From bcb115b55b92b75d862d743629b29ac0b84d2fd0 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Tue, 15 Jul 2014 13:35:36 +0200
Subject: [PATCH 01/15] added benchmark for gemv

---
 benchmark/Makefile |  77 ++++++++++++++-
 benchmark/gemv.c   | 229 +++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 302 insertions(+), 4 deletions(-)
 create mode 100644 benchmark/gemv.c

diff --git a/benchmark/Makefile b/benchmark/Makefile
index e3910ee96..db183c8ad 100644
--- a/benchmark/Makefile
+++ b/benchmark/Makefile
@@ -2,12 +2,12 @@ TOPDIR	= ..
 include $(TOPDIR)/Makefile.system
 
 # ACML standard
-ACML=/opt/acml5.3.1/gfortran64_mp/lib
-LIBACML = -fopenmp $(ACML)/libacml_mp.a -lgfortran -lm
+#ACML=/opt/acml5.3.1/gfortran64_mp/lib
+#LIBACML = -fopenmp $(ACML)/libacml_mp.a -lgfortran -lm
 
 # ACML custom
-#ACML=/opt/pb/acml-5-3-1-gfortran-64bit/gfortran64_fma4_mp/lib
-#LIBACML = -fopenmp $(ACML)/libacml_mp.a -lgfortran -lm
+ACML=/opt/pb/acml-5-3-1-gfortran-64bit/gfortran64_fma4_mp/lib
+LIBACML = -fopenmp $(ACML)/libacml_mp.a -lgfortran -lm
 
 # Atlas Ubuntu 
 #ATLAS=/usr/lib/atlas-base
@@ -37,6 +37,7 @@ goto :: slinpack.goto dlinpack.goto clinpack.goto zlinpack.goto \
        chemm.goto zhemm.goto \
        cherk.goto zherk.goto \
        cher2k.goto zher2k.goto \
+       sgemv.goto dgemv.goto cgemv.goto zgemv.goto \
        ssymm.goto dsymm.goto csymm.goto zsymm.goto
 
 acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \
@@ -49,6 +50,7 @@ acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \
        chemm.acml zhemm.acml \
        cherk.acml zherk.acml \
        cher2k.acml zher2k.acml \
+       sgemv.acml dgemv.acml cgemv.acml zgemv.acml \
        ssymm.acml dsymm.acml csymm.acml zsymm.acml
 
 atlas :: slinpack.atlas dlinpack.atlas clinpack.atlas zlinpack.atlas \
@@ -61,6 +63,7 @@ atlas :: slinpack.atlas dlinpack.atlas clinpack.atlas zlinpack.atlas \
        chemm.atlas zhemm.atlas \
        cherk.atlas zherk.atlas \
        cher2k.atlas zher2k.atlas \
+       sgemv.atlas dgemv.atlas cgemv.atlas zgemv.atlas \
        ssymm.atlas dsymm.atlas csymm.atlas zsymm.atlas
 
 mkl :: slinpack.mkl dlinpack.mkl clinpack.mkl zlinpack.mkl \
@@ -73,6 +76,7 @@ mkl :: slinpack.mkl dlinpack.mkl clinpack.mkl zlinpack.mkl \
        chemm.mkl zhemm.mkl \
        cherk.mkl zherk.mkl \
        cher2k.mkl zher2k.mkl \
+       sgemv.mkl dgemv.mkl cgemv.mkl zgemv.mkl \
        ssymm.mkl dsymm.mkl csymm.mkl zsymm.mkl
 
 all :: goto atlas acml mkl
@@ -601,6 +605,61 @@ zher2k.atlas : zher2k.$(SUFFIX)
 zher2k.mkl : zher2k.$(SUFFIX)
 	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
 
+##################################### Sgemv ####################################################
+sgemv.goto : sgemv.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+sgemv.acml : sgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+sgemv.atlas : sgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+sgemv.mkl : sgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Dgemv ####################################################
+dgemv.goto : dgemv.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+dgemv.acml : dgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+dgemv.atlas : dgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+dgemv.mkl : dgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Cgemv ####################################################
+
+cgemv.goto : cgemv.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+cgemv.acml : cgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+cgemv.atlas : cgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+cgemv.mkl : cgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Zgemv ####################################################
+
+zgemv.goto : zgemv.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+zgemv.acml : zgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+zgemv.atlas : zgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+zgemv.mkl : zgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+
 ###################################################################################################
 
 slinpack.$(SUFFIX) : linpack.c
@@ -717,7 +776,17 @@ cher2k.$(SUFFIX) : her2k.c
 zher2k.$(SUFFIX) : her2k.c
 	$(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^
 
+sgemv.$(SUFFIX) : gemv.c
+	$(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^
 
+dgemv.$(SUFFIX) : gemv.c
+	$(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^
+
+cgemv.$(SUFFIX) : gemv.c
+	$(CC) $(CFLAGS) -c -DCOMPLEX -UDOUBLE -o $(@F) $^
+
+zgemv.$(SUFFIX) : gemv.c
+	$(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^
 
 
 clean ::
diff --git a/benchmark/gemv.c b/benchmark/gemv.c
new file mode 100644
index 000000000..e26a36ac1
--- /dev/null
+++ b/benchmark/gemv.c
@@ -0,0 +1,229 @@
+/***************************************************************************
+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 GEMV
+
+#ifndef COMPLEX
+
+#ifdef DOUBLE
+#define GEMV   BLASFUNC(dgemv)
+#else
+#define GEMV   BLASFUNC(sgemv)
+#endif
+
+#else
+
+#ifdef DOUBLE
+#define GEMV   BLASFUNC(zgemv)
+#else
+#define GEMV   BLASFUNC(cgemv)
+#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 *a, *x, *y;
+  FLOAT alpha[] = {1.0, 1.0};
+  FLOAT beta [] = {1.0, 1.0};
+  char trans='N';
+  blasint m, i, j;
+  blasint inc_x=1,inc_y=1;
+  blasint n=0;
+  int has_param_n = 0;
+  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_TRANS")))  trans=*p;
+  if ((p = getenv("OPENBLAS_PARAM_N"))) {
+	  n = atoi(p);
+	  if ((n>0) && (n<=to)) has_param_n = 1;
+  }
+
+  if ( has_param_n == 1 )
+    fprintf(stderr, "From : %3d  To : %3d Step = %3d Trans = '%c' N = %d Inc_x = %d Inc_y = %d Loops = %d\n", from, to, step,trans,n,inc_x,inc_y,loops);
+  else
+    fprintf(stderr, "From : %3d  To : %3d Step = %3d Trans = '%c' Inc_x = %d Inc_y = %d Loops = %d\n", from, to, step,trans,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;
+
+   if ( has_param_n == 0 ) n = m;
+
+   fprintf(stderr, " %6dx%d : ", (int)m,(int)n);
+
+   for(j = 0; j < m; j++){
+      		for(i = 0; i < n * COMPSIZE; i++){
+			a[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+      		}
+   }
+
+
+    for (l=0; l<loops; l++)
+    {
+
+   	for(i = 0; i < n * COMPSIZE * abs(inc_x); i++){
+			x[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+   	}
+
+   	for(i = 0; i < n * COMPSIZE * abs(inc_y); i++){
+			y[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+   	}
+    	gettimeofday( &start, (struct timezone *)0);
+
+    	GEMV (&trans, &m, &n, 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)n / timeg * 1.e-6);
+
+  }
+
+  return 0;
+}
+
+void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));

From d286daa2ba4a6461b0e587592459db6765d3745b Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Tue, 15 Jul 2014 14:41:35 +0200
Subject: [PATCH 02/15] adjusted number of threads for small size

---
 interface/gemm.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/interface/gemm.c b/interface/gemm.c
index 07fea153c..74908e842 100644
--- a/interface/gemm.c
+++ b/interface/gemm.c
@@ -405,11 +405,11 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANS
 
 #ifndef COMPLEX
   double MNK = (double) args.m * (double) args.n * (double) args.k;
-  if ( MNK <= (1024.0  * (double) GEMM_MULTITHREAD_THRESHOLD)  )
+  if ( MNK <= (16.0 * 1024.0  * (double) GEMM_MULTITHREAD_THRESHOLD)  )
 	nthreads_max = 1;
   else
   {
-  	if ( MNK <= (65536.0 * (double) GEMM_MULTITHREAD_THRESHOLD) )
+  	if ( MNK <= (2.0 * 65536.0 * (double) GEMM_MULTITHREAD_THRESHOLD) )
 	{
 		nthreads_max = 4;
 		if ( args.m < 16 * GEMM_MULTITHREAD_THRESHOLD )

From b985cea65dbb4d60b51d204fd6144741fb9a7f0b Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Tue, 15 Jul 2014 16:04:46 +0200
Subject: [PATCH 03/15] adjust number of threads for sgemv and dgemv

---
 interface/gemv.c | 13 ++++++++++++-
 1 file changed, 12 insertions(+), 1 deletion(-)

diff --git a/interface/gemv.c b/interface/gemv.c
index 562ceee9f..08553ad21 100644
--- a/interface/gemv.c
+++ b/interface/gemv.c
@@ -211,7 +211,18 @@ void CNAME(enum CBLAS_ORDER order,
   buffer = (FLOAT *)blas_memory_alloc(1);
 
 #ifdef SMP
-  nthreads = num_cpu_avail(2);
+
+  int  nthreads_max = num_cpu_avail(2);
+  int  nthreads_avail = nthreads_max;
+
+  double MNK = (double) m * (double) n;
+  if ( MNK <= (500.0 * 100.0  * (double) GEMM_MULTITHREAD_THRESHOLD)  )
+        nthreads_max = 1;
+
+  if ( nthreads_max > nthreads_avail )
+        nthreads = nthreads_avail;
+  else
+        nthreads = nthreads_max;
 
   if (nthreads == 1) {
 #endif

From 51413925bdcc1fceec46e58acbd7cf03b7762aa1 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Tue, 15 Jul 2014 16:27:02 +0200
Subject: [PATCH 04/15] adjust number of threads for small size in cgemv and
 zgemv

---
 interface/zgemv.c | 14 +++++++++++++-
 1 file changed, 13 insertions(+), 1 deletion(-)

diff --git a/interface/zgemv.c b/interface/zgemv.c
index fcc2fda54..50513a8e4 100644
--- a/interface/zgemv.c
+++ b/interface/zgemv.c
@@ -233,7 +233,19 @@ void CNAME(enum CBLAS_ORDER order,
   buffer = (FLOAT *)blas_memory_alloc(1);
 
 #ifdef SMP
-  nthreads = num_cpu_avail(2);
+
+  int  nthreads_max = num_cpu_avail(2);
+  int  nthreads_avail = nthreads_max;
+
+  double MNK = (double) m * (double) n;
+  if ( MNK <= (80.0 * 20.0  * (double) GEMM_MULTITHREAD_THRESHOLD)  )
+        nthreads_max = 1;
+
+  if ( nthreads_max > nthreads_avail )
+        nthreads = nthreads_avail;
+  else
+        nthreads = nthreads_max;
+
 
   if (nthreads == 1) {
 #endif

From 7ceb25d7b370f87f89ee900a47015d33dcaab8bf Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Wed, 16 Jul 2014 17:08:43 +0200
Subject: [PATCH 05/15] changed string GFORTRAN to lowercase

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

diff --git a/Makefile b/Makefile
index 2e378883b..3aaf092fc 100644
--- a/Makefile
+++ b/Makefile
@@ -247,7 +247,7 @@ ifndef NOFORTRAN
 	-@echo "SUFFIX      = $(SUFFIX)" >> $(NETLIB_LAPACK_DIR)/make.inc
 	-@echo "PSUFFIX     = $(PSUFFIX)" >> $(NETLIB_LAPACK_DIR)/make.inc
 	-@echo "CEXTRALIB   = $(EXTRALIB)" >> $(NETLIB_LAPACK_DIR)/make.inc
-ifeq ($(FC), GFORTRAN)
+ifeq ($(FC), gfortran)
 	-@echo "TIMER       = INT_ETIME" >> $(NETLIB_LAPACK_DIR)/make.inc
 ifdef SMP
 	-@echo "LOADER      = $(FC) -pthread" >> $(NETLIB_LAPACK_DIR)/make.inc

From 3c5732615dd01e4d865d6c5d516e75889b165347 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Thu, 17 Jul 2014 23:15:07 +0200
Subject: [PATCH 06/15] added blocked sgemv_n and microkernel for bulldozer and
 piledriver

---
 kernel/x86_64/KERNEL.BULLDOZER           |   1 +
 kernel/x86_64/KERNEL.PILEDRIVER          |   1 +
 kernel/x86_64/sgemv_n_avx.c              | 194 +++++++++++++
 kernel/x86_64/sgemv_n_microk_bulldozer.c | 346 +++++++++++++++++++++++
 4 files changed, 542 insertions(+)
 create mode 100644 kernel/x86_64/sgemv_n_avx.c
 create mode 100644 kernel/x86_64/sgemv_n_microk_bulldozer.c

diff --git a/kernel/x86_64/KERNEL.BULLDOZER b/kernel/x86_64/KERNEL.BULLDOZER
index 70370a73c..55932e69f 100644
--- a/kernel/x86_64/KERNEL.BULLDOZER
+++ b/kernel/x86_64/KERNEL.BULLDOZER
@@ -1,3 +1,4 @@
+SGEMVNKERNEL = sgemv_n_avx.c
 
 ZGEMVNKERNEL = zgemv_n_dup.S
 ZGEMVTKERNEL = zgemv_t.S
diff --git a/kernel/x86_64/KERNEL.PILEDRIVER b/kernel/x86_64/KERNEL.PILEDRIVER
index 92b5dc7c9..145d9fb2f 100644
--- a/kernel/x86_64/KERNEL.PILEDRIVER
+++ b/kernel/x86_64/KERNEL.PILEDRIVER
@@ -1,3 +1,4 @@
+SGEMVNKERNEL = sgemv_n_avx.c
 
 ZGEMVNKERNEL = zgemv_n_dup.S
 ZGEMVTKERNEL = zgemv_t.S
diff --git a/kernel/x86_64/sgemv_n_avx.c b/kernel/x86_64/sgemv_n_avx.c
new file mode 100644
index 000000000..8c263543c
--- /dev/null
+++ b/kernel/x86_64/sgemv_n_avx.c
@@ -0,0 +1,194 @@
+/***************************************************************************
+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.c"
+#endif
+
+static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
+{
+	BLASLONG i;
+	for ( i=0; i<n; i++ )
+	{
+		*dest = *src;
+		dest++;
+		src += inc_src;
+	}
+}
+
+static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest)
+{
+	BLASLONG i;
+	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;
+	BLASLONG n1,n2;
+	BLASLONG m1,m2;
+	FLOAT *xbuffer,*ybuffer;
+	xbuffer = buffer;
+	ybuffer = xbuffer + 2048 + 256;
+	
+	n1 = n / 512 ;
+	n2 = n % 512 ;
+
+	m1 = m / 32;
+	m2 = m % 32;
+
+	x_ptr = x;
+	a_ptr = a;
+	y_ptr = y;
+
+	for (j=0; j<n1; j++)
+	{
+
+		if ( inc_x == 1 )
+			xbuffer = x_ptr;
+		else
+			copy_x(512,x_ptr,xbuffer,inc_x);
+
+		x_ptr += 512 * inc_x;
+		a_ptr += j * 512;
+		y_ptr = y;
+
+
+		for(i = 0; i<m1; i++ )
+		{
+			sgemv_kernel_32(512,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(32,ybuffer,y_ptr,inc_y);
+			y_ptr += 32 * inc_y;
+			a_ptr += 32;			
+
+		}
+		if ( m2 & 16 )
+		{
+			sgemv_kernel_16(512,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(16,ybuffer,y_ptr,inc_y);
+			y_ptr += 16 * inc_y;
+			a_ptr += 16;			
+		}
+		if ( m2 & 8 )
+		{
+			sgemv_kernel_8(512,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(8,ybuffer,y_ptr,inc_y);
+			y_ptr += 8 * inc_y;
+			a_ptr += 8;			
+		}
+		if ( m2 & 4 )
+		{
+			sgemv_kernel_4(512,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(4,ybuffer,y_ptr,inc_y);
+			y_ptr += 4 * inc_y;
+			a_ptr += 4;			
+		}
+		if ( m2 & 2 )
+		{
+			sgemv_kernel_2(512,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(2,ybuffer,y_ptr,inc_y);
+			y_ptr += 2 * inc_y;
+			a_ptr += 2;			
+		}
+		if ( m2 & 1 )
+		{
+			sgemv_kernel_1(512,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(1,ybuffer,y_ptr,inc_y);
+		}
+
+	}
+
+	if ( n2 > 0 )
+	{
+
+		if ( inc_x == 1 )
+			xbuffer = x_ptr;
+		else
+			copy_x(n2,x_ptr,xbuffer,inc_x);
+
+		y_ptr = y;
+
+		for(i = 0; i<m1; i++ )
+		{
+			sgemv_kernel_32(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(32,ybuffer,y_ptr,inc_y);
+			y_ptr += 32 * inc_y;
+			a_ptr += 32;			
+
+		}
+		if ( m2 & 16 )
+		{
+			sgemv_kernel_16(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(16,ybuffer,y_ptr,inc_y);
+			y_ptr += 16 * inc_y;
+			a_ptr += 16;			
+		}
+		if ( m2 & 8 )
+		{
+			sgemv_kernel_8(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(8,ybuffer,y_ptr,inc_y);
+			y_ptr += 8 * inc_y;
+			a_ptr += 8;			
+		}
+		if ( m2 & 4 )
+		{
+			sgemv_kernel_4(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(4,ybuffer,y_ptr,inc_y);
+			y_ptr += 4 * inc_y;
+			a_ptr += 4;			
+		}
+		if ( m2 & 2 )
+		{
+			sgemv_kernel_2(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(2,ybuffer,y_ptr,inc_y);
+			y_ptr += 2 * inc_y;
+			a_ptr += 2;			
+		}
+		if ( m2 & 1 )
+		{
+			sgemv_kernel_1(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(1,ybuffer,y_ptr,inc_y);
+		}
+
+
+	}
+	return(0);
+}
+
+
diff --git a/kernel/x86_64/sgemv_n_microk_bulldozer.c b/kernel/x86_64/sgemv_n_microk_bulldozer.c
new file mode 100644
index 000000000..3dad43643
--- /dev/null
+++ b/kernel/x86_64/sgemv_n_microk_bulldozer.c
@@ -0,0 +1,346 @@
+/***************************************************************************
+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.
+*****************************************************************************/
+
+
+static void sgemv_kernel_32( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	float *pre = a + lda*4*2;
+
+	__asm __volatile
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+	"movq		%6,	 %%r8\n\t"		// address for prefetch
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+
+	"vxorps		%%ymm12, %%ymm12, %%ymm12\n\t"	// set to zero
+	"vxorps		%%ymm13, %%ymm13, %%ymm13\n\t"	// set to zero
+	"vxorps		%%ymm14, %%ymm14, %%ymm14\n\t"	// set to zero
+	"vxorps		%%ymm15, %%ymm15, %%ymm15\n\t"	// set to zero
+
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%ymm0	 \n\t"	// load values of c
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+
+	"leaq     (%%r8 , %%rcx, 4), %%r8	 \n\t"	// add lda to pointer for prefetch
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+
+	"vfmaddps %%ymm12,   0*4(%%rsi), %%ymm0, %%ymm12\n\t" // multiply a and c and add to temp
+	"vfmaddps %%ymm13,   8*4(%%rsi), %%ymm0, %%ymm13\n\t" // multiply a and c and add to temp
+	"vfmaddps %%ymm14,  16*4(%%rsi), %%ymm0, %%ymm14\n\t" // multiply a and c and add to temp
+	"vfmaddps %%ymm15,  24*4(%%rsi), %%ymm0, %%ymm15\n\t" // multiply a and c and add to temp
+
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%ymm12, %%ymm1,  %%ymm12\n\t"  // scale by alpha
+	"vmulps		%%ymm13, %%ymm1,  %%ymm13\n\t"  // scale by alpha
+	"vmulps		%%ymm14, %%ymm1,  %%ymm14\n\t"  // scale by alpha
+	"vmulps		%%ymm15, %%ymm1,  %%ymm15\n\t"  // scale by alpha
+
+	"vmovups	%%ymm12,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm13,  8*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm14, 16*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm15, 24*4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y),      // 5
+	  "m" (pre)	// 6
+	: "rax", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11",
+	  "xmm0" , "xmm1", 
+	  "xmm12", "xmm13", "xmm14", "xmm15",
+	  "memory"
+	);
+
+} 
+
+static void sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+	float *pre = a + lda*4*3;
+
+	__asm __volatile
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+	"movq		%6,	 %%r8\n\t"		// address for prefetch
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+
+	"vxorps		%%ymm12, %%ymm12, %%ymm12\n\t"	// set to zero
+	"vxorps		%%ymm13, %%ymm13, %%ymm13\n\t"	// set to zero
+
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%ymm0	 \n\t"	// load values of c
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+
+	"leaq     (%%r8 , %%rcx, 4), %%r8	 \n\t"	// add lda to pointer for prefetch
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+
+	"vfmaddps %%ymm12,   0*4(%%rsi), %%ymm0, %%ymm12\n\t" // multiply a and c and add to temp
+	"vfmaddps %%ymm13,   8*4(%%rsi), %%ymm0, %%ymm13\n\t" // multiply a and c and add to temp
+
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%ymm12, %%ymm1,  %%ymm12\n\t"  // scale by alpha
+	"vmulps		%%ymm13, %%ymm1,  %%ymm13\n\t"  // scale by alpha
+
+	"vmovups	%%ymm12,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm13,  8*4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y),      // 5
+	  "m" (pre)	// 6
+	: "rax", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11",
+	  "xmm0" , "xmm1", 
+	  "xmm12", "xmm13", "xmm14", "xmm15",
+	  "memory"
+	);
+
+} 
+
+
+static void sgemv_kernel_8( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	__asm __volatile
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%ymm12, %%ymm12, %%ymm12\n\t"	// set to zero
+
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%ymm0	 \n\t"	// load values of c
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+
+	"vfmaddps %%ymm12,   0*4(%%rsi), %%ymm0, %%ymm12\n\t" // multiply a and c and add to temp
+
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%ymm12, %%ymm1,  %%ymm12\n\t"  // scale by alpha
+
+	"vmovups	%%ymm12,     (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "rax", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11",
+	  "xmm0" , "xmm1", 
+	  "xmm12", "xmm13", "xmm14", "xmm15",
+	  "memory"
+	);
+
+} 
+
+
+static void sgemv_kernel_4( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	__asm __volatile
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%xmm0	 \n\t"	// load values of c
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+
+	"vfmaddps %%xmm12,   0*4(%%rsi), %%xmm0, %%xmm12\n\t" // multiply a and c and add to temp
+
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%xmm12, %%xmm1,  %%xmm12\n\t"  // scale by alpha
+
+	"vmovups	%%xmm12,     (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "rax", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11",
+	  "xmm0" , "xmm1", 
+	  "xmm12", "xmm13", "xmm14", "xmm15",
+	  "memory"
+	);
+
+} 
+
+static void sgemv_kernel_2( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	__asm __volatile
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vmovss         %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+	"vxorps		%%xmm13, %%xmm13, %%xmm13\n\t"	// set to zero
+
+	".L01LOOP%=:				 \n\t"
+	"vmovss      	(%%rdi),   %%xmm0	 \n\t"	// load values of c
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+
+	"vfmaddss %%xmm12,   0*4(%%rsi), %%xmm0, %%xmm12\n\t" // multiply a and c and add to temp
+	"vfmaddss %%xmm13,   1*4(%%rsi), %%xmm0, %%xmm13\n\t" // multiply a and c and add to temp
+
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulss		%%xmm12, %%xmm1,  %%xmm12\n\t"  // scale by alpha
+	"vmulss		%%xmm13, %%xmm1,  %%xmm13\n\t"  // scale by alpha
+
+	"vmovss 	%%xmm12,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovss 	%%xmm13,    4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "rax", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11",
+	  "xmm0" , "xmm1", 
+	  "xmm12", "xmm13", "xmm14", "xmm15",
+	  "memory"
+	);
+
+} 
+
+
+
+static void sgemv_kernel_1( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	__asm __volatile
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vmovss         %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+
+	".L01LOOP%=:				 \n\t"
+	"vmovss      	(%%rdi),   %%xmm0	 \n\t"	// load values of c
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+
+	"vfmaddss %%xmm12,   0*4(%%rsi), %%xmm0, %%xmm12\n\t" // multiply a and c and add to temp
+
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulss		%%xmm12, %%xmm1,  %%xmm12\n\t"  // scale by alpha
+
+	"vmovss 	%%xmm12,     (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "rax", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11",
+	  "xmm0" , "xmm1", 
+	  "xmm12", "xmm13", "xmm14", "xmm15",
+	  "memory"
+	);
+
+} 
+
+

From c8a4a561773dcd4b905b7618b2518539d467daaf Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Fri, 18 Jul 2014 11:25:21 +0200
Subject: [PATCH 07/15] performance optimizations for sgemv_n

---
 kernel/x86_64/sgemv_n_avx.c              |  32 ++++--
 kernel/x86_64/sgemv_n_microk_bulldozer.c | 137 ++++++++++++++++++++---
 2 files changed, 146 insertions(+), 23 deletions(-)

diff --git a/kernel/x86_64/sgemv_n_avx.c b/kernel/x86_64/sgemv_n_avx.c
index 8c263543c..dc8d015d8 100644
--- a/kernel/x86_64/sgemv_n_avx.c
+++ b/kernel/x86_64/sgemv_n_avx.c
@@ -70,12 +70,11 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
 	n1 = n / 512 ;
 	n2 = n % 512 ;
 
-	m1 = m / 32;
-	m2 = m % 32;
+	m1 = m / 64;
+	m2 = m % 64;
 
-	x_ptr = x;
-	a_ptr = a;
 	y_ptr = y;
+	x_ptr = x;
 
 	for (j=0; j<n1; j++)
 	{
@@ -85,12 +84,19 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
 		else
 			copy_x(512,x_ptr,xbuffer,inc_x);
 
-		x_ptr += 512 * inc_x;
-		a_ptr += j * 512;
+		a_ptr = a + j * 512 * lda;
 		y_ptr = y;
 
-
 		for(i = 0; i<m1; i++ )
+		{
+			sgemv_kernel_64(512,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(64,ybuffer,y_ptr,inc_y);
+			y_ptr += 64 * inc_y;
+			a_ptr += 64;			
+
+		}
+
+		if ( m2 & 32 )
 		{
 			sgemv_kernel_32(512,alpha,a_ptr,lda,xbuffer,ybuffer);
 			add_y(32,ybuffer,y_ptr,inc_y);
@@ -98,6 +104,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
 			a_ptr += 32;			
 
 		}
+
 		if ( m2 & 16 )
 		{
 			sgemv_kernel_16(512,alpha,a_ptr,lda,xbuffer,ybuffer);
@@ -131,6 +138,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
 			sgemv_kernel_1(512,alpha,a_ptr,lda,xbuffer,ybuffer);
 			add_y(1,ybuffer,y_ptr,inc_y);
 		}
+		x_ptr += 512 * inc_x;
 
 	}
 
@@ -142,9 +150,19 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
 		else
 			copy_x(n2,x_ptr,xbuffer,inc_x);
 
+		a_ptr = a + n1 * 512 * lda;
 		y_ptr = y;
 
 		for(i = 0; i<m1; i++ )
+		{
+			sgemv_kernel_64(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(64,ybuffer,y_ptr,inc_y);
+			y_ptr += 64 * inc_y;
+			a_ptr += 64;			
+
+		}
+
+		if ( m2 & 32 )
 		{
 			sgemv_kernel_32(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
 			add_y(32,ybuffer,y_ptr,inc_y);
diff --git a/kernel/x86_64/sgemv_n_microk_bulldozer.c b/kernel/x86_64/sgemv_n_microk_bulldozer.c
index 3dad43643..1cecd96c5 100644
--- a/kernel/x86_64/sgemv_n_microk_bulldozer.c
+++ b/kernel/x86_64/sgemv_n_microk_bulldozer.c
@@ -25,12 +25,11 @@ 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.
 *****************************************************************************/
 
-
-static void sgemv_kernel_32( long n, float alpha, float *a, long lda, float *x, float *y)
+static void sgemv_kernel_64( long n, float alpha, float *a, long lda, float *x, float *y)
 {
 
 
-	float *pre = a + lda*4*2;
+	float *pre = a + lda*3;
 
 	__asm __volatile
 	(
@@ -44,38 +43,56 @@ static void sgemv_kernel_32( long n, float alpha, float *a, long lda, float *x,
 	"prefetcht0	(%%r8)\n\t"			// Prefetch
 	"prefetcht0   64(%%r8)\n\t"			// Prefetch
 
+	"vxorps		%%ymm8 , %%ymm8 , %%ymm8 \n\t"	// set to zero
+	"vxorps		%%ymm9 , %%ymm9 , %%ymm9 \n\t"	// set to zero
+	"vxorps		%%ymm10, %%ymm10, %%ymm10\n\t"	// set to zero
+	"vxorps		%%ymm11, %%ymm11, %%ymm11\n\t"	// set to zero
 	"vxorps		%%ymm12, %%ymm12, %%ymm12\n\t"	// set to zero
 	"vxorps		%%ymm13, %%ymm13, %%ymm13\n\t"	// set to zero
 	"vxorps		%%ymm14, %%ymm14, %%ymm14\n\t"	// set to zero
 	"vxorps		%%ymm15, %%ymm15, %%ymm15\n\t"	// set to zero
-
+	".align 16				 \n\t"
 	".L01LOOP%=:				 \n\t"
 	"vbroadcastss	(%%rdi),   %%ymm0	 \n\t"	// load values of c
-        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
-
+	"nop					 \n\t"
 	"leaq     (%%r8 , %%rcx, 4), %%r8	 \n\t"	// add lda to pointer for prefetch
+
 	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"vfmaddps %%ymm8 ,   0*4(%%rsi), %%ymm0, %%ymm8 \n\t" // multiply a and c and add to temp
 	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+	"vfmaddps %%ymm9 ,   8*4(%%rsi), %%ymm0, %%ymm9 \n\t" // multiply a and c and add to temp
+	"prefetcht0   128(%%r8)\n\t"			// Prefetch
+	"vfmaddps %%ymm10,  16*4(%%rsi), %%ymm0, %%ymm10\n\t" // multiply a and c and add to temp
+	"vfmaddps %%ymm11,  24*4(%%rsi), %%ymm0, %%ymm11\n\t" // multiply a and c and add to temp
+	"prefetcht0   192(%%r8)\n\t"			// Prefetch
+	"vfmaddps %%ymm12,  32*4(%%rsi), %%ymm0, %%ymm12\n\t" // multiply a and c and add to temp
+	"vfmaddps %%ymm13,  40*4(%%rsi), %%ymm0, %%ymm13\n\t" // multiply a and c and add to temp
+	"vfmaddps %%ymm14,  48*4(%%rsi), %%ymm0, %%ymm14\n\t" // multiply a and c and add to temp
+	"vfmaddps %%ymm15,  56*4(%%rsi), %%ymm0, %%ymm15\n\t" // multiply a and c and add to temp
 
-	"vfmaddps %%ymm12,   0*4(%%rsi), %%ymm0, %%ymm12\n\t" // multiply a and c and add to temp
-	"vfmaddps %%ymm13,   8*4(%%rsi), %%ymm0, %%ymm13\n\t" // multiply a and c and add to temp
-	"vfmaddps %%ymm14,  16*4(%%rsi), %%ymm0, %%ymm14\n\t" // multiply a and c and add to temp
-	"vfmaddps %%ymm15,  24*4(%%rsi), %%ymm0, %%ymm15\n\t" // multiply a and c and add to temp
-
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
 	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
 
 	"dec		%%rax			 \n\t"  // n = n -1
 	"jnz		.L01LOOP%=		 \n\t"
 
+	"vmulps		%%ymm8 , %%ymm1,  %%ymm8 \n\t"  // scale by alpha
+	"vmulps		%%ymm9 , %%ymm1,  %%ymm9 \n\t"  // scale by alpha
+	"vmulps		%%ymm10, %%ymm1,  %%ymm10\n\t"  // scale by alpha
+	"vmulps		%%ymm11, %%ymm1,  %%ymm11\n\t"  // scale by alpha
 	"vmulps		%%ymm12, %%ymm1,  %%ymm12\n\t"  // scale by alpha
 	"vmulps		%%ymm13, %%ymm1,  %%ymm13\n\t"  // scale by alpha
 	"vmulps		%%ymm14, %%ymm1,  %%ymm14\n\t"  // scale by alpha
 	"vmulps		%%ymm15, %%ymm1,  %%ymm15\n\t"  // scale by alpha
 
-	"vmovups	%%ymm12,     (%%rdx)	 \n\t"  // store temp -> y
-	"vmovups	%%ymm13,  8*4(%%rdx)	 \n\t"  // store temp -> y
-	"vmovups	%%ymm14, 16*4(%%rdx)	 \n\t"  // store temp -> y
-	"vmovups	%%ymm15, 24*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm8 ,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm9 ,  8*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm10, 16*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm11, 24*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm12, 32*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm13, 40*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm14, 48*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm15, 56*4(%%rdx)	 \n\t"  // store temp -> y
 
 	:
         :
@@ -88,6 +105,94 @@ static void sgemv_kernel_32( long n, float alpha, float *a, long lda, float *x,
 	  "m" (pre)	// 6
 	: "rax", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11",
 	  "xmm0" , "xmm1", 
+	  "xmm8", "xmm9", "xmm10", "xmm11",
+	  "xmm12", "xmm13", "xmm14", "xmm15",
+	  "memory"
+	);
+
+} 
+
+
+
+static void sgemv_kernel_32( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	float *pre = a + lda*3;
+
+	__asm __volatile
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+	"movq		%6,	 %%r8\n\t"		// address for prefetch
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+
+	"vxorps		%%xmm8 , %%xmm8 , %%xmm8 \n\t"	// set to zero
+	"vxorps		%%xmm9 , %%xmm9 , %%xmm9 \n\t"	// set to zero
+	"vxorps		%%xmm10, %%xmm10, %%xmm10\n\t"	// set to zero
+	"vxorps		%%xmm11, %%xmm11, %%xmm11\n\t"	// set to zero
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+	"vxorps		%%xmm13, %%xmm13, %%xmm13\n\t"	// set to zero
+	"vxorps		%%xmm14, %%xmm14, %%xmm14\n\t"	// set to zero
+	"vxorps		%%xmm15, %%xmm15, %%xmm15\n\t"	// set to zero
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%xmm0	 \n\t"	// load values of c
+	"nop					 \n\t"
+	"leaq     (%%r8 , %%rcx, 4), %%r8	 \n\t"	// add lda to pointer for prefetch
+
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"vfmaddps %%xmm8 ,   0*4(%%rsi), %%xmm0, %%xmm8 \n\t" // multiply a and c and add to temp
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+	"vfmaddps %%xmm9 ,   4*4(%%rsi), %%xmm0, %%xmm9 \n\t" // multiply a and c and add to temp
+	"vfmaddps %%xmm10,   8*4(%%rsi), %%xmm0, %%xmm10\n\t" // multiply a and c and add to temp
+	"vfmaddps %%xmm11,  12*4(%%rsi), %%xmm0, %%xmm11\n\t" // multiply a and c and add to temp
+	"vfmaddps %%xmm12,  16*4(%%rsi), %%xmm0, %%xmm12\n\t" // multiply a and c and add to temp
+	"vfmaddps %%xmm13,  20*4(%%rsi), %%xmm0, %%xmm13\n\t" // multiply a and c and add to temp
+	"vfmaddps %%xmm14,  24*4(%%rsi), %%xmm0, %%xmm14\n\t" // multiply a and c and add to temp
+	"vfmaddps %%xmm15,  28*4(%%rsi), %%xmm0, %%xmm15\n\t" // multiply a and c and add to temp
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%xmm8 , %%xmm1,  %%xmm8 \n\t"  // scale by alpha
+	"vmulps		%%xmm9 , %%xmm1,  %%xmm9 \n\t"  // scale by alpha
+	"vmulps		%%xmm10, %%xmm1,  %%xmm10\n\t"  // scale by alpha
+	"vmulps		%%xmm11, %%xmm1,  %%xmm11\n\t"  // scale by alpha
+	"vmulps		%%xmm12, %%xmm1,  %%xmm12\n\t"  // scale by alpha
+	"vmulps		%%xmm13, %%xmm1,  %%xmm13\n\t"  // scale by alpha
+	"vmulps		%%xmm14, %%xmm1,  %%xmm14\n\t"  // scale by alpha
+	"vmulps		%%xmm15, %%xmm1,  %%xmm15\n\t"  // scale by alpha
+
+	"vmovups	%%xmm8 ,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%xmm9 ,  4*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%xmm10,  8*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%xmm11, 12*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%xmm12, 16*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%xmm13, 20*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%xmm14, 24*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%xmm15, 28*4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y),      // 5
+	  "m" (pre)	// 6
+	: "rax", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11",
+	  "xmm0" , "xmm1", 
+	  "xmm8", "xmm9", "xmm10", "xmm11",
 	  "xmm12", "xmm13", "xmm14", "xmm15",
 	  "memory"
 	);
@@ -97,7 +202,7 @@ static void sgemv_kernel_32( long n, float alpha, float *a, long lda, float *x,
 static void sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
 {
 
-	float *pre = a + lda*4*3;
+	float *pre = a + lda*1;
 
 	__asm __volatile
 	(

From b3938fe371d1806233b06eff23eda4456d2f763a Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Sat, 19 Jul 2014 07:15:34 +0200
Subject: [PATCH 08/15] don't use this sgemv_n on Windows

---
 kernel/x86_64/KERNEL.BULLDOZER           |  5 ++
 kernel/x86_64/KERNEL.PILEDRIVER          |  5 ++
 kernel/x86_64/sgemv_n_avx.c              |  6 +-
 kernel/x86_64/sgemv_n_microk_bulldozer.c | 78 ++++++++++++------------
 4 files changed, 53 insertions(+), 41 deletions(-)

diff --git a/kernel/x86_64/KERNEL.BULLDOZER b/kernel/x86_64/KERNEL.BULLDOZER
index 55932e69f..fac8016a6 100644
--- a/kernel/x86_64/KERNEL.BULLDOZER
+++ b/kernel/x86_64/KERNEL.BULLDOZER
@@ -1,4 +1,9 @@
+ifdef OS_WINDOWS
+SGEMVNKERNEL = ../arm/gemv_n.c
+else
 SGEMVNKERNEL = sgemv_n_avx.c
+endif
+
 
 ZGEMVNKERNEL = zgemv_n_dup.S
 ZGEMVTKERNEL = zgemv_t.S
diff --git a/kernel/x86_64/KERNEL.PILEDRIVER b/kernel/x86_64/KERNEL.PILEDRIVER
index 145d9fb2f..555c8053d 100644
--- a/kernel/x86_64/KERNEL.PILEDRIVER
+++ b/kernel/x86_64/KERNEL.PILEDRIVER
@@ -1,4 +1,9 @@
+ifdef OS_WINDOWS
+SGEMVNKERNEL = ../arm/gemv_n.c
+else
 SGEMVNKERNEL = sgemv_n_avx.c
+endif
+
 
 ZGEMVNKERNEL = zgemv_n_dup.S
 ZGEMVTKERNEL = zgemv_t.S
diff --git a/kernel/x86_64/sgemv_n_avx.c b/kernel/x86_64/sgemv_n_avx.c
index dc8d015d8..91e3ee424 100644
--- a/kernel/x86_64/sgemv_n_avx.c
+++ b/kernel/x86_64/sgemv_n_avx.c
@@ -61,8 +61,10 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
 	FLOAT *a_ptr;
 	FLOAT *x_ptr;
 	FLOAT *y_ptr;
-	BLASLONG n1,n2;
-	BLASLONG m1,m2;
+	BLASLONG n1;
+	BLASLONG m1;
+	BLASLONG register m2;
+	BLASLONG register n2;
 	FLOAT *xbuffer,*ybuffer;
 	xbuffer = buffer;
 	ybuffer = xbuffer + 2048 + 256;
diff --git a/kernel/x86_64/sgemv_n_microk_bulldozer.c b/kernel/x86_64/sgemv_n_microk_bulldozer.c
index 1cecd96c5..1b07f0291 100644
--- a/kernel/x86_64/sgemv_n_microk_bulldozer.c
+++ b/kernel/x86_64/sgemv_n_microk_bulldozer.c
@@ -25,13 +25,13 @@ 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.
 *****************************************************************************/
 
-static void sgemv_kernel_64( long n, float alpha, float *a, long lda, float *x, float *y)
+static void  sgemv_kernel_64( long n, float alpha, float *a, long lda, float *x, float *y)
 {
 
 
 	float *pre = a + lda*3;
 
-	__asm __volatile
+	__asm__  __volatile__
 	(
 	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
 	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
@@ -103,10 +103,10 @@ static void sgemv_kernel_64( long n, float alpha, float *a, long lda, float *x,
           "m" (x),      // 4
           "m" (y),      // 5
 	  "m" (pre)	// 6
-	: "rax", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11",
-	  "xmm0" , "xmm1", 
-	  "xmm8", "xmm9", "xmm10", "xmm11",
-	  "xmm12", "xmm13", "xmm14", "xmm15",
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
 	  "memory"
 	);
 
@@ -114,13 +114,13 @@ static void sgemv_kernel_64( long n, float alpha, float *a, long lda, float *x,
 
 
 
-static void sgemv_kernel_32( long n, float alpha, float *a, long lda, float *x, float *y)
+static void  sgemv_kernel_32( long n, float alpha, float *a, long lda, float *x, float *y)
 {
 
 
 	float *pre = a + lda*3;
 
-	__asm __volatile
+	__asm__ __volatile__
 	(
 	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
 	"vbroadcastss   %1, 	 %%xmm1\n\t"		// alpha -> xmm1
@@ -190,21 +190,16 @@ static void sgemv_kernel_32( long n, float alpha, float *a, long lda, float *x,
           "m" (x),      // 4
           "m" (y),      // 5
 	  "m" (pre)	// 6
-	: "rax", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11",
-	  "xmm0" , "xmm1", 
-	  "xmm8", "xmm9", "xmm10", "xmm11",
-	  "xmm12", "xmm13", "xmm14", "xmm15",
-	  "memory"
 	);
 
 } 
 
-static void sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
+static void  sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
 {
 
-	float *pre = a + lda*1;
+	float *pre = a + lda*3;
 
-	__asm __volatile
+	__asm__ __volatile__
 	(
 	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
 	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
@@ -248,20 +243,21 @@ static void sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x,
           "m" (x),      // 4
           "m" (y),      // 5
 	  "m" (pre)	// 6
-	: "rax", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11",
-	  "xmm0" , "xmm1", 
-	  "xmm12", "xmm13", "xmm14", "xmm15",
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
 	  "memory"
 	);
 
 } 
 
 
-static void sgemv_kernel_8( long n, float alpha, float *a, long lda, float *x, float *y)
+static void  sgemv_kernel_8( long n, float alpha, float *a, long lda, float *x, float *y)
 {
 
 
-	__asm __volatile
+	__asm__ __volatile__
 	(
 	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
 	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
@@ -295,20 +291,21 @@ static void sgemv_kernel_8( long n, float alpha, float *a, long lda, float *x, f
           "m" (lda),    // 3
           "m" (x),      // 4
           "m" (y)       // 5
-	: "rax", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11",
-	  "xmm0" , "xmm1", 
-	  "xmm12", "xmm13", "xmm14", "xmm15",
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
 	  "memory"
 	);
 
 } 
 
 
-static void sgemv_kernel_4( long n, float alpha, float *a, long lda, float *x, float *y)
+static void  sgemv_kernel_4( long n, float alpha, float *a, long lda, float *x, float *y)
 {
 
 
-	__asm __volatile
+	__asm__ __volatile__
 	(
 	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
 	"vbroadcastss   %1, 	 %%xmm1\n\t"		// alpha -> xmm1
@@ -342,19 +339,20 @@ static void sgemv_kernel_4( long n, float alpha, float *a, long lda, float *x, f
           "m" (lda),    // 3
           "m" (x),      // 4
           "m" (y)       // 5
-	: "rax", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11",
-	  "xmm0" , "xmm1", 
-	  "xmm12", "xmm13", "xmm14", "xmm15",
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
 	  "memory"
 	);
 
 } 
 
-static void sgemv_kernel_2( long n, float alpha, float *a, long lda, float *x, float *y)
+static void  sgemv_kernel_2( long n, float alpha, float *a, long lda, float *x, float *y)
 {
 
 
-	__asm __volatile
+	__asm__ __volatile__
 	(
 	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
 	"vmovss         %1, 	 %%xmm1\n\t"		// alpha -> xmm1
@@ -392,9 +390,10 @@ static void sgemv_kernel_2( long n, float alpha, float *a, long lda, float *x, f
           "m" (lda),    // 3
           "m" (x),      // 4
           "m" (y)       // 5
-	: "rax", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11",
-	  "xmm0" , "xmm1", 
-	  "xmm12", "xmm13", "xmm14", "xmm15",
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
 	  "memory"
 	);
 
@@ -402,11 +401,11 @@ static void sgemv_kernel_2( long n, float alpha, float *a, long lda, float *x, f
 
 
 
-static void sgemv_kernel_1( long n, float alpha, float *a, long lda, float *x, float *y)
+static void  sgemv_kernel_1( long n, float alpha, float *a, long lda, float *x, float *y)
 {
 
 
-	__asm __volatile
+	__asm__ __volatile__
 	(
 	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
 	"vmovss         %1, 	 %%xmm1\n\t"		// alpha -> xmm1
@@ -440,9 +439,10 @@ static void sgemv_kernel_1( long n, float alpha, float *a, long lda, float *x, f
           "m" (lda),    // 3
           "m" (x),      // 4
           "m" (y)       // 5
-	: "rax", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11",
-	  "xmm0" , "xmm1", 
-	  "xmm12", "xmm13", "xmm14", "xmm15",
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
 	  "memory"
 	);
 

From 2cce125c795632e4dd6f209e5b9703ed39a7ef10 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Sat, 19 Jul 2014 15:48:07 +0200
Subject: [PATCH 09/15] added optimized sgemv_t for bulldozer and piledriver

---
 kernel/x86_64/KERNEL.BULLDOZER           |   2 +
 kernel/x86_64/KERNEL.PILEDRIVER          |   2 +
 kernel/x86_64/sgemv_t_avx.c              | 228 +++++++++++++++++++++++
 kernel/x86_64/sgemv_t_microk_bulldozer.c |  99 ++++++++++
 4 files changed, 331 insertions(+)
 create mode 100644 kernel/x86_64/sgemv_t_avx.c
 create mode 100644 kernel/x86_64/sgemv_t_microk_bulldozer.c

diff --git a/kernel/x86_64/KERNEL.BULLDOZER b/kernel/x86_64/KERNEL.BULLDOZER
index fac8016a6..73a9ad2ec 100644
--- a/kernel/x86_64/KERNEL.BULLDOZER
+++ b/kernel/x86_64/KERNEL.BULLDOZER
@@ -1,7 +1,9 @@
 ifdef OS_WINDOWS
 SGEMVNKERNEL = ../arm/gemv_n.c
+SGEMVTKERNEL = ../arm/gemv_t.c
 else
 SGEMVNKERNEL = sgemv_n_avx.c
+SGEMVTKERNEL = sgemv_t_avx.c
 endif
 
 
diff --git a/kernel/x86_64/KERNEL.PILEDRIVER b/kernel/x86_64/KERNEL.PILEDRIVER
index 555c8053d..453e7b762 100644
--- a/kernel/x86_64/KERNEL.PILEDRIVER
+++ b/kernel/x86_64/KERNEL.PILEDRIVER
@@ -1,7 +1,9 @@
 ifdef OS_WINDOWS
 SGEMVNKERNEL = ../arm/gemv_n.c
+SGEMVTKERNEL = ../arm/gemv_t.c
 else
 SGEMVNKERNEL = sgemv_n_avx.c
+SGEMVTKERNEL = sgemv_t_avx.c
 endif
 
 
diff --git a/kernel/x86_64/sgemv_t_avx.c b/kernel/x86_64/sgemv_t_avx.c
new file mode 100644
index 000000000..c9cdb60cd
--- /dev/null
+++ b/kernel/x86_64/sgemv_t_avx.c
@@ -0,0 +1,228 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include "common.h"
+
+#if defined(BULLDOZER) || defined(PILEDRIVER)
+#include "sgemv_t_microk_bulldozer.c"
+#endif
+
+static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
+{
+	BLASLONG i;
+	for ( i=0; i<n; i++ )
+	{
+		*dest = *src;
+		dest++;
+		src += inc_src;
+	}
+}
+
+static void  sgemv_kernel_1( BLASLONG n, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, FLOAT *y)
+{
+
+	FLOAT register temp0 = 0.0;
+	BLASLONG i;
+	for ( i=0; i<n ; i++)
+	{
+		temp0 += a[i] * x[i];
+	}
+	temp0 *= alpha ;
+	*y += temp0;
+}
+
+
+
+
+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 *a_ptrl;
+	BLASLONG m1;
+	BLASLONG register m2;
+	FLOAT *xbuffer;
+	xbuffer = buffer;
+	BLASLONG register Mblock;
+
+	m1 = m / 1024 ;
+	m2 = m % 1024 ;
+
+	x_ptr = x;
+	a_ptr = a;
+
+	for (j=0; j<m1; j++)
+	{
+
+		if ( inc_x == 1 )
+			xbuffer = x_ptr;
+		else
+			copy_x(1024,x_ptr,xbuffer,inc_x);
+
+		y_ptr = y;
+		a_ptrl = a_ptr;
+
+		for(i = 0; i<n; i++ )
+		{
+			sgemv_kernel_16(1024,alpha,a_ptrl,lda,xbuffer,y_ptr);
+			y_ptr += inc_y;
+			a_ptrl += lda;
+		}
+		a_ptr += 1024;	
+		x_ptr += 1024 * inc_x;
+	}
+
+	if ( m2 == 0 ) return(0);
+
+	Mblock = 512;
+	while ( Mblock >= 16 )
+	{
+	  if ( m2 & Mblock)
+	  {
+
+		if ( inc_x == 1 )
+			xbuffer = x_ptr;
+		else
+			copy_x(Mblock,x_ptr,xbuffer,inc_x);
+
+		y_ptr = y;
+		a_ptrl = a_ptr;
+
+		for(i = 0; i<n; i++ )
+		{
+			sgemv_kernel_16(Mblock,alpha,a_ptrl,lda,xbuffer,y_ptr);
+			y_ptr += inc_y;
+			a_ptrl += lda;
+		}
+		a_ptr += Mblock;	
+		x_ptr += Mblock * inc_x;
+
+
+ 	  }
+	  Mblock /= 2;
+
+	}
+
+        if ( m2 & Mblock)
+	{
+
+		if ( inc_x == 1 )
+			xbuffer = x_ptr;
+		else
+			copy_x(Mblock,x_ptr,xbuffer,inc_x);
+
+		y_ptr = y;
+		a_ptrl = a_ptr;
+
+		for(i = 0; i<n; i++ )
+		{
+			sgemv_kernel_1(Mblock,alpha,a_ptrl,lda,xbuffer,y_ptr);
+			y_ptr += inc_y;
+			a_ptrl += lda;
+		}
+		a_ptr += Mblock;	
+		x_ptr += Mblock * inc_x;
+
+
+ 	}
+	Mblock /= 2;
+
+
+        if ( m2 & Mblock)
+	{
+
+		if ( inc_x == 1 )
+			xbuffer = x_ptr;
+		else
+			copy_x(Mblock,x_ptr,xbuffer,inc_x);
+
+		y_ptr = y;
+		a_ptrl = a_ptr;
+
+		for(i = 0; i<n; i++ )
+		{
+			sgemv_kernel_1(Mblock,alpha,a_ptrl,lda,xbuffer,y_ptr);
+			y_ptr += inc_y;
+			a_ptrl += lda;
+		}
+		a_ptr += Mblock;	
+		x_ptr += Mblock * inc_x;
+
+
+ 	}
+	Mblock /= 2;
+
+        if ( m2 & Mblock)
+	{
+
+		if ( inc_x == 1 )
+			xbuffer = x_ptr;
+		else
+			copy_x(Mblock,x_ptr,xbuffer,inc_x);
+
+		y_ptr = y;
+		a_ptrl = a_ptr;
+
+		for(i = 0; i<n; i++ )
+		{
+			sgemv_kernel_1(Mblock,alpha,a_ptrl,lda,xbuffer,y_ptr);
+			y_ptr += inc_y;
+			a_ptrl += lda;
+		}
+		a_ptr += Mblock;	
+		x_ptr += Mblock * inc_x;
+
+
+ 	}
+	Mblock /= 2;
+
+        if ( m2 & Mblock)
+	{
+
+		xbuffer = x_ptr;
+
+		y_ptr = y;
+		a_ptrl = a_ptr;
+
+		for(i = 0; i<n; i++ )
+		{
+			sgemv_kernel_1(Mblock,alpha,a_ptrl,lda,xbuffer,y_ptr);
+			y_ptr += inc_y;
+			a_ptrl += lda;
+		}
+
+
+ 	}
+
+	return(0);
+}
+
+
diff --git a/kernel/x86_64/sgemv_t_microk_bulldozer.c b/kernel/x86_64/sgemv_t_microk_bulldozer.c
new file mode 100644
index 000000000..56b12a1e8
--- /dev/null
+++ b/kernel/x86_64/sgemv_t_microk_bulldozer.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.
+*****************************************************************************/
+
+static void  sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+	//n = n / 16;
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vmovss         %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"leaq	(, %%rcx,4), %%rcx       \n\t"		// scale lda by size of float
+	"leaq	(%%rsi,%%rcx,1), %%r8    \n\t"		// pointer to next line
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+	"vxorps		%%xmm13, %%xmm13, %%xmm13\n\t"	// set to zero
+	"vxorps		%%xmm14, %%xmm14, %%xmm14\n\t"	// set to zero
+	"vxorps		%%xmm15, %%xmm15, %%xmm15\n\t"	// set to zero
+
+	"sarq		$4, %%rax		 \n\t"	// n = n / 16
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	// "prefetcht0	512(%%rsi)		 \n\t"
+	"prefetcht0	(%%r8)		 	 \n\t" //prefetch next line of a
+	"vmovups	(%%rsi), %%xmm4		 \n\t"
+	"vmovups     4*4(%%rsi), %%xmm5		 \n\t"
+	"vmovups     8*4(%%rsi), %%xmm6		 \n\t"
+	"vmovups    12*4(%%rsi), %%xmm7		 \n\t"
+
+	"vfmaddps %%xmm12,   0*4(%%rdi), %%xmm4, %%xmm12\n\t" // multiply a and c and add to temp
+	"vfmaddps %%xmm13,   4*4(%%rdi), %%xmm5, %%xmm13\n\t" // multiply a and c and add to temp
+	"vfmaddps %%xmm14,   8*4(%%rdi), %%xmm6, %%xmm14\n\t" // multiply a and c and add to temp
+	"vfmaddps %%xmm15,  12*4(%%rdi), %%xmm7, %%xmm15\n\t" // multiply a and c and add to temp
+
+        "addq		$16*4    ,   %%r8	 \n\t"  // increment prefetch pointer 
+        "addq		$16*4    ,   %%rsi	 \n\t"  // increment pointer of a 
+        "addq		$16*4    ,   %%rdi	 \n\t"  // increment pointer of c 
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vaddps		%%xmm12, %%xmm14, %%xmm12\n\t"	
+	"vaddps		%%xmm13, %%xmm15, %%xmm13\n\t"	
+	"vaddps		%%xmm12, %%xmm13, %%xmm12\n\t"	
+	"vhaddps	%%xmm12, %%xmm12, %%xmm12\n\t"	
+	"vhaddps	%%xmm12, %%xmm12, %%xmm12\n\t"	
+
+	"vfmaddss	(%%rdx), %%xmm12, %%xmm1, %%xmm12\n\t"
+	"vmovss		%%xmm12, (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
+	  "%xmm0", "%xmm1", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
+

From c06f9986d449bdaa109f742a3e5f7114b4ff50ad Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Sun, 20 Jul 2014 10:21:08 +0200
Subject: [PATCH 10/15] added sgemv_t microkernel for sandybridge

---
 kernel/x86_64/sgemv_t_avx.c          |   2 +
 kernel/x86_64/sgemv_t_microk_sandy.c | 105 +++++++++++++++++++++++++++
 2 files changed, 107 insertions(+)
 create mode 100644 kernel/x86_64/sgemv_t_microk_sandy.c

diff --git a/kernel/x86_64/sgemv_t_avx.c b/kernel/x86_64/sgemv_t_avx.c
index c9cdb60cd..7a9efa35e 100644
--- a/kernel/x86_64/sgemv_t_avx.c
+++ b/kernel/x86_64/sgemv_t_avx.c
@@ -30,6 +30,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #if defined(BULLDOZER) || defined(PILEDRIVER)
 #include "sgemv_t_microk_bulldozer.c"
+#else
+#include "sgemv_t_microk_sandy.c"
 #endif
 
 static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
diff --git a/kernel/x86_64/sgemv_t_microk_sandy.c b/kernel/x86_64/sgemv_t_microk_sandy.c
new file mode 100644
index 000000000..1745db3a7
--- /dev/null
+++ b/kernel/x86_64/sgemv_t_microk_sandy.c
@@ -0,0 +1,105 @@
+/***************************************************************************
+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.
+*****************************************************************************/
+
+static void  sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+	//n = n / 16;
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vmovss         %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"leaq	(, %%rcx,4), %%rcx       \n\t"		// scale lda by size of float
+	"leaq	(%%rsi,%%rcx,1), %%r8    \n\t"		// pointer to next line
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+	"vxorps		%%xmm13, %%xmm13, %%xmm13\n\t"	// set to zero
+	"vxorps		%%xmm14, %%xmm14, %%xmm14\n\t"	// set to zero
+	"vxorps		%%xmm15, %%xmm15, %%xmm15\n\t"	// set to zero
+
+	"sarq		$4, %%rax		 \n\t"	// n = n / 16
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	// "prefetcht0	512(%%rsi)		 \n\t"
+	"prefetcht0	(%%r8)		 	 \n\t" //prefetch next line of a
+	"vmovups	(%%rsi), %%xmm4		 \n\t"
+	"vmovups     4*4(%%rsi), %%xmm5		 \n\t"
+	"vmovups     8*4(%%rsi), %%xmm6		 \n\t"
+	"vmovups    12*4(%%rsi), %%xmm7		 \n\t"
+
+	"vmulps      0*4(%%rdi), %%xmm4, %%xmm8 \n\t" // multiply a and c and add to temp
+	"vmulps      4*4(%%rdi), %%xmm5, %%xmm9 \n\t" // multiply a and c and add to temp
+	"vmulps      8*4(%%rdi), %%xmm6, %%xmm10\n\t" // multiply a and c and add to temp
+	"vmulps     12*4(%%rdi), %%xmm7, %%xmm11\n\t" // multiply a and c and add to temp
+
+	"vaddps		%%xmm12, %%xmm8 , %%xmm12\n\t"	
+	"vaddps		%%xmm13, %%xmm9 , %%xmm13\n\t"	
+	"vaddps		%%xmm14, %%xmm10, %%xmm14\n\t"	
+	"vaddps		%%xmm15, %%xmm11, %%xmm15\n\t"	
+
+        "addq		$16*4    ,   %%r8	 \n\t"  // increment prefetch pointer 
+        "addq		$16*4    ,   %%rsi	 \n\t"  // increment pointer of a 
+        "addq		$16*4    ,   %%rdi	 \n\t"  // increment pointer of c 
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vaddps		%%xmm12, %%xmm14, %%xmm12\n\t"	
+	"vaddps		%%xmm13, %%xmm15, %%xmm13\n\t"	
+	"vaddps		%%xmm12, %%xmm13, %%xmm12\n\t"	
+	"vhaddps	%%xmm12, %%xmm12, %%xmm12\n\t"	
+	"vhaddps	%%xmm12, %%xmm12, %%xmm12\n\t"	
+
+	"vfmaddss	(%%rdx), %%xmm12, %%xmm1, %%xmm12\n\t"
+	"vmovss		%%xmm12, (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
+	  "%xmm0", "%xmm1", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
+

From 02eb72ac426226566b6b9d3cffd4beaacde88672 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Sun, 20 Jul 2014 10:48:41 +0200
Subject: [PATCH 11/15] bugfix in sgemv_t_microk_sandy.c

---
 kernel/x86_64/KERNEL.SANDYBRIDGE     | 8 ++++++++
 kernel/x86_64/sgemv_t_microk_sandy.c | 3 ++-
 2 files changed, 10 insertions(+), 1 deletion(-)

diff --git a/kernel/x86_64/KERNEL.SANDYBRIDGE b/kernel/x86_64/KERNEL.SANDYBRIDGE
index 7228357ce..7d6b81d54 100644
--- a/kernel/x86_64/KERNEL.SANDYBRIDGE
+++ b/kernel/x86_64/KERNEL.SANDYBRIDGE
@@ -1,3 +1,11 @@
+ifdef OS_WINDOWS
+#SGEMVNKERNEL = ../arm/gemv_n.c
+SGEMVTKERNEL = ../arm/gemv_t.c
+else
+#SGEMVNKERNEL = sgemv_n_avx.c
+SGEMVTKERNEL = sgemv_t_avx.c
+endif
+
 
 SGEMMKERNEL    =  sgemm_kernel_16x4_sandy.S
 SGEMMINCOPY    =  ../generic/gemm_ncopy_16.c
diff --git a/kernel/x86_64/sgemv_t_microk_sandy.c b/kernel/x86_64/sgemv_t_microk_sandy.c
index 1745db3a7..4ecd6d3d0 100644
--- a/kernel/x86_64/sgemv_t_microk_sandy.c
+++ b/kernel/x86_64/sgemv_t_microk_sandy.c
@@ -80,7 +80,8 @@ static void  sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x,
 	"vhaddps	%%xmm12, %%xmm12, %%xmm12\n\t"	
 	"vhaddps	%%xmm12, %%xmm12, %%xmm12\n\t"	
 
-	"vfmaddss	(%%rdx), %%xmm12, %%xmm1, %%xmm12\n\t"
+	"vmulss		%%xmm12, %%xmm1, %%xmm12 \n\t"
+	"vaddss	       (%%rdx), %%xmm12, %%xmm12\n\t"
 	"vmovss		%%xmm12, (%%rdx)	 \n\t"  // store temp -> y
 
 	:

From d9d4077c9317b0c283dbce0547ea299dc5f1df82 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Sun, 20 Jul 2014 11:30:32 +0200
Subject: [PATCH 12/15] added sgemv_t microkernel for haswell

---
 kernel/x86_64/KERNEL.HASWELL           |   8 ++
 kernel/x86_64/sgemv_t_avx.c            |   2 +
 kernel/x86_64/sgemv_t_microk_haswell.c | 100 +++++++++++++++++++++++++
 3 files changed, 110 insertions(+)
 create mode 100644 kernel/x86_64/sgemv_t_microk_haswell.c

diff --git a/kernel/x86_64/KERNEL.HASWELL b/kernel/x86_64/KERNEL.HASWELL
index ae316cff0..288e39537 100644
--- a/kernel/x86_64/KERNEL.HASWELL
+++ b/kernel/x86_64/KERNEL.HASWELL
@@ -1,3 +1,11 @@
+ifdef OS_WINDOWS
+#SGEMVNKERNEL = ../arm/gemv_n.c
+SGEMVTKERNEL = ../arm/gemv_t.c
+else
+#SGEMVNKERNEL = sgemv_n_avx.c
+SGEMVTKERNEL = sgemv_t_avx.c
+endif
+
 
 SGEMMKERNEL    =  sgemm_kernel_16x4_haswell.S
 SGEMMINCOPY    =  ../generic/gemm_ncopy_16.c
diff --git a/kernel/x86_64/sgemv_t_avx.c b/kernel/x86_64/sgemv_t_avx.c
index 7a9efa35e..55fb3d623 100644
--- a/kernel/x86_64/sgemv_t_avx.c
+++ b/kernel/x86_64/sgemv_t_avx.c
@@ -30,6 +30,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #if defined(BULLDOZER) || defined(PILEDRIVER)
 #include "sgemv_t_microk_bulldozer.c"
+#elif defined(HASWELL)
+#include "sgemv_t_microk_haswell.c"
 #else
 #include "sgemv_t_microk_sandy.c"
 #endif
diff --git a/kernel/x86_64/sgemv_t_microk_haswell.c b/kernel/x86_64/sgemv_t_microk_haswell.c
new file mode 100644
index 000000000..ecb9845bb
--- /dev/null
+++ b/kernel/x86_64/sgemv_t_microk_haswell.c
@@ -0,0 +1,100 @@
+/***************************************************************************
+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.
+*****************************************************************************/
+
+static void  sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+	//n = n / 16;
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vmovss         %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"leaq	(, %%rcx,4), %%rcx       \n\t"		// scale lda by size of float
+	"leaq	(%%rsi,%%rcx,1), %%r8    \n\t"		// pointer to next line
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+	"vxorps		%%xmm13, %%xmm13, %%xmm13\n\t"	// set to zero
+	"vxorps		%%xmm14, %%xmm14, %%xmm14\n\t"	// set to zero
+	"vxorps		%%xmm15, %%xmm15, %%xmm15\n\t"	// set to zero
+
+	"sarq		$4, %%rax		 \n\t"	// n = n / 16
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	// "prefetcht0	512(%%rsi)		 \n\t"
+	"prefetcht0	(%%r8)		 	 \n\t" //prefetch next line of a
+	"vmovups	(%%rsi), %%xmm4		 \n\t"
+	"vmovups     4*4(%%rsi), %%xmm5		 \n\t"
+	"vmovups     8*4(%%rsi), %%xmm6		 \n\t"
+	"vmovups    12*4(%%rsi), %%xmm7		 \n\t"
+
+	"vfmadd231ps   0*4(%%rdi), %%xmm4, %%xmm12\n\t" // multiply a and c and add to temp
+	"vfmadd231ps   4*4(%%rdi), %%xmm5, %%xmm13\n\t" // multiply a and c and add to temp
+	"vfmadd231ps   8*4(%%rdi), %%xmm6, %%xmm14\n\t" // multiply a and c and add to temp
+	"vfmadd231ps  12*4(%%rdi), %%xmm7, %%xmm15\n\t" // multiply a and c and add to temp
+
+        "addq		$16*4    ,   %%r8	 \n\t"  // increment prefetch pointer 
+        "addq		$16*4    ,   %%rsi	 \n\t"  // increment pointer of a 
+        "addq		$16*4    ,   %%rdi	 \n\t"  // increment pointer of c 
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vaddps		%%xmm12, %%xmm14, %%xmm12\n\t"	
+	"vaddps		%%xmm13, %%xmm15, %%xmm13\n\t"	
+	"vaddps		%%xmm12, %%xmm13, %%xmm12\n\t"	
+	"vhaddps	%%xmm12, %%xmm12, %%xmm12\n\t"	
+	"vhaddps	%%xmm12, %%xmm12, %%xmm12\n\t"	
+
+	"vmulss	        %%xmm12, %%xmm1, %%xmm12\n\t"
+	"vaddss		(%%rdx), %%xmm12,%%xmm12\n\t"
+	"vmovss		%%xmm12, (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
+	  "%xmm0", "%xmm1", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
+

From c0fe95fb725aba06ce08a114a0b79f91d91ec64a Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Sun, 20 Jul 2014 13:17:47 +0200
Subject: [PATCH 13/15] added sgemv_n microkernel for sandybridge

---
 kernel/x86_64/KERNEL.SANDYBRIDGE     |   4 +-
 kernel/x86_64/sgemv_n_avx.c          |   2 +
 kernel/x86_64/sgemv_n_microk_sandy.c | 474 +++++++++++++++++++++++++++
 3 files changed, 478 insertions(+), 2 deletions(-)
 create mode 100644 kernel/x86_64/sgemv_n_microk_sandy.c

diff --git a/kernel/x86_64/KERNEL.SANDYBRIDGE b/kernel/x86_64/KERNEL.SANDYBRIDGE
index 7d6b81d54..9d7a49562 100644
--- a/kernel/x86_64/KERNEL.SANDYBRIDGE
+++ b/kernel/x86_64/KERNEL.SANDYBRIDGE
@@ -1,8 +1,8 @@
 ifdef OS_WINDOWS
-#SGEMVNKERNEL = ../arm/gemv_n.c
+SGEMVNKERNEL = ../arm/gemv_n.c
 SGEMVTKERNEL = ../arm/gemv_t.c
 else
-#SGEMVNKERNEL = sgemv_n_avx.c
+SGEMVNKERNEL = sgemv_n_avx.c
 SGEMVTKERNEL = sgemv_t_avx.c
 endif
 
diff --git a/kernel/x86_64/sgemv_n_avx.c b/kernel/x86_64/sgemv_n_avx.c
index 91e3ee424..96a03ec57 100644
--- a/kernel/x86_64/sgemv_n_avx.c
+++ b/kernel/x86_64/sgemv_n_avx.c
@@ -30,6 +30,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #if defined(BULLDOZER) || defined(PILEDRIVER)
 #include "sgemv_n_microk_bulldozer.c"
+#else
+#include "sgemv_n_microk_sandy.c"
 #endif
 
 static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
diff --git a/kernel/x86_64/sgemv_n_microk_sandy.c b/kernel/x86_64/sgemv_n_microk_sandy.c
new file mode 100644
index 000000000..7d9360f94
--- /dev/null
+++ b/kernel/x86_64/sgemv_n_microk_sandy.c
@@ -0,0 +1,474 @@
+/***************************************************************************
+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.
+*****************************************************************************/
+
+static void  sgemv_kernel_64( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	float *pre = a + lda*3;
+
+	__asm__  __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+	"movq		%6,	 %%r8\n\t"		// address for prefetch
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+
+	"vxorps		%%ymm8 , %%ymm8 , %%ymm8 \n\t"	// set to zero
+	"vxorps		%%ymm9 , %%ymm9 , %%ymm9 \n\t"	// set to zero
+	"vxorps		%%ymm10, %%ymm10, %%ymm10\n\t"	// set to zero
+	"vxorps		%%ymm11, %%ymm11, %%ymm11\n\t"	// set to zero
+	"vxorps		%%ymm12, %%ymm12, %%ymm12\n\t"	// set to zero
+	"vxorps		%%ymm13, %%ymm13, %%ymm13\n\t"	// set to zero
+	"vxorps		%%ymm14, %%ymm14, %%ymm14\n\t"	// set to zero
+	"vxorps		%%ymm15, %%ymm15, %%ymm15\n\t"	// set to zero
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%ymm0	 \n\t"	// load values of c
+	"nop					 \n\t"
+	"leaq     (%%r8 , %%rcx, 4), %%r8	 \n\t"	// add lda to pointer for prefetch
+
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+	"prefetcht0  128(%%r8)\n\t"			// Prefetch
+	"prefetcht0  192(%%r8)\n\t"			// Prefetch
+
+	"vmulps   0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
+	"vmulps   8*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
+	"vmulps  16*4(%%rsi), %%ymm0, %%ymm6 \n\t" // multiply a and c and add to temp
+	"vmulps  24*4(%%rsi), %%ymm0, %%ymm7 \n\t" // multiply a and c and add to temp
+
+	"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
+	"vaddps %%ymm9 , %%ymm5, %%ymm9 \n\t" // multiply a and c and add to temp
+	"vaddps %%ymm10, %%ymm6, %%ymm10\n\t" // multiply a and c and add to temp
+	"vaddps %%ymm11, %%ymm7, %%ymm11\n\t" // multiply a and c and add to temp
+
+	"vmulps  32*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
+	"vmulps  40*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
+	"vmulps  48*4(%%rsi), %%ymm0, %%ymm6 \n\t" // multiply a and c and add to temp
+	"vmulps  56*4(%%rsi), %%ymm0, %%ymm7 \n\t" // multiply a and c and add to temp
+
+	"vaddps %%ymm12, %%ymm4, %%ymm12\n\t" // multiply a and c and add to temp
+	"vaddps %%ymm13, %%ymm5, %%ymm13\n\t" // multiply a and c and add to temp
+	"vaddps %%ymm14, %%ymm6, %%ymm14\n\t" // multiply a and c and add to temp
+	"vaddps %%ymm15, %%ymm7, %%ymm15\n\t" // multiply a and c and add to temp
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%ymm8 , %%ymm1,  %%ymm8 \n\t"  // scale by alpha
+	"vmulps		%%ymm9 , %%ymm1,  %%ymm9 \n\t"  // scale by alpha
+	"vmulps		%%ymm10, %%ymm1,  %%ymm10\n\t"  // scale by alpha
+	"vmulps		%%ymm11, %%ymm1,  %%ymm11\n\t"  // scale by alpha
+	"vmulps		%%ymm12, %%ymm1,  %%ymm12\n\t"  // scale by alpha
+	"vmulps		%%ymm13, %%ymm1,  %%ymm13\n\t"  // scale by alpha
+	"vmulps		%%ymm14, %%ymm1,  %%ymm14\n\t"  // scale by alpha
+	"vmulps		%%ymm15, %%ymm1,  %%ymm15\n\t"  // scale by alpha
+
+	"vmovups	%%ymm8 ,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm9 ,  8*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm10, 16*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm11, 24*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm12, 32*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm13, 40*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm14, 48*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm15, 56*4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y),      // 5
+	  "m" (pre)	// 6
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
+	  "%xmm0", "%xmm1", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
+
+static void  sgemv_kernel_32( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	float *pre = a + lda*3;
+
+	__asm__  __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+	"movq		%6,	 %%r8\n\t"		// address for prefetch
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+
+	"vxorps		%%ymm8 , %%ymm8 , %%ymm8 \n\t"	// set to zero
+	"vxorps		%%ymm9 , %%ymm9 , %%ymm9 \n\t"	// set to zero
+	"vxorps		%%ymm10, %%ymm10, %%ymm10\n\t"	// set to zero
+	"vxorps		%%ymm11, %%ymm11, %%ymm11\n\t"	// set to zero
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%ymm0	 \n\t"	// load values of c
+	"nop					 \n\t"
+	"leaq     (%%r8 , %%rcx, 4), %%r8	 \n\t"	// add lda to pointer for prefetch
+
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+
+	"vmulps   0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
+	"vmulps   8*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
+	"vmulps  16*4(%%rsi), %%ymm0, %%ymm6 \n\t" // multiply a and c and add to temp
+	"vmulps  24*4(%%rsi), %%ymm0, %%ymm7 \n\t" // multiply a and c and add to temp
+
+	"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
+	"vaddps %%ymm9 , %%ymm5, %%ymm9 \n\t" // multiply a and c and add to temp
+	"vaddps %%ymm10, %%ymm6, %%ymm10\n\t" // multiply a and c and add to temp
+	"vaddps %%ymm11, %%ymm7, %%ymm11\n\t" // multiply a and c and add to temp
+
+
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%ymm8 , %%ymm1,  %%ymm8 \n\t"  // scale by alpha
+	"vmulps		%%ymm9 , %%ymm1,  %%ymm9 \n\t"  // scale by alpha
+	"vmulps		%%ymm10, %%ymm1,  %%ymm10\n\t"  // scale by alpha
+	"vmulps		%%ymm11, %%ymm1,  %%ymm11\n\t"  // scale by alpha
+
+	"vmovups	%%ymm8 ,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm9 ,  8*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm10, 16*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm11, 24*4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y),      // 5
+	  "m" (pre)	// 6
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
+	  "%xmm0", "%xmm1", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "memory"
+	);
+
+
+
+} 
+
+static void  sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+	float *pre = a + lda*3;
+	
+	__asm__  __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+	"movq		%6,	 %%r8\n\t"		// address for prefetch
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+
+	"vxorps		%%ymm8 , %%ymm8 , %%ymm8 \n\t"	// set to zero
+	"vxorps		%%ymm9 , %%ymm9 , %%ymm9 \n\t"	// set to zero
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%ymm0	 \n\t"	// load values of c
+	"nop					 \n\t"
+	"leaq     (%%r8 , %%rcx, 4), %%r8	 \n\t"	// add lda to pointer for prefetch
+
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+
+	"vmulps   0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
+	"vmulps   8*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
+
+	"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
+	"vaddps %%ymm9 , %%ymm5, %%ymm9 \n\t" // multiply a and c and add to temp
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%ymm8 , %%ymm1,  %%ymm8 \n\t"  // scale by alpha
+	"vmulps		%%ymm9 , %%ymm1,  %%ymm9 \n\t"  // scale by alpha
+
+	"vmovups	%%ymm8 ,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm9 ,  8*4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y),      // 5
+	  "m" (pre)	// 6
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
+	  "%xmm0", "%xmm1", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "memory"
+	);
+
+
+} 
+
+
+static void  sgemv_kernel_8( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+	
+	__asm__  __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%ymm8 , %%ymm8 , %%ymm8 \n\t"	// set to zero
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%ymm0	 \n\t"	// load values of c
+
+	"vmulps   0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
+	"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%ymm8 , %%ymm1,  %%ymm8 \n\t"  // scale by alpha
+	"vmovups	%%ymm8 ,     (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
+	  "%xmm0", "%xmm1", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "memory"
+	);
+
+
+} 
+
+
+static void  sgemv_kernel_4( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%xmm0	 \n\t"	// load values of c
+
+	"vmulps   0*4(%%rsi), %%xmm0, %%xmm4 \n\t" 		// multiply a and c and add to temp
+	"vaddps %%xmm12, %%xmm4, %%xmm12     \n\t" 		// multiply a and c and add to temp
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%xmm12, %%xmm1,  %%xmm12\n\t"  // scale by alpha
+
+	"vmovups	%%xmm12,     (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+static void  sgemv_kernel_2( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vmovss         %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+	"vxorps		%%xmm13, %%xmm13, %%xmm13\n\t"	// set to zero
+
+	".L01LOOP%=:				 \n\t"
+	"vmovss      	(%%rdi),   %%xmm0	 \n\t"	// load values of c
+
+	"vmulps   0*4(%%rsi), %%xmm0, %%xmm4 \n\t" 		// multiply a and c and add to temp
+	"vmulps   1*4(%%rsi), %%xmm0, %%xmm5 \n\t" 		// multiply a and c and add to temp
+
+	"vaddps %%xmm12, %%xmm4, %%xmm12     \n\t" 		// multiply a and c and add to temp
+	"vaddps %%xmm13, %%xmm5, %%xmm13     \n\t" 		// multiply a and c and add to temp
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulss		%%xmm12, %%xmm1,  %%xmm12\n\t"  // scale by alpha
+	"vmulss		%%xmm13, %%xmm1,  %%xmm13\n\t"  // scale by alpha
+
+	"vmovss 	%%xmm12,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovss 	%%xmm13,    4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
+
+static void  sgemv_kernel_1( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vmovss         %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+
+	".L01LOOP%=:				 \n\t"
+	"vmovss      	(%%rdi),   %%xmm0	 \n\t"	// load values of c
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+
+	"vmulss   0*4(%%rsi), %%xmm0, %%xmm4 \n\t" 		// multiply a and c and add to temp
+	"vaddss %%xmm12, %%xmm4, %%xmm12     \n\t" 		// multiply a and c and add to temp
+
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulss		%%xmm12, %%xmm1,  %%xmm12\n\t"  // scale by alpha
+
+	"vmovss 	%%xmm12,     (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

From 5392d11b045abddfe51c45e69848c807121486e8 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Sun, 20 Jul 2014 14:08:04 +0200
Subject: [PATCH 14/15] optimized sgemv_n_microk_sandy.c

---
 kernel/x86_64/sgemv_n_microk_sandy.c | 9 ++++-----
 1 file changed, 4 insertions(+), 5 deletions(-)

diff --git a/kernel/x86_64/sgemv_n_microk_sandy.c b/kernel/x86_64/sgemv_n_microk_sandy.c
index 7d9360f94..9bdb06600 100644
--- a/kernel/x86_64/sgemv_n_microk_sandy.c
+++ b/kernel/x86_64/sgemv_n_microk_sandy.c
@@ -29,7 +29,7 @@ static void  sgemv_kernel_64( long n, float alpha, float *a, long lda, float *x,
 {
 
 
-	float *pre = a + lda*3;
+	float *pre = a + lda*2;
 
 	__asm__  __volatile__
 	(
@@ -58,20 +58,19 @@ static void  sgemv_kernel_64( long n, float alpha, float *a, long lda, float *x,
 	"leaq     (%%r8 , %%rcx, 4), %%r8	 \n\t"	// add lda to pointer for prefetch
 
 	"prefetcht0	(%%r8)\n\t"			// Prefetch
-	"prefetcht0   64(%%r8)\n\t"			// Prefetch
-	"prefetcht0  128(%%r8)\n\t"			// Prefetch
-	"prefetcht0  192(%%r8)\n\t"			// Prefetch
-
 	"vmulps   0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
 	"vmulps   8*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
 	"vmulps  16*4(%%rsi), %%ymm0, %%ymm6 \n\t" // multiply a and c and add to temp
 	"vmulps  24*4(%%rsi), %%ymm0, %%ymm7 \n\t" // multiply a and c and add to temp
 
 	"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
 	"vaddps %%ymm9 , %%ymm5, %%ymm9 \n\t" // multiply a and c and add to temp
+	"prefetcht0  128(%%r8)\n\t"			// Prefetch
 	"vaddps %%ymm10, %%ymm6, %%ymm10\n\t" // multiply a and c and add to temp
 	"vaddps %%ymm11, %%ymm7, %%ymm11\n\t" // multiply a and c and add to temp
 
+	"prefetcht0  192(%%r8)\n\t"			// Prefetch
 	"vmulps  32*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
 	"vmulps  40*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
 	"vmulps  48*4(%%rsi), %%ymm0, %%ymm6 \n\t" // multiply a and c and add to temp

From 6acbafe45b732b3410e67db5b1a5f05ed5e90f1e Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Sun, 20 Jul 2014 14:52:25 +0200
Subject: [PATCH 15/15] added sgemv_n microkernel for haswell

---
 kernel/x86_64/KERNEL.HASWELL           |   4 +-
 kernel/x86_64/sgemv_n_avx.c            |   2 +
 kernel/x86_64/sgemv_n_microk_haswell.c | 461 +++++++++++++++++++++++++
 3 files changed, 465 insertions(+), 2 deletions(-)
 create mode 100644 kernel/x86_64/sgemv_n_microk_haswell.c

diff --git a/kernel/x86_64/KERNEL.HASWELL b/kernel/x86_64/KERNEL.HASWELL
index 288e39537..871a7d490 100644
--- a/kernel/x86_64/KERNEL.HASWELL
+++ b/kernel/x86_64/KERNEL.HASWELL
@@ -1,8 +1,8 @@
 ifdef OS_WINDOWS
-#SGEMVNKERNEL = ../arm/gemv_n.c
+SGEMVNKERNEL = ../arm/gemv_n.c
 SGEMVTKERNEL = ../arm/gemv_t.c
 else
-#SGEMVNKERNEL = sgemv_n_avx.c
+SGEMVNKERNEL = sgemv_n_avx.c
 SGEMVTKERNEL = sgemv_t_avx.c
 endif
 
diff --git a/kernel/x86_64/sgemv_n_avx.c b/kernel/x86_64/sgemv_n_avx.c
index 96a03ec57..57aaad4b4 100644
--- a/kernel/x86_64/sgemv_n_avx.c
+++ b/kernel/x86_64/sgemv_n_avx.c
@@ -30,6 +30,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #if defined(BULLDOZER) || defined(PILEDRIVER)
 #include "sgemv_n_microk_bulldozer.c"
+#elif defined(HASWELL)
+#include "sgemv_n_microk_haswell.c"
 #else
 #include "sgemv_n_microk_sandy.c"
 #endif
diff --git a/kernel/x86_64/sgemv_n_microk_haswell.c b/kernel/x86_64/sgemv_n_microk_haswell.c
new file mode 100644
index 000000000..9db3869d2
--- /dev/null
+++ b/kernel/x86_64/sgemv_n_microk_haswell.c
@@ -0,0 +1,461 @@
+/***************************************************************************
+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.
+*****************************************************************************/
+
+static void  sgemv_kernel_64( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	float *pre = a + lda*2;
+
+	__asm__  __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+	"movq		%6,	 %%r8\n\t"		// address for prefetch
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+
+	"vxorps		%%ymm8 , %%ymm8 , %%ymm8 \n\t"	// set to zero
+	"vxorps		%%ymm9 , %%ymm9 , %%ymm9 \n\t"	// set to zero
+	"vxorps		%%ymm10, %%ymm10, %%ymm10\n\t"	// set to zero
+	"vxorps		%%ymm11, %%ymm11, %%ymm11\n\t"	// set to zero
+	"vxorps		%%ymm12, %%ymm12, %%ymm12\n\t"	// set to zero
+	"vxorps		%%ymm13, %%ymm13, %%ymm13\n\t"	// set to zero
+	"vxorps		%%ymm14, %%ymm14, %%ymm14\n\t"	// set to zero
+	"vxorps		%%ymm15, %%ymm15, %%ymm15\n\t"	// set to zero
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%ymm0	 \n\t"	// load values of c
+	"leaq     (%%r8 , %%rcx, 4), %%r8	 \n\t"	// add lda to pointer for prefetch
+
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"vfmadd231ps   0*4(%%rsi), %%ymm0, %%ymm8 \n\t" // multiply a and c and add to temp
+	"vfmadd231ps   8*4(%%rsi), %%ymm0, %%ymm9 \n\t" // multiply a and c and add to temp
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+	"vfmadd231ps  16*4(%%rsi), %%ymm0, %%ymm10\n\t" // multiply a and c and add to temp
+	"vfmadd231ps  24*4(%%rsi), %%ymm0, %%ymm11\n\t" // multiply a and c and add to temp
+	"prefetcht0  128(%%r8)\n\t"			// Prefetch
+	"vfmadd231ps  32*4(%%rsi), %%ymm0, %%ymm12\n\t" // multiply a and c and add to temp
+	"vfmadd231ps  40*4(%%rsi), %%ymm0, %%ymm13\n\t" // multiply a and c and add to temp
+	"prefetcht0  192(%%r8)\n\t"			// Prefetch
+	"vfmadd231ps  48*4(%%rsi), %%ymm0, %%ymm14\n\t" // multiply a and c and add to temp
+	"vfmadd231ps  56*4(%%rsi), %%ymm0, %%ymm15\n\t" // multiply a and c and add to temp
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%ymm8 , %%ymm1,  %%ymm8 \n\t"  // scale by alpha
+	"vmulps		%%ymm9 , %%ymm1,  %%ymm9 \n\t"  // scale by alpha
+	"vmulps		%%ymm10, %%ymm1,  %%ymm10\n\t"  // scale by alpha
+	"vmulps		%%ymm11, %%ymm1,  %%ymm11\n\t"  // scale by alpha
+	"vmulps		%%ymm12, %%ymm1,  %%ymm12\n\t"  // scale by alpha
+	"vmulps		%%ymm13, %%ymm1,  %%ymm13\n\t"  // scale by alpha
+	"vmulps		%%ymm14, %%ymm1,  %%ymm14\n\t"  // scale by alpha
+	"vmulps		%%ymm15, %%ymm1,  %%ymm15\n\t"  // scale by alpha
+
+	"vmovups	%%ymm8 ,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm9 ,  8*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm10, 16*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm11, 24*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm12, 32*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm13, 40*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm14, 48*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm15, 56*4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y),      // 5
+	  "m" (pre)	// 6
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
+	  "%xmm0", "%xmm1", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
+
+static void  sgemv_kernel_32( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	float *pre = a + lda*3;
+
+	__asm__  __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+	"movq		%6,	 %%r8\n\t"		// address for prefetch
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+
+	"vxorps		%%ymm8 , %%ymm8 , %%ymm8 \n\t"	// set to zero
+	"vxorps		%%ymm9 , %%ymm9 , %%ymm9 \n\t"	// set to zero
+	"vxorps		%%ymm10, %%ymm10, %%ymm10\n\t"	// set to zero
+	"vxorps		%%ymm11, %%ymm11, %%ymm11\n\t"	// set to zero
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%ymm0	 \n\t"	// load values of c
+	"nop					 \n\t"
+	"leaq     (%%r8 , %%rcx, 4), %%r8	 \n\t"	// add lda to pointer for prefetch
+
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+
+	"vmulps   0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
+	"vmulps   8*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
+	"vmulps  16*4(%%rsi), %%ymm0, %%ymm6 \n\t" // multiply a and c and add to temp
+	"vmulps  24*4(%%rsi), %%ymm0, %%ymm7 \n\t" // multiply a and c and add to temp
+
+	"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
+	"vaddps %%ymm9 , %%ymm5, %%ymm9 \n\t" // multiply a and c and add to temp
+	"vaddps %%ymm10, %%ymm6, %%ymm10\n\t" // multiply a and c and add to temp
+	"vaddps %%ymm11, %%ymm7, %%ymm11\n\t" // multiply a and c and add to temp
+
+
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%ymm8 , %%ymm1,  %%ymm8 \n\t"  // scale by alpha
+	"vmulps		%%ymm9 , %%ymm1,  %%ymm9 \n\t"  // scale by alpha
+	"vmulps		%%ymm10, %%ymm1,  %%ymm10\n\t"  // scale by alpha
+	"vmulps		%%ymm11, %%ymm1,  %%ymm11\n\t"  // scale by alpha
+
+	"vmovups	%%ymm8 ,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm9 ,  8*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm10, 16*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm11, 24*4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y),      // 5
+	  "m" (pre)	// 6
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
+	  "%xmm0", "%xmm1", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "memory"
+	);
+
+
+
+} 
+
+static void  sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+	float *pre = a + lda*3;
+	
+	__asm__  __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+	"movq		%6,	 %%r8\n\t"		// address for prefetch
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+
+	"vxorps		%%ymm8 , %%ymm8 , %%ymm8 \n\t"	// set to zero
+	"vxorps		%%ymm9 , %%ymm9 , %%ymm9 \n\t"	// set to zero
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%ymm0	 \n\t"	// load values of c
+	"nop					 \n\t"
+	"leaq     (%%r8 , %%rcx, 4), %%r8	 \n\t"	// add lda to pointer for prefetch
+
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+
+	"vmulps   0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
+	"vmulps   8*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
+
+	"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
+	"vaddps %%ymm9 , %%ymm5, %%ymm9 \n\t" // multiply a and c and add to temp
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%ymm8 , %%ymm1,  %%ymm8 \n\t"  // scale by alpha
+	"vmulps		%%ymm9 , %%ymm1,  %%ymm9 \n\t"  // scale by alpha
+
+	"vmovups	%%ymm8 ,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm9 ,  8*4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y),      // 5
+	  "m" (pre)	// 6
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
+	  "%xmm0", "%xmm1", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "memory"
+	);
+
+
+} 
+
+
+static void  sgemv_kernel_8( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+	
+	__asm__  __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%ymm8 , %%ymm8 , %%ymm8 \n\t"	// set to zero
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%ymm0	 \n\t"	// load values of c
+
+	"vmulps   0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
+	"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%ymm8 , %%ymm1,  %%ymm8 \n\t"  // scale by alpha
+	"vmovups	%%ymm8 ,     (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
+	  "%xmm0", "%xmm1", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "memory"
+	);
+
+
+} 
+
+
+static void  sgemv_kernel_4( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%xmm0	 \n\t"	// load values of c
+
+	"vmulps   0*4(%%rsi), %%xmm0, %%xmm4 \n\t" 		// multiply a and c and add to temp
+	"vaddps %%xmm12, %%xmm4, %%xmm12     \n\t" 		// multiply a and c and add to temp
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%xmm12, %%xmm1,  %%xmm12\n\t"  // scale by alpha
+
+	"vmovups	%%xmm12,     (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+static void  sgemv_kernel_2( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vmovss         %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+	"vxorps		%%xmm13, %%xmm13, %%xmm13\n\t"	// set to zero
+
+	".L01LOOP%=:				 \n\t"
+	"vmovss      	(%%rdi),   %%xmm0	 \n\t"	// load values of c
+
+	"vmulps   0*4(%%rsi), %%xmm0, %%xmm4 \n\t" 		// multiply a and c and add to temp
+	"vmulps   1*4(%%rsi), %%xmm0, %%xmm5 \n\t" 		// multiply a and c and add to temp
+
+	"vaddps %%xmm12, %%xmm4, %%xmm12     \n\t" 		// multiply a and c and add to temp
+	"vaddps %%xmm13, %%xmm5, %%xmm13     \n\t" 		// multiply a and c and add to temp
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulss		%%xmm12, %%xmm1,  %%xmm12\n\t"  // scale by alpha
+	"vmulss		%%xmm13, %%xmm1,  %%xmm13\n\t"  // scale by alpha
+
+	"vmovss 	%%xmm12,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovss 	%%xmm13,    4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
+
+static void  sgemv_kernel_1( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vmovss         %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+
+	".L01LOOP%=:				 \n\t"
+	"vmovss      	(%%rdi),   %%xmm0	 \n\t"	// load values of c
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+
+	"vmulss   0*4(%%rsi), %%xmm0, %%xmm4 \n\t" 		// multiply a and c and add to temp
+	"vaddss %%xmm12, %%xmm4, %%xmm12     \n\t" 		// multiply a and c and add to temp
+
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulss		%%xmm12, %%xmm1,  %%xmm12\n\t"  // scale by alpha
+
+	"vmovss 	%%xmm12,     (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+