def setUp(self):
self.old_flag = get_config().use_local_memory
get_config().use_local_memory = True
super(TestAccelerationEval1DGPUOctreeNonCached, self).setUp()
def get_lmem_loop_kernel(self, g_idx, sg_idx, group, dest, source,
eq_group):
kind = 'loop'
sub_grp = '' if sg_idx == -1 else 's{idx}'.format(idx=sg_idx)
kernel = 'g{g_idx}{sg}_{source}_on_{dest}_loop'.format(g_idx=g_idx,
sg=sub_grp,
source=source,
dest=dest)
sph_k_name = self.object.kernel.__class__.__name__
context = eq_group.context
all_args, py_args = [], []
setup_code = self._declare_precomp_vars(context)
setup_code.append('GLOBAL_MEM %s* SPH_KERNEL = kern;' % sph_k_name)
if eq_group.has_loop_all():
raise NotImplementedError("loop_all not suported with local "
"memory")
loop_code = []
pre = []
for p, cb in eq_group.precomputed.items():
src = cb.code.strip().splitlines()
pre.extend([' ' * 4 + x + ';' for x in src])
if len(pre) > 0:
pre.append('')
loop_code.extend(pre)
_all_args, _py_args, _calls = self._get_equation_method_calls(
eq_group, kind, indent=' ')
loop_code.extend(_calls)
for arg, py_arg in zip(_all_args, _py_args):
if arg not in all_args:
all_args.append(arg)
py_args.append(py_arg)
s_ary, d_ary = eq_group.get_array_names()
source_vars = set(s_ary)
source_var_types = self._get_arg_base_types(source_vars)
def modify_var_name(x):
if x.startswith('s_'):
return x + '_global'
else:
return x
s_ary.update(d_ary)
_args = list(s_ary)
py_args.extend(_args)
_args_modified = [modify_var_name(x) for x in _args]
all_args.extend(self._get_typed_args(_args_modified))
setup_body = '\n'.join([' ' * 4 + x for x in setup_code])
setup_body = self._set_kernel(setup_body, self.object.kernel)
loop_body = '\n'.join([' ' * 4 + x for x in loop_code])
loop_body = self._set_kernel(loop_body, self.object.kernel)
all_args.extend([
'GLOBAL_MEM {kernel}* kern'.format(kernel=sph_k_name), 'double t',
'double dt'
])
all_args.extend(get_kernel_args_list())
self.data.append(
dict(kernel=kernel,
args=py_args,
dest=dest,
source=source,
loop=True,
real=group.real,
type='kernel'))
body = generate_body(setup=setup_body,
loop=loop_body,
vars=source_vars,
types=source_var_types,
wgs=get_config().wgs)
sig = get_kernel_definition(kernel, all_args)
return ('{sig}\n{{\n{body}\n\n}}\n'.format(sig=sig, body=body))
def _cleanup():
get_config().use_double = orig
def convert_to_float_if_needed(code):
use_double = get_config().use_double
if not use_double:
code = re.sub(r'\bdouble\b', 'float', code)
return code
def get_loop_kernel(self, g_idx, sg_idx, group, dest, source, eq_group):
if get_config().use_local_memory:
return self.get_lmem_loop_kernel(g_idx, sg_idx, group, dest,
source, eq_group)
kind = 'loop'
sub_grp = '' if sg_idx == -1 else 's{idx}'.format(idx=sg_idx)
kernel = 'g{g_idx}{sg}_{source}_on_{dest}_loop'.format(g_idx=g_idx,
sg=sub_grp,
source=source,
dest=dest)
sph_k_name = self.object.kernel.__class__.__name__
context = eq_group.context
all_args, py_args = [], []
code = self._declare_precomp_vars(context)
code.extend([
'unsigned int d_idx = GID_0 * LDIM_0 + LID_0;',
'/* Guard for padded threads. */',
'if (d_idx > NP_MAX) {return;};', 'unsigned int s_idx, i;',
'GLOBAL_MEM %s* SPH_KERNEL = kern;' % sph_k_name,
'unsigned int start = start_idx[d_idx];',
'GLOBAL_MEM unsigned int* NBRS = &(neighbors[start]);',
'int N_NBRS = nbr_length[d_idx];',
'unsigned int end = start + N_NBRS;'
])
if eq_group.has_loop_all():
_all_args, _py_args, _calls = self._get_equation_method_calls(
eq_group, kind='loop_all', indent='')
code.extend(['', '// Calling loop_all of equations.'])
code.extend(_calls)
code.append('')
all_args.extend(_all_args)
py_args.extend(_py_args)
if eq_group.has_loop():
code.append('// Calling loop of equations.')
code.append('for (i=start; i<end; i++) {')
code.append(' s_idx = neighbors[i];')
pre = []
for p, cb in eq_group.precomputed.items():
src = cb.code.strip().splitlines()
pre.extend([' ' * 4 + x + ';' for x in src])
if len(pre) > 0:
pre.append('')
code.extend(pre)
_all_args, _py_args, _calls = self._get_equation_method_calls(
eq_group, kind, indent=' ')
code.extend(_calls)
for arg, py_arg in zip(_all_args, _py_args):
if arg not in all_args:
all_args.append(arg)
py_args.append(py_arg)
code.append('}')
s_ary, d_ary = eq_group.get_array_names()
s_ary.update(d_ary)
_args = list(s_ary)
py_args.extend(_args)
all_args.extend(self._get_typed_args(_args))
body = '\n'.join([' ' * 4 + x for x in code])
body = self._set_kernel(body, self.object.kernel)
all_args.extend([
'GLOBAL_MEM {kernel}* kern'.format(kernel=sph_k_name),
'GLOBAL_MEM unsigned int *nbr_length',
'GLOBAL_MEM unsigned int *start_idx',
'GLOBAL_MEM unsigned int *neighbors', 'double t', 'double dt',
'unsigned int NP_MAX'
])
self.data.append(
dict(kernel=kernel,
args=py_args,
dest=dest,
source=source,
loop=True,
real=group.real,
type='kernel'))
sig = get_kernel_definition(kernel, all_args)
return ('{sig}\n{{\n{body}\n\n}}\n'.format(sig=sig, body=body))
def tearDown(self):
super(OctreeGPUNNPSTestCase, self).tearDown()
get_config().use_double = self._orig_use_double
def setup_module():
get_config().use_openmp = True
def setUp(self):
cu = pytest.importorskip("pycuda")
cfg = get_config()
self.orig_use_double = cfg.use_double
cfg.use_double = True
self.backend = 'cuda'
def tearDown(self):
get_config().use_double = self.orig_use_double
def __init__(self,
all_particles,
scheme,
domain=None,
innerloop=True,
updates=True,
parallel=False,
steps=None,
D=0):
"""The second integrator is a simple Euler-Integrator (accurate
enough due to very small time steps; very fast) using EBGSteps.
EBGSteps are basically the same as EulerSteps, exept for the fact
that they work with an intermediate ebg velocity [eu, ev, ew].
This velocity does not interfere with the actual velocity, which
is neseccery to not disturb the real velocity through artificial
damping in this step. The ebg velocity is initialized for each
inner loop again and reset in the outer loop."""
from math import ceil
from pysph.base.kernels import CubicSpline
from pysph.sph.integrator_step import EBGStep
from compyle.config import get_config
from pysph.sph.integrator import EulerIntegrator
from pysph.sph.scheme import BeadChainScheme
from pysph.sph.equation import Group
from pysph.sph.fiber.utils import (HoldPoints, Contact,
ComputeDistance)
from pysph.sph.fiber.beadchain import (Tension, Bending,
ArtificialDamping)
from pysph.base.nnps import DomainManager, LinkedListNNPS
from pysph.sph.acceleration_eval import AccelerationEval
from pysph.sph.sph_compiler import SPHCompiler
if not isinstance(scheme, BeadChainScheme):
raise TypeError("Scheme must be BeadChainScheme")
self.innerloop = innerloop
self.dt = scheme.dt
self.fiber_dt = scheme.fiber_dt
self.domain_updates = updates
self.steps = steps
self.D = D
self.eta0 = scheme.rho0 * scheme.nu
# if there are more than 1 particles involved, elastic equations are
# iterated in an inner loop.
if self.innerloop:
# second integrator
# self.fiber_integrator = EulerIntegrator(fiber=EBGStep())
steppers = {}
for f in scheme.fibers:
steppers[f] = EBGStep()
self.fiber_integrator = EulerIntegrator(**steppers)
# The type of spline has no influence here. It must be large enough
# to contain the next particle though.
kernel = CubicSpline(dim=scheme.dim)
equations = []
g1 = []
for fiber in scheme.fibers:
g1.append(ComputeDistance(dest=fiber, sources=[fiber]))
equations.append(Group(equations=g1))
g2 = []
for fiber in scheme.fibers:
g2.append(
Tension(dest=fiber, sources=None, ea=scheme.E * scheme.A))
g2.append(
Bending(dest=fiber, sources=None, ei=scheme.E * scheme.Ip))
g2.append(
Contact(dest=fiber,
sources=scheme.fibers,
E=scheme.E,
d=scheme.dx,
dim=scheme.dim,
k=scheme.k,
lim=scheme.lim,
eta0=self.eta0))
g2.append(ArtificialDamping(dest=fiber, sources=None,
d=self.D))
equations.append(Group(equations=g2))
g3 = []
for fiber in scheme.fibers:
g3.append(HoldPoints(dest=fiber, sources=None, tag=100))
equations.append(Group(equations=g3))
# These equations are applied to fiber particles only - that's the
# reason for computational speed up.
particles = [p for p in all_particles if p.name in scheme.fibers]
# A seperate DomainManager is needed to ensure that particles don't
# leave the domain.
if domain:
xmin = domain.manager.xmin
ymin = domain.manager.ymin
zmin = domain.manager.zmin
xmax = domain.manager.xmax
ymax = domain.manager.ymax
zmax = domain.manager.zmax
periodic_in_x = domain.manager.periodic_in_x
periodic_in_y = domain.manager.periodic_in_y
periodic_in_z = domain.manager.periodic_in_z
gamma_yx = domain.manager.gamma_yx
gamma_zx = domain.manager.gamma_zx
gamma_zy = domain.manager.gamma_zy
n_layers = domain.manager.n_layers
N = self.steps or int(ceil(self.dt / self.fiber_dt))
# dt = self.dt/N
self.domain = DomainManager(xmin=xmin,
xmax=xmax,
ymin=ymin,
ymax=ymax,
zmin=zmin,
zmax=zmax,
periodic_in_x=periodic_in_x,
periodic_in_y=periodic_in_y,
periodic_in_z=periodic_in_z,
gamma_yx=gamma_yx,
gamma_zx=gamma_zx,
gamma_zy=gamma_zy,
n_layers=n_layers,
dt=self.dt,
calls_per_step=N)
else:
self.domain = None
# A seperate list for the nearest neighbourhood search is
# benefitial since it is much smaller than the original one.
nnps = LinkedListNNPS(dim=scheme.dim,
particles=particles,
radius_scale=kernel.radius_scale,
domain=self.domain,
fixed_h=False,
cache=False,
sort_gids=False)
# The acceleration evaluator needs to be set up in order to compile
# it together with the integrator.
if parallel:
self.acceleration_eval = AccelerationEval(
particle_arrays=particles,
equations=equations,
kernel=kernel)
else:
self.acceleration_eval = AccelerationEval(
particle_arrays=particles,
equations=equations,
kernel=kernel,
mode='serial')
# Compilation of the integrator not using openmp, because the
# overhead is too large for those few fiber particles.
comp = SPHCompiler(self.acceleration_eval, self.fiber_integrator)
if parallel:
comp.compile()
else:
config = get_config()
config.use_openmp = False
comp.compile()
config.use_openmp = True
self.acceleration_eval.set_nnps(nnps)
# Connecting neighbourhood list to integrator.
self.fiber_integrator.set_nnps(nnps)
def setUp(self):
ocl = pytest.importorskip("pyopencl")
cfg = get_config()
self.orig_use_double = cfg.use_double
cfg.use_double = True
self.backend = 'opencl'
def tearDown(self):
super(TestZOrderGPUNNPSWithSorting, self).tearDown()
get_config().use_double = self._orig_use_double
choices=['gpu_comp', 'omp_comp', 'comp_algo'],
help='Choose the comparison.')
p.add_argument('--nnps',
action='store',
dest='nnps',
default='linear',
choices=['linear', 'simple'],
help='Choose algorithm.')
p.add_argument('--use-double',
action='store_true',
dest='use_double',
default=False,
help='Use double precision on the GPU.')
o = p.parse_args()
get_config().use_double = o.use_double
solver_algo = (md_nnps.MDNNPSSolver
if o.nnps == 'linear' else md_simple.MDSolver)
n_list = [10000 * (2 ** i) for i in range(10)] if o.nnps == 'linear' else \
[500 * (2 ** i) for i in range(8)]
if o.comp == "gpu_comp":
backends = ["opencl", "cuda", "cython"]
print("Running for", n_list)
speedups, t_list = compare(backends, n_list, solver_algo)
plot(n_list, speedups, t_list, o.nnps)
elif o.comp == "omp_comp":
backends = ["cython_omp", "cython"]
print("Running for", n_list)
speedups, t_list = compare(backends, n_list, solver_algo)
plot(n_list, speedups, t_list, o.nnps)
if __name__ == '__main__':
from argparse import ArgumentParser
p = ArgumentParser()
p.add_argument('-b',
'--backend',
action='store',
dest='backend',
default='cython',
help='Choose the backend.')
p.add_argument('--openmp',
action='store_true',
dest='openmp',
default=False,
help='Use OpenMP.')
p.add_argument('--use-double',
action='store_true',
dest='use_double',
default=False,
help='Use double precision on the GPU.')
p.add_argument('-n',
action='store',
type=int,
dest='n',
default=10000,
help='Number of particles.')
o = p.parse_args()
get_config().use_openmp = o.openmp
get_config().use_double = o.use_double
run(o.n, o.backend)
def tearDown(self):
super(TestAccelerationEval1DGPUOctreeNonCached, self).tearDown()
get_config().use_local_memory = self.old_flag
def setUp(self):
get_config().use_opencl = False
def tearDown(self):
super(BruteForceNNPSTestCase, self).tearDown()
get_config().use_double = self._orig_use_double
def setUp(self):
get_config().use_opencl = False
self.backend = None
def tearDown(self):
super(ZOrderGPUDoubleNNPSTestCase, self).tearDown()
get_config().use_double = self._orig_use_double
def teardown_module():
get_config().use_openmp = False