def plot_perf_vs_sublattice(dirnames):
fig = pl.figure()
ax = fig.add_subplot(1, 1, 1)
to_plot = collections.defaultdict(lambda: collections.defaultdict(list))
for dirname in dirnames:
filename = os.path.join(dirname, 'extract-log.json')
if not os.path.isfile(filename):
print('File is missing:', filename)
continue
with open(filename) as f:
data = json.load(f)
for update_no, update_data in sorted(data.items()):
nodes = update_data['nodes']
subgrid_volume = update_data['subgrid_volume']
for solver, solver_data in update_data['solvers'].items():
gflops = solver_data['gflops']
iters = solver_data['iters']
to_plot[solver][subgrid_volume] += gflops
for solver, tuples in sorted(to_plot.items()):
x = sorted(tuples.keys())
y = np.array([
np.percentile(gflops, 50) / nodes
for subgrid_volume, gflops in sorted(tuples.items())
])
yerr_down = y - np.array([
np.percentile(gflops, 50 - 34.13) / nodes
for subgrid_volume, gflops in sorted(tuples.items())
])
yerr_up = np.array([
np.percentile(gflops, 50 + 34.13) / nodes
for subgrid_volume, gflops in sorted(tuples.items())
]) - y
ax.errorbar(x,
y, (yerr_down, yerr_up),
marker='o',
linestyle='none',
label=solver)
ax.set_title('Solver Scaling')
ax.set_xlabel('Subgrid Volume')
ax.set_ylabel(r'Gflop/s per Node')
util.dandify_axes(ax)
util.dandify_figure(fig)
pl.savefig('plot-gflops-vs-subgrid_volume.pdf')
pl.savefig('plot-gflops-vs-subgrid_volume.png')
def plot_perf(dirnames):
fig = pl.figure()
ax = fig.add_subplot(1, 1, 1)
for dirname in dirnames:
filename = os.path.join(dirname, 'extract-log.json')
if not os.path.isfile(filename):
print('File is missing:', filename)
continue
with open(filename) as f:
data = json.load(f)
solvers = collections.defaultdict(list)
for update_no, update_data in sorted(data.items()):
nodes = update_data['nodes']
for solver, solver_data in update_data['solvers'].items():
gflops = solver_data['gflops']
iters = solver_data['iters']
solvers[solver].append((
int(update_no),
np.median(gflops) / nodes,
np.percentile(gflops, transforms.PERCENTILE_LOW) / nodes,
np.percentile(gflops, transforms.PERCENTILE_HIGH) / nodes,
))
for solver, tuples in sorted(solvers.items()):
x, y, yerr_low, yerr_high = zip(*sorted(tuples))
label = '{} -- {}'.format(
os.path.basename(os.path.realpath(dirname)), solver)
line, = ax.plot(x, y, marker='o', label=label)
ax.fill_between(x,
np.array(y) - np.array(yerr_low),
np.array(y) + np.array(yerr_high),
alpha=0.2,
color=line.get_color())
ax.set_title('Solver Performance')
ax.set_xlabel('Update Number (shifted for visibility)')
ax.set_ylabel(r'Gflop/s per Node')
util.dandify_axes(ax)
util.dandify_figure(fig)
pl.savefig('plot-perf.pdf')
pl.savefig('plot-perf.png')
def visualize(xml_file, root=None):
t, t2e = util.load_columns(xml_file + '.t2e.tsv')
fig = pl.figure(figsize=(16, 9))
ax1 = fig.add_subplot(1, 2, 1)
ax2 = fig.add_subplot(2, 2, 2)
ax3 = fig.add_subplot(2, 2, 4)
ax1.plot(t, t2e)
ax1.plot(t, np.ones(t.shape) * 0.3)
if root is not None:
x, y = root
#ax2 = pl.axes([.65, .6, .3, .3], axisbg='0.9')
ax2.plot(t, t2e, marker='o')
ax2.plot(t, np.ones(t.shape) * 0.3)
sel = t < 3 * x
ax3.plot(t[sel], t2e[sel])
ax3.plot(t[sel], np.ones(t[sel].shape) * 0.3)
ax1.plot([x], [y], marker='o', alpha=0.3, color='red', markersize=20)
ax2.plot([x], [y], marker='o', alpha=0.3, color='red', markersize=20)
ax3.plot([x], [y], marker='o', alpha=0.3, color='red', markersize=20)
ax2.set_xlim(0.7 * x, 1.3 * x)
ax2.set_ylim(0.22, 0.38)
#ax2.locator_params(nbins=6)
#ax3.locator_params(nbins=6)
ax1.set_title('All Computed Data')
ax2.set_title('Interpolation')
ax3.set_title('First $3 t_0$')
for ax in [ax1, ax2, ax3]:
ax.set_xlabel('Wilson Flow Time $t$')
ax.set_ylabel(r'$t^2 \langle E \rangle$')
util.dandify_axes(ax)
util.dandify_figure(fig)
fig.savefig(xml_file + '.t2e.pdf')
def plot_solver_data(path_in,
path_out,
ylabel,
title='Solver Data',
log_scale=False):
fig, ax = util.make_figure()
with open(path_in) as f:
data = json.load(f)
for solver, (x, y, yerr_down, yerr_up) in data.items():
x = np.array(x)
y = np.array(y)
yerr_down = np.array(yerr_down)
yerr_up = np.array(yerr_up)
label = solver
p = ax.plot(x, y, label=label)
ax.fill_between(x,
y - yerr_down,
y + yerr_up,
alpha=0.3,
color=p[0].get_color())
ax.set_title(title)
ax.set_xlabel('Update Number')
ax.set_ylabel(ylabel)
if log_scale:
ax.set_yscale('log')
util.dandify_axes(ax)
if log_scale:
start, end = ax.get_ylim()
print(start, end)
print(end / start)
if end / start < 15:
start = 10**int(np.log10(start) - 1)
end = 10**int(np.log10(end) + 1)
print('{:.10g} {:.10g}'.format(start, end))
ax.set_ylim(start, end)
util.dandify_figure(fig)
fig.savefig(path_out)
def plot_generic(path_in,
path_out,
xlabel,
ylabel,
title,
use_auto_ylim=False):
fig = pl.figure()
ax = fig.add_subplot(1, 1, 1)
x, y = util.load_columns(path_in, 2)
if len(x) > 0:
ax.plot(x, y, marker='o', markersize=2)
ax.set_title(title)
ax.set_xlabel(xlabel)
ax.set_ylabel(ylabel)
if use_auto_ylim:
ax.set_ylim(get_auto_ylim(y))
util.dandify_axes(ax)
util.dandify_figure(fig)
pl.savefig(path_out)
def main():
options = _parse_args()
R = 300
# Read in the data from the paper.
a_inv_val = 1616
a_inv_err = 20
a_inv_dist = bootstrap.make_dist(a_inv_val, a_inv_err, n=R)
aml, ams, l, t, trajectories, ampi_val, ampi_err, amk_val, amk_err, f_k_f_pi_val, f_k_f_pi_err = util.load_columns(
'physical_point/gmor.txt')
ampi_dist = bootstrap.make_dist(ampi_val, ampi_err, n=R)
amk_dist = bootstrap.make_dist(amk_val, amk_err, n=R)
mpi_dist = [ampi * a_inv for ampi, a_inv in zip(ampi_dist, a_inv_dist)]
mk_dist = [amk * a_inv for amk, a_inv in zip(amk_dist, a_inv_dist)]
# Convert the data in lattice units into physical units.
mpi_dist = [a_inv * ampi for ampi, a_inv in zip(ampi_dist, a_inv_dist)]
mpi_val, mpi_avg, mpi_err = bootstrap.average_and_std_arrays(mpi_dist)
mpi_sq_dist = [mpi**2 for mpi in mpi_dist]
mpi_sq_val, mpi_sq_avg, mpi_sq_err = bootstrap.average_and_std_arrays(
mpi_sq_dist)
ampi_sq_dist = [ampi**2 for ampi in ampi_dist]
ampi_sq_val, ampi_sq_avg, ampi_sq_err = bootstrap.average_and_std_arrays(
ampi_sq_dist)
# Do a GMOR fit in order to extract `a B` and `a m_cr`.
popt_dist = [
op.curve_fit(gmor_pion, aml, ampi_sq)[0] for ampi_sq in ampi_sq_dist
]
aB_dist = [popt[0] for popt in popt_dist]
amcr_dist = [popt[1] for popt in popt_dist]
aB_val, aB_avg, aB_err = bootstrap.average_and_std_arrays(aB_dist)
amcr_val, amcr_avg, amcr_err = bootstrap.average_and_std_arrays(amcr_dist)
print('aB =', siunitx(aB_val, aB_err))
print('am_cr =', siunitx(amcr_val, amcr_err))
ams_paper = -0.057
ams_phys = ams_paper - amcr_val
ams_red = 0.9 * ams_phys
ams_bare_red = ams_red + amcr_val
print(ams_paper, ams_phys, ams_red, ams_bare_red)
print()
print('Mass preconditioning masses:')
amlq = aml - amcr_val
for i in range(3):
amprec = amlq * 10**i + amcr_val
kappa = 1 / (amprec * 2 + 8)
print('a m_prec:', amprec)
print('κ', kappa)
exit()
diff_dist = [
np.sqrt(2) * np.sqrt(mk**2 - 0.5 * mpi**2)
for mpi, mk in zip(mpi_dist, mk_dist)
]
diff_val, diff_avg, diff_err = bootstrap.average_and_std_arrays(diff_dist)
popt_dist = [
op.curve_fit(linear, mpi, diff)[0]
for mpi, diff in zip(mpi_dist, diff_dist)
]
fit_x = np.linspace(np.min(mpi_dist), np.max(mpi_dist), 100)
fit_y_dist = [linear(fit_x, *popt) for popt in popt_dist]
fit_y_val, fit_y_avg, fit_y_err = bootstrap.average_and_std_arrays(
fit_y_dist)
# Physical meson masses from FLAG paper.
mpi_phys_dist = bootstrap.make_dist(134.8, 0.3, R)
mk_phys_dist = bootstrap.make_dist(494.2, 0.3, R)
mpi_phys_val, mpi_phys_avg, mpi_phys_err = bootstrap.average_and_std_arrays(
mpi_phys_dist)
ampi_phys_dist = [
mpi_phys / a_inv for a_inv, mpi_phys in zip(a_inv_dist, mpi_phys_dist)
]
amk_phys_dist = [
mk_phys / a_inv for a_inv, mk_phys in zip(a_inv_dist, mk_phys_dist)
]
ampi_phys_val, ampi_phys_avg, ampi_phys_err = bootstrap.average_and_std_arrays(
ampi_phys_dist)
amk_phys_val, amk_phys_avg, amk_phys_err = bootstrap.average_and_std_arrays(
amk_phys_dist)
print('aM_pi phys =', siunitx(ampi_phys_val, ampi_phys_err))
print('aM_k phys =', siunitx(amk_phys_val, amk_phys_err))
new_b_dist = [
np.sqrt(mk_phys**2 - 0.5 * mpi_phys**2) - popt[0] * mpi_phys for
mpi_phys, mk_phys, popt in zip(mpi_phys_dist, mk_phys_dist, popt_dist)
]
diff_sqrt_phys_dist = [
np.sqrt(mk_phys**2 - 0.5 * mpi_phys**2)
for mpi_phys, mk_phys in zip(mpi_phys_dist, mk_phys_dist)
]
diff_sqrt_phys_val, diff_sqrt_phys_avg, diff_sqrt_phys_err = bootstrap.average_and_std_arrays(
diff_sqrt_phys_dist)
ex_x = np.linspace(120, 700, 100)
ex_y_dist = [
linear(ex_x, popt[0], b) for popt, b in zip(popt_dist, new_b_dist)
]
ex_y_val, ex_y_avg, ex_y_err = bootstrap.average_and_std_arrays(ex_y_dist)
ams_art_dist = [
linear(mpi, popt[0], b)**2 / a_inv**2 / aB - amcr
for mpi, popt, b, a_inv, aB, amcr in zip(
mpi_dist, popt_dist, new_b_dist, a_inv_dist, aB_dist, amcr_dist)
]
ams_art_val, ams_art_avg, ams_art_err = bootstrap.average_and_std_arrays(
ams_art_dist)
print('a m_s with artifacts', siunitx(ams_art_val, ams_art_err))
fig, ax = util.make_figure()
ax.fill_between(fit_x,
fit_y_val + fit_y_err,
fit_y_val - fit_y_err,
color='red',
alpha=0.2)
ax.plot(fit_x, fit_y_val, label='Fit', color='red')
ax.fill_between(ex_x,
ex_y_val + ex_y_err,
ex_y_val - ex_y_err,
color='orange',
alpha=0.2)
ax.plot(ex_x, ex_y_val, label='Extrapolation', color='orange')
ax.errorbar(mpi_val,
diff_val,
xerr=mpi_err,
yerr=diff_err,
linestyle='none',
label='Data (Dürr 2010)')
ax.errorbar([mpi_phys_val], [diff_sqrt_phys_val],
xerr=[mpi_phys_err],
yerr=[diff_sqrt_phys_err],
label='Physical Point (Aoki)')
util.save_figure(fig, 'test')
np.savetxt('artifact-bmw-data.tsv',
np.column_stack([mpi_val, diff_val, mpi_err, diff_err]))
np.savetxt('artifact-bmw-fit.tsv', np.column_stack([fit_x, fit_y_val]))
np.savetxt('artifact-bmw-band.tsv',
bootstrap.pgfplots_error_band(fit_x, fit_y_val, fit_y_err))
np.savetxt(
'artifact-phys-data.tsv',
np.column_stack([[mpi_phys_val], [diff_sqrt_phys_val], [mpi_phys_err],
[diff_sqrt_phys_err]]))
np.savetxt('artifact-phys-fit.tsv', np.column_stack([ex_x, ex_y_val]))
np.savetxt('artifact-phys-band.tsv',
bootstrap.pgfplots_error_band(ex_x, ex_y_val, ex_y_err))
np.savetxt('artifact-ms.tsv',
np.column_stack([mpi_val, ams_art_val, mpi_err, ams_art_err]))
# Compute the strange quark mass that is needed to obtain a physical meson
# mass difference, ignoring lattice artifacts.
ams_phys_dist = [(amk_phys**2 - 0.5 * ampi_phys**2) / aB - amcr
for ampi_phys, amk_phys, aB, amcr in zip(
ampi_phys_dist, amk_phys_dist, aB_dist, amcr_dist)]
ams_phys_cen, ams_phys_val, ams_phys_err = bootstrap.average_and_std_arrays(
ams_phys_dist)
print('M_K = {} MeV <== am_s ='.format(siunitx(494.2, 0.3)),
siunitx(ams_phys_cen, ams_phys_err))
aml_phys_dist = [
op.newton(lambda aml: gmor_pion(aml, *popt) - ampi_phys**2,
np.min(aml))
for popt, ampi_phys in zip(popt_dist, ampi_phys_dist)
]
fit_x = np.linspace(np.min(aml_phys_dist), np.max(aml), 100)
fit_y_dist = [
np.sqrt(gmor_pion(fit_x, *popt)) * a_inv
for popt, a_inv in zip(popt_dist, a_inv_dist)
]
fit_y_cen, fit_y_val, fit_y_err = bootstrap.average_and_std_arrays(
fit_y_dist)
np.savetxt('physical_point/mpi-vs-aml-data.tsv',
np.column_stack([aml, mpi_val, mpi_err]))
np.savetxt('physical_point/mpi-vs-aml-fit.tsv',
np.column_stack([fit_x, fit_y_cen]))
np.savetxt('physical_point/mpi-vs-aml-band.tsv',
bootstrap.pgfplots_error_band(fit_x, fit_y_cen, fit_y_err))
aml_phys_val, aml_phys_avg, aml_phys_err = bootstrap.average_and_std_arrays(
aml_phys_dist)
mpi_cen, mpi_val, mpi_err = bootstrap.average_and_std_arrays(mpi_dist)
#aml_240_val, aml_240_avg, aml_240_err = bootstrap.average_and_std_arrays(aml_240_dist)
print('M_pi = {} MeV <== am_l ='.format(siunitx(134.8, 0.3)),
siunitx(aml_phys_val, aml_phys_err))
#print('M_pi = 240 MeV <== am_l =', siunitx(aml_240_val, aml_240_err))
fig = pl.figure()
ax = fig.add_subplot(2, 1, 1)
ax.fill_between(fit_x,
fit_y_val - fit_y_err,
fit_y_val + fit_y_err,
color='0.8')
ax.plot(fit_x, fit_y_val, color='black', label='GMOR Fit')
ax.errorbar(aml,
mpi_val,
yerr=mpi_err,
color='blue',
marker='+',
linestyle='none',
label='Data')
ax.errorbar([aml_phys_val], [135],
xerr=[aml_phys_err],
marker='+',
color='red',
label='Extrapolation')
#ax.errorbar([aml_240_val], [240], xerr=[aml_240_err], marker='+', color='red')
ax.set_title('Extrapolation to the Physical Point')
ax.set_xlabel(r'$a m_\mathrm{ud}$')
ax.set_ylabel(r'$M_\pi / \mathrm{MeV}$')
util.dandify_axes(ax)
ax = fig.add_subplot(2, 1, 2)
ax.hist(aml_phys_dist - aml_phys_val, bins=50)
ax.locator_params(nbins=6)
ax.set_title('Bootstrap Bias')
ax.set_xlabel(
r'$(a m_\mathrm{ud}^\mathrm{phys})^* - a m_\mathrm{ud}^\mathrm{phys}$')
util.dandify_axes(ax)
util.dandify_figure(fig)
fig.savefig('physical_point/GMOR.pdf')
np.savetxt('physical_point/ampi-sq-vs-aml.tsv',
np.column_stack([aml, ampi_sq_val, ampi_sq_err]))
np.savetxt('physical_point/mpi-sq-vs-aml.tsv',
np.column_stack([aml, mpi_sq_val, mpi_sq_err]))