From 9c89757562f43af48645a6563161909321077646 Mon Sep 17 00:00:00 2001
From: wjc404 <52632443+wjc404@users.noreply.github.com>
Date: Fri, 19 Jul 2019 23:47:58 +0800
Subject: [PATCH] Add files via upload

---
 kernel/x86_64/dgemm_kernel_4x8_haswell.S | 29 +++++++++++++++++++++++-
 1 file changed, 28 insertions(+), 1 deletion(-)

diff --git a/kernel/x86_64/dgemm_kernel_4x8_haswell.S b/kernel/x86_64/dgemm_kernel_4x8_haswell.S
index 42692f33b..e26bddea3 100644
--- a/kernel/x86_64/dgemm_kernel_4x8_haswell.S
+++ b/kernel/x86_64/dgemm_kernel_4x8_haswell.S
@@ -1865,6 +1865,14 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 	SAVE4x12
 
+        salq    $3, K
+        prefetcht2 32(B)
+        prefetcht2 32(B, K, 8)
+        prefetcht2 96(B)
+        prefetcht2 96(B, K, 8)
+        addq    $128, B
+        sarq    $3, K
+
 	decq	I			# i --
 	jne	.L12_11
 	ALIGN_4	
@@ -1872,6 +1880,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 /**************************************************************************
 * Rest of M 
 ***************************************************************************/
+        movq    M, I
+        sarq    $2, I
+        salq    $7, I
+        subq    I, B
+
 .L12_20:
 	// Test rest of M
 
@@ -2102,7 +2115,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 	jmp .L13_16
 
-
+        PREFETCHT0_C
 .L13_13:
 
 	test $1, %rax
@@ -2147,6 +2160,14 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 	SAVE4x12
 
+        salq    $3, K
+        prefetcht2 (B)
+        prefetcht2 (B, K, 8)
+        prefetcht2 64(B)
+        prefetcht2 64(B, K, 8)
+        addq    $128, B
+        sarq    $3, K
+
 	decq	I			# i --
 	jne	.L13_11
 	ALIGN_4	
@@ -2154,6 +2175,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 /**************************************************************************
 * Rest of M 
 ***************************************************************************/
+
+        movq    M, I
+        sarq    $2, I
+        salq    $7, I
+        subq    I, B
+
 .L13_20:
 	// Test rest of M