added optimized dgemv_t kernel for haswell
authorwernsaar <wernsaar@googlemail.com>
Thu, 7 Aug 2014 08:08:54 +0000 (10:08 +0200)
committerwernsaar <wernsaar@googlemail.com>
Thu, 7 Aug 2014 08:08:54 +0000 (10:08 +0200)
kernel/x86_64/KERNEL.HASWELL
kernel/x86_64/dgemv_t.c [new file with mode: 0644]
kernel/x86_64/dgemv_t_microk_haswell-2.c [new file with mode: 0644]

index bcff2e2..cd280e4 100644 (file)
@@ -1,6 +1,7 @@
 SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t.c
 DGEMVNKERNEL = dgemv_n.c
+DGEMVTKERNEL = dgemv_t.c
 
 SGEMMKERNEL    =  sgemm_kernel_16x4_haswell.S
 SGEMMINCOPY    =  ../generic/gemm_ncopy_16.c
diff --git a/kernel/x86_64/dgemv_t.c b/kernel/x86_64/dgemv_t.c
new file mode 100644 (file)
index 0000000..76aacd3
--- /dev/null
@@ -0,0 +1,188 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include "common.h"
+
+#if defined(HASWELL)
+#include "dgemv_t_microk_haswell-2.c"
+#endif
+
+#define NBMAX 2048
+
+#ifndef HAVE_KERNEL_16x4
+
+static void dgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+       BLASLONG i;
+       FLOAT *a0,*a1,*a2,*a3;
+       a0 = ap[0];
+       a1 = ap[1];
+       a2 = ap[2];
+       a3 = ap[3];
+       FLOAT temp0 = 0.0;
+       FLOAT temp1 = 0.0;
+       FLOAT temp2 = 0.0;
+       FLOAT temp3 = 0.0;
+
+       for ( i=0; i< n; i+=4 )
+       {
+               temp0 += a0[i]*x[i] + a0[i+1]*x[i+1] + a0[i+2]*x[i+2] + a0[i+3]*x[i+3];         
+               temp1 += a1[i]*x[i] + a1[i+1]*x[i+1] + a1[i+2]*x[i+2] + a1[i+3]*x[i+3];         
+               temp2 += a2[i]*x[i] + a2[i+1]*x[i+1] + a2[i+2]*x[i+2] + a2[i+3]*x[i+3];         
+               temp3 += a3[i]*x[i] + a3[i+1]*x[i+1] + a3[i+2]*x[i+2] + a3[i+3]*x[i+3];         
+       }
+       y[0] = temp0;
+       y[1] = temp1;
+       y[2] = temp2;
+       y[3] = temp3;
+}
+       
+#endif
+
+static void dgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
+{
+       BLASLONG i;
+       FLOAT *a0;
+       a0 = ap;
+       FLOAT temp = 0.0;
+
+       for ( i=0; i< n; i+=4 )
+       {
+               temp += a0[i]*x[i] + a0[i+1]*x[i+1] + a0[i+2]*x[i+2] + a0[i+3]*x[i+3];          
+       }
+       *y = temp;
+}
+       
+static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
+{
+        BLASLONG i;
+        for ( i=0; i<n; i++ )
+        {
+                *dest = *src;
+                dest++;
+                src += inc_src;
+        }
+}
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+       BLASLONG i;
+       BLASLONG j;
+       FLOAT *a_ptr;
+       FLOAT *x_ptr;
+       FLOAT *y_ptr;
+       FLOAT *ap[4];
+       BLASLONG n1;
+       BLASLONG m1;
+       BLASLONG m2;
+       BLASLONG n2;
+       FLOAT ybuffer[4],*xbuffer;
+
+       xbuffer = buffer;
+       
+       n1 = n / 4 ;
+       n2 = n % 4 ;
+       
+       m1 = m - ( m % 16 );
+       m2 = (m % NBMAX) - (m % 16) ;
+       
+
+       BLASLONG NB = NBMAX;
+
+       while ( NB == NBMAX )
+       {
+               
+               m1 -= NB;
+               if ( m1 < 0)
+               {
+                       if ( m2 == 0 ) break;   
+                       NB = m2;
+               }
+               
+               y_ptr = y;
+               a_ptr = a;
+               x_ptr = x;
+               copy_x(NB,x_ptr,xbuffer,inc_x);
+               for( i = 0; i < n1 ; i++)
+               {
+                       ap[0] = a_ptr;
+                       ap[1] = a_ptr + lda;
+                       ap[2] = ap[1] + lda;
+                       ap[3] = ap[2] + lda;
+                       dgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
+                       a_ptr += 4 * lda;
+                       *y_ptr += ybuffer[0]*alpha;
+                       y_ptr  += inc_y;
+                       *y_ptr += ybuffer[1]*alpha;
+                       y_ptr  += inc_y;
+                       *y_ptr += ybuffer[2]*alpha;
+                       y_ptr  += inc_y;
+                       *y_ptr += ybuffer[3]*alpha;
+                       y_ptr  += inc_y;
+               }
+
+               for( i = 0; i < n2 ; i++)
+               {
+                       dgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
+                       a_ptr += 1 * lda;
+                       *y_ptr += ybuffer[0]*alpha;
+                       y_ptr  += inc_y;
+
+               }
+               a += NB;
+               x += NB * inc_x;        
+       }
+
+       BLASLONG m3 = m % 16;
+       if ( m3 == 0 ) return(0);
+       x_ptr = x;
+       for ( i=0; i< m3; i++ )
+       {
+               xbuffer[i] = *x_ptr;
+               x_ptr += inc_x;
+       }
+       j=0;
+       a_ptr = a;
+       y_ptr = y;
+       while ( j < n)
+       {
+               FLOAT temp = 0.0;
+               for( i = 0; i < m3; i++ )
+               {
+                       temp += a_ptr[i] * xbuffer[i];
+               }
+               a_ptr += lda;
+               y_ptr[0] += alpha * temp;
+               y_ptr += inc_y;
+               j++;
+       }
+       return(0);
+}
+
+
diff --git a/kernel/x86_64/dgemv_t_microk_haswell-2.c b/kernel/x86_64/dgemv_t_microk_haswell-2.c
new file mode 100644 (file)
index 0000000..94d4c31
--- /dev/null
@@ -0,0 +1,107 @@
+/***************************************************************************
+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_16x4 1
+static void dgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y) __attribute__ ((noinline));
+
+static void dgemv_kernel_16x4( BLASLONG n, float **ap, float *x, float *y)
+{
+
+       BLASLONG register i = 0;
+
+       __asm__  __volatile__
+       (
+       "vzeroupper                              \n\t"
+       "vxorpd         %%ymm4 , %%ymm4, %%ymm4  \n\t"
+       "vxorpd         %%ymm5 , %%ymm5, %%ymm5  \n\t"
+       "vxorpd         %%ymm6 , %%ymm6, %%ymm6  \n\t"
+       "vxorpd         %%ymm7 , %%ymm7, %%ymm7  \n\t"
+
+       ".align 16                               \n\t"
+       ".L01LOOP%=:                             \n\t"
+       "prefetcht0      384(%2,%0,8)            \n\t"
+       "vmovups        (%2,%0,8), %%ymm12       \n\t"  // 4 * x
+       "vmovups      32(%2,%0,8), %%ymm13       \n\t"  // 4 * x
+
+       "prefetcht0      384(%4,%0,8)                  \n\t"
+       "vfmadd231pd   (%4,%0,8), %%ymm12, %%ymm4      \n\t" 
+       "vfmadd231pd   (%5,%0,8), %%ymm12, %%ymm5      \n\t" 
+       "prefetcht0      384(%5,%0,8)                  \n\t"
+       "vfmadd231pd 32(%4,%0,8), %%ymm13, %%ymm4      \n\t" 
+       "vfmadd231pd 32(%5,%0,8), %%ymm13, %%ymm5      \n\t" 
+       "prefetcht0      384(%6,%0,8)                  \n\t"
+       "vfmadd231pd   (%6,%0,8), %%ymm12, %%ymm6      \n\t" 
+       "vfmadd231pd   (%7,%0,8), %%ymm12, %%ymm7      \n\t" 
+       "prefetcht0      384(%7,%0,8)                  \n\t"
+       "vfmadd231pd 32(%6,%0,8), %%ymm13, %%ymm6      \n\t" 
+       "vfmadd231pd 32(%7,%0,8), %%ymm13, %%ymm7      \n\t" 
+
+        "addq          $8 , %0                       \n\t"
+       "subq           $8 , %1                       \n\t"             
+       "jnz            .L01LOOP%=                    \n\t"
+
+       "vextractf128   $1 , %%ymm4, %%xmm12          \n\t"
+       "vextractf128   $1 , %%ymm5, %%xmm13          \n\t"
+       "vextractf128   $1 , %%ymm6, %%xmm14          \n\t"
+       "vextractf128   $1 , %%ymm7, %%xmm15          \n\t"
+
+       "vaddpd         %%xmm4, %%xmm12, %%xmm4       \n\t"
+       "vaddpd         %%xmm5, %%xmm13, %%xmm5       \n\t"
+       "vaddpd         %%xmm6, %%xmm14, %%xmm6       \n\t"
+       "vaddpd         %%xmm7, %%xmm15, %%xmm7       \n\t"
+
+        "vhaddpd        %%xmm4, %%xmm4, %%xmm4  \n\t"
+        "vhaddpd        %%xmm5, %%xmm5, %%xmm5  \n\t"
+        "vhaddpd        %%xmm6, %%xmm6, %%xmm6  \n\t"
+        "vhaddpd        %%xmm7, %%xmm7, %%xmm7  \n\t"
+
+        "vmovsd         %%xmm4,    (%3)         \n\t"
+        "vmovsd         %%xmm5,   8(%3)         \n\t"
+        "vmovsd         %%xmm6,  16(%3)         \n\t"
+        "vmovsd         %%xmm7,  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
+       : "cc", 
+         "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+         "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+         "memory"
+       );
+
+} 
+
+