Merge pull request #540 from wernsaar/develop

Optimized dot- and axpy-kernels

kernel/x86_64/KERNEL.PILEDRIVER

@@ -1,3 +1,4 @@
+SAXPYKERNEL = saxpy.c
 DAXPYKERNEL = daxpy.c
 CAXPYKERNEL = caxpy.c
 ZAXPYKERNEL = zaxpy.c
@@ -12,10 +13,10 @@ DGEMVNKERNEL = dgemv_n_bulldozer.S
 DGEMVTKERNEL = dgemv_t_bulldozer.S
 
 SDOTKERNEL   = sdot.c
+DDOTKERNEL   = ddot.c
 CDOTKERNEL   = cdot.c
 ZDOTKERNEL   = zdot.c
 
-DDOTKERNEL   = ddot_bulldozer.S
 DCOPYKERNEL  = dcopy_bulldozer.S
 
 SGEMMKERNEL    =  sgemm_kernel_16x2_piledriver.S

kernel/x86_64/KERNEL.STEAMROLLER

@@ -1,3 +1,4 @@
+SAXPYKERNEL = saxpy.c
 DAXPYKERNEL = daxpy.c
 CAXPYKERNEL = caxpy.c
 ZAXPYKERNEL = zaxpy.c

kernel/x86_64/daxpy.c

@@ -31,8 +31,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #if defined(NEHALEM)
 #include "daxpy_microk_nehalem-2.c"
-#elif defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER)
+#elif defined(BULLDOZER)
 #include "daxpy_microk_bulldozer-2.c"
+#elif defined(STEAMROLLER)
+#include "daxpy_microk_steamroller-2.c"
+#elif defined(PILEDRIVER)
+#include "daxpy_microk_piledriver-2.c"
 #elif defined(HASWELL)
 #include "daxpy_microk_haswell-2.c"
 #elif defined(SANDYBRIDGE)

kernel/x86_64/daxpy_microk_piledriver-2.c

@@ -0,0 +1,160 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_8 1
+static void daxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *alpha) __attribute__ ((noinline));
+
+static void daxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
+{
+
+
+	BLASLONG register i = 0;
+
+  if ( n < 640 )
+  {
+
+	__asm__  __volatile__
+	(
+	"vmovddup		(%4), %%xmm0		    		\n\t"  // alpha	
+
+	".align 16				            		\n\t"
+	"1:				            			\n\t"
+
+        "vmovups                  (%3,%0,8), %%xmm8          		\n\t"  // 2  y
+        "vmovups                16(%3,%0,8), %%xmm9         		\n\t"  // 2  y
+        "vmovups                32(%3,%0,8), %%xmm10         		\n\t"  // 2  y
+        "vmovups                48(%3,%0,8), %%xmm11         		\n\t"  // 2  y
+
+        "vmovups                64(%3,%0,8), %%xmm12         		\n\t"  // 2  y
+        "vmovups                80(%3,%0,8), %%xmm13        		\n\t"  // 2  y
+        "vmovups                96(%3,%0,8), %%xmm14         		\n\t"  // 2  y
+        "vmovups               112(%3,%0,8), %%xmm15         		\n\t"  // 2  y
+
+	"vfmadd231pd       (%2,%0,8), %%xmm0 , %%xmm8  			\n\t"   // y += alpha * x
+	"vfmadd231pd     16(%2,%0,8), %%xmm0 , %%xmm9 		  	\n\t"   // y += alpha * x
+	"vfmadd231pd     32(%2,%0,8), %%xmm0 , %%xmm10 			\n\t"   // y += alpha * x
+	"vfmadd231pd     48(%2,%0,8), %%xmm0 , %%xmm11 			\n\t"   // y += alpha * x
+
+	"vfmadd231pd     64(%2,%0,8), %%xmm0 , %%xmm12 			\n\t"   // y += alpha * x
+	"vfmadd231pd     80(%2,%0,8), %%xmm0 , %%xmm13		  	\n\t"   // y += alpha * x
+	"vfmadd231pd     96(%2,%0,8), %%xmm0 , %%xmm14 			\n\t"   // y += alpha * x
+	"vfmadd231pd    112(%2,%0,8), %%xmm0 , %%xmm15 			\n\t"   // y += alpha * x
+
+	"vmovups		%%xmm8 ,   (%3,%0,8)                 	\n\t"
+	"vmovups		%%xmm9 , 16(%3,%0,8)                 	\n\t"
+	"vmovups		%%xmm10, 32(%3,%0,8)           		\n\t"
+	"vmovups		%%xmm11, 48(%3,%0,8)                 	\n\t"
+
+	"vmovups		%%xmm12, 64(%3,%0,8)                 	\n\t"
+	"vmovups		%%xmm13, 80(%3,%0,8)                 	\n\t"
+	"vmovups		%%xmm14, 96(%3,%0,8)           		\n\t"
+	"vmovups		%%xmm15,112(%3,%0,8)                 	\n\t"
+
+	"addq		$16, %0	  	 	             		\n\t"
+	"subq	        $16, %1			             		\n\t"		
+	"jnz		1b		             			\n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (alpha)   // 4
+	: "cc", 
+	  "%xmm0", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+	return;
+  }
+
+
+	__asm__  __volatile__
+	(
+	"vmovddup		(%4), %%xmm0		    		\n\t"  // alpha	
+
+	".align 16				            		\n\t"
+	"1:				            			\n\t"
+
+	"prefetcht0	       512(%3,%0,8)				\n\t"
+        "vmovups                  (%3,%0,8), %%xmm8          		\n\t"  // 2  y
+        "vmovups                16(%3,%0,8), %%xmm9         		\n\t"  // 2  y
+        "vmovups                32(%3,%0,8), %%xmm10         		\n\t"  // 2  y
+        "vmovups                48(%3,%0,8), %%xmm11         		\n\t"  // 2  y
+
+	"prefetcht0	       576(%3,%0,8)				\n\t"
+        "vmovups                64(%3,%0,8), %%xmm12         		\n\t"  // 2  y
+        "vmovups                80(%3,%0,8), %%xmm13        		\n\t"  // 2  y
+        "vmovups                96(%3,%0,8), %%xmm14         		\n\t"  // 2  y
+        "vmovups               112(%3,%0,8), %%xmm15         		\n\t"  // 2  y
+
+	"prefetcht0	       512(%2,%0,8)				\n\t"
+	"vfmadd231pd       (%2,%0,8), %%xmm0 , %%xmm8  			\n\t"   // y += alpha * x
+	"vfmadd231pd     16(%2,%0,8), %%xmm0 , %%xmm9 		  	\n\t"   // y += alpha * x
+	"vfmadd231pd     32(%2,%0,8), %%xmm0 , %%xmm10 			\n\t"   // y += alpha * x
+	"vfmadd231pd     48(%2,%0,8), %%xmm0 , %%xmm11 			\n\t"   // y += alpha * x
+
+	"prefetcht0	       576(%2,%0,8)				\n\t"
+	"vfmadd231pd     64(%2,%0,8), %%xmm0 , %%xmm12 			\n\t"   // y += alpha * x
+	"vfmadd231pd     80(%2,%0,8), %%xmm0 , %%xmm13		  	\n\t"   // y += alpha * x
+	"vfmadd231pd     96(%2,%0,8), %%xmm0 , %%xmm14 			\n\t"   // y += alpha * x
+	"vfmadd231pd    112(%2,%0,8), %%xmm0 , %%xmm15 			\n\t"   // y += alpha * x
+
+	"vmovups		%%xmm8 ,   (%3,%0,8)                 	\n\t"
+	"vmovups		%%xmm9 , 16(%3,%0,8)                 	\n\t"
+	"vmovups		%%xmm10, 32(%3,%0,8)           		\n\t"
+	"vmovups		%%xmm11, 48(%3,%0,8)                 	\n\t"
+
+	"vmovups		%%xmm12, 64(%3,%0,8)                 	\n\t"
+	"vmovups		%%xmm13, 80(%3,%0,8)                 	\n\t"
+	"vmovups		%%xmm14, 96(%3,%0,8)           		\n\t"
+	"vmovups		%%xmm15,112(%3,%0,8)                 	\n\t"
+
+	"addq		$16, %0	  	 	             		\n\t"
+	"subq	        $16, %1			             		\n\t"		
+	"jnz		1b		             			\n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (alpha)   // 4
+	: "cc", 
+	  "%xmm0", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+
+} 
+
+

kernel/x86_64/daxpy_microk_steamroller-2.c

@@ -0,0 +1,160 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_8 1
+static void daxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *alpha) __attribute__ ((noinline));
+
+static void daxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
+{
+
+
+	BLASLONG register i = 0;
+
+  if ( n < 2048 )
+  {
+
+	__asm__  __volatile__
+	(
+	"vmovddup		(%4), %%xmm0		    		\n\t"  // alpha	
+
+	".align 16				            		\n\t"
+	"1:				            			\n\t"
+
+        "vmovups                  (%3,%0,8), %%xmm8          		\n\t"  // 2  y
+        "vmovups                16(%3,%0,8), %%xmm9         		\n\t"  // 2  y
+        "vmovups                32(%3,%0,8), %%xmm10         		\n\t"  // 2  y
+        "vmovups                48(%3,%0,8), %%xmm11         		\n\t"  // 2  y
+
+        "vmovups                64(%3,%0,8), %%xmm12         		\n\t"  // 2  y
+        "vmovups                80(%3,%0,8), %%xmm13        		\n\t"  // 2  y
+        "vmovups                96(%3,%0,8), %%xmm14         		\n\t"  // 2  y
+        "vmovups               112(%3,%0,8), %%xmm15         		\n\t"  // 2  y
+
+	"vfmadd231pd       (%2,%0,8), %%xmm0 , %%xmm8  			\n\t"   // y += alpha * x
+	"vfmadd231pd     16(%2,%0,8), %%xmm0 , %%xmm9 		  	\n\t"   // y += alpha * x
+	"vfmadd231pd     32(%2,%0,8), %%xmm0 , %%xmm10 			\n\t"   // y += alpha * x
+	"vfmadd231pd     48(%2,%0,8), %%xmm0 , %%xmm11 			\n\t"   // y += alpha * x
+
+	"vfmadd231pd     64(%2,%0,8), %%xmm0 , %%xmm12 			\n\t"   // y += alpha * x
+	"vfmadd231pd     80(%2,%0,8), %%xmm0 , %%xmm13		  	\n\t"   // y += alpha * x
+	"vfmadd231pd     96(%2,%0,8), %%xmm0 , %%xmm14 			\n\t"   // y += alpha * x
+	"vfmadd231pd    112(%2,%0,8), %%xmm0 , %%xmm15 			\n\t"   // y += alpha * x
+
+	"vmovups		%%xmm8 ,   (%3,%0,8)                 	\n\t"
+	"vmovups		%%xmm9 , 16(%3,%0,8)                 	\n\t"
+	"vmovups		%%xmm10, 32(%3,%0,8)           		\n\t"
+	"vmovups		%%xmm11, 48(%3,%0,8)                 	\n\t"
+
+	"vmovups		%%xmm12, 64(%3,%0,8)                 	\n\t"
+	"vmovups		%%xmm13, 80(%3,%0,8)                 	\n\t"
+	"vmovups		%%xmm14, 96(%3,%0,8)           		\n\t"
+	"vmovups		%%xmm15,112(%3,%0,8)                 	\n\t"
+
+	"addq		$16, %0	  	 	             		\n\t"
+	"subq	        $16, %1			             		\n\t"		
+	"jnz		1b		             			\n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (alpha)   // 4
+	: "cc", 
+	  "%xmm0", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+	return;
+  }
+
+
+	__asm__  __volatile__
+	(
+	"vmovddup		(%4), %%xmm0		    		\n\t"  // alpha	
+
+	".align 16				            		\n\t"
+	"1:				            			\n\t"
+
+	"prefetcht0	       512(%3,%0,8)				\n\t"
+        "vmovups                  (%3,%0,8), %%xmm8          		\n\t"  // 2  y
+        "vmovups                16(%3,%0,8), %%xmm9         		\n\t"  // 2  y
+        "vmovups                32(%3,%0,8), %%xmm10         		\n\t"  // 2  y
+        "vmovups                48(%3,%0,8), %%xmm11         		\n\t"  // 2  y
+
+	"prefetcht0	       576(%3,%0,8)				\n\t"
+        "vmovups                64(%3,%0,8), %%xmm12         		\n\t"  // 2  y
+        "vmovups                80(%3,%0,8), %%xmm13        		\n\t"  // 2  y
+        "vmovups                96(%3,%0,8), %%xmm14         		\n\t"  // 2  y
+        "vmovups               112(%3,%0,8), %%xmm15         		\n\t"  // 2  y
+
+	"prefetcht0	       512(%2,%0,8)				\n\t"
+	"vfmadd231pd       (%2,%0,8), %%xmm0 , %%xmm8  			\n\t"   // y += alpha * x
+	"vfmadd231pd     16(%2,%0,8), %%xmm0 , %%xmm9 		  	\n\t"   // y += alpha * x
+	"vfmadd231pd     32(%2,%0,8), %%xmm0 , %%xmm10 			\n\t"   // y += alpha * x
+	"vfmadd231pd     48(%2,%0,8), %%xmm0 , %%xmm11 			\n\t"   // y += alpha * x
+
+	"prefetcht0	       576(%2,%0,8)				\n\t"
+	"vfmadd231pd     64(%2,%0,8), %%xmm0 , %%xmm12 			\n\t"   // y += alpha * x
+	"vfmadd231pd     80(%2,%0,8), %%xmm0 , %%xmm13		  	\n\t"   // y += alpha * x
+	"vfmadd231pd     96(%2,%0,8), %%xmm0 , %%xmm14 			\n\t"   // y += alpha * x
+	"vfmadd231pd    112(%2,%0,8), %%xmm0 , %%xmm15 			\n\t"   // y += alpha * x
+
+	"vmovups		%%xmm8 ,   (%3,%0,8)                 	\n\t"
+	"vmovups		%%xmm9 , 16(%3,%0,8)                 	\n\t"
+	"vmovups		%%xmm10, 32(%3,%0,8)           		\n\t"
+	"vmovups		%%xmm11, 48(%3,%0,8)                 	\n\t"
+
+	"vmovups		%%xmm12, 64(%3,%0,8)                 	\n\t"
+	"vmovups		%%xmm13, 80(%3,%0,8)                 	\n\t"
+	"vmovups		%%xmm14, 96(%3,%0,8)           		\n\t"
+	"vmovups		%%xmm15,112(%3,%0,8)                 	\n\t"
+
+	"addq		$16, %0	  	 	             		\n\t"
+	"subq	        $16, %1			             		\n\t"		
+	"jnz		1b		             			\n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (alpha)   // 4
+	: "cc", 
+	  "%xmm0", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+
+} 
+
+

kernel/x86_64/ddot.c

@@ -29,8 +29,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "common.h"
 
 
-#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER)
+#if defined(BULLDOZER) 
 #include "ddot_microk_bulldozer-2.c"
+#elif defined(STEAMROLLER)
+#include "ddot_microk_steamroller-2.c"
+#elif defined(PILEDRIVER)
+#include "ddot_microk_piledriver-2.c"
 #elif defined(NEHALEM) 
 #include "ddot_microk_nehalem-2.c"
 #elif defined(HASWELL)

kernel/x86_64/ddot_microk_piledriver-2.c

@@ -0,0 +1,165 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_8 1
+static void ddot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *dot) __attribute__ ((noinline));
+
+static void ddot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
+{
+
+
+	BLASLONG register i = 0;
+
+  if ( n < 1408 )
+  {
+
+	__asm__  __volatile__
+	(
+	"vzeroupper					     \n\t"
+	"vxorpd		%%xmm4, %%xmm4, %%xmm4	             \n\t"
+	"vxorpd		%%xmm5, %%xmm5, %%xmm5	             \n\t"
+	"vxorpd		%%xmm6, %%xmm6, %%xmm6	             \n\t"
+	"vxorpd		%%xmm7, %%xmm7, %%xmm7	             \n\t"
+
+	".align 16				             \n\t"
+	"1:				             \n\t"
+        "vmovups                  (%2,%0,8), %%xmm12         \n\t"  // 2 * x
+        "vmovups                16(%2,%0,8), %%xmm13         \n\t"  // 2 * x
+        "vmovups                32(%2,%0,8), %%xmm14         \n\t"  // 2 * x
+        "vmovups                48(%2,%0,8), %%xmm15         \n\t"  // 2 * x
+
+	"vfmadd231pd         (%3,%0,8), %%xmm12, %%xmm4 \n\t"  // 2 * y
+        "vmovups                64(%2,%0,8), %%xmm0         \n\t"  // 2 * x
+        "vmovups                80(%2,%0,8), %%xmm1         \n\t"  // 2 * x
+	"vfmadd231pd       16(%3,%0,8), %%xmm13, %%xmm5 \n\t"  // 2 * y
+        "vmovups                96(%2,%0,8), %%xmm2         \n\t"  // 2 * x
+        "vmovups               112(%2,%0,8), %%xmm3         \n\t"  // 2 * x
+	"vfmadd231pd       32(%3,%0,8), %%xmm14, %%xmm6 \n\t"  // 2 * y
+	"vfmadd231pd       48(%3,%0,8), %%xmm15, %%xmm7 \n\t"  // 2 * y
+
+
+	"vfmadd231pd       64(%3,%0,8), %%xmm0 , %%xmm4 \n\t"  // 2 * y
+	"vfmadd231pd       80(%3,%0,8), %%xmm1 , %%xmm5 \n\t"  // 2 * y
+	"vfmadd231pd       96(%3,%0,8), %%xmm2 , %%xmm6 \n\t"  // 2 * y
+	"vfmadd231pd      112(%3,%0,8), %%xmm3 , %%xmm7 \n\t"  // 2 * y
+
+	"addq		$16 , %0	  	 	             \n\t"
+	"subq	        $16 , %1			             \n\t"		
+
+	"jnz		1b		             \n\t"
+
+	"vaddpd        %%xmm4, %%xmm5, %%xmm4	\n\t"
+	"vaddpd        %%xmm6, %%xmm7, %%xmm6	\n\t"
+	"vaddpd        %%xmm4, %%xmm6, %%xmm4	\n\t"
+
+	"vhaddpd        %%xmm4, %%xmm4, %%xmm4	\n\t"
+
+	"vmovsd		%%xmm4,    (%4)		\n\t"
+	"vzeroupper					     \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (dot)     // 4
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+	return;
+  }
+
+	__asm__  __volatile__
+	(
+	"vzeroupper					     \n\t"
+	"vxorpd		%%xmm4, %%xmm4, %%xmm4	             \n\t"
+	"vxorpd		%%xmm5, %%xmm5, %%xmm5	             \n\t"
+	"vxorpd		%%xmm6, %%xmm6, %%xmm6	             \n\t"
+	"vxorpd		%%xmm7, %%xmm7, %%xmm7	             \n\t"
+
+	".align 16				             \n\t"
+	"1:				             \n\t"
+	"prefetcht0	768(%2,%0,8)			     \n\t"
+	"prefetcht0	832(%2,%0,8)			     \n\t"
+        "vmovups                  (%2,%0,8), %%xmm12         \n\t"  // 2 * x
+        "vmovups                16(%2,%0,8), %%xmm13         \n\t"  // 2 * x
+        "vmovups                32(%2,%0,8), %%xmm14         \n\t"  // 2 * x
+        "vmovups                48(%2,%0,8), %%xmm15         \n\t"  // 2 * x
+
+	"prefetcht0	768(%3,%0,8)			     \n\t"
+	"prefetcht0	832(%3,%0,8)			     \n\t"
+	"vfmadd231pd         (%3,%0,8), %%xmm12, %%xmm4 \n\t"  // 2 * y
+        "vmovups                64(%2,%0,8), %%xmm0         \n\t"  // 2 * x
+        "vmovups                80(%2,%0,8), %%xmm1         \n\t"  // 2 * x
+	"vfmadd231pd       16(%3,%0,8), %%xmm13, %%xmm5 \n\t"  // 2 * y
+        "vmovups                96(%2,%0,8), %%xmm2         \n\t"  // 2 * x
+        "vmovups               112(%2,%0,8), %%xmm3         \n\t"  // 2 * x
+	"vfmadd231pd       32(%3,%0,8), %%xmm14, %%xmm6 \n\t"  // 2 * y
+	"vfmadd231pd       48(%3,%0,8), %%xmm15, %%xmm7 \n\t"  // 2 * y
+
+
+	"vfmadd231pd       64(%3,%0,8), %%xmm0 , %%xmm4 \n\t"  // 2 * y
+	"vfmadd231pd       80(%3,%0,8), %%xmm1 , %%xmm5 \n\t"  // 2 * y
+	"vfmadd231pd       96(%3,%0,8), %%xmm2 , %%xmm6 \n\t"  // 2 * y
+	"vfmadd231pd      112(%3,%0,8), %%xmm3 , %%xmm7 \n\t"  // 2 * y
+
+	"addq		$16 , %0	  	 	             \n\t"
+	"subq	        $16 , %1			             \n\t"		
+
+	"jnz		1b		             \n\t"
+
+	"vaddpd        %%xmm4, %%xmm5, %%xmm4	\n\t"
+	"vaddpd        %%xmm6, %%xmm7, %%xmm6	\n\t"
+	"vaddpd        %%xmm4, %%xmm6, %%xmm4	\n\t"
+
+	"vhaddpd        %%xmm4, %%xmm4, %%xmm4	\n\t"
+
+	"vmovsd		%%xmm4,    (%4)		\n\t"
+	"vzeroupper					     \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (dot)     // 4
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+
+} 
+
+

kernel/x86_64/ddot_microk_steamroller-2.c

@@ -0,0 +1,97 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_8 1
+static void ddot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *dot) __attribute__ ((noinline));
+
+static void ddot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
+{
+
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper					     \n\t"
+	"vxorpd		%%xmm4, %%xmm4, %%xmm4	             \n\t"
+	"vxorpd		%%xmm5, %%xmm5, %%xmm5	             \n\t"
+	"vxorpd		%%xmm6, %%xmm6, %%xmm6	             \n\t"
+	"vxorpd		%%xmm7, %%xmm7, %%xmm7	             \n\t"
+
+	".align 16				             \n\t"
+	"1:				             \n\t"
+        "vmovups                  (%2,%0,8), %%xmm12         \n\t"  // 2 * x
+        "vmovups                16(%2,%0,8), %%xmm13         \n\t"  // 2 * x
+        "vmovups                32(%2,%0,8), %%xmm14         \n\t"  // 2 * x
+        "vmovups                48(%2,%0,8), %%xmm15         \n\t"  // 2 * x
+
+	"vfmadd231pd         (%3,%0,8), %%xmm12, %%xmm4 \n\t"  // 2 * y
+        "vmovups                64(%2,%0,8), %%xmm0         \n\t"  // 2 * x
+        "vmovups                80(%2,%0,8), %%xmm1         \n\t"  // 2 * x
+	"vfmadd231pd       16(%3,%0,8), %%xmm13, %%xmm5 \n\t"  // 2 * y
+        "vmovups                96(%2,%0,8), %%xmm2         \n\t"  // 2 * x
+        "vmovups               112(%2,%0,8), %%xmm3         \n\t"  // 2 * x
+	"vfmadd231pd       32(%3,%0,8), %%xmm14, %%xmm6 \n\t"  // 2 * y
+	"vfmadd231pd       48(%3,%0,8), %%xmm15, %%xmm7 \n\t"  // 2 * y
+
+
+	"vfmadd231pd       64(%3,%0,8), %%xmm0 , %%xmm4 \n\t"  // 2 * y
+	"vfmadd231pd       80(%3,%0,8), %%xmm1 , %%xmm5 \n\t"  // 2 * y
+	"vfmadd231pd       96(%3,%0,8), %%xmm2 , %%xmm6 \n\t"  // 2 * y
+	"vfmadd231pd      112(%3,%0,8), %%xmm3 , %%xmm7 \n\t"  // 2 * y
+
+	"addq		$16 , %0	  	 	             \n\t"
+	"subq	        $16 , %1			             \n\t"		
+
+	"jnz		1b		             \n\t"
+
+	"vaddpd        %%xmm4, %%xmm5, %%xmm4	\n\t"
+	"vaddpd        %%xmm6, %%xmm7, %%xmm6	\n\t"
+	"vaddpd        %%xmm4, %%xmm6, %%xmm4	\n\t"
+
+	"vhaddpd        %%xmm4, %%xmm4, %%xmm4	\n\t"
+
+	"vmovsd		%%xmm4,    (%4)		\n\t"
+	"vzeroupper					     \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (dot)     // 4
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

kernel/x86_64/saxpy.c

@@ -35,6 +35,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "saxpy_microk_haswell-2.c"
 #elif defined(SANDYBRIDGE)
 #include "saxpy_microk_sandy-2.c"
+#elif defined(PILEDRIVER) || defined(STEAMROLLER)
+#include "saxpy_microk_piledriver-2.c"
 #endif
 
 

kernel/x86_64/saxpy_microk_piledriver-2.c

@@ -0,0 +1,159 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16 1
+static void saxpy_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *alpha) __attribute__ ((noinline));
+
+static void saxpy_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
+{
+
+
+	BLASLONG register i = 0;
+
+  if ( n < 1024 )
+  {
+
+
+	__asm__  __volatile__
+	(
+	"vzeroupper				     \n\t"
+	"vbroadcastss		(%4), %%xmm0		    \n\t"  // alpha	
+
+	".align 16				            \n\t"
+	"1:				            \n\t"
+
+        "vmovups                  (%3,%0,4), %%xmm8          \n\t"  // 4 * y
+        "vmovups                16(%3,%0,4), %%xmm9          \n\t"  // 4 * y
+        "vmovups                32(%3,%0,4), %%xmm10         \n\t"  // 4 * y
+        "vmovups                48(%3,%0,4), %%xmm11         \n\t"  // 4 * y
+        "vmovups                64(%3,%0,4), %%xmm12         \n\t"  // 4 * y
+        "vmovups                80(%3,%0,4), %%xmm13         \n\t"  // 4 * y
+        "vmovups                96(%3,%0,4), %%xmm14         \n\t"  // 4 * y
+        "vmovups               112(%3,%0,4), %%xmm15         \n\t"  // 4 * y
+
+	"vfmadd231ps       (%2,%0,4), %%xmm0  , %%xmm8   	     \n\t"   // y += alpha * x
+	"vfmadd231ps     16(%2,%0,4), %%xmm0  , %%xmm9   	     \n\t"   // y += alpha * x
+	"vfmadd231ps     32(%2,%0,4), %%xmm0  , %%xmm10  	     \n\t"   // y += alpha * x
+	"vfmadd231ps     48(%2,%0,4), %%xmm0  , %%xmm11  	     \n\t"   // y += alpha * x
+	"vfmadd231ps     64(%2,%0,4), %%xmm0  , %%xmm12  	     \n\t"   // y += alpha * x
+	"vfmadd231ps     80(%2,%0,4), %%xmm0  , %%xmm13  	     \n\t"   // y += alpha * x
+	"vfmadd231ps     96(%2,%0,4), %%xmm0  , %%xmm14  	     \n\t"   // y += alpha * x
+	"vfmadd231ps    112(%2,%0,4), %%xmm0  , %%xmm15  	     \n\t"   // y += alpha * x
+
+	"vmovups	%%xmm8 ,   (%3,%0,4)		     \n\t"
+	"vmovups	%%xmm9 , 16(%3,%0,4)		     \n\t"
+	"vmovups	%%xmm10, 32(%3,%0,4)		     \n\t"
+	"vmovups	%%xmm11, 48(%3,%0,4)		     \n\t"
+	"vmovups	%%xmm12, 64(%3,%0,4)		     \n\t"
+	"vmovups	%%xmm13, 80(%3,%0,4)		     \n\t"
+	"vmovups	%%xmm14, 96(%3,%0,4)		     \n\t"
+	"vmovups	%%xmm15,112(%3,%0,4)		     \n\t"
+
+	"addq		$32, %0	  	 	             \n\t"
+	"subq	        $32, %1			             \n\t"		
+	"jnz		1b		             \n\t"
+	"vzeroupper				     \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (alpha)   // 4
+	: "cc", 
+	  "%xmm0", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+	return;
+
+  }
+
+	__asm__  __volatile__
+	(
+	"vzeroupper				     \n\t"
+	"vbroadcastss		(%4), %%xmm0		    \n\t"  // alpha	
+
+	".align 16				            \n\t"
+	"1:				            \n\t"
+	"prefetcht0		512(%3,%0,4)	             \n\t"
+	"prefetcht0		576(%3,%0,4)	             \n\t"
+
+        "vmovups                  (%3,%0,4), %%xmm8          \n\t"  // 4 * y
+        "vmovups                16(%3,%0,4), %%xmm9          \n\t"  // 4 * y
+        "vmovups                32(%3,%0,4), %%xmm10         \n\t"  // 4 * y
+        "vmovups                48(%3,%0,4), %%xmm11         \n\t"  // 4 * y
+        "vmovups                64(%3,%0,4), %%xmm12         \n\t"  // 4 * y
+        "vmovups                80(%3,%0,4), %%xmm13         \n\t"  // 4 * y
+        "vmovups                96(%3,%0,4), %%xmm14         \n\t"  // 4 * y
+        "vmovups               112(%3,%0,4), %%xmm15         \n\t"  // 4 * y
+
+	"prefetcht0		512(%2,%0,4)	             \n\t"
+	"prefetcht0		576(%2,%0,4)	             \n\t"
+	"vfmadd231ps       (%2,%0,4), %%xmm0  , %%xmm8   	     \n\t"   // y += alpha * x
+	"vfmadd231ps     16(%2,%0,4), %%xmm0  , %%xmm9   	     \n\t"   // y += alpha * x
+	"vfmadd231ps     32(%2,%0,4), %%xmm0  , %%xmm10  	     \n\t"   // y += alpha * x
+	"vfmadd231ps     48(%2,%0,4), %%xmm0  , %%xmm11  	     \n\t"   // y += alpha * x
+	"vfmadd231ps     64(%2,%0,4), %%xmm0  , %%xmm12  	     \n\t"   // y += alpha * x
+	"vfmadd231ps     80(%2,%0,4), %%xmm0  , %%xmm13  	     \n\t"   // y += alpha * x
+	"vfmadd231ps     96(%2,%0,4), %%xmm0  , %%xmm14  	     \n\t"   // y += alpha * x
+	"vfmadd231ps    112(%2,%0,4), %%xmm0  , %%xmm15  	     \n\t"   // y += alpha * x
+
+	"vmovups	%%xmm8 ,   (%3,%0,4)		     \n\t"
+	"vmovups	%%xmm9 , 16(%3,%0,4)		     \n\t"
+	"vmovups	%%xmm10, 32(%3,%0,4)		     \n\t"
+	"vmovups	%%xmm11, 48(%3,%0,4)		     \n\t"
+	"vmovups	%%xmm12, 64(%3,%0,4)		     \n\t"
+	"vmovups	%%xmm13, 80(%3,%0,4)		     \n\t"
+	"vmovups	%%xmm14, 96(%3,%0,4)		     \n\t"
+	"vmovups	%%xmm15,112(%3,%0,4)		     \n\t"
+
+	"addq		$32, %0	  	 	             \n\t"
+	"subq	        $32, %1			             \n\t"		
+	"jnz		1b		             \n\t"
+	"vzeroupper				     \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (alpha)   // 4
+	: "cc", 
+	  "%xmm0", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+
+} 
+
+

kernel/x86_64/sdot.c

@@ -28,8 +28,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include "common.h"
 
-#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER)
+#if defined(BULLDOZER) 
 #include "sdot_microk_bulldozer-2.c"
+#elif defined(STEAMROLLER) || defined(PILEDRIVER)
+#include "sdot_microk_steamroller-2.c"
 #elif defined(NEHALEM)
 #include "sdot_microk_nehalem-2.c"
 #elif defined(HASWELL)

kernel/x86_64/sdot_microk_steamroller-2.c

@@ -0,0 +1,163 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16 1
+static void sdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *dot) __attribute__ ((noinline));
+
+static void sdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
+{
+
+
+	BLASLONG register i = 0;
+
+
+  if ( n < 4096 )
+  {
+
+	__asm__  __volatile__
+	(
+	"vxorps		%%xmm4, %%xmm4, %%xmm4	             \n\t"
+	"vxorps		%%xmm5, %%xmm5, %%xmm5	             \n\t"
+	"vxorps		%%xmm6, %%xmm6, %%xmm6	             \n\t"
+	"vxorps		%%xmm7, %%xmm7, %%xmm7	             \n\t"
+
+	".align 16				             \n\t"
+	"1:				             \n\t"
+        "vmovups                  (%2,%0,4), %%xmm12         \n\t"  // 4 * x
+        "vmovups                16(%2,%0,4), %%xmm13         \n\t"  // 4 * x
+        "vmovups                32(%2,%0,4), %%xmm14         \n\t"  // 4 * x
+        "vmovups                48(%2,%0,4), %%xmm15         \n\t"  // 4 * x
+
+	"vfmadd231ps       (%3,%0,4), %%xmm12, %%xmm4 \n\t"  // 4 * y
+	"vfmadd231ps     16(%3,%0,4), %%xmm13, %%xmm5 \n\t"  // 4 * y
+        "vmovups                64(%2,%0,4), %%xmm0          \n\t"  // 4 * x
+        "vmovups                80(%2,%0,4), %%xmm1          \n\t"  // 4 * x
+	"vfmadd231ps     32(%3,%0,4), %%xmm14, %%xmm6 \n\t"  // 4 * y
+	"vfmadd231ps     48(%3,%0,4), %%xmm15, %%xmm7 \n\t"  // 4 * y
+
+        "vmovups                96(%2,%0,4), %%xmm2          \n\t"  // 4 * x
+        "vmovups               112(%2,%0,4), %%xmm3          \n\t"  // 4 * x
+
+	"vfmadd231ps     64(%3,%0,4), %%xmm0 , %%xmm4 \n\t"  // 4 * y
+	"vfmadd231ps     80(%3,%0,4), %%xmm1 , %%xmm5 \n\t"  // 4 * y
+	"vfmadd231ps     96(%3,%0,4), %%xmm2 , %%xmm6 \n\t"  // 4 * y
+	"vfmadd231ps    112(%3,%0,4), %%xmm3 , %%xmm7 \n\t"  // 4 * y
+
+	"addq		$32, %0	  	 	             \n\t"
+	"subq	        $32, %1			             \n\t"		
+	"jnz		1b		             \n\t"
+
+	"vaddps        %%xmm4, %%xmm5, %%xmm4	\n\t"
+	"vaddps        %%xmm6, %%xmm7, %%xmm6	\n\t"
+	"vaddps        %%xmm4, %%xmm6, %%xmm4	\n\t"
+
+	"vhaddps        %%xmm4, %%xmm4, %%xmm4	\n\t"
+	"vhaddps        %%xmm4, %%xmm4, %%xmm4	\n\t"
+
+	"vmovss		%%xmm4,    (%4)		\n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (dot)     // 4
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+	return;
+
+  }
+
+	__asm__  __volatile__
+	(
+	"vxorps		%%xmm4, %%xmm4, %%xmm4	             \n\t"
+	"vxorps		%%xmm5, %%xmm5, %%xmm5	             \n\t"
+	"vxorps		%%xmm6, %%xmm6, %%xmm6	             \n\t"
+	"vxorps		%%xmm7, %%xmm7, %%xmm7	             \n\t"
+
+	".align 16				             \n\t"
+	"1:				             \n\t"
+	"prefetcht0	512(%2,%0,4)		     \n\t"
+        "vmovups                  (%2,%0,4), %%xmm12         \n\t"  // 4 * x
+        "vmovups                16(%2,%0,4), %%xmm13         \n\t"  // 4 * x
+        "vmovups                32(%2,%0,4), %%xmm14         \n\t"  // 4 * x
+        "vmovups                48(%2,%0,4), %%xmm15         \n\t"  // 4 * x
+
+	"prefetcht0	512(%3,%0,4)		     \n\t"
+	"vfmadd231ps       (%3,%0,4), %%xmm12, %%xmm4 \n\t"  // 4 * y
+	"vfmadd231ps     16(%3,%0,4), %%xmm13, %%xmm5 \n\t"  // 4 * y
+	"prefetcht0	576(%2,%0,4)		     \n\t"
+        "vmovups                64(%2,%0,4), %%xmm0          \n\t"  // 4 * x
+        "vmovups                80(%2,%0,4), %%xmm1          \n\t"  // 4 * x
+	"prefetcht0	576(%3,%0,4)		     \n\t"
+	"vfmadd231ps     32(%3,%0,4), %%xmm14, %%xmm6 \n\t"  // 4 * y
+	"vfmadd231ps     48(%3,%0,4), %%xmm15, %%xmm7 \n\t"  // 4 * y
+
+        "vmovups                96(%2,%0,4), %%xmm2          \n\t"  // 4 * x
+        "vmovups               112(%2,%0,4), %%xmm3          \n\t"  // 4 * x
+
+	"vfmadd231ps     64(%3,%0,4), %%xmm0 , %%xmm4 \n\t"  // 4 * y
+	"vfmadd231ps     80(%3,%0,4), %%xmm1 , %%xmm5 \n\t"  // 4 * y
+	"vfmadd231ps     96(%3,%0,4), %%xmm2 , %%xmm6 \n\t"  // 4 * y
+	"vfmadd231ps    112(%3,%0,4), %%xmm3 , %%xmm7 \n\t"  // 4 * y
+
+	"addq		$32, %0	  	 	             \n\t"
+	"subq	        $32, %1			             \n\t"		
+	"jnz		1b		             \n\t"
+
+	"vaddps        %%xmm4, %%xmm5, %%xmm4	\n\t"
+	"vaddps        %%xmm6, %%xmm7, %%xmm6	\n\t"
+	"vaddps        %%xmm4, %%xmm6, %%xmm4	\n\t"
+
+	"vhaddps        %%xmm4, %%xmm4, %%xmm4	\n\t"
+	"vhaddps        %%xmm4, %%xmm4, %%xmm4	\n\t"
+
+	"vmovss		%%xmm4,    (%4)		\n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (dot)     // 4
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+
+} 
+
+