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

daxpy_haswell: Change to C+instrinsics + AVX512 to mimic the change to saxpy_haswell 

Use the same transformation as was done to saxpy for daxpy

gives a low double digit performance increase
This commit is contained in:
Arjan van de Ven 2018-08-05 18:29:34 +00:00
parent 7af8a5445d
commit 93aa18b1a8
1 changed files with 40 additions and 45 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

73
kernel/x86_64/daxpy_microk_haswell-2.c
View File

@ -25,54 +25,49 @@ 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. USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/ *****************************************************************************/
#ifndef __AVX512CD__
#pragma GCC target("avx2,fma")
#endif
#ifdef __AVX2__
#include <immintrin.h>
#define HAVE_KERNEL_8 1 #define HAVE_KERNEL_8 1
static void daxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *alpha) __attribute__ ((noinline));
static void daxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha) static void daxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
{ {
BLASLONG i = 0;
__m256d __alpha;
BLASLONG register i = 0; __alpha = _mm256_broadcastsd_pd(_mm_load_sd(alpha));
__asm__ __volatile__ #ifdef __AVX512CD__
( BLASLONG n32;
"vbroadcastsd (%4), %%ymm0 \n\t" // alpha __m512d __alpha5;
__alpha5 = _mm512_broadcastsd_pd(_mm_load_sd(alpha));
".p2align 4 \n\t" n32 = n & ~31;
"1: \n\t"
"vmovups (%3,%0,8), %%ymm12 \n\t" // 4 * y
"vmovups 32(%3,%0,8), %%ymm13 \n\t" // 4 * y
"vmovups 64(%3,%0,8), %%ymm14 \n\t" // 4 * y
"vmovups 96(%3,%0,8), %%ymm15 \n\t" // 4 * y
"vfmadd231pd (%2,%0,8), %%ymm0 , %%ymm12 \n\t" // y += alpha * x
"vfmadd231pd 32(%2,%0,8), %%ymm0 , %%ymm13 \n\t" // y += alpha * x
"vfmadd231pd 64(%2,%0,8), %%ymm0 , %%ymm14 \n\t" // y += alpha * x
"vfmadd231pd 96(%2,%0,8), %%ymm0 , %%ymm15 \n\t" // y += alpha * x
"vmovups %%ymm12, (%3,%0,8) \n\t"
"vmovups %%ymm13, 32(%3,%0,8) \n\t"
"vmovups %%ymm14, 64(%3,%0,8) \n\t"
"vmovups %%ymm15, 96(%3,%0,8) \n\t"
"addq $16, %0 \n\t"
"subq $16, %1 \n\t"
"jnz 1b \n\t"
"vzeroupper \n\t"
:
:
"r" (i), // 0
"r" (n), // 1
"r" (x), // 2
"r" (y), // 3
"r" (alpha) // 4
: "cc",
"%xmm0",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
for (; i < n32; i+= 32) {
_mm512_storeu_pd(&y[i + 0], _mm512_loadu_pd(&y[i + 0]) + __alpha5 * _mm512_loadu_pd(&x[i + 0]));
_mm512_storeu_pd(&y[i + 8], _mm512_loadu_pd(&y[i + 8]) + __alpha5 * _mm512_loadu_pd(&x[i + 8]));
_mm512_storeu_pd(&y[i + 16], _mm512_loadu_pd(&y[i + 16]) + __alpha5 * _mm512_loadu_pd(&x[i + 16]));
_mm512_storeu_pd(&y[i + 24], _mm512_loadu_pd(&y[i + 24]) + __alpha5 * _mm512_loadu_pd(&x[i + 24]));
} }
#endif
for (; i < n; i+= 16) {
_mm256_storeu_pd(&y[i + 0], _mm256_loadu_pd(&y[i + 0]) + __alpha * _mm256_loadu_pd(&x[i + 0]));
_mm256_storeu_pd(&y[i + 4], _mm256_loadu_pd(&y[i + 4]) + __alpha * _mm256_loadu_pd(&x[i + 4]));
_mm256_storeu_pd(&y[i + 8], _mm256_loadu_pd(&y[i + 8]) + __alpha * _mm256_loadu_pd(&x[i + 8]));
_mm256_storeu_pd(&y[i + 12], _mm256_loadu_pd(&y[i + 12]) + __alpha * _mm256_loadu_pd(&x[i + 12]));
}
}
#endif