void dedisperse_gbt_jon(Data *dat, float *outdata)
{
//remap_data(dat);
//float **tmp=matrix(dat->nchan, dat->ndata);
float **tmp=(float **)malloc(sizeof(float *)*dat->nchan);
for (int i=0;i<dat->nchan;i++)
tmp[i]=outdata+i*dat->ndata;
memset(tmp[0],0,sizeof(tmp[0][0])*dat->ndata*dat->nchan);
double t1=omp_get_wtime();
//dedisperse_dual(dat->data,tmp,dat->nchan,dat->ndata);
//memcpy(ip_dat[0],dat->data[0],dat->nchan*dat->ndata*sizeof(dat->data[0][0]));
//dedisperse_inplace(ip_dat,dat->nchan,dat->ndata);
//dedisperse_single(dat->data,tmp,dat->nchan,dat->ndata);
dedisperse_blocked_cached(dat->data,tmp,dat->nchan,dat->ndata);
//printf("took %12.4f seconds to dedisperse.\n",omp_get_wtime()-t1);
for(int i = 0; i < dat->nchan; i++){
memcpy(outdata + i*dat->ndata,dat->data[i],dat->ndata*sizeof(outdata[0]));
}
free(tmp);
if (0) {
printf("element 500,300 is %12.5e\n",dat->data[500][300]);
FILE *outfile=fopen("burst_dump.dat","w");
fwrite(&(dat->nchan),sizeof(dat->nchan),1,outfile);
fwrite(&(dat->ndata),sizeof(dat->ndata),1,outfile);
fwrite(outdata,sizeof(outdata[0]),dat->nchan*dat->ndata,outfile);
fclose(outfile);
}
}
int main(int argc, char *argv[])
{
//int nchan=4096;
//int ndat=12000;
int nchan=1024;
int ndat=327680;
int nrep=1;
if (argc>1)
nchan=atoi(argv[1]);
if (argc>2)
ndat=atoi(argv[2]);
if (argc>3)
nrep=atoi(argv[3]);
float **dat=matrix(nchan,ndat+nchan);
float **dat2=matrix(nchan,ndat+nchan);
if (1)
for (int i=0;i<nchan;i++)
dat[i][(int)(0.8317*i+160.2)]=1;
else
for (int i=0;i<nchan;i++)
dat[i][ndat/2]=1;
#if 0
write_mat(dat,nchan,ndat,"dat_starting.dat");
dedisperse_kernel(dat,dat2,nchan,ndat);
write_mat(dat2,nchan,ndat,"dat_1pass.dat");
dedisperse_2pass(dat,dat2,nchan,ndat);
write_mat(dat,nchan,ndat,"dat_2pass.dat");
#endif
double t1=omp_get_wtime();
//dedisperse(dat,dat2,nchan,ndat);
//dedisperse_blocked(dat,dat2,nchan,ndat);
dedisperse_blocked_cached(dat,dat2,nchan,ndat);
double t2=omp_get_wtime();
printf("took %12.4f seconds.\n",t2-t1);
int ichan,idat;
find_peak(dat,nchan,ndat,&ichan,&idat);
t1=omp_get_wtime();
printf("took %12.4f seconds to find peak.\n",t1-t2);
for (int i=0;i<10;i++) {
t1=omp_get_wtime();
for (int j=0;j<nrep;j++) {
dedisperse_blocked_cached(dat,dat2,nchan,ndat);
//dedisperse(dat,dat2,nchan,ndat);
}
t2=omp_get_wtime();
double nops=get_npass(nchan)*(nchan+0.0)*(ndat+0.0)*(nrep+0.0);
printf("took %12.6f seconds at rate %12.6f.\n",t2-t1,nops/(t2-t1)/1024/1024);
//printf("took %12.4f seconds.\n",t2-t1);
}
//write_mat(dat,nchan,ndat,"dat_final1.dat");
//write_mat(dat2,nchan,ndat,"dat_final2.dat");
}