] += (timer.getTime());
}
if (best_time >= (mean_time[i] / 5)) {
best_time = mean_time[i] / 5;
NEW_BLOCK_SIZE = NB_THREAD[i];
NEW_kernel_Dim = tmp_kernel_Dim;
}
i++;
}while (i < 6);
#endif
}
protected:
Eval & eval;
};
#endif

84
branches/ParadisEO-GPU/paradiseo-gpu/src/eval/moGPUEvalFunc.h Normal file
View file

@ -0,0 +1,84 @@
/*
<moGPUEvalFunc.h>
Copyright (C) DOLPHIN Project-Team, INRIA Lille - Nord Europe, 2006-2010
Karima Boufaras, Thé Van LUONG
This software is governed by the CeCILL license under French law and
abiding by the rules of distribution of free software. You can use,
modify and/ or redistribute the software under the terms of the CeCILL
license as circulated by CEA, CNRS and INRIA at the following URL
"http://www.cecill.info".
As a counterpart to the access to the source code and rights to copy,
modify and redistribute granted by the license, users are provided only
with a limited warranty and the software's author, the holder of the
economic rights, and the successive licensors have only limited liability.
In this respect, the user's attention is drawn to the risks associated
with loading, using, modifying and/or developing or reproducing the
software by the user in light of its specific status of free software,
that may mean that it is complicated to manipulate, and that also
therefore means that it is reserved for developers and experienced
professionals having in-depth computer knowledge. Users are therefore
encouraged to load and test the software's suitability as regards their
requirements in conditions enabling the security of their systems and/or
data to be ensured and, more generally, to use and operate it in the
same conditions as regards security.
The fact that you are presently reading this means that you have had
knowledge of the CeCILL license and that you accept its terms.
ParadisEO WebSite : http://paradiseo.gforge.inria.fr
Contact: paradiseo-help@lists.gforge.inria.fr
*/
#ifndef __moGPUEvalFunc_H
#define __moGPUEvalFunc_H
/**
* Abstract class for GPU evaluation of neighbor
*/
template<class Neighbor>
class moGPUEvalFunc {
public:
/**
* Define type of a solution corresponding to Neighbor
*/
typedef typename Neighbor::EOT EOT;
typedef typename EOT::ElemType T;
typedef typename EOT::Fitness Fitness;
/**
* Constructor
*/
moGPUEvalFunc() {
}
/**
* Destructor
*/
virtual ~moGPUEvalFunc() {
}
/**
*Virtual functor to compute fitness of a solution neighbor
*@param _solution the solution which generate the neighborhood
*@param _fitness the current solution fitness
*@param _index an array that contains a set of indexes corresponding to the current thread identifier neighbor
* the last element of this array contains neighborhood size
*/
virtual inline __host__ __device__ Fitness operator() (T * _solution,Fitness _fitness, unsigned int * _index) {
return _fitness;
}
};
#endif

71
branches/ParadisEO-GPU/paradiseo-gpu/src/eval/moGPUKernelEvalByCpy.h Normal file
View file

@ -0,0 +1,71 @@
/*
<moGPUKernelEvalByCpy.h>
Copyright (C) DOLPHIN Project-Team, INRIA Lille - Nord Europe, 2006-2010
Karima Boufaras, Thé Van LUONG
This software is governed by the CeCILL license under French law and
abiding by the rules of distribution of free software. You can use,
modify and/ or redistribute the software under the terms of the CeCILL
license as circulated by CEA, CNRS and INRIA at the following URL
"http://www.cecill.info".
As a counterpart to the access to the source code and rights to copy,
modify and redistribute granted by the license, users are provided only
with a limited warranty and the software's author, the holder of the
economic rights, and the successive licensors have only limited liability.
In this respect, the user's attention is drawn to the risks associated
with loading, using, modifying and/or developing or reproducing the
software by the user in light of its specific status of free software,
that may mean that it is complicated to manipulate, and that also
therefore means that it is reserved for developers and experienced
professionals having in-depth computer knowledge. Users are therefore
encouraged to load and test the software's suitability as regards their
requirements in conditions enabling the security of their systems and/or
data to be ensured and, more generally, to use and operate it in the
same conditions as regards security.
The fact that you are presently reading this means that you have had
knowledge of the CeCILL license and that you accept its terms.
ParadisEO WebSite : http://paradiseo.gforge.inria.fr
Contact: paradiseo-help@lists.gforge.inria.fr
*/
#ifndef __moGPUKernelEvalByCpy_H
#define __moGPUKernelEvalByCpy_H
///////////////////////////////////////////////////////////////////////////////////////////////////////////////
/**
* The kernel function called from the host and executed in device to compute all neighbors fitness at one time
* without mapping, each thread id compute one fitness by copying localy the solution
* @param _eval how to evaluate each neighbor
* @param _solution the representation of solution( vector of int,float....)
* @param _allFitness the array of Fitness to save all neighbors fitness
* @param _fitness the current solution fitness
* @param _neighborhoodsize the size of the neighborhood
*/
template<class T, class Fitness, class Eval>
__global__ void moGPUKernelEvalByCpy(Eval _eval, T * _solution, Fitness* _allFitness,
Fitness _fitness, unsigned _neighborhoodsize) {
// The thread identifier within a grid block's
int id = blockIdx.x * blockDim.x + threadIdx.x;
// array to save index to be changed
unsigned index[2];
T sol_tmp[SIZE];
// In this representation each id identify one and only one neighbor in neighborhood
if (id < _neighborhoodsize) {
for(unsigned i=0;i<SIZE;i++)
sol_tmp[i]=_solution[i];
//Change the id'th element of solution
index[0]=id;
index[1]=_neighborhoodsize;
//Compute fitness for id'th neighbor
_allFitness[id] = _eval(sol_tmp,_fitness,index);
}
}
#endif

68
branches/ParadisEO-GPU/paradiseo-gpu/src/eval/moGPUKernelEvalByModif.h Normal file
View file

@ -0,0 +1,68 @@
/*
<moGPUKernelEvalByModif.h>
Copyright (C) DOLPHIN Project-Team, INRIA Lille - Nord Europe, 2006-2010
Karima Boufaras, Thé Van LUONG
This software is governed by the CeCILL license under French law and
abiding by the rules of distribution of free software. You can use,
modify and/ or redistribute the software under the terms of the CeCILL
license as circulated by CEA, CNRS and INRIA at the following URL
"http://www.cecill.info".
As a counterpart to the access to the source code and rights to copy,
modify and redistribute granted by the license, users are provided only
with a limited warranty and the software's author, the holder of the
economic rights, and the successive licensors have only limited liability.
In this respect, the user's attention is drawn to the risks associated
with loading, using, modifying and/or developing or reproducing the
software by the user in light of its specific status of free software,
that may mean that it is complicated to manipulate, and that also
therefore means that it is reserved for developers and experienced
professionals having in-depth computer knowledge. Users are therefore
encouraged to load and test the software's suitability as regards their
requirements in conditions enabling the security of their systems and/or
data to be ensured and, more generally, to use and operate it in the
same conditions as regards security.
The fact that you are presently reading this means that you have had
knowledge of the CeCILL license and that you accept its terms.
ParadisEO WebSite : http://paradiseo.gforge.inria.fr
Contact: paradiseo-help@lists.gforge.inria.fr
*/
#ifndef __moGPUKernelEvalByModif_H
#define __moGPUKernelEvalByModif_H
///////////////////////////////////////////////////////////////////////////////////////////////////////////////
/**
* The kernel function called from the host and executed in device to compute all neighbors fitness at one time
* without mapping, each thread id compute one fitness by modif of solution
* @param _eval how to evaluate each neighbor
* @param _solution the representation of solution( vector of int,float....)
* @param _allFitness the array of Fitness to save all neighbors fitness
* @param _fitness the current solution fitness
* @param _neighborhoodsize the size of the neighborhood
*/
template<class T, class Fitness, class Eval>
__global__ void moGPUKernelEvalByModif(Eval _eval, T * _solution, Fitness* _allFitness,
Fitness _fitness, unsigned _neighborhoodsize) {
// The thread identifier within a grid block's
int id = blockIdx.x * blockDim.x + threadIdx.x;
// array to save index to be changed
unsigned index[2];
// In this representation each id identify one and only one neighbor in neighborhood
if (id < _neighborhoodsize) {
//Change the id'th element of solution
index[0]=id;
index[1]=_neighborhoodsize;
//Compute fitness for id'th neighbor
_allFitness[id] = _eval(_solution,_fitness,index);
}
}
#endif

202
branches/ParadisEO-GPU/paradiseo-gpu/src/eval/moGPUMappingEvalByCpy.h Normal file
View file

@ -0,0 +1,202 @@
/*
<moGPUMappingEvalByCpy.h>
Copyright (C) DOLPHIN Project-Team, INRIA Lille - Nord Europe, 2006-2010
Karima Boufaras, Thé Van LUONG
This software is governed by the CeCILL license under French law and
abiding by the rules of distribution of free software. You can use,
modify and/ or redistribute the software under the terms of the CeCILL
license as circulated by CEA, CNRS and INRIA at the following URL
"http://www.cecill.info".
As a counterpart to the access to the source code and rights to copy,
modify and redistribute granted by the license, users are provided only
with a limited warranty and the software's author, the holder of the
economic rights, and the successive licensors have only limited liability.
In this respect, the user's attention is drawn to the risks associated
with loading, using, modifying and/or developing or reproducing the
software by the user in light of its specific status of free software,
that may mean that it is complicated to manipulate, and that also
therefore means that it is reserved for developers and experienced
professionals having in-depth computer knowledge. Users are therefore
encouraged to load and test the software's suitability as regards their
requirements in conditions enabling the security of their systems and/or
data to be ensured and, more generally, to use and operate it in the
same conditions as regards security.
The fact that you are presently reading this means that you have had
knowledge of the CeCILL license and that you accept its terms.
ParadisEO WebSite : http://paradiseo.gforge.inria.fr
Contact: paradiseo-help@lists.gforge.inria.fr
*/
#ifndef __moGPUMappingEvalByCpy_H
#define __moGPUMappingEvalByCpy_H
#include <eval/moGPUEval.h>
#include <eval/moGPUMappingKernelEvalByCpy.h>
#include <performance/moGPUTimer.h>
/**
* class for the Mapping neighborhood evaluation
*/
template<class Neighbor, class Eval>
class moGPUMappingEvalByCpy: public moGPUEval<Neighbor> {
public:
/**
* Define type of a solution corresponding to Neighbor
*/
typedef typename Neighbor::EOT EOT;
/**
* Define type of a vector corresponding to Solution
*/
typedef typename EOT::ElemType T;
/**
* Define type of a fitness corresponding to Solution
*/
typedef typename EOT::Fitness Fitness;
using moGPUEval<Neighbor>::neighborhoodSize;
using moGPUEval<Neighbor>::host_FitnessArray;
using moGPUEval<Neighbor>::device_FitnessArray;
using moGPUEval<Neighbor>::device_solution;
using moGPUEval<Neighbor>::NEW_BLOCK_SIZE;
using moGPUEval<Neighbor>::NEW_kernel_Dim;
using moGPUEval<Neighbor>::mutex;
/**
* Constructor
* @param _neighborhoodSize the size of the neighborhood
* @param _eval how to evaluate a neighbor
*/
moGPUMappingEvalByCpy(unsigned int _neighborhoodSize, Eval & _eval) :
moGPUEval<Neighbor> (_neighborhoodSize), eval(_eval) {
}
/**
* Destructor
*/
~moGPUMappingEvalByCpy() {
}
/**
* Compute fitness for all solution neighbors in device with associated mapping
* @param _sol the solution that generate the neighborhood to evaluate parallely
* @param _mapping the array of mapping indexes that associate a neighbor identifier to X-position
* @param _cpySolution Launch kernel with local copy option of solution in each thread if it's set to true
* @param _withCalibration an automatic kernel configuration, fix nbr of thread by block and nbr of grid by kernel
*/
void neighborhoodEval(EOT & _sol, unsigned int * _mapping,
bool _cpySolution, bool _withCalibration) {
if (_cpySolution) {
unsigned size = _sol.size();
// Get Current solution fitness
Fitness fitness = _sol.fitness();
if (!mutex) {
//Allocate the space for solution in the device global memory
cudaMalloc((void**) &device_solution.vect, size * sizeof(T));
if (_withCalibration)
calibration(_sol, _mapping);
mutex = true;
}
//Copy the solution vector from the host to device
cudaMemcpy(device_solution.vect, _sol.vect, size * sizeof(T),
cudaMemcpyHostToDevice);
//Launch the Kernel to compute all neighbors fitness,using a given mapping
moGPUMappingKernelEvalByCpy<T,Fitness,Eval><<<NEW_kernel_Dim,NEW_BLOCK_SIZE >>>(eval,device_solution.vect,device_FitnessArray,fitness,_mapping,neighborhoodSize);
cudaMemcpy(host_FitnessArray, device_FitnessArray, neighborhoodSize
* sizeof(Fitness), cudaMemcpyDeviceToHost);
} else
cout << "It's evaluation by copy set cpySolution to true" << endl;
}
/**
* Compute the best combination of number of block by grid and number of thread within block
* @param _sol the solution that generate the neighborhood to evaluate parallely
* @param _mapping the array of mapping indexes that associate a neighbor identifier to X-position
*/
virtual void calibration(EOT & _sol, unsigned int * _mapping) {
unsigned size = _sol.size();
Fitness fitness = _sol.fitness();
unsigned NB_THREAD[6] = { 16, 32, 64, 128, 256, 512 };
double mean_time[7] = { 0, 0, 0, 0, 0, 0 };
unsigned i = 0;
double best_time = 0;
unsigned tmp_kernel_Dim;
best_time = RAND_MAX;
#ifndef BLOCK_SIZE
do {
tmp_kernel_Dim = neighborhoodSize / NB_THREAD[i]
+ ((neighborhoodSize % NB_THREAD[i] == 0) ? 0 : 1);
for (unsigned k = 0; k < 5; k++) {
cudaMemcpy(device_solution.vect, _sol.vect, size * sizeof(T),
cudaMemcpyHostToDevice);
moGPUTimer timer;
timer.start();
moGPUMappingKernelEvalByCpy<T,Fitness,Eval><<<tmp_kernel_Dim,NB_THREAD[i]>>>(eval,device_solution.vect,device_FitnessArray,fitness,_mapping,neighborhoodSize);
timer.stop();
mean_time[i] += (timer.getTime());
}
if (best_time >= (mean_time[i] / 5)) {
best_time = mean_time[i] / 5;
NEW_BLOCK_SIZE = NB_THREAD[i];
NEW_kernel_Dim = tmp_kernel_Dim;
}
i++;
} while (i < 6);
#else
tmp_kernel_Dim =NEW_kernel_Dim;
for (unsigned k = 0; k < 5; k++) {
cudaMemcpy(device_solution.vect, _sol.vect, size * sizeof(T),
cudaMemcpyHostToDevice);
moGPUTimer timer;
timer.start();
moGPUMappingKernelEvalByCpy<T,Fitness,Eval><<<tmp_kernel_Dim,BLOCK_SIZE>>>(eval,device_solution.vect,device_FitnessArray,fitness,_mapping,neighborhoodSize);
timer.stop();
mean_time[6] += (timer.getTime());
}
if (best_time >= (mean_time[6] / 5))
best_time = mean_time[6] / 5;
do {
tmp_kernel_Dim = neighborhoodSize / NB_THREAD[i]
+ ((neighborhoodSize % NB_THREAD[i] == 0) ? 0 : 1);
for (unsigned k = 0; k < 5; k++) {
cudaMemcpy(device_solution.vect, _sol.vect, size * sizeof(T),
cudaMemcpyHostToDevice);
moGPUTimer timer;
timer.start();
moGPUMappingKernelEvalByCpy<T,Fitness,Eval><<<tmp_kernel_Dim,NB_THREAD[i]>>>(eval,device_solution.vect,device_FitnessArray,fitness,_mapping,neighborhoodSize);
timer.stop();
mean_time[i] += (timer.getTime());
}
if (best_time >= (mean_time[i] / 5)) {
best_time = mean_time[i] / 5;
NEW_BLOCK_SIZE = NB_THREAD[i];
NEW_kernel_Dim = tmp_kernel_Dim;
}
i++;
}while (i < 6);
#endif
}
protected:
Eval & eval;
};
#endif

200
branches/ParadisEO-GPU/paradiseo-gpu/src/eval/moGPUMappingEvalByModif.h Normal file
View file

@ -0,0 +1,200 @@
/*
<moGPUMappingEvalByModif.h>
Copyright (C) DOLPHIN Project-Team, INRIA Lille - Nord Europe, 2006-2010
Karima Boufaras, Thé Van LUONG
This software is governed by the CeCILL license under French law and
abiding by the rules of distribution of free software. You can use,
modify and/ or redistribute the software under the terms of the CeCILL
license as circulated by CEA, CNRS and INRIA at the following URL
"http://www.cecill.info".
As a counterpart to the access to the source code and rights to copy,
modify and redistribute granted by the license, users are provided only
with a limited warranty and the software's author, the holder of the
economic rights, and the successive licensors have only limited liability.
In this respect, the user's attention is drawn to the risks associated
with loading, using, modifying and/or developing or reproducing the
software by the user in light of its specific status of free software,
that may mean that it is complicated to manipulate, and that also
therefore means that it is reserved for developers and experienced
professionals having in-depth computer knowledge. Users are therefore
encouraged to load and test the software's suitability as regards their
requirements in conditions enabling the security of their systems and/or
data to be ensured and, more generally, to use and operate it in the
same conditions as regards security.
The fact that you are presently reading this means that you have had
knowledge of the CeCILL license and that you accept its terms.
ParadisEO WebSite : http://paradiseo.gforge.inria.fr
Contact: paradiseo-help@lists.gforge.inria.fr
*/
#ifndef __moGPUMappingEvalByModif_H
#define __moGPUMappingEvalByModif_H
#include <eval/moGPUEval.h>
#include <eval/moGPUMappingKernelEvalByModif.h>
#include <performance/moGPUTimer.h>
/**
* class for the Mapping neighborhood evaluation
*/
template<class Neighbor, class Eval>
class moGPUMappingEvalByModif: public moGPUEval<Neighbor> {
public:
/**
* Define type of a solution corresponding to Neighbor
*/
typedef typename Neighbor::EOT EOT;
/**
* Define type of a vector corresponding to Solution
*/
typedef typename EOT::ElemType T;
/**
* Define type of a fitness corresponding to Solution
*/
typedef typename EOT::Fitness Fitness;
using moGPUEval<Neighbor>::neighborhoodSize;
using moGPUEval<Neighbor>::host_FitnessArray;
using moGPUEval<Neighbor>::device_FitnessArray;
using moGPUEval<Neighbor>::device_solution;
using moGPUEval<Neighbor>::NEW_BLOCK_SIZE;
using moGPUEval<Neighbor>::NEW_kernel_Dim;
using moGPUEval<Neighbor>::mutex;
/**
* Constructor
* @param _neighborhoodSize the size of the neighborhood
* @param _eval the incremental evaluation
*/
moGPUMappingEvalByModif(unsigned int _neighborhoodSize, Eval & _eval) :
moGPUEval<Neighbor> (_neighborhoodSize), eval(_eval) {
}
/**
* Destructor
*/
~moGPUMappingEvalByModif() {
}
/**
* Compute fitness for all solution neighbors in device with associated mapping
* @param _sol the solution that generate the neighborhood to evaluate parallely
* @param _mapping the array of mapping indexes that associate a neighbor identifier to X-position
* @param _cpySolution Launch kernel with local copy option of solution in each thread if it's set to true
* @param _withCalibration an automatic kernel configuration, fix nbr of thread by block and nbr of grid by kernel
*/
void neighborhoodEval(EOT & _sol, unsigned int * _mapping,
bool _cpySolution, bool _withCalibration) {
if (!_cpySolution) {
unsigned size = _sol.size();
// Get Current solution fitness
Fitness fitness = _sol.fitness();
if (!mutex) {
//Allocate the space for solution in the device global memory
cudaMalloc((void**) &device_solution.vect, size * sizeof(T));
if (_withCalibration)
calibration(_sol, _mapping);
mutex = true;
}
//Copy the solution vector from the host to device
cudaMemcpy(device_solution.vect, _sol.vect, size * sizeof(T),
cudaMemcpyHostToDevice);
//Launch the Kernel to compute all neighbors fitness,using a given mapping
moGPUMappingKernelEvalByModif<T,Fitness,Eval><<<NEW_kernel_Dim,NEW_BLOCK_SIZE >>>(eval,device_solution.vect,device_FitnessArray,fitness,_mapping,neighborhoodSize);
cudaMemcpy(host_FitnessArray, device_FitnessArray, neighborhoodSize
* sizeof(Fitness), cudaMemcpyDeviceToHost);
} else
cout << "It's evaluation by Modif set cpySolution to false" << endl;
}
/**
* Compute the best combination of number of block by grid and number of thread within block
* @param _sol the solution that generate the neighborhood to evaluate parallely
* @param _mapping the array of mapping indexes that associate a neighbor identifier to X-position
*/
virtual void calibration(EOT & _sol, unsigned int * _mapping) {
unsigned size = _sol.size();
Fitness fitness = _sol.fitness();
unsigned NB_THREAD[6] = { 16, 32, 64, 128, 256, 512 };
double mean_time[7] = { 0, 0, 0, 0, 0, 0 };
unsigned i = 0;
double best_time = 0;
unsigned tmp_kernel_Dim;
best_time = RAND_MAX;
#ifndef BLOCK_SIZE
do {
tmp_kernel_Dim = neighborhoodSize / NB_THREAD[i]
+ ((neighborhoodSize % NB_THREAD[i] == 0) ? 0 : 1);
for (unsigned k = 0; k < 5; k++) {
cudaMemcpy(device_solution.vect, _sol.vect, size * sizeof(T),
cudaMemcpyHostToDevice);
moGPUTimer timer;
timer.start();
moGPUMappingKernelEvalByModif<T,Fitness,Eval><<<tmp_kernel_Dim,NB_THREAD[i]>>>(eval,device_solution.vect,device_FitnessArray,fitness,_mapping,neighborhoodSize);
timer.stop();
mean_time[i] += (timer.getTime());
}
if (best_time >= (mean_time[i] / 5)) {
best_time = mean_time[i] / 5;
NEW_BLOCK_SIZE = NB_THREAD[i];
NEW_kernel_Dim = tmp_kernel_Dim;
}
i++;
} while (i < 6);
#else
tmp_kernel_Dim =NEW_kernel_Dim;
for (unsigned k = 0; k < 5; k++) {
cudaMemcpy(device_solution.vect, _sol.vect, size * sizeof(T),
cudaMemcpyHostToDevice);
moGPUTimer timer;
timer.start();
moGPUMappingKernelEvalByModif<T,Fitness,Eval><<<tmp_kernel_Dim,NEW_BLOCK_SIZE >>>(eval,device_solution.vect,device_FitnessArray,fitness,_mapping,neighborhoodSize);
timer.stop();
mean_time[6] += (timer.getTime());
}
if (best_time >= (mean_time[6] / 5))
best_time = mean_time[6] / 5;
do {
tmp_kernel_Dim = neighborhoodSize / NB_THREAD[i]
+ ((neighborhoodSize % NB_THREAD[i] == 0) ? 0 : 1);
for (unsigned k = 0; k < 5; k++) {
cudaMemcpy(device_solution.vect, _sol.vect, size * sizeof(T),
cudaMemcpyHostToDevice);
moGPUTimer timer;
timer.start();
moGPUMappingKernelEvalByModif<T,Fitness,Eval><<<tmp_kernel_Dim,NB_THREAD[i] >>>(eval,device_solution.vect,device_FitnessArray,fitness,_mapping,neighborhoodSize);
timer.stop();
mean_time[i] += (timer.getTime());
}
if (best_time >= (mean_time[i] / 5)) {
best_time = mean_time[i] / 5;
NEW_BLOCK_SIZE = NB_THREAD[i];
NEW_kernel_Dim = tmp_kernel_Dim;
}
i++;
}while (i < 6);
#endif
}
protected:
Eval & eval;
};
#endif

76
branches/ParadisEO-GPU/paradiseo-gpu/src/eval/moGPUMappingKernelEvalByCpy.h Normal file
View file

@ -0,0 +1,76 @@
/*
<moGPUMappingKernelEvalByCpy.h>
Copyright (C) DOLPHIN Project-Team, INRIA Lille - Nord Europe, 2006-2010
Karima Boufaras, Thé Van LUONG
This software is governed by the CeCILL license under French law and
abiding by the rules of distribution of free software. You can use,
modify and/ or redistribute the software under the terms of the CeCILL
license as circulated by CEA, CNRS and INRIA at the following URL
"http://www.cecill.info".
As a counterpart to the access to the source code and rights to copy,
modify and redistribute granted by the license, users are provided only
with a limited warranty and the software's author, the holder of the
economic rights, and the successive licensors have only limited liability.
In this respect, the user's attention is drawn to the risks associated
with loading, using, modifying and/or developing or reproducing the
software by the user in light of its specific status of free software,
that may mean that it is complicated to manipulate, and that also
therefore means that it is reserved for developers and experienced
professionals having in-depth computer knowledge. Users are therefore
encouraged to load and test the software's suitability as regards their
requirements in conditions enabling the security of their systems and/or
data to be ensured and, more generally, to use and operate it in the
same conditions as regards security.
The fact that you are presently reading this means that you have had
knowledge of the CeCILL license and that you accept its terms.
ParadisEO WebSite : http://paradiseo.gforge.inria.fr
Contact: paradiseo-help@lists.gforge.inria.fr
*/
#ifndef __moGPUMappingKernelEvalByCpy_H
#define __moGPUMappingKernelEvalByCpy_H
///////////////////////////////////////////////////////////////////////////////////////////////////////////////
/**
* The kernel function called from the host and executed in device to compute all neighbors fitness at one time
* without mapping, each thread id compute one fitness by modif of solution
* @param _eval how to evaluate each neighbor
* @param _solution the representation of solution( vector of int,float....)
* @param _allFitness the array of Fitness to save all neighbors fitness
* @param _fitness the current solution fitness
* @param _mapping associate to each threadID a set of correspondent indexes
* @param _neighborhoodsize the size of the neighborhood
*/
template<class T, class Fitness, class Eval>
__global__ void moGPUMappingKernelEvalByCpy(Eval _eval, T * _solution, Fitness* _allFitness,
Fitness _fitness,unsigned * _mapping,unsigned _neighborhoodsize) {
// The thread identifier within a grid block's
int id = blockIdx.x * blockDim.x + threadIdx.x;
//counter of number of x-change
unsigned i;
// array to save set a set of indexes corresponding to the current thread identifier
unsigned index[NB_POS+2];
T sol_tmp[SIZE];
// In this representation each id identify one and only one neighbor in neighborhood
if (id < _neighborhoodsize) {
for(i=0;i<SIZE;i++)
sol_tmp[i]=_solution[i];
for(i=0;i<NB_POS;i++)
index[i]=_mapping[id + i * _neighborhoodsize];
index[NB_POS]=_neighborhoodsize;
index[NB_POS+1]=id;
//Evaluate by Modif Id'th neighbor with index mapping
_allFitness[id]=_eval(sol_tmp,_fitness, index);
}
}
#endif

73
branches/ParadisEO-GPU/paradiseo-gpu/src/eval/moGPUMappingKernelEvalByModif.h Normal file
View file

@ -0,0 +1,73 @@
/*
<moGPUMappingKernelEvalByModif.h>
Copyright (C) DOLPHIN Project-Team, INRIA Lille - Nord Europe, 2006-2010
Karima Boufaras, Thé Van LUONG
This software is governed by the CeCILL license under French law and
abiding by the rules of distribution of free software. You can use,
modify and/ or redistribute the software under the terms of the CeCILL
license as circulated by CEA, CNRS and INRIA at the following URL
"http://www.cecill.info".
As a counterpart to the access to the source code and rights to copy,
modify and redistribute granted by the license, users are provided only
with a limited warranty and the software's author, the holder of the
economic rights, and the successive licensors have only limited liability.
In this respect, the user's attention is drawn to the risks associated
with loading, using, modifying and/or developing or reproducing the
software by the user in light of its specific status of free software,
that may mean that it is complicated to manipulate, and that also
therefore means that it is reserved for developers and experienced
professionals having in-depth computer knowledge. Users are therefore
encouraged to load and test the software's suitability as regards their
requirements in conditions enabling the security of their systems and/or
data to be ensured and, more generally, to use and operate it in the
same conditions as regards security.
The fact that you are presently reading this means that you have had
knowledge of the CeCILL license and that you accept its terms.
ParadisEO WebSite : http://paradiseo.gforge.inria.fr
Contact: paradiseo-help@lists.gforge.inria.fr
*/
#ifndef __moGPUMappingKernelEvalByModif_H
#define __moGPUMappingKernelEvalByModif_H
///////////////////////////////////////////////////////////////////////////////////////////////////////////////
/**
* The kernel function called from the host and executed in device to compute all neighbors fitness at one time
* without mapping, each thread id compute one fitness by modif of solution
* @param _eval how to evaluate each neighbor
* @param _solution the representation of solution( vector of int,float....)
* @param _allFitness the array of Fitness to save all neighbors fitness
* @param _fitness the current solution fitness
* @param _mapping associate to each threadID a set of correspondent indexes
* @param _neighborhoodsize the size of the neighborhood
*/
template<class T, class Fitness, class Eval>
__global__ void moGPUMappingKernelEvalByModif(Eval _eval, T * _solution, Fitness* _allFitness,
Fitness _fitness,unsigned int * _mapping,unsigned _neighborhoodsize) {
// The thread identifier within a grid block's
int id = blockIdx.x * blockDim.x + threadIdx.x;
//counter of number of x-change
unsigned i;
// array to save set a set of indexes corresponding to the current thread identifier
unsigned index[NB_POS+2];
// In this representation each id identify one and only one neighbor in neighborhood
if (id < _neighborhoodsize) {
for(i=0;i<NB_POS;i++)
index[i]=_mapping[id + i * _neighborhoodsize];
index[NB_POS]=_neighborhoodsize;
index[NB_POS+1]=id;
//Evaluate by Modif Id'th neighbor with index mapping
_allFitness[id]=_eval(_solution,_fitness, index);
}
}
#endif