From c1a6374c6fe7df294aeca2c550bc58d61acfa654 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Wed, 13 Aug 2014 16:10:03 +0200
Subject: [PATCH] optimized zgemv_n kernel for sandybridge

---
 kernel/x86_64/KERNEL.SANDYBRIDGE       |  3 +++
 kernel/x86_64/zgemv_n.c                |  3 +++
 kernel/x86_64/zgemv_n_microk_sandy-2.c | 13 ++++++++-----
 3 files changed, 14 insertions(+), 5 deletions(-)

diff --git a/kernel/x86_64/KERNEL.SANDYBRIDGE b/kernel/x86_64/KERNEL.SANDYBRIDGE
index d4fbca7f2..b654d3564 100644
--- a/kernel/x86_64/KERNEL.SANDYBRIDGE
+++ b/kernel/x86_64/KERNEL.SANDYBRIDGE
@@ -1,6 +1,9 @@
 SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t.c
 
+ZGEMVNKERNEL = zgemv_n.c
+
+
 SGEMMKERNEL    =  sgemm_kernel_16x4_sandy.S
 SGEMMINCOPY    =  ../generic/gemm_ncopy_16.c
 SGEMMITCOPY    =  ../generic/gemm_tcopy_16.c
diff --git a/kernel/x86_64/zgemv_n.c b/kernel/x86_64/zgemv_n.c
index 7b8907044..9098368a5 100644
--- a/kernel/x86_64/zgemv_n.c
+++ b/kernel/x86_64/zgemv_n.c
@@ -31,9 +31,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #if defined(HASWELL)
 #include "zgemv_n_microk_haswell-2.c"
+#elif defined(SANDYBRIDGE)
+#include "zgemv_n_microk_sandy-2.c"
 #endif
 
 
+
 #define NBMAX 1024
 
 #ifndef HAVE_KERNEL_16x4
diff --git a/kernel/x86_64/zgemv_n_microk_sandy-2.c b/kernel/x86_64/zgemv_n_microk_sandy-2.c
index f90e2210a..352c60f87 100644
--- a/kernel/x86_64/zgemv_n_microk_sandy-2.c
+++ b/kernel/x86_64/zgemv_n_microk_sandy-2.c
@@ -50,39 +50,42 @@ static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	".align 16				        \n\t"
 	".L01LOOP%=:				        \n\t"
 
+        "prefetcht0      256(%4,%0,8)                   \n\t"
 	"vmovups	(%4,%0,8), %%ymm8	        \n\t" // 2 complex values form a0
 	"vmovups      32(%4,%0,8), %%ymm9	        \n\t" // 2 complex values form a0
 
 	"vmulpd		  %%ymm8 , %%ymm0 , %%ymm12	\n\t"
 	"vmulpd		  %%ymm8 , %%ymm1 , %%ymm13	\n\t"
+        "prefetcht0      256(%5,%0,8)                   \n\t"
 	"vmulpd		  %%ymm9 , %%ymm0 , %%ymm14	\n\t"
-	"vmulpd		  %%ymm9 , %%ymm1 , %%ymm15	\n\t"
-
 	"vmovups	(%5,%0,8), %%ymm8	        \n\t" // 2 complex values form a0
+	"vmulpd		  %%ymm9 , %%ymm1 , %%ymm15	\n\t"
 	"vmovups      32(%5,%0,8), %%ymm9	        \n\t" // 2 complex values form a0
 
 	"vmulpd		  %%ymm8 , %%ymm2 , %%ymm10	\n\t"
 	"vaddpd		  %%ymm12, %%ymm10, %%ymm12	\n\t"
 	"vmulpd		  %%ymm8 , %%ymm3 , %%ymm11	\n\t"
 	"vaddpd		  %%ymm13, %%ymm11, %%ymm13	\n\t"
+        "prefetcht0      256(%6,%0,8)                   \n\t"
 	"vmulpd		  %%ymm9 , %%ymm2 , %%ymm10	\n\t"
 	"vaddpd		  %%ymm14, %%ymm10, %%ymm14	\n\t"
+	"vmovups	(%6,%0,8), %%ymm8	        \n\t" // 2 complex values form a0
 	"vmulpd		  %%ymm9 , %%ymm3 , %%ymm11	\n\t"
 	"vaddpd		  %%ymm15, %%ymm11, %%ymm15	\n\t"
 
-	"vmovups	(%6,%0,8), %%ymm8	        \n\t" // 2 complex values form a0
 	"vmovups      32(%6,%0,8), %%ymm9	        \n\t" // 2 complex values form a0
 
 	"vmulpd		  %%ymm8 , %%ymm4 , %%ymm10	\n\t"
 	"vaddpd		  %%ymm12, %%ymm10, %%ymm12	\n\t"
 	"vmulpd		  %%ymm8 , %%ymm5 , %%ymm11	\n\t"
 	"vaddpd		  %%ymm13, %%ymm11, %%ymm13	\n\t"
+        "prefetcht0      256(%7,%0,8)                   \n\t"
 	"vmulpd		  %%ymm9 , %%ymm4 , %%ymm10	\n\t"
 	"vaddpd		  %%ymm14, %%ymm10, %%ymm14	\n\t"
+	"vmovups	(%7,%0,8), %%ymm8	        \n\t" // 2 complex values form a0
 	"vmulpd		  %%ymm9 , %%ymm5 , %%ymm11	\n\t"
 	"vaddpd		  %%ymm15, %%ymm11, %%ymm15	\n\t"
 
-	"vmovups	(%7,%0,8), %%ymm8	        \n\t" // 2 complex values form a0
 	"vmovups      32(%7,%0,8), %%ymm9	        \n\t" // 2 complex values form a0
 
 	"vmulpd		  %%ymm8 , %%ymm6 , %%ymm10	\n\t"
@@ -94,7 +97,7 @@ static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	"vmulpd		  %%ymm9 , %%ymm7 , %%ymm11	\n\t"
 	"vaddpd		  %%ymm15, %%ymm11, %%ymm15	\n\t"
 
-	"prefetcht0      192(%3,%0,8)			\n\t"
+	"prefetcht0      256(%3,%0,8)			\n\t"
 	"vmovups	  (%3,%0,8),  %%ymm10           \n\t"
 	"vmovups	32(%3,%0,8),  %%ymm11           \n\t"