Unify 128bit kernels real block2, block4, and block6

e04b11b2 · Andreas Marek · b6369e22 · e04b11b2 · e04b11b2 · e04b11b2
Commit e04b11b2 authored May 06, 2019 by Andreas Marek
--- a/Doxyfile.in
+++ b/Doxyfile.in
@@ -900,7 +900,7 @@ EXCLUDE                = @top_srcdir@/src/GPU/check_for_gpu.F90 \
 			@top_srcdir@/src/elpa2/kernels/real_avx-avx2_6hv_single_precision.c \
 			@top_srcdir@/src/elpa2/kernels/real_avx-avx2_6hv_double_precision.c \
 			@top_srcdir@/src/elpa2/kernels/complex_sse_1hv_double_precision.c \
-			@top_srcdir@/src/elpa2/kernels/real_sse_6hv_template.c \
+			@top_srcdir@/src/elpa2/kernels/real_128_BLOCK_template.c \
 			@top_srcdir@/src/elpa2/kernels/complex_template.F90 \
 			@top_srcdir@/src/elpa2/kernels/complex_avx-avx2_2hv_double_precision.c \
 			@top_srcdir@/src/elpa2/kernels/real_avx512_2hv_double_precision.c \

--- a/Makefile.am
+++ b/Makefile.am
@@ -785,7 +785,6 @@ EXTRA_DIST = \
  src/elpa2/kernels/real_vsx_4hv_template.c \
  src/elpa2/kernels/real_vsx_6hv_template.c \
  src/elpa2/kernels/real_128bit_BLOCK_template.c \
-  src/elpa2/kernels/real_sse_6hv_template.c \
  src/elpa2/kernels/real_template.F90 \
  src/elpa2/kernels/simple_template.F90 \
  src/elpa2/kernels/simple_block4_template.F90 \

--- a/src/elpa2/kernels/real_128bit_BLOCK_template.c
+++ b/src/elpa2/kernels/real_128bit_BLOCK_template.c
--- a/src/elpa2/kernels/real_sparc64_2hv_double_precision.c
+++ b/src/elpa2/kernels/real_sparc64_2hv_double_precision.c
@@ -48,7 +48,7 @@

 #define REALCASE 1
 #define DOUBLE_PRECISION 1
-#define BLOCK2
+#define BLOCK2 1
 #include "../../general/precision_macros.h"
 #include "real_128bit_BLOCK_template.c"
 #undef REALCASE

--- a/src/elpa2/kernels/real_sparc64_2hv_single_precision.c
+++ b/src/elpa2/kernels/real_sparc64_2hv_single_precision.c
@@ -48,7 +48,7 @@

 #define REALCASE 1
 #define SINGLE_PRECISION 1
-#define BLOCK2
+#define BLOCK2 1
 #include "../../general/precision_macros.h"
 #include "real_128bit_BLOCK_template.c"
 #undef BLOCK2

--- a/src/elpa2/kernels/real_sparc64_4hv_single_precision.c
+++ b/src/elpa2/kernels/real_sparc64_4hv_single_precision.c
@@ -48,7 +48,7 @@

 #define REALCASE 1
 #define SINGLE_PRECISION 1
-#define BLOCK4
+#define BLOCK4 1
 #include "../../general/precision_macros.h"
 #include "real_128bit_BLOCK_template.c"
 #undef BLOCK4

--- a/src/elpa2/kernels/real_sparc64_6hv_double_precision.c
+++ b/src/elpa2/kernels/real_sparc64_6hv_double_precision.c
@@ -48,8 +48,10 @@

 #define REALCASE 1
 #define DOUBLE_PRECISION 1
+#define BLOCK6 1
 #include "../../general/precision_macros.h"
-#include "real_sse_6hv_template.c"
+#include "real_128bit_BLOCK_template.c"
 #undef REALCASE
+#undef BLOCK6
 #undef DOUBLE_PRECISION

--- a/src/elpa2/kernels/real_sparc64_6hv_single_precision.c
+++ b/src/elpa2/kernels/real_sparc64_6hv_single_precision.c
@@ -48,8 +48,10 @@

 #define REALCASE 1
 #define SINGLE_PRECISION 1
+#define BLOCK6 1
 #include "../../general/precision_macros.h"
-#include "real_sse_6hv_template.c"
+#include "real_128bit_BLOCK_template.c"
 #undef REALCASE
+#undef BLOCK6
 #undef SINGLE_PRECISION

--- a/src/elpa2/kernels/real_sse_6hv_double_precision.c
+++ b/src/elpa2/kernels/real_sse_6hv_double_precision.c
@@ -48,8 +48,10 @@

 #define REALCASE 1
 #define DOUBLE_PRECISION 1
+#define BLOCK6 1
 #include "../../general/precision_macros.h"
-#include "real_sse_6hv_template.c"
+#include "real_128bit_BLOCK_template.c"
 #undef REALCASE
+#undef BLOCK6
 #undef DOUBLE_PRECISION

--- a/src/elpa2/kernels/real_sse_6hv_single_precision.c
+++ b/src/elpa2/kernels/real_sse_6hv_single_precision.c
@@ -48,8 +48,10 @@

 #define REALCASE 1
 #define SINGLE_PRECISION 1
+#define BLOCK6 1
 #include "../../general/precision_macros.h"
-#include "real_sse_6hv_template.c"
+#include "real_128bit_BLOCK_template.c"
 #undef REALCASE
+#undef BLOCK6
 #undef SINGLE__PRECISION

--- a/src/elpa2/kernels/real_sse_6hv_template.c
+++ b/src/elpa2/kernels/real_sse_6hv_template.c