Split the microkernel workload into chunks of 32 floats for dsdot mode to limit loss of precision
This commit is contained in:
parent
28c3fa8950
commit
5e3e91d0fc
|
@ -78,7 +78,12 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
|
|||
BLASLONG ix=0,iy=0;
|
||||
double dot = 0.0 ;
|
||||
|
||||
#if defined (DSDOT)
|
||||
double mydot = 0.0;
|
||||
FLOAT asmdot = 0.0;
|
||||
#else
|
||||
FLOAT mydot=0.0;
|
||||
#endif
|
||||
BLASLONG n1;
|
||||
|
||||
if ( n <= 0 ) return(dot);
|
||||
|
@ -89,9 +94,23 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
|
|||
n1 = n & (BLASLONG)(-32);
|
||||
|
||||
if ( n1 )
|
||||
#if defined(DSDOT)
|
||||
{
|
||||
FLOAT *x1=x;
|
||||
FLOAT *y1=y;
|
||||
BLASLONG n2 = 32;
|
||||
while (i<n1) {
|
||||
sdot_kernel_16(n2, x1, y1 , &asmdot );
|
||||
mydot += (double)asmdot;
|
||||
asmdot=0.;
|
||||
x1+=32;
|
||||
y1+=32;
|
||||
i+=32;
|
||||
}
|
||||
}
|
||||
#else
|
||||
sdot_kernel_16(n1, x, y , &mydot );
|
||||
|
||||
|
||||
#endif
|
||||
i = n1;
|
||||
while(i < n)
|
||||
{
|
||||
|
@ -103,11 +122,8 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
|
|||
i++ ;
|
||||
|
||||
}
|
||||
#if defined(DSDOT)
|
||||
dot+=(double)mydot;
|
||||
#else
|
||||
|
||||
dot+=mydot;
|
||||
#endif
|
||||
return(dot);
|
||||
|
||||
|
||||
|
|
Loading…
Reference in New Issue