Optimize s/drot function for POWER10
authorRajalakshmi Srinivasaraghavan <rajis@linux.ibm.com>
Thu, 21 Jan 2021 19:24:45 +0000 (13:24 -0600)
committerRajalakshmi Srinivasaraghavan <rajis@linux.ibm.com>
Thu, 21 Jan 2021 19:24:45 +0000 (13:24 -0600)
This patch makes use of new POWER10 vector pair instructions for
loads and stores.

kernel/power/drot.c
kernel/power/drot_microk_power10.c [new file with mode: 0644]
kernel/power/srot.c
kernel/power/srot_microk_power10.c [new file with mode: 0644]

index 951c2f9..94d9d95 100644 (file)
@@ -39,9 +39,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #pragma GCC optimize "O1"
 
-#if defined(POWER8) || defined(POWER9) || defined(POWER10)
 #if defined(__VEC__) || defined(__ALTIVEC__)
+#if defined(POWER8) || defined(POWER9)
 #include "drot_microk_power8.c"
+#elif defined(POWER10)
+#include "drot_microk_power10.c"
 #endif
 #endif
 
@@ -115,12 +117,30 @@ int CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT
        if ( (inc_x == 1) && (inc_y == 1) )
        {
 
+#if defined(POWER10)
+               if ( n >= 16 )
+               {
+                       BLASLONG align = ((32 - ((uintptr_t)y & (uintptr_t)0x1F)) >> 3) & 0x3;
+                       for (i = 0; i < align; i++) {
+                               temp  = c*x[i] + s*y[i] ;
+                               y[i]  = c*y[i] - s*x[i] ;
+                               x[i]  = temp ;
+                       }
+               }
+               BLASLONG n1 = (n-i) & -16;
+               if ( n1 > 0 )
+               {
+                       drot_kernel_16(n1,&x[i], &y[i], c, s);
+                       i+=n1;
+               }
+#else
                BLASLONG n1 = n & -16;
                if ( n1 > 0 )
                {
                        drot_kernel_16(n1, x1, y1, c, s);
                        i=n1;
                }
+#endif
 
                while(i < n)
                {
diff --git a/kernel/power/drot_microk_power10.c b/kernel/power/drot_microk_power10.c
new file mode 100644 (file)
index 0000000..e34e745
--- /dev/null
@@ -0,0 +1,148 @@
+/***************************************************************************
+Copyright (c) 2021, 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 drot_kernel_16 (long n, double *x, double *y, double c, double s)
+{
+  __asm__
+    (
+       XXSPLTD_S(36,%x5,0)     // load c to both dwords
+       XXSPLTD_S(37,%x6,0)     // load s to both dwords
+       "lxvp            32, 0(%3)       \n\t"   // load x
+       "lxvp            34, 32(%3)      \n\t"
+       "lxvp            48, 0(%4)       \n\t"   // load y
+       "lxvp            50, 32(%4)      \n\t"
+
+       "addic.         %2, %2, -8      \n\t"
+       "ble            two%=           \n\t"
+
+       ".align 5               \n"
+     "one%=:                           \n\t"
+
+       "xvmuldp                40, 32, 36      \n\t"   // c * x
+       "xvmuldp                41, 33, 36      \n\t"
+       "xvmuldp                42, 34, 36      \n\t"
+       "xvmuldp                43, 35, 36      \n\t"
+
+       "xvmuldp                44, 32, 37      \n\t"   // s * x
+       "xvmuldp                45, 33, 37      \n\t"
+       "xvmuldp                46, 34, 37      \n\t"
+       "xvmuldp                47, 35, 37      \n\t"
+
+       "lxvp            32, 64(%3)       \n\t"   // load x
+       "lxvp            34, 96(%3)      \n\t"
+       "xvmuldp                52, 48, 36      \n\t"   // c * y
+       "xvmuldp                53, 49, 36      \n\t"
+       "xvmuldp                54, 50, 36      \n\t"
+       "xvmuldp                55, 51, 36      \n\t"
+
+       "xvmuldp                38, 48, 37      \n\t"   // s * y
+       "xvmuldp                39, 49, 37      \n\t"
+       "xvmuldp                56, 50, 37      \n\t"
+       "xvmuldp                57, 51, 37      \n\t"
+
+       "lxvp            48, 64(%4)       \n\t"   // load y
+       "lxvp            50, 96(%4)      \n\t"
+
+       "xvadddp                40, 40, 38      \n\t"   // c * x + s * y
+       "xvadddp                41, 41, 39      \n\t"   // c * x + s * y
+       "xvadddp                42, 42, 56      \n\t"   // c * x + s * y
+       "xvadddp                43, 43, 57      \n\t"   // c * x + s * y
+
+       "stxvp           40, 0(%3)       \n\t"   // store x
+       "stxvp           42, 32(%3)      \n\t"
+
+       "xvsubdp         52, 52, 44      \n\t"   // c * y - s * x
+       "xvsubdp         53, 53, 45      \n\t"   // c * y - s * x
+       "xvsubdp         54, 54, 46      \n\t"   // c * y - s * x
+       "xvsubdp         55, 55, 47      \n\t"   // c * y - s * x
+
+       "stxvp           52, 0(%4)       \n\t"   // store y
+       "stxvp           54, 32(%4)      \n\t"
+
+       "addi           %3, %3, 64      \n\t"
+       "addi           %4, %4, 64      \n\t"
+
+       "addic.         %2, %2, -8      \n\t"
+       "bgt            one%=           \n"
+
+     "two%=:                           \n\t"
+
+       "xvmuldp                40, 32, 36      \n\t"   // c * x
+       "xvmuldp                41, 33, 36      \n\t"
+       "xvmuldp                42, 34, 36      \n\t"
+       "xvmuldp                43, 35, 36      \n\t"
+
+       "xvmuldp         52, 48, 36      \n\t"   // c * y
+       "xvmuldp         53, 49, 36      \n\t"
+       "xvmuldp         54, 50, 36      \n\t"
+       "xvmuldp         55, 51, 36      \n\t"
+
+       "xvmuldp         44, 32, 37      \n\t"   // s * x
+       "xvmuldp         45, 33, 37      \n\t"
+       "xvmuldp         46, 34, 37      \n\t"
+       "xvmuldp         47, 35, 37      \n\t"
+
+       "xvmuldp         38, 48, 37     \n\t"   // s * y
+       "xvmuldp         39, 49, 37     \n\t"
+       "xvmuldp         56, 50, 37     \n\t"
+       "xvmuldp         57, 51, 37     \n\t"
+
+       "xvadddp         40, 40, 38     \n\t"   // c * x + s * y
+       "xvadddp         41, 41, 39     \n\t"   // c * x + s * y
+       "xvadddp         42, 42, 56     \n\t"   // c * x + s * y
+       "xvadddp         43, 43, 57     \n\t"   // c * x + s * y
+
+       "stxvp           40, 0(%3)       \n\t"   // store x
+       "stxvp           42, 32(%3)      \n\t"
+       "xvsubdp         52, 52, 44      \n\t"   // c * y - s * x
+       "xvsubdp         53, 53, 45      \n\t"   // c * y - s * x
+       "xvsubdp         54, 54, 46      \n\t"   // c * y - s * x
+       "xvsubdp         55, 55, 47      \n\t"   // c * y - s * x
+
+       "stxvp           52, 0(%4)       \n\t"   // store y
+       "stxvp           54, 32(%4)      \n\t"
+
+     "#n=%2 x=%0=%3 y=%1=%4 c=%5 s=%6\n"
+     :
+       "+m" (*x),
+       "+m" (*y),
+       "+r" (n),       // 2
+       "+b" (x),       // 3
+       "+b" (y)        // 4
+     :
+       "d" (c),                // 5 
+       "d" (s)         // 6 
+     :
+       "cr0",
+       "vs32","vs33","vs34","vs35","vs36","vs37","vs38","vs39",
+       "vs40","vs41","vs42","vs43","vs44","vs45","vs46","vs47",
+       "vs48","vs49","vs50","vs51","vs52","vs53","vs54","vs55",
+       "vs56","vs57"
+     );
+}
index a53342f..3e4f93e 100644 (file)
@@ -39,9 +39,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #pragma GCC optimize "O1"
 
-#if defined(POWER8) || defined(POWER9) || defined(POWER10)
 #if defined(__VEC__) || defined(__ALTIVEC__)
+#if defined(POWER8) || defined(POWER9)
 #include "srot_microk_power8.c"
+#elif defined(POWER10)
+#include "srot_microk_power10.c"
 #endif
 #endif
 
@@ -115,6 +117,23 @@ int CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT
        if ( (inc_x == 1) && (inc_y == 1) )
        {
 
+#if defined(POWER10)
+               if ( n >= 16 )
+               {
+                       BLASLONG align = ((32 - ((uintptr_t)y & (uintptr_t)0x1F)) >> 2) & 0x7;
+                       for (i = 0; i < align; i++) {
+                               temp  = c*x[i] + s*y[i] ;
+                               y[i]  = c*y[i] - s*x[i] ;
+                               x[i]  = temp ;
+                       }
+               }
+               BLASLONG n1 = (n-i) & -16;
+               if ( n1 > 0 )
+               {
+                       srot_kernel_16(n1, &x1[i], &y1[i], c, s);
+                       i+=n1;
+               }
+#else
                BLASLONG n1 = n & -16;
                if ( n1 > 0 )
                {
@@ -122,6 +141,7 @@ int CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT
                        i=n1;
                }
 
+#endif
                while(i < n)
                {
                        temp  = c*x[i] + s*y[i] ;
diff --git a/kernel/power/srot_microk_power10.c b/kernel/power/srot_microk_power10.c
new file mode 100644 (file)
index 0000000..c54c307
--- /dev/null
@@ -0,0 +1,151 @@
+/***************************************************************************
+Copyright (c) 2021, 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 srot_kernel_16 (long n, float *x, float *y, float c, float s)
+{
+  __asm__
+    (
+       "xscvdpspn       36, %x5        \n\t"   // load c to all words
+       "xxspltw         36, 36, 0       \n\t"
+
+       "xscvdpspn       37, %x6        \n\t"   // load s to all words
+       "xxspltw         37, 37, 0       \n\t"
+       "lxvp            32, 0(%3)       \n\t"   // load x
+       "lxvp            34, 32(%3)      \n\t"
+       "lxvp            48, 0(%4)       \n\t"   // load y
+       "lxvp            50, 32(%4)      \n\t"
+
+       "addic.         %2, %2, -16     \n\t"
+       "ble            two%=           \n\t"
+
+       ".align 5               \n"
+     "one%=:                           \n\t"
+
+       "xvmulsp                40, 32, 36      \n\t"   // c * x
+       "xvmulsp                41, 33, 36      \n\t"
+       "xvmulsp                42, 34, 36      \n\t"
+       "xvmulsp                43, 35, 36      \n\t"
+
+       "xvmulsp                44, 32, 37      \n\t"   // s * x
+       "xvmulsp                45, 33, 37      \n\t"
+       "xvmulsp                46, 34, 37      \n\t"
+       "xvmulsp                47, 35, 37      \n\t"
+
+       "lxvp            32, 64(%3)       \n\t"   // load x
+       "lxvp            34, 96(%3)      \n\t"
+       "xvmulsp                52, 48, 36      \n\t"   // c * y
+       "xvmulsp                53, 49, 36      \n\t"
+       "xvmulsp                54, 50, 36      \n\t"
+       "xvmulsp                55, 51, 36      \n\t"
+
+       "xvmulsp                38, 48, 37      \n\t"   // s * y
+       "xvmulsp                39, 49, 37      \n\t"
+       "xvmulsp                56, 50, 37      \n\t"
+       "xvmulsp                57, 51, 37      \n\t"
+
+       "lxvp            48, 64(%4)       \n\t"   // load y
+       "lxvp            50, 96(%4)      \n\t"
+
+       "xvaddsp                40, 40, 38      \n\t"   // c * x + s * y
+       "xvaddsp                41, 41, 39      \n\t"   // c * x + s * y
+       "xvaddsp                42, 42, 56      \n\t"   // c * x + s * y
+       "xvaddsp                43, 43, 57      \n\t"   // c * x + s * y
+
+       "stxvp           40, 0(%3)       \n\t"   // store x
+       "stxvp           42, 32(%3)      \n\t"
+
+       "xvsubsp         52, 52, 44      \n\t"   // c * y - s * x
+       "xvsubsp         53, 53, 45      \n\t"   // c * y - s * x
+       "xvsubsp         54, 54, 46      \n\t"   // c * y - s * x
+       "xvsubsp         55, 55, 47      \n\t"   // c * y - s * x
+
+       "stxvp           52, 0(%4)       \n\t"   // store y
+       "stxvp           54, 32(%4)      \n\t"
+
+       "addi           %3, %3, 64      \n\t"
+       "addi           %4, %4, 64      \n\t"
+
+       "addic.         %2, %2, -16     \n\t"
+       "bgt            one%=           \n"
+
+     "two%=:                           \n\t"
+
+       "xvmulsp                40, 32, 36      \n\t"   // c * x
+       "xvmulsp                41, 33, 36      \n\t"
+       "xvmulsp                42, 34, 36      \n\t"
+       "xvmulsp                43, 35, 36      \n\t"
+
+       "xvmulsp         52, 48, 36      \n\t"   // c * y
+       "xvmulsp         53, 49, 36      \n\t"
+       "xvmulsp         54, 50, 36      \n\t"
+       "xvmulsp         55, 51, 36      \n\t"
+
+       "xvmulsp         44, 32, 37      \n\t"   // s * x
+       "xvmulsp         45, 33, 37      \n\t"
+       "xvmulsp         46, 34, 37      \n\t"
+       "xvmulsp         47, 35, 37      \n\t"
+
+       "xvmulsp         38, 48, 37     \n\t"   // s * y
+       "xvmulsp         39, 49, 37     \n\t"
+       "xvmulsp         56, 50, 37     \n\t"
+       "xvmulsp         57, 51, 37     \n\t"
+
+       "xvaddsp         40, 40, 38     \n\t"   // c * x + s * y
+       "xvaddsp         41, 41, 39     \n\t"   // c * x + s * y
+       "xvaddsp         42, 42, 56     \n\t"   // c * x + s * y
+       "xvaddsp         43, 43, 57     \n\t"   // c * x + s * y
+
+       "stxvp           40, 0(%3)       \n\t"   // store x
+       "stxvp           42, 32(%3)      \n\t"
+       "xvsubsp         52, 52, 44      \n\t"   // c * y - s * x
+       "xvsubsp         53, 53, 45      \n\t"   // c * y - s * x
+       "xvsubsp         54, 54, 46      \n\t"   // c * y - s * x
+       "xvsubsp         55, 55, 47      \n\t"   // c * y - s * x
+
+       "stxvp           52, 0(%4)       \n\t"   // store y
+       "stxvp           54, 32(%4)      \n\t"
+
+     "#n=%2 x=%0=%3 y=%1=%4 c=%5 s=%6\n"
+     :
+       "+m" (*x),
+       "+m" (*y),
+       "+r" (n),       // 2
+       "+b" (x),       // 3
+       "+b" (y)        // 4
+     :
+       "f" (c),                // 5 
+       "f" (s)         // 6 
+     :
+       "cr0",
+       "vs32","vs33","vs34","vs35","vs36","vs37","vs38","vs39",
+       "vs40","vs41","vs42","vs43","vs44","vs45","vs46","vs47",
+       "vs48","vs49","vs50","vs51","vs52","vs53","vs54","vs55",
+       "vs56","vs57"
+     );
+}