def translate_expression(self, expr):
for varname, var in self.variables.iteritems():
if isinstance(var, Function):
impl_name = var.implementations[self.codeobj_class].name
if impl_name is not None:
expr = word_substitute(expr, {varname: impl_name})
return CPPNodeRenderer().render_expr(expr).strip()
def freeze(code, ns):
# this is a bit of a hack, it should be passed to the template somehow
for k, v in ns.items():
if (isinstance(v, Variable) and not isinstance(v, AttributeVariable) and
v.scalar and v.constant and v.read_only):
v = v.get_value()
if isinstance(v, basestring):
code = word_substitute(code, {k: v})
elif isinstance(v, numbers.Number):
# Use a renderer to correctly transform constants such as True or inf
renderer = CPPNodeRenderer()
string_value = renderer.render_expr(repr(v))
if v < 0:
string_value = '(%s)' % string_value
code = word_substitute(code, {k: string_value})
else:
pass # don't deal with this object
return code
def freeze(code, ns):
# this is a bit of a hack, it should be passed to the template somehow
for k, v in ns.items():
if (isinstance(v, Variable) and not isinstance(v, AttributeVariable)
and v.scalar and v.constant and v.read_only):
v = v.get_value()
if isinstance(v, basestring):
code = word_substitute(code, {k: v})
elif isinstance(v, numbers.Number):
# Use a renderer to correctly transform constants such as True or inf
renderer = CPPNodeRenderer()
string_value = renderer.render_expr(repr(v))
if v < 0:
string_value = '(%s)' % string_value
code = word_substitute(code, {k: string_value})
else:
pass # don't deal with this object
return code
def translate_expression(self, expr):
for varname, var in self.variables.iteritems():
if isinstance(var, Function):
try:
impl_name = var.implementations[self.codeobj_class].name
except KeyError:
raise NotImplementedError('Function {} is not available '
'for use with '
'GeNN.'.format(var.name))
if impl_name is not None:
expr = word_substitute(expr, {varname: impl_name})
return CPPNodeRenderer().render_expr(expr).strip()
def fill_with_array(self, var, arr):
arr = np.asarray(arr)
array_name = self.get_array_name(var, access_data=False)
arr = np.asarray(arr)
if arr.shape == ():
if var.size == 1:
value = CPPNodeRenderer().render_expr(repr(arr.item(0)))
# For a single assignment, generate a code line instead of storing the array
self.main_queue.append(
('set_by_single_value', (array_name, 0, value)))
else:
self.main_queue.append(
('set_by_constant', (array_name, arr.item())))
else:
# Using the std::vector instead of a pointer to the underlying
# data for dynamic arrays is fast enough here and it saves us some
# additional work to set up the pointer
static_array_name = self.static_array(array_name, arr)
self.main_queue.append(
('set_by_array', (array_name, static_array_name)))
def translate_expression(self, expr):
expr = word_substitute(expr, self.func_name_replacements)
return CPPNodeRenderer().render_expr(expr).strip()
def test_parse_expressions_cpp():
if prefs.codegen.target != 'weave':
pytest.skip('weave-only test')
parse_expressions(CPPNodeRenderer(), cpp_evaluator)
def test_parse_expressions_cpp():
parse_expressions(CPPNodeRenderer(), cpp_evaluator)
def generate_main_source(self, writer):
main_lines = []
procedures = [('', main_lines)]
runfuncs = {}
for func, args in self.main_queue:
if func=='run_code_object':
codeobj, = args
main_lines.append('_run_%s();' % codeobj.name)
elif func=='run_network':
net, netcode = args
main_lines.extend(netcode)
elif func=='set_by_constant':
arrayname, value, is_dynamic = args
size_str = arrayname+'.size()' if is_dynamic else '_num_'+arrayname
code = '''
{pragma}
for(int i=0; i<{size_str}; i++)
{{
{arrayname}[i] = {value};
}}
'''.format(arrayname=arrayname, size_str=size_str,
value=CPPNodeRenderer().render_expr(repr(value)),
pragma=openmp_pragma('static'))
main_lines.extend(code.split('\n'))
elif func=='set_by_array':
arrayname, staticarrayname, is_dynamic = args
size_str = arrayname+'.size()' if is_dynamic else '_num_'+arrayname
code = '''
{pragma}
for(int i=0; i<{size_str}; i++)
{{
{arrayname}[i] = {staticarrayname}[i];
}}
'''.format(arrayname=arrayname, size_str=size_str,
staticarrayname=staticarrayname,
pragma=openmp_pragma('static'))
main_lines.extend(code.split('\n'))
elif func=='set_by_single_value':
arrayname, item, value = args
code = '{arrayname}[{item}] = {value};'.format(arrayname=arrayname,
item=item,
value=value)
main_lines.extend([code])
elif func=='set_array_by_array':
arrayname, staticarrayname_index, staticarrayname_value = args
code = '''
{pragma}
for(int i=0; i<_num_{staticarrayname_index}; i++)
{{
{arrayname}[{staticarrayname_index}[i]] = {staticarrayname_value}[i];
}}
'''.format(arrayname=arrayname, staticarrayname_index=staticarrayname_index,
staticarrayname_value=staticarrayname_value, pragma=openmp_pragma('static'))
main_lines.extend(code.split('\n'))
elif func=='resize_array':
array_name, new_size = args
main_lines.append("{array_name}.resize({new_size});".format(array_name=array_name,
new_size=new_size))
elif func=='insert_code':
main_lines.append(args)
elif func=='start_run_func':
name, include_in_parent = args
if include_in_parent:
main_lines.append('%s();' % name)
main_lines = []
procedures.append((name, main_lines))
elif func=='end_run_func':
name, include_in_parent = args
name, main_lines = procedures.pop(-1)
runfuncs[name] = main_lines
name, main_lines = procedures[-1]
elif func=='seed':
seed = args
nb_threads = prefs.devices.cpp_standalone.openmp_threads
if nb_threads == 0: # no OpenMP
nb_threads = 1
main_lines.append('for (int _i=0; _i<{nb_threads}; _i++)'.format(nb_threads=nb_threads))
if seed is None: # random
main_lines.append(' rk_randomseed(brian::_mersenne_twister_states[_i]);')
else:
main_lines.append(' rk_seed({seed!r}L + _i, brian::_mersenne_twister_states[_i]);'.format(seed=seed))
else:
raise NotImplementedError("Unknown main queue function type "+func)
self.runfuncs = runfuncs
# generate the finalisations
for codeobj in self.code_objects.itervalues():
if hasattr(codeobj.code, 'main_finalise'):
main_lines.append(codeobj.code.main_finalise)
# The code_objects are passed in the right order to run them because they were
# sorted by the Network object. To support multiple clocks we'll need to be
# smarter about that.
main_tmp = self.code_object_class().templater.main(None, None,
main_lines=main_lines,
code_objects=self.code_objects.values(),
report_func=self.report_func,
dt=float(self.defaultclock.dt),
user_headers=self.headers
)
writer.write('main.cpp', main_tmp)
def generate_main_source(self, writer):
main_lines = []
procedures = [('', main_lines)]
runfuncs = {}
for func, args in self.main_queue:
if func == 'run_code_object':
codeobj, = args
main_lines.append('_run_%s();' % codeobj.name)
elif func == 'run_network':
net, netcode = args
main_lines.extend(netcode)
elif func == 'set_by_constant':
arrayname, value = args
code = '''
{pragma}
for(int i=0; i<_num_{arrayname}; i++)
{{
{arrayname}[i] = {value};
}}
'''.format(arrayname=arrayname,
value=CPPNodeRenderer().render_expr(repr(value)),
pragma=openmp_pragma('static'))
main_lines.extend(code.split('\n'))
elif func == 'set_by_array':
arrayname, staticarrayname = args
code = '''
{pragma}
for(int i=0; i<_num_{staticarrayname}; i++)
{{
{arrayname}[i] = {staticarrayname}[i];
}}
'''.format(arrayname=arrayname,
staticarrayname=staticarrayname,
pragma=openmp_pragma('static'))
main_lines.extend(code.split('\n'))
elif func == 'set_by_single_value':
arrayname, item, value = args
code = '{arrayname}[{item}] = {value};'.format(
arrayname=arrayname, item=item, value=value)
main_lines.extend([code])
elif func == 'set_array_by_array':
arrayname, staticarrayname_index, staticarrayname_value = args
code = '''
{pragma}
for(int i=0; i<_num_{staticarrayname_index}; i++)
{{
{arrayname}[{staticarrayname_index}[i]] = {staticarrayname_value}[i];
}}
'''.format(arrayname=arrayname,
staticarrayname_index=staticarrayname_index,
staticarrayname_value=staticarrayname_value,
pragma=openmp_pragma('static'))
main_lines.extend(code.split('\n'))
elif func == 'resize_array':
array_name, new_size = args
main_lines.append("{array_name}.resize({new_size});".format(
array_name=array_name, new_size=new_size))
elif func == 'insert_code':
main_lines.append(args)
elif func == 'start_run_func':
name, include_in_parent = args
if include_in_parent:
main_lines.append('%s();' % name)
main_lines = []
procedures.append((name, main_lines))
elif func == 'end_run_func':
name, include_in_parent = args
name, main_lines = procedures.pop(-1)
runfuncs[name] = main_lines
name, main_lines = procedures[-1]
else:
raise NotImplementedError("Unknown main queue function type " +
func)
self.runfuncs = runfuncs
# generate the finalisations
for codeobj in self.code_objects.itervalues():
if hasattr(codeobj.code, 'main_finalise'):
main_lines.append(codeobj.code.main_finalise)
# The code_objects are passed in the right order to run them because they were
# sorted by the Network object. To support multiple clocks we'll need to be
# smarter about that.
main_tmp = CPPStandaloneCodeObject.templater.main(
None,
None,
main_lines=main_lines,
code_objects=self.code_objects.values(),
report_func=self.report_func,
dt=float(defaultclock.dt),
user_headers=prefs['codegen.cpp.headers'])
writer.write('main.cpp', main_tmp)
def variableview_set_with_index_array(self, variableview, item, value,
check_units):
if isinstance(item, slice) and item == slice(None):
item = 'True'
value = Quantity(value)
if (isinstance(item, int) or
(isinstance(item, np.ndarray)
and item.shape == ())) and value.size == 1:
array_name = self.get_array_name(variableview.variable,
access_data=False)
value = CPPNodeRenderer().render_expr(
repr(np.asarray(value).item(0)))
# For a single assignment, generate a code line instead of storing the array
self.main_queue.append(
('set_by_single_value', (array_name, item, value)))
elif (value.size == 1 and item == 'True'
and variableview.index_var_name == '_idx'):
# set the whole array to a scalar value
if have_same_dimensions(value, 1):
# Avoid a representation as "Quantity(...)" or "array(...)"
value = float(value)
variableview.set_with_expression_conditional(
cond=item, code=repr(value), check_units=check_units)
# Simple case where we don't have to do any indexing
elif (item == 'True' and variableview.index_var == '_idx'):
self.fill_with_array(variableview.variable, value)
else:
# We have to calculate indices. This will not work for synaptic
# variables
try:
indices = np.asarray(
variableview.indexing(item,
index_var=variableview.index_var))
except NotImplementedError:
raise NotImplementedError(
('Cannot set variable "%s" this way in '
'standalone, try using string '
'expressions.') % variableview.name)
# Using the std::vector instead of a pointer to the underlying
# data for dynamic arrays is fast enough here and it saves us some
# additional work to set up the pointer
arrayname = self.get_array_name(variableview.variable,
access_data=False)
if indices.shape == ():
self.main_queue.append(
('set_by_single_value', (arrayname, indices.item(),
float(value))))
else:
staticarrayname_index = self.static_array(
'_index_' + arrayname, indices)
staticarrayname_value = self.static_array(
'_value_' + arrayname, value)
self.main_queue.append(
('set_array_by_array', (arrayname, staticarrayname_index,
staticarrayname_value)))
staticarrayname_index = self.static_array('_index_' + arrayname,
indices)
if (indices.shape != () and
(value.shape == () or (value.size == 1 and indices.size > 1))):
value = np.repeat(value, indices.size)
staticarrayname_value = self.static_array('_value_' + arrayname,
value)
self.main_queue.append(
('set_array_by_array', (arrayname, staticarrayname_index,
staticarrayname_value)))
def generate_main_source(self, writer, main_includes):
main_lines = []
procedures = [('', main_lines)]
runfuncs = {}
for func, args in self.main_queue:
if func == 'run_code_object':
codeobj, = args
codeobj.runs_every_tick = False
main_lines.append('_run_%s();' % codeobj.name)
elif func == 'run_network':
net, netcode = args
for clock in net._clocks:
line = "{net.name}.add(&{clock.name}, _sync_clocks);".format(
clock=clock, net=net)
netcode.insert(1, line)
main_lines.extend(netcode)
elif func == 'set_by_constant':
arrayname, value, is_dynamic = args
size_str = arrayname + ".size()" if is_dynamic else "_num_" + arrayname
code = '''
for(int i=0; i<{size_str}; i++)
{{
{arrayname}[i] = {value};
}}
'''.format(arrayname=arrayname,
size_str=size_str,
value=CPPNodeRenderer().render_expr(repr(value)))
main_lines.extend(code.split('\n'))
pointer_arrayname = "dev{arrayname}".format(
arrayname=arrayname)
if arrayname.endswith('space'): # eventspace
pointer_arrayname += '[current_idx{arrayname}]'.format(
arrayname=arrayname)
if is_dynamic:
pointer_arrayname = "thrust::raw_pointer_cast(&dev{arrayname}[0])".format(
arrayname=arrayname)
line = "cudaMemcpy({pointer_arrayname}, &{arrayname}[0], sizeof({arrayname}[0])*{size_str}, cudaMemcpyHostToDevice);".format(
arrayname=arrayname,
size_str=size_str,
pointer_arrayname=pointer_arrayname,
value=CPPNodeRenderer().render_expr(repr(value)))
main_lines.extend([line])
elif func == 'set_by_single_value':
arrayname, item, value = args
pointer_arrayname = "dev{arrayname}".format(
arrayname=arrayname)
if arrayname in self.dynamic_arrays.values():
pointer_arrayname = "thrust::raw_pointer_cast(&dev{arrayname}[0])".format(
arrayname=arrayname)
code = '''
{arrayname}[{item}] = {value};
cudaMemcpy(&{pointer_arrayname}[{item}], &{arrayname}[{item}], sizeof({arrayname}[0]), cudaMemcpyHostToDevice);
'''.format(pointer_arrayname=pointer_arrayname,
arrayname=arrayname,
item=item,
value=value)
main_lines.extend([code])
elif func == 'set_by_array':
arrayname, staticarrayname, is_dynamic = args
size = "_num_" + arrayname
if is_dynamic:
size = arrayname + ".size()"
code = '''
for(int i=0; i<_num_{staticarrayname}; i++)
{{
{arrayname}[i] = {staticarrayname}[i];
}}
'''.format(arrayname=arrayname,
staticarrayname=staticarrayname,
size_str=size)
pointer_arrayname = "dev{arrayname}".format(
arrayname=arrayname)
if arrayname in self.dynamic_arrays.values():
pointer_arrayname = "thrust::raw_pointer_cast(&dev{arrayname}[0])".format(
arrayname=arrayname)
main_lines.extend(code.split('\n'))
line = '''
cudaMemcpy({pointer_arrayname}, &{arrayname}[0], sizeof({arrayname}[0])*{size_str}, cudaMemcpyHostToDevice);
'''.format(pointer_arrayname=pointer_arrayname,
staticarrayname=staticarrayname,
size_str=size,
arrayname=arrayname)
main_lines.extend([line])
elif func == 'set_array_by_array':
arrayname, staticarrayname_index, staticarrayname_value = args
code = '''
for(int i=0; i<_num_{staticarrayname_index}; i++)
{{
{arrayname}[{staticarrayname_index}[i]] = {staticarrayname_value}[i];
}}
cudaMemcpy(dev{arrayname}, &{arrayname}[0], sizeof({arrayname}[0])*_num_{arrayname}, cudaMemcpyHostToDevice);
'''.format(arrayname=arrayname,
staticarrayname_index=staticarrayname_index,
staticarrayname_value=staticarrayname_value)
main_lines.extend(code.split('\n'))
elif func == 'resize_array':
array_name, new_size = args
main_lines.append('''
{array_name}.resize({new_size});
dev{array_name}.resize({new_size});
'''.format(array_name=array_name, new_size=new_size))
elif func == 'insert_code':
main_lines.append(args)
elif func == 'start_run_func':
name, include_in_parent = args
if include_in_parent:
main_lines.append('%s();' % name)
main_lines = []
procedures.append((name, main_lines))
elif func == 'end_run_func':
name, include_in_parent = args
name, main_lines = procedures.pop(-1)
runfuncs[name] = main_lines
name, main_lines = procedures[-1]
elif func == 'seed':
seed = args
if seed is not None:
main_lines.append(
'curandSetPseudoRandomGeneratorSeed(random_float_generator, {seed!r}ULL);'
.format(seed=seed))
# generator offset needs to be reset to its default (=0)
main_lines.append(
'curandSetGeneratorOffset(random_float_generator, 0ULL);'
)
# else a random seed is set in objects.cu::_init_arrays()
else:
raise NotImplementedError("Unknown main queue function type " +
func)
# generate the finalisations
for codeobj in self.code_objects.itervalues():
if hasattr(codeobj.code, 'main_finalise'):
main_lines.append(codeobj.code.main_finalise)
main_tmp = CUDAStandaloneCodeObject.templater.main(
None,
None,
main_lines=main_lines,
code_objects=self.code_objects.values(),
report_func=self.report_func,
dt=float(defaultclock.dt),
additional_headers=main_includes)
writer.write('main.cu', main_tmp)
def translate_expression(self, expr):
return CPPNodeRenderer().render_expr(expr).strip()