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 zgemv_n kernel

tags/v0.2.12^2
wernsaar 11 years ago
parent
cafba99b6b
commit
bced4594bb
2 changed files with 920 additions and 0 deletions
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +623
    -0
      kernel/x86_64/zgemv_n_4.c
  2. +297
    -0
      kernel/x86_64/zgemv_n_microk_haswell-4.c

+ 623
- 0
kernel/x86_64/zgemv_n_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 <stdlib.h>
#include <stdio.h>
#include "common.h"

#if defined(HASWELL)
#include "zgemv_n_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)
{
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*8);

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);
}



+ 297
- 0
kernel/x86_64/zgemv_n_microk_haswell-4.c View File

@@ -0,0 +1,297 @@
/***************************************************************************
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"
);

}




Write Preview
Loading…
Cancel
Save