long double sum_to_all(long double in) {
long double out = in;
#ifdef HAVE_MPI
if (MPI_LONG_DOUBLE == MPI_DATATYPE_NULL)
out = sum_to_all(double(in));
else
MPI_Allreduce(&in,&out,1,MPI_LONG_DOUBLE,MPI_SUM,mycomm);
#endif
return out;
}
double *dft_ldos::ldos() const {
// we try to get the overall scale factor right (at least for a point source)
// so that we can compare against the analytical formula for testing
// ... in most practical cases, the scale factor won't matter because
// the user will compute the relative LDOS of 2 cases (e.g. LDOS/vacuum)
// overall scale factor
double Jsum_all = sum_to_all(Jsum);
double scale = 4.0/pi // from definition of LDOS comparison to power
* -0.5 // power = -1/2 Re[E* J]
/ (Jsum_all * Jsum_all); // normalize to unit-integral current
double *sum = new double[Nomega];
for (int i = 0; i < Nomega; ++i) /* 4/pi * work done by unit dipole */
sum[i] = scale * real(Fdft[i] * conj(Jdft[i])) / abs2(Jdft[i]);
double *out = new double[Nomega];
sum_to_all(sum, out, Nomega);
delete[] sum;
return out;
}
complex<long double> sum_to_all(complex<long double> in) {
complex<long double> out = in;
#ifdef HAVE_MPI
if (MPI_LONG_DOUBLE == MPI_DATATYPE_NULL) {
complex<double> dout;
dout = sum_to_all(complex<double>(double(in.real()), double(in.imag())));
out = complex<long double>(dout.real(), dout.imag());
}
else
MPI_Allreduce(&in,&out,2,MPI_LONG_DOUBLE,MPI_SUM,mycomm);
#endif
return out;
}
static double norm2(size_t n, const realnum *x) {
// note: we don't just do sqrt(dot(n, x, x)) in order to avoid overflow
size_t i;
double xmax = 0, scale;
long double sum = 0;
for (i = 0; i < n; ++i) {
double xabs = fabs(x[i]);
if (xabs > xmax) xmax = xabs;
}
xmax = max_to_all(xmax);
if (xmax == 0) return 0;
scale = 1.0 / xmax;
for (i = 0; i < n; ++i) {
double xs = scale * x[i];
sum += xs * xs;
}
return xmax * sqrt(sum_to_all(sum));
}
int
main(int argc, char* argv[])
{
int c, i, mype, num_pes, tests, passed;
char *pgm;
shmem_init();
mype = shmem_my_pe();
num_pes = shmem_n_pes();
if ((pgm=strrchr(argv[0],'/')))
pgm++;
else
pgm = argv[0];
while((c=getopt(argc,argv,"ampsSoxhv")) != -1) {
switch(c) {
case 'a':
And++; // do not run and_to_all
break;
case 'm':
Min++; // do not run min_to_all
break;
case 'o':
Or++; // do not run or_to_all
break;
case 'p':
Prod++; // do not run prod_to_all
break;
case 's':
Sum++; // do not run sum_to_all
break;
case 'x':
Xor++; // do not run xor_to_all
break;
case 'S':
Serialize++;
break;
case 'v':
Verbose++;
break;
case 'h':
default:
Rfprintf(stderr,"usage: %s {-v(verbose)|h(help)}\n",pgm);
shmem_finalize();
return 1;
}
}
for (i = 0; i < SHMEM_REDUCE_SYNC_SIZE; i++) {
pSync[i] = SHMEM_SYNC_VALUE;
pSync1[i] = SHMEM_SYNC_VALUE;
}
tests = passed = 0;
shmem_barrier_all();
passed += max_to_all(mype, num_pes);
tests++;
if (!Min) {
passed += min_to_all(mype, num_pes);
tests++;
}
if (!Sum) {
passed += sum_to_all(mype, num_pes);
tests++;
}
if (!And) {
passed += and_to_all(mype, num_pes);
tests++;
}
if (!Prod) {
passed += prod_to_all(mype, num_pes);
tests++;
}
if (!Or) {
passed += or_to_all(mype, num_pes);
tests++;
}
if (!Xor) {
passed += xor_to_all(mype, num_pes);
tests++;
}
c = 0;
if (mype == 0) {
if ((Verbose || tests != passed))
fprintf(stderr,"to_all[%d] %d of %d tests passed\n",
mype,passed,tests);
c = (tests == passed ? 0 : 1);
}
shmem_finalize();
return c;
}
void sum_to_all(const complex<float> *in, complex<double> *out, int size) {
sum_to_all((const float*) in, (double*) out, 2*size);
}
void sum_to_all(const float *in, double *out, int size) {
double *in2 = new double[size];
for (int i = 0; i < size; ++i) in2[i] = in[i];
sum_to_all(in2, out, size);
delete[] in2;
}
complex<double> *dft_ldos::J() const {
complex<double> *out = new complex<double>[Nomega];
sum_to_all(Jdft, out, Nomega);
return out;
}
static double dot(size_t n, const realnum *x, const realnum *y) {
double sum = 0;
for (size_t i = 0; i < n; ++i)
sum += x[i] * y[i];
return sum_to_all(sum);
}
