floraachy
/
OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Remove the need for most locking in memory.c. 

Using thread local storage for tracking memory allocations means that threads
no longer have to lock at all when doing memory allocations / frees. This
particularly helps the gemm driver since it does an allocation per invocation.
Even without threading at all, this helps, since even calling a lock with
no contention has a cost:

Before this change, no threading:
```
----------------------------------------------------
Benchmark             Time           CPU Iterations
----------------------------------------------------
BM_SGEMM/4          102 ns        102 ns   13504412
BM_SGEMM/6          175 ns        175 ns    7997580
BM_SGEMM/8          205 ns        205 ns    6842073
BM_SGEMM/10         266 ns        266 ns    5294919
BM_SGEMM/16         478 ns        478 ns    2963441
BM_SGEMM/20         690 ns        690 ns    2144755
BM_SGEMM/32        1906 ns       1906 ns     716981
BM_SGEMM/40        2983 ns       2983 ns     473218
BM_SGEMM/64        9421 ns       9422 ns     148450
BM_SGEMM/72       12630 ns      12631 ns     112105
BM_SGEMM/80       15845 ns      15846 ns      89118
BM_SGEMM/90       25675 ns      25676 ns      54332
BM_SGEMM/100      29864 ns      29865 ns      47120
BM_SGEMM/112      37841 ns      37842 ns      36717
BM_SGEMM/128      56531 ns      56532 ns      25361
BM_SGEMM/140      75886 ns      75888 ns      18143
BM_SGEMM/150      98493 ns      98496 ns      14299
BM_SGEMM/160     102620 ns     102622 ns      13381
BM_SGEMM/170     135169 ns     135173 ns      10231
BM_SGEMM/180     146170 ns     146172 ns       9535
BM_SGEMM/189     190226 ns     190231 ns       7397
BM_SGEMM/200     194513 ns     194519 ns       7210
BM_SGEMM/256     396561 ns     396573 ns       3531
```
with this change:
```
----------------------------------------------------
Benchmark             Time           CPU Iterations
----------------------------------------------------
BM_SGEMM/4           95 ns         95 ns   14500387
BM_SGEMM/6          166 ns        166 ns    8381763
BM_SGEMM/8          196 ns        196 ns    7277044
BM_SGEMM/10         256 ns        256 ns    5515721
BM_SGEMM/16         463 ns        463 ns    3025197
BM_SGEMM/20         636 ns        636 ns    2070213
BM_SGEMM/32        1885 ns       1885 ns     739444
BM_SGEMM/40        2969 ns       2969 ns     472152
BM_SGEMM/64        9371 ns       9372 ns     148932
BM_SGEMM/72       12431 ns      12431 ns     112919
BM_SGEMM/80       15615 ns      15616 ns      89978
BM_SGEMM/90       25397 ns      25398 ns      55041
BM_SGEMM/100      29445 ns      29446 ns      47540
BM_SGEMM/112      37530 ns      37531 ns      37286
BM_SGEMM/128      55373 ns      55375 ns      25277
BM_SGEMM/140      76241 ns      76241 ns      18259
BM_SGEMM/150     102196 ns     102200 ns      13736
BM_SGEMM/160     101521 ns     101525 ns      13556
BM_SGEMM/170     136182 ns     136184 ns      10567
BM_SGEMM/180     146861 ns     146864 ns       9035
BM_SGEMM/189     192632 ns     192632 ns       7231
BM_SGEMM/200     198547 ns     198555 ns       6995
BM_SGEMM/256     392316 ns     392330 ns       3539
```

Before, when built with USE_THREAD=1, GEMM_MULTITHREAD_THRESHOLD = 4, the cost
of small matrix operations was overshadowed by thread locking (look smaller than
32) even when not explicitly spawning threads:
```
----------------------------------------------------
Benchmark             Time           CPU Iterations
----------------------------------------------------
BM_SGEMM/4          328 ns        328 ns    4170562
BM_SGEMM/6          396 ns        396 ns    3536400
BM_SGEMM/8          418 ns        418 ns    3330102
BM_SGEMM/10         491 ns        491 ns    2863047
BM_SGEMM/16         710 ns        710 ns    2028314
BM_SGEMM/20         871 ns        871 ns    1581546
BM_SGEMM/32        2132 ns       2132 ns     657089
BM_SGEMM/40        3197 ns       3196 ns     437969
BM_SGEMM/64        9645 ns       9645 ns     144987
BM_SGEMM/72       35064 ns      32881 ns      50264
BM_SGEMM/80       37661 ns      35787 ns      42080
BM_SGEMM/90       36507 ns      36077 ns      40091
BM_SGEMM/100      32513 ns      31850 ns      48607
BM_SGEMM/112      41742 ns      41207 ns      37273
BM_SGEMM/128      67211 ns      65095 ns      21933
BM_SGEMM/140      68263 ns      67943 ns      19245
BM_SGEMM/150     121854 ns     115439 ns      10660
BM_SGEMM/160     116826 ns     115539 ns      10000
BM_SGEMM/170     126566 ns     122798 ns      11960
BM_SGEMM/180     130088 ns     127292 ns      11503
BM_SGEMM/189     120309 ns     116634 ns      13162
BM_SGEMM/200     114559 ns     110993 ns      10000
BM_SGEMM/256     217063 ns     207806 ns       6417
```
and after, it's gone (note this includes my other change which reduces calls
to num_cpu_avail):
```
----------------------------------------------------
Benchmark             Time           CPU Iterations
----------------------------------------------------
BM_SGEMM/4           95 ns         95 ns   12347650
BM_SGEMM/6          166 ns        166 ns    8259683
BM_SGEMM/8          193 ns        193 ns    7162210
BM_SGEMM/10         258 ns        258 ns    5415657
BM_SGEMM/16         471 ns        471 ns    2981009
BM_SGEMM/20         666 ns        666 ns    2148002
BM_SGEMM/32        1903 ns       1903 ns     738245
BM_SGEMM/40        2969 ns       2969 ns     473239
BM_SGEMM/64        9440 ns       9440 ns     148442
BM_SGEMM/72       37239 ns      33330 ns      46813
BM_SGEMM/80       57350 ns      55949 ns      32251
BM_SGEMM/90       36275 ns      36249 ns      42259
BM_SGEMM/100      31111 ns      31008 ns      45270
BM_SGEMM/112      43782 ns      40912 ns      34749
BM_SGEMM/128      67375 ns      64406 ns      22443
BM_SGEMM/140      76389 ns      67003 ns      21430
BM_SGEMM/150      72952 ns      71830 ns      19793
BM_SGEMM/160      97039 ns      96858 ns      11498
BM_SGEMM/170     123272 ns     122007 ns      11855
BM_SGEMM/180     126828 ns     126505 ns      11567
BM_SGEMM/189     115179 ns     114665 ns      11044
BM_SGEMM/200      89289 ns      87259 ns      16147
BM_SGEMM/256     226252 ns     222677 ns       7375
```

I've also tested this with ThreadSanitizer and found no data races during
execution.  I'm not sure why 200 is always faster than it's neighbors, we must
be hitting some optimal cache size or something.
This commit is contained in:
Craig Donner 2018-06-14 12:18:04 +01:00
parent ed682a4a0c
commit bf40f806ef
1 changed files with 43 additions and 156 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

171
driver/others/memory.c
View File

@ -139,6 +139,14 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define FIXED_PAGESIZE 4096 #define FIXED_PAGESIZE 4096
#endif #endif
#ifndef BUFFERS_PER_THREAD
#ifdef USE_OPENMP
#define BUFFERS_PER_THREAD (MAX_CPU_NUMBER * 2 * MAX_PARALLEL_NUMBER)
#else
#define BUFFERS_PER_THREAD NUM_BUFFERS
#endif
#endif
#define BITMASK(a, b, c) ((((a) >> (b)) & (c))) #define BITMASK(a, b, c) ((((a) >> (b)) & (c)))
#if defined(_MSC_VER) && !defined(__clang__) #if defined(_MSC_VER) && !defined(__clang__)
@ -415,8 +423,15 @@ struct release_t {
int hugetlb_allocated = 0; int hugetlb_allocated = 0;
static struct release_t release_info[NUM_BUFFERS]; #if defined(OS_WINDOWS)
static int release_pos = 0; #define THREAD_LOCAL __declspec(thread)
#define UNLIKELY_TO_BE_ZERO(x) (x)
#else
#define THREAD_LOCAL __thread
#define UNLIKELY_TO_BE_ZERO(x) (__builtin_expect(x, 0))
#endif
static struct release_t THREAD_LOCAL release_info[BUFFERS_PER_THREAD];
static int THREAD_LOCAL release_pos = 0;
#if defined(OS_LINUX) && !defined(NO_WARMUP) #if defined(OS_LINUX) && !defined(NO_WARMUP)
static int hot_alloc = 0; static int hot_alloc = 0;
@ -459,15 +474,9 @@ static void *alloc_mmap(void *address){
} }
if (map_address != (void *)-1) { if (map_address != (void *)-1) {
#if defined(SMP) && !defined(USE_OPENMP)
LOCK_COMMAND(&alloc_lock);
#endif
release_info[release_pos].address = map_address; release_info[release_pos].address = map_address;
release_info[release_pos].func = alloc_mmap_free; release_info[release_pos].func = alloc_mmap_free;
release_pos ++; release_pos ++;
#if defined(SMP) && !defined(USE_OPENMP)
UNLOCK_COMMAND(&alloc_lock);
#endif
} }
#ifdef OS_LINUX #ifdef OS_LINUX
@ -611,15 +620,9 @@ static void *alloc_mmap(void *address){
#endif #endif
if (map_address != (void *)-1) { if (map_address != (void *)-1) {
#if defined(SMP) && !defined(USE_OPENMP)
LOCK_COMMAND(&alloc_lock);
#endif
release_info[release_pos].address = map_address; release_info[release_pos].address = map_address;
release_info[release_pos].func = alloc_mmap_free; release_info[release_pos].func = alloc_mmap_free;
release_pos ++; release_pos ++;
#if defined(SMP) && !defined(USE_OPENMP)
UNLOCK_COMMAND(&alloc_lock);
#endif
} }
return map_address; return map_address;
@ -961,20 +964,17 @@ static BLASULONG base_address = 0UL;
static BLASULONG base_address = BASE_ADDRESS; static BLASULONG base_address = BASE_ADDRESS;
#endif #endif
static volatile struct { struct memory_t {
BLASULONG lock;
void *addr; void *addr;
#if defined(WHEREAMI) && !defined(USE_OPENMP)
int pos;
#endif
int used; int used;
#ifndef __64BIT__ #ifndef __64BIT__
char dummy[48]; char dummy[48];
#else #else
char dummy[40]; char dummy[40];
#endif #endif
};
} memory[NUM_BUFFERS]; static struct memory_t THREAD_LOCAL memory[BUFFERS_PER_THREAD];
static int memory_initialized = 0; static int memory_initialized = 0;
@ -987,9 +987,6 @@ static int memory_initialized = 0;
void *blas_memory_alloc(int procpos){ void *blas_memory_alloc(int procpos){
int position; int position;
#if defined(WHEREAMI) && !defined(USE_OPENMP)
int mypos;
#endif
void *map_address; void *map_address;
@ -1020,23 +1017,13 @@ void *blas_memory_alloc(int procpos){
}; };
void *(**func)(void *address); void *(**func)(void *address);
#if defined(USE_OPENMP) if (UNLIKELY_TO_BE_ZERO(memory_initialized)) {
if (!memory_initialized) {
#endif
/* Only allow a single thread to initialize memory system */
LOCK_COMMAND(&alloc_lock); LOCK_COMMAND(&alloc_lock);
if (!memory_initialized) { if (!memory_initialized) {
#if defined(WHEREAMI) && !defined(USE_OPENMP)
for (position = 0; position < NUM_BUFFERS; position ++){
memory[position].addr = (void *)0;
memory[position].pos = -1;
memory[position].used = 0;
memory[position].lock = 0;
}
#endif
#ifdef DYNAMIC_ARCH #ifdef DYNAMIC_ARCH
gotoblas_dynamic_init(); gotoblas_dynamic_init();
#endif #endif
@ -1059,63 +1046,19 @@ void *blas_memory_alloc(int procpos){
} }
UNLOCK_COMMAND(&alloc_lock); UNLOCK_COMMAND(&alloc_lock);
#if defined(USE_OPENMP)
} }
#endif
#ifdef DEBUG #ifdef DEBUG
printf("Alloc Start ...\n"); printf("Alloc Start ...\n");
#endif
#if defined(WHEREAMI) && !defined(USE_OPENMP)
mypos = WhereAmI();
position = mypos;
while (position >= NUM_BUFFERS) position >>= 1;
do {
if (!memory[position].used && (memory[position].pos == mypos)) {
#if defined(SMP) && !defined(USE_OPENMP)
LOCK_COMMAND(&alloc_lock);
#else
blas_lock(&memory[position].lock);
#endif
if (!memory[position].used) goto allocation;
#if defined(SMP) && !defined(USE_OPENMP)
UNLOCK_COMMAND(&alloc_lock);
#else
blas_unlock(&memory[position].lock);
#endif
}
position ++;
} while (position < NUM_BUFFERS);
#endif #endif
position = 0; position = 0;
do { do {
#if defined(SMP) && !defined(USE_OPENMP)
LOCK_COMMAND(&alloc_lock);
#else
if (!memory[position].used) {
blas_lock(&memory[position].lock);
#endif
if (!memory[position].used) goto allocation; if (!memory[position].used) goto allocation;
#if defined(SMP) && !defined(USE_OPENMP)
UNLOCK_COMMAND(&alloc_lock);
#else
blas_unlock(&memory[position].lock);
}
#endif
position ++; position ++;
} while (position < NUM_BUFFERS); } while (position < BUFFERS_PER_THREAD);
goto error; goto error;
@ -1126,11 +1069,6 @@ void *blas_memory_alloc(int procpos){
#endif #endif
memory[position].used = 1; memory[position].used = 1;
#if defined(SMP) && !defined(USE_OPENMP)
UNLOCK_COMMAND(&alloc_lock);
#else
blas_unlock(&memory[position].lock);
#endif
if (!memory[position].addr) { if (!memory[position].addr) {
do { do {
@ -1148,14 +1086,14 @@ void *blas_memory_alloc(int procpos){
#ifdef ALLOC_DEVICEDRIVER #ifdef ALLOC_DEVICEDRIVER
if ((*func == alloc_devicedirver) && (map_address == (void *)-1)) { if ((*func == alloc_devicedirver) && (map_address == (void *)-1)) {
fprintf(stderr, "OpenBLAS Warning ... Physically contigous allocation was failed.\n"); fprintf(stderr, "OpenBLAS Warning ... Physically contiguous allocation failed.\n");
} }
#endif #endif
#ifdef ALLOC_HUGETLBFILE #ifdef ALLOC_HUGETLBFILE
if ((*func == alloc_hugetlbfile) && (map_address == (void *)-1)) { if ((*func == alloc_hugetlbfile) && (map_address == (void *)-1)) {
#ifndef OS_WINDOWS #ifndef OS_WINDOWS
fprintf(stderr, "OpenBLAS Warning ... HugeTLB(File) allocation was failed.\n"); fprintf(stderr, "OpenBLAS Warning ... HugeTLB(File) allocation failed.\n");
#endif #endif
} }
#endif #endif
@ -1176,44 +1114,13 @@ void *blas_memory_alloc(int procpos){
} while ((BLASLONG)map_address == -1); } while ((BLASLONG)map_address == -1);
#if defined(SMP) && !defined(USE_OPENMP)
LOCK_COMMAND(&alloc_lock);
#endif
memory[position].addr = map_address; memory[position].addr = map_address;
#if defined(SMP) && !defined(USE_OPENMP)
UNLOCK_COMMAND(&alloc_lock);
#endif
#ifdef DEBUG #ifdef DEBUG
printf(" Mapping Succeeded. %p(%d)\n", (void *)memory[position].addr, position); printf(" Mapping Succeeded. %p(%d)\n", (void *)memory[position].addr, position);
#endif #endif
} }
#if defined(WHEREAMI) && !defined(USE_OPENMP)
if (memory[position].pos == -1) memory[position].pos = mypos;
#endif
#ifdef DYNAMIC_ARCH
if (memory_initialized == 1) {
LOCK_COMMAND(&alloc_lock);
if (memory_initialized == 1) {
if (!gotoblas) gotoblas_dynamic_init();
memory_initialized = 2;
}
UNLOCK_COMMAND(&alloc_lock);
}
#endif
#ifdef DEBUG #ifdef DEBUG
printf("Mapped : %p %3d\n\n", printf("Mapped : %p %3d\n\n",
(void *)memory[position].addr, position); (void *)memory[position].addr, position);
@ -1222,7 +1129,7 @@ void *blas_memory_alloc(int procpos){
return (void *)memory[position].addr; return (void *)memory[position].addr;
error: error:
printf("BLAS : Program is Terminated. Because you tried to allocate too many memory regions.\n"); printf("OpenBLAS : Program will terminate because you tried to allocate too many memory regions.\n");
return NULL; return NULL;
} }
@ -1236,10 +1143,7 @@ void blas_memory_free(void *free_area){
#endif #endif
position = 0; position = 0;
#if defined(SMP) && !defined(USE_OPENMP) while ((position < BUFFERS_PER_THREAD) && (memory[position].addr != free_area))
LOCK_COMMAND(&alloc_lock);
#endif
while ((position < NUM_BUFFERS) && (memory[position].addr != free_area))
position++; position++;
if (memory[position].addr != free_area) goto error; if (memory[position].addr != free_area) goto error;
@ -1248,13 +1152,7 @@ void blas_memory_free(void *free_area){
printf(" Position : %d\n", position); printf(" Position : %d\n", position);
#endif #endif
// arm: ensure all writes are finished before other thread takes this memory
WMB;
memory[position].used = 0; memory[position].used = 0;
#if defined(SMP) && !defined(USE_OPENMP)
UNLOCK_COMMAND(&alloc_lock);
#endif
#ifdef DEBUG #ifdef DEBUG
printf("Unmap Succeeded.\n\n"); printf("Unmap Succeeded.\n\n");
@ -1266,11 +1164,8 @@ void blas_memory_free(void *free_area){
printf("BLAS : Bad memory unallocation! : %4d %p\n", position, free_area); printf("BLAS : Bad memory unallocation! : %4d %p\n", position, free_area);
#ifdef DEBUG #ifdef DEBUG
for (position = 0; position < NUM_BUFFERS; position++) for (position = 0; position < BUFFERS_PER_THREAD; position++)
printf("%4ld %p : %d\n", position, memory[position].addr, memory[position].used); printf("%4ld %p : %d\n", position, memory[position].addr, memory[position].used);
#endif
#if defined(SMP) && !defined(USE_OPENMP)
UNLOCK_COMMAND(&alloc_lock);
#endif #endif
return; return;
} }
@ -1293,8 +1188,6 @@ void blas_shutdown(void){
BLASFUNC(blas_thread_shutdown)(); BLASFUNC(blas_thread_shutdown)();
#endif #endif
LOCK_COMMAND(&alloc_lock);
for (pos = 0; pos < release_pos; pos ++) { for (pos = 0; pos < release_pos; pos ++) {
release_info[pos].func(&release_info[pos]); release_info[pos].func(&release_info[pos]);
} }
@ -1305,17 +1198,11 @@ void blas_shutdown(void){
base_address = BASE_ADDRESS; base_address = BASE_ADDRESS;
#endif #endif
for (pos = 0; pos < NUM_BUFFERS; pos ++){ for (pos = 0; pos < BUFFERS_PER_THREAD; pos ++){
memory[pos].addr = (void *)0; memory[pos].addr = (void *)0;
memory[pos].used = 0; memory[pos].used = 0;
#if defined(WHEREAMI) && !defined(USE_OPENMP)
memory[pos].pos = -1;
#endif
memory[pos].lock = 0;
} }
UNLOCK_COMMAND(&alloc_lock);
return; return;
} }