From f719152295d6066dfeeba51ead8a684e3d522e2d Mon Sep 17 00:00:00 2001
From: Dominic Kempf <dominic.kempf@iwr.uni-heidelberg.de>
Date: Thu, 24 Nov 2016 14:03:18 +0100
Subject: [PATCH] Make sumfact data structure loopy-controlled
---
python/dune/perftool/cgen/clazz.py | 2 +-
python/dune/perftool/generation/__init__.py | 1 +
python/dune/perftool/generation/loopy.py | 14 +++++++++++++
python/dune/perftool/loopy/target.py | 9 ++++++++-
python/dune/perftool/pdelab/localoperator.py | 20 ++++++++++++++++++-
python/dune/perftool/sumfact/amatrix.py | 21 ++++++++++----------
6 files changed, 54 insertions(+), 13 deletions(-)
diff --git a/python/dune/perftool/cgen/clazz.py b/python/dune/perftool/cgen/clazz.py
index f265638e..dca212ec 100644
--- a/python/dune/perftool/cgen/clazz.py
+++ b/python/dune/perftool/cgen/clazz.py
@@ -72,7 +72,7 @@ class Class(Generable):
for bc in base_classes:
assert isinstance(bc, BaseClass)
for mem in members:
- assert isinstance(mem, ClassMember)
+ assert isinstance(mem, Generable)
def generate(self):
# define the class header
diff --git a/python/dune/perftool/generation/__init__.py b/python/dune/perftool/generation/__init__.py
index 72fad902..89a98fd7 100644
--- a/python/dune/perftool/generation/__init__.py
+++ b/python/dune/perftool/generation/__init__.py
@@ -36,6 +36,7 @@ from dune.perftool.generation.loopy import (barrier,
globalarg,
iname,
instruction,
+ loopy_class_member,
kernel_cached,
noop_instruction,
silenced_warning,
diff --git a/python/dune/perftool/generation/loopy.py b/python/dune/perftool/generation/loopy.py
index c61fec6a..16c9d4f5 100644
--- a/python/dune/perftool/generation/loopy.py
+++ b/python/dune/perftool/generation/loopy.py
@@ -164,3 +164,17 @@ def barrier(**kwargs):
name = 'barrier_{}'.format(get_counter('barrier'))
_barrier(id=name, **kwargs)
return name
+
+
+def loopy_class_member(name, classtag=None, **kwargs):
+ """ A class member is based on loopy! It is an
+ * temporary variable of the constructor kernel
+ * A globalarg of the requesting kernel (to make things pass)
+ """
+ assert classtag
+ temporary_variable(name, kernel=classtag, **kwargs)
+ silenced_warning("read_no_write({})".format(name), kernel=classtag)
+
+ kwargs.pop("decl_method", None)
+ # TODO I guess some filtering has to be applied here.
+ globalarg(name, **kwargs)
\ No newline at end of file
diff --git a/python/dune/perftool/loopy/target.py b/python/dune/perftool/loopy/target.py
index 57ff1d23..f3b7d3e3 100644
--- a/python/dune/perftool/loopy/target.py
+++ b/python/dune/perftool/loopy/target.py
@@ -115,11 +115,18 @@ class DuneASTBuilder(CASTBuilder):
post_include("#define BARRIER asm volatile(\"\": : :\"memory\")", filetag="operatorfile")
return cgen.Line("BARRIER;")
+ def get_temporary_decls(self, codegen_state, schedule_index):
+ if self.target.declare_temporaries:
+ return CASTBuilder.get_temporary_decls(self, codegen_state, schedule_index)
+ else:
+ return []
+
class DuneTarget(TargetBase):
- def __init__(self):
+ def __init__(self, declare_temporaries=True):
# Set fortran_abi to allow reusing CASTBuilder for the moment
self.fortran_abi = False
+ self.declare_temporaries = declare_temporaries
def split_kernel_at_global_barriers(self):
return False
diff --git a/python/dune/perftool/pdelab/localoperator.py b/python/dune/perftool/pdelab/localoperator.py
index 0234b422..30454153 100644
--- a/python/dune/perftool/pdelab/localoperator.py
+++ b/python/dune/perftool/pdelab/localoperator.py
@@ -644,11 +644,29 @@ def cgen_class_from_cache(tag, members=[]):
# Construct the constructor
constructor_knl = extract_kernel_from_cache(tag)
+ from dune.perftool.loopy.target import DuneTarget
+ constructor_knl = constructor_knl.copy(target=DuneTarget(declare_temporaries=False))
signature = "{}({})".format(basename, ", ".join(next(iter(p.generate(with_semicolon=False))) for p in constructor_params))
constructor = LoopyKernelMethod([signature], constructor_knl, add_timings=False, initializer_list=il)
+ # Take any temporary declarations from the kernel and make them class members
+ target = DuneTarget()
+ from loopy.codegen import CodeGenerationState
+ codegen_state = CodeGenerationState(kernel=constructor_knl,
+ implemented_data_info=None,
+ implemented_domain=None,
+ implemented_predicates=frozenset(),
+ seen_dtypes=frozenset(),
+ seen_functions=frozenset(),
+ seen_atomic_dtypes=frozenset(),
+ var_subst_map={},
+ allow_complex=False,
+ is_generating_device_code=True,
+ )
+ decls = target.get_device_ast_builder().get_temporary_decls(codegen_state, 0)
+
from dune.perftool.cgen import Class
- return Class(basename, base_classes=base_classes, members=[constructor] + members + pm, tparam_decls=tparams)
+ return Class(basename, base_classes=base_classes, members=[constructor] + members + pm + decls, tparam_decls=tparams)
def generate_localoperator_kernels(formdata, data):
diff --git a/python/dune/perftool/sumfact/amatrix.py b/python/dune/perftool/sumfact/amatrix.py
index 387fb400..e919951b 100644
--- a/python/dune/perftool/sumfact/amatrix.py
+++ b/python/dune/perftool/sumfact/amatrix.py
@@ -13,6 +13,7 @@ from dune.perftool.generation import (class_member,
include_file,
initializer_list,
instruction,
+ loopy_class_member,
preamble,
silenced_warning,
temporary_variable,
@@ -93,30 +94,30 @@ def basis_functions_per_direction():
return polynomial_degree() + 1
-@class_member(classtag="operator")
def define_oned_quadrature_weights(name):
- range_field = lop_template_range_field()
- number_qp = quadrature_points_per_direction()
- return "{} {}[{}];".format(range_field, name, number_qp)
+ loopy_class_member(name,
+ dtype=numpy.float64,
+ classtag="operator",
+ shape=(quadrature_points_per_direction(),),
+ )
def name_oned_quadrature_weights():
name = "qw"
- globalarg(name, shape=(quadrature_points_per_direction(),), dtype=NumpyType(numpy.float64))
define_oned_quadrature_weights(name)
return name
-@class_member(classtag="operator")
def define_oned_quadrature_points(name):
- range_field = lop_template_range_field()
- number_qp = quadrature_points_per_direction()
- return "{} {}[{}];".format(range_field, name, number_qp)
+ loopy_class_member(name,
+ dtype=numpy.float64,
+ classtag="operator",
+ shape=(quadrature_points_per_direction(),),
+ )
def name_oned_quadrature_points():
name = "qp"
- globalarg(name, shape=(quadrature_points_per_direction(),), dtype=NumpyType(numpy.float64))
define_oned_quadrature_points(name)
return name
--
GitLab