added optimized sger kernel for sandybridge
authorWerner Saar <wernsaar@googlemail.com>
Tue, 28 Apr 2015 13:33:38 +0000 (15:33 +0200)
committerWerner Saar <wernsaar@googlemail.com>
Tue, 28 Apr 2015 13:33:38 +0000 (15:33 +0200)
kernel/x86_64/KERNEL.SANDYBRIDGE
kernel/x86_64/sger.c [new file with mode: 0644]
kernel/x86_64/sger_microk_sandy-2.c [new file with mode: 0644]

index 055072c..06bb7f3 100644 (file)
@@ -1,3 +1,5 @@
+SGERKERNEL = sger.c
+
 SGEMVNKERNEL = sgemv_n_4.c
 SGEMVTKERNEL = sgemv_t_4.c
 
diff --git a/kernel/x86_64/sger.c b/kernel/x86_64/sger.c
new file mode 100644 (file)
index 0000000..84c056c
--- /dev/null
@@ -0,0 +1,84 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* 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.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``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 UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN 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.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#include <stdio.h>
+#include "common.h"
+
+#if defined(SANDYBRIDGE)
+#include "sger_microk_sandy-2.c"
+#endif
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha,
+        FLOAT *x, BLASLONG incx,
+        FLOAT *y, BLASLONG incy,
+        FLOAT *a, BLASLONG lda, FLOAT *buffer){
+
+       FLOAT *X    = x;
+
+       if (incx != 1) {
+               X = buffer;
+               COPY_K(m, x, incx, X, 1);
+       }
+
+       BLASLONG m1 = m & -16;
+
+       while (n > 0) 
+       {
+               FLOAT y0 = alpha * *y;
+               if ( m1 > 0 )
+               {
+                       #ifdef HAVE_KERNEL_16
+                               sger_kernel_16(m1, X, a, &y0);
+                       #else
+                               AXPYU_K(m1, 0, 0, y0, X, 1, a, 1, NULL, 0);
+                       #endif
+               }
+       
+               if ( m > m1 )
+               {
+                       AXPYU_K(m-m1, 0, 0, y0, X+m1 , 1, a+m1, 1, NULL, 0);
+               }
+       
+               a += lda;
+               y += incy;
+               n --;
+       }
+
+       return 0;
+}
+
diff --git a/kernel/x86_64/sger_microk_sandy-2.c b/kernel/x86_64/sger_microk_sandy-2.c
new file mode 100644 (file)
index 0000000..51c3bef
--- /dev/null
@@ -0,0 +1,124 @@
+/***************************************************************************
+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 sger_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *alpha) __attribute__ ((noinline));
+
+static void sger_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
+{
+
+
+       BLASLONG register i = 0;
+
+       __asm__  __volatile__
+       (
+       "vbroadcastss           (%4), %%xmm0                \n\t"  // alpha     
+       "prefetcht0     256(%3,%0,4)                        \n\t"
+       "vmovups          (%3,%0,4), %%xmm8         \n\t"
+       "vmovups        16(%3,%0,4), %%xmm9         \n\t"
+       "vmovups        32(%3,%0,4), %%xmm10        \n\t"
+       "vmovups        48(%3,%0,4), %%xmm11        \n\t"
+
+       "prefetcht0     256(%2,%0,4)                        \n\t"
+       "vmovups          (%2,%0,4), %%xmm4         \n\t"
+       "vmovups        16(%2,%0,4), %%xmm5         \n\t"
+       "vmovups        32(%2,%0,4), %%xmm6         \n\t"
+       "vmovups        48(%2,%0,4), %%xmm7        \n\t"
+
+       "addq           $16, %0                              \n\t"
+       "subq           $16, %1                              \n\t"              
+       "jz             2f                           \n\t"
+
+       ".align 16                                          \n\t"
+       "1:                                         \n\t"
+
+       "vmulps         %%xmm4, %%xmm0, %%xmm4          \n\t"
+       "vaddps         %%xmm8 , %%xmm4, %%xmm12             \n\t"
+       "vmulps         %%xmm5, %%xmm0, %%xmm5          \n\t"
+       "vaddps         %%xmm9 , %%xmm5, %%xmm13             \n\t"
+       "vmulps         %%xmm6, %%xmm0, %%xmm6          \n\t"
+       "vaddps         %%xmm10, %%xmm6, %%xmm14             \n\t"
+       "vmulps         %%xmm7, %%xmm0, %%xmm7          \n\t"
+       "vaddps         %%xmm11, %%xmm7, %%xmm15             \n\t"
+
+       "prefetcht0     256(%3,%0,4)                        \n\t"
+       "vmovups          (%3,%0,4), %%xmm8         \n\t"
+       "vmovups        16(%3,%0,4), %%xmm9         \n\t"
+       "vmovups        32(%3,%0,4), %%xmm10        \n\t"
+       "vmovups        48(%3,%0,4), %%xmm11        \n\t"
+
+       "prefetcht0     256(%2,%0,4)                        \n\t"
+       "vmovups          (%2,%0,4), %%xmm4         \n\t"
+       "vmovups        16(%2,%0,4), %%xmm5         \n\t"
+       "vmovups        32(%2,%0,4), %%xmm6         \n\t"
+       "vmovups        48(%2,%0,4), %%xmm7        \n\t"
+
+       "vmovups        %%xmm12,  -64(%3,%0,4)               \n\t"
+       "vmovups        %%xmm13,  -48(%3,%0,4)               \n\t"
+       "vmovups        %%xmm14,  -32(%3,%0,4)               \n\t"
+       "vmovups        %%xmm15,  -16(%3,%0,4)               \n\t"
+
+       "addq           $16, %0                              \n\t"
+       "subq           $16, %1                              \n\t"              
+       "jnz            1b                           \n\t"
+
+       "2:                                         \n\t"
+       "vmulps         %%xmm4, %%xmm0, %%xmm4          \n\t"
+       "vmulps         %%xmm5, %%xmm0, %%xmm5          \n\t"
+       "vmulps         %%xmm6, %%xmm0, %%xmm6          \n\t"
+       "vmulps         %%xmm7, %%xmm0, %%xmm7          \n\t"
+
+       "vaddps         %%xmm8 , %%xmm4, %%xmm12             \n\t"
+       "vaddps         %%xmm9 , %%xmm5, %%xmm13             \n\t"
+       "vaddps         %%xmm10, %%xmm6, %%xmm14             \n\t"
+       "vaddps         %%xmm11, %%xmm7, %%xmm15             \n\t"
+
+       "vmovups        %%xmm12,  -64(%3,%0,4)               \n\t"
+       "vmovups        %%xmm13,  -48(%3,%0,4)               \n\t"
+       "vmovups        %%xmm14,  -32(%3,%0,4)               \n\t"
+       "vmovups        %%xmm15,  -16(%3,%0,4)               \n\t"
+
+       "vzeroupper                                          \n\t"
+
+       :
+        : 
+          "r" (i),     // 0    
+         "r" (n),      // 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (alpha)   // 4
+       : "cc", 
+         "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+         "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+         "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+         "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+         "memory"
+       );
+
+} 
+
+