floraachy/OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
f9d67bb5e8e895fd5fe7e36e43febef7aa06ef35
OpenBLAS/kernel/x86_64/dgemv_n_4.c
Arjan van de Ven 87bebdbd8a Add an AVX512 enabled DGEMV (n) function 
written in C intrinsics for best readability.
(the same C code works for Haswell as well)

For logistical reasons the code falls back to the existing
haswell AVX2 implementation if the GCC or LLVM compiler is not new enough
2018-08-11 17:38:12 +00:00

487 lines
11 KiB
C
Raw Blame History

/***************************************************************************
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_n_microk_nehalem-4.c"
#elif defined(HASWELL) || defined(ZEN) || defined(STEAMROLLER) || defined(EXCAVATOR)
#include "dgemv_n_microk_haswell-4.c"
#elif defined (SKYLAKEX)
#include "dgemv_n_microk_skylakex-4.c"
#endif
#define NBMAX 2048
#ifndef HAVE_KERNEL_4x4
static void dgemv_kernel_4x4(BLASLONG n, FLOAT **ap, FLOAT *xo, FLOAT *y, FLOAT *alpha)
{
BLASLONG i;
FLOAT *a0,*a1,*a2,*a3;
FLOAT x[4];
a0 = ap[0];
a1 = ap[1];
a2 = ap[2];
a3 = ap[3];
for ( i=0; i<4; i++)
x[i] = xo[i] * *alpha;
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
#ifndef HAVE_KERNEL_4x2
static void dgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha) __attribute__ ((noinline));
static void dgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"movsd (%2) , %%xmm12 \n\t" // x0
"movsd (%6) , %%xmm4 \n\t" // alpha
"movsd 8(%2) , %%xmm13 \n\t" // x1
"mulsd %%xmm4 , %%xmm12 \n\t" // alpha
"mulsd %%xmm4 , %%xmm13 \n\t" // alpha
"shufpd $0, %%xmm12, %%xmm12 \n\t"
"shufpd $0, %%xmm13, %%xmm13 \n\t"
// ".align 16 \n\t"
"1: \n\t"
"movups (%3,%0,8), %%xmm4 \n\t" // 2 * y
"movups 16(%3,%0,8), %%xmm5 \n\t" // 2 * y
"movups (%4,%0,8), %%xmm8 \n\t"
"movups (%5,%0,8), %%xmm9 \n\t"
"mulpd %%xmm12, %%xmm8 \n\t"
"mulpd %%xmm13, %%xmm9 \n\t"
"addpd %%xmm8 , %%xmm4 \n\t"
"addpd %%xmm9 , %%xmm4 \n\t"
"movups 16(%4,%0,8), %%xmm8 \n\t"
"movups 16(%5,%0,8), %%xmm9 \n\t"
"mulpd %%xmm12, %%xmm8 \n\t"
"mulpd %%xmm13, %%xmm9 \n\t"
"addpd %%xmm8 , %%xmm5 \n\t"
"addpd %%xmm9 , %%xmm5 \n\t"
"movups %%xmm4 , (%3,%0,8) \n\t" // 2 * y
"movups %%xmm5 , 16(%3,%0,8) \n\t" // 2 * y
"addq $4 , %0 \n\t"
"subq $4 , %1 \n\t"
"jnz 1b \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",
"%xmm4", "%xmm5",
"%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
#endif
#ifndef HAVE_KERNEL_4x1
static void dgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT *alpha) __attribute__ ((noinline));
static void dgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{
BLASLONG register i = 0;
__asm__ __volatile__
(
"movsd (%2), %%xmm12 \n\t" // x0
"mulsd (%5), %%xmm12 \n\t" // alpha
"shufpd $0, %%xmm12, %%xmm12 \n\t"
// ".align 16 \n\t"
"1: \n\t"
"movups (%4,%0,8), %%xmm8 \n\t" // 2 * a
"movups 16(%4,%0,8), %%xmm9 \n\t" // 2 * a
"movups (%3,%0,8), %%xmm4 \n\t" // 2 * y
"movups 16(%3,%0,8), %%xmm5 \n\t" // 2 * y
"mulpd %%xmm12, %%xmm8 \n\t"
"mulpd %%xmm12, %%xmm9 \n\t"
"addpd %%xmm8 , %%xmm4 \n\t"
"addpd %%xmm9 , %%xmm5 \n\t"
"movups %%xmm4 , (%3,%0,8) \n\t" // 2 * y
"movups %%xmm5 , 16(%3,%0,8) \n\t" // 2 * y
"addq $4 , %0 \n\t"
"subq $4 , %1 \n\t"
"jnz 1b \n\t"
:
:
"r" (i), // 0
"r" (n), // 1
"r" (x), // 2
"r" (y), // 3
"r" (ap), // 4
"r" (alpha) // 5
: "cc",
"%xmm4", "%xmm5",
"%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
#endif
static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest) __attribute__ ((noinline));
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++ )
{
*dest += *src;
src++;
dest += inc_dest;
}
return;
}
}
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;
FLOAT *a_ptr;
FLOAT *x_ptr;
FLOAT *y_ptr;
FLOAT *ap[4];
BLASLONG n1;
BLASLONG m1;
BLASLONG m2;
BLASLONG m3;
BLASLONG n2;
BLASLONG lda4 = lda << 2;
FLOAT xbuffer[8],*ybuffer;
if ( m < 1 ) return(0);
if ( n < 1 ) return(0);
ybuffer = buffer;
n1 = n >> 2 ;
n2 = n & 3 ;
m3 = m & 3 ;
m1 = m & -4 ;
m2 = (m & (NBMAX-1)) - m3 ;
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;
ap[0] = a_ptr;
ap[1] = a_ptr + lda;
ap[2] = ap[1] + lda;
ap[3] = ap[2] + lda;
if ( inc_y != 1 )
memset(ybuffer,0,NB*8);
else
ybuffer = y_ptr;
if ( inc_x == 1 )
{
for( i = 0; i < n1 ; i++)
{
dgemv_kernel_4x4(NB,ap,x_ptr,ybuffer,&alpha);
ap[0] += lda4;
ap[1] += lda4;
ap[2] += lda4;
ap[3] += lda4;
a_ptr += lda4;
x_ptr += 4;
}
if ( n2 & 2 )
{
dgemv_kernel_4x2(NB,ap,x_ptr,ybuffer,&alpha);
a_ptr += lda*2;
x_ptr += 2;
}
if ( n2 & 1 )
{
dgemv_kernel_4x1(NB,a_ptr,x_ptr,ybuffer,&alpha);
/* a_ptr += lda;
x_ptr += 1; */
}
}
else
{
for( i = 0; i < n1 ; i++)
{
xbuffer[0] = x_ptr[0];
x_ptr += inc_x;
xbuffer[1] = x_ptr[0];
x_ptr += inc_x;
xbuffer[2] = x_ptr[0];
x_ptr += inc_x;
xbuffer[3] = x_ptr[0];
x_ptr += inc_x;
dgemv_kernel_4x4(NB,ap,xbuffer,ybuffer,&alpha);
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];
x_ptr += inc_x;
dgemv_kernel_4x1(NB,a_ptr,xbuffer,ybuffer,&alpha);
a_ptr += lda;
}
}
a += NB;
if ( inc_y != 1 )
{
add_y(NB,ybuffer,y_ptr,inc_y);
y_ptr += NB * inc_y;
}
else
y_ptr += NB ;
}
if ( m3 == 0 ) return(0);
if ( m3 == 3 )
{
a_ptr = a;
x_ptr = x;
FLOAT temp0 = 0.0;
FLOAT temp1 = 0.0;
FLOAT temp2 = 0.0;
if ( lda == 3 && inc_x ==1 )
{
for( i = 0; i < ( n & -4 ); i+=4 )
{
temp0 += a_ptr[0] * x_ptr[0] + a_ptr[3] * x_ptr[1];
temp1 += a_ptr[1] * x_ptr[0] + a_ptr[4] * x_ptr[1];
temp2 += a_ptr[2] * x_ptr[0] + a_ptr[5] * x_ptr[1];
temp0 += a_ptr[6] * x_ptr[2] + a_ptr[9] * x_ptr[3];
temp1 += a_ptr[7] * x_ptr[2] + a_ptr[10] * x_ptr[3];
temp2 += a_ptr[8] * x_ptr[2] + a_ptr[11] * x_ptr[3];
a_ptr += 12;
x_ptr += 4;
}
for( ; i < n; i++ )
{
temp0 += a_ptr[0] * x_ptr[0];
temp1 += a_ptr[1] * x_ptr[0];
temp2 += a_ptr[2] * x_ptr[0];
a_ptr += 3;
x_ptr ++;
}
}
else
{
for( i = 0; i < n; i++ )
{
temp0 += a_ptr[0] * x_ptr[0];
temp1 += a_ptr[1] * x_ptr[0];
temp2 += a_ptr[2] * x_ptr[0];
a_ptr += lda;
x_ptr += inc_x;
}
}
y_ptr[0] += alpha * temp0;
y_ptr += inc_y;
y_ptr[0] += alpha * temp1;
y_ptr += inc_y;
y_ptr[0] += alpha * temp2;
return(0);
}
if ( m3 == 2 )
{
a_ptr = a;
x_ptr = x;
FLOAT temp0 = 0.0;
FLOAT temp1 = 0.0;
if ( lda == 2 && inc_x ==1 )
{
for( i = 0; i < (n & -4) ; i+=4 )
{
temp0 += a_ptr[0] * x_ptr[0] + a_ptr[2] * x_ptr[1];
temp1 += a_ptr[1] * x_ptr[0] + a_ptr[3] * x_ptr[1];
temp0 += a_ptr[4] * x_ptr[2] + a_ptr[6] * x_ptr[3];
temp1 += a_ptr[5] * x_ptr[2] + a_ptr[7] * x_ptr[3];
a_ptr += 8;
x_ptr += 4;
}
for( ; i < n; i++ )
{
temp0 += a_ptr[0] * x_ptr[0];
temp1 += a_ptr[1] * x_ptr[0];
a_ptr += 2;
x_ptr ++;
}
}
else
{
for( i = 0; i < n; i++ )
{
temp0 += a_ptr[0] * x_ptr[0];
temp1 += a_ptr[1] * x_ptr[0];
a_ptr += lda;
x_ptr += inc_x;
}
}
y_ptr[0] += alpha * temp0;
y_ptr += inc_y;
y_ptr[0] += alpha * temp1;
return(0);
}
if ( m3 == 1 )
{
a_ptr = a;
x_ptr = x;
FLOAT temp = 0.0;
if ( lda == 1 && inc_x ==1 )
{
for( i = 0; i < (n & -4); i+=4 )
{
temp += a_ptr[i] * x_ptr[i] + a_ptr[i+1] * x_ptr[i+1] + a_ptr[i+2] * x_ptr[i+2] + a_ptr[i+3] * x_ptr[i+3];
}
for( ; i < n; i++ )
{
temp += a_ptr[i] * x_ptr[i];
}
}
else
{
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;
return(0);
}
return(0);
}
Reference in New Issue View Git Blame Copy Permalink