3.3.3.2 : Sans les transfert de données

Développons le fichier main_kernel.cpp :

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57

#include <iostream>
#include "timer.h"
#include "asterics_cuda.h"

#include "hadamard_product_cuda.h"

///Get the number of cycles per elements of the Hadamard product
/**	@param nbElement : number of elements of the tables
 * 	@param nbRepetition : number of repetition to evaluate the function hadamard_product
 * 	@param maxNbThreadPerBlockX : maximum number of thread per block on X
 * 	@param maxNbBlockX : maximum number of block in the grid on X
 * 	@param warpSize : number of thread per warp
*/
void evaluateHadamardProduct(long unsigned int nbElement, long unsigned int nbRepetition, int maxNbThreadPerBlockX, int maxNbBlockX, int warpSize){
	//Allocation of the tables
	float * tabResult = new float[nbElement];
	float * tabX = new float[nbElement];
	float * tabY = new float[nbElement];
	//Initialisation of the tables
	for(long unsigned int i(0lu); i < nbElement; ++i){
		tabX[i] = (float)(i*32lu%17lu);
		tabY[i] = (float)(i*57lu%31lu);
	}
	hadamard_product_cuda_clock(tabResult, tabX, tabY, nbElement, nbRepetition,
				maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	
	//Deallocate the tables
	delete[] tabResult;
	delete[] tabX;
	delete[] tabY;
}


int main(int argc, char** argv){
	std::cout << "Hadamard product Kernel" << std::endl;
#ifdef SELECTED_GPU
	if(!asterics_setDevice(SELECTED_GPU)){std::cerr << "Device '" SELECTED_GPU "' not found" << std::endl;return -1;}
#else
	int deviceCount(asterics_getNbCudaDevice());
	int maxNbThreadPerBlockX(0), maxNbBlockX(0), warpSize(0);
	asterics_getGpuInfo(maxNbThreadPerBlockX, maxNbBlockX, warpSize, 0);
#endif
	evaluateHadamardProduct(1000lu, 100000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(2000lu, 100000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(3000lu, 100000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(5000lu, 100000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(10000lu, 100000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(20000lu, 100000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(50000lu, 100000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(100000lu, 100000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(500000lu, 100000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(1000000lu, 10000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(10000000lu, 10000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(100000000lu, 1000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(500000000lu, 100lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	return 0;
}


Le fichier main_kernel.cpp complet.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63

/***************************************
	Auteur : Pierre Aubert
	Mail : pierre.aubert@lapp.in2p3.fr
	Licence : CeCILL-C
****************************************/

#include <iostream>
#include "timer.h"
#include "asterics_cuda.h"

#include "hadamard_product_cuda.h"

///Get the number of cycles per elements of the Hadamard product
/**	@param nbElement : number of elements of the tables
 * 	@param nbRepetition : number of repetition to evaluate the function hadamard_product
 * 	@param maxNbThreadPerBlockX : maximum number of thread per block on X
 * 	@param maxNbBlockX : maximum number of block in the grid on X
 * 	@param warpSize : number of thread per warp
*/
void evaluateHadamardProduct(long unsigned int nbElement, long unsigned int nbRepetition, int maxNbThreadPerBlockX, int maxNbBlockX, int warpSize){
	//Allocation of the tables
	float * tabResult = new float[nbElement];
	float * tabX = new float[nbElement];
	float * tabY = new float[nbElement];
	//Initialisation of the tables
	for(long unsigned int i(0lu); i < nbElement; ++i){
		tabX[i] = (float)(i*32lu%17lu);
		tabY[i] = (float)(i*57lu%31lu);
	}
	hadamard_product_cuda_clock(tabResult, tabX, tabY, nbElement, nbRepetition,
				maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	
	//Deallocate the tables
	delete[] tabResult;
	delete[] tabX;
	delete[] tabY;
}


int main(int argc, char** argv){
	std::cout << "Hadamard product Kernel" << std::endl;
#ifdef SELECTED_GPU
	if(!asterics_setDevice(SELECTED_GPU)){std::cerr << "Device '" SELECTED_GPU "' not found" << std::endl;return -1;}
#else
	int deviceCount(asterics_getNbCudaDevice());
	int maxNbThreadPerBlockX(0), maxNbBlockX(0), warpSize(0);
	asterics_getGpuInfo(maxNbThreadPerBlockX, maxNbBlockX, warpSize, 0);
#endif
	evaluateHadamardProduct(1000lu, 100000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(2000lu, 100000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(3000lu, 100000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(5000lu, 100000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(10000lu, 100000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(20000lu, 100000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(50000lu, 100000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(100000lu, 100000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(500000lu, 100000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(1000000lu, 10000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(10000000lu, 10000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(100000000lu, 1000lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	evaluateHadamardProduct(500000000lu, 100lu, maxNbThreadPerBlockX, maxNbBlockX, warpSize);
	return 0;
}


Vous pouvez télécharger le fichier ici.