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

better optimzations for sgemv_t kernel

This commit is contained in:
wernsaar 2014-09-06 21:28:57 +02:00
parent c8eaf3ae2d
commit 5ae1731fe6
1 changed files with 99 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

87
kernel/x86_64/sgemv_t_4.c
View File

@ -446,12 +446,45 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
} }
else else
{ {
if ( inc_y == 1 )
{
BLASLONG register lda2 = lda << 1;
BLASLONG register lda4 = lda << 2;
BLASLONG register lda3 = lda2 + lda;
for ( j=0; j< ( n & -4 ); j+=4 )
{
y_ptr[j] += *aj * xtemp0 + *(aj+1) * xtemp1 + *(aj+2) * xtemp2;
y_ptr[j+1] += *(aj+lda) * xtemp0 + *(aj+lda+1) * xtemp1 + *(aj+lda+2) * xtemp2;
y_ptr[j+2] += *(aj+lda2) * xtemp0 + *(aj+lda2+1) * xtemp1 + *(aj+lda2+2) * xtemp2;
y_ptr[j+3] += *(aj+lda3) * xtemp0 + *(aj+lda3+1) * xtemp1 + *(aj+lda3+2) * xtemp2;
aj += lda4;
}
for ( ; j< n ; j++ )
{
y_ptr[j] += *aj * xtemp0 + *(aj+1) * xtemp1 + *(aj+2) * xtemp2 ;
aj += lda;
}
}
else
{
for ( j=0; j<n; j++ ) for ( j=0; j<n; j++ )
{ {
*y_ptr += *aj * xtemp0 + *(aj+1) * xtemp1 + *(aj+2) * xtemp2; *y_ptr += *aj * xtemp0 + *(aj+1) * xtemp1 + *(aj+2) * xtemp2;
y_ptr += inc_y; y_ptr += inc_y;
aj += lda; aj += lda;
} }
}
} }
return(0); return(0);
} }
@ -486,15 +519,41 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
} }
else else
{
if ( inc_y == 1 )
{ {
BLASLONG register lda2 = lda << 1;
BLASLONG register lda4 = lda << 2;
BLASLONG register lda3 = lda2 + lda;
for ( j=0; j< ( n & -4 ); j+=4 )
{
y_ptr[j] += *aj * xtemp0 + *(aj+1) * xtemp1 ;
y_ptr[j+1] += *(aj+lda) * xtemp0 + *(aj+lda+1) * xtemp1 ;
y_ptr[j+2] += *(aj+lda2) * xtemp0 + *(aj+lda2+1) * xtemp1 ;
y_ptr[j+3] += *(aj+lda3) * xtemp0 + *(aj+lda3+1) * xtemp1 ;
aj += lda4;
}
for ( ; j< n ; j++ )
{
y_ptr[j] += *aj * xtemp0 + *(aj+1) * xtemp1 ;
aj += lda;
}
}
else
{
for ( j=0; j<n; j++ ) for ( j=0; j<n; j++ )
{ {
*y_ptr += *aj * xtemp0 + *(aj+1) * xtemp1 ; *y_ptr += *aj * xtemp0 + *(aj+1) * xtemp1 ;
y_ptr += inc_y; y_ptr += inc_y;
aj += lda; aj += lda;
} }
}
} }
return(0); return(0);
@ -520,6 +579,30 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
}
else
{
if ( inc_y == 1 )
{
BLASLONG register lda2 = lda << 1;
BLASLONG register lda4 = lda << 2;
BLASLONG register lda3 = lda2 + lda;
for ( j=0; j< ( n & -4 ); j+=4 )
{
y_ptr[j] += *aj * xtemp;
y_ptr[j+1] += *(aj+lda) * xtemp;
y_ptr[j+2] += *(aj+lda2) * xtemp;
y_ptr[j+3] += *(aj+lda3) * xtemp;
aj += lda4 ;
}
for ( ; j<n; j++ )
{
y_ptr[j] += *aj * xtemp;
aj += lda;
}
} }
else else
{ {
@ -529,6 +612,8 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
y_ptr += inc_y; y_ptr += inc_y;
aj += lda; aj += lda;
} }
}
} }
return(0); return(0);