def generate_loop(self, fields):
"""
The functions to generate stress loops and velocity loops are identical,
save for a single parameter. Moved the common code to this function to reduce repetition of code.
"""
if self.eval_const:
self.create_const_dict()
m = self.margin.value
body = []
for d in range(self.dimension-1, -1, -1):
i = self.index[d]
i0 = m
i1 = ccode(self.dim[d]-m)
if d == self.dimension-1:
# inner loop
if not self.fission:
body = self.simple_kernel(fields, [d, i, i0, i1])
else:
body = self.fission_kernel(fields, [d, i, i0, i1])
if not d == self.dimension-1:
body = [cgen.For(cgen.InlineInitializer(cgen.Value('int', i), i0), cgen.Line('%s<%s' % (i, i1)), cgen.Line('++%s' % i), cgen.Block(body))]
if not self.pluto and self.omp:
body.insert(0, cgen.Pragma('omp for schedule(static,1)'))
return cgen.Module(body)
def save_field_block(self, filename, field):
statements = []
statements.append(cgen.Initializer(cgen.Value("int", "dims[]"), "{dim1, dim1, dim1}"))
statements.append(cgen.Initializer(cgen.Value("float", "spacing[]"), "{dx1, dx2, dx3}"))
statements.append(cgen.Assign("std::string vtkfile", "\""+filename+"\" + std::to_string(_ti)"))
statements.append(cgen.Statement("opesci_dump_field_vts_3d(vtkfile, dims, spacing, 2, &"+field+"["+ccode(self.time[len(self.time)-1])+"][0][0][0])"))
return cgen.Module([cgen.Pragma("omp single"), cgen.Block(statements)])
def generate_second_initialisation(self):
loop = [Symbol('_'+x.name) for x in self.index] # symbols for loop
m = self.margin.value
statements = []
v = symbols("v")
for field in self.fields:
body = []
if self.omp:
statements.append(cgen.Pragma('omp for schedule(static,1)'))
# populate xvalue, yvalue zvalue code
for d in range(self.dimension-1, -1, -1):
i = loop[d]
i0 = m
i1 = ccode(self.dim[d]-m)
pre = []
post = []
if d == self.dimension-1:
# inner loop
# first time step
kernel = self.transform_kernel(field)
for arg in kernel.args:
if str(arg).startswith("-") and str(self.t - 1) in str(arg):
kernel = kernel.subs({arg: 0}, simultaneous=True)
arg = 2*v*self.dt
kernel = 0.5*(kernel + arg)
kernel = kernel.subs({self.t: self.time[0]})
for idx in self.index:
kernel = kernel.subs(idx, '_'+idx.name)
body = [cgen.Assign(ccode(field[[self.time[1]]+loop]), ccode(kernel))]
body = pre + body + post
body = [cgen.For(cgen.InlineInitializer(cgen.Value('int', i), i0), cgen.Line('%s<%s' % (i, i1)), cgen.Line('++%s' % i), cgen.Block(body))]
statements.append(body[0])
return statements
def initialise(self):
loop = [Symbol('_'+x.name) for x in self.index] # symbols for loop
statements = []
for field in self.fields:
body = []
if self.omp:
statements.append(cgen.Pragma('omp for schedule(static,1)'))
# populate xvalue, yvalue zvalue code
for d in range(self.dimension-1, -1, -1):
i = loop[d]
i0 = 0
i1 = ccode(self.dim[d])
pre = []
#velocity_initialisation = cgen.Assign(ccode())
post = []
if d == self.dimension-1:
# inner loop
# first time step
t0 = 0
sol = field.sol.subs(self.t, t0)
for idx in self.index:
sol = sol.subs(idx, '_'+idx.name)
body = [cgen.Assign(ccode(field[[0]+loop]), ccode(sol))]
body = pre + body + post
body = [cgen.For(cgen.InlineInitializer(cgen.Value('int', i), i0), cgen.Line('%s<%s' % (i, i1)), cgen.Line('++%s' % i), cgen.Block(body))]
statements.append(body[0])
statements += self.generate_second_initialisation()
return cgen.Module(statements)
def time_stepping(self):
"""
generate time index variable for time stepping
e.g. for 2nd order time-accurate scheme, varibales are t0, t1
for 4th order time-accurate scheme, variables are t0, t1, t2, t3
the variables are used to address the field arrays
e.g. in 2nd order scheme, U[t1] will be updated using U[t0]
the variables are calculated by taking mod with time periodicity
return generated code as string
"""
_ti = Symbol('_ti')
body = []
for i in range(len(self.time)):
lhs = self.time[i].name
if i == 0:
rhs = ccode(_ti % self.tp)
else:
rhs = ccode((self.time[i-1]+1) % self.tp)
body.append(cgen.Assign(lhs, rhs))
body = cgen.Block(body)
body = cgen.Module([cgen.Pragma('omp single'), body])
return body
def simple_kernel(self, grid_field, indexes):
"""
Generate the inner loop with all fields from stress or velocity
:param grid_field: stress or velocity field array
:param indexes: array with dimension, dimension var, initial margin, final margin
- iterate through fields and replace mako template
- return inner loop code as string
"""
body = []
idx = [self.time[len(self.time)-1]] + self.index
# This loop create the most inner loop with all fields
for field in grid_field:
body.append(cgen.Assign(ccode(field[idx]), ccode(self.kernel_sympy(field))))
body = [cgen.For(cgen.InlineInitializer(cgen.Value('int', indexes[1]), indexes[2]), cgen.Line('%s<%s' % (indexes[1], indexes[3])), cgen.Line('++%s' % indexes[1]), cgen.Block(body))]
if not self.pluto and self.ivdep and indexes[0] == self.dimension-1:
body.insert(0, self.compiler._ivdep)
if not self.pluto and self.simd and indexes[0] == self.dimension-1:
body.insert(0, cgen.Pragma('simd'))
return body
def _ivdep(self):
return cgen.Pragma('ivdep')