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 16012767f4 into 646e168d26

pull/235/merge
wernsaar 13 years ago
parent
646e168d26 16012767f4
commit
b5aecb6dd4
6 changed files with 3254 additions and 304 deletions
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +4
    -0
      kernel/x86_64/KERNEL.BULLDOZER
  2. +408
    -0
      kernel/x86_64/daxpy_bulldozer.S
  3. +291
    -0
      kernel/x86_64/dcopy_bulldozer.S
  4. +311
    -0
      kernel/x86_64/ddot_bulldozer.S
  5. +302
    -304
      kernel/x86_64/dgemm_ncopy_8_bulldozer.S
  6. +1938
    -0
      kernel/x86_64/dgemv_t_bulldozer.S

+ 4
- 0
kernel/x86_64/KERNEL.BULLDOZER View File

@@ -2,6 +2,10 @@ ZGEMVNKERNEL = zgemv_n_dup.S
ZGEMVTKERNEL = zgemv_t_dup.S

DGEMVNKERNEL = dgemv_n_bulldozer.S
DGEMVTKERNEL = dgemv_t_bulldozer.S
DAXPYKERNEL = daxpy_bulldozer.S
DDOTKERNEL = ddot_bulldozer.S
DCOPYKERNEL = dcopy_bulldozer.S

SGEMMKERNEL = sgemm_kernel_16x2_bulldozer.S
SGEMMINCOPY = ../generic/gemm_ncopy_16.c


+ 408
- 0
kernel/x86_64/daxpy_bulldozer.S View File

@@ -0,0 +1,408 @@
/*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */
/* 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. */
/* */
/* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
/* AUSTIN ``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 UNIVERSITY OF TEXAS AT */
/* AUSTIN 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. */
/* */
/* The views and conclusions contained in the software and */
/* documentation are those of the authors and should not be */
/* interpreted as representing official policies, either expressed */
/* or implied, of The University of Texas at Austin. */
/*********************************************************************/

#define ASSEMBLER
#include "common.h"

#ifndef WINDOWS_ABI
#define M ARG1
#define X ARG4
#define INCX ARG5
#define Y ARG6
#define INCY ARG2
#else
#define M ARG1
#define X ARG2
#define INCX ARG3
#define Y ARG4
#define INCY %r10
#endif

#define YY %r11
#define ALPHA %xmm15

#define A_PRE 640

#include "l1param.h"
PROLOGUE
PROFCODE

#ifndef WINDOWS_ABI
#ifndef XDOUBLE
movq 8(%rsp), INCY
#else
movq 24(%rsp), INCY
#endif
vmovups %xmm0, ALPHA
#else
vmovups %xmm3, ALPHA

movq 40(%rsp), X
movq 48(%rsp), INCX
movq 56(%rsp), Y
movq 64(%rsp), INCY
#endif

SAVEREGISTERS

unpcklpd ALPHA, ALPHA

leaq (, INCX, SIZE), INCX
leaq (, INCY, SIZE), INCY

testq M, M
jle .L47
cmpq $SIZE, INCX
jne .L40
cmpq $SIZE, INCY
jne .L40

testq $SIZE, Y
je .L10

movsd (X), %xmm0
mulsd ALPHA, %xmm0
addsd (Y), %xmm0
movsd %xmm0, (Y)
addq $1 * SIZE, X
addq $1 * SIZE, Y
decq M
jle .L19
ALIGN_4

.L10:
subq $-16 * SIZE, X
subq $-16 * SIZE, Y

movq M, %rax
sarq $4, %rax
jle .L13

vmovups -16 * SIZE(X), %xmm0
vmovups -14 * SIZE(X), %xmm1
vmovups -12 * SIZE(X), %xmm2
vmovups -10 * SIZE(X), %xmm3

decq %rax
jle .L12
ALIGN_3

.L11:

prefetchnta A_PRE(Y)

vmovups -8 * SIZE(X), %xmm4
vfmaddpd -16 * SIZE(Y), ALPHA, %xmm0 , %xmm0
vfmaddpd -14 * SIZE(Y), ALPHA, %xmm1 , %xmm1
vmovups -6 * SIZE(X), %xmm5
vmovups -4 * SIZE(X), %xmm6
vfmaddpd -12 * SIZE(Y), ALPHA, %xmm2 , %xmm2
vfmaddpd -10 * SIZE(Y), ALPHA, %xmm3 , %xmm3
vmovups -2 * SIZE(X), %xmm7


vmovups %xmm0, -16 * SIZE(Y)
vmovups %xmm1, -14 * SIZE(Y)
prefetchnta A_PRE(X)
nop
vmovups %xmm2, -12 * SIZE(Y)
vmovups %xmm3, -10 * SIZE(Y)

prefetchnta A_PRE+64(Y)

vmovups 0 * SIZE(X), %xmm0
vfmaddpd -8 * SIZE(Y), ALPHA, %xmm4 , %xmm4
vfmaddpd -6 * SIZE(Y), ALPHA, %xmm5 , %xmm5
vmovups 2 * SIZE(X), %xmm1
vmovups 4 * SIZE(X), %xmm2
vfmaddpd -4 * SIZE(Y), ALPHA, %xmm6 , %xmm6
vfmaddpd -2 * SIZE(Y), ALPHA, %xmm7 , %xmm7
vmovups 6 * SIZE(X), %xmm3


vmovups %xmm4, -8 * SIZE(Y)
vmovups %xmm5, -6 * SIZE(Y)
prefetchnta A_PRE+64(X)
nop
vmovups %xmm6, -4 * SIZE(Y)
vmovups %xmm7, -2 * SIZE(Y)

subq $-16 * SIZE, Y
subq $-16 * SIZE, X
decq %rax
jg .L11
ALIGN_3

.L12:

vmovups -8 * SIZE(X), %xmm4
vfmaddpd -16 * SIZE(Y), ALPHA, %xmm0 , %xmm0
vfmaddpd -14 * SIZE(Y), ALPHA, %xmm1 , %xmm1
vmovups -6 * SIZE(X), %xmm5
vmovups -4 * SIZE(X), %xmm6
vfmaddpd -12 * SIZE(Y), ALPHA, %xmm2 , %xmm2
vfmaddpd -10 * SIZE(Y), ALPHA, %xmm3 , %xmm3
vmovups -2 * SIZE(X), %xmm7


vmovups %xmm0, -16 * SIZE(Y)
vmovups %xmm1, -14 * SIZE(Y)
vmovups %xmm2, -12 * SIZE(Y)
vmovups %xmm3, -10 * SIZE(Y)

vfmaddpd -8 * SIZE(Y), ALPHA, %xmm4 , %xmm4
vfmaddpd -6 * SIZE(Y), ALPHA, %xmm5 , %xmm5
vfmaddpd -4 * SIZE(Y), ALPHA, %xmm6 , %xmm6
vfmaddpd -2 * SIZE(Y), ALPHA, %xmm7 , %xmm7

vmovups %xmm4, -8 * SIZE(Y)
vmovups %xmm5, -6 * SIZE(Y)
vmovups %xmm6, -4 * SIZE(Y)
vmovups %xmm7, -2 * SIZE(Y)

subq $-16 * SIZE, Y
subq $-16 * SIZE, X
ALIGN_3

.L13:


movq M, %rax
andq $8, %rax
jle .L14
ALIGN_3

vmovups -16 * SIZE(X), %xmm0
vmovups -14 * SIZE(X), %xmm1
vmovups -12 * SIZE(X), %xmm2
vmovups -10 * SIZE(X), %xmm3

vfmaddpd -16 * SIZE(Y), ALPHA, %xmm0 , %xmm0
vfmaddpd -14 * SIZE(Y), ALPHA, %xmm1 , %xmm1
vfmaddpd -12 * SIZE(Y), ALPHA, %xmm2 , %xmm2
vfmaddpd -10 * SIZE(Y), ALPHA, %xmm3 , %xmm3

vmovups %xmm0, -16 * SIZE(Y)
vmovups %xmm1, -14 * SIZE(Y)
vmovups %xmm2, -12 * SIZE(Y)
vmovups %xmm3, -10 * SIZE(Y)

addq $8 * SIZE, X
addq $8 * SIZE, Y
ALIGN_3

.L14:
movq M, %rax
andq $4, %rax
jle .L15
ALIGN_3

vmovups -16 * SIZE(X), %xmm0
vmovups -14 * SIZE(X), %xmm1

vfmaddpd -16 * SIZE(Y), ALPHA, %xmm0 , %xmm0
vfmaddpd -14 * SIZE(Y), ALPHA, %xmm1 , %xmm1

vmovups %xmm0, -16 * SIZE(Y)
vmovups %xmm1, -14 * SIZE(Y)

addq $4 * SIZE, X
addq $4 * SIZE, Y
ALIGN_3

.L15:
movq M, %rax
andq $2, %rax
jle .L16
ALIGN_3

vmovups -16 * SIZE(X), %xmm0
vfmaddpd -16 * SIZE(Y), ALPHA, %xmm0 , %xmm0
vmovups %xmm0, -16 * SIZE(Y)

addq $2 * SIZE, X
addq $2 * SIZE, Y
ALIGN_3

.L16:
movq M, %rax
andq $1, %rax
jle .L19
ALIGN_3

vmovsd -16 * SIZE(X), %xmm0
vfmaddsd -16 * SIZE(Y), ALPHA, %xmm0 , %xmm0

vmovsd %xmm0, -16 * SIZE(Y)
ALIGN_3

.L19:
xorq %rax,%rax

RESTOREREGISTERS

ret
ALIGN_3


.L40:
movq Y, YY
movq M, %rax
//If incx==0 || incy==0, avoid unloop.
cmpq $0, INCX
je .L46
cmpq $0, INCY
je .L46
sarq $3, %rax
jle .L45

prefetchnta 512(X)
prefetchnta 512+64(X)
prefetchnta 512+128(X)
prefetchnta 512+192(X)

prefetchnta 512(Y)
prefetchnta 512+64(Y)
prefetchnta 512+128(Y)
prefetchnta 512+192(Y)
ALIGN_3

.L41:

vmovsd 0 * SIZE(X), %xmm0
addq INCX, X
vmovhpd 0 * SIZE(X), %xmm0 , %xmm0
addq INCX, X

vmovsd 0 * SIZE(YY), %xmm6
addq INCY, YY
vmovhpd 0 * SIZE(YY), %xmm6 , %xmm6
addq INCY, YY


vmovsd 0 * SIZE(X), %xmm1
addq INCX, X
vmovhpd 0 * SIZE(X), %xmm1 , %xmm1
addq INCX, X

vmovsd 0 * SIZE(YY), %xmm7
addq INCY, YY
vmovhpd 0 * SIZE(YY), %xmm7 , %xmm7
addq INCY, YY

vfmaddpd %xmm6 , ALPHA , %xmm0 , %xmm0

vmovsd 0 * SIZE(X), %xmm2
addq INCX, X
vmovhpd 0 * SIZE(X), %xmm2 , %xmm2
addq INCX, X

vmovsd 0 * SIZE(YY), %xmm8
addq INCY, YY
vmovhpd 0 * SIZE(YY), %xmm8 , %xmm8
addq INCY, YY

vfmaddpd %xmm7 , ALPHA , %xmm1 , %xmm1

vmovsd 0 * SIZE(X), %xmm3
addq INCX, X
vmovhpd 0 * SIZE(X), %xmm3 , %xmm3
addq INCX, X

vfmaddpd %xmm8 , ALPHA , %xmm2 , %xmm2

vmovsd 0 * SIZE(YY), %xmm9
addq INCY, YY
vmovhpd 0 * SIZE(YY), %xmm9 , %xmm9
addq INCY, YY


vmovsd %xmm0, 0 * SIZE(Y)
addq INCY, Y
vmovhpd %xmm0, 0 * SIZE(Y)
addq INCY, Y
vmovsd %xmm1, 0 * SIZE(Y)
addq INCY, Y
vmovhpd %xmm1, 0 * SIZE(Y)
addq INCY, Y
vmovsd %xmm2, 0 * SIZE(Y)
addq INCY, Y
vmovhpd %xmm2, 0 * SIZE(Y)
addq INCY, Y

vfmaddpd %xmm9 , ALPHA , %xmm3 , %xmm3

vmovsd %xmm3, 0 * SIZE(Y)
addq INCY, Y
vmovhpd %xmm3, 0 * SIZE(Y)
addq INCY, Y

decq %rax
jg .L41
ALIGN_3

.L45:
movq M, %rax
andq $7, %rax
jle .L47
ALIGN_3

.L46:
vmovsd (X), %xmm0
addq INCX, X

vfmaddsd (Y) , ALPHA , %xmm0 , %xmm0

vmovsd %xmm0, (Y)
addq INCY, Y

decq %rax
jg .L46
ALIGN_3

.L47:
xorq %rax, %rax

RESTOREREGISTERS

ret

EPILOGUE

+ 291
- 0
kernel/x86_64/dcopy_bulldozer.S View File

@@ -0,0 +1,291 @@
/*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */
/* 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. */
/* */
/* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
/* AUSTIN ``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 UNIVERSITY OF TEXAS AT */
/* AUSTIN 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. */
/* */
/* The views and conclusions contained in the software and */
/* documentation are those of the authors and should not be */
/* interpreted as representing official policies, either expressed */
/* or implied, of The University of Texas at Austin. */
/*********************************************************************/

#define ASSEMBLER
#include "common.h"

#define M ARG1 /* rdi */
#define X ARG2 /* rsi */
#define INCX ARG3 /* rdx */
#define Y ARG4 /* rcx */
#ifndef WINDOWS_ABI
#define INCY ARG5 /* r8 */
#else
#define INCY %r10
#endif

#include "l1param.h"

#define VLOAD(OFFSET, ADDR, REG) vmovups OFFSET(ADDR), REG
#define VSHUFPD_1(REG1 , REG2) vshufpd $0x01, REG1, REG2, REG2
#define A_PRE 640
#define B_PRE 640

PROLOGUE
PROFCODE

#ifdef WINDOWS_ABI
movq 40(%rsp), INCY
#endif

SAVEREGISTERS

leaq (, INCX, SIZE), INCX
leaq (, INCY, SIZE), INCY

cmpq $SIZE, INCX
jne .L40
cmpq $SIZE, INCY
jne .L40

testq $SIZE, X
je .L10

vmovsd (X), %xmm0
vmovsd %xmm0, (Y)
addq $1 * SIZE, X
addq $1 * SIZE, Y
decq M
jle .L19
ALIGN_4

.L10:
subq $-16 * SIZE, X
subq $-16 * SIZE, Y


movq M, %rax
sarq $4, %rax
jle .L13

vmovups -16 * SIZE(X), %xmm0
vmovups -14 * SIZE(X), %xmm1
vmovups -12 * SIZE(X), %xmm2
vmovups -10 * SIZE(X), %xmm3
vmovups -8 * SIZE(X), %xmm4
vmovups -6 * SIZE(X), %xmm5
vmovups -4 * SIZE(X), %xmm6
vmovups -2 * SIZE(X), %xmm7

decq %rax
jle .L12
ALIGN_4

.L11:

prefetchnta A_PRE(X)
nop
vmovups %xmm0, -16 * SIZE(Y)
vmovups %xmm1, -14 * SIZE(Y)
prefetchnta B_PRE(Y)
nop
vmovups %xmm2, -12 * SIZE(Y)
vmovups %xmm3, -10 * SIZE(Y)

VLOAD( 0 * SIZE, X, %xmm0)
VLOAD( 2 * SIZE, X, %xmm1)
VLOAD( 4 * SIZE, X, %xmm2)
VLOAD( 6 * SIZE, X, %xmm3)

prefetchnta A_PRE+64(X)
nop
vmovups %xmm4, -8 * SIZE(Y)
vmovups %xmm5, -6 * SIZE(Y)
prefetchnta B_PRE+64(Y)
nop
vmovups %xmm6, -4 * SIZE(Y)
vmovups %xmm7, -2 * SIZE(Y)

VLOAD( 8 * SIZE, X, %xmm4)
VLOAD(10 * SIZE, X, %xmm5)
subq $-16 * SIZE, Y
VLOAD(12 * SIZE, X, %xmm6)
VLOAD(14 * SIZE, X, %xmm7)

subq $-16 * SIZE, X
decq %rax
jg .L11
ALIGN_3

.L12:
vmovups %xmm0, -16 * SIZE(Y)
vmovups %xmm1, -14 * SIZE(Y)
vmovups %xmm2, -12 * SIZE(Y)
vmovups %xmm3, -10 * SIZE(Y)
vmovups %xmm4, -8 * SIZE(Y)
vmovups %xmm5, -6 * SIZE(Y)
vmovups %xmm6, -4 * SIZE(Y)
vmovups %xmm7, -2 * SIZE(Y)

subq $-16 * SIZE, Y
subq $-16 * SIZE, X
ALIGN_3

.L13:
testq $8, M
jle .L14
ALIGN_3

vmovups -16 * SIZE(X), %xmm0
vmovups -14 * SIZE(X), %xmm1
vmovups -12 * SIZE(X), %xmm2
vmovups -10 * SIZE(X), %xmm3

vmovups %xmm0, -16 * SIZE(Y)
vmovups %xmm1, -14 * SIZE(Y)
vmovups %xmm2, -12 * SIZE(Y)
vmovups %xmm3, -10 * SIZE(Y)

addq $8 * SIZE, X
addq $8 * SIZE, Y
ALIGN_3

.L14:
testq $4, M
jle .L15
ALIGN_3

vmovups -16 * SIZE(X), %xmm0
vmovups -14 * SIZE(X), %xmm1

vmovups %xmm0, -16 * SIZE(Y)
vmovups %xmm1, -14 * SIZE(Y)

addq $4 * SIZE, X
addq $4 * SIZE, Y
ALIGN_3

.L15:
testq $2, M
jle .L16
ALIGN_3

vmovups -16 * SIZE(X), %xmm0
vmovups %xmm0, -16 * SIZE(Y)

addq $2 * SIZE, X
addq $2 * SIZE, Y
ALIGN_3

.L16:
testq $1, M
jle .L19
ALIGN_3

vmovsd -16 * SIZE(X), %xmm0
vmovsd %xmm0, -16 * SIZE(Y)
ALIGN_3

.L19:
xorq %rax,%rax

RESTOREREGISTERS

ret
ALIGN_3



.L40:
movq M, %rax
sarq $3, %rax
jle .L45
ALIGN_3

.L41:
vmovsd (X), %xmm0
addq INCX, X
vmovsd (X), %xmm4
addq INCX, X
vmovsd (X), %xmm1
addq INCX, X
vmovsd (X), %xmm5
addq INCX, X
vmovsd (X), %xmm2
addq INCX, X
vmovsd (X), %xmm6
addq INCX, X
vmovsd (X), %xmm3
addq INCX, X
vmovsd (X), %xmm7
addq INCX, X

vmovsd %xmm0, (Y)
addq INCY, Y
vmovsd %xmm4, (Y)
addq INCY, Y
vmovsd %xmm1, (Y)
addq INCY, Y
vmovsd %xmm5, (Y)
addq INCY, Y
vmovsd %xmm2, (Y)
addq INCY, Y
vmovsd %xmm6, (Y)
addq INCY, Y
vmovsd %xmm3, (Y)
addq INCY, Y
vmovsd %xmm7, (Y)
addq INCY, Y

decq %rax
jg .L41
ALIGN_3

.L45:
movq M, %rax
andq $7, %rax
jle .L47
ALIGN_3

.L46:
vmovsd (X), %xmm0
addq INCX, X
vmovsd %xmm0, (Y)
addq INCY, Y
decq %rax
jg .L46
ALIGN_3

.L47:
xorq %rax, %rax

RESTOREREGISTERS

ret

EPILOGUE

+ 311
- 0
kernel/x86_64/ddot_bulldozer.S View File

@@ -0,0 +1,311 @@
/*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */
/* 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. */
/* */
/* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
/* AUSTIN ``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 UNIVERSITY OF TEXAS AT */
/* AUSTIN 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. */
/* */
/* The views and conclusions contained in the software and */
/* documentation are those of the authors and should not be */
/* interpreted as representing official policies, either expressed */
/* or implied, of The University of Texas at Austin. */
/*********************************************************************/

#define ASSEMBLER
#include "common.h"

#define N ARG1 /* rdi */
#define X ARG2 /* rsi */
#define INCX ARG3 /* rdx */
#define Y ARG4 /* rcx */
#ifndef WINDOWS_ABI
#define INCY ARG5 /* r8 */
#else
#define INCY %r10
#endif

#define A_PRE 512

#include "l1param.h"

PROLOGUE
PROFCODE

#ifdef WINDOWS_ABI
movq 40(%rsp), INCY
#endif

SAVEREGISTERS

leaq (, INCX, SIZE), INCX
leaq (, INCY, SIZE), INCY

vxorps %xmm0, %xmm0 , %xmm0
vxorps %xmm1, %xmm1 , %xmm1
vxorps %xmm2, %xmm2 , %xmm2
vxorps %xmm3, %xmm3 , %xmm3

cmpq $0, N
jle .L999

cmpq $SIZE, INCX
jne .L50
cmpq $SIZE, INCY
jne .L50

subq $-16 * SIZE, X
subq $-16 * SIZE, Y

testq $SIZE, Y
je .L10

vmovsd -16 * SIZE(X), %xmm0
vmulsd -16 * SIZE(Y), %xmm0 , %xmm0
addq $1 * SIZE, X
addq $1 * SIZE, Y
decq N
ALIGN_2

.L10:

movq N, %rax
sarq $4, %rax
jle .L14

vmovups -16 * SIZE(X), %xmm4
vmovups -14 * SIZE(X), %xmm5
vmovups -12 * SIZE(X), %xmm6
vmovups -10 * SIZE(X), %xmm7

vmovups -8 * SIZE(X), %xmm8
vmovups -6 * SIZE(X), %xmm9
vmovups -4 * SIZE(X), %xmm10
vmovups -2 * SIZE(X), %xmm11

decq %rax
jle .L12

ALIGN_3

.L11:
prefetchnta A_PRE(Y)

vfmaddpd %xmm0 , -16 * SIZE(Y), %xmm4 , %xmm0
vfmaddpd %xmm1 , -14 * SIZE(Y), %xmm5 , %xmm1
prefetchnta A_PRE(X)
vfmaddpd %xmm2 , -12 * SIZE(Y), %xmm6 , %xmm2
vfmaddpd %xmm3 , -10 * SIZE(Y), %xmm7 , %xmm3

vmovups 0 * SIZE(X), %xmm4
vfmaddpd %xmm0 , -8 * SIZE(Y), %xmm8 , %xmm0
vfmaddpd %xmm1 , -6 * SIZE(Y), %xmm9 , %xmm1
vmovups 2 * SIZE(X), %xmm5
vmovups 4 * SIZE(X), %xmm6
vfmaddpd %xmm2 , -4 * SIZE(Y), %xmm10, %xmm2
vfmaddpd %xmm3 , -2 * SIZE(Y), %xmm11, %xmm3
vmovups 6 * SIZE(X), %xmm7

prefetchnta A_PRE+64(Y)

vmovups 8 * SIZE(X), %xmm8
vmovups 10 * SIZE(X), %xmm9
prefetchnta A_PRE+64(X)
vmovups 12 * SIZE(X), %xmm10
vmovups 14 * SIZE(X), %xmm11

subq $-16 * SIZE, X
subq $-16 * SIZE, Y

decq %rax
jg .L11
ALIGN_3

.L12:

vfmaddpd %xmm0 , -16 * SIZE(Y), %xmm4 , %xmm0
vfmaddpd %xmm1 , -14 * SIZE(Y), %xmm5 , %xmm1
vfmaddpd %xmm2 , -12 * SIZE(Y), %xmm6 , %xmm2
vfmaddpd %xmm3 , -10 * SIZE(Y), %xmm7 , %xmm3

vfmaddpd %xmm0 , -8 * SIZE(Y), %xmm8 , %xmm0
vfmaddpd %xmm1 , -6 * SIZE(Y), %xmm9 , %xmm1
vfmaddpd %xmm2 , -4 * SIZE(Y), %xmm10, %xmm2
vfmaddpd %xmm3 , -2 * SIZE(Y), %xmm11, %xmm3

subq $-16 * SIZE, X
subq $-16 * SIZE, Y
ALIGN_3

.L14:
testq $15, N
jle .L999

testq $8, N
jle .L15

vmovups -16 * SIZE(X), %xmm4
vmovups -14 * SIZE(X), %xmm5
vmovups -12 * SIZE(X), %xmm6
vmovups -10 * SIZE(X), %xmm7

vfmaddpd %xmm0 , -16 * SIZE(Y), %xmm4 , %xmm0
vfmaddpd %xmm1 , -14 * SIZE(Y), %xmm5 , %xmm1
vfmaddpd %xmm2 , -12 * SIZE(Y), %xmm6 , %xmm2
vfmaddpd %xmm3 , -10 * SIZE(Y), %xmm7 , %xmm3

addq $8 * SIZE, X
addq $8 * SIZE, Y
ALIGN_3

.L15:
testq $4, N
jle .L16

vmovups -16 * SIZE(X), %xmm4
vmovups -14 * SIZE(X), %xmm5

vfmaddpd %xmm0 , -16 * SIZE(Y), %xmm4 , %xmm0
vfmaddpd %xmm1 , -14 * SIZE(Y), %xmm5 , %xmm1

addq $4 * SIZE, X
addq $4 * SIZE, Y
ALIGN_3

.L16:
testq $2, N
jle .L17

vmovups -16 * SIZE(X), %xmm4
vfmaddpd %xmm0 , -16 * SIZE(Y), %xmm4 , %xmm0


addq $2 * SIZE, X
addq $2 * SIZE, Y
ALIGN_3

.L17:
testq $1, N
jle .L999

vmovsd -16 * SIZE(X), %xmm4
vmovsd -16 * SIZE(Y), %xmm5
vfmaddpd %xmm0, %xmm4 , %xmm5 , %xmm0
jmp .L999
ALIGN_3


.L50:
movq N, %rax
sarq $3, %rax
jle .L55
ALIGN_3

.L53:


vmovsd 0 * SIZE(X), %xmm4
addq INCX, X
vmovsd 0 * SIZE(Y), %xmm8
addq INCY, Y
vmovsd 0 * SIZE(X), %xmm5
addq INCX, X
vmovsd 0 * SIZE(Y), %xmm9
addq INCY, Y

vmovsd 0 * SIZE(X), %xmm6
addq INCX, X
vmovsd 0 * SIZE(Y), %xmm10
addq INCY, Y
vmovsd 0 * SIZE(X), %xmm7
addq INCX, X
vmovsd 0 * SIZE(Y), %xmm11
addq INCY, Y

vfmaddpd %xmm0 , %xmm4 , %xmm8 , %xmm0
vfmaddpd %xmm1 , %xmm5 , %xmm9 , %xmm1
vfmaddpd %xmm2 , %xmm6 , %xmm10, %xmm2
vfmaddpd %xmm3 , %xmm7 , %xmm11, %xmm3


vmovsd 0 * SIZE(X), %xmm4
addq INCX, X
vmovsd 0 * SIZE(Y), %xmm8
addq INCY, Y
vmovsd 0 * SIZE(X), %xmm5
addq INCX, X
vmovsd 0 * SIZE(Y), %xmm9
addq INCY, Y

vmovsd 0 * SIZE(X), %xmm6
addq INCX, X
vmovsd 0 * SIZE(Y), %xmm10
addq INCY, Y
vmovsd 0 * SIZE(X), %xmm7
addq INCX, X
vmovsd 0 * SIZE(Y), %xmm11
addq INCY, Y

vfmaddpd %xmm0 , %xmm4 , %xmm8 , %xmm0
vfmaddpd %xmm1 , %xmm5 , %xmm9 , %xmm1
vfmaddpd %xmm2 , %xmm6 , %xmm10, %xmm2
vfmaddpd %xmm3 , %xmm7 , %xmm11, %xmm3

decq %rax
jg .L53
ALIGN_3

.L55:
movq N, %rax
andq $7, %rax
jle .L999
ALIGN_3

.L56:
vmovsd 0 * SIZE(X), %xmm4
addq INCX, X
vmovsd 0 * SIZE(Y), %xmm8
addq INCY, Y

vfmaddpd %xmm0 , %xmm4 , %xmm8 , %xmm0

decq %rax
jg .L56
ALIGN_3

.L999:
vaddpd %xmm1, %xmm0 , %xmm0
vaddpd %xmm3, %xmm2 , %xmm2
vaddpd %xmm2, %xmm0 , %xmm0

vhaddpd %xmm0, %xmm0 , %xmm0

RESTOREREGISTERS

ret

EPILOGUE

+ 302
- 304
kernel/x86_64/dgemm_ncopy_8_bulldozer.S
File diff suppressed because it is too large
View File


+ 1938
- 0
kernel/x86_64/dgemv_t_bulldozer.S
File diff suppressed because it is too large
View File


Write Preview
Loading…
Cancel
Save