Error in setting specific kernel at compile time

In case of SSE/AVX/AVX2 it could happen that more than one kernel
(since some depend on other kernels, e.g. block 6 on block 4 and
block 2) were called

configure.ac

@@ -800,10 +800,14 @@ fi
 
 if test x"${use_specific_complex_kernel}" = x"no" ; then
  AC_DEFINE([WITH_NO_SPECIFIC_COMPLEX_KERNEL],[1],[do not use only one specific complex kernel (set at compile time)])
+else
+ AC_DEFINE([WITH_ONE_SPECIFIC_COMPLEX_KERNEL],[1],[use only one specific complex kernel (set at compile time)])
 fi
 
 if test x"${use_specific_real_kernel}" = x"no" ; then
   AC_DEFINE([WITH_NO_SPECIFIC_REAL_KERNEL],[1],[do not use only one specific real kernel (set at compile time)])
+else
+  AC_DEFINE([WITH_ONE_SPECIFIC_REAL_KERNEL],[1],[use only one specific real kernel (set at compile time)])
 fi
 
 LT_INIT

src/elpa2.F90

@@ -214,7 +214,7 @@ function solve_evp_real_2stage(na, nev, a, lda, ev, q, ldq, nblk,        &
      THIS_REAL_ELPA_KERNEL = get_actual_real_kernel()
    endif
 
-   ! check whether choosen kernel is allowed
+   ! check whether choosen kernel is allowed: function returns true if NOT allowed! change this
    if (check_allowed_real_kernels(THIS_REAL_ELPA_KERNEL)) then
 
      if (my_pe == 0) then
@@ -230,10 +230,18 @@ function solve_evp_real_2stage(na, nev, a, lda, ev, q, ldq, nblk,        &
        enddo
 
        write(error_unit,*) " "
-       write(error_unit,*) "The defaul kernel REAL_ELPA_KERNEL_GENERIC will be used !"
+       ! check whether generic kernel is defined
+       if (AVAILABLE_REAL_ELPA_KERNELS(REAL_ELPA_KERNEL_GENERIC) .eq. 1) then
+         write(error_unit,*) "The default kernel REAL_ELPA_KERNEL_GENERIC will be used !"
+       else
+         write(error_unit,*) "As default kernel ",REAL_ELPA_KERNEL_NAMES(DEFAULT_REAL_ELPA_KERNEL)," will be used"
+       endif
+     endif  ! my_pe == 0
+     if (AVAILABLE_REAL_ELPA_KERNELS(REAL_ELPA_KERNEL_GENERIC) .eq. 1) then
+       THIS_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_GENERIC
+     else
+       THIS_REAL_ELPA_KERNEL = DEFAULT_REAL_ELPA_KERNEL
      endif
-     THIS_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_GENERIC
-
    endif
 
    ! Choose bandwidth, must be a multiple of nblk, set to a value >= 32
@@ -433,9 +441,18 @@ function solve_evp_complex_2stage(na, nev, a, lda, ev, q, ldq, nblk, &
        enddo
 
        write(error_unit,*) " "
-       write(error_unit,*) "The defaul kernel COMPLEX_ELPA_KERNEL_GENERIC will be used !"
+       ! check whether generic kernel is defined
+       if (AVAILABLE_COMPLEX_ELPA_KERNELS(COMPLEX_ELPA_KERNEL_GENERIC) .eq. 1) then
+         write(error_unit,*) "The default kernel COMPLEX_ELPA_KERNEL_GENERIC will be used !"
+       else
+         write(error_unit,*) "As default kernel ",COMPLEX_ELPA_KERNEL_NAMES(DEFAULT_COMPLEX_ELPA_KERNEL)," will be used"
+       endif
+     endif  ! my_pe == 0
+     if (AVAILABLE_COMPLEX_ELPA_KERNELS(COMPLEX_ELPA_KERNEL_GENERIC) .eq. 1) then
+       THIS_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_GENERIC
+     else
+       THIS_COMPLEX_ELPA_KERNEL = DEFAULT_COMPLEX_ELPA_KERNEL
      endif
-     THIS_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_GENERIC
    endif
    ! Choose bandwidth, must be a multiple of nblk, set to a value >= 32
 

src/elpa2_utilities.F90

@@ -76,14 +76,16 @@ module ELPA2_utilities
             REAL_ELPA_KERNEL_AVX_BLOCK2,                               &
             REAL_ELPA_KERNEL_AVX_BLOCK4, REAL_ELPA_KERNEL_AVX_BLOCK6,  &
             REAL_ELPA_KERNEL_AVX2_BLOCK2,                              &
-            REAL_ELPA_KERNEL_AVX2_BLOCK4, REAL_ELPA_KERNEL_AVX2_BLOCK6
+            REAL_ELPA_KERNEL_AVX2_BLOCK4, REAL_ELPA_KERNEL_AVX2_BLOCK6,&
+            DEFAULT_REAL_ELPA_KERNEL
 
   public :: COMPLEX_ELPA_KERNEL_GENERIC, COMPLEX_ELPA_KERNEL_GENERIC_SIMPLE, &
             COMPLEX_ELPA_KERNEL_BGP, COMPLEX_ELPA_KERNEL_BGQ,                &
             COMPLEX_ELPA_KERNEL_SSE, COMPLEX_ELPA_KERNEL_SSE_BLOCK1,         &
             COMPLEX_ELPA_KERNEL_SSE_BLOCK2,                                  &
             COMPLEX_ELPA_KERNEL_AVX_BLOCK1,COMPLEX_ELPA_KERNEL_AVX_BLOCK2,   &
-            COMPLEX_ELPA_KERNEL_AVX2_BLOCK1,COMPLEX_ELPA_KERNEL_AVX2_BLOCK2
+            COMPLEX_ELPA_KERNEL_AVX2_BLOCK1,COMPLEX_ELPA_KERNEL_AVX2_BLOCK2, &
+            DEFAULT_COMPLEX_ELPA_KERNEL
 
   public :: REAL_ELPA_KERNEL_NAMES, COMPLEX_ELPA_KERNEL_NAMES
 
@@ -115,10 +117,114 @@ module ELPA2_utilities
   integer, parameter :: REAL_ELPA_KERNEL_AVX2_BLOCK6     = ELPA2_REAL_KERNEL_AVX2_BLOCK6
 
 #if defined(WITH_REAL_AVX_BLOCK2_KERNEL)
+
+#ifndef WITH_ONE_SPECIFIC_REAL_KERNEL
+  integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_GENERIC
+#else /* WITH_ONE_SPECIFIC_REAL_KERNEL */
+
+#ifdef WITH_REAL_GENERIC_KERNEL
   integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_GENERIC
+#endif
+#ifdef WITH_REAL_GENERIC_SIMPLE_KERNEL
+  integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_GENERIC_SIMPLE
+#endif
+#ifdef WITH_REAL_SSE_ASSEMBLY_KERNEL
+  integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_SSE
+#endif
+#if defined(WITH_REAL_SSE_BLOCK2_KERNEL) || defined(WITH_REAL_SSE_BLOCK4_KERNEL) || defined(WITH_REAL_SSE_BLOCK6_KERNEL)
+
+#ifdef WITH_REAL_SSE_BLOCK6_KERNEL
+  integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_SSE_BLOCK6
+#else
+
+#ifdef WITH_REAL_SSE_BLOCK4_KERNEL
+  integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_SSE_BLOCK4
+#else
+#ifdef WITH_REAL_SSE_BLOCK2_KERNEL
+  integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_SSE_BLOCK2
+#endif
+#endif
+#endif
+#endif /*  #if defined(WITH_REAL_SSE_BLOCK2_KERNEL) || defined(WITH_REAL_SSE_BLOCK4_KERNEL) || defined(WITH_REAL_SSE_BLOCK6_KERNEL) */
+
+#if defined(WITH_REAL_AVX_BLOCK2_KERNEL) || defined(WITH_REAL_AVX_BLOCK4_KERNEL) || defined(WITH_REAL_AVX_BLOCK6_KERNEL)
+#ifdef WITH_REAL_AVX_BLOCK6_KERNEL
+  integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_AVX_BLOCK6
+#else
+#ifdef WITH_REAL_AVX_BLOCK4_KERNEL
+  integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_AVX_BLOCK4
 #else
+#ifdef WITH_REAL_AVX_BLOCK2_KERNEL
+  integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_AVX_BLOCK2
+#endif
+#endif
+#endif
+#endif /*  #if defined(WITH_REAL_AVX_BLOCK2_KERNEL) || defined(WITH_REAL_AVX_BLOCK4_KERNEL) || defined(WITH_REAL_AVX_BLOCK6_KERNEL) */
+
+#ifdef WITH_REAL_BGP_KERNEL
+  integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_AVX_BGP
+#endif
+#ifdef WITH_REAL_BGQ_KERNEL
+  integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_AVX_BGQ
+#endif
+
+#endif /* WITH_ONE_SPECIFIC_REAL_KERNEL */
+
+#else /* WITH_REAL_AVX_BLOCK2_KERNEL */
+
+#ifndef WITH_ONE_SPECIFIC_REAL_KERNEL
+  integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_GENERIC
+#else /* WITH_ONE_SPECIFIC_REAL_KERNEL */
+
+#ifdef WITH_REAL_GENERIC_KERNEL
   integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_GENERIC
 #endif
+#ifdef WITH_REAL_GENERIC_SIMPLE_KERNEL
+  integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_GENERIC_SIMPLE
+#endif
+#ifdef WITH_REAL_SSE_ASSEMBLY_KERNEL
+  integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_SSE
+#endif
+
+#if defined(WITH_REAL_SSE_BLOCK2_KERNEL) || defined(WITH_REAL_SSE_BLOCK4_KERNEL) || defined(WITH_REAL_SSE_BLOCK6_KERNEL)
+#ifdef WITH_REAL_SSE_BLOCK6_KERNEL
+  integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_SSE_BLOCK6
+#else
+#ifdef WITH_REAL_SSE_BLOCK4_KERNEL
+  integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_SSE_BLOCK4
+#else
+#ifdef WITH_REAL_SSE_BLOCK2_KERNEL
+  integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_SSE_BLOCK2
+#endif
+#endif
+#endif
+#endif /*  #if defined(WITH_REAL_SSE_BLOCK2_KERNEL) || defined(WITH_REAL_SSE_BLOCK4_KERNEL) || defined(WITH_REAL_SSE_BLOCK6_KERNEL) */
+
+#if defined(WITH_REAL_AVX_BLOCK2_KERNEL) || defined(WITH_REAL_AVX_BLOCK4_KERNEL) || defined(WITH_REAL_AVX_BLOCK6_KERNEL)
+#ifdef WITH_REAL_AVX_BLOCK6_KERNEL
+  integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_AVX_BLOCK6
+#else
+#ifdef WITH_REAL_AVX_BLOCK4_KERNEL
+  integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_AVX_BLOCK4
+#else
+#ifdef WITH_REAL_AVX_BLOCK2_KERNEL
+  integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_AVX_BLOCK2
+#endif
+#endif
+#endif
+#endif /*  #if defined(WITH_REAL_AVX_BLOCK2_KERNEL) || defined(WITH_REAL_AVX_BLOCK4_KERNEL) || defined(WITH_REAL_AVX_BLOCK6_KERNEL) */
+
+#ifdef WITH_REAL_BGP_KERNEL
+  integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_AVX_BGP
+#endif
+#ifdef WITH_REAL_BGQ_KERNEL
+  integer, parameter :: DEFAULT_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_AVX_BGQ
+#endif
+
+#endif  /* WITH_ONE_SPECIFIC_REAL_KERNEL */
+
+#endif /* WITH_REAL_AVX_BLOCK2_KERNEL */
+
   character(35), parameter, dimension(number_of_real_kernels) :: &
   REAL_ELPA_KERNEL_NAMES =    (/"REAL_ELPA_KERNEL_GENERIC         ", &
                                 "REAL_ELPA_KERNEL_GENERIC_SIMPLE  ", &
@@ -149,10 +255,86 @@ module ELPA2_utilities
   integer, parameter :: COMPLEX_ELPA_KERNEL_AVX2_BLOCK2     = ELPA2_COMPLEX_KERNEL_AVX2_BLOCK2
 
 #if defined(WITH_COMPLEX_AVX_BLOCK1_KERNEL)
+
+#ifndef WITH_ONE_SPECIFIC_COMPLEX_KERNEL
+  integer, parameter :: DEFAULT_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_GENERIC
+#else /* WITH_ONE_SPECIFIC_COMPLEX_KERNEL */
+
+! go through all kernels and set them
+#ifdef WITH_COMPLEX_GENERIC_KERNEL
   integer, parameter :: DEFAULT_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_GENERIC
+#endif
+#ifdef WITH_COMPLEX_GENERIC_SIMPLE_KERNEL
+  integer, parameter :: DEFAULT_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_GENERIC_SIMPLE
+#endif
+#ifdef WITH_COMPLEX_SSE_ASSEMBLY_KERNEL
+  integer, parameter :: DEFAULT_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_SSE
+#endif
+
+#if defined(WITH_COMPLEX_SSE_BLOCK1_KERNEL) || defined(WITH_COMPLEX_SSE_BLOCK2_KERNEL)
+#ifdef WITH_COMPLEX_SSE_BLOCK2_KERNEL
+  integer, parameter :: DEFAULT_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_SSE_BLOCK2
+#else
+#ifdef WITH_COMPLEX_SSE_BLOCK1_KERNEL
+  integer, parameter :: DEFAULT_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_SSE_BLOCK1
+#endif
+#endif
+#endif /* defined(WITH_COMPLEXL_SSE_BLOCK1_KERNEL) || defined(WITH_COMPLEX_SSE_BLOCK2_KERNEL) */
+
+#if defined(WITH_COMPLEX_AVX_BLOCK1_KERNEL) || defined(WITH_COMPLEX_AVX_BLOCK2_KERNEL)
+#ifdef WITH_COMPLEX_AVX_BLOCK2_KERNEL
+  integer, parameter :: DEFAULT_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_AVX_BLOCK2
 #else
+#ifdef WITH_COMPLEX_AVX_BLOCK1_KERNEL
+  integer, parameter :: DEFAULT_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_AVX_BLOCK1
+#endif
+#endif
+#endif /* defined(WITH_COMPLEX_AVX_BLOCK1_KERNEL) || defined(WITH_COMPLEX_AVX_BLOCK2_KERNEL) */
+
+#endif /* WITH_ONE_SPECIFIC_COMPLEX_KERNEL */
+
+#else /* WITH_COMPLEX_AVX_BLOCK1_KERNEL */
+
+#ifndef WITH_ONE_SPECIFIC_COMPLEX_KERNEL
+  integer, parameter :: DEFAULT_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_GENERIC
+
+#else /* WITH_ONE_SPECIFIC_COMPLEX_KERNEL */
+
+! go through all kernels and set them
+#ifdef WITH_COMPLEX_GENERIC_KERNEL
   integer, parameter :: DEFAULT_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_GENERIC
 #endif
+#ifdef WITH_COMPLEX_GENERIC_SIMPLE_KERNEL
+  integer, parameter :: DEFAULT_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_GENERIC_SIMPLE
+#endif
+#ifdef WITH_COMPLEX_SSE_ASSEMBLY_KERNEL
+  integer, parameter :: DEFAULT_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_SSE
+#endif
+
+#if defined(WITH_COMPLEX_SSE_BLOCK1_KERNEL) || defined(WITH_COMPLEX_SSE_BLOCK2_KERNEL)
+#ifdef WITH_COMPLEX_SSE_BLOCK2_KERNEL
+  integer, parameter :: DEFAULT_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_SSE_BLOCK2
+#else
+#ifdef WITH_COMPLEX_SSE_BLOCK1_KERNEL
+  integer, parameter :: DEFAULT_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_SSE_BLOCK1
+#endif
+#endif
+#endif /* defined(WITH_COMPLEXL_SSE_BLOCK1_KERNEL) || defined(WITH_COMPLEX_SSE_BLOCK2_KERNEL) */
+
+#if defined(WITH_COMPLEX_AVX_BLOCK1_KERNEL) || defined(WITH_COMPLEX_AVX_BLOCK2_KERNEL)
+#ifdef WITH_COMPLEX_AVX_BLOCK2_KERNEL
+  integer, parameter :: DEFAULT_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_AVX_BLOCK2
+#else
+#ifdef WITH_COMPLEX_AVX_BLOCK1_KERNEL
+  integer, parameter :: DEFAULT_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_AVX_BLOCK1
+#endif
+#endif
+#endif /* defined(WITH_COMPLEX_AVX_BLOCK1_KERNEL) || defined(WITH_COMPLEX_AVX_BLOCK2_KERNEL) */
+
+#endif /* WITH_ONE_SPECIFIC_COMPLEX_KERNEL */
+
+#endif /* WITH_COMPLEX_AVX_BLOCK1_KERNEL */
+
   character(35), parameter, dimension(number_of_complex_kernels) :: &
   COMPLEX_ELPA_KERNEL_NAMES = (/"COMPLEX_ELPA_KERNEL_GENERIC         ", &
                                 "COMPLEX_ELPA_KERNEL_GENERIC_SIMPLE  ", &

src/mod_compute_hh_trafo_complex.F90

@@ -250,6 +250,8 @@ module compute_hh_trafo_complex
 #if defined(WITH_NO_SPECIFIC_COMPLEX_KERNEL)
           if (THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_SSE_BLOCK1) then
 #endif /* WITH_NO_SPECIFIC_COMPLEX_KERNEL */
+
+#if defined(WITH_NO_SPECIFIC_COMPLEX_KERNEL) || (defined(WITH_ONE_SPECIFIC_COMPLEX_KERNEL) && !defined(COMPLEX_ELPA_KERNEL_SSE_BLOCK2))
             ttt = mpi_wtime()
             do j = ncols, 1, -1
 #ifdef WITH_OPENMP
@@ -260,6 +262,8 @@ module compute_hh_trafo_complex
                                                        bcast_buffer(1,j+off),nbw,nl,stripe_width)
 #endif
             enddo
+#endif /* defined(WITH_NO_SPECIFIC_COMPLEX_KERNEL) || (defined(WITH_ONE_SPECIFIC_COMPLEX_KERNEL) && !defined(COMPLEX_ELPA_KERNEL_SSE_BLOCK2)) */
+
 #if defined(WITH_NO_SPECIFIC_COMPLEX_KERNEL)
           endif
 #endif /* WITH_NO_SPECIFIC_COMPLEX_KERNEL */
@@ -270,6 +274,8 @@ module compute_hh_trafo_complex
           if ((THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_AVX_BLOCK1) .or. &
               (THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_AVX2_BLOCK1)) then
 #endif /* WITH_NO_SPECIFIC_COMPLEX_KERNEL */
+
+#if defined(WITH_NO_SPECIFIC_COMPLEX_KERNEL) || (defined(WITH_ONE_SPECIFIC_COMPLEX_KERNEL) && !defined(COMPLEX_ELPA_KERNEL_AVX_BLOCK2) && !defined(COMPLEX_ELPA_KERNEL_AVX2_BLOCK2))
             ttt = mpi_wtime()
             do j = ncols, 1, -1
 #ifdef WITH_OPENMP
@@ -280,6 +286,8 @@ module compute_hh_trafo_complex
                                                        bcast_buffer(1,j+off),nbw,nl,stripe_width)
 #endif
             enddo
+#endif /* defined(WITH_NO_SPECIFIC_COMPLEX_KERNEL) || (defined(WITH_ONE_SPECIFIC_COMPLEX_KERNEL) && !defined(COMPLEX_ELPA_KERNEL_AVX_BLOCK2) && !defined(COMPLEX_ELPA_KERNEL_AVX2_BLOCK2)) */
+
 #if defined(WITH_NO_SPECIFIC_COMPLEX_KERNEL)
           endif
 #endif /* WITH_NO_SPECIFIC_COMPLEX_KERNEL */

src/mod_compute_hh_trafo_real.F90

@@ -218,6 +218,8 @@ module compute_hh_trafo_real
 #if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
            if (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_SSE_BLOCK2) then
 #endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
+
+#if defined(WITH_NO_SPECIFIC_REAL_KERNEL) || (defined(WITH_ONE_SPECIFIC_REAL_KERNEL) && !defined(WITH_REAL_SSE_BLOCK6_KERNEL) && !defined(WITH_REAL_SSE_BLOCK4_KERNEL))
              do j = ncols, 2, -2
                w(:,1) = bcast_buffer(1:nbw,j+off)
                w(:,2) = bcast_buffer(1:nbw,j+off-1)
@@ -229,6 +231,8 @@ module compute_hh_trafo_real
                                                        w, nbw, nl, stripe_width, nbw)
 #endif
              enddo
+#endif /* defined(WITH_NO_SPECIFIC_REAL_KERNEL) || (defined(WITH_ONE_SPECIFIC_REAL_KERNEL) && !defined(WITH_REAL_SSE_BLOCK6_KERNEL) && !defined(WITH_REAL_SSE_BLOCK4_KERNEL)) */
+
 #if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
            endif
 #endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
@@ -239,6 +243,8 @@ module compute_hh_trafo_real
            if ((THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX_BLOCK2) .or. &
                (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX2_BLOCK2))  then
 #endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
+
+#if defined(WITH_NO_SPECIFIC_REAL_KERNEL) || (defined(WITH_ONE_SPECIFIC_REAL_KERNEL) && !defined(WITH_REAL_AVX_BLOCK6_KERNEL) && !defined(WITH_REAL_AVX_BLOCK4_KERNEL) && !defined(WITH_REAL_AVX2_BLOCK6_KERNEL) && !defined(WITH_REAL_AVX2_BLOCK4_KERNEL))
              do j = ncols, 2, -2
                w(:,1) = bcast_buffer(1:nbw,j+off)
                w(:,2) = bcast_buffer(1:nbw,j+off-1)
@@ -250,6 +256,8 @@ module compute_hh_trafo_real
                                                        w, nbw, nl, stripe_width, nbw)
 #endif
              enddo
+#endif /* defined(WITH_NO_SPECIFIC_REAL_KERNEL) || (defined(WITH_ONE_SPECIFIC_REAL_KERNEL) ... */
+
 #if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
            endif
 #endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
@@ -322,6 +330,8 @@ module compute_hh_trafo_real
 #if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
          if (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_SSE_BLOCK4) then
 #endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
+
+#if defined(WITH_NO_SPECIFIC_REAL_KERNEL) || (defined(WITH_ONE_SPECIFIC_REAL_KERNEL) && !defined(WITH_REAL_SSE_BLOCK6_KERNEL))
            ! X86 INTRINSIC CODE, USING 4 HOUSEHOLDER VECTORS
            do j = ncols, 4, -4
              w(:,1) = bcast_buffer(1:nbw,j+off)
@@ -354,6 +364,9 @@ module compute_hh_trafo_real
            if (jj==1) call single_hh_trafo_real_cpu(a(1:stripe_width,1+off+a_off:1+off+a_off+nbw-1,istripe), &
                                           bcast_buffer(1:nbw,off+1), nbw, nl, stripe_width)
 #endif
+
+#endif /* defined(WITH_NO_SPECIFIC_REAL_KERNEL) || (defined(WITH_ONE_SPECIFIC_REAL_KERNEL) && !defined(WITH_REAL_SSE_BLOCK6_KERNEL)) */
+
 #if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
          endif
 #endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
@@ -364,6 +377,8 @@ module compute_hh_trafo_real
          if ((THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX_BLOCK4) .or. &
              (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX2_BLOCK4)) then
 #endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
+
+#if defined(WITH_NO_SPECIFIC_REAL_KERNEL) || (defined(WITH_ONE_SPECIFIC_REAL_KERNEL) && !defined(WITH_REAL_AVX_BLOCK6_KERNEL) && !defined(WITH_REAL_AVX2_BLOCK6_KERNEL))
            ! X86 INTRINSIC CODE, USING 4 HOUSEHOLDER VECTORS
            do j = ncols, 4, -4
              w(:,1) = bcast_buffer(1:nbw,j+off)
@@ -396,6 +411,9 @@ module compute_hh_trafo_real
            if (jj==1) call single_hh_trafo_real_cpu(a(1:stripe_width,1+off+a_off:1+off+a_off+nbw-1,istripe), &
                                           bcast_buffer(1:nbw,off+1), nbw, nl, stripe_width)
 #endif
+
+#endif /* defined(WITH_NO_SPECIFIC_REAL_KERNEL) || (defined(WITH_ONE_SPECIFIC_REAL_KERNEL) && !defined(WITH_REAL_AVX_BLOCK6_KERNEL) && !defined(WITH_REAL_AVX2_BLOCK6_KERNEL)) */
+
 #if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
          endif
 #endif /* WITH_NO_SPECIFIC_REAL_KERNEL */

test/fortran_test_programs/elpa_test_programs_print_headers.X90

@@ -129,7 +129,7 @@
 #ifdef WITH_REAL_GENERIC_SIMPLE_KERNEL
      print *,"GENERIC SIMPLE kernel for real matrices"
 #endif
-#ifdef WITH_REAL_SSE_KERNEL
+#ifdef WITH_REAL_SSE_ASSEMBLY_KERNEL
      print *,"SSE ASSEMBLER kernel for real matrices"
 #endif
 #ifdef WITH_REAL_BGP_KERNEL
@@ -174,7 +174,7 @@
 #ifdef WITH_COMPLEX_GENERIC_SIMPLE_KERNEL
      print *,"GENERIC SIMPLE kernel for complex matrices"
 #endif
-#ifdef WITH_COMPLEX_SSE_KERNEL
+#ifdef WITH_COMPLEX_SSE_ASSEMBLY_KERNEL
      print *,"SSE ASSEMBLER kernel for complex matrices"
 #endif
 

test/fortran_test_programs/test_complex2_choose_kernel_with_api.F90

@@ -287,8 +287,61 @@ program test_complex2
    success = solve_evp_complex_2stage(na, nev, a, na_rows, ev, z, na_rows, nblk, &
                                  na_cols, &
                                  mpi_comm_rows, mpi_comm_cols, mpi_comm_world, &
+#ifndef WITH_ONE_SPECIFIC_COMPLEX_KERNEL
                                  COMPLEX_ELPA_KERNEL_GENERIC_SIMPLE)
+#else /* WITH_ONE_SPECIFIC_COMPLEX_KERNEL */
 
+#ifdef  WITH_COMPLEX_GENERIC_KERNEL
+                                 COMPLEX_ELPA_KERNEL_GENERIC)
+#endif
+
+#ifdef  WITH_COMPLEX_GENERIC_SIMPLE_KERNEL
+                                 COMPLEX_ELPA_KERNEL_GENERIC_SIMPLE)
+#endif
+
+#ifdef  WITH_COMPLEX_SSE_ASSEMBLY_KERNEL
+                                 COMPLEX_ELPA_KERNEL_SSE)
+#endif
+
+#ifdef WITH_ONE_SPECIFIC_COMPLEX_KERNEL
+
+#ifdef  WITH_COMPLEX_SSE_BLOCK2_KERNEL
+                                 COMPLEX_ELPA_KERNEL_SSE_BLOCK2)
+#else
+#ifdef  WITH_COMPLEX_SSE_BLOCK1_KERNEL
+                                 COMPLEX_ELPA_KERNEL_SSE_BLOCK1)
+#endif
+#endif
+
+#ifdef  WITH_COMPLEX_AVX_BLOCK2_KERNEL
+                                 COMPLEX_ELPA_KERNEL_AVX_BLOCK2)
+#else
+#ifdef  WITH_COMPLEX_AVX_BLOCK1_KERNEL
+                                 COMPLEX_ELPA_KERNEL_AVX_BLOCK1)
+#endif
+#endif
+
+#else /* WITH_ONE_SPECIFIC_COMPLEX_KERNEL */
+
+#ifdef  WITH_COMPLEX_SSE_BLOCK1_KERNEL
+                                 COMPLEX_ELPA_KERNEL_SSE_BLOCK1)
+#endif
+
+#ifdef  WITH_COMPLEX_SSE_BLOCK2_KERNEL
+                                 COMPLEX_ELPA_KERNEL_SSE_BLOCK2)
+#endif
+
+#ifdef  WITH_COMPLEX_AVX_BLOCK1_KERNEL
+                                 COMPLEX_ELPA_KERNEL_AVX_BLOCK1)
+#endif
+
+#ifdef  WITH_COMPLEX_AVX_BLOCK2_KERNEL
+                                 COMPLEX_ELPA_KERNEL_AVX_BLOCK2)
+#endif
+
+#endif  /*   WITH_ONE_SPECIFIC_COMPLEX_KERNEL */
+
+#endif /* WITH_ONE_SPECIFIC_COMPLEX_KERNEL */
 
    if (.not.(success)) then
       write(error_unit,*) "solve_evp_complex_2stage produced an error! Aborting..."

test/fortran_test_programs/test_real2_choose_kernel_with_api.F90

@@ -278,9 +278,85 @@ program test_real2
    success = solve_evp_real_2stage(na, nev, a, na_rows, ev, z, na_rows, nblk, &
                               na_cols, &
                               mpi_comm_rows, mpi_comm_cols, mpi_comm_world, &
+#ifndef WITH_ONE_SPECIFIC_REAL_KERNEL
+                             REAL_ELPA_KERNEL_GENERIC_SIMPLE)
+#else /* WITH_ONE_SPECIFIC_COMPLEX_KERNEL */
+
+#ifdef WITH_REAL_GENERIC_KERNEL
+                              REAL_ELPA_KERNEL_GENERIC)
+#endif
+
+#ifdef WITH_REAL_GENERIC_SIMPLE_KERNEL
                               REAL_ELPA_KERNEL_GENERIC_SIMPLE)
+#endif
+
+#ifdef WITH_REAL_SSE_ASSEMBLY_KERNEL
+                              REAL_ELPA_KERNEL_SSE)
+#endif
+#ifdef WITH_ONE_SPECIFIC_REAL_KERNEL
+
+#ifdef WITH_REAL_SSE_BLOCK6_KERNEL
+                              REAL_ELPA_KERNEL_SSE_BLOCK6)
+#else
+#ifdef WITH_REAL_SSE_BLOCK4_KERNEL
+                              REAL_ELPA_KERNEL_SSE_BLOCK4)
+#else
+#ifdef WITH_REAL_SSE_BLOCK2_KERNEL
+                              REAL_ELPA_KERNEL_SSE_BLOCK2)
+#endif
+#endif
+#endif
+#ifdef WITH_REAL_AVX_BLOCK6_KERNEL
+                              REAL_ELPA_KERNEL_AVX_BLOCK6)
+#else
+#ifdef WITH_REAL_AVX_BLOCK4_KERNEL
+                              REAL_ELPA_KERNEL_AVX_BLOCK4)
+#else
+#ifdef WITH_REAL_AVX_BLOCK2_KERNEL
+                              REAL_ELPA_KERNEL_AVX_BLOCK2)
+#endif
+#endif
+#endif
+
+#else /* WITH_ONE_SPECIFIC_REAL_KERNEL */
+
+#ifdef WITH_REAL_SSE_BLOCK2_KERNEL
+                              REAL_ELPA_KERNEL_SSE_BLOCK2)
+#endif
+
+#ifdef WITH_REAL_SSE_BLOCK4_KERNEL
+                              REAL_ELPA_KERNEL_SSE_BLOCK4)
+#endif
+
+#ifdef WITH_REAL_SSE_BLOCK6_KERNEL
+                              REAL_ELPA_KERNEL_SSE_BLOCK6)
+#endif
+
+#ifdef WITH_REAL_AVX_BLOCK2_KERNEL
+                              REAL_ELPA_KERNEL_AVX_BLOCK2)
+#endif
+
+#ifdef WITH_REAL_AVX_BLOCK4_KERNEL
+                              REAL_ELPA_KERNEL_AVX_BLOCK4)
+#endif
+
+#ifdef WITH_REAL_AVX_BLOCK6_KERNEL
+                              REAL_ELPA_KERNEL_AVX_BLOCK6)
+#endif
+
+#endif /* WITH_ONE_SPECIFIC_REAL_KERNEL */
+
+#ifdef WITH_REAL_BGP_KERNEL
+                              REAL_ELPA_KERNEL_BGP)
+#endif
+
+#ifdef WITH_REAL_BGQ_KERNEL
+                              REAL_ELPA_KERNEL_BGQ)
+#endif
+
+#endif /* WITH_ONE_SPECIFIC_REAL_KERNEL */
 
-   if (.not.(success)) then
+    if (.not.(success)) then
       write(error_unit,*) "solve_evp_real_2stage produced an error! Aborting..."
 #ifdef WITH_MPI
       call MPI_ABORT(mpi_comm_world, 1, mpierr)