Intrinsic for the high-performance templated operator [x86]

[camierjs]

Enables the performance templated classes to use specific vector intrinsics on the following architectures:

• x86 (SSE/AVX/AVX2/AVX512),
• Power8 & Power9 (VSX), tested on Lassen with xlc++,
• BG/Q (QPX), tested on Vulcan.

It can be enabled with MFEM_USE_SIMD=YES.

Results on Vulcan, master:

Vulcan, x86, with the SIMD enabled:

Vulcan, x86, with SIMD and JIT to use vectorization only when needed:

PR	Author	Editor	Reviewers	Assignment	Approval	Merge
#1047	@camierjs	@tzanio	@tzanio + @v-dobrev	03/12/20	05/18/20	05/25/20

[tzanio]

Merged in next for testing...

[tzanio]

There are errors from this PR in the autotest runs on tux429, e.g.

nvcc  -O3 -std=c++11 -x=cu --expt-extended-lambda -arch=sm_60 -ccbin g++ -Xcompiler="-march=native  -Wall" -I../.. -I../../../occa/include -I../../../libCEED/include -I../../../raja/include -Xcompiler=-fopenmp ex1.cpp -o ex1 -L../.. -lmfem -Xlinker=-rpath,../../../occa/lib -L../../../occa/lib -locca -Xlinker=-rpath,../../../libCEED/lib -L../../../libCEED/lib -lceed -Xlinker=-rpath,../../../raja/lib -L../../../raja/lib -lRAJA   -lrt
../../linalg/ttensor.hpp(300): error: "mfem::AutoSIMD<double, 8, 64> &(int)" contains a vector, which is not supported in device code
          detected during:
            instantiation of class "mfem::TVector<S, data_t, align> [with S=750, data_t=mfem::AutoSIMD<double, 8, 64>, align=false]" 
(350): here
            instantiation of class "mfem::TMatrix<N1, N2, data_t, align> [with N1=125, N2=6, data_t=mfem::AutoSIMD<double, 8, 64>, align=false]" 
../../general/mem_manager.hpp(780): here
            instantiation of "void mfem::Memory<T>::Delete() [with T=mfem::TMatrix<125, 6, mfem::AutoSIMD<double, 8, 64>, false>]" 
../../fem/tbilinearform.hpp(129): here
            instantiation of "mfem::TBilinearForm<meshType, solFESpace, IR, IntegratorType, solVecLayout_t, complex_t, real_t, impl_traits_t>::~TBilinearForm() [with meshType=mesh_t, solFESpace=sol_fes_t, IR=int_rule_t, IntegratorType=integ_t, solVecLayout_t=mfem::ScalarLayout, complex_t=double, real_t=double, impl_traits_t=mfem::AutoSIMDTraits<double, double>]" 
../../fem/tbilinearform.hpp(125): here
            instantiation of "mfem::TBilinearForm<meshType, solFESpace, IR, IntegratorType, solVecLayout_t, complex_t, real_t, impl_traits_t>::TBilinearForm(const IntegratorType &, const mfem::FiniteElementSpace &) [with meshType=mesh_t, solFESpace=sol_fes_t, IR=int_rule_t, IntegratorType=integ_t, solVecLayout_t=mfem::ScalarLayout, complex_t=double, real_t=double, impl_traits_t=mfem::AutoSIMDTraits<double, double>]" 
ex1.cpp(288): here

[v-dobrev]

I'm not sure why, on my desktop, I did not get this error.

It maybe okay to remove the MFEM_HOST_DEVICE from the methods that generate the error:

mfem/linalg/ttensor.hpp

Lines 300 to 301 in 069ad59

	MFEM_HOST_DEVICE data_t &operator[](int i) { return data[i]; }
	MFEM_HOST_DEVICE const data_t &operator[](int i) const { return data[i]; }

[v-dobrev]

Re-merged in next.

[tzanio]

The tux429 runs look OK now, but there are new errors on tux426 ☹️

g++  -O3 -std=c++11 -I..  ex21.cpp -o ex21 -L.. -lmfem -lrt
g++ -g -c /home/kolev1/autotest-mfem/tux426/mfem/tests/unit/unit_test_main.cpp  -O3 -std=c++11 -I../..  -I. -I../.. -o unit_test_main.o
make[1]: Entering directory `/home/kolev1/autotest-mfem/tux426/mfem/miniapps/performance'
g++  -O3 -std=c++11 -march=native -pedantic -Wall -I../..  ex1.cpp -o ex1 -L../.. -lmfem -lrt
In file included from ../../linalg/../linalg/simd/x86.hpp:19:0,
                 from ../../linalg/../linalg/simd.hpp:25,
                 from ../../linalg/ttensor.hpp:16,
                 from ../../mfem-performance.hpp:19,
                 from ex1.cpp:32:
../../linalg/../linalg/simd/m512.hpp: In member function ‘mfem::AutoSIMD<double, 8, 64> mfem::AutoSIMD<double, 8, 64>::operator-() const’:
../../linalg/../linalg/simd/m512.hpp:115:58: error: ‘_mm512_xor_pd’ was not declared in this scope
       r.m512d = _mm512_xor_pd(_mm512_set1_pd(-0.0), m512d);
                                                          ^
make[1]: *** [ex1] Error 1
make[1]: Leaving directory `/home/kolev1/autotest-mfem/tux426/mfem/miniapps/performance'

[v-dobrev]

It's good we are testing AVX512 too. 😄

It looks like this particular intrinsic is available only when __AVX512DQ__ is defined. I'm trying to figure out what alternative intrinsic can be used when __AVX512DQ__ is not defined.

[v-dobrev]

I guess we can always fall back on return (0.0-(*this)); which will be the same as

   AutoSIMD<double,8,64> r;
   r.m512d = _mm512_sub_pd(_mm512_set1_pd(0.0),v.m512d);
   return r;

However, there may be a better/faster alternative.

[v-dobrev]

Here is one alternative:
https://github.com/vectorclass/version2/blob/c1e56e55371eab3e423efd90f4bdd66bef0d75b2/vectorf512.h#L900-L904

which expanded becomes:

_mm512_castsi512_pd(
   _mm512_xor_epi32(
      _mm512_castpd_si512(a),
      _mm512_set1_epi64(0x8000000000000000)));

I'll push this in a moment.

@tzanio, can you try it on tux426?

[v-dobrev]

Re-merged in next.

[tzanio]

Re-merged in next for testing ...