From faf7b3d1bb34854ad05e6da0a2f658276490b316 Mon Sep 17 00:00:00 2001
From: Martin Kroeker <martin@ruby.chemie.uni-freiburg.de>
Date: Wed, 24 Apr 2024 17:49:40 +0200
Subject: [PATCH 1/3] Document the two HUGETLB options for buffer allocation
---
Makefile.rule | 10 ++++++++++
1 file changed, 10 insertions(+)
diff --git a/Makefile.rule b/Makefile.rule
index bb239be2a4..d25299775d 100644
--- a/Makefile.rule
+++ b/Makefile.rule
@@ -219,6 +219,16 @@ NO_AFFINITY = 1
# to the user space. If bigphysarea is enabled, it will use it.
# DEVICEDRIVER_ALLOCATION = 1
+# Use large page allocation (called hugepage support in Linux context)
+# for the thread buffers (with access by shared memory operations)
+# HUGETLB_ALLOCATION = 1
+
+# Use large page allocation called hugepages in Linux) based on mmap accessing
+# a memory-backed pseudofile (requires hugetlbfs to be mounted in the system,
+# the example below has it mounted on /hugepages. OpenBLAS will create the backing
+# file as gotoblas.processid in that path)
+# HUGETLBFILE_ALLOCATION = /hugepages
+
# If you need to synchronize FP CSR between threads (for x86/x86_64 and aarch64 only).
# CONSISTENT_FPCSR = 1
From 9c4e10fbd155daa0ee8206a8313436f14f7f7282 Mon Sep 17 00:00:00 2001
From: Martin Kroeker <martin@ruby.chemie.uni-freiburg.de>
Date: Sat, 4 May 2024 14:48:02 +0200
Subject: [PATCH 2/3] sort hugetlb and shm alloc options
---
Makefile.system | 21 ++++++++++++++-------
1 file changed, 14 insertions(+), 7 deletions(-)
diff --git a/Makefile.system b/Makefile.system
index 840389db73..67830f2a3e 100644
--- a/Makefile.system
+++ b/Makefile.system
@@ -355,7 +355,10 @@ OBJCONV = $(CROSS_SUFFIX)objconv
# When fortran support was either not detected or actively deselected, only build BLAS.
ifeq ($(NOFORTRAN), 1)
C_LAPACK = 1
-override FEXTRALIB =
+override FEXTRALIB =
+ifeq ($(C_COMPILER), GCC)
+CCOMMON_OPT += -Wno-error=incompatible-pointer-types
+endif
endif
ifeq ($(C_COMPILER), GCC)
@@ -959,18 +962,12 @@ endif
ifeq ($(ARCH), loongarch64)
LA64_ABI=$(shell $(CC) -mabi=lp64d -c $(TOPDIR)/cpuid_loongarch64.c -o /dev/null > /dev/null 2> /dev/null && echo lp64d)
-LA64_ARCH=$(shell $(CC) -march=loongarch64 -c $(TOPDIR)/cpuid_loongarch64.c -o /dev/null > /dev/null 2> /dev/null && echo loongarch64)
ifneq ($(LA64_ABI), lp64d)
LA64_ABI=lp64
endif
-ifneq ($(LA64_ARCH), loongarch64)
-CCOMMON_OPT += -mabi=$(LA64_ABI)
-FCOMMON_OPT += -mabi=$(LA64_ABI)
-else
CCOMMON_OPT += -march=loongarch64 -mabi=$(LA64_ABI)
FCOMMON_OPT += -march=loongarch64 -mabi=$(LA64_ABI)
endif
-endif
endif
@@ -1589,13 +1586,23 @@ ifdef FUNCTION_PROFILE
CCOMMON_OPT += -DFUNCTION_PROFILE
endif
+ifdef SHMEM_ALLOCATION
+ifneq ($(SHMEM_ALLOCATION), 0)
+CCOMMON_OPT += -DALLOC_SHM
+endif
+endif
+
ifdef HUGETLB_ALLOCATION
+ifneq ($(HUGETLB_ALLOCATION), 0)
CCOMMON_OPT += -DALLOC_HUGETLB
endif
+endif
ifdef HUGETLBFILE_ALLOCATION
+ifneq ($(HUGETLBFILE_ALLOCATION), 0)
CCOMMON_OPT += -DALLOC_HUGETLBFILE -DHUGETLB_FILE_NAME=$(HUGETLBFILE_ALLOCATION)
endif
+endif
ifdef STATIC_ALLOCATION
CCOMMON_OPT += -DALLOC_STATIC
From dc99b613805fd15f1e71836b0f3e37c17770d697 Mon Sep 17 00:00:00 2001
From: Martin Kroeker <martin@ruby.chemie.uni-freiburg.de>
Date: Sat, 4 May 2024 14:49:00 +0200
Subject: [PATCH 3/3] sort unwanted interdependencies of alloc_shm and
alloc_hugetlb
---
driver/others/memory.c | 58 ++++++++++++++++++++++++++++++++----------
1 file changed, 44 insertions(+), 14 deletions(-)
diff --git a/driver/others/memory.c b/driver/others/memory.c
index c55688aced..94b2f1a6ad 100644
--- a/driver/others/memory.c
+++ b/driver/others/memory.c
@@ -1165,11 +1165,10 @@ void *blas_memory_alloc(int procpos){
#ifdef ALLOC_DEVICEDRIVER
alloc_devicedirver,
#endif
-/* Hugetlb implicitly assumes ALLOC_SHM */
-#ifdef ALLOC_SHM
+#ifdef ALLOC_SHM && !defined(ALLOC_HUGETLB)
alloc_shm,
#endif
-#if ((defined ALLOC_SHM) && (defined OS_LINUX || defined OS_AIX || defined __sun__ || defined OS_WINDOWS))
+#if ((defined ALLOC_HUGETLB) && (defined OS_LINUX || defined OS_AIX || defined __sun__ || defined OS_WINDOWS))
alloc_hugetlb,
#endif
#ifdef ALLOC_MMAP
@@ -1190,7 +1189,6 @@ void *blas_memory_alloc(int procpos){
struct alloc_t * alloc_info;
struct alloc_t ** alloc_table;
-
#if defined(SMP) && !defined(USE_OPENMP)
int mi;
LOCK_COMMAND(&alloc_lock);
@@ -1282,7 +1280,7 @@ UNLOCK_COMMAND(&alloc_lock);
}
#endif
-#if (defined ALLOC_SHM) && (defined OS_LINUX || defined OS_AIX || defined __sun__ || defined OS_WINDOWS)
+#if (defined ALLOC_HUGETLB) && (defined OS_LINUX || defined OS_AIX || defined __sun__ || defined OS_WINDOWS)
if ((*func == alloc_hugetlb) && (map_address != (void *)-1)) hugetlb_allocated = 1;
#endif
@@ -2494,7 +2492,7 @@ static void *alloc_devicedirver(void *address){
#endif
-#ifdef ALLOC_SHM
+#if defined(ALLOC_SHM) && !defined(ALLOC_HUGETLB)
static void alloc_shm_free(struct release_t *release){
@@ -2506,7 +2504,9 @@ static void alloc_shm_free(struct release_t *release){
static void *alloc_shm(void *address){
void *map_address;
int shmid;
-
+#ifdef DEBUG
+ fprintf(stderr,"alloc_shm got called\n");
+#endif
shmid = shmget(IPC_PRIVATE, BUFFER_SIZE,IPC_CREAT | 0600);
map_address = (void *)shmat(shmid, address, 0);
@@ -2533,6 +2533,7 @@ static void *alloc_shm(void *address){
return map_address;
}
+#endif
#if defined OS_LINUX || defined OS_AIX || defined __sun__ || defined OS_WINDOWS
@@ -2562,6 +2563,10 @@ static void *alloc_hugetlb(void *address){
void *map_address = (void *)-1;
+#ifdef DEBUG
+fprintf(stderr,"alloc_hugetlb got called\n");
+#endif
+
#if defined(OS_LINUX) || defined(OS_AIX)
int shmid;
@@ -2583,7 +2588,7 @@ static void *alloc_hugetlb(void *address){
if (map_address != (void *)-1){
shmctl(shmid, IPC_RMID, 0);
- }
+ }else printf("alloc_hugetlb failed\n");
}
#endif
@@ -2645,7 +2650,6 @@ static void *alloc_hugetlb(void *address){
}
#endif
-#endif
#ifdef ALLOC_HUGETLBFILE
@@ -2762,11 +2766,10 @@ void *blas_memory_alloc(int procpos){
#ifdef ALLOC_DEVICEDRIVER
alloc_devicedirver,
#endif
-/* Hugetlb implicitly assumes ALLOC_SHM */
-#ifdef ALLOC_SHM
+#ifdef ALLOC_SHM && !defined(ALLOC_HUGETLB)
alloc_shm,
#endif
-#if ((defined ALLOC_SHM) && (defined OS_LINUX || defined OS_AIX || defined __sun__ || defined OS_WINDOWS))
+#if ((defined ALLOC_HUGETLB) && (defined OS_LINUX || defined OS_AIX || defined __sun__ || defined OS_WINDOWS))
alloc_hugetlb,
#endif
#ifdef ALLOC_MMAP
@@ -2945,8 +2948,22 @@ void *blas_memory_alloc(int procpos){
}
#endif
-#if (defined ALLOC_SHM) && (defined OS_LINUX || defined OS_AIX || defined __sun__ || defined OS_WINDOWS)
+#if (defined ALLOC_HUGETLB) && (defined OS_LINUX || defined OS_AIX || defined __sun__ || defined OS_WINDOWS)
if ((*func == alloc_hugetlb) && (map_address != (void *)-1)) hugetlb_allocated = 1;
+#ifdef DEBUG
+ if (hugetlb_allocated) printf("allocating via shared memory with large page support (hugetlb)\n");
+#endif
+#endif
+
+#if (defined ALLOC_SHM) && (defined OS_LINUX || defined OS_AIX || defined __sun__ || defined OS_WINDOWS)
+#ifdef DEBUG
+ printf("allocating via shared memory\n");
+#endif
+ if ((*func == alloc_shm) && (map_address == (void *)-1)) {
+#ifndef OS_WINDOWS
+ fprintf(stderr, "OpenBLAS Warning ... shared memory allocation was failed.\n");
+#endif
+ }
#endif
func ++;
@@ -3061,10 +3078,23 @@ void *blas_memory_alloc(int procpos){
}
#endif
-#if (defined ALLOC_SHM) && (defined OS_LINUX || defined OS_AIX || defined __sun__ || defined OS_WINDOWS)
+#if (defined ALLOC_HUGETLB) && (defined OS_LINUX || defined OS_AIX || defined __sun__ || defined OS_WINDOWS)
+#ifdef DEBUG
+ fprintf(stderr,"OpenBLAS: allocating via shared memory with large page support (hugetlb)\n");
+#endif
if ((*func == alloc_hugetlb) && (map_address != (void *)-1)) hugetlb_allocated = 1;
#endif
+#if (defined ALLOC_SHM) && (defined OS_LINUX || defined OS_AIX || defined __sun__ || defined OS_WINDOWS)
+#ifdef DEBUG
+ fprintf(stderr,"allocating via shared memory\n");
+#endif
+ if ((*func == alloc_shm) && (map_address == (void *)-1)) {
+#ifndef OS_WINDOWS
+ fprintf(stderr, "OpenBLAS Warning ... shared memory allocation was failed.\n");
+#endif
+ }
+#endif
func ++;
}