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

POWER10: Improve axpy performance

Browse Source 
This patch aligns the stores to 32 byte boundary for saxpy and daxpy
before entering into vector pair loop. Fox caxpy, changed the store
instructions to stxv to improve performance of unaligned cases.
This commit is contained in:
Rajalakshmi Srinivasaraghavan
2020-12-10 11:51:42 -06:00
parent 83de62c20d
commit 346e30a46a
3 changed files with 37 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
kernel/power/daxpy_power10.c
+11 -4
View File
@@ -66,12 +66,19 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
if ( (inc_x == 1) && (inc_y == 1) )
{
BLASLONG n1 = n & -16;
if ( n >= 16 )
{
BLASLONG align = ((32 - ((uintptr_t)y & (uintptr_t)0x1F)) >> 3) & 0x3;
for (i = 0; i < align; i++) {
y[i] += da * x[i] ;
}
}
BLASLONG n1 = (n-i) & -16;
if ( n1 )
daxpy_kernel_8(n1, &x[i], &y[i], da);
if ( n1 )
daxpy_kernel_8(n1, x, y, da);
i += n1;
i = n1;
while(i < n)
{