Browse Source

fix sve ztrmm kernel

tags/v0.3.20
Bine Brank 4 years ago
parent
commit
2e2c02b762
1 changed files with 4 additions and 4 deletions
  1. +4
    -4
      kernel/arm64/ztrmm_kernel_sve_v1x4.S

+ 4
- 4
kernel/arm64/ztrmm_kernel_sve_v1x4.S View File

@@ -723,7 +723,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
add pB, pB, temp
#endif
#if defined(LEFT)
add tempOffset, tempOffset, #4
add tempOffset, tempOffset, lanes
#endif

prfm PLDL1KEEP, [pA]
@@ -856,7 +856,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
add pB, pB, temp
#endif
#if defined(LEFT)
add tempOffset, tempOffset, #4
add tempOffset, tempOffset, lanes
#endif

.Lztrmm_kernel_L2_Mv1_END:
@@ -923,7 +923,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
add tempK, tempOffset, #1
#endif

asr counterL , temp, #3 // counterL = counterL / 8
asr counterL , tempK, #3 // counterL = counterL / 8
cmp counterL , #0
ble .Lztrmm_kernel_L1_Mv1_40
.align 5
@@ -972,7 +972,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
add pB, pB, temp
#endif
#if defined(LEFT)
add tempOffset, tempOffset, #4
add tempOffset, tempOffset, lanes
#endif

.Lztrmm_kernel_L1_Mv1_END:


Loading…
Cancel
Save