floraachy
/
OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Optimize s/dasum function for POWER10 

This patch makes use of new POWER10 vector pair instructions for
loads and stores.
This commit is contained in:
Rajalakshmi Srinivasaraghavan 2021-03-05 16:22:36 -06:00
parent 3679781872
commit 41646ed006
4 changed files with 343 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
kernel/power/dasum.c
View File

@ -46,9 +46,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#endif
#if defined(POWER8) || defined(POWER9) || defined(POWER10)
#if defined(__VEC__) || defined(__ALTIVEC__)
#if defined(POWER8) || defined(POWER9)
#include "dasum_microk_power8.c"
#elif defined(POWER10)
#include "dasum_microk_power10.c"
#endif
#endif
@ -110,6 +112,21 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
if ( inc_x == 1 )
{
#if defined(POWER10)
if ( n >= 16 )
{
BLASLONG align = ((32 - ((uintptr_t)x & (uintptr_t)0x1F)) >> 3) & 0x3;
for (i = 0; i < align; i++) {
sumf += ABS(x[i]);
}
}
n1 = (n-i) & -16;
if ( n1 > 0 )
{
sumf += dasum_kernel_16(n1, &x[i]);
i+=n1;
}
#else
n1 = n & -16;
if ( n1 > 0 )
{
@ -117,6 +134,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
sumf = dasum_kernel_16(n1, x);
i=n1;
}
#endif
while(i < n)
{

152
kernel/power/dasum_microk_power10.c Normal file
View File

@ -0,0 +1,152 @@
/***************************************************************************
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 double dasum_kernel_16 (long n, double *x)
{
double sum;
__vector double t0;
__vector double t1;
__vector double t2;
__vector double t3;
__asm__
(
"dcbt 0, %2 \n\t"
"xxlxor 32, 32, 32 \n\t"
"xxlxor 33, 33, 33 \n\t"
"xxlxor 34, 34, 34 \n\t"
"xxlxor 35, 35, 35 \n\t"
"xxlxor 36, 36, 36 \n\t"
"xxlxor 37, 37, 37 \n\t"
"xxlxor 38, 38, 38 \n\t"
"xxlxor 39, 39, 39 \n\t"
"lxvp 40, 0(%2) \n\t"
"lxvp 42, 32(%2) \n\t"
"lxvp 44, 64(%2) \n\t"
"lxvp 46, 96(%2) \n\t"
"addi %2, %2, 128 \n\t"
"addic. %1, %1, -16 \n\t"
"ble two%= \n\t"
".align 5 \n"
"one%=: \n\t"
"xvabsdp 48, 40 \n\t"
"xvabsdp 49, 41 \n\t"
"xvabsdp 50, 42 \n\t"
"xvabsdp 51, 43 \n\t"
"lxvp 40, 0(%2) \n\t"
"xvabsdp %x3, 44 \n\t"
"xvabsdp %x4, 45 \n\t"
"lxvp 42, 32(%2) \n\t"
"xvabsdp %x5, 46 \n\t"
"xvabsdp %x6, 47 \n\t"
"lxvp 44, 64(%2) \n\t"
"xvadddp 32, 32, 48 \n\t"
"xvadddp 33, 33, 49 \n\t"
"lxvp 46, 96(%2) \n\t"
"xvadddp 34, 34, 50 \n\t"
"xvadddp 35, 35, 51 \n\t"
"addi %2, %2, 128 \n\t"
"xvadddp 36, 36, %x3 \n\t"
"xvadddp 37, 37, %x4 \n\t"
"addic. %1, %1, -16 \n\t"
"xvadddp 38, 38, %x5 \n\t"
"xvadddp 39, 39, %x6 \n\t"
"bgt one%= \n"
"two%=: \n\t"
"xvabsdp 48, 40 \n\t"
"xvabsdp 49, 41 \n\t"
"xvabsdp 50, 42 \n\t"
"xvabsdp 51, 43 \n\t"
"xvabsdp %x3, 44 \n\t"
"xvabsdp %x4, 45 \n\t"
"xvabsdp %x5, 46 \n\t"
"xvabsdp %x6, 47 \n\t"
"xvadddp 32, 32, 48 \n\t"
"xvadddp 33, 33, 49 \n\t"
"xvadddp 34, 34, 50 \n\t"
"xvadddp 35, 35, 51 \n\t"
"xvadddp 36, 36, %x3 \n\t"
"xvadddp 37, 37, %x4 \n\t"
"xvadddp 38, 38, %x5 \n\t"
"xvadddp 39, 39, %x6 \n\t"
"xvadddp 32, 32, 33 \n\t"
"xvadddp 34, 34, 35 \n\t"
"xvadddp 36, 36, 37 \n\t"
"xvadddp 38, 38, 39 \n\t"
"xvadddp 32, 32, 34 \n\t"
"xvadddp 36, 36, 38 \n\t"
"xvadddp 32, 32, 36 \n\t"
XXSWAPD_S(33,32)
"xsadddp %x0, 32, 33 \n"
"#n=%1 x=%3=%2 sum=%0\n"
"#t0=%x3 t1=%x4 t2=%x5 t3=%x6"
:
"=d" (sum), // 0
"+r" (n), // 1
"+b" (x), // 2
"=wa" (t0), // 3
"=wa" (t1), // 4
"=wa" (t2), // 5
"=wa" (t3) // 6
:
"m" (*x)
:
"cr0",
"vs32","vs33","vs34","vs35","vs36","vs37","vs38","vs39",
"vs40","vs41","vs42","vs43","vs44","vs45","vs46","vs47",
"vs48","vs49","vs50","vs51"
);
return sum;
}

20
kernel/power/sasum.c
View File

@ -46,9 +46,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#endif
#if defined(POWER8) || defined(POWER9) || defined(POWER10)
#if defined(__VEC__) || defined(__ALTIVEC__)
#if defined(POWER8) || defined(POWER9)
#include "sasum_microk_power8.c"
#elif defined(POWER10)
#include "sasum_microk_power10.c"
#endif
#endif
@ -110,6 +112,21 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
if ( inc_x == 1 )
{
#if defined(POWER10)
if ( n >= 32 )
{
BLASLONG align = ((32 - ((uintptr_t)x & (uintptr_t)0x1F)) >> 2) & 0x7;
for (i = 0; i < align; i++) {
sumf += ABS(x[i]);
}
}
n1 = (n-i) & -32;
if ( n1 > 0 )
{
sumf += sasum_kernel_32(n1, &x[i]);
i+=n1;
}
#else
n1 = n & -32;
if ( n1 > 0 )
{
@ -117,6 +134,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
sumf = sasum_kernel_32(n1, x);
i=n1;
}
#endif
while(i < n)
{

153
kernel/power/sasum_microk_power10.c Normal file
View File

@ -0,0 +1,153 @@
/***************************************************************************
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_32 1
static float sasum_kernel_32 (long n, float *x)
{
float sum;
__vector float t0;
__vector float t1;
__vector float t2;
__vector float t3;
__asm__
(
"dcbt 0, %2 \n\t"
"xxlxor 32, 32, 32 \n\t"
"xxlxor 33, 33, 33 \n\t"
"xxlxor 34, 34, 34 \n\t"
"xxlxor 35, 35, 35 \n\t"
"xxlxor 36, 36, 36 \n\t"
"xxlxor 37, 37, 37 \n\t"
"xxlxor 38, 38, 38 \n\t"
"xxlxor 39, 39, 39 \n\t"
"lxvp 40, 0(%2) \n\t"
"lxvp 42, 32(%2) \n\t"
"lxvp 44, 64(%2) \n\t"
"lxvp 46, 96(%2) \n\t"
"addi %2, %2, 128 \n\t"
"addic. %1, %1, -32 \n\t"
"ble two%= \n\t"
".align 5 \n"
"one%=: \n\t"
"xvabssp 48, 40 \n\t"
"xvabssp 49, 41 \n\t"
"xvabssp 50, 42 \n\t"
"xvabssp 51, 43 \n\t"
"lxvp 40, 0(%2) \n\t"
"xvabssp %x3, 44 \n\t"
"xvabssp %x4, 45 \n\t"
"lxvp 42, 32(%2) \n\t"
"xvabssp %x5, 46 \n\t"
"xvabssp %x6, 47 \n\t"
"lxvp 44, 64(%2) \n\t"
"xvaddsp 32, 32, 48 \n\t"
"xvaddsp 33, 33, 49 \n\t"
"lxvp 46, 96(%2) \n\t"
"xvaddsp 34, 34, 50 \n\t"
"xvaddsp 35, 35, 51 \n\t"
"addi %2, %2, 128 \n\t"
"xvaddsp 36, 36, %x3 \n\t"
"xvaddsp 37, 37, %x4 \n\t"
"addic. %1, %1, -32 \n\t"
"xvaddsp 38, 38, %x5 \n\t"
"xvaddsp 39, 39, %x6 \n\t"
"bgt one%= \n"
"two%=: \n\t"
"xvabssp 48, 40 \n\t"
"xvabssp 49, 41 \n\t"
"xvabssp 50, 42 \n\t"
"xvabssp 51, 43 \n\t"
"xvabssp %x3, 44 \n\t"
"xvabssp %x4, 45 \n\t"
"xvabssp %x5, 46 \n\t"
"xvabssp %x6, 47 \n\t"
"xvaddsp 32, 32, 48 \n\t"
"xvaddsp 33, 33, 49 \n\t"
"xvaddsp 34, 34, 50 \n\t"
"xvaddsp 35, 35, 51 \n\t"
"xvaddsp 36, 36, %x3 \n\t"
"xvaddsp 37, 37, %x4 \n\t"
"xvaddsp 38, 38, %x5 \n\t"
"xvaddsp 39, 39, %x6 \n\t"
"xvaddsp 32, 32, 33 \n\t"
"xvaddsp 34, 34, 35 \n\t"
"xvaddsp 36, 36, 37 \n\t"
"xvaddsp 38, 38, 39 \n\t"
"xvaddsp 32, 32, 34 \n\t"
"xvaddsp 36, 36, 38 \n\t"
"xvaddsp 32, 32, 36 \n\t"
"xxsldwi 33, 32, 32, 2 \n\t"
"xvaddsp 32, 32, 33 \n\t"
"xxsldwi 33, 32, 32, 1 \n\t"
"xvaddsp 32, 32, 33 \n\t"
"xscvspdp %x0, 32 \n"
"#n=%1 x=%3=%2 sum=%0\n"
"#t0=%x3 t1=%x4 t2=%x5 t3=%x6"
:
"=f" (sum), // 0
"+r" (n), // 1
"+b" (x), // 2
"=wa" (t0), // 3
"=wa" (t1), // 4
"=wa" (t2), // 5
"=wa" (t3) // 6
:
"m" (*x)
:
"cr0",
"vs32","vs33","vs34","vs35","vs36","vs37","vs38","vs39",
"vs40","vs41","vs42","vs43","vs44","vs45","vs46","vs47",
"vs48","vs49","vs50","vs51"
);
return sum;
}