Start to unify complex SSE kernels

Doxyfile.in

@@ -926,7 +926,7 @@ EXCLUDE                = @top_srcdir@/src/GPU/check_for_gpu.F90 \
 			@top_srcdir@/src/elpa2/kernels/complex_avx-avx2_1hv_double_precision.c \
 			@top_srcdir@/src/elpa2/kernels/complex_avx512_2hv_single_precision.c \
 			@top_srcdir@/src/elpa2/kernels/mod_single_hh_trafo_real.F90 \
-			@top_srcdir@/src/elpa2/kernels/complex_sse_1hv_template.c \
+			@top_srcdir@/src/elpa2/kernels/complex_128bit_256bit_512bit_BLOCK_template.c \
 			@top_srcdir@/src/elpa2/kernels/real_avx-avx2_4hv_double_precision.c \
 			@top_srcdir@/src/elpa2/kernels/complex_avx512_2hv_double_precision.c \
 			@top_srcdir@/src/elpa2/kernels/complex_sse_2hv_template.c \

Makefile.am

@@ -795,7 +795,7 @@ EXTRA_DIST = \
   src/elpa2/kernels/complex_avx-avx2_2hv_template.c \
   src/elpa2/kernels/complex_avx512_1hv_template.c \
   src/elpa2/kernels/complex_avx512_2hv_template.c \
-  src/elpa2/kernels/complex_sse_1hv_template.c \
+  src/elpa2/kernels/complex_128bit_256bit_512bit_BLOCK_template.c \
   src/elpa2/kernels/complex_sse_2hv_template.c \
   src/elpa2/kernels/complex_template.F90 \
   src/elpa2/kernels/real_vsx_4hv_template.c \

src/elpa2/kernels/complex_sse_1hv_double_precision.c

@@ -48,8 +48,12 @@
 
 #define COMPLEXCASE 1
 #define DOUBLE_PRECISION 1
+#define BLOCK1 1
+#define VEC_SET SSE_128
 #include "../../general/precision_macros.h"
-#include "complex_sse_1hv_template.c"
+#include "complex_128bit_256bit_512bit_BLOCK_template.c"
+#undef BLOCK1
+#undef VEC_SET
 #undef DOUBLE_PRECISION
 #undef COMPLEXCASE
 

src/elpa2/kernels/complex_sse_1hv_single_precision.c

@@ -48,8 +48,12 @@
 
 #define COMPLEXCASE 1
 #define SINGLE_PRECISION 1
+#define BLOCK1 1
+#define VEC_SET SSE_128
 #include "../../general/precision_macros.h"
-#include "complex_sse_1hv_template.c"
+#include "complex_128bit_256bit_512bit_BLOCK_template.c"
+#undef BLOCK1
+#undef VEC_SET
 #undef COMPLEXCASE
 #undef SINGLE_PRECISION