matrix-free FEEvaluation: initialize with nan #14152

tjhei · 2022-07-19T23:05:11Z

Initialize the internal scratch_data of FEEvaluation with signaling_nans
in debug mode.

tjhei · 2022-07-19T23:06:41Z

This is for trying to fix a bug I am seeing in ASPECT and @bangerth gave me the idea to initialize the internal data structures with NaNs to be able to more reliably see when things go wrong.

FYI @kronbichler

tjhei · 2022-07-20T00:59:16Z

And the following tests fail with a FPE:

The following tests FAILED:

	5711 - matrix_free/matrix_vector_hessians_cells.debug (Failed)

	5753 - matrix_free/matrix_vector_15.debug (Failed)

	5804 - matrix_free/matrix_vector_faces_02.debug (Failed)

	5878 - matrix_free/matrix_vector_hessians_faces.debug (Failed)

😃

bangerth

I'm glad this was useful. Let's hope that someone can see what the problem is -- that's exactly why we invented the signaling_nan() infrastructure.

bangerth · 2022-07-20T02:17:17Z

include/deal.II/matrix_free/fe_evaluation_data.h

+#  ifdef DEBUG
+  scratch_data_array->clear();
+  scratch_data_array->resize(allocated_size,
+                             Number(numbers::signaling_nan<double>()));


Should this be the following instead?

Suggested change

Number(numbers::signaling_nan<double>()));

numbers::signaling_nan<Number>());

Won't work because this will be instantiated for Number=VectorizedArray and .
I was too lazy to extend signaling_nan.

Ah, but then what are the semantics of casting a signaling nan double to a VectorizedArray?

Leave it as is for the moment -- the patch isn't going to get merged for now while there are failures, but if you don't mind, add the specializations of signaling_nan for VectorizedArray objects.

It is calling the constructor that puts this value into all values:

dealii/include/deal.II/base/vectorization.h

Line 440 in cb83302

VectorizedArray(const Number scalar)

But you should be able to use

Suggested change

Number(numbers::signaling_nan<double>()));

Number(numbers::signaling_nan<ScalarNumber>()));

which is defined here:

dealii/include/deal.II/matrix_free/fe_evaluation_data.h

Lines 125 to 126 in b8f23f1

using ScalarNumber =

typename internal::VectorizedArrayTrait<Number>::value_type;

kronbichler · 2022-07-20T11:24:38Z

This approach currently breaks the use case in

dealii/include/deal.II/matrix_free/mapping_info.templates.h

Lines 2106 to 2108 in b8f23f1

    
           AlignedVector<VectorizedDouble> evaluation_data; 
        
           eval_int.set_data_pointers(&evaluation_data, dim); 
        
           eval_ext.set_data_pointers(&evaluation_data, dim);

where we attach two different evaluator objects to the same storage array, so the second one frees the first item's memory. Let me try to fix it locally and see what happens.

tjhei · 2022-07-20T14:43:05Z

Is that the reason for the test failures?

kronbichler · 2022-07-20T16:10:10Z

Indeed, with this fixed I see the tests passing on my local machine. Let me push this fix, it is better coding style to attach separate arrays anyway, irrespective of this PR.

tjhei · 2022-07-20T19:45:12Z

adjusted and rebased.

bangerth · 2022-07-20T20:05:06Z

Let's hope this works now!

bangerth · 2022-07-21T02:48:01Z

The Windows compiler has some trouble:

D:\a\dealii\dealii\include\deal.II/base/numbers.h(926,34): error C2039: 'NumberType': is not a member of 'dealii::numbers::internal' [D:\a\dealii\dealii\build\source\fe\obj_fe_debug.vcxproj]
D:\a\dealii\dealii\include\deal.II/base/signaling_nan.h(33): message : see declaration of 'dealii::numbers::internal' [D:\a\dealii\dealii\build\source\fe\obj_fe_debug.vcxproj]
D:\a\dealii\dealii\include\deal.II/base/numbers.h(942): message : see reference to function template instantiation 'bool dealii::numbers::values_are_equal<Number,double>(const Number1 &,const Number2 &)' being compiled [D:\a\dealii\dealii\build\source\fe\obj_fe_debug.vcxproj]
          with
          [
              Number=double,
              Number1=double,
              Number2=double
          ]
D:\a\dealii\dealii\include\deal.II/base/tensor.h(1548): message : see reference to function template instantiation 'bool dealii::numbers::value_is_zero<Number>(const Number &)' being compiled [D:\a\dealii\dealii\build\source\fe\obj_fe_debug.vcxproj]
          with
          [
              Number=double
          ]
D:\a\dealii\dealii\include\deal.II/base/tensor.h(1547): message : while compiling class template member function 'dealii::Tensor<1,1,double> &dealii::Tensor<1,1,double>::operator =(const Number &)' [D:\a\dealii\dealii\build\source\fe\obj_fe_debug.vcxproj]
          with
          [
              Number=double
          ]
D:\a\dealii\dealii\include\deal.II/base/tensor.h(949): message : see reference to function template instantiation 'dealii::Tensor<1,1,double> &dealii::Tensor<1,1,double>::operator =(const Number &)' being compiled [D:\a\dealii\dealii\build\source\fe\obj_fe_debug.vcxproj]
          with
          [
              Number=double
          ]
D:\a\dealii\dealii\include\deal.II/base/tensor.h(2172): message : see reference to function template instantiation 'dealii::Tensor<1,1,double> dealii::internal::TensorImplementation::division_operator<1,1,double,OtherNumber,0>(const dealii::Tensor<1,1,double> &,const OtherNumber &)' being compiled [D:\a\dealii\dealii\build\source\fe\obj_fe_debug.vcxproj]
          with
          [
              OtherNumber=double
          ]
D:\a\dealii\dealii\include\deal.II/fe/mapping_q_internal.h(2013): message : see reference to function template instantiation 'dealii::Tensor<1,1,double> dealii::operator /<1,1,double,double>(const dealii::Tensor<1,1,double> &,const OtherNumber &)' being compiled [D:\a\dealii\dealii\build\source\fe\obj_fe_debug.vcxproj]
          with
          [
              OtherNumber=double
          ]
D:\a\dealii\dealii\include\deal.II/fe/mapping_q_internal.h(2098): message : see reference to function template instantiation 'void dealii::internal::MappingQImplementation::maybe_compute_face_data<1,1>(const dealii::MappingQ<1,1> &,const dealii::TriaIterator<dealii::CellAccessor<1,1>> &,const unsigned int,const unsigned int,const unsigned int,const std::vector<double,std::allocator<double>> &,const dealii::MappingQ<1,1>::InternalData &,dealii::internal::FEValuesImplementation::MappingRelatedData<1,1> &)' being compiled [D:\a\dealii\dealii\build\source\fe\obj_fe_debug.vcxproj]

tjhei · 2022-07-21T13:35:37Z

I am not sure how to make MSVC happy, so I am disabling the NaNs there. 🤷

Initialize the internal scratch_data of FEEvaluation with signaling_nans in debug mode.

otherwise MSVC tries to use numbers::internal from signaling_nan.h

tamiko · 2022-07-27T21:51:44Z

include/deal.II/base/numbers.h

@@ -892,7 +892,7 @@ namespace numbers
  {
    // Use the specialized definition for two ADOL-C taped types
    return value_is_less_than(value_1,
-                              internal::NumberType<adouble>::value(value_2));
+                              ::internal::NumberType<adouble>::value(value_2));


(oops My bad - the regression tester had been out of commission for a week. I missed a reboot.)

The global namespace qualifier ::internal is a typo. Fixed in #14162

tjhei added the ready to test label Jul 19, 2022

bangerth reviewed Jul 20, 2022

View reviewed changes

kronbichler mentioned this pull request Jul 20, 2022

MatrixFree: Use separate data arrays for two separate evaluators #14153

Merged

tjhei force-pushed the mf_nans branch from abb1a32 to e0c5524 Compare July 20, 2022 19:44

bangerth approved these changes Jul 20, 2022

View reviewed changes

kronbichler approved these changes Jul 20, 2022

View reviewed changes

masterleinad approved these changes Jul 21, 2022

View reviewed changes

tjhei added 2 commits July 21, 2022 10:59

matrix-free FEEvaluation: initialize with nan

8d15a97

Initialize the internal scratch_data of FEEvaluation with signaling_nans in debug mode.

qualify namspace for MSVC

b1883d9

otherwise MSVC tries to use numbers::internal from signaling_nan.h

tjhei force-pushed the mf_nans branch from c69001a to b1883d9 Compare July 21, 2022 15:00

masterleinad merged commit 9407279 into dealii:master Jul 21, 2022

tjhei deleted the mf_nans branch July 21, 2022 20:00

tamiko reviewed Jul 27, 2022

View reviewed changes

mkghadban pushed a commit to OpenFCST/dealii that referenced this pull request Sep 8, 2022

Merge pull request dealii#14152 from tjhei/mf_nans

01598a7

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

matrix-free FEEvaluation: initialize with nan #14152

matrix-free FEEvaluation: initialize with nan #14152

tjhei commented Jul 19, 2022

tjhei commented Jul 19, 2022

tjhei commented Jul 20, 2022

bangerth left a comment

bangerth Jul 20, 2022

tjhei Jul 20, 2022

bangerth Jul 20, 2022

tjhei Jul 20, 2022

kronbichler Jul 20, 2022

kronbichler commented Jul 20, 2022

tjhei commented Jul 20, 2022

kronbichler commented Jul 20, 2022

tjhei commented Jul 20, 2022

bangerth commented Jul 20, 2022

bangerth commented Jul 21, 2022

tjhei commented Jul 21, 2022

tamiko Jul 27, 2022

	Number(numbers::signaling_nan<double>()));
	numbers::signaling_nan<Number>());

	Number(numbers::signaling_nan<double>()));
	Number(numbers::signaling_nan<ScalarNumber>()));

	using ScalarNumber =
	typename internal::VectorizedArrayTrait<Number>::value_type;

matrix-free FEEvaluation: initialize with nan #14152

matrix-free FEEvaluation: initialize with nan #14152

Conversation

tjhei commented Jul 19, 2022

tjhei commented Jul 19, 2022

tjhei commented Jul 20, 2022

bangerth left a comment

Choose a reason for hiding this comment

bangerth Jul 20, 2022

Choose a reason for hiding this comment

tjhei Jul 20, 2022

Choose a reason for hiding this comment

bangerth Jul 20, 2022

Choose a reason for hiding this comment

tjhei Jul 20, 2022

Choose a reason for hiding this comment

kronbichler Jul 20, 2022

Choose a reason for hiding this comment

kronbichler commented Jul 20, 2022

tjhei commented Jul 20, 2022

kronbichler commented Jul 20, 2022

tjhei commented Jul 20, 2022

bangerth commented Jul 20, 2022

bangerth commented Jul 21, 2022

tjhei commented Jul 21, 2022

tamiko Jul 27, 2022

Choose a reason for hiding this comment