Welcome to mirror list, hosted at ThFree Co, Russian Federation.

git.blender.org/blender.git - Unnamed repository; edit this file 'description' to name the repository.
summaryrefslogtreecommitdiff
path: root/intern
diff options
context:
space:
mode:
authorSergey Sharybin <sergey.vfx@gmail.com>2013-08-19 14:51:40 +0400
committerSergey Sharybin <sergey.vfx@gmail.com>2013-08-19 14:51:40 +0400
commitc0f8e15295371ae69b38185d3ee5659cdd1af386 (patch)
tree7a550dcc6b6e5026efc850be4a69023708379d75 /intern
parentefa836531e171028870ee50bcebc4e56a17b536b (diff)
Speedup for guarded allocator
- Re-arrange locks, so no actual memory allocation (which is relatively slow) happens from inside the lock. operation system will take care of locks which might be needed there on it's own. - Use spin lock instead of mutex, since it's just list operations happens from inside lock, no need in mutex here. - Use atomic operations for memory in use and total used blocks counters. This makes guarded allocator almost the same speed as non-guarded one in files from Tube project. There're still MemHead/MemTail overhead which might be bad for CPU cache utilization
Diffstat (limited to 'intern')
-rw-r--r--intern/atomic/atomic_ops.h3
-rw-r--r--intern/guardedalloc/CMakeLists.txt1
-rw-r--r--intern/guardedalloc/SConscript2
-rw-r--r--intern/guardedalloc/intern/mallocn.c46
4 files changed, 23 insertions, 29 deletions
diff --git a/intern/atomic/atomic_ops.h b/intern/atomic/atomic_ops.h
index 845e517a42f..7baa5f6ecac 100644
--- a/intern/atomic/atomic_ops.h
+++ b/intern/atomic/atomic_ops.h
@@ -44,7 +44,8 @@
# endif
#endif
-#if defined(_M_X64) || defined(__amd64__) || defined(__x86_64__)
+/* TODO(sergey): check on other 64bit platforms. */
+#if defined(_M_X64) || defined(_M_AMD64) || defined(__amd64__) || defined(__x86_64__)
# define LG_SIZEOF_PTR 3
# define LG_SIZEOF_INT 3
#else
diff --git a/intern/guardedalloc/CMakeLists.txt b/intern/guardedalloc/CMakeLists.txt
index 4f6c177ef7d..7ac978b0edd 100644
--- a/intern/guardedalloc/CMakeLists.txt
+++ b/intern/guardedalloc/CMakeLists.txt
@@ -25,6 +25,7 @@
set(INC
.
+ ../atomic
)
set(INC_SYS
diff --git a/intern/guardedalloc/SConscript b/intern/guardedalloc/SConscript
index eb558763dbe..e3f787fe6c2 100644
--- a/intern/guardedalloc/SConscript
+++ b/intern/guardedalloc/SConscript
@@ -38,6 +38,6 @@ if env['WITH_BF_CXX_GUARDEDALLOC']:
sources.append('cpp/mallocn.cpp')
defs.append('WITH_CXX_GUARDEDALLOC')
-incs = '.'
+incs = '. ../atomic'
env.BlenderLib ('bf_intern_guardedalloc', sources, Split(incs), defs, libtype=['intern','player'], priority = [5,150] )
diff --git a/intern/guardedalloc/intern/mallocn.c b/intern/guardedalloc/intern/mallocn.c
index 520df788055..0af16e58f0a 100644
--- a/intern/guardedalloc/intern/mallocn.c
+++ b/intern/guardedalloc/intern/mallocn.c
@@ -53,6 +53,8 @@
/* should always be defined except for experimental cases */
#ifdef WITH_GUARDEDALLOC
+#include "atomic_ops.h"
+
/* Blame Microsoft for LLP64 and no inttypes.h, quick workaround needed: */
#if defined(WIN64)
# define SIZET_FORMAT "%I64u"
@@ -210,8 +212,8 @@ static const char *check_memlist(MemHead *memh);
/* --------------------------------------------------------------------- */
-static volatile int totblock = 0;
-static volatile uintptr_t mem_in_use = 0, mmap_in_use = 0, peak_mem = 0;
+static unsigned int totblock = 0;
+static size_t mem_in_use = 0, mmap_in_use = 0, peak_mem = 0;
static volatile struct localListBase _membase;
static volatile struct localListBase *membase = &_membase;
@@ -493,31 +495,29 @@ static void make_memhead_header(MemHead *memh, size_t len, const char *str)
memt = (MemTail *)(((char *) memh) + sizeof(MemHead) + len);
memt->tag3 = MEMTAG3;
-
+
+ atomic_add_u(&totblock, 1);
+ atomic_add_z(&mem_in_use, len);
+
+ mem_lock_thread();
addtail(membase, &memh->next);
if (memh->next) {
memh->nextname = MEMNEXT(memh->next)->name;
}
-
- totblock++;
- mem_in_use += len;
-
peak_mem = mem_in_use > peak_mem ? mem_in_use : peak_mem;
+ mem_unlock_thread();
}
void *MEM_mallocN(size_t len, const char *str)
{
MemHead *memh;
- mem_lock_thread();
-
len = (len + 3) & ~3; /* allocate in units of 4 */
memh = (MemHead *)malloc(len + sizeof(MemHead) + sizeof(MemTail));
if (memh) {
make_memhead_header(memh, len, str);
- mem_unlock_thread();
if (malloc_debug_memset && len)
memset(memh + 1, 255, len);
@@ -528,7 +528,6 @@ void *MEM_mallocN(size_t len, const char *str)
#endif
return (++memh);
}
- mem_unlock_thread();
print_error("Malloc returns null: len=" SIZET_FORMAT " in %s, total %u\n",
SIZET_ARG(len), str, (unsigned int) mem_in_use);
return NULL;
@@ -538,15 +537,12 @@ void *MEM_callocN(size_t len, const char *str)
{
MemHead *memh;
- mem_lock_thread();
-
len = (len + 3) & ~3; /* allocate in units of 4 */
memh = (MemHead *)calloc(len + sizeof(MemHead) + sizeof(MemTail), 1);
if (memh) {
make_memhead_header(memh, len, str);
- mem_unlock_thread();
#ifdef DEBUG_MEMCOUNTER
if (_mallocn_count == DEBUG_MEMCOUNTER_ERROR_VAL)
memcount_raise(__func__);
@@ -554,7 +550,6 @@ void *MEM_callocN(size_t len, const char *str)
#endif
return (++memh);
}
- mem_unlock_thread();
print_error("Calloc returns null: len=" SIZET_FORMAT " in %s, total %u\n",
SIZET_ARG(len), str, (unsigned int) mem_in_use);
return NULL;
@@ -565,8 +560,6 @@ void *MEM_mapallocN(size_t len, const char *str)
{
MemHead *memh;
- mem_lock_thread();
-
len = (len + 3) & ~3; /* allocate in units of 4 */
memh = mmap(NULL, len + sizeof(MemHead) + sizeof(MemTail),
@@ -575,7 +568,8 @@ void *MEM_mapallocN(size_t len, const char *str)
if (memh != (MemHead *)-1) {
make_memhead_header(memh, len, str);
memh->mmap = 1;
- mmap_in_use += len;
+ atomic_add_z(&mmap_in_use, len);
+ mem_lock_thread();
peak_mem = mmap_in_use > peak_mem ? mmap_in_use : peak_mem;
mem_unlock_thread();
#ifdef DEBUG_MEMCOUNTER
@@ -586,7 +580,6 @@ void *MEM_mapallocN(size_t len, const char *str)
return (++memh);
}
else {
- mem_unlock_thread();
print_error("Mapalloc returns null, fallback to regular malloc: "
"len=" SIZET_FORMAT " in %s, total %u\n",
SIZET_ARG(len), str, (unsigned int) mmap_in_use);
@@ -844,7 +837,6 @@ void MEM_freeN(void *vmemh)
return;
}
- mem_lock_thread();
if ((memh->tag1 == MEMTAG1) &&
(memh->tag2 == MEMTAG2) &&
((memh->len & 0x3) == 0))
@@ -858,8 +850,6 @@ void MEM_freeN(void *vmemh)
/* after tags !!! */
rem_memblock(memh);
- mem_unlock_thread();
-
return;
}
MemorY_ErroR(memh->name, "end corrupt");
@@ -869,7 +859,9 @@ void MEM_freeN(void *vmemh)
}
}
else {
+ mem_lock_thread();
name = check_memlist(memh);
+ mem_unlock_thread();
if (name == NULL)
MemorY_ErroR("free", "pointer not in memlist");
else
@@ -879,8 +871,6 @@ void MEM_freeN(void *vmemh)
totblock--;
/* here a DUMP should happen */
- mem_unlock_thread();
-
return;
}
@@ -927,6 +917,7 @@ static void remlink(volatile localListBase *listbase, void *vlink)
static void rem_memblock(MemHead *memh)
{
+ mem_lock_thread();
remlink(membase, &memh->next);
if (memh->prev) {
if (memh->next)
@@ -934,9 +925,10 @@ static void rem_memblock(MemHead *memh)
else
MEMNEXT(memh->prev)->nextname = NULL;
}
+ mem_unlock_thread();
- totblock--;
- mem_in_use -= memh->len;
+ atomic_sub_u(&totblock, 1);
+ atomic_sub_z(&mem_in_use, memh->len);
#ifdef DEBUG_MEMDUPLINAME
if (memh->need_free_name)
@@ -944,7 +936,7 @@ static void rem_memblock(MemHead *memh)
#endif
if (memh->mmap) {
- mmap_in_use -= memh->len;
+ atomic_sub_z(&mmap_in_use, memh->len);
if (munmap(memh, memh->len + sizeof(MemHead) + sizeof(MemTail)))
printf("Couldn't unmap memory %s\n", memh->name);
}