Allow to do gemv and ger buffer allocation on the stack

ger and gemv call blas_memory_alloc/free which in their turn
call blas_lock. blas_lock create thread contention when matrices
are small and the number of thread is high enough. We avoid
call blas_memory_alloc by replacing it with stack allocation.
This can be enabled with:
make -DMAX_STACK_ALLOC=2048
The given size (in byte) must be high enough to avoid thread contention
and small enough to avoid stack overflow.

Fix #478
This commit is contained in:
Jerome Robert 2014-12-26 14:42:00 +01:00
parent cbb3ab80e7
commit e9d9a8eae3
4 changed files with 32 additions and 3 deletions

View File

@ -305,6 +305,10 @@ ifdef SANITY_CHECK
CCOMMON_OPT += -DSANITY_CHECK -DREFNAME=$(*F)f$(BU)
endif
ifdef MAX_STACK_ALLOC
CCOMMON_OPT += -DMAX_STACK_ALLOC=$(MAX_STACK_ALLOC)
endif
#
# Architecture dependent settings
#

View File

@ -208,7 +208,18 @@ void CNAME(enum CBLAS_ORDER order,
if (incx < 0) x -= (lenx - 1) * incx;
if (incy < 0) y -= (leny - 1) * incy;
#ifdef MAX_STACK_ALLOC
int 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))
stack_alloc_size = 0;
FLOAT stack_buffer[stack_alloc_size];
buffer = stack_alloc_size ? stack_buffer : (FLOAT *)blas_memory_alloc(1);
#else
buffer = (FLOAT *)blas_memory_alloc(1);
#endif
#ifdef SMP
@ -237,7 +248,10 @@ void CNAME(enum CBLAS_ORDER order,
}
#endif
blas_memory_free(buffer);
#ifdef MAX_STACK_ALLOC
if(!stack_alloc_size)
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(1, m * n + m + n, 2 * m * n);

View File

@ -171,7 +171,15 @@ void CNAME(enum CBLAS_ORDER order,
if (incy < 0) y -= (n - 1) * incy;
if (incx < 0) x -= (m - 1) * incx;
#ifdef MAX_STACK_ALLOC
int stack_alloc_size = m;
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);
#else
buffer = (FLOAT *)blas_memory_alloc(1);
#endif
#ifdef SMPTEST
nthreads = num_cpu_avail(2);
@ -190,7 +198,10 @@ void CNAME(enum CBLAS_ORDER order,
}
#endif
blas_memory_free(buffer);
#ifdef MAX_STACK_ALLOC
if(!stack_alloc_size)
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(1, m * n + m + n, 2 * m * n);

View File

@ -302,7 +302,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
if ( n < 1 ) return(0);
xbuffer = buffer;
ytemp = buffer + NBMAX;
ytemp = buffer + (m < NBMAX ? m : NBMAX);
n0 = n / NBMAX;
n1 = (n % NBMAX) >> 2 ;