Prepare implementation of AVX-512 kernels

Makefile.am

@@ -142,6 +142,15 @@ endif
 endif
 endif
 
+if WITH_REAL_AVX512_BLOCK2_KERNEL
+  libelpa@SUFFIX@_private_la_SOURCES += src/elpa2_kernels/elpa2_kernels_real_avx512_2hv_double_precision.c
+if WANT_SINGLE_PRECISION_REAL
+  libelpa@SUFFIX@_private_la_SOURCES += src/elpa2_kernels/elpa2_kernels_real_avx512_2hv_single_precision.c
+endif
+
+endif
+
+
 if WITH_REAL_SSE_BLOCK4_KERNEL
   libelpa@SUFFIX@_private_la_SOURCES += src/elpa2_kernels/elpa2_kernels_real_sse_4hv_double_precision.c
 if WANT_SINGLE_PRECISION_REAL
@@ -163,6 +172,15 @@ endif
 endif
 endif
 
+if WITH_REAL_AVX512_BLOCK4_KERNEL
+  libelpa@SUFFIX@_private_la_SOURCES += src/elpa2_kernels/elpa2_kernels_real_avx512_4hv_double_precision.c
+if WANT_SINGLE_PRECISION_REAL
+  libelpa@SUFFIX@_private_la_SOURCES += src/elpa2_kernels/elpa2_kernels_real_avx-512_4hv_single_precision.c
+endif
+endif
+
+
+
 if WITH_REAL_SSE_BLOCK6_KERNEL
   libelpa@SUFFIX@_private_la_SOURCES += src/elpa2_kernels/elpa2_kernels_real_sse_6hv_double_precision.c
 if WANT_SINGLE_PRECISION_REAL
@@ -184,6 +202,14 @@ endif
 endif
 endif
 
+if WITH_REAL_AVX512_BLOCK6_KERNEL
+  libelpa@SUFFIX@_private_la_SOURCES += src/elpa2_kernels/elpa2_kernels_real_avx512_6hv_double_precision.c
+if WANT_SINGLE_PRECISION_REAL
+  libelpa@SUFFIX@_private_la_SOURCES += src/elpa2_kernels/elpa2_kernels_real_avx512_6hv_single_precision.c
+endif
+endif
+
+
 if WITH_COMPLEX_SSE_BLOCK1_KERNEL
   libelpa@SUFFIX@_private_la_SOURCES += src/elpa2_kernels/elpa2_kernels_complex_sse_1hv_double_precision.c
 if WANT_SINGLE_PRECISION_COMPLEX
@@ -205,6 +231,14 @@ endif
 endif
 endif
 
+
+if WITH_COMPLEX_AVX512_BLOCK1_KERNEL
+  libelpa@SUFFIX@_private_la_SOURCES += src/elpa2_kernels/elpa2_kernels_complex_avx512_1hv_double_precision.c
+if WANT_SINGLE_PRECISION_COMPLEX
+  libelpa@SUFFIX@_private_la_SOURCES += src/elpa2_kernels/elpa2_kernels_complex_avx512_1hv_single_precision.c
+endif
+endif
+
 if WITH_COMPLEX_SSE_BLOCK2_KERNEL
   libelpa@SUFFIX@_private_la_SOURCES += src/elpa2_kernels/elpa2_kernels_complex_sse_2hv_double_precision.c
 if WANT_SINGLE_PRECISION_COMPLEX
@@ -226,6 +260,13 @@ endif
 endif
 endif
 
+if WITH_COMPLEX_AVX512_BLOCK2_KERNEL
+  libelpa@SUFFIX@_private_la_SOURCES += src/elpa2_kernels/elpa2_kernels_complex_avx512_2hv_double_precision.c
+if WANT_SINGLE_PRECISION_COMPLEX
+  libelpa@SUFFIX@_private_la_SOURCES += src/elpa2_kernels/elpa2_kernels_complex_avx-512_2hv_single_precision.c
+endif
+endif
+
 .cu.lo:
 	NVCC="$(NVCC)" libtool --mode=compile --tag=CC $(top_srcdir)/nvcc_wrap $(NVCCFLAGS) $(LDFLAGS) -I$(top_builddir)/ -I$(top_srcdir)/  -c $< -o $@
 

configure.ac

@@ -863,6 +863,21 @@ if test x"${install_real_avx2_block6}" = x"yes" ; then
  AC_DEFINE([WITH_REAL_AVX2_BLOCK6_KERNEL],[1],[can use real_avx2_block6 kernel])
 fi
 
+AM_CONDITIONAL([WITH_REAL_AVX512_BLOCK2_KERNEL],[test x"$install_real_avx512_block2" = x"yes"])
+if test x"${install_real_avx512_block2}" = x"yes" ; then
+ AC_DEFINE([WITH_REAL_AVX512_BLOCK2_KERNEL],[1],[can use real_avx512_block2 kernel])
+fi
+
+AM_CONDITIONAL([WITH_REAL_AVX512_BLOCK4_KERNEL],[test x"$install_real_avx512_block4" = x"yes"])
+if test x"${install_real_avx512_block4}" = x"yes" ; then
+ AC_DEFINE([WITH_REAL_AVX512_BLOCK4_KERNEL],[1],[can use real_avx512_block4 kernel])
+fi
+
+AM_CONDITIONAL([WITH_REAL_AVX512_BLOCK6_KERNEL],[test x"$install_real_avx512_block6" = x"yes"])
+if test x"${install_real_avx512_block6}" = x"yes" ; then
+ AC_DEFINE([WITH_REAL_AVX512_BLOCK6_KERNEL],[1],[can use real_avx512_block6 kernel])
+fi
+
 AM_CONDITIONAL([WITH_COMPLEX_SSE_BLOCK1_KERNEL],[test x"$install_complex_sse_block1" = x"yes"])
 if test x"${install_complex_sse_block1}" = x"yes" ; then
  AC_DEFINE([WITH_COMPLEX_SSE_BLOCK1_KERNEL],[1],[can use complex_sse_block1 kernel])
@@ -893,6 +908,16 @@ if test x"${install_complex_avx2_block2}" = x"yes" ; then
  AC_DEFINE([WITH_COMPLEX_AVX2_BLOCK2_KERNEL],[1],[can use complex_avx2_block2 kernel])
 fi
 
+AM_CONDITIONAL([WITH_COMPLEX_AVX512_BLOCK1_KERNEL],[test x"$install_complex_avx512_block1" = x"yes"])
+if test x"${install_complex_avx512_block1}" = x"yes" ; then
+ AC_DEFINE([WITH_COMPLEX_AVX512_BLOCK1_KERNEL],[1],[can use complex_avx512_block1 kernel])
+fi
+
+AM_CONDITIONAL([WITH_COMPLEX_AVX512_BLOCK2_KERNEL],[test x"$install_complex_avx512_block2" = x"yes"])
+if test x"${install_complex_avx512_block2}" = x"yes" ; then
+ AC_DEFINE([WITH_COMPLEX_AVX512_BLOCK2_KERNEL],[1],[can use complex_avx512_block2 kernel])
+fi
+
 AM_CONDITIONAL([WITH_REAL_BGP_KERNEL],[test x"$install_real_bgp" = x"yes"])
 if test x"${install_real_bgp}" = x"yes" ; then
  AC_DEFINE([WITH_REAL_BGP_KERNEL],[1],[can use real BGP kernel])