From 8109d8232c66c4119044fa3111947d88afae9eb6 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Tue, 9 Sep 2014 14:38:08 +0200
Subject: [PATCH] optimized dgemv_t kernel for haswell

---
 kernel/x86_64/dgemv_t_4.c                | 10 +---------
 kernel/x86_64/dgemv_t_microk_haswell-4.c | 22 +++++++++++-----------
 2 files changed, 12 insertions(+), 20 deletions(-)

diff --git a/kernel/x86_64/dgemv_t_4.c b/kernel/x86_64/dgemv_t_4.c
index 0d0409bec..ebec7d2c3 100644
--- a/kernel/x86_64/dgemv_t_4.c
+++ b/kernel/x86_64/dgemv_t_4.c
@@ -28,17 +28,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include "common.h"
 
-/*
-#if defined(NEHALEM)
-#include "dgemv_t_microk_nehalem-4.c"
-#elif defined(BULLDOZER) || defined(PILEDRIVER)
-#include "dgemv_t_microk_bulldozer-4.c"
-#elif defined(SANDYBRIDGE)
-#include "dgemv_t_microk_sandy-4.c"
-#elif defined(HASWELL)
+#if defined(HASWELL)
 #include "dgemv_t_microk_haswell-4.c"
 #endif
-*/
 
 #define NBMAX 2048
 
diff --git a/kernel/x86_64/dgemv_t_microk_haswell-4.c b/kernel/x86_64/dgemv_t_microk_haswell-4.c
index 410225500..33b43515d 100644
--- a/kernel/x86_64/dgemv_t_microk_haswell-4.c
+++ b/kernel/x86_64/dgemv_t_microk_haswell-4.c
@@ -61,25 +61,25 @@ static void dgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 
 	".align 16				 \n\t"
 	".L01LOOP%=:				 \n\t"
-	"prefetcht0	 384(%2,%0,8)		 \n\t"
+	// "prefetcht0	 384(%2,%0,8)		 \n\t"
 	"vmovups	(%2,%0,8), %%ymm12       \n\t"	// 4 * x
 	"vmovups      32(%2,%0,8), %%ymm13       \n\t"	// 4 * x
 
-	"prefetcht0	 384(%4,%0,8)		       \n\t"
+	// "prefetcht0	 384(%4,%0,8)		       \n\t"
 	"vfmadd231pd   (%4,%0,8), %%ymm12, %%ymm4      \n\t" 
 	"vfmadd231pd   (%5,%0,8), %%ymm12, %%ymm5      \n\t" 
-	"prefetcht0	 384(%5,%0,8)		       \n\t"
-	"vfmadd231pd 32(%4,%0,8), %%ymm13, %%ymm4      \n\t" 
-	"vfmadd231pd 32(%5,%0,8), %%ymm13, %%ymm5      \n\t" 
-	"prefetcht0	 384(%6,%0,8)		       \n\t"
+	// "prefetcht0	 384(%5,%0,8)		       \n\t"
 	"vfmadd231pd   (%6,%0,8), %%ymm12, %%ymm6      \n\t" 
 	"vfmadd231pd   (%7,%0,8), %%ymm12, %%ymm7      \n\t" 
-	"prefetcht0	 384(%7,%0,8)		       \n\t"
-	"vfmadd231pd 32(%6,%0,8), %%ymm13, %%ymm6      \n\t" 
-	"vfmadd231pd 32(%7,%0,8), %%ymm13, %%ymm7      \n\t" 
+	// "prefetcht0	 384(%6,%0,8)		       \n\t"
+	"vfmadd231pd 32(%4,%0,8), %%ymm13, %%ymm4      \n\t" 
+	"vfmadd231pd 32(%5,%0,8), %%ymm13, %%ymm5      \n\t" 
+        "addq		$8 , %0	  	 	       \n\t"
+	// "prefetcht0	 384(%7,%0,8)		       \n\t"
+	"vfmadd231pd -32(%6,%0,8), %%ymm13, %%ymm6     \n\t" 
+	"subq	        $8 , %1			       \n\t"		
+	"vfmadd231pd -32(%7,%0,8), %%ymm13, %%ymm7     \n\t" 
 
-        "addq		$8 , %0	  	 	      \n\t"
-	"subq	        $8 , %1			      \n\t"		
 	"jnz		.L01LOOP%=		      \n\t"
 
         ".L16END%=:                                   \n\t"