Add files via upload

diff --git a/kernel/x86_64/dgemm_kernel_4x8_haswell.S b/kernel/x86_64/dgemm_kernel_4x8_haswell.S
index 6d1460b..6a8619e 100644 (file)
--- a/kernel/x86_64/dgemm_kernel_4x8_haswell.S
+++ b/kernel/x86_64/dgemm_kernel_4x8_haswell.S
@@ -1622,35 +1622,13 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 \r
 \r
 .macro PREFETCHT0_C\r
+        prefetcht0 ALPHA\r
         prefetcht0 (CO1)\r
         prefetcht0 24(CO1)\r
         prefetcht0 (CO1,LDC,4)\r
         prefetcht0 24(CO1,LDC,4)\r
         prefetcht0 (CO1,LDC,8)\r
         prefetcht0 24(CO1,LDC,8)\r
-        addq  LDC,CO1\r
-        prefetcht0 (CO1)\r
-        prefetcht0 24(CO1)\r
-        prefetcht0 (CO1,LDC,4)\r
-        prefetcht0 24(CO1,LDC,4)\r
-        prefetcht0 (CO1,LDC,8)\r
-        prefetcht0 24(CO1,LDC,8)\r
-        leaq  (CO1,LDC,2),CO1\r
-        prefetcht0 (CO1)\r
-        prefetcht0 24(CO1)\r
-        prefetcht0 (CO1,LDC,4)\r
-        prefetcht0 24(CO1,LDC,4)\r
-        prefetcht0 (CO1,LDC,8)\r
-        prefetcht0 24(CO1,LDC,8)\r
-        subq  LDC,CO1\r
-        prefetcht0 (CO1)\r
-        prefetcht0 24(CO1)\r
-        prefetcht0 (CO1,LDC,4)\r
-        prefetcht0 24(CO1,LDC,4)\r
-        prefetcht0 (CO1,LDC,8)\r
-        prefetcht0 24(CO1,LDC,8)\r
-        subq  LDC,CO1\r
-        subq  LDC,CO1\r
 .endm\r
 /*******************************************************************************************/\r
 \r
@@ -1820,12 +1798,19 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
        dec     %rax\r
        jne     .L12_12\r
        \r
-        PREFETCHT0_C\r
 .L12_12a:\r
-\r
+        PREFETCHT0_C\r
+        addq  LDC,CO1\r
        KERNEL4x12_M1\r
+        PREFETCHT0_C\r
+        leaq  (CO1,LDC,2),CO1\r
        KERNEL4x12_M2\r
+        PREFETCHT0_C\r
+        subq  LDC,CO1\r
        KERNEL4x12_M1\r
+        PREFETCHT0_C\r
+        subq  LDC,CO1\r
+        subq  LDC,CO1\r
        KERNEL4x12_M2\r
 \r
        KERNEL4x12_M1\r
@@ -2133,9 +2118,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 \r
 .L13_12a:\r
 \r
+        PREFETCHT0_C\r
+        addq  LDC,CO1\r
        KERNEL4x12_M1\r
+        PREFETCHT0_C\r
+        leaq  (CO1,LDC,2),CO1\r
        KERNEL4x12_M2\r
+        PREFETCHT0_C\r
+        subq  LDC,CO1\r
        KERNEL4x12_M1\r
+        PREFETCHT0_C\r
+        subq  LDC,CO1\r
+        subq  LDC,CO1\r
        KERNEL4x12_M2\r
 \r
        KERNEL4x12_M1\r
@@ -2145,7 +2139,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 \r
        jmp .L13_16\r
 \r
-        PREFETCHT0_C\r
 .L13_13:\r
 \r
        test $1, %rax\r