Make double precision real Power8 kernels available

Makefile.am

@@ -687,6 +687,8 @@ EXTRA_DIST = \
   src/elpa2/kernels/real_avx512_4hv_template.c \
   src/elpa2/kernels/real_avx512_6hv_template.c \
   src/elpa2/kernels/real_vsx_2hv_template.c \
+  src/elpa2/kernels/real_vsx_4hv_template.c \
+  src/elpa2/kernels/real_vsx_6hv_template.c \
   src/elpa2/kernels/real_sse_2hv_template.c \
   src/elpa2/kernels/real_sse_4hv_template.c \
   src/elpa2/kernels/real_sse_6hv_template.c \

configure.ac

@@ -478,8 +478,6 @@ m4_define(elpa_m4_vsx_kernels, [
         real_vsx_block2
         real_vsx_block4
         real_vsx_block6
-        complex_vsx_block1
-        complex_vsx_block2
 ])
 
 m4_define(elpa_m4_avx_kernels, [

elpa/elpa_constants.h.in

@@ -74,9 +74,7 @@ enum ELPA_REAL_KERNELS {
         X(ELPA_2STAGE_COMPLEX_AVX512_BLOCK2, 13, @ELPA_2STAGE_COMPLEX_AVX512_BLOCK2_COMPILED@, __VA_ARGS__) \
         X(ELPA_2STAGE_COMPLEX_GPU, 14, @ELPA_2STAGE_COMPLEX_GPU_COMPILED@, __VA_ARGS__) \
         X(ELPA_2STAGE_COMPLEX_SPARC64_BLOCK1, 15, @ELPA_2STAGE_COMPLEX_SPARC64_BLOCK1_COMPILED@, __VA_ARGS__) \
-        X(ELPA_2STAGE_COMPLEX_SPARC64_BLOCK2, 16, @ELPA_2STAGE_COMPLEX_SPARC64_BLOCK2_COMPILED@, __VA_ARGS__) \
-        X(ELPA_2STAGE_COMPLEX_VSX_BLOCK1, 17, @ELPA_2STAGE_COMPLEX_VSX_BLOCK1_COMPILED@, __VA_ARGS__) \
-        X(ELPA_2STAGE_COMPLEX_VSX_BLOCK2, 18, @ELPA_2STAGE_COMPLEX_VSX_BLOCK2_COMPILED@, __VA_ARGS__)
+        X(ELPA_2STAGE_COMPLEX_SPARC64_BLOCK2, 16, @ELPA_2STAGE_COMPLEX_SPARC64_BLOCK2_COMPILED@, __VA_ARGS__)
 
 #define ELPA_FOR_ALL_2STAGE_COMPLEX_KERNELS_AND_DEFAULT(X) \
         ELPA_FOR_ALL_2STAGE_COMPLEX_KERNELS(X) \

src/elpa2/compute_hh_trafo.F90

@@ -663,13 +663,13 @@
 
 ! vsx block1 complex kernel
 #if defined(WITH_COMPLEX_VSX_BLOCK1_KERNEL)
-#ifndef WITH_FIXED_COMPLEX_KERNEL
-          if (kernel .eq. ELPA_2STAGE_COMPLEX_VSX_BLOCK1) then
-#endif /* not WITH_FIXED_COMPLEX_KERNEL */
-
-#if (!defined(WITH_FIXED_COMPLEX_KERNEL)) || (defined(WITH_FIXED_COMPLEX_KERNEL) && !defined(WITH_COMPLEX_VSX_BLOCK2_KERNEL))
-            ttt = mpi_wtime()
-            do j = ncols, 1, -1
+!#ifndef WITH_FIXED_COMPLEX_KERNEL
+!          if (kernel .eq. ELPA_2STAGE_COMPLEX_VSX_BLOCK1) then
+!#endif /* not WITH_FIXED_COMPLEX_KERNEL */
+!
+!#if (!defined(WITH_FIXED_COMPLEX_KERNEL)) || (defined(WITH_FIXED_COMPLEX_KERNEL) && !defined(WITH_COMPLEX_VSX_BLOCK2_KERNEL))
+!            ttt = mpi_wtime()
+!            do j = ncols, 1, -1
 !#ifdef WITH_OPENMP
 !              call single_hh_trafo_&
 !                   &MATH_DATATYPE&
@@ -683,12 +683,12 @@
 !                   &PRECISION&
 !                   & (c_loc(a(1,j+off+a_off,istripe)), bcast_buffer(1,j+off),nbw,nl,stripe_width)
 !#endif
-            enddo
-#endif /* (!defined(WITH_FIXED_COMPLEX_KERNEL)) || (defined(WITH_FIXED_COMPLEX_KERNEL) && !defined(WITH_COMPLEX_VSX_BLOCK2_KERNEL)) */
-
-#ifndef WITH_FIXED_COMPLEX_KERNEL
-          endif ! (kernel .eq. ELPA_2STAGE_COMPLEX_VSX_BLOCK1)
-#endif /* not WITH_FIXED_COMPLEX_KERNEL */
+!            enddo
+!#endif /* (!defined(WITH_FIXED_COMPLEX_KERNEL)) || (defined(WITH_FIXED_COMPLEX_KERNEL) && !defined(WITH_COMPLEX_VSX_BLOCK2_KERNEL)) */
+!
+!#ifndef WITH_FIXED_COMPLEX_KERNEL
+!          endif ! (kernel .eq. ELPA_2STAGE_COMPLEX_VSX_BLOCK1)
+!#endif /* not WITH_FIXED_COMPLEX_KERNEL */
 #endif /* WITH_COMPLEX_VSX_BLOCK1_KERNEL */
 
 #endif /* COMPLEXCASE */
@@ -965,14 +965,14 @@
 ! implementation of vsx block 2 complex case
 
 #if defined(WITH_COMPLEX_VSX_BLOCK2_KERNEL)
-#ifndef WITH_FIXED_COMPLEX_KERNEL
-           if (kernel .eq. ELPA_2STAGE_COMPLEX_VSX_BLOCK2) then
-#endif  /* not WITH_FIXED_COMPLEX_KERNEL */
-
-             ttt = mpi_wtime()
-             do j = ncols, 2, -2
-               w(:,1) = bcast_buffer(1:nbw,j+off)
-               w(:,2) = bcast_buffer(1:nbw,j+off-1)
+!#ifndef WITH_FIXED_COMPLEX_KERNEL
+!           if (kernel .eq. ELPA_2STAGE_COMPLEX_VSX_BLOCK2) then
+!#endif  /* not WITH_FIXED_COMPLEX_KERNEL */
+!
+!             ttt = mpi_wtime()
+!             do j = ncols, 2, -2
+!               w(:,1) = bcast_buffer(1:nbw,j+off)
+!               w(:,2) = bcast_buffer(1:nbw,j+off-1)
 !#ifdef WITH_OPENMP
 !               call double_hh_trafo_&
 !                    &MATH_DATATYPE&
@@ -986,7 +986,7 @@
 !                    &PRECISION&
 !                    & (c_loc(a(1,j+off+a_off-1,istripe)), w, nbw, nl, stripe_width, nbw)
 !#endif
-             enddo
+!             enddo
 !#ifdef WITH_OPENMP
 !             if (j==1) call single_hh_trafo_&
 !                 &MATH_DATATYPE&
@@ -1000,10 +1000,10 @@
 !                            &PRECISION&
 !                            & (c_loc(a(1,1+off+a_off,istripe)), bcast_buffer(1,off+1), nbw, nl, stripe_width)
 !#endif
-
-#ifndef WITH_FIXED_COMPLEX_KERNEL
-           endif ! (kernel .eq. ELPA_2STAGE_COMPLEX_VSX_BLOCK2)
-#endif  /* not WITH_FIXED_COMPLEX_KERNEL */
+!
+!#ifndef WITH_FIXED_COMPLEX_KERNEL
+!           endif ! (kernel .eq. ELPA_2STAGE_COMPLEX_VSX_BLOCK2)
+!#endif  /* not WITH_FIXED_COMPLEX_KERNEL */
 #endif /* WITH_COMPLEX_VSX_BLOCK2_KERNEL */
 #endif /* COMPLEXCASE == 1 */
 

src/elpa2/elpa2_template.F90

@@ -172,7 +172,7 @@
     if (gpu == 1) then
       if (kernel .ne. ELPA_2STAGE_REAL_GPU) then
         write(error_unit,*) "ELPA: Warning, GPU usage has been requested but compute kernel is defined as non-GPU!"
-	write(error_unit,*) "The compute kernel will be executed on CPUs!"
+        write(error_unit,*) "The compute kernel will be executed on CPUs!"
       else if (nblk .ne. 128) then
         kernel = ELPA_2STAGE_REAL_GENERIC
       endif
@@ -182,6 +182,18 @@
         write(error_unit,*) "ELPA: Warning, GPU usage has been requested but compute kernel is defined as non-GPU!"
       endif
     endif
+
+#ifdef SINGLE_PRECISION_REAL
+    ! special case at the moment NO single precision kernels on POWER 8 -> set GENERIC for now
+    if (kernel .eq. ELPA_2STAGE_REAL_VSX_BLOCK2 .or. &
+        kernel .eq. ELPA_2STAGE_REAL_VSX_BLOCK4 .or. &
+        kernel .eq. ELPA_2STAGE_REAL_VSX_BLOCK6        ) then
+        write(error_unit,*) "ELPA: At the moment there exist no specific SINGLE precision kernels for POWER8"
+        write(error_unit,*) "The GENERIC kernel will be used at the moment"
+        kernel = ELPA_2STAGE_REAL_GENERIC
+    endif
+#endif
+
 #endif
 
 #if COMPLEXCASE == 1
@@ -191,7 +203,7 @@
     if (gpu == 1) then
       if (kernel .ne. ELPA_2STAGE_COMPLEX_GPU) then
         write(error_unit,*) "ELPA: Warning, GPU usage has been requested but compute kernel is defined as non-GPU!"
-	write(error_unit,*) "The compute kernel will be executed on CPUs!"
+        write(error_unit,*) "The compute kernel will be executed on CPUs!"
       else if (nblk .ne. 128) then
         kernel = ELPA_2STAGE_COMPLEX_GENERIC
       endif