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

LoongArch64: Fixed {s/d}axpby

This commit is contained in:
gxw 2024-02-01 19:57:05 +08:00
parent 3597827c93
commit 1e1f487dc7
2 changed files with 14 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
kernel/loongarch64/axpby_lasx.S
View File

@ -57,10 +57,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
PROLOGUE PROLOGUE
bge $r0, N, .L999 bge $r0, N, .L999
li.d TEMP, 1
movgr2fr.d a1, $r0 movgr2fr.d a1, $r0
ffint.s.l a1, a1 ffint.s.l a1, a1
slli.d TEMP, TEMP, BASE_SHIFT
slli.d INCX, INCX, BASE_SHIFT slli.d INCX, INCX, BASE_SHIFT
slli.d INCY, INCY, BASE_SHIFT slli.d INCY, INCY, BASE_SHIFT
MTG t1, ALPHA MTG t1, ALPHA
@ -75,6 +73,13 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
xvreplgr2vr.w VXB, t2 xvreplgr2vr.w VXB, t2
xvreplgr2vr.w VXZ, t3 xvreplgr2vr.w VXZ, t3
#endif #endif
// If incx == 0 || incy == 0, do one by one
and TEMP, INCX, INCY
or I, N, N
beqz TEMP, .L998
li.d TEMP, 1
slli.d TEMP, TEMP, BASE_SHIFT
srai.d I, N, 3 srai.d I, N, 3
bne INCX, TEMP, .L20 bne INCX, TEMP, .L20
bne INCY, TEMP, .L12 // INCX==1 and INCY!=1 bne INCY, TEMP, .L12 // INCX==1 and INCY!=1

9
kernel/loongarch64/axpby_lsx.S
View File

@ -57,10 +57,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
PROLOGUE PROLOGUE
bge $r0, N, .L999 bge $r0, N, .L999
li.d TEMP, 1
movgr2fr.d a1, $r0 movgr2fr.d a1, $r0
ffint.s.l a1, a1 ffint.s.l a1, a1
slli.d TEMP, TEMP, BASE_SHIFT
slli.d INCX, INCX, BASE_SHIFT slli.d INCX, INCX, BASE_SHIFT
slli.d INCY, INCY, BASE_SHIFT slli.d INCY, INCY, BASE_SHIFT
MTG t1, ALPHA MTG t1, ALPHA
@ -75,6 +73,13 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
vreplgr2vr.w VXB, t2 vreplgr2vr.w VXB, t2
vreplgr2vr.w VXZ, t3 vreplgr2vr.w VXZ, t3
#endif #endif
// If incx == 0 || incy == 0, do one by one
and TEMP, INCX, INCY
or I, N, N
beqz TEMP, .L998
li.d TEMP, 1
slli.d TEMP, TEMP, BASE_SHIFT
srai.d I, N, 3 srai.d I, N, 3
bne INCX, TEMP, .L20 bne INCX, TEMP, .L20
bne INCY, TEMP, .L12 // INCX==1 and INCY!=1 bne INCY, TEMP, .L12 // INCX==1 and INCY!=1