def benchmark_theory_threads_all(numpart_frac=[
0.001, 0.005, 0.01, 0.05, 0.1, 0.2, 0.25, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0
],
nrepeats=3,
keys=None,
isa=None):
from Corrfunc.theory import DD, DDrppi, wp, xi
allkeys = [ #'DD', 'DDrppi',
'wp', 'xi'
]
allisa = ['avx512f', 'avx', 'sse42', 'fallback']
if keys is None:
keys = allkeys
else:
for k in keys:
if k not in allkeys:
msg = "Valid routines to benchmark are: {0}\nFound routine"\
" = {1}".format(allkeys, k)
raise ValueError(msg)
if isa is None:
isa = allisa
else:
for i in isa:
if i not in allisa:
msg = "Valid instructions sets benchmark are: {0}\n"\
"Found routine = {1}".format(allisa, i)
raise ValueError(msg)
numpart_frac = np.array(numpart_frac)
print("Benchmarking theory routines {0} for isa = {1}".format(keys, isa))
allx, ally, allz = read_catalog()
rmin = 0.1
rmax = 84.0
nbins = 20
bins = np.logspace(np.log10(rmin), np.log10(rmax), nbins)
pimax = rmax # Set to rmax for comparisons between wp and xi
autocorr = 1
boxsize = 420.0
nthreads = max_threads
dtype = np.dtype([('repeat', np.int), ('name', 'U16'), ('isa', 'U16'),
('rmax', np.float), ('ndata', np.int), ('nrand', np.int),
('nthreads', np.int), ('runtime', np.float),
('serial_time', np.float), ('pair_time', np.float),
('api_time', np.float)])
totN = len(numpart_frac) * len(keys) * len(isa) * nrepeats
runtimes = np.empty(totN, dtype=dtype)
runtimes['nthreads'][:] = nthreads
runtimes['rmax'][:] = rmax
index = 0
stderr_filename = 'stderr.txt'
for run_isa in isa:
for frac in numpart_frac:
npts = np.int(frac * len(allx))
print("Working with N = {0}".format(npts), file=sys.stderr)
x = np.random.choice(allx, npts, replace=False)
y = np.random.choice(ally, npts, replace=False)
z = np.random.choice(allz, npts, replace=False)
start_thread_index = index
if 'DD' in keys:
for repeat in range(nrepeats):
runtimes['repeat'][index] = repeat
with stderr_redirected(to=stderr_filename):
t0 = time.time()
_, api_time = DD(autocorr,
nthreads,
bins,
x,
y,
z,
verbose=True,
c_api_timer=True,
isa=run_isa)
t1 = time.time()
runtimes['name'][index] = 'DD'
runtimes['repeat'][index] = repeat
runtimes['isa'][index] = run_isa
runtimes['ndata'][index] = npts
runtimes['nrand'][index] = npts
runtimes['rmax'][index] = rmax
runtimes['nthreads'][index] = nthreads
runtimes['runtime'][index] = t1 - t0
serial_time, pair_time = _get_times(stderr_filename)
runtimes['serial_time'][index] = serial_time
runtimes['pair_time'][index] = pair_time
runtimes['api_time'][index] = api_time
index += 1
if 'DDrppi' in keys:
for repeat in range(nrepeats):
runtimes['repeat'][index] = repeat
with stderr_redirected(to=stderr_filename):
t0 = time.time()
_, api_time = DDrppi(autocorr,
nthreads,
pimax,
bins,
x,
y,
z,
verbose=True,
c_api_timer=True,
isa=run_isa)
t1 = time.time()
runtimes['name'][index] = 'DDrppi'
runtimes['isa'][index] = run_isa
runtimes['ndata'][index] = npts
runtimes['nrand'][index] = npts
runtimes['rmax'][index] = rmax
runtimes['nthreads'][index] = nthreads
runtimes['runtime'][index] = t1 - t0
serial_time, pair_time = _get_times(stderr_filename)
runtimes['serial_time'][index] = serial_time
runtimes['pair_time'][index] = pair_time
runtimes['api_time'][index] = api_time
index += 1
if 'wp' in keys:
for repeat in range(nrepeats):
runtimes['repeat'][index] = repeat
with stderr_redirected(to=stderr_filename):
t0 = time.time()
_, api_time = wp(boxsize,
pimax,
nthreads,
bins,
x,
y,
z,
verbose=True,
c_api_timer=True,
isa=run_isa)
t1 = time.time()
runtimes['name'][index] = 'wp'
runtimes['repeat'][index] = repeat
runtimes['isa'][index] = run_isa
runtimes['ndata'][index] = npts
runtimes['nrand'][index] = npts
runtimes['rmax'][index] = rmax
runtimes['nthreads'][index] = nthreads
runtimes['runtime'][index] = t1 - t0
serial_time, pair_time = _get_times(stderr_filename)
runtimes['serial_time'][index] = serial_time
runtimes['pair_time'][index] = pair_time
runtimes['api_time'][index] = api_time
index += 1
if 'xi' in keys:
for repeat in range(nrepeats):
runtimes['repeat'][index] = repeat
with stderr_redirected(to=stderr_filename):
t0 = time.time()
_, api_time = xi(boxsize,
nthreads,
bins,
x,
y,
z,
verbose=True,
c_api_timer=True,
isa=run_isa)
t1 = time.time()
runtimes['name'][index] = 'xi'
runtimes['repeat'][index] = repeat
runtimes['isa'][index] = run_isa
runtimes['ndata'][index] = npts
runtimes['nrand'][index] = npts
runtimes['rmax'][index] = rmax
runtimes['nthreads'][index] = nthreads
runtimes['runtime'][index] = t1 - t0
serial_time, pair_time = _get_times(stderr_filename)
runtimes['serial_time'][index] = serial_time
runtimes['pair_time'][index] = pair_time
runtimes['api_time'][index] = api_time
index += 1
print("{0}".format(runtimes[start_thread_index:index]))
sys.stdout.flush()
print("index = {0} totN = {1}".format(index, totN))
# autocorr is always 1 for theory routines -> 'nrand' == 'ndata'
runtimes['nrand'][:] = (runtimes['ndata'][:]).copy()
return keys, isa, runtimes
def calc_xirppi_fast(x1,
y1,
z1,
rpbins,
pimax,
pi_bin_size,
lbox,
Nthread,
num_cells=20,
x2=None,
y2=None,
z2=None): # all r assumed to be in h-1 mpc units.
start = time.time()
if not isinstance(pimax, int):
raise ValueError("pimax needs to be an integer")
if not isinstance(pi_bin_size, int):
raise ValueError("pi_bin_size needs to be an integer")
if not pimax % pi_bin_size == 0:
raise ValueError(
"pi_bin_size needs to be an integer divisor of pimax, current values are ",
pi_bin_size, pimax)
ND1 = float(len(x1))
if x2 is not None:
ND2 = len(x2)
autocorr = 0
else:
autocorr = 1
ND2 = ND1
# single precision mode
# to do: make this native
cf_start = time.time()
rpbins = rpbins.astype(np.float32)
pimax = np.float32(pimax)
x1 = x1.astype(np.float32)
y1 = y1.astype(np.float32)
z1 = z1.astype(np.float32)
lbox = np.float32(lbox)
if autocorr == 1:
results = DDrppi(autocorr,
Nthread,
pimax,
rpbins,
x1,
y1,
z1,
boxsize=lbox,
periodic=True,
max_cells_per_dim=num_cells)
DD_counts = results['npairs']
else:
x2 = x2.astype(np.float32)
y2 = y2.astype(np.float32)
z2 = z2.astype(np.float32)
results = DDrppi(autocorr,
Nthread,
pimax,
rpbins,
x1,
y1,
z1,
X2=x2,
Y2=y2,
Z2=z2,
boxsize=lbox,
periodic=True,
max_cells_per_dim=num_cells)
DD_counts = results['npairs']
print("corrfunc took time ", time.time() - cf_start)
DD_counts_new = np.array([
np.sum(DD_counts[i:i + pi_bin_size])
for i in range(0, len(DD_counts), pi_bin_size)
])
DD_counts_new = DD_counts_new.reshape(
(len(rpbins) - 1, int(pimax / pi_bin_size)))
# RR_counts_new = np.zeros((len(rpbins) - 1, int(pimax/pi_bin_size)))
RR_counts_new = np.pi * (
rpbins[1:]**2 - rpbins[:-1]**2) * pi_bin_size / lbox**3 * ND1 * ND2 * 2
xirppi = DD_counts_new / RR_counts_new[:, None] - 1
print("corrfunc took ", time.time() - start, "ngal ", len(x1))
return xirppi
def benchmark_theory_threads_all(min_threads=1,
max_threads=max_threads,
nrepeats=1,
keys=None,
isa=None):
from Corrfunc.theory import DD, DDrppi, wp, xi
allkeys = [ #'DDrppi', 'DD',
'wp', 'xi'
]
allisa = ['avx', 'sse42', 'fallback']
if keys is None:
keys = allkeys
else:
for k in keys:
if k not in allkeys:
msg = "Valid routines to benchmark are: {0}\nFound routine"\
" = {1}".format(allkeys, k)
raise ValueError(msg)
if isa is None:
isa = allisa
else:
for i in isa:
if i not in allisa:
msg = "Valid instructions sets benchmark are: {0}\n"\
"Found routine = {1}".format(allisa, i)
raise ValueError(msg)
print("Benchmarking theory routines = {0} with isa = {1}".format(
keys, isa))
x, y, z = read_catalog()
rmax = 42.0
rmin = 0.1
nbins = 20
bins = np.logspace(np.log10(rmin), np.log10(rmax), nbins)
autocorr = 1
pimax = rmax # Set to rmax for comparisons between wp and xi
boxsize = 420.0
dtype = np.dtype([('repeat', np.int), ('name', 'S16'), ('isa', 'S16'),
('nthreads', np.int), ('runtime', np.float),
('serial_time', np.float), ('pair_time', np.float),
('api_time', np.float)])
totN = (max_threads - min_threads + 1) * len(keys) * len(isa) * nrepeats
runtimes = np.empty(totN, dtype=dtype)
index = 0
stderr_filename = 'stderr.txt'
for run_isa in isa:
for nthreads in range(min_threads, max_threads + 1):
print("Working on nthreads = {0}".format(nthreads),
file=sys.stderr)
start_thread_index = index
if 'DD' in keys:
for repeat in range(nrepeats):
runtimes['repeat'][index] = repeat
with stderr_redirected(to=stderr_filename):
t0 = time.time()
_, api_time = DD(autocorr,
nthreads,
bins,
x,
y,
z,
verbose=True,
c_api_timer=True,
isa=run_isa)
t1 = time.time()
runtimes['name'][index] = 'DD'
runtimes['isa'][index] = run_isa
runtimes['nthreads'][index] = nthreads
runtimes['runtime'][index] = t1 - t0
serial_time, pair_time = _get_times(stderr_filename)
runtimes['serial_time'][index] = serial_time
runtimes['pair_time'][index] = pair_time
runtimes['api_time'][index] = api_time
index += 1
if 'DDrppi' in keys:
for repeat in range(nrepeats):
runtimes['repeat'][index] = repeat
with stderr_redirected(to=stderr_filename):
t0 = time.time()
_, api_time = DDrppi(autocorr,
nthreads,
pimax,
bins,
x,
y,
z,
verbose=True,
c_api_timer=True,
isa=run_isa)
t1 = time.time()
runtimes['name'][index] = 'DDrppi'
runtimes['isa'][index] = run_isa
runtimes['nthreads'][index] = nthreads
runtimes['runtime'][index] = t1 - t0
serial_time, pair_time = _get_times(stderr_filename)
runtimes['serial_time'][index] = serial_time
runtimes['pair_time'][index] = pair_time
runtimes['api_time'][index] = api_time
index += 1
if 'wp' in keys:
for repeat in range(nrepeats):
runtimes['repeat'][index] = repeat
with stderr_redirected(to=stderr_filename):
t0 = time.time()
_, api_time = wp(boxsize,
pimax,
nthreads,
bins,
x,
y,
z,
verbose=True,
c_api_timer=True,
isa=run_isa)
t1 = time.time()
runtimes['name'][index] = 'wp'
runtimes['isa'][index] = run_isa
runtimes['nthreads'][index] = nthreads
runtimes['runtime'][index] = t1 - t0
serial_time, pair_time = _get_times(stderr_filename)
runtimes['serial_time'][index] = serial_time
runtimes['pair_time'][index] = pair_time
runtimes['api_time'][index] = api_time
index += 1
if 'xi' in keys:
for repeat in range(nrepeats):
runtimes['repeat'][index] = repeat
with stderr_redirected(to=stderr_filename):
t0 = time.time()
_, api_time = xi(boxsize,
nthreads,
bins,
x,
y,
z,
verbose=True,
c_api_timer=True,
isa=run_isa)
t1 = time.time()
runtimes['name'][index] = 'xi'
runtimes['isa'][index] = run_isa
runtimes['nthreads'][index] = nthreads
runtimes['runtime'][index] = t1 - t0
serial_time, pair_time = _get_times(stderr_filename)
runtimes['serial_time'][index] = serial_time
runtimes['pair_time'][index] = pair_time
runtimes['api_time'][index] = api_time
index += 1
print("{0}".format(runtimes[start_thread_index:index]))
sys.stdout.flush()
print("index = {0} totN = {1}".format(index, totN))
return keys, isa, runtimes
def calc_wp_fast(x1,
y1,
z1,
rpbins,
pimax,
lbox,
Nthread,
num_cells=30,
x2=None,
y2=None,
z2=None): # all r assumed to be in h-1 mpc units.
if not isinstance(pimax, int):
raise ValueError("pimax needs to be an integer")
ND1 = float(len(x1))
if x2 is not None:
ND2 = len(x2)
autocorr = 0
else:
autocorr = 1
ND2 = ND1
# single precision mode
# to do: make this native
cf_start = time.time()
rpbins = rpbins.astype(np.float32)
pimax = np.float32(pimax)
x1 = x1.astype(np.float32)
y1 = y1.astype(np.float32)
z1 = z1.astype(np.float32)
lbox = np.float32(lbox)
if autocorr == 1:
print("sample size", len(x1))
results = DDrppi(autocorr,
Nthread,
pimax,
rpbins,
x1,
y1,
z1,
boxsize=lbox,
periodic=True,
max_cells_per_dim=num_cells)
DD_counts = results['npairs']
else:
print("sample size", len(x1), len(x2))
x2 = x2.astype(np.float32)
y2 = y2.astype(np.float32)
z2 = z2.astype(np.float32)
results = DDrppi(autocorr,
Nthread,
pimax,
rpbins,
x1,
y1,
z1,
X2=x2,
Y2=y2,
Z2=z2,
boxsize=lbox,
periodic=True,
max_cells_per_dim=num_cells)
DD_counts = results['npairs']
print("corrfunc took time ", time.time() - cf_start)
DD_counts = DD_counts.reshape((len(rpbins) - 1, int(pimax)))
# RR_counts = np.zeros((len(rpbins) - 1, int(pimax)))
# for i in range(len(rpbins) - 1):
RR_counts = np.pi * (rpbins[1:]**2 -
rpbins[:-1]**2) / lbox**3 * ND1 * ND2 * 2
xirppi = DD_counts / RR_counts[:, None] - 1
return 2 * np.sum(xirppi, axis=1)