void test_question_four_one(){
double* A = matrix_alloc(2,2);
double* B = matrix_alloc(2,2);
double* C = matrix_alloc(2,2);
matrix_init_ones(C,2,2);
int i =0;
for (i=0 ; i<4 ; i++) {
A[i] = i+1;
B[i] = i+5;
}
printf("Matrix A:\n");
affiche(2,2,A,2,stdout);
printf("Matrix B:\n");
affiche(2,2,B,2,stdout);
cblas_dgemm_scalaire(CblasColMajor, CblasTrans, CblasNoTrans, 2, 2, 2, 1, A, 2, B, 2, 0, C, 2);
printf("\nMatrix C:\n");
affiche(2,2,C,2,stdout);
// 17 39
// 23 53
free(A);
free(B);
free(C);
}
void prod_matrix(int N, int Nb, int myrank,
double* bl_a, double* bl_b, double* bl_c,
MPI_Comm comm_grid, MPI_Comm comm_col, MPI_Comm comm_row)
{
int k;
double my_a[Nb*Nb];
int coords[2];
MPI_Status st;
int gd = N/Nb;
for (int i = 0; i < Nb*Nb; ++i)
{
my_a[i] = bl_a[i];
}
MPI_Cart_coords(comm_grid, myrank, 2, coords);
int sndto = (((coords[0]-1)%gd) +gd) %gd;
int recvfrom = (coords[0]+1)%gd;
int myrow = coords[0];
int mycol = coords[1];
for (k = 0; k < gd; k++)
{
/* If I am i+k%N proc of the line
* Bcast_line(A[i][i+k%N])
* Else
* recv(A) from the i+k%N proc of the line
*/
if(mycol == (myrow+k)%gd)
{
for (int i = 0; i < Nb*Nb; ++i)
{
bl_a[i] = my_a[i];
}
}
MPI_Bcast(bl_a, Nb*Nb, MPI_DOUBLE, (myrow+k)%gd, comm_row);
cblas_dgemm_scalaire(Nb, bl_a, bl_b, bl_c); //Cij = A[i][i+k%N]*B[i+k%N][j]
/* send(B) to upper neighbour
*/
MPI_Sendrecv_replace(bl_b, Nb*Nb, MPI_DOUBLE, sndto, 0,
recvfrom, 0, comm_col, &st);
}
}
int main(int argc, char* argv[])
{
char dummy[L2_CACHE_SIZE];
// Tests de performances de ddot
int size = 50;
blas_t *matriceD, *matriceE;
alloc_vecteur(&matriceD, size);
alloc_vecteur(&matriceE, size);
printf("Tests de performance de la fonction ddot\n");
perf_t *t1, *t2,*t3, *t4,*t5, *t6,*t7, *t8, *t9, *t10;
t1 = malloc(sizeof(perf_t));
t2 = malloc(sizeof(perf_t));
t3 = malloc(sizeof(perf_t));
t4 = malloc(sizeof(perf_t));
t5 = malloc(sizeof(perf_t));
t6 = malloc(sizeof(perf_t));
t7 = malloc(sizeof(perf_t));
t8 = malloc(sizeof(perf_t));
t9 = malloc(sizeof(perf_t));
t10 = malloc(sizeof(perf_t));
double mflops, mflops1,mflops2,mflops3,mflops4, mflops5;
char command[200];
system("rm results/ddot_perf.txt");
for(size = 50; size < 100000000; size += size/4)
{
printf("M: %d ", size);
if(size != 50)
{
free(matriceD);
free(matriceE);
alloc_vecteur(&matriceD, size);
alloc_vecteur(&matriceE, size);
}
memset(dummy, 0, sizeof(dummy));
perf(t1);
blas_t res = cblas_ddot(size, matriceD, 1, matriceE, 1);
perf(t2);
perf_diff(t1, t2);
mflops = perf_mflops(t2, 2 * size);
printf("Mflops/s: %le\n", mflops);
sprintf(command, "echo %d %lf >> results/ddot_perf.txt", size, mflops);
system(command);
}
// Test de performance dgemm
//////////////////////////////////////////
long m = 100;
blas_t *matriceA, *matriceB, *matriceC;
alloc_matrice(&matriceA, m, m);
alloc_matrice(&matriceB, m, m);
matriceC = calloc(m*m,sizeof(blas_t));
system("rm results/dgemm_perf.txt");
for(; m< 1000; m+=20)
{
printf("M: %d ", m);
if(m != 100)
{
free(matriceA);
free(matriceB);
free(matriceC);
alloc_matrice(&matriceA, m, m);
alloc_matrice(&matriceB, m, m);
alloc_matrice(&matriceC, m, m);
}
memset(dummy, 0, sizeof(dummy));
perf(t1);
cblas_dgemm_scalaire( CblasNoTrans, CblasNoTrans ,m, m, m, 1, matriceA, m, matriceB, m, 1, matriceC, m);
perf(t2);
perf_diff(t1, t2);
mflops1 = perf_mflops(t2, m * m * m * 3 + m * m );
perf(t3);
cblas_dgemm_scalaire1(matriceC, m, matriceA, m, matriceB, m, m);
perf(t4);
perf_diff(t3, t4);
mflops2 = perf_mflops(t4, m * m * m * 3);
perf(t5);
cblas_dgemm_scalaire2(matriceC, m, matriceA, m, matriceB, m, m);
perf(t6);
perf_diff(t5, t6);
mflops3 = perf_mflops(t6, m * m * m * 3);
perf(t7);
cblas_dgemm_scalaire3(matriceC, m, matriceA, m, matriceB, m, m);
perf(t8);
perf_diff(t7, t8);
mflops4 = perf_mflops(t8, m * m * m * 3);
perf(t9);
cblas_dgemm(CblasColMajor, CblasTrans, CblasNoTrans, m, m,m, 1, matriceA, m, matriceB, m, 1, matriceC, m);
perf(t10);
perf_diff(t9, t10);
mflops5 = perf_mflops(t10, m * m * m * 3);
sprintf(command, "echo %d %lf %lf %lf %lf %lf >> results/dgemm_perf.txt", m * m, mflops1, mflops2, mflops3, mflops4, mflops5);
system(command);
printf("Mflops/s : %d %lf %lf %lf %lf %lf\n", m * m, mflops1, mflops2, mflops3, mflops4, mflops5 );
}
free(matriceA);
free(matriceB);
free(matriceC);
free(matriceD);
free(matriceE);
return EXIT_SUCCESS;
}
