From 95a8caa2f340b5936ef3a8106f04df07a07e4d93 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Wed, 6 Aug 2014 12:12:17 +0200
Subject: [PATCH 01/27] added optimized sgemv_t kernel

---
 kernel/x86_64/KERNEL                       |   2 +-
 kernel/x86_64/KERNEL.BULLDOZER             |   2 +-
 kernel/x86_64/KERNEL.HASWELL               |   2 +-
 kernel/x86_64/KERNEL.NEHALEM               |   2 +-
 kernel/x86_64/KERNEL.PILEDRIVER            |   2 +-
 kernel/x86_64/KERNEL.SANDYBRIDGE           |   2 +-
 kernel/x86_64/sgemv_t.c                    | 200 +++++++++++++++++++++
 kernel/x86_64/sgemv_t_microk_bulldozer-2.c | 109 +++++++++++
 8 files changed, 315 insertions(+), 6 deletions(-)
 create mode 100644 kernel/x86_64/sgemv_t.c
 create mode 100644 kernel/x86_64/sgemv_t_microk_bulldozer-2.c

diff --git a/kernel/x86_64/KERNEL b/kernel/x86_64/KERNEL
index 9d0080ae7..3508753ee 100644
--- a/kernel/x86_64/KERNEL
+++ b/kernel/x86_64/KERNEL
@@ -377,7 +377,7 @@ SGEMVNKERNEL = sgemv_n.c
 endif
 
 ifndef SGEMVTKERNEL
-SGEMVTKERNEL = ../arm/gemv_t.c
+SGEMVTKERNEL = sgemv_t.c
 endif
 
 ifndef DGEMVNKERNEL
diff --git a/kernel/x86_64/KERNEL.BULLDOZER b/kernel/x86_64/KERNEL.BULLDOZER
index 21fc94701..accdddf0e 100644
--- a/kernel/x86_64/KERNEL.BULLDOZER
+++ b/kernel/x86_64/KERNEL.BULLDOZER
@@ -3,7 +3,7 @@ SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = ../arm/gemv_t.c
 else
 SGEMVNKERNEL = sgemv_n.c
-SGEMVTKERNEL = sgemv_t_avx.c
+SGEMVTKERNEL = sgemv_t.c
 endif
 
 
diff --git a/kernel/x86_64/KERNEL.HASWELL b/kernel/x86_64/KERNEL.HASWELL
index 3e20fcfc7..878a56b04 100644
--- a/kernel/x86_64/KERNEL.HASWELL
+++ b/kernel/x86_64/KERNEL.HASWELL
@@ -3,7 +3,7 @@ SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = ../arm/gemv_t.c
 else
 SGEMVNKERNEL = sgemv_n.c
-SGEMVTKERNEL = sgemv_t_avx.c
+SGEMVTKERNEL = sgemv_t.c
 endif
 
 
diff --git a/kernel/x86_64/KERNEL.NEHALEM b/kernel/x86_64/KERNEL.NEHALEM
index 04efa391a..8276150c6 100644
--- a/kernel/x86_64/KERNEL.NEHALEM
+++ b/kernel/x86_64/KERNEL.NEHALEM
@@ -3,7 +3,7 @@ SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = ../arm/gemv_t.c
 else
 SGEMVNKERNEL = sgemv_n.c
-SGEMVTKERNEL = ../arm/gemv_t.c
+SGEMVTKERNEL = sgemv_t.c
 endif
 
 
diff --git a/kernel/x86_64/KERNEL.PILEDRIVER b/kernel/x86_64/KERNEL.PILEDRIVER
index b7565edeb..7b3c9a7b8 100644
--- a/kernel/x86_64/KERNEL.PILEDRIVER
+++ b/kernel/x86_64/KERNEL.PILEDRIVER
@@ -3,7 +3,7 @@ SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = ../arm/gemv_t.c
 else
 SGEMVNKERNEL = sgemv_n.c
-SGEMVTKERNEL = sgemv_t_avx.c
+SGEMVTKERNEL = sgemv_t.c
 endif
 
 
diff --git a/kernel/x86_64/KERNEL.SANDYBRIDGE b/kernel/x86_64/KERNEL.SANDYBRIDGE
index 9dae3a41d..26706b61d 100644
--- a/kernel/x86_64/KERNEL.SANDYBRIDGE
+++ b/kernel/x86_64/KERNEL.SANDYBRIDGE
@@ -3,7 +3,7 @@ SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = ../arm/gemv_t.c
 else
 SGEMVNKERNEL = sgemv_n.c
-SGEMVTKERNEL = sgemv_t_avx.c
+SGEMVTKERNEL = sgemv_t.c
 endif
 
 
diff --git a/kernel/x86_64/sgemv_t.c b/kernel/x86_64/sgemv_t.c
new file mode 100644
index 000000000..89254c256
--- /dev/null
+++ b/kernel/x86_64/sgemv_t.c
@@ -0,0 +1,200 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include "common.h"
+
+#if defined(BULLDOZER) || defined(PILEDRIVER)
+#include "sgemv_t_microk_bulldozer-2.c"
+#endif
+
+/*
+#if defined(BULLDOZER) || defined(PILEDRIVER)
+#include "sgemv_n_microk_bulldozer-2.c"
+#elif defined(HASWELL)
+#include "sgemv_n_microk_haswell-2.c"
+#elif defined(SANDYBRIDGE)
+#include "sgemv_n_microk_sandy-2.c"
+#elif defined(NEHALEM)
+#include "sgemv_n_microk_nehalem-2.c"
+#endif
+*/
+
+#define NBMAX 4096
+
+#ifndef HAVE_KERNEL_16x4
+
+static void sgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0,*a1,*a2,*a3;
+	a0 = ap[0];
+	a1 = ap[1];
+	a2 = ap[2];
+	a3 = ap[3];
+	FLOAT temp0 = 0.0;
+	FLOAT temp1 = 0.0;
+	FLOAT temp2 = 0.0;
+	FLOAT temp3 = 0.0;
+
+	for ( i=0; i< n; i+=4 )
+	{
+		temp0 += a0[i]*x[i] + a0[i+1]*x[i+1] + a0[i+2]*x[i+2] + a0[i+3]*x[i+3];		
+		temp1 += a1[i]*x[i] + a1[i+1]*x[i+1] + a1[i+2]*x[i+2] + a1[i+3]*x[i+3];		
+		temp2 += a2[i]*x[i] + a2[i+1]*x[i+1] + a2[i+2]*x[i+2] + a2[i+3]*x[i+3];		
+		temp3 += a3[i]*x[i] + a3[i+1]*x[i+1] + a3[i+2]*x[i+2] + a3[i+3]*x[i+3];		
+	}
+	y[0] = temp0;
+	y[1] = temp1;
+	y[2] = temp2;
+	y[3] = temp3;
+}
+	
+#endif
+
+static void sgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0;
+	a0 = ap;
+	FLOAT temp = 0.0;
+
+	for ( i=0; i< n; i+=4 )
+	{
+		temp += a0[i]*x[i] + a0[i+1]*x[i+1] + a0[i+2]*x[i+2] + a0[i+3]*x[i+3];		
+	}
+	*y = temp;
+}
+	
+static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
+{
+        BLASLONG i;
+        for ( i=0; i<n; i++ )
+        {
+                *dest = *src;
+                dest++;
+                src += inc_src;
+        }
+}
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG j;
+	FLOAT *a_ptr;
+	FLOAT *x_ptr;
+	FLOAT *y_ptr;
+	FLOAT *ap[4];
+	BLASLONG n1;
+	BLASLONG m1;
+	BLASLONG m2;
+	BLASLONG n2;
+	FLOAT ybuffer[4],*xbuffer;
+
+	xbuffer = buffer;
+	
+	n1 = n / 4 ;
+	n2 = n % 4 ;
+	
+	m1 = m - ( m % 16 );
+	m2 = (m % NBMAX) - (m % 16) ;
+	
+
+	BLASLONG NB = NBMAX;
+
+	while ( NB == NBMAX )
+	{
+		
+		m1 -= NB;
+		if ( m1 < 0)
+		{
+			if ( m2 == 0 ) break;	
+			NB = m2;
+		}
+		
+		y_ptr = y;
+		a_ptr = a;
+		x_ptr = x;
+		copy_x(NB,x_ptr,xbuffer,inc_x);
+		for( i = 0; i < n1 ; i++)
+		{
+			ap[0] = a_ptr;
+			ap[1] = a_ptr + lda;
+			ap[2] = ap[1] + lda;
+			ap[3] = ap[2] + lda;
+			sgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
+			a_ptr += 4 * lda;
+			*y_ptr += ybuffer[0]*alpha;
+			y_ptr  += inc_y;
+			*y_ptr += ybuffer[1]*alpha;
+			y_ptr  += inc_y;
+			*y_ptr += ybuffer[2]*alpha;
+			y_ptr  += inc_y;
+			*y_ptr += ybuffer[3]*alpha;
+			y_ptr  += inc_y;
+		}
+
+		for( i = 0; i < n2 ; i++)
+		{
+			sgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
+			a_ptr += 1 * lda;
+			*y_ptr += ybuffer[0]*alpha;
+			y_ptr  += inc_y;
+
+		}
+		a += NB;
+		x += NB * inc_x;	
+	}
+
+	BLASLONG m3 = m % 16;
+	if ( m3 == 0 ) return(0);
+	x_ptr = x;
+	for ( i=0; i< m3; i++ )
+	{
+		xbuffer[i] = *x_ptr;
+		x_ptr += inc_x;
+	}
+	j=0;
+	a_ptr = a;
+	y_ptr = y;
+	while ( j < n)
+	{
+		FLOAT temp = 0.0;
+		for( i = 0; i < m3; i++ )
+		{
+			temp += a_ptr[i] * xbuffer[i];
+		}
+		a_ptr += lda;
+		y_ptr[0] += alpha * temp;
+		y_ptr += inc_y;
+		j++;
+	}
+	return(0);
+}
+
+
diff --git a/kernel/x86_64/sgemv_t_microk_bulldozer-2.c b/kernel/x86_64/sgemv_t_microk_bulldozer-2.c
new file mode 100644
index 000000000..54bdca63a
--- /dev/null
+++ b/kernel/x86_64/sgemv_t_microk_bulldozer-2.c
@@ -0,0 +1,109 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y) __attribute__ ((noinline));
+
+static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vxorps		%%xmm4, %%xmm4, %%xmm4	 \n\t"
+	"vxorps		%%xmm5, %%xmm5, %%xmm5	 \n\t"
+	"vxorps		%%xmm6, %%xmm6, %%xmm6	 \n\t"
+	"vxorps		%%xmm7, %%xmm7, %%xmm7	 \n\t"
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+        "vmovups        (%2,%0,4), %%xmm12       \n\t"  // 4 * x
+
+	"prefetcht0	 384(%4,%0,4)		       \n\t"
+	"vfmaddps %%xmm4,   (%4,%0,4), %%xmm12, %%xmm4 \n\t" 
+	"vfmaddps %%xmm5,   (%5,%0,4), %%xmm12, %%xmm5 \n\t" 
+        "vmovups      16(%2,%0,4), %%xmm13       \n\t"  // 4 * x
+	"vfmaddps %%xmm6,   (%6,%0,4), %%xmm12, %%xmm6 \n\t" 
+	"vfmaddps %%xmm7,   (%7,%0,4), %%xmm12, %%xmm7 \n\t" 
+	"prefetcht0	 384(%5,%0,4)		       \n\t"
+	"vfmaddps %%xmm4, 16(%4,%0,4), %%xmm13, %%xmm4 \n\t" 
+	"vfmaddps %%xmm5, 16(%5,%0,4), %%xmm13, %%xmm5 \n\t" 
+        "vmovups      32(%2,%0,4), %%xmm14       \n\t"  // 4 * x
+	"vfmaddps %%xmm6, 16(%6,%0,4), %%xmm13, %%xmm6 \n\t" 
+	"vfmaddps %%xmm7, 16(%7,%0,4), %%xmm13, %%xmm7 \n\t" 
+	"prefetcht0	 384(%6,%0,4)		       \n\t"
+	"vfmaddps %%xmm4, 32(%4,%0,4), %%xmm14, %%xmm4 \n\t" 
+	"vfmaddps %%xmm5, 32(%5,%0,4), %%xmm14, %%xmm5 \n\t" 
+        "vmovups      48(%2,%0,4), %%xmm15       \n\t"  // 4 * x
+	"vfmaddps %%xmm6, 32(%6,%0,4), %%xmm14, %%xmm6 \n\t" 
+	"vfmaddps %%xmm7, 32(%7,%0,4), %%xmm14, %%xmm7 \n\t" 
+	"prefetcht0	 384(%7,%0,4)		       \n\t"
+	"vfmaddps %%xmm4, 48(%4,%0,4), %%xmm15, %%xmm4 \n\t" 
+	"vfmaddps %%xmm5, 48(%5,%0,4), %%xmm15, %%xmm5 \n\t" 
+	"vfmaddps %%xmm6, 48(%6,%0,4), %%xmm15, %%xmm6 \n\t" 
+	"vfmaddps %%xmm7, 48(%7,%0,4), %%xmm15, %%xmm7 \n\t" 
+
+        "addq		$16, %0	  	 	      \n\t"
+	"subq	        $16, %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \n\t"
+
+	"vhaddps        %%xmm4, %%xmm4, %%xmm4	\n\t"
+	"vhaddps        %%xmm5, %%xmm5, %%xmm5	\n\t"
+	"vhaddps        %%xmm6, %%xmm6, %%xmm6	\n\t"
+	"vhaddps        %%xmm7, %%xmm7, %%xmm7	\n\t"
+
+	"vhaddps        %%xmm4, %%xmm4, %%xmm4	\n\t"
+	"vhaddps        %%xmm5, %%xmm5, %%xmm5	\n\t"
+	"vhaddps        %%xmm6, %%xmm6, %%xmm6	\n\t"
+	"vhaddps        %%xmm7, %%xmm7, %%xmm7	\n\t"
+
+	"vmovss		%%xmm4,    (%3)		\n\t"
+	"vmovss		%%xmm5,   4(%3)		\n\t"
+	"vmovss		%%xmm6,   8(%3)		\n\t"
+	"vmovss		%%xmm7,  12(%3)		\n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm4", "%xmm5", 
+	  "%xmm6", "%xmm7", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From 891b9608546b7b0dba350f8a1672b49334acfe7f Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Wed, 6 Aug 2014 13:42:41 +0200
Subject: [PATCH 02/27] added optimized sgemv_t kernel for haswell

---
 kernel/x86_64/sgemv_t.c                  |   2 +
 kernel/x86_64/sgemv_t_microk_haswell-2.c | 112 +++++++++++++++++++++++
 2 files changed, 114 insertions(+)
 create mode 100644 kernel/x86_64/sgemv_t_microk_haswell-2.c

diff --git a/kernel/x86_64/sgemv_t.c b/kernel/x86_64/sgemv_t.c
index 89254c256..566311a69 100644
--- a/kernel/x86_64/sgemv_t.c
+++ b/kernel/x86_64/sgemv_t.c
@@ -30,6 +30,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #if defined(BULLDOZER) || defined(PILEDRIVER)
 #include "sgemv_t_microk_bulldozer-2.c"
+#elif defined(HASWELL)
+#include "sgemv_t_microk_haswell-2.c"
 #endif
 
 /*
diff --git a/kernel/x86_64/sgemv_t_microk_haswell-2.c b/kernel/x86_64/sgemv_t_microk_haswell-2.c
new file mode 100644
index 000000000..cef703483
--- /dev/null
+++ b/kernel/x86_64/sgemv_t_microk_haswell-2.c
@@ -0,0 +1,112 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y) __attribute__ ((noinline));
+
+static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			         \n\t"
+	"vxorps		%%ymm4 , %%ymm4, %%ymm4  \n\t"
+	"vxorps		%%ymm5 , %%ymm5, %%ymm5  \n\t"
+	"vxorps		%%ymm6 , %%ymm6, %%ymm6  \n\t"
+	"vxorps		%%ymm7 , %%ymm7, %%ymm7  \n\t"
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"prefetcht0	 384(%2,%0,4)		 \n\t"
+	"vmovups	(%2,%0,4), %%ymm12       \n\t"	// 8 * x
+	"vmovups      32(%2,%0,4), %%ymm13       \n\t"	// 8 * x
+
+	"prefetcht0	 384(%4,%0,4)		       \n\t"
+	"vfmadd231ps   (%4,%0,4), %%ymm12, %%ymm4      \n\t" 
+	"vfmadd231ps   (%5,%0,4), %%ymm12, %%ymm5      \n\t" 
+	"prefetcht0	 384(%5,%0,4)		       \n\t"
+	"vfmadd231ps 32(%4,%0,4), %%ymm13, %%ymm4      \n\t" 
+	"vfmadd231ps 32(%5,%0,4), %%ymm13, %%ymm5      \n\t" 
+	"prefetcht0	 384(%6,%0,4)		       \n\t"
+	"vfmadd231ps   (%6,%0,4), %%ymm12, %%ymm6      \n\t" 
+	"vfmadd231ps   (%7,%0,4), %%ymm12, %%ymm7      \n\t" 
+	"prefetcht0	 384(%7,%0,4)		       \n\t"
+	"vfmadd231ps 32(%6,%0,4), %%ymm13, %%ymm6      \n\t" 
+	"vfmadd231ps 32(%7,%0,4), %%ymm13, %%ymm7      \n\t" 
+
+        "addq		$16, %0	  	 	      \n\t"
+	"subq	        $16, %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \n\t"
+
+	"vextractf128   $1 , %%ymm4, %%xmm12	      \n\t"
+	"vextractf128   $1 , %%ymm5, %%xmm13	      \n\t"
+	"vextractf128   $1 , %%ymm6, %%xmm14	      \n\t"
+	"vextractf128   $1 , %%ymm7, %%xmm15	      \n\t"
+
+	"vaddps		%%xmm4, %%xmm12, %%xmm4       \n\t"
+	"vaddps		%%xmm5, %%xmm13, %%xmm5       \n\t"
+	"vaddps		%%xmm6, %%xmm14, %%xmm6       \n\t"
+	"vaddps		%%xmm7, %%xmm15, %%xmm7       \n\t"
+
+        "vhaddps        %%xmm4, %%xmm4, %%xmm4  \n\t"
+        "vhaddps        %%xmm5, %%xmm5, %%xmm5  \n\t"
+        "vhaddps        %%xmm6, %%xmm6, %%xmm6  \n\t"
+        "vhaddps        %%xmm7, %%xmm7, %%xmm7  \n\t"
+
+        "vhaddps        %%xmm4, %%xmm4, %%xmm4  \n\t"
+        "vhaddps        %%xmm5, %%xmm5, %%xmm5  \n\t"
+        "vhaddps        %%xmm6, %%xmm6, %%xmm6  \n\t"
+        "vhaddps        %%xmm7, %%xmm7, %%xmm7  \n\t"
+
+        "vmovss         %%xmm4,    (%3)         \n\t"
+        "vmovss         %%xmm5,   4(%3)         \n\t"
+        "vmovss         %%xmm6,   8(%3)         \n\t"
+        "vmovss         %%xmm7,  12(%3)         \n\t"
+
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From 7aa43c8928bb1c69373ed426aa565c207a99470f Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Wed, 6 Aug 2014 14:06:30 +0200
Subject: [PATCH 03/27] enabled optimized sgemv kernels for windows

---
 kernel/x86_64/KERNEL.HASWELL | 6 ------
 1 file changed, 6 deletions(-)

diff --git a/kernel/x86_64/KERNEL.HASWELL b/kernel/x86_64/KERNEL.HASWELL
index 878a56b04..ff22954c6 100644
--- a/kernel/x86_64/KERNEL.HASWELL
+++ b/kernel/x86_64/KERNEL.HASWELL
@@ -1,11 +1,5 @@
-ifdef OS_WINDOWS
-SGEMVNKERNEL = sgemv_n.c
-SGEMVTKERNEL = ../arm/gemv_t.c
-else
 SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t.c
-endif
-
 
 SGEMMKERNEL    =  sgemm_kernel_16x4_haswell.S
 SGEMMINCOPY    =  ../generic/gemm_ncopy_16.c

From ca6c8d06cea5bb2fbe1b544f0da1dabefd256422 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Wed, 6 Aug 2014 14:24:36 +0200
Subject: [PATCH 04/27] enabled optimized sgemv kernels for windows

---
 kernel/x86_64/KERNEL.BULLDOZER  | 6 ------
 kernel/x86_64/KERNEL.PILEDRIVER | 6 ------
 2 files changed, 12 deletions(-)

diff --git a/kernel/x86_64/KERNEL.BULLDOZER b/kernel/x86_64/KERNEL.BULLDOZER
index accdddf0e..893f13064 100644
--- a/kernel/x86_64/KERNEL.BULLDOZER
+++ b/kernel/x86_64/KERNEL.BULLDOZER
@@ -1,11 +1,5 @@
-ifdef OS_WINDOWS
-SGEMVNKERNEL = sgemv_n.c
-SGEMVTKERNEL = ../arm/gemv_t.c
-else
 SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t.c
-endif
-
 
 ZGEMVNKERNEL = zgemv_n_dup.S
 ZGEMVTKERNEL = zgemv_t.S
diff --git a/kernel/x86_64/KERNEL.PILEDRIVER b/kernel/x86_64/KERNEL.PILEDRIVER
index 7b3c9a7b8..146a8768b 100644
--- a/kernel/x86_64/KERNEL.PILEDRIVER
+++ b/kernel/x86_64/KERNEL.PILEDRIVER
@@ -1,11 +1,5 @@
-ifdef OS_WINDOWS
-SGEMVNKERNEL = sgemv_n.c
-SGEMVTKERNEL = ../arm/gemv_t.c
-else
 SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t.c
-endif
-
 
 ZGEMVNKERNEL = zgemv_n_dup.S
 ZGEMVTKERNEL = zgemv_t.S

From d945a2b06d131995d4184311d397b17a6f1c7afb Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Wed, 6 Aug 2014 16:21:48 +0200
Subject: [PATCH 05/27] added optimized sgemv_t kernel for nehalem

---
 kernel/x86_64/sgemv_t.c                  |   2 +
 kernel/x86_64/sgemv_t_microk_nehalem-2.c | 159 +++++++++++++++++++++++
 2 files changed, 161 insertions(+)
 create mode 100644 kernel/x86_64/sgemv_t_microk_nehalem-2.c

diff --git a/kernel/x86_64/sgemv_t.c b/kernel/x86_64/sgemv_t.c
index 566311a69..47e749e58 100644
--- a/kernel/x86_64/sgemv_t.c
+++ b/kernel/x86_64/sgemv_t.c
@@ -32,6 +32,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "sgemv_t_microk_bulldozer-2.c"
 #elif defined(HASWELL)
 #include "sgemv_t_microk_haswell-2.c"
+#elif defined(NEHALEM)
+#include "sgemv_t_microk_nehalem-2.c"
 #endif
 
 /*
diff --git a/kernel/x86_64/sgemv_t_microk_nehalem-2.c b/kernel/x86_64/sgemv_t_microk_nehalem-2.c
new file mode 100644
index 000000000..e1f2b81bd
--- /dev/null
+++ b/kernel/x86_64/sgemv_t_microk_nehalem-2.c
@@ -0,0 +1,159 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y) __attribute__ ((noinline));
+
+static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"xorps		%%xmm0 , %%xmm0	         \n\t"
+	"xorps		%%xmm1 , %%xmm1	         \n\t"
+	"xorps		%%xmm2 , %%xmm2	         \n\t"
+	"xorps		%%xmm3 , %%xmm3	         \n\t"
+	"xorps		%%xmm4 , %%xmm4	         \n\t"
+	"xorps		%%xmm5 , %%xmm5	         \n\t"
+	"xorps		%%xmm6 , %%xmm6	         \n\t"
+	"xorps		%%xmm7 , %%xmm7	         \n\t"
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"prefetcht0	 384(%2,%0,4)		       \n\t"
+	"movups	       (%2,%0,4), %%xmm12        \n\t"	// 4 * x
+	"movups      16(%2,%0,4), %%xmm13        \n\t"	// 4 * x
+	"movups             (%4,%0,4), %%xmm8          \n\t" 
+	"movups      32(%2,%0,4), %%xmm14        \n\t"	// 4 * x
+	"movups      48(%2,%0,4), %%xmm15        \n\t"	// 4 * x
+
+	"prefetcht0	 384(%4,%0,4)		       \n\t"
+
+	"movups           16(%4,%0,4), %%xmm9          \n\t" 
+	"movups           32(%4,%0,4), %%xmm10         \n\t" 
+	"movups           48(%4,%0,4), %%xmm11         \n\t" 
+	"mulps		%%xmm12, %%xmm8		       \n\t"
+	"addps		%%xmm8 , %%xmm0		       \n\t"
+	"mulps		%%xmm13, %%xmm9		       \n\t"
+	"addps		%%xmm9 , %%xmm4		       \n\t"
+	"movups             (%5,%0,4), %%xmm8          \n\t" 
+	"mulps		%%xmm14, %%xmm10	       \n\t"
+	"addps		%%xmm10, %%xmm0		       \n\t"
+	"mulps		%%xmm15, %%xmm11	       \n\t"
+	"addps		%%xmm11, %%xmm4		       \n\t"
+
+	"prefetcht0	 384(%5,%0,4)		       \n\t"
+
+	"movups           16(%5,%0,4), %%xmm9          \n\t" 
+	"movups           32(%5,%0,4), %%xmm10         \n\t" 
+	"movups           48(%5,%0,4), %%xmm11         \n\t" 
+	"mulps		%%xmm12, %%xmm8		       \n\t"
+	"addps		%%xmm8 , %%xmm1		       \n\t"
+	"mulps		%%xmm13, %%xmm9		       \n\t"
+	"addps		%%xmm9 , %%xmm5		       \n\t"
+	"movups             (%6,%0,4), %%xmm8          \n\t" 
+	"mulps		%%xmm14, %%xmm10	       \n\t"
+	"addps		%%xmm10, %%xmm1		       \n\t"
+	"mulps		%%xmm15, %%xmm11	       \n\t"
+	"addps		%%xmm11, %%xmm5		       \n\t"
+
+	"prefetcht0	 384(%6,%0,4)		       \n\t"
+
+	"movups           16(%6,%0,4), %%xmm9          \n\t" 
+	"movups           32(%6,%0,4), %%xmm10         \n\t" 
+	"movups           48(%6,%0,4), %%xmm11         \n\t" 
+	"mulps		%%xmm12, %%xmm8		       \n\t"
+	"addps		%%xmm8 , %%xmm2		       \n\t"
+	"mulps		%%xmm13, %%xmm9		       \n\t"
+	"addps		%%xmm9 , %%xmm6		       \n\t"
+	"movups             (%7,%0,4), %%xmm8          \n\t" 
+	"mulps		%%xmm14, %%xmm10	       \n\t"
+	"addps		%%xmm10, %%xmm2		       \n\t"
+	"mulps		%%xmm15, %%xmm11	       \n\t"
+	"addps		%%xmm11, %%xmm6		       \n\t"
+
+	"prefetcht0	 384(%7,%0,4)		       \n\t"
+
+	"movups           16(%7,%0,4), %%xmm9          \n\t" 
+	"movups           32(%7,%0,4), %%xmm10         \n\t" 
+	"movups           48(%7,%0,4), %%xmm11         \n\t" 
+	"mulps		%%xmm12, %%xmm8		       \n\t"
+	"addps		%%xmm8 , %%xmm3		       \n\t"
+	"mulps		%%xmm13, %%xmm9		       \n\t"
+	"addps		%%xmm9 , %%xmm7		       \n\t"
+	"mulps		%%xmm14, %%xmm10	       \n\t"
+	"addps		%%xmm10, %%xmm3		       \n\t"
+	"mulps		%%xmm15, %%xmm11	       \n\t"
+	"addps		%%xmm11, %%xmm7		       \n\t"
+
+        "addq		$16, %0	  	 	      \n\t"
+	"subq	        $16, %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \n\t"
+
+	"addps	       %%xmm0, %%xmm4		      \n\t"
+	"addps	       %%xmm1, %%xmm5		      \n\t"
+	"addps	       %%xmm2, %%xmm6		      \n\t"
+	"addps	       %%xmm3, %%xmm7		      \n\t"
+
+        "haddps        %%xmm4, %%xmm4  \n\t"
+        "haddps        %%xmm5, %%xmm5  \n\t"
+        "haddps        %%xmm6, %%xmm6  \n\t"
+        "haddps        %%xmm7, %%xmm7  \n\t"
+
+        "haddps        %%xmm4, %%xmm4  \n\t"
+        "haddps        %%xmm5, %%xmm5  \n\t"
+        "haddps        %%xmm6, %%xmm6  \n\t"
+        "haddps        %%xmm7, %%xmm7  \n\t"
+
+        "movss         %%xmm4,    (%3)         \n\t"
+        "movss         %%xmm5,   4(%3)         \n\t"
+        "movss         %%xmm6,   8(%3)         \n\t"
+        "movss         %%xmm7,  12(%3)         \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From 2f8927376f50445f800b995ba6010e7ed571ecba Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Wed, 6 Aug 2014 16:58:21 +0200
Subject: [PATCH 06/27] enabled optimized nehalem sgemv_t kernel for windows

---
 kernel/x86_64/KERNEL.NEHALEM | 6 ------
 1 file changed, 6 deletions(-)

diff --git a/kernel/x86_64/KERNEL.NEHALEM b/kernel/x86_64/KERNEL.NEHALEM
index 8276150c6..ca9ff252d 100644
--- a/kernel/x86_64/KERNEL.NEHALEM
+++ b/kernel/x86_64/KERNEL.NEHALEM
@@ -1,11 +1,5 @@
-ifdef OS_WINDOWS
-SGEMVNKERNEL = sgemv_n.c
-SGEMVTKERNEL = ../arm/gemv_t.c
-else
 SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t.c
-endif
-
 
 SGEMMKERNEL    =  gemm_kernel_4x8_nehalem.S
 SGEMMINCOPY    =  gemm_ncopy_4.S

From c9bad1403ad4ff8b170bab16affcc1de1a6b66ce Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Thu, 7 Aug 2014 07:49:33 +0200
Subject: [PATCH 07/27] added optimized sgemv_t kernel for sandybridge

---
 kernel/x86_64/KERNEL.SANDYBRIDGE       |   6 --
 kernel/x86_64/sgemv_t.c                |  14 +--
 kernel/x86_64/sgemv_t_microk_sandy-2.c | 132 +++++++++++++++++++++++++
 3 files changed, 134 insertions(+), 18 deletions(-)
 create mode 100644 kernel/x86_64/sgemv_t_microk_sandy-2.c

diff --git a/kernel/x86_64/KERNEL.SANDYBRIDGE b/kernel/x86_64/KERNEL.SANDYBRIDGE
index 26706b61d..d4fbca7f2 100644
--- a/kernel/x86_64/KERNEL.SANDYBRIDGE
+++ b/kernel/x86_64/KERNEL.SANDYBRIDGE
@@ -1,11 +1,5 @@
-ifdef OS_WINDOWS
-SGEMVNKERNEL = sgemv_n.c
-SGEMVTKERNEL = ../arm/gemv_t.c
-else
 SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t.c
-endif
-
 
 SGEMMKERNEL    =  sgemm_kernel_16x4_sandy.S
 SGEMMINCOPY    =  ../generic/gemm_ncopy_16.c
diff --git a/kernel/x86_64/sgemv_t.c b/kernel/x86_64/sgemv_t.c
index 47e749e58..adfaa9925 100644
--- a/kernel/x86_64/sgemv_t.c
+++ b/kernel/x86_64/sgemv_t.c
@@ -32,22 +32,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "sgemv_t_microk_bulldozer-2.c"
 #elif defined(HASWELL)
 #include "sgemv_t_microk_haswell-2.c"
+#elif defined(SANDYBRIDGE)
+#include "sgemv_t_microk_sandy-2.c"
 #elif defined(NEHALEM)
 #include "sgemv_t_microk_nehalem-2.c"
 #endif
 
-/*
-#if defined(BULLDOZER) || defined(PILEDRIVER)
-#include "sgemv_n_microk_bulldozer-2.c"
-#elif defined(HASWELL)
-#include "sgemv_n_microk_haswell-2.c"
-#elif defined(SANDYBRIDGE)
-#include "sgemv_n_microk_sandy-2.c"
-#elif defined(NEHALEM)
-#include "sgemv_n_microk_nehalem-2.c"
-#endif
-*/
-
 #define NBMAX 4096
 
 #ifndef HAVE_KERNEL_16x4
diff --git a/kernel/x86_64/sgemv_t_microk_sandy-2.c b/kernel/x86_64/sgemv_t_microk_sandy-2.c
new file mode 100644
index 000000000..6a3748238
--- /dev/null
+++ b/kernel/x86_64/sgemv_t_microk_sandy-2.c
@@ -0,0 +1,132 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y) __attribute__ ((noinline));
+
+static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+        "vxorps         %%ymm0 , %%ymm0, %%ymm0  \n\t"
+        "vxorps         %%ymm1 , %%ymm1, %%ymm1  \n\t"
+        "vxorps         %%ymm2 , %%ymm2, %%ymm2  \n\t"
+        "vxorps         %%ymm3 , %%ymm3, %%ymm3  \n\t"
+        "vxorps         %%ymm4 , %%ymm4, %%ymm4  \n\t"
+        "vxorps         %%ymm5 , %%ymm5, %%ymm5  \n\t"
+        "vxorps         %%ymm6 , %%ymm6, %%ymm6  \n\t"
+        "vxorps         %%ymm7 , %%ymm7, %%ymm7  \n\t"
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"prefetcht0	 384(%2,%0,4)		       \n\t"
+        "vmovups        (%2,%0,4), %%ymm12       \n\t"  // 8 * x
+        "vmovups      32(%2,%0,4), %%ymm13       \n\t"  // 8 * x
+
+	"prefetcht0	 384(%4,%0,4)		       \n\t"
+	"vmulps   (%4,%0,4), %%ymm12, %%ymm8      \n\t" 
+	"vaddps	  %%ymm4, %%ymm8 , %%ymm4	  \n\t"
+	"vmulps 32(%4,%0,4), %%ymm13, %%ymm9      \n\t" 
+	"vaddps	  %%ymm0, %%ymm9 , %%ymm0	  \n\t"
+	"prefetcht0	 384(%5,%0,4)		       \n\t"
+	"vmulps   (%5,%0,4), %%ymm12, %%ymm10     \n\t" 
+	"vaddps	  %%ymm1, %%ymm10, %%ymm1	  \n\t"
+	"vmulps 32(%5,%0,4), %%ymm13, %%ymm11     \n\t" 
+	"vaddps	  %%ymm5, %%ymm11, %%ymm5	  \n\t"
+	"prefetcht0	 384(%6,%0,4)		       \n\t"
+	"vmulps   (%6,%0,4), %%ymm12, %%ymm8      \n\t" 
+	"vaddps	  %%ymm6, %%ymm8 , %%ymm6	  \n\t"
+	"vmulps 32(%6,%0,4), %%ymm13, %%ymm9      \n\t" 
+	"vaddps	  %%ymm2, %%ymm9 , %%ymm2	  \n\t"
+	"prefetcht0	 384(%7,%0,4)		       \n\t"
+	"vmulps   (%7,%0,4), %%ymm12, %%ymm10     \n\t" 
+	"vaddps	  %%ymm7, %%ymm10, %%ymm7	  \n\t"
+	"vmulps 32(%7,%0,4), %%ymm13, %%ymm11     \n\t" 
+	"vaddps	  %%ymm3, %%ymm11, %%ymm3	  \n\t"
+
+        "addq		$16, %0	  	 	      \n\t"
+	"subq	        $16, %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \n\t"
+
+        "vaddps         %%ymm4, %%ymm0, %%ymm4       \n\t"
+        "vaddps         %%ymm5, %%ymm1, %%ymm5       \n\t"
+        "vaddps         %%ymm6, %%ymm2, %%ymm6       \n\t"
+        "vaddps         %%ymm7, %%ymm3, %%ymm7       \n\t"
+
+        "vextractf128   $1 , %%ymm4, %%xmm12          \n\t"
+        "vextractf128   $1 , %%ymm5, %%xmm13          \n\t"
+        "vextractf128   $1 , %%ymm6, %%xmm14          \n\t"
+        "vextractf128   $1 , %%ymm7, %%xmm15          \n\t"
+
+        "vaddps         %%xmm4, %%xmm12, %%xmm4       \n\t"
+        "vaddps         %%xmm5, %%xmm13, %%xmm5       \n\t"
+        "vaddps         %%xmm6, %%xmm14, %%xmm6       \n\t"
+        "vaddps         %%xmm7, %%xmm15, %%xmm7       \n\t"
+
+        "vhaddps        %%xmm4, %%xmm4, %%xmm4  \n\t"
+        "vhaddps        %%xmm5, %%xmm5, %%xmm5  \n\t"
+        "vhaddps        %%xmm6, %%xmm6, %%xmm6  \n\t"
+        "vhaddps        %%xmm7, %%xmm7, %%xmm7  \n\t"
+
+        "vhaddps        %%xmm4, %%xmm4, %%xmm4  \n\t"
+        "vhaddps        %%xmm5, %%xmm5, %%xmm5  \n\t"
+        "vhaddps        %%xmm6, %%xmm6, %%xmm6  \n\t"
+        "vhaddps        %%xmm7, %%xmm7, %%xmm7  \n\t"
+
+        "vmovss         %%xmm4,    (%3)         \n\t"
+        "vmovss         %%xmm5,   4(%3)         \n\t"
+        "vmovss         %%xmm6,   8(%3)         \n\t"
+        "vmovss         %%xmm7,  12(%3)         \n\t"
+
+
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From 60f17628cc65387f474160462511f91f93d50c3e Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Thu, 7 Aug 2014 09:18:02 +0200
Subject: [PATCH 08/27] added optimized dgemv_n kernel for haswell

---
 kernel/x86_64/KERNEL.HASWELL             |   1 +
 kernel/x86_64/dgemv_n.c                  | 203 +++++++++++++++++++++++
 kernel/x86_64/dgemv_n_microk_haswell-2.c |  89 ++++++++++
 3 files changed, 293 insertions(+)
 create mode 100644 kernel/x86_64/dgemv_n.c
 create mode 100644 kernel/x86_64/dgemv_n_microk_haswell-2.c

diff --git a/kernel/x86_64/KERNEL.HASWELL b/kernel/x86_64/KERNEL.HASWELL
index ff22954c6..bcff2e224 100644
--- a/kernel/x86_64/KERNEL.HASWELL
+++ b/kernel/x86_64/KERNEL.HASWELL
@@ -1,5 +1,6 @@
 SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t.c
+DGEMVNKERNEL = dgemv_n.c
 
 SGEMMKERNEL    =  sgemm_kernel_16x4_haswell.S
 SGEMMINCOPY    =  ../generic/gemm_ncopy_16.c
diff --git a/kernel/x86_64/dgemv_n.c b/kernel/x86_64/dgemv_n.c
new file mode 100644
index 000000000..5192ba193
--- /dev/null
+++ b/kernel/x86_64/dgemv_n.c
@@ -0,0 +1,203 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include "common.h"
+
+
+#if defined(HASWELL)
+#include "dgemv_n_microk_haswell-2.c"
+#endif
+
+
+#define NBMAX 2048
+
+#ifndef HAVE_KERNEL_16x4
+
+static void dgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0,*a1,*a2,*a3;
+	a0 = ap[0];
+	a1 = ap[1];
+	a2 = ap[2];
+	a3 = ap[3];
+
+	for ( i=0; i< n; i+=4 )
+	{
+		y[i] += a0[i]*x[0] + a1[i]*x[1] + a2[i]*x[2] + a3[i]*x[3];		
+		y[i+1] += a0[i+1]*x[0] + a1[i+1]*x[1] + a2[i+1]*x[2] + a3[i+1]*x[3];		
+		y[i+2] += a0[i+2]*x[0] + a1[i+2]*x[1] + a2[i+2]*x[2] + a3[i+2]*x[3];		
+		y[i+3] += a0[i+3]*x[0] + a1[i+3]*x[1] + a2[i+3]*x[2] + a3[i+3]*x[3];		
+	}
+}
+	
+#endif
+
+static void dgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0;
+	a0 = ap;
+
+	for ( i=0; i< n; i+=4 )
+	{
+		y[i] += a0[i]*x[0];		
+		y[i+1] += a0[i+1]*x[0];		
+		y[i+2] += a0[i+2]*x[0];		
+		y[i+3] += a0[i+3]*x[0];		
+	}
+}
+	
+
+static void zero_y(BLASLONG n, FLOAT *dest)
+{
+	BLASLONG i;
+	for ( i=0; i<n; i++ )
+	{
+		*dest = 0.0;
+		dest++;
+	}
+}
+
+
+
+static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest)
+{
+	BLASLONG i;
+	if ( inc_dest == 1 )
+	{
+		for ( i=0; i<n; i+=4 )
+		{
+			dest[i] += src[i];
+			dest[i+1] += src[i+1];
+			dest[i+2] += src[i+2];
+			dest[i+3] += src[i+3];
+		}
+
+	}
+	else
+	{
+		for ( i=0; i<n; i++ )
+		{
+			*dest += *src;
+			src++;
+			dest += inc_dest;
+		}
+	}
+}
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG j;
+	FLOAT *a_ptr;
+	FLOAT *x_ptr;
+	FLOAT *y_ptr;
+	FLOAT *ap[4];
+	BLASLONG n1;
+	BLASLONG m1;
+	BLASLONG m2;
+	BLASLONG n2;
+	FLOAT xbuffer[4],*ybuffer;
+
+	ybuffer = buffer;
+	
+	n1 = n / 4 ;
+	n2 = n % 4 ;
+	
+	m1 = m - ( m % 16 );
+	m2 = (m % NBMAX) - (m % 16) ;
+	
+	y_ptr = y;
+
+	BLASLONG NB = NBMAX;
+
+	while ( NB == NBMAX )
+	{
+		
+		m1 -= NB;
+		if ( m1 < 0)
+		{
+			if ( m2 == 0 ) break;	
+			NB = m2;
+		}
+		
+		a_ptr = a;
+		x_ptr = x;
+		zero_y(NB,ybuffer);
+		for( i = 0; i < n1 ; i++)
+		{
+			xbuffer[0] = alpha * x_ptr[0];
+			x_ptr += inc_x;	
+			xbuffer[1] = alpha * x_ptr[0];
+			x_ptr += inc_x;	
+			xbuffer[2] = alpha * x_ptr[0];
+			x_ptr += inc_x;	
+			xbuffer[3] = alpha * x_ptr[0];
+			x_ptr += inc_x;	
+			ap[0] = a_ptr;
+			ap[1] = a_ptr + lda;
+			ap[2] = ap[1] + lda;
+			ap[3] = ap[2] + lda;
+			dgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
+			a_ptr += 4 * lda;
+		}
+
+		for( i = 0; i < n2 ; i++)
+		{
+			xbuffer[0] = alpha * x_ptr[0];
+			x_ptr += inc_x;	
+			dgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
+			a_ptr += 1 * lda;
+
+		}
+		add_y(NB,ybuffer,y_ptr,inc_y);
+		a     += NB;
+		y_ptr += NB * inc_y;
+	}
+	j=0;
+	while ( j < (m % 16))
+	{
+		a_ptr = a;
+		x_ptr = x;
+		FLOAT temp = 0.0;
+		for( i = 0; i < n; i++ )
+		{
+			temp += a_ptr[0] * x_ptr[0];
+			a_ptr += lda;
+			x_ptr += inc_x;
+		}
+		y_ptr[0] += alpha * temp;
+		y_ptr += inc_y;
+		a++;
+		j++;
+	}
+	return(0);
+}
+
+
diff --git a/kernel/x86_64/dgemv_n_microk_haswell-2.c b/kernel/x86_64/dgemv_n_microk_haswell-2.c
new file mode 100644
index 000000000..88c7d4163
--- /dev/null
+++ b/kernel/x86_64/dgemv_n_microk_haswell-2.c
@@ -0,0 +1,89 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void dgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y) __attribute__ ((noinline));
+
+static void dgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+	"vbroadcastsd    (%2), %%ymm12	 \n\t"	// x0 
+	"vbroadcastsd   8(%2), %%ymm13	 \n\t"	// x1 
+	"vbroadcastsd  16(%2), %%ymm14	 \n\t"	// x2 
+	"vbroadcastsd  24(%2), %%ymm15	 \n\t"	// x3 
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"prefetcht0	 192(%4,%0,8)		 \n\t"
+	"vmovups	(%3,%0,8), %%ymm4	 \n\t"	// 4 * y
+	"vmovups      32(%3,%0,8), %%ymm5	 \n\t"	// 4 * y
+
+	"prefetcht0	 192(%4,%0,8)		       \n\t"
+	"vfmadd231pd   (%4,%0,8), %%ymm12, %%ymm4      \n\t" 
+	"vfmadd231pd 32(%4,%0,8), %%ymm12, %%ymm5      \n\t" 
+	"prefetcht0	 192(%5,%0,8)		       \n\t"
+	"vfmadd231pd   (%5,%0,8), %%ymm13, %%ymm4      \n\t" 
+	"vfmadd231pd 32(%5,%0,8), %%ymm13, %%ymm5      \n\t" 
+	"prefetcht0	 192(%6,%0,8)		       \n\t"
+	"vfmadd231pd   (%6,%0,8), %%ymm14, %%ymm4      \n\t" 
+	"vfmadd231pd 32(%6,%0,8), %%ymm14, %%ymm5      \n\t" 
+	"prefetcht0	 192(%7,%0,8)		       \n\t"
+	"vfmadd231pd   (%7,%0,8), %%ymm15, %%ymm4      \n\t" 
+	"vfmadd231pd 32(%7,%0,8), %%ymm15, %%ymm5      \n\t" 
+
+	"vmovups  %%ymm4,   (%3,%0,8)		      \n\t"	// 4 * y
+	"vmovups  %%ymm5, 32(%3,%0,8)		      \n\t"	// 4 * y
+
+        "addq		$8 , %0	  	 	      \n\t"
+	"subq	        $8 , %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \n\t"
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm4", "%xmm5", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From 006ef3ea012b4aa09e775724bcf7094500c597d5 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Thu, 7 Aug 2014 10:08:54 +0200
Subject: [PATCH 09/27] added optimized dgemv_t kernel for haswell

---
 kernel/x86_64/KERNEL.HASWELL             |   1 +
 kernel/x86_64/dgemv_t.c                  | 188 +++++++++++++++++++++++
 kernel/x86_64/dgemv_t_microk_haswell-2.c | 107 +++++++++++++
 3 files changed, 296 insertions(+)
 create mode 100644 kernel/x86_64/dgemv_t.c
 create mode 100644 kernel/x86_64/dgemv_t_microk_haswell-2.c

diff --git a/kernel/x86_64/KERNEL.HASWELL b/kernel/x86_64/KERNEL.HASWELL
index bcff2e224..cd280e4b6 100644
--- a/kernel/x86_64/KERNEL.HASWELL
+++ b/kernel/x86_64/KERNEL.HASWELL
@@ -1,6 +1,7 @@
 SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t.c
 DGEMVNKERNEL = dgemv_n.c
+DGEMVTKERNEL = dgemv_t.c
 
 SGEMMKERNEL    =  sgemm_kernel_16x4_haswell.S
 SGEMMINCOPY    =  ../generic/gemm_ncopy_16.c
diff --git a/kernel/x86_64/dgemv_t.c b/kernel/x86_64/dgemv_t.c
new file mode 100644
index 000000000..76aacd349
--- /dev/null
+++ b/kernel/x86_64/dgemv_t.c
@@ -0,0 +1,188 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include "common.h"
+
+#if defined(HASWELL)
+#include "dgemv_t_microk_haswell-2.c"
+#endif
+
+#define NBMAX 2048
+
+#ifndef HAVE_KERNEL_16x4
+
+static void dgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0,*a1,*a2,*a3;
+	a0 = ap[0];
+	a1 = ap[1];
+	a2 = ap[2];
+	a3 = ap[3];
+	FLOAT temp0 = 0.0;
+	FLOAT temp1 = 0.0;
+	FLOAT temp2 = 0.0;
+	FLOAT temp3 = 0.0;
+
+	for ( i=0; i< n; i+=4 )
+	{
+		temp0 += a0[i]*x[i] + a0[i+1]*x[i+1] + a0[i+2]*x[i+2] + a0[i+3]*x[i+3];		
+		temp1 += a1[i]*x[i] + a1[i+1]*x[i+1] + a1[i+2]*x[i+2] + a1[i+3]*x[i+3];		
+		temp2 += a2[i]*x[i] + a2[i+1]*x[i+1] + a2[i+2]*x[i+2] + a2[i+3]*x[i+3];		
+		temp3 += a3[i]*x[i] + a3[i+1]*x[i+1] + a3[i+2]*x[i+2] + a3[i+3]*x[i+3];		
+	}
+	y[0] = temp0;
+	y[1] = temp1;
+	y[2] = temp2;
+	y[3] = temp3;
+}
+	
+#endif
+
+static void dgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0;
+	a0 = ap;
+	FLOAT temp = 0.0;
+
+	for ( i=0; i< n; i+=4 )
+	{
+		temp += a0[i]*x[i] + a0[i+1]*x[i+1] + a0[i+2]*x[i+2] + a0[i+3]*x[i+3];		
+	}
+	*y = temp;
+}
+	
+static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
+{
+        BLASLONG i;
+        for ( i=0; i<n; i++ )
+        {
+                *dest = *src;
+                dest++;
+                src += inc_src;
+        }
+}
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG j;
+	FLOAT *a_ptr;
+	FLOAT *x_ptr;
+	FLOAT *y_ptr;
+	FLOAT *ap[4];
+	BLASLONG n1;
+	BLASLONG m1;
+	BLASLONG m2;
+	BLASLONG n2;
+	FLOAT ybuffer[4],*xbuffer;
+
+	xbuffer = buffer;
+	
+	n1 = n / 4 ;
+	n2 = n % 4 ;
+	
+	m1 = m - ( m % 16 );
+	m2 = (m % NBMAX) - (m % 16) ;
+	
+
+	BLASLONG NB = NBMAX;
+
+	while ( NB == NBMAX )
+	{
+		
+		m1 -= NB;
+		if ( m1 < 0)
+		{
+			if ( m2 == 0 ) break;	
+			NB = m2;
+		}
+		
+		y_ptr = y;
+		a_ptr = a;
+		x_ptr = x;
+		copy_x(NB,x_ptr,xbuffer,inc_x);
+		for( i = 0; i < n1 ; i++)
+		{
+			ap[0] = a_ptr;
+			ap[1] = a_ptr + lda;
+			ap[2] = ap[1] + lda;
+			ap[3] = ap[2] + lda;
+			dgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
+			a_ptr += 4 * lda;
+			*y_ptr += ybuffer[0]*alpha;
+			y_ptr  += inc_y;
+			*y_ptr += ybuffer[1]*alpha;
+			y_ptr  += inc_y;
+			*y_ptr += ybuffer[2]*alpha;
+			y_ptr  += inc_y;
+			*y_ptr += ybuffer[3]*alpha;
+			y_ptr  += inc_y;
+		}
+
+		for( i = 0; i < n2 ; i++)
+		{
+			dgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
+			a_ptr += 1 * lda;
+			*y_ptr += ybuffer[0]*alpha;
+			y_ptr  += inc_y;
+
+		}
+		a += NB;
+		x += NB * inc_x;	
+	}
+
+	BLASLONG m3 = m % 16;
+	if ( m3 == 0 ) return(0);
+	x_ptr = x;
+	for ( i=0; i< m3; i++ )
+	{
+		xbuffer[i] = *x_ptr;
+		x_ptr += inc_x;
+	}
+	j=0;
+	a_ptr = a;
+	y_ptr = y;
+	while ( j < n)
+	{
+		FLOAT temp = 0.0;
+		for( i = 0; i < m3; i++ )
+		{
+			temp += a_ptr[i] * xbuffer[i];
+		}
+		a_ptr += lda;
+		y_ptr[0] += alpha * temp;
+		y_ptr += inc_y;
+		j++;
+	}
+	return(0);
+}
+
+
diff --git a/kernel/x86_64/dgemv_t_microk_haswell-2.c b/kernel/x86_64/dgemv_t_microk_haswell-2.c
new file mode 100644
index 000000000..94d4c319e
--- /dev/null
+++ b/kernel/x86_64/dgemv_t_microk_haswell-2.c
@@ -0,0 +1,107 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void dgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y) __attribute__ ((noinline));
+
+static void dgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			         \n\t"
+	"vxorpd		%%ymm4 , %%ymm4, %%ymm4  \n\t"
+	"vxorpd		%%ymm5 , %%ymm5, %%ymm5  \n\t"
+	"vxorpd		%%ymm6 , %%ymm6, %%ymm6  \n\t"
+	"vxorpd		%%ymm7 , %%ymm7, %%ymm7  \n\t"
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \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"
+	"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"
+	"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" 
+
+        "addq		$8 , %0	  	 	      \n\t"
+	"subq	        $8 , %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \n\t"
+
+	"vextractf128   $1 , %%ymm4, %%xmm12	      \n\t"
+	"vextractf128   $1 , %%ymm5, %%xmm13	      \n\t"
+	"vextractf128   $1 , %%ymm6, %%xmm14	      \n\t"
+	"vextractf128   $1 , %%ymm7, %%xmm15	      \n\t"
+
+	"vaddpd		%%xmm4, %%xmm12, %%xmm4       \n\t"
+	"vaddpd		%%xmm5, %%xmm13, %%xmm5       \n\t"
+	"vaddpd		%%xmm6, %%xmm14, %%xmm6       \n\t"
+	"vaddpd		%%xmm7, %%xmm15, %%xmm7       \n\t"
+
+        "vhaddpd        %%xmm4, %%xmm4, %%xmm4  \n\t"
+        "vhaddpd        %%xmm5, %%xmm5, %%xmm5  \n\t"
+        "vhaddpd        %%xmm6, %%xmm6, %%xmm6  \n\t"
+        "vhaddpd        %%xmm7, %%xmm7, %%xmm7  \n\t"
+
+        "vmovsd         %%xmm4,    (%3)         \n\t"
+        "vmovsd         %%xmm5,   8(%3)         \n\t"
+        "vmovsd         %%xmm6,  16(%3)         \n\t"
+        "vmovsd         %%xmm7,  24(%3)         \n\t"
+
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From aa54fe064c1056e4f87266d1d64f10c51d9558a9 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Thu, 7 Aug 2014 22:30:20 +0200
Subject: [PATCH 10/27] added zgemv_n c-function

---
 kernel/x86_64/zgemv_n.c | 302 ++++++++++++++++++++++++++++++++++++++++
 1 file changed, 302 insertions(+)
 create mode 100644 kernel/x86_64/zgemv_n.c

diff --git a/kernel/x86_64/zgemv_n.c b/kernel/x86_64/zgemv_n.c
new file mode 100644
index 000000000..be5b08dcd
--- /dev/null
+++ b/kernel/x86_64/zgemv_n.c
@@ -0,0 +1,302 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include "common.h"
+
+
+#define NBMAX 1024
+
+#ifndef HAVE_KERNEL_16x4
+
+static void zgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0,*a1,*a2,*a3;
+	a0 = ap[0];
+	a1 = ap[1];
+	a2 = ap[2];
+	a3 = ap[3];
+
+	for ( i=0; i< 2*n; i+=2 )
+	{
+#if !defined(CONJ) 
+#if !defined(XCONJ)
+		y[i]   += a0[i]*x[0] - a0[i+1] * x[1];
+		y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
+		y[i]   += a1[i]*x[2] - a1[i+1] * x[3];
+		y[i+1] += a1[i]*x[3] + a1[i+1] * x[2];
+		y[i]   += a2[i]*x[4] - a2[i+1] * x[5];
+		y[i+1] += a2[i]*x[5] + a2[i+1] * x[4];
+		y[i]   += a3[i]*x[6] - a3[i+1] * x[7];
+		y[i+1] += a3[i]*x[7] + a3[i+1] * x[6];
+#else 
+		y[i]   += a0[i]*x[0] + a0[i+1] * x[1];
+		y[i+1] += a0[i]*x[1] - a0[i+1] * x[0];
+		y[i]   += a1[i]*x[2] + a1[i+1] * x[3];
+		y[i+1] += a1[i]*x[3] - a1[i+1] * x[2];
+		y[i]   += a2[i]*x[4] + a2[i+1] * x[5];
+		y[i+1] += a2[i]*x[5] - a2[i+1] * x[4];
+		y[i]   += a3[i]*x[6] + a3[i+1] * x[7];
+		y[i+1] += a3[i]*x[7] - a3[i+1] * x[6];
+#endif
+#else
+#if !defined(XCONJ)
+		y[i]   += a0[i]*x[0] + a0[i+1] * x[1];
+		y[i+1] -= a0[i]*x[1] - a0[i+1] * x[0];
+		y[i]   += a1[i]*x[2] + a1[i+1] * x[3];
+		y[i+1] -= a1[i]*x[3] - a1[i+1] * x[2];
+		y[i]   += a2[i]*x[4] + a2[i+1] * x[5];
+		y[i+1] -= a2[i]*x[5] - a2[i+1] * x[4];
+		y[i]   += a3[i]*x[6] + a3[i+1] * x[7];
+		y[i+1] -= a3[i]*x[7] - a3[i+1] * x[6];
+
+#else
+		y[i]   += a0[i]*x[0] - a0[i+1] * x[1];
+		y[i+1] -= a0[i]*x[1] + a0[i+1] * x[0];
+		y[i]   += a1[i]*x[2] - a1[i+1] * x[3];
+		y[i+1] -= a1[i]*x[3] + a1[i+1] * x[2];
+		y[i]   += a2[i]*x[4] - a2[i+1] * x[5];
+		y[i+1] -= a2[i]*x[5] + a2[i+1] * x[4];
+		y[i]   += a3[i]*x[6] - a3[i+1] * x[7];
+		y[i+1] -= a3[i]*x[7] + a3[i+1] * x[6];
+
+#endif
+#endif
+	}
+}
+	
+#endif
+
+static void zgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0;
+	a0 = ap;
+
+	for ( i=0; i< 2*n; i+=2 )
+	{
+#if !defined(CONJ) 
+#if !defined(XCONJ)
+		y[i]   += a0[i]*x[0] - a0[i+1] * x[1];
+		y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
+#else 
+		y[i]   += a0[i]*x[0] + a0[i+1] * x[1];
+		y[i+1] += a0[i]*x[1] - a0[i+1] * x[0];
+#endif
+#else
+#if !defined(XCONJ)
+		y[i]   += a0[i]*x[0] + a0[i+1] * x[1];
+		y[i+1] -= a0[i]*x[1] - a0[i+1] * x[0];
+
+#else
+		y[i]   += a0[i]*x[0] - a0[i+1] * x[1];
+		y[i+1] -= a0[i]*x[1] + a0[i+1] * x[0];
+#endif
+#endif
+
+	}
+}
+	
+
+static void zero_y(BLASLONG n, FLOAT *dest)
+{
+	BLASLONG i;
+	for ( i=0; i<2*n; i++ )
+	{
+		*dest = 0.0;
+		dest++;
+	}
+}
+
+
+
+static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest)
+{
+	BLASLONG i;
+	for ( i=0; i<n; i++ )
+	{
+		*dest += *src;
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+		*(dest+1) += *(src+1);
+#else
+		*(dest+1) -= *(src+1);
+#endif
+		src+=2;
+		dest += inc_dest;
+	}
+}
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r,FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG j;
+	FLOAT *a_ptr;
+	FLOAT *x_ptr;
+	FLOAT *y_ptr;
+	FLOAT *ap[4];
+	BLASLONG n1;
+	BLASLONG m1;
+	BLASLONG m2;
+	BLASLONG n2;
+	FLOAT xbuffer[8],*ybuffer;
+
+	ybuffer = buffer;
+	
+	inc_x *= 2;
+	inc_y *= 2;
+	lda   *= 2;
+
+	n1 = n / 4 ;
+	n2 = n % 4 ;
+	
+	m1 = m - ( m % 16 );
+	m2 = (m % NBMAX) - (m % 16) ;
+	
+	y_ptr = y;
+
+	BLASLONG NB = NBMAX;
+
+	while ( NB == NBMAX )
+	{
+		
+		m1 -= NB;
+		if ( m1 < 0)
+		{
+			if ( m2 == 0 ) break;	
+			NB = m2;
+		}
+		
+		a_ptr = a;
+		x_ptr = x;
+		zero_y(NB,ybuffer);
+		for( i = 0; i < n1 ; i++)
+		{
+
+#if !defined(XCONJ)
+			xbuffer[0] = alpha_r * x_ptr[0] - alpha_i * x_ptr[1];
+			xbuffer[1] = alpha_r * x_ptr[1] + alpha_i * x_ptr[0];
+			x_ptr += inc_x;	
+			xbuffer[2] = alpha_r * x_ptr[0] - alpha_i * x_ptr[1];
+			xbuffer[3] = alpha_r * x_ptr[1] + alpha_i * x_ptr[0];
+			x_ptr += inc_x;	
+			xbuffer[4] = alpha_r * x_ptr[0] - alpha_i * x_ptr[1];
+			xbuffer[5] = alpha_r * x_ptr[1] + alpha_i * x_ptr[0];
+			x_ptr += inc_x;	
+			xbuffer[6] = alpha_r * x_ptr[0] - alpha_i * x_ptr[1];
+			xbuffer[7] = alpha_r * x_ptr[1] + alpha_i * x_ptr[0];
+			x_ptr += inc_x;	
+#else
+			xbuffer[0] = alpha_r * x_ptr[0] + alpha_i * x_ptr[1];
+			xbuffer[1] = alpha_r * x_ptr[1] - alpha_i * x_ptr[0];
+			x_ptr += inc_x;	
+			xbuffer[2] = alpha_r * x_ptr[0] + alpha_i * x_ptr[1];
+			xbuffer[3] = alpha_r * x_ptr[1] - alpha_i * x_ptr[0];
+			x_ptr += inc_x;	
+			xbuffer[4] = alpha_r * x_ptr[0] + alpha_i * x_ptr[1];
+			xbuffer[5] = alpha_r * x_ptr[1] - alpha_i * x_ptr[0];
+			x_ptr += inc_x;	
+			xbuffer[6] = alpha_r * x_ptr[0] + alpha_i * x_ptr[1];
+			xbuffer[7] = alpha_r * x_ptr[1] - alpha_i * x_ptr[0];
+			x_ptr += inc_x;	
+#endif
+
+			ap[0] = a_ptr;
+			ap[1] = a_ptr + lda;
+			ap[2] = ap[1] + lda;
+			ap[3] = ap[2] + lda;
+			zgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
+			a_ptr += 4 * lda;
+		}
+
+		for( i = 0; i < n2 ; i++)
+		{
+#if !defined(XCONJ)
+			xbuffer[0] = alpha_r * x_ptr[0] - alpha_i * x_ptr[1];
+			xbuffer[1] = alpha_r * x_ptr[1] + alpha_i * x_ptr[0];
+			x_ptr += inc_x;	
+#else
+			xbuffer[0] = alpha_r * x_ptr[0] + alpha_i * x_ptr[1];
+			xbuffer[1] = alpha_r * x_ptr[1] - alpha_i * x_ptr[0];
+			x_ptr += inc_x;	
+#endif
+
+			zgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
+			a_ptr += 1 * lda;
+
+		}
+		add_y(NB,ybuffer,y_ptr,inc_y);
+		a     += 2 * NB;
+		y_ptr += NB * inc_y;
+	}
+	j=0;
+	while ( j < (m % 16))
+	{
+		a_ptr = a;
+		x_ptr = x;
+		FLOAT temp_r = 0.0;
+		FLOAT temp_i = 0.0;
+		for( i = 0; i < n; i++ )
+		{
+#if  !defined(CONJ)
+#if  !defined(XCONJ)
+			temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
+			temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
+#else
+			temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
+			temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
+#endif
+#else
+#if  !defined(XCONJ)
+			temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
+			temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
+#else
+			temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
+			temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
+#endif
+#endif
+
+
+			a_ptr += lda;
+			x_ptr += inc_x;
+		}
+
+#if !defined(XCONJ) 
+		y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
+		y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
+#else
+		y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
+		y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
+#endif
+		y_ptr += inc_y;
+		a+=2;
+		j++;
+	}
+	return(0);
+}
+
+

From 462b4885ff471da4e525a2f768a734151aab774b Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Sun, 10 Aug 2014 08:39:17 +0200
Subject: [PATCH 11/27] added optimized zgemv_n kernel for haswell

---
 kernel/x86_64/KERNEL.HASWELL             |   1 +
 kernel/x86_64/dgemv_n_microk_haswell-2.c |   2 +-
 kernel/x86_64/zgemv_n.c                  |   4 +
 kernel/x86_64/zgemv_n_microk_haswell-2.c | 149 +++++++++++++++++++++++
 4 files changed, 155 insertions(+), 1 deletion(-)
 create mode 100644 kernel/x86_64/zgemv_n_microk_haswell-2.c

diff --git a/kernel/x86_64/KERNEL.HASWELL b/kernel/x86_64/KERNEL.HASWELL
index cd280e4b6..d126eb6f4 100644
--- a/kernel/x86_64/KERNEL.HASWELL
+++ b/kernel/x86_64/KERNEL.HASWELL
@@ -2,6 +2,7 @@ SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t.c
 DGEMVNKERNEL = dgemv_n.c
 DGEMVTKERNEL = dgemv_t.c
+ZGEMVNKERNEL = zgemv_n.c
 
 SGEMMKERNEL    =  sgemm_kernel_16x4_haswell.S
 SGEMMINCOPY    =  ../generic/gemm_ncopy_16.c
diff --git a/kernel/x86_64/dgemv_n_microk_haswell-2.c b/kernel/x86_64/dgemv_n_microk_haswell-2.c
index 88c7d4163..b9f462cb2 100644
--- a/kernel/x86_64/dgemv_n_microk_haswell-2.c
+++ b/kernel/x86_64/dgemv_n_microk_haswell-2.c
@@ -43,7 +43,7 @@ static void dgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
 
 	".align 16				 \n\t"
 	".L01LOOP%=:				 \n\t"
-	"prefetcht0	 192(%4,%0,8)		 \n\t"
+	"prefetcht0	 192(%3,%0,8)		 \n\t"
 	"vmovups	(%3,%0,8), %%ymm4	 \n\t"	// 4 * y
 	"vmovups      32(%3,%0,8), %%ymm5	 \n\t"	// 4 * y
 
diff --git a/kernel/x86_64/zgemv_n.c b/kernel/x86_64/zgemv_n.c
index be5b08dcd..141cb35df 100644
--- a/kernel/x86_64/zgemv_n.c
+++ b/kernel/x86_64/zgemv_n.c
@@ -28,6 +28,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include "common.h"
 
+#if defined(HASWELL)
+#include "zgemv_n_microk_haswell-2.c"
+#endif
+
 
 #define NBMAX 1024
 
diff --git a/kernel/x86_64/zgemv_n_microk_haswell-2.c b/kernel/x86_64/zgemv_n_microk_haswell-2.c
new file mode 100644
index 000000000..8583f96b3
--- /dev/null
+++ b/kernel/x86_64/zgemv_n_microk_haswell-2.c
@@ -0,0 +1,149 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void zgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y) __attribute__ ((noinline));
+
+static void zgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+
+	"vbroadcastsd	  (%2), %%ymm0                  \n\t"  // real part x0
+	"vbroadcastsd	 8(%2), %%ymm1                  \n\t"  // imag part x0
+	"vbroadcastsd	16(%2), %%ymm2                  \n\t"  // real part x1
+	"vbroadcastsd	24(%2), %%ymm3                  \n\t"  // imag part x1
+	"vbroadcastsd	32(%2), %%ymm4                  \n\t"  // real part x2
+	"vbroadcastsd	40(%2), %%ymm5                  \n\t"  // imag part x2
+	"vbroadcastsd	48(%2), %%ymm6                  \n\t"  // real part x3
+	"vbroadcastsd	56(%2), %%ymm7                  \n\t"  // imag part x3
+
+
+	".align 16				        \n\t"
+	".L01LOOP%=:				        \n\t"
+	"prefetcht0      192(%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
+
+	"vxorpd		%%ymm12, %%ymm12, %%ymm12	\n\t"
+	"vxorpd		%%ymm13, %%ymm13, %%ymm13	\n\t"
+	"vxorpd		%%ymm14, %%ymm14, %%ymm14	\n\t"
+	"vxorpd		%%ymm15, %%ymm15, %%ymm15	\n\t"
+
+	"prefetcht0      192(%5,%0,8)			\n\t"
+	"vmovups	(%5,%0,8), %%ymm10              \n\t" // 2 complex values form a1
+	"vmovups      32(%5,%0,8), %%ymm11              \n\t" // 2 complex values form a1
+
+	"vfmadd231pd      %%ymm8 , %%ymm0, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmadd231pd      %%ymm8 , %%ymm1, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vfmadd231pd      %%ymm9 , %%ymm0, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vfmadd231pd      %%ymm9 , %%ymm1, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"prefetcht0      192(%6,%0,8)			\n\t"
+	"vmovups	(%6,%0,8), %%ymm8	        \n\t" // 2 complex values form a2
+	"vmovups      32(%6,%0,8), %%ymm9	        \n\t" // 2 complex values form a2
+
+	"vfmadd231pd      %%ymm10, %%ymm2, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmadd231pd      %%ymm10, %%ymm3, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vfmadd231pd      %%ymm11, %%ymm2, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vfmadd231pd      %%ymm11, %%ymm3, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"prefetcht0      192(%7,%0,8)			\n\t"
+	"vmovups	(%7,%0,8), %%ymm10              \n\t" // 2 complex values form a3
+	"vmovups      32(%7,%0,8), %%ymm11              \n\t" // 2 complex values form a3
+
+	"vfmadd231pd      %%ymm8 , %%ymm4, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmadd231pd      %%ymm8 , %%ymm5, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vfmadd231pd      %%ymm9 , %%ymm4, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vfmadd231pd      %%ymm9 , %%ymm5, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"vfmadd231pd      %%ymm10, %%ymm6, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmadd231pd      %%ymm10, %%ymm7, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vfmadd231pd      %%ymm11, %%ymm6, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vfmadd231pd      %%ymm11, %%ymm7, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilpd      $0x5 , %%ymm13, %%ymm13               \n\t"
+        "vpermilpd      $0x5 , %%ymm15, %%ymm15               \n\t"
+        "vaddsubpd      %%ymm13, %%ymm12, %%ymm8              \n\t"
+        "vaddsubpd      %%ymm15, %%ymm14, %%ymm9              \n\t"
+#else
+        "vpermilpd      $0x5 , %%ymm12, %%ymm12               \n\t"
+        "vpermilpd      $0x5 , %%ymm14, %%ymm14               \n\t"
+        "vaddsubpd      %%ymm13, %%ymm12, %%ymm8              \n\t"
+        "vaddsubpd      %%ymm15, %%ymm14, %%ymm9              \n\t"
+        "vpermilpd      $0x5 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilpd      $0x5 , %%ymm9 , %%ymm9                \n\t"
+#endif
+
+	"prefetcht0      192(%3,%0,8)			\n\t"
+	"vmovups	  (%3,%0,8),  %%ymm12           \n\t"
+	"vmovups	32(%3,%0,8),  %%ymm13           \n\t"
+
+#if !defined(XCONJ)
+        "vaddpd         %%ymm8, %%ymm12, %%ymm12              \n\t"
+        "vaddpd         %%ymm9, %%ymm13, %%ymm13              \n\t"
+#else
+        "vaddsubpd              %%ymm12, %%ymm8, %%ymm12              \n\t"
+        "vaddsubpd              %%ymm13, %%ymm9, %%ymm13              \n\t"
+#endif
+
+
+	"vmovups  %%ymm12,   (%3,%0,8)		        \n\t" // 2 complex values to y	
+	"vmovups  %%ymm13, 32(%3,%0,8)		        \n\t"	
+
+        "addq		$8 , %0	  	 	        \n\t"
+	"subq	        $4 , %1			        \n\t"		
+	"jnz		.L01LOOP%=		        \n\t"
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From dbc2eff029b298e324c175b698ec132436e6df43 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Sun, 10 Aug 2014 11:57:24 +0200
Subject: [PATCH 12/27] disabled optimized haswell zgemv_n kernel for windows (
 bad rounding )

---
 kernel/x86_64/KERNEL.HASWELL | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/kernel/x86_64/KERNEL.HASWELL b/kernel/x86_64/KERNEL.HASWELL
index d126eb6f4..7d4cddbcc 100644
--- a/kernel/x86_64/KERNEL.HASWELL
+++ b/kernel/x86_64/KERNEL.HASWELL
@@ -2,7 +2,10 @@ SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t.c
 DGEMVNKERNEL = dgemv_n.c
 DGEMVTKERNEL = dgemv_t.c
+
+ifndef OS_WINDOWS 
 ZGEMVNKERNEL = zgemv_n.c
+endif
 
 SGEMMKERNEL    =  sgemm_kernel_16x4_haswell.S
 SGEMMINCOPY    =  ../generic/gemm_ncopy_16.c

From 6fe416976d1ad2ff8f60829cb01a63a11d876429 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Mon, 11 Aug 2014 09:13:18 +0200
Subject: [PATCH 13/27] added optimimized zgemv_t c-kernel

---
 kernel/x86_64/zgemv_t.c | 267 ++++++++++++++++++++++++++++++++++++++++
 1 file changed, 267 insertions(+)
 create mode 100644 kernel/x86_64/zgemv_t.c

diff --git a/kernel/x86_64/zgemv_t.c b/kernel/x86_64/zgemv_t.c
new file mode 100644
index 000000000..a2dc45c45
--- /dev/null
+++ b/kernel/x86_64/zgemv_t.c
@@ -0,0 +1,267 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include "common.h"
+
+/*
+#if defined(HASWELL)
+#include "zgemv_t_microk_haswell-2.c"
+#endif
+*/
+
+#define NBMAX 1028
+
+#ifndef HAVE_KERNEL_16x4
+
+static void zgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0,*a1,*a2,*a3;
+	a0 = ap[0];
+	a1 = ap[1];
+	a2 = ap[2];
+	a3 = ap[3];
+	FLOAT temp_r0 = 0.0;
+	FLOAT temp_r1 = 0.0;
+	FLOAT temp_r2 = 0.0;
+	FLOAT temp_r3 = 0.0;
+	FLOAT temp_i0 = 0.0;
+	FLOAT temp_i1 = 0.0;
+	FLOAT temp_i2 = 0.0;
+	FLOAT temp_i3 = 0.0;
+
+
+	for ( i=0; i< 2*n; i+=2 )
+	{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+		temp_r0 += a0[i]*x[i]   - a0[i+1]*x[i+1];		
+		temp_i0 += a0[i]*x[i+1] + a0[i+1]*x[i];		
+		temp_r1 += a1[i]*x[i]   - a1[i+1]*x[i+1];		
+		temp_i1 += a1[i]*x[i+1] + a1[i+1]*x[i];		
+		temp_r2 += a2[i]*x[i]   - a2[i+1]*x[i+1];		
+		temp_i2 += a2[i]*x[i+1] + a2[i+1]*x[i];		
+		temp_r3 += a3[i]*x[i]   - a3[i+1]*x[i+1];		
+		temp_i3 += a3[i]*x[i+1] + a3[i+1]*x[i];		
+#else
+		temp_r0 += a0[i]*x[i]   + a0[i+1]*x[i+1];		
+		temp_i0 += a0[i]*x[i+1] - a0[i+1]*x[i];		
+		temp_r1 += a1[i]*x[i]   + a1[i+1]*x[i+1];		
+		temp_i1 += a1[i]*x[i+1] - a1[i+1]*x[i];		
+		temp_r2 += a2[i]*x[i]   + a2[i+1]*x[i+1];		
+		temp_i2 += a2[i]*x[i+1] - a2[i+1]*x[i];		
+		temp_r3 += a3[i]*x[i]   + a3[i+1]*x[i+1];		
+		temp_i3 += a3[i]*x[i+1] - a3[i+1]*x[i];		
+#endif
+	}
+	y[0] = temp_r0;
+	y[1] = temp_i0;
+	y[2] = temp_r1;
+	y[3] = temp_i1;
+	y[4] = temp_r2;
+	y[5] = temp_i2;
+	y[6] = temp_r3;
+	y[7] = temp_i3;
+}
+	
+#endif
+
+static void zgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0;
+	a0 = ap;
+	FLOAT temp_r = 0.0;
+	FLOAT temp_i = 0.0;
+
+	for ( i=0; i< 2*n; i+=2 )
+	{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+		temp_r += a0[i]*x[i]   - a0[i+1]*x[i+1];		
+		temp_i += a0[i]*x[i+1] + a0[i+1]*x[i];		
+#else
+		temp_r += a0[i]*x[i]   + a0[i+1]*x[i+1];		
+		temp_i += a0[i]*x[i+1] - a0[i+1]*x[i];		
+#endif
+	}
+	*y      = temp_r;
+	*(y+1)  = temp_i;
+}
+	
+static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
+{
+        BLASLONG i;
+        for ( i=0; i<n; i++ )
+        {
+                *dest     = *src;
+                *(dest+1) = *(src+1);
+                dest+=2;
+                src += inc_src;
+        }
+}
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG j;
+	FLOAT *a_ptr;
+	FLOAT *x_ptr;
+	FLOAT *y_ptr;
+	FLOAT *ap[8];
+	BLASLONG n1;
+	BLASLONG m1;
+	BLASLONG m2;
+	BLASLONG n2;
+	FLOAT ybuffer[8],*xbuffer;
+
+        inc_x *= 2;
+        inc_y *= 2;
+        lda   *= 2;
+
+	xbuffer = buffer;
+	
+	n1 = n / 4 ;
+	n2 = n % 4 ;
+	
+	m1 = m - ( m % 16 );
+	m2 = (m % NBMAX) - (m % 16) ;
+	
+
+	BLASLONG NB = NBMAX;
+
+	while ( NB == NBMAX )
+	{
+		
+		m1 -= NB;
+		if ( m1 < 0)
+		{
+			if ( m2 == 0 ) break;	
+			NB = m2;
+		}
+		
+		y_ptr = y;
+		a_ptr = a;
+		x_ptr = x;
+		copy_x(NB,x_ptr,xbuffer,inc_x);
+		for( i = 0; i < n1 ; i++)
+		{
+			ap[0] = a_ptr;
+			ap[1] = a_ptr + lda;
+			ap[2] = ap[1] + lda;
+			ap[3] = ap[2] + lda;
+			zgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
+			a_ptr += 4 * lda;
+
+#if !defined(XCONJ)
+			y_ptr[0] += alpha_r * ybuffer[0] - alpha_i * ybuffer[1];
+			y_ptr[1] += alpha_r * ybuffer[1] + alpha_i * ybuffer[0];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[2] - alpha_i * ybuffer[3];
+			y_ptr[1] += alpha_r * ybuffer[3] + alpha_i * ybuffer[2];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[4] - alpha_i * ybuffer[5];
+			y_ptr[1] += alpha_r * ybuffer[5] + alpha_i * ybuffer[4];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[6] - alpha_i * ybuffer[7];
+			y_ptr[1] += alpha_r * ybuffer[7] + alpha_i * ybuffer[6];
+			y_ptr  += inc_y;
+#else
+			y_ptr[0] += alpha_r * ybuffer[0] + alpha_i * ybuffer[1];
+			y_ptr[1] -= alpha_r * ybuffer[1] - alpha_i * ybuffer[0];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[2] + alpha_i * ybuffer[3];
+			y_ptr[1] -= alpha_r * ybuffer[3] - alpha_i * ybuffer[2];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[4] + alpha_i * ybuffer[5];
+			y_ptr[1] -= alpha_r * ybuffer[5] - alpha_i * ybuffer[4];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[6] + alpha_i * ybuffer[7];
+			y_ptr[1] -= alpha_r * ybuffer[7] - alpha_i * ybuffer[6];
+			y_ptr  += inc_y;
+#endif
+		}
+
+		for( i = 0; i < n2 ; i++)
+		{
+			zgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
+			a_ptr += 1 * lda;
+
+#if !defined(XCONJ)
+			y_ptr[0] += alpha_r * ybuffer[0] - alpha_i * ybuffer[1];
+			y_ptr[1] += alpha_r * ybuffer[1] + alpha_i * ybuffer[0];
+			y_ptr  += inc_y;
+#else
+			y_ptr[0] += alpha_r * ybuffer[0] + alpha_i * ybuffer[1];
+			y_ptr[1] -= alpha_r * ybuffer[1] - alpha_i * ybuffer[0];
+			y_ptr  += inc_y;
+#endif
+
+		}
+		a += 2* NB;
+		x += NB * inc_x;	
+	}
+
+	BLASLONG m3 = m % 16;
+	if ( m3 == 0 ) return(0);
+
+	x_ptr = x;
+	copy_x(m3,x_ptr,xbuffer,inc_x);
+	j=0;
+	a_ptr = a;
+	y_ptr = y;
+	while ( j < n)
+	{
+		FLOAT temp_r = 0.0;
+		FLOAT temp_i = 0.0;
+		for( i = 0; i < m3*2; i+=2 )
+		{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+			temp_r += a_ptr[i] * xbuffer[i]   - a_ptr[i+1] * xbuffer[i+1];
+			temp_i += a_ptr[i] * xbuffer[i+1] + a_ptr[i+1] * xbuffer[i];
+#else
+			temp_r += a_ptr[i] * xbuffer[i]   + a_ptr[i+1] * xbuffer[i+1];
+			temp_i += a_ptr[i] * xbuffer[i+1] - a_ptr[i+1] * xbuffer[i];
+#endif
+		}
+		a_ptr += lda;
+
+#if !defined(XCONJ) 
+                y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
+                y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
+#else
+                y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
+                y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
+#endif
+
+		y_ptr += inc_y;
+		j++;
+	}
+	return(0);
+}
+
+

From 726ad085cb00af5c13a7742a4b73df598953bf80 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Mon, 11 Aug 2014 13:10:12 +0200
Subject: [PATCH 14/27] added optimized zgemv_t for haswell

---
 kernel/x86_64/dgemv_n_microk_haswell-2.c   |   4 +-
 kernel/x86_64/dgemv_t_microk_haswell-2.c   |   4 +-
 kernel/x86_64/sgemv_n_microk_bulldozer-2.c |   4 +-
 kernel/x86_64/sgemv_n_microk_haswell-2.c   |   4 +-
 kernel/x86_64/sgemv_n_microk_nehalem-2.c   |   4 +-
 kernel/x86_64/sgemv_n_microk_sandy-2.c     |   4 +-
 kernel/x86_64/sgemv_t_microk_bulldozer-2.c |   4 +-
 kernel/x86_64/sgemv_t_microk_haswell-2.c   |   4 +-
 kernel/x86_64/sgemv_t_microk_nehalem-2.c   |   4 +-
 kernel/x86_64/sgemv_t_microk_sandy-2.c     |   4 +-
 kernel/x86_64/zgemv_n_microk_haswell-2.c   |   4 +-
 kernel/x86_64/zgemv_t_microk_bulldozer-2.c | 139 +++++++++++++++++++++
 kernel/x86_64/zgemv_t_microk_haswell-2.c   | 139 +++++++++++++++++++++
 13 files changed, 300 insertions(+), 22 deletions(-)
 create mode 100644 kernel/x86_64/zgemv_t_microk_bulldozer-2.c
 create mode 100644 kernel/x86_64/zgemv_t_microk_haswell-2.c

diff --git a/kernel/x86_64/dgemv_n_microk_haswell-2.c b/kernel/x86_64/dgemv_n_microk_haswell-2.c
index b9f462cb2..28e2fe4f6 100644
--- a/kernel/x86_64/dgemv_n_microk_haswell-2.c
+++ b/kernel/x86_64/dgemv_n_microk_haswell-2.c
@@ -26,9 +26,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *****************************************************************************/
 
 #define HAVE_KERNEL_16x4 1
-static void dgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y) __attribute__ ((noinline));
+static void dgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
 
-static void dgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
+static void dgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 {
 
 	BLASLONG register i = 0;
diff --git a/kernel/x86_64/dgemv_t_microk_haswell-2.c b/kernel/x86_64/dgemv_t_microk_haswell-2.c
index 94d4c319e..1a4ba37d7 100644
--- a/kernel/x86_64/dgemv_t_microk_haswell-2.c
+++ b/kernel/x86_64/dgemv_t_microk_haswell-2.c
@@ -26,9 +26,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *****************************************************************************/
 
 #define HAVE_KERNEL_16x4 1
-static void dgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y) __attribute__ ((noinline));
+static void dgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
 
-static void dgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
+static void dgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 {
 
 	BLASLONG register i = 0;
diff --git a/kernel/x86_64/sgemv_n_microk_bulldozer-2.c b/kernel/x86_64/sgemv_n_microk_bulldozer-2.c
index d50fa4268..c4a490587 100644
--- a/kernel/x86_64/sgemv_n_microk_bulldozer-2.c
+++ b/kernel/x86_64/sgemv_n_microk_bulldozer-2.c
@@ -26,9 +26,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *****************************************************************************/
 
 #define HAVE_KERNEL_16x4 1
-static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y) __attribute__ ((noinline));
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
 
-static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 {
 
 	BLASLONG register i = 0;
diff --git a/kernel/x86_64/sgemv_n_microk_haswell-2.c b/kernel/x86_64/sgemv_n_microk_haswell-2.c
index d3fee67c3..19888d150 100644
--- a/kernel/x86_64/sgemv_n_microk_haswell-2.c
+++ b/kernel/x86_64/sgemv_n_microk_haswell-2.c
@@ -26,9 +26,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *****************************************************************************/
 
 #define HAVE_KERNEL_16x4 1
-static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y) __attribute__ ((noinline));
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
 
-static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 {
 
 	BLASLONG register i = 0;
diff --git a/kernel/x86_64/sgemv_n_microk_nehalem-2.c b/kernel/x86_64/sgemv_n_microk_nehalem-2.c
index 3cfb82a45..40ccbb78f 100644
--- a/kernel/x86_64/sgemv_n_microk_nehalem-2.c
+++ b/kernel/x86_64/sgemv_n_microk_nehalem-2.c
@@ -26,9 +26,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *****************************************************************************/
 
 #define HAVE_KERNEL_16x4 1
-static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y) __attribute__ ((noinline));
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
 
-static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 {
 
 	BLASLONG register i = 0;
diff --git a/kernel/x86_64/sgemv_n_microk_sandy-2.c b/kernel/x86_64/sgemv_n_microk_sandy-2.c
index 21eff1c5e..b255ddbcb 100644
--- a/kernel/x86_64/sgemv_n_microk_sandy-2.c
+++ b/kernel/x86_64/sgemv_n_microk_sandy-2.c
@@ -26,9 +26,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *****************************************************************************/
 
 #define HAVE_KERNEL_16x4 1
-static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y) __attribute__ ((noinline));
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
 
-static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 {
 
 	BLASLONG register i = 0;
diff --git a/kernel/x86_64/sgemv_t_microk_bulldozer-2.c b/kernel/x86_64/sgemv_t_microk_bulldozer-2.c
index 54bdca63a..e4498afa3 100644
--- a/kernel/x86_64/sgemv_t_microk_bulldozer-2.c
+++ b/kernel/x86_64/sgemv_t_microk_bulldozer-2.c
@@ -26,9 +26,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *****************************************************************************/
 
 #define HAVE_KERNEL_16x4 1
-static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y) __attribute__ ((noinline));
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
 
-static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 {
 
 	BLASLONG register i = 0;
diff --git a/kernel/x86_64/sgemv_t_microk_haswell-2.c b/kernel/x86_64/sgemv_t_microk_haswell-2.c
index cef703483..e6d47270d 100644
--- a/kernel/x86_64/sgemv_t_microk_haswell-2.c
+++ b/kernel/x86_64/sgemv_t_microk_haswell-2.c
@@ -26,9 +26,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *****************************************************************************/
 
 #define HAVE_KERNEL_16x4 1
-static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y) __attribute__ ((noinline));
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
 
-static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 {
 
 	BLASLONG register i = 0;
diff --git a/kernel/x86_64/sgemv_t_microk_nehalem-2.c b/kernel/x86_64/sgemv_t_microk_nehalem-2.c
index e1f2b81bd..db5a1448b 100644
--- a/kernel/x86_64/sgemv_t_microk_nehalem-2.c
+++ b/kernel/x86_64/sgemv_t_microk_nehalem-2.c
@@ -26,9 +26,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *****************************************************************************/
 
 #define HAVE_KERNEL_16x4 1
-static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y) __attribute__ ((noinline));
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
 
-static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 {
 
 	BLASLONG register i = 0;
diff --git a/kernel/x86_64/sgemv_t_microk_sandy-2.c b/kernel/x86_64/sgemv_t_microk_sandy-2.c
index 6a3748238..841522302 100644
--- a/kernel/x86_64/sgemv_t_microk_sandy-2.c
+++ b/kernel/x86_64/sgemv_t_microk_sandy-2.c
@@ -26,9 +26,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *****************************************************************************/
 
 #define HAVE_KERNEL_16x4 1
-static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y) __attribute__ ((noinline));
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
 
-static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 {
 
 	BLASLONG register i = 0;
diff --git a/kernel/x86_64/zgemv_n_microk_haswell-2.c b/kernel/x86_64/zgemv_n_microk_haswell-2.c
index 8583f96b3..833983fe0 100644
--- a/kernel/x86_64/zgemv_n_microk_haswell-2.c
+++ b/kernel/x86_64/zgemv_n_microk_haswell-2.c
@@ -26,9 +26,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *****************************************************************************/
 
 #define HAVE_KERNEL_16x4 1
-static void zgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y) __attribute__ ((noinline));
+static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
 
-static void zgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
+static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 {
 
 	BLASLONG register i = 0;
diff --git a/kernel/x86_64/zgemv_t_microk_bulldozer-2.c b/kernel/x86_64/zgemv_t_microk_bulldozer-2.c
new file mode 100644
index 000000000..efb6d784e
--- /dev/null
+++ b/kernel/x86_64/zgemv_t_microk_bulldozer-2.c
@@ -0,0 +1,139 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary froms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary from must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+
+	"vxorpd		%%ymm8 , %%ymm8 , %%ymm8 	\n\t" // temp
+	"vxorpd		%%ymm9 , %%ymm9 , %%ymm9 	\n\t" // temp
+	"vxorpd		%%ymm10, %%ymm10, %%ymm10	\n\t" // temp
+	"vxorpd		%%ymm11, %%ymm11, %%ymm11	\n\t" // temp
+	"vxorpd		%%ymm12, %%ymm12, %%ymm12	\n\t" // temp
+	"vxorpd		%%ymm13, %%ymm13, %%ymm13	\n\t"
+	"vxorpd		%%ymm14, %%ymm14, %%ymm14	\n\t"
+	"vxorpd		%%ymm15, %%ymm15, %%ymm15	\n\t"
+
+	".align 16				        \n\t"
+	".L01LOOP%=:				        \n\t"
+
+	"vmovddup	   (%2,%0,8), %%xmm0            \n\t"  // real value from x0
+	"vmovddup	  8(%2,%0,8), %%xmm1            \n\t"  // imag value from x0
+	"vmovddup	 16(%2,%0,8), %%xmm2            \n\t"  // real value from x1
+	"vmovddup	 24(%2,%0,8), %%xmm3            \n\t"  // imag value from x1
+	"vinsertf128	 $1, %%xmm2, %%ymm0 , %%ymm0	\n\t"  // real values from x0 and x1
+	"vinsertf128	 $1, %%xmm3, %%ymm1 , %%ymm1	\n\t"  // imag values from x0 and x1
+
+	"vmovups	(%4,%0,8), %%ymm4	        \n\t" // 2 complex values from a0
+	"vmovups	(%5,%0,8), %%ymm5               \n\t" // 2 complex values from a1
+	"vmovups	(%6,%0,8), %%ymm6	        \n\t" // 2 complex values from a2
+	"vmovups	(%7,%0,8), %%ymm7               \n\t" // 2 complex values from a3
+
+	"vfmaddpd   %%ymm8 ,   %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddpd   %%ymm9 ,   %%ymm4 , %%ymm1, %%ymm9       \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmaddpd   %%ymm10,   %%ymm5 , %%ymm0, %%ymm10      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddpd   %%ymm11,   %%ymm5 , %%ymm1, %%ymm11      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmaddpd   %%ymm12,   %%ymm6 , %%ymm0, %%ymm12      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddpd   %%ymm13,   %%ymm6 , %%ymm1, %%ymm13      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmaddpd   %%ymm14,   %%ymm7 , %%ymm0, %%ymm14      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddpd   %%ymm15,   %%ymm7 , %%ymm1, %%ymm15      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+
+
+        "addq		$4 , %0	  	 	        \n\t"
+	"subq	        $2 , %1			        \n\t"		
+	"jnz		.L01LOOP%=		        \n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilpd      $0x5 , %%ymm9 , %%ymm9                \n\t"
+        "vpermilpd      $0x5 , %%ymm11, %%ymm11               \n\t"
+        "vpermilpd      $0x5 , %%ymm13, %%ymm13               \n\t"
+        "vpermilpd      $0x5 , %%ymm15, %%ymm15               \n\t"
+        "vaddsubpd      %%ymm9 , %%ymm8, %%ymm8               \n\t" 
+        "vaddsubpd      %%ymm11, %%ymm10, %%ymm10             \n\t"
+        "vaddsubpd      %%ymm13, %%ymm12, %%ymm12             \n\t"
+        "vaddsubpd      %%ymm15, %%ymm14, %%ymm14             \n\t"
+#else
+        "vpermilpd      $0x5 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilpd      $0x5 , %%ymm10, %%ymm10               \n\t"
+        "vpermilpd      $0x5 , %%ymm12, %%ymm12               \n\t"
+        "vpermilpd      $0x5 , %%ymm14, %%ymm14               \n\t"
+        "vaddsubpd      %%ymm8 , %%ymm9 , %%ymm8              \n\t"
+        "vaddsubpd      %%ymm10, %%ymm11, %%ymm10             \n\t"
+        "vaddsubpd      %%ymm12, %%ymm13, %%ymm12             \n\t"
+        "vaddsubpd      %%ymm14, %%ymm15, %%ymm14             \n\t"
+        "vpermilpd      $0x5 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilpd      $0x5 , %%ymm10, %%ymm10               \n\t"
+        "vpermilpd      $0x5 , %%ymm12, %%ymm12               \n\t"
+        "vpermilpd      $0x5 , %%ymm14, %%ymm14               \n\t"
+#endif
+
+	"vextractf128   $1, %%ymm8 , %%xmm9		      \n\t"
+	"vextractf128   $1, %%ymm10, %%xmm11	      	      \n\t"
+	"vextractf128   $1, %%ymm12, %%xmm13		      \n\t"
+	"vextractf128   $1, %%ymm14, %%xmm15		      \n\t"
+
+	"vaddpd		%%xmm8 , %%xmm9 , %%xmm8       \n\t"
+	"vaddpd		%%xmm10, %%xmm11, %%xmm10      \n\t"
+	"vaddpd		%%xmm12, %%xmm13, %%xmm12      \n\t"
+	"vaddpd		%%xmm14, %%xmm15, %%xmm14      \n\t"
+
+	"vmovups	%%xmm8 ,   (%3)			\n\t"
+	"vmovups	%%xmm10, 16(%3)			\n\t"
+	"vmovups	%%xmm12, 32(%3)			\n\t"
+	"vmovups	%%xmm14, 48(%3)			\n\t"
+
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
diff --git a/kernel/x86_64/zgemv_t_microk_haswell-2.c b/kernel/x86_64/zgemv_t_microk_haswell-2.c
new file mode 100644
index 000000000..2dddef27d
--- /dev/null
+++ b/kernel/x86_64/zgemv_t_microk_haswell-2.c
@@ -0,0 +1,139 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary froms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary from must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+
+	"vxorpd		%%ymm8 , %%ymm8 , %%ymm8 	\n\t" // temp
+	"vxorpd		%%ymm9 , %%ymm9 , %%ymm9 	\n\t" // temp
+	"vxorpd		%%ymm10, %%ymm10, %%ymm10	\n\t" // temp
+	"vxorpd		%%ymm11, %%ymm11, %%ymm11	\n\t" // temp
+	"vxorpd		%%ymm12, %%ymm12, %%ymm12	\n\t" // temp
+	"vxorpd		%%ymm13, %%ymm13, %%ymm13	\n\t"
+	"vxorpd		%%ymm14, %%ymm14, %%ymm14	\n\t"
+	"vxorpd		%%ymm15, %%ymm15, %%ymm15	\n\t"
+
+	".align 16				        \n\t"
+	".L01LOOP%=:				        \n\t"
+
+	"vmovddup	   (%2,%0,8), %%xmm0            \n\t"  // real value from x0
+	"vmovddup	  8(%2,%0,8), %%xmm1            \n\t"  // imag value from x0
+	"vmovddup	 16(%2,%0,8), %%xmm2            \n\t"  // real value from x1
+	"vmovddup	 24(%2,%0,8), %%xmm3            \n\t"  // imag value from x1
+	"vinsertf128	 $1, %%xmm2, %%ymm0 , %%ymm0	\n\t"  // real values from x0 and x1
+	"vinsertf128	 $1, %%xmm3, %%ymm1 , %%ymm1	\n\t"  // imag values from x0 and x1
+
+	"vmovups	(%4,%0,8), %%ymm4	        \n\t" // 2 complex values from a0
+	"vmovups	(%5,%0,8), %%ymm5               \n\t" // 2 complex values from a1
+	"vmovups	(%6,%0,8), %%ymm6	        \n\t" // 2 complex values from a2
+	"vmovups	(%7,%0,8), %%ymm7               \n\t" // 2 complex values from a3
+
+	"vfmadd231pd      %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231pd      %%ymm4 , %%ymm1, %%ymm9       \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231pd      %%ymm5 , %%ymm0, %%ymm10      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231pd      %%ymm5 , %%ymm1, %%ymm11      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231pd      %%ymm6 , %%ymm0, %%ymm12      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231pd      %%ymm6 , %%ymm1, %%ymm13      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231pd      %%ymm7 , %%ymm0, %%ymm14      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231pd      %%ymm7 , %%ymm1, %%ymm15      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+
+
+        "addq		$4 , %0	  	 	        \n\t"
+	"subq	        $2 , %1			        \n\t"		
+	"jnz		.L01LOOP%=		        \n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilpd      $0x5 , %%ymm9 , %%ymm9                \n\t"
+        "vpermilpd      $0x5 , %%ymm11, %%ymm11               \n\t"
+        "vpermilpd      $0x5 , %%ymm13, %%ymm13               \n\t"
+        "vpermilpd      $0x5 , %%ymm15, %%ymm15               \n\t"
+        "vaddsubpd      %%ymm9 , %%ymm8, %%ymm8               \n\t" 
+        "vaddsubpd      %%ymm11, %%ymm10, %%ymm10             \n\t"
+        "vaddsubpd      %%ymm13, %%ymm12, %%ymm12             \n\t"
+        "vaddsubpd      %%ymm15, %%ymm14, %%ymm14             \n\t"
+#else
+        "vpermilpd      $0x5 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilpd      $0x5 , %%ymm10, %%ymm10               \n\t"
+        "vpermilpd      $0x5 , %%ymm12, %%ymm12               \n\t"
+        "vpermilpd      $0x5 , %%ymm14, %%ymm14               \n\t"
+        "vaddsubpd      %%ymm8 , %%ymm9 , %%ymm8              \n\t"
+        "vaddsubpd      %%ymm10, %%ymm11, %%ymm10             \n\t"
+        "vaddsubpd      %%ymm12, %%ymm13, %%ymm12             \n\t"
+        "vaddsubpd      %%ymm14, %%ymm15, %%ymm14             \n\t"
+        "vpermilpd      $0x5 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilpd      $0x5 , %%ymm10, %%ymm10               \n\t"
+        "vpermilpd      $0x5 , %%ymm12, %%ymm12               \n\t"
+        "vpermilpd      $0x5 , %%ymm14, %%ymm14               \n\t"
+#endif
+
+	"vextractf128   $1, %%ymm8 , %%xmm9		      \n\t"
+	"vextractf128   $1, %%ymm10, %%xmm11	      	      \n\t"
+	"vextractf128   $1, %%ymm12, %%xmm13		      \n\t"
+	"vextractf128   $1, %%ymm14, %%xmm15		      \n\t"
+
+	"vaddpd		%%xmm8 , %%xmm9 , %%xmm8       \n\t"
+	"vaddpd		%%xmm10, %%xmm11, %%xmm10      \n\t"
+	"vaddpd		%%xmm12, %%xmm13, %%xmm12      \n\t"
+	"vaddpd		%%xmm14, %%xmm15, %%xmm14      \n\t"
+
+	"vmovups	%%xmm8 ,   (%3)			\n\t"
+	"vmovups	%%xmm10, 16(%3)			\n\t"
+	"vmovups	%%xmm12, 32(%3)			\n\t"
+	"vmovups	%%xmm14, 48(%3)			\n\t"
+
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From 09fcd3a34135ad86f4b17f8e15893fe99b9f0171 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Mon, 11 Aug 2014 14:19:25 +0200
Subject: [PATCH 15/27] add optimized zgemv_t kernel for bulldozer

---
 kernel/x86_64/KERNEL.BULLDOZER             |   2 +-
 kernel/x86_64/zgemv_t.c                    |   8 +-
 kernel/x86_64/zgemv_t_microk_bulldozer-2.c | 151 +++++++++++++--------
 3 files changed, 101 insertions(+), 60 deletions(-)

diff --git a/kernel/x86_64/KERNEL.BULLDOZER b/kernel/x86_64/KERNEL.BULLDOZER
index 893f13064..19bf7fd32 100644
--- a/kernel/x86_64/KERNEL.BULLDOZER
+++ b/kernel/x86_64/KERNEL.BULLDOZER
@@ -2,7 +2,7 @@ SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t.c
 
 ZGEMVNKERNEL = zgemv_n_dup.S
-ZGEMVTKERNEL = zgemv_t.S
+ZGEMVTKERNEL = zgemv_t.c
 
 DGEMVNKERNEL = dgemv_n_bulldozer.S
 DGEMVTKERNEL = dgemv_t_bulldozer.S
diff --git a/kernel/x86_64/zgemv_t.c b/kernel/x86_64/zgemv_t.c
index a2dc45c45..b54d5f4e2 100644
--- a/kernel/x86_64/zgemv_t.c
+++ b/kernel/x86_64/zgemv_t.c
@@ -28,11 +28,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include "common.h"
 
-/*
-#if defined(HASWELL)
-#include "zgemv_t_microk_haswell-2.c"
+
+#if defined(BULLDOZER)
+#include "zgemv_t_microk_bulldozer-2.c"
 #endif
-*/
+
 
 #define NBMAX 1028
 
diff --git a/kernel/x86_64/zgemv_t_microk_bulldozer-2.c b/kernel/x86_64/zgemv_t_microk_bulldozer-2.c
index efb6d784e..65d5a10a2 100644
--- a/kernel/x86_64/zgemv_t_microk_bulldozer-2.c
+++ b/kernel/x86_64/zgemv_t_microk_bulldozer-2.c
@@ -37,77 +37,118 @@ static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	(
 	"vzeroupper			 \n\t"
 
-	"vxorpd		%%ymm8 , %%ymm8 , %%ymm8 	\n\t" // temp
-	"vxorpd		%%ymm9 , %%ymm9 , %%ymm9 	\n\t" // temp
-	"vxorpd		%%ymm10, %%ymm10, %%ymm10	\n\t" // temp
-	"vxorpd		%%ymm11, %%ymm11, %%ymm11	\n\t" // temp
-	"vxorpd		%%ymm12, %%ymm12, %%ymm12	\n\t" // temp
-	"vxorpd		%%ymm13, %%ymm13, %%ymm13	\n\t"
-	"vxorpd		%%ymm14, %%ymm14, %%ymm14	\n\t"
-	"vxorpd		%%ymm15, %%ymm15, %%ymm15	\n\t"
+	"vxorpd		%%xmm8 , %%xmm8 , %%xmm8 	\n\t" // temp
+	"vxorpd		%%xmm9 , %%xmm9 , %%xmm9 	\n\t" // temp
+	"vxorpd		%%xmm10, %%xmm10, %%xmm10	\n\t" // temp
+	"vxorpd		%%xmm11, %%xmm11, %%xmm11	\n\t" // temp
+	"vxorpd		%%xmm12, %%xmm12, %%xmm12	\n\t" // temp
+	"vxorpd		%%xmm13, %%xmm13, %%xmm13	\n\t"
+	"vxorpd		%%xmm14, %%xmm14, %%xmm14	\n\t"
+	"vxorpd		%%xmm15, %%xmm15, %%xmm15	\n\t"
 
 	".align 16				        \n\t"
 	".L01LOOP%=:				        \n\t"
 
 	"vmovddup	   (%2,%0,8), %%xmm0            \n\t"  // real value from x0
 	"vmovddup	  8(%2,%0,8), %%xmm1            \n\t"  // imag value from x0
-	"vmovddup	 16(%2,%0,8), %%xmm2            \n\t"  // real value from x1
-	"vmovddup	 24(%2,%0,8), %%xmm3            \n\t"  // imag value from x1
-	"vinsertf128	 $1, %%xmm2, %%ymm0 , %%ymm0	\n\t"  // real values from x0 and x1
-	"vinsertf128	 $1, %%xmm3, %%ymm1 , %%ymm1	\n\t"  // imag values from x0 and x1
 
-	"vmovups	(%4,%0,8), %%ymm4	        \n\t" // 2 complex values from a0
-	"vmovups	(%5,%0,8), %%ymm5               \n\t" // 2 complex values from a1
-	"vmovups	(%6,%0,8), %%ymm6	        \n\t" // 2 complex values from a2
-	"vmovups	(%7,%0,8), %%ymm7               \n\t" // 2 complex values from a3
+	"prefetcht0  192(%4,%0,8)                       \n\t"
+	"vmovups	(%4,%0,8), %%xmm4	        \n\t" // 1 complex values from a0
+	"prefetcht0  192(%5,%0,8)                       \n\t"
+	"vmovups	(%5,%0,8), %%xmm5               \n\t" // 1 complex values from a1
+	"prefetcht0  192(%6,%0,8)                       \n\t"
+	"vmovups	(%6,%0,8), %%xmm6	        \n\t" // 1 complex values from a2
+	"prefetcht0  192(%7,%0,8)                       \n\t"
+	"vmovups	(%7,%0,8), %%xmm7               \n\t" // 1 complex values from a3
 
-	"vfmaddpd   %%ymm8 ,   %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
-	"vfmaddpd   %%ymm9 ,   %%ymm4 , %%ymm1, %%ymm9       \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
-	"vfmaddpd   %%ymm10,   %%ymm5 , %%ymm0, %%ymm10      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
-	"vfmaddpd   %%ymm11,   %%ymm5 , %%ymm1, %%ymm11      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
-	"vfmaddpd   %%ymm12,   %%ymm6 , %%ymm0, %%ymm12      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
-	"vfmaddpd   %%ymm13,   %%ymm6 , %%ymm1, %%ymm13      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
-	"vfmaddpd   %%ymm14,   %%ymm7 , %%ymm0, %%ymm14      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
-	"vfmaddpd   %%ymm15,   %%ymm7 , %%ymm1, %%ymm15      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmaddpd   %%xmm8 ,   %%xmm4 , %%xmm0, %%xmm8       \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm9 ,   %%xmm4 , %%xmm1, %%xmm9       \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm10,   %%xmm5 , %%xmm0, %%xmm10      \n\t" // ar0*xr0,al0*xr0
+	"vfmaddpd   %%xmm11,   %%xmm5 , %%xmm1, %%xmm11      \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm12,   %%xmm6 , %%xmm0, %%xmm12      \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm13,   %%xmm6 , %%xmm1, %%xmm13      \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm14,   %%xmm7 , %%xmm0, %%xmm14      \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm15,   %%xmm7 , %%xmm1, %%xmm15      \n\t" // ar0*xl0,al0*xl0 
 
+	"vmovddup	 16(%2,%0,8), %%xmm0            \n\t"  // real value from x0
+	"vmovddup	 24(%2,%0,8), %%xmm1            \n\t"  // imag value from x0
 
-        "addq		$4 , %0	  	 	        \n\t"
-	"subq	        $2 , %1			        \n\t"		
+	"vmovups      16(%4,%0,8), %%xmm4	        \n\t" // 1 complex values from a0
+	"vmovups      16(%5,%0,8), %%xmm5               \n\t" // 1 complex values from a1
+	"vmovups      16(%6,%0,8), %%xmm6	        \n\t" // 1 complex values from a2
+	"vmovups      16(%7,%0,8), %%xmm7               \n\t" // 1 complex values from a3
+
+	"vfmaddpd   %%xmm8 ,   %%xmm4 , %%xmm0, %%xmm8       \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm9 ,   %%xmm4 , %%xmm1, %%xmm9       \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm10,   %%xmm5 , %%xmm0, %%xmm10      \n\t" // ar0*xr0,al0*xr0
+	"vfmaddpd   %%xmm11,   %%xmm5 , %%xmm1, %%xmm11      \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm12,   %%xmm6 , %%xmm0, %%xmm12      \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm13,   %%xmm6 , %%xmm1, %%xmm13      \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm14,   %%xmm7 , %%xmm0, %%xmm14      \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm15,   %%xmm7 , %%xmm1, %%xmm15      \n\t" // ar0*xl0,al0*xl0 
+
+	"vmovddup	 32(%2,%0,8), %%xmm0            \n\t"  // real value from x0
+	"vmovddup	 40(%2,%0,8), %%xmm1            \n\t"  // imag value from x0
+
+	"vmovups      32(%4,%0,8), %%xmm4	        \n\t" // 1 complex values from a0
+	"vmovups      32(%5,%0,8), %%xmm5               \n\t" // 1 complex values from a1
+	"vmovups      32(%6,%0,8), %%xmm6	        \n\t" // 1 complex values from a2
+	"vmovups      32(%7,%0,8), %%xmm7               \n\t" // 1 complex values from a3
+
+	"vfmaddpd   %%xmm8 ,   %%xmm4 , %%xmm0, %%xmm8       \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm9 ,   %%xmm4 , %%xmm1, %%xmm9       \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm10,   %%xmm5 , %%xmm0, %%xmm10      \n\t" // ar0*xr0,al0*xr0
+	"vfmaddpd   %%xmm11,   %%xmm5 , %%xmm1, %%xmm11      \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm12,   %%xmm6 , %%xmm0, %%xmm12      \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm13,   %%xmm6 , %%xmm1, %%xmm13      \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm14,   %%xmm7 , %%xmm0, %%xmm14      \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm15,   %%xmm7 , %%xmm1, %%xmm15      \n\t" // ar0*xl0,al0*xl0 
+
+	"vmovddup	 48(%2,%0,8), %%xmm0            \n\t"  // real value from x0
+	"vmovddup	 56(%2,%0,8), %%xmm1            \n\t"  // imag value from x0
+
+	"vmovups      48(%4,%0,8), %%xmm4	        \n\t" // 1 complex values from a0
+	"vmovups      48(%5,%0,8), %%xmm5               \n\t" // 1 complex values from a1
+	"vmovups      48(%6,%0,8), %%xmm6	        \n\t" // 1 complex values from a2
+	"vmovups      48(%7,%0,8), %%xmm7               \n\t" // 1 complex values from a3
+
+	"vfmaddpd   %%xmm8 ,   %%xmm4 , %%xmm0, %%xmm8       \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm9 ,   %%xmm4 , %%xmm1, %%xmm9       \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm10,   %%xmm5 , %%xmm0, %%xmm10      \n\t" // ar0*xr0,al0*xr0
+	"vfmaddpd   %%xmm11,   %%xmm5 , %%xmm1, %%xmm11      \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm12,   %%xmm6 , %%xmm0, %%xmm12      \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm13,   %%xmm6 , %%xmm1, %%xmm13      \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm14,   %%xmm7 , %%xmm0, %%xmm14      \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm15,   %%xmm7 , %%xmm1, %%xmm15      \n\t" // ar0*xl0,al0*xl0 
+
+        "addq		$8 , %0	  	 	        \n\t"
+	"subq	        $4 , %1			        \n\t"		
 	"jnz		.L01LOOP%=		        \n\t"
 
 #if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
-        "vpermilpd      $0x5 , %%ymm9 , %%ymm9                \n\t"
-        "vpermilpd      $0x5 , %%ymm11, %%ymm11               \n\t"
-        "vpermilpd      $0x5 , %%ymm13, %%ymm13               \n\t"
-        "vpermilpd      $0x5 , %%ymm15, %%ymm15               \n\t"
-        "vaddsubpd      %%ymm9 , %%ymm8, %%ymm8               \n\t" 
-        "vaddsubpd      %%ymm11, %%ymm10, %%ymm10             \n\t"
-        "vaddsubpd      %%ymm13, %%ymm12, %%ymm12             \n\t"
-        "vaddsubpd      %%ymm15, %%ymm14, %%ymm14             \n\t"
+        "vpermilpd      $0x1 , %%xmm9 , %%xmm9                \n\t"
+        "vpermilpd      $0x1 , %%xmm11, %%xmm11               \n\t"
+        "vpermilpd      $0x1 , %%xmm13, %%xmm13               \n\t"
+        "vpermilpd      $0x1 , %%xmm15, %%xmm15               \n\t"
+        "vaddsubpd      %%xmm9 , %%xmm8, %%xmm8               \n\t" 
+        "vaddsubpd      %%xmm11, %%xmm10, %%xmm10             \n\t"
+        "vaddsubpd      %%xmm13, %%xmm12, %%xmm12             \n\t"
+        "vaddsubpd      %%xmm15, %%xmm14, %%xmm14             \n\t"
 #else
-        "vpermilpd      $0x5 , %%ymm8 , %%ymm8                \n\t"
-        "vpermilpd      $0x5 , %%ymm10, %%ymm10               \n\t"
-        "vpermilpd      $0x5 , %%ymm12, %%ymm12               \n\t"
-        "vpermilpd      $0x5 , %%ymm14, %%ymm14               \n\t"
-        "vaddsubpd      %%ymm8 , %%ymm9 , %%ymm8              \n\t"
-        "vaddsubpd      %%ymm10, %%ymm11, %%ymm10             \n\t"
-        "vaddsubpd      %%ymm12, %%ymm13, %%ymm12             \n\t"
-        "vaddsubpd      %%ymm14, %%ymm15, %%ymm14             \n\t"
-        "vpermilpd      $0x5 , %%ymm8 , %%ymm8                \n\t"
-        "vpermilpd      $0x5 , %%ymm10, %%ymm10               \n\t"
-        "vpermilpd      $0x5 , %%ymm12, %%ymm12               \n\t"
-        "vpermilpd      $0x5 , %%ymm14, %%ymm14               \n\t"
+        "vpermilpd      $0x1 , %%xmm8 , %%xmm8                \n\t"
+        "vpermilpd      $0x1 , %%xmm10, %%xmm10               \n\t"
+        "vpermilpd      $0x1 , %%xmm12, %%xmm12               \n\t"
+        "vpermilpd      $0x1 , %%xmm14, %%xmm14               \n\t"
+        "vaddsubpd      %%xmm8 , %%xmm9 , %%xmm8              \n\t"
+        "vaddsubpd      %%xmm10, %%xmm11, %%xmm10             \n\t"
+        "vaddsubpd      %%xmm12, %%xmm13, %%xmm12             \n\t"
+        "vaddsubpd      %%xmm14, %%xmm15, %%xmm14             \n\t"
+        "vpermilpd      $0x1 , %%xmm8 , %%xmm8                \n\t"
+        "vpermilpd      $0x1 , %%xmm10, %%xmm10               \n\t"
+        "vpermilpd      $0x1 , %%xmm12, %%xmm12               \n\t"
+        "vpermilpd      $0x1 , %%xmm14, %%xmm14               \n\t"
 #endif
 
-	"vextractf128   $1, %%ymm8 , %%xmm9		      \n\t"
-	"vextractf128   $1, %%ymm10, %%xmm11	      	      \n\t"
-	"vextractf128   $1, %%ymm12, %%xmm13		      \n\t"
-	"vextractf128   $1, %%ymm14, %%xmm15		      \n\t"
-
-	"vaddpd		%%xmm8 , %%xmm9 , %%xmm8       \n\t"
-	"vaddpd		%%xmm10, %%xmm11, %%xmm10      \n\t"
-	"vaddpd		%%xmm12, %%xmm13, %%xmm12      \n\t"
-	"vaddpd		%%xmm14, %%xmm15, %%xmm14      \n\t"
 
 	"vmovups	%%xmm8 ,   (%3)			\n\t"
 	"vmovups	%%xmm10, 16(%3)			\n\t"

From 58b075daef4c65b02951ed8a8fd78dc53cab0893 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Mon, 11 Aug 2014 16:57:52 +0200
Subject: [PATCH 16/27] added optimized zgemv_t kernel for haswell

---
 kernel/x86_64/KERNEL.HASWELL             |  1 +
 kernel/x86_64/zgemv_t.c                  |  2 ++
 kernel/x86_64/zgemv_t_microk_haswell-2.c | 29 ++++++++++++++++++++++--
 3 files changed, 30 insertions(+), 2 deletions(-)

diff --git a/kernel/x86_64/KERNEL.HASWELL b/kernel/x86_64/KERNEL.HASWELL
index 7d4cddbcc..9a48289c5 100644
--- a/kernel/x86_64/KERNEL.HASWELL
+++ b/kernel/x86_64/KERNEL.HASWELL
@@ -6,6 +6,7 @@ DGEMVTKERNEL = dgemv_t.c
 ifndef OS_WINDOWS 
 ZGEMVNKERNEL = zgemv_n.c
 endif
+ZGEMVTKERNEL = zgemv_t.c
 
 SGEMMKERNEL    =  sgemm_kernel_16x4_haswell.S
 SGEMMINCOPY    =  ../generic/gemm_ncopy_16.c
diff --git a/kernel/x86_64/zgemv_t.c b/kernel/x86_64/zgemv_t.c
index b54d5f4e2..df75afeff 100644
--- a/kernel/x86_64/zgemv_t.c
+++ b/kernel/x86_64/zgemv_t.c
@@ -31,6 +31,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #if defined(BULLDOZER)
 #include "zgemv_t_microk_bulldozer-2.c"
+#elif defined(HASWELL)
+#include "zgemv_t_microk_haswell-2.c"
 #endif
 
 
diff --git a/kernel/x86_64/zgemv_t_microk_haswell-2.c b/kernel/x86_64/zgemv_t_microk_haswell-2.c
index 2dddef27d..99a620e44 100644
--- a/kernel/x86_64/zgemv_t_microk_haswell-2.c
+++ b/kernel/x86_64/zgemv_t_microk_haswell-2.c
@@ -49,6 +49,7 @@ static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	".align 16				        \n\t"
 	".L01LOOP%=:				        \n\t"
 
+        "prefetcht0      384(%2,%0,8)                   \n\t"
 	"vmovddup	   (%2,%0,8), %%xmm0            \n\t"  // real value from x0
 	"vmovddup	  8(%2,%0,8), %%xmm1            \n\t"  // imag value from x0
 	"vmovddup	 16(%2,%0,8), %%xmm2            \n\t"  // real value from x1
@@ -56,9 +57,13 @@ static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	"vinsertf128	 $1, %%xmm2, %%ymm0 , %%ymm0	\n\t"  // real values from x0 and x1
 	"vinsertf128	 $1, %%xmm3, %%ymm1 , %%ymm1	\n\t"  // imag values from x0 and x1
 
+        "prefetcht0      384(%4,%0,8)                   \n\t"
 	"vmovups	(%4,%0,8), %%ymm4	        \n\t" // 2 complex values from a0
+        "prefetcht0      384(%5,%0,8)                   \n\t"
 	"vmovups	(%5,%0,8), %%ymm5               \n\t" // 2 complex values from a1
+        "prefetcht0      384(%6,%0,8)                   \n\t"
 	"vmovups	(%6,%0,8), %%ymm6	        \n\t" // 2 complex values from a2
+        "prefetcht0      384(%7,%0,8)                   \n\t"
 	"vmovups	(%7,%0,8), %%ymm7               \n\t" // 2 complex values from a3
 
 	"vfmadd231pd      %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
@@ -70,9 +75,29 @@ static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	"vfmadd231pd      %%ymm7 , %%ymm0, %%ymm14      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
 	"vfmadd231pd      %%ymm7 , %%ymm1, %%ymm15      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
 
+	"vmovddup	 32(%2,%0,8), %%xmm0            \n\t"  // real value from x0
+	"vmovddup	 40(%2,%0,8), %%xmm1            \n\t"  // imag value from x0
+	"vmovddup	 48(%2,%0,8), %%xmm2            \n\t"  // real value from x1
+	"vmovddup	 56(%2,%0,8), %%xmm3            \n\t"  // imag value from x1
+	"vinsertf128	 $1, %%xmm2, %%ymm0 , %%ymm0	\n\t"  // real values from x0 and x1
+	"vinsertf128	 $1, %%xmm3, %%ymm1 , %%ymm1	\n\t"  // imag values from x0 and x1
 
-        "addq		$4 , %0	  	 	        \n\t"
-	"subq	        $2 , %1			        \n\t"		
+	"vmovups       32(%4,%0,8), %%ymm4	        \n\t" // 2 complex values from a0
+	"vmovups       32(%5,%0,8), %%ymm5              \n\t" // 2 complex values from a1
+	"vmovups       32(%6,%0,8), %%ymm6	        \n\t" // 2 complex values from a2
+	"vmovups       32(%7,%0,8), %%ymm7               \n\t" // 2 complex values from a3
+
+	"vfmadd231pd      %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231pd      %%ymm4 , %%ymm1, %%ymm9       \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231pd      %%ymm5 , %%ymm0, %%ymm10      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231pd      %%ymm5 , %%ymm1, %%ymm11      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231pd      %%ymm6 , %%ymm0, %%ymm12      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231pd      %%ymm6 , %%ymm1, %%ymm13      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231pd      %%ymm7 , %%ymm0, %%ymm14      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231pd      %%ymm7 , %%ymm1, %%ymm15      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+
+        "addq		$8 , %0	  	 	        \n\t"
+	"subq	        $4 , %1			        \n\t"		
 	"jnz		.L01LOOP%=		        \n\t"
 
 #if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )

From 07c66b196093b3bc124f674a341ce304939eccde Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Tue, 12 Aug 2014 08:35:42 +0200
Subject: [PATCH 17/27] modified algorithm for better numerical stability

---
 kernel/x86_64/zgemv_n.c | 114 ++++++++++------------------------------
 1 file changed, 27 insertions(+), 87 deletions(-)

diff --git a/kernel/x86_64/zgemv_n.c b/kernel/x86_64/zgemv_n.c
index 141cb35df..75e40eccb 100644
--- a/kernel/x86_64/zgemv_n.c
+++ b/kernel/x86_64/zgemv_n.c
@@ -48,8 +48,7 @@ static void zgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 
 	for ( i=0; i< 2*n; i+=2 )
 	{
-#if !defined(CONJ) 
-#if !defined(XCONJ)
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
 		y[i]   += a0[i]*x[0] - a0[i+1] * x[1];
 		y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
 		y[i]   += a1[i]*x[2] - a1[i+1] * x[3];
@@ -67,29 +66,6 @@ static void zgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 		y[i+1] += a2[i]*x[5] - a2[i+1] * x[4];
 		y[i]   += a3[i]*x[6] + a3[i+1] * x[7];
 		y[i+1] += a3[i]*x[7] - a3[i+1] * x[6];
-#endif
-#else
-#if !defined(XCONJ)
-		y[i]   += a0[i]*x[0] + a0[i+1] * x[1];
-		y[i+1] -= a0[i]*x[1] - a0[i+1] * x[0];
-		y[i]   += a1[i]*x[2] + a1[i+1] * x[3];
-		y[i+1] -= a1[i]*x[3] - a1[i+1] * x[2];
-		y[i]   += a2[i]*x[4] + a2[i+1] * x[5];
-		y[i+1] -= a2[i]*x[5] - a2[i+1] * x[4];
-		y[i]   += a3[i]*x[6] + a3[i+1] * x[7];
-		y[i+1] -= a3[i]*x[7] - a3[i+1] * x[6];
-
-#else
-		y[i]   += a0[i]*x[0] - a0[i+1] * x[1];
-		y[i+1] -= a0[i]*x[1] + a0[i+1] * x[0];
-		y[i]   += a1[i]*x[2] - a1[i+1] * x[3];
-		y[i+1] -= a1[i]*x[3] + a1[i+1] * x[2];
-		y[i]   += a2[i]*x[4] - a2[i+1] * x[5];
-		y[i+1] -= a2[i]*x[5] + a2[i+1] * x[4];
-		y[i]   += a3[i]*x[6] - a3[i+1] * x[7];
-		y[i+1] -= a3[i]*x[7] + a3[i+1] * x[6];
-
-#endif
 #endif
 	}
 }
@@ -104,23 +80,12 @@ static void zgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
 
 	for ( i=0; i< 2*n; i+=2 )
 	{
-#if !defined(CONJ) 
-#if !defined(XCONJ)
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
 		y[i]   += a0[i]*x[0] - a0[i+1] * x[1];
 		y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
 #else 
 		y[i]   += a0[i]*x[0] + a0[i+1] * x[1];
 		y[i+1] += a0[i]*x[1] - a0[i+1] * x[0];
-#endif
-#else
-#if !defined(XCONJ)
-		y[i]   += a0[i]*x[0] + a0[i+1] * x[1];
-		y[i+1] -= a0[i]*x[1] - a0[i+1] * x[0];
-
-#else
-		y[i]   += a0[i]*x[0] - a0[i+1] * x[1];
-		y[i+1] -= a0[i]*x[1] + a0[i+1] * x[0];
-#endif
 #endif
 
 	}
@@ -139,17 +104,24 @@ static void zero_y(BLASLONG n, FLOAT *dest)
 
 
 
-static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest)
+static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i)
 {
 	BLASLONG i;
+	FLOAT temp_r;
+	FLOAT temp_i;
 	for ( i=0; i<n; i++ )
 	{
-		*dest += *src;
-#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
-		*(dest+1) += *(src+1);
+#if !defined(XCONJ) 
+		temp_r = alpha_r * src[0] - alpha_i * src[1];
+		temp_i = alpha_r * src[1] + alpha_i * src[0];
 #else
-		*(dest+1) -= *(src+1);
+		temp_r =  alpha_r * src[0] + alpha_i * src[1];
+		temp_i = -alpha_r * src[1] + alpha_i * src[0];
 #endif
+
+		*dest += temp_r;
+		*(dest+1) += temp_i;
+
 		src+=2;
 		dest += inc_dest;
 	}
@@ -201,33 +173,18 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r,FLOAT alpha_i,
 		for( i = 0; i < n1 ; i++)
 		{
 
-#if !defined(XCONJ)
-			xbuffer[0] = alpha_r * x_ptr[0] - alpha_i * x_ptr[1];
-			xbuffer[1] = alpha_r * x_ptr[1] + alpha_i * x_ptr[0];
+			xbuffer[0] = x_ptr[0];
+			xbuffer[1] = x_ptr[1];
 			x_ptr += inc_x;	
-			xbuffer[2] = alpha_r * x_ptr[0] - alpha_i * x_ptr[1];
-			xbuffer[3] = alpha_r * x_ptr[1] + alpha_i * x_ptr[0];
+			xbuffer[2] = x_ptr[0];
+			xbuffer[3] = x_ptr[1];
 			x_ptr += inc_x;	
-			xbuffer[4] = alpha_r * x_ptr[0] - alpha_i * x_ptr[1];
-			xbuffer[5] = alpha_r * x_ptr[1] + alpha_i * x_ptr[0];
+			xbuffer[4] = x_ptr[0];
+			xbuffer[5] = x_ptr[1];
 			x_ptr += inc_x;	
-			xbuffer[6] = alpha_r * x_ptr[0] - alpha_i * x_ptr[1];
-			xbuffer[7] = alpha_r * x_ptr[1] + alpha_i * x_ptr[0];
+			xbuffer[6] = x_ptr[0];
+			xbuffer[7] = x_ptr[1];
 			x_ptr += inc_x;	
-#else
-			xbuffer[0] = alpha_r * x_ptr[0] + alpha_i * x_ptr[1];
-			xbuffer[1] = alpha_r * x_ptr[1] - alpha_i * x_ptr[0];
-			x_ptr += inc_x;	
-			xbuffer[2] = alpha_r * x_ptr[0] + alpha_i * x_ptr[1];
-			xbuffer[3] = alpha_r * x_ptr[1] - alpha_i * x_ptr[0];
-			x_ptr += inc_x;	
-			xbuffer[4] = alpha_r * x_ptr[0] + alpha_i * x_ptr[1];
-			xbuffer[5] = alpha_r * x_ptr[1] - alpha_i * x_ptr[0];
-			x_ptr += inc_x;	
-			xbuffer[6] = alpha_r * x_ptr[0] + alpha_i * x_ptr[1];
-			xbuffer[7] = alpha_r * x_ptr[1] - alpha_i * x_ptr[0];
-			x_ptr += inc_x;	
-#endif
 
 			ap[0] = a_ptr;
 			ap[1] = a_ptr + lda;
@@ -239,24 +196,18 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r,FLOAT alpha_i,
 
 		for( i = 0; i < n2 ; i++)
 		{
-#if !defined(XCONJ)
-			xbuffer[0] = alpha_r * x_ptr[0] - alpha_i * x_ptr[1];
-			xbuffer[1] = alpha_r * x_ptr[1] + alpha_i * x_ptr[0];
+			xbuffer[0] = x_ptr[0];
+			xbuffer[1] = x_ptr[1];
 			x_ptr += inc_x;	
-#else
-			xbuffer[0] = alpha_r * x_ptr[0] + alpha_i * x_ptr[1];
-			xbuffer[1] = alpha_r * x_ptr[1] - alpha_i * x_ptr[0];
-			x_ptr += inc_x;	
-#endif
-
 			zgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
 			a_ptr += 1 * lda;
 
 		}
-		add_y(NB,ybuffer,y_ptr,inc_y);
+		add_y(NB,ybuffer,y_ptr,inc_y,alpha_r,alpha_i);
 		a     += 2 * NB;
 		y_ptr += NB * inc_y;
 	}
+
 	j=0;
 	while ( j < (m % 16))
 	{
@@ -266,24 +217,13 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r,FLOAT alpha_i,
 		FLOAT temp_i = 0.0;
 		for( i = 0; i < n; i++ )
 		{
-#if  !defined(CONJ)
-#if  !defined(XCONJ)
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
 			temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
 			temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
 #else
 			temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
 			temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
 #endif
-#else
-#if  !defined(XCONJ)
-			temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
-			temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
-#else
-			temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
-			temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
-#endif
-#endif
-
 
 			a_ptr += lda;
 			x_ptr += inc_x;

From 6093ee53639a77477fd55b7cfdea56df0e93ba4a Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Tue, 12 Aug 2014 10:02:25 +0200
Subject: [PATCH 18/27] bugfix in zgemv_n_microk_haswell-2.c

---
 kernel/x86_64/zgemv_n_microk_haswell-2.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/kernel/x86_64/zgemv_n_microk_haswell-2.c b/kernel/x86_64/zgemv_n_microk_haswell-2.c
index 833983fe0..bb40ec3ac 100644
--- a/kernel/x86_64/zgemv_n_microk_haswell-2.c
+++ b/kernel/x86_64/zgemv_n_microk_haswell-2.c
@@ -99,8 +99,8 @@ static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 #else
         "vpermilpd      $0x5 , %%ymm12, %%ymm12               \n\t"
         "vpermilpd      $0x5 , %%ymm14, %%ymm14               \n\t"
-        "vaddsubpd      %%ymm13, %%ymm12, %%ymm8              \n\t"
-        "vaddsubpd      %%ymm15, %%ymm14, %%ymm9              \n\t"
+        "vaddsubpd      %%ymm12, %%ymm13, %%ymm8              \n\t"
+        "vaddsubpd      %%ymm14, %%ymm15, %%ymm9              \n\t"
         "vpermilpd      $0x5 , %%ymm8 , %%ymm8                \n\t"
         "vpermilpd      $0x5 , %%ymm9 , %%ymm9                \n\t"
 #endif

From b06550519eb6a02d87f73a48b73fcef2fdedb9c9 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Tue, 12 Aug 2014 12:15:41 +0200
Subject: [PATCH 19/27] added optimized cgemv_t c-kernel

---
 kernel/x86_64/cgemv_t.c                | 269 +++++++++++++++++++++++++
 kernel/x86_64/zgemv_n_microk_sandy-2.c | 161 +++++++++++++++
 2 files changed, 430 insertions(+)
 create mode 100644 kernel/x86_64/cgemv_t.c
 create mode 100644 kernel/x86_64/zgemv_n_microk_sandy-2.c

diff --git a/kernel/x86_64/cgemv_t.c b/kernel/x86_64/cgemv_t.c
new file mode 100644
index 000000000..ccdf13a57
--- /dev/null
+++ b/kernel/x86_64/cgemv_t.c
@@ -0,0 +1,269 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include "common.h"
+
+/*
+#if defined(BULLDOZER)
+#include "zgemv_t_microk_bulldozer-2.c"
+#elif defined(HASWELL)
+#include "zgemv_t_microk_haswell-2.c"
+#endif
+*/
+
+#define NBMAX 2048
+
+#ifndef HAVE_KERNEL_16x4
+
+static void zgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0,*a1,*a2,*a3;
+	a0 = ap[0];
+	a1 = ap[1];
+	a2 = ap[2];
+	a3 = ap[3];
+	FLOAT temp_r0 = 0.0;
+	FLOAT temp_r1 = 0.0;
+	FLOAT temp_r2 = 0.0;
+	FLOAT temp_r3 = 0.0;
+	FLOAT temp_i0 = 0.0;
+	FLOAT temp_i1 = 0.0;
+	FLOAT temp_i2 = 0.0;
+	FLOAT temp_i3 = 0.0;
+
+
+	for ( i=0; i< 2*n; i+=2 )
+	{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+		temp_r0 += a0[i]*x[i]   - a0[i+1]*x[i+1];		
+		temp_i0 += a0[i]*x[i+1] + a0[i+1]*x[i];		
+		temp_r1 += a1[i]*x[i]   - a1[i+1]*x[i+1];		
+		temp_i1 += a1[i]*x[i+1] + a1[i+1]*x[i];		
+		temp_r2 += a2[i]*x[i]   - a2[i+1]*x[i+1];		
+		temp_i2 += a2[i]*x[i+1] + a2[i+1]*x[i];		
+		temp_r3 += a3[i]*x[i]   - a3[i+1]*x[i+1];		
+		temp_i3 += a3[i]*x[i+1] + a3[i+1]*x[i];		
+#else
+		temp_r0 += a0[i]*x[i]   + a0[i+1]*x[i+1];		
+		temp_i0 += a0[i]*x[i+1] - a0[i+1]*x[i];		
+		temp_r1 += a1[i]*x[i]   + a1[i+1]*x[i+1];		
+		temp_i1 += a1[i]*x[i+1] - a1[i+1]*x[i];		
+		temp_r2 += a2[i]*x[i]   + a2[i+1]*x[i+1];		
+		temp_i2 += a2[i]*x[i+1] - a2[i+1]*x[i];		
+		temp_r3 += a3[i]*x[i]   + a3[i+1]*x[i+1];		
+		temp_i3 += a3[i]*x[i+1] - a3[i+1]*x[i];		
+#endif
+	}
+	y[0] = temp_r0;
+	y[1] = temp_i0;
+	y[2] = temp_r1;
+	y[3] = temp_i1;
+	y[4] = temp_r2;
+	y[5] = temp_i2;
+	y[6] = temp_r3;
+	y[7] = temp_i3;
+}
+	
+#endif
+
+static void zgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0;
+	a0 = ap;
+	FLOAT temp_r = 0.0;
+	FLOAT temp_i = 0.0;
+
+	for ( i=0; i< 2*n; i+=2 )
+	{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+		temp_r += a0[i]*x[i]   - a0[i+1]*x[i+1];		
+		temp_i += a0[i]*x[i+1] + a0[i+1]*x[i];		
+#else
+		temp_r += a0[i]*x[i]   + a0[i+1]*x[i+1];		
+		temp_i += a0[i]*x[i+1] - a0[i+1]*x[i];		
+#endif
+	}
+	*y      = temp_r;
+	*(y+1)  = temp_i;
+}
+	
+static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
+{
+        BLASLONG i;
+        for ( i=0; i<n; i++ )
+        {
+                *dest     = *src;
+                *(dest+1) = *(src+1);
+                dest+=2;
+                src += inc_src;
+        }
+}
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG j;
+	FLOAT *a_ptr;
+	FLOAT *x_ptr;
+	FLOAT *y_ptr;
+	FLOAT *ap[8];
+	BLASLONG n1;
+	BLASLONG m1;
+	BLASLONG m2;
+	BLASLONG n2;
+	FLOAT ybuffer[8],*xbuffer;
+
+        inc_x *= 2;
+        inc_y *= 2;
+        lda   *= 2;
+
+	xbuffer = buffer;
+	
+	n1 = n / 4 ;
+	n2 = n % 4 ;
+	
+	m1 = m - ( m % 16 );
+	m2 = (m % NBMAX) - (m % 16) ;
+	
+
+	BLASLONG NB = NBMAX;
+
+	while ( NB == NBMAX )
+	{
+		
+		m1 -= NB;
+		if ( m1 < 0)
+		{
+			if ( m2 == 0 ) break;	
+			NB = m2;
+		}
+		
+		y_ptr = y;
+		a_ptr = a;
+		x_ptr = x;
+		copy_x(NB,x_ptr,xbuffer,inc_x);
+		for( i = 0; i < n1 ; i++)
+		{
+			ap[0] = a_ptr;
+			ap[1] = a_ptr + lda;
+			ap[2] = ap[1] + lda;
+			ap[3] = ap[2] + lda;
+			zgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
+			a_ptr += 4 * lda;
+
+#if !defined(XCONJ)
+			y_ptr[0] += alpha_r * ybuffer[0] - alpha_i * ybuffer[1];
+			y_ptr[1] += alpha_r * ybuffer[1] + alpha_i * ybuffer[0];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[2] - alpha_i * ybuffer[3];
+			y_ptr[1] += alpha_r * ybuffer[3] + alpha_i * ybuffer[2];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[4] - alpha_i * ybuffer[5];
+			y_ptr[1] += alpha_r * ybuffer[5] + alpha_i * ybuffer[4];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[6] - alpha_i * ybuffer[7];
+			y_ptr[1] += alpha_r * ybuffer[7] + alpha_i * ybuffer[6];
+			y_ptr  += inc_y;
+#else
+			y_ptr[0] += alpha_r * ybuffer[0] + alpha_i * ybuffer[1];
+			y_ptr[1] -= alpha_r * ybuffer[1] - alpha_i * ybuffer[0];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[2] + alpha_i * ybuffer[3];
+			y_ptr[1] -= alpha_r * ybuffer[3] - alpha_i * ybuffer[2];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[4] + alpha_i * ybuffer[5];
+			y_ptr[1] -= alpha_r * ybuffer[5] - alpha_i * ybuffer[4];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[6] + alpha_i * ybuffer[7];
+			y_ptr[1] -= alpha_r * ybuffer[7] - alpha_i * ybuffer[6];
+			y_ptr  += inc_y;
+#endif
+		}
+
+		for( i = 0; i < n2 ; i++)
+		{
+			zgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
+			a_ptr += 1 * lda;
+
+#if !defined(XCONJ)
+			y_ptr[0] += alpha_r * ybuffer[0] - alpha_i * ybuffer[1];
+			y_ptr[1] += alpha_r * ybuffer[1] + alpha_i * ybuffer[0];
+			y_ptr  += inc_y;
+#else
+			y_ptr[0] += alpha_r * ybuffer[0] + alpha_i * ybuffer[1];
+			y_ptr[1] -= alpha_r * ybuffer[1] - alpha_i * ybuffer[0];
+			y_ptr  += inc_y;
+#endif
+
+		}
+		a += 2* NB;
+		x += NB * inc_x;	
+	}
+
+	BLASLONG m3 = m % 16;
+	if ( m3 == 0 ) return(0);
+
+	x_ptr = x;
+	copy_x(m3,x_ptr,xbuffer,inc_x);
+	j=0;
+	a_ptr = a;
+	y_ptr = y;
+	while ( j < n)
+	{
+		FLOAT temp_r = 0.0;
+		FLOAT temp_i = 0.0;
+		for( i = 0; i < m3*2; i+=2 )
+		{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+			temp_r += a_ptr[i] * xbuffer[i]   - a_ptr[i+1] * xbuffer[i+1];
+			temp_i += a_ptr[i] * xbuffer[i+1] + a_ptr[i+1] * xbuffer[i];
+#else
+			temp_r += a_ptr[i] * xbuffer[i]   + a_ptr[i+1] * xbuffer[i+1];
+			temp_i += a_ptr[i] * xbuffer[i+1] - a_ptr[i+1] * xbuffer[i];
+#endif
+		}
+		a_ptr += lda;
+
+#if !defined(XCONJ) 
+                y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
+                y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
+#else
+                y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
+                y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
+#endif
+
+		y_ptr += inc_y;
+		j++;
+	}
+	return(0);
+}
+
+
diff --git a/kernel/x86_64/zgemv_n_microk_sandy-2.c b/kernel/x86_64/zgemv_n_microk_sandy-2.c
new file mode 100644
index 000000000..8061ed4fa
--- /dev/null
+++ b/kernel/x86_64/zgemv_n_microk_sandy-2.c
@@ -0,0 +1,161 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+
+	"vbroadcastsd	  (%2), %%ymm0                  \n\t"  // real part x0
+	"vbroadcastsd	 8(%2), %%ymm1                  \n\t"  // imag part x0
+	"vbroadcastsd	16(%2), %%ymm2                  \n\t"  // real part x1
+	"vbroadcastsd	24(%2), %%ymm3                  \n\t"  // imag part x1
+	"vbroadcastsd	32(%2), %%ymm4                  \n\t"  // real part x2
+	"vbroadcastsd	40(%2), %%ymm5                  \n\t"  // imag part x2
+	"vbroadcastsd	48(%2), %%ymm6                  \n\t"  // real part x3
+	"vbroadcastsd	56(%2), %%ymm7                  \n\t"  // imag part x3
+
+
+	".align 16				        \n\t"
+	".L01LOOP%=:				        \n\t"
+
+	"vxorpd		%%ymm12, %%ymm12, %%ymm12	\n\t"
+	"vxorpd		%%ymm13, %%ymm13, %%ymm13	\n\t"
+	"vxorpd		%%ymm14, %%ymm14, %%ymm14	\n\t"
+	"vxorpd		%%ymm15, %%ymm15, %%ymm15	\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 , %%ymm10	\n\t"
+	"vaddpd		  %%ymm12, %%ymm10, %%ymm12	\n\t"
+	"vmulpd		  %%ymm8 , %%ymm1 , %%ymm11	\n\t"
+	"vaddpd		  %%ymm13, %%ymm11, %%ymm13	\n\t"
+	"vmulpd		  %%ymm9 , %%ymm0 , %%ymm10	\n\t"
+	"vaddpd		  %%ymm14, %%ymm10, %%ymm14	\n\t"
+	"vmulpd		  %%ymm9 , %%ymm1 , %%ymm11	\n\t"
+	"vaddpd		  %%ymm15, %%ymm11, %%ymm15	\n\t"
+
+	"vmovups	(%5,%0,8), %%ymm8	        \n\t" // 2 complex values form a0
+	"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"
+	"vmulpd		  %%ymm9 , %%ymm2 , %%ymm10	\n\t"
+	"vaddpd		  %%ymm14, %%ymm10, %%ymm14	\n\t"
+	"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"
+	"vmulpd		  %%ymm9 , %%ymm4 , %%ymm10	\n\t"
+	"vaddpd		  %%ymm14, %%ymm10, %%ymm14	\n\t"
+	"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"
+	"vaddpd		  %%ymm12, %%ymm10, %%ymm12	\n\t"
+	"vmulpd		  %%ymm8 , %%ymm7 , %%ymm11	\n\t"
+	"vaddpd		  %%ymm13, %%ymm11, %%ymm13	\n\t"
+	"vmulpd		  %%ymm9 , %%ymm6 , %%ymm10	\n\t"
+	"vaddpd		  %%ymm14, %%ymm10, %%ymm14	\n\t"
+	"vmulpd		  %%ymm9 , %%ymm7 , %%ymm11	\n\t"
+	"vaddpd		  %%ymm15, %%ymm11, %%ymm15	\n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilpd      $0x5 , %%ymm13, %%ymm13               \n\t"
+        "vpermilpd      $0x5 , %%ymm15, %%ymm15               \n\t"
+        "vaddsubpd      %%ymm13, %%ymm12, %%ymm8              \n\t"
+        "vaddsubpd      %%ymm15, %%ymm14, %%ymm9              \n\t"
+#else
+        "vpermilpd      $0x5 , %%ymm12, %%ymm12               \n\t"
+        "vpermilpd      $0x5 , %%ymm14, %%ymm14               \n\t"
+        "vaddsubpd      %%ymm12, %%ymm13, %%ymm8              \n\t"
+        "vaddsubpd      %%ymm14, %%ymm15, %%ymm9              \n\t"
+        "vpermilpd      $0x5 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilpd      $0x5 , %%ymm9 , %%ymm9                \n\t"
+#endif
+
+	"prefetcht0      192(%3,%0,8)			\n\t"
+	"vmovups	  (%3,%0,8),  %%ymm12           \n\t"
+	"vmovups	32(%3,%0,8),  %%ymm13           \n\t"
+
+#if !defined(XCONJ)
+        "vaddpd         %%ymm8, %%ymm12, %%ymm12              \n\t"
+        "vaddpd         %%ymm9, %%ymm13, %%ymm13              \n\t"
+#else
+        "vaddsubpd              %%ymm12, %%ymm8, %%ymm12              \n\t"
+        "vaddsubpd              %%ymm13, %%ymm9, %%ymm13              \n\t"
+#endif
+
+
+	"vmovups  %%ymm12,   (%3,%0,8)		        \n\t" // 2 complex values to y	
+	"vmovups  %%ymm13, 32(%3,%0,8)		        \n\t"	
+
+        "addq		$8 , %0	  	 	        \n\t"
+	"subq	        $4 , %1			        \n\t"		
+	"jnz		.L01LOOP%=		        \n\t"
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From 9528f0d9ee39d6050dc8330842e00a59fb751839 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Wed, 13 Aug 2014 12:18:03 +0200
Subject: [PATCH 20/27] bugfix in zgemv_n_microk_sandy-2.c

---
 kernel/x86_64/zgemv_n_microk_sandy-2.c | 35 ++++++++------------------
 1 file changed, 10 insertions(+), 25 deletions(-)

diff --git a/kernel/x86_64/zgemv_n_microk_sandy-2.c b/kernel/x86_64/zgemv_n_microk_sandy-2.c
index 8061ed4fa..f90e2210a 100644
--- a/kernel/x86_64/zgemv_n_microk_sandy-2.c
+++ b/kernel/x86_64/zgemv_n_microk_sandy-2.c
@@ -50,22 +50,13 @@ static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	".align 16				        \n\t"
 	".L01LOOP%=:				        \n\t"
 
-	"vxorpd		%%ymm12, %%ymm12, %%ymm12	\n\t"
-	"vxorpd		%%ymm13, %%ymm13, %%ymm13	\n\t"
-	"vxorpd		%%ymm14, %%ymm14, %%ymm14	\n\t"
-	"vxorpd		%%ymm15, %%ymm15, %%ymm15	\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 , %%ymm10	\n\t"
-	"vaddpd		  %%ymm12, %%ymm10, %%ymm12	\n\t"
-	"vmulpd		  %%ymm8 , %%ymm1 , %%ymm11	\n\t"
-	"vaddpd		  %%ymm13, %%ymm11, %%ymm13	\n\t"
-	"vmulpd		  %%ymm9 , %%ymm0 , %%ymm10	\n\t"
-	"vaddpd		  %%ymm14, %%ymm10, %%ymm14	\n\t"
-	"vmulpd		  %%ymm9 , %%ymm1 , %%ymm11	\n\t"
-	"vaddpd		  %%ymm15, %%ymm11, %%ymm15	\n\t"
+	"vmulpd		  %%ymm8 , %%ymm0 , %%ymm12	\n\t"
+	"vmulpd		  %%ymm8 , %%ymm1 , %%ymm13	\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
 	"vmovups      32(%5,%0,8), %%ymm9	        \n\t" // 2 complex values form a0
@@ -103,6 +94,10 @@ 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"
+	"vmovups	  (%3,%0,8),  %%ymm10           \n\t"
+	"vmovups	32(%3,%0,8),  %%ymm11           \n\t"
+
 #if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
         "vpermilpd      $0x5 , %%ymm13, %%ymm13               \n\t"
         "vpermilpd      $0x5 , %%ymm15, %%ymm15               \n\t"
@@ -117,18 +112,8 @@ static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
         "vpermilpd      $0x5 , %%ymm9 , %%ymm9                \n\t"
 #endif
 
-	"prefetcht0      192(%3,%0,8)			\n\t"
-	"vmovups	  (%3,%0,8),  %%ymm12           \n\t"
-	"vmovups	32(%3,%0,8),  %%ymm13           \n\t"
-
-#if !defined(XCONJ)
-        "vaddpd         %%ymm8, %%ymm12, %%ymm12              \n\t"
-        "vaddpd         %%ymm9, %%ymm13, %%ymm13              \n\t"
-#else
-        "vaddsubpd              %%ymm12, %%ymm8, %%ymm12              \n\t"
-        "vaddsubpd              %%ymm13, %%ymm9, %%ymm13              \n\t"
-#endif
-
+        "vaddpd         %%ymm8, %%ymm10, %%ymm12              \n\t"
+        "vaddpd         %%ymm9, %%ymm11, %%ymm13              \n\t"
 
 	"vmovups  %%ymm12,   (%3,%0,8)		        \n\t" // 2 complex values to y	
 	"vmovups  %%ymm13, 32(%3,%0,8)		        \n\t"	

From 11e34ddd1b45832606d5ef000d07519410f30676 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Wed, 13 Aug 2014 12:54:18 +0200
Subject: [PATCH 21/27] bugfix for zgemv_n_microk_haswell-2.c

---
 kernel/x86_64/KERNEL.HASWELL             |  4 +---
 kernel/x86_64/zgemv_n.c                  | 11 ++++++++-
 kernel/x86_64/zgemv_n_microk_haswell-2.c | 30 +++++++-----------------
 3 files changed, 20 insertions(+), 25 deletions(-)

diff --git a/kernel/x86_64/KERNEL.HASWELL b/kernel/x86_64/KERNEL.HASWELL
index 9a48289c5..2d54920cc 100644
--- a/kernel/x86_64/KERNEL.HASWELL
+++ b/kernel/x86_64/KERNEL.HASWELL
@@ -3,10 +3,8 @@ SGEMVTKERNEL = sgemv_t.c
 DGEMVNKERNEL = dgemv_n.c
 DGEMVTKERNEL = dgemv_t.c
 
-ifndef OS_WINDOWS 
 ZGEMVNKERNEL = zgemv_n.c
-endif
-ZGEMVTKERNEL = zgemv_t.c
+#ZGEMVTKERNEL = zgemv_t.c
 
 SGEMMKERNEL    =  sgemm_kernel_16x4_haswell.S
 SGEMMINCOPY    =  ../generic/gemm_ncopy_16.c
diff --git a/kernel/x86_64/zgemv_n.c b/kernel/x86_64/zgemv_n.c
index 75e40eccb..7b8907044 100644
--- a/kernel/x86_64/zgemv_n.c
+++ b/kernel/x86_64/zgemv_n.c
@@ -25,7 +25,8 @@ OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
 USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *****************************************************************************/
 
-
+#include <stdlib.h>
+#include <stdio.h>
 #include "common.h"
 
 #if defined(HASWELL)
@@ -141,6 +142,14 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r,FLOAT alpha_i,
 	BLASLONG n2;
 	FLOAT xbuffer[8],*ybuffer;
 
+
+#if 0
+printf("%s %d %d %.16f %.16f %d %d %d\n","zgemv_n",m,n,alpha_r,alpha_i,lda,inc_x,inc_y);
+#endif
+
+	if ( m < 1 ) return(0);
+	if ( n < 1 ) return(0);
+
 	ybuffer = buffer;
 	
 	inc_x *= 2;
diff --git a/kernel/x86_64/zgemv_n_microk_haswell-2.c b/kernel/x86_64/zgemv_n_microk_haswell-2.c
index bb40ec3ac..e1c5838f7 100644
--- a/kernel/x86_64/zgemv_n_microk_haswell-2.c
+++ b/kernel/x86_64/zgemv_n_microk_haswell-2.c
@@ -53,19 +53,14 @@ static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	"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
 
-	"vxorpd		%%ymm12, %%ymm12, %%ymm12	\n\t"
-	"vxorpd		%%ymm13, %%ymm13, %%ymm13	\n\t"
-	"vxorpd		%%ymm14, %%ymm14, %%ymm14	\n\t"
-	"vxorpd		%%ymm15, %%ymm15, %%ymm15	\n\t"
-
 	"prefetcht0      192(%5,%0,8)			\n\t"
 	"vmovups	(%5,%0,8), %%ymm10              \n\t" // 2 complex values form a1
 	"vmovups      32(%5,%0,8), %%ymm11              \n\t" // 2 complex values form a1
 
-	"vfmadd231pd      %%ymm8 , %%ymm0, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
-	"vfmadd231pd      %%ymm8 , %%ymm1, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
-	"vfmadd231pd      %%ymm9 , %%ymm0, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
-	"vfmadd231pd      %%ymm9 , %%ymm1, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+	"vmulpd      %%ymm8 , %%ymm0, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vmulpd      %%ymm8 , %%ymm1, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vmulpd      %%ymm9 , %%ymm0, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vmulpd      %%ymm9 , %%ymm1, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
 
 	"prefetcht0      192(%6,%0,8)			\n\t"
 	"vmovups	(%6,%0,8), %%ymm8	        \n\t" // 2 complex values form a2
@@ -90,6 +85,9 @@ static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	"vfmadd231pd      %%ymm11, %%ymm6, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
 	"vfmadd231pd      %%ymm11, %%ymm7, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
 
+	"prefetcht0      192(%3,%0,8)			\n\t"
+	"vmovups	  (%3,%0,8),  %%ymm10           \n\t"
+	"vmovups	32(%3,%0,8),  %%ymm11           \n\t"
 
 #if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
         "vpermilpd      $0x5 , %%ymm13, %%ymm13               \n\t"
@@ -105,18 +103,8 @@ static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
         "vpermilpd      $0x5 , %%ymm9 , %%ymm9                \n\t"
 #endif
 
-	"prefetcht0      192(%3,%0,8)			\n\t"
-	"vmovups	  (%3,%0,8),  %%ymm12           \n\t"
-	"vmovups	32(%3,%0,8),  %%ymm13           \n\t"
-
-#if !defined(XCONJ)
-        "vaddpd         %%ymm8, %%ymm12, %%ymm12              \n\t"
-        "vaddpd         %%ymm9, %%ymm13, %%ymm13              \n\t"
-#else
-        "vaddsubpd              %%ymm12, %%ymm8, %%ymm12              \n\t"
-        "vaddsubpd              %%ymm13, %%ymm9, %%ymm13              \n\t"
-#endif
-
+        "vaddpd         %%ymm8, %%ymm10, %%ymm12              \n\t"
+        "vaddpd         %%ymm9, %%ymm11, %%ymm13              \n\t"
 
 	"vmovups  %%ymm12,   (%3,%0,8)		        \n\t" // 2 complex values to y	
 	"vmovups  %%ymm13, 32(%3,%0,8)		        \n\t"	

From 8c582d362d0f8a53c222dc4c9cbb7919cdb32116 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Wed, 13 Aug 2014 13:42:22 +0200
Subject: [PATCH 22/27] optimized zgemv_t_microk_haswell-2.c

---
 kernel/x86_64/KERNEL.HASWELL             |  2 +-
 kernel/x86_64/zgemv_t.c                  |  3 +++
 kernel/x86_64/zgemv_t_microk_haswell-2.c | 28 +++++++++++-------------
 3 files changed, 17 insertions(+), 16 deletions(-)

diff --git a/kernel/x86_64/KERNEL.HASWELL b/kernel/x86_64/KERNEL.HASWELL
index 2d54920cc..6d0792f16 100644
--- a/kernel/x86_64/KERNEL.HASWELL
+++ b/kernel/x86_64/KERNEL.HASWELL
@@ -4,7 +4,7 @@ DGEMVNKERNEL = dgemv_n.c
 DGEMVTKERNEL = dgemv_t.c
 
 ZGEMVNKERNEL = zgemv_n.c
-#ZGEMVTKERNEL = zgemv_t.c
+ZGEMVTKERNEL = zgemv_t.c
 
 SGEMMKERNEL    =  sgemm_kernel_16x4_haswell.S
 SGEMMINCOPY    =  ../generic/gemm_ncopy_16.c
diff --git a/kernel/x86_64/zgemv_t.c b/kernel/x86_64/zgemv_t.c
index df75afeff..bb3f90420 100644
--- a/kernel/x86_64/zgemv_t.c
+++ b/kernel/x86_64/zgemv_t.c
@@ -141,6 +141,9 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i,
 	BLASLONG n2;
 	FLOAT ybuffer[8],*xbuffer;
 
+        if ( m < 1 ) return(0);
+        if ( n < 1 ) return(0);
+
         inc_x *= 2;
         inc_y *= 2;
         lda   *= 2;
diff --git a/kernel/x86_64/zgemv_t_microk_haswell-2.c b/kernel/x86_64/zgemv_t_microk_haswell-2.c
index 99a620e44..8325db5cf 100644
--- a/kernel/x86_64/zgemv_t_microk_haswell-2.c
+++ b/kernel/x86_64/zgemv_t_microk_haswell-2.c
@@ -49,23 +49,22 @@ static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	".align 16				        \n\t"
 	".L01LOOP%=:				        \n\t"
 
-        "prefetcht0      384(%2,%0,8)                   \n\t"
+        "prefetcht0      192(%2,%0,8)                   \n\t"
 	"vmovddup	   (%2,%0,8), %%xmm0            \n\t"  // real value from x0
+        "prefetcht0      192(%4,%0,8)                   \n\t"
+	"vmovups	(%5,%0,8), %%ymm5               \n\t" // 2 complex values from a1
 	"vmovddup	  8(%2,%0,8), %%xmm1            \n\t"  // imag value from x0
+	"vmovups	(%4,%0,8), %%ymm4	        \n\t" // 2 complex values from a0
+        "prefetcht0      192(%5,%0,8)                   \n\t"
 	"vmovddup	 16(%2,%0,8), %%xmm2            \n\t"  // real value from x1
+        "prefetcht0      192(%6,%0,8)                   \n\t"
+	"vmovups	(%6,%0,8), %%ymm6	        \n\t" // 2 complex values from a2
 	"vmovddup	 24(%2,%0,8), %%xmm3            \n\t"  // imag value from x1
+        "prefetcht0      192(%7,%0,8)                   \n\t"
+	"vmovups	(%7,%0,8), %%ymm7               \n\t" // 2 complex values from a3
 	"vinsertf128	 $1, %%xmm2, %%ymm0 , %%ymm0	\n\t"  // real values from x0 and x1
 	"vinsertf128	 $1, %%xmm3, %%ymm1 , %%ymm1	\n\t"  // imag values from x0 and x1
 
-        "prefetcht0      384(%4,%0,8)                   \n\t"
-	"vmovups	(%4,%0,8), %%ymm4	        \n\t" // 2 complex values from a0
-        "prefetcht0      384(%5,%0,8)                   \n\t"
-	"vmovups	(%5,%0,8), %%ymm5               \n\t" // 2 complex values from a1
-        "prefetcht0      384(%6,%0,8)                   \n\t"
-	"vmovups	(%6,%0,8), %%ymm6	        \n\t" // 2 complex values from a2
-        "prefetcht0      384(%7,%0,8)                   \n\t"
-	"vmovups	(%7,%0,8), %%ymm7               \n\t" // 2 complex values from a3
-
 	"vfmadd231pd      %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
 	"vfmadd231pd      %%ymm4 , %%ymm1, %%ymm9       \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
 	"vfmadd231pd      %%ymm5 , %%ymm0, %%ymm10      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
@@ -75,17 +74,16 @@ static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	"vfmadd231pd      %%ymm7 , %%ymm0, %%ymm14      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
 	"vfmadd231pd      %%ymm7 , %%ymm1, %%ymm15      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
 
+	"vmovups       32(%4,%0,8), %%ymm4	        \n\t" // 2 complex values from a0
+	"vmovups       32(%5,%0,8), %%ymm5              \n\t" // 2 complex values from a1
 	"vmovddup	 32(%2,%0,8), %%xmm0            \n\t"  // real value from x0
 	"vmovddup	 40(%2,%0,8), %%xmm1            \n\t"  // imag value from x0
 	"vmovddup	 48(%2,%0,8), %%xmm2            \n\t"  // real value from x1
 	"vmovddup	 56(%2,%0,8), %%xmm3            \n\t"  // imag value from x1
-	"vinsertf128	 $1, %%xmm2, %%ymm0 , %%ymm0	\n\t"  // real values from x0 and x1
-	"vinsertf128	 $1, %%xmm3, %%ymm1 , %%ymm1	\n\t"  // imag values from x0 and x1
-
-	"vmovups       32(%4,%0,8), %%ymm4	        \n\t" // 2 complex values from a0
-	"vmovups       32(%5,%0,8), %%ymm5              \n\t" // 2 complex values from a1
 	"vmovups       32(%6,%0,8), %%ymm6	        \n\t" // 2 complex values from a2
 	"vmovups       32(%7,%0,8), %%ymm7               \n\t" // 2 complex values from a3
+	"vinsertf128	 $1, %%xmm2, %%ymm0 , %%ymm0	\n\t"  // real values from x0 and x1
+	"vinsertf128	 $1, %%xmm3, %%ymm1 , %%ymm1	\n\t"  // imag values from x0 and x1
 
 	"vfmadd231pd      %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
 	"vfmadd231pd      %%ymm4 , %%ymm1, %%ymm9       \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 

From 2470129132df121aa922e4abe955c64a5d1385cb Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Wed, 13 Aug 2014 13:54:19 +0200
Subject: [PATCH 23/27] added fast return, if m or n < 1

---
 kernel/x86_64/dgemv_n.c | 3 +++
 kernel/x86_64/dgemv_t.c | 3 +++
 kernel/x86_64/sgemv_n.c | 3 +++
 kernel/x86_64/sgemv_t.c | 3 +++
 kernel/x86_64/zgemv_t.c | 2 +-
 5 files changed, 13 insertions(+), 1 deletion(-)

diff --git a/kernel/x86_64/dgemv_n.c b/kernel/x86_64/dgemv_n.c
index 5192ba193..5d826dc63 100644
--- a/kernel/x86_64/dgemv_n.c
+++ b/kernel/x86_64/dgemv_n.c
@@ -125,6 +125,9 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
 	BLASLONG n2;
 	FLOAT xbuffer[4],*ybuffer;
 
+        if ( m < 1 ) return(0);
+        if ( n < 1 ) return(0);
+
 	ybuffer = buffer;
 	
 	n1 = n / 4 ;
diff --git a/kernel/x86_64/dgemv_t.c b/kernel/x86_64/dgemv_t.c
index 76aacd349..0fa8378fe 100644
--- a/kernel/x86_64/dgemv_t.c
+++ b/kernel/x86_64/dgemv_t.c
@@ -104,6 +104,9 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
 	BLASLONG n2;
 	FLOAT ybuffer[4],*xbuffer;
 
+        if ( m < 1 ) return(0);
+        if ( n < 1 ) return(0);
+
 	xbuffer = buffer;
 	
 	n1 = n / 4 ;
diff --git a/kernel/x86_64/sgemv_n.c b/kernel/x86_64/sgemv_n.c
index f2de1b76a..faa8e1f8c 100644
--- a/kernel/x86_64/sgemv_n.c
+++ b/kernel/x86_64/sgemv_n.c
@@ -131,6 +131,9 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
 	BLASLONG n2;
 	FLOAT xbuffer[4],*ybuffer;
 
+        if ( m < 1 ) return(0);
+        if ( n < 1 ) return(0);
+
 	ybuffer = buffer;
 	
 	n1 = n / 4 ;
diff --git a/kernel/x86_64/sgemv_t.c b/kernel/x86_64/sgemv_t.c
index adfaa9925..532afee5d 100644
--- a/kernel/x86_64/sgemv_t.c
+++ b/kernel/x86_64/sgemv_t.c
@@ -110,6 +110,9 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
 	BLASLONG n2;
 	FLOAT ybuffer[4],*xbuffer;
 
+        if ( m < 1 ) return(0);
+        if ( n < 1 ) return(0);
+
 	xbuffer = buffer;
 	
 	n1 = n / 4 ;
diff --git a/kernel/x86_64/zgemv_t.c b/kernel/x86_64/zgemv_t.c
index bb3f90420..9f5444a72 100644
--- a/kernel/x86_64/zgemv_t.c
+++ b/kernel/x86_64/zgemv_t.c
@@ -29,7 +29,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "common.h"
 
 
-#if defined(BULLDOZER)
+#if defined(BULLDOZER) || defined(PILEDRIVER)
 #include "zgemv_t_microk_bulldozer-2.c"
 #elif defined(HASWELL)
 #include "zgemv_t_microk_haswell-2.c"

From dc0593731365c17a694d297afd5ba8bf1bfaf0cd Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Wed, 13 Aug 2014 14:54:50 +0200
Subject: [PATCH 24/27] added additional test values

---
 lapack-netlib/TESTING/dstest.in | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/lapack-netlib/TESTING/dstest.in b/lapack-netlib/TESTING/dstest.in
index 6ec68d13a..4a31076a6 100644
--- a/lapack-netlib/TESTING/dstest.in
+++ b/lapack-netlib/TESTING/dstest.in
@@ -1,8 +1,8 @@
 Data file for testing DSGESV/DSPOSV LAPACK routines
 12                                      Number of values of M
 0 1 2 13 17 45 78 91 101 119 120 132    values of M (row dimension)
-4                                       Number of values of NRHS
-1 2 14 16                               Values of NRHS (number of right hand sides)
+6                                       Number of values of NRHS
+1 2 14 15 16 13                         Values of NRHS (number of right hand sides)
 30.0                                    Threshold value of test ratio
 T                                       Put T to test the driver routine
 T                                       Put T to test the error exits

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 25/27] 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"
 

From 4568d32b6bb1ad27882268b8866ef35def75605e Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Thu, 14 Aug 2014 14:10:29 +0200
Subject: [PATCH 26/27] added optimized cgemv_t kernel for haswell

---
 kernel/x86_64/KERNEL.HASWELL             |   2 +
 kernel/x86_64/cgemv_t.c                  |  18 +--
 kernel/x86_64/cgemv_t_microk_haswell-2.c | 171 +++++++++++++++++++++++
 3 files changed, 180 insertions(+), 11 deletions(-)
 create mode 100644 kernel/x86_64/cgemv_t_microk_haswell-2.c

diff --git a/kernel/x86_64/KERNEL.HASWELL b/kernel/x86_64/KERNEL.HASWELL
index 6d0792f16..e07448abb 100644
--- a/kernel/x86_64/KERNEL.HASWELL
+++ b/kernel/x86_64/KERNEL.HASWELL
@@ -6,6 +6,8 @@ DGEMVTKERNEL = dgemv_t.c
 ZGEMVNKERNEL = zgemv_n.c
 ZGEMVTKERNEL = zgemv_t.c
 
+CGEMVTKERNEL = cgemv_t.c
+
 SGEMMKERNEL    =  sgemm_kernel_16x4_haswell.S
 SGEMMINCOPY    =  ../generic/gemm_ncopy_16.c
 SGEMMITCOPY    =  ../generic/gemm_tcopy_16.c
diff --git a/kernel/x86_64/cgemv_t.c b/kernel/x86_64/cgemv_t.c
index ccdf13a57..e40fd349e 100644
--- a/kernel/x86_64/cgemv_t.c
+++ b/kernel/x86_64/cgemv_t.c
@@ -28,19 +28,15 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include "common.h"
 
-/*
-#if defined(BULLDOZER)
-#include "zgemv_t_microk_bulldozer-2.c"
-#elif defined(HASWELL)
-#include "zgemv_t_microk_haswell-2.c"
+#if defined(HASWELL)
+#include "cgemv_t_microk_haswell-2.c"
 #endif
-*/
 
 #define NBMAX 2048
 
 #ifndef HAVE_KERNEL_16x4
 
-static void zgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+static void cgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 {
 	BLASLONG i;
 	FLOAT *a0,*a1,*a2,*a3;
@@ -92,7 +88,7 @@ static void zgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	
 #endif
 
-static void zgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
+static void cgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
 {
 	BLASLONG i;
 	FLOAT *a0;
@@ -113,7 +109,7 @@ static void zgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
 	*y      = temp_r;
 	*(y+1)  = temp_i;
 }
-	
+
 static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
 {
         BLASLONG i;
@@ -176,7 +172,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i,
 			ap[1] = a_ptr + lda;
 			ap[2] = ap[1] + lda;
 			ap[3] = ap[2] + lda;
-			zgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
+			cgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
 			a_ptr += 4 * lda;
 
 #if !defined(XCONJ)
@@ -210,7 +206,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i,
 
 		for( i = 0; i < n2 ; i++)
 		{
-			zgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
+			cgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
 			a_ptr += 1 * lda;
 
 #if !defined(XCONJ)
diff --git a/kernel/x86_64/cgemv_t_microk_haswell-2.c b/kernel/x86_64/cgemv_t_microk_haswell-2.c
new file mode 100644
index 000000000..0d79714af
--- /dev/null
+++ b/kernel/x86_64/cgemv_t_microk_haswell-2.c
@@ -0,0 +1,171 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary froms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary from must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void cgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void cgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+
+	"vxorps		%%ymm8 , %%ymm8 , %%ymm8 	\n\t" // temp
+	"vxorps		%%ymm9 , %%ymm9 , %%ymm9 	\n\t" // temp
+	"vxorps		%%ymm10, %%ymm10, %%ymm10	\n\t" // temp
+	"vxorps		%%ymm11, %%ymm11, %%ymm11	\n\t" // temp
+	"vxorps		%%ymm12, %%ymm12, %%ymm12	\n\t" // temp
+	"vxorps		%%ymm13, %%ymm13, %%ymm13	\n\t"
+	"vxorps		%%ymm14, %%ymm14, %%ymm14	\n\t"
+	"vxorps		%%ymm15, %%ymm15, %%ymm15	\n\t"
+
+	".align 16				        \n\t"
+	".L01LOOP%=:				        \n\t"
+        "prefetcht0      192(%4,%0,4)                   \n\t"
+	"vmovups	(%4,%0,4), %%ymm4	        \n\t" // 4 complex values from a0
+        "prefetcht0      192(%5,%0,4)                   \n\t"
+	"vmovups	(%5,%0,4), %%ymm5               \n\t" // 4 complex values from a1
+
+        "prefetcht0      192(%2,%0,4)                   \n\t"
+	"vmovups	    (%2,%0,4)  , %%ymm6		\n\t" // 4 complex values from x
+	"vpermilps        $0xb1, %%ymm6, %%ymm7		\n\t" // exchange real and imap parts
+	"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0         \n\t" // only the real parts
+	"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1         \n\t" // only the imag parts
+	
+        "prefetcht0      192(%6,%0,4)                   \n\t"
+	"vmovups	(%6,%0,4), %%ymm6	        \n\t" // 4 complex values from a2
+        "prefetcht0      192(%7,%0,4)                   \n\t"
+	"vmovups	(%7,%0,4), %%ymm7               \n\t" // 4 complex values from a3
+
+	"vfmadd231ps      %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231ps      %%ymm4 , %%ymm1, %%ymm9       \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231ps      %%ymm5 , %%ymm0, %%ymm10      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231ps      %%ymm5 , %%ymm1, %%ymm11      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231ps      %%ymm6 , %%ymm0, %%ymm12      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231ps      %%ymm6 , %%ymm1, %%ymm13      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231ps      %%ymm7 , %%ymm0, %%ymm14      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231ps      %%ymm7 , %%ymm1, %%ymm15      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+
+	"vmovups       32(%4,%0,4), %%ymm4	        \n\t" // 2 complex values from a0
+	"vmovups       32(%5,%0,4), %%ymm5              \n\t" // 2 complex values from a1
+
+	"vmovups	  32(%2,%0,4)  , %%ymm6		\n\t" // 4 complex values from x
+	"vpermilps        $0xb1, %%ymm6, %%ymm7		\n\t" // exchange real and imap parts
+	"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0         \n\t" // only the real parts
+	"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1         \n\t" // only the imag parts
+
+	"vmovups       32(%6,%0,4), %%ymm6	        \n\t" // 2 complex values from a2
+	"vmovups       32(%7,%0,4), %%ymm7              \n\t" // 2 complex values from a3
+
+	"vfmadd231ps      %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231ps      %%ymm4 , %%ymm1, %%ymm9       \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231ps      %%ymm5 , %%ymm0, %%ymm10      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231ps      %%ymm5 , %%ymm1, %%ymm11      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231ps      %%ymm6 , %%ymm0, %%ymm12      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231ps      %%ymm6 , %%ymm1, %%ymm13      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231ps      %%ymm7 , %%ymm0, %%ymm14      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231ps      %%ymm7 , %%ymm1, %%ymm15      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+
+        "addq		$16 , %0	  	 	        \n\t"
+	"subq	        $8  , %1			        \n\t"		
+	"jnz		.L01LOOP%=		        \n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilps      $0xb1 , %%ymm9 , %%ymm9                \n\t"
+        "vpermilps      $0xb1 , %%ymm11, %%ymm11               \n\t"
+        "vpermilps      $0xb1 , %%ymm13, %%ymm13               \n\t"
+        "vpermilps      $0xb1 , %%ymm15, %%ymm15               \n\t"
+        "vaddsubps      %%ymm9 , %%ymm8, %%ymm8               \n\t" 
+        "vaddsubps      %%ymm11, %%ymm10, %%ymm10             \n\t"
+        "vaddsubps      %%ymm13, %%ymm12, %%ymm12             \n\t"
+        "vaddsubps      %%ymm15, %%ymm14, %%ymm14             \n\t"
+#else
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilps      $0xb1 , %%ymm10, %%ymm10               \n\t"
+        "vpermilps      $0xb1 , %%ymm12, %%ymm12               \n\t"
+        "vpermilps      $0xb1 , %%ymm14, %%ymm14               \n\t"
+        "vaddsubps      %%ymm8 , %%ymm9 , %%ymm8              \n\t"
+        "vaddsubps      %%ymm10, %%ymm11, %%ymm10             \n\t"
+        "vaddsubps      %%ymm12, %%ymm13, %%ymm12             \n\t"
+        "vaddsubps      %%ymm14, %%ymm15, %%ymm14             \n\t"
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilps      $0xb1 , %%ymm10, %%ymm10               \n\t"
+        "vpermilps      $0xb1 , %%ymm12, %%ymm12               \n\t"
+        "vpermilps      $0xb1 , %%ymm14, %%ymm14               \n\t"
+#endif
+
+	"vextractf128   $1, %%ymm8 , %%xmm9		      \n\t"
+	"vextractf128   $1, %%ymm10, %%xmm11	      	      \n\t"
+	"vextractf128   $1, %%ymm12, %%xmm13		      \n\t"
+	"vextractf128   $1, %%ymm14, %%xmm15		      \n\t"
+
+	"vaddps		%%xmm8 , %%xmm9 , %%xmm8       \n\t"
+	"vaddps		%%xmm10, %%xmm11, %%xmm10      \n\t"
+	"vaddps		%%xmm12, %%xmm13, %%xmm12      \n\t"
+	"vaddps		%%xmm14, %%xmm15, %%xmm14      \n\t"
+
+	"vshufpd        $0x1, %%xmm8 , %%xmm8 , %%xmm9   \n\t"
+	"vshufpd        $0x1, %%xmm10, %%xmm10, %%xmm11  \n\t"
+	"vshufpd        $0x1, %%xmm12, %%xmm12, %%xmm13  \n\t"
+	"vshufpd        $0x1, %%xmm14, %%xmm14, %%xmm15  \n\t"
+
+	"vaddps		%%xmm8 , %%xmm9 , %%xmm8       \n\t"
+	"vaddps		%%xmm10, %%xmm11, %%xmm10      \n\t"
+	"vaddps		%%xmm12, %%xmm13, %%xmm12      \n\t"
+	"vaddps		%%xmm14, %%xmm15, %%xmm14      \n\t"
+
+	"vmovsd	%%xmm8 ,   (%3)			\n\t"
+	"vmovsd	%%xmm10,  8(%3)			\n\t"
+	"vmovsd	%%xmm12, 16(%3)			\n\t"
+	"vmovsd	%%xmm14, 24(%3)			\n\t"
+
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From 11eab4c0199f85ece37453f351ffee6450bf8c7c Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Thu, 14 Aug 2014 19:00:30 +0200
Subject: [PATCH 27/27] added optimized cgemv_n for haswell

---
 kernel/x86_64/KERNEL.HASWELL             |   2 +
 kernel/x86_64/cgemv_n.c                  | 255 +++++++++++++++++++++++
 kernel/x86_64/cgemv_n_microk_haswell-2.c | 137 ++++++++++++
 3 files changed, 394 insertions(+)
 create mode 100644 kernel/x86_64/cgemv_n.c
 create mode 100644 kernel/x86_64/cgemv_n_microk_haswell-2.c

diff --git a/kernel/x86_64/KERNEL.HASWELL b/kernel/x86_64/KERNEL.HASWELL
index e07448abb..d0ac9c72f 100644
--- a/kernel/x86_64/KERNEL.HASWELL
+++ b/kernel/x86_64/KERNEL.HASWELL
@@ -1,11 +1,13 @@
 SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t.c
+
 DGEMVNKERNEL = dgemv_n.c
 DGEMVTKERNEL = dgemv_t.c
 
 ZGEMVNKERNEL = zgemv_n.c
 ZGEMVTKERNEL = zgemv_t.c
 
+CGEMVNKERNEL = cgemv_n.c
 CGEMVTKERNEL = cgemv_t.c
 
 SGEMMKERNEL    =  sgemm_kernel_16x4_haswell.S
diff --git a/kernel/x86_64/cgemv_n.c b/kernel/x86_64/cgemv_n.c
new file mode 100644
index 000000000..47ef0d447
--- /dev/null
+++ b/kernel/x86_64/cgemv_n.c
@@ -0,0 +1,255 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include <stdlib.h>
+#include <stdio.h>
+#include "common.h"
+
+#if defined(HASWELL)
+#include "cgemv_n_microk_haswell-2.c"
+#endif
+
+
+#define NBMAX 2048
+
+#ifndef HAVE_KERNEL_16x4
+
+static void cgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0,*a1,*a2,*a3;
+	a0 = ap[0];
+	a1 = ap[1];
+	a2 = ap[2];
+	a3 = ap[3];
+
+	for ( i=0; i< 2*n; i+=2 )
+	{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+		y[i]   += a0[i]*x[0] - a0[i+1] * x[1];
+		y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
+		y[i]   += a1[i]*x[2] - a1[i+1] * x[3];
+		y[i+1] += a1[i]*x[3] + a1[i+1] * x[2];
+		y[i]   += a2[i]*x[4] - a2[i+1] * x[5];
+		y[i+1] += a2[i]*x[5] + a2[i+1] * x[4];
+		y[i]   += a3[i]*x[6] - a3[i+1] * x[7];
+		y[i+1] += a3[i]*x[7] + a3[i+1] * x[6];
+#else 
+		y[i]   += a0[i]*x[0] + a0[i+1] * x[1];
+		y[i+1] += a0[i]*x[1] - a0[i+1] * x[0];
+		y[i]   += a1[i]*x[2] + a1[i+1] * x[3];
+		y[i+1] += a1[i]*x[3] - a1[i+1] * x[2];
+		y[i]   += a2[i]*x[4] + a2[i+1] * x[5];
+		y[i+1] += a2[i]*x[5] - a2[i+1] * x[4];
+		y[i]   += a3[i]*x[6] + a3[i+1] * x[7];
+		y[i+1] += a3[i]*x[7] - a3[i+1] * x[6];
+#endif
+	}
+}
+	
+#endif
+
+static void cgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0;
+	a0 = ap;
+
+	for ( i=0; i< 2*n; i+=2 )
+	{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+		y[i]   += a0[i]*x[0] - a0[i+1] * x[1];
+		y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
+#else 
+		y[i]   += a0[i]*x[0] + a0[i+1] * x[1];
+		y[i+1] += a0[i]*x[1] - a0[i+1] * x[0];
+#endif
+
+	}
+}
+	
+
+static void zero_y(BLASLONG n, FLOAT *dest)
+{
+	BLASLONG i;
+	for ( i=0; i<2*n; i++ )
+	{
+		*dest = 0.0;
+		dest++;
+	}
+}
+
+
+
+static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i)
+{
+	BLASLONG i;
+	FLOAT temp_r;
+	FLOAT temp_i;
+	for ( i=0; i<n; i++ )
+	{
+#if !defined(XCONJ) 
+		temp_r = alpha_r * src[0] - alpha_i * src[1];
+		temp_i = alpha_r * src[1] + alpha_i * src[0];
+#else
+		temp_r =  alpha_r * src[0] + alpha_i * src[1];
+		temp_i = -alpha_r * src[1] + alpha_i * src[0];
+#endif
+
+		*dest += temp_r;
+		*(dest+1) += temp_i;
+
+		src+=2;
+		dest += inc_dest;
+	}
+}
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r,FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG j;
+	FLOAT *a_ptr;
+	FLOAT *x_ptr;
+	FLOAT *y_ptr;
+	FLOAT *ap[4];
+	BLASLONG n1;
+	BLASLONG m1;
+	BLASLONG m2;
+	BLASLONG n2;
+	FLOAT xbuffer[8],*ybuffer;
+
+
+#if 0
+printf("%s %d %d %.16f %.16f %d %d %d\n","zgemv_n",m,n,alpha_r,alpha_i,lda,inc_x,inc_y);
+#endif
+
+	if ( m < 1 ) return(0);
+	if ( n < 1 ) return(0);
+
+	ybuffer = buffer;
+	
+	inc_x *= 2;
+	inc_y *= 2;
+	lda   *= 2;
+
+	n1 = n / 4 ;
+	n2 = n % 4 ;
+	
+	m1 = m - ( m % 16 );
+	m2 = (m % NBMAX) - (m % 16) ;
+	
+	y_ptr = y;
+
+	BLASLONG NB = NBMAX;
+
+	while ( NB == NBMAX )
+	{
+		
+		m1 -= NB;
+		if ( m1 < 0)
+		{
+			if ( m2 == 0 ) break;	
+			NB = m2;
+		}
+		
+		a_ptr = a;
+		x_ptr = x;
+		zero_y(NB,ybuffer);
+		for( i = 0; i < n1 ; i++)
+		{
+
+			xbuffer[0] = x_ptr[0];
+			xbuffer[1] = x_ptr[1];
+			x_ptr += inc_x;	
+			xbuffer[2] = x_ptr[0];
+			xbuffer[3] = x_ptr[1];
+			x_ptr += inc_x;	
+			xbuffer[4] = x_ptr[0];
+			xbuffer[5] = x_ptr[1];
+			x_ptr += inc_x;	
+			xbuffer[6] = x_ptr[0];
+			xbuffer[7] = x_ptr[1];
+			x_ptr += inc_x;	
+
+			ap[0] = a_ptr;
+			ap[1] = a_ptr + lda;
+			ap[2] = ap[1] + lda;
+			ap[3] = ap[2] + lda;
+			cgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
+			a_ptr += 4 * lda;
+		}
+
+		for( i = 0; i < n2 ; i++)
+		{
+			xbuffer[0] = x_ptr[0];
+			xbuffer[1] = x_ptr[1];
+			x_ptr += inc_x;	
+			cgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
+			a_ptr += 1 * lda;
+
+		}
+		add_y(NB,ybuffer,y_ptr,inc_y,alpha_r,alpha_i);
+		a     += 2 * NB;
+		y_ptr += NB * inc_y;
+	}
+
+	j=0;
+	while ( j < (m % 16))
+	{
+		a_ptr = a;
+		x_ptr = x;
+		FLOAT temp_r = 0.0;
+		FLOAT temp_i = 0.0;
+		for( i = 0; i < n; i++ )
+		{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+			temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
+			temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
+#else
+			temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
+			temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
+#endif
+
+			a_ptr += lda;
+			x_ptr += inc_x;
+		}
+
+#if !defined(XCONJ) 
+		y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
+		y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
+#else
+		y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
+		y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
+#endif
+		y_ptr += inc_y;
+		a+=2;
+		j++;
+	}
+	return(0);
+}
+
+
diff --git a/kernel/x86_64/cgemv_n_microk_haswell-2.c b/kernel/x86_64/cgemv_n_microk_haswell-2.c
new file mode 100644
index 000000000..9b1501013
--- /dev/null
+++ b/kernel/x86_64/cgemv_n_microk_haswell-2.c
@@ -0,0 +1,137 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void cgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void cgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+
+	"vbroadcastss	  (%2), %%ymm0                  \n\t"  // real part x0
+	"vbroadcastss	 4(%2), %%ymm1                  \n\t"  // imag part x0
+	"vbroadcastss	 8(%2), %%ymm2                  \n\t"  // real part x1
+	"vbroadcastss	12(%2), %%ymm3                  \n\t"  // imag part x1
+	"vbroadcastss	16(%2), %%ymm4                  \n\t"  // real part x2
+	"vbroadcastss	20(%2), %%ymm5                  \n\t"  // imag part x2
+	"vbroadcastss	24(%2), %%ymm6                  \n\t"  // real part x3
+	"vbroadcastss	28(%2), %%ymm7                  \n\t"  // imag part x3
+
+
+	".align 16				        \n\t"
+	".L01LOOP%=:				        \n\t"
+	"prefetcht0      320(%4,%0,4)			\n\t"
+	"vmovups	(%4,%0,4), %%ymm8	        \n\t" // 4 complex values form a0
+	"vmovups      32(%4,%0,4), %%ymm9	        \n\t" // 4 complex values form a0
+
+	"prefetcht0      320(%5,%0,4)			\n\t"
+	"vmovups	(%5,%0,4), %%ymm10              \n\t" // 4 complex values form a1
+	"vmovups      32(%5,%0,4), %%ymm11              \n\t" // 4 complex values form a1
+
+	"vmulps      %%ymm8 , %%ymm0, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vmulps      %%ymm8 , %%ymm1, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vmulps      %%ymm9 , %%ymm0, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vmulps      %%ymm9 , %%ymm1, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"prefetcht0      320(%6,%0,4)			\n\t"
+	"vmovups	(%6,%0,4), %%ymm8	        \n\t" // 4 complex values form a2
+	"vmovups      32(%6,%0,4), %%ymm9	        \n\t" // 4 complex values form a2
+
+	"vfmadd231ps      %%ymm10, %%ymm2, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmadd231ps      %%ymm10, %%ymm3, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vfmadd231ps      %%ymm11, %%ymm2, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vfmadd231ps      %%ymm11, %%ymm3, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"prefetcht0      320(%7,%0,4)			\n\t"
+	"vmovups	(%7,%0,4), %%ymm10              \n\t" // 4 complex values form a3
+	"vmovups      32(%7,%0,4), %%ymm11              \n\t" // 4 complex values form a3
+
+	"vfmadd231ps      %%ymm8 , %%ymm4, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmadd231ps      %%ymm8 , %%ymm5, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vfmadd231ps      %%ymm9 , %%ymm4, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vfmadd231ps      %%ymm9 , %%ymm5, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"vfmadd231ps      %%ymm10, %%ymm6, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmadd231ps      %%ymm10, %%ymm7, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vfmadd231ps      %%ymm11, %%ymm6, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vfmadd231ps      %%ymm11, %%ymm7, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"prefetcht0      320(%3,%0,4)			\n\t"
+	"vmovups	  (%3,%0,4),  %%ymm10           \n\t"
+	"vmovups	32(%3,%0,4),  %%ymm11           \n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilps      $0xb1 , %%ymm13, %%ymm13               \n\t"
+        "vpermilps      $0xb1 , %%ymm15, %%ymm15               \n\t"
+        "vaddsubps      %%ymm13, %%ymm12, %%ymm8              \n\t"
+        "vaddsubps      %%ymm15, %%ymm14, %%ymm9              \n\t"
+#else
+        "vpermilps      $0xb1 , %%ymm12, %%ymm12               \n\t"
+        "vpermilps      $0xb1 , %%ymm14, %%ymm14               \n\t"
+        "vaddsubps      %%ymm12, %%ymm13, %%ymm8              \n\t"
+        "vaddsubps      %%ymm14, %%ymm15, %%ymm9              \n\t"
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilps      $0xb1 , %%ymm9 , %%ymm9                \n\t"
+#endif
+
+        "vaddps         %%ymm8, %%ymm10, %%ymm12              \n\t"
+        "vaddps         %%ymm9, %%ymm11, %%ymm13              \n\t"
+
+	"vmovups  %%ymm12,   (%3,%0,4)		        \n\t" // 4 complex values to y	
+	"vmovups  %%ymm13, 32(%3,%0,4)		        \n\t"	
+
+        "addq		$16, %0	  	 	        \n\t"
+	"subq	        $8 , %1			        \n\t"		
+	"jnz		.L01LOOP%=		        \n\t"
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+