openblas/openblas-ppc64be_up2_p8.patch

115 lines
3.4 KiB
Diff
Raw Normal View History

From: Michel Normand <normand@linux.vnet.ibm.com>
Subject: openblas ppc64be up2 p8
Date: Wed, 03 Feb 2021 15:39:25 +0100
openblas ppc64be up2 p8
because:
* openblas build failed for ppc64 (BE) in openSUSE
since version 0.3.12
* ppc64 (BE) not supported by IBM after P8.
Signed-off-by: Michel Normand <normand@linux.vnet.ibm.com>
---
Makefile.system | 10 +++++++---
driver/others/dynamic_power.c | 11 +++++++++++
2 files changed, 18 insertions(+), 3 deletions(-)
Index: OpenBLAS-0.3.25/driver/others/dynamic_power.c
===================================================================
--- OpenBLAS-0.3.25.orig/driver/others/dynamic_power.c
+++ OpenBLAS-0.3.25/driver/others/dynamic_power.c
@@ -3,12 +3,14 @@
extern gotoblas_t gotoblas_POWER6;
extern gotoblas_t gotoblas_POWER8;
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
#if ((!defined __GNUC__) || ( __GNUC__ >= 6)) || defined(__clang__)
extern gotoblas_t gotoblas_POWER9;
#endif
#ifdef HAVE_P10_SUPPORT
extern gotoblas_t gotoblas_POWER10;
#endif
+#endif
extern void openblas_warning(int verbose, const char *msg);
@@ -28,11 +30,13 @@ char *gotoblas_corename(void) {
#endif
if (gotoblas == &gotoblas_POWER8) return corename[2];
#if ((!defined __GNUC__) || ( __GNUC__ >= 6)) || defined(__clang__)
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
if (gotoblas == &gotoblas_POWER9) return corename[3];
#endif
#ifdef HAVE_P10_SUPPORT
if (gotoblas == &gotoblas_POWER10) return corename[4];
#endif
+#endif
return corename[0];
}
@@ -243,6 +247,10 @@ static gotoblas_t *get_coretype(void) {
#endif
if (__builtin_cpu_is("power8"))
return &gotoblas_POWER8;
+ /* Fall back to the POWER8 implementation for big endian */
+#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
+ return &gotoblas_POWER8;
+#else
#if ((!defined __GNUC__) || ( __GNUC__ >= 6)) || defined(__clang__)
if (__builtin_cpu_is("power9"))
return &gotoblas_POWER9;
@@ -260,6 +268,7 @@ static gotoblas_t *get_coretype(void) {
if (__builtin_cpu_is("power10"))
return &gotoblas_POWER9;
#endif
+#endif
return NULL;
}
@@ -284,12 +293,14 @@ static gotoblas_t *force_coretype(char *
case 1: return (&gotoblas_POWER6);
#endif
case 2: return (&gotoblas_POWER8);
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
#if ((!defined __GNUC__) || ( __GNUC__ >= 6)) || defined(__clang__)
case 3: return (&gotoblas_POWER9);
#endif
#ifdef HAVE_P10_SUPPORT
case 4: return (&gotoblas_POWER10);
#endif
+#endif
default: return NULL;
}
snprintf(message, 128, "Core not found: %s\n", coretype);
Index: OpenBLAS-0.3.25/Makefile.system
===================================================================
--- OpenBLAS-0.3.25.orig/Makefile.system
+++ OpenBLAS-0.3.25/Makefile.system
@@ -748,6 +748,9 @@ ifeq ($(ARCH), power)
ifneq ($(C_COMPILER), PGI)
DYNAMIC_CORE = POWER6
DYNAMIC_CORE += POWER8
+ifeq ($(__BYTE_ORDER__),__ORDER_BIG_ENDIAN__)
+$(info, OpenBLAS: for big endian limit to POWER8 kernels.)
+else
ifneq ($(C_COMPILER), GCC)
DYNAMIC_CORE += POWER9
DYNAMIC_CORE += POWER10
@@ -776,11 +779,12 @@ else
$(info, OpenBLAS: Your gcc version is too old to build the POWER10 kernels.)
endif
endif
-else
+endif # __ORDER_BIG_ENDIAN__
+else # C_COMPILER PGI
DYNAMIC_CORE = POWER8
DYNAMIC_CORE += POWER9
-endif
-endif
+endif # C_COMPILER PGI
+endif # ARCH power
# If DYNAMIC_CORE is not set, DYNAMIC_ARCH cannot do anything, so force it to empty
ifndef DYNAMIC_CORE