Skip to content

Commit

Permalink
Working version. Incomplete Yukawa potential and madelung constant. I…
Browse files Browse the repository at this point in the history 
…ll-conditioning in Yukawa lattice computation
  • Loading branch information
@kcroker
kcroker committed Jan 22, 2016
1 parent 60cf031 commit ef42f9f
Show file tree
Hide file tree
Showing 4 changed files with 68 additions and 60 deletions.
8 changes: 2 additions & 6 deletions forcetree.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1826,8 +1826,7 @@ int force_treeevaluate_shortrange(int target, int mode)

sG = -1;

// XXX!
if(0) //r2min > rcut2)
if(r2min > rcut2)
{
/* check whether we can stop walking along this branch */
eff_dist = rcut + 0.5 * nop->len;
Expand Down Expand Up @@ -2124,9 +2123,6 @@ int force_treeevaluate_lattice_correction(int target, int mode, double pos_x, do
r2[sG] = dx[sG] * dx[sG] + dy[sG] * dy[sG] + dz[sG] * dz[sG];

// KC 1/5/16
// XXX
// WHOA BUDDY
// r2min, r2max are not set if we only have an actual particle!
r2min = r2max = r2[sG];
}
else
Expand Down Expand Up @@ -3273,7 +3269,7 @@ void force_treeallocate(int maxnodes, int maxpart)
// Anything above this is oversampling, because the integrand MUST be zero as far as the machine is concerned
// (because the normalized greens is bounded above by 1).
//
ngravsPeriodicTable = ngravsConvolutionInit(NTAB, 3, 8);
ngravsPeriodicTable = ngravsConvolutionInit(NTAB, 4, 12);
Z = 1.0/2.0;

if(!ThisTask)
Expand Down
75 changes: 36 additions & 39 deletions ngravs.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -824,19 +824,8 @@ double ewald_psi(double x[3])
*
*/

// KC 1/10/16
// XXX
// ym/2 should be on all Yukawa quantities so that a Gauss computation
// gives a unit enclosed charge!!
//
// For some reason though, puting this in really fucks everything up
//

/*! A pure Yukawa force
*
* Radii take values in [0, BoxLength] for periodic, unconstrained otherwise.
* If we are in PERIODIC mode, so that the tables do not become sensitive to
* the BoxLength as given in the configuration files, YUKAWA_IMASS is in units of EN.
*/
double yukawa(double target, double source, double h, double r, long N) {

Expand All @@ -850,16 +839,20 @@ double yukawa(double target, double source, double h, double r, long N) {
*/
double pgyukawa(double target, double source, double k2, double k, long N) {

//double ym = YUKAWA_IMASS/(2*M_PI);
double ym = YUKAWA_IMASS/(2*M_PI);
return 1.0 / (k2 + ym*ym);
double asmth2;

asmth2 = (2 * M_PI) * All.Asmth[0] / All.BoxSize;
asmth2 *= asmth2;

return 1.0 / (k2 + ym*ym) * exp(-ym*ym*asmth2);
}

double normed_pgyukawa(double target, double source, double k2, double k, long N) {

// This converts from PMGRID units into shortrange interpolation table units
double ym = gridKtoNormK(YUKAWA_IMASS/(2*M_PI));
return k2 / (k2 + ym*ym);
return k2 / (k2 + ym*ym) * exp(-ym*ym*0.25);
}

/*! This function computes the Madelung constant for the yukawa potential
Expand Down Expand Up @@ -1003,10 +996,11 @@ void yukawa_lattice_force(int iii, int jjj, int kkk, double x[3], double force[3
int i, h[3], n[3], h2;
double ym;
double fac;

// KC 11/16/15
// Note our use of Salin's optimal 'alpha', and our excessive momentum-space
alpha = 5.64;
//alpha = 1.0;

if(iii == 0 && jjj == 0 && kkk == 0)
return;
Expand Down Expand Up @@ -1039,6 +1033,12 @@ void yukawa_lattice_force(int iii, int jjj, int kkk, double x[3], double force[3
// Looks like this takes the first four images out in position space in each direction (so
// bracketing by 8 overall)

// KC 1/19/16
// This does not group expressions with the same order of n together in additions.
// Change so that we do.
//
// Note that erfc(x > 27) = 0 in double precision
//
for(n[0] = -5; n[0] <= 5; n[0]++)
for(n[1] = -5; n[1] <= 5; n[1]++)
for(n[2] = -5; n[2] <= 5; n[2]++)
Expand All @@ -1051,29 +1051,37 @@ void yukawa_lattice_force(int iii, int jjj, int kkk, double x[3], double force[3
// Note, as YUKAWA_IMASS \to zero, we regenerate the Ewald for Coloumb
// val = erfc(alpha * r) + 2 * alpha * r / sqrt(M_PI) * exp(-alpha * alpha * r * r);

// TYPE I
// 0.5*(A + B) eventually /r^3
val = 0.5000000000000*( exp(ym*r)*gsl_sf_erfc(alpha*r + ym/(2*alpha)) +
exp(-ym*r)*gsl_sf_erfc(alpha*r - ym/(2*alpha)));
val = 0.5*( exp(ym*r)*gsl_sf_erfc(alpha*r + ym/(2*alpha)) +
exp(-ym*r)*gsl_sf_erfc(alpha*r - ym/(2*alpha)));

// KC 1/7/16
// For r > rcut, approx = 0.5*exp(ym*r)*erfc(alpha*r + ym/(2*alpha))
// Overall, it will carry a factor of 1/r^2.

// 0.5*(A + C*r) eventually /r^3
/* val = 0.5*gsl_sf_erfc(alpha*r - ym/(2*alpha))*exp(-ym*r)*(1 + ym*r); */
/* val += 0.5*gsl_sf_erfc(alpha*r + ym/(2*alpha))*exp(ym*r)*(1 - ym*r); */
/* val = 0.5*gsl_sf_erfc(alpha*r - ym/(2*alpha))*exp(-ym*r)*(1.0/r + ym); */
/* val += 0.5*gsl_sf_erfc(alpha*r + ym/(2*alpha))*exp(ym*r)*(1.0/r - ym); */

// 0.5*(B + D*r) eventually /r^3

// TYPE I
for(i = 0; i < 3; i++)
force[i] -= dx[i] / (r * r * r) * val;

/* for(i = 0; i < 3; i++) */
/* force[i] -= dx[i] / (r * r) * val; */

// TYPE II
// Now E
/* val = 2*alpha*exp(-alpha*alpha*r*r-ym*ym/(4*alpha*alpha))/sqrt(M_PI); */
/* val += 2*alpha*exp(-alpha*alpha*r*r-ym*ym/(4*alpha*alpha))/sqrt(M_PI); */

// 0.5*ym*(C +1 D) + E eventually /r^2
val = 0.5000000000000*ym*(-exp(ym*r)*gsl_sf_erfc(alpha*r + ym/(2*alpha)) +
exp(-ym*r)*gsl_sf_erfc(alpha*r - ym/(2*alpha))) +
//
// The subtraction here could destroy accuracy?
val = 0.5*ym*(-exp(ym*r)*gsl_sf_erfc(alpha*r + ym/(2*alpha)) +
exp(-ym*r)*gsl_sf_erfc(alpha*r - ym/(2*alpha))) +
2*alpha*exp(-alpha*alpha*r*r-ym*ym/(4*alpha*alpha))/sqrt(M_PI);

// KC 1/7/16
Expand Down Expand Up @@ -1320,11 +1328,6 @@ FLOAT fourierIntegrand(FLOAT k, gravity normKGreen, FLOAT Z) {

FLOAT k2 = k*k;

// XXX?
// But if we fuck with this, we'll break newton. The exponential factor
// seems to be correct because it works with newton, where the normgreens = 1.0
//
// But if this were the problem, it would only be on the lowend of the force...
return (*normKGreen)(1, 1, k2, k, 1) * exp(-k2 * Z * Z);
}

Expand Down Expand Up @@ -1369,9 +1372,8 @@ int performConvolution(struct ngravsInterpolant *s, gravity normKGreen, FLOAT Z,
}

// 1) FFTW needs this loaded in wonk order
// XXX? Should this be explicitly zero'd to avoid a constant force contribution
// to the corretion tabulations? Apparently not, because the k = 0 for
// normed newton is 1 and that works...
// Note we need zero power in order to compute the
// potential term correctly.
in[0].re = fourierIntegrand(jTok(0, Z, s), normKGreen, Z);

for(j = 1; j < s->ngravs_tpm_n/2; ++j) {
Expand All @@ -1390,23 +1392,18 @@ int performConvolution(struct ngravsInterpolant *s, gravity normKGreen, FLOAT Z,
/* printf("%.15f %.15f\n", mTox(m, s), out[m].re*norm); */
/* exit(0); */

// ???
sum = s->ngravs_tpm_n;

for(m = 0; m < s->ntab; ++m)
oRes[m] = out[gadgetToFourier(m, s)].re * norm;

// 3) Integrate so as to constrain the error correctly:
// Newton-Cotes 4-point rule
// Run the sum at double precision, though we may assign to lower precision
//
// KC 1/5/16
// XXX?
// The first term of the integrated quantity should be zero, as it corresponds to r=0?
//
// Is there an off by one error here?
//sum = 0.0;
// Remove the lowest
// First term of in[] should be zero.
in[0].re = 0.0;
sum = 0.0;
for(m = 0; m < s->ngravs_tpm_n-3; m += 3) {
sum += (mTox(m+3, s) - mTox(m, s)) * 0.125 * norm * (out[m].re + 3.0*out[m+1].re + 3.0*out[m+2].re + out[m+3].re);

Expand Down
17 changes: 7 additions & 10 deletions pm_periodic.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -138,7 +138,6 @@ void pm_init_periodic_allocate(int dimprod)
int dimprodmax;
double bytes_tot = 0;
size_t bytes;
int n, m;

MPI_Allreduce(&dimprod, &dimprodmax, 1, MPI_INT, MPI_MAX, MPI_COMM_WORLD);

Expand Down Expand Up @@ -184,7 +183,6 @@ void pm_init_periodic_allocate(int dimprod)
*/
void pm_init_periodic_free(void)
{
int n,m;
/* allocate the memory to hold the FFT fields */
free(workspace);
free(forcegrid);
Expand Down Expand Up @@ -488,14 +486,13 @@ void pmforce_periodic(void)
//
// We also apply the short range truncation, here in k-space
//
//
// This is the only place asmth has entered anywhere!
smth = (*GreensFxns[nA][nB])(All.MassTable[nA], All.MassTable[nB], k2, sqrt(k2), 1);
smth *= -exp(-k2 * asmth2) * ff * ff * ff * ff;
smth *= -exp(-k2*asmth2) * ff * ff * ff * ff;

// NOTE: Transposed order of indicies due to FFTW in k-space
ip = PMGRID * (PMGRID / 2 + 1) * (y - slabstart_y) + (PMGRID / 2 + 1) * x + z;

// KC 27.9.15
//
// CONSTRAINT:
Expand Down Expand Up @@ -1061,11 +1058,11 @@ void pmpotential_periodic(void)
}

// 12/31/15
// XXX! Need to be adjusted to check for nan. If so, then do this. Otherwise,
// we need the DC power.
// This looks like a check to set the DC power to zero....
/* if(slabstart_y == 0) */
/* fft_of_rhogrid[0].re = fft_of_rhogrid[0].im = 0.0; */
// Check for nan. If so, set DC power to zero (this was what was done previously)
if(slabstart_y == 0) {
if(fft_of_rhogrid[0].re != fft_of_rhogrid[0].re) // IEEE spec!
fft_of_rhogrid[0].re = fft_of_rhogrid[0].im = 0.0;
}

/* Do the FFT to get the potential */

Expand Down
28 changes: 23 additions & 5 deletions rdep.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -46,11 +46,29 @@ def __init__(self, ptype, crap2, crap3, x, y, z, fx, fy, fz, jx, jy, jz, cfx, cf
# cen = np.array([4967.0, 4967.0, 4967.0])
cen = np.array([random.random()*L for x in range(3)])

# These separations probe the default TPM transition scale
seps_pmxition = [10 + n*(400 - 10)/(N/2) for n in range(N/2)]
seps_longrange = [400 + n*(L - 400)/(N/2) for n in range(N/2)]
seps_pmxition = []
seps_longrange = []
rutile = raw_input("Enter custom range (min,max) [Enter] for default: ")
bounds = [float(x) for x in rutile.split(',')]

if len(bounds) > 1:

# Reorder if necessary
if bounds[0] > bounds[1]:
tmp = bounds[0]
bounds[0] = bounds[1]
bounds[1] = tmp

# Assign the partitioning
seps_pmxition = []
seps_longrange = [bounds[0] + n*(bounds[1] - bounds[0])/N for n in range(N)]
else:
print "Using default ranges over entire length"

# These separations probe the default TPM transition scale
seps_pmxition = [10 + n*(400 - 10)/(N/2) for n in range(N/2)]
seps_longrange = [400 + n*(L - 400)/(N/2) for n in range(N/2)]

#seps = [L/s for s in range(1,N+1)]
seps = [x for x in seps_pmxition + seps_longrange]
print seps

Expand All @@ -60,7 +78,7 @@ def __init__(self, ptype, crap2, crap3, x, y, z, fx, fy, fz, jx, jy, jz, cfx, cf
if not os.path.exists("./%s/%s" % (proggyName, label)):
os.mkdir("./%s/%s" % (proggyName, label))
else:
print "%s run already exists. Refusing to overwrite."
print "%s run already exists. Refusing to overwrite." % label
sys.exit()

for i,d in enumerate(seps):
Expand Down

0 comments on commit ef42f9f

Please sign in to comment.