diff --git a/python/dune/perftool/generation/loopy.py b/python/dune/perftool/generation/loopy.py
index a113cb6a203fead3a704d4ce74f2b46bb08aa16a..1345af2486c6b1e7c77a4f5317ef7f6b1823ec83 100644
--- a/python/dune/perftool/generation/loopy.py
+++ b/python/dune/perftool/generation/loopy.py
@@ -115,6 +115,8 @@ def temporary_variable(name, **kwargs):
)
def c_instruction_impl(**kw):
kw.setdefault('assignees', [])
+ from pymbolic.primitives import Variable
+ kw['assignees'] = frozenset(Variable(i) for i in kw['assignees'])
inames = kw.pop('inames', kw.get('forced_iname_deps', []))
return loopy.CInstruction(inames, **kw)
@@ -124,6 +126,9 @@ def c_instruction_impl(**kw):
cache_key_generator=lambda *a, **kw: kw['expression'],
)
def expr_instruction_impl(**kw):
+ if 'assignees' in kw:
+ from pymbolic.primitives import Variable
+ kw['assignees'] = frozenset(Variable(i) for i in kw['assignees'])
return loopy.ExpressionInstruction(**kw)
diff --git a/python/dune/perftool/loopy/target.py b/python/dune/perftool/loopy/target.py
index 71bb18d067122592ec867ddb56bd32ee3df4ef08..3c717f2de5d87a3070fbec434cfb815c9676ab93 100644
--- a/python/dune/perftool/loopy/target.py
+++ b/python/dune/perftool/loopy/target.py
@@ -1,28 +1,18 @@
-import numpy
-import six
-
-from loopy.target import TargetBase
-from loopy.target.c.codegen.expression import LoopyCCodeMapper
-
-
-from loopy.library.reduction import (ReductionOperation,
- register_reduction_parser,
- )
-
-
-class AllToDouble(dict):
- """ This imitates a dict that maps everything to double and logs the requested keys """
- def __getitem__(self, key):
- self.__setitem__(key, numpy.float64)
- return numpy.float64
+from loopy.target import (TargetBase,
+ ASTBuilderBase,
+ DummyHostASTBuilder,
+ )
+from loopy.target.c import CASTBuilder
+from loopy.target.c.codegen.expression import ExpressionToCMapper
_registry = {'float32': 'float',
'int32': 'int',
- 'float64': 'double'}
+ 'float64': 'double',
+ 'string': 'std::string'}
-class MyMapper(LoopyCCodeMapper):
+class MyMapper(ExpressionToCMapper):
def map_subscript(self, expr, enclosing_prec, type_context):
ret = str(expr.aggregate)
from pymbolic.primitives import Variable
@@ -41,50 +31,17 @@ class MyMapper(LoopyCCodeMapper):
return super(MyMapper, self).map_variable(expr, enclosing_prec, type_context)
-class DuneTarget(TargetBase):
-
- def get_or_register_dtype(self, names, dtype=None):
- return dtype
-
- def dtype_to_typename(self, dtype):
- # For now, we do this the simplest possible way
- return _registry[dtype.dtype.name]
-
- def is_vector_dtype(self, dtype):
- return False
-
+class DuneASTBuilder(CASTBuilder):
def get_expression_to_code_mapper(self, codegen_state):
return MyMapper(codegen_state)
- def generate_code(self, kernel, codegen_state, impl_arg_info):
- from cgen import Block
- body = Block()
-
- from loopy.codegen.loop import set_up_hw_parallel_loops
- gen_code = set_up_hw_parallel_loops(kernel, 0, codegen_state)
-
- from cgen import Line
- body.append(Line())
-
- if isinstance(gen_code.ast, Block):
- body.extend(gen_code.ast.contents)
- else:
- body.append(gen_code.ast)
-
- return str(body), gen_code.implemented_domains
-
- def get_value_arg_decl(self, name, shape, dtype, is_written):
- assert shape == ()
-
- return "blubb"
-
- def get_global_arg_decl(self, name, shape, dtype, is_written):
- return "blubb"
+ def get_temporary_decls(self, codegen_state, schedule_index):
+ # Currently we do *not* want to build the temporary declarations
+ # through loopy, as this would involve lots of work on a proper
+ # type system!
+ return []
def emit_sequential_loop(self, codegen_state, iname, iname_dtype, static_lbound, static_ubound, inner):
-
- from loopy.codegen import wrap_in
-
# Some of our loops are special as they use PDELab specific concepts.
# Fortunately those loops are tied to specific inames.
from dune.perftool.pdelab.quadrature import quadrature_iname
@@ -92,32 +49,28 @@ class DuneTarget(TargetBase):
from cgen import CustomLoop
from dune.perftool.pdelab.quadrature import quadrature_loop_statement
loop_stmt = quadrature_loop_statement()
- return wrap_in(CustomLoop,
- "for({})".format(loop_stmt),
- inner)
-
- # From here on it is the default implementation taken from loopys CTarget.
- ecm = codegen_state.expression_to_code_mapper
-
- from loopy.symbolic import aff_to_expr
-
- from pymbolic.mapper.stringifier import PREC_NONE
- from cgen import For
-
- return wrap_in(For,
- "{} {} = {}".format(self.dtype_to_typename(iname_dtype),
- iname,
- ecm(aff_to_expr(static_lbound),
- PREC_NONE,
- "i",
- ),
- ),
- "{} <= {}".format(iname,
- ecm(aff_to_expr(static_ubound),
- PREC_NONE,
- "i",
- ),
- ),
- "++{}".format(iname),
- inner,
- )
+ return CustomLoop("for({})".format(loop_stmt), inner)
+ else:
+ return CASTBuilder.emit_sequential_loop(self, codegen_state, iname, iname_dtype, static_lbound, static_ubound, inner)
+
+
+class DuneTarget(TargetBase):
+ def __init__(self):
+ # Set fortran_abi to allow reusing CASTBuilder for the moment
+ self.fortran_abi = False
+
+ def split_kernel_at_global_barriers(self):
+ return False
+
+ def get_host_ast_builder(self):
+ return DummyHostASTBuilder(self)
+
+ def get_device_ast_builder(self):
+ return DuneASTBuilder(self)
+
+ def dtype_to_typename(self, dtype):
+ # For now, we do this the simplest possible way
+ return _registry[dtype.dtype.name]
+
+ def is_vector_dtype(self, dtype):
+ return False
diff --git a/python/dune/perftool/pdelab/basis.py b/python/dune/perftool/pdelab/basis.py
index db50a9053d24bf83417630589894524b2d998685..0031f468087c22de132ba186404188299e26a24f 100644
--- a/python/dune/perftool/pdelab/basis.py
+++ b/python/dune/perftool/pdelab/basis.py
@@ -290,7 +290,7 @@ def evaluate_basis_gradient(leaf_element, name, restriction):
name,
index,
),
- assignees=name,
+ assignees=frozenset({name}),
read_variables=frozenset({reference_gradients}),
)
diff --git a/python/dune/perftool/pdelab/localoperator.py b/python/dune/perftool/pdelab/localoperator.py
index 596b1414fb67582362920acf54b2ac3a61f98af2..958e0974ddf24d2b3b7c8889d6ccfabd5207dd94 100644
--- a/python/dune/perftool/pdelab/localoperator.py
+++ b/python/dune/perftool/pdelab/localoperator.py
@@ -16,7 +16,6 @@ from dune.perftool.cgen.clazz import (AccessModifier,
ClassMember,
)
from dune.perftool import Restriction
-from pytools import memoize
@template_parameter("operator")
@@ -218,12 +217,14 @@ def generate_kernel(integrals):
class AssemblyMethod(ClassMember):
def __init__(self, signature, kernel):
- from loopy import generate_code
+ from loopy import generate_body
from cgen import LiteralLines, Block
content = signature
content.append('{')
if kernel is not None:
- content.extend(' ' + l for l in generate_code(kernel)[0].split('\n'))
+ for i, p in kernel.preambles:
+ content.append(p)
+ content.extend(l for l in generate_body(kernel).split('\n')[1:-1])
content.append('}')
ClassMember.__init__(self, content)
diff --git a/python/dune/perftool/pdelab/parameter.py b/python/dune/perftool/pdelab/parameter.py
index 5ba55552f880f736ef6d8647451f3c04b07cd239..c4e3c5488015445dc954258525370cece2c82d65 100644
--- a/python/dune/perftool/pdelab/parameter.py
+++ b/python/dune/perftool/pdelab/parameter.py
@@ -107,7 +107,7 @@ def construct_nested_fieldvector(t, shape):
@cached
def cell_parameter_function(name, expr, restriction, t='double'):
shape = expr.ufl_element().value_shape()
- shape_impl = ('fv',)*len(shape)
+ shape_impl = ('fv',) * len(shape)
t = construct_nested_fieldvector(t, shape)
temporary_variable(name, shape=shape, shape_impl=shape_impl)
define_parameter_function_class_member(name, expr, t, True)
@@ -117,7 +117,7 @@ def cell_parameter_function(name, expr, restriction, t='double'):
@cached
def intersection_parameter_function(name, expr, t='double'):
shape = expr.ufl_element().value_shape()
- shape_impl = ('fv',)*len(shape)
+ shape_impl = ('fv',) * len(shape)
t = construct_nested_fieldvector(t, shape)
temporary_variable(name, shape=shape, shape_impl=shape_impl)
define_parameter_function_class_member(name, expr, t, False)
diff --git a/python/loopy b/python/loopy
index 6b32fb790fb7c4947da03f6c8f1a3694ee90da92..db5dd7408b03e4cb453120cee3476f44b4179dfc 160000
--- a/python/loopy
+++ b/python/loopy
@@ -1 +1 @@
-Subproject commit 6b32fb790fb7c4947da03f6c8f1a3694ee90da92
+Subproject commit db5dd7408b03e4cb453120cee3476f44b4179dfc