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

removed obsolete files

tags/v0.2.12^2
wernsaar 11 years ago
parent
658939faaa
commit
cd34e9701b
8 changed files with 0 additions and 2140 deletions
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +0
    -218
      kernel/x86_64/sgemv_n_avx.c
  2. +0
    -451
      kernel/x86_64/sgemv_n_microk_bulldozer.c
  3. +0
    -461
      kernel/x86_64/sgemv_n_microk_haswell.c
  4. +0
    -473
      kernel/x86_64/sgemv_n_microk_sandy.c
  5. +0
    -232
      kernel/x86_64/sgemv_t_avx.c
  6. +0
    -99
      kernel/x86_64/sgemv_t_microk_bulldozer.c
  7. +0
    -100
      kernel/x86_64/sgemv_t_microk_haswell.c
  8. +0
    -106
      kernel/x86_64/sgemv_t_microk_sandy.c

+ 0
- 218
kernel/x86_64/sgemv_n_avx.c View File

@@ -1,218 +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 "sgemv_n_microk_bulldozer.c"
#elif defined(HASWELL)
#include "sgemv_n_microk_haswell.c"
#else
#include "sgemv_n_microk_sandy.c"
#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++;
src += inc_src;
}
}

static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest)
{
BLASLONG i;
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;
BLASLONG n1;
BLASLONG m1;
BLASLONG register m2;
BLASLONG register n2;
FLOAT *xbuffer,*ybuffer;
xbuffer = buffer;
ybuffer = xbuffer + 2048 + 256;
n1 = n / 512 ;
n2 = n % 512 ;

m1 = m / 64;
m2 = m % 64;

y_ptr = y;
x_ptr = x;

for (j=0; j<n1; j++)
{

if ( inc_x == 1 )
xbuffer = x_ptr;
else
copy_x(512,x_ptr,xbuffer,inc_x);

a_ptr = a + j * 512 * lda;
y_ptr = y;

for(i = 0; i<m1; i++ )
{
sgemv_kernel_64(512,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(64,ybuffer,y_ptr,inc_y);
y_ptr += 64 * inc_y;
a_ptr += 64;

}

if ( m2 & 32 )
{
sgemv_kernel_32(512,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(32,ybuffer,y_ptr,inc_y);
y_ptr += 32 * inc_y;
a_ptr += 32;

}

if ( m2 & 16 )
{
sgemv_kernel_16(512,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(16,ybuffer,y_ptr,inc_y);
y_ptr += 16 * inc_y;
a_ptr += 16;
}
if ( m2 & 8 )
{
sgemv_kernel_8(512,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(8,ybuffer,y_ptr,inc_y);
y_ptr += 8 * inc_y;
a_ptr += 8;
}
if ( m2 & 4 )
{
sgemv_kernel_4(512,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(4,ybuffer,y_ptr,inc_y);
y_ptr += 4 * inc_y;
a_ptr += 4;
}
if ( m2 & 2 )
{
sgemv_kernel_2(512,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(2,ybuffer,y_ptr,inc_y);
y_ptr += 2 * inc_y;
a_ptr += 2;
}
if ( m2 & 1 )
{
sgemv_kernel_1(512,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(1,ybuffer,y_ptr,inc_y);
}
x_ptr += 512 * inc_x;

}

if ( n2 > 0 )
{

if ( inc_x == 1 )
xbuffer = x_ptr;
else
copy_x(n2,x_ptr,xbuffer,inc_x);

a_ptr = a + n1 * 512 * lda;
y_ptr = y;

for(i = 0; i<m1; i++ )
{
sgemv_kernel_64(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(64,ybuffer,y_ptr,inc_y);
y_ptr += 64 * inc_y;
a_ptr += 64;

}

if ( m2 & 32 )
{
sgemv_kernel_32(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(32,ybuffer,y_ptr,inc_y);
y_ptr += 32 * inc_y;
a_ptr += 32;

}
if ( m2 & 16 )
{
sgemv_kernel_16(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(16,ybuffer,y_ptr,inc_y);
y_ptr += 16 * inc_y;
a_ptr += 16;
}
if ( m2 & 8 )
{
sgemv_kernel_8(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(8,ybuffer,y_ptr,inc_y);
y_ptr += 8 * inc_y;
a_ptr += 8;
}
if ( m2 & 4 )
{
sgemv_kernel_4(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(4,ybuffer,y_ptr,inc_y);
y_ptr += 4 * inc_y;
a_ptr += 4;
}
if ( m2 & 2 )
{
sgemv_kernel_2(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(2,ybuffer,y_ptr,inc_y);
y_ptr += 2 * inc_y;
a_ptr += 2;
}
if ( m2 & 1 )
{
sgemv_kernel_1(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(1,ybuffer,y_ptr,inc_y);
}


}
return(0);
}



+ 0
- 451
kernel/x86_64/sgemv_n_microk_bulldozer.c View File

@@ -1,451 +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.
*****************************************************************************/

static void sgemv_kernel_64( long n, float alpha, float *a, long lda, float *x, float *y)
{


float *pre = a + lda*3;

__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%ymm1\n\t" // alpha -> ymm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"movq %6, %%r8\n\t" // address for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"prefetcht0 64(%%r8)\n\t" // Prefetch

"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // set to zero
"vxorps %%ymm9 , %%ymm9 , %%ymm9 \n\t" // set to zero
"vxorps %%ymm10, %%ymm10, %%ymm10\n\t" // set to zero
"vxorps %%ymm11, %%ymm11, %%ymm11\n\t" // set to zero
"vxorps %%ymm12, %%ymm12, %%ymm12\n\t" // set to zero
"vxorps %%ymm13, %%ymm13, %%ymm13\n\t" // set to zero
"vxorps %%ymm14, %%ymm14, %%ymm14\n\t" // set to zero
"vxorps %%ymm15, %%ymm15, %%ymm15\n\t" // set to zero
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%ymm0 \n\t" // load values of c
"nop \n\t"
"leaq (%%r8 , %%rcx, 4), %%r8 \n\t" // add lda to pointer for prefetch

"prefetcht0 (%%r8)\n\t" // Prefetch
"vfmaddps %%ymm8 , 0*4(%%rsi), %%ymm0, %%ymm8 \n\t" // multiply a and c and add to temp
"prefetcht0 64(%%r8)\n\t" // Prefetch
"vfmaddps %%ymm9 , 8*4(%%rsi), %%ymm0, %%ymm9 \n\t" // multiply a and c and add to temp
"prefetcht0 128(%%r8)\n\t" // Prefetch
"vfmaddps %%ymm10, 16*4(%%rsi), %%ymm0, %%ymm10\n\t" // multiply a and c and add to temp
"vfmaddps %%ymm11, 24*4(%%rsi), %%ymm0, %%ymm11\n\t" // multiply a and c and add to temp
"prefetcht0 192(%%r8)\n\t" // Prefetch
"vfmaddps %%ymm12, 32*4(%%rsi), %%ymm0, %%ymm12\n\t" // multiply a and c and add to temp
"vfmaddps %%ymm13, 40*4(%%rsi), %%ymm0, %%ymm13\n\t" // multiply a and c and add to temp
"vfmaddps %%ymm14, 48*4(%%rsi), %%ymm0, %%ymm14\n\t" // multiply a and c and add to temp
"vfmaddps %%ymm15, 56*4(%%rsi), %%ymm0, %%ymm15\n\t" // multiply a and c and add to temp

"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a

"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vmulps %%ymm8 , %%ymm1, %%ymm8 \n\t" // scale by alpha
"vmulps %%ymm9 , %%ymm1, %%ymm9 \n\t" // scale by alpha
"vmulps %%ymm10, %%ymm1, %%ymm10\n\t" // scale by alpha
"vmulps %%ymm11, %%ymm1, %%ymm11\n\t" // scale by alpha
"vmulps %%ymm12, %%ymm1, %%ymm12\n\t" // scale by alpha
"vmulps %%ymm13, %%ymm1, %%ymm13\n\t" // scale by alpha
"vmulps %%ymm14, %%ymm1, %%ymm14\n\t" // scale by alpha
"vmulps %%ymm15, %%ymm1, %%ymm15\n\t" // scale by alpha

"vmovups %%ymm8 , (%%rdx) \n\t" // store temp -> y
"vmovups %%ymm9 , 8*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm10, 16*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm11, 24*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm12, 32*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm13, 40*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm14, 48*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm15, 56*4(%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y), // 5
"m" (pre) // 6
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);

}



static void sgemv_kernel_32( long n, float alpha, float *a, long lda, float *x, float *y)
{


float *pre = a + lda*3;

__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"movq %6, %%r8\n\t" // address for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"prefetcht0 64(%%r8)\n\t" // Prefetch

"vxorps %%xmm8 , %%xmm8 , %%xmm8 \n\t" // set to zero
"vxorps %%xmm9 , %%xmm9 , %%xmm9 \n\t" // set to zero
"vxorps %%xmm10, %%xmm10, %%xmm10\n\t" // set to zero
"vxorps %%xmm11, %%xmm11, %%xmm11\n\t" // set to zero
"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero
"vxorps %%xmm13, %%xmm13, %%xmm13\n\t" // set to zero
"vxorps %%xmm14, %%xmm14, %%xmm14\n\t" // set to zero
"vxorps %%xmm15, %%xmm15, %%xmm15\n\t" // set to zero
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%xmm0 \n\t" // load values of c
"nop \n\t"
"leaq (%%r8 , %%rcx, 4), %%r8 \n\t" // add lda to pointer for prefetch

"prefetcht0 (%%r8)\n\t" // Prefetch
"vfmaddps %%xmm8 , 0*4(%%rsi), %%xmm0, %%xmm8 \n\t" // multiply a and c and add to temp
"prefetcht0 64(%%r8)\n\t" // Prefetch
"vfmaddps %%xmm9 , 4*4(%%rsi), %%xmm0, %%xmm9 \n\t" // multiply a and c and add to temp
"vfmaddps %%xmm10, 8*4(%%rsi), %%xmm0, %%xmm10\n\t" // multiply a and c and add to temp
"vfmaddps %%xmm11, 12*4(%%rsi), %%xmm0, %%xmm11\n\t" // multiply a and c and add to temp
"vfmaddps %%xmm12, 16*4(%%rsi), %%xmm0, %%xmm12\n\t" // multiply a and c and add to temp
"vfmaddps %%xmm13, 20*4(%%rsi), %%xmm0, %%xmm13\n\t" // multiply a and c and add to temp
"vfmaddps %%xmm14, 24*4(%%rsi), %%xmm0, %%xmm14\n\t" // multiply a and c and add to temp
"vfmaddps %%xmm15, 28*4(%%rsi), %%xmm0, %%xmm15\n\t" // multiply a and c and add to temp

"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a

"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vmulps %%xmm8 , %%xmm1, %%xmm8 \n\t" // scale by alpha
"vmulps %%xmm9 , %%xmm1, %%xmm9 \n\t" // scale by alpha
"vmulps %%xmm10, %%xmm1, %%xmm10\n\t" // scale by alpha
"vmulps %%xmm11, %%xmm1, %%xmm11\n\t" // scale by alpha
"vmulps %%xmm12, %%xmm1, %%xmm12\n\t" // scale by alpha
"vmulps %%xmm13, %%xmm1, %%xmm13\n\t" // scale by alpha
"vmulps %%xmm14, %%xmm1, %%xmm14\n\t" // scale by alpha
"vmulps %%xmm15, %%xmm1, %%xmm15\n\t" // scale by alpha

"vmovups %%xmm8 , (%%rdx) \n\t" // store temp -> y
"vmovups %%xmm9 , 4*4(%%rdx) \n\t" // store temp -> y
"vmovups %%xmm10, 8*4(%%rdx) \n\t" // store temp -> y
"vmovups %%xmm11, 12*4(%%rdx) \n\t" // store temp -> y
"vmovups %%xmm12, 16*4(%%rdx) \n\t" // store temp -> y
"vmovups %%xmm13, 20*4(%%rdx) \n\t" // store temp -> y
"vmovups %%xmm14, 24*4(%%rdx) \n\t" // store temp -> y
"vmovups %%xmm15, 28*4(%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y), // 5
"m" (pre) // 6
);

}

static void sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
{

float *pre = a + lda*3;

__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%ymm1\n\t" // alpha -> ymm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"movq %6, %%r8\n\t" // address for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch

"vxorps %%ymm12, %%ymm12, %%ymm12\n\t" // set to zero
"vxorps %%ymm13, %%ymm13, %%ymm13\n\t" // set to zero

".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%ymm0 \n\t" // load values of c
"addq $4 , %%rdi \n\t" // increment pointer of c

"leaq (%%r8 , %%rcx, 4), %%r8 \n\t" // add lda to pointer for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch

"vfmaddps %%ymm12, 0*4(%%rsi), %%ymm0, %%ymm12\n\t" // multiply a and c and add to temp
"vfmaddps %%ymm13, 8*4(%%rsi), %%ymm0, %%ymm13\n\t" // multiply a and c and add to temp

"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a

"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vmulps %%ymm12, %%ymm1, %%ymm12\n\t" // scale by alpha
"vmulps %%ymm13, %%ymm1, %%ymm13\n\t" // scale by alpha

"vmovups %%ymm12, (%%rdx) \n\t" // store temp -> y
"vmovups %%ymm13, 8*4(%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y), // 5
"m" (pre) // 6
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);

}


static void sgemv_kernel_8( long n, float alpha, float *a, long lda, float *x, float *y)
{


__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%ymm1\n\t" // alpha -> ymm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx

"vxorps %%ymm12, %%ymm12, %%ymm12\n\t" // set to zero

".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%ymm0 \n\t" // load values of c
"addq $4 , %%rdi \n\t" // increment pointer of c

"vfmaddps %%ymm12, 0*4(%%rsi), %%ymm0, %%ymm12\n\t" // multiply a and c and add to temp

"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a

"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vmulps %%ymm12, %%ymm1, %%ymm12\n\t" // scale by alpha

"vmovups %%ymm12, (%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);

}


static void sgemv_kernel_4( long n, float alpha, float *a, long lda, float *x, float *y)
{


__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx

"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero

".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%xmm0 \n\t" // load values of c
"addq $4 , %%rdi \n\t" // increment pointer of c

"vfmaddps %%xmm12, 0*4(%%rsi), %%xmm0, %%xmm12\n\t" // multiply a and c and add to temp

"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a

"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vmulps %%xmm12, %%xmm1, %%xmm12\n\t" // scale by alpha

"vmovups %%xmm12, (%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);

}

static void sgemv_kernel_2( long n, float alpha, float *a, long lda, float *x, float *y)
{


__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vmovss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx

"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero
"vxorps %%xmm13, %%xmm13, %%xmm13\n\t" // set to zero

".L01LOOP%=: \n\t"
"vmovss (%%rdi), %%xmm0 \n\t" // load values of c
"addq $4 , %%rdi \n\t" // increment pointer of c

"vfmaddss %%xmm12, 0*4(%%rsi), %%xmm0, %%xmm12\n\t" // multiply a and c and add to temp
"vfmaddss %%xmm13, 1*4(%%rsi), %%xmm0, %%xmm13\n\t" // multiply a and c and add to temp

"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a

"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vmulss %%xmm12, %%xmm1, %%xmm12\n\t" // scale by alpha
"vmulss %%xmm13, %%xmm1, %%xmm13\n\t" // scale by alpha

"vmovss %%xmm12, (%%rdx) \n\t" // store temp -> y
"vmovss %%xmm13, 4(%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);

}



static void sgemv_kernel_1( long n, float alpha, float *a, long lda, float *x, float *y)
{


__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vmovss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx

"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero

".L01LOOP%=: \n\t"
"vmovss (%%rdi), %%xmm0 \n\t" // load values of c
"addq $4 , %%rdi \n\t" // increment pointer of c

"vfmaddss %%xmm12, 0*4(%%rsi), %%xmm0, %%xmm12\n\t" // multiply a and c and add to temp

"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a

"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vmulss %%xmm12, %%xmm1, %%xmm12\n\t" // scale by alpha

"vmovss %%xmm12, (%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);

}



+ 0
- 461
kernel/x86_64/sgemv_n_microk_haswell.c View File

@@ -1,461 +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.
*****************************************************************************/

static void sgemv_kernel_64( long n, float alpha, float *a, long lda, float *x, float *y)
{


float *pre = a + lda*2;

__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%ymm1\n\t" // alpha -> ymm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"movq %6, %%r8\n\t" // address for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"prefetcht0 64(%%r8)\n\t" // Prefetch

"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // set to zero
"vxorps %%ymm9 , %%ymm9 , %%ymm9 \n\t" // set to zero
"vxorps %%ymm10, %%ymm10, %%ymm10\n\t" // set to zero
"vxorps %%ymm11, %%ymm11, %%ymm11\n\t" // set to zero
"vxorps %%ymm12, %%ymm12, %%ymm12\n\t" // set to zero
"vxorps %%ymm13, %%ymm13, %%ymm13\n\t" // set to zero
"vxorps %%ymm14, %%ymm14, %%ymm14\n\t" // set to zero
"vxorps %%ymm15, %%ymm15, %%ymm15\n\t" // set to zero
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%ymm0 \n\t" // load values of c
"leaq (%%r8 , %%rcx, 4), %%r8 \n\t" // add lda to pointer for prefetch

"prefetcht0 (%%r8)\n\t" // Prefetch
"vfmadd231ps 0*4(%%rsi), %%ymm0, %%ymm8 \n\t" // multiply a and c and add to temp
"vfmadd231ps 8*4(%%rsi), %%ymm0, %%ymm9 \n\t" // multiply a and c and add to temp
"prefetcht0 64(%%r8)\n\t" // Prefetch
"vfmadd231ps 16*4(%%rsi), %%ymm0, %%ymm10\n\t" // multiply a and c and add to temp
"vfmadd231ps 24*4(%%rsi), %%ymm0, %%ymm11\n\t" // multiply a and c and add to temp
"prefetcht0 128(%%r8)\n\t" // Prefetch
"vfmadd231ps 32*4(%%rsi), %%ymm0, %%ymm12\n\t" // multiply a and c and add to temp
"vfmadd231ps 40*4(%%rsi), %%ymm0, %%ymm13\n\t" // multiply a and c and add to temp
"prefetcht0 192(%%r8)\n\t" // Prefetch
"vfmadd231ps 48*4(%%rsi), %%ymm0, %%ymm14\n\t" // multiply a and c and add to temp
"vfmadd231ps 56*4(%%rsi), %%ymm0, %%ymm15\n\t" // multiply a and c and add to temp

"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a

"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vmulps %%ymm8 , %%ymm1, %%ymm8 \n\t" // scale by alpha
"vmulps %%ymm9 , %%ymm1, %%ymm9 \n\t" // scale by alpha
"vmulps %%ymm10, %%ymm1, %%ymm10\n\t" // scale by alpha
"vmulps %%ymm11, %%ymm1, %%ymm11\n\t" // scale by alpha
"vmulps %%ymm12, %%ymm1, %%ymm12\n\t" // scale by alpha
"vmulps %%ymm13, %%ymm1, %%ymm13\n\t" // scale by alpha
"vmulps %%ymm14, %%ymm1, %%ymm14\n\t" // scale by alpha
"vmulps %%ymm15, %%ymm1, %%ymm15\n\t" // scale by alpha

"vmovups %%ymm8 , (%%rdx) \n\t" // store temp -> y
"vmovups %%ymm9 , 8*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm10, 16*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm11, 24*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm12, 32*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm13, 40*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm14, 48*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm15, 56*4(%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y), // 5
"m" (pre) // 6
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);

}



static void sgemv_kernel_32( long n, float alpha, float *a, long lda, float *x, float *y)
{


float *pre = a + lda*3;

__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%ymm1\n\t" // alpha -> ymm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"movq %6, %%r8\n\t" // address for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"prefetcht0 64(%%r8)\n\t" // Prefetch

"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // set to zero
"vxorps %%ymm9 , %%ymm9 , %%ymm9 \n\t" // set to zero
"vxorps %%ymm10, %%ymm10, %%ymm10\n\t" // set to zero
"vxorps %%ymm11, %%ymm11, %%ymm11\n\t" // set to zero
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%ymm0 \n\t" // load values of c
"nop \n\t"
"leaq (%%r8 , %%rcx, 4), %%r8 \n\t" // add lda to pointer for prefetch

"prefetcht0 (%%r8)\n\t" // Prefetch
"prefetcht0 64(%%r8)\n\t" // Prefetch

"vmulps 0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
"vmulps 8*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
"vmulps 16*4(%%rsi), %%ymm0, %%ymm6 \n\t" // multiply a and c and add to temp
"vmulps 24*4(%%rsi), %%ymm0, %%ymm7 \n\t" // multiply a and c and add to temp

"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
"vaddps %%ymm9 , %%ymm5, %%ymm9 \n\t" // multiply a and c and add to temp
"vaddps %%ymm10, %%ymm6, %%ymm10\n\t" // multiply a and c and add to temp
"vaddps %%ymm11, %%ymm7, %%ymm11\n\t" // multiply a and c and add to temp



"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a

"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vmulps %%ymm8 , %%ymm1, %%ymm8 \n\t" // scale by alpha
"vmulps %%ymm9 , %%ymm1, %%ymm9 \n\t" // scale by alpha
"vmulps %%ymm10, %%ymm1, %%ymm10\n\t" // scale by alpha
"vmulps %%ymm11, %%ymm1, %%ymm11\n\t" // scale by alpha

"vmovups %%ymm8 , (%%rdx) \n\t" // store temp -> y
"vmovups %%ymm9 , 8*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm10, 16*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm11, 24*4(%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y), // 5
"m" (pre) // 6
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"memory"
);



}

static void sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
{

float *pre = a + lda*3;
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%ymm1\n\t" // alpha -> ymm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"movq %6, %%r8\n\t" // address for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"prefetcht0 64(%%r8)\n\t" // Prefetch

"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // set to zero
"vxorps %%ymm9 , %%ymm9 , %%ymm9 \n\t" // set to zero
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%ymm0 \n\t" // load values of c
"nop \n\t"
"leaq (%%r8 , %%rcx, 4), %%r8 \n\t" // add lda to pointer for prefetch

"prefetcht0 (%%r8)\n\t" // Prefetch

"vmulps 0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
"vmulps 8*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp

"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
"vaddps %%ymm9 , %%ymm5, %%ymm9 \n\t" // multiply a and c and add to temp

"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a

"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vmulps %%ymm8 , %%ymm1, %%ymm8 \n\t" // scale by alpha
"vmulps %%ymm9 , %%ymm1, %%ymm9 \n\t" // scale by alpha

"vmovups %%ymm8 , (%%rdx) \n\t" // store temp -> y
"vmovups %%ymm9 , 8*4(%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y), // 5
"m" (pre) // 6
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"memory"
);


}


static void sgemv_kernel_8( long n, float alpha, float *a, long lda, float *x, float *y)
{
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%ymm1\n\t" // alpha -> ymm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx

"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // set to zero
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%ymm0 \n\t" // load values of c

"vmulps 0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp

"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a

"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vmulps %%ymm8 , %%ymm1, %%ymm8 \n\t" // scale by alpha
"vmovups %%ymm8 , (%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"memory"
);


}


static void sgemv_kernel_4( long n, float alpha, float *a, long lda, float *x, float *y)
{


__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx

"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero

".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%xmm0 \n\t" // load values of c

"vmulps 0*4(%%rsi), %%xmm0, %%xmm4 \n\t" // multiply a and c and add to temp
"vaddps %%xmm12, %%xmm4, %%xmm12 \n\t" // multiply a and c and add to temp

"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a

"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vmulps %%xmm12, %%xmm1, %%xmm12\n\t" // scale by alpha

"vmovups %%xmm12, (%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);

}

static void sgemv_kernel_2( long n, float alpha, float *a, long lda, float *x, float *y)
{


__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vmovss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx

"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero
"vxorps %%xmm13, %%xmm13, %%xmm13\n\t" // set to zero

".L01LOOP%=: \n\t"
"vmovss (%%rdi), %%xmm0 \n\t" // load values of c

"vmulps 0*4(%%rsi), %%xmm0, %%xmm4 \n\t" // multiply a and c and add to temp
"vmulps 1*4(%%rsi), %%xmm0, %%xmm5 \n\t" // multiply a and c and add to temp

"vaddps %%xmm12, %%xmm4, %%xmm12 \n\t" // multiply a and c and add to temp
"vaddps %%xmm13, %%xmm5, %%xmm13 \n\t" // multiply a and c and add to temp

"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a

"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vmulss %%xmm12, %%xmm1, %%xmm12\n\t" // scale by alpha
"vmulss %%xmm13, %%xmm1, %%xmm13\n\t" // scale by alpha

"vmovss %%xmm12, (%%rdx) \n\t" // store temp -> y
"vmovss %%xmm13, 4(%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);

}



static void sgemv_kernel_1( long n, float alpha, float *a, long lda, float *x, float *y)
{


__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vmovss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx

"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero

".L01LOOP%=: \n\t"
"vmovss (%%rdi), %%xmm0 \n\t" // load values of c
"addq $4 , %%rdi \n\t" // increment pointer of c

"vmulss 0*4(%%rsi), %%xmm0, %%xmm4 \n\t" // multiply a and c and add to temp
"vaddss %%xmm12, %%xmm4, %%xmm12 \n\t" // multiply a and c and add to temp

"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a

"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vmulss %%xmm12, %%xmm1, %%xmm12\n\t" // scale by alpha

"vmovss %%xmm12, (%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);

}



+ 0
- 473
kernel/x86_64/sgemv_n_microk_sandy.c View File

@@ -1,473 +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.
*****************************************************************************/

static void sgemv_kernel_64( long n, float alpha, float *a, long lda, float *x, float *y)
{


float *pre = a + lda*2;

__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%ymm1\n\t" // alpha -> ymm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"movq %6, %%r8\n\t" // address for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"prefetcht0 64(%%r8)\n\t" // Prefetch

"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // set to zero
"vxorps %%ymm9 , %%ymm9 , %%ymm9 \n\t" // set to zero
"vxorps %%ymm10, %%ymm10, %%ymm10\n\t" // set to zero
"vxorps %%ymm11, %%ymm11, %%ymm11\n\t" // set to zero
"vxorps %%ymm12, %%ymm12, %%ymm12\n\t" // set to zero
"vxorps %%ymm13, %%ymm13, %%ymm13\n\t" // set to zero
"vxorps %%ymm14, %%ymm14, %%ymm14\n\t" // set to zero
"vxorps %%ymm15, %%ymm15, %%ymm15\n\t" // set to zero
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%ymm0 \n\t" // load values of c
"nop \n\t"
"leaq (%%r8 , %%rcx, 4), %%r8 \n\t" // add lda to pointer for prefetch

"prefetcht0 (%%r8)\n\t" // Prefetch
"vmulps 0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
"vmulps 8*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
"prefetcht0 64(%%r8)\n\t" // Prefetch
"vmulps 16*4(%%rsi), %%ymm0, %%ymm6 \n\t" // multiply a and c and add to temp
"vmulps 24*4(%%rsi), %%ymm0, %%ymm7 \n\t" // multiply a and c and add to temp

"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
"vaddps %%ymm9 , %%ymm5, %%ymm9 \n\t" // multiply a and c and add to temp
"prefetcht0 128(%%r8)\n\t" // Prefetch
"vaddps %%ymm10, %%ymm6, %%ymm10\n\t" // multiply a and c and add to temp
"vaddps %%ymm11, %%ymm7, %%ymm11\n\t" // multiply a and c and add to temp

"prefetcht0 192(%%r8)\n\t" // Prefetch
"vmulps 32*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
"vmulps 40*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
"vmulps 48*4(%%rsi), %%ymm0, %%ymm6 \n\t" // multiply a and c and add to temp
"vmulps 56*4(%%rsi), %%ymm0, %%ymm7 \n\t" // multiply a and c and add to temp

"vaddps %%ymm12, %%ymm4, %%ymm12\n\t" // multiply a and c and add to temp
"vaddps %%ymm13, %%ymm5, %%ymm13\n\t" // multiply a and c and add to temp
"vaddps %%ymm14, %%ymm6, %%ymm14\n\t" // multiply a and c and add to temp
"vaddps %%ymm15, %%ymm7, %%ymm15\n\t" // multiply a and c and add to temp

"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a

"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vmulps %%ymm8 , %%ymm1, %%ymm8 \n\t" // scale by alpha
"vmulps %%ymm9 , %%ymm1, %%ymm9 \n\t" // scale by alpha
"vmulps %%ymm10, %%ymm1, %%ymm10\n\t" // scale by alpha
"vmulps %%ymm11, %%ymm1, %%ymm11\n\t" // scale by alpha
"vmulps %%ymm12, %%ymm1, %%ymm12\n\t" // scale by alpha
"vmulps %%ymm13, %%ymm1, %%ymm13\n\t" // scale by alpha
"vmulps %%ymm14, %%ymm1, %%ymm14\n\t" // scale by alpha
"vmulps %%ymm15, %%ymm1, %%ymm15\n\t" // scale by alpha

"vmovups %%ymm8 , (%%rdx) \n\t" // store temp -> y
"vmovups %%ymm9 , 8*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm10, 16*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm11, 24*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm12, 32*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm13, 40*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm14, 48*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm15, 56*4(%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y), // 5
"m" (pre) // 6
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);

}



static void sgemv_kernel_32( long n, float alpha, float *a, long lda, float *x, float *y)
{


float *pre = a + lda*3;

__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%ymm1\n\t" // alpha -> ymm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"movq %6, %%r8\n\t" // address for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"prefetcht0 64(%%r8)\n\t" // Prefetch

"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // set to zero
"vxorps %%ymm9 , %%ymm9 , %%ymm9 \n\t" // set to zero
"vxorps %%ymm10, %%ymm10, %%ymm10\n\t" // set to zero
"vxorps %%ymm11, %%ymm11, %%ymm11\n\t" // set to zero
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%ymm0 \n\t" // load values of c
"nop \n\t"
"leaq (%%r8 , %%rcx, 4), %%r8 \n\t" // add lda to pointer for prefetch

"prefetcht0 (%%r8)\n\t" // Prefetch
"prefetcht0 64(%%r8)\n\t" // Prefetch

"vmulps 0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
"vmulps 8*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
"vmulps 16*4(%%rsi), %%ymm0, %%ymm6 \n\t" // multiply a and c and add to temp
"vmulps 24*4(%%rsi), %%ymm0, %%ymm7 \n\t" // multiply a and c and add to temp

"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
"vaddps %%ymm9 , %%ymm5, %%ymm9 \n\t" // multiply a and c and add to temp
"vaddps %%ymm10, %%ymm6, %%ymm10\n\t" // multiply a and c and add to temp
"vaddps %%ymm11, %%ymm7, %%ymm11\n\t" // multiply a and c and add to temp



"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a

"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vmulps %%ymm8 , %%ymm1, %%ymm8 \n\t" // scale by alpha
"vmulps %%ymm9 , %%ymm1, %%ymm9 \n\t" // scale by alpha
"vmulps %%ymm10, %%ymm1, %%ymm10\n\t" // scale by alpha
"vmulps %%ymm11, %%ymm1, %%ymm11\n\t" // scale by alpha

"vmovups %%ymm8 , (%%rdx) \n\t" // store temp -> y
"vmovups %%ymm9 , 8*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm10, 16*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm11, 24*4(%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y), // 5
"m" (pre) // 6
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"memory"
);



}

static void sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
{

float *pre = a + lda*3;
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%ymm1\n\t" // alpha -> ymm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"movq %6, %%r8\n\t" // address for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"prefetcht0 64(%%r8)\n\t" // Prefetch

"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // set to zero
"vxorps %%ymm9 , %%ymm9 , %%ymm9 \n\t" // set to zero
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%ymm0 \n\t" // load values of c
"nop \n\t"
"leaq (%%r8 , %%rcx, 4), %%r8 \n\t" // add lda to pointer for prefetch

"prefetcht0 (%%r8)\n\t" // Prefetch

"vmulps 0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
"vmulps 8*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp

"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
"vaddps %%ymm9 , %%ymm5, %%ymm9 \n\t" // multiply a and c and add to temp

"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a

"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vmulps %%ymm8 , %%ymm1, %%ymm8 \n\t" // scale by alpha
"vmulps %%ymm9 , %%ymm1, %%ymm9 \n\t" // scale by alpha

"vmovups %%ymm8 , (%%rdx) \n\t" // store temp -> y
"vmovups %%ymm9 , 8*4(%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y), // 5
"m" (pre) // 6
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"memory"
);


}


static void sgemv_kernel_8( long n, float alpha, float *a, long lda, float *x, float *y)
{
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%ymm1\n\t" // alpha -> ymm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx

"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // set to zero
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%ymm0 \n\t" // load values of c

"vmulps 0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp

"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a

"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vmulps %%ymm8 , %%ymm1, %%ymm8 \n\t" // scale by alpha
"vmovups %%ymm8 , (%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"memory"
);


}


static void sgemv_kernel_4( long n, float alpha, float *a, long lda, float *x, float *y)
{


__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx

"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero

".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%xmm0 \n\t" // load values of c

"vmulps 0*4(%%rsi), %%xmm0, %%xmm4 \n\t" // multiply a and c and add to temp
"vaddps %%xmm12, %%xmm4, %%xmm12 \n\t" // multiply a and c and add to temp

"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a

"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vmulps %%xmm12, %%xmm1, %%xmm12\n\t" // scale by alpha

"vmovups %%xmm12, (%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);

}

static void sgemv_kernel_2( long n, float alpha, float *a, long lda, float *x, float *y)
{


__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vmovss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx

"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero
"vxorps %%xmm13, %%xmm13, %%xmm13\n\t" // set to zero

".L01LOOP%=: \n\t"
"vmovss (%%rdi), %%xmm0 \n\t" // load values of c

"vmulps 0*4(%%rsi), %%xmm0, %%xmm4 \n\t" // multiply a and c and add to temp
"vmulps 1*4(%%rsi), %%xmm0, %%xmm5 \n\t" // multiply a and c and add to temp

"vaddps %%xmm12, %%xmm4, %%xmm12 \n\t" // multiply a and c and add to temp
"vaddps %%xmm13, %%xmm5, %%xmm13 \n\t" // multiply a and c and add to temp

"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a

"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vmulss %%xmm12, %%xmm1, %%xmm12\n\t" // scale by alpha
"vmulss %%xmm13, %%xmm1, %%xmm13\n\t" // scale by alpha

"vmovss %%xmm12, (%%rdx) \n\t" // store temp -> y
"vmovss %%xmm13, 4(%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);

}



static void sgemv_kernel_1( long n, float alpha, float *a, long lda, float *x, float *y)
{


__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vmovss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx

"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero

".L01LOOP%=: \n\t"
"vmovss (%%rdi), %%xmm0 \n\t" // load values of c
"addq $4 , %%rdi \n\t" // increment pointer of c

"vmulss 0*4(%%rsi), %%xmm0, %%xmm4 \n\t" // multiply a and c and add to temp
"vaddss %%xmm12, %%xmm4, %%xmm12 \n\t" // multiply a and c and add to temp

"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a

"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vmulss %%xmm12, %%xmm1, %%xmm12\n\t" // scale by alpha

"vmovss %%xmm12, (%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);

}



+ 0
- 232
kernel/x86_64/sgemv_t_avx.c View File

@@ -1,232 +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 "sgemv_t_microk_bulldozer.c"
#elif defined(HASWELL)
#include "sgemv_t_microk_haswell.c"
#else
#include "sgemv_t_microk_sandy.c"
#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++;
src += inc_src;
}
}

static void sgemv_kernel_1( BLASLONG n, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, FLOAT *y)
{

FLOAT register temp0 = 0.0;
BLASLONG i;
for ( i=0; i<n ; i++)
{
temp0 += a[i] * x[i];
}
temp0 *= alpha ;
*y += temp0;
}




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 *a_ptrl;
BLASLONG m1;
BLASLONG register m2;
FLOAT *xbuffer;
xbuffer = buffer;
BLASLONG register Mblock;

m1 = m / 1024 ;
m2 = m % 1024 ;

x_ptr = x;
a_ptr = a;

for (j=0; j<m1; j++)
{

if ( inc_x == 1 )
xbuffer = x_ptr;
else
copy_x(1024,x_ptr,xbuffer,inc_x);

y_ptr = y;
a_ptrl = a_ptr;

for(i = 0; i<n; i++ )
{
sgemv_kernel_16(1024,alpha,a_ptrl,lda,xbuffer,y_ptr);
y_ptr += inc_y;
a_ptrl += lda;
}
a_ptr += 1024;
x_ptr += 1024 * inc_x;
}

if ( m2 == 0 ) return(0);

Mblock = 512;
while ( Mblock >= 16 )
{
if ( m2 & Mblock)
{

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

y_ptr = y;
a_ptrl = a_ptr;

for(i = 0; i<n; i++ )
{
sgemv_kernel_16(Mblock,alpha,a_ptrl,lda,xbuffer,y_ptr);
y_ptr += inc_y;
a_ptrl += lda;
}
a_ptr += Mblock;
x_ptr += Mblock * inc_x;


}
Mblock /= 2;

}

if ( m2 & Mblock)
{

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

y_ptr = y;
a_ptrl = a_ptr;

for(i = 0; i<n; i++ )
{
sgemv_kernel_1(Mblock,alpha,a_ptrl,lda,xbuffer,y_ptr);
y_ptr += inc_y;
a_ptrl += lda;
}
a_ptr += Mblock;
x_ptr += Mblock * inc_x;


}
Mblock /= 2;


if ( m2 & Mblock)
{

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

y_ptr = y;
a_ptrl = a_ptr;

for(i = 0; i<n; i++ )
{
sgemv_kernel_1(Mblock,alpha,a_ptrl,lda,xbuffer,y_ptr);
y_ptr += inc_y;
a_ptrl += lda;
}
a_ptr += Mblock;
x_ptr += Mblock * inc_x;


}
Mblock /= 2;

if ( m2 & Mblock)
{

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

y_ptr = y;
a_ptrl = a_ptr;

for(i = 0; i<n; i++ )
{
sgemv_kernel_1(Mblock,alpha,a_ptrl,lda,xbuffer,y_ptr);
y_ptr += inc_y;
a_ptrl += lda;
}
a_ptr += Mblock;
x_ptr += Mblock * inc_x;


}
Mblock /= 2;

if ( m2 & Mblock)
{

xbuffer = x_ptr;

y_ptr = y;
a_ptrl = a_ptr;

for(i = 0; i<n; i++ )
{
sgemv_kernel_1(Mblock,alpha,a_ptrl,lda,xbuffer,y_ptr);
y_ptr += inc_y;
a_ptrl += lda;
}


}

return(0);
}



+ 0
- 99
kernel/x86_64/sgemv_t_microk_bulldozer.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.
*****************************************************************************/

static void sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
{

//n = n / 16;

__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vmovss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx

"leaq (, %%rcx,4), %%rcx \n\t" // scale lda by size of float
"leaq (%%rsi,%%rcx,1), %%r8 \n\t" // pointer to next line

"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero
"vxorps %%xmm13, %%xmm13, %%xmm13\n\t" // set to zero
"vxorps %%xmm14, %%xmm14, %%xmm14\n\t" // set to zero
"vxorps %%xmm15, %%xmm15, %%xmm15\n\t" // set to zero

"sarq $4, %%rax \n\t" // n = n / 16

".align 16 \n\t"
".L01LOOP%=: \n\t"
// "prefetcht0 512(%%rsi) \n\t"
"prefetcht0 (%%r8) \n\t" //prefetch next line of a
"vmovups (%%rsi), %%xmm4 \n\t"
"vmovups 4*4(%%rsi), %%xmm5 \n\t"
"vmovups 8*4(%%rsi), %%xmm6 \n\t"
"vmovups 12*4(%%rsi), %%xmm7 \n\t"

"vfmaddps %%xmm12, 0*4(%%rdi), %%xmm4, %%xmm12\n\t" // multiply a and c and add to temp
"vfmaddps %%xmm13, 4*4(%%rdi), %%xmm5, %%xmm13\n\t" // multiply a and c and add to temp
"vfmaddps %%xmm14, 8*4(%%rdi), %%xmm6, %%xmm14\n\t" // multiply a and c and add to temp
"vfmaddps %%xmm15, 12*4(%%rdi), %%xmm7, %%xmm15\n\t" // multiply a and c and add to temp

"addq $16*4 , %%r8 \n\t" // increment prefetch pointer
"addq $16*4 , %%rsi \n\t" // increment pointer of a
"addq $16*4 , %%rdi \n\t" // increment pointer of c
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vaddps %%xmm12, %%xmm14, %%xmm12\n\t"
"vaddps %%xmm13, %%xmm15, %%xmm13\n\t"
"vaddps %%xmm12, %%xmm13, %%xmm12\n\t"
"vhaddps %%xmm12, %%xmm12, %%xmm12\n\t"
"vhaddps %%xmm12, %%xmm12, %%xmm12\n\t"

"vfmaddss (%%rdx), %%xmm12, %%xmm1, %%xmm12\n\t"
"vmovss %%xmm12, (%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);

}




+ 0
- 100
kernel/x86_64/sgemv_t_microk_haswell.c View File

@@ -1,100 +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.
*****************************************************************************/

static void sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
{

//n = n / 16;

__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vmovss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx

"leaq (, %%rcx,4), %%rcx \n\t" // scale lda by size of float
"leaq (%%rsi,%%rcx,1), %%r8 \n\t" // pointer to next line

"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero
"vxorps %%xmm13, %%xmm13, %%xmm13\n\t" // set to zero
"vxorps %%xmm14, %%xmm14, %%xmm14\n\t" // set to zero
"vxorps %%xmm15, %%xmm15, %%xmm15\n\t" // set to zero

"sarq $4, %%rax \n\t" // n = n / 16

".align 16 \n\t"
".L01LOOP%=: \n\t"
// "prefetcht0 512(%%rsi) \n\t"
"prefetcht0 (%%r8) \n\t" //prefetch next line of a
"vmovups (%%rsi), %%xmm4 \n\t"
"vmovups 4*4(%%rsi), %%xmm5 \n\t"
"vmovups 8*4(%%rsi), %%xmm6 \n\t"
"vmovups 12*4(%%rsi), %%xmm7 \n\t"

"vfmadd231ps 0*4(%%rdi), %%xmm4, %%xmm12\n\t" // multiply a and c and add to temp
"vfmadd231ps 4*4(%%rdi), %%xmm5, %%xmm13\n\t" // multiply a and c and add to temp
"vfmadd231ps 8*4(%%rdi), %%xmm6, %%xmm14\n\t" // multiply a and c and add to temp
"vfmadd231ps 12*4(%%rdi), %%xmm7, %%xmm15\n\t" // multiply a and c and add to temp

"addq $16*4 , %%r8 \n\t" // increment prefetch pointer
"addq $16*4 , %%rsi \n\t" // increment pointer of a
"addq $16*4 , %%rdi \n\t" // increment pointer of c
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vaddps %%xmm12, %%xmm14, %%xmm12\n\t"
"vaddps %%xmm13, %%xmm15, %%xmm13\n\t"
"vaddps %%xmm12, %%xmm13, %%xmm12\n\t"
"vhaddps %%xmm12, %%xmm12, %%xmm12\n\t"
"vhaddps %%xmm12, %%xmm12, %%xmm12\n\t"

"vmulss %%xmm12, %%xmm1, %%xmm12\n\t"
"vaddss (%%rdx), %%xmm12,%%xmm12\n\t"
"vmovss %%xmm12, (%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);

}




+ 0
- 106
kernel/x86_64/sgemv_t_microk_sandy.c View File

@@ -1,106 +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.
*****************************************************************************/

static void sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
{

//n = n / 16;

__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vmovss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx

"leaq (, %%rcx,4), %%rcx \n\t" // scale lda by size of float
"leaq (%%rsi,%%rcx,1), %%r8 \n\t" // pointer to next line

"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero
"vxorps %%xmm13, %%xmm13, %%xmm13\n\t" // set to zero
"vxorps %%xmm14, %%xmm14, %%xmm14\n\t" // set to zero
"vxorps %%xmm15, %%xmm15, %%xmm15\n\t" // set to zero

"sarq $4, %%rax \n\t" // n = n / 16

".align 16 \n\t"
".L01LOOP%=: \n\t"
// "prefetcht0 512(%%rsi) \n\t"
"prefetcht0 (%%r8) \n\t" //prefetch next line of a
"vmovups (%%rsi), %%xmm4 \n\t"
"vmovups 4*4(%%rsi), %%xmm5 \n\t"
"vmovups 8*4(%%rsi), %%xmm6 \n\t"
"vmovups 12*4(%%rsi), %%xmm7 \n\t"

"vmulps 0*4(%%rdi), %%xmm4, %%xmm8 \n\t" // multiply a and c and add to temp
"vmulps 4*4(%%rdi), %%xmm5, %%xmm9 \n\t" // multiply a and c and add to temp
"vmulps 8*4(%%rdi), %%xmm6, %%xmm10\n\t" // multiply a and c and add to temp
"vmulps 12*4(%%rdi), %%xmm7, %%xmm11\n\t" // multiply a and c and add to temp

"vaddps %%xmm12, %%xmm8 , %%xmm12\n\t"
"vaddps %%xmm13, %%xmm9 , %%xmm13\n\t"
"vaddps %%xmm14, %%xmm10, %%xmm14\n\t"
"vaddps %%xmm15, %%xmm11, %%xmm15\n\t"

"addq $16*4 , %%r8 \n\t" // increment prefetch pointer
"addq $16*4 , %%rsi \n\t" // increment pointer of a
"addq $16*4 , %%rdi \n\t" // increment pointer of c
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"

"vaddps %%xmm12, %%xmm14, %%xmm12\n\t"
"vaddps %%xmm13, %%xmm15, %%xmm13\n\t"
"vaddps %%xmm12, %%xmm13, %%xmm12\n\t"
"vhaddps %%xmm12, %%xmm12, %%xmm12\n\t"
"vhaddps %%xmm12, %%xmm12, %%xmm12\n\t"

"vmulss %%xmm12, %%xmm1, %%xmm12 \n\t"
"vaddss (%%rdx), %%xmm12, %%xmm12\n\t"
"vmovss %%xmm12, (%%rdx) \n\t" // store temp -> y

:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);

}




Write Preview
Loading…
Cancel
Save