From db6917303f12a64fe11430a3b46aaabdefd36234 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Mon, 4 Aug 2014 14:29:01 +0200
Subject: [PATCH 1/9] added a better optimized sgemv_n kernel for bulldozer and
 piledriver

---
 kernel/x86_64/KERNEL.BULLDOZER             |   2 +-
 kernel/x86_64/KERNEL.PILEDRIVER            |   2 +-
 kernel/x86_64/sgemv_n.c                    | 203 +++++++++++++++++++++
 kernel/x86_64/sgemv_n_microk_bulldozer-2.c |  99 ++++++++++
 4 files changed, 304 insertions(+), 2 deletions(-)
 create mode 100644 kernel/x86_64/sgemv_n.c
 create mode 100644 kernel/x86_64/sgemv_n_microk_bulldozer-2.c

diff --git a/kernel/x86_64/KERNEL.BULLDOZER b/kernel/x86_64/KERNEL.BULLDOZER
index 73a9ad2ec..ca7a0b7c4 100644
--- a/kernel/x86_64/KERNEL.BULLDOZER
+++ b/kernel/x86_64/KERNEL.BULLDOZER
@@ -2,7 +2,7 @@ ifdef OS_WINDOWS
 SGEMVNKERNEL = ../arm/gemv_n.c
 SGEMVTKERNEL = ../arm/gemv_t.c
 else
-SGEMVNKERNEL = sgemv_n_avx.c
+SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t_avx.c
 endif
 
diff --git a/kernel/x86_64/KERNEL.PILEDRIVER b/kernel/x86_64/KERNEL.PILEDRIVER
index 453e7b762..b9d680c43 100644
--- a/kernel/x86_64/KERNEL.PILEDRIVER
+++ b/kernel/x86_64/KERNEL.PILEDRIVER
@@ -2,7 +2,7 @@ ifdef OS_WINDOWS
 SGEMVNKERNEL = ../arm/gemv_n.c
 SGEMVTKERNEL = ../arm/gemv_t.c
 else
-SGEMVNKERNEL = sgemv_n_avx.c
+SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t_avx.c
 endif
 
diff --git a/kernel/x86_64/sgemv_n.c b/kernel/x86_64/sgemv_n.c
new file mode 100644
index 000000000..6f240797d
--- /dev/null
+++ b/kernel/x86_64/sgemv_n.c
@@ -0,0 +1,203 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include "common.h"
+
+
+#if defined(BULLDOZER) || defined(PILEDRIVER)
+#include "sgemv_n_microk_bulldozer-2.c"
+#endif
+
+
+#define NBMAX 4096
+
+#ifndef HAVE_KERNEL_16x4
+
+static void sgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+	int i;
+	float *a0,*a1,*a2,*a3;
+	a0 = ap[0];
+	a1 = ap[1];
+	a2 = ap[2];
+	a3 = ap[3];
+
+	for ( i=0; i< n; i+=4 )
+	{
+		y[i] += a0[i]*x[0] + a1[i]*x[1] + a2[i]*x[2] + a3[i]*x[3];		
+		y[i+1] += a0[i+1]*x[0] + a1[i+1]*x[1] + a2[i+1]*x[2] + a3[i+1]*x[3];		
+		y[i+2] += a0[i+2]*x[0] + a1[i+2]*x[1] + a2[i+2]*x[2] + a3[i+2]*x[3];		
+		y[i+3] += a0[i+3]*x[0] + a1[i+3]*x[1] + a2[i+3]*x[2] + a3[i+3]*x[3];		
+	}
+}
+	
+#endif
+
+static void sgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
+{
+	int i;
+	float *a0;
+	a0 = ap;
+
+	for ( i=0; i< n; i+=4 )
+	{
+		y[i] += a0[i]*x[0];		
+		y[i+1] += a0[i+1]*x[0];		
+		y[i+2] += a0[i+2]*x[0];		
+		y[i+3] += a0[i+3]*x[0];		
+	}
+}
+	
+
+static void zero_y(BLASLONG n, FLOAT *dest)
+{
+	BLASLONG i;
+	for ( i=0; i<n; i++ )
+	{
+		*dest = 0.0;
+		dest++;
+	}
+}
+
+
+
+static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest)
+{
+	BLASLONG i;
+	if ( inc_dest == 1 )
+	{
+		for ( i=0; i<n; i+=4 )
+		{
+			dest[i] += src[i];
+			dest[i+1] += src[i+1];
+			dest[i+2] += src[i+2];
+			dest[i+3] += src[i+3];
+		}
+
+	}
+	else
+	{
+		for ( i=0; i<n; i++ )
+		{
+			*dest += *src;
+			src++;
+			dest += inc_dest;
+		}
+	}
+}
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG j;
+	FLOAT *a_ptr;
+	FLOAT *x_ptr;
+	FLOAT *y_ptr;
+	FLOAT *ap[4];
+	BLASLONG n1;
+	BLASLONG m1;
+	BLASLONG m2;
+	BLASLONG n2;
+	FLOAT xbuffer[4],*ybuffer;
+	ybuffer = buffer;
+	
+	n1 = n / 4 ;
+	n2 = n % 4 ;
+
+	
+	m1 = m - ( m % 16 );
+	m2 = (m % NBMAX) - (m % 16) ;
+	
+	y_ptr = y;
+
+	BLASLONG NB = NBMAX;
+
+	while ( NB == NBMAX )
+	{
+		
+		m1 -= NB;
+		if ( m1 < 0)
+		{
+			if ( m2 == 0 ) break;	
+			NB = m2;
+		}
+		
+		a_ptr = a;
+		x_ptr = x;
+		zero_y(NB,ybuffer);
+		for( i = 0; i < n1 ; i++)
+		{
+			xbuffer[0] = alpha * x_ptr[0];
+			x_ptr += inc_x;	
+			xbuffer[1] = alpha * x_ptr[0];
+			x_ptr += inc_x;	
+			xbuffer[2] = alpha * x_ptr[0];
+			x_ptr += inc_x;	
+			xbuffer[3] = alpha * x_ptr[0];
+			x_ptr += inc_x;	
+			ap[0] = a_ptr;
+			ap[1] = a_ptr + lda;
+			ap[2] = ap[1] + lda;
+			ap[3] = ap[2] + lda;
+			sgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
+			a_ptr += 4 * lda;
+		}
+
+		for( i = 0; i < n2 ; i++)
+		{
+			xbuffer[0] = alpha * x_ptr[0];
+			x_ptr += inc_x;	
+			sgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
+			a_ptr += 1 * lda;
+
+		}
+		add_y(NB,ybuffer,y_ptr,inc_y);
+		a     += NB;
+		y_ptr += NB * inc_y;
+	}
+	j=0;
+	while ( j < (m % 16))
+	{
+		a_ptr = a;
+		x_ptr = x;
+		FLOAT temp = 0.0;
+		for( i = 0; i < n; i++ )
+		{
+			temp += a_ptr[0] * x_ptr[0];
+			a_ptr += lda;
+			x_ptr += inc_x;
+		}
+		y_ptr[0] += alpha * temp;
+		y_ptr += inc_y;
+		a++;
+		j++;
+	}
+	return(0);
+}
+
+
diff --git a/kernel/x86_64/sgemv_n_microk_bulldozer-2.c b/kernel/x86_64/sgemv_n_microk_bulldozer-2.c
new file mode 100644
index 000000000..7e9ee5cc9
--- /dev/null
+++ b/kernel/x86_64/sgemv_n_microk_bulldozer-2.c
@@ -0,0 +1,99 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void sgemv_kernel_16x4( long n, float **ap, float *x, float *y) __attribute__ ((noinline));
+
+static void sgemv_kernel_16x4( long n, float **ap, float *x, float *y)
+{
+
+	long register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vbroadcastss    (%2), %%xmm12	 \n\t"	// x0 
+	"vbroadcastss   4(%2), %%xmm13	 \n\t"	// x1 
+	"vbroadcastss   8(%2), %%xmm14	 \n\t"	// x2 
+	"vbroadcastss  12(%2), %%xmm15	 \n\t"	// x3 
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vmovups	(%3,%0,4), %%xmm4	 \n\t"	// 4 * y
+	"vmovups      16(%3,%0,4), %%xmm5	 \n\t"	// 4 * y
+	"vmovups      32(%3,%0,4), %%xmm6	 \n\t"	// 4 * y
+	"vmovups      48(%3,%0,4), %%xmm7	 \n\t"	// 4 * y
+
+	"prefetcht0	 192(%4,%0,4)		       \n\t"
+	"vfmaddps %%xmm4,   (%4,%0,4), %%xmm12, %%xmm4 \n\t" 
+	"vfmaddps %%xmm5, 16(%4,%0,4), %%xmm12, %%xmm5 \n\t" 
+	"vfmaddps %%xmm6, 32(%4,%0,4), %%xmm12, %%xmm6 \n\t" 
+	"vfmaddps %%xmm7, 48(%4,%0,4), %%xmm12, %%xmm7 \n\t" 
+	"prefetcht0	 192(%5,%0,4)		       \n\t"
+	"vfmaddps %%xmm4,   (%5,%0,4), %%xmm13, %%xmm4 \n\t" 
+	"vfmaddps %%xmm5, 16(%5,%0,4), %%xmm13, %%xmm5 \n\t" 
+	"vfmaddps %%xmm6, 32(%5,%0,4), %%xmm13, %%xmm6 \n\t" 
+	"vfmaddps %%xmm7, 48(%5,%0,4), %%xmm13, %%xmm7 \n\t" 
+	"prefetcht0	 192(%6,%0,4)		       \n\t"
+	"vfmaddps %%xmm4,   (%6,%0,4), %%xmm14, %%xmm4 \n\t" 
+	"vfmaddps %%xmm5, 16(%6,%0,4), %%xmm14, %%xmm5 \n\t" 
+	"vfmaddps %%xmm6, 32(%6,%0,4), %%xmm14, %%xmm6 \n\t" 
+	"vfmaddps %%xmm7, 48(%6,%0,4), %%xmm14, %%xmm7 \n\t" 
+	"prefetcht0	 192(%7,%0,4)		       \n\t"
+	"vfmaddps %%xmm4,   (%7,%0,4), %%xmm15, %%xmm4 \n\t" 
+	"vfmaddps %%xmm5, 16(%7,%0,4), %%xmm15, %%xmm5 \n\t" 
+	"vfmaddps %%xmm6, 32(%7,%0,4), %%xmm15, %%xmm6 \n\t" 
+	"vfmaddps %%xmm7, 48(%7,%0,4), %%xmm15, %%xmm7 \n\t" 
+
+	"vmovups  %%xmm4,   (%3,%0,4)		      \n\t"	// 4 * y
+	"vmovups  %%xmm5, 16(%3,%0,4)		      \n\t"	// 4 * y
+	"vmovups  %%xmm6, 32(%3,%0,4)		      \n\t"	// 4 * y
+	"vmovups  %%xmm7, 48(%3,%0,4)		      \n\t"	// 4 * y
+
+        "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" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm4", "%xmm5", 
+	  "%xmm6", "%xmm7", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From 3fbc13eb652fbeff928c6259203373bc7b359be2 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Mon, 4 Aug 2014 16:22:11 +0200
Subject: [PATCH 2/9] modified sgemv_n for haswell

---
 kernel/x86_64/KERNEL.HASWELL             |  2 +-
 kernel/x86_64/sgemv_n.c                  |  2 +
 kernel/x86_64/sgemv_n_microk_haswell-2.c | 86 ++++++++++++++++++++++++
 3 files changed, 89 insertions(+), 1 deletion(-)
 create mode 100644 kernel/x86_64/sgemv_n_microk_haswell-2.c

diff --git a/kernel/x86_64/KERNEL.HASWELL b/kernel/x86_64/KERNEL.HASWELL
index 871a7d490..df4aad92f 100644
--- a/kernel/x86_64/KERNEL.HASWELL
+++ b/kernel/x86_64/KERNEL.HASWELL
@@ -2,7 +2,7 @@ ifdef OS_WINDOWS
 SGEMVNKERNEL = ../arm/gemv_n.c
 SGEMVTKERNEL = ../arm/gemv_t.c
 else
-SGEMVNKERNEL = sgemv_n_avx.c
+SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t_avx.c
 endif
 
diff --git a/kernel/x86_64/sgemv_n.c b/kernel/x86_64/sgemv_n.c
index 6f240797d..d1f4d5f60 100644
--- a/kernel/x86_64/sgemv_n.c
+++ b/kernel/x86_64/sgemv_n.c
@@ -31,6 +31,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #if defined(BULLDOZER) || defined(PILEDRIVER)
 #include "sgemv_n_microk_bulldozer-2.c"
+#elif defined(HASWELL)
+#include "sgemv_n_microk_haswell-2.c"
 #endif
 
 
diff --git a/kernel/x86_64/sgemv_n_microk_haswell-2.c b/kernel/x86_64/sgemv_n_microk_haswell-2.c
new file mode 100644
index 000000000..0bad0ec79
--- /dev/null
+++ b/kernel/x86_64/sgemv_n_microk_haswell-2.c
@@ -0,0 +1,86 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void sgemv_kernel_16x4( long n, float **ap, float *x, float *y) __attribute__ ((noinline));
+
+static void sgemv_kernel_16x4( long n, float **ap, float *x, float *y)
+{
+
+	long register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vbroadcastss    (%2), %%ymm12	 \n\t"	// x0 
+	"vbroadcastss   4(%2), %%ymm13	 \n\t"	// x1 
+	"vbroadcastss   8(%2), %%ymm14	 \n\t"	// x2 
+	"vbroadcastss  12(%2), %%ymm15	 \n\t"	// x3 
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vmovups	(%3,%0,4), %%ymm4	 \n\t"	// 8 * y
+	"vmovups      32(%3,%0,4), %%ymm5	 \n\t"	// 8 * y
+
+	"prefetcht0	 192(%4,%0,4)		       \n\t"
+	"vfmadd231ps   (%4,%0,4), %%ymm12, %%ymm4      \n\t" 
+	"vfmadd231ps 32(%4,%0,4), %%ymm12, %%ymm5      \n\t" 
+	"prefetcht0	 192(%5,%0,4)		       \n\t"
+	"vfmadd231ps   (%5,%0,4), %%ymm13, %%ymm4      \n\t" 
+	"vfmadd231ps 32(%5,%0,4), %%ymm13, %%ymm5      \n\t" 
+	"prefetcht0	 192(%6,%0,4)		       \n\t"
+	"vfmadd231ps   (%6,%0,4), %%ymm14, %%ymm4      \n\t" 
+	"vfmadd231ps 32(%6,%0,4), %%ymm14, %%ymm5      \n\t" 
+	"prefetcht0	 192(%7,%0,4)		       \n\t"
+	"vfmadd231ps   (%7,%0,4), %%ymm15, %%ymm4      \n\t" 
+	"vfmadd231ps 32(%7,%0,4), %%ymm15, %%ymm5      \n\t" 
+
+	"vmovups  %%ymm4,   (%3,%0,4)		      \n\t"	// 8 * y
+	"vmovups  %%ymm5, 32(%3,%0,4)		      \n\t"	// 8 * y
+
+        "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" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm4", "%xmm5", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From 7fa7ea3e1e73a79edc6a9facaa573a5f94193827 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Tue, 5 Aug 2014 08:04:47 +0200
Subject: [PATCH 3/9] updated haswell optimized sgmv_n kernel

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

diff --git a/kernel/x86_64/sgemv_n_microk_haswell-2.c b/kernel/x86_64/sgemv_n_microk_haswell-2.c
index 0bad0ec79..b19db9e1e 100644
--- a/kernel/x86_64/sgemv_n_microk_haswell-2.c
+++ b/kernel/x86_64/sgemv_n_microk_haswell-2.c
@@ -35,6 +35,7 @@ static void sgemv_kernel_16x4( long n, float **ap, float *x, float *y)
 
 	__asm__  __volatile__
 	(
+	"vzeroupper			 \n\t"
 	"vbroadcastss    (%2), %%ymm12	 \n\t"	// x0 
 	"vbroadcastss   4(%2), %%ymm13	 \n\t"	// x1 
 	"vbroadcastss   8(%2), %%ymm14	 \n\t"	// x2 
@@ -64,6 +65,7 @@ static void sgemv_kernel_16x4( long n, float **ap, float *x, float *y)
         "addq		$16, %0	  	 	      \n\t"
 	"subq	        $16, %1			      \n\t"		
 	"jnz		.L01LOOP%=		      \n\t"
+	"vzeroupper			 \n\t"
 
 	:
         : 

From a4dde45f879494387f8a6a6015fefc0174b86238 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Tue, 5 Aug 2014 08:53:09 +0200
Subject: [PATCH 4/9] optimized sgemv_n kernel for sandybridge

---
 kernel/x86_64/KERNEL.SANDYBRIDGE       |  2 +-
 kernel/x86_64/sgemv_n.c                |  2 +
 kernel/x86_64/sgemv_n_microk_sandy-2.c | 97 ++++++++++++++++++++++++++
 3 files changed, 100 insertions(+), 1 deletion(-)
 create mode 100644 kernel/x86_64/sgemv_n_microk_sandy-2.c

diff --git a/kernel/x86_64/KERNEL.SANDYBRIDGE b/kernel/x86_64/KERNEL.SANDYBRIDGE
index 9d7a49562..7869e37a8 100644
--- a/kernel/x86_64/KERNEL.SANDYBRIDGE
+++ b/kernel/x86_64/KERNEL.SANDYBRIDGE
@@ -2,7 +2,7 @@ ifdef OS_WINDOWS
 SGEMVNKERNEL = ../arm/gemv_n.c
 SGEMVTKERNEL = ../arm/gemv_t.c
 else
-SGEMVNKERNEL = sgemv_n_avx.c
+SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t_avx.c
 endif
 
diff --git a/kernel/x86_64/sgemv_n.c b/kernel/x86_64/sgemv_n.c
index d1f4d5f60..2f30fd9de 100644
--- a/kernel/x86_64/sgemv_n.c
+++ b/kernel/x86_64/sgemv_n.c
@@ -33,6 +33,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "sgemv_n_microk_bulldozer-2.c"
 #elif defined(HASWELL)
 #include "sgemv_n_microk_haswell-2.c"
+#elif defined(SANDYBRIDGE)
+#include "sgemv_n_microk_sandy-2.c"
 #endif
 
 
diff --git a/kernel/x86_64/sgemv_n_microk_sandy-2.c b/kernel/x86_64/sgemv_n_microk_sandy-2.c
new file mode 100644
index 000000000..dfcb0e17b
--- /dev/null
+++ b/kernel/x86_64/sgemv_n_microk_sandy-2.c
@@ -0,0 +1,97 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void sgemv_kernel_16x4( long n, float **ap, float *x, float *y) __attribute__ ((noinline));
+
+static void sgemv_kernel_16x4( long n, float **ap, float *x, float *y)
+{
+
+	long register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+	"vbroadcastss    (%2), %%ymm12	 \n\t"	// x0 
+	"vbroadcastss   4(%2), %%ymm13	 \n\t"	// x1 
+	"vbroadcastss   8(%2), %%ymm14	 \n\t"	// x2 
+	"vbroadcastss  12(%2), %%ymm15	 \n\t"	// x3 
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vmovups	(%3,%0,4), %%ymm4	 \n\t"	// 8 * y
+	"vmovups      32(%3,%0,4), %%ymm5	 \n\t"	// 8 * y
+
+	"prefetcht0	 192(%4,%0,4)		       \n\t"
+	"vmulps   (%4,%0,4), %%ymm12, %%ymm8      \n\t" 
+	"vaddps	  %%ymm4, %%ymm8 , %%ymm4	  \n\t"
+	"vmulps 32(%4,%0,4), %%ymm12, %%ymm9      \n\t" 
+	"vaddps	  %%ymm5, %%ymm9 , %%ymm5	  \n\t"
+	"prefetcht0	 192(%5,%0,4)		       \n\t"
+	"vmulps   (%5,%0,4), %%ymm13, %%ymm10     \n\t" 
+	"vaddps	  %%ymm4, %%ymm10, %%ymm4	  \n\t"
+	"vmulps 32(%5,%0,4), %%ymm13, %%ymm11     \n\t" 
+	"vaddps	  %%ymm5, %%ymm11, %%ymm5	  \n\t"
+	"prefetcht0	 192(%6,%0,4)		       \n\t"
+	"vmulps   (%6,%0,4), %%ymm14, %%ymm8      \n\t" 
+	"vaddps	  %%ymm4, %%ymm8 , %%ymm4	  \n\t"
+	"vmulps 32(%6,%0,4), %%ymm14, %%ymm9      \n\t" 
+	"vaddps	  %%ymm5, %%ymm9 , %%ymm5	  \n\t"
+	"prefetcht0	 192(%7,%0,4)		       \n\t"
+	"vmulps   (%7,%0,4), %%ymm15, %%ymm10     \n\t" 
+	"vaddps	  %%ymm4, %%ymm10, %%ymm4	  \n\t"
+	"vmulps 32(%7,%0,4), %%ymm15, %%ymm11     \n\t" 
+	"vaddps	  %%ymm5, %%ymm11, %%ymm5	  \n\t"
+
+	"vmovups  %%ymm4,   (%3,%0,4)		      \n\t"	// 8 * y
+	"vmovups  %%ymm5, 32(%3,%0,4)		      \n\t"	// 8 * y
+
+        "addq		$16, %0	  	 	      \n\t"
+	"subq	        $16, %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \n\t"
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm4", "%xmm5", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From 793f2d43b0a557163f556f744f81fd9031779919 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Tue, 5 Aug 2014 10:50:08 +0200
Subject: [PATCH 5/9] added optimized sgemv_n kernel for nehalem

---
 kernel/x86_64/KERNEL.NEHALEM             |   7 ++
 kernel/x86_64/sgemv_n.c                  |   2 +
 kernel/x86_64/sgemv_n_microk_nehalem-2.c | 144 +++++++++++++++++++++++
 3 files changed, 153 insertions(+)
 create mode 100644 kernel/x86_64/sgemv_n_microk_nehalem-2.c

diff --git a/kernel/x86_64/KERNEL.NEHALEM b/kernel/x86_64/KERNEL.NEHALEM
index 2f9c20583..bb1264316 100644
--- a/kernel/x86_64/KERNEL.NEHALEM
+++ b/kernel/x86_64/KERNEL.NEHALEM
@@ -1,3 +1,10 @@
+ifdef OS_WINDOWS
+SGEMVNKERNEL = ../arm/gemv_n.c
+SGEMVTKERNEL = ../arm/gemv_t.c
+else
+SGEMVNKERNEL = sgemv_n.c
+endif
+
 
 SGEMMKERNEL    =  gemm_kernel_4x8_nehalem.S
 SGEMMINCOPY    =  gemm_ncopy_4.S
diff --git a/kernel/x86_64/sgemv_n.c b/kernel/x86_64/sgemv_n.c
index 2f30fd9de..4961deffb 100644
--- a/kernel/x86_64/sgemv_n.c
+++ b/kernel/x86_64/sgemv_n.c
@@ -35,6 +35,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "sgemv_n_microk_haswell-2.c"
 #elif defined(SANDYBRIDGE)
 #include "sgemv_n_microk_sandy-2.c"
+#elif defined(NEHALEM)
+#include "sgemv_n_microk_nehalem-2.c"
 #endif
 
 
diff --git a/kernel/x86_64/sgemv_n_microk_nehalem-2.c b/kernel/x86_64/sgemv_n_microk_nehalem-2.c
new file mode 100644
index 000000000..8499a01a6
--- /dev/null
+++ b/kernel/x86_64/sgemv_n_microk_nehalem-2.c
@@ -0,0 +1,144 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void sgemv_kernel_16x4( long n, float **ap, float *x, float *y) __attribute__ ((noinline));
+
+static void sgemv_kernel_16x4( long n, float **ap, float *x, float *y)
+{
+
+	long register i = 0;
+
+	__asm__  __volatile__
+	(
+	"movss    (%2), %%xmm12	 \n\t"	// x0 
+	"movss   4(%2), %%xmm13	 \n\t"	// x1 
+	"movss   8(%2), %%xmm14	 \n\t"	// x2 
+	"movss  12(%2), %%xmm15	 \n\t"	// x3 
+	"shufps $0,  %%xmm12, %%xmm12\n\t"	
+	"shufps $0,  %%xmm13, %%xmm13\n\t"	
+	"shufps $0,  %%xmm14, %%xmm14\n\t"	
+	"shufps $0,  %%xmm15, %%xmm15\n\t"	
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"movups	       (%3,%0,4), %%xmm4	 \n\t"	// 4 * y
+	"movups      16(%3,%0,4), %%xmm5	 \n\t"	// 4 * y
+	"movups      32(%3,%0,4), %%xmm6	 \n\t"	// 4 * y
+	"movups      48(%3,%0,4), %%xmm7	 \n\t"	// 4 * y
+
+	"prefetcht0	 192(%4,%0,4)		       \n\t"
+
+	"movups             (%4,%0,4), %%xmm8          \n\t" 
+	"movups           16(%4,%0,4), %%xmm9          \n\t" 
+	"movups           32(%4,%0,4), %%xmm10         \n\t" 
+	"movups           48(%4,%0,4), %%xmm11         \n\t" 
+	"mulps		%%xmm12, %%xmm8		       \n\t"
+	"addps		%%xmm8 , %%xmm4		       \n\t"
+	"mulps		%%xmm12, %%xmm9		       \n\t"
+	"addps		%%xmm9 , %%xmm5		       \n\t"
+	"mulps		%%xmm12, %%xmm10	       \n\t"
+	"addps		%%xmm10, %%xmm6		       \n\t"
+	"mulps		%%xmm12, %%xmm11	       \n\t"
+	"addps		%%xmm11, %%xmm7		       \n\t"
+
+	"prefetcht0	 192(%5,%0,4)		       \n\t"
+
+	"movups             (%5,%0,4), %%xmm8          \n\t" 
+	"movups           16(%5,%0,4), %%xmm9          \n\t" 
+	"movups           32(%5,%0,4), %%xmm10         \n\t" 
+	"movups           48(%5,%0,4), %%xmm11         \n\t" 
+	"mulps		%%xmm13, %%xmm8		       \n\t"
+	"addps		%%xmm8 , %%xmm4		       \n\t"
+	"mulps		%%xmm13, %%xmm9		       \n\t"
+	"addps		%%xmm9 , %%xmm5		       \n\t"
+	"mulps		%%xmm13, %%xmm10	       \n\t"
+	"addps		%%xmm10, %%xmm6		       \n\t"
+	"mulps		%%xmm13, %%xmm11	       \n\t"
+	"addps		%%xmm11, %%xmm7		       \n\t"
+
+	"prefetcht0	 192(%6,%0,4)		       \n\t"
+
+	"movups             (%6,%0,4), %%xmm8          \n\t" 
+	"movups           16(%6,%0,4), %%xmm9          \n\t" 
+	"movups           32(%6,%0,4), %%xmm10         \n\t" 
+	"movups           48(%6,%0,4), %%xmm11         \n\t" 
+	"mulps		%%xmm14, %%xmm8		       \n\t"
+	"addps		%%xmm8 , %%xmm4		       \n\t"
+	"mulps		%%xmm14, %%xmm9		       \n\t"
+	"addps		%%xmm9 , %%xmm5		       \n\t"
+	"mulps		%%xmm14, %%xmm10	       \n\t"
+	"addps		%%xmm10, %%xmm6		       \n\t"
+	"mulps		%%xmm14, %%xmm11	       \n\t"
+	"addps		%%xmm11, %%xmm7		       \n\t"
+
+	"prefetcht0	 192(%7,%0,4)		       \n\t"
+
+	"movups             (%7,%0,4), %%xmm8          \n\t" 
+	"movups           16(%7,%0,4), %%xmm9          \n\t" 
+	"movups           32(%7,%0,4), %%xmm10         \n\t" 
+	"movups           48(%7,%0,4), %%xmm11         \n\t" 
+	"mulps		%%xmm15, %%xmm8		       \n\t"
+	"addps		%%xmm8 , %%xmm4		       \n\t"
+	"mulps		%%xmm15, %%xmm9		       \n\t"
+	"addps		%%xmm9 , %%xmm5		       \n\t"
+	"mulps		%%xmm15, %%xmm10	       \n\t"
+	"addps		%%xmm10, %%xmm6		       \n\t"
+	"mulps		%%xmm15, %%xmm11	       \n\t"
+	"addps		%%xmm11, %%xmm7		       \n\t"
+
+
+	"movups  %%xmm4,   (%3,%0,4)		      \n\t"	// 4 * y
+	"movups  %%xmm5, 16(%3,%0,4)		      \n\t"	// 4 * y
+	"movups  %%xmm6, 32(%3,%0,4)		      \n\t"	// 4 * y
+	"movups  %%xmm7, 48(%3,%0,4)		      \n\t"	// 4 * y
+
+        "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" (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 9175b8bd5f8df59146b8539f9760d7ed42611648 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Tue, 5 Aug 2014 13:28:39 +0200
Subject: [PATCH 6/9] changed long to blaslong for windows compatibility

---
 kernel/x86_64/sgemv_n_microk_bulldozer-2.c | 6 +++---
 kernel/x86_64/sgemv_n_microk_haswell-2.c   | 6 +++---
 kernel/x86_64/sgemv_n_microk_nehalem-2.c   | 6 +++---
 kernel/x86_64/sgemv_n_microk_sandy-2.c     | 6 +++---
 4 files changed, 12 insertions(+), 12 deletions(-)

diff --git a/kernel/x86_64/sgemv_n_microk_bulldozer-2.c b/kernel/x86_64/sgemv_n_microk_bulldozer-2.c
index 7e9ee5cc9..d50fa4268 100644
--- a/kernel/x86_64/sgemv_n_microk_bulldozer-2.c
+++ b/kernel/x86_64/sgemv_n_microk_bulldozer-2.c
@@ -26,12 +26,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *****************************************************************************/
 
 #define HAVE_KERNEL_16x4 1
-static void sgemv_kernel_16x4( long n, float **ap, float *x, float *y) __attribute__ ((noinline));
+static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y) __attribute__ ((noinline));
 
-static void sgemv_kernel_16x4( long n, float **ap, float *x, float *y)
+static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
 {
 
-	long register i = 0;
+	BLASLONG register i = 0;
 
 	__asm__  __volatile__
 	(
diff --git a/kernel/x86_64/sgemv_n_microk_haswell-2.c b/kernel/x86_64/sgemv_n_microk_haswell-2.c
index b19db9e1e..d3fee67c3 100644
--- a/kernel/x86_64/sgemv_n_microk_haswell-2.c
+++ b/kernel/x86_64/sgemv_n_microk_haswell-2.c
@@ -26,12 +26,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *****************************************************************************/
 
 #define HAVE_KERNEL_16x4 1
-static void sgemv_kernel_16x4( long n, float **ap, float *x, float *y) __attribute__ ((noinline));
+static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y) __attribute__ ((noinline));
 
-static void sgemv_kernel_16x4( long n, float **ap, float *x, float *y)
+static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
 {
 
-	long register i = 0;
+	BLASLONG register i = 0;
 
 	__asm__  __volatile__
 	(
diff --git a/kernel/x86_64/sgemv_n_microk_nehalem-2.c b/kernel/x86_64/sgemv_n_microk_nehalem-2.c
index 8499a01a6..3cfb82a45 100644
--- a/kernel/x86_64/sgemv_n_microk_nehalem-2.c
+++ b/kernel/x86_64/sgemv_n_microk_nehalem-2.c
@@ -26,12 +26,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *****************************************************************************/
 
 #define HAVE_KERNEL_16x4 1
-static void sgemv_kernel_16x4( long n, float **ap, float *x, float *y) __attribute__ ((noinline));
+static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y) __attribute__ ((noinline));
 
-static void sgemv_kernel_16x4( long n, float **ap, float *x, float *y)
+static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
 {
 
-	long register i = 0;
+	BLASLONG register i = 0;
 
 	__asm__  __volatile__
 	(
diff --git a/kernel/x86_64/sgemv_n_microk_sandy-2.c b/kernel/x86_64/sgemv_n_microk_sandy-2.c
index dfcb0e17b..21eff1c5e 100644
--- a/kernel/x86_64/sgemv_n_microk_sandy-2.c
+++ b/kernel/x86_64/sgemv_n_microk_sandy-2.c
@@ -26,12 +26,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *****************************************************************************/
 
 #define HAVE_KERNEL_16x4 1
-static void sgemv_kernel_16x4( long n, float **ap, float *x, float *y) __attribute__ ((noinline));
+static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y) __attribute__ ((noinline));
 
-static void sgemv_kernel_16x4( long n, float **ap, float *x, float *y)
+static void sgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
 {
 
-	long register i = 0;
+	BLASLONG register i = 0;
 
 	__asm__  __volatile__
 	(

From 2bab92961f5031f8881d5d6635e470b6e7c0a416 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Tue, 5 Aug 2014 14:52:54 +0200
Subject: [PATCH 7/9] enabled optimized sgemv_n kernels for windows

---
 kernel/x86_64/KERNEL.BULLDOZER   | 2 +-
 kernel/x86_64/KERNEL.HASWELL     | 2 +-
 kernel/x86_64/KERNEL.NEHALEM     | 3 ++-
 kernel/x86_64/KERNEL.PILEDRIVER  | 2 +-
 kernel/x86_64/KERNEL.SANDYBRIDGE | 2 +-
 5 files changed, 6 insertions(+), 5 deletions(-)

diff --git a/kernel/x86_64/KERNEL.BULLDOZER b/kernel/x86_64/KERNEL.BULLDOZER
index ca7a0b7c4..21fc94701 100644
--- a/kernel/x86_64/KERNEL.BULLDOZER
+++ b/kernel/x86_64/KERNEL.BULLDOZER
@@ -1,5 +1,5 @@
 ifdef OS_WINDOWS
-SGEMVNKERNEL = ../arm/gemv_n.c
+SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = ../arm/gemv_t.c
 else
 SGEMVNKERNEL = sgemv_n.c
diff --git a/kernel/x86_64/KERNEL.HASWELL b/kernel/x86_64/KERNEL.HASWELL
index df4aad92f..3e20fcfc7 100644
--- a/kernel/x86_64/KERNEL.HASWELL
+++ b/kernel/x86_64/KERNEL.HASWELL
@@ -1,5 +1,5 @@
 ifdef OS_WINDOWS
-SGEMVNKERNEL = ../arm/gemv_n.c
+SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = ../arm/gemv_t.c
 else
 SGEMVNKERNEL = sgemv_n.c
diff --git a/kernel/x86_64/KERNEL.NEHALEM b/kernel/x86_64/KERNEL.NEHALEM
index bb1264316..04efa391a 100644
--- a/kernel/x86_64/KERNEL.NEHALEM
+++ b/kernel/x86_64/KERNEL.NEHALEM
@@ -1,8 +1,9 @@
 ifdef OS_WINDOWS
-SGEMVNKERNEL = ../arm/gemv_n.c
+SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = ../arm/gemv_t.c
 else
 SGEMVNKERNEL = sgemv_n.c
+SGEMVTKERNEL = ../arm/gemv_t.c
 endif
 
 
diff --git a/kernel/x86_64/KERNEL.PILEDRIVER b/kernel/x86_64/KERNEL.PILEDRIVER
index b9d680c43..b7565edeb 100644
--- a/kernel/x86_64/KERNEL.PILEDRIVER
+++ b/kernel/x86_64/KERNEL.PILEDRIVER
@@ -1,5 +1,5 @@
 ifdef OS_WINDOWS
-SGEMVNKERNEL = ../arm/gemv_n.c
+SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = ../arm/gemv_t.c
 else
 SGEMVNKERNEL = sgemv_n.c
diff --git a/kernel/x86_64/KERNEL.SANDYBRIDGE b/kernel/x86_64/KERNEL.SANDYBRIDGE
index 7869e37a8..9dae3a41d 100644
--- a/kernel/x86_64/KERNEL.SANDYBRIDGE
+++ b/kernel/x86_64/KERNEL.SANDYBRIDGE
@@ -1,5 +1,5 @@
 ifdef OS_WINDOWS
-SGEMVNKERNEL = ../arm/gemv_n.c
+SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = ../arm/gemv_t.c
 else
 SGEMVNKERNEL = sgemv_n.c

From c80084a98f2882a7bb959bad26cc66ebd65bef63 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Tue, 5 Aug 2014 19:42:56 +0200
Subject: [PATCH 8/9] changed default x86_64 sgemv_n kernel to sgemv_n.c

---
 kernel/x86_64/KERNEL | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/kernel/x86_64/KERNEL b/kernel/x86_64/KERNEL
index ec21826d7..9d0080ae7 100644
--- a/kernel/x86_64/KERNEL
+++ b/kernel/x86_64/KERNEL
@@ -373,7 +373,7 @@ endif
 GEMVDEP = ../l2param.h
 
 ifndef SGEMVNKERNEL
-SGEMVNKERNEL = ../arm/gemv_n.c
+SGEMVNKERNEL = sgemv_n.c
 endif
 
 ifndef SGEMVTKERNEL

From 8c05b8105b1dad9b9f1ae16d7e0774ce153b6bdd Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Tue, 5 Aug 2014 20:14:29 +0200
Subject: [PATCH 9/9] bugfix in sgemv_n.c

---
 kernel/x86_64/sgemv_n.c | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/kernel/x86_64/sgemv_n.c b/kernel/x86_64/sgemv_n.c
index 4961deffb..f2de1b76a 100644
--- a/kernel/x86_64/sgemv_n.c
+++ b/kernel/x86_64/sgemv_n.c
@@ -46,8 +46,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 static void sgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 {
-	int i;
-	float *a0,*a1,*a2,*a3;
+	BLASLONG i;
+	FLOAT *a0,*a1,*a2,*a3;
 	a0 = ap[0];
 	a1 = ap[1];
 	a2 = ap[2];
@@ -66,8 +66,8 @@ static void sgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 
 static void sgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
 {
-	int i;
-	float *a0;
+	BLASLONG i;
+	FLOAT *a0;
 	a0 = ap;
 
 	for ( i=0; i< n; i+=4 )
@@ -130,11 +130,11 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
 	BLASLONG m2;
 	BLASLONG n2;
 	FLOAT xbuffer[4],*ybuffer;
+
 	ybuffer = buffer;
 	
 	n1 = n / 4 ;
 	n2 = n % 4 ;
-
 	
 	m1 = m - ( m % 16 );
 	m2 = (m % NBMAX) - (m % 16) ;