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 b30eccc..3f7f9a9 100644 (file)
--- a/kernel/x86_64/dgemm_kernel_4x8_haswell.S
+++ b/kernel/x86_64/dgemm_kernel_4x8_haswell.S
@@ -267,23 +267,24 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 \r
 .macro SAVE4x12\r
 \r
+        prefetcht0      128(%rsp) /*BUFFER 1*/\r
        vbroadcastsd    ALPHA, %ymm0\r
 \r
        vmulpd  %ymm0 , %ymm4 , %ymm4\r
        vmulpd  %ymm0 , %ymm5 , %ymm5\r
        vmulpd  %ymm0 , %ymm6 , %ymm6\r
        vmulpd  %ymm0 , %ymm7 , %ymm7\r
-\r
+        prefetcht0      192(%rsp)\r
        vmulpd  %ymm0 , %ymm8 , %ymm8\r
        vmulpd  %ymm0 , %ymm9 , %ymm9\r
        vmulpd  %ymm0 , %ymm10, %ymm10\r
        vmulpd  %ymm0 , %ymm11, %ymm11\r
-\r
+        prefetcht0      256(%rsp)\r
        vmulpd  %ymm0 , %ymm12, %ymm12\r
        vmulpd  %ymm0 , %ymm13, %ymm13\r
        vmulpd  %ymm0 , %ymm14, %ymm14\r
        vmulpd  %ymm0 , %ymm15, %ymm15\r
-\r
+        prefetcht0      320(%rsp)\r
        vpermilpd $ 0x05 , %ymm5, %ymm5\r
        vpermilpd $ 0x05 , %ymm7, %ymm7\r
 \r