def load_fields(self):
"""Code fragment that loads field arrays from 'grid' struct"""
idxs = ''.join(['[%d]' % d.value for d in self.dim])
return '\n'.join(['%s (*%s)%s = (%s (*)%s) grid->%s;' %
(self.real_t, ccode(f.label), idxs,
self.real_t, idxs, ccode(f.label))
for f in self.fields])
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 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
"""
result = ''
tmpl = self.lookup.get_template('time_stepping.txt')
_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 += lhs + ' = ' + rhs + ';\n'
dict1 = {'body': body}
result = render(tmpl, dict1)
return result
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 declare_fields(self):
"""
- generate code for delcaring fields
- the generated code first declare fields as std::vector
of size=vec_size, then cast to multidimensional array
- return the generated code as string
"""
result = ''
arr = '' # = [dim1][dim2][dim3]...
for d in self.dim:
arr += '[' + d.name + ']'
vsize = 1
for d in self.dim:
vsize *= d.value
vsize *= self.order[0]*2
for field in self.sfields + self.vfields:
vec = '_' + ccode(field.label) + '_vec'
# alloc aligned memory (on windows and linux)
result += self.real_t + ' *' + vec + ';\n'
result += '#ifdef _MSC_VER\n'
result += vec + ' = (' + self.real_t + '*) _aligned_malloc(' + str(vsize) \
+ '*sizeof(' + self.real_t + '), ' + str(self.alignment) + ');\n'
result += '#else\n'
result += 'posix_memalign((void **)(&' + vec + '), ' + str(self.alignment) \
+ ', ' + str(vsize) + '*sizeof(' + self.real_t + '));\n'
result += '#endif\n'
# cast pointer to multidimensional array
result += self.real_t + ' (*' + ccode(field.label) + ')' + arr \
+ '= (' + self.real_t + ' (*)' + arr + ') ' + vec + ';\n'
if self.read:
# add code to read data
result += self.read_data()
return result
def declare_fields(self):
"""
- generate code for declaring fields
- the generated code first declare fields as std::vector
of size=vec_size, then cast to multidimensional array
- return the generated code as string
"""
result = []
arr = '' # = [dim1][dim2][dim3]...
for d in self.dim:
arr += '[' + d.name + ']'
vsize = 1
for d in self.dim:
vsize *= d.value
vsize *= len(self.time)
statements = []
for field in self.fields:
vec = "_%s_vec" % ccode(field.label)
vec_value = cgen.Pointer(cgen.Value(self.real_t, vec))
# alloc aligned memory (on windows and linux)
statements.append(vec_value)
ifdef = cgen.IfDef('_MSC_VER', [cgen.Assign(vec, '(%s*) _aligned_malloc(%s*sizeof(%s), %s)' % (self.real_t, str(vsize), self.real_t, str(self.alignment)))],
[cgen.Statement('posix_memalign((void **)(&%s), %d, %d*sizeof(%s))' % (vec, self.alignment, vsize, self.real_t))])
statements.append(ifdef)
# cast pointer to multidimensional array
cast_pointer = cgen.Initializer(cgen.Value(self.real_t, "(*%s)%s" % (ccode(field.label), arr)), '(%s (*)%s) %s' % (self.real_t, arr, vec))
statements.append(cast_pointer)
vec = "_%s_vec" % ccode("m")
vec_value = cgen.Pointer(cgen.Value(self.real_t, vec))
statements.append(vec_value)
result += statements
return cgen.Module(result)
def output_step(self):
"""
- generate code for output at each time step
- typically output selected fields in vtk format
- return generated code as string
"""
if self.output_vts:
return self.save_field_block(ccode(self.fields[0].label)+"_", ccode(self.fields[0].label))
return None
def vtk_save_field(self):
"""
generate code to output this field with vtk
uses Mako template
returns the generated code as string
"""
tmpl = self.lookup.get_template('save_field.txt')
result = ''
dict1 = {'filename': ccode(self.label)+'_', 'field': ccode(self.label)}
result = render(tmpl, dict1)
return result
def converge_test(self):
"""
- generate code for convergence test
- convergence test implemented by calculating L2 norm
of the simulation against analytical solution
- L2 norm of each field is calculated and output with printf()
- return generated code as string
"""
result = []
if not self.converge:
return cgen.Module(result)
m = self.margin.value
ti = self.ntsteps.value % 2 # last updated grid
loop = [Symbol('_'+x.name) for x in self.index] # symbols for loop
for i in range(len(self.spacing)):
result.append(cgen.Statement('printf("%d\\n")' % self.spacing[i].value))
for field in self.fields:
body = []
l2 = ccode(field.label)+'_l2'
idx = [ti] + loop
result.append(cgen.Initializer(cgen.Value(self.real_t, l2), 0.0))
# 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)
expr = self.spacing[d]*(loop[d] - self.margin.value)
pre = [cgen.Initializer(cgen.Value(self.real_t, self.index[d].name), ccode(expr))]
if d == self.dimension-1:
# inner loop
tn = self.dt.value*self.ntsteps.value \
if not field.staggered[0] \
else self.dt.value*self.ntsteps.value \
+ self.dt.value/2.0
body = [cgen.Statement('%s += %s' % (l2, ccode((field[idx] - (field.sol.subs(self.t, tn)))**2.0)))]
body = pre+body
body = [cgen.For(cgen.InlineInitializer(cgen.Value('int', i), i0), cgen.Line('%s<%s' % (i, i1)), cgen.Line('++%s' % i), cgen.Block(body))]
result += body
volume = 1.0
for i in range(len(self.spacing)):
volume *= self.spacing[i].value
l2_value = 'pow(' + l2 + '*' + ccode(volume) + ', 0.5)'
result.append(cgen.Statement('conv->%s = %s' % (l2, l2_value)))
return cgen.Module(result)
def initialise(self):
"""
generate code for initialisation of the fields
- substitute starting time to the analytical function of the fields
- substitute field coordinates calculated from array indices
to the analytical function of the fields
- generate inner loop by inserting kernel into Mako template
- recursive insertion to generate nested loop
return generated code as string
"""
tmpl = self.lookup.get_template('generic_loop.txt')
result = ''
m = self.margin.value
loop = [Symbol('_'+x.name) for x in self.index] # symbols for loop
for field in self.sfields+self.vfields:
body = ''
if self.omp:
result += '#pragma omp for\n'
# populate xvalue, yvalue zvalue code
for d in range(self.dimension-1, -1, -1):
i = loop[d]
i0 = m
if field.staggered[d+1]:
i1 = ccode(self.dim[d]-m-1)
expr = self.spacing[d]*(loop[d] - self.margin.value + 0.5)
else:
i1 = ccode(self.dim[d]-m)
expr = self.spacing[d]*(loop[d] - self.margin.value)
pre = self.real_t + ' ' + self.index[d].name + '= ' \
+ ccode(expr) + ';\n'
post = ''
if d == self.dimension-1:
# inner loop
# first time step
t0 = self.dt.value/2 if field.staggered[0] else 0
sol = field.sol.subs(self.t, t0)
if self.read:
sol = sol.subs(field.media_param)
for idx in self.index:
sol = sol.subs(idx, '_'+idx.name)
body = ccode(field[[0]+loop]) + '=' \
+ ccode(sol) + ';\n'
body = pre + body + post
dict1 = {'i': i, 'i0': i0, 'i1': i1, 'body': body}
body = render(tmpl, dict1)
result += body
return result
def vtk_save_field(self):
"""
generate code to output this field with vtk
uses Mako template
returns the generated code as string
"""
tmpl = self.lookup.get_template('save_field.txt')
result = ''
dict1 = {
'filename': ccode(self.label) + '_',
'field': ccode(self.label)
}
result = render(tmpl, dict1)
return result
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 initialise(self):
"""
generate code for initialisation of the fields
- substitute starting time to the analytical function of the fields
- substitute field coordinates calculated from array indices
to the analytical function of the fields
- generate inner loop by inserting kernel into Mako template
- recursive insertion to generate nested loop
return generated code as string
"""
m = self.margin.value
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 = m
if field.staggered[d+1]:
i1 = ccode(self.dim[d]-m-1)
expr = self.spacing[d]*(loop[d] - self.margin.value + 0.5)
else:
i1 = ccode(self.dim[d]-m)
expr = self.spacing[d]*(loop[d] - self.margin.value)
pre = [cgen.Initializer(cgen.Value(self.real_t, self.index[d].name), ccode(expr))]
post = []
if d == self.dimension-1:
# inner loop
# first time step
t0 = self.dt.value/2 if field.staggered[0] else 0
sol = field.sol.subs(self.t, t0)
if self.read:
sol = sol.subs(self.media_dict)
sol = self.resolve_media_params(sol)
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])
return cgen.Module(statements)
def define_fields(self):
"""Code fragment that defines field arrays"""
result = []
for f in self.fields:
var = cgen.Pointer(cgen.Value(self.real_t, ccode(f.label)))
result.append(var)
return cgen.Module(result)
def store_fields(self):
"""Code fragment that stores field arrays to 'grid' struct"""
result = []
for f in self.fields:
assignment = cgen.Assign('grid->%s' % ccode(f.label), '(%s*) %s' % (self.real_t, ccode(f.label))) # There must be a better way of doing this. This hardly seems better than string manipulation
result.append(assignment)
return cgen.Module(result)
def load_fields(self):
"""Code fragment that loads field arrays from 'grid' struct"""
idxs = ''.join(['[%d]' % d.value for d in self.dim])
result = []
for f in self.fields:
back_assign = cgen.Initializer(cgen.Value(self.real_t, "(*%s)%s" % (ccode(f.label), idxs)), '(%s (*)%s) grid->%s' % (self.real_t, idxs, ccode(f.label))) # Another hackish attempt.
result.append(back_assign)
return cgen.Module(result)
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 set_free_surface(self, indices, d, b, side, read=False):
"""
- set free surface boundary condition to boundary d, at index b
:param indices: list of indices, e.g. [t,x,y,z] for 3D
:param d: direction of the boundary surface normal
:param b: location of the boundary (index)
:param side: lower boundary (0) or upper boundary (1)
- e.g. set_free_surface([t,x,y,z],1,2,0)
set y-z plane at x=2 to be lower free surface
- ghost cells are calculated using reflection of stress fields
- store the code to populate ghost cells to self.bc
"""
# use this stress field to solve for ghost cell expression
field = self.sfields[d]
idx = list(indices)
if not field.staggered[d]:
# if not staggered, solve ghost cell using T'[b]=0
eq = Eq(field.dt)
shift = hf
t = b - hf
else:
# if staggered, solve ghost cell using T'[b-1/2]=T[b+1/2]
eq = Eq(field.dt.subs(indices[d], indices[d]-hf),
field.dt.subs(indices[d], indices[d]+hf))
shift = 1
t = b
idx[d] -= ((-1)**side)*shift
lhs = self[idx]
rhs = solve(eq, lhs)[0]
lhs = lhs.subs(indices[d], t)
rhs = self.align(rhs.subs(indices[d], t))
# if read data from file, replace media parameters with array
# replace particular index with boundary
if read:
rhs = rhs.subs(self.media_param)
rhs = rhs.subs(indices[d], b)
self.bc[d][side] = ccode(lhs) + ' = ' + ccode(rhs) + ';\n'
def free_memory(self):
"""
- generate code for free allocated memory
- return the generated code as string
"""
statements = []
for field in self.fields:
# alloc aligned memory (on windows and linux)
ifdef = cgen.IfDef('_MSC_VER', [cgen.Statement('_aligned_free(grid->%s)' % (ccode(field.label)))],
[cgen.Statement('free(grid->%s)' % (ccode(field.label)))])
statements.append(ifdef)
return cgen.Module(statements)
def converge_test(self):
"""
- generate code for convergence test
- convergence test implemented by calculating L2 norm
of the simulation against analytical solution
- L2 norm of each field is calculated and output with printf()
- return generated code as string
"""
result = ''
if not self.converge:
return result
tmpl = self.lookup.get_template('generic_loop.txt')
m = self.margin.value
ti = self.ntsteps.value % 2 # last updated grid
loop = [Symbol('_'+x.name) for x in self.index] # symbols for loop
for i in range(len(self.spacing)):
result += 'printf("' + str(self.spacing[i].value) + '\\n");\n'
for field in self.sfields+self.vfields:
body = ''
l2 = ccode(field.label)+'_l2'
idx = [ti] + loop
result += self.real_t + ' ' + l2 + ' = 0.0;\n'
# populate xvalue, yvalue zvalue code
for d in range(self.dimension-1, -1, -1):
i = loop[d]
i0 = m
if field.staggered[d+1]:
i1 = ccode(self.dim[d]-m-1)
expr = self.spacing[d]*(loop[d] - self.margin.value + 0.5)
else:
i1 = ccode(self.dim[d]-m)
expr = self.spacing[d]*(loop[d] - self.margin.value)
pre = self.real_t + ' ' + self.index[d].name + '= ' \
+ ccode(expr) + ';\n'
if d == self.dimension-1:
# inner loop
tn = self.dt.value*self.ntsteps.value \
if not field.staggered[0] \
else self.dt.value*self.ntsteps.value \
+ self.dt.value/2.0
body = l2 + '+=' \
+ ccode((field[idx] -
(field.sol.subs(self.t, tn)))**2.0) + ';\n'
body = pre + body
dict1 = {'i': i, 'i0': i0, 'i1': i1, 'body': body}
body = render(tmpl, dict1)
result += body
volume = 1.0
for i in range(len(self.spacing)):
volume *= self.spacing[i].value
l2_value = 'pow(' + l2 + '*' + ccode(volume) + ', 0.5)'
result += 'conv->%s = %s;\n' % (l2, l2_value)
return result
def velocity_loop(self):
"""
generate code for velocity field update loop
- loop through velocity fields to generate code of computation kernel
- generate inner loop by inserting kernel into Mako template
- recursive insertion to generate nested loop
return generated code as string
"""
tmpl = self.lookup.get_template('generic_loop.txt')
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
idx = [self.time[1]] + self.index
for field in self.vfields:
body += ccode(field[idx]) + '=' \
+ ccode(field.fd_align.xreplace({self.t+1:
self.time[1],
self.t:
self.time[0]})) \
+ ';\n'
dict1 = {'i': i, 'i0': i0, 'i1': i1, 'body': body}
body = render(tmpl, dict1)
if self.ivdep and d == self.dimension-1:
body = '%s\n' % self.compiler._ivdep + body
if self.simd and d == self.dimension-1:
body = '#pragma simd\n' + body
if self.omp:
body = '#pragma omp for\n' + body
return body
def simple_loop(self, kernel):
"""
- helper function to generate simple nested loop over the entire domain
(not including ghost cells) with kernel at the inner loop
- variables defined in self.index are used as loop variables
"""
result = ''
tmpl = self.lookup.get_template('generic_loop.txt')
m = self.margin.value
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
result += kernel + ';\n'
dict1 = {'i': i, 'i0': i0, 'i1': i1, 'body': result}
result = render(tmpl, dict1)
return result
class OpesciConvergence(Structure):
_fields_ = [('%s_l2' % ccode(f.label), c_float)
for f in self.fields]
class OpesciGrid(Structure):
_fields_ = [(ccode(f.label), POINTER(c_float))
for f in self.fields]
def read_data(self):
"""
- generate code for reading data (rho, Vp, Vs) from input files
- calculate effective media parameters beta, lambda, mu from the data
"""
result = ''
if self.read:
arr = '' # =[dim2][dim3]...
for d in self.dim[1:]:
arr += '[' + d.name + ']'
vsize = 1
for d in self.dim:
vsize *= d.value
# declare fields to read physical parameters from file
# always use float not double
loop = [self.rho, self.vp, self.vs] + self.beta + [self.lam] + self.mu
for field in loop:
vec = '_' + ccode(field.label) + '_vec'
# alloc aligned memory (on windows and linux)
result += self.real_t + ' *' + vec + ';\n'
result += '#ifdef _MSC_VER\n'
result += vec + ' = (' + self.real_t + '*) _aligned_malloc(' + str(vsize) \
+ '*sizeof(' + self.real_t + '), ' + str(self.alignment) + ');\n'
result += '#else\n'
result += 'posix_memalign((void **)(&' + vec + '), ' + str(self.alignment) \
+ ', ' + str(vsize) + '*sizeof(' + self.real_t + '));\n'
result += '#endif\n'
# cast pointer to multidimensional array
result += self.real_t + ' (*' + ccode(field.label) + ')' + arr \
+ '= (' + self.real_t + ' (*)' + arr + ') ' + vec + ';\n'
# read from file
result += 'opesci_read_simple_binary_ptr("' + self.rho_file + '",_' \
+ ccode(self.rho.label) + '_vec, ' + str(vsize) + ');\n'
result += 'opesci_read_simple_binary_ptr("' + self.vp_file + '",_' \
+ ccode(self.vp.label) + '_vec, ' + str(vsize) + ');\n'
result += 'opesci_read_simple_binary_ptr("' + self.vs_file + '",_' \
+ ccode(self.vs.label) + '_vec, ' + str(vsize) + ');\n'
# calculated effective media parameter
idx = self.index
# make copies of index
idx100 = list(idx)
idx010 = list(idx)
idx001 = list(idx)
idx110 = list(idx)
idx101 = list(idx)
idx011 = list(idx)
# shift the indices to obtain idx100=[x+1,y,z] etc
idx100[0] += 1
idx010[1] += 1
idx001[2] += 1
idx110[0] += 1
idx110[1] += 1
idx101[0] += 1
idx101[2] += 1
idx011[1] += 1
idx011[2] += 1
# beta
kernel = ccode(self.beta[0][idx]) + '=' + ccode(1.0/self.rho[idx])
result += self.simple_loop(kernel)
# beta1 (effective bouyancy in x direction)
kernel = ccode(self.beta[1][idx]) + '=' \
+ ccode((self.beta[0][idx] + self.beta[0][idx100])/2.0)
result += self.simple_loop(kernel)
# beta2 (effective bouyancy in y direction)
kernel = ccode(self.beta[2][idx]) + '=' \
+ ccode((self.beta[0][idx] + self.beta[0][idx010])/2.0)
result += self.simple_loop(kernel)
# beta3 (effective bouyancy in z direction)
kernel = ccode(self.beta[3][idx]) + '=' + \
ccode((self.beta[0][idx] + self.beta[0][idx001])/2.0)
result += self.simple_loop(kernel)
# lambda
kernel = ccode(self.lam[idx]) + '=' + \
ccode(self.rho[idx]*(self.vp[idx]**2-2*self.vs[idx]**2))
result += self.simple_loop(kernel)
# mu
kernel = ccode(self.mu[0][idx]) + '=' \
+ ccode(self.rho[idx]*(self.vs[idx]**2))
result += self.simple_loop(kernel)
# mu12 (effective shear modulus for shear stress sigma_xy)
kernel = ccode(self.mu[1][idx]) + '=' \
+ ccode(1.0/(0.25*(1.0/self.mu[0][idx]+1.0/self.mu[0][idx100]
+ 1.0/self.mu[0][idx010]+1.0/self.mu[0][idx110])))
result += self.simple_loop(kernel)
# mu13 (effective shear modulus for shear stress sigma_xz)
kernel = ccode(self.mu[2][idx]) + '=' \
+ ccode(1.0/(0.25*(1.0/self.mu[0][idx]+1.0/self.mu[0][idx100]
+ 1.0/self.mu[0][idx001]+1.0/self.mu[0][idx101])))
result += self.simple_loop(kernel)
# mu23 (effective shear modulus for shear stress sigma_yz)
kernel = ccode(self.mu[3][idx]) + '=' \
+ ccode(1.0/(0.25*(1.0/self.mu[0][idx]+1.0/self.mu[0][idx010]
+ 1.0/self.mu[0][idx001]+1.0/self.mu[0][idx011])))
result += self.simple_loop(kernel)
return result
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 print_convergence(self):
"""Code fragment that prints convergence norms"""
statements = [cgen.Statement('printf("%s %s\\n", conv.%s_l2)' % (ccode(f.label), '\t%.10f', ccode(f.label))) for f in self.fields]
return cgen.Module(statements)
def define_convergence(self):
"""Code fragment that defines convergence norms"""
result = []
for f in self.fields:
result.append(cgen.Value(self.real_t, '%s_l2' % ccode(f.label)))
return cgen.Module(result)
def read_data(self):
"""
- generate code for reading data (rho, Vp, Vs) from input files
- calculate effective media parameters beta, lambda, mu from the data
"""
statements = []
if self.read:
arr = '' # =[dim2][dim3]...
for d in self.dim[1:]:
arr += '[' + d.name + ']'
vsize = 1
for d in self.dim:
vsize *= d.value
# declare fields to read physical parameters from file
# always use float not double
loop = [self.rho, self.vp, self.vs] + self.beta + [self.lam] + self.mu
for field in loop:
vec = "_%s_vec" % ccode(field.label)
vec_value = cgen.Pointer(cgen.Value(self.real_t, vec))
# alloc aligned memory (on windows and linux)
statements.append(vec_value)
ifdef = cgen.IfDef('_MSC_VER', [cgen.Assign(vec, '(%s*) _aligned_malloc(%d * sizeof(%s), %d)' % (self.real_t, vsize, self.real_t, self.alignment))],
[cgen.Statement('posix_memalign((void **)(&%s), %d, %d*sizeof(%s))' % (vec, self.alignment, vsize, self.real_t))])
statements.append(ifdef)
cast_pointer = cgen.Initializer(cgen.Value(self.real_t, "(*%s)%s" % (ccode(field.label), arr)), '(%s (*)%s) %s' % (self.real_t, arr, vec))
statements.append(cast_pointer)
# read from file
statements.append(cgen.Statement('opesci_read_simple_binary_ptr("%s", _%s_vec, %d)' % (self.rho_file, self.rho.label, vsize)))
statements.append(cgen.Statement('opesci_read_simple_binary_ptr("%s", _%s_vec, %d)' % (self.vp_file, self.vp.label, vsize)))
statements.append(cgen.Statement('opesci_read_simple_binary_ptr("%s", _%s_vec, %d)' % (self.vs_file, self.vs.label, vsize)))
# calculated effective media parameter
idx = self.index
# make copies of index
idx100 = list(idx)
idx010 = list(idx)
idx001 = list(idx)
idx110 = list(idx)
idx101 = list(idx)
idx011 = list(idx)
# shift the indices to obtain idx100=[x+1,y,z] etc
idx100[0] += 1
idx010[1] += 1
idx001[2] += 1
idx110[0] += 1
idx110[1] += 1
idx101[0] += 1
idx101[2] += 1
idx011[1] += 1
idx011[2] += 1
# beta
kernel = cgen.Assign(ccode(self.beta[0][idx]), ccode(1.0/self.rho[idx]))
statements.append(self.simple_loop(kernel))
# beta1 (effective bouyancy in x direction)
kernel = cgen.Assign(ccode(self.beta[1][idx]), ccode((self.beta[0][idx] + self.beta[0][idx100])/2.0))
statements.append(self.simple_loop(kernel))
# beta2 (effective bouyancy in y direction)
kernel = cgen.Assign(ccode(self.beta[2][idx]), ccode((self.beta[0][idx] + self.beta[0][idx010])/2.0))
statements.append(self.simple_loop(kernel))
# beta3 (effective bouyancy in z direction)
kernel = cgen.Assign(ccode(self.beta[3][idx]), ccode((self.beta[0][idx] + self.beta[0][idx001])/2.0))
statements.append(self.simple_loop(kernel))
# lambda
kernel = cgen.Assign(ccode(self.lam[idx]), ccode(self.rho[idx]*(self.vp[idx]**2-2*self.vs[idx]**2)))
statements.append(self.simple_loop(kernel))
# mu
kernel = cgen.Assign(ccode(self.mu[0][idx]), ccode(self.rho[idx]*(self.vs[idx]**2)))
statements.append(self.simple_loop(kernel))
# mu12 (effective shear modulus for shear stress sigma_xy)
kernel = cgen.Assign(ccode(self.mu[1][idx]), ccode(1.0/(0.25*(1.0/self.mu[0][idx]+1.0/self.mu[0][idx100] + 1.0/self.mu[0][idx010]+1.0/self.mu[0][idx110]))))
statements.append(self.simple_loop(kernel))
# mu13 (effective shear modulus for shear stress sigma_xz)
kernel = cgen.Assign(ccode(self.mu[2][idx]), ccode(1.0/(0.25*(1.0/self.mu[0][idx]+1.0/self.mu[0][idx100] + 1.0/self.mu[0][idx001]+1.0/self.mu[0][idx101]))))
statements.append(self.simple_loop(kernel))
# mu23 (effective shear modulus for shear stress sigma_yz)
kernel = cgen.Assign(ccode(self.mu[3][idx]), ccode(1.0/(0.25*(1.0/self.mu[0][idx]+1.0/self.mu[0][idx010] + 1.0/self.mu[0][idx001]+1.0/self.mu[0][idx011]))))
statements.append(self.simple_loop(kernel))
return statements
def simple_loop(self, kernel):
"""
- helper function to generate simple nested loop over the entire domain
(not including ghost cells) with kernel at the inner loop
- variables defined in self.index are used as loop variables
"""
result = kernel
m = self.margin.value
for d in range(self.dimension-1, -1, -1):
result = cgen.For(cgen.InlineInitializer(cgen.Value('int', self.index[d]), m), cgen.Line('%s<%s' % (self.index[d], ccode(self.dim[d]-m))), cgen.Line('++%s' % self.index[d]), result)
return result
def fission_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 for each dimension separate minus, plus and unitary strides
on its own loop replacing it on mako template
- return inner loop code as string
"""
body = []
body_tmp = []
operator = ['=', '+=']
idx = [self.time[1]] + self.index
for field in grid_field:
remainder_kernel = ()
operator_idx = 0
kernel = self.transform_kernel(field)
if self.read:
kernel = self.resolve_media_params(kernel)
kernel_args = kernel.args
for dim in range(self.dimension-1, -1, -1):
dimension = self.index[dim]
kernel_stmt_pos = kernel
kernel_stmt_neg = kernel
# For each dimension in each field iterate through its expressions to separate
# positive and negative strides
for arg in kernel_args:
if not (str(dimension) + " -" in str(arg)):
kernel_stmt_neg = kernel_stmt_neg.subs({arg: 0}, simultaneous=True)
if not (str(dimension) + " +" in str(arg)):
kernel_stmt_pos = kernel_stmt_pos.subs({arg: 0}, simultaneous=True)
remainder_kernel += kernel_stmt_pos.args
remainder_kernel += kernel_stmt_neg.args
# Create the inner loop for with negative strides expressions
if not (len(kernel_stmt_neg.args) == 0):
body_tmp = [cgen.Statement(ccode(field[idx]) + operator[operator_idx] + ccode(kernel_stmt_neg.xreplace({self.t+1: self.time[1], self.t: self.time[0]})))]
body_tmp = [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_tmp))]
if not self.pluto and self.ivdep and indexes[0] == self.dimension-1:
body_tmp.insert(0, self.compiler._ivdep)
if not self.pluto and self.simd and indexes[0] == self.dimension-1:
body_tmp.insert(0, cgen.Pragma('simd'))
body = body + body_tmp
operator_idx = 1
# Create the inner loop for with positive strides expressions
if not (len(kernel_stmt_pos.args) == 0):
body_tmp = [cgen.Statement(ccode(field[idx]) + operator[operator_idx] + ccode(kernel_stmt_pos.xreplace({self.t+1: self.time[1], self.t: self.time[0]})))]
body_tmp = [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_tmp))]
if not self.pluto and self.ivdep and indexes[0] == self.dimension-1:
body_tmp.insert(0, self.compiler._ivdep)
if not self.pluto and self.simd and indexes[0] == self.dimension-1:
body_tmp.insert(0, cgen.Pragma('simd'))
body = body + body_tmp
operator_idx = 1
# Create the inner loop for unit strided array access
kernel_stmt = kernel
for arg in remainder_kernel:
kernel_stmt = kernel_stmt.subs({arg: 0}, simultaneous=True)
body_tmp = [cgen.Statement(ccode(field[idx]) + '+=' + ccode(kernel_stmt.xreplace({self.t+1: self.time[1], self.t: self.time[0]})))]
body_tmp = [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_tmp))]
if not self.pluto and self.ivdep and indexes[0] == self.dimension-1:
body_tmp.insert(0, self.compiler._ivdep)
if not self.pluto and self.simd and indexes[0] == self.dimension-1:
body_tmp.insert(0, cgen.Pragma('simd'))
body = body + body_tmp
return body
def copy_memory(self):
#data, rowcount, colcount
vec = "_%s_vec" % ccode("m")
statements = [cgen.Assign(vec, 'data')]
return statements
