def visit_Return(self, node):
self.visit(node.value)
node.ccode = c.Statement('return %s' % node.value.ccode)
def generate(self, funcname, field_args, const_args, kernel_ast,
c_include):
ccode = []
pname = self.ptype.name + 'p'
# ==== Add include for Parcels and math header ==== #
ccode += [str(c.Include("parcels.h", system=False))]
ccode += [str(c.Include("math.h", system=False))]
ccode += [str(c.Assign('double _next_dt', '0'))]
ccode += [str(c.Assign('size_t _next_dt_set', '0'))]
ccode += [
str(
c.Assign(
'const int ngrid',
str(self.fieldset.gridset.size if self.
fieldset is not None else 1)))
]
# ==== Generate type definition for particle type ==== #
vdeclp = [
c.Pointer(c.POD(v.dtype, v.name)) for v in self.ptype.variables
]
ccode += [
str(c.Typedef(c.GenerableStruct("", vdeclp, declname=pname)))
]
# Generate type definition for single particle type
vdecl = [
c.POD(v.dtype, v.name) for v in self.ptype.variables
if v.dtype != np.uint64
]
ccode += [
str(
c.Typedef(
c.GenerableStruct("", vdecl, declname=self.ptype.name)))
]
args = [
c.Pointer(c.Value(self.ptype.name, "particle_backup")),
c.Pointer(c.Value(pname, "particles")),
c.Value("int", "pnum")
]
p_back_set_decl = c.FunctionDeclaration(
c.Static(
c.DeclSpecifier(c.Value("void", "set_particle_backup"),
spec='inline')), args)
body = []
for v in self.ptype.variables:
if v.dtype != np.uint64 and v.name not in ['dt', 'state']:
body += [
c.Assign(("particle_backup->%s" % v.name),
("particles->%s[pnum]" % v.name))
]
p_back_set_body = c.Block(body)
p_back_set = str(c.FunctionBody(p_back_set_decl, p_back_set_body))
ccode += [p_back_set]
args = [
c.Pointer(c.Value(self.ptype.name, "particle_backup")),
c.Pointer(c.Value(pname, "particles")),
c.Value("int", "pnum")
]
p_back_get_decl = c.FunctionDeclaration(
c.Static(
c.DeclSpecifier(c.Value("void", "get_particle_backup"),
spec='inline')), args)
body = []
for v in self.ptype.variables:
if v.dtype != np.uint64 and v.name not in ['dt', 'state']:
body += [
c.Assign(("particles->%s[pnum]" % v.name),
("particle_backup->%s" % v.name))
]
p_back_get_body = c.Block(body)
p_back_get = str(c.FunctionBody(p_back_get_decl, p_back_get_body))
ccode += [p_back_get]
update_next_dt_decl = c.FunctionDeclaration(
c.Static(
c.DeclSpecifier(c.Value("void", "update_next_dt"),
spec='inline')), [c.Value('double', 'dt')])
if 'update_next_dt' in str(kernel_ast):
body = []
body += [c.Assign("_next_dt", "dt")]
body += [c.Assign("_next_dt_set", "1")]
update_next_dt_body = c.Block(body)
update_next_dt = str(
c.FunctionBody(update_next_dt_decl, update_next_dt_body))
ccode += [update_next_dt]
if c_include:
ccode += [c_include]
# ==== Insert kernel code ==== #
ccode += [str(kernel_ast)]
# Generate outer loop for repeated kernel invocation
args = [
c.Value("int", "num_particles"),
c.Pointer(c.Value(pname, "particles")),
c.Value("double", "endtime"),
c.Value("double", "dt")
]
for field, _ in field_args.items():
args += [c.Pointer(c.Value("CField", "%s" % field))]
for const, _ in const_args.items():
args += [c.Value("double", const)]
fargs_str = ", ".join(['particles->time[pnum]'] +
list(field_args.keys()) +
list(const_args.keys()))
# ==== statement clusters use to compose 'body' variable and variables 'time_loop' and 'part_loop' ==== ##
sign_dt = c.Assign("sign_dt", "dt > 0 ? 1 : -1")
particle_backup = c.Statement("%s particle_backup" % self.ptype.name)
sign_end_part = c.Assign(
"sign_end_part", "(endtime - particles->time[pnum]) > 0 ? 1 : -1")
reset_res_state = c.Assign("res", "particles->state[pnum]")
update_state = c.Assign("particles->state[pnum]", "res")
update_pdt = c.If(
"_next_dt_set == 1",
c.Block([
c.Assign("_next_dt_set", "0"),
c.Assign("particles->dt[pnum]", "_next_dt")
]))
dt_pos = c.Assign(
"__dt",
"fmin(fabs(particles->dt[pnum]), fabs(endtime - particles->time[pnum]))"
) # original
pdt_eq_dt_pos = c.Assign("__pdt_prekernels", "__dt * sign_dt")
partdt = c.Assign("particles->dt[pnum]", "__pdt_prekernels")
check_pdt = c.If(
"(res == SUCCESS) & !is_equal_dbl(__pdt_prekernels, particles->dt[pnum])",
c.Assign("res", "REPEAT"))
dt_0_break = c.If("is_zero_dbl(particles->dt[pnum])",
c.Statement("break"))
notstarted_continue = c.If(
"(( sign_end_part != sign_dt) || is_close_dbl(__dt, 0) ) && !is_zero_dbl(particles->dt[pnum])",
c.Block([
c.If("fabs(particles->time[pnum]) >= fabs(endtime)",
c.Assign("particles->state[pnum]", "SUCCESS")),
c.Statement("continue")
]))
# ==== main computation body ==== #
body = [
c.Statement(
"set_particle_backup(&particle_backup, particles, pnum)")
]
body += [pdt_eq_dt_pos]
body += [partdt]
body += [
c.Value("StatusCode", "state_prev"),
c.Assign("state_prev", "particles->state[pnum]")
]
body += [
c.Assign("res", "%s(particles, pnum, %s)" % (funcname, fargs_str))
]
body += [
c.If("(res==SUCCESS) && (particles->state[pnum] != state_prev)",
c.Assign("res", "particles->state[pnum]"))
]
body += [check_pdt]
body += [
c.If(
"res == SUCCESS || res == DELETE",
c.Block([
c.Statement(
"particles->time[pnum] += particles->dt[pnum]"),
update_pdt, dt_pos, sign_end_part,
c.If(
"(res != DELETE) && !is_close_dbl(__dt, 0) && (sign_dt == sign_end_part)",
c.Assign("res", "EVALUATE")),
c.If("sign_dt != sign_end_part",
c.Assign("__dt", "0")), update_state, dt_0_break
]),
c.Block([
c.Statement(
"get_particle_backup(&particle_backup, particles, pnum)"
), dt_pos, sign_end_part,
c.If("sign_dt != sign_end_part", c.Assign("__dt", "0")),
update_state,
c.Statement("break")
]))
]
time_loop = c.While(
"(particles->state[pnum] == EVALUATE || particles->state[pnum] == REPEAT) || is_zero_dbl(particles->dt[pnum])",
c.Block(body))
part_loop = c.For(
"pnum = 0", "pnum < num_particles", "++pnum",
c.Block([
sign_end_part, reset_res_state, dt_pos, notstarted_continue,
time_loop
]))
fbody = c.Block([
c.Value("int", "pnum, sign_dt, sign_end_part"),
c.Value("StatusCode", "res"),
c.Value("double", "__pdt_prekernels"),
c.Value("double",
"__dt"), # 1e-8 = built-in tolerance for np.isclose()
sign_dt,
particle_backup,
part_loop
])
fdecl = c.FunctionDeclaration(c.Value("void", "particle_loop"), args)
ccode += [str(c.FunctionBody(fdecl, fbody))]
return "\n\n".join(ccode)
def __init__(self, header=None, body=None, footer=None):
b = [Element(c.Comment('Flush denormal numbers to zero in hardware')),
Element(c.Statement('_MM_SET_DENORMALS_ZERO_MODE(_MM_DENORMALS_ZERO_ON)')),
Element(c.Statement('_MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_ON)'))]
super(Denormals, self).__init__(header, b, footer)
def ccode(self):
return c.Statement('%s(%s)' % (self.name, ','.join(self.params)))
def visit_Call(self, o):
arguments = self._args_call(o.params)
return c.Statement('%s(%s)' % (o.name, ','.join(arguments)))
def visit_ForeignExpression(self, o):
return c.Statement(ccode(o.expr))
def visit_Expr(self, node):
self.visit(node.value)
node.ccode = c.Statement(node.value.ccode)
def visit_Break(self, node):
node.ccode = c.Statement("break")
"""Generate C++ code from an expression calling CodePrinter class
:param expr: The expression
:param settings: A dictionary of settings for code printing
:returns: The resulting code as a string. If it fails, then it returns the expr
"""
if isinstance(expr, Eq):
return ccode_eq(expr)
try:
return CodePrinter(settings).doprint(expr, None)
except:
return expr
def ccode_eq(eq, **settings):
"""Generate C++ assignment from an equation assigning RHS to LHS
:param eq: The equation
:param settings: A dictionary of settings for code printing
:returns: The resulting code as a string
"""
return CodePrinter(settings).doprint(eq.lhs, None) \
+ ' = ' + CodePrinter(settings).doprint(eq.rhs, None)
blankline = c.Line("")
printmark = lambda i: c.Line('printf("Here: %s\\n"); fflush(stdout);' % i)
printvar = lambda i: c.Statement('printf("%s=%%s\\n", %s); fflush(stdout);' % (i, i))
INT = Function('INT')
FLOAT = Function('FLOAT')
def map_Continue(self, node):
return cgen.Statement("label_%s:" % node.label)
def map_Return(self, node):
return cgen.Statement("return")
def map_Goto(self, node):
return cgen.Statement("goto label_%s" % node.label)
omplang = {
'for': c.Pragma('omp for schedule(static)'),
'collapse': lambda i: c.Pragma('omp for collapse(%d) schedule(static)' % i),
'par-region': lambda i: c.Pragma('omp parallel %s' % i),
'par-for': c.Pragma('omp parallel for schedule(static)'),
'simd-for': c.Pragma('omp simd'),
'simd-for-aligned': lambda i, j: c.Pragma('omp simd aligned(%s:%d)' % (i, j))
}
"""
Compiler-specific language
"""
complang_ALL = {
'IntelCompiler': {'ignore-deps': c.Pragma('ivdep'),
'ntstores': c.Pragma('vector nontemporal'),
'storefence': c.Statement('_mm_sfence()'),
'noinline': c.Pragma('noinline')}
}
complang_ALL['IntelKNLCompiler'] = complang_ALL['IntelCompiler']
"""
SIMD generic info
"""
simdinfo = {
# Sizes in bytes of a vector register
'sse': 16, 'see4_2': 16,
'avx': 32, 'avx2': 32,
'avx512f': 64
}