/* sign function as defined in http://www.netlib.org/lapack/lawnspdf/lawn148.pdf */
static vsip_scalar_f sign_f(vsip_scalar_f a_in)
{
if(a_in < 0.0)
return -1.0;
else
return 1.0;
} /* same */
static void biDiagPhaseToZero_f( svdObj_f *svd)
{
vsip_mview_f *L = svd->L;
vsip_vview_f *d = svd->d;
vsip_vview_f *f = svd->f;
vsip_mview_f *R = svd->R;
vsip_scalar_f eps0 = svd->eps0;
vsip_length n_d=vsip_vgetlength_f(d);
vsip_length n_f=vsip_vgetlength_f(f);
vsip_index i,j;
vsip_scalar_f ps;
vsip_scalar_f m;
vsip_vview_f *l = svd->ls_one;
vsip_vview_f *r = svd->rs_one;
for(i=0; i<n_d; i++){
ps=vsip_vget_f(d,i);
m = vsip_mag_f(ps);
ps=sign_f(ps);
if(m > eps0){
col_sv_f(L,l,i);vsip_svmul_f(ps,l,l);
vsip_vput_f(d,i,m);
if (i < n_f)
vsip_vput_f(f,i,ps*vsip_vget_f(f,i));
} else {
vsip_vput_f(d,i,0.0);
}
}
svdZeroCheckAndSet_f(eps0,d,f);
for (i=0; i<n_f-1; i++){
j=i+1;
ps = vsip_vget_f(f,i);
m = vsip_mag_f(ps);
ps=sign_f(ps);
col_sv_f(L, l, j);vsip_svmul_f(ps,l,l);
row_sv_f(R,r,j);vsip_svmul_f(ps,r,r);
vsip_vput_f(f,i,m);
vsip_vput_f(f,j,ps * vsip_vget_f(f,j));
}
j=n_f;
i=j-1;
ps=vsip_vget_f(f,i);
m=vsip_mag_f(ps);
ps=sign_f(ps);
vsip_vput_f(f,i,m);
col_sv_f(L, l, j);vsip_svmul_f(ps,l,l);
row_sv_f(R,r,j);vsip_svmul_f(ps,r,r);
}
static void phaseCheck_f(svdObj_f *svd)
{
biDiagPhaseToZero_f(svd);
}
void houseProd_f(vsip_vview_f *v, vsip_mview_f *A)
{
vsip_mattr_f a_atr;
vsip_vview_f *w;
vsip_mview_f *B;
vsip_mgetattrib_f(A,&a_atr);
B=vsip_mcreate_f(a_atr.col_length,a_atr.row_length,VSIP_ROW,VSIP_MEM_NONE);
w = vsip_vcreate_f(a_atr.row_length,VSIP_MEM_NONE);
vsip_scalar_f beta = 2.0/vsip_vdot_f(v,v);
vsip_vmprod_f(v,A,w);
vsip_vouter_f(beta,v,w,B);
vsip_msub_f(A,B,A);
vsip_valldestroy_f(w);
vsip_malldestroy_f(B);
} /* need to remove create */
void prodHouse_f(vsip_mview_f *A, vsip_vview_f *v)
{
vsip_mattr_f a_atr;
vsip_vview_f *w;
vsip_mview_f *B;
vsip_mgetattrib_f(A,&a_atr);
B=vsip_mcreate_f(a_atr.col_length,a_atr.row_length,VSIP_ROW,VSIP_MEM_NONE);
w = vsip_vcreate_f(a_atr.col_length,VSIP_MEM_NONE);
vsip_scalar_f beta = 2.0/vsip_vdot_f(v,v);
vsip_mvprod_f(A,v,w);
vsip_vouter_f(beta,w,v,B);
vsip_msub_f(A,B,A);
vsip_valldestroy_f(w);
vsip_malldestroy_f(B);
} /* need to remove create */
int main() {
vsip_init((void*)0);{
vsip_mview_f *A = vsip_mcreate_f(NN,NN,VSIP_COL,0);
vsip_vview_f *x0 = vsip_vcreate_f(NN,0);
vsip_vview_f *b = vsip_vcreate_f(NN,0);
vsip_mview_f *X = vsip_mcreate_f(NN,3,VSIP_ROW,0);
vsip_mview_f *XT = vsip_mcreate_f(NN,3,VSIP_COL,0);
vsip_vramp_f(1,1,x0); vsip_vmul_f(x0,x0,x0);
vsip_mput_f(A,0,0,-3); vsip_mput_f(A,0,1,7); vsip_mput_f(A,0,2,10); vsip_mput_f(A,0,3,12);
vsip_mput_f(A,1,0,0); vsip_mput_f(A,1,1,13); vsip_mput_f(A,1,2,18); vsip_mput_f(A,1,3,6);
vsip_mput_f(A,2,0,2); vsip_mput_f(A,2,1,-9); vsip_mput_f(A,2,2,6); vsip_mput_f(A,2,3,3);
vsip_mput_f(A,3,0,1); vsip_mput_f(A,3,1,2); vsip_mput_f(A,3,2,3); vsip_mput_f(A,3,3,4);
{int i,j; printf("A = [\n"); for(i=0; i<NN; i++){
for(j=0; j<NN; j++) printf("%9.2f%s",vsip_mget_f(A,i,j),(j == NN-1) ? "":",");
printf(";\n");
}
printf("]\n");
}
{ int k;
vsip_vview_f *x;
vsip_length L = vsip_mgetrowlength_f(X);
for(k=0; k<L; k++){
x = vsip_mcolview_f(X,k);
vsip_mvprod_f(A,x0,b);
vsip_vcopy_f_f(b,x);
vsip_svmul_f(2.0,x0,x0);
vsip_vdestroy_f(x);
}
{int i,j; printf("X = [\n");for(i=0; i<NN; i++){
for(j=0; j<3; j++) printf("%9.2f%s",vsip_mget_f(X,i,j),(j == 2) ? "":",");
printf(";\n");
}
printf("]\n");
}
{
vsip_lu_f* luAop = vsip_lud_create_f(NN);
vsip_mcopy_f_f(X,XT);
if(luAop == NULL) exit(1);
vsip_lud_f(luAop,A);
{ vsip_lu_attr_f attr;
vsip_lud_getattr_f(luAop,&attr);
printf("lud size %lu\n",attr.n);
}
vsip_lusol_f(luAop,0,X);
vsip_lusol_f(luAop,1,XT);
vsip_lud_destroy_f(luAop);
}
}
{int i,j; printf("A\\X\n"); for(i=0; i<NN; i++){
for(j=0; j<3; j++) printf("%9.2f%s",vsip_mget_f(X,i,j),(j == 2) ? "":",");
printf(";\n");
}
}
{int i,j; printf("A'\\X\n"); for(i=0; i<NN; i++){
for(j=0; j<3; j++) printf("%9.2f%s",vsip_mget_f(XT,i,j),(j == 2) ? "":",");
printf(";\n");
}
}
{
vsip_valldestroy_f(b);
vsip_valldestroy_f(x0);
vsip_malldestroy_f(X);
vsip_malldestroy_f(A);
}
}vsip_finalize((void*)0);return 1;
}
int main(){vsip_init((void*)0);
{
vsip_mview_f *Adummy = vsip_mcreate_f(5*NN,5*NN,VSIP_COL,0);
vsip_mview_f *A = vsip_msubview_f(Adummy,3,2,NN,NN);
vsip_vview_f *x0 = vsip_vcreate_f(NN,0);
vsip_mview_f *X = vsip_mcreate_f(NN,3,VSIP_ROW,0);
vsip_mview_f *XT = vsip_mcreate_f(NN,3,VSIP_COL,0);
vsip_mputrowstride_f(A,2*vsip_mgetrowstride_f(A));
vsip_mputcolstride_f(A,3*vsip_mgetcolstride_f(A));
/* matrix data */
vsip_mput_f(A,0,0,0.5); vsip_mput_f(A,0,1,7); vsip_mput_f(A,0,2,10); vsip_mput_f(A,0,3,12);
vsip_mput_f(A,0,4,-3); vsip_mput_f(A,0,5,0); vsip_mput_f(A,0,6,.05);
vsip_mput_f(A,1,0,2); vsip_mput_f(A,1,1,13); vsip_mput_f(A,1,2,18); vsip_mput_f(A,1,3,6);
vsip_mput_f(A,1,4,0); vsip_mput_f(A,1,5,130); vsip_mput_f(A,1,6,8);
vsip_mput_f(A,2,0,3); vsip_mput_f(A,2,1,-9); vsip_mput_f(A,2,2,2); vsip_mput_f(A,2,3,3);
vsip_mput_f(A,2,4,2); vsip_mput_f(A,2,5,-9); vsip_mput_f(A,2,6,6);
vsip_mput_f(A,3,0,4); vsip_mput_f(A,3,1,2); vsip_mput_f(A,3,2,2); vsip_mput_f(A,3,3,4);
vsip_mput_f(A,3,4,1); vsip_mput_f(A,3,5,2); vsip_mput_f(A,3,6,3);
vsip_mput_f(A,4,0,.2); vsip_mput_f(A,4,1,2); vsip_mput_f(A,4,2,9); vsip_mput_f(A,4,3,4);
vsip_mput_f(A,4,4,1); vsip_mput_f(A,4,5,2); vsip_mput_f(A,4,6,3);
vsip_mput_f(A,5,0,.1); vsip_mput_f(A,5,1,2); vsip_mput_f(A,5,2,.3); vsip_mput_f(A,5,3,4);
vsip_mput_f(A,5,4,1); vsip_mput_f(A,5,5,2); vsip_mput_f(A,5,6,3);
vsip_mput_f(A,6,0,.01); vsip_mput_f(A,6,1,.2); vsip_mput_f(A,6,2,3); vsip_mput_f(A,6,3,4);
vsip_mput_f(A,6,4,1); vsip_mput_f(A,6,5,2); vsip_mput_f(A,6,6,3);
{int i,j; printf("A = [\n"); for(i=0; i<NN; i++){
for(j=0; j<NN; j++) printf("%9.2f%s",vsip_mget_f(A,i,j),(j == NN-1) ? "":",");
printf(";\n");
}
printf("]\n");
}
{ /* were solving for NTRANS Ax = B */
/* use a known X, calculate B using Ax */
int k;
vsip_vview_f *x;
vsip_mview_f *AT = vsip_mtransview_f(A);
vsip_length L = vsip_mgetrowlength_f(X);
vsip_vramp_f(1,1,x0);
for(k=0; k<L; k++){
x = vsip_mcolview_f(X,k);
vsip_mvprod_f(A,x0,x);
vsip_svmul_f(2.0,x0,x0);
vsip_vdestroy_f(x);
}
vsip_vramp_f(1,1,x0);
for(k=0; k<L; k++){
x = vsip_mcolview_f(XT,k);
vsip_mvprod_f(AT,x0,x);
vsip_svmul_f(2.0,x0,x0);
vsip_vdestroy_f(x);
}
vsip_mdestroy_f(AT);
printf("X = 1 2 4\n 2 4 8\n 3 6 12\n 4 8 16\n 5 10 20\n 6 12 24\n 7 14 28\n");
{
int i,j; printf("B = [\n");for(i=0; i<NN; i++){
for(j=0; j<3; j++) printf("%9.2f%s",vsip_mget_f(X,i,j),(j == 2) ? "":",");
printf(";\n");
}
printf("]\n");
}
{
/* then solve for B to see if we get X back */
vsip_lu_f* luAop = vsip_lud_create_f(NN);
if(luAop == NULL) exit(1);
vsip_lud_f(luAop,A);
{ vsip_lu_attr_f attr;
vsip_lud_getattr_f(luAop,&attr);
printf("lud size %lu\n",attr.n);
}
vsip_lusol_f(luAop,VSIP_MAT_NTRANS,X);
vsip_lusol_f(luAop,VSIP_MAT_TRANS,XT);
vsip_lud_destroy_f(luAop);
}
}
{int i,j; printf("A\\X\n"); for(i=0; i<NN; i++){
for(j=0; j<3; j++) printf("%9.2f%s",vsip_mget_f(X,i,j),(j == 2) ? "":",");
printf(";\n");
}
}
{int i,j; printf("A'\\X\n"); for(i=0; i<NN; i++){
for(j=0; j<3; j++) printf("%9.2f%s",vsip_mget_f(XT,i,j),(j == 2) ? "":",");
printf(";\n");
}
}
{
vsip_malldestroy_f(XT);
vsip_valldestroy_f(x0);
vsip_malldestroy_f(X);
vsip_mdestroy_f(A);
vsip_malldestroy_f(Adummy);
}
} vsip_finalize((void*)0); return 1;
}
int main(){
int init = vsip_init((void*)0);
int i,j, cholsol_retval,chold_retval;
double t0 = VU_ansi_c_clock(); /* for doing some timeing */
vsip_cscalar_f czero = vsip_cmplx_f((vsip_scalar_f)0.0,(vsip_scalar_f)0.0);
vsip_cmview_f *A = vsip_cmcreate_f(N,N,VSIP_COL,0);
vsip_cmview_f *RU = vsip_cmcreate_f(N,N,VSIP_COL,0);
vsip_cmview_f *RL = vsip_cmcreate_f(N,N,VSIP_COL,0);
vsip_cmview_f *XB = vsip_cmcreate_f(N,M,VSIP_ROW,0);
vsip_cchol_f* chol = vsip_cchold_create_f(UPORLO,N); /* NOTE: UPORLO macro above main() */
/* to make sure we have a valid Positive Symetric define */
/* an upper triangular (RU) with positive pivots and */
/* zero below the main diagonal. */
/* Then initialize RL with hermitian of RU */
/* finally create A as the matrix product of RL and RU */
/* Initialise matrix RU */
/* time this */
t0 = VU_ansi_c_clock();
for (i=0; i<N; i++){
for(j = i; j < N; j++){
#ifdef OBNOXIOUS
/* make up some reasonably obnoxious data */
vsip_scalar_f a = cos(1.5/((j+1)*(i+1)))+sqrt(i*j);
vsip_scalar_f b = (i + j + 1) * cos(M_PI * a);
#else
/* the above was to obnoxious for bigger than about N = 10 */
/* the following works for N > 100 */
vsip_scalar_f a = 1; vsip_scalar_f b = 1;
#endif
if(i == j) /* fill diagonal */
vsip_cmput_f(RU,i,j, vsip_cmplx_f(sqrt(N) + sqrt(i),0));
else { /* fill off diagonal */
vsip_cmput_f(RU,i,j,vsip_cmplx_f(b,a));
vsip_cmput_f(RU,j,i,czero);
}
}
}
/* initialize RL */
vsip_cmherm_f(RU,RL);
#ifdef PRINT
VU_cmprintm_f("7.4",RU);
VU_cmprintm_f("7.4",RL);
#endif
printf("Matrix initialize for RU and RL = %f seconds\n",VU_ansi_c_clock() - t0);
/* initialize A */
/* this step will take a long time so time it */
t0 = VU_ansi_c_clock();
vsip_cmprod_f(RL,RU,A);
#ifdef OBNOXIOUS
for(i=0; i<N; i++){
vsip_cvview_f *aview = vsip_cmrowview_f(A,i);
vsip_cvrsdiv_f(aview,vsip_cmag_f(vsip_cvmeanval_f(aview)),aview);
vsip_cvdestroy_f(aview);
}
#endif
printf("Matrix multiply for initialization of A = %f seconds\n",VU_ansi_c_clock() - t0);
/* print A */
/* we only want to do this if A is something reasonable to print */
/* selected as an option in the make file */
#ifdef PRINT
printf("Matrix A =\n");
VU_cmprintm_f("4.2",A);
fflush(stdout);
#endif
/* initialise rhs */
/* start out with XB = {1,2,3,...,M} */
/* calculate what B must be using A */
/* then solve to see if we get XB back */
{ vsip_index i;
vsip_vview_f *y = vsip_vcreate_f(vsip_cmgetcollength_f(A),VSIP_MEM_NONE);
vsip_vview_f *x_r,*x_i;
vsip_cvview_f *x;
vsip_mview_f *A_r = vsip_mrealview_f(A),
*A_i = vsip_mimagview_f(A);
/* time this */
t0 = VU_ansi_c_clock();
for(i=0; i<M; i++){
vsip_vfill_f((vsip_scalar_f)i+1.0,y);
x = vsip_cmcolview_f(XB,i);
x_r = vsip_vrealview_f(x);
x_i = vsip_vimagview_f(x);
vsip_mvprod_f(A_r,y,x_r);
vsip_mvprod_f(A_i,y,x_i);
vsip_cvdestroy_f(x);
vsip_vdestroy_f(x_r);
vsip_vdestroy_f(x_i);
}
vsip_mdestroy_f(A_r);
vsip_mdestroy_f(A_i);
printf("Matrix init for B = %f seconds\n",VU_ansi_c_clock() - t0);
}
/* print XB */
/* we only want to do this if XB is something reasonable to print */
/* selected as an option in the make file */
#ifdef PRINT
printf("Matrix B = \n");
VU_cmprintm_f("7.4",XB);
fflush(stdout);
#endif
if(chol != NULL){
t0 = VU_ansi_c_clock(); /* we want to time the decomposition */
chold_retval = vsip_cchold_f(chol,A);
printf("time decomp %f\n",VU_ansi_c_clock() - t0);
printf("decompostion returns %d\n",chold_retval);
/* now do the solution */
t0 = VU_ansi_c_clock(); /* we want to time the solution */
cholsol_retval=vsip_ccholsol_f(chol,XB);
printf("time solution %f\n",VU_ansi_c_clock() - t0);
printf("cholsol returns %d\n",cholsol_retval);
/* print XB */
/* we only want to do this if XB is something reasonable to print */
/* selected as an option in the make file; otherwise */
/* we print a single row of XB if the matrix is to large since */
/* M is usally reasonable. Printed as a column vector */
#ifdef PRINT
printf("Matrix X = \n");
VU_cmprintm_f("7.4",XB);
fflush(stdout);
#else
{ /* pick a row in the middle */
vsip_cvview_f *x = vsip_cmrowview_f(XB,N/2);
printf("This output sould be 1,2,...,M\n");
VU_cvprintm_f("7.4",x);
fflush(stdout);
vsip_cvdestroy_f(x);
}
#endif
} else {
printf("failed to create cholesky object \n");
}
vsip_cmalldestroy_f(XB);
vsip_cmalldestroy_f(A);
vsip_cmalldestroy_f(RL);
vsip_cmalldestroy_f(RU);
vsip_cchold_destroy_f(chol);
vsip_finalize((void*)0);
return 1;
}