Merge branch 'master' into ELPA_GPU

09d13e2b · Andreas Marek · 62fe6edc · 0d256c1b · 09d13e2b · 09d13e2b
Commit 09d13e2b authored Apr 20, 2016 by Andreas Marek
17 changed files
--- a/.gitignore
+++ b/.gitignore
@@ -12,6 +12,7 @@ autom4te.cache
 compile
 config.guess
 config.h.in
+config.h.in~
 config.sub
 configure
 depcomp
@@ -19,3 +20,8 @@ install-sh
 ltmain.sh
 missing
 test-driver
+m4/libtool.m4
+m4/ltoptions.m4
+m4/ltsugar.m4
+m4/ltversion.m4
+m4/lt~obsolete.m4
--- a/.gitlab-ci.yml
+++ b/.gitlab-ci.yml
 jobs:
-  script: ./autogen.sh && ./configure && make && make check TEST_FLAGS='1500 50 16'
+  script:
+    - export LANG=C
+    - module load impi intel gcc mkl autotools
+    - ./autogen.sh
+    - ./configure CFLAGS="-O3 -mavx" CXXFLAGS="-O3 -mavx" FCFLAGS="-O3 -mavx" SCALAPACK_LDFLAGS="-L/afs/@cell/common/soft/intel/ics2015/15.0/mkl/lib/intel64 -lmkl_scalapack_lp64 -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lmkl_blacs_intelmpi_lp64 -lpthread -lm -Wl,-rpath,$MKL_HOME/lib/intel64" SCALAPACK_FCFLAGS="-L/afs/@cell/common/soft/intel/ics2015/15.0/mkl/lib/intel64 -lmkl_scalapack_lp64 -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lmkl_blacs_intelmpi_lp64 -lpthread -lm -I$MKLROOT/include/intel64/lp64"
+    - make -j 8
+    - make check TEST_FLAGS='1500 50 16'
--- a/Makefile.am
+++ b/Makefile.am
@@ -82,18 +82,18 @@ if WITH_REAL_BGQ_KERNEL
  libelpa@SUFFIX@_la_SOURCES += src/elpa2_kernels/elpa2_kernels_real_bgq.f90
 endif

-if WITH_REAL_SSE_KERNEL
+if WITH_REAL_SSE_ASSEMBLY_KERNEL
  libelpa@SUFFIX@_la_SOURCES += src/elpa2_kernels/elpa2_kernels_asm_x86_64_double_precision.s
 if WANT_SINGLE_PRECISION_REAL
  libelpa@SUFFIX@_la_SOURCES += src/elpa2_kernels/elpa2_kernels_asm_x86_64_single_precision.s
 endif
+
 else
-if WITH_COMPLEX_SSE_KERNEL
+if WITH_COMPLEX_SSE_ASSEMBLY_KERNEL
  libelpa@SUFFIX@_la_SOURCES += src/elpa2_kernels/elpa2_kernels_asm_x86_64_double_precision.s
 if WANT_SINGLE_PRECISION_COMPLEX
  libelpa@SUFFIX@_la_SOURCES += src/elpa2_kernels/elpa2_kernels_asm_x86_64_single_precision.s
 endif
-
 endif
 endif


--- a/configure.ac
+++ b/configure.ac
@@ -202,55 +202,77 @@ if test x"${with_ftimings}" = x"yes"; then
 fi
 AM_CONDITIONAL([HAVE_DETAILED_TIMINGS],[test x"$with_ftimings" = x"yes"])

-AC_MSG_CHECKING(whether double-precision SSE assembler kernel can be compiled)
+AC_MSG_CHECKING(whether double-precision SSE assembly kernel can be compiled)

 $CC -c $srcdir/src/elpa2_kernels/elpa2_kernels_asm_x86_64_double_precision.s -o test.o 2>/dev/null
 if test "$?" == 0; then
-  can_compile_sse=yes
-  install_real_sse=yes
+  can_compile_sse_assembly=yes
+  install_real_sse_assembly=yes
+  install_complex_sse_assembly=yes
+else
+  can_compile_sse_assembly=no
+  install_real_sse_assembly=no
+  install_complex_sse_assembly=no
+fi
+
+rm -f ./test.o
+AC_MSG_RESULT([${can_compile_sse_assembly}])
+
+if test x"${want_single_precision}" = x"yes" ; then
+  AC_MSG_CHECKING(whether single-precision SSE assembly kernel can be compiled)
+
+  $CC -c $srcdir/src/elpa2_kernels/elpa2_kernels_asm_x86_64_single_precision.s -o test.o 2>/dev/null
+  if test "$?" == 0; then
+    can_compile_sse_assembly=yes
+    install_real_sse_assembly=yes
+    install_complex_sse_assembly=yes
+  else
+    can_compile_sse_assembly=no
+    install_real_sse_assembly=no
+    install_complex_sse_assembly=no
+  fi
+  rm -f ./test.o
+  AC_MSG_RESULT([${can_compile_sse_assembly}])
+  if test x"${can_compile_sse_assembly}" = x"no" ; then
+    AC_MSG_WARN([Cannot compile single-precision SSE assembly kernel: disabling SSE assembly kernels alltogether])
+  fi
+fi
+
+dnl check whether on can compile with sse-gcc intrinsics
+AC_MSG_CHECKING(whether we can compile SSE with gcc intrinsics in C)
+AC_COMPILE_IFELSE([AC_LANG_SOURCE([
+ #include <x86intrin.h>
+ int main(int argc, char **argv){
+ double* q;
+ __m128d h1 = _mm_loaddup_pd(q);
+ return 0;
+ }
+ ])],
+ [can_compile_sse_intrinsics=yes],
+ [can_compile_sse_intrinsics=no]
+)
+AC_MSG_RESULT([${can_compile_sse_intrinsics}])
+
+if test "${can_compile_sse_intrinsics}" = "yes"; then
+  install_real_sse_intrinsics=yes
  install_real_sse_block2=yes
  install_real_sse_block4=yes
  install_real_sse_block6=yes

-  install_complex_sse=yes
+  install_complex_sse_intrinsics=yes
  install_complex_sse_block1=yes
  install_complex_sse_block2=yes
-
 else
-  can_compile_sse=no
-  install_real_sse=no
+  install_real_sse_intrinsics=no
  install_real_sse_block2=no
  install_real_sse_block4=no
  install_real_sse_block6=no

-  install_complex_sse=no
+  install_complex_sse_intrinsics=no
  install_complex_sse_block1=no
  install_complex_sse_block2=no
 fi

-rm -f ./test.o
-AC_MSG_RESULT([${can_compile_sse}])
-
-if test x"${want_single_precision}" = x"yes" ; then
-  AC_MSG_CHECKING(whether single-precision SSE assembler kernel can be compiled)
-
-  $CC -c $srcdir/src/elpa2_kernels/elpa2_kernels_asm_x86_64_double_precision.s -o test.o 2>/dev/null
-  if test "$?" == 0; then
-    can_compile_sse=yes
-    install_real_sse=yes
-    install_complex_sse=yes
-  else
-    can_compile_sse=no
-    install_real_sse=no
-    install_complex_sse=no
-  fi
-  rm -f ./test.o
-  AC_MSG_RESULT([${can_compile_sse}])
-  if test x"${can_compile_sse}" = x"no" ; then
-    AC_MSG_WARN([Cannot compile single-precision SSE kernel: disabling SSE kernels alltogether])
-  fi
-fi
-
 dnl check whether one can compile with avx - gcc intrinsics

 dnl first pass: try with specified CFLAGS and CXXFLAGS
@@ -356,10 +378,16 @@ else
  install_complex_avx2_block1=no
  install_complex_avx2_block2=no
 fi
-AM_CONDITIONAL([HAVE_SSE],[test x"$can_compile_sse" = x"yes"])
-if test x"${can_compile_sse}" = x"yes" ; then
-  AC_DEFINE([HAVE_SSE],[1],[SSE is supported on this CPU])
+
+AM_CONDITIONAL([HAVE_SSE_ASSEMBLY],[test x"$can_compile_sse_assembly" = x"yes"])
+if test x"${can_compile_sse_assembly}" = x"yes" ; then
+  AC_DEFINE([HAVE_SSE_ASSEMBLY],[1],[assembly SSE is supported on this CPU])
+fi
+AM_CONDITIONAL([HAVE_SSE_INTRINSICS],[test x"$can_compile_sse_intrinsics" = x"yes"])
+if test x"${can_compile_sse_intrinsics}" = x"yes" ; then
+  AC_DEFINE([HAVE_SSE_INTRINSICS],[1],[gcc intrinsics SSE is supported on this CPU])
 fi
+
 AM_CONDITIONAL([HAVE_AVX],[test x"$can_compile_avx" = x"yes"])
 if test x"${can_compile_avx}" = x"yes" ; then
  AC_DEFINE([HAVE_AVX],[1],[AVX is supported on this CPU])
@@ -671,7 +699,7 @@ dnl generic-simple kernel
 DEFINE_OPTION_SPECIFIC_REAL_KERNEL([real-generic-simple-kernel-only],[generic-simple-kernel],[install_real_generic_simple])

 dnl sse kernel
-DEFINE_OPTION_SPECIFIC_REAL_KERNEL([real-sse-kernel-only],[sse-kernel],[install_real_sse])
+DEFINE_OPTION_SPECIFIC_REAL_KERNEL([real-sse-assembly-kernel-only],[sse-assembly-kernel],[install_real_sse_assembly])

 dnl bgp kernel
 DEFINE_OPTION_SPECIFIC_REAL_KERNEL([real-bgp-kernel-only],[bgp-kernel],[install_real_bgp])
@@ -706,7 +734,7 @@ dnl generic-simple kernel
 DEFINE_OPTION_SPECIFIC_COMPLEX_KERNEL([complex-generic-simple-kernel-only],[generic-simple-kernel],[install_complex_generic_simple])

 dnl sse kernel
-DEFINE_OPTION_SPECIFIC_COMPLEX_KERNEL([complex-sse-kernel-only],[sse-kernel],[install_complex_sse])
+DEFINE_OPTION_SPECIFIC_COMPLEX_KERNEL([complex-sse-assembly-kernel-only],[sse-assembly-kernel],[install_complex_sse_assembly])

 dnl complex-bqp kernel
 DEFINE_OPTION_SPECIFIC_COMPLEX_KERNEL([complex-bgp-kernel-only],[bgp-kernel],[install_complex_bgp])
@@ -757,14 +785,14 @@ if test x"${install_complex_generic_simple}" = x"yes" ; then
 AC_DEFINE([WITH_COMPLEX_GENERIC_SIMPLE_KERNEL],[1],[can use complex generic-simple kernel])
 fi

-AM_CONDITIONAL([WITH_REAL_SSE_KERNEL],[test x"$install_real_sse" = x"yes"])
-if test x"${install_real_sse}" = x"yes" ; then
- AC_DEFINE([WITH_REAL_SSE_KERNEL],[1],[can use real SSE kernel])
+AM_CONDITIONAL([WITH_REAL_SSE_ASSEMBLY_KERNEL],[test x"$install_real_sse_assembly" = x"yes"])
+if test x"${install_real_sse_assembly}" = x"yes" ; then
+ AC_DEFINE([WITH_REAL_SSE_ASSEMBLY_KERNEL],[1],[can use real SSE assembly kernel])
 fi

-AM_CONDITIONAL([WITH_COMPLEX_SSE_KERNEL],[test x"$install_complex_sse" = x"yes"])
-if test x"${install_complex_sse}" = x"yes" ; then
- AC_DEFINE([WITH_COMPLEX_SSE_KERNEL],[1],[can use complex SSE kernel])
+AM_CONDITIONAL([WITH_COMPLEX_SSE_ASSEMBLY_KERNEL],[test x"$install_complex_sse_assembly" = x"yes"])
+if test x"${install_complex_sse_assembly}" = x"yes" ; then
+ AC_DEFINE([WITH_COMPLEX_SSE_ASSEMBLY_KERNEL],[1],[can use complex SSE assembly kernel])
 fi

 AM_CONDITIONAL([WITH_REAL_SSE_BLOCK2_KERNEL],[test x"$install_real_sse_block2" = x"yes"])

--- a/src/elpa2_kernels/elpa2_kernels_complex_sse_1hv_double_precision.c
+++ b/src/elpa2_kernels/elpa2_kernels_complex_sse_1hv_double_precision.c
@@ -67,7 +67,7 @@

 #define __forceinline __attribute__((always_inline))

-#ifdef HAVE_SSE
+#ifdef HAVE_SSE_INTRINSICS
 #undef __AVX__
 #endif

@@ -78,7 +78,7 @@ static __forceinline void hh_trafo_complex_kernel_4_SSE_1hv_double(double comple
 static __forceinline void hh_trafo_complex_kernel_2_SSE_1hv_double(double complex* q, double complex* hh, int nb, int ldq);

 /*
-!f>#ifdef HAVE_SSE
+!f>#ifdef HAVE_SSE_INTRINSICS
 !f> interface
 !f>   subroutine single_hh_trafo_complex_sse_1hv_double(q, hh, pnb, pnq, pldq) &
 !f>                             bind(C, name="single_hh_trafo_complex_sse_1hv_double")

--- a/src/elpa2_kernels/elpa2_kernels_complex_sse_1hv_single_precision.c
+++ b/src/elpa2_kernels/elpa2_kernels_complex_sse_1hv_single_precision.c
@@ -67,7 +67,7 @@

 #define __forceinline __attribute__((always_inline))

-#ifdef HAVE_SSE
+#ifdef HAVE_SSE_INTRINSICS
 #undef __AVX__
 #endif

@@ -78,7 +78,7 @@ static __forceinline void hh_trafo_complex_kernel_4_SSE_1hv_single(complex* q, c
 static __forceinline void hh_trafo_complex_kernel_2_SSE_1hv_single(complex* q, complex* hh, int nb, int ldq);

 /*
-!f>#ifdef HAVE_SSE
+!f>#ifdef HAVE_SSE_INTRINSICS
 !f> interface
 !f>   subroutine single_hh_trafo_complex_sse_1hv_single(q, hh, pnb, pnq, pldq) &
 !f>                             bind(C, name="single_hh_trafo_complex_sse_1hv_single")

--- a/src/elpa2_kernels/elpa2_kernels_complex_sse_2hv_double_precision.c
+++ b/src/elpa2_kernels/elpa2_kernels_complex_sse_2hv_double_precision.c
@@ -66,7 +66,7 @@

 #define __forceinline __attribute__((always_inline))

-#ifdef HAVE_SSE
+#ifdef HAVE_SSE_INTRINSICS
 #undef __AVX__
 #endif

@@ -78,7 +78,7 @@ static __forceinline void hh_trafo_complex_kernel_2_SSE_2hv_double(double comple
 static __forceinline void hh_trafo_complex_kernel_1_SSE_2hv_double(double complex* q, double complex* hh, int nb, int ldq, int ldh, double complex s);

 /*
-!f>#ifdef HAVE_SSE
+!f>#ifdef HAVE_SSE_INTRINSICS
 !f> interface
 !f>   subroutine double_hh_trafo_complex_sse_2hv_double(q, hh, pnb, pnq, pldq, pldh) &
 !f>                             bind(C, name="double_hh_trafo_complex_sse_2hv_double")

--- a/src/elpa2_kernels/elpa2_kernels_complex_sse_2hv_single_precision.c
+++ b/src/elpa2_kernels/elpa2_kernels_complex_sse_2hv_single_precision.c
@@ -66,7 +66,7 @@

 #define __forceinline __attribute__((always_inline))

-#ifdef HAVE_SSE
+#ifdef HAVE_SSE_INTRINSICS
 #undef __AVX__
 #endif

@@ -74,7 +74,7 @@
 static __forceinline void hh_trafo_complex_kernel_4_SSE_2hv_single(complex* q, complex* hh, int nb, int ldq, int ldh, complex s, complex s1);

 /*
-!f>#ifdef HAVE_SSE
+!f>#ifdef HAVE_SSE_INTRINSICS
 !f> interface
 !f>   subroutine double_hh_trafo_complex_sse_2hv_single(q, hh, pnb, pnq, pldq, pldh) &
 !f>                             bind(C, name="double_hh_trafo_complex_sse_2hv_single")

--- a/src/elpa2_kernels/elpa2_kernels_real_sse_2hv_double_precision.c
+++ b/src/elpa2_kernels/elpa2_kernels_real_sse_2hv_double_precision.c
@@ -67,7 +67,7 @@
 #define __forceinline __attribute__((always_inline)) static


-#ifdef HAVE_SSE
+#ifdef HAVE_SSE_INTRINSICS
 #undef __AVX__
 #endif

@@ -79,7 +79,7 @@ __forceinline void hh_trafo_kernel_12_SSE_2hv_double(double* q, double* hh, int
 void double_hh_trafo_real_sse_2hv_double(double* q, double* hh, int* pnb, int* pnq, int* pldq, int* pldh);

 /*
-!f>#ifdef HAVE_SSE
+!f>#ifdef HAVE_SSE_INTRINSICS
 !f> interface
 !f>   subroutine double_hh_trafo_real_sse_2hv_double(q, hh, pnb, pnq, pldq, pldh) &
 !f>                             bind(C, name="double_hh_trafo_real_sse_2hv_double")

--- a/src/elpa2_kernels/elpa2_kernels_real_sse_2hv_single_precision.c
+++ b/src/elpa2_kernels/elpa2_kernels_real_sse_2hv_single_precision.c
@@ -67,7 +67,7 @@
 #define __forceinline __attribute__((always_inline)) static


-#ifdef HAVE_SSE
+#ifdef HAVE_SSE_INTRINSICS
 #undef __AVX__
 #endif

@@ -79,7 +79,7 @@ __forceinline void hh_trafo_kernel_12_SSE_2hv_single(float* q, float* hh, int nb
 void double_hh_trafo_real_sse_2hv_single_(float* q, float* hh, int* pnb, int* pnq, int* pldq, int* pldh);

 /*
-!f>#ifdef HAVE_SSE
+!f>#ifdef HAVE_SSE_INTRINSICS
 !f> interface
 !f>   subroutine double_hh_trafo_real_sse_2hv_single(q, hh, pnb, pnq, pldq, pldh) &
 !f>                             bind(C, name="double_hh_trafo_real_sse_2hv_single")

--- a/src/elpa2_kernels/elpa2_kernels_real_sse_4hv_double_precision.c
+++ b/src/elpa2_kernels/elpa2_kernels_real_sse_4hv_double_precision.c
@@ -66,7 +66,7 @@

 #define __forceinline __attribute__((always_inline)) static

-#ifdef HAVE_SSE
+#ifdef HAVE_SSE_INTRINSICS
 #undef __AVX__
 #endif

@@ -78,7 +78,7 @@ __forceinline void hh_trafo_kernel_6_SSE_4hv_double(double* q, double* hh, int n
 void quad_hh_trafo_real_sse_4hv_double(double* q, double* hh, int* pnb, int* pnq, int* pldq, int* pldh);

 /*
-!f>#ifdef HAVE_SSE
+!f>#ifdef HAVE_SSE_INTRINSICS
 !f> interface
 !f>   subroutine quad_hh_trafo_real_sse_4hv_double(q, hh, pnb, pnq, pldq, pldh) &
 !f>                             bind(C, name="quad_hh_trafo_real_sse_4hv_double")

--- a/src/elpa2_kernels/elpa2_kernels_real_sse_4hv_single_precision.c
+++ b/src/elpa2_kernels/elpa2_kernels_real_sse_4hv_single_precision.c
@@ -66,7 +66,7 @@

 #define __forceinline __attribute__((always_inline)) static

-#ifdef HAVE_SSE
+#ifdef HAVE_SSE_INTRINSICS
 #undef __AVX__
 #endif
 __forceinline void hh_trafo_kernel_4_SSE_4hv_single(float* q, float* hh, int nb, int ldq, int ldh, float s_1_2, float s_1_3, float s_2_3, float s_1_4, float s_2_4, float s_3_4);
@@ -76,7 +76,7 @@ __forceinline void hh_trafo_kernel_12_SSE_4hv_single(float* q, float* hh, int nb
 void quad_hh_trafo_real_sse_4hv_single_(float* q, float* hh, int* pnb, int* pnq, int* pldq, int* pldh);

 /*
-!f>#ifdef HAVE_SSE
+!f>#ifdef HAVE_SSE_INTRINSICS
 !f> interface
 !f>   subroutine quad_hh_trafo_real_sse_4hv_single(q, hh, pnb, pnq, pldq, pldh) &
 !f>                             bind(C, name="quad_hh_trafo_real_sse_4hv_single")

--- a/src/elpa2_kernels/elpa2_kernels_real_sse_6hv_double_precision.c
+++ b/src/elpa2_kernels/elpa2_kernels_real_sse_6hv_double_precision.c
@@ -66,7 +66,7 @@

 #define __forceinline __attribute__((always_inline)) static

-#ifdef HAVE_SSE
+#ifdef HAVE_SSE_INTRINSICS
 #undef __AVX__
 #endif

@@ -76,7 +76,7 @@ static void hh_trafo_kernel_4_SSE_6hv_double(double* q, double* hh, int nb, int
 void hexa_hh_trafo_real_sse_6hv_double(double* q, double* hh, int* pnb, int* pnq, int* pldq, int* pldh);

 /*
-!f>#ifdef HAVE_SSE
+!f>#ifdef HAVE_SSE_INTRINSICS
 !f> interface
 !f>   subroutine hexa_hh_trafo_real_sse_6hv_double(q, hh, pnb, pnq, pldq, pldh) &
 !f>                             bind(C, name="hexa_hh_trafo_real_sse_6hv_double")

--- a/src/elpa2_kernels/elpa2_kernels_real_sse_6hv_single_precision.c
+++ b/src/elpa2_kernels/elpa2_kernels_real_sse_6hv_single_precision.c
@@ -66,7 +66,7 @@

 #define __forceinline __attribute__((always_inline)) static

-#ifdef HAVE_SSE
+#ifdef HAVE_SSE_INTRINSICS
 #undef __AVX__
 #endif

@@ -80,7 +80,7 @@ static void hh_trafo_kernel_8_SSE_6hv_single(float* q, float* hh, int nb, int ld
 void hexa_hh_trafo_real_sse_6hv_single_(float* q, float* hh, int* pnb, int* pnq, int* pldq, int* pldh);

 /*
-!f>#ifdef HAVE_SSE
+!f>#ifdef HAVE_SSE_INTRINSICS
 !f> interface
 !f>   subroutine hexa_hh_trafo_real_sse_6hv_single(q, hh, pnb, pnq, pldq, pldh) &
 !f>                             bind(C, name="hexa_hh_trafo_real_sse_6hv_single")

--- a/src/elpa2_utilities.F90
+++ b/src/elpa2_utilities.F90
@@ -132,7 +132,7 @@ module ELPA2_utilities
 #ifdef WITH_REAL_GENERIC_SIMPLE_KERNEL
  integer(kind=ik), parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_GENERIC_SIMPLE
 #endif
-#ifdef WITH_REAL_SSE_KERNEL
+#ifdef WITH_REAL_SSE_ASSEMBLY_KERNEL
  integer(kind=ik), parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_SSE
 #endif
 #ifdef WITH_REAL_AVX_BLOCK2_KERNEL
@@ -168,7 +168,7 @@ module ELPA2_utilities
 #ifdef WITH_REAL_GENERIC_SIMPLE_KERNEL
  integer(kind=ik), parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_GENERIC_SIMPLE
 #endif
-#ifdef WITH_REAL_SSE_KERNEL
+#ifdef WITH_REAL_SSE_ASSEMBLY_KERNEL
  integer(kind=ik), parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_SSE
 #endif
 #ifdef WITH_REAL_AVX_BLOCK2_KERNEL
@@ -238,7 +238,7 @@ module ELPA2_utilities
 #ifdef WITH_COMPLEX_GENERIC_SIMPLE_KERNEL
  integer(kind=ik), parameter :: DEFAULT_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_GENERIC_SIMPLE
 #endif
-#ifdef WITH_COMPLEX_SSE_KERNEL
+#ifdef WITH_COMPLEX_SSE_ASSEMBLY_KERNEL
  integer(kind=ik), parameter :: DEFAULT_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_SSE
 #endif
 #ifdef WITH_COMPLEX_AVX1_BLOCK1_KERNEL
@@ -267,7 +267,7 @@ module ELPA2_utilities
 #ifdef WITH_COMPLEX_GENERIC_SIMPLE_KERNEL
  integer(kind=ik), parameter :: DEFAULT_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_GENERIC_SIMPLE
 #endif
-#ifdef WITH_COMPLEX_SSE_KERNEL
+#ifdef WITH_COMPLEX_SSE_ASSEMBLY_KERNEL
  integer(kind=ik), parameter :: DEFAULT_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_SSE
 #endif
 #ifdef WITH_COMPLEX_AVX1_BLOCK1_KERNEL
@@ -321,7 +321,7 @@ module ELPA2_utilities
 #else
                                              ,0                       &
 #endif
-#if WITH_REAL_SSE_KERNEL
+#if WITH_REAL_SSE_ASSEMBLY_KERNEL
                                                ,1                     &
 #else
                                                ,0                     &
@@ -402,7 +402,7 @@ module ELPA2_utilities
 #else
                                              ,0                          &
 #endif
-#if WITH_COMPLEX_SSE_KERNEL
+#if WITH_COMPLEX_SSE_ASSEMBLY_KERNEL
                                                ,1                        &
 #else
                                                ,0                        &

--- a/src/mod_compute_hh_trafo_complex.F90
+++ b/src/mod_compute_hh_trafo_complex.F90
@@ -87,7 +87,7 @@ module compute_hh_trafo_complex
 #endif
           use iso_c_binding

-#if defined(HAVE_AVX) || defined(HAVE_SSE)
+#if defined(HAVE_AVX) || defined(HAVE_SSE_INTRINSICS) || defined(HAVE_SSE_ASSEMBLY)
         use kernel_interfaces
 #endif
           implicit none
@@ -271,7 +271,7 @@ module compute_hh_trafo_complex
 #endif /* WITH_NO_SPECIFIC_COMPLEX_KERNEL */
 #endif /* WITH_COMPLEX_GENERIC_KERNEL */

-#if defined(WITH_COMPLEX_SSE_KERNEL)
+#if defined(WITH_COMPLEX_SSE_ASSEMBLY_KERNEL)
 #if defined(WITH_NO_SPECIFIC_COMPLEX_KERNEL)
           if (THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_SSE) then
 #endif /* WITH_NO_SPECIFIC_COMPLEX_KERNEL */
@@ -288,7 +288,7 @@ module compute_hh_trafo_complex
 #if defined(WITH_NO_SPECIFIC_COMPLEX_KERNEL)
          endif
 #endif /* WITH_NO_SPECIFIC_COMPLEX_KERNEL */
-#endif /* WITH_COMPLEX_SSE_KERNEL */
+#endif /* WITH_COMPLEX_SSE_ASSEMBLY_KERNEL */


 !#if defined(WITH_AVX_SANDYBRIDGE)
@@ -386,7 +386,7 @@ module compute_hh_trafo_complex
           use timings
 #endif

-#if defined(HAVE_AVX) || defined(HAVE_SSE)
+#if defined(HAVE_AVX) || defined(HAVE_SSE_INTRINSICS) || defined(HAVE_SSE_ASSEMBLY)
           use kernel_interfaces
 #endif
           use iso_c_binding
@@ -569,7 +569,7 @@ module compute_hh_trafo_complex
 #endif /* WITH_NO_SPECIFIC_COMPLEX_KERNEL */
 #endif /* WITH_COMPLEX_GENERIC_KERNEL */

-#if defined(WITH_COMPLEX_SSE_KERNEL)
+#if defined(WITH_COMPLEX_SSE_ASSEMBLY_KERNEL)
 #if defined(WITH_NO_SPECIFIC_COMPLEX_KERNEL)
           if (THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_SSE) then
 #endif /* WITH_NO_SPECIFIC_COMPLEX_KERNEL */
@@ -586,7 +586,7 @@ module compute_hh_trafo_complex
 #if defined(WITH_NO_SPECIFIC_COMPLEX_KERNEL)
          endif
 #endif /* WITH_NO_SPECIFIC_COMPLEX_KERNEL */
-#endif /* WITH_COMPLEX_SSE_KERNEL */
+#endif /* WITH_COMPLEX_SSE_ASSEMBLY_KERNEL */


 !#if defined(WITH_AVX_SANDYBRIDGE)

--- a/src/mod_compute_hh_trafo_real.F90
+++ b/src/mod_compute_hh_trafo_real.F90
@@ -104,7 +104,7 @@ module compute_hh_trafo_real
         use timings
 #endif

-#if defined(HAVE_AVX) || defined(HAVE_SSE)
+#if defined(HAVE_AVX) || defined(HAVE_SSE_INTRINSICS) || defined(HAVE_SSE_ASSEMBLY)
         use kernel_interfaces
 #endif
         implicit none
@@ -275,7 +275,7 @@ module compute_hh_trafo_real
 #endif /* WITH_REAL_GENERIC_SIMPLE_KERNEL */


-#if defined(WITH_REAL_SSE_KERNEL)
+#if defined(WITH_REAL_SSE_ASSEMBLY_KERNEL)
 #if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
             if (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_SSE) then
 #endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
@@ -293,7 +293,7 @@ module compute_hh_trafo_real
 #if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
             endif
 #endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
-#endif /* WITH_REAL_SSE_KERNEL */
+#endif /* WITH_REAL_SSE_ASSEMBLY_KERNEL */

 #if defined(WITH_REAL_SSE_BLOCK2_KERNEL)
 #if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
@@ -661,7 +661,7 @@ module compute_hh_trafo_real
         use timings
 #endif

-#if defined(HAVE_AVX) || defined(HAVE_SSE)
+#if defined(HAVE_AVX) || defined(HAVE_SSE_INTRINSICS) || defined(HAVE_SSE_ASSEMBLY)
         use kernel_interfaces
 #endif
         implicit none
@@ -831,7 +831,7 @@ module compute_hh_trafo_real
 #endif /* WITH_REAL_GENERIC_SIMPLE_KERNEL */


-#if defined(WITH_REAL_SSE_KERNEL)
+#if defined(WITH_REAL_SSE_ASSEMBLY_KERNEL)
 #if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
             if (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_SSE) then
 #endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
@@ -849,7 +849,7 @@ module compute_hh_trafo_real
 #if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
             endif
 #endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
-#endif /* WITH_REAL_SSE_KERNEL */
+#endif /* WITH_REAL_SSE_ASSEMBLY_KERNEL */

 #if defined(WITH_REAL_SSE_BLOCK2_KERNEL)
 #if defined(WITH_NO_SPECIFIC_REAL_KERNEL)