`performance_ex1` with triangles

[aschaf]

`performance_ex1` with triangles

[v-dobrev]

Your suggested fixes look good to me. Thanks @aschaf!

[aschaf]

I opened #4206 to resolve the issue.

A related question: if I want to further speed up the assembly (eg. on my laptop) should I use the MPI parallel version to utilize all available cores, or would it make sense to somehow extend the TBilinearform to utilize eg. OpenMP for the assembly loop?

[v-dobrev]

A related question: if I want to further speed up the assembly (eg. on my laptop) should I use the MPI parallel version to utilize all available cores, or would it make sense to somehow extend the TBilinearform to utilize eg. OpenMP for the assembly loop?

The MPI-parallel version should be easier to try -- just look at how miniapps/performance/ex1p.cpp differes from miniapps/performance/ex1.cpp. Implementing an OpenMP version will require some effort and in the end may actually be slower.

[aschaf]

I noticed another issue I somewhat missed in the pull request. When turning on SIMD, it again crashes at line 103 in teltrans.hpp

mfem/fem/teltrans.hpp

Lines 94 to 106 in 7c296d0

	template <typename vint_t, int NE>
	inline MFEM_ALWAYS_INLINE
	void SetAttributes(int el, vint_t (&attrib)[NE]) const
	{
	const int vsize = sizeof(vint_t)/sizeof(attrib[0][0]);
	for (int i = 0; i < NE; i++)
	{
	for (int j = 0; j < vsize; j++)
	{
	attrib[i][j] = elements[el+j+i*vsize]->GetAttribute();
	}
	}
	}

specifically because at some point el+j+i*vsize is greater than the number of elements. A quick and dirty fix was just to replace line 103 with

attrib[i][j] = ((el+j+i*vsize) < fes.GetNE()) ? elements[el+j+i*vsize]->GetAttribute() : 1;

[v-dobrev]

@aschaf, you are right -- this is a bug. In the respective situation in VectorExtract, we stop copying values when we reach the last element, see

mfem/fem/tfespace.hpp

Line 214 in 7c296d0

void VectorExtract(const vec_layout_t &vl,

A better approach for here and for VectorExtract is probably to replicate the last element attribute or values, so that all SIMD entries have meaningful values.

[aschaf]

@v-dobrev Thank you for the explanation, I will open a PR once this is fixed.