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

Add [cz]scal microkernels for SKYLAKEX 

These are as similar to dscal_microk_skylakex-2.c as possible
for consistency.

Note that before this change SKYLAKEX+ uses generic C functions for
cscal/zscal via commit 2271c350 from #2610 (which is masked by
commit 086d87a30). However now #3799 disables FMAs (in turn enabled
by `-march=skylake-avx512`) in the plain C code which fixes excessive
LAPACK test failures more nicely.
This commit is contained in:
Bart Oldeman 2022-11-09 08:28:23 -05:00
parent 4743d80c22
commit 6c1043eb41
5 changed files with 310 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
kernel/x86_64/KERNEL.SKYLAKEX
View File

@ -44,8 +44,5 @@ DGEMM_BETA = dgemm_beta_skylakex.c
CGEMMKERNEL = cgemm_kernel_8x2_skylakex.c
ZGEMMKERNEL = zgemm_kernel_4x2_skylakex.c
CSCALKERNEL = ../arm/zscal.c
ZSCALKERNEL = ../arm/zscal.c
CASUMKERNEL = casum.c
ZASUMKERNEL = zasum.c

4
kernel/x86_64/cscal.c
View File

@ -41,7 +41,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "common.h"
#if defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX) || defined (COOPERLAKE) || defined (SAPPHIRERAPIDS)
#if defined (SKYLAKEX) || defined (COOPERLAKE) || defined (SAPPHIRERAPIDS)
#include "cscal_microk_skylakex-2.c"
#elif defined(HASWELL) || defined(ZEN)
#include "cscal_microk_haswell-2.c"
#elif defined(BULLDOZER) || defined(PILEDRIVER)
#include "cscal_microk_bulldozer-2.c"

152
kernel/x86_64/cscal_microk_skylakex-2.c Normal file
View File

@ -0,0 +1,152 @@
/***************************************************************************
Copyright (c) 2014-2015, 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.
*****************************************************************************/
/* need a new enough GCC for avx512 support */
#if (( defined(__GNUC__) && __GNUC__ > 6 && defined(__AVX2__)) || (defined(__clang__) && __clang_major__ >= 6))
#include <immintrin.h>
#define HAVE_KERNEL_16 1
static void cscal_kernel_16( BLASLONG n, FLOAT *alpha, FLOAT *x)
{
BLASLONG i = 0;
BLASLONG n2 = n + n;
#ifdef __AVX512CD__
/* _mm512_addsub_ps does not exist so we flip signs for odd elements of da_i */
__m512 da_r = _mm512_set1_ps(alpha[0]);
__m512 da_i = _mm512_set1_ps(alpha[1]) * _mm512_set4_ps(1, -1, 1, -1);
for (; i < n2; i += 32) {
__m512 x0 = _mm512_loadu_ps(&x[i + 0]);
__m512 x1 = _mm512_loadu_ps(&x[i + 16]);
__m512 y0 = _mm512_permute_ps(x0, 0xb1);
__m512 y1 = _mm512_permute_ps(x1, 0xb1);
_mm512_storeu_ps(&x[i + 0], _mm512_add_ps(da_r * x0, da_i * y0));
_mm512_storeu_ps(&x[i + 16], _mm512_add_ps(da_r * x1, da_i * y1));
}
#else
__m256 da_r = _mm256_set1_ps(alpha[0]);
__m256 da_i = _mm256_set1_ps(alpha[1]);
for (; i < n2; i += 32) {
__m256 x0 = _mm256_loadu_ps(&x[i + 0]);
__m256 x1 = _mm256_loadu_ps(&x[i + 8]);
__m256 x2 = _mm256_loadu_ps(&x[i + 16]);
__m256 x3 = _mm256_loadu_ps(&x[i + 24]);
__m256 y0 = _mm256_permute_ps(x0, 0xb1);
__m256 y1 = _mm256_permute_ps(x1, 0xb1);
__m256 y2 = _mm256_permute_ps(x2, 0xb1);
__m256 y3 = _mm256_permute_ps(x3, 0xb1);
_mm256_storeu_ps(&x[i + 0], _mm256_addsub_ps(da_r * x0, da_i * y0));
_mm256_storeu_ps(&x[i + 8], _mm256_addsub_ps(da_r * x1, da_i * y1));
_mm256_storeu_ps(&x[i + 16], _mm256_addsub_ps(da_r * x2, da_i * y2));
_mm256_storeu_ps(&x[i + 24], _mm256_addsub_ps(da_r * x3, da_i * y3));
}
#endif
}
static void cscal_kernel_16_zero_r( BLASLONG n, FLOAT *alpha, FLOAT *x)
{
BLASLONG i = 0;
BLASLONG n2 = n + n;
#ifdef __AVX512CD__
__m512 da_i = _mm512_set1_ps(alpha[1]) * _mm512_set4_ps(1, -1, 1, -1);
for (; i < n2; i += 32) {
__m512 y0 = _mm512_permute_ps(_mm512_loadu_ps(&x[i + 0]), 0xb1);
__m512 y1 = _mm512_permute_ps(_mm512_loadu_ps(&x[i + 16]), 0xb1);
_mm512_storeu_ps(&x[i + 0], da_i * y0);
_mm512_storeu_ps(&x[i + 16], da_i * y1);
}
#else
__m256 da_i = _mm256_set1_ps(alpha[1]) * _mm256_set_ps(1, -1, 1, -1, 1, -1, 1, -1);
for (; i < n2; i += 32) {
__m256 y0 = _mm256_permute_ps(_mm256_loadu_ps(&x[i + 0]), 0xb1);
__m256 y1 = _mm256_permute_ps(_mm256_loadu_ps(&x[i + 8]), 0xb1);
__m256 y2 = _mm256_permute_ps(_mm256_loadu_ps(&x[i + 16]), 0xb1);
__m256 y3 = _mm256_permute_ps(_mm256_loadu_ps(&x[i + 24]), 0xb1);
_mm256_storeu_ps(&x[i + 0], da_i * y0);
_mm256_storeu_ps(&x[i + 8], da_i * y1);
_mm256_storeu_ps(&x[i + 16], da_i * y2);
_mm256_storeu_ps(&x[i + 24], da_i * y3);
}
#endif
}
static void cscal_kernel_16_zero_i( BLASLONG n, FLOAT *alpha, FLOAT *x)
{
BLASLONG i = 0;
BLASLONG n2 = n + n;
#ifdef __AVX512CD__
__m512 da_r = _mm512_set1_ps(alpha[0]);
for (; i < n2; i += 32) {
_mm512_storeu_ps(&x[i + 0], da_r * _mm512_loadu_ps(&x[i + 0]));
_mm512_storeu_ps(&x[i + 16], da_r * _mm512_loadu_ps(&x[i + 16]));
}
#else
__m256 da_r = _mm256_set1_ps(alpha[0]);
for (; i < n2; i += 32) {
_mm256_storeu_ps(&x[i + 0], da_r * _mm256_loadu_ps(&x[i + 0]));
_mm256_storeu_ps(&x[i + 8], da_r * _mm256_loadu_ps(&x[i + 8]));
_mm256_storeu_ps(&x[i + 16], da_r * _mm256_loadu_ps(&x[i + 16]));
_mm256_storeu_ps(&x[i + 24], da_r * _mm256_loadu_ps(&x[i + 24]));
}
#endif
}
static void cscal_kernel_16_zero( BLASLONG n, FLOAT *alpha, FLOAT *x)
{
BLASLONG i = 0;
BLASLONG n2 = n + n;
/* question to self: Why is this not just memset() */
#ifdef __AVX512CD__
__m512 zero = _mm512_setzero_ps();
for (; i < n2; i += 32) {
_mm512_storeu_ps(&x[i], zero);
_mm512_storeu_ps(&x[i + 16], zero);
}
#else
__m256 zero = _mm256_setzero_ps();
for (; i < n2; i += 32) {
_mm256_storeu_ps(&x[i + 0], zero);
_mm256_storeu_ps(&x[i + 8], zero);
_mm256_storeu_ps(&x[i + 16], zero);
_mm256_storeu_ps(&x[i + 24], zero);
}
#endif
}
#else
#include "cscal_microk_haswell-2.c"
#endif

4
kernel/x86_64/zscal.c
View File

@ -41,7 +41,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "common.h"
#if defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX) || defined (COOPERLAKE) || defined (SAPPHIRERAPIDS)
#if defined (SKYLAKEX) || defined (COOPERLAKE) || defined (SAPPHIRERAPIDS)
#include "zscal_microk_skylakex-2.c"
#elif defined(HASWELL) || defined(ZEN)
#include "zscal_microk_haswell-2.c"
#elif defined(BULLDOZER) || defined(PILEDRIVER)
#include "zscal_microk_bulldozer-2.c"

152
kernel/x86_64/zscal_microk_skylakex-2.c Normal file
View File

@ -0,0 +1,152 @@
/***************************************************************************
Copyright (c) 2014-2015, 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.
*****************************************************************************/
/* need a new enough GCC for avx512 support */
#if (( defined(__GNUC__) && __GNUC__ > 6 && defined(__AVX2__)) || (defined(__clang__) && __clang_major__ >= 6))
#include <immintrin.h>
#define HAVE_KERNEL_8 1
static void zscal_kernel_8( BLASLONG n, FLOAT *alpha, FLOAT *x)
{
BLASLONG i = 0;
BLASLONG n2 = n + n;
#ifdef __AVX512CD__
/* _mm512_addsub_pd does not exist so we flip signs for odd elements of da_i */
__m512d da_r = _mm512_set1_pd(alpha[0]);
__m512d da_i = _mm512_set1_pd(alpha[1]) * _mm512_set4_pd(1, -1, 1, -1);
for (; i < n2; i += 16) {
__m512d x0 = _mm512_loadu_pd(&x[i + 0]);
__m512d x1 = _mm512_loadu_pd(&x[i + 8]);
__m512d y0 = _mm512_permute_pd(x0, 0x55);
__m512d y1 = _mm512_permute_pd(x1, 0x55);
_mm512_storeu_pd(&x[i + 0], _mm512_add_pd(da_r * x0, da_i * y0));
_mm512_storeu_pd(&x[i + 8], _mm512_add_pd(da_r * x1, da_i * y1));
}
#else
__m256d da_r = _mm256_set1_pd(alpha[0]);
__m256d da_i = _mm256_set1_pd(alpha[1]);
for (; i < n2; i += 16) {
__m256d x0 = _mm256_loadu_pd(&x[i + 0]);
__m256d x1 = _mm256_loadu_pd(&x[i + 4]);
__m256d x2 = _mm256_loadu_pd(&x[i + 8]);
__m256d x3 = _mm256_loadu_pd(&x[i + 12]);
__m256d y0 = _mm256_permute_pd(x0, 0x05);
__m256d y1 = _mm256_permute_pd(x1, 0x05);
__m256d y2 = _mm256_permute_pd(x2, 0x05);
__m256d y3 = _mm256_permute_pd(x3, 0x05);
_mm256_storeu_pd(&x[i + 0], _mm256_addsub_pd(da_r * x0, da_i * y0));
_mm256_storeu_pd(&x[i + 4], _mm256_addsub_pd(da_r * x1, da_i * y1));
_mm256_storeu_pd(&x[i + 8], _mm256_addsub_pd(da_r * x2, da_i * y2));
_mm256_storeu_pd(&x[i + 12], _mm256_addsub_pd(da_r * x3, da_i * y3));
}
#endif
}
static void zscal_kernel_8_zero_r( BLASLONG n, FLOAT *alpha, FLOAT *x)
{
BLASLONG i = 0;
BLASLONG n2 = n + n;
#ifdef __AVX512CD__
__m512d da_i = _mm512_set1_pd(alpha[1]) * _mm512_set4_pd(1, -1, 1, -1);
for (; i < n2; i += 16) {
__m512d y0 = _mm512_permute_pd(_mm512_loadu_pd(&x[i + 0]), 0x55);
__m512d y1 = _mm512_permute_pd(_mm512_loadu_pd(&x[i + 8]), 0x55);
_mm512_storeu_pd(&x[i + 0], da_i * y0);
_mm512_storeu_pd(&x[i + 8], da_i * y1);
}
#else
__m256d da_i = _mm256_set1_pd(alpha[1]) * _mm256_set_pd(1, -1, 1, -1);
for (; i < n2; i += 16) {
__m256d y0 = _mm256_permute_pd(_mm256_loadu_pd(&x[i + 0]), 0x05);
__m256d y1 = _mm256_permute_pd(_mm256_loadu_pd(&x[i + 8]), 0x05);
__m256d y2 = _mm256_permute_pd(_mm256_loadu_pd(&x[i + 16]), 0x05);
__m256d y3 = _mm256_permute_pd(_mm256_loadu_pd(&x[i + 24]), 0x05);
_mm256_storeu_pd(&x[i + 0], da_i * y0);
_mm256_storeu_pd(&x[i + 4], da_i * y1);
_mm256_storeu_pd(&x[i + 8], da_i * y2);
_mm256_storeu_pd(&x[i + 12], da_i * y3);
}
#endif
}
static void zscal_kernel_8_zero_i( BLASLONG n, FLOAT *alpha, FLOAT *x)
{
BLASLONG i = 0;
BLASLONG n2 = n + n;
#ifdef __AVX512CD__
__m512d da_r = _mm512_set1_pd(alpha[0]);
for (; i < n2; i += 16) {
_mm512_storeu_pd(&x[i + 0], da_r * _mm512_loadu_pd(&x[i + 0]));
_mm512_storeu_pd(&x[i + 8], da_r * _mm512_loadu_pd(&x[i + 8]));
}
#else
__m256d da_r = _mm256_set1_pd(alpha[0]);
for (; i < n2; i += 16) {
_mm256_storeu_pd(&x[i + 0], da_r * _mm256_loadu_pd(&x[i + 0]));
_mm256_storeu_pd(&x[i + 4], da_r * _mm256_loadu_pd(&x[i + 4]));
_mm256_storeu_pd(&x[i + 8], da_r * _mm256_loadu_pd(&x[i + 8]));
_mm256_storeu_pd(&x[i + 12], da_r * _mm256_loadu_pd(&x[i + 12]));
}
#endif
}
static void zscal_kernel_8_zero( BLASLONG n, FLOAT *alpha, FLOAT *x)
{
BLASLONG i = 0;
BLASLONG n2 = n + n;
/* question to self: Why is this not just memset() */
#ifdef __AVX512CD__
__m512d zero = _mm512_setzero_pd();
for (; i < n2; i += 16) {
_mm512_storeu_pd(&x[i], zero);
_mm512_storeu_pd(&x[i + 8], zero);
}
#else
__m256d zero = _mm256_setzero_pd();
for (; i < n2; i += 16) {
_mm256_storeu_pd(&x[i + 0], zero);
_mm256_storeu_pd(&x[i + 4], zero);
_mm256_storeu_pd(&x[i + 8], zero);
_mm256_storeu_pd(&x[i + 12], zero);
}
#endif
}
#else
#include "zscal_microk_haswell-2.c"
#endif