def get_pbc_kernels(self, runner):
dist_kernels = super(LBSingleFluidShanChen,
self).get_pbc_kernels(runner)
macro_kernels = defaultdict(lambda: defaultdict(list))
if self.config.node_addressing == 'indirect':
signature = 'PPi'
args = [runner.gpu_indirect_address()]
else:
signature = 'Pi'
args = []
for i in range(0, 3):
for field_pair in self._scalar_fields:
if not field_pair.abstract.need_nn:
continue
macro_kernels[0][i].append(
runner.get_kernel(
'ApplyMacroPeriodicBoundaryConditions', args +
[runner.gpu_field(field_pair.buffer),
np.uint32(i)], signature))
for i in range(0, 3):
for field_pair in self._scalar_fields:
if not field_pair.abstract.need_nn:
continue
macro_kernels[1][i].append(
runner.get_kernel(
'ApplyMacroPeriodicBoundaryConditions', args +
[runner.gpu_field(field_pair.buffer),
np.uint32(i)], signature))
ret = subdomain_runner.MacroKernels(macro=macro_kernels,
distributions=dist_kernels)
return ret
def get_pbc_kernels(self, runner):
gpu_dist1a = runner.gpu_dist(0, 0)
gpu_dist1b = runner.gpu_dist(0, 1)
gpu_dist2a = runner.gpu_dist(1, 0)
gpu_dist2b = runner.gpu_dist(1, 1)
# grid type (primary, secondary) -> axis -> kernels
dist_kernels = defaultdict(lambda: defaultdict(list))
macro_kernels = defaultdict(lambda: defaultdict(list))
if self.config.node_addressing == 'indirect':
args = [runner.gpu_indirect_address()]
signature = 'PPi'
else:
args = []
signature = 'Pi'
for i in range(0, 3):
dist_kernels[0][i] = [
runner.get_kernel('ApplyPeriodicBoundaryConditions',
args + [gpu_dist1a, np.uint32(i)],
signature),
runner.get_kernel('ApplyPeriodicBoundaryConditions',
args + [gpu_dist2a, np.uint32(i)], signature)
]
for field_pair in self._scalar_fields:
if not field_pair.abstract.need_nn:
continue
macro_kernels[0][i].append(
runner.get_kernel(
'ApplyMacroPeriodicBoundaryConditions', args +
[runner.gpu_field(field_pair.buffer),
np.uint32(i)], signature))
if self.config.access_pattern == 'AB':
gpu_dist1 = gpu_dist1b
gpu_dist2 = gpu_dist2b
kernel = 'ApplyPeriodicBoundaryConditions'
else:
gpu_dist1 = gpu_dist1a
gpu_dist2 = gpu_dist2a
kernel = 'ApplyPeriodicBoundaryConditionsWithSwap'
for i in range(0, 3):
if self.config.node_addressing == 'indirect':
args2 = [runner.gpu_indirect_address()]
sig2 = 'PPi'
else:
args2 = []
sig2 = 'Pi'
dist_kernels[1][i] = [
runner.get_kernel(kernel,
args2 + [gpu_dist1, np.uint32(i)], sig2),
runner.get_kernel(kernel,
args2 + [gpu_dist2, np.uint32(i)], sig2)
]
# This is the same as above -- for macroscopic fields, there is no
# distinction between primary and secondary buffers.
for field_pair in self._scalar_fields:
if not field_pair.abstract.need_nn:
continue
macro_kernels[1][i].append(
runner.get_kernel(
'ApplyMacroPeriodicBoundaryConditions', args +
[runner.gpu_field(field_pair.buffer),
np.uint32(i)], signature))
ret = subdomain_runner.MacroKernels(macro=macro_kernels,
distributions=dist_kernels)
return ret