void dedisperse_single(float **inin, float **outout, int nchan,int ndat)
{
//omp_set_num_threads(8);
int npass=get_npass(nchan);
//printf("need %d passes.\n",npass);
//npass=2;
int bs=nchan;
float **in=inin;
float **out=outout;
//FILE *fout;
//fout = fopen('/var/log/burst_bench.log', 'w');
//fclose(fout);
// omp_set_dynamic(0);
// omp_set_num_threads(8);
for (int i=0;i<npass;i++) {
#pragma omp parallel for
for (int j=0;j<nchan;j+=bs) {
//printf("dedisperse using %i threads\n",omp_get_num_threads());
dedisperse_kernel(in+j,out+j,bs,ndat);
}
bs/=2;
float **tmp=in;
in=out;
out=tmp;
}
memcpy(out[0],in[0],nchan*ndat*sizeof(float));
}
void dedisperse_dual(float **inin, float **outout, int nchan,int ndat)
{
int npass=get_npass(nchan);
//printf("need %d passes from %d channels..\n",npass,nchan);
//npass=2;
int bs=nchan;
float **in=inin;
float **out=outout;
//the npasss-1 is so that we stop in time to hand the final pass to
//the single-step kernel in the event of an odd depth.
for (int i=0;i<npass-1;i+=2) {
#pragma omp parallel for
for (int j=0;j<nchan;j+=bs) {
//dedisperse_kernel_2pass_v2(in+j,out+j,bs,ndat);
dedisperse_block_kernel_2pass((const float **)(in+j),out+j,bs,ndat);
}
bs/=4;
float **tmp=in;
in=out;
out=tmp;
}
if (npass%2==1) {
//do a single step if we come in with odd depth
//printf("doing final step for odd depth with block size %d.\n",bs);
#pragma omp parallel for
for (int j=0;j<nchan;j+=bs)
dedisperse_kernel(in+j,out+j,bs,ndat);
float **tmp=in;
in=out;
out=tmp;
}
memcpy(out[0],in[0],nchan*ndat*sizeof(float));
}
void dedisperse(float **inin, float **outout, int nchan,int ndat)
{
//return;
int npass=get_npass(nchan);
//printf("need %d passes.\n",npass);
//npass=2;
int bs=nchan;
float **in=inin;
float **out=outout;
for (int i=0;i<npass;i++) {
//#pragma omp parallel for
for (int j=0;j<nchan;j+=bs) {
dedisperse_kernel(in+j,out+j,bs,ndat);
}
bs/=2;
float **tmp=in;
in=out;
out=tmp;
}
memcpy(out[0],in[0],nchan*ndat*sizeof(float));
}
/*--------------------------------------------------------------------------------*/
void dedisperse_2pass(float **dat, float **dat2, int nchan, int ndat)
{
dedisperse_kernel(dat,dat2,nchan,ndat);
dedisperse_kernel(dat2,dat,nchan/2,ndat);
dedisperse_kernel(dat2+nchan/2,dat+nchan/2,nchan/2,ndat);
}
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");
}
