For gemv_t, directly use malloc to create the buffer.
This commit is contained in:
parent
62231ab337
commit
9798481979
2
common.h
2
common.h
|
@ -499,6 +499,8 @@ void blas_set_parameter(void);
|
|||
int blas_get_cpu_number(void);
|
||||
void *blas_memory_alloc (int);
|
||||
void blas_memory_free (void *);
|
||||
void *blas_memory_alloc_nolock (int); //use malloc without blas_lock
|
||||
void blas_memory_free_nolock (void *);
|
||||
|
||||
int get_num_procs (void);
|
||||
|
||||
|
|
|
@ -1161,6 +1161,16 @@ void blas_memory_free(void *free_area){
|
|||
return;
|
||||
}
|
||||
|
||||
void *blas_memory_alloc_nolock(int unused) {
|
||||
void *map_address;
|
||||
map_address = (void *)malloc(BUFFER_SIZE + FIXED_PAGESIZE);
|
||||
return map_address;
|
||||
}
|
||||
|
||||
void blas_memory_free_nolock(void * map_address) {
|
||||
free(map_address);
|
||||
}
|
||||
|
||||
void blas_shutdown(void){
|
||||
|
||||
int pos;
|
||||
|
|
|
@ -211,15 +211,24 @@ void CNAME(enum CBLAS_ORDER order,
|
|||
#ifdef MAX_STACK_ALLOC
|
||||
// make it volatile because some gemv implementation (ex: dgemv_n.S)
|
||||
// do not restore all register
|
||||
volatile int stack_alloc_size = m + n;
|
||||
if(stack_alloc_size < 128)
|
||||
volatile int stack_alloc_size = 0;
|
||||
if (trans == 0) {
|
||||
stack_alloc_size = m + n;
|
||||
if(stack_alloc_size < 128)
|
||||
//dgemv_n.S require a 128 bytes buffer
|
||||
stack_alloc_size = 128;
|
||||
if(stack_alloc_size > MAX_STACK_ALLOC / sizeof(FLOAT))
|
||||
|
||||
if(stack_alloc_size > MAX_STACK_ALLOC / sizeof(FLOAT))
|
||||
stack_alloc_size = 0;
|
||||
FLOAT stack_buffer[stack_alloc_size];
|
||||
buffer = stack_alloc_size ? stack_buffer : (FLOAT *)blas_memory_alloc(1);
|
||||
FLOAT stack_buffer[stack_alloc_size];
|
||||
buffer = stack_alloc_size ? stack_buffer : (FLOAT *)blas_memory_alloc_nolock(1);
|
||||
|
||||
}else{
|
||||
//for gemv_t, only malloc
|
||||
buffer = (FLOAT *)blas_memory_alloc_nolock(1);
|
||||
}
|
||||
#else
|
||||
//Original OpenBLAS/GotoBLAS codes.
|
||||
buffer = (FLOAT *)blas_memory_alloc(1);
|
||||
#endif
|
||||
|
||||
|
@ -251,10 +260,13 @@ void CNAME(enum CBLAS_ORDER order,
|
|||
#endif
|
||||
|
||||
#ifdef MAX_STACK_ALLOC
|
||||
if(!stack_alloc_size)
|
||||
#endif
|
||||
if(!stack_alloc_size){
|
||||
blas_memory_free_nolock(buffer);
|
||||
}
|
||||
#else
|
||||
blas_memory_free(buffer);
|
||||
|
||||
#endif
|
||||
|
||||
FUNCTION_PROFILE_END(1, m * n + m + n, 2 * m * n);
|
||||
|
||||
IDEBUG_END;
|
||||
|
|
Loading…
Reference in New Issue