def exercise(flags, space_group_info, n_sampled):
symbol = space_group_info.type().hall_symbol()
print symbol,
if flags.fix_seed:
random.seed(0)
if not flags.include_high_symmetry:
if space_group_info.group().order_p() > 8:
if len(symbol) > 15: print
print " [ Omitted, rerun with --include_high_symmetry to override ]"
return
print
n = int(flags.repeats)
if n == 0: n = 1
progress = progress_displayed_as_fraction(n)
for i in xrange(n):
xs = random_structure.xray_structure(space_group_info=space_group_info,
elements=['C'] * 5 + ['O'] * 2 +
['N'],
use_u_iso=True,
random_u_iso=True,
random_u_iso_scale=0.04,
use_u_aniso=False)
f_c = miller.build_set(xs, anomalous_flag=False,
d_min=0.8).structure_factors_from_scatterers(
xs, algorithm='direct').f_calc()
f_o = f_c.as_amplitude_array()
f_c_in_p1 = f_c.expand_to_p1()
exercise_value(f_c_in_p1, f_o, flags, n_sampled)
exercise_gradient(f_c_in_p1, f_o, flags, n_sampled)
progress.advance()
progress.done()
def exercise(flags, space_group_info, n_sampled):
symbol = space_group_info.type().hall_symbol()
print symbol,
if flags.fix_seed:
random.seed(0)
if not flags.include_high_symmetry:
if space_group_info.group().order_p() > 8:
if len(symbol) > 15: print
print " [ Omitted, rerun with --include_high_symmetry to override ]"
return
print
n = int(flags.repeats)
if n == 0: n = 1
progress = progress_displayed_as_fraction(n)
for i in xrange(n):
xs = random_structure.xray_structure(
space_group_info=space_group_info,
elements=['C']*5 + ['O']*2 + ['N'],
use_u_iso=True,
random_u_iso=True,
random_u_iso_scale=0.04,
use_u_aniso=False)
f_c = miller.build_set(
xs, anomalous_flag=False, d_min=0.8
).structure_factors_from_scatterers(
xs, algorithm='direct').f_calc()
f_o = f_c.as_amplitude_array()
f_c_in_p1 = f_c.expand_to_p1()
exercise_value(f_c_in_p1, f_o, flags, n_sampled)
exercise_gradient(f_c_in_p1, f_o, flags, n_sampled)
progress.advance()
progress.done()
def compare_times(comprehensive=False,
svd_impl_name="dgesvd",
use_fortran=False):
import scitbx.linalg.svd
from scitbx.array_family import flex
import time
from libtbx.utils import progress_displayed_as_fraction
mt = flex.mersenne_twister(seed=0)
samples = []
if not comprehensive:
dims = (100, 200)
for m in dims:
for n in dims:
samples.append((m, n))
else:
if comprehensive == "timing-1":
dims = range(100, 600, 100)
for m in dims:
for n in dims:
samples.append((m, n))
elif comprehensive == "timing-2":
for k in (1, 1.1, 1.2, 1.3, 1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2, 3, 4,
5, 6, 7, 8, 9, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100):
for n in (10, 20, 30, 40, 50, 60, 70, 80, 90, 100):
samples.append((int(k * n), n))
else:
raise RuntimeError(comprehensive)
if not comprehensive:
header = " m n real %s %s" % (svd_impl_name, ["fem", "for"
][int(use_fortran)])
else:
handwritten_wrt_lapack = []
progress = progress_displayed_as_fraction(len(samples))
lapack_svd_impl = getattr(scitbx.linalg, "lapack_%s" % svd_impl_name)
try:
for m, n in samples:
if comprehensive: progress.advance()
a = mt.random_double(size=m * n) * 4 - 2
a.reshape(flex.grid(m, n))
ac = a.deep_copy()
t0 = time.time()
svd_real = scitbx.linalg.svd.real(ac,
accumulate_u=True,
accumulate_v=True)
time_svd_real = time.time() - t0
t0 = time.time()
at = a.matrix_transpose()
svd_lapack = lapack_svd_impl(a=at, use_fortran=use_fortran)
if (svd_lapack is None):
return
svd_lapack.u.matrix_transpose()
svd_lapack.vt.matrix_transpose()
time_dgesvd = time.time() - t0
if not comprehensive:
if (header is not None):
print header
header = None
print "%3d %3d %4.2f %4.2f" % (m, n, time_svd_real,
time_dgesvd)
else:
handwritten_wrt_lapack.append(
(m, n, time_svd_real / time_dgesvd))
finally:
if comprehensive:
print "handwrittenwrtlapack%s%s={" % (svd_impl_name,
['', 'fortran'][use_fortran])
print ",".join([
"{%3d, %3d, %4.2f}" % (m, n, t)
for (m, n, t) in handwritten_wrt_lapack
])
print "}"
def compare_times(comprehensive=False, svd_impl_name="dgesvd", use_fortran=False):
import scitbx.linalg.svd
from scitbx.array_family import flex
import time
from libtbx.utils import progress_displayed_as_fraction
mt = flex.mersenne_twister(seed=0)
samples = []
if not comprehensive:
dims = (100, 200)
for m in dims:
for n in dims:
samples.append((m, n))
else:
if comprehensive == "timing-1":
dims = range(100, 600, 100)
for m in dims:
for n in dims:
samples.append((m, n))
elif comprehensive == "timing-2":
for k in (
1,
1.1,
1.2,
1.3,
1.4,
1.5,
1.6,
1.7,
1.8,
1.9,
2,
3,
4,
5,
6,
7,
8,
9,
10,
20,
30,
40,
50,
60,
70,
80,
90,
100,
):
for n in (10, 20, 30, 40, 50, 60, 70, 80, 90, 100):
samples.append((int(k * n), n))
else:
raise RuntimeError(comprehensive)
if not comprehensive:
header = " m n real %s %s" % (svd_impl_name, ["fem", "for"][int(use_fortran)])
else:
handwritten_wrt_lapack = []
progress = progress_displayed_as_fraction(len(samples))
lapack_svd_impl = getattr(scitbx.linalg, "lapack_%s" % svd_impl_name)
try:
for m, n in samples:
if comprehensive:
progress.advance()
a = mt.random_double(size=m * n) * 4 - 2
a.reshape(flex.grid(m, n))
ac = a.deep_copy()
t0 = time.time()
svd_real = scitbx.linalg.svd.real(ac, accumulate_u=True, accumulate_v=True)
time_svd_real = time.time() - t0
t0 = time.time()
at = a.matrix_transpose()
svd_lapack = lapack_svd_impl(a=at, use_fortran=use_fortran)
if svd_lapack is None:
return
svd_lapack.u.matrix_transpose()
svd_lapack.vt.matrix_transpose()
time_dgesvd = time.time() - t0
if not comprehensive:
if header is not None:
print header
header = None
print "%3d %3d %4.2f %4.2f" % (m, n, time_svd_real, time_dgesvd)
else:
handwritten_wrt_lapack.append((m, n, time_svd_real / time_dgesvd))
finally:
if comprehensive:
print "handwrittenwrtlapack%s%s={" % (svd_impl_name, ["", "fortran"][use_fortran])
print ",".join(["{%3d, %3d, %4.2f}" % (m, n, t) for (m, n, t) in handwritten_wrt_lapack])
print "}"
