From f49f8047acbea636eb2a3542f306803a1285793b Mon Sep 17 00:00:00 2001
From: wjc404 <52632443+wjc404@users.noreply.github.com>
Date: Sat, 20 Jul 2019 14:33:37 +0800
Subject: [PATCH] Add files via upload

---
 kernel/x86_64/dgemm_kernel_4x8_haswell.S | 50 ++++++++++++++++++++----
 1 file changed, 43 insertions(+), 7 deletions(-)

diff --git a/kernel/x86_64/dgemm_kernel_4x8_haswell.S b/kernel/x86_64/dgemm_kernel_4x8_haswell.S
index 225af3673..6d1460bb2 100644
--- a/kernel/x86_64/dgemm_kernel_4x8_haswell.S
+++ b/kernel/x86_64/dgemm_kernel_4x8_haswell.S
@@ -279,30 +279,45 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	vmulpd	%ymm0 , %ymm9 , %ymm9
 	vmulpd	%ymm0 , %ymm10, %ymm10
 	vmulpd	%ymm0 , %ymm11, %ymm11
+#if B_PR1 >= 96
         prefetcht0      128 + BUFFER1
+#endif
 	vmulpd	%ymm0 , %ymm12, %ymm12
 	vmulpd	%ymm0 , %ymm13, %ymm13
 	vmulpd	%ymm0 , %ymm14, %ymm14
 	vmulpd	%ymm0 , %ymm15, %ymm15
+#if B_PR1 >= 160
         prefetcht0      192 + BUFFER1
+#endif
 	vpermilpd $ 0x05 , %ymm5, %ymm5
 	vpermilpd $ 0x05 , %ymm7, %ymm7
-
+#if B_PR1 >= 224
+        prefetcht0      256 + BUFFER1
+#endif
 	vblendpd $ 0x0a, %ymm5, %ymm4, %ymm0
 	vblendpd $ 0x05, %ymm5, %ymm4, %ymm1
 	vblendpd $ 0x0a, %ymm7, %ymm6, %ymm2
 	vblendpd $ 0x05, %ymm7, %ymm6, %ymm3
-
+#if B_PR1 >= 288
+        prefetcht0      320 + BUFFER1
+#endif
 	vperm2f128 $ 0x01 , %ymm2, %ymm2 , %ymm2
 	vperm2f128 $ 0x01 , %ymm3, %ymm3 , %ymm3
-
+#if B_PR1 >= 352
+        prefetcht0      384 + BUFFER1
+#endif
 	vblendpd $ 0x03, %ymm0, %ymm2 , %ymm4
 	vblendpd $ 0x03, %ymm1, %ymm3 , %ymm5
 	vblendpd $ 0x03, %ymm2, %ymm0 , %ymm6
 	vblendpd $ 0x03, %ymm3, %ymm1 , %ymm7
-
+#if B_PR1 >= 416
+        prefetcht0      448 + BUFFER1
+#endif
         leaq    (CO1, LDC, 2), %rax     
 	
+#if B_PR1 >= 480
+        prefetcht0      512 + BUFFER1
+#endif
 
 #if !defined(TRMMKERNEL)
 
@@ -1867,13 +1882,19 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
         /* here for the prefetch of next b source block */
 	/* the increment should be proportional to the ratio of GEMM_Q/GEMM_P */
-	/* currently an increment of 128 byte is suitable */
+
         salq    $3, K
+#ifdef WINDOWS_ABI /* GEMM_P == GEMM_Q * 4 */
+        prefetcht2 32(B)
+        prefetcht2 32(B, K, 8)
+        addq    $64, B /* increment */
+#else /* GEMM_P == GEMM_Q * 2 under linux x86_64 */
         prefetcht2 32(B)
         prefetcht2 32(B, K, 8)
         prefetcht2 96(B)
         prefetcht2 96(B, K, 8)
         addq    $128, B /* increment */
+#endif
         sarq    $3, K
 
 	decq	I			# i --
@@ -1883,10 +1904,15 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 /**************************************************************************
 * Rest of M 
 ***************************************************************************/
-        /* recover the original value of pointer B */
+
+        /* recover the original value of pointer B after prefetch */
         movq    M, I
         sarq    $2, I
+#ifdef WINDOWS_ABI /* GEMM_P == GEMM_Q * 4 */
+        salq    $6, I
+#else /* GEMM_P == GEMM_Q * 2 under linux x86_64 */
         salq    $7, I
+#endif
         subq    I, B
 
 .L12_20:
@@ -2166,13 +2192,19 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
         /* here for the prefetch of next b source block */
 	/* the increment should be proportional to the ratio of GEMM_Q/GEMM_P */
-	/* currently an increment of 128 byte is suitable */
+
         salq    $3, K
+#ifdef WINDOWS_ABI /* GEMM_P == GEMM_Q * 4 */
+        prefetcht2 (B)
+        prefetcht2 (B, K, 8)
+        addq    $64, B
+#else /* GEMM_P == GEMM_Q * 2 under linux x86_64 */
         prefetcht2 (B)
         prefetcht2 (B, K, 8)
         prefetcht2 64(B)
         prefetcht2 64(B, K, 8)
         addq    $128, B
+#endif
         sarq    $3, K
 
 	decq	I			# i --
@@ -2185,7 +2217,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         /* recover the original value of pointer B */
         movq    M, I
         sarq    $2, I
+#ifdef WINDOWS_ABI /* GEMM_P == GEMM_Q * 4 */
+        salq    $6, I
+#else /* GEMM_P == GEMM_Q * 2 under linux x86_64 */
         salq    $7, I
+#endif
         subq    I, B
 
 .L13_20: