void wls_cpu(int_t y_cols, int_t y_rows, double* wts, double* y, double* out_beta_cpu){
double *xtwx = (double *)calloc((y_rows+y_cols-1)*(y_rows+y_cols-1),sizeof(double));
double *xtwy = (double *)calloc((y_rows+y_cols),sizeof(double));
//XTWX start
XTWX(y_rows, y_cols, wts, xtwx);
// printf("\nXTWX done\n"); fflush(stdout);
//XTWXinv
XTWXinv(y_rows, y_cols, xtwx);
// printf("\nXTWXinv done\n"); fflush(stdout);
//XTWY
XTWY(y_rows, y_cols, wts, y, xtwy);
// printf("\nXTWY done\n"); fflush(stdout);
int i, j;
//OUT_BETA
for (i=0; i < y_rows+y_cols-1; i++){
out_beta_cpu[i] = 0.0;
for (j=0; j < y_rows+y_cols-1; j++){
out_beta_cpu[i] += xtwx[j*(y_rows+y_cols -1)+i]*xtwy[j];
}
}
free(xtwx);
free(xtwy);
return 0;
}
int main(){
struct timeval start, end;
long utime;
int_t y_rows = Y_ROWS;
int_t y_cols = Y_COLS;
double *wts = (double *)calloc((y_rows*y_cols),sizeof(double));
double *y = (double *)calloc((y_rows*y_cols),sizeof(double));
printf("testing here\n");
#ifdef CPU_COMPUTE
double *xtwx = (double *)calloc((y_rows+y_cols-1)*(y_rows+y_cols-1),sizeof(double));
double *xtwy = (double *)calloc((y_rows+y_cols),sizeof(double));
double *out_beta = (double *)calloc((y_rows+y_cols),sizeof(double));
#endif
double *out_beta_gpu = (double *)calloc((y_rows+y_cols),sizeof(double));
int_t i,j;
/* initialize random seed: */
srand (time(NULL));
for(j=0; j<y_cols; j++){
for(i=0; i<y_rows; i++){
wts[j*y_rows+i] = rand()%(RANDOM_MAX - RANDOM_MIN + 1) + RANDOM_MIN;
wts[j*y_rows+i]=1.0/wts[j*y_rows+i];
y[j*y_rows+i] = rand()%(RANDOM_MAX - RANDOM_MIN + 1) + RANDOM_MIN;
y[j*y_rows+i]=1.0/y[j*y_rows+i];
}
}
#ifdef CPU_COMPUTE
printf("\n Starting CPU Computation\n");
gettimeofday(&start, NULL);
//XTWX start
XTWX(y_rows,y_cols,wts,xtwx);
printf("\nXTWX done\n"); fflush(stdout);
//XTWXinv
XTWXinv(y_rows, y_cols,xtwx);
printf("\nXTWXinv done\n"); fflush(stdout);
//XTWY
XTWY(y_rows, y_cols, wts,y, xtwy);
printf("\nXTWY done\n"); fflush(stdout);
//OUT_BETA
for (i=0;i < y_rows+y_cols-1; i++){
out_beta[i] = 0.0;
for (j=0;j < y_rows+y_cols -1; j++){
out_beta[i] += xtwx[j*(y_rows+y_cols -1)+i]*xtwy[j];
}
}
gettimeofday(&end, NULL);
utime = ((end.tv_sec - start.tv_sec) * 1000000 + end.tv_usec - start.tv_usec);
printf("\n CPU Computation done \n");
printf("\nTime CPU = %ld us\n",utime);
#endif
printf("\n Starting GPU Computation\n");
//GPU start
gettimeofday(&start, NULL);
wls_gpu(y_cols, y_rows, wts, y, out_beta_gpu);
gettimeofday(&end, NULL);
utime = ((end.tv_sec - start.tv_sec) * 1000000 + end.tv_usec - start.tv_usec);
printf("\n GPU Computation done \n");
printf("\nTime GPU = %ld us\n",utime);
#if 0
//check
int_t M_size = y_cols+y_rows-1;
//check for A
double Ainv_err = 0.0;
for(i=0; i<y_cols; i++){
Ainv_err+=fabs(((1.0/h_Ainv[i]) - xtwx[i*M_size+i])/xtwx[i*M_size+i]);
}
Ainv_err/=y_cols;
printf("\nError Ainv = %e\n",Ainv_err);
//check for B
double B_err = 0.0;
for(j=0; j<(y_rows-1); j++){
for(i=0; i<y_cols; i++){
B_err+=fabs((h_B[j*y_cols+i] - xtwx[(y_cols+j)*M_size+i])/xtwx[(y_cols+j)*M_size+i]);
}
}
B_err/=(y_cols*(y_rows-1));
printf("\nError B = %e\n",B_err);
//check for D
double D_err = 0.0;
for(j=0; j<(y_rows-1); j++){
for(i=0; i<(y_rows-1); i++){
D_err+=fabs((h_D[j*(y_rows-1)+i] - xtwx[(y_cols+j)*M_size+y_cols+i])/xtwx[(y_cols+j)*M_size+y_cols+i]);
}
}
D_err/=((y_rows-1)*(y_rows-1));
printf("\nError D = %e\n",D_err);
//check for Q
double Q_err = 0.0;
for(j=0; j<(y_rows-1); j++){
for(i=0; i<y_cols; i++){
Q_err+=fabs((h_Q[j*y_cols+i] - xtwx[(y_cols+j)*M_size+i])/xtwx[(y_cols+j)*M_size+i]);
}
}
Q_err/=(y_cols*(y_rows-1));
printf("\nError Q = %e\n",Q_err);
//check for S
double S_err = 0.0;
for(j=0; j<(y_rows-1); j++){
for(i=0; i<(y_rows-1); i++){
S_err+=fabs((h_S[j*(y_rows-1)+i] - xtwx[(y_cols+j)*M_size+y_cols+i])/xtwx[(y_cols+j)*M_size+y_cols+i]);
}
}
S_err/=((y_rows-1)*(y_rows-1));
printf("\nError S = %e\n",S_err);
#endif
#ifdef CPU_COMPUTE
//check for out_beta
double out_beta_err = 0.0;
for(j=0; j<(y_rows+y_cols-1); j++){
out_beta_err+=fabs((out_beta_gpu[j] - out_beta[j])/out_beta[j]);
}
out_beta_err/=(y_rows+y_cols-1);
printf("\nError Out_beta_error = %e\n",out_beta_err);
#endif
free(wts);
free(y);
#ifdef CPU_COMPUTE
free(xtwx);
free(xtwy);
free(out_beta);
#endif
free(out_beta_gpu);
return 0;
}
void XTWY_R(int *rows, int *cols, double *out_weights, double *y,double *xtwy){
XTWY(*rows, *cols, out_weights,y,xtwy);
}
void rlm_wfit_anova_engine(double *y, int y_rows, int y_cols, double *input_scale, double *w, double *out_beta, double *out_resids, double *out_weights,double (* PsiFn)(double, double, int), double psi_k,int max_iter, int initialized){
int i,j,iter;
/* double tol = 1e-7; */
double acc = 1e-4;
double scale =0.0;
double conv;
double endprobe;
double *wts = out_weights;
double *resids = out_resids;
double *old_resids = Calloc(y_rows*y_cols,double);
double *rowmeans = Calloc(y_rows,double);
double *xtwx = Calloc((y_rows+y_cols-1)*(y_rows+y_cols-1),double);
double *xtwy = Calloc((y_rows+y_cols),double);
double sumweights, rows;
rows = y_rows*y_cols;
if (!initialized){
/* intially use equal weights */
for (i=0; i < rows; i++){
wts[i] = w[i]*1.0;
}
}
/* starting matrix */
for (i=0; i < y_rows; i++){
for (j=0; j < y_cols; j++){
resids[j*y_rows + i] = y[j*y_rows + i];
}
}
/* sweep columns (ie chip effects) */
for (j=0; j < y_cols; j++){
out_beta[j] = 0.0;
sumweights = 0.0;
for (i=0; i < y_rows; i++){
out_beta[j] += wts[j*y_rows + i]* resids[j*y_rows + i];
sumweights += wts[j*y_rows + i];
}
out_beta[j]/=sumweights;
for (i=0; i < y_rows; i++){
resids[j*y_rows + i] = resids[j*y_rows + i] - out_beta[j];
}
}
/* sweep rows (ie probe effects) */
for (i=0; i < y_rows; i++){
rowmeans[i] = 0.0;
sumweights = 0.0;
for (j=0; j < y_cols; j++){
rowmeans[i] += wts[j*y_rows + i]* resids[j*y_rows + i];
sumweights += wts[j*y_rows + i];
}
rowmeans[i]/=sumweights;
for (j=0; j < y_cols; j++){
resids[j*y_rows + i] = resids[j*y_rows + i] - rowmeans[i];
}
}
for (i=0; i < y_rows-1; i++){
out_beta[i+y_cols] = rowmeans[i];
}
for (iter = 0; iter < max_iter; iter++){
if (*input_scale < 0){
scale = med_abs(resids,rows)/0.6745;
} else {
scale = *input_scale;
}
if (fabs(scale) < 1e-10){
/*printf("Scale too small \n"); */
break;
}
for (i =0; i < rows; i++){
old_resids[i] = resids[i];
}
for (i=0; i < rows; i++){
wts[i] = w[i]*PsiFn(resids[i]/scale,psi_k,0); /* psi_huber(resids[i]/scale,k,0); */
}
/* printf("%f\n",scale); */
/* weighted least squares */
memset(xtwx,0,(y_rows+y_cols-1)*(y_rows+y_cols-1)*sizeof(double));
XTWX(y_rows,y_cols,wts,xtwx);
XTWXinv(y_rows, y_cols,xtwx);
XTWY(y_rows, y_cols, wts,y, xtwy);
for (i=0;i < y_rows+y_cols-1; i++){
out_beta[i] = 0.0;
for (j=0;j < y_rows+y_cols -1; j++){
out_beta[i] += xtwx[j*(y_rows+y_cols -1)+i]*xtwy[j];
}
}
/* residuals */
for (i=0; i < y_rows-1; i++){
for (j=0; j < y_cols; j++){
resids[j*y_rows +i] = y[j*y_rows + i]- (out_beta[j] + out_beta[i + y_cols]);
}
}
for (j=0; j < y_cols; j++){
endprobe=0.0;
for (i=0; i < y_rows-1; i++){
endprobe+= out_beta[i + y_cols];
}
resids[j*y_rows + y_rows-1] = y[j*y_rows + y_rows-1]- (out_beta[j] - endprobe);
}
/*check convergence based on residuals */
conv = irls_delta(old_resids,resids, rows);
if (conv < acc){
/* printf("Converged \n");*/
break;
}
}
if (*input_scale < 0){
scale = med_abs(resids,rows)/0.6745;
} else {
scale = *input_scale;
}
Free(xtwx);
Free(xtwy);
Free(old_resids);
Free(rowmeans);
input_scale[0] = scale;
}
