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

Merge pull request #448 from wernsaar/develop 

Optimized cgemv and zgemv kernels
This commit is contained in:
Zhang Xianyi 2014-09-15 13:12:14 +08:00
parent 1cba8e7b11 b7c9566eea
commit 59e2c20557
40 changed files with 5203 additions and 3687 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
interface/zgemv.c
View File

@ -238,7 +238,7 @@ void CNAME(enum CBLAS_ORDER order,
int nthreads_avail = nthreads_max;
double MNK = (double) m * (double) n;
if ( MNK <= (80.0 * 20.0 * (double) GEMM_MULTITHREAD_THRESHOLD) )
if ( MNK <= ( 256.0 * (double) (GEMM_MULTITHREAD_THRESHOLD * GEMM_MULTITHREAD_THRESHOLD) ))
nthreads_max = 1;
if ( nthreads_max > nthreads_avail )

2
kernel/x86_64/KERNEL.BULLDOZER
View File

@ -14,7 +14,7 @@ SGEMVNKERNEL = sgemv_n_4.c
SGEMVTKERNEL = sgemv_t_4.c
ZGEMVNKERNEL = zgemv_n_dup.S
ZGEMVTKERNEL = zgemv_t.c
ZGEMVTKERNEL = zgemv_t_4.c
DGEMVNKERNEL = dgemv_n_bulldozer.S
DGEMVTKERNEL = dgemv_t_bulldozer.S

8
kernel/x86_64/KERNEL.HASWELL
View File

@ -4,11 +4,11 @@ SGEMVTKERNEL = sgemv_t_4.c
DGEMVNKERNEL = dgemv_n_4.c
DGEMVTKERNEL = dgemv_t_4.c
ZGEMVNKERNEL = zgemv_n.c
ZGEMVTKERNEL = zgemv_t.c
ZGEMVNKERNEL = zgemv_n_4.c
ZGEMVTKERNEL = zgemv_t_4.c
CGEMVNKERNEL = cgemv_n.c
CGEMVTKERNEL = cgemv_t.c
CGEMVNKERNEL = cgemv_n_4.c
CGEMVTKERNEL = cgemv_t_4.c
SGEMMKERNEL = sgemm_kernel_16x4_haswell.S
SGEMMINCOPY = ../generic/gemm_ncopy_16.c

3
kernel/x86_64/KERNEL.PILEDRIVER
View File

@ -2,10 +2,11 @@ SGEMVNKERNEL = sgemv_n_4.c
SGEMVTKERNEL = sgemv_t_4.c
ZGEMVNKERNEL = zgemv_n_dup.S
ZGEMVTKERNEL = zgemv_t.S
ZGEMVTKERNEL = zgemv_t_4.c
DGEMVNKERNEL = dgemv_n_bulldozer.S
DGEMVTKERNEL = dgemv_t_bulldozer.S
DDOTKERNEL = ddot_bulldozer.S
DCOPYKERNEL = dcopy_bulldozer.S

2
kernel/x86_64/KERNEL.SANDYBRIDGE
View File

@ -1,7 +1,7 @@
SGEMVNKERNEL = sgemv_n_4.c
SGEMVTKERNEL = sgemv_t_4.c
ZGEMVNKERNEL = zgemv_n.c
ZGEMVNKERNEL = zgemv_n_4.c
SGEMMKERNEL = sgemm_kernel_16x4_sandy.S

255
kernel/x86_64/cgemv_n.c
View File

@ -1,255 +0,0 @@
/***************************************************************************
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 <stdlib.h>
#include <stdio.h>
#include "common.h"
#if defined(HASWELL)
#include "cgemv_n_microk_haswell-2.c"
#endif
#define NBMAX 2048
#ifndef HAVE_KERNEL_16x4
static void cgemv_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];
for ( i=0; i< 2*n; i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
y[i] += a0[i]*x[0] - a0[i+1] * x[1];
y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
y[i] += a1[i]*x[2] - a1[i+1] * x[3];
y[i+1] += a1[i]*x[3] + a1[i+1] * x[2];
y[i] += a2[i]*x[4] - a2[i+1] * x[5];
y[i+1] += a2[i]*x[5] + a2[i+1] * x[4];
y[i] += a3[i]*x[6] - a3[i+1] * x[7];
y[i+1] += a3[i]*x[7] + a3[i+1] * x[6];
#else
y[i] += a0[i]*x[0] + a0[i+1] * x[1];
y[i+1] += a0[i]*x[1] - a0[i+1] * x[0];
y[i] += a1[i]*x[2] + a1[i+1] * x[3];
y[i+1] += a1[i]*x[3] - a1[i+1] * x[2];
y[i] += a2[i]*x[4] + a2[i+1] * x[5];
y[i+1] += a2[i]*x[5] - a2[i+1] * x[4];
y[i] += a3[i]*x[6] + a3[i+1] * x[7];
y[i+1] += a3[i]*x[7] - a3[i+1] * x[6];
#endif
}
}
#endif
static void cgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
{
BLASLONG i;
FLOAT *a0;
a0 = ap;
for ( i=0; i< 2*n; i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
y[i] += a0[i]*x[0] - a0[i+1] * x[1];
y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
#else
y[i] += a0[i]*x[0] + a0[i+1] * x[1];
y[i+1] += a0[i]*x[1] - a0[i+1] * x[0];
#endif
}
}
static void zero_y(BLASLONG n, FLOAT *dest)
{
BLASLONG i;
for ( i=0; i<2*n; i++ )
{
*dest = 0.0;
dest++;
}
}
static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i)
{
BLASLONG i;
FLOAT temp_r;
FLOAT temp_i;
for ( i=0; i<n; i++ )
{
#if !defined(XCONJ)
temp_r = alpha_r * src[0] - alpha_i * src[1];
temp_i = alpha_r * src[1] + alpha_i * src[0];
#else
temp_r = alpha_r * src[0] + alpha_i * src[1];
temp_i = -alpha_r * src[1] + alpha_i * src[0];
#endif
*dest += temp_r;
*(dest+1) += temp_i;
src+=2;
dest += inc_dest;
}
}
int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r,FLOAT alpha_i, 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 xbuffer[8],*ybuffer;
#if 0
printf("%s %d %d %.16f %.16f %d %d %d\n","zgemv_n",m,n,alpha_r,alpha_i,lda,inc_x,inc_y);
#endif
if ( m < 1 ) return(0);
if ( n < 1 ) return(0);
ybuffer = buffer;
inc_x *= 2;
inc_y *= 2;
lda *= 2;
n1 = n / 4 ;
n2 = n % 4 ;
m1 = m - ( m % 16 );
m2 = (m % NBMAX) - (m % 16) ;
y_ptr = y;
BLASLONG NB = NBMAX;
while ( NB == NBMAX )
{
m1 -= NB;
if ( m1 < 0)
{
if ( m2 == 0 ) break;
NB = m2;
}
a_ptr = a;
x_ptr = x;
zero_y(NB,ybuffer);
for( i = 0; i < n1 ; i++)
{
xbuffer[0] = x_ptr[0];
xbuffer[1] = x_ptr[1];
x_ptr += inc_x;
xbuffer[2] = x_ptr[0];
xbuffer[3] = x_ptr[1];
x_ptr += inc_x;
xbuffer[4] = x_ptr[0];
xbuffer[5] = x_ptr[1];
x_ptr += inc_x;
xbuffer[6] = x_ptr[0];
xbuffer[7] = x_ptr[1];
x_ptr += inc_x;
ap[0] = a_ptr;
ap[1] = a_ptr + lda;
ap[2] = ap[1] + lda;
ap[3] = ap[2] + lda;
cgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
a_ptr += 4 * lda;
}
for( i = 0; i < n2 ; i++)
{
xbuffer[0] = x_ptr[0];
xbuffer[1] = x_ptr[1];
x_ptr += inc_x;
cgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
a_ptr += 1 * lda;
}
add_y(NB,ybuffer,y_ptr,inc_y,alpha_r,alpha_i);
a += 2 * NB;
y_ptr += NB * inc_y;
}
j=0;
while ( j < (m % 16))
{
a_ptr = a;
x_ptr = x;
FLOAT temp_r = 0.0;
FLOAT temp_i = 0.0;
for( i = 0; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
#else
temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
#endif
a_ptr += lda;
x_ptr += inc_x;
}
#if !defined(XCONJ)
y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
#else
y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
#endif
y_ptr += inc_y;
a+=2;
j++;
}
return(0);
}

623
kernel/x86_64/cgemv_n_4.c Normal file
View File

@ -0,0 +1,623 @@
/***************************************************************************
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 <stdlib.h>
#include <stdio.h>
#include "common.h"
#if defined(HASWELL)
#include "cgemv_n_microk_haswell-4.c"
#endif
#define NBMAX 2048
#ifndef HAVE_KERNEL_4x4
static void cgemv_kernel_4x4(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];
for ( i=0; i< 2*n; i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
y[i] += a0[i]*x[0] - a0[i+1] * x[1];
y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
y[i] += a1[i]*x[2] - a1[i+1] * x[3];
y[i+1] += a1[i]*x[3] + a1[i+1] * x[2];
y[i] += a2[i]*x[4] - a2[i+1] * x[5];
y[i+1] += a2[i]*x[5] + a2[i+1] * x[4];
y[i] += a3[i]*x[6] - a3[i+1] * x[7];
y[i+1] += a3[i]*x[7] + a3[i+1] * x[6];
#else
y[i] += a0[i]*x[0] + a0[i+1] * x[1];
y[i+1] += a0[i]*x[1] - a0[i+1] * x[0];
y[i] += a1[i]*x[2] + a1[i+1] * x[3];
y[i+1] += a1[i]*x[3] - a1[i+1] * x[2];
y[i] += a2[i]*x[4] + a2[i+1] * x[5];
y[i+1] += a2[i]*x[5] - a2[i+1] * x[4];
y[i] += a3[i]*x[6] + a3[i+1] * x[7];
y[i+1] += a3[i]*x[7] - a3[i+1] * x[6];
#endif
}
}
#endif
#ifndef HAVE_KERNEL_4x2
static void cgemv_kernel_4x2(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
BLASLONG i;
FLOAT *a0,*a1;
a0 = ap[0];
a1 = ap[1];
for ( i=0; i< 2*n; i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
y[i] += a0[i]*x[0] - a0[i+1] * x[1];
y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
y[i] += a1[i]*x[2] - a1[i+1] * x[3];
y[i+1] += a1[i]*x[3] + a1[i+1] * x[2];
#else
y[i] += a0[i]*x[0] + a0[i+1] * x[1];
y[i+1] += a0[i]*x[1] - a0[i+1] * x[0];
y[i] += a1[i]*x[2] + a1[i+1] * x[3];
y[i+1] += a1[i]*x[3] - a1[i+1] * x[2];
#endif
}
}
#endif
#ifndef HAVE_KERNEL_4x1
static void cgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
{
BLASLONG i;
FLOAT *a0;
a0 = ap;
for ( i=0; i< 2*n; i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
y[i] += a0[i]*x[0] - a0[i+1] * x[1];
y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
#else
y[i] += a0[i]*x[0] + a0[i+1] * x[1];
y[i+1] += a0[i]*x[1] - a0[i+1] * x[0];
#endif
}
}
#endif
#ifndef HAVE_KERNEL_ADDY
static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i) __attribute__ ((noinline));
static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i)
{
BLASLONG i;
if ( inc_dest != 2 )
{
FLOAT temp_r;
FLOAT temp_i;
for ( i=0; i<n; i++ )
{
#if !defined(XCONJ)
temp_r = alpha_r * src[0] - alpha_i * src[1];
temp_i = alpha_r * src[1] + alpha_i * src[0];
#else
temp_r = alpha_r * src[0] + alpha_i * src[1];
temp_i = -alpha_r * src[1] + alpha_i * src[0];
#endif
*dest += temp_r;
*(dest+1) += temp_i;
src+=2;
dest += inc_dest;
}
return;
}
FLOAT temp_r0;
FLOAT temp_i0;
FLOAT temp_r1;
FLOAT temp_i1;
FLOAT temp_r2;
FLOAT temp_i2;
FLOAT temp_r3;
FLOAT temp_i3;
for ( i=0; i<n; i+=4 )
{
#if !defined(XCONJ)
temp_r0 = alpha_r * src[0] - alpha_i * src[1];
temp_i0 = alpha_r * src[1] + alpha_i * src[0];
temp_r1 = alpha_r * src[2] - alpha_i * src[3];
temp_i1 = alpha_r * src[3] + alpha_i * src[2];
temp_r2 = alpha_r * src[4] - alpha_i * src[5];
temp_i2 = alpha_r * src[5] + alpha_i * src[4];
temp_r3 = alpha_r * src[6] - alpha_i * src[7];
temp_i3 = alpha_r * src[7] + alpha_i * src[6];
#else
temp_r0 = alpha_r * src[0] + alpha_i * src[1];
temp_i0 = -alpha_r * src[1] + alpha_i * src[0];
temp_r1 = alpha_r * src[2] + alpha_i * src[3];
temp_i1 = -alpha_r * src[3] + alpha_i * src[2];
temp_r2 = alpha_r * src[4] + alpha_i * src[5];
temp_i2 = -alpha_r * src[5] + alpha_i * src[4];
temp_r3 = alpha_r * src[6] + alpha_i * src[7];
temp_i3 = -alpha_r * src[7] + alpha_i * src[6];
#endif
dest[0] += temp_r0;
dest[1] += temp_i0;
dest[2] += temp_r1;
dest[3] += temp_i1;
dest[4] += temp_r2;
dest[5] += temp_i2;
dest[6] += temp_r3;
dest[7] += temp_i3;
src += 8;
dest += 8;
}
return;
}
#endif
int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r,FLOAT alpha_i, 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 m3;
BLASLONG n2;
BLASLONG lda4;
FLOAT xbuffer[8],*ybuffer;
#if 0
printf("%s %d %d %.16f %.16f %d %d %d\n","zgemv_n",m,n,alpha_r,alpha_i,lda,inc_x,inc_y);
#endif
if ( m < 1 ) return(0);
if ( n < 1 ) return(0);
ybuffer = buffer;
inc_x *= 2;
inc_y *= 2;
lda *= 2;
lda4 = 4 * lda;
n1 = n / 4 ;
n2 = n % 4 ;
m3 = m % 4;
m1 = m - ( m % 4 );
m2 = (m % NBMAX) - (m % 4) ;
y_ptr = y;
BLASLONG NB = NBMAX;
while ( NB == NBMAX )
{
m1 -= NB;
if ( m1 < 0)
{
if ( m2 == 0 ) break;
NB = m2;
}
a_ptr = a;
ap[0] = a_ptr;
ap[1] = a_ptr + lda;
ap[2] = ap[1] + lda;
ap[3] = ap[2] + lda;
x_ptr = x;
//zero_y(NB,ybuffer);
memset(ybuffer,0,NB*8);
if ( inc_x == 2 )
{
for( i = 0; i < n1 ; i++)
{
cgemv_kernel_4x4(NB,ap,x_ptr,ybuffer);
ap[0] += lda4;
ap[1] += lda4;
ap[2] += lda4;
ap[3] += lda4;
a_ptr += lda4;
x_ptr += 8;
}
if ( n2 & 2 )
{
cgemv_kernel_4x2(NB,ap,x_ptr,ybuffer);
x_ptr += 4;
a_ptr += 2 * lda;
}
if ( n2 & 1 )
{
cgemv_kernel_4x1(NB,a_ptr,x_ptr,ybuffer);
x_ptr += 2;
a_ptr += lda;
}
}
else
{
for( i = 0; i < n1 ; i++)
{
xbuffer[0] = x_ptr[0];
xbuffer[1] = x_ptr[1];
x_ptr += inc_x;
xbuffer[2] = x_ptr[0];
xbuffer[3] = x_ptr[1];
x_ptr += inc_x;
xbuffer[4] = x_ptr[0];
xbuffer[5] = x_ptr[1];
x_ptr += inc_x;
xbuffer[6] = x_ptr[0];
xbuffer[7] = x_ptr[1];
x_ptr += inc_x;
cgemv_kernel_4x4(NB,ap,xbuffer,ybuffer);
ap[0] += lda4;
ap[1] += lda4;
ap[2] += lda4;
ap[3] += lda4;
a_ptr += lda4;
}
for( i = 0; i < n2 ; i++)
{
xbuffer[0] = x_ptr[0];
xbuffer[1] = x_ptr[1];
x_ptr += inc_x;
cgemv_kernel_4x1(NB,a_ptr,xbuffer,ybuffer);
a_ptr += 1 * lda;
}
}
add_y(NB,ybuffer,y_ptr,inc_y,alpha_r,alpha_i);
a += 2 * NB;
y_ptr += NB * inc_y;
}
if ( m3 == 0 ) return(0);
if ( m3 == 1 )
{
a_ptr = a;
x_ptr = x;
FLOAT temp_r = 0.0;
FLOAT temp_i = 0.0;
if ( lda == 2 && inc_x == 2 )
{
for( i=0 ; i < (n & -2); i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r += a_ptr[2] * x_ptr[2] - a_ptr[3] * x_ptr[3];
temp_i += a_ptr[2] * x_ptr[3] + a_ptr[3] * x_ptr[2];
#else
temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r += a_ptr[2] * x_ptr[2] + a_ptr[3] * x_ptr[3];
temp_i += a_ptr[2] * x_ptr[3] - a_ptr[3] * x_ptr[2];
#endif
a_ptr += 4;
x_ptr += 4;
}
for( ; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
#else
temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
#endif
a_ptr += 2;
x_ptr += 2;
}
}
else
{
for( i = 0; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
#else
temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
#endif
a_ptr += lda;
x_ptr += inc_x;
}
}
#if !defined(XCONJ)
y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
#else
y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
#endif
return(0);
}
if ( m3 == 2 )
{
a_ptr = a;
x_ptr = x;
FLOAT temp_r0 = 0.0;
FLOAT temp_i0 = 0.0;
FLOAT temp_r1 = 0.0;
FLOAT temp_i1 = 0.0;
if ( lda == 4 && inc_x == 2 )
{
for( i = 0; i < (n & -2); i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
temp_r0 += a_ptr[4] * x_ptr[2] - a_ptr[5] * x_ptr[3];
temp_i0 += a_ptr[4] * x_ptr[3] + a_ptr[5] * x_ptr[2];
temp_r1 += a_ptr[6] * x_ptr[2] - a_ptr[7] * x_ptr[3];
temp_i1 += a_ptr[6] * x_ptr[3] + a_ptr[7] * x_ptr[2];
#else
temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
temp_r0 += a_ptr[4] * x_ptr[2] + a_ptr[5] * x_ptr[3];
temp_i0 += a_ptr[4] * x_ptr[3] - a_ptr[5] * x_ptr[2];
temp_r1 += a_ptr[6] * x_ptr[2] + a_ptr[7] * x_ptr[3];
temp_i1 += a_ptr[6] * x_ptr[3] - a_ptr[7] * x_ptr[2];
#endif
a_ptr += 8;
x_ptr += 4;
}
for( ; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
#else
temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
#endif
a_ptr += 4;
x_ptr += 2;
}
}
else
{
for( i=0 ; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
#else
temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
#endif
a_ptr += lda;
x_ptr += inc_x;
}
}
#if !defined(XCONJ)
y_ptr[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
y_ptr[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r1 - alpha_i * temp_i1;
y_ptr[1] += alpha_r * temp_i1 + alpha_i * temp_r1;
#else
y_ptr[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
y_ptr[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r1 + alpha_i * temp_i1;
y_ptr[1] -= alpha_r * temp_i1 - alpha_i * temp_r1;
#endif
return(0);
}
if ( m3 == 3 )
{
a_ptr = a;
x_ptr = x;
FLOAT temp_r0 = 0.0;
FLOAT temp_i0 = 0.0;
FLOAT temp_r1 = 0.0;
FLOAT temp_i1 = 0.0;
FLOAT temp_r2 = 0.0;
FLOAT temp_i2 = 0.0;
if ( lda == 6 && inc_x == 2 )
{
for( i=0 ; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
temp_r2 += a_ptr[4] * x_ptr[0] - a_ptr[5] * x_ptr[1];
temp_i2 += a_ptr[4] * x_ptr[1] + a_ptr[5] * x_ptr[0];
#else
temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
temp_r2 += a_ptr[4] * x_ptr[0] + a_ptr[5] * x_ptr[1];
temp_i2 += a_ptr[4] * x_ptr[1] - a_ptr[5] * x_ptr[0];
#endif
a_ptr += 6;
x_ptr += 2;
}
}
else
{
for( i = 0; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
temp_r2 += a_ptr[4] * x_ptr[0] - a_ptr[5] * x_ptr[1];
temp_i2 += a_ptr[4] * x_ptr[1] + a_ptr[5] * x_ptr[0];
#else
temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
temp_r2 += a_ptr[4] * x_ptr[0] + a_ptr[5] * x_ptr[1];
temp_i2 += a_ptr[4] * x_ptr[1] - a_ptr[5] * x_ptr[0];
#endif
a_ptr += lda;
x_ptr += inc_x;
}
}
#if !defined(XCONJ)
y_ptr[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
y_ptr[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r1 - alpha_i * temp_i1;
y_ptr[1] += alpha_r * temp_i1 + alpha_i * temp_r1;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r2 - alpha_i * temp_i2;
y_ptr[1] += alpha_r * temp_i2 + alpha_i * temp_r2;
#else
y_ptr[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
y_ptr[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r1 + alpha_i * temp_i1;
y_ptr[1] -= alpha_r * temp_i1 - alpha_i * temp_r1;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r2 + alpha_i * temp_i2;
y_ptr[1] -= alpha_r * temp_i2 - alpha_i * temp_r2;
#endif
return(0);
}
return(0);
}

137
kernel/x86_64/cgemv_n_microk_haswell-2.c
View File

@ -1,137 +0,0 @@
/***************************************************************************
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 cgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void cgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastss (%2), %%ymm0 \n\t" // real part x0
"vbroadcastss 4(%2), %%ymm1 \n\t" // imag part x0
"vbroadcastss 8(%2), %%ymm2 \n\t" // real part x1
"vbroadcastss 12(%2), %%ymm3 \n\t" // imag part x1
"vbroadcastss 16(%2), %%ymm4 \n\t" // real part x2
"vbroadcastss 20(%2), %%ymm5 \n\t" // imag part x2
"vbroadcastss 24(%2), %%ymm6 \n\t" // real part x3
"vbroadcastss 28(%2), %%ymm7 \n\t" // imag part x3
".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 320(%4,%0,4) \n\t"
"vmovups (%4,%0,4), %%ymm8 \n\t" // 4 complex values form a0
"vmovups 32(%4,%0,4), %%ymm9 \n\t" // 4 complex values form a0
"prefetcht0 320(%5,%0,4) \n\t"
"vmovups (%5,%0,4), %%ymm10 \n\t" // 4 complex values form a1
"vmovups 32(%5,%0,4), %%ymm11 \n\t" // 4 complex values form a1
"vmulps %%ymm8 , %%ymm0, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vmulps %%ymm8 , %%ymm1, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vmulps %%ymm9 , %%ymm0, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vmulps %%ymm9 , %%ymm1, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"prefetcht0 320(%6,%0,4) \n\t"
"vmovups (%6,%0,4), %%ymm8 \n\t" // 4 complex values form a2
"vmovups 32(%6,%0,4), %%ymm9 \n\t" // 4 complex values form a2
"vfmadd231ps %%ymm10, %%ymm2, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vfmadd231ps %%ymm10, %%ymm3, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vfmadd231ps %%ymm11, %%ymm2, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vfmadd231ps %%ymm11, %%ymm3, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"prefetcht0 320(%7,%0,4) \n\t"
"vmovups (%7,%0,4), %%ymm10 \n\t" // 4 complex values form a3
"vmovups 32(%7,%0,4), %%ymm11 \n\t" // 4 complex values form a3
"vfmadd231ps %%ymm8 , %%ymm4, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vfmadd231ps %%ymm8 , %%ymm5, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vfmadd231ps %%ymm9 , %%ymm4, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vfmadd231ps %%ymm9 , %%ymm5, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"vfmadd231ps %%ymm10, %%ymm6, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vfmadd231ps %%ymm10, %%ymm7, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vfmadd231ps %%ymm11, %%ymm6, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vfmadd231ps %%ymm11, %%ymm7, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"prefetcht0 320(%3,%0,4) \n\t"
"vmovups (%3,%0,4), %%ymm10 \n\t"
"vmovups 32(%3,%0,4), %%ymm11 \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilps $0xb1 , %%ymm13, %%ymm13 \n\t"
"vpermilps $0xb1 , %%ymm15, %%ymm15 \n\t"
"vaddsubps %%ymm13, %%ymm12, %%ymm8 \n\t"
"vaddsubps %%ymm15, %%ymm14, %%ymm9 \n\t"
#else
"vpermilps $0xb1 , %%ymm12, %%ymm12 \n\t"
"vpermilps $0xb1 , %%ymm14, %%ymm14 \n\t"
"vaddsubps %%ymm12, %%ymm13, %%ymm8 \n\t"
"vaddsubps %%ymm14, %%ymm15, %%ymm9 \n\t"
"vpermilps $0xb1 , %%ymm8 , %%ymm8 \n\t"
"vpermilps $0xb1 , %%ymm9 , %%ymm9 \n\t"
#endif
"vaddps %%ymm8, %%ymm10, %%ymm12 \n\t"
"vaddps %%ymm9, %%ymm11, %%ymm13 \n\t"
"vmovups %%ymm12, (%3,%0,4) \n\t" // 4 complex values to y
"vmovups %%ymm13, 32(%3,%0,4) \n\t"
"addq $16, %0 \n\t"
"subq $8 , %1 \n\t"
"jnz .L01LOOP%= \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",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}

542
kernel/x86_64/cgemv_n_microk_haswell-4.c Normal file
View File

@ -0,0 +1,542 @@
/***************************************************************************
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_4x4 1
static void cgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void cgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
BLASLONG register i = 0;
BLASLONG register n1 = n & -8 ;
BLASLONG register n2 = n & 4 ;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastss (%2), %%ymm0 \n\t" // real part x0
"vbroadcastss 4(%2), %%ymm1 \n\t" // imag part x0
"vbroadcastss 8(%2), %%ymm2 \n\t" // real part x1
"vbroadcastss 12(%2), %%ymm3 \n\t" // imag part x1
"vbroadcastss 16(%2), %%ymm4 \n\t" // real part x2
"vbroadcastss 20(%2), %%ymm5 \n\t" // imag part x2
"vbroadcastss 24(%2), %%ymm6 \n\t" // real part x3
"vbroadcastss 28(%2), %%ymm7 \n\t" // imag part x3
"cmpq $0 , %1 \n\t"
"je .L01END%= \n\t"
".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 320(%4,%0,4) \n\t"
"vmovups (%4,%0,4), %%ymm8 \n\t" // 4 complex values form a0
"vmovups 32(%4,%0,4), %%ymm9 \n\t" // 4 complex values form a0
"prefetcht0 320(%5,%0,4) \n\t"
"vmovups (%5,%0,4), %%ymm10 \n\t" // 4 complex values form a1
"vmovups 32(%5,%0,4), %%ymm11 \n\t" // 4 complex values form a1
"vmulps %%ymm8 , %%ymm0, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vmulps %%ymm8 , %%ymm1, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vmulps %%ymm9 , %%ymm0, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vmulps %%ymm9 , %%ymm1, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"prefetcht0 320(%6,%0,4) \n\t"
"vmovups (%6,%0,4), %%ymm8 \n\t" // 4 complex values form a2
"vmovups 32(%6,%0,4), %%ymm9 \n\t" // 4 complex values form a2
"vfmadd231ps %%ymm10, %%ymm2, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vfmadd231ps %%ymm10, %%ymm3, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vfmadd231ps %%ymm11, %%ymm2, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vfmadd231ps %%ymm11, %%ymm3, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"prefetcht0 320(%7,%0,4) \n\t"
"vmovups (%7,%0,4), %%ymm10 \n\t" // 4 complex values form a3
"vmovups 32(%7,%0,4), %%ymm11 \n\t" // 4 complex values form a3
"vfmadd231ps %%ymm8 , %%ymm4, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vfmadd231ps %%ymm8 , %%ymm5, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vfmadd231ps %%ymm9 , %%ymm4, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vfmadd231ps %%ymm9 , %%ymm5, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"vfmadd231ps %%ymm10, %%ymm6, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vfmadd231ps %%ymm10, %%ymm7, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vfmadd231ps %%ymm11, %%ymm6, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vfmadd231ps %%ymm11, %%ymm7, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"prefetcht0 320(%3,%0,4) \n\t"
"vmovups (%3,%0,4), %%ymm10 \n\t"
"vmovups 32(%3,%0,4), %%ymm11 \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilps $0xb1 , %%ymm13, %%ymm13 \n\t"
"vpermilps $0xb1 , %%ymm15, %%ymm15 \n\t"
"vaddsubps %%ymm13, %%ymm12, %%ymm8 \n\t"
"vaddsubps %%ymm15, %%ymm14, %%ymm9 \n\t"
#else
"vpermilps $0xb1 , %%ymm12, %%ymm12 \n\t"
"vpermilps $0xb1 , %%ymm14, %%ymm14 \n\t"
"vaddsubps %%ymm12, %%ymm13, %%ymm8 \n\t"
"vaddsubps %%ymm14, %%ymm15, %%ymm9 \n\t"
"vpermilps $0xb1 , %%ymm8 , %%ymm8 \n\t"
"vpermilps $0xb1 , %%ymm9 , %%ymm9 \n\t"
#endif
"vaddps %%ymm8, %%ymm10, %%ymm12 \n\t"
"vaddps %%ymm9, %%ymm11, %%ymm13 \n\t"
"vmovups %%ymm12, (%3,%0,4) \n\t" // 4 complex values to y
"vmovups %%ymm13, 32(%3,%0,4) \n\t"
"addq $16, %0 \n\t"
"subq $8 , %1 \n\t"
"jnz .L01LOOP%= \n\t"
".L01END%=: \n\t"
"cmpq $4, %8 \n\t"
"jne .L02END%= \n\t"
"vmovups (%4,%0,4), %%ymm8 \n\t" // 4 complex values form a0
"vmovups (%5,%0,4), %%ymm10 \n\t" // 4 complex values form a1
"vmulps %%ymm8 , %%ymm0, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vmulps %%ymm8 , %%ymm1, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vfmadd231ps %%ymm10, %%ymm2, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vfmadd231ps %%ymm10, %%ymm3, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vmovups (%6,%0,4), %%ymm8 \n\t" // 4 complex values form a2
"vmovups (%7,%0,4), %%ymm10 \n\t" // 4 complex values form a3
"vfmadd231ps %%ymm8 , %%ymm4, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vfmadd231ps %%ymm8 , %%ymm5, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vfmadd231ps %%ymm10, %%ymm6, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vfmadd231ps %%ymm10, %%ymm7, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vmovups (%3,%0,4), %%ymm10 \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilps $0xb1 , %%ymm13, %%ymm13 \n\t"
"vaddsubps %%ymm13, %%ymm12, %%ymm8 \n\t"
#else
"vpermilps $0xb1 , %%ymm12, %%ymm12 \n\t"
"vaddsubps %%ymm12, %%ymm13, %%ymm8 \n\t"
"vpermilps $0xb1 , %%ymm8 , %%ymm8 \n\t"
#endif
"vaddps %%ymm8, %%ymm10, %%ymm12 \n\t"
"vmovups %%ymm12, (%3,%0,4) \n\t" // 4 complex values to y
".L02END%=: \n\t"
"vzeroupper \n\t"
:
:
"r" (i), // 0
"r" (n1), // 1
"r" (x), // 2
"r" (y), // 3
"r" (ap[0]), // 4
"r" (ap[1]), // 5
"r" (ap[2]), // 6
"r" (ap[3]), // 7
"r" (n2) // 8
: "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
#define HAVE_KERNEL_4x2 1
static void cgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void cgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
BLASLONG register i = 0;
BLASLONG register n1 = n & -8 ;
BLASLONG register n2 = n & 4 ;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastss (%2), %%ymm0 \n\t" // real part x0
"vbroadcastss 4(%2), %%ymm1 \n\t" // imag part x0
"vbroadcastss 8(%2), %%ymm2 \n\t" // real part x1
"vbroadcastss 12(%2), %%ymm3 \n\t" // imag part x1
"cmpq $0 , %1 \n\t"
"je .L01END%= \n\t"
".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 320(%4,%0,4) \n\t"
"vmovups (%4,%0,4), %%ymm8 \n\t" // 4 complex values form a0
"vmovups 32(%4,%0,4), %%ymm9 \n\t" // 4 complex values form a0
"prefetcht0 320(%5,%0,4) \n\t"
"vmovups (%5,%0,4), %%ymm10 \n\t" // 4 complex values form a1
"vmovups 32(%5,%0,4), %%ymm11 \n\t" // 4 complex values form a1
"vmulps %%ymm8 , %%ymm0, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vmulps %%ymm8 , %%ymm1, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vmulps %%ymm9 , %%ymm0, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vmulps %%ymm9 , %%ymm1, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"vfmadd231ps %%ymm10, %%ymm2, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vfmadd231ps %%ymm10, %%ymm3, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vfmadd231ps %%ymm11, %%ymm2, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vfmadd231ps %%ymm11, %%ymm3, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"prefetcht0 320(%3,%0,4) \n\t"
"vmovups (%3,%0,4), %%ymm10 \n\t"
"vmovups 32(%3,%0,4), %%ymm11 \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilps $0xb1 , %%ymm13, %%ymm13 \n\t"
"vpermilps $0xb1 , %%ymm15, %%ymm15 \n\t"
"vaddsubps %%ymm13, %%ymm12, %%ymm8 \n\t"
"vaddsubps %%ymm15, %%ymm14, %%ymm9 \n\t"
#else
"vpermilps $0xb1 , %%ymm12, %%ymm12 \n\t"
"vpermilps $0xb1 , %%ymm14, %%ymm14 \n\t"
"vaddsubps %%ymm12, %%ymm13, %%ymm8 \n\t"
"vaddsubps %%ymm14, %%ymm15, %%ymm9 \n\t"
"vpermilps $0xb1 , %%ymm8 , %%ymm8 \n\t"
"vpermilps $0xb1 , %%ymm9 , %%ymm9 \n\t"
#endif
"vaddps %%ymm8, %%ymm10, %%ymm12 \n\t"
"vaddps %%ymm9, %%ymm11, %%ymm13 \n\t"
"vmovups %%ymm12, (%3,%0,4) \n\t" // 4 complex values to y
"vmovups %%ymm13, 32(%3,%0,4) \n\t"
"addq $16, %0 \n\t"
"subq $8 , %1 \n\t"
"jnz .L01LOOP%= \n\t"
".L01END%=: \n\t"
"cmpq $4, %6 \n\t"
"jne .L02END%= \n\t"
"vmovups (%4,%0,4), %%ymm8 \n\t" // 4 complex values form a0
"vmovups (%5,%0,4), %%ymm10 \n\t" // 4 complex values form a1
"vmulps %%ymm8 , %%ymm0, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vmulps %%ymm8 , %%ymm1, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vfmadd231ps %%ymm10, %%ymm2, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vfmadd231ps %%ymm10, %%ymm3, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vmovups (%3,%0,4), %%ymm10 \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilps $0xb1 , %%ymm13, %%ymm13 \n\t"
"vaddsubps %%ymm13, %%ymm12, %%ymm8 \n\t"
#else
"vpermilps $0xb1 , %%ymm12, %%ymm12 \n\t"
"vaddsubps %%ymm12, %%ymm13, %%ymm8 \n\t"
"vpermilps $0xb1 , %%ymm8 , %%ymm8 \n\t"
#endif
"vaddps %%ymm8, %%ymm10, %%ymm12 \n\t"
"vmovups %%ymm12, (%3,%0,4) \n\t" // 4 complex values to y
".L02END%=: \n\t"
"vzeroupper \n\t"
:
:
"r" (i), // 0
"r" (n1), // 1
"r" (x), // 2
"r" (y), // 3
"r" (ap[0]), // 4
"r" (ap[1]), // 5
"r" (n2) // 6
: "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
#define HAVE_KERNEL_4x1 1
static void cgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void cgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
{
BLASLONG register i = 0;
BLASLONG register n1 = n & -8 ;
BLASLONG register n2 = n & 4 ;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastss (%2), %%ymm0 \n\t" // real part x0
"vbroadcastss 4(%2), %%ymm1 \n\t" // imag part x0
"cmpq $0 , %1 \n\t"
"je .L01END%= \n\t"
".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 320(%4,%0,4) \n\t"
"vmovups (%4,%0,4), %%ymm8 \n\t" // 4 complex values form a0
"vmovups 32(%4,%0,4), %%ymm9 \n\t" // 4 complex values form a0
"vmulps %%ymm8 , %%ymm0, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vmulps %%ymm8 , %%ymm1, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vmulps %%ymm9 , %%ymm0, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vmulps %%ymm9 , %%ymm1, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"prefetcht0 320(%3,%0,4) \n\t"
"vmovups (%3,%0,4), %%ymm10 \n\t"
"vmovups 32(%3,%0,4), %%ymm11 \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilps $0xb1 , %%ymm13, %%ymm13 \n\t"
"vpermilps $0xb1 , %%ymm15, %%ymm15 \n\t"
"vaddsubps %%ymm13, %%ymm12, %%ymm8 \n\t"
"vaddsubps %%ymm15, %%ymm14, %%ymm9 \n\t"
#else
"vpermilps $0xb1 , %%ymm12, %%ymm12 \n\t"
"vpermilps $0xb1 , %%ymm14, %%ymm14 \n\t"
"vaddsubps %%ymm12, %%ymm13, %%ymm8 \n\t"
"vaddsubps %%ymm14, %%ymm15, %%ymm9 \n\t"
"vpermilps $0xb1 , %%ymm8 , %%ymm8 \n\t"
"vpermilps $0xb1 , %%ymm9 , %%ymm9 \n\t"
#endif
"addq $16, %0 \n\t"
"vaddps %%ymm8, %%ymm10, %%ymm12 \n\t"
"vaddps %%ymm9, %%ymm11, %%ymm13 \n\t"
"subq $8 , %1 \n\t"
"vmovups %%ymm12,-64(%3,%0,4) \n\t" // 4 complex values to y
"vmovups %%ymm13,-32(%3,%0,4) \n\t"
"jnz .L01LOOP%= \n\t"
".L01END%=: \n\t"
"cmpq $4, %5 \n\t"
"jne .L02END%= \n\t"
"vmovups (%4,%0,4), %%ymm8 \n\t" // 4 complex values form a0
"vmulps %%ymm8 , %%ymm0, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vmulps %%ymm8 , %%ymm1, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vmovups (%3,%0,4), %%ymm10 \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilps $0xb1 , %%ymm13, %%ymm13 \n\t"
"vaddsubps %%ymm13, %%ymm12, %%ymm8 \n\t"
#else
"vpermilps $0xb1 , %%ymm12, %%ymm12 \n\t"
"vaddsubps %%ymm12, %%ymm13, %%ymm8 \n\t"
"vpermilps $0xb1 , %%ymm8 , %%ymm8 \n\t"
#endif
"vaddps %%ymm8, %%ymm10, %%ymm12 \n\t"
"vmovups %%ymm12, (%3,%0,4) \n\t" // 4 complex values to y
".L02END%=: \n\t"
"vzeroupper \n\t"
:
:
"r" (i), // 0
"r" (n1), // 1
"r" (x), // 2
"r" (y), // 3
"r" (ap), // 4
"r" (n2) // 5
: "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
#define HAVE_KERNEL_ADDY 1
static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i) __attribute__ ((noinline));
static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i)
{
BLASLONG i;
if ( inc_dest != 2 )
{
FLOAT temp_r;
FLOAT temp_i;
for ( i=0; i<n; i++ )
{
#if !defined(XCONJ)
temp_r = alpha_r * src[0] - alpha_i * src[1];
temp_i = alpha_r * src[1] + alpha_i * src[0];
#else
temp_r = alpha_r * src[0] + alpha_i * src[1];
temp_i = -alpha_r * src[1] + alpha_i * src[0];
#endif
*dest += temp_r;
*(dest+1) += temp_i;
src+=2;
dest += inc_dest;
}
return;
}
i=0;
BLASLONG register n1 = n & -8 ;
BLASLONG register n2 = n & 4 ;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastss (%4), %%ymm0 \n\t" // alpha_r
"vbroadcastss (%5), %%ymm1 \n\t" // alpha_i
"cmpq $0 , %1 \n\t"
"je .L01END%= \n\t"
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vmovups (%2,%0,4), %%ymm8 \n\t" // 4 complex values from src
"vmovups 32(%2,%0,4), %%ymm9 \n\t"
"vmulps %%ymm8 , %%ymm0, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vmulps %%ymm8 , %%ymm1, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vmulps %%ymm9 , %%ymm0, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vmulps %%ymm9 , %%ymm1, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"vmovups (%3,%0,4), %%ymm10 \n\t" // 4 complex values from dest
"vmovups 32(%3,%0,4), %%ymm11 \n\t"
#if !defined(XCONJ)
"vpermilps $0xb1 , %%ymm13, %%ymm13 \n\t"
"vpermilps $0xb1 , %%ymm15, %%ymm15 \n\t"
"vaddsubps %%ymm13, %%ymm12, %%ymm8 \n\t"
"vaddsubps %%ymm15, %%ymm14, %%ymm9 \n\t"
#else
"vpermilps $0xb1 , %%ymm12, %%ymm12 \n\t"
"vpermilps $0xb1 , %%ymm14, %%ymm14 \n\t"
"vaddsubps %%ymm12, %%ymm13, %%ymm8 \n\t"
"vaddsubps %%ymm14, %%ymm15, %%ymm9 \n\t"
"vpermilps $0xb1 , %%ymm8 , %%ymm8 \n\t"
"vpermilps $0xb1 , %%ymm9 , %%ymm9 \n\t"
#endif
"addq $16, %0 \n\t"
"vaddps %%ymm8, %%ymm10, %%ymm12 \n\t"
"vaddps %%ymm9, %%ymm11, %%ymm13 \n\t"
"subq $8 , %1 \n\t"
"vmovups %%ymm12,-64(%3,%0,4) \n\t" // 4 complex values to y
"vmovups %%ymm13,-32(%3,%0,4) \n\t"
"jnz .L01LOOP%= \n\t"
".L01END%=: \n\t"
"cmpq $4, %6 \n\t"
"jne .L02END%= \n\t"
"vmovups (%2,%0,4), %%ymm8 \n\t" // 4 complex values src
"vmulps %%ymm8 , %%ymm0, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vmulps %%ymm8 , %%ymm1, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vmovups (%3,%0,4), %%ymm10 \n\t"
#if !defined(XCONJ)
"vpermilps $0xb1 , %%ymm13, %%ymm13 \n\t"
"vaddsubps %%ymm13, %%ymm12, %%ymm8 \n\t"
#else
"vpermilps $0xb1 , %%ymm12, %%ymm12 \n\t"
"vaddsubps %%ymm12, %%ymm13, %%ymm8 \n\t"
"vpermilps $0xb1 , %%ymm8 , %%ymm8 \n\t"
#endif
"vaddps %%ymm8, %%ymm10, %%ymm12 \n\t"
"vmovups %%ymm12, (%3,%0,4) \n\t" // 4 complex values to y
".L02END%=: \n\t"
"vzeroupper \n\t"
:
:
"r" (i), // 0
"r" (n1), // 1
"r" (src), // 2
"r" (dest), // 3
"r" (&alpha_r), // 4
"r" (&alpha_i), // 5
"r" (n2) // 6
: "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
return;
}

265
kernel/x86_64/cgemv_t.c
View File

@ -1,265 +0,0 @@
/***************************************************************************
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 "cgemv_t_microk_haswell-2.c"
#endif
#define NBMAX 2048
#ifndef HAVE_KERNEL_16x4
static void cgemv_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 temp_r0 = 0.0;
FLOAT temp_r1 = 0.0;
FLOAT temp_r2 = 0.0;
FLOAT temp_r3 = 0.0;
FLOAT temp_i0 = 0.0;
FLOAT temp_i1 = 0.0;
FLOAT temp_i2 = 0.0;
FLOAT temp_i3 = 0.0;
for ( i=0; i< 2*n; i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a0[i]*x[i] - a0[i+1]*x[i+1];
temp_i0 += a0[i]*x[i+1] + a0[i+1]*x[i];
temp_r1 += a1[i]*x[i] - a1[i+1]*x[i+1];
temp_i1 += a1[i]*x[i+1] + a1[i+1]*x[i];
temp_r2 += a2[i]*x[i] - a2[i+1]*x[i+1];
temp_i2 += a2[i]*x[i+1] + a2[i+1]*x[i];
temp_r3 += a3[i]*x[i] - a3[i+1]*x[i+1];
temp_i3 += a3[i]*x[i+1] + a3[i+1]*x[i];
#else
temp_r0 += a0[i]*x[i] + a0[i+1]*x[i+1];
temp_i0 += a0[i]*x[i+1] - a0[i+1]*x[i];
temp_r1 += a1[i]*x[i] + a1[i+1]*x[i+1];
temp_i1 += a1[i]*x[i+1] - a1[i+1]*x[i];
temp_r2 += a2[i]*x[i] + a2[i+1]*x[i+1];
temp_i2 += a2[i]*x[i+1] - a2[i+1]*x[i];
temp_r3 += a3[i]*x[i] + a3[i+1]*x[i+1];
temp_i3 += a3[i]*x[i+1] - a3[i+1]*x[i];
#endif
}
y[0] = temp_r0;
y[1] = temp_i0;
y[2] = temp_r1;
y[3] = temp_i1;
y[4] = temp_r2;
y[5] = temp_i2;
y[6] = temp_r3;
y[7] = temp_i3;
}
#endif
static void cgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
{
BLASLONG i;
FLOAT *a0;
a0 = ap;
FLOAT temp_r = 0.0;
FLOAT temp_i = 0.0;
for ( i=0; i< 2*n; i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r += a0[i]*x[i] - a0[i+1]*x[i+1];
temp_i += a0[i]*x[i+1] + a0[i+1]*x[i];
#else
temp_r += a0[i]*x[i] + a0[i+1]*x[i+1];
temp_i += a0[i]*x[i+1] - a0[i+1]*x[i];
#endif
}
*y = temp_r;
*(y+1) = temp_i;
}
static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
{
BLASLONG i;
for ( i=0; i<n; i++ )
{
*dest = *src;
*(dest+1) = *(src+1);
dest+=2;
src += inc_src;
}
}
int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i, 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[8];
BLASLONG n1;
BLASLONG m1;
BLASLONG m2;
BLASLONG n2;
FLOAT ybuffer[8],*xbuffer;
inc_x *= 2;
inc_y *= 2;
lda *= 2;
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;
cgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
a_ptr += 4 * lda;
#if !defined(XCONJ)
y_ptr[0] += alpha_r * ybuffer[0] - alpha_i * ybuffer[1];
y_ptr[1] += alpha_r * ybuffer[1] + alpha_i * ybuffer[0];
y_ptr += inc_y;
y_ptr[0] += alpha_r * ybuffer[2] - alpha_i * ybuffer[3];
y_ptr[1] += alpha_r * ybuffer[3] + alpha_i * ybuffer[2];
y_ptr += inc_y;
y_ptr[0] += alpha_r * ybuffer[4] - alpha_i * ybuffer[5];
y_ptr[1] += alpha_r * ybuffer[5] + alpha_i * ybuffer[4];
y_ptr += inc_y;
y_ptr[0] += alpha_r * ybuffer[6] - alpha_i * ybuffer[7];
y_ptr[1] += alpha_r * ybuffer[7] + alpha_i * ybuffer[6];
y_ptr += inc_y;
#else
y_ptr[0] += alpha_r * ybuffer[0] + alpha_i * ybuffer[1];
y_ptr[1] -= alpha_r * ybuffer[1] - alpha_i * ybuffer[0];
y_ptr += inc_y;
y_ptr[0] += alpha_r * ybuffer[2] + alpha_i * ybuffer[3];
y_ptr[1] -= alpha_r * ybuffer[3] - alpha_i * ybuffer[2];
y_ptr += inc_y;
y_ptr[0] += alpha_r * ybuffer[4] + alpha_i * ybuffer[5];
y_ptr[1] -= alpha_r * ybuffer[5] - alpha_i * ybuffer[4];
y_ptr += inc_y;
y_ptr[0] += alpha_r * ybuffer[6] + alpha_i * ybuffer[7];
y_ptr[1] -= alpha_r * ybuffer[7] - alpha_i * ybuffer[6];
y_ptr += inc_y;
#endif
}
for( i = 0; i < n2 ; i++)
{
cgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
a_ptr += 1 * lda;
#if !defined(XCONJ)
y_ptr[0] += alpha_r * ybuffer[0] - alpha_i * ybuffer[1];
y_ptr[1] += alpha_r * ybuffer[1] + alpha_i * ybuffer[0];
y_ptr += inc_y;
#else
y_ptr[0] += alpha_r * ybuffer[0] + alpha_i * ybuffer[1];
y_ptr[1] -= alpha_r * ybuffer[1] - alpha_i * ybuffer[0];
y_ptr += inc_y;
#endif
}
a += 2* NB;
x += NB * inc_x;
}
BLASLONG m3 = m % 16;
if ( m3 == 0 ) return(0);
x_ptr = x;
copy_x(m3,x_ptr,xbuffer,inc_x);
j=0;
a_ptr = a;
y_ptr = y;
while ( j < n)
{
FLOAT temp_r = 0.0;
FLOAT temp_i = 0.0;
for( i = 0; i < m3*2; i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r += a_ptr[i] * xbuffer[i] - a_ptr[i+1] * xbuffer[i+1];
temp_i += a_ptr[i] * xbuffer[i+1] + a_ptr[i+1] * xbuffer[i];
#else
temp_r += a_ptr[i] * xbuffer[i] + a_ptr[i+1] * xbuffer[i+1];
temp_i += a_ptr[i] * xbuffer[i+1] - a_ptr[i+1] * xbuffer[i];
#endif
}
a_ptr += lda;
#if !defined(XCONJ)
y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
#else
y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
#endif
y_ptr += inc_y;
j++;
}
return(0);
}

579
kernel/x86_64/cgemv_t_4.c Normal file
View File

@ -0,0 +1,579 @@
/***************************************************************************
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 "cgemv_t_microk_haswell-4.c"
#endif
#define NBMAX 2048
#ifndef HAVE_KERNEL_4x4
static void cgemv_kernel_4x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{
BLASLONG i;
FLOAT *a0,*a1,*a2,*a3;
a0 = ap[0];
a1 = ap[1];
a2 = ap[2];
a3 = ap[3];
FLOAT alpha_r = alpha[0];
FLOAT alpha_i = alpha[1];
FLOAT temp_r0 = 0.0;
FLOAT temp_r1 = 0.0;
FLOAT temp_r2 = 0.0;
FLOAT temp_r3 = 0.0;
FLOAT temp_i0 = 0.0;
FLOAT temp_i1 = 0.0;
FLOAT temp_i2 = 0.0;
FLOAT temp_i3 = 0.0;
for ( i=0; i< 2*n; i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a0[i]*x[i] - a0[i+1]*x[i+1];
temp_i0 += a0[i]*x[i+1] + a0[i+1]*x[i];
temp_r1 += a1[i]*x[i] - a1[i+1]*x[i+1];
temp_i1 += a1[i]*x[i+1] + a1[i+1]*x[i];
temp_r2 += a2[i]*x[i] - a2[i+1]*x[i+1];
temp_i2 += a2[i]*x[i+1] + a2[i+1]*x[i];
temp_r3 += a3[i]*x[i] - a3[i+1]*x[i+1];
temp_i3 += a3[i]*x[i+1] + a3[i+1]*x[i];
#else
temp_r0 += a0[i]*x[i] + a0[i+1]*x[i+1];
temp_i0 += a0[i]*x[i+1] - a0[i+1]*x[i];
temp_r1 += a1[i]*x[i] + a1[i+1]*x[i+1];
temp_i1 += a1[i]*x[i+1] - a1[i+1]*x[i];
temp_r2 += a2[i]*x[i] + a2[i+1]*x[i+1];
temp_i2 += a2[i]*x[i+1] - a2[i+1]*x[i];
temp_r3 += a3[i]*x[i] + a3[i+1]*x[i+1];
temp_i3 += a3[i]*x[i+1] - a3[i+1]*x[i];
#endif
}
#if !defined(XCONJ)
y[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
y[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
y[2] += alpha_r * temp_r1 - alpha_i * temp_i1;
y[3] += alpha_r * temp_i1 + alpha_i * temp_r1;
y[4] += alpha_r * temp_r2 - alpha_i * temp_i2;
y[5] += alpha_r * temp_i2 + alpha_i * temp_r2;
y[6] += alpha_r * temp_r3 - alpha_i * temp_i3;
y[7] += alpha_r * temp_i3 + alpha_i * temp_r3;
#else
y[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
y[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
y[2] += alpha_r * temp_r1 + alpha_i * temp_i1;
y[3] -= alpha_r * temp_i1 - alpha_i * temp_r1;
y[4] += alpha_r * temp_r2 + alpha_i * temp_i2;
y[5] -= alpha_r * temp_i2 - alpha_i * temp_r2;
y[6] += alpha_r * temp_r3 + alpha_i * temp_i3;
y[7] -= alpha_r * temp_i3 - alpha_i * temp_r3;
#endif
}
#endif
#ifndef HAVE_KERNEL_4x2
static void cgemv_kernel_4x2(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{
BLASLONG i;
FLOAT *a0,*a1;
a0 = ap[0];
a1 = ap[1];
FLOAT alpha_r = alpha[0];
FLOAT alpha_i = alpha[1];
FLOAT temp_r0 = 0.0;
FLOAT temp_r1 = 0.0;
FLOAT temp_i0 = 0.0;
FLOAT temp_i1 = 0.0;
for ( i=0; i< 2*n; i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a0[i]*x[i] - a0[i+1]*x[i+1];
temp_i0 += a0[i]*x[i+1] + a0[i+1]*x[i];
temp_r1 += a1[i]*x[i] - a1[i+1]*x[i+1];
temp_i1 += a1[i]*x[i+1] + a1[i+1]*x[i];
#else
temp_r0 += a0[i]*x[i] + a0[i+1]*x[i+1];
temp_i0 += a0[i]*x[i+1] - a0[i+1]*x[i];
temp_r1 += a1[i]*x[i] + a1[i+1]*x[i+1];
temp_i1 += a1[i]*x[i+1] - a1[i+1]*x[i];
#endif
}
#if !defined(XCONJ)
y[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
y[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
y[2] += alpha_r * temp_r1 - alpha_i * temp_i1;
y[3] += alpha_r * temp_i1 + alpha_i * temp_r1;
#else
y[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
y[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
y[2] += alpha_r * temp_r1 + alpha_i * temp_i1;
y[3] -= alpha_r * temp_i1 - alpha_i * temp_r1;
#endif
}
#endif
#ifndef HAVE_KERNEL_4x1
static void cgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{
BLASLONG i;
FLOAT *a0;
a0 = ap;
FLOAT alpha_r = alpha[0];
FLOAT alpha_i = alpha[1];
FLOAT temp_r0 = 0.0;
FLOAT temp_i0 = 0.0;
for ( i=0; i< 2*n; i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a0[i]*x[i] - a0[i+1]*x[i+1];
temp_i0 += a0[i]*x[i+1] + a0[i+1]*x[i];
#else
temp_r0 += a0[i]*x[i] + a0[i+1]*x[i+1];
temp_i0 += a0[i]*x[i+1] - a0[i+1]*x[i];
#endif
}
#if !defined(XCONJ)
y[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
y[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
#else
y[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
y[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
#endif
}
#endif
static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
{
BLASLONG i;
for ( i=0; i<n; i++ )
{
*dest = *src;
*(dest+1) = *(src+1);
dest+=2;
src += inc_src;
}
}
int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i, 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[8];
BLASLONG n1;
BLASLONG m1;
BLASLONG m2;
BLASLONG m3;
BLASLONG n2;
BLASLONG lda4;
FLOAT ybuffer[8],*xbuffer;
FLOAT alpha[2];
if ( m < 1 ) return(0);
if ( n < 1 ) return(0);
inc_x <<= 1;
inc_y <<= 1;
lda <<= 1;
lda4 = lda << 2;
xbuffer = buffer;
n1 = n >> 2 ;
n2 = n & 3 ;
m3 = m & 3 ;
m1 = m - m3;
m2 = (m & (NBMAX-1)) - m3 ;
alpha[0] = alpha_r;
alpha[1] = alpha_i;
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;
ap[0] = a_ptr;
ap[1] = a_ptr + lda;
ap[2] = ap[1] + lda;
ap[3] = ap[2] + lda;
if ( inc_x != 2 )
copy_x(NB,x_ptr,xbuffer,inc_x);
else
xbuffer = x_ptr;
if ( inc_y == 2 )
{
for( i = 0; i < n1 ; i++)
{
cgemv_kernel_4x4(NB,ap,xbuffer,y_ptr,alpha);
ap[0] += lda4;
ap[1] += lda4;
ap[2] += lda4;
ap[3] += lda4;
a_ptr += lda4;
y_ptr += 8;
}
if ( n2 & 2 )
{
cgemv_kernel_4x2(NB,ap,xbuffer,y_ptr,alpha);
a_ptr += lda * 2;
y_ptr += 4;
}
if ( n2 & 1 )
{
cgemv_kernel_4x1(NB,a_ptr,xbuffer,y_ptr,alpha);
a_ptr += lda;
y_ptr += 2;
}
}
else
{
for( i = 0; i < n1 ; i++)
{
memset(ybuffer,0,32);
cgemv_kernel_4x4(NB,ap,xbuffer,ybuffer,alpha);
ap[0] += lda4;
ap[1] += lda4;
ap[2] += lda4;
ap[3] += lda4;
a_ptr += lda4;
y_ptr[0] += ybuffer[0];
y_ptr[1] += ybuffer[1];
y_ptr += inc_y;
y_ptr[0] += ybuffer[2];
y_ptr[1] += ybuffer[3];
y_ptr += inc_y;
y_ptr[0] += ybuffer[4];
y_ptr[1] += ybuffer[5];
y_ptr += inc_y;
y_ptr[0] += ybuffer[6];
y_ptr[1] += ybuffer[7];
y_ptr += inc_y;
}
for( i = 0; i < n2 ; i++)
{
memset(ybuffer,0,32);
cgemv_kernel_4x1(NB,a_ptr,xbuffer,ybuffer,alpha);
a_ptr += lda;
y_ptr[0] += ybuffer[0];
y_ptr[1] += ybuffer[1];
y_ptr += inc_y;
}
}
a += 2 * NB;
x += NB * inc_x;
}
if ( m3 == 0 ) return(0);
x_ptr = x;
j=0;
a_ptr = a;
y_ptr = y;
if ( m3 == 3 )
{
FLOAT temp_r ;
FLOAT temp_i ;
FLOAT x0 = x_ptr[0];
FLOAT x1 = x_ptr[1];
x_ptr += inc_x;
FLOAT x2 = x_ptr[0];
FLOAT x3 = x_ptr[1];
x_ptr += inc_x;
FLOAT x4 = x_ptr[0];
FLOAT x5 = x_ptr[1];
while ( j < n)
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
temp_r += a_ptr[2] * x2 - a_ptr[3] * x3;
temp_i += a_ptr[2] * x3 + a_ptr[3] * x2;
temp_r += a_ptr[4] * x4 - a_ptr[5] * x5;
temp_i += a_ptr[4] * x5 + a_ptr[5] * x4;
#else
temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
temp_r += a_ptr[2] * x2 + a_ptr[3] * x3;
temp_i += a_ptr[2] * x3 - a_ptr[3] * x2;
temp_r += a_ptr[4] * x4 + a_ptr[5] * x5;
temp_i += a_ptr[4] * x5 - a_ptr[5] * x4;
#endif
#if !defined(XCONJ)
y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
#else
y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
#endif
a_ptr += lda;
y_ptr += inc_y;
j++;
}
return(0);
}
if ( m3 == 2 )
{
FLOAT temp_r ;
FLOAT temp_i ;
FLOAT temp_r1 ;
FLOAT temp_i1 ;
FLOAT x0 = x_ptr[0];
FLOAT x1 = x_ptr[1];
x_ptr += inc_x;
FLOAT x2 = x_ptr[0];
FLOAT x3 = x_ptr[1];
FLOAT ar = alpha[0];
FLOAT ai = alpha[1];
while ( j < ( n & -2 ))
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
temp_r += a_ptr[2] * x2 - a_ptr[3] * x3;
temp_i += a_ptr[2] * x3 + a_ptr[3] * x2;
a_ptr += lda;
temp_r1 = a_ptr[0] * x0 - a_ptr[1] * x1;
temp_i1 = a_ptr[0] * x1 + a_ptr[1] * x0;
temp_r1 += a_ptr[2] * x2 - a_ptr[3] * x3;
temp_i1 += a_ptr[2] * x3 + a_ptr[3] * x2;
#else
temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
temp_r += a_ptr[2] * x2 + a_ptr[3] * x3;
temp_i += a_ptr[2] * x3 - a_ptr[3] * x2;
a_ptr += lda;
temp_r1 = a_ptr[0] * x0 + a_ptr[1] * x1;
temp_i1 = a_ptr[0] * x1 - a_ptr[1] * x0;
temp_r1 += a_ptr[2] * x2 + a_ptr[3] * x3;
temp_i1 += a_ptr[2] * x3 - a_ptr[3] * x2;
#endif
#if !defined(XCONJ)
y_ptr[0] += ar * temp_r - ai * temp_i;
y_ptr[1] += ar * temp_i + ai * temp_r;
y_ptr += inc_y;
y_ptr[0] += ar * temp_r1 - ai * temp_i1;
y_ptr[1] += ar * temp_i1 + ai * temp_r1;
#else
y_ptr[0] += ar * temp_r + ai * temp_i;
y_ptr[1] -= ar * temp_i - ai * temp_r;
y_ptr += inc_y;
y_ptr[0] += ar * temp_r1 + ai * temp_i1;
y_ptr[1] -= ar * temp_i1 - ai * temp_r1;
#endif
a_ptr += lda;
y_ptr += inc_y;
j+=2;
}
while ( j < n)
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
temp_r += a_ptr[2] * x2 - a_ptr[3] * x3;
temp_i += a_ptr[2] * x3 + a_ptr[3] * x2;
#else
temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
temp_r += a_ptr[2] * x2 + a_ptr[3] * x3;
temp_i += a_ptr[2] * x3 - a_ptr[3] * x2;
#endif
#if !defined(XCONJ)
y_ptr[0] += ar * temp_r - ai * temp_i;
y_ptr[1] += ar * temp_i + ai * temp_r;
#else
y_ptr[0] += ar * temp_r + ai * temp_i;
y_ptr[1] -= ar * temp_i - ai * temp_r;
#endif
a_ptr += lda;
y_ptr += inc_y;
j++;
}
return(0);
}
if ( m3 == 1 )
{
FLOAT temp_r ;
FLOAT temp_i ;
FLOAT temp_r1 ;
FLOAT temp_i1 ;
FLOAT x0 = x_ptr[0];
FLOAT x1 = x_ptr[1];
FLOAT ar = alpha[0];
FLOAT ai = alpha[1];
while ( j < ( n & -2 ))
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
a_ptr += lda;
temp_r1 = a_ptr[0] * x0 - a_ptr[1] * x1;
temp_i1 = a_ptr[0] * x1 + a_ptr[1] * x0;
#else
temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
a_ptr += lda;
temp_r1 = a_ptr[0] * x0 + a_ptr[1] * x1;
temp_i1 = a_ptr[0] * x1 - a_ptr[1] * x0;
#endif
#if !defined(XCONJ)
y_ptr[0] += ar * temp_r - ai * temp_i;
y_ptr[1] += ar * temp_i + ai * temp_r;
y_ptr += inc_y;
y_ptr[0] += ar * temp_r1 - ai * temp_i1;
y_ptr[1] += ar * temp_i1 + ai * temp_r1;
#else
y_ptr[0] += ar * temp_r + ai * temp_i;
y_ptr[1] -= ar * temp_i - ai * temp_r;
y_ptr += inc_y;
y_ptr[0] += ar * temp_r1 + ai * temp_i1;
y_ptr[1] -= ar * temp_i1 - ai * temp_r1;
#endif
a_ptr += lda;
y_ptr += inc_y;
j+=2;
}
while ( j < n)
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
#else
temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
#endif
#if !defined(XCONJ)
y_ptr[0] += ar * temp_r - ai * temp_i;
y_ptr[1] += ar * temp_i + ai * temp_r;
#else
y_ptr[0] += ar * temp_r + ai * temp_i;
y_ptr[1] -= ar * temp_i - ai * temp_r;
#endif
a_ptr += lda;
y_ptr += inc_y;
j++;
}
return(0);
}
return(0);
}

171
kernel/x86_64/cgemv_t_microk_haswell-2.c
View File

@ -1,171 +0,0 @@
/***************************************************************************
Copyright (c) 2014, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary froms, 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 from 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 cgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void cgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // temp
"vxorps %%ymm9 , %%ymm9 , %%ymm9 \n\t" // temp
"vxorps %%ymm10, %%ymm10, %%ymm10 \n\t" // temp
"vxorps %%ymm11, %%ymm11, %%ymm11 \n\t" // temp
"vxorps %%ymm12, %%ymm12, %%ymm12 \n\t" // temp
"vxorps %%ymm13, %%ymm13, %%ymm13 \n\t"
"vxorps %%ymm14, %%ymm14, %%ymm14 \n\t"
"vxorps %%ymm15, %%ymm15, %%ymm15 \n\t"
".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 192(%4,%0,4) \n\t"
"vmovups (%4,%0,4), %%ymm4 \n\t" // 4 complex values from a0
"prefetcht0 192(%5,%0,4) \n\t"
"vmovups (%5,%0,4), %%ymm5 \n\t" // 4 complex values from a1
"prefetcht0 192(%2,%0,4) \n\t"
"vmovups (%2,%0,4) , %%ymm6 \n\t" // 4 complex values from x
"vpermilps $0xb1, %%ymm6, %%ymm7 \n\t" // exchange real and imap parts
"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0 \n\t" // only the real parts
"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1 \n\t" // only the imag parts
"prefetcht0 192(%6,%0,4) \n\t"
"vmovups (%6,%0,4), %%ymm6 \n\t" // 4 complex values from a2
"prefetcht0 192(%7,%0,4) \n\t"
"vmovups (%7,%0,4), %%ymm7 \n\t" // 4 complex values from a3
"vfmadd231ps %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231ps %%ymm5 , %%ymm0, %%ymm10 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm5 , %%ymm1, %%ymm11 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231ps %%ymm6 , %%ymm0, %%ymm12 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm6 , %%ymm1, %%ymm13 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231ps %%ymm7 , %%ymm0, %%ymm14 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm7 , %%ymm1, %%ymm15 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vmovups 32(%4,%0,4), %%ymm4 \n\t" // 2 complex values from a0
"vmovups 32(%5,%0,4), %%ymm5 \n\t" // 2 complex values from a1
"vmovups 32(%2,%0,4) , %%ymm6 \n\t" // 4 complex values from x
"vpermilps $0xb1, %%ymm6, %%ymm7 \n\t" // exchange real and imap parts
"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0 \n\t" // only the real parts
"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1 \n\t" // only the imag parts
"vmovups 32(%6,%0,4), %%ymm6 \n\t" // 2 complex values from a2
"vmovups 32(%7,%0,4), %%ymm7 \n\t" // 2 complex values from a3
"vfmadd231ps %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231ps %%ymm5 , %%ymm0, %%ymm10 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm5 , %%ymm1, %%ymm11 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231ps %%ymm6 , %%ymm0, %%ymm12 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm6 , %%ymm1, %%ymm13 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231ps %%ymm7 , %%ymm0, %%ymm14 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm7 , %%ymm1, %%ymm15 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"addq $16 , %0 \n\t"
"subq $8 , %1 \n\t"
"jnz .L01LOOP%= \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilps $0xb1 , %%ymm9 , %%ymm9 \n\t"
"vpermilps $0xb1 , %%ymm11, %%ymm11 \n\t"
"vpermilps $0xb1 , %%ymm13, %%ymm13 \n\t"
"vpermilps $0xb1 , %%ymm15, %%ymm15 \n\t"
"vaddsubps %%ymm9 , %%ymm8, %%ymm8 \n\t"
"vaddsubps %%ymm11, %%ymm10, %%ymm10 \n\t"
"vaddsubps %%ymm13, %%ymm12, %%ymm12 \n\t"
"vaddsubps %%ymm15, %%ymm14, %%ymm14 \n\t"
#else
"vpermilps $0xb1 , %%ymm8 , %%ymm8 \n\t"
"vpermilps $0xb1 , %%ymm10, %%ymm10 \n\t"
"vpermilps $0xb1 , %%ymm12, %%ymm12 \n\t"
"vpermilps $0xb1 , %%ymm14, %%ymm14 \n\t"
"vaddsubps %%ymm8 , %%ymm9 , %%ymm8 \n\t"
"vaddsubps %%ymm10, %%ymm11, %%ymm10 \n\t"
"vaddsubps %%ymm12, %%ymm13, %%ymm12 \n\t"
"vaddsubps %%ymm14, %%ymm15, %%ymm14 \n\t"
"vpermilps $0xb1 , %%ymm8 , %%ymm8 \n\t"
"vpermilps $0xb1 , %%ymm10, %%ymm10 \n\t"
"vpermilps $0xb1 , %%ymm12, %%ymm12 \n\t"
"vpermilps $0xb1 , %%ymm14, %%ymm14 \n\t"
#endif
"vextractf128 $1, %%ymm8 , %%xmm9 \n\t"
"vextractf128 $1, %%ymm10, %%xmm11 \n\t"
"vextractf128 $1, %%ymm12, %%xmm13 \n\t"
"vextractf128 $1, %%ymm14, %%xmm15 \n\t"
"vaddps %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vaddps %%xmm10, %%xmm11, %%xmm10 \n\t"
"vaddps %%xmm12, %%xmm13, %%xmm12 \n\t"
"vaddps %%xmm14, %%xmm15, %%xmm14 \n\t"
"vshufpd $0x1, %%xmm8 , %%xmm8 , %%xmm9 \n\t"
"vshufpd $0x1, %%xmm10, %%xmm10, %%xmm11 \n\t"
"vshufpd $0x1, %%xmm12, %%xmm12, %%xmm13 \n\t"
"vshufpd $0x1, %%xmm14, %%xmm14, %%xmm15 \n\t"
"vaddps %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vaddps %%xmm10, %%xmm11, %%xmm10 \n\t"
"vaddps %%xmm12, %%xmm13, %%xmm12 \n\t"
"vaddps %%xmm14, %%xmm15, %%xmm14 \n\t"
"vmovsd %%xmm8 , (%3) \n\t"
"vmovsd %%xmm10, 8(%3) \n\t"
"vmovsd %%xmm12, 16(%3) \n\t"
"vmovsd %%xmm14, 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",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}

539
kernel/x86_64/cgemv_t_microk_haswell-4.c Normal file
View File

@ -0,0 +1,539 @@
/***************************************************************************
Copyright (c) 2014, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary froms, 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 from 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_4x4 1
static void cgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha) __attribute__ ((noinline));
static void cgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // temp
"vxorps %%ymm9 , %%ymm9 , %%ymm9 \n\t" // temp
"vxorps %%ymm10, %%ymm10, %%ymm10 \n\t" // temp
"vxorps %%ymm11, %%ymm11, %%ymm11 \n\t" // temp
"vxorps %%ymm12, %%ymm12, %%ymm12 \n\t" // temp
"vxorps %%ymm13, %%ymm13, %%ymm13 \n\t"
"vxorps %%ymm14, %%ymm14, %%ymm14 \n\t"
"vxorps %%ymm15, %%ymm15, %%ymm15 \n\t"
"testq $0x04, %1 \n\t"
"jz .L08LABEL%= \n\t"
"vmovups (%4,%0,4), %%ymm4 \n\t" // 4 complex values from a0
"vmovups (%5,%0,4), %%ymm5 \n\t" // 4 complex values from a1
"vmovups (%2,%0,4) , %%ymm6 \n\t" // 4 complex values from x
"vpermilps $0xb1, %%ymm6, %%ymm7 \n\t" // exchange real and imap parts
"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0 \n\t" // only the real parts
"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1 \n\t" // only the imag parts
"vmovups (%6,%0,4), %%ymm6 \n\t" // 4 complex values from a2
"vmovups (%7,%0,4), %%ymm7 \n\t" // 4 complex values from a3
"vfmadd231ps %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231ps %%ymm5 , %%ymm0, %%ymm10 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm5 , %%ymm1, %%ymm11 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231ps %%ymm6 , %%ymm0, %%ymm12 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm6 , %%ymm1, %%ymm13 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231ps %%ymm7 , %%ymm0, %%ymm14 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm7 , %%ymm1, %%ymm15 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"addq $8 , %0 \n\t"
"subq $4 , %1 \n\t"
".L08LABEL%=: \n\t"
"cmpq $0, %1 \n\t"
"je .L08END%= \n\t"
".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 192(%4,%0,4) \n\t"
"vmovups (%4,%0,4), %%ymm4 \n\t" // 4 complex values from a0
"prefetcht0 192(%5,%0,4) \n\t"
"vmovups (%5,%0,4), %%ymm5 \n\t" // 4 complex values from a1
"prefetcht0 192(%2,%0,4) \n\t"
"vmovups (%2,%0,4) , %%ymm6 \n\t" // 4 complex values from x
"vpermilps $0xb1, %%ymm6, %%ymm7 \n\t" // exchange real and imap parts
"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0 \n\t" // only the real parts
"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1 \n\t" // only the imag parts
"prefetcht0 192(%6,%0,4) \n\t"
"vmovups (%6,%0,4), %%ymm6 \n\t" // 4 complex values from a2
"prefetcht0 192(%7,%0,4) \n\t"
"vmovups (%7,%0,4), %%ymm7 \n\t" // 4 complex values from a3
"vfmadd231ps %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231ps %%ymm5 , %%ymm0, %%ymm10 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm5 , %%ymm1, %%ymm11 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231ps %%ymm6 , %%ymm0, %%ymm12 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm6 , %%ymm1, %%ymm13 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231ps %%ymm7 , %%ymm0, %%ymm14 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm7 , %%ymm1, %%ymm15 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vmovups 32(%4,%0,4), %%ymm4 \n\t" // 4 complex values from a0
"vmovups 32(%5,%0,4), %%ymm5 \n\t" // 4 complex values from a1
"vmovups 32(%2,%0,4) , %%ymm6 \n\t" // 4 complex values from x
"vpermilps $0xb1, %%ymm6, %%ymm7 \n\t" // exchange real and imap parts
"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0 \n\t" // only the real parts
"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1 \n\t" // only the imag parts
"vmovups 32(%6,%0,4), %%ymm6 \n\t" // 4 complex values from a2
"vmovups 32(%7,%0,4), %%ymm7 \n\t" // 4 complex values from a3
"vfmadd231ps %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231ps %%ymm5 , %%ymm0, %%ymm10 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm5 , %%ymm1, %%ymm11 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231ps %%ymm6 , %%ymm0, %%ymm12 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm6 , %%ymm1, %%ymm13 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231ps %%ymm7 , %%ymm0, %%ymm14 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm7 , %%ymm1, %%ymm15 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"addq $16 , %0 \n\t"
"subq $8 , %1 \n\t"
"jnz .L01LOOP%= \n\t"
".L08END%=: \n\t"
"vbroadcastss (%8) , %%xmm0 \n\t" // value from alpha
"vbroadcastss 4(%8) , %%xmm1 \n\t" // value from alpha
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilps $0xb1 , %%ymm9 , %%ymm9 \n\t"
"vpermilps $0xb1 , %%ymm11, %%ymm11 \n\t"
"vpermilps $0xb1 , %%ymm13, %%ymm13 \n\t"
"vpermilps $0xb1 , %%ymm15, %%ymm15 \n\t"
"vaddsubps %%ymm9 , %%ymm8, %%ymm8 \n\t"
"vaddsubps %%ymm11, %%ymm10, %%ymm10 \n\t"
"vaddsubps %%ymm13, %%ymm12, %%ymm12 \n\t"
"vaddsubps %%ymm15, %%ymm14, %%ymm14 \n\t"
#else
"vpermilps $0xb1 , %%ymm8 , %%ymm8 \n\t"
"vpermilps $0xb1 , %%ymm10, %%ymm10 \n\t"
"vpermilps $0xb1 , %%ymm12, %%ymm12 \n\t"
"vpermilps $0xb1 , %%ymm14, %%ymm14 \n\t"
"vaddsubps %%ymm8 , %%ymm9 , %%ymm8 \n\t"
"vaddsubps %%ymm10, %%ymm11, %%ymm10 \n\t"
"vaddsubps %%ymm12, %%ymm13, %%ymm12 \n\t"
"vaddsubps %%ymm14, %%ymm15, %%ymm14 \n\t"
"vpermilps $0xb1 , %%ymm8 , %%ymm8 \n\t"
"vpermilps $0xb1 , %%ymm10, %%ymm10 \n\t"
"vpermilps $0xb1 , %%ymm12, %%ymm12 \n\t"
"vpermilps $0xb1 , %%ymm14, %%ymm14 \n\t"
#endif
"vmovsd (%3), %%xmm4 \n\t" // read y
"vmovsd 8(%3), %%xmm5 \n\t"
"vmovsd 16(%3), %%xmm6 \n\t"
"vmovsd 24(%3), %%xmm7 \n\t"
"vextractf128 $1, %%ymm8 , %%xmm9 \n\t"
"vextractf128 $1, %%ymm10, %%xmm11 \n\t"
"vextractf128 $1, %%ymm12, %%xmm13 \n\t"
"vextractf128 $1, %%ymm14, %%xmm15 \n\t"
"vaddps %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vaddps %%xmm10, %%xmm11, %%xmm10 \n\t"
"vaddps %%xmm12, %%xmm13, %%xmm12 \n\t"
"vaddps %%xmm14, %%xmm15, %%xmm14 \n\t"
"vshufpd $0x1, %%xmm8 , %%xmm8 , %%xmm9 \n\t"
"vshufpd $0x1, %%xmm10, %%xmm10, %%xmm11 \n\t"
"vshufpd $0x1, %%xmm12, %%xmm12, %%xmm13 \n\t"
"vshufpd $0x1, %%xmm14, %%xmm14, %%xmm15 \n\t"
"vaddps %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vaddps %%xmm10, %%xmm11, %%xmm10 \n\t"
"vaddps %%xmm12, %%xmm13, %%xmm12 \n\t"
"vaddps %%xmm14, %%xmm15, %%xmm14 \n\t"
"vmulps %%xmm8 , %%xmm1 , %%xmm9 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulps %%xmm8 , %%xmm0 , %%xmm8 \n\t" // t_r * alpha_r , t_i * alpha_r
"vmulps %%xmm10, %%xmm1 , %%xmm11 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulps %%xmm10, %%xmm0 , %%xmm10 \n\t" // t_r * alpha_r , t_i * alpha_r
"vmulps %%xmm12, %%xmm1 , %%xmm13 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulps %%xmm12, %%xmm0 , %%xmm12 \n\t" // t_r * alpha_r , t_i * alpha_r
"vmulps %%xmm14, %%xmm1 , %%xmm15 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulps %%xmm14, %%xmm0 , %%xmm14 \n\t" // t_r * alpha_r , t_i * alpha_r
#if !defined(XCONJ)
"vpermilps $0xb1 , %%xmm9 , %%xmm9 \n\t"
"vpermilps $0xb1 , %%xmm11, %%xmm11 \n\t"
"vpermilps $0xb1 , %%xmm13, %%xmm13 \n\t"
"vpermilps $0xb1 , %%xmm15, %%xmm15 \n\t"
"vaddsubps %%xmm9 , %%xmm8, %%xmm8 \n\t"
"vaddsubps %%xmm11, %%xmm10, %%xmm10 \n\t"
"vaddsubps %%xmm13, %%xmm12, %%xmm12 \n\t"
"vaddsubps %%xmm15, %%xmm14, %%xmm14 \n\t"
#else
"vpermilps $0xb1 , %%xmm8 , %%xmm8 \n\t"
"vpermilps $0xb1 , %%xmm10, %%xmm10 \n\t"
"vpermilps $0xb1 , %%xmm12, %%xmm12 \n\t"
"vpermilps $0xb1 , %%xmm14, %%xmm14 \n\t"
"vaddsubps %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vaddsubps %%xmm10, %%xmm11, %%xmm10 \n\t"
"vaddsubps %%xmm12, %%xmm13, %%xmm12 \n\t"
"vaddsubps %%xmm14, %%xmm15, %%xmm14 \n\t"
"vpermilps $0xb1 , %%xmm8 , %%xmm8 \n\t"
"vpermilps $0xb1 , %%xmm10, %%xmm10 \n\t"
"vpermilps $0xb1 , %%xmm12, %%xmm12 \n\t"
"vpermilps $0xb1 , %%xmm14, %%xmm14 \n\t"
#endif
"vaddps %%xmm8 , %%xmm4 , %%xmm8 \n\t"
"vaddps %%xmm10, %%xmm5 , %%xmm10 \n\t"
"vaddps %%xmm12, %%xmm6 , %%xmm12 \n\t"
"vaddps %%xmm14, %%xmm7 , %%xmm14 \n\t"
"vmovsd %%xmm8 , (%3) \n\t"
"vmovsd %%xmm10, 8(%3) \n\t"
"vmovsd %%xmm12, 16(%3) \n\t"
"vmovsd %%xmm14, 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
"r" (alpha) // 8
: "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
#define HAVE_KERNEL_4x2 1
static void cgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha) __attribute__ ((noinline));
static void cgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // temp
"vxorps %%ymm9 , %%ymm9 , %%ymm9 \n\t" // temp
"vxorps %%ymm10, %%ymm10, %%ymm10 \n\t" // temp
"vxorps %%ymm11, %%ymm11, %%ymm11 \n\t" // temp
"testq $0x04, %1 \n\t"
"jz .L08LABEL%= \n\t"
"vmovups (%4,%0,4), %%ymm4 \n\t" // 4 complex values from a0
"vmovups (%5,%0,4), %%ymm5 \n\t" // 4 complex values from a1
"vmovups (%2,%0,4) , %%ymm6 \n\t" // 4 complex values from x
"vpermilps $0xb1, %%ymm6, %%ymm7 \n\t" // exchange real and imap parts
"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0 \n\t" // only the real parts
"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1 \n\t" // only the imag parts
"vfmadd231ps %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231ps %%ymm5 , %%ymm0, %%ymm10 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm5 , %%ymm1, %%ymm11 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"addq $8 , %0 \n\t"
"subq $4 , %1 \n\t"
".L08LABEL%=: \n\t"
"cmpq $0, %1 \n\t"
"je .L08END%= \n\t"
".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 192(%4,%0,4) \n\t"
"vmovups (%4,%0,4), %%ymm4 \n\t" // 4 complex values from a0
"prefetcht0 192(%5,%0,4) \n\t"
"vmovups (%5,%0,4), %%ymm5 \n\t" // 4 complex values from a1
"prefetcht0 192(%2,%0,4) \n\t"
"vmovups (%2,%0,4) , %%ymm6 \n\t" // 4 complex values from x
"vpermilps $0xb1, %%ymm6, %%ymm7 \n\t" // exchange real and imap parts
"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0 \n\t" // only the real parts
"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1 \n\t" // only the imag parts
"vfmadd231ps %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231ps %%ymm5 , %%ymm0, %%ymm10 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm5 , %%ymm1, %%ymm11 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vmovups 32(%4,%0,4), %%ymm4 \n\t" // 4 complex values from a0
"vmovups 32(%5,%0,4), %%ymm5 \n\t" // 4 complex values from a1
"vmovups 32(%2,%0,4) , %%ymm6 \n\t" // 4 complex values from x
"vpermilps $0xb1, %%ymm6, %%ymm7 \n\t" // exchange real and imap parts
"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0 \n\t" // only the real parts
"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1 \n\t" // only the imag parts
"vfmadd231ps %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231ps %%ymm5 , %%ymm0, %%ymm10 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm5 , %%ymm1, %%ymm11 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"addq $16 , %0 \n\t"
"subq $8 , %1 \n\t"
"jnz .L01LOOP%= \n\t"
".L08END%=: \n\t"
"vbroadcastss (%6) , %%xmm0 \n\t" // value from alpha
"vbroadcastss 4(%6) , %%xmm1 \n\t" // value from alpha
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilps $0xb1 , %%ymm9 , %%ymm9 \n\t"
"vpermilps $0xb1 , %%ymm11, %%ymm11 \n\t"
"vaddsubps %%ymm9 , %%ymm8, %%ymm8 \n\t"
"vaddsubps %%ymm11, %%ymm10, %%ymm10 \n\t"
#else
"vpermilps $0xb1 , %%ymm8 , %%ymm8 \n\t"
"vpermilps $0xb1 , %%ymm10, %%ymm10 \n\t"
"vaddsubps %%ymm8 , %%ymm9 , %%ymm8 \n\t"
"vaddsubps %%ymm10, %%ymm11, %%ymm10 \n\t"
"vpermilps $0xb1 , %%ymm8 , %%ymm8 \n\t"
"vpermilps $0xb1 , %%ymm10, %%ymm10 \n\t"
#endif
"vmovsd (%3), %%xmm4 \n\t" // read y
"vmovsd 8(%3), %%xmm5 \n\t"
"vextractf128 $1, %%ymm8 , %%xmm9 \n\t"
"vextractf128 $1, %%ymm10, %%xmm11 \n\t"
"vaddps %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vaddps %%xmm10, %%xmm11, %%xmm10 \n\t"
"vshufpd $0x1, %%xmm8 , %%xmm8 , %%xmm9 \n\t"
"vshufpd $0x1, %%xmm10, %%xmm10, %%xmm11 \n\t"
"vaddps %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vaddps %%xmm10, %%xmm11, %%xmm10 \n\t"
"vmulps %%xmm8 , %%xmm1 , %%xmm9 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulps %%xmm8 , %%xmm0 , %%xmm8 \n\t" // t_r * alpha_r , t_i * alpha_r
"vmulps %%xmm10, %%xmm1 , %%xmm11 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulps %%xmm10, %%xmm0 , %%xmm10 \n\t" // t_r * alpha_r , t_i * alpha_r
#if !defined(XCONJ)
"vpermilps $0xb1 , %%xmm9 , %%xmm9 \n\t"
"vpermilps $0xb1 , %%xmm11, %%xmm11 \n\t"
"vaddsubps %%xmm9 , %%xmm8, %%xmm8 \n\t"
"vaddsubps %%xmm11, %%xmm10, %%xmm10 \n\t"
#else
"vpermilps $0xb1 , %%xmm8 , %%xmm8 \n\t"
"vpermilps $0xb1 , %%xmm10, %%xmm10 \n\t"
"vaddsubps %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vaddsubps %%xmm10, %%xmm11, %%xmm10 \n\t"
"vpermilps $0xb1 , %%xmm8 , %%xmm8 \n\t"
"vpermilps $0xb1 , %%xmm10, %%xmm10 \n\t"
#endif
"vaddps %%xmm8 , %%xmm4 , %%xmm8 \n\t"
"vaddps %%xmm10, %%xmm5 , %%xmm10 \n\t"
"vmovsd %%xmm8 , (%3) \n\t"
"vmovsd %%xmm10, 8(%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" (alpha) // 6
: "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
#define HAVE_KERNEL_4x1 1
static void cgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT *alpha) __attribute__ ((noinline));
static void cgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // temp
"vxorps %%ymm9 , %%ymm9 , %%ymm9 \n\t" // temp
"testq $0x04, %1 \n\t"
"jz .L08LABEL%= \n\t"
"vmovups (%4,%0,4), %%ymm4 \n\t" // 4 complex values from a0
"vmovups (%2,%0,4) , %%ymm6 \n\t" // 4 complex values from x
"vpermilps $0xb1, %%ymm6, %%ymm7 \n\t" // exchange real and imap parts
"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0 \n\t" // only the real parts
"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1 \n\t" // only the imag parts
"vfmadd231ps %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"addq $8 , %0 \n\t"
"subq $4 , %1 \n\t"
".L08LABEL%=: \n\t"
"cmpq $0, %1 \n\t"
"je .L08END%= \n\t"
".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 192(%4,%0,4) \n\t"
"vmovups (%4,%0,4), %%ymm4 \n\t" // 4 complex values from a0
"prefetcht0 192(%2,%0,4) \n\t"
"vmovups (%2,%0,4) , %%ymm6 \n\t" // 4 complex values from x
"vpermilps $0xb1, %%ymm6, %%ymm7 \n\t" // exchange real and imap parts
"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0 \n\t" // only the real parts
"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1 \n\t" // only the imag parts
"vfmadd231ps %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vmovups 32(%4,%0,4), %%ymm4 \n\t" // 4 complex values from a0
"vmovups 32(%2,%0,4) , %%ymm6 \n\t" // 4 complex values from x
"vpermilps $0xb1, %%ymm6, %%ymm7 \n\t" // exchange real and imap parts
"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0 \n\t" // only the real parts
"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1 \n\t" // only the imag parts
"vfmadd231ps %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"addq $16 , %0 \n\t"
"subq $8 , %1 \n\t"
"jnz .L01LOOP%= \n\t"
".L08END%=: \n\t"
"vbroadcastss (%5) , %%xmm0 \n\t" // value from alpha
"vbroadcastss 4(%5) , %%xmm1 \n\t" // value from alpha
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilps $0xb1 , %%ymm9 , %%ymm9 \n\t"
"vaddsubps %%ymm9 , %%ymm8, %%ymm8 \n\t"
#else
"vpermilps $0xb1 , %%ymm8 , %%ymm8 \n\t"
"vaddsubps %%ymm8 , %%ymm9 , %%ymm8 \n\t"
"vpermilps $0xb1 , %%ymm8 , %%ymm8 \n\t"
#endif
"vmovsd (%3), %%xmm4 \n\t" // read y
"vextractf128 $1, %%ymm8 , %%xmm9 \n\t"
"vaddps %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vshufpd $0x1, %%xmm8 , %%xmm8 , %%xmm9 \n\t"
"vaddps %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vmulps %%xmm8 , %%xmm1 , %%xmm9 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulps %%xmm8 , %%xmm0 , %%xmm8 \n\t" // t_r * alpha_r , t_i * alpha_r
#if !defined(XCONJ)
"vpermilps $0xb1 , %%xmm9 , %%xmm9 \n\t"
"vaddsubps %%xmm9 , %%xmm8, %%xmm8 \n\t"
#else
"vpermilps $0xb1 , %%xmm8 , %%xmm8 \n\t"
"vaddsubps %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vpermilps $0xb1 , %%xmm8 , %%xmm8 \n\t"
#endif
"vaddps %%xmm8 , %%xmm4 , %%xmm8 \n\t"
"vmovsd %%xmm8 , (%3) \n\t"
"vzeroupper \n\t"
:
:
"r" (i), // 0
"r" (n), // 1
"r" (x), // 2
"r" (y), // 3
"r" (ap), // 4
"r" (alpha) // 5
: "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}

208
kernel/x86_64/dgemv_n.c
View File

@ -1,208 +0,0 @@
/***************************************************************************
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_n_microk_haswell-2.c"
#elif defined(NEHALEM)
#include "dgemv_n_microk_nehalem-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];
for ( i=0; i< n; i+=4 )
{
y[i] += a0[i]*x[0] + a1[i]*x[1] + a2[i]*x[2] + a3[i]*x[3];
y[i+1] += a0[i+1]*x[0] + a1[i+1]*x[1] + a2[i+1]*x[2] + a3[i+1]*x[3];
y[i+2] += a0[i+2]*x[0] + a1[i+2]*x[1] + a2[i+2]*x[2] + a3[i+2]*x[3];
y[i+3] += a0[i+3]*x[0] + a1[i+3]*x[1] + a2[i+3]*x[2] + a3[i+3]*x[3];
}
}
#endif
static void dgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
{
BLASLONG i;
FLOAT *a0;
a0 = ap;
for ( i=0; i< n; i+=4 )
{
y[i] += a0[i]*x[0];
y[i+1] += a0[i+1]*x[0];
y[i+2] += a0[i+2]*x[0];
y[i+3] += a0[i+3]*x[0];
}
}
static void zero_y(BLASLONG n, FLOAT *dest)
{
BLASLONG i;
for ( i=0; i<n; i++ )
{
*dest = 0.0;
dest++;
}
}
static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest)
{
BLASLONG i;
if ( inc_dest == 1 )
{
for ( i=0; i<n; i+=4 )
{
dest[i] += src[i];
dest[i+1] += src[i+1];
dest[i+2] += src[i+2];
dest[i+3] += src[i+3];
}
}
else
{
for ( i=0; i<n; i++ )
{
*dest += *src;
src++;
dest += inc_dest;
}
}
}
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 xbuffer[4],*ybuffer;
if ( m < 1 ) return(0);
if ( n < 1 ) return(0);
ybuffer = buffer;
n1 = n / 4 ;
n2 = n % 4 ;
m1 = m - ( m % 16 );
m2 = (m % NBMAX) - (m % 16) ;
y_ptr = y;
BLASLONG NB = NBMAX;
while ( NB == NBMAX )
{
m1 -= NB;
if ( m1 < 0)
{
if ( m2 == 0 ) break;
NB = m2;
}
a_ptr = a;
x_ptr = x;
zero_y(NB,ybuffer);
for( i = 0; i < n1 ; i++)
{
xbuffer[0] = alpha * x_ptr[0];
x_ptr += inc_x;
xbuffer[1] = alpha * x_ptr[0];
x_ptr += inc_x;
xbuffer[2] = alpha * x_ptr[0];
x_ptr += inc_x;
xbuffer[3] = alpha * x_ptr[0];
x_ptr += inc_x;
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;
}
for( i = 0; i < n2 ; i++)
{
xbuffer[0] = alpha * x_ptr[0];
x_ptr += inc_x;
dgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
a_ptr += 1 * lda;
}
add_y(NB,ybuffer,y_ptr,inc_y);
a += NB;
y_ptr += NB * inc_y;
}
j=0;
while ( j < (m % 16))
{
a_ptr = a;
x_ptr = x;
FLOAT temp = 0.0;
for( i = 0; i < n; i++ )
{
temp += a_ptr[0] * x_ptr[0];
a_ptr += lda;
x_ptr += inc_x;
}
y_ptr[0] += alpha * temp;
y_ptr += inc_y;
a++;
j++;
}
return(0);
}

89
kernel/x86_64/dgemv_n_microk_haswell-2.c
View File

@ -1,89 +0,0 @@
/***************************************************************************
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"
"vbroadcastsd (%2), %%ymm12 \n\t" // x0
"vbroadcastsd 8(%2), %%ymm13 \n\t" // x1
"vbroadcastsd 16(%2), %%ymm14 \n\t" // x2
"vbroadcastsd 24(%2), %%ymm15 \n\t" // x3
".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 192(%3,%0,8) \n\t"
"vmovups (%3,%0,8), %%ymm4 \n\t" // 4 * y
"vmovups 32(%3,%0,8), %%ymm5 \n\t" // 4 * y
"prefetcht0 192(%4,%0,8) \n\t"
"vfmadd231pd (%4,%0,8), %%ymm12, %%ymm4 \n\t"
"vfmadd231pd 32(%4,%0,8), %%ymm12, %%ymm5 \n\t"
"prefetcht0 192(%5,%0,8) \n\t"
"vfmadd231pd (%5,%0,8), %%ymm13, %%ymm4 \n\t"
"vfmadd231pd 32(%5,%0,8), %%ymm13, %%ymm5 \n\t"
"prefetcht0 192(%6,%0,8) \n\t"
"vfmadd231pd (%6,%0,8), %%ymm14, %%ymm4 \n\t"
"vfmadd231pd 32(%6,%0,8), %%ymm14, %%ymm5 \n\t"
"prefetcht0 192(%7,%0,8) \n\t"
"vfmadd231pd (%7,%0,8), %%ymm15, %%ymm4 \n\t"
"vfmadd231pd 32(%7,%0,8), %%ymm15, %%ymm5 \n\t"
"vmovups %%ymm4, (%3,%0,8) \n\t" // 4 * y
"vmovups %%ymm5, 32(%3,%0,8) \n\t" // 4 * y
"addq $8 , %0 \n\t"
"subq $8 , %1 \n\t"
"jnz .L01LOOP%= \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",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}

137
kernel/x86_64/dgemv_n_microk_nehalem-2.c
View File

@ -1,137 +0,0 @@
/***************************************************************************
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__
(
"movddup (%2), %%xmm12 \n\t" // x0
"movddup 8(%2), %%xmm13 \n\t" // x1
"movddup 16(%2), %%xmm14 \n\t" // x2
"movddup 24(%2), %%xmm15 \n\t" // x3
".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 192(%3,%0,8) \n\t"
"movups (%3,%0,8), %%xmm4 \n\t" // 2 * y
"movups 16(%3,%0,8), %%xmm5 \n\t" // 2 * y
"movups 32(%3,%0,8), %%xmm6 \n\t" // 2 * y
"movups 48(%3,%0,8), %%xmm7 \n\t" // 2 * y
"movups (%4,%0,8), %%xmm8 \n\t" // 2 * a
"movups 16(%4,%0,8), %%xmm9 \n\t" // 2 * a
"movups 32(%4,%0,8), %%xmm10 \n\t" // 2 * a
"movups 48(%4,%0,8), %%xmm11 \n\t" // 2 * a
"prefetcht0 192(%4,%0,8) \n\t"
"mulpd %%xmm12 , %%xmm8 \n\t" // a * x
"mulpd %%xmm12 , %%xmm9 \n\t" // a * x
"mulpd %%xmm12 , %%xmm10 \n\t" // a * x
"mulpd %%xmm12 , %%xmm11 \n\t" // a * x
"addpd %%xmm8 , %%xmm4 \n\t" // y += a * x
"addpd %%xmm9 , %%xmm5 \n\t" // y += a * x
"addpd %%xmm10 , %%xmm6 \n\t" // y += a * x
"addpd %%xmm11 , %%xmm7 \n\t" // y += a * x
"prefetcht0 192(%5,%0,8) \n\t"
"movups (%5,%0,8), %%xmm8 \n\t" // 2 * a
"movups 16(%5,%0,8), %%xmm9 \n\t" // 2 * a
"movups 32(%5,%0,8), %%xmm10 \n\t" // 2 * a
"movups 48(%5,%0,8), %%xmm11 \n\t" // 2 * a
"mulpd %%xmm13 , %%xmm8 \n\t" // a * x
"mulpd %%xmm13 , %%xmm9 \n\t" // a * x
"mulpd %%xmm13 , %%xmm10 \n\t" // a * x
"mulpd %%xmm13 , %%xmm11 \n\t" // a * x
"addpd %%xmm8 , %%xmm4 \n\t" // y += a * x
"addpd %%xmm9 , %%xmm5 \n\t" // y += a * x
"addpd %%xmm10 , %%xmm6 \n\t" // y += a * x
"addpd %%xmm11 , %%xmm7 \n\t" // y += a * x
"prefetcht0 192(%6,%0,8) \n\t"
"movups (%6,%0,8), %%xmm8 \n\t" // 2 * a
"movups 16(%6,%0,8), %%xmm9 \n\t" // 2 * a
"movups 32(%6,%0,8), %%xmm10 \n\t" // 2 * a
"movups 48(%6,%0,8), %%xmm11 \n\t" // 2 * a
"mulpd %%xmm14 , %%xmm8 \n\t" // a * x
"mulpd %%xmm14 , %%xmm9 \n\t" // a * x
"mulpd %%xmm14 , %%xmm10 \n\t" // a * x
"mulpd %%xmm14 , %%xmm11 \n\t" // a * x
"addpd %%xmm8 , %%xmm4 \n\t" // y += a * x
"addpd %%xmm9 , %%xmm5 \n\t" // y += a * x
"addpd %%xmm10 , %%xmm6 \n\t" // y += a * x
"addpd %%xmm11 , %%xmm7 \n\t" // y += a * x
"prefetcht0 192(%7,%0,8) \n\t"
"movups (%7,%0,8), %%xmm8 \n\t" // 2 * a
"movups 16(%7,%0,8), %%xmm9 \n\t" // 2 * a
"movups 32(%7,%0,8), %%xmm10 \n\t" // 2 * a
"movups 48(%7,%0,8), %%xmm11 \n\t" // 2 * a
"mulpd %%xmm15 , %%xmm8 \n\t" // a * x
"mulpd %%xmm15 , %%xmm9 \n\t" // a * x
"mulpd %%xmm15 , %%xmm10 \n\t" // a * x
"mulpd %%xmm15 , %%xmm11 \n\t" // a * x
"addpd %%xmm8 , %%xmm4 \n\t" // y += a * x
"addpd %%xmm9 , %%xmm5 \n\t" // y += a * x
"addpd %%xmm10 , %%xmm6 \n\t" // y += a * x
"addpd %%xmm11 , %%xmm7 \n\t" // y += a * x
"movups %%xmm4, (%3,%0,8) \n\t" // 4 * y
"movups %%xmm5, 16(%3,%0,8) \n\t" // 4 * y
"movups %%xmm6, 32(%3,%0,8) \n\t" // 4 * y
"movups %%xmm7, 48(%3,%0,8) \n\t" // 4 * y
"addq $8 , %0 \n\t"
"subq $8 , %1 \n\t"
"jnz .L01LOOP%= \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",
"%xmm8", "%xmm9",
"%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}

191
kernel/x86_64/dgemv_t.c
View File

@ -1,191 +0,0 @@
/***************************************************************************
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;
if ( m < 1 ) return(0);
if ( n < 1 ) return(0);
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);
}

107
kernel/x86_64/dgemv_t_microk_haswell-2.c
View File

@ -1,107 +0,0 @@
/***************************************************************************
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"
);
}

11
kernel/x86_64/sgemv_n.c
View File

@ -29,17 +29,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "common.h"
#if defined(BULLDOZER) || defined(PILEDRIVER)
#include "sgemv_n_microk_bulldozer-2.c"
#elif defined(HASWELL)
#include "sgemv_n_microk_haswell-2.c"
#elif defined(SANDYBRIDGE)
#include "sgemv_n_microk_sandy-2.c"
#elif defined(NEHALEM)
#include "sgemv_n_microk_nehalem-2.c"
#endif
#define NBMAX 4096
#ifndef HAVE_KERNEL_16x4

99
kernel/x86_64/sgemv_n_microk_bulldozer-2.c
View File

@ -1,99 +0,0 @@
/***************************************************************************
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 sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vbroadcastss (%2), %%xmm12 \n\t" // x0
"vbroadcastss 4(%2), %%xmm13 \n\t" // x1
"vbroadcastss 8(%2), %%xmm14 \n\t" // x2
"vbroadcastss 12(%2), %%xmm15 \n\t" // x3
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vmovups (%3,%0,4), %%xmm4 \n\t" // 4 * y
"vmovups 16(%3,%0,4), %%xmm5 \n\t" // 4 * y
"vmovups 32(%3,%0,4), %%xmm6 \n\t" // 4 * y
"vmovups 48(%3,%0,4), %%xmm7 \n\t" // 4 * y
"prefetcht0 192(%4,%0,4) \n\t"
"vfmaddps %%xmm4, (%4,%0,4), %%xmm12, %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%4,%0,4), %%xmm12, %%xmm5 \n\t"
"vfmaddps %%xmm6, 32(%4,%0,4), %%xmm12, %%xmm6 \n\t"
"vfmaddps %%xmm7, 48(%4,%0,4), %%xmm12, %%xmm7 \n\t"
"prefetcht0 192(%5,%0,4) \n\t"
"vfmaddps %%xmm4, (%5,%0,4), %%xmm13, %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%5,%0,4), %%xmm13, %%xmm5 \n\t"
"vfmaddps %%xmm6, 32(%5,%0,4), %%xmm13, %%xmm6 \n\t"
"vfmaddps %%xmm7, 48(%5,%0,4), %%xmm13, %%xmm7 \n\t"
"prefetcht0 192(%6,%0,4) \n\t"
"vfmaddps %%xmm4, (%6,%0,4), %%xmm14, %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%6,%0,4), %%xmm14, %%xmm5 \n\t"
"vfmaddps %%xmm6, 32(%6,%0,4), %%xmm14, %%xmm6 \n\t"
"vfmaddps %%xmm7, 48(%6,%0,4), %%xmm14, %%xmm7 \n\t"
"prefetcht0 192(%7,%0,4) \n\t"
"vfmaddps %%xmm4, (%7,%0,4), %%xmm15, %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%7,%0,4), %%xmm15, %%xmm5 \n\t"
"vfmaddps %%xmm6, 32(%7,%0,4), %%xmm15, %%xmm6 \n\t"
"vfmaddps %%xmm7, 48(%7,%0,4), %%xmm15, %%xmm7 \n\t"
"vmovups %%xmm4, (%3,%0,4) \n\t" // 4 * y
"vmovups %%xmm5, 16(%3,%0,4) \n\t" // 4 * y
"vmovups %%xmm6, 32(%3,%0,4) \n\t" // 4 * y
"vmovups %%xmm7, 48(%3,%0,4) \n\t" // 4 * y
"addq $16, %0 \n\t"
"subq $16, %1 \n\t"
"jnz .L01LOOP%= \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"
);
}

88
kernel/x86_64/sgemv_n_microk_haswell-2.c
View File

@ -1,88 +0,0 @@
/***************************************************************************
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 sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastss (%2), %%ymm12 \n\t" // x0
"vbroadcastss 4(%2), %%ymm13 \n\t" // x1
"vbroadcastss 8(%2), %%ymm14 \n\t" // x2
"vbroadcastss 12(%2), %%ymm15 \n\t" // x3
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vmovups (%3,%0,4), %%ymm4 \n\t" // 8 * y
"vmovups 32(%3,%0,4), %%ymm5 \n\t" // 8 * y
"prefetcht0 192(%4,%0,4) \n\t"
"vfmadd231ps (%4,%0,4), %%ymm12, %%ymm4 \n\t"
"vfmadd231ps 32(%4,%0,4), %%ymm12, %%ymm5 \n\t"
"prefetcht0 192(%5,%0,4) \n\t"
"vfmadd231ps (%5,%0,4), %%ymm13, %%ymm4 \n\t"
"vfmadd231ps 32(%5,%0,4), %%ymm13, %%ymm5 \n\t"
"prefetcht0 192(%6,%0,4) \n\t"
"vfmadd231ps (%6,%0,4), %%ymm14, %%ymm4 \n\t"
"vfmadd231ps 32(%6,%0,4), %%ymm14, %%ymm5 \n\t"
"prefetcht0 192(%7,%0,4) \n\t"
"vfmadd231ps (%7,%0,4), %%ymm15, %%ymm4 \n\t"
"vfmadd231ps 32(%7,%0,4), %%ymm15, %%ymm5 \n\t"
"vmovups %%ymm4, (%3,%0,4) \n\t" // 8 * y
"vmovups %%ymm5, 32(%3,%0,4) \n\t" // 8 * y
"addq $16, %0 \n\t"
"subq $16, %1 \n\t"
"jnz .L01LOOP%= \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",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}

144
kernel/x86_64/sgemv_n_microk_nehalem-2.c
View File

@ -1,144 +0,0 @@
/***************************************************************************
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 sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"movss (%2), %%xmm12 \n\t" // x0
"movss 4(%2), %%xmm13 \n\t" // x1
"movss 8(%2), %%xmm14 \n\t" // x2
"movss 12(%2), %%xmm15 \n\t" // x3
"shufps $0, %%xmm12, %%xmm12\n\t"
"shufps $0, %%xmm13, %%xmm13\n\t"
"shufps $0, %%xmm14, %%xmm14\n\t"
"shufps $0, %%xmm15, %%xmm15\n\t"
".align 16 \n\t"
".L01LOOP%=: \n\t"
"movups (%3,%0,4), %%xmm4 \n\t" // 4 * y
"movups 16(%3,%0,4), %%xmm5 \n\t" // 4 * y
"movups 32(%3,%0,4), %%xmm6 \n\t" // 4 * y
"movups 48(%3,%0,4), %%xmm7 \n\t" // 4 * y
"prefetcht0 192(%4,%0,4) \n\t"
"movups (%4,%0,4), %%xmm8 \n\t"
"movups 16(%4,%0,4), %%xmm9 \n\t"
"movups 32(%4,%0,4), %%xmm10 \n\t"
"movups 48(%4,%0,4), %%xmm11 \n\t"
"mulps %%xmm12, %%xmm8 \n\t"
"addps %%xmm8 , %%xmm4 \n\t"
"mulps %%xmm12, %%xmm9 \n\t"
"addps %%xmm9 , %%xmm5 \n\t"
"mulps %%xmm12, %%xmm10 \n\t"
"addps %%xmm10, %%xmm6 \n\t"
"mulps %%xmm12, %%xmm11 \n\t"
"addps %%xmm11, %%xmm7 \n\t"
"prefetcht0 192(%5,%0,4) \n\t"
"movups (%5,%0,4), %%xmm8 \n\t"
"movups 16(%5,%0,4), %%xmm9 \n\t"
"movups 32(%5,%0,4), %%xmm10 \n\t"
"movups 48(%5,%0,4), %%xmm11 \n\t"
"mulps %%xmm13, %%xmm8 \n\t"
"addps %%xmm8 , %%xmm4 \n\t"
"mulps %%xmm13, %%xmm9 \n\t"
"addps %%xmm9 , %%xmm5 \n\t"
"mulps %%xmm13, %%xmm10 \n\t"
"addps %%xmm10, %%xmm6 \n\t"
"mulps %%xmm13, %%xmm11 \n\t"
"addps %%xmm11, %%xmm7 \n\t"
"prefetcht0 192(%6,%0,4) \n\t"
"movups (%6,%0,4), %%xmm8 \n\t"
"movups 16(%6,%0,4), %%xmm9 \n\t"
"movups 32(%6,%0,4), %%xmm10 \n\t"
"movups 48(%6,%0,4), %%xmm11 \n\t"
"mulps %%xmm14, %%xmm8 \n\t"
"addps %%xmm8 , %%xmm4 \n\t"
"mulps %%xmm14, %%xmm9 \n\t"
"addps %%xmm9 , %%xmm5 \n\t"
"mulps %%xmm14, %%xmm10 \n\t"
"addps %%xmm10, %%xmm6 \n\t"
"mulps %%xmm14, %%xmm11 \n\t"
"addps %%xmm11, %%xmm7 \n\t"
"prefetcht0 192(%7,%0,4) \n\t"
"movups (%7,%0,4), %%xmm8 \n\t"
"movups 16(%7,%0,4), %%xmm9 \n\t"
"movups 32(%7,%0,4), %%xmm10 \n\t"
"movups 48(%7,%0,4), %%xmm11 \n\t"
"mulps %%xmm15, %%xmm8 \n\t"
"addps %%xmm8 , %%xmm4 \n\t"
"mulps %%xmm15, %%xmm9 \n\t"
"addps %%xmm9 , %%xmm5 \n\t"
"mulps %%xmm15, %%xmm10 \n\t"
"addps %%xmm10, %%xmm6 \n\t"
"mulps %%xmm15, %%xmm11 \n\t"
"addps %%xmm11, %%xmm7 \n\t"
"movups %%xmm4, (%3,%0,4) \n\t" // 4 * y
"movups %%xmm5, 16(%3,%0,4) \n\t" // 4 * y
"movups %%xmm6, 32(%3,%0,4) \n\t" // 4 * y
"movups %%xmm7, 48(%3,%0,4) \n\t" // 4 * y
"addq $16, %0 \n\t"
"subq $16, %1 \n\t"
"jnz .L01LOOP%= \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",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}

97
kernel/x86_64/sgemv_n_microk_sandy-2.c
View File

@ -1,97 +0,0 @@
/***************************************************************************
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 sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastss (%2), %%ymm12 \n\t" // x0
"vbroadcastss 4(%2), %%ymm13 \n\t" // x1
"vbroadcastss 8(%2), %%ymm14 \n\t" // x2
"vbroadcastss 12(%2), %%ymm15 \n\t" // x3
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vmovups (%3,%0,4), %%ymm4 \n\t" // 8 * y
"vmovups 32(%3,%0,4), %%ymm5 \n\t" // 8 * y
"prefetcht0 192(%4,%0,4) \n\t"
"vmulps (%4,%0,4), %%ymm12, %%ymm8 \n\t"
"vaddps %%ymm4, %%ymm8 , %%ymm4 \n\t"
"vmulps 32(%4,%0,4), %%ymm12, %%ymm9 \n\t"
"vaddps %%ymm5, %%ymm9 , %%ymm5 \n\t"
"prefetcht0 192(%5,%0,4) \n\t"
"vmulps (%5,%0,4), %%ymm13, %%ymm10 \n\t"
"vaddps %%ymm4, %%ymm10, %%ymm4 \n\t"
"vmulps 32(%5,%0,4), %%ymm13, %%ymm11 \n\t"
"vaddps %%ymm5, %%ymm11, %%ymm5 \n\t"
"prefetcht0 192(%6,%0,4) \n\t"
"vmulps (%6,%0,4), %%ymm14, %%ymm8 \n\t"
"vaddps %%ymm4, %%ymm8 , %%ymm4 \n\t"
"vmulps 32(%6,%0,4), %%ymm14, %%ymm9 \n\t"
"vaddps %%ymm5, %%ymm9 , %%ymm5 \n\t"
"prefetcht0 192(%7,%0,4) \n\t"
"vmulps (%7,%0,4), %%ymm15, %%ymm10 \n\t"
"vaddps %%ymm4, %%ymm10, %%ymm4 \n\t"
"vmulps 32(%7,%0,4), %%ymm15, %%ymm11 \n\t"
"vaddps %%ymm5, %%ymm11, %%ymm5 \n\t"
"vmovups %%ymm4, (%3,%0,4) \n\t" // 8 * y
"vmovups %%ymm5, 32(%3,%0,4) \n\t" // 8 * y
"addq $16, %0 \n\t"
"subq $16, %1 \n\t"
"jnz .L01LOOP%= \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",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}

10
kernel/x86_64/sgemv_t.c
View File

@ -28,16 +28,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "common.h"
#if defined(BULLDOZER) || defined(PILEDRIVER)
#include "sgemv_t_microk_bulldozer-2.c"
#elif defined(HASWELL)
#include "sgemv_t_microk_haswell-2.c"
#elif defined(SANDYBRIDGE)
#include "sgemv_t_microk_sandy-2.c"
#elif defined(NEHALEM)
#include "sgemv_t_microk_nehalem-2.c"
#endif
#define NBMAX 4096
#ifndef HAVE_KERNEL_16x4

109
kernel/x86_64/sgemv_t_microk_bulldozer-2.c
View File

@ -1,109 +0,0 @@
/***************************************************************************
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 sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vxorps %%xmm4, %%xmm4, %%xmm4 \n\t"
"vxorps %%xmm5, %%xmm5, %%xmm5 \n\t"
"vxorps %%xmm6, %%xmm6, %%xmm6 \n\t"
"vxorps %%xmm7, %%xmm7, %%xmm7 \n\t"
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vmovups (%2,%0,4), %%xmm12 \n\t" // 4 * x
"prefetcht0 384(%4,%0,4) \n\t"
"vfmaddps %%xmm4, (%4,%0,4), %%xmm12, %%xmm4 \n\t"
"vfmaddps %%xmm5, (%5,%0,4), %%xmm12, %%xmm5 \n\t"
"vmovups 16(%2,%0,4), %%xmm13 \n\t" // 4 * x
"vfmaddps %%xmm6, (%6,%0,4), %%xmm12, %%xmm6 \n\t"
"vfmaddps %%xmm7, (%7,%0,4), %%xmm12, %%xmm7 \n\t"
"prefetcht0 384(%5,%0,4) \n\t"
"vfmaddps %%xmm4, 16(%4,%0,4), %%xmm13, %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%5,%0,4), %%xmm13, %%xmm5 \n\t"
"vmovups 32(%2,%0,4), %%xmm14 \n\t" // 4 * x
"vfmaddps %%xmm6, 16(%6,%0,4), %%xmm13, %%xmm6 \n\t"
"vfmaddps %%xmm7, 16(%7,%0,4), %%xmm13, %%xmm7 \n\t"
"prefetcht0 384(%6,%0,4) \n\t"
"vfmaddps %%xmm4, 32(%4,%0,4), %%xmm14, %%xmm4 \n\t"
"vfmaddps %%xmm5, 32(%5,%0,4), %%xmm14, %%xmm5 \n\t"
"vmovups 48(%2,%0,4), %%xmm15 \n\t" // 4 * x
"vfmaddps %%xmm6, 32(%6,%0,4), %%xmm14, %%xmm6 \n\t"
"vfmaddps %%xmm7, 32(%7,%0,4), %%xmm14, %%xmm7 \n\t"
"prefetcht0 384(%7,%0,4) \n\t"
"vfmaddps %%xmm4, 48(%4,%0,4), %%xmm15, %%xmm4 \n\t"
"vfmaddps %%xmm5, 48(%5,%0,4), %%xmm15, %%xmm5 \n\t"
"vfmaddps %%xmm6, 48(%6,%0,4), %%xmm15, %%xmm6 \n\t"
"vfmaddps %%xmm7, 48(%7,%0,4), %%xmm15, %%xmm7 \n\t"
"addq $16, %0 \n\t"
"subq $16, %1 \n\t"
"jnz .L01LOOP%= \n\t"
"vhaddps %%xmm4, %%xmm4, %%xmm4 \n\t"
"vhaddps %%xmm5, %%xmm5, %%xmm5 \n\t"
"vhaddps %%xmm6, %%xmm6, %%xmm6 \n\t"
"vhaddps %%xmm7, %%xmm7, %%xmm7 \n\t"
"vhaddps %%xmm4, %%xmm4, %%xmm4 \n\t"
"vhaddps %%xmm5, %%xmm5, %%xmm5 \n\t"
"vhaddps %%xmm6, %%xmm6, %%xmm6 \n\t"
"vhaddps %%xmm7, %%xmm7, %%xmm7 \n\t"
"vmovss %%xmm4, (%3) \n\t"
"vmovss %%xmm5, 4(%3) \n\t"
"vmovss %%xmm6, 8(%3) \n\t"
"vmovss %%xmm7, 12(%3) \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"
);
}

112
kernel/x86_64/sgemv_t_microk_haswell-2.c
View File

@ -1,112 +0,0 @@
/***************************************************************************
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 sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vxorps %%ymm4 , %%ymm4, %%ymm4 \n\t"
"vxorps %%ymm5 , %%ymm5, %%ymm5 \n\t"
"vxorps %%ymm6 , %%ymm6, %%ymm6 \n\t"
"vxorps %%ymm7 , %%ymm7, %%ymm7 \n\t"
".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 384(%2,%0,4) \n\t"
"vmovups (%2,%0,4), %%ymm12 \n\t" // 8 * x
"vmovups 32(%2,%0,4), %%ymm13 \n\t" // 8 * x
"prefetcht0 384(%4,%0,4) \n\t"
"vfmadd231ps (%4,%0,4), %%ymm12, %%ymm4 \n\t"
"vfmadd231ps (%5,%0,4), %%ymm12, %%ymm5 \n\t"
"prefetcht0 384(%5,%0,4) \n\t"
"vfmadd231ps 32(%4,%0,4), %%ymm13, %%ymm4 \n\t"
"vfmadd231ps 32(%5,%0,4), %%ymm13, %%ymm5 \n\t"
"prefetcht0 384(%6,%0,4) \n\t"
"vfmadd231ps (%6,%0,4), %%ymm12, %%ymm6 \n\t"
"vfmadd231ps (%7,%0,4), %%ymm12, %%ymm7 \n\t"
"prefetcht0 384(%7,%0,4) \n\t"
"vfmadd231ps 32(%6,%0,4), %%ymm13, %%ymm6 \n\t"
"vfmadd231ps 32(%7,%0,4), %%ymm13, %%ymm7 \n\t"
"addq $16, %0 \n\t"
"subq $16, %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"
"vaddps %%xmm4, %%xmm12, %%xmm4 \n\t"
"vaddps %%xmm5, %%xmm13, %%xmm5 \n\t"
"vaddps %%xmm6, %%xmm14, %%xmm6 \n\t"
"vaddps %%xmm7, %%xmm15, %%xmm7 \n\t"
"vhaddps %%xmm4, %%xmm4, %%xmm4 \n\t"
"vhaddps %%xmm5, %%xmm5, %%xmm5 \n\t"
"vhaddps %%xmm6, %%xmm6, %%xmm6 \n\t"
"vhaddps %%xmm7, %%xmm7, %%xmm7 \n\t"
"vhaddps %%xmm4, %%xmm4, %%xmm4 \n\t"
"vhaddps %%xmm5, %%xmm5, %%xmm5 \n\t"
"vhaddps %%xmm6, %%xmm6, %%xmm6 \n\t"
"vhaddps %%xmm7, %%xmm7, %%xmm7 \n\t"
"vmovss %%xmm4, (%3) \n\t"
"vmovss %%xmm5, 4(%3) \n\t"
"vmovss %%xmm6, 8(%3) \n\t"
"vmovss %%xmm7, 12(%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"
);
}

159
kernel/x86_64/sgemv_t_microk_nehalem-2.c
View File

@ -1,159 +0,0 @@
/***************************************************************************
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 sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"xorps %%xmm0 , %%xmm0 \n\t"
"xorps %%xmm1 , %%xmm1 \n\t"
"xorps %%xmm2 , %%xmm2 \n\t"
"xorps %%xmm3 , %%xmm3 \n\t"
"xorps %%xmm4 , %%xmm4 \n\t"
"xorps %%xmm5 , %%xmm5 \n\t"
"xorps %%xmm6 , %%xmm6 \n\t"
"xorps %%xmm7 , %%xmm7 \n\t"
".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 384(%2,%0,4) \n\t"
"movups (%2,%0,4), %%xmm12 \n\t" // 4 * x
"movups 16(%2,%0,4), %%xmm13 \n\t" // 4 * x
"movups (%4,%0,4), %%xmm8 \n\t"
"movups 32(%2,%0,4), %%xmm14 \n\t" // 4 * x
"movups 48(%2,%0,4), %%xmm15 \n\t" // 4 * x
"prefetcht0 384(%4,%0,4) \n\t"
"movups 16(%4,%0,4), %%xmm9 \n\t"
"movups 32(%4,%0,4), %%xmm10 \n\t"
"movups 48(%4,%0,4), %%xmm11 \n\t"
"mulps %%xmm12, %%xmm8 \n\t"
"addps %%xmm8 , %%xmm0 \n\t"
"mulps %%xmm13, %%xmm9 \n\t"
"addps %%xmm9 , %%xmm4 \n\t"
"movups (%5,%0,4), %%xmm8 \n\t"
"mulps %%xmm14, %%xmm10 \n\t"
"addps %%xmm10, %%xmm0 \n\t"
"mulps %%xmm15, %%xmm11 \n\t"
"addps %%xmm11, %%xmm4 \n\t"
"prefetcht0 384(%5,%0,4) \n\t"
"movups 16(%5,%0,4), %%xmm9 \n\t"
"movups 32(%5,%0,4), %%xmm10 \n\t"
"movups 48(%5,%0,4), %%xmm11 \n\t"
"mulps %%xmm12, %%xmm8 \n\t"
"addps %%xmm8 , %%xmm1 \n\t"
"mulps %%xmm13, %%xmm9 \n\t"
"addps %%xmm9 , %%xmm5 \n\t"
"movups (%6,%0,4), %%xmm8 \n\t"
"mulps %%xmm14, %%xmm10 \n\t"
"addps %%xmm10, %%xmm1 \n\t"
"mulps %%xmm15, %%xmm11 \n\t"
"addps %%xmm11, %%xmm5 \n\t"
"prefetcht0 384(%6,%0,4) \n\t"
"movups 16(%6,%0,4), %%xmm9 \n\t"
"movups 32(%6,%0,4), %%xmm10 \n\t"
"movups 48(%6,%0,4), %%xmm11 \n\t"
"mulps %%xmm12, %%xmm8 \n\t"
"addps %%xmm8 , %%xmm2 \n\t"
"mulps %%xmm13, %%xmm9 \n\t"
"addps %%xmm9 , %%xmm6 \n\t"
"movups (%7,%0,4), %%xmm8 \n\t"
"mulps %%xmm14, %%xmm10 \n\t"
"addps %%xmm10, %%xmm2 \n\t"
"mulps %%xmm15, %%xmm11 \n\t"
"addps %%xmm11, %%xmm6 \n\t"
"prefetcht0 384(%7,%0,4) \n\t"
"movups 16(%7,%0,4), %%xmm9 \n\t"
"movups 32(%7,%0,4), %%xmm10 \n\t"
"movups 48(%7,%0,4), %%xmm11 \n\t"
"mulps %%xmm12, %%xmm8 \n\t"
"addps %%xmm8 , %%xmm3 \n\t"
"mulps %%xmm13, %%xmm9 \n\t"
"addps %%xmm9 , %%xmm7 \n\t"
"mulps %%xmm14, %%xmm10 \n\t"
"addps %%xmm10, %%xmm3 \n\t"
"mulps %%xmm15, %%xmm11 \n\t"
"addps %%xmm11, %%xmm7 \n\t"
"addq $16, %0 \n\t"
"subq $16, %1 \n\t"
"jnz .L01LOOP%= \n\t"
"addps %%xmm0, %%xmm4 \n\t"
"addps %%xmm1, %%xmm5 \n\t"
"addps %%xmm2, %%xmm6 \n\t"
"addps %%xmm3, %%xmm7 \n\t"
"haddps %%xmm4, %%xmm4 \n\t"
"haddps %%xmm5, %%xmm5 \n\t"
"haddps %%xmm6, %%xmm6 \n\t"
"haddps %%xmm7, %%xmm7 \n\t"
"haddps %%xmm4, %%xmm4 \n\t"
"haddps %%xmm5, %%xmm5 \n\t"
"haddps %%xmm6, %%xmm6 \n\t"
"haddps %%xmm7, %%xmm7 \n\t"
"movss %%xmm4, (%3) \n\t"
"movss %%xmm5, 4(%3) \n\t"
"movss %%xmm6, 8(%3) \n\t"
"movss %%xmm7, 12(%3) \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",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}

132
kernel/x86_64/sgemv_t_microk_sandy-2.c
View File

@ -1,132 +0,0 @@
/***************************************************************************
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 sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vxorps %%ymm0 , %%ymm0, %%ymm0 \n\t"
"vxorps %%ymm1 , %%ymm1, %%ymm1 \n\t"
"vxorps %%ymm2 , %%ymm2, %%ymm2 \n\t"
"vxorps %%ymm3 , %%ymm3, %%ymm3 \n\t"
"vxorps %%ymm4 , %%ymm4, %%ymm4 \n\t"
"vxorps %%ymm5 , %%ymm5, %%ymm5 \n\t"
"vxorps %%ymm6 , %%ymm6, %%ymm6 \n\t"
"vxorps %%ymm7 , %%ymm7, %%ymm7 \n\t"
".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 384(%2,%0,4) \n\t"
"vmovups (%2,%0,4), %%ymm12 \n\t" // 8 * x
"vmovups 32(%2,%0,4), %%ymm13 \n\t" // 8 * x
"prefetcht0 384(%4,%0,4) \n\t"
"vmulps (%4,%0,4), %%ymm12, %%ymm8 \n\t"
"vaddps %%ymm4, %%ymm8 , %%ymm4 \n\t"
"vmulps 32(%4,%0,4), %%ymm13, %%ymm9 \n\t"
"vaddps %%ymm0, %%ymm9 , %%ymm0 \n\t"
"prefetcht0 384(%5,%0,4) \n\t"
"vmulps (%5,%0,4), %%ymm12, %%ymm10 \n\t"
"vaddps %%ymm1, %%ymm10, %%ymm1 \n\t"
"vmulps 32(%5,%0,4), %%ymm13, %%ymm11 \n\t"
"vaddps %%ymm5, %%ymm11, %%ymm5 \n\t"
"prefetcht0 384(%6,%0,4) \n\t"
"vmulps (%6,%0,4), %%ymm12, %%ymm8 \n\t"
"vaddps %%ymm6, %%ymm8 , %%ymm6 \n\t"
"vmulps 32(%6,%0,4), %%ymm13, %%ymm9 \n\t"
"vaddps %%ymm2, %%ymm9 , %%ymm2 \n\t"
"prefetcht0 384(%7,%0,4) \n\t"
"vmulps (%7,%0,4), %%ymm12, %%ymm10 \n\t"
"vaddps %%ymm7, %%ymm10, %%ymm7 \n\t"
"vmulps 32(%7,%0,4), %%ymm13, %%ymm11 \n\t"
"vaddps %%ymm3, %%ymm11, %%ymm3 \n\t"
"addq $16, %0 \n\t"
"subq $16, %1 \n\t"
"jnz .L01LOOP%= \n\t"
"vaddps %%ymm4, %%ymm0, %%ymm4 \n\t"
"vaddps %%ymm5, %%ymm1, %%ymm5 \n\t"
"vaddps %%ymm6, %%ymm2, %%ymm6 \n\t"
"vaddps %%ymm7, %%ymm3, %%ymm7 \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"
"vaddps %%xmm4, %%xmm12, %%xmm4 \n\t"
"vaddps %%xmm5, %%xmm13, %%xmm5 \n\t"
"vaddps %%xmm6, %%xmm14, %%xmm6 \n\t"
"vaddps %%xmm7, %%xmm15, %%xmm7 \n\t"
"vhaddps %%xmm4, %%xmm4, %%xmm4 \n\t"
"vhaddps %%xmm5, %%xmm5, %%xmm5 \n\t"
"vhaddps %%xmm6, %%xmm6, %%xmm6 \n\t"
"vhaddps %%xmm7, %%xmm7, %%xmm7 \n\t"
"vhaddps %%xmm4, %%xmm4, %%xmm4 \n\t"
"vhaddps %%xmm5, %%xmm5, %%xmm5 \n\t"
"vhaddps %%xmm6, %%xmm6, %%xmm6 \n\t"
"vhaddps %%xmm7, %%xmm7, %%xmm7 \n\t"
"vmovss %%xmm4, (%3) \n\t"
"vmovss %%xmm5, 4(%3) \n\t"
"vmovss %%xmm6, 8(%3) \n\t"
"vmovss %%xmm7, 12(%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",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}

258
kernel/x86_64/zgemv_n.c
View File

@ -1,258 +0,0 @@
/***************************************************************************
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 <stdlib.h>
#include <stdio.h>
#include "common.h"
#if defined(HASWELL)
#include "zgemv_n_microk_haswell-2.c"
#elif defined(SANDYBRIDGE)
#include "zgemv_n_microk_sandy-2.c"
#endif
#define NBMAX 1024
#ifndef HAVE_KERNEL_16x4
static void zgemv_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];
for ( i=0; i< 2*n; i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
y[i] += a0[i]*x[0] - a0[i+1] * x[1];
y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
y[i] += a1[i]*x[2] - a1[i+1] * x[3];
y[i+1] += a1[i]*x[3] + a1[i+1] * x[2];
y[i] += a2[i]*x[4] - a2[i+1] * x[5];
y[i+1] += a2[i]*x[5] + a2[i+1] * x[4];
y[i] += a3[i]*x[6] - a3[i+1] * x[7];
y[i+1] += a3[i]*x[7] + a3[i+1] * x[6];
#else
y[i] += a0[i]*x[0] + a0[i+1] * x[1];
y[i+1] += a0[i]*x[1] - a0[i+1] * x[0];
y[i] += a1[i]*x[2] + a1[i+1] * x[3];
y[i+1] += a1[i]*x[3] - a1[i+1] * x[2];
y[i] += a2[i]*x[4] + a2[i+1] * x[5];
y[i+1] += a2[i]*x[5] - a2[i+1] * x[4];
y[i] += a3[i]*x[6] + a3[i+1] * x[7];
y[i+1] += a3[i]*x[7] - a3[i+1] * x[6];
#endif
}
}
#endif
static void zgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
{
BLASLONG i;
FLOAT *a0;
a0 = ap;
for ( i=0; i< 2*n; i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
y[i] += a0[i]*x[0] - a0[i+1] * x[1];
y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
#else
y[i] += a0[i]*x[0] + a0[i+1] * x[1];
y[i+1] += a0[i]*x[1] - a0[i+1] * x[0];
#endif
}
}
static void zero_y(BLASLONG n, FLOAT *dest)
{
BLASLONG i;
for ( i=0; i<2*n; i++ )
{
*dest = 0.0;
dest++;
}
}
static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i)
{
BLASLONG i;
FLOAT temp_r;
FLOAT temp_i;
for ( i=0; i<n; i++ )
{
#if !defined(XCONJ)
temp_r = alpha_r * src[0] - alpha_i * src[1];
temp_i = alpha_r * src[1] + alpha_i * src[0];
#else
temp_r = alpha_r * src[0] + alpha_i * src[1];
temp_i = -alpha_r * src[1] + alpha_i * src[0];
#endif
*dest += temp_r;
*(dest+1) += temp_i;
src+=2;
dest += inc_dest;
}
}
int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r,FLOAT alpha_i, 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 xbuffer[8],*ybuffer;
#if 0
printf("%s %d %d %.16f %.16f %d %d %d\n","zgemv_n",m,n,alpha_r,alpha_i,lda,inc_x,inc_y);
#endif
if ( m < 1 ) return(0);
if ( n < 1 ) return(0);
ybuffer = buffer;
inc_x *= 2;
inc_y *= 2;
lda *= 2;
n1 = n / 4 ;
n2 = n % 4 ;
m1 = m - ( m % 16 );
m2 = (m % NBMAX) - (m % 16) ;
y_ptr = y;
BLASLONG NB = NBMAX;
while ( NB == NBMAX )
{
m1 -= NB;
if ( m1 < 0)
{
if ( m2 == 0 ) break;
NB = m2;
}
a_ptr = a;
x_ptr = x;
zero_y(NB,ybuffer);
for( i = 0; i < n1 ; i++)
{
xbuffer[0] = x_ptr[0];
xbuffer[1] = x_ptr[1];
x_ptr += inc_x;
xbuffer[2] = x_ptr[0];
xbuffer[3] = x_ptr[1];
x_ptr += inc_x;
xbuffer[4] = x_ptr[0];
xbuffer[5] = x_ptr[1];
x_ptr += inc_x;
xbuffer[6] = x_ptr[0];
xbuffer[7] = x_ptr[1];
x_ptr += inc_x;
ap[0] = a_ptr;
ap[1] = a_ptr + lda;
ap[2] = ap[1] + lda;
ap[3] = ap[2] + lda;
zgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
a_ptr += 4 * lda;
}
for( i = 0; i < n2 ; i++)
{
xbuffer[0] = x_ptr[0];
xbuffer[1] = x_ptr[1];
x_ptr += inc_x;
zgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
a_ptr += 1 * lda;
}
add_y(NB,ybuffer,y_ptr,inc_y,alpha_r,alpha_i);
a += 2 * NB;
y_ptr += NB * inc_y;
}
j=0;
while ( j < (m % 16))
{
a_ptr = a;
x_ptr = x;
FLOAT temp_r = 0.0;
FLOAT temp_i = 0.0;
for( i = 0; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
#else
temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
#endif
a_ptr += lda;
x_ptr += inc_x;
}
#if !defined(XCONJ)
y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
#else
y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
#endif
y_ptr += inc_y;
a+=2;
j++;
}
return(0);
}

626
kernel/x86_64/zgemv_n_4.c Normal file
View File

@ -0,0 +1,626 @@
/***************************************************************************
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 <stdlib.h>
#include <stdio.h>
#include "common.h"
#if defined(HASWELL)
#include "zgemv_n_microk_haswell-4.c"
#elif defined(SANDYBRIDGE)
#include "zgemv_n_microk_sandy-4.c"
#endif
#define NBMAX 1024
#ifndef HAVE_KERNEL_4x4
static void zgemv_kernel_4x4(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];
for ( i=0; i< 2*n; i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
y[i] += a0[i]*x[0] - a0[i+1] * x[1];
y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
y[i] += a1[i]*x[2] - a1[i+1] * x[3];
y[i+1] += a1[i]*x[3] + a1[i+1] * x[2];
y[i] += a2[i]*x[4] - a2[i+1] * x[5];
y[i+1] += a2[i]*x[5] + a2[i+1] * x[4];
y[i] += a3[i]*x[6] - a3[i+1] * x[7];
y[i+1] += a3[i]*x[7] + a3[i+1] * x[6];
#else
y[i] += a0[i]*x[0] + a0[i+1] * x[1];
y[i+1] += a0[i]*x[1] - a0[i+1] * x[0];
y[i] += a1[i]*x[2] + a1[i+1] * x[3];
y[i+1] += a1[i]*x[3] - a1[i+1] * x[2];
y[i] += a2[i]*x[4] + a2[i+1] * x[5];
y[i+1] += a2[i]*x[5] - a2[i+1] * x[4];
y[i] += a3[i]*x[6] + a3[i+1] * x[7];
y[i+1] += a3[i]*x[7] - a3[i+1] * x[6];
#endif
}
}
#endif
#ifndef HAVE_KERNEL_4x2
static void zgemv_kernel_4x2(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
BLASLONG i;
FLOAT *a0,*a1;
a0 = ap[0];
a1 = ap[1];
for ( i=0; i< 2*n; i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
y[i] += a0[i]*x[0] - a0[i+1] * x[1];
y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
y[i] += a1[i]*x[2] - a1[i+1] * x[3];
y[i+1] += a1[i]*x[3] + a1[i+1] * x[2];
#else
y[i] += a0[i]*x[0] + a0[i+1] * x[1];
y[i+1] += a0[i]*x[1] - a0[i+1] * x[0];
y[i] += a1[i]*x[2] + a1[i+1] * x[3];
y[i+1] += a1[i]*x[3] - a1[i+1] * x[2];
#endif
}
}
#endif
#ifndef HAVE_KERNEL_4x1
static void zgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
{
BLASLONG i;
FLOAT *a0;
a0 = ap;
for ( i=0; i< 2*n; i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
y[i] += a0[i]*x[0] - a0[i+1] * x[1];
y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
#else
y[i] += a0[i]*x[0] + a0[i+1] * x[1];
y[i+1] += a0[i]*x[1] - a0[i+1] * x[0];
#endif
}
}
#endif
#ifndef HAVE_KERNEL_ADDY
static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i) __attribute__ ((noinline));
static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i)
{
BLASLONG i;
if ( inc_dest != 2 )
{
FLOAT temp_r;
FLOAT temp_i;
for ( i=0; i<n; i++ )
{
#if !defined(XCONJ)
temp_r = alpha_r * src[0] - alpha_i * src[1];
temp_i = alpha_r * src[1] + alpha_i * src[0];
#else
temp_r = alpha_r * src[0] + alpha_i * src[1];
temp_i = -alpha_r * src[1] + alpha_i * src[0];
#endif
*dest += temp_r;
*(dest+1) += temp_i;
src+=2;
dest += inc_dest;
}
return;
}
FLOAT temp_r0;
FLOAT temp_i0;
FLOAT temp_r1;
FLOAT temp_i1;
FLOAT temp_r2;
FLOAT temp_i2;
FLOAT temp_r3;
FLOAT temp_i3;
for ( i=0; i<n; i+=4 )
{
#if !defined(XCONJ)
temp_r0 = alpha_r * src[0] - alpha_i * src[1];
temp_i0 = alpha_r * src[1] + alpha_i * src[0];
temp_r1 = alpha_r * src[2] - alpha_i * src[3];
temp_i1 = alpha_r * src[3] + alpha_i * src[2];
temp_r2 = alpha_r * src[4] - alpha_i * src[5];
temp_i2 = alpha_r * src[5] + alpha_i * src[4];
temp_r3 = alpha_r * src[6] - alpha_i * src[7];
temp_i3 = alpha_r * src[7] + alpha_i * src[6];
#else
temp_r0 = alpha_r * src[0] + alpha_i * src[1];
temp_i0 = -alpha_r * src[1] + alpha_i * src[0];
temp_r1 = alpha_r * src[2] + alpha_i * src[3];
temp_i1 = -alpha_r * src[3] + alpha_i * src[2];
temp_r2 = alpha_r * src[4] + alpha_i * src[5];
temp_i2 = -alpha_r * src[5] + alpha_i * src[4];
temp_r3 = alpha_r * src[6] + alpha_i * src[7];
temp_i3 = -alpha_r * src[7] + alpha_i * src[6];
#endif
dest[0] += temp_r0;
dest[1] += temp_i0;
dest[2] += temp_r1;
dest[3] += temp_i1;
dest[4] += temp_r2;
dest[5] += temp_i2;
dest[6] += temp_r3;
dest[7] += temp_i3;
src += 8;
dest += 8;
}
return;
}
#endif
int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r,FLOAT alpha_i, 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 m3;
BLASLONG n2;
BLASLONG lda4;
FLOAT xbuffer[8],*ybuffer;
#if 0
printf("%s %d %d %.16f %.16f %d %d %d\n","zgemv_n",m,n,alpha_r,alpha_i,lda,inc_x,inc_y);
#endif
if ( m < 1 ) return(0);
if ( n < 1 ) return(0);
ybuffer = buffer;
inc_x *= 2;
inc_y *= 2;
lda *= 2;
lda4 = 4 * lda;
n1 = n / 4 ;
n2 = n % 4 ;
m3 = m % 4;
m1 = m - ( m % 4 );
m2 = (m % NBMAX) - (m % 4) ;
y_ptr = y;
BLASLONG NB = NBMAX;
while ( NB == NBMAX )
{
m1 -= NB;
if ( m1 < 0)
{
if ( m2 == 0 ) break;
NB = m2;
}
a_ptr = a;
ap[0] = a_ptr;
ap[1] = a_ptr + lda;
ap[2] = ap[1] + lda;
ap[3] = ap[2] + lda;
x_ptr = x;
//zero_y(NB,ybuffer);
memset(ybuffer,0,NB*16);
if ( inc_x == 2 )
{
for( i = 0; i < n1 ; i++)
{
zgemv_kernel_4x4(NB,ap,x_ptr,ybuffer);
ap[0] += lda4;
ap[1] += lda4;
ap[2] += lda4;
ap[3] += lda4;
a_ptr += lda4;
x_ptr += 8;
}
if ( n2 & 2 )
{
zgemv_kernel_4x2(NB,ap,x_ptr,ybuffer);
x_ptr += 4;
a_ptr += 2 * lda;
}
if ( n2 & 1 )
{
zgemv_kernel_4x1(NB,a_ptr,x_ptr,ybuffer);
x_ptr += 2;
a_ptr += lda;
}
}
else
{
for( i = 0; i < n1 ; i++)
{
xbuffer[0] = x_ptr[0];
xbuffer[1] = x_ptr[1];
x_ptr += inc_x;
xbuffer[2] = x_ptr[0];
xbuffer[3] = x_ptr[1];
x_ptr += inc_x;
xbuffer[4] = x_ptr[0];
xbuffer[5] = x_ptr[1];
x_ptr += inc_x;
xbuffer[6] = x_ptr[0];
xbuffer[7] = x_ptr[1];
x_ptr += inc_x;
zgemv_kernel_4x4(NB,ap,xbuffer,ybuffer);
ap[0] += lda4;
ap[1] += lda4;
ap[2] += lda4;
ap[3] += lda4;
a_ptr += lda4;
}
for( i = 0; i < n2 ; i++)
{
xbuffer[0] = x_ptr[0];
xbuffer[1] = x_ptr[1];
x_ptr += inc_x;
zgemv_kernel_4x1(NB,a_ptr,xbuffer,ybuffer);
a_ptr += 1 * lda;
}
}
add_y(NB,ybuffer,y_ptr,inc_y,alpha_r,alpha_i);
a += 2 * NB;
y_ptr += NB * inc_y;
}
if ( m3 == 0 ) return(0);
if ( m3 == 1 )
{
a_ptr = a;
x_ptr = x;
FLOAT temp_r = 0.0;
FLOAT temp_i = 0.0;
if ( lda == 2 && inc_x == 2 )
{
for( i=0 ; i < (n & -2); i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r += a_ptr[2] * x_ptr[2] - a_ptr[3] * x_ptr[3];
temp_i += a_ptr[2] * x_ptr[3] + a_ptr[3] * x_ptr[2];
#else
temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r += a_ptr[2] * x_ptr[2] + a_ptr[3] * x_ptr[3];
temp_i += a_ptr[2] * x_ptr[3] - a_ptr[3] * x_ptr[2];
#endif
a_ptr += 4;
x_ptr += 4;
}
for( ; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
#else
temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
#endif
a_ptr += 2;
x_ptr += 2;
}
}
else
{
for( i = 0; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
#else
temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
#endif
a_ptr += lda;
x_ptr += inc_x;
}
}
#if !defined(XCONJ)
y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
#else
y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
#endif
return(0);
}
if ( m3 == 2 )
{
a_ptr = a;
x_ptr = x;
FLOAT temp_r0 = 0.0;
FLOAT temp_i0 = 0.0;
FLOAT temp_r1 = 0.0;
FLOAT temp_i1 = 0.0;
if ( lda == 4 && inc_x == 2 )
{
for( i = 0; i < (n & -2); i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
temp_r0 += a_ptr[4] * x_ptr[2] - a_ptr[5] * x_ptr[3];
temp_i0 += a_ptr[4] * x_ptr[3] + a_ptr[5] * x_ptr[2];
temp_r1 += a_ptr[6] * x_ptr[2] - a_ptr[7] * x_ptr[3];
temp_i1 += a_ptr[6] * x_ptr[3] + a_ptr[7] * x_ptr[2];
#else
temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
temp_r0 += a_ptr[4] * x_ptr[2] + a_ptr[5] * x_ptr[3];
temp_i0 += a_ptr[4] * x_ptr[3] - a_ptr[5] * x_ptr[2];
temp_r1 += a_ptr[6] * x_ptr[2] + a_ptr[7] * x_ptr[3];
temp_i1 += a_ptr[6] * x_ptr[3] - a_ptr[7] * x_ptr[2];
#endif
a_ptr += 8;
x_ptr += 4;
}
for( ; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
#else
temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
#endif
a_ptr += 4;
x_ptr += 2;
}
}
else
{
for( i=0 ; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
#else
temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
#endif
a_ptr += lda;
x_ptr += inc_x;
}
}
#if !defined(XCONJ)
y_ptr[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
y_ptr[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r1 - alpha_i * temp_i1;
y_ptr[1] += alpha_r * temp_i1 + alpha_i * temp_r1;
#else
y_ptr[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
y_ptr[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r1 + alpha_i * temp_i1;
y_ptr[1] -= alpha_r * temp_i1 - alpha_i * temp_r1;
#endif
return(0);
}
if ( m3 == 3 )
{
a_ptr = a;
x_ptr = x;
FLOAT temp_r0 = 0.0;
FLOAT temp_i0 = 0.0;
FLOAT temp_r1 = 0.0;
FLOAT temp_i1 = 0.0;
FLOAT temp_r2 = 0.0;
FLOAT temp_i2 = 0.0;
if ( lda == 6 && inc_x == 2 )
{
for( i=0 ; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
temp_r2 += a_ptr[4] * x_ptr[0] - a_ptr[5] * x_ptr[1];
temp_i2 += a_ptr[4] * x_ptr[1] + a_ptr[5] * x_ptr[0];
#else
temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
temp_r2 += a_ptr[4] * x_ptr[0] + a_ptr[5] * x_ptr[1];
temp_i2 += a_ptr[4] * x_ptr[1] - a_ptr[5] * x_ptr[0];
#endif
a_ptr += 6;
x_ptr += 2;
}
}
else
{
for( i = 0; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
temp_r2 += a_ptr[4] * x_ptr[0] - a_ptr[5] * x_ptr[1];
temp_i2 += a_ptr[4] * x_ptr[1] + a_ptr[5] * x_ptr[0];
#else
temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
temp_r2 += a_ptr[4] * x_ptr[0] + a_ptr[5] * x_ptr[1];
temp_i2 += a_ptr[4] * x_ptr[1] - a_ptr[5] * x_ptr[0];
#endif
a_ptr += lda;
x_ptr += inc_x;
}
}
#if !defined(XCONJ)
y_ptr[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
y_ptr[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r1 - alpha_i * temp_i1;
y_ptr[1] += alpha_r * temp_i1 + alpha_i * temp_r1;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r2 - alpha_i * temp_i2;
y_ptr[1] += alpha_r * temp_i2 + alpha_i * temp_r2;
#else
y_ptr[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
y_ptr[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r1 + alpha_i * temp_i1;
y_ptr[1] -= alpha_r * temp_i1 - alpha_i * temp_r1;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r2 + alpha_i * temp_i2;
y_ptr[1] -= alpha_r * temp_i2 - alpha_i * temp_r2;
#endif
return(0);
}
return(0);
}

137
kernel/x86_64/zgemv_n_microk_haswell-2.c
View File

@ -1,137 +0,0 @@
/***************************************************************************
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 zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastsd (%2), %%ymm0 \n\t" // real part x0
"vbroadcastsd 8(%2), %%ymm1 \n\t" // imag part x0
"vbroadcastsd 16(%2), %%ymm2 \n\t" // real part x1
"vbroadcastsd 24(%2), %%ymm3 \n\t" // imag part x1
"vbroadcastsd 32(%2), %%ymm4 \n\t" // real part x2
"vbroadcastsd 40(%2), %%ymm5 \n\t" // imag part x2
"vbroadcastsd 48(%2), %%ymm6 \n\t" // real part x3
"vbroadcastsd 56(%2), %%ymm7 \n\t" // imag part x3
".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 192(%4,%0,8) \n\t"
"vmovups (%4,%0,8), %%ymm8 \n\t" // 2 complex values form a0
"vmovups 32(%4,%0,8), %%ymm9 \n\t" // 2 complex values form a0
"prefetcht0 192(%5,%0,8) \n\t"
"vmovups (%5,%0,8), %%ymm10 \n\t" // 2 complex values form a1
"vmovups 32(%5,%0,8), %%ymm11 \n\t" // 2 complex values form a1
"vmulpd %%ymm8 , %%ymm0, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vmulpd %%ymm8 , %%ymm1, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vmulpd %%ymm9 , %%ymm0, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vmulpd %%ymm9 , %%ymm1, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"prefetcht0 192(%6,%0,8) \n\t"
"vmovups (%6,%0,8), %%ymm8 \n\t" // 2 complex values form a2
"vmovups 32(%6,%0,8), %%ymm9 \n\t" // 2 complex values form a2
"vfmadd231pd %%ymm10, %%ymm2, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vfmadd231pd %%ymm10, %%ymm3, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vfmadd231pd %%ymm11, %%ymm2, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vfmadd231pd %%ymm11, %%ymm3, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"prefetcht0 192(%7,%0,8) \n\t"
"vmovups (%7,%0,8), %%ymm10 \n\t" // 2 complex values form a3
"vmovups 32(%7,%0,8), %%ymm11 \n\t" // 2 complex values form a3
"vfmadd231pd %%ymm8 , %%ymm4, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vfmadd231pd %%ymm8 , %%ymm5, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vfmadd231pd %%ymm9 , %%ymm4, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vfmadd231pd %%ymm9 , %%ymm5, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"vfmadd231pd %%ymm10, %%ymm6, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vfmadd231pd %%ymm10, %%ymm7, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vfmadd231pd %%ymm11, %%ymm6, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vfmadd231pd %%ymm11, %%ymm7, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"prefetcht0 192(%3,%0,8) \n\t"
"vmovups (%3,%0,8), %%ymm10 \n\t"
"vmovups 32(%3,%0,8), %%ymm11 \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilpd $0x5 , %%ymm13, %%ymm13 \n\t"
"vpermilpd $0x5 , %%ymm15, %%ymm15 \n\t"
"vaddsubpd %%ymm13, %%ymm12, %%ymm8 \n\t"
"vaddsubpd %%ymm15, %%ymm14, %%ymm9 \n\t"
#else
"vpermilpd $0x5 , %%ymm12, %%ymm12 \n\t"
"vpermilpd $0x5 , %%ymm14, %%ymm14 \n\t"
"vaddsubpd %%ymm12, %%ymm13, %%ymm8 \n\t"
"vaddsubpd %%ymm14, %%ymm15, %%ymm9 \n\t"
"vpermilpd $0x5 , %%ymm8 , %%ymm8 \n\t"
"vpermilpd $0x5 , %%ymm9 , %%ymm9 \n\t"
#endif
"vaddpd %%ymm8, %%ymm10, %%ymm12 \n\t"
"vaddpd %%ymm9, %%ymm11, %%ymm13 \n\t"
"vmovups %%ymm12, (%3,%0,8) \n\t" // 2 complex values to y
"vmovups %%ymm13, 32(%3,%0,8) \n\t"
"addq $8 , %0 \n\t"
"subq $4 , %1 \n\t"
"jnz .L01LOOP%= \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",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}

400
kernel/x86_64/zgemv_n_microk_haswell-4.c Normal file
View File

@ -0,0 +1,400 @@
/***************************************************************************
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_4x4 1
static void zgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void zgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastsd (%2), %%ymm0 \n\t" // real part x0
"vbroadcastsd 8(%2), %%ymm1 \n\t" // imag part x0
"vbroadcastsd 16(%2), %%ymm2 \n\t" // real part x1
"vbroadcastsd 24(%2), %%ymm3 \n\t" // imag part x1
"vbroadcastsd 32(%2), %%ymm4 \n\t" // real part x2
"vbroadcastsd 40(%2), %%ymm5 \n\t" // imag part x2
"vbroadcastsd 48(%2), %%ymm6 \n\t" // real part x3
"vbroadcastsd 56(%2), %%ymm7 \n\t" // imag part x3
".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 192(%4,%0,8) \n\t"
"vmovups (%4,%0,8), %%ymm8 \n\t" // 2 complex values form a0
"vmovups 32(%4,%0,8), %%ymm9 \n\t" // 2 complex values form a0
"prefetcht0 192(%5,%0,8) \n\t"
"vmovups (%5,%0,8), %%ymm10 \n\t" // 2 complex values form a1
"vmovups 32(%5,%0,8), %%ymm11 \n\t" // 2 complex values form a1
"vmulpd %%ymm8 , %%ymm0, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vmulpd %%ymm8 , %%ymm1, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vmulpd %%ymm9 , %%ymm0, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vmulpd %%ymm9 , %%ymm1, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"prefetcht0 192(%6,%0,8) \n\t"
"vmovups (%6,%0,8), %%ymm8 \n\t" // 2 complex values form a2
"vmovups 32(%6,%0,8), %%ymm9 \n\t" // 2 complex values form a2
"vfmadd231pd %%ymm10, %%ymm2, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vfmadd231pd %%ymm10, %%ymm3, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vfmadd231pd %%ymm11, %%ymm2, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vfmadd231pd %%ymm11, %%ymm3, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"prefetcht0 192(%7,%0,8) \n\t"
"vmovups (%7,%0,8), %%ymm10 \n\t" // 2 complex values form a3
"vmovups 32(%7,%0,8), %%ymm11 \n\t" // 2 complex values form a3
"vfmadd231pd %%ymm8 , %%ymm4, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vfmadd231pd %%ymm8 , %%ymm5, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vfmadd231pd %%ymm9 , %%ymm4, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vfmadd231pd %%ymm9 , %%ymm5, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"vfmadd231pd %%ymm10, %%ymm6, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vfmadd231pd %%ymm10, %%ymm7, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vfmadd231pd %%ymm11, %%ymm6, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vfmadd231pd %%ymm11, %%ymm7, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"prefetcht0 192(%3,%0,8) \n\t"
"vmovups (%3,%0,8), %%ymm10 \n\t"
"vmovups 32(%3,%0,8), %%ymm11 \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilpd $0x5 , %%ymm13, %%ymm13 \n\t"
"vpermilpd $0x5 , %%ymm15, %%ymm15 \n\t"
"vaddsubpd %%ymm13, %%ymm12, %%ymm8 \n\t"
"vaddsubpd %%ymm15, %%ymm14, %%ymm9 \n\t"
#else
"vpermilpd $0x5 , %%ymm12, %%ymm12 \n\t"
"vpermilpd $0x5 , %%ymm14, %%ymm14 \n\t"
"vaddsubpd %%ymm12, %%ymm13, %%ymm8 \n\t"
"vaddsubpd %%ymm14, %%ymm15, %%ymm9 \n\t"
"vpermilpd $0x5 , %%ymm8 , %%ymm8 \n\t"
"vpermilpd $0x5 , %%ymm9 , %%ymm9 \n\t"
#endif
"vaddpd %%ymm8, %%ymm10, %%ymm12 \n\t"
"vaddpd %%ymm9, %%ymm11, %%ymm13 \n\t"
"vmovups %%ymm12, (%3,%0,8) \n\t" // 2 complex values to y
"vmovups %%ymm13, 32(%3,%0,8) \n\t"
"addq $8 , %0 \n\t"
"subq $4 , %1 \n\t"
"jnz .L01LOOP%= \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",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
#define HAVE_KERNEL_4x2 1
static void zgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void zgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastsd (%2), %%ymm0 \n\t" // real part x0
"vbroadcastsd 8(%2), %%ymm1 \n\t" // imag part x0
"vbroadcastsd 16(%2), %%ymm2 \n\t" // real part x1
"vbroadcastsd 24(%2), %%ymm3 \n\t" // imag part x1
".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 192(%4,%0,8) \n\t"
"vmovups (%4,%0,8), %%ymm8 \n\t" // 2 complex values form a0
"vmovups 32(%4,%0,8), %%ymm9 \n\t" // 2 complex values form a0
"prefetcht0 192(%5,%0,8) \n\t"
"vmovups (%5,%0,8), %%ymm10 \n\t" // 2 complex values form a1
"vmovups 32(%5,%0,8), %%ymm11 \n\t" // 2 complex values form a1
"vmulpd %%ymm8 , %%ymm0, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vmulpd %%ymm8 , %%ymm1, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vmulpd %%ymm9 , %%ymm0, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vmulpd %%ymm9 , %%ymm1, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"vfmadd231pd %%ymm10, %%ymm2, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vfmadd231pd %%ymm10, %%ymm3, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vfmadd231pd %%ymm11, %%ymm2, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vfmadd231pd %%ymm11, %%ymm3, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"prefetcht0 192(%3,%0,8) \n\t"
"vmovups (%3,%0,8), %%ymm10 \n\t"
"vmovups 32(%3,%0,8), %%ymm11 \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilpd $0x5 , %%ymm13, %%ymm13 \n\t"
"vpermilpd $0x5 , %%ymm15, %%ymm15 \n\t"
"vaddsubpd %%ymm13, %%ymm12, %%ymm8 \n\t"
"vaddsubpd %%ymm15, %%ymm14, %%ymm9 \n\t"
#else
"vpermilpd $0x5 , %%ymm12, %%ymm12 \n\t"
"vpermilpd $0x5 , %%ymm14, %%ymm14 \n\t"
"vaddsubpd %%ymm12, %%ymm13, %%ymm8 \n\t"
"vaddsubpd %%ymm14, %%ymm15, %%ymm9 \n\t"
"vpermilpd $0x5 , %%ymm8 , %%ymm8 \n\t"
"vpermilpd $0x5 , %%ymm9 , %%ymm9 \n\t"
#endif
"vaddpd %%ymm8, %%ymm10, %%ymm12 \n\t"
"vaddpd %%ymm9, %%ymm11, %%ymm13 \n\t"
"vmovups %%ymm12, (%3,%0,8) \n\t" // 2 complex values to y
"vmovups %%ymm13, 32(%3,%0,8) \n\t"
"addq $8 , %0 \n\t"
"subq $4 , %1 \n\t"
"jnz .L01LOOP%= \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
: "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
#define HAVE_KERNEL_4x1 1
static void zgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void zgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastsd (%2), %%ymm0 \n\t" // real part x0
"vbroadcastsd 8(%2), %%ymm1 \n\t" // imag part x0
".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 192(%4,%0,8) \n\t"
"vmovups (%4,%0,8), %%ymm8 \n\t" // 2 complex values form a0
"vmovups 32(%4,%0,8), %%ymm9 \n\t" // 2 complex values form a0
"vmulpd %%ymm8 , %%ymm0, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vmulpd %%ymm8 , %%ymm1, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vmulpd %%ymm9 , %%ymm0, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vmulpd %%ymm9 , %%ymm1, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"prefetcht0 192(%3,%0,8) \n\t"
"vmovups (%3,%0,8), %%ymm10 \n\t"
"vmovups 32(%3,%0,8), %%ymm11 \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilpd $0x5 , %%ymm13, %%ymm13 \n\t"
"vpermilpd $0x5 , %%ymm15, %%ymm15 \n\t"
"vaddsubpd %%ymm13, %%ymm12, %%ymm8 \n\t"
"vaddsubpd %%ymm15, %%ymm14, %%ymm9 \n\t"
#else
"vpermilpd $0x5 , %%ymm12, %%ymm12 \n\t"
"vpermilpd $0x5 , %%ymm14, %%ymm14 \n\t"
"vaddsubpd %%ymm12, %%ymm13, %%ymm8 \n\t"
"vaddsubpd %%ymm14, %%ymm15, %%ymm9 \n\t"
"vpermilpd $0x5 , %%ymm8 , %%ymm8 \n\t"
"vpermilpd $0x5 , %%ymm9 , %%ymm9 \n\t"
#endif
"vaddpd %%ymm8, %%ymm10, %%ymm12 \n\t"
"vaddpd %%ymm9, %%ymm11, %%ymm13 \n\t"
"vmovups %%ymm12, (%3,%0,8) \n\t" // 2 complex values to y
"vmovups %%ymm13, 32(%3,%0,8) \n\t"
"addq $8 , %0 \n\t"
"subq $4 , %1 \n\t"
"jnz .L01LOOP%= \n\t"
"vzeroupper \n\t"
:
:
"r" (i), // 0
"r" (n), // 1
"r" (x), // 2
"r" (y), // 3
"r" (ap) // 4
: "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
#define HAVE_KERNEL_ADDY 1
static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i) __attribute__ ((noinline));
static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i)
{
BLASLONG i;
if ( inc_dest != 2 )
{
FLOAT temp_r;
FLOAT temp_i;
for ( i=0; i<n; i++ )
{
#if !defined(XCONJ)
temp_r = alpha_r * src[0] - alpha_i * src[1];
temp_i = alpha_r * src[1] + alpha_i * src[0];
#else
temp_r = alpha_r * src[0] + alpha_i * src[1];
temp_i = -alpha_r * src[1] + alpha_i * src[0];
#endif
*dest += temp_r;
*(dest+1) += temp_i;
src+=2;
dest += inc_dest;
}
return;
}
i=0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastsd (%4), %%ymm0 \n\t" // alpha_r
"vbroadcastsd (%5), %%ymm1 \n\t" // alpha_i
".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 192(%2,%0,8) \n\t"
"vmovups (%2,%0,8), %%ymm8 \n\t" // 2 complex values from src
"vmovups 32(%2,%0,8), %%ymm9 \n\t"
"vmulpd %%ymm8 , %%ymm0, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vmulpd %%ymm8 , %%ymm1, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vmulpd %%ymm9 , %%ymm0, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vmulpd %%ymm9 , %%ymm1, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
"prefetcht0 192(%3,%0,8) \n\t"
"vmovups (%3,%0,8), %%ymm10 \n\t" // 2 complex values from dest
"vmovups 32(%3,%0,8), %%ymm11 \n\t"
#if !defined(XCONJ)
"vpermilpd $0x5 , %%ymm13, %%ymm13 \n\t"
"vpermilpd $0x5 , %%ymm15, %%ymm15 \n\t"
"vaddsubpd %%ymm13, %%ymm12, %%ymm8 \n\t"
"vaddsubpd %%ymm15, %%ymm14, %%ymm9 \n\t"
#else
"vpermilpd $0x5 , %%ymm12, %%ymm12 \n\t"
"vpermilpd $0x5 , %%ymm14, %%ymm14 \n\t"
"vaddsubpd %%ymm12, %%ymm13, %%ymm8 \n\t"
"vaddsubpd %%ymm14, %%ymm15, %%ymm9 \n\t"
"vpermilpd $0x5 , %%ymm8 , %%ymm8 \n\t"
"vpermilpd $0x5 , %%ymm9 , %%ymm9 \n\t"
#endif
"vaddpd %%ymm8, %%ymm10, %%ymm12 \n\t"
"vaddpd %%ymm9, %%ymm11, %%ymm13 \n\t"
"vmovups %%ymm12, (%3,%0,8) \n\t" // 2 complex values to y
"vmovups %%ymm13, 32(%3,%0,8) \n\t"
"addq $8 , %0 \n\t"
"subq $4 , %1 \n\t"
"jnz .L01LOOP%= \n\t"
"vzeroupper \n\t"
:
:
"r" (i), // 0
"r" (n), // 1
"r" (src), // 2
"r" (dest), // 3
"r" (&alpha_r), // 4
"r" (&alpha_i) // 5
: "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
return;
}

149
kernel/x86_64/zgemv_n_microk_sandy-2.c
View File

@ -1,149 +0,0 @@
/***************************************************************************
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 zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastsd (%2), %%ymm0 \n\t" // real part x0
"vbroadcastsd 8(%2), %%ymm1 \n\t" // imag part x0
"vbroadcastsd 16(%2), %%ymm2 \n\t" // real part x1
"vbroadcastsd 24(%2), %%ymm3 \n\t" // imag part x1
"vbroadcastsd 32(%2), %%ymm4 \n\t" // real part x2
"vbroadcastsd 40(%2), %%ymm5 \n\t" // imag part x2
"vbroadcastsd 48(%2), %%ymm6 \n\t" // real part x3
"vbroadcastsd 56(%2), %%ymm7 \n\t" // imag part x3
".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 256(%4,%0,8) \n\t"
"vmovups (%4,%0,8), %%ymm8 \n\t" // 2 complex values form a0
"vmovups 32(%4,%0,8), %%ymm9 \n\t" // 2 complex values form a0
"vmulpd %%ymm8 , %%ymm0 , %%ymm12 \n\t"
"vmulpd %%ymm8 , %%ymm1 , %%ymm13 \n\t"
"prefetcht0 256(%5,%0,8) \n\t"
"vmulpd %%ymm9 , %%ymm0 , %%ymm14 \n\t"
"vmovups (%5,%0,8), %%ymm8 \n\t" // 2 complex values form a0
"vmulpd %%ymm9 , %%ymm1 , %%ymm15 \n\t"
"vmovups 32(%5,%0,8), %%ymm9 \n\t" // 2 complex values form a0
"vmulpd %%ymm8 , %%ymm2 , %%ymm10 \n\t"
"vaddpd %%ymm12, %%ymm10, %%ymm12 \n\t"
"vmulpd %%ymm8 , %%ymm3 , %%ymm11 \n\t"
"vaddpd %%ymm13, %%ymm11, %%ymm13 \n\t"
"prefetcht0 256(%6,%0,8) \n\t"
"vmulpd %%ymm9 , %%ymm2 , %%ymm10 \n\t"
"vaddpd %%ymm14, %%ymm10, %%ymm14 \n\t"
"vmovups (%6,%0,8), %%ymm8 \n\t" // 2 complex values form a0
"vmulpd %%ymm9 , %%ymm3 , %%ymm11 \n\t"
"vaddpd %%ymm15, %%ymm11, %%ymm15 \n\t"
"vmovups 32(%6,%0,8), %%ymm9 \n\t" // 2 complex values form a0
"vmulpd %%ymm8 , %%ymm4 , %%ymm10 \n\t"
"vaddpd %%ymm12, %%ymm10, %%ymm12 \n\t"
"vmulpd %%ymm8 , %%ymm5 , %%ymm11 \n\t"
"vaddpd %%ymm13, %%ymm11, %%ymm13 \n\t"
"prefetcht0 256(%7,%0,8) \n\t"
"vmulpd %%ymm9 , %%ymm4 , %%ymm10 \n\t"
"vaddpd %%ymm14, %%ymm10, %%ymm14 \n\t"
"vmovups (%7,%0,8), %%ymm8 \n\t" // 2 complex values form a0
"vmulpd %%ymm9 , %%ymm5 , %%ymm11 \n\t"
"vaddpd %%ymm15, %%ymm11, %%ymm15 \n\t"
"vmovups 32(%7,%0,8), %%ymm9 \n\t" // 2 complex values form a0
"vmulpd %%ymm8 , %%ymm6 , %%ymm10 \n\t"
"vaddpd %%ymm12, %%ymm10, %%ymm12 \n\t"
"vmulpd %%ymm8 , %%ymm7 , %%ymm11 \n\t"
"vaddpd %%ymm13, %%ymm11, %%ymm13 \n\t"
"vmulpd %%ymm9 , %%ymm6 , %%ymm10 \n\t"
"vaddpd %%ymm14, %%ymm10, %%ymm14 \n\t"
"vmulpd %%ymm9 , %%ymm7 , %%ymm11 \n\t"
"vaddpd %%ymm15, %%ymm11, %%ymm15 \n\t"
"prefetcht0 256(%3,%0,8) \n\t"
"vmovups (%3,%0,8), %%ymm10 \n\t"
"vmovups 32(%3,%0,8), %%ymm11 \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilpd $0x5 , %%ymm13, %%ymm13 \n\t"
"vpermilpd $0x5 , %%ymm15, %%ymm15 \n\t"
"vaddsubpd %%ymm13, %%ymm12, %%ymm8 \n\t"
"vaddsubpd %%ymm15, %%ymm14, %%ymm9 \n\t"
#else
"vpermilpd $0x5 , %%ymm12, %%ymm12 \n\t"
"vpermilpd $0x5 , %%ymm14, %%ymm14 \n\t"
"vaddsubpd %%ymm12, %%ymm13, %%ymm8 \n\t"
"vaddsubpd %%ymm14, %%ymm15, %%ymm9 \n\t"
"vpermilpd $0x5 , %%ymm8 , %%ymm8 \n\t"
"vpermilpd $0x5 , %%ymm9 , %%ymm9 \n\t"
#endif
"vaddpd %%ymm8, %%ymm10, %%ymm12 \n\t"
"vaddpd %%ymm9, %%ymm11, %%ymm13 \n\t"
"vmovups %%ymm12, (%3,%0,8) \n\t" // 2 complex values to y
"vmovups %%ymm13, 32(%3,%0,8) \n\t"
"addq $8 , %0 \n\t"
"subq $4 , %1 \n\t"
"jnz .L01LOOP%= \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",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}

417
kernel/x86_64/zgemv_n_microk_sandy-4.c Normal file
View File

@ -0,0 +1,417 @@
/***************************************************************************
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_4x4 1
static void zgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void zgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastsd (%2), %%ymm0 \n\t" // real part x0
"vbroadcastsd 8(%2), %%ymm1 \n\t" // imag part x0
"vbroadcastsd 16(%2), %%ymm2 \n\t" // real part x1
"vbroadcastsd 24(%2), %%ymm3 \n\t" // imag part x1
"vbroadcastsd 32(%2), %%ymm4 \n\t" // real part x2
"vbroadcastsd 40(%2), %%ymm5 \n\t" // imag part x2
"vbroadcastsd 48(%2), %%ymm6 \n\t" // real part x3
"vbroadcastsd 56(%2), %%ymm7 \n\t" // imag part x3
".align 16 \n\t"
".L01LOOP%=: \n\t"
//"prefetcht0 256(%4,%0,8) \n\t"
"vmovups (%4,%0,8), %%ymm8 \n\t" // 2 complex values form a0
"vmovups 32(%4,%0,8), %%ymm9 \n\t" // 2 complex values form a0
"vmulpd %%ymm8 , %%ymm0 , %%ymm12 \n\t"
"vmulpd %%ymm8 , %%ymm1 , %%ymm13 \n\t"
//"prefetcht0 256(%5,%0,8) \n\t"
"vmulpd %%ymm9 , %%ymm0 , %%ymm14 \n\t"
"vmovups (%5,%0,8), %%ymm8 \n\t" // 2 complex values form a0
"vmulpd %%ymm9 , %%ymm1 , %%ymm15 \n\t"
"vmovups 32(%5,%0,8), %%ymm9 \n\t" // 2 complex values form a0
"vmulpd %%ymm8 , %%ymm2 , %%ymm10 \n\t"
"vmulpd %%ymm8 , %%ymm3 , %%ymm11 \n\t"
"vaddpd %%ymm12, %%ymm10, %%ymm12 \n\t"
"vaddpd %%ymm13, %%ymm11, %%ymm13 \n\t"
//"prefetcht0 256(%6,%0,8) \n\t"
"vmulpd %%ymm9 , %%ymm2 , %%ymm10 \n\t"
"vmovups (%6,%0,8), %%ymm8 \n\t" // 2 complex values form a0
"vaddpd %%ymm14, %%ymm10, %%ymm14 \n\t"
"vmulpd %%ymm9 , %%ymm3 , %%ymm11 \n\t"
"vmovups 32(%6,%0,8), %%ymm9 \n\t" // 2 complex values form a0
"vaddpd %%ymm15, %%ymm11, %%ymm15 \n\t"
"vmulpd %%ymm8 , %%ymm4 , %%ymm10 \n\t"
"vmulpd %%ymm8 , %%ymm5 , %%ymm11 \n\t"
"vaddpd %%ymm12, %%ymm10, %%ymm12 \n\t"
"vaddpd %%ymm13, %%ymm11, %%ymm13 \n\t"
// "prefetcht0 256(%7,%0,8) \n\t"
"vmulpd %%ymm9 , %%ymm4 , %%ymm10 \n\t"
"vmovups (%7,%0,8), %%ymm8 \n\t" // 2 complex values form a0
"vaddpd %%ymm14, %%ymm10, %%ymm14 \n\t"
"vmulpd %%ymm9 , %%ymm5 , %%ymm11 \n\t"
"vmovups 32(%7,%0,8), %%ymm9 \n\t" // 2 complex values form a0
"vaddpd %%ymm15, %%ymm11, %%ymm15 \n\t"
"vmulpd %%ymm8 , %%ymm6 , %%ymm10 \n\t"
"vmulpd %%ymm8 , %%ymm7 , %%ymm11 \n\t"
"vaddpd %%ymm12, %%ymm10, %%ymm12 \n\t"
"vaddpd %%ymm13, %%ymm11, %%ymm13 \n\t"
"vmulpd %%ymm9 , %%ymm6 , %%ymm10 \n\t"
"vmulpd %%ymm9 , %%ymm7 , %%ymm11 \n\t"
"vaddpd %%ymm14, %%ymm10, %%ymm14 \n\t"
"vaddpd %%ymm15, %%ymm11, %%ymm15 \n\t"
// "prefetcht0 256(%3,%0,8) \n\t"
"vmovups (%3,%0,8), %%ymm10 \n\t"
"vmovups 32(%3,%0,8), %%ymm11 \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilpd $0x5 , %%ymm13, %%ymm13 \n\t"
"vpermilpd $0x5 , %%ymm15, %%ymm15 \n\t"
"vaddsubpd %%ymm13, %%ymm12, %%ymm8 \n\t"
"vaddsubpd %%ymm15, %%ymm14, %%ymm9 \n\t"
#else
"vpermilpd $0x5 , %%ymm12, %%ymm12 \n\t"
"vpermilpd $0x5 , %%ymm14, %%ymm14 \n\t"
"vaddsubpd %%ymm12, %%ymm13, %%ymm8 \n\t"
"vaddsubpd %%ymm14, %%ymm15, %%ymm9 \n\t"
"vpermilpd $0x5 , %%ymm8 , %%ymm8 \n\t"
"vpermilpd $0x5 , %%ymm9 , %%ymm9 \n\t"
#endif
"vaddpd %%ymm8, %%ymm10, %%ymm12 \n\t"
"vaddpd %%ymm9, %%ymm11, %%ymm13 \n\t"
"vmovups %%ymm12, (%3,%0,8) \n\t" // 2 complex values to y
"vmovups %%ymm13, 32(%3,%0,8) \n\t"
"addq $8 , %0 \n\t"
"subq $4 , %1 \n\t"
"jnz .L01LOOP%= \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",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
#define HAVE_KERNEL_4x2 1
static void zgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void zgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastsd (%2), %%ymm0 \n\t" // real part x0
"vbroadcastsd 8(%2), %%ymm1 \n\t" // imag part x0
"vbroadcastsd 16(%2), %%ymm2 \n\t" // real part x1
"vbroadcastsd 24(%2), %%ymm3 \n\t" // imag part x1
".align 16 \n\t"
".L01LOOP%=: \n\t"
// "prefetcht0 256(%4,%0,8) \n\t"
"vmovups (%4,%0,8), %%ymm8 \n\t" // 2 complex values form a0
"vmovups 32(%4,%0,8), %%ymm9 \n\t" // 2 complex values form a0
"vmulpd %%ymm8 , %%ymm0 , %%ymm12 \n\t"
"vmulpd %%ymm8 , %%ymm1 , %%ymm13 \n\t"
// "prefetcht0 256(%5,%0,8) \n\t"
"vmulpd %%ymm9 , %%ymm0 , %%ymm14 \n\t"
"vmovups (%5,%0,8), %%ymm8 \n\t" // 2 complex values form a0
"vmulpd %%ymm9 , %%ymm1 , %%ymm15 \n\t"
"vmovups 32(%5,%0,8), %%ymm9 \n\t" // 2 complex values form a0
"vmulpd %%ymm8 , %%ymm2 , %%ymm10 \n\t"
"vmulpd %%ymm8 , %%ymm3 , %%ymm11 \n\t"
"vaddpd %%ymm12, %%ymm10, %%ymm12 \n\t"
"vaddpd %%ymm13, %%ymm11, %%ymm13 \n\t"
"vmulpd %%ymm9 , %%ymm2 , %%ymm10 \n\t"
"vmulpd %%ymm9 , %%ymm3 , %%ymm11 \n\t"
"vaddpd %%ymm14, %%ymm10, %%ymm14 \n\t"
"vaddpd %%ymm15, %%ymm11, %%ymm15 \n\t"
// "prefetcht0 256(%3,%0,8) \n\t"
"vmovups (%3,%0,8), %%ymm10 \n\t"
"vmovups 32(%3,%0,8), %%ymm11 \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilpd $0x5 , %%ymm13, %%ymm13 \n\t"
"vpermilpd $0x5 , %%ymm15, %%ymm15 \n\t"
"vaddsubpd %%ymm13, %%ymm12, %%ymm8 \n\t"
"vaddsubpd %%ymm15, %%ymm14, %%ymm9 \n\t"
#else
"vpermilpd $0x5 , %%ymm12, %%ymm12 \n\t"
"vpermilpd $0x5 , %%ymm14, %%ymm14 \n\t"
"vaddsubpd %%ymm12, %%ymm13, %%ymm8 \n\t"
"vaddsubpd %%ymm14, %%ymm15, %%ymm9 \n\t"
"vpermilpd $0x5 , %%ymm8 , %%ymm8 \n\t"
"vpermilpd $0x5 , %%ymm9 , %%ymm9 \n\t"
#endif
"vaddpd %%ymm8, %%ymm10, %%ymm12 \n\t"
"vaddpd %%ymm9, %%ymm11, %%ymm13 \n\t"
"vmovups %%ymm12, (%3,%0,8) \n\t" // 2 complex values to y
"vmovups %%ymm13, 32(%3,%0,8) \n\t"
"addq $8 , %0 \n\t"
"subq $4 , %1 \n\t"
"jnz .L01LOOP%= \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
: "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
#define HAVE_KERNEL_4x1 1
static void zgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void zgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastsd (%2), %%ymm0 \n\t" // real part x0
"vbroadcastsd 8(%2), %%ymm1 \n\t" // imag part x0
".align 16 \n\t"
".L01LOOP%=: \n\t"
// "prefetcht0 256(%4,%0,8) \n\t"
"vmovups (%4,%0,8), %%ymm8 \n\t" // 2 complex values form a0
"vmovups 32(%4,%0,8), %%ymm9 \n\t" // 2 complex values form a0
"vmulpd %%ymm8 , %%ymm0 , %%ymm12 \n\t"
"vmulpd %%ymm8 , %%ymm1 , %%ymm13 \n\t"
"vmulpd %%ymm9 , %%ymm0 , %%ymm14 \n\t"
"vmulpd %%ymm9 , %%ymm1 , %%ymm15 \n\t"
// "prefetcht0 256(%3,%0,8) \n\t"
"vmovups (%3,%0,8), %%ymm10 \n\t"
"vmovups 32(%3,%0,8), %%ymm11 \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilpd $0x5 , %%ymm13, %%ymm13 \n\t"
"vpermilpd $0x5 , %%ymm15, %%ymm15 \n\t"
"vaddsubpd %%ymm13, %%ymm12, %%ymm8 \n\t"
"vaddsubpd %%ymm15, %%ymm14, %%ymm9 \n\t"
#else
"vpermilpd $0x5 , %%ymm12, %%ymm12 \n\t"
"vpermilpd $0x5 , %%ymm14, %%ymm14 \n\t"
"vaddsubpd %%ymm12, %%ymm13, %%ymm8 \n\t"
"vaddsubpd %%ymm14, %%ymm15, %%ymm9 \n\t"
"vpermilpd $0x5 , %%ymm8 , %%ymm8 \n\t"
"vpermilpd $0x5 , %%ymm9 , %%ymm9 \n\t"
#endif
"vaddpd %%ymm8, %%ymm10, %%ymm12 \n\t"
"vaddpd %%ymm9, %%ymm11, %%ymm13 \n\t"
"vmovups %%ymm12, (%3,%0,8) \n\t" // 2 complex values to y
"vmovups %%ymm13, 32(%3,%0,8) \n\t"
"addq $8 , %0 \n\t"
"subq $4 , %1 \n\t"
"jnz .L01LOOP%= \n\t"
"vzeroupper \n\t"
:
:
"r" (i), // 0
"r" (n), // 1
"r" (x), // 2
"r" (y), // 3
"r" (ap) // 4
: "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
#define HAVE_KERNEL_ADDY 1
static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i) __attribute__ ((noinline));
static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i)
{
BLASLONG i;
if ( inc_dest != 2 )
{
FLOAT temp_r;
FLOAT temp_i;
for ( i=0; i<n; i++ )
{
#if !defined(XCONJ)
temp_r = alpha_r * src[0] - alpha_i * src[1];
temp_i = alpha_r * src[1] + alpha_i * src[0];
#else
temp_r = alpha_r * src[0] + alpha_i * src[1];
temp_i = -alpha_r * src[1] + alpha_i * src[0];
#endif
*dest += temp_r;
*(dest+1) += temp_i;
src+=2;
dest += inc_dest;
}
return;
}
i=0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastsd (%4), %%ymm0 \n\t" // alpha_r
"vbroadcastsd (%5), %%ymm1 \n\t" // alpha_i
".align 16 \n\t"
".L01LOOP%=: \n\t"
// "prefetcht0 192(%2,%0,8) \n\t"
"vmovups (%2,%0,8), %%ymm8 \n\t" // 2 complex values from src
"vmovups 32(%2,%0,8), %%ymm9 \n\t"
"vmulpd %%ymm8 , %%ymm0, %%ymm12 \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
"vmulpd %%ymm8 , %%ymm1, %%ymm13 \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
"vmulpd %%ymm9 , %%ymm0, %%ymm14 \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
"vmulpd %%ymm9 , %%ymm1, %%ymm15 \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
// "prefetcht0 192(%3,%0,8) \n\t"
"vmovups (%3,%0,8), %%ymm10 \n\t" // 2 complex values from dest
"vmovups 32(%3,%0,8), %%ymm11 \n\t"
#if !defined(XCONJ)
"vpermilpd $0x5 , %%ymm13, %%ymm13 \n\t"
"vpermilpd $0x5 , %%ymm15, %%ymm15 \n\t"
"vaddsubpd %%ymm13, %%ymm12, %%ymm8 \n\t"
"vaddsubpd %%ymm15, %%ymm14, %%ymm9 \n\t"
#else
"vpermilpd $0x5 , %%ymm12, %%ymm12 \n\t"
"vpermilpd $0x5 , %%ymm14, %%ymm14 \n\t"
"vaddsubpd %%ymm12, %%ymm13, %%ymm8 \n\t"
"vaddsubpd %%ymm14, %%ymm15, %%ymm9 \n\t"
"vpermilpd $0x5 , %%ymm8 , %%ymm8 \n\t"
"vpermilpd $0x5 , %%ymm9 , %%ymm9 \n\t"
#endif
"vaddpd %%ymm8, %%ymm10, %%ymm12 \n\t"
"vaddpd %%ymm9, %%ymm11, %%ymm13 \n\t"
"vmovups %%ymm12, (%3,%0,8) \n\t" // 2 complex values to y
"vmovups %%ymm13, 32(%3,%0,8) \n\t"
"addq $8 , %0 \n\t"
"subq $4 , %1 \n\t"
"jnz .L01LOOP%= \n\t"
"vzeroupper \n\t"
:
:
"r" (i), // 0
"r" (n), // 1
"r" (src), // 2
"r" (dest), // 3
"r" (&alpha_r), // 4
"r" (&alpha_i) // 5
: "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
return;
}

272
kernel/x86_64/zgemv_t.c
View File

@ -1,272 +0,0 @@
/***************************************************************************
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(BULLDOZER) || defined(PILEDRIVER)
#include "zgemv_t_microk_bulldozer-2.c"
#elif defined(HASWELL)
#include "zgemv_t_microk_haswell-2.c"
#endif
#define NBMAX 1028
#ifndef HAVE_KERNEL_16x4
static void zgemv_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 temp_r0 = 0.0;
FLOAT temp_r1 = 0.0;
FLOAT temp_r2 = 0.0;
FLOAT temp_r3 = 0.0;
FLOAT temp_i0 = 0.0;
FLOAT temp_i1 = 0.0;
FLOAT temp_i2 = 0.0;
FLOAT temp_i3 = 0.0;
for ( i=0; i< 2*n; i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a0[i]*x[i] - a0[i+1]*x[i+1];
temp_i0 += a0[i]*x[i+1] + a0[i+1]*x[i];
temp_r1 += a1[i]*x[i] - a1[i+1]*x[i+1];
temp_i1 += a1[i]*x[i+1] + a1[i+1]*x[i];
temp_r2 += a2[i]*x[i] - a2[i+1]*x[i+1];
temp_i2 += a2[i]*x[i+1] + a2[i+1]*x[i];
temp_r3 += a3[i]*x[i] - a3[i+1]*x[i+1];
temp_i3 += a3[i]*x[i+1] + a3[i+1]*x[i];
#else
temp_r0 += a0[i]*x[i] + a0[i+1]*x[i+1];
temp_i0 += a0[i]*x[i+1] - a0[i+1]*x[i];
temp_r1 += a1[i]*x[i] + a1[i+1]*x[i+1];
temp_i1 += a1[i]*x[i+1] - a1[i+1]*x[i];
temp_r2 += a2[i]*x[i] + a2[i+1]*x[i+1];
temp_i2 += a2[i]*x[i+1] - a2[i+1]*x[i];
temp_r3 += a3[i]*x[i] + a3[i+1]*x[i+1];
temp_i3 += a3[i]*x[i+1] - a3[i+1]*x[i];
#endif
}
y[0] = temp_r0;
y[1] = temp_i0;
y[2] = temp_r1;
y[3] = temp_i1;
y[4] = temp_r2;
y[5] = temp_i2;
y[6] = temp_r3;
y[7] = temp_i3;
}
#endif
static void zgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
{
BLASLONG i;
FLOAT *a0;
a0 = ap;
FLOAT temp_r = 0.0;
FLOAT temp_i = 0.0;
for ( i=0; i< 2*n; i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r += a0[i]*x[i] - a0[i+1]*x[i+1];
temp_i += a0[i]*x[i+1] + a0[i+1]*x[i];
#else
temp_r += a0[i]*x[i] + a0[i+1]*x[i+1];
temp_i += a0[i]*x[i+1] - a0[i+1]*x[i];
#endif
}
*y = temp_r;
*(y+1) = temp_i;
}
static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
{
BLASLONG i;
for ( i=0; i<n; i++ )
{
*dest = *src;
*(dest+1) = *(src+1);
dest+=2;
src += inc_src;
}
}
int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i, 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[8];
BLASLONG n1;
BLASLONG m1;
BLASLONG m2;
BLASLONG n2;
FLOAT ybuffer[8],*xbuffer;
if ( m < 1 ) return(0);
if ( n < 1 ) return(0);
inc_x *= 2;
inc_y *= 2;
lda *= 2;
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;
zgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
a_ptr += 4 * lda;
#if !defined(XCONJ)
y_ptr[0] += alpha_r * ybuffer[0] - alpha_i * ybuffer[1];
y_ptr[1] += alpha_r * ybuffer[1] + alpha_i * ybuffer[0];
y_ptr += inc_y;
y_ptr[0] += alpha_r * ybuffer[2] - alpha_i * ybuffer[3];
y_ptr[1] += alpha_r * ybuffer[3] + alpha_i * ybuffer[2];
y_ptr += inc_y;
y_ptr[0] += alpha_r * ybuffer[4] - alpha_i * ybuffer[5];
y_ptr[1] += alpha_r * ybuffer[5] + alpha_i * ybuffer[4];
y_ptr += inc_y;
y_ptr[0] += alpha_r * ybuffer[6] - alpha_i * ybuffer[7];
y_ptr[1] += alpha_r * ybuffer[7] + alpha_i * ybuffer[6];
y_ptr += inc_y;
#else
y_ptr[0] += alpha_r * ybuffer[0] + alpha_i * ybuffer[1];
y_ptr[1] -= alpha_r * ybuffer[1] - alpha_i * ybuffer[0];
y_ptr += inc_y;
y_ptr[0] += alpha_r * ybuffer[2] + alpha_i * ybuffer[3];
y_ptr[1] -= alpha_r * ybuffer[3] - alpha_i * ybuffer[2];
y_ptr += inc_y;
y_ptr[0] += alpha_r * ybuffer[4] + alpha_i * ybuffer[5];
y_ptr[1] -= alpha_r * ybuffer[5] - alpha_i * ybuffer[4];
y_ptr += inc_y;
y_ptr[0] += alpha_r * ybuffer[6] + alpha_i * ybuffer[7];
y_ptr[1] -= alpha_r * ybuffer[7] - alpha_i * ybuffer[6];
y_ptr += inc_y;
#endif
}
for( i = 0; i < n2 ; i++)
{
zgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
a_ptr += 1 * lda;
#if !defined(XCONJ)
y_ptr[0] += alpha_r * ybuffer[0] - alpha_i * ybuffer[1];
y_ptr[1] += alpha_r * ybuffer[1] + alpha_i * ybuffer[0];
y_ptr += inc_y;
#else
y_ptr[0] += alpha_r * ybuffer[0] + alpha_i * ybuffer[1];
y_ptr[1] -= alpha_r * ybuffer[1] - alpha_i * ybuffer[0];
y_ptr += inc_y;
#endif
}
a += 2* NB;
x += NB * inc_x;
}
BLASLONG m3 = m % 16;
if ( m3 == 0 ) return(0);
x_ptr = x;
copy_x(m3,x_ptr,xbuffer,inc_x);
j=0;
a_ptr = a;
y_ptr = y;
while ( j < n)
{
FLOAT temp_r = 0.0;
FLOAT temp_i = 0.0;
for( i = 0; i < m3*2; i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r += a_ptr[i] * xbuffer[i] - a_ptr[i+1] * xbuffer[i+1];
temp_i += a_ptr[i] * xbuffer[i+1] + a_ptr[i+1] * xbuffer[i];
#else
temp_r += a_ptr[i] * xbuffer[i] + a_ptr[i+1] * xbuffer[i+1];
temp_i += a_ptr[i] * xbuffer[i+1] - a_ptr[i+1] * xbuffer[i];
#endif
}
a_ptr += lda;
#if !defined(XCONJ)
y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
#else
y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
#endif
y_ptr += inc_y;
j++;
}
return(0);
}

583
kernel/x86_64/zgemv_t_4.c Normal file
View File

@ -0,0 +1,583 @@
/***************************************************************************
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(BULLDOZER) || defined(PILEDRIVER)
#include "zgemv_t_microk_bulldozer-4.c"
#elif defined(HASWELL)
#include "zgemv_t_microk_haswell-4.c"
#endif
#define NBMAX 1024
#ifndef HAVE_KERNEL_4x4
static void zgemv_kernel_4x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{
BLASLONG i;
FLOAT *a0,*a1,*a2,*a3;
a0 = ap[0];
a1 = ap[1];
a2 = ap[2];
a3 = ap[3];
FLOAT alpha_r = alpha[0];
FLOAT alpha_i = alpha[1];
FLOAT temp_r0 = 0.0;
FLOAT temp_r1 = 0.0;
FLOAT temp_r2 = 0.0;
FLOAT temp_r3 = 0.0;
FLOAT temp_i0 = 0.0;
FLOAT temp_i1 = 0.0;
FLOAT temp_i2 = 0.0;
FLOAT temp_i3 = 0.0;
for ( i=0; i< 2*n; i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a0[i]*x[i] - a0[i+1]*x[i+1];
temp_i0 += a0[i]*x[i+1] + a0[i+1]*x[i];
temp_r1 += a1[i]*x[i] - a1[i+1]*x[i+1];
temp_i1 += a1[i]*x[i+1] + a1[i+1]*x[i];
temp_r2 += a2[i]*x[i] - a2[i+1]*x[i+1];
temp_i2 += a2[i]*x[i+1] + a2[i+1]*x[i];
temp_r3 += a3[i]*x[i] - a3[i+1]*x[i+1];
temp_i3 += a3[i]*x[i+1] + a3[i+1]*x[i];
#else
temp_r0 += a0[i]*x[i] + a0[i+1]*x[i+1];
temp_i0 += a0[i]*x[i+1] - a0[i+1]*x[i];
temp_r1 += a1[i]*x[i] + a1[i+1]*x[i+1];
temp_i1 += a1[i]*x[i+1] - a1[i+1]*x[i];
temp_r2 += a2[i]*x[i] + a2[i+1]*x[i+1];
temp_i2 += a2[i]*x[i+1] - a2[i+1]*x[i];
temp_r3 += a3[i]*x[i] + a3[i+1]*x[i+1];
temp_i3 += a3[i]*x[i+1] - a3[i+1]*x[i];
#endif
}
#if !defined(XCONJ)
y[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
y[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
y[2] += alpha_r * temp_r1 - alpha_i * temp_i1;
y[3] += alpha_r * temp_i1 + alpha_i * temp_r1;
y[4] += alpha_r * temp_r2 - alpha_i * temp_i2;
y[5] += alpha_r * temp_i2 + alpha_i * temp_r2;
y[6] += alpha_r * temp_r3 - alpha_i * temp_i3;
y[7] += alpha_r * temp_i3 + alpha_i * temp_r3;
#else
y[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
y[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
y[2] += alpha_r * temp_r1 + alpha_i * temp_i1;
y[3] -= alpha_r * temp_i1 - alpha_i * temp_r1;
y[4] += alpha_r * temp_r2 + alpha_i * temp_i2;
y[5] -= alpha_r * temp_i2 - alpha_i * temp_r2;
y[6] += alpha_r * temp_r3 + alpha_i * temp_i3;
y[7] -= alpha_r * temp_i3 - alpha_i * temp_r3;
#endif
}
#endif
#ifndef HAVE_KERNEL_4x2
static void zgemv_kernel_4x2(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{
BLASLONG i;
FLOAT *a0,*a1;
a0 = ap[0];
a1 = ap[1];
FLOAT alpha_r = alpha[0];
FLOAT alpha_i = alpha[1];
FLOAT temp_r0 = 0.0;
FLOAT temp_r1 = 0.0;
FLOAT temp_i0 = 0.0;
FLOAT temp_i1 = 0.0;
for ( i=0; i< 2*n; i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a0[i]*x[i] - a0[i+1]*x[i+1];
temp_i0 += a0[i]*x[i+1] + a0[i+1]*x[i];
temp_r1 += a1[i]*x[i] - a1[i+1]*x[i+1];
temp_i1 += a1[i]*x[i+1] + a1[i+1]*x[i];
#else
temp_r0 += a0[i]*x[i] + a0[i+1]*x[i+1];
temp_i0 += a0[i]*x[i+1] - a0[i+1]*x[i];
temp_r1 += a1[i]*x[i] + a1[i+1]*x[i+1];
temp_i1 += a1[i]*x[i+1] - a1[i+1]*x[i];
#endif
}
#if !defined(XCONJ)
y[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
y[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
y[2] += alpha_r * temp_r1 - alpha_i * temp_i1;
y[3] += alpha_r * temp_i1 + alpha_i * temp_r1;
#else
y[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
y[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
y[2] += alpha_r * temp_r1 + alpha_i * temp_i1;
y[3] -= alpha_r * temp_i1 - alpha_i * temp_r1;
#endif
}
#endif
#ifndef HAVE_KERNEL_4x1
static void zgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{
BLASLONG i;
FLOAT *a0;
a0 = ap;
FLOAT alpha_r = alpha[0];
FLOAT alpha_i = alpha[1];
FLOAT temp_r0 = 0.0;
FLOAT temp_i0 = 0.0;
for ( i=0; i< 2*n; i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a0[i]*x[i] - a0[i+1]*x[i+1];
temp_i0 += a0[i]*x[i+1] + a0[i+1]*x[i];
#else
temp_r0 += a0[i]*x[i] + a0[i+1]*x[i+1];
temp_i0 += a0[i]*x[i+1] - a0[i+1]*x[i];
#endif
}
#if !defined(XCONJ)
y[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
y[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
#else
y[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
y[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
#endif
}
#endif
static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
{
BLASLONG i;
for ( i=0; i<n; i++ )
{
*dest = *src;
*(dest+1) = *(src+1);
dest+=2;
src += inc_src;
}
}
int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i, 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[8];
BLASLONG n1;
BLASLONG m1;
BLASLONG m2;
BLASLONG m3;
BLASLONG n2;
BLASLONG lda4;
FLOAT ybuffer[8],*xbuffer;
FLOAT alpha[2];
if ( m < 1 ) return(0);
if ( n < 1 ) return(0);
inc_x <<= 1;
inc_y <<= 1;
lda <<= 1;
lda4 = lda << 2;
xbuffer = buffer;
n1 = n >> 2 ;
n2 = n & 3 ;
m3 = m & 3 ;
m1 = m - m3;
m2 = (m & (NBMAX-1)) - m3 ;
alpha[0] = alpha_r;
alpha[1] = alpha_i;
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;
ap[0] = a_ptr;
ap[1] = a_ptr + lda;
ap[2] = ap[1] + lda;
ap[3] = ap[2] + lda;
if ( inc_x != 2 )
copy_x(NB,x_ptr,xbuffer,inc_x);
else
xbuffer = x_ptr;
if ( inc_y == 2 )
{
for( i = 0; i < n1 ; i++)
{
zgemv_kernel_4x4(NB,ap,xbuffer,y_ptr,alpha);
ap[0] += lda4;
ap[1] += lda4;
ap[2] += lda4;
ap[3] += lda4;
a_ptr += lda4;
y_ptr += 8;
}
if ( n2 & 2 )
{
zgemv_kernel_4x2(NB,ap,xbuffer,y_ptr,alpha);
a_ptr += lda * 2;
y_ptr += 4;
}
if ( n2 & 1 )
{
zgemv_kernel_4x1(NB,a_ptr,xbuffer,y_ptr,alpha);
a_ptr += lda;
y_ptr += 2;
}
}
else
{
for( i = 0; i < n1 ; i++)
{
memset(ybuffer,0,64);
zgemv_kernel_4x4(NB,ap,xbuffer,ybuffer,alpha);
ap[0] += lda4;
ap[1] += lda4;
ap[2] += lda4;
ap[3] += lda4;
a_ptr += lda4;
y_ptr[0] += ybuffer[0];
y_ptr[1] += ybuffer[1];
y_ptr += inc_y;
y_ptr[0] += ybuffer[2];
y_ptr[1] += ybuffer[3];
y_ptr += inc_y;
y_ptr[0] += ybuffer[4];
y_ptr[1] += ybuffer[5];
y_ptr += inc_y;
y_ptr[0] += ybuffer[6];
y_ptr[1] += ybuffer[7];
y_ptr += inc_y;
}
for( i = 0; i < n2 ; i++)
{
memset(ybuffer,0,64);
zgemv_kernel_4x1(NB,a_ptr,xbuffer,ybuffer,alpha);
a_ptr += lda;
y_ptr[0] += ybuffer[0];
y_ptr[1] += ybuffer[1];
y_ptr += inc_y;
}
}
a += 2 * NB;
x += NB * inc_x;
}
if ( m3 == 0 ) return(0);
x_ptr = x;
j=0;
a_ptr = a;
y_ptr = y;
if ( m3 == 3 )
{
FLOAT temp_r ;
FLOAT temp_i ;
FLOAT x0 = x_ptr[0];
FLOAT x1 = x_ptr[1];
x_ptr += inc_x;
FLOAT x2 = x_ptr[0];
FLOAT x3 = x_ptr[1];
x_ptr += inc_x;
FLOAT x4 = x_ptr[0];
FLOAT x5 = x_ptr[1];
while ( j < n)
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
temp_r += a_ptr[2] * x2 - a_ptr[3] * x3;
temp_i += a_ptr[2] * x3 + a_ptr[3] * x2;
temp_r += a_ptr[4] * x4 - a_ptr[5] * x5;
temp_i += a_ptr[4] * x5 + a_ptr[5] * x4;
#else
temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
temp_r += a_ptr[2] * x2 + a_ptr[3] * x3;
temp_i += a_ptr[2] * x3 - a_ptr[3] * x2;
temp_r += a_ptr[4] * x4 + a_ptr[5] * x5;
temp_i += a_ptr[4] * x5 - a_ptr[5] * x4;
#endif
#if !defined(XCONJ)
y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
#else
y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
#endif
a_ptr += lda;
y_ptr += inc_y;
j++;
}
return(0);
}
if ( m3 == 2 )
{
FLOAT temp_r ;
FLOAT temp_i ;
FLOAT temp_r1 ;
FLOAT temp_i1 ;
FLOAT x0 = x_ptr[0];
FLOAT x1 = x_ptr[1];
x_ptr += inc_x;
FLOAT x2 = x_ptr[0];
FLOAT x3 = x_ptr[1];
FLOAT ar = alpha[0];
FLOAT ai = alpha[1];
while ( j < ( n & -2 ))
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
temp_r += a_ptr[2] * x2 - a_ptr[3] * x3;
temp_i += a_ptr[2] * x3 + a_ptr[3] * x2;
a_ptr += lda;
temp_r1 = a_ptr[0] * x0 - a_ptr[1] * x1;
temp_i1 = a_ptr[0] * x1 + a_ptr[1] * x0;
temp_r1 += a_ptr[2] * x2 - a_ptr[3] * x3;
temp_i1 += a_ptr[2] * x3 + a_ptr[3] * x2;
#else
temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
temp_r += a_ptr[2] * x2 + a_ptr[3] * x3;
temp_i += a_ptr[2] * x3 - a_ptr[3] * x2;
a_ptr += lda;
temp_r1 = a_ptr[0] * x0 + a_ptr[1] * x1;
temp_i1 = a_ptr[0] * x1 - a_ptr[1] * x0;
temp_r1 += a_ptr[2] * x2 + a_ptr[3] * x3;
temp_i1 += a_ptr[2] * x3 - a_ptr[3] * x2;
#endif
#if !defined(XCONJ)
y_ptr[0] += ar * temp_r - ai * temp_i;
y_ptr[1] += ar * temp_i + ai * temp_r;
y_ptr += inc_y;
y_ptr[0] += ar * temp_r1 - ai * temp_i1;
y_ptr[1] += ar * temp_i1 + ai * temp_r1;
#else
y_ptr[0] += ar * temp_r + ai * temp_i;
y_ptr[1] -= ar * temp_i - ai * temp_r;
y_ptr += inc_y;
y_ptr[0] += ar * temp_r1 + ai * temp_i1;
y_ptr[1] -= ar * temp_i1 - ai * temp_r1;
#endif
a_ptr += lda;
y_ptr += inc_y;
j+=2;
}
while ( j < n)
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
temp_r += a_ptr[2] * x2 - a_ptr[3] * x3;
temp_i += a_ptr[2] * x3 + a_ptr[3] * x2;
#else
temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
temp_r += a_ptr[2] * x2 + a_ptr[3] * x3;
temp_i += a_ptr[2] * x3 - a_ptr[3] * x2;
#endif
#if !defined(XCONJ)
y_ptr[0] += ar * temp_r - ai * temp_i;
y_ptr[1] += ar * temp_i + ai * temp_r;
#else
y_ptr[0] += ar * temp_r + ai * temp_i;
y_ptr[1] -= ar * temp_i - ai * temp_r;
#endif
a_ptr += lda;
y_ptr += inc_y;
j++;
}
return(0);
}
if ( m3 == 1 )
{
FLOAT temp_r ;
FLOAT temp_i ;
FLOAT temp_r1 ;
FLOAT temp_i1 ;
FLOAT x0 = x_ptr[0];
FLOAT x1 = x_ptr[1];
FLOAT ar = alpha[0];
FLOAT ai = alpha[1];
while ( j < ( n & -2 ))
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
a_ptr += lda;
temp_r1 = a_ptr[0] * x0 - a_ptr[1] * x1;
temp_i1 = a_ptr[0] * x1 + a_ptr[1] * x0;
#else
temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
a_ptr += lda;
temp_r1 = a_ptr[0] * x0 + a_ptr[1] * x1;
temp_i1 = a_ptr[0] * x1 - a_ptr[1] * x0;
#endif
#if !defined(XCONJ)
y_ptr[0] += ar * temp_r - ai * temp_i;
y_ptr[1] += ar * temp_i + ai * temp_r;
y_ptr += inc_y;
y_ptr[0] += ar * temp_r1 - ai * temp_i1;
y_ptr[1] += ar * temp_i1 + ai * temp_r1;
#else
y_ptr[0] += ar * temp_r + ai * temp_i;
y_ptr[1] -= ar * temp_i - ai * temp_r;
y_ptr += inc_y;
y_ptr[0] += ar * temp_r1 + ai * temp_i1;
y_ptr[1] -= ar * temp_i1 - ai * temp_r1;
#endif
a_ptr += lda;
y_ptr += inc_y;
j+=2;
}
while ( j < n)
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
#else
temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
#endif
#if !defined(XCONJ)
y_ptr[0] += ar * temp_r - ai * temp_i;
y_ptr[1] += ar * temp_i + ai * temp_r;
#else
y_ptr[0] += ar * temp_r + ai * temp_i;
y_ptr[1] -= ar * temp_i - ai * temp_r;
#endif
a_ptr += lda;
y_ptr += inc_y;
j++;
}
return(0);
}
return(0);
}

180
kernel/x86_64/zgemv_t_microk_bulldozer-2.c
View File

@ -1,180 +0,0 @@
/***************************************************************************
Copyright (c) 2014, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary froms, 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 from 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 zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vxorpd %%xmm8 , %%xmm8 , %%xmm8 \n\t" // temp
"vxorpd %%xmm9 , %%xmm9 , %%xmm9 \n\t" // temp
"vxorpd %%xmm10, %%xmm10, %%xmm10 \n\t" // temp
"vxorpd %%xmm11, %%xmm11, %%xmm11 \n\t" // temp
"vxorpd %%xmm12, %%xmm12, %%xmm12 \n\t" // temp
"vxorpd %%xmm13, %%xmm13, %%xmm13 \n\t"
"vxorpd %%xmm14, %%xmm14, %%xmm14 \n\t"
"vxorpd %%xmm15, %%xmm15, %%xmm15 \n\t"
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vmovddup (%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 8(%2,%0,8), %%xmm1 \n\t" // imag value from x0
"prefetcht0 192(%4,%0,8) \n\t"
"vmovups (%4,%0,8), %%xmm4 \n\t" // 1 complex values from a0
"prefetcht0 192(%5,%0,8) \n\t"
"vmovups (%5,%0,8), %%xmm5 \n\t" // 1 complex values from a1
"prefetcht0 192(%6,%0,8) \n\t"
"vmovups (%6,%0,8), %%xmm6 \n\t" // 1 complex values from a2
"prefetcht0 192(%7,%0,8) \n\t"
"vmovups (%7,%0,8), %%xmm7 \n\t" // 1 complex values from a3
"vfmaddpd %%xmm8 , %%xmm4 , %%xmm0, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm4 , %%xmm1, %%xmm9 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm10, %%xmm5 , %%xmm0, %%xmm10 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm11, %%xmm5 , %%xmm1, %%xmm11 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm12, %%xmm6 , %%xmm0, %%xmm12 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm13, %%xmm6 , %%xmm1, %%xmm13 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm14, %%xmm7 , %%xmm0, %%xmm14 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm15, %%xmm7 , %%xmm1, %%xmm15 \n\t" // ar0*xl0,al0*xl0
"vmovddup 16(%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 24(%2,%0,8), %%xmm1 \n\t" // imag value from x0
"vmovups 16(%4,%0,8), %%xmm4 \n\t" // 1 complex values from a0
"vmovups 16(%5,%0,8), %%xmm5 \n\t" // 1 complex values from a1
"vmovups 16(%6,%0,8), %%xmm6 \n\t" // 1 complex values from a2
"vmovups 16(%7,%0,8), %%xmm7 \n\t" // 1 complex values from a3
"vfmaddpd %%xmm8 , %%xmm4 , %%xmm0, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm4 , %%xmm1, %%xmm9 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm10, %%xmm5 , %%xmm0, %%xmm10 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm11, %%xmm5 , %%xmm1, %%xmm11 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm12, %%xmm6 , %%xmm0, %%xmm12 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm13, %%xmm6 , %%xmm1, %%xmm13 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm14, %%xmm7 , %%xmm0, %%xmm14 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm15, %%xmm7 , %%xmm1, %%xmm15 \n\t" // ar0*xl0,al0*xl0
"vmovddup 32(%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 40(%2,%0,8), %%xmm1 \n\t" // imag value from x0
"vmovups 32(%4,%0,8), %%xmm4 \n\t" // 1 complex values from a0
"vmovups 32(%5,%0,8), %%xmm5 \n\t" // 1 complex values from a1
"vmovups 32(%6,%0,8), %%xmm6 \n\t" // 1 complex values from a2
"vmovups 32(%7,%0,8), %%xmm7 \n\t" // 1 complex values from a3
"vfmaddpd %%xmm8 , %%xmm4 , %%xmm0, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm4 , %%xmm1, %%xmm9 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm10, %%xmm5 , %%xmm0, %%xmm10 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm11, %%xmm5 , %%xmm1, %%xmm11 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm12, %%xmm6 , %%xmm0, %%xmm12 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm13, %%xmm6 , %%xmm1, %%xmm13 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm14, %%xmm7 , %%xmm0, %%xmm14 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm15, %%xmm7 , %%xmm1, %%xmm15 \n\t" // ar0*xl0,al0*xl0
"vmovddup 48(%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 56(%2,%0,8), %%xmm1 \n\t" // imag value from x0
"vmovups 48(%4,%0,8), %%xmm4 \n\t" // 1 complex values from a0
"vmovups 48(%5,%0,8), %%xmm5 \n\t" // 1 complex values from a1
"vmovups 48(%6,%0,8), %%xmm6 \n\t" // 1 complex values from a2
"vmovups 48(%7,%0,8), %%xmm7 \n\t" // 1 complex values from a3
"vfmaddpd %%xmm8 , %%xmm4 , %%xmm0, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm4 , %%xmm1, %%xmm9 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm10, %%xmm5 , %%xmm0, %%xmm10 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm11, %%xmm5 , %%xmm1, %%xmm11 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm12, %%xmm6 , %%xmm0, %%xmm12 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm13, %%xmm6 , %%xmm1, %%xmm13 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm14, %%xmm7 , %%xmm0, %%xmm14 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm15, %%xmm7 , %%xmm1, %%xmm15 \n\t" // ar0*xl0,al0*xl0
"addq $8 , %0 \n\t"
"subq $4 , %1 \n\t"
"jnz .L01LOOP%= \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilpd $0x1 , %%xmm9 , %%xmm9 \n\t"
"vpermilpd $0x1 , %%xmm11, %%xmm11 \n\t"
"vpermilpd $0x1 , %%xmm13, %%xmm13 \n\t"
"vpermilpd $0x1 , %%xmm15, %%xmm15 \n\t"
"vaddsubpd %%xmm9 , %%xmm8, %%xmm8 \n\t"
"vaddsubpd %%xmm11, %%xmm10, %%xmm10 \n\t"
"vaddsubpd %%xmm13, %%xmm12, %%xmm12 \n\t"
"vaddsubpd %%xmm15, %%xmm14, %%xmm14 \n\t"
#else
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
"vpermilpd $0x1 , %%xmm10, %%xmm10 \n\t"
"vpermilpd $0x1 , %%xmm12, %%xmm12 \n\t"
"vpermilpd $0x1 , %%xmm14, %%xmm14 \n\t"
"vaddsubpd %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vaddsubpd %%xmm10, %%xmm11, %%xmm10 \n\t"
"vaddsubpd %%xmm12, %%xmm13, %%xmm12 \n\t"
"vaddsubpd %%xmm14, %%xmm15, %%xmm14 \n\t"
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
"vpermilpd $0x1 , %%xmm10, %%xmm10 \n\t"
"vpermilpd $0x1 , %%xmm12, %%xmm12 \n\t"
"vpermilpd $0x1 , %%xmm14, %%xmm14 \n\t"
#endif
"vmovups %%xmm8 , (%3) \n\t"
"vmovups %%xmm10, 16(%3) \n\t"
"vmovups %%xmm12, 32(%3) \n\t"
"vmovups %%xmm14, 48(%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",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}

457
kernel/x86_64/zgemv_t_microk_bulldozer-4.c Normal file
View File

@ -0,0 +1,457 @@
/***************************************************************************
Copyright (c) 2014, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary froms, 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 from 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_4x4 1
static void zgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha) __attribute__ ((noinline));
static void zgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vxorpd %%xmm8 , %%xmm8 , %%xmm8 \n\t" // temp
"vxorpd %%xmm9 , %%xmm9 , %%xmm9 \n\t" // temp
"vxorpd %%xmm10, %%xmm10, %%xmm10 \n\t" // temp
"vxorpd %%xmm11, %%xmm11, %%xmm11 \n\t" // temp
"vxorpd %%xmm12, %%xmm12, %%xmm12 \n\t" // temp
"vxorpd %%xmm13, %%xmm13, %%xmm13 \n\t"
"vxorpd %%xmm14, %%xmm14, %%xmm14 \n\t"
"vxorpd %%xmm15, %%xmm15, %%xmm15 \n\t"
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vmovddup (%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 8(%2,%0,8), %%xmm1 \n\t" // imag value from x0
"prefetcht0 192(%4,%0,8) \n\t"
"vmovups (%4,%0,8), %%xmm4 \n\t" // 1 complex values from a0
"prefetcht0 192(%5,%0,8) \n\t"
"vmovups (%5,%0,8), %%xmm5 \n\t" // 1 complex values from a1
"prefetcht0 192(%6,%0,8) \n\t"
"vmovups (%6,%0,8), %%xmm6 \n\t" // 1 complex values from a2
"prefetcht0 192(%7,%0,8) \n\t"
"vmovups (%7,%0,8), %%xmm7 \n\t" // 1 complex values from a3
"vfmaddpd %%xmm8 , %%xmm4 , %%xmm0, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm4 , %%xmm1, %%xmm9 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm10, %%xmm5 , %%xmm0, %%xmm10 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm11, %%xmm5 , %%xmm1, %%xmm11 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm12, %%xmm6 , %%xmm0, %%xmm12 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm13, %%xmm6 , %%xmm1, %%xmm13 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm14, %%xmm7 , %%xmm0, %%xmm14 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm15, %%xmm7 , %%xmm1, %%xmm15 \n\t" // ar0*xl0,al0*xl0
"vmovddup 16(%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 24(%2,%0,8), %%xmm1 \n\t" // imag value from x0
"vmovups 16(%4,%0,8), %%xmm4 \n\t" // 1 complex values from a0
"vmovups 16(%5,%0,8), %%xmm5 \n\t" // 1 complex values from a1
"vmovups 16(%6,%0,8), %%xmm6 \n\t" // 1 complex values from a2
"vmovups 16(%7,%0,8), %%xmm7 \n\t" // 1 complex values from a3
"vfmaddpd %%xmm8 , %%xmm4 , %%xmm0, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm4 , %%xmm1, %%xmm9 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm10, %%xmm5 , %%xmm0, %%xmm10 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm11, %%xmm5 , %%xmm1, %%xmm11 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm12, %%xmm6 , %%xmm0, %%xmm12 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm13, %%xmm6 , %%xmm1, %%xmm13 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm14, %%xmm7 , %%xmm0, %%xmm14 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm15, %%xmm7 , %%xmm1, %%xmm15 \n\t" // ar0*xl0,al0*xl0
"vmovddup 32(%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 40(%2,%0,8), %%xmm1 \n\t" // imag value from x0
"vmovups 32(%4,%0,8), %%xmm4 \n\t" // 1 complex values from a0
"vmovups 32(%5,%0,8), %%xmm5 \n\t" // 1 complex values from a1
"vmovups 32(%6,%0,8), %%xmm6 \n\t" // 1 complex values from a2
"vmovups 32(%7,%0,8), %%xmm7 \n\t" // 1 complex values from a3
"vfmaddpd %%xmm8 , %%xmm4 , %%xmm0, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm4 , %%xmm1, %%xmm9 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm10, %%xmm5 , %%xmm0, %%xmm10 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm11, %%xmm5 , %%xmm1, %%xmm11 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm12, %%xmm6 , %%xmm0, %%xmm12 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm13, %%xmm6 , %%xmm1, %%xmm13 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm14, %%xmm7 , %%xmm0, %%xmm14 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm15, %%xmm7 , %%xmm1, %%xmm15 \n\t" // ar0*xl0,al0*xl0
"vmovddup 48(%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 56(%2,%0,8), %%xmm1 \n\t" // imag value from x0
"vmovups 48(%4,%0,8), %%xmm4 \n\t" // 1 complex values from a0
"vmovups 48(%5,%0,8), %%xmm5 \n\t" // 1 complex values from a1
"vmovups 48(%6,%0,8), %%xmm6 \n\t" // 1 complex values from a2
"vmovups 48(%7,%0,8), %%xmm7 \n\t" // 1 complex values from a3
"vfmaddpd %%xmm8 , %%xmm4 , %%xmm0, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm4 , %%xmm1, %%xmm9 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm10, %%xmm5 , %%xmm0, %%xmm10 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm11, %%xmm5 , %%xmm1, %%xmm11 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm12, %%xmm6 , %%xmm0, %%xmm12 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm13, %%xmm6 , %%xmm1, %%xmm13 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm14, %%xmm7 , %%xmm0, %%xmm14 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm15, %%xmm7 , %%xmm1, %%xmm15 \n\t" // ar0*xl0,al0*xl0
"addq $8 , %0 \n\t"
"subq $4 , %1 \n\t"
"jnz .L01LOOP%= \n\t"
"vmovddup (%8) , %%xmm0 \n\t" // value from alpha
"vmovddup 8(%8) , %%xmm1 \n\t" // value from alpha
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilpd $0x1 , %%xmm9 , %%xmm9 \n\t"
"vpermilpd $0x1 , %%xmm11, %%xmm11 \n\t"
"vpermilpd $0x1 , %%xmm13, %%xmm13 \n\t"
"vpermilpd $0x1 , %%xmm15, %%xmm15 \n\t"
"vaddsubpd %%xmm9 , %%xmm8, %%xmm8 \n\t"
"vaddsubpd %%xmm11, %%xmm10, %%xmm10 \n\t"
"vaddsubpd %%xmm13, %%xmm12, %%xmm12 \n\t"
"vaddsubpd %%xmm15, %%xmm14, %%xmm14 \n\t"
#else
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
"vpermilpd $0x1 , %%xmm10, %%xmm10 \n\t"
"vpermilpd $0x1 , %%xmm12, %%xmm12 \n\t"
"vpermilpd $0x1 , %%xmm14, %%xmm14 \n\t"
"vaddsubpd %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vaddsubpd %%xmm10, %%xmm11, %%xmm10 \n\t"
"vaddsubpd %%xmm12, %%xmm13, %%xmm12 \n\t"
"vaddsubpd %%xmm14, %%xmm15, %%xmm14 \n\t"
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
"vpermilpd $0x1 , %%xmm10, %%xmm10 \n\t"
"vpermilpd $0x1 , %%xmm12, %%xmm12 \n\t"
"vpermilpd $0x1 , %%xmm14, %%xmm14 \n\t"
#endif
"vmulpd %%xmm8 , %%xmm1 , %%xmm9 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulpd %%xmm8 , %%xmm0 , %%xmm8 \n\t" // t_r * alpha_r , t_i * alpha_r
"vmulpd %%xmm10, %%xmm1 , %%xmm11 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulpd %%xmm10, %%xmm0 , %%xmm10 \n\t" // t_r * alpha_r , t_i * alpha_r
"vmulpd %%xmm12, %%xmm1 , %%xmm13 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulpd %%xmm12, %%xmm0 , %%xmm12 \n\t" // t_r * alpha_r , t_i * alpha_r
"vmulpd %%xmm14, %%xmm1 , %%xmm15 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulpd %%xmm14, %%xmm0 , %%xmm14 \n\t" // t_r * alpha_r , t_i * alpha_r
#if !defined(XCONJ)
"vpermilpd $0x1 , %%xmm9 , %%xmm9 \n\t"
"vpermilpd $0x1 , %%xmm11, %%xmm11 \n\t"
"vpermilpd $0x1 , %%xmm13, %%xmm13 \n\t"
"vpermilpd $0x1 , %%xmm15, %%xmm15 \n\t"
"vaddsubpd %%xmm9 , %%xmm8, %%xmm8 \n\t"
"vaddsubpd %%xmm11, %%xmm10, %%xmm10 \n\t"
"vaddsubpd %%xmm13, %%xmm12, %%xmm12 \n\t"
"vaddsubpd %%xmm15, %%xmm14, %%xmm14 \n\t"
#else
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
"vpermilpd $0x1 , %%xmm10, %%xmm10 \n\t"
"vpermilpd $0x1 , %%xmm12, %%xmm12 \n\t"
"vpermilpd $0x1 , %%xmm14, %%xmm14 \n\t"
"vaddsubpd %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vaddsubpd %%xmm10, %%xmm11, %%xmm10 \n\t"
"vaddsubpd %%xmm12, %%xmm13, %%xmm12 \n\t"
"vaddsubpd %%xmm14, %%xmm15, %%xmm14 \n\t"
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
"vpermilpd $0x1 , %%xmm10, %%xmm10 \n\t"
"vpermilpd $0x1 , %%xmm12, %%xmm12 \n\t"
"vpermilpd $0x1 , %%xmm14, %%xmm14 \n\t"
#endif
"vaddpd (%3) , %%xmm8 , %%xmm8 \n\t"
"vaddpd 16(%3) , %%xmm10, %%xmm10 \n\t"
"vaddpd 32(%3) , %%xmm12, %%xmm12 \n\t"
"vaddpd 48(%3) , %%xmm14, %%xmm14 \n\t"
"vmovups %%xmm8 , (%3) \n\t"
"vmovups %%xmm10, 16(%3) \n\t"
"vmovups %%xmm12, 32(%3) \n\t"
"vmovups %%xmm14, 48(%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
"r" (alpha) // 8
: "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
#define HAVE_KERNEL_4x2 1
static void zgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha) __attribute__ ((noinline));
static void zgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vxorpd %%xmm8 , %%xmm8 , %%xmm8 \n\t" // temp
"vxorpd %%xmm9 , %%xmm9 , %%xmm9 \n\t" // temp
"vxorpd %%xmm10, %%xmm10, %%xmm10 \n\t" // temp
"vxorpd %%xmm11, %%xmm11, %%xmm11 \n\t" // temp
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vmovddup (%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 8(%2,%0,8), %%xmm1 \n\t" // imag value from x0
"prefetcht0 192(%4,%0,8) \n\t"
"vmovups (%4,%0,8), %%xmm4 \n\t" // 1 complex values from a0
"prefetcht0 192(%5,%0,8) \n\t"
"vmovups (%5,%0,8), %%xmm5 \n\t" // 1 complex values from a1
"vfmaddpd %%xmm8 , %%xmm4 , %%xmm0, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm4 , %%xmm1, %%xmm9 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm10, %%xmm5 , %%xmm0, %%xmm10 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm11, %%xmm5 , %%xmm1, %%xmm11 \n\t" // ar0*xl0,al0*xl0
"vmovddup 16(%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 24(%2,%0,8), %%xmm1 \n\t" // imag value from x0
"vmovups 16(%4,%0,8), %%xmm4 \n\t" // 1 complex values from a0
"vmovups 16(%5,%0,8), %%xmm5 \n\t" // 1 complex values from a1
"vfmaddpd %%xmm8 , %%xmm4 , %%xmm0, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm4 , %%xmm1, %%xmm9 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm10, %%xmm5 , %%xmm0, %%xmm10 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm11, %%xmm5 , %%xmm1, %%xmm11 \n\t" // ar0*xl0,al0*xl0
"vmovddup 32(%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 40(%2,%0,8), %%xmm1 \n\t" // imag value from x0
"vmovups 32(%4,%0,8), %%xmm4 \n\t" // 1 complex values from a0
"vmovups 32(%5,%0,8), %%xmm5 \n\t" // 1 complex values from a1
"vfmaddpd %%xmm8 , %%xmm4 , %%xmm0, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm4 , %%xmm1, %%xmm9 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm10, %%xmm5 , %%xmm0, %%xmm10 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm11, %%xmm5 , %%xmm1, %%xmm11 \n\t" // ar0*xl0,al0*xl0
"vmovddup 48(%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 56(%2,%0,8), %%xmm1 \n\t" // imag value from x0
"vmovups 48(%4,%0,8), %%xmm4 \n\t" // 1 complex values from a0
"vmovups 48(%5,%0,8), %%xmm5 \n\t" // 1 complex values from a1
"vfmaddpd %%xmm8 , %%xmm4 , %%xmm0, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm4 , %%xmm1, %%xmm9 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm10, %%xmm5 , %%xmm0, %%xmm10 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm11, %%xmm5 , %%xmm1, %%xmm11 \n\t" // ar0*xl0,al0*xl0
"addq $8 , %0 \n\t"
"subq $4 , %1 \n\t"
"jnz .L01LOOP%= \n\t"
"vmovddup (%6) , %%xmm0 \n\t" // value from alpha
"vmovddup 8(%6) , %%xmm1 \n\t" // value from alpha
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilpd $0x1 , %%xmm9 , %%xmm9 \n\t"
"vpermilpd $0x1 , %%xmm11, %%xmm11 \n\t"
"vaddsubpd %%xmm9 , %%xmm8, %%xmm8 \n\t"
"vaddsubpd %%xmm11, %%xmm10, %%xmm10 \n\t"
#else
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
"vpermilpd $0x1 , %%xmm10, %%xmm10 \n\t"
"vaddsubpd %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vaddsubpd %%xmm10, %%xmm11, %%xmm10 \n\t"
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
"vpermilpd $0x1 , %%xmm10, %%xmm10 \n\t"
#endif
"vmulpd %%xmm8 , %%xmm1 , %%xmm9 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulpd %%xmm8 , %%xmm0 , %%xmm8 \n\t" // t_r * alpha_r , t_i * alpha_r
"vmulpd %%xmm10, %%xmm1 , %%xmm11 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulpd %%xmm10, %%xmm0 , %%xmm10 \n\t" // t_r * alpha_r , t_i * alpha_r
#if !defined(XCONJ)
"vpermilpd $0x1 , %%xmm9 , %%xmm9 \n\t"
"vpermilpd $0x1 , %%xmm11, %%xmm11 \n\t"
"vaddsubpd %%xmm9 , %%xmm8, %%xmm8 \n\t"
"vaddsubpd %%xmm11, %%xmm10, %%xmm10 \n\t"
#else
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
"vpermilpd $0x1 , %%xmm10, %%xmm10 \n\t"
"vaddsubpd %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vaddsubpd %%xmm10, %%xmm11, %%xmm10 \n\t"
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
"vpermilpd $0x1 , %%xmm10, %%xmm10 \n\t"
#endif
"vaddpd (%3) , %%xmm8 , %%xmm8 \n\t"
"vaddpd 16(%3) , %%xmm10, %%xmm10 \n\t"
"vmovups %%xmm8 , (%3) \n\t"
"vmovups %%xmm10, 16(%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" (alpha) // 6
: "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
#define HAVE_KERNEL_4x1 1
static void zgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT *alpha) __attribute__ ((noinline));
static void zgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vxorpd %%xmm8 , %%xmm8 , %%xmm8 \n\t" // temp
"vxorpd %%xmm9 , %%xmm9 , %%xmm9 \n\t" // temp
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vmovddup (%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 8(%2,%0,8), %%xmm1 \n\t" // imag value from x0
"prefetcht0 192(%4,%0,8) \n\t"
"vmovups (%4,%0,8), %%xmm4 \n\t" // 1 complex values from a0
"vmovups 16(%4,%0,8), %%xmm5 \n\t" // 1 complex values from a0
"vmovddup 16(%2,%0,8), %%xmm2 \n\t" // real value from x0
"vmovddup 24(%2,%0,8), %%xmm3 \n\t" // imag value from x0
"vfmaddpd %%xmm8 , %%xmm4 , %%xmm0, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm4 , %%xmm1, %%xmm9 \n\t" // ar0*xl0,al0*xl0
"vmovddup 32(%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 40(%2,%0,8), %%xmm1 \n\t" // imag value from x0
"vfmaddpd %%xmm8 , %%xmm5 , %%xmm2, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm5 , %%xmm3, %%xmm9 \n\t" // ar0*xl0,al0*xl0
"vmovups 32(%4,%0,8), %%xmm4 \n\t" // 1 complex values from a0
"vmovups 48(%4,%0,8), %%xmm5 \n\t" // 1 complex values from a0
"vmovddup 48(%2,%0,8), %%xmm2 \n\t" // real value from x0
"vmovddup 56(%2,%0,8), %%xmm3 \n\t" // imag value from x0
"addq $8 , %0 \n\t"
"vfmaddpd %%xmm8 , %%xmm4 , %%xmm0, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm4 , %%xmm1, %%xmm9 \n\t" // ar0*xl0,al0*xl0
"subq $4 , %1 \n\t"
"vfmaddpd %%xmm8 , %%xmm5 , %%xmm2, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm5 , %%xmm3, %%xmm9 \n\t" // ar0*xl0,al0*xl0
"jnz .L01LOOP%= \n\t"
"vmovddup (%5) , %%xmm0 \n\t" // value from alpha
"vmovddup 8(%5) , %%xmm1 \n\t" // value from alpha
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilpd $0x1 , %%xmm9 , %%xmm9 \n\t"
"vaddsubpd %%xmm9 , %%xmm8, %%xmm8 \n\t"
#else
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
"vaddsubpd %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
#endif
"vmulpd %%xmm8 , %%xmm1 , %%xmm9 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulpd %%xmm8 , %%xmm0 , %%xmm8 \n\t" // t_r * alpha_r , t_i * alpha_r
#if !defined(XCONJ)
"vpermilpd $0x1 , %%xmm9 , %%xmm9 \n\t"
"vaddsubpd %%xmm9 , %%xmm8, %%xmm8 \n\t"
#else
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
"vaddsubpd %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
#endif
"vaddpd (%3) , %%xmm8 , %%xmm8 \n\t"
"vmovups %%xmm8 , (%3) \n\t"
"vzeroupper \n\t"
:
:
"r" (i), // 0
"r" (n), // 1
"r" (x), // 2
"r" (y), // 3
"r" (ap), // 4
"r" (alpha) // 5
: "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}

162
kernel/x86_64/zgemv_t_microk_haswell-2.c
View File

@ -1,162 +0,0 @@
/***************************************************************************
Copyright (c) 2014, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary froms, 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 from 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 zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vxorpd %%ymm8 , %%ymm8 , %%ymm8 \n\t" // temp
"vxorpd %%ymm9 , %%ymm9 , %%ymm9 \n\t" // temp
"vxorpd %%ymm10, %%ymm10, %%ymm10 \n\t" // temp
"vxorpd %%ymm11, %%ymm11, %%ymm11 \n\t" // temp
"vxorpd %%ymm12, %%ymm12, %%ymm12 \n\t" // temp
"vxorpd %%ymm13, %%ymm13, %%ymm13 \n\t"
"vxorpd %%ymm14, %%ymm14, %%ymm14 \n\t"
"vxorpd %%ymm15, %%ymm15, %%ymm15 \n\t"
".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 192(%2,%0,8) \n\t"
"vmovddup (%2,%0,8), %%xmm0 \n\t" // real value from x0
"prefetcht0 192(%4,%0,8) \n\t"
"vmovups (%5,%0,8), %%ymm5 \n\t" // 2 complex values from a1
"vmovddup 8(%2,%0,8), %%xmm1 \n\t" // imag value from x0
"vmovups (%4,%0,8), %%ymm4 \n\t" // 2 complex values from a0
"prefetcht0 192(%5,%0,8) \n\t"
"vmovddup 16(%2,%0,8), %%xmm2 \n\t" // real value from x1
"prefetcht0 192(%6,%0,8) \n\t"
"vmovups (%6,%0,8), %%ymm6 \n\t" // 2 complex values from a2
"vmovddup 24(%2,%0,8), %%xmm3 \n\t" // imag value from x1
"prefetcht0 192(%7,%0,8) \n\t"
"vmovups (%7,%0,8), %%ymm7 \n\t" // 2 complex values from a3
"vinsertf128 $1, %%xmm2, %%ymm0 , %%ymm0 \n\t" // real values from x0 and x1
"vinsertf128 $1, %%xmm3, %%ymm1 , %%ymm1 \n\t" // imag values from x0 and x1
"vfmadd231pd %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231pd %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231pd %%ymm5 , %%ymm0, %%ymm10 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231pd %%ymm5 , %%ymm1, %%ymm11 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231pd %%ymm6 , %%ymm0, %%ymm12 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231pd %%ymm6 , %%ymm1, %%ymm13 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231pd %%ymm7 , %%ymm0, %%ymm14 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231pd %%ymm7 , %%ymm1, %%ymm15 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vmovups 32(%4,%0,8), %%ymm4 \n\t" // 2 complex values from a0
"vmovups 32(%5,%0,8), %%ymm5 \n\t" // 2 complex values from a1
"vmovddup 32(%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 40(%2,%0,8), %%xmm1 \n\t" // imag value from x0
"vmovddup 48(%2,%0,8), %%xmm2 \n\t" // real value from x1
"vmovddup 56(%2,%0,8), %%xmm3 \n\t" // imag value from x1
"vmovups 32(%6,%0,8), %%ymm6 \n\t" // 2 complex values from a2
"vmovups 32(%7,%0,8), %%ymm7 \n\t" // 2 complex values from a3
"vinsertf128 $1, %%xmm2, %%ymm0 , %%ymm0 \n\t" // real values from x0 and x1
"vinsertf128 $1, %%xmm3, %%ymm1 , %%ymm1 \n\t" // imag values from x0 and x1
"vfmadd231pd %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231pd %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231pd %%ymm5 , %%ymm0, %%ymm10 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231pd %%ymm5 , %%ymm1, %%ymm11 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231pd %%ymm6 , %%ymm0, %%ymm12 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231pd %%ymm6 , %%ymm1, %%ymm13 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231pd %%ymm7 , %%ymm0, %%ymm14 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231pd %%ymm7 , %%ymm1, %%ymm15 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"addq $8 , %0 \n\t"
"subq $4 , %1 \n\t"
"jnz .L01LOOP%= \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilpd $0x5 , %%ymm9 , %%ymm9 \n\t"
"vpermilpd $0x5 , %%ymm11, %%ymm11 \n\t"
"vpermilpd $0x5 , %%ymm13, %%ymm13 \n\t"
"vpermilpd $0x5 , %%ymm15, %%ymm15 \n\t"
"vaddsubpd %%ymm9 , %%ymm8, %%ymm8 \n\t"
"vaddsubpd %%ymm11, %%ymm10, %%ymm10 \n\t"
"vaddsubpd %%ymm13, %%ymm12, %%ymm12 \n\t"
"vaddsubpd %%ymm15, %%ymm14, %%ymm14 \n\t"
#else
"vpermilpd $0x5 , %%ymm8 , %%ymm8 \n\t"
"vpermilpd $0x5 , %%ymm10, %%ymm10 \n\t"
"vpermilpd $0x5 , %%ymm12, %%ymm12 \n\t"
"vpermilpd $0x5 , %%ymm14, %%ymm14 \n\t"
"vaddsubpd %%ymm8 , %%ymm9 , %%ymm8 \n\t"
"vaddsubpd %%ymm10, %%ymm11, %%ymm10 \n\t"
"vaddsubpd %%ymm12, %%ymm13, %%ymm12 \n\t"
"vaddsubpd %%ymm14, %%ymm15, %%ymm14 \n\t"
"vpermilpd $0x5 , %%ymm8 , %%ymm8 \n\t"
"vpermilpd $0x5 , %%ymm10, %%ymm10 \n\t"
"vpermilpd $0x5 , %%ymm12, %%ymm12 \n\t"
"vpermilpd $0x5 , %%ymm14, %%ymm14 \n\t"
#endif
"vextractf128 $1, %%ymm8 , %%xmm9 \n\t"
"vextractf128 $1, %%ymm10, %%xmm11 \n\t"
"vextractf128 $1, %%ymm12, %%xmm13 \n\t"
"vextractf128 $1, %%ymm14, %%xmm15 \n\t"
"vaddpd %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vaddpd %%xmm10, %%xmm11, %%xmm10 \n\t"
"vaddpd %%xmm12, %%xmm13, %%xmm12 \n\t"
"vaddpd %%xmm14, %%xmm15, %%xmm14 \n\t"
"vmovups %%xmm8 , (%3) \n\t"
"vmovups %%xmm10, 16(%3) \n\t"
"vmovups %%xmm12, 32(%3) \n\t"
"vmovups %%xmm14, 48(%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",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}

428
kernel/x86_64/zgemv_t_microk_haswell-4.c Normal file
View File

@ -0,0 +1,428 @@
/***************************************************************************
Copyright (c) 2014, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary froms, 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 from 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_4x4 1
static void zgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha) __attribute__ ((noinline));
static void zgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vxorpd %%ymm8 , %%ymm8 , %%ymm8 \n\t" // temp
"vxorpd %%ymm9 , %%ymm9 , %%ymm9 \n\t" // temp
"vxorpd %%ymm10, %%ymm10, %%ymm10 \n\t" // temp
"vxorpd %%ymm11, %%ymm11, %%ymm11 \n\t" // temp
"vxorpd %%ymm12, %%ymm12, %%ymm12 \n\t" // temp
"vxorpd %%ymm13, %%ymm13, %%ymm13 \n\t"
"vxorpd %%ymm14, %%ymm14, %%ymm14 \n\t"
"vxorpd %%ymm15, %%ymm15, %%ymm15 \n\t"
".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 192(%2,%0,8) \n\t"
"vmovddup (%2,%0,8), %%xmm0 \n\t" // real value from x0
"prefetcht0 192(%4,%0,8) \n\t"
"vmovups (%5,%0,8), %%ymm5 \n\t" // 2 complex values from a1
"vmovddup 8(%2,%0,8), %%xmm1 \n\t" // imag value from x0
"vmovups (%4,%0,8), %%ymm4 \n\t" // 2 complex values from a0
"prefetcht0 192(%5,%0,8) \n\t"
"vmovddup 16(%2,%0,8), %%xmm2 \n\t" // real value from x1
"prefetcht0 192(%6,%0,8) \n\t"
"vmovups (%6,%0,8), %%ymm6 \n\t" // 2 complex values from a2
"vmovddup 24(%2,%0,8), %%xmm3 \n\t" // imag value from x1
"prefetcht0 192(%7,%0,8) \n\t"
"vmovups (%7,%0,8), %%ymm7 \n\t" // 2 complex values from a3
"vinsertf128 $1, %%xmm2, %%ymm0 , %%ymm0 \n\t" // real values from x0 and x1
"vinsertf128 $1, %%xmm3, %%ymm1 , %%ymm1 \n\t" // imag values from x0 and x1
"vfmadd231pd %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231pd %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231pd %%ymm5 , %%ymm0, %%ymm10 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231pd %%ymm5 , %%ymm1, %%ymm11 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231pd %%ymm6 , %%ymm0, %%ymm12 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231pd %%ymm6 , %%ymm1, %%ymm13 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231pd %%ymm7 , %%ymm0, %%ymm14 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231pd %%ymm7 , %%ymm1, %%ymm15 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vmovups 32(%4,%0,8), %%ymm4 \n\t" // 2 complex values from a0
"vmovups 32(%5,%0,8), %%ymm5 \n\t" // 2 complex values from a1
"vmovddup 32(%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 40(%2,%0,8), %%xmm1 \n\t" // imag value from x0
"vmovddup 48(%2,%0,8), %%xmm2 \n\t" // real value from x1
"vmovddup 56(%2,%0,8), %%xmm3 \n\t" // imag value from x1
"vmovups 32(%6,%0,8), %%ymm6 \n\t" // 2 complex values from a2
"vmovups 32(%7,%0,8), %%ymm7 \n\t" // 2 complex values from a3
"vinsertf128 $1, %%xmm2, %%ymm0 , %%ymm0 \n\t" // real values from x0 and x1
"vinsertf128 $1, %%xmm3, %%ymm1 , %%ymm1 \n\t" // imag values from x0 and x1
"vfmadd231pd %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231pd %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231pd %%ymm5 , %%ymm0, %%ymm10 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231pd %%ymm5 , %%ymm1, %%ymm11 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231pd %%ymm6 , %%ymm0, %%ymm12 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231pd %%ymm6 , %%ymm1, %%ymm13 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231pd %%ymm7 , %%ymm0, %%ymm14 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231pd %%ymm7 , %%ymm1, %%ymm15 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"addq $8 , %0 \n\t"
"subq $4 , %1 \n\t"
"jnz .L01LOOP%= \n\t"
"vmovddup (%8) , %%xmm0 \n\t" // value from alpha
"vmovddup 8(%8) , %%xmm1 \n\t" // value from alpha
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilpd $0x5 , %%ymm9 , %%ymm9 \n\t"
"vpermilpd $0x5 , %%ymm11, %%ymm11 \n\t"
"vpermilpd $0x5 , %%ymm13, %%ymm13 \n\t"
"vpermilpd $0x5 , %%ymm15, %%ymm15 \n\t"
"vaddsubpd %%ymm9 , %%ymm8, %%ymm8 \n\t"
"vaddsubpd %%ymm11, %%ymm10, %%ymm10 \n\t"
"vaddsubpd %%ymm13, %%ymm12, %%ymm12 \n\t"
"vaddsubpd %%ymm15, %%ymm14, %%ymm14 \n\t"
#else
"vpermilpd $0x5 , %%ymm8 , %%ymm8 \n\t"
"vpermilpd $0x5 , %%ymm10, %%ymm10 \n\t"
"vpermilpd $0x5 , %%ymm12, %%ymm12 \n\t"
"vpermilpd $0x5 , %%ymm14, %%ymm14 \n\t"
"vaddsubpd %%ymm8 , %%ymm9 , %%ymm8 \n\t"
"vaddsubpd %%ymm10, %%ymm11, %%ymm10 \n\t"
"vaddsubpd %%ymm12, %%ymm13, %%ymm12 \n\t"
"vaddsubpd %%ymm14, %%ymm15, %%ymm14 \n\t"
"vpermilpd $0x5 , %%ymm8 , %%ymm8 \n\t"
"vpermilpd $0x5 , %%ymm10, %%ymm10 \n\t"
"vpermilpd $0x5 , %%ymm12, %%ymm12 \n\t"
"vpermilpd $0x5 , %%ymm14, %%ymm14 \n\t"
#endif
"vextractf128 $1, %%ymm8 , %%xmm9 \n\t"
"vextractf128 $1, %%ymm10, %%xmm11 \n\t"
"vextractf128 $1, %%ymm12, %%xmm13 \n\t"
"vextractf128 $1, %%ymm14, %%xmm15 \n\t"
"vaddpd %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vaddpd %%xmm10, %%xmm11, %%xmm10 \n\t"
"vaddpd %%xmm12, %%xmm13, %%xmm12 \n\t"
"vaddpd %%xmm14, %%xmm15, %%xmm14 \n\t"
"vmulpd %%xmm8 , %%xmm1 , %%xmm9 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulpd %%xmm8 , %%xmm0 , %%xmm8 \n\t" // t_r * alpha_r , t_i * alpha_r
"vmulpd %%xmm10, %%xmm1 , %%xmm11 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulpd %%xmm10, %%xmm0 , %%xmm10 \n\t" // t_r * alpha_r , t_i * alpha_r
"vmulpd %%xmm12, %%xmm1 , %%xmm13 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulpd %%xmm12, %%xmm0 , %%xmm12 \n\t" // t_r * alpha_r , t_i * alpha_r
"vmulpd %%xmm14, %%xmm1 , %%xmm15 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulpd %%xmm14, %%xmm0 , %%xmm14 \n\t" // t_r * alpha_r , t_i * alpha_r
#if !defined(XCONJ)
"vpermilpd $0x1 , %%xmm9 , %%xmm9 \n\t"
"vpermilpd $0x1 , %%xmm11, %%xmm11 \n\t"
"vpermilpd $0x1 , %%xmm13, %%xmm13 \n\t"
"vpermilpd $0x1 , %%xmm15, %%xmm15 \n\t"
"vaddsubpd %%xmm9 , %%xmm8, %%xmm8 \n\t"
"vaddsubpd %%xmm11, %%xmm10, %%xmm10 \n\t"
"vaddsubpd %%xmm13, %%xmm12, %%xmm12 \n\t"
"vaddsubpd %%xmm15, %%xmm14, %%xmm14 \n\t"
#else
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
"vpermilpd $0x1 , %%xmm10, %%xmm10 \n\t"
"vpermilpd $0x1 , %%xmm12, %%xmm12 \n\t"
"vpermilpd $0x1 , %%xmm14, %%xmm14 \n\t"
"vaddsubpd %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vaddsubpd %%xmm10, %%xmm11, %%xmm10 \n\t"
"vaddsubpd %%xmm12, %%xmm13, %%xmm12 \n\t"
"vaddsubpd %%xmm14, %%xmm15, %%xmm14 \n\t"
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
"vpermilpd $0x1 , %%xmm10, %%xmm10 \n\t"
"vpermilpd $0x1 , %%xmm12, %%xmm12 \n\t"
"vpermilpd $0x1 , %%xmm14, %%xmm14 \n\t"
#endif
"vaddpd (%3) , %%xmm8 , %%xmm8 \n\t"
"vaddpd 16(%3) , %%xmm10, %%xmm10 \n\t"
"vaddpd 32(%3) , %%xmm12, %%xmm12 \n\t"
"vaddpd 48(%3) , %%xmm14, %%xmm14 \n\t"
"vmovups %%xmm8 , (%3) \n\t"
"vmovups %%xmm10, 16(%3) \n\t"
"vmovups %%xmm12, 32(%3) \n\t"
"vmovups %%xmm14, 48(%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
"r" (alpha) // 8
: "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
#define HAVE_KERNEL_4x2 1
static void zgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha) __attribute__ ((noinline));
static void zgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vxorpd %%ymm8 , %%ymm8 , %%ymm8 \n\t" // temp
"vxorpd %%ymm9 , %%ymm9 , %%ymm9 \n\t" // temp
"vxorpd %%ymm10, %%ymm10, %%ymm10 \n\t" // temp
"vxorpd %%ymm11, %%ymm11, %%ymm11 \n\t" // temp
".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 192(%2,%0,8) \n\t"
"vmovddup (%2,%0,8), %%xmm0 \n\t" // real value from x0
"prefetcht0 192(%4,%0,8) \n\t"
"vmovups (%5,%0,8), %%ymm5 \n\t" // 2 complex values from a1
"vmovddup 8(%2,%0,8), %%xmm1 \n\t" // imag value from x0
"vmovups (%4,%0,8), %%ymm4 \n\t" // 2 complex values from a0
"prefetcht0 192(%5,%0,8) \n\t"
"vmovddup 16(%2,%0,8), %%xmm2 \n\t" // real value from x1
"vmovddup 24(%2,%0,8), %%xmm3 \n\t" // imag value from x1
"vinsertf128 $1, %%xmm2, %%ymm0 , %%ymm0 \n\t" // real values from x0 and x1
"vinsertf128 $1, %%xmm3, %%ymm1 , %%ymm1 \n\t" // imag values from x0 and x1
"vfmadd231pd %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231pd %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231pd %%ymm5 , %%ymm0, %%ymm10 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231pd %%ymm5 , %%ymm1, %%ymm11 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vmovups 32(%4,%0,8), %%ymm4 \n\t" // 2 complex values from a0
"vmovups 32(%5,%0,8), %%ymm5 \n\t" // 2 complex values from a1
"vmovddup 32(%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 40(%2,%0,8), %%xmm1 \n\t" // imag value from x0
"vmovddup 48(%2,%0,8), %%xmm2 \n\t" // real value from x1
"vmovddup 56(%2,%0,8), %%xmm3 \n\t" // imag value from x1
"vinsertf128 $1, %%xmm2, %%ymm0 , %%ymm0 \n\t" // real values from x0 and x1
"vinsertf128 $1, %%xmm3, %%ymm1 , %%ymm1 \n\t" // imag values from x0 and x1
"vfmadd231pd %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231pd %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231pd %%ymm5 , %%ymm0, %%ymm10 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231pd %%ymm5 , %%ymm1, %%ymm11 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"addq $8 , %0 \n\t"
"subq $4 , %1 \n\t"
"jnz .L01LOOP%= \n\t"
"vmovddup (%6) , %%xmm0 \n\t" // value from alpha
"vmovddup 8(%6) , %%xmm1 \n\t" // value from alpha
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilpd $0x5 , %%ymm9 , %%ymm9 \n\t"
"vpermilpd $0x5 , %%ymm11, %%ymm11 \n\t"
"vaddsubpd %%ymm9 , %%ymm8, %%ymm8 \n\t"
"vaddsubpd %%ymm11, %%ymm10, %%ymm10 \n\t"
#else
"vpermilpd $0x5 , %%ymm8 , %%ymm8 \n\t"
"vpermilpd $0x5 , %%ymm10, %%ymm10 \n\t"
"vaddsubpd %%ymm8 , %%ymm9 , %%ymm8 \n\t"
"vaddsubpd %%ymm10, %%ymm11, %%ymm10 \n\t"
"vpermilpd $0x5 , %%ymm8 , %%ymm8 \n\t"
"vpermilpd $0x5 , %%ymm10, %%ymm10 \n\t"
#endif
"vextractf128 $1, %%ymm8 , %%xmm9 \n\t"
"vextractf128 $1, %%ymm10, %%xmm11 \n\t"
"vaddpd %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vaddpd %%xmm10, %%xmm11, %%xmm10 \n\t"
"vmulpd %%xmm8 , %%xmm1 , %%xmm9 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulpd %%xmm8 , %%xmm0 , %%xmm8 \n\t" // t_r * alpha_r , t_i * alpha_r
"vmulpd %%xmm10, %%xmm1 , %%xmm11 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulpd %%xmm10, %%xmm0 , %%xmm10 \n\t" // t_r * alpha_r , t_i * alpha_r
#if !defined(XCONJ)
"vpermilpd $0x1 , %%xmm9 , %%xmm9 \n\t"
"vpermilpd $0x1 , %%xmm11, %%xmm11 \n\t"
"vaddsubpd %%xmm9 , %%xmm8, %%xmm8 \n\t"
"vaddsubpd %%xmm11, %%xmm10, %%xmm10 \n\t"
#else
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
"vpermilpd $0x1 , %%xmm10, %%xmm10 \n\t"
"vaddsubpd %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vaddsubpd %%xmm10, %%xmm11, %%xmm10 \n\t"
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
"vpermilpd $0x1 , %%xmm10, %%xmm10 \n\t"
#endif
"vaddpd (%3) , %%xmm8 , %%xmm8 \n\t"
"vaddpd 16(%3) , %%xmm10, %%xmm10 \n\t"
"vmovups %%xmm8 , (%3) \n\t"
"vmovups %%xmm10, 16(%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" (alpha) // 6
: "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
#define HAVE_KERNEL_4x1 1
static void zgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT *alpha) __attribute__ ((noinline));
static void zgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"vzeroupper \n\t"
"vxorpd %%ymm8 , %%ymm8 , %%ymm8 \n\t" // temp
"vxorpd %%ymm9 , %%ymm9 , %%ymm9 \n\t" // temp
".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 192(%2,%0,8) \n\t"
"vmovddup (%2,%0,8), %%xmm0 \n\t" // real value from x0
"prefetcht0 192(%4,%0,8) \n\t"
"vmovddup 8(%2,%0,8), %%xmm1 \n\t" // imag value from x0
"vmovups (%4,%0,8), %%ymm4 \n\t" // 2 complex values from a0
"vmovddup 16(%2,%0,8), %%xmm2 \n\t" // real value from x1
"vmovddup 24(%2,%0,8), %%xmm3 \n\t" // imag value from x1
"vinsertf128 $1, %%xmm2, %%ymm0 , %%ymm0 \n\t" // real values from x0 and x1
"vinsertf128 $1, %%xmm3, %%ymm1 , %%ymm1 \n\t" // imag values from x0 and x1
"vfmadd231pd %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231pd %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vmovups 32(%4,%0,8), %%ymm4 \n\t" // 2 complex values from a0
"vmovddup 32(%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 40(%2,%0,8), %%xmm1 \n\t" // imag value from x0
"vmovddup 48(%2,%0,8), %%xmm2 \n\t" // real value from x1
"vmovddup 56(%2,%0,8), %%xmm3 \n\t" // imag value from x1
"vinsertf128 $1, %%xmm2, %%ymm0 , %%ymm0 \n\t" // real values from x0 and x1
"vinsertf128 $1, %%xmm3, %%ymm1 , %%ymm1 \n\t" // imag values from x0 and x1
"vfmadd231pd %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231pd %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"addq $8 , %0 \n\t"
"subq $4 , %1 \n\t"
"jnz .L01LOOP%= \n\t"
"vmovddup (%5) , %%xmm0 \n\t" // value from alpha
"vmovddup 8(%5) , %%xmm1 \n\t" // value from alpha
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilpd $0x5 , %%ymm9 , %%ymm9 \n\t"
"vaddsubpd %%ymm9 , %%ymm8, %%ymm8 \n\t"
#else
"vpermilpd $0x5 , %%ymm8 , %%ymm8 \n\t"
"vaddsubpd %%ymm8 , %%ymm9 , %%ymm8 \n\t"
"vpermilpd $0x5 , %%ymm8 , %%ymm8 \n\t"
#endif
"vextractf128 $1, %%ymm8 , %%xmm9 \n\t"
"vaddpd %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vmulpd %%xmm8 , %%xmm1 , %%xmm9 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulpd %%xmm8 , %%xmm0 , %%xmm8 \n\t" // t_r * alpha_r , t_i * alpha_r
#if !defined(XCONJ)
"vpermilpd $0x1 , %%xmm9 , %%xmm9 \n\t"
"vaddsubpd %%xmm9 , %%xmm8, %%xmm8 \n\t"
#else
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
"vaddsubpd %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
#endif
"vaddpd (%3) , %%xmm8 , %%xmm8 \n\t"
"vmovups %%xmm8 , (%3) \n\t"
"vzeroupper \n\t"
:
:
"r" (i), // 0
"r" (n), // 1
"r" (x), // 2
"r" (y), // 3
"r" (ap), // 4
"r" (alpha) // 5
: "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}