diff --git a/python/dune/codegen/sumfact/accumulation.py b/python/dune/codegen/sumfact/accumulation.py
index f327fceb0ba5e31d14dadaa8d977e7147c6655ed..ae4a763063f0dc25303fc517356b58b02551eebc 100644
--- a/python/dune/codegen/sumfact/accumulation.py
+++ b/python/dune/codegen/sumfact/accumulation.py
@@ -44,7 +44,7 @@ from dune.codegen.sumfact.switch import (get_facedir,
)
from dune.codegen.sumfact.symbolic import SumfactKernel, SumfactKernelInterfaceBase
from dune.codegen.ufl.modified_terminals import extract_modified_arguments
-from dune.codegen.tools import get_pymbolic_basename, get_leaf
+from dune.codegen.tools import get_pymbolic_basename, get_leaf, ImmutableCuttingRecord
from dune.codegen.error import CodegenError
from pytools import ImmutableRecord, product
@@ -90,7 +90,7 @@ def accum_iname(element, bound, i):
return sumfact_iname(bound, "accum{}".format(suffix))
-class AccumulationOutput(SumfactKernelInterfaceBase, ImmutableRecord):
+class AccumulationOutput(SumfactKernelInterfaceBase, ImmutableCuttingRecord):
def __init__(self,
matrix_sequence=None,
accumvar=None,
@@ -112,26 +112,20 @@ class AccumulationOutput(SumfactKernelInterfaceBase, ImmutableRecord):
# recalculating it in the property.
dim = world_dimension()
quadrature_permutation = sumfact_quadrature_permutation_strategy(dim, restriction[0])
-
- # Calculate cost optimal permutation
matrix_sequence = permute_forward(matrix_sequence, quadrature_permutation)
- cost_permutation = sumfact_cost_permutation_strategy(matrix_sequence, self.stage)
# TODO: Isnt accumvar superfluous in the presence of all the other infos?
# Note: Do not put matrix_sequence into the Record. That screws up the vectorization strategy!
- ImmutableRecord.__init__(self,
- accumvar=accumvar,
- restriction=restriction,
- test_element=test_element,
- test_element_index=test_element_index,
- trial_element=trial_element,
- trial_element_index=trial_element_index,
- _quadrature_permutation=quadrature_permutation,
- _cost_permutation=cost_permutation,
- )
-
- def __repr__(self):
- return ImmutableRecord.__repr__(self)
+ ImmutableCuttingRecord.__init__(self,
+ accumvar=accumvar,
+ restriction=restriction,
+ test_element=test_element,
+ test_element_index=test_element_index,
+ trial_element=trial_element,
+ trial_element_index=trial_element_index,
+ _quadrature_permutation=quadrature_permutation,
+ _permuted_matrix_sequence=matrix_sequence,
+ )
def get_keyword_arguments(self):
"""Get dictionary of keyword arguments needed to initialize this class
@@ -141,7 +135,7 @@ class AccumulationOutput(SumfactKernelInterfaceBase, ImmutableRecord):
this dict to create an interface.
"""
dict = self.get_copy_kwargs()
- del dict['_cost_permutation']
+ del dict['_permuted_matrix_sequence']
del dict['_quadrature_permutation']
dict['matrix_sequence'] = None
return dict
@@ -152,7 +146,7 @@ class AccumulationOutput(SumfactKernelInterfaceBase, ImmutableRecord):
@property
def cost_permutation(self):
- return self._cost_permutation
+ return sumfact_cost_permutation_strategy(self._permuted_matrix_sequence, self.stage)
@property
def stage(self):
diff --git a/python/dune/codegen/sumfact/autotune.py b/python/dune/codegen/sumfact/autotune.py
index 4b9a1d132214652a9258cb59b46364117aeece1f..68d81957e00917d55383282c748879f7f96db587 100644
--- a/python/dune/codegen/sumfact/autotune.py
+++ b/python/dune/codegen/sumfact/autotune.py
@@ -13,6 +13,7 @@ import os
import re
import subprocess
import filelock
+import hashlib
def get_cmake_cache_entry(entry):
@@ -192,10 +193,12 @@ def autotune_realization(sf):
os.mkdir(dir)
basename = "autotune_sumfact_{}".format(sf.function_name)
- name = os.path.join(dir, "autotune_sumfact_{}".format(sf.function_name))
+ basename = hashlib.sha256(basename.encode()).hexdigest()
+
filename = os.path.join(dir, "{}.cc".format(basename))
logname = os.path.join(dir, "{}.log".format(basename))
- lock = "{}.lock".format(name)
+ lock = os.path.join(dir, "{}.lock".format(basename))
+ executable = os.path.join(dir, basename)
# Generate and compile a benchmark program
with cache_restoring():
@@ -203,20 +206,20 @@ def autotune_realization(sf):
if not os.path.isfile(logname):
generate_standalone_code(sf, filename)
- ret = subprocess.call(compiler_invocation(name, filename))
+ devnull = open(os.devnull, 'w')
+ ret = subprocess.call(compiler_invocation(executable, filename), stdout=devnull, stderr=subprocess.STDOUT)
if ret != 0:
raise CodegenAutotuneError("Compilation of autotune executable failed. Invocation: {}".format(" ".join(compiler_invocation(name, filename))))
# Check whether the user specified an execution wrapper
call = []
- wrapper = get_cmake_cache_entry("DUNE_PERFTOOL_BENCHMARK_WRAPPER")
+ wrapper = get_cmake_cache_entry("DUNE_CODEGEN_BENCHMARK_WRAPPER")
if wrapper:
call.append(wrapper)
# Run the benchmark program
- call.append(name)
+ call.append(executable)
call.append(logname)
- devnull = open(os.devnull, 'w')
ret = subprocess.call(call, stdout=devnull, stderr=subprocess.STDOUT)
if ret != 0:
raise CodegenAutotuneError("Execution of autotune benchmark failed. Invocation: {}".format(" ".join(call)))
diff --git a/python/dune/codegen/sumfact/basis.py b/python/dune/codegen/sumfact/basis.py
index 9fac39426df4339a5804237418e5edfcd6e092c6..c9b75eb445af01e9420e850650b12054dd0115f3 100644
--- a/python/dune/codegen/sumfact/basis.py
+++ b/python/dune/codegen/sumfact/basis.py
@@ -44,13 +44,13 @@ from dune.codegen.options import get_form_option
from dune.codegen.pdelab.driver import FEM_name_mangling
from dune.codegen.pdelab.restriction import restricted_name
from dune.codegen.pdelab.spaces import name_lfs, name_lfs_bound, name_leaf_lfs
-from dune.codegen.tools import maybe_wrap_subscript
+from dune.codegen.tools import maybe_wrap_subscript, ImmutableCuttingRecord
from dune.codegen.pdelab.basis import shape_as_pymbolic
from dune.codegen.sumfact.accumulation import sumfact_iname
from ufl import MixedElement, VectorElement, TensorElement, TensorProductElement
-from pytools import product, ImmutableRecord
+from pytools import product
from loopy.match import Writes
@@ -235,7 +235,7 @@ class SumfactBasisMixin(GenericBasisMixin):
return prim.Subscript(var, vsf.quadrature_index(sf, self))
-class LFSSumfactKernelInput(SumfactKernelInterfaceBase, ImmutableRecord):
+class LFSSumfactKernelInput(SumfactKernelInterfaceBase, ImmutableCuttingRecord):
def __init__(self,
matrix_sequence=None,
coeff_func=None,
@@ -252,25 +252,17 @@ class LFSSumfactKernelInput(SumfactKernelInterfaceBase, ImmutableRecord):
# recalculating it in the property.
dim = world_dimension()
quadrature_permutation = sumfact_quadrature_permutation_strategy(dim, restriction)
-
matrix_sequence = permute_forward(matrix_sequence, quadrature_permutation)
- cost_permutation = sumfact_cost_permutation_strategy(matrix_sequence, self.stage)
# Note: Do not put matrix_sequence into the Record. That screws up the vectorization strategy!
- ImmutableRecord.__init__(self,
- coeff_func=coeff_func,
- element=element,
- element_index=element_index,
- restriction=restriction,
- _quadrature_permutation=quadrature_permutation,
- _cost_permutation=cost_permutation,
- )
-
- def __repr__(self):
- return ImmutableRecord.__repr__(self)
-
- def __str__(self):
- return repr(self)
+ ImmutableCuttingRecord.__init__(self,
+ coeff_func=coeff_func,
+ element=element,
+ element_index=element_index,
+ restriction=restriction,
+ _quadrature_permutation=quadrature_permutation,
+ _permuted_matrix_sequence=matrix_sequence,
+ )
def get_keyword_arguments(self):
"""Get dictionary of keyword arguments needed to initialize this class
@@ -280,7 +272,7 @@ class LFSSumfactKernelInput(SumfactKernelInterfaceBase, ImmutableRecord):
this dict to create an interface.
"""
dict = self.get_copy_kwargs()
- del dict['_cost_permutation']
+ del dict['_permuted_matrix_sequence']
del dict['_quadrature_permutation']
dict['matrix_sequence'] = None
return dict
@@ -291,7 +283,7 @@ class LFSSumfactKernelInput(SumfactKernelInterfaceBase, ImmutableRecord):
@property
def cost_permutation(self):
- return self._cost_permutation
+ return sumfact_cost_permutation_strategy(self._permuted_matrix_sequence, self.stage)
@property
def stage(self):
diff --git a/python/dune/codegen/sumfact/geometry.py b/python/dune/codegen/sumfact/geometry.py
index 08e9d43e8fe685b3b24b60bd10be418b5f078ad7..d37558f5c40f42355b5fa99ed9a4ea25ab7ade28 100644
--- a/python/dune/codegen/sumfact/geometry.py
+++ b/python/dune/codegen/sumfact/geometry.py
@@ -41,12 +41,10 @@ from dune.codegen.sumfact.permutation import (permute_backward,
from dune.codegen.sumfact.quadrature import additional_inames
from dune.codegen.sumfact.switch import get_facedir, get_facemod
from dune.codegen.sumfact.symbolic import SumfactKernelInterfaceBase, SumfactKernel
-from dune.codegen.tools import get_pymbolic_basename
+from dune.codegen.tools import get_pymbolic_basename, ImmutableCuttingRecord
from dune.codegen.options import get_form_option, option_switch
from dune.codegen.ufl.modified_terminals import Restriction
-from pytools import ImmutableRecord
-
from loopy.match import Writes
import pymbolic.primitives as prim
@@ -337,7 +335,7 @@ def global_corner_iname(restriction):
return name
-class GeoCornersInput(SumfactKernelInterfaceBase, ImmutableRecord):
+class GeoCornersInput(SumfactKernelInterfaceBase, ImmutableCuttingRecord):
def __init__(self,
matrix_sequence=None,
direction=None,
@@ -361,23 +359,15 @@ class GeoCornersInput(SumfactKernelInterfaceBase, ImmutableRecord):
# recalculating it in the property.
dim = world_dimension()
quadrature_permutation = sumfact_quadrature_permutation_strategy(dim, restriction)
-
matrix_sequence = permute_forward(matrix_sequence, quadrature_permutation)
- cost_permutation = sumfact_cost_permutation_strategy(matrix_sequence, self.stage)
# Note: Do not put matrix_sequence into the Record. That screws up the vectorization strategy!
- ImmutableRecord.__init__(self,
- direction=direction,
- restriction=restriction,
- _quadrature_permutation=quadrature_permutation,
- _cost_permutation=cost_permutation,
- )
-
- def __repr__(self):
- return ImmutableRecord.__repr__(self)
-
- def __str__(self):
- return repr(self)
+ ImmutableCuttingRecord.__init__(self,
+ direction=direction,
+ restriction=restriction,
+ _quadrature_permutation=quadrature_permutation,
+ _permuted_matrix_sequence=matrix_sequence,
+ )
def get_keyword_arguments(self):
"""Get dictionary of keyword arguments needed to initialize this class
@@ -387,7 +377,7 @@ class GeoCornersInput(SumfactKernelInterfaceBase, ImmutableRecord):
this dict to create an interface.
"""
dict = self.get_copy_kwargs()
- del dict['_cost_permutation']
+ del dict['_permuted_matrix_sequence']
del dict['_quadrature_permutation']
dict['matrix_sequence'] = None
return dict
@@ -398,7 +388,7 @@ class GeoCornersInput(SumfactKernelInterfaceBase, ImmutableRecord):
@property
def cost_permutation(self):
- return self._cost_permutation
+ return sumfact_cost_permutation_strategy(self._permuted_matrix_sequence, self.stage)
@property
def stage(self):
diff --git a/python/dune/codegen/sumfact/symbolic.py b/python/dune/codegen/sumfact/symbolic.py
index 2768ece8993b01565a2eeacb4311fa976174c236..60641b0ce4af4ddef525a4c03a438c4ebe277c0c 100644
--- a/python/dune/codegen/sumfact/symbolic.py
+++ b/python/dune/codegen/sumfact/symbolic.py
@@ -214,9 +214,6 @@ class SumfactKernelInterfaceBase(object):
def function_name_suffix(self):
return ""
- def __repr__(self):
- return "SumfactKernelInterfaceBase()"
-
class VectorSumfactKernelInput(SumfactKernelInterfaceBase):
def __init__(self, interfaces):
@@ -240,8 +237,6 @@ class VectorSumfactKernelInput(SumfactKernelInterfaceBase):
# permutation. For both structured and unstructured grids the order of
# the global directions should be the same leading to the same cost
# permutation for all those sum factorization kernels.
- for i in self.interfaces:
- assert i.cost_permutation == self.interfaces[0].cost_permutation
return self.interfaces[0].cost_permutation
@property
@@ -608,7 +603,7 @@ class SumfactKernel(SumfactKernelBase, ImmutableRecord, prim.Variable):
# TODO: For now we do not vectorize SumfactKernels with different
# quadrature_permutation. This should be handled like upper/lower
# vectorization
- return tuple(m.quadrature_size for m in self.matrix_sequence_quadrature_permuted) + tuple(m.basis_size for m in self.matrix_sequence_quadrature_permuted) + (self.stage, self.buffer, self.interface.within_inames) + (self.interface.direct_is_possible, self.interface.quadrature_permutation)
+ return tuple(m.basis_size for m in self.matrix_sequence_quadrature_permuted) + (self.stage, self.buffer, self.interface.within_inames) + (self.interface.direct_is_possible, self.interface.quadrature_permutation)
@property
def cache_key(self):
@@ -847,20 +842,9 @@ class VectorizedSumfactKernel(SumfactKernelBase, ImmutableRecord, prim.Variable)
# For now we don't mix direct and non_direct input. Could be done in an upper/lower way.
assert len(set(tuple(k.interface.direct_is_possible for k in kernels))) == 1
- # Assert properties of the matrix sequence of the underlying kernels
- for i in range(kernels[0].length):
- assert len(set(tuple(k.matrix_sequence_quadrature_permuted[i].rows for k in kernels))) == 1
- assert len(set(tuple(k.matrix_sequence_quadrature_permuted[i].cols for k in kernels))) == 1
- assert len(set(tuple(k.matrix_sequence_quadrature_permuted[i].direction for k in kernels))) == 1
- assert len(set(tuple(k.matrix_sequence_quadrature_permuted[i].transpose for k in kernels))) == 1
-
# Join the instruction dependencies of all subkernels
insn_dep = insn_dep.union(k.insn_dep for k in kernels)
- # Assert that the cost_permutation is the same for all kernels
- for k in kernels:
- assert k.interface.cost_permutation == kernels[0].interface.cost_permutation
-
# We currently assume that all subkernels are consecutive, 0-based within the vector
assert None not in kernels
@@ -1013,16 +997,7 @@ class VectorizedSumfactKernel(SumfactKernelBase, ImmutableRecord, prim.Variable)
def horizontal_index(self, sf):
for i, k in enumerate(self.kernels):
- # We need to identify to which part of the vectorized kernel sf
- # corresponds. Since splitting might change the cost_permutation we
- # exclude it in the comparison below. We also make sure to check
- # that derivatives are the same.
- from copy import deepcopy
- sf_interface = deepcopy(sf.interface)
- sf_interface._cost_permutation = None
- k_interface = deepcopy(k.interface)
- k_interface._cost_permutation = None
- if repr(sf_interface) == repr(k_interface):
+ if sf.interface == k.interface:
if tuple(mat.derivative for mat in sf.matrix_sequence_quadrature_permuted) == tuple(mat.derivative for mat in k.matrix_sequence_quadrature_permuted):
return i
diff --git a/python/dune/codegen/sumfact/tabulation.py b/python/dune/codegen/sumfact/tabulation.py
index 7540cc89790392c9ce55cd17b69fcd96370d2de5..9def97eb3ba4fdae280cc65e7f12ca73164d1146 100644
--- a/python/dune/codegen/sumfact/tabulation.py
+++ b/python/dune/codegen/sumfact/tabulation.py
@@ -145,11 +145,8 @@ class BasisTabulationMatrixArray(BasisTabulationMatrixBase):
assert isinstance(tabs, tuple)
# Assert that all the basis tabulations match in size!
- assert len(set(t.quadrature_size for t in tabs)) == 1
assert len(set(t.basis_size for t in tabs)) == 1
assert len(set(t.transpose for t in tabs)) == 1
- assert len(set(t.direction for t in tabs)) == 1
- assert len(set(t.slice_size for t in tabs)) == 1
self.tabs = tabs
if width is None:
diff --git a/python/dune/codegen/sumfact/vectorization.py b/python/dune/codegen/sumfact/vectorization.py
index 5788cbe84693255d81c15c2745344f3eaea35450..e753652b10b3ac9b1765ee071bce1ccebd15f6b1 100644
--- a/python/dune/codegen/sumfact/vectorization.py
+++ b/python/dune/codegen/sumfact/vectorization.py
@@ -367,8 +367,6 @@ def level1_optimal_vectorization_strategy(sumfacts, width):
def level2_optimal_vectorization_strategy(sumfacts, width, qp):
- set_quadrature_points(qp)
-
# Find the sets of simultaneously realizable kernels
keys = frozenset(sf.parallel_key for sf in sumfacts)
diff --git a/python/dune/codegen/tools.py b/python/dune/codegen/tools.py
index d5c0a18ebe8f7e32316aa459c959ab1eb9ba75f8..0726adc8ff612825e1270f41ee95f2943e04738f 100644
--- a/python/dune/codegen/tools.py
+++ b/python/dune/codegen/tools.py
@@ -4,6 +4,22 @@ from __future__ import absolute_import
import loopy as lp
import pymbolic.primitives as prim
import frozendict
+import pytools
+
+
+class ImmutableCuttingRecord(pytools.ImmutableRecord):
+ """
+ A record implementation that drops fields starting with an underscore
+ from hash and equality computation
+ """
+ def __repr__(self):
+ return "{}({})".format(type(self), ",".join(repr(getattr(self, f)) for f in self.__class__.fields if not f.startswith("_")))
+
+ def __hash__(self):
+ return hash((type(self),) + tuple(getattr(self, field) for field in self.__class__.fields if not field.startswith("_")))
+
+ def __eq__(self, other):
+ return type(self) == type(other) and all(getattr(self, field) == getattr(other, field) for field in self.__class__.fields if not field.startswith("_"))
def get_pymbolic_basename(expr):