OSchip
/
OpenBLAS
Not watched
1
0
Fork 0
Code
Releases 66 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
Browse Source

added optimized dgemv_t kernel for haswell

tags/v0.2.12^2
wernsaar 11 years ago
parent
cd34e9701b
commit
44f2bf9bae
2 changed files with 750 additions and 0 deletions
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +623
    -0
      kernel/x86_64/dgemv_t_4.c
  2. +127
    -0
      kernel/x86_64/dgemv_t_microk_haswell-4.c

+ 623
- 0
kernel/x86_64/dgemv_t_4.c 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 "common.h"

/*
#if defined(NEHALEM)
#include "dgemv_t_microk_nehalem-4.c"
#elif defined(BULLDOZER) || defined(PILEDRIVER)
#include "dgemv_t_microk_bulldozer-4.c"
#elif defined(SANDYBRIDGE)
#include "dgemv_t_microk_sandy-4.c"
#elif defined(HASWELL)
#include "dgemv_t_microk_haswell-4.c"
#endif
*/

#define NBMAX 2048

#ifndef HAVE_KERNEL_4x4

static void dgemv_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];
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_4x2(BLASLONG n, FLOAT *ap0, FLOAT *ap1, FLOAT *x, FLOAT *y) __attribute__ ((noinline));

static void dgemv_kernel_4x2(BLASLONG n, FLOAT *ap0, FLOAT *ap1, FLOAT *x, FLOAT *y)
{
BLASLONG i;

i=0;

__asm__ __volatile__
(
"xorpd %%xmm10 , %%xmm10 \n\t"
"xorpd %%xmm11 , %%xmm11 \n\t"
"testq $2 , %1 \n\t"
"jz .L01LABEL%= \n\t"

"movups (%5,%0,8) , %%xmm14 \n\t" // x
"movups (%3,%0,8) , %%xmm12 \n\t" // ap0
"movups (%4,%0,8) , %%xmm13 \n\t" // ap1
"mulpd %%xmm14 , %%xmm12 \n\t"
"mulpd %%xmm14 , %%xmm13 \n\t"
"addq $2 , %0 \n\t"
"addpd %%xmm12 , %%xmm10 \n\t"
"subq $2 , %1 \n\t"
"addpd %%xmm13 , %%xmm11 \n\t"

".L01LABEL%=: \n\t"

"cmpq $0, %1 \n\t"
"je .L01END%= \n\t"

".align 16 \n\t"
".L01LOOP%=: \n\t"

"movups (%5,%0,8) , %%xmm14 \n\t" // x
"movups (%3,%0,8) , %%xmm12 \n\t" // ap0
"movups (%4,%0,8) , %%xmm13 \n\t" // ap1
"mulpd %%xmm14 , %%xmm12 \n\t"
"mulpd %%xmm14 , %%xmm13 \n\t"
"addpd %%xmm12 , %%xmm10 \n\t"
"addpd %%xmm13 , %%xmm11 \n\t"

"movups 16(%5,%0,8) , %%xmm14 \n\t" // x
"movups 16(%3,%0,8) , %%xmm12 \n\t" // ap0
"movups 16(%4,%0,8) , %%xmm13 \n\t" // ap1
"mulpd %%xmm14 , %%xmm12 \n\t"
"mulpd %%xmm14 , %%xmm13 \n\t"
"addpd %%xmm12 , %%xmm10 \n\t"
"addpd %%xmm13 , %%xmm11 \n\t"

"addq $4 , %0 \n\t"
"subq $4 , %1 \n\t"
"jnz .L01LOOP%= \n\t"

".L01END%=: \n\t"

"haddpd %%xmm10, %%xmm10 \n\t"
"haddpd %%xmm11, %%xmm11 \n\t"

"movsd %%xmm10, (%2) \n\t"
"movsd %%xmm11,8(%2) \n\t"

:
:
"r" (i), // 0
"r" (n), // 1
"r" (y), // 2
"r" (ap0), // 3
"r" (ap1), // 4
"r" (x) // 5
: "cc",
"%xmm4", "%xmm5", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);


}
static void dgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));

static void dgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
{
BLASLONG i;

i=0;

__asm__ __volatile__
(
"xorpd %%xmm9 , %%xmm9 \n\t"
"xorpd %%xmm10 , %%xmm10 \n\t"
"testq $2 , %1 \n\t"
"jz .L01LABEL%= \n\t"

"movups (%3,%0,8) , %%xmm12 \n\t"
"movups (%4,%0,8) , %%xmm11 \n\t"
"mulpd %%xmm11 , %%xmm12 \n\t"
"addq $2 , %0 \n\t"
"addpd %%xmm12 , %%xmm10 \n\t"
"subq $2 , %1 \n\t"

".L01LABEL%=: \n\t"

"cmpq $0, %1 \n\t"
"je .L01END%= \n\t"

".align 16 \n\t"
".L01LOOP%=: \n\t"

"movups (%3,%0,8) , %%xmm12 \n\t"
"movups 16(%3,%0,8) , %%xmm14 \n\t"
"movups (%4,%0,8) , %%xmm11 \n\t"
"movups 16(%4,%0,8) , %%xmm13 \n\t"
"mulpd %%xmm11 , %%xmm12 \n\t"
"mulpd %%xmm13 , %%xmm14 \n\t"
"addq $4 , %0 \n\t"
"addpd %%xmm12 , %%xmm10 \n\t"
"subq $4 , %1 \n\t"
"addpd %%xmm14 , %%xmm9 \n\t"

"jnz .L01LOOP%= \n\t"

".L01END%=: \n\t"

"addpd %%xmm9 , %%xmm10 \n\t"
"haddpd %%xmm10, %%xmm10 \n\t"

"movsd %%xmm10, (%2) \n\t"

:
:
"r" (i), // 0
"r" (n), // 1
"r" (y), // 2
"r" (ap), // 3
"r" (x) // 4
: "cc",
"%xmm9", "%xmm10" ,
"%xmm11", "%xmm12", "%xmm13", "%xmm14",
"memory"
);


}
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;
}
}

static void add_y(BLASLONG n, FLOAT da , FLOAT *src, FLOAT *dest, BLASLONG inc_dest) __attribute__ ((noinline));

static void add_y(BLASLONG n, FLOAT da , FLOAT *src, FLOAT *dest, BLASLONG inc_dest)
{

BLASLONG i;

if ( inc_dest != 1 )
{
for ( i=0; i<n; i++ )
{
*dest += src[i] * da;
dest += inc_dest;
}
return;
}

i=0;

__asm__ __volatile__
(
"movsd (%2) , %%xmm10 \n\t"
"shufpd $0 , %%xmm10 , %%xmm10 \n\t"

".align 16 \n\t"
".L01LOOP%=: \n\t"

"movups (%3,%0,8) , %%xmm12 \n\t"
"movups (%4,%0,8) , %%xmm11 \n\t"
"mulpd %%xmm10 , %%xmm12 \n\t"
"addq $2 , %0 \n\t"
"addpd %%xmm12 , %%xmm11 \n\t"
"subq $2 , %1 \n\t"
"movups %%xmm11, -16(%4,%0,8) \n\t"

"jnz .L01LOOP%= \n\t"

:
:
"r" (i), // 0
"r" (n), // 1
"r" (&da), // 2
"r" (src), // 3
"r" (dest) // 4
: "cc",
"%xmm10", "%xmm11", "%xmm12",
"memory"
);


}

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 register i;
BLASLONG register j;
FLOAT *a_ptr;
FLOAT *x_ptr;
FLOAT *y_ptr;
BLASLONG n0;
BLASLONG n1;
BLASLONG m1;
BLASLONG m2;
BLASLONG m3;
BLASLONG n2;
FLOAT ybuffer[4],*xbuffer;
FLOAT *ytemp;

if ( m < 1 ) return(0);
if ( n < 1 ) return(0);

xbuffer = buffer;
ytemp = buffer + NBMAX;
n0 = n / NBMAX;
n1 = (n % NBMAX) >> 2 ;
n2 = n & 3 ;

m3 = m & 3 ;
m1 = m & -4 ;
m2 = (m & (NBMAX-1)) - m3 ;


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;

if ( inc_x == 1 )
xbuffer = x_ptr;
else
copy_x(NB,x_ptr,xbuffer,inc_x);


FLOAT *ap[4];
FLOAT *yp;
BLASLONG register lda4 = 4 * lda;
ap[0] = a_ptr;
ap[1] = a_ptr + lda;
ap[2] = ap[1] + lda;
ap[3] = ap[2] + lda;

if ( n0 > 0 )
{
BLASLONG nb1 = NBMAX / 4;
for( j=0; j<n0; j++)
{

yp = ytemp;
for( i = 0; i < nb1 ; i++)
{
dgemv_kernel_4x4(NB,ap,xbuffer,yp);
ap[0] += lda4 ;
ap[1] += lda4 ;
ap[2] += lda4 ;
ap[3] += lda4 ;
yp += 4;
}
add_y(nb1*4, alpha, ytemp, y_ptr, inc_y );
y_ptr += nb1 * inc_y * 4;
a_ptr += nb1 * lda4 ;

}

}


yp = ytemp;

for( i = 0; i < n1 ; i++)
{
dgemv_kernel_4x4(NB,ap,xbuffer,yp);
ap[0] += lda4 ;
ap[1] += lda4 ;
ap[2] += lda4 ;
ap[3] += lda4 ;
yp += 4;
}
if ( n1 > 0 )
{
add_y(n1*4, alpha, ytemp, y_ptr, inc_y );
y_ptr += n1 * inc_y * 4;
a_ptr += n1 * lda4 ;
}

if ( n2 & 2 )
{

dgemv_kernel_4x2(NB,ap[0],ap[1],xbuffer,ybuffer);
a_ptr += lda * 2;
*y_ptr += ybuffer[0] * alpha;
y_ptr += inc_y;
*y_ptr += ybuffer[1] * alpha;
y_ptr += inc_y;

}

if ( n2 & 1 )
{

dgemv_kernel_4x1(NB,a_ptr,xbuffer,ybuffer);
a_ptr += lda;
*y_ptr += ybuffer[0] * alpha;
y_ptr += inc_y;

}
a += NB;
x += NB * inc_x;
}

if ( m3 == 0 ) return(0);

x_ptr = x;
a_ptr = a;
if ( m3 == 3 )
{
FLOAT xtemp0 = *x_ptr * alpha;
x_ptr += inc_x;
FLOAT xtemp1 = *x_ptr * alpha;
x_ptr += inc_x;
FLOAT xtemp2 = *x_ptr * alpha;

FLOAT *aj = a_ptr;
y_ptr = y;

if ( lda == 3 && inc_y == 1 )
{

for ( j=0; j< ( n & -4) ; j+=4 )
{

y_ptr[j] += aj[0] * xtemp0 + aj[1] * xtemp1 + aj[2] * xtemp2;
y_ptr[j+1] += aj[3] * xtemp0 + aj[4] * xtemp1 + aj[5] * xtemp2;
y_ptr[j+2] += aj[6] * xtemp0 + aj[7] * xtemp1 + aj[8] * xtemp2;
y_ptr[j+3] += aj[9] * xtemp0 + aj[10] * xtemp1 + aj[11] * xtemp2;
aj += 12;
}

for ( ; j<n; j++ )
{
y_ptr[j] += aj[0] * xtemp0 + aj[1] * xtemp1 + aj[2] * xtemp2;
aj += 3;
}

}
else
{

if ( inc_y == 1 )
{

BLASLONG register lda2 = lda << 1;
BLASLONG register lda4 = lda << 2;
BLASLONG register lda3 = lda2 + lda;

for ( j=0; j< ( n & -4 ); j+=4 )
{

y_ptr[j] += *aj * xtemp0 + *(aj+1) * xtemp1 + *(aj+2) * xtemp2;
y_ptr[j+1] += *(aj+lda) * xtemp0 + *(aj+lda+1) * xtemp1 + *(aj+lda+2) * xtemp2;
y_ptr[j+2] += *(aj+lda2) * xtemp0 + *(aj+lda2+1) * xtemp1 + *(aj+lda2+2) * xtemp2;
y_ptr[j+3] += *(aj+lda3) * xtemp0 + *(aj+lda3+1) * xtemp1 + *(aj+lda3+2) * xtemp2;
aj += lda4;
}

for ( ; j< n ; j++ )
{

y_ptr[j] += *aj * xtemp0 + *(aj+1) * xtemp1 + *(aj+2) * xtemp2 ;
aj += lda;
}

}
else
{

for ( j=0; j<n; j++ )
{
*y_ptr += *aj * xtemp0 + *(aj+1) * xtemp1 + *(aj+2) * xtemp2;
y_ptr += inc_y;
aj += lda;
}


}

}
return(0);
}

if ( m3 == 2 )
{
FLOAT xtemp0 = *x_ptr * alpha;
x_ptr += inc_x;
FLOAT xtemp1 = *x_ptr * alpha;

FLOAT *aj = a_ptr;
y_ptr = y;

if ( lda == 2 && inc_y == 1 )
{

for ( j=0; j< ( n & -4) ; j+=4 )
{
y_ptr[j] += aj[0] * xtemp0 + aj[1] * xtemp1 ;
y_ptr[j+1] += aj[2] * xtemp0 + aj[3] * xtemp1 ;
y_ptr[j+2] += aj[4] * xtemp0 + aj[5] * xtemp1 ;
y_ptr[j+3] += aj[6] * xtemp0 + aj[7] * xtemp1 ;
aj += 8;

}

for ( ; j<n; j++ )
{
y_ptr[j] += aj[0] * xtemp0 + aj[1] * xtemp1 ;
aj += 2;
}

}
else
{
if ( inc_y == 1 )
{

BLASLONG register lda2 = lda << 1;
BLASLONG register lda4 = lda << 2;
BLASLONG register lda3 = lda2 + lda;

for ( j=0; j< ( n & -4 ); j+=4 )
{

y_ptr[j] += *aj * xtemp0 + *(aj+1) * xtemp1 ;
y_ptr[j+1] += *(aj+lda) * xtemp0 + *(aj+lda+1) * xtemp1 ;
y_ptr[j+2] += *(aj+lda2) * xtemp0 + *(aj+lda2+1) * xtemp1 ;
y_ptr[j+3] += *(aj+lda3) * xtemp0 + *(aj+lda3+1) * xtemp1 ;
aj += lda4;
}

for ( ; j< n ; j++ )
{

y_ptr[j] += *aj * xtemp0 + *(aj+1) * xtemp1 ;
aj += lda;
}

}
else
{
for ( j=0; j<n; j++ )
{
*y_ptr += *aj * xtemp0 + *(aj+1) * xtemp1 ;
y_ptr += inc_y;
aj += lda;
}
}

}
return(0);

}

FLOAT xtemp = *x_ptr * alpha;
FLOAT *aj = a_ptr;
y_ptr = y;
if ( lda == 1 && inc_y == 1 )
{
for ( j=0; j< ( n & -4) ; j+=4 )
{
y_ptr[j] += aj[j] * xtemp;
y_ptr[j+1] += aj[j+1] * xtemp;
y_ptr[j+2] += aj[j+2] * xtemp;
y_ptr[j+3] += aj[j+3] * xtemp;
}
for ( ; j<n ; j++ )
{
y_ptr[j] += aj[j] * xtemp;
}



}
else
{
if ( inc_y == 1 )
{

BLASLONG register lda2 = lda << 1;
BLASLONG register lda4 = lda << 2;
BLASLONG register lda3 = lda2 + lda;
for ( j=0; j< ( n & -4 ); j+=4 )
{
y_ptr[j] += *aj * xtemp;
y_ptr[j+1] += *(aj+lda) * xtemp;
y_ptr[j+2] += *(aj+lda2) * xtemp;
y_ptr[j+3] += *(aj+lda3) * xtemp;
aj += lda4 ;
}

for ( ; j<n; j++ )
{
y_ptr[j] += *aj * xtemp;
aj += lda;
}

}
else
{
for ( j=0; j<n; j++ )
{
*y_ptr += *aj * xtemp;
y_ptr += inc_y;
aj += lda;
}

}
}

return(0);
}



+ 127
- 0
kernel/x86_64/dgemv_t_microk_haswell-4.c View File

@@ -0,0 +1,127 @@
/***************************************************************************
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 dgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));

static void dgemv_kernel_4x4( 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"

"testq $0x04, %1 \n\t"
"jz .L08LABEL%= \n\t"

"vmovups (%2,%0,8), %%ymm12 \n\t" // 4 * x

"vfmadd231pd (%4,%0,8), %%ymm12, %%ymm4 \n\t"
"vfmadd231pd (%5,%0,8), %%ymm12, %%ymm5 \n\t"
"vfmadd231pd (%6,%0,8), %%ymm12, %%ymm6 \n\t"
"vfmadd231pd (%7,%0,8), %%ymm12, %%ymm7 \n\t"

"addq $4 , %0 \n\t"
"subq $4 , %1 \n\t"

".L08LABEL%=: \n\t"

"cmpq $0, %1 \n\t"
"je .L16END%= \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"

".L16END%=: \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"
);

}



Write Preview
Loading…
Cancel
Save