Update dgemm_kernel_4x8_haswell.S

diff --git a/kernel/x86_64/dgemm_kernel_4x8_haswell.S b/kernel/x86_64/dgemm_kernel_4x8_haswell.S
index b986105..814a1c3 100644 (file)
--- a/kernel/x86_64/dgemm_kernel_4x8_haswell.S
+++ b/kernel/x86_64/dgemm_kernel_4x8_haswell.S
@@ -317,10 +317,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
        vmovups %ymm6 ,         (%rax)\r
        vmovups %ymm7 ,         (%rax, LDC)\r
 \r
-       prefetcht0      32(CO1)\r
-       prefetcht0      32(CO1,LDC)\r
-       prefetcht0      32(%rax)\r
-       prefetcht0      32(%rax,LDC)\r
+       prefetcht0      56(CO1)\r
+       prefetcht0      56(CO1,LDC)\r
+       prefetcht0      56(%rax)\r
+       prefetcht0      56(%rax,LDC)\r
 \r
        vpermilpd $ 0x05 , %ymm9 , %ymm9\r
        vpermilpd $ 0x05 , %ymm11, %ymm11\r
@@ -356,10 +356,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
        vmovups %ymm6 ,         (%rbp)\r
        vmovups %ymm7 ,         (%rbp, LDC)\r
 \r
-       prefetcht0      32(%rax)\r
-       prefetcht0      32(%rax,LDC)\r
-       prefetcht0      32(%rbp)\r
-       prefetcht0      32(%rbp,LDC)\r
+       prefetcht0      56(%rax)\r
+       prefetcht0      56(%rax,LDC)\r
+       prefetcht0      56(%rbp)\r
+       prefetcht0      56(%rbp,LDC)\r
 \r
        vpermilpd $ 0x05 , %ymm13, %ymm13\r
        vpermilpd $ 0x05 , %ymm15, %ymm15\r
@@ -395,10 +395,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
        vmovups %ymm6 ,         (%rbp)\r
        vmovups %ymm7 ,         (%rbp, LDC)\r
 \r
-       prefetcht0      32(%rax)\r
-       prefetcht0      32(%rax,LDC)\r
-       prefetcht0      32(%rbp)\r
-       prefetcht0      32(%rbp,LDC)\r
+       prefetcht0      56(%rax)\r
+       prefetcht0      56(%rax,LDC)\r
+       prefetcht0      56(%rbp)\r
+       prefetcht0      56(%rbp,LDC)\r
 \r
        addq    $ 4*SIZE, CO1\r
 .endm\r
@@ -826,10 +826,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
        vmovups %ymm6 ,         (%rax)\r
        vmovups %ymm7 ,         (%rax, LDC)\r
 \r
-       prefetcht0      32(CO1)\r
-       prefetcht0      32(CO1,LDC)\r
-       prefetcht0      32(%rax)\r
-       prefetcht0      32(%rax,LDC)\r
+       prefetcht0      56(CO1)\r
+       prefetcht0      56(CO1,LDC)\r
+       prefetcht0      56(%rax)\r
+       prefetcht0      56(%rax,LDC)\r
 \r
        vpermilpd $ 0x05 , %ymm9 , %ymm9\r
        vpermilpd $ 0x05 , %ymm11, %ymm11\r
@@ -865,10 +865,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
        vmovups %ymm6 ,         (%rbp)\r
        vmovups %ymm7 ,         (%rbp, LDC)\r
 \r
-       prefetcht0      32(%rax)\r
-       prefetcht0      32(%rax,LDC)\r
-       prefetcht0      32(%rbp)\r
-       prefetcht0      32(%rbp,LDC)\r
+       prefetcht0      56(%rax)\r
+       prefetcht0      56(%rax,LDC)\r
+       prefetcht0      56(%rbp)\r
+       prefetcht0      56(%rbp,LDC)\r
 \r
        addq    $ 4*SIZE, CO1\r
 .endm\r