Unify compute_hh_trafo complex/real

Makefile.am

@@ -29,7 +29,6 @@ libelpa@SUFFIX@_private_la_SOURCES = \
 	src/mod_redist_band.F90 \
         src/mod_pack_unpack_cpu.F90 \
         src/mod_compute_hh_trafo.F90 \
-        src/mod_compute_hh_trafo_complex.F90 \
         src/aligned_mem.F90 \
         src/elpa1_compute_private.F90 \
 	src/elpa2_determine_workload.F90 \
@@ -904,11 +903,8 @@ elpa2_compute.i: $(top_srcdir)/src/elpa2_compute.F90
 elpa2_kernels_real.i: $(top_srcdir)/src/elpa2_kernels/elpa2_kernels_real.F90
 	$(CPP) $(CPPFLAGS) -I$(top_builddir)/ -c $(top_srcdir)/src/elpa2_kernels/elpa2_kernels_real.F90 -o $@
 
-mod_compute_hh_trafo_real.i: $(top_srcdir)/src/mod_compute_hh_trafo_real.F90
-	$(CPP) $(CPPFLAGS) -I$(top_builddir)/ -c $(top_srcdir)/src/mod_compute_hh_trafo_real.F90 -o $@
-
-mod_compute_hh_trafo_complex.i: $(top_srcdir)/src/mod_compute_hh_trafo_complex.F90
-	$(CPP) $(CPPFLAGS) -I$(top_builddir)/ -c $(top_srcdir)/src/mod_compute_hh_trafo_complex.F90 -o $@
+mod_compute_hh_trafo.i: $(top_srcdir)/src/mod_compute_hh_trafo.F90
+	$(CPP) $(CPPFLAGS) -I$(top_builddir)/ -c $(top_srcdir)/src/mod_compute_hh_trafo.F90 -o $@
 
 test_real.i: $(top_srcdir)/test/Fortran/test_real1.F90
 	$(CPP) $(CPPFLAGS) -I$(top_builddir)/ -c $(top_srcdir)/test/Fortran/test_real1.F90 -o $@

src/elpa2_trans_ev_tridi_to_band_template.X90

@@ -47,13 +47,7 @@
       use elpa2_workload
       use pack_unpack_cpu
       use pack_unpack_gpu
-#if REALCASE == 1
-      use compute_hh_trafo_real
-
-#endif
-#if COMPLEXCASE == 1
-      use compute_hh_trafo_complex
-#endif
+      use compute_hh_trafo
       use cuda_functions
       use precision
       use iso_c_binding

src/mod_compute_hh_trafo.F90

@@ -41,7 +41,7 @@
 !
 ! This file was written by A. Marek, MPCDF
 
-module compute_hh_trafo_real
+module compute_hh_trafo
 #include "config-f90.h"
   use elpa_mpi
   implicit none
@@ -52,6 +52,14 @@ module compute_hh_trafo_real
   public compute_hh_trafo_real_cpu_double
 #endif
 
+#ifdef WITH_OPENMP
+  public compute_hh_trafo_complex_cpu_openmp_double
+#else
+  public compute_hh_trafo_complex_cpu_double
+#endif
+  public compute_hh_trafo_complex_gpu_double
+
+
 #ifdef WANT_SINGLE_PRECISION_REAL
 #ifdef WITH_OPENMP
   public compute_hh_trafo_real_cpu_openmp_single
@@ -60,6 +68,16 @@ module compute_hh_trafo_real
 #endif
 #endif
 
+#ifdef WANT_SINGLE_PRECISION_COMPLEX
+
+
+#ifdef WITH_OPENMP
+  public compute_hh_trafo_complex_cpu_openmp_single
+#else
+  public compute_hh_trafo_complex_cpu_single
+#endif
+  public compute_hh_trafo_complex_gpu_single
+#endif
   contains
 
   !real double precision
@@ -80,5 +98,41 @@ module compute_hh_trafo_real
 #undef SINGLE_PRECISION
 #endif
 
+  !complex double precision
+#define COMPLEXCASE 1
+#define DOUBLE_PRECISION 1
+#include "precision_macros.h"
+#include "compute_hh_trafo.X90"
+#undef COMPLEXCASE
+#undef DOUBLE_PRECISION
+
+ ! complex single precision
+#if defined(WANT_SINGLE_PRECISION_COMPLEX)
+#define COMPLEXCASE 1
+#define SINGLE_PRECISION 1
+#include "precision_macros.h"
+#include "compute_hh_trafo.X90"
+#undef COMPLEXCASE
+#undef SINGLE_PRECISION
+#endif
+
+
+  !complex double precision
+#define COMPLEXCASE 1
+#define DOUBLE_PRECISION 1
+#include "precision_macros.h"
+#include "compute_hh_trafo_complex_gpu.X90"
+#undef COMPLEXCASE
+#undef DOUBLE_PRECISION
+
+ ! complex single precision
+#if defined(WANT_SINGLE_PRECISION_COMPLEX)
+#define COMPLEXCASE 1
+#define SINGLE_PRECISION 1
+#include "precision_macros.h"
+#include "compute_hh_trafo_complex_gpu.X90"
+#undef COMPLEXCASE
+#undef SINGLE_PRECISION
+#endif
 
 end module