diff --git a/benchmark/Makefile b/benchmark/Makefile
index cf219cef1..402a2e07b 100644
--- a/benchmark/Makefile
+++ b/benchmark/Makefile
@@ -6,8 +6,13 @@ include $(TOPDIR)/Makefile.system
 #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
+
+# ACML 6.1 custom
+ACML=/home/saar/acml6.1/gfortran64_mp/lib
+LIBACML = -fopenmp $(ACML)/libacml_mp.so -lgfortran -lm
+ 
 
 # Atlas Ubuntu 
 #ATLAS=/usr/lib/atlas-base
diff --git a/benchmark/axpy.c b/benchmark/axpy.c
index ef3b5ae4f..a7206b690 100644
--- a/benchmark/axpy.c
+++ b/benchmark/axpy.c
@@ -114,7 +114,7 @@ static void *huge_malloc(BLASLONG size){
 
 #endif
 
-int MAIN__(int argc, char *argv[]){
+int main(int argc, char *argv[]){
 
   FLOAT *x, *y;
   FLOAT alpha[2] = { 2.0, 2.0 };
@@ -198,4 +198,4 @@ int MAIN__(int argc, char *argv[]){
   return 0;
 }
 
-void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
+// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
diff --git a/benchmark/cholesky.c b/benchmark/cholesky.c
index 76c368eda..c8b96d80f 100644
--- a/benchmark/cholesky.c
+++ b/benchmark/cholesky.c
@@ -117,7 +117,7 @@ static __inline double getmflops(int ratio, int m, double secs){
 }
 
 
-int MAIN__(int argc, char *argv[]){
+int main(int argc, char *argv[]){
 
 #ifndef COMPLEX
   char *trans[] = {"T", "N"};
@@ -273,4 +273,4 @@ int MAIN__(int argc, char *argv[]){
   return 0;
 }
 
-void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
+// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
diff --git a/benchmark/dot.c b/benchmark/dot.c
index 6132ed324..4c8d6cc38 100644
--- a/benchmark/dot.c
+++ b/benchmark/dot.c
@@ -108,7 +108,7 @@ static void *huge_malloc(BLASLONG size){
 
 #endif
 
-int MAIN__(int argc, char *argv[]){
+int main(int argc, char *argv[]){
 
   FLOAT *x, *y;
   FLOAT result;
@@ -192,4 +192,4 @@ int MAIN__(int argc, char *argv[]){
   return 0;
 }
 
-void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
+// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
diff --git a/benchmark/geev.c b/benchmark/geev.c
index 3b7465360..a2ca2c315 100644
--- a/benchmark/geev.c
+++ b/benchmark/geev.c
@@ -139,7 +139,7 @@ static void *huge_malloc(BLASLONG size){
 
 #endif
 
-int MAIN__(int argc, char *argv[]){
+int main(int argc, char *argv[]){
 
   FLOAT *a,*vl,*vr,*wi,*wr,*work,*rwork;
   FLOAT wkopt[4];
@@ -257,4 +257,4 @@ int MAIN__(int argc, char *argv[]){
   return 0;
 }
 
-void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
+// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
diff --git a/benchmark/gemm.c b/benchmark/gemm.c
index 4f9a58825..5a3587622 100644
--- a/benchmark/gemm.c
+++ b/benchmark/gemm.c
@@ -118,14 +118,15 @@ static void *huge_malloc(BLASLONG size){
 
 #endif
 
-int MAIN__(int argc, char *argv[]){
+int main(int argc, char *argv[]){
 
   FLOAT *a, *b, *c;
   FLOAT alpha[] = {1.0, 1.0};
   FLOAT beta [] = {1.0, 1.0};
   char trans='N';
-  blasint m, i, j;
+  blasint m, n, i, j;
   int loops = 1;
+  int has_param_n=0;
   int l;
   char *p;
 
@@ -162,6 +163,11 @@ int MAIN__(int argc, char *argv[]){
   if ( p != NULL )
 	loops = atoi(p);
 
+  if ((p = getenv("OPENBLAS_PARAM_N"))) {
+          n = atoi(p);
+	  has_param_n=1;	  
+  }
+
 
 #ifdef linux
   srandom(getpid());
@@ -174,7 +180,14 @@ int MAIN__(int argc, char *argv[]){
 
     timeg=0;
 
-    fprintf(stderr, " %6d : ", (int)m);
+    if ( has_param_n == 1 && n <= m )
+    	n=n;
+    else
+    	n=m;
+
+
+
+    fprintf(stderr, " %6dx%d : ", (int)m, (int)n);
 
     for (l=0; l<loops; l++)
     {
@@ -189,7 +202,7 @@ int MAIN__(int argc, char *argv[]){
 
     	gettimeofday( &start, (struct timezone *)0);
 
-    	GEMM (&trans, &trans, &m, &m, &m, alpha, a, &m, b, &m, beta, c, &m );
+    	GEMM (&trans, &trans, &m, &n, &m, alpha, a, &m, b, &m, beta, c, &m );
 
     	gettimeofday( &stop, (struct timezone *)0);
 
@@ -202,11 +215,11 @@ int MAIN__(int argc, char *argv[]){
     timeg /= loops;
     fprintf(stderr,
 	    " %10.2f MFlops\n",
-	    COMPSIZE * COMPSIZE * 2. * (double)m * (double)m * (double)m / timeg * 1.e-6);
+	    COMPSIZE * COMPSIZE * 2. * (double)m * (double)m * (double)n / timeg * 1.e-6);
 
   }
 
   return 0;
 }
 
-void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
+// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
diff --git a/benchmark/gemm3m.c b/benchmark/gemm3m.c
index 048d74be6..d39543585 100644
--- a/benchmark/gemm3m.c
+++ b/benchmark/gemm3m.c
@@ -118,7 +118,7 @@ static void *huge_malloc(BLASLONG size){
 
 #endif
 
-int MAIN__(int argc, char *argv[]){
+int main(int argc, char *argv[]){
 
   FLOAT *a, *b, *c;
   FLOAT alpha[] = {1.0, 1.0};
@@ -209,4 +209,4 @@ int MAIN__(int argc, char *argv[]){
   return 0;
 }
 
-void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
+// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
diff --git a/benchmark/gemv.c b/benchmark/gemv.c
index e21868259..42af2825a 100644
--- a/benchmark/gemv.c
+++ b/benchmark/gemv.c
@@ -118,7 +118,7 @@ static void *huge_malloc(BLASLONG size){
 
 #endif
 
-int MAIN__(int argc, char *argv[]){
+int main(int argc, char *argv[]){
 
   FLOAT *a, *x, *y;
   FLOAT alpha[] = {1.0, 1.0};
@@ -266,4 +266,4 @@ int MAIN__(int argc, char *argv[]){
   return 0;
 }
 
-void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
+// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
diff --git a/benchmark/ger.c b/benchmark/ger.c
index 5085389da..354281006 100644
--- a/benchmark/ger.c
+++ b/benchmark/ger.c
@@ -108,7 +108,7 @@ static void *huge_malloc(BLASLONG size){
 
 #endif
 
-int MAIN__(int argc, char *argv[]){
+int main(int argc, char *argv[]){
 
   FLOAT *a, *x, *y;
   FLOAT alpha[] = {1.0, 1.0};
@@ -214,5 +214,5 @@ int MAIN__(int argc, char *argv[]){
   return 0;
 }
 
-void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
+// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
 
diff --git a/benchmark/getri.c b/benchmark/getri.c
index 897f1ff04..083cdc9aa 100644
--- a/benchmark/getri.c
+++ b/benchmark/getri.c
@@ -137,7 +137,7 @@ static void *huge_malloc(BLASLONG size){
 
 #endif
 
-int MAIN__(int argc, char *argv[]){
+int main(int argc, char *argv[]){
 
   FLOAT *a,*work;
   FLOAT wkopt[4];
@@ -231,4 +231,4 @@ int MAIN__(int argc, char *argv[]){
   return 0;
 }
 
-void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
+// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
diff --git a/benchmark/hemm.c b/benchmark/hemm.c
index f5d4b4fd9..318c407ba 100644
--- a/benchmark/hemm.c
+++ b/benchmark/hemm.c
@@ -107,7 +107,7 @@ static void *huge_malloc(BLASLONG size){
 
 #endif
 
-int MAIN__(int argc, char *argv[]){
+int main(int argc, char *argv[]){
 
   FLOAT *a, *b, *c;
   FLOAT alpha[] = {1.0, 1.0};
@@ -189,4 +189,4 @@ int MAIN__(int argc, char *argv[]){
   return 0;
 }
 
-void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
+// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
diff --git a/benchmark/hemv.c b/benchmark/hemv.c
index 79b7679cc..05028e3cf 100644
--- a/benchmark/hemv.c
+++ b/benchmark/hemv.c
@@ -108,7 +108,7 @@ static void *huge_malloc(BLASLONG size){
 
 #endif
 
-int MAIN__(int argc, char *argv[]){
+int main(int argc, char *argv[]){
 
   FLOAT *a, *x, *y;
   FLOAT alpha[] = {1.0, 1.0};
@@ -205,4 +205,4 @@ int MAIN__(int argc, char *argv[]){
   return 0;
 }
 
-void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
+// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
diff --git a/benchmark/her2k.c b/benchmark/her2k.c
index 49ab8d214..028e2718f 100644
--- a/benchmark/her2k.c
+++ b/benchmark/her2k.c
@@ -106,7 +106,7 @@ static void *huge_malloc(BLASLONG size){
 
 #endif
 
-int MAIN__(int argc, char *argv[]){
+int main(int argc, char *argv[]){
 
   FLOAT *a, *b, *c;
   FLOAT alpha[] = {1.0, 1.0};
@@ -188,4 +188,4 @@ int MAIN__(int argc, char *argv[]){
   return 0;
 }
 
-void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
+// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
diff --git a/benchmark/herk.c b/benchmark/herk.c
index 8c053b019..d2e25ff46 100644
--- a/benchmark/herk.c
+++ b/benchmark/herk.c
@@ -108,7 +108,7 @@ static void *huge_malloc(BLASLONG size){
 
 #endif
 
-int MAIN__(int argc, char *argv[]){
+int main(int argc, char *argv[]){
 
   FLOAT *a, *c;
   FLOAT alpha[] = {1.0, 1.0};
@@ -186,4 +186,4 @@ int MAIN__(int argc, char *argv[]){
   return 0;
 }
 
-void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
+// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
diff --git a/benchmark/linpack.c b/benchmark/linpack.c
index 98a874208..7d5c87163 100644
--- a/benchmark/linpack.c
+++ b/benchmark/linpack.c
@@ -137,7 +137,7 @@ static void *huge_malloc(BLASLONG size){
 
 #endif
 
-int MAIN__(int argc, char *argv[]){
+int main(int argc, char *argv[]){
 
   FLOAT *a, *b;
   blasint *ipiv;
@@ -270,4 +270,4 @@ int MAIN__(int argc, char *argv[]){
   return 0;
 }
 
-void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
+// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
diff --git a/benchmark/potrf.c b/benchmark/potrf.c
index 7b6cdd799..3caf61caa 100644
--- a/benchmark/potrf.c
+++ b/benchmark/potrf.c
@@ -114,7 +114,7 @@ int gettimeofday(struct timeval *tv, void *tz){
 
 #endif
 
-int MAIN__(int argc, char *argv[]){
+int main(int argc, char *argv[]){
 
 #ifndef COMPLEX
   char *trans[] = {"T", "N"};
@@ -278,5 +278,5 @@ int MAIN__(int argc, char *argv[]){
   return 0;
 }
 
-void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
+// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
 
diff --git a/benchmark/symm.c b/benchmark/symm.c
index 187dfe2ae..35ebcee97 100644
--- a/benchmark/symm.c
+++ b/benchmark/symm.c
@@ -118,7 +118,7 @@ static void *huge_malloc(BLASLONG size){
 
 #endif
 
-int MAIN__(int argc, char *argv[]){
+int main(int argc, char *argv[]){
 
   FLOAT *a, *b, *c;
   FLOAT alpha[] = {1.0, 1.0};
@@ -200,4 +200,4 @@ int MAIN__(int argc, char *argv[]){
   return 0;
 }
 
-void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
+// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
diff --git a/benchmark/symv.c b/benchmark/symv.c
index 4bcfb411b..df2a5d301 100644
--- a/benchmark/symv.c
+++ b/benchmark/symv.c
@@ -118,7 +118,7 @@ static void *huge_malloc(BLASLONG size){
 
 #endif
 
-int MAIN__(int argc, char *argv[]){
+int main(int argc, char *argv[]){
 
   FLOAT *a, *x, *y;
   FLOAT alpha[] = {1.0, 1.0};
@@ -215,4 +215,4 @@ int MAIN__(int argc, char *argv[]){
   return 0;
 }
 
-void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
+// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
diff --git a/benchmark/syr2k.c b/benchmark/syr2k.c
index e11b04e42..9840b5f3e 100644
--- a/benchmark/syr2k.c
+++ b/benchmark/syr2k.c
@@ -118,7 +118,7 @@ static void *huge_malloc(BLASLONG size){
 
 #endif
 
-int MAIN__(int argc, char *argv[]){
+int main(int argc, char *argv[]){
 
   FLOAT *a, *b, *c;
   FLOAT alpha[] = {1.0, 1.0};
@@ -200,4 +200,4 @@ int MAIN__(int argc, char *argv[]){
   return 0;
 }
 
-void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
+// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
diff --git a/benchmark/syrk.c b/benchmark/syrk.c
index f01549688..34817f2bb 100644
--- a/benchmark/syrk.c
+++ b/benchmark/syrk.c
@@ -118,7 +118,7 @@ static void *huge_malloc(BLASLONG size){
 
 #endif
 
-int MAIN__(int argc, char *argv[]){
+int main(int argc, char *argv[]){
 
   FLOAT *a, *c;
   FLOAT alpha[] = {1.0, 1.0};
@@ -196,4 +196,4 @@ int MAIN__(int argc, char *argv[]){
   return 0;
 }
 
-void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
+// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
diff --git a/benchmark/trmm.c b/benchmark/trmm.c
index 328dc9a10..f81e9d912 100644
--- a/benchmark/trmm.c
+++ b/benchmark/trmm.c
@@ -118,7 +118,7 @@ static void *huge_malloc(BLASLONG size){
 
 #endif
 
-int MAIN__(int argc, char *argv[]){
+int main(int argc, char *argv[]){
 
   FLOAT *a, *b;
   FLOAT alpha[] = {1.0, 1.0};
@@ -199,4 +199,4 @@ int MAIN__(int argc, char *argv[]){
   return 0;
 }
 
-void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
+// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
diff --git a/benchmark/trsm.c b/benchmark/trsm.c
index 908a0fcb7..ed969b707 100644
--- a/benchmark/trsm.c
+++ b/benchmark/trsm.c
@@ -118,7 +118,7 @@ static void *huge_malloc(BLASLONG size){
 
 #endif
 
-int MAIN__(int argc, char *argv[]){
+int main(int argc, char *argv[]){
 
   FLOAT *a, *b;
   FLOAT alpha[] = {1.0, 1.0};
@@ -199,4 +199,4 @@ int MAIN__(int argc, char *argv[]){
   return 0;
 }
 
-void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
+// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
diff --git a/kernel/x86_64/KERNEL.SANDYBRIDGE b/kernel/x86_64/KERNEL.SANDYBRIDGE
index 61e13a116..ff96cd011 100644
--- a/kernel/x86_64/KERNEL.SANDYBRIDGE
+++ b/kernel/x86_64/KERNEL.SANDYBRIDGE
@@ -34,17 +34,17 @@ CGEMMITCOPYOBJ =  cgemm_itcopy$(TSUFFIX).$(SUFFIX)
 CGEMMONCOPYOBJ =  cgemm_oncopy$(TSUFFIX).$(SUFFIX)
 CGEMMOTCOPYOBJ =  cgemm_otcopy$(TSUFFIX).$(SUFFIX)
 
-
-ZGEMMKERNEL    =  zgemm_kernel_4x4_sandy.S
-ZGEMMINCOPY    =
-ZGEMMITCOPY    =
+ZGEMMKERNEL    =  zgemm_kernel_1x4_nehalem.S
+ZGEMMINCOPY    =  zgemm_ncopy_1.S
+ZGEMMITCOPY    =  zgemm_tcopy_1.S
 ZGEMMONCOPY    =  ../generic/zgemm_ncopy_4.c
 ZGEMMOTCOPY    =  ../generic/zgemm_tcopy_4.c
-ZGEMMINCOPYOBJ =
-ZGEMMITCOPYOBJ =
+ZGEMMINCOPYOBJ =  zgemm_incopy$(TSUFFIX).$(SUFFIX)
+ZGEMMITCOPYOBJ =  zgemm_itcopy$(TSUFFIX).$(SUFFIX)
 ZGEMMONCOPYOBJ =  zgemm_oncopy$(TSUFFIX).$(SUFFIX)
 ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)
 
+
 #STRSMKERNEL_LN	=  trsm_kernel_LN_4x8_nehalem.S
 #STRSMKERNEL_LT	=  trsm_kernel_LT_4x8_nehalem.S
 #STRSMKERNEL_RN	=  trsm_kernel_LT_4x8_nehalem.S
diff --git a/kernel/x86_64/dgemm_kernel_4x4_haswell.S b/kernel/x86_64/dgemm_kernel_4x4_haswell.S
index a49a51ee9..0a2ca7ae3 100644
--- a/kernel/x86_64/dgemm_kernel_4x4_haswell.S
+++ b/kernel/x86_64/dgemm_kernel_4x4_haswell.S
@@ -1092,18 +1092,48 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x1
 
-	vxorpd		%xmm4 , %xmm4 , %xmm4
-	vxorpd		%xmm5 , %xmm5 , %xmm5
+	vxorpd		%ymm4 , %ymm4 , %ymm4
+	vxorpd		%ymm5 , %ymm5 , %ymm5
+	vxorpd		%ymm6 , %ymm6 , %ymm6
+	vxorpd		%ymm7 , %ymm7 , %ymm7
+
+.endm
+
+
+.macro KERNEL4x1
+
+	vbroadcastsd	-12 * SIZE(BO), %ymm0
+	vbroadcastsd	-11 * SIZE(BO), %ymm1
+	vbroadcastsd	-10 * SIZE(BO), %ymm2
+	vbroadcastsd	-9  * SIZE(BO), %ymm3
+
+	vfmadd231pd  	-16 * SIZE(AO) ,%ymm0  , %ymm4
+	vfmadd231pd  	-12 * SIZE(AO) ,%ymm1  , %ymm5
+
+	vbroadcastsd	-8  * SIZE(BO), %ymm0
+	vbroadcastsd	-7  * SIZE(BO), %ymm1
+
+	vfmadd231pd  	-8  * SIZE(AO) ,%ymm2  , %ymm6
+	vfmadd231pd  	-4  * SIZE(AO) ,%ymm3  , %ymm7
+
+	vbroadcastsd	-6  * SIZE(BO), %ymm2
+	vbroadcastsd	-5  * SIZE(BO), %ymm3
+
+	vfmadd231pd  	 0  * SIZE(AO) ,%ymm0  , %ymm4
+	vfmadd231pd  	 4  * SIZE(AO) ,%ymm1  , %ymm5
+	vfmadd231pd  	 8  * SIZE(AO) ,%ymm2  , %ymm6
+	vfmadd231pd  	 12 * SIZE(AO) ,%ymm3  , %ymm7
+
+	addq		$ 8 *SIZE, BO
+	addq		$ 32*SIZE, AO
 
 .endm
 
 
 .macro KERNEL4x1_SUB
-	vmovddup	-12 * SIZE(BO), %xmm2
-	vmovups 	-16 * SIZE(AO), %xmm0
-	vmovups 	-14 * SIZE(AO), %xmm1
-	vfmadd231pd  	%xmm0 ,%xmm2  , %xmm4
-	vfmadd231pd  	%xmm1 ,%xmm2  , %xmm5
+	vbroadcastsd	-12 * SIZE(BO), %ymm2
+	vmovups 	-16 * SIZE(AO), %ymm0
+	vfmadd231pd  	%ymm0 ,%ymm2  , %ymm4
 	addq		$ 1*SIZE, BO
 	addq		$ 4*SIZE, AO
 
@@ -1112,21 +1142,22 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro SAVE4x1
 
-	vmovddup	ALPHA, %xmm0
+	vbroadcastsd	ALPHA, %ymm0
 
-	vmulpd	%xmm0 , %xmm4 , %xmm4
-	vmulpd	%xmm0 , %xmm5 , %xmm5
+	vaddpd	%ymm4,%ymm5, %ymm4 
+	vaddpd	%ymm6,%ymm7, %ymm6 
+	vaddpd	%ymm4,%ymm6, %ymm4 
+
+	vmulpd	%ymm0 , %ymm4 , %ymm4
 
 
 #if !defined(TRMMKERNEL)
 
-	vaddpd 	                (CO1)     , %xmm4, %xmm4
-	vaddpd 	        2 * SIZE(CO1)     , %xmm5, %xmm5
+	vaddpd 	                (CO1)     , %ymm4, %ymm4
 
 #endif
 
-	vmovups	%xmm4 ,  	(CO1)
-	vmovups	%xmm5 , 2 * SIZE(CO1)
+	vmovups	%ymm4 ,  	(CO1)
 
 	addq	$ 4*SIZE, CO1
 .endm
@@ -2112,15 +2143,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .L1_12:
 
-	KERNEL4x1_SUB
-	KERNEL4x1_SUB
-	KERNEL4x1_SUB
-	KERNEL4x1_SUB
-
-	KERNEL4x1_SUB
-	KERNEL4x1_SUB
-	KERNEL4x1_SUB
-	KERNEL4x1_SUB
+	KERNEL4x1
 
 	dec	%rax
 	jne	.L1_12
@@ -3180,15 +3203,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .L1_12:
 
-	KERNEL4x1_SUB
-	KERNEL4x1_SUB
-	KERNEL4x1_SUB
-	KERNEL4x1_SUB
-
-	KERNEL4x1_SUB
-	KERNEL4x1_SUB
-	KERNEL4x1_SUB
-	KERNEL4x1_SUB
+	KERNEL4x1
 
 	dec	%rax
 	jne	.L1_12
diff --git a/lapack-netlib/BLAS/TESTING/cblat2.f b/lapack-netlib/BLAS/TESTING/cblat2.f
index 5833ea81a..2a6edd382 100644
--- a/lapack-netlib/BLAS/TESTING/cblat2.f
+++ b/lapack-netlib/BLAS/TESTING/cblat2.f
@@ -120,7 +120,7 @@
       REAL               RZERO
       PARAMETER          ( RZERO = 0.0 )
       INTEGER            NMAX, INCMAX
-      PARAMETER          ( NMAX = 65, INCMAX = 2 )
+      PARAMETER          ( NMAX = 128, INCMAX = 2 )
       INTEGER            NINMAX, NIDMAX, NKBMAX, NALMAX, NBEMAX
       PARAMETER          ( NINMAX = 7, NIDMAX = 9, NKBMAX = 7,
      $                   NALMAX = 7, NBEMAX = 7 )
diff --git a/lapack-netlib/BLAS/TESTING/cblat3.f b/lapack-netlib/BLAS/TESTING/cblat3.f
index 09f2cb9c5..fb2aa4ece 100644
--- a/lapack-netlib/BLAS/TESTING/cblat3.f
+++ b/lapack-netlib/BLAS/TESTING/cblat3.f
@@ -102,7 +102,7 @@
       REAL               RZERO
       PARAMETER          ( RZERO = 0.0 )
       INTEGER            NMAX
-      PARAMETER          ( NMAX = 65 )
+      PARAMETER          ( NMAX = 128 )
       INTEGER            NIDMAX, NALMAX, NBEMAX
       PARAMETER          ( NIDMAX = 9, NALMAX = 7, NBEMAX = 7 )
 *     .. Local Scalars ..
diff --git a/lapack-netlib/BLAS/TESTING/dblat2.f b/lapack-netlib/BLAS/TESTING/dblat2.f
index 0fa80afa4..80623b260 100644
--- a/lapack-netlib/BLAS/TESTING/dblat2.f
+++ b/lapack-netlib/BLAS/TESTING/dblat2.f
@@ -117,7 +117,7 @@
       DOUBLE PRECISION   ZERO, ONE
       PARAMETER          ( ZERO = 0.0D0, ONE = 1.0D0 )
       INTEGER            NMAX, INCMAX
-      PARAMETER          ( NMAX = 65, INCMAX = 2 )
+      PARAMETER          ( NMAX = 128, INCMAX = 2 )
       INTEGER            NINMAX, NIDMAX, NKBMAX, NALMAX, NBEMAX
       PARAMETER          ( NINMAX = 7, NIDMAX = 9, NKBMAX = 7,
      $                   NALMAX = 7, NBEMAX = 7 )
diff --git a/lapack-netlib/BLAS/TESTING/dblat3.f b/lapack-netlib/BLAS/TESTING/dblat3.f
index 8d37c7453..72c17ed3b 100644
--- a/lapack-netlib/BLAS/TESTING/dblat3.f
+++ b/lapack-netlib/BLAS/TESTING/dblat3.f
@@ -97,7 +97,7 @@
       DOUBLE PRECISION   ZERO, ONE
       PARAMETER          ( ZERO = 0.0D0, ONE = 1.0D0 )
       INTEGER            NMAX
-      PARAMETER          ( NMAX = 65 )
+      PARAMETER          ( NMAX = 128 )
       INTEGER            NIDMAX, NALMAX, NBEMAX
       PARAMETER          ( NIDMAX = 9, NALMAX = 7, NBEMAX = 7 )
 *     .. Local Scalars ..
diff --git a/lapack-netlib/BLAS/TESTING/sblat2.f b/lapack-netlib/BLAS/TESTING/sblat2.f
index 71605ed31..601add7e9 100644
--- a/lapack-netlib/BLAS/TESTING/sblat2.f
+++ b/lapack-netlib/BLAS/TESTING/sblat2.f
@@ -117,7 +117,7 @@
       REAL               ZERO, ONE
       PARAMETER          ( ZERO = 0.0, ONE = 1.0 )
       INTEGER            NMAX, INCMAX
-      PARAMETER          ( NMAX = 65, INCMAX = 2 )
+      PARAMETER          ( NMAX = 128, INCMAX = 2 )
       INTEGER            NINMAX, NIDMAX, NKBMAX, NALMAX, NBEMAX
       PARAMETER          ( NINMAX = 7, NIDMAX = 9, NKBMAX = 7,
      $                   NALMAX = 7, NBEMAX = 7 )
diff --git a/lapack-netlib/BLAS/TESTING/sblat3.f b/lapack-netlib/BLAS/TESTING/sblat3.f
index 879269633..78d809379 100644
--- a/lapack-netlib/BLAS/TESTING/sblat3.f
+++ b/lapack-netlib/BLAS/TESTING/sblat3.f
@@ -97,7 +97,7 @@
       REAL               ZERO, ONE
       PARAMETER          ( ZERO = 0.0, ONE = 1.0 )
       INTEGER            NMAX
-      PARAMETER          ( NMAX = 65 )
+      PARAMETER          ( NMAX = 128 )
       INTEGER            NIDMAX, NALMAX, NBEMAX
       PARAMETER          ( NIDMAX = 9, NALMAX = 7, NBEMAX = 7 )
 *     .. Local Scalars ..
diff --git a/lapack-netlib/BLAS/TESTING/zblat2.f b/lapack-netlib/BLAS/TESTING/zblat2.f
index 53129a11e..2e3e08e7c 100644
--- a/lapack-netlib/BLAS/TESTING/zblat2.f
+++ b/lapack-netlib/BLAS/TESTING/zblat2.f
@@ -121,7 +121,7 @@
       DOUBLE PRECISION   RZERO
       PARAMETER          ( RZERO = 0.0D0 )
       INTEGER            NMAX, INCMAX
-      PARAMETER          ( NMAX = 65, INCMAX = 2 )
+      PARAMETER          ( NMAX = 128, INCMAX = 2 )
       INTEGER            NINMAX, NIDMAX, NKBMAX, NALMAX, NBEMAX
       PARAMETER          ( NINMAX = 7, NIDMAX = 9, NKBMAX = 7,
      $                   NALMAX = 7, NBEMAX = 7 )
diff --git a/lapack-netlib/BLAS/TESTING/zblat3.f b/lapack-netlib/BLAS/TESTING/zblat3.f
index 59ca24145..39ce06b99 100644
--- a/lapack-netlib/BLAS/TESTING/zblat3.f
+++ b/lapack-netlib/BLAS/TESTING/zblat3.f
@@ -104,7 +104,7 @@
       DOUBLE PRECISION   RZERO
       PARAMETER          ( RZERO = 0.0D0 )
       INTEGER            NMAX
-      PARAMETER          ( NMAX = 65 )
+      PARAMETER          ( NMAX = 128 )
       INTEGER            NIDMAX, NALMAX, NBEMAX
       PARAMETER          ( NIDMAX = 9, NALMAX = 7, NBEMAX = 7 )
 *     .. Local Scalars ..
diff --git a/lapack-netlib/TESTING/sep.in b/lapack-netlib/TESTING/sep.in
index e0ed58512..19bd7c3da 100644
--- a/lapack-netlib/TESTING/sep.in
+++ b/lapack-netlib/TESTING/sep.in
@@ -1,11 +1,11 @@
 SEP:  Data file for testing Symmetric Eigenvalue Problem routines
-6                                 Number of values of N
-0 1 2 3 5 20                      Values of N (dimension)
+8                                 Number of values of N
+0 1 2 3 5 19 20 21                Values of N (dimension)
 5                                 Number of values of NB
 1 3  3  3 10                      Values of NB (blocksize)
 2 2  2  2  2                      Values of NBMIN (minimum blocksize)
 1 0  5  9  1                      Values of NX (crossover point)
-60.0                              Threshold value
+160.0                             Threshold value
 T                                 Put T to test the LAPACK routines
 T                                 Put T to test the driver routines
 T                                 Put T to test the error exits
diff --git a/param.h b/param.h
index 28ed91e60..bce05c957 100644
--- a/param.h
+++ b/param.h
@@ -1129,7 +1129,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define DGEMM_DEFAULT_UNROLL_M 8
 #define QGEMM_DEFAULT_UNROLL_M 2
 #define CGEMM_DEFAULT_UNROLL_M 8
-#define ZGEMM_DEFAULT_UNROLL_M 4
+#define ZGEMM_DEFAULT_UNROLL_M 1
 #define XGEMM_DEFAULT_UNROLL_M 1
 
 #define SGEMM_DEFAULT_UNROLL_N 4