From fdf291be30dfe3312cc7e7039a770a9e30f4c515 Mon Sep 17 00:00:00 2001
From: Werner Saar <wernsaar@googlemail.com>
Date: Thu, 10 Mar 2016 09:42:07 +0100
Subject: [PATCH] Added optimized cgemv_n and cgemv_t kernels for bulldozer,
 piledriver and steamroller

---
 kernel/x86_64/cgemv_n_microk_bulldozer-4.c | 541 +++++++++++++++++++++
 kernel/x86_64/cgemv_t_microk_bulldozer-4.c | 541 +++++++++++++++++++++
 2 files changed, 1082 insertions(+)
 create mode 100644 kernel/x86_64/cgemv_n_microk_bulldozer-4.c
 create mode 100644 kernel/x86_64/cgemv_t_microk_bulldozer-4.c

diff --git a/kernel/x86_64/cgemv_n_microk_bulldozer-4.c b/kernel/x86_64/cgemv_n_microk_bulldozer-4.c
new file mode 100644
index 000000000..a74b41269
--- /dev/null
+++ b/kernel/x86_64/cgemv_n_microk_bulldozer-4.c
@@ -0,0 +1,541 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_4x4 1
+static void cgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void cgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+	BLASLONG register n1 = n & -8 ;
+	BLASLONG register n2 = n &  4 ;
+
+	__asm__  __volatile__
+	(
+
+	"vbroadcastss	  (%2), %%ymm0                  \n\t"  // real part x0
+	"vbroadcastss	 4(%2), %%ymm1                  \n\t"  // imag part x0
+	"vbroadcastss	 8(%2), %%ymm2                  \n\t"  // real part x1
+	"vbroadcastss	12(%2), %%ymm3                  \n\t"  // imag part x1
+	"vbroadcastss	16(%2), %%ymm4                  \n\t"  // real part x2
+	"vbroadcastss	20(%2), %%ymm5                  \n\t"  // imag part x2
+	"vbroadcastss	24(%2), %%ymm6                  \n\t"  // real part x3
+	"vbroadcastss	28(%2), %%ymm7                  \n\t"  // imag part x3
+
+	"cmpq		$0 , %1				\n\t"
+	"je		2f			        \n\t"
+
+	".align 16				        \n\t"
+	"1:				        \n\t"
+	"prefetcht0      384(%4,%0,4)			\n\t"
+	"vmovups	(%4,%0,4), %%ymm8	        \n\t" // 4 complex values form a0
+	"vmovups      32(%4,%0,4), %%ymm9	        \n\t" // 4 complex values form a0
+
+	"prefetcht0      384(%5,%0,4)			\n\t"
+	"vmovups	(%5,%0,4), %%ymm10              \n\t" // 4 complex values form a1
+	"vmovups      32(%5,%0,4), %%ymm11              \n\t" // 4 complex values form a1
+
+	"vmulps      %%ymm8 , %%ymm0, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vmulps      %%ymm8 , %%ymm1, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vmulps      %%ymm9 , %%ymm0, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vmulps      %%ymm9 , %%ymm1, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"prefetcht0      384(%6,%0,4)			\n\t"
+	"vmovups	(%6,%0,4), %%ymm8	        \n\t" // 4 complex values form a2
+	"vmovups      32(%6,%0,4), %%ymm9	        \n\t" // 4 complex values form a2
+
+	"vfmaddps      %%ymm12, %%ymm10, %%ymm2, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmaddps      %%ymm13, %%ymm10, %%ymm3, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vfmaddps      %%ymm14, %%ymm11, %%ymm2, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vfmaddps      %%ymm15, %%ymm11, %%ymm3, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"prefetcht0      384(%7,%0,4)			\n\t"
+	"vmovups	(%7,%0,4), %%ymm10              \n\t" // 4 complex values form a3
+	"vmovups      32(%7,%0,4), %%ymm11              \n\t" // 4 complex values form a3
+
+	"vfmaddps      %%ymm12, %%ymm8 , %%ymm4, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmaddps      %%ymm13, %%ymm8 , %%ymm5, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vfmaddps      %%ymm14, %%ymm9 , %%ymm4, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vfmaddps      %%ymm15, %%ymm9 , %%ymm5, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"vfmaddps      %%ymm12, %%ymm10, %%ymm6, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmaddps      %%ymm13, %%ymm10, %%ymm7, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vfmaddps      %%ymm14, %%ymm11, %%ymm6, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vfmaddps      %%ymm15, %%ymm11, %%ymm7, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"prefetcht0      384(%3,%0,4)			\n\t"
+	"vmovups	  (%3,%0,4),  %%ymm10           \n\t"
+	"vmovups	32(%3,%0,4),  %%ymm11           \n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilps      $0xb1 , %%ymm13, %%ymm13               \n\t"
+        "vpermilps      $0xb1 , %%ymm15, %%ymm15               \n\t"
+        "vaddsubps      %%ymm13, %%ymm12, %%ymm8              \n\t"
+        "vaddsubps      %%ymm15, %%ymm14, %%ymm9              \n\t"
+#else
+        "vpermilps      $0xb1 , %%ymm12, %%ymm12               \n\t"
+        "vpermilps      $0xb1 , %%ymm14, %%ymm14               \n\t"
+        "vaddsubps      %%ymm12, %%ymm13, %%ymm8              \n\t"
+        "vaddsubps      %%ymm14, %%ymm15, %%ymm9              \n\t"
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilps      $0xb1 , %%ymm9 , %%ymm9                \n\t"
+#endif
+
+        "vaddps         %%ymm8, %%ymm10, %%ymm12              \n\t"
+        "vaddps         %%ymm9, %%ymm11, %%ymm13              \n\t"
+
+	"vmovups  %%ymm12,   (%3,%0,4)		        \n\t" // 4 complex values to y	
+	"vmovups  %%ymm13, 32(%3,%0,4)		        \n\t"	
+
+        "addq		$16, %0	  	 	        \n\t"
+	"subq	        $8 , %1			        \n\t"		
+	"jnz		1b		        \n\t"
+
+	"2:				        \n\t"
+
+	"cmpq		$4, %8				\n\t"
+	"jne		3f				\n\t"
+
+	"vmovups	(%4,%0,4), %%ymm8	        \n\t" // 4 complex values form a0
+	"vmovups	(%5,%0,4), %%ymm10              \n\t" // 4 complex values form a1
+
+	"vmulps      %%ymm8 , %%ymm0, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vmulps      %%ymm8 , %%ymm1, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+
+	"vfmaddps      %%ymm12, %%ymm10, %%ymm2, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmaddps      %%ymm13, %%ymm10, %%ymm3, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+
+	"vmovups	(%6,%0,4), %%ymm8	        \n\t" // 4 complex values form a2
+	"vmovups	(%7,%0,4), %%ymm10              \n\t" // 4 complex values form a3
+
+	"vfmaddps      %%ymm12, %%ymm8 , %%ymm4, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmaddps      %%ymm13, %%ymm8 , %%ymm5, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+
+	"vfmaddps      %%ymm12, %%ymm10, %%ymm6, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmaddps      %%ymm13, %%ymm10, %%ymm7, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+
+	"vmovups	  (%3,%0,4),  %%ymm10           \n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilps      $0xb1 , %%ymm13, %%ymm13               \n\t"
+        "vaddsubps      %%ymm13, %%ymm12, %%ymm8              \n\t"
+#else
+        "vpermilps      $0xb1 , %%ymm12, %%ymm12               \n\t"
+        "vaddsubps      %%ymm12, %%ymm13, %%ymm8              \n\t"
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+#endif
+
+        "vaddps         %%ymm8, %%ymm10, %%ymm12              \n\t"
+
+	"vmovups  %%ymm12,   (%3,%0,4)		        \n\t" // 4 complex values to y	
+
+	"3:				        \n\t"
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n1),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3]),  // 7
+          "r" (n2)      // 8
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
+#define HAVE_KERNEL_4x2 1
+static void cgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void cgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+	BLASLONG register n1 = n & -8 ;
+	BLASLONG register n2 = n &  4 ;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+
+	"vbroadcastss	  (%2), %%ymm0                  \n\t"  // real part x0
+	"vbroadcastss	 4(%2), %%ymm1                  \n\t"  // imag part x0
+	"vbroadcastss	 8(%2), %%ymm2                  \n\t"  // real part x1
+	"vbroadcastss	12(%2), %%ymm3                  \n\t"  // imag part x1
+
+	"cmpq		$0 , %1				\n\t"
+	"je		2f			        \n\t"
+
+	//	".align 16				        \n\t"
+	"1:				        \n\t"
+	"prefetcht0      384(%4,%0,4)			\n\t"
+	"vmovups	(%4,%0,4), %%ymm8	        \n\t" // 4 complex values form a0
+	"vmovups      32(%4,%0,4), %%ymm9	        \n\t" // 4 complex values form a0
+
+	"prefetcht0      384(%5,%0,4)			\n\t"
+	"vmovups	(%5,%0,4), %%ymm10              \n\t" // 4 complex values form a1
+	"vmovups      32(%5,%0,4), %%ymm11              \n\t" // 4 complex values form a1
+
+	"vmulps      %%ymm8 , %%ymm0, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vmulps      %%ymm8 , %%ymm1, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vmulps      %%ymm9 , %%ymm0, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vmulps      %%ymm9 , %%ymm1, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"vfmaddps      %%ymm12, %%ymm10, %%ymm2, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmaddps      %%ymm13, %%ymm10, %%ymm3, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vfmaddps      %%ymm14, %%ymm11, %%ymm2, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vfmaddps      %%ymm15, %%ymm11, %%ymm3, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"prefetcht0      384(%3,%0,4)			\n\t"
+	"vmovups	  (%3,%0,4),  %%ymm10           \n\t"
+	"vmovups	32(%3,%0,4),  %%ymm11           \n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilps      $0xb1 , %%ymm13, %%ymm13               \n\t"
+        "vpermilps      $0xb1 , %%ymm15, %%ymm15               \n\t"
+        "vaddsubps      %%ymm13, %%ymm12, %%ymm8              \n\t"
+        "vaddsubps      %%ymm15, %%ymm14, %%ymm9              \n\t"
+#else
+        "vpermilps      $0xb1 , %%ymm12, %%ymm12               \n\t"
+        "vpermilps      $0xb1 , %%ymm14, %%ymm14               \n\t"
+        "vaddsubps      %%ymm12, %%ymm13, %%ymm8              \n\t"
+        "vaddsubps      %%ymm14, %%ymm15, %%ymm9              \n\t"
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilps      $0xb1 , %%ymm9 , %%ymm9                \n\t"
+#endif
+
+        "vaddps         %%ymm8, %%ymm10, %%ymm12              \n\t"
+        "vaddps         %%ymm9, %%ymm11, %%ymm13              \n\t"
+
+	"vmovups  %%ymm12,   (%3,%0,4)		        \n\t" // 4 complex values to y	
+	"vmovups  %%ymm13, 32(%3,%0,4)		        \n\t"	
+
+        "addq		$16, %0	  	 	        \n\t"
+	"subq	        $8 , %1			        \n\t"		
+	"jnz		1b		        \n\t"
+
+	"2:				        \n\t"
+
+	"cmpq		$4, %6				\n\t"
+	"jne		3f				\n\t"
+
+	"vmovups	(%4,%0,4), %%ymm8	        \n\t" // 4 complex values form a0
+	"vmovups	(%5,%0,4), %%ymm10              \n\t" // 4 complex values form a1
+
+	"vmulps      %%ymm8 , %%ymm0, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vmulps      %%ymm8 , %%ymm1, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+
+	"vfmaddps      %%ymm12, %%ymm10, %%ymm2, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmaddps      %%ymm13, %%ymm10, %%ymm3, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+
+	"vmovups	  (%3,%0,4),  %%ymm10           \n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilps      $0xb1 , %%ymm13, %%ymm13               \n\t"
+        "vaddsubps      %%ymm13, %%ymm12, %%ymm8              \n\t"
+#else
+        "vpermilps      $0xb1 , %%ymm12, %%ymm12               \n\t"
+        "vaddsubps      %%ymm12, %%ymm13, %%ymm8              \n\t"
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+#endif
+
+        "vaddps         %%ymm8, %%ymm10, %%ymm12              \n\t"
+
+	"vmovups  %%ymm12,   (%3,%0,4)		        \n\t" // 4 complex values to y	
+
+	"3:				        \n\t"
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n1),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (n2)      // 6
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
+#define HAVE_KERNEL_4x1 1
+static void cgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void cgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+	BLASLONG register n1 = n & -8 ;
+	BLASLONG register n2 = n &  4 ;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+
+	"vbroadcastss	  (%2), %%ymm0                  \n\t"  // real part x0
+	"vbroadcastss	 4(%2), %%ymm1                  \n\t"  // imag part x0
+
+	"cmpq		$0 , %1				\n\t"
+	"je		2f			        \n\t"
+
+	//	".align 16				        \n\t"
+	"1:				        \n\t"
+	"prefetcht0      384(%4,%0,4)			\n\t"
+	"vmovups	(%4,%0,4), %%ymm8	        \n\t" // 4 complex values form a0
+	"vmovups      32(%4,%0,4), %%ymm9	        \n\t" // 4 complex values form a0
+
+	"vmulps      %%ymm8 , %%ymm0, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vmulps      %%ymm8 , %%ymm1, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vmulps      %%ymm9 , %%ymm0, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vmulps      %%ymm9 , %%ymm1, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"prefetcht0      384(%3,%0,4)			\n\t"
+	"vmovups	  (%3,%0,4),  %%ymm10           \n\t"
+	"vmovups	32(%3,%0,4),  %%ymm11           \n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilps      $0xb1 , %%ymm13, %%ymm13               \n\t"
+        "vpermilps      $0xb1 , %%ymm15, %%ymm15               \n\t"
+        "vaddsubps      %%ymm13, %%ymm12, %%ymm8              \n\t"
+        "vaddsubps      %%ymm15, %%ymm14, %%ymm9              \n\t"
+#else
+        "vpermilps      $0xb1 , %%ymm12, %%ymm12               \n\t"
+        "vpermilps      $0xb1 , %%ymm14, %%ymm14               \n\t"
+        "vaddsubps      %%ymm12, %%ymm13, %%ymm8              \n\t"
+        "vaddsubps      %%ymm14, %%ymm15, %%ymm9              \n\t"
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilps      $0xb1 , %%ymm9 , %%ymm9                \n\t"
+#endif
+
+        "addq		$16, %0	  	 	        \n\t"
+        "vaddps         %%ymm8, %%ymm10, %%ymm12              \n\t"
+        "vaddps         %%ymm9, %%ymm11, %%ymm13              \n\t"
+
+	"subq	        $8 , %1			        \n\t"		
+	"vmovups  %%ymm12,-64(%3,%0,4)		        \n\t" // 4 complex values to y	
+	"vmovups  %%ymm13,-32(%3,%0,4)		        \n\t"	
+
+	"jnz		1b		        \n\t"
+
+	"2:				        \n\t"
+
+	"cmpq		$4, %5				\n\t"
+	"jne		3f				\n\t"
+
+	"vmovups	(%4,%0,4), %%ymm8	        \n\t" // 4 complex values form a0
+
+	"vmulps      %%ymm8 , %%ymm0, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vmulps      %%ymm8 , %%ymm1, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+
+	"vmovups	  (%3,%0,4),  %%ymm10           \n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilps      $0xb1 , %%ymm13, %%ymm13               \n\t"
+        "vaddsubps      %%ymm13, %%ymm12, %%ymm8              \n\t"
+#else
+        "vpermilps      $0xb1 , %%ymm12, %%ymm12               \n\t"
+        "vaddsubps      %%ymm12, %%ymm13, %%ymm8              \n\t"
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+#endif
+
+        "vaddps         %%ymm8, %%ymm10, %%ymm12              \n\t"
+
+	"vmovups  %%ymm12,   (%3,%0,4)		        \n\t" // 4 complex values to y	
+
+	"3:				        \n\t"
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n1),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap),     // 4
+          "r" (n2)      // 5
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
+#define HAVE_KERNEL_ADDY 1
+
+static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i)  __attribute__ ((noinline));
+
+static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i)
+{
+	BLASLONG i;
+
+	if ( inc_dest != 2 )
+	{
+
+		FLOAT temp_r;
+		FLOAT temp_i;
+		for ( i=0; i<n; i++ )
+		{
+#if !defined(XCONJ) 
+			temp_r = alpha_r * src[0] - alpha_i * src[1];
+			temp_i = alpha_r * src[1] + alpha_i * src[0];
+#else
+			temp_r =  alpha_r * src[0] + alpha_i * src[1];
+			temp_i = -alpha_r * src[1] + alpha_i * src[0];
+#endif
+
+			*dest += temp_r;
+			*(dest+1) += temp_i;
+
+			src+=2;
+			dest += inc_dest;
+		}
+		return;
+	}
+
+	i=0;
+	BLASLONG register n1 = n & -8 ;
+	BLASLONG register n2 = n &  4 ;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+
+	"vbroadcastss	  (%4), %%ymm0                  \n\t"  // alpha_r
+	"vbroadcastss	  (%5), %%ymm1                  \n\t"  // alpha_i
+
+	"cmpq		$0 , %1				\n\t"
+	"je		2f			        \n\t"
+
+	//	".align 16				        \n\t"
+	"1:				        \n\t"
+	"vmovups	(%2,%0,4), %%ymm8	        \n\t" // 4 complex values from src
+	"vmovups      32(%2,%0,4), %%ymm9	        \n\t" 
+
+	"vmulps      %%ymm8 , %%ymm0, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vmulps      %%ymm8 , %%ymm1, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vmulps      %%ymm9 , %%ymm0, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vmulps      %%ymm9 , %%ymm1, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"vmovups	  (%3,%0,4),  %%ymm10           \n\t" // 4 complex values from dest
+	"vmovups	32(%3,%0,4),  %%ymm11           \n\t"
+
+#if  !defined(XCONJ)  
+        "vpermilps      $0xb1 , %%ymm13, %%ymm13               \n\t"
+        "vpermilps      $0xb1 , %%ymm15, %%ymm15               \n\t"
+        "vaddsubps      %%ymm13, %%ymm12, %%ymm8              \n\t"
+        "vaddsubps      %%ymm15, %%ymm14, %%ymm9              \n\t"
+#else
+        "vpermilps      $0xb1 , %%ymm12, %%ymm12               \n\t"
+        "vpermilps      $0xb1 , %%ymm14, %%ymm14               \n\t"
+        "vaddsubps      %%ymm12, %%ymm13, %%ymm8              \n\t"
+        "vaddsubps      %%ymm14, %%ymm15, %%ymm9              \n\t"
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilps      $0xb1 , %%ymm9 , %%ymm9                \n\t"
+#endif
+
+        "addq		$16, %0	  	 	        \n\t"
+        "vaddps         %%ymm8, %%ymm10, %%ymm12              \n\t"
+        "vaddps         %%ymm9, %%ymm11, %%ymm13              \n\t"
+
+	"subq	        $8 , %1			        \n\t"		
+	"vmovups  %%ymm12,-64(%3,%0,4)		        \n\t" // 4 complex values to y	
+	"vmovups  %%ymm13,-32(%3,%0,4)		        \n\t"	
+
+	"jnz		1b		        \n\t"
+
+	"2:				        \n\t"
+
+	"cmpq		$4, %6				\n\t"
+	"jne		3f				\n\t"
+
+	"vmovups	(%2,%0,4), %%ymm8	        \n\t" // 4 complex values src
+
+	"vmulps      %%ymm8 , %%ymm0, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vmulps      %%ymm8 , %%ymm1, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+
+	"vmovups	  (%3,%0,4),  %%ymm10           \n\t"
+
+#if !defined(XCONJ)
+        "vpermilps      $0xb1 , %%ymm13, %%ymm13               \n\t"
+        "vaddsubps      %%ymm13, %%ymm12, %%ymm8              \n\t"
+#else
+        "vpermilps      $0xb1 , %%ymm12, %%ymm12               \n\t"
+        "vaddsubps      %%ymm12, %%ymm13, %%ymm8              \n\t"
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+#endif
+
+        "vaddps         %%ymm8, %%ymm10, %%ymm12              \n\t"
+
+	"vmovups  %%ymm12,   (%3,%0,4)		        \n\t" // 4 complex values to y	
+
+	"3:				        \n\t"
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	      // 0	
+	  "r" (n1),  	      // 1
+          "r" (src),          // 2
+          "r" (dest),         // 3
+          "r" (&alpha_r),     // 4
+          "r" (&alpha_i),     // 5
+          "r" (n2)            // 6
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+	return;
+
+}
+
diff --git a/kernel/x86_64/cgemv_t_microk_bulldozer-4.c b/kernel/x86_64/cgemv_t_microk_bulldozer-4.c
new file mode 100644
index 000000000..941b9cfc7
--- /dev/null
+++ b/kernel/x86_64/cgemv_t_microk_bulldozer-4.c
@@ -0,0 +1,541 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary froms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary from must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_4x4 1
+static void cgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha) __attribute__ ((noinline));
+
+static void cgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+
+	"vxorps		%%ymm8 , %%ymm8 , %%ymm8 	\n\t" // temp
+	"vxorps		%%ymm9 , %%ymm9 , %%ymm9 	\n\t" // temp
+	"vxorps		%%ymm10, %%ymm10, %%ymm10	\n\t" // temp
+	"vxorps		%%ymm11, %%ymm11, %%ymm11	\n\t" // temp
+	"vxorps		%%ymm12, %%ymm12, %%ymm12	\n\t" // temp
+	"vxorps		%%ymm13, %%ymm13, %%ymm13	\n\t"
+	"vxorps		%%ymm14, %%ymm14, %%ymm14	\n\t"
+	"vxorps		%%ymm15, %%ymm15, %%ymm15	\n\t"
+
+        "testq          $0x04, %1                      \n\t"
+        "jz             2f                      \n\t"
+
+	"vmovups	(%4,%0,4), %%ymm4	        \n\t" // 4 complex values from a0
+	"vmovups	(%5,%0,4), %%ymm5               \n\t" // 4 complex values from a1
+
+	"vmovups	    (%2,%0,4)  , %%ymm6		\n\t" // 4 complex values from x
+	"vpermilps        $0xb1, %%ymm6, %%ymm7		\n\t" // exchange real and imap parts
+	"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0         \n\t" // only the real parts
+	"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1         \n\t" // only the imag parts
+	
+	"vmovups	(%6,%0,4), %%ymm6	        \n\t" // 4 complex values from a2
+	"vmovups	(%7,%0,4), %%ymm7               \n\t" // 4 complex values from a3
+
+	"vfmaddps      %%ymm8 , %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddps      %%ymm9 , %%ymm4 , %%ymm1, %%ymm9       \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmaddps      %%ymm10, %%ymm5 , %%ymm0, %%ymm10      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddps      %%ymm11, %%ymm5 , %%ymm1, %%ymm11      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmaddps      %%ymm12, %%ymm6 , %%ymm0, %%ymm12      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddps      %%ymm13, %%ymm6 , %%ymm1, %%ymm13      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmaddps      %%ymm14, %%ymm7 , %%ymm0, %%ymm14      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddps      %%ymm15, %%ymm7 , %%ymm1, %%ymm15      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	
+        "addq		$8  , %0	  	 	        \n\t"
+	"subq	        $4  , %1			        \n\t"		
+
+        "2:                                  \n\t"
+	"cmpq           $0, %1                         \n\t"
+        "je             3f                      \n\t"
+
+	//	".align 16				        \n\t"
+	"1:				        \n\t"
+        "prefetcht0      384(%4,%0,4)                   \n\t"
+	"vmovups	(%4,%0,4), %%ymm4	        \n\t" // 4 complex values from a0
+        "prefetcht0      384(%5,%0,4)                   \n\t"
+	"vmovups	(%5,%0,4), %%ymm5               \n\t" // 4 complex values from a1
+
+        "prefetcht0      384(%2,%0,4)                   \n\t"
+	"vmovups	    (%2,%0,4)  , %%ymm6		\n\t" // 4 complex values from x
+	"vpermilps        $0xb1, %%ymm6, %%ymm7		\n\t" // exchange real and imap parts
+	"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0         \n\t" // only the real parts
+	"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1         \n\t" // only the imag parts
+	
+        "prefetcht0      384(%6,%0,4)                   \n\t"
+	"vmovups	(%6,%0,4), %%ymm6	        \n\t" // 4 complex values from a2
+        "prefetcht0      384(%7,%0,4)                   \n\t"
+	"vmovups	(%7,%0,4), %%ymm7               \n\t" // 4 complex values from a3
+
+	"vfmaddps      %%ymm8 , %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddps      %%ymm9 , %%ymm4 , %%ymm1, %%ymm9       \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmaddps      %%ymm10, %%ymm5 , %%ymm0, %%ymm10      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddps      %%ymm11, %%ymm5 , %%ymm1, %%ymm11      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmaddps      %%ymm12, %%ymm6 , %%ymm0, %%ymm12      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddps      %%ymm13, %%ymm6 , %%ymm1, %%ymm13      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmaddps      %%ymm14, %%ymm7 , %%ymm0, %%ymm14      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddps      %%ymm15, %%ymm7 , %%ymm1, %%ymm15      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	
+
+	"vmovups       32(%4,%0,4), %%ymm4	        \n\t" // 4 complex values from a0
+	"vmovups       32(%5,%0,4), %%ymm5              \n\t" // 4 complex values from a1
+
+	"vmovups	  32(%2,%0,4)  , %%ymm6		\n\t" // 4 complex values from x
+	"vpermilps        $0xb1, %%ymm6, %%ymm7		\n\t" // exchange real and imap parts
+	"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0         \n\t" // only the real parts
+	"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1         \n\t" // only the imag parts
+
+	"vmovups       32(%6,%0,4), %%ymm6	        \n\t" // 4 complex values from a2
+	"vmovups       32(%7,%0,4), %%ymm7              \n\t" // 4 complex values from a3
+
+	"vfmaddps      %%ymm8 , %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddps      %%ymm9 , %%ymm4 , %%ymm1, %%ymm9       \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmaddps      %%ymm10, %%ymm5 , %%ymm0, %%ymm10      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddps      %%ymm11, %%ymm5 , %%ymm1, %%ymm11      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmaddps      %%ymm12, %%ymm6 , %%ymm0, %%ymm12      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddps      %%ymm13, %%ymm6 , %%ymm1, %%ymm13      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmaddps      %%ymm14, %%ymm7 , %%ymm0, %%ymm14      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddps      %%ymm15, %%ymm7 , %%ymm1, %%ymm15      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	
+        "addq		$16 , %0	  	 	        \n\t"
+	"subq	        $8  , %1			        \n\t"		
+	"jnz		1b		        \n\t"
+
+        "3:                                   \n\t"
+
+        "vbroadcastss    (%8)  , %%xmm0                \n\t"  // value from alpha
+        "vbroadcastss   4(%8)  , %%xmm1                \n\t"  // value from alpha
+
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilps      $0xb1 , %%ymm9 , %%ymm9                \n\t"
+        "vpermilps      $0xb1 , %%ymm11, %%ymm11               \n\t"
+        "vpermilps      $0xb1 , %%ymm13, %%ymm13               \n\t"
+        "vpermilps      $0xb1 , %%ymm15, %%ymm15               \n\t"
+        "vaddsubps      %%ymm9 , %%ymm8, %%ymm8                \n\t" 
+        "vaddsubps      %%ymm11, %%ymm10, %%ymm10              \n\t"
+        "vaddsubps      %%ymm13, %%ymm12, %%ymm12              \n\t"
+        "vaddsubps      %%ymm15, %%ymm14, %%ymm14              \n\t"
+#else
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilps      $0xb1 , %%ymm10, %%ymm10               \n\t"
+        "vpermilps      $0xb1 , %%ymm12, %%ymm12               \n\t"
+        "vpermilps      $0xb1 , %%ymm14, %%ymm14               \n\t"
+        "vaddsubps      %%ymm8 , %%ymm9 , %%ymm8               \n\t"
+        "vaddsubps      %%ymm10, %%ymm11, %%ymm10              \n\t"
+        "vaddsubps      %%ymm12, %%ymm13, %%ymm12              \n\t"
+        "vaddsubps      %%ymm14, %%ymm15, %%ymm14              \n\t"
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilps      $0xb1 , %%ymm10, %%ymm10               \n\t"
+        "vpermilps      $0xb1 , %%ymm12, %%ymm12               \n\t"
+        "vpermilps      $0xb1 , %%ymm14, %%ymm14               \n\t"
+#endif
+
+	"vmovsd         (%3), %%xmm4			\n\t" // read y
+	"vmovsd        8(%3), %%xmm5			\n\t"
+	"vmovsd       16(%3), %%xmm6			\n\t"
+	"vmovsd       24(%3), %%xmm7			\n\t"
+
+	"vextractf128   $1, %%ymm8 , %%xmm9		      \n\t"
+	"vextractf128   $1, %%ymm10, %%xmm11	      	      \n\t"
+	"vextractf128   $1, %%ymm12, %%xmm13		      \n\t"
+	"vextractf128   $1, %%ymm14, %%xmm15		      \n\t"
+
+	"vaddps		%%xmm8 , %%xmm9 , %%xmm8       \n\t"
+	"vaddps		%%xmm10, %%xmm11, %%xmm10      \n\t"
+	"vaddps		%%xmm12, %%xmm13, %%xmm12      \n\t"
+	"vaddps		%%xmm14, %%xmm15, %%xmm14      \n\t"
+
+	"vshufpd        $0x1, %%xmm8 , %%xmm8 , %%xmm9   \n\t"
+	"vshufpd        $0x1, %%xmm10, %%xmm10, %%xmm11  \n\t"
+	"vshufpd        $0x1, %%xmm12, %%xmm12, %%xmm13  \n\t"
+	"vshufpd        $0x1, %%xmm14, %%xmm14, %%xmm15  \n\t"
+
+	"vaddps		%%xmm8 , %%xmm9 , %%xmm8       \n\t"
+	"vaddps		%%xmm10, %%xmm11, %%xmm10      \n\t"
+	"vaddps		%%xmm12, %%xmm13, %%xmm12      \n\t"
+	"vaddps		%%xmm14, %%xmm15, %%xmm14      \n\t"
+
+
+        "vmulps         %%xmm8 , %%xmm1 , %%xmm9              \n\t"  // t_r * alpha_i , t_i * alpha_i
+        "vmulps         %%xmm8 , %%xmm0 , %%xmm8              \n\t"  // t_r * alpha_r , t_i * alpha_r
+        "vmulps         %%xmm10, %%xmm1 , %%xmm11             \n\t"  // t_r * alpha_i , t_i * alpha_i
+        "vmulps         %%xmm10, %%xmm0 , %%xmm10             \n\t"  // t_r * alpha_r , t_i * alpha_r
+        "vmulps         %%xmm12, %%xmm1 , %%xmm13             \n\t"  // t_r * alpha_i , t_i * alpha_i
+        "vmulps         %%xmm12, %%xmm0 , %%xmm12             \n\t"  // t_r * alpha_r , t_i * alpha_r
+        "vmulps         %%xmm14, %%xmm1 , %%xmm15             \n\t"  // t_r * alpha_i , t_i * alpha_i
+        "vmulps         %%xmm14, %%xmm0 , %%xmm14             \n\t"  // t_r * alpha_r , t_i * alpha_r
+
+#if !defined(XCONJ)
+        "vpermilps      $0xb1 , %%xmm9 , %%xmm9                \n\t"
+        "vpermilps      $0xb1 , %%xmm11, %%xmm11               \n\t"
+        "vpermilps      $0xb1 , %%xmm13, %%xmm13               \n\t"
+        "vpermilps      $0xb1 , %%xmm15, %%xmm15               \n\t"
+        "vaddsubps      %%xmm9 , %%xmm8, %%xmm8               \n\t"
+        "vaddsubps      %%xmm11, %%xmm10, %%xmm10             \n\t"
+        "vaddsubps      %%xmm13, %%xmm12, %%xmm12             \n\t"
+        "vaddsubps      %%xmm15, %%xmm14, %%xmm14             \n\t"
+#else
+        "vpermilps      $0xb1 , %%xmm8 , %%xmm8                \n\t"
+        "vpermilps      $0xb1 , %%xmm10, %%xmm10               \n\t"
+        "vpermilps      $0xb1 , %%xmm12, %%xmm12               \n\t"
+        "vpermilps      $0xb1 , %%xmm14, %%xmm14               \n\t"
+        "vaddsubps      %%xmm8 , %%xmm9 , %%xmm8              \n\t"
+        "vaddsubps      %%xmm10, %%xmm11, %%xmm10             \n\t"
+        "vaddsubps      %%xmm12, %%xmm13, %%xmm12             \n\t"
+        "vaddsubps      %%xmm14, %%xmm15, %%xmm14             \n\t"
+        "vpermilps      $0xb1 , %%xmm8 , %%xmm8                \n\t"
+        "vpermilps      $0xb1 , %%xmm10, %%xmm10               \n\t"
+        "vpermilps      $0xb1 , %%xmm12, %%xmm12               \n\t"
+        "vpermilps      $0xb1 , %%xmm14, %%xmm14               \n\t"
+#endif
+
+
+	"vaddps		%%xmm8 , %%xmm4 , %%xmm8       \n\t"
+	"vaddps		%%xmm10, %%xmm5 , %%xmm10      \n\t"
+	"vaddps		%%xmm12, %%xmm6 , %%xmm12      \n\t"
+	"vaddps		%%xmm14, %%xmm7 , %%xmm14      \n\t"
+
+	"vmovsd	%%xmm8 ,   (%3)			\n\t"
+	"vmovsd	%%xmm10,  8(%3)			\n\t"
+	"vmovsd	%%xmm12, 16(%3)			\n\t"
+	"vmovsd	%%xmm14, 24(%3)			\n\t"
+
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3]),  // 7
+          "r" (alpha)   // 8
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
+#define HAVE_KERNEL_4x2 1
+static void cgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha) __attribute__ ((noinline));
+
+static void cgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+
+	"vxorps		%%ymm8 , %%ymm8 , %%ymm8 	\n\t" // temp
+	"vxorps		%%ymm9 , %%ymm9 , %%ymm9 	\n\t" // temp
+	"vxorps		%%ymm10, %%ymm10, %%ymm10	\n\t" // temp
+	"vxorps		%%ymm11, %%ymm11, %%ymm11	\n\t" // temp
+
+        "testq          $0x04, %1                      \n\t"
+        "jz             2f                    \n\t"
+
+	"vmovups	(%4,%0,4), %%ymm4	        \n\t" // 4 complex values from a0
+	"vmovups	(%5,%0,4), %%ymm5               \n\t" // 4 complex values from a1
+
+	"vmovups	    (%2,%0,4)  , %%ymm6		\n\t" // 4 complex values from x
+	"vpermilps        $0xb1, %%ymm6, %%ymm7		\n\t" // exchange real and imap parts
+	"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0         \n\t" // only the real parts
+	"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1         \n\t" // only the imag parts
+	
+
+	"vfmaddps      %%ymm8 , %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddps      %%ymm9 , %%ymm4 , %%ymm1, %%ymm9       \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmaddps      %%ymm10, %%ymm5 , %%ymm0, %%ymm10      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddps      %%ymm11, %%ymm5 , %%ymm1, %%ymm11      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	
+        "addq		$8  , %0	  	 	        \n\t"
+	"subq	        $4  , %1			        \n\t"		
+
+        "2:                                  \n\t"
+	"cmpq           $0, %1                         \n\t"
+        "je             3f                      \n\t"
+
+	//	".align 16				        \n\t"
+	"1:				        \n\t"
+        "prefetcht0      384(%4,%0,4)                   \n\t"
+	"vmovups	(%4,%0,4), %%ymm4	        \n\t" // 4 complex values from a0
+        "prefetcht0      384(%5,%0,4)                   \n\t"
+	"vmovups	(%5,%0,4), %%ymm5               \n\t" // 4 complex values from a1
+
+        "prefetcht0      384(%2,%0,4)                   \n\t"
+	"vmovups	    (%2,%0,4)  , %%ymm6		\n\t" // 4 complex values from x
+	"vpermilps        $0xb1, %%ymm6, %%ymm7		\n\t" // exchange real and imap parts
+	"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0         \n\t" // only the real parts
+	"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1         \n\t" // only the imag parts
+	
+	"vfmaddps      %%ymm8 , %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddps      %%ymm9 , %%ymm4 , %%ymm1, %%ymm9       \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmaddps      %%ymm10, %%ymm5 , %%ymm0, %%ymm10      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddps      %%ymm11, %%ymm5 , %%ymm1, %%ymm11      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+
+	"vmovups       32(%4,%0,4), %%ymm4	        \n\t" // 4 complex values from a0
+	"vmovups       32(%5,%0,4), %%ymm5              \n\t" // 4 complex values from a1
+
+	"vmovups	  32(%2,%0,4)  , %%ymm6		\n\t" // 4 complex values from x
+	"vpermilps        $0xb1, %%ymm6, %%ymm7		\n\t" // exchange real and imap parts
+	"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0         \n\t" // only the real parts
+	"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1         \n\t" // only the imag parts
+
+	"vfmaddps      %%ymm8 , %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddps      %%ymm9 , %%ymm4 , %%ymm1, %%ymm9       \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmaddps      %%ymm10, %%ymm5 , %%ymm0, %%ymm10      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddps      %%ymm11, %%ymm5 , %%ymm1, %%ymm11      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	
+        "addq		$16 , %0	  	 	        \n\t"
+	"subq	        $8  , %1			        \n\t"		
+	"jnz		1b		        \n\t"
+
+        "3:                                   \n\t"
+
+        "vbroadcastss    (%6)  , %%xmm0                \n\t"  // value from alpha
+        "vbroadcastss   4(%6)  , %%xmm1                \n\t"  // value from alpha
+
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilps      $0xb1 , %%ymm9 , %%ymm9                \n\t"
+        "vpermilps      $0xb1 , %%ymm11, %%ymm11               \n\t"
+        "vaddsubps      %%ymm9 , %%ymm8, %%ymm8                \n\t" 
+        "vaddsubps      %%ymm11, %%ymm10, %%ymm10              \n\t"
+#else
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilps      $0xb1 , %%ymm10, %%ymm10               \n\t"
+        "vaddsubps      %%ymm8 , %%ymm9 , %%ymm8               \n\t"
+        "vaddsubps      %%ymm10, %%ymm11, %%ymm10              \n\t"
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilps      $0xb1 , %%ymm10, %%ymm10               \n\t"
+#endif
+
+	"vmovsd         (%3), %%xmm4			\n\t" // read y
+	"vmovsd        8(%3), %%xmm5			\n\t"
+
+	"vextractf128   $1, %%ymm8 , %%xmm9		      \n\t"
+	"vextractf128   $1, %%ymm10, %%xmm11	      	      \n\t"
+
+	"vaddps		%%xmm8 , %%xmm9 , %%xmm8       \n\t"
+	"vaddps		%%xmm10, %%xmm11, %%xmm10      \n\t"
+
+	"vshufpd        $0x1, %%xmm8 , %%xmm8 , %%xmm9   \n\t"
+	"vshufpd        $0x1, %%xmm10, %%xmm10, %%xmm11  \n\t"
+
+	"vaddps		%%xmm8 , %%xmm9 , %%xmm8       \n\t"
+	"vaddps		%%xmm10, %%xmm11, %%xmm10      \n\t"
+
+        "vmulps         %%xmm8 , %%xmm1 , %%xmm9              \n\t"  // t_r * alpha_i , t_i * alpha_i
+        "vmulps         %%xmm8 , %%xmm0 , %%xmm8              \n\t"  // t_r * alpha_r , t_i * alpha_r
+        "vmulps         %%xmm10, %%xmm1 , %%xmm11             \n\t"  // t_r * alpha_i , t_i * alpha_i
+        "vmulps         %%xmm10, %%xmm0 , %%xmm10             \n\t"  // t_r * alpha_r , t_i * alpha_r
+
+#if !defined(XCONJ)
+        "vpermilps      $0xb1 , %%xmm9 , %%xmm9                \n\t"
+        "vpermilps      $0xb1 , %%xmm11, %%xmm11               \n\t"
+        "vaddsubps      %%xmm9 , %%xmm8, %%xmm8               \n\t"
+        "vaddsubps      %%xmm11, %%xmm10, %%xmm10             \n\t"
+#else
+        "vpermilps      $0xb1 , %%xmm8 , %%xmm8                \n\t"
+        "vpermilps      $0xb1 , %%xmm10, %%xmm10               \n\t"
+        "vaddsubps      %%xmm8 , %%xmm9 , %%xmm8              \n\t"
+        "vaddsubps      %%xmm10, %%xmm11, %%xmm10             \n\t"
+        "vpermilps      $0xb1 , %%xmm8 , %%xmm8                \n\t"
+        "vpermilps      $0xb1 , %%xmm10, %%xmm10               \n\t"
+#endif
+
+
+	"vaddps		%%xmm8 , %%xmm4 , %%xmm8       \n\t"
+	"vaddps		%%xmm10, %%xmm5 , %%xmm10      \n\t"
+
+	"vmovsd	%%xmm8 ,   (%3)			\n\t"
+	"vmovsd	%%xmm10,  8(%3)			\n\t"
+
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (alpha)   // 6
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
+#define HAVE_KERNEL_4x1 1
+static void cgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT *alpha) __attribute__ ((noinline));
+
+static void cgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+
+	"vxorps		%%ymm8 , %%ymm8 , %%ymm8 	\n\t" // temp
+	"vxorps		%%ymm9 , %%ymm9 , %%ymm9 	\n\t" // temp
+
+        "testq          $0x04, %1                      \n\t"
+        "jz             2f                    \n\t"
+
+	"vmovups	(%4,%0,4), %%ymm4	        \n\t" // 4 complex values from a0
+
+	"vmovups	    (%2,%0,4)  , %%ymm6		\n\t" // 4 complex values from x
+	"vpermilps        $0xb1, %%ymm6, %%ymm7		\n\t" // exchange real and imap parts
+	"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0         \n\t" // only the real parts
+	"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1         \n\t" // only the imag parts
+	
+
+	"vfmaddps      %%ymm8 , %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddps      %%ymm9 , %%ymm4 , %%ymm1, %%ymm9       \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	
+        "addq		$8  , %0	  	 	        \n\t"
+	"subq	        $4  , %1			        \n\t"		
+
+        "2:                                  \n\t"
+	"cmpq           $0, %1                         \n\t"
+        "je             3f                      \n\t"
+
+	//	".align 16				        \n\t"
+	"1:				        \n\t"
+        "prefetcht0      384(%4,%0,4)                   \n\t"
+	"vmovups	(%4,%0,4), %%ymm4	        \n\t" // 4 complex values from a0
+
+        "prefetcht0      384(%2,%0,4)                   \n\t"
+	"vmovups	    (%2,%0,4)  , %%ymm6		\n\t" // 4 complex values from x
+	"vpermilps        $0xb1, %%ymm6, %%ymm7		\n\t" // exchange real and imap parts
+	"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0         \n\t" // only the real parts
+	"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1         \n\t" // only the imag parts
+	
+	"vfmaddps      %%ymm8 , %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddps      %%ymm9 , %%ymm4 , %%ymm1, %%ymm9       \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	
+	"vmovups       32(%4,%0,4), %%ymm4	        \n\t" // 4 complex values from a0
+
+	"vmovups	  32(%2,%0,4)  , %%ymm6		\n\t" // 4 complex values from x
+	"vpermilps        $0xb1, %%ymm6, %%ymm7		\n\t" // exchange real and imap parts
+	"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0         \n\t" // only the real parts
+	"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1         \n\t" // only the imag parts
+
+	"vfmaddps      %%ymm8 , %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmaddps      %%ymm9 , %%ymm4 , %%ymm1, %%ymm9       \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	
+
+        "addq		$16 , %0	  	 	        \n\t"
+	"subq	        $8  , %1			        \n\t"		
+	"jnz		1b		        \n\t"
+
+        "3:                                   \n\t"
+
+        "vbroadcastss    (%5)  , %%xmm0                \n\t"  // value from alpha
+        "vbroadcastss   4(%5)  , %%xmm1                \n\t"  // value from alpha
+
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilps      $0xb1 , %%ymm9 , %%ymm9                \n\t"
+        "vaddsubps      %%ymm9 , %%ymm8, %%ymm8                \n\t" 
+#else
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+        "vaddsubps      %%ymm8 , %%ymm9 , %%ymm8               \n\t"
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+#endif
+
+	"vmovsd         (%3), %%xmm4			\n\t" // read y
+
+	"vextractf128   $1, %%ymm8 , %%xmm9		      \n\t"
+
+	"vaddps		%%xmm8 , %%xmm9 , %%xmm8       \n\t"
+
+	"vshufpd        $0x1, %%xmm8 , %%xmm8 , %%xmm9   \n\t"
+
+	"vaddps		%%xmm8 , %%xmm9 , %%xmm8       \n\t"
+
+        "vmulps         %%xmm8 , %%xmm1 , %%xmm9              \n\t"  // t_r * alpha_i , t_i * alpha_i
+        "vmulps         %%xmm8 , %%xmm0 , %%xmm8              \n\t"  // t_r * alpha_r , t_i * alpha_r
+
+#if !defined(XCONJ)
+        "vpermilps      $0xb1 , %%xmm9 , %%xmm9                \n\t"
+        "vaddsubps      %%xmm9 , %%xmm8, %%xmm8               \n\t"
+#else
+        "vpermilps      $0xb1 , %%xmm8 , %%xmm8                \n\t"
+        "vaddsubps      %%xmm8 , %%xmm9 , %%xmm8              \n\t"
+        "vpermilps      $0xb1 , %%xmm8 , %%xmm8                \n\t"
+#endif
+
+
+	"vaddps		%%xmm8 , %%xmm4 , %%xmm8       \n\t"
+
+	"vmovsd	%%xmm8 ,   (%3)			\n\t"
+
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap),     // 4
+          "r" (alpha)   // 5
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+