def ufunc_at_vectorisation(self, statement, variables, indices,
conditional_write_vars, created_vars,
used_variables):
if not self._use_ufunc_at_vectorisation:
raise VectorisationError()
# Avoids circular import
from brian2.devices.device import device
# See https://github.com/brian-team/brian2/pull/531 for explanation
used = set(get_identifiers(statement.expr))
used = used.intersection(k for k in list(variables.keys())
if k in indices and indices[k] != '_idx')
used_variables.update(used)
if statement.var in used_variables:
raise VectorisationError()
expr = NumpyNodeRenderer(
auto_vectorise=self.auto_vectorise).render_expr(statement.expr)
if statement.op == ':=' or indices[
statement.var] == '_idx' or not statement.inplace:
if statement.op == ':=':
op = '='
else:
op = statement.op
line = '{var} {op} {expr}'.format(var=statement.var,
op=op,
expr=expr)
elif statement.inplace:
if statement.op == '+=':
ufunc_name = '_numpy.add'
elif statement.op == '*=':
ufunc_name = '_numpy.multiply'
elif statement.op == '/=':
ufunc_name = '_numpy.divide'
elif statement.op == '-=':
ufunc_name = '_numpy.subtract'
else:
raise VectorisationError()
line = '{ufunc_name}.at({array_name}, {idx}, {expr})'.format(
ufunc_name=ufunc_name,
array_name=device.get_array_name(variables[statement.var]),
idx=indices[statement.var],
expr=expr)
line = self.conditional_write(
line,
statement,
variables,
conditional_write_vars=conditional_write_vars,
created_vars=created_vars)
else:
raise VectorisationError()
if len(statement.comment):
line += ' # ' + statement.comment
return line
def ufunc_at_vectorisation(self, statement, variables, indices,
conditional_write_vars, created_vars, index):
'''
'''
# Avoids circular import
from brian2.devices.device import device
# We assume that the code has passed the test for synapse order independence
expr = NumpyNodeRenderer().render_expr(statement.expr)
if statement.op == ':=' or indices[
statement.var] == '_idx' or not statement.inplace:
if statement.op == ':=':
op = '='
else:
op = statement.op
line = '{var} {op} {expr}'.format(var=statement.var,
op=op,
expr=expr)
elif statement.inplace:
if statement.op == '+=':
ufunc_name = '_numpy.add'
elif statement.op == '*=':
ufunc_name = '_numpy.multiply'
elif statement.op == '/=':
ufunc_name = '_numpy.divide'
elif statement.op == '-=':
ufunc_name = '_numpy.subtract'
else:
raise VectorisationError()
line = '{ufunc_name}.at({array_name}, {idx}, {expr})'.format(
ufunc_name=ufunc_name,
array_name=device.get_array_name(variables[statement.var]),
idx=indices[statement.var],
expr=expr)
line = self.conditional_write(
line,
statement,
variables,
conditional_write_vars=conditional_write_vars,
created_vars=created_vars)
else:
raise VectorisationError()
if len(statement.comment):
line += ' # ' + statement.comment
return line
def ufunc_at_vectorisation(self, statement, variables, indices,
conditional_write_vars, created_vars, used_variables):
if not self._use_ufunc_at_vectorisation:
raise VectorisationError()
# Avoids circular import
from brian2.devices.device import device
# See https://github.com/brian-team/brian2/pull/531 for explanation
used = set(get_identifiers(statement.expr))
used = used.intersection(k for k in variables.keys() if k in indices and indices[k]!='_idx')
used_variables.update(used)
if statement.var in used_variables:
raise VectorisationError()
expr = NumpyNodeRenderer().render_expr(statement.expr)
if statement.op == ':=' or indices[statement.var] == '_idx' or not statement.inplace:
if statement.op == ':=':
op = '='
else:
op = statement.op
line = '{var} {op} {expr}'.format(var=statement.var, op=op, expr=expr)
elif statement.inplace:
if statement.op == '+=':
ufunc_name = '_numpy.add'
elif statement.op == '*=':
ufunc_name = '_numpy.multiply'
elif statement.op == '/=':
ufunc_name = '_numpy.divide'
elif statement.op == '-=':
ufunc_name = '_numpy.subtract'
else:
raise VectorisationError()
line = '{ufunc_name}.at({array_name}, {idx}, {expr})'.format(
ufunc_name=ufunc_name,
array_name=device.get_array_name(variables[statement.var]),
idx=indices[statement.var],
expr=expr)
line = self.conditional_write(line, statement, variables,
conditional_write_vars=conditional_write_vars,
created_vars=created_vars)
else:
raise VectorisationError()
if len(statement.comment):
line += ' # ' + statement.comment
return line
def ufunc_at_vectorisation(self, statement, variables, indices,
conditional_write_vars, created_vars, index):
'''
'''
# Avoids circular import
from brian2.devices.device import device
# We assume that the code has passed the test for synapse order independence
expr = NumpyNodeRenderer().render_expr(statement.expr)
if statement.op == ':=' or indices[statement.var] == '_idx' or not statement.inplace:
if statement.op == ':=':
op = '='
else:
op = statement.op
line = '{var} {op} {expr}'.format(var=statement.var,
op=op,
expr=expr)
elif statement.inplace:
if statement.op == '+=':
ufunc_name = '_numpy.add'
elif statement.op == '*=':
ufunc_name = '_numpy.multiply'
elif statement.op == '/=':
ufunc_name = '_numpy.divide'
elif statement.op == '-=':
ufunc_name = '_numpy.subtract'
else:
raise VectorisationError()
line = '{ufunc_name}.at({array_name}, {idx}, {expr})'.format(ufunc_name=ufunc_name,
array_name=device.get_array_name(variables[statement.var]),
idx=indices[statement.var],
expr=expr)
line = self.conditional_write(line, statement, variables,
conditional_write_vars=conditional_write_vars,
created_vars=created_vars)
else:
raise VectorisationError()
if len(statement.comment):
line += ' # ' + statement.comment
return line
def determine_keywords(self):
# set up the restricted pointers, these are used so that the compiler
# knows there is no aliasing in the pointers, for optimisation
lines = []
# it is possible that several different variable names refer to the
# same array. E.g. in gapjunction code, v_pre and v_post refer to the
# same array if a group is connected to itself
handled_pointers = set()
template_kwds = {}
# again, do the import here to avoid a circular dependency.
from brian2.devices.device import get_device
device = get_device()
for varname, var in self.variables.iteritems():
if isinstance(var, ArrayVariable):
# This is the "true" array name, not the restricted pointer.
array_name = device.get_array_name(var)
pointer_name = self.get_array_name(var)
if pointer_name in handled_pointers:
continue
if getattr(var, 'ndim', 1) > 1:
continue # multidimensional (dynamic) arrays have to be treated differently
line = self.c_data_type(var.dtype) + ' * ' + self.restrict + pointer_name + ' = ' + array_name + ';'
lines.append(line)
handled_pointers.add(pointer_name)
pointers = '\n'.join(lines)
# set up the functions
user_functions = []
support_code = ''
hash_defines = ''
# set convertion types for standard C99 functions in device code
if prefs.codegen.generators.cuda.default_functions_integral_convertion == np.float64:
default_func_type = 'double'
other_func_type = 'float'
else: # np.float32
default_func_type = 'float'
other_func_type = 'double'
# set clip function to either use all float or all double arguments
# see #51 for details
if prefs['core.default_float_dtype'] == np.float64:
float_dtype = 'float'
else: # np.float32
float_dtype = 'double'
for varname, variable in self.variables.items():
if isinstance(variable, Function):
user_functions.append((varname, variable))
funccode = variable.implementations[self.codeobj_class].get_code(self.owner)
if varname in functions_C99:
funccode = funccode.format(default_type=default_func_type, other_type=other_func_type)
if varname == 'clip':
funccode = funccode.format(float_dtype=float_dtype)
if isinstance(funccode, basestring):
funccode = {'support_code': funccode}
if funccode is not None:
support_code += '\n' + deindent(funccode.get('support_code', ''))
hash_defines += '\n' + deindent(funccode.get('hashdefine_code', ''))
# add the Python function with a leading '_python', if it
# exists. This allows the function to make use of the Python
# function via weave if necessary (e.g. in the case of randn)
if not variable.pyfunc is None:
pyfunc_name = '_python_' + varname
if pyfunc_name in self.variables:
logger.warn(('Namespace already contains function %s, '
'not replacing it') % pyfunc_name)
else:
self.variables[pyfunc_name] = variable.pyfunc
# delete the user-defined functions from the namespace and add the
# function namespaces (if any)
for funcname, func in user_functions:
del self.variables[funcname]
func_namespace = func.implementations[self.codeobj_class].get_namespace(self.owner)
if func_namespace is not None:
self.variables.update(func_namespace)
support_code += '\n' + deindent(self.universal_support_code)
keywords = {'pointers_lines': stripped_deindented_lines(pointers),
'support_code_lines': stripped_deindented_lines(support_code),
'hashdefine_lines': stripped_deindented_lines(hash_defines),
'denormals_code_lines': stripped_deindented_lines(self.denormals_to_zero_code()),
'uses_atomics': self.uses_atomics
}
keywords.update(template_kwds)
return keywords
def determine_keywords(self):
# set up the restricted pointers, these are used so that the compiler
# knows there is no aliasing in the pointers, for optimisation
pointers = []
# Add additional lines inside the kernel functions
kernel_lines = []
# It is possible that several different variable names refer to the
# same array. E.g. in gapjunction code, v_pre and v_post refer to the
# same array if a group is connected to itself
handled_pointers = set()
template_kwds = {}
# Again, do the import here to avoid a circular dependency.
from brian2.devices.device import get_device
device = get_device()
for varname, var in self.variables.iteritems():
if isinstance(var, ArrayVariable):
# This is the "true" array name, not the restricted pointer.
array_name = device.get_array_name(var)
pointer_name = self.get_array_name(var)
if pointer_name in handled_pointers:
continue
if getattr(var, 'ndim', 1) > 1:
continue # multidimensional (dynamic) arrays have to be treated differently
restrict = self.restrict
# turn off restricted pointers for scalars for safety
if var.scalar:
restrict = ' '
line = '{0}* {1} {2} = {3};'.format(
self.c_data_type(var.dtype), restrict, pointer_name,
array_name)
pointers.append(line)
handled_pointers.add(pointer_name)
# set up the functions
user_functions = []
support_code = []
hash_defines = []
for varname, variable in self.variables.items():
if isinstance(variable, Function):
hd, ps, sc, uf, kl = self._add_user_function(varname, variable)
user_functions.extend(uf)
support_code.extend(sc)
pointers.extend(ps)
hash_defines.extend(hd)
kernel_lines.extend(kl)
support_code.append(self.universal_support_code)
# Clock variables (t, dt, timestep) are passed by value to kernels and
# need to be translated back into pointers for scalar/vector code.
for varname, variable in self.variables.iteritems():
if hasattr(variable, 'owner') and isinstance(
variable.owner, Clock):
# get arrayname without _ptr suffix (e.g. _array_defaultclock_dt)
arrayname = self.get_array_name(variable, pointer=False)
line = "const {dtype}* _ptr{arrayname} = &_value{arrayname};"
line = line.format(dtype=c_data_type(variable.dtype),
arrayname=arrayname)
if line not in kernel_lines:
kernel_lines.append(line)
keywords = {
'pointers_lines':
stripped_deindented_lines('\n'.join(pointers)),
'support_code_lines':
stripped_deindented_lines('\n'.join(support_code)),
'hashdefine_lines':
stripped_deindented_lines('\n'.join(hash_defines)),
'denormals_code_lines':
stripped_deindented_lines('\n'.join(
self.denormals_to_zero_code())),
'kernel_lines':
stripped_deindented_lines('\n'.join(kernel_lines)),
'uses_atomics':
self.uses_atomics
}
keywords.update(template_kwds)
return keywords
def determine_keywords(self):
# set up the restricted pointers, these are used so that the compiler
# knows there is no aliasing in the pointers, for optimisation
pointers = []
# Add additional lines inside the kernel functions
kernel_lines = []
# It is possible that several different variable names refer to the
# same array. E.g. in gapjunction code, v_pre and v_post refer to the
# same array if a group is connected to itself
handled_pointers = set()
template_kwds = {}
# Again, do the import here to avoid a circular dependency.
from brian2.devices.device import get_device
device = get_device()
for varname, var in self.variables.items():
if isinstance(var, ArrayVariable):
# This is the "true" array name, not the restricted pointer.
array_name = device.get_array_name(var)
pointer_name = self.get_array_name(var)
if pointer_name in handled_pointers:
continue
if getattr(var, 'ndim', 1) > 1:
continue # multidimensional (dynamic) arrays have to be treated differently
restrict = self.restrict
# turn off restricted pointers for scalars for safety
if var.scalar:
restrict = ' '
# Need to use correct dt type in pointers_lines for single precision,
# see #148
if varname == "dt" and prefs.core.default_float_dtype == np.float32:
# c_data_type(variable.dtype) is float, but we need double
dtype = "double"
else:
dtype = self.c_data_type(var.dtype)
line = '{0}* {1} {2} = {3};'.format(dtype,
restrict,
pointer_name,
array_name)
pointers.append(line)
handled_pointers.add(pointer_name)
# set up the functions
user_functions = []
support_code = []
hash_defines = []
added = set() # keep track of functions that were added
for varname, variable in list(self.variables.items()):
if isinstance(variable, Function):
user_func = self._add_user_function(varname, variable, added)
if user_func is not None:
hd, ps, sc, uf, kl = user_func
user_functions.extend(uf)
support_code.extend(sc)
pointers.extend(ps)
hash_defines.extend(hd)
kernel_lines.extend(kl)
# Generate universal_support_code once when the first codeobject is created.
# Can't do it at import time since need access to user preferences
# This is a class attribute (not instance attribute).
if CUDACodeGenerator.universal_support_code is None:
_atomic_support_code = _generate_atomic_support_code()
CUDACodeGenerator.universal_support_code = (
_hightype_support_code
+ _mod_support_code
+ _floordiv_support_code
+ _pow_support_code
+ _atomic_support_code
)
support_code.append(CUDACodeGenerator.universal_support_code)
# Clock variables (t, dt, timestep) are passed by value to kernels and
# need to be translated back into pointers for scalar/vector code.
for varname, variable in self.variables.items():
if hasattr(variable, 'owner') and isinstance(variable.owner, Clock):
# get arrayname without _ptr suffix (e.g. _array_defaultclock_dt)
arrayname = self.get_array_name(variable, prefix='')
# kernel_lines appear before dt is cast to float (in scalar_code), hence
# we need to still use double (used in kernel parameters), see #148
if varname == "dt" and prefs.core.default_float_dtype == np.float32:
# c_data_type(variable.dtype) is float, but we need double
dtype = "double"
else:
dtype = dtype=c_data_type(variable.dtype)
line = f"const {dtype}* _ptr{arrayname} = &_value{arrayname};"
if line not in kernel_lines:
kernel_lines.append(line)
keywords = {'pointers_lines': stripped_deindented_lines('\n'.join(pointers)),
'support_code_lines': stripped_deindented_lines('\n'.join(support_code)),
'hashdefine_lines': stripped_deindented_lines('\n'.join(hash_defines)),
'denormals_code_lines': stripped_deindented_lines('\n'.join(self.denormals_to_zero_code())),
'kernel_lines': stripped_deindented_lines('\n'.join(kernel_lines)),
'uses_atomics': self.uses_atomics
}
keywords.update(template_kwds)
return keywords