|
kusano |
2b45e8 |
/*********************************************************************/
|
|
kusano |
2b45e8 |
/* Copyright 2009, 2010 The University of Texas at Austin. */
|
|
kusano |
2b45e8 |
/* All rights reserved. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* Redistribution and use in source and binary forms, with or */
|
|
kusano |
2b45e8 |
/* without modification, are permitted provided that the following */
|
|
kusano |
2b45e8 |
/* conditions are met: */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* 1. Redistributions of source code must retain the above */
|
|
kusano |
2b45e8 |
/* copyright notice, this list of conditions and the following */
|
|
kusano |
2b45e8 |
/* disclaimer. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* 2. Redistributions in binary form must reproduce the above */
|
|
kusano |
2b45e8 |
/* copyright notice, this list of conditions and the following */
|
|
kusano |
2b45e8 |
/* disclaimer in the documentation and/or other materials */
|
|
kusano |
2b45e8 |
/* provided with the distribution. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
|
|
kusano |
2b45e8 |
/* AUSTIN ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, */
|
|
kusano |
2b45e8 |
/* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
|
|
kusano |
2b45e8 |
/* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
|
|
kusano |
2b45e8 |
/* DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY OF TEXAS AT */
|
|
kusano |
2b45e8 |
/* AUSTIN OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, */
|
|
kusano |
2b45e8 |
/* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES */
|
|
kusano |
2b45e8 |
/* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE */
|
|
kusano |
2b45e8 |
/* GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR */
|
|
kusano |
2b45e8 |
/* BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF */
|
|
kusano |
2b45e8 |
/* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT */
|
|
kusano |
2b45e8 |
/* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT */
|
|
kusano |
2b45e8 |
/* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE */
|
|
kusano |
2b45e8 |
/* POSSIBILITY OF SUCH DAMAGE. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* The views and conclusions contained in the software and */
|
|
kusano |
2b45e8 |
/* documentation are those of the authors and should not be */
|
|
kusano |
2b45e8 |
/* interpreted as representing official policies, either expressed */
|
|
kusano |
2b45e8 |
/* or implied, of The University of Texas at Austin. */
|
|
kusano |
2b45e8 |
/*********************************************************************/
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define ASSEMBLER
|
|
kusano |
2b45e8 |
#include "common.h"
|
|
kusano |
2b45e8 |
#include "l2param.h"
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define P 32
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define STACKSIZE 80
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define ALPHA 8 + STACKSIZE(%rsp)
|
|
kusano |
2b45e8 |
#define OLD_INCX 24 + STACKSIZE(%rsp)
|
|
kusano |
2b45e8 |
#define OLD_Y 32 + STACKSIZE(%rsp)
|
|
kusano |
2b45e8 |
#define OLD_INCY 40 + STACKSIZE(%rsp)
|
|
kusano |
2b45e8 |
#define BUFFER 48 + STACKSIZE(%rsp)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define PLDA_M 56 (%rsp)
|
|
kusano |
2b45e8 |
#define IS 64 (%rsp)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define M %rdi
|
|
kusano |
2b45e8 |
#define N %rsi
|
|
kusano |
2b45e8 |
#define A %rcx
|
|
kusano |
2b45e8 |
#define LDA %r8
|
|
kusano |
2b45e8 |
#define X %r9
|
|
kusano |
2b45e8 |
#define INCX %rdx
|
|
kusano |
2b45e8 |
#define Y %rbp
|
|
kusano |
2b45e8 |
#define INCY %r10
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define TEMP %rax
|
|
kusano |
2b45e8 |
#define I %rax
|
|
kusano |
2b45e8 |
#define J %r11
|
|
kusano |
2b45e8 |
#define A1 %r12
|
|
kusano |
2b45e8 |
#define X1 %r13
|
|
kusano |
2b45e8 |
#define Y1 %r14
|
|
kusano |
2b45e8 |
#define XP %r15
|
|
kusano |
2b45e8 |
/* #define BUFFER %r15 */
|
|
kusano |
2b45e8 |
#define MIN_N %rbx
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PROLOGUE
|
|
kusano |
2b45e8 |
PROFCODE
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subq $STACKSIZE, %rsp
|
|
kusano |
2b45e8 |
movq %rbx, 0(%rsp)
|
|
kusano |
2b45e8 |
movq %rbp, 8(%rsp)
|
|
kusano |
2b45e8 |
movq %r12, 16(%rsp)
|
|
kusano |
2b45e8 |
movq %r13, 24(%rsp)
|
|
kusano |
2b45e8 |
movq %r14, 32(%rsp)
|
|
kusano |
2b45e8 |
movq %r15, 40(%rsp)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq OLD_INCX, INCX
|
|
kusano |
2b45e8 |
movq OLD_Y, Y
|
|
kusano |
2b45e8 |
movq OLD_INCY, INCY
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FLD ALPHA
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
salq $BASE_SHIFT, INCX
|
|
kusano |
2b45e8 |
salq $BASE_SHIFT, INCY
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq $0, IS
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
test M, M
|
|
kusano |
2b45e8 |
jle .L79 # goto END
|
|
kusano |
2b45e8 |
test N, N
|
|
kusano |
2b45e8 |
jle .L79 # goto END
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq LDA, %rax
|
|
kusano |
2b45e8 |
imulq $P, %rax # P * lda
|
|
kusano |
2b45e8 |
subq M ,%rax # P * lda - m
|
|
kusano |
2b45e8 |
salq $BASE_SHIFT, %rax
|
|
kusano |
2b45e8 |
movq %rax, PLDA_M
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
salq $BASE_SHIFT, LDA
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L32:
|
|
kusano |
2b45e8 |
movq $P, %rax
|
|
kusano |
2b45e8 |
movq N, MIN_N
|
|
kusano |
2b45e8 |
subq IS, MIN_N
|
|
kusano |
2b45e8 |
cmpq %rax, MIN_N
|
|
kusano |
2b45e8 |
cmovg %rax, MIN_N
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq IS, XP
|
|
kusano |
2b45e8 |
salq $BASE_SHIFT, XP
|
|
kusano |
2b45e8 |
leaq (X,XP, 1), XP
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
cmpq $SIZE, INCX
|
|
kusano |
2b45e8 |
je .L34 # if incx == 1 goto L34
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq BUFFER, XP
|
|
kusano |
2b45e8 |
movq XP, X1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq MIN_N, I
|
|
kusano |
2b45e8 |
sarq $2,I
|
|
kusano |
2b45e8 |
jle .L35
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L36:
|
|
kusano |
2b45e8 |
FLD (X)
|
|
kusano |
2b45e8 |
addq INCX,X
|
|
kusano |
2b45e8 |
FLD (X)
|
|
kusano |
2b45e8 |
addq INCX,X
|
|
kusano |
2b45e8 |
FLD (X)
|
|
kusano |
2b45e8 |
addq INCX,X
|
|
kusano |
2b45e8 |
FLD (X)
|
|
kusano |
2b45e8 |
addq INCX,X
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FST 3 * SIZE(X1)
|
|
kusano |
2b45e8 |
FST 2 * SIZE(X1)
|
|
kusano |
2b45e8 |
FST 1 * SIZE(X1)
|
|
kusano |
2b45e8 |
FST 0 * SIZE(X1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $4 * SIZE, X1
|
|
kusano |
2b45e8 |
decq I
|
|
kusano |
2b45e8 |
jg .L36
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L35:
|
|
kusano |
2b45e8 |
movq MIN_N, I
|
|
kusano |
2b45e8 |
andq $3, I
|
|
kusano |
2b45e8 |
jle .L34
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L42:
|
|
kusano |
2b45e8 |
FLD (X)
|
|
kusano |
2b45e8 |
addq INCX, X
|
|
kusano |
2b45e8 |
FST (X1)
|
|
kusano |
2b45e8 |
addq $SIZE, X1
|
|
kusano |
2b45e8 |
decq I
|
|
kusano |
2b45e8 |
jg .L42
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
/* Main Routine */
|
|
kusano |
2b45e8 |
.L34:
|
|
kusano |
2b45e8 |
movq Y, Y1
|
|
kusano |
2b45e8 |
movq M, J
|
|
kusano |
2b45e8 |
sarq $2, J
|
|
kusano |
2b45e8 |
jle .L47
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L48:
|
|
kusano |
2b45e8 |
movq A, A1 # a_offset = a
|
|
kusano |
2b45e8 |
fldz
|
|
kusano |
2b45e8 |
addq $4 * SIZE, A # a += 4
|
|
kusano |
2b45e8 |
fldz
|
|
kusano |
2b45e8 |
movq XP, X1 # b_offset = xp
|
|
kusano |
2b45e8 |
fldz
|
|
kusano |
2b45e8 |
movq MIN_N, I # i = min_n
|
|
kusano |
2b45e8 |
fldz
|
|
kusano |
2b45e8 |
FLD (X1) # bt1 = b_offset
|
|
kusano |
2b45e8 |
sarq $1, I
|
|
kusano |
2b45e8 |
jle .L51
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L80:
|
|
kusano |
2b45e8 |
FLD 0 * SIZE(A1) # at1 = *(a_offset + 0)
|
|
kusano |
2b45e8 |
fmul %st(1), %st # at1 *= bt1
|
|
kusano |
2b45e8 |
faddp %st, %st(2) # ct1 += at1
|
|
kusano |
2b45e8 |
FLD 1 * SIZE(A1) # at1 = *(a_offset + 1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmul %st(1), %st # at1 *= bt1
|
|
kusano |
2b45e8 |
faddp %st, %st(3) # ct2 += at1
|
|
kusano |
2b45e8 |
FLD 2 * SIZE(A1) # at1 = *(a_offset + 2)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmul %st(1), %st # at1 *= bt1
|
|
kusano |
2b45e8 |
faddp %st, %st(4) # ct3 += at1
|
|
kusano |
2b45e8 |
FLD 3 * SIZE(A1) # bt1 *= *(a_offset + 3)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmulp %st, %st(1)
|
|
kusano |
2b45e8 |
faddp %st, %st(4) # ct4 += at1
|
|
kusano |
2b45e8 |
FLD 1 * SIZE(X1) # bt1 = b_offset
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq LDA, A1 # a_offset += lda
|
|
kusano |
2b45e8 |
FLD 0 * SIZE(A1) # at1 = *(a_offset + 0)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmul %st(1), %st # at1 *= bt1
|
|
kusano |
2b45e8 |
faddp %st, %st(2) # ct1 += at1
|
|
kusano |
2b45e8 |
FLD 1 * SIZE(A1) # at1 = *(a_offset + 1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmul %st(1), %st # at1 *= bt1
|
|
kusano |
2b45e8 |
faddp %st, %st(3) # ct2 += at1
|
|
kusano |
2b45e8 |
FLD 2 * SIZE(A1) # at1 = *(a_offset + 2)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmul %st(1), %st # at1 *= bt1
|
|
kusano |
2b45e8 |
faddp %st, %st(4) # ct3 += at1
|
|
kusano |
2b45e8 |
FLD 3 * SIZE(A1) # bt1 *= *(a_offset + 3)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmulp %st, %st(1)
|
|
kusano |
2b45e8 |
addq LDA, A1
|
|
kusano |
2b45e8 |
faddp %st, %st(4) # ct4 += at1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FLD 2 * SIZE(X1) # bt1 = b_offset
|
|
kusano |
2b45e8 |
addq $2 * SIZE, X1 # b_offset += 2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
decq I
|
|
kusano |
2b45e8 |
jg .L80
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L51:
|
|
kusano |
2b45e8 |
movq MIN_N, I
|
|
kusano |
2b45e8 |
andq $1, I
|
|
kusano |
2b45e8 |
je .L57
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FLD 0 * SIZE(A1) # at1 = *(a_offset + 0)
|
|
kusano |
2b45e8 |
fmul %st(1), %st # at1 *= bt1
|
|
kusano |
2b45e8 |
faddp %st, %st(2) # ct1 += at1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FLD 1 * SIZE(A1) # at1 = *(a_offset + 1)
|
|
kusano |
2b45e8 |
fmul %st(1), %st # at1 *= bt1
|
|
kusano |
2b45e8 |
faddp %st, %st(3) # ct2 += at1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FLD 2 * SIZE(A1) # at1 = *(a_offset + 2)
|
|
kusano |
2b45e8 |
fmul %st(1), %st # at1 *= bt1
|
|
kusano |
2b45e8 |
faddp %st, %st(4) # ct3 += at1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FLD 3 * SIZE(A1) # bt1 *= *(a_offset + 3)
|
|
kusano |
2b45e8 |
fmulp %st, %st(1)
|
|
kusano |
2b45e8 |
faddp %st, %st(4) # ct4 += at1
|
|
kusano |
2b45e8 |
fldz
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L57:
|
|
kusano |
2b45e8 |
ffreep %st(0)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fxch %st(4)
|
|
kusano |
2b45e8 |
fmul %st, %st(4)
|
|
kusano |
2b45e8 |
fmul %st, %st(1)
|
|
kusano |
2b45e8 |
fmul %st, %st(2)
|
|
kusano |
2b45e8 |
fmul %st, %st(3)
|
|
kusano |
2b45e8 |
fxch %st(4)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FLD (Y1)
|
|
kusano |
2b45e8 |
faddp %st, %st(1)
|
|
kusano |
2b45e8 |
FST (Y1)
|
|
kusano |
2b45e8 |
addq INCY, Y1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FLD (Y1)
|
|
kusano |
2b45e8 |
faddp %st, %st(1)
|
|
kusano |
2b45e8 |
FST (Y1)
|
|
kusano |
2b45e8 |
addq INCY, Y1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FLD (Y1)
|
|
kusano |
2b45e8 |
faddp %st, %st(1)
|
|
kusano |
2b45e8 |
FST (Y1)
|
|
kusano |
2b45e8 |
addq INCY, Y1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FLD (Y1)
|
|
kusano |
2b45e8 |
faddp %st, %st(1)
|
|
kusano |
2b45e8 |
FST (Y1)
|
|
kusano |
2b45e8 |
addq INCY, Y1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
decq J # j --
|
|
kusano |
2b45e8 |
jg .L48
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L47:
|
|
kusano |
2b45e8 |
movq M, J
|
|
kusano |
2b45e8 |
andq $3, J # j = (m & 3)
|
|
kusano |
2b45e8 |
jle .L60
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L61:
|
|
kusano |
2b45e8 |
movq A, A1 # a_offset = a
|
|
kusano |
2b45e8 |
fldz
|
|
kusano |
2b45e8 |
addq $SIZE, A # a++
|
|
kusano |
2b45e8 |
fldz
|
|
kusano |
2b45e8 |
movq XP, X1
|
|
kusano |
2b45e8 |
fldz
|
|
kusano |
2b45e8 |
fldz
|
|
kusano |
2b45e8 |
movq MIN_N, I
|
|
kusano |
2b45e8 |
sarq $3, I
|
|
kusano |
2b45e8 |
jle .L64
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L65:
|
|
kusano |
2b45e8 |
FLD 0 * SIZE(X1)
|
|
kusano |
2b45e8 |
FLD (A1)
|
|
kusano |
2b45e8 |
fmulp %st, %st(1)
|
|
kusano |
2b45e8 |
faddp %st, %st(1)
|
|
kusano |
2b45e8 |
addq LDA, A1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FLD 1 * SIZE(X1)
|
|
kusano |
2b45e8 |
FLD (A1)
|
|
kusano |
2b45e8 |
fmulp %st, %st(1)
|
|
kusano |
2b45e8 |
faddp %st, %st(2)
|
|
kusano |
2b45e8 |
addq LDA ,A1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FLD 2 * SIZE(X1)
|
|
kusano |
2b45e8 |
FLD (A1)
|
|
kusano |
2b45e8 |
fmulp %st, %st(1)
|
|
kusano |
2b45e8 |
faddp %st, %st(3)
|
|
kusano |
2b45e8 |
addq LDA, A1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FLD 3 * SIZE(X1)
|
|
kusano |
2b45e8 |
FLD (A1)
|
|
kusano |
2b45e8 |
fmulp %st, %st(1)
|
|
kusano |
2b45e8 |
faddp %st, %st(4)
|
|
kusano |
2b45e8 |
addq LDA, A1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FLD 4 * SIZE(X1)
|
|
kusano |
2b45e8 |
FLD (A1)
|
|
kusano |
2b45e8 |
fmulp %st, %st(1)
|
|
kusano |
2b45e8 |
faddp %st,%st(1)
|
|
kusano |
2b45e8 |
addq LDA, A1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FLD 5 * SIZE(X1)
|
|
kusano |
2b45e8 |
FLD (A1)
|
|
kusano |
2b45e8 |
fmulp %st, %st(1)
|
|
kusano |
2b45e8 |
faddp %st, %st(2)
|
|
kusano |
2b45e8 |
addq LDA, A1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FLD 6 * SIZE(X1)
|
|
kusano |
2b45e8 |
FLD (A1)
|
|
kusano |
2b45e8 |
fmulp %st, %st(1)
|
|
kusano |
2b45e8 |
faddp %st,%st(3)
|
|
kusano |
2b45e8 |
addq LDA, A1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FLD 7 * SIZE(X1)
|
|
kusano |
2b45e8 |
FLD (A1)
|
|
kusano |
2b45e8 |
fmulp %st, %st(1)
|
|
kusano |
2b45e8 |
faddp %st,%st(4)
|
|
kusano |
2b45e8 |
addq LDA, A1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $8 * SIZE, X1
|
|
kusano |
2b45e8 |
decq I
|
|
kusano |
2b45e8 |
jg .L65
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L64:
|
|
kusano |
2b45e8 |
movq MIN_N,I
|
|
kusano |
2b45e8 |
andq $7, I
|
|
kusano |
2b45e8 |
jle .L70
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L71:
|
|
kusano |
2b45e8 |
FLD (X1)
|
|
kusano |
2b45e8 |
addq $SIZE, X1
|
|
kusano |
2b45e8 |
FLD (A1)
|
|
kusano |
2b45e8 |
fmulp %st, %st(1)
|
|
kusano |
2b45e8 |
addq LDA, A1 # a_offset += lda
|
|
kusano |
2b45e8 |
faddp %st, %st(1)
|
|
kusano |
2b45e8 |
decq I
|
|
kusano |
2b45e8 |
jg .L71
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L70:
|
|
kusano |
2b45e8 |
faddp %st, %st(1)
|
|
kusano |
2b45e8 |
faddp %st, %st(1)
|
|
kusano |
2b45e8 |
faddp %st, %st(1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmul %st(1), %st
|
|
kusano |
2b45e8 |
FLD (Y1)
|
|
kusano |
2b45e8 |
faddp %st, %st(1)
|
|
kusano |
2b45e8 |
FST (Y1)
|
|
kusano |
2b45e8 |
addq INCY, Y1
|
|
kusano |
2b45e8 |
decq J
|
|
kusano |
2b45e8 |
jg .L61
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L60:
|
|
kusano |
2b45e8 |
addq PLDA_M, A
|
|
kusano |
2b45e8 |
addq $P, IS
|
|
kusano |
2b45e8 |
cmpq N, IS
|
|
kusano |
2b45e8 |
jl .L32
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L79:
|
|
kusano |
2b45e8 |
EMMS
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq 0(%rsp), %rbx
|
|
kusano |
2b45e8 |
movq 8(%rsp), %rbp
|
|
kusano |
2b45e8 |
movq 16(%rsp), %r12
|
|
kusano |
2b45e8 |
movq 24(%rsp), %r13
|
|
kusano |
2b45e8 |
movq 32(%rsp), %r14
|
|
kusano |
2b45e8 |
movq 40(%rsp), %r15
|
|
kusano |
2b45e8 |
addq $STACKSIZE, %rsp
|
|
kusano |
2b45e8 |
ret
|
|
kusano |
2b45e8 |
EPILOGUE
|