From 493d4fe7e5bfdc5a237df2c6c1fe489904d5bf7d Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Sat, 16 Aug 2014 11:36:48 +0200
Subject: [PATCH 01/32] added reference in C for symv_L

---
 kernel/arm/symv_L.c | 70 +++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 70 insertions(+)
 create mode 100644 kernel/arm/symv_L.c

diff --git a/kernel/arm/symv_L.c b/kernel/arm/symv_L.c
new file mode 100644
index 000000000..8f48d03f5
--- /dev/null
+++ b/kernel/arm/symv_L.c
@@ -0,0 +1,70 @@
+/***************************************************************************
+Copyright (c) 2013, 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"
+
+int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG ix,iy;
+	BLASLONG jx,jy;
+	BLASLONG j;
+	FLOAT temp1;
+	FLOAT temp2;
+
+#if 0
+	if ( m != offset )
+		printf("Symv_L: m=%d offset=%d\n",m,offset);
+#endif
+
+	jx = 0;
+	jy = 0;
+
+	for (j=0; j<offset; j++)
+	{
+		temp1 = alpha * x[jx];
+		temp2 = 0.0;
+		y[jy] += temp1 * a[j*lda+j];
+		iy = jy;
+		ix = jx;
+		for (i=j+1; i<m; i++)
+		{
+			ix += inc_x;
+			iy += inc_y;
+			y[iy] += temp1 * a[j*lda+i];
+			temp2 += a[j*lda+i] * x[ix];
+			
+		}
+		y[jy] += alpha * temp2;
+		jx    += inc_x;
+		jy    += inc_y;
+	}
+	return(0);
+}
+
+

From 101dd08173a1fa32fdb52a471a591ff7e2b53d41 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Sat, 16 Aug 2014 13:52:50 +0200
Subject: [PATCH 02/32] add reference in C for symv_U

---
 kernel/arm/symv_U.c | 71 +++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 71 insertions(+)
 create mode 100644 kernel/arm/symv_U.c

diff --git a/kernel/arm/symv_U.c b/kernel/arm/symv_U.c
new file mode 100644
index 000000000..b5a0c96e9
--- /dev/null
+++ b/kernel/arm/symv_U.c
@@ -0,0 +1,71 @@
+/***************************************************************************
+Copyright (c) 2013, 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"
+
+int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG ix,iy;
+	BLASLONG jx,jy;
+	BLASLONG j;
+	FLOAT temp1;
+	FLOAT temp2;
+
+#if 0
+	if( m != offset )
+		printf("Symv_U: m=%d offset=%d\n",m,offset);
+#endif
+
+	BLASLONG m1 = m - offset;
+
+	jx = m1 * inc_x;
+	jy = m1 * inc_y;
+
+	for (j=m1; j<m; j++)
+	{
+		temp1 = alpha * x[jx];
+		temp2 = 0.0;
+		iy = 0;
+		ix = 0;
+		for (i=0; i<j; i++)
+		{
+			y[iy] += temp1 * a[j*lda+i];
+			temp2 += a[j*lda+i] * x[ix];
+			ix += inc_x;
+			iy += inc_y;
+			
+		}
+		y[jy] += temp1 * a[j*lda+j] + alpha * temp2;
+		jx    += inc_x;
+		jy    += inc_y;
+	}
+	return(0);
+}
+
+

From 74c9465672f8df5da6ca47cc2c1dd5dedb921e2b Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Mon, 18 Aug 2014 12:18:10 +0200
Subject: [PATCH 03/32] added optimized dsymv_U kernel for bulldozer

---
 kernel/x86_64/KERNEL.BULLDOZER             |   2 +
 kernel/x86_64/dsymv_U.c                    | 208 +++++++++++++++++++++
 kernel/x86_64/dsymv_U_microk_bulldozer-2.c | 117 ++++++++++++
 3 files changed, 327 insertions(+)
 create mode 100644 kernel/x86_64/dsymv_U.c
 create mode 100644 kernel/x86_64/dsymv_U_microk_bulldozer-2.c

diff --git a/kernel/x86_64/KERNEL.BULLDOZER b/kernel/x86_64/KERNEL.BULLDOZER
index 19bf7fd32..03925cc19 100644
--- a/kernel/x86_64/KERNEL.BULLDOZER
+++ b/kernel/x86_64/KERNEL.BULLDOZER
@@ -1,3 +1,5 @@
+DSYMV_U_KERNEL = dsymv_U.c
+
 SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t.c
 
diff --git a/kernel/x86_64/dsymv_U.c b/kernel/x86_64/dsymv_U.c
new file mode 100644
index 000000000..1f22abe8d
--- /dev/null
+++ b/kernel/x86_64/dsymv_U.c
@@ -0,0 +1,208 @@
+/***************************************************************************
+Copyright (c) 2013, 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 "dsymv_U_microk_bulldozer-2.c"
+#endif
+
+
+#ifndef HAVE_KERNEL_8x2
+
+static void dsymv_kernel_8x2(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *xp, FLOAT *yp, FLOAT *temp1, FLOAT *temp2)
+{
+	FLOAT at0,at1,at2,at3;
+	FLOAT tmp2[2] = { 0.0, 0.0 };
+	FLOAT tp0;
+	FLOAT tp1;
+	BLASLONG i;
+
+	tp0 = temp1[0];
+	tp1 = temp1[1];
+	
+	for (i=0; i<n; i+=2)
+	{
+		at0     = a0[i];
+		at1     = a1[i];
+		at2     = a0[i+1];
+		at3     = a1[i+1];
+		yp[i]   += tp0 * at0 + tp1 *at1;
+		yp[i+1] += tp0 * at2 + tp1 *at3;
+		tmp2[0] += at0 * xp[i] + at2 * xp[i+1];
+		tmp2[1] += at1 * xp[i] + at3 * xp[i+1];
+
+	}
+	temp2[0] = tmp2[0];
+	temp2[1] = tmp2[1];
+
+}
+
+#endif
+
+static void dsymv_kernel_8x1(BLASLONG n, FLOAT *a0, FLOAT *xp, FLOAT *yp, FLOAT *temp1, FLOAT *temp2)
+{
+	FLOAT at0,at1,at2,at3;
+	FLOAT temp = 0.0;
+	FLOAT t1   = *temp1;
+	BLASLONG i;
+	
+	for (i=0; i<(n/4)*4; i+=4)
+	{
+		at0     = a0[i];
+		at1     = a0[i+1];
+		at2     = a0[i+2];
+		at3     = a0[i+3];
+
+		yp[i]   += t1    * at0;
+		temp    += at0   * xp[i];
+		yp[i+1] += t1    * at1;
+		temp    += at1   * xp[i+1];
+
+		yp[i+2] += t1    * at2;
+		temp    += at2   * xp[i+2];
+		yp[i+3] += t1    * at3;
+		temp    += at3   * xp[i+3];
+
+	}
+	*temp2 = temp;
+}
+
+int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG ix,iy;
+	BLASLONG jx,jy;
+	BLASLONG j;
+	BLASLONG m2;
+	FLOAT temp1;
+	FLOAT temp2;
+	FLOAT *xp, *yp;
+	FLOAT *a0,*a1;
+	FLOAT tmp1[2];
+	FLOAT tmp2[2];
+
+#if 0
+	if( m != offset )
+		printf("Symv_U: m=%d offset=%d\n",m,offset);
+#endif
+
+	BLASLONG m1 = m - offset;
+	BLASLONG mrange = m -m1;
+
+	if ( (inc_x!=1) || (inc_y!=1) || (mrange<16) )
+	{
+
+		jx = m1 * inc_x;
+		jy = m1 * inc_y;
+
+		for (j=m1; j<m; j++)
+		{
+			temp1 = alpha * x[jx];
+			temp2 = 0.0;
+			iy = 0;
+			ix = 0;
+			for (i=0; i<j; i++)
+			{
+				y[iy] += temp1 * a[j*lda+i];
+				temp2 += a[j*lda+i] * x[ix];
+				ix += inc_x;
+				iy += inc_y;
+			
+			}
+			y[jy] += temp1 * a[j*lda+j] + alpha * temp2;
+			jx    += inc_x;
+			jy    += inc_y;
+		}
+		return(0);
+	}
+
+	xp = x;
+	yp = y;
+
+	m2 = m - ( mrange % 2 );
+
+	for (j=m1; j<m2; j+=2)
+	{
+		tmp1[0] = alpha * xp[j];
+		tmp1[1] = alpha * xp[j+1];
+		tmp2[0] = 0.0;
+		tmp2[1] = 0.0;
+		a0    = &a[j*lda];
+		a1    = a0+lda;
+		FLOAT at0,at1;
+		BLASLONG j1 = (j/8)*8;		
+		if ( j1 )
+			dsymv_kernel_8x2(j1, a0, a1, xp, yp, tmp1, tmp2);
+
+		for (i=j1; i<j; i++)
+		{
+			at0     = a0[i];
+			at1     = a1[i];
+			yp[i]   += tmp1[0] * at0 + tmp1[1] *at1;
+			tmp2[0] += at0 * xp[i];
+			tmp2[1] += at1 * xp[i];
+			
+		}
+
+		at1     = a1[j];
+		yp[j]   += tmp1[1] * at1;
+		tmp2[1] += at1 * xp[j];
+
+		yp[j]   += tmp1[0] * a0[j]   + alpha * tmp2[0];
+		yp[j+1] += tmp1[1] * a1[j+1] + alpha * tmp2[1];
+	}
+
+	for ( ; j<m; j++)
+	{
+		temp1 = alpha * xp[j];
+		temp2 = 0.0;
+		a0    = &a[j*lda];
+		FLOAT at0;
+		BLASLONG j1 = (j/8)*8;		
+
+		if ( j1 )
+			dsymv_kernel_8x1(j1, a0, xp, yp, &temp1, &temp2);
+
+		for (i=j1 ; i<j; i++)
+		{
+			at0     = a0[i];
+			yp[i] += temp1 * at0;
+			temp2 += at0 * xp[i];
+			
+		}
+
+		yp[j] += temp1 * a0[j] + alpha * temp2;
+	}
+
+	return(0);
+	
+
+}
+
+
diff --git a/kernel/x86_64/dsymv_U_microk_bulldozer-2.c b/kernel/x86_64/dsymv_U_microk_bulldozer-2.c
new file mode 100644
index 000000000..3b035222b
--- /dev/null
+++ b/kernel/x86_64/dsymv_U_microk_bulldozer-2.c
@@ -0,0 +1,117 @@
+/***************************************************************************
+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_8x2 1
+static void dsymv_kernel_8x2( BLASLONG n, FLOAT *a1, FLOAT *a2, FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2) __attribute__ ((noinline));
+
+static void dsymv_kernel_8x2(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vxorpd		%%xmm0 , %%xmm0 , %%xmm0     \n\t"	// temp2[0]
+	"vxorpd		%%xmm1 , %%xmm1 , %%xmm1     \n\t"	// temp2[1]
+	"vmovddup	(%6),    %%xmm2	             \n\t"	// temp1[0]
+	"vmovddup      8(%6),    %%xmm3	             \n\t"	// temp1[1]
+
+	"xorq		%0,%0			     \n\t"
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+
+	"prefetcht0	 192(%4,%0,8)		       \n\t"
+	"vmovups	(%4,%0,8), %%xmm4	 \n\t"	// 2 * a0
+	"vmovups      16(%4,%0,8), %%xmm5	 \n\t"	// 2 * a0
+	"prefetcht0	 192(%2,%0,8)		       \n\t"
+	"vmovups	(%2,%0,8), %%xmm8	 \n\t"	// 2 * x
+	"vmovups      16(%2,%0,8), %%xmm9	 \n\t"	// 2 * x
+	"prefetcht0	 192(%3,%0,8)		       \n\t"
+	"vmovups      32(%4,%0,8), %%xmm6	 \n\t"	// 2 * a0
+	"vmovups      48(%4,%0,8), %%xmm7	 \n\t"	// 2 * a0
+	"vmovups      32(%2,%0,8), %%xmm10	 \n\t"	// 2 * x
+	"vmovups      48(%2,%0,8), %%xmm11	 \n\t"	// 2 * x
+
+	"prefetcht0	 192(%5,%0,8)		            \n\t"
+	"vfmaddpd    (%3,%0,8), %%xmm2 , %%xmm4 , %%xmm12   \n\t" // y += temp1 * a0
+	"vfmaddpd      %%xmm0 , %%xmm8 , %%xmm4 , %%xmm0    \n\t" // temp2 += a0 * x
+	"vfmaddpd  16(%3,%0,8), %%xmm2 , %%xmm5 , %%xmm13   \n\t" // y += temp1 * a0
+	"vmovups	(%5,%0,8), %%xmm4	            \n\t"	// 2 * a1
+	"vfmaddpd      %%xmm0 , %%xmm9 , %%xmm5 , %%xmm0    \n\t" // temp2 += a0 * x
+	"vfmaddpd  32(%3,%0,8), %%xmm2 , %%xmm6 , %%xmm14   \n\t" // y += temp1 * a0
+	"vmovups      16(%5,%0,8), %%xmm5	            \n\t"	// 2 * a1
+	"vfmaddpd      %%xmm0 , %%xmm10, %%xmm6 , %%xmm0    \n\t" // temp2 += a0 * x
+	"vfmaddpd  48(%3,%0,8), %%xmm2 , %%xmm7 , %%xmm15   \n\t" // y += temp1 * a0
+	"vmovups      32(%5,%0,8), %%xmm6	            \n\t"	// 2 * a1
+	"vfmaddpd      %%xmm0 , %%xmm11, %%xmm7 , %%xmm0    \n\t" // temp2 += a0 * x
+	"vmovups      48(%5,%0,8), %%xmm7	            \n\t"	// 2 * a1
+
+	"vfmaddpd %%xmm12, %%xmm3 , %%xmm4 , %%xmm12   \n\t" // y += temp1 * a1
+	"vfmaddpd %%xmm13, %%xmm3 , %%xmm5 , %%xmm13   \n\t" // y += temp1 * a1
+	"vmovups  %%xmm12,   (%3,%0,8)		       \n\t"	// 2 * y
+	"vfmaddpd %%xmm14, %%xmm3 , %%xmm6 , %%xmm14   \n\t" // y += temp1 * a1
+	"vmovups  %%xmm13, 16(%3,%0,8)		       \n\t"	// 2 * y
+	"vfmaddpd %%xmm15, %%xmm3 , %%xmm7 , %%xmm15   \n\t" // y += temp1 * a1
+	"vmovups  %%xmm14, 32(%3,%0,8)		       \n\t"	// 2 * y
+
+	"vfmaddpd %%xmm1 , %%xmm8 , %%xmm4 , %%xmm1   \n\t" // temp2 += a1 * x
+	"vfmaddpd %%xmm1 , %%xmm9 , %%xmm5 , %%xmm1   \n\t" // temp2 += a1 * x
+	"vmovups  %%xmm15, 48(%3,%0,8)		      \n\t"	// 2 * y
+	"vfmaddpd %%xmm1 , %%xmm10, %%xmm6 , %%xmm1   \n\t" // temp2 += a1 * x
+	"vfmaddpd %%xmm1 , %%xmm11, %%xmm7 , %%xmm1   \n\t" // temp2 += a1 * x
+
+        "addq		$8, %0	  	 	      \n\t"
+	"subq	        $8, %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \n\t"
+
+	"vhaddpd        %%xmm0, %%xmm0, %%xmm0  \n\t"
+	"vhaddpd        %%xmm1, %%xmm1, %%xmm1  \n\t"
+	"vmovsd         %%xmm0 , (%7)		\n\t"	// save temp2
+	"vmovsd         %%xmm1 ,8(%7)		\n\t"	// save temp2
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (a0),  // 4
+          "r" (a1),  // 5
+          "r" (temp1),  // 6
+          "r" (temp2)   // 7
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From 4f39447c05cfa6772e707a81adee2ae5a3087e78 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Mon, 18 Aug 2014 13:52:24 +0200
Subject: [PATCH 04/32] added optimized ssymv_U kernel for bulldozer

---
 kernel/x86_64/KERNEL.BULLDOZER             |   1 +
 kernel/x86_64/ssymv_U.c                    | 209 +++++++++++++++++++++
 kernel/x86_64/ssymv_U_microk_bulldozer-2.c | 119 ++++++++++++
 3 files changed, 329 insertions(+)
 create mode 100644 kernel/x86_64/ssymv_U.c
 create mode 100644 kernel/x86_64/ssymv_U_microk_bulldozer-2.c

diff --git a/kernel/x86_64/KERNEL.BULLDOZER b/kernel/x86_64/KERNEL.BULLDOZER
index 03925cc19..a078528cd 100644
--- a/kernel/x86_64/KERNEL.BULLDOZER
+++ b/kernel/x86_64/KERNEL.BULLDOZER
@@ -1,4 +1,5 @@
 DSYMV_U_KERNEL = dsymv_U.c
+SSYMV_U_KERNEL = ssymv_U.c
 
 SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t.c
diff --git a/kernel/x86_64/ssymv_U.c b/kernel/x86_64/ssymv_U.c
new file mode 100644
index 000000000..75b8e2c3e
--- /dev/null
+++ b/kernel/x86_64/ssymv_U.c
@@ -0,0 +1,209 @@
+/***************************************************************************
+Copyright (c) 2013, 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 "ssymv_U_microk_bulldozer-2.c"
+#endif
+
+
+#ifndef HAVE_KERNEL_16x2
+
+static void ssymv_kernel_16x2(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *xp, FLOAT *yp, FLOAT *temp1, FLOAT *temp2)
+{
+	FLOAT at0,at1,at2,at3;
+	FLOAT tmp2[2] = { 0.0, 0.0 };
+	FLOAT tp0;
+	FLOAT tp1;
+	BLASLONG i;
+
+	tp0 = temp1[0];
+	tp1 = temp1[1];
+	
+	for (i=0; i<n; i+=2)
+	{
+		at0     = a0[i];
+		at1     = a1[i];
+		at2     = a0[i+1];
+		at3     = a1[i+1];
+		yp[i]   += tp0 * at0 + tp1 *at1;
+		yp[i+1] += tp0 * at2 + tp1 *at3;
+		tmp2[0] += at0 * xp[i] + at2 * xp[i+1];
+		tmp2[1] += at1 * xp[i] + at3 * xp[i+1];
+
+	}
+	temp2[0] = tmp2[0];
+	temp2[1] = tmp2[1];
+
+}
+
+#endif
+
+static void ssymv_kernel_8x1(BLASLONG n, FLOAT *a0, FLOAT *xp, FLOAT *yp, FLOAT *temp1, FLOAT *temp2)
+{
+	FLOAT at0,at1,at2,at3;
+	FLOAT temp = 0.0;
+	FLOAT t1   = *temp1;
+	BLASLONG i;
+	
+	for (i=0; i<(n/4)*4; i+=4)
+	{
+		at0     = a0[i];
+		at1     = a0[i+1];
+		at2     = a0[i+2];
+		at3     = a0[i+3];
+
+		yp[i]   += t1    * at0;
+		temp    += at0   * xp[i];
+		yp[i+1] += t1    * at1;
+		temp    += at1   * xp[i+1];
+
+		yp[i+2] += t1    * at2;
+		temp    += at2   * xp[i+2];
+		yp[i+3] += t1    * at3;
+		temp    += at3   * xp[i+3];
+
+	}
+	*temp2 = temp;
+}
+
+int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG ix,iy;
+	BLASLONG jx,jy;
+	BLASLONG j;
+	BLASLONG m2;
+	FLOAT temp1;
+	FLOAT temp2;
+	FLOAT *xp, *yp;
+	FLOAT *a0,*a1;
+	FLOAT tmp1[2];
+	FLOAT tmp2[2];
+
+#if 0
+	if( m != offset )
+		printf("Symv_U: m=%d offset=%d\n",m,offset);
+#endif
+
+	BLASLONG m1 = m - offset;
+	BLASLONG mrange = m -m1;
+
+	if ( (inc_x!=1) || (inc_y!=1) || (mrange<16) )
+	{
+
+		jx = m1 * inc_x;
+		jy = m1 * inc_y;
+
+		for (j=m1; j<m; j++)
+		{
+			temp1 = alpha * x[jx];
+			temp2 = 0.0;
+			iy = 0;
+			ix = 0;
+			for (i=0; i<j; i++)
+			{
+				y[iy] += temp1 * a[j*lda+i];
+				temp2 += a[j*lda+i] * x[ix];
+				ix += inc_x;
+				iy += inc_y;
+			
+			}
+			y[jy] += temp1 * a[j*lda+j] + alpha * temp2;
+			jx    += inc_x;
+			jy    += inc_y;
+		}
+		return(0);
+	}
+
+	xp = x;
+	yp = y;
+
+	m2 = m - ( mrange % 2 );
+
+	for (j=m1; j<m2; j+=2)
+	{
+		tmp1[0] = alpha * xp[j];
+		tmp1[1] = alpha * xp[j+1];
+		tmp2[0] = 0.0;
+		tmp2[1] = 0.0;
+		a0    = &a[j*lda];
+		a1    = a0+lda;
+		FLOAT at0,at1;
+		BLASLONG j1 = (j/16)*16;		
+		if ( j1 )
+			ssymv_kernel_16x2(j1, a0, a1, xp, yp, tmp1, tmp2);
+
+		for (i=j1; i<j; i++)
+		{
+			at0     = a0[i];
+			at1     = a1[i];
+			yp[i]   += tmp1[0] * at0 + tmp1[1] *at1;
+			tmp2[0] += at0 * xp[i];
+			tmp2[1] += at1 * xp[i];
+			
+		}
+
+		at1     = a1[j];
+		yp[j]   += tmp1[1] * at1;
+		tmp2[1] += at1 * xp[j];
+
+		yp[j]   += tmp1[0] * a0[j]   + alpha * tmp2[0];
+		yp[j+1] += tmp1[1] * a1[j+1] + alpha * tmp2[1];
+	}
+
+	for ( ; j<m; j++)
+	{
+		temp1 = alpha * xp[j];
+		temp2 = 0.0;
+		a0    = &a[j*lda];
+		FLOAT at0;
+		BLASLONG j1 = (j/8)*8;		
+
+		if ( j1 )
+			ssymv_kernel_8x1(j1, a0, xp, yp, &temp1, &temp2);
+
+		for (i=j1 ; i<j; i++)
+		{
+			at0     = a0[i];
+			yp[i] += temp1 * at0;
+			temp2 += at0 * xp[i];
+			
+		}
+
+		yp[j] += temp1 * a0[j] + alpha * temp2;
+	}
+
+	return(0);
+	
+
+}
+
+
diff --git a/kernel/x86_64/ssymv_U_microk_bulldozer-2.c b/kernel/x86_64/ssymv_U_microk_bulldozer-2.c
new file mode 100644
index 000000000..b0b0bed65
--- /dev/null
+++ b/kernel/x86_64/ssymv_U_microk_bulldozer-2.c
@@ -0,0 +1,119 @@
+/***************************************************************************
+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_16x2 1
+static void ssymv_kernel_16x2( BLASLONG n, FLOAT *a1, FLOAT *a2, FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2) __attribute__ ((noinline));
+
+static void ssymv_kernel_16x2(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vxorps		%%xmm0 , %%xmm0 , %%xmm0     \n\t"	// temp2[0]
+	"vxorps		%%xmm1 , %%xmm1 , %%xmm1     \n\t"	// temp2[1]
+	"vbroadcastss	(%6),    %%xmm2	             \n\t"	// temp1[0]
+	"vbroadcastss  4(%6),    %%xmm3	             \n\t"	// temp1[1]
+
+	"xorq		%0,%0			     \n\t"
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+
+	"prefetcht0	 192(%4,%0,4)		       \n\t"
+	"vmovups	(%4,%0,4), %%xmm4	 \n\t"	// 2 * a0
+	"vmovups      16(%4,%0,4), %%xmm5	 \n\t"	// 2 * a0
+	"prefetcht0	 192(%2,%0,4)		       \n\t"
+	"vmovups	(%2,%0,4), %%xmm8	 \n\t"	// 2 * x
+	"vmovups      16(%2,%0,4), %%xmm9	 \n\t"	// 2 * x
+	"prefetcht0	 192(%3,%0,4)		       \n\t"
+	"vmovups      32(%4,%0,4), %%xmm6	 \n\t"	// 2 * a0
+	"vmovups      48(%4,%0,4), %%xmm7	 \n\t"	// 2 * a0
+	"vmovups      32(%2,%0,4), %%xmm10	 \n\t"	// 2 * x
+	"vmovups      48(%2,%0,4), %%xmm11	 \n\t"	// 2 * x
+
+	"prefetcht0	 192(%5,%0,4)		            \n\t"
+	"vfmaddps    (%3,%0,4), %%xmm2 , %%xmm4 , %%xmm12   \n\t" // y += temp1 * a0
+	"vfmaddps      %%xmm0 , %%xmm8 , %%xmm4 , %%xmm0    \n\t" // temp2 += a0 * x
+	"vfmaddps  16(%3,%0,4), %%xmm2 , %%xmm5 , %%xmm13   \n\t" // y += temp1 * a0
+	"vmovups	(%5,%0,4), %%xmm4	            \n\t"	// 2 * a1
+	"vfmaddps      %%xmm0 , %%xmm9 , %%xmm5 , %%xmm0    \n\t" // temp2 += a0 * x
+	"vfmaddps  32(%3,%0,4), %%xmm2 , %%xmm6 , %%xmm14   \n\t" // y += temp1 * a0
+	"vmovups      16(%5,%0,4), %%xmm5	            \n\t"	// 2 * a1
+	"vfmaddps      %%xmm0 , %%xmm10, %%xmm6 , %%xmm0    \n\t" // temp2 += a0 * x
+	"vfmaddps  48(%3,%0,4), %%xmm2 , %%xmm7 , %%xmm15   \n\t" // y += temp1 * a0
+	"vmovups      32(%5,%0,4), %%xmm6	            \n\t"	// 2 * a1
+	"vfmaddps      %%xmm0 , %%xmm11, %%xmm7 , %%xmm0    \n\t" // temp2 += a0 * x
+	"vmovups      48(%5,%0,4), %%xmm7	            \n\t"	// 2 * a1
+
+	"vfmaddps %%xmm12, %%xmm3 , %%xmm4 , %%xmm12   \n\t" // y += temp1 * a1
+	"vfmaddps %%xmm13, %%xmm3 , %%xmm5 , %%xmm13   \n\t" // y += temp1 * a1
+	"vmovups  %%xmm12,   (%3,%0,4)		       \n\t"	// 2 * y
+	"vfmaddps %%xmm14, %%xmm3 , %%xmm6 , %%xmm14   \n\t" // y += temp1 * a1
+	"vmovups  %%xmm13, 16(%3,%0,4)		       \n\t"	// 2 * y
+	"vfmaddps %%xmm15, %%xmm3 , %%xmm7 , %%xmm15   \n\t" // y += temp1 * a1
+	"vmovups  %%xmm14, 32(%3,%0,4)		       \n\t"	// 2 * y
+
+	"vfmaddps %%xmm1 , %%xmm8 , %%xmm4 , %%xmm1   \n\t" // temp2 += a1 * x
+	"vfmaddps %%xmm1 , %%xmm9 , %%xmm5 , %%xmm1   \n\t" // temp2 += a1 * x
+	"vmovups  %%xmm15, 48(%3,%0,4)		      \n\t"	// 2 * y
+	"vfmaddps %%xmm1 , %%xmm10, %%xmm6 , %%xmm1   \n\t" // temp2 += a1 * x
+	"vfmaddps %%xmm1 , %%xmm11, %%xmm7 , %%xmm1   \n\t" // temp2 += a1 * x
+
+        "addq		$16, %0	  	 	      \n\t"
+	"subq	        $16, %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \n\t"
+
+	"vhaddps        %%xmm0, %%xmm0, %%xmm0  \n\t"
+	"vhaddps        %%xmm1, %%xmm1, %%xmm1  \n\t"
+	"vhaddps        %%xmm0, %%xmm0, %%xmm0  \n\t"
+	"vhaddps        %%xmm1, %%xmm1, %%xmm1  \n\t"
+	"vmovss         %%xmm0 , (%7)		\n\t"	// save temp2
+	"vmovss         %%xmm1 ,4(%7)		\n\t"	// save temp2
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (a0),  // 4
+          "r" (a1),  // 5
+          "r" (temp1),  // 6
+          "r" (temp2)   // 7
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From 4ba4ab623f045c54d68bd55519b55a3ac5596b74 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Tue, 19 Aug 2014 17:09:45 +0200
Subject: [PATCH 05/32] added optimized ssymv_U kernel for nehalem

---
 kernel/x86_64/KERNEL.NEHALEM             |   3 +
 kernel/x86_64/ssymv_U.c                  | 134 +++++++++++++++++------
 kernel/x86_64/ssymv_U_microk_nehalem-2.c | 130 ++++++++++++++++++++++
 3 files changed, 232 insertions(+), 35 deletions(-)
 create mode 100644 kernel/x86_64/ssymv_U_microk_nehalem-2.c

diff --git a/kernel/x86_64/KERNEL.NEHALEM b/kernel/x86_64/KERNEL.NEHALEM
index ca9ff252d..353514449 100644
--- a/kernel/x86_64/KERNEL.NEHALEM
+++ b/kernel/x86_64/KERNEL.NEHALEM
@@ -1,3 +1,6 @@
+#DSYMV_U_KERNEL = dsymv_U.c
+SSYMV_U_KERNEL = ssymv_U.c
+
 SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t.c
 
diff --git a/kernel/x86_64/ssymv_U.c b/kernel/x86_64/ssymv_U.c
index 75b8e2c3e..61127aa3d 100644
--- a/kernel/x86_64/ssymv_U.c
+++ b/kernel/x86_64/ssymv_U.c
@@ -31,41 +31,94 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #if defined(BULLDOZER)
 #include "ssymv_U_microk_bulldozer-2.c"
+#elif defined(NEHALEM)
+#include "ssymv_U_microk_nehalem-2.c"
 #endif
 
+#ifndef HAVE_KERNEL_4x4
 
-#ifndef HAVE_KERNEL_16x2
-
-static void ssymv_kernel_16x2(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *xp, FLOAT *yp, FLOAT *temp1, FLOAT *temp2)
+static void ssymv_kernel_4x4(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *a2, FLOAT *a3, FLOAT *xp, FLOAT *yp, FLOAT *temp1, FLOAT *temp2)
 {
 	FLOAT at0,at1,at2,at3;
-	FLOAT tmp2[2] = { 0.0, 0.0 };
+	FLOAT x;
+	FLOAT tmp2[4] = { 0.0, 0.0, 0.0, 0.0 };
 	FLOAT tp0;
 	FLOAT tp1;
+	FLOAT tp2;
+	FLOAT tp3;
 	BLASLONG i;
 
 	tp0 = temp1[0];
 	tp1 = temp1[1];
+	tp2 = temp1[2];
+	tp3 = temp1[3];
 	
-	for (i=0; i<n; i+=2)
+	for (i=0; i<n; i++)
 	{
 		at0     = a0[i];
 		at1     = a1[i];
-		at2     = a0[i+1];
-		at3     = a1[i+1];
-		yp[i]   += tp0 * at0 + tp1 *at1;
-		yp[i+1] += tp0 * at2 + tp1 *at3;
-		tmp2[0] += at0 * xp[i] + at2 * xp[i+1];
-		tmp2[1] += at1 * xp[i] + at3 * xp[i+1];
+		at2     = a2[i];
+		at3     = a3[i];
+		x       = xp[i];
+		yp[i]   += tp0 * at0 + tp1 *at1 + tp2 * at2 + tp3 * at3;
+		tmp2[0] += at0 * x;
+		tmp2[1] += at1 * x;
+		tmp2[2] += at2 * x;
+		tmp2[3] += at3 * x;
 
 	}
-	temp2[0] = tmp2[0];
-	temp2[1] = tmp2[1];
 
+	temp2[0] += tmp2[0];
+	temp2[1] += tmp2[1];
+	temp2[2] += tmp2[2];
+	temp2[3] += tmp2[3];
 }
 
 #endif
 
+
+#ifndef HAVE_KERNEL_1x4
+
+static void ssymv_kernel_1x4(BLASLONG from, BLASLONG to, FLOAT *a0, FLOAT *a1, FLOAT *a2, FLOAT *a3, FLOAT *xp, FLOAT *yp, FLOAT *temp1, FLOAT *temp2)
+{
+	FLOAT at0,at1,at2,at3;
+	FLOAT x;
+	FLOAT tmp2[4] = { 0.0, 0.0, 0.0, 0.0 };
+	FLOAT tp0;
+	FLOAT tp1;
+	FLOAT tp2;
+	FLOAT tp3;
+	BLASLONG i;
+
+	tp0 = temp1[0];
+	tp1 = temp1[1];
+	tp2 = temp1[2];
+	tp3 = temp1[3];
+	
+	for (i=from; i<to; i++)
+	{
+		at0     = a0[i];
+		at1     = a1[i];
+		at2     = a2[i];
+		at3     = a3[i];
+		x       = xp[i];
+		yp[i]   += tp0 * at0 + tp1 *at1 + tp2 * at2 + tp3 * at3;
+		tmp2[0] += at0 * x;
+		tmp2[1] += at1 * x;
+		tmp2[2] += at2 * x;
+		tmp2[3] += at3 * x;
+
+	}
+
+	temp2[0] += tmp2[0];
+	temp2[1] += tmp2[1];
+	temp2[2] += tmp2[2];
+	temp2[3] += tmp2[3];
+}
+
+#endif
+
+
 static void ssymv_kernel_8x1(BLASLONG n, FLOAT *a0, FLOAT *xp, FLOAT *yp, FLOAT *temp1, FLOAT *temp2)
 {
 	FLOAT at0,at1,at2,at3;
@@ -100,13 +153,16 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOA
 	BLASLONG ix,iy;
 	BLASLONG jx,jy;
 	BLASLONG j;
+	BLASLONG j1;
+	BLASLONG j2;
 	BLASLONG m2;
 	FLOAT temp1;
 	FLOAT temp2;
 	FLOAT *xp, *yp;
-	FLOAT *a0,*a1;
-	FLOAT tmp1[2];
-	FLOAT tmp2[2];
+	FLOAT *a0,*a1,*a2,*a3;
+	FLOAT at0,at1,at2,at3;
+	FLOAT tmp1[4];
+	FLOAT tmp2[4];
 
 #if 0
 	if( m != offset )
@@ -146,37 +202,45 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOA
 	xp = x;
 	yp = y;
 
-	m2 = m - ( mrange % 2 );
+	m2 = m - ( mrange % 4 );
 
-	for (j=m1; j<m2; j+=2)
+	for (j=m1; j<m2; j+=4)
 	{
 		tmp1[0] = alpha * xp[j];
 		tmp1[1] = alpha * xp[j+1];
+		tmp1[2] = alpha * xp[j+2];
+		tmp1[3] = alpha * xp[j+3];
 		tmp2[0] = 0.0;
 		tmp2[1] = 0.0;
+		tmp2[2] = 0.0;
+		tmp2[3] = 0.0;
 		a0    = &a[j*lda];
 		a1    = a0+lda;
-		FLOAT at0,at1;
-		BLASLONG j1 = (j/16)*16;		
+		a2    = a1+lda;
+		a3    = a2+lda;
+		j1 = (j/8)*8;		
 		if ( j1 )
-			ssymv_kernel_16x2(j1, a0, a1, xp, yp, tmp1, tmp2);
+			ssymv_kernel_4x4(j1, a0, a1, a2, a3, xp, yp, tmp1, tmp2);
+		if ( j1 < j )
+			ssymv_kernel_1x4(j1, j,  a0, a1, a2, a3, xp, yp, tmp1, tmp2);
 
-		for (i=j1; i<j; i++)
+		j2 = 0;
+		for ( j1 = j ; j1 < j+4 ; j1++ )
 		{
-			at0     = a0[i];
-			at1     = a1[i];
-			yp[i]   += tmp1[0] * at0 + tmp1[1] *at1;
-			tmp2[0] += at0 * xp[i];
-			tmp2[1] += at1 * xp[i];
-			
+			temp1 = tmp1[j2];
+			temp2 = tmp2[j2];
+			a0    = &a[j1*lda];
+			for ( i=j ; i<j1; i++ )
+			{
+				yp[i] += temp1 * a0[i];	
+				temp2 += a0[i] * xp[i];
+				
+			}
+			y[j1] += temp1 * a0[j1] + alpha * temp2;
+			j2++;
+
 		}
 
-		at1     = a1[j];
-		yp[j]   += tmp1[1] * at1;
-		tmp2[1] += at1 * xp[j];
-
-		yp[j]   += tmp1[0] * a0[j]   + alpha * tmp2[0];
-		yp[j+1] += tmp1[1] * a1[j+1] + alpha * tmp2[1];
 	}
 
 	for ( ; j<m; j++)
@@ -185,7 +249,7 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOA
 		temp2 = 0.0;
 		a0    = &a[j*lda];
 		FLOAT at0;
-		BLASLONG j1 = (j/8)*8;		
+		j1 = (j/8)*8;		
 
 		if ( j1 )
 			ssymv_kernel_8x1(j1, a0, xp, yp, &temp1, &temp2);
diff --git a/kernel/x86_64/ssymv_U_microk_nehalem-2.c b/kernel/x86_64/ssymv_U_microk_nehalem-2.c
new file mode 100644
index 000000000..9505a395a
--- /dev/null
+++ b/kernel/x86_64/ssymv_U_microk_nehalem-2.c
@@ -0,0 +1,130 @@
+/***************************************************************************
+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_4x4 1
+static void ssymv_kernel_4x4( BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *a2, FLOAT *a3,  FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2) __attribute__ ((noinline));
+
+static void ssymv_kernel_4x4(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *a2, FLOAT *a3, FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"xorps          %%xmm0 , %%xmm0              \n\t"	// temp2[0]
+	"xorps          %%xmm1 , %%xmm1              \n\t"	// temp2[1]
+	"xorps          %%xmm2 , %%xmm2              \n\t"	// temp2[2]
+	"xorps          %%xmm3 , %%xmm3              \n\t"	// temp2[3]
+	"movss	       (%8),    %%xmm4	             \n\t"	// temp1[0]
+	"movss        4(%8),    %%xmm5	             \n\t"	// temp1[1]
+	"movss        8(%8),    %%xmm6	             \n\t"	// temp1[2]
+	"movss       12(%8),    %%xmm7	             \n\t"	// temp1[3]
+	"shufps $0,  %%xmm4, %%xmm4                  \n\t"
+	"shufps $0,  %%xmm5, %%xmm5                  \n\t"
+	"shufps $0,  %%xmm6, %%xmm6                  \n\t"
+	"shufps $0,  %%xmm7, %%xmm7                  \n\t"
+
+	"xorq		%0,%0			     \n\t"
+
+	".align 16		  		       \n\t"
+	".L01LOOP%=:				       \n\t"
+	"movups	            (%2,%0,4), %%xmm8	       \n\t"	// 4 * x
+	"movups	            (%3,%0,4), %%xmm9         \n\t"	// 4 * y
+
+	"movups	            (%4,%0,4), %%xmm12	       \n\t"	// 4 * a
+	"movups	            (%5,%0,4), %%xmm13	       \n\t"	// 4 * a
+
+	"movups		    %%xmm12  , %%xmm11	       \n\t"
+	"mulps		    %%xmm4   , %%xmm11	       \n\t"    // temp1 * a
+	"addps		    %%xmm11  , %%xmm9	       \n\t"    // y += temp1 * a
+	"mulps		    %%xmm8   , %%xmm12	       \n\t"    // a * x
+	"addps		    %%xmm12  , %%xmm0 	       \n\t"    // temp2 += x * a
+
+	"movups	            (%6,%0,4), %%xmm14	       \n\t"	// 4 * a
+	"movups	            (%7,%0,4), %%xmm15	       \n\t"	// 4 * a
+
+	"movups		    %%xmm13  , %%xmm11	       \n\t"
+	"mulps		    %%xmm5   , %%xmm11	       \n\t"    // temp1 * a
+	"addps		    %%xmm11  , %%xmm9	       \n\t"    // y += temp1 * a
+	"mulps		    %%xmm8   , %%xmm13	       \n\t"    // a * x
+	"addps		    %%xmm13  , %%xmm1 	       \n\t"    // temp2 += x * a
+
+	"movups		    %%xmm14  , %%xmm11	       \n\t"
+	"mulps		    %%xmm6   , %%xmm11	       \n\t"    // temp1 * a
+	"addps		    %%xmm11  , %%xmm9	       \n\t"    // y += temp1 * a
+	"mulps		    %%xmm8   , %%xmm14	       \n\t"    // a * x
+	"addps		    %%xmm14  , %%xmm2 	       \n\t"    // temp2 += x * a
+
+	"movups		    %%xmm15  , %%xmm11	       \n\t"
+	"mulps		    %%xmm7   , %%xmm11	       \n\t"    // temp1 * a
+	"addps		    %%xmm11  , %%xmm9	       \n\t"    // y += temp1 * a
+	"mulps		    %%xmm8   , %%xmm15	       \n\t"    // a * x
+	"addps		    %%xmm15  , %%xmm3 	       \n\t"    // temp2 += x * a
+
+	"movups             %%xmm9,   (%3,%0,4)       \n\t"    // 4 * y
+
+        "addq		$4 , %0	  	 	      \n\t"
+	"subq	        $4 , %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \n\t"
+
+	"haddps        %%xmm0, %%xmm0  \n\t"
+	"haddps        %%xmm1, %%xmm1  \n\t"
+	"haddps        %%xmm2, %%xmm2  \n\t"
+	"haddps        %%xmm3, %%xmm3  \n\t"
+	"haddps        %%xmm0, %%xmm0  \n\t"
+	"haddps        %%xmm1, %%xmm1  \n\t"
+	"haddps        %%xmm2, %%xmm2  \n\t"
+	"haddps        %%xmm3, %%xmm3  \n\t"
+
+	"movss         %%xmm0 ,   (%9)		\n\t"	// save temp2
+	"movss         %%xmm1 ,  4(%9)		\n\t"	// save temp2
+	"movss         %%xmm2 ,  8(%9)		\n\t"	// save temp2
+	"movss         %%xmm3 , 12(%9)		\n\t"	// save temp2
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (a0),     // 4
+          "r" (a1),     // 5
+          "r" (a2),     // 6
+          "r" (a3),     // 7
+          "r" (temp1),  // 8
+          "r" (temp2)   // 9
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From f824c2b75167717389d889447968cc6ee5c60368 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Tue, 19 Aug 2014 19:25:03 +0200
Subject: [PATCH 06/32] updated optimized ssymv_U for bulldozer

---
 kernel/x86_64/ssymv_U_microk_bulldozer-2.c | 97 ++++++++++------------
 1 file changed, 46 insertions(+), 51 deletions(-)

diff --git a/kernel/x86_64/ssymv_U_microk_bulldozer-2.c b/kernel/x86_64/ssymv_U_microk_bulldozer-2.c
index b0b0bed65..b8b3b73e9 100644
--- a/kernel/x86_64/ssymv_U_microk_bulldozer-2.c
+++ b/kernel/x86_64/ssymv_U_microk_bulldozer-2.c
@@ -25,10 +25,10 @@ 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_16x2 1
-static void ssymv_kernel_16x2( BLASLONG n, FLOAT *a1, FLOAT *a2, FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2) __attribute__ ((noinline));
+#define HAVE_KERNEL_4x4 1
+static void ssymv_kernel_4x4( BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *a2, FLOAT *a3, FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2) __attribute__ ((noinline));
 
-static void ssymv_kernel_16x2(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2)
+static void ssymv_kernel_4x4(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *a2, FLOAT *a3, FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2)
 {
 
 	BLASLONG register i = 0;
@@ -37,64 +37,57 @@ static void ssymv_kernel_16x2(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *x, FLOAT
 	(
 	"vxorps		%%xmm0 , %%xmm0 , %%xmm0     \n\t"	// temp2[0]
 	"vxorps		%%xmm1 , %%xmm1 , %%xmm1     \n\t"	// temp2[1]
-	"vbroadcastss	(%6),    %%xmm2	             \n\t"	// temp1[0]
-	"vbroadcastss  4(%6),    %%xmm3	             \n\t"	// temp1[1]
+	"vxorps		%%xmm2 , %%xmm2 , %%xmm2     \n\t"	// temp2[2]
+	"vxorps		%%xmm3 , %%xmm3 , %%xmm3     \n\t"	// temp2[3]
+	"vbroadcastss	(%8),    %%xmm4	             \n\t"	// temp1[0]
+	"vbroadcastss  4(%8),    %%xmm5	             \n\t"	// temp1[1]
+	"vbroadcastss  8(%8),    %%xmm6	             \n\t"	// temp1[1]
+	"vbroadcastss 12(%8),    %%xmm7	             \n\t"	// temp1[1]
 
 	"xorq		%0,%0			     \n\t"
 
-	".align 16				 \n\t"
-	".L01LOOP%=:				 \n\t"
+	".align 16				     \n\t"
+	".L01LOOP%=:				     \n\t"
 
-	"prefetcht0	 192(%4,%0,4)		       \n\t"
-	"vmovups	(%4,%0,4), %%xmm4	 \n\t"	// 2 * a0
-	"vmovups      16(%4,%0,4), %%xmm5	 \n\t"	// 2 * a0
-	"prefetcht0	 192(%2,%0,4)		       \n\t"
-	"vmovups	(%2,%0,4), %%xmm8	 \n\t"	// 2 * x
-	"vmovups      16(%2,%0,4), %%xmm9	 \n\t"	// 2 * x
-	"prefetcht0	 192(%3,%0,4)		       \n\t"
-	"vmovups      32(%4,%0,4), %%xmm6	 \n\t"	// 2 * a0
-	"vmovups      48(%4,%0,4), %%xmm7	 \n\t"	// 2 * a0
-	"vmovups      32(%2,%0,4), %%xmm10	 \n\t"	// 2 * x
-	"vmovups      48(%2,%0,4), %%xmm11	 \n\t"	// 2 * x
+	"vmovups	(%2,%0,4), %%xmm8	           \n\t"  // 4 * x
+	"vmovups	(%3,%0,4), %%xmm9	           \n\t"  // 4 * y
 
-	"prefetcht0	 192(%5,%0,4)		            \n\t"
-	"vfmaddps    (%3,%0,4), %%xmm2 , %%xmm4 , %%xmm12   \n\t" // y += temp1 * a0
-	"vfmaddps      %%xmm0 , %%xmm8 , %%xmm4 , %%xmm0    \n\t" // temp2 += a0 * x
-	"vfmaddps  16(%3,%0,4), %%xmm2 , %%xmm5 , %%xmm13   \n\t" // y += temp1 * a0
-	"vmovups	(%5,%0,4), %%xmm4	            \n\t"	// 2 * a1
-	"vfmaddps      %%xmm0 , %%xmm9 , %%xmm5 , %%xmm0    \n\t" // temp2 += a0 * x
-	"vfmaddps  32(%3,%0,4), %%xmm2 , %%xmm6 , %%xmm14   \n\t" // y += temp1 * a0
-	"vmovups      16(%5,%0,4), %%xmm5	            \n\t"	// 2 * a1
-	"vfmaddps      %%xmm0 , %%xmm10, %%xmm6 , %%xmm0    \n\t" // temp2 += a0 * x
-	"vfmaddps  48(%3,%0,4), %%xmm2 , %%xmm7 , %%xmm15   \n\t" // y += temp1 * a0
-	"vmovups      32(%5,%0,4), %%xmm6	            \n\t"	// 2 * a1
-	"vfmaddps      %%xmm0 , %%xmm11, %%xmm7 , %%xmm0    \n\t" // temp2 += a0 * x
-	"vmovups      48(%5,%0,4), %%xmm7	            \n\t"	// 2 * a1
+	"vmovups	(%4,%0,4), %%xmm12	           \n\t"  // 4 * a
+	"vmovups	(%5,%0,4), %%xmm13	           \n\t"  // 4 * a
 
-	"vfmaddps %%xmm12, %%xmm3 , %%xmm4 , %%xmm12   \n\t" // y += temp1 * a1
-	"vfmaddps %%xmm13, %%xmm3 , %%xmm5 , %%xmm13   \n\t" // y += temp1 * a1
-	"vmovups  %%xmm12,   (%3,%0,4)		       \n\t"	// 2 * y
-	"vfmaddps %%xmm14, %%xmm3 , %%xmm6 , %%xmm14   \n\t" // y += temp1 * a1
-	"vmovups  %%xmm13, 16(%3,%0,4)		       \n\t"	// 2 * y
-	"vfmaddps %%xmm15, %%xmm3 , %%xmm7 , %%xmm15   \n\t" // y += temp1 * a1
-	"vmovups  %%xmm14, 32(%3,%0,4)		       \n\t"	// 2 * y
+	"vfmaddps	%%xmm0 , %%xmm8, %%xmm12 , %%xmm0  \n\t"  // temp2 += x * a
+	"vfmaddps	%%xmm9 , %%xmm4, %%xmm12 , %%xmm9  \n\t"  // y     += temp1 * a
 
-	"vfmaddps %%xmm1 , %%xmm8 , %%xmm4 , %%xmm1   \n\t" // temp2 += a1 * x
-	"vfmaddps %%xmm1 , %%xmm9 , %%xmm5 , %%xmm1   \n\t" // temp2 += a1 * x
-	"vmovups  %%xmm15, 48(%3,%0,4)		      \n\t"	// 2 * y
-	"vfmaddps %%xmm1 , %%xmm10, %%xmm6 , %%xmm1   \n\t" // temp2 += a1 * x
-	"vfmaddps %%xmm1 , %%xmm11, %%xmm7 , %%xmm1   \n\t" // temp2 += a1 * x
+	"vfmaddps	%%xmm1 , %%xmm8, %%xmm13 , %%xmm1  \n\t"  // temp2 += x * a
+	"vmovups	(%6,%0,4), %%xmm14	           \n\t"  // 4 * a
+	"vfmaddps	%%xmm9 , %%xmm5, %%xmm13 , %%xmm9  \n\t"  // y     += temp1 * a
 
-        "addq		$16, %0	  	 	      \n\t"
-	"subq	        $16, %1			      \n\t"		
+	"vfmaddps	%%xmm2 , %%xmm8, %%xmm14 , %%xmm2  \n\t"  // temp2 += x * a
+	"vmovups	(%7,%0,4), %%xmm15	           \n\t"  // 4 * a
+	"vfmaddps	%%xmm9 , %%xmm6, %%xmm14 , %%xmm9  \n\t"  // y     += temp1 * a
+
+	"vfmaddps	%%xmm3 , %%xmm8, %%xmm15 , %%xmm3  \n\t"  // temp2 += x * a
+	"vfmaddps	%%xmm9 , %%xmm7, %%xmm15 , %%xmm9  \n\t"  // y     += temp1 * a
+
+	"vmovups	%%xmm9 , (%3,%0,4)		   \n\t"
+
+	"addq		$4 , %0	  	 	      \n\t"
+	"subq	        $4 , %1			      \n\t"		
 	"jnz		.L01LOOP%=		      \n\t"
 
 	"vhaddps        %%xmm0, %%xmm0, %%xmm0  \n\t"
 	"vhaddps        %%xmm1, %%xmm1, %%xmm1  \n\t"
+	"vhaddps        %%xmm2, %%xmm2, %%xmm2  \n\t"
+	"vhaddps        %%xmm3, %%xmm3, %%xmm3  \n\t"
 	"vhaddps        %%xmm0, %%xmm0, %%xmm0  \n\t"
 	"vhaddps        %%xmm1, %%xmm1, %%xmm1  \n\t"
-	"vmovss         %%xmm0 , (%7)		\n\t"	// save temp2
-	"vmovss         %%xmm1 ,4(%7)		\n\t"	// save temp2
+	"vhaddps        %%xmm2, %%xmm2, %%xmm2  \n\t"
+	"vhaddps        %%xmm3, %%xmm3, %%xmm3  \n\t"
+
+	"vmovss         %%xmm0 ,  (%9)		\n\t"	// save temp2
+	"vmovss         %%xmm1 , 4(%9)		\n\t"	// save temp2
+	"vmovss         %%xmm2 , 8(%9)		\n\t"	// save temp2
+	"vmovss         %%xmm3 ,12(%9)		\n\t"	// save temp2
 
 	:
         : 
@@ -102,10 +95,12 @@ static void ssymv_kernel_16x2(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *x, FLOAT
 	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
-          "r" (a0),  // 4
-          "r" (a1),  // 5
-          "r" (temp1),  // 6
-          "r" (temp2)   // 7
+          "r" (a0),     // 4
+          "r" (a1),     // 5
+          "r" (a2),     // 6
+          "r" (a3),     // 7
+          "r" (temp1),  // 8
+          "r" (temp2)   // 9
 	: "cc", 
 	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
 	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 

From ef6374196d0fbb69d0720c973abad9ef39a89253 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Wed, 20 Aug 2014 09:00:56 +0200
Subject: [PATCH 07/32] updated optimized dsymv_U kernel for bulldozer

---
 kernel/x86_64/dsymv_U.c                    | 135 +++++++++++++++------
 kernel/x86_64/dsymv_U_microk_bulldozer-2.c | 115 ++++++++++--------
 2 files changed, 164 insertions(+), 86 deletions(-)

diff --git a/kernel/x86_64/dsymv_U.c b/kernel/x86_64/dsymv_U.c
index 1f22abe8d..267755c2f 100644
--- a/kernel/x86_64/dsymv_U.c
+++ b/kernel/x86_64/dsymv_U.c
@@ -28,43 +28,97 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include "common.h"
 
+
 #if defined(BULLDOZER)
 #include "dsymv_U_microk_bulldozer-2.c"
+#elif defined(NEHALEM)
+#include "dsymv_U_microk_nehalem-2.c"
 #endif
 
+#ifndef HAVE_KERNEL_4x4
 
-#ifndef HAVE_KERNEL_8x2
-
-static void dsymv_kernel_8x2(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *xp, FLOAT *yp, FLOAT *temp1, FLOAT *temp2)
+static void dsymv_kernel_4x4(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *a2, FLOAT *a3, FLOAT *xp, FLOAT *yp, FLOAT *temp1, FLOAT *temp2)
 {
 	FLOAT at0,at1,at2,at3;
-	FLOAT tmp2[2] = { 0.0, 0.0 };
+	FLOAT x;
+	FLOAT tmp2[4] = { 0.0, 0.0, 0.0, 0.0 };
 	FLOAT tp0;
 	FLOAT tp1;
+	FLOAT tp2;
+	FLOAT tp3;
 	BLASLONG i;
 
 	tp0 = temp1[0];
 	tp1 = temp1[1];
+	tp2 = temp1[2];
+	tp3 = temp1[3];
 	
-	for (i=0; i<n; i+=2)
+	for (i=0; i<n; i++)
 	{
 		at0     = a0[i];
 		at1     = a1[i];
-		at2     = a0[i+1];
-		at3     = a1[i+1];
-		yp[i]   += tp0 * at0 + tp1 *at1;
-		yp[i+1] += tp0 * at2 + tp1 *at3;
-		tmp2[0] += at0 * xp[i] + at2 * xp[i+1];
-		tmp2[1] += at1 * xp[i] + at3 * xp[i+1];
+		at2     = a2[i];
+		at3     = a3[i];
+		x       = xp[i];
+		yp[i]   += tp0 * at0 + tp1 *at1 + tp2 * at2 + tp3 * at3;
+		tmp2[0] += at0 * x;
+		tmp2[1] += at1 * x;
+		tmp2[2] += at2 * x;
+		tmp2[3] += at3 * x;
 
 	}
-	temp2[0] = tmp2[0];
-	temp2[1] = tmp2[1];
 
+	temp2[0] += tmp2[0];
+	temp2[1] += tmp2[1];
+	temp2[2] += tmp2[2];
+	temp2[3] += tmp2[3];
 }
 
 #endif
 
+
+#ifndef HAVE_KERNEL_1x4
+
+static void dsymv_kernel_1x4(BLASLONG from, BLASLONG to, FLOAT *a0, FLOAT *a1, FLOAT *a2, FLOAT *a3, FLOAT *xp, FLOAT *yp, FLOAT *temp1, FLOAT *temp2)
+{
+	FLOAT at0,at1,at2,at3;
+	FLOAT x;
+	FLOAT tmp2[4] = { 0.0, 0.0, 0.0, 0.0 };
+	FLOAT tp0;
+	FLOAT tp1;
+	FLOAT tp2;
+	FLOAT tp3;
+	BLASLONG i;
+
+	tp0 = temp1[0];
+	tp1 = temp1[1];
+	tp2 = temp1[2];
+	tp3 = temp1[3];
+	
+	for (i=from; i<to; i++)
+	{
+		at0     = a0[i];
+		at1     = a1[i];
+		at2     = a2[i];
+		at3     = a3[i];
+		x       = xp[i];
+		yp[i]   += tp0 * at0 + tp1 *at1 + tp2 * at2 + tp3 * at3;
+		tmp2[0] += at0 * x;
+		tmp2[1] += at1 * x;
+		tmp2[2] += at2 * x;
+		tmp2[3] += at3 * x;
+
+	}
+
+	temp2[0] += tmp2[0];
+	temp2[1] += tmp2[1];
+	temp2[2] += tmp2[2];
+	temp2[3] += tmp2[3];
+}
+
+#endif
+
+
 static void dsymv_kernel_8x1(BLASLONG n, FLOAT *a0, FLOAT *xp, FLOAT *yp, FLOAT *temp1, FLOAT *temp2)
 {
 	FLOAT at0,at1,at2,at3;
@@ -99,13 +153,16 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOA
 	BLASLONG ix,iy;
 	BLASLONG jx,jy;
 	BLASLONG j;
+	BLASLONG j1;
+	BLASLONG j2;
 	BLASLONG m2;
 	FLOAT temp1;
 	FLOAT temp2;
 	FLOAT *xp, *yp;
-	FLOAT *a0,*a1;
-	FLOAT tmp1[2];
-	FLOAT tmp2[2];
+	FLOAT *a0,*a1,*a2,*a3;
+	FLOAT at0,at1,at2,at3;
+	FLOAT tmp1[4];
+	FLOAT tmp2[4];
 
 #if 0
 	if( m != offset )
@@ -145,37 +202,45 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOA
 	xp = x;
 	yp = y;
 
-	m2 = m - ( mrange % 2 );
+	m2 = m - ( mrange % 4 );
 
-	for (j=m1; j<m2; j+=2)
+	for (j=m1; j<m2; j+=4)
 	{
 		tmp1[0] = alpha * xp[j];
 		tmp1[1] = alpha * xp[j+1];
+		tmp1[2] = alpha * xp[j+2];
+		tmp1[3] = alpha * xp[j+3];
 		tmp2[0] = 0.0;
 		tmp2[1] = 0.0;
+		tmp2[2] = 0.0;
+		tmp2[3] = 0.0;
 		a0    = &a[j*lda];
 		a1    = a0+lda;
-		FLOAT at0,at1;
-		BLASLONG j1 = (j/8)*8;		
+		a2    = a1+lda;
+		a3    = a2+lda;
+		j1 = (j/8)*8;		
 		if ( j1 )
-			dsymv_kernel_8x2(j1, a0, a1, xp, yp, tmp1, tmp2);
+			dsymv_kernel_4x4(j1, a0, a1, a2, a3, xp, yp, tmp1, tmp2);
+		if ( j1 < j )
+			dsymv_kernel_1x4(j1, j,  a0, a1, a2, a3, xp, yp, tmp1, tmp2);
 
-		for (i=j1; i<j; i++)
+		j2 = 0;
+		for ( j1 = j ; j1 < j+4 ; j1++ )
 		{
-			at0     = a0[i];
-			at1     = a1[i];
-			yp[i]   += tmp1[0] * at0 + tmp1[1] *at1;
-			tmp2[0] += at0 * xp[i];
-			tmp2[1] += at1 * xp[i];
-			
+			temp1 = tmp1[j2];
+			temp2 = tmp2[j2];
+			a0    = &a[j1*lda];
+			for ( i=j ; i<j1; i++ )
+			{
+				yp[i] += temp1 * a0[i];	
+				temp2 += a0[i] * xp[i];
+				
+			}
+			y[j1] += temp1 * a0[j1] + alpha * temp2;
+			j2++;
+
 		}
 
-		at1     = a1[j];
-		yp[j]   += tmp1[1] * at1;
-		tmp2[1] += at1 * xp[j];
-
-		yp[j]   += tmp1[0] * a0[j]   + alpha * tmp2[0];
-		yp[j+1] += tmp1[1] * a1[j+1] + alpha * tmp2[1];
 	}
 
 	for ( ; j<m; j++)
@@ -184,7 +249,7 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOA
 		temp2 = 0.0;
 		a0    = &a[j*lda];
 		FLOAT at0;
-		BLASLONG j1 = (j/8)*8;		
+		j1 = (j/8)*8;		
 
 		if ( j1 )
 			dsymv_kernel_8x1(j1, a0, xp, yp, &temp1, &temp2);
diff --git a/kernel/x86_64/dsymv_U_microk_bulldozer-2.c b/kernel/x86_64/dsymv_U_microk_bulldozer-2.c
index 3b035222b..492920253 100644
--- a/kernel/x86_64/dsymv_U_microk_bulldozer-2.c
+++ b/kernel/x86_64/dsymv_U_microk_bulldozer-2.c
@@ -25,10 +25,10 @@ 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_8x2 1
-static void dsymv_kernel_8x2( BLASLONG n, FLOAT *a1, FLOAT *a2, FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2) __attribute__ ((noinline));
+#define HAVE_KERNEL_4x4 1
+static void dsymv_kernel_4x4( BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *a2, FLOAT *a3, FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2) __attribute__ ((noinline));
 
-static void dsymv_kernel_8x2(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2)
+static void dsymv_kernel_4x4(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *a2, FLOAT *a3, FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2)
 {
 
 	BLASLONG register i = 0;
@@ -37,62 +37,73 @@ static void dsymv_kernel_8x2(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *x, FLOAT *
 	(
 	"vxorpd		%%xmm0 , %%xmm0 , %%xmm0     \n\t"	// temp2[0]
 	"vxorpd		%%xmm1 , %%xmm1 , %%xmm1     \n\t"	// temp2[1]
-	"vmovddup	(%6),    %%xmm2	             \n\t"	// temp1[0]
-	"vmovddup      8(%6),    %%xmm3	             \n\t"	// temp1[1]
+	"vxorpd		%%xmm2 , %%xmm2 , %%xmm2     \n\t"	// temp2[2]
+	"vxorpd		%%xmm3 , %%xmm3 , %%xmm3     \n\t"	// temp2[3]
+	"vmovddup	(%8),    %%xmm4	             \n\t"	// temp1[0]
+	"vmovddup  8(%8),    %%xmm5	             \n\t"	// temp1[1]
+	"vmovddup 16(%8),    %%xmm6	             \n\t"	// temp1[1]
+	"vmovddup 24(%8),    %%xmm7	             \n\t"	// temp1[1]
 
 	"xorq		%0,%0			     \n\t"
 
-	".align 16				 \n\t"
-	".L01LOOP%=:				 \n\t"
+	".align 16				     \n\t"
+	".L01LOOP%=:				     \n\t"
 
-	"prefetcht0	 192(%4,%0,8)		       \n\t"
-	"vmovups	(%4,%0,8), %%xmm4	 \n\t"	// 2 * a0
-	"vmovups      16(%4,%0,8), %%xmm5	 \n\t"	// 2 * a0
-	"prefetcht0	 192(%2,%0,8)		       \n\t"
-	"vmovups	(%2,%0,8), %%xmm8	 \n\t"	// 2 * x
-	"vmovups      16(%2,%0,8), %%xmm9	 \n\t"	// 2 * x
-	"prefetcht0	 192(%3,%0,8)		       \n\t"
-	"vmovups      32(%4,%0,8), %%xmm6	 \n\t"	// 2 * a0
-	"vmovups      48(%4,%0,8), %%xmm7	 \n\t"	// 2 * a0
-	"vmovups      32(%2,%0,8), %%xmm10	 \n\t"	// 2 * x
-	"vmovups      48(%2,%0,8), %%xmm11	 \n\t"	// 2 * x
+	"vmovups	(%4,%0,8), %%xmm12	           \n\t"  // 2 * a
+	"vmovups	(%2,%0,8), %%xmm8	           \n\t"  // 2 * x
+	"vmovups	(%3,%0,8), %%xmm9	           \n\t"  // 2 * y
 
-	"prefetcht0	 192(%5,%0,8)		            \n\t"
-	"vfmaddpd    (%3,%0,8), %%xmm2 , %%xmm4 , %%xmm12   \n\t" // y += temp1 * a0
-	"vfmaddpd      %%xmm0 , %%xmm8 , %%xmm4 , %%xmm0    \n\t" // temp2 += a0 * x
-	"vfmaddpd  16(%3,%0,8), %%xmm2 , %%xmm5 , %%xmm13   \n\t" // y += temp1 * a0
-	"vmovups	(%5,%0,8), %%xmm4	            \n\t"	// 2 * a1
-	"vfmaddpd      %%xmm0 , %%xmm9 , %%xmm5 , %%xmm0    \n\t" // temp2 += a0 * x
-	"vfmaddpd  32(%3,%0,8), %%xmm2 , %%xmm6 , %%xmm14   \n\t" // y += temp1 * a0
-	"vmovups      16(%5,%0,8), %%xmm5	            \n\t"	// 2 * a1
-	"vfmaddpd      %%xmm0 , %%xmm10, %%xmm6 , %%xmm0    \n\t" // temp2 += a0 * x
-	"vfmaddpd  48(%3,%0,8), %%xmm2 , %%xmm7 , %%xmm15   \n\t" // y += temp1 * a0
-	"vmovups      32(%5,%0,8), %%xmm6	            \n\t"	// 2 * a1
-	"vfmaddpd      %%xmm0 , %%xmm11, %%xmm7 , %%xmm0    \n\t" // temp2 += a0 * x
-	"vmovups      48(%5,%0,8), %%xmm7	            \n\t"	// 2 * a1
+	"vmovups	(%5,%0,8), %%xmm13	           \n\t"  // 2 * a
 
-	"vfmaddpd %%xmm12, %%xmm3 , %%xmm4 , %%xmm12   \n\t" // y += temp1 * a1
-	"vfmaddpd %%xmm13, %%xmm3 , %%xmm5 , %%xmm13   \n\t" // y += temp1 * a1
-	"vmovups  %%xmm12,   (%3,%0,8)		       \n\t"	// 2 * y
-	"vfmaddpd %%xmm14, %%xmm3 , %%xmm6 , %%xmm14   \n\t" // y += temp1 * a1
-	"vmovups  %%xmm13, 16(%3,%0,8)		       \n\t"	// 2 * y
-	"vfmaddpd %%xmm15, %%xmm3 , %%xmm7 , %%xmm15   \n\t" // y += temp1 * a1
-	"vmovups  %%xmm14, 32(%3,%0,8)		       \n\t"	// 2 * y
+	"vfmaddpd	%%xmm0 , %%xmm8, %%xmm12 , %%xmm0  \n\t"  // temp2 += x * a
+	"vfmaddpd	%%xmm9 , %%xmm4, %%xmm12 , %%xmm9  \n\t"  // y     += temp1 * a
+	"vmovups	(%6,%0,8), %%xmm14	           \n\t"  // 2 * a
 
-	"vfmaddpd %%xmm1 , %%xmm8 , %%xmm4 , %%xmm1   \n\t" // temp2 += a1 * x
-	"vfmaddpd %%xmm1 , %%xmm9 , %%xmm5 , %%xmm1   \n\t" // temp2 += a1 * x
-	"vmovups  %%xmm15, 48(%3,%0,8)		      \n\t"	// 2 * y
-	"vfmaddpd %%xmm1 , %%xmm10, %%xmm6 , %%xmm1   \n\t" // temp2 += a1 * x
-	"vfmaddpd %%xmm1 , %%xmm11, %%xmm7 , %%xmm1   \n\t" // temp2 += a1 * x
+	"vfmaddpd	%%xmm1 , %%xmm8, %%xmm13 , %%xmm1  \n\t"  // temp2 += x * a
+	"vfmaddpd	%%xmm9 , %%xmm5, %%xmm13 , %%xmm9  \n\t"  // y     += temp1 * a
+	"vmovups	(%7,%0,8), %%xmm15	           \n\t"  // 2 * a
+
+	"vmovups	16(%3,%0,8), %%xmm11	           \n\t"  // 2 * y
+	"vfmaddpd	%%xmm2 , %%xmm8, %%xmm14 , %%xmm2  \n\t"  // temp2 += x * a
+	"vmovups	16(%4,%0,8), %%xmm12	           \n\t"  // 2 * a
+	"vfmaddpd	%%xmm9 , %%xmm6, %%xmm14 , %%xmm9  \n\t"  // y     += temp1 * a
+	"vmovups	16(%2,%0,8), %%xmm10	           \n\t"  // 2 * x
+
+	"vfmaddpd	%%xmm3 , %%xmm8, %%xmm15 , %%xmm3  \n\t"  // temp2 += x * a
+	"vfmaddpd	%%xmm9 , %%xmm7, %%xmm15 , %%xmm9  \n\t"  // y     += temp1 * a
+
+	"vmovups	16(%5,%0,8), %%xmm13	           \n\t"  // 2 * a
+	"vmovups	16(%6,%0,8), %%xmm14	           \n\t"  // 2 * a
+
+	"vfmaddpd	%%xmm0 , %%xmm10, %%xmm12 , %%xmm0  \n\t"  // temp2 += x * a
+	"vfmaddpd	%%xmm11 , %%xmm4, %%xmm12 , %%xmm11  \n\t"  // y     += temp1 * a
+
+	"vmovups	16(%7,%0,8), %%xmm15	           \n\t"  // 2 * a
+	"vfmaddpd	%%xmm1 , %%xmm10, %%xmm13 , %%xmm1  \n\t"  // temp2 += x * a
+	"vfmaddpd	%%xmm11 , %%xmm5, %%xmm13 , %%xmm11  \n\t"  // y     += temp1 * a
+
+	"vfmaddpd	%%xmm2 , %%xmm10, %%xmm14 , %%xmm2  \n\t"  // temp2 += x * a
+	"addq		$4 , %0	  	 	      \n\t"
+	"vfmaddpd	%%xmm11 , %%xmm6, %%xmm14 , %%xmm11  \n\t"  // y     += temp1 * a
+
+	"vfmaddpd	%%xmm3 , %%xmm10, %%xmm15 , %%xmm3  \n\t"  // temp2 += x * a
+	"vfmaddpd	%%xmm11 , %%xmm7, %%xmm15 , %%xmm11  \n\t"  // y     += temp1 * a
+	"subq	        $4 , %1			      \n\t"		
+
+	"vmovups	%%xmm9 ,  -32(%3,%0,8)		   \n\t"
+	"vmovups	%%xmm11 , -16(%3,%0,8)		   \n\t"
 
-        "addq		$8, %0	  	 	      \n\t"
-	"subq	        $8, %1			      \n\t"		
 	"jnz		.L01LOOP%=		      \n\t"
 
 	"vhaddpd        %%xmm0, %%xmm0, %%xmm0  \n\t"
 	"vhaddpd        %%xmm1, %%xmm1, %%xmm1  \n\t"
-	"vmovsd         %%xmm0 , (%7)		\n\t"	// save temp2
-	"vmovsd         %%xmm1 ,8(%7)		\n\t"	// save temp2
+	"vhaddpd        %%xmm2, %%xmm2, %%xmm2  \n\t"
+	"vhaddpd        %%xmm3, %%xmm3, %%xmm3  \n\t"
+
+	"vmovsd         %%xmm0 ,  (%9)		\n\t"	// save temp2
+	"vmovsd         %%xmm1 , 8(%9)		\n\t"	// save temp2
+	"vmovsd         %%xmm2 ,16(%9)		\n\t"	// save temp2
+	"vmovsd         %%xmm3 ,24(%9)		\n\t"	// save temp2
 
 	:
         : 
@@ -100,10 +111,12 @@ static void dsymv_kernel_8x2(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *x, FLOAT *
 	  "r" (n),  	// 1
           "r" (x),      // 2
           "r" (y),      // 3
-          "r" (a0),  // 4
-          "r" (a1),  // 5
-          "r" (temp1),  // 6
-          "r" (temp2)   // 7
+          "r" (a0),     // 4
+          "r" (a1),     // 5
+          "r" (a2),     // 6
+          "r" (a3),     // 7
+          "r" (temp1),  // 8
+          "r" (temp2)   // 9
 	: "cc", 
 	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
 	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 

From 8247f38dc1159b46c769495b6d37f844d33d414a Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Wed, 20 Aug 2014 09:58:04 +0200
Subject: [PATCH 08/32] added optimized dsymv_U kernel for nehalem

---
 kernel/x86_64/KERNEL.NEHALEM             |   2 +-
 kernel/x86_64/dsymv_U_microk_nehalem-2.c | 125 +++++++++++++++++++++++
 2 files changed, 126 insertions(+), 1 deletion(-)
 create mode 100644 kernel/x86_64/dsymv_U_microk_nehalem-2.c

diff --git a/kernel/x86_64/KERNEL.NEHALEM b/kernel/x86_64/KERNEL.NEHALEM
index 353514449..35be648a4 100644
--- a/kernel/x86_64/KERNEL.NEHALEM
+++ b/kernel/x86_64/KERNEL.NEHALEM
@@ -1,4 +1,4 @@
-#DSYMV_U_KERNEL = dsymv_U.c
+DSYMV_U_KERNEL = dsymv_U.c
 SSYMV_U_KERNEL = ssymv_U.c
 
 SGEMVNKERNEL = sgemv_n.c
diff --git a/kernel/x86_64/dsymv_U_microk_nehalem-2.c b/kernel/x86_64/dsymv_U_microk_nehalem-2.c
new file mode 100644
index 000000000..6aab57500
--- /dev/null
+++ b/kernel/x86_64/dsymv_U_microk_nehalem-2.c
@@ -0,0 +1,125 @@
+/***************************************************************************
+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_4x4 1
+static void dsymv_kernel_4x4( BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *a2, FLOAT *a3,  FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2) __attribute__ ((noinline));
+
+static void dsymv_kernel_4x4(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *a2, FLOAT *a3, FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"xorpd          %%xmm0 , %%xmm0              \n\t"	// temp2[0]
+	"xorpd          %%xmm1 , %%xmm1              \n\t"	// temp2[1]
+	"xorpd          %%xmm2 , %%xmm2              \n\t"	// temp2[2]
+	"xorpd          %%xmm3 , %%xmm3              \n\t"	// temp2[3]
+	"movsd	       (%8),    %%xmm4	             \n\t"	// temp1[0]
+	"movsd        8(%8),    %%xmm5	             \n\t"	// temp1[1]
+	"movsd       16(%8),    %%xmm6	             \n\t"	// temp1[2]
+	"movsd       24(%8),    %%xmm7	             \n\t"	// temp1[3]
+	"shufpd $0,  %%xmm4, %%xmm4                  \n\t"
+	"shufpd $0,  %%xmm5, %%xmm5                  \n\t"
+	"shufpd $0,  %%xmm6, %%xmm6                  \n\t"
+	"shufpd $0,  %%xmm7, %%xmm7                  \n\t"
+
+	"xorq		%0,%0			     \n\t"
+
+	".align 16		  		       \n\t"
+	".L01LOOP%=:				       \n\t"
+	"movups	            (%4,%0,8), %%xmm12	       \n\t"	// 2 * a
+	"movups	            (%2,%0,8), %%xmm8	       \n\t"	// 2 * x
+	"movups		    %%xmm12  , %%xmm11	       \n\t"
+	"movups	            (%3,%0,8), %%xmm9         \n\t"	// 2 * y
+	"movups	            (%5,%0,8), %%xmm13	       \n\t"	// 2 * a
+
+	"mulpd		    %%xmm4   , %%xmm11	       \n\t"    // temp1 * a
+	"addpd		    %%xmm11  , %%xmm9	       \n\t"    // y += temp1 * a
+	"mulpd		    %%xmm8   , %%xmm12	       \n\t"    // a * x
+	"addpd		    %%xmm12  , %%xmm0 	       \n\t"    // temp2 += x * a
+
+	"movups	            (%6,%0,8), %%xmm14	       \n\t"	// 2 * a
+	"movups	            (%7,%0,8), %%xmm15	       \n\t"	// 2 * a
+
+	"movups		    %%xmm13  , %%xmm11	       \n\t"
+	"mulpd		    %%xmm5   , %%xmm11	       \n\t"    // temp1 * a
+	"addpd		    %%xmm11  , %%xmm9	       \n\t"    // y += temp1 * a
+	"mulpd		    %%xmm8   , %%xmm13	       \n\t"    // a * x
+	"addpd		    %%xmm13  , %%xmm1 	       \n\t"    // temp2 += x * a
+
+	"movups		    %%xmm14  , %%xmm11	       \n\t"
+	"mulpd		    %%xmm6   , %%xmm11	       \n\t"    // temp1 * a
+	"addpd		    %%xmm11  , %%xmm9	       \n\t"    // y += temp1 * a
+	"mulpd		    %%xmm8   , %%xmm14	       \n\t"    // a * x
+	"addpd		    %%xmm14  , %%xmm2 	       \n\t"    // temp2 += x * a
+
+        "addq		$2 , %0	  	 	      \n\t"
+	"movups		    %%xmm15  , %%xmm11	       \n\t"
+	"mulpd		    %%xmm7   , %%xmm11	       \n\t"    // temp1 * a
+	"addpd		    %%xmm11  , %%xmm9	       \n\t"    // y += temp1 * a
+	"mulpd		    %%xmm8   , %%xmm15	       \n\t"    // a * x
+	"addpd		    %%xmm15  , %%xmm3 	       \n\t"    // temp2 += x * a
+
+	"movups             %%xmm9,-16(%3,%0,8)       \n\t"    // 2 * y
+
+	"subq	        $2 , %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \n\t"
+
+	"haddpd        %%xmm0, %%xmm0  \n\t"
+	"haddpd        %%xmm1, %%xmm1  \n\t"
+	"haddpd        %%xmm2, %%xmm2  \n\t"
+	"haddpd        %%xmm3, %%xmm3  \n\t"
+
+	"movsd         %%xmm0 ,   (%9)		\n\t"	// save temp2
+	"movsd         %%xmm1 ,  8(%9)		\n\t"	// save temp2
+	"movsd         %%xmm2 , 16(%9)		\n\t"	// save temp2
+	"movsd         %%xmm3 , 24(%9)		\n\t"	// save temp2
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (a0),     // 4
+          "r" (a1),     // 5
+          "r" (a2),     // 6
+          "r" (a3),     // 7
+          "r" (temp1),  // 8
+          "r" (temp2)   // 9
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From f6f9122660ee8be175b61dd91abc1a86e9bb4a7e Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Thu, 21 Aug 2014 13:02:53 +0200
Subject: [PATCH 09/32] added optimized dsymv_L kernel for bulldozer

---
 kernel/x86_64/dsymv_L.c                    | 299 +++++++++++++++++++++
 kernel/x86_64/dsymv_L_microk_bulldozer-2.c | 137 ++++++++++
 2 files changed, 436 insertions(+)
 create mode 100644 kernel/x86_64/dsymv_L.c
 create mode 100644 kernel/x86_64/dsymv_L_microk_bulldozer-2.c

diff --git a/kernel/x86_64/dsymv_L.c b/kernel/x86_64/dsymv_L.c
new file mode 100644
index 000000000..29e3aa2f8
--- /dev/null
+++ b/kernel/x86_64/dsymv_L.c
@@ -0,0 +1,299 @@
+/***************************************************************************
+Copyright (c) 2013, 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 "dsymv_L_microk_bulldozer-2.c"
+#elif defined(NEHALEM)
+#include "dsymv_U_microk_nehalem-2.c"
+#endif
+
+
+#ifndef HAVE_KERNEL_4x4
+
+static void dsymv_kernel_4x4(BLASLONG from, BLASLONG to, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *tmp1, FLOAT *temp2)
+{
+        FLOAT tmp2[4] = { 0.0, 0.0, 0.0, 0.0 };
+        BLASLONG i;
+
+        for (i=from; i<to; i+=4)
+        {
+
+		y[i]    += tmp1[0] * ap[0][i];
+		tmp2[0] += ap[0][i] * x[i];
+		y[i]    += tmp1[1] * ap[1][i];
+		tmp2[1] += ap[1][i] * x[i];
+		y[i]    += tmp1[2] * ap[2][i];
+		tmp2[2] += ap[2][i] * x[i];
+		y[i]    += tmp1[3] * ap[3][i];
+		tmp2[3] += ap[3][i] * x[i];
+
+		y[i+1]  += tmp1[0] * ap[0][i+1];
+		tmp2[0] += ap[0][i+1] * x[i+1];
+		y[i+1]  += tmp1[1] * ap[1][i+1];
+		tmp2[1] += ap[1][i+1] * x[i+1];
+		y[i+1]  += tmp1[2] * ap[2][i+1];
+		tmp2[2] += ap[2][i+1] * x[i+1];
+		y[i+1]  += tmp1[3] * ap[3][i+1];
+		tmp2[3] += ap[3][i+1] * x[i+1];
+
+		y[i+2]  += tmp1[0] * ap[0][i+2];
+		tmp2[0] += ap[0][i+2] * x[i+2];
+		y[i+2]  += tmp1[1] * ap[1][i+2];
+		tmp2[1] += ap[1][i+2] * x[i+2];
+		y[i+2]  += tmp1[2] * ap[2][i+2];
+		tmp2[2] += ap[2][i+2] * x[i+2];
+		y[i+2]  += tmp1[3] * ap[3][i+2];
+		tmp2[3] += ap[3][i+2] * x[i+2];
+
+		y[i+3]  += tmp1[0] * ap[0][i+3];
+		tmp2[0] += ap[0][i+3] * x[i+3];
+		y[i+3]  += tmp1[1] * ap[1][i+3];
+		tmp2[1] += ap[1][i+3] * x[i+3];
+		y[i+3]  += tmp1[2] * ap[2][i+3];
+		tmp2[2] += ap[2][i+3] * x[i+3];
+		y[i+3]  += tmp1[3] * ap[3][i+3];
+		tmp2[3] += ap[3][i+3] * x[i+3];
+
+        }
+
+        temp2[0] += tmp2[0];
+        temp2[1] += tmp2[1];
+        temp2[2] += tmp2[2];
+        temp2[3] += tmp2[3];
+}
+
+#endif
+
+
+
+
+int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG ix,iy;
+	BLASLONG jx,jy;
+	BLASLONG j;
+	FLOAT temp1;
+	FLOAT temp2;
+	FLOAT tmp1[4];
+	FLOAT tmp2[4];
+	FLOAT *ap[4];
+
+#if 0
+	if ( m != offset )
+		printf("Symv_L: m=%d offset=%d\n",m,offset);
+#endif
+
+
+	if ( (inc_x != 1) || (inc_y != 1) )
+	{
+
+		jx = 0;
+		jy = 0;
+
+		for (j=0; j<offset; j++)
+		{
+			temp1 = alpha * x[jx];
+			temp2 = 0.0;
+			y[jy] += temp1 * a[j*lda+j];
+			iy = jy;
+			ix = jx;
+			for (i=j+1; i<m; i++)
+			{
+				ix += inc_x;
+				iy += inc_y;
+				y[iy] += temp1 * a[j*lda+i];
+				temp2 += a[j*lda+i] * x[ix];
+			
+			}
+			y[jy] += alpha * temp2;
+			jx    += inc_x;
+			jy    += inc_y;
+		}
+		return(0);
+	}
+
+	BLASLONG offset1 = (offset/4)*4;
+
+	for (j=0; j<offset1; j+=4)
+	{
+		tmp1[0] = alpha * x[j];
+		tmp1[1] = alpha * x[j+1];
+		tmp1[2] = alpha * x[j+2];
+		tmp1[3] = alpha * x[j+3];
+		tmp2[0] = 0.0;
+		tmp2[1] = 0.0;
+		tmp2[2] = 0.0;
+		tmp2[3] = 0.0;
+		ap[0]   = &a[j*lda];
+		ap[1]   = ap[0] + lda;
+		ap[2]   = ap[1] + lda;
+		ap[3]   = ap[2] + lda;
+		y[j]   += tmp1[0] * ap[0][j];
+		y[j+1] += tmp1[1] * ap[1][j+1];
+		y[j+2] += tmp1[2] * ap[2][j+2];
+		y[j+3] += tmp1[3] * ap[3][j+3];
+		BLASLONG from = j+1;
+		if ( m - from >=12 )
+		{
+			BLASLONG m2 = (m/4)*4;
+			for (i=j+1; i<j+4; i++)
+			{
+				y[i] += tmp1[0] * ap[0][i];
+				tmp2[0] += ap[0][i] * x[i];
+			}
+
+			for (i=j+2; i<j+4; i++)
+			{
+				y[i] += tmp1[1] * ap[1][i];
+				tmp2[1] += ap[1][i] * x[i];
+			}
+
+			for (i=j+3; i<j+4; i++)
+			{
+				y[i] += tmp1[2] * ap[2][i];
+				tmp2[2] += ap[2][i] * x[i];
+			}
+
+			if ( m2 > j+4 )
+				dsymv_kernel_4x4(j+4,m2,ap,x,y,tmp1,tmp2);
+
+
+			for (i=m2; i<m; i++)
+			{
+				y[i] += tmp1[0] * ap[0][i];
+				tmp2[0] += ap[0][i] * x[i];
+
+				y[i] += tmp1[1] * ap[1][i];
+				tmp2[1] += ap[1][i] * x[i];
+
+				y[i] += tmp1[2] * ap[2][i];
+				tmp2[2] += ap[2][i] * x[i];
+
+				y[i] += tmp1[3] * ap[3][i];
+				tmp2[3] += ap[3][i] * x[i];
+
+			}
+
+
+		}
+		else
+		{
+
+			for (i=j+1; i<j+4; i++)
+			{
+				y[i] += tmp1[0] * ap[0][i];
+				tmp2[0] += ap[0][i] * x[i];
+			}
+
+			for (i=j+2; i<j+4; i++)
+			{
+				y[i] += tmp1[1] * ap[1][i];
+				tmp2[1] += ap[1][i] * x[i];
+			}
+
+			for (i=j+3; i<j+4; i++)
+			{
+				y[i] += tmp1[2] * ap[2][i];
+				tmp2[2] += ap[2][i] * x[i];
+			}
+
+			for (i=j+4; i<m; i++)
+			{
+				y[i] += tmp1[0] * ap[0][i];
+				tmp2[0] += ap[0][i] * x[i];
+
+				y[i] += tmp1[1] * ap[1][i];
+				tmp2[1] += ap[1][i] * x[i];
+
+				y[i] += tmp1[2] * ap[2][i];
+				tmp2[2] += ap[2][i] * x[i];
+
+				y[i] += tmp1[3] * ap[3][i];
+				tmp2[3] += ap[3][i] * x[i];
+
+			}
+
+		}
+		y[j]   += alpha * tmp2[0];
+		y[j+1] += alpha * tmp2[1];
+		y[j+2] += alpha * tmp2[2];
+		y[j+3] += alpha * tmp2[3];
+	}
+
+
+	for (j=offset1; j<offset; j++)
+	{
+		temp1 = alpha * x[j];
+		temp2 = 0.0;
+		y[j] += temp1 * a[j*lda+j];
+		BLASLONG from = j+1;
+		if ( m - from >=8 )
+		{
+			BLASLONG j1 = ((from + 4)/4)*4;
+			BLASLONG j2 = (m/4)*4;
+			for (i=from; i<j1; i++)
+			{
+				y[i] += temp1 * a[j*lda+i];
+				temp2 += a[j*lda+i] * x[i];
+			
+			}
+
+			for (i=j1; i<j2; i++)
+			{
+				y[i] += temp1 * a[j*lda+i];
+				temp2 += a[j*lda+i] * x[i];
+			
+			}
+
+			for (i=j2; i<m; i++)
+			{
+				y[i] += temp1 * a[j*lda+i];
+				temp2 += a[j*lda+i] * x[i];
+			
+			}
+
+		}
+		else
+		{
+			for (i=from; i<m; i++)
+			{
+				y[i] += temp1 * a[j*lda+i];
+				temp2 += a[j*lda+i] * x[i];
+			
+			}
+
+		}
+		y[j] += alpha * temp2;
+	}
+	return(0);
+}
+
+
diff --git a/kernel/x86_64/dsymv_L_microk_bulldozer-2.c b/kernel/x86_64/dsymv_L_microk_bulldozer-2.c
new file mode 100644
index 000000000..70d8df36b
--- /dev/null
+++ b/kernel/x86_64/dsymv_L_microk_bulldozer-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_4x4 1
+static void dsymv_kernel_4x4( BLASLONG from, BLASLONG to, FLOAT **a, FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2) __attribute__ ((noinline));
+
+static void dsymv_kernel_4x4(BLASLONG from, BLASLONG to, FLOAT **a, FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2)
+{
+
+
+	__asm__  __volatile__
+	(
+	"vxorpd		%%xmm0 , %%xmm0 , %%xmm0     \n\t"	// temp2[0]
+	"vxorpd		%%xmm1 , %%xmm1 , %%xmm1     \n\t"	// temp2[1]
+	"vxorpd		%%xmm2 , %%xmm2 , %%xmm2     \n\t"	// temp2[2]
+	"vxorpd		%%xmm3 , %%xmm3 , %%xmm3     \n\t"	// temp2[3]
+	"vmovddup   (%8),    %%xmm4	             \n\t"	// temp1[0]
+	"vmovddup  8(%8),    %%xmm5	             \n\t"	// temp1[1]
+	"vmovddup 16(%8),    %%xmm6	             \n\t"	// temp1[1]
+	"vmovddup 24(%8),    %%xmm7	             \n\t"	// temp1[1]
+
+	".align 16				     \n\t"
+	".L01LOOP%=:				     \n\t"
+
+	"vmovups	(%4,%0,8), %%xmm12	           \n\t"  // 2 * a
+	"vmovups	(%2,%0,8), %%xmm8	           \n\t"  // 2 * x
+	"vmovups	(%3,%0,8), %%xmm9	           \n\t"  // 2 * y
+
+	"vmovups	(%5,%0,8), %%xmm13	           \n\t"  // 2 * a
+
+	"vfmaddpd	%%xmm0 , %%xmm8, %%xmm12 , %%xmm0  \n\t"  // temp2 += x * a
+	"vfmaddpd	%%xmm9 , %%xmm4, %%xmm12 , %%xmm9  \n\t"  // y     += temp1 * a
+	"vmovups	(%6,%0,8), %%xmm14	           \n\t"  // 2 * a
+
+	"vfmaddpd	%%xmm1 , %%xmm8, %%xmm13 , %%xmm1  \n\t"  // temp2 += x * a
+	"vfmaddpd	%%xmm9 , %%xmm5, %%xmm13 , %%xmm9  \n\t"  // y     += temp1 * a
+	"vmovups	(%7,%0,8), %%xmm15	           \n\t"  // 2 * a
+
+	"vmovups	16(%3,%0,8), %%xmm11	           \n\t"  // 2 * y
+	"vfmaddpd	%%xmm2 , %%xmm8, %%xmm14 , %%xmm2  \n\t"  // temp2 += x * a
+	"vmovups	16(%4,%0,8), %%xmm12	           \n\t"  // 2 * a
+	"vfmaddpd	%%xmm9 , %%xmm6, %%xmm14 , %%xmm9  \n\t"  // y     += temp1 * a
+	"vmovups	16(%2,%0,8), %%xmm10	           \n\t"  // 2 * x
+
+	"vfmaddpd	%%xmm3 , %%xmm8, %%xmm15 , %%xmm3  \n\t"  // temp2 += x * a
+	"vfmaddpd	%%xmm9 , %%xmm7, %%xmm15 , %%xmm9  \n\t"  // y     += temp1 * a
+
+	"vmovups	16(%5,%0,8), %%xmm13	           \n\t"  // 2 * a
+	"vmovups	16(%6,%0,8), %%xmm14	           \n\t"  // 2 * a
+
+	"vfmaddpd	%%xmm0 , %%xmm10, %%xmm12 , %%xmm0  \n\t"  // temp2 += x * a
+	"vfmaddpd	%%xmm11 , %%xmm4, %%xmm12 , %%xmm11  \n\t"  // y     += temp1 * a
+
+	"vmovups	16(%7,%0,8), %%xmm15	           \n\t"  // 2 * a
+	"vfmaddpd	%%xmm1 , %%xmm10, %%xmm13 , %%xmm1  \n\t"  // temp2 += x * a
+	"vfmaddpd	%%xmm11 , %%xmm5, %%xmm13 , %%xmm11  \n\t"  // y     += temp1 * a
+
+	"vfmaddpd	%%xmm2 , %%xmm10, %%xmm14 , %%xmm2  \n\t"  // temp2 += x * a
+	"vfmaddpd	%%xmm11 , %%xmm6, %%xmm14 , %%xmm11  \n\t"  // y     += temp1 * a
+
+	"vfmaddpd	%%xmm3 , %%xmm10, %%xmm15 , %%xmm3  \n\t"  // temp2 += x * a
+	"vfmaddpd	%%xmm11 , %%xmm7, %%xmm15 , %%xmm11  \n\t"  // y     += temp1 * a
+	"addq		$4 , %0	  	 	      \n\t"
+
+	"vmovups	%%xmm9 ,  -32(%3,%0,8)		   \n\t"
+	"vmovups	%%xmm11 , -16(%3,%0,8)		   \n\t"
+
+	"cmpq		%0 , %1			      \n\t"
+	"jnz		.L01LOOP%=		      \n\t"
+
+	"vmovsd		  (%9), %%xmm4		      \n\t"
+	"vmovsd		 8(%9), %%xmm5		      \n\t"
+	"vmovsd		16(%9), %%xmm6		      \n\t"
+	"vmovsd		24(%9), %%xmm7		      \n\t"
+
+	"vhaddpd        %%xmm0, %%xmm0, %%xmm0  \n\t"
+	"vhaddpd        %%xmm1, %%xmm1, %%xmm1  \n\t"
+	"vhaddpd        %%xmm2, %%xmm2, %%xmm2  \n\t"
+	"vhaddpd        %%xmm3, %%xmm3, %%xmm3  \n\t"
+
+	"vaddsd		%%xmm4, %%xmm0, %%xmm0  \n\t"
+	"vaddsd		%%xmm5, %%xmm1, %%xmm1  \n\t"
+	"vaddsd		%%xmm6, %%xmm2, %%xmm2  \n\t"
+	"vaddsd		%%xmm7, %%xmm3, %%xmm3  \n\t"
+
+	"vmovsd         %%xmm0 ,  (%9)		\n\t"	// save temp2
+	"vmovsd         %%xmm1 , 8(%9)		\n\t"	// save temp2
+	"vmovsd         %%xmm2 ,16(%9)		\n\t"	// save temp2
+	"vmovsd         %%xmm3 ,24(%9)		\n\t"	// save temp2
+
+	:
+        : 
+          "r" (from),	// 0	
+	  "r" (to),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (a[0]),	// 4
+          "r" (a[1]),	// 5
+          "r" (a[2]),	// 6
+          "r" (a[3]),	// 8
+          "r" (temp1),  // 8
+          "r" (temp2)   // 9
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From ec05ff3f64ce95c065ad30cfe8cf18a56ec8d4cd Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Thu, 21 Aug 2014 13:32:06 +0200
Subject: [PATCH 10/32] added optimized ssymv_L kernel for bulldozer

---
 kernel/x86_64/ssymv_L.c                    | 299 +++++++++++++++++++++
 kernel/x86_64/ssymv_L_microk_bulldozer-2.c | 122 +++++++++
 2 files changed, 421 insertions(+)
 create mode 100644 kernel/x86_64/ssymv_L.c
 create mode 100644 kernel/x86_64/ssymv_L_microk_bulldozer-2.c

diff --git a/kernel/x86_64/ssymv_L.c b/kernel/x86_64/ssymv_L.c
new file mode 100644
index 000000000..352f1c862
--- /dev/null
+++ b/kernel/x86_64/ssymv_L.c
@@ -0,0 +1,299 @@
+/***************************************************************************
+Copyright (c) 2013, 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 "ssymv_L_microk_bulldozer-2.c"
+#elif defined(NEHALEM)
+#include "ssymv_U_microk_nehalem-2.c"
+#endif
+
+
+#ifndef HAVE_KERNEL_4x4
+
+static void ssymv_kernel_4x4(BLASLONG from, BLASLONG to, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *tmp1, FLOAT *temp2)
+{
+        FLOAT tmp2[4] = { 0.0, 0.0, 0.0, 0.0 };
+        BLASLONG i;
+
+        for (i=from; i<to; i+=4)
+        {
+
+		y[i]    += tmp1[0] * ap[0][i];
+		tmp2[0] += ap[0][i] * x[i];
+		y[i]    += tmp1[1] * ap[1][i];
+		tmp2[1] += ap[1][i] * x[i];
+		y[i]    += tmp1[2] * ap[2][i];
+		tmp2[2] += ap[2][i] * x[i];
+		y[i]    += tmp1[3] * ap[3][i];
+		tmp2[3] += ap[3][i] * x[i];
+
+		y[i+1]  += tmp1[0] * ap[0][i+1];
+		tmp2[0] += ap[0][i+1] * x[i+1];
+		y[i+1]  += tmp1[1] * ap[1][i+1];
+		tmp2[1] += ap[1][i+1] * x[i+1];
+		y[i+1]  += tmp1[2] * ap[2][i+1];
+		tmp2[2] += ap[2][i+1] * x[i+1];
+		y[i+1]  += tmp1[3] * ap[3][i+1];
+		tmp2[3] += ap[3][i+1] * x[i+1];
+
+		y[i+2]  += tmp1[0] * ap[0][i+2];
+		tmp2[0] += ap[0][i+2] * x[i+2];
+		y[i+2]  += tmp1[1] * ap[1][i+2];
+		tmp2[1] += ap[1][i+2] * x[i+2];
+		y[i+2]  += tmp1[2] * ap[2][i+2];
+		tmp2[2] += ap[2][i+2] * x[i+2];
+		y[i+2]  += tmp1[3] * ap[3][i+2];
+		tmp2[3] += ap[3][i+2] * x[i+2];
+
+		y[i+3]  += tmp1[0] * ap[0][i+3];
+		tmp2[0] += ap[0][i+3] * x[i+3];
+		y[i+3]  += tmp1[1] * ap[1][i+3];
+		tmp2[1] += ap[1][i+3] * x[i+3];
+		y[i+3]  += tmp1[2] * ap[2][i+3];
+		tmp2[2] += ap[2][i+3] * x[i+3];
+		y[i+3]  += tmp1[3] * ap[3][i+3];
+		tmp2[3] += ap[3][i+3] * x[i+3];
+
+        }
+
+        temp2[0] += tmp2[0];
+        temp2[1] += tmp2[1];
+        temp2[2] += tmp2[2];
+        temp2[3] += tmp2[3];
+}
+
+#endif
+
+
+
+
+int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG ix,iy;
+	BLASLONG jx,jy;
+	BLASLONG j;
+	FLOAT temp1;
+	FLOAT temp2;
+	FLOAT tmp1[4];
+	FLOAT tmp2[4];
+	FLOAT *ap[4];
+
+#if 0
+	if ( m != offset )
+		printf("Symv_L: m=%d offset=%d\n",m,offset);
+#endif
+
+
+	if ( (inc_x != 1) || (inc_y != 1) )
+	{
+
+		jx = 0;
+		jy = 0;
+
+		for (j=0; j<offset; j++)
+		{
+			temp1 = alpha * x[jx];
+			temp2 = 0.0;
+			y[jy] += temp1 * a[j*lda+j];
+			iy = jy;
+			ix = jx;
+			for (i=j+1; i<m; i++)
+			{
+				ix += inc_x;
+				iy += inc_y;
+				y[iy] += temp1 * a[j*lda+i];
+				temp2 += a[j*lda+i] * x[ix];
+			
+			}
+			y[jy] += alpha * temp2;
+			jx    += inc_x;
+			jy    += inc_y;
+		}
+		return(0);
+	}
+
+	BLASLONG offset1 = (offset/4)*4;
+
+	for (j=0; j<offset1; j+=4)
+	{
+		tmp1[0] = alpha * x[j];
+		tmp1[1] = alpha * x[j+1];
+		tmp1[2] = alpha * x[j+2];
+		tmp1[3] = alpha * x[j+3];
+		tmp2[0] = 0.0;
+		tmp2[1] = 0.0;
+		tmp2[2] = 0.0;
+		tmp2[3] = 0.0;
+		ap[0]   = &a[j*lda];
+		ap[1]   = ap[0] + lda;
+		ap[2]   = ap[1] + lda;
+		ap[3]   = ap[2] + lda;
+		y[j]   += tmp1[0] * ap[0][j];
+		y[j+1] += tmp1[1] * ap[1][j+1];
+		y[j+2] += tmp1[2] * ap[2][j+2];
+		y[j+3] += tmp1[3] * ap[3][j+3];
+		BLASLONG from = j+1;
+		if ( m - from >=12 )
+		{
+			BLASLONG m2 = (m/4)*4;
+			for (i=j+1; i<j+4; i++)
+			{
+				y[i] += tmp1[0] * ap[0][i];
+				tmp2[0] += ap[0][i] * x[i];
+			}
+
+			for (i=j+2; i<j+4; i++)
+			{
+				y[i] += tmp1[1] * ap[1][i];
+				tmp2[1] += ap[1][i] * x[i];
+			}
+
+			for (i=j+3; i<j+4; i++)
+			{
+				y[i] += tmp1[2] * ap[2][i];
+				tmp2[2] += ap[2][i] * x[i];
+			}
+
+			if ( m2 > j+4 )
+				ssymv_kernel_4x4(j+4,m2,ap,x,y,tmp1,tmp2);
+
+
+			for (i=m2; i<m; i++)
+			{
+				y[i] += tmp1[0] * ap[0][i];
+				tmp2[0] += ap[0][i] * x[i];
+
+				y[i] += tmp1[1] * ap[1][i];
+				tmp2[1] += ap[1][i] * x[i];
+
+				y[i] += tmp1[2] * ap[2][i];
+				tmp2[2] += ap[2][i] * x[i];
+
+				y[i] += tmp1[3] * ap[3][i];
+				tmp2[3] += ap[3][i] * x[i];
+
+			}
+
+
+		}
+		else
+		{
+
+			for (i=j+1; i<j+4; i++)
+			{
+				y[i] += tmp1[0] * ap[0][i];
+				tmp2[0] += ap[0][i] * x[i];
+			}
+
+			for (i=j+2; i<j+4; i++)
+			{
+				y[i] += tmp1[1] * ap[1][i];
+				tmp2[1] += ap[1][i] * x[i];
+			}
+
+			for (i=j+3; i<j+4; i++)
+			{
+				y[i] += tmp1[2] * ap[2][i];
+				tmp2[2] += ap[2][i] * x[i];
+			}
+
+			for (i=j+4; i<m; i++)
+			{
+				y[i] += tmp1[0] * ap[0][i];
+				tmp2[0] += ap[0][i] * x[i];
+
+				y[i] += tmp1[1] * ap[1][i];
+				tmp2[1] += ap[1][i] * x[i];
+
+				y[i] += tmp1[2] * ap[2][i];
+				tmp2[2] += ap[2][i] * x[i];
+
+				y[i] += tmp1[3] * ap[3][i];
+				tmp2[3] += ap[3][i] * x[i];
+
+			}
+
+		}
+		y[j]   += alpha * tmp2[0];
+		y[j+1] += alpha * tmp2[1];
+		y[j+2] += alpha * tmp2[2];
+		y[j+3] += alpha * tmp2[3];
+	}
+
+
+	for (j=offset1; j<offset; j++)
+	{
+		temp1 = alpha * x[j];
+		temp2 = 0.0;
+		y[j] += temp1 * a[j*lda+j];
+		BLASLONG from = j+1;
+		if ( m - from >=8 )
+		{
+			BLASLONG j1 = ((from + 4)/4)*4;
+			BLASLONG j2 = (m/4)*4;
+			for (i=from; i<j1; i++)
+			{
+				y[i] += temp1 * a[j*lda+i];
+				temp2 += a[j*lda+i] * x[i];
+			
+			}
+
+			for (i=j1; i<j2; i++)
+			{
+				y[i] += temp1 * a[j*lda+i];
+				temp2 += a[j*lda+i] * x[i];
+			
+			}
+
+			for (i=j2; i<m; i++)
+			{
+				y[i] += temp1 * a[j*lda+i];
+				temp2 += a[j*lda+i] * x[i];
+			
+			}
+
+		}
+		else
+		{
+			for (i=from; i<m; i++)
+			{
+				y[i] += temp1 * a[j*lda+i];
+				temp2 += a[j*lda+i] * x[i];
+			
+			}
+
+		}
+		y[j] += alpha * temp2;
+	}
+	return(0);
+}
+
+
diff --git a/kernel/x86_64/ssymv_L_microk_bulldozer-2.c b/kernel/x86_64/ssymv_L_microk_bulldozer-2.c
new file mode 100644
index 000000000..c9206c1be
--- /dev/null
+++ b/kernel/x86_64/ssymv_L_microk_bulldozer-2.c
@@ -0,0 +1,122 @@
+/***************************************************************************
+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_4x4 1
+static void ssymv_kernel_4x4( BLASLONG from, BLASLONG to, FLOAT **a, FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2) __attribute__ ((noinline));
+
+static void ssymv_kernel_4x4(BLASLONG from, BLASLONG to, FLOAT **a, FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2)
+{
+
+
+	__asm__  __volatile__
+	(
+	"vxorps		%%xmm0 , %%xmm0 , %%xmm0     \n\t"	// temp2[0]
+	"vxorps		%%xmm1 , %%xmm1 , %%xmm1     \n\t"	// temp2[1]
+	"vxorps		%%xmm2 , %%xmm2 , %%xmm2     \n\t"	// temp2[2]
+	"vxorps		%%xmm3 , %%xmm3 , %%xmm3     \n\t"	// temp2[3]
+	"vbroadcastss   (%8),    %%xmm4	             \n\t"	// temp1[0]
+	"vbroadcastss  4(%8),    %%xmm5	             \n\t"	// temp1[1]
+	"vbroadcastss  8(%8),    %%xmm6	             \n\t"	// temp1[2]
+	"vbroadcastss 12(%8),    %%xmm7	             \n\t"	// temp1[3]
+
+	".align 16				     \n\t"
+	".L01LOOP%=:				     \n\t"
+
+	"vmovups	(%4,%0,4), %%xmm12	           \n\t"  // 2 * a
+	"vmovups	(%2,%0,4), %%xmm8	           \n\t"  // 2 * x
+	"vmovups	(%3,%0,4), %%xmm9	           \n\t"  // 2 * y
+
+	"vmovups	(%5,%0,4), %%xmm13	           \n\t"  // 2 * a
+
+	"vfmaddps	%%xmm0 , %%xmm8, %%xmm12 , %%xmm0  \n\t"  // temp2 += x * a
+	"vfmaddps	%%xmm9 , %%xmm4, %%xmm12 , %%xmm9  \n\t"  // y     += temp1 * a
+	"vmovups	(%6,%0,4), %%xmm14	           \n\t"  // 2 * a
+
+	"vfmaddps	%%xmm1 , %%xmm8, %%xmm13 , %%xmm1  \n\t"  // temp2 += x * a
+	"vfmaddps	%%xmm9 , %%xmm5, %%xmm13 , %%xmm9  \n\t"  // y     += temp1 * a
+	"vmovups	(%7,%0,4), %%xmm15	           \n\t"  // 2 * a
+
+	"vfmaddps	%%xmm2 , %%xmm8, %%xmm14 , %%xmm2  \n\t"  // temp2 += x * a
+	"vfmaddps	%%xmm9 , %%xmm6, %%xmm14 , %%xmm9  \n\t"  // y     += temp1 * a
+
+	"vfmaddps	%%xmm3 , %%xmm8, %%xmm15 , %%xmm3  \n\t"  // temp2 += x * a
+	"vfmaddps	%%xmm9 , %%xmm7, %%xmm15 , %%xmm9  \n\t"  // y     += temp1 * a
+
+	"addq		$4 , %0	  	 	      \n\t"
+
+	"vmovups	%%xmm9 ,  -16(%3,%0,4)		   \n\t"
+
+	"cmpq		%0 , %1			      \n\t"
+	"jnz		.L01LOOP%=		      \n\t"
+
+	"vmovss		  (%9), %%xmm4		      \n\t"
+	"vmovss		 4(%9), %%xmm5		      \n\t"
+	"vmovss		 8(%9), %%xmm6		      \n\t"
+	"vmovss		12(%9), %%xmm7		      \n\t"
+
+	"vhaddps        %%xmm0, %%xmm0, %%xmm0  \n\t"
+	"vhaddps        %%xmm1, %%xmm1, %%xmm1  \n\t"
+	"vhaddps        %%xmm2, %%xmm2, %%xmm2  \n\t"
+	"vhaddps        %%xmm3, %%xmm3, %%xmm3  \n\t"
+	"vhaddps        %%xmm0, %%xmm0, %%xmm0  \n\t"
+	"vhaddps        %%xmm1, %%xmm1, %%xmm1  \n\t"
+	"vhaddps        %%xmm2, %%xmm2, %%xmm2  \n\t"
+	"vhaddps        %%xmm3, %%xmm3, %%xmm3  \n\t"
+
+	"vaddss		%%xmm4, %%xmm0, %%xmm0  \n\t"
+	"vaddss		%%xmm5, %%xmm1, %%xmm1  \n\t"
+	"vaddss		%%xmm6, %%xmm2, %%xmm2  \n\t"
+	"vaddss		%%xmm7, %%xmm3, %%xmm3  \n\t"
+
+	"vmovss         %%xmm0 ,  (%9)		\n\t"	// save temp2
+	"vmovss         %%xmm1 , 4(%9)		\n\t"	// save temp2
+	"vmovss         %%xmm2 , 8(%9)		\n\t"	// save temp2
+	"vmovss         %%xmm3 ,12(%9)		\n\t"	// save temp2
+
+	:
+        : 
+          "r" (from),	// 0	
+	  "r" (to),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (a[0]),	// 4
+          "r" (a[1]),	// 5
+          "r" (a[2]),	// 6
+          "r" (a[3]),	// 8
+          "r" (temp1),  // 8
+          "r" (temp2)   // 9
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From c8b06452665e09de319599deebdc5a9c0bc7ea5f Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Thu, 21 Aug 2014 14:27:00 +0200
Subject: [PATCH 11/32] added optimized symv_L kernels for nehalem

---
 kernel/x86_64/KERNEL.BULLDOZER           |   2 +
 kernel/x86_64/KERNEL.NEHALEM             |   2 +
 kernel/x86_64/dsymv_L.c                  |   2 +-
 kernel/x86_64/dsymv_L_microk_nehalem-2.c | 132 ++++++++++++++++++++++
 kernel/x86_64/ssymv_L.c                  |   2 +-
 kernel/x86_64/ssymv_L_microk_nehalem-2.c | 137 +++++++++++++++++++++++
 6 files changed, 275 insertions(+), 2 deletions(-)
 create mode 100644 kernel/x86_64/dsymv_L_microk_nehalem-2.c
 create mode 100644 kernel/x86_64/ssymv_L_microk_nehalem-2.c

diff --git a/kernel/x86_64/KERNEL.BULLDOZER b/kernel/x86_64/KERNEL.BULLDOZER
index a078528cd..3ee1978b8 100644
--- a/kernel/x86_64/KERNEL.BULLDOZER
+++ b/kernel/x86_64/KERNEL.BULLDOZER
@@ -1,5 +1,7 @@
 DSYMV_U_KERNEL = dsymv_U.c
+DSYMV_L_KERNEL = dsymv_L.c
 SSYMV_U_KERNEL = ssymv_U.c
+SSYMV_L_KERNEL = ssymv_L.c
 
 SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t.c
diff --git a/kernel/x86_64/KERNEL.NEHALEM b/kernel/x86_64/KERNEL.NEHALEM
index 35be648a4..b16fd9c49 100644
--- a/kernel/x86_64/KERNEL.NEHALEM
+++ b/kernel/x86_64/KERNEL.NEHALEM
@@ -1,5 +1,7 @@
 DSYMV_U_KERNEL = dsymv_U.c
+DSYMV_L_KERNEL = dsymv_L.c
 SSYMV_U_KERNEL = ssymv_U.c
+SSYMV_L_KERNEL = ssymv_L.c
 
 SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t.c
diff --git a/kernel/x86_64/dsymv_L.c b/kernel/x86_64/dsymv_L.c
index 29e3aa2f8..8d1337746 100644
--- a/kernel/x86_64/dsymv_L.c
+++ b/kernel/x86_64/dsymv_L.c
@@ -31,7 +31,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #if defined(BULLDOZER)
 #include "dsymv_L_microk_bulldozer-2.c"
 #elif defined(NEHALEM)
-#include "dsymv_U_microk_nehalem-2.c"
+#include "dsymv_L_microk_nehalem-2.c"
 #endif
 
 
diff --git a/kernel/x86_64/dsymv_L_microk_nehalem-2.c b/kernel/x86_64/dsymv_L_microk_nehalem-2.c
new file mode 100644
index 000000000..3ba596c5e
--- /dev/null
+++ b/kernel/x86_64/dsymv_L_microk_nehalem-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_4x4 1
+static void dsymv_kernel_4x4( BLASLONG from, BLASLONG to, FLOAT **a, FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2) __attribute__ ((noinline));
+
+static void dsymv_kernel_4x4(BLASLONG from, BLASLONG to, FLOAT **a, FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2)
+{
+
+
+	__asm__  __volatile__
+	(
+	"xorpd          %%xmm0 , %%xmm0              \n\t"	// temp2[0]
+	"xorpd          %%xmm1 , %%xmm1              \n\t"	// temp2[1]
+	"xorpd          %%xmm2 , %%xmm2              \n\t"	// temp2[2]
+	"xorpd          %%xmm3 , %%xmm3              \n\t"	// temp2[3]
+	"movsd	       (%8),    %%xmm4	             \n\t"	// temp1[0]
+	"movsd        8(%8),    %%xmm5	             \n\t"	// temp1[1]
+	"movsd       16(%8),    %%xmm6	             \n\t"	// temp1[2]
+	"movsd       24(%8),    %%xmm7	             \n\t"	// temp1[3]
+	"shufpd $0,  %%xmm4, %%xmm4                  \n\t"
+	"shufpd $0,  %%xmm5, %%xmm5                  \n\t"
+	"shufpd $0,  %%xmm6, %%xmm6                  \n\t"
+	"shufpd $0,  %%xmm7, %%xmm7                  \n\t"
+
+	".align 16		  		       \n\t"
+	".L01LOOP%=:				       \n\t"
+	"movups	            (%4,%0,8), %%xmm12	       \n\t"	// 2 * a
+	"movups	            (%2,%0,8), %%xmm8	       \n\t"	// 2 * x
+	"movups		    %%xmm12  , %%xmm11	       \n\t"
+	"movups	            (%3,%0,8), %%xmm9         \n\t"	// 2 * y
+	"movups	            (%5,%0,8), %%xmm13	       \n\t"	// 2 * a
+
+	"mulpd		    %%xmm4   , %%xmm11	       \n\t"    // temp1 * a
+	"addpd		    %%xmm11  , %%xmm9	       \n\t"    // y += temp1 * a
+	"mulpd		    %%xmm8   , %%xmm12	       \n\t"    // a * x
+	"addpd		    %%xmm12  , %%xmm0 	       \n\t"    // temp2 += x * a
+
+	"movups	            (%6,%0,8), %%xmm14	       \n\t"	// 2 * a
+	"movups	            (%7,%0,8), %%xmm15	       \n\t"	// 2 * a
+
+	"movups		    %%xmm13  , %%xmm11	       \n\t"
+	"mulpd		    %%xmm5   , %%xmm11	       \n\t"    // temp1 * a
+	"addpd		    %%xmm11  , %%xmm9	       \n\t"    // y += temp1 * a
+	"mulpd		    %%xmm8   , %%xmm13	       \n\t"    // a * x
+	"addpd		    %%xmm13  , %%xmm1 	       \n\t"    // temp2 += x * a
+
+	"movups		    %%xmm14  , %%xmm11	       \n\t"
+	"mulpd		    %%xmm6   , %%xmm11	       \n\t"    // temp1 * a
+	"addpd		    %%xmm11  , %%xmm9	       \n\t"    // y += temp1 * a
+	"mulpd		    %%xmm8   , %%xmm14	       \n\t"    // a * x
+	"addpd		    %%xmm14  , %%xmm2 	       \n\t"    // temp2 += x * a
+
+        "addq		$2 , %0	  	 	      \n\t"
+	"movups		    %%xmm15  , %%xmm11	       \n\t"
+	"mulpd		    %%xmm7   , %%xmm11	       \n\t"    // temp1 * a
+	"addpd		    %%xmm11  , %%xmm9	       \n\t"    // y += temp1 * a
+	"mulpd		    %%xmm8   , %%xmm15	       \n\t"    // a * x
+	"addpd		    %%xmm15  , %%xmm3 	       \n\t"    // temp2 += x * a
+
+	"movups             %%xmm9,-16(%3,%0,8)       \n\t"    // 2 * y
+
+	"cmpq	        %0 , %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \n\t"
+
+	"movsd	       (%9),    %%xmm4	             \n\t"	// temp1[0]
+	"movsd        8(%9),    %%xmm5	             \n\t"	// temp1[1]
+	"movsd       16(%9),    %%xmm6	             \n\t"	// temp1[2]
+	"movsd       24(%9),    %%xmm7	             \n\t"	// temp1[3]
+
+	"haddpd        %%xmm0, %%xmm0  \n\t"
+	"haddpd        %%xmm1, %%xmm1  \n\t"
+	"haddpd        %%xmm2, %%xmm2  \n\t"
+	"haddpd        %%xmm3, %%xmm3  \n\t"
+
+	"addsd         %%xmm4, %%xmm0  \n\t"
+	"addsd         %%xmm5, %%xmm1  \n\t"
+	"addsd         %%xmm6, %%xmm2  \n\t"
+	"addsd         %%xmm7, %%xmm3  \n\t"
+
+	"movsd         %%xmm0 ,   (%9)		\n\t"	// save temp2
+	"movsd         %%xmm1 ,  8(%9)		\n\t"	// save temp2
+	"movsd         %%xmm2 , 16(%9)		\n\t"	// save temp2
+	"movsd         %%xmm3 , 24(%9)		\n\t"	// save temp2
+
+	:
+        : 
+          "r" (from),	// 0	
+	  "r" (to),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (a[0]),   // 4
+          "r" (a[1]),   // 5
+          "r" (a[2]),   // 6
+          "r" (a[3]),   // 7
+          "r" (temp1),  // 8
+          "r" (temp2)   // 9
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
diff --git a/kernel/x86_64/ssymv_L.c b/kernel/x86_64/ssymv_L.c
index 352f1c862..096adc6ca 100644
--- a/kernel/x86_64/ssymv_L.c
+++ b/kernel/x86_64/ssymv_L.c
@@ -31,7 +31,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #if defined(BULLDOZER)
 #include "ssymv_L_microk_bulldozer-2.c"
 #elif defined(NEHALEM)
-#include "ssymv_U_microk_nehalem-2.c"
+#include "ssymv_L_microk_nehalem-2.c"
 #endif
 
 
diff --git a/kernel/x86_64/ssymv_L_microk_nehalem-2.c b/kernel/x86_64/ssymv_L_microk_nehalem-2.c
new file mode 100644
index 000000000..a1c62caf6
--- /dev/null
+++ b/kernel/x86_64/ssymv_L_microk_nehalem-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_4x4 1
+static void ssymv_kernel_4x4( BLASLONG from, BLASLONG to, FLOAT **a, FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2) __attribute__ ((noinline));
+
+static void ssymv_kernel_4x4(BLASLONG from, BLASLONG to,  FLOAT **a, FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2)
+{
+
+
+	__asm__  __volatile__
+	(
+	"xorps          %%xmm0 , %%xmm0              \n\t"	// temp2[0]
+	"xorps          %%xmm1 , %%xmm1              \n\t"	// temp2[1]
+	"xorps          %%xmm2 , %%xmm2              \n\t"	// temp2[2]
+	"xorps          %%xmm3 , %%xmm3              \n\t"	// temp2[3]
+	"movss	       (%8),    %%xmm4	             \n\t"	// temp1[0]
+	"movss        4(%8),    %%xmm5	             \n\t"	// temp1[1]
+	"movss        8(%8),    %%xmm6	             \n\t"	// temp1[2]
+	"movss       12(%8),    %%xmm7	             \n\t"	// temp1[3]
+	"shufps $0,  %%xmm4, %%xmm4                  \n\t"
+	"shufps $0,  %%xmm5, %%xmm5                  \n\t"
+	"shufps $0,  %%xmm6, %%xmm6                  \n\t"
+	"shufps $0,  %%xmm7, %%xmm7                  \n\t"
+
+	".align 16		  		       \n\t"
+	".L01LOOP%=:				       \n\t"
+	"movups	            (%2,%0,4), %%xmm8	       \n\t"	// 4 * x
+	"movups	            (%3,%0,4), %%xmm9         \n\t"	// 4 * y
+
+	"movups	            (%4,%0,4), %%xmm12	       \n\t"	// 4 * a
+	"movups	            (%5,%0,4), %%xmm13	       \n\t"	// 4 * a
+
+	"movups		    %%xmm12  , %%xmm11	       \n\t"
+	"mulps		    %%xmm4   , %%xmm11	       \n\t"    // temp1 * a
+	"addps		    %%xmm11  , %%xmm9	       \n\t"    // y += temp1 * a
+	"mulps		    %%xmm8   , %%xmm12	       \n\t"    // a * x
+	"addps		    %%xmm12  , %%xmm0 	       \n\t"    // temp2 += x * a
+
+	"movups	            (%6,%0,4), %%xmm14	       \n\t"	// 4 * a
+	"movups	            (%7,%0,4), %%xmm15	       \n\t"	// 4 * a
+
+	"movups		    %%xmm13  , %%xmm11	       \n\t"
+	"mulps		    %%xmm5   , %%xmm11	       \n\t"    // temp1 * a
+	"addps		    %%xmm11  , %%xmm9	       \n\t"    // y += temp1 * a
+	"mulps		    %%xmm8   , %%xmm13	       \n\t"    // a * x
+	"addps		    %%xmm13  , %%xmm1 	       \n\t"    // temp2 += x * a
+
+	"movups		    %%xmm14  , %%xmm11	       \n\t"
+	"mulps		    %%xmm6   , %%xmm11	       \n\t"    // temp1 * a
+	"addps		    %%xmm11  , %%xmm9	       \n\t"    // y += temp1 * a
+	"mulps		    %%xmm8   , %%xmm14	       \n\t"    // a * x
+	"addps		    %%xmm14  , %%xmm2 	       \n\t"    // temp2 += x * a
+
+	"movups		    %%xmm15  , %%xmm11	       \n\t"
+	"mulps		    %%xmm7   , %%xmm11	       \n\t"    // temp1 * a
+	"addps		    %%xmm11  , %%xmm9	       \n\t"    // y += temp1 * a
+	"mulps		    %%xmm8   , %%xmm15	       \n\t"    // a * x
+	"addps		    %%xmm15  , %%xmm3 	       \n\t"    // temp2 += x * a
+
+	"movups             %%xmm9,   (%3,%0,4)       \n\t"    // 4 * y
+
+        "addq		$4 , %0	  	 	      \n\t"
+	"cmpq	        %0 , %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \n\t"
+
+	"movss	       (%9),    %%xmm4	             \n\t"	// temp1[0]
+	"movss        4(%9),    %%xmm5	             \n\t"	// temp1[1]
+	"movss        8(%9),    %%xmm6	             \n\t"	// temp1[2]
+	"movss       12(%9),    %%xmm7	             \n\t"	// temp1[3]
+
+	"haddps        %%xmm0, %%xmm0  \n\t"
+	"haddps        %%xmm1, %%xmm1  \n\t"
+	"haddps        %%xmm2, %%xmm2  \n\t"
+	"haddps        %%xmm3, %%xmm3  \n\t"
+	"haddps        %%xmm0, %%xmm0  \n\t"
+	"haddps        %%xmm1, %%xmm1  \n\t"
+	"haddps        %%xmm2, %%xmm2  \n\t"
+	"haddps        %%xmm3, %%xmm3  \n\t"
+
+	"addss         %%xmm4, %%xmm0  \n\t"
+	"addss         %%xmm5, %%xmm1  \n\t"
+	"addss         %%xmm6, %%xmm2  \n\t"
+	"addss         %%xmm7, %%xmm3  \n\t"
+
+	"movss         %%xmm0 ,   (%9)		\n\t"	// save temp2
+	"movss         %%xmm1 ,  4(%9)		\n\t"	// save temp2
+	"movss         %%xmm2 ,  8(%9)		\n\t"	// save temp2
+	"movss         %%xmm3 , 12(%9)		\n\t"	// save temp2
+
+	:
+        : 
+          "r" (from),	// 0	
+	  "r" (to),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (a[0]),   // 4
+          "r" (a[1]),   // 5
+          "r" (a[2]),   // 6
+          "r" (a[3]),   // 7
+          "r" (temp1),  // 8
+          "r" (temp2)   // 9
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From 6f73ffc114a6d56b85c3023f6e8f7856e4198aec Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Thu, 21 Aug 2014 19:33:57 +0200
Subject: [PATCH 12/32] added benchmarks for csymv and zsymv

---
 benchmark/Makefile | 40 ++++++++++++++++++++++++++++++++++++----
 1 file changed, 36 insertions(+), 4 deletions(-)

diff --git a/benchmark/Makefile b/benchmark/Makefile
index de94dcc59..3b2d263c8 100644
--- a/benchmark/Makefile
+++ b/benchmark/Makefile
@@ -35,7 +35,7 @@ goto :: slinpack.goto dlinpack.goto clinpack.goto zlinpack.goto \
        ssyrk.goto dsyrk.goto csyrk.goto zsyrk.goto \
        ssyr2k.goto dsyr2k.goto csyr2k.goto zsyr2k.goto \
        sger.goto dger.goto \
-       ssymv.goto dsymv.goto \
+       ssymv.goto dsymv.goto csymv.goto zsymv.goto \
        chemm.goto zhemm.goto \
        cherk.goto zherk.goto \
        cher2k.goto zher2k.goto \
@@ -53,7 +53,7 @@ acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \
        ssyrk.acml dsyrk.acml csyrk.acml zsyrk.acml \
        ssyr2k.acml dsyr2k.acml csyr2k.acml zsyr2k.acml \
        sger.acml dger.acml \
-       ssymv.acml dsymv.acml \
+       ssymv.acml dsymv.acml csymv.acml zsymv.acml \
        chemm.acml zhemm.acml \
        cherk.acml zherk.acml \
        cher2k.acml zher2k.acml \
@@ -71,7 +71,7 @@ atlas :: slinpack.atlas dlinpack.atlas clinpack.atlas zlinpack.atlas \
        ssyrk.atlas dsyrk.atlas csyrk.atlas zsyrk.atlas \
        ssyr2k.atlas dsyr2k.atlas csyr2k.atlas zsyr2k.atlas \
        sger.atlas dger.atlas \
-       ssymv.atlas dsymv.atlas \
+       ssymv.atlas dsymv.atlas csymv.atlas zsymv.atlas \
        chemm.acml zhemm.acml \
        chemm.atlas zhemm.atlas \
        cherk.atlas zherk.atlas \
@@ -90,7 +90,7 @@ mkl :: slinpack.mkl dlinpack.mkl clinpack.mkl zlinpack.mkl \
        ssyrk.mkl dsyrk.mkl csyrk.mkl zsyrk.mkl \
        ssyr2k.mkl dsyr2k.mkl csyr2k.mkl zsyr2k.mkl \
        sger.mkl dger.mkl \
-       ssymv.mkl dsymv.mkl \
+       ssymv.mkl dsymv.mkl csymv.mkl zsymv.mkl \
        chemm.mkl zhemm.mkl \
        cherk.mkl zherk.mkl \
        cher2k.mkl zher2k.mkl \
@@ -732,6 +732,32 @@ dsymv.atlas : dsymv.$(SUFFIX)
 dsymv.mkl : dsymv.$(SUFFIX)
 	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
 
+##################################### Csymv ####################################################
+csymv.goto : csymv.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+csymv.acml : csymv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+csymv.atlas : csymv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+csymv.mkl : csymv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Dsymv ####################################################
+zsymv.goto : zsymv.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+zsymv.acml : zsymv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+zsymv.atlas : zsymv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+zsymv.mkl : zsymv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
 ##################################### Sgeev ####################################################
 sgeev.goto : sgeev.$(SUFFIX) ../$(LIBNAME)
 	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
@@ -1037,6 +1063,12 @@ ssymv.$(SUFFIX) : symv.c
 dsymv.$(SUFFIX) : symv.c
 	$(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^
 
+csymv.$(SUFFIX) : symv.c
+	$(CC) $(CFLAGS) -c -DCOMPLEX -UDOUBLE -o $(@F) $^
+
+zsymv.$(SUFFIX) : symv.c
+	$(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^
+
 sgeev.$(SUFFIX) : geev.c
 	$(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^
 

From fb0b4552a5dd5e83434f159aaf4d44b8a0377b49 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Fri, 22 Aug 2014 10:00:09 +0200
Subject: [PATCH 13/32] added hemv benchmark

---
 benchmark/Makefile |  38 +++++++++
 benchmark/hemv.c   | 208 +++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 246 insertions(+)
 create mode 100644 benchmark/hemv.c

diff --git a/benchmark/Makefile b/benchmark/Makefile
index 3b2d263c8..6c364ddfe 100644
--- a/benchmark/Makefile
+++ b/benchmark/Makefile
@@ -36,6 +36,7 @@ goto :: slinpack.goto dlinpack.goto clinpack.goto zlinpack.goto \
        ssyr2k.goto dsyr2k.goto csyr2k.goto zsyr2k.goto \
        sger.goto dger.goto \
        ssymv.goto dsymv.goto csymv.goto zsymv.goto \
+       chemv.goto zhemv.goto \
        chemm.goto zhemm.goto \
        cherk.goto zherk.goto \
        cher2k.goto zher2k.goto \
@@ -54,6 +55,7 @@ acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \
        ssyr2k.acml dsyr2k.acml csyr2k.acml zsyr2k.acml \
        sger.acml dger.acml \
        ssymv.acml dsymv.acml csymv.acml zsymv.acml \
+       chemv.acml zhemv.acml \
        chemm.acml zhemm.acml \
        cherk.acml zherk.acml \
        cher2k.acml zher2k.acml \
@@ -72,6 +74,7 @@ atlas :: slinpack.atlas dlinpack.atlas clinpack.atlas zlinpack.atlas \
        ssyr2k.atlas dsyr2k.atlas csyr2k.atlas zsyr2k.atlas \
        sger.atlas dger.atlas \
        ssymv.atlas dsymv.atlas csymv.atlas zsymv.atlas \
+       chemv.atlas zhemv.atlas \
        chemm.acml zhemm.acml \
        chemm.atlas zhemm.atlas \
        cherk.atlas zherk.atlas \
@@ -91,6 +94,7 @@ mkl :: slinpack.mkl dlinpack.mkl clinpack.mkl zlinpack.mkl \
        ssyr2k.mkl dsyr2k.mkl csyr2k.mkl zsyr2k.mkl \
        sger.mkl dger.mkl \
        ssymv.mkl dsymv.mkl csymv.mkl zsymv.mkl \
+       chemv.mkl zhemv.mkl \
        chemm.mkl zhemm.mkl \
        cherk.mkl zherk.mkl \
        cher2k.mkl zher2k.mkl \
@@ -922,6 +926,33 @@ zpotrf.atlas : zpotrf.$(SUFFIX)
 zpotrf.mkl : zpotrf.$(SUFFIX)
 	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
 
+##################################### Chemv ####################################################
+
+chemv.goto : chemv.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+chemv.acml : chemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+chemv.atlas : chemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+chemv.mkl : chemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Zhemv ####################################################
+
+zhemv.goto : zhemv.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+zhemv.acml : zhemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+zhemv.atlas : zhemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+zhemv.mkl : zhemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
 
 ###################################################################################################
 
@@ -1105,6 +1136,13 @@ cpotrf.$(SUFFIX) : potrf.c
 zpotrf.$(SUFFIX) : potrf.c
 	$(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^
 
+chemv.$(SUFFIX) : hemv.c
+	$(CC) $(CFLAGS) -c -DCOMPLEX -UDOUBLE -o $(@F) $^
+
+zhemv.$(SUFFIX) : hemv.c
+	$(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^
+
+
 
 
 
diff --git a/benchmark/hemv.c b/benchmark/hemv.c
new file mode 100644
index 000000000..79b7679cc
--- /dev/null
+++ b/benchmark/hemv.c
@@ -0,0 +1,208 @@
+/***************************************************************************
+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 <stdio.h>
+#include <stdlib.h>
+#ifdef __CYGWIN32__
+#include <sys/time.h>
+#endif
+#include "common.h"
+
+
+#undef HEMV
+
+
+#ifdef DOUBLE
+#define HEMV   BLASFUNC(zhemv)
+#else
+#define HEMV   BLASFUNC(chemv)
+#endif
+
+
+#if defined(__WIN32__) || defined(__WIN64__)
+
+#ifndef DELTA_EPOCH_IN_MICROSECS
+#define DELTA_EPOCH_IN_MICROSECS 11644473600000000ULL
+#endif
+
+int gettimeofday(struct timeval *tv, void *tz){
+
+  FILETIME ft;
+  unsigned __int64 tmpres = 0;
+  static int tzflag;
+
+  if (NULL != tv)
+    {
+      GetSystemTimeAsFileTime(&ft);
+
+      tmpres |= ft.dwHighDateTime;
+      tmpres <<= 32;
+      tmpres |= ft.dwLowDateTime;
+
+      /*converting file time to unix epoch*/
+      tmpres /= 10;  /*convert into microseconds*/
+      tmpres -= DELTA_EPOCH_IN_MICROSECS;
+      tv->tv_sec = (long)(tmpres / 1000000UL);
+      tv->tv_usec = (long)(tmpres % 1000000UL);
+    }
+
+  return 0;
+}
+
+#endif
+
+#if !defined(__WIN32__) && !defined(__WIN64__) && !defined(__CYGWIN32__) && 0
+
+static void *huge_malloc(BLASLONG size){
+  int shmid;
+  void *address;
+
+#ifndef SHM_HUGETLB
+#define SHM_HUGETLB 04000
+#endif
+
+  if ((shmid =shmget(IPC_PRIVATE,
+		     (size + HUGE_PAGESIZE) & ~(HUGE_PAGESIZE - 1),
+		     SHM_HUGETLB | IPC_CREAT |0600)) < 0) {
+    printf( "Memory allocation failed(shmget).\n");
+    exit(1);
+  }
+
+  address = shmat(shmid, NULL, SHM_RND);
+
+  if ((BLASLONG)address == -1){
+    printf( "Memory allocation failed(shmat).\n");
+    exit(1);
+  }
+
+  shmctl(shmid, IPC_RMID, 0);
+
+  return address;
+}
+
+#define malloc huge_malloc
+
+#endif
+
+int MAIN__(int argc, char *argv[]){
+
+  FLOAT *a, *x, *y;
+  FLOAT alpha[] = {1.0, 1.0};
+  FLOAT beta [] = {1.0, 1.0};
+  char uplo='L';
+  blasint m, i, j;
+  blasint inc_x=1,inc_y=1;
+  int loops = 1;
+  int l;
+  char *p;
+
+  int from =   1;
+  int to   = 200;
+  int step =   1;
+
+  struct timeval start, stop;
+  double time1,timeg;
+
+  argc--;argv++;
+
+  if (argc > 0) { from     = atol(*argv);		argc--; argv++;}
+  if (argc > 0) { to       = MAX(atol(*argv), from);	argc--; argv++;}
+  if (argc > 0) { step     = atol(*argv);		argc--; argv++;}
+
+  if ((p = getenv("OPENBLAS_LOOPS")))  loops = atoi(p);
+  if ((p = getenv("OPENBLAS_INCX")))   inc_x = atoi(p);
+  if ((p = getenv("OPENBLAS_INCY")))   inc_y = atoi(p);
+  if ((p = getenv("OPENBLAS_UPLO")))  uplo=*p;
+
+  fprintf(stderr, "From : %3d  To : %3d Step = %3d Uplo = '%c' Inc_x = %d Inc_y = %d Loops = %d\n", from, to, step,uplo,inc_x,inc_y,loops);
+
+  if (( a = (FLOAT *)malloc(sizeof(FLOAT) * to * to * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+  if (( x = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_x) * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+  if (( y = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_y) * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+#ifdef linux
+  srandom(getpid());
+#endif
+
+  fprintf(stderr, "   SIZE       Flops\n");
+
+  for(m = from; m <= to; m += step)
+  {
+
+   timeg=0;
+
+   fprintf(stderr, " %6dx%d : ", (int)m,(int)m);
+
+   for(j = 0; j < m; j++){
+      		for(i = 0; i < m * COMPSIZE; i++){
+			a[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+      		}
+   }
+
+
+    for (l=0; l<loops; l++)
+    {
+
+   	for(i = 0; i < m * COMPSIZE * abs(inc_x); i++){
+			x[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+   	}
+
+   	for(i = 0; i < m * COMPSIZE * abs(inc_y); i++){
+			y[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+   	}
+    	gettimeofday( &start, (struct timezone *)0);
+
+    	HEMV (&uplo, &m, alpha, a, &m, x, &inc_x, beta, y, &inc_y );
+
+    	gettimeofday( &stop, (struct timezone *)0);
+
+    	time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6;
+
+	timeg += time1;
+
+    }
+
+    timeg /= loops;
+
+    fprintf(stderr,
+	    " %10.2f MFlops\n",
+	    COMPSIZE * COMPSIZE * 2. * (double)m * (double)m / timeg * 1.e-6);
+
+  }
+
+  return 0;
+}
+
+void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));

From e4472ad8507ed24eb4fe67532a55ed773253b8ee Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Fri, 22 Aug 2014 11:42:07 +0200
Subject: [PATCH 14/32] added sdot and ddot benchmarks

---
 benchmark/Makefile |  37 +++++++++
 benchmark/dot.c    | 195 +++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 232 insertions(+)
 create mode 100644 benchmark/dot.c

diff --git a/benchmark/Makefile b/benchmark/Makefile
index 6c364ddfe..1bc631aa4 100644
--- a/benchmark/Makefile
+++ b/benchmark/Makefile
@@ -35,6 +35,7 @@ goto :: slinpack.goto dlinpack.goto clinpack.goto zlinpack.goto \
        ssyrk.goto dsyrk.goto csyrk.goto zsyrk.goto \
        ssyr2k.goto dsyr2k.goto csyr2k.goto zsyr2k.goto \
        sger.goto dger.goto \
+       sdot.goto ddot.goto \
        ssymv.goto dsymv.goto csymv.goto zsymv.goto \
        chemv.goto zhemv.goto \
        chemm.goto zhemm.goto \
@@ -54,6 +55,7 @@ acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \
        ssyrk.acml dsyrk.acml csyrk.acml zsyrk.acml \
        ssyr2k.acml dsyr2k.acml csyr2k.acml zsyr2k.acml \
        sger.acml dger.acml \
+       sdot.acml ddot.acml \
        ssymv.acml dsymv.acml csymv.acml zsymv.acml \
        chemv.acml zhemv.acml \
        chemm.acml zhemm.acml \
@@ -73,6 +75,7 @@ atlas :: slinpack.atlas dlinpack.atlas clinpack.atlas zlinpack.atlas \
        ssyrk.atlas dsyrk.atlas csyrk.atlas zsyrk.atlas \
        ssyr2k.atlas dsyr2k.atlas csyr2k.atlas zsyr2k.atlas \
        sger.atlas dger.atlas \
+       sdot.atlas ddot.atlas \
        ssymv.atlas dsymv.atlas csymv.atlas zsymv.atlas \
        chemv.atlas zhemv.atlas \
        chemm.acml zhemm.acml \
@@ -93,6 +96,7 @@ mkl :: slinpack.mkl dlinpack.mkl clinpack.mkl zlinpack.mkl \
        ssyrk.mkl dsyrk.mkl csyrk.mkl zsyrk.mkl \
        ssyr2k.mkl dsyr2k.mkl csyr2k.mkl zsyr2k.mkl \
        sger.mkl dger.mkl \
+       sdot.mkl ddot.mkl \
        ssymv.mkl dsymv.mkl csymv.mkl zsymv.mkl \
        chemv.mkl zhemv.mkl \
        chemm.mkl zhemm.mkl \
@@ -954,6 +958,33 @@ zhemv.atlas : zhemv.$(SUFFIX)
 zhemv.mkl : zhemv.$(SUFFIX)
 	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
 
+##################################### Sdot ####################################################
+sdot.goto : sdot.$(SUFFIX) ../$(LIBNAME)
+        $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+sdot.acml : sdot.$(SUFFIX)
+        -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+sdot.atlas : sdot.$(SUFFIX)
+        -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+sdot.mkl : sdot.$(SUFFIX)
+        -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Ddot ####################################################
+ddot.goto : ddot.$(SUFFIX) ../$(LIBNAME)
+        $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+ddot.acml : ddot.$(SUFFIX)
+        -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+ddot.atlas : ddot.$(SUFFIX)
+        -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+ddot.mkl : ddot.$(SUFFIX)
+        -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+
 ###################################################################################################
 
 slinpack.$(SUFFIX) : linpack.c
@@ -1142,6 +1173,12 @@ chemv.$(SUFFIX) : hemv.c
 zhemv.$(SUFFIX) : hemv.c
 	$(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^
 
+sdot.$(SUFFIX) : dot.c
+        $(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^
+
+ddot.$(SUFFIX) : dot.c
+        $(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^
+
 
 
 
diff --git a/benchmark/dot.c b/benchmark/dot.c
new file mode 100644
index 000000000..6132ed324
--- /dev/null
+++ b/benchmark/dot.c
@@ -0,0 +1,195 @@
+/***************************************************************************
+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 <stdio.h>
+#include <stdlib.h>
+#ifdef __CYGWIN32__
+#include <sys/time.h>
+#endif
+#include "common.h"
+
+
+#undef DOT
+
+
+#ifdef DOUBLE
+#define DOT   BLASFUNC(ddot)
+#else
+#define DOT   BLASFUNC(sdot)
+#endif
+
+
+#if defined(__WIN32__) || defined(__WIN64__)
+
+#ifndef DELTA_EPOCH_IN_MICROSECS
+#define DELTA_EPOCH_IN_MICROSECS 11644473600000000ULL
+#endif
+
+int gettimeofday(struct timeval *tv, void *tz){
+
+  FILETIME ft;
+  unsigned __int64 tmpres = 0;
+  static int tzflag;
+
+  if (NULL != tv)
+    {
+      GetSystemTimeAsFileTime(&ft);
+
+      tmpres |= ft.dwHighDateTime;
+      tmpres <<= 32;
+      tmpres |= ft.dwLowDateTime;
+
+      /*converting file time to unix epoch*/
+      tmpres /= 10;  /*convert into microseconds*/
+      tmpres -= DELTA_EPOCH_IN_MICROSECS;
+      tv->tv_sec = (long)(tmpres / 1000000UL);
+      tv->tv_usec = (long)(tmpres % 1000000UL);
+    }
+
+  return 0;
+}
+
+#endif
+
+#if !defined(__WIN32__) && !defined(__WIN64__) && !defined(__CYGWIN32__) && 0
+
+static void *huge_malloc(BLASLONG size){
+  int shmid;
+  void *address;
+
+#ifndef SHM_HUGETLB
+#define SHM_HUGETLB 04000
+#endif
+
+  if ((shmid =shmget(IPC_PRIVATE,
+		     (size + HUGE_PAGESIZE) & ~(HUGE_PAGESIZE - 1),
+		     SHM_HUGETLB | IPC_CREAT |0600)) < 0) {
+    printf( "Memory allocation failed(shmget).\n");
+    exit(1);
+  }
+
+  address = shmat(shmid, NULL, SHM_RND);
+
+  if ((BLASLONG)address == -1){
+    printf( "Memory allocation failed(shmat).\n");
+    exit(1);
+  }
+
+  shmctl(shmid, IPC_RMID, 0);
+
+  return address;
+}
+
+#define malloc huge_malloc
+
+#endif
+
+int MAIN__(int argc, char *argv[]){
+
+  FLOAT *x, *y;
+  FLOAT result;
+  blasint m, i;
+  blasint inc_x=1,inc_y=1;
+  int loops = 1;
+  int l;
+  char *p;
+
+  int from =   1;
+  int to   = 200;
+  int step =   1;
+
+  struct timeval start, stop;
+  double time1,timeg;
+
+  argc--;argv++;
+
+  if (argc > 0) { from     = atol(*argv);		argc--; argv++;}
+  if (argc > 0) { to       = MAX(atol(*argv), from);	argc--; argv++;}
+  if (argc > 0) { step     = atol(*argv);		argc--; argv++;}
+
+  if ((p = getenv("OPENBLAS_LOOPS")))  loops = atoi(p);
+  if ((p = getenv("OPENBLAS_INCX")))   inc_x = atoi(p);
+  if ((p = getenv("OPENBLAS_INCY")))   inc_y = atoi(p);
+
+  fprintf(stderr, "From : %3d  To : %3d Step = %3d Inc_x = %d Inc_y = %d Loops = %d\n", from, to, step,inc_x,inc_y,loops);
+
+  if (( x = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_x) * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+  if (( y = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_y) * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+#ifdef linux
+  srandom(getpid());
+#endif
+
+  fprintf(stderr, "   SIZE       Flops\n");
+
+  for(m = from; m <= to; m += step)
+  {
+
+   timeg=0;
+
+   fprintf(stderr, " %6d : ", (int)m);
+
+
+   for (l=0; l<loops; l++)
+   {
+
+   	for(i = 0; i < m * COMPSIZE * abs(inc_x); i++){
+			x[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+   	}
+
+   	for(i = 0; i < m * COMPSIZE * abs(inc_y); i++){
+			y[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+   	}
+    	gettimeofday( &start, (struct timezone *)0);
+
+    	result = DOT (&m, x, &inc_x, y, &inc_y );
+
+    	gettimeofday( &stop, (struct timezone *)0);
+
+    	time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6;
+
+	timeg += time1;
+
+    }
+
+    timeg /= loops;
+
+    fprintf(stderr,
+	    " %10.2f MFlops\n",
+	    COMPSIZE * COMPSIZE * 2. * (double)m / timeg * 1.e-6);
+
+  }
+
+  return 0;
+}
+
+void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));

From 7e404de3de19cd7652325f41fe5dcc7a70456144 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Fri, 22 Aug 2014 11:51:30 +0200
Subject: [PATCH 15/32] bugfix in Makefile

---
 benchmark/Makefile | 20 ++++++++++----------
 1 file changed, 10 insertions(+), 10 deletions(-)

diff --git a/benchmark/Makefile b/benchmark/Makefile
index 1bc631aa4..679a98c32 100644
--- a/benchmark/Makefile
+++ b/benchmark/Makefile
@@ -960,29 +960,29 @@ zhemv.mkl : zhemv.$(SUFFIX)
 
 ##################################### Sdot ####################################################
 sdot.goto : sdot.$(SUFFIX) ../$(LIBNAME)
-        $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
 
 sdot.acml : sdot.$(SUFFIX)
-        -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
 
 sdot.atlas : sdot.$(SUFFIX)
-        -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
 
 sdot.mkl : sdot.$(SUFFIX)
-        -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
 
 ##################################### Ddot ####################################################
 ddot.goto : ddot.$(SUFFIX) ../$(LIBNAME)
-        $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
 
 ddot.acml : ddot.$(SUFFIX)
-        -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
 
 ddot.atlas : ddot.$(SUFFIX)
-        -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
 
 ddot.mkl : ddot.$(SUFFIX)
-        -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
 
 
 ###################################################################################################
@@ -1174,10 +1174,10 @@ zhemv.$(SUFFIX) : hemv.c
 	$(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^
 
 sdot.$(SUFFIX) : dot.c
-        $(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^
+	$(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^
 
 ddot.$(SUFFIX) : dot.c
-        $(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^
+	$(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^
 
 
 

From 8a9e868919c50c407057f1e79b5a4c94a24e54e8 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Fri, 22 Aug 2014 14:29:17 +0200
Subject: [PATCH 16/32] added optimized sdot for bulldozer

---
 kernel/x86_64/sdot.c                    | 107 ++++++++++++++++++++++++
 kernel/x86_64/sdot_microk_bulldozer-2.c |  85 +++++++++++++++++++
 2 files changed, 192 insertions(+)
 create mode 100644 kernel/x86_64/sdot.c
 create mode 100644 kernel/x86_64/sdot_microk_bulldozer-2.c

diff --git a/kernel/x86_64/sdot.c b/kernel/x86_64/sdot.c
new file mode 100644
index 000000000..8c60b954a
--- /dev/null
+++ b/kernel/x86_64/sdot.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.
+*****************************************************************************/
+
+
+#include "common.h"
+
+#if defined(BULLDOZER) || defined(PILEDRIVER)
+#include "sdot_microk_bulldozer-2.c"
+#endif
+
+
+#ifndef HAVE_KERNEL_16
+
+void sdot_kernel_16(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *d)
+{
+	BLASLONG register i = 0;
+	FLOAT dot = 0.0;
+
+	while(i < n)
+        {
+              dot += y[i]  * x[i]
+                  + y[i+1] * x[i+1]
+                  + y[i+2] * x[i+2]
+                  + y[i+3] * x[i+3]
+                  + y[i+4] * x[i+4]
+                  + y[i+5] * x[i+5]
+                  + y[i+6] * x[i+6]
+                  + y[i+7] * x[i+7] ;
+
+              i+=8 ;
+
+       }
+       *d += dot;
+
+}
+
+#endif
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
+{
+	BLASLONG i=0;
+	BLASLONG ix=0,iy=0;
+
+	FLOAT  dot = 0.0 ;
+
+	if ( n <= 0 )  return(dot);
+
+	if ( (inc_x == 1) && (inc_y == 1) )
+	{
+
+		int n1 = n & -16;
+
+		if ( n1 )
+			sdot_kernel_16(n1, x, y , &dot );
+
+
+		i = n1;
+		while(i < n)
+		{
+
+			dot += y[i] * x[i] ;
+			i++ ;
+
+		}
+		return(dot);
+
+
+	}
+
+	while(i < n)
+	{
+
+		dot += y[iy] * x[ix] ;
+		ix  += inc_x ;
+		iy  += inc_y ;
+		i++ ;
+
+	}
+	return(dot);
+
+}
+
+
diff --git a/kernel/x86_64/sdot_microk_bulldozer-2.c b/kernel/x86_64/sdot_microk_bulldozer-2.c
new file mode 100644
index 000000000..024b2ce6d
--- /dev/null
+++ b/kernel/x86_64/sdot_microk_bulldozer-2.c
@@ -0,0 +1,85 @@
+/***************************************************************************
+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_16 1
+static void sdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *dot) __attribute__ ((noinline));
+
+static void sdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
+{
+
+
+	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
+        "vmovups                16(%2,%0,4), %%xmm13         \n\t"  // 4 * x
+        "vmovups                32(%2,%0,4), %%xmm14         \n\t"  // 4 * x
+        "vmovups                48(%2,%0,4), %%xmm15         \n\t"  // 4 * x
+
+	"vfmaddps       %%xmm4,   (%3,%0,4), %%xmm12, %%xmm4 \n\t"  // 4 * y
+	"vfmaddps       %%xmm5, 16(%3,%0,4), %%xmm13, %%xmm5 \n\t"  // 4 * y
+	"vfmaddps       %%xmm6, 32(%3,%0,4), %%xmm14, %%xmm6 \n\t"  // 4 * y
+	"vfmaddps       %%xmm7, 48(%3,%0,4), %%xmm15, %%xmm7 \n\t"  // 4 * y
+
+	"addq		$16, %0	  	 	             \n\t"
+	"subq	        $16, %1			             \n\t"		
+	"jnz		.L01LOOP%=		             \n\t"
+
+	"vaddps        %%xmm4, %%xmm5, %%xmm4	\n\t"
+	"vaddps        %%xmm6, %%xmm7, %%xmm6	\n\t"
+	"vaddps        %%xmm4, %%xmm6, %%xmm4	\n\t"
+
+	"vhaddps        %%xmm4, %%xmm4, %%xmm4	\n\t"
+	"vhaddps        %%xmm4, %%xmm4, %%xmm4	\n\t"
+
+	"vmovss		%%xmm4,    (%4)		\n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (dot)     // 4
+	: "cc", 
+	  "%xmm4", "%xmm5", 
+	  "%xmm6", "%xmm7", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From 5d97b0754c3c1ef975365410b37d9cda360b73eb Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Fri, 22 Aug 2014 17:00:26 +0200
Subject: [PATCH 17/32] added optimized sdot kernel for nehalem

---
 kernel/x86_64/KERNEL.NEHALEM          |  2 +
 kernel/x86_64/sdot.c                  |  2 +
 kernel/x86_64/sdot_microk_nehalem-2.c | 94 +++++++++++++++++++++++++++
 3 files changed, 98 insertions(+)
 create mode 100644 kernel/x86_64/sdot_microk_nehalem-2.c

diff --git a/kernel/x86_64/KERNEL.NEHALEM b/kernel/x86_64/KERNEL.NEHALEM
index b16fd9c49..aee622d0d 100644
--- a/kernel/x86_64/KERNEL.NEHALEM
+++ b/kernel/x86_64/KERNEL.NEHALEM
@@ -1,3 +1,5 @@
+SDOTKERNEL     = sdot.c
+
 DSYMV_U_KERNEL = dsymv_U.c
 DSYMV_L_KERNEL = dsymv_L.c
 SSYMV_U_KERNEL = ssymv_U.c
diff --git a/kernel/x86_64/sdot.c b/kernel/x86_64/sdot.c
index 8c60b954a..a13d65d25 100644
--- a/kernel/x86_64/sdot.c
+++ b/kernel/x86_64/sdot.c
@@ -30,6 +30,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #if defined(BULLDOZER) || defined(PILEDRIVER)
 #include "sdot_microk_bulldozer-2.c"
+#elif defined(NEHALEM)
+#include "sdot_microk_nehalem-2.c"
 #endif
 
 
diff --git a/kernel/x86_64/sdot_microk_nehalem-2.c b/kernel/x86_64/sdot_microk_nehalem-2.c
new file mode 100644
index 000000000..3548ace88
--- /dev/null
+++ b/kernel/x86_64/sdot_microk_nehalem-2.c
@@ -0,0 +1,94 @@
+/***************************************************************************
+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_16 1
+static void sdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *dot) __attribute__ ((noinline));
+
+static void sdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
+{
+
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"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"
+        "movups                  (%2,%0,4), %%xmm12         \n\t"  // 4 * x
+        "movups                  (%3,%0,4), %%xmm8          \n\t"  // 4 * x
+        "movups                16(%2,%0,4), %%xmm13         \n\t"  // 4 * x
+        "movups                16(%3,%0,4), %%xmm9          \n\t"  // 4 * x
+        "movups                32(%2,%0,4), %%xmm14         \n\t"  // 4 * x
+        "movups                32(%3,%0,4), %%xmm10         \n\t"  // 4 * x
+        "movups                48(%2,%0,4), %%xmm15         \n\t"  // 4 * x
+        "movups                48(%3,%0,4), %%xmm11         \n\t"  // 4 * x
+
+	"mulps          %%xmm8 , %%xmm12 		    \n\t"  
+	"mulps          %%xmm9 , %%xmm13 		    \n\t"  
+	"mulps          %%xmm10, %%xmm14 		    \n\t"  
+	"mulps          %%xmm11, %%xmm15 		    \n\t"  
+
+	"addps		%%xmm12, %%xmm4   \n\t"
+	"addps		%%xmm13, %%xmm5   \n\t"
+	"addps		%%xmm14, %%xmm6   \n\t"
+	"addps		%%xmm15, %%xmm7   \n\t"
+
+	"addq		$16, %0	  	 	             \n\t"
+	"subq	        $16, %1			             \n\t"		
+	"jnz		.L01LOOP%=		             \n\t"
+
+	"addps        %%xmm5, %%xmm4	\n\t"
+	"addps        %%xmm7, %%xmm6	\n\t"
+	"addps        %%xmm6, %%xmm4	\n\t"
+
+	"haddps        %%xmm4, %%xmm4	\n\t"
+	"haddps        %%xmm4, %%xmm4	\n\t"
+
+	"movss	       %%xmm4,    (%4)	\n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (dot)     // 4
+	: "cc", 
+	  "%xmm4", "%xmm5", 
+	  "%xmm6", "%xmm7", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From 95a707ced3a332221d7539e51e0051a9d07edd91 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Fri, 22 Aug 2014 17:01:27 +0200
Subject: [PATCH 18/32] update of KERNEL.BULLDOZER

---
 kernel/x86_64/KERNEL.BULLDOZER | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/kernel/x86_64/KERNEL.BULLDOZER b/kernel/x86_64/KERNEL.BULLDOZER
index 3ee1978b8..7aa597ea0 100644
--- a/kernel/x86_64/KERNEL.BULLDOZER
+++ b/kernel/x86_64/KERNEL.BULLDOZER
@@ -1,3 +1,5 @@
+SDOTKERNEL     = sdot.c
+
 DSYMV_U_KERNEL = dsymv_U.c
 DSYMV_L_KERNEL = dsymv_L.c
 SSYMV_U_KERNEL = ssymv_U.c

From 53ec5789e2aeeb4bc5eee2de65047ec3a8169c8a Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Fri, 22 Aug 2014 17:02:55 +0200
Subject: [PATCH 19/32] bugfix for Makefile

---
 benchmark/Makefile | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/benchmark/Makefile b/benchmark/Makefile
index 679a98c32..933b4bd44 100644
--- a/benchmark/Makefile
+++ b/benchmark/Makefile
@@ -108,7 +108,7 @@ mkl :: slinpack.mkl dlinpack.mkl clinpack.mkl zlinpack.mkl \
        spotrf.mkl dpotrf.mkl cpotrf.mkl zpotrf.mkl \
        ssymm.mkl dsymm.mkl csymm.mkl zsymm.mkl
 
-all :: goto atlas acml mkl
+all :: goto mkl atlas acml
 
 ##################################### Slinpack ####################################################
 slinpack.goto : slinpack.$(SUFFIX) ../$(LIBNAME)

From 16d6be852dbcb09365fd61e1e11b3cac66901b76 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Fri, 22 Aug 2014 20:34:41 +0200
Subject: [PATCH 20/32] added optimized ddot kernel for nehalem

---
 kernel/x86_64/KERNEL.NEHALEM          |   1 +
 kernel/x86_64/ddot.c                  | 110 ++++++++++++++++++++++++++
 kernel/x86_64/ddot_microk_nehalem-2.c |  94 ++++++++++++++++++++++
 kernel/x86_64/sdot_microk_nehalem-2.c |   4 +-
 4 files changed, 207 insertions(+), 2 deletions(-)
 create mode 100644 kernel/x86_64/ddot.c
 create mode 100644 kernel/x86_64/ddot_microk_nehalem-2.c

diff --git a/kernel/x86_64/KERNEL.NEHALEM b/kernel/x86_64/KERNEL.NEHALEM
index aee622d0d..e0f3c3336 100644
--- a/kernel/x86_64/KERNEL.NEHALEM
+++ b/kernel/x86_64/KERNEL.NEHALEM
@@ -1,4 +1,5 @@
 SDOTKERNEL     = sdot.c
+DDOTKERNEL     = ddot.c
 
 DSYMV_U_KERNEL = dsymv_U.c
 DSYMV_L_KERNEL = dsymv_L.c
diff --git a/kernel/x86_64/ddot.c b/kernel/x86_64/ddot.c
new file mode 100644
index 000000000..ee6785f6a
--- /dev/null
+++ b/kernel/x86_64/ddot.c
@@ -0,0 +1,110 @@
+/***************************************************************************
+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 "ddot_microk_bulldozer-2.c"
+#elif defined(NEHALEM)
+#include "ddot_microk_nehalem-2.c"
+#endif
+
+
+#ifndef HAVE_KERNEL_8
+
+void ddot_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *d)
+{
+	BLASLONG register i = 0;
+	FLOAT dot = 0.0;
+
+	while(i < n)
+        {
+              dot += y[i]  * x[i]
+                  + y[i+1] * x[i+1]
+                  + y[i+2] * x[i+2]
+                  + y[i+3] * x[i+3]
+                  + y[i+4] * x[i+4]
+                  + y[i+5] * x[i+5]
+                  + y[i+6] * x[i+6]
+                  + y[i+7] * x[i+7] ;
+
+              i+=8 ;
+
+       }
+       *d += dot;
+
+}
+
+#endif
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
+{
+	BLASLONG i=0;
+	BLASLONG ix=0,iy=0;
+
+	FLOAT  dot = 0.0 ;
+
+	if ( n <= 0 )  return(dot);
+
+	if ( (inc_x == 1) && (inc_y == 1) )
+	{
+
+		int n1 = n & -8;
+
+		if ( n1 )
+			ddot_kernel_8(n1, x, y , &dot );
+
+
+		i = n1;
+		while(i < n)
+		{
+
+			dot += y[i] * x[i] ;
+			i++ ;
+
+		}
+		return(dot);
+
+
+	}
+
+	while(i < n)
+	{
+
+		dot += y[iy] * x[ix] ;
+		ix  += inc_x ;
+		iy  += inc_y ;
+		i++ ;
+
+	}
+	return(dot);
+
+}
+
+
diff --git a/kernel/x86_64/ddot_microk_nehalem-2.c b/kernel/x86_64/ddot_microk_nehalem-2.c
new file mode 100644
index 000000000..dd05053f7
--- /dev/null
+++ b/kernel/x86_64/ddot_microk_nehalem-2.c
@@ -0,0 +1,94 @@
+/***************************************************************************
+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_8 1
+static void ddot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *dot) __attribute__ ((noinline));
+
+static void ddot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
+{
+
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"xorpd		%%xmm4, %%xmm4	             \n\t"
+	"xorpd		%%xmm5, %%xmm5	             \n\t"
+	"xorpd		%%xmm6, %%xmm6	             \n\t"
+	"xorpd		%%xmm7, %%xmm7	             \n\t"
+
+	".align 16				            \n\t"
+	".L01LOOP%=:				            \n\t"
+
+        "movups                  (%2,%0,8), %%xmm12         \n\t"  // 2 * x
+        "movups                  (%3,%0,8), %%xmm8          \n\t"  // 2 * y
+        "movups                16(%2,%0,8), %%xmm13         \n\t"  // 2 * x
+        "movups                16(%3,%0,8), %%xmm9          \n\t"  // 2 * y
+        "movups                32(%2,%0,8), %%xmm14         \n\t"  // 2 * x
+        "movups                32(%3,%0,8), %%xmm10         \n\t"  // 2 * y
+        "movups                48(%2,%0,8), %%xmm15         \n\t"  // 2 * x
+        "movups                48(%3,%0,8), %%xmm11         \n\t"  // 2 * y
+
+	"mulpd          %%xmm8 , %%xmm12 		    \n\t"  
+	"mulpd          %%xmm9 , %%xmm13 		    \n\t"  
+	"mulpd          %%xmm10, %%xmm14 		    \n\t"  
+	"mulpd          %%xmm11, %%xmm15 		    \n\t"  
+
+	"addpd		%%xmm12, %%xmm4   \n\t"
+	"addpd		%%xmm13, %%xmm5   \n\t"
+	"addpd		%%xmm14, %%xmm6   \n\t"
+	"addpd		%%xmm15, %%xmm7   \n\t"
+
+	"addq		$8 , %0	  	 	             \n\t"
+	"subq	        $8 , %1			             \n\t"		
+	"jnz		.L01LOOP%=		             \n\t"
+
+	"addpd        %%xmm5, %%xmm4	\n\t"
+	"addpd        %%xmm7, %%xmm6	\n\t"
+	"addpd        %%xmm6, %%xmm4	\n\t"
+
+	"haddpd        %%xmm4, %%xmm4	\n\t"
+
+	"movsd	       %%xmm4,    (%4)	\n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (dot)     // 4
+	: "cc",  
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
diff --git a/kernel/x86_64/sdot_microk_nehalem-2.c b/kernel/x86_64/sdot_microk_nehalem-2.c
index 3548ace88..2a918b5ea 100644
--- a/kernel/x86_64/sdot_microk_nehalem-2.c
+++ b/kernel/x86_64/sdot_microk_nehalem-2.c
@@ -83,8 +83,8 @@ static void sdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
           "r" (y),      // 3
           "r" (dot)     // 4
 	: "cc", 
-	  "%xmm4", "%xmm5", 
-	  "%xmm6", "%xmm7", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
 	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
 	  "memory"
 	);

From f1b96c4846f9fd110b65335600445fae50fb1d26 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Fri, 22 Aug 2014 21:19:29 +0200
Subject: [PATCH 21/32] added optimized ddot kernel for bulldozer

---
 kernel/x86_64/KERNEL.BULLDOZER          |  1 +
 kernel/x86_64/ddot_microk_bulldozer-2.c | 84 +++++++++++++++++++++++++
 2 files changed, 85 insertions(+)
 create mode 100644 kernel/x86_64/ddot_microk_bulldozer-2.c

diff --git a/kernel/x86_64/KERNEL.BULLDOZER b/kernel/x86_64/KERNEL.BULLDOZER
index 7aa597ea0..9e6fc2e28 100644
--- a/kernel/x86_64/KERNEL.BULLDOZER
+++ b/kernel/x86_64/KERNEL.BULLDOZER
@@ -1,4 +1,5 @@
 SDOTKERNEL     = sdot.c
+DDOTKERNEL     = ddot.c
 
 DSYMV_U_KERNEL = dsymv_U.c
 DSYMV_L_KERNEL = dsymv_L.c
diff --git a/kernel/x86_64/ddot_microk_bulldozer-2.c b/kernel/x86_64/ddot_microk_bulldozer-2.c
new file mode 100644
index 000000000..0c77b6349
--- /dev/null
+++ b/kernel/x86_64/ddot_microk_bulldozer-2.c
@@ -0,0 +1,84 @@
+/***************************************************************************
+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_8 1
+static void ddot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *dot) __attribute__ ((noinline));
+
+static void ddot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
+{
+
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vxorpd		%%xmm4, %%xmm4, %%xmm4	             \n\t"
+	"vxorpd		%%xmm5, %%xmm5, %%xmm5	             \n\t"
+	"vxorpd		%%xmm6, %%xmm6, %%xmm6	             \n\t"
+	"vxorpd		%%xmm7, %%xmm7, %%xmm7	             \n\t"
+
+	".align 16				             \n\t"
+	".L01LOOP%=:				             \n\t"
+        "vmovups                  (%2,%0,8), %%xmm12         \n\t"  // 2 * x
+        "vmovups                16(%2,%0,8), %%xmm13         \n\t"  // 2 * x
+        "vmovups                32(%2,%0,8), %%xmm14         \n\t"  // 2 * x
+        "vmovups                48(%2,%0,8), %%xmm15         \n\t"  // 2 * x
+
+	"vfmaddpd       %%xmm4,   (%3,%0,8), %%xmm12, %%xmm4 \n\t"  // 2 * y
+	"vfmaddpd       %%xmm5, 16(%3,%0,8), %%xmm13, %%xmm5 \n\t"  // 2 * y
+	"vfmaddpd       %%xmm6, 32(%3,%0,8), %%xmm14, %%xmm6 \n\t"  // 2 * y
+	"vfmaddpd       %%xmm7, 48(%3,%0,8), %%xmm15, %%xmm7 \n\t"  // 2 * y
+
+	"addq		$8 , %0	  	 	             \n\t"
+	"subq	        $8 , %1			             \n\t"		
+	"jnz		.L01LOOP%=		             \n\t"
+
+	"vaddpd        %%xmm4, %%xmm5, %%xmm4	\n\t"
+	"vaddpd        %%xmm6, %%xmm7, %%xmm6	\n\t"
+	"vaddpd        %%xmm4, %%xmm6, %%xmm4	\n\t"
+
+	"vhaddpd        %%xmm4, %%xmm4, %%xmm4	\n\t"
+
+	"vmovsd		%%xmm4,    (%4)		\n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (dot)     // 4
+	: "cc", 
+	  "%xmm4", "%xmm5", 
+	  "%xmm6", "%xmm7", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From ac76b6267f1d7de2cbbc0b7fb22c825281c36b44 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Sat, 23 Aug 2014 10:40:57 +0200
Subject: [PATCH 22/32] added optimized dgemv_n kernel for nehalem

---
 kernel/x86_64/KERNEL.NEHALEM             |   1 +
 kernel/x86_64/dgemv_n.c                  |   2 +
 kernel/x86_64/dgemv_n_microk_nehalem-2.c | 137 +++++++++++++++++++++++
 3 files changed, 140 insertions(+)
 create mode 100644 kernel/x86_64/dgemv_n_microk_nehalem-2.c

diff --git a/kernel/x86_64/KERNEL.NEHALEM b/kernel/x86_64/KERNEL.NEHALEM
index e0f3c3336..5523b4571 100644
--- a/kernel/x86_64/KERNEL.NEHALEM
+++ b/kernel/x86_64/KERNEL.NEHALEM
@@ -8,6 +8,7 @@ SSYMV_L_KERNEL = ssymv_L.c
 
 SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t.c
+DGEMVNKERNEL = dgemv_n.c
 
 SGEMMKERNEL    =  gemm_kernel_4x8_nehalem.S
 SGEMMINCOPY    =  gemm_ncopy_4.S
diff --git a/kernel/x86_64/dgemv_n.c b/kernel/x86_64/dgemv_n.c
index 5d826dc63..cecb8d3fc 100644
--- a/kernel/x86_64/dgemv_n.c
+++ b/kernel/x86_64/dgemv_n.c
@@ -31,6 +31,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #if defined(HASWELL)
 #include "dgemv_n_microk_haswell-2.c"
+#elif defined(NEHALEM)
+#include "dgemv_n_microk_nehalem-2.c"
 #endif
 
 
diff --git a/kernel/x86_64/dgemv_n_microk_nehalem-2.c b/kernel/x86_64/dgemv_n_microk_nehalem-2.c
new file mode 100644
index 000000000..84b82f805
--- /dev/null
+++ b/kernel/x86_64/dgemv_n_microk_nehalem-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 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__
+	(
+	"movddup    (%2), %%xmm12	 \n\t"	// x0 
+	"movddup   8(%2), %%xmm13	 \n\t"	// x1 
+	"movddup  16(%2), %%xmm14	 \n\t"	// x2 
+	"movddup  24(%2), %%xmm15	 \n\t"	// x3 
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"prefetcht0	 192(%3,%0,8)		 \n\t"
+	"movups	       (%3,%0,8), %%xmm4	 \n\t"	// 2 * y
+	"movups      16(%3,%0,8), %%xmm5	 \n\t"	// 2 * y
+	"movups      32(%3,%0,8), %%xmm6	 \n\t"	// 2 * y
+	"movups      48(%3,%0,8), %%xmm7	 \n\t"	// 2 * y
+	"movups	       (%4,%0,8), %%xmm8	 \n\t"	// 2 * a
+	"movups      16(%4,%0,8), %%xmm9	 \n\t"	// 2 * a
+	"movups      32(%4,%0,8), %%xmm10        \n\t"	// 2 * a
+	"movups      48(%4,%0,8), %%xmm11        \n\t"  // 2 * a
+
+	"prefetcht0	 192(%4,%0,8)		       \n\t"
+	"mulpd           %%xmm12 , %%xmm8	       \n\t" // a * x
+	"mulpd           %%xmm12 , %%xmm9	       \n\t" // a * x
+	"mulpd           %%xmm12 , %%xmm10             \n\t" // a * x
+	"mulpd           %%xmm12 , %%xmm11             \n\t" // a * x
+	"addpd		 %%xmm8  , %%xmm4              \n\t" // y += a * x
+	"addpd		 %%xmm9  , %%xmm5              \n\t" // y += a * x
+	"addpd		 %%xmm10 , %%xmm6              \n\t" // y += a * x
+	"addpd		 %%xmm11 , %%xmm7              \n\t" // y += a * x
+
+	"prefetcht0	 192(%5,%0,8)		       \n\t"
+	"movups	       (%5,%0,8), %%xmm8	       \n\t" // 2 * a
+	"movups      16(%5,%0,8), %%xmm9	       \n\t" // 2 * a
+	"movups      32(%5,%0,8), %%xmm10              \n\t" // 2 * a
+	"movups      48(%5,%0,8), %%xmm11              \n\t" // 2 * a
+	"mulpd           %%xmm13 , %%xmm8	       \n\t" // a * x
+	"mulpd           %%xmm13 , %%xmm9	       \n\t" // a * x
+	"mulpd           %%xmm13 , %%xmm10             \n\t" // a * x
+	"mulpd           %%xmm13 , %%xmm11             \n\t" // a * x
+	"addpd		 %%xmm8  , %%xmm4              \n\t" // y += a * x
+	"addpd		 %%xmm9  , %%xmm5              \n\t" // y += a * x
+	"addpd		 %%xmm10 , %%xmm6              \n\t" // y += a * x
+	"addpd		 %%xmm11 , %%xmm7              \n\t" // y += a * x
+
+	"prefetcht0	 192(%6,%0,8)		       \n\t"
+	"movups	       (%6,%0,8), %%xmm8	       \n\t" // 2 * a
+	"movups      16(%6,%0,8), %%xmm9	       \n\t" // 2 * a
+	"movups      32(%6,%0,8), %%xmm10              \n\t" // 2 * a
+	"movups      48(%6,%0,8), %%xmm11              \n\t" // 2 * a
+	"mulpd           %%xmm14 , %%xmm8	       \n\t" // a * x
+	"mulpd           %%xmm14 , %%xmm9	       \n\t" // a * x
+	"mulpd           %%xmm14 , %%xmm10             \n\t" // a * x
+	"mulpd           %%xmm14 , %%xmm11             \n\t" // a * x
+	"addpd		 %%xmm8  , %%xmm4              \n\t" // y += a * x
+	"addpd		 %%xmm9  , %%xmm5              \n\t" // y += a * x
+	"addpd		 %%xmm10 , %%xmm6              \n\t" // y += a * x
+	"addpd		 %%xmm11 , %%xmm7              \n\t" // y += a * x
+
+	"prefetcht0	 192(%7,%0,8)		       \n\t"
+	"movups	       (%7,%0,8), %%xmm8	       \n\t" // 2 * a
+	"movups      16(%7,%0,8), %%xmm9	       \n\t" // 2 * a
+	"movups      32(%7,%0,8), %%xmm10              \n\t" // 2 * a
+	"movups      48(%7,%0,8), %%xmm11              \n\t" // 2 * a
+	"mulpd           %%xmm15 , %%xmm8	       \n\t" // a * x
+	"mulpd           %%xmm15 , %%xmm9	       \n\t" // a * x
+	"mulpd           %%xmm15 , %%xmm10             \n\t" // a * x
+	"mulpd           %%xmm15 , %%xmm11             \n\t" // a * x
+	"addpd		 %%xmm8  , %%xmm4              \n\t" // y += a * x
+	"addpd		 %%xmm9  , %%xmm5              \n\t" // y += a * x
+	"addpd		 %%xmm10 , %%xmm6              \n\t" // y += a * x
+	"addpd		 %%xmm11 , %%xmm7              \n\t" // y += a * x
+
+	"movups  %%xmm4,   (%3,%0,8)		      \n\t"	// 4 * y
+	"movups  %%xmm5, 16(%3,%0,8)		      \n\t"	// 4 * y
+	"movups  %%xmm6, 32(%3,%0,8)		      \n\t"	// 4 * y
+	"movups  %%xmm7, 48(%3,%0,8)		      \n\t"	// 4 * y
+
+        "addq		$8 , %0	  	 	      \n\t"
+	"subq	        $8 , %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \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", 
+	  "%xmm8", "%xmm9", 
+	  "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From 55e81da379ebf839b28a100f194e03589959ab91 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Sat, 23 Aug 2014 13:12:44 +0200
Subject: [PATCH 23/32] added axpy benchmark-test

---
 benchmark/Makefile |  68 +++++++++++++++
 benchmark/axpy.c   | 201 +++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 269 insertions(+)
 create mode 100644 benchmark/axpy.c

diff --git a/benchmark/Makefile b/benchmark/Makefile
index 933b4bd44..07bf5a792 100644
--- a/benchmark/Makefile
+++ b/benchmark/Makefile
@@ -36,6 +36,7 @@ goto :: slinpack.goto dlinpack.goto clinpack.goto zlinpack.goto \
        ssyr2k.goto dsyr2k.goto csyr2k.goto zsyr2k.goto \
        sger.goto dger.goto \
        sdot.goto ddot.goto \
+       saxpy.goto daxpy.goto caxpy.goto zaxpy.goto \
        ssymv.goto dsymv.goto csymv.goto zsymv.goto \
        chemv.goto zhemv.goto \
        chemm.goto zhemm.goto \
@@ -56,6 +57,7 @@ acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \
        ssyr2k.acml dsyr2k.acml csyr2k.acml zsyr2k.acml \
        sger.acml dger.acml \
        sdot.acml ddot.acml \
+       saxpy.acml daxpy.acml caxpy.acml zaxpy.acml \
        ssymv.acml dsymv.acml csymv.acml zsymv.acml \
        chemv.acml zhemv.acml \
        chemm.acml zhemm.acml \
@@ -76,6 +78,7 @@ atlas :: slinpack.atlas dlinpack.atlas clinpack.atlas zlinpack.atlas \
        ssyr2k.atlas dsyr2k.atlas csyr2k.atlas zsyr2k.atlas \
        sger.atlas dger.atlas \
        sdot.atlas ddot.atlas \
+       saxpy.atlas daxpy.atlas caxpy.atlas zaxpy.atlas \
        ssymv.atlas dsymv.atlas csymv.atlas zsymv.atlas \
        chemv.atlas zhemv.atlas \
        chemm.acml zhemm.acml \
@@ -97,6 +100,7 @@ mkl :: slinpack.mkl dlinpack.mkl clinpack.mkl zlinpack.mkl \
        ssyr2k.mkl dsyr2k.mkl csyr2k.mkl zsyr2k.mkl \
        sger.mkl dger.mkl \
        sdot.mkl ddot.mkl \
+       saxpy.mkl daxpy.mkl caxpy.mkl zaxpy.mkl \
        ssymv.mkl dsymv.mkl csymv.mkl zsymv.mkl \
        chemv.mkl zhemv.mkl \
        chemm.mkl zhemm.mkl \
@@ -984,6 +988,61 @@ ddot.atlas : ddot.$(SUFFIX)
 ddot.mkl : ddot.$(SUFFIX)
 	$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
 
+##################################### Saxpy ####################################################
+saxpy.goto : saxpy.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+saxpy.acml : saxpy.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+saxpy.atlas : saxpy.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+saxpy.mkl : saxpy.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Daxpy ####################################################
+daxpy.goto : daxpy.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+daxpy.acml : daxpy.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+daxpy.atlas : daxpy.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+daxpy.mkl : daxpy.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Caxpy ####################################################
+
+caxpy.goto : caxpy.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+caxpy.acml : caxpy.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+caxpy.atlas : caxpy.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+caxpy.mkl : caxpy.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Zaxpy ####################################################
+
+zaxpy.goto : zaxpy.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+zaxpy.acml : zaxpy.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+zaxpy.atlas : zaxpy.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+zaxpy.mkl : zaxpy.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+
 
 ###################################################################################################
 
@@ -1179,8 +1238,17 @@ sdot.$(SUFFIX) : dot.c
 ddot.$(SUFFIX) : dot.c
 	$(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^
 
+saxpy.$(SUFFIX) : axpy.c
+	$(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^
 
+daxpy.$(SUFFIX) : axpy.c
+	$(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^
 
+caxpy.$(SUFFIX) : axpy.c
+	$(CC) $(CFLAGS) -c -DCOMPLEX -UDOUBLE -o $(@F) $^
+
+zaxpy.$(SUFFIX) : axpy.c
+	$(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^
 
 
 
diff --git a/benchmark/axpy.c b/benchmark/axpy.c
new file mode 100644
index 000000000..ef3b5ae4f
--- /dev/null
+++ b/benchmark/axpy.c
@@ -0,0 +1,201 @@
+/***************************************************************************
+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 <stdio.h>
+#include <stdlib.h>
+#ifdef __CYGWIN32__
+#include <sys/time.h>
+#endif
+#include "common.h"
+
+
+#undef AXPY
+
+#ifdef COMPLEX
+#ifdef DOUBLE
+#define AXPY   BLASFUNC(zaxpy)
+#else
+#define AXPY   BLASFUNC(caxpy)
+#endif
+#else
+#ifdef DOUBLE
+#define AXPY   BLASFUNC(daxpy)
+#else
+#define AXPY   BLASFUNC(saxpy)
+#endif
+#endif
+
+#if defined(__WIN32__) || defined(__WIN64__)
+
+#ifndef DELTA_EPOCH_IN_MICROSECS
+#define DELTA_EPOCH_IN_MICROSECS 11644473600000000ULL
+#endif
+
+int gettimeofday(struct timeval *tv, void *tz){
+
+  FILETIME ft;
+  unsigned __int64 tmpres = 0;
+  static int tzflag;
+
+  if (NULL != tv)
+    {
+      GetSystemTimeAsFileTime(&ft);
+
+      tmpres |= ft.dwHighDateTime;
+      tmpres <<= 32;
+      tmpres |= ft.dwLowDateTime;
+
+      /*converting file time to unix epoch*/
+      tmpres /= 10;  /*convert into microseconds*/
+      tmpres -= DELTA_EPOCH_IN_MICROSECS;
+      tv->tv_sec = (long)(tmpres / 1000000UL);
+      tv->tv_usec = (long)(tmpres % 1000000UL);
+    }
+
+  return 0;
+}
+
+#endif
+
+#if !defined(__WIN32__) && !defined(__WIN64__) && !defined(__CYGWIN32__) && 0
+
+static void *huge_malloc(BLASLONG size){
+  int shmid;
+  void *address;
+
+#ifndef SHM_HUGETLB
+#define SHM_HUGETLB 04000
+#endif
+
+  if ((shmid =shmget(IPC_PRIVATE,
+		     (size + HUGE_PAGESIZE) & ~(HUGE_PAGESIZE - 1),
+		     SHM_HUGETLB | IPC_CREAT |0600)) < 0) {
+    printf( "Memory allocation failed(shmget).\n");
+    exit(1);
+  }
+
+  address = shmat(shmid, NULL, SHM_RND);
+
+  if ((BLASLONG)address == -1){
+    printf( "Memory allocation failed(shmat).\n");
+    exit(1);
+  }
+
+  shmctl(shmid, IPC_RMID, 0);
+
+  return address;
+}
+
+#define malloc huge_malloc
+
+#endif
+
+int MAIN__(int argc, char *argv[]){
+
+  FLOAT *x, *y;
+  FLOAT alpha[2] = { 2.0, 2.0 };
+  blasint m, i;
+  blasint inc_x=1,inc_y=1;
+  int loops = 1;
+  int l;
+  char *p;
+
+  int from =   1;
+  int to   = 200;
+  int step =   1;
+
+  struct timeval start, stop;
+  double time1,timeg;
+
+  argc--;argv++;
+
+  if (argc > 0) { from     = atol(*argv);		argc--; argv++;}
+  if (argc > 0) { to       = MAX(atol(*argv), from);	argc--; argv++;}
+  if (argc > 0) { step     = atol(*argv);		argc--; argv++;}
+
+  if ((p = getenv("OPENBLAS_LOOPS")))  loops = atoi(p);
+  if ((p = getenv("OPENBLAS_INCX")))   inc_x = atoi(p);
+  if ((p = getenv("OPENBLAS_INCY")))   inc_y = atoi(p);
+
+  fprintf(stderr, "From : %3d  To : %3d Step = %3d Inc_x = %d Inc_y = %d Loops = %d\n", from, to, step,inc_x,inc_y,loops);
+
+  if (( x = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_x) * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+  if (( y = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_y) * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+#ifdef linux
+  srandom(getpid());
+#endif
+
+  fprintf(stderr, "   SIZE       Flops\n");
+
+  for(m = from; m <= to; m += step)
+  {
+
+   timeg=0;
+
+   fprintf(stderr, " %6d : ", (int)m);
+
+
+   for (l=0; l<loops; l++)
+   {
+
+   	for(i = 0; i < m * COMPSIZE * abs(inc_x); i++){
+			x[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+   	}
+
+   	for(i = 0; i < m * COMPSIZE * abs(inc_y); i++){
+			y[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+   	}
+    	gettimeofday( &start, (struct timezone *)0);
+
+    	AXPY (&m, alpha, x, &inc_x, y, &inc_y );
+
+    	gettimeofday( &stop, (struct timezone *)0);
+
+    	time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6;
+
+	timeg += time1;
+
+    }
+
+    timeg /= loops;
+
+    fprintf(stderr,
+	    " %10.2f MFlops\n",
+	    COMPSIZE * COMPSIZE * 2. * (double)m / timeg * 1.e-6);
+
+  }
+
+  return 0;
+}
+
+void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));

From e45c960c2c7f151f76b6a01478ebb7574337fdce Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Sat, 23 Aug 2014 17:15:21 +0200
Subject: [PATCH 24/32] added optimized saxpy kernel for nehalem

---
 kernel/x86_64/KERNEL.NEHALEM           |   2 +
 kernel/x86_64/saxpy.c                  | 103 +++++++++++++++++++++++++
 kernel/x86_64/saxpy_microk_nehalem-2.c |  91 ++++++++++++++++++++++
 3 files changed, 196 insertions(+)
 create mode 100644 kernel/x86_64/saxpy.c
 create mode 100644 kernel/x86_64/saxpy_microk_nehalem-2.c

diff --git a/kernel/x86_64/KERNEL.NEHALEM b/kernel/x86_64/KERNEL.NEHALEM
index 5523b4571..3cbaa471f 100644
--- a/kernel/x86_64/KERNEL.NEHALEM
+++ b/kernel/x86_64/KERNEL.NEHALEM
@@ -1,3 +1,5 @@
+SAXPYKERNEL = saxpy.c
+
 SDOTKERNEL     = sdot.c
 DDOTKERNEL     = ddot.c
 
diff --git a/kernel/x86_64/saxpy.c b/kernel/x86_64/saxpy.c
new file mode 100644
index 000000000..e6c016ee3
--- /dev/null
+++ b/kernel/x86_64/saxpy.c
@@ -0,0 +1,103 @@
+/***************************************************************************
+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(NEHALEM)
+#include "saxpy_microk_nehalem-2.c"
+#endif
+
+
+#ifndef HAVE_KERNEL_16
+
+void saxpy_kernel_16(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
+{
+	BLASLONG register i = 0;
+	FLOAT a = *alpha;
+
+	while(i < n)
+        {
+              y[i]   += a * x[i];
+              y[i+1] += a * x[i+1];
+              y[i+2] += a * x[i+2];
+              y[i+3] += a * x[i+3];
+              y[i+4] += a * x[i+4];
+              y[i+5] += a * x[i+5];
+              y[i+6] += a * x[i+6];
+              y[i+7] += a * x[i+7];
+              i+=8 ;
+
+       }
+
+}
+
+#endif
+
+int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2)
+{
+	BLASLONG i=0;
+	BLASLONG ix=0,iy=0;
+
+	if ( n <= 0 )  return(0);
+
+	if ( (inc_x == 1) && (inc_y == 1) )
+	{
+
+		int n1 = n & -16;
+
+		if ( n1 )
+			saxpy_kernel_16(n1, x, y , &da );
+
+		i = n1;
+		while(i < n)
+		{
+
+			y[i] += da * x[i] ;
+			i++ ;
+
+		}
+		return(0);
+
+
+	}
+
+	while(i < n)
+	{
+
+		y[iy] += da * x[ix] ;
+		ix  += inc_x ;
+		iy  += inc_y ;
+		i++ ;
+
+	}
+	return(0);
+
+}
+
+
diff --git a/kernel/x86_64/saxpy_microk_nehalem-2.c b/kernel/x86_64/saxpy_microk_nehalem-2.c
new file mode 100644
index 000000000..14ff51a0d
--- /dev/null
+++ b/kernel/x86_64/saxpy_microk_nehalem-2.c
@@ -0,0 +1,91 @@
+/***************************************************************************
+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_16 1
+static void saxpy_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *alpha) __attribute__ ((noinline));
+
+static void saxpy_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
+{
+
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"movss			(%4), %%xmm0		    \n\t"  // alpha	
+	"shufps          $0,  %%xmm0, %%xmm0                \n\t"
+
+	".align 16				            \n\t"
+	".L01LOOP%=:				            \n\t"
+        // "prefetcht0      192(%2,%0,4)                       \n\t"
+        // "prefetcht0      192(%3,%0,4)                       \n\t"
+
+        "movups                  (%2,%0,4), %%xmm12         \n\t"  // 4 * x
+        "movups                16(%2,%0,4), %%xmm13         \n\t"  // 4 * x
+        "movups                32(%2,%0,4), %%xmm14         \n\t"  // 4 * x
+        "movups                48(%2,%0,4), %%xmm15         \n\t"  // 4 * x
+        "movups                  (%3,%0,4), %%xmm8          \n\t"  // 4 * y
+        "movups                16(%3,%0,4), %%xmm9          \n\t"  // 4 * y
+        "movups                32(%3,%0,4), %%xmm10         \n\t"  // 4 * y
+        "movups                48(%3,%0,4), %%xmm11         \n\t"  // 4 * y
+
+	"mulps          %%xmm0 , %%xmm12 		    \n\t"  // alpha * x
+	"mulps          %%xmm0 , %%xmm13 		    \n\t"  
+	"mulps          %%xmm0 , %%xmm14 		    \n\t"  
+	"mulps          %%xmm0 , %%xmm15 		    \n\t"  
+
+	"addps		%%xmm12, %%xmm8   \n\t" // y += alpha *x
+	"addps		%%xmm13, %%xmm9   \n\t"
+	"addps		%%xmm14, %%xmm10  \n\t"
+	"addps		%%xmm15, %%xmm11  \n\t"
+
+	"movups		%%xmm8 ,   (%3,%0,4)                 \n\t"
+	"movups		%%xmm9 , 16(%3,%0,4)                 \n\t"
+	"movups		%%xmm10, 32(%3,%0,4)                 \n\t"
+	"movups		%%xmm11, 48(%3,%0,4)                 \n\t"
+
+	"addq		$16, %0	  	 	             \n\t"
+	"subq	        $16, %1			             \n\t"		
+	"jnz		.L01LOOP%=		             \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (alpha)   // 4
+	: "cc", 
+	  "%xmm0", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From 29125864b30e730f7c81abf98818fb3f5734a9c4 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Sat, 23 Aug 2014 17:28:01 +0200
Subject: [PATCH 25/32] updated gemm.c

---
 benchmark/gemm.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/benchmark/gemm.c b/benchmark/gemm.c
index fc482c075..4f9a58825 100644
--- a/benchmark/gemm.c
+++ b/benchmark/gemm.c
@@ -142,7 +142,9 @@ int MAIN__(int argc, char *argv[]){
   if (argc > 0) { to       = MAX(atol(*argv), from);	argc--; argv++;}
   if (argc > 0) { step     = atol(*argv);		argc--; argv++;}
 
-  fprintf(stderr, "From : %3d  To : %3d Step = %3d\n", from, to, step);
+  if ((p = getenv("OPENBLAS_TRANS")))  trans=*p;
+
+  fprintf(stderr, "From : %3d  To : %3d Step=%d : Trans=%c\n", from, to, step, trans);
 
   if (( a = (FLOAT *)malloc(sizeof(FLOAT) * to * to * COMPSIZE)) == NULL){
     fprintf(stderr,"Out of Memory!!\n");exit(1);

From b55f99730231d7d33eb3bf0ab49a55c5f22c92d0 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Sat, 23 Aug 2014 17:53:07 +0200
Subject: [PATCH 26/32] added optimized daxpy kernel for nehalem

---
 kernel/x86_64/KERNEL.NEHALEM           |   1 +
 kernel/x86_64/daxpy.c                  | 103 +++++++++++++++++++++++++
 kernel/x86_64/daxpy_microk_nehalem-2.c |  91 ++++++++++++++++++++++
 3 files changed, 195 insertions(+)
 create mode 100644 kernel/x86_64/daxpy.c
 create mode 100644 kernel/x86_64/daxpy_microk_nehalem-2.c

diff --git a/kernel/x86_64/KERNEL.NEHALEM b/kernel/x86_64/KERNEL.NEHALEM
index 3cbaa471f..8adb579cf 100644
--- a/kernel/x86_64/KERNEL.NEHALEM
+++ b/kernel/x86_64/KERNEL.NEHALEM
@@ -1,4 +1,5 @@
 SAXPYKERNEL = saxpy.c
+DAXPYKERNEL = daxpy.c
 
 SDOTKERNEL     = sdot.c
 DDOTKERNEL     = ddot.c
diff --git a/kernel/x86_64/daxpy.c b/kernel/x86_64/daxpy.c
new file mode 100644
index 000000000..feda045d0
--- /dev/null
+++ b/kernel/x86_64/daxpy.c
@@ -0,0 +1,103 @@
+/***************************************************************************
+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(NEHALEM)
+#include "daxpy_microk_nehalem-2.c"
+#endif
+
+
+#ifndef HAVE_KERNEL_8
+
+void daxpy_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
+{
+	BLASLONG register i = 0;
+	FLOAT a = *alpha;
+
+	while(i < n)
+        {
+              y[i]   += a * x[i];
+              y[i+1] += a * x[i+1];
+              y[i+2] += a * x[i+2];
+              y[i+3] += a * x[i+3];
+              y[i+4] += a * x[i+4];
+              y[i+5] += a * x[i+5];
+              y[i+6] += a * x[i+6];
+              y[i+7] += a * x[i+7];
+              i+=8 ;
+
+       }
+
+}
+
+#endif
+
+int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2)
+{
+	BLASLONG i=0;
+	BLASLONG ix=0,iy=0;
+
+	if ( n <= 0 )  return(0);
+
+	if ( (inc_x == 1) && (inc_y == 1) )
+	{
+
+		int n1 = n & -8;
+
+		if ( n1 )
+			daxpy_kernel_8(n1, x, y , &da );
+
+		i = n1;
+		while(i < n)
+		{
+
+			y[i] += da * x[i] ;
+			i++ ;
+
+		}
+		return(0);
+
+
+	}
+
+	while(i < n)
+	{
+
+		y[iy] += da * x[ix] ;
+		ix  += inc_x ;
+		iy  += inc_y ;
+		i++ ;
+
+	}
+	return(0);
+
+}
+
+
diff --git a/kernel/x86_64/daxpy_microk_nehalem-2.c b/kernel/x86_64/daxpy_microk_nehalem-2.c
new file mode 100644
index 000000000..32ed1857c
--- /dev/null
+++ b/kernel/x86_64/daxpy_microk_nehalem-2.c
@@ -0,0 +1,91 @@
+/***************************************************************************
+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_8 1
+static void daxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *alpha) __attribute__ ((noinline));
+
+static void daxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
+{
+
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"movsd			(%4), %%xmm0		    \n\t"  // alpha	
+	"shufpd          $0,  %%xmm0, %%xmm0                \n\t"
+
+	".align 16				            \n\t"
+	".L01LOOP%=:				            \n\t"
+        // "prefetcht0      192(%2,%0,8)                       \n\t"
+        // "prefetcht0      192(%3,%0,8)                       \n\t"
+
+        "movups                  (%2,%0,8), %%xmm12         \n\t"  // 2 * x
+        "movups                16(%2,%0,8), %%xmm13         \n\t"  // 2 * x
+        "movups                32(%2,%0,8), %%xmm14         \n\t"  // 2 * x
+        "movups                48(%2,%0,8), %%xmm15         \n\t"  // 2 * x
+        "movups                  (%3,%0,8), %%xmm8          \n\t"  // 2 * y
+        "movups                16(%3,%0,8), %%xmm9          \n\t"  // 2 * y
+        "movups                32(%3,%0,8), %%xmm10         \n\t"  // 2 * y
+        "movups                48(%3,%0,8), %%xmm11         \n\t"  // 2 * y
+
+	"mulpd          %%xmm0 , %%xmm12 		    \n\t"  // alpha * x
+	"mulpd          %%xmm0 , %%xmm13 		    \n\t"  
+	"mulpd          %%xmm0 , %%xmm14 		    \n\t"  
+	"mulpd          %%xmm0 , %%xmm15 		    \n\t"  
+
+	"addpd		%%xmm12, %%xmm8   \n\t" // y += alpha *x
+	"addpd		%%xmm13, %%xmm9   \n\t"
+	"addpd		%%xmm14, %%xmm10  \n\t"
+	"addpd		%%xmm15, %%xmm11  \n\t"
+
+	"movups		%%xmm8 ,   (%3,%0,8)                 \n\t"
+	"movups		%%xmm9 , 16(%3,%0,8)                 \n\t"
+	"movups		%%xmm10, 32(%3,%0,8)                 \n\t"
+	"movups		%%xmm11, 48(%3,%0,8)                 \n\t"
+
+	"addq		$8 , %0	  	 	             \n\t"
+	"subq	        $8 , %1			             \n\t"		
+	"jnz		.L01LOOP%=		             \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (alpha)   // 4
+	: "cc", 
+	  "%xmm0", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From 9d2ace8bac646d423528b307528043446a9c98ee Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Sun, 24 Aug 2014 10:57:12 +0200
Subject: [PATCH 27/32] added optimized daxpy kernel for bulldozer

---
 kernel/x86_64/daxpy.c                    |  2 +
 kernel/x86_64/daxpy_microk_bulldozer-2.c | 82 ++++++++++++++++++++++++
 2 files changed, 84 insertions(+)
 create mode 100644 kernel/x86_64/daxpy_microk_bulldozer-2.c

diff --git a/kernel/x86_64/daxpy.c b/kernel/x86_64/daxpy.c
index feda045d0..83754cbd3 100644
--- a/kernel/x86_64/daxpy.c
+++ b/kernel/x86_64/daxpy.c
@@ -31,6 +31,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #if defined(NEHALEM)
 #include "daxpy_microk_nehalem-2.c"
+#elif defined(BULLDOZER)
+#include "daxpy_microk_bulldozer-2.c"
 #endif
 
 
diff --git a/kernel/x86_64/daxpy_microk_bulldozer-2.c b/kernel/x86_64/daxpy_microk_bulldozer-2.c
new file mode 100644
index 000000000..b1ef84a18
--- /dev/null
+++ b/kernel/x86_64/daxpy_microk_bulldozer-2.c
@@ -0,0 +1,82 @@
+/***************************************************************************
+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_8 1
+static void daxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *alpha) __attribute__ ((noinline));
+
+static void daxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
+{
+
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vmovddup		(%4), %%xmm0		    \n\t"  // alpha	
+
+	".align 16				            \n\t"
+	".L01LOOP%=:				            \n\t"
+
+        "prefetcht0      768(%3,%0,8)                       \n\t"
+        "vmovups                  (%2,%0,8), %%xmm12         \n\t"  // 2 * x
+	"vfmaddpd       (%3,%0,8), %%xmm0 , %%xmm12, %%xmm8  \n\t"   // y += alpha * x
+        "vmovups                16(%2,%0,8), %%xmm13         \n\t"  // 2 * x
+	".align 2				            \n\t"
+	"vmovups		%%xmm8 ,   (%3,%0,8)                 \n\t"
+	"vfmaddpd     16(%3,%0,8), %%xmm0 , %%xmm13, %%xmm9  \n\t"   // y += alpha * x
+	".align 2				            \n\t"
+        "vmovups                32(%2,%0,8), %%xmm14         \n\t"  // 2 * x
+	"vmovups		%%xmm9 , 16(%3,%0,8)                 \n\t"
+        "prefetcht0      768(%2,%0,8)                       \n\t"
+	".align 2				            \n\t"
+	"vfmaddpd     32(%3,%0,8), %%xmm0 , %%xmm14, %%xmm10 \n\t"   // y += alpha * x
+        "vmovups                48(%2,%0,8), %%xmm15         \n\t"  // 2 * x
+	"vmovups		%%xmm10, 32(%3,%0,8)                 \n\t"
+	"vfmaddpd     48(%3,%0,8), %%xmm0 , %%xmm15, %%xmm11 \n\t"   // y += alpha * x
+	"vmovups		%%xmm11, 48(%3,%0,8)                 \n\t"
+
+	"addq		$8 , %0	  	 	             \n\t"
+	"subq	        $8 , %1			             \n\t"		
+	"jnz		.L01LOOP%=		             \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (alpha)   // 4
+	: "cc", 
+	  "%xmm0", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From ee744451559d7e41c8bab16d2bd82f6f2ab8c103 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Mon, 25 Aug 2014 14:53:28 +0200
Subject: [PATCH 28/32] added optimized caxpy kernel for bulldozer

---
 kernel/x86_64/KERNEL.BULLDOZER           |   3 +
 kernel/x86_64/caxpy.c                    | 131 ++++++++++++++++++++++
 kernel/x86_64/caxpy_microk_bulldozer-2.c | 135 +++++++++++++++++++++++
 kernel/x86_64/daxpy.c                    |   2 +-
 kernel/x86_64/ddot.c                     |   2 +-
 kernel/x86_64/saxpy.c                    |   2 +-
 kernel/x86_64/sdot.c                     |   2 +-
 7 files changed, 273 insertions(+), 4 deletions(-)
 create mode 100644 kernel/x86_64/caxpy.c
 create mode 100644 kernel/x86_64/caxpy_microk_bulldozer-2.c

diff --git a/kernel/x86_64/KERNEL.BULLDOZER b/kernel/x86_64/KERNEL.BULLDOZER
index 9e6fc2e28..701eea310 100644
--- a/kernel/x86_64/KERNEL.BULLDOZER
+++ b/kernel/x86_64/KERNEL.BULLDOZER
@@ -1,3 +1,6 @@
+DAXPYKERNEL = daxpy.c
+CAXPYKERNEL = caxpy.c
+
 SDOTKERNEL     = sdot.c
 DDOTKERNEL     = ddot.c
 
diff --git a/kernel/x86_64/caxpy.c b/kernel/x86_64/caxpy.c
new file mode 100644
index 000000000..fa8924ae9
--- /dev/null
+++ b/kernel/x86_64/caxpy.c
@@ -0,0 +1,131 @@
+/***************************************************************************
+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 "caxpy_microk_bulldozer-2.c"
+#endif
+
+
+#ifndef HAVE_KERNEL_8
+
+static void caxpy_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
+{
+	BLASLONG register i  = 0;
+	BLASLONG register ix = 0;
+	FLOAT da_r = alpha[0];
+	FLOAT da_i = alpha[1];
+	
+
+	while(i < n)
+        {
+#if !defined(CONJ)
+              y[ix]   += ( da_r * x[ix]   - da_i * x[ix+1] ) ;
+              y[ix+1] += ( da_r * x[ix+1] + da_i * x[ix]   ) ;
+              y[ix+2] += ( da_r * x[ix+2] - da_i * x[ix+3] ) ;
+              y[ix+3] += ( da_r * x[ix+3] + da_i * x[ix+2] ) ;
+#else
+              y[ix]   += ( da_r * x[ix]   + da_i * x[ix+1] ) ;
+              y[ix+1] -= ( da_r * x[ix+1] - da_i * x[ix]   ) ;
+              y[ix+2] += ( da_r * x[ix+2] + da_i * x[ix+3] ) ;
+              y[ix+3] -= ( da_r * x[ix+3] - da_i * x[ix+2] ) ;
+#endif
+
+              ix+=4 ;
+              i+=2 ;
+
+       }
+
+}
+
+#endif
+
+int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r, FLOAT da_i, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2)
+{
+	BLASLONG i=0;
+	BLASLONG ix=0,iy=0;
+	FLOAT da[2];
+
+	if ( n <= 0 )  return(0);
+
+	if ( (inc_x == 1) && (inc_y == 1) )
+	{
+
+		int n1 = n & -8;
+
+		if ( n1 )
+		{
+			da[0] = da_r;
+			da[1] = da_i;
+			caxpy_kernel_8(n1, x, y , &da );
+			ix = 2 * n1;
+		}
+		i = n1;
+		while(i < n)
+		{
+#if !defined(CONJ)
+                	y[ix]   += ( da_r * x[ix]   - da_i * x[ix+1] ) ;
+                	y[ix+1] += ( da_r * x[ix+1] + da_i * x[ix]   ) ;
+#else
+                	y[ix]   += ( da_r * x[ix]   + da_i * x[ix+1] ) ;
+                	y[ix+1] -= ( da_r * x[ix+1] - da_i * x[ix]   ) ;
+#endif
+			i++ ;
+			ix += 2;
+
+		}
+		return(0);
+
+
+	}
+
+	inc_x *=2;
+	inc_y *=2;
+
+	while(i < n)
+	{
+
+#if !defined(CONJ)
+                y[iy]   += ( da_r * x[ix]   - da_i * x[ix+1] ) ;
+                y[iy+1] += ( da_r * x[ix+1] + da_i * x[ix]   ) ;
+#else
+                y[iy]   += ( da_r * x[ix]   + da_i * x[ix+1] ) ;
+                y[iy+1] -= ( da_r * x[ix+1] - da_i * x[ix]   ) ;
+#endif
+		ix  += inc_x ;
+		iy  += inc_y ;
+		i++ ;
+
+	}
+	return(0);
+
+}
+
+
diff --git a/kernel/x86_64/caxpy_microk_bulldozer-2.c b/kernel/x86_64/caxpy_microk_bulldozer-2.c
new file mode 100644
index 000000000..86407028c
--- /dev/null
+++ b/kernel/x86_64/caxpy_microk_bulldozer-2.c
@@ -0,0 +1,135 @@
+/***************************************************************************
+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_8 1
+static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *alpha) __attribute__ ((noinline));
+
+static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
+{
+
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vbroadcastss		(%4), %%xmm0		    \n\t"  // real part of alpha
+	"vbroadcastss	       4(%4), %%xmm1		    \n\t"  // imag part of alpha
+
+	".align 16				            \n\t"
+	".L01LOOP%=:				            \n\t"
+
+	"prefetcht0  768(%2,%0,4)                           \n\t"
+	"vmovups        (%2,%0,4), %%xmm5                   \n\t" // 2 complex values from x
+	"vmovups      16(%2,%0,4), %%xmm7                   \n\t" // 2 complex values from x
+	"vmovups      32(%2,%0,4), %%xmm9                   \n\t" // 2 complex values from x
+	"vmovups      48(%2,%0,4), %%xmm11                  \n\t" // 2 complex values from x
+	"prefetcht0  768(%3,%0,4)                           \n\t"
+
+#if !defined(CONJ)
+	"vfmaddps    (%3,%0,4), %%xmm0 , %%xmm5, %%xmm12    \n\t"
+	"vpermilps	$0xb1 , %%xmm5 , %%xmm4 	    \n\t"  // exchange real and imag part
+	"vmulps         %%xmm1, %%xmm4 , %%xmm4             \n\t"
+
+	"vfmaddps  16(%3,%0,4), %%xmm0 , %%xmm7, %%xmm13    \n\t"
+	"vpermilps	$0xb1 , %%xmm7 , %%xmm6 	    \n\t"  // exchange real and imag part
+	"vmulps         %%xmm1, %%xmm6 , %%xmm6             \n\t"
+
+	"vfmaddps  32(%3,%0,4), %%xmm0 , %%xmm9, %%xmm14    \n\t"
+	"vpermilps	$0xb1 , %%xmm9 , %%xmm8 	    \n\t"  // exchange real and imag part
+	"vmulps         %%xmm1, %%xmm8 , %%xmm8             \n\t"
+
+	"vfmaddps  48(%3,%0,4), %%xmm0 , %%xmm11,%%xmm15    \n\t"
+	"vpermilps	$0xb1 , %%xmm11, %%xmm10 	    \n\t"  // exchange real and imag part
+	"vmulps         %%xmm1, %%xmm10, %%xmm10            \n\t"
+
+        "vaddsubps      %%xmm4, %%xmm12, %%xmm12            \n\t"
+        "vaddsubps      %%xmm6, %%xmm13, %%xmm13            \n\t"
+        "vaddsubps      %%xmm8, %%xmm14, %%xmm14            \n\t"
+        "vaddsubps      %%xmm10,%%xmm15, %%xmm15            \n\t"
+
+#else
+
+	"vmulps		%%xmm0, %%xmm5, %%xmm4		    \n\t" // a_r*x_r, a_r*x_i
+	"vmulps		%%xmm1, %%xmm5, %%xmm5 		    \n\t" // a_i*x_r, a_i*x_i
+	"vmulps		%%xmm0, %%xmm7, %%xmm6		    \n\t" // a_r*x_r, a_r*x_i
+	"vmulps		%%xmm1, %%xmm7, %%xmm7 		    \n\t" // a_i*x_r, a_i*x_i
+	"vmulps		%%xmm0, %%xmm9, %%xmm8		    \n\t" // a_r*x_r, a_r*x_i
+	"vmulps		%%xmm1, %%xmm9, %%xmm9 		    \n\t" // a_i*x_r, a_i*x_i
+	"vmulps		%%xmm0, %%xmm11, %%xmm10	    \n\t" // a_r*x_r, a_r*x_i
+	"vmulps		%%xmm1, %%xmm11, %%xmm11	    \n\t" // a_i*x_r, a_i*x_i
+
+	"vpermilps	$0xb1 , %%xmm4 , %%xmm4 	    \n\t"  // exchange real and imag part
+	"vaddsubps	%%xmm4 ,%%xmm5 , %%xmm4 	    \n\t"
+	"vpermilps	$0xb1 , %%xmm4 , %%xmm4 	    \n\t"  // exchange real and imag part
+
+	"vpermilps	$0xb1 , %%xmm6 , %%xmm6 	    \n\t"  // exchange real and imag part
+	"vaddsubps	%%xmm6 ,%%xmm7 , %%xmm6 	    \n\t"
+	"vpermilps	$0xb1 , %%xmm6 , %%xmm6 	    \n\t"  // exchange real and imag part
+
+	"vpermilps	$0xb1 , %%xmm8 , %%xmm8 	    \n\t"  // exchange real and imag part
+	"vaddsubps	%%xmm8 ,%%xmm9 , %%xmm8 	    \n\t"
+	"vpermilps	$0xb1 , %%xmm8 , %%xmm8 	    \n\t"  // exchange real and imag part
+
+	"vpermilps	$0xb1 , %%xmm10, %%xmm10	    \n\t"  // exchange real and imag part
+	"vaddsubps	%%xmm10,%%xmm11, %%xmm10	    \n\t"
+	"vpermilps	$0xb1 , %%xmm10, %%xmm10	    \n\t"  // exchange real and imag part
+
+	"vaddps	     (%3,%0,4) ,%%xmm4 , %%xmm12	    \n\t"
+	"vaddps	   16(%3,%0,4) ,%%xmm6 , %%xmm13	    \n\t"
+	"vaddps	   32(%3,%0,4) ,%%xmm8 , %%xmm14	    \n\t"
+	"vaddps	   48(%3,%0,4) ,%%xmm10, %%xmm15	    \n\t"
+
+
+#endif
+
+	"vmovups	%%xmm12,   (%3,%0,4)		    \n\t"
+	"vmovups	%%xmm13, 16(%3,%0,4)		    \n\t"
+	"vmovups	%%xmm14, 32(%3,%0,4)		    \n\t"
+	"vmovups	%%xmm15, 48(%3,%0,4)		    \n\t"
+
+	"addq		$16, %0	  	 	             \n\t"
+	"subq	        $8 , %1			             \n\t"		
+	"jnz		.L01LOOP%=		             \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (alpha)   // 4
+	: "cc", 
+	  "%xmm0", "%xmm1",
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
diff --git a/kernel/x86_64/daxpy.c b/kernel/x86_64/daxpy.c
index 83754cbd3..f1d50c909 100644
--- a/kernel/x86_64/daxpy.c
+++ b/kernel/x86_64/daxpy.c
@@ -38,7 +38,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #ifndef HAVE_KERNEL_8
 
-void daxpy_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
+static void daxpy_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 {
 	BLASLONG register i = 0;
 	FLOAT a = *alpha;
diff --git a/kernel/x86_64/ddot.c b/kernel/x86_64/ddot.c
index ee6785f6a..b3aad438f 100644
--- a/kernel/x86_64/ddot.c
+++ b/kernel/x86_64/ddot.c
@@ -38,7 +38,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #ifndef HAVE_KERNEL_8
 
-void ddot_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *d)
+static void ddot_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *d)
 {
 	BLASLONG register i = 0;
 	FLOAT dot = 0.0;
diff --git a/kernel/x86_64/saxpy.c b/kernel/x86_64/saxpy.c
index e6c016ee3..da81f1354 100644
--- a/kernel/x86_64/saxpy.c
+++ b/kernel/x86_64/saxpy.c
@@ -36,7 +36,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #ifndef HAVE_KERNEL_16
 
-void saxpy_kernel_16(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
+static void saxpy_kernel_16(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 {
 	BLASLONG register i = 0;
 	FLOAT a = *alpha;
diff --git a/kernel/x86_64/sdot.c b/kernel/x86_64/sdot.c
index a13d65d25..632d16810 100644
--- a/kernel/x86_64/sdot.c
+++ b/kernel/x86_64/sdot.c
@@ -37,7 +37,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #ifndef HAVE_KERNEL_16
 
-void sdot_kernel_16(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *d)
+static void sdot_kernel_16(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *d)
 {
 	BLASLONG register i = 0;
 	FLOAT dot = 0.0;

From 3885eebdb82787b452b532e6cb7a23d9711a514e Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Mon, 25 Aug 2014 15:52:35 +0200
Subject: [PATCH 29/32] added optimized zaxpy bulldozer kernel

---
 kernel/x86_64/KERNEL.BULLDOZER           |   1 +
 kernel/x86_64/zaxpy.c                    | 131 ++++++++++++++++++++++
 kernel/x86_64/zaxpy_microk_bulldozer-2.c | 135 +++++++++++++++++++++++
 3 files changed, 267 insertions(+)
 create mode 100644 kernel/x86_64/zaxpy.c
 create mode 100644 kernel/x86_64/zaxpy_microk_bulldozer-2.c

diff --git a/kernel/x86_64/KERNEL.BULLDOZER b/kernel/x86_64/KERNEL.BULLDOZER
index 701eea310..6318b202c 100644
--- a/kernel/x86_64/KERNEL.BULLDOZER
+++ b/kernel/x86_64/KERNEL.BULLDOZER
@@ -1,5 +1,6 @@
 DAXPYKERNEL = daxpy.c
 CAXPYKERNEL = caxpy.c
+ZAXPYKERNEL = zaxpy.c
 
 SDOTKERNEL     = sdot.c
 DDOTKERNEL     = ddot.c
diff --git a/kernel/x86_64/zaxpy.c b/kernel/x86_64/zaxpy.c
new file mode 100644
index 000000000..ca2f03dd0
--- /dev/null
+++ b/kernel/x86_64/zaxpy.c
@@ -0,0 +1,131 @@
+/***************************************************************************
+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 "zaxpy_microk_bulldozer-2.c"
+#endif
+
+
+#ifndef HAVE_KERNEL_4
+
+static void zaxpy_kernel_4(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
+{
+	BLASLONG register i  = 0;
+	BLASLONG register ix = 0;
+	FLOAT da_r = alpha[0];
+	FLOAT da_i = alpha[1];
+	
+
+	while(i < n)
+        {
+#if !defined(CONJ)
+              y[ix]   += ( da_r * x[ix]   - da_i * x[ix+1] ) ;
+              y[ix+1] += ( da_r * x[ix+1] + da_i * x[ix]   ) ;
+              y[ix+2] += ( da_r * x[ix+2] - da_i * x[ix+3] ) ;
+              y[ix+3] += ( da_r * x[ix+3] + da_i * x[ix+2] ) ;
+#else
+              y[ix]   += ( da_r * x[ix]   + da_i * x[ix+1] ) ;
+              y[ix+1] -= ( da_r * x[ix+1] - da_i * x[ix]   ) ;
+              y[ix+2] += ( da_r * x[ix+2] + da_i * x[ix+3] ) ;
+              y[ix+3] -= ( da_r * x[ix+3] - da_i * x[ix+2] ) ;
+#endif
+
+              ix+=4 ;
+              i+=2 ;
+
+       }
+
+}
+
+#endif
+
+int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r, FLOAT da_i, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2)
+{
+	BLASLONG i=0;
+	BLASLONG ix=0,iy=0;
+	FLOAT da[2];
+
+	if ( n <= 0 )  return(0);
+
+	if ( (inc_x == 1) && (inc_y == 1) )
+	{
+
+		int n1 = n & -4;
+
+		if ( n1 )
+		{
+			da[0] = da_r;
+			da[1] = da_i;
+			zaxpy_kernel_4(n1, x, y , &da );
+			ix = 2 * n1;
+		}
+		i = n1;
+		while(i < n)
+		{
+#if !defined(CONJ)
+                	y[ix]   += ( da_r * x[ix]   - da_i * x[ix+1] ) ;
+                	y[ix+1] += ( da_r * x[ix+1] + da_i * x[ix]   ) ;
+#else
+                	y[ix]   += ( da_r * x[ix]   + da_i * x[ix+1] ) ;
+                	y[ix+1] -= ( da_r * x[ix+1] - da_i * x[ix]   ) ;
+#endif
+			i++ ;
+			ix += 2;
+
+		}
+		return(0);
+
+
+	}
+
+	inc_x *=2;
+	inc_y *=2;
+
+	while(i < n)
+	{
+
+#if !defined(CONJ)
+                y[iy]   += ( da_r * x[ix]   - da_i * x[ix+1] ) ;
+                y[iy+1] += ( da_r * x[ix+1] + da_i * x[ix]   ) ;
+#else
+                y[iy]   += ( da_r * x[ix]   + da_i * x[ix+1] ) ;
+                y[iy+1] -= ( da_r * x[ix+1] - da_i * x[ix]   ) ;
+#endif
+		ix  += inc_x ;
+		iy  += inc_y ;
+		i++ ;
+
+	}
+	return(0);
+
+}
+
+
diff --git a/kernel/x86_64/zaxpy_microk_bulldozer-2.c b/kernel/x86_64/zaxpy_microk_bulldozer-2.c
new file mode 100644
index 000000000..780109b69
--- /dev/null
+++ b/kernel/x86_64/zaxpy_microk_bulldozer-2.c
@@ -0,0 +1,135 @@
+/***************************************************************************
+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_4 1
+static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *alpha) __attribute__ ((noinline));
+
+static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
+{
+
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vmovddup		(%4), %%xmm0		    \n\t"  // real part of alpha
+	"vmovddup	       8(%4), %%xmm1		    \n\t"  // imag part of alpha
+
+	".align 16				            \n\t"
+	".L01LOOP%=:				            \n\t"
+
+	"prefetcht0  768(%2,%0,8)                           \n\t"
+	"vmovups        (%2,%0,8), %%xmm5                   \n\t" // 1 complex values from x
+	"vmovups      16(%2,%0,8), %%xmm7                   \n\t" // 1 complex values from x
+	"vmovups      32(%2,%0,8), %%xmm9                   \n\t" // 1 complex values from x
+	"vmovups      48(%2,%0,8), %%xmm11                  \n\t" // 1 complex values from x
+	"prefetcht0  768(%3,%0,8)                           \n\t"
+
+#if !defined(CONJ)
+	"vfmaddpd    (%3,%0,8), %%xmm0 , %%xmm5, %%xmm12    \n\t"
+	"vpermilpd	$0x1 , %%xmm5 , %%xmm4 	    \n\t"  // exchange real and imag part
+	"vmulpd         %%xmm1, %%xmm4 , %%xmm4             \n\t"
+
+	"vfmaddpd  16(%3,%0,8), %%xmm0 , %%xmm7, %%xmm13    \n\t"
+	"vpermilpd	$0x1 , %%xmm7 , %%xmm6 	    \n\t"  // exchange real and imag part
+	"vmulpd         %%xmm1, %%xmm6 , %%xmm6             \n\t"
+
+	"vfmaddpd  32(%3,%0,8), %%xmm0 , %%xmm9, %%xmm14    \n\t"
+	"vpermilpd	$0x1 , %%xmm9 , %%xmm8 	    \n\t"  // exchange real and imag part
+	"vmulpd         %%xmm1, %%xmm8 , %%xmm8             \n\t"
+
+	"vfmaddpd  48(%3,%0,8), %%xmm0 , %%xmm11,%%xmm15    \n\t"
+	"vpermilpd	$0x1 , %%xmm11, %%xmm10 	    \n\t"  // exchange real and imag part
+	"vmulpd         %%xmm1, %%xmm10, %%xmm10            \n\t"
+
+        "vaddsubpd      %%xmm4, %%xmm12, %%xmm12            \n\t"
+        "vaddsubpd      %%xmm6, %%xmm13, %%xmm13            \n\t"
+        "vaddsubpd      %%xmm8, %%xmm14, %%xmm14            \n\t"
+        "vaddsubpd      %%xmm10,%%xmm15, %%xmm15            \n\t"
+
+#else
+
+	"vmulpd		%%xmm0, %%xmm5, %%xmm4		    \n\t" // a_r*x_r, a_r*x_i
+	"vmulpd		%%xmm1, %%xmm5, %%xmm5 		    \n\t" // a_i*x_r, a_i*x_i
+	"vmulpd		%%xmm0, %%xmm7, %%xmm6		    \n\t" // a_r*x_r, a_r*x_i
+	"vmulpd		%%xmm1, %%xmm7, %%xmm7 		    \n\t" // a_i*x_r, a_i*x_i
+	"vmulpd		%%xmm0, %%xmm9, %%xmm8		    \n\t" // a_r*x_r, a_r*x_i
+	"vmulpd		%%xmm1, %%xmm9, %%xmm9 		    \n\t" // a_i*x_r, a_i*x_i
+	"vmulpd		%%xmm0, %%xmm11, %%xmm10	    \n\t" // a_r*x_r, a_r*x_i
+	"vmulpd		%%xmm1, %%xmm11, %%xmm11	    \n\t" // a_i*x_r, a_i*x_i
+
+	"vpermilpd	$0x1 , %%xmm4 , %%xmm4 	    \n\t"  // exchange real and imag part
+	"vaddsubpd	%%xmm4 ,%%xmm5 , %%xmm4 	    \n\t"
+	"vpermilpd	$0x1 , %%xmm4 , %%xmm4 	    \n\t"  // exchange real and imag part
+
+	"vpermilpd	$0x1 , %%xmm6 , %%xmm6 	    \n\t"  // exchange real and imag part
+	"vaddsubpd	%%xmm6 ,%%xmm7 , %%xmm6 	    \n\t"
+	"vpermilpd	$0x1 , %%xmm6 , %%xmm6 	    \n\t"  // exchange real and imag part
+
+	"vpermilpd	$0x1 , %%xmm8 , %%xmm8 	    \n\t"  // exchange real and imag part
+	"vaddsubpd	%%xmm8 ,%%xmm9 , %%xmm8 	    \n\t"
+	"vpermilpd	$0x1 , %%xmm8 , %%xmm8 	    \n\t"  // exchange real and imag part
+
+	"vpermilpd	$0x1 , %%xmm10, %%xmm10	    \n\t"  // exchange real and imag part
+	"vaddsubpd	%%xmm10,%%xmm11, %%xmm10	    \n\t"
+	"vpermilpd	$0x1 , %%xmm10, %%xmm10	    \n\t"  // exchange real and imag part
+
+	"vaddpd	     (%3,%0,8) ,%%xmm4 , %%xmm12	    \n\t"
+	"vaddpd	   16(%3,%0,8) ,%%xmm6 , %%xmm13	    \n\t"
+	"vaddpd	   32(%3,%0,8) ,%%xmm8 , %%xmm14	    \n\t"
+	"vaddpd	   48(%3,%0,8) ,%%xmm10, %%xmm15	    \n\t"
+
+
+#endif
+
+	"vmovups	%%xmm12,   (%3,%0,8)		    \n\t"
+	"vmovups	%%xmm13, 16(%3,%0,8)		    \n\t"
+	"vmovups	%%xmm14, 32(%3,%0,8)		    \n\t"
+	"vmovups	%%xmm15, 48(%3,%0,8)		    \n\t"
+
+	"addq		$8 , %0	  	 	             \n\t"
+	"subq	        $4 , %1			             \n\t"		
+	"jnz		.L01LOOP%=		             \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (alpha)   // 4
+	: "cc", 
+	  "%xmm0", "%xmm1",
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From 84badf80866d300eecdb62f49689df1b6ca2ccd7 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Tue, 26 Aug 2014 17:36:32 +0200
Subject: [PATCH 30/32] EXPERIMENTAL: added the flag -no-integrated-as for
 clang compiler in Makefile.system

---
 Makefile.system | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/Makefile.system b/Makefile.system
index ccde8e9ce..385ad47c3 100644
--- a/Makefile.system
+++ b/Makefile.system
@@ -339,7 +339,7 @@ FCOMMON_OPT	+= -m128bit-long-double
 endif
 ifeq ($(C_COMPILER), CLANG)
 EXPRECISION	= 1
-CCOMMON_OPT	+= -DEXPRECISION
+CCOMMON_OPT	+= -DEXPRECISION 
 FCOMMON_OPT	+= -m128bit-long-double
 endif
 endif
@@ -350,6 +350,12 @@ ifeq ($(C_COMPILER), INTEL)
 CCOMMON_OPT    += -wd981
 endif
 
+ifeq ($(ARCH), x86_64)
+ifeq ($(C_COMPILER), CLANG)
+CCOMMON_OPT    += -no-integrated-as
+endif
+endif
+
 ifeq ($(USE_OPENMP), 1)
 # ifeq logical or. GCC or LSB
 ifeq ($(C_COMPILER), $(filter $(C_COMPILER),GCC LSB))

From 5fa61587312bd6221fe4050233b7135f3cc73a3f Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Tue, 26 Aug 2014 18:29:40 +0200
Subject: [PATCH 31/32] renoved flag no-integrated-as, because not working on
 macosx

---
 Makefile.system | 5 -----
 1 file changed, 5 deletions(-)

diff --git a/Makefile.system b/Makefile.system
index 385ad47c3..d2ff74146 100644
--- a/Makefile.system
+++ b/Makefile.system
@@ -350,11 +350,6 @@ ifeq ($(C_COMPILER), INTEL)
 CCOMMON_OPT    += -wd981
 endif
 
-ifeq ($(ARCH), x86_64)
-ifeq ($(C_COMPILER), CLANG)
-CCOMMON_OPT    += -no-integrated-as
-endif
-endif
 
 ifeq ($(USE_OPENMP), 1)
 # ifeq logical or. GCC or LSB

From 20cd85012509993f2e1ed9b9ab83ae63297e0e6d Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Wed, 27 Aug 2014 09:00:20 +0200
Subject: [PATCH 32/32] modification for clang compiler

---
 kernel/x86_64/cgemm_kernel_8x2_haswell.S  | 16 ++++++++--------
 kernel/x86_64/sgemm_kernel_16x4_haswell.S | 20 ++++++++++----------
 kernel/x86_64/zgemm_kernel_4x2_haswell.S  | 12 ++++++------
 3 files changed, 24 insertions(+), 24 deletions(-)

diff --git a/kernel/x86_64/cgemm_kernel_8x2_haswell.S b/kernel/x86_64/cgemm_kernel_8x2_haswell.S
index 98f40054e..a608071db 100644
--- a/kernel/x86_64/cgemm_kernel_8x2_haswell.S
+++ b/kernel/x86_64/cgemm_kernel_8x2_haswell.S
@@ -227,8 +227,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         VFMADDPS_I(        %ymm7 ,%ymm3,%ymm1 )
 
 
-        addq    $6*SIZE, BO                           
-        addq    $16*SIZE, AO                         
+        addq    $ 6*SIZE, BO                           
+        addq    $ 16*SIZE, AO                         
 	decq	%rax
 .endm
 
@@ -356,8 +356,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         VFMADDPS_R(        %ymm4 ,%ymm2,%ymm0 )
         VFMADDPS_I(        %ymm5 ,%ymm3,%ymm0 )
 
-        addq    $6*SIZE, BO                           
-        addq    $8*SIZE, AO                         
+        addq    $ 6*SIZE, BO                           
+        addq    $ 8*SIZE, AO                         
 	decq	%rax
 .endm
 
@@ -447,8 +447,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         VFMADDPS_R(        %xmm4 ,%xmm2,%xmm0 )
         VFMADDPS_I(        %xmm5 ,%xmm3,%xmm0 )
 
-        addq    $6*SIZE, BO                           
-        addq    $4*SIZE, AO                         
+        addq    $ 6*SIZE, BO                           
+        addq    $ 4*SIZE, AO                         
 	decq	%rax
 
 .endm
@@ -540,8 +540,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         VFMADDPS_R(        %xmm4 ,%xmm2,%xmm0 )
         VFMADDPS_I(        %xmm5 ,%xmm3,%xmm0 )
 
-        addq    $6*SIZE, BO                           
-        addq    $2*SIZE, AO                         
+        addq    $ 6*SIZE, BO                           
+        addq    $ 2*SIZE, AO                         
 	decq	%rax
 
 .endm
diff --git a/kernel/x86_64/sgemm_kernel_16x4_haswell.S b/kernel/x86_64/sgemm_kernel_16x4_haswell.S
index d88add02b..ef156fd27 100644
--- a/kernel/x86_64/sgemm_kernel_16x4_haswell.S
+++ b/kernel/x86_64/sgemm_kernel_16x4_haswell.S
@@ -181,8 +181,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	VFMADD231PS_(  	%ymm14,%ymm3,%ymm0 )
 	VFMADD231PS_(  	%ymm15,%ymm3,%ymm1 )
 
-	addq	$6*SIZE, BO 
-	addq	$16*SIZE, AO 
+	addq	$ 6*SIZE, BO 
+	addq	$ 16*SIZE, AO 
 	decq	%rax 
 .endm
 
@@ -268,8 +268,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	VFMADD231PS_(  	%ymm12,%ymm2,%ymm0  )
 	VFMADD231PS_(  	%ymm14,%ymm3,%ymm0 )
 
-	addq	$6*SIZE, BO 
-	addq	$8*SIZE, AO 
+	addq	$ 6*SIZE, BO 
+	addq	$ 8*SIZE, AO 
 	decq	%rax 
 .endm
 
@@ -327,8 +327,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	VFMADD231PS_(  	%xmm12,%xmm2,%xmm0  )
 	VFMADD231PS_(  	%xmm14,%xmm3,%xmm0 )
 
-	addq	$6*SIZE, BO 
-	addq	$4*SIZE, AO 
+	addq	$ 6*SIZE, BO 
+	addq	$ 4*SIZE, AO 
 	decq	%rax 
 .endm
 
@@ -392,8 +392,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	VFMADD231SS_(  	%xmm14,%xmm3,%xmm0 )
 	VFMADD231SS_(  	%xmm15,%xmm3,%xmm1 )
 
-	addq	$6*SIZE, BO 
-	addq	$2*SIZE, AO 
+	addq	$ 6*SIZE, BO 
+	addq	$ 2*SIZE, AO 
 	decq	%rax 
 .endm
 
@@ -478,8 +478,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	VFMADD231SS_(  	%xmm12,%xmm2,%xmm0  )
 	VFMADD231SS_(  	%xmm14,%xmm3,%xmm0 )
 
-	addq	$6*SIZE, BO 
-	addq	$1*SIZE, AO 
+	addq	$ 6*SIZE, BO 
+	addq	$ 1*SIZE, AO 
 	decq	%rax 
 .endm
 
diff --git a/kernel/x86_64/zgemm_kernel_4x2_haswell.S b/kernel/x86_64/zgemm_kernel_4x2_haswell.S
index e23e09ecc..f91bfa89b 100644
--- a/kernel/x86_64/zgemm_kernel_4x2_haswell.S
+++ b/kernel/x86_64/zgemm_kernel_4x2_haswell.S
@@ -222,8 +222,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         VFMADDPD_I(        %ymm5 ,%ymm3,%ymm0 )
         VFMADDPD_I(        %ymm7 ,%ymm3,%ymm1 )
 
-        addq    $6*SIZE, BO                           
-        addq    $8*SIZE, AO                           
+        addq    $ 6*SIZE, BO                           
+        addq    $ 8*SIZE, AO                           
         decq	%rax                         
 .endm
 
@@ -362,8 +362,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         VFMADDPD_I(        %xmm5 ,%xmm3,%xmm0 )
         VFMADDPD_I(        %xmm7 ,%xmm3,%xmm1 )
 
-        addq    $6*SIZE, BO                           
-        addq    $4*SIZE, AO                           
+        addq    $ 6*SIZE, BO                           
+        addq    $ 4*SIZE, AO                           
         decq    %rax                         
 .endm
 
@@ -491,8 +491,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         VFMADDPD_R(        %xmm4 ,%xmm2,%xmm0 )
         VFMADDPD_I(        %xmm5 ,%xmm3,%xmm0 )
 
-        addq    $6*SIZE, BO                           
-        addq    $2*SIZE, AO                           
+        addq    $ 6*SIZE, BO                           
+        addq    $ 2*SIZE, AO                           
         decq    %rax                         
 .endm