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

Merge pull request #544 from wernsaar/develop 

Optimized  caxpy-kernels
tags/v0.2.15^2
wernsaar 11 years ago
parent
fb02cb0a41 13889515b3
commit
59e7a518c6
7 changed files with 580 additions and 59 deletions
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +2
    -2
      kernel/x86_64/KERNEL.HASWELL
  2. +1
    -0
      kernel/x86_64/KERNEL.SANDYBRIDGE
  3. +10
    -3
      kernel/x86_64/caxpy.c
  4. +119
    -54
      kernel/x86_64/caxpy_microk_bulldozer-2.c
  5. +132
    -0
      kernel/x86_64/caxpy_microk_haswell-2.c
  6. +116
    -0
      kernel/x86_64/caxpy_microk_sandy-2.c
  7. +200
    -0
      kernel/x86_64/caxpy_microk_steamroller-2.c

+ 2
- 2
kernel/x86_64/KERNEL.HASWELL View File

@@ -15,9 +15,9 @@ DDOTKERNEL = ddot.c
CDOTKERNEL = cdot.c
ZDOTKERNEL = zdot.c


DAXPYKERNEL = daxpy.c
SAXPYKERNEL = saxpy.c
DAXPYKERNEL = daxpy.c
CAXPYKERNEL = caxpy.c

SGEMMKERNEL = sgemm_kernel_16x4_haswell.S
SGEMMINCOPY = ../generic/gemm_ncopy_16.c


+ 1
- 0
kernel/x86_64/KERNEL.SANDYBRIDGE View File

@@ -11,6 +11,7 @@ ZDOTKERNEL = zdot.c

SAXPYKERNEL = saxpy.c
DAXPYKERNEL = daxpy.c
CAXPYKERNEL = caxpy.c

SGEMMKERNEL = sgemm_kernel_16x4_sandy.S
SGEMMINCOPY = ../generic/gemm_ncopy_16.c


+ 10
- 3
kernel/x86_64/caxpy.c View File

@@ -29,8 +29,14 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "common.h"


#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER)
#if defined(PILEDRIVER) || defined(STEAMROLLER)
#include "caxpy_microk_steamroller-2.c"
#elif defined(BULLDOZER)
#include "caxpy_microk_bulldozer-2.c"
#elif defined(HASWELL)
#include "caxpy_microk_haswell-2.c"
#elif defined(SANDYBRIDGE)
#include "caxpy_microk_sandy-2.c"
#endif


@@ -78,15 +84,16 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r, FLOAT da_i,
if ( (inc_x == 1) && (inc_y == 1) )
{

int n1 = n & -8;
int n1 = n & -32;

if ( n1 )
{
da[0] = da_r;
da[1] = da_i;
caxpy_kernel_8(n1, x, y , &da );
caxpy_kernel_8(n1, x, y , da );
ix = 2 * n1;
}

i = n1;
while(i < n)
{


+ 119
- 54
kernel/x86_64/caxpy_microk_bulldozer-2.c View File

@@ -31,89 +31,154 @@ static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *alpha) __att
static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
{

#if !defined(CONJ)
FLOAT mvec[4] = { -1.0, 1.0, -1.0, 1.0 };
#else
FLOAT mvec[4] = { 1.0, -1.0, 1.0, -1.0 };
#endif

BLASLONG register i = 0;

if ( n < 640 )
{

__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastss (%4), %%xmm0 \n\t" // real part of alpha
"vbroadcastss 4(%4), %%xmm1 \n\t" // imag part of alpha
#if !defined(CONJ)
"vmulps (%5), %%xmm1 , %%xmm1 \n\t"
#else
"vmulps (%5), %%xmm0 , %%xmm0 \n\t"
#endif

".align 16 \n\t"
"1: \n\t"

"prefetcht0 768(%2,%0,4) \n\t"
"vmovups (%2,%0,4), %%xmm5 \n\t" // 2 complex values from x
".align 2 \n\t"
"vmovups 16(%2,%0,4), %%xmm7 \n\t" // 2 complex values from x
"vmovups 32(%2,%0,4), %%xmm9 \n\t" // 2 complex values from x
"vmovups 48(%2,%0,4), %%xmm11 \n\t" // 2 complex values from x
"prefetcht0 768(%3,%0,4) \n\t"

#if !defined(CONJ)
"vfmaddps (%3,%0,4), %%xmm0 , %%xmm5, %%xmm12 \n\t"
"vpermilps $0xb1 , %%xmm5 , %%xmm4 \n\t" // exchange real and imag part
"vmulps %%xmm1, %%xmm4 , %%xmm4 \n\t"
"vmovups 64(%2,%0,4), %%xmm12 \n\t" // 2 complex values from x
"vmovups 80(%2,%0,4), %%xmm13 \n\t" // 2 complex values from x
"vmovups 96(%2,%0,4), %%xmm14 \n\t" // 2 complex values from x
"vmovups 112(%2,%0,4), %%xmm15 \n\t" // 2 complex values from x

"vfmaddps 16(%3,%0,4), %%xmm0 , %%xmm7, %%xmm13 \n\t"
"vpermilps $0xb1 , %%xmm5 , %%xmm4 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%xmm7 , %%xmm6 \n\t" // exchange real and imag part
"vmulps %%xmm1, %%xmm6 , %%xmm6 \n\t"

"vfmaddps 32(%3,%0,4), %%xmm0 , %%xmm9, %%xmm14 \n\t"
"vpermilps $0xb1 , %%xmm9 , %%xmm8 \n\t" // exchange real and imag part
"vmulps %%xmm1, %%xmm8 , %%xmm8 \n\t"

"vfmaddps 48(%3,%0,4), %%xmm0 , %%xmm11,%%xmm15 \n\t"
"vpermilps $0xb1 , %%xmm11, %%xmm10 \n\t" // exchange real and imag part
"vmulps %%xmm1, %%xmm10, %%xmm10 \n\t"

"vaddsubps %%xmm4, %%xmm12, %%xmm12 \n\t"
"vaddsubps %%xmm6, %%xmm13, %%xmm13 \n\t"
"vaddsubps %%xmm8, %%xmm14, %%xmm14 \n\t"
"vaddsubps %%xmm10,%%xmm15, %%xmm15 \n\t"

#else

"vmulps %%xmm0, %%xmm5, %%xmm4 \n\t" // a_r*x_r, a_r*x_i
"vmulps %%xmm1, %%xmm5, %%xmm5 \n\t" // a_i*x_r, a_i*x_i
"vmulps %%xmm0, %%xmm7, %%xmm6 \n\t" // a_r*x_r, a_r*x_i
"vmulps %%xmm1, %%xmm7, %%xmm7 \n\t" // a_i*x_r, a_i*x_i
"vmulps %%xmm0, %%xmm9, %%xmm8 \n\t" // a_r*x_r, a_r*x_i
"vmulps %%xmm1, %%xmm9, %%xmm9 \n\t" // a_i*x_r, a_i*x_i
"vmulps %%xmm0, %%xmm11, %%xmm10 \n\t" // a_r*x_r, a_r*x_i
"vmulps %%xmm1, %%xmm11, %%xmm11 \n\t" // a_i*x_r, a_i*x_i
"vfmaddps (%3,%0,4), %%xmm0 , %%xmm5, %%xmm5 \n\t"
".align 2 \n\t"
"vfmaddps 16(%3,%0,4), %%xmm0 , %%xmm7, %%xmm7 \n\t"
"vfmaddps 32(%3,%0,4), %%xmm0 , %%xmm9, %%xmm9 \n\t"
"vfmaddps 48(%3,%0,4), %%xmm0 , %%xmm11,%%xmm11 \n\t"

"vfmaddps %%xmm5 , %%xmm1 , %%xmm4 , %%xmm5 \n\t"
"vfmaddps %%xmm7 , %%xmm1 , %%xmm6 , %%xmm7 \n\t"
"vfmaddps %%xmm9 , %%xmm1 , %%xmm8 , %%xmm9 \n\t"
"vfmaddps %%xmm11, %%xmm1 , %%xmm10, %%xmm11 \n\t"

"vpermilps $0xb1 , %%xmm12, %%xmm4 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%xmm13, %%xmm6 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%xmm14, %%xmm8 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%xmm15, %%xmm10 \n\t" // exchange real and imag part

"vfmaddps 64(%3,%0,4), %%xmm0 , %%xmm12, %%xmm12 \n\t"
"vfmaddps 80(%3,%0,4), %%xmm0 , %%xmm13, %%xmm13 \n\t"
"vfmaddps 96(%3,%0,4), %%xmm0 , %%xmm14, %%xmm14 \n\t"
"vfmaddps 112(%3,%0,4), %%xmm0 , %%xmm15, %%xmm15 \n\t"

"vfmaddps %%xmm12, %%xmm1 , %%xmm4 , %%xmm12 \n\t"
"vfmaddps %%xmm13, %%xmm1 , %%xmm6 , %%xmm13 \n\t"
"vfmaddps %%xmm14, %%xmm1 , %%xmm8 , %%xmm14 \n\t"
"vfmaddps %%xmm15, %%xmm1 , %%xmm10, %%xmm15 \n\t"

"vmovups %%xmm5 , (%3,%0,4) \n\t"
".align 2 \n\t"
"vmovups %%xmm7 , 16(%3,%0,4) \n\t"
"vmovups %%xmm9 , 32(%3,%0,4) \n\t"
"vmovups %%xmm11, 48(%3,%0,4) \n\t"
"vmovups %%xmm12, 64(%3,%0,4) \n\t"
"vmovups %%xmm13, 80(%3,%0,4) \n\t"
"vmovups %%xmm14, 96(%3,%0,4) \n\t"
"vmovups %%xmm15,112(%3,%0,4) \n\t"

"addq $32, %0 \n\t"
"subq $16, %1 \n\t"
"jnz 1b \n\t"
"vzeroupper \n\t"

"vpermilps $0xb1 , %%xmm4 , %%xmm4 \n\t" // exchange real and imag part
"vaddsubps %%xmm4 ,%%xmm5 , %%xmm4 \n\t"
"vpermilps $0xb1 , %%xmm4 , %%xmm4 \n\t" // exchange real and imag part
:
:
"r" (i), // 0
"r" (n), // 1
"r" (x), // 2
"r" (y), // 3
"r" (alpha), // 4
"r" (mvec) // 5
: "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
return;
}

"vpermilps $0xb1 , %%xmm6 , %%xmm6 \n\t" // exchange real and imag part
"vaddsubps %%xmm6 ,%%xmm7 , %%xmm6 \n\t"
"vpermilps $0xb1 , %%xmm6 , %%xmm6 \n\t" // exchange real and imag part
__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastss (%4), %%xmm0 \n\t" // real part of alpha
"vbroadcastss 4(%4), %%xmm1 \n\t" // imag part of alpha
#if !defined(CONJ)
"vmulps (%5), %%xmm1 , %%xmm1 \n\t"
#else
"vmulps (%5), %%xmm0 , %%xmm0 \n\t"
#endif

"vpermilps $0xb1 , %%xmm8 , %%xmm8 \n\t" // exchange real and imag part
"vaddsubps %%xmm8 ,%%xmm9 , %%xmm8 \n\t"
"vpermilps $0xb1 , %%xmm8 , %%xmm8 \n\t" // exchange real and imag part
".align 16 \n\t"
"1: \n\t"

"vpermilps $0xb1 , %%xmm10, %%xmm10 \n\t" // exchange real and imag part
"vaddsubps %%xmm10,%%xmm11, %%xmm10 \n\t"
"vpermilps $0xb1 , %%xmm10, %%xmm10 \n\t" // exchange real and imag part
"prefetcht0 512(%2,%0,4) \n\t"
"vmovups (%2,%0,4), %%xmm5 \n\t" // 2 complex values from x
".align 2 \n\t"
"vmovups 16(%2,%0,4), %%xmm7 \n\t" // 2 complex values from x
"vmovups 32(%2,%0,4), %%xmm9 \n\t" // 2 complex values from x
"vmovups 48(%2,%0,4), %%xmm11 \n\t" // 2 complex values from x

"vaddps (%3,%0,4) ,%%xmm4 , %%xmm12 \n\t"
"vaddps 16(%3,%0,4) ,%%xmm6 , %%xmm13 \n\t"
"vaddps 32(%3,%0,4) ,%%xmm8 , %%xmm14 \n\t"
"vaddps 48(%3,%0,4) ,%%xmm10, %%xmm15 \n\t"
"vpermilps $0xb1 , %%xmm5 , %%xmm4 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%xmm7 , %%xmm6 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%xmm9 , %%xmm8 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%xmm11, %%xmm10 \n\t" // exchange real and imag part

"prefetcht0 512(%3,%0,4) \n\t"
"vfmaddps (%3,%0,4), %%xmm0 , %%xmm5, %%xmm5 \n\t"
".align 2 \n\t"
"vfmaddps 16(%3,%0,4), %%xmm0 , %%xmm7, %%xmm7 \n\t"
"vfmaddps 32(%3,%0,4), %%xmm0 , %%xmm9, %%xmm9 \n\t"
"vfmaddps 48(%3,%0,4), %%xmm0 , %%xmm11,%%xmm11 \n\t"

#endif
"vfmaddps %%xmm5 , %%xmm1 , %%xmm4 , %%xmm5 \n\t"
"vfmaddps %%xmm7 , %%xmm1 , %%xmm6 , %%xmm7 \n\t"
"vfmaddps %%xmm9 , %%xmm1 , %%xmm8 , %%xmm9 \n\t"
"vfmaddps %%xmm11, %%xmm1 , %%xmm10, %%xmm11 \n\t"

"vmovups %%xmm12, (%3,%0,4) \n\t"
"vmovups %%xmm13, 16(%3,%0,4) \n\t"
"vmovups %%xmm14, 32(%3,%0,4) \n\t"
"vmovups %%xmm15, 48(%3,%0,4) \n\t"
"vmovups %%xmm5 , (%3,%0,4) \n\t"
".align 2 \n\t"
"vmovups %%xmm7 , 16(%3,%0,4) \n\t"
"vmovups %%xmm9 , 32(%3,%0,4) \n\t"
"vmovups %%xmm11, 48(%3,%0,4) \n\t"

"addq $16, %0 \n\t"
"subq $8 , %1 \n\t"
"subq $8, %1 \n\t"
"jnz 1b \n\t"
"vzeroupper \n\t"

:
:
@@ -121,15 +186,15 @@ static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"r" (n), // 1
"r" (x), // 2
"r" (y), // 3
"r" (alpha) // 4
"r" (alpha), // 4
"r" (mvec) // 5
: "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);

}

}


+ 132
- 0
kernel/x86_64/caxpy_microk_haswell-2.c View File

@@ -0,0 +1,132 @@
/***************************************************************************
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_8 1
static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *alpha) __attribute__ ((noinline));

static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
{

#if !defined(CONJ)
FLOAT mvec[8] = { -1.0, 1.0, -1.0, 1.0, -1.0, 1.0, -1.0, 1.0 };
#else
FLOAT mvec[8] = { 1.0, -1.0, 1.0, -1.0, 1.0, -1.0, 1.0, -1.0 };
#endif

BLASLONG register i = 0;

__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastss (%4), %%ymm0 \n\t" // real part of alpha
"vbroadcastss 4(%4), %%ymm1 \n\t" // imag part of alpha
#if !defined(CONJ)
"vmulps (%5), %%ymm1 , %%ymm1 \n\t"
#else
"vmulps (%5), %%ymm0 , %%ymm0 \n\t"
#endif

".align 16 \n\t"
"1: \n\t"

"vmovups (%2,%0,4), %%ymm5 \n\t" // 4 complex values from x
".align 2 \n\t"
"vmovups 32(%2,%0,4), %%ymm7 \n\t" // 4 complex values from x
"vmovups 64(%2,%0,4), %%ymm9 \n\t" // 4 complex values from x
"vmovups 96(%2,%0,4), %%ymm11 \n\t" // 4 complex values from x

"vmovups 128(%2,%0,4), %%ymm12 \n\t" // 4 complex values from x
"vmovups 160(%2,%0,4), %%ymm13 \n\t" // 4 complex values from x
"vmovups 192(%2,%0,4), %%ymm14 \n\t" // 4 complex values from x
"vmovups 224(%2,%0,4), %%ymm15 \n\t" // 4 complex values from x

"vpermilps $0xb1 , %%ymm5 , %%ymm4 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%ymm7 , %%ymm6 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%ymm9 , %%ymm8 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%ymm11, %%ymm10 \n\t" // exchange real and imag part

"vfmadd213ps (%3,%0,4), %%ymm0 , %%ymm5 \n\t"
".align 2 \n\t"
"vfmadd213ps 32(%3,%0,4), %%ymm0 , %%ymm7 \n\t"
"vfmadd213ps 64(%3,%0,4), %%ymm0 , %%ymm9 \n\t"
"vfmadd213ps 96(%3,%0,4), %%ymm0 , %%ymm11 \n\t"

"vfmadd231ps %%ymm1 , %%ymm4 , %%ymm5 \n\t"
"vfmadd231ps %%ymm1 , %%ymm6 , %%ymm7 \n\t"
"vfmadd231ps %%ymm1 , %%ymm8 , %%ymm9 \n\t"
"vfmadd231ps %%ymm1 , %%ymm10, %%ymm11 \n\t"

"vpermilps $0xb1 , %%ymm12, %%ymm4 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%ymm13, %%ymm6 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%ymm14, %%ymm8 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%ymm15, %%ymm10 \n\t" // exchange real and imag part

"vfmadd213ps 128(%3,%0,4), %%ymm0 , %%ymm12 \n\t"
"vfmadd213ps 160(%3,%0,4), %%ymm0 , %%ymm13 \n\t"
"vfmadd213ps 192(%3,%0,4), %%ymm0 , %%ymm14 \n\t"
"vfmadd213ps 224(%3,%0,4), %%ymm0 , %%ymm15 \n\t"

"vfmadd231ps %%ymm1 , %%ymm4 , %%ymm12 \n\t"
"vfmadd231ps %%ymm1 , %%ymm6 , %%ymm13 \n\t"
"vfmadd231ps %%ymm1 , %%ymm8 , %%ymm14 \n\t"
"vfmadd231ps %%ymm1 , %%ymm10, %%ymm15 \n\t"

"vmovups %%ymm5 , (%3,%0,4) \n\t"
".align 2 \n\t"
"vmovups %%ymm7 , 32(%3,%0,4) \n\t"
"vmovups %%ymm9 , 64(%3,%0,4) \n\t"
"vmovups %%ymm11, 96(%3,%0,4) \n\t"

"vmovups %%ymm12,128(%3,%0,4) \n\t"
"vmovups %%ymm13,160(%3,%0,4) \n\t"
"vmovups %%ymm14,192(%3,%0,4) \n\t"
"vmovups %%ymm15,224(%3,%0,4) \n\t"

"addq $64, %0 \n\t"
"subq $32, %1 \n\t"
"jnz 1b \n\t"
"vzeroupper \n\t"

:
:
"r" (i), // 0
"r" (n), // 1
"r" (x), // 2
"r" (y), // 3
"r" (alpha), // 4
"r" (mvec) // 5
: "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);


}


+ 116
- 0
kernel/x86_64/caxpy_microk_sandy-2.c View File

@@ -0,0 +1,116 @@
/***************************************************************************
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_8 1
static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *alpha) __attribute__ ((noinline));

static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
{

#if !defined(CONJ)
FLOAT mvec[8] = { -1.0, 1.0, -1.0, 1.0, -1.0, 1.0, -1.0, 1.0 };
#else
FLOAT mvec[8] = { 1.0, -1.0, 1.0, -1.0, 1.0, -1.0, 1.0, -1.0 };
#endif

BLASLONG register i = 0;

__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastss (%4), %%ymm0 \n\t" // real part of alpha
"vbroadcastss 4(%4), %%ymm1 \n\t" // imag part of alpha
#if !defined(CONJ)
"vmulps (%5), %%ymm1 , %%ymm1 \n\t"
#else
"vmulps (%5), %%ymm0 , %%ymm0 \n\t"
#endif

".align 16 \n\t"
"1: \n\t"

"vmovups (%2,%0,4), %%ymm5 \n\t" // 4 complex values from x
".align 2 \n\t"
"vmovups 32(%2,%0,4), %%ymm7 \n\t" // 4 complex values from x
"vmovups 64(%2,%0,4), %%ymm9 \n\t" // 4 complex values from x
"vmovups 96(%2,%0,4), %%ymm11 \n\t" // 4 complex values from x

"vpermilps $0xb1 , %%ymm5 , %%ymm4 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%ymm7 , %%ymm6 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%ymm9 , %%ymm8 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%ymm11, %%ymm10 \n\t" // exchange real and imag part

"vmulps %%ymm5 , %%ymm0 , %%ymm5 \n\t"
"vmulps %%ymm7 , %%ymm0 , %%ymm7 \n\t"
"vmulps %%ymm9 , %%ymm0 , %%ymm9 \n\t"
"vmulps %%ymm11, %%ymm0 , %%ymm11 \n\t"

"vaddps (%3,%0,4), %%ymm5 , %%ymm5 \n\t"
"vaddps 32(%3,%0,4), %%ymm7 , %%ymm7 \n\t"
"vaddps 64(%3,%0,4), %%ymm9 , %%ymm9 \n\t"
"vaddps 96(%3,%0,4), %%ymm11, %%ymm11 \n\t"

"vmulps %%ymm4 , %%ymm1 , %%ymm4 \n\t"
"vmulps %%ymm6 , %%ymm1 , %%ymm6 \n\t"
"vmulps %%ymm8 , %%ymm1 , %%ymm8 \n\t"
"vmulps %%ymm10, %%ymm1 , %%ymm10 \n\t"

"vaddps %%ymm4 , %%ymm5 , %%ymm5 \n\t"
"vaddps %%ymm6 , %%ymm7 , %%ymm7 \n\t"
"vaddps %%ymm8 , %%ymm9 , %%ymm9 \n\t"
"vaddps %%ymm10, %%ymm11, %%ymm11 \n\t"

"vmovups %%ymm5 , (%3,%0,4) \n\t"
".align 2 \n\t"
"vmovups %%ymm7 , 32(%3,%0,4) \n\t"
"vmovups %%ymm9 , 64(%3,%0,4) \n\t"
"vmovups %%ymm11, 96(%3,%0,4) \n\t"

"addq $32, %0 \n\t"
"subq $16, %1 \n\t"
"jnz 1b \n\t"
"vzeroupper \n\t"

:
:
"r" (i), // 0
"r" (n), // 1
"r" (x), // 2
"r" (y), // 3
"r" (alpha), // 4
"r" (mvec) // 5
: "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);


}


+ 200
- 0
kernel/x86_64/caxpy_microk_steamroller-2.c View File

@@ -0,0 +1,200 @@
/***************************************************************************
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_8 1
static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *alpha) __attribute__ ((noinline));

static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
{

#if !defined(CONJ)
FLOAT mvec[4] = { -1.0, 1.0, -1.0, 1.0 };
#else
FLOAT mvec[4] = { 1.0, -1.0, 1.0, -1.0 };
#endif

BLASLONG register i = 0;

if ( n <= 2048 )
{

__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastss (%4), %%xmm0 \n\t" // real part of alpha
"vbroadcastss 4(%4), %%xmm1 \n\t" // imag part of alpha
#if !defined(CONJ)
"vmulps (%5), %%xmm1 , %%xmm1 \n\t"
#else
"vmulps (%5), %%xmm0 , %%xmm0 \n\t"
#endif

".align 16 \n\t"
"1: \n\t"

"vmovups (%2,%0,4), %%xmm5 \n\t" // 2 complex values from x
".align 2 \n\t"
"vmovups 16(%2,%0,4), %%xmm7 \n\t" // 2 complex values from x
"vmovups 32(%2,%0,4), %%xmm9 \n\t" // 2 complex values from x
"vmovups 48(%2,%0,4), %%xmm11 \n\t" // 2 complex values from x

"vmovups 64(%2,%0,4), %%xmm12 \n\t" // 2 complex values from x
"vmovups 80(%2,%0,4), %%xmm13 \n\t" // 2 complex values from x
"vmovups 96(%2,%0,4), %%xmm14 \n\t" // 2 complex values from x
"vmovups 112(%2,%0,4), %%xmm15 \n\t" // 2 complex values from x

"vpermilps $0xb1 , %%xmm5 , %%xmm4 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%xmm7 , %%xmm6 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%xmm9 , %%xmm8 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%xmm11, %%xmm10 \n\t" // exchange real and imag part

"vfmadd213ps (%3,%0,4), %%xmm0 , %%xmm5 \n\t"
".align 2 \n\t"
"vfmadd213ps 16(%3,%0,4), %%xmm0 , %%xmm7 \n\t"
"vfmadd213ps 32(%3,%0,4), %%xmm0 , %%xmm9 \n\t"
"vfmadd213ps 48(%3,%0,4), %%xmm0 , %%xmm11 \n\t"

"vfmadd231ps %%xmm1 , %%xmm4 , %%xmm5 \n\t"
"vfmadd231ps %%xmm1 , %%xmm6 , %%xmm7 \n\t"
"vfmadd231ps %%xmm1 , %%xmm8 , %%xmm9 \n\t"
"vfmadd231ps %%xmm1 , %%xmm10, %%xmm11 \n\t"

"vpermilps $0xb1 , %%xmm12, %%xmm4 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%xmm13, %%xmm6 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%xmm14, %%xmm8 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%xmm15, %%xmm10 \n\t" // exchange real and imag part

"vfmadd213ps 64(%3,%0,4), %%xmm0 , %%xmm12 \n\t"
"vfmadd213ps 80(%3,%0,4), %%xmm0 , %%xmm13 \n\t"
"vfmadd213ps 96(%3,%0,4), %%xmm0 , %%xmm14 \n\t"
"vfmadd213ps 112(%3,%0,4), %%xmm0 , %%xmm15 \n\t"

"vfmadd231ps %%xmm1 , %%xmm4 , %%xmm12 \n\t"
"vfmadd231ps %%xmm1 , %%xmm6 , %%xmm13 \n\t"
"vfmadd231ps %%xmm1 , %%xmm8 , %%xmm14 \n\t"
"vfmadd231ps %%xmm1 , %%xmm10, %%xmm15 \n\t"

"vmovups %%xmm5 , (%3,%0,4) \n\t"
".align 2 \n\t"
"vmovups %%xmm7 , 16(%3,%0,4) \n\t"
"vmovups %%xmm9 , 32(%3,%0,4) \n\t"
"vmovups %%xmm11, 48(%3,%0,4) \n\t"
"vmovups %%xmm12, 64(%3,%0,4) \n\t"
"vmovups %%xmm13, 80(%3,%0,4) \n\t"
"vmovups %%xmm14, 96(%3,%0,4) \n\t"
"vmovups %%xmm15,112(%3,%0,4) \n\t"

"addq $32, %0 \n\t"
"subq $16, %1 \n\t"
"jnz 1b \n\t"
"vzeroupper \n\t"

:
:
"r" (i), // 0
"r" (n), // 1
"r" (x), // 2
"r" (y), // 3
"r" (alpha), // 4
"r" (mvec) // 5
: "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
return;
}

__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastss (%4), %%xmm0 \n\t" // real part of alpha
"vbroadcastss 4(%4), %%xmm1 \n\t" // imag part of alpha
#if !defined(CONJ)
"vmulps (%5), %%xmm1 , %%xmm1 \n\t"
#else
"vmulps (%5), %%xmm0 , %%xmm0 \n\t"
#endif

".align 16 \n\t"
"1: \n\t"

"prefetcht0 512(%2,%0,4) \n\t"
"vmovups (%2,%0,4), %%xmm5 \n\t" // 2 complex values from x
".align 2 \n\t"
"vmovups 16(%2,%0,4), %%xmm7 \n\t" // 2 complex values from x
"vmovups 32(%2,%0,4), %%xmm9 \n\t" // 2 complex values from x
"vmovups 48(%2,%0,4), %%xmm11 \n\t" // 2 complex values from x

"vpermilps $0xb1 , %%xmm5 , %%xmm4 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%xmm7 , %%xmm6 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%xmm9 , %%xmm8 \n\t" // exchange real and imag part
"vpermilps $0xb1 , %%xmm11, %%xmm10 \n\t" // exchange real and imag part

"prefetcht0 512(%3,%0,4) \n\t"
"vfmadd213ps (%3,%0,4), %%xmm0 , %%xmm5 \n\t"
".align 2 \n\t"
"vfmadd213ps 16(%3,%0,4), %%xmm0 , %%xmm7 \n\t"
"vfmadd213ps 32(%3,%0,4), %%xmm0 , %%xmm9 \n\t"
"vfmadd213ps 48(%3,%0,4), %%xmm0 , %%xmm11 \n\t"

"vfmadd231ps %%xmm1 , %%xmm4 , %%xmm5 \n\t"
"vfmadd231ps %%xmm1 , %%xmm6 , %%xmm7 \n\t"
"vfmadd231ps %%xmm1 , %%xmm8 , %%xmm9 \n\t"
"vfmadd231ps %%xmm1 , %%xmm10, %%xmm11 \n\t"

"vmovups %%xmm5 , (%3,%0,4) \n\t"
".align 2 \n\t"
"vmovups %%xmm7 , 16(%3,%0,4) \n\t"
"vmovups %%xmm9 , 32(%3,%0,4) \n\t"
"vmovups %%xmm11, 48(%3,%0,4) \n\t"

"addq $16, %0 \n\t"
"subq $8, %1 \n\t"
"jnz 1b \n\t"
"vzeroupper \n\t"

:
:
"r" (i), // 0
"r" (n), // 1
"r" (x), // 2
"r" (y), // 3
"r" (alpha), // 4
"r" (mvec) // 5
: "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"memory"
);


}


Write Preview
Loading…
Cancel
Save